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, 2014 by Delphix. All rights reserved.
25 /* Copyright (c) 2013 by Saso Kiselkov. All rights reserved. */
26 /* Copyright (c) 2013, Joyent, Inc. All rights reserved. */
27 /* Copyright (c) 2014, Nexenta Systems, Inc. All rights reserved. */
30 #include <sys/dmu_impl.h>
31 #include <sys/dmu_tx.h>
33 #include <sys/dnode.h>
34 #include <sys/zfs_context.h>
35 #include <sys/dmu_objset.h>
36 #include <sys/dmu_traverse.h>
37 #include <sys/dsl_dataset.h>
38 #include <sys/dsl_dir.h>
39 #include <sys/dsl_pool.h>
40 #include <sys/dsl_synctask.h>
41 #include <sys/dsl_prop.h>
42 #include <sys/dmu_zfetch.h>
43 #include <sys/zfs_ioctl.h>
45 #include <sys/zio_checksum.h>
46 #include <sys/zio_compress.h>
48 #include <sys/zfeature.h>
51 #include <sys/zfs_znode.h>
55 * Enable/disable nopwrite feature.
57 int zfs_nopwrite_enabled = 1;
58 SYSCTL_DECL(_vfs_zfs);
59 TUNABLE_INT("vfs.zfs.nopwrite_enabled", &zfs_nopwrite_enabled);
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(objset_t *os, uint64_t object, uint64_t offset,
135 void *tag, dmu_buf_t **dbp)
142 err = dnode_hold(os, object, FTAG, &dn);
145 blkid = dbuf_whichblock(dn, 0, offset);
146 rw_enter(&dn->dn_struct_rwlock, RW_READER);
147 db = dbuf_hold(dn, blkid, tag);
148 rw_exit(&dn->dn_struct_rwlock);
149 dnode_rele(dn, FTAG);
153 return (SET_ERROR(EIO));
161 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
162 void *tag, dmu_buf_t **dbp, int flags)
165 int db_flags = DB_RF_CANFAIL;
167 if (flags & DMU_READ_NO_PREFETCH)
168 db_flags |= DB_RF_NOPREFETCH;
170 err = dmu_buf_hold_noread(os, object, offset, tag, dbp);
172 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
173 err = dbuf_read(db, NULL, db_flags);
186 return (DN_MAX_BONUSLEN);
190 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
192 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
199 if (dn->dn_bonus != db) {
200 error = SET_ERROR(EINVAL);
201 } else if (newsize < 0 || newsize > db_fake->db_size) {
202 error = SET_ERROR(EINVAL);
204 dnode_setbonuslen(dn, newsize, tx);
213 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
215 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
222 if (!DMU_OT_IS_VALID(type)) {
223 error = SET_ERROR(EINVAL);
224 } else if (dn->dn_bonus != db) {
225 error = SET_ERROR(EINVAL);
227 dnode_setbonus_type(dn, type, tx);
236 dmu_get_bonustype(dmu_buf_t *db_fake)
238 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
240 dmu_object_type_t type;
244 type = dn->dn_bonustype;
251 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
256 error = dnode_hold(os, object, FTAG, &dn);
257 dbuf_rm_spill(dn, tx);
258 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
259 dnode_rm_spill(dn, tx);
260 rw_exit(&dn->dn_struct_rwlock);
261 dnode_rele(dn, FTAG);
266 * returns ENOENT, EIO, or 0.
269 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
275 error = dnode_hold(os, object, FTAG, &dn);
279 rw_enter(&dn->dn_struct_rwlock, RW_READER);
280 if (dn->dn_bonus == NULL) {
281 rw_exit(&dn->dn_struct_rwlock);
282 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
283 if (dn->dn_bonus == NULL)
284 dbuf_create_bonus(dn);
288 /* as long as the bonus buf is held, the dnode will be held */
289 if (refcount_add(&db->db_holds, tag) == 1) {
290 VERIFY(dnode_add_ref(dn, db));
291 atomic_inc_32(&dn->dn_dbufs_count);
295 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
296 * hold and incrementing the dbuf count to ensure that dnode_move() sees
297 * a dnode hold for every dbuf.
299 rw_exit(&dn->dn_struct_rwlock);
301 dnode_rele(dn, FTAG);
303 VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH));
310 * returns ENOENT, EIO, or 0.
312 * This interface will allocate a blank spill dbuf when a spill blk
313 * doesn't already exist on the dnode.
315 * if you only want to find an already existing spill db, then
316 * dmu_spill_hold_existing() should be used.
319 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
321 dmu_buf_impl_t *db = NULL;
324 if ((flags & DB_RF_HAVESTRUCT) == 0)
325 rw_enter(&dn->dn_struct_rwlock, RW_READER);
327 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
329 if ((flags & DB_RF_HAVESTRUCT) == 0)
330 rw_exit(&dn->dn_struct_rwlock);
333 err = dbuf_read(db, NULL, flags);
342 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
344 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
351 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
352 err = SET_ERROR(EINVAL);
354 rw_enter(&dn->dn_struct_rwlock, RW_READER);
356 if (!dn->dn_have_spill) {
357 err = SET_ERROR(ENOENT);
359 err = dmu_spill_hold_by_dnode(dn,
360 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
363 rw_exit(&dn->dn_struct_rwlock);
371 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
373 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
379 err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp);
386 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
387 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
388 * and can induce severe lock contention when writing to several files
389 * whose dnodes are in the same block.
392 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
393 int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
396 uint64_t blkid, nblks, i;
401 ASSERT(length <= DMU_MAX_ACCESS);
403 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT;
404 if (flags & DMU_READ_NO_PREFETCH || length > zfetch_array_rd_sz)
405 dbuf_flags |= DB_RF_NOPREFETCH;
407 rw_enter(&dn->dn_struct_rwlock, RW_READER);
408 if (dn->dn_datablkshift) {
409 int blkshift = dn->dn_datablkshift;
410 nblks = (P2ROUNDUP(offset+length, 1ULL<<blkshift) -
411 P2ALIGN(offset, 1ULL<<blkshift)) >> blkshift;
413 if (offset + length > dn->dn_datablksz) {
414 zfs_panic_recover("zfs: accessing past end of object "
415 "%llx/%llx (size=%u access=%llu+%llu)",
416 (longlong_t)dn->dn_objset->
417 os_dsl_dataset->ds_object,
418 (longlong_t)dn->dn_object, dn->dn_datablksz,
419 (longlong_t)offset, (longlong_t)length);
420 rw_exit(&dn->dn_struct_rwlock);
421 return (SET_ERROR(EIO));
425 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
427 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
428 blkid = dbuf_whichblock(dn, 0, offset);
429 for (i = 0; i < nblks; i++) {
430 dmu_buf_impl_t *db = dbuf_hold(dn, blkid+i, tag);
432 rw_exit(&dn->dn_struct_rwlock);
433 dmu_buf_rele_array(dbp, nblks, tag);
435 return (SET_ERROR(EIO));
437 /* initiate async i/o */
439 (void) dbuf_read(db, zio, dbuf_flags);
442 curthread->td_ru.ru_oublock++;
446 rw_exit(&dn->dn_struct_rwlock);
448 /* wait for async i/o */
451 dmu_buf_rele_array(dbp, nblks, tag);
455 /* wait for other io to complete */
457 for (i = 0; i < nblks; i++) {
458 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
459 mutex_enter(&db->db_mtx);
460 while (db->db_state == DB_READ ||
461 db->db_state == DB_FILL)
462 cv_wait(&db->db_changed, &db->db_mtx);
463 if (db->db_state == DB_UNCACHED)
464 err = SET_ERROR(EIO);
465 mutex_exit(&db->db_mtx);
467 dmu_buf_rele_array(dbp, nblks, tag);
479 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
480 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
485 err = dnode_hold(os, object, FTAG, &dn);
489 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
490 numbufsp, dbpp, DMU_READ_PREFETCH);
492 dnode_rele(dn, FTAG);
498 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
499 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
501 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
507 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
508 numbufsp, dbpp, DMU_READ_PREFETCH);
515 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
518 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
523 for (i = 0; i < numbufs; i++) {
525 dbuf_rele(dbp[i], tag);
528 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
532 * Issue prefetch i/os for the given blocks. If level is greater than 0, the
533 * indirect blocks prefeteched will be those that point to the blocks containing
534 * the data starting at offset, and continuing to offset + len.
536 * Note that if the indirect blocks above the blocks being prefetched are not in
537 * cache, they will be asychronously read in.
540 dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset,
541 uint64_t len, zio_priority_t pri)
547 if (zfs_prefetch_disable)
550 if (len == 0) { /* they're interested in the bonus buffer */
551 dn = DMU_META_DNODE(os);
553 if (object == 0 || object >= DN_MAX_OBJECT)
556 rw_enter(&dn->dn_struct_rwlock, RW_READER);
557 blkid = dbuf_whichblock(dn, level,
558 object * sizeof (dnode_phys_t));
559 dbuf_prefetch(dn, level, blkid, pri, 0);
560 rw_exit(&dn->dn_struct_rwlock);
565 * XXX - Note, if the dnode for the requested object is not
566 * already cached, we will do a *synchronous* read in the
567 * dnode_hold() call. The same is true for any indirects.
569 err = dnode_hold(os, object, FTAG, &dn);
573 rw_enter(&dn->dn_struct_rwlock, RW_READER);
575 * offset + len - 1 is the last byte we want to prefetch for, and offset
576 * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the
577 * last block we want to prefetch, and dbuf_whichblock(dn, level,
578 * offset) is the first. Then the number we need to prefetch is the
581 if (level > 0 || dn->dn_datablkshift != 0) {
582 nblks = dbuf_whichblock(dn, level, offset + len - 1) -
583 dbuf_whichblock(dn, level, offset) + 1;
585 nblks = (offset < dn->dn_datablksz);
589 blkid = dbuf_whichblock(dn, level, offset);
590 for (int i = 0; i < nblks; i++)
591 dbuf_prefetch(dn, level, blkid + i, pri, 0);
594 rw_exit(&dn->dn_struct_rwlock);
596 dnode_rele(dn, FTAG);
600 * Get the next "chunk" of file data to free. We traverse the file from
601 * the end so that the file gets shorter over time (if we crashes in the
602 * middle, this will leave us in a better state). We find allocated file
603 * data by simply searching the allocated level 1 indirects.
605 * On input, *start should be the first offset that does not need to be
606 * freed (e.g. "offset + length"). On return, *start will be the first
607 * offset that should be freed.
610 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum)
612 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
613 /* bytes of data covered by a level-1 indirect block */
615 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
617 ASSERT3U(minimum, <=, *start);
619 if (*start - minimum <= iblkrange * maxblks) {
623 ASSERT(ISP2(iblkrange));
625 for (uint64_t blks = 0; *start > minimum && blks < maxblks; blks++) {
629 * dnode_next_offset(BACKWARDS) will find an allocated L1
630 * indirect block at or before the input offset. We must
631 * decrement *start so that it is at the end of the region
635 err = dnode_next_offset(dn,
636 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
638 /* if there are no indirect blocks before start, we are done */
642 } else if (err != 0) {
646 /* set start to the beginning of this L1 indirect */
647 *start = P2ALIGN(*start, iblkrange);
649 if (*start < minimum)
655 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
658 uint64_t object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
661 if (offset >= object_size)
664 if (length == DMU_OBJECT_END || offset + length > object_size)
665 length = object_size - offset;
667 while (length != 0) {
668 uint64_t chunk_end, chunk_begin;
670 chunk_end = chunk_begin = offset + length;
672 /* move chunk_begin backwards to the beginning of this chunk */
673 err = get_next_chunk(dn, &chunk_begin, offset);
676 ASSERT3U(chunk_begin, >=, offset);
677 ASSERT3U(chunk_begin, <=, chunk_end);
679 dmu_tx_t *tx = dmu_tx_create(os);
680 dmu_tx_hold_free(tx, dn->dn_object,
681 chunk_begin, chunk_end - chunk_begin);
684 * Mark this transaction as typically resulting in a net
685 * reduction in space used.
687 dmu_tx_mark_netfree(tx);
688 err = dmu_tx_assign(tx, TXG_WAIT);
693 dnode_free_range(dn, chunk_begin, chunk_end - chunk_begin, tx);
696 length -= chunk_end - chunk_begin;
702 dmu_free_long_range(objset_t *os, uint64_t object,
703 uint64_t offset, uint64_t length)
708 err = dnode_hold(os, object, FTAG, &dn);
711 err = dmu_free_long_range_impl(os, dn, offset, length);
714 * It is important to zero out the maxblkid when freeing the entire
715 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
716 * will take the fast path, and (b) dnode_reallocate() can verify
717 * that the entire file has been freed.
719 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
722 dnode_rele(dn, FTAG);
727 dmu_free_long_object(objset_t *os, uint64_t object)
732 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
736 tx = dmu_tx_create(os);
737 dmu_tx_hold_bonus(tx, object);
738 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
739 dmu_tx_mark_netfree(tx);
740 err = dmu_tx_assign(tx, TXG_WAIT);
742 err = dmu_object_free(os, object, tx);
752 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
753 uint64_t size, dmu_tx_t *tx)
756 int err = dnode_hold(os, object, FTAG, &dn);
759 ASSERT(offset < UINT64_MAX);
760 ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
761 dnode_free_range(dn, offset, size, tx);
762 dnode_rele(dn, FTAG);
767 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
768 void *buf, uint32_t flags)
774 err = dnode_hold(os, object, FTAG, &dn);
779 * Deal with odd block sizes, where there can't be data past the first
780 * block. If we ever do the tail block optimization, we will need to
781 * handle that here as well.
783 if (dn->dn_maxblkid == 0) {
784 int newsz = offset > dn->dn_datablksz ? 0 :
785 MIN(size, dn->dn_datablksz - offset);
786 bzero((char *)buf + newsz, size - newsz);
791 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
795 * NB: we could do this block-at-a-time, but it's nice
796 * to be reading in parallel.
798 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
799 TRUE, FTAG, &numbufs, &dbp, flags);
803 for (i = 0; i < numbufs; i++) {
806 dmu_buf_t *db = dbp[i];
810 bufoff = offset - db->db_offset;
811 tocpy = (int)MIN(db->db_size - bufoff, size);
813 bcopy((char *)db->db_data + bufoff, buf, tocpy);
817 buf = (char *)buf + tocpy;
819 dmu_buf_rele_array(dbp, numbufs, FTAG);
821 dnode_rele(dn, FTAG);
826 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
827 const void *buf, dmu_tx_t *tx)
835 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
836 FALSE, FTAG, &numbufs, &dbp));
838 for (i = 0; i < numbufs; i++) {
841 dmu_buf_t *db = dbp[i];
845 bufoff = offset - db->db_offset;
846 tocpy = (int)MIN(db->db_size - bufoff, size);
848 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
850 if (tocpy == db->db_size)
851 dmu_buf_will_fill(db, tx);
853 dmu_buf_will_dirty(db, tx);
855 bcopy(buf, (char *)db->db_data + bufoff, tocpy);
857 if (tocpy == db->db_size)
858 dmu_buf_fill_done(db, tx);
862 buf = (char *)buf + tocpy;
864 dmu_buf_rele_array(dbp, numbufs, FTAG);
868 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
877 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
878 FALSE, FTAG, &numbufs, &dbp));
880 for (i = 0; i < numbufs; i++) {
881 dmu_buf_t *db = dbp[i];
883 dmu_buf_will_not_fill(db, tx);
885 dmu_buf_rele_array(dbp, numbufs, FTAG);
889 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
890 void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
891 int compressed_size, int byteorder, dmu_tx_t *tx)
895 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
896 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
897 VERIFY0(dmu_buf_hold_noread(os, object, offset,
900 dmu_buf_write_embedded(db,
901 data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
902 uncompressed_size, compressed_size, byteorder, tx);
904 dmu_buf_rele(db, FTAG);
908 * DMU support for xuio
910 kstat_t *xuio_ksp = NULL;
913 dmu_xuio_init(xuio_t *xuio, int nblk)
916 uio_t *uio = &xuio->xu_uio;
918 uio->uio_iovcnt = nblk;
919 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
921 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
923 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
924 priv->iovp = uio->uio_iov;
925 XUIO_XUZC_PRIV(xuio) = priv;
927 if (XUIO_XUZC_RW(xuio) == UIO_READ)
928 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
930 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
936 dmu_xuio_fini(xuio_t *xuio)
938 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
939 int nblk = priv->cnt;
941 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
942 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
943 kmem_free(priv, sizeof (dmu_xuio_t));
945 if (XUIO_XUZC_RW(xuio) == UIO_READ)
946 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
948 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
952 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
953 * and increase priv->next by 1.
956 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
959 uio_t *uio = &xuio->xu_uio;
960 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
961 int i = priv->next++;
963 ASSERT(i < priv->cnt);
964 ASSERT(off + n <= arc_buf_size(abuf));
965 iov = uio->uio_iov + i;
966 iov->iov_base = (char *)abuf->b_data + off;
968 priv->bufs[i] = abuf;
973 dmu_xuio_cnt(xuio_t *xuio)
975 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
980 dmu_xuio_arcbuf(xuio_t *xuio, int i)
982 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
984 ASSERT(i < priv->cnt);
985 return (priv->bufs[i]);
989 dmu_xuio_clear(xuio_t *xuio, int i)
991 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
993 ASSERT(i < priv->cnt);
994 priv->bufs[i] = NULL;
1000 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
1001 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
1002 KSTAT_FLAG_VIRTUAL);
1003 if (xuio_ksp != NULL) {
1004 xuio_ksp->ks_data = &xuio_stats;
1005 kstat_install(xuio_ksp);
1010 xuio_stat_fini(void)
1012 if (xuio_ksp != NULL) {
1013 kstat_delete(xuio_ksp);
1019 xuio_stat_wbuf_copied()
1021 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1025 xuio_stat_wbuf_nocopy()
1027 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
1032 dmu_read_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size)
1035 int numbufs, i, err;
1036 xuio_t *xuio = NULL;
1039 * NB: we could do this block-at-a-time, but it's nice
1040 * to be reading in parallel.
1042 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1043 TRUE, FTAG, &numbufs, &dbp, 0);
1048 if (uio->uio_extflg == UIO_XUIO)
1049 xuio = (xuio_t *)uio;
1052 for (i = 0; i < numbufs; i++) {
1055 dmu_buf_t *db = dbp[i];
1059 bufoff = uio->uio_loffset - db->db_offset;
1060 tocpy = (int)MIN(db->db_size - bufoff, size);
1063 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
1064 arc_buf_t *dbuf_abuf = dbi->db_buf;
1065 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
1066 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
1068 uio->uio_resid -= tocpy;
1069 uio->uio_loffset += tocpy;
1072 if (abuf == dbuf_abuf)
1073 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
1075 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
1077 err = uiomove((char *)db->db_data + bufoff, tocpy,
1085 dmu_buf_rele_array(dbp, numbufs, FTAG);
1091 * Read 'size' bytes into the uio buffer.
1092 * From object zdb->db_object.
1093 * Starting at offset uio->uio_loffset.
1095 * If the caller already has a dbuf in the target object
1096 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1097 * because we don't have to find the dnode_t for the object.
1100 dmu_read_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size)
1102 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1111 err = dmu_read_uio_dnode(dn, uio, size);
1118 * Read 'size' bytes into the uio buffer.
1119 * From the specified object
1120 * Starting at offset uio->uio_loffset.
1123 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
1131 err = dnode_hold(os, object, FTAG, &dn);
1135 err = dmu_read_uio_dnode(dn, uio, size);
1137 dnode_rele(dn, FTAG);
1143 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1150 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1151 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1155 for (i = 0; i < numbufs; i++) {
1158 dmu_buf_t *db = dbp[i];
1162 bufoff = uio->uio_loffset - db->db_offset;
1163 tocpy = (int)MIN(db->db_size - bufoff, size);
1165 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1167 if (tocpy == db->db_size)
1168 dmu_buf_will_fill(db, tx);
1170 dmu_buf_will_dirty(db, tx);
1173 * XXX uiomove could block forever (eg. nfs-backed
1174 * pages). There needs to be a uiolockdown() function
1175 * to lock the pages in memory, so that uiomove won't
1178 err = uiomove((char *)db->db_data + bufoff, tocpy,
1181 if (tocpy == db->db_size)
1182 dmu_buf_fill_done(db, tx);
1190 dmu_buf_rele_array(dbp, numbufs, FTAG);
1195 * Write 'size' bytes from the uio buffer.
1196 * To object zdb->db_object.
1197 * Starting at offset uio->uio_loffset.
1199 * If the caller already has a dbuf in the target object
1200 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1201 * because we don't have to find the dnode_t for the object.
1204 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1207 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1216 err = dmu_write_uio_dnode(dn, uio, size, tx);
1223 * Write 'size' bytes from the uio buffer.
1224 * To the specified object.
1225 * Starting at offset uio->uio_loffset.
1228 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1237 err = dnode_hold(os, object, FTAG, &dn);
1241 err = dmu_write_uio_dnode(dn, uio, size, tx);
1243 dnode_rele(dn, FTAG);
1250 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1251 page_t *pp, dmu_tx_t *tx)
1260 err = dmu_buf_hold_array(os, object, offset, size,
1261 FALSE, FTAG, &numbufs, &dbp);
1265 for (i = 0; i < numbufs; i++) {
1266 int tocpy, copied, thiscpy;
1268 dmu_buf_t *db = dbp[i];
1272 ASSERT3U(db->db_size, >=, PAGESIZE);
1274 bufoff = offset - db->db_offset;
1275 tocpy = (int)MIN(db->db_size - bufoff, size);
1277 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1279 if (tocpy == db->db_size)
1280 dmu_buf_will_fill(db, tx);
1282 dmu_buf_will_dirty(db, tx);
1284 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1285 ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff);
1286 thiscpy = MIN(PAGESIZE, tocpy - copied);
1287 va = zfs_map_page(pp, S_READ);
1288 bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1289 zfs_unmap_page(pp, va);
1294 if (tocpy == db->db_size)
1295 dmu_buf_fill_done(db, tx);
1300 dmu_buf_rele_array(dbp, numbufs, FTAG);
1307 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1308 vm_page_t *ma, dmu_tx_t *tx)
1318 err = dmu_buf_hold_array(os, object, offset, size,
1319 FALSE, FTAG, &numbufs, &dbp);
1323 for (i = 0; i < numbufs; i++) {
1324 int tocpy, copied, thiscpy;
1326 dmu_buf_t *db = dbp[i];
1330 ASSERT3U(db->db_size, >=, PAGESIZE);
1332 bufoff = offset - db->db_offset;
1333 tocpy = (int)MIN(db->db_size - bufoff, size);
1335 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1337 if (tocpy == db->db_size)
1338 dmu_buf_will_fill(db, tx);
1340 dmu_buf_will_dirty(db, tx);
1342 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1343 ASSERT3U(ptoa((*ma)->pindex), ==, db->db_offset + bufoff);
1344 thiscpy = MIN(PAGESIZE, tocpy - copied);
1345 va = zfs_map_page(*ma, &sf);
1346 bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1352 if (tocpy == db->db_size)
1353 dmu_buf_fill_done(db, tx);
1358 dmu_buf_rele_array(dbp, numbufs, FTAG);
1365 * Allocate a loaned anonymous arc buffer.
1368 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1370 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1372 return (arc_loan_buf(db->db_objset->os_spa, size));
1376 * Free a loaned arc buffer.
1379 dmu_return_arcbuf(arc_buf_t *buf)
1381 arc_return_buf(buf, FTAG);
1382 VERIFY(arc_buf_remove_ref(buf, FTAG));
1386 * When possible directly assign passed loaned arc buffer to a dbuf.
1387 * If this is not possible copy the contents of passed arc buf via
1391 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1394 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1397 uint32_t blksz = (uint32_t)arc_buf_size(buf);
1400 DB_DNODE_ENTER(dbuf);
1401 dn = DB_DNODE(dbuf);
1402 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1403 blkid = dbuf_whichblock(dn, 0, offset);
1404 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1405 rw_exit(&dn->dn_struct_rwlock);
1406 DB_DNODE_EXIT(dbuf);
1409 * We can only assign if the offset is aligned, the arc buf is the
1410 * same size as the dbuf, and the dbuf is not metadata. It
1411 * can't be metadata because the loaned arc buf comes from the
1412 * user-data kmem arena.
1414 if (offset == db->db.db_offset && blksz == db->db.db_size &&
1415 DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA) {
1416 dbuf_assign_arcbuf(db, buf, tx);
1417 dbuf_rele(db, FTAG);
1422 DB_DNODE_ENTER(dbuf);
1423 dn = DB_DNODE(dbuf);
1425 object = dn->dn_object;
1426 DB_DNODE_EXIT(dbuf);
1428 dbuf_rele(db, FTAG);
1429 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1430 dmu_return_arcbuf(buf);
1431 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1436 dbuf_dirty_record_t *dsa_dr;
1437 dmu_sync_cb_t *dsa_done;
1444 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1446 dmu_sync_arg_t *dsa = varg;
1447 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1448 blkptr_t *bp = zio->io_bp;
1450 if (zio->io_error == 0) {
1451 if (BP_IS_HOLE(bp)) {
1453 * A block of zeros may compress to a hole, but the
1454 * block size still needs to be known for replay.
1456 BP_SET_LSIZE(bp, db->db_size);
1457 } else if (!BP_IS_EMBEDDED(bp)) {
1458 ASSERT(BP_GET_LEVEL(bp) == 0);
1465 dmu_sync_late_arrival_ready(zio_t *zio)
1467 dmu_sync_ready(zio, NULL, zio->io_private);
1472 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1474 dmu_sync_arg_t *dsa = varg;
1475 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1476 dmu_buf_impl_t *db = dr->dr_dbuf;
1478 mutex_enter(&db->db_mtx);
1479 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1480 if (zio->io_error == 0) {
1481 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1482 if (dr->dt.dl.dr_nopwrite) {
1483 blkptr_t *bp = zio->io_bp;
1484 blkptr_t *bp_orig = &zio->io_bp_orig;
1485 uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1487 ASSERT(BP_EQUAL(bp, bp_orig));
1488 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1489 ASSERT(zio_checksum_table[chksum].ci_dedup);
1491 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1492 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1493 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1496 * Old style holes are filled with all zeros, whereas
1497 * new-style holes maintain their lsize, type, level,
1498 * and birth time (see zio_write_compress). While we
1499 * need to reset the BP_SET_LSIZE() call that happened
1500 * in dmu_sync_ready for old style holes, we do *not*
1501 * want to wipe out the information contained in new
1502 * style holes. Thus, only zero out the block pointer if
1503 * it's an old style hole.
1505 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) &&
1506 dr->dt.dl.dr_overridden_by.blk_birth == 0)
1507 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1509 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1511 cv_broadcast(&db->db_changed);
1512 mutex_exit(&db->db_mtx);
1514 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1516 kmem_free(dsa, sizeof (*dsa));
1520 dmu_sync_late_arrival_done(zio_t *zio)
1522 blkptr_t *bp = zio->io_bp;
1523 dmu_sync_arg_t *dsa = zio->io_private;
1524 blkptr_t *bp_orig = &zio->io_bp_orig;
1526 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1528 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE)
1529 * then there is nothing to do here. Otherwise, free the
1530 * newly allocated block in this txg.
1532 if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
1533 ASSERT(BP_EQUAL(bp, bp_orig));
1535 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1536 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1537 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1538 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1542 dmu_tx_commit(dsa->dsa_tx);
1544 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1546 kmem_free(dsa, sizeof (*dsa));
1550 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1551 zio_prop_t *zp, zbookmark_phys_t *zb)
1553 dmu_sync_arg_t *dsa;
1556 tx = dmu_tx_create(os);
1557 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1558 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1560 /* Make zl_get_data do txg_waited_synced() */
1561 return (SET_ERROR(EIO));
1564 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1566 dsa->dsa_done = done;
1570 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1571 zgd->zgd_db->db_data, zgd->zgd_db->db_size, zp,
1572 dmu_sync_late_arrival_ready, NULL, dmu_sync_late_arrival_done, dsa,
1573 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1579 * Intent log support: sync the block associated with db to disk.
1580 * N.B. and XXX: the caller is responsible for making sure that the
1581 * data isn't changing while dmu_sync() is writing it.
1585 * EEXIST: this txg has already been synced, so there's nothing to do.
1586 * The caller should not log the write.
1588 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1589 * The caller should not log the write.
1591 * EALREADY: this block is already in the process of being synced.
1592 * The caller should track its progress (somehow).
1594 * EIO: could not do the I/O.
1595 * The caller should do a txg_wait_synced().
1597 * 0: the I/O has been initiated.
1598 * The caller should log this blkptr in the done callback.
1599 * It is possible that the I/O will fail, in which case
1600 * the error will be reported to the done callback and
1601 * propagated to pio from zio_done().
1604 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1606 blkptr_t *bp = zgd->zgd_bp;
1607 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1608 objset_t *os = db->db_objset;
1609 dsl_dataset_t *ds = os->os_dsl_dataset;
1610 dbuf_dirty_record_t *dr;
1611 dmu_sync_arg_t *dsa;
1612 zbookmark_phys_t zb;
1616 ASSERT(pio != NULL);
1619 SET_BOOKMARK(&zb, ds->ds_object,
1620 db->db.db_object, db->db_level, db->db_blkid);
1624 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1628 * If we're frozen (running ziltest), we always need to generate a bp.
1630 if (txg > spa_freeze_txg(os->os_spa))
1631 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1634 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1635 * and us. If we determine that this txg is not yet syncing,
1636 * but it begins to sync a moment later, that's OK because the
1637 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1639 mutex_enter(&db->db_mtx);
1641 if (txg <= spa_last_synced_txg(os->os_spa)) {
1643 * This txg has already synced. There's nothing to do.
1645 mutex_exit(&db->db_mtx);
1646 return (SET_ERROR(EEXIST));
1649 if (txg <= spa_syncing_txg(os->os_spa)) {
1651 * This txg is currently syncing, so we can't mess with
1652 * the dirty record anymore; just write a new log block.
1654 mutex_exit(&db->db_mtx);
1655 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1658 dr = db->db_last_dirty;
1659 while (dr && dr->dr_txg != txg)
1664 * There's no dr for this dbuf, so it must have been freed.
1665 * There's no need to log writes to freed blocks, so we're done.
1667 mutex_exit(&db->db_mtx);
1668 return (SET_ERROR(ENOENT));
1671 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
1674 * Assume the on-disk data is X, the current syncing data (in
1675 * txg - 1) is Y, and the current in-memory data is Z (currently
1678 * We usually want to perform a nopwrite if X and Z are the
1679 * same. However, if Y is different (i.e. the BP is going to
1680 * change before this write takes effect), then a nopwrite will
1681 * be incorrect - we would override with X, which could have
1682 * been freed when Y was written.
1684 * (Note that this is not a concern when we are nop-writing from
1685 * syncing context, because X and Y must be identical, because
1686 * all previous txgs have been synced.)
1688 * Therefore, we disable nopwrite if the current BP could change
1689 * before this TXG. There are two ways it could change: by
1690 * being dirty (dr_next is non-NULL), or by being freed
1691 * (dnode_block_freed()). This behavior is verified by
1692 * zio_done(), which VERIFYs that the override BP is identical
1693 * to the on-disk BP.
1697 if (dr->dr_next != NULL || dnode_block_freed(dn, db->db_blkid))
1698 zp.zp_nopwrite = B_FALSE;
1701 ASSERT(dr->dr_txg == txg);
1702 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1703 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1705 * We have already issued a sync write for this buffer,
1706 * or this buffer has already been synced. It could not
1707 * have been dirtied since, or we would have cleared the state.
1709 mutex_exit(&db->db_mtx);
1710 return (SET_ERROR(EALREADY));
1713 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1714 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1715 mutex_exit(&db->db_mtx);
1717 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1719 dsa->dsa_done = done;
1723 zio_nowait(arc_write(pio, os->os_spa, txg,
1724 bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
1725 DBUF_IS_L2COMPRESSIBLE(db), &zp, dmu_sync_ready,
1726 NULL, dmu_sync_done, dsa, ZIO_PRIORITY_SYNC_WRITE,
1727 ZIO_FLAG_CANFAIL, &zb));
1733 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1739 err = dnode_hold(os, object, FTAG, &dn);
1742 err = dnode_set_blksz(dn, size, ibs, tx);
1743 dnode_rele(dn, FTAG);
1748 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1754 * Send streams include each object's checksum function. This
1755 * check ensures that the receiving system can understand the
1756 * checksum function transmitted.
1758 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS);
1760 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1761 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS);
1762 dn->dn_checksum = checksum;
1763 dnode_setdirty(dn, tx);
1764 dnode_rele(dn, FTAG);
1768 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1774 * Send streams include each object's compression function. This
1775 * check ensures that the receiving system can understand the
1776 * compression function transmitted.
1778 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS);
1780 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1781 dn->dn_compress = compress;
1782 dnode_setdirty(dn, tx);
1783 dnode_rele(dn, FTAG);
1786 int zfs_mdcomp_disable = 0;
1787 TUNABLE_INT("vfs.zfs.mdcomp_disable", &zfs_mdcomp_disable);
1788 SYSCTL_INT(_vfs_zfs, OID_AUTO, mdcomp_disable, CTLFLAG_RW,
1789 &zfs_mdcomp_disable, 0, "Disable metadata compression");
1792 * When the "redundant_metadata" property is set to "most", only indirect
1793 * blocks of this level and higher will have an additional ditto block.
1795 int zfs_redundant_metadata_most_ditto_level = 2;
1798 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
1800 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
1801 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
1803 enum zio_checksum checksum = os->os_checksum;
1804 enum zio_compress compress = os->os_compress;
1805 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
1806 boolean_t dedup = B_FALSE;
1807 boolean_t nopwrite = B_FALSE;
1808 boolean_t dedup_verify = os->os_dedup_verify;
1809 int copies = os->os_copies;
1812 * We maintain different write policies for each of the following
1815 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1816 * 3. all other level 0 blocks
1819 if (zfs_mdcomp_disable) {
1820 compress = ZIO_COMPRESS_EMPTY;
1823 * XXX -- we should design a compression algorithm
1824 * that specializes in arrays of bps.
1826 compress = zio_compress_select(os->os_spa,
1827 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON);
1831 * Metadata always gets checksummed. If the data
1832 * checksum is multi-bit correctable, and it's not a
1833 * ZBT-style checksum, then it's suitable for metadata
1834 * as well. Otherwise, the metadata checksum defaults
1837 if (zio_checksum_table[checksum].ci_correctable < 1 ||
1838 zio_checksum_table[checksum].ci_eck)
1839 checksum = ZIO_CHECKSUM_FLETCHER_4;
1841 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL ||
1842 (os->os_redundant_metadata ==
1843 ZFS_REDUNDANT_METADATA_MOST &&
1844 (level >= zfs_redundant_metadata_most_ditto_level ||
1845 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))))
1847 } else if (wp & WP_NOFILL) {
1851 * If we're writing preallocated blocks, we aren't actually
1852 * writing them so don't set any policy properties. These
1853 * blocks are currently only used by an external subsystem
1854 * outside of zfs (i.e. dump) and not written by the zio
1857 compress = ZIO_COMPRESS_OFF;
1858 checksum = ZIO_CHECKSUM_NOPARITY;
1860 compress = zio_compress_select(os->os_spa, dn->dn_compress,
1863 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
1864 zio_checksum_select(dn->dn_checksum, checksum) :
1868 * Determine dedup setting. If we are in dmu_sync(),
1869 * we won't actually dedup now because that's all
1870 * done in syncing context; but we do want to use the
1871 * dedup checkum. If the checksum is not strong
1872 * enough to ensure unique signatures, force
1875 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
1876 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
1877 if (!zio_checksum_table[checksum].ci_dedup)
1878 dedup_verify = B_TRUE;
1882 * Enable nopwrite if we have a cryptographically secure
1883 * checksum that has no known collisions (i.e. SHA-256)
1884 * and compression is enabled. We don't enable nopwrite if
1885 * dedup is enabled as the two features are mutually exclusive.
1887 nopwrite = (!dedup && zio_checksum_table[checksum].ci_dedup &&
1888 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
1891 zp->zp_checksum = checksum;
1892 zp->zp_compress = compress;
1893 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
1894 zp->zp_level = level;
1895 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
1896 zp->zp_dedup = dedup;
1897 zp->zp_dedup_verify = dedup && dedup_verify;
1898 zp->zp_nopwrite = nopwrite;
1902 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
1908 * Sync any current changes before
1909 * we go trundling through the block pointers.
1911 err = dmu_object_wait_synced(os, object);
1916 err = dnode_hold(os, object, FTAG, &dn);
1921 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
1922 dnode_rele(dn, FTAG);
1928 * Given the ZFS object, if it contains any dirty nodes
1929 * this function flushes all dirty blocks to disk. This
1930 * ensures the DMU object info is updated. A more efficient
1931 * future version might just find the TXG with the maximum
1932 * ID and wait for that to be synced.
1935 dmu_object_wait_synced(objset_t *os, uint64_t object) {
1939 error = dnode_hold(os, object, FTAG, &dn);
1944 for (i = 0; i < TXG_SIZE; i++) {
1945 if (list_link_active(&dn->dn_dirty_link[i])) {
1949 dnode_rele(dn, FTAG);
1950 if (i != TXG_SIZE) {
1951 txg_wait_synced(dmu_objset_pool(os), 0);
1958 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
1962 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1963 mutex_enter(&dn->dn_mtx);
1967 doi->doi_data_block_size = dn->dn_datablksz;
1968 doi->doi_metadata_block_size = dn->dn_indblkshift ?
1969 1ULL << dn->dn_indblkshift : 0;
1970 doi->doi_type = dn->dn_type;
1971 doi->doi_bonus_type = dn->dn_bonustype;
1972 doi->doi_bonus_size = dn->dn_bonuslen;
1973 doi->doi_indirection = dn->dn_nlevels;
1974 doi->doi_checksum = dn->dn_checksum;
1975 doi->doi_compress = dn->dn_compress;
1976 doi->doi_nblkptr = dn->dn_nblkptr;
1977 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
1978 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
1979 doi->doi_fill_count = 0;
1980 for (int i = 0; i < dnp->dn_nblkptr; i++)
1981 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
1983 mutex_exit(&dn->dn_mtx);
1984 rw_exit(&dn->dn_struct_rwlock);
1988 * Get information on a DMU object.
1989 * If doi is NULL, just indicates whether the object exists.
1992 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
1995 int err = dnode_hold(os, object, FTAG, &dn);
2001 dmu_object_info_from_dnode(dn, doi);
2003 dnode_rele(dn, FTAG);
2008 * As above, but faster; can be used when you have a held dbuf in hand.
2011 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
2013 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2016 dmu_object_info_from_dnode(DB_DNODE(db), doi);
2021 * Faster still when you only care about the size.
2022 * This is specifically optimized for zfs_getattr().
2025 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
2026 u_longlong_t *nblk512)
2028 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2034 *blksize = dn->dn_datablksz;
2035 /* add 1 for dnode space */
2036 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
2037 SPA_MINBLOCKSHIFT) + 1;
2042 byteswap_uint64_array(void *vbuf, size_t size)
2044 uint64_t *buf = vbuf;
2045 size_t count = size >> 3;
2048 ASSERT((size & 7) == 0);
2050 for (i = 0; i < count; i++)
2051 buf[i] = BSWAP_64(buf[i]);
2055 byteswap_uint32_array(void *vbuf, size_t size)
2057 uint32_t *buf = vbuf;
2058 size_t count = size >> 2;
2061 ASSERT((size & 3) == 0);
2063 for (i = 0; i < count; i++)
2064 buf[i] = BSWAP_32(buf[i]);
2068 byteswap_uint16_array(void *vbuf, size_t size)
2070 uint16_t *buf = vbuf;
2071 size_t count = size >> 1;
2074 ASSERT((size & 1) == 0);
2076 for (i = 0; i < count; i++)
2077 buf[i] = BSWAP_16(buf[i]);
2082 byteswap_uint8_array(void *vbuf, size_t size)
2096 zio_compress_init();
2104 arc_fini(); /* arc depends on l2arc, so arc must go first */
2107 zio_compress_fini();