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");
63 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
64 { DMU_BSWAP_UINT8, TRUE, "unallocated" },
65 { DMU_BSWAP_ZAP, TRUE, "object directory" },
66 { DMU_BSWAP_UINT64, TRUE, "object array" },
67 { DMU_BSWAP_UINT8, TRUE, "packed nvlist" },
68 { DMU_BSWAP_UINT64, TRUE, "packed nvlist size" },
69 { DMU_BSWAP_UINT64, TRUE, "bpobj" },
70 { DMU_BSWAP_UINT64, TRUE, "bpobj header" },
71 { DMU_BSWAP_UINT64, TRUE, "SPA space map header" },
72 { DMU_BSWAP_UINT64, TRUE, "SPA space map" },
73 { DMU_BSWAP_UINT64, TRUE, "ZIL intent log" },
74 { DMU_BSWAP_DNODE, TRUE, "DMU dnode" },
75 { DMU_BSWAP_OBJSET, TRUE, "DMU objset" },
76 { DMU_BSWAP_UINT64, TRUE, "DSL directory" },
77 { DMU_BSWAP_ZAP, TRUE, "DSL directory child map"},
78 { DMU_BSWAP_ZAP, TRUE, "DSL dataset snap map" },
79 { DMU_BSWAP_ZAP, TRUE, "DSL props" },
80 { DMU_BSWAP_UINT64, TRUE, "DSL dataset" },
81 { DMU_BSWAP_ZNODE, TRUE, "ZFS znode" },
82 { DMU_BSWAP_OLDACL, TRUE, "ZFS V0 ACL" },
83 { DMU_BSWAP_UINT8, FALSE, "ZFS plain file" },
84 { DMU_BSWAP_ZAP, TRUE, "ZFS directory" },
85 { DMU_BSWAP_ZAP, TRUE, "ZFS master node" },
86 { DMU_BSWAP_ZAP, TRUE, "ZFS delete queue" },
87 { DMU_BSWAP_UINT8, FALSE, "zvol object" },
88 { DMU_BSWAP_ZAP, TRUE, "zvol prop" },
89 { DMU_BSWAP_UINT8, FALSE, "other uint8[]" },
90 { DMU_BSWAP_UINT64, FALSE, "other uint64[]" },
91 { DMU_BSWAP_ZAP, TRUE, "other ZAP" },
92 { DMU_BSWAP_ZAP, TRUE, "persistent error log" },
93 { DMU_BSWAP_UINT8, TRUE, "SPA history" },
94 { DMU_BSWAP_UINT64, TRUE, "SPA history offsets" },
95 { DMU_BSWAP_ZAP, TRUE, "Pool properties" },
96 { DMU_BSWAP_ZAP, TRUE, "DSL permissions" },
97 { DMU_BSWAP_ACL, TRUE, "ZFS ACL" },
98 { DMU_BSWAP_UINT8, TRUE, "ZFS SYSACL" },
99 { DMU_BSWAP_UINT8, TRUE, "FUID table" },
100 { DMU_BSWAP_UINT64, TRUE, "FUID table size" },
101 { DMU_BSWAP_ZAP, TRUE, "DSL dataset next clones"},
102 { DMU_BSWAP_ZAP, TRUE, "scan work queue" },
103 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group used" },
104 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group quota" },
105 { DMU_BSWAP_ZAP, TRUE, "snapshot refcount tags"},
106 { DMU_BSWAP_ZAP, TRUE, "DDT ZAP algorithm" },
107 { DMU_BSWAP_ZAP, TRUE, "DDT statistics" },
108 { DMU_BSWAP_UINT8, TRUE, "System attributes" },
109 { DMU_BSWAP_ZAP, TRUE, "SA master node" },
110 { DMU_BSWAP_ZAP, TRUE, "SA attr registration" },
111 { DMU_BSWAP_ZAP, TRUE, "SA attr layouts" },
112 { DMU_BSWAP_ZAP, TRUE, "scan translations" },
113 { DMU_BSWAP_UINT8, FALSE, "deduplicated block" },
114 { DMU_BSWAP_ZAP, TRUE, "DSL deadlist map" },
115 { DMU_BSWAP_UINT64, TRUE, "DSL deadlist map hdr" },
116 { DMU_BSWAP_ZAP, TRUE, "DSL dir clones" },
117 { DMU_BSWAP_UINT64, TRUE, "bpobj subobj" }
120 const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
121 { byteswap_uint8_array, "uint8" },
122 { byteswap_uint16_array, "uint16" },
123 { byteswap_uint32_array, "uint32" },
124 { byteswap_uint64_array, "uint64" },
125 { zap_byteswap, "zap" },
126 { dnode_buf_byteswap, "dnode" },
127 { dmu_objset_byteswap, "objset" },
128 { zfs_znode_byteswap, "znode" },
129 { zfs_oldacl_byteswap, "oldacl" },
130 { zfs_acl_byteswap, "acl" }
134 dmu_buf_hold_noread_by_dnode(dnode_t *dn, uint64_t offset,
135 void *tag, dmu_buf_t **dbp)
140 blkid = dbuf_whichblock(dn, 0, offset);
141 rw_enter(&dn->dn_struct_rwlock, RW_READER);
142 db = dbuf_hold(dn, blkid, tag);
143 rw_exit(&dn->dn_struct_rwlock);
147 return (SET_ERROR(EIO));
154 dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset,
155 void *tag, dmu_buf_t **dbp)
162 err = dnode_hold(os, object, FTAG, &dn);
165 blkid = dbuf_whichblock(dn, 0, offset);
166 rw_enter(&dn->dn_struct_rwlock, RW_READER);
167 db = dbuf_hold(dn, blkid, tag);
168 rw_exit(&dn->dn_struct_rwlock);
169 dnode_rele(dn, FTAG);
173 return (SET_ERROR(EIO));
181 dmu_buf_hold_by_dnode(dnode_t *dn, uint64_t offset,
182 void *tag, dmu_buf_t **dbp, int flags)
185 int db_flags = DB_RF_CANFAIL;
187 if (flags & DMU_READ_NO_PREFETCH)
188 db_flags |= DB_RF_NOPREFETCH;
190 err = dmu_buf_hold_noread_by_dnode(dn, offset, tag, dbp);
192 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
193 err = dbuf_read(db, NULL, db_flags);
204 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
205 void *tag, dmu_buf_t **dbp, int flags)
208 int db_flags = DB_RF_CANFAIL;
210 if (flags & DMU_READ_NO_PREFETCH)
211 db_flags |= DB_RF_NOPREFETCH;
213 err = dmu_buf_hold_noread(os, object, offset, tag, dbp);
215 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
216 err = dbuf_read(db, NULL, db_flags);
229 return (DN_MAX_BONUSLEN);
233 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
235 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
242 if (dn->dn_bonus != db) {
243 error = SET_ERROR(EINVAL);
244 } else if (newsize < 0 || newsize > db_fake->db_size) {
245 error = SET_ERROR(EINVAL);
247 dnode_setbonuslen(dn, newsize, tx);
256 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
258 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
265 if (!DMU_OT_IS_VALID(type)) {
266 error = SET_ERROR(EINVAL);
267 } else if (dn->dn_bonus != db) {
268 error = SET_ERROR(EINVAL);
270 dnode_setbonus_type(dn, type, tx);
279 dmu_get_bonustype(dmu_buf_t *db_fake)
281 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
283 dmu_object_type_t type;
287 type = dn->dn_bonustype;
294 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
299 error = dnode_hold(os, object, FTAG, &dn);
300 dbuf_rm_spill(dn, tx);
301 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
302 dnode_rm_spill(dn, tx);
303 rw_exit(&dn->dn_struct_rwlock);
304 dnode_rele(dn, FTAG);
309 * returns ENOENT, EIO, or 0.
312 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
318 error = dnode_hold(os, object, FTAG, &dn);
322 rw_enter(&dn->dn_struct_rwlock, RW_READER);
323 if (dn->dn_bonus == NULL) {
324 rw_exit(&dn->dn_struct_rwlock);
325 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
326 if (dn->dn_bonus == NULL)
327 dbuf_create_bonus(dn);
331 /* as long as the bonus buf is held, the dnode will be held */
332 if (refcount_add(&db->db_holds, tag) == 1) {
333 VERIFY(dnode_add_ref(dn, db));
334 atomic_inc_32(&dn->dn_dbufs_count);
338 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
339 * hold and incrementing the dbuf count to ensure that dnode_move() sees
340 * a dnode hold for every dbuf.
342 rw_exit(&dn->dn_struct_rwlock);
344 dnode_rele(dn, FTAG);
346 VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH));
353 * returns ENOENT, EIO, or 0.
355 * This interface will allocate a blank spill dbuf when a spill blk
356 * doesn't already exist on the dnode.
358 * if you only want to find an already existing spill db, then
359 * dmu_spill_hold_existing() should be used.
362 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
364 dmu_buf_impl_t *db = NULL;
367 if ((flags & DB_RF_HAVESTRUCT) == 0)
368 rw_enter(&dn->dn_struct_rwlock, RW_READER);
370 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
372 if ((flags & DB_RF_HAVESTRUCT) == 0)
373 rw_exit(&dn->dn_struct_rwlock);
376 err = dbuf_read(db, NULL, flags);
385 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
387 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
394 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
395 err = SET_ERROR(EINVAL);
397 rw_enter(&dn->dn_struct_rwlock, RW_READER);
399 if (!dn->dn_have_spill) {
400 err = SET_ERROR(ENOENT);
402 err = dmu_spill_hold_by_dnode(dn,
403 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
406 rw_exit(&dn->dn_struct_rwlock);
414 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
416 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
422 err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp);
429 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
430 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
431 * and can induce severe lock contention when writing to several files
432 * whose dnodes are in the same block.
435 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
436 boolean_t read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
439 uint64_t blkid, nblks, i;
444 ASSERT(length <= DMU_MAX_ACCESS);
447 * Note: We directly notify the prefetch code of this read, so that
448 * we can tell it about the multi-block read. dbuf_read() only knows
449 * about the one block it is accessing.
451 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT |
454 rw_enter(&dn->dn_struct_rwlock, RW_READER);
455 if (dn->dn_datablkshift) {
456 int blkshift = dn->dn_datablkshift;
457 nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) -
458 P2ALIGN(offset, 1ULL << blkshift)) >> blkshift;
460 if (offset + length > dn->dn_datablksz) {
461 zfs_panic_recover("zfs: accessing past end of object "
462 "%llx/%llx (size=%u access=%llu+%llu)",
463 (longlong_t)dn->dn_objset->
464 os_dsl_dataset->ds_object,
465 (longlong_t)dn->dn_object, dn->dn_datablksz,
466 (longlong_t)offset, (longlong_t)length);
467 rw_exit(&dn->dn_struct_rwlock);
468 return (SET_ERROR(EIO));
472 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
474 #if defined(_KERNEL) && defined(RACCT)
475 if (racct_enable && !read) {
477 racct_add_force(curproc, RACCT_WRITEBPS, length);
478 racct_add_force(curproc, RACCT_WRITEIOPS, nblks);
479 PROC_UNLOCK(curproc);
483 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
484 blkid = dbuf_whichblock(dn, 0, offset);
485 for (i = 0; i < nblks; i++) {
486 dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag);
488 rw_exit(&dn->dn_struct_rwlock);
489 dmu_buf_rele_array(dbp, nblks, tag);
491 return (SET_ERROR(EIO));
494 /* initiate async i/o */
496 (void) dbuf_read(db, zio, dbuf_flags);
499 curthread->td_ru.ru_oublock++;
504 if ((flags & DMU_READ_NO_PREFETCH) == 0 &&
505 DNODE_META_IS_CACHEABLE(dn) && length <= zfetch_array_rd_sz) {
506 dmu_zfetch(&dn->dn_zfetch, blkid, nblks,
507 read && DNODE_IS_CACHEABLE(dn));
509 rw_exit(&dn->dn_struct_rwlock);
511 /* wait for async i/o */
514 dmu_buf_rele_array(dbp, nblks, tag);
518 /* wait for other io to complete */
520 for (i = 0; i < nblks; i++) {
521 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
522 mutex_enter(&db->db_mtx);
523 while (db->db_state == DB_READ ||
524 db->db_state == DB_FILL)
525 cv_wait(&db->db_changed, &db->db_mtx);
526 if (db->db_state == DB_UNCACHED)
527 err = SET_ERROR(EIO);
528 mutex_exit(&db->db_mtx);
530 dmu_buf_rele_array(dbp, nblks, tag);
542 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
543 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
548 err = dnode_hold(os, object, FTAG, &dn);
552 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
553 numbufsp, dbpp, DMU_READ_PREFETCH);
555 dnode_rele(dn, FTAG);
561 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
562 uint64_t length, boolean_t read, void *tag, int *numbufsp,
565 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
571 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
572 numbufsp, dbpp, DMU_READ_PREFETCH);
579 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
582 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
587 for (i = 0; i < numbufs; i++) {
589 dbuf_rele(dbp[i], tag);
592 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
596 * Issue prefetch i/os for the given blocks. If level is greater than 0, the
597 * indirect blocks prefeteched will be those that point to the blocks containing
598 * the data starting at offset, and continuing to offset + len.
600 * Note that if the indirect blocks above the blocks being prefetched are not in
601 * cache, they will be asychronously read in.
604 dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset,
605 uint64_t len, zio_priority_t pri)
611 if (len == 0) { /* they're interested in the bonus buffer */
612 dn = DMU_META_DNODE(os);
614 if (object == 0 || object >= DN_MAX_OBJECT)
617 rw_enter(&dn->dn_struct_rwlock, RW_READER);
618 blkid = dbuf_whichblock(dn, level,
619 object * sizeof (dnode_phys_t));
620 dbuf_prefetch(dn, level, blkid, pri, 0);
621 rw_exit(&dn->dn_struct_rwlock);
626 * XXX - Note, if the dnode for the requested object is not
627 * already cached, we will do a *synchronous* read in the
628 * dnode_hold() call. The same is true for any indirects.
630 err = dnode_hold(os, object, FTAG, &dn);
634 rw_enter(&dn->dn_struct_rwlock, RW_READER);
636 * offset + len - 1 is the last byte we want to prefetch for, and offset
637 * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the
638 * last block we want to prefetch, and dbuf_whichblock(dn, level,
639 * offset) is the first. Then the number we need to prefetch is the
642 if (level > 0 || dn->dn_datablkshift != 0) {
643 nblks = dbuf_whichblock(dn, level, offset + len - 1) -
644 dbuf_whichblock(dn, level, offset) + 1;
646 nblks = (offset < dn->dn_datablksz);
650 blkid = dbuf_whichblock(dn, level, offset);
651 for (int i = 0; i < nblks; i++)
652 dbuf_prefetch(dn, level, blkid + i, pri, 0);
655 rw_exit(&dn->dn_struct_rwlock);
657 dnode_rele(dn, FTAG);
661 * Get the next "chunk" of file data to free. We traverse the file from
662 * the end so that the file gets shorter over time (if we crashes in the
663 * middle, this will leave us in a better state). We find allocated file
664 * data by simply searching the allocated level 1 indirects.
666 * On input, *start should be the first offset that does not need to be
667 * freed (e.g. "offset + length"). On return, *start will be the first
668 * offset that should be freed.
671 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum)
673 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
674 /* bytes of data covered by a level-1 indirect block */
676 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
678 ASSERT3U(minimum, <=, *start);
680 if (*start - minimum <= iblkrange * maxblks) {
684 ASSERT(ISP2(iblkrange));
686 for (uint64_t blks = 0; *start > minimum && blks < maxblks; blks++) {
690 * dnode_next_offset(BACKWARDS) will find an allocated L1
691 * indirect block at or before the input offset. We must
692 * decrement *start so that it is at the end of the region
696 err = dnode_next_offset(dn,
697 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
699 /* if there are no indirect blocks before start, we are done */
703 } else if (err != 0) {
707 /* set start to the beginning of this L1 indirect */
708 *start = P2ALIGN(*start, iblkrange);
710 if (*start < minimum)
716 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
719 uint64_t object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
722 if (offset >= object_size)
725 if (length == DMU_OBJECT_END || offset + length > object_size)
726 length = object_size - offset;
728 while (length != 0) {
729 uint64_t chunk_end, chunk_begin;
731 chunk_end = chunk_begin = offset + length;
733 /* move chunk_begin backwards to the beginning of this chunk */
734 err = get_next_chunk(dn, &chunk_begin, offset);
737 ASSERT3U(chunk_begin, >=, offset);
738 ASSERT3U(chunk_begin, <=, chunk_end);
740 dmu_tx_t *tx = dmu_tx_create(os);
741 dmu_tx_hold_free(tx, dn->dn_object,
742 chunk_begin, chunk_end - chunk_begin);
745 * Mark this transaction as typically resulting in a net
746 * reduction in space used.
748 dmu_tx_mark_netfree(tx);
749 err = dmu_tx_assign(tx, TXG_WAIT);
754 dnode_free_range(dn, chunk_begin, chunk_end - chunk_begin, tx);
757 length -= chunk_end - chunk_begin;
763 dmu_free_long_range(objset_t *os, uint64_t object,
764 uint64_t offset, uint64_t length)
769 err = dnode_hold(os, object, FTAG, &dn);
772 err = dmu_free_long_range_impl(os, dn, offset, length);
775 * It is important to zero out the maxblkid when freeing the entire
776 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
777 * will take the fast path, and (b) dnode_reallocate() can verify
778 * that the entire file has been freed.
780 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
783 dnode_rele(dn, FTAG);
788 dmu_free_long_object(objset_t *os, uint64_t object)
793 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
797 tx = dmu_tx_create(os);
798 dmu_tx_hold_bonus(tx, object);
799 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
800 dmu_tx_mark_netfree(tx);
801 err = dmu_tx_assign(tx, TXG_WAIT);
803 err = dmu_object_free(os, object, tx);
813 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
814 uint64_t size, dmu_tx_t *tx)
817 int err = dnode_hold(os, object, FTAG, &dn);
820 ASSERT(offset < UINT64_MAX);
821 ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
822 dnode_free_range(dn, offset, size, tx);
823 dnode_rele(dn, FTAG);
828 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
829 void *buf, uint32_t flags)
835 err = dnode_hold(os, object, FTAG, &dn);
840 * Deal with odd block sizes, where there can't be data past the first
841 * block. If we ever do the tail block optimization, we will need to
842 * handle that here as well.
844 if (dn->dn_maxblkid == 0) {
845 int newsz = offset > dn->dn_datablksz ? 0 :
846 MIN(size, dn->dn_datablksz - offset);
847 bzero((char *)buf + newsz, size - newsz);
852 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
856 * NB: we could do this block-at-a-time, but it's nice
857 * to be reading in parallel.
859 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
860 TRUE, FTAG, &numbufs, &dbp, flags);
864 for (i = 0; i < numbufs; i++) {
867 dmu_buf_t *db = dbp[i];
871 bufoff = offset - db->db_offset;
872 tocpy = (int)MIN(db->db_size - bufoff, size);
874 bcopy((char *)db->db_data + bufoff, buf, tocpy);
878 buf = (char *)buf + tocpy;
880 dmu_buf_rele_array(dbp, numbufs, FTAG);
882 dnode_rele(dn, FTAG);
887 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
888 const void *buf, dmu_tx_t *tx)
896 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
897 FALSE, FTAG, &numbufs, &dbp));
899 for (i = 0; i < numbufs; i++) {
902 dmu_buf_t *db = dbp[i];
906 bufoff = offset - db->db_offset;
907 tocpy = (int)MIN(db->db_size - bufoff, size);
909 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
911 if (tocpy == db->db_size)
912 dmu_buf_will_fill(db, tx);
914 dmu_buf_will_dirty(db, tx);
916 bcopy(buf, (char *)db->db_data + bufoff, tocpy);
918 if (tocpy == db->db_size)
919 dmu_buf_fill_done(db, tx);
923 buf = (char *)buf + tocpy;
925 dmu_buf_rele_array(dbp, numbufs, FTAG);
929 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
938 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
939 FALSE, FTAG, &numbufs, &dbp));
941 for (i = 0; i < numbufs; i++) {
942 dmu_buf_t *db = dbp[i];
944 dmu_buf_will_not_fill(db, tx);
946 dmu_buf_rele_array(dbp, numbufs, FTAG);
950 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
951 void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
952 int compressed_size, int byteorder, dmu_tx_t *tx)
956 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
957 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
958 VERIFY0(dmu_buf_hold_noread(os, object, offset,
961 dmu_buf_write_embedded(db,
962 data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
963 uncompressed_size, compressed_size, byteorder, tx);
965 dmu_buf_rele(db, FTAG);
969 * DMU support for xuio
971 kstat_t *xuio_ksp = NULL;
974 dmu_xuio_init(xuio_t *xuio, int nblk)
977 uio_t *uio = &xuio->xu_uio;
979 uio->uio_iovcnt = nblk;
980 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
982 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
984 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
985 priv->iovp = uio->uio_iov;
986 XUIO_XUZC_PRIV(xuio) = priv;
988 if (XUIO_XUZC_RW(xuio) == UIO_READ)
989 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
991 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
997 dmu_xuio_fini(xuio_t *xuio)
999 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1000 int nblk = priv->cnt;
1002 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
1003 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
1004 kmem_free(priv, sizeof (dmu_xuio_t));
1006 if (XUIO_XUZC_RW(xuio) == UIO_READ)
1007 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
1009 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
1013 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
1014 * and increase priv->next by 1.
1017 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
1020 uio_t *uio = &xuio->xu_uio;
1021 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1022 int i = priv->next++;
1024 ASSERT(i < priv->cnt);
1025 ASSERT(off + n <= arc_buf_size(abuf));
1026 iov = uio->uio_iov + i;
1027 iov->iov_base = (char *)abuf->b_data + off;
1029 priv->bufs[i] = abuf;
1034 dmu_xuio_cnt(xuio_t *xuio)
1036 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1041 dmu_xuio_arcbuf(xuio_t *xuio, int i)
1043 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1045 ASSERT(i < priv->cnt);
1046 return (priv->bufs[i]);
1050 dmu_xuio_clear(xuio_t *xuio, int i)
1052 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1054 ASSERT(i < priv->cnt);
1055 priv->bufs[i] = NULL;
1059 xuio_stat_init(void)
1061 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
1062 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
1063 KSTAT_FLAG_VIRTUAL);
1064 if (xuio_ksp != NULL) {
1065 xuio_ksp->ks_data = &xuio_stats;
1066 kstat_install(xuio_ksp);
1071 xuio_stat_fini(void)
1073 if (xuio_ksp != NULL) {
1074 kstat_delete(xuio_ksp);
1080 xuio_stat_wbuf_copied()
1082 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1086 xuio_stat_wbuf_nocopy()
1088 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
1093 dmu_read_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size)
1096 int numbufs, i, err;
1097 xuio_t *xuio = NULL;
1100 * NB: we could do this block-at-a-time, but it's nice
1101 * to be reading in parallel.
1103 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1104 TRUE, FTAG, &numbufs, &dbp, 0);
1109 if (uio->uio_extflg == UIO_XUIO)
1110 xuio = (xuio_t *)uio;
1113 for (i = 0; i < numbufs; i++) {
1116 dmu_buf_t *db = dbp[i];
1120 bufoff = uio->uio_loffset - db->db_offset;
1121 tocpy = (int)MIN(db->db_size - bufoff, size);
1124 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
1125 arc_buf_t *dbuf_abuf = dbi->db_buf;
1126 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
1127 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
1129 uio->uio_resid -= tocpy;
1130 uio->uio_loffset += tocpy;
1133 if (abuf == dbuf_abuf)
1134 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
1136 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
1139 err = uiomove((char *)db->db_data + bufoff, tocpy,
1142 err = vn_io_fault_uiomove((char *)db->db_data + bufoff,
1151 dmu_buf_rele_array(dbp, numbufs, FTAG);
1157 * Read 'size' bytes into the uio buffer.
1158 * From object zdb->db_object.
1159 * Starting at offset uio->uio_loffset.
1161 * If the caller already has a dbuf in the target object
1162 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1163 * because we don't have to find the dnode_t for the object.
1166 dmu_read_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size)
1168 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1177 err = dmu_read_uio_dnode(dn, uio, size);
1184 * Read 'size' bytes into the uio buffer.
1185 * From the specified object
1186 * Starting at offset uio->uio_loffset.
1189 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
1197 err = dnode_hold(os, object, FTAG, &dn);
1201 err = dmu_read_uio_dnode(dn, uio, size);
1203 dnode_rele(dn, FTAG);
1209 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1216 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1217 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1221 for (i = 0; i < numbufs; i++) {
1224 dmu_buf_t *db = dbp[i];
1228 bufoff = uio->uio_loffset - db->db_offset;
1229 tocpy = (int)MIN(db->db_size - bufoff, size);
1231 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1233 if (tocpy == db->db_size)
1234 dmu_buf_will_fill(db, tx);
1236 dmu_buf_will_dirty(db, tx);
1240 * XXX uiomove could block forever (eg. nfs-backed
1241 * pages). There needs to be a uiolockdown() function
1242 * to lock the pages in memory, so that uiomove won't
1245 err = uiomove((char *)db->db_data + bufoff, tocpy,
1248 err = vn_io_fault_uiomove((char *)db->db_data + bufoff, tocpy,
1252 if (tocpy == db->db_size)
1253 dmu_buf_fill_done(db, tx);
1261 dmu_buf_rele_array(dbp, numbufs, FTAG);
1266 * Write 'size' bytes from the uio buffer.
1267 * To object zdb->db_object.
1268 * Starting at offset uio->uio_loffset.
1270 * If the caller already has a dbuf in the target object
1271 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1272 * because we don't have to find the dnode_t for the object.
1275 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1278 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1287 err = dmu_write_uio_dnode(dn, uio, size, tx);
1294 * Write 'size' bytes from the uio buffer.
1295 * To the specified object.
1296 * Starting at offset uio->uio_loffset.
1299 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1308 err = dnode_hold(os, object, FTAG, &dn);
1312 err = dmu_write_uio_dnode(dn, uio, size, tx);
1314 dnode_rele(dn, FTAG);
1321 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1322 page_t *pp, dmu_tx_t *tx)
1331 err = dmu_buf_hold_array(os, object, offset, size,
1332 FALSE, FTAG, &numbufs, &dbp);
1336 for (i = 0; i < numbufs; i++) {
1337 int tocpy, copied, thiscpy;
1339 dmu_buf_t *db = dbp[i];
1343 ASSERT3U(db->db_size, >=, PAGESIZE);
1345 bufoff = offset - db->db_offset;
1346 tocpy = (int)MIN(db->db_size - bufoff, size);
1348 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1350 if (tocpy == db->db_size)
1351 dmu_buf_will_fill(db, tx);
1353 dmu_buf_will_dirty(db, tx);
1355 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1356 ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff);
1357 thiscpy = MIN(PAGESIZE, tocpy - copied);
1358 va = zfs_map_page(pp, S_READ);
1359 bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1360 zfs_unmap_page(pp, va);
1365 if (tocpy == db->db_size)
1366 dmu_buf_fill_done(db, tx);
1371 dmu_buf_rele_array(dbp, numbufs, FTAG);
1375 #else /* !illumos */
1378 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1379 vm_page_t *ma, dmu_tx_t *tx)
1389 err = dmu_buf_hold_array(os, object, offset, size,
1390 FALSE, FTAG, &numbufs, &dbp);
1394 for (i = 0; i < numbufs; i++) {
1395 int tocpy, copied, thiscpy;
1397 dmu_buf_t *db = dbp[i];
1401 ASSERT3U(db->db_size, >=, PAGESIZE);
1403 bufoff = offset - db->db_offset;
1404 tocpy = (int)MIN(db->db_size - bufoff, size);
1406 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1408 if (tocpy == db->db_size)
1409 dmu_buf_will_fill(db, tx);
1411 dmu_buf_will_dirty(db, tx);
1413 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1414 ASSERT3U(ptoa((*ma)->pindex), ==, db->db_offset + bufoff);
1415 thiscpy = MIN(PAGESIZE, tocpy - copied);
1416 va = zfs_map_page(*ma, &sf);
1417 bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1423 if (tocpy == db->db_size)
1424 dmu_buf_fill_done(db, tx);
1429 dmu_buf_rele_array(dbp, numbufs, FTAG);
1432 #endif /* illumos */
1433 #endif /* _KERNEL */
1436 * Allocate a loaned anonymous arc buffer.
1439 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1441 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1443 return (arc_loan_buf(db->db_objset->os_spa, size));
1447 * Free a loaned arc buffer.
1450 dmu_return_arcbuf(arc_buf_t *buf)
1452 arc_return_buf(buf, FTAG);
1453 arc_buf_destroy(buf, FTAG);
1457 * When possible directly assign passed loaned arc buffer to a dbuf.
1458 * If this is not possible copy the contents of passed arc buf via
1462 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1465 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1468 uint32_t blksz = (uint32_t)arc_buf_size(buf);
1471 DB_DNODE_ENTER(dbuf);
1472 dn = DB_DNODE(dbuf);
1473 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1474 blkid = dbuf_whichblock(dn, 0, offset);
1475 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1476 rw_exit(&dn->dn_struct_rwlock);
1477 DB_DNODE_EXIT(dbuf);
1480 * We can only assign if the offset is aligned, the arc buf is the
1481 * same size as the dbuf, and the dbuf is not metadata. It
1482 * can't be metadata because the loaned arc buf comes from the
1483 * user-data kmem arena.
1485 if (offset == db->db.db_offset && blksz == db->db.db_size &&
1486 DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA) {
1488 curthread->td_ru.ru_oublock++;
1492 racct_add_force(curproc, RACCT_WRITEBPS, blksz);
1493 racct_add_force(curproc, RACCT_WRITEIOPS, 1);
1494 PROC_UNLOCK(curproc);
1497 #endif /* _KERNEL */
1498 dbuf_assign_arcbuf(db, buf, tx);
1499 dbuf_rele(db, FTAG);
1504 DB_DNODE_ENTER(dbuf);
1505 dn = DB_DNODE(dbuf);
1507 object = dn->dn_object;
1508 DB_DNODE_EXIT(dbuf);
1510 dbuf_rele(db, FTAG);
1511 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1512 dmu_return_arcbuf(buf);
1513 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1518 dbuf_dirty_record_t *dsa_dr;
1519 dmu_sync_cb_t *dsa_done;
1526 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1528 dmu_sync_arg_t *dsa = varg;
1529 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1530 blkptr_t *bp = zio->io_bp;
1532 if (zio->io_error == 0) {
1533 if (BP_IS_HOLE(bp)) {
1535 * A block of zeros may compress to a hole, but the
1536 * block size still needs to be known for replay.
1538 BP_SET_LSIZE(bp, db->db_size);
1539 } else if (!BP_IS_EMBEDDED(bp)) {
1540 ASSERT(BP_GET_LEVEL(bp) == 0);
1547 dmu_sync_late_arrival_ready(zio_t *zio)
1549 dmu_sync_ready(zio, NULL, zio->io_private);
1554 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1556 dmu_sync_arg_t *dsa = varg;
1557 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1558 dmu_buf_impl_t *db = dr->dr_dbuf;
1560 mutex_enter(&db->db_mtx);
1561 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1562 if (zio->io_error == 0) {
1563 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1564 if (dr->dt.dl.dr_nopwrite) {
1565 blkptr_t *bp = zio->io_bp;
1566 blkptr_t *bp_orig = &zio->io_bp_orig;
1567 uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1569 ASSERT(BP_EQUAL(bp, bp_orig));
1570 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1571 ASSERT(zio_checksum_table[chksum].ci_flags &
1572 ZCHECKSUM_FLAG_NOPWRITE);
1574 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1575 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1576 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1579 * Old style holes are filled with all zeros, whereas
1580 * new-style holes maintain their lsize, type, level,
1581 * and birth time (see zio_write_compress). While we
1582 * need to reset the BP_SET_LSIZE() call that happened
1583 * in dmu_sync_ready for old style holes, we do *not*
1584 * want to wipe out the information contained in new
1585 * style holes. Thus, only zero out the block pointer if
1586 * it's an old style hole.
1588 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) &&
1589 dr->dt.dl.dr_overridden_by.blk_birth == 0)
1590 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1592 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1594 cv_broadcast(&db->db_changed);
1595 mutex_exit(&db->db_mtx);
1597 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1599 kmem_free(dsa, sizeof (*dsa));
1603 dmu_sync_late_arrival_done(zio_t *zio)
1605 blkptr_t *bp = zio->io_bp;
1606 dmu_sync_arg_t *dsa = zio->io_private;
1607 blkptr_t *bp_orig = &zio->io_bp_orig;
1609 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1611 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE)
1612 * then there is nothing to do here. Otherwise, free the
1613 * newly allocated block in this txg.
1615 if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
1616 ASSERT(BP_EQUAL(bp, bp_orig));
1618 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1619 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1620 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1621 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1625 dmu_tx_commit(dsa->dsa_tx);
1627 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1629 kmem_free(dsa, sizeof (*dsa));
1633 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1634 zio_prop_t *zp, zbookmark_phys_t *zb)
1636 dmu_sync_arg_t *dsa;
1639 tx = dmu_tx_create(os);
1640 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1641 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1643 /* Make zl_get_data do txg_waited_synced() */
1644 return (SET_ERROR(EIO));
1647 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1649 dsa->dsa_done = done;
1653 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx),
1654 zgd->zgd_bp, zgd->zgd_db->db_data, zgd->zgd_db->db_size,
1655 zp, dmu_sync_late_arrival_ready, NULL,
1656 NULL, dmu_sync_late_arrival_done, dsa, ZIO_PRIORITY_SYNC_WRITE,
1657 ZIO_FLAG_CANFAIL, zb));
1663 * Intent log support: sync the block associated with db to disk.
1664 * N.B. and XXX: the caller is responsible for making sure that the
1665 * data isn't changing while dmu_sync() is writing it.
1669 * EEXIST: this txg has already been synced, so there's nothing to do.
1670 * The caller should not log the write.
1672 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1673 * The caller should not log the write.
1675 * EALREADY: this block is already in the process of being synced.
1676 * The caller should track its progress (somehow).
1678 * EIO: could not do the I/O.
1679 * The caller should do a txg_wait_synced().
1681 * 0: the I/O has been initiated.
1682 * The caller should log this blkptr in the done callback.
1683 * It is possible that the I/O will fail, in which case
1684 * the error will be reported to the done callback and
1685 * propagated to pio from zio_done().
1688 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1690 blkptr_t *bp = zgd->zgd_bp;
1691 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1692 objset_t *os = db->db_objset;
1693 dsl_dataset_t *ds = os->os_dsl_dataset;
1694 dbuf_dirty_record_t *dr;
1695 dmu_sync_arg_t *dsa;
1696 zbookmark_phys_t zb;
1700 ASSERT(pio != NULL);
1703 SET_BOOKMARK(&zb, ds->ds_object,
1704 db->db.db_object, db->db_level, db->db_blkid);
1708 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1712 * If we're frozen (running ziltest), we always need to generate a bp.
1714 if (txg > spa_freeze_txg(os->os_spa))
1715 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1718 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1719 * and us. If we determine that this txg is not yet syncing,
1720 * but it begins to sync a moment later, that's OK because the
1721 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1723 mutex_enter(&db->db_mtx);
1725 if (txg <= spa_last_synced_txg(os->os_spa)) {
1727 * This txg has already synced. There's nothing to do.
1729 mutex_exit(&db->db_mtx);
1730 return (SET_ERROR(EEXIST));
1733 if (txg <= spa_syncing_txg(os->os_spa)) {
1735 * This txg is currently syncing, so we can't mess with
1736 * the dirty record anymore; just write a new log block.
1738 mutex_exit(&db->db_mtx);
1739 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1742 dr = db->db_last_dirty;
1743 while (dr && dr->dr_txg != txg)
1748 * There's no dr for this dbuf, so it must have been freed.
1749 * There's no need to log writes to freed blocks, so we're done.
1751 mutex_exit(&db->db_mtx);
1752 return (SET_ERROR(ENOENT));
1755 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
1758 * Assume the on-disk data is X, the current syncing data (in
1759 * txg - 1) is Y, and the current in-memory data is Z (currently
1762 * We usually want to perform a nopwrite if X and Z are the
1763 * same. However, if Y is different (i.e. the BP is going to
1764 * change before this write takes effect), then a nopwrite will
1765 * be incorrect - we would override with X, which could have
1766 * been freed when Y was written.
1768 * (Note that this is not a concern when we are nop-writing from
1769 * syncing context, because X and Y must be identical, because
1770 * all previous txgs have been synced.)
1772 * Therefore, we disable nopwrite if the current BP could change
1773 * before this TXG. There are two ways it could change: by
1774 * being dirty (dr_next is non-NULL), or by being freed
1775 * (dnode_block_freed()). This behavior is verified by
1776 * zio_done(), which VERIFYs that the override BP is identical
1777 * to the on-disk BP.
1781 if (dr->dr_next != NULL || dnode_block_freed(dn, db->db_blkid))
1782 zp.zp_nopwrite = B_FALSE;
1785 ASSERT(dr->dr_txg == txg);
1786 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1787 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1789 * We have already issued a sync write for this buffer,
1790 * or this buffer has already been synced. It could not
1791 * have been dirtied since, or we would have cleared the state.
1793 mutex_exit(&db->db_mtx);
1794 return (SET_ERROR(EALREADY));
1797 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1798 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1799 mutex_exit(&db->db_mtx);
1801 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1803 dsa->dsa_done = done;
1807 zio_nowait(arc_write(pio, os->os_spa, txg,
1808 bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
1809 &zp, dmu_sync_ready, NULL, NULL, dmu_sync_done, dsa,
1810 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb));
1816 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1822 err = dnode_hold(os, object, FTAG, &dn);
1825 err = dnode_set_blksz(dn, size, ibs, tx);
1826 dnode_rele(dn, FTAG);
1831 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1837 * Send streams include each object's checksum function. This
1838 * check ensures that the receiving system can understand the
1839 * checksum function transmitted.
1841 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS);
1843 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1844 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS);
1845 dn->dn_checksum = checksum;
1846 dnode_setdirty(dn, tx);
1847 dnode_rele(dn, FTAG);
1851 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1857 * Send streams include each object's compression function. This
1858 * check ensures that the receiving system can understand the
1859 * compression function transmitted.
1861 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS);
1863 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1864 dn->dn_compress = compress;
1865 dnode_setdirty(dn, tx);
1866 dnode_rele(dn, FTAG);
1869 int zfs_mdcomp_disable = 0;
1870 SYSCTL_INT(_vfs_zfs, OID_AUTO, mdcomp_disable, CTLFLAG_RWTUN,
1871 &zfs_mdcomp_disable, 0, "Disable metadata compression");
1874 * When the "redundant_metadata" property is set to "most", only indirect
1875 * blocks of this level and higher will have an additional ditto block.
1877 int zfs_redundant_metadata_most_ditto_level = 2;
1880 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
1882 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
1883 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
1885 enum zio_checksum checksum = os->os_checksum;
1886 enum zio_compress compress = os->os_compress;
1887 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
1888 boolean_t dedup = B_FALSE;
1889 boolean_t nopwrite = B_FALSE;
1890 boolean_t dedup_verify = os->os_dedup_verify;
1891 int copies = os->os_copies;
1894 * We maintain different write policies for each of the following
1897 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1898 * 3. all other level 0 blocks
1901 if (zfs_mdcomp_disable) {
1902 compress = ZIO_COMPRESS_EMPTY;
1905 * XXX -- we should design a compression algorithm
1906 * that specializes in arrays of bps.
1908 compress = zio_compress_select(os->os_spa,
1909 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON);
1913 * Metadata always gets checksummed. If the data
1914 * checksum is multi-bit correctable, and it's not a
1915 * ZBT-style checksum, then it's suitable for metadata
1916 * as well. Otherwise, the metadata checksum defaults
1919 if (!(zio_checksum_table[checksum].ci_flags &
1920 ZCHECKSUM_FLAG_METADATA) ||
1921 (zio_checksum_table[checksum].ci_flags &
1922 ZCHECKSUM_FLAG_EMBEDDED))
1923 checksum = ZIO_CHECKSUM_FLETCHER_4;
1925 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL ||
1926 (os->os_redundant_metadata ==
1927 ZFS_REDUNDANT_METADATA_MOST &&
1928 (level >= zfs_redundant_metadata_most_ditto_level ||
1929 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))))
1931 } else if (wp & WP_NOFILL) {
1935 * If we're writing preallocated blocks, we aren't actually
1936 * writing them so don't set any policy properties. These
1937 * blocks are currently only used by an external subsystem
1938 * outside of zfs (i.e. dump) and not written by the zio
1941 compress = ZIO_COMPRESS_OFF;
1942 checksum = ZIO_CHECKSUM_NOPARITY;
1944 compress = zio_compress_select(os->os_spa, dn->dn_compress,
1947 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
1948 zio_checksum_select(dn->dn_checksum, checksum) :
1952 * Determine dedup setting. If we are in dmu_sync(),
1953 * we won't actually dedup now because that's all
1954 * done in syncing context; but we do want to use the
1955 * dedup checkum. If the checksum is not strong
1956 * enough to ensure unique signatures, force
1959 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
1960 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
1961 if (!(zio_checksum_table[checksum].ci_flags &
1962 ZCHECKSUM_FLAG_DEDUP))
1963 dedup_verify = B_TRUE;
1967 * Enable nopwrite if we have secure enough checksum
1968 * algorithm (see comment in zio_nop_write) and
1969 * compression is enabled. We don't enable nopwrite if
1970 * dedup is enabled as the two features are mutually
1973 nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags &
1974 ZCHECKSUM_FLAG_NOPWRITE) &&
1975 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
1978 zp->zp_checksum = checksum;
1979 zp->zp_compress = compress;
1980 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
1981 zp->zp_level = level;
1982 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
1983 zp->zp_dedup = dedup;
1984 zp->zp_dedup_verify = dedup && dedup_verify;
1985 zp->zp_nopwrite = nopwrite;
1989 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
1995 * Sync any current changes before
1996 * we go trundling through the block pointers.
1998 err = dmu_object_wait_synced(os, object);
2003 err = dnode_hold(os, object, FTAG, &dn);
2008 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
2009 dnode_rele(dn, FTAG);
2015 * Given the ZFS object, if it contains any dirty nodes
2016 * this function flushes all dirty blocks to disk. This
2017 * ensures the DMU object info is updated. A more efficient
2018 * future version might just find the TXG with the maximum
2019 * ID and wait for that to be synced.
2022 dmu_object_wait_synced(objset_t *os, uint64_t object)
2027 error = dnode_hold(os, object, FTAG, &dn);
2032 for (i = 0; i < TXG_SIZE; i++) {
2033 if (list_link_active(&dn->dn_dirty_link[i])) {
2037 dnode_rele(dn, FTAG);
2038 if (i != TXG_SIZE) {
2039 txg_wait_synced(dmu_objset_pool(os), 0);
2046 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2050 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2051 mutex_enter(&dn->dn_mtx);
2055 doi->doi_data_block_size = dn->dn_datablksz;
2056 doi->doi_metadata_block_size = dn->dn_indblkshift ?
2057 1ULL << dn->dn_indblkshift : 0;
2058 doi->doi_type = dn->dn_type;
2059 doi->doi_bonus_type = dn->dn_bonustype;
2060 doi->doi_bonus_size = dn->dn_bonuslen;
2061 doi->doi_indirection = dn->dn_nlevels;
2062 doi->doi_checksum = dn->dn_checksum;
2063 doi->doi_compress = dn->dn_compress;
2064 doi->doi_nblkptr = dn->dn_nblkptr;
2065 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
2066 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
2067 doi->doi_fill_count = 0;
2068 for (int i = 0; i < dnp->dn_nblkptr; i++)
2069 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
2071 mutex_exit(&dn->dn_mtx);
2072 rw_exit(&dn->dn_struct_rwlock);
2076 * Get information on a DMU object.
2077 * If doi is NULL, just indicates whether the object exists.
2080 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
2083 int err = dnode_hold(os, object, FTAG, &dn);
2089 dmu_object_info_from_dnode(dn, doi);
2091 dnode_rele(dn, FTAG);
2096 * As above, but faster; can be used when you have a held dbuf in hand.
2099 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
2101 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2104 dmu_object_info_from_dnode(DB_DNODE(db), doi);
2109 * Faster still when you only care about the size.
2110 * This is specifically optimized for zfs_getattr().
2113 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
2114 u_longlong_t *nblk512)
2116 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2122 *blksize = dn->dn_datablksz;
2123 /* add 1 for dnode space */
2124 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
2125 SPA_MINBLOCKSHIFT) + 1;
2130 byteswap_uint64_array(void *vbuf, size_t size)
2132 uint64_t *buf = vbuf;
2133 size_t count = size >> 3;
2136 ASSERT((size & 7) == 0);
2138 for (i = 0; i < count; i++)
2139 buf[i] = BSWAP_64(buf[i]);
2143 byteswap_uint32_array(void *vbuf, size_t size)
2145 uint32_t *buf = vbuf;
2146 size_t count = size >> 2;
2149 ASSERT((size & 3) == 0);
2151 for (i = 0; i < count; i++)
2152 buf[i] = BSWAP_32(buf[i]);
2156 byteswap_uint16_array(void *vbuf, size_t size)
2158 uint16_t *buf = vbuf;
2159 size_t count = size >> 1;
2162 ASSERT((size & 1) == 0);
2164 for (i = 0; i < count; i++)
2165 buf[i] = BSWAP_16(buf[i]);
2170 byteswap_uint8_array(void *vbuf, size_t size)
2183 zio_compress_init();
2192 arc_fini(); /* arc depends on l2arc, so arc must go first */
2195 zio_compress_fini();