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, 2016 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>
50 #include <sys/racct.h>
52 #include <sys/zfs_znode.h>
56 * Enable/disable nopwrite feature.
58 int zfs_nopwrite_enabled = 1;
59 SYSCTL_DECL(_vfs_zfs);
60 SYSCTL_INT(_vfs_zfs, OID_AUTO, nopwrite_enabled, CTLFLAG_RDTUN,
61 &zfs_nopwrite_enabled, 0, "Enable nopwrite feature");
64 * Tunable to control percentage of dirtied blocks from frees in one TXG.
65 * After this threshold is crossed, additional dirty blocks from frees
66 * wait until the next TXG.
67 * A value of zero will disable this throttle.
69 uint32_t zfs_per_txg_dirty_frees_percent = 30;
70 SYSCTL_INT(_vfs_zfs, OID_AUTO, per_txg_dirty_frees_percent, CTLFLAG_RWTUN,
71 &zfs_per_txg_dirty_frees_percent, 0, "Percentage of dirtied blocks from frees in one txg");
73 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
74 { DMU_BSWAP_UINT8, TRUE, "unallocated" },
75 { DMU_BSWAP_ZAP, TRUE, "object directory" },
76 { DMU_BSWAP_UINT64, TRUE, "object array" },
77 { DMU_BSWAP_UINT8, TRUE, "packed nvlist" },
78 { DMU_BSWAP_UINT64, TRUE, "packed nvlist size" },
79 { DMU_BSWAP_UINT64, TRUE, "bpobj" },
80 { DMU_BSWAP_UINT64, TRUE, "bpobj header" },
81 { DMU_BSWAP_UINT64, TRUE, "SPA space map header" },
82 { DMU_BSWAP_UINT64, TRUE, "SPA space map" },
83 { DMU_BSWAP_UINT64, TRUE, "ZIL intent log" },
84 { DMU_BSWAP_DNODE, TRUE, "DMU dnode" },
85 { DMU_BSWAP_OBJSET, TRUE, "DMU objset" },
86 { DMU_BSWAP_UINT64, TRUE, "DSL directory" },
87 { DMU_BSWAP_ZAP, TRUE, "DSL directory child map"},
88 { DMU_BSWAP_ZAP, TRUE, "DSL dataset snap map" },
89 { DMU_BSWAP_ZAP, TRUE, "DSL props" },
90 { DMU_BSWAP_UINT64, TRUE, "DSL dataset" },
91 { DMU_BSWAP_ZNODE, TRUE, "ZFS znode" },
92 { DMU_BSWAP_OLDACL, TRUE, "ZFS V0 ACL" },
93 { DMU_BSWAP_UINT8, FALSE, "ZFS plain file" },
94 { DMU_BSWAP_ZAP, TRUE, "ZFS directory" },
95 { DMU_BSWAP_ZAP, TRUE, "ZFS master node" },
96 { DMU_BSWAP_ZAP, TRUE, "ZFS delete queue" },
97 { DMU_BSWAP_UINT8, FALSE, "zvol object" },
98 { DMU_BSWAP_ZAP, TRUE, "zvol prop" },
99 { DMU_BSWAP_UINT8, FALSE, "other uint8[]" },
100 { DMU_BSWAP_UINT64, FALSE, "other uint64[]" },
101 { DMU_BSWAP_ZAP, TRUE, "other ZAP" },
102 { DMU_BSWAP_ZAP, TRUE, "persistent error log" },
103 { DMU_BSWAP_UINT8, TRUE, "SPA history" },
104 { DMU_BSWAP_UINT64, TRUE, "SPA history offsets" },
105 { DMU_BSWAP_ZAP, TRUE, "Pool properties" },
106 { DMU_BSWAP_ZAP, TRUE, "DSL permissions" },
107 { DMU_BSWAP_ACL, TRUE, "ZFS ACL" },
108 { DMU_BSWAP_UINT8, TRUE, "ZFS SYSACL" },
109 { DMU_BSWAP_UINT8, TRUE, "FUID table" },
110 { DMU_BSWAP_UINT64, TRUE, "FUID table size" },
111 { DMU_BSWAP_ZAP, TRUE, "DSL dataset next clones"},
112 { DMU_BSWAP_ZAP, TRUE, "scan work queue" },
113 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group used" },
114 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group quota" },
115 { DMU_BSWAP_ZAP, TRUE, "snapshot refcount tags"},
116 { DMU_BSWAP_ZAP, TRUE, "DDT ZAP algorithm" },
117 { DMU_BSWAP_ZAP, TRUE, "DDT statistics" },
118 { DMU_BSWAP_UINT8, TRUE, "System attributes" },
119 { DMU_BSWAP_ZAP, TRUE, "SA master node" },
120 { DMU_BSWAP_ZAP, TRUE, "SA attr registration" },
121 { DMU_BSWAP_ZAP, TRUE, "SA attr layouts" },
122 { DMU_BSWAP_ZAP, TRUE, "scan translations" },
123 { DMU_BSWAP_UINT8, FALSE, "deduplicated block" },
124 { DMU_BSWAP_ZAP, TRUE, "DSL deadlist map" },
125 { DMU_BSWAP_UINT64, TRUE, "DSL deadlist map hdr" },
126 { DMU_BSWAP_ZAP, TRUE, "DSL dir clones" },
127 { DMU_BSWAP_UINT64, TRUE, "bpobj subobj" }
130 const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
131 { byteswap_uint8_array, "uint8" },
132 { byteswap_uint16_array, "uint16" },
133 { byteswap_uint32_array, "uint32" },
134 { byteswap_uint64_array, "uint64" },
135 { zap_byteswap, "zap" },
136 { dnode_buf_byteswap, "dnode" },
137 { dmu_objset_byteswap, "objset" },
138 { zfs_znode_byteswap, "znode" },
139 { zfs_oldacl_byteswap, "oldacl" },
140 { zfs_acl_byteswap, "acl" }
144 dmu_buf_hold_noread_by_dnode(dnode_t *dn, uint64_t offset,
145 void *tag, dmu_buf_t **dbp)
150 blkid = dbuf_whichblock(dn, 0, offset);
151 rw_enter(&dn->dn_struct_rwlock, RW_READER);
152 db = dbuf_hold(dn, blkid, tag);
153 rw_exit(&dn->dn_struct_rwlock);
157 return (SET_ERROR(EIO));
164 dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset,
165 void *tag, dmu_buf_t **dbp)
172 err = dnode_hold(os, object, FTAG, &dn);
175 blkid = dbuf_whichblock(dn, 0, offset);
176 rw_enter(&dn->dn_struct_rwlock, RW_READER);
177 db = dbuf_hold(dn, blkid, tag);
178 rw_exit(&dn->dn_struct_rwlock);
179 dnode_rele(dn, FTAG);
183 return (SET_ERROR(EIO));
191 dmu_buf_hold_by_dnode(dnode_t *dn, uint64_t offset,
192 void *tag, dmu_buf_t **dbp, int flags)
195 int db_flags = DB_RF_CANFAIL;
197 if (flags & DMU_READ_NO_PREFETCH)
198 db_flags |= DB_RF_NOPREFETCH;
200 err = dmu_buf_hold_noread_by_dnode(dn, offset, tag, dbp);
202 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
203 err = dbuf_read(db, NULL, db_flags);
214 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
215 void *tag, dmu_buf_t **dbp, int flags)
218 int db_flags = DB_RF_CANFAIL;
220 if (flags & DMU_READ_NO_PREFETCH)
221 db_flags |= DB_RF_NOPREFETCH;
223 err = dmu_buf_hold_noread(os, object, offset, tag, dbp);
225 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
226 err = dbuf_read(db, NULL, db_flags);
239 return (DN_MAX_BONUSLEN);
243 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
245 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
252 if (dn->dn_bonus != db) {
253 error = SET_ERROR(EINVAL);
254 } else if (newsize < 0 || newsize > db_fake->db_size) {
255 error = SET_ERROR(EINVAL);
257 dnode_setbonuslen(dn, newsize, tx);
266 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
268 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
275 if (!DMU_OT_IS_VALID(type)) {
276 error = SET_ERROR(EINVAL);
277 } else if (dn->dn_bonus != db) {
278 error = SET_ERROR(EINVAL);
280 dnode_setbonus_type(dn, type, tx);
289 dmu_get_bonustype(dmu_buf_t *db_fake)
291 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
293 dmu_object_type_t type;
297 type = dn->dn_bonustype;
304 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
309 error = dnode_hold(os, object, FTAG, &dn);
310 dbuf_rm_spill(dn, tx);
311 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
312 dnode_rm_spill(dn, tx);
313 rw_exit(&dn->dn_struct_rwlock);
314 dnode_rele(dn, FTAG);
319 * returns ENOENT, EIO, or 0.
322 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
328 error = dnode_hold(os, object, FTAG, &dn);
332 rw_enter(&dn->dn_struct_rwlock, RW_READER);
333 if (dn->dn_bonus == NULL) {
334 rw_exit(&dn->dn_struct_rwlock);
335 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
336 if (dn->dn_bonus == NULL)
337 dbuf_create_bonus(dn);
341 /* as long as the bonus buf is held, the dnode will be held */
342 if (refcount_add(&db->db_holds, tag) == 1) {
343 VERIFY(dnode_add_ref(dn, db));
344 atomic_inc_32(&dn->dn_dbufs_count);
348 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
349 * hold and incrementing the dbuf count to ensure that dnode_move() sees
350 * a dnode hold for every dbuf.
352 rw_exit(&dn->dn_struct_rwlock);
354 dnode_rele(dn, FTAG);
356 VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH));
363 * returns ENOENT, EIO, or 0.
365 * This interface will allocate a blank spill dbuf when a spill blk
366 * doesn't already exist on the dnode.
368 * if you only want to find an already existing spill db, then
369 * dmu_spill_hold_existing() should be used.
372 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
374 dmu_buf_impl_t *db = NULL;
377 if ((flags & DB_RF_HAVESTRUCT) == 0)
378 rw_enter(&dn->dn_struct_rwlock, RW_READER);
380 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
382 if ((flags & DB_RF_HAVESTRUCT) == 0)
383 rw_exit(&dn->dn_struct_rwlock);
386 err = dbuf_read(db, NULL, flags);
395 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
397 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
404 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
405 err = SET_ERROR(EINVAL);
407 rw_enter(&dn->dn_struct_rwlock, RW_READER);
409 if (!dn->dn_have_spill) {
410 err = SET_ERROR(ENOENT);
412 err = dmu_spill_hold_by_dnode(dn,
413 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
416 rw_exit(&dn->dn_struct_rwlock);
424 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
426 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
432 err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp);
439 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
440 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
441 * and can induce severe lock contention when writing to several files
442 * whose dnodes are in the same block.
445 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
446 boolean_t read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
449 uint64_t blkid, nblks, i;
454 ASSERT(length <= DMU_MAX_ACCESS);
457 * Note: We directly notify the prefetch code of this read, so that
458 * we can tell it about the multi-block read. dbuf_read() only knows
459 * about the one block it is accessing.
461 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT |
464 rw_enter(&dn->dn_struct_rwlock, RW_READER);
465 if (dn->dn_datablkshift) {
466 int blkshift = dn->dn_datablkshift;
467 nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) -
468 P2ALIGN(offset, 1ULL << blkshift)) >> blkshift;
470 if (offset + length > dn->dn_datablksz) {
471 zfs_panic_recover("zfs: accessing past end of object "
472 "%llx/%llx (size=%u access=%llu+%llu)",
473 (longlong_t)dn->dn_objset->
474 os_dsl_dataset->ds_object,
475 (longlong_t)dn->dn_object, dn->dn_datablksz,
476 (longlong_t)offset, (longlong_t)length);
477 rw_exit(&dn->dn_struct_rwlock);
478 return (SET_ERROR(EIO));
482 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
484 #if defined(_KERNEL) && defined(RACCT)
485 if (racct_enable && !read) {
487 racct_add_force(curproc, RACCT_WRITEBPS, length);
488 racct_add_force(curproc, RACCT_WRITEIOPS, nblks);
489 PROC_UNLOCK(curproc);
493 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
494 blkid = dbuf_whichblock(dn, 0, offset);
495 for (i = 0; i < nblks; i++) {
496 dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag);
498 rw_exit(&dn->dn_struct_rwlock);
499 dmu_buf_rele_array(dbp, nblks, tag);
501 return (SET_ERROR(EIO));
504 /* initiate async i/o */
506 (void) dbuf_read(db, zio, dbuf_flags);
509 curthread->td_ru.ru_oublock++;
514 if ((flags & DMU_READ_NO_PREFETCH) == 0 &&
515 DNODE_META_IS_CACHEABLE(dn) && length <= zfetch_array_rd_sz) {
516 dmu_zfetch(&dn->dn_zfetch, blkid, nblks,
517 read && DNODE_IS_CACHEABLE(dn));
519 rw_exit(&dn->dn_struct_rwlock);
521 /* wait for async i/o */
524 dmu_buf_rele_array(dbp, nblks, tag);
528 /* wait for other io to complete */
530 for (i = 0; i < nblks; i++) {
531 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
532 mutex_enter(&db->db_mtx);
533 while (db->db_state == DB_READ ||
534 db->db_state == DB_FILL)
535 cv_wait(&db->db_changed, &db->db_mtx);
536 if (db->db_state == DB_UNCACHED)
537 err = SET_ERROR(EIO);
538 mutex_exit(&db->db_mtx);
540 dmu_buf_rele_array(dbp, nblks, tag);
552 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
553 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
558 err = dnode_hold(os, object, FTAG, &dn);
562 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
563 numbufsp, dbpp, DMU_READ_PREFETCH);
565 dnode_rele(dn, FTAG);
571 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
572 uint64_t length, boolean_t read, void *tag, int *numbufsp,
575 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
581 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
582 numbufsp, dbpp, DMU_READ_PREFETCH);
589 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
592 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
597 for (i = 0; i < numbufs; i++) {
599 dbuf_rele(dbp[i], tag);
602 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
606 * Issue prefetch i/os for the given blocks. If level is greater than 0, the
607 * indirect blocks prefeteched will be those that point to the blocks containing
608 * the data starting at offset, and continuing to offset + len.
610 * Note that if the indirect blocks above the blocks being prefetched are not in
611 * cache, they will be asychronously read in.
614 dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset,
615 uint64_t len, zio_priority_t pri)
621 if (len == 0) { /* they're interested in the bonus buffer */
622 dn = DMU_META_DNODE(os);
624 if (object == 0 || object >= DN_MAX_OBJECT)
627 rw_enter(&dn->dn_struct_rwlock, RW_READER);
628 blkid = dbuf_whichblock(dn, level,
629 object * sizeof (dnode_phys_t));
630 dbuf_prefetch(dn, level, blkid, pri, 0);
631 rw_exit(&dn->dn_struct_rwlock);
636 * XXX - Note, if the dnode for the requested object is not
637 * already cached, we will do a *synchronous* read in the
638 * dnode_hold() call. The same is true for any indirects.
640 err = dnode_hold(os, object, FTAG, &dn);
644 rw_enter(&dn->dn_struct_rwlock, RW_READER);
646 * offset + len - 1 is the last byte we want to prefetch for, and offset
647 * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the
648 * last block we want to prefetch, and dbuf_whichblock(dn, level,
649 * offset) is the first. Then the number we need to prefetch is the
652 if (level > 0 || dn->dn_datablkshift != 0) {
653 nblks = dbuf_whichblock(dn, level, offset + len - 1) -
654 dbuf_whichblock(dn, level, offset) + 1;
656 nblks = (offset < dn->dn_datablksz);
660 blkid = dbuf_whichblock(dn, level, offset);
661 for (int i = 0; i < nblks; i++)
662 dbuf_prefetch(dn, level, blkid + i, pri, 0);
665 rw_exit(&dn->dn_struct_rwlock);
667 dnode_rele(dn, FTAG);
671 * Get the next "chunk" of file data to free. We traverse the file from
672 * the end so that the file gets shorter over time (if we crashes in the
673 * middle, this will leave us in a better state). We find allocated file
674 * data by simply searching the allocated level 1 indirects.
676 * On input, *start should be the first offset that does not need to be
677 * freed (e.g. "offset + length"). On return, *start will be the first
678 * offset that should be freed.
681 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum)
683 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
684 /* bytes of data covered by a level-1 indirect block */
686 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
688 ASSERT3U(minimum, <=, *start);
690 if (*start - minimum <= iblkrange * maxblks) {
694 ASSERT(ISP2(iblkrange));
696 for (uint64_t blks = 0; *start > minimum && blks < maxblks; blks++) {
700 * dnode_next_offset(BACKWARDS) will find an allocated L1
701 * indirect block at or before the input offset. We must
702 * decrement *start so that it is at the end of the region
706 err = dnode_next_offset(dn,
707 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
709 /* if there are no indirect blocks before start, we are done */
713 } else if (err != 0) {
717 /* set start to the beginning of this L1 indirect */
718 *start = P2ALIGN(*start, iblkrange);
720 if (*start < minimum)
726 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
729 uint64_t object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
731 uint64_t dirty_frees_threshold;
732 dsl_pool_t *dp = dmu_objset_pool(os);
734 if (offset >= object_size)
737 if (zfs_per_txg_dirty_frees_percent <= 100)
738 dirty_frees_threshold =
739 zfs_per_txg_dirty_frees_percent * zfs_dirty_data_max / 100;
741 dirty_frees_threshold = zfs_dirty_data_max / 4;
743 if (length == DMU_OBJECT_END || offset + length > object_size)
744 length = object_size - offset;
746 while (length != 0) {
747 uint64_t chunk_end, chunk_begin, chunk_len;
748 uint64_t long_free_dirty_all_txgs = 0;
751 chunk_end = chunk_begin = offset + length;
753 /* move chunk_begin backwards to the beginning of this chunk */
754 err = get_next_chunk(dn, &chunk_begin, offset);
757 ASSERT3U(chunk_begin, >=, offset);
758 ASSERT3U(chunk_begin, <=, chunk_end);
760 chunk_len = chunk_end - chunk_begin;
762 mutex_enter(&dp->dp_lock);
763 for (int t = 0; t < TXG_SIZE; t++) {
764 long_free_dirty_all_txgs +=
765 dp->dp_long_free_dirty_pertxg[t];
767 mutex_exit(&dp->dp_lock);
770 * To avoid filling up a TXG with just frees wait for
771 * the next TXG to open before freeing more chunks if
772 * we have reached the threshold of frees
774 if (dirty_frees_threshold != 0 &&
775 long_free_dirty_all_txgs >= dirty_frees_threshold) {
776 txg_wait_open(dp, 0);
780 tx = dmu_tx_create(os);
781 dmu_tx_hold_free(tx, dn->dn_object, chunk_begin, chunk_len);
784 * Mark this transaction as typically resulting in a net
785 * reduction in space used.
787 dmu_tx_mark_netfree(tx);
788 err = dmu_tx_assign(tx, TXG_WAIT);
794 mutex_enter(&dp->dp_lock);
795 dp->dp_long_free_dirty_pertxg[dmu_tx_get_txg(tx) & TXG_MASK] +=
797 mutex_exit(&dp->dp_lock);
798 DTRACE_PROBE3(free__long__range,
799 uint64_t, long_free_dirty_all_txgs, uint64_t, chunk_len,
800 uint64_t, dmu_tx_get_txg(tx));
801 dnode_free_range(dn, chunk_begin, chunk_len, tx);
810 dmu_free_long_range(objset_t *os, uint64_t object,
811 uint64_t offset, uint64_t length)
816 err = dnode_hold(os, object, FTAG, &dn);
819 err = dmu_free_long_range_impl(os, dn, offset, length);
822 * It is important to zero out the maxblkid when freeing the entire
823 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
824 * will take the fast path, and (b) dnode_reallocate() can verify
825 * that the entire file has been freed.
827 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
830 dnode_rele(dn, FTAG);
835 dmu_free_long_object(objset_t *os, uint64_t object)
840 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
844 tx = dmu_tx_create(os);
845 dmu_tx_hold_bonus(tx, object);
846 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
847 dmu_tx_mark_netfree(tx);
848 err = dmu_tx_assign(tx, TXG_WAIT);
850 err = dmu_object_free(os, object, tx);
860 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
861 uint64_t size, dmu_tx_t *tx)
864 int err = dnode_hold(os, object, FTAG, &dn);
867 ASSERT(offset < UINT64_MAX);
868 ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
869 dnode_free_range(dn, offset, size, tx);
870 dnode_rele(dn, FTAG);
875 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
876 void *buf, uint32_t flags)
882 err = dnode_hold(os, object, FTAG, &dn);
887 * Deal with odd block sizes, where there can't be data past the first
888 * block. If we ever do the tail block optimization, we will need to
889 * handle that here as well.
891 if (dn->dn_maxblkid == 0) {
892 int newsz = offset > dn->dn_datablksz ? 0 :
893 MIN(size, dn->dn_datablksz - offset);
894 bzero((char *)buf + newsz, size - newsz);
899 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
903 * NB: we could do this block-at-a-time, but it's nice
904 * to be reading in parallel.
906 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
907 TRUE, FTAG, &numbufs, &dbp, flags);
911 for (i = 0; i < numbufs; i++) {
914 dmu_buf_t *db = dbp[i];
918 bufoff = offset - db->db_offset;
919 tocpy = (int)MIN(db->db_size - bufoff, size);
921 bcopy((char *)db->db_data + bufoff, buf, tocpy);
925 buf = (char *)buf + tocpy;
927 dmu_buf_rele_array(dbp, numbufs, FTAG);
929 dnode_rele(dn, FTAG);
934 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
935 const void *buf, dmu_tx_t *tx)
943 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
944 FALSE, FTAG, &numbufs, &dbp));
946 for (i = 0; i < numbufs; i++) {
949 dmu_buf_t *db = dbp[i];
953 bufoff = offset - db->db_offset;
954 tocpy = (int)MIN(db->db_size - bufoff, size);
956 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
958 if (tocpy == db->db_size)
959 dmu_buf_will_fill(db, tx);
961 dmu_buf_will_dirty(db, tx);
963 bcopy(buf, (char *)db->db_data + bufoff, tocpy);
965 if (tocpy == db->db_size)
966 dmu_buf_fill_done(db, tx);
970 buf = (char *)buf + tocpy;
972 dmu_buf_rele_array(dbp, numbufs, FTAG);
976 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
985 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
986 FALSE, FTAG, &numbufs, &dbp));
988 for (i = 0; i < numbufs; i++) {
989 dmu_buf_t *db = dbp[i];
991 dmu_buf_will_not_fill(db, tx);
993 dmu_buf_rele_array(dbp, numbufs, FTAG);
997 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
998 void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
999 int compressed_size, int byteorder, dmu_tx_t *tx)
1003 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
1004 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
1005 VERIFY0(dmu_buf_hold_noread(os, object, offset,
1008 dmu_buf_write_embedded(db,
1009 data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
1010 uncompressed_size, compressed_size, byteorder, tx);
1012 dmu_buf_rele(db, FTAG);
1016 * DMU support for xuio
1018 kstat_t *xuio_ksp = NULL;
1021 dmu_xuio_init(xuio_t *xuio, int nblk)
1024 uio_t *uio = &xuio->xu_uio;
1026 uio->uio_iovcnt = nblk;
1027 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
1029 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
1031 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
1032 priv->iovp = uio->uio_iov;
1033 XUIO_XUZC_PRIV(xuio) = priv;
1035 if (XUIO_XUZC_RW(xuio) == UIO_READ)
1036 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
1038 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
1044 dmu_xuio_fini(xuio_t *xuio)
1046 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1047 int nblk = priv->cnt;
1049 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
1050 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
1051 kmem_free(priv, sizeof (dmu_xuio_t));
1053 if (XUIO_XUZC_RW(xuio) == UIO_READ)
1054 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
1056 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
1060 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
1061 * and increase priv->next by 1.
1064 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
1067 uio_t *uio = &xuio->xu_uio;
1068 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1069 int i = priv->next++;
1071 ASSERT(i < priv->cnt);
1072 ASSERT(off + n <= arc_buf_size(abuf));
1073 iov = uio->uio_iov + i;
1074 iov->iov_base = (char *)abuf->b_data + off;
1076 priv->bufs[i] = abuf;
1081 dmu_xuio_cnt(xuio_t *xuio)
1083 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1088 dmu_xuio_arcbuf(xuio_t *xuio, int i)
1090 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1092 ASSERT(i < priv->cnt);
1093 return (priv->bufs[i]);
1097 dmu_xuio_clear(xuio_t *xuio, int i)
1099 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1101 ASSERT(i < priv->cnt);
1102 priv->bufs[i] = NULL;
1106 xuio_stat_init(void)
1108 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
1109 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
1110 KSTAT_FLAG_VIRTUAL);
1111 if (xuio_ksp != NULL) {
1112 xuio_ksp->ks_data = &xuio_stats;
1113 kstat_install(xuio_ksp);
1118 xuio_stat_fini(void)
1120 if (xuio_ksp != NULL) {
1121 kstat_delete(xuio_ksp);
1127 xuio_stat_wbuf_copied()
1129 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1133 xuio_stat_wbuf_nocopy()
1135 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
1140 dmu_read_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size)
1143 int numbufs, i, err;
1144 xuio_t *xuio = NULL;
1147 * NB: we could do this block-at-a-time, but it's nice
1148 * to be reading in parallel.
1150 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1151 TRUE, FTAG, &numbufs, &dbp, 0);
1156 if (uio->uio_extflg == UIO_XUIO)
1157 xuio = (xuio_t *)uio;
1160 for (i = 0; i < numbufs; i++) {
1163 dmu_buf_t *db = dbp[i];
1167 bufoff = uio->uio_loffset - db->db_offset;
1168 tocpy = (int)MIN(db->db_size - bufoff, size);
1171 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
1172 arc_buf_t *dbuf_abuf = dbi->db_buf;
1173 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
1174 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
1176 uio->uio_resid -= tocpy;
1177 uio->uio_loffset += tocpy;
1180 if (abuf == dbuf_abuf)
1181 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
1183 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
1186 err = uiomove((char *)db->db_data + bufoff, tocpy,
1189 err = vn_io_fault_uiomove((char *)db->db_data + bufoff,
1198 dmu_buf_rele_array(dbp, numbufs, FTAG);
1204 * Read 'size' bytes into the uio buffer.
1205 * From object zdb->db_object.
1206 * Starting at offset uio->uio_loffset.
1208 * If the caller already has a dbuf in the target object
1209 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1210 * because we don't have to find the dnode_t for the object.
1213 dmu_read_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size)
1215 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1224 err = dmu_read_uio_dnode(dn, uio, size);
1231 * Read 'size' bytes into the uio buffer.
1232 * From the specified object
1233 * Starting at offset uio->uio_loffset.
1236 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
1244 err = dnode_hold(os, object, FTAG, &dn);
1248 err = dmu_read_uio_dnode(dn, uio, size);
1250 dnode_rele(dn, FTAG);
1256 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1263 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1264 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1268 for (i = 0; i < numbufs; i++) {
1271 dmu_buf_t *db = dbp[i];
1275 bufoff = uio->uio_loffset - db->db_offset;
1276 tocpy = (int)MIN(db->db_size - bufoff, size);
1278 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1280 if (tocpy == db->db_size)
1281 dmu_buf_will_fill(db, tx);
1283 dmu_buf_will_dirty(db, tx);
1287 * XXX uiomove could block forever (eg. nfs-backed
1288 * pages). There needs to be a uiolockdown() function
1289 * to lock the pages in memory, so that uiomove won't
1292 err = uiomove((char *)db->db_data + bufoff, tocpy,
1295 err = vn_io_fault_uiomove((char *)db->db_data + bufoff, tocpy,
1299 if (tocpy == db->db_size)
1300 dmu_buf_fill_done(db, tx);
1308 dmu_buf_rele_array(dbp, numbufs, FTAG);
1313 * Write 'size' bytes from the uio buffer.
1314 * To object zdb->db_object.
1315 * Starting at offset uio->uio_loffset.
1317 * If the caller already has a dbuf in the target object
1318 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1319 * because we don't have to find the dnode_t for the object.
1322 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1325 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1334 err = dmu_write_uio_dnode(dn, uio, size, tx);
1341 * Write 'size' bytes from the uio buffer.
1342 * To the specified object.
1343 * Starting at offset uio->uio_loffset.
1346 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1355 err = dnode_hold(os, object, FTAG, &dn);
1359 err = dmu_write_uio_dnode(dn, uio, size, tx);
1361 dnode_rele(dn, FTAG);
1368 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1369 page_t *pp, dmu_tx_t *tx)
1378 err = dmu_buf_hold_array(os, object, offset, size,
1379 FALSE, FTAG, &numbufs, &dbp);
1383 for (i = 0; i < numbufs; i++) {
1384 int tocpy, copied, thiscpy;
1386 dmu_buf_t *db = dbp[i];
1390 ASSERT3U(db->db_size, >=, PAGESIZE);
1392 bufoff = offset - db->db_offset;
1393 tocpy = (int)MIN(db->db_size - bufoff, size);
1395 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1397 if (tocpy == db->db_size)
1398 dmu_buf_will_fill(db, tx);
1400 dmu_buf_will_dirty(db, tx);
1402 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1403 ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff);
1404 thiscpy = MIN(PAGESIZE, tocpy - copied);
1405 va = zfs_map_page(pp, S_READ);
1406 bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1407 zfs_unmap_page(pp, va);
1412 if (tocpy == db->db_size)
1413 dmu_buf_fill_done(db, tx);
1418 dmu_buf_rele_array(dbp, numbufs, FTAG);
1422 #else /* !illumos */
1425 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1426 vm_page_t *ma, dmu_tx_t *tx)
1436 err = dmu_buf_hold_array(os, object, offset, size,
1437 FALSE, FTAG, &numbufs, &dbp);
1441 for (i = 0; i < numbufs; i++) {
1442 int tocpy, copied, thiscpy;
1444 dmu_buf_t *db = dbp[i];
1448 ASSERT3U(db->db_size, >=, PAGESIZE);
1450 bufoff = offset - db->db_offset;
1451 tocpy = (int)MIN(db->db_size - bufoff, size);
1453 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1455 if (tocpy == db->db_size)
1456 dmu_buf_will_fill(db, tx);
1458 dmu_buf_will_dirty(db, tx);
1460 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1461 ASSERT3U(ptoa((*ma)->pindex), ==, db->db_offset + bufoff);
1462 thiscpy = MIN(PAGESIZE, tocpy - copied);
1463 va = zfs_map_page(*ma, &sf);
1464 bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1470 if (tocpy == db->db_size)
1471 dmu_buf_fill_done(db, tx);
1476 dmu_buf_rele_array(dbp, numbufs, FTAG);
1479 #endif /* illumos */
1480 #endif /* _KERNEL */
1483 * Allocate a loaned anonymous arc buffer.
1486 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1488 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1490 return (arc_loan_buf(db->db_objset->os_spa, size));
1494 * Free a loaned arc buffer.
1497 dmu_return_arcbuf(arc_buf_t *buf)
1499 arc_return_buf(buf, FTAG);
1500 arc_buf_destroy(buf, FTAG);
1504 * When possible directly assign passed loaned arc buffer to a dbuf.
1505 * If this is not possible copy the contents of passed arc buf via
1509 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1512 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1515 uint32_t blksz = (uint32_t)arc_buf_size(buf);
1518 DB_DNODE_ENTER(dbuf);
1519 dn = DB_DNODE(dbuf);
1520 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1521 blkid = dbuf_whichblock(dn, 0, offset);
1522 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1523 rw_exit(&dn->dn_struct_rwlock);
1524 DB_DNODE_EXIT(dbuf);
1527 * We can only assign if the offset is aligned, the arc buf is the
1528 * same size as the dbuf, and the dbuf is not metadata. It
1529 * can't be metadata because the loaned arc buf comes from the
1530 * user-data kmem arena.
1532 if (offset == db->db.db_offset && blksz == db->db.db_size &&
1533 DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA) {
1535 curthread->td_ru.ru_oublock++;
1539 racct_add_force(curproc, RACCT_WRITEBPS, blksz);
1540 racct_add_force(curproc, RACCT_WRITEIOPS, 1);
1541 PROC_UNLOCK(curproc);
1544 #endif /* _KERNEL */
1545 dbuf_assign_arcbuf(db, buf, tx);
1546 dbuf_rele(db, FTAG);
1551 DB_DNODE_ENTER(dbuf);
1552 dn = DB_DNODE(dbuf);
1554 object = dn->dn_object;
1555 DB_DNODE_EXIT(dbuf);
1557 dbuf_rele(db, FTAG);
1558 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1559 dmu_return_arcbuf(buf);
1560 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1565 dbuf_dirty_record_t *dsa_dr;
1566 dmu_sync_cb_t *dsa_done;
1573 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1575 dmu_sync_arg_t *dsa = varg;
1576 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1577 blkptr_t *bp = zio->io_bp;
1579 if (zio->io_error == 0) {
1580 if (BP_IS_HOLE(bp)) {
1582 * A block of zeros may compress to a hole, but the
1583 * block size still needs to be known for replay.
1585 BP_SET_LSIZE(bp, db->db_size);
1586 } else if (!BP_IS_EMBEDDED(bp)) {
1587 ASSERT(BP_GET_LEVEL(bp) == 0);
1594 dmu_sync_late_arrival_ready(zio_t *zio)
1596 dmu_sync_ready(zio, NULL, zio->io_private);
1601 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1603 dmu_sync_arg_t *dsa = varg;
1604 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1605 dmu_buf_impl_t *db = dr->dr_dbuf;
1607 mutex_enter(&db->db_mtx);
1608 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1609 if (zio->io_error == 0) {
1610 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1611 if (dr->dt.dl.dr_nopwrite) {
1612 blkptr_t *bp = zio->io_bp;
1613 blkptr_t *bp_orig = &zio->io_bp_orig;
1614 uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1616 ASSERT(BP_EQUAL(bp, bp_orig));
1617 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1618 ASSERT(zio_checksum_table[chksum].ci_flags &
1619 ZCHECKSUM_FLAG_NOPWRITE);
1621 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1622 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1623 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1626 * Old style holes are filled with all zeros, whereas
1627 * new-style holes maintain their lsize, type, level,
1628 * and birth time (see zio_write_compress). While we
1629 * need to reset the BP_SET_LSIZE() call that happened
1630 * in dmu_sync_ready for old style holes, we do *not*
1631 * want to wipe out the information contained in new
1632 * style holes. Thus, only zero out the block pointer if
1633 * it's an old style hole.
1635 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) &&
1636 dr->dt.dl.dr_overridden_by.blk_birth == 0)
1637 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1639 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1641 cv_broadcast(&db->db_changed);
1642 mutex_exit(&db->db_mtx);
1644 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1646 kmem_free(dsa, sizeof (*dsa));
1650 dmu_sync_late_arrival_done(zio_t *zio)
1652 blkptr_t *bp = zio->io_bp;
1653 dmu_sync_arg_t *dsa = zio->io_private;
1654 blkptr_t *bp_orig = &zio->io_bp_orig;
1656 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1658 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE)
1659 * then there is nothing to do here. Otherwise, free the
1660 * newly allocated block in this txg.
1662 if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
1663 ASSERT(BP_EQUAL(bp, bp_orig));
1665 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1666 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1667 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1668 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1672 dmu_tx_commit(dsa->dsa_tx);
1674 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1676 kmem_free(dsa, sizeof (*dsa));
1680 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1681 zio_prop_t *zp, zbookmark_phys_t *zb)
1683 dmu_sync_arg_t *dsa;
1686 tx = dmu_tx_create(os);
1687 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1688 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1690 /* Make zl_get_data do txg_waited_synced() */
1691 return (SET_ERROR(EIO));
1694 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1696 dsa->dsa_done = done;
1700 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx),
1701 zgd->zgd_bp, zgd->zgd_db->db_data, zgd->zgd_db->db_size,
1702 zp, dmu_sync_late_arrival_ready, NULL,
1703 NULL, dmu_sync_late_arrival_done, dsa, ZIO_PRIORITY_SYNC_WRITE,
1704 ZIO_FLAG_CANFAIL, zb));
1710 * Intent log support: sync the block associated with db to disk.
1711 * N.B. and XXX: the caller is responsible for making sure that the
1712 * data isn't changing while dmu_sync() is writing it.
1716 * EEXIST: this txg has already been synced, so there's nothing to do.
1717 * The caller should not log the write.
1719 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1720 * The caller should not log the write.
1722 * EALREADY: this block is already in the process of being synced.
1723 * The caller should track its progress (somehow).
1725 * EIO: could not do the I/O.
1726 * The caller should do a txg_wait_synced().
1728 * 0: the I/O has been initiated.
1729 * The caller should log this blkptr in the done callback.
1730 * It is possible that the I/O will fail, in which case
1731 * the error will be reported to the done callback and
1732 * propagated to pio from zio_done().
1735 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1737 blkptr_t *bp = zgd->zgd_bp;
1738 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1739 objset_t *os = db->db_objset;
1740 dsl_dataset_t *ds = os->os_dsl_dataset;
1741 dbuf_dirty_record_t *dr;
1742 dmu_sync_arg_t *dsa;
1743 zbookmark_phys_t zb;
1747 ASSERT(pio != NULL);
1750 SET_BOOKMARK(&zb, ds->ds_object,
1751 db->db.db_object, db->db_level, db->db_blkid);
1755 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1759 * If we're frozen (running ziltest), we always need to generate a bp.
1761 if (txg > spa_freeze_txg(os->os_spa))
1762 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1765 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1766 * and us. If we determine that this txg is not yet syncing,
1767 * but it begins to sync a moment later, that's OK because the
1768 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1770 mutex_enter(&db->db_mtx);
1772 if (txg <= spa_last_synced_txg(os->os_spa)) {
1774 * This txg has already synced. There's nothing to do.
1776 mutex_exit(&db->db_mtx);
1777 return (SET_ERROR(EEXIST));
1780 if (txg <= spa_syncing_txg(os->os_spa)) {
1782 * This txg is currently syncing, so we can't mess with
1783 * the dirty record anymore; just write a new log block.
1785 mutex_exit(&db->db_mtx);
1786 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1789 dr = db->db_last_dirty;
1790 while (dr && dr->dr_txg != txg)
1795 * There's no dr for this dbuf, so it must have been freed.
1796 * There's no need to log writes to freed blocks, so we're done.
1798 mutex_exit(&db->db_mtx);
1799 return (SET_ERROR(ENOENT));
1802 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
1805 * Assume the on-disk data is X, the current syncing data (in
1806 * txg - 1) is Y, and the current in-memory data is Z (currently
1809 * We usually want to perform a nopwrite if X and Z are the
1810 * same. However, if Y is different (i.e. the BP is going to
1811 * change before this write takes effect), then a nopwrite will
1812 * be incorrect - we would override with X, which could have
1813 * been freed when Y was written.
1815 * (Note that this is not a concern when we are nop-writing from
1816 * syncing context, because X and Y must be identical, because
1817 * all previous txgs have been synced.)
1819 * Therefore, we disable nopwrite if the current BP could change
1820 * before this TXG. There are two ways it could change: by
1821 * being dirty (dr_next is non-NULL), or by being freed
1822 * (dnode_block_freed()). This behavior is verified by
1823 * zio_done(), which VERIFYs that the override BP is identical
1824 * to the on-disk BP.
1828 if (dr->dr_next != NULL || dnode_block_freed(dn, db->db_blkid))
1829 zp.zp_nopwrite = B_FALSE;
1832 ASSERT(dr->dr_txg == txg);
1833 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1834 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1836 * We have already issued a sync write for this buffer,
1837 * or this buffer has already been synced. It could not
1838 * have been dirtied since, or we would have cleared the state.
1840 mutex_exit(&db->db_mtx);
1841 return (SET_ERROR(EALREADY));
1844 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1845 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1846 mutex_exit(&db->db_mtx);
1848 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1850 dsa->dsa_done = done;
1854 zio_nowait(arc_write(pio, os->os_spa, txg,
1855 bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
1856 &zp, dmu_sync_ready, NULL, NULL, dmu_sync_done, dsa,
1857 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb));
1863 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1869 err = dnode_hold(os, object, FTAG, &dn);
1872 err = dnode_set_blksz(dn, size, ibs, tx);
1873 dnode_rele(dn, FTAG);
1878 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1884 * Send streams include each object's checksum function. This
1885 * check ensures that the receiving system can understand the
1886 * checksum function transmitted.
1888 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS);
1890 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1891 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS);
1892 dn->dn_checksum = checksum;
1893 dnode_setdirty(dn, tx);
1894 dnode_rele(dn, FTAG);
1898 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1904 * Send streams include each object's compression function. This
1905 * check ensures that the receiving system can understand the
1906 * compression function transmitted.
1908 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS);
1910 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1911 dn->dn_compress = compress;
1912 dnode_setdirty(dn, tx);
1913 dnode_rele(dn, FTAG);
1916 int zfs_mdcomp_disable = 0;
1917 SYSCTL_INT(_vfs_zfs, OID_AUTO, mdcomp_disable, CTLFLAG_RWTUN,
1918 &zfs_mdcomp_disable, 0, "Disable metadata compression");
1921 * When the "redundant_metadata" property is set to "most", only indirect
1922 * blocks of this level and higher will have an additional ditto block.
1924 int zfs_redundant_metadata_most_ditto_level = 2;
1927 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
1929 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
1930 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
1932 enum zio_checksum checksum = os->os_checksum;
1933 enum zio_compress compress = os->os_compress;
1934 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
1935 boolean_t dedup = B_FALSE;
1936 boolean_t nopwrite = B_FALSE;
1937 boolean_t dedup_verify = os->os_dedup_verify;
1938 int copies = os->os_copies;
1941 * We maintain different write policies for each of the following
1944 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1945 * 3. all other level 0 blocks
1948 if (zfs_mdcomp_disable) {
1949 compress = ZIO_COMPRESS_EMPTY;
1952 * XXX -- we should design a compression algorithm
1953 * that specializes in arrays of bps.
1955 compress = zio_compress_select(os->os_spa,
1956 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON);
1960 * Metadata always gets checksummed. If the data
1961 * checksum is multi-bit correctable, and it's not a
1962 * ZBT-style checksum, then it's suitable for metadata
1963 * as well. Otherwise, the metadata checksum defaults
1966 if (!(zio_checksum_table[checksum].ci_flags &
1967 ZCHECKSUM_FLAG_METADATA) ||
1968 (zio_checksum_table[checksum].ci_flags &
1969 ZCHECKSUM_FLAG_EMBEDDED))
1970 checksum = ZIO_CHECKSUM_FLETCHER_4;
1972 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL ||
1973 (os->os_redundant_metadata ==
1974 ZFS_REDUNDANT_METADATA_MOST &&
1975 (level >= zfs_redundant_metadata_most_ditto_level ||
1976 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))))
1978 } else if (wp & WP_NOFILL) {
1982 * If we're writing preallocated blocks, we aren't actually
1983 * writing them so don't set any policy properties. These
1984 * blocks are currently only used by an external subsystem
1985 * outside of zfs (i.e. dump) and not written by the zio
1988 compress = ZIO_COMPRESS_OFF;
1989 checksum = ZIO_CHECKSUM_NOPARITY;
1991 compress = zio_compress_select(os->os_spa, dn->dn_compress,
1994 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
1995 zio_checksum_select(dn->dn_checksum, checksum) :
1999 * Determine dedup setting. If we are in dmu_sync(),
2000 * we won't actually dedup now because that's all
2001 * done in syncing context; but we do want to use the
2002 * dedup checkum. If the checksum is not strong
2003 * enough to ensure unique signatures, force
2006 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
2007 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
2008 if (!(zio_checksum_table[checksum].ci_flags &
2009 ZCHECKSUM_FLAG_DEDUP))
2010 dedup_verify = B_TRUE;
2014 * Enable nopwrite if we have secure enough checksum
2015 * algorithm (see comment in zio_nop_write) and
2016 * compression is enabled. We don't enable nopwrite if
2017 * dedup is enabled as the two features are mutually
2020 nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags &
2021 ZCHECKSUM_FLAG_NOPWRITE) &&
2022 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
2025 zp->zp_checksum = checksum;
2026 zp->zp_compress = compress;
2027 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
2028 zp->zp_level = level;
2029 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
2030 zp->zp_dedup = dedup;
2031 zp->zp_dedup_verify = dedup && dedup_verify;
2032 zp->zp_nopwrite = nopwrite;
2036 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
2042 * Sync any current changes before
2043 * we go trundling through the block pointers.
2045 err = dmu_object_wait_synced(os, object);
2050 err = dnode_hold(os, object, FTAG, &dn);
2055 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
2056 dnode_rele(dn, FTAG);
2062 * Given the ZFS object, if it contains any dirty nodes
2063 * this function flushes all dirty blocks to disk. This
2064 * ensures the DMU object info is updated. A more efficient
2065 * future version might just find the TXG with the maximum
2066 * ID and wait for that to be synced.
2069 dmu_object_wait_synced(objset_t *os, uint64_t object)
2074 error = dnode_hold(os, object, FTAG, &dn);
2079 for (i = 0; i < TXG_SIZE; i++) {
2080 if (list_link_active(&dn->dn_dirty_link[i])) {
2084 dnode_rele(dn, FTAG);
2085 if (i != TXG_SIZE) {
2086 txg_wait_synced(dmu_objset_pool(os), 0);
2093 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2097 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2098 mutex_enter(&dn->dn_mtx);
2102 doi->doi_data_block_size = dn->dn_datablksz;
2103 doi->doi_metadata_block_size = dn->dn_indblkshift ?
2104 1ULL << dn->dn_indblkshift : 0;
2105 doi->doi_type = dn->dn_type;
2106 doi->doi_bonus_type = dn->dn_bonustype;
2107 doi->doi_bonus_size = dn->dn_bonuslen;
2108 doi->doi_indirection = dn->dn_nlevels;
2109 doi->doi_checksum = dn->dn_checksum;
2110 doi->doi_compress = dn->dn_compress;
2111 doi->doi_nblkptr = dn->dn_nblkptr;
2112 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
2113 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
2114 doi->doi_fill_count = 0;
2115 for (int i = 0; i < dnp->dn_nblkptr; i++)
2116 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
2118 mutex_exit(&dn->dn_mtx);
2119 rw_exit(&dn->dn_struct_rwlock);
2123 * Get information on a DMU object.
2124 * If doi is NULL, just indicates whether the object exists.
2127 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
2130 int err = dnode_hold(os, object, FTAG, &dn);
2136 dmu_object_info_from_dnode(dn, doi);
2138 dnode_rele(dn, FTAG);
2143 * As above, but faster; can be used when you have a held dbuf in hand.
2146 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
2148 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2151 dmu_object_info_from_dnode(DB_DNODE(db), doi);
2156 * Faster still when you only care about the size.
2157 * This is specifically optimized for zfs_getattr().
2160 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
2161 u_longlong_t *nblk512)
2163 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2169 *blksize = dn->dn_datablksz;
2170 /* add 1 for dnode space */
2171 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
2172 SPA_MINBLOCKSHIFT) + 1;
2177 byteswap_uint64_array(void *vbuf, size_t size)
2179 uint64_t *buf = vbuf;
2180 size_t count = size >> 3;
2183 ASSERT((size & 7) == 0);
2185 for (i = 0; i < count; i++)
2186 buf[i] = BSWAP_64(buf[i]);
2190 byteswap_uint32_array(void *vbuf, size_t size)
2192 uint32_t *buf = vbuf;
2193 size_t count = size >> 2;
2196 ASSERT((size & 3) == 0);
2198 for (i = 0; i < count; i++)
2199 buf[i] = BSWAP_32(buf[i]);
2203 byteswap_uint16_array(void *vbuf, size_t size)
2205 uint16_t *buf = vbuf;
2206 size_t count = size >> 1;
2209 ASSERT((size & 1) == 0);
2211 for (i = 0; i < count; i++)
2212 buf[i] = BSWAP_16(buf[i]);
2217 byteswap_uint8_array(void *vbuf, size_t size)
2230 zio_compress_init();
2239 arc_fini(); /* arc depends on l2arc, so arc must go first */
2242 zio_compress_fini();