5960 zfs recv should prefetch indirect blocks
[illumos-gate.git] / usr / src / uts / common / fs / zfs / dmu.c
blob582542eab7a0a7c6e54fbb20e3965ca45054f99b
1 /*
2 * CDDL HEADER START
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]
19 * CDDL HEADER END
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2014 by Delphix. All rights reserved.
25 /* Copyright (c) 2013 by Saso Kiselkov. All rights reserved. */
26 /* Copyright (c) 2013, Joyent, Inc. All rights reserved. */
27 /* Copyright (c) 2014, Nexenta Systems, Inc. All rights reserved. */
29 #include <sys/dmu.h>
30 #include <sys/dmu_impl.h>
31 #include <sys/dmu_tx.h>
32 #include <sys/dbuf.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>
44 #include <sys/zap.h>
45 #include <sys/zio_checksum.h>
46 #include <sys/zio_compress.h>
47 #include <sys/sa.h>
48 #include <sys/zfeature.h>
49 #ifdef _KERNEL
50 #include <sys/vmsystm.h>
51 #include <sys/zfs_znode.h>
52 #endif
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)
133 dnode_t *dn;
134 uint64_t blkid;
135 dmu_buf_impl_t *db;
136 int err;
138 err = dnode_hold(os, object, FTAG, &dn);
139 if (err)
140 return (err);
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);
147 if (db == NULL) {
148 *dbp = NULL;
149 return (SET_ERROR(EIO));
152 *dbp = &db->db;
153 return (err);
157 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
158 void *tag, dmu_buf_t **dbp, int flags)
160 int err;
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);
167 if (err == 0) {
168 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
169 err = dbuf_read(db, NULL, db_flags);
170 if (err != 0) {
171 dbuf_rele(db, tag);
172 *dbp = NULL;
176 return (err);
180 dmu_bonus_max(void)
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;
189 dnode_t *dn;
190 int error;
192 DB_DNODE_ENTER(db);
193 dn = DB_DNODE(db);
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);
199 } else {
200 dnode_setbonuslen(dn, newsize, tx);
201 error = 0;
204 DB_DNODE_EXIT(db);
205 return (error);
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;
212 dnode_t *dn;
213 int error;
215 DB_DNODE_ENTER(db);
216 dn = DB_DNODE(db);
218 if (!DMU_OT_IS_VALID(type)) {
219 error = SET_ERROR(EINVAL);
220 } else if (dn->dn_bonus != db) {
221 error = SET_ERROR(EINVAL);
222 } else {
223 dnode_setbonus_type(dn, type, tx);
224 error = 0;
227 DB_DNODE_EXIT(db);
228 return (error);
231 dmu_object_type_t
232 dmu_get_bonustype(dmu_buf_t *db_fake)
234 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
235 dnode_t *dn;
236 dmu_object_type_t type;
238 DB_DNODE_ENTER(db);
239 dn = DB_DNODE(db);
240 type = dn->dn_bonustype;
241 DB_DNODE_EXIT(db);
243 return (type);
247 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
249 dnode_t *dn;
250 int error;
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);
258 return (error);
262 * returns ENOENT, EIO, or 0.
265 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
267 dnode_t *dn;
268 dmu_buf_impl_t *db;
269 int error;
271 error = dnode_hold(os, object, FTAG, &dn);
272 if (error)
273 return (error);
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);
282 db = dn->dn_bonus;
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));
301 *dbp = &db->db;
302 return (0);
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;
318 int err;
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);
328 ASSERT(db != NULL);
329 err = dbuf_read(db, NULL, flags);
330 if (err == 0)
331 *dbp = &db->db;
332 else
333 dbuf_rele(db, tag);
334 return (err);
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;
341 dnode_t *dn;
342 int err;
344 DB_DNODE_ENTER(db);
345 dn = DB_DNODE(db);
347 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
348 err = SET_ERROR(EINVAL);
349 } else {
350 rw_enter(&dn->dn_struct_rwlock, RW_READER);
352 if (!dn->dn_have_spill) {
353 err = SET_ERROR(ENOENT);
354 } else {
355 err = dmu_spill_hold_by_dnode(dn,
356 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
359 rw_exit(&dn->dn_struct_rwlock);
362 DB_DNODE_EXIT(db);
363 return (err);
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;
370 dnode_t *dn;
371 int err;
373 DB_DNODE_ENTER(db);
374 dn = DB_DNODE(db);
375 err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp);
376 DB_DNODE_EXIT(db);
378 return (err);
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.
387 static int
388 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
389 int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
391 dmu_buf_t **dbp;
392 uint64_t blkid, nblks, i;
393 uint32_t dbuf_flags;
394 int err;
395 zio_t *zio;
397 ASSERT(length <= DMU_MAX_ACCESS);
399 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT;
400 if (flags & DMU_READ_NO_PREFETCH || length > zfetch_array_rd_sz)
401 dbuf_flags |= DB_RF_NOPREFETCH;
403 rw_enter(&dn->dn_struct_rwlock, RW_READER);
404 if (dn->dn_datablkshift) {
405 int blkshift = dn->dn_datablkshift;
406 nblks = (P2ROUNDUP(offset+length, 1ULL<<blkshift) -
407 P2ALIGN(offset, 1ULL<<blkshift)) >> blkshift;
408 } else {
409 if (offset + length > dn->dn_datablksz) {
410 zfs_panic_recover("zfs: accessing past end of object "
411 "%llx/%llx (size=%u access=%llu+%llu)",
412 (longlong_t)dn->dn_objset->
413 os_dsl_dataset->ds_object,
414 (longlong_t)dn->dn_object, dn->dn_datablksz,
415 (longlong_t)offset, (longlong_t)length);
416 rw_exit(&dn->dn_struct_rwlock);
417 return (SET_ERROR(EIO));
419 nblks = 1;
421 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
423 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
424 blkid = dbuf_whichblock(dn, 0, offset);
425 for (i = 0; i < nblks; i++) {
426 dmu_buf_impl_t *db = dbuf_hold(dn, blkid+i, tag);
427 if (db == NULL) {
428 rw_exit(&dn->dn_struct_rwlock);
429 dmu_buf_rele_array(dbp, nblks, tag);
430 zio_nowait(zio);
431 return (SET_ERROR(EIO));
433 /* initiate async i/o */
434 if (read) {
435 (void) dbuf_read(db, zio, dbuf_flags);
437 dbp[i] = &db->db;
439 rw_exit(&dn->dn_struct_rwlock);
441 /* wait for async i/o */
442 err = zio_wait(zio);
443 if (err) {
444 dmu_buf_rele_array(dbp, nblks, tag);
445 return (err);
448 /* wait for other io to complete */
449 if (read) {
450 for (i = 0; i < nblks; i++) {
451 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
452 mutex_enter(&db->db_mtx);
453 while (db->db_state == DB_READ ||
454 db->db_state == DB_FILL)
455 cv_wait(&db->db_changed, &db->db_mtx);
456 if (db->db_state == DB_UNCACHED)
457 err = SET_ERROR(EIO);
458 mutex_exit(&db->db_mtx);
459 if (err) {
460 dmu_buf_rele_array(dbp, nblks, tag);
461 return (err);
466 *numbufsp = nblks;
467 *dbpp = dbp;
468 return (0);
471 static int
472 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
473 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
475 dnode_t *dn;
476 int err;
478 err = dnode_hold(os, object, FTAG, &dn);
479 if (err)
480 return (err);
482 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
483 numbufsp, dbpp, DMU_READ_PREFETCH);
485 dnode_rele(dn, FTAG);
487 return (err);
491 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
492 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
494 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
495 dnode_t *dn;
496 int err;
498 DB_DNODE_ENTER(db);
499 dn = DB_DNODE(db);
500 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
501 numbufsp, dbpp, DMU_READ_PREFETCH);
502 DB_DNODE_EXIT(db);
504 return (err);
507 void
508 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
510 int i;
511 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
513 if (numbufs == 0)
514 return;
516 for (i = 0; i < numbufs; i++) {
517 if (dbp[i])
518 dbuf_rele(dbp[i], tag);
521 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
525 * Issue prefetch i/os for the given blocks. If level is greater than 0, the
526 * indirect blocks prefeteched will be those that point to the blocks containing
527 * the data starting at offset, and continuing to offset + len.
529 * Note that if the indirect blocks above the blocks being prefetched are not in
530 * cache, they will be asychronously read in.
532 void
533 dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset,
534 uint64_t len, zio_priority_t pri)
536 dnode_t *dn;
537 uint64_t blkid;
538 int nblks, err;
540 if (zfs_prefetch_disable)
541 return;
543 if (len == 0) { /* they're interested in the bonus buffer */
544 dn = DMU_META_DNODE(os);
546 if (object == 0 || object >= DN_MAX_OBJECT)
547 return;
549 rw_enter(&dn->dn_struct_rwlock, RW_READER);
550 blkid = dbuf_whichblock(dn, level,
551 object * sizeof (dnode_phys_t));
552 dbuf_prefetch(dn, level, blkid, pri, 0);
553 rw_exit(&dn->dn_struct_rwlock);
554 return;
558 * XXX - Note, if the dnode for the requested object is not
559 * already cached, we will do a *synchronous* read in the
560 * dnode_hold() call. The same is true for any indirects.
562 err = dnode_hold(os, object, FTAG, &dn);
563 if (err != 0)
564 return;
566 rw_enter(&dn->dn_struct_rwlock, RW_READER);
568 * offset + len - 1 is the last byte we want to prefetch for, and offset
569 * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the
570 * last block we want to prefetch, and dbuf_whichblock(dn, level,
571 * offset) is the first. Then the number we need to prefetch is the
572 * last - first + 1.
574 if (level > 0 || dn->dn_datablkshift != 0) {
575 nblks = dbuf_whichblock(dn, level, offset + len - 1) -
576 dbuf_whichblock(dn, level, offset) + 1;
577 } else {
578 nblks = (offset < dn->dn_datablksz);
581 if (nblks != 0) {
582 blkid = dbuf_whichblock(dn, level, offset);
583 for (int i = 0; i < nblks; i++)
584 dbuf_prefetch(dn, level, blkid + i, pri, 0);
587 rw_exit(&dn->dn_struct_rwlock);
589 dnode_rele(dn, FTAG);
593 * Get the next "chunk" of file data to free. We traverse the file from
594 * the end so that the file gets shorter over time (if we crashes in the
595 * middle, this will leave us in a better state). We find allocated file
596 * data by simply searching the allocated level 1 indirects.
598 * On input, *start should be the first offset that does not need to be
599 * freed (e.g. "offset + length"). On return, *start will be the first
600 * offset that should be freed.
602 static int
603 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum)
605 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
606 /* bytes of data covered by a level-1 indirect block */
607 uint64_t iblkrange =
608 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
610 ASSERT3U(minimum, <=, *start);
612 if (*start - minimum <= iblkrange * maxblks) {
613 *start = minimum;
614 return (0);
616 ASSERT(ISP2(iblkrange));
618 for (uint64_t blks = 0; *start > minimum && blks < maxblks; blks++) {
619 int err;
622 * dnode_next_offset(BACKWARDS) will find an allocated L1
623 * indirect block at or before the input offset. We must
624 * decrement *start so that it is at the end of the region
625 * to search.
627 (*start)--;
628 err = dnode_next_offset(dn,
629 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
631 /* if there are no indirect blocks before start, we are done */
632 if (err == ESRCH) {
633 *start = minimum;
634 break;
635 } else if (err != 0) {
636 return (err);
639 /* set start to the beginning of this L1 indirect */
640 *start = P2ALIGN(*start, iblkrange);
642 if (*start < minimum)
643 *start = minimum;
644 return (0);
647 static int
648 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
649 uint64_t length)
651 uint64_t object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
652 int err;
654 if (offset >= object_size)
655 return (0);
657 if (length == DMU_OBJECT_END || offset + length > object_size)
658 length = object_size - offset;
660 while (length != 0) {
661 uint64_t chunk_end, chunk_begin;
663 chunk_end = chunk_begin = offset + length;
665 /* move chunk_begin backwards to the beginning of this chunk */
666 err = get_next_chunk(dn, &chunk_begin, offset);
667 if (err)
668 return (err);
669 ASSERT3U(chunk_begin, >=, offset);
670 ASSERT3U(chunk_begin, <=, chunk_end);
672 dmu_tx_t *tx = dmu_tx_create(os);
673 dmu_tx_hold_free(tx, dn->dn_object,
674 chunk_begin, chunk_end - chunk_begin);
677 * Mark this transaction as typically resulting in a net
678 * reduction in space used.
680 dmu_tx_mark_netfree(tx);
681 err = dmu_tx_assign(tx, TXG_WAIT);
682 if (err) {
683 dmu_tx_abort(tx);
684 return (err);
686 dnode_free_range(dn, chunk_begin, chunk_end - chunk_begin, tx);
687 dmu_tx_commit(tx);
689 length -= chunk_end - chunk_begin;
691 return (0);
695 dmu_free_long_range(objset_t *os, uint64_t object,
696 uint64_t offset, uint64_t length)
698 dnode_t *dn;
699 int err;
701 err = dnode_hold(os, object, FTAG, &dn);
702 if (err != 0)
703 return (err);
704 err = dmu_free_long_range_impl(os, dn, offset, length);
707 * It is important to zero out the maxblkid when freeing the entire
708 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
709 * will take the fast path, and (b) dnode_reallocate() can verify
710 * that the entire file has been freed.
712 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
713 dn->dn_maxblkid = 0;
715 dnode_rele(dn, FTAG);
716 return (err);
720 dmu_free_long_object(objset_t *os, uint64_t object)
722 dmu_tx_t *tx;
723 int err;
725 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
726 if (err != 0)
727 return (err);
729 tx = dmu_tx_create(os);
730 dmu_tx_hold_bonus(tx, object);
731 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
732 dmu_tx_mark_netfree(tx);
733 err = dmu_tx_assign(tx, TXG_WAIT);
734 if (err == 0) {
735 err = dmu_object_free(os, object, tx);
736 dmu_tx_commit(tx);
737 } else {
738 dmu_tx_abort(tx);
741 return (err);
745 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
746 uint64_t size, dmu_tx_t *tx)
748 dnode_t *dn;
749 int err = dnode_hold(os, object, FTAG, &dn);
750 if (err)
751 return (err);
752 ASSERT(offset < UINT64_MAX);
753 ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
754 dnode_free_range(dn, offset, size, tx);
755 dnode_rele(dn, FTAG);
756 return (0);
760 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
761 void *buf, uint32_t flags)
763 dnode_t *dn;
764 dmu_buf_t **dbp;
765 int numbufs, err;
767 err = dnode_hold(os, object, FTAG, &dn);
768 if (err)
769 return (err);
772 * Deal with odd block sizes, where there can't be data past the first
773 * block. If we ever do the tail block optimization, we will need to
774 * handle that here as well.
776 if (dn->dn_maxblkid == 0) {
777 int newsz = offset > dn->dn_datablksz ? 0 :
778 MIN(size, dn->dn_datablksz - offset);
779 bzero((char *)buf + newsz, size - newsz);
780 size = newsz;
783 while (size > 0) {
784 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
785 int i;
788 * NB: we could do this block-at-a-time, but it's nice
789 * to be reading in parallel.
791 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
792 TRUE, FTAG, &numbufs, &dbp, flags);
793 if (err)
794 break;
796 for (i = 0; i < numbufs; i++) {
797 int tocpy;
798 int bufoff;
799 dmu_buf_t *db = dbp[i];
801 ASSERT(size > 0);
803 bufoff = offset - db->db_offset;
804 tocpy = (int)MIN(db->db_size - bufoff, size);
806 bcopy((char *)db->db_data + bufoff, buf, tocpy);
808 offset += tocpy;
809 size -= tocpy;
810 buf = (char *)buf + tocpy;
812 dmu_buf_rele_array(dbp, numbufs, FTAG);
814 dnode_rele(dn, FTAG);
815 return (err);
818 void
819 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
820 const void *buf, dmu_tx_t *tx)
822 dmu_buf_t **dbp;
823 int numbufs, i;
825 if (size == 0)
826 return;
828 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
829 FALSE, FTAG, &numbufs, &dbp));
831 for (i = 0; i < numbufs; i++) {
832 int tocpy;
833 int bufoff;
834 dmu_buf_t *db = dbp[i];
836 ASSERT(size > 0);
838 bufoff = offset - db->db_offset;
839 tocpy = (int)MIN(db->db_size - bufoff, size);
841 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
843 if (tocpy == db->db_size)
844 dmu_buf_will_fill(db, tx);
845 else
846 dmu_buf_will_dirty(db, tx);
848 bcopy(buf, (char *)db->db_data + bufoff, tocpy);
850 if (tocpy == db->db_size)
851 dmu_buf_fill_done(db, tx);
853 offset += tocpy;
854 size -= tocpy;
855 buf = (char *)buf + tocpy;
857 dmu_buf_rele_array(dbp, numbufs, FTAG);
860 void
861 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
862 dmu_tx_t *tx)
864 dmu_buf_t **dbp;
865 int numbufs, i;
867 if (size == 0)
868 return;
870 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
871 FALSE, FTAG, &numbufs, &dbp));
873 for (i = 0; i < numbufs; i++) {
874 dmu_buf_t *db = dbp[i];
876 dmu_buf_will_not_fill(db, tx);
878 dmu_buf_rele_array(dbp, numbufs, FTAG);
881 void
882 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
883 void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
884 int compressed_size, int byteorder, dmu_tx_t *tx)
886 dmu_buf_t *db;
888 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
889 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
890 VERIFY0(dmu_buf_hold_noread(os, object, offset,
891 FTAG, &db));
893 dmu_buf_write_embedded(db,
894 data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
895 uncompressed_size, compressed_size, byteorder, tx);
897 dmu_buf_rele(db, FTAG);
901 * DMU support for xuio
903 kstat_t *xuio_ksp = NULL;
906 dmu_xuio_init(xuio_t *xuio, int nblk)
908 dmu_xuio_t *priv;
909 uio_t *uio = &xuio->xu_uio;
911 uio->uio_iovcnt = nblk;
912 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
914 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
915 priv->cnt = nblk;
916 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
917 priv->iovp = uio->uio_iov;
918 XUIO_XUZC_PRIV(xuio) = priv;
920 if (XUIO_XUZC_RW(xuio) == UIO_READ)
921 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
922 else
923 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
925 return (0);
928 void
929 dmu_xuio_fini(xuio_t *xuio)
931 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
932 int nblk = priv->cnt;
934 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
935 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
936 kmem_free(priv, sizeof (dmu_xuio_t));
938 if (XUIO_XUZC_RW(xuio) == UIO_READ)
939 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
940 else
941 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
945 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
946 * and increase priv->next by 1.
949 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
951 struct iovec *iov;
952 uio_t *uio = &xuio->xu_uio;
953 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
954 int i = priv->next++;
956 ASSERT(i < priv->cnt);
957 ASSERT(off + n <= arc_buf_size(abuf));
958 iov = uio->uio_iov + i;
959 iov->iov_base = (char *)abuf->b_data + off;
960 iov->iov_len = n;
961 priv->bufs[i] = abuf;
962 return (0);
966 dmu_xuio_cnt(xuio_t *xuio)
968 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
969 return (priv->cnt);
972 arc_buf_t *
973 dmu_xuio_arcbuf(xuio_t *xuio, int i)
975 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
977 ASSERT(i < priv->cnt);
978 return (priv->bufs[i]);
981 void
982 dmu_xuio_clear(xuio_t *xuio, int i)
984 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
986 ASSERT(i < priv->cnt);
987 priv->bufs[i] = NULL;
990 static void
991 xuio_stat_init(void)
993 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
994 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
995 KSTAT_FLAG_VIRTUAL);
996 if (xuio_ksp != NULL) {
997 xuio_ksp->ks_data = &xuio_stats;
998 kstat_install(xuio_ksp);
1002 static void
1003 xuio_stat_fini(void)
1005 if (xuio_ksp != NULL) {
1006 kstat_delete(xuio_ksp);
1007 xuio_ksp = NULL;
1011 void
1012 xuio_stat_wbuf_copied()
1014 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1017 void
1018 xuio_stat_wbuf_nocopy()
1020 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
1023 #ifdef _KERNEL
1024 static int
1025 dmu_read_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size)
1027 dmu_buf_t **dbp;
1028 int numbufs, i, err;
1029 xuio_t *xuio = NULL;
1032 * NB: we could do this block-at-a-time, but it's nice
1033 * to be reading in parallel.
1035 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1036 TRUE, FTAG, &numbufs, &dbp, 0);
1037 if (err)
1038 return (err);
1040 if (uio->uio_extflg == UIO_XUIO)
1041 xuio = (xuio_t *)uio;
1043 for (i = 0; i < numbufs; i++) {
1044 int tocpy;
1045 int bufoff;
1046 dmu_buf_t *db = dbp[i];
1048 ASSERT(size > 0);
1050 bufoff = uio->uio_loffset - db->db_offset;
1051 tocpy = (int)MIN(db->db_size - bufoff, size);
1053 if (xuio) {
1054 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
1055 arc_buf_t *dbuf_abuf = dbi->db_buf;
1056 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
1057 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
1058 if (!err) {
1059 uio->uio_resid -= tocpy;
1060 uio->uio_loffset += tocpy;
1063 if (abuf == dbuf_abuf)
1064 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
1065 else
1066 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
1067 } else {
1068 err = uiomove((char *)db->db_data + bufoff, tocpy,
1069 UIO_READ, uio);
1071 if (err)
1072 break;
1074 size -= tocpy;
1076 dmu_buf_rele_array(dbp, numbufs, FTAG);
1078 return (err);
1082 * Read 'size' bytes into the uio buffer.
1083 * From object zdb->db_object.
1084 * Starting at offset uio->uio_loffset.
1086 * If the caller already has a dbuf in the target object
1087 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1088 * because we don't have to find the dnode_t for the object.
1091 dmu_read_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size)
1093 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1094 dnode_t *dn;
1095 int err;
1097 if (size == 0)
1098 return (0);
1100 DB_DNODE_ENTER(db);
1101 dn = DB_DNODE(db);
1102 err = dmu_read_uio_dnode(dn, uio, size);
1103 DB_DNODE_EXIT(db);
1105 return (err);
1109 * Read 'size' bytes into the uio buffer.
1110 * From the specified object
1111 * Starting at offset uio->uio_loffset.
1114 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
1116 dnode_t *dn;
1117 int err;
1119 if (size == 0)
1120 return (0);
1122 err = dnode_hold(os, object, FTAG, &dn);
1123 if (err)
1124 return (err);
1126 err = dmu_read_uio_dnode(dn, uio, size);
1128 dnode_rele(dn, FTAG);
1130 return (err);
1133 static int
1134 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1136 dmu_buf_t **dbp;
1137 int numbufs;
1138 int err = 0;
1139 int i;
1141 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1142 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1143 if (err)
1144 return (err);
1146 for (i = 0; i < numbufs; i++) {
1147 int tocpy;
1148 int bufoff;
1149 dmu_buf_t *db = dbp[i];
1151 ASSERT(size > 0);
1153 bufoff = uio->uio_loffset - db->db_offset;
1154 tocpy = (int)MIN(db->db_size - bufoff, size);
1156 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1158 if (tocpy == db->db_size)
1159 dmu_buf_will_fill(db, tx);
1160 else
1161 dmu_buf_will_dirty(db, tx);
1164 * XXX uiomove could block forever (eg. nfs-backed
1165 * pages). There needs to be a uiolockdown() function
1166 * to lock the pages in memory, so that uiomove won't
1167 * block.
1169 err = uiomove((char *)db->db_data + bufoff, tocpy,
1170 UIO_WRITE, uio);
1172 if (tocpy == db->db_size)
1173 dmu_buf_fill_done(db, tx);
1175 if (err)
1176 break;
1178 size -= tocpy;
1181 dmu_buf_rele_array(dbp, numbufs, FTAG);
1182 return (err);
1186 * Write 'size' bytes from the uio buffer.
1187 * To object zdb->db_object.
1188 * Starting at offset uio->uio_loffset.
1190 * If the caller already has a dbuf in the target object
1191 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1192 * because we don't have to find the dnode_t for the object.
1195 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1196 dmu_tx_t *tx)
1198 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1199 dnode_t *dn;
1200 int err;
1202 if (size == 0)
1203 return (0);
1205 DB_DNODE_ENTER(db);
1206 dn = DB_DNODE(db);
1207 err = dmu_write_uio_dnode(dn, uio, size, tx);
1208 DB_DNODE_EXIT(db);
1210 return (err);
1214 * Write 'size' bytes from the uio buffer.
1215 * To the specified object.
1216 * Starting at offset uio->uio_loffset.
1219 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1220 dmu_tx_t *tx)
1222 dnode_t *dn;
1223 int err;
1225 if (size == 0)
1226 return (0);
1228 err = dnode_hold(os, object, FTAG, &dn);
1229 if (err)
1230 return (err);
1232 err = dmu_write_uio_dnode(dn, uio, size, tx);
1234 dnode_rele(dn, FTAG);
1236 return (err);
1240 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1241 page_t *pp, dmu_tx_t *tx)
1243 dmu_buf_t **dbp;
1244 int numbufs, i;
1245 int err;
1247 if (size == 0)
1248 return (0);
1250 err = dmu_buf_hold_array(os, object, offset, size,
1251 FALSE, FTAG, &numbufs, &dbp);
1252 if (err)
1253 return (err);
1255 for (i = 0; i < numbufs; i++) {
1256 int tocpy, copied, thiscpy;
1257 int bufoff;
1258 dmu_buf_t *db = dbp[i];
1259 caddr_t va;
1261 ASSERT(size > 0);
1262 ASSERT3U(db->db_size, >=, PAGESIZE);
1264 bufoff = offset - db->db_offset;
1265 tocpy = (int)MIN(db->db_size - bufoff, size);
1267 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1269 if (tocpy == db->db_size)
1270 dmu_buf_will_fill(db, tx);
1271 else
1272 dmu_buf_will_dirty(db, tx);
1274 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1275 ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff);
1276 thiscpy = MIN(PAGESIZE, tocpy - copied);
1277 va = zfs_map_page(pp, S_READ);
1278 bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1279 zfs_unmap_page(pp, va);
1280 pp = pp->p_next;
1281 bufoff += PAGESIZE;
1284 if (tocpy == db->db_size)
1285 dmu_buf_fill_done(db, tx);
1287 offset += tocpy;
1288 size -= tocpy;
1290 dmu_buf_rele_array(dbp, numbufs, FTAG);
1291 return (err);
1293 #endif
1296 * Allocate a loaned anonymous arc buffer.
1298 arc_buf_t *
1299 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1301 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1303 return (arc_loan_buf(db->db_objset->os_spa, size));
1307 * Free a loaned arc buffer.
1309 void
1310 dmu_return_arcbuf(arc_buf_t *buf)
1312 arc_return_buf(buf, FTAG);
1313 VERIFY(arc_buf_remove_ref(buf, FTAG));
1317 * When possible directly assign passed loaned arc buffer to a dbuf.
1318 * If this is not possible copy the contents of passed arc buf via
1319 * dmu_write().
1321 void
1322 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1323 dmu_tx_t *tx)
1325 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1326 dnode_t *dn;
1327 dmu_buf_impl_t *db;
1328 uint32_t blksz = (uint32_t)arc_buf_size(buf);
1329 uint64_t blkid;
1331 DB_DNODE_ENTER(dbuf);
1332 dn = DB_DNODE(dbuf);
1333 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1334 blkid = dbuf_whichblock(dn, 0, offset);
1335 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1336 rw_exit(&dn->dn_struct_rwlock);
1337 DB_DNODE_EXIT(dbuf);
1340 * We can only assign if the offset is aligned, the arc buf is the
1341 * same size as the dbuf, and the dbuf is not metadata. It
1342 * can't be metadata because the loaned arc buf comes from the
1343 * user-data kmem arena.
1345 if (offset == db->db.db_offset && blksz == db->db.db_size &&
1346 DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA) {
1347 dbuf_assign_arcbuf(db, buf, tx);
1348 dbuf_rele(db, FTAG);
1349 } else {
1350 objset_t *os;
1351 uint64_t object;
1353 DB_DNODE_ENTER(dbuf);
1354 dn = DB_DNODE(dbuf);
1355 os = dn->dn_objset;
1356 object = dn->dn_object;
1357 DB_DNODE_EXIT(dbuf);
1359 dbuf_rele(db, FTAG);
1360 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1361 dmu_return_arcbuf(buf);
1362 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1366 typedef struct {
1367 dbuf_dirty_record_t *dsa_dr;
1368 dmu_sync_cb_t *dsa_done;
1369 zgd_t *dsa_zgd;
1370 dmu_tx_t *dsa_tx;
1371 } dmu_sync_arg_t;
1373 /* ARGSUSED */
1374 static void
1375 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1377 dmu_sync_arg_t *dsa = varg;
1378 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1379 blkptr_t *bp = zio->io_bp;
1381 if (zio->io_error == 0) {
1382 if (BP_IS_HOLE(bp)) {
1384 * A block of zeros may compress to a hole, but the
1385 * block size still needs to be known for replay.
1387 BP_SET_LSIZE(bp, db->db_size);
1388 } else if (!BP_IS_EMBEDDED(bp)) {
1389 ASSERT(BP_GET_LEVEL(bp) == 0);
1390 bp->blk_fill = 1;
1395 static void
1396 dmu_sync_late_arrival_ready(zio_t *zio)
1398 dmu_sync_ready(zio, NULL, zio->io_private);
1401 /* ARGSUSED */
1402 static void
1403 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1405 dmu_sync_arg_t *dsa = varg;
1406 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1407 dmu_buf_impl_t *db = dr->dr_dbuf;
1409 mutex_enter(&db->db_mtx);
1410 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1411 if (zio->io_error == 0) {
1412 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1413 if (dr->dt.dl.dr_nopwrite) {
1414 blkptr_t *bp = zio->io_bp;
1415 blkptr_t *bp_orig = &zio->io_bp_orig;
1416 uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1418 ASSERT(BP_EQUAL(bp, bp_orig));
1419 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1420 ASSERT(zio_checksum_table[chksum].ci_dedup);
1422 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1423 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1424 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1427 * Old style holes are filled with all zeros, whereas
1428 * new-style holes maintain their lsize, type, level,
1429 * and birth time (see zio_write_compress). While we
1430 * need to reset the BP_SET_LSIZE() call that happened
1431 * in dmu_sync_ready for old style holes, we do *not*
1432 * want to wipe out the information contained in new
1433 * style holes. Thus, only zero out the block pointer if
1434 * it's an old style hole.
1436 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) &&
1437 dr->dt.dl.dr_overridden_by.blk_birth == 0)
1438 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1439 } else {
1440 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1442 cv_broadcast(&db->db_changed);
1443 mutex_exit(&db->db_mtx);
1445 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1447 kmem_free(dsa, sizeof (*dsa));
1450 static void
1451 dmu_sync_late_arrival_done(zio_t *zio)
1453 blkptr_t *bp = zio->io_bp;
1454 dmu_sync_arg_t *dsa = zio->io_private;
1455 blkptr_t *bp_orig = &zio->io_bp_orig;
1457 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1459 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE)
1460 * then there is nothing to do here. Otherwise, free the
1461 * newly allocated block in this txg.
1463 if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
1464 ASSERT(BP_EQUAL(bp, bp_orig));
1465 } else {
1466 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1467 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1468 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1469 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1473 dmu_tx_commit(dsa->dsa_tx);
1475 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1477 kmem_free(dsa, sizeof (*dsa));
1480 static int
1481 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1482 zio_prop_t *zp, zbookmark_phys_t *zb)
1484 dmu_sync_arg_t *dsa;
1485 dmu_tx_t *tx;
1487 tx = dmu_tx_create(os);
1488 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1489 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1490 dmu_tx_abort(tx);
1491 /* Make zl_get_data do txg_waited_synced() */
1492 return (SET_ERROR(EIO));
1495 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1496 dsa->dsa_dr = NULL;
1497 dsa->dsa_done = done;
1498 dsa->dsa_zgd = zgd;
1499 dsa->dsa_tx = tx;
1501 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1502 zgd->zgd_db->db_data, zgd->zgd_db->db_size, zp,
1503 dmu_sync_late_arrival_ready, NULL, dmu_sync_late_arrival_done, dsa,
1504 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1506 return (0);
1510 * Intent log support: sync the block associated with db to disk.
1511 * N.B. and XXX: the caller is responsible for making sure that the
1512 * data isn't changing while dmu_sync() is writing it.
1514 * Return values:
1516 * EEXIST: this txg has already been synced, so there's nothing to do.
1517 * The caller should not log the write.
1519 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1520 * The caller should not log the write.
1522 * EALREADY: this block is already in the process of being synced.
1523 * The caller should track its progress (somehow).
1525 * EIO: could not do the I/O.
1526 * The caller should do a txg_wait_synced().
1528 * 0: the I/O has been initiated.
1529 * The caller should log this blkptr in the done callback.
1530 * It is possible that the I/O will fail, in which case
1531 * the error will be reported to the done callback and
1532 * propagated to pio from zio_done().
1535 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1537 blkptr_t *bp = zgd->zgd_bp;
1538 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1539 objset_t *os = db->db_objset;
1540 dsl_dataset_t *ds = os->os_dsl_dataset;
1541 dbuf_dirty_record_t *dr;
1542 dmu_sync_arg_t *dsa;
1543 zbookmark_phys_t zb;
1544 zio_prop_t zp;
1545 dnode_t *dn;
1547 ASSERT(pio != NULL);
1548 ASSERT(txg != 0);
1550 SET_BOOKMARK(&zb, ds->ds_object,
1551 db->db.db_object, db->db_level, db->db_blkid);
1553 DB_DNODE_ENTER(db);
1554 dn = DB_DNODE(db);
1555 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1556 DB_DNODE_EXIT(db);
1559 * If we're frozen (running ziltest), we always need to generate a bp.
1561 if (txg > spa_freeze_txg(os->os_spa))
1562 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1565 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1566 * and us. If we determine that this txg is not yet syncing,
1567 * but it begins to sync a moment later, that's OK because the
1568 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1570 mutex_enter(&db->db_mtx);
1572 if (txg <= spa_last_synced_txg(os->os_spa)) {
1574 * This txg has already synced. There's nothing to do.
1576 mutex_exit(&db->db_mtx);
1577 return (SET_ERROR(EEXIST));
1580 if (txg <= spa_syncing_txg(os->os_spa)) {
1582 * This txg is currently syncing, so we can't mess with
1583 * the dirty record anymore; just write a new log block.
1585 mutex_exit(&db->db_mtx);
1586 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1589 dr = db->db_last_dirty;
1590 while (dr && dr->dr_txg != txg)
1591 dr = dr->dr_next;
1593 if (dr == NULL) {
1595 * There's no dr for this dbuf, so it must have been freed.
1596 * There's no need to log writes to freed blocks, so we're done.
1598 mutex_exit(&db->db_mtx);
1599 return (SET_ERROR(ENOENT));
1602 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
1605 * Assume the on-disk data is X, the current syncing data (in
1606 * txg - 1) is Y, and the current in-memory data is Z (currently
1607 * in dmu_sync).
1609 * We usually want to perform a nopwrite if X and Z are the
1610 * same. However, if Y is different (i.e. the BP is going to
1611 * change before this write takes effect), then a nopwrite will
1612 * be incorrect - we would override with X, which could have
1613 * been freed when Y was written.
1615 * (Note that this is not a concern when we are nop-writing from
1616 * syncing context, because X and Y must be identical, because
1617 * all previous txgs have been synced.)
1619 * Therefore, we disable nopwrite if the current BP could change
1620 * before this TXG. There are two ways it could change: by
1621 * being dirty (dr_next is non-NULL), or by being freed
1622 * (dnode_block_freed()). This behavior is verified by
1623 * zio_done(), which VERIFYs that the override BP is identical
1624 * to the on-disk BP.
1626 DB_DNODE_ENTER(db);
1627 dn = DB_DNODE(db);
1628 if (dr->dr_next != NULL || dnode_block_freed(dn, db->db_blkid))
1629 zp.zp_nopwrite = B_FALSE;
1630 DB_DNODE_EXIT(db);
1632 ASSERT(dr->dr_txg == txg);
1633 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1634 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1636 * We have already issued a sync write for this buffer,
1637 * or this buffer has already been synced. It could not
1638 * have been dirtied since, or we would have cleared the state.
1640 mutex_exit(&db->db_mtx);
1641 return (SET_ERROR(EALREADY));
1644 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1645 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1646 mutex_exit(&db->db_mtx);
1648 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1649 dsa->dsa_dr = dr;
1650 dsa->dsa_done = done;
1651 dsa->dsa_zgd = zgd;
1652 dsa->dsa_tx = NULL;
1654 zio_nowait(arc_write(pio, os->os_spa, txg,
1655 bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
1656 DBUF_IS_L2COMPRESSIBLE(db), &zp, dmu_sync_ready,
1657 NULL, dmu_sync_done, dsa, ZIO_PRIORITY_SYNC_WRITE,
1658 ZIO_FLAG_CANFAIL, &zb));
1660 return (0);
1664 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1665 dmu_tx_t *tx)
1667 dnode_t *dn;
1668 int err;
1670 err = dnode_hold(os, object, FTAG, &dn);
1671 if (err)
1672 return (err);
1673 err = dnode_set_blksz(dn, size, ibs, tx);
1674 dnode_rele(dn, FTAG);
1675 return (err);
1678 void
1679 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1680 dmu_tx_t *tx)
1682 dnode_t *dn;
1685 * Send streams include each object's checksum function. This
1686 * check ensures that the receiving system can understand the
1687 * checksum function transmitted.
1689 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS);
1691 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1692 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS);
1693 dn->dn_checksum = checksum;
1694 dnode_setdirty(dn, tx);
1695 dnode_rele(dn, FTAG);
1698 void
1699 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1700 dmu_tx_t *tx)
1702 dnode_t *dn;
1705 * Send streams include each object's compression function. This
1706 * check ensures that the receiving system can understand the
1707 * compression function transmitted.
1709 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS);
1711 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1712 dn->dn_compress = compress;
1713 dnode_setdirty(dn, tx);
1714 dnode_rele(dn, FTAG);
1717 int zfs_mdcomp_disable = 0;
1720 * When the "redundant_metadata" property is set to "most", only indirect
1721 * blocks of this level and higher will have an additional ditto block.
1723 int zfs_redundant_metadata_most_ditto_level = 2;
1725 void
1726 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
1728 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
1729 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
1730 (wp & WP_SPILL));
1731 enum zio_checksum checksum = os->os_checksum;
1732 enum zio_compress compress = os->os_compress;
1733 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
1734 boolean_t dedup = B_FALSE;
1735 boolean_t nopwrite = B_FALSE;
1736 boolean_t dedup_verify = os->os_dedup_verify;
1737 int copies = os->os_copies;
1740 * We maintain different write policies for each of the following
1741 * types of data:
1742 * 1. metadata
1743 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1744 * 3. all other level 0 blocks
1746 if (ismd) {
1747 if (zfs_mdcomp_disable) {
1748 compress = ZIO_COMPRESS_EMPTY;
1749 } else {
1751 * XXX -- we should design a compression algorithm
1752 * that specializes in arrays of bps.
1754 compress = zio_compress_select(os->os_spa,
1755 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON);
1759 * Metadata always gets checksummed. If the data
1760 * checksum is multi-bit correctable, and it's not a
1761 * ZBT-style checksum, then it's suitable for metadata
1762 * as well. Otherwise, the metadata checksum defaults
1763 * to fletcher4.
1765 if (zio_checksum_table[checksum].ci_correctable < 1 ||
1766 zio_checksum_table[checksum].ci_eck)
1767 checksum = ZIO_CHECKSUM_FLETCHER_4;
1769 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL ||
1770 (os->os_redundant_metadata ==
1771 ZFS_REDUNDANT_METADATA_MOST &&
1772 (level >= zfs_redundant_metadata_most_ditto_level ||
1773 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))))
1774 copies++;
1775 } else if (wp & WP_NOFILL) {
1776 ASSERT(level == 0);
1779 * If we're writing preallocated blocks, we aren't actually
1780 * writing them so don't set any policy properties. These
1781 * blocks are currently only used by an external subsystem
1782 * outside of zfs (i.e. dump) and not written by the zio
1783 * pipeline.
1785 compress = ZIO_COMPRESS_OFF;
1786 checksum = ZIO_CHECKSUM_NOPARITY;
1787 } else {
1788 compress = zio_compress_select(os->os_spa, dn->dn_compress,
1789 compress);
1791 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
1792 zio_checksum_select(dn->dn_checksum, checksum) :
1793 dedup_checksum;
1796 * Determine dedup setting. If we are in dmu_sync(),
1797 * we won't actually dedup now because that's all
1798 * done in syncing context; but we do want to use the
1799 * dedup checkum. If the checksum is not strong
1800 * enough to ensure unique signatures, force
1801 * dedup_verify.
1803 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
1804 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
1805 if (!zio_checksum_table[checksum].ci_dedup)
1806 dedup_verify = B_TRUE;
1810 * Enable nopwrite if we have a cryptographically secure
1811 * checksum that has no known collisions (i.e. SHA-256)
1812 * and compression is enabled. We don't enable nopwrite if
1813 * dedup is enabled as the two features are mutually exclusive.
1815 nopwrite = (!dedup && zio_checksum_table[checksum].ci_dedup &&
1816 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
1819 zp->zp_checksum = checksum;
1820 zp->zp_compress = compress;
1821 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
1822 zp->zp_level = level;
1823 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
1824 zp->zp_dedup = dedup;
1825 zp->zp_dedup_verify = dedup && dedup_verify;
1826 zp->zp_nopwrite = nopwrite;
1830 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
1832 dnode_t *dn;
1833 int err;
1836 * Sync any current changes before
1837 * we go trundling through the block pointers.
1839 err = dmu_object_wait_synced(os, object);
1840 if (err) {
1841 return (err);
1844 err = dnode_hold(os, object, FTAG, &dn);
1845 if (err) {
1846 return (err);
1849 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
1850 dnode_rele(dn, FTAG);
1852 return (err);
1856 * Given the ZFS object, if it contains any dirty nodes
1857 * this function flushes all dirty blocks to disk. This
1858 * ensures the DMU object info is updated. A more efficient
1859 * future version might just find the TXG with the maximum
1860 * ID and wait for that to be synced.
1863 dmu_object_wait_synced(objset_t *os, uint64_t object) {
1864 dnode_t *dn;
1865 int error, i;
1867 error = dnode_hold(os, object, FTAG, &dn);
1868 if (error) {
1869 return (error);
1872 for (i = 0; i < TXG_SIZE; i++) {
1873 if (list_link_active(&dn->dn_dirty_link[i])) {
1874 break;
1877 dnode_rele(dn, FTAG);
1878 if (i != TXG_SIZE) {
1879 txg_wait_synced(dmu_objset_pool(os), 0);
1882 return (0);
1885 void
1886 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
1888 dnode_phys_t *dnp;
1890 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1891 mutex_enter(&dn->dn_mtx);
1893 dnp = dn->dn_phys;
1895 doi->doi_data_block_size = dn->dn_datablksz;
1896 doi->doi_metadata_block_size = dn->dn_indblkshift ?
1897 1ULL << dn->dn_indblkshift : 0;
1898 doi->doi_type = dn->dn_type;
1899 doi->doi_bonus_type = dn->dn_bonustype;
1900 doi->doi_bonus_size = dn->dn_bonuslen;
1901 doi->doi_indirection = dn->dn_nlevels;
1902 doi->doi_checksum = dn->dn_checksum;
1903 doi->doi_compress = dn->dn_compress;
1904 doi->doi_nblkptr = dn->dn_nblkptr;
1905 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
1906 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
1907 doi->doi_fill_count = 0;
1908 for (int i = 0; i < dnp->dn_nblkptr; i++)
1909 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
1911 mutex_exit(&dn->dn_mtx);
1912 rw_exit(&dn->dn_struct_rwlock);
1916 * Get information on a DMU object.
1917 * If doi is NULL, just indicates whether the object exists.
1920 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
1922 dnode_t *dn;
1923 int err = dnode_hold(os, object, FTAG, &dn);
1925 if (err)
1926 return (err);
1928 if (doi != NULL)
1929 dmu_object_info_from_dnode(dn, doi);
1931 dnode_rele(dn, FTAG);
1932 return (0);
1936 * As above, but faster; can be used when you have a held dbuf in hand.
1938 void
1939 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
1941 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1943 DB_DNODE_ENTER(db);
1944 dmu_object_info_from_dnode(DB_DNODE(db), doi);
1945 DB_DNODE_EXIT(db);
1949 * Faster still when you only care about the size.
1950 * This is specifically optimized for zfs_getattr().
1952 void
1953 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
1954 u_longlong_t *nblk512)
1956 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1957 dnode_t *dn;
1959 DB_DNODE_ENTER(db);
1960 dn = DB_DNODE(db);
1962 *blksize = dn->dn_datablksz;
1963 /* add 1 for dnode space */
1964 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
1965 SPA_MINBLOCKSHIFT) + 1;
1966 DB_DNODE_EXIT(db);
1969 void
1970 byteswap_uint64_array(void *vbuf, size_t size)
1972 uint64_t *buf = vbuf;
1973 size_t count = size >> 3;
1974 int i;
1976 ASSERT((size & 7) == 0);
1978 for (i = 0; i < count; i++)
1979 buf[i] = BSWAP_64(buf[i]);
1982 void
1983 byteswap_uint32_array(void *vbuf, size_t size)
1985 uint32_t *buf = vbuf;
1986 size_t count = size >> 2;
1987 int i;
1989 ASSERT((size & 3) == 0);
1991 for (i = 0; i < count; i++)
1992 buf[i] = BSWAP_32(buf[i]);
1995 void
1996 byteswap_uint16_array(void *vbuf, size_t size)
1998 uint16_t *buf = vbuf;
1999 size_t count = size >> 1;
2000 int i;
2002 ASSERT((size & 1) == 0);
2004 for (i = 0; i < count; i++)
2005 buf[i] = BSWAP_16(buf[i]);
2008 /* ARGSUSED */
2009 void
2010 byteswap_uint8_array(void *vbuf, size_t size)
2014 void
2015 dmu_init(void)
2017 zfs_dbgmsg_init();
2018 sa_cache_init();
2019 xuio_stat_init();
2020 dmu_objset_init();
2021 dnode_init();
2022 dbuf_init();
2023 zfetch_init();
2024 l2arc_init();
2025 arc_init();
2028 void
2029 dmu_fini(void)
2031 arc_fini(); /* arc depends on l2arc, so arc must go first */
2032 l2arc_fini();
2033 zfetch_fini();
2034 dbuf_fini();
2035 dnode_fini();
2036 dmu_objset_fini();
2037 xuio_stat_fini();
2038 sa_cache_fini();
2039 zfs_dbgmsg_fini();