]> 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
Extending FreeBSD UIO Struct
[mirror_zfs.git] / module / zfs / dmu.c
1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 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 */
32
33 #include <sys/dmu.h>
34 #include <sys/dmu_impl.h>
35 #include <sys/dmu_tx.h>
36 #include <sys/dbuf.h>
37 #include <sys/dnode.h>
38 #include <sys/zfs_context.h>
39 #include <sys/dmu_objset.h>
40 #include <sys/dmu_traverse.h>
41 #include <sys/dsl_dataset.h>
42 #include <sys/dsl_dir.h>
43 #include <sys/dsl_pool.h>
44 #include <sys/dsl_synctask.h>
45 #include <sys/dsl_prop.h>
46 #include <sys/dmu_zfetch.h>
47 #include <sys/zfs_ioctl.h>
48 #include <sys/zap.h>
49 #include <sys/zio_checksum.h>
50 #include <sys/zio_compress.h>
51 #include <sys/sa.h>
52 #include <sys/zfeature.h>
53 #include <sys/abd.h>
54 #include <sys/trace_zfs.h>
55 #include <sys/zfs_rlock.h>
56 #ifdef _KERNEL
57 #include <sys/vmsystm.h>
58 #include <sys/zfs_znode.h>
59 #endif
60
61 /*
62 * Enable/disable nopwrite feature.
63 */
64 int zfs_nopwrite_enabled = 1;
65
66 /*
67 * Tunable to control percentage of dirtied L1 blocks from frees allowed into
68 * one TXG. After this threshold is crossed, additional dirty blocks from frees
69 * will wait until the next TXG.
70 * A value of zero will disable this throttle.
71 */
72 unsigned long zfs_per_txg_dirty_frees_percent = 5;
73
74 /*
75 * Enable/disable forcing txg sync when dirty in dmu_offset_next.
76 */
77 int zfs_dmu_offset_next_sync = 0;
78
79 /*
80 * Limit the amount we can prefetch with one call to this amount. This
81 * helps to limit the amount of memory that can be used by prefetching.
82 * Larger objects should be prefetched a bit at a time.
83 */
84 int dmu_prefetch_max = 8 * SPA_MAXBLOCKSIZE;
85
86 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
87 {DMU_BSWAP_UINT8, TRUE, FALSE, FALSE, "unallocated" },
88 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "object directory" },
89 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "object array" },
90 {DMU_BSWAP_UINT8, TRUE, FALSE, FALSE, "packed nvlist" },
91 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "packed nvlist size" },
92 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "bpobj" },
93 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "bpobj header" },
94 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "SPA space map header" },
95 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "SPA space map" },
96 {DMU_BSWAP_UINT64, TRUE, FALSE, TRUE, "ZIL intent log" },
97 {DMU_BSWAP_DNODE, TRUE, FALSE, TRUE, "DMU dnode" },
98 {DMU_BSWAP_OBJSET, TRUE, TRUE, FALSE, "DMU objset" },
99 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "DSL directory" },
100 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL directory child map"},
101 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL dataset snap map" },
102 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL props" },
103 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "DSL dataset" },
104 {DMU_BSWAP_ZNODE, TRUE, FALSE, FALSE, "ZFS znode" },
105 {DMU_BSWAP_OLDACL, TRUE, FALSE, TRUE, "ZFS V0 ACL" },
106 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "ZFS plain file" },
107 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS directory" },
108 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "ZFS master node" },
109 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS delete queue" },
110 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "zvol object" },
111 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "zvol prop" },
112 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "other uint8[]" },
113 {DMU_BSWAP_UINT64, FALSE, FALSE, TRUE, "other uint64[]" },
114 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "other ZAP" },
115 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "persistent error log" },
116 {DMU_BSWAP_UINT8, TRUE, FALSE, FALSE, "SPA history" },
117 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "SPA history offsets" },
118 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "Pool properties" },
119 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL permissions" },
120 {DMU_BSWAP_ACL, TRUE, FALSE, TRUE, "ZFS ACL" },
121 {DMU_BSWAP_UINT8, TRUE, FALSE, TRUE, "ZFS SYSACL" },
122 {DMU_BSWAP_UINT8, TRUE, FALSE, TRUE, "FUID table" },
123 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "FUID table size" },
124 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL dataset next clones"},
125 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "scan work queue" },
126 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS user/group/project used" },
127 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS user/group/project quota"},
128 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "snapshot refcount tags"},
129 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "DDT ZAP algorithm" },
130 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "DDT statistics" },
131 {DMU_BSWAP_UINT8, TRUE, FALSE, TRUE, "System attributes" },
132 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "SA master node" },
133 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "SA attr registration" },
134 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "SA attr layouts" },
135 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "scan translations" },
136 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "deduplicated block" },
137 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL deadlist map" },
138 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "DSL deadlist map hdr" },
139 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL dir clones" },
140 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "bpobj subobj" }
141 };
142
143 const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
144 { byteswap_uint8_array, "uint8" },
145 { byteswap_uint16_array, "uint16" },
146 { byteswap_uint32_array, "uint32" },
147 { byteswap_uint64_array, "uint64" },
148 { zap_byteswap, "zap" },
149 { dnode_buf_byteswap, "dnode" },
150 { dmu_objset_byteswap, "objset" },
151 { zfs_znode_byteswap, "znode" },
152 { zfs_oldacl_byteswap, "oldacl" },
153 { zfs_acl_byteswap, "acl" }
154 };
155
156 static int
157 dmu_buf_hold_noread_by_dnode(dnode_t *dn, uint64_t offset,
158 void *tag, dmu_buf_t **dbp)
159 {
160 uint64_t blkid;
161 dmu_buf_impl_t *db;
162
163 rw_enter(&dn->dn_struct_rwlock, RW_READER);
164 blkid = dbuf_whichblock(dn, 0, offset);
165 db = dbuf_hold(dn, blkid, tag);
166 rw_exit(&dn->dn_struct_rwlock);
167
168 if (db == NULL) {
169 *dbp = NULL;
170 return (SET_ERROR(EIO));
171 }
172
173 *dbp = &db->db;
174 return (0);
175 }
176 int
177 dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset,
178 void *tag, dmu_buf_t **dbp)
179 {
180 dnode_t *dn;
181 uint64_t blkid;
182 dmu_buf_impl_t *db;
183 int err;
184
185 err = dnode_hold(os, object, FTAG, &dn);
186 if (err)
187 return (err);
188 rw_enter(&dn->dn_struct_rwlock, RW_READER);
189 blkid = dbuf_whichblock(dn, 0, offset);
190 db = dbuf_hold(dn, blkid, tag);
191 rw_exit(&dn->dn_struct_rwlock);
192 dnode_rele(dn, FTAG);
193
194 if (db == NULL) {
195 *dbp = NULL;
196 return (SET_ERROR(EIO));
197 }
198
199 *dbp = &db->db;
200 return (err);
201 }
202
203 int
204 dmu_buf_hold_by_dnode(dnode_t *dn, uint64_t offset,
205 void *tag, dmu_buf_t **dbp, int flags)
206 {
207 int err;
208 int db_flags = DB_RF_CANFAIL;
209
210 if (flags & DMU_READ_NO_PREFETCH)
211 db_flags |= DB_RF_NOPREFETCH;
212 if (flags & DMU_READ_NO_DECRYPT)
213 db_flags |= DB_RF_NO_DECRYPT;
214
215 err = dmu_buf_hold_noread_by_dnode(dn, offset, tag, dbp);
216 if (err == 0) {
217 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
218 err = dbuf_read(db, NULL, db_flags);
219 if (err != 0) {
220 dbuf_rele(db, tag);
221 *dbp = NULL;
222 }
223 }
224
225 return (err);
226 }
227
228 int
229 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
230 void *tag, dmu_buf_t **dbp, int flags)
231 {
232 int err;
233 int db_flags = DB_RF_CANFAIL;
234
235 if (flags & DMU_READ_NO_PREFETCH)
236 db_flags |= DB_RF_NOPREFETCH;
237 if (flags & DMU_READ_NO_DECRYPT)
238 db_flags |= DB_RF_NO_DECRYPT;
239
240 err = dmu_buf_hold_noread(os, object, offset, tag, dbp);
241 if (err == 0) {
242 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
243 err = dbuf_read(db, NULL, db_flags);
244 if (err != 0) {
245 dbuf_rele(db, tag);
246 *dbp = NULL;
247 }
248 }
249
250 return (err);
251 }
252
253 int
254 dmu_bonus_max(void)
255 {
256 return (DN_OLD_MAX_BONUSLEN);
257 }
258
259 int
260 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
261 {
262 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
263 dnode_t *dn;
264 int error;
265
266 DB_DNODE_ENTER(db);
267 dn = DB_DNODE(db);
268
269 if (dn->dn_bonus != db) {
270 error = SET_ERROR(EINVAL);
271 } else if (newsize < 0 || newsize > db_fake->db_size) {
272 error = SET_ERROR(EINVAL);
273 } else {
274 dnode_setbonuslen(dn, newsize, tx);
275 error = 0;
276 }
277
278 DB_DNODE_EXIT(db);
279 return (error);
280 }
281
282 int
283 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
284 {
285 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
286 dnode_t *dn;
287 int error;
288
289 DB_DNODE_ENTER(db);
290 dn = DB_DNODE(db);
291
292 if (!DMU_OT_IS_VALID(type)) {
293 error = SET_ERROR(EINVAL);
294 } else if (dn->dn_bonus != db) {
295 error = SET_ERROR(EINVAL);
296 } else {
297 dnode_setbonus_type(dn, type, tx);
298 error = 0;
299 }
300
301 DB_DNODE_EXIT(db);
302 return (error);
303 }
304
305 dmu_object_type_t
306 dmu_get_bonustype(dmu_buf_t *db_fake)
307 {
308 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
309 dnode_t *dn;
310 dmu_object_type_t type;
311
312 DB_DNODE_ENTER(db);
313 dn = DB_DNODE(db);
314 type = dn->dn_bonustype;
315 DB_DNODE_EXIT(db);
316
317 return (type);
318 }
319
320 int
321 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
322 {
323 dnode_t *dn;
324 int error;
325
326 error = dnode_hold(os, object, FTAG, &dn);
327 dbuf_rm_spill(dn, tx);
328 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
329 dnode_rm_spill(dn, tx);
330 rw_exit(&dn->dn_struct_rwlock);
331 dnode_rele(dn, FTAG);
332 return (error);
333 }
334
335 /*
336 * Lookup and hold the bonus buffer for the provided dnode. If the dnode
337 * has not yet been allocated a new bonus dbuf a will be allocated.
338 * Returns ENOENT, EIO, or 0.
339 */
340 int dmu_bonus_hold_by_dnode(dnode_t *dn, void *tag, dmu_buf_t **dbp,
341 uint32_t flags)
342 {
343 dmu_buf_impl_t *db;
344 int error;
345 uint32_t db_flags = DB_RF_MUST_SUCCEED;
346
347 if (flags & DMU_READ_NO_PREFETCH)
348 db_flags |= DB_RF_NOPREFETCH;
349 if (flags & DMU_READ_NO_DECRYPT)
350 db_flags |= DB_RF_NO_DECRYPT;
351
352 rw_enter(&dn->dn_struct_rwlock, RW_READER);
353 if (dn->dn_bonus == NULL) {
354 rw_exit(&dn->dn_struct_rwlock);
355 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
356 if (dn->dn_bonus == NULL)
357 dbuf_create_bonus(dn);
358 }
359 db = dn->dn_bonus;
360
361 /* as long as the bonus buf is held, the dnode will be held */
362 if (zfs_refcount_add(&db->db_holds, tag) == 1) {
363 VERIFY(dnode_add_ref(dn, db));
364 atomic_inc_32(&dn->dn_dbufs_count);
365 }
366
367 /*
368 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
369 * hold and incrementing the dbuf count to ensure that dnode_move() sees
370 * a dnode hold for every dbuf.
371 */
372 rw_exit(&dn->dn_struct_rwlock);
373
374 error = dbuf_read(db, NULL, db_flags);
375 if (error) {
376 dnode_evict_bonus(dn);
377 dbuf_rele(db, tag);
378 *dbp = NULL;
379 return (error);
380 }
381
382 *dbp = &db->db;
383 return (0);
384 }
385
386 int
387 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
388 {
389 dnode_t *dn;
390 int error;
391
392 error = dnode_hold(os, object, FTAG, &dn);
393 if (error)
394 return (error);
395
396 error = dmu_bonus_hold_by_dnode(dn, tag, dbp, DMU_READ_NO_PREFETCH);
397 dnode_rele(dn, FTAG);
398
399 return (error);
400 }
401
402 /*
403 * returns ENOENT, EIO, or 0.
404 *
405 * This interface will allocate a blank spill dbuf when a spill blk
406 * doesn't already exist on the dnode.
407 *
408 * if you only want to find an already existing spill db, then
409 * dmu_spill_hold_existing() should be used.
410 */
411 int
412 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
413 {
414 dmu_buf_impl_t *db = NULL;
415 int err;
416
417 if ((flags & DB_RF_HAVESTRUCT) == 0)
418 rw_enter(&dn->dn_struct_rwlock, RW_READER);
419
420 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
421
422 if ((flags & DB_RF_HAVESTRUCT) == 0)
423 rw_exit(&dn->dn_struct_rwlock);
424
425 if (db == NULL) {
426 *dbp = NULL;
427 return (SET_ERROR(EIO));
428 }
429 err = dbuf_read(db, NULL, flags);
430 if (err == 0)
431 *dbp = &db->db;
432 else {
433 dbuf_rele(db, tag);
434 *dbp = NULL;
435 }
436 return (err);
437 }
438
439 int
440 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
441 {
442 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
443 dnode_t *dn;
444 int err;
445
446 DB_DNODE_ENTER(db);
447 dn = DB_DNODE(db);
448
449 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
450 err = SET_ERROR(EINVAL);
451 } else {
452 rw_enter(&dn->dn_struct_rwlock, RW_READER);
453
454 if (!dn->dn_have_spill) {
455 err = SET_ERROR(ENOENT);
456 } else {
457 err = dmu_spill_hold_by_dnode(dn,
458 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
459 }
460
461 rw_exit(&dn->dn_struct_rwlock);
462 }
463
464 DB_DNODE_EXIT(db);
465 return (err);
466 }
467
468 int
469 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, uint32_t flags, void *tag,
470 dmu_buf_t **dbp)
471 {
472 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
473 dnode_t *dn;
474 int err;
475 uint32_t db_flags = DB_RF_CANFAIL;
476
477 if (flags & DMU_READ_NO_DECRYPT)
478 db_flags |= DB_RF_NO_DECRYPT;
479
480 DB_DNODE_ENTER(db);
481 dn = DB_DNODE(db);
482 err = dmu_spill_hold_by_dnode(dn, db_flags, tag, dbp);
483 DB_DNODE_EXIT(db);
484
485 return (err);
486 }
487
488 /*
489 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
490 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
491 * and can induce severe lock contention when writing to several files
492 * whose dnodes are in the same block.
493 */
494 int
495 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
496 boolean_t read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
497 {
498 dmu_buf_t **dbp;
499 uint64_t blkid, nblks, i;
500 uint32_t dbuf_flags;
501 int err;
502 zio_t *zio = NULL;
503
504 ASSERT(length <= DMU_MAX_ACCESS);
505
506 /*
507 * Note: We directly notify the prefetch code of this read, so that
508 * we can tell it about the multi-block read. dbuf_read() only knows
509 * about the one block it is accessing.
510 */
511 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT |
512 DB_RF_NOPREFETCH;
513
514 rw_enter(&dn->dn_struct_rwlock, RW_READER);
515 if (dn->dn_datablkshift) {
516 int blkshift = dn->dn_datablkshift;
517 nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) -
518 P2ALIGN(offset, 1ULL << blkshift)) >> blkshift;
519 } else {
520 if (offset + length > dn->dn_datablksz) {
521 zfs_panic_recover("zfs: accessing past end of object "
522 "%llx/%llx (size=%u access=%llu+%llu)",
523 (longlong_t)dn->dn_objset->
524 os_dsl_dataset->ds_object,
525 (longlong_t)dn->dn_object, dn->dn_datablksz,
526 (longlong_t)offset, (longlong_t)length);
527 rw_exit(&dn->dn_struct_rwlock);
528 return (SET_ERROR(EIO));
529 }
530 nblks = 1;
531 }
532 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
533
534 if (read)
535 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL,
536 ZIO_FLAG_CANFAIL);
537 blkid = dbuf_whichblock(dn, 0, offset);
538 for (i = 0; i < nblks; i++) {
539 dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag);
540 if (db == NULL) {
541 rw_exit(&dn->dn_struct_rwlock);
542 dmu_buf_rele_array(dbp, nblks, tag);
543 if (read)
544 zio_nowait(zio);
545 return (SET_ERROR(EIO));
546 }
547
548 /* initiate async i/o */
549 if (read)
550 (void) dbuf_read(db, zio, dbuf_flags);
551 dbp[i] = &db->db;
552 }
553
554 if ((flags & DMU_READ_NO_PREFETCH) == 0 &&
555 DNODE_META_IS_CACHEABLE(dn) && length <= zfetch_array_rd_sz) {
556 dmu_zfetch(&dn->dn_zfetch, blkid, nblks,
557 read && DNODE_IS_CACHEABLE(dn), B_TRUE);
558 }
559 rw_exit(&dn->dn_struct_rwlock);
560
561 if (read) {
562 /* wait for async read i/o */
563 err = zio_wait(zio);
564 if (err) {
565 dmu_buf_rele_array(dbp, nblks, tag);
566 return (err);
567 }
568
569 /* wait for other io to complete */
570 for (i = 0; i < nblks; i++) {
571 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
572 mutex_enter(&db->db_mtx);
573 while (db->db_state == DB_READ ||
574 db->db_state == DB_FILL)
575 cv_wait(&db->db_changed, &db->db_mtx);
576 if (db->db_state == DB_UNCACHED)
577 err = SET_ERROR(EIO);
578 mutex_exit(&db->db_mtx);
579 if (err) {
580 dmu_buf_rele_array(dbp, nblks, tag);
581 return (err);
582 }
583 }
584 }
585
586 *numbufsp = nblks;
587 *dbpp = dbp;
588 return (0);
589 }
590
591 static int
592 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
593 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
594 {
595 dnode_t *dn;
596 int err;
597
598 err = dnode_hold(os, object, FTAG, &dn);
599 if (err)
600 return (err);
601
602 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
603 numbufsp, dbpp, DMU_READ_PREFETCH);
604
605 dnode_rele(dn, FTAG);
606
607 return (err);
608 }
609
610 int
611 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
612 uint64_t length, boolean_t read, void *tag, int *numbufsp,
613 dmu_buf_t ***dbpp)
614 {
615 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
616 dnode_t *dn;
617 int err;
618
619 DB_DNODE_ENTER(db);
620 dn = DB_DNODE(db);
621 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
622 numbufsp, dbpp, DMU_READ_PREFETCH);
623 DB_DNODE_EXIT(db);
624
625 return (err);
626 }
627
628 void
629 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
630 {
631 int i;
632 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
633
634 if (numbufs == 0)
635 return;
636
637 for (i = 0; i < numbufs; i++) {
638 if (dbp[i])
639 dbuf_rele(dbp[i], tag);
640 }
641
642 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
643 }
644
645 /*
646 * Issue prefetch i/os for the given blocks. If level is greater than 0, the
647 * indirect blocks prefetched will be those that point to the blocks containing
648 * the data starting at offset, and continuing to offset + len.
649 *
650 * Note that if the indirect blocks above the blocks being prefetched are not
651 * in cache, they will be asynchronously read in.
652 */
653 void
654 dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset,
655 uint64_t len, zio_priority_t pri)
656 {
657 dnode_t *dn;
658 uint64_t blkid;
659 int nblks, err;
660
661 if (len == 0) { /* they're interested in the bonus buffer */
662 dn = DMU_META_DNODE(os);
663
664 if (object == 0 || object >= DN_MAX_OBJECT)
665 return;
666
667 rw_enter(&dn->dn_struct_rwlock, RW_READER);
668 blkid = dbuf_whichblock(dn, level,
669 object * sizeof (dnode_phys_t));
670 dbuf_prefetch(dn, level, blkid, pri, 0);
671 rw_exit(&dn->dn_struct_rwlock);
672 return;
673 }
674
675 /*
676 * See comment before the definition of dmu_prefetch_max.
677 */
678 len = MIN(len, dmu_prefetch_max);
679
680 /*
681 * XXX - Note, if the dnode for the requested object is not
682 * already cached, we will do a *synchronous* read in the
683 * dnode_hold() call. The same is true for any indirects.
684 */
685 err = dnode_hold(os, object, FTAG, &dn);
686 if (err != 0)
687 return;
688
689 /*
690 * offset + len - 1 is the last byte we want to prefetch for, and offset
691 * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the
692 * last block we want to prefetch, and dbuf_whichblock(dn, level,
693 * offset) is the first. Then the number we need to prefetch is the
694 * last - first + 1.
695 */
696 rw_enter(&dn->dn_struct_rwlock, RW_READER);
697 if (level > 0 || dn->dn_datablkshift != 0) {
698 nblks = dbuf_whichblock(dn, level, offset + len - 1) -
699 dbuf_whichblock(dn, level, offset) + 1;
700 } else {
701 nblks = (offset < dn->dn_datablksz);
702 }
703
704 if (nblks != 0) {
705 blkid = dbuf_whichblock(dn, level, offset);
706 for (int i = 0; i < nblks; i++)
707 dbuf_prefetch(dn, level, blkid + i, pri, 0);
708 }
709 rw_exit(&dn->dn_struct_rwlock);
710
711 dnode_rele(dn, FTAG);
712 }
713
714 /*
715 * Get the next "chunk" of file data to free. We traverse the file from
716 * the end so that the file gets shorter over time (if we crashes in the
717 * middle, this will leave us in a better state). We find allocated file
718 * data by simply searching the allocated level 1 indirects.
719 *
720 * On input, *start should be the first offset that does not need to be
721 * freed (e.g. "offset + length"). On return, *start will be the first
722 * offset that should be freed and l1blks is set to the number of level 1
723 * indirect blocks found within the chunk.
724 */
725 static int
726 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum, uint64_t *l1blks)
727 {
728 uint64_t blks;
729 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
730 /* bytes of data covered by a level-1 indirect block */
731 uint64_t iblkrange = (uint64_t)dn->dn_datablksz *
732 EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
733
734 ASSERT3U(minimum, <=, *start);
735
736 /*
737 * Check if we can free the entire range assuming that all of the
738 * L1 blocks in this range have data. If we can, we use this
739 * worst case value as an estimate so we can avoid having to look
740 * at the object's actual data.
741 */
742 uint64_t total_l1blks =
743 (roundup(*start, iblkrange) - (minimum / iblkrange * iblkrange)) /
744 iblkrange;
745 if (total_l1blks <= maxblks) {
746 *l1blks = total_l1blks;
747 *start = minimum;
748 return (0);
749 }
750 ASSERT(ISP2(iblkrange));
751
752 for (blks = 0; *start > minimum && blks < maxblks; blks++) {
753 int err;
754
755 /*
756 * dnode_next_offset(BACKWARDS) will find an allocated L1
757 * indirect block at or before the input offset. We must
758 * decrement *start so that it is at the end of the region
759 * to search.
760 */
761 (*start)--;
762
763 err = dnode_next_offset(dn,
764 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
765
766 /* if there are no indirect blocks before start, we are done */
767 if (err == ESRCH) {
768 *start = minimum;
769 break;
770 } else if (err != 0) {
771 *l1blks = blks;
772 return (err);
773 }
774
775 /* set start to the beginning of this L1 indirect */
776 *start = P2ALIGN(*start, iblkrange);
777 }
778 if (*start < minimum)
779 *start = minimum;
780 *l1blks = blks;
781
782 return (0);
783 }
784
785 /*
786 * If this objset is of type OST_ZFS return true if vfs's unmounted flag is set,
787 * otherwise return false.
788 * Used below in dmu_free_long_range_impl() to enable abort when unmounting
789 */
790 /*ARGSUSED*/
791 static boolean_t
792 dmu_objset_zfs_unmounting(objset_t *os)
793 {
794 #ifdef _KERNEL
795 if (dmu_objset_type(os) == DMU_OST_ZFS)
796 return (zfs_get_vfs_flag_unmounted(os));
797 #endif
798 return (B_FALSE);
799 }
800
801 static int
802 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
803 uint64_t length)
804 {
805 uint64_t object_size;
806 int err;
807 uint64_t dirty_frees_threshold;
808 dsl_pool_t *dp = dmu_objset_pool(os);
809
810 if (dn == NULL)
811 return (SET_ERROR(EINVAL));
812
813 object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
814 if (offset >= object_size)
815 return (0);
816
817 if (zfs_per_txg_dirty_frees_percent <= 100)
818 dirty_frees_threshold =
819 zfs_per_txg_dirty_frees_percent * zfs_dirty_data_max / 100;
820 else
821 dirty_frees_threshold = zfs_dirty_data_max / 20;
822
823 if (length == DMU_OBJECT_END || offset + length > object_size)
824 length = object_size - offset;
825
826 while (length != 0) {
827 uint64_t chunk_end, chunk_begin, chunk_len;
828 uint64_t l1blks;
829 dmu_tx_t *tx;
830
831 if (dmu_objset_zfs_unmounting(dn->dn_objset))
832 return (SET_ERROR(EINTR));
833
834 chunk_end = chunk_begin = offset + length;
835
836 /* move chunk_begin backwards to the beginning of this chunk */
837 err = get_next_chunk(dn, &chunk_begin, offset, &l1blks);
838 if (err)
839 return (err);
840 ASSERT3U(chunk_begin, >=, offset);
841 ASSERT3U(chunk_begin, <=, chunk_end);
842
843 chunk_len = chunk_end - chunk_begin;
844
845 tx = dmu_tx_create(os);
846 dmu_tx_hold_free(tx, dn->dn_object, chunk_begin, chunk_len);
847
848 /*
849 * Mark this transaction as typically resulting in a net
850 * reduction in space used.
851 */
852 dmu_tx_mark_netfree(tx);
853 err = dmu_tx_assign(tx, TXG_WAIT);
854 if (err) {
855 dmu_tx_abort(tx);
856 return (err);
857 }
858
859 uint64_t txg = dmu_tx_get_txg(tx);
860
861 mutex_enter(&dp->dp_lock);
862 uint64_t long_free_dirty =
863 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK];
864 mutex_exit(&dp->dp_lock);
865
866 /*
867 * To avoid filling up a TXG with just frees, wait for
868 * the next TXG to open before freeing more chunks if
869 * we have reached the threshold of frees.
870 */
871 if (dirty_frees_threshold != 0 &&
872 long_free_dirty >= dirty_frees_threshold) {
873 DMU_TX_STAT_BUMP(dmu_tx_dirty_frees_delay);
874 dmu_tx_commit(tx);
875 txg_wait_open(dp, 0, B_TRUE);
876 continue;
877 }
878
879 /*
880 * In order to prevent unnecessary write throttling, for each
881 * TXG, we track the cumulative size of L1 blocks being dirtied
882 * in dnode_free_range() below. We compare this number to a
883 * tunable threshold, past which we prevent new L1 dirty freeing
884 * blocks from being added into the open TXG. See
885 * dmu_free_long_range_impl() for details. The threshold
886 * prevents write throttle activation due to dirty freeing L1
887 * blocks taking up a large percentage of zfs_dirty_data_max.
888 */
889 mutex_enter(&dp->dp_lock);
890 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] +=
891 l1blks << dn->dn_indblkshift;
892 mutex_exit(&dp->dp_lock);
893 DTRACE_PROBE3(free__long__range,
894 uint64_t, long_free_dirty, uint64_t, chunk_len,
895 uint64_t, txg);
896 dnode_free_range(dn, chunk_begin, chunk_len, tx);
897
898 dmu_tx_commit(tx);
899
900 length -= chunk_len;
901 }
902 return (0);
903 }
904
905 int
906 dmu_free_long_range(objset_t *os, uint64_t object,
907 uint64_t offset, uint64_t length)
908 {
909 dnode_t *dn;
910 int err;
911
912 err = dnode_hold(os, object, FTAG, &dn);
913 if (err != 0)
914 return (err);
915 err = dmu_free_long_range_impl(os, dn, offset, length);
916
917 /*
918 * It is important to zero out the maxblkid when freeing the entire
919 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
920 * will take the fast path, and (b) dnode_reallocate() can verify
921 * that the entire file has been freed.
922 */
923 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
924 dn->dn_maxblkid = 0;
925
926 dnode_rele(dn, FTAG);
927 return (err);
928 }
929
930 int
931 dmu_free_long_object(objset_t *os, uint64_t object)
932 {
933 dmu_tx_t *tx;
934 int err;
935
936 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
937 if (err != 0)
938 return (err);
939
940 tx = dmu_tx_create(os);
941 dmu_tx_hold_bonus(tx, object);
942 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
943 dmu_tx_mark_netfree(tx);
944 err = dmu_tx_assign(tx, TXG_WAIT);
945 if (err == 0) {
946 if (err == 0)
947 err = dmu_object_free(os, object, tx);
948
949 dmu_tx_commit(tx);
950 } else {
951 dmu_tx_abort(tx);
952 }
953
954 return (err);
955 }
956
957 int
958 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
959 uint64_t size, dmu_tx_t *tx)
960 {
961 dnode_t *dn;
962 int err = dnode_hold(os, object, FTAG, &dn);
963 if (err)
964 return (err);
965 ASSERT(offset < UINT64_MAX);
966 ASSERT(size == DMU_OBJECT_END || size <= UINT64_MAX - offset);
967 dnode_free_range(dn, offset, size, tx);
968 dnode_rele(dn, FTAG);
969 return (0);
970 }
971
972 static int
973 dmu_read_impl(dnode_t *dn, uint64_t offset, uint64_t size,
974 void *buf, uint32_t flags)
975 {
976 dmu_buf_t **dbp;
977 int numbufs, err = 0;
978
979 /*
980 * Deal with odd block sizes, where there can't be data past the first
981 * block. If we ever do the tail block optimization, we will need to
982 * handle that here as well.
983 */
984 if (dn->dn_maxblkid == 0) {
985 uint64_t newsz = offset > dn->dn_datablksz ? 0 :
986 MIN(size, dn->dn_datablksz - offset);
987 bzero((char *)buf + newsz, size - newsz);
988 size = newsz;
989 }
990
991 while (size > 0) {
992 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
993 int i;
994
995 /*
996 * NB: we could do this block-at-a-time, but it's nice
997 * to be reading in parallel.
998 */
999 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
1000 TRUE, FTAG, &numbufs, &dbp, flags);
1001 if (err)
1002 break;
1003
1004 for (i = 0; i < numbufs; i++) {
1005 uint64_t tocpy;
1006 int64_t bufoff;
1007 dmu_buf_t *db = dbp[i];
1008
1009 ASSERT(size > 0);
1010
1011 bufoff = offset - db->db_offset;
1012 tocpy = MIN(db->db_size - bufoff, size);
1013
1014 (void) memcpy(buf, (char *)db->db_data + bufoff, tocpy);
1015
1016 offset += tocpy;
1017 size -= tocpy;
1018 buf = (char *)buf + tocpy;
1019 }
1020 dmu_buf_rele_array(dbp, numbufs, FTAG);
1021 }
1022 return (err);
1023 }
1024
1025 int
1026 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1027 void *buf, uint32_t flags)
1028 {
1029 dnode_t *dn;
1030 int err;
1031
1032 err = dnode_hold(os, object, FTAG, &dn);
1033 if (err != 0)
1034 return (err);
1035
1036 err = dmu_read_impl(dn, offset, size, buf, flags);
1037 dnode_rele(dn, FTAG);
1038 return (err);
1039 }
1040
1041 int
1042 dmu_read_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, void *buf,
1043 uint32_t flags)
1044 {
1045 return (dmu_read_impl(dn, offset, size, buf, flags));
1046 }
1047
1048 static void
1049 dmu_write_impl(dmu_buf_t **dbp, int numbufs, uint64_t offset, uint64_t size,
1050 const void *buf, dmu_tx_t *tx)
1051 {
1052 int i;
1053
1054 for (i = 0; i < numbufs; i++) {
1055 uint64_t tocpy;
1056 int64_t bufoff;
1057 dmu_buf_t *db = dbp[i];
1058
1059 ASSERT(size > 0);
1060
1061 bufoff = offset - db->db_offset;
1062 tocpy = MIN(db->db_size - bufoff, size);
1063
1064 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1065
1066 if (tocpy == db->db_size)
1067 dmu_buf_will_fill(db, tx);
1068 else
1069 dmu_buf_will_dirty(db, tx);
1070
1071 (void) memcpy((char *)db->db_data + bufoff, buf, tocpy);
1072
1073 if (tocpy == db->db_size)
1074 dmu_buf_fill_done(db, tx);
1075
1076 offset += tocpy;
1077 size -= tocpy;
1078 buf = (char *)buf + tocpy;
1079 }
1080 }
1081
1082 void
1083 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1084 const void *buf, dmu_tx_t *tx)
1085 {
1086 dmu_buf_t **dbp;
1087 int numbufs;
1088
1089 if (size == 0)
1090 return;
1091
1092 VERIFY0(dmu_buf_hold_array(os, object, offset, size,
1093 FALSE, FTAG, &numbufs, &dbp));
1094 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1095 dmu_buf_rele_array(dbp, numbufs, FTAG);
1096 }
1097
1098 /*
1099 * Note: Lustre is an external consumer of this interface.
1100 */
1101 void
1102 dmu_write_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size,
1103 const void *buf, dmu_tx_t *tx)
1104 {
1105 dmu_buf_t **dbp;
1106 int numbufs;
1107
1108 if (size == 0)
1109 return;
1110
1111 VERIFY0(dmu_buf_hold_array_by_dnode(dn, offset, size,
1112 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH));
1113 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1114 dmu_buf_rele_array(dbp, numbufs, FTAG);
1115 }
1116
1117 void
1118 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1119 dmu_tx_t *tx)
1120 {
1121 dmu_buf_t **dbp;
1122 int numbufs, i;
1123
1124 if (size == 0)
1125 return;
1126
1127 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
1128 FALSE, FTAG, &numbufs, &dbp));
1129
1130 for (i = 0; i < numbufs; i++) {
1131 dmu_buf_t *db = dbp[i];
1132
1133 dmu_buf_will_not_fill(db, tx);
1134 }
1135 dmu_buf_rele_array(dbp, numbufs, FTAG);
1136 }
1137
1138 void
1139 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
1140 void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
1141 int compressed_size, int byteorder, dmu_tx_t *tx)
1142 {
1143 dmu_buf_t *db;
1144
1145 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
1146 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
1147 VERIFY0(dmu_buf_hold_noread(os, object, offset,
1148 FTAG, &db));
1149
1150 dmu_buf_write_embedded(db,
1151 data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
1152 uncompressed_size, compressed_size, byteorder, tx);
1153
1154 dmu_buf_rele(db, FTAG);
1155 }
1156
1157 void
1158 dmu_redact(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1159 dmu_tx_t *tx)
1160 {
1161 int numbufs, i;
1162 dmu_buf_t **dbp;
1163
1164 VERIFY0(dmu_buf_hold_array(os, object, offset, size, FALSE, FTAG,
1165 &numbufs, &dbp));
1166 for (i = 0; i < numbufs; i++)
1167 dmu_buf_redact(dbp[i], tx);
1168 dmu_buf_rele_array(dbp, numbufs, FTAG);
1169 }
1170
1171 #ifdef _KERNEL
1172 int
1173 dmu_read_uio_dnode(dnode_t *dn, zfs_uio_t *uio, uint64_t size)
1174 {
1175 dmu_buf_t **dbp;
1176 int numbufs, i, err;
1177
1178 /*
1179 * NB: we could do this block-at-a-time, but it's nice
1180 * to be reading in parallel.
1181 */
1182 err = dmu_buf_hold_array_by_dnode(dn, zfs_uio_offset(uio), size,
1183 TRUE, FTAG, &numbufs, &dbp, 0);
1184 if (err)
1185 return (err);
1186
1187 for (i = 0; i < numbufs; i++) {
1188 uint64_t tocpy;
1189 int64_t bufoff;
1190 dmu_buf_t *db = dbp[i];
1191
1192 ASSERT(size > 0);
1193
1194 bufoff = zfs_uio_offset(uio) - db->db_offset;
1195 tocpy = MIN(db->db_size - bufoff, size);
1196
1197 err = zfs_uio_fault_move((char *)db->db_data + bufoff, tocpy,
1198 UIO_READ, uio);
1199
1200 if (err)
1201 break;
1202
1203 size -= tocpy;
1204 }
1205 dmu_buf_rele_array(dbp, numbufs, FTAG);
1206
1207 return (err);
1208 }
1209
1210 /*
1211 * Read 'size' bytes into the uio buffer.
1212 * From object zdb->db_object.
1213 * Starting at zfs_uio_offset(uio).
1214 *
1215 * If the caller already has a dbuf in the target object
1216 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1217 * because we don't have to find the dnode_t for the object.
1218 */
1219 int
1220 dmu_read_uio_dbuf(dmu_buf_t *zdb, zfs_uio_t *uio, uint64_t size)
1221 {
1222 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1223 dnode_t *dn;
1224 int err;
1225
1226 if (size == 0)
1227 return (0);
1228
1229 DB_DNODE_ENTER(db);
1230 dn = DB_DNODE(db);
1231 err = dmu_read_uio_dnode(dn, uio, size);
1232 DB_DNODE_EXIT(db);
1233
1234 return (err);
1235 }
1236
1237 /*
1238 * Read 'size' bytes into the uio buffer.
1239 * From the specified object
1240 * Starting at offset zfs_uio_offset(uio).
1241 */
1242 int
1243 dmu_read_uio(objset_t *os, uint64_t object, zfs_uio_t *uio, uint64_t size)
1244 {
1245 dnode_t *dn;
1246 int err;
1247
1248 if (size == 0)
1249 return (0);
1250
1251 err = dnode_hold(os, object, FTAG, &dn);
1252 if (err)
1253 return (err);
1254
1255 err = dmu_read_uio_dnode(dn, uio, size);
1256
1257 dnode_rele(dn, FTAG);
1258
1259 return (err);
1260 }
1261
1262 int
1263 dmu_write_uio_dnode(dnode_t *dn, zfs_uio_t *uio, uint64_t size, dmu_tx_t *tx)
1264 {
1265 dmu_buf_t **dbp;
1266 int numbufs;
1267 int err = 0;
1268 int i;
1269
1270 err = dmu_buf_hold_array_by_dnode(dn, zfs_uio_offset(uio), size,
1271 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1272 if (err)
1273 return (err);
1274
1275 for (i = 0; i < numbufs; i++) {
1276 uint64_t tocpy;
1277 int64_t bufoff;
1278 dmu_buf_t *db = dbp[i];
1279
1280 ASSERT(size > 0);
1281
1282 bufoff = zfs_uio_offset(uio) - db->db_offset;
1283 tocpy = MIN(db->db_size - bufoff, size);
1284
1285 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1286
1287 if (tocpy == db->db_size)
1288 dmu_buf_will_fill(db, tx);
1289 else
1290 dmu_buf_will_dirty(db, tx);
1291
1292 /*
1293 * XXX zfs_uiomove could block forever (eg.nfs-backed
1294 * pages). There needs to be a uiolockdown() function
1295 * to lock the pages in memory, so that zfs_uiomove won't
1296 * block.
1297 */
1298 err = zfs_uio_fault_move((char *)db->db_data + bufoff,
1299 tocpy, UIO_WRITE, uio);
1300
1301 if (tocpy == db->db_size)
1302 dmu_buf_fill_done(db, tx);
1303
1304 if (err)
1305 break;
1306
1307 size -= tocpy;
1308 }
1309
1310 dmu_buf_rele_array(dbp, numbufs, FTAG);
1311 return (err);
1312 }
1313
1314 /*
1315 * Write 'size' bytes from the uio buffer.
1316 * To object zdb->db_object.
1317 * Starting at offset zfs_uio_offset(uio).
1318 *
1319 * If the caller already has a dbuf in the target object
1320 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1321 * because we don't have to find the dnode_t for the object.
1322 */
1323 int
1324 dmu_write_uio_dbuf(dmu_buf_t *zdb, zfs_uio_t *uio, uint64_t size,
1325 dmu_tx_t *tx)
1326 {
1327 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1328 dnode_t *dn;
1329 int err;
1330
1331 if (size == 0)
1332 return (0);
1333
1334 DB_DNODE_ENTER(db);
1335 dn = DB_DNODE(db);
1336 err = dmu_write_uio_dnode(dn, uio, size, tx);
1337 DB_DNODE_EXIT(db);
1338
1339 return (err);
1340 }
1341
1342 /*
1343 * Write 'size' bytes from the uio buffer.
1344 * To the specified object.
1345 * Starting at offset zfs_uio_offset(uio).
1346 */
1347 int
1348 dmu_write_uio(objset_t *os, uint64_t object, zfs_uio_t *uio, uint64_t size,
1349 dmu_tx_t *tx)
1350 {
1351 dnode_t *dn;
1352 int err;
1353
1354 if (size == 0)
1355 return (0);
1356
1357 err = dnode_hold(os, object, FTAG, &dn);
1358 if (err)
1359 return (err);
1360
1361 err = dmu_write_uio_dnode(dn, uio, size, tx);
1362
1363 dnode_rele(dn, FTAG);
1364
1365 return (err);
1366 }
1367 #endif /* _KERNEL */
1368
1369 /*
1370 * Allocate a loaned anonymous arc buffer.
1371 */
1372 arc_buf_t *
1373 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1374 {
1375 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1376
1377 return (arc_loan_buf(db->db_objset->os_spa, B_FALSE, size));
1378 }
1379
1380 /*
1381 * Free a loaned arc buffer.
1382 */
1383 void
1384 dmu_return_arcbuf(arc_buf_t *buf)
1385 {
1386 arc_return_buf(buf, FTAG);
1387 arc_buf_destroy(buf, FTAG);
1388 }
1389
1390 /*
1391 * A "lightweight" write is faster than a regular write (e.g.
1392 * dmu_write_by_dnode() or dmu_assign_arcbuf_by_dnode()), because it avoids the
1393 * CPU cost of creating a dmu_buf_impl_t and arc_buf_[hdr_]_t. However, the
1394 * data can not be read or overwritten until the transaction's txg has been
1395 * synced. This makes it appropriate for workloads that are known to be
1396 * (temporarily) write-only, like "zfs receive".
1397 *
1398 * A single block is written, starting at the specified offset in bytes. If
1399 * the call is successful, it returns 0 and the provided abd has been
1400 * consumed (the caller should not free it).
1401 */
1402 int
1403 dmu_lightweight_write_by_dnode(dnode_t *dn, uint64_t offset, abd_t *abd,
1404 const zio_prop_t *zp, enum zio_flag flags, dmu_tx_t *tx)
1405 {
1406 dbuf_dirty_record_t *dr =
1407 dbuf_dirty_lightweight(dn, dbuf_whichblock(dn, 0, offset), tx);
1408 if (dr == NULL)
1409 return (SET_ERROR(EIO));
1410 dr->dt.dll.dr_abd = abd;
1411 dr->dt.dll.dr_props = *zp;
1412 dr->dt.dll.dr_flags = flags;
1413 return (0);
1414 }
1415
1416 /*
1417 * When possible directly assign passed loaned arc buffer to a dbuf.
1418 * If this is not possible copy the contents of passed arc buf via
1419 * dmu_write().
1420 */
1421 int
1422 dmu_assign_arcbuf_by_dnode(dnode_t *dn, uint64_t offset, arc_buf_t *buf,
1423 dmu_tx_t *tx)
1424 {
1425 dmu_buf_impl_t *db;
1426 objset_t *os = dn->dn_objset;
1427 uint64_t object = dn->dn_object;
1428 uint32_t blksz = (uint32_t)arc_buf_lsize(buf);
1429 uint64_t blkid;
1430
1431 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1432 blkid = dbuf_whichblock(dn, 0, offset);
1433 db = dbuf_hold(dn, blkid, FTAG);
1434 if (db == NULL)
1435 return (SET_ERROR(EIO));
1436 rw_exit(&dn->dn_struct_rwlock);
1437
1438 /*
1439 * We can only assign if the offset is aligned and the arc buf is the
1440 * same size as the dbuf.
1441 */
1442 if (offset == db->db.db_offset && blksz == db->db.db_size) {
1443 dbuf_assign_arcbuf(db, buf, tx);
1444 dbuf_rele(db, FTAG);
1445 } else {
1446 /* compressed bufs must always be assignable to their dbuf */
1447 ASSERT3U(arc_get_compression(buf), ==, ZIO_COMPRESS_OFF);
1448 ASSERT(!(buf->b_flags & ARC_BUF_FLAG_COMPRESSED));
1449
1450 dbuf_rele(db, FTAG);
1451 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1452 dmu_return_arcbuf(buf);
1453 }
1454
1455 return (0);
1456 }
1457
1458 int
1459 dmu_assign_arcbuf_by_dbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1460 dmu_tx_t *tx)
1461 {
1462 int err;
1463 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1464
1465 DB_DNODE_ENTER(dbuf);
1466 err = dmu_assign_arcbuf_by_dnode(DB_DNODE(dbuf), offset, buf, tx);
1467 DB_DNODE_EXIT(dbuf);
1468
1469 return (err);
1470 }
1471
1472 typedef struct {
1473 dbuf_dirty_record_t *dsa_dr;
1474 dmu_sync_cb_t *dsa_done;
1475 zgd_t *dsa_zgd;
1476 dmu_tx_t *dsa_tx;
1477 } dmu_sync_arg_t;
1478
1479 /* ARGSUSED */
1480 static void
1481 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1482 {
1483 dmu_sync_arg_t *dsa = varg;
1484 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1485 blkptr_t *bp = zio->io_bp;
1486
1487 if (zio->io_error == 0) {
1488 if (BP_IS_HOLE(bp)) {
1489 /*
1490 * A block of zeros may compress to a hole, but the
1491 * block size still needs to be known for replay.
1492 */
1493 BP_SET_LSIZE(bp, db->db_size);
1494 } else if (!BP_IS_EMBEDDED(bp)) {
1495 ASSERT(BP_GET_LEVEL(bp) == 0);
1496 BP_SET_FILL(bp, 1);
1497 }
1498 }
1499 }
1500
1501 static void
1502 dmu_sync_late_arrival_ready(zio_t *zio)
1503 {
1504 dmu_sync_ready(zio, NULL, zio->io_private);
1505 }
1506
1507 /* ARGSUSED */
1508 static void
1509 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1510 {
1511 dmu_sync_arg_t *dsa = varg;
1512 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1513 dmu_buf_impl_t *db = dr->dr_dbuf;
1514 zgd_t *zgd = dsa->dsa_zgd;
1515
1516 /*
1517 * Record the vdev(s) backing this blkptr so they can be flushed after
1518 * the writes for the lwb have completed.
1519 */
1520 if (zio->io_error == 0) {
1521 zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp);
1522 }
1523
1524 mutex_enter(&db->db_mtx);
1525 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1526 if (zio->io_error == 0) {
1527 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1528 if (dr->dt.dl.dr_nopwrite) {
1529 blkptr_t *bp = zio->io_bp;
1530 blkptr_t *bp_orig = &zio->io_bp_orig;
1531 uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1532
1533 ASSERT(BP_EQUAL(bp, bp_orig));
1534 VERIFY(BP_EQUAL(bp, db->db_blkptr));
1535 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1536 VERIFY(zio_checksum_table[chksum].ci_flags &
1537 ZCHECKSUM_FLAG_NOPWRITE);
1538 }
1539 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1540 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1541 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1542
1543 /*
1544 * Old style holes are filled with all zeros, whereas
1545 * new-style holes maintain their lsize, type, level,
1546 * and birth time (see zio_write_compress). While we
1547 * need to reset the BP_SET_LSIZE() call that happened
1548 * in dmu_sync_ready for old style holes, we do *not*
1549 * want to wipe out the information contained in new
1550 * style holes. Thus, only zero out the block pointer if
1551 * it's an old style hole.
1552 */
1553 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) &&
1554 dr->dt.dl.dr_overridden_by.blk_birth == 0)
1555 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1556 } else {
1557 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1558 }
1559 cv_broadcast(&db->db_changed);
1560 mutex_exit(&db->db_mtx);
1561
1562 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1563
1564 kmem_free(dsa, sizeof (*dsa));
1565 }
1566
1567 static void
1568 dmu_sync_late_arrival_done(zio_t *zio)
1569 {
1570 blkptr_t *bp = zio->io_bp;
1571 dmu_sync_arg_t *dsa = zio->io_private;
1572 zgd_t *zgd = dsa->dsa_zgd;
1573
1574 if (zio->io_error == 0) {
1575 /*
1576 * Record the vdev(s) backing this blkptr so they can be
1577 * flushed after the writes for the lwb have completed.
1578 */
1579 zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp);
1580
1581 if (!BP_IS_HOLE(bp)) {
1582 blkptr_t *bp_orig __maybe_unused = &zio->io_bp_orig;
1583 ASSERT(!(zio->io_flags & ZIO_FLAG_NOPWRITE));
1584 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1585 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1586 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1587 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1588 }
1589 }
1590
1591 dmu_tx_commit(dsa->dsa_tx);
1592
1593 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1594
1595 abd_free(zio->io_abd);
1596 kmem_free(dsa, sizeof (*dsa));
1597 }
1598
1599 static int
1600 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1601 zio_prop_t *zp, zbookmark_phys_t *zb)
1602 {
1603 dmu_sync_arg_t *dsa;
1604 dmu_tx_t *tx;
1605
1606 tx = dmu_tx_create(os);
1607 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1608 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1609 dmu_tx_abort(tx);
1610 /* Make zl_get_data do txg_waited_synced() */
1611 return (SET_ERROR(EIO));
1612 }
1613
1614 /*
1615 * In order to prevent the zgd's lwb from being free'd prior to
1616 * dmu_sync_late_arrival_done() being called, we have to ensure
1617 * the lwb's "max txg" takes this tx's txg into account.
1618 */
1619 zil_lwb_add_txg(zgd->zgd_lwb, dmu_tx_get_txg(tx));
1620
1621 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1622 dsa->dsa_dr = NULL;
1623 dsa->dsa_done = done;
1624 dsa->dsa_zgd = zgd;
1625 dsa->dsa_tx = tx;
1626
1627 /*
1628 * Since we are currently syncing this txg, it's nontrivial to
1629 * determine what BP to nopwrite against, so we disable nopwrite.
1630 *
1631 * When syncing, the db_blkptr is initially the BP of the previous
1632 * txg. We can not nopwrite against it because it will be changed
1633 * (this is similar to the non-late-arrival case where the dbuf is
1634 * dirty in a future txg).
1635 *
1636 * Then dbuf_write_ready() sets bp_blkptr to the location we will write.
1637 * We can not nopwrite against it because although the BP will not
1638 * (typically) be changed, the data has not yet been persisted to this
1639 * location.
1640 *
1641 * Finally, when dbuf_write_done() is called, it is theoretically
1642 * possible to always nopwrite, because the data that was written in
1643 * this txg is the same data that we are trying to write. However we
1644 * would need to check that this dbuf is not dirty in any future
1645 * txg's (as we do in the normal dmu_sync() path). For simplicity, we
1646 * don't nopwrite in this case.
1647 */
1648 zp->zp_nopwrite = B_FALSE;
1649
1650 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1651 abd_get_from_buf(zgd->zgd_db->db_data, zgd->zgd_db->db_size),
1652 zgd->zgd_db->db_size, zgd->zgd_db->db_size, zp,
1653 dmu_sync_late_arrival_ready, NULL, NULL, dmu_sync_late_arrival_done,
1654 dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1655
1656 return (0);
1657 }
1658
1659 /*
1660 * Intent log support: sync the block associated with db to disk.
1661 * N.B. and XXX: the caller is responsible for making sure that the
1662 * data isn't changing while dmu_sync() is writing it.
1663 *
1664 * Return values:
1665 *
1666 * EEXIST: this txg has already been synced, so there's nothing to do.
1667 * The caller should not log the write.
1668 *
1669 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1670 * The caller should not log the write.
1671 *
1672 * EALREADY: this block is already in the process of being synced.
1673 * The caller should track its progress (somehow).
1674 *
1675 * EIO: could not do the I/O.
1676 * The caller should do a txg_wait_synced().
1677 *
1678 * 0: the I/O has been initiated.
1679 * The caller should log this blkptr in the done callback.
1680 * It is possible that the I/O will fail, in which case
1681 * the error will be reported to the done callback and
1682 * propagated to pio from zio_done().
1683 */
1684 int
1685 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1686 {
1687 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1688 objset_t *os = db->db_objset;
1689 dsl_dataset_t *ds = os->os_dsl_dataset;
1690 dbuf_dirty_record_t *dr, *dr_next;
1691 dmu_sync_arg_t *dsa;
1692 zbookmark_phys_t zb;
1693 zio_prop_t zp;
1694 dnode_t *dn;
1695
1696 ASSERT(pio != NULL);
1697 ASSERT(txg != 0);
1698
1699 SET_BOOKMARK(&zb, ds->ds_object,
1700 db->db.db_object, db->db_level, db->db_blkid);
1701
1702 DB_DNODE_ENTER(db);
1703 dn = DB_DNODE(db);
1704 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1705 DB_DNODE_EXIT(db);
1706
1707 /*
1708 * If we're frozen (running ziltest), we always need to generate a bp.
1709 */
1710 if (txg > spa_freeze_txg(os->os_spa))
1711 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1712
1713 /*
1714 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1715 * and us. If we determine that this txg is not yet syncing,
1716 * but it begins to sync a moment later, that's OK because the
1717 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1718 */
1719 mutex_enter(&db->db_mtx);
1720
1721 if (txg <= spa_last_synced_txg(os->os_spa)) {
1722 /*
1723 * This txg has already synced. There's nothing to do.
1724 */
1725 mutex_exit(&db->db_mtx);
1726 return (SET_ERROR(EEXIST));
1727 }
1728
1729 if (txg <= spa_syncing_txg(os->os_spa)) {
1730 /*
1731 * This txg is currently syncing, so we can't mess with
1732 * the dirty record anymore; just write a new log block.
1733 */
1734 mutex_exit(&db->db_mtx);
1735 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1736 }
1737
1738 dr = dbuf_find_dirty_eq(db, txg);
1739
1740 if (dr == NULL) {
1741 /*
1742 * There's no dr for this dbuf, so it must have been freed.
1743 * There's no need to log writes to freed blocks, so we're done.
1744 */
1745 mutex_exit(&db->db_mtx);
1746 return (SET_ERROR(ENOENT));
1747 }
1748
1749 dr_next = list_next(&db->db_dirty_records, dr);
1750 ASSERT(dr_next == NULL || dr_next->dr_txg < txg);
1751
1752 if (db->db_blkptr != NULL) {
1753 /*
1754 * We need to fill in zgd_bp with the current blkptr so that
1755 * the nopwrite code can check if we're writing the same
1756 * data that's already on disk. We can only nopwrite if we
1757 * are sure that after making the copy, db_blkptr will not
1758 * change until our i/o completes. We ensure this by
1759 * holding the db_mtx, and only allowing nopwrite if the
1760 * block is not already dirty (see below). This is verified
1761 * by dmu_sync_done(), which VERIFYs that the db_blkptr has
1762 * not changed.
1763 */
1764 *zgd->zgd_bp = *db->db_blkptr;
1765 }
1766
1767 /*
1768 * Assume the on-disk data is X, the current syncing data (in
1769 * txg - 1) is Y, and the current in-memory data is Z (currently
1770 * in dmu_sync).
1771 *
1772 * We usually want to perform a nopwrite if X and Z are the
1773 * same. However, if Y is different (i.e. the BP is going to
1774 * change before this write takes effect), then a nopwrite will
1775 * be incorrect - we would override with X, which could have
1776 * been freed when Y was written.
1777 *
1778 * (Note that this is not a concern when we are nop-writing from
1779 * syncing context, because X and Y must be identical, because
1780 * all previous txgs have been synced.)
1781 *
1782 * Therefore, we disable nopwrite if the current BP could change
1783 * before this TXG. There are two ways it could change: by
1784 * being dirty (dr_next is non-NULL), or by being freed
1785 * (dnode_block_freed()). This behavior is verified by
1786 * zio_done(), which VERIFYs that the override BP is identical
1787 * to the on-disk BP.
1788 */
1789 DB_DNODE_ENTER(db);
1790 dn = DB_DNODE(db);
1791 if (dr_next != NULL || dnode_block_freed(dn, db->db_blkid))
1792 zp.zp_nopwrite = B_FALSE;
1793 DB_DNODE_EXIT(db);
1794
1795 ASSERT(dr->dr_txg == txg);
1796 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1797 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1798 /*
1799 * We have already issued a sync write for this buffer,
1800 * or this buffer has already been synced. It could not
1801 * have been dirtied since, or we would have cleared the state.
1802 */
1803 mutex_exit(&db->db_mtx);
1804 return (SET_ERROR(EALREADY));
1805 }
1806
1807 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1808 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1809 mutex_exit(&db->db_mtx);
1810
1811 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1812 dsa->dsa_dr = dr;
1813 dsa->dsa_done = done;
1814 dsa->dsa_zgd = zgd;
1815 dsa->dsa_tx = NULL;
1816
1817 zio_nowait(arc_write(pio, os->os_spa, txg,
1818 zgd->zgd_bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
1819 &zp, dmu_sync_ready, NULL, NULL, dmu_sync_done, dsa,
1820 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb));
1821
1822 return (0);
1823 }
1824
1825 int
1826 dmu_object_set_nlevels(objset_t *os, uint64_t object, int nlevels, dmu_tx_t *tx)
1827 {
1828 dnode_t *dn;
1829 int err;
1830
1831 err = dnode_hold(os, object, FTAG, &dn);
1832 if (err)
1833 return (err);
1834 err = dnode_set_nlevels(dn, nlevels, tx);
1835 dnode_rele(dn, FTAG);
1836 return (err);
1837 }
1838
1839 int
1840 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1841 dmu_tx_t *tx)
1842 {
1843 dnode_t *dn;
1844 int err;
1845
1846 err = dnode_hold(os, object, FTAG, &dn);
1847 if (err)
1848 return (err);
1849 err = dnode_set_blksz(dn, size, ibs, tx);
1850 dnode_rele(dn, FTAG);
1851 return (err);
1852 }
1853
1854 int
1855 dmu_object_set_maxblkid(objset_t *os, uint64_t object, uint64_t maxblkid,
1856 dmu_tx_t *tx)
1857 {
1858 dnode_t *dn;
1859 int err;
1860
1861 err = dnode_hold(os, object, FTAG, &dn);
1862 if (err)
1863 return (err);
1864 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1865 dnode_new_blkid(dn, maxblkid, tx, B_FALSE, B_TRUE);
1866 rw_exit(&dn->dn_struct_rwlock);
1867 dnode_rele(dn, FTAG);
1868 return (0);
1869 }
1870
1871 void
1872 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1873 dmu_tx_t *tx)
1874 {
1875 dnode_t *dn;
1876
1877 /*
1878 * Send streams include each object's checksum function. This
1879 * check ensures that the receiving system can understand the
1880 * checksum function transmitted.
1881 */
1882 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS);
1883
1884 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1885 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS);
1886 dn->dn_checksum = checksum;
1887 dnode_setdirty(dn, tx);
1888 dnode_rele(dn, FTAG);
1889 }
1890
1891 void
1892 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1893 dmu_tx_t *tx)
1894 {
1895 dnode_t *dn;
1896
1897 /*
1898 * Send streams include each object's compression function. This
1899 * check ensures that the receiving system can understand the
1900 * compression function transmitted.
1901 */
1902 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS);
1903
1904 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1905 dn->dn_compress = compress;
1906 dnode_setdirty(dn, tx);
1907 dnode_rele(dn, FTAG);
1908 }
1909
1910 /*
1911 * When the "redundant_metadata" property is set to "most", only indirect
1912 * blocks of this level and higher will have an additional ditto block.
1913 */
1914 int zfs_redundant_metadata_most_ditto_level = 2;
1915
1916 void
1917 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
1918 {
1919 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
1920 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
1921 (wp & WP_SPILL));
1922 enum zio_checksum checksum = os->os_checksum;
1923 enum zio_compress compress = os->os_compress;
1924 uint8_t complevel = os->os_complevel;
1925 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
1926 boolean_t dedup = B_FALSE;
1927 boolean_t nopwrite = B_FALSE;
1928 boolean_t dedup_verify = os->os_dedup_verify;
1929 boolean_t encrypt = B_FALSE;
1930 int copies = os->os_copies;
1931
1932 /*
1933 * We maintain different write policies for each of the following
1934 * types of data:
1935 * 1. metadata
1936 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1937 * 3. all other level 0 blocks
1938 */
1939 if (ismd) {
1940 /*
1941 * XXX -- we should design a compression algorithm
1942 * that specializes in arrays of bps.
1943 */
1944 compress = zio_compress_select(os->os_spa,
1945 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON);
1946
1947 /*
1948 * Metadata always gets checksummed. If the data
1949 * checksum is multi-bit correctable, and it's not a
1950 * ZBT-style checksum, then it's suitable for metadata
1951 * as well. Otherwise, the metadata checksum defaults
1952 * to fletcher4.
1953 */
1954 if (!(zio_checksum_table[checksum].ci_flags &
1955 ZCHECKSUM_FLAG_METADATA) ||
1956 (zio_checksum_table[checksum].ci_flags &
1957 ZCHECKSUM_FLAG_EMBEDDED))
1958 checksum = ZIO_CHECKSUM_FLETCHER_4;
1959
1960 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL ||
1961 (os->os_redundant_metadata ==
1962 ZFS_REDUNDANT_METADATA_MOST &&
1963 (level >= zfs_redundant_metadata_most_ditto_level ||
1964 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))))
1965 copies++;
1966 } else if (wp & WP_NOFILL) {
1967 ASSERT(level == 0);
1968
1969 /*
1970 * If we're writing preallocated blocks, we aren't actually
1971 * writing them so don't set any policy properties. These
1972 * blocks are currently only used by an external subsystem
1973 * outside of zfs (i.e. dump) and not written by the zio
1974 * pipeline.
1975 */
1976 compress = ZIO_COMPRESS_OFF;
1977 checksum = ZIO_CHECKSUM_OFF;
1978 } else {
1979 compress = zio_compress_select(os->os_spa, dn->dn_compress,
1980 compress);
1981 complevel = zio_complevel_select(os->os_spa, compress,
1982 complevel, complevel);
1983
1984 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
1985 zio_checksum_select(dn->dn_checksum, checksum) :
1986 dedup_checksum;
1987
1988 /*
1989 * Determine dedup setting. If we are in dmu_sync(),
1990 * we won't actually dedup now because that's all
1991 * done in syncing context; but we do want to use the
1992 * dedup checksum. If the checksum is not strong
1993 * enough to ensure unique signatures, force
1994 * dedup_verify.
1995 */
1996 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
1997 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
1998 if (!(zio_checksum_table[checksum].ci_flags &
1999 ZCHECKSUM_FLAG_DEDUP))
2000 dedup_verify = B_TRUE;
2001 }
2002
2003 /*
2004 * Enable nopwrite if we have secure enough checksum
2005 * algorithm (see comment in zio_nop_write) and
2006 * compression is enabled. We don't enable nopwrite if
2007 * dedup is enabled as the two features are mutually
2008 * exclusive.
2009 */
2010 nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags &
2011 ZCHECKSUM_FLAG_NOPWRITE) &&
2012 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
2013 }
2014
2015 /*
2016 * All objects in an encrypted objset are protected from modification
2017 * via a MAC. Encrypted objects store their IV and salt in the last DVA
2018 * in the bp, so we cannot use all copies. Encrypted objects are also
2019 * not subject to nopwrite since writing the same data will still
2020 * result in a new ciphertext. Only encrypted blocks can be dedup'd
2021 * to avoid ambiguity in the dedup code since the DDT does not store
2022 * object types.
2023 */
2024 if (os->os_encrypted && (wp & WP_NOFILL) == 0) {
2025 encrypt = B_TRUE;
2026
2027 if (DMU_OT_IS_ENCRYPTED(type)) {
2028 copies = MIN(copies, SPA_DVAS_PER_BP - 1);
2029 nopwrite = B_FALSE;
2030 } else {
2031 dedup = B_FALSE;
2032 }
2033
2034 if (level <= 0 &&
2035 (type == DMU_OT_DNODE || type == DMU_OT_OBJSET)) {
2036 compress = ZIO_COMPRESS_EMPTY;
2037 }
2038 }
2039
2040 zp->zp_compress = compress;
2041 zp->zp_complevel = complevel;
2042 zp->zp_checksum = checksum;
2043 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
2044 zp->zp_level = level;
2045 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
2046 zp->zp_dedup = dedup;
2047 zp->zp_dedup_verify = dedup && dedup_verify;
2048 zp->zp_nopwrite = nopwrite;
2049 zp->zp_encrypt = encrypt;
2050 zp->zp_byteorder = ZFS_HOST_BYTEORDER;
2051 bzero(zp->zp_salt, ZIO_DATA_SALT_LEN);
2052 bzero(zp->zp_iv, ZIO_DATA_IV_LEN);
2053 bzero(zp->zp_mac, ZIO_DATA_MAC_LEN);
2054 zp->zp_zpl_smallblk = DMU_OT_IS_FILE(zp->zp_type) ?
2055 os->os_zpl_special_smallblock : 0;
2056
2057 ASSERT3U(zp->zp_compress, !=, ZIO_COMPRESS_INHERIT);
2058 }
2059
2060 /*
2061 * This function is only called from zfs_holey_common() for zpl_llseek()
2062 * in order to determine the location of holes. In order to accurately
2063 * report holes all dirty data must be synced to disk. This causes extremely
2064 * poor performance when seeking for holes in a dirty file. As a compromise,
2065 * only provide hole data when the dnode is clean. When a dnode is dirty
2066 * report the dnode as having no holes which is always a safe thing to do.
2067 */
2068 int
2069 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
2070 {
2071 dnode_t *dn;
2072 int i, err;
2073 boolean_t clean = B_TRUE;
2074
2075 err = dnode_hold(os, object, FTAG, &dn);
2076 if (err)
2077 return (err);
2078
2079 /*
2080 * Check if dnode is dirty
2081 */
2082 for (i = 0; i < TXG_SIZE; i++) {
2083 if (multilist_link_active(&dn->dn_dirty_link[i])) {
2084 clean = B_FALSE;
2085 break;
2086 }
2087 }
2088
2089 /*
2090 * If compatibility option is on, sync any current changes before
2091 * we go trundling through the block pointers.
2092 */
2093 if (!clean && zfs_dmu_offset_next_sync) {
2094 clean = B_TRUE;
2095 dnode_rele(dn, FTAG);
2096 txg_wait_synced(dmu_objset_pool(os), 0);
2097 err = dnode_hold(os, object, FTAG, &dn);
2098 if (err)
2099 return (err);
2100 }
2101
2102 if (clean)
2103 err = dnode_next_offset(dn,
2104 (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
2105 else
2106 err = SET_ERROR(EBUSY);
2107
2108 dnode_rele(dn, FTAG);
2109
2110 return (err);
2111 }
2112
2113 void
2114 __dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2115 {
2116 dnode_phys_t *dnp = dn->dn_phys;
2117
2118 doi->doi_data_block_size = dn->dn_datablksz;
2119 doi->doi_metadata_block_size = dn->dn_indblkshift ?
2120 1ULL << dn->dn_indblkshift : 0;
2121 doi->doi_type = dn->dn_type;
2122 doi->doi_bonus_type = dn->dn_bonustype;
2123 doi->doi_bonus_size = dn->dn_bonuslen;
2124 doi->doi_dnodesize = dn->dn_num_slots << DNODE_SHIFT;
2125 doi->doi_indirection = dn->dn_nlevels;
2126 doi->doi_checksum = dn->dn_checksum;
2127 doi->doi_compress = dn->dn_compress;
2128 doi->doi_nblkptr = dn->dn_nblkptr;
2129 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
2130 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
2131 doi->doi_fill_count = 0;
2132 for (int i = 0; i < dnp->dn_nblkptr; i++)
2133 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
2134 }
2135
2136 void
2137 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2138 {
2139 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2140 mutex_enter(&dn->dn_mtx);
2141
2142 __dmu_object_info_from_dnode(dn, doi);
2143
2144 mutex_exit(&dn->dn_mtx);
2145 rw_exit(&dn->dn_struct_rwlock);
2146 }
2147
2148 /*
2149 * Get information on a DMU object.
2150 * If doi is NULL, just indicates whether the object exists.
2151 */
2152 int
2153 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
2154 {
2155 dnode_t *dn;
2156 int err = dnode_hold(os, object, FTAG, &dn);
2157
2158 if (err)
2159 return (err);
2160
2161 if (doi != NULL)
2162 dmu_object_info_from_dnode(dn, doi);
2163
2164 dnode_rele(dn, FTAG);
2165 return (0);
2166 }
2167
2168 /*
2169 * As above, but faster; can be used when you have a held dbuf in hand.
2170 */
2171 void
2172 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
2173 {
2174 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2175
2176 DB_DNODE_ENTER(db);
2177 dmu_object_info_from_dnode(DB_DNODE(db), doi);
2178 DB_DNODE_EXIT(db);
2179 }
2180
2181 /*
2182 * Faster still when you only care about the size.
2183 */
2184 void
2185 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
2186 u_longlong_t *nblk512)
2187 {
2188 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2189 dnode_t *dn;
2190
2191 DB_DNODE_ENTER(db);
2192 dn = DB_DNODE(db);
2193
2194 *blksize = dn->dn_datablksz;
2195 /* add in number of slots used for the dnode itself */
2196 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
2197 SPA_MINBLOCKSHIFT) + dn->dn_num_slots;
2198 DB_DNODE_EXIT(db);
2199 }
2200
2201 void
2202 dmu_object_dnsize_from_db(dmu_buf_t *db_fake, int *dnsize)
2203 {
2204 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2205 dnode_t *dn;
2206
2207 DB_DNODE_ENTER(db);
2208 dn = DB_DNODE(db);
2209 *dnsize = dn->dn_num_slots << DNODE_SHIFT;
2210 DB_DNODE_EXIT(db);
2211 }
2212
2213 void
2214 byteswap_uint64_array(void *vbuf, size_t size)
2215 {
2216 uint64_t *buf = vbuf;
2217 size_t count = size >> 3;
2218 int i;
2219
2220 ASSERT((size & 7) == 0);
2221
2222 for (i = 0; i < count; i++)
2223 buf[i] = BSWAP_64(buf[i]);
2224 }
2225
2226 void
2227 byteswap_uint32_array(void *vbuf, size_t size)
2228 {
2229 uint32_t *buf = vbuf;
2230 size_t count = size >> 2;
2231 int i;
2232
2233 ASSERT((size & 3) == 0);
2234
2235 for (i = 0; i < count; i++)
2236 buf[i] = BSWAP_32(buf[i]);
2237 }
2238
2239 void
2240 byteswap_uint16_array(void *vbuf, size_t size)
2241 {
2242 uint16_t *buf = vbuf;
2243 size_t count = size >> 1;
2244 int i;
2245
2246 ASSERT((size & 1) == 0);
2247
2248 for (i = 0; i < count; i++)
2249 buf[i] = BSWAP_16(buf[i]);
2250 }
2251
2252 /* ARGSUSED */
2253 void
2254 byteswap_uint8_array(void *vbuf, size_t size)
2255 {
2256 }
2257
2258 void
2259 dmu_init(void)
2260 {
2261 abd_init();
2262 zfs_dbgmsg_init();
2263 sa_cache_init();
2264 dmu_objset_init();
2265 dnode_init();
2266 zfetch_init();
2267 dmu_tx_init();
2268 l2arc_init();
2269 arc_init();
2270 dbuf_init();
2271 }
2272
2273 void
2274 dmu_fini(void)
2275 {
2276 arc_fini(); /* arc depends on l2arc, so arc must go first */
2277 l2arc_fini();
2278 dmu_tx_fini();
2279 zfetch_fini();
2280 dbuf_fini();
2281 dnode_fini();
2282 dmu_objset_fini();
2283 sa_cache_fini();
2284 zfs_dbgmsg_fini();
2285 abd_fini();
2286 }
2287
2288 EXPORT_SYMBOL(dmu_bonus_hold);
2289 EXPORT_SYMBOL(dmu_bonus_hold_by_dnode);
2290 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus);
2291 EXPORT_SYMBOL(dmu_buf_rele_array);
2292 EXPORT_SYMBOL(dmu_prefetch);
2293 EXPORT_SYMBOL(dmu_free_range);
2294 EXPORT_SYMBOL(dmu_free_long_range);
2295 EXPORT_SYMBOL(dmu_free_long_object);
2296 EXPORT_SYMBOL(dmu_read);
2297 EXPORT_SYMBOL(dmu_read_by_dnode);
2298 EXPORT_SYMBOL(dmu_write);
2299 EXPORT_SYMBOL(dmu_write_by_dnode);
2300 EXPORT_SYMBOL(dmu_prealloc);
2301 EXPORT_SYMBOL(dmu_object_info);
2302 EXPORT_SYMBOL(dmu_object_info_from_dnode);
2303 EXPORT_SYMBOL(dmu_object_info_from_db);
2304 EXPORT_SYMBOL(dmu_object_size_from_db);
2305 EXPORT_SYMBOL(dmu_object_dnsize_from_db);
2306 EXPORT_SYMBOL(dmu_object_set_nlevels);
2307 EXPORT_SYMBOL(dmu_object_set_blocksize);
2308 EXPORT_SYMBOL(dmu_object_set_maxblkid);
2309 EXPORT_SYMBOL(dmu_object_set_checksum);
2310 EXPORT_SYMBOL(dmu_object_set_compress);
2311 EXPORT_SYMBOL(dmu_offset_next);
2312 EXPORT_SYMBOL(dmu_write_policy);
2313 EXPORT_SYMBOL(dmu_sync);
2314 EXPORT_SYMBOL(dmu_request_arcbuf);
2315 EXPORT_SYMBOL(dmu_return_arcbuf);
2316 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dnode);
2317 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dbuf);
2318 EXPORT_SYMBOL(dmu_buf_hold);
2319 EXPORT_SYMBOL(dmu_ot);
2320
2321 /* BEGIN CSTYLED */
2322 ZFS_MODULE_PARAM(zfs, zfs_, nopwrite_enabled, INT, ZMOD_RW,
2323 "Enable NOP writes");
2324
2325 ZFS_MODULE_PARAM(zfs, zfs_, per_txg_dirty_frees_percent, ULONG, ZMOD_RW,
2326 "Percentage of dirtied blocks from frees in one TXG");
2327
2328 ZFS_MODULE_PARAM(zfs, zfs_, dmu_offset_next_sync, INT, ZMOD_RW,
2329 "Enable forcing txg sync to find holes");
2330
2331 ZFS_MODULE_PARAM(zfs, , dmu_prefetch_max, INT, ZMOD_RW,
2332 "Limit one prefetch call to this size");
2333 /* END CSTYLED */
ZFS On Linux mirror for PVE