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 SYSCTL_INT(_vfs_zfs, OID_AUTO, nopwrite_enabled, CTLFLAG_RDTUN,
60 &zfs_nopwrite_enabled, 0, "Enable nopwrite feature");
62 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
63 { DMU_BSWAP_UINT8, TRUE, "unallocated" },
64 { DMU_BSWAP_ZAP, TRUE, "object directory" },
65 { DMU_BSWAP_UINT64, TRUE, "object array" },
66 { DMU_BSWAP_UINT8, TRUE, "packed nvlist" },
67 { DMU_BSWAP_UINT64, TRUE, "packed nvlist size" },
68 { DMU_BSWAP_UINT64, TRUE, "bpobj" },
69 { DMU_BSWAP_UINT64, TRUE, "bpobj header" },
70 { DMU_BSWAP_UINT64, TRUE, "SPA space map header" },
71 { DMU_BSWAP_UINT64, TRUE, "SPA space map" },
72 { DMU_BSWAP_UINT64, TRUE, "ZIL intent log" },
73 { DMU_BSWAP_DNODE, TRUE, "DMU dnode" },
74 { DMU_BSWAP_OBJSET, TRUE, "DMU objset" },
75 { DMU_BSWAP_UINT64, TRUE, "DSL directory" },
76 { DMU_BSWAP_ZAP, TRUE, "DSL directory child map"},
77 { DMU_BSWAP_ZAP, TRUE, "DSL dataset snap map" },
78 { DMU_BSWAP_ZAP, TRUE, "DSL props" },
79 { DMU_BSWAP_UINT64, TRUE, "DSL dataset" },
80 { DMU_BSWAP_ZNODE, TRUE, "ZFS znode" },
81 { DMU_BSWAP_OLDACL, TRUE, "ZFS V0 ACL" },
82 { DMU_BSWAP_UINT8, FALSE, "ZFS plain file" },
83 { DMU_BSWAP_ZAP, TRUE, "ZFS directory" },
84 { DMU_BSWAP_ZAP, TRUE, "ZFS master node" },
85 { DMU_BSWAP_ZAP, TRUE, "ZFS delete queue" },
86 { DMU_BSWAP_UINT8, FALSE, "zvol object" },
87 { DMU_BSWAP_ZAP, TRUE, "zvol prop" },
88 { DMU_BSWAP_UINT8, FALSE, "other uint8[]" },
89 { DMU_BSWAP_UINT64, FALSE, "other uint64[]" },
90 { DMU_BSWAP_ZAP, TRUE, "other ZAP" },
91 { DMU_BSWAP_ZAP, TRUE, "persistent error log" },
92 { DMU_BSWAP_UINT8, TRUE, "SPA history" },
93 { DMU_BSWAP_UINT64, TRUE, "SPA history offsets" },
94 { DMU_BSWAP_ZAP, TRUE, "Pool properties" },
95 { DMU_BSWAP_ZAP, TRUE, "DSL permissions" },
96 { DMU_BSWAP_ACL, TRUE, "ZFS ACL" },
97 { DMU_BSWAP_UINT8, TRUE, "ZFS SYSACL" },
98 { DMU_BSWAP_UINT8, TRUE, "FUID table" },
99 { DMU_BSWAP_UINT64, TRUE, "FUID table size" },
100 { DMU_BSWAP_ZAP, TRUE, "DSL dataset next clones"},
101 { DMU_BSWAP_ZAP, TRUE, "scan work queue" },
102 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group used" },
103 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group quota" },
104 { DMU_BSWAP_ZAP, TRUE, "snapshot refcount tags"},
105 { DMU_BSWAP_ZAP, TRUE, "DDT ZAP algorithm" },
106 { DMU_BSWAP_ZAP, TRUE, "DDT statistics" },
107 { DMU_BSWAP_UINT8, TRUE, "System attributes" },
108 { DMU_BSWAP_ZAP, TRUE, "SA master node" },
109 { DMU_BSWAP_ZAP, TRUE, "SA attr registration" },
110 { DMU_BSWAP_ZAP, TRUE, "SA attr layouts" },
111 { DMU_BSWAP_ZAP, TRUE, "scan translations" },
112 { DMU_BSWAP_UINT8, FALSE, "deduplicated block" },
113 { DMU_BSWAP_ZAP, TRUE, "DSL deadlist map" },
114 { DMU_BSWAP_UINT64, TRUE, "DSL deadlist map hdr" },
115 { DMU_BSWAP_ZAP, TRUE, "DSL dir clones" },
116 { DMU_BSWAP_UINT64, TRUE, "bpobj subobj" }
119 const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
120 { byteswap_uint8_array, "uint8" },
121 { byteswap_uint16_array, "uint16" },
122 { byteswap_uint32_array, "uint32" },
123 { byteswap_uint64_array, "uint64" },
124 { zap_byteswap, "zap" },
125 { dnode_buf_byteswap, "dnode" },
126 { dmu_objset_byteswap, "objset" },
127 { zfs_znode_byteswap, "znode" },
128 { zfs_oldacl_byteswap, "oldacl" },
129 { zfs_acl_byteswap, "acl" }
133 dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset,
134 void *tag, dmu_buf_t **dbp)
141 err = dnode_hold(os, object, FTAG, &dn);
144 blkid = dbuf_whichblock(dn, offset);
145 rw_enter(&dn->dn_struct_rwlock, RW_READER);
146 db = dbuf_hold(dn, blkid, tag);
147 rw_exit(&dn->dn_struct_rwlock);
148 dnode_rele(dn, FTAG);
152 return (SET_ERROR(EIO));
160 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
161 void *tag, dmu_buf_t **dbp, int flags)
164 int db_flags = DB_RF_CANFAIL;
166 if (flags & DMU_READ_NO_PREFETCH)
167 db_flags |= DB_RF_NOPREFETCH;
169 err = dmu_buf_hold_noread(os, object, offset, tag, dbp);
171 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
172 err = dbuf_read(db, NULL, db_flags);
185 return (DN_MAX_BONUSLEN);
189 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
191 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
198 if (dn->dn_bonus != db) {
199 error = SET_ERROR(EINVAL);
200 } else if (newsize < 0 || newsize > db_fake->db_size) {
201 error = SET_ERROR(EINVAL);
203 dnode_setbonuslen(dn, newsize, tx);
212 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
214 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
221 if (!DMU_OT_IS_VALID(type)) {
222 error = SET_ERROR(EINVAL);
223 } else if (dn->dn_bonus != db) {
224 error = SET_ERROR(EINVAL);
226 dnode_setbonus_type(dn, type, tx);
235 dmu_get_bonustype(dmu_buf_t *db_fake)
237 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
239 dmu_object_type_t type;
243 type = dn->dn_bonustype;
250 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
255 error = dnode_hold(os, object, FTAG, &dn);
256 dbuf_rm_spill(dn, tx);
257 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
258 dnode_rm_spill(dn, tx);
259 rw_exit(&dn->dn_struct_rwlock);
260 dnode_rele(dn, FTAG);
265 * returns ENOENT, EIO, or 0.
268 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
274 error = dnode_hold(os, object, FTAG, &dn);
278 rw_enter(&dn->dn_struct_rwlock, RW_READER);
279 if (dn->dn_bonus == NULL) {
280 rw_exit(&dn->dn_struct_rwlock);
281 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
282 if (dn->dn_bonus == NULL)
283 dbuf_create_bonus(dn);
287 /* as long as the bonus buf is held, the dnode will be held */
288 if (refcount_add(&db->db_holds, tag) == 1) {
289 VERIFY(dnode_add_ref(dn, db));
290 atomic_inc_32(&dn->dn_dbufs_count);
294 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
295 * hold and incrementing the dbuf count to ensure that dnode_move() sees
296 * a dnode hold for every dbuf.
298 rw_exit(&dn->dn_struct_rwlock);
300 dnode_rele(dn, FTAG);
302 VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH));
309 * returns ENOENT, EIO, or 0.
311 * This interface will allocate a blank spill dbuf when a spill blk
312 * doesn't already exist on the dnode.
314 * if you only want to find an already existing spill db, then
315 * dmu_spill_hold_existing() should be used.
318 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
320 dmu_buf_impl_t *db = NULL;
323 if ((flags & DB_RF_HAVESTRUCT) == 0)
324 rw_enter(&dn->dn_struct_rwlock, RW_READER);
326 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
328 if ((flags & DB_RF_HAVESTRUCT) == 0)
329 rw_exit(&dn->dn_struct_rwlock);
332 err = dbuf_read(db, NULL, flags);
341 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
343 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
350 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
351 err = SET_ERROR(EINVAL);
353 rw_enter(&dn->dn_struct_rwlock, RW_READER);
355 if (!dn->dn_have_spill) {
356 err = SET_ERROR(ENOENT);
358 err = dmu_spill_hold_by_dnode(dn,
359 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
362 rw_exit(&dn->dn_struct_rwlock);
370 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
372 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
378 err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp);
385 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
386 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
387 * and can induce severe lock contention when writing to several files
388 * whose dnodes are in the same block.
391 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
392 int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
395 uint64_t blkid, nblks, i;
400 ASSERT(length <= DMU_MAX_ACCESS);
402 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT;
403 if (flags & DMU_READ_NO_PREFETCH || length > zfetch_array_rd_sz)
404 dbuf_flags |= DB_RF_NOPREFETCH;
406 rw_enter(&dn->dn_struct_rwlock, RW_READER);
407 if (dn->dn_datablkshift) {
408 int blkshift = dn->dn_datablkshift;
409 nblks = (P2ROUNDUP(offset+length, 1ULL<<blkshift) -
410 P2ALIGN(offset, 1ULL<<blkshift)) >> blkshift;
412 if (offset + length > dn->dn_datablksz) {
413 zfs_panic_recover("zfs: accessing past end of object "
414 "%llx/%llx (size=%u access=%llu+%llu)",
415 (longlong_t)dn->dn_objset->
416 os_dsl_dataset->ds_object,
417 (longlong_t)dn->dn_object, dn->dn_datablksz,
418 (longlong_t)offset, (longlong_t)length);
419 rw_exit(&dn->dn_struct_rwlock);
420 return (SET_ERROR(EIO));
424 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
426 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
427 blkid = dbuf_whichblock(dn, offset);
428 for (i = 0; i < nblks; i++) {
429 dmu_buf_impl_t *db = dbuf_hold(dn, blkid+i, tag);
431 rw_exit(&dn->dn_struct_rwlock);
432 dmu_buf_rele_array(dbp, nblks, tag);
434 return (SET_ERROR(EIO));
436 /* initiate async i/o */
438 (void) dbuf_read(db, zio, dbuf_flags);
441 curthread->td_ru.ru_oublock++;
445 rw_exit(&dn->dn_struct_rwlock);
447 /* wait for async i/o */
450 dmu_buf_rele_array(dbp, nblks, tag);
454 /* wait for other io to complete */
456 for (i = 0; i < nblks; i++) {
457 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
458 mutex_enter(&db->db_mtx);
459 while (db->db_state == DB_READ ||
460 db->db_state == DB_FILL)
461 cv_wait(&db->db_changed, &db->db_mtx);
462 if (db->db_state == DB_UNCACHED)
463 err = SET_ERROR(EIO);
464 mutex_exit(&db->db_mtx);
466 dmu_buf_rele_array(dbp, nblks, tag);
478 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
479 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
484 err = dnode_hold(os, object, FTAG, &dn);
488 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
489 numbufsp, dbpp, DMU_READ_PREFETCH);
491 dnode_rele(dn, FTAG);
497 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
498 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
500 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
506 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
507 numbufsp, dbpp, DMU_READ_PREFETCH);
514 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
517 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
522 for (i = 0; i < numbufs; i++) {
524 dbuf_rele(dbp[i], tag);
527 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
531 * Issue prefetch i/os for the given blocks.
533 * Note: The assumption is that we *know* these blocks will be needed
534 * almost immediately. Therefore, the prefetch i/os will be issued at
535 * ZIO_PRIORITY_SYNC_READ
537 * Note: indirect blocks and other metadata will be read synchronously,
538 * causing this function to block if they are not already cached.
541 dmu_prefetch(objset_t *os, uint64_t object, uint64_t offset, uint64_t len)
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, object * sizeof (dnode_phys_t));
558 dbuf_prefetch(dn, blkid, ZIO_PRIORITY_SYNC_READ);
559 rw_exit(&dn->dn_struct_rwlock);
564 * XXX - Note, if the dnode for the requested object is not
565 * already cached, we will do a *synchronous* read in the
566 * dnode_hold() call. The same is true for any indirects.
568 err = dnode_hold(os, object, FTAG, &dn);
572 rw_enter(&dn->dn_struct_rwlock, RW_READER);
573 if (dn->dn_datablkshift) {
574 int blkshift = dn->dn_datablkshift;
575 nblks = (P2ROUNDUP(offset + len, 1 << blkshift) -
576 P2ALIGN(offset, 1 << blkshift)) >> blkshift;
578 nblks = (offset < dn->dn_datablksz);
582 blkid = dbuf_whichblock(dn, offset);
583 for (int i = 0; i < nblks; i++)
584 dbuf_prefetch(dn, blkid + i, ZIO_PRIORITY_SYNC_READ);
587 rw_exit(&dn->dn_struct_rwlock);
589 dnode_rele(dn, FTAG);
593 * Get the next "chunk" of file data to free. We traverse the file from
594 * the end so that the file gets shorter over time (if we crashes in the
595 * middle, this will leave us in a better state). We find allocated file
596 * data by simply searching the allocated level 1 indirects.
598 * On input, *start should be the first offset that does not need to be
599 * freed (e.g. "offset + length"). On return, *start will be the first
600 * offset that should be freed.
603 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum)
605 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
606 /* bytes of data covered by a level-1 indirect block */
608 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
610 ASSERT3U(minimum, <=, *start);
612 if (*start - minimum <= iblkrange * maxblks) {
616 ASSERT(ISP2(iblkrange));
618 for (uint64_t blks = 0; *start > minimum && blks < maxblks; blks++) {
622 * dnode_next_offset(BACKWARDS) will find an allocated L1
623 * indirect block at or before the input offset. We must
624 * decrement *start so that it is at the end of the region
628 err = dnode_next_offset(dn,
629 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
631 /* if there are no indirect blocks before start, we are done */
635 } else if (err != 0) {
639 /* set start to the beginning of this L1 indirect */
640 *start = P2ALIGN(*start, iblkrange);
642 if (*start < minimum)
648 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
651 uint64_t object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
654 if (offset >= object_size)
657 if (length == DMU_OBJECT_END || offset + length > object_size)
658 length = object_size - offset;
660 while (length != 0) {
661 uint64_t chunk_end, chunk_begin;
663 chunk_end = chunk_begin = offset + length;
665 /* move chunk_begin backwards to the beginning of this chunk */
666 err = get_next_chunk(dn, &chunk_begin, offset);
669 ASSERT3U(chunk_begin, >=, offset);
670 ASSERT3U(chunk_begin, <=, chunk_end);
672 dmu_tx_t *tx = dmu_tx_create(os);
673 dmu_tx_hold_free(tx, dn->dn_object,
674 chunk_begin, chunk_end - chunk_begin);
677 * Mark this transaction as typically resulting in a net
678 * reduction in space used.
680 dmu_tx_mark_netfree(tx);
681 err = dmu_tx_assign(tx, TXG_WAIT);
686 dnode_free_range(dn, chunk_begin, chunk_end - chunk_begin, tx);
689 length -= chunk_end - chunk_begin;
695 dmu_free_long_range(objset_t *os, uint64_t object,
696 uint64_t offset, uint64_t length)
701 err = dnode_hold(os, object, FTAG, &dn);
704 err = dmu_free_long_range_impl(os, dn, offset, length);
707 * It is important to zero out the maxblkid when freeing the entire
708 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
709 * will take the fast path, and (b) dnode_reallocate() can verify
710 * that the entire file has been freed.
712 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
715 dnode_rele(dn, FTAG);
720 dmu_free_long_object(objset_t *os, uint64_t object)
725 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
729 tx = dmu_tx_create(os);
730 dmu_tx_hold_bonus(tx, object);
731 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
732 dmu_tx_mark_netfree(tx);
733 err = dmu_tx_assign(tx, TXG_WAIT);
735 err = dmu_object_free(os, object, tx);
745 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
746 uint64_t size, dmu_tx_t *tx)
749 int err = dnode_hold(os, object, FTAG, &dn);
752 ASSERT(offset < UINT64_MAX);
753 ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
754 dnode_free_range(dn, offset, size, tx);
755 dnode_rele(dn, FTAG);
760 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
761 void *buf, uint32_t flags)
767 err = dnode_hold(os, object, FTAG, &dn);
772 * Deal with odd block sizes, where there can't be data past the first
773 * block. If we ever do the tail block optimization, we will need to
774 * handle that here as well.
776 if (dn->dn_maxblkid == 0) {
777 int newsz = offset > dn->dn_datablksz ? 0 :
778 MIN(size, dn->dn_datablksz - offset);
779 bzero((char *)buf + newsz, size - newsz);
784 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
788 * NB: we could do this block-at-a-time, but it's nice
789 * to be reading in parallel.
791 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
792 TRUE, FTAG, &numbufs, &dbp, flags);
796 for (i = 0; i < numbufs; i++) {
799 dmu_buf_t *db = dbp[i];
803 bufoff = offset - db->db_offset;
804 tocpy = (int)MIN(db->db_size - bufoff, size);
806 bcopy((char *)db->db_data + bufoff, buf, tocpy);
810 buf = (char *)buf + tocpy;
812 dmu_buf_rele_array(dbp, numbufs, FTAG);
814 dnode_rele(dn, FTAG);
819 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
820 const void *buf, dmu_tx_t *tx)
828 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
829 FALSE, FTAG, &numbufs, &dbp));
831 for (i = 0; i < numbufs; i++) {
834 dmu_buf_t *db = dbp[i];
838 bufoff = offset - db->db_offset;
839 tocpy = (int)MIN(db->db_size - bufoff, size);
841 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
843 if (tocpy == db->db_size)
844 dmu_buf_will_fill(db, tx);
846 dmu_buf_will_dirty(db, tx);
848 bcopy(buf, (char *)db->db_data + bufoff, tocpy);
850 if (tocpy == db->db_size)
851 dmu_buf_fill_done(db, tx);
855 buf = (char *)buf + tocpy;
857 dmu_buf_rele_array(dbp, numbufs, FTAG);
861 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
870 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
871 FALSE, FTAG, &numbufs, &dbp));
873 for (i = 0; i < numbufs; i++) {
874 dmu_buf_t *db = dbp[i];
876 dmu_buf_will_not_fill(db, tx);
878 dmu_buf_rele_array(dbp, numbufs, FTAG);
882 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
883 void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
884 int compressed_size, int byteorder, dmu_tx_t *tx)
888 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
889 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
890 VERIFY0(dmu_buf_hold_noread(os, object, offset,
893 dmu_buf_write_embedded(db,
894 data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
895 uncompressed_size, compressed_size, byteorder, tx);
897 dmu_buf_rele(db, FTAG);
901 * DMU support for xuio
903 kstat_t *xuio_ksp = NULL;
906 dmu_xuio_init(xuio_t *xuio, int nblk)
909 uio_t *uio = &xuio->xu_uio;
911 uio->uio_iovcnt = nblk;
912 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
914 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
916 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
917 priv->iovp = uio->uio_iov;
918 XUIO_XUZC_PRIV(xuio) = priv;
920 if (XUIO_XUZC_RW(xuio) == UIO_READ)
921 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
923 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
929 dmu_xuio_fini(xuio_t *xuio)
931 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
932 int nblk = priv->cnt;
934 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
935 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
936 kmem_free(priv, sizeof (dmu_xuio_t));
938 if (XUIO_XUZC_RW(xuio) == UIO_READ)
939 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
941 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
945 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
946 * and increase priv->next by 1.
949 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
952 uio_t *uio = &xuio->xu_uio;
953 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
954 int i = priv->next++;
956 ASSERT(i < priv->cnt);
957 ASSERT(off + n <= arc_buf_size(abuf));
958 iov = uio->uio_iov + i;
959 iov->iov_base = (char *)abuf->b_data + off;
961 priv->bufs[i] = abuf;
966 dmu_xuio_cnt(xuio_t *xuio)
968 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
973 dmu_xuio_arcbuf(xuio_t *xuio, int i)
975 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
977 ASSERT(i < priv->cnt);
978 return (priv->bufs[i]);
982 dmu_xuio_clear(xuio_t *xuio, int i)
984 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
986 ASSERT(i < priv->cnt);
987 priv->bufs[i] = NULL;
993 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
994 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
996 if (xuio_ksp != NULL) {
997 xuio_ksp->ks_data = &xuio_stats;
998 kstat_install(xuio_ksp);
1003 xuio_stat_fini(void)
1005 if (xuio_ksp != NULL) {
1006 kstat_delete(xuio_ksp);
1012 xuio_stat_wbuf_copied()
1014 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1018 xuio_stat_wbuf_nocopy()
1020 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
1025 dmu_read_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size)
1028 int numbufs, i, err;
1029 xuio_t *xuio = NULL;
1032 * NB: we could do this block-at-a-time, but it's nice
1033 * to be reading in parallel.
1035 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1036 TRUE, FTAG, &numbufs, &dbp, 0);
1041 if (uio->uio_extflg == UIO_XUIO)
1042 xuio = (xuio_t *)uio;
1045 for (i = 0; i < numbufs; i++) {
1048 dmu_buf_t *db = dbp[i];
1052 bufoff = uio->uio_loffset - db->db_offset;
1053 tocpy = (int)MIN(db->db_size - bufoff, size);
1056 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
1057 arc_buf_t *dbuf_abuf = dbi->db_buf;
1058 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
1059 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
1061 uio->uio_resid -= tocpy;
1062 uio->uio_loffset += tocpy;
1065 if (abuf == dbuf_abuf)
1066 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
1068 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
1070 err = uiomove((char *)db->db_data + bufoff, tocpy,
1078 dmu_buf_rele_array(dbp, numbufs, FTAG);
1084 * Read 'size' bytes into the uio buffer.
1085 * From object zdb->db_object.
1086 * Starting at offset uio->uio_loffset.
1088 * If the caller already has a dbuf in the target object
1089 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1090 * because we don't have to find the dnode_t for the object.
1093 dmu_read_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size)
1095 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1104 err = dmu_read_uio_dnode(dn, uio, size);
1111 * Read 'size' bytes into the uio buffer.
1112 * From the specified object
1113 * Starting at offset uio->uio_loffset.
1116 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
1124 err = dnode_hold(os, object, FTAG, &dn);
1128 err = dmu_read_uio_dnode(dn, uio, size);
1130 dnode_rele(dn, FTAG);
1136 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1143 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1144 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1148 for (i = 0; i < numbufs; i++) {
1151 dmu_buf_t *db = dbp[i];
1155 bufoff = uio->uio_loffset - db->db_offset;
1156 tocpy = (int)MIN(db->db_size - bufoff, size);
1158 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1160 if (tocpy == db->db_size)
1161 dmu_buf_will_fill(db, tx);
1163 dmu_buf_will_dirty(db, tx);
1166 * XXX uiomove could block forever (eg. nfs-backed
1167 * pages). There needs to be a uiolockdown() function
1168 * to lock the pages in memory, so that uiomove won't
1171 err = uiomove((char *)db->db_data + bufoff, tocpy,
1174 if (tocpy == db->db_size)
1175 dmu_buf_fill_done(db, tx);
1183 dmu_buf_rele_array(dbp, numbufs, FTAG);
1188 * Write 'size' bytes from the uio buffer.
1189 * To object zdb->db_object.
1190 * Starting at offset uio->uio_loffset.
1192 * If the caller already has a dbuf in the target object
1193 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1194 * because we don't have to find the dnode_t for the object.
1197 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1200 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1209 err = dmu_write_uio_dnode(dn, uio, size, tx);
1216 * Write 'size' bytes from the uio buffer.
1217 * To the specified object.
1218 * Starting at offset uio->uio_loffset.
1221 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1230 err = dnode_hold(os, object, FTAG, &dn);
1234 err = dmu_write_uio_dnode(dn, uio, size, tx);
1236 dnode_rele(dn, FTAG);
1243 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1244 page_t *pp, dmu_tx_t *tx)
1253 err = dmu_buf_hold_array(os, object, offset, size,
1254 FALSE, FTAG, &numbufs, &dbp);
1258 for (i = 0; i < numbufs; i++) {
1259 int tocpy, copied, thiscpy;
1261 dmu_buf_t *db = dbp[i];
1265 ASSERT3U(db->db_size, >=, PAGESIZE);
1267 bufoff = offset - db->db_offset;
1268 tocpy = (int)MIN(db->db_size - bufoff, size);
1270 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1272 if (tocpy == db->db_size)
1273 dmu_buf_will_fill(db, tx);
1275 dmu_buf_will_dirty(db, tx);
1277 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1278 ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff);
1279 thiscpy = MIN(PAGESIZE, tocpy - copied);
1280 va = zfs_map_page(pp, S_READ);
1281 bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1282 zfs_unmap_page(pp, va);
1287 if (tocpy == db->db_size)
1288 dmu_buf_fill_done(db, tx);
1293 dmu_buf_rele_array(dbp, numbufs, FTAG);
1297 #else /* !illumos */
1300 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1301 vm_page_t *ma, dmu_tx_t *tx)
1311 err = dmu_buf_hold_array(os, object, offset, size,
1312 FALSE, FTAG, &numbufs, &dbp);
1316 for (i = 0; i < numbufs; i++) {
1317 int tocpy, copied, thiscpy;
1319 dmu_buf_t *db = dbp[i];
1323 ASSERT3U(db->db_size, >=, PAGESIZE);
1325 bufoff = offset - db->db_offset;
1326 tocpy = (int)MIN(db->db_size - bufoff, size);
1328 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1330 if (tocpy == db->db_size)
1331 dmu_buf_will_fill(db, tx);
1333 dmu_buf_will_dirty(db, tx);
1335 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1336 ASSERT3U(ptoa((*ma)->pindex), ==, db->db_offset + bufoff);
1337 thiscpy = MIN(PAGESIZE, tocpy - copied);
1338 va = zfs_map_page(*ma, &sf);
1339 bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1345 if (tocpy == db->db_size)
1346 dmu_buf_fill_done(db, tx);
1351 dmu_buf_rele_array(dbp, numbufs, FTAG);
1354 #endif /* illumos */
1355 #endif /* _KERNEL */
1358 * Allocate a loaned anonymous arc buffer.
1361 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1363 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1365 return (arc_loan_buf(db->db_objset->os_spa, size));
1369 * Free a loaned arc buffer.
1372 dmu_return_arcbuf(arc_buf_t *buf)
1374 arc_return_buf(buf, FTAG);
1375 VERIFY(arc_buf_remove_ref(buf, FTAG));
1379 * When possible directly assign passed loaned arc buffer to a dbuf.
1380 * If this is not possible copy the contents of passed arc buf via
1384 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1387 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1390 uint32_t blksz = (uint32_t)arc_buf_size(buf);
1393 DB_DNODE_ENTER(dbuf);
1394 dn = DB_DNODE(dbuf);
1395 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1396 blkid = dbuf_whichblock(dn, offset);
1397 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1398 rw_exit(&dn->dn_struct_rwlock);
1399 DB_DNODE_EXIT(dbuf);
1402 * We can only assign if the offset is aligned, the arc buf is the
1403 * same size as the dbuf, and the dbuf is not metadata. It
1404 * can't be metadata because the loaned arc buf comes from the
1405 * user-data kmem arena.
1407 if (offset == db->db.db_offset && blksz == db->db.db_size &&
1408 DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA) {
1409 dbuf_assign_arcbuf(db, buf, tx);
1410 dbuf_rele(db, FTAG);
1415 DB_DNODE_ENTER(dbuf);
1416 dn = DB_DNODE(dbuf);
1418 object = dn->dn_object;
1419 DB_DNODE_EXIT(dbuf);
1421 dbuf_rele(db, FTAG);
1422 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1423 dmu_return_arcbuf(buf);
1424 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1429 dbuf_dirty_record_t *dsa_dr;
1430 dmu_sync_cb_t *dsa_done;
1437 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1439 dmu_sync_arg_t *dsa = varg;
1440 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1441 blkptr_t *bp = zio->io_bp;
1443 if (zio->io_error == 0) {
1444 if (BP_IS_HOLE(bp)) {
1446 * A block of zeros may compress to a hole, but the
1447 * block size still needs to be known for replay.
1449 BP_SET_LSIZE(bp, db->db_size);
1450 } else if (!BP_IS_EMBEDDED(bp)) {
1451 ASSERT(BP_GET_LEVEL(bp) == 0);
1458 dmu_sync_late_arrival_ready(zio_t *zio)
1460 dmu_sync_ready(zio, NULL, zio->io_private);
1465 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1467 dmu_sync_arg_t *dsa = varg;
1468 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1469 dmu_buf_impl_t *db = dr->dr_dbuf;
1471 mutex_enter(&db->db_mtx);
1472 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1473 if (zio->io_error == 0) {
1474 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1475 if (dr->dt.dl.dr_nopwrite) {
1476 blkptr_t *bp = zio->io_bp;
1477 blkptr_t *bp_orig = &zio->io_bp_orig;
1478 uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1480 ASSERT(BP_EQUAL(bp, bp_orig));
1481 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1482 ASSERT(zio_checksum_table[chksum].ci_dedup);
1484 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1485 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1486 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1487 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by))
1488 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1490 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1492 cv_broadcast(&db->db_changed);
1493 mutex_exit(&db->db_mtx);
1495 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1497 kmem_free(dsa, sizeof (*dsa));
1501 dmu_sync_late_arrival_done(zio_t *zio)
1503 blkptr_t *bp = zio->io_bp;
1504 dmu_sync_arg_t *dsa = zio->io_private;
1505 blkptr_t *bp_orig = &zio->io_bp_orig;
1507 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1509 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE)
1510 * then there is nothing to do here. Otherwise, free the
1511 * newly allocated block in this txg.
1513 if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
1514 ASSERT(BP_EQUAL(bp, bp_orig));
1516 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1517 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1518 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1519 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1523 dmu_tx_commit(dsa->dsa_tx);
1525 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1527 kmem_free(dsa, sizeof (*dsa));
1531 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1532 zio_prop_t *zp, zbookmark_phys_t *zb)
1534 dmu_sync_arg_t *dsa;
1537 tx = dmu_tx_create(os);
1538 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1539 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1541 /* Make zl_get_data do txg_waited_synced() */
1542 return (SET_ERROR(EIO));
1545 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1547 dsa->dsa_done = done;
1551 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1552 zgd->zgd_db->db_data, zgd->zgd_db->db_size, zp,
1553 dmu_sync_late_arrival_ready, NULL, dmu_sync_late_arrival_done, dsa,
1554 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1560 * Intent log support: sync the block associated with db to disk.
1561 * N.B. and XXX: the caller is responsible for making sure that the
1562 * data isn't changing while dmu_sync() is writing it.
1566 * EEXIST: this txg has already been synced, so there's nothing to do.
1567 * The caller should not log the write.
1569 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1570 * The caller should not log the write.
1572 * EALREADY: this block is already in the process of being synced.
1573 * The caller should track its progress (somehow).
1575 * EIO: could not do the I/O.
1576 * The caller should do a txg_wait_synced().
1578 * 0: the I/O has been initiated.
1579 * The caller should log this blkptr in the done callback.
1580 * It is possible that the I/O will fail, in which case
1581 * the error will be reported to the done callback and
1582 * propagated to pio from zio_done().
1585 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1587 blkptr_t *bp = zgd->zgd_bp;
1588 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1589 objset_t *os = db->db_objset;
1590 dsl_dataset_t *ds = os->os_dsl_dataset;
1591 dbuf_dirty_record_t *dr;
1592 dmu_sync_arg_t *dsa;
1593 zbookmark_phys_t zb;
1597 ASSERT(pio != NULL);
1600 SET_BOOKMARK(&zb, ds->ds_object,
1601 db->db.db_object, db->db_level, db->db_blkid);
1605 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1609 * If we're frozen (running ziltest), we always need to generate a bp.
1611 if (txg > spa_freeze_txg(os->os_spa))
1612 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1615 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1616 * and us. If we determine that this txg is not yet syncing,
1617 * but it begins to sync a moment later, that's OK because the
1618 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1620 mutex_enter(&db->db_mtx);
1622 if (txg <= spa_last_synced_txg(os->os_spa)) {
1624 * This txg has already synced. There's nothing to do.
1626 mutex_exit(&db->db_mtx);
1627 return (SET_ERROR(EEXIST));
1630 if (txg <= spa_syncing_txg(os->os_spa)) {
1632 * This txg is currently syncing, so we can't mess with
1633 * the dirty record anymore; just write a new log block.
1635 mutex_exit(&db->db_mtx);
1636 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1639 dr = db->db_last_dirty;
1640 while (dr && dr->dr_txg != txg)
1645 * There's no dr for this dbuf, so it must have been freed.
1646 * There's no need to log writes to freed blocks, so we're done.
1648 mutex_exit(&db->db_mtx);
1649 return (SET_ERROR(ENOENT));
1652 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
1655 * Assume the on-disk data is X, the current syncing data is Y,
1656 * and the current in-memory data is Z (currently in dmu_sync).
1657 * X and Z are identical but Y is has been modified. Normally,
1658 * when X and Z are the same we will perform a nopwrite but if Y
1659 * is different we must disable nopwrite since the resulting write
1660 * of Y to disk can free the block containing X. If we allowed a
1661 * nopwrite to occur the block pointing to Z would reference a freed
1662 * block. Since this is a rare case we simplify this by disabling
1663 * nopwrite if the current dmu_sync-ing dbuf has been modified in
1664 * a previous transaction.
1667 zp.zp_nopwrite = B_FALSE;
1669 ASSERT(dr->dr_txg == txg);
1670 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1671 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1673 * We have already issued a sync write for this buffer,
1674 * or this buffer has already been synced. It could not
1675 * have been dirtied since, or we would have cleared the state.
1677 mutex_exit(&db->db_mtx);
1678 return (SET_ERROR(EALREADY));
1681 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1682 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1683 mutex_exit(&db->db_mtx);
1685 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1687 dsa->dsa_done = done;
1691 zio_nowait(arc_write(pio, os->os_spa, txg,
1692 bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
1693 DBUF_IS_L2COMPRESSIBLE(db), &zp, dmu_sync_ready,
1694 NULL, dmu_sync_done, dsa, ZIO_PRIORITY_SYNC_WRITE,
1695 ZIO_FLAG_CANFAIL, &zb));
1701 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1707 err = dnode_hold(os, object, FTAG, &dn);
1710 err = dnode_set_blksz(dn, size, ibs, tx);
1711 dnode_rele(dn, FTAG);
1716 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1722 * Send streams include each object's checksum function. This
1723 * check ensures that the receiving system can understand the
1724 * checksum function transmitted.
1726 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS);
1728 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1729 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS);
1730 dn->dn_checksum = checksum;
1731 dnode_setdirty(dn, tx);
1732 dnode_rele(dn, FTAG);
1736 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1742 * Send streams include each object's compression function. This
1743 * check ensures that the receiving system can understand the
1744 * compression function transmitted.
1746 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS);
1748 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1749 dn->dn_compress = compress;
1750 dnode_setdirty(dn, tx);
1751 dnode_rele(dn, FTAG);
1754 int zfs_mdcomp_disable = 0;
1755 SYSCTL_INT(_vfs_zfs, OID_AUTO, mdcomp_disable, CTLFLAG_RWTUN,
1756 &zfs_mdcomp_disable, 0, "Disable metadata compression");
1759 * When the "redundant_metadata" property is set to "most", only indirect
1760 * blocks of this level and higher will have an additional ditto block.
1762 int zfs_redundant_metadata_most_ditto_level = 2;
1765 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
1767 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
1768 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
1770 enum zio_checksum checksum = os->os_checksum;
1771 enum zio_compress compress = os->os_compress;
1772 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
1773 boolean_t dedup = B_FALSE;
1774 boolean_t nopwrite = B_FALSE;
1775 boolean_t dedup_verify = os->os_dedup_verify;
1776 int copies = os->os_copies;
1779 * We maintain different write policies for each of the following
1782 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1783 * 3. all other level 0 blocks
1787 * XXX -- we should design a compression algorithm
1788 * that specializes in arrays of bps.
1790 boolean_t lz4_ac = spa_feature_is_active(os->os_spa,
1791 SPA_FEATURE_LZ4_COMPRESS);
1793 if (zfs_mdcomp_disable) {
1794 compress = ZIO_COMPRESS_EMPTY;
1795 } else if (lz4_ac) {
1796 compress = ZIO_COMPRESS_LZ4;
1798 compress = ZIO_COMPRESS_LZJB;
1802 * Metadata always gets checksummed. If the data
1803 * checksum is multi-bit correctable, and it's not a
1804 * ZBT-style checksum, then it's suitable for metadata
1805 * as well. Otherwise, the metadata checksum defaults
1808 if (zio_checksum_table[checksum].ci_correctable < 1 ||
1809 zio_checksum_table[checksum].ci_eck)
1810 checksum = ZIO_CHECKSUM_FLETCHER_4;
1812 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL ||
1813 (os->os_redundant_metadata ==
1814 ZFS_REDUNDANT_METADATA_MOST &&
1815 (level >= zfs_redundant_metadata_most_ditto_level ||
1816 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))))
1818 } else if (wp & WP_NOFILL) {
1822 * If we're writing preallocated blocks, we aren't actually
1823 * writing them so don't set any policy properties. These
1824 * blocks are currently only used by an external subsystem
1825 * outside of zfs (i.e. dump) and not written by the zio
1828 compress = ZIO_COMPRESS_OFF;
1829 checksum = ZIO_CHECKSUM_NOPARITY;
1831 compress = zio_compress_select(dn->dn_compress, compress);
1833 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
1834 zio_checksum_select(dn->dn_checksum, checksum) :
1838 * Determine dedup setting. If we are in dmu_sync(),
1839 * we won't actually dedup now because that's all
1840 * done in syncing context; but we do want to use the
1841 * dedup checkum. If the checksum is not strong
1842 * enough to ensure unique signatures, force
1845 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
1846 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
1847 if (!zio_checksum_table[checksum].ci_dedup)
1848 dedup_verify = B_TRUE;
1852 * Enable nopwrite if we have a cryptographically secure
1853 * checksum that has no known collisions (i.e. SHA-256)
1854 * and compression is enabled. We don't enable nopwrite if
1855 * dedup is enabled as the two features are mutually exclusive.
1857 nopwrite = (!dedup && zio_checksum_table[checksum].ci_dedup &&
1858 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
1861 zp->zp_checksum = checksum;
1862 zp->zp_compress = compress;
1863 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
1864 zp->zp_level = level;
1865 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
1866 zp->zp_dedup = dedup;
1867 zp->zp_dedup_verify = dedup && dedup_verify;
1868 zp->zp_nopwrite = nopwrite;
1872 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
1877 err = dnode_hold(os, object, FTAG, &dn);
1881 * Sync any current changes before
1882 * we go trundling through the block pointers.
1884 for (i = 0; i < TXG_SIZE; i++) {
1885 if (list_link_active(&dn->dn_dirty_link[i]))
1888 if (i != TXG_SIZE) {
1889 dnode_rele(dn, FTAG);
1890 txg_wait_synced(dmu_objset_pool(os), 0);
1891 err = dnode_hold(os, object, FTAG, &dn);
1896 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
1897 dnode_rele(dn, FTAG);
1903 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
1907 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1908 mutex_enter(&dn->dn_mtx);
1912 doi->doi_data_block_size = dn->dn_datablksz;
1913 doi->doi_metadata_block_size = dn->dn_indblkshift ?
1914 1ULL << dn->dn_indblkshift : 0;
1915 doi->doi_type = dn->dn_type;
1916 doi->doi_bonus_type = dn->dn_bonustype;
1917 doi->doi_bonus_size = dn->dn_bonuslen;
1918 doi->doi_indirection = dn->dn_nlevels;
1919 doi->doi_checksum = dn->dn_checksum;
1920 doi->doi_compress = dn->dn_compress;
1921 doi->doi_nblkptr = dn->dn_nblkptr;
1922 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
1923 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
1924 doi->doi_fill_count = 0;
1925 for (int i = 0; i < dnp->dn_nblkptr; i++)
1926 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
1928 mutex_exit(&dn->dn_mtx);
1929 rw_exit(&dn->dn_struct_rwlock);
1933 * Get information on a DMU object.
1934 * If doi is NULL, just indicates whether the object exists.
1937 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
1940 int err = dnode_hold(os, object, FTAG, &dn);
1946 dmu_object_info_from_dnode(dn, doi);
1948 dnode_rele(dn, FTAG);
1953 * As above, but faster; can be used when you have a held dbuf in hand.
1956 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
1958 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1961 dmu_object_info_from_dnode(DB_DNODE(db), doi);
1966 * Faster still when you only care about the size.
1967 * This is specifically optimized for zfs_getattr().
1970 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
1971 u_longlong_t *nblk512)
1973 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1979 *blksize = dn->dn_datablksz;
1980 /* add 1 for dnode space */
1981 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
1982 SPA_MINBLOCKSHIFT) + 1;
1987 byteswap_uint64_array(void *vbuf, size_t size)
1989 uint64_t *buf = vbuf;
1990 size_t count = size >> 3;
1993 ASSERT((size & 7) == 0);
1995 for (i = 0; i < count; i++)
1996 buf[i] = BSWAP_64(buf[i]);
2000 byteswap_uint32_array(void *vbuf, size_t size)
2002 uint32_t *buf = vbuf;
2003 size_t count = size >> 2;
2006 ASSERT((size & 3) == 0);
2008 for (i = 0; i < count; i++)
2009 buf[i] = BSWAP_32(buf[i]);
2013 byteswap_uint16_array(void *vbuf, size_t size)
2015 uint16_t *buf = vbuf;
2016 size_t count = size >> 1;
2019 ASSERT((size & 1) == 0);
2021 for (i = 0; i < count; i++)
2022 buf[i] = BSWAP_16(buf[i]);
2027 byteswap_uint8_array(void *vbuf, size_t size)
2041 zio_compress_init();
2049 arc_fini(); /* arc depends on l2arc, so arc must go first */
2052 zio_compress_fini();