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) 2012, 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. */
29 #include <sys/dmu_impl.h>
30 #include <sys/dmu_tx.h>
32 #include <sys/dnode.h>
33 #include <sys/zfs_context.h>
34 #include <sys/dmu_objset.h>
35 #include <sys/dmu_traverse.h>
36 #include <sys/dsl_dataset.h>
37 #include <sys/dsl_dir.h>
38 #include <sys/dsl_pool.h>
39 #include <sys/dsl_synctask.h>
40 #include <sys/dsl_prop.h>
41 #include <sys/dmu_zfetch.h>
42 #include <sys/zfs_ioctl.h>
44 #include <sys/zio_checksum.h>
45 #include <sys/zio_compress.h>
49 #include <sys/zfs_znode.h>
53 * Enable/disable nopwrite feature.
55 int zfs_nopwrite_enabled = 1;
56 SYSCTL_DECL(_vfs_zfs);
57 TUNABLE_INT("vfs.zfs.nopwrite_enabled", &zfs_nopwrite_enabled);
58 SYSCTL_INT(_vfs_zfs, OID_AUTO, nopwrite_enabled, CTLFLAG_RDTUN,
59 &zfs_nopwrite_enabled, 0, "Enable nopwrite feature");
61 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
62 { DMU_BSWAP_UINT8, TRUE, "unallocated" },
63 { DMU_BSWAP_ZAP, TRUE, "object directory" },
64 { DMU_BSWAP_UINT64, TRUE, "object array" },
65 { DMU_BSWAP_UINT8, TRUE, "packed nvlist" },
66 { DMU_BSWAP_UINT64, TRUE, "packed nvlist size" },
67 { DMU_BSWAP_UINT64, TRUE, "bpobj" },
68 { DMU_BSWAP_UINT64, TRUE, "bpobj header" },
69 { DMU_BSWAP_UINT64, TRUE, "SPA space map header" },
70 { DMU_BSWAP_UINT64, TRUE, "SPA space map" },
71 { DMU_BSWAP_UINT64, TRUE, "ZIL intent log" },
72 { DMU_BSWAP_DNODE, TRUE, "DMU dnode" },
73 { DMU_BSWAP_OBJSET, TRUE, "DMU objset" },
74 { DMU_BSWAP_UINT64, TRUE, "DSL directory" },
75 { DMU_BSWAP_ZAP, TRUE, "DSL directory child map"},
76 { DMU_BSWAP_ZAP, TRUE, "DSL dataset snap map" },
77 { DMU_BSWAP_ZAP, TRUE, "DSL props" },
78 { DMU_BSWAP_UINT64, TRUE, "DSL dataset" },
79 { DMU_BSWAP_ZNODE, TRUE, "ZFS znode" },
80 { DMU_BSWAP_OLDACL, TRUE, "ZFS V0 ACL" },
81 { DMU_BSWAP_UINT8, FALSE, "ZFS plain file" },
82 { DMU_BSWAP_ZAP, TRUE, "ZFS directory" },
83 { DMU_BSWAP_ZAP, TRUE, "ZFS master node" },
84 { DMU_BSWAP_ZAP, TRUE, "ZFS delete queue" },
85 { DMU_BSWAP_UINT8, FALSE, "zvol object" },
86 { DMU_BSWAP_ZAP, TRUE, "zvol prop" },
87 { DMU_BSWAP_UINT8, FALSE, "other uint8[]" },
88 { DMU_BSWAP_UINT64, FALSE, "other uint64[]" },
89 { DMU_BSWAP_ZAP, TRUE, "other ZAP" },
90 { DMU_BSWAP_ZAP, TRUE, "persistent error log" },
91 { DMU_BSWAP_UINT8, TRUE, "SPA history" },
92 { DMU_BSWAP_UINT64, TRUE, "SPA history offsets" },
93 { DMU_BSWAP_ZAP, TRUE, "Pool properties" },
94 { DMU_BSWAP_ZAP, TRUE, "DSL permissions" },
95 { DMU_BSWAP_ACL, TRUE, "ZFS ACL" },
96 { DMU_BSWAP_UINT8, TRUE, "ZFS SYSACL" },
97 { DMU_BSWAP_UINT8, TRUE, "FUID table" },
98 { DMU_BSWAP_UINT64, TRUE, "FUID table size" },
99 { DMU_BSWAP_ZAP, TRUE, "DSL dataset next clones"},
100 { DMU_BSWAP_ZAP, TRUE, "scan work queue" },
101 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group used" },
102 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group quota" },
103 { DMU_BSWAP_ZAP, TRUE, "snapshot refcount tags"},
104 { DMU_BSWAP_ZAP, TRUE, "DDT ZAP algorithm" },
105 { DMU_BSWAP_ZAP, TRUE, "DDT statistics" },
106 { DMU_BSWAP_UINT8, TRUE, "System attributes" },
107 { DMU_BSWAP_ZAP, TRUE, "SA master node" },
108 { DMU_BSWAP_ZAP, TRUE, "SA attr registration" },
109 { DMU_BSWAP_ZAP, TRUE, "SA attr layouts" },
110 { DMU_BSWAP_ZAP, TRUE, "scan translations" },
111 { DMU_BSWAP_UINT8, FALSE, "deduplicated block" },
112 { DMU_BSWAP_ZAP, TRUE, "DSL deadlist map" },
113 { DMU_BSWAP_UINT64, TRUE, "DSL deadlist map hdr" },
114 { DMU_BSWAP_ZAP, TRUE, "DSL dir clones" },
115 { DMU_BSWAP_UINT64, TRUE, "bpobj subobj" }
118 const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
119 { byteswap_uint8_array, "uint8" },
120 { byteswap_uint16_array, "uint16" },
121 { byteswap_uint32_array, "uint32" },
122 { byteswap_uint64_array, "uint64" },
123 { zap_byteswap, "zap" },
124 { dnode_buf_byteswap, "dnode" },
125 { dmu_objset_byteswap, "objset" },
126 { zfs_znode_byteswap, "znode" },
127 { zfs_oldacl_byteswap, "oldacl" },
128 { zfs_acl_byteswap, "acl" }
132 dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset,
133 void *tag, dmu_buf_t **dbp)
140 err = dnode_hold(os, object, FTAG, &dn);
143 blkid = dbuf_whichblock(dn, offset);
144 rw_enter(&dn->dn_struct_rwlock, RW_READER);
145 db = dbuf_hold(dn, blkid, tag);
146 rw_exit(&dn->dn_struct_rwlock);
147 dnode_rele(dn, FTAG);
151 return (SET_ERROR(EIO));
159 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
160 void *tag, dmu_buf_t **dbp, int flags)
163 int db_flags = DB_RF_CANFAIL;
165 if (flags & DMU_READ_NO_PREFETCH)
166 db_flags |= DB_RF_NOPREFETCH;
168 err = dmu_buf_hold_noread(os, object, offset, tag, dbp);
170 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
171 err = dbuf_read(db, NULL, db_flags);
184 return (DN_MAX_BONUSLEN);
188 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
190 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
197 if (dn->dn_bonus != db) {
198 error = SET_ERROR(EINVAL);
199 } else if (newsize < 0 || newsize > db_fake->db_size) {
200 error = SET_ERROR(EINVAL);
202 dnode_setbonuslen(dn, newsize, tx);
211 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
213 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
220 if (!DMU_OT_IS_VALID(type)) {
221 error = SET_ERROR(EINVAL);
222 } else if (dn->dn_bonus != db) {
223 error = SET_ERROR(EINVAL);
225 dnode_setbonus_type(dn, type, tx);
234 dmu_get_bonustype(dmu_buf_t *db_fake)
236 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
238 dmu_object_type_t type;
242 type = dn->dn_bonustype;
249 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
254 error = dnode_hold(os, object, FTAG, &dn);
255 dbuf_rm_spill(dn, tx);
256 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
257 dnode_rm_spill(dn, tx);
258 rw_exit(&dn->dn_struct_rwlock);
259 dnode_rele(dn, FTAG);
264 * returns ENOENT, EIO, or 0.
267 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
273 error = dnode_hold(os, object, FTAG, &dn);
277 rw_enter(&dn->dn_struct_rwlock, RW_READER);
278 if (dn->dn_bonus == NULL) {
279 rw_exit(&dn->dn_struct_rwlock);
280 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
281 if (dn->dn_bonus == NULL)
282 dbuf_create_bonus(dn);
286 /* as long as the bonus buf is held, the dnode will be held */
287 if (refcount_add(&db->db_holds, tag) == 1) {
288 VERIFY(dnode_add_ref(dn, db));
289 (void) atomic_inc_32_nv(&dn->dn_dbufs_count);
293 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
294 * hold and incrementing the dbuf count to ensure that dnode_move() sees
295 * a dnode hold for every dbuf.
297 rw_exit(&dn->dn_struct_rwlock);
299 dnode_rele(dn, FTAG);
301 VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH));
308 * returns ENOENT, EIO, or 0.
310 * This interface will allocate a blank spill dbuf when a spill blk
311 * doesn't already exist on the dnode.
313 * if you only want to find an already existing spill db, then
314 * dmu_spill_hold_existing() should be used.
317 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
319 dmu_buf_impl_t *db = NULL;
322 if ((flags & DB_RF_HAVESTRUCT) == 0)
323 rw_enter(&dn->dn_struct_rwlock, RW_READER);
325 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
327 if ((flags & DB_RF_HAVESTRUCT) == 0)
328 rw_exit(&dn->dn_struct_rwlock);
331 err = dbuf_read(db, NULL, flags);
340 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
342 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
349 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
350 err = SET_ERROR(EINVAL);
352 rw_enter(&dn->dn_struct_rwlock, RW_READER);
354 if (!dn->dn_have_spill) {
355 err = SET_ERROR(ENOENT);
357 err = dmu_spill_hold_by_dnode(dn,
358 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
361 rw_exit(&dn->dn_struct_rwlock);
369 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
371 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
377 err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp);
384 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
385 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
386 * and can induce severe lock contention when writing to several files
387 * whose dnodes are in the same block.
390 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
391 int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
394 uint64_t blkid, nblks, i;
399 ASSERT(length <= DMU_MAX_ACCESS);
401 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT;
402 if (flags & DMU_READ_NO_PREFETCH || length > zfetch_array_rd_sz)
403 dbuf_flags |= DB_RF_NOPREFETCH;
405 rw_enter(&dn->dn_struct_rwlock, RW_READER);
406 if (dn->dn_datablkshift) {
407 int blkshift = dn->dn_datablkshift;
408 nblks = (P2ROUNDUP(offset+length, 1ULL<<blkshift) -
409 P2ALIGN(offset, 1ULL<<blkshift)) >> blkshift;
411 if (offset + length > dn->dn_datablksz) {
412 zfs_panic_recover("zfs: accessing past end of object "
413 "%llx/%llx (size=%u access=%llu+%llu)",
414 (longlong_t)dn->dn_objset->
415 os_dsl_dataset->ds_object,
416 (longlong_t)dn->dn_object, dn->dn_datablksz,
417 (longlong_t)offset, (longlong_t)length);
418 rw_exit(&dn->dn_struct_rwlock);
419 return (SET_ERROR(EIO));
423 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
425 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
426 blkid = dbuf_whichblock(dn, offset);
427 for (i = 0; i < nblks; i++) {
428 dmu_buf_impl_t *db = dbuf_hold(dn, blkid+i, tag);
430 rw_exit(&dn->dn_struct_rwlock);
431 dmu_buf_rele_array(dbp, nblks, tag);
433 return (SET_ERROR(EIO));
435 /* initiate async i/o */
437 (void) dbuf_read(db, zio, dbuf_flags);
440 curthread->td_ru.ru_oublock++;
444 rw_exit(&dn->dn_struct_rwlock);
446 /* wait for async i/o */
449 dmu_buf_rele_array(dbp, nblks, tag);
453 /* wait for other io to complete */
455 for (i = 0; i < nblks; i++) {
456 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
457 mutex_enter(&db->db_mtx);
458 while (db->db_state == DB_READ ||
459 db->db_state == DB_FILL)
460 cv_wait(&db->db_changed, &db->db_mtx);
461 if (db->db_state == DB_UNCACHED)
462 err = SET_ERROR(EIO);
463 mutex_exit(&db->db_mtx);
465 dmu_buf_rele_array(dbp, nblks, tag);
477 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
478 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
483 err = dnode_hold(os, object, FTAG, &dn);
487 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
488 numbufsp, dbpp, DMU_READ_PREFETCH);
490 dnode_rele(dn, FTAG);
496 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
497 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
499 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
505 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
506 numbufsp, dbpp, DMU_READ_PREFETCH);
513 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
516 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
521 for (i = 0; i < numbufs; i++) {
523 dbuf_rele(dbp[i], tag);
526 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
530 * Issue prefetch i/os for the given blocks.
532 * Note: The assumption is that we *know* these blocks will be needed
533 * almost immediately. Therefore, the prefetch i/os will be issued at
534 * ZIO_PRIORITY_SYNC_READ
536 * Note: indirect blocks and other metadata will be read synchronously,
537 * causing this function to block if they are not already cached.
540 dmu_prefetch(objset_t *os, uint64_t object, uint64_t offset, uint64_t len)
546 if (zfs_prefetch_disable)
549 if (len == 0) { /* they're interested in the bonus buffer */
550 dn = DMU_META_DNODE(os);
552 if (object == 0 || object >= DN_MAX_OBJECT)
555 rw_enter(&dn->dn_struct_rwlock, RW_READER);
556 blkid = dbuf_whichblock(dn, object * sizeof (dnode_phys_t));
557 dbuf_prefetch(dn, blkid, ZIO_PRIORITY_SYNC_READ);
558 rw_exit(&dn->dn_struct_rwlock);
563 * XXX - Note, if the dnode for the requested object is not
564 * already cached, we will do a *synchronous* read in the
565 * dnode_hold() call. The same is true for any indirects.
567 err = dnode_hold(os, object, FTAG, &dn);
571 rw_enter(&dn->dn_struct_rwlock, RW_READER);
572 if (dn->dn_datablkshift) {
573 int blkshift = dn->dn_datablkshift;
574 nblks = (P2ROUNDUP(offset + len, 1 << blkshift) -
575 P2ALIGN(offset, 1 << blkshift)) >> blkshift;
577 nblks = (offset < dn->dn_datablksz);
581 blkid = dbuf_whichblock(dn, offset);
582 for (int i = 0; i < nblks; i++)
583 dbuf_prefetch(dn, blkid + i, ZIO_PRIORITY_SYNC_READ);
586 rw_exit(&dn->dn_struct_rwlock);
588 dnode_rele(dn, FTAG);
592 * Get the next "chunk" of file data to free. We traverse the file from
593 * the end so that the file gets shorter over time (if we crashes in the
594 * middle, this will leave us in a better state). We find allocated file
595 * data by simply searching the allocated level 1 indirects.
597 * On input, *start should be the first offset that does not need to be
598 * freed (e.g. "offset + length"). On return, *start will be the first
599 * offset that should be freed.
602 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum)
604 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
605 /* bytes of data covered by a level-1 indirect block */
607 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
609 ASSERT3U(minimum, <=, *start);
611 if (*start - minimum <= iblkrange * maxblks) {
615 ASSERT(ISP2(iblkrange));
617 for (uint64_t blks = 0; *start > minimum && blks < maxblks; blks++) {
621 * dnode_next_offset(BACKWARDS) will find an allocated L1
622 * indirect block at or before the input offset. We must
623 * decrement *start so that it is at the end of the region
627 err = dnode_next_offset(dn,
628 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
630 /* if there are no indirect blocks before start, we are done */
634 } else if (err != 0) {
638 /* set start to the beginning of this L1 indirect */
639 *start = P2ALIGN(*start, iblkrange);
641 if (*start < minimum)
647 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
650 uint64_t object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
653 if (offset >= object_size)
656 if (length == DMU_OBJECT_END || offset + length > object_size)
657 length = object_size - offset;
659 while (length != 0) {
660 uint64_t chunk_end, chunk_begin;
662 chunk_end = chunk_begin = offset + length;
664 /* move chunk_begin backwards to the beginning of this chunk */
665 err = get_next_chunk(dn, &chunk_begin, offset);
668 ASSERT3U(chunk_begin, >=, offset);
669 ASSERT3U(chunk_begin, <=, chunk_end);
671 dmu_tx_t *tx = dmu_tx_create(os);
672 dmu_tx_hold_free(tx, dn->dn_object,
673 chunk_begin, chunk_end - chunk_begin);
674 err = dmu_tx_assign(tx, TXG_WAIT);
679 dnode_free_range(dn, chunk_begin, chunk_end - chunk_begin, tx);
682 length -= chunk_end - chunk_begin;
688 dmu_free_long_range(objset_t *os, uint64_t object,
689 uint64_t offset, uint64_t length)
694 err = dnode_hold(os, object, FTAG, &dn);
697 err = dmu_free_long_range_impl(os, dn, offset, length);
700 * It is important to zero out the maxblkid when freeing the entire
701 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
702 * will take the fast path, and (b) dnode_reallocate() can verify
703 * that the entire file has been freed.
705 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
708 dnode_rele(dn, FTAG);
713 dmu_free_long_object(objset_t *os, uint64_t object)
718 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
722 tx = dmu_tx_create(os);
723 dmu_tx_hold_bonus(tx, object);
724 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
725 err = dmu_tx_assign(tx, TXG_WAIT);
727 err = dmu_object_free(os, object, tx);
737 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
738 uint64_t size, dmu_tx_t *tx)
741 int err = dnode_hold(os, object, FTAG, &dn);
744 ASSERT(offset < UINT64_MAX);
745 ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
746 dnode_free_range(dn, offset, size, tx);
747 dnode_rele(dn, FTAG);
752 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
753 void *buf, uint32_t flags)
759 err = dnode_hold(os, object, FTAG, &dn);
764 * Deal with odd block sizes, where there can't be data past the first
765 * block. If we ever do the tail block optimization, we will need to
766 * handle that here as well.
768 if (dn->dn_maxblkid == 0) {
769 int newsz = offset > dn->dn_datablksz ? 0 :
770 MIN(size, dn->dn_datablksz - offset);
771 bzero((char *)buf + newsz, size - newsz);
776 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
780 * NB: we could do this block-at-a-time, but it's nice
781 * to be reading in parallel.
783 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
784 TRUE, FTAG, &numbufs, &dbp, flags);
788 for (i = 0; i < numbufs; i++) {
791 dmu_buf_t *db = dbp[i];
795 bufoff = offset - db->db_offset;
796 tocpy = (int)MIN(db->db_size - bufoff, size);
798 bcopy((char *)db->db_data + bufoff, buf, tocpy);
802 buf = (char *)buf + tocpy;
804 dmu_buf_rele_array(dbp, numbufs, FTAG);
806 dnode_rele(dn, FTAG);
811 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
812 const void *buf, dmu_tx_t *tx)
820 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
821 FALSE, FTAG, &numbufs, &dbp));
823 for (i = 0; i < numbufs; i++) {
826 dmu_buf_t *db = dbp[i];
830 bufoff = offset - db->db_offset;
831 tocpy = (int)MIN(db->db_size - bufoff, size);
833 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
835 if (tocpy == db->db_size)
836 dmu_buf_will_fill(db, tx);
838 dmu_buf_will_dirty(db, tx);
840 bcopy(buf, (char *)db->db_data + bufoff, tocpy);
842 if (tocpy == db->db_size)
843 dmu_buf_fill_done(db, tx);
847 buf = (char *)buf + tocpy;
849 dmu_buf_rele_array(dbp, numbufs, FTAG);
853 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
862 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
863 FALSE, FTAG, &numbufs, &dbp));
865 for (i = 0; i < numbufs; i++) {
866 dmu_buf_t *db = dbp[i];
868 dmu_buf_will_not_fill(db, tx);
870 dmu_buf_rele_array(dbp, numbufs, FTAG);
874 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
875 void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
876 int compressed_size, int byteorder, dmu_tx_t *tx)
880 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
881 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
882 VERIFY0(dmu_buf_hold_noread(os, object, offset,
885 dmu_buf_write_embedded(db,
886 data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
887 uncompressed_size, compressed_size, byteorder, tx);
889 dmu_buf_rele(db, FTAG);
893 * DMU support for xuio
895 kstat_t *xuio_ksp = NULL;
898 dmu_xuio_init(xuio_t *xuio, int nblk)
901 uio_t *uio = &xuio->xu_uio;
903 uio->uio_iovcnt = nblk;
904 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
906 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
908 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
909 priv->iovp = uio->uio_iov;
910 XUIO_XUZC_PRIV(xuio) = priv;
912 if (XUIO_XUZC_RW(xuio) == UIO_READ)
913 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
915 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
921 dmu_xuio_fini(xuio_t *xuio)
923 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
924 int nblk = priv->cnt;
926 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
927 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
928 kmem_free(priv, sizeof (dmu_xuio_t));
930 if (XUIO_XUZC_RW(xuio) == UIO_READ)
931 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
933 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
937 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
938 * and increase priv->next by 1.
941 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
944 uio_t *uio = &xuio->xu_uio;
945 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
946 int i = priv->next++;
948 ASSERT(i < priv->cnt);
949 ASSERT(off + n <= arc_buf_size(abuf));
950 iov = uio->uio_iov + i;
951 iov->iov_base = (char *)abuf->b_data + off;
953 priv->bufs[i] = abuf;
958 dmu_xuio_cnt(xuio_t *xuio)
960 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
965 dmu_xuio_arcbuf(xuio_t *xuio, int i)
967 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
969 ASSERT(i < priv->cnt);
970 return (priv->bufs[i]);
974 dmu_xuio_clear(xuio_t *xuio, int i)
976 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
978 ASSERT(i < priv->cnt);
979 priv->bufs[i] = NULL;
985 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
986 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
988 if (xuio_ksp != NULL) {
989 xuio_ksp->ks_data = &xuio_stats;
990 kstat_install(xuio_ksp);
997 if (xuio_ksp != NULL) {
998 kstat_delete(xuio_ksp);
1004 xuio_stat_wbuf_copied()
1006 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1010 xuio_stat_wbuf_nocopy()
1012 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
1017 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
1020 int numbufs, i, err;
1021 xuio_t *xuio = NULL;
1024 * NB: we could do this block-at-a-time, but it's nice
1025 * to be reading in parallel.
1027 err = dmu_buf_hold_array(os, object, uio->uio_loffset, size, TRUE, FTAG,
1033 if (uio->uio_extflg == UIO_XUIO)
1034 xuio = (xuio_t *)uio;
1037 for (i = 0; i < numbufs; i++) {
1040 dmu_buf_t *db = dbp[i];
1044 bufoff = uio->uio_loffset - db->db_offset;
1045 tocpy = (int)MIN(db->db_size - bufoff, size);
1048 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
1049 arc_buf_t *dbuf_abuf = dbi->db_buf;
1050 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
1051 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
1053 uio->uio_resid -= tocpy;
1054 uio->uio_loffset += tocpy;
1057 if (abuf == dbuf_abuf)
1058 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
1060 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
1062 err = uiomove((char *)db->db_data + bufoff, tocpy,
1070 dmu_buf_rele_array(dbp, numbufs, FTAG);
1076 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1083 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1084 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1088 for (i = 0; i < numbufs; i++) {
1091 dmu_buf_t *db = dbp[i];
1095 bufoff = uio->uio_loffset - db->db_offset;
1096 tocpy = (int)MIN(db->db_size - bufoff, size);
1098 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1100 if (tocpy == db->db_size)
1101 dmu_buf_will_fill(db, tx);
1103 dmu_buf_will_dirty(db, tx);
1106 * XXX uiomove could block forever (eg. nfs-backed
1107 * pages). There needs to be a uiolockdown() function
1108 * to lock the pages in memory, so that uiomove won't
1111 err = uiomove((char *)db->db_data + bufoff, tocpy,
1114 if (tocpy == db->db_size)
1115 dmu_buf_fill_done(db, tx);
1123 dmu_buf_rele_array(dbp, numbufs, FTAG);
1128 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1131 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1140 err = dmu_write_uio_dnode(dn, uio, size, tx);
1147 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1156 err = dnode_hold(os, object, FTAG, &dn);
1160 err = dmu_write_uio_dnode(dn, uio, size, tx);
1162 dnode_rele(dn, FTAG);
1169 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1170 page_t *pp, dmu_tx_t *tx)
1179 err = dmu_buf_hold_array(os, object, offset, size,
1180 FALSE, FTAG, &numbufs, &dbp);
1184 for (i = 0; i < numbufs; i++) {
1185 int tocpy, copied, thiscpy;
1187 dmu_buf_t *db = dbp[i];
1191 ASSERT3U(db->db_size, >=, PAGESIZE);
1193 bufoff = offset - db->db_offset;
1194 tocpy = (int)MIN(db->db_size - bufoff, size);
1196 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1198 if (tocpy == db->db_size)
1199 dmu_buf_will_fill(db, tx);
1201 dmu_buf_will_dirty(db, tx);
1203 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1204 ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff);
1205 thiscpy = MIN(PAGESIZE, tocpy - copied);
1206 va = zfs_map_page(pp, S_READ);
1207 bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1208 zfs_unmap_page(pp, va);
1213 if (tocpy == db->db_size)
1214 dmu_buf_fill_done(db, tx);
1219 dmu_buf_rele_array(dbp, numbufs, FTAG);
1226 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1227 vm_page_t *ma, dmu_tx_t *tx)
1237 err = dmu_buf_hold_array(os, object, offset, size,
1238 FALSE, FTAG, &numbufs, &dbp);
1242 for (i = 0; i < numbufs; i++) {
1243 int tocpy, copied, thiscpy;
1245 dmu_buf_t *db = dbp[i];
1249 ASSERT3U(db->db_size, >=, PAGESIZE);
1251 bufoff = offset - db->db_offset;
1252 tocpy = (int)MIN(db->db_size - bufoff, size);
1254 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1256 if (tocpy == db->db_size)
1257 dmu_buf_will_fill(db, tx);
1259 dmu_buf_will_dirty(db, tx);
1261 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1262 ASSERT3U(ptoa((*ma)->pindex), ==, db->db_offset + bufoff);
1263 thiscpy = MIN(PAGESIZE, tocpy - copied);
1264 va = zfs_map_page(*ma, &sf);
1265 bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1271 if (tocpy == db->db_size)
1272 dmu_buf_fill_done(db, tx);
1277 dmu_buf_rele_array(dbp, numbufs, FTAG);
1284 * Allocate a loaned anonymous arc buffer.
1287 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1289 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1291 return (arc_loan_buf(db->db_objset->os_spa, size));
1295 * Free a loaned arc buffer.
1298 dmu_return_arcbuf(arc_buf_t *buf)
1300 arc_return_buf(buf, FTAG);
1301 VERIFY(arc_buf_remove_ref(buf, FTAG));
1305 * When possible directly assign passed loaned arc buffer to a dbuf.
1306 * If this is not possible copy the contents of passed arc buf via
1310 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1313 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1316 uint32_t blksz = (uint32_t)arc_buf_size(buf);
1319 DB_DNODE_ENTER(dbuf);
1320 dn = DB_DNODE(dbuf);
1321 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1322 blkid = dbuf_whichblock(dn, offset);
1323 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1324 rw_exit(&dn->dn_struct_rwlock);
1325 DB_DNODE_EXIT(dbuf);
1327 if (offset == db->db.db_offset && blksz == db->db.db_size) {
1328 dbuf_assign_arcbuf(db, buf, tx);
1329 dbuf_rele(db, FTAG);
1334 DB_DNODE_ENTER(dbuf);
1335 dn = DB_DNODE(dbuf);
1337 object = dn->dn_object;
1338 DB_DNODE_EXIT(dbuf);
1340 dbuf_rele(db, FTAG);
1341 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1342 dmu_return_arcbuf(buf);
1343 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1348 dbuf_dirty_record_t *dsa_dr;
1349 dmu_sync_cb_t *dsa_done;
1356 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1358 dmu_sync_arg_t *dsa = varg;
1359 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1360 blkptr_t *bp = zio->io_bp;
1362 if (zio->io_error == 0) {
1363 if (BP_IS_HOLE(bp)) {
1365 * A block of zeros may compress to a hole, but the
1366 * block size still needs to be known for replay.
1368 BP_SET_LSIZE(bp, db->db_size);
1369 } else if (!BP_IS_EMBEDDED(bp)) {
1370 ASSERT(BP_GET_LEVEL(bp) == 0);
1377 dmu_sync_late_arrival_ready(zio_t *zio)
1379 dmu_sync_ready(zio, NULL, zio->io_private);
1384 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1386 dmu_sync_arg_t *dsa = varg;
1387 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1388 dmu_buf_impl_t *db = dr->dr_dbuf;
1390 mutex_enter(&db->db_mtx);
1391 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1392 if (zio->io_error == 0) {
1393 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1394 if (dr->dt.dl.dr_nopwrite) {
1395 blkptr_t *bp = zio->io_bp;
1396 blkptr_t *bp_orig = &zio->io_bp_orig;
1397 uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1399 ASSERT(BP_EQUAL(bp, bp_orig));
1400 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1401 ASSERT(zio_checksum_table[chksum].ci_dedup);
1403 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1404 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1405 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1406 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by))
1407 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1409 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1411 cv_broadcast(&db->db_changed);
1412 mutex_exit(&db->db_mtx);
1414 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1416 kmem_free(dsa, sizeof (*dsa));
1420 dmu_sync_late_arrival_done(zio_t *zio)
1422 blkptr_t *bp = zio->io_bp;
1423 dmu_sync_arg_t *dsa = zio->io_private;
1424 blkptr_t *bp_orig = &zio->io_bp_orig;
1426 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1428 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE)
1429 * then there is nothing to do here. Otherwise, free the
1430 * newly allocated block in this txg.
1432 if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
1433 ASSERT(BP_EQUAL(bp, bp_orig));
1435 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1436 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1437 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1438 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1442 dmu_tx_commit(dsa->dsa_tx);
1444 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1446 kmem_free(dsa, sizeof (*dsa));
1450 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1451 zio_prop_t *zp, zbookmark_t *zb)
1453 dmu_sync_arg_t *dsa;
1456 tx = dmu_tx_create(os);
1457 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1458 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1460 /* Make zl_get_data do txg_waited_synced() */
1461 return (SET_ERROR(EIO));
1464 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1466 dsa->dsa_done = done;
1470 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1471 zgd->zgd_db->db_data, zgd->zgd_db->db_size, zp,
1472 dmu_sync_late_arrival_ready, NULL, dmu_sync_late_arrival_done, dsa,
1473 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1479 * Intent log support: sync the block associated with db to disk.
1480 * N.B. and XXX: the caller is responsible for making sure that the
1481 * data isn't changing while dmu_sync() is writing it.
1485 * EEXIST: this txg has already been synced, so there's nothing to do.
1486 * The caller should not log the write.
1488 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1489 * The caller should not log the write.
1491 * EALREADY: this block is already in the process of being synced.
1492 * The caller should track its progress (somehow).
1494 * EIO: could not do the I/O.
1495 * The caller should do a txg_wait_synced().
1497 * 0: the I/O has been initiated.
1498 * The caller should log this blkptr in the done callback.
1499 * It is possible that the I/O will fail, in which case
1500 * the error will be reported to the done callback and
1501 * propagated to pio from zio_done().
1504 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1506 blkptr_t *bp = zgd->zgd_bp;
1507 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1508 objset_t *os = db->db_objset;
1509 dsl_dataset_t *ds = os->os_dsl_dataset;
1510 dbuf_dirty_record_t *dr;
1511 dmu_sync_arg_t *dsa;
1516 ASSERT(pio != NULL);
1519 SET_BOOKMARK(&zb, ds->ds_object,
1520 db->db.db_object, db->db_level, db->db_blkid);
1524 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1528 * If we're frozen (running ziltest), we always need to generate a bp.
1530 if (txg > spa_freeze_txg(os->os_spa))
1531 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1534 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1535 * and us. If we determine that this txg is not yet syncing,
1536 * but it begins to sync a moment later, that's OK because the
1537 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1539 mutex_enter(&db->db_mtx);
1541 if (txg <= spa_last_synced_txg(os->os_spa)) {
1543 * This txg has already synced. There's nothing to do.
1545 mutex_exit(&db->db_mtx);
1546 return (SET_ERROR(EEXIST));
1549 if (txg <= spa_syncing_txg(os->os_spa)) {
1551 * This txg is currently syncing, so we can't mess with
1552 * the dirty record anymore; just write a new log block.
1554 mutex_exit(&db->db_mtx);
1555 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1558 dr = db->db_last_dirty;
1559 while (dr && dr->dr_txg != txg)
1564 * There's no dr for this dbuf, so it must have been freed.
1565 * There's no need to log writes to freed blocks, so we're done.
1567 mutex_exit(&db->db_mtx);
1568 return (SET_ERROR(ENOENT));
1571 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
1574 * Assume the on-disk data is X, the current syncing data is Y,
1575 * and the current in-memory data is Z (currently in dmu_sync).
1576 * X and Z are identical but Y is has been modified. Normally,
1577 * when X and Z are the same we will perform a nopwrite but if Y
1578 * is different we must disable nopwrite since the resulting write
1579 * of Y to disk can free the block containing X. If we allowed a
1580 * nopwrite to occur the block pointing to Z would reference a freed
1581 * block. Since this is a rare case we simplify this by disabling
1582 * nopwrite if the current dmu_sync-ing dbuf has been modified in
1583 * a previous transaction.
1586 zp.zp_nopwrite = B_FALSE;
1588 ASSERT(dr->dr_txg == txg);
1589 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1590 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1592 * We have already issued a sync write for this buffer,
1593 * or this buffer has already been synced. It could not
1594 * have been dirtied since, or we would have cleared the state.
1596 mutex_exit(&db->db_mtx);
1597 return (SET_ERROR(EALREADY));
1600 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1601 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1602 mutex_exit(&db->db_mtx);
1604 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1606 dsa->dsa_done = done;
1610 zio_nowait(arc_write(pio, os->os_spa, txg,
1611 bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
1612 DBUF_IS_L2COMPRESSIBLE(db), &zp, dmu_sync_ready,
1613 NULL, dmu_sync_done, dsa, ZIO_PRIORITY_SYNC_WRITE,
1614 ZIO_FLAG_CANFAIL, &zb));
1620 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1626 err = dnode_hold(os, object, FTAG, &dn);
1629 err = dnode_set_blksz(dn, size, ibs, tx);
1630 dnode_rele(dn, FTAG);
1635 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1641 * Send streams include each object's checksum function. This
1642 * check ensures that the receiving system can understand the
1643 * checksum function transmitted.
1645 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS);
1647 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1648 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS);
1649 dn->dn_checksum = checksum;
1650 dnode_setdirty(dn, tx);
1651 dnode_rele(dn, FTAG);
1655 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1661 * Send streams include each object's compression function. This
1662 * check ensures that the receiving system can understand the
1663 * compression function transmitted.
1665 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS);
1667 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1668 dn->dn_compress = compress;
1669 dnode_setdirty(dn, tx);
1670 dnode_rele(dn, FTAG);
1673 int zfs_mdcomp_disable = 0;
1674 TUNABLE_INT("vfs.zfs.mdcomp_disable", &zfs_mdcomp_disable);
1675 SYSCTL_INT(_vfs_zfs, OID_AUTO, mdcomp_disable, CTLFLAG_RW,
1676 &zfs_mdcomp_disable, 0, "Disable metadata compression");
1679 * When the "redundant_metadata" property is set to "most", only indirect
1680 * blocks of this level and higher will have an additional ditto block.
1682 int zfs_redundant_metadata_most_ditto_level = 2;
1685 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
1687 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
1688 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
1690 enum zio_checksum checksum = os->os_checksum;
1691 enum zio_compress compress = os->os_compress;
1692 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
1693 boolean_t dedup = B_FALSE;
1694 boolean_t nopwrite = B_FALSE;
1695 boolean_t dedup_verify = os->os_dedup_verify;
1696 int copies = os->os_copies;
1699 * We maintain different write policies for each of the following
1702 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1703 * 3. all other level 0 blocks
1707 * XXX -- we should design a compression algorithm
1708 * that specializes in arrays of bps.
1710 compress = zfs_mdcomp_disable ? ZIO_COMPRESS_EMPTY :
1714 * Metadata always gets checksummed. If the data
1715 * checksum is multi-bit correctable, and it's not a
1716 * ZBT-style checksum, then it's suitable for metadata
1717 * as well. Otherwise, the metadata checksum defaults
1720 if (zio_checksum_table[checksum].ci_correctable < 1 ||
1721 zio_checksum_table[checksum].ci_eck)
1722 checksum = ZIO_CHECKSUM_FLETCHER_4;
1724 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL ||
1725 (os->os_redundant_metadata ==
1726 ZFS_REDUNDANT_METADATA_MOST &&
1727 (level >= zfs_redundant_metadata_most_ditto_level ||
1728 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))))
1730 } else if (wp & WP_NOFILL) {
1734 * If we're writing preallocated blocks, we aren't actually
1735 * writing them so don't set any policy properties. These
1736 * blocks are currently only used by an external subsystem
1737 * outside of zfs (i.e. dump) and not written by the zio
1740 compress = ZIO_COMPRESS_OFF;
1741 checksum = ZIO_CHECKSUM_NOPARITY;
1743 compress = zio_compress_select(dn->dn_compress, compress);
1745 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
1746 zio_checksum_select(dn->dn_checksum, checksum) :
1750 * Determine dedup setting. If we are in dmu_sync(),
1751 * we won't actually dedup now because that's all
1752 * done in syncing context; but we do want to use the
1753 * dedup checkum. If the checksum is not strong
1754 * enough to ensure unique signatures, force
1757 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
1758 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
1759 if (!zio_checksum_table[checksum].ci_dedup)
1760 dedup_verify = B_TRUE;
1764 * Enable nopwrite if we have a cryptographically secure
1765 * checksum that has no known collisions (i.e. SHA-256)
1766 * and compression is enabled. We don't enable nopwrite if
1767 * dedup is enabled as the two features are mutually exclusive.
1769 nopwrite = (!dedup && zio_checksum_table[checksum].ci_dedup &&
1770 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
1773 zp->zp_checksum = checksum;
1774 zp->zp_compress = compress;
1775 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
1776 zp->zp_level = level;
1777 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
1778 zp->zp_dedup = dedup;
1779 zp->zp_dedup_verify = dedup && dedup_verify;
1780 zp->zp_nopwrite = nopwrite;
1784 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
1789 err = dnode_hold(os, object, FTAG, &dn);
1793 * Sync any current changes before
1794 * we go trundling through the block pointers.
1796 for (i = 0; i < TXG_SIZE; i++) {
1797 if (list_link_active(&dn->dn_dirty_link[i]))
1800 if (i != TXG_SIZE) {
1801 dnode_rele(dn, FTAG);
1802 txg_wait_synced(dmu_objset_pool(os), 0);
1803 err = dnode_hold(os, object, FTAG, &dn);
1808 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
1809 dnode_rele(dn, FTAG);
1815 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
1819 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1820 mutex_enter(&dn->dn_mtx);
1824 doi->doi_data_block_size = dn->dn_datablksz;
1825 doi->doi_metadata_block_size = dn->dn_indblkshift ?
1826 1ULL << dn->dn_indblkshift : 0;
1827 doi->doi_type = dn->dn_type;
1828 doi->doi_bonus_type = dn->dn_bonustype;
1829 doi->doi_bonus_size = dn->dn_bonuslen;
1830 doi->doi_indirection = dn->dn_nlevels;
1831 doi->doi_checksum = dn->dn_checksum;
1832 doi->doi_compress = dn->dn_compress;
1833 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
1834 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
1835 doi->doi_fill_count = 0;
1836 for (int i = 0; i < dnp->dn_nblkptr; i++)
1837 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
1839 mutex_exit(&dn->dn_mtx);
1840 rw_exit(&dn->dn_struct_rwlock);
1844 * Get information on a DMU object.
1845 * If doi is NULL, just indicates whether the object exists.
1848 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
1851 int err = dnode_hold(os, object, FTAG, &dn);
1857 dmu_object_info_from_dnode(dn, doi);
1859 dnode_rele(dn, FTAG);
1864 * As above, but faster; can be used when you have a held dbuf in hand.
1867 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
1869 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1872 dmu_object_info_from_dnode(DB_DNODE(db), doi);
1877 * Faster still when you only care about the size.
1878 * This is specifically optimized for zfs_getattr().
1881 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
1882 u_longlong_t *nblk512)
1884 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1890 *blksize = dn->dn_datablksz;
1891 /* add 1 for dnode space */
1892 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
1893 SPA_MINBLOCKSHIFT) + 1;
1898 byteswap_uint64_array(void *vbuf, size_t size)
1900 uint64_t *buf = vbuf;
1901 size_t count = size >> 3;
1904 ASSERT((size & 7) == 0);
1906 for (i = 0; i < count; i++)
1907 buf[i] = BSWAP_64(buf[i]);
1911 byteswap_uint32_array(void *vbuf, size_t size)
1913 uint32_t *buf = vbuf;
1914 size_t count = size >> 2;
1917 ASSERT((size & 3) == 0);
1919 for (i = 0; i < count; i++)
1920 buf[i] = BSWAP_32(buf[i]);
1924 byteswap_uint16_array(void *vbuf, size_t size)
1926 uint16_t *buf = vbuf;
1927 size_t count = size >> 1;
1930 ASSERT((size & 1) == 0);
1932 for (i = 0; i < count; i++)
1933 buf[i] = BSWAP_16(buf[i]);
1938 byteswap_uint8_array(void *vbuf, size_t size)
1952 zio_compress_init();
1960 arc_fini(); /* arc depends on l2arc, so arc must go first */
1963 zio_compress_fini();