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 (c) 2014, Nexenta Systems, Inc. All rights reserved. */
30 #include <sys/dmu_impl.h>
31 #include <sys/dmu_tx.h>
33 #include <sys/dnode.h>
34 #include <sys/zfs_context.h>
35 #include <sys/dmu_objset.h>
36 #include <sys/dmu_traverse.h>
37 #include <sys/dsl_dataset.h>
38 #include <sys/dsl_dir.h>
39 #include <sys/dsl_pool.h>
40 #include <sys/dsl_synctask.h>
41 #include <sys/dsl_prop.h>
42 #include <sys/dmu_zfetch.h>
43 #include <sys/zfs_ioctl.h>
45 #include <sys/zio_checksum.h>
46 #include <sys/zio_compress.h>
48 #include <sys/zfeature.h>
51 #include <sys/racct.h>
53 #include <sys/zfs_znode.h>
57 * Enable/disable nopwrite feature.
59 int zfs_nopwrite_enabled = 1;
60 SYSCTL_DECL(_vfs_zfs);
61 SYSCTL_INT(_vfs_zfs, OID_AUTO, nopwrite_enabled, CTLFLAG_RDTUN,
62 &zfs_nopwrite_enabled, 0, "Enable nopwrite feature");
65 * Tunable to control percentage of dirtied blocks from frees in one TXG.
66 * After this threshold is crossed, additional dirty blocks from frees
67 * wait until the next TXG.
68 * A value of zero will disable this throttle.
70 uint32_t zfs_per_txg_dirty_frees_percent = 30;
71 SYSCTL_INT(_vfs_zfs, OID_AUTO, per_txg_dirty_frees_percent, CTLFLAG_RWTUN,
72 &zfs_per_txg_dirty_frees_percent, 0, "Percentage of dirtied blocks from frees in one txg");
74 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
75 { DMU_BSWAP_UINT8, TRUE, "unallocated" },
76 { DMU_BSWAP_ZAP, TRUE, "object directory" },
77 { DMU_BSWAP_UINT64, TRUE, "object array" },
78 { DMU_BSWAP_UINT8, TRUE, "packed nvlist" },
79 { DMU_BSWAP_UINT64, TRUE, "packed nvlist size" },
80 { DMU_BSWAP_UINT64, TRUE, "bpobj" },
81 { DMU_BSWAP_UINT64, TRUE, "bpobj header" },
82 { DMU_BSWAP_UINT64, TRUE, "SPA space map header" },
83 { DMU_BSWAP_UINT64, TRUE, "SPA space map" },
84 { DMU_BSWAP_UINT64, TRUE, "ZIL intent log" },
85 { DMU_BSWAP_DNODE, TRUE, "DMU dnode" },
86 { DMU_BSWAP_OBJSET, TRUE, "DMU objset" },
87 { DMU_BSWAP_UINT64, TRUE, "DSL directory" },
88 { DMU_BSWAP_ZAP, TRUE, "DSL directory child map"},
89 { DMU_BSWAP_ZAP, TRUE, "DSL dataset snap map" },
90 { DMU_BSWAP_ZAP, TRUE, "DSL props" },
91 { DMU_BSWAP_UINT64, TRUE, "DSL dataset" },
92 { DMU_BSWAP_ZNODE, TRUE, "ZFS znode" },
93 { DMU_BSWAP_OLDACL, TRUE, "ZFS V0 ACL" },
94 { DMU_BSWAP_UINT8, FALSE, "ZFS plain file" },
95 { DMU_BSWAP_ZAP, TRUE, "ZFS directory" },
96 { DMU_BSWAP_ZAP, TRUE, "ZFS master node" },
97 { DMU_BSWAP_ZAP, TRUE, "ZFS delete queue" },
98 { DMU_BSWAP_UINT8, FALSE, "zvol object" },
99 { DMU_BSWAP_ZAP, TRUE, "zvol prop" },
100 { DMU_BSWAP_UINT8, FALSE, "other uint8[]" },
101 { DMU_BSWAP_UINT64, FALSE, "other uint64[]" },
102 { DMU_BSWAP_ZAP, TRUE, "other ZAP" },
103 { DMU_BSWAP_ZAP, TRUE, "persistent error log" },
104 { DMU_BSWAP_UINT8, TRUE, "SPA history" },
105 { DMU_BSWAP_UINT64, TRUE, "SPA history offsets" },
106 { DMU_BSWAP_ZAP, TRUE, "Pool properties" },
107 { DMU_BSWAP_ZAP, TRUE, "DSL permissions" },
108 { DMU_BSWAP_ACL, TRUE, "ZFS ACL" },
109 { DMU_BSWAP_UINT8, TRUE, "ZFS SYSACL" },
110 { DMU_BSWAP_UINT8, TRUE, "FUID table" },
111 { DMU_BSWAP_UINT64, TRUE, "FUID table size" },
112 { DMU_BSWAP_ZAP, TRUE, "DSL dataset next clones"},
113 { DMU_BSWAP_ZAP, TRUE, "scan work queue" },
114 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group used" },
115 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group quota" },
116 { DMU_BSWAP_ZAP, TRUE, "snapshot refcount tags"},
117 { DMU_BSWAP_ZAP, TRUE, "DDT ZAP algorithm" },
118 { DMU_BSWAP_ZAP, TRUE, "DDT statistics" },
119 { DMU_BSWAP_UINT8, TRUE, "System attributes" },
120 { DMU_BSWAP_ZAP, TRUE, "SA master node" },
121 { DMU_BSWAP_ZAP, TRUE, "SA attr registration" },
122 { DMU_BSWAP_ZAP, TRUE, "SA attr layouts" },
123 { DMU_BSWAP_ZAP, TRUE, "scan translations" },
124 { DMU_BSWAP_UINT8, FALSE, "deduplicated block" },
125 { DMU_BSWAP_ZAP, TRUE, "DSL deadlist map" },
126 { DMU_BSWAP_UINT64, TRUE, "DSL deadlist map hdr" },
127 { DMU_BSWAP_ZAP, TRUE, "DSL dir clones" },
128 { DMU_BSWAP_UINT64, TRUE, "bpobj subobj" }
131 const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
132 { byteswap_uint8_array, "uint8" },
133 { byteswap_uint16_array, "uint16" },
134 { byteswap_uint32_array, "uint32" },
135 { byteswap_uint64_array, "uint64" },
136 { zap_byteswap, "zap" },
137 { dnode_buf_byteswap, "dnode" },
138 { dmu_objset_byteswap, "objset" },
139 { zfs_znode_byteswap, "znode" },
140 { zfs_oldacl_byteswap, "oldacl" },
141 { zfs_acl_byteswap, "acl" }
145 dmu_buf_hold_noread_by_dnode(dnode_t *dn, uint64_t offset,
146 void *tag, dmu_buf_t **dbp)
151 blkid = dbuf_whichblock(dn, 0, offset);
152 rw_enter(&dn->dn_struct_rwlock, RW_READER);
153 db = dbuf_hold(dn, blkid, tag);
154 rw_exit(&dn->dn_struct_rwlock);
158 return (SET_ERROR(EIO));
165 dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset,
166 void *tag, dmu_buf_t **dbp)
173 err = dnode_hold(os, object, FTAG, &dn);
176 blkid = dbuf_whichblock(dn, 0, offset);
177 rw_enter(&dn->dn_struct_rwlock, RW_READER);
178 db = dbuf_hold(dn, blkid, tag);
179 rw_exit(&dn->dn_struct_rwlock);
180 dnode_rele(dn, FTAG);
184 return (SET_ERROR(EIO));
192 dmu_buf_hold_by_dnode(dnode_t *dn, uint64_t offset,
193 void *tag, dmu_buf_t **dbp, int flags)
196 int db_flags = DB_RF_CANFAIL;
198 if (flags & DMU_READ_NO_PREFETCH)
199 db_flags |= DB_RF_NOPREFETCH;
201 err = dmu_buf_hold_noread_by_dnode(dn, offset, tag, dbp);
203 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
204 err = dbuf_read(db, NULL, db_flags);
215 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
216 void *tag, dmu_buf_t **dbp, int flags)
219 int db_flags = DB_RF_CANFAIL;
221 if (flags & DMU_READ_NO_PREFETCH)
222 db_flags |= DB_RF_NOPREFETCH;
224 err = dmu_buf_hold_noread(os, object, offset, tag, dbp);
226 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
227 err = dbuf_read(db, NULL, db_flags);
240 return (DN_MAX_BONUSLEN);
244 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
246 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
253 if (dn->dn_bonus != db) {
254 error = SET_ERROR(EINVAL);
255 } else if (newsize < 0 || newsize > db_fake->db_size) {
256 error = SET_ERROR(EINVAL);
258 dnode_setbonuslen(dn, newsize, tx);
267 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
269 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
276 if (!DMU_OT_IS_VALID(type)) {
277 error = SET_ERROR(EINVAL);
278 } else if (dn->dn_bonus != db) {
279 error = SET_ERROR(EINVAL);
281 dnode_setbonus_type(dn, type, tx);
290 dmu_get_bonustype(dmu_buf_t *db_fake)
292 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
294 dmu_object_type_t type;
298 type = dn->dn_bonustype;
305 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
310 error = dnode_hold(os, object, FTAG, &dn);
311 dbuf_rm_spill(dn, tx);
312 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
313 dnode_rm_spill(dn, tx);
314 rw_exit(&dn->dn_struct_rwlock);
315 dnode_rele(dn, FTAG);
320 * returns ENOENT, EIO, or 0.
323 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
329 error = dnode_hold(os, object, FTAG, &dn);
333 rw_enter(&dn->dn_struct_rwlock, RW_READER);
334 if (dn->dn_bonus == NULL) {
335 rw_exit(&dn->dn_struct_rwlock);
336 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
337 if (dn->dn_bonus == NULL)
338 dbuf_create_bonus(dn);
342 /* as long as the bonus buf is held, the dnode will be held */
343 if (refcount_add(&db->db_holds, tag) == 1) {
344 VERIFY(dnode_add_ref(dn, db));
345 atomic_inc_32(&dn->dn_dbufs_count);
349 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
350 * hold and incrementing the dbuf count to ensure that dnode_move() sees
351 * a dnode hold for every dbuf.
353 rw_exit(&dn->dn_struct_rwlock);
355 dnode_rele(dn, FTAG);
357 VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH));
364 * returns ENOENT, EIO, or 0.
366 * This interface will allocate a blank spill dbuf when a spill blk
367 * doesn't already exist on the dnode.
369 * if you only want to find an already existing spill db, then
370 * dmu_spill_hold_existing() should be used.
373 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
375 dmu_buf_impl_t *db = NULL;
378 if ((flags & DB_RF_HAVESTRUCT) == 0)
379 rw_enter(&dn->dn_struct_rwlock, RW_READER);
381 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
383 if ((flags & DB_RF_HAVESTRUCT) == 0)
384 rw_exit(&dn->dn_struct_rwlock);
387 err = dbuf_read(db, NULL, flags);
396 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
398 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
405 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
406 err = SET_ERROR(EINVAL);
408 rw_enter(&dn->dn_struct_rwlock, RW_READER);
410 if (!dn->dn_have_spill) {
411 err = SET_ERROR(ENOENT);
413 err = dmu_spill_hold_by_dnode(dn,
414 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
417 rw_exit(&dn->dn_struct_rwlock);
425 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
427 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
433 err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp);
440 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
441 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
442 * and can induce severe lock contention when writing to several files
443 * whose dnodes are in the same block.
446 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
447 boolean_t read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
450 uint64_t blkid, nblks, i;
455 ASSERT(length <= DMU_MAX_ACCESS);
458 * Note: We directly notify the prefetch code of this read, so that
459 * we can tell it about the multi-block read. dbuf_read() only knows
460 * about the one block it is accessing.
462 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT |
465 rw_enter(&dn->dn_struct_rwlock, RW_READER);
466 if (dn->dn_datablkshift) {
467 int blkshift = dn->dn_datablkshift;
468 nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) -
469 P2ALIGN(offset, 1ULL << blkshift)) >> blkshift;
471 if (offset + length > dn->dn_datablksz) {
472 zfs_panic_recover("zfs: accessing past end of object "
473 "%llx/%llx (size=%u access=%llu+%llu)",
474 (longlong_t)dn->dn_objset->
475 os_dsl_dataset->ds_object,
476 (longlong_t)dn->dn_object, dn->dn_datablksz,
477 (longlong_t)offset, (longlong_t)length);
478 rw_exit(&dn->dn_struct_rwlock);
479 return (SET_ERROR(EIO));
483 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
485 #if defined(_KERNEL) && defined(RACCT)
486 if (racct_enable && !read) {
488 racct_add_force(curproc, RACCT_WRITEBPS, length);
489 racct_add_force(curproc, RACCT_WRITEIOPS, nblks);
490 PROC_UNLOCK(curproc);
494 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
495 blkid = dbuf_whichblock(dn, 0, offset);
496 for (i = 0; i < nblks; i++) {
497 dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag);
499 rw_exit(&dn->dn_struct_rwlock);
500 dmu_buf_rele_array(dbp, nblks, tag);
502 return (SET_ERROR(EIO));
505 /* initiate async i/o */
507 (void) dbuf_read(db, zio, dbuf_flags);
510 curthread->td_ru.ru_oublock++;
515 if ((flags & DMU_READ_NO_PREFETCH) == 0 &&
516 DNODE_META_IS_CACHEABLE(dn) && length <= zfetch_array_rd_sz) {
517 dmu_zfetch(&dn->dn_zfetch, blkid, nblks,
518 read && DNODE_IS_CACHEABLE(dn));
520 rw_exit(&dn->dn_struct_rwlock);
522 /* wait for async i/o */
525 dmu_buf_rele_array(dbp, nblks, tag);
529 /* wait for other io to complete */
531 for (i = 0; i < nblks; i++) {
532 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
533 mutex_enter(&db->db_mtx);
534 while (db->db_state == DB_READ ||
535 db->db_state == DB_FILL)
536 cv_wait(&db->db_changed, &db->db_mtx);
537 if (db->db_state == DB_UNCACHED)
538 err = SET_ERROR(EIO);
539 mutex_exit(&db->db_mtx);
541 dmu_buf_rele_array(dbp, nblks, tag);
553 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
554 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
559 err = dnode_hold(os, object, FTAG, &dn);
563 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
564 numbufsp, dbpp, DMU_READ_PREFETCH);
566 dnode_rele(dn, FTAG);
572 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
573 uint64_t length, boolean_t read, void *tag, int *numbufsp,
576 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
582 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
583 numbufsp, dbpp, DMU_READ_PREFETCH);
590 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
593 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
598 for (i = 0; i < numbufs; i++) {
600 dbuf_rele(dbp[i], tag);
603 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
607 * Issue prefetch i/os for the given blocks. If level is greater than 0, the
608 * indirect blocks prefeteched will be those that point to the blocks containing
609 * the data starting at offset, and continuing to offset + len.
611 * Note that if the indirect blocks above the blocks being prefetched are not in
612 * cache, they will be asychronously read in.
615 dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset,
616 uint64_t len, zio_priority_t pri)
622 if (len == 0) { /* they're interested in the bonus buffer */
623 dn = DMU_META_DNODE(os);
625 if (object == 0 || object >= DN_MAX_OBJECT)
628 rw_enter(&dn->dn_struct_rwlock, RW_READER);
629 blkid = dbuf_whichblock(dn, level,
630 object * sizeof (dnode_phys_t));
631 dbuf_prefetch(dn, level, blkid, pri, 0);
632 rw_exit(&dn->dn_struct_rwlock);
637 * XXX - Note, if the dnode for the requested object is not
638 * already cached, we will do a *synchronous* read in the
639 * dnode_hold() call. The same is true for any indirects.
641 err = dnode_hold(os, object, FTAG, &dn);
645 rw_enter(&dn->dn_struct_rwlock, RW_READER);
647 * offset + len - 1 is the last byte we want to prefetch for, and offset
648 * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the
649 * last block we want to prefetch, and dbuf_whichblock(dn, level,
650 * offset) is the first. Then the number we need to prefetch is the
653 if (level > 0 || dn->dn_datablkshift != 0) {
654 nblks = dbuf_whichblock(dn, level, offset + len - 1) -
655 dbuf_whichblock(dn, level, offset) + 1;
657 nblks = (offset < dn->dn_datablksz);
661 blkid = dbuf_whichblock(dn, level, offset);
662 for (int i = 0; i < nblks; i++)
663 dbuf_prefetch(dn, level, blkid + i, pri, 0);
666 rw_exit(&dn->dn_struct_rwlock);
668 dnode_rele(dn, FTAG);
672 * Get the next "chunk" of file data to free. We traverse the file from
673 * the end so that the file gets shorter over time (if we crashes in the
674 * middle, this will leave us in a better state). We find allocated file
675 * data by simply searching the allocated level 1 indirects.
677 * On input, *start should be the first offset that does not need to be
678 * freed (e.g. "offset + length"). On return, *start will be the first
679 * offset that should be freed.
682 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum)
684 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
685 /* bytes of data covered by a level-1 indirect block */
687 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
689 ASSERT3U(minimum, <=, *start);
691 if (*start - minimum <= iblkrange * maxblks) {
695 ASSERT(ISP2(iblkrange));
697 for (uint64_t blks = 0; *start > minimum && blks < maxblks; blks++) {
701 * dnode_next_offset(BACKWARDS) will find an allocated L1
702 * indirect block at or before the input offset. We must
703 * decrement *start so that it is at the end of the region
707 err = dnode_next_offset(dn,
708 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
710 /* if there are no indirect blocks before start, we are done */
714 } else if (err != 0) {
718 /* set start to the beginning of this L1 indirect */
719 *start = P2ALIGN(*start, iblkrange);
721 if (*start < minimum)
727 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
730 uint64_t object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
732 uint64_t dirty_frees_threshold;
733 dsl_pool_t *dp = dmu_objset_pool(os);
735 if (offset >= object_size)
738 if (zfs_per_txg_dirty_frees_percent <= 100)
739 dirty_frees_threshold =
740 zfs_per_txg_dirty_frees_percent * zfs_dirty_data_max / 100;
742 dirty_frees_threshold = zfs_dirty_data_max / 4;
744 if (length == DMU_OBJECT_END || offset + length > object_size)
745 length = object_size - offset;
747 while (length != 0) {
748 uint64_t chunk_end, chunk_begin, chunk_len;
749 uint64_t long_free_dirty_all_txgs = 0;
752 chunk_end = chunk_begin = offset + length;
754 /* move chunk_begin backwards to the beginning of this chunk */
755 err = get_next_chunk(dn, &chunk_begin, offset);
758 ASSERT3U(chunk_begin, >=, offset);
759 ASSERT3U(chunk_begin, <=, chunk_end);
761 chunk_len = chunk_end - chunk_begin;
763 mutex_enter(&dp->dp_lock);
764 for (int t = 0; t < TXG_SIZE; t++) {
765 long_free_dirty_all_txgs +=
766 dp->dp_long_free_dirty_pertxg[t];
768 mutex_exit(&dp->dp_lock);
771 * To avoid filling up a TXG with just frees wait for
772 * the next TXG to open before freeing more chunks if
773 * we have reached the threshold of frees
775 if (dirty_frees_threshold != 0 &&
776 long_free_dirty_all_txgs >= dirty_frees_threshold) {
777 txg_wait_open(dp, 0);
781 tx = dmu_tx_create(os);
782 dmu_tx_hold_free(tx, dn->dn_object, chunk_begin, chunk_len);
785 * Mark this transaction as typically resulting in a net
786 * reduction in space used.
788 dmu_tx_mark_netfree(tx);
789 err = dmu_tx_assign(tx, TXG_WAIT);
795 mutex_enter(&dp->dp_lock);
796 dp->dp_long_free_dirty_pertxg[dmu_tx_get_txg(tx) & TXG_MASK] +=
798 mutex_exit(&dp->dp_lock);
799 DTRACE_PROBE3(free__long__range,
800 uint64_t, long_free_dirty_all_txgs, uint64_t, chunk_len,
801 uint64_t, dmu_tx_get_txg(tx));
802 dnode_free_range(dn, chunk_begin, chunk_len, tx);
811 dmu_free_long_range(objset_t *os, uint64_t object,
812 uint64_t offset, uint64_t length)
817 err = dnode_hold(os, object, FTAG, &dn);
820 err = dmu_free_long_range_impl(os, dn, offset, length);
823 * It is important to zero out the maxblkid when freeing the entire
824 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
825 * will take the fast path, and (b) dnode_reallocate() can verify
826 * that the entire file has been freed.
828 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
831 dnode_rele(dn, FTAG);
836 dmu_free_long_object(objset_t *os, uint64_t object)
841 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
845 tx = dmu_tx_create(os);
846 dmu_tx_hold_bonus(tx, object);
847 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
848 dmu_tx_mark_netfree(tx);
849 err = dmu_tx_assign(tx, TXG_WAIT);
851 err = dmu_object_free(os, object, tx);
861 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
862 uint64_t size, dmu_tx_t *tx)
865 int err = dnode_hold(os, object, FTAG, &dn);
868 ASSERT(offset < UINT64_MAX);
869 ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
870 dnode_free_range(dn, offset, size, tx);
871 dnode_rele(dn, FTAG);
876 dmu_read_impl(dnode_t *dn, uint64_t offset, uint64_t size,
877 void *buf, uint32_t flags)
880 int numbufs, err = 0;
883 * Deal with odd block sizes, where there can't be data past the first
884 * block. If we ever do the tail block optimization, we will need to
885 * handle that here as well.
887 if (dn->dn_maxblkid == 0) {
888 int newsz = offset > dn->dn_datablksz ? 0 :
889 MIN(size, dn->dn_datablksz - offset);
890 bzero((char *)buf + newsz, size - newsz);
895 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
899 * NB: we could do this block-at-a-time, but it's nice
900 * to be reading in parallel.
902 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
903 TRUE, FTAG, &numbufs, &dbp, flags);
907 for (i = 0; i < numbufs; i++) {
910 dmu_buf_t *db = dbp[i];
914 bufoff = offset - db->db_offset;
915 tocpy = (int)MIN(db->db_size - bufoff, size);
917 bcopy((char *)db->db_data + bufoff, buf, tocpy);
921 buf = (char *)buf + tocpy;
923 dmu_buf_rele_array(dbp, numbufs, FTAG);
929 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
930 void *buf, uint32_t flags)
935 err = dnode_hold(os, object, FTAG, &dn);
939 err = dmu_read_impl(dn, offset, size, buf, flags);
940 dnode_rele(dn, FTAG);
945 dmu_read_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, void *buf,
948 return (dmu_read_impl(dn, offset, size, buf, flags));
952 dmu_write_impl(dmu_buf_t **dbp, int numbufs, uint64_t offset, uint64_t size,
953 const void *buf, dmu_tx_t *tx)
957 for (i = 0; i < numbufs; i++) {
960 dmu_buf_t *db = dbp[i];
964 bufoff = offset - db->db_offset;
965 tocpy = (int)MIN(db->db_size - bufoff, size);
967 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
969 if (tocpy == db->db_size)
970 dmu_buf_will_fill(db, tx);
972 dmu_buf_will_dirty(db, tx);
974 bcopy(buf, (char *)db->db_data + bufoff, tocpy);
976 if (tocpy == db->db_size)
977 dmu_buf_fill_done(db, tx);
981 buf = (char *)buf + tocpy;
986 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
987 const void *buf, dmu_tx_t *tx)
995 VERIFY0(dmu_buf_hold_array(os, object, offset, size,
996 FALSE, FTAG, &numbufs, &dbp));
997 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
998 dmu_buf_rele_array(dbp, numbufs, FTAG);
1002 dmu_write_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size,
1003 const void *buf, dmu_tx_t *tx)
1011 VERIFY0(dmu_buf_hold_array_by_dnode(dn, offset, size,
1012 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH));
1013 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1014 dmu_buf_rele_array(dbp, numbufs, FTAG);
1018 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1027 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
1028 FALSE, FTAG, &numbufs, &dbp));
1030 for (i = 0; i < numbufs; i++) {
1031 dmu_buf_t *db = dbp[i];
1033 dmu_buf_will_not_fill(db, tx);
1035 dmu_buf_rele_array(dbp, numbufs, FTAG);
1039 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
1040 void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
1041 int compressed_size, int byteorder, dmu_tx_t *tx)
1045 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
1046 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
1047 VERIFY0(dmu_buf_hold_noread(os, object, offset,
1050 dmu_buf_write_embedded(db,
1051 data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
1052 uncompressed_size, compressed_size, byteorder, tx);
1054 dmu_buf_rele(db, FTAG);
1058 * DMU support for xuio
1060 kstat_t *xuio_ksp = NULL;
1063 dmu_xuio_init(xuio_t *xuio, int nblk)
1066 uio_t *uio = &xuio->xu_uio;
1068 uio->uio_iovcnt = nblk;
1069 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
1071 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
1073 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
1074 priv->iovp = uio->uio_iov;
1075 XUIO_XUZC_PRIV(xuio) = priv;
1077 if (XUIO_XUZC_RW(xuio) == UIO_READ)
1078 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
1080 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
1086 dmu_xuio_fini(xuio_t *xuio)
1088 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1089 int nblk = priv->cnt;
1091 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
1092 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
1093 kmem_free(priv, sizeof (dmu_xuio_t));
1095 if (XUIO_XUZC_RW(xuio) == UIO_READ)
1096 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
1098 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
1102 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
1103 * and increase priv->next by 1.
1106 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
1109 uio_t *uio = &xuio->xu_uio;
1110 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1111 int i = priv->next++;
1113 ASSERT(i < priv->cnt);
1114 ASSERT(off + n <= arc_buf_lsize(abuf));
1115 iov = uio->uio_iov + i;
1116 iov->iov_base = (char *)abuf->b_data + off;
1118 priv->bufs[i] = abuf;
1123 dmu_xuio_cnt(xuio_t *xuio)
1125 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1130 dmu_xuio_arcbuf(xuio_t *xuio, int i)
1132 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1134 ASSERT(i < priv->cnt);
1135 return (priv->bufs[i]);
1139 dmu_xuio_clear(xuio_t *xuio, int i)
1141 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1143 ASSERT(i < priv->cnt);
1144 priv->bufs[i] = NULL;
1148 xuio_stat_init(void)
1150 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
1151 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
1152 KSTAT_FLAG_VIRTUAL);
1153 if (xuio_ksp != NULL) {
1154 xuio_ksp->ks_data = &xuio_stats;
1155 kstat_install(xuio_ksp);
1160 xuio_stat_fini(void)
1162 if (xuio_ksp != NULL) {
1163 kstat_delete(xuio_ksp);
1169 xuio_stat_wbuf_copied(void)
1171 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1175 xuio_stat_wbuf_nocopy(void)
1177 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
1182 dmu_read_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size)
1185 int numbufs, i, err;
1186 xuio_t *xuio = NULL;
1189 * NB: we could do this block-at-a-time, but it's nice
1190 * to be reading in parallel.
1192 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1193 TRUE, FTAG, &numbufs, &dbp, 0);
1198 if (uio->uio_extflg == UIO_XUIO)
1199 xuio = (xuio_t *)uio;
1202 for (i = 0; i < numbufs; i++) {
1205 dmu_buf_t *db = dbp[i];
1209 bufoff = uio->uio_loffset - db->db_offset;
1210 tocpy = (int)MIN(db->db_size - bufoff, size);
1213 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
1214 arc_buf_t *dbuf_abuf = dbi->db_buf;
1215 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
1216 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
1218 uio->uio_resid -= tocpy;
1219 uio->uio_loffset += tocpy;
1222 if (abuf == dbuf_abuf)
1223 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
1225 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
1228 err = uiomove((char *)db->db_data + bufoff, tocpy,
1231 err = vn_io_fault_uiomove((char *)db->db_data + bufoff,
1240 dmu_buf_rele_array(dbp, numbufs, FTAG);
1246 * Read 'size' bytes into the uio buffer.
1247 * From object zdb->db_object.
1248 * Starting at offset uio->uio_loffset.
1250 * If the caller already has a dbuf in the target object
1251 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1252 * because we don't have to find the dnode_t for the object.
1255 dmu_read_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size)
1257 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1266 err = dmu_read_uio_dnode(dn, uio, size);
1273 * Read 'size' bytes into the uio buffer.
1274 * From the specified object
1275 * Starting at offset uio->uio_loffset.
1278 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
1286 err = dnode_hold(os, object, FTAG, &dn);
1290 err = dmu_read_uio_dnode(dn, uio, size);
1292 dnode_rele(dn, FTAG);
1298 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1305 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1306 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1310 for (i = 0; i < numbufs; i++) {
1313 dmu_buf_t *db = dbp[i];
1317 bufoff = uio->uio_loffset - db->db_offset;
1318 tocpy = (int)MIN(db->db_size - bufoff, size);
1320 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1322 if (tocpy == db->db_size)
1323 dmu_buf_will_fill(db, tx);
1325 dmu_buf_will_dirty(db, tx);
1329 * XXX uiomove could block forever (eg. nfs-backed
1330 * pages). There needs to be a uiolockdown() function
1331 * to lock the pages in memory, so that uiomove won't
1334 err = uiomove((char *)db->db_data + bufoff, tocpy,
1337 err = vn_io_fault_uiomove((char *)db->db_data + bufoff, tocpy,
1341 if (tocpy == db->db_size)
1342 dmu_buf_fill_done(db, tx);
1350 dmu_buf_rele_array(dbp, numbufs, FTAG);
1355 * Write 'size' bytes from the uio buffer.
1356 * To object zdb->db_object.
1357 * Starting at offset uio->uio_loffset.
1359 * If the caller already has a dbuf in the target object
1360 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1361 * because we don't have to find the dnode_t for the object.
1364 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1367 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1376 err = dmu_write_uio_dnode(dn, uio, size, tx);
1383 * Write 'size' bytes from the uio buffer.
1384 * To the specified object.
1385 * Starting at offset uio->uio_loffset.
1388 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1397 err = dnode_hold(os, object, FTAG, &dn);
1401 err = dmu_write_uio_dnode(dn, uio, size, tx);
1403 dnode_rele(dn, FTAG);
1410 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1411 page_t *pp, dmu_tx_t *tx)
1420 err = dmu_buf_hold_array(os, object, offset, size,
1421 FALSE, FTAG, &numbufs, &dbp);
1425 for (i = 0; i < numbufs; i++) {
1426 int tocpy, copied, thiscpy;
1428 dmu_buf_t *db = dbp[i];
1432 ASSERT3U(db->db_size, >=, PAGESIZE);
1434 bufoff = offset - db->db_offset;
1435 tocpy = (int)MIN(db->db_size - bufoff, size);
1437 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1439 if (tocpy == db->db_size)
1440 dmu_buf_will_fill(db, tx);
1442 dmu_buf_will_dirty(db, tx);
1444 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1445 ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff);
1446 thiscpy = MIN(PAGESIZE, tocpy - copied);
1447 va = zfs_map_page(pp, S_READ);
1448 bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1449 zfs_unmap_page(pp, va);
1454 if (tocpy == db->db_size)
1455 dmu_buf_fill_done(db, tx);
1460 dmu_buf_rele_array(dbp, numbufs, FTAG);
1464 #else /* !illumos */
1467 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1468 vm_page_t *ma, dmu_tx_t *tx)
1478 err = dmu_buf_hold_array(os, object, offset, size,
1479 FALSE, FTAG, &numbufs, &dbp);
1483 for (i = 0; i < numbufs; i++) {
1484 int tocpy, copied, thiscpy;
1486 dmu_buf_t *db = dbp[i];
1490 ASSERT3U(db->db_size, >=, PAGESIZE);
1492 bufoff = offset - db->db_offset;
1493 tocpy = (int)MIN(db->db_size - bufoff, size);
1495 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1497 if (tocpy == db->db_size)
1498 dmu_buf_will_fill(db, tx);
1500 dmu_buf_will_dirty(db, tx);
1502 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1503 ASSERT3U(ptoa((*ma)->pindex), ==, db->db_offset + bufoff);
1504 thiscpy = MIN(PAGESIZE, tocpy - copied);
1505 va = zfs_map_page(*ma, &sf);
1506 bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1512 if (tocpy == db->db_size)
1513 dmu_buf_fill_done(db, tx);
1518 dmu_buf_rele_array(dbp, numbufs, FTAG);
1521 #endif /* illumos */
1522 #endif /* _KERNEL */
1525 * Allocate a loaned anonymous arc buffer.
1528 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1530 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1532 return (arc_loan_buf(db->db_objset->os_spa, B_FALSE, size));
1536 * Free a loaned arc buffer.
1539 dmu_return_arcbuf(arc_buf_t *buf)
1541 arc_return_buf(buf, FTAG);
1542 arc_buf_destroy(buf, FTAG);
1546 * When possible directly assign passed loaned arc buffer to a dbuf.
1547 * If this is not possible copy the contents of passed arc buf via
1551 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1554 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1557 uint32_t blksz = (uint32_t)arc_buf_lsize(buf);
1560 DB_DNODE_ENTER(dbuf);
1561 dn = DB_DNODE(dbuf);
1562 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1563 blkid = dbuf_whichblock(dn, 0, offset);
1564 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1565 rw_exit(&dn->dn_struct_rwlock);
1566 DB_DNODE_EXIT(dbuf);
1569 * We can only assign if the offset is aligned, the arc buf is the
1570 * same size as the dbuf, and the dbuf is not metadata.
1572 if (offset == db->db.db_offset && blksz == db->db.db_size) {
1574 curthread->td_ru.ru_oublock++;
1578 racct_add_force(curproc, RACCT_WRITEBPS, blksz);
1579 racct_add_force(curproc, RACCT_WRITEIOPS, 1);
1580 PROC_UNLOCK(curproc);
1583 #endif /* _KERNEL */
1584 dbuf_assign_arcbuf(db, buf, tx);
1585 dbuf_rele(db, FTAG);
1590 /* compressed bufs must always be assignable to their dbuf */
1591 ASSERT3U(arc_get_compression(buf), ==, ZIO_COMPRESS_OFF);
1592 ASSERT(!(buf->b_flags & ARC_BUF_FLAG_COMPRESSED));
1594 DB_DNODE_ENTER(dbuf);
1595 dn = DB_DNODE(dbuf);
1597 object = dn->dn_object;
1598 DB_DNODE_EXIT(dbuf);
1600 dbuf_rele(db, FTAG);
1601 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1602 dmu_return_arcbuf(buf);
1603 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1608 dbuf_dirty_record_t *dsa_dr;
1609 dmu_sync_cb_t *dsa_done;
1616 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1618 dmu_sync_arg_t *dsa = varg;
1619 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1620 blkptr_t *bp = zio->io_bp;
1622 if (zio->io_error == 0) {
1623 if (BP_IS_HOLE(bp)) {
1625 * A block of zeros may compress to a hole, but the
1626 * block size still needs to be known for replay.
1628 BP_SET_LSIZE(bp, db->db_size);
1629 } else if (!BP_IS_EMBEDDED(bp)) {
1630 ASSERT(BP_GET_LEVEL(bp) == 0);
1637 dmu_sync_late_arrival_ready(zio_t *zio)
1639 dmu_sync_ready(zio, NULL, zio->io_private);
1644 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1646 dmu_sync_arg_t *dsa = varg;
1647 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1648 dmu_buf_impl_t *db = dr->dr_dbuf;
1650 mutex_enter(&db->db_mtx);
1651 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1652 if (zio->io_error == 0) {
1653 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1654 if (dr->dt.dl.dr_nopwrite) {
1655 blkptr_t *bp = zio->io_bp;
1656 blkptr_t *bp_orig = &zio->io_bp_orig;
1657 uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1659 ASSERT(BP_EQUAL(bp, bp_orig));
1660 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1661 ASSERT(zio_checksum_table[chksum].ci_flags &
1662 ZCHECKSUM_FLAG_NOPWRITE);
1664 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1665 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1666 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1669 * Old style holes are filled with all zeros, whereas
1670 * new-style holes maintain their lsize, type, level,
1671 * and birth time (see zio_write_compress). While we
1672 * need to reset the BP_SET_LSIZE() call that happened
1673 * in dmu_sync_ready for old style holes, we do *not*
1674 * want to wipe out the information contained in new
1675 * style holes. Thus, only zero out the block pointer if
1676 * it's an old style hole.
1678 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) &&
1679 dr->dt.dl.dr_overridden_by.blk_birth == 0)
1680 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1682 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1684 cv_broadcast(&db->db_changed);
1685 mutex_exit(&db->db_mtx);
1687 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1689 kmem_free(dsa, sizeof (*dsa));
1693 dmu_sync_late_arrival_done(zio_t *zio)
1695 blkptr_t *bp = zio->io_bp;
1696 dmu_sync_arg_t *dsa = zio->io_private;
1697 blkptr_t *bp_orig = &zio->io_bp_orig;
1699 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1701 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE)
1702 * then there is nothing to do here. Otherwise, free the
1703 * newly allocated block in this txg.
1705 if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
1706 ASSERT(BP_EQUAL(bp, bp_orig));
1708 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1709 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1710 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1711 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1715 dmu_tx_commit(dsa->dsa_tx);
1717 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1719 abd_put(zio->io_abd);
1720 kmem_free(dsa, sizeof (*dsa));
1724 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1725 zio_prop_t *zp, zbookmark_phys_t *zb)
1727 dmu_sync_arg_t *dsa;
1730 tx = dmu_tx_create(os);
1731 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1732 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1734 /* Make zl_get_data do txg_waited_synced() */
1735 return (SET_ERROR(EIO));
1738 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1740 dsa->dsa_done = done;
1744 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1745 abd_get_from_buf(zgd->zgd_db->db_data, zgd->zgd_db->db_size),
1746 zgd->zgd_db->db_size, zgd->zgd_db->db_size, zp,
1747 dmu_sync_late_arrival_ready, NULL, NULL, dmu_sync_late_arrival_done,
1748 dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1754 * Intent log support: sync the block associated with db to disk.
1755 * N.B. and XXX: the caller is responsible for making sure that the
1756 * data isn't changing while dmu_sync() is writing it.
1760 * EEXIST: this txg has already been synced, so there's nothing to do.
1761 * The caller should not log the write.
1763 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1764 * The caller should not log the write.
1766 * EALREADY: this block is already in the process of being synced.
1767 * The caller should track its progress (somehow).
1769 * EIO: could not do the I/O.
1770 * The caller should do a txg_wait_synced().
1772 * 0: the I/O has been initiated.
1773 * The caller should log this blkptr in the done callback.
1774 * It is possible that the I/O will fail, in which case
1775 * the error will be reported to the done callback and
1776 * propagated to pio from zio_done().
1779 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1781 blkptr_t *bp = zgd->zgd_bp;
1782 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1783 objset_t *os = db->db_objset;
1784 dsl_dataset_t *ds = os->os_dsl_dataset;
1785 dbuf_dirty_record_t *dr;
1786 dmu_sync_arg_t *dsa;
1787 zbookmark_phys_t zb;
1791 ASSERT(pio != NULL);
1794 SET_BOOKMARK(&zb, ds->ds_object,
1795 db->db.db_object, db->db_level, db->db_blkid);
1799 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1803 * If we're frozen (running ziltest), we always need to generate a bp.
1805 if (txg > spa_freeze_txg(os->os_spa))
1806 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1809 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1810 * and us. If we determine that this txg is not yet syncing,
1811 * but it begins to sync a moment later, that's OK because the
1812 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1814 mutex_enter(&db->db_mtx);
1816 if (txg <= spa_last_synced_txg(os->os_spa)) {
1818 * This txg has already synced. There's nothing to do.
1820 mutex_exit(&db->db_mtx);
1821 return (SET_ERROR(EEXIST));
1824 if (txg <= spa_syncing_txg(os->os_spa)) {
1826 * This txg is currently syncing, so we can't mess with
1827 * the dirty record anymore; just write a new log block.
1829 mutex_exit(&db->db_mtx);
1830 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1833 dr = db->db_last_dirty;
1834 while (dr && dr->dr_txg != txg)
1839 * There's no dr for this dbuf, so it must have been freed.
1840 * There's no need to log writes to freed blocks, so we're done.
1842 mutex_exit(&db->db_mtx);
1843 return (SET_ERROR(ENOENT));
1846 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
1849 * Assume the on-disk data is X, the current syncing data (in
1850 * txg - 1) is Y, and the current in-memory data is Z (currently
1853 * We usually want to perform a nopwrite if X and Z are the
1854 * same. However, if Y is different (i.e. the BP is going to
1855 * change before this write takes effect), then a nopwrite will
1856 * be incorrect - we would override with X, which could have
1857 * been freed when Y was written.
1859 * (Note that this is not a concern when we are nop-writing from
1860 * syncing context, because X and Y must be identical, because
1861 * all previous txgs have been synced.)
1863 * Therefore, we disable nopwrite if the current BP could change
1864 * before this TXG. There are two ways it could change: by
1865 * being dirty (dr_next is non-NULL), or by being freed
1866 * (dnode_block_freed()). This behavior is verified by
1867 * zio_done(), which VERIFYs that the override BP is identical
1868 * to the on-disk BP.
1872 if (dr->dr_next != NULL || dnode_block_freed(dn, db->db_blkid))
1873 zp.zp_nopwrite = B_FALSE;
1876 ASSERT(dr->dr_txg == txg);
1877 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1878 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1880 * We have already issued a sync write for this buffer,
1881 * or this buffer has already been synced. It could not
1882 * have been dirtied since, or we would have cleared the state.
1884 mutex_exit(&db->db_mtx);
1885 return (SET_ERROR(EALREADY));
1888 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1889 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1890 mutex_exit(&db->db_mtx);
1892 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1894 dsa->dsa_done = done;
1898 zio_nowait(arc_write(pio, os->os_spa, txg,
1899 bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
1900 &zp, dmu_sync_ready, NULL, NULL, dmu_sync_done, dsa,
1901 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb));
1907 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1913 err = dnode_hold(os, object, FTAG, &dn);
1916 err = dnode_set_blksz(dn, size, ibs, tx);
1917 dnode_rele(dn, FTAG);
1922 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1928 * Send streams include each object's checksum function. This
1929 * check ensures that the receiving system can understand the
1930 * checksum function transmitted.
1932 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS);
1934 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1935 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS);
1936 dn->dn_checksum = checksum;
1937 dnode_setdirty(dn, tx);
1938 dnode_rele(dn, FTAG);
1942 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1948 * Send streams include each object's compression function. This
1949 * check ensures that the receiving system can understand the
1950 * compression function transmitted.
1952 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS);
1954 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1955 dn->dn_compress = compress;
1956 dnode_setdirty(dn, tx);
1957 dnode_rele(dn, FTAG);
1960 int zfs_mdcomp_disable = 0;
1961 SYSCTL_INT(_vfs_zfs, OID_AUTO, mdcomp_disable, CTLFLAG_RWTUN,
1962 &zfs_mdcomp_disable, 0, "Disable metadata compression");
1965 * When the "redundant_metadata" property is set to "most", only indirect
1966 * blocks of this level and higher will have an additional ditto block.
1968 int zfs_redundant_metadata_most_ditto_level = 2;
1971 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
1973 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
1974 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
1976 enum zio_checksum checksum = os->os_checksum;
1977 enum zio_compress compress = os->os_compress;
1978 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
1979 boolean_t dedup = B_FALSE;
1980 boolean_t nopwrite = B_FALSE;
1981 boolean_t dedup_verify = os->os_dedup_verify;
1982 int copies = os->os_copies;
1985 * We maintain different write policies for each of the following
1988 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1989 * 3. all other level 0 blocks
1992 if (zfs_mdcomp_disable) {
1993 compress = ZIO_COMPRESS_EMPTY;
1996 * XXX -- we should design a compression algorithm
1997 * that specializes in arrays of bps.
1999 compress = zio_compress_select(os->os_spa,
2000 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON);
2004 * Metadata always gets checksummed. If the data
2005 * checksum is multi-bit correctable, and it's not a
2006 * ZBT-style checksum, then it's suitable for metadata
2007 * as well. Otherwise, the metadata checksum defaults
2010 if (!(zio_checksum_table[checksum].ci_flags &
2011 ZCHECKSUM_FLAG_METADATA) ||
2012 (zio_checksum_table[checksum].ci_flags &
2013 ZCHECKSUM_FLAG_EMBEDDED))
2014 checksum = ZIO_CHECKSUM_FLETCHER_4;
2016 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL ||
2017 (os->os_redundant_metadata ==
2018 ZFS_REDUNDANT_METADATA_MOST &&
2019 (level >= zfs_redundant_metadata_most_ditto_level ||
2020 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))))
2022 } else if (wp & WP_NOFILL) {
2026 * If we're writing preallocated blocks, we aren't actually
2027 * writing them so don't set any policy properties. These
2028 * blocks are currently only used by an external subsystem
2029 * outside of zfs (i.e. dump) and not written by the zio
2032 compress = ZIO_COMPRESS_OFF;
2033 checksum = ZIO_CHECKSUM_NOPARITY;
2035 compress = zio_compress_select(os->os_spa, dn->dn_compress,
2038 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
2039 zio_checksum_select(dn->dn_checksum, checksum) :
2043 * Determine dedup setting. If we are in dmu_sync(),
2044 * we won't actually dedup now because that's all
2045 * done in syncing context; but we do want to use the
2046 * dedup checkum. If the checksum is not strong
2047 * enough to ensure unique signatures, force
2050 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
2051 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
2052 if (!(zio_checksum_table[checksum].ci_flags &
2053 ZCHECKSUM_FLAG_DEDUP))
2054 dedup_verify = B_TRUE;
2058 * Enable nopwrite if we have secure enough checksum
2059 * algorithm (see comment in zio_nop_write) and
2060 * compression is enabled. We don't enable nopwrite if
2061 * dedup is enabled as the two features are mutually
2064 nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags &
2065 ZCHECKSUM_FLAG_NOPWRITE) &&
2066 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
2069 zp->zp_checksum = checksum;
2070 zp->zp_compress = compress;
2071 ASSERT3U(zp->zp_compress, !=, ZIO_COMPRESS_INHERIT);
2073 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
2074 zp->zp_level = level;
2075 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
2076 zp->zp_dedup = dedup;
2077 zp->zp_dedup_verify = dedup && dedup_verify;
2078 zp->zp_nopwrite = nopwrite;
2082 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
2088 * Sync any current changes before
2089 * we go trundling through the block pointers.
2091 err = dmu_object_wait_synced(os, object);
2096 err = dnode_hold(os, object, FTAG, &dn);
2101 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
2102 dnode_rele(dn, FTAG);
2108 * Given the ZFS object, if it contains any dirty nodes
2109 * this function flushes all dirty blocks to disk. This
2110 * ensures the DMU object info is updated. A more efficient
2111 * future version might just find the TXG with the maximum
2112 * ID and wait for that to be synced.
2115 dmu_object_wait_synced(objset_t *os, uint64_t object)
2120 error = dnode_hold(os, object, FTAG, &dn);
2125 for (i = 0; i < TXG_SIZE; i++) {
2126 if (list_link_active(&dn->dn_dirty_link[i])) {
2130 dnode_rele(dn, FTAG);
2131 if (i != TXG_SIZE) {
2132 txg_wait_synced(dmu_objset_pool(os), 0);
2139 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2143 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2144 mutex_enter(&dn->dn_mtx);
2148 doi->doi_data_block_size = dn->dn_datablksz;
2149 doi->doi_metadata_block_size = dn->dn_indblkshift ?
2150 1ULL << dn->dn_indblkshift : 0;
2151 doi->doi_type = dn->dn_type;
2152 doi->doi_bonus_type = dn->dn_bonustype;
2153 doi->doi_bonus_size = dn->dn_bonuslen;
2154 doi->doi_indirection = dn->dn_nlevels;
2155 doi->doi_checksum = dn->dn_checksum;
2156 doi->doi_compress = dn->dn_compress;
2157 doi->doi_nblkptr = dn->dn_nblkptr;
2158 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
2159 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
2160 doi->doi_fill_count = 0;
2161 for (int i = 0; i < dnp->dn_nblkptr; i++)
2162 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
2164 mutex_exit(&dn->dn_mtx);
2165 rw_exit(&dn->dn_struct_rwlock);
2169 * Get information on a DMU object.
2170 * If doi is NULL, just indicates whether the object exists.
2173 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
2176 int err = dnode_hold(os, object, FTAG, &dn);
2182 dmu_object_info_from_dnode(dn, doi);
2184 dnode_rele(dn, FTAG);
2189 * As above, but faster; can be used when you have a held dbuf in hand.
2192 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
2194 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2197 dmu_object_info_from_dnode(DB_DNODE(db), doi);
2202 * Faster still when you only care about the size.
2203 * This is specifically optimized for zfs_getattr().
2206 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
2207 u_longlong_t *nblk512)
2209 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2215 *blksize = dn->dn_datablksz;
2216 /* add 1 for dnode space */
2217 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
2218 SPA_MINBLOCKSHIFT) + 1;
2223 byteswap_uint64_array(void *vbuf, size_t size)
2225 uint64_t *buf = vbuf;
2226 size_t count = size >> 3;
2229 ASSERT((size & 7) == 0);
2231 for (i = 0; i < count; i++)
2232 buf[i] = BSWAP_64(buf[i]);
2236 byteswap_uint32_array(void *vbuf, size_t size)
2238 uint32_t *buf = vbuf;
2239 size_t count = size >> 2;
2242 ASSERT((size & 3) == 0);
2244 for (i = 0; i < count; i++)
2245 buf[i] = BSWAP_32(buf[i]);
2249 byteswap_uint16_array(void *vbuf, size_t size)
2251 uint16_t *buf = vbuf;
2252 size_t count = size >> 1;
2255 ASSERT((size & 1) == 0);
2257 for (i = 0; i < count; i++)
2258 buf[i] = BSWAP_16(buf[i]);
2263 byteswap_uint8_array(void *vbuf, size_t size)
2277 zio_compress_init();
2286 arc_fini(); /* arc depends on l2arc, so arc must go first */
2289 zio_compress_fini();