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.
26 #include <sys/dmu_impl.h>
27 #include <sys/dmu_tx.h>
29 #include <sys/dnode.h>
30 #include <sys/zfs_context.h>
31 #include <sys/dmu_objset.h>
32 #include <sys/dmu_traverse.h>
33 #include <sys/dsl_dataset.h>
34 #include <sys/dsl_dir.h>
35 #include <sys/dsl_pool.h>
36 #include <sys/dsl_synctask.h>
37 #include <sys/dsl_prop.h>
38 #include <sys/dmu_zfetch.h>
39 #include <sys/zfs_ioctl.h>
41 #include <sys/zio_checksum.h>
44 #include <sys/zfs_znode.h>
47 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
48 { byteswap_uint8_array, TRUE, "unallocated" },
49 { zap_byteswap, TRUE, "object directory" },
50 { byteswap_uint64_array, TRUE, "object array" },
51 { byteswap_uint8_array, TRUE, "packed nvlist" },
52 { byteswap_uint64_array, TRUE, "packed nvlist size" },
53 { byteswap_uint64_array, TRUE, "bpobj" },
54 { byteswap_uint64_array, TRUE, "bpobj header" },
55 { byteswap_uint64_array, TRUE, "SPA space map header" },
56 { byteswap_uint64_array, TRUE, "SPA space map" },
57 { byteswap_uint64_array, TRUE, "ZIL intent log" },
58 { dnode_buf_byteswap, TRUE, "DMU dnode" },
59 { dmu_objset_byteswap, TRUE, "DMU objset" },
60 { byteswap_uint64_array, TRUE, "DSL directory" },
61 { zap_byteswap, TRUE, "DSL directory child map"},
62 { zap_byteswap, TRUE, "DSL dataset snap map" },
63 { zap_byteswap, TRUE, "DSL props" },
64 { byteswap_uint64_array, TRUE, "DSL dataset" },
65 { zfs_znode_byteswap, TRUE, "ZFS znode" },
66 { zfs_oldacl_byteswap, TRUE, "ZFS V0 ACL" },
67 { byteswap_uint8_array, FALSE, "ZFS plain file" },
68 { zap_byteswap, TRUE, "ZFS directory" },
69 { zap_byteswap, TRUE, "ZFS master node" },
70 { zap_byteswap, TRUE, "ZFS delete queue" },
71 { byteswap_uint8_array, FALSE, "zvol object" },
72 { zap_byteswap, TRUE, "zvol prop" },
73 { byteswap_uint8_array, FALSE, "other uint8[]" },
74 { byteswap_uint64_array, FALSE, "other uint64[]" },
75 { zap_byteswap, TRUE, "other ZAP" },
76 { zap_byteswap, TRUE, "persistent error log" },
77 { byteswap_uint8_array, TRUE, "SPA history" },
78 { byteswap_uint64_array, TRUE, "SPA history offsets" },
79 { zap_byteswap, TRUE, "Pool properties" },
80 { zap_byteswap, TRUE, "DSL permissions" },
81 { zfs_acl_byteswap, TRUE, "ZFS ACL" },
82 { byteswap_uint8_array, TRUE, "ZFS SYSACL" },
83 { byteswap_uint8_array, TRUE, "FUID table" },
84 { byteswap_uint64_array, TRUE, "FUID table size" },
85 { zap_byteswap, TRUE, "DSL dataset next clones"},
86 { zap_byteswap, TRUE, "scan work queue" },
87 { zap_byteswap, TRUE, "ZFS user/group used" },
88 { zap_byteswap, TRUE, "ZFS user/group quota" },
89 { zap_byteswap, TRUE, "snapshot refcount tags"},
90 { zap_byteswap, TRUE, "DDT ZAP algorithm" },
91 { zap_byteswap, TRUE, "DDT statistics" },
92 { byteswap_uint8_array, TRUE, "System attributes" },
93 { zap_byteswap, TRUE, "SA master node" },
94 { zap_byteswap, TRUE, "SA attr registration" },
95 { zap_byteswap, TRUE, "SA attr layouts" },
96 { zap_byteswap, TRUE, "scan translations" },
97 { byteswap_uint8_array, FALSE, "deduplicated block" },
98 { zap_byteswap, TRUE, "DSL deadlist map" },
99 { byteswap_uint64_array, TRUE, "DSL deadlist map hdr" },
100 { zap_byteswap, TRUE, "DSL dir clones" },
101 { byteswap_uint64_array, TRUE, "bpobj subobj" },
105 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
106 void *tag, dmu_buf_t **dbp, int flags)
112 int db_flags = DB_RF_CANFAIL;
114 if (flags & DMU_READ_NO_PREFETCH)
115 db_flags |= DB_RF_NOPREFETCH;
117 err = dnode_hold(os, object, FTAG, &dn);
120 blkid = dbuf_whichblock(dn, offset);
121 rw_enter(&dn->dn_struct_rwlock, RW_READER);
122 db = dbuf_hold(dn, blkid, tag);
123 rw_exit(&dn->dn_struct_rwlock);
127 err = dbuf_read(db, NULL, db_flags);
134 dnode_rele(dn, FTAG);
135 *dbp = &db->db; /* NULL db plus first field offset is NULL */
142 return (DN_MAX_BONUSLEN);
146 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
148 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
155 if (dn->dn_bonus != db) {
157 } else if (newsize < 0 || newsize > db_fake->db_size) {
160 dnode_setbonuslen(dn, newsize, tx);
169 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
171 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
178 if (type > DMU_OT_NUMTYPES) {
180 } else if (dn->dn_bonus != db) {
183 dnode_setbonus_type(dn, type, tx);
192 dmu_get_bonustype(dmu_buf_t *db_fake)
194 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
196 dmu_object_type_t type;
200 type = dn->dn_bonustype;
207 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
212 error = dnode_hold(os, object, FTAG, &dn);
213 dbuf_rm_spill(dn, tx);
214 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
215 dnode_rm_spill(dn, tx);
216 rw_exit(&dn->dn_struct_rwlock);
217 dnode_rele(dn, FTAG);
222 * returns ENOENT, EIO, or 0.
225 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
231 error = dnode_hold(os, object, FTAG, &dn);
235 rw_enter(&dn->dn_struct_rwlock, RW_READER);
236 if (dn->dn_bonus == NULL) {
237 rw_exit(&dn->dn_struct_rwlock);
238 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
239 if (dn->dn_bonus == NULL)
240 dbuf_create_bonus(dn);
244 /* as long as the bonus buf is held, the dnode will be held */
245 if (refcount_add(&db->db_holds, tag) == 1) {
246 VERIFY(dnode_add_ref(dn, db));
247 (void) atomic_inc_32_nv(&dn->dn_dbufs_count);
251 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
252 * hold and incrementing the dbuf count to ensure that dnode_move() sees
253 * a dnode hold for every dbuf.
255 rw_exit(&dn->dn_struct_rwlock);
257 dnode_rele(dn, FTAG);
259 VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH));
266 * returns ENOENT, EIO, or 0.
268 * This interface will allocate a blank spill dbuf when a spill blk
269 * doesn't already exist on the dnode.
271 * if you only want to find an already existing spill db, then
272 * dmu_spill_hold_existing() should be used.
275 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
277 dmu_buf_impl_t *db = NULL;
280 if ((flags & DB_RF_HAVESTRUCT) == 0)
281 rw_enter(&dn->dn_struct_rwlock, RW_READER);
283 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
285 if ((flags & DB_RF_HAVESTRUCT) == 0)
286 rw_exit(&dn->dn_struct_rwlock);
289 err = dbuf_read(db, NULL, flags);
298 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
300 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
307 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
310 rw_enter(&dn->dn_struct_rwlock, RW_READER);
312 if (!dn->dn_have_spill) {
315 err = dmu_spill_hold_by_dnode(dn,
316 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
319 rw_exit(&dn->dn_struct_rwlock);
327 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
329 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
335 err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp);
342 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
343 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
344 * and can induce severe lock contention when writing to several files
345 * whose dnodes are in the same block.
348 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
349 int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
351 dsl_pool_t *dp = NULL;
353 uint64_t blkid, nblks, i;
359 ASSERT(length <= DMU_MAX_ACCESS);
361 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT;
362 if (flags & DMU_READ_NO_PREFETCH || length > zfetch_array_rd_sz)
363 dbuf_flags |= DB_RF_NOPREFETCH;
365 rw_enter(&dn->dn_struct_rwlock, RW_READER);
366 if (dn->dn_datablkshift) {
367 int blkshift = dn->dn_datablkshift;
368 nblks = (P2ROUNDUP(offset+length, 1ULL<<blkshift) -
369 P2ALIGN(offset, 1ULL<<blkshift)) >> blkshift;
371 if (offset + length > dn->dn_datablksz) {
372 zfs_panic_recover("zfs: accessing past end of object "
373 "%llx/%llx (size=%u access=%llu+%llu)",
374 (longlong_t)dn->dn_objset->
375 os_dsl_dataset->ds_object,
376 (longlong_t)dn->dn_object, dn->dn_datablksz,
377 (longlong_t)offset, (longlong_t)length);
378 rw_exit(&dn->dn_struct_rwlock);
383 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
385 if (dn->dn_objset->os_dsl_dataset)
386 dp = dn->dn_objset->os_dsl_dataset->ds_dir->dd_pool;
387 if (dp && dsl_pool_sync_context(dp))
389 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
390 blkid = dbuf_whichblock(dn, offset);
391 for (i = 0; i < nblks; i++) {
392 dmu_buf_impl_t *db = dbuf_hold(dn, blkid+i, tag);
394 rw_exit(&dn->dn_struct_rwlock);
395 dmu_buf_rele_array(dbp, nblks, tag);
399 /* initiate async i/o */
401 (void) dbuf_read(db, zio, dbuf_flags);
405 rw_exit(&dn->dn_struct_rwlock);
407 /* wait for async i/o */
409 /* track read overhead when we are in sync context */
410 if (dp && dsl_pool_sync_context(dp))
411 dp->dp_read_overhead += gethrtime() - start;
413 dmu_buf_rele_array(dbp, nblks, tag);
417 /* wait for other io to complete */
419 for (i = 0; i < nblks; i++) {
420 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
421 mutex_enter(&db->db_mtx);
422 while (db->db_state == DB_READ ||
423 db->db_state == DB_FILL)
424 cv_wait(&db->db_changed, &db->db_mtx);
425 if (db->db_state == DB_UNCACHED)
427 mutex_exit(&db->db_mtx);
429 dmu_buf_rele_array(dbp, nblks, tag);
441 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
442 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
447 err = dnode_hold(os, object, FTAG, &dn);
451 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
452 numbufsp, dbpp, DMU_READ_PREFETCH);
454 dnode_rele(dn, FTAG);
460 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
461 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
463 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
469 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
470 numbufsp, dbpp, DMU_READ_PREFETCH);
477 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
480 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
485 for (i = 0; i < numbufs; i++) {
487 dbuf_rele(dbp[i], tag);
490 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
494 dmu_prefetch(objset_t *os, uint64_t object, uint64_t offset, uint64_t len)
500 if (zfs_prefetch_disable)
503 if (len == 0) { /* they're interested in the bonus buffer */
504 dn = DMU_META_DNODE(os);
506 if (object == 0 || object >= DN_MAX_OBJECT)
509 rw_enter(&dn->dn_struct_rwlock, RW_READER);
510 blkid = dbuf_whichblock(dn, object * sizeof (dnode_phys_t));
511 dbuf_prefetch(dn, blkid);
512 rw_exit(&dn->dn_struct_rwlock);
517 * XXX - Note, if the dnode for the requested object is not
518 * already cached, we will do a *synchronous* read in the
519 * dnode_hold() call. The same is true for any indirects.
521 err = dnode_hold(os, object, FTAG, &dn);
525 rw_enter(&dn->dn_struct_rwlock, RW_READER);
526 if (dn->dn_datablkshift) {
527 int blkshift = dn->dn_datablkshift;
528 nblks = (P2ROUNDUP(offset+len, 1<<blkshift) -
529 P2ALIGN(offset, 1<<blkshift)) >> blkshift;
531 nblks = (offset < dn->dn_datablksz);
535 blkid = dbuf_whichblock(dn, offset);
536 for (i = 0; i < nblks; i++)
537 dbuf_prefetch(dn, blkid+i);
540 rw_exit(&dn->dn_struct_rwlock);
542 dnode_rele(dn, FTAG);
546 * Get the next "chunk" of file data to free. We traverse the file from
547 * the end so that the file gets shorter over time (if we crashes in the
548 * middle, this will leave us in a better state). We find allocated file
549 * data by simply searching the allocated level 1 indirects.
552 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t limit)
554 uint64_t len = *start - limit;
556 uint64_t maxblks = DMU_MAX_ACCESS / (1ULL << (dn->dn_indblkshift + 1));
558 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
560 ASSERT(limit <= *start);
562 if (len <= iblkrange * maxblks) {
566 ASSERT(ISP2(iblkrange));
568 while (*start > limit && blkcnt < maxblks) {
571 /* find next allocated L1 indirect */
572 err = dnode_next_offset(dn,
573 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
575 /* if there are no more, then we are done */
584 /* reset offset to end of "next" block back */
585 *start = P2ALIGN(*start, iblkrange);
595 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
596 uint64_t length, boolean_t free_dnode)
599 uint64_t object_size, start, end, len;
600 boolean_t trunc = (length == DMU_OBJECT_END);
603 align = 1 << dn->dn_datablkshift;
605 object_size = align == 1 ? dn->dn_datablksz :
606 (dn->dn_maxblkid + 1) << dn->dn_datablkshift;
608 end = offset + length;
609 if (trunc || end > object_size)
613 length = end - offset;
617 /* assert(offset <= start) */
618 err = get_next_chunk(dn, &start, offset);
621 len = trunc ? DMU_OBJECT_END : end - start;
623 tx = dmu_tx_create(os);
624 dmu_tx_hold_free(tx, dn->dn_object, start, len);
625 err = dmu_tx_assign(tx, TXG_WAIT);
631 dnode_free_range(dn, start, trunc ? -1 : len, tx);
633 if (start == 0 && free_dnode) {
638 length -= end - start;
647 dmu_free_long_range(objset_t *os, uint64_t object,
648 uint64_t offset, uint64_t length)
653 err = dnode_hold(os, object, FTAG, &dn);
656 err = dmu_free_long_range_impl(os, dn, offset, length, FALSE);
657 dnode_rele(dn, FTAG);
662 dmu_free_object(objset_t *os, uint64_t object)
668 err = dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED,
672 if (dn->dn_nlevels == 1) {
673 tx = dmu_tx_create(os);
674 dmu_tx_hold_bonus(tx, object);
675 dmu_tx_hold_free(tx, dn->dn_object, 0, DMU_OBJECT_END);
676 err = dmu_tx_assign(tx, TXG_WAIT);
678 dnode_free_range(dn, 0, DMU_OBJECT_END, tx);
685 err = dmu_free_long_range_impl(os, dn, 0, DMU_OBJECT_END, TRUE);
687 dnode_rele(dn, FTAG);
692 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
693 uint64_t size, dmu_tx_t *tx)
696 int err = dnode_hold(os, object, FTAG, &dn);
699 ASSERT(offset < UINT64_MAX);
700 ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
701 dnode_free_range(dn, offset, size, tx);
702 dnode_rele(dn, FTAG);
707 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
708 void *buf, uint32_t flags)
714 err = dnode_hold(os, object, FTAG, &dn);
719 * Deal with odd block sizes, where there can't be data past the first
720 * block. If we ever do the tail block optimization, we will need to
721 * handle that here as well.
723 if (dn->dn_maxblkid == 0) {
724 int newsz = offset > dn->dn_datablksz ? 0 :
725 MIN(size, dn->dn_datablksz - offset);
726 bzero((char *)buf + newsz, size - newsz);
731 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
735 * NB: we could do this block-at-a-time, but it's nice
736 * to be reading in parallel.
738 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
739 TRUE, FTAG, &numbufs, &dbp, flags);
743 for (i = 0; i < numbufs; i++) {
746 dmu_buf_t *db = dbp[i];
750 bufoff = offset - db->db_offset;
751 tocpy = (int)MIN(db->db_size - bufoff, size);
753 bcopy((char *)db->db_data + bufoff, buf, tocpy);
757 buf = (char *)buf + tocpy;
759 dmu_buf_rele_array(dbp, numbufs, FTAG);
761 dnode_rele(dn, FTAG);
766 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
767 const void *buf, dmu_tx_t *tx)
775 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
776 FALSE, FTAG, &numbufs, &dbp));
778 for (i = 0; i < numbufs; i++) {
781 dmu_buf_t *db = dbp[i];
785 bufoff = offset - db->db_offset;
786 tocpy = (int)MIN(db->db_size - bufoff, size);
788 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
790 if (tocpy == db->db_size)
791 dmu_buf_will_fill(db, tx);
793 dmu_buf_will_dirty(db, tx);
795 bcopy(buf, (char *)db->db_data + bufoff, tocpy);
797 if (tocpy == db->db_size)
798 dmu_buf_fill_done(db, tx);
802 buf = (char *)buf + tocpy;
804 dmu_buf_rele_array(dbp, numbufs, FTAG);
808 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
817 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
818 FALSE, FTAG, &numbufs, &dbp));
820 for (i = 0; i < numbufs; i++) {
821 dmu_buf_t *db = dbp[i];
823 dmu_buf_will_not_fill(db, tx);
825 dmu_buf_rele_array(dbp, numbufs, FTAG);
829 * DMU support for xuio
831 kstat_t *xuio_ksp = NULL;
834 dmu_xuio_init(xuio_t *xuio, int nblk)
837 uio_t *uio = &xuio->xu_uio;
839 uio->uio_iovcnt = nblk;
840 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
842 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
844 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
845 priv->iovp = uio->uio_iov;
846 XUIO_XUZC_PRIV(xuio) = priv;
848 if (XUIO_XUZC_RW(xuio) == UIO_READ)
849 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
851 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
857 dmu_xuio_fini(xuio_t *xuio)
859 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
860 int nblk = priv->cnt;
862 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
863 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
864 kmem_free(priv, sizeof (dmu_xuio_t));
866 if (XUIO_XUZC_RW(xuio) == UIO_READ)
867 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
869 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
873 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
874 * and increase priv->next by 1.
877 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
880 uio_t *uio = &xuio->xu_uio;
881 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
882 int i = priv->next++;
884 ASSERT(i < priv->cnt);
885 ASSERT(off + n <= arc_buf_size(abuf));
886 iov = uio->uio_iov + i;
887 iov->iov_base = (char *)abuf->b_data + off;
889 priv->bufs[i] = abuf;
894 dmu_xuio_cnt(xuio_t *xuio)
896 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
901 dmu_xuio_arcbuf(xuio_t *xuio, int i)
903 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
905 ASSERT(i < priv->cnt);
906 return (priv->bufs[i]);
910 dmu_xuio_clear(xuio_t *xuio, int i)
912 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
914 ASSERT(i < priv->cnt);
915 priv->bufs[i] = NULL;
921 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
922 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
924 if (xuio_ksp != NULL) {
925 xuio_ksp->ks_data = &xuio_stats;
926 kstat_install(xuio_ksp);
933 if (xuio_ksp != NULL) {
934 kstat_delete(xuio_ksp);
940 xuio_stat_wbuf_copied()
942 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
946 xuio_stat_wbuf_nocopy()
948 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
953 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
960 * NB: we could do this block-at-a-time, but it's nice
961 * to be reading in parallel.
963 err = dmu_buf_hold_array(os, object, uio->uio_loffset, size, TRUE, FTAG,
969 if (uio->uio_extflg == UIO_XUIO)
970 xuio = (xuio_t *)uio;
973 for (i = 0; i < numbufs; i++) {
976 dmu_buf_t *db = dbp[i];
980 bufoff = uio->uio_loffset - db->db_offset;
981 tocpy = (int)MIN(db->db_size - bufoff, size);
984 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
985 arc_buf_t *dbuf_abuf = dbi->db_buf;
986 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
987 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
989 uio->uio_resid -= tocpy;
990 uio->uio_loffset += tocpy;
993 if (abuf == dbuf_abuf)
994 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
996 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
998 err = uiomove((char *)db->db_data + bufoff, tocpy,
1006 dmu_buf_rele_array(dbp, numbufs, FTAG);
1012 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1019 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1020 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1024 for (i = 0; i < numbufs; i++) {
1027 dmu_buf_t *db = dbp[i];
1031 bufoff = uio->uio_loffset - db->db_offset;
1032 tocpy = (int)MIN(db->db_size - bufoff, size);
1034 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1036 if (tocpy == db->db_size)
1037 dmu_buf_will_fill(db, tx);
1039 dmu_buf_will_dirty(db, tx);
1042 * XXX uiomove could block forever (eg. nfs-backed
1043 * pages). There needs to be a uiolockdown() function
1044 * to lock the pages in memory, so that uiomove won't
1047 err = uiomove((char *)db->db_data + bufoff, tocpy,
1050 if (tocpy == db->db_size)
1051 dmu_buf_fill_done(db, tx);
1059 dmu_buf_rele_array(dbp, numbufs, FTAG);
1064 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1067 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1076 err = dmu_write_uio_dnode(dn, uio, size, tx);
1083 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1092 err = dnode_hold(os, object, FTAG, &dn);
1096 err = dmu_write_uio_dnode(dn, uio, size, tx);
1098 dnode_rele(dn, FTAG);
1105 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1106 page_t *pp, dmu_tx_t *tx)
1115 err = dmu_buf_hold_array(os, object, offset, size,
1116 FALSE, FTAG, &numbufs, &dbp);
1120 for (i = 0; i < numbufs; i++) {
1121 int tocpy, copied, thiscpy;
1123 dmu_buf_t *db = dbp[i];
1127 ASSERT3U(db->db_size, >=, PAGESIZE);
1129 bufoff = offset - db->db_offset;
1130 tocpy = (int)MIN(db->db_size - bufoff, size);
1132 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1134 if (tocpy == db->db_size)
1135 dmu_buf_will_fill(db, tx);
1137 dmu_buf_will_dirty(db, tx);
1139 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1140 ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff);
1141 thiscpy = MIN(PAGESIZE, tocpy - copied);
1142 va = zfs_map_page(pp, S_READ);
1143 bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1144 zfs_unmap_page(pp, va);
1149 if (tocpy == db->db_size)
1150 dmu_buf_fill_done(db, tx);
1155 dmu_buf_rele_array(dbp, numbufs, FTAG);
1162 * Allocate a loaned anonymous arc buffer.
1165 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1167 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1170 DB_GET_SPA(&spa, db);
1171 return (arc_loan_buf(spa, size));
1175 * Free a loaned arc buffer.
1178 dmu_return_arcbuf(arc_buf_t *buf)
1180 arc_return_buf(buf, FTAG);
1181 VERIFY(arc_buf_remove_ref(buf, FTAG) == 1);
1185 * When possible directly assign passed loaned arc buffer to a dbuf.
1186 * If this is not possible copy the contents of passed arc buf via
1190 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1193 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1196 uint32_t blksz = (uint32_t)arc_buf_size(buf);
1199 DB_DNODE_ENTER(dbuf);
1200 dn = DB_DNODE(dbuf);
1201 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1202 blkid = dbuf_whichblock(dn, offset);
1203 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1204 rw_exit(&dn->dn_struct_rwlock);
1205 DB_DNODE_EXIT(dbuf);
1207 if (offset == db->db.db_offset && blksz == db->db.db_size) {
1208 dbuf_assign_arcbuf(db, buf, tx);
1209 dbuf_rele(db, FTAG);
1214 DB_DNODE_ENTER(dbuf);
1215 dn = DB_DNODE(dbuf);
1217 object = dn->dn_object;
1218 DB_DNODE_EXIT(dbuf);
1220 dbuf_rele(db, FTAG);
1221 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1222 dmu_return_arcbuf(buf);
1223 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1228 dbuf_dirty_record_t *dsa_dr;
1229 dmu_sync_cb_t *dsa_done;
1236 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1238 dmu_sync_arg_t *dsa = varg;
1239 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1240 blkptr_t *bp = zio->io_bp;
1242 if (zio->io_error == 0) {
1243 if (BP_IS_HOLE(bp)) {
1245 * A block of zeros may compress to a hole, but the
1246 * block size still needs to be known for replay.
1248 BP_SET_LSIZE(bp, db->db_size);
1250 ASSERT(BP_GET_LEVEL(bp) == 0);
1257 dmu_sync_late_arrival_ready(zio_t *zio)
1259 dmu_sync_ready(zio, NULL, zio->io_private);
1264 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1266 dmu_sync_arg_t *dsa = varg;
1267 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1268 dmu_buf_impl_t *db = dr->dr_dbuf;
1270 mutex_enter(&db->db_mtx);
1271 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1272 if (zio->io_error == 0) {
1273 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1274 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1275 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1276 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by))
1277 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1279 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1281 cv_broadcast(&db->db_changed);
1282 mutex_exit(&db->db_mtx);
1284 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1286 kmem_free(dsa, sizeof (*dsa));
1290 dmu_sync_late_arrival_done(zio_t *zio)
1292 blkptr_t *bp = zio->io_bp;
1293 dmu_sync_arg_t *dsa = zio->io_private;
1295 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1296 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1297 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1298 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1301 dmu_tx_commit(dsa->dsa_tx);
1303 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1305 kmem_free(dsa, sizeof (*dsa));
1309 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1310 zio_prop_t *zp, zbookmark_t *zb)
1312 dmu_sync_arg_t *dsa;
1315 tx = dmu_tx_create(os);
1316 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1317 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1319 return (EIO); /* Make zl_get_data do txg_waited_synced() */
1322 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1324 dsa->dsa_done = done;
1328 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1329 zgd->zgd_db->db_data, zgd->zgd_db->db_size, zp,
1330 dmu_sync_late_arrival_ready, dmu_sync_late_arrival_done, dsa,
1331 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1337 * Intent log support: sync the block associated with db to disk.
1338 * N.B. and XXX: the caller is responsible for making sure that the
1339 * data isn't changing while dmu_sync() is writing it.
1343 * EEXIST: this txg has already been synced, so there's nothing to to.
1344 * The caller should not log the write.
1346 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1347 * The caller should not log the write.
1349 * EALREADY: this block is already in the process of being synced.
1350 * The caller should track its progress (somehow).
1352 * EIO: could not do the I/O.
1353 * The caller should do a txg_wait_synced().
1355 * 0: the I/O has been initiated.
1356 * The caller should log this blkptr in the done callback.
1357 * It is possible that the I/O will fail, in which case
1358 * the error will be reported to the done callback and
1359 * propagated to pio from zio_done().
1362 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1364 blkptr_t *bp = zgd->zgd_bp;
1365 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1366 objset_t *os = db->db_objset;
1367 dsl_dataset_t *ds = os->os_dsl_dataset;
1368 dbuf_dirty_record_t *dr;
1369 dmu_sync_arg_t *dsa;
1374 ASSERT(pio != NULL);
1375 ASSERT(BP_IS_HOLE(bp));
1378 SET_BOOKMARK(&zb, ds->ds_object,
1379 db->db.db_object, db->db_level, db->db_blkid);
1383 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1387 * If we're frozen (running ziltest), we always need to generate a bp.
1389 if (txg > spa_freeze_txg(os->os_spa))
1390 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1393 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1394 * and us. If we determine that this txg is not yet syncing,
1395 * but it begins to sync a moment later, that's OK because the
1396 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1398 mutex_enter(&db->db_mtx);
1400 if (txg <= spa_last_synced_txg(os->os_spa)) {
1402 * This txg has already synced. There's nothing to do.
1404 mutex_exit(&db->db_mtx);
1408 if (txg <= spa_syncing_txg(os->os_spa)) {
1410 * This txg is currently syncing, so we can't mess with
1411 * the dirty record anymore; just write a new log block.
1413 mutex_exit(&db->db_mtx);
1414 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1417 dr = db->db_last_dirty;
1418 while (dr && dr->dr_txg != txg)
1423 * There's no dr for this dbuf, so it must have been freed.
1424 * There's no need to log writes to freed blocks, so we're done.
1426 mutex_exit(&db->db_mtx);
1430 ASSERT(dr->dr_txg == txg);
1431 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1432 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1434 * We have already issued a sync write for this buffer,
1435 * or this buffer has already been synced. It could not
1436 * have been dirtied since, or we would have cleared the state.
1438 mutex_exit(&db->db_mtx);
1442 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1443 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1444 mutex_exit(&db->db_mtx);
1446 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1448 dsa->dsa_done = done;
1452 zio_nowait(arc_write(pio, os->os_spa, txg,
1453 bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db), &zp,
1454 dmu_sync_ready, dmu_sync_done, dsa,
1455 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb));
1461 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1467 err = dnode_hold(os, object, FTAG, &dn);
1470 err = dnode_set_blksz(dn, size, ibs, tx);
1471 dnode_rele(dn, FTAG);
1476 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1481 /* XXX assumes dnode_hold will not get an i/o error */
1482 (void) dnode_hold(os, object, FTAG, &dn);
1483 ASSERT(checksum < ZIO_CHECKSUM_FUNCTIONS);
1484 dn->dn_checksum = checksum;
1485 dnode_setdirty(dn, tx);
1486 dnode_rele(dn, FTAG);
1490 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1495 /* XXX assumes dnode_hold will not get an i/o error */
1496 (void) dnode_hold(os, object, FTAG, &dn);
1497 ASSERT(compress < ZIO_COMPRESS_FUNCTIONS);
1498 dn->dn_compress = compress;
1499 dnode_setdirty(dn, tx);
1500 dnode_rele(dn, FTAG);
1503 int zfs_mdcomp_disable = 0;
1504 TUNABLE_INT("vfs.zfs.mdcomp_disable", &zfs_mdcomp_disable);
1505 SYSCTL_DECL(_vfs_zfs);
1506 SYSCTL_INT(_vfs_zfs, OID_AUTO, mdcomp_disable, CTLFLAG_RW,
1507 &zfs_mdcomp_disable, 0, "Disable metadata compression");
1510 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
1512 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
1513 boolean_t ismd = (level > 0 || dmu_ot[type].ot_metadata ||
1515 enum zio_checksum checksum = os->os_checksum;
1516 enum zio_compress compress = os->os_compress;
1517 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
1519 boolean_t dedup_verify = os->os_dedup_verify;
1520 int copies = os->os_copies;
1523 * Determine checksum setting.
1527 * Metadata always gets checksummed. If the data
1528 * checksum is multi-bit correctable, and it's not a
1529 * ZBT-style checksum, then it's suitable for metadata
1530 * as well. Otherwise, the metadata checksum defaults
1533 if (zio_checksum_table[checksum].ci_correctable < 1 ||
1534 zio_checksum_table[checksum].ci_eck)
1535 checksum = ZIO_CHECKSUM_FLETCHER_4;
1537 checksum = zio_checksum_select(dn->dn_checksum, checksum);
1541 * Determine compression setting.
1545 * XXX -- we should design a compression algorithm
1546 * that specializes in arrays of bps.
1548 compress = zfs_mdcomp_disable ? ZIO_COMPRESS_EMPTY :
1551 compress = zio_compress_select(dn->dn_compress, compress);
1555 * Determine dedup setting. If we are in dmu_sync(), we won't
1556 * actually dedup now because that's all done in syncing context;
1557 * but we do want to use the dedup checkum. If the checksum is not
1558 * strong enough to ensure unique signatures, force dedup_verify.
1560 dedup = (!ismd && dedup_checksum != ZIO_CHECKSUM_OFF);
1562 checksum = dedup_checksum;
1563 if (!zio_checksum_table[checksum].ci_dedup)
1567 if (wp & WP_DMU_SYNC)
1570 if (wp & WP_NOFILL) {
1571 ASSERT(!ismd && level == 0);
1572 checksum = ZIO_CHECKSUM_OFF;
1573 compress = ZIO_COMPRESS_OFF;
1577 zp->zp_checksum = checksum;
1578 zp->zp_compress = compress;
1579 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
1580 zp->zp_level = level;
1581 zp->zp_copies = MIN(copies + ismd, spa_max_replication(os->os_spa));
1582 zp->zp_dedup = dedup;
1583 zp->zp_dedup_verify = dedup && dedup_verify;
1587 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
1592 err = dnode_hold(os, object, FTAG, &dn);
1596 * Sync any current changes before
1597 * we go trundling through the block pointers.
1599 for (i = 0; i < TXG_SIZE; i++) {
1600 if (list_link_active(&dn->dn_dirty_link[i]))
1603 if (i != TXG_SIZE) {
1604 dnode_rele(dn, FTAG);
1605 txg_wait_synced(dmu_objset_pool(os), 0);
1606 err = dnode_hold(os, object, FTAG, &dn);
1611 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
1612 dnode_rele(dn, FTAG);
1618 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
1622 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1623 mutex_enter(&dn->dn_mtx);
1627 doi->doi_data_block_size = dn->dn_datablksz;
1628 doi->doi_metadata_block_size = dn->dn_indblkshift ?
1629 1ULL << dn->dn_indblkshift : 0;
1630 doi->doi_type = dn->dn_type;
1631 doi->doi_bonus_type = dn->dn_bonustype;
1632 doi->doi_bonus_size = dn->dn_bonuslen;
1633 doi->doi_indirection = dn->dn_nlevels;
1634 doi->doi_checksum = dn->dn_checksum;
1635 doi->doi_compress = dn->dn_compress;
1636 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
1637 doi->doi_max_offset = (dnp->dn_maxblkid + 1) * dn->dn_datablksz;
1638 doi->doi_fill_count = 0;
1639 for (int i = 0; i < dnp->dn_nblkptr; i++)
1640 doi->doi_fill_count += dnp->dn_blkptr[i].blk_fill;
1642 mutex_exit(&dn->dn_mtx);
1643 rw_exit(&dn->dn_struct_rwlock);
1647 * Get information on a DMU object.
1648 * If doi is NULL, just indicates whether the object exists.
1651 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
1654 int err = dnode_hold(os, object, FTAG, &dn);
1660 dmu_object_info_from_dnode(dn, doi);
1662 dnode_rele(dn, FTAG);
1667 * As above, but faster; can be used when you have a held dbuf in hand.
1670 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
1672 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1675 dmu_object_info_from_dnode(DB_DNODE(db), doi);
1680 * Faster still when you only care about the size.
1681 * This is specifically optimized for zfs_getattr().
1684 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
1685 u_longlong_t *nblk512)
1687 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1693 *blksize = dn->dn_datablksz;
1694 /* add 1 for dnode space */
1695 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
1696 SPA_MINBLOCKSHIFT) + 1;
1701 byteswap_uint64_array(void *vbuf, size_t size)
1703 uint64_t *buf = vbuf;
1704 size_t count = size >> 3;
1707 ASSERT((size & 7) == 0);
1709 for (i = 0; i < count; i++)
1710 buf[i] = BSWAP_64(buf[i]);
1714 byteswap_uint32_array(void *vbuf, size_t size)
1716 uint32_t *buf = vbuf;
1717 size_t count = size >> 2;
1720 ASSERT((size & 3) == 0);
1722 for (i = 0; i < count; i++)
1723 buf[i] = BSWAP_32(buf[i]);
1727 byteswap_uint16_array(void *vbuf, size_t size)
1729 uint16_t *buf = vbuf;
1730 size_t count = size >> 1;
1733 ASSERT((size & 1) == 0);
1735 for (i = 0; i < count; i++)
1736 buf[i] = BSWAP_16(buf[i]);
1741 byteswap_uint8_array(void *vbuf, size_t size)