]> 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
Fix cloning into already dirty dbufs.
[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 https://opensource.org/licenses/CDDL-1.0.
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, 2020 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 * Copyright (c) 2019 Datto Inc.
29 * Copyright (c) 2019, Klara Inc.
30 * Copyright (c) 2019, Allan Jude
31 * Copyright (c) 2022 Hewlett Packard Enterprise Development LP.
32 * Copyright (c) 2021, 2022 by Pawel Jakub Dawidek
33 */
34
35 #include <sys/dmu.h>
36 #include <sys/dmu_impl.h>
37 #include <sys/dmu_tx.h>
38 #include <sys/dbuf.h>
39 #include <sys/dnode.h>
40 #include <sys/zfs_context.h>
41 #include <sys/dmu_objset.h>
42 #include <sys/dmu_traverse.h>
43 #include <sys/dsl_dataset.h>
44 #include <sys/dsl_dir.h>
45 #include <sys/dsl_pool.h>
46 #include <sys/dsl_synctask.h>
47 #include <sys/dsl_prop.h>
48 #include <sys/dmu_zfetch.h>
49 #include <sys/zfs_ioctl.h>
50 #include <sys/zap.h>
51 #include <sys/zio_checksum.h>
52 #include <sys/zio_compress.h>
53 #include <sys/sa.h>
54 #include <sys/zfeature.h>
55 #include <sys/abd.h>
56 #include <sys/brt.h>
57 #include <sys/trace_zfs.h>
58 #include <sys/zfs_racct.h>
59 #include <sys/zfs_rlock.h>
60 #ifdef _KERNEL
61 #include <sys/vmsystm.h>
62 #include <sys/zfs_znode.h>
63 #endif
64
65 /*
66 * Enable/disable nopwrite feature.
67 */
68 static int zfs_nopwrite_enabled = 1;
69
70 /*
71 * Tunable to control percentage of dirtied L1 blocks from frees allowed into
72 * one TXG. After this threshold is crossed, additional dirty blocks from frees
73 * will wait until the next TXG.
74 * A value of zero will disable this throttle.
75 */
76 static uint_t zfs_per_txg_dirty_frees_percent = 30;
77
78 /*
79 * Enable/disable forcing txg sync when dirty checking for holes with lseek().
80 * By default this is enabled to ensure accurate hole reporting, it can result
81 * in a significant performance penalty for lseek(SEEK_HOLE) heavy workloads.
82 * Disabling this option will result in holes never being reported in dirty
83 * files which is always safe.
84 */
85 static int zfs_dmu_offset_next_sync = 1;
86
87 /*
88 * Limit the amount we can prefetch with one call to this amount. This
89 * helps to limit the amount of memory that can be used by prefetching.
90 * Larger objects should be prefetched a bit at a time.
91 */
92 uint_t dmu_prefetch_max = 8 * SPA_MAXBLOCKSIZE;
93
94 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
95 {DMU_BSWAP_UINT8, TRUE, FALSE, FALSE, "unallocated" },
96 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "object directory" },
97 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "object array" },
98 {DMU_BSWAP_UINT8, TRUE, FALSE, FALSE, "packed nvlist" },
99 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "packed nvlist size" },
100 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "bpobj" },
101 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "bpobj header" },
102 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "SPA space map header" },
103 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "SPA space map" },
104 {DMU_BSWAP_UINT64, TRUE, FALSE, TRUE, "ZIL intent log" },
105 {DMU_BSWAP_DNODE, TRUE, FALSE, TRUE, "DMU dnode" },
106 {DMU_BSWAP_OBJSET, TRUE, TRUE, FALSE, "DMU objset" },
107 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "DSL directory" },
108 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL directory child map"},
109 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL dataset snap map" },
110 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL props" },
111 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "DSL dataset" },
112 {DMU_BSWAP_ZNODE, TRUE, FALSE, FALSE, "ZFS znode" },
113 {DMU_BSWAP_OLDACL, TRUE, FALSE, TRUE, "ZFS V0 ACL" },
114 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "ZFS plain file" },
115 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS directory" },
116 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "ZFS master node" },
117 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS delete queue" },
118 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "zvol object" },
119 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "zvol prop" },
120 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "other uint8[]" },
121 {DMU_BSWAP_UINT64, FALSE, FALSE, TRUE, "other uint64[]" },
122 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "other ZAP" },
123 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "persistent error log" },
124 {DMU_BSWAP_UINT8, TRUE, FALSE, FALSE, "SPA history" },
125 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "SPA history offsets" },
126 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "Pool properties" },
127 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL permissions" },
128 {DMU_BSWAP_ACL, TRUE, FALSE, TRUE, "ZFS ACL" },
129 {DMU_BSWAP_UINT8, TRUE, FALSE, TRUE, "ZFS SYSACL" },
130 {DMU_BSWAP_UINT8, TRUE, FALSE, TRUE, "FUID table" },
131 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "FUID table size" },
132 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL dataset next clones"},
133 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "scan work queue" },
134 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS user/group/project used" },
135 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS user/group/project quota"},
136 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "snapshot refcount tags"},
137 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "DDT ZAP algorithm" },
138 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "DDT statistics" },
139 {DMU_BSWAP_UINT8, TRUE, FALSE, TRUE, "System attributes" },
140 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "SA master node" },
141 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "SA attr registration" },
142 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "SA attr layouts" },
143 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "scan translations" },
144 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "deduplicated block" },
145 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL deadlist map" },
146 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "DSL deadlist map hdr" },
147 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL dir clones" },
148 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "bpobj subobj" }
149 };
150
151 dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
152 { byteswap_uint8_array, "uint8" },
153 { byteswap_uint16_array, "uint16" },
154 { byteswap_uint32_array, "uint32" },
155 { byteswap_uint64_array, "uint64" },
156 { zap_byteswap, "zap" },
157 { dnode_buf_byteswap, "dnode" },
158 { dmu_objset_byteswap, "objset" },
159 { zfs_znode_byteswap, "znode" },
160 { zfs_oldacl_byteswap, "oldacl" },
161 { zfs_acl_byteswap, "acl" }
162 };
163
164 static int
165 dmu_buf_hold_noread_by_dnode(dnode_t *dn, uint64_t offset,
166 const void *tag, dmu_buf_t **dbp)
167 {
168 uint64_t blkid;
169 dmu_buf_impl_t *db;
170
171 rw_enter(&dn->dn_struct_rwlock, RW_READER);
172 blkid = dbuf_whichblock(dn, 0, offset);
173 db = dbuf_hold(dn, blkid, tag);
174 rw_exit(&dn->dn_struct_rwlock);
175
176 if (db == NULL) {
177 *dbp = NULL;
178 return (SET_ERROR(EIO));
179 }
180
181 *dbp = &db->db;
182 return (0);
183 }
184 int
185 dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset,
186 const void *tag, dmu_buf_t **dbp)
187 {
188 dnode_t *dn;
189 uint64_t blkid;
190 dmu_buf_impl_t *db;
191 int err;
192
193 err = dnode_hold(os, object, FTAG, &dn);
194 if (err)
195 return (err);
196 rw_enter(&dn->dn_struct_rwlock, RW_READER);
197 blkid = dbuf_whichblock(dn, 0, offset);
198 db = dbuf_hold(dn, blkid, tag);
199 rw_exit(&dn->dn_struct_rwlock);
200 dnode_rele(dn, FTAG);
201
202 if (db == NULL) {
203 *dbp = NULL;
204 return (SET_ERROR(EIO));
205 }
206
207 *dbp = &db->db;
208 return (err);
209 }
210
211 int
212 dmu_buf_hold_by_dnode(dnode_t *dn, uint64_t offset,
213 const void *tag, dmu_buf_t **dbp, int flags)
214 {
215 int err;
216 int db_flags = DB_RF_CANFAIL;
217
218 if (flags & DMU_READ_NO_PREFETCH)
219 db_flags |= DB_RF_NOPREFETCH;
220 if (flags & DMU_READ_NO_DECRYPT)
221 db_flags |= DB_RF_NO_DECRYPT;
222
223 err = dmu_buf_hold_noread_by_dnode(dn, offset, tag, dbp);
224 if (err == 0) {
225 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
226 err = dbuf_read(db, NULL, db_flags);
227 if (err != 0) {
228 dbuf_rele(db, tag);
229 *dbp = NULL;
230 }
231 }
232
233 return (err);
234 }
235
236 int
237 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
238 const void *tag, dmu_buf_t **dbp, int flags)
239 {
240 int err;
241 int db_flags = DB_RF_CANFAIL;
242
243 if (flags & DMU_READ_NO_PREFETCH)
244 db_flags |= DB_RF_NOPREFETCH;
245 if (flags & DMU_READ_NO_DECRYPT)
246 db_flags |= DB_RF_NO_DECRYPT;
247
248 err = dmu_buf_hold_noread(os, object, offset, tag, dbp);
249 if (err == 0) {
250 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
251 err = dbuf_read(db, NULL, db_flags);
252 if (err != 0) {
253 dbuf_rele(db, tag);
254 *dbp = NULL;
255 }
256 }
257
258 return (err);
259 }
260
261 int
262 dmu_bonus_max(void)
263 {
264 return (DN_OLD_MAX_BONUSLEN);
265 }
266
267 int
268 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
269 {
270 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
271 dnode_t *dn;
272 int error;
273
274 DB_DNODE_ENTER(db);
275 dn = DB_DNODE(db);
276
277 if (dn->dn_bonus != db) {
278 error = SET_ERROR(EINVAL);
279 } else if (newsize < 0 || newsize > db_fake->db_size) {
280 error = SET_ERROR(EINVAL);
281 } else {
282 dnode_setbonuslen(dn, newsize, tx);
283 error = 0;
284 }
285
286 DB_DNODE_EXIT(db);
287 return (error);
288 }
289
290 int
291 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
292 {
293 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
294 dnode_t *dn;
295 int error;
296
297 DB_DNODE_ENTER(db);
298 dn = DB_DNODE(db);
299
300 if (!DMU_OT_IS_VALID(type)) {
301 error = SET_ERROR(EINVAL);
302 } else if (dn->dn_bonus != db) {
303 error = SET_ERROR(EINVAL);
304 } else {
305 dnode_setbonus_type(dn, type, tx);
306 error = 0;
307 }
308
309 DB_DNODE_EXIT(db);
310 return (error);
311 }
312
313 dmu_object_type_t
314 dmu_get_bonustype(dmu_buf_t *db_fake)
315 {
316 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
317 dnode_t *dn;
318 dmu_object_type_t type;
319
320 DB_DNODE_ENTER(db);
321 dn = DB_DNODE(db);
322 type = dn->dn_bonustype;
323 DB_DNODE_EXIT(db);
324
325 return (type);
326 }
327
328 int
329 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
330 {
331 dnode_t *dn;
332 int error;
333
334 error = dnode_hold(os, object, FTAG, &dn);
335 dbuf_rm_spill(dn, tx);
336 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
337 dnode_rm_spill(dn, tx);
338 rw_exit(&dn->dn_struct_rwlock);
339 dnode_rele(dn, FTAG);
340 return (error);
341 }
342
343 /*
344 * Lookup and hold the bonus buffer for the provided dnode. If the dnode
345 * has not yet been allocated a new bonus dbuf a will be allocated.
346 * Returns ENOENT, EIO, or 0.
347 */
348 int dmu_bonus_hold_by_dnode(dnode_t *dn, const void *tag, dmu_buf_t **dbp,
349 uint32_t flags)
350 {
351 dmu_buf_impl_t *db;
352 int error;
353 uint32_t db_flags = DB_RF_MUST_SUCCEED;
354
355 if (flags & DMU_READ_NO_PREFETCH)
356 db_flags |= DB_RF_NOPREFETCH;
357 if (flags & DMU_READ_NO_DECRYPT)
358 db_flags |= DB_RF_NO_DECRYPT;
359
360 rw_enter(&dn->dn_struct_rwlock, RW_READER);
361 if (dn->dn_bonus == NULL) {
362 if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
363 rw_exit(&dn->dn_struct_rwlock);
364 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
365 }
366 if (dn->dn_bonus == NULL)
367 dbuf_create_bonus(dn);
368 }
369 db = dn->dn_bonus;
370
371 /* as long as the bonus buf is held, the dnode will be held */
372 if (zfs_refcount_add(&db->db_holds, tag) == 1) {
373 VERIFY(dnode_add_ref(dn, db));
374 atomic_inc_32(&dn->dn_dbufs_count);
375 }
376
377 /*
378 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
379 * hold and incrementing the dbuf count to ensure that dnode_move() sees
380 * a dnode hold for every dbuf.
381 */
382 rw_exit(&dn->dn_struct_rwlock);
383
384 error = dbuf_read(db, NULL, db_flags);
385 if (error) {
386 dnode_evict_bonus(dn);
387 dbuf_rele(db, tag);
388 *dbp = NULL;
389 return (error);
390 }
391
392 *dbp = &db->db;
393 return (0);
394 }
395
396 int
397 dmu_bonus_hold(objset_t *os, uint64_t object, const void *tag, dmu_buf_t **dbp)
398 {
399 dnode_t *dn;
400 int error;
401
402 error = dnode_hold(os, object, FTAG, &dn);
403 if (error)
404 return (error);
405
406 error = dmu_bonus_hold_by_dnode(dn, tag, dbp, DMU_READ_NO_PREFETCH);
407 dnode_rele(dn, FTAG);
408
409 return (error);
410 }
411
412 /*
413 * returns ENOENT, EIO, or 0.
414 *
415 * This interface will allocate a blank spill dbuf when a spill blk
416 * doesn't already exist on the dnode.
417 *
418 * if you only want to find an already existing spill db, then
419 * dmu_spill_hold_existing() should be used.
420 */
421 int
422 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, const void *tag,
423 dmu_buf_t **dbp)
424 {
425 dmu_buf_impl_t *db = NULL;
426 int err;
427
428 if ((flags & DB_RF_HAVESTRUCT) == 0)
429 rw_enter(&dn->dn_struct_rwlock, RW_READER);
430
431 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
432
433 if ((flags & DB_RF_HAVESTRUCT) == 0)
434 rw_exit(&dn->dn_struct_rwlock);
435
436 if (db == NULL) {
437 *dbp = NULL;
438 return (SET_ERROR(EIO));
439 }
440 err = dbuf_read(db, NULL, flags);
441 if (err == 0)
442 *dbp = &db->db;
443 else {
444 dbuf_rele(db, tag);
445 *dbp = NULL;
446 }
447 return (err);
448 }
449
450 int
451 dmu_spill_hold_existing(dmu_buf_t *bonus, const void *tag, dmu_buf_t **dbp)
452 {
453 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
454 dnode_t *dn;
455 int err;
456
457 DB_DNODE_ENTER(db);
458 dn = DB_DNODE(db);
459
460 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
461 err = SET_ERROR(EINVAL);
462 } else {
463 rw_enter(&dn->dn_struct_rwlock, RW_READER);
464
465 if (!dn->dn_have_spill) {
466 err = SET_ERROR(ENOENT);
467 } else {
468 err = dmu_spill_hold_by_dnode(dn,
469 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
470 }
471
472 rw_exit(&dn->dn_struct_rwlock);
473 }
474
475 DB_DNODE_EXIT(db);
476 return (err);
477 }
478
479 int
480 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, uint32_t flags, const void *tag,
481 dmu_buf_t **dbp)
482 {
483 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
484 dnode_t *dn;
485 int err;
486 uint32_t db_flags = DB_RF_CANFAIL;
487
488 if (flags & DMU_READ_NO_DECRYPT)
489 db_flags |= DB_RF_NO_DECRYPT;
490
491 DB_DNODE_ENTER(db);
492 dn = DB_DNODE(db);
493 err = dmu_spill_hold_by_dnode(dn, db_flags, tag, dbp);
494 DB_DNODE_EXIT(db);
495
496 return (err);
497 }
498
499 /*
500 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
501 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
502 * and can induce severe lock contention when writing to several files
503 * whose dnodes are in the same block.
504 */
505 int
506 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
507 boolean_t read, const void *tag, int *numbufsp, dmu_buf_t ***dbpp,
508 uint32_t flags)
509 {
510 dmu_buf_t **dbp;
511 zstream_t *zs = NULL;
512 uint64_t blkid, nblks, i;
513 uint32_t dbuf_flags;
514 int err;
515 zio_t *zio = NULL;
516 boolean_t missed = B_FALSE;
517
518 ASSERT(!read || length <= DMU_MAX_ACCESS);
519
520 /*
521 * Note: We directly notify the prefetch code of this read, so that
522 * we can tell it about the multi-block read. dbuf_read() only knows
523 * about the one block it is accessing.
524 */
525 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT |
526 DB_RF_NOPREFETCH;
527
528 if ((flags & DMU_READ_NO_DECRYPT) != 0)
529 dbuf_flags |= DB_RF_NO_DECRYPT;
530
531 rw_enter(&dn->dn_struct_rwlock, RW_READER);
532 if (dn->dn_datablkshift) {
533 int blkshift = dn->dn_datablkshift;
534 nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) -
535 P2ALIGN(offset, 1ULL << blkshift)) >> blkshift;
536 } else {
537 if (offset + length > dn->dn_datablksz) {
538 zfs_panic_recover("zfs: accessing past end of object "
539 "%llx/%llx (size=%u access=%llu+%llu)",
540 (longlong_t)dn->dn_objset->
541 os_dsl_dataset->ds_object,
542 (longlong_t)dn->dn_object, dn->dn_datablksz,
543 (longlong_t)offset, (longlong_t)length);
544 rw_exit(&dn->dn_struct_rwlock);
545 return (SET_ERROR(EIO));
546 }
547 nblks = 1;
548 }
549 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
550
551 if (read)
552 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL,
553 ZIO_FLAG_CANFAIL);
554 blkid = dbuf_whichblock(dn, 0, offset);
555 if ((flags & DMU_READ_NO_PREFETCH) == 0 &&
556 length <= zfetch_array_rd_sz) {
557 /*
558 * Prepare the zfetch before initiating the demand reads, so
559 * that if multiple threads block on same indirect block, we
560 * base predictions on the original less racy request order.
561 */
562 zs = dmu_zfetch_prepare(&dn->dn_zfetch, blkid, nblks, read,
563 B_TRUE);
564 }
565 for (i = 0; i < nblks; i++) {
566 dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag);
567 if (db == NULL) {
568 if (zs)
569 dmu_zfetch_run(zs, missed, B_TRUE);
570 rw_exit(&dn->dn_struct_rwlock);
571 dmu_buf_rele_array(dbp, nblks, tag);
572 if (read)
573 zio_nowait(zio);
574 return (SET_ERROR(EIO));
575 }
576
577 /*
578 * Initiate async demand data read.
579 * We check the db_state after calling dbuf_read() because
580 * (1) dbuf_read() may change the state to CACHED due to a
581 * hit in the ARC, and (2) on a cache miss, a child will
582 * have been added to "zio" but not yet completed, so the
583 * state will not yet be CACHED.
584 */
585 if (read) {
586 if (i == nblks - 1 && blkid + i < dn->dn_maxblkid &&
587 offset + length < db->db.db_offset +
588 db->db.db_size) {
589 if (offset <= db->db.db_offset)
590 dbuf_flags |= DB_RF_PARTIAL_FIRST;
591 else
592 dbuf_flags |= DB_RF_PARTIAL_MORE;
593 }
594 (void) dbuf_read(db, zio, dbuf_flags);
595 if (db->db_state != DB_CACHED)
596 missed = B_TRUE;
597 }
598 dbp[i] = &db->db;
599 }
600
601 if (!read)
602 zfs_racct_write(length, nblks);
603
604 if (zs)
605 dmu_zfetch_run(zs, missed, B_TRUE);
606 rw_exit(&dn->dn_struct_rwlock);
607
608 if (read) {
609 /* wait for async read i/o */
610 err = zio_wait(zio);
611 if (err) {
612 dmu_buf_rele_array(dbp, nblks, tag);
613 return (err);
614 }
615
616 /* wait for other io to complete */
617 for (i = 0; i < nblks; i++) {
618 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
619 mutex_enter(&db->db_mtx);
620 while (db->db_state == DB_READ ||
621 db->db_state == DB_FILL)
622 cv_wait(&db->db_changed, &db->db_mtx);
623 if (db->db_state == DB_UNCACHED)
624 err = SET_ERROR(EIO);
625 mutex_exit(&db->db_mtx);
626 if (err) {
627 dmu_buf_rele_array(dbp, nblks, tag);
628 return (err);
629 }
630 }
631 }
632
633 *numbufsp = nblks;
634 *dbpp = dbp;
635 return (0);
636 }
637
638 int
639 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
640 uint64_t length, int read, const void *tag, int *numbufsp,
641 dmu_buf_t ***dbpp)
642 {
643 dnode_t *dn;
644 int err;
645
646 err = dnode_hold(os, object, FTAG, &dn);
647 if (err)
648 return (err);
649
650 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
651 numbufsp, dbpp, DMU_READ_PREFETCH);
652
653 dnode_rele(dn, FTAG);
654
655 return (err);
656 }
657
658 int
659 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
660 uint64_t length, boolean_t read, const void *tag, int *numbufsp,
661 dmu_buf_t ***dbpp)
662 {
663 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
664 dnode_t *dn;
665 int err;
666
667 DB_DNODE_ENTER(db);
668 dn = DB_DNODE(db);
669 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
670 numbufsp, dbpp, DMU_READ_PREFETCH);
671 DB_DNODE_EXIT(db);
672
673 return (err);
674 }
675
676 void
677 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, const void *tag)
678 {
679 int i;
680 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
681
682 if (numbufs == 0)
683 return;
684
685 for (i = 0; i < numbufs; i++) {
686 if (dbp[i])
687 dbuf_rele(dbp[i], tag);
688 }
689
690 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
691 }
692
693 /*
694 * Issue prefetch i/os for the given blocks. If level is greater than 0, the
695 * indirect blocks prefetched will be those that point to the blocks containing
696 * the data starting at offset, and continuing to offset + len.
697 *
698 * Note that if the indirect blocks above the blocks being prefetched are not
699 * in cache, they will be asynchronously read in.
700 */
701 void
702 dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset,
703 uint64_t len, zio_priority_t pri)
704 {
705 dnode_t *dn;
706 uint64_t blkid;
707 int nblks, err;
708
709 if (len == 0) { /* they're interested in the bonus buffer */
710 dn = DMU_META_DNODE(os);
711
712 if (object == 0 || object >= DN_MAX_OBJECT)
713 return;
714
715 rw_enter(&dn->dn_struct_rwlock, RW_READER);
716 blkid = dbuf_whichblock(dn, level,
717 object * sizeof (dnode_phys_t));
718 dbuf_prefetch(dn, level, blkid, pri, 0);
719 rw_exit(&dn->dn_struct_rwlock);
720 return;
721 }
722
723 /*
724 * See comment before the definition of dmu_prefetch_max.
725 */
726 len = MIN(len, dmu_prefetch_max);
727
728 /*
729 * XXX - Note, if the dnode for the requested object is not
730 * already cached, we will do a *synchronous* read in the
731 * dnode_hold() call. The same is true for any indirects.
732 */
733 err = dnode_hold(os, object, FTAG, &dn);
734 if (err != 0)
735 return;
736
737 /*
738 * offset + len - 1 is the last byte we want to prefetch for, and offset
739 * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the
740 * last block we want to prefetch, and dbuf_whichblock(dn, level,
741 * offset) is the first. Then the number we need to prefetch is the
742 * last - first + 1.
743 */
744 rw_enter(&dn->dn_struct_rwlock, RW_READER);
745 if (level > 0 || dn->dn_datablkshift != 0) {
746 nblks = dbuf_whichblock(dn, level, offset + len - 1) -
747 dbuf_whichblock(dn, level, offset) + 1;
748 } else {
749 nblks = (offset < dn->dn_datablksz);
750 }
751
752 if (nblks != 0) {
753 blkid = dbuf_whichblock(dn, level, offset);
754 for (int i = 0; i < nblks; i++)
755 dbuf_prefetch(dn, level, blkid + i, pri, 0);
756 }
757 rw_exit(&dn->dn_struct_rwlock);
758
759 dnode_rele(dn, FTAG);
760 }
761
762 /*
763 * Get the next "chunk" of file data to free. We traverse the file from
764 * the end so that the file gets shorter over time (if we crashes in the
765 * middle, this will leave us in a better state). We find allocated file
766 * data by simply searching the allocated level 1 indirects.
767 *
768 * On input, *start should be the first offset that does not need to be
769 * freed (e.g. "offset + length"). On return, *start will be the first
770 * offset that should be freed and l1blks is set to the number of level 1
771 * indirect blocks found within the chunk.
772 */
773 static int
774 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum, uint64_t *l1blks)
775 {
776 uint64_t blks;
777 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
778 /* bytes of data covered by a level-1 indirect block */
779 uint64_t iblkrange = (uint64_t)dn->dn_datablksz *
780 EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
781
782 ASSERT3U(minimum, <=, *start);
783
784 /*
785 * Check if we can free the entire range assuming that all of the
786 * L1 blocks in this range have data. If we can, we use this
787 * worst case value as an estimate so we can avoid having to look
788 * at the object's actual data.
789 */
790 uint64_t total_l1blks =
791 (roundup(*start, iblkrange) - (minimum / iblkrange * iblkrange)) /
792 iblkrange;
793 if (total_l1blks <= maxblks) {
794 *l1blks = total_l1blks;
795 *start = minimum;
796 return (0);
797 }
798 ASSERT(ISP2(iblkrange));
799
800 for (blks = 0; *start > minimum && blks < maxblks; blks++) {
801 int err;
802
803 /*
804 * dnode_next_offset(BACKWARDS) will find an allocated L1
805 * indirect block at or before the input offset. We must
806 * decrement *start so that it is at the end of the region
807 * to search.
808 */
809 (*start)--;
810
811 err = dnode_next_offset(dn,
812 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
813
814 /* if there are no indirect blocks before start, we are done */
815 if (err == ESRCH) {
816 *start = minimum;
817 break;
818 } else if (err != 0) {
819 *l1blks = blks;
820 return (err);
821 }
822
823 /* set start to the beginning of this L1 indirect */
824 *start = P2ALIGN(*start, iblkrange);
825 }
826 if (*start < minimum)
827 *start = minimum;
828 *l1blks = blks;
829
830 return (0);
831 }
832
833 /*
834 * If this objset is of type OST_ZFS return true if vfs's unmounted flag is set,
835 * otherwise return false.
836 * Used below in dmu_free_long_range_impl() to enable abort when unmounting
837 */
838 static boolean_t
839 dmu_objset_zfs_unmounting(objset_t *os)
840 {
841 #ifdef _KERNEL
842 if (dmu_objset_type(os) == DMU_OST_ZFS)
843 return (zfs_get_vfs_flag_unmounted(os));
844 #else
845 (void) os;
846 #endif
847 return (B_FALSE);
848 }
849
850 static int
851 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
852 uint64_t length)
853 {
854 uint64_t object_size;
855 int err;
856 uint64_t dirty_frees_threshold;
857 dsl_pool_t *dp = dmu_objset_pool(os);
858
859 if (dn == NULL)
860 return (SET_ERROR(EINVAL));
861
862 object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
863 if (offset >= object_size)
864 return (0);
865
866 if (zfs_per_txg_dirty_frees_percent <= 100)
867 dirty_frees_threshold =
868 zfs_per_txg_dirty_frees_percent * zfs_dirty_data_max / 100;
869 else
870 dirty_frees_threshold = zfs_dirty_data_max / 20;
871
872 if (length == DMU_OBJECT_END || offset + length > object_size)
873 length = object_size - offset;
874
875 while (length != 0) {
876 uint64_t chunk_end, chunk_begin, chunk_len;
877 uint64_t l1blks;
878 dmu_tx_t *tx;
879
880 if (dmu_objset_zfs_unmounting(dn->dn_objset))
881 return (SET_ERROR(EINTR));
882
883 chunk_end = chunk_begin = offset + length;
884
885 /* move chunk_begin backwards to the beginning of this chunk */
886 err = get_next_chunk(dn, &chunk_begin, offset, &l1blks);
887 if (err)
888 return (err);
889 ASSERT3U(chunk_begin, >=, offset);
890 ASSERT3U(chunk_begin, <=, chunk_end);
891
892 chunk_len = chunk_end - chunk_begin;
893
894 tx = dmu_tx_create(os);
895 dmu_tx_hold_free(tx, dn->dn_object, chunk_begin, chunk_len);
896
897 /*
898 * Mark this transaction as typically resulting in a net
899 * reduction in space used.
900 */
901 dmu_tx_mark_netfree(tx);
902 err = dmu_tx_assign(tx, TXG_WAIT);
903 if (err) {
904 dmu_tx_abort(tx);
905 return (err);
906 }
907
908 uint64_t txg = dmu_tx_get_txg(tx);
909
910 mutex_enter(&dp->dp_lock);
911 uint64_t long_free_dirty =
912 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK];
913 mutex_exit(&dp->dp_lock);
914
915 /*
916 * To avoid filling up a TXG with just frees, wait for
917 * the next TXG to open before freeing more chunks if
918 * we have reached the threshold of frees.
919 */
920 if (dirty_frees_threshold != 0 &&
921 long_free_dirty >= dirty_frees_threshold) {
922 DMU_TX_STAT_BUMP(dmu_tx_dirty_frees_delay);
923 dmu_tx_commit(tx);
924 txg_wait_open(dp, 0, B_TRUE);
925 continue;
926 }
927
928 /*
929 * In order to prevent unnecessary write throttling, for each
930 * TXG, we track the cumulative size of L1 blocks being dirtied
931 * in dnode_free_range() below. We compare this number to a
932 * tunable threshold, past which we prevent new L1 dirty freeing
933 * blocks from being added into the open TXG. See
934 * dmu_free_long_range_impl() for details. The threshold
935 * prevents write throttle activation due to dirty freeing L1
936 * blocks taking up a large percentage of zfs_dirty_data_max.
937 */
938 mutex_enter(&dp->dp_lock);
939 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] +=
940 l1blks << dn->dn_indblkshift;
941 mutex_exit(&dp->dp_lock);
942 DTRACE_PROBE3(free__long__range,
943 uint64_t, long_free_dirty, uint64_t, chunk_len,
944 uint64_t, txg);
945 dnode_free_range(dn, chunk_begin, chunk_len, tx);
946
947 dmu_tx_commit(tx);
948
949 length -= chunk_len;
950 }
951 return (0);
952 }
953
954 int
955 dmu_free_long_range(objset_t *os, uint64_t object,
956 uint64_t offset, uint64_t length)
957 {
958 dnode_t *dn;
959 int err;
960
961 err = dnode_hold(os, object, FTAG, &dn);
962 if (err != 0)
963 return (err);
964 err = dmu_free_long_range_impl(os, dn, offset, length);
965
966 /*
967 * It is important to zero out the maxblkid when freeing the entire
968 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
969 * will take the fast path, and (b) dnode_reallocate() can verify
970 * that the entire file has been freed.
971 */
972 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
973 dn->dn_maxblkid = 0;
974
975 dnode_rele(dn, FTAG);
976 return (err);
977 }
978
979 int
980 dmu_free_long_object(objset_t *os, uint64_t object)
981 {
982 dmu_tx_t *tx;
983 int err;
984
985 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
986 if (err != 0)
987 return (err);
988
989 tx = dmu_tx_create(os);
990 dmu_tx_hold_bonus(tx, object);
991 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
992 dmu_tx_mark_netfree(tx);
993 err = dmu_tx_assign(tx, TXG_WAIT);
994 if (err == 0) {
995 err = dmu_object_free(os, object, tx);
996 dmu_tx_commit(tx);
997 } else {
998 dmu_tx_abort(tx);
999 }
1000
1001 return (err);
1002 }
1003
1004 int
1005 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
1006 uint64_t size, dmu_tx_t *tx)
1007 {
1008 dnode_t *dn;
1009 int err = dnode_hold(os, object, FTAG, &dn);
1010 if (err)
1011 return (err);
1012 ASSERT(offset < UINT64_MAX);
1013 ASSERT(size == DMU_OBJECT_END || size <= UINT64_MAX - offset);
1014 dnode_free_range(dn, offset, size, tx);
1015 dnode_rele(dn, FTAG);
1016 return (0);
1017 }
1018
1019 static int
1020 dmu_read_impl(dnode_t *dn, uint64_t offset, uint64_t size,
1021 void *buf, uint32_t flags)
1022 {
1023 dmu_buf_t **dbp;
1024 int numbufs, err = 0;
1025
1026 /*
1027 * Deal with odd block sizes, where there can't be data past the first
1028 * block. If we ever do the tail block optimization, we will need to
1029 * handle that here as well.
1030 */
1031 if (dn->dn_maxblkid == 0) {
1032 uint64_t newsz = offset > dn->dn_datablksz ? 0 :
1033 MIN(size, dn->dn_datablksz - offset);
1034 memset((char *)buf + newsz, 0, size - newsz);
1035 size = newsz;
1036 }
1037
1038 while (size > 0) {
1039 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
1040 int i;
1041
1042 /*
1043 * NB: we could do this block-at-a-time, but it's nice
1044 * to be reading in parallel.
1045 */
1046 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
1047 TRUE, FTAG, &numbufs, &dbp, flags);
1048 if (err)
1049 break;
1050
1051 for (i = 0; i < numbufs; i++) {
1052 uint64_t tocpy;
1053 int64_t bufoff;
1054 dmu_buf_t *db = dbp[i];
1055
1056 ASSERT(size > 0);
1057
1058 bufoff = offset - db->db_offset;
1059 tocpy = MIN(db->db_size - bufoff, size);
1060
1061 (void) memcpy(buf, (char *)db->db_data + bufoff, tocpy);
1062
1063 offset += tocpy;
1064 size -= tocpy;
1065 buf = (char *)buf + tocpy;
1066 }
1067 dmu_buf_rele_array(dbp, numbufs, FTAG);
1068 }
1069 return (err);
1070 }
1071
1072 int
1073 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1074 void *buf, uint32_t flags)
1075 {
1076 dnode_t *dn;
1077 int err;
1078
1079 err = dnode_hold(os, object, FTAG, &dn);
1080 if (err != 0)
1081 return (err);
1082
1083 err = dmu_read_impl(dn, offset, size, buf, flags);
1084 dnode_rele(dn, FTAG);
1085 return (err);
1086 }
1087
1088 int
1089 dmu_read_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, void *buf,
1090 uint32_t flags)
1091 {
1092 return (dmu_read_impl(dn, offset, size, buf, flags));
1093 }
1094
1095 static void
1096 dmu_write_impl(dmu_buf_t **dbp, int numbufs, uint64_t offset, uint64_t size,
1097 const void *buf, dmu_tx_t *tx)
1098 {
1099 int i;
1100
1101 for (i = 0; i < numbufs; i++) {
1102 uint64_t tocpy;
1103 int64_t bufoff;
1104 dmu_buf_t *db = dbp[i];
1105
1106 ASSERT(size > 0);
1107
1108 bufoff = offset - db->db_offset;
1109 tocpy = MIN(db->db_size - bufoff, size);
1110
1111 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1112
1113 if (tocpy == db->db_size)
1114 dmu_buf_will_fill(db, tx);
1115 else
1116 dmu_buf_will_dirty(db, tx);
1117
1118 (void) memcpy((char *)db->db_data + bufoff, buf, tocpy);
1119
1120 if (tocpy == db->db_size)
1121 dmu_buf_fill_done(db, tx);
1122
1123 offset += tocpy;
1124 size -= tocpy;
1125 buf = (char *)buf + tocpy;
1126 }
1127 }
1128
1129 void
1130 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1131 const void *buf, dmu_tx_t *tx)
1132 {
1133 dmu_buf_t **dbp;
1134 int numbufs;
1135
1136 if (size == 0)
1137 return;
1138
1139 VERIFY0(dmu_buf_hold_array(os, object, offset, size,
1140 FALSE, FTAG, &numbufs, &dbp));
1141 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1142 dmu_buf_rele_array(dbp, numbufs, FTAG);
1143 }
1144
1145 /*
1146 * Note: Lustre is an external consumer of this interface.
1147 */
1148 void
1149 dmu_write_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size,
1150 const void *buf, dmu_tx_t *tx)
1151 {
1152 dmu_buf_t **dbp;
1153 int numbufs;
1154
1155 if (size == 0)
1156 return;
1157
1158 VERIFY0(dmu_buf_hold_array_by_dnode(dn, offset, size,
1159 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH));
1160 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1161 dmu_buf_rele_array(dbp, numbufs, FTAG);
1162 }
1163
1164 void
1165 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1166 dmu_tx_t *tx)
1167 {
1168 dmu_buf_t **dbp;
1169 int numbufs, i;
1170
1171 if (size == 0)
1172 return;
1173
1174 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
1175 FALSE, FTAG, &numbufs, &dbp));
1176
1177 for (i = 0; i < numbufs; i++) {
1178 dmu_buf_t *db = dbp[i];
1179
1180 dmu_buf_will_not_fill(db, tx);
1181 }
1182 dmu_buf_rele_array(dbp, numbufs, FTAG);
1183 }
1184
1185 void
1186 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
1187 void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
1188 int compressed_size, int byteorder, dmu_tx_t *tx)
1189 {
1190 dmu_buf_t *db;
1191
1192 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
1193 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
1194 VERIFY0(dmu_buf_hold_noread(os, object, offset,
1195 FTAG, &db));
1196
1197 dmu_buf_write_embedded(db,
1198 data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
1199 uncompressed_size, compressed_size, byteorder, tx);
1200
1201 dmu_buf_rele(db, FTAG);
1202 }
1203
1204 void
1205 dmu_redact(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1206 dmu_tx_t *tx)
1207 {
1208 int numbufs, i;
1209 dmu_buf_t **dbp;
1210
1211 VERIFY0(dmu_buf_hold_array(os, object, offset, size, FALSE, FTAG,
1212 &numbufs, &dbp));
1213 for (i = 0; i < numbufs; i++)
1214 dmu_buf_redact(dbp[i], tx);
1215 dmu_buf_rele_array(dbp, numbufs, FTAG);
1216 }
1217
1218 #ifdef _KERNEL
1219 int
1220 dmu_read_uio_dnode(dnode_t *dn, zfs_uio_t *uio, uint64_t size)
1221 {
1222 dmu_buf_t **dbp;
1223 int numbufs, i, err;
1224
1225 /*
1226 * NB: we could do this block-at-a-time, but it's nice
1227 * to be reading in parallel.
1228 */
1229 err = dmu_buf_hold_array_by_dnode(dn, zfs_uio_offset(uio), size,
1230 TRUE, FTAG, &numbufs, &dbp, 0);
1231 if (err)
1232 return (err);
1233
1234 for (i = 0; i < numbufs; i++) {
1235 uint64_t tocpy;
1236 int64_t bufoff;
1237 dmu_buf_t *db = dbp[i];
1238
1239 ASSERT(size > 0);
1240
1241 bufoff = zfs_uio_offset(uio) - db->db_offset;
1242 tocpy = MIN(db->db_size - bufoff, size);
1243
1244 err = zfs_uio_fault_move((char *)db->db_data + bufoff, tocpy,
1245 UIO_READ, uio);
1246
1247 if (err)
1248 break;
1249
1250 size -= tocpy;
1251 }
1252 dmu_buf_rele_array(dbp, numbufs, FTAG);
1253
1254 return (err);
1255 }
1256
1257 /*
1258 * Read 'size' bytes into the uio buffer.
1259 * From object zdb->db_object.
1260 * Starting at zfs_uio_offset(uio).
1261 *
1262 * If the caller already has a dbuf in the target object
1263 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1264 * because we don't have to find the dnode_t for the object.
1265 */
1266 int
1267 dmu_read_uio_dbuf(dmu_buf_t *zdb, zfs_uio_t *uio, uint64_t size)
1268 {
1269 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1270 dnode_t *dn;
1271 int err;
1272
1273 if (size == 0)
1274 return (0);
1275
1276 DB_DNODE_ENTER(db);
1277 dn = DB_DNODE(db);
1278 err = dmu_read_uio_dnode(dn, uio, size);
1279 DB_DNODE_EXIT(db);
1280
1281 return (err);
1282 }
1283
1284 /*
1285 * Read 'size' bytes into the uio buffer.
1286 * From the specified object
1287 * Starting at offset zfs_uio_offset(uio).
1288 */
1289 int
1290 dmu_read_uio(objset_t *os, uint64_t object, zfs_uio_t *uio, uint64_t size)
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_read_uio_dnode(dn, uio, size);
1303
1304 dnode_rele(dn, FTAG);
1305
1306 return (err);
1307 }
1308
1309 int
1310 dmu_write_uio_dnode(dnode_t *dn, zfs_uio_t *uio, uint64_t size, dmu_tx_t *tx)
1311 {
1312 dmu_buf_t **dbp;
1313 int numbufs;
1314 int err = 0;
1315 int i;
1316
1317 err = dmu_buf_hold_array_by_dnode(dn, zfs_uio_offset(uio), size,
1318 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1319 if (err)
1320 return (err);
1321
1322 for (i = 0; i < numbufs; i++) {
1323 uint64_t tocpy;
1324 int64_t bufoff;
1325 dmu_buf_t *db = dbp[i];
1326
1327 ASSERT(size > 0);
1328
1329 bufoff = zfs_uio_offset(uio) - db->db_offset;
1330 tocpy = MIN(db->db_size - bufoff, size);
1331
1332 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1333
1334 if (tocpy == db->db_size)
1335 dmu_buf_will_fill(db, tx);
1336 else
1337 dmu_buf_will_dirty(db, tx);
1338
1339 /*
1340 * XXX zfs_uiomove could block forever (eg.nfs-backed
1341 * pages). There needs to be a uiolockdown() function
1342 * to lock the pages in memory, so that zfs_uiomove won't
1343 * block.
1344 */
1345 err = zfs_uio_fault_move((char *)db->db_data + bufoff,
1346 tocpy, UIO_WRITE, uio);
1347
1348 if (tocpy == db->db_size)
1349 dmu_buf_fill_done(db, tx);
1350
1351 if (err)
1352 break;
1353
1354 size -= tocpy;
1355 }
1356
1357 dmu_buf_rele_array(dbp, numbufs, FTAG);
1358 return (err);
1359 }
1360
1361 /*
1362 * Write 'size' bytes from the uio buffer.
1363 * To object zdb->db_object.
1364 * Starting at offset zfs_uio_offset(uio).
1365 *
1366 * If the caller already has a dbuf in the target object
1367 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1368 * because we don't have to find the dnode_t for the object.
1369 */
1370 int
1371 dmu_write_uio_dbuf(dmu_buf_t *zdb, zfs_uio_t *uio, uint64_t size,
1372 dmu_tx_t *tx)
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_write_uio_dnode(dn, uio, size, tx);
1384 DB_DNODE_EXIT(db);
1385
1386 return (err);
1387 }
1388
1389 /*
1390 * Write 'size' bytes from the uio buffer.
1391 * To the specified object.
1392 * Starting at offset zfs_uio_offset(uio).
1393 */
1394 int
1395 dmu_write_uio(objset_t *os, uint64_t object, zfs_uio_t *uio, uint64_t size,
1396 dmu_tx_t *tx)
1397 {
1398 dnode_t *dn;
1399 int err;
1400
1401 if (size == 0)
1402 return (0);
1403
1404 err = dnode_hold(os, object, FTAG, &dn);
1405 if (err)
1406 return (err);
1407
1408 err = dmu_write_uio_dnode(dn, uio, size, tx);
1409
1410 dnode_rele(dn, FTAG);
1411
1412 return (err);
1413 }
1414 #endif /* _KERNEL */
1415
1416 /*
1417 * Allocate a loaned anonymous arc buffer.
1418 */
1419 arc_buf_t *
1420 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1421 {
1422 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1423
1424 return (arc_loan_buf(db->db_objset->os_spa, B_FALSE, size));
1425 }
1426
1427 /*
1428 * Free a loaned arc buffer.
1429 */
1430 void
1431 dmu_return_arcbuf(arc_buf_t *buf)
1432 {
1433 arc_return_buf(buf, FTAG);
1434 arc_buf_destroy(buf, FTAG);
1435 }
1436
1437 /*
1438 * A "lightweight" write is faster than a regular write (e.g.
1439 * dmu_write_by_dnode() or dmu_assign_arcbuf_by_dnode()), because it avoids the
1440 * CPU cost of creating a dmu_buf_impl_t and arc_buf_[hdr_]_t. However, the
1441 * data can not be read or overwritten until the transaction's txg has been
1442 * synced. This makes it appropriate for workloads that are known to be
1443 * (temporarily) write-only, like "zfs receive".
1444 *
1445 * A single block is written, starting at the specified offset in bytes. If
1446 * the call is successful, it returns 0 and the provided abd has been
1447 * consumed (the caller should not free it).
1448 */
1449 int
1450 dmu_lightweight_write_by_dnode(dnode_t *dn, uint64_t offset, abd_t *abd,
1451 const zio_prop_t *zp, zio_flag_t flags, dmu_tx_t *tx)
1452 {
1453 dbuf_dirty_record_t *dr =
1454 dbuf_dirty_lightweight(dn, dbuf_whichblock(dn, 0, offset), tx);
1455 if (dr == NULL)
1456 return (SET_ERROR(EIO));
1457 dr->dt.dll.dr_abd = abd;
1458 dr->dt.dll.dr_props = *zp;
1459 dr->dt.dll.dr_flags = flags;
1460 return (0);
1461 }
1462
1463 /*
1464 * When possible directly assign passed loaned arc buffer to a dbuf.
1465 * If this is not possible copy the contents of passed arc buf via
1466 * dmu_write().
1467 */
1468 int
1469 dmu_assign_arcbuf_by_dnode(dnode_t *dn, uint64_t offset, arc_buf_t *buf,
1470 dmu_tx_t *tx)
1471 {
1472 dmu_buf_impl_t *db;
1473 objset_t *os = dn->dn_objset;
1474 uint64_t object = dn->dn_object;
1475 uint32_t blksz = (uint32_t)arc_buf_lsize(buf);
1476 uint64_t blkid;
1477
1478 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1479 blkid = dbuf_whichblock(dn, 0, offset);
1480 db = dbuf_hold(dn, blkid, FTAG);
1481 if (db == NULL)
1482 return (SET_ERROR(EIO));
1483 rw_exit(&dn->dn_struct_rwlock);
1484
1485 /*
1486 * We can only assign if the offset is aligned and the arc buf is the
1487 * same size as the dbuf.
1488 */
1489 if (offset == db->db.db_offset && blksz == db->db.db_size) {
1490 zfs_racct_write(blksz, 1);
1491 dbuf_assign_arcbuf(db, buf, tx);
1492 dbuf_rele(db, FTAG);
1493 } else {
1494 /* compressed bufs must always be assignable to their dbuf */
1495 ASSERT3U(arc_get_compression(buf), ==, ZIO_COMPRESS_OFF);
1496 ASSERT(!(buf->b_flags & ARC_BUF_FLAG_COMPRESSED));
1497
1498 dbuf_rele(db, FTAG);
1499 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1500 dmu_return_arcbuf(buf);
1501 }
1502
1503 return (0);
1504 }
1505
1506 int
1507 dmu_assign_arcbuf_by_dbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1508 dmu_tx_t *tx)
1509 {
1510 int err;
1511 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1512
1513 DB_DNODE_ENTER(dbuf);
1514 err = dmu_assign_arcbuf_by_dnode(DB_DNODE(dbuf), offset, buf, tx);
1515 DB_DNODE_EXIT(dbuf);
1516
1517 return (err);
1518 }
1519
1520 typedef struct {
1521 dbuf_dirty_record_t *dsa_dr;
1522 dmu_sync_cb_t *dsa_done;
1523 zgd_t *dsa_zgd;
1524 dmu_tx_t *dsa_tx;
1525 } dmu_sync_arg_t;
1526
1527 static void
1528 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1529 {
1530 (void) buf;
1531 dmu_sync_arg_t *dsa = varg;
1532 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1533 blkptr_t *bp = zio->io_bp;
1534
1535 if (zio->io_error == 0) {
1536 if (BP_IS_HOLE(bp)) {
1537 /*
1538 * A block of zeros may compress to a hole, but the
1539 * block size still needs to be known for replay.
1540 */
1541 BP_SET_LSIZE(bp, db->db_size);
1542 } else if (!BP_IS_EMBEDDED(bp)) {
1543 ASSERT(BP_GET_LEVEL(bp) == 0);
1544 BP_SET_FILL(bp, 1);
1545 }
1546 }
1547 }
1548
1549 static void
1550 dmu_sync_late_arrival_ready(zio_t *zio)
1551 {
1552 dmu_sync_ready(zio, NULL, zio->io_private);
1553 }
1554
1555 static void
1556 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1557 {
1558 (void) buf;
1559 dmu_sync_arg_t *dsa = varg;
1560 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1561 dmu_buf_impl_t *db = dr->dr_dbuf;
1562 zgd_t *zgd = dsa->dsa_zgd;
1563
1564 /*
1565 * Record the vdev(s) backing this blkptr so they can be flushed after
1566 * the writes for the lwb have completed.
1567 */
1568 if (zio->io_error == 0) {
1569 zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp);
1570 }
1571
1572 mutex_enter(&db->db_mtx);
1573 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1574 if (zio->io_error == 0) {
1575 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1576 if (dr->dt.dl.dr_nopwrite) {
1577 blkptr_t *bp = zio->io_bp;
1578 blkptr_t *bp_orig = &zio->io_bp_orig;
1579 uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1580
1581 ASSERT(BP_EQUAL(bp, bp_orig));
1582 VERIFY(BP_EQUAL(bp, db->db_blkptr));
1583 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1584 VERIFY(zio_checksum_table[chksum].ci_flags &
1585 ZCHECKSUM_FLAG_NOPWRITE);
1586 }
1587 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1588 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1589 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1590
1591 /*
1592 * Old style holes are filled with all zeros, whereas
1593 * new-style holes maintain their lsize, type, level,
1594 * and birth time (see zio_write_compress). While we
1595 * need to reset the BP_SET_LSIZE() call that happened
1596 * in dmu_sync_ready for old style holes, we do *not*
1597 * want to wipe out the information contained in new
1598 * style holes. Thus, only zero out the block pointer if
1599 * it's an old style hole.
1600 */
1601 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) &&
1602 dr->dt.dl.dr_overridden_by.blk_birth == 0)
1603 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1604 } else {
1605 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1606 }
1607 cv_broadcast(&db->db_changed);
1608 mutex_exit(&db->db_mtx);
1609
1610 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1611
1612 kmem_free(dsa, sizeof (*dsa));
1613 }
1614
1615 static void
1616 dmu_sync_late_arrival_done(zio_t *zio)
1617 {
1618 blkptr_t *bp = zio->io_bp;
1619 dmu_sync_arg_t *dsa = zio->io_private;
1620 zgd_t *zgd = dsa->dsa_zgd;
1621
1622 if (zio->io_error == 0) {
1623 /*
1624 * Record the vdev(s) backing this blkptr so they can be
1625 * flushed after the writes for the lwb have completed.
1626 */
1627 zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp);
1628
1629 if (!BP_IS_HOLE(bp)) {
1630 blkptr_t *bp_orig __maybe_unused = &zio->io_bp_orig;
1631 ASSERT(!(zio->io_flags & ZIO_FLAG_NOPWRITE));
1632 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1633 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1634 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1635 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1636 }
1637 }
1638
1639 dmu_tx_commit(dsa->dsa_tx);
1640
1641 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1642
1643 abd_free(zio->io_abd);
1644 kmem_free(dsa, sizeof (*dsa));
1645 }
1646
1647 static int
1648 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1649 zio_prop_t *zp, zbookmark_phys_t *zb)
1650 {
1651 dmu_sync_arg_t *dsa;
1652 dmu_tx_t *tx;
1653
1654 tx = dmu_tx_create(os);
1655 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1656 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1657 dmu_tx_abort(tx);
1658 /* Make zl_get_data do txg_waited_synced() */
1659 return (SET_ERROR(EIO));
1660 }
1661
1662 /*
1663 * In order to prevent the zgd's lwb from being free'd prior to
1664 * dmu_sync_late_arrival_done() being called, we have to ensure
1665 * the lwb's "max txg" takes this tx's txg into account.
1666 */
1667 zil_lwb_add_txg(zgd->zgd_lwb, dmu_tx_get_txg(tx));
1668
1669 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1670 dsa->dsa_dr = NULL;
1671 dsa->dsa_done = done;
1672 dsa->dsa_zgd = zgd;
1673 dsa->dsa_tx = tx;
1674
1675 /*
1676 * Since we are currently syncing this txg, it's nontrivial to
1677 * determine what BP to nopwrite against, so we disable nopwrite.
1678 *
1679 * When syncing, the db_blkptr is initially the BP of the previous
1680 * txg. We can not nopwrite against it because it will be changed
1681 * (this is similar to the non-late-arrival case where the dbuf is
1682 * dirty in a future txg).
1683 *
1684 * Then dbuf_write_ready() sets bp_blkptr to the location we will write.
1685 * We can not nopwrite against it because although the BP will not
1686 * (typically) be changed, the data has not yet been persisted to this
1687 * location.
1688 *
1689 * Finally, when dbuf_write_done() is called, it is theoretically
1690 * possible to always nopwrite, because the data that was written in
1691 * this txg is the same data that we are trying to write. However we
1692 * would need to check that this dbuf is not dirty in any future
1693 * txg's (as we do in the normal dmu_sync() path). For simplicity, we
1694 * don't nopwrite in this case.
1695 */
1696 zp->zp_nopwrite = B_FALSE;
1697
1698 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1699 abd_get_from_buf(zgd->zgd_db->db_data, zgd->zgd_db->db_size),
1700 zgd->zgd_db->db_size, zgd->zgd_db->db_size, zp,
1701 dmu_sync_late_arrival_ready, NULL, NULL, dmu_sync_late_arrival_done,
1702 dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1703
1704 return (0);
1705 }
1706
1707 /*
1708 * Intent log support: sync the block associated with db to disk.
1709 * N.B. and XXX: the caller is responsible for making sure that the
1710 * data isn't changing while dmu_sync() is writing it.
1711 *
1712 * Return values:
1713 *
1714 * EEXIST: this txg has already been synced, so there's nothing to do.
1715 * The caller should not log the write.
1716 *
1717 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1718 * The caller should not log the write.
1719 *
1720 * EALREADY: this block is already in the process of being synced.
1721 * The caller should track its progress (somehow).
1722 *
1723 * EIO: could not do the I/O.
1724 * The caller should do a txg_wait_synced().
1725 *
1726 * 0: the I/O has been initiated.
1727 * The caller should log this blkptr in the done callback.
1728 * It is possible that the I/O will fail, in which case
1729 * the error will be reported to the done callback and
1730 * propagated to pio from zio_done().
1731 */
1732 int
1733 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1734 {
1735 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1736 objset_t *os = db->db_objset;
1737 dsl_dataset_t *ds = os->os_dsl_dataset;
1738 dbuf_dirty_record_t *dr, *dr_next;
1739 dmu_sync_arg_t *dsa;
1740 zbookmark_phys_t zb;
1741 zio_prop_t zp;
1742 dnode_t *dn;
1743
1744 ASSERT(pio != NULL);
1745 ASSERT(txg != 0);
1746
1747 SET_BOOKMARK(&zb, ds->ds_object,
1748 db->db.db_object, db->db_level, db->db_blkid);
1749
1750 DB_DNODE_ENTER(db);
1751 dn = DB_DNODE(db);
1752 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1753 DB_DNODE_EXIT(db);
1754
1755 /*
1756 * If we're frozen (running ziltest), we always need to generate a bp.
1757 */
1758 if (txg > spa_freeze_txg(os->os_spa))
1759 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1760
1761 /*
1762 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1763 * and us. If we determine that this txg is not yet syncing,
1764 * but it begins to sync a moment later, that's OK because the
1765 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1766 */
1767 mutex_enter(&db->db_mtx);
1768
1769 if (txg <= spa_last_synced_txg(os->os_spa)) {
1770 /*
1771 * This txg has already synced. There's nothing to do.
1772 */
1773 mutex_exit(&db->db_mtx);
1774 return (SET_ERROR(EEXIST));
1775 }
1776
1777 if (txg <= spa_syncing_txg(os->os_spa)) {
1778 /*
1779 * This txg is currently syncing, so we can't mess with
1780 * the dirty record anymore; just write a new log block.
1781 */
1782 mutex_exit(&db->db_mtx);
1783 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1784 }
1785
1786 dr = dbuf_find_dirty_eq(db, txg);
1787
1788 if (dr == NULL) {
1789 /*
1790 * There's no dr for this dbuf, so it must have been freed.
1791 * There's no need to log writes to freed blocks, so we're done.
1792 */
1793 mutex_exit(&db->db_mtx);
1794 return (SET_ERROR(ENOENT));
1795 }
1796
1797 dr_next = list_next(&db->db_dirty_records, dr);
1798 ASSERT(dr_next == NULL || dr_next->dr_txg < txg);
1799
1800 if (db->db_blkptr != NULL) {
1801 /*
1802 * We need to fill in zgd_bp with the current blkptr so that
1803 * the nopwrite code can check if we're writing the same
1804 * data that's already on disk. We can only nopwrite if we
1805 * are sure that after making the copy, db_blkptr will not
1806 * change until our i/o completes. We ensure this by
1807 * holding the db_mtx, and only allowing nopwrite if the
1808 * block is not already dirty (see below). This is verified
1809 * by dmu_sync_done(), which VERIFYs that the db_blkptr has
1810 * not changed.
1811 */
1812 *zgd->zgd_bp = *db->db_blkptr;
1813 }
1814
1815 /*
1816 * Assume the on-disk data is X, the current syncing data (in
1817 * txg - 1) is Y, and the current in-memory data is Z (currently
1818 * in dmu_sync).
1819 *
1820 * We usually want to perform a nopwrite if X and Z are the
1821 * same. However, if Y is different (i.e. the BP is going to
1822 * change before this write takes effect), then a nopwrite will
1823 * be incorrect - we would override with X, which could have
1824 * been freed when Y was written.
1825 *
1826 * (Note that this is not a concern when we are nop-writing from
1827 * syncing context, because X and Y must be identical, because
1828 * all previous txgs have been synced.)
1829 *
1830 * Therefore, we disable nopwrite if the current BP could change
1831 * before this TXG. There are two ways it could change: by
1832 * being dirty (dr_next is non-NULL), or by being freed
1833 * (dnode_block_freed()). This behavior is verified by
1834 * zio_done(), which VERIFYs that the override BP is identical
1835 * to the on-disk BP.
1836 */
1837 DB_DNODE_ENTER(db);
1838 dn = DB_DNODE(db);
1839 if (dr_next != NULL || dnode_block_freed(dn, db->db_blkid))
1840 zp.zp_nopwrite = B_FALSE;
1841 DB_DNODE_EXIT(db);
1842
1843 ASSERT(dr->dr_txg == txg);
1844 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1845 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1846 /*
1847 * We have already issued a sync write for this buffer,
1848 * or this buffer has already been synced. It could not
1849 * have been dirtied since, or we would have cleared the state.
1850 */
1851 mutex_exit(&db->db_mtx);
1852 return (SET_ERROR(EALREADY));
1853 }
1854
1855 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1856 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1857 mutex_exit(&db->db_mtx);
1858
1859 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1860 dsa->dsa_dr = dr;
1861 dsa->dsa_done = done;
1862 dsa->dsa_zgd = zgd;
1863 dsa->dsa_tx = NULL;
1864
1865 zio_nowait(arc_write(pio, os->os_spa, txg, zgd->zgd_bp,
1866 dr->dt.dl.dr_data, !DBUF_IS_CACHEABLE(db), dbuf_is_l2cacheable(db),
1867 &zp, dmu_sync_ready, NULL, NULL, dmu_sync_done, dsa,
1868 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb));
1869
1870 return (0);
1871 }
1872
1873 int
1874 dmu_object_set_nlevels(objset_t *os, uint64_t object, int nlevels, dmu_tx_t *tx)
1875 {
1876 dnode_t *dn;
1877 int err;
1878
1879 err = dnode_hold(os, object, FTAG, &dn);
1880 if (err)
1881 return (err);
1882 err = dnode_set_nlevels(dn, nlevels, tx);
1883 dnode_rele(dn, FTAG);
1884 return (err);
1885 }
1886
1887 int
1888 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1889 dmu_tx_t *tx)
1890 {
1891 dnode_t *dn;
1892 int err;
1893
1894 err = dnode_hold(os, object, FTAG, &dn);
1895 if (err)
1896 return (err);
1897 err = dnode_set_blksz(dn, size, ibs, tx);
1898 dnode_rele(dn, FTAG);
1899 return (err);
1900 }
1901
1902 int
1903 dmu_object_set_maxblkid(objset_t *os, uint64_t object, uint64_t maxblkid,
1904 dmu_tx_t *tx)
1905 {
1906 dnode_t *dn;
1907 int err;
1908
1909 err = dnode_hold(os, object, FTAG, &dn);
1910 if (err)
1911 return (err);
1912 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1913 dnode_new_blkid(dn, maxblkid, tx, B_FALSE, B_TRUE);
1914 rw_exit(&dn->dn_struct_rwlock);
1915 dnode_rele(dn, FTAG);
1916 return (0);
1917 }
1918
1919 void
1920 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1921 dmu_tx_t *tx)
1922 {
1923 dnode_t *dn;
1924
1925 /*
1926 * Send streams include each object's checksum function. This
1927 * check ensures that the receiving system can understand the
1928 * checksum function transmitted.
1929 */
1930 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS);
1931
1932 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1933 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS);
1934 dn->dn_checksum = checksum;
1935 dnode_setdirty(dn, tx);
1936 dnode_rele(dn, FTAG);
1937 }
1938
1939 void
1940 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1941 dmu_tx_t *tx)
1942 {
1943 dnode_t *dn;
1944
1945 /*
1946 * Send streams include each object's compression function. This
1947 * check ensures that the receiving system can understand the
1948 * compression function transmitted.
1949 */
1950 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS);
1951
1952 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1953 dn->dn_compress = compress;
1954 dnode_setdirty(dn, tx);
1955 dnode_rele(dn, FTAG);
1956 }
1957
1958 /*
1959 * When the "redundant_metadata" property is set to "most", only indirect
1960 * blocks of this level and higher will have an additional ditto block.
1961 */
1962 static const int zfs_redundant_metadata_most_ditto_level = 2;
1963
1964 void
1965 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
1966 {
1967 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
1968 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
1969 (wp & WP_SPILL));
1970 enum zio_checksum checksum = os->os_checksum;
1971 enum zio_compress compress = os->os_compress;
1972 uint8_t complevel = os->os_complevel;
1973 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
1974 boolean_t dedup = B_FALSE;
1975 boolean_t nopwrite = B_FALSE;
1976 boolean_t dedup_verify = os->os_dedup_verify;
1977 boolean_t encrypt = B_FALSE;
1978 int copies = os->os_copies;
1979
1980 /*
1981 * We maintain different write policies for each of the following
1982 * types of data:
1983 * 1. metadata
1984 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1985 * 3. all other level 0 blocks
1986 */
1987 if (ismd) {
1988 /*
1989 * XXX -- we should design a compression algorithm
1990 * that specializes in arrays of bps.
1991 */
1992 compress = zio_compress_select(os->os_spa,
1993 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON);
1994
1995 /*
1996 * Metadata always gets checksummed. If the data
1997 * checksum is multi-bit correctable, and it's not a
1998 * ZBT-style checksum, then it's suitable for metadata
1999 * as well. Otherwise, the metadata checksum defaults
2000 * to fletcher4.
2001 */
2002 if (!(zio_checksum_table[checksum].ci_flags &
2003 ZCHECKSUM_FLAG_METADATA) ||
2004 (zio_checksum_table[checksum].ci_flags &
2005 ZCHECKSUM_FLAG_EMBEDDED))
2006 checksum = ZIO_CHECKSUM_FLETCHER_4;
2007
2008 switch (os->os_redundant_metadata) {
2009 case ZFS_REDUNDANT_METADATA_ALL:
2010 copies++;
2011 break;
2012 case ZFS_REDUNDANT_METADATA_MOST:
2013 if (level >= zfs_redundant_metadata_most_ditto_level ||
2014 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))
2015 copies++;
2016 break;
2017 case ZFS_REDUNDANT_METADATA_SOME:
2018 if (DMU_OT_IS_CRITICAL(type))
2019 copies++;
2020 break;
2021 case ZFS_REDUNDANT_METADATA_NONE:
2022 break;
2023 }
2024 } else if (wp & WP_NOFILL) {
2025 ASSERT(level == 0);
2026
2027 /*
2028 * If we're writing preallocated blocks, we aren't actually
2029 * writing them so don't set any policy properties. These
2030 * blocks are currently only used by an external subsystem
2031 * outside of zfs (i.e. dump) and not written by the zio
2032 * pipeline.
2033 */
2034 compress = ZIO_COMPRESS_OFF;
2035 checksum = ZIO_CHECKSUM_OFF;
2036 } else {
2037 compress = zio_compress_select(os->os_spa, dn->dn_compress,
2038 compress);
2039 complevel = zio_complevel_select(os->os_spa, compress,
2040 complevel, complevel);
2041
2042 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
2043 zio_checksum_select(dn->dn_checksum, checksum) :
2044 dedup_checksum;
2045
2046 /*
2047 * Determine dedup setting. If we are in dmu_sync(),
2048 * we won't actually dedup now because that's all
2049 * done in syncing context; but we do want to use the
2050 * dedup checksum. If the checksum is not strong
2051 * enough to ensure unique signatures, force
2052 * dedup_verify.
2053 */
2054 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
2055 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
2056 if (!(zio_checksum_table[checksum].ci_flags &
2057 ZCHECKSUM_FLAG_DEDUP))
2058 dedup_verify = B_TRUE;
2059 }
2060
2061 /*
2062 * Enable nopwrite if we have secure enough checksum
2063 * algorithm (see comment in zio_nop_write) and
2064 * compression is enabled. We don't enable nopwrite if
2065 * dedup is enabled as the two features are mutually
2066 * exclusive.
2067 */
2068 nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags &
2069 ZCHECKSUM_FLAG_NOPWRITE) &&
2070 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
2071 }
2072
2073 /*
2074 * All objects in an encrypted objset are protected from modification
2075 * via a MAC. Encrypted objects store their IV and salt in the last DVA
2076 * in the bp, so we cannot use all copies. Encrypted objects are also
2077 * not subject to nopwrite since writing the same data will still
2078 * result in a new ciphertext. Only encrypted blocks can be dedup'd
2079 * to avoid ambiguity in the dedup code since the DDT does not store
2080 * object types.
2081 */
2082 if (os->os_encrypted && (wp & WP_NOFILL) == 0) {
2083 encrypt = B_TRUE;
2084
2085 if (DMU_OT_IS_ENCRYPTED(type)) {
2086 copies = MIN(copies, SPA_DVAS_PER_BP - 1);
2087 nopwrite = B_FALSE;
2088 } else {
2089 dedup = B_FALSE;
2090 }
2091
2092 if (level <= 0 &&
2093 (type == DMU_OT_DNODE || type == DMU_OT_OBJSET)) {
2094 compress = ZIO_COMPRESS_EMPTY;
2095 }
2096 }
2097
2098 zp->zp_compress = compress;
2099 zp->zp_complevel = complevel;
2100 zp->zp_checksum = checksum;
2101 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
2102 zp->zp_level = level;
2103 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
2104 zp->zp_dedup = dedup;
2105 zp->zp_dedup_verify = dedup && dedup_verify;
2106 zp->zp_nopwrite = nopwrite;
2107 zp->zp_encrypt = encrypt;
2108 zp->zp_byteorder = ZFS_HOST_BYTEORDER;
2109 memset(zp->zp_salt, 0, ZIO_DATA_SALT_LEN);
2110 memset(zp->zp_iv, 0, ZIO_DATA_IV_LEN);
2111 memset(zp->zp_mac, 0, ZIO_DATA_MAC_LEN);
2112 zp->zp_zpl_smallblk = DMU_OT_IS_FILE(zp->zp_type) ?
2113 os->os_zpl_special_smallblock : 0;
2114
2115 ASSERT3U(zp->zp_compress, !=, ZIO_COMPRESS_INHERIT);
2116 }
2117
2118 /*
2119 * This function is only called from zfs_holey_common() for zpl_llseek()
2120 * in order to determine the location of holes. In order to accurately
2121 * report holes all dirty data must be synced to disk. This causes extremely
2122 * poor performance when seeking for holes in a dirty file. As a compromise,
2123 * only provide hole data when the dnode is clean. When a dnode is dirty
2124 * report the dnode as having no holes which is always a safe thing to do.
2125 */
2126 int
2127 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
2128 {
2129 dnode_t *dn;
2130 int err;
2131
2132 restart:
2133 err = dnode_hold(os, object, FTAG, &dn);
2134 if (err)
2135 return (err);
2136
2137 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2138
2139 if (dnode_is_dirty(dn)) {
2140 /*
2141 * If the zfs_dmu_offset_next_sync module option is enabled
2142 * then strict hole reporting has been requested. Dirty
2143 * dnodes must be synced to disk to accurately report all
2144 * holes. When disabled dirty dnodes are reported to not
2145 * have any holes which is always safe.
2146 *
2147 * When called by zfs_holey_common() the zp->z_rangelock
2148 * is held to prevent zfs_write() and mmap writeback from
2149 * re-dirtying the dnode after txg_wait_synced().
2150 */
2151 if (zfs_dmu_offset_next_sync) {
2152 rw_exit(&dn->dn_struct_rwlock);
2153 dnode_rele(dn, FTAG);
2154 txg_wait_synced(dmu_objset_pool(os), 0);
2155 goto restart;
2156 }
2157
2158 err = SET_ERROR(EBUSY);
2159 } else {
2160 err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK |
2161 (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
2162 }
2163
2164 rw_exit(&dn->dn_struct_rwlock);
2165 dnode_rele(dn, FTAG);
2166
2167 return (err);
2168 }
2169
2170 int
2171 dmu_read_l0_bps(objset_t *os, uint64_t object, uint64_t offset, uint64_t length,
2172 dmu_tx_t *tx, blkptr_t *bps, size_t *nbpsp)
2173 {
2174 dmu_buf_t **dbp, *dbuf;
2175 dmu_buf_impl_t *db;
2176 blkptr_t *bp;
2177 int error, numbufs;
2178
2179 error = dmu_buf_hold_array(os, object, offset, length, FALSE, FTAG,
2180 &numbufs, &dbp);
2181 if (error != 0) {
2182 if (error == ESRCH) {
2183 error = SET_ERROR(ENXIO);
2184 }
2185 return (error);
2186 }
2187
2188 ASSERT3U(numbufs, <=, *nbpsp);
2189
2190 for (int i = 0; i < numbufs; i++) {
2191 dbuf = dbp[i];
2192 db = (dmu_buf_impl_t *)dbuf;
2193
2194 mutex_enter(&db->db_mtx);
2195
2196 /*
2197 * If the block is not on the disk yet, it has no BP assigned.
2198 * There is not much we can do...
2199 */
2200 if (!list_is_empty(&db->db_dirty_records)) {
2201 dbuf_dirty_record_t *dr;
2202
2203 dr = list_head(&db->db_dirty_records);
2204 if (dr->dt.dl.dr_brtwrite) {
2205 /*
2206 * This is very special case where we clone a
2207 * block and in the same transaction group we
2208 * read its BP (most likely to clone the clone).
2209 */
2210 bp = &dr->dt.dl.dr_overridden_by;
2211 } else {
2212 /*
2213 * The block was modified in the same
2214 * transaction group.
2215 */
2216 mutex_exit(&db->db_mtx);
2217 error = SET_ERROR(EAGAIN);
2218 goto out;
2219 }
2220 } else {
2221 bp = db->db_blkptr;
2222 }
2223
2224 mutex_exit(&db->db_mtx);
2225
2226 if (bp == NULL) {
2227 /*
2228 * The block was created in this transaction group,
2229 * so it has no BP yet.
2230 */
2231 error = SET_ERROR(EAGAIN);
2232 goto out;
2233 }
2234 if (dmu_buf_is_dirty(dbuf, tx)) {
2235 error = SET_ERROR(EAGAIN);
2236 goto out;
2237 }
2238 /*
2239 * Make sure we clone only data blocks.
2240 */
2241 if (BP_IS_METADATA(bp) && !BP_IS_HOLE(bp)) {
2242 error = SET_ERROR(EINVAL);
2243 goto out;
2244 }
2245
2246 bps[i] = *bp;
2247 }
2248
2249 *nbpsp = numbufs;
2250 out:
2251 dmu_buf_rele_array(dbp, numbufs, FTAG);
2252
2253 return (error);
2254 }
2255
2256 void
2257 dmu_brt_clone(objset_t *os, uint64_t object, uint64_t offset, uint64_t length,
2258 dmu_tx_t *tx, const blkptr_t *bps, size_t nbps, boolean_t replay)
2259 {
2260 spa_t *spa;
2261 dmu_buf_t **dbp, *dbuf;
2262 dmu_buf_impl_t *db;
2263 struct dirty_leaf *dl;
2264 dbuf_dirty_record_t *dr;
2265 const blkptr_t *bp;
2266 int numbufs;
2267
2268 spa = os->os_spa;
2269
2270 VERIFY0(dmu_buf_hold_array(os, object, offset, length, FALSE, FTAG,
2271 &numbufs, &dbp));
2272 ASSERT3U(nbps, ==, numbufs);
2273
2274 for (int i = 0; i < numbufs; i++) {
2275 dbuf = dbp[i];
2276 db = (dmu_buf_impl_t *)dbuf;
2277 bp = &bps[i];
2278
2279 ASSERT0(db->db_level);
2280 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2281 ASSERT(BP_IS_HOLE(bp) || dbuf->db_size == BP_GET_LSIZE(bp));
2282
2283 mutex_enter(&db->db_mtx);
2284
2285 VERIFY(!dbuf_undirty(db, tx));
2286 ASSERT(list_head(&db->db_dirty_records) == NULL);
2287 if (db->db_buf != NULL) {
2288 arc_buf_destroy(db->db_buf, db);
2289 db->db_buf = NULL;
2290 }
2291
2292 mutex_exit(&db->db_mtx);
2293
2294 dmu_buf_will_not_fill(dbuf, tx);
2295
2296 mutex_enter(&db->db_mtx);
2297
2298 dr = list_head(&db->db_dirty_records);
2299 VERIFY(dr != NULL);
2300 ASSERT3U(dr->dr_txg, ==, tx->tx_txg);
2301 dl = &dr->dt.dl;
2302 dl->dr_overridden_by = *bp;
2303 dl->dr_brtwrite = B_TRUE;
2304
2305 dl->dr_override_state = DR_OVERRIDDEN;
2306 if (BP_IS_HOLE(bp)) {
2307 dl->dr_overridden_by.blk_birth = 0;
2308 dl->dr_overridden_by.blk_phys_birth = 0;
2309 } else {
2310 dl->dr_overridden_by.blk_birth = dr->dr_txg;
2311 dl->dr_overridden_by.blk_phys_birth =
2312 BP_PHYSICAL_BIRTH(bp);
2313 }
2314
2315 mutex_exit(&db->db_mtx);
2316
2317 /*
2318 * When data in embedded into BP there is no need to create
2319 * BRT entry as there is no data block. Just copy the BP as
2320 * it contains the data.
2321 * Also, when replaying ZIL we don't want to bump references
2322 * in the BRT as it was already done during ZIL claim.
2323 */
2324 if (!replay && !BP_IS_HOLE(bp) && !BP_IS_EMBEDDED(bp)) {
2325 brt_pending_add(spa, bp, tx);
2326 }
2327 }
2328
2329 dmu_buf_rele_array(dbp, numbufs, FTAG);
2330 }
2331
2332 void
2333 __dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2334 {
2335 dnode_phys_t *dnp = dn->dn_phys;
2336
2337 doi->doi_data_block_size = dn->dn_datablksz;
2338 doi->doi_metadata_block_size = dn->dn_indblkshift ?
2339 1ULL << dn->dn_indblkshift : 0;
2340 doi->doi_type = dn->dn_type;
2341 doi->doi_bonus_type = dn->dn_bonustype;
2342 doi->doi_bonus_size = dn->dn_bonuslen;
2343 doi->doi_dnodesize = dn->dn_num_slots << DNODE_SHIFT;
2344 doi->doi_indirection = dn->dn_nlevels;
2345 doi->doi_checksum = dn->dn_checksum;
2346 doi->doi_compress = dn->dn_compress;
2347 doi->doi_nblkptr = dn->dn_nblkptr;
2348 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
2349 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
2350 doi->doi_fill_count = 0;
2351 for (int i = 0; i < dnp->dn_nblkptr; i++)
2352 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
2353 }
2354
2355 void
2356 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2357 {
2358 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2359 mutex_enter(&dn->dn_mtx);
2360
2361 __dmu_object_info_from_dnode(dn, doi);
2362
2363 mutex_exit(&dn->dn_mtx);
2364 rw_exit(&dn->dn_struct_rwlock);
2365 }
2366
2367 /*
2368 * Get information on a DMU object.
2369 * If doi is NULL, just indicates whether the object exists.
2370 */
2371 int
2372 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
2373 {
2374 dnode_t *dn;
2375 int err = dnode_hold(os, object, FTAG, &dn);
2376
2377 if (err)
2378 return (err);
2379
2380 if (doi != NULL)
2381 dmu_object_info_from_dnode(dn, doi);
2382
2383 dnode_rele(dn, FTAG);
2384 return (0);
2385 }
2386
2387 /*
2388 * As above, but faster; can be used when you have a held dbuf in hand.
2389 */
2390 void
2391 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
2392 {
2393 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2394
2395 DB_DNODE_ENTER(db);
2396 dmu_object_info_from_dnode(DB_DNODE(db), doi);
2397 DB_DNODE_EXIT(db);
2398 }
2399
2400 /*
2401 * Faster still when you only care about the size.
2402 */
2403 void
2404 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
2405 u_longlong_t *nblk512)
2406 {
2407 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2408 dnode_t *dn;
2409
2410 DB_DNODE_ENTER(db);
2411 dn = DB_DNODE(db);
2412
2413 *blksize = dn->dn_datablksz;
2414 /* add in number of slots used for the dnode itself */
2415 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
2416 SPA_MINBLOCKSHIFT) + dn->dn_num_slots;
2417 DB_DNODE_EXIT(db);
2418 }
2419
2420 void
2421 dmu_object_dnsize_from_db(dmu_buf_t *db_fake, int *dnsize)
2422 {
2423 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2424 dnode_t *dn;
2425
2426 DB_DNODE_ENTER(db);
2427 dn = DB_DNODE(db);
2428 *dnsize = dn->dn_num_slots << DNODE_SHIFT;
2429 DB_DNODE_EXIT(db);
2430 }
2431
2432 void
2433 byteswap_uint64_array(void *vbuf, size_t size)
2434 {
2435 uint64_t *buf = vbuf;
2436 size_t count = size >> 3;
2437 int i;
2438
2439 ASSERT((size & 7) == 0);
2440
2441 for (i = 0; i < count; i++)
2442 buf[i] = BSWAP_64(buf[i]);
2443 }
2444
2445 void
2446 byteswap_uint32_array(void *vbuf, size_t size)
2447 {
2448 uint32_t *buf = vbuf;
2449 size_t count = size >> 2;
2450 int i;
2451
2452 ASSERT((size & 3) == 0);
2453
2454 for (i = 0; i < count; i++)
2455 buf[i] = BSWAP_32(buf[i]);
2456 }
2457
2458 void
2459 byteswap_uint16_array(void *vbuf, size_t size)
2460 {
2461 uint16_t *buf = vbuf;
2462 size_t count = size >> 1;
2463 int i;
2464
2465 ASSERT((size & 1) == 0);
2466
2467 for (i = 0; i < count; i++)
2468 buf[i] = BSWAP_16(buf[i]);
2469 }
2470
2471 void
2472 byteswap_uint8_array(void *vbuf, size_t size)
2473 {
2474 (void) vbuf, (void) size;
2475 }
2476
2477 void
2478 dmu_init(void)
2479 {
2480 abd_init();
2481 zfs_dbgmsg_init();
2482 sa_cache_init();
2483 dmu_objset_init();
2484 dnode_init();
2485 zfetch_init();
2486 dmu_tx_init();
2487 l2arc_init();
2488 arc_init();
2489 dbuf_init();
2490 }
2491
2492 void
2493 dmu_fini(void)
2494 {
2495 arc_fini(); /* arc depends on l2arc, so arc must go first */
2496 l2arc_fini();
2497 dmu_tx_fini();
2498 zfetch_fini();
2499 dbuf_fini();
2500 dnode_fini();
2501 dmu_objset_fini();
2502 sa_cache_fini();
2503 zfs_dbgmsg_fini();
2504 abd_fini();
2505 }
2506
2507 EXPORT_SYMBOL(dmu_bonus_hold);
2508 EXPORT_SYMBOL(dmu_bonus_hold_by_dnode);
2509 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus);
2510 EXPORT_SYMBOL(dmu_buf_rele_array);
2511 EXPORT_SYMBOL(dmu_prefetch);
2512 EXPORT_SYMBOL(dmu_free_range);
2513 EXPORT_SYMBOL(dmu_free_long_range);
2514 EXPORT_SYMBOL(dmu_free_long_object);
2515 EXPORT_SYMBOL(dmu_read);
2516 EXPORT_SYMBOL(dmu_read_by_dnode);
2517 EXPORT_SYMBOL(dmu_write);
2518 EXPORT_SYMBOL(dmu_write_by_dnode);
2519 EXPORT_SYMBOL(dmu_prealloc);
2520 EXPORT_SYMBOL(dmu_object_info);
2521 EXPORT_SYMBOL(dmu_object_info_from_dnode);
2522 EXPORT_SYMBOL(dmu_object_info_from_db);
2523 EXPORT_SYMBOL(dmu_object_size_from_db);
2524 EXPORT_SYMBOL(dmu_object_dnsize_from_db);
2525 EXPORT_SYMBOL(dmu_object_set_nlevels);
2526 EXPORT_SYMBOL(dmu_object_set_blocksize);
2527 EXPORT_SYMBOL(dmu_object_set_maxblkid);
2528 EXPORT_SYMBOL(dmu_object_set_checksum);
2529 EXPORT_SYMBOL(dmu_object_set_compress);
2530 EXPORT_SYMBOL(dmu_offset_next);
2531 EXPORT_SYMBOL(dmu_write_policy);
2532 EXPORT_SYMBOL(dmu_sync);
2533 EXPORT_SYMBOL(dmu_request_arcbuf);
2534 EXPORT_SYMBOL(dmu_return_arcbuf);
2535 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dnode);
2536 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dbuf);
2537 EXPORT_SYMBOL(dmu_buf_hold);
2538 EXPORT_SYMBOL(dmu_ot);
2539
2540 ZFS_MODULE_PARAM(zfs, zfs_, nopwrite_enabled, INT, ZMOD_RW,
2541 "Enable NOP writes");
2542
2543 ZFS_MODULE_PARAM(zfs, zfs_, per_txg_dirty_frees_percent, UINT, ZMOD_RW,
2544 "Percentage of dirtied blocks from frees in one TXG");
2545
2546 ZFS_MODULE_PARAM(zfs, zfs_, dmu_offset_next_sync, INT, ZMOD_RW,
2547 "Enable forcing txg sync to find holes");
2548
2549 /* CSTYLED */
2550 ZFS_MODULE_PARAM(zfs, , dmu_prefetch_max, UINT, ZMOD_RW,
2551 "Limit one prefetch call to this size");
ZFS On Linux mirror for PVE