]> git.proxmox.com Git - mirror_zfs.git/blob - module/zfs/dmu.c
projects / mirror_zfs.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Illumos 4370, 4371
[mirror_zfs.git] / module / zfs / dmu.c
1 /*
2 * CDDL HEADER START
3 *
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.
7 *
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.
12 *
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]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2013 by Delphix. All rights reserved.
24 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
25 */
26
27 #include <sys/dmu.h>
28 #include <sys/dmu_impl.h>
29 #include <sys/dmu_tx.h>
30 #include <sys/dbuf.h>
31 #include <sys/dnode.h>
32 #include <sys/zfs_context.h>
33 #include <sys/dmu_objset.h>
34 #include <sys/dmu_traverse.h>
35 #include <sys/dsl_dataset.h>
36 #include <sys/dsl_dir.h>
37 #include <sys/dsl_pool.h>
38 #include <sys/dsl_synctask.h>
39 #include <sys/dsl_prop.h>
40 #include <sys/dmu_zfetch.h>
41 #include <sys/zfs_ioctl.h>
42 #include <sys/zap.h>
43 #include <sys/zio_checksum.h>
44 #include <sys/zio_compress.h>
45 #include <sys/sa.h>
46 #ifdef _KERNEL
47 #include <sys/vmsystm.h>
48 #include <sys/zfs_znode.h>
49 #endif
50
51 /*
52 * Enable/disable nopwrite feature.
53 */
54 int zfs_nopwrite_enabled = 1;
55
56 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
57 { DMU_BSWAP_UINT8, TRUE, "unallocated" },
58 { DMU_BSWAP_ZAP, TRUE, "object directory" },
59 { DMU_BSWAP_UINT64, TRUE, "object array" },
60 { DMU_BSWAP_UINT8, TRUE, "packed nvlist" },
61 { DMU_BSWAP_UINT64, TRUE, "packed nvlist size" },
62 { DMU_BSWAP_UINT64, TRUE, "bpobj" },
63 { DMU_BSWAP_UINT64, TRUE, "bpobj header" },
64 { DMU_BSWAP_UINT64, TRUE, "SPA space map header" },
65 { DMU_BSWAP_UINT64, TRUE, "SPA space map" },
66 { DMU_BSWAP_UINT64, TRUE, "ZIL intent log" },
67 { DMU_BSWAP_DNODE, TRUE, "DMU dnode" },
68 { DMU_BSWAP_OBJSET, TRUE, "DMU objset" },
69 { DMU_BSWAP_UINT64, TRUE, "DSL directory" },
70 { DMU_BSWAP_ZAP, TRUE, "DSL directory child map"},
71 { DMU_BSWAP_ZAP, TRUE, "DSL dataset snap map" },
72 { DMU_BSWAP_ZAP, TRUE, "DSL props" },
73 { DMU_BSWAP_UINT64, TRUE, "DSL dataset" },
74 { DMU_BSWAP_ZNODE, TRUE, "ZFS znode" },
75 { DMU_BSWAP_OLDACL, TRUE, "ZFS V0 ACL" },
76 { DMU_BSWAP_UINT8, FALSE, "ZFS plain file" },
77 { DMU_BSWAP_ZAP, TRUE, "ZFS directory" },
78 { DMU_BSWAP_ZAP, TRUE, "ZFS master node" },
79 { DMU_BSWAP_ZAP, TRUE, "ZFS delete queue" },
80 { DMU_BSWAP_UINT8, FALSE, "zvol object" },
81 { DMU_BSWAP_ZAP, TRUE, "zvol prop" },
82 { DMU_BSWAP_UINT8, FALSE, "other uint8[]" },
83 { DMU_BSWAP_UINT64, FALSE, "other uint64[]" },
84 { DMU_BSWAP_ZAP, TRUE, "other ZAP" },
85 { DMU_BSWAP_ZAP, TRUE, "persistent error log" },
86 { DMU_BSWAP_UINT8, TRUE, "SPA history" },
87 { DMU_BSWAP_UINT64, TRUE, "SPA history offsets" },
88 { DMU_BSWAP_ZAP, TRUE, "Pool properties" },
89 { DMU_BSWAP_ZAP, TRUE, "DSL permissions" },
90 { DMU_BSWAP_ACL, TRUE, "ZFS ACL" },
91 { DMU_BSWAP_UINT8, TRUE, "ZFS SYSACL" },
92 { DMU_BSWAP_UINT8, TRUE, "FUID table" },
93 { DMU_BSWAP_UINT64, TRUE, "FUID table size" },
94 { DMU_BSWAP_ZAP, TRUE, "DSL dataset next clones"},
95 { DMU_BSWAP_ZAP, TRUE, "scan work queue" },
96 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group used" },
97 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group quota" },
98 { DMU_BSWAP_ZAP, TRUE, "snapshot refcount tags"},
99 { DMU_BSWAP_ZAP, TRUE, "DDT ZAP algorithm" },
100 { DMU_BSWAP_ZAP, TRUE, "DDT statistics" },
101 { DMU_BSWAP_UINT8, TRUE, "System attributes" },
102 { DMU_BSWAP_ZAP, TRUE, "SA master node" },
103 { DMU_BSWAP_ZAP, TRUE, "SA attr registration" },
104 { DMU_BSWAP_ZAP, TRUE, "SA attr layouts" },
105 { DMU_BSWAP_ZAP, TRUE, "scan translations" },
106 { DMU_BSWAP_UINT8, FALSE, "deduplicated block" },
107 { DMU_BSWAP_ZAP, TRUE, "DSL deadlist map" },
108 { DMU_BSWAP_UINT64, TRUE, "DSL deadlist map hdr" },
109 { DMU_BSWAP_ZAP, TRUE, "DSL dir clones" },
110 { DMU_BSWAP_UINT64, TRUE, "bpobj subobj" }
111 };
112
113 const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
114 { byteswap_uint8_array, "uint8" },
115 { byteswap_uint16_array, "uint16" },
116 { byteswap_uint32_array, "uint32" },
117 { byteswap_uint64_array, "uint64" },
118 { zap_byteswap, "zap" },
119 { dnode_buf_byteswap, "dnode" },
120 { dmu_objset_byteswap, "objset" },
121 { zfs_znode_byteswap, "znode" },
122 { zfs_oldacl_byteswap, "oldacl" },
123 { zfs_acl_byteswap, "acl" }
124 };
125
126 int
127 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
128 void *tag, dmu_buf_t **dbp, int flags)
129 {
130 dnode_t *dn;
131 uint64_t blkid;
132 dmu_buf_impl_t *db;
133 int err;
134 int db_flags = DB_RF_CANFAIL;
135
136 if (flags & DMU_READ_NO_PREFETCH)
137 db_flags |= DB_RF_NOPREFETCH;
138
139 err = dnode_hold(os, object, FTAG, &dn);
140 if (err)
141 return (err);
142 blkid = dbuf_whichblock(dn, offset);
143 rw_enter(&dn->dn_struct_rwlock, RW_READER);
144 db = dbuf_hold(dn, blkid, tag);
145 rw_exit(&dn->dn_struct_rwlock);
146 if (db == NULL) {
147 err = SET_ERROR(EIO);
148 } else {
149 err = dbuf_read(db, NULL, db_flags);
150 if (err) {
151 dbuf_rele(db, tag);
152 db = NULL;
153 }
154 }
155
156 dnode_rele(dn, FTAG);
157 *dbp = &db->db; /* NULL db plus first field offset is NULL */
158 return (err);
159 }
160
161 int
162 dmu_bonus_max(void)
163 {
164 return (DN_MAX_BONUSLEN);
165 }
166
167 int
168 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
169 {
170 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
171 dnode_t *dn;
172 int error;
173
174 DB_DNODE_ENTER(db);
175 dn = DB_DNODE(db);
176
177 if (dn->dn_bonus != db) {
178 error = SET_ERROR(EINVAL);
179 } else if (newsize < 0 || newsize > db_fake->db_size) {
180 error = SET_ERROR(EINVAL);
181 } else {
182 dnode_setbonuslen(dn, newsize, tx);
183 error = 0;
184 }
185
186 DB_DNODE_EXIT(db);
187 return (error);
188 }
189
190 int
191 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
192 {
193 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
194 dnode_t *dn;
195 int error;
196
197 DB_DNODE_ENTER(db);
198 dn = DB_DNODE(db);
199
200 if (!DMU_OT_IS_VALID(type)) {
201 error = SET_ERROR(EINVAL);
202 } else if (dn->dn_bonus != db) {
203 error = SET_ERROR(EINVAL);
204 } else {
205 dnode_setbonus_type(dn, type, tx);
206 error = 0;
207 }
208
209 DB_DNODE_EXIT(db);
210 return (error);
211 }
212
213 dmu_object_type_t
214 dmu_get_bonustype(dmu_buf_t *db_fake)
215 {
216 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
217 dnode_t *dn;
218 dmu_object_type_t type;
219
220 DB_DNODE_ENTER(db);
221 dn = DB_DNODE(db);
222 type = dn->dn_bonustype;
223 DB_DNODE_EXIT(db);
224
225 return (type);
226 }
227
228 int
229 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
230 {
231 dnode_t *dn;
232 int error;
233
234 error = dnode_hold(os, object, FTAG, &dn);
235 dbuf_rm_spill(dn, tx);
236 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
237 dnode_rm_spill(dn, tx);
238 rw_exit(&dn->dn_struct_rwlock);
239 dnode_rele(dn, FTAG);
240 return (error);
241 }
242
243 /*
244 * returns ENOENT, EIO, or 0.
245 */
246 int
247 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
248 {
249 dnode_t *dn;
250 dmu_buf_impl_t *db;
251 int error;
252
253 error = dnode_hold(os, object, FTAG, &dn);
254 if (error)
255 return (error);
256
257 rw_enter(&dn->dn_struct_rwlock, RW_READER);
258 if (dn->dn_bonus == NULL) {
259 rw_exit(&dn->dn_struct_rwlock);
260 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
261 if (dn->dn_bonus == NULL)
262 dbuf_create_bonus(dn);
263 }
264 db = dn->dn_bonus;
265
266 /* as long as the bonus buf is held, the dnode will be held */
267 if (refcount_add(&db->db_holds, tag) == 1) {
268 VERIFY(dnode_add_ref(dn, db));
269 (void) atomic_inc_32_nv(&dn->dn_dbufs_count);
270 }
271
272 /*
273 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
274 * hold and incrementing the dbuf count to ensure that dnode_move() sees
275 * a dnode hold for every dbuf.
276 */
277 rw_exit(&dn->dn_struct_rwlock);
278
279 dnode_rele(dn, FTAG);
280
281 VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH));
282
283 *dbp = &db->db;
284 return (0);
285 }
286
287 /*
288 * returns ENOENT, EIO, or 0.
289 *
290 * This interface will allocate a blank spill dbuf when a spill blk
291 * doesn't already exist on the dnode.
292 *
293 * if you only want to find an already existing spill db, then
294 * dmu_spill_hold_existing() should be used.
295 */
296 int
297 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
298 {
299 dmu_buf_impl_t *db = NULL;
300 int err;
301
302 if ((flags & DB_RF_HAVESTRUCT) == 0)
303 rw_enter(&dn->dn_struct_rwlock, RW_READER);
304
305 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
306
307 if ((flags & DB_RF_HAVESTRUCT) == 0)
308 rw_exit(&dn->dn_struct_rwlock);
309
310 ASSERT(db != NULL);
311 err = dbuf_read(db, NULL, flags);
312 if (err == 0)
313 *dbp = &db->db;
314 else
315 dbuf_rele(db, tag);
316 return (err);
317 }
318
319 int
320 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
321 {
322 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
323 dnode_t *dn;
324 int err;
325
326 DB_DNODE_ENTER(db);
327 dn = DB_DNODE(db);
328
329 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
330 err = SET_ERROR(EINVAL);
331 } else {
332 rw_enter(&dn->dn_struct_rwlock, RW_READER);
333
334 if (!dn->dn_have_spill) {
335 err = SET_ERROR(ENOENT);
336 } else {
337 err = dmu_spill_hold_by_dnode(dn,
338 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
339 }
340
341 rw_exit(&dn->dn_struct_rwlock);
342 }
343
344 DB_DNODE_EXIT(db);
345 return (err);
346 }
347
348 int
349 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
350 {
351 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
352 dnode_t *dn;
353 int err;
354
355 DB_DNODE_ENTER(db);
356 dn = DB_DNODE(db);
357 err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp);
358 DB_DNODE_EXIT(db);
359
360 return (err);
361 }
362
363 /*
364 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
365 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
366 * and can induce severe lock contention when writing to several files
367 * whose dnodes are in the same block.
368 */
369 static int
370 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
371 int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
372 {
373 dmu_buf_t **dbp;
374 uint64_t blkid, nblks, i;
375 uint32_t dbuf_flags;
376 int err;
377 zio_t *zio;
378
379 ASSERT(length <= DMU_MAX_ACCESS);
380
381 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT;
382 if (flags & DMU_READ_NO_PREFETCH || length > zfetch_array_rd_sz)
383 dbuf_flags |= DB_RF_NOPREFETCH;
384
385 rw_enter(&dn->dn_struct_rwlock, RW_READER);
386 if (dn->dn_datablkshift) {
387 int blkshift = dn->dn_datablkshift;
388 nblks = (P2ROUNDUP(offset+length, 1ULL<<blkshift) -
389 P2ALIGN(offset, 1ULL<<blkshift)) >> blkshift;
390 } else {
391 if (offset + length > dn->dn_datablksz) {
392 zfs_panic_recover("zfs: accessing past end of object "
393 "%llx/%llx (size=%u access=%llu+%llu)",
394 (longlong_t)dn->dn_objset->
395 os_dsl_dataset->ds_object,
396 (longlong_t)dn->dn_object, dn->dn_datablksz,
397 (longlong_t)offset, (longlong_t)length);
398 rw_exit(&dn->dn_struct_rwlock);
399 return (SET_ERROR(EIO));
400 }
401 nblks = 1;
402 }
403 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks,
404 KM_PUSHPAGE | KM_NODEBUG);
405
406 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
407 blkid = dbuf_whichblock(dn, offset);
408 for (i = 0; i < nblks; i++) {
409 dmu_buf_impl_t *db = dbuf_hold(dn, blkid+i, tag);
410 if (db == NULL) {
411 rw_exit(&dn->dn_struct_rwlock);
412 dmu_buf_rele_array(dbp, nblks, tag);
413 zio_nowait(zio);
414 return (SET_ERROR(EIO));
415 }
416 /* initiate async i/o */
417 if (read) {
418 (void) dbuf_read(db, zio, dbuf_flags);
419 }
420 dbp[i] = &db->db;
421 }
422 rw_exit(&dn->dn_struct_rwlock);
423
424 /* wait for async i/o */
425 err = zio_wait(zio);
426 if (err) {
427 dmu_buf_rele_array(dbp, nblks, tag);
428 return (err);
429 }
430
431 /* wait for other io to complete */
432 if (read) {
433 for (i = 0; i < nblks; i++) {
434 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
435 mutex_enter(&db->db_mtx);
436 while (db->db_state == DB_READ ||
437 db->db_state == DB_FILL)
438 cv_wait(&db->db_changed, &db->db_mtx);
439 if (db->db_state == DB_UNCACHED)
440 err = SET_ERROR(EIO);
441 mutex_exit(&db->db_mtx);
442 if (err) {
443 dmu_buf_rele_array(dbp, nblks, tag);
444 return (err);
445 }
446 }
447 }
448
449 *numbufsp = nblks;
450 *dbpp = dbp;
451 return (0);
452 }
453
454 static int
455 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
456 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
457 {
458 dnode_t *dn;
459 int err;
460
461 err = dnode_hold(os, object, FTAG, &dn);
462 if (err)
463 return (err);
464
465 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
466 numbufsp, dbpp, DMU_READ_PREFETCH);
467
468 dnode_rele(dn, FTAG);
469
470 return (err);
471 }
472
473 int
474 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
475 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
476 {
477 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
478 dnode_t *dn;
479 int err;
480
481 DB_DNODE_ENTER(db);
482 dn = DB_DNODE(db);
483 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
484 numbufsp, dbpp, DMU_READ_PREFETCH);
485 DB_DNODE_EXIT(db);
486
487 return (err);
488 }
489
490 void
491 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
492 {
493 int i;
494 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
495
496 if (numbufs == 0)
497 return;
498
499 for (i = 0; i < numbufs; i++) {
500 if (dbp[i])
501 dbuf_rele(dbp[i], tag);
502 }
503
504 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
505 }
506
507 /*
508 * Issue prefetch i/os for the given blocks.
509 *
510 * Note: The assumption is that we *know* these blocks will be needed
511 * almost immediately. Therefore, the prefetch i/os will be issued at
512 * ZIO_PRIORITY_SYNC_READ
513 *
514 * Note: indirect blocks and other metadata will be read synchronously,
515 * causing this function to block if they are not already cached.
516 */
517 void
518 dmu_prefetch(objset_t *os, uint64_t object, uint64_t offset, uint64_t len)
519 {
520 dnode_t *dn;
521 uint64_t blkid;
522 int nblks, err;
523
524 if (zfs_prefetch_disable)
525 return;
526
527 if (len == 0) { /* they're interested in the bonus buffer */
528 dn = DMU_META_DNODE(os);
529
530 if (object == 0 || object >= DN_MAX_OBJECT)
531 return;
532
533 rw_enter(&dn->dn_struct_rwlock, RW_READER);
534 blkid = dbuf_whichblock(dn, object * sizeof (dnode_phys_t));
535 dbuf_prefetch(dn, blkid, ZIO_PRIORITY_SYNC_READ);
536 rw_exit(&dn->dn_struct_rwlock);
537 return;
538 }
539
540 /*
541 * XXX - Note, if the dnode for the requested object is not
542 * already cached, we will do a *synchronous* read in the
543 * dnode_hold() call. The same is true for any indirects.
544 */
545 err = dnode_hold(os, object, FTAG, &dn);
546 if (err != 0)
547 return;
548
549 rw_enter(&dn->dn_struct_rwlock, RW_READER);
550 if (dn->dn_datablkshift) {
551 int blkshift = dn->dn_datablkshift;
552 nblks = (P2ROUNDUP(offset + len, 1 << blkshift) -
553 P2ALIGN(offset, 1 << blkshift)) >> blkshift;
554 } else {
555 nblks = (offset < dn->dn_datablksz);
556 }
557
558 if (nblks != 0) {
559 int i;
560
561 blkid = dbuf_whichblock(dn, offset);
562 for (i = 0; i < nblks; i++)
563 dbuf_prefetch(dn, blkid + i, ZIO_PRIORITY_SYNC_READ);
564 }
565
566 rw_exit(&dn->dn_struct_rwlock);
567
568 dnode_rele(dn, FTAG);
569 }
570
571 /*
572 * Get the next "chunk" of file data to free. We traverse the file from
573 * the end so that the file gets shorter over time (if we crashes in the
574 * middle, this will leave us in a better state). We find allocated file
575 * data by simply searching the allocated level 1 indirects.
576 *
577 * On input, *start should be the first offset that does not need to be
578 * freed (e.g. "offset + length"). On return, *start will be the first
579 * offset that should be freed.
580 */
581 static int
582 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum)
583 {
584 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
585 /* bytes of data covered by a level-1 indirect block */
586 uint64_t iblkrange =
587 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
588 uint64_t blks;
589
590 ASSERT3U(minimum, <=, *start);
591
592 if (*start - minimum <= iblkrange * maxblks) {
593 *start = minimum;
594 return (0);
595 }
596 ASSERT(ISP2(iblkrange));
597
598 for (blks = 0; *start > minimum && blks < maxblks; blks++) {
599 int err;
600
601 /*
602 * dnode_next_offset(BACKWARDS) will find an allocated L1
603 * indirect block at or before the input offset. We must
604 * decrement *start so that it is at the end of the region
605 * to search.
606 */
607 (*start)--;
608 err = dnode_next_offset(dn,
609 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
610
611 /* if there are no indirect blocks before start, we are done */
612 if (err == ESRCH) {
613 *start = minimum;
614 break;
615 } else if (err != 0) {
616 return (err);
617 }
618
619 /* set start to the beginning of this L1 indirect */
620 *start = P2ALIGN(*start, iblkrange);
621 }
622 if (*start < minimum)
623 *start = minimum;
624 return (0);
625 }
626
627 static int
628 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
629 uint64_t length)
630 {
631 uint64_t object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
632 int err;
633
634 if (offset >= object_size)
635 return (0);
636
637 if (length == DMU_OBJECT_END || offset + length > object_size)
638 length = object_size - offset;
639
640 while (length != 0) {
641 uint64_t chunk_end, chunk_begin;
642 dmu_tx_t *tx;
643
644 chunk_end = chunk_begin = offset + length;
645
646 /* move chunk_begin backwards to the beginning of this chunk */
647 err = get_next_chunk(dn, &chunk_begin, offset);
648 if (err)
649 return (err);
650 ASSERT3U(chunk_begin, >=, offset);
651 ASSERT3U(chunk_begin, <=, chunk_end);
652
653 tx = dmu_tx_create(os);
654 dmu_tx_hold_free(tx, dn->dn_object,
655 chunk_begin, chunk_end - chunk_begin);
656 err = dmu_tx_assign(tx, TXG_WAIT);
657 if (err) {
658 dmu_tx_abort(tx);
659 return (err);
660 }
661 dnode_free_range(dn, chunk_begin, chunk_end - chunk_begin, tx);
662 dmu_tx_commit(tx);
663
664 length -= chunk_end - chunk_begin;
665 }
666 return (0);
667 }
668
669 int
670 dmu_free_long_range(objset_t *os, uint64_t object,
671 uint64_t offset, uint64_t length)
672 {
673 dnode_t *dn;
674 int err;
675
676 err = dnode_hold(os, object, FTAG, &dn);
677 if (err != 0)
678 return (err);
679 err = dmu_free_long_range_impl(os, dn, offset, length);
680
681 /*
682 * It is important to zero out the maxblkid when freeing the entire
683 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
684 * will take the fast path, and (b) dnode_reallocate() can verify
685 * that the entire file has been freed.
686 */
687 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
688 dn->dn_maxblkid = 0;
689
690 dnode_rele(dn, FTAG);
691 return (err);
692 }
693
694 int
695 dmu_free_long_object(objset_t *os, uint64_t object)
696 {
697 dmu_tx_t *tx;
698 int err;
699
700 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
701 if (err != 0)
702 return (err);
703
704 tx = dmu_tx_create(os);
705 dmu_tx_hold_bonus(tx, object);
706 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
707 err = dmu_tx_assign(tx, TXG_WAIT);
708 if (err == 0) {
709 err = dmu_object_free(os, object, tx);
710 dmu_tx_commit(tx);
711 } else {
712 dmu_tx_abort(tx);
713 }
714
715 return (err);
716 }
717
718 int
719 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
720 uint64_t size, dmu_tx_t *tx)
721 {
722 dnode_t *dn;
723 int err = dnode_hold(os, object, FTAG, &dn);
724 if (err)
725 return (err);
726 ASSERT(offset < UINT64_MAX);
727 ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
728 dnode_free_range(dn, offset, size, tx);
729 dnode_rele(dn, FTAG);
730 return (0);
731 }
732
733 int
734 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
735 void *buf, uint32_t flags)
736 {
737 dnode_t *dn;
738 dmu_buf_t **dbp;
739 int numbufs, err;
740
741 err = dnode_hold(os, object, FTAG, &dn);
742 if (err)
743 return (err);
744
745 /*
746 * Deal with odd block sizes, where there can't be data past the first
747 * block. If we ever do the tail block optimization, we will need to
748 * handle that here as well.
749 */
750 if (dn->dn_maxblkid == 0) {
751 int newsz = offset > dn->dn_datablksz ? 0 :
752 MIN(size, dn->dn_datablksz - offset);
753 bzero((char *)buf + newsz, size - newsz);
754 size = newsz;
755 }
756
757 while (size > 0) {
758 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
759 int i;
760
761 /*
762 * NB: we could do this block-at-a-time, but it's nice
763 * to be reading in parallel.
764 */
765 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
766 TRUE, FTAG, &numbufs, &dbp, flags);
767 if (err)
768 break;
769
770 for (i = 0; i < numbufs; i++) {
771 int tocpy;
772 int bufoff;
773 dmu_buf_t *db = dbp[i];
774
775 ASSERT(size > 0);
776
777 bufoff = offset - db->db_offset;
778 tocpy = (int)MIN(db->db_size - bufoff, size);
779
780 bcopy((char *)db->db_data + bufoff, buf, tocpy);
781
782 offset += tocpy;
783 size -= tocpy;
784 buf = (char *)buf + tocpy;
785 }
786 dmu_buf_rele_array(dbp, numbufs, FTAG);
787 }
788 dnode_rele(dn, FTAG);
789 return (err);
790 }
791
792 void
793 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
794 const void *buf, dmu_tx_t *tx)
795 {
796 dmu_buf_t **dbp;
797 int numbufs, i;
798
799 if (size == 0)
800 return;
801
802 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
803 FALSE, FTAG, &numbufs, &dbp));
804
805 for (i = 0; i < numbufs; i++) {
806 int tocpy;
807 int bufoff;
808 dmu_buf_t *db = dbp[i];
809
810 ASSERT(size > 0);
811
812 bufoff = offset - db->db_offset;
813 tocpy = (int)MIN(db->db_size - bufoff, size);
814
815 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
816
817 if (tocpy == db->db_size)
818 dmu_buf_will_fill(db, tx);
819 else
820 dmu_buf_will_dirty(db, tx);
821
822 (void) memcpy((char *)db->db_data + bufoff, buf, tocpy);
823
824 if (tocpy == db->db_size)
825 dmu_buf_fill_done(db, tx);
826
827 offset += tocpy;
828 size -= tocpy;
829 buf = (char *)buf + tocpy;
830 }
831 dmu_buf_rele_array(dbp, numbufs, FTAG);
832 }
833
834 void
835 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
836 dmu_tx_t *tx)
837 {
838 dmu_buf_t **dbp;
839 int numbufs, i;
840
841 if (size == 0)
842 return;
843
844 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
845 FALSE, FTAG, &numbufs, &dbp));
846
847 for (i = 0; i < numbufs; i++) {
848 dmu_buf_t *db = dbp[i];
849
850 dmu_buf_will_not_fill(db, tx);
851 }
852 dmu_buf_rele_array(dbp, numbufs, FTAG);
853 }
854
855 /*
856 * DMU support for xuio
857 */
858 kstat_t *xuio_ksp = NULL;
859
860 typedef struct xuio_stats {
861 /* loaned yet not returned arc_buf */
862 kstat_named_t xuiostat_onloan_rbuf;
863 kstat_named_t xuiostat_onloan_wbuf;
864 /* whether a copy is made when loaning out a read buffer */
865 kstat_named_t xuiostat_rbuf_copied;
866 kstat_named_t xuiostat_rbuf_nocopy;
867 /* whether a copy is made when assigning a write buffer */
868 kstat_named_t xuiostat_wbuf_copied;
869 kstat_named_t xuiostat_wbuf_nocopy;
870 } xuio_stats_t;
871
872 static xuio_stats_t xuio_stats = {
873 { "onloan_read_buf", KSTAT_DATA_UINT64 },
874 { "onloan_write_buf", KSTAT_DATA_UINT64 },
875 { "read_buf_copied", KSTAT_DATA_UINT64 },
876 { "read_buf_nocopy", KSTAT_DATA_UINT64 },
877 { "write_buf_copied", KSTAT_DATA_UINT64 },
878 { "write_buf_nocopy", KSTAT_DATA_UINT64 }
879 };
880
881 #define XUIOSTAT_INCR(stat, val) \
882 atomic_add_64(&xuio_stats.stat.value.ui64, (val))
883 #define XUIOSTAT_BUMP(stat) XUIOSTAT_INCR(stat, 1)
884
885 int
886 dmu_xuio_init(xuio_t *xuio, int nblk)
887 {
888 dmu_xuio_t *priv;
889 uio_t *uio = &xuio->xu_uio;
890
891 uio->uio_iovcnt = nblk;
892 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_PUSHPAGE);
893
894 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_PUSHPAGE);
895 priv->cnt = nblk;
896 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_PUSHPAGE);
897 priv->iovp = uio->uio_iov;
898 XUIO_XUZC_PRIV(xuio) = priv;
899
900 if (XUIO_XUZC_RW(xuio) == UIO_READ)
901 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
902 else
903 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
904
905 return (0);
906 }
907
908 void
909 dmu_xuio_fini(xuio_t *xuio)
910 {
911 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
912 int nblk = priv->cnt;
913
914 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
915 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
916 kmem_free(priv, sizeof (dmu_xuio_t));
917
918 if (XUIO_XUZC_RW(xuio) == UIO_READ)
919 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
920 else
921 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
922 }
923
924 /*
925 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
926 * and increase priv->next by 1.
927 */
928 int
929 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
930 {
931 struct iovec *iov;
932 uio_t *uio = &xuio->xu_uio;
933 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
934 int i = priv->next++;
935
936 ASSERT(i < priv->cnt);
937 ASSERT(off + n <= arc_buf_size(abuf));
938 iov = uio->uio_iov + i;
939 iov->iov_base = (char *)abuf->b_data + off;
940 iov->iov_len = n;
941 priv->bufs[i] = abuf;
942 return (0);
943 }
944
945 int
946 dmu_xuio_cnt(xuio_t *xuio)
947 {
948 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
949 return (priv->cnt);
950 }
951
952 arc_buf_t *
953 dmu_xuio_arcbuf(xuio_t *xuio, int i)
954 {
955 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
956
957 ASSERT(i < priv->cnt);
958 return (priv->bufs[i]);
959 }
960
961 void
962 dmu_xuio_clear(xuio_t *xuio, int i)
963 {
964 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
965
966 ASSERT(i < priv->cnt);
967 priv->bufs[i] = NULL;
968 }
969
970 static void
971 xuio_stat_init(void)
972 {
973 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
974 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
975 KSTAT_FLAG_VIRTUAL);
976 if (xuio_ksp != NULL) {
977 xuio_ksp->ks_data = &xuio_stats;
978 kstat_install(xuio_ksp);
979 }
980 }
981
982 static void
983 xuio_stat_fini(void)
984 {
985 if (xuio_ksp != NULL) {
986 kstat_delete(xuio_ksp);
987 xuio_ksp = NULL;
988 }
989 }
990
991 void
992 xuio_stat_wbuf_copied()
993 {
994 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
995 }
996
997 void
998 xuio_stat_wbuf_nocopy()
999 {
1000 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
1001 }
1002
1003 #ifdef _KERNEL
1004
1005 /*
1006 * Copy up to size bytes between arg_buf and req based on the data direction
1007 * described by the req. If an entire req's data cannot be transfered in one
1008 * pass, you should pass in @req_offset to indicate where to continue. The
1009 * return value is the number of bytes successfully copied to arg_buf.
1010 */
1011 static int
1012 dmu_req_copy(void *arg_buf, int size, struct request *req, size_t req_offset)
1013 {
1014 struct bio_vec bv, *bvp;
1015 struct req_iterator iter;
1016 char *bv_buf;
1017 int tocpy, bv_len, bv_offset;
1018 int offset = 0;
1019
1020 rq_for_each_segment4(bv, bvp, req, iter) {
1021 /*
1022 * Fully consumed the passed arg_buf. We use goto here because
1023 * rq_for_each_segment is a double loop
1024 */
1025 ASSERT3S(offset, <=, size);
1026 if (size == offset)
1027 goto out;
1028
1029 /* Skip already copied bv */
1030 if (req_offset >= bv.bv_len) {
1031 req_offset -= bv.bv_len;
1032 continue;
1033 }
1034
1035 bv_len = bv.bv_len - req_offset;
1036 bv_offset = bv.bv_offset + req_offset;
1037 req_offset = 0;
1038
1039 tocpy = MIN(bv_len, size - offset);
1040 ASSERT3S(tocpy, >=, 0);
1041
1042 bv_buf = page_address(bv.bv_page) + bv_offset;
1043 ASSERT3P(bv_buf, !=, NULL);
1044
1045 if (rq_data_dir(req) == WRITE)
1046 memcpy(arg_buf + offset, bv_buf, tocpy);
1047 else
1048 memcpy(bv_buf, arg_buf + offset, tocpy);
1049
1050 offset += tocpy;
1051 }
1052 out:
1053 return (offset);
1054 }
1055
1056 int
1057 dmu_read_req(objset_t *os, uint64_t object, struct request *req)
1058 {
1059 uint64_t size = blk_rq_bytes(req);
1060 uint64_t offset = blk_rq_pos(req) << 9;
1061 dmu_buf_t **dbp;
1062 int numbufs, i, err;
1063 size_t req_offset;
1064
1065 /*
1066 * NB: we could do this block-at-a-time, but it's nice
1067 * to be reading in parallel.
1068 */
1069 err = dmu_buf_hold_array(os, object, offset, size, TRUE, FTAG,
1070 &numbufs, &dbp);
1071 if (err)
1072 return (err);
1073
1074 req_offset = 0;
1075 for (i = 0; i < numbufs; i++) {
1076 int tocpy, didcpy, bufoff;
1077 dmu_buf_t *db = dbp[i];
1078
1079 bufoff = offset - db->db_offset;
1080 ASSERT3S(bufoff, >=, 0);
1081
1082 tocpy = (int)MIN(db->db_size - bufoff, size);
1083 if (tocpy == 0)
1084 break;
1085
1086 didcpy = dmu_req_copy(db->db_data + bufoff, tocpy, req,
1087 req_offset);
1088
1089 if (didcpy < tocpy)
1090 err = EIO;
1091
1092 if (err)
1093 break;
1094
1095 size -= tocpy;
1096 offset += didcpy;
1097 req_offset += didcpy;
1098 err = 0;
1099 }
1100 dmu_buf_rele_array(dbp, numbufs, FTAG);
1101
1102 return (err);
1103 }
1104
1105 int
1106 dmu_write_req(objset_t *os, uint64_t object, struct request *req, dmu_tx_t *tx)
1107 {
1108 uint64_t size = blk_rq_bytes(req);
1109 uint64_t offset = blk_rq_pos(req) << 9;
1110 dmu_buf_t **dbp;
1111 int numbufs, i, err;
1112 size_t req_offset;
1113
1114 if (size == 0)
1115 return (0);
1116
1117 err = dmu_buf_hold_array(os, object, offset, size, FALSE, FTAG,
1118 &numbufs, &dbp);
1119 if (err)
1120 return (err);
1121
1122 req_offset = 0;
1123 for (i = 0; i < numbufs; i++) {
1124 int tocpy, didcpy, bufoff;
1125 dmu_buf_t *db = dbp[i];
1126
1127 bufoff = offset - db->db_offset;
1128 ASSERT3S(bufoff, >=, 0);
1129
1130 tocpy = (int)MIN(db->db_size - bufoff, size);
1131 if (tocpy == 0)
1132 break;
1133
1134 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1135
1136 if (tocpy == db->db_size)
1137 dmu_buf_will_fill(db, tx);
1138 else
1139 dmu_buf_will_dirty(db, tx);
1140
1141 didcpy = dmu_req_copy(db->db_data + bufoff, tocpy, req,
1142 req_offset);
1143
1144 if (tocpy == db->db_size)
1145 dmu_buf_fill_done(db, tx);
1146
1147 if (didcpy < tocpy)
1148 err = EIO;
1149
1150 if (err)
1151 break;
1152
1153 size -= tocpy;
1154 offset += didcpy;
1155 req_offset += didcpy;
1156 err = 0;
1157 }
1158
1159 dmu_buf_rele_array(dbp, numbufs, FTAG);
1160 return (err);
1161 }
1162
1163 int
1164 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
1165 {
1166 dmu_buf_t **dbp;
1167 int numbufs, i, err;
1168 xuio_t *xuio = NULL;
1169
1170 /*
1171 * NB: we could do this block-at-a-time, but it's nice
1172 * to be reading in parallel.
1173 */
1174 err = dmu_buf_hold_array(os, object, uio->uio_loffset, size, TRUE, FTAG,
1175 &numbufs, &dbp);
1176 if (err)
1177 return (err);
1178
1179 for (i = 0; i < numbufs; i++) {
1180 int tocpy;
1181 int bufoff;
1182 dmu_buf_t *db = dbp[i];
1183
1184 ASSERT(size > 0);
1185
1186 bufoff = uio->uio_loffset - db->db_offset;
1187 tocpy = (int)MIN(db->db_size - bufoff, size);
1188
1189 if (xuio) {
1190 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
1191 arc_buf_t *dbuf_abuf = dbi->db_buf;
1192 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
1193 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
1194 if (!err) {
1195 uio->uio_resid -= tocpy;
1196 uio->uio_loffset += tocpy;
1197 }
1198
1199 if (abuf == dbuf_abuf)
1200 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
1201 else
1202 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
1203 } else {
1204 err = uiomove((char *)db->db_data + bufoff, tocpy,
1205 UIO_READ, uio);
1206 }
1207 if (err)
1208 break;
1209
1210 size -= tocpy;
1211 }
1212 dmu_buf_rele_array(dbp, numbufs, FTAG);
1213
1214 return (err);
1215 }
1216
1217 static int
1218 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1219 {
1220 dmu_buf_t **dbp;
1221 int numbufs;
1222 int err = 0;
1223 int i;
1224
1225 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1226 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1227 if (err)
1228 return (err);
1229
1230 for (i = 0; i < numbufs; i++) {
1231 int tocpy;
1232 int bufoff;
1233 dmu_buf_t *db = dbp[i];
1234
1235 ASSERT(size > 0);
1236
1237 bufoff = uio->uio_loffset - db->db_offset;
1238 tocpy = (int)MIN(db->db_size - bufoff, size);
1239
1240 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1241
1242 if (tocpy == db->db_size)
1243 dmu_buf_will_fill(db, tx);
1244 else
1245 dmu_buf_will_dirty(db, tx);
1246
1247 /*
1248 * XXX uiomove could block forever (eg.nfs-backed
1249 * pages). There needs to be a uiolockdown() function
1250 * to lock the pages in memory, so that uiomove won't
1251 * block.
1252 */
1253 err = uiomove((char *)db->db_data + bufoff, tocpy,
1254 UIO_WRITE, uio);
1255
1256 if (tocpy == db->db_size)
1257 dmu_buf_fill_done(db, tx);
1258
1259 if (err)
1260 break;
1261
1262 size -= tocpy;
1263 }
1264
1265 dmu_buf_rele_array(dbp, numbufs, FTAG);
1266 return (err);
1267 }
1268
1269 int
1270 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1271 dmu_tx_t *tx)
1272 {
1273 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1274 dnode_t *dn;
1275 int err;
1276
1277 if (size == 0)
1278 return (0);
1279
1280 DB_DNODE_ENTER(db);
1281 dn = DB_DNODE(db);
1282 err = dmu_write_uio_dnode(dn, uio, size, tx);
1283 DB_DNODE_EXIT(db);
1284
1285 return (err);
1286 }
1287
1288 int
1289 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1290 dmu_tx_t *tx)
1291 {
1292 dnode_t *dn;
1293 int err;
1294
1295 if (size == 0)
1296 return (0);
1297
1298 err = dnode_hold(os, object, FTAG, &dn);
1299 if (err)
1300 return (err);
1301
1302 err = dmu_write_uio_dnode(dn, uio, size, tx);
1303
1304 dnode_rele(dn, FTAG);
1305
1306 return (err);
1307 }
1308 #endif /* _KERNEL */
1309
1310 /*
1311 * Allocate a loaned anonymous arc buffer.
1312 */
1313 arc_buf_t *
1314 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1315 {
1316 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1317
1318 return (arc_loan_buf(db->db_objset->os_spa, size));
1319 }
1320
1321 /*
1322 * Free a loaned arc buffer.
1323 */
1324 void
1325 dmu_return_arcbuf(arc_buf_t *buf)
1326 {
1327 arc_return_buf(buf, FTAG);
1328 VERIFY(arc_buf_remove_ref(buf, FTAG));
1329 }
1330
1331 /*
1332 * When possible directly assign passed loaned arc buffer to a dbuf.
1333 * If this is not possible copy the contents of passed arc buf via
1334 * dmu_write().
1335 */
1336 void
1337 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1338 dmu_tx_t *tx)
1339 {
1340 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1341 dnode_t *dn;
1342 dmu_buf_impl_t *db;
1343 uint32_t blksz = (uint32_t)arc_buf_size(buf);
1344 uint64_t blkid;
1345
1346 DB_DNODE_ENTER(dbuf);
1347 dn = DB_DNODE(dbuf);
1348 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1349 blkid = dbuf_whichblock(dn, offset);
1350 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1351 rw_exit(&dn->dn_struct_rwlock);
1352 DB_DNODE_EXIT(dbuf);
1353
1354 if (offset == db->db.db_offset && blksz == db->db.db_size) {
1355 dbuf_assign_arcbuf(db, buf, tx);
1356 dbuf_rele(db, FTAG);
1357 } else {
1358 objset_t *os;
1359 uint64_t object;
1360
1361 DB_DNODE_ENTER(dbuf);
1362 dn = DB_DNODE(dbuf);
1363 os = dn->dn_objset;
1364 object = dn->dn_object;
1365 DB_DNODE_EXIT(dbuf);
1366
1367 dbuf_rele(db, FTAG);
1368 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1369 dmu_return_arcbuf(buf);
1370 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1371 }
1372 }
1373
1374 typedef struct {
1375 dbuf_dirty_record_t *dsa_dr;
1376 dmu_sync_cb_t *dsa_done;
1377 zgd_t *dsa_zgd;
1378 dmu_tx_t *dsa_tx;
1379 } dmu_sync_arg_t;
1380
1381 /* ARGSUSED */
1382 static void
1383 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1384 {
1385 dmu_sync_arg_t *dsa = varg;
1386 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1387 blkptr_t *bp = zio->io_bp;
1388
1389 if (zio->io_error == 0) {
1390 if (BP_IS_HOLE(bp)) {
1391 /*
1392 * A block of zeros may compress to a hole, but the
1393 * block size still needs to be known for replay.
1394 */
1395 BP_SET_LSIZE(bp, db->db_size);
1396 } else {
1397 ASSERT(BP_GET_LEVEL(bp) == 0);
1398 bp->blk_fill = 1;
1399 }
1400 }
1401 }
1402
1403 static void
1404 dmu_sync_late_arrival_ready(zio_t *zio)
1405 {
1406 dmu_sync_ready(zio, NULL, zio->io_private);
1407 }
1408
1409 /* ARGSUSED */
1410 static void
1411 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1412 {
1413 dmu_sync_arg_t *dsa = varg;
1414 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1415 dmu_buf_impl_t *db = dr->dr_dbuf;
1416
1417 mutex_enter(&db->db_mtx);
1418 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1419 if (zio->io_error == 0) {
1420 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1421 if (dr->dt.dl.dr_nopwrite) {
1422 ASSERTV(blkptr_t *bp = zio->io_bp);
1423 ASSERTV(blkptr_t *bp_orig = &zio->io_bp_orig);
1424 ASSERTV(uint8_t chksum = BP_GET_CHECKSUM(bp_orig));
1425
1426 ASSERT(BP_EQUAL(bp, bp_orig));
1427 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1428 ASSERT(zio_checksum_table[chksum].ci_dedup);
1429 }
1430 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1431 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1432 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1433 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by))
1434 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1435 } else {
1436 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1437 }
1438 cv_broadcast(&db->db_changed);
1439 mutex_exit(&db->db_mtx);
1440
1441 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1442
1443 kmem_free(dsa, sizeof (*dsa));
1444 }
1445
1446 static void
1447 dmu_sync_late_arrival_done(zio_t *zio)
1448 {
1449 blkptr_t *bp = zio->io_bp;
1450 dmu_sync_arg_t *dsa = zio->io_private;
1451 ASSERTV(blkptr_t *bp_orig = &zio->io_bp_orig);
1452
1453 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1454 /*
1455 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE)
1456 * then there is nothing to do here. Otherwise, free the
1457 * newly allocated block in this txg.
1458 */
1459 if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
1460 ASSERT(BP_EQUAL(bp, bp_orig));
1461 } else {
1462 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1463 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1464 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1465 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1466 }
1467 }
1468
1469 dmu_tx_commit(dsa->dsa_tx);
1470
1471 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1472
1473 kmem_free(dsa, sizeof (*dsa));
1474 }
1475
1476 static int
1477 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1478 zio_prop_t *zp, zbookmark_t *zb)
1479 {
1480 dmu_sync_arg_t *dsa;
1481 dmu_tx_t *tx;
1482
1483 tx = dmu_tx_create(os);
1484 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1485 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1486 dmu_tx_abort(tx);
1487 /* Make zl_get_data do txg_waited_synced() */
1488 return (SET_ERROR(EIO));
1489 }
1490
1491 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_PUSHPAGE);
1492 dsa->dsa_dr = NULL;
1493 dsa->dsa_done = done;
1494 dsa->dsa_zgd = zgd;
1495 dsa->dsa_tx = tx;
1496
1497 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1498 zgd->zgd_db->db_data, zgd->zgd_db->db_size, zp,
1499 dmu_sync_late_arrival_ready, NULL, dmu_sync_late_arrival_done, dsa,
1500 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL|ZIO_FLAG_FASTWRITE, zb));
1501
1502 return (0);
1503 }
1504
1505 /*
1506 * Intent log support: sync the block associated with db to disk.
1507 * N.B. and XXX: the caller is responsible for making sure that the
1508 * data isn't changing while dmu_sync() is writing it.
1509 *
1510 * Return values:
1511 *
1512 * EEXIST: this txg has already been synced, so there's nothing to do.
1513 * The caller should not log the write.
1514 *
1515 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1516 * The caller should not log the write.
1517 *
1518 * EALREADY: this block is already in the process of being synced.
1519 * The caller should track its progress (somehow).
1520 *
1521 * EIO: could not do the I/O.
1522 * The caller should do a txg_wait_synced().
1523 *
1524 * 0: the I/O has been initiated.
1525 * The caller should log this blkptr in the done callback.
1526 * It is possible that the I/O will fail, in which case
1527 * the error will be reported to the done callback and
1528 * propagated to pio from zio_done().
1529 */
1530 int
1531 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1532 {
1533 blkptr_t *bp = zgd->zgd_bp;
1534 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1535 objset_t *os = db->db_objset;
1536 dsl_dataset_t *ds = os->os_dsl_dataset;
1537 dbuf_dirty_record_t *dr;
1538 dmu_sync_arg_t *dsa;
1539 zbookmark_t zb;
1540 zio_prop_t zp;
1541 dnode_t *dn;
1542
1543 ASSERT(pio != NULL);
1544 ASSERT(txg != 0);
1545
1546 SET_BOOKMARK(&zb, ds->ds_object,
1547 db->db.db_object, db->db_level, db->db_blkid);
1548
1549 DB_DNODE_ENTER(db);
1550 dn = DB_DNODE(db);
1551 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1552 DB_DNODE_EXIT(db);
1553
1554 /*
1555 * If we're frozen (running ziltest), we always need to generate a bp.
1556 */
1557 if (txg > spa_freeze_txg(os->os_spa))
1558 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1559
1560 /*
1561 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1562 * and us. If we determine that this txg is not yet syncing,
1563 * but it begins to sync a moment later, that's OK because the
1564 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1565 */
1566 mutex_enter(&db->db_mtx);
1567
1568 if (txg <= spa_last_synced_txg(os->os_spa)) {
1569 /*
1570 * This txg has already synced. There's nothing to do.
1571 */
1572 mutex_exit(&db->db_mtx);
1573 return (SET_ERROR(EEXIST));
1574 }
1575
1576 if (txg <= spa_syncing_txg(os->os_spa)) {
1577 /*
1578 * This txg is currently syncing, so we can't mess with
1579 * the dirty record anymore; just write a new log block.
1580 */
1581 mutex_exit(&db->db_mtx);
1582 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1583 }
1584
1585 dr = db->db_last_dirty;
1586 while (dr && dr->dr_txg != txg)
1587 dr = dr->dr_next;
1588
1589 if (dr == NULL) {
1590 /*
1591 * There's no dr for this dbuf, so it must have been freed.
1592 * There's no need to log writes to freed blocks, so we're done.
1593 */
1594 mutex_exit(&db->db_mtx);
1595 return (SET_ERROR(ENOENT));
1596 }
1597
1598 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
1599
1600 /*
1601 * Assume the on-disk data is X, the current syncing data is Y,
1602 * and the current in-memory data is Z (currently in dmu_sync).
1603 * X and Z are identical but Y is has been modified. Normally,
1604 * when X and Z are the same we will perform a nopwrite but if Y
1605 * is different we must disable nopwrite since the resulting write
1606 * of Y to disk can free the block containing X. If we allowed a
1607 * nopwrite to occur the block pointing to Z would reference a freed
1608 * block. Since this is a rare case we simplify this by disabling
1609 * nopwrite if the current dmu_sync-ing dbuf has been modified in
1610 * a previous transaction.
1611 */
1612 if (dr->dr_next)
1613 zp.zp_nopwrite = B_FALSE;
1614
1615 ASSERT(dr->dr_txg == txg);
1616 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1617 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1618 /*
1619 * We have already issued a sync write for this buffer,
1620 * or this buffer has already been synced. It could not
1621 * have been dirtied since, or we would have cleared the state.
1622 */
1623 mutex_exit(&db->db_mtx);
1624 return (SET_ERROR(EALREADY));
1625 }
1626
1627 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1628 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1629 mutex_exit(&db->db_mtx);
1630
1631 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_PUSHPAGE);
1632 dsa->dsa_dr = dr;
1633 dsa->dsa_done = done;
1634 dsa->dsa_zgd = zgd;
1635 dsa->dsa_tx = NULL;
1636
1637 zio_nowait(arc_write(pio, os->os_spa, txg,
1638 bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
1639 DBUF_IS_L2COMPRESSIBLE(db), &zp, dmu_sync_ready,
1640 NULL, dmu_sync_done, dsa, ZIO_PRIORITY_SYNC_WRITE,
1641 ZIO_FLAG_CANFAIL, &zb));
1642
1643 return (0);
1644 }
1645
1646 int
1647 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1648 dmu_tx_t *tx)
1649 {
1650 dnode_t *dn;
1651 int err;
1652
1653 err = dnode_hold(os, object, FTAG, &dn);
1654 if (err)
1655 return (err);
1656 err = dnode_set_blksz(dn, size, ibs, tx);
1657 dnode_rele(dn, FTAG);
1658 return (err);
1659 }
1660
1661 void
1662 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1663 dmu_tx_t *tx)
1664 {
1665 dnode_t *dn;
1666
1667 /* XXX assumes dnode_hold will not get an i/o error */
1668 (void) dnode_hold(os, object, FTAG, &dn);
1669 ASSERT(checksum < ZIO_CHECKSUM_FUNCTIONS);
1670 dn->dn_checksum = checksum;
1671 dnode_setdirty(dn, tx);
1672 dnode_rele(dn, FTAG);
1673 }
1674
1675 void
1676 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1677 dmu_tx_t *tx)
1678 {
1679 dnode_t *dn;
1680
1681 /* XXX assumes dnode_hold will not get an i/o error */
1682 (void) dnode_hold(os, object, FTAG, &dn);
1683 ASSERT(compress < ZIO_COMPRESS_FUNCTIONS);
1684 dn->dn_compress = compress;
1685 dnode_setdirty(dn, tx);
1686 dnode_rele(dn, FTAG);
1687 }
1688
1689 int zfs_mdcomp_disable = 0;
1690
1691 void
1692 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
1693 {
1694 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
1695 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
1696 (wp & WP_SPILL));
1697 enum zio_checksum checksum = os->os_checksum;
1698 enum zio_compress compress = os->os_compress;
1699 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
1700 boolean_t dedup = B_FALSE;
1701 boolean_t nopwrite = B_FALSE;
1702 boolean_t dedup_verify = os->os_dedup_verify;
1703 int copies = os->os_copies;
1704
1705 /*
1706 * We maintain different write policies for each of the following
1707 * types of data:
1708 * 1. metadata
1709 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1710 * 3. all other level 0 blocks
1711 */
1712 if (ismd) {
1713 /*
1714 * XXX -- we should design a compression algorithm
1715 * that specializes in arrays of bps.
1716 */
1717 compress = zfs_mdcomp_disable ? ZIO_COMPRESS_EMPTY :
1718 ZIO_COMPRESS_LZJB;
1719
1720 /*
1721 * Metadata always gets checksummed. If the data
1722 * checksum is multi-bit correctable, and it's not a
1723 * ZBT-style checksum, then it's suitable for metadata
1724 * as well. Otherwise, the metadata checksum defaults
1725 * to fletcher4.
1726 */
1727 if (zio_checksum_table[checksum].ci_correctable < 1 ||
1728 zio_checksum_table[checksum].ci_eck)
1729 checksum = ZIO_CHECKSUM_FLETCHER_4;
1730 } else if (wp & WP_NOFILL) {
1731 ASSERT(level == 0);
1732
1733 /*
1734 * If we're writing preallocated blocks, we aren't actually
1735 * writing them so don't set any policy properties. These
1736 * blocks are currently only used by an external subsystem
1737 * outside of zfs (i.e. dump) and not written by the zio
1738 * pipeline.
1739 */
1740 compress = ZIO_COMPRESS_OFF;
1741 checksum = ZIO_CHECKSUM_OFF;
1742 } else {
1743 compress = zio_compress_select(dn->dn_compress, compress);
1744
1745 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
1746 zio_checksum_select(dn->dn_checksum, checksum) :
1747 dedup_checksum;
1748
1749 /*
1750 * Determine dedup setting. If we are in dmu_sync(),
1751 * we won't actually dedup now because that's all
1752 * done in syncing context; but we do want to use the
1753 * dedup checkum. If the checksum is not strong
1754 * enough to ensure unique signatures, force
1755 * dedup_verify.
1756 */
1757 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
1758 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
1759 if (!zio_checksum_table[checksum].ci_dedup)
1760 dedup_verify = B_TRUE;
1761 }
1762
1763 /*
1764 * Enable nopwrite if we have a cryptographically secure
1765 * checksum that has no known collisions (i.e. SHA-256)
1766 * and compression is enabled. We don't enable nopwrite if
1767 * dedup is enabled as the two features are mutually exclusive.
1768 */
1769 nopwrite = (!dedup && zio_checksum_table[checksum].ci_dedup &&
1770 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
1771 }
1772
1773 zp->zp_checksum = checksum;
1774 zp->zp_compress = compress;
1775 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
1776 zp->zp_level = level;
1777 zp->zp_copies = MIN(copies + ismd, spa_max_replication(os->os_spa));
1778 zp->zp_dedup = dedup;
1779 zp->zp_dedup_verify = dedup && dedup_verify;
1780 zp->zp_nopwrite = nopwrite;
1781 }
1782
1783 int
1784 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
1785 {
1786 dnode_t *dn;
1787 int i, err;
1788
1789 err = dnode_hold(os, object, FTAG, &dn);
1790 if (err)
1791 return (err);
1792 /*
1793 * Sync any current changes before
1794 * we go trundling through the block pointers.
1795 */
1796 for (i = 0; i < TXG_SIZE; i++) {
1797 if (list_link_active(&dn->dn_dirty_link[i]))
1798 break;
1799 }
1800 if (i != TXG_SIZE) {
1801 dnode_rele(dn, FTAG);
1802 txg_wait_synced(dmu_objset_pool(os), 0);
1803 err = dnode_hold(os, object, FTAG, &dn);
1804 if (err)
1805 return (err);
1806 }
1807
1808 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
1809 dnode_rele(dn, FTAG);
1810
1811 return (err);
1812 }
1813
1814 void
1815 __dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
1816 {
1817 dnode_phys_t *dnp = dn->dn_phys;
1818 int i;
1819
1820 doi->doi_data_block_size = dn->dn_datablksz;
1821 doi->doi_metadata_block_size = dn->dn_indblkshift ?
1822 1ULL << dn->dn_indblkshift : 0;
1823 doi->doi_type = dn->dn_type;
1824 doi->doi_bonus_type = dn->dn_bonustype;
1825 doi->doi_bonus_size = dn->dn_bonuslen;
1826 doi->doi_indirection = dn->dn_nlevels;
1827 doi->doi_checksum = dn->dn_checksum;
1828 doi->doi_compress = dn->dn_compress;
1829 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
1830 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
1831 doi->doi_fill_count = 0;
1832 for (i = 0; i < dnp->dn_nblkptr; i++)
1833 doi->doi_fill_count += dnp->dn_blkptr[i].blk_fill;
1834 }
1835
1836 void
1837 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
1838 {
1839 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1840 mutex_enter(&dn->dn_mtx);
1841
1842 __dmu_object_info_from_dnode(dn, doi);
1843
1844 mutex_exit(&dn->dn_mtx);
1845 rw_exit(&dn->dn_struct_rwlock);
1846 }
1847
1848 /*
1849 * Get information on a DMU object.
1850 * If doi is NULL, just indicates whether the object exists.
1851 */
1852 int
1853 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
1854 {
1855 dnode_t *dn;
1856 int err = dnode_hold(os, object, FTAG, &dn);
1857
1858 if (err)
1859 return (err);
1860
1861 if (doi != NULL)
1862 dmu_object_info_from_dnode(dn, doi);
1863
1864 dnode_rele(dn, FTAG);
1865 return (0);
1866 }
1867
1868 /*
1869 * As above, but faster; can be used when you have a held dbuf in hand.
1870 */
1871 void
1872 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
1873 {
1874 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1875
1876 DB_DNODE_ENTER(db);
1877 dmu_object_info_from_dnode(DB_DNODE(db), doi);
1878 DB_DNODE_EXIT(db);
1879 }
1880
1881 /*
1882 * Faster still when you only care about the size.
1883 * This is specifically optimized for zfs_getattr().
1884 */
1885 void
1886 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
1887 u_longlong_t *nblk512)
1888 {
1889 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1890 dnode_t *dn;
1891
1892 DB_DNODE_ENTER(db);
1893 dn = DB_DNODE(db);
1894
1895 *blksize = dn->dn_datablksz;
1896 /* add 1 for dnode space */
1897 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
1898 SPA_MINBLOCKSHIFT) + 1;
1899 DB_DNODE_EXIT(db);
1900 }
1901
1902 void
1903 byteswap_uint64_array(void *vbuf, size_t size)
1904 {
1905 uint64_t *buf = vbuf;
1906 size_t count = size >> 3;
1907 int i;
1908
1909 ASSERT((size & 7) == 0);
1910
1911 for (i = 0; i < count; i++)
1912 buf[i] = BSWAP_64(buf[i]);
1913 }
1914
1915 void
1916 byteswap_uint32_array(void *vbuf, size_t size)
1917 {
1918 uint32_t *buf = vbuf;
1919 size_t count = size >> 2;
1920 int i;
1921
1922 ASSERT((size & 3) == 0);
1923
1924 for (i = 0; i < count; i++)
1925 buf[i] = BSWAP_32(buf[i]);
1926 }
1927
1928 void
1929 byteswap_uint16_array(void *vbuf, size_t size)
1930 {
1931 uint16_t *buf = vbuf;
1932 size_t count = size >> 1;
1933 int i;
1934
1935 ASSERT((size & 1) == 0);
1936
1937 for (i = 0; i < count; i++)
1938 buf[i] = BSWAP_16(buf[i]);
1939 }
1940
1941 /* ARGSUSED */
1942 void
1943 byteswap_uint8_array(void *vbuf, size_t size)
1944 {
1945 }
1946
1947 void
1948 dmu_init(void)
1949 {
1950 zfs_dbgmsg_init();
1951 sa_cache_init();
1952 xuio_stat_init();
1953 dmu_objset_init();
1954 dnode_init();
1955 dbuf_init();
1956 zfetch_init();
1957 dmu_tx_init();
1958 l2arc_init();
1959 arc_init();
1960 }
1961
1962 void
1963 dmu_fini(void)
1964 {
1965 arc_fini(); /* arc depends on l2arc, so arc must go first */
1966 l2arc_fini();
1967 dmu_tx_fini();
1968 zfetch_fini();
1969 dbuf_fini();
1970 dnode_fini();
1971 dmu_objset_fini();
1972 xuio_stat_fini();
1973 sa_cache_fini();
1974 zfs_dbgmsg_fini();
1975 }
1976
1977 #if defined(_KERNEL) && defined(HAVE_SPL)
1978 EXPORT_SYMBOL(dmu_bonus_hold);
1979 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus);
1980 EXPORT_SYMBOL(dmu_buf_rele_array);
1981 EXPORT_SYMBOL(dmu_prefetch);
1982 EXPORT_SYMBOL(dmu_free_range);
1983 EXPORT_SYMBOL(dmu_free_long_range);
1984 EXPORT_SYMBOL(dmu_free_long_object);
1985 EXPORT_SYMBOL(dmu_read);
1986 EXPORT_SYMBOL(dmu_write);
1987 EXPORT_SYMBOL(dmu_prealloc);
1988 EXPORT_SYMBOL(dmu_object_info);
1989 EXPORT_SYMBOL(dmu_object_info_from_dnode);
1990 EXPORT_SYMBOL(dmu_object_info_from_db);
1991 EXPORT_SYMBOL(dmu_object_size_from_db);
1992 EXPORT_SYMBOL(dmu_object_set_blocksize);
1993 EXPORT_SYMBOL(dmu_object_set_checksum);
1994 EXPORT_SYMBOL(dmu_object_set_compress);
1995 EXPORT_SYMBOL(dmu_write_policy);
1996 EXPORT_SYMBOL(dmu_sync);
1997 EXPORT_SYMBOL(dmu_request_arcbuf);
1998 EXPORT_SYMBOL(dmu_return_arcbuf);
1999 EXPORT_SYMBOL(dmu_assign_arcbuf);
2000 EXPORT_SYMBOL(dmu_buf_hold);
2001 EXPORT_SYMBOL(dmu_ot);
2002
2003 module_param(zfs_mdcomp_disable, int, 0644);
2004 MODULE_PARM_DESC(zfs_mdcomp_disable, "Disable meta data compression");
2005
2006 module_param(zfs_nopwrite_enabled, int, 0644);
2007 MODULE_PARM_DESC(zfs_nopwrite_enabled, "Enable NOP writes");
2008
2009 #endif
ZFS On Linux mirror for PVE