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