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, 2015 by Delphix. All rights reserved.
24 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
25 * Copyright (c) 2014, Nexenta Systems, Inc. All rights reserved.
26 * Copyright (c) 2015 by Chunwei Chen. All rights reserved.
30 #include <sys/dmu_impl.h>
31 #include <sys/dmu_tx.h>
33 #include <sys/dnode.h>
34 #include <sys/zfs_context.h>
35 #include <sys/dmu_objset.h>
36 #include <sys/dmu_traverse.h>
37 #include <sys/dsl_dataset.h>
38 #include <sys/dsl_dir.h>
39 #include <sys/dsl_pool.h>
40 #include <sys/dsl_synctask.h>
41 #include <sys/dsl_prop.h>
42 #include <sys/dmu_zfetch.h>
43 #include <sys/zfs_ioctl.h>
45 #include <sys/zio_checksum.h>
46 #include <sys/zio_compress.h>
48 #include <sys/zfeature.h>
50 #include <sys/vmsystm.h>
51 #include <sys/zfs_znode.h>
55 * Enable/disable nopwrite feature.
57 int zfs_nopwrite_enabled = 1;
59 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
60 { DMU_BSWAP_UINT8, TRUE, "unallocated" },
61 { DMU_BSWAP_ZAP, TRUE, "object directory" },
62 { DMU_BSWAP_UINT64, TRUE, "object array" },
63 { DMU_BSWAP_UINT8, TRUE, "packed nvlist" },
64 { DMU_BSWAP_UINT64, TRUE, "packed nvlist size" },
65 { DMU_BSWAP_UINT64, TRUE, "bpobj" },
66 { DMU_BSWAP_UINT64, TRUE, "bpobj header" },
67 { DMU_BSWAP_UINT64, TRUE, "SPA space map header" },
68 { DMU_BSWAP_UINT64, TRUE, "SPA space map" },
69 { DMU_BSWAP_UINT64, TRUE, "ZIL intent log" },
70 { DMU_BSWAP_DNODE, TRUE, "DMU dnode" },
71 { DMU_BSWAP_OBJSET, TRUE, "DMU objset" },
72 { DMU_BSWAP_UINT64, TRUE, "DSL directory" },
73 { DMU_BSWAP_ZAP, TRUE, "DSL directory child map"},
74 { DMU_BSWAP_ZAP, TRUE, "DSL dataset snap map" },
75 { DMU_BSWAP_ZAP, TRUE, "DSL props" },
76 { DMU_BSWAP_UINT64, TRUE, "DSL dataset" },
77 { DMU_BSWAP_ZNODE, TRUE, "ZFS znode" },
78 { DMU_BSWAP_OLDACL, TRUE, "ZFS V0 ACL" },
79 { DMU_BSWAP_UINT8, FALSE, "ZFS plain file" },
80 { DMU_BSWAP_ZAP, TRUE, "ZFS directory" },
81 { DMU_BSWAP_ZAP, TRUE, "ZFS master node" },
82 { DMU_BSWAP_ZAP, TRUE, "ZFS delete queue" },
83 { DMU_BSWAP_UINT8, FALSE, "zvol object" },
84 { DMU_BSWAP_ZAP, TRUE, "zvol prop" },
85 { DMU_BSWAP_UINT8, FALSE, "other uint8[]" },
86 { DMU_BSWAP_UINT64, FALSE, "other uint64[]" },
87 { DMU_BSWAP_ZAP, TRUE, "other ZAP" },
88 { DMU_BSWAP_ZAP, TRUE, "persistent error log" },
89 { DMU_BSWAP_UINT8, TRUE, "SPA history" },
90 { DMU_BSWAP_UINT64, TRUE, "SPA history offsets" },
91 { DMU_BSWAP_ZAP, TRUE, "Pool properties" },
92 { DMU_BSWAP_ZAP, TRUE, "DSL permissions" },
93 { DMU_BSWAP_ACL, TRUE, "ZFS ACL" },
94 { DMU_BSWAP_UINT8, TRUE, "ZFS SYSACL" },
95 { DMU_BSWAP_UINT8, TRUE, "FUID table" },
96 { DMU_BSWAP_UINT64, TRUE, "FUID table size" },
97 { DMU_BSWAP_ZAP, TRUE, "DSL dataset next clones"},
98 { DMU_BSWAP_ZAP, TRUE, "scan work queue" },
99 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group used" },
100 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group quota" },
101 { DMU_BSWAP_ZAP, TRUE, "snapshot refcount tags"},
102 { DMU_BSWAP_ZAP, TRUE, "DDT ZAP algorithm" },
103 { DMU_BSWAP_ZAP, TRUE, "DDT statistics" },
104 { DMU_BSWAP_UINT8, TRUE, "System attributes" },
105 { DMU_BSWAP_ZAP, TRUE, "SA master node" },
106 { DMU_BSWAP_ZAP, TRUE, "SA attr registration" },
107 { DMU_BSWAP_ZAP, TRUE, "SA attr layouts" },
108 { DMU_BSWAP_ZAP, TRUE, "scan translations" },
109 { DMU_BSWAP_UINT8, FALSE, "deduplicated block" },
110 { DMU_BSWAP_ZAP, TRUE, "DSL deadlist map" },
111 { DMU_BSWAP_UINT64, TRUE, "DSL deadlist map hdr" },
112 { DMU_BSWAP_ZAP, TRUE, "DSL dir clones" },
113 { DMU_BSWAP_UINT64, TRUE, "bpobj subobj" }
116 const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
117 { byteswap_uint8_array, "uint8" },
118 { byteswap_uint16_array, "uint16" },
119 { byteswap_uint32_array, "uint32" },
120 { byteswap_uint64_array, "uint64" },
121 { zap_byteswap, "zap" },
122 { dnode_buf_byteswap, "dnode" },
123 { dmu_objset_byteswap, "objset" },
124 { zfs_znode_byteswap, "znode" },
125 { zfs_oldacl_byteswap, "oldacl" },
126 { zfs_acl_byteswap, "acl" }
130 dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset,
131 void *tag, dmu_buf_t **dbp)
138 err = dnode_hold(os, object, FTAG, &dn);
141 blkid = dbuf_whichblock(dn, 0, offset);
142 rw_enter(&dn->dn_struct_rwlock, RW_READER);
143 db = dbuf_hold(dn, blkid, tag);
144 rw_exit(&dn->dn_struct_rwlock);
145 dnode_rele(dn, FTAG);
149 return (SET_ERROR(EIO));
157 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
158 void *tag, dmu_buf_t **dbp, int flags)
161 int db_flags = DB_RF_CANFAIL;
163 if (flags & DMU_READ_NO_PREFETCH)
164 db_flags |= DB_RF_NOPREFETCH;
166 err = dmu_buf_hold_noread(os, object, offset, tag, dbp);
168 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
169 err = dbuf_read(db, NULL, db_flags);
182 return (DN_MAX_BONUSLEN);
186 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
188 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
195 if (dn->dn_bonus != db) {
196 error = SET_ERROR(EINVAL);
197 } else if (newsize < 0 || newsize > db_fake->db_size) {
198 error = SET_ERROR(EINVAL);
200 dnode_setbonuslen(dn, newsize, tx);
209 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
211 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
218 if (!DMU_OT_IS_VALID(type)) {
219 error = SET_ERROR(EINVAL);
220 } else if (dn->dn_bonus != db) {
221 error = SET_ERROR(EINVAL);
223 dnode_setbonus_type(dn, type, tx);
232 dmu_get_bonustype(dmu_buf_t *db_fake)
234 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
236 dmu_object_type_t type;
240 type = dn->dn_bonustype;
247 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
252 error = dnode_hold(os, object, FTAG, &dn);
253 dbuf_rm_spill(dn, tx);
254 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
255 dnode_rm_spill(dn, tx);
256 rw_exit(&dn->dn_struct_rwlock);
257 dnode_rele(dn, FTAG);
262 * returns ENOENT, EIO, or 0.
265 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
271 error = dnode_hold(os, object, FTAG, &dn);
275 rw_enter(&dn->dn_struct_rwlock, RW_READER);
276 if (dn->dn_bonus == NULL) {
277 rw_exit(&dn->dn_struct_rwlock);
278 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
279 if (dn->dn_bonus == NULL)
280 dbuf_create_bonus(dn);
284 /* as long as the bonus buf is held, the dnode will be held */
285 if (refcount_add(&db->db_holds, tag) == 1) {
286 VERIFY(dnode_add_ref(dn, db));
287 atomic_inc_32(&dn->dn_dbufs_count);
291 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
292 * hold and incrementing the dbuf count to ensure that dnode_move() sees
293 * a dnode hold for every dbuf.
295 rw_exit(&dn->dn_struct_rwlock);
297 dnode_rele(dn, FTAG);
299 VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH));
306 * returns ENOENT, EIO, or 0.
308 * This interface will allocate a blank spill dbuf when a spill blk
309 * doesn't already exist on the dnode.
311 * if you only want to find an already existing spill db, then
312 * dmu_spill_hold_existing() should be used.
315 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
317 dmu_buf_impl_t *db = NULL;
320 if ((flags & DB_RF_HAVESTRUCT) == 0)
321 rw_enter(&dn->dn_struct_rwlock, RW_READER);
323 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
325 if ((flags & DB_RF_HAVESTRUCT) == 0)
326 rw_exit(&dn->dn_struct_rwlock);
329 err = dbuf_read(db, NULL, flags);
338 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
340 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
347 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
348 err = SET_ERROR(EINVAL);
350 rw_enter(&dn->dn_struct_rwlock, RW_READER);
352 if (!dn->dn_have_spill) {
353 err = SET_ERROR(ENOENT);
355 err = dmu_spill_hold_by_dnode(dn,
356 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
359 rw_exit(&dn->dn_struct_rwlock);
367 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
369 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
375 err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp);
382 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
383 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
384 * and can induce severe lock contention when writing to several files
385 * whose dnodes are in the same block.
388 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
389 boolean_t read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
392 uint64_t blkid, nblks, i;
397 ASSERT(length <= DMU_MAX_ACCESS);
400 * Note: We directly notify the prefetch code of this read, so that
401 * we can tell it about the multi-block read. dbuf_read() only knows
402 * about the one block it is accessing.
404 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT |
407 rw_enter(&dn->dn_struct_rwlock, RW_READER);
408 if (dn->dn_datablkshift) {
409 int blkshift = dn->dn_datablkshift;
410 nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) -
411 P2ALIGN(offset, 1ULL << blkshift)) >> blkshift;
413 if (offset + length > dn->dn_datablksz) {
414 zfs_panic_recover("zfs: accessing past end of object "
415 "%llx/%llx (size=%u access=%llu+%llu)",
416 (longlong_t)dn->dn_objset->
417 os_dsl_dataset->ds_object,
418 (longlong_t)dn->dn_object, dn->dn_datablksz,
419 (longlong_t)offset, (longlong_t)length);
420 rw_exit(&dn->dn_struct_rwlock);
421 return (SET_ERROR(EIO));
425 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
427 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
428 blkid = dbuf_whichblock(dn, 0, offset);
429 for (i = 0; i < nblks; i++) {
430 dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag);
432 rw_exit(&dn->dn_struct_rwlock);
433 dmu_buf_rele_array(dbp, nblks, tag);
435 return (SET_ERROR(EIO));
438 /* initiate async i/o */
440 (void) dbuf_read(db, zio, dbuf_flags);
444 if ((flags & DMU_READ_NO_PREFETCH) == 0 && read &&
445 length < zfetch_array_rd_sz) {
446 dmu_zfetch(&dn->dn_zfetch, blkid, nblks);
448 rw_exit(&dn->dn_struct_rwlock);
450 /* wait for async i/o */
453 dmu_buf_rele_array(dbp, nblks, tag);
457 /* wait for other io to complete */
459 for (i = 0; i < nblks; i++) {
460 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
461 mutex_enter(&db->db_mtx);
462 while (db->db_state == DB_READ ||
463 db->db_state == DB_FILL)
464 cv_wait(&db->db_changed, &db->db_mtx);
465 if (db->db_state == DB_UNCACHED)
466 err = SET_ERROR(EIO);
467 mutex_exit(&db->db_mtx);
469 dmu_buf_rele_array(dbp, nblks, tag);
481 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
482 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
487 err = dnode_hold(os, object, FTAG, &dn);
491 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
492 numbufsp, dbpp, DMU_READ_PREFETCH);
494 dnode_rele(dn, FTAG);
500 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
501 uint64_t length, boolean_t read, void *tag, int *numbufsp,
504 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
510 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
511 numbufsp, dbpp, DMU_READ_PREFETCH);
518 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
521 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
526 for (i = 0; i < numbufs; i++) {
528 dbuf_rele(dbp[i], tag);
531 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
535 * Issue prefetch i/os for the given blocks. If level is greater than 0, the
536 * indirect blocks prefeteched will be those that point to the blocks containing
537 * the data starting at offset, and continuing to offset + len.
539 * Note that if the indirect blocks above the blocks being prefetched are not in
540 * cache, they will be asychronously read in.
543 dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset,
544 uint64_t len, zio_priority_t pri)
550 if (len == 0) { /* they're interested in the bonus buffer */
551 dn = DMU_META_DNODE(os);
553 if (object == 0 || object >= DN_MAX_OBJECT)
556 rw_enter(&dn->dn_struct_rwlock, RW_READER);
557 blkid = dbuf_whichblock(dn, level,
558 object * sizeof (dnode_phys_t));
559 dbuf_prefetch(dn, level, blkid, pri, 0);
560 rw_exit(&dn->dn_struct_rwlock);
565 * XXX - Note, if the dnode for the requested object is not
566 * already cached, we will do a *synchronous* read in the
567 * dnode_hold() call. The same is true for any indirects.
569 err = dnode_hold(os, object, FTAG, &dn);
573 rw_enter(&dn->dn_struct_rwlock, RW_READER);
575 * offset + len - 1 is the last byte we want to prefetch for, and offset
576 * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the
577 * last block we want to prefetch, and dbuf_whichblock(dn, level,
578 * offset) is the first. Then the number we need to prefetch is the
581 if (level > 0 || dn->dn_datablkshift != 0) {
582 nblks = dbuf_whichblock(dn, level, offset + len - 1) -
583 dbuf_whichblock(dn, level, offset) + 1;
585 nblks = (offset < dn->dn_datablksz);
591 blkid = dbuf_whichblock(dn, level, offset);
592 for (i = 0; i < nblks; i++)
593 dbuf_prefetch(dn, level, blkid + i, pri, 0);
596 rw_exit(&dn->dn_struct_rwlock);
598 dnode_rele(dn, FTAG);
602 * Get the next "chunk" of file data to free. We traverse the file from
603 * the end so that the file gets shorter over time (if we crashes in the
604 * middle, this will leave us in a better state). We find allocated file
605 * data by simply searching the allocated level 1 indirects.
607 * On input, *start should be the first offset that does not need to be
608 * freed (e.g. "offset + length"). On return, *start will be the first
609 * offset that should be freed.
612 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum)
614 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
615 /* bytes of data covered by a level-1 indirect block */
617 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
620 ASSERT3U(minimum, <=, *start);
622 if (*start - minimum <= iblkrange * maxblks) {
626 ASSERT(ISP2(iblkrange));
628 for (blks = 0; *start > minimum && blks < maxblks; blks++) {
632 * dnode_next_offset(BACKWARDS) will find an allocated L1
633 * indirect block at or before the input offset. We must
634 * decrement *start so that it is at the end of the region
638 err = dnode_next_offset(dn,
639 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
641 /* if there are no indirect blocks before start, we are done */
645 } else if (err != 0) {
649 /* set start to the beginning of this L1 indirect */
650 *start = P2ALIGN(*start, iblkrange);
652 if (*start < minimum)
658 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
661 uint64_t object_size;
665 return (SET_ERROR(EINVAL));
667 object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
668 if (offset >= object_size)
671 if (length == DMU_OBJECT_END || offset + length > object_size)
672 length = object_size - offset;
674 while (length != 0) {
675 uint64_t chunk_end, chunk_begin;
678 chunk_end = chunk_begin = offset + length;
680 /* move chunk_begin backwards to the beginning of this chunk */
681 err = get_next_chunk(dn, &chunk_begin, offset);
684 ASSERT3U(chunk_begin, >=, offset);
685 ASSERT3U(chunk_begin, <=, chunk_end);
687 tx = dmu_tx_create(os);
688 dmu_tx_hold_free(tx, dn->dn_object,
689 chunk_begin, chunk_end - chunk_begin);
690 err = dmu_tx_assign(tx, TXG_WAIT);
695 dnode_free_range(dn, chunk_begin, chunk_end - chunk_begin, tx);
698 length -= chunk_end - chunk_begin;
704 dmu_free_long_range(objset_t *os, uint64_t object,
705 uint64_t offset, uint64_t length)
710 err = dnode_hold(os, object, FTAG, &dn);
713 err = dmu_free_long_range_impl(os, dn, offset, length);
716 * It is important to zero out the maxblkid when freeing the entire
717 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
718 * will take the fast path, and (b) dnode_reallocate() can verify
719 * that the entire file has been freed.
721 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
724 dnode_rele(dn, FTAG);
729 dmu_free_long_object(objset_t *os, uint64_t object)
734 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
738 tx = dmu_tx_create(os);
739 dmu_tx_hold_bonus(tx, object);
740 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
741 err = dmu_tx_assign(tx, TXG_WAIT);
743 err = dmu_object_free(os, object, tx);
753 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
754 uint64_t size, dmu_tx_t *tx)
757 int err = dnode_hold(os, object, FTAG, &dn);
760 ASSERT(offset < UINT64_MAX);
761 ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
762 dnode_free_range(dn, offset, size, tx);
763 dnode_rele(dn, FTAG);
768 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
769 void *buf, uint32_t flags)
775 err = dnode_hold(os, object, FTAG, &dn);
780 * Deal with odd block sizes, where there can't be data past the first
781 * block. If we ever do the tail block optimization, we will need to
782 * handle that here as well.
784 if (dn->dn_maxblkid == 0) {
785 uint64_t newsz = offset > dn->dn_datablksz ? 0 :
786 MIN(size, dn->dn_datablksz - offset);
787 bzero((char *)buf + newsz, size - newsz);
792 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
796 * NB: we could do this block-at-a-time, but it's nice
797 * to be reading in parallel.
799 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
800 TRUE, FTAG, &numbufs, &dbp, flags);
804 for (i = 0; i < numbufs; i++) {
807 dmu_buf_t *db = dbp[i];
811 bufoff = offset - db->db_offset;
812 tocpy = MIN(db->db_size - bufoff, size);
814 (void) memcpy(buf, (char *)db->db_data + bufoff, tocpy);
818 buf = (char *)buf + tocpy;
820 dmu_buf_rele_array(dbp, numbufs, FTAG);
822 dnode_rele(dn, FTAG);
827 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
828 const void *buf, dmu_tx_t *tx)
836 VERIFY0(dmu_buf_hold_array(os, object, offset, size,
837 FALSE, FTAG, &numbufs, &dbp));
839 for (i = 0; i < numbufs; i++) {
842 dmu_buf_t *db = dbp[i];
846 bufoff = offset - db->db_offset;
847 tocpy = MIN(db->db_size - bufoff, size);
849 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
851 if (tocpy == db->db_size)
852 dmu_buf_will_fill(db, tx);
854 dmu_buf_will_dirty(db, tx);
856 (void) memcpy((char *)db->db_data + bufoff, buf, tocpy);
858 if (tocpy == db->db_size)
859 dmu_buf_fill_done(db, tx);
863 buf = (char *)buf + tocpy;
865 dmu_buf_rele_array(dbp, numbufs, FTAG);
869 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
878 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
879 FALSE, FTAG, &numbufs, &dbp));
881 for (i = 0; i < numbufs; i++) {
882 dmu_buf_t *db = dbp[i];
884 dmu_buf_will_not_fill(db, tx);
886 dmu_buf_rele_array(dbp, numbufs, FTAG);
890 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
891 void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
892 int compressed_size, int byteorder, dmu_tx_t *tx)
896 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
897 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
898 VERIFY0(dmu_buf_hold_noread(os, object, offset,
901 dmu_buf_write_embedded(db,
902 data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
903 uncompressed_size, compressed_size, byteorder, tx);
905 dmu_buf_rele(db, FTAG);
909 * DMU support for xuio
911 kstat_t *xuio_ksp = NULL;
913 typedef struct xuio_stats {
914 /* loaned yet not returned arc_buf */
915 kstat_named_t xuiostat_onloan_rbuf;
916 kstat_named_t xuiostat_onloan_wbuf;
917 /* whether a copy is made when loaning out a read buffer */
918 kstat_named_t xuiostat_rbuf_copied;
919 kstat_named_t xuiostat_rbuf_nocopy;
920 /* whether a copy is made when assigning a write buffer */
921 kstat_named_t xuiostat_wbuf_copied;
922 kstat_named_t xuiostat_wbuf_nocopy;
925 static xuio_stats_t xuio_stats = {
926 { "onloan_read_buf", KSTAT_DATA_UINT64 },
927 { "onloan_write_buf", KSTAT_DATA_UINT64 },
928 { "read_buf_copied", KSTAT_DATA_UINT64 },
929 { "read_buf_nocopy", KSTAT_DATA_UINT64 },
930 { "write_buf_copied", KSTAT_DATA_UINT64 },
931 { "write_buf_nocopy", KSTAT_DATA_UINT64 }
934 #define XUIOSTAT_INCR(stat, val) \
935 atomic_add_64(&xuio_stats.stat.value.ui64, (val))
936 #define XUIOSTAT_BUMP(stat) XUIOSTAT_INCR(stat, 1)
939 dmu_xuio_init(xuio_t *xuio, int nblk)
942 uio_t *uio = &xuio->xu_uio;
944 uio->uio_iovcnt = nblk;
945 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
947 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
949 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
950 priv->iovp = (iovec_t *)uio->uio_iov;
951 XUIO_XUZC_PRIV(xuio) = priv;
953 if (XUIO_XUZC_RW(xuio) == UIO_READ)
954 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
956 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
962 dmu_xuio_fini(xuio_t *xuio)
964 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
965 int nblk = priv->cnt;
967 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
968 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
969 kmem_free(priv, sizeof (dmu_xuio_t));
971 if (XUIO_XUZC_RW(xuio) == UIO_READ)
972 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
974 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
978 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
979 * and increase priv->next by 1.
982 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
985 uio_t *uio = &xuio->xu_uio;
986 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
987 int i = priv->next++;
989 ASSERT(i < priv->cnt);
990 ASSERT(off + n <= arc_buf_size(abuf));
991 iov = (iovec_t *)uio->uio_iov + i;
992 iov->iov_base = (char *)abuf->b_data + off;
994 priv->bufs[i] = abuf;
999 dmu_xuio_cnt(xuio_t *xuio)
1001 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1006 dmu_xuio_arcbuf(xuio_t *xuio, int i)
1008 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1010 ASSERT(i < priv->cnt);
1011 return (priv->bufs[i]);
1015 dmu_xuio_clear(xuio_t *xuio, int i)
1017 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1019 ASSERT(i < priv->cnt);
1020 priv->bufs[i] = NULL;
1024 xuio_stat_init(void)
1026 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
1027 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
1028 KSTAT_FLAG_VIRTUAL);
1029 if (xuio_ksp != NULL) {
1030 xuio_ksp->ks_data = &xuio_stats;
1031 kstat_install(xuio_ksp);
1036 xuio_stat_fini(void)
1038 if (xuio_ksp != NULL) {
1039 kstat_delete(xuio_ksp);
1045 xuio_stat_wbuf_copied()
1047 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1051 xuio_stat_wbuf_nocopy()
1053 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
1058 dmu_read_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size)
1061 int numbufs, i, err;
1062 xuio_t *xuio = NULL;
1065 * NB: we could do this block-at-a-time, but it's nice
1066 * to be reading in parallel.
1068 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1069 TRUE, FTAG, &numbufs, &dbp, 0);
1073 for (i = 0; i < numbufs; i++) {
1076 dmu_buf_t *db = dbp[i];
1080 bufoff = uio->uio_loffset - db->db_offset;
1081 tocpy = MIN(db->db_size - bufoff, size);
1084 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
1085 arc_buf_t *dbuf_abuf = dbi->db_buf;
1086 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
1087 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
1089 uio->uio_resid -= tocpy;
1090 uio->uio_loffset += tocpy;
1093 if (abuf == dbuf_abuf)
1094 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
1096 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
1098 err = uiomove((char *)db->db_data + bufoff, tocpy,
1106 dmu_buf_rele_array(dbp, numbufs, FTAG);
1112 * Read 'size' bytes into the uio buffer.
1113 * From object zdb->db_object.
1114 * Starting at offset uio->uio_loffset.
1116 * If the caller already has a dbuf in the target object
1117 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1118 * because we don't have to find the dnode_t for the object.
1121 dmu_read_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size)
1123 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1132 err = dmu_read_uio_dnode(dn, uio, size);
1139 * Read 'size' bytes into the uio buffer.
1140 * From the specified object
1141 * Starting at offset uio->uio_loffset.
1144 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
1152 err = dnode_hold(os, object, FTAG, &dn);
1156 err = dmu_read_uio_dnode(dn, uio, size);
1158 dnode_rele(dn, FTAG);
1164 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1171 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1172 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1176 for (i = 0; i < numbufs; i++) {
1179 dmu_buf_t *db = dbp[i];
1183 bufoff = uio->uio_loffset - db->db_offset;
1184 tocpy = MIN(db->db_size - bufoff, size);
1186 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1188 if (tocpy == db->db_size)
1189 dmu_buf_will_fill(db, tx);
1191 dmu_buf_will_dirty(db, tx);
1194 * XXX uiomove could block forever (eg.nfs-backed
1195 * pages). There needs to be a uiolockdown() function
1196 * to lock the pages in memory, so that uiomove won't
1199 err = uiomove((char *)db->db_data + bufoff, tocpy,
1202 if (tocpy == db->db_size)
1203 dmu_buf_fill_done(db, tx);
1211 dmu_buf_rele_array(dbp, numbufs, FTAG);
1216 * Write 'size' bytes from the uio buffer.
1217 * To object zdb->db_object.
1218 * Starting at offset uio->uio_loffset.
1220 * If the caller already has a dbuf in the target object
1221 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1222 * because we don't have to find the dnode_t for the object.
1225 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1228 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1237 err = dmu_write_uio_dnode(dn, uio, size, tx);
1244 * Write 'size' bytes from the uio buffer.
1245 * To the specified object.
1246 * Starting at offset uio->uio_loffset.
1249 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1258 err = dnode_hold(os, object, FTAG, &dn);
1262 err = dmu_write_uio_dnode(dn, uio, size, tx);
1264 dnode_rele(dn, FTAG);
1268 #endif /* _KERNEL */
1271 * Allocate a loaned anonymous arc buffer.
1274 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1276 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1278 return (arc_loan_buf(db->db_objset->os_spa, size));
1282 * Free a loaned arc buffer.
1285 dmu_return_arcbuf(arc_buf_t *buf)
1287 arc_return_buf(buf, FTAG);
1288 VERIFY(arc_buf_remove_ref(buf, FTAG));
1292 * When possible directly assign passed loaned arc buffer to a dbuf.
1293 * If this is not possible copy the contents of passed arc buf via
1297 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1300 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1303 uint32_t blksz = (uint32_t)arc_buf_size(buf);
1306 DB_DNODE_ENTER(dbuf);
1307 dn = DB_DNODE(dbuf);
1308 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1309 blkid = dbuf_whichblock(dn, 0, offset);
1310 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1311 rw_exit(&dn->dn_struct_rwlock);
1312 DB_DNODE_EXIT(dbuf);
1315 * We can only assign if the offset is aligned, the arc buf is the
1316 * same size as the dbuf, and the dbuf is not metadata. It
1317 * can't be metadata because the loaned arc buf comes from the
1318 * user-data kmem area.
1320 if (offset == db->db.db_offset && blksz == db->db.db_size &&
1321 DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA) {
1322 dbuf_assign_arcbuf(db, buf, tx);
1323 dbuf_rele(db, FTAG);
1328 DB_DNODE_ENTER(dbuf);
1329 dn = DB_DNODE(dbuf);
1331 object = dn->dn_object;
1332 DB_DNODE_EXIT(dbuf);
1334 dbuf_rele(db, FTAG);
1335 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1336 dmu_return_arcbuf(buf);
1337 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1342 dbuf_dirty_record_t *dsa_dr;
1343 dmu_sync_cb_t *dsa_done;
1350 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1352 dmu_sync_arg_t *dsa = varg;
1353 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1354 blkptr_t *bp = zio->io_bp;
1356 if (zio->io_error == 0) {
1357 if (BP_IS_HOLE(bp)) {
1359 * A block of zeros may compress to a hole, but the
1360 * block size still needs to be known for replay.
1362 BP_SET_LSIZE(bp, db->db_size);
1363 } else if (!BP_IS_EMBEDDED(bp)) {
1364 ASSERT(BP_GET_LEVEL(bp) == 0);
1371 dmu_sync_late_arrival_ready(zio_t *zio)
1373 dmu_sync_ready(zio, NULL, zio->io_private);
1378 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1380 dmu_sync_arg_t *dsa = varg;
1381 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1382 dmu_buf_impl_t *db = dr->dr_dbuf;
1384 mutex_enter(&db->db_mtx);
1385 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1386 if (zio->io_error == 0) {
1387 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1388 if (dr->dt.dl.dr_nopwrite) {
1389 ASSERTV(blkptr_t *bp = zio->io_bp);
1390 ASSERTV(blkptr_t *bp_orig = &zio->io_bp_orig);
1391 ASSERTV(uint8_t chksum = BP_GET_CHECKSUM(bp_orig));
1393 ASSERT(BP_EQUAL(bp, bp_orig));
1394 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1395 ASSERT(zio_checksum_table[chksum].ci_dedup);
1397 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1398 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1399 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1402 * Old style holes are filled with all zeros, whereas
1403 * new-style holes maintain their lsize, type, level,
1404 * and birth time (see zio_write_compress). While we
1405 * need to reset the BP_SET_LSIZE() call that happened
1406 * in dmu_sync_ready for old style holes, we do *not*
1407 * want to wipe out the information contained in new
1408 * style holes. Thus, only zero out the block pointer if
1409 * it's an old style hole.
1411 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) &&
1412 dr->dt.dl.dr_overridden_by.blk_birth == 0)
1413 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1415 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1417 cv_broadcast(&db->db_changed);
1418 mutex_exit(&db->db_mtx);
1420 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1422 kmem_free(dsa, sizeof (*dsa));
1426 dmu_sync_late_arrival_done(zio_t *zio)
1428 blkptr_t *bp = zio->io_bp;
1429 dmu_sync_arg_t *dsa = zio->io_private;
1430 ASSERTV(blkptr_t *bp_orig = &zio->io_bp_orig);
1432 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1434 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE)
1435 * then there is nothing to do here. Otherwise, free the
1436 * newly allocated block in this txg.
1438 if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
1439 ASSERT(BP_EQUAL(bp, bp_orig));
1441 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1442 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1443 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1444 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1448 dmu_tx_commit(dsa->dsa_tx);
1450 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1452 kmem_free(dsa, sizeof (*dsa));
1456 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1457 zio_prop_t *zp, zbookmark_phys_t *zb)
1459 dmu_sync_arg_t *dsa;
1462 tx = dmu_tx_create(os);
1463 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1464 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1466 /* Make zl_get_data do txg_waited_synced() */
1467 return (SET_ERROR(EIO));
1470 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1472 dsa->dsa_done = done;
1476 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1477 zgd->zgd_db->db_data, zgd->zgd_db->db_size, zp,
1478 dmu_sync_late_arrival_ready, NULL, dmu_sync_late_arrival_done, dsa,
1479 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL|ZIO_FLAG_FASTWRITE, zb));
1485 * Intent log support: sync the block associated with db to disk.
1486 * N.B. and XXX: the caller is responsible for making sure that the
1487 * data isn't changing while dmu_sync() is writing it.
1491 * EEXIST: this txg has already been synced, so there's nothing to do.
1492 * The caller should not log the write.
1494 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1495 * The caller should not log the write.
1497 * EALREADY: this block is already in the process of being synced.
1498 * The caller should track its progress (somehow).
1500 * EIO: could not do the I/O.
1501 * The caller should do a txg_wait_synced().
1503 * 0: the I/O has been initiated.
1504 * The caller should log this blkptr in the done callback.
1505 * It is possible that the I/O will fail, in which case
1506 * the error will be reported to the done callback and
1507 * propagated to pio from zio_done().
1510 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1512 blkptr_t *bp = zgd->zgd_bp;
1513 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1514 objset_t *os = db->db_objset;
1515 dsl_dataset_t *ds = os->os_dsl_dataset;
1516 dbuf_dirty_record_t *dr;
1517 dmu_sync_arg_t *dsa;
1518 zbookmark_phys_t zb;
1522 ASSERT(pio != NULL);
1525 SET_BOOKMARK(&zb, ds->ds_object,
1526 db->db.db_object, db->db_level, db->db_blkid);
1530 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1534 * If we're frozen (running ziltest), we always need to generate a bp.
1536 if (txg > spa_freeze_txg(os->os_spa))
1537 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1540 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1541 * and us. If we determine that this txg is not yet syncing,
1542 * but it begins to sync a moment later, that's OK because the
1543 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1545 mutex_enter(&db->db_mtx);
1547 if (txg <= spa_last_synced_txg(os->os_spa)) {
1549 * This txg has already synced. There's nothing to do.
1551 mutex_exit(&db->db_mtx);
1552 return (SET_ERROR(EEXIST));
1555 if (txg <= spa_syncing_txg(os->os_spa)) {
1557 * This txg is currently syncing, so we can't mess with
1558 * the dirty record anymore; just write a new log block.
1560 mutex_exit(&db->db_mtx);
1561 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1564 dr = db->db_last_dirty;
1565 while (dr && dr->dr_txg != txg)
1570 * There's no dr for this dbuf, so it must have been freed.
1571 * There's no need to log writes to freed blocks, so we're done.
1573 mutex_exit(&db->db_mtx);
1574 return (SET_ERROR(ENOENT));
1577 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
1580 * Assume the on-disk data is X, the current syncing data (in
1581 * txg - 1) is Y, and the current in-memory data is Z (currently
1584 * We usually want to perform a nopwrite if X and Z are the
1585 * same. However, if Y is different (i.e. the BP is going to
1586 * change before this write takes effect), then a nopwrite will
1587 * be incorrect - we would override with X, which could have
1588 * been freed when Y was written.
1590 * (Note that this is not a concern when we are nop-writing from
1591 * syncing context, because X and Y must be identical, because
1592 * all previous txgs have been synced.)
1594 * Therefore, we disable nopwrite if the current BP could change
1595 * before this TXG. There are two ways it could change: by
1596 * being dirty (dr_next is non-NULL), or by being freed
1597 * (dnode_block_freed()). This behavior is verified by
1598 * zio_done(), which VERIFYs that the override BP is identical
1599 * to the on-disk BP.
1603 if (dr->dr_next != NULL || dnode_block_freed(dn, db->db_blkid))
1604 zp.zp_nopwrite = B_FALSE;
1607 ASSERT(dr->dr_txg == txg);
1608 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1609 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1611 * We have already issued a sync write for this buffer,
1612 * or this buffer has already been synced. It could not
1613 * have been dirtied since, or we would have cleared the state.
1615 mutex_exit(&db->db_mtx);
1616 return (SET_ERROR(EALREADY));
1619 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1620 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1621 mutex_exit(&db->db_mtx);
1623 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1625 dsa->dsa_done = done;
1629 zio_nowait(arc_write(pio, os->os_spa, txg,
1630 bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
1631 DBUF_IS_L2COMPRESSIBLE(db), &zp, dmu_sync_ready,
1632 NULL, dmu_sync_done, dsa, ZIO_PRIORITY_SYNC_WRITE,
1633 ZIO_FLAG_CANFAIL, &zb));
1639 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1645 err = dnode_hold(os, object, FTAG, &dn);
1648 err = dnode_set_blksz(dn, size, ibs, tx);
1649 dnode_rele(dn, FTAG);
1654 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1660 * Send streams include each object's checksum function. This
1661 * check ensures that the receiving system can understand the
1662 * checksum function transmitted.
1664 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS);
1666 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1667 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS);
1668 dn->dn_checksum = checksum;
1669 dnode_setdirty(dn, tx);
1670 dnode_rele(dn, FTAG);
1674 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1680 * Send streams include each object's compression function. This
1681 * check ensures that the receiving system can understand the
1682 * compression function transmitted.
1684 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS);
1686 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1687 dn->dn_compress = compress;
1688 dnode_setdirty(dn, tx);
1689 dnode_rele(dn, FTAG);
1692 int zfs_mdcomp_disable = 0;
1695 * When the "redundant_metadata" property is set to "most", only indirect
1696 * blocks of this level and higher will have an additional ditto block.
1698 int zfs_redundant_metadata_most_ditto_level = 2;
1701 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
1703 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
1704 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
1706 enum zio_checksum checksum = os->os_checksum;
1707 enum zio_compress compress = os->os_compress;
1708 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
1709 boolean_t dedup = B_FALSE;
1710 boolean_t nopwrite = B_FALSE;
1711 boolean_t dedup_verify = os->os_dedup_verify;
1712 int copies = os->os_copies;
1715 * We maintain different write policies for each of the following
1718 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1719 * 3. all other level 0 blocks
1722 if (zfs_mdcomp_disable) {
1723 compress = ZIO_COMPRESS_EMPTY;
1726 * XXX -- we should design a compression algorithm
1727 * that specializes in arrays of bps.
1729 compress = zio_compress_select(os->os_spa,
1730 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON);
1734 * Metadata always gets checksummed. If the data
1735 * checksum is multi-bit correctable, and it's not a
1736 * ZBT-style checksum, then it's suitable for metadata
1737 * as well. Otherwise, the metadata checksum defaults
1740 if (zio_checksum_table[checksum].ci_correctable < 1 ||
1741 zio_checksum_table[checksum].ci_eck)
1742 checksum = ZIO_CHECKSUM_FLETCHER_4;
1744 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL ||
1745 (os->os_redundant_metadata ==
1746 ZFS_REDUNDANT_METADATA_MOST &&
1747 (level >= zfs_redundant_metadata_most_ditto_level ||
1748 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))))
1750 } else if (wp & WP_NOFILL) {
1754 * If we're writing preallocated blocks, we aren't actually
1755 * writing them so don't set any policy properties. These
1756 * blocks are currently only used by an external subsystem
1757 * outside of zfs (i.e. dump) and not written by the zio
1760 compress = ZIO_COMPRESS_OFF;
1761 checksum = ZIO_CHECKSUM_OFF;
1763 compress = zio_compress_select(os->os_spa, dn->dn_compress,
1766 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
1767 zio_checksum_select(dn->dn_checksum, checksum) :
1771 * Determine dedup setting. If we are in dmu_sync(),
1772 * we won't actually dedup now because that's all
1773 * done in syncing context; but we do want to use the
1774 * dedup checkum. If the checksum is not strong
1775 * enough to ensure unique signatures, force
1778 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
1779 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
1780 if (!zio_checksum_table[checksum].ci_dedup)
1781 dedup_verify = B_TRUE;
1785 * Enable nopwrite if we have a cryptographically secure
1786 * checksum that has no known collisions (i.e. SHA-256)
1787 * and compression is enabled. We don't enable nopwrite if
1788 * dedup is enabled as the two features are mutually exclusive.
1790 nopwrite = (!dedup && zio_checksum_table[checksum].ci_dedup &&
1791 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
1794 zp->zp_checksum = checksum;
1795 zp->zp_compress = compress;
1796 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
1797 zp->zp_level = level;
1798 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
1799 zp->zp_dedup = dedup;
1800 zp->zp_dedup_verify = dedup && dedup_verify;
1801 zp->zp_nopwrite = nopwrite;
1805 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
1810 err = dnode_hold(os, object, FTAG, &dn);
1814 * Sync any current changes before
1815 * we go trundling through the block pointers.
1817 for (i = 0; i < TXG_SIZE; i++) {
1818 if (list_link_active(&dn->dn_dirty_link[i]))
1821 if (i != TXG_SIZE) {
1822 dnode_rele(dn, FTAG);
1823 txg_wait_synced(dmu_objset_pool(os), 0);
1824 err = dnode_hold(os, object, FTAG, &dn);
1829 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
1830 dnode_rele(dn, FTAG);
1836 __dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
1838 dnode_phys_t *dnp = dn->dn_phys;
1841 doi->doi_data_block_size = dn->dn_datablksz;
1842 doi->doi_metadata_block_size = dn->dn_indblkshift ?
1843 1ULL << dn->dn_indblkshift : 0;
1844 doi->doi_type = dn->dn_type;
1845 doi->doi_bonus_type = dn->dn_bonustype;
1846 doi->doi_bonus_size = dn->dn_bonuslen;
1847 doi->doi_indirection = dn->dn_nlevels;
1848 doi->doi_checksum = dn->dn_checksum;
1849 doi->doi_compress = dn->dn_compress;
1850 doi->doi_nblkptr = dn->dn_nblkptr;
1851 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
1852 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
1853 doi->doi_fill_count = 0;
1854 for (i = 0; i < dnp->dn_nblkptr; i++)
1855 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
1859 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
1861 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1862 mutex_enter(&dn->dn_mtx);
1864 __dmu_object_info_from_dnode(dn, doi);
1866 mutex_exit(&dn->dn_mtx);
1867 rw_exit(&dn->dn_struct_rwlock);
1871 * Get information on a DMU object.
1872 * If doi is NULL, just indicates whether the object exists.
1875 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
1878 int err = dnode_hold(os, object, FTAG, &dn);
1884 dmu_object_info_from_dnode(dn, doi);
1886 dnode_rele(dn, FTAG);
1891 * As above, but faster; can be used when you have a held dbuf in hand.
1894 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
1896 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1899 dmu_object_info_from_dnode(DB_DNODE(db), doi);
1904 * Faster still when you only care about the size.
1905 * This is specifically optimized for zfs_getattr().
1908 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
1909 u_longlong_t *nblk512)
1911 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1917 *blksize = dn->dn_datablksz;
1918 /* add 1 for dnode space */
1919 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
1920 SPA_MINBLOCKSHIFT) + 1;
1925 byteswap_uint64_array(void *vbuf, size_t size)
1927 uint64_t *buf = vbuf;
1928 size_t count = size >> 3;
1931 ASSERT((size & 7) == 0);
1933 for (i = 0; i < count; i++)
1934 buf[i] = BSWAP_64(buf[i]);
1938 byteswap_uint32_array(void *vbuf, size_t size)
1940 uint32_t *buf = vbuf;
1941 size_t count = size >> 2;
1944 ASSERT((size & 3) == 0);
1946 for (i = 0; i < count; i++)
1947 buf[i] = BSWAP_32(buf[i]);
1951 byteswap_uint16_array(void *vbuf, size_t size)
1953 uint16_t *buf = vbuf;
1954 size_t count = size >> 1;
1957 ASSERT((size & 1) == 0);
1959 for (i = 0; i < count; i++)
1960 buf[i] = BSWAP_16(buf[i]);
1965 byteswap_uint8_array(void *vbuf, size_t size)
1987 arc_fini(); /* arc depends on l2arc, so arc must go first */
1999 #if defined(_KERNEL) && defined(HAVE_SPL)
2000 EXPORT_SYMBOL(dmu_bonus_hold);
2001 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus);
2002 EXPORT_SYMBOL(dmu_buf_rele_array);
2003 EXPORT_SYMBOL(dmu_prefetch);
2004 EXPORT_SYMBOL(dmu_free_range);
2005 EXPORT_SYMBOL(dmu_free_long_range);
2006 EXPORT_SYMBOL(dmu_free_long_object);
2007 EXPORT_SYMBOL(dmu_read);
2008 EXPORT_SYMBOL(dmu_write);
2009 EXPORT_SYMBOL(dmu_prealloc);
2010 EXPORT_SYMBOL(dmu_object_info);
2011 EXPORT_SYMBOL(dmu_object_info_from_dnode);
2012 EXPORT_SYMBOL(dmu_object_info_from_db);
2013 EXPORT_SYMBOL(dmu_object_size_from_db);
2014 EXPORT_SYMBOL(dmu_object_set_blocksize);
2015 EXPORT_SYMBOL(dmu_object_set_checksum);
2016 EXPORT_SYMBOL(dmu_object_set_compress);
2017 EXPORT_SYMBOL(dmu_write_policy);
2018 EXPORT_SYMBOL(dmu_sync);
2019 EXPORT_SYMBOL(dmu_request_arcbuf);
2020 EXPORT_SYMBOL(dmu_return_arcbuf);
2021 EXPORT_SYMBOL(dmu_assign_arcbuf);
2022 EXPORT_SYMBOL(dmu_buf_hold);
2023 EXPORT_SYMBOL(dmu_ot);
2025 module_param(zfs_mdcomp_disable, int, 0644);
2026 MODULE_PARM_DESC(zfs_mdcomp_disable, "Disable meta data compression");
2028 module_param(zfs_nopwrite_enabled, int, 0644);
2029 MODULE_PARM_DESC(zfs_nopwrite_enabled, "Enable NOP writes");