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, 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 (void) atomic_inc_32_nv(&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, 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.
534 * Note: The assumption is that we *know* these blocks will be needed
535 * almost immediately. Therefore, the prefetch i/os will be issued at
536 * ZIO_PRIORITY_SYNC_READ
538 * Note: indirect blocks and other metadata will be read synchronously,
539 * causing this function to block if they are not already cached.
542 dmu_prefetch(objset_t *os, uint64_t object, uint64_t offset, uint64_t len)
548 if (zfs_prefetch_disable)
551 if (len == 0) { /* they're interested in the bonus buffer */
552 dn = DMU_META_DNODE(os);
554 if (object == 0 || object >= DN_MAX_OBJECT)
557 rw_enter(&dn->dn_struct_rwlock, RW_READER);
558 blkid = dbuf_whichblock(dn, object * sizeof (dnode_phys_t));
559 dbuf_prefetch(dn, blkid, ZIO_PRIORITY_SYNC_READ);
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);
574 if (dn->dn_datablkshift) {
575 int blkshift = dn->dn_datablkshift;
576 nblks = (P2ROUNDUP(offset + len, 1 << blkshift) -
577 P2ALIGN(offset, 1 << blkshift)) >> blkshift;
579 nblks = (offset < dn->dn_datablksz);
583 blkid = dbuf_whichblock(dn, offset);
584 for (int i = 0; i < nblks; i++)
585 dbuf_prefetch(dn, blkid + i, ZIO_PRIORITY_SYNC_READ);
588 rw_exit(&dn->dn_struct_rwlock);
590 dnode_rele(dn, FTAG);
594 * Get the next "chunk" of file data to free. We traverse the file from
595 * the end so that the file gets shorter over time (if we crashes in the
596 * middle, this will leave us in a better state). We find allocated file
597 * data by simply searching the allocated level 1 indirects.
599 * On input, *start should be the first offset that does not need to be
600 * freed (e.g. "offset + length"). On return, *start will be the first
601 * offset that should be freed.
604 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum)
606 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
607 /* bytes of data covered by a level-1 indirect block */
609 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
611 ASSERT3U(minimum, <=, *start);
613 if (*start - minimum <= iblkrange * maxblks) {
617 ASSERT(ISP2(iblkrange));
619 for (uint64_t blks = 0; *start > minimum && blks < maxblks; blks++) {
623 * dnode_next_offset(BACKWARDS) will find an allocated L1
624 * indirect block at or before the input offset. We must
625 * decrement *start so that it is at the end of the region
629 err = dnode_next_offset(dn,
630 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
632 /* if there are no indirect blocks before start, we are done */
636 } else if (err != 0) {
640 /* set start to the beginning of this L1 indirect */
641 *start = P2ALIGN(*start, iblkrange);
643 if (*start < minimum)
649 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
652 uint64_t object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
655 if (offset >= object_size)
658 if (length == DMU_OBJECT_END || offset + length > object_size)
659 length = object_size - offset;
661 while (length != 0) {
662 uint64_t chunk_end, chunk_begin;
664 chunk_end = chunk_begin = offset + length;
666 /* move chunk_begin backwards to the beginning of this chunk */
667 err = get_next_chunk(dn, &chunk_begin, offset);
670 ASSERT3U(chunk_begin, >=, offset);
671 ASSERT3U(chunk_begin, <=, chunk_end);
673 dmu_tx_t *tx = dmu_tx_create(os);
674 dmu_tx_hold_free(tx, dn->dn_object,
675 chunk_begin, chunk_end - chunk_begin);
676 err = dmu_tx_assign(tx, TXG_WAIT);
681 dnode_free_range(dn, chunk_begin, chunk_end - chunk_begin, tx);
684 length -= chunk_end - chunk_begin;
690 dmu_free_long_range(objset_t *os, uint64_t object,
691 uint64_t offset, uint64_t length)
696 err = dnode_hold(os, object, FTAG, &dn);
699 err = dmu_free_long_range_impl(os, dn, offset, length);
702 * It is important to zero out the maxblkid when freeing the entire
703 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
704 * will take the fast path, and (b) dnode_reallocate() can verify
705 * that the entire file has been freed.
707 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
710 dnode_rele(dn, FTAG);
715 dmu_free_long_object(objset_t *os, uint64_t object)
720 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
724 tx = dmu_tx_create(os);
725 dmu_tx_hold_bonus(tx, object);
726 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
727 err = dmu_tx_assign(tx, TXG_WAIT);
729 err = dmu_object_free(os, object, tx);
739 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
740 uint64_t size, dmu_tx_t *tx)
743 int err = dnode_hold(os, object, FTAG, &dn);
746 ASSERT(offset < UINT64_MAX);
747 ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
748 dnode_free_range(dn, offset, size, tx);
749 dnode_rele(dn, FTAG);
754 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
755 void *buf, uint32_t flags)
761 err = dnode_hold(os, object, FTAG, &dn);
766 * Deal with odd block sizes, where there can't be data past the first
767 * block. If we ever do the tail block optimization, we will need to
768 * handle that here as well.
770 if (dn->dn_maxblkid == 0) {
771 int newsz = offset > dn->dn_datablksz ? 0 :
772 MIN(size, dn->dn_datablksz - offset);
773 bzero((char *)buf + newsz, size - newsz);
778 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
782 * NB: we could do this block-at-a-time, but it's nice
783 * to be reading in parallel.
785 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
786 TRUE, FTAG, &numbufs, &dbp, flags);
790 for (i = 0; i < numbufs; i++) {
793 dmu_buf_t *db = dbp[i];
797 bufoff = offset - db->db_offset;
798 tocpy = (int)MIN(db->db_size - bufoff, size);
800 bcopy((char *)db->db_data + bufoff, buf, tocpy);
804 buf = (char *)buf + tocpy;
806 dmu_buf_rele_array(dbp, numbufs, FTAG);
808 dnode_rele(dn, FTAG);
813 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
814 const void *buf, dmu_tx_t *tx)
822 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
823 FALSE, FTAG, &numbufs, &dbp));
825 for (i = 0; i < numbufs; i++) {
828 dmu_buf_t *db = dbp[i];
832 bufoff = offset - db->db_offset;
833 tocpy = (int)MIN(db->db_size - bufoff, size);
835 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
837 if (tocpy == db->db_size)
838 dmu_buf_will_fill(db, tx);
840 dmu_buf_will_dirty(db, tx);
842 bcopy(buf, (char *)db->db_data + bufoff, tocpy);
844 if (tocpy == db->db_size)
845 dmu_buf_fill_done(db, tx);
849 buf = (char *)buf + tocpy;
851 dmu_buf_rele_array(dbp, numbufs, FTAG);
855 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
864 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
865 FALSE, FTAG, &numbufs, &dbp));
867 for (i = 0; i < numbufs; i++) {
868 dmu_buf_t *db = dbp[i];
870 dmu_buf_will_not_fill(db, tx);
872 dmu_buf_rele_array(dbp, numbufs, FTAG);
876 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
877 void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
878 int compressed_size, int byteorder, dmu_tx_t *tx)
882 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
883 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
884 VERIFY0(dmu_buf_hold_noread(os, object, offset,
887 dmu_buf_write_embedded(db,
888 data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
889 uncompressed_size, compressed_size, byteorder, tx);
891 dmu_buf_rele(db, FTAG);
895 * DMU support for xuio
897 kstat_t *xuio_ksp = NULL;
900 dmu_xuio_init(xuio_t *xuio, int nblk)
903 uio_t *uio = &xuio->xu_uio;
905 uio->uio_iovcnt = nblk;
906 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
908 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
910 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
911 priv->iovp = uio->uio_iov;
912 XUIO_XUZC_PRIV(xuio) = priv;
914 if (XUIO_XUZC_RW(xuio) == UIO_READ)
915 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
917 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
923 dmu_xuio_fini(xuio_t *xuio)
925 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
926 int nblk = priv->cnt;
928 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
929 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
930 kmem_free(priv, sizeof (dmu_xuio_t));
932 if (XUIO_XUZC_RW(xuio) == UIO_READ)
933 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
935 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
939 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
940 * and increase priv->next by 1.
943 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
946 uio_t *uio = &xuio->xu_uio;
947 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
948 int i = priv->next++;
950 ASSERT(i < priv->cnt);
951 ASSERT(off + n <= arc_buf_size(abuf));
952 iov = uio->uio_iov + i;
953 iov->iov_base = (char *)abuf->b_data + off;
955 priv->bufs[i] = abuf;
960 dmu_xuio_cnt(xuio_t *xuio)
962 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
967 dmu_xuio_arcbuf(xuio_t *xuio, int i)
969 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
971 ASSERT(i < priv->cnt);
972 return (priv->bufs[i]);
976 dmu_xuio_clear(xuio_t *xuio, int i)
978 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
980 ASSERT(i < priv->cnt);
981 priv->bufs[i] = NULL;
987 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
988 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
990 if (xuio_ksp != NULL) {
991 xuio_ksp->ks_data = &xuio_stats;
992 kstat_install(xuio_ksp);
999 if (xuio_ksp != NULL) {
1000 kstat_delete(xuio_ksp);
1006 xuio_stat_wbuf_copied()
1008 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1012 xuio_stat_wbuf_nocopy()
1014 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
1019 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
1022 int numbufs, i, err;
1023 xuio_t *xuio = NULL;
1026 * NB: we could do this block-at-a-time, but it's nice
1027 * to be reading in parallel.
1029 err = dmu_buf_hold_array(os, object, uio->uio_loffset, size, TRUE, FTAG,
1035 if (uio->uio_extflg == UIO_XUIO)
1036 xuio = (xuio_t *)uio;
1039 for (i = 0; i < numbufs; i++) {
1042 dmu_buf_t *db = dbp[i];
1046 bufoff = uio->uio_loffset - db->db_offset;
1047 tocpy = (int)MIN(db->db_size - bufoff, size);
1050 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
1051 arc_buf_t *dbuf_abuf = dbi->db_buf;
1052 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
1053 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
1055 uio->uio_resid -= tocpy;
1056 uio->uio_loffset += tocpy;
1059 if (abuf == dbuf_abuf)
1060 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
1062 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
1064 err = uiomove((char *)db->db_data + bufoff, tocpy,
1072 dmu_buf_rele_array(dbp, numbufs, FTAG);
1078 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1085 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1086 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1090 for (i = 0; i < numbufs; i++) {
1093 dmu_buf_t *db = dbp[i];
1097 bufoff = uio->uio_loffset - db->db_offset;
1098 tocpy = (int)MIN(db->db_size - bufoff, size);
1100 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1102 if (tocpy == db->db_size)
1103 dmu_buf_will_fill(db, tx);
1105 dmu_buf_will_dirty(db, tx);
1108 * XXX uiomove could block forever (eg. nfs-backed
1109 * pages). There needs to be a uiolockdown() function
1110 * to lock the pages in memory, so that uiomove won't
1113 err = uiomove((char *)db->db_data + bufoff, tocpy,
1116 if (tocpy == db->db_size)
1117 dmu_buf_fill_done(db, tx);
1125 dmu_buf_rele_array(dbp, numbufs, FTAG);
1130 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1133 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1142 err = dmu_write_uio_dnode(dn, uio, size, tx);
1149 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1158 err = dnode_hold(os, object, FTAG, &dn);
1162 err = dmu_write_uio_dnode(dn, uio, size, tx);
1164 dnode_rele(dn, FTAG);
1171 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1172 page_t *pp, dmu_tx_t *tx)
1181 err = dmu_buf_hold_array(os, object, offset, size,
1182 FALSE, FTAG, &numbufs, &dbp);
1186 for (i = 0; i < numbufs; i++) {
1187 int tocpy, copied, thiscpy;
1189 dmu_buf_t *db = dbp[i];
1193 ASSERT3U(db->db_size, >=, PAGESIZE);
1195 bufoff = offset - db->db_offset;
1196 tocpy = (int)MIN(db->db_size - bufoff, size);
1198 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1200 if (tocpy == db->db_size)
1201 dmu_buf_will_fill(db, tx);
1203 dmu_buf_will_dirty(db, tx);
1205 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1206 ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff);
1207 thiscpy = MIN(PAGESIZE, tocpy - copied);
1208 va = zfs_map_page(pp, S_READ);
1209 bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1210 zfs_unmap_page(pp, va);
1215 if (tocpy == db->db_size)
1216 dmu_buf_fill_done(db, tx);
1221 dmu_buf_rele_array(dbp, numbufs, FTAG);
1228 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1229 vm_page_t *ma, dmu_tx_t *tx)
1239 err = dmu_buf_hold_array(os, object, offset, size,
1240 FALSE, FTAG, &numbufs, &dbp);
1244 for (i = 0; i < numbufs; i++) {
1245 int tocpy, copied, thiscpy;
1247 dmu_buf_t *db = dbp[i];
1251 ASSERT3U(db->db_size, >=, PAGESIZE);
1253 bufoff = offset - db->db_offset;
1254 tocpy = (int)MIN(db->db_size - bufoff, size);
1256 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1258 if (tocpy == db->db_size)
1259 dmu_buf_will_fill(db, tx);
1261 dmu_buf_will_dirty(db, tx);
1263 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1264 ASSERT3U(ptoa((*ma)->pindex), ==, db->db_offset + bufoff);
1265 thiscpy = MIN(PAGESIZE, tocpy - copied);
1266 va = zfs_map_page(*ma, &sf);
1267 bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1273 if (tocpy == db->db_size)
1274 dmu_buf_fill_done(db, tx);
1279 dmu_buf_rele_array(dbp, numbufs, FTAG);
1286 * Allocate a loaned anonymous arc buffer.
1289 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1291 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1293 return (arc_loan_buf(db->db_objset->os_spa, size));
1297 * Free a loaned arc buffer.
1300 dmu_return_arcbuf(arc_buf_t *buf)
1302 arc_return_buf(buf, FTAG);
1303 VERIFY(arc_buf_remove_ref(buf, FTAG));
1307 * When possible directly assign passed loaned arc buffer to a dbuf.
1308 * If this is not possible copy the contents of passed arc buf via
1312 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1315 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1318 uint32_t blksz = (uint32_t)arc_buf_size(buf);
1321 DB_DNODE_ENTER(dbuf);
1322 dn = DB_DNODE(dbuf);
1323 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1324 blkid = dbuf_whichblock(dn, offset);
1325 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1326 rw_exit(&dn->dn_struct_rwlock);
1327 DB_DNODE_EXIT(dbuf);
1329 if (offset == db->db.db_offset && blksz == db->db.db_size) {
1330 dbuf_assign_arcbuf(db, buf, tx);
1331 dbuf_rele(db, FTAG);
1336 DB_DNODE_ENTER(dbuf);
1337 dn = DB_DNODE(dbuf);
1339 object = dn->dn_object;
1340 DB_DNODE_EXIT(dbuf);
1342 dbuf_rele(db, FTAG);
1343 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1344 dmu_return_arcbuf(buf);
1345 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1350 dbuf_dirty_record_t *dsa_dr;
1351 dmu_sync_cb_t *dsa_done;
1358 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1360 dmu_sync_arg_t *dsa = varg;
1361 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1362 blkptr_t *bp = zio->io_bp;
1364 if (zio->io_error == 0) {
1365 if (BP_IS_HOLE(bp)) {
1367 * A block of zeros may compress to a hole, but the
1368 * block size still needs to be known for replay.
1370 BP_SET_LSIZE(bp, db->db_size);
1371 } else if (!BP_IS_EMBEDDED(bp)) {
1372 ASSERT(BP_GET_LEVEL(bp) == 0);
1379 dmu_sync_late_arrival_ready(zio_t *zio)
1381 dmu_sync_ready(zio, NULL, zio->io_private);
1386 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1388 dmu_sync_arg_t *dsa = varg;
1389 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1390 dmu_buf_impl_t *db = dr->dr_dbuf;
1392 mutex_enter(&db->db_mtx);
1393 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1394 if (zio->io_error == 0) {
1395 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1396 if (dr->dt.dl.dr_nopwrite) {
1397 blkptr_t *bp = zio->io_bp;
1398 blkptr_t *bp_orig = &zio->io_bp_orig;
1399 uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1401 ASSERT(BP_EQUAL(bp, bp_orig));
1402 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1403 ASSERT(zio_checksum_table[chksum].ci_dedup);
1405 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1406 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1407 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1408 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by))
1409 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1411 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1413 cv_broadcast(&db->db_changed);
1414 mutex_exit(&db->db_mtx);
1416 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1418 kmem_free(dsa, sizeof (*dsa));
1422 dmu_sync_late_arrival_done(zio_t *zio)
1424 blkptr_t *bp = zio->io_bp;
1425 dmu_sync_arg_t *dsa = zio->io_private;
1426 blkptr_t *bp_orig = &zio->io_bp_orig;
1428 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1430 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE)
1431 * then there is nothing to do here. Otherwise, free the
1432 * newly allocated block in this txg.
1434 if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
1435 ASSERT(BP_EQUAL(bp, bp_orig));
1437 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1438 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1439 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1440 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1444 dmu_tx_commit(dsa->dsa_tx);
1446 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1448 kmem_free(dsa, sizeof (*dsa));
1452 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1453 zio_prop_t *zp, zbookmark_phys_t *zb)
1455 dmu_sync_arg_t *dsa;
1458 tx = dmu_tx_create(os);
1459 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1460 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1462 /* Make zl_get_data do txg_waited_synced() */
1463 return (SET_ERROR(EIO));
1466 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1468 dsa->dsa_done = done;
1472 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1473 zgd->zgd_db->db_data, zgd->zgd_db->db_size, zp,
1474 dmu_sync_late_arrival_ready, NULL, dmu_sync_late_arrival_done, dsa,
1475 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1481 * Intent log support: sync the block associated with db to disk.
1482 * N.B. and XXX: the caller is responsible for making sure that the
1483 * data isn't changing while dmu_sync() is writing it.
1487 * EEXIST: this txg has already been synced, so there's nothing to do.
1488 * The caller should not log the write.
1490 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1491 * The caller should not log the write.
1493 * EALREADY: this block is already in the process of being synced.
1494 * The caller should track its progress (somehow).
1496 * EIO: could not do the I/O.
1497 * The caller should do a txg_wait_synced().
1499 * 0: the I/O has been initiated.
1500 * The caller should log this blkptr in the done callback.
1501 * It is possible that the I/O will fail, in which case
1502 * the error will be reported to the done callback and
1503 * propagated to pio from zio_done().
1506 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1508 blkptr_t *bp = zgd->zgd_bp;
1509 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1510 objset_t *os = db->db_objset;
1511 dsl_dataset_t *ds = os->os_dsl_dataset;
1512 dbuf_dirty_record_t *dr;
1513 dmu_sync_arg_t *dsa;
1514 zbookmark_phys_t zb;
1518 ASSERT(pio != NULL);
1521 SET_BOOKMARK(&zb, ds->ds_object,
1522 db->db.db_object, db->db_level, db->db_blkid);
1526 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1530 * If we're frozen (running ziltest), we always need to generate a bp.
1532 if (txg > spa_freeze_txg(os->os_spa))
1533 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1536 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1537 * and us. If we determine that this txg is not yet syncing,
1538 * but it begins to sync a moment later, that's OK because the
1539 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1541 mutex_enter(&db->db_mtx);
1543 if (txg <= spa_last_synced_txg(os->os_spa)) {
1545 * This txg has already synced. There's nothing to do.
1547 mutex_exit(&db->db_mtx);
1548 return (SET_ERROR(EEXIST));
1551 if (txg <= spa_syncing_txg(os->os_spa)) {
1553 * This txg is currently syncing, so we can't mess with
1554 * the dirty record anymore; just write a new log block.
1556 mutex_exit(&db->db_mtx);
1557 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1560 dr = db->db_last_dirty;
1561 while (dr && dr->dr_txg != txg)
1566 * There's no dr for this dbuf, so it must have been freed.
1567 * There's no need to log writes to freed blocks, so we're done.
1569 mutex_exit(&db->db_mtx);
1570 return (SET_ERROR(ENOENT));
1573 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
1576 * Assume the on-disk data is X, the current syncing data is Y,
1577 * and the current in-memory data is Z (currently in dmu_sync).
1578 * X and Z are identical but Y is has been modified. Normally,
1579 * when X and Z are the same we will perform a nopwrite but if Y
1580 * is different we must disable nopwrite since the resulting write
1581 * of Y to disk can free the block containing X. If we allowed a
1582 * nopwrite to occur the block pointing to Z would reference a freed
1583 * block. Since this is a rare case we simplify this by disabling
1584 * nopwrite if the current dmu_sync-ing dbuf has been modified in
1585 * a previous transaction.
1588 zp.zp_nopwrite = B_FALSE;
1590 ASSERT(dr->dr_txg == txg);
1591 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1592 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1594 * We have already issued a sync write for this buffer,
1595 * or this buffer has already been synced. It could not
1596 * have been dirtied since, or we would have cleared the state.
1598 mutex_exit(&db->db_mtx);
1599 return (SET_ERROR(EALREADY));
1602 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1603 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1604 mutex_exit(&db->db_mtx);
1606 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1608 dsa->dsa_done = done;
1612 zio_nowait(arc_write(pio, os->os_spa, txg,
1613 bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
1614 DBUF_IS_L2COMPRESSIBLE(db), &zp, dmu_sync_ready,
1615 NULL, dmu_sync_done, dsa, ZIO_PRIORITY_SYNC_WRITE,
1616 ZIO_FLAG_CANFAIL, &zb));
1622 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1628 err = dnode_hold(os, object, FTAG, &dn);
1631 err = dnode_set_blksz(dn, size, ibs, tx);
1632 dnode_rele(dn, FTAG);
1637 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1643 * Send streams include each object's checksum function. This
1644 * check ensures that the receiving system can understand the
1645 * checksum function transmitted.
1647 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS);
1649 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1650 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS);
1651 dn->dn_checksum = checksum;
1652 dnode_setdirty(dn, tx);
1653 dnode_rele(dn, FTAG);
1657 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1663 * Send streams include each object's compression function. This
1664 * check ensures that the receiving system can understand the
1665 * compression function transmitted.
1667 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS);
1669 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1670 dn->dn_compress = compress;
1671 dnode_setdirty(dn, tx);
1672 dnode_rele(dn, FTAG);
1675 int zfs_mdcomp_disable = 0;
1676 TUNABLE_INT("vfs.zfs.mdcomp_disable", &zfs_mdcomp_disable);
1677 SYSCTL_INT(_vfs_zfs, OID_AUTO, mdcomp_disable, CTLFLAG_RW,
1678 &zfs_mdcomp_disable, 0, "Disable metadata compression");
1681 * When the "redundant_metadata" property is set to "most", only indirect
1682 * blocks of this level and higher will have an additional ditto block.
1684 int zfs_redundant_metadata_most_ditto_level = 2;
1687 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
1689 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
1690 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
1692 enum zio_checksum checksum = os->os_checksum;
1693 enum zio_compress compress = os->os_compress;
1694 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
1695 boolean_t dedup = B_FALSE;
1696 boolean_t nopwrite = B_FALSE;
1697 boolean_t dedup_verify = os->os_dedup_verify;
1698 int copies = os->os_copies;
1701 * We maintain different write policies for each of the following
1704 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1705 * 3. all other level 0 blocks
1709 * XXX -- we should design a compression algorithm
1710 * that specializes in arrays of bps.
1712 boolean_t lz4_ac = spa_feature_is_active(os->os_spa,
1713 SPA_FEATURE_LZ4_COMPRESS);
1715 if (zfs_mdcomp_disable) {
1716 compress = ZIO_COMPRESS_EMPTY;
1717 } else if (lz4_ac) {
1718 compress = ZIO_COMPRESS_LZ4;
1720 compress = ZIO_COMPRESS_LZJB;
1724 * Metadata always gets checksummed. If the data
1725 * checksum is multi-bit correctable, and it's not a
1726 * ZBT-style checksum, then it's suitable for metadata
1727 * as well. Otherwise, the metadata checksum defaults
1730 if (zio_checksum_table[checksum].ci_correctable < 1 ||
1731 zio_checksum_table[checksum].ci_eck)
1732 checksum = ZIO_CHECKSUM_FLETCHER_4;
1734 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL ||
1735 (os->os_redundant_metadata ==
1736 ZFS_REDUNDANT_METADATA_MOST &&
1737 (level >= zfs_redundant_metadata_most_ditto_level ||
1738 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))))
1740 } else if (wp & WP_NOFILL) {
1744 * If we're writing preallocated blocks, we aren't actually
1745 * writing them so don't set any policy properties. These
1746 * blocks are currently only used by an external subsystem
1747 * outside of zfs (i.e. dump) and not written by the zio
1750 compress = ZIO_COMPRESS_OFF;
1751 checksum = ZIO_CHECKSUM_NOPARITY;
1753 compress = zio_compress_select(dn->dn_compress, compress);
1755 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
1756 zio_checksum_select(dn->dn_checksum, checksum) :
1760 * Determine dedup setting. If we are in dmu_sync(),
1761 * we won't actually dedup now because that's all
1762 * done in syncing context; but we do want to use the
1763 * dedup checkum. If the checksum is not strong
1764 * enough to ensure unique signatures, force
1767 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
1768 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
1769 if (!zio_checksum_table[checksum].ci_dedup)
1770 dedup_verify = B_TRUE;
1774 * Enable nopwrite if we have a cryptographically secure
1775 * checksum that has no known collisions (i.e. SHA-256)
1776 * and compression is enabled. We don't enable nopwrite if
1777 * dedup is enabled as the two features are mutually exclusive.
1779 nopwrite = (!dedup && zio_checksum_table[checksum].ci_dedup &&
1780 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
1783 zp->zp_checksum = checksum;
1784 zp->zp_compress = compress;
1785 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
1786 zp->zp_level = level;
1787 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
1788 zp->zp_dedup = dedup;
1789 zp->zp_dedup_verify = dedup && dedup_verify;
1790 zp->zp_nopwrite = nopwrite;
1794 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
1799 err = dnode_hold(os, object, FTAG, &dn);
1803 * Sync any current changes before
1804 * we go trundling through the block pointers.
1806 for (i = 0; i < TXG_SIZE; i++) {
1807 if (list_link_active(&dn->dn_dirty_link[i]))
1810 if (i != TXG_SIZE) {
1811 dnode_rele(dn, FTAG);
1812 txg_wait_synced(dmu_objset_pool(os), 0);
1813 err = dnode_hold(os, object, FTAG, &dn);
1818 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
1819 dnode_rele(dn, FTAG);
1825 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
1829 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1830 mutex_enter(&dn->dn_mtx);
1834 doi->doi_data_block_size = dn->dn_datablksz;
1835 doi->doi_metadata_block_size = dn->dn_indblkshift ?
1836 1ULL << dn->dn_indblkshift : 0;
1837 doi->doi_type = dn->dn_type;
1838 doi->doi_bonus_type = dn->dn_bonustype;
1839 doi->doi_bonus_size = dn->dn_bonuslen;
1840 doi->doi_indirection = dn->dn_nlevels;
1841 doi->doi_checksum = dn->dn_checksum;
1842 doi->doi_compress = dn->dn_compress;
1843 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
1844 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
1845 doi->doi_fill_count = 0;
1846 for (int i = 0; i < dnp->dn_nblkptr; i++)
1847 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
1849 mutex_exit(&dn->dn_mtx);
1850 rw_exit(&dn->dn_struct_rwlock);
1854 * Get information on a DMU object.
1855 * If doi is NULL, just indicates whether the object exists.
1858 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
1861 int err = dnode_hold(os, object, FTAG, &dn);
1867 dmu_object_info_from_dnode(dn, doi);
1869 dnode_rele(dn, FTAG);
1874 * As above, but faster; can be used when you have a held dbuf in hand.
1877 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
1879 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1882 dmu_object_info_from_dnode(DB_DNODE(db), doi);
1887 * Faster still when you only care about the size.
1888 * This is specifically optimized for zfs_getattr().
1891 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
1892 u_longlong_t *nblk512)
1894 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1900 *blksize = dn->dn_datablksz;
1901 /* add 1 for dnode space */
1902 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
1903 SPA_MINBLOCKSHIFT) + 1;
1908 byteswap_uint64_array(void *vbuf, size_t size)
1910 uint64_t *buf = vbuf;
1911 size_t count = size >> 3;
1914 ASSERT((size & 7) == 0);
1916 for (i = 0; i < count; i++)
1917 buf[i] = BSWAP_64(buf[i]);
1921 byteswap_uint32_array(void *vbuf, size_t size)
1923 uint32_t *buf = vbuf;
1924 size_t count = size >> 2;
1927 ASSERT((size & 3) == 0);
1929 for (i = 0; i < count; i++)
1930 buf[i] = BSWAP_32(buf[i]);
1934 byteswap_uint16_array(void *vbuf, size_t size)
1936 uint16_t *buf = vbuf;
1937 size_t count = size >> 1;
1940 ASSERT((size & 1) == 0);
1942 for (i = 0; i < count; i++)
1943 buf[i] = BSWAP_16(buf[i]);
1948 byteswap_uint8_array(void *vbuf, size_t size)
1962 zio_compress_init();
1970 arc_fini(); /* arc depends on l2arc, so arc must go first */
1973 zio_compress_fini();