]> 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
Undo c89 workarounds to match with upstream
[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) 2011, 2017 by Delphix. All rights reserved.
24 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
25 * Copyright (c) 2013, Joyent, Inc. All rights reserved.
26 * Copyright (c) 2016, Nexenta Systems, Inc. All rights reserved.
27 * Copyright (c) 2015 by Chunwei Chen. All rights reserved.
28 */
29
30 #include <sys/dmu.h>
31 #include <sys/dmu_impl.h>
32 #include <sys/dmu_tx.h>
33 #include <sys/dbuf.h>
34 #include <sys/dnode.h>
35 #include <sys/zfs_context.h>
36 #include <sys/dmu_objset.h>
37 #include <sys/dmu_traverse.h>
38 #include <sys/dsl_dataset.h>
39 #include <sys/dsl_dir.h>
40 #include <sys/dsl_pool.h>
41 #include <sys/dsl_synctask.h>
42 #include <sys/dsl_prop.h>
43 #include <sys/dmu_zfetch.h>
44 #include <sys/zfs_ioctl.h>
45 #include <sys/zap.h>
46 #include <sys/zio_checksum.h>
47 #include <sys/zio_compress.h>
48 #include <sys/sa.h>
49 #include <sys/zfeature.h>
50 #include <sys/abd.h>
51 #include <sys/trace_dmu.h>
52 #include <sys/zfs_rlock.h>
53 #ifdef _KERNEL
54 #include <sys/vmsystm.h>
55 #include <sys/zfs_znode.h>
56 #endif
57
58 /*
59 * Enable/disable nopwrite feature.
60 */
61 int zfs_nopwrite_enabled = 1;
62
63 /*
64 * Tunable to control percentage of dirtied blocks from frees in one TXG.
65 * After this threshold is crossed, additional dirty blocks from frees
66 * wait until the next TXG.
67 * A value of zero will disable this throttle.
68 */
69 unsigned long zfs_per_txg_dirty_frees_percent = 30;
70
71 /*
72 * Enable/disable forcing txg sync when dirty in dmu_offset_next.
73 */
74 int zfs_dmu_offset_next_sync = 0;
75
76 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
77 { DMU_BSWAP_UINT8, TRUE, FALSE, "unallocated" },
78 { DMU_BSWAP_ZAP, TRUE, FALSE, "object directory" },
79 { DMU_BSWAP_UINT64, TRUE, FALSE, "object array" },
80 { DMU_BSWAP_UINT8, TRUE, FALSE, "packed nvlist" },
81 { DMU_BSWAP_UINT64, TRUE, FALSE, "packed nvlist size" },
82 { DMU_BSWAP_UINT64, TRUE, FALSE, "bpobj" },
83 { DMU_BSWAP_UINT64, TRUE, FALSE, "bpobj header" },
84 { DMU_BSWAP_UINT64, TRUE, FALSE, "SPA space map header" },
85 { DMU_BSWAP_UINT64, TRUE, FALSE, "SPA space map" },
86 { DMU_BSWAP_UINT64, TRUE, TRUE, "ZIL intent log" },
87 { DMU_BSWAP_DNODE, TRUE, TRUE, "DMU dnode" },
88 { DMU_BSWAP_OBJSET, TRUE, FALSE, "DMU objset" },
89 { DMU_BSWAP_UINT64, TRUE, FALSE, "DSL directory" },
90 { DMU_BSWAP_ZAP, TRUE, FALSE, "DSL directory child map"},
91 { DMU_BSWAP_ZAP, TRUE, FALSE, "DSL dataset snap map" },
92 { DMU_BSWAP_ZAP, TRUE, FALSE, "DSL props" },
93 { DMU_BSWAP_UINT64, TRUE, FALSE, "DSL dataset" },
94 { DMU_BSWAP_ZNODE, TRUE, FALSE, "ZFS znode" },
95 { DMU_BSWAP_OLDACL, TRUE, TRUE, "ZFS V0 ACL" },
96 { DMU_BSWAP_UINT8, FALSE, TRUE, "ZFS plain file" },
97 { DMU_BSWAP_ZAP, TRUE, TRUE, "ZFS directory" },
98 { DMU_BSWAP_ZAP, TRUE, FALSE, "ZFS master node" },
99 { DMU_BSWAP_ZAP, TRUE, TRUE, "ZFS delete queue" },
100 { DMU_BSWAP_UINT8, FALSE, TRUE, "zvol object" },
101 { DMU_BSWAP_ZAP, TRUE, FALSE, "zvol prop" },
102 { DMU_BSWAP_UINT8, FALSE, TRUE, "other uint8[]" },
103 { DMU_BSWAP_UINT64, FALSE, TRUE, "other uint64[]" },
104 { DMU_BSWAP_ZAP, TRUE, FALSE, "other ZAP" },
105 { DMU_BSWAP_ZAP, TRUE, FALSE, "persistent error log" },
106 { DMU_BSWAP_UINT8, TRUE, FALSE, "SPA history" },
107 { DMU_BSWAP_UINT64, TRUE, FALSE, "SPA history offsets" },
108 { DMU_BSWAP_ZAP, TRUE, FALSE, "Pool properties" },
109 { DMU_BSWAP_ZAP, TRUE, FALSE, "DSL permissions" },
110 { DMU_BSWAP_ACL, TRUE, TRUE, "ZFS ACL" },
111 { DMU_BSWAP_UINT8, TRUE, TRUE, "ZFS SYSACL" },
112 { DMU_BSWAP_UINT8, TRUE, TRUE, "FUID table" },
113 { DMU_BSWAP_UINT64, TRUE, FALSE, "FUID table size" },
114 { DMU_BSWAP_ZAP, TRUE, FALSE, "DSL dataset next clones"},
115 { DMU_BSWAP_ZAP, TRUE, FALSE, "scan work queue" },
116 { DMU_BSWAP_ZAP, TRUE, TRUE, "ZFS user/group used" },
117 { DMU_BSWAP_ZAP, TRUE, TRUE, "ZFS user/group quota" },
118 { DMU_BSWAP_ZAP, TRUE, FALSE, "snapshot refcount tags"},
119 { DMU_BSWAP_ZAP, TRUE, FALSE, "DDT ZAP algorithm" },
120 { DMU_BSWAP_ZAP, TRUE, FALSE, "DDT statistics" },
121 { DMU_BSWAP_UINT8, TRUE, TRUE, "System attributes" },
122 { DMU_BSWAP_ZAP, TRUE, TRUE, "SA master node" },
123 { DMU_BSWAP_ZAP, TRUE, TRUE, "SA attr registration" },
124 { DMU_BSWAP_ZAP, TRUE, TRUE, "SA attr layouts" },
125 { DMU_BSWAP_ZAP, TRUE, FALSE, "scan translations" },
126 { DMU_BSWAP_UINT8, FALSE, TRUE, "deduplicated block" },
127 { DMU_BSWAP_ZAP, TRUE, FALSE, "DSL deadlist map" },
128 { DMU_BSWAP_UINT64, TRUE, FALSE, "DSL deadlist map hdr" },
129 { DMU_BSWAP_ZAP, TRUE, FALSE, "DSL dir clones" },
130 { DMU_BSWAP_UINT64, TRUE, FALSE, "bpobj subobj" }
131 };
132
133 const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
134 { byteswap_uint8_array, "uint8" },
135 { byteswap_uint16_array, "uint16" },
136 { byteswap_uint32_array, "uint32" },
137 { byteswap_uint64_array, "uint64" },
138 { zap_byteswap, "zap" },
139 { dnode_buf_byteswap, "dnode" },
140 { dmu_objset_byteswap, "objset" },
141 { zfs_znode_byteswap, "znode" },
142 { zfs_oldacl_byteswap, "oldacl" },
143 { zfs_acl_byteswap, "acl" }
144 };
145
146 int
147 dmu_buf_hold_noread_by_dnode(dnode_t *dn, uint64_t offset,
148 void *tag, dmu_buf_t **dbp)
149 {
150 uint64_t blkid;
151 dmu_buf_impl_t *db;
152
153 blkid = dbuf_whichblock(dn, 0, offset);
154 rw_enter(&dn->dn_struct_rwlock, RW_READER);
155 db = dbuf_hold(dn, blkid, tag);
156 rw_exit(&dn->dn_struct_rwlock);
157
158 if (db == NULL) {
159 *dbp = NULL;
160 return (SET_ERROR(EIO));
161 }
162
163 *dbp = &db->db;
164 return (0);
165 }
166 int
167 dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset,
168 void *tag, dmu_buf_t **dbp)
169 {
170 dnode_t *dn;
171 uint64_t blkid;
172 dmu_buf_impl_t *db;
173 int err;
174
175 err = dnode_hold(os, object, FTAG, &dn);
176 if (err)
177 return (err);
178 blkid = dbuf_whichblock(dn, 0, offset);
179 rw_enter(&dn->dn_struct_rwlock, RW_READER);
180 db = dbuf_hold(dn, blkid, tag);
181 rw_exit(&dn->dn_struct_rwlock);
182 dnode_rele(dn, FTAG);
183
184 if (db == NULL) {
185 *dbp = NULL;
186 return (SET_ERROR(EIO));
187 }
188
189 *dbp = &db->db;
190 return (err);
191 }
192
193 int
194 dmu_buf_hold_by_dnode(dnode_t *dn, uint64_t offset,
195 void *tag, dmu_buf_t **dbp, int flags)
196 {
197 int err;
198 int db_flags = DB_RF_CANFAIL;
199
200 if (flags & DMU_READ_NO_PREFETCH)
201 db_flags |= DB_RF_NOPREFETCH;
202 if (flags & DMU_READ_NO_DECRYPT)
203 db_flags |= DB_RF_NO_DECRYPT;
204
205 err = dmu_buf_hold_noread_by_dnode(dn, offset, tag, dbp);
206 if (err == 0) {
207 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
208 err = dbuf_read(db, NULL, db_flags);
209 if (err != 0) {
210 dbuf_rele(db, tag);
211 *dbp = NULL;
212 }
213 }
214
215 return (err);
216 }
217
218 int
219 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
220 void *tag, dmu_buf_t **dbp, int flags)
221 {
222 int err;
223 int db_flags = DB_RF_CANFAIL;
224
225 if (flags & DMU_READ_NO_PREFETCH)
226 db_flags |= DB_RF_NOPREFETCH;
227 if (flags & DMU_READ_NO_DECRYPT)
228 db_flags |= DB_RF_NO_DECRYPT;
229
230 err = dmu_buf_hold_noread(os, object, offset, tag, dbp);
231 if (err == 0) {
232 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
233 err = dbuf_read(db, NULL, db_flags);
234 if (err != 0) {
235 dbuf_rele(db, tag);
236 *dbp = NULL;
237 }
238 }
239
240 return (err);
241 }
242
243 int
244 dmu_bonus_max(void)
245 {
246 return (DN_OLD_MAX_BONUSLEN);
247 }
248
249 int
250 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
251 {
252 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
253 dnode_t *dn;
254 int error;
255
256 DB_DNODE_ENTER(db);
257 dn = DB_DNODE(db);
258
259 if (dn->dn_bonus != db) {
260 error = SET_ERROR(EINVAL);
261 } else if (newsize < 0 || newsize > db_fake->db_size) {
262 error = SET_ERROR(EINVAL);
263 } else {
264 dnode_setbonuslen(dn, newsize, tx);
265 error = 0;
266 }
267
268 DB_DNODE_EXIT(db);
269 return (error);
270 }
271
272 int
273 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
274 {
275 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
276 dnode_t *dn;
277 int error;
278
279 DB_DNODE_ENTER(db);
280 dn = DB_DNODE(db);
281
282 if (!DMU_OT_IS_VALID(type)) {
283 error = SET_ERROR(EINVAL);
284 } else if (dn->dn_bonus != db) {
285 error = SET_ERROR(EINVAL);
286 } else {
287 dnode_setbonus_type(dn, type, tx);
288 error = 0;
289 }
290
291 DB_DNODE_EXIT(db);
292 return (error);
293 }
294
295 dmu_object_type_t
296 dmu_get_bonustype(dmu_buf_t *db_fake)
297 {
298 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
299 dnode_t *dn;
300 dmu_object_type_t type;
301
302 DB_DNODE_ENTER(db);
303 dn = DB_DNODE(db);
304 type = dn->dn_bonustype;
305 DB_DNODE_EXIT(db);
306
307 return (type);
308 }
309
310 int
311 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
312 {
313 dnode_t *dn;
314 int error;
315
316 error = dnode_hold(os, object, FTAG, &dn);
317 dbuf_rm_spill(dn, tx);
318 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
319 dnode_rm_spill(dn, tx);
320 rw_exit(&dn->dn_struct_rwlock);
321 dnode_rele(dn, FTAG);
322 return (error);
323 }
324
325 /*
326 * returns ENOENT, EIO, or 0.
327 */
328 int
329 dmu_bonus_hold_impl(objset_t *os, uint64_t object, void *tag, uint32_t flags,
330 dmu_buf_t **dbp)
331 {
332 dnode_t *dn;
333 dmu_buf_impl_t *db;
334 int error;
335 uint32_t db_flags = DB_RF_MUST_SUCCEED;
336
337 if (flags & DMU_READ_NO_PREFETCH)
338 db_flags |= DB_RF_NOPREFETCH;
339 if (flags & DMU_READ_NO_DECRYPT)
340 db_flags |= DB_RF_NO_DECRYPT;
341
342 error = dnode_hold(os, object, FTAG, &dn);
343 if (error)
344 return (error);
345
346 rw_enter(&dn->dn_struct_rwlock, RW_READER);
347 if (dn->dn_bonus == NULL) {
348 rw_exit(&dn->dn_struct_rwlock);
349 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
350 if (dn->dn_bonus == NULL)
351 dbuf_create_bonus(dn);
352 }
353 db = dn->dn_bonus;
354
355 /* as long as the bonus buf is held, the dnode will be held */
356 if (refcount_add(&db->db_holds, tag) == 1) {
357 VERIFY(dnode_add_ref(dn, db));
358 atomic_inc_32(&dn->dn_dbufs_count);
359 }
360
361 /*
362 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
363 * hold and incrementing the dbuf count to ensure that dnode_move() sees
364 * a dnode hold for every dbuf.
365 */
366 rw_exit(&dn->dn_struct_rwlock);
367
368 dnode_rele(dn, FTAG);
369
370 error = dbuf_read(db, NULL, db_flags);
371 if (error) {
372 dnode_evict_bonus(dn);
373 dbuf_rele(db, tag);
374 *dbp = NULL;
375 return (error);
376 }
377
378 *dbp = &db->db;
379 return (0);
380 }
381
382 int
383 dmu_bonus_hold(objset_t *os, uint64_t obj, void *tag, dmu_buf_t **dbp)
384 {
385 return (dmu_bonus_hold_impl(os, obj, tag, DMU_READ_NO_PREFETCH, dbp));
386 }
387
388 /*
389 * returns ENOENT, EIO, or 0.
390 *
391 * This interface will allocate a blank spill dbuf when a spill blk
392 * doesn't already exist on the dnode.
393 *
394 * if you only want to find an already existing spill db, then
395 * dmu_spill_hold_existing() should be used.
396 */
397 int
398 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
399 {
400 dmu_buf_impl_t *db = NULL;
401 int err;
402
403 if ((flags & DB_RF_HAVESTRUCT) == 0)
404 rw_enter(&dn->dn_struct_rwlock, RW_READER);
405
406 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
407
408 if ((flags & DB_RF_HAVESTRUCT) == 0)
409 rw_exit(&dn->dn_struct_rwlock);
410
411 if (db == NULL) {
412 *dbp = NULL;
413 return (SET_ERROR(EIO));
414 }
415 err = dbuf_read(db, NULL, flags);
416 if (err == 0)
417 *dbp = &db->db;
418 else {
419 dbuf_rele(db, tag);
420 *dbp = NULL;
421 }
422 return (err);
423 }
424
425 int
426 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
427 {
428 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
429 dnode_t *dn;
430 int err;
431
432 DB_DNODE_ENTER(db);
433 dn = DB_DNODE(db);
434
435 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
436 err = SET_ERROR(EINVAL);
437 } else {
438 rw_enter(&dn->dn_struct_rwlock, RW_READER);
439
440 if (!dn->dn_have_spill) {
441 err = SET_ERROR(ENOENT);
442 } else {
443 err = dmu_spill_hold_by_dnode(dn,
444 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
445 }
446
447 rw_exit(&dn->dn_struct_rwlock);
448 }
449
450 DB_DNODE_EXIT(db);
451 return (err);
452 }
453
454 int
455 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
456 {
457 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
458 dnode_t *dn;
459 int err;
460
461 DB_DNODE_ENTER(db);
462 dn = DB_DNODE(db);
463 err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp);
464 DB_DNODE_EXIT(db);
465
466 return (err);
467 }
468
469 /*
470 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
471 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
472 * and can induce severe lock contention when writing to several files
473 * whose dnodes are in the same block.
474 */
475 static int
476 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
477 boolean_t read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
478 {
479 dmu_buf_t **dbp;
480 uint64_t blkid, nblks, i;
481 uint32_t dbuf_flags;
482 int err;
483 zio_t *zio;
484
485 ASSERT(length <= DMU_MAX_ACCESS);
486
487 /*
488 * Note: We directly notify the prefetch code of this read, so that
489 * we can tell it about the multi-block read. dbuf_read() only knows
490 * about the one block it is accessing.
491 */
492 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT |
493 DB_RF_NOPREFETCH;
494
495 rw_enter(&dn->dn_struct_rwlock, RW_READER);
496 if (dn->dn_datablkshift) {
497 int blkshift = dn->dn_datablkshift;
498 nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) -
499 P2ALIGN(offset, 1ULL << blkshift)) >> blkshift;
500 } else {
501 if (offset + length > dn->dn_datablksz) {
502 zfs_panic_recover("zfs: accessing past end of object "
503 "%llx/%llx (size=%u access=%llu+%llu)",
504 (longlong_t)dn->dn_objset->
505 os_dsl_dataset->ds_object,
506 (longlong_t)dn->dn_object, dn->dn_datablksz,
507 (longlong_t)offset, (longlong_t)length);
508 rw_exit(&dn->dn_struct_rwlock);
509 return (SET_ERROR(EIO));
510 }
511 nblks = 1;
512 }
513 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
514
515 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
516 blkid = dbuf_whichblock(dn, 0, offset);
517 for (i = 0; i < nblks; i++) {
518 dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag);
519 if (db == NULL) {
520 rw_exit(&dn->dn_struct_rwlock);
521 dmu_buf_rele_array(dbp, nblks, tag);
522 zio_nowait(zio);
523 return (SET_ERROR(EIO));
524 }
525
526 /* initiate async i/o */
527 if (read)
528 (void) dbuf_read(db, zio, dbuf_flags);
529 dbp[i] = &db->db;
530 }
531
532 if ((flags & DMU_READ_NO_PREFETCH) == 0 &&
533 DNODE_META_IS_CACHEABLE(dn) && length <= zfetch_array_rd_sz) {
534 dmu_zfetch(&dn->dn_zfetch, blkid, nblks,
535 read && DNODE_IS_CACHEABLE(dn));
536 }
537 rw_exit(&dn->dn_struct_rwlock);
538
539 /* wait for async i/o */
540 err = zio_wait(zio);
541 if (err) {
542 dmu_buf_rele_array(dbp, nblks, tag);
543 return (err);
544 }
545
546 /* wait for other io to complete */
547 if (read) {
548 for (i = 0; i < nblks; i++) {
549 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
550 mutex_enter(&db->db_mtx);
551 while (db->db_state == DB_READ ||
552 db->db_state == DB_FILL)
553 cv_wait(&db->db_changed, &db->db_mtx);
554 if (db->db_state == DB_UNCACHED)
555 err = SET_ERROR(EIO);
556 mutex_exit(&db->db_mtx);
557 if (err) {
558 dmu_buf_rele_array(dbp, nblks, tag);
559 return (err);
560 }
561 }
562 }
563
564 *numbufsp = nblks;
565 *dbpp = dbp;
566 return (0);
567 }
568
569 static int
570 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
571 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
572 {
573 dnode_t *dn;
574 int err;
575
576 err = dnode_hold(os, object, FTAG, &dn);
577 if (err)
578 return (err);
579
580 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
581 numbufsp, dbpp, DMU_READ_PREFETCH);
582
583 dnode_rele(dn, FTAG);
584
585 return (err);
586 }
587
588 int
589 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
590 uint64_t length, boolean_t read, void *tag, int *numbufsp,
591 dmu_buf_t ***dbpp)
592 {
593 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
594 dnode_t *dn;
595 int err;
596
597 DB_DNODE_ENTER(db);
598 dn = DB_DNODE(db);
599 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
600 numbufsp, dbpp, DMU_READ_PREFETCH);
601 DB_DNODE_EXIT(db);
602
603 return (err);
604 }
605
606 void
607 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
608 {
609 int i;
610 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
611
612 if (numbufs == 0)
613 return;
614
615 for (i = 0; i < numbufs; i++) {
616 if (dbp[i])
617 dbuf_rele(dbp[i], tag);
618 }
619
620 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
621 }
622
623 /*
624 * Issue prefetch i/os for the given blocks. If level is greater than 0, the
625 * indirect blocks prefeteched will be those that point to the blocks containing
626 * the data starting at offset, and continuing to offset + len.
627 *
628 * Note that if the indirect blocks above the blocks being prefetched are not
629 * in cache, they will be asychronously read in.
630 */
631 void
632 dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset,
633 uint64_t len, zio_priority_t pri)
634 {
635 dnode_t *dn;
636 uint64_t blkid;
637 int nblks, err;
638
639 if (len == 0) { /* they're interested in the bonus buffer */
640 dn = DMU_META_DNODE(os);
641
642 if (object == 0 || object >= DN_MAX_OBJECT)
643 return;
644
645 rw_enter(&dn->dn_struct_rwlock, RW_READER);
646 blkid = dbuf_whichblock(dn, level,
647 object * sizeof (dnode_phys_t));
648 dbuf_prefetch(dn, level, blkid, pri, 0);
649 rw_exit(&dn->dn_struct_rwlock);
650 return;
651 }
652
653 /*
654 * XXX - Note, if the dnode for the requested object is not
655 * already cached, we will do a *synchronous* read in the
656 * dnode_hold() call. The same is true for any indirects.
657 */
658 err = dnode_hold(os, object, FTAG, &dn);
659 if (err != 0)
660 return;
661
662 rw_enter(&dn->dn_struct_rwlock, RW_READER);
663 /*
664 * offset + len - 1 is the last byte we want to prefetch for, and offset
665 * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the
666 * last block we want to prefetch, and dbuf_whichblock(dn, level,
667 * offset) is the first. Then the number we need to prefetch is the
668 * last - first + 1.
669 */
670 if (level > 0 || dn->dn_datablkshift != 0) {
671 nblks = dbuf_whichblock(dn, level, offset + len - 1) -
672 dbuf_whichblock(dn, level, offset) + 1;
673 } else {
674 nblks = (offset < dn->dn_datablksz);
675 }
676
677 if (nblks != 0) {
678 blkid = dbuf_whichblock(dn, level, offset);
679 for (int i = 0; i < nblks; i++)
680 dbuf_prefetch(dn, level, blkid + i, pri, 0);
681 }
682
683 rw_exit(&dn->dn_struct_rwlock);
684
685 dnode_rele(dn, FTAG);
686 }
687
688 /*
689 * Get the next "chunk" of file data to free. We traverse the file from
690 * the end so that the file gets shorter over time (if we crashes in the
691 * middle, this will leave us in a better state). We find allocated file
692 * data by simply searching the allocated level 1 indirects.
693 *
694 * On input, *start should be the first offset that does not need to be
695 * freed (e.g. "offset + length"). On return, *start will be the first
696 * offset that should be freed.
697 */
698 static int
699 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum)
700 {
701 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
702 /* bytes of data covered by a level-1 indirect block */
703 uint64_t iblkrange =
704 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
705
706 ASSERT3U(minimum, <=, *start);
707
708 if (*start - minimum <= iblkrange * maxblks) {
709 *start = minimum;
710 return (0);
711 }
712 ASSERT(ISP2(iblkrange));
713
714 for (uint64_t blks = 0; *start > minimum && blks < maxblks; blks++) {
715 int err;
716
717 /*
718 * dnode_next_offset(BACKWARDS) will find an allocated L1
719 * indirect block at or before the input offset. We must
720 * decrement *start so that it is at the end of the region
721 * to search.
722 */
723 (*start)--;
724 err = dnode_next_offset(dn,
725 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
726
727 /* if there are no indirect blocks before start, we are done */
728 if (err == ESRCH) {
729 *start = minimum;
730 break;
731 } else if (err != 0) {
732 return (err);
733 }
734
735 /* set start to the beginning of this L1 indirect */
736 *start = P2ALIGN(*start, iblkrange);
737 }
738 if (*start < minimum)
739 *start = minimum;
740 return (0);
741 }
742
743 /*
744 * If this objset is of type OST_ZFS return true if vfs's unmounted flag is set,
745 * otherwise return false.
746 * Used below in dmu_free_long_range_impl() to enable abort when unmounting
747 */
748 /*ARGSUSED*/
749 static boolean_t
750 dmu_objset_zfs_unmounting(objset_t *os)
751 {
752 #ifdef _KERNEL
753 if (dmu_objset_type(os) == DMU_OST_ZFS)
754 return (zfs_get_vfs_flag_unmounted(os));
755 #endif
756 return (B_FALSE);
757 }
758
759 static int
760 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
761 uint64_t length, boolean_t raw)
762 {
763 uint64_t object_size;
764 int err;
765 uint64_t dirty_frees_threshold;
766 dsl_pool_t *dp = dmu_objset_pool(os);
767
768 if (dn == NULL)
769 return (SET_ERROR(EINVAL));
770
771 object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
772 if (offset >= object_size)
773 return (0);
774
775 if (zfs_per_txg_dirty_frees_percent <= 100)
776 dirty_frees_threshold =
777 zfs_per_txg_dirty_frees_percent * zfs_dirty_data_max / 100;
778 else
779 dirty_frees_threshold = zfs_dirty_data_max / 4;
780
781 if (length == DMU_OBJECT_END || offset + length > object_size)
782 length = object_size - offset;
783
784 while (length != 0) {
785 uint64_t chunk_end, chunk_begin, chunk_len;
786 uint64_t long_free_dirty_all_txgs = 0;
787 dmu_tx_t *tx;
788
789 if (dmu_objset_zfs_unmounting(dn->dn_objset))
790 return (SET_ERROR(EINTR));
791
792 chunk_end = chunk_begin = offset + length;
793
794 /* move chunk_begin backwards to the beginning of this chunk */
795 err = get_next_chunk(dn, &chunk_begin, offset);
796 if (err)
797 return (err);
798 ASSERT3U(chunk_begin, >=, offset);
799 ASSERT3U(chunk_begin, <=, chunk_end);
800
801 chunk_len = chunk_end - chunk_begin;
802
803 mutex_enter(&dp->dp_lock);
804 for (int t = 0; t < TXG_SIZE; t++) {
805 long_free_dirty_all_txgs +=
806 dp->dp_long_free_dirty_pertxg[t];
807 }
808 mutex_exit(&dp->dp_lock);
809
810 /*
811 * To avoid filling up a TXG with just frees wait for
812 * the next TXG to open before freeing more chunks if
813 * we have reached the threshold of frees
814 */
815 if (dirty_frees_threshold != 0 &&
816 long_free_dirty_all_txgs >= dirty_frees_threshold) {
817 txg_wait_open(dp, 0);
818 continue;
819 }
820
821 tx = dmu_tx_create(os);
822 dmu_tx_hold_free(tx, dn->dn_object, chunk_begin, chunk_len);
823
824 /*
825 * Mark this transaction as typically resulting in a net
826 * reduction in space used.
827 */
828 dmu_tx_mark_netfree(tx);
829 err = dmu_tx_assign(tx, TXG_WAIT);
830 if (err) {
831 dmu_tx_abort(tx);
832 return (err);
833 }
834
835 mutex_enter(&dp->dp_lock);
836 dp->dp_long_free_dirty_pertxg[dmu_tx_get_txg(tx) & TXG_MASK] +=
837 chunk_len;
838 mutex_exit(&dp->dp_lock);
839 DTRACE_PROBE3(free__long__range,
840 uint64_t, long_free_dirty_all_txgs, uint64_t, chunk_len,
841 uint64_t, dmu_tx_get_txg(tx));
842 dnode_free_range(dn, chunk_begin, chunk_len, tx);
843
844 /* if this is a raw free, mark the dirty record as such */
845 if (raw) {
846 dbuf_dirty_record_t *dr = dn->dn_dbuf->db_last_dirty;
847
848 while (dr != NULL && dr->dr_txg > tx->tx_txg)
849 dr = dr->dr_next;
850 if (dr != NULL && dr->dr_txg == tx->tx_txg)
851 dr->dt.dl.dr_raw = B_TRUE;
852 }
853
854 dmu_tx_commit(tx);
855
856 length -= chunk_len;
857 }
858 return (0);
859 }
860
861 int
862 dmu_free_long_range(objset_t *os, uint64_t object,
863 uint64_t offset, uint64_t length)
864 {
865 dnode_t *dn;
866 int err;
867
868 err = dnode_hold(os, object, FTAG, &dn);
869 if (err != 0)
870 return (err);
871 err = dmu_free_long_range_impl(os, dn, offset, length, B_FALSE);
872
873 /*
874 * It is important to zero out the maxblkid when freeing the entire
875 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
876 * will take the fast path, and (b) dnode_reallocate() can verify
877 * that the entire file has been freed.
878 */
879 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
880 dn->dn_maxblkid = 0;
881
882 dnode_rele(dn, FTAG);
883 return (err);
884 }
885
886 /*
887 * This function is equivalent to dmu_free_long_range(), but also
888 * marks the new dirty record as a raw write.
889 */
890 int
891 dmu_free_long_range_raw(objset_t *os, uint64_t object,
892 uint64_t offset, uint64_t length)
893 {
894 dnode_t *dn;
895 int err;
896
897 err = dnode_hold(os, object, FTAG, &dn);
898 if (err != 0)
899 return (err);
900 err = dmu_free_long_range_impl(os, dn, offset, length, B_TRUE);
901
902 /*
903 * It is important to zero out the maxblkid when freeing the entire
904 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
905 * will take the fast path, and (b) dnode_reallocate() can verify
906 * that the entire file has been freed.
907 */
908 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
909 dn->dn_maxblkid = 0;
910
911 dnode_rele(dn, FTAG);
912 return (err);
913 }
914
915 static int
916 dmu_free_long_object_impl(objset_t *os, uint64_t object, boolean_t raw)
917 {
918 dmu_tx_t *tx;
919 int err;
920
921 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
922 if (err != 0)
923 return (err);
924
925 tx = dmu_tx_create(os);
926 dmu_tx_hold_bonus(tx, object);
927 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
928 dmu_tx_mark_netfree(tx);
929 err = dmu_tx_assign(tx, TXG_WAIT);
930 if (err == 0) {
931 if (raw)
932 err = dmu_object_dirty_raw(os, object, tx);
933 if (err == 0)
934 err = dmu_object_free(os, object, tx);
935
936 dmu_tx_commit(tx);
937 } else {
938 dmu_tx_abort(tx);
939 }
940
941 return (err);
942 }
943
944 int
945 dmu_free_long_object(objset_t *os, uint64_t object)
946 {
947 return (dmu_free_long_object_impl(os, object, B_FALSE));
948 }
949
950 int
951 dmu_free_long_object_raw(objset_t *os, uint64_t object)
952 {
953 return (dmu_free_long_object_impl(os, object, B_TRUE));
954 }
955
956
957 int
958 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
959 uint64_t size, dmu_tx_t *tx)
960 {
961 dnode_t *dn;
962 int err = dnode_hold(os, object, FTAG, &dn);
963 if (err)
964 return (err);
965 ASSERT(offset < UINT64_MAX);
966 ASSERT(size == DMU_OBJECT_END || size <= UINT64_MAX - offset);
967 dnode_free_range(dn, offset, size, tx);
968 dnode_rele(dn, FTAG);
969 return (0);
970 }
971
972 static int
973 dmu_read_impl(dnode_t *dn, uint64_t offset, uint64_t size,
974 void *buf, uint32_t flags)
975 {
976 dmu_buf_t **dbp;
977 int numbufs, err = 0;
978
979 /*
980 * Deal with odd block sizes, where there can't be data past the first
981 * block. If we ever do the tail block optimization, we will need to
982 * handle that here as well.
983 */
984 if (dn->dn_maxblkid == 0) {
985 uint64_t newsz = offset > dn->dn_datablksz ? 0 :
986 MIN(size, dn->dn_datablksz - offset);
987 bzero((char *)buf + newsz, size - newsz);
988 size = newsz;
989 }
990
991 while (size > 0) {
992 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
993 int i;
994
995 /*
996 * NB: we could do this block-at-a-time, but it's nice
997 * to be reading in parallel.
998 */
999 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
1000 TRUE, FTAG, &numbufs, &dbp, flags);
1001 if (err)
1002 break;
1003
1004 for (i = 0; i < numbufs; i++) {
1005 uint64_t tocpy;
1006 int64_t bufoff;
1007 dmu_buf_t *db = dbp[i];
1008
1009 ASSERT(size > 0);
1010
1011 bufoff = offset - db->db_offset;
1012 tocpy = MIN(db->db_size - bufoff, size);
1013
1014 (void) memcpy(buf, (char *)db->db_data + bufoff, tocpy);
1015
1016 offset += tocpy;
1017 size -= tocpy;
1018 buf = (char *)buf + tocpy;
1019 }
1020 dmu_buf_rele_array(dbp, numbufs, FTAG);
1021 }
1022 return (err);
1023 }
1024
1025 int
1026 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1027 void *buf, uint32_t flags)
1028 {
1029 dnode_t *dn;
1030 int err;
1031
1032 err = dnode_hold(os, object, FTAG, &dn);
1033 if (err != 0)
1034 return (err);
1035
1036 err = dmu_read_impl(dn, offset, size, buf, flags);
1037 dnode_rele(dn, FTAG);
1038 return (err);
1039 }
1040
1041 int
1042 dmu_read_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, void *buf,
1043 uint32_t flags)
1044 {
1045 return (dmu_read_impl(dn, offset, size, buf, flags));
1046 }
1047
1048 static void
1049 dmu_write_impl(dmu_buf_t **dbp, int numbufs, uint64_t offset, uint64_t size,
1050 const void *buf, dmu_tx_t *tx)
1051 {
1052 int i;
1053
1054 for (i = 0; i < numbufs; i++) {
1055 uint64_t tocpy;
1056 int64_t bufoff;
1057 dmu_buf_t *db = dbp[i];
1058
1059 ASSERT(size > 0);
1060
1061 bufoff = offset - db->db_offset;
1062 tocpy = MIN(db->db_size - bufoff, size);
1063
1064 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1065
1066 if (tocpy == db->db_size)
1067 dmu_buf_will_fill(db, tx);
1068 else
1069 dmu_buf_will_dirty(db, tx);
1070
1071 (void) memcpy((char *)db->db_data + bufoff, buf, tocpy);
1072
1073 if (tocpy == db->db_size)
1074 dmu_buf_fill_done(db, tx);
1075
1076 offset += tocpy;
1077 size -= tocpy;
1078 buf = (char *)buf + tocpy;
1079 }
1080 }
1081
1082 void
1083 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1084 const void *buf, dmu_tx_t *tx)
1085 {
1086 dmu_buf_t **dbp;
1087 int numbufs;
1088
1089 if (size == 0)
1090 return;
1091
1092 VERIFY0(dmu_buf_hold_array(os, object, offset, size,
1093 FALSE, FTAG, &numbufs, &dbp));
1094 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1095 dmu_buf_rele_array(dbp, numbufs, FTAG);
1096 }
1097
1098 void
1099 dmu_write_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size,
1100 const void *buf, dmu_tx_t *tx)
1101 {
1102 dmu_buf_t **dbp;
1103 int numbufs;
1104
1105 if (size == 0)
1106 return;
1107
1108 VERIFY0(dmu_buf_hold_array_by_dnode(dn, offset, size,
1109 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH));
1110 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1111 dmu_buf_rele_array(dbp, numbufs, FTAG);
1112 }
1113
1114 void
1115 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1116 dmu_tx_t *tx)
1117 {
1118 dmu_buf_t **dbp;
1119 int numbufs, i;
1120
1121 if (size == 0)
1122 return;
1123
1124 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
1125 FALSE, FTAG, &numbufs, &dbp));
1126
1127 for (i = 0; i < numbufs; i++) {
1128 dmu_buf_t *db = dbp[i];
1129
1130 dmu_buf_will_not_fill(db, tx);
1131 }
1132 dmu_buf_rele_array(dbp, numbufs, FTAG);
1133 }
1134
1135 void
1136 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
1137 void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
1138 int compressed_size, int byteorder, dmu_tx_t *tx)
1139 {
1140 dmu_buf_t *db;
1141
1142 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
1143 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
1144 VERIFY0(dmu_buf_hold_noread(os, object, offset,
1145 FTAG, &db));
1146
1147 dmu_buf_write_embedded(db,
1148 data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
1149 uncompressed_size, compressed_size, byteorder, tx);
1150
1151 dmu_buf_rele(db, FTAG);
1152 }
1153
1154 /*
1155 * DMU support for xuio
1156 */
1157 kstat_t *xuio_ksp = NULL;
1158
1159 typedef struct xuio_stats {
1160 /* loaned yet not returned arc_buf */
1161 kstat_named_t xuiostat_onloan_rbuf;
1162 kstat_named_t xuiostat_onloan_wbuf;
1163 /* whether a copy is made when loaning out a read buffer */
1164 kstat_named_t xuiostat_rbuf_copied;
1165 kstat_named_t xuiostat_rbuf_nocopy;
1166 /* whether a copy is made when assigning a write buffer */
1167 kstat_named_t xuiostat_wbuf_copied;
1168 kstat_named_t xuiostat_wbuf_nocopy;
1169 } xuio_stats_t;
1170
1171 static xuio_stats_t xuio_stats = {
1172 { "onloan_read_buf", KSTAT_DATA_UINT64 },
1173 { "onloan_write_buf", KSTAT_DATA_UINT64 },
1174 { "read_buf_copied", KSTAT_DATA_UINT64 },
1175 { "read_buf_nocopy", KSTAT_DATA_UINT64 },
1176 { "write_buf_copied", KSTAT_DATA_UINT64 },
1177 { "write_buf_nocopy", KSTAT_DATA_UINT64 }
1178 };
1179
1180 #define XUIOSTAT_INCR(stat, val) \
1181 atomic_add_64(&xuio_stats.stat.value.ui64, (val))
1182 #define XUIOSTAT_BUMP(stat) XUIOSTAT_INCR(stat, 1)
1183
1184 #ifdef HAVE_UIO_ZEROCOPY
1185 int
1186 dmu_xuio_init(xuio_t *xuio, int nblk)
1187 {
1188 dmu_xuio_t *priv;
1189 uio_t *uio = &xuio->xu_uio;
1190
1191 uio->uio_iovcnt = nblk;
1192 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
1193
1194 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
1195 priv->cnt = nblk;
1196 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
1197 priv->iovp = (iovec_t *)uio->uio_iov;
1198 XUIO_XUZC_PRIV(xuio) = priv;
1199
1200 if (XUIO_XUZC_RW(xuio) == UIO_READ)
1201 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
1202 else
1203 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
1204
1205 return (0);
1206 }
1207
1208 void
1209 dmu_xuio_fini(xuio_t *xuio)
1210 {
1211 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1212 int nblk = priv->cnt;
1213
1214 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
1215 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
1216 kmem_free(priv, sizeof (dmu_xuio_t));
1217
1218 if (XUIO_XUZC_RW(xuio) == UIO_READ)
1219 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
1220 else
1221 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
1222 }
1223
1224 /*
1225 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
1226 * and increase priv->next by 1.
1227 */
1228 int
1229 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
1230 {
1231 struct iovec *iov;
1232 uio_t *uio = &xuio->xu_uio;
1233 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1234 int i = priv->next++;
1235
1236 ASSERT(i < priv->cnt);
1237 ASSERT(off + n <= arc_buf_lsize(abuf));
1238 iov = (iovec_t *)uio->uio_iov + i;
1239 iov->iov_base = (char *)abuf->b_data + off;
1240 iov->iov_len = n;
1241 priv->bufs[i] = abuf;
1242 return (0);
1243 }
1244
1245 int
1246 dmu_xuio_cnt(xuio_t *xuio)
1247 {
1248 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1249 return (priv->cnt);
1250 }
1251
1252 arc_buf_t *
1253 dmu_xuio_arcbuf(xuio_t *xuio, int i)
1254 {
1255 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1256
1257 ASSERT(i < priv->cnt);
1258 return (priv->bufs[i]);
1259 }
1260
1261 void
1262 dmu_xuio_clear(xuio_t *xuio, int i)
1263 {
1264 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1265
1266 ASSERT(i < priv->cnt);
1267 priv->bufs[i] = NULL;
1268 }
1269 #endif /* HAVE_UIO_ZEROCOPY */
1270
1271 static void
1272 xuio_stat_init(void)
1273 {
1274 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
1275 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
1276 KSTAT_FLAG_VIRTUAL);
1277 if (xuio_ksp != NULL) {
1278 xuio_ksp->ks_data = &xuio_stats;
1279 kstat_install(xuio_ksp);
1280 }
1281 }
1282
1283 static void
1284 xuio_stat_fini(void)
1285 {
1286 if (xuio_ksp != NULL) {
1287 kstat_delete(xuio_ksp);
1288 xuio_ksp = NULL;
1289 }
1290 }
1291
1292 void
1293 xuio_stat_wbuf_copied(void)
1294 {
1295 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1296 }
1297
1298 void
1299 xuio_stat_wbuf_nocopy(void)
1300 {
1301 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
1302 }
1303
1304 #ifdef _KERNEL
1305 int
1306 dmu_read_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size)
1307 {
1308 dmu_buf_t **dbp;
1309 int numbufs, i, err;
1310 #ifdef HAVE_UIO_ZEROCOPY
1311 xuio_t *xuio = NULL;
1312 #endif
1313
1314 /*
1315 * NB: we could do this block-at-a-time, but it's nice
1316 * to be reading in parallel.
1317 */
1318 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1319 TRUE, FTAG, &numbufs, &dbp, 0);
1320 if (err)
1321 return (err);
1322
1323 for (i = 0; i < numbufs; i++) {
1324 uint64_t tocpy;
1325 int64_t bufoff;
1326 dmu_buf_t *db = dbp[i];
1327
1328 ASSERT(size > 0);
1329
1330 bufoff = uio->uio_loffset - db->db_offset;
1331 tocpy = MIN(db->db_size - bufoff, size);
1332
1333 #ifdef HAVE_UIO_ZEROCOPY
1334 if (xuio) {
1335 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
1336 arc_buf_t *dbuf_abuf = dbi->db_buf;
1337 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
1338 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
1339 if (!err) {
1340 uio->uio_resid -= tocpy;
1341 uio->uio_loffset += tocpy;
1342 }
1343
1344 if (abuf == dbuf_abuf)
1345 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
1346 else
1347 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
1348 } else
1349 #endif
1350 err = uiomove((char *)db->db_data + bufoff, tocpy,
1351 UIO_READ, uio);
1352 if (err)
1353 break;
1354
1355 size -= tocpy;
1356 }
1357 dmu_buf_rele_array(dbp, numbufs, FTAG);
1358
1359 return (err);
1360 }
1361
1362 /*
1363 * Read 'size' bytes into the uio buffer.
1364 * From object zdb->db_object.
1365 * Starting at offset uio->uio_loffset.
1366 *
1367 * If the caller already has a dbuf in the target object
1368 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1369 * because we don't have to find the dnode_t for the object.
1370 */
1371 int
1372 dmu_read_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size)
1373 {
1374 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1375 dnode_t *dn;
1376 int err;
1377
1378 if (size == 0)
1379 return (0);
1380
1381 DB_DNODE_ENTER(db);
1382 dn = DB_DNODE(db);
1383 err = dmu_read_uio_dnode(dn, uio, size);
1384 DB_DNODE_EXIT(db);
1385
1386 return (err);
1387 }
1388
1389 /*
1390 * Read 'size' bytes into the uio buffer.
1391 * From the specified object
1392 * Starting at offset uio->uio_loffset.
1393 */
1394 int
1395 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
1396 {
1397 dnode_t *dn;
1398 int err;
1399
1400 if (size == 0)
1401 return (0);
1402
1403 err = dnode_hold(os, object, FTAG, &dn);
1404 if (err)
1405 return (err);
1406
1407 err = dmu_read_uio_dnode(dn, uio, size);
1408
1409 dnode_rele(dn, FTAG);
1410
1411 return (err);
1412 }
1413
1414 int
1415 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1416 {
1417 dmu_buf_t **dbp;
1418 int numbufs;
1419 int err = 0;
1420 int i;
1421
1422 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1423 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1424 if (err)
1425 return (err);
1426
1427 for (i = 0; i < numbufs; i++) {
1428 uint64_t tocpy;
1429 int64_t bufoff;
1430 dmu_buf_t *db = dbp[i];
1431
1432 ASSERT(size > 0);
1433
1434 bufoff = uio->uio_loffset - db->db_offset;
1435 tocpy = MIN(db->db_size - bufoff, size);
1436
1437 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1438
1439 if (tocpy == db->db_size)
1440 dmu_buf_will_fill(db, tx);
1441 else
1442 dmu_buf_will_dirty(db, tx);
1443
1444 /*
1445 * XXX uiomove could block forever (eg.nfs-backed
1446 * pages). There needs to be a uiolockdown() function
1447 * to lock the pages in memory, so that uiomove won't
1448 * block.
1449 */
1450 err = uiomove((char *)db->db_data + bufoff, tocpy,
1451 UIO_WRITE, uio);
1452
1453 if (tocpy == db->db_size)
1454 dmu_buf_fill_done(db, tx);
1455
1456 if (err)
1457 break;
1458
1459 size -= tocpy;
1460 }
1461
1462 dmu_buf_rele_array(dbp, numbufs, FTAG);
1463 return (err);
1464 }
1465
1466 /*
1467 * Write 'size' bytes from the uio buffer.
1468 * To object zdb->db_object.
1469 * Starting at offset uio->uio_loffset.
1470 *
1471 * If the caller already has a dbuf in the target object
1472 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1473 * because we don't have to find the dnode_t for the object.
1474 */
1475 int
1476 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1477 dmu_tx_t *tx)
1478 {
1479 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1480 dnode_t *dn;
1481 int err;
1482
1483 if (size == 0)
1484 return (0);
1485
1486 DB_DNODE_ENTER(db);
1487 dn = DB_DNODE(db);
1488 err = dmu_write_uio_dnode(dn, uio, size, tx);
1489 DB_DNODE_EXIT(db);
1490
1491 return (err);
1492 }
1493
1494 /*
1495 * Write 'size' bytes from the uio buffer.
1496 * To the specified object.
1497 * Starting at offset uio->uio_loffset.
1498 */
1499 int
1500 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1501 dmu_tx_t *tx)
1502 {
1503 dnode_t *dn;
1504 int err;
1505
1506 if (size == 0)
1507 return (0);
1508
1509 err = dnode_hold(os, object, FTAG, &dn);
1510 if (err)
1511 return (err);
1512
1513 err = dmu_write_uio_dnode(dn, uio, size, tx);
1514
1515 dnode_rele(dn, FTAG);
1516
1517 return (err);
1518 }
1519 #endif /* _KERNEL */
1520
1521 /*
1522 * Allocate a loaned anonymous arc buffer.
1523 */
1524 arc_buf_t *
1525 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1526 {
1527 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1528
1529 return (arc_loan_buf(db->db_objset->os_spa, B_FALSE, size));
1530 }
1531
1532 /*
1533 * Free a loaned arc buffer.
1534 */
1535 void
1536 dmu_return_arcbuf(arc_buf_t *buf)
1537 {
1538 arc_return_buf(buf, FTAG);
1539 arc_buf_destroy(buf, FTAG);
1540 }
1541
1542 void
1543 dmu_convert_to_raw(dmu_buf_t *handle, boolean_t byteorder, const uint8_t *salt,
1544 const uint8_t *iv, const uint8_t *mac, dmu_tx_t *tx)
1545 {
1546 dmu_object_type_t type;
1547 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1548 uint64_t dsobj = dmu_objset_id(db->db_objset);
1549
1550 ASSERT3P(db->db_buf, !=, NULL);
1551 ASSERT3U(dsobj, !=, 0);
1552
1553 dmu_buf_will_change_crypt_params(handle, tx);
1554
1555 DB_DNODE_ENTER(db);
1556 type = DB_DNODE(db)->dn_type;
1557 DB_DNODE_EXIT(db);
1558
1559 /*
1560 * This technically violates the assumption the dmu code makes
1561 * that dnode blocks are only released in syncing context.
1562 */
1563 (void) arc_release(db->db_buf, db);
1564 arc_convert_to_raw(db->db_buf, dsobj, byteorder, type, salt, iv, mac);
1565 }
1566
1567 void
1568 dmu_copy_from_buf(objset_t *os, uint64_t object, uint64_t offset,
1569 dmu_buf_t *handle, dmu_tx_t *tx)
1570 {
1571 dmu_buf_t *dst_handle;
1572 dmu_buf_impl_t *dstdb;
1573 dmu_buf_impl_t *srcdb = (dmu_buf_impl_t *)handle;
1574 arc_buf_t *abuf;
1575 uint64_t datalen;
1576 boolean_t byteorder;
1577 uint8_t salt[ZIO_DATA_SALT_LEN];
1578 uint8_t iv[ZIO_DATA_IV_LEN];
1579 uint8_t mac[ZIO_DATA_MAC_LEN];
1580
1581 ASSERT3P(srcdb->db_buf, !=, NULL);
1582
1583 /* hold the db that we want to write to */
1584 VERIFY0(dmu_buf_hold(os, object, offset, FTAG, &dst_handle,
1585 DMU_READ_NO_DECRYPT));
1586 dstdb = (dmu_buf_impl_t *)dst_handle;
1587 datalen = arc_buf_size(srcdb->db_buf);
1588
1589 /* allocated an arc buffer that matches the type of srcdb->db_buf */
1590 if (arc_is_encrypted(srcdb->db_buf)) {
1591 arc_get_raw_params(srcdb->db_buf, &byteorder, salt, iv, mac);
1592 abuf = arc_loan_raw_buf(os->os_spa, dmu_objset_id(os),
1593 byteorder, salt, iv, mac, DB_DNODE(dstdb)->dn_type,
1594 datalen, arc_buf_lsize(srcdb->db_buf),
1595 arc_get_compression(srcdb->db_buf));
1596 } else {
1597 /* we won't get a compressed db back from dmu_buf_hold() */
1598 ASSERT3U(arc_get_compression(srcdb->db_buf),
1599 ==, ZIO_COMPRESS_OFF);
1600 abuf = arc_loan_buf(os->os_spa,
1601 DMU_OT_IS_METADATA(DB_DNODE(dstdb)->dn_type), datalen);
1602 }
1603
1604 ASSERT3U(datalen, ==, arc_buf_size(abuf));
1605
1606 /* copy the data to the new buffer and assign it to the dstdb */
1607 bcopy(srcdb->db_buf->b_data, abuf->b_data, datalen);
1608 dbuf_assign_arcbuf(dstdb, abuf, tx);
1609 dmu_buf_rele(dst_handle, FTAG);
1610 }
1611
1612 /*
1613 * When possible directly assign passed loaned arc buffer to a dbuf.
1614 * If this is not possible copy the contents of passed arc buf via
1615 * dmu_write().
1616 */
1617 void
1618 dmu_assign_arcbuf_by_dnode(dnode_t *dn, uint64_t offset, arc_buf_t *buf,
1619 dmu_tx_t *tx)
1620 {
1621 dmu_buf_impl_t *db;
1622 objset_t *os = dn->dn_objset;
1623 uint64_t object = dn->dn_object;
1624 uint32_t blksz = (uint32_t)arc_buf_lsize(buf);
1625 uint64_t blkid;
1626
1627 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1628 blkid = dbuf_whichblock(dn, 0, offset);
1629 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1630 rw_exit(&dn->dn_struct_rwlock);
1631
1632 /*
1633 * We can only assign if the offset is aligned, the arc buf is the
1634 * same size as the dbuf, and the dbuf is not metadata.
1635 */
1636 if (offset == db->db.db_offset && blksz == db->db.db_size) {
1637 dbuf_assign_arcbuf(db, buf, tx);
1638 dbuf_rele(db, FTAG);
1639 } else {
1640 /* compressed bufs must always be assignable to their dbuf */
1641 ASSERT3U(arc_get_compression(buf), ==, ZIO_COMPRESS_OFF);
1642 ASSERT(!(buf->b_flags & ARC_BUF_FLAG_COMPRESSED));
1643
1644 dbuf_rele(db, FTAG);
1645 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1646 dmu_return_arcbuf(buf);
1647 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1648 }
1649 }
1650
1651 void
1652 dmu_assign_arcbuf_by_dbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1653 dmu_tx_t *tx)
1654 {
1655 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1656
1657 DB_DNODE_ENTER(dbuf);
1658 dmu_assign_arcbuf_by_dnode(DB_DNODE(dbuf), offset, buf, tx);
1659 DB_DNODE_EXIT(dbuf);
1660 }
1661
1662 typedef struct {
1663 dbuf_dirty_record_t *dsa_dr;
1664 dmu_sync_cb_t *dsa_done;
1665 zgd_t *dsa_zgd;
1666 dmu_tx_t *dsa_tx;
1667 } dmu_sync_arg_t;
1668
1669 /* ARGSUSED */
1670 static void
1671 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1672 {
1673 dmu_sync_arg_t *dsa = varg;
1674 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1675 blkptr_t *bp = zio->io_bp;
1676
1677 if (zio->io_error == 0) {
1678 if (BP_IS_HOLE(bp)) {
1679 /*
1680 * A block of zeros may compress to a hole, but the
1681 * block size still needs to be known for replay.
1682 */
1683 BP_SET_LSIZE(bp, db->db_size);
1684 } else if (!BP_IS_EMBEDDED(bp)) {
1685 ASSERT(BP_GET_LEVEL(bp) == 0);
1686 BP_SET_FILL(bp, 1);
1687 }
1688 }
1689 }
1690
1691 static void
1692 dmu_sync_late_arrival_ready(zio_t *zio)
1693 {
1694 dmu_sync_ready(zio, NULL, zio->io_private);
1695 }
1696
1697 /* ARGSUSED */
1698 static void
1699 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1700 {
1701 dmu_sync_arg_t *dsa = varg;
1702 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1703 dmu_buf_impl_t *db = dr->dr_dbuf;
1704
1705 mutex_enter(&db->db_mtx);
1706 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1707 if (zio->io_error == 0) {
1708 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1709 if (dr->dt.dl.dr_nopwrite) {
1710 blkptr_t *bp = zio->io_bp;
1711 blkptr_t *bp_orig = &zio->io_bp_orig;
1712 uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1713
1714 ASSERT(BP_EQUAL(bp, bp_orig));
1715 VERIFY(BP_EQUAL(bp, db->db_blkptr));
1716 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1717 VERIFY(zio_checksum_table[chksum].ci_flags &
1718 ZCHECKSUM_FLAG_NOPWRITE);
1719 }
1720 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1721 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1722 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1723
1724 /*
1725 * Old style holes are filled with all zeros, whereas
1726 * new-style holes maintain their lsize, type, level,
1727 * and birth time (see zio_write_compress). While we
1728 * need to reset the BP_SET_LSIZE() call that happened
1729 * in dmu_sync_ready for old style holes, we do *not*
1730 * want to wipe out the information contained in new
1731 * style holes. Thus, only zero out the block pointer if
1732 * it's an old style hole.
1733 */
1734 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) &&
1735 dr->dt.dl.dr_overridden_by.blk_birth == 0)
1736 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1737 } else {
1738 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1739 }
1740 cv_broadcast(&db->db_changed);
1741 mutex_exit(&db->db_mtx);
1742
1743 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1744
1745 kmem_free(dsa, sizeof (*dsa));
1746 }
1747
1748 static void
1749 dmu_sync_late_arrival_done(zio_t *zio)
1750 {
1751 blkptr_t *bp = zio->io_bp;
1752 dmu_sync_arg_t *dsa = zio->io_private;
1753 ASSERTV(blkptr_t *bp_orig = &zio->io_bp_orig);
1754
1755 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1756 ASSERT(!(zio->io_flags & ZIO_FLAG_NOPWRITE));
1757 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1758 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1759 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1760 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1761 }
1762
1763 dmu_tx_commit(dsa->dsa_tx);
1764
1765 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1766
1767 abd_put(zio->io_abd);
1768 kmem_free(dsa, sizeof (*dsa));
1769 }
1770
1771 static int
1772 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1773 zio_prop_t *zp, zbookmark_phys_t *zb)
1774 {
1775 dmu_sync_arg_t *dsa;
1776 dmu_tx_t *tx;
1777
1778 tx = dmu_tx_create(os);
1779 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1780 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1781 dmu_tx_abort(tx);
1782 /* Make zl_get_data do txg_waited_synced() */
1783 return (SET_ERROR(EIO));
1784 }
1785
1786 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1787 dsa->dsa_dr = NULL;
1788 dsa->dsa_done = done;
1789 dsa->dsa_zgd = zgd;
1790 dsa->dsa_tx = tx;
1791
1792 /*
1793 * Since we are currently syncing this txg, it's nontrivial to
1794 * determine what BP to nopwrite against, so we disable nopwrite.
1795 *
1796 * When syncing, the db_blkptr is initially the BP of the previous
1797 * txg. We can not nopwrite against it because it will be changed
1798 * (this is similar to the non-late-arrival case where the dbuf is
1799 * dirty in a future txg).
1800 *
1801 * Then dbuf_write_ready() sets bp_blkptr to the location we will write.
1802 * We can not nopwrite against it because although the BP will not
1803 * (typically) be changed, the data has not yet been persisted to this
1804 * location.
1805 *
1806 * Finally, when dbuf_write_done() is called, it is theoretically
1807 * possible to always nopwrite, because the data that was written in
1808 * this txg is the same data that we are trying to write. However we
1809 * would need to check that this dbuf is not dirty in any future
1810 * txg's (as we do in the normal dmu_sync() path). For simplicity, we
1811 * don't nopwrite in this case.
1812 */
1813 zp->zp_nopwrite = B_FALSE;
1814
1815 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1816 abd_get_from_buf(zgd->zgd_db->db_data, zgd->zgd_db->db_size),
1817 zgd->zgd_db->db_size, zgd->zgd_db->db_size, zp,
1818 dmu_sync_late_arrival_ready, NULL, NULL, dmu_sync_late_arrival_done,
1819 dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1820
1821 return (0);
1822 }
1823
1824 /*
1825 * Intent log support: sync the block associated with db to disk.
1826 * N.B. and XXX: the caller is responsible for making sure that the
1827 * data isn't changing while dmu_sync() is writing it.
1828 *
1829 * Return values:
1830 *
1831 * EEXIST: this txg has already been synced, so there's nothing to do.
1832 * The caller should not log the write.
1833 *
1834 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1835 * The caller should not log the write.
1836 *
1837 * EALREADY: this block is already in the process of being synced.
1838 * The caller should track its progress (somehow).
1839 *
1840 * EIO: could not do the I/O.
1841 * The caller should do a txg_wait_synced().
1842 *
1843 * 0: the I/O has been initiated.
1844 * The caller should log this blkptr in the done callback.
1845 * It is possible that the I/O will fail, in which case
1846 * the error will be reported to the done callback and
1847 * propagated to pio from zio_done().
1848 */
1849 int
1850 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1851 {
1852 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1853 objset_t *os = db->db_objset;
1854 dsl_dataset_t *ds = os->os_dsl_dataset;
1855 dbuf_dirty_record_t *dr;
1856 dmu_sync_arg_t *dsa;
1857 zbookmark_phys_t zb;
1858 zio_prop_t zp;
1859 dnode_t *dn;
1860
1861 ASSERT(pio != NULL);
1862 ASSERT(txg != 0);
1863
1864 /* dbuf is within the locked range */
1865 ASSERT3U(db->db.db_offset, >=, zgd->zgd_rl->r_off);
1866 ASSERT3U(db->db.db_offset + db->db.db_size, <=,
1867 zgd->zgd_rl->r_off + zgd->zgd_rl->r_len);
1868
1869 SET_BOOKMARK(&zb, ds->ds_object,
1870 db->db.db_object, db->db_level, db->db_blkid);
1871
1872 DB_DNODE_ENTER(db);
1873 dn = DB_DNODE(db);
1874 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1875 DB_DNODE_EXIT(db);
1876
1877 /*
1878 * If we're frozen (running ziltest), we always need to generate a bp.
1879 */
1880 if (txg > spa_freeze_txg(os->os_spa))
1881 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1882
1883 /*
1884 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1885 * and us. If we determine that this txg is not yet syncing,
1886 * but it begins to sync a moment later, that's OK because the
1887 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1888 */
1889 mutex_enter(&db->db_mtx);
1890
1891 if (txg <= spa_last_synced_txg(os->os_spa)) {
1892 /*
1893 * This txg has already synced. There's nothing to do.
1894 */
1895 mutex_exit(&db->db_mtx);
1896 return (SET_ERROR(EEXIST));
1897 }
1898
1899 if (txg <= spa_syncing_txg(os->os_spa)) {
1900 /*
1901 * This txg is currently syncing, so we can't mess with
1902 * the dirty record anymore; just write a new log block.
1903 */
1904 mutex_exit(&db->db_mtx);
1905 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1906 }
1907
1908 dr = db->db_last_dirty;
1909 while (dr && dr->dr_txg != txg)
1910 dr = dr->dr_next;
1911
1912 if (dr == NULL) {
1913 /*
1914 * There's no dr for this dbuf, so it must have been freed.
1915 * There's no need to log writes to freed blocks, so we're done.
1916 */
1917 mutex_exit(&db->db_mtx);
1918 return (SET_ERROR(ENOENT));
1919 }
1920
1921 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
1922
1923 if (db->db_blkptr != NULL) {
1924 /*
1925 * We need to fill in zgd_bp with the current blkptr so that
1926 * the nopwrite code can check if we're writing the same
1927 * data that's already on disk. We can only nopwrite if we
1928 * are sure that after making the copy, db_blkptr will not
1929 * change until our i/o completes. We ensure this by
1930 * holding the db_mtx, and only allowing nopwrite if the
1931 * block is not already dirty (see below). This is verified
1932 * by dmu_sync_done(), which VERIFYs that the db_blkptr has
1933 * not changed.
1934 */
1935 *zgd->zgd_bp = *db->db_blkptr;
1936 }
1937
1938 /*
1939 * Assume the on-disk data is X, the current syncing data (in
1940 * txg - 1) is Y, and the current in-memory data is Z (currently
1941 * in dmu_sync).
1942 *
1943 * We usually want to perform a nopwrite if X and Z are the
1944 * same. However, if Y is different (i.e. the BP is going to
1945 * change before this write takes effect), then a nopwrite will
1946 * be incorrect - we would override with X, which could have
1947 * been freed when Y was written.
1948 *
1949 * (Note that this is not a concern when we are nop-writing from
1950 * syncing context, because X and Y must be identical, because
1951 * all previous txgs have been synced.)
1952 *
1953 * Therefore, we disable nopwrite if the current BP could change
1954 * before this TXG. There are two ways it could change: by
1955 * being dirty (dr_next is non-NULL), or by being freed
1956 * (dnode_block_freed()). This behavior is verified by
1957 * zio_done(), which VERIFYs that the override BP is identical
1958 * to the on-disk BP.
1959 */
1960 DB_DNODE_ENTER(db);
1961 dn = DB_DNODE(db);
1962 if (dr->dr_next != NULL || dnode_block_freed(dn, db->db_blkid))
1963 zp.zp_nopwrite = B_FALSE;
1964 DB_DNODE_EXIT(db);
1965
1966 ASSERT(dr->dr_txg == txg);
1967 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1968 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1969 /*
1970 * We have already issued a sync write for this buffer,
1971 * or this buffer has already been synced. It could not
1972 * have been dirtied since, or we would have cleared the state.
1973 */
1974 mutex_exit(&db->db_mtx);
1975 return (SET_ERROR(EALREADY));
1976 }
1977
1978 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1979 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1980 mutex_exit(&db->db_mtx);
1981
1982 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1983 dsa->dsa_dr = dr;
1984 dsa->dsa_done = done;
1985 dsa->dsa_zgd = zgd;
1986 dsa->dsa_tx = NULL;
1987
1988 zio_nowait(arc_write(pio, os->os_spa, txg,
1989 zgd->zgd_bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
1990 &zp, dmu_sync_ready, NULL, NULL, dmu_sync_done, dsa,
1991 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb));
1992
1993 return (0);
1994 }
1995
1996 int
1997 dmu_object_set_nlevels(objset_t *os, uint64_t object, int nlevels, dmu_tx_t *tx)
1998 {
1999 dnode_t *dn;
2000 int err;
2001
2002 err = dnode_hold(os, object, FTAG, &dn);
2003 if (err)
2004 return (err);
2005 err = dnode_set_nlevels(dn, nlevels, tx);
2006 dnode_rele(dn, FTAG);
2007 return (err);
2008 }
2009
2010 int
2011 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
2012 dmu_tx_t *tx)
2013 {
2014 dnode_t *dn;
2015 int err;
2016
2017 err = dnode_hold(os, object, FTAG, &dn);
2018 if (err)
2019 return (err);
2020 err = dnode_set_blksz(dn, size, ibs, tx);
2021 dnode_rele(dn, FTAG);
2022 return (err);
2023 }
2024
2025 void
2026 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
2027 dmu_tx_t *tx)
2028 {
2029 dnode_t *dn;
2030
2031 /*
2032 * Send streams include each object's checksum function. This
2033 * check ensures that the receiving system can understand the
2034 * checksum function transmitted.
2035 */
2036 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS);
2037
2038 VERIFY0(dnode_hold(os, object, FTAG, &dn));
2039 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS);
2040 dn->dn_checksum = checksum;
2041 dnode_setdirty(dn, tx);
2042 dnode_rele(dn, FTAG);
2043 }
2044
2045 void
2046 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
2047 dmu_tx_t *tx)
2048 {
2049 dnode_t *dn;
2050
2051 /*
2052 * Send streams include each object's compression function. This
2053 * check ensures that the receiving system can understand the
2054 * compression function transmitted.
2055 */
2056 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS);
2057
2058 VERIFY0(dnode_hold(os, object, FTAG, &dn));
2059 dn->dn_compress = compress;
2060 dnode_setdirty(dn, tx);
2061 dnode_rele(dn, FTAG);
2062 }
2063
2064 /*
2065 * Dirty an object and set the dirty record's raw flag. This is used
2066 * when writing raw data to an object that will not effect the
2067 * encryption parameters, specifically during raw receives.
2068 */
2069 int
2070 dmu_object_dirty_raw(objset_t *os, uint64_t object, dmu_tx_t *tx)
2071 {
2072 dnode_t *dn;
2073 int err;
2074
2075 err = dnode_hold(os, object, FTAG, &dn);
2076 if (err)
2077 return (err);
2078 dmu_buf_will_change_crypt_params((dmu_buf_t *)dn->dn_dbuf, tx);
2079 dnode_rele(dn, FTAG);
2080 return (err);
2081 }
2082
2083 int zfs_mdcomp_disable = 0;
2084
2085 /*
2086 * When the "redundant_metadata" property is set to "most", only indirect
2087 * blocks of this level and higher will have an additional ditto block.
2088 */
2089 int zfs_redundant_metadata_most_ditto_level = 2;
2090
2091 void
2092 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
2093 {
2094 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
2095 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
2096 (wp & WP_SPILL));
2097 enum zio_checksum checksum = os->os_checksum;
2098 enum zio_compress compress = os->os_compress;
2099 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
2100 boolean_t dedup = B_FALSE;
2101 boolean_t nopwrite = B_FALSE;
2102 boolean_t dedup_verify = os->os_dedup_verify;
2103 boolean_t encrypt = B_FALSE;
2104 int copies = os->os_copies;
2105
2106 /*
2107 * We maintain different write policies for each of the following
2108 * types of data:
2109 * 1. metadata
2110 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
2111 * 3. all other level 0 blocks
2112 */
2113 if (ismd) {
2114 if (zfs_mdcomp_disable) {
2115 compress = ZIO_COMPRESS_EMPTY;
2116 } else {
2117 /*
2118 * XXX -- we should design a compression algorithm
2119 * that specializes in arrays of bps.
2120 */
2121 compress = zio_compress_select(os->os_spa,
2122 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON);
2123 }
2124
2125 /*
2126 * Metadata always gets checksummed. If the data
2127 * checksum is multi-bit correctable, and it's not a
2128 * ZBT-style checksum, then it's suitable for metadata
2129 * as well. Otherwise, the metadata checksum defaults
2130 * to fletcher4.
2131 */
2132 if (!(zio_checksum_table[checksum].ci_flags &
2133 ZCHECKSUM_FLAG_METADATA) ||
2134 (zio_checksum_table[checksum].ci_flags &
2135 ZCHECKSUM_FLAG_EMBEDDED))
2136 checksum = ZIO_CHECKSUM_FLETCHER_4;
2137
2138 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL ||
2139 (os->os_redundant_metadata ==
2140 ZFS_REDUNDANT_METADATA_MOST &&
2141 (level >= zfs_redundant_metadata_most_ditto_level ||
2142 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))))
2143 copies++;
2144 } else if (wp & WP_NOFILL) {
2145 ASSERT(level == 0);
2146
2147 /*
2148 * If we're writing preallocated blocks, we aren't actually
2149 * writing them so don't set any policy properties. These
2150 * blocks are currently only used by an external subsystem
2151 * outside of zfs (i.e. dump) and not written by the zio
2152 * pipeline.
2153 */
2154 compress = ZIO_COMPRESS_OFF;
2155 checksum = ZIO_CHECKSUM_OFF;
2156 } else {
2157 compress = zio_compress_select(os->os_spa, dn->dn_compress,
2158 compress);
2159
2160 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
2161 zio_checksum_select(dn->dn_checksum, checksum) :
2162 dedup_checksum;
2163
2164 /*
2165 * Determine dedup setting. If we are in dmu_sync(),
2166 * we won't actually dedup now because that's all
2167 * done in syncing context; but we do want to use the
2168 * dedup checkum. If the checksum is not strong
2169 * enough to ensure unique signatures, force
2170 * dedup_verify.
2171 */
2172 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
2173 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
2174 if (!(zio_checksum_table[checksum].ci_flags &
2175 ZCHECKSUM_FLAG_DEDUP))
2176 dedup_verify = B_TRUE;
2177 }
2178
2179 /*
2180 * Enable nopwrite if we have secure enough checksum
2181 * algorithm (see comment in zio_nop_write) and
2182 * compression is enabled. We don't enable nopwrite if
2183 * dedup is enabled as the two features are mutually
2184 * exclusive.
2185 */
2186 nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags &
2187 ZCHECKSUM_FLAG_NOPWRITE) &&
2188 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
2189 }
2190
2191 /*
2192 * All objects in an encrypted objset are protected from modification
2193 * via a MAC. Encrypted objects store their IV and salt in the last DVA
2194 * in the bp, so we cannot use all copies. Encrypted objects are also
2195 * not subject to nopwrite since writing the same data will still
2196 * result in a new ciphertext. Only encrypted blocks can be dedup'd
2197 * to avoid ambiguity in the dedup code since the DDT does not store
2198 * object types.
2199 */
2200 if (os->os_encrypted && (wp & WP_NOFILL) == 0) {
2201 encrypt = B_TRUE;
2202
2203 if (DMU_OT_IS_ENCRYPTED(type)) {
2204 copies = MIN(copies, SPA_DVAS_PER_BP - 1);
2205 nopwrite = B_FALSE;
2206 } else {
2207 dedup = B_FALSE;
2208 }
2209
2210 if (type == DMU_OT_DNODE || type == DMU_OT_OBJSET)
2211 compress = ZIO_COMPRESS_EMPTY;
2212 }
2213
2214 zp->zp_compress = compress;
2215 zp->zp_checksum = checksum;
2216 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
2217 zp->zp_level = level;
2218 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
2219 zp->zp_dedup = dedup;
2220 zp->zp_dedup_verify = dedup && dedup_verify;
2221 zp->zp_nopwrite = nopwrite;
2222 zp->zp_encrypt = encrypt;
2223 zp->zp_byteorder = ZFS_HOST_BYTEORDER;
2224 bzero(zp->zp_salt, ZIO_DATA_SALT_LEN);
2225 bzero(zp->zp_iv, ZIO_DATA_IV_LEN);
2226 bzero(zp->zp_mac, ZIO_DATA_MAC_LEN);
2227
2228 ASSERT3U(zp->zp_compress, !=, ZIO_COMPRESS_INHERIT);
2229 }
2230
2231 /*
2232 * This function is only called from zfs_holey_common() for zpl_llseek()
2233 * in order to determine the location of holes. In order to accurately
2234 * report holes all dirty data must be synced to disk. This causes extremely
2235 * poor performance when seeking for holes in a dirty file. As a compromise,
2236 * only provide hole data when the dnode is clean. When a dnode is dirty
2237 * report the dnode as having no holes which is always a safe thing to do.
2238 */
2239 int
2240 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
2241 {
2242 dnode_t *dn;
2243 int i, err;
2244 boolean_t clean = B_TRUE;
2245
2246 err = dnode_hold(os, object, FTAG, &dn);
2247 if (err)
2248 return (err);
2249
2250 /*
2251 * Check if dnode is dirty
2252 */
2253 if (dn->dn_dirtyctx != DN_UNDIRTIED) {
2254 for (i = 0; i < TXG_SIZE; i++) {
2255 if (!list_is_empty(&dn->dn_dirty_records[i])) {
2256 clean = B_FALSE;
2257 break;
2258 }
2259 }
2260 }
2261
2262 /*
2263 * If compatibility option is on, sync any current changes before
2264 * we go trundling through the block pointers.
2265 */
2266 if (!clean && zfs_dmu_offset_next_sync) {
2267 clean = B_TRUE;
2268 dnode_rele(dn, FTAG);
2269 txg_wait_synced(dmu_objset_pool(os), 0);
2270 err = dnode_hold(os, object, FTAG, &dn);
2271 if (err)
2272 return (err);
2273 }
2274
2275 if (clean)
2276 err = dnode_next_offset(dn,
2277 (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
2278 else
2279 err = SET_ERROR(EBUSY);
2280
2281 dnode_rele(dn, FTAG);
2282
2283 return (err);
2284 }
2285
2286 void
2287 __dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2288 {
2289 dnode_phys_t *dnp = dn->dn_phys;
2290
2291 doi->doi_data_block_size = dn->dn_datablksz;
2292 doi->doi_metadata_block_size = dn->dn_indblkshift ?
2293 1ULL << dn->dn_indblkshift : 0;
2294 doi->doi_type = dn->dn_type;
2295 doi->doi_bonus_type = dn->dn_bonustype;
2296 doi->doi_bonus_size = dn->dn_bonuslen;
2297 doi->doi_dnodesize = dn->dn_num_slots << DNODE_SHIFT;
2298 doi->doi_indirection = dn->dn_nlevels;
2299 doi->doi_checksum = dn->dn_checksum;
2300 doi->doi_compress = dn->dn_compress;
2301 doi->doi_nblkptr = dn->dn_nblkptr;
2302 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
2303 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
2304 doi->doi_fill_count = 0;
2305 for (int i = 0; i < dnp->dn_nblkptr; i++)
2306 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
2307 }
2308
2309 void
2310 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2311 {
2312 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2313 mutex_enter(&dn->dn_mtx);
2314
2315 __dmu_object_info_from_dnode(dn, doi);
2316
2317 mutex_exit(&dn->dn_mtx);
2318 rw_exit(&dn->dn_struct_rwlock);
2319 }
2320
2321 /*
2322 * Get information on a DMU object.
2323 * If doi is NULL, just indicates whether the object exists.
2324 */
2325 int
2326 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
2327 {
2328 dnode_t *dn;
2329 int err = dnode_hold(os, object, FTAG, &dn);
2330
2331 if (err)
2332 return (err);
2333
2334 if (doi != NULL)
2335 dmu_object_info_from_dnode(dn, doi);
2336
2337 dnode_rele(dn, FTAG);
2338 return (0);
2339 }
2340
2341 /*
2342 * As above, but faster; can be used when you have a held dbuf in hand.
2343 */
2344 void
2345 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
2346 {
2347 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2348
2349 DB_DNODE_ENTER(db);
2350 dmu_object_info_from_dnode(DB_DNODE(db), doi);
2351 DB_DNODE_EXIT(db);
2352 }
2353
2354 /*
2355 * Faster still when you only care about the size.
2356 * This is specifically optimized for zfs_getattr().
2357 */
2358 void
2359 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
2360 u_longlong_t *nblk512)
2361 {
2362 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2363 dnode_t *dn;
2364
2365 DB_DNODE_ENTER(db);
2366 dn = DB_DNODE(db);
2367
2368 *blksize = dn->dn_datablksz;
2369 /* add in number of slots used for the dnode itself */
2370 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
2371 SPA_MINBLOCKSHIFT) + dn->dn_num_slots;
2372 DB_DNODE_EXIT(db);
2373 }
2374
2375 void
2376 dmu_object_dnsize_from_db(dmu_buf_t *db_fake, int *dnsize)
2377 {
2378 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2379 dnode_t *dn;
2380
2381 DB_DNODE_ENTER(db);
2382 dn = DB_DNODE(db);
2383 *dnsize = dn->dn_num_slots << DNODE_SHIFT;
2384 DB_DNODE_EXIT(db);
2385 }
2386
2387 void
2388 byteswap_uint64_array(void *vbuf, size_t size)
2389 {
2390 uint64_t *buf = vbuf;
2391 size_t count = size >> 3;
2392 int i;
2393
2394 ASSERT((size & 7) == 0);
2395
2396 for (i = 0; i < count; i++)
2397 buf[i] = BSWAP_64(buf[i]);
2398 }
2399
2400 void
2401 byteswap_uint32_array(void *vbuf, size_t size)
2402 {
2403 uint32_t *buf = vbuf;
2404 size_t count = size >> 2;
2405 int i;
2406
2407 ASSERT((size & 3) == 0);
2408
2409 for (i = 0; i < count; i++)
2410 buf[i] = BSWAP_32(buf[i]);
2411 }
2412
2413 void
2414 byteswap_uint16_array(void *vbuf, size_t size)
2415 {
2416 uint16_t *buf = vbuf;
2417 size_t count = size >> 1;
2418 int i;
2419
2420 ASSERT((size & 1) == 0);
2421
2422 for (i = 0; i < count; i++)
2423 buf[i] = BSWAP_16(buf[i]);
2424 }
2425
2426 /* ARGSUSED */
2427 void
2428 byteswap_uint8_array(void *vbuf, size_t size)
2429 {
2430 }
2431
2432 void
2433 dmu_init(void)
2434 {
2435 abd_init();
2436 zfs_dbgmsg_init();
2437 sa_cache_init();
2438 xuio_stat_init();
2439 dmu_objset_init();
2440 dnode_init();
2441 zfetch_init();
2442 dmu_tx_init();
2443 l2arc_init();
2444 arc_init();
2445 dbuf_init();
2446 }
2447
2448 void
2449 dmu_fini(void)
2450 {
2451 arc_fini(); /* arc depends on l2arc, so arc must go first */
2452 l2arc_fini();
2453 dmu_tx_fini();
2454 zfetch_fini();
2455 dbuf_fini();
2456 dnode_fini();
2457 dmu_objset_fini();
2458 xuio_stat_fini();
2459 sa_cache_fini();
2460 zfs_dbgmsg_fini();
2461 abd_fini();
2462 }
2463
2464 #if defined(_KERNEL) && defined(HAVE_SPL)
2465 EXPORT_SYMBOL(dmu_bonus_hold);
2466 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus);
2467 EXPORT_SYMBOL(dmu_buf_rele_array);
2468 EXPORT_SYMBOL(dmu_prefetch);
2469 EXPORT_SYMBOL(dmu_free_range);
2470 EXPORT_SYMBOL(dmu_free_long_range);
2471 EXPORT_SYMBOL(dmu_free_long_range_raw);
2472 EXPORT_SYMBOL(dmu_free_long_object);
2473 EXPORT_SYMBOL(dmu_free_long_object_raw);
2474 EXPORT_SYMBOL(dmu_read);
2475 EXPORT_SYMBOL(dmu_read_by_dnode);
2476 EXPORT_SYMBOL(dmu_write);
2477 EXPORT_SYMBOL(dmu_write_by_dnode);
2478 EXPORT_SYMBOL(dmu_prealloc);
2479 EXPORT_SYMBOL(dmu_object_info);
2480 EXPORT_SYMBOL(dmu_object_info_from_dnode);
2481 EXPORT_SYMBOL(dmu_object_info_from_db);
2482 EXPORT_SYMBOL(dmu_object_size_from_db);
2483 EXPORT_SYMBOL(dmu_object_dnsize_from_db);
2484 EXPORT_SYMBOL(dmu_object_set_nlevels);
2485 EXPORT_SYMBOL(dmu_object_set_blocksize);
2486 EXPORT_SYMBOL(dmu_object_set_checksum);
2487 EXPORT_SYMBOL(dmu_object_set_compress);
2488 EXPORT_SYMBOL(dmu_write_policy);
2489 EXPORT_SYMBOL(dmu_sync);
2490 EXPORT_SYMBOL(dmu_request_arcbuf);
2491 EXPORT_SYMBOL(dmu_return_arcbuf);
2492 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dnode);
2493 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dbuf);
2494 EXPORT_SYMBOL(dmu_buf_hold);
2495 EXPORT_SYMBOL(dmu_ot);
2496
2497 /* BEGIN CSTYLED */
2498 module_param(zfs_mdcomp_disable, int, 0644);
2499 MODULE_PARM_DESC(zfs_mdcomp_disable, "Disable meta data compression");
2500
2501 module_param(zfs_nopwrite_enabled, int, 0644);
2502 MODULE_PARM_DESC(zfs_nopwrite_enabled, "Enable NOP writes");
2503
2504 module_param(zfs_per_txg_dirty_frees_percent, ulong, 0644);
2505 MODULE_PARM_DESC(zfs_per_txg_dirty_frees_percent,
2506 "percentage of dirtied blocks from frees in one TXG");
2507
2508 module_param(zfs_dmu_offset_next_sync, int, 0644);
2509 MODULE_PARM_DESC(zfs_dmu_offset_next_sync,
2510 "Enable forcing txg sync to find holes");
2511
2512 /* END CSTYLED */
2513
2514 #endif
ZFS On Linux mirror for PVE