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
dmu_buf_hold_noread_by_dnode(dnode_t * dn,uint64_t offset,const void * tag,dmu_buf_t ** dbp)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
dmu_buf_hold_noread(objset_t * os,uint64_t object,uint64_t offset,const void * tag,dmu_buf_t ** dbp)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
dmu_buf_hold_by_dnode(dnode_t * dn,uint64_t offset,const void * tag,dmu_buf_t ** dbp,int flags)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
dmu_buf_hold(objset_t * os,uint64_t object,uint64_t offset,const void * tag,dmu_buf_t ** dbp,int flags)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
dmu_bonus_max(void)267 dmu_bonus_max(void)
268 {
269 return (DN_OLD_MAX_BONUSLEN);
270 }
271
272 int
dmu_set_bonus(dmu_buf_t * db_fake,int newsize,dmu_tx_t * tx)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 if (newsize < 0 || newsize > db_fake->db_size)
280 return (SET_ERROR(EINVAL));
281
282 DB_DNODE_ENTER(db);
283 dn = DB_DNODE(db);
284
285 if (dn->dn_bonus != db) {
286 error = SET_ERROR(EINVAL);
287 } else {
288 dnode_setbonuslen(dn, newsize, tx);
289 error = 0;
290 }
291
292 DB_DNODE_EXIT(db);
293 return (error);
294 }
295
296 int
dmu_set_bonustype(dmu_buf_t * db_fake,dmu_object_type_t type,dmu_tx_t * tx)297 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
298 {
299 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
300 dnode_t *dn;
301 int error;
302
303 if (!DMU_OT_IS_VALID(type))
304 return (SET_ERROR(EINVAL));
305
306 DB_DNODE_ENTER(db);
307 dn = DB_DNODE(db);
308
309 if (dn->dn_bonus != db) {
310 error = SET_ERROR(EINVAL);
311 } else {
312 dnode_setbonus_type(dn, type, tx);
313 error = 0;
314 }
315
316 DB_DNODE_EXIT(db);
317 return (error);
318 }
319
320 dmu_object_type_t
dmu_get_bonustype(dmu_buf_t * db_fake)321 dmu_get_bonustype(dmu_buf_t *db_fake)
322 {
323 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
324 dmu_object_type_t type;
325
326 DB_DNODE_ENTER(db);
327 type = DB_DNODE(db)->dn_bonustype;
328 DB_DNODE_EXIT(db);
329
330 return (type);
331 }
332
333 int
dmu_rm_spill(objset_t * os,uint64_t object,dmu_tx_t * tx)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 */
dmu_bonus_hold_by_dnode(dnode_t * dn,const void * tag,dmu_buf_t ** dbp,uint32_t flags)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
dmu_bonus_hold(objset_t * os,uint64_t object,const void * tag,dmu_buf_t ** dbp)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
dmu_spill_hold_by_dnode(dnode_t * dn,uint32_t flags,const void * tag,dmu_buf_t ** dbp)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
dmu_spill_hold_existing(dmu_buf_t * bonus,const void * tag,dmu_buf_t ** dbp)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
dmu_spill_hold_by_bonus(dmu_buf_t * bonus,uint32_t flags,const void * tag,dmu_buf_t ** dbp)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 int err;
490 uint32_t db_flags = DB_RF_CANFAIL;
491
492 if (flags & DMU_READ_NO_DECRYPT)
493 db_flags |= DB_RF_NO_DECRYPT;
494
495 DB_DNODE_ENTER(db);
496 err = dmu_spill_hold_by_dnode(DB_DNODE(db), db_flags, tag, dbp);
497 DB_DNODE_EXIT(db);
498
499 return (err);
500 }
501
502 /*
503 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
504 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
505 * and can induce severe lock contention when writing to several files
506 * whose dnodes are in the same block.
507 */
508 int
dmu_buf_hold_array_by_dnode(dnode_t * dn,uint64_t offset,uint64_t length,boolean_t read,const void * tag,int * numbufsp,dmu_buf_t *** dbpp,uint32_t flags)509 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
510 boolean_t read, const void *tag, int *numbufsp, dmu_buf_t ***dbpp,
511 uint32_t flags)
512 {
513 dmu_buf_t **dbp;
514 zstream_t *zs = NULL;
515 uint64_t blkid, nblks, i;
516 uint32_t dbuf_flags;
517 int err;
518 zio_t *zio = NULL;
519 boolean_t missed = B_FALSE;
520
521 ASSERT(!read || length <= DMU_MAX_ACCESS);
522
523 /*
524 * Note: We directly notify the prefetch code of this read, so that
525 * we can tell it about the multi-block read. dbuf_read() only knows
526 * about the one block it is accessing.
527 */
528 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT |
529 DB_RF_NOPREFETCH;
530
531 if ((flags & DMU_READ_NO_DECRYPT) != 0)
532 dbuf_flags |= DB_RF_NO_DECRYPT;
533
534 rw_enter(&dn->dn_struct_rwlock, RW_READER);
535 if (dn->dn_datablkshift) {
536 int blkshift = dn->dn_datablkshift;
537 nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) -
538 P2ALIGN_TYPED(offset, 1ULL << blkshift, uint64_t))
539 >> blkshift;
540 } else {
541 if (offset + length > dn->dn_datablksz) {
542 zfs_panic_recover("zfs: accessing past end of object "
543 "%llx/%llx (size=%u access=%llu+%llu)",
544 (longlong_t)dn->dn_objset->
545 os_dsl_dataset->ds_object,
546 (longlong_t)dn->dn_object, dn->dn_datablksz,
547 (longlong_t)offset, (longlong_t)length);
548 rw_exit(&dn->dn_struct_rwlock);
549 return (SET_ERROR(EIO));
550 }
551 nblks = 1;
552 }
553 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
554
555 if (read)
556 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL,
557 ZIO_FLAG_CANFAIL);
558 blkid = dbuf_whichblock(dn, 0, offset);
559 if ((flags & DMU_READ_NO_PREFETCH) == 0) {
560 /*
561 * Prepare the zfetch before initiating the demand reads, so
562 * that if multiple threads block on same indirect block, we
563 * base predictions on the original less racy request order.
564 */
565 zs = dmu_zfetch_prepare(&dn->dn_zfetch, blkid, nblks, read,
566 B_TRUE);
567 }
568 for (i = 0; i < nblks; i++) {
569 dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag);
570 if (db == NULL) {
571 if (zs) {
572 dmu_zfetch_run(&dn->dn_zfetch, zs, missed,
573 B_TRUE);
574 }
575 rw_exit(&dn->dn_struct_rwlock);
576 dmu_buf_rele_array(dbp, nblks, tag);
577 if (read)
578 zio_nowait(zio);
579 return (SET_ERROR(EIO));
580 }
581
582 /*
583 * Initiate async demand data read.
584 * We check the db_state after calling dbuf_read() because
585 * (1) dbuf_read() may change the state to CACHED due to a
586 * hit in the ARC, and (2) on a cache miss, a child will
587 * have been added to "zio" but not yet completed, so the
588 * state will not yet be CACHED.
589 */
590 if (read) {
591 if (i == nblks - 1 && blkid + i < dn->dn_maxblkid &&
592 offset + length < db->db.db_offset +
593 db->db.db_size) {
594 if (offset <= db->db.db_offset)
595 dbuf_flags |= DB_RF_PARTIAL_FIRST;
596 else
597 dbuf_flags |= DB_RF_PARTIAL_MORE;
598 }
599 (void) dbuf_read(db, zio, dbuf_flags);
600 if (db->db_state != DB_CACHED)
601 missed = B_TRUE;
602 }
603 dbp[i] = &db->db;
604 }
605
606 if (!read)
607 zfs_racct_write(length, nblks);
608
609 if (zs)
610 dmu_zfetch_run(&dn->dn_zfetch, zs, missed, B_TRUE);
611 rw_exit(&dn->dn_struct_rwlock);
612
613 if (read) {
614 /* wait for async read i/o */
615 err = zio_wait(zio);
616 if (err) {
617 dmu_buf_rele_array(dbp, nblks, tag);
618 return (err);
619 }
620
621 /* wait for other io to complete */
622 for (i = 0; i < nblks; i++) {
623 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
624 mutex_enter(&db->db_mtx);
625 while (db->db_state == DB_READ ||
626 db->db_state == DB_FILL)
627 cv_wait(&db->db_changed, &db->db_mtx);
628 if (db->db_state == DB_UNCACHED)
629 err = SET_ERROR(EIO);
630 mutex_exit(&db->db_mtx);
631 if (err) {
632 dmu_buf_rele_array(dbp, nblks, tag);
633 return (err);
634 }
635 }
636 }
637
638 *numbufsp = nblks;
639 *dbpp = dbp;
640 return (0);
641 }
642
643 int
dmu_buf_hold_array(objset_t * os,uint64_t object,uint64_t offset,uint64_t length,int read,const void * tag,int * numbufsp,dmu_buf_t *** dbpp)644 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
645 uint64_t length, int read, const void *tag, int *numbufsp,
646 dmu_buf_t ***dbpp)
647 {
648 dnode_t *dn;
649 int err;
650
651 err = dnode_hold(os, object, FTAG, &dn);
652 if (err)
653 return (err);
654
655 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
656 numbufsp, dbpp, DMU_READ_PREFETCH);
657
658 dnode_rele(dn, FTAG);
659
660 return (err);
661 }
662
663 int
dmu_buf_hold_array_by_bonus(dmu_buf_t * db_fake,uint64_t offset,uint64_t length,boolean_t read,const void * tag,int * numbufsp,dmu_buf_t *** dbpp)664 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
665 uint64_t length, boolean_t read, const void *tag, int *numbufsp,
666 dmu_buf_t ***dbpp)
667 {
668 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
669 int err;
670
671 DB_DNODE_ENTER(db);
672 err = dmu_buf_hold_array_by_dnode(DB_DNODE(db), offset, length, read,
673 tag, numbufsp, dbpp, DMU_READ_PREFETCH);
674 DB_DNODE_EXIT(db);
675
676 return (err);
677 }
678
679 void
dmu_buf_rele_array(dmu_buf_t ** dbp_fake,int numbufs,const void * tag)680 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, const void *tag)
681 {
682 int i;
683 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
684
685 if (numbufs == 0)
686 return;
687
688 for (i = 0; i < numbufs; i++) {
689 if (dbp[i])
690 dbuf_rele(dbp[i], tag);
691 }
692
693 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
694 }
695
696 /*
697 * Issue prefetch I/Os for the given blocks. If level is greater than 0, the
698 * indirect blocks prefetched will be those that point to the blocks containing
699 * the data starting at offset, and continuing to offset + len. If the range
700 * it too long, prefetch the first dmu_prefetch_max bytes as requested, while
701 * for the rest only a higher level, also fitting within dmu_prefetch_max. It
702 * should primarily help random reads, since for long sequential reads there is
703 * a speculative prefetcher.
704 *
705 * Note that if the indirect blocks above the blocks being prefetched are not
706 * in cache, they will be asynchronously read in. Dnode read by dnode_hold()
707 * is currently synchronous.
708 */
709 void
dmu_prefetch(objset_t * os,uint64_t object,int64_t level,uint64_t offset,uint64_t len,zio_priority_t pri)710 dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset,
711 uint64_t len, zio_priority_t pri)
712 {
713 dnode_t *dn;
714
715 if (dmu_prefetch_max == 0 || len == 0) {
716 dmu_prefetch_dnode(os, object, pri);
717 return;
718 }
719
720 if (dnode_hold(os, object, FTAG, &dn) != 0)
721 return;
722
723 dmu_prefetch_by_dnode(dn, level, offset, len, pri);
724
725 dnode_rele(dn, FTAG);
726 }
727
728 void
dmu_prefetch_by_dnode(dnode_t * dn,int64_t level,uint64_t offset,uint64_t len,zio_priority_t pri)729 dmu_prefetch_by_dnode(dnode_t *dn, int64_t level, uint64_t offset,
730 uint64_t len, zio_priority_t pri)
731 {
732 int64_t level2 = level;
733 uint64_t start, end, start2, end2;
734
735 /*
736 * Depending on len we may do two prefetches: blocks [start, end) at
737 * level, and following blocks [start2, end2) at higher level2.
738 */
739 rw_enter(&dn->dn_struct_rwlock, RW_READER);
740 if (dn->dn_datablkshift != 0) {
741 /*
742 * The object has multiple blocks. Calculate the full range
743 * of blocks [start, end2) and then split it into two parts,
744 * so that the first [start, end) fits into dmu_prefetch_max.
745 */
746 start = dbuf_whichblock(dn, level, offset);
747 end2 = dbuf_whichblock(dn, level, offset + len - 1) + 1;
748 uint8_t ibs = dn->dn_indblkshift;
749 uint8_t bs = (level == 0) ? dn->dn_datablkshift : ibs;
750 uint_t limit = P2ROUNDUP(dmu_prefetch_max, 1 << bs) >> bs;
751 start2 = end = MIN(end2, start + limit);
752
753 /*
754 * Find level2 where [start2, end2) fits into dmu_prefetch_max.
755 */
756 uint8_t ibps = ibs - SPA_BLKPTRSHIFT;
757 limit = P2ROUNDUP(dmu_prefetch_max, 1 << ibs) >> ibs;
758 do {
759 level2++;
760 start2 = P2ROUNDUP(start2, 1 << ibps) >> ibps;
761 end2 = P2ROUNDUP(end2, 1 << ibps) >> ibps;
762 } while (end2 - start2 > limit);
763 } else {
764 /* There is only one block. Prefetch it or nothing. */
765 start = start2 = end2 = 0;
766 end = start + (level == 0 && offset < dn->dn_datablksz);
767 }
768
769 for (uint64_t i = start; i < end; i++)
770 dbuf_prefetch(dn, level, i, pri, 0);
771 for (uint64_t i = start2; i < end2; i++)
772 dbuf_prefetch(dn, level2, i, pri, 0);
773 rw_exit(&dn->dn_struct_rwlock);
774 }
775
776 /*
777 * Issue prefetch I/Os for the given object's dnode.
778 */
779 void
dmu_prefetch_dnode(objset_t * os,uint64_t object,zio_priority_t pri)780 dmu_prefetch_dnode(objset_t *os, uint64_t object, zio_priority_t pri)
781 {
782 if (object == 0 || object >= DN_MAX_OBJECT)
783 return;
784
785 dnode_t *dn = DMU_META_DNODE(os);
786 rw_enter(&dn->dn_struct_rwlock, RW_READER);
787 uint64_t blkid = dbuf_whichblock(dn, 0, object * sizeof (dnode_phys_t));
788 dbuf_prefetch(dn, 0, blkid, pri, 0);
789 rw_exit(&dn->dn_struct_rwlock);
790 }
791
792 /*
793 * Get the next "chunk" of file data to free. We traverse the file from
794 * the end so that the file gets shorter over time (if we crashes in the
795 * middle, this will leave us in a better state). We find allocated file
796 * data by simply searching the allocated level 1 indirects.
797 *
798 * On input, *start should be the first offset that does not need to be
799 * freed (e.g. "offset + length"). On return, *start will be the first
800 * offset that should be freed and l1blks is set to the number of level 1
801 * indirect blocks found within the chunk.
802 */
803 static int
get_next_chunk(dnode_t * dn,uint64_t * start,uint64_t minimum,uint64_t * l1blks)804 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum, uint64_t *l1blks)
805 {
806 uint64_t blks;
807 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
808 /* bytes of data covered by a level-1 indirect block */
809 uint64_t iblkrange = (uint64_t)dn->dn_datablksz *
810 EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
811
812 ASSERT3U(minimum, <=, *start);
813
814 /* dn_nlevels == 1 means we don't have any L1 blocks */
815 if (dn->dn_nlevels <= 1) {
816 *l1blks = 0;
817 *start = minimum;
818 return (0);
819 }
820
821 /*
822 * Check if we can free the entire range assuming that all of the
823 * L1 blocks in this range have data. If we can, we use this
824 * worst case value as an estimate so we can avoid having to look
825 * at the object's actual data.
826 */
827 uint64_t total_l1blks =
828 (roundup(*start, iblkrange) - (minimum / iblkrange * iblkrange)) /
829 iblkrange;
830 if (total_l1blks <= maxblks) {
831 *l1blks = total_l1blks;
832 *start = minimum;
833 return (0);
834 }
835 ASSERT(ISP2(iblkrange));
836
837 for (blks = 0; *start > minimum && blks < maxblks; blks++) {
838 int err;
839
840 /*
841 * dnode_next_offset(BACKWARDS) will find an allocated L1
842 * indirect block at or before the input offset. We must
843 * decrement *start so that it is at the end of the region
844 * to search.
845 */
846 (*start)--;
847
848 err = dnode_next_offset(dn,
849 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
850
851 /* if there are no indirect blocks before start, we are done */
852 if (err == ESRCH) {
853 *start = minimum;
854 break;
855 } else if (err != 0) {
856 *l1blks = blks;
857 return (err);
858 }
859
860 /* set start to the beginning of this L1 indirect */
861 *start = P2ALIGN_TYPED(*start, iblkrange, uint64_t);
862 }
863 if (*start < minimum)
864 *start = minimum;
865 *l1blks = blks;
866
867 return (0);
868 }
869
870 /*
871 * If this objset is of type OST_ZFS return true if vfs's unmounted flag is set,
872 * otherwise return false.
873 * Used below in dmu_free_long_range_impl() to enable abort when unmounting
874 */
875 static boolean_t
dmu_objset_zfs_unmounting(objset_t * os)876 dmu_objset_zfs_unmounting(objset_t *os)
877 {
878 #ifdef _KERNEL
879 if (dmu_objset_type(os) == DMU_OST_ZFS)
880 return (zfs_get_vfs_flag_unmounted(os));
881 #else
882 (void) os;
883 #endif
884 return (B_FALSE);
885 }
886
887 static int
dmu_free_long_range_impl(objset_t * os,dnode_t * dn,uint64_t offset,uint64_t length)888 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
889 uint64_t length)
890 {
891 uint64_t object_size;
892 int err;
893 uint64_t dirty_frees_threshold;
894 dsl_pool_t *dp = dmu_objset_pool(os);
895
896 if (dn == NULL)
897 return (SET_ERROR(EINVAL));
898
899 object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
900 if (offset >= object_size)
901 return (0);
902
903 if (zfs_per_txg_dirty_frees_percent <= 100)
904 dirty_frees_threshold =
905 zfs_per_txg_dirty_frees_percent * zfs_dirty_data_max / 100;
906 else
907 dirty_frees_threshold = zfs_dirty_data_max / 20;
908
909 if (length == DMU_OBJECT_END || offset + length > object_size)
910 length = object_size - offset;
911
912 while (length != 0) {
913 uint64_t chunk_end, chunk_begin, chunk_len;
914 uint64_t l1blks;
915 dmu_tx_t *tx;
916
917 if (dmu_objset_zfs_unmounting(dn->dn_objset))
918 return (SET_ERROR(EINTR));
919
920 chunk_end = chunk_begin = offset + length;
921
922 /* move chunk_begin backwards to the beginning of this chunk */
923 err = get_next_chunk(dn, &chunk_begin, offset, &l1blks);
924 if (err)
925 return (err);
926 ASSERT3U(chunk_begin, >=, offset);
927 ASSERT3U(chunk_begin, <=, chunk_end);
928
929 chunk_len = chunk_end - chunk_begin;
930
931 tx = dmu_tx_create(os);
932 dmu_tx_hold_free(tx, dn->dn_object, chunk_begin, chunk_len);
933
934 /*
935 * Mark this transaction as typically resulting in a net
936 * reduction in space used.
937 */
938 dmu_tx_mark_netfree(tx);
939 err = dmu_tx_assign(tx, TXG_WAIT);
940 if (err) {
941 dmu_tx_abort(tx);
942 return (err);
943 }
944
945 uint64_t txg = dmu_tx_get_txg(tx);
946
947 mutex_enter(&dp->dp_lock);
948 uint64_t long_free_dirty =
949 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK];
950 mutex_exit(&dp->dp_lock);
951
952 /*
953 * To avoid filling up a TXG with just frees, wait for
954 * the next TXG to open before freeing more chunks if
955 * we have reached the threshold of frees.
956 */
957 if (dirty_frees_threshold != 0 &&
958 long_free_dirty >= dirty_frees_threshold) {
959 DMU_TX_STAT_BUMP(dmu_tx_dirty_frees_delay);
960 dmu_tx_commit(tx);
961 txg_wait_open(dp, 0, B_TRUE);
962 continue;
963 }
964
965 /*
966 * In order to prevent unnecessary write throttling, for each
967 * TXG, we track the cumulative size of L1 blocks being dirtied
968 * in dnode_free_range() below. We compare this number to a
969 * tunable threshold, past which we prevent new L1 dirty freeing
970 * blocks from being added into the open TXG. See
971 * dmu_free_long_range_impl() for details. The threshold
972 * prevents write throttle activation due to dirty freeing L1
973 * blocks taking up a large percentage of zfs_dirty_data_max.
974 */
975 mutex_enter(&dp->dp_lock);
976 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] +=
977 l1blks << dn->dn_indblkshift;
978 mutex_exit(&dp->dp_lock);
979 DTRACE_PROBE3(free__long__range,
980 uint64_t, long_free_dirty, uint64_t, chunk_len,
981 uint64_t, txg);
982 dnode_free_range(dn, chunk_begin, chunk_len, tx);
983
984 dmu_tx_commit(tx);
985
986 length -= chunk_len;
987 }
988 return (0);
989 }
990
991 int
dmu_free_long_range(objset_t * os,uint64_t object,uint64_t offset,uint64_t length)992 dmu_free_long_range(objset_t *os, uint64_t object,
993 uint64_t offset, uint64_t length)
994 {
995 dnode_t *dn;
996 int err;
997
998 err = dnode_hold(os, object, FTAG, &dn);
999 if (err != 0)
1000 return (err);
1001 err = dmu_free_long_range_impl(os, dn, offset, length);
1002
1003 /*
1004 * It is important to zero out the maxblkid when freeing the entire
1005 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
1006 * will take the fast path, and (b) dnode_reallocate() can verify
1007 * that the entire file has been freed.
1008 */
1009 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
1010 dn->dn_maxblkid = 0;
1011
1012 dnode_rele(dn, FTAG);
1013 return (err);
1014 }
1015
1016 int
dmu_free_long_object(objset_t * os,uint64_t object)1017 dmu_free_long_object(objset_t *os, uint64_t object)
1018 {
1019 dmu_tx_t *tx;
1020 int err;
1021
1022 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
1023 if (err != 0)
1024 return (err);
1025
1026 tx = dmu_tx_create(os);
1027 dmu_tx_hold_bonus(tx, object);
1028 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
1029 dmu_tx_mark_netfree(tx);
1030 err = dmu_tx_assign(tx, TXG_WAIT);
1031 if (err == 0) {
1032 err = dmu_object_free(os, object, tx);
1033 dmu_tx_commit(tx);
1034 } else {
1035 dmu_tx_abort(tx);
1036 }
1037
1038 return (err);
1039 }
1040
1041 int
dmu_free_range(objset_t * os,uint64_t object,uint64_t offset,uint64_t size,dmu_tx_t * tx)1042 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
1043 uint64_t size, dmu_tx_t *tx)
1044 {
1045 dnode_t *dn;
1046 int err = dnode_hold(os, object, FTAG, &dn);
1047 if (err)
1048 return (err);
1049 ASSERT(offset < UINT64_MAX);
1050 ASSERT(size == DMU_OBJECT_END || size <= UINT64_MAX - offset);
1051 dnode_free_range(dn, offset, size, tx);
1052 dnode_rele(dn, FTAG);
1053 return (0);
1054 }
1055
1056 static int
dmu_read_impl(dnode_t * dn,uint64_t offset,uint64_t size,void * buf,uint32_t flags)1057 dmu_read_impl(dnode_t *dn, uint64_t offset, uint64_t size,
1058 void *buf, uint32_t flags)
1059 {
1060 dmu_buf_t **dbp;
1061 int numbufs, err = 0;
1062
1063 /*
1064 * Deal with odd block sizes, where there can't be data past the first
1065 * block. If we ever do the tail block optimization, we will need to
1066 * handle that here as well.
1067 */
1068 if (dn->dn_maxblkid == 0) {
1069 uint64_t newsz = offset > dn->dn_datablksz ? 0 :
1070 MIN(size, dn->dn_datablksz - offset);
1071 memset((char *)buf + newsz, 0, size - newsz);
1072 size = newsz;
1073 }
1074
1075 while (size > 0) {
1076 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
1077 int i;
1078
1079 /*
1080 * NB: we could do this block-at-a-time, but it's nice
1081 * to be reading in parallel.
1082 */
1083 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
1084 TRUE, FTAG, &numbufs, &dbp, flags);
1085 if (err)
1086 break;
1087
1088 for (i = 0; i < numbufs; i++) {
1089 uint64_t tocpy;
1090 int64_t bufoff;
1091 dmu_buf_t *db = dbp[i];
1092
1093 ASSERT(size > 0);
1094
1095 bufoff = offset - db->db_offset;
1096 tocpy = MIN(db->db_size - bufoff, size);
1097
1098 (void) memcpy(buf, (char *)db->db_data + bufoff, tocpy);
1099
1100 offset += tocpy;
1101 size -= tocpy;
1102 buf = (char *)buf + tocpy;
1103 }
1104 dmu_buf_rele_array(dbp, numbufs, FTAG);
1105 }
1106 return (err);
1107 }
1108
1109 int
dmu_read(objset_t * os,uint64_t object,uint64_t offset,uint64_t size,void * buf,uint32_t flags)1110 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1111 void *buf, uint32_t flags)
1112 {
1113 dnode_t *dn;
1114 int err;
1115
1116 err = dnode_hold(os, object, FTAG, &dn);
1117 if (err != 0)
1118 return (err);
1119
1120 err = dmu_read_impl(dn, offset, size, buf, flags);
1121 dnode_rele(dn, FTAG);
1122 return (err);
1123 }
1124
1125 int
dmu_read_by_dnode(dnode_t * dn,uint64_t offset,uint64_t size,void * buf,uint32_t flags)1126 dmu_read_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, void *buf,
1127 uint32_t flags)
1128 {
1129 return (dmu_read_impl(dn, offset, size, buf, flags));
1130 }
1131
1132 static void
dmu_write_impl(dmu_buf_t ** dbp,int numbufs,uint64_t offset,uint64_t size,const void * buf,dmu_tx_t * tx)1133 dmu_write_impl(dmu_buf_t **dbp, int numbufs, uint64_t offset, uint64_t size,
1134 const void *buf, dmu_tx_t *tx)
1135 {
1136 int i;
1137
1138 for (i = 0; i < numbufs; i++) {
1139 uint64_t tocpy;
1140 int64_t bufoff;
1141 dmu_buf_t *db = dbp[i];
1142
1143 ASSERT(size > 0);
1144
1145 bufoff = offset - db->db_offset;
1146 tocpy = MIN(db->db_size - bufoff, size);
1147
1148 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1149
1150 if (tocpy == db->db_size)
1151 dmu_buf_will_fill(db, tx, B_FALSE);
1152 else
1153 dmu_buf_will_dirty(db, tx);
1154
1155 (void) memcpy((char *)db->db_data + bufoff, buf, tocpy);
1156
1157 if (tocpy == db->db_size)
1158 dmu_buf_fill_done(db, tx, B_FALSE);
1159
1160 offset += tocpy;
1161 size -= tocpy;
1162 buf = (char *)buf + tocpy;
1163 }
1164 }
1165
1166 void
dmu_write(objset_t * os,uint64_t object,uint64_t offset,uint64_t size,const void * buf,dmu_tx_t * tx)1167 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1168 const void *buf, dmu_tx_t *tx)
1169 {
1170 dmu_buf_t **dbp;
1171 int numbufs;
1172
1173 if (size == 0)
1174 return;
1175
1176 VERIFY0(dmu_buf_hold_array(os, object, offset, size,
1177 FALSE, FTAG, &numbufs, &dbp));
1178 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1179 dmu_buf_rele_array(dbp, numbufs, FTAG);
1180 }
1181
1182 /*
1183 * Note: Lustre is an external consumer of this interface.
1184 */
1185 void
dmu_write_by_dnode(dnode_t * dn,uint64_t offset,uint64_t size,const void * buf,dmu_tx_t * tx)1186 dmu_write_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size,
1187 const void *buf, dmu_tx_t *tx)
1188 {
1189 dmu_buf_t **dbp;
1190 int numbufs;
1191
1192 if (size == 0)
1193 return;
1194
1195 VERIFY0(dmu_buf_hold_array_by_dnode(dn, offset, size,
1196 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH));
1197 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1198 dmu_buf_rele_array(dbp, numbufs, FTAG);
1199 }
1200
1201 void
dmu_prealloc(objset_t * os,uint64_t object,uint64_t offset,uint64_t size,dmu_tx_t * tx)1202 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1203 dmu_tx_t *tx)
1204 {
1205 dmu_buf_t **dbp;
1206 int numbufs, i;
1207
1208 if (size == 0)
1209 return;
1210
1211 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
1212 FALSE, FTAG, &numbufs, &dbp));
1213
1214 for (i = 0; i < numbufs; i++) {
1215 dmu_buf_t *db = dbp[i];
1216
1217 dmu_buf_will_not_fill(db, tx);
1218 }
1219 dmu_buf_rele_array(dbp, numbufs, FTAG);
1220 }
1221
1222 void
dmu_write_embedded(objset_t * os,uint64_t object,uint64_t offset,void * data,uint8_t etype,uint8_t comp,int uncompressed_size,int compressed_size,int byteorder,dmu_tx_t * tx)1223 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
1224 void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
1225 int compressed_size, int byteorder, dmu_tx_t *tx)
1226 {
1227 dmu_buf_t *db;
1228
1229 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
1230 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
1231 VERIFY0(dmu_buf_hold_noread(os, object, offset,
1232 FTAG, &db));
1233
1234 dmu_buf_write_embedded(db,
1235 data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
1236 uncompressed_size, compressed_size, byteorder, tx);
1237
1238 dmu_buf_rele(db, FTAG);
1239 }
1240
1241 void
dmu_redact(objset_t * os,uint64_t object,uint64_t offset,uint64_t size,dmu_tx_t * tx)1242 dmu_redact(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1243 dmu_tx_t *tx)
1244 {
1245 int numbufs, i;
1246 dmu_buf_t **dbp;
1247
1248 VERIFY0(dmu_buf_hold_array(os, object, offset, size, FALSE, FTAG,
1249 &numbufs, &dbp));
1250 for (i = 0; i < numbufs; i++)
1251 dmu_buf_redact(dbp[i], tx);
1252 dmu_buf_rele_array(dbp, numbufs, FTAG);
1253 }
1254
1255 #ifdef _KERNEL
1256 int
dmu_read_uio_dnode(dnode_t * dn,zfs_uio_t * uio,uint64_t size)1257 dmu_read_uio_dnode(dnode_t *dn, zfs_uio_t *uio, uint64_t size)
1258 {
1259 dmu_buf_t **dbp;
1260 int numbufs, i, err;
1261
1262 /*
1263 * NB: we could do this block-at-a-time, but it's nice
1264 * to be reading in parallel.
1265 */
1266 err = dmu_buf_hold_array_by_dnode(dn, zfs_uio_offset(uio), size,
1267 TRUE, FTAG, &numbufs, &dbp, 0);
1268 if (err)
1269 return (err);
1270
1271 for (i = 0; i < numbufs; i++) {
1272 uint64_t tocpy;
1273 int64_t bufoff;
1274 dmu_buf_t *db = dbp[i];
1275
1276 ASSERT(size > 0);
1277
1278 bufoff = zfs_uio_offset(uio) - db->db_offset;
1279 tocpy = MIN(db->db_size - bufoff, size);
1280
1281 err = zfs_uio_fault_move((char *)db->db_data + bufoff, tocpy,
1282 UIO_READ, uio);
1283
1284 if (err)
1285 break;
1286
1287 size -= tocpy;
1288 }
1289 dmu_buf_rele_array(dbp, numbufs, FTAG);
1290
1291 return (err);
1292 }
1293
1294 /*
1295 * Read 'size' bytes into the uio buffer.
1296 * From object zdb->db_object.
1297 * Starting at zfs_uio_offset(uio).
1298 *
1299 * If the caller already has a dbuf in the target object
1300 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1301 * because we don't have to find the dnode_t for the object.
1302 */
1303 int
dmu_read_uio_dbuf(dmu_buf_t * zdb,zfs_uio_t * uio,uint64_t size)1304 dmu_read_uio_dbuf(dmu_buf_t *zdb, zfs_uio_t *uio, uint64_t size)
1305 {
1306 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1307 int err;
1308
1309 if (size == 0)
1310 return (0);
1311
1312 DB_DNODE_ENTER(db);
1313 err = dmu_read_uio_dnode(DB_DNODE(db), uio, size);
1314 DB_DNODE_EXIT(db);
1315
1316 return (err);
1317 }
1318
1319 /*
1320 * Read 'size' bytes into the uio buffer.
1321 * From the specified object
1322 * Starting at offset zfs_uio_offset(uio).
1323 */
1324 int
dmu_read_uio(objset_t * os,uint64_t object,zfs_uio_t * uio,uint64_t size)1325 dmu_read_uio(objset_t *os, uint64_t object, zfs_uio_t *uio, uint64_t size)
1326 {
1327 dnode_t *dn;
1328 int err;
1329
1330 if (size == 0)
1331 return (0);
1332
1333 err = dnode_hold(os, object, FTAG, &dn);
1334 if (err)
1335 return (err);
1336
1337 err = dmu_read_uio_dnode(dn, uio, size);
1338
1339 dnode_rele(dn, FTAG);
1340
1341 return (err);
1342 }
1343
1344 int
dmu_write_uio_dnode(dnode_t * dn,zfs_uio_t * uio,uint64_t size,dmu_tx_t * tx)1345 dmu_write_uio_dnode(dnode_t *dn, zfs_uio_t *uio, uint64_t size, dmu_tx_t *tx)
1346 {
1347 dmu_buf_t **dbp;
1348 int numbufs;
1349 int err = 0;
1350 int i;
1351
1352 err = dmu_buf_hold_array_by_dnode(dn, zfs_uio_offset(uio), size,
1353 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1354 if (err)
1355 return (err);
1356
1357 for (i = 0; i < numbufs; i++) {
1358 uint64_t tocpy;
1359 int64_t bufoff;
1360 dmu_buf_t *db = dbp[i];
1361
1362 ASSERT(size > 0);
1363
1364 offset_t off = zfs_uio_offset(uio);
1365 bufoff = off - db->db_offset;
1366 tocpy = MIN(db->db_size - bufoff, size);
1367
1368 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1369
1370 if (tocpy == db->db_size)
1371 dmu_buf_will_fill(db, tx, B_TRUE);
1372 else
1373 dmu_buf_will_dirty(db, tx);
1374
1375 err = zfs_uio_fault_move((char *)db->db_data + bufoff,
1376 tocpy, UIO_WRITE, uio);
1377
1378 if (tocpy == db->db_size && dmu_buf_fill_done(db, tx, err)) {
1379 /* The fill was reverted. Undo any uio progress. */
1380 zfs_uio_advance(uio, off - zfs_uio_offset(uio));
1381 }
1382
1383 if (err)
1384 break;
1385
1386 size -= tocpy;
1387 }
1388
1389 dmu_buf_rele_array(dbp, numbufs, FTAG);
1390 return (err);
1391 }
1392
1393 /*
1394 * Write 'size' bytes from the uio buffer.
1395 * To object zdb->db_object.
1396 * Starting at offset zfs_uio_offset(uio).
1397 *
1398 * If the caller already has a dbuf in the target object
1399 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1400 * because we don't have to find the dnode_t for the object.
1401 */
1402 int
dmu_write_uio_dbuf(dmu_buf_t * zdb,zfs_uio_t * uio,uint64_t size,dmu_tx_t * tx)1403 dmu_write_uio_dbuf(dmu_buf_t *zdb, zfs_uio_t *uio, uint64_t size,
1404 dmu_tx_t *tx)
1405 {
1406 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1407 int err;
1408
1409 if (size == 0)
1410 return (0);
1411
1412 DB_DNODE_ENTER(db);
1413 err = dmu_write_uio_dnode(DB_DNODE(db), uio, size, tx);
1414 DB_DNODE_EXIT(db);
1415
1416 return (err);
1417 }
1418
1419 /*
1420 * Write 'size' bytes from the uio buffer.
1421 * To the specified object.
1422 * Starting at offset zfs_uio_offset(uio).
1423 */
1424 int
dmu_write_uio(objset_t * os,uint64_t object,zfs_uio_t * uio,uint64_t size,dmu_tx_t * tx)1425 dmu_write_uio(objset_t *os, uint64_t object, zfs_uio_t *uio, uint64_t size,
1426 dmu_tx_t *tx)
1427 {
1428 dnode_t *dn;
1429 int err;
1430
1431 if (size == 0)
1432 return (0);
1433
1434 err = dnode_hold(os, object, FTAG, &dn);
1435 if (err)
1436 return (err);
1437
1438 err = dmu_write_uio_dnode(dn, uio, size, tx);
1439
1440 dnode_rele(dn, FTAG);
1441
1442 return (err);
1443 }
1444 #endif /* _KERNEL */
1445
1446 /*
1447 * Allocate a loaned anonymous arc buffer.
1448 */
1449 arc_buf_t *
dmu_request_arcbuf(dmu_buf_t * handle,int size)1450 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1451 {
1452 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1453
1454 return (arc_loan_buf(db->db_objset->os_spa, B_FALSE, size));
1455 }
1456
1457 /*
1458 * Free a loaned arc buffer.
1459 */
1460 void
dmu_return_arcbuf(arc_buf_t * buf)1461 dmu_return_arcbuf(arc_buf_t *buf)
1462 {
1463 arc_return_buf(buf, FTAG);
1464 arc_buf_destroy(buf, FTAG);
1465 }
1466
1467 /*
1468 * A "lightweight" write is faster than a regular write (e.g.
1469 * dmu_write_by_dnode() or dmu_assign_arcbuf_by_dnode()), because it avoids the
1470 * CPU cost of creating a dmu_buf_impl_t and arc_buf_[hdr_]_t. However, the
1471 * data can not be read or overwritten until the transaction's txg has been
1472 * synced. This makes it appropriate for workloads that are known to be
1473 * (temporarily) write-only, like "zfs receive".
1474 *
1475 * A single block is written, starting at the specified offset in bytes. If
1476 * the call is successful, it returns 0 and the provided abd has been
1477 * consumed (the caller should not free it).
1478 */
1479 int
dmu_lightweight_write_by_dnode(dnode_t * dn,uint64_t offset,abd_t * abd,const zio_prop_t * zp,zio_flag_t flags,dmu_tx_t * tx)1480 dmu_lightweight_write_by_dnode(dnode_t *dn, uint64_t offset, abd_t *abd,
1481 const zio_prop_t *zp, zio_flag_t flags, dmu_tx_t *tx)
1482 {
1483 dbuf_dirty_record_t *dr =
1484 dbuf_dirty_lightweight(dn, dbuf_whichblock(dn, 0, offset), tx);
1485 if (dr == NULL)
1486 return (SET_ERROR(EIO));
1487 dr->dt.dll.dr_abd = abd;
1488 dr->dt.dll.dr_props = *zp;
1489 dr->dt.dll.dr_flags = flags;
1490 return (0);
1491 }
1492
1493 /*
1494 * When possible directly assign passed loaned arc buffer to a dbuf.
1495 * If this is not possible copy the contents of passed arc buf via
1496 * dmu_write().
1497 */
1498 int
dmu_assign_arcbuf_by_dnode(dnode_t * dn,uint64_t offset,arc_buf_t * buf,dmu_tx_t * tx)1499 dmu_assign_arcbuf_by_dnode(dnode_t *dn, uint64_t offset, arc_buf_t *buf,
1500 dmu_tx_t *tx)
1501 {
1502 dmu_buf_impl_t *db;
1503 objset_t *os = dn->dn_objset;
1504 uint64_t object = dn->dn_object;
1505 uint32_t blksz = (uint32_t)arc_buf_lsize(buf);
1506 uint64_t blkid;
1507
1508 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1509 blkid = dbuf_whichblock(dn, 0, offset);
1510 db = dbuf_hold(dn, blkid, FTAG);
1511 rw_exit(&dn->dn_struct_rwlock);
1512 if (db == NULL)
1513 return (SET_ERROR(EIO));
1514
1515 /*
1516 * We can only assign if the offset is aligned and the arc buf is the
1517 * same size as the dbuf.
1518 */
1519 if (offset == db->db.db_offset && blksz == db->db.db_size) {
1520 zfs_racct_write(blksz, 1);
1521 dbuf_assign_arcbuf(db, buf, tx);
1522 dbuf_rele(db, FTAG);
1523 } else {
1524 /* compressed bufs must always be assignable to their dbuf */
1525 ASSERT3U(arc_get_compression(buf), ==, ZIO_COMPRESS_OFF);
1526 ASSERT(!(buf->b_flags & ARC_BUF_FLAG_COMPRESSED));
1527
1528 dbuf_rele(db, FTAG);
1529 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1530 dmu_return_arcbuf(buf);
1531 }
1532
1533 return (0);
1534 }
1535
1536 int
dmu_assign_arcbuf_by_dbuf(dmu_buf_t * handle,uint64_t offset,arc_buf_t * buf,dmu_tx_t * tx)1537 dmu_assign_arcbuf_by_dbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1538 dmu_tx_t *tx)
1539 {
1540 int err;
1541 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1542
1543 DB_DNODE_ENTER(db);
1544 err = dmu_assign_arcbuf_by_dnode(DB_DNODE(db), offset, buf, tx);
1545 DB_DNODE_EXIT(db);
1546
1547 return (err);
1548 }
1549
1550 typedef struct {
1551 dbuf_dirty_record_t *dsa_dr;
1552 dmu_sync_cb_t *dsa_done;
1553 zgd_t *dsa_zgd;
1554 dmu_tx_t *dsa_tx;
1555 } dmu_sync_arg_t;
1556
1557 static void
dmu_sync_ready(zio_t * zio,arc_buf_t * buf,void * varg)1558 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1559 {
1560 (void) buf;
1561 dmu_sync_arg_t *dsa = varg;
1562 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1563 blkptr_t *bp = zio->io_bp;
1564
1565 if (zio->io_error == 0) {
1566 if (BP_IS_HOLE(bp)) {
1567 /*
1568 * A block of zeros may compress to a hole, but the
1569 * block size still needs to be known for replay.
1570 */
1571 BP_SET_LSIZE(bp, db->db_size);
1572 } else if (!BP_IS_EMBEDDED(bp)) {
1573 ASSERT(BP_GET_LEVEL(bp) == 0);
1574 BP_SET_FILL(bp, 1);
1575 }
1576 }
1577 }
1578
1579 static void
dmu_sync_late_arrival_ready(zio_t * zio)1580 dmu_sync_late_arrival_ready(zio_t *zio)
1581 {
1582 dmu_sync_ready(zio, NULL, zio->io_private);
1583 }
1584
1585 static void
dmu_sync_done(zio_t * zio,arc_buf_t * buf,void * varg)1586 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1587 {
1588 (void) buf;
1589 dmu_sync_arg_t *dsa = varg;
1590 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1591 dmu_buf_impl_t *db = dr->dr_dbuf;
1592 zgd_t *zgd = dsa->dsa_zgd;
1593
1594 /*
1595 * Record the vdev(s) backing this blkptr so they can be flushed after
1596 * the writes for the lwb have completed.
1597 */
1598 if (zio->io_error == 0) {
1599 zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp);
1600 }
1601
1602 mutex_enter(&db->db_mtx);
1603 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1604 if (zio->io_error == 0) {
1605 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1606 if (dr->dt.dl.dr_nopwrite) {
1607 blkptr_t *bp = zio->io_bp;
1608 blkptr_t *bp_orig = &zio->io_bp_orig;
1609 uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1610
1611 ASSERT(BP_EQUAL(bp, bp_orig));
1612 VERIFY(BP_EQUAL(bp, db->db_blkptr));
1613 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1614 VERIFY(zio_checksum_table[chksum].ci_flags &
1615 ZCHECKSUM_FLAG_NOPWRITE);
1616 }
1617 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1618 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1619 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1620
1621 /*
1622 * Old style holes are filled with all zeros, whereas
1623 * new-style holes maintain their lsize, type, level,
1624 * and birth time (see zio_write_compress). While we
1625 * need to reset the BP_SET_LSIZE() call that happened
1626 * in dmu_sync_ready for old style holes, we do *not*
1627 * want to wipe out the information contained in new
1628 * style holes. Thus, only zero out the block pointer if
1629 * it's an old style hole.
1630 */
1631 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) &&
1632 dr->dt.dl.dr_overridden_by.blk_birth == 0)
1633 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1634 } else {
1635 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1636 }
1637 cv_broadcast(&db->db_changed);
1638 mutex_exit(&db->db_mtx);
1639
1640 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1641
1642 kmem_free(dsa, sizeof (*dsa));
1643 }
1644
1645 static void
dmu_sync_late_arrival_done(zio_t * zio)1646 dmu_sync_late_arrival_done(zio_t *zio)
1647 {
1648 blkptr_t *bp = zio->io_bp;
1649 dmu_sync_arg_t *dsa = zio->io_private;
1650 zgd_t *zgd = dsa->dsa_zgd;
1651
1652 if (zio->io_error == 0) {
1653 /*
1654 * Record the vdev(s) backing this blkptr so they can be
1655 * flushed after the writes for the lwb have completed.
1656 */
1657 zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp);
1658
1659 if (!BP_IS_HOLE(bp)) {
1660 blkptr_t *bp_orig __maybe_unused = &zio->io_bp_orig;
1661 ASSERT(!(zio->io_flags & ZIO_FLAG_NOPWRITE));
1662 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1663 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1664 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1665 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1666 }
1667 }
1668
1669 dmu_tx_commit(dsa->dsa_tx);
1670
1671 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1672
1673 abd_free(zio->io_abd);
1674 kmem_free(dsa, sizeof (*dsa));
1675 }
1676
1677 static int
dmu_sync_late_arrival(zio_t * pio,objset_t * os,dmu_sync_cb_t * done,zgd_t * zgd,zio_prop_t * zp,zbookmark_phys_t * zb)1678 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1679 zio_prop_t *zp, zbookmark_phys_t *zb)
1680 {
1681 dmu_sync_arg_t *dsa;
1682 dmu_tx_t *tx;
1683 int error;
1684
1685 error = dbuf_read((dmu_buf_impl_t *)zgd->zgd_db, NULL,
1686 DB_RF_CANFAIL | DB_RF_NOPREFETCH);
1687 if (error != 0)
1688 return (error);
1689
1690 tx = dmu_tx_create(os);
1691 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1692 /*
1693 * This transaction does not produce any dirty data or log blocks, so
1694 * it should not be throttled. All other cases wait for TXG sync, by
1695 * which time the log block we are writing will be obsolete, so we can
1696 * skip waiting and just return error here instead.
1697 */
1698 if (dmu_tx_assign(tx, TXG_NOWAIT | TXG_NOTHROTTLE) != 0) {
1699 dmu_tx_abort(tx);
1700 /* Make zl_get_data do txg_waited_synced() */
1701 return (SET_ERROR(EIO));
1702 }
1703
1704 /*
1705 * In order to prevent the zgd's lwb from being free'd prior to
1706 * dmu_sync_late_arrival_done() being called, we have to ensure
1707 * the lwb's "max txg" takes this tx's txg into account.
1708 */
1709 zil_lwb_add_txg(zgd->zgd_lwb, dmu_tx_get_txg(tx));
1710
1711 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1712 dsa->dsa_dr = NULL;
1713 dsa->dsa_done = done;
1714 dsa->dsa_zgd = zgd;
1715 dsa->dsa_tx = tx;
1716
1717 /*
1718 * Since we are currently syncing this txg, it's nontrivial to
1719 * determine what BP to nopwrite against, so we disable nopwrite.
1720 *
1721 * When syncing, the db_blkptr is initially the BP of the previous
1722 * txg. We can not nopwrite against it because it will be changed
1723 * (this is similar to the non-late-arrival case where the dbuf is
1724 * dirty in a future txg).
1725 *
1726 * Then dbuf_write_ready() sets bp_blkptr to the location we will write.
1727 * We can not nopwrite against it because although the BP will not
1728 * (typically) be changed, the data has not yet been persisted to this
1729 * location.
1730 *
1731 * Finally, when dbuf_write_done() is called, it is theoretically
1732 * possible to always nopwrite, because the data that was written in
1733 * this txg is the same data that we are trying to write. However we
1734 * would need to check that this dbuf is not dirty in any future
1735 * txg's (as we do in the normal dmu_sync() path). For simplicity, we
1736 * don't nopwrite in this case.
1737 */
1738 zp->zp_nopwrite = B_FALSE;
1739
1740 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1741 abd_get_from_buf(zgd->zgd_db->db_data, zgd->zgd_db->db_size),
1742 zgd->zgd_db->db_size, zgd->zgd_db->db_size, zp,
1743 dmu_sync_late_arrival_ready, NULL, dmu_sync_late_arrival_done,
1744 dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1745
1746 return (0);
1747 }
1748
1749 /*
1750 * Intent log support: sync the block associated with db to disk.
1751 * N.B. and XXX: the caller is responsible for making sure that the
1752 * data isn't changing while dmu_sync() is writing it.
1753 *
1754 * Return values:
1755 *
1756 * EEXIST: this txg has already been synced, so there's nothing to do.
1757 * The caller should not log the write.
1758 *
1759 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1760 * The caller should not log the write.
1761 *
1762 * EALREADY: this block is already in the process of being synced.
1763 * The caller should track its progress (somehow).
1764 *
1765 * EIO: could not do the I/O.
1766 * The caller should do a txg_wait_synced().
1767 *
1768 * 0: the I/O has been initiated.
1769 * The caller should log this blkptr in the done callback.
1770 * It is possible that the I/O will fail, in which case
1771 * the error will be reported to the done callback and
1772 * propagated to pio from zio_done().
1773 */
1774 int
dmu_sync(zio_t * pio,uint64_t txg,dmu_sync_cb_t * done,zgd_t * zgd)1775 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1776 {
1777 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1778 objset_t *os = db->db_objset;
1779 dsl_dataset_t *ds = os->os_dsl_dataset;
1780 dbuf_dirty_record_t *dr, *dr_next;
1781 dmu_sync_arg_t *dsa;
1782 zbookmark_phys_t zb;
1783 zio_prop_t zp;
1784
1785 ASSERT(pio != NULL);
1786 ASSERT(txg != 0);
1787
1788 SET_BOOKMARK(&zb, ds->ds_object,
1789 db->db.db_object, db->db_level, db->db_blkid);
1790
1791 DB_DNODE_ENTER(db);
1792 dmu_write_policy(os, DB_DNODE(db), db->db_level, WP_DMU_SYNC, &zp);
1793 DB_DNODE_EXIT(db);
1794
1795 /*
1796 * If we're frozen (running ziltest), we always need to generate a bp.
1797 */
1798 if (txg > spa_freeze_txg(os->os_spa))
1799 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1800
1801 /*
1802 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1803 * and us. If we determine that this txg is not yet syncing,
1804 * but it begins to sync a moment later, that's OK because the
1805 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1806 */
1807 mutex_enter(&db->db_mtx);
1808
1809 if (txg <= spa_last_synced_txg(os->os_spa)) {
1810 /*
1811 * This txg has already synced. There's nothing to do.
1812 */
1813 mutex_exit(&db->db_mtx);
1814 return (SET_ERROR(EEXIST));
1815 }
1816
1817 if (txg <= spa_syncing_txg(os->os_spa)) {
1818 /*
1819 * This txg is currently syncing, so we can't mess with
1820 * the dirty record anymore; just write a new log block.
1821 */
1822 mutex_exit(&db->db_mtx);
1823 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1824 }
1825
1826 dr = dbuf_find_dirty_eq(db, txg);
1827
1828 if (dr == NULL) {
1829 /*
1830 * There's no dr for this dbuf, so it must have been freed.
1831 * There's no need to log writes to freed blocks, so we're done.
1832 */
1833 mutex_exit(&db->db_mtx);
1834 return (SET_ERROR(ENOENT));
1835 }
1836
1837 dr_next = list_next(&db->db_dirty_records, dr);
1838 ASSERT(dr_next == NULL || dr_next->dr_txg < txg);
1839
1840 if (db->db_blkptr != NULL) {
1841 /*
1842 * We need to fill in zgd_bp with the current blkptr so that
1843 * the nopwrite code can check if we're writing the same
1844 * data that's already on disk. We can only nopwrite if we
1845 * are sure that after making the copy, db_blkptr will not
1846 * change until our i/o completes. We ensure this by
1847 * holding the db_mtx, and only allowing nopwrite if the
1848 * block is not already dirty (see below). This is verified
1849 * by dmu_sync_done(), which VERIFYs that the db_blkptr has
1850 * not changed.
1851 */
1852 *zgd->zgd_bp = *db->db_blkptr;
1853 }
1854
1855 /*
1856 * Assume the on-disk data is X, the current syncing data (in
1857 * txg - 1) is Y, and the current in-memory data is Z (currently
1858 * in dmu_sync).
1859 *
1860 * We usually want to perform a nopwrite if X and Z are the
1861 * same. However, if Y is different (i.e. the BP is going to
1862 * change before this write takes effect), then a nopwrite will
1863 * be incorrect - we would override with X, which could have
1864 * been freed when Y was written.
1865 *
1866 * (Note that this is not a concern when we are nop-writing from
1867 * syncing context, because X and Y must be identical, because
1868 * all previous txgs have been synced.)
1869 *
1870 * Therefore, we disable nopwrite if the current BP could change
1871 * before this TXG. There are two ways it could change: by
1872 * being dirty (dr_next is non-NULL), or by being freed
1873 * (dnode_block_freed()). This behavior is verified by
1874 * zio_done(), which VERIFYs that the override BP is identical
1875 * to the on-disk BP.
1876 */
1877 if (dr_next != NULL) {
1878 zp.zp_nopwrite = B_FALSE;
1879 } else {
1880 DB_DNODE_ENTER(db);
1881 if (dnode_block_freed(DB_DNODE(db), db->db_blkid))
1882 zp.zp_nopwrite = B_FALSE;
1883 DB_DNODE_EXIT(db);
1884 }
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
dmu_object_set_nlevels(objset_t * os,uint64_t object,int nlevels,dmu_tx_t * tx)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
dmu_object_set_blocksize(objset_t * os,uint64_t object,uint64_t size,int ibs,dmu_tx_t * tx)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
dmu_object_set_maxblkid(objset_t * os,uint64_t object,uint64_t maxblkid,dmu_tx_t * tx)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
dmu_object_set_checksum(objset_t * os,uint64_t object,uint8_t checksum,dmu_tx_t * tx)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
dmu_object_set_compress(objset_t * os,uint64_t object,uint8_t compress,dmu_tx_t * tx)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
dmu_write_policy(objset_t * os,dnode_t * dn,int level,int wp,zio_prop_t * zp)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
dmu_offset_next(objset_t * os,uint64_t object,boolean_t hole,uint64_t * off)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
dmu_read_l0_bps(objset_t * os,uint64_t object,uint64_t offset,uint64_t length,blkptr_t * bps,size_t * nbpsp)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_PHYSICAL_BIRTH(bp) > spa_freeze_txg(os->os_spa)) {
2294 error = SET_ERROR(EINVAL);
2295 goto out;
2296 }
2297 if (BP_PHYSICAL_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
dmu_brt_clone(objset_t * os,uint64_t object,uint64_t offset,uint64_t length,dmu_tx_t * tx,const blkptr_t * bps,size_t nbps)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->blk_birth != 0) {
2370 if (!BP_IS_EMBEDDED(bp)) {
2371 BP_SET_BIRTH(&dl->dr_overridden_by, dr->dr_txg,
2372 BP_PHYSICAL_BIRTH(bp));
2373 } else {
2374 dl->dr_overridden_by.blk_birth = dr->dr_txg;
2375 }
2376 }
2377 dl->dr_brtwrite = B_TRUE;
2378 dl->dr_override_state = DR_OVERRIDDEN;
2379
2380 mutex_exit(&db->db_mtx);
2381
2382 /*
2383 * When data in embedded into BP there is no need to create
2384 * BRT entry as there is no data block. Just copy the BP as
2385 * it contains the data.
2386 */
2387 if (!BP_IS_HOLE(bp) && !BP_IS_EMBEDDED(bp)) {
2388 brt_pending_add(spa, bp, tx);
2389 }
2390 }
2391 out:
2392 dmu_buf_rele_array(dbp, numbufs, FTAG);
2393
2394 return (error);
2395 }
2396
2397 void
__dmu_object_info_from_dnode(dnode_t * dn,dmu_object_info_t * doi)2398 __dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2399 {
2400 dnode_phys_t *dnp = dn->dn_phys;
2401
2402 doi->doi_data_block_size = dn->dn_datablksz;
2403 doi->doi_metadata_block_size = dn->dn_indblkshift ?
2404 1ULL << dn->dn_indblkshift : 0;
2405 doi->doi_type = dn->dn_type;
2406 doi->doi_bonus_type = dn->dn_bonustype;
2407 doi->doi_bonus_size = dn->dn_bonuslen;
2408 doi->doi_dnodesize = dn->dn_num_slots << DNODE_SHIFT;
2409 doi->doi_indirection = dn->dn_nlevels;
2410 doi->doi_checksum = dn->dn_checksum;
2411 doi->doi_compress = dn->dn_compress;
2412 doi->doi_nblkptr = dn->dn_nblkptr;
2413 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
2414 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
2415 doi->doi_fill_count = 0;
2416 for (int i = 0; i < dnp->dn_nblkptr; i++)
2417 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
2418 }
2419
2420 void
dmu_object_info_from_dnode(dnode_t * dn,dmu_object_info_t * doi)2421 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2422 {
2423 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2424 mutex_enter(&dn->dn_mtx);
2425
2426 __dmu_object_info_from_dnode(dn, doi);
2427
2428 mutex_exit(&dn->dn_mtx);
2429 rw_exit(&dn->dn_struct_rwlock);
2430 }
2431
2432 /*
2433 * Get information on a DMU object.
2434 * If doi is NULL, just indicates whether the object exists.
2435 */
2436 int
dmu_object_info(objset_t * os,uint64_t object,dmu_object_info_t * doi)2437 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
2438 {
2439 dnode_t *dn;
2440 int err = dnode_hold(os, object, FTAG, &dn);
2441
2442 if (err)
2443 return (err);
2444
2445 if (doi != NULL)
2446 dmu_object_info_from_dnode(dn, doi);
2447
2448 dnode_rele(dn, FTAG);
2449 return (0);
2450 }
2451
2452 /*
2453 * As above, but faster; can be used when you have a held dbuf in hand.
2454 */
2455 void
dmu_object_info_from_db(dmu_buf_t * db_fake,dmu_object_info_t * doi)2456 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
2457 {
2458 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2459
2460 DB_DNODE_ENTER(db);
2461 dmu_object_info_from_dnode(DB_DNODE(db), doi);
2462 DB_DNODE_EXIT(db);
2463 }
2464
2465 /*
2466 * Faster still when you only care about the size.
2467 */
2468 void
dmu_object_size_from_db(dmu_buf_t * db_fake,uint32_t * blksize,u_longlong_t * nblk512)2469 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
2470 u_longlong_t *nblk512)
2471 {
2472 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2473 dnode_t *dn;
2474
2475 DB_DNODE_ENTER(db);
2476 dn = DB_DNODE(db);
2477
2478 *blksize = dn->dn_datablksz;
2479 /* add in number of slots used for the dnode itself */
2480 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
2481 SPA_MINBLOCKSHIFT) + dn->dn_num_slots;
2482 DB_DNODE_EXIT(db);
2483 }
2484
2485 void
dmu_object_dnsize_from_db(dmu_buf_t * db_fake,int * dnsize)2486 dmu_object_dnsize_from_db(dmu_buf_t *db_fake, int *dnsize)
2487 {
2488 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2489
2490 DB_DNODE_ENTER(db);
2491 *dnsize = DB_DNODE(db)->dn_num_slots << DNODE_SHIFT;
2492 DB_DNODE_EXIT(db);
2493 }
2494
2495 void
byteswap_uint64_array(void * vbuf,size_t size)2496 byteswap_uint64_array(void *vbuf, size_t size)
2497 {
2498 uint64_t *buf = vbuf;
2499 size_t count = size >> 3;
2500 int i;
2501
2502 ASSERT((size & 7) == 0);
2503
2504 for (i = 0; i < count; i++)
2505 buf[i] = BSWAP_64(buf[i]);
2506 }
2507
2508 void
byteswap_uint32_array(void * vbuf,size_t size)2509 byteswap_uint32_array(void *vbuf, size_t size)
2510 {
2511 uint32_t *buf = vbuf;
2512 size_t count = size >> 2;
2513 int i;
2514
2515 ASSERT((size & 3) == 0);
2516
2517 for (i = 0; i < count; i++)
2518 buf[i] = BSWAP_32(buf[i]);
2519 }
2520
2521 void
byteswap_uint16_array(void * vbuf,size_t size)2522 byteswap_uint16_array(void *vbuf, size_t size)
2523 {
2524 uint16_t *buf = vbuf;
2525 size_t count = size >> 1;
2526 int i;
2527
2528 ASSERT((size & 1) == 0);
2529
2530 for (i = 0; i < count; i++)
2531 buf[i] = BSWAP_16(buf[i]);
2532 }
2533
2534 void
byteswap_uint8_array(void * vbuf,size_t size)2535 byteswap_uint8_array(void *vbuf, size_t size)
2536 {
2537 (void) vbuf, (void) size;
2538 }
2539
2540 void
dmu_init(void)2541 dmu_init(void)
2542 {
2543 abd_init();
2544 zfs_dbgmsg_init();
2545 sa_cache_init();
2546 dmu_objset_init();
2547 dnode_init();
2548 zfetch_init();
2549 dmu_tx_init();
2550 l2arc_init();
2551 arc_init();
2552 dbuf_init();
2553 }
2554
2555 void
dmu_fini(void)2556 dmu_fini(void)
2557 {
2558 arc_fini(); /* arc depends on l2arc, so arc must go first */
2559 l2arc_fini();
2560 dmu_tx_fini();
2561 zfetch_fini();
2562 dbuf_fini();
2563 dnode_fini();
2564 dmu_objset_fini();
2565 sa_cache_fini();
2566 zfs_dbgmsg_fini();
2567 abd_fini();
2568 }
2569
2570 EXPORT_SYMBOL(dmu_bonus_hold);
2571 EXPORT_SYMBOL(dmu_bonus_hold_by_dnode);
2572 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus);
2573 EXPORT_SYMBOL(dmu_buf_rele_array);
2574 EXPORT_SYMBOL(dmu_prefetch);
2575 EXPORT_SYMBOL(dmu_prefetch_by_dnode);
2576 EXPORT_SYMBOL(dmu_prefetch_dnode);
2577 EXPORT_SYMBOL(dmu_free_range);
2578 EXPORT_SYMBOL(dmu_free_long_range);
2579 EXPORT_SYMBOL(dmu_free_long_object);
2580 EXPORT_SYMBOL(dmu_read);
2581 EXPORT_SYMBOL(dmu_read_by_dnode);
2582 EXPORT_SYMBOL(dmu_write);
2583 EXPORT_SYMBOL(dmu_write_by_dnode);
2584 EXPORT_SYMBOL(dmu_prealloc);
2585 EXPORT_SYMBOL(dmu_object_info);
2586 EXPORT_SYMBOL(dmu_object_info_from_dnode);
2587 EXPORT_SYMBOL(dmu_object_info_from_db);
2588 EXPORT_SYMBOL(dmu_object_size_from_db);
2589 EXPORT_SYMBOL(dmu_object_dnsize_from_db);
2590 EXPORT_SYMBOL(dmu_object_set_nlevels);
2591 EXPORT_SYMBOL(dmu_object_set_blocksize);
2592 EXPORT_SYMBOL(dmu_object_set_maxblkid);
2593 EXPORT_SYMBOL(dmu_object_set_checksum);
2594 EXPORT_SYMBOL(dmu_object_set_compress);
2595 EXPORT_SYMBOL(dmu_offset_next);
2596 EXPORT_SYMBOL(dmu_write_policy);
2597 EXPORT_SYMBOL(dmu_sync);
2598 EXPORT_SYMBOL(dmu_request_arcbuf);
2599 EXPORT_SYMBOL(dmu_return_arcbuf);
2600 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dnode);
2601 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dbuf);
2602 EXPORT_SYMBOL(dmu_buf_hold);
2603 EXPORT_SYMBOL(dmu_ot);
2604
2605 ZFS_MODULE_PARAM(zfs, zfs_, nopwrite_enabled, INT, ZMOD_RW,
2606 "Enable NOP writes");
2607
2608 ZFS_MODULE_PARAM(zfs, zfs_, per_txg_dirty_frees_percent, UINT, ZMOD_RW,
2609 "Percentage of dirtied blocks from frees in one TXG");
2610
2611 ZFS_MODULE_PARAM(zfs, zfs_, dmu_offset_next_sync, INT, ZMOD_RW,
2612 "Enable forcing txg sync to find holes");
2613
2614 /* CSTYLED */
2615 ZFS_MODULE_PARAM(zfs, , dmu_prefetch_max, UINT, ZMOD_RW,
2616 "Limit one prefetch call to this size");
2617