]> 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
FreeBSD: Add zfs_link_create() error handling
[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 #ifdef _ILP32
93 uint_t dmu_prefetch_max = 8 * 1024 * 1024;
94 #else
95 uint_t dmu_prefetch_max = 8 * SPA_MAXBLOCKSIZE;
96 #endif
97
98 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
99 {DMU_BSWAP_UINT8, TRUE, FALSE, FALSE, "unallocated" },
100 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "object directory" },
101 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "object array" },
102 {DMU_BSWAP_UINT8, TRUE, FALSE, FALSE, "packed nvlist" },
103 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "packed nvlist size" },
104 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "bpobj" },
105 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "bpobj header" },
106 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "SPA space map header" },
107 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "SPA space map" },
108 {DMU_BSWAP_UINT64, TRUE, FALSE, TRUE, "ZIL intent log" },
109 {DMU_BSWAP_DNODE, TRUE, FALSE, TRUE, "DMU dnode" },
110 {DMU_BSWAP_OBJSET, TRUE, TRUE, FALSE, "DMU objset" },
111 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "DSL directory" },
112 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL directory child map"},
113 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL dataset snap map" },
114 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL props" },
115 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "DSL dataset" },
116 {DMU_BSWAP_ZNODE, TRUE, FALSE, FALSE, "ZFS znode" },
117 {DMU_BSWAP_OLDACL, TRUE, FALSE, TRUE, "ZFS V0 ACL" },
118 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "ZFS plain file" },
119 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS directory" },
120 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "ZFS master node" },
121 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS delete queue" },
122 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "zvol object" },
123 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "zvol prop" },
124 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "other uint8[]" },
125 {DMU_BSWAP_UINT64, FALSE, FALSE, TRUE, "other uint64[]" },
126 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "other ZAP" },
127 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "persistent error log" },
128 {DMU_BSWAP_UINT8, TRUE, FALSE, FALSE, "SPA history" },
129 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "SPA history offsets" },
130 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "Pool properties" },
131 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL permissions" },
132 {DMU_BSWAP_ACL, TRUE, FALSE, TRUE, "ZFS ACL" },
133 {DMU_BSWAP_UINT8, TRUE, FALSE, TRUE, "ZFS SYSACL" },
134 {DMU_BSWAP_UINT8, TRUE, FALSE, TRUE, "FUID table" },
135 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "FUID table size" },
136 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL dataset next clones"},
137 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "scan work queue" },
138 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS user/group/project used" },
139 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS user/group/project quota"},
140 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "snapshot refcount tags"},
141 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "DDT ZAP algorithm" },
142 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "DDT statistics" },
143 {DMU_BSWAP_UINT8, TRUE, FALSE, TRUE, "System attributes" },
144 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "SA master node" },
145 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "SA attr registration" },
146 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "SA attr layouts" },
147 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "scan translations" },
148 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "deduplicated block" },
149 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL deadlist map" },
150 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "DSL deadlist map hdr" },
151 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL dir clones" },
152 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "bpobj subobj" }
153 };
154
155 dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
156 { byteswap_uint8_array, "uint8" },
157 { byteswap_uint16_array, "uint16" },
158 { byteswap_uint32_array, "uint32" },
159 { byteswap_uint64_array, "uint64" },
160 { zap_byteswap, "zap" },
161 { dnode_buf_byteswap, "dnode" },
162 { dmu_objset_byteswap, "objset" },
163 { zfs_znode_byteswap, "znode" },
164 { zfs_oldacl_byteswap, "oldacl" },
165 { zfs_acl_byteswap, "acl" }
166 };
167
168 int
169 dmu_buf_hold_noread_by_dnode(dnode_t *dn, uint64_t offset,
170 const void *tag, dmu_buf_t **dbp)
171 {
172 uint64_t blkid;
173 dmu_buf_impl_t *db;
174
175 rw_enter(&dn->dn_struct_rwlock, RW_READER);
176 blkid = dbuf_whichblock(dn, 0, offset);
177 db = dbuf_hold(dn, blkid, tag);
178 rw_exit(&dn->dn_struct_rwlock);
179
180 if (db == NULL) {
181 *dbp = NULL;
182 return (SET_ERROR(EIO));
183 }
184
185 *dbp = &db->db;
186 return (0);
187 }
188
189 int
190 dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset,
191 const void *tag, dmu_buf_t **dbp)
192 {
193 dnode_t *dn;
194 uint64_t blkid;
195 dmu_buf_impl_t *db;
196 int err;
197
198 err = dnode_hold(os, object, FTAG, &dn);
199 if (err)
200 return (err);
201 rw_enter(&dn->dn_struct_rwlock, RW_READER);
202 blkid = dbuf_whichblock(dn, 0, offset);
203 db = dbuf_hold(dn, blkid, tag);
204 rw_exit(&dn->dn_struct_rwlock);
205 dnode_rele(dn, FTAG);
206
207 if (db == NULL) {
208 *dbp = NULL;
209 return (SET_ERROR(EIO));
210 }
211
212 *dbp = &db->db;
213 return (err);
214 }
215
216 int
217 dmu_buf_hold_by_dnode(dnode_t *dn, uint64_t offset,
218 const void *tag, dmu_buf_t **dbp, int flags)
219 {
220 int err;
221 int db_flags = DB_RF_CANFAIL;
222
223 if (flags & DMU_READ_NO_PREFETCH)
224 db_flags |= DB_RF_NOPREFETCH;
225 if (flags & DMU_READ_NO_DECRYPT)
226 db_flags |= DB_RF_NO_DECRYPT;
227
228 err = dmu_buf_hold_noread_by_dnode(dn, offset, tag, dbp);
229 if (err == 0) {
230 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
231 err = dbuf_read(db, NULL, db_flags);
232 if (err != 0) {
233 dbuf_rele(db, tag);
234 *dbp = NULL;
235 }
236 }
237
238 return (err);
239 }
240
241 int
242 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
243 const void *tag, dmu_buf_t **dbp, int flags)
244 {
245 int err;
246 int db_flags = DB_RF_CANFAIL;
247
248 if (flags & DMU_READ_NO_PREFETCH)
249 db_flags |= DB_RF_NOPREFETCH;
250 if (flags & DMU_READ_NO_DECRYPT)
251 db_flags |= DB_RF_NO_DECRYPT;
252
253 err = dmu_buf_hold_noread(os, object, offset, tag, dbp);
254 if (err == 0) {
255 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
256 err = dbuf_read(db, NULL, db_flags);
257 if (err != 0) {
258 dbuf_rele(db, tag);
259 *dbp = NULL;
260 }
261 }
262
263 return (err);
264 }
265
266 int
267 dmu_bonus_max(void)
268 {
269 return (DN_OLD_MAX_BONUSLEN);
270 }
271
272 int
273 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, 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 (dn->dn_bonus != db) {
283 error = SET_ERROR(EINVAL);
284 } else if (newsize < 0 || newsize > db_fake->db_size) {
285 error = SET_ERROR(EINVAL);
286 } else {
287 dnode_setbonuslen(dn, newsize, tx);
288 error = 0;
289 }
290
291 DB_DNODE_EXIT(db);
292 return (error);
293 }
294
295 int
296 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
297 {
298 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
299 dnode_t *dn;
300 int error;
301
302 DB_DNODE_ENTER(db);
303 dn = DB_DNODE(db);
304
305 if (!DMU_OT_IS_VALID(type)) {
306 error = SET_ERROR(EINVAL);
307 } else if (dn->dn_bonus != db) {
308 error = SET_ERROR(EINVAL);
309 } else {
310 dnode_setbonus_type(dn, type, tx);
311 error = 0;
312 }
313
314 DB_DNODE_EXIT(db);
315 return (error);
316 }
317
318 dmu_object_type_t
319 dmu_get_bonustype(dmu_buf_t *db_fake)
320 {
321 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
322 dnode_t *dn;
323 dmu_object_type_t type;
324
325 DB_DNODE_ENTER(db);
326 dn = DB_DNODE(db);
327 type = dn->dn_bonustype;
328 DB_DNODE_EXIT(db);
329
330 return (type);
331 }
332
333 int
334 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
335 {
336 dnode_t *dn;
337 int error;
338
339 error = dnode_hold(os, object, FTAG, &dn);
340 dbuf_rm_spill(dn, tx);
341 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
342 dnode_rm_spill(dn, tx);
343 rw_exit(&dn->dn_struct_rwlock);
344 dnode_rele(dn, FTAG);
345 return (error);
346 }
347
348 /*
349 * Lookup and hold the bonus buffer for the provided dnode. If the dnode
350 * has not yet been allocated a new bonus dbuf a will be allocated.
351 * Returns ENOENT, EIO, or 0.
352 */
353 int dmu_bonus_hold_by_dnode(dnode_t *dn, const void *tag, dmu_buf_t **dbp,
354 uint32_t flags)
355 {
356 dmu_buf_impl_t *db;
357 int error;
358 uint32_t db_flags = DB_RF_MUST_SUCCEED;
359
360 if (flags & DMU_READ_NO_PREFETCH)
361 db_flags |= DB_RF_NOPREFETCH;
362 if (flags & DMU_READ_NO_DECRYPT)
363 db_flags |= DB_RF_NO_DECRYPT;
364
365 rw_enter(&dn->dn_struct_rwlock, RW_READER);
366 if (dn->dn_bonus == NULL) {
367 if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
368 rw_exit(&dn->dn_struct_rwlock);
369 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
370 }
371 if (dn->dn_bonus == NULL)
372 dbuf_create_bonus(dn);
373 }
374 db = dn->dn_bonus;
375
376 /* as long as the bonus buf is held, the dnode will be held */
377 if (zfs_refcount_add(&db->db_holds, tag) == 1) {
378 VERIFY(dnode_add_ref(dn, db));
379 atomic_inc_32(&dn->dn_dbufs_count);
380 }
381
382 /*
383 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
384 * hold and incrementing the dbuf count to ensure that dnode_move() sees
385 * a dnode hold for every dbuf.
386 */
387 rw_exit(&dn->dn_struct_rwlock);
388
389 error = dbuf_read(db, NULL, db_flags);
390 if (error) {
391 dnode_evict_bonus(dn);
392 dbuf_rele(db, tag);
393 *dbp = NULL;
394 return (error);
395 }
396
397 *dbp = &db->db;
398 return (0);
399 }
400
401 int
402 dmu_bonus_hold(objset_t *os, uint64_t object, const void *tag, dmu_buf_t **dbp)
403 {
404 dnode_t *dn;
405 int error;
406
407 error = dnode_hold(os, object, FTAG, &dn);
408 if (error)
409 return (error);
410
411 error = dmu_bonus_hold_by_dnode(dn, tag, dbp, DMU_READ_NO_PREFETCH);
412 dnode_rele(dn, FTAG);
413
414 return (error);
415 }
416
417 /*
418 * returns ENOENT, EIO, or 0.
419 *
420 * This interface will allocate a blank spill dbuf when a spill blk
421 * doesn't already exist on the dnode.
422 *
423 * if you only want to find an already existing spill db, then
424 * dmu_spill_hold_existing() should be used.
425 */
426 int
427 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, const void *tag,
428 dmu_buf_t **dbp)
429 {
430 dmu_buf_impl_t *db = NULL;
431 int err;
432
433 if ((flags & DB_RF_HAVESTRUCT) == 0)
434 rw_enter(&dn->dn_struct_rwlock, RW_READER);
435
436 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
437
438 if ((flags & DB_RF_HAVESTRUCT) == 0)
439 rw_exit(&dn->dn_struct_rwlock);
440
441 if (db == NULL) {
442 *dbp = NULL;
443 return (SET_ERROR(EIO));
444 }
445 err = dbuf_read(db, NULL, flags);
446 if (err == 0)
447 *dbp = &db->db;
448 else {
449 dbuf_rele(db, tag);
450 *dbp = NULL;
451 }
452 return (err);
453 }
454
455 int
456 dmu_spill_hold_existing(dmu_buf_t *bonus, const void *tag, dmu_buf_t **dbp)
457 {
458 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
459 dnode_t *dn;
460 int err;
461
462 DB_DNODE_ENTER(db);
463 dn = DB_DNODE(db);
464
465 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
466 err = SET_ERROR(EINVAL);
467 } else {
468 rw_enter(&dn->dn_struct_rwlock, RW_READER);
469
470 if (!dn->dn_have_spill) {
471 err = SET_ERROR(ENOENT);
472 } else {
473 err = dmu_spill_hold_by_dnode(dn,
474 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
475 }
476
477 rw_exit(&dn->dn_struct_rwlock);
478 }
479
480 DB_DNODE_EXIT(db);
481 return (err);
482 }
483
484 int
485 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, uint32_t flags, const void *tag,
486 dmu_buf_t **dbp)
487 {
488 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
489 dnode_t *dn;
490 int err;
491 uint32_t db_flags = DB_RF_CANFAIL;
492
493 if (flags & DMU_READ_NO_DECRYPT)
494 db_flags |= DB_RF_NO_DECRYPT;
495
496 DB_DNODE_ENTER(db);
497 dn = DB_DNODE(db);
498 err = dmu_spill_hold_by_dnode(dn, db_flags, tag, dbp);
499 DB_DNODE_EXIT(db);
500
501 return (err);
502 }
503
504 /*
505 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
506 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
507 * and can induce severe lock contention when writing to several files
508 * whose dnodes are in the same block.
509 */
510 int
511 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
512 boolean_t read, const void *tag, int *numbufsp, dmu_buf_t ***dbpp,
513 uint32_t flags)
514 {
515 dmu_buf_t **dbp;
516 zstream_t *zs = NULL;
517 uint64_t blkid, nblks, i;
518 uint32_t dbuf_flags;
519 int err;
520 zio_t *zio = NULL;
521 boolean_t missed = B_FALSE;
522
523 ASSERT(!read || length <= DMU_MAX_ACCESS);
524
525 /*
526 * Note: We directly notify the prefetch code of this read, so that
527 * we can tell it about the multi-block read. dbuf_read() only knows
528 * about the one block it is accessing.
529 */
530 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT |
531 DB_RF_NOPREFETCH;
532
533 if ((flags & DMU_READ_NO_DECRYPT) != 0)
534 dbuf_flags |= DB_RF_NO_DECRYPT;
535
536 rw_enter(&dn->dn_struct_rwlock, RW_READER);
537 if (dn->dn_datablkshift) {
538 int blkshift = dn->dn_datablkshift;
539 nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) -
540 P2ALIGN_TYPED(offset, 1ULL << blkshift, uint64_t))
541 >> blkshift;
542 } else {
543 if (offset + length > dn->dn_datablksz) {
544 zfs_panic_recover("zfs: accessing past end of object "
545 "%llx/%llx (size=%u access=%llu+%llu)",
546 (longlong_t)dn->dn_objset->
547 os_dsl_dataset->ds_object,
548 (longlong_t)dn->dn_object, dn->dn_datablksz,
549 (longlong_t)offset, (longlong_t)length);
550 rw_exit(&dn->dn_struct_rwlock);
551 return (SET_ERROR(EIO));
552 }
553 nblks = 1;
554 }
555 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
556
557 if (read)
558 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL,
559 ZIO_FLAG_CANFAIL);
560 blkid = dbuf_whichblock(dn, 0, offset);
561 if ((flags & DMU_READ_NO_PREFETCH) == 0) {
562 /*
563 * Prepare the zfetch before initiating the demand reads, so
564 * that if multiple threads block on same indirect block, we
565 * base predictions on the original less racy request order.
566 */
567 zs = dmu_zfetch_prepare(&dn->dn_zfetch, blkid, nblks, read,
568 B_TRUE);
569 }
570 for (i = 0; i < nblks; i++) {
571 dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag);
572 if (db == NULL) {
573 if (zs) {
574 dmu_zfetch_run(&dn->dn_zfetch, zs, missed,
575 B_TRUE);
576 }
577 rw_exit(&dn->dn_struct_rwlock);
578 dmu_buf_rele_array(dbp, nblks, tag);
579 if (read)
580 zio_nowait(zio);
581 return (SET_ERROR(EIO));
582 }
583
584 /*
585 * Initiate async demand data read.
586 * We check the db_state after calling dbuf_read() because
587 * (1) dbuf_read() may change the state to CACHED due to a
588 * hit in the ARC, and (2) on a cache miss, a child will
589 * have been added to "zio" but not yet completed, so the
590 * state will not yet be CACHED.
591 */
592 if (read) {
593 if (i == nblks - 1 && blkid + i < dn->dn_maxblkid &&
594 offset + length < db->db.db_offset +
595 db->db.db_size) {
596 if (offset <= db->db.db_offset)
597 dbuf_flags |= DB_RF_PARTIAL_FIRST;
598 else
599 dbuf_flags |= DB_RF_PARTIAL_MORE;
600 }
601 (void) dbuf_read(db, zio, dbuf_flags);
602 if (db->db_state != DB_CACHED)
603 missed = B_TRUE;
604 }
605 dbp[i] = &db->db;
606 }
607
608 if (!read)
609 zfs_racct_write(length, nblks);
610
611 if (zs)
612 dmu_zfetch_run(&dn->dn_zfetch, zs, missed, B_TRUE);
613 rw_exit(&dn->dn_struct_rwlock);
614
615 if (read) {
616 /* wait for async read i/o */
617 err = zio_wait(zio);
618 if (err) {
619 dmu_buf_rele_array(dbp, nblks, tag);
620 return (err);
621 }
622
623 /* wait for other io to complete */
624 for (i = 0; i < nblks; i++) {
625 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
626 mutex_enter(&db->db_mtx);
627 while (db->db_state == DB_READ ||
628 db->db_state == DB_FILL)
629 cv_wait(&db->db_changed, &db->db_mtx);
630 if (db->db_state == DB_UNCACHED)
631 err = SET_ERROR(EIO);
632 mutex_exit(&db->db_mtx);
633 if (err) {
634 dmu_buf_rele_array(dbp, nblks, tag);
635 return (err);
636 }
637 }
638 }
639
640 *numbufsp = nblks;
641 *dbpp = dbp;
642 return (0);
643 }
644
645 int
646 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
647 uint64_t length, int read, const void *tag, int *numbufsp,
648 dmu_buf_t ***dbpp)
649 {
650 dnode_t *dn;
651 int err;
652
653 err = dnode_hold(os, object, FTAG, &dn);
654 if (err)
655 return (err);
656
657 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
658 numbufsp, dbpp, DMU_READ_PREFETCH);
659
660 dnode_rele(dn, FTAG);
661
662 return (err);
663 }
664
665 int
666 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
667 uint64_t length, boolean_t read, const void *tag, int *numbufsp,
668 dmu_buf_t ***dbpp)
669 {
670 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
671 dnode_t *dn;
672 int err;
673
674 DB_DNODE_ENTER(db);
675 dn = DB_DNODE(db);
676 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
677 numbufsp, dbpp, DMU_READ_PREFETCH);
678 DB_DNODE_EXIT(db);
679
680 return (err);
681 }
682
683 void
684 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, const void *tag)
685 {
686 int i;
687 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
688
689 if (numbufs == 0)
690 return;
691
692 for (i = 0; i < numbufs; i++) {
693 if (dbp[i])
694 dbuf_rele(dbp[i], tag);
695 }
696
697 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
698 }
699
700 /*
701 * Issue prefetch I/Os for the given blocks. If level is greater than 0, the
702 * indirect blocks prefetched will be those that point to the blocks containing
703 * the data starting at offset, and continuing to offset + len. If the range
704 * it too long, prefetch the first dmu_prefetch_max bytes as requested, while
705 * for the rest only a higher level, also fitting within dmu_prefetch_max. It
706 * should primarily help random reads, since for long sequential reads there is
707 * a speculative prefetcher.
708 *
709 * Note that if the indirect blocks above the blocks being prefetched are not
710 * in cache, they will be asynchronously read in. Dnode read by dnode_hold()
711 * is currently synchronous.
712 */
713 void
714 dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset,
715 uint64_t len, zio_priority_t pri)
716 {
717 dnode_t *dn;
718
719 if (dmu_prefetch_max == 0 || len == 0) {
720 dmu_prefetch_dnode(os, object, pri);
721 return;
722 }
723
724 if (dnode_hold(os, object, FTAG, &dn) != 0)
725 return;
726
727 dmu_prefetch_by_dnode(dn, level, offset, len, pri);
728
729 dnode_rele(dn, FTAG);
730 }
731
732 void
733 dmu_prefetch_by_dnode(dnode_t *dn, int64_t level, uint64_t offset,
734 uint64_t len, zio_priority_t pri)
735 {
736 int64_t level2 = level;
737 uint64_t start, end, start2, end2;
738
739 /*
740 * Depending on len we may do two prefetches: blocks [start, end) at
741 * level, and following blocks [start2, end2) at higher level2.
742 */
743 rw_enter(&dn->dn_struct_rwlock, RW_READER);
744 if (dn->dn_datablkshift != 0) {
745 /*
746 * The object has multiple blocks. Calculate the full range
747 * of blocks [start, end2) and then split it into two parts,
748 * so that the first [start, end) fits into dmu_prefetch_max.
749 */
750 start = dbuf_whichblock(dn, level, offset);
751 end2 = dbuf_whichblock(dn, level, offset + len - 1) + 1;
752 uint8_t ibs = dn->dn_indblkshift;
753 uint8_t bs = (level == 0) ? dn->dn_datablkshift : ibs;
754 uint_t limit = P2ROUNDUP(dmu_prefetch_max, 1 << bs) >> bs;
755 start2 = end = MIN(end2, start + limit);
756
757 /*
758 * Find level2 where [start2, end2) fits into dmu_prefetch_max.
759 */
760 uint8_t ibps = ibs - SPA_BLKPTRSHIFT;
761 limit = P2ROUNDUP(dmu_prefetch_max, 1 << ibs) >> ibs;
762 do {
763 level2++;
764 start2 = P2ROUNDUP(start2, 1 << ibps) >> ibps;
765 end2 = P2ROUNDUP(end2, 1 << ibps) >> ibps;
766 } while (end2 - start2 > limit);
767 } else {
768 /* There is only one block. Prefetch it or nothing. */
769 start = start2 = end2 = 0;
770 end = start + (level == 0 && offset < dn->dn_datablksz);
771 }
772
773 for (uint64_t i = start; i < end; i++)
774 dbuf_prefetch(dn, level, i, pri, 0);
775 for (uint64_t i = start2; i < end2; i++)
776 dbuf_prefetch(dn, level2, i, pri, 0);
777 rw_exit(&dn->dn_struct_rwlock);
778 }
779
780 /*
781 * Issue prefetch I/Os for the given object's dnode.
782 */
783 void
784 dmu_prefetch_dnode(objset_t *os, uint64_t object, zio_priority_t pri)
785 {
786 if (object == 0 || object >= DN_MAX_OBJECT)
787 return;
788
789 dnode_t *dn = DMU_META_DNODE(os);
790 rw_enter(&dn->dn_struct_rwlock, RW_READER);
791 uint64_t blkid = dbuf_whichblock(dn, 0, object * sizeof (dnode_phys_t));
792 dbuf_prefetch(dn, 0, blkid, pri, 0);
793 rw_exit(&dn->dn_struct_rwlock);
794 }
795
796 /*
797 * Get the next "chunk" of file data to free. We traverse the file from
798 * the end so that the file gets shorter over time (if we crashes in the
799 * middle, this will leave us in a better state). We find allocated file
800 * data by simply searching the allocated level 1 indirects.
801 *
802 * On input, *start should be the first offset that does not need to be
803 * freed (e.g. "offset + length"). On return, *start will be the first
804 * offset that should be freed and l1blks is set to the number of level 1
805 * indirect blocks found within the chunk.
806 */
807 static int
808 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum, uint64_t *l1blks)
809 {
810 uint64_t blks;
811 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
812 /* bytes of data covered by a level-1 indirect block */
813 uint64_t iblkrange = (uint64_t)dn->dn_datablksz *
814 EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
815
816 ASSERT3U(minimum, <=, *start);
817
818 /*
819 * Check if we can free the entire range assuming that all of the
820 * L1 blocks in this range have data. If we can, we use this
821 * worst case value as an estimate so we can avoid having to look
822 * at the object's actual data.
823 */
824 uint64_t total_l1blks =
825 (roundup(*start, iblkrange) - (minimum / iblkrange * iblkrange)) /
826 iblkrange;
827 if (total_l1blks <= maxblks) {
828 *l1blks = total_l1blks;
829 *start = minimum;
830 return (0);
831 }
832 ASSERT(ISP2(iblkrange));
833
834 for (blks = 0; *start > minimum && blks < maxblks; blks++) {
835 int err;
836
837 /*
838 * dnode_next_offset(BACKWARDS) will find an allocated L1
839 * indirect block at or before the input offset. We must
840 * decrement *start so that it is at the end of the region
841 * to search.
842 */
843 (*start)--;
844
845 err = dnode_next_offset(dn,
846 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
847
848 /* if there are no indirect blocks before start, we are done */
849 if (err == ESRCH) {
850 *start = minimum;
851 break;
852 } else if (err != 0) {
853 *l1blks = blks;
854 return (err);
855 }
856
857 /* set start to the beginning of this L1 indirect */
858 *start = P2ALIGN_TYPED(*start, iblkrange, uint64_t);
859 }
860 if (*start < minimum)
861 *start = minimum;
862 *l1blks = blks;
863
864 return (0);
865 }
866
867 /*
868 * If this objset is of type OST_ZFS return true if vfs's unmounted flag is set,
869 * otherwise return false.
870 * Used below in dmu_free_long_range_impl() to enable abort when unmounting
871 */
872 static boolean_t
873 dmu_objset_zfs_unmounting(objset_t *os)
874 {
875 #ifdef _KERNEL
876 if (dmu_objset_type(os) == DMU_OST_ZFS)
877 return (zfs_get_vfs_flag_unmounted(os));
878 #else
879 (void) os;
880 #endif
881 return (B_FALSE);
882 }
883
884 static int
885 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
886 uint64_t length)
887 {
888 uint64_t object_size;
889 int err;
890 uint64_t dirty_frees_threshold;
891 dsl_pool_t *dp = dmu_objset_pool(os);
892
893 if (dn == NULL)
894 return (SET_ERROR(EINVAL));
895
896 object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
897 if (offset >= object_size)
898 return (0);
899
900 if (zfs_per_txg_dirty_frees_percent <= 100)
901 dirty_frees_threshold =
902 zfs_per_txg_dirty_frees_percent * zfs_dirty_data_max / 100;
903 else
904 dirty_frees_threshold = zfs_dirty_data_max / 20;
905
906 if (length == DMU_OBJECT_END || offset + length > object_size)
907 length = object_size - offset;
908
909 while (length != 0) {
910 uint64_t chunk_end, chunk_begin, chunk_len;
911 uint64_t l1blks;
912 dmu_tx_t *tx;
913
914 if (dmu_objset_zfs_unmounting(dn->dn_objset))
915 return (SET_ERROR(EINTR));
916
917 chunk_end = chunk_begin = offset + length;
918
919 /* move chunk_begin backwards to the beginning of this chunk */
920 err = get_next_chunk(dn, &chunk_begin, offset, &l1blks);
921 if (err)
922 return (err);
923 ASSERT3U(chunk_begin, >=, offset);
924 ASSERT3U(chunk_begin, <=, chunk_end);
925
926 chunk_len = chunk_end - chunk_begin;
927
928 tx = dmu_tx_create(os);
929 dmu_tx_hold_free(tx, dn->dn_object, chunk_begin, chunk_len);
930
931 /*
932 * Mark this transaction as typically resulting in a net
933 * reduction in space used.
934 */
935 dmu_tx_mark_netfree(tx);
936 err = dmu_tx_assign(tx, TXG_WAIT);
937 if (err) {
938 dmu_tx_abort(tx);
939 return (err);
940 }
941
942 uint64_t txg = dmu_tx_get_txg(tx);
943
944 mutex_enter(&dp->dp_lock);
945 uint64_t long_free_dirty =
946 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK];
947 mutex_exit(&dp->dp_lock);
948
949 /*
950 * To avoid filling up a TXG with just frees, wait for
951 * the next TXG to open before freeing more chunks if
952 * we have reached the threshold of frees.
953 */
954 if (dirty_frees_threshold != 0 &&
955 long_free_dirty >= dirty_frees_threshold) {
956 DMU_TX_STAT_BUMP(dmu_tx_dirty_frees_delay);
957 dmu_tx_commit(tx);
958 txg_wait_open(dp, 0, B_TRUE);
959 continue;
960 }
961
962 /*
963 * In order to prevent unnecessary write throttling, for each
964 * TXG, we track the cumulative size of L1 blocks being dirtied
965 * in dnode_free_range() below. We compare this number to a
966 * tunable threshold, past which we prevent new L1 dirty freeing
967 * blocks from being added into the open TXG. See
968 * dmu_free_long_range_impl() for details. The threshold
969 * prevents write throttle activation due to dirty freeing L1
970 * blocks taking up a large percentage of zfs_dirty_data_max.
971 */
972 mutex_enter(&dp->dp_lock);
973 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] +=
974 l1blks << dn->dn_indblkshift;
975 mutex_exit(&dp->dp_lock);
976 DTRACE_PROBE3(free__long__range,
977 uint64_t, long_free_dirty, uint64_t, chunk_len,
978 uint64_t, txg);
979 dnode_free_range(dn, chunk_begin, chunk_len, tx);
980
981 dmu_tx_commit(tx);
982
983 length -= chunk_len;
984 }
985 return (0);
986 }
987
988 int
989 dmu_free_long_range(objset_t *os, uint64_t object,
990 uint64_t offset, uint64_t length)
991 {
992 dnode_t *dn;
993 int err;
994
995 err = dnode_hold(os, object, FTAG, &dn);
996 if (err != 0)
997 return (err);
998 err = dmu_free_long_range_impl(os, dn, offset, length);
999
1000 /*
1001 * It is important to zero out the maxblkid when freeing the entire
1002 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
1003 * will take the fast path, and (b) dnode_reallocate() can verify
1004 * that the entire file has been freed.
1005 */
1006 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
1007 dn->dn_maxblkid = 0;
1008
1009 dnode_rele(dn, FTAG);
1010 return (err);
1011 }
1012
1013 int
1014 dmu_free_long_object(objset_t *os, uint64_t object)
1015 {
1016 dmu_tx_t *tx;
1017 int err;
1018
1019 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
1020 if (err != 0)
1021 return (err);
1022
1023 tx = dmu_tx_create(os);
1024 dmu_tx_hold_bonus(tx, object);
1025 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
1026 dmu_tx_mark_netfree(tx);
1027 err = dmu_tx_assign(tx, TXG_WAIT);
1028 if (err == 0) {
1029 err = dmu_object_free(os, object, tx);
1030 dmu_tx_commit(tx);
1031 } else {
1032 dmu_tx_abort(tx);
1033 }
1034
1035 return (err);
1036 }
1037
1038 int
1039 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
1040 uint64_t size, dmu_tx_t *tx)
1041 {
1042 dnode_t *dn;
1043 int err = dnode_hold(os, object, FTAG, &dn);
1044 if (err)
1045 return (err);
1046 ASSERT(offset < UINT64_MAX);
1047 ASSERT(size == DMU_OBJECT_END || size <= UINT64_MAX - offset);
1048 dnode_free_range(dn, offset, size, tx);
1049 dnode_rele(dn, FTAG);
1050 return (0);
1051 }
1052
1053 static int
1054 dmu_read_impl(dnode_t *dn, uint64_t offset, uint64_t size,
1055 void *buf, uint32_t flags)
1056 {
1057 dmu_buf_t **dbp;
1058 int numbufs, err = 0;
1059
1060 /*
1061 * Deal with odd block sizes, where there can't be data past the first
1062 * block. If we ever do the tail block optimization, we will need to
1063 * handle that here as well.
1064 */
1065 if (dn->dn_maxblkid == 0) {
1066 uint64_t newsz = offset > dn->dn_datablksz ? 0 :
1067 MIN(size, dn->dn_datablksz - offset);
1068 memset((char *)buf + newsz, 0, size - newsz);
1069 size = newsz;
1070 }
1071
1072 while (size > 0) {
1073 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
1074 int i;
1075
1076 /*
1077 * NB: we could do this block-at-a-time, but it's nice
1078 * to be reading in parallel.
1079 */
1080 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
1081 TRUE, FTAG, &numbufs, &dbp, flags);
1082 if (err)
1083 break;
1084
1085 for (i = 0; i < numbufs; i++) {
1086 uint64_t tocpy;
1087 int64_t bufoff;
1088 dmu_buf_t *db = dbp[i];
1089
1090 ASSERT(size > 0);
1091
1092 bufoff = offset - db->db_offset;
1093 tocpy = MIN(db->db_size - bufoff, size);
1094
1095 (void) memcpy(buf, (char *)db->db_data + bufoff, tocpy);
1096
1097 offset += tocpy;
1098 size -= tocpy;
1099 buf = (char *)buf + tocpy;
1100 }
1101 dmu_buf_rele_array(dbp, numbufs, FTAG);
1102 }
1103 return (err);
1104 }
1105
1106 int
1107 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1108 void *buf, uint32_t flags)
1109 {
1110 dnode_t *dn;
1111 int err;
1112
1113 err = dnode_hold(os, object, FTAG, &dn);
1114 if (err != 0)
1115 return (err);
1116
1117 err = dmu_read_impl(dn, offset, size, buf, flags);
1118 dnode_rele(dn, FTAG);
1119 return (err);
1120 }
1121
1122 int
1123 dmu_read_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, void *buf,
1124 uint32_t flags)
1125 {
1126 return (dmu_read_impl(dn, offset, size, buf, flags));
1127 }
1128
1129 static void
1130 dmu_write_impl(dmu_buf_t **dbp, int numbufs, uint64_t offset, uint64_t size,
1131 const void *buf, dmu_tx_t *tx)
1132 {
1133 int i;
1134
1135 for (i = 0; i < numbufs; i++) {
1136 uint64_t tocpy;
1137 int64_t bufoff;
1138 dmu_buf_t *db = dbp[i];
1139
1140 ASSERT(size > 0);
1141
1142 bufoff = offset - db->db_offset;
1143 tocpy = MIN(db->db_size - bufoff, size);
1144
1145 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1146
1147 if (tocpy == db->db_size)
1148 dmu_buf_will_fill(db, tx, B_FALSE);
1149 else
1150 dmu_buf_will_dirty(db, tx);
1151
1152 (void) memcpy((char *)db->db_data + bufoff, buf, tocpy);
1153
1154 if (tocpy == db->db_size)
1155 dmu_buf_fill_done(db, tx, B_FALSE);
1156
1157 offset += tocpy;
1158 size -= tocpy;
1159 buf = (char *)buf + tocpy;
1160 }
1161 }
1162
1163 void
1164 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1165 const void *buf, dmu_tx_t *tx)
1166 {
1167 dmu_buf_t **dbp;
1168 int numbufs;
1169
1170 if (size == 0)
1171 return;
1172
1173 VERIFY0(dmu_buf_hold_array(os, object, offset, size,
1174 FALSE, FTAG, &numbufs, &dbp));
1175 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1176 dmu_buf_rele_array(dbp, numbufs, FTAG);
1177 }
1178
1179 /*
1180 * Note: Lustre is an external consumer of this interface.
1181 */
1182 void
1183 dmu_write_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size,
1184 const void *buf, dmu_tx_t *tx)
1185 {
1186 dmu_buf_t **dbp;
1187 int numbufs;
1188
1189 if (size == 0)
1190 return;
1191
1192 VERIFY0(dmu_buf_hold_array_by_dnode(dn, offset, size,
1193 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH));
1194 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1195 dmu_buf_rele_array(dbp, numbufs, FTAG);
1196 }
1197
1198 void
1199 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1200 dmu_tx_t *tx)
1201 {
1202 dmu_buf_t **dbp;
1203 int numbufs, i;
1204
1205 if (size == 0)
1206 return;
1207
1208 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
1209 FALSE, FTAG, &numbufs, &dbp));
1210
1211 for (i = 0; i < numbufs; i++) {
1212 dmu_buf_t *db = dbp[i];
1213
1214 dmu_buf_will_not_fill(db, tx);
1215 }
1216 dmu_buf_rele_array(dbp, numbufs, FTAG);
1217 }
1218
1219 void
1220 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
1221 void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
1222 int compressed_size, int byteorder, dmu_tx_t *tx)
1223 {
1224 dmu_buf_t *db;
1225
1226 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
1227 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
1228 VERIFY0(dmu_buf_hold_noread(os, object, offset,
1229 FTAG, &db));
1230
1231 dmu_buf_write_embedded(db,
1232 data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
1233 uncompressed_size, compressed_size, byteorder, tx);
1234
1235 dmu_buf_rele(db, FTAG);
1236 }
1237
1238 void
1239 dmu_redact(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1240 dmu_tx_t *tx)
1241 {
1242 int numbufs, i;
1243 dmu_buf_t **dbp;
1244
1245 VERIFY0(dmu_buf_hold_array(os, object, offset, size, FALSE, FTAG,
1246 &numbufs, &dbp));
1247 for (i = 0; i < numbufs; i++)
1248 dmu_buf_redact(dbp[i], tx);
1249 dmu_buf_rele_array(dbp, numbufs, FTAG);
1250 }
1251
1252 #ifdef _KERNEL
1253 int
1254 dmu_read_uio_dnode(dnode_t *dn, zfs_uio_t *uio, uint64_t size)
1255 {
1256 dmu_buf_t **dbp;
1257 int numbufs, i, err;
1258
1259 /*
1260 * NB: we could do this block-at-a-time, but it's nice
1261 * to be reading in parallel.
1262 */
1263 err = dmu_buf_hold_array_by_dnode(dn, zfs_uio_offset(uio), size,
1264 TRUE, FTAG, &numbufs, &dbp, 0);
1265 if (err)
1266 return (err);
1267
1268 for (i = 0; i < numbufs; i++) {
1269 uint64_t tocpy;
1270 int64_t bufoff;
1271 dmu_buf_t *db = dbp[i];
1272
1273 ASSERT(size > 0);
1274
1275 bufoff = zfs_uio_offset(uio) - db->db_offset;
1276 tocpy = MIN(db->db_size - bufoff, size);
1277
1278 err = zfs_uio_fault_move((char *)db->db_data + bufoff, tocpy,
1279 UIO_READ, uio);
1280
1281 if (err)
1282 break;
1283
1284 size -= tocpy;
1285 }
1286 dmu_buf_rele_array(dbp, numbufs, FTAG);
1287
1288 return (err);
1289 }
1290
1291 /*
1292 * Read 'size' bytes into the uio buffer.
1293 * From object zdb->db_object.
1294 * Starting at zfs_uio_offset(uio).
1295 *
1296 * If the caller already has a dbuf in the target object
1297 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1298 * because we don't have to find the dnode_t for the object.
1299 */
1300 int
1301 dmu_read_uio_dbuf(dmu_buf_t *zdb, zfs_uio_t *uio, uint64_t size)
1302 {
1303 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1304 dnode_t *dn;
1305 int err;
1306
1307 if (size == 0)
1308 return (0);
1309
1310 DB_DNODE_ENTER(db);
1311 dn = DB_DNODE(db);
1312 err = dmu_read_uio_dnode(dn, uio, size);
1313 DB_DNODE_EXIT(db);
1314
1315 return (err);
1316 }
1317
1318 /*
1319 * Read 'size' bytes into the uio buffer.
1320 * From the specified object
1321 * Starting at offset zfs_uio_offset(uio).
1322 */
1323 int
1324 dmu_read_uio(objset_t *os, uint64_t object, zfs_uio_t *uio, uint64_t size)
1325 {
1326 dnode_t *dn;
1327 int err;
1328
1329 if (size == 0)
1330 return (0);
1331
1332 err = dnode_hold(os, object, FTAG, &dn);
1333 if (err)
1334 return (err);
1335
1336 err = dmu_read_uio_dnode(dn, uio, size);
1337
1338 dnode_rele(dn, FTAG);
1339
1340 return (err);
1341 }
1342
1343 int
1344 dmu_write_uio_dnode(dnode_t *dn, zfs_uio_t *uio, uint64_t size, dmu_tx_t *tx)
1345 {
1346 dmu_buf_t **dbp;
1347 int numbufs;
1348 int err = 0;
1349 int i;
1350
1351 err = dmu_buf_hold_array_by_dnode(dn, zfs_uio_offset(uio), size,
1352 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1353 if (err)
1354 return (err);
1355
1356 for (i = 0; i < numbufs; i++) {
1357 uint64_t tocpy;
1358 int64_t bufoff;
1359 dmu_buf_t *db = dbp[i];
1360
1361 ASSERT(size > 0);
1362
1363 offset_t off = zfs_uio_offset(uio);
1364 bufoff = off - db->db_offset;
1365 tocpy = MIN(db->db_size - bufoff, size);
1366
1367 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1368
1369 if (tocpy == db->db_size)
1370 dmu_buf_will_fill(db, tx, B_TRUE);
1371 else
1372 dmu_buf_will_dirty(db, tx);
1373
1374 err = zfs_uio_fault_move((char *)db->db_data + bufoff,
1375 tocpy, UIO_WRITE, uio);
1376
1377 if (tocpy == db->db_size && dmu_buf_fill_done(db, tx, err)) {
1378 /* The fill was reverted. Undo any uio progress. */
1379 zfs_uio_advance(uio, off - zfs_uio_offset(uio));
1380 }
1381
1382 if (err)
1383 break;
1384
1385 size -= tocpy;
1386 }
1387
1388 dmu_buf_rele_array(dbp, numbufs, FTAG);
1389 return (err);
1390 }
1391
1392 /*
1393 * Write 'size' bytes from the uio buffer.
1394 * To object zdb->db_object.
1395 * Starting at offset zfs_uio_offset(uio).
1396 *
1397 * If the caller already has a dbuf in the target object
1398 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1399 * because we don't have to find the dnode_t for the object.
1400 */
1401 int
1402 dmu_write_uio_dbuf(dmu_buf_t *zdb, zfs_uio_t *uio, uint64_t size,
1403 dmu_tx_t *tx)
1404 {
1405 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1406 dnode_t *dn;
1407 int err;
1408
1409 if (size == 0)
1410 return (0);
1411
1412 DB_DNODE_ENTER(db);
1413 dn = DB_DNODE(db);
1414 err = dmu_write_uio_dnode(dn, uio, size, tx);
1415 DB_DNODE_EXIT(db);
1416
1417 return (err);
1418 }
1419
1420 /*
1421 * Write 'size' bytes from the uio buffer.
1422 * To the specified object.
1423 * Starting at offset zfs_uio_offset(uio).
1424 */
1425 int
1426 dmu_write_uio(objset_t *os, uint64_t object, zfs_uio_t *uio, uint64_t size,
1427 dmu_tx_t *tx)
1428 {
1429 dnode_t *dn;
1430 int err;
1431
1432 if (size == 0)
1433 return (0);
1434
1435 err = dnode_hold(os, object, FTAG, &dn);
1436 if (err)
1437 return (err);
1438
1439 err = dmu_write_uio_dnode(dn, uio, size, tx);
1440
1441 dnode_rele(dn, FTAG);
1442
1443 return (err);
1444 }
1445 #endif /* _KERNEL */
1446
1447 /*
1448 * Allocate a loaned anonymous arc buffer.
1449 */
1450 arc_buf_t *
1451 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1452 {
1453 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1454
1455 return (arc_loan_buf(db->db_objset->os_spa, B_FALSE, size));
1456 }
1457
1458 /*
1459 * Free a loaned arc buffer.
1460 */
1461 void
1462 dmu_return_arcbuf(arc_buf_t *buf)
1463 {
1464 arc_return_buf(buf, FTAG);
1465 arc_buf_destroy(buf, FTAG);
1466 }
1467
1468 /*
1469 * A "lightweight" write is faster than a regular write (e.g.
1470 * dmu_write_by_dnode() or dmu_assign_arcbuf_by_dnode()), because it avoids the
1471 * CPU cost of creating a dmu_buf_impl_t and arc_buf_[hdr_]_t. However, the
1472 * data can not be read or overwritten until the transaction's txg has been
1473 * synced. This makes it appropriate for workloads that are known to be
1474 * (temporarily) write-only, like "zfs receive".
1475 *
1476 * A single block is written, starting at the specified offset in bytes. If
1477 * the call is successful, it returns 0 and the provided abd has been
1478 * consumed (the caller should not free it).
1479 */
1480 int
1481 dmu_lightweight_write_by_dnode(dnode_t *dn, uint64_t offset, abd_t *abd,
1482 const zio_prop_t *zp, zio_flag_t flags, dmu_tx_t *tx)
1483 {
1484 dbuf_dirty_record_t *dr =
1485 dbuf_dirty_lightweight(dn, dbuf_whichblock(dn, 0, offset), tx);
1486 if (dr == NULL)
1487 return (SET_ERROR(EIO));
1488 dr->dt.dll.dr_abd = abd;
1489 dr->dt.dll.dr_props = *zp;
1490 dr->dt.dll.dr_flags = flags;
1491 return (0);
1492 }
1493
1494 /*
1495 * When possible directly assign passed loaned arc buffer to a dbuf.
1496 * If this is not possible copy the contents of passed arc buf via
1497 * dmu_write().
1498 */
1499 int
1500 dmu_assign_arcbuf_by_dnode(dnode_t *dn, uint64_t offset, arc_buf_t *buf,
1501 dmu_tx_t *tx)
1502 {
1503 dmu_buf_impl_t *db;
1504 objset_t *os = dn->dn_objset;
1505 uint64_t object = dn->dn_object;
1506 uint32_t blksz = (uint32_t)arc_buf_lsize(buf);
1507 uint64_t blkid;
1508
1509 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1510 blkid = dbuf_whichblock(dn, 0, offset);
1511 db = dbuf_hold(dn, blkid, FTAG);
1512 rw_exit(&dn->dn_struct_rwlock);
1513 if (db == NULL)
1514 return (SET_ERROR(EIO));
1515
1516 /*
1517 * We can only assign if the offset is aligned and the arc buf is the
1518 * same size as the dbuf.
1519 */
1520 if (offset == db->db.db_offset && blksz == db->db.db_size) {
1521 zfs_racct_write(blksz, 1);
1522 dbuf_assign_arcbuf(db, buf, tx);
1523 dbuf_rele(db, FTAG);
1524 } else {
1525 /* compressed bufs must always be assignable to their dbuf */
1526 ASSERT3U(arc_get_compression(buf), ==, ZIO_COMPRESS_OFF);
1527 ASSERT(!(buf->b_flags & ARC_BUF_FLAG_COMPRESSED));
1528
1529 dbuf_rele(db, FTAG);
1530 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1531 dmu_return_arcbuf(buf);
1532 }
1533
1534 return (0);
1535 }
1536
1537 int
1538 dmu_assign_arcbuf_by_dbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1539 dmu_tx_t *tx)
1540 {
1541 int err;
1542 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1543
1544 DB_DNODE_ENTER(dbuf);
1545 err = dmu_assign_arcbuf_by_dnode(DB_DNODE(dbuf), offset, buf, tx);
1546 DB_DNODE_EXIT(dbuf);
1547
1548 return (err);
1549 }
1550
1551 typedef struct {
1552 dbuf_dirty_record_t *dsa_dr;
1553 dmu_sync_cb_t *dsa_done;
1554 zgd_t *dsa_zgd;
1555 dmu_tx_t *dsa_tx;
1556 } dmu_sync_arg_t;
1557
1558 static void
1559 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1560 {
1561 (void) buf;
1562 dmu_sync_arg_t *dsa = varg;
1563 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1564 blkptr_t *bp = zio->io_bp;
1565
1566 if (zio->io_error == 0) {
1567 if (BP_IS_HOLE(bp)) {
1568 /*
1569 * A block of zeros may compress to a hole, but the
1570 * block size still needs to be known for replay.
1571 */
1572 BP_SET_LSIZE(bp, db->db_size);
1573 } else if (!BP_IS_EMBEDDED(bp)) {
1574 ASSERT(BP_GET_LEVEL(bp) == 0);
1575 BP_SET_FILL(bp, 1);
1576 }
1577 }
1578 }
1579
1580 static void
1581 dmu_sync_late_arrival_ready(zio_t *zio)
1582 {
1583 dmu_sync_ready(zio, NULL, zio->io_private);
1584 }
1585
1586 static void
1587 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1588 {
1589 (void) buf;
1590 dmu_sync_arg_t *dsa = varg;
1591 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1592 dmu_buf_impl_t *db = dr->dr_dbuf;
1593 zgd_t *zgd = dsa->dsa_zgd;
1594
1595 /*
1596 * Record the vdev(s) backing this blkptr so they can be flushed after
1597 * the writes for the lwb have completed.
1598 */
1599 if (zio->io_error == 0) {
1600 zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp);
1601 }
1602
1603 mutex_enter(&db->db_mtx);
1604 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1605 if (zio->io_error == 0) {
1606 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1607 if (dr->dt.dl.dr_nopwrite) {
1608 blkptr_t *bp = zio->io_bp;
1609 blkptr_t *bp_orig = &zio->io_bp_orig;
1610 uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1611
1612 ASSERT(BP_EQUAL(bp, bp_orig));
1613 VERIFY(BP_EQUAL(bp, db->db_blkptr));
1614 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1615 VERIFY(zio_checksum_table[chksum].ci_flags &
1616 ZCHECKSUM_FLAG_NOPWRITE);
1617 }
1618 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1619 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1620 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1621
1622 /*
1623 * Old style holes are filled with all zeros, whereas
1624 * new-style holes maintain their lsize, type, level,
1625 * and birth time (see zio_write_compress). While we
1626 * need to reset the BP_SET_LSIZE() call that happened
1627 * in dmu_sync_ready for old style holes, we do *not*
1628 * want to wipe out the information contained in new
1629 * style holes. Thus, only zero out the block pointer if
1630 * it's an old style hole.
1631 */
1632 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) &&
1633 BP_GET_LOGICAL_BIRTH(&dr->dt.dl.dr_overridden_by) == 0)
1634 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1635 } else {
1636 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1637 }
1638 cv_broadcast(&db->db_changed);
1639 mutex_exit(&db->db_mtx);
1640
1641 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1642
1643 kmem_free(dsa, sizeof (*dsa));
1644 }
1645
1646 static void
1647 dmu_sync_late_arrival_done(zio_t *zio)
1648 {
1649 blkptr_t *bp = zio->io_bp;
1650 dmu_sync_arg_t *dsa = zio->io_private;
1651 zgd_t *zgd = dsa->dsa_zgd;
1652
1653 if (zio->io_error == 0) {
1654 /*
1655 * Record the vdev(s) backing this blkptr so they can be
1656 * flushed after the writes for the lwb have completed.
1657 */
1658 zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp);
1659
1660 if (!BP_IS_HOLE(bp)) {
1661 blkptr_t *bp_orig __maybe_unused = &zio->io_bp_orig;
1662 ASSERT(!(zio->io_flags & ZIO_FLAG_NOPWRITE));
1663 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1664 ASSERT(BP_GET_LOGICAL_BIRTH(zio->io_bp) == zio->io_txg);
1665 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1666 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1667 }
1668 }
1669
1670 dmu_tx_commit(dsa->dsa_tx);
1671
1672 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1673
1674 abd_free(zio->io_abd);
1675 kmem_free(dsa, sizeof (*dsa));
1676 }
1677
1678 static int
1679 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1680 zio_prop_t *zp, zbookmark_phys_t *zb)
1681 {
1682 dmu_sync_arg_t *dsa;
1683 dmu_tx_t *tx;
1684 int error;
1685
1686 error = dbuf_read((dmu_buf_impl_t *)zgd->zgd_db, NULL,
1687 DB_RF_CANFAIL | DB_RF_NOPREFETCH);
1688 if (error != 0)
1689 return (error);
1690
1691 tx = dmu_tx_create(os);
1692 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1693 /*
1694 * This transaction does not produce any dirty data or log blocks, so
1695 * it should not be throttled. All other cases wait for TXG sync, by
1696 * which time the log block we are writing will be obsolete, so we can
1697 * skip waiting and just return error here instead.
1698 */
1699 if (dmu_tx_assign(tx, TXG_NOWAIT | TXG_NOTHROTTLE) != 0) {
1700 dmu_tx_abort(tx);
1701 /* Make zl_get_data do txg_waited_synced() */
1702 return (SET_ERROR(EIO));
1703 }
1704
1705 /*
1706 * In order to prevent the zgd's lwb from being free'd prior to
1707 * dmu_sync_late_arrival_done() being called, we have to ensure
1708 * the lwb's "max txg" takes this tx's txg into account.
1709 */
1710 zil_lwb_add_txg(zgd->zgd_lwb, dmu_tx_get_txg(tx));
1711
1712 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1713 dsa->dsa_dr = NULL;
1714 dsa->dsa_done = done;
1715 dsa->dsa_zgd = zgd;
1716 dsa->dsa_tx = tx;
1717
1718 /*
1719 * Since we are currently syncing this txg, it's nontrivial to
1720 * determine what BP to nopwrite against, so we disable nopwrite.
1721 *
1722 * When syncing, the db_blkptr is initially the BP of the previous
1723 * txg. We can not nopwrite against it because it will be changed
1724 * (this is similar to the non-late-arrival case where the dbuf is
1725 * dirty in a future txg).
1726 *
1727 * Then dbuf_write_ready() sets bp_blkptr to the location we will write.
1728 * We can not nopwrite against it because although the BP will not
1729 * (typically) be changed, the data has not yet been persisted to this
1730 * location.
1731 *
1732 * Finally, when dbuf_write_done() is called, it is theoretically
1733 * possible to always nopwrite, because the data that was written in
1734 * this txg is the same data that we are trying to write. However we
1735 * would need to check that this dbuf is not dirty in any future
1736 * txg's (as we do in the normal dmu_sync() path). For simplicity, we
1737 * don't nopwrite in this case.
1738 */
1739 zp->zp_nopwrite = B_FALSE;
1740
1741 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1742 abd_get_from_buf(zgd->zgd_db->db_data, zgd->zgd_db->db_size),
1743 zgd->zgd_db->db_size, zgd->zgd_db->db_size, zp,
1744 dmu_sync_late_arrival_ready, NULL, dmu_sync_late_arrival_done,
1745 dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1746
1747 return (0);
1748 }
1749
1750 /*
1751 * Intent log support: sync the block associated with db to disk.
1752 * N.B. and XXX: the caller is responsible for making sure that the
1753 * data isn't changing while dmu_sync() is writing it.
1754 *
1755 * Return values:
1756 *
1757 * EEXIST: this txg has already been synced, so there's nothing to do.
1758 * The caller should not log the write.
1759 *
1760 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1761 * The caller should not log the write.
1762 *
1763 * EALREADY: this block is already in the process of being synced.
1764 * The caller should track its progress (somehow).
1765 *
1766 * EIO: could not do the I/O.
1767 * The caller should do a txg_wait_synced().
1768 *
1769 * 0: the I/O has been initiated.
1770 * The caller should log this blkptr in the done callback.
1771 * It is possible that the I/O will fail, in which case
1772 * the error will be reported to the done callback and
1773 * propagated to pio from zio_done().
1774 */
1775 int
1776 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1777 {
1778 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1779 objset_t *os = db->db_objset;
1780 dsl_dataset_t *ds = os->os_dsl_dataset;
1781 dbuf_dirty_record_t *dr, *dr_next;
1782 dmu_sync_arg_t *dsa;
1783 zbookmark_phys_t zb;
1784 zio_prop_t zp;
1785 dnode_t *dn;
1786
1787 ASSERT(pio != NULL);
1788 ASSERT(txg != 0);
1789
1790 SET_BOOKMARK(&zb, ds->ds_object,
1791 db->db.db_object, db->db_level, db->db_blkid);
1792
1793 DB_DNODE_ENTER(db);
1794 dn = DB_DNODE(db);
1795 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1796 DB_DNODE_EXIT(db);
1797
1798 /*
1799 * If we're frozen (running ziltest), we always need to generate a bp.
1800 */
1801 if (txg > spa_freeze_txg(os->os_spa))
1802 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1803
1804 /*
1805 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1806 * and us. If we determine that this txg is not yet syncing,
1807 * but it begins to sync a moment later, that's OK because the
1808 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1809 */
1810 mutex_enter(&db->db_mtx);
1811
1812 if (txg <= spa_last_synced_txg(os->os_spa)) {
1813 /*
1814 * This txg has already synced. There's nothing to do.
1815 */
1816 mutex_exit(&db->db_mtx);
1817 return (SET_ERROR(EEXIST));
1818 }
1819
1820 if (txg <= spa_syncing_txg(os->os_spa)) {
1821 /*
1822 * This txg is currently syncing, so we can't mess with
1823 * the dirty record anymore; just write a new log block.
1824 */
1825 mutex_exit(&db->db_mtx);
1826 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1827 }
1828
1829 dr = dbuf_find_dirty_eq(db, txg);
1830
1831 if (dr == NULL) {
1832 /*
1833 * There's no dr for this dbuf, so it must have been freed.
1834 * There's no need to log writes to freed blocks, so we're done.
1835 */
1836 mutex_exit(&db->db_mtx);
1837 return (SET_ERROR(ENOENT));
1838 }
1839
1840 dr_next = list_next(&db->db_dirty_records, dr);
1841 ASSERT(dr_next == NULL || dr_next->dr_txg < txg);
1842
1843 if (db->db_blkptr != NULL) {
1844 /*
1845 * We need to fill in zgd_bp with the current blkptr so that
1846 * the nopwrite code can check if we're writing the same
1847 * data that's already on disk. We can only nopwrite if we
1848 * are sure that after making the copy, db_blkptr will not
1849 * change until our i/o completes. We ensure this by
1850 * holding the db_mtx, and only allowing nopwrite if the
1851 * block is not already dirty (see below). This is verified
1852 * by dmu_sync_done(), which VERIFYs that the db_blkptr has
1853 * not changed.
1854 */
1855 *zgd->zgd_bp = *db->db_blkptr;
1856 }
1857
1858 /*
1859 * Assume the on-disk data is X, the current syncing data (in
1860 * txg - 1) is Y, and the current in-memory data is Z (currently
1861 * in dmu_sync).
1862 *
1863 * We usually want to perform a nopwrite if X and Z are the
1864 * same. However, if Y is different (i.e. the BP is going to
1865 * change before this write takes effect), then a nopwrite will
1866 * be incorrect - we would override with X, which could have
1867 * been freed when Y was written.
1868 *
1869 * (Note that this is not a concern when we are nop-writing from
1870 * syncing context, because X and Y must be identical, because
1871 * all previous txgs have been synced.)
1872 *
1873 * Therefore, we disable nopwrite if the current BP could change
1874 * before this TXG. There are two ways it could change: by
1875 * being dirty (dr_next is non-NULL), or by being freed
1876 * (dnode_block_freed()). This behavior is verified by
1877 * zio_done(), which VERIFYs that the override BP is identical
1878 * to the on-disk BP.
1879 */
1880 DB_DNODE_ENTER(db);
1881 dn = DB_DNODE(db);
1882 if (dr_next != NULL || dnode_block_freed(dn, db->db_blkid))
1883 zp.zp_nopwrite = B_FALSE;
1884 DB_DNODE_EXIT(db);
1885
1886 ASSERT(dr->dr_txg == txg);
1887 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1888 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1889 /*
1890 * We have already issued a sync write for this buffer,
1891 * or this buffer has already been synced. It could not
1892 * have been dirtied since, or we would have cleared the state.
1893 */
1894 mutex_exit(&db->db_mtx);
1895 return (SET_ERROR(EALREADY));
1896 }
1897
1898 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1899 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1900 mutex_exit(&db->db_mtx);
1901
1902 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1903 dsa->dsa_dr = dr;
1904 dsa->dsa_done = done;
1905 dsa->dsa_zgd = zgd;
1906 dsa->dsa_tx = NULL;
1907
1908 zio_nowait(arc_write(pio, os->os_spa, txg, zgd->zgd_bp,
1909 dr->dt.dl.dr_data, !DBUF_IS_CACHEABLE(db), dbuf_is_l2cacheable(db),
1910 &zp, dmu_sync_ready, NULL, dmu_sync_done, dsa,
1911 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb));
1912
1913 return (0);
1914 }
1915
1916 int
1917 dmu_object_set_nlevels(objset_t *os, uint64_t object, int nlevels, dmu_tx_t *tx)
1918 {
1919 dnode_t *dn;
1920 int err;
1921
1922 err = dnode_hold(os, object, FTAG, &dn);
1923 if (err)
1924 return (err);
1925 err = dnode_set_nlevels(dn, nlevels, tx);
1926 dnode_rele(dn, FTAG);
1927 return (err);
1928 }
1929
1930 int
1931 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1932 dmu_tx_t *tx)
1933 {
1934 dnode_t *dn;
1935 int err;
1936
1937 err = dnode_hold(os, object, FTAG, &dn);
1938 if (err)
1939 return (err);
1940 err = dnode_set_blksz(dn, size, ibs, tx);
1941 dnode_rele(dn, FTAG);
1942 return (err);
1943 }
1944
1945 int
1946 dmu_object_set_maxblkid(objset_t *os, uint64_t object, uint64_t maxblkid,
1947 dmu_tx_t *tx)
1948 {
1949 dnode_t *dn;
1950 int err;
1951
1952 err = dnode_hold(os, object, FTAG, &dn);
1953 if (err)
1954 return (err);
1955 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1956 dnode_new_blkid(dn, maxblkid, tx, B_FALSE, B_TRUE);
1957 rw_exit(&dn->dn_struct_rwlock);
1958 dnode_rele(dn, FTAG);
1959 return (0);
1960 }
1961
1962 void
1963 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1964 dmu_tx_t *tx)
1965 {
1966 dnode_t *dn;
1967
1968 /*
1969 * Send streams include each object's checksum function. This
1970 * check ensures that the receiving system can understand the
1971 * checksum function transmitted.
1972 */
1973 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS);
1974
1975 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1976 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS);
1977 dn->dn_checksum = checksum;
1978 dnode_setdirty(dn, tx);
1979 dnode_rele(dn, FTAG);
1980 }
1981
1982 void
1983 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1984 dmu_tx_t *tx)
1985 {
1986 dnode_t *dn;
1987
1988 /*
1989 * Send streams include each object's compression function. This
1990 * check ensures that the receiving system can understand the
1991 * compression function transmitted.
1992 */
1993 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS);
1994
1995 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1996 dn->dn_compress = compress;
1997 dnode_setdirty(dn, tx);
1998 dnode_rele(dn, FTAG);
1999 }
2000
2001 /*
2002 * When the "redundant_metadata" property is set to "most", only indirect
2003 * blocks of this level and higher will have an additional ditto block.
2004 */
2005 static const int zfs_redundant_metadata_most_ditto_level = 2;
2006
2007 void
2008 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
2009 {
2010 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
2011 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
2012 (wp & WP_SPILL));
2013 enum zio_checksum checksum = os->os_checksum;
2014 enum zio_compress compress = os->os_compress;
2015 uint8_t complevel = os->os_complevel;
2016 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
2017 boolean_t dedup = B_FALSE;
2018 boolean_t nopwrite = B_FALSE;
2019 boolean_t dedup_verify = os->os_dedup_verify;
2020 boolean_t encrypt = B_FALSE;
2021 int copies = os->os_copies;
2022
2023 /*
2024 * We maintain different write policies for each of the following
2025 * types of data:
2026 * 1. metadata
2027 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
2028 * 3. all other level 0 blocks
2029 */
2030 if (ismd) {
2031 /*
2032 * XXX -- we should design a compression algorithm
2033 * that specializes in arrays of bps.
2034 */
2035 compress = zio_compress_select(os->os_spa,
2036 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON);
2037
2038 /*
2039 * Metadata always gets checksummed. If the data
2040 * checksum is multi-bit correctable, and it's not a
2041 * ZBT-style checksum, then it's suitable for metadata
2042 * as well. Otherwise, the metadata checksum defaults
2043 * to fletcher4.
2044 */
2045 if (!(zio_checksum_table[checksum].ci_flags &
2046 ZCHECKSUM_FLAG_METADATA) ||
2047 (zio_checksum_table[checksum].ci_flags &
2048 ZCHECKSUM_FLAG_EMBEDDED))
2049 checksum = ZIO_CHECKSUM_FLETCHER_4;
2050
2051 switch (os->os_redundant_metadata) {
2052 case ZFS_REDUNDANT_METADATA_ALL:
2053 copies++;
2054 break;
2055 case ZFS_REDUNDANT_METADATA_MOST:
2056 if (level >= zfs_redundant_metadata_most_ditto_level ||
2057 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))
2058 copies++;
2059 break;
2060 case ZFS_REDUNDANT_METADATA_SOME:
2061 if (DMU_OT_IS_CRITICAL(type))
2062 copies++;
2063 break;
2064 case ZFS_REDUNDANT_METADATA_NONE:
2065 break;
2066 }
2067 } else if (wp & WP_NOFILL) {
2068 ASSERT(level == 0);
2069
2070 /*
2071 * If we're writing preallocated blocks, we aren't actually
2072 * writing them so don't set any policy properties. These
2073 * blocks are currently only used by an external subsystem
2074 * outside of zfs (i.e. dump) and not written by the zio
2075 * pipeline.
2076 */
2077 compress = ZIO_COMPRESS_OFF;
2078 checksum = ZIO_CHECKSUM_OFF;
2079 } else {
2080 compress = zio_compress_select(os->os_spa, dn->dn_compress,
2081 compress);
2082 complevel = zio_complevel_select(os->os_spa, compress,
2083 complevel, complevel);
2084
2085 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
2086 zio_checksum_select(dn->dn_checksum, checksum) :
2087 dedup_checksum;
2088
2089 /*
2090 * Determine dedup setting. If we are in dmu_sync(),
2091 * we won't actually dedup now because that's all
2092 * done in syncing context; but we do want to use the
2093 * dedup checksum. If the checksum is not strong
2094 * enough to ensure unique signatures, force
2095 * dedup_verify.
2096 */
2097 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
2098 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
2099 if (!(zio_checksum_table[checksum].ci_flags &
2100 ZCHECKSUM_FLAG_DEDUP))
2101 dedup_verify = B_TRUE;
2102 }
2103
2104 /*
2105 * Enable nopwrite if we have secure enough checksum
2106 * algorithm (see comment in zio_nop_write) and
2107 * compression is enabled. We don't enable nopwrite if
2108 * dedup is enabled as the two features are mutually
2109 * exclusive.
2110 */
2111 nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags &
2112 ZCHECKSUM_FLAG_NOPWRITE) &&
2113 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
2114 }
2115
2116 /*
2117 * All objects in an encrypted objset are protected from modification
2118 * via a MAC. Encrypted objects store their IV and salt in the last DVA
2119 * in the bp, so we cannot use all copies. Encrypted objects are also
2120 * not subject to nopwrite since writing the same data will still
2121 * result in a new ciphertext. Only encrypted blocks can be dedup'd
2122 * to avoid ambiguity in the dedup code since the DDT does not store
2123 * object types.
2124 */
2125 if (os->os_encrypted && (wp & WP_NOFILL) == 0) {
2126 encrypt = B_TRUE;
2127
2128 if (DMU_OT_IS_ENCRYPTED(type)) {
2129 copies = MIN(copies, SPA_DVAS_PER_BP - 1);
2130 nopwrite = B_FALSE;
2131 } else {
2132 dedup = B_FALSE;
2133 }
2134
2135 if (level <= 0 &&
2136 (type == DMU_OT_DNODE || type == DMU_OT_OBJSET)) {
2137 compress = ZIO_COMPRESS_EMPTY;
2138 }
2139 }
2140
2141 zp->zp_compress = compress;
2142 zp->zp_complevel = complevel;
2143 zp->zp_checksum = checksum;
2144 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
2145 zp->zp_level = level;
2146 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
2147 zp->zp_dedup = dedup;
2148 zp->zp_dedup_verify = dedup && dedup_verify;
2149 zp->zp_nopwrite = nopwrite;
2150 zp->zp_encrypt = encrypt;
2151 zp->zp_byteorder = ZFS_HOST_BYTEORDER;
2152 memset(zp->zp_salt, 0, ZIO_DATA_SALT_LEN);
2153 memset(zp->zp_iv, 0, ZIO_DATA_IV_LEN);
2154 memset(zp->zp_mac, 0, ZIO_DATA_MAC_LEN);
2155 zp->zp_zpl_smallblk = DMU_OT_IS_FILE(zp->zp_type) ?
2156 os->os_zpl_special_smallblock : 0;
2157
2158 ASSERT3U(zp->zp_compress, !=, ZIO_COMPRESS_INHERIT);
2159 }
2160
2161 /*
2162 * Reports the location of data and holes in an object. In order to
2163 * accurately report holes all dirty data must be synced to disk. This
2164 * causes extremely poor performance when seeking for holes in a dirty file.
2165 * As a compromise, only provide hole data when the dnode is clean. When
2166 * a dnode is dirty report the dnode as having no holes by returning EBUSY
2167 * which is always safe to do.
2168 */
2169 int
2170 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
2171 {
2172 dnode_t *dn;
2173 int restarted = 0, err;
2174
2175 restart:
2176 err = dnode_hold(os, object, FTAG, &dn);
2177 if (err)
2178 return (err);
2179
2180 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2181
2182 if (dnode_is_dirty(dn)) {
2183 /*
2184 * If the zfs_dmu_offset_next_sync module option is enabled
2185 * then hole reporting has been requested. Dirty dnodes
2186 * must be synced to disk to accurately report holes.
2187 *
2188 * Provided a RL_READER rangelock spanning 0-UINT64_MAX is
2189 * held by the caller only a single restart will be required.
2190 * We tolerate callers which do not hold the rangelock by
2191 * returning EBUSY and not reporting holes after one restart.
2192 */
2193 if (zfs_dmu_offset_next_sync) {
2194 rw_exit(&dn->dn_struct_rwlock);
2195 dnode_rele(dn, FTAG);
2196
2197 if (restarted)
2198 return (SET_ERROR(EBUSY));
2199
2200 txg_wait_synced(dmu_objset_pool(os), 0);
2201 restarted = 1;
2202 goto restart;
2203 }
2204
2205 err = SET_ERROR(EBUSY);
2206 } else {
2207 err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK |
2208 (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
2209 }
2210
2211 rw_exit(&dn->dn_struct_rwlock);
2212 dnode_rele(dn, FTAG);
2213
2214 return (err);
2215 }
2216
2217 int
2218 dmu_read_l0_bps(objset_t *os, uint64_t object, uint64_t offset, uint64_t length,
2219 blkptr_t *bps, size_t *nbpsp)
2220 {
2221 dmu_buf_t **dbp, *dbuf;
2222 dmu_buf_impl_t *db;
2223 blkptr_t *bp;
2224 int error, numbufs;
2225
2226 error = dmu_buf_hold_array(os, object, offset, length, FALSE, FTAG,
2227 &numbufs, &dbp);
2228 if (error != 0) {
2229 if (error == ESRCH) {
2230 error = SET_ERROR(ENXIO);
2231 }
2232 return (error);
2233 }
2234
2235 ASSERT3U(numbufs, <=, *nbpsp);
2236
2237 for (int i = 0; i < numbufs; i++) {
2238 dbuf = dbp[i];
2239 db = (dmu_buf_impl_t *)dbuf;
2240
2241 mutex_enter(&db->db_mtx);
2242
2243 if (!list_is_empty(&db->db_dirty_records)) {
2244 dbuf_dirty_record_t *dr;
2245
2246 dr = list_head(&db->db_dirty_records);
2247 if (dr->dt.dl.dr_brtwrite) {
2248 /*
2249 * This is very special case where we clone a
2250 * block and in the same transaction group we
2251 * read its BP (most likely to clone the clone).
2252 */
2253 bp = &dr->dt.dl.dr_overridden_by;
2254 } else {
2255 /*
2256 * The block was modified in the same
2257 * transaction group.
2258 */
2259 mutex_exit(&db->db_mtx);
2260 error = SET_ERROR(EAGAIN);
2261 goto out;
2262 }
2263 } else {
2264 bp = db->db_blkptr;
2265 }
2266
2267 mutex_exit(&db->db_mtx);
2268
2269 if (bp == NULL) {
2270 /*
2271 * The file size was increased, but the block was never
2272 * written, otherwise we would either have the block
2273 * pointer or the dirty record and would not get here.
2274 * It is effectively a hole, so report it as such.
2275 */
2276 BP_ZERO(&bps[i]);
2277 continue;
2278 }
2279 /*
2280 * Make sure we clone only data blocks.
2281 */
2282 if (BP_IS_METADATA(bp) && !BP_IS_HOLE(bp)) {
2283 error = SET_ERROR(EINVAL);
2284 goto out;
2285 }
2286
2287 /*
2288 * If the block was allocated in transaction group that is not
2289 * yet synced, we could clone it, but we couldn't write this
2290 * operation into ZIL, or it may be impossible to replay, since
2291 * the block may appear not yet allocated at that point.
2292 */
2293 if (BP_GET_BIRTH(bp) > spa_freeze_txg(os->os_spa)) {
2294 error = SET_ERROR(EINVAL);
2295 goto out;
2296 }
2297 if (BP_GET_BIRTH(bp) > spa_last_synced_txg(os->os_spa)) {
2298 error = SET_ERROR(EAGAIN);
2299 goto out;
2300 }
2301
2302 bps[i] = *bp;
2303 }
2304
2305 *nbpsp = numbufs;
2306 out:
2307 dmu_buf_rele_array(dbp, numbufs, FTAG);
2308
2309 return (error);
2310 }
2311
2312 int
2313 dmu_brt_clone(objset_t *os, uint64_t object, uint64_t offset, uint64_t length,
2314 dmu_tx_t *tx, const blkptr_t *bps, size_t nbps)
2315 {
2316 spa_t *spa;
2317 dmu_buf_t **dbp, *dbuf;
2318 dmu_buf_impl_t *db;
2319 struct dirty_leaf *dl;
2320 dbuf_dirty_record_t *dr;
2321 const blkptr_t *bp;
2322 int error = 0, i, numbufs;
2323
2324 spa = os->os_spa;
2325
2326 VERIFY0(dmu_buf_hold_array(os, object, offset, length, FALSE, FTAG,
2327 &numbufs, &dbp));
2328 ASSERT3U(nbps, ==, numbufs);
2329
2330 /*
2331 * Before we start cloning make sure that the dbufs sizes match new BPs
2332 * sizes. If they don't, that's a no-go, as we are not able to shrink
2333 * dbufs.
2334 */
2335 for (i = 0; i < numbufs; i++) {
2336 dbuf = dbp[i];
2337 db = (dmu_buf_impl_t *)dbuf;
2338 bp = &bps[i];
2339
2340 ASSERT0(db->db_level);
2341 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2342 ASSERT(db->db_blkid != DMU_SPILL_BLKID);
2343
2344 if (!BP_IS_HOLE(bp) && BP_GET_LSIZE(bp) != dbuf->db_size) {
2345 error = SET_ERROR(EXDEV);
2346 goto out;
2347 }
2348 }
2349
2350 for (i = 0; i < numbufs; i++) {
2351 dbuf = dbp[i];
2352 db = (dmu_buf_impl_t *)dbuf;
2353 bp = &bps[i];
2354
2355 ASSERT0(db->db_level);
2356 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2357 ASSERT(db->db_blkid != DMU_SPILL_BLKID);
2358 ASSERT(BP_IS_HOLE(bp) || dbuf->db_size == BP_GET_LSIZE(bp));
2359
2360 dmu_buf_will_clone(dbuf, tx);
2361
2362 mutex_enter(&db->db_mtx);
2363
2364 dr = list_head(&db->db_dirty_records);
2365 VERIFY(dr != NULL);
2366 ASSERT3U(dr->dr_txg, ==, tx->tx_txg);
2367 dl = &dr->dt.dl;
2368 dl->dr_overridden_by = *bp;
2369 if (!BP_IS_HOLE(bp) || BP_GET_LOGICAL_BIRTH(bp) != 0) {
2370 if (!BP_IS_EMBEDDED(bp)) {
2371 BP_SET_BIRTH(&dl->dr_overridden_by, dr->dr_txg,
2372 BP_GET_BIRTH(bp));
2373 } else {
2374 BP_SET_LOGICAL_BIRTH(&dl->dr_overridden_by,
2375 dr->dr_txg);
2376 }
2377 }
2378 dl->dr_brtwrite = B_TRUE;
2379 dl->dr_override_state = DR_OVERRIDDEN;
2380
2381 mutex_exit(&db->db_mtx);
2382
2383 /*
2384 * When data in embedded into BP there is no need to create
2385 * BRT entry as there is no data block. Just copy the BP as
2386 * it contains the data.
2387 */
2388 if (!BP_IS_HOLE(bp) && !BP_IS_EMBEDDED(bp)) {
2389 brt_pending_add(spa, bp, tx);
2390 }
2391 }
2392 out:
2393 dmu_buf_rele_array(dbp, numbufs, FTAG);
2394
2395 return (error);
2396 }
2397
2398 void
2399 __dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2400 {
2401 dnode_phys_t *dnp = dn->dn_phys;
2402
2403 doi->doi_data_block_size = dn->dn_datablksz;
2404 doi->doi_metadata_block_size = dn->dn_indblkshift ?
2405 1ULL << dn->dn_indblkshift : 0;
2406 doi->doi_type = dn->dn_type;
2407 doi->doi_bonus_type = dn->dn_bonustype;
2408 doi->doi_bonus_size = dn->dn_bonuslen;
2409 doi->doi_dnodesize = dn->dn_num_slots << DNODE_SHIFT;
2410 doi->doi_indirection = dn->dn_nlevels;
2411 doi->doi_checksum = dn->dn_checksum;
2412 doi->doi_compress = dn->dn_compress;
2413 doi->doi_nblkptr = dn->dn_nblkptr;
2414 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
2415 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
2416 doi->doi_fill_count = 0;
2417 for (int i = 0; i < dnp->dn_nblkptr; i++)
2418 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
2419 }
2420
2421 void
2422 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2423 {
2424 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2425 mutex_enter(&dn->dn_mtx);
2426
2427 __dmu_object_info_from_dnode(dn, doi);
2428
2429 mutex_exit(&dn->dn_mtx);
2430 rw_exit(&dn->dn_struct_rwlock);
2431 }
2432
2433 /*
2434 * Get information on a DMU object.
2435 * If doi is NULL, just indicates whether the object exists.
2436 */
2437 int
2438 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
2439 {
2440 dnode_t *dn;
2441 int err = dnode_hold(os, object, FTAG, &dn);
2442
2443 if (err)
2444 return (err);
2445
2446 if (doi != NULL)
2447 dmu_object_info_from_dnode(dn, doi);
2448
2449 dnode_rele(dn, FTAG);
2450 return (0);
2451 }
2452
2453 /*
2454 * As above, but faster; can be used when you have a held dbuf in hand.
2455 */
2456 void
2457 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
2458 {
2459 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2460
2461 DB_DNODE_ENTER(db);
2462 dmu_object_info_from_dnode(DB_DNODE(db), doi);
2463 DB_DNODE_EXIT(db);
2464 }
2465
2466 /*
2467 * Faster still when you only care about the size.
2468 */
2469 void
2470 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
2471 u_longlong_t *nblk512)
2472 {
2473 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2474 dnode_t *dn;
2475
2476 DB_DNODE_ENTER(db);
2477 dn = DB_DNODE(db);
2478
2479 *blksize = dn->dn_datablksz;
2480 /* add in number of slots used for the dnode itself */
2481 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
2482 SPA_MINBLOCKSHIFT) + dn->dn_num_slots;
2483 DB_DNODE_EXIT(db);
2484 }
2485
2486 void
2487 dmu_object_dnsize_from_db(dmu_buf_t *db_fake, int *dnsize)
2488 {
2489 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2490 dnode_t *dn;
2491
2492 DB_DNODE_ENTER(db);
2493 dn = DB_DNODE(db);
2494 *dnsize = dn->dn_num_slots << DNODE_SHIFT;
2495 DB_DNODE_EXIT(db);
2496 }
2497
2498 void
2499 byteswap_uint64_array(void *vbuf, size_t size)
2500 {
2501 uint64_t *buf = vbuf;
2502 size_t count = size >> 3;
2503 int i;
2504
2505 ASSERT((size & 7) == 0);
2506
2507 for (i = 0; i < count; i++)
2508 buf[i] = BSWAP_64(buf[i]);
2509 }
2510
2511 void
2512 byteswap_uint32_array(void *vbuf, size_t size)
2513 {
2514 uint32_t *buf = vbuf;
2515 size_t count = size >> 2;
2516 int i;
2517
2518 ASSERT((size & 3) == 0);
2519
2520 for (i = 0; i < count; i++)
2521 buf[i] = BSWAP_32(buf[i]);
2522 }
2523
2524 void
2525 byteswap_uint16_array(void *vbuf, size_t size)
2526 {
2527 uint16_t *buf = vbuf;
2528 size_t count = size >> 1;
2529 int i;
2530
2531 ASSERT((size & 1) == 0);
2532
2533 for (i = 0; i < count; i++)
2534 buf[i] = BSWAP_16(buf[i]);
2535 }
2536
2537 void
2538 byteswap_uint8_array(void *vbuf, size_t size)
2539 {
2540 (void) vbuf, (void) size;
2541 }
2542
2543 void
2544 dmu_init(void)
2545 {
2546 abd_init();
2547 zfs_dbgmsg_init();
2548 sa_cache_init();
2549 dmu_objset_init();
2550 dnode_init();
2551 zfetch_init();
2552 dmu_tx_init();
2553 l2arc_init();
2554 arc_init();
2555 dbuf_init();
2556 }
2557
2558 void
2559 dmu_fini(void)
2560 {
2561 arc_fini(); /* arc depends on l2arc, so arc must go first */
2562 l2arc_fini();
2563 dmu_tx_fini();
2564 zfetch_fini();
2565 dbuf_fini();
2566 dnode_fini();
2567 dmu_objset_fini();
2568 sa_cache_fini();
2569 zfs_dbgmsg_fini();
2570 abd_fini();
2571 }
2572
2573 EXPORT_SYMBOL(dmu_bonus_hold);
2574 EXPORT_SYMBOL(dmu_bonus_hold_by_dnode);
2575 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus);
2576 EXPORT_SYMBOL(dmu_buf_rele_array);
2577 EXPORT_SYMBOL(dmu_prefetch);
2578 EXPORT_SYMBOL(dmu_prefetch_by_dnode);
2579 EXPORT_SYMBOL(dmu_prefetch_dnode);
2580 EXPORT_SYMBOL(dmu_free_range);
2581 EXPORT_SYMBOL(dmu_free_long_range);
2582 EXPORT_SYMBOL(dmu_free_long_object);
2583 EXPORT_SYMBOL(dmu_read);
2584 EXPORT_SYMBOL(dmu_read_by_dnode);
2585 EXPORT_SYMBOL(dmu_write);
2586 EXPORT_SYMBOL(dmu_write_by_dnode);
2587 EXPORT_SYMBOL(dmu_prealloc);
2588 EXPORT_SYMBOL(dmu_object_info);
2589 EXPORT_SYMBOL(dmu_object_info_from_dnode);
2590 EXPORT_SYMBOL(dmu_object_info_from_db);
2591 EXPORT_SYMBOL(dmu_object_size_from_db);
2592 EXPORT_SYMBOL(dmu_object_dnsize_from_db);
2593 EXPORT_SYMBOL(dmu_object_set_nlevels);
2594 EXPORT_SYMBOL(dmu_object_set_blocksize);
2595 EXPORT_SYMBOL(dmu_object_set_maxblkid);
2596 EXPORT_SYMBOL(dmu_object_set_checksum);
2597 EXPORT_SYMBOL(dmu_object_set_compress);
2598 EXPORT_SYMBOL(dmu_offset_next);
2599 EXPORT_SYMBOL(dmu_write_policy);
2600 EXPORT_SYMBOL(dmu_sync);
2601 EXPORT_SYMBOL(dmu_request_arcbuf);
2602 EXPORT_SYMBOL(dmu_return_arcbuf);
2603 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dnode);
2604 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dbuf);
2605 EXPORT_SYMBOL(dmu_buf_hold);
2606 EXPORT_SYMBOL(dmu_ot);
2607
2608 ZFS_MODULE_PARAM(zfs, zfs_, nopwrite_enabled, INT, ZMOD_RW,
2609 "Enable NOP writes");
2610
2611 ZFS_MODULE_PARAM(zfs, zfs_, per_txg_dirty_frees_percent, UINT, ZMOD_RW,
2612 "Percentage of dirtied blocks from frees in one TXG");
2613
2614 ZFS_MODULE_PARAM(zfs, zfs_, dmu_offset_next_sync, INT, ZMOD_RW,
2615 "Enable forcing txg sync to find holes");
2616
2617 /* CSTYLED */
2618 ZFS_MODULE_PARAM(zfs, , dmu_prefetch_max, UINT, ZMOD_RW,
2619 "Limit one prefetch call to this size");
ZFS On Linux mirror for PVE