]> 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
OpenZFS 6322 - ZFS indirect block predictive prefetch
[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 #ifdef _KERNEL
51 #include <sys/vmsystm.h>
52 #include <sys/zfs_znode.h>
53 #endif
54
55 /*
56 * Enable/disable nopwrite feature.
57 */
58 int zfs_nopwrite_enabled = 1;
59
60 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
61 { DMU_BSWAP_UINT8, TRUE, "unallocated" },
62 { DMU_BSWAP_ZAP, TRUE, "object directory" },
63 { DMU_BSWAP_UINT64, TRUE, "object array" },
64 { DMU_BSWAP_UINT8, TRUE, "packed nvlist" },
65 { DMU_BSWAP_UINT64, TRUE, "packed nvlist size" },
66 { DMU_BSWAP_UINT64, TRUE, "bpobj" },
67 { DMU_BSWAP_UINT64, TRUE, "bpobj header" },
68 { DMU_BSWAP_UINT64, TRUE, "SPA space map header" },
69 { DMU_BSWAP_UINT64, TRUE, "SPA space map" },
70 { DMU_BSWAP_UINT64, TRUE, "ZIL intent log" },
71 { DMU_BSWAP_DNODE, TRUE, "DMU dnode" },
72 { DMU_BSWAP_OBJSET, TRUE, "DMU objset" },
73 { DMU_BSWAP_UINT64, TRUE, "DSL directory" },
74 { DMU_BSWAP_ZAP, TRUE, "DSL directory child map"},
75 { DMU_BSWAP_ZAP, TRUE, "DSL dataset snap map" },
76 { DMU_BSWAP_ZAP, TRUE, "DSL props" },
77 { DMU_BSWAP_UINT64, TRUE, "DSL dataset" },
78 { DMU_BSWAP_ZNODE, TRUE, "ZFS znode" },
79 { DMU_BSWAP_OLDACL, TRUE, "ZFS V0 ACL" },
80 { DMU_BSWAP_UINT8, FALSE, "ZFS plain file" },
81 { DMU_BSWAP_ZAP, TRUE, "ZFS directory" },
82 { DMU_BSWAP_ZAP, TRUE, "ZFS master node" },
83 { DMU_BSWAP_ZAP, TRUE, "ZFS delete queue" },
84 { DMU_BSWAP_UINT8, FALSE, "zvol object" },
85 { DMU_BSWAP_ZAP, TRUE, "zvol prop" },
86 { DMU_BSWAP_UINT8, FALSE, "other uint8[]" },
87 { DMU_BSWAP_UINT64, FALSE, "other uint64[]" },
88 { DMU_BSWAP_ZAP, TRUE, "other ZAP" },
89 { DMU_BSWAP_ZAP, TRUE, "persistent error log" },
90 { DMU_BSWAP_UINT8, TRUE, "SPA history" },
91 { DMU_BSWAP_UINT64, TRUE, "SPA history offsets" },
92 { DMU_BSWAP_ZAP, TRUE, "Pool properties" },
93 { DMU_BSWAP_ZAP, TRUE, "DSL permissions" },
94 { DMU_BSWAP_ACL, TRUE, "ZFS ACL" },
95 { DMU_BSWAP_UINT8, TRUE, "ZFS SYSACL" },
96 { DMU_BSWAP_UINT8, TRUE, "FUID table" },
97 { DMU_BSWAP_UINT64, TRUE, "FUID table size" },
98 { DMU_BSWAP_ZAP, TRUE, "DSL dataset next clones"},
99 { DMU_BSWAP_ZAP, TRUE, "scan work queue" },
100 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group used" },
101 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group quota" },
102 { DMU_BSWAP_ZAP, TRUE, "snapshot refcount tags"},
103 { DMU_BSWAP_ZAP, TRUE, "DDT ZAP algorithm" },
104 { DMU_BSWAP_ZAP, TRUE, "DDT statistics" },
105 { DMU_BSWAP_UINT8, TRUE, "System attributes" },
106 { DMU_BSWAP_ZAP, TRUE, "SA master node" },
107 { DMU_BSWAP_ZAP, TRUE, "SA attr registration" },
108 { DMU_BSWAP_ZAP, TRUE, "SA attr layouts" },
109 { DMU_BSWAP_ZAP, TRUE, "scan translations" },
110 { DMU_BSWAP_UINT8, FALSE, "deduplicated block" },
111 { DMU_BSWAP_ZAP, TRUE, "DSL deadlist map" },
112 { DMU_BSWAP_UINT64, TRUE, "DSL deadlist map hdr" },
113 { DMU_BSWAP_ZAP, TRUE, "DSL dir clones" },
114 { DMU_BSWAP_UINT64, TRUE, "bpobj subobj" }
115 };
116
117 const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
118 { byteswap_uint8_array, "uint8" },
119 { byteswap_uint16_array, "uint16" },
120 { byteswap_uint32_array, "uint32" },
121 { byteswap_uint64_array, "uint64" },
122 { zap_byteswap, "zap" },
123 { dnode_buf_byteswap, "dnode" },
124 { dmu_objset_byteswap, "objset" },
125 { zfs_znode_byteswap, "znode" },
126 { zfs_oldacl_byteswap, "oldacl" },
127 { zfs_acl_byteswap, "acl" }
128 };
129
130 int
131 dmu_buf_hold_noread_by_dnode(dnode_t *dn, uint64_t offset,
132 void *tag, dmu_buf_t **dbp)
133 {
134 uint64_t blkid;
135 dmu_buf_impl_t *db;
136
137 blkid = dbuf_whichblock(dn, 0, offset);
138 rw_enter(&dn->dn_struct_rwlock, RW_READER);
139 db = dbuf_hold(dn, blkid, tag);
140 rw_exit(&dn->dn_struct_rwlock);
141
142 if (db == NULL) {
143 *dbp = NULL;
144 return (SET_ERROR(EIO));
145 }
146
147 *dbp = &db->db;
148 return (0);
149 }
150 int
151 dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset,
152 void *tag, dmu_buf_t **dbp)
153 {
154 dnode_t *dn;
155 uint64_t blkid;
156 dmu_buf_impl_t *db;
157 int err;
158
159 err = dnode_hold(os, object, FTAG, &dn);
160 if (err)
161 return (err);
162 blkid = dbuf_whichblock(dn, 0, offset);
163 rw_enter(&dn->dn_struct_rwlock, RW_READER);
164 db = dbuf_hold(dn, blkid, tag);
165 rw_exit(&dn->dn_struct_rwlock);
166 dnode_rele(dn, FTAG);
167
168 if (db == NULL) {
169 *dbp = NULL;
170 return (SET_ERROR(EIO));
171 }
172
173 *dbp = &db->db;
174 return (err);
175 }
176
177 int
178 dmu_buf_hold_by_dnode(dnode_t *dn, uint64_t offset,
179 void *tag, dmu_buf_t **dbp, int flags)
180 {
181 int err;
182 int db_flags = DB_RF_CANFAIL;
183
184 if (flags & DMU_READ_NO_PREFETCH)
185 db_flags |= DB_RF_NOPREFETCH;
186
187 err = dmu_buf_hold_noread_by_dnode(dn, offset, tag, dbp);
188 if (err == 0) {
189 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
190 err = dbuf_read(db, NULL, db_flags);
191 if (err != 0) {
192 dbuf_rele(db, tag);
193 *dbp = NULL;
194 }
195 }
196
197 return (err);
198 }
199
200 int
201 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
202 void *tag, dmu_buf_t **dbp, int flags)
203 {
204 int err;
205 int db_flags = DB_RF_CANFAIL;
206
207 if (flags & DMU_READ_NO_PREFETCH)
208 db_flags |= DB_RF_NOPREFETCH;
209
210 err = dmu_buf_hold_noread(os, object, offset, tag, dbp);
211 if (err == 0) {
212 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
213 err = dbuf_read(db, NULL, db_flags);
214 if (err != 0) {
215 dbuf_rele(db, tag);
216 *dbp = NULL;
217 }
218 }
219
220 return (err);
221 }
222
223 int
224 dmu_bonus_max(void)
225 {
226 return (DN_OLD_MAX_BONUSLEN);
227 }
228
229 int
230 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
231 {
232 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
233 dnode_t *dn;
234 int error;
235
236 DB_DNODE_ENTER(db);
237 dn = DB_DNODE(db);
238
239 if (dn->dn_bonus != db) {
240 error = SET_ERROR(EINVAL);
241 } else if (newsize < 0 || newsize > db_fake->db_size) {
242 error = SET_ERROR(EINVAL);
243 } else {
244 dnode_setbonuslen(dn, newsize, tx);
245 error = 0;
246 }
247
248 DB_DNODE_EXIT(db);
249 return (error);
250 }
251
252 int
253 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
254 {
255 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
256 dnode_t *dn;
257 int error;
258
259 DB_DNODE_ENTER(db);
260 dn = DB_DNODE(db);
261
262 if (!DMU_OT_IS_VALID(type)) {
263 error = SET_ERROR(EINVAL);
264 } else if (dn->dn_bonus != db) {
265 error = SET_ERROR(EINVAL);
266 } else {
267 dnode_setbonus_type(dn, type, tx);
268 error = 0;
269 }
270
271 DB_DNODE_EXIT(db);
272 return (error);
273 }
274
275 dmu_object_type_t
276 dmu_get_bonustype(dmu_buf_t *db_fake)
277 {
278 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
279 dnode_t *dn;
280 dmu_object_type_t type;
281
282 DB_DNODE_ENTER(db);
283 dn = DB_DNODE(db);
284 type = dn->dn_bonustype;
285 DB_DNODE_EXIT(db);
286
287 return (type);
288 }
289
290 int
291 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
292 {
293 dnode_t *dn;
294 int error;
295
296 error = dnode_hold(os, object, FTAG, &dn);
297 dbuf_rm_spill(dn, tx);
298 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
299 dnode_rm_spill(dn, tx);
300 rw_exit(&dn->dn_struct_rwlock);
301 dnode_rele(dn, FTAG);
302 return (error);
303 }
304
305 /*
306 * returns ENOENT, EIO, or 0.
307 */
308 int
309 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
310 {
311 dnode_t *dn;
312 dmu_buf_impl_t *db;
313 int error;
314
315 error = dnode_hold(os, object, FTAG, &dn);
316 if (error)
317 return (error);
318
319 rw_enter(&dn->dn_struct_rwlock, RW_READER);
320 if (dn->dn_bonus == NULL) {
321 rw_exit(&dn->dn_struct_rwlock);
322 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
323 if (dn->dn_bonus == NULL)
324 dbuf_create_bonus(dn);
325 }
326 db = dn->dn_bonus;
327
328 /* as long as the bonus buf is held, the dnode will be held */
329 if (refcount_add(&db->db_holds, tag) == 1) {
330 VERIFY(dnode_add_ref(dn, db));
331 atomic_inc_32(&dn->dn_dbufs_count);
332 }
333
334 /*
335 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
336 * hold and incrementing the dbuf count to ensure that dnode_move() sees
337 * a dnode hold for every dbuf.
338 */
339 rw_exit(&dn->dn_struct_rwlock);
340
341 dnode_rele(dn, FTAG);
342
343 VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH));
344
345 *dbp = &db->db;
346 return (0);
347 }
348
349 /*
350 * returns ENOENT, EIO, or 0.
351 *
352 * This interface will allocate a blank spill dbuf when a spill blk
353 * doesn't already exist on the dnode.
354 *
355 * if you only want to find an already existing spill db, then
356 * dmu_spill_hold_existing() should be used.
357 */
358 int
359 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
360 {
361 dmu_buf_impl_t *db = NULL;
362 int err;
363
364 if ((flags & DB_RF_HAVESTRUCT) == 0)
365 rw_enter(&dn->dn_struct_rwlock, RW_READER);
366
367 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
368
369 if ((flags & DB_RF_HAVESTRUCT) == 0)
370 rw_exit(&dn->dn_struct_rwlock);
371
372 ASSERT(db != NULL);
373 err = dbuf_read(db, NULL, flags);
374 if (err == 0)
375 *dbp = &db->db;
376 else
377 dbuf_rele(db, tag);
378 return (err);
379 }
380
381 int
382 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
383 {
384 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
385 dnode_t *dn;
386 int err;
387
388 DB_DNODE_ENTER(db);
389 dn = DB_DNODE(db);
390
391 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
392 err = SET_ERROR(EINVAL);
393 } else {
394 rw_enter(&dn->dn_struct_rwlock, RW_READER);
395
396 if (!dn->dn_have_spill) {
397 err = SET_ERROR(ENOENT);
398 } else {
399 err = dmu_spill_hold_by_dnode(dn,
400 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
401 }
402
403 rw_exit(&dn->dn_struct_rwlock);
404 }
405
406 DB_DNODE_EXIT(db);
407 return (err);
408 }
409
410 int
411 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
412 {
413 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
414 dnode_t *dn;
415 int err;
416
417 DB_DNODE_ENTER(db);
418 dn = DB_DNODE(db);
419 err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp);
420 DB_DNODE_EXIT(db);
421
422 return (err);
423 }
424
425 /*
426 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
427 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
428 * and can induce severe lock contention when writing to several files
429 * whose dnodes are in the same block.
430 */
431 static int
432 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
433 boolean_t read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
434 {
435 dmu_buf_t **dbp;
436 uint64_t blkid, nblks, i;
437 uint32_t dbuf_flags;
438 int err;
439 zio_t *zio;
440
441 ASSERT(length <= DMU_MAX_ACCESS);
442
443 /*
444 * Note: We directly notify the prefetch code of this read, so that
445 * we can tell it about the multi-block read. dbuf_read() only knows
446 * about the one block it is accessing.
447 */
448 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT |
449 DB_RF_NOPREFETCH;
450
451 rw_enter(&dn->dn_struct_rwlock, RW_READER);
452 if (dn->dn_datablkshift) {
453 int blkshift = dn->dn_datablkshift;
454 nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) -
455 P2ALIGN(offset, 1ULL << blkshift)) >> blkshift;
456 } else {
457 if (offset + length > dn->dn_datablksz) {
458 zfs_panic_recover("zfs: accessing past end of object "
459 "%llx/%llx (size=%u access=%llu+%llu)",
460 (longlong_t)dn->dn_objset->
461 os_dsl_dataset->ds_object,
462 (longlong_t)dn->dn_object, dn->dn_datablksz,
463 (longlong_t)offset, (longlong_t)length);
464 rw_exit(&dn->dn_struct_rwlock);
465 return (SET_ERROR(EIO));
466 }
467 nblks = 1;
468 }
469 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
470
471 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
472 blkid = dbuf_whichblock(dn, 0, offset);
473 for (i = 0; i < nblks; i++) {
474 dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag);
475 if (db == NULL) {
476 rw_exit(&dn->dn_struct_rwlock);
477 dmu_buf_rele_array(dbp, nblks, tag);
478 zio_nowait(zio);
479 return (SET_ERROR(EIO));
480 }
481
482 /* initiate async i/o */
483 if (read)
484 (void) dbuf_read(db, zio, dbuf_flags);
485 dbp[i] = &db->db;
486 }
487
488 if ((flags & DMU_READ_NO_PREFETCH) == 0 &&
489 DNODE_META_IS_CACHEABLE(dn) && length <= zfetch_array_rd_sz) {
490 dmu_zfetch(&dn->dn_zfetch, blkid, nblks,
491 read && DNODE_IS_CACHEABLE(dn));
492 }
493 rw_exit(&dn->dn_struct_rwlock);
494
495 /* wait for async i/o */
496 err = zio_wait(zio);
497 if (err) {
498 dmu_buf_rele_array(dbp, nblks, tag);
499 return (err);
500 }
501
502 /* wait for other io to complete */
503 if (read) {
504 for (i = 0; i < nblks; i++) {
505 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
506 mutex_enter(&db->db_mtx);
507 while (db->db_state == DB_READ ||
508 db->db_state == DB_FILL)
509 cv_wait(&db->db_changed, &db->db_mtx);
510 if (db->db_state == DB_UNCACHED)
511 err = SET_ERROR(EIO);
512 mutex_exit(&db->db_mtx);
513 if (err) {
514 dmu_buf_rele_array(dbp, nblks, tag);
515 return (err);
516 }
517 }
518 }
519
520 *numbufsp = nblks;
521 *dbpp = dbp;
522 return (0);
523 }
524
525 static int
526 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
527 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
528 {
529 dnode_t *dn;
530 int err;
531
532 err = dnode_hold(os, object, FTAG, &dn);
533 if (err)
534 return (err);
535
536 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
537 numbufsp, dbpp, DMU_READ_PREFETCH);
538
539 dnode_rele(dn, FTAG);
540
541 return (err);
542 }
543
544 int
545 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
546 uint64_t length, boolean_t read, void *tag, int *numbufsp,
547 dmu_buf_t ***dbpp)
548 {
549 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
550 dnode_t *dn;
551 int err;
552
553 DB_DNODE_ENTER(db);
554 dn = DB_DNODE(db);
555 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
556 numbufsp, dbpp, DMU_READ_PREFETCH);
557 DB_DNODE_EXIT(db);
558
559 return (err);
560 }
561
562 void
563 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
564 {
565 int i;
566 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
567
568 if (numbufs == 0)
569 return;
570
571 for (i = 0; i < numbufs; i++) {
572 if (dbp[i])
573 dbuf_rele(dbp[i], tag);
574 }
575
576 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
577 }
578
579 /*
580 * Issue prefetch i/os for the given blocks. If level is greater than 0, the
581 * indirect blocks prefeteched will be those that point to the blocks containing
582 * the data starting at offset, and continuing to offset + len.
583 *
584 * Note that if the indirect blocks above the blocks being prefetched are not in
585 * cache, they will be asychronously read in.
586 */
587 void
588 dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset,
589 uint64_t len, zio_priority_t pri)
590 {
591 dnode_t *dn;
592 uint64_t blkid;
593 int nblks, err;
594
595 if (len == 0) { /* they're interested in the bonus buffer */
596 dn = DMU_META_DNODE(os);
597
598 if (object == 0 || object >= DN_MAX_OBJECT)
599 return;
600
601 rw_enter(&dn->dn_struct_rwlock, RW_READER);
602 blkid = dbuf_whichblock(dn, level,
603 object * sizeof (dnode_phys_t));
604 dbuf_prefetch(dn, level, blkid, pri, 0);
605 rw_exit(&dn->dn_struct_rwlock);
606 return;
607 }
608
609 /*
610 * XXX - Note, if the dnode for the requested object is not
611 * already cached, we will do a *synchronous* read in the
612 * dnode_hold() call. The same is true for any indirects.
613 */
614 err = dnode_hold(os, object, FTAG, &dn);
615 if (err != 0)
616 return;
617
618 rw_enter(&dn->dn_struct_rwlock, RW_READER);
619 /*
620 * offset + len - 1 is the last byte we want to prefetch for, and offset
621 * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the
622 * last block we want to prefetch, and dbuf_whichblock(dn, level,
623 * offset) is the first. Then the number we need to prefetch is the
624 * last - first + 1.
625 */
626 if (level > 0 || dn->dn_datablkshift != 0) {
627 nblks = dbuf_whichblock(dn, level, offset + len - 1) -
628 dbuf_whichblock(dn, level, offset) + 1;
629 } else {
630 nblks = (offset < dn->dn_datablksz);
631 }
632
633 if (nblks != 0) {
634 int i;
635
636 blkid = dbuf_whichblock(dn, level, offset);
637 for (i = 0; i < nblks; i++)
638 dbuf_prefetch(dn, level, blkid + i, pri, 0);
639 }
640
641 rw_exit(&dn->dn_struct_rwlock);
642
643 dnode_rele(dn, FTAG);
644 }
645
646 /*
647 * Get the next "chunk" of file data to free. We traverse the file from
648 * the end so that the file gets shorter over time (if we crashes in the
649 * middle, this will leave us in a better state). We find allocated file
650 * data by simply searching the allocated level 1 indirects.
651 *
652 * On input, *start should be the first offset that does not need to be
653 * freed (e.g. "offset + length"). On return, *start will be the first
654 * offset that should be freed.
655 */
656 static int
657 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum)
658 {
659 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
660 /* bytes of data covered by a level-1 indirect block */
661 uint64_t iblkrange =
662 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
663 uint64_t blks;
664
665 ASSERT3U(minimum, <=, *start);
666
667 if (*start - minimum <= iblkrange * maxblks) {
668 *start = minimum;
669 return (0);
670 }
671 ASSERT(ISP2(iblkrange));
672
673 for (blks = 0; *start > minimum && blks < maxblks; blks++) {
674 int err;
675
676 /*
677 * dnode_next_offset(BACKWARDS) will find an allocated L1
678 * indirect block at or before the input offset. We must
679 * decrement *start so that it is at the end of the region
680 * to search.
681 */
682 (*start)--;
683 err = dnode_next_offset(dn,
684 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
685
686 /* if there are no indirect blocks before start, we are done */
687 if (err == ESRCH) {
688 *start = minimum;
689 break;
690 } else if (err != 0) {
691 return (err);
692 }
693
694 /* set start to the beginning of this L1 indirect */
695 *start = P2ALIGN(*start, iblkrange);
696 }
697 if (*start < minimum)
698 *start = minimum;
699 return (0);
700 }
701
702 static int
703 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
704 uint64_t length)
705 {
706 uint64_t object_size;
707 int err;
708
709 if (dn == NULL)
710 return (SET_ERROR(EINVAL));
711
712 object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
713 if (offset >= object_size)
714 return (0);
715
716 if (length == DMU_OBJECT_END || offset + length > object_size)
717 length = object_size - offset;
718
719 while (length != 0) {
720 uint64_t chunk_end, chunk_begin;
721 dmu_tx_t *tx;
722
723 chunk_end = chunk_begin = offset + length;
724
725 /* move chunk_begin backwards to the beginning of this chunk */
726 err = get_next_chunk(dn, &chunk_begin, offset);
727 if (err)
728 return (err);
729 ASSERT3U(chunk_begin, >=, offset);
730 ASSERT3U(chunk_begin, <=, chunk_end);
731
732 tx = dmu_tx_create(os);
733 dmu_tx_hold_free(tx, dn->dn_object,
734 chunk_begin, chunk_end - chunk_begin);
735
736 /*
737 * Mark this transaction as typically resulting in a net
738 * reduction in space used.
739 */
740 dmu_tx_mark_netfree(tx);
741 err = dmu_tx_assign(tx, TXG_WAIT);
742 if (err) {
743 dmu_tx_abort(tx);
744 return (err);
745 }
746 dnode_free_range(dn, chunk_begin, chunk_end - chunk_begin, tx);
747 dmu_tx_commit(tx);
748
749 length -= chunk_end - chunk_begin;
750 }
751 return (0);
752 }
753
754 int
755 dmu_free_long_range(objset_t *os, uint64_t object,
756 uint64_t offset, uint64_t length)
757 {
758 dnode_t *dn;
759 int err;
760
761 err = dnode_hold(os, object, FTAG, &dn);
762 if (err != 0)
763 return (err);
764 err = dmu_free_long_range_impl(os, dn, offset, length);
765
766 /*
767 * It is important to zero out the maxblkid when freeing the entire
768 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
769 * will take the fast path, and (b) dnode_reallocate() can verify
770 * that the entire file has been freed.
771 */
772 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
773 dn->dn_maxblkid = 0;
774
775 dnode_rele(dn, FTAG);
776 return (err);
777 }
778
779 int
780 dmu_free_long_object(objset_t *os, uint64_t object)
781 {
782 dmu_tx_t *tx;
783 int err;
784
785 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
786 if (err != 0)
787 return (err);
788
789 tx = dmu_tx_create(os);
790 dmu_tx_hold_bonus(tx, object);
791 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
792 dmu_tx_mark_netfree(tx);
793 err = dmu_tx_assign(tx, TXG_WAIT);
794 if (err == 0) {
795 err = dmu_object_free(os, object, tx);
796 dmu_tx_commit(tx);
797 } else {
798 dmu_tx_abort(tx);
799 }
800
801 return (err);
802 }
803
804 int
805 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
806 uint64_t size, dmu_tx_t *tx)
807 {
808 dnode_t *dn;
809 int err = dnode_hold(os, object, FTAG, &dn);
810 if (err)
811 return (err);
812 ASSERT(offset < UINT64_MAX);
813 ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
814 dnode_free_range(dn, offset, size, tx);
815 dnode_rele(dn, FTAG);
816 return (0);
817 }
818
819 int
820 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
821 void *buf, uint32_t flags)
822 {
823 dnode_t *dn;
824 dmu_buf_t **dbp;
825 int numbufs, err;
826
827 err = dnode_hold(os, object, FTAG, &dn);
828 if (err)
829 return (err);
830
831 /*
832 * Deal with odd block sizes, where there can't be data past the first
833 * block. If we ever do the tail block optimization, we will need to
834 * handle that here as well.
835 */
836 if (dn->dn_maxblkid == 0) {
837 uint64_t newsz = offset > dn->dn_datablksz ? 0 :
838 MIN(size, dn->dn_datablksz - offset);
839 bzero((char *)buf + newsz, size - newsz);
840 size = newsz;
841 }
842
843 while (size > 0) {
844 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
845 int i;
846
847 /*
848 * NB: we could do this block-at-a-time, but it's nice
849 * to be reading in parallel.
850 */
851 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
852 TRUE, FTAG, &numbufs, &dbp, flags);
853 if (err)
854 break;
855
856 for (i = 0; i < numbufs; i++) {
857 uint64_t tocpy;
858 int64_t bufoff;
859 dmu_buf_t *db = dbp[i];
860
861 ASSERT(size > 0);
862
863 bufoff = offset - db->db_offset;
864 tocpy = MIN(db->db_size - bufoff, size);
865
866 (void) memcpy(buf, (char *)db->db_data + bufoff, tocpy);
867
868 offset += tocpy;
869 size -= tocpy;
870 buf = (char *)buf + tocpy;
871 }
872 dmu_buf_rele_array(dbp, numbufs, FTAG);
873 }
874 dnode_rele(dn, FTAG);
875 return (err);
876 }
877
878 void
879 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
880 const void *buf, dmu_tx_t *tx)
881 {
882 dmu_buf_t **dbp;
883 int numbufs, i;
884
885 if (size == 0)
886 return;
887
888 VERIFY0(dmu_buf_hold_array(os, object, offset, size,
889 FALSE, FTAG, &numbufs, &dbp));
890
891 for (i = 0; i < numbufs; i++) {
892 uint64_t tocpy;
893 int64_t bufoff;
894 dmu_buf_t *db = dbp[i];
895
896 ASSERT(size > 0);
897
898 bufoff = offset - db->db_offset;
899 tocpy = MIN(db->db_size - bufoff, size);
900
901 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
902
903 if (tocpy == db->db_size)
904 dmu_buf_will_fill(db, tx);
905 else
906 dmu_buf_will_dirty(db, tx);
907
908 (void) memcpy((char *)db->db_data + bufoff, buf, tocpy);
909
910 if (tocpy == db->db_size)
911 dmu_buf_fill_done(db, tx);
912
913 offset += tocpy;
914 size -= tocpy;
915 buf = (char *)buf + tocpy;
916 }
917 dmu_buf_rele_array(dbp, numbufs, FTAG);
918 }
919
920 void
921 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
922 dmu_tx_t *tx)
923 {
924 dmu_buf_t **dbp;
925 int numbufs, i;
926
927 if (size == 0)
928 return;
929
930 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
931 FALSE, FTAG, &numbufs, &dbp));
932
933 for (i = 0; i < numbufs; i++) {
934 dmu_buf_t *db = dbp[i];
935
936 dmu_buf_will_not_fill(db, tx);
937 }
938 dmu_buf_rele_array(dbp, numbufs, FTAG);
939 }
940
941 void
942 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
943 void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
944 int compressed_size, int byteorder, dmu_tx_t *tx)
945 {
946 dmu_buf_t *db;
947
948 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
949 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
950 VERIFY0(dmu_buf_hold_noread(os, object, offset,
951 FTAG, &db));
952
953 dmu_buf_write_embedded(db,
954 data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
955 uncompressed_size, compressed_size, byteorder, tx);
956
957 dmu_buf_rele(db, FTAG);
958 }
959
960 /*
961 * DMU support for xuio
962 */
963 kstat_t *xuio_ksp = NULL;
964
965 typedef struct xuio_stats {
966 /* loaned yet not returned arc_buf */
967 kstat_named_t xuiostat_onloan_rbuf;
968 kstat_named_t xuiostat_onloan_wbuf;
969 /* whether a copy is made when loaning out a read buffer */
970 kstat_named_t xuiostat_rbuf_copied;
971 kstat_named_t xuiostat_rbuf_nocopy;
972 /* whether a copy is made when assigning a write buffer */
973 kstat_named_t xuiostat_wbuf_copied;
974 kstat_named_t xuiostat_wbuf_nocopy;
975 } xuio_stats_t;
976
977 static xuio_stats_t xuio_stats = {
978 { "onloan_read_buf", KSTAT_DATA_UINT64 },
979 { "onloan_write_buf", KSTAT_DATA_UINT64 },
980 { "read_buf_copied", KSTAT_DATA_UINT64 },
981 { "read_buf_nocopy", KSTAT_DATA_UINT64 },
982 { "write_buf_copied", KSTAT_DATA_UINT64 },
983 { "write_buf_nocopy", KSTAT_DATA_UINT64 }
984 };
985
986 #define XUIOSTAT_INCR(stat, val) \
987 atomic_add_64(&xuio_stats.stat.value.ui64, (val))
988 #define XUIOSTAT_BUMP(stat) XUIOSTAT_INCR(stat, 1)
989
990 int
991 dmu_xuio_init(xuio_t *xuio, int nblk)
992 {
993 dmu_xuio_t *priv;
994 uio_t *uio = &xuio->xu_uio;
995
996 uio->uio_iovcnt = nblk;
997 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
998
999 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
1000 priv->cnt = nblk;
1001 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
1002 priv->iovp = (iovec_t *)uio->uio_iov;
1003 XUIO_XUZC_PRIV(xuio) = priv;
1004
1005 if (XUIO_XUZC_RW(xuio) == UIO_READ)
1006 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
1007 else
1008 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
1009
1010 return (0);
1011 }
1012
1013 void
1014 dmu_xuio_fini(xuio_t *xuio)
1015 {
1016 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1017 int nblk = priv->cnt;
1018
1019 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
1020 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
1021 kmem_free(priv, sizeof (dmu_xuio_t));
1022
1023 if (XUIO_XUZC_RW(xuio) == UIO_READ)
1024 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
1025 else
1026 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
1027 }
1028
1029 /*
1030 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
1031 * and increase priv->next by 1.
1032 */
1033 int
1034 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
1035 {
1036 struct iovec *iov;
1037 uio_t *uio = &xuio->xu_uio;
1038 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1039 int i = priv->next++;
1040
1041 ASSERT(i < priv->cnt);
1042 ASSERT(off + n <= arc_buf_size(abuf));
1043 iov = (iovec_t *)uio->uio_iov + i;
1044 iov->iov_base = (char *)abuf->b_data + off;
1045 iov->iov_len = n;
1046 priv->bufs[i] = abuf;
1047 return (0);
1048 }
1049
1050 int
1051 dmu_xuio_cnt(xuio_t *xuio)
1052 {
1053 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1054 return (priv->cnt);
1055 }
1056
1057 arc_buf_t *
1058 dmu_xuio_arcbuf(xuio_t *xuio, int i)
1059 {
1060 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1061
1062 ASSERT(i < priv->cnt);
1063 return (priv->bufs[i]);
1064 }
1065
1066 void
1067 dmu_xuio_clear(xuio_t *xuio, int i)
1068 {
1069 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1070
1071 ASSERT(i < priv->cnt);
1072 priv->bufs[i] = NULL;
1073 }
1074
1075 static void
1076 xuio_stat_init(void)
1077 {
1078 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
1079 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
1080 KSTAT_FLAG_VIRTUAL);
1081 if (xuio_ksp != NULL) {
1082 xuio_ksp->ks_data = &xuio_stats;
1083 kstat_install(xuio_ksp);
1084 }
1085 }
1086
1087 static void
1088 xuio_stat_fini(void)
1089 {
1090 if (xuio_ksp != NULL) {
1091 kstat_delete(xuio_ksp);
1092 xuio_ksp = NULL;
1093 }
1094 }
1095
1096 void
1097 xuio_stat_wbuf_copied()
1098 {
1099 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1100 }
1101
1102 void
1103 xuio_stat_wbuf_nocopy()
1104 {
1105 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
1106 }
1107
1108 #ifdef _KERNEL
1109 static int
1110 dmu_read_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size)
1111 {
1112 dmu_buf_t **dbp;
1113 int numbufs, i, err;
1114 xuio_t *xuio = NULL;
1115
1116 /*
1117 * NB: we could do this block-at-a-time, but it's nice
1118 * to be reading in parallel.
1119 */
1120 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1121 TRUE, FTAG, &numbufs, &dbp, 0);
1122 if (err)
1123 return (err);
1124
1125 for (i = 0; i < numbufs; i++) {
1126 uint64_t tocpy;
1127 int64_t bufoff;
1128 dmu_buf_t *db = dbp[i];
1129
1130 ASSERT(size > 0);
1131
1132 bufoff = uio->uio_loffset - db->db_offset;
1133 tocpy = MIN(db->db_size - bufoff, size);
1134
1135 if (xuio) {
1136 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
1137 arc_buf_t *dbuf_abuf = dbi->db_buf;
1138 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
1139 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
1140 if (!err) {
1141 uio->uio_resid -= tocpy;
1142 uio->uio_loffset += tocpy;
1143 }
1144
1145 if (abuf == dbuf_abuf)
1146 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
1147 else
1148 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
1149 } else {
1150 err = uiomove((char *)db->db_data + bufoff, tocpy,
1151 UIO_READ, uio);
1152 }
1153 if (err)
1154 break;
1155
1156 size -= tocpy;
1157 }
1158 dmu_buf_rele_array(dbp, numbufs, FTAG);
1159
1160 return (err);
1161 }
1162
1163 /*
1164 * Read 'size' bytes into the uio buffer.
1165 * From object zdb->db_object.
1166 * Starting at offset uio->uio_loffset.
1167 *
1168 * If the caller already has a dbuf in the target object
1169 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1170 * because we don't have to find the dnode_t for the object.
1171 */
1172 int
1173 dmu_read_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size)
1174 {
1175 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1176 dnode_t *dn;
1177 int err;
1178
1179 if (size == 0)
1180 return (0);
1181
1182 DB_DNODE_ENTER(db);
1183 dn = DB_DNODE(db);
1184 err = dmu_read_uio_dnode(dn, uio, size);
1185 DB_DNODE_EXIT(db);
1186
1187 return (err);
1188 }
1189
1190 /*
1191 * Read 'size' bytes into the uio buffer.
1192 * From the specified object
1193 * Starting at offset uio->uio_loffset.
1194 */
1195 int
1196 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
1197 {
1198 dnode_t *dn;
1199 int err;
1200
1201 if (size == 0)
1202 return (0);
1203
1204 err = dnode_hold(os, object, FTAG, &dn);
1205 if (err)
1206 return (err);
1207
1208 err = dmu_read_uio_dnode(dn, uio, size);
1209
1210 dnode_rele(dn, FTAG);
1211
1212 return (err);
1213 }
1214
1215 static int
1216 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1217 {
1218 dmu_buf_t **dbp;
1219 int numbufs;
1220 int err = 0;
1221 int i;
1222
1223 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1224 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1225 if (err)
1226 return (err);
1227
1228 for (i = 0; i < numbufs; i++) {
1229 uint64_t tocpy;
1230 int64_t bufoff;
1231 dmu_buf_t *db = dbp[i];
1232
1233 ASSERT(size > 0);
1234
1235 bufoff = uio->uio_loffset - db->db_offset;
1236 tocpy = MIN(db->db_size - bufoff, size);
1237
1238 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1239
1240 if (tocpy == db->db_size)
1241 dmu_buf_will_fill(db, tx);
1242 else
1243 dmu_buf_will_dirty(db, tx);
1244
1245 /*
1246 * XXX uiomove could block forever (eg.nfs-backed
1247 * pages). There needs to be a uiolockdown() function
1248 * to lock the pages in memory, so that uiomove won't
1249 * block.
1250 */
1251 err = uiomove((char *)db->db_data + bufoff, tocpy,
1252 UIO_WRITE, uio);
1253
1254 if (tocpy == db->db_size)
1255 dmu_buf_fill_done(db, tx);
1256
1257 if (err)
1258 break;
1259
1260 size -= tocpy;
1261 }
1262
1263 dmu_buf_rele_array(dbp, numbufs, FTAG);
1264 return (err);
1265 }
1266
1267 /*
1268 * Write 'size' bytes from the uio buffer.
1269 * To object zdb->db_object.
1270 * Starting at offset uio->uio_loffset.
1271 *
1272 * If the caller already has a dbuf in the target object
1273 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1274 * because we don't have to find the dnode_t for the object.
1275 */
1276 int
1277 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1278 dmu_tx_t *tx)
1279 {
1280 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1281 dnode_t *dn;
1282 int err;
1283
1284 if (size == 0)
1285 return (0);
1286
1287 DB_DNODE_ENTER(db);
1288 dn = DB_DNODE(db);
1289 err = dmu_write_uio_dnode(dn, uio, size, tx);
1290 DB_DNODE_EXIT(db);
1291
1292 return (err);
1293 }
1294
1295 /*
1296 * Write 'size' bytes from the uio buffer.
1297 * To the specified object.
1298 * Starting at offset uio->uio_loffset.
1299 */
1300 int
1301 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1302 dmu_tx_t *tx)
1303 {
1304 dnode_t *dn;
1305 int err;
1306
1307 if (size == 0)
1308 return (0);
1309
1310 err = dnode_hold(os, object, FTAG, &dn);
1311 if (err)
1312 return (err);
1313
1314 err = dmu_write_uio_dnode(dn, uio, size, tx);
1315
1316 dnode_rele(dn, FTAG);
1317
1318 return (err);
1319 }
1320 #endif /* _KERNEL */
1321
1322 /*
1323 * Allocate a loaned anonymous arc buffer.
1324 */
1325 arc_buf_t *
1326 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1327 {
1328 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1329
1330 return (arc_loan_buf(db->db_objset->os_spa, size));
1331 }
1332
1333 /*
1334 * Free a loaned arc buffer.
1335 */
1336 void
1337 dmu_return_arcbuf(arc_buf_t *buf)
1338 {
1339 arc_return_buf(buf, FTAG);
1340 VERIFY(arc_buf_remove_ref(buf, FTAG));
1341 }
1342
1343 /*
1344 * When possible directly assign passed loaned arc buffer to a dbuf.
1345 * If this is not possible copy the contents of passed arc buf via
1346 * dmu_write().
1347 */
1348 void
1349 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1350 dmu_tx_t *tx)
1351 {
1352 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1353 dnode_t *dn;
1354 dmu_buf_impl_t *db;
1355 uint32_t blksz = (uint32_t)arc_buf_size(buf);
1356 uint64_t blkid;
1357
1358 DB_DNODE_ENTER(dbuf);
1359 dn = DB_DNODE(dbuf);
1360 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1361 blkid = dbuf_whichblock(dn, 0, offset);
1362 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1363 rw_exit(&dn->dn_struct_rwlock);
1364 DB_DNODE_EXIT(dbuf);
1365
1366 /*
1367 * We can only assign if the offset is aligned, the arc buf is the
1368 * same size as the dbuf, and the dbuf is not metadata. It
1369 * can't be metadata because the loaned arc buf comes from the
1370 * user-data kmem area.
1371 */
1372 if (offset == db->db.db_offset && blksz == db->db.db_size &&
1373 DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA) {
1374 dbuf_assign_arcbuf(db, buf, tx);
1375 dbuf_rele(db, FTAG);
1376 } else {
1377 objset_t *os;
1378 uint64_t object;
1379
1380 DB_DNODE_ENTER(dbuf);
1381 dn = DB_DNODE(dbuf);
1382 os = dn->dn_objset;
1383 object = dn->dn_object;
1384 DB_DNODE_EXIT(dbuf);
1385
1386 dbuf_rele(db, FTAG);
1387 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1388 dmu_return_arcbuf(buf);
1389 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1390 }
1391 }
1392
1393 typedef struct {
1394 dbuf_dirty_record_t *dsa_dr;
1395 dmu_sync_cb_t *dsa_done;
1396 zgd_t *dsa_zgd;
1397 dmu_tx_t *dsa_tx;
1398 } dmu_sync_arg_t;
1399
1400 /* ARGSUSED */
1401 static void
1402 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1403 {
1404 dmu_sync_arg_t *dsa = varg;
1405 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1406 blkptr_t *bp = zio->io_bp;
1407
1408 if (zio->io_error == 0) {
1409 if (BP_IS_HOLE(bp)) {
1410 /*
1411 * A block of zeros may compress to a hole, but the
1412 * block size still needs to be known for replay.
1413 */
1414 BP_SET_LSIZE(bp, db->db_size);
1415 } else if (!BP_IS_EMBEDDED(bp)) {
1416 ASSERT(BP_GET_LEVEL(bp) == 0);
1417 bp->blk_fill = 1;
1418 }
1419 }
1420 }
1421
1422 static void
1423 dmu_sync_late_arrival_ready(zio_t *zio)
1424 {
1425 dmu_sync_ready(zio, NULL, zio->io_private);
1426 }
1427
1428 /* ARGSUSED */
1429 static void
1430 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1431 {
1432 dmu_sync_arg_t *dsa = varg;
1433 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1434 dmu_buf_impl_t *db = dr->dr_dbuf;
1435
1436 mutex_enter(&db->db_mtx);
1437 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1438 if (zio->io_error == 0) {
1439 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1440 if (dr->dt.dl.dr_nopwrite) {
1441 ASSERTV(blkptr_t *bp = zio->io_bp);
1442 ASSERTV(blkptr_t *bp_orig = &zio->io_bp_orig);
1443 ASSERTV(uint8_t chksum = BP_GET_CHECKSUM(bp_orig));
1444
1445 ASSERT(BP_EQUAL(bp, bp_orig));
1446 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1447 ASSERT(zio_checksum_table[chksum].ci_dedup);
1448 }
1449 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1450 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1451 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1452
1453 /*
1454 * Old style holes are filled with all zeros, whereas
1455 * new-style holes maintain their lsize, type, level,
1456 * and birth time (see zio_write_compress). While we
1457 * need to reset the BP_SET_LSIZE() call that happened
1458 * in dmu_sync_ready for old style holes, we do *not*
1459 * want to wipe out the information contained in new
1460 * style holes. Thus, only zero out the block pointer if
1461 * it's an old style hole.
1462 */
1463 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) &&
1464 dr->dt.dl.dr_overridden_by.blk_birth == 0)
1465 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1466 } else {
1467 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1468 }
1469 cv_broadcast(&db->db_changed);
1470 mutex_exit(&db->db_mtx);
1471
1472 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1473
1474 kmem_free(dsa, sizeof (*dsa));
1475 }
1476
1477 static void
1478 dmu_sync_late_arrival_done(zio_t *zio)
1479 {
1480 blkptr_t *bp = zio->io_bp;
1481 dmu_sync_arg_t *dsa = zio->io_private;
1482 ASSERTV(blkptr_t *bp_orig = &zio->io_bp_orig);
1483
1484 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1485 /*
1486 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE)
1487 * then there is nothing to do here. Otherwise, free the
1488 * newly allocated block in this txg.
1489 */
1490 if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
1491 ASSERT(BP_EQUAL(bp, bp_orig));
1492 } else {
1493 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1494 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1495 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1496 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1497 }
1498 }
1499
1500 dmu_tx_commit(dsa->dsa_tx);
1501
1502 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1503
1504 kmem_free(dsa, sizeof (*dsa));
1505 }
1506
1507 static int
1508 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1509 zio_prop_t *zp, zbookmark_phys_t *zb)
1510 {
1511 dmu_sync_arg_t *dsa;
1512 dmu_tx_t *tx;
1513
1514 tx = dmu_tx_create(os);
1515 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1516 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1517 dmu_tx_abort(tx);
1518 /* Make zl_get_data do txg_waited_synced() */
1519 return (SET_ERROR(EIO));
1520 }
1521
1522 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1523 dsa->dsa_dr = NULL;
1524 dsa->dsa_done = done;
1525 dsa->dsa_zgd = zgd;
1526 dsa->dsa_tx = tx;
1527
1528 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx),
1529 zgd->zgd_bp, zgd->zgd_db->db_data, zgd->zgd_db->db_size,
1530 zp, dmu_sync_late_arrival_ready, NULL,
1531 NULL, dmu_sync_late_arrival_done, dsa, ZIO_PRIORITY_SYNC_WRITE,
1532 ZIO_FLAG_CANFAIL, zb));
1533
1534 return (0);
1535 }
1536
1537 /*
1538 * Intent log support: sync the block associated with db to disk.
1539 * N.B. and XXX: the caller is responsible for making sure that the
1540 * data isn't changing while dmu_sync() is writing it.
1541 *
1542 * Return values:
1543 *
1544 * EEXIST: this txg has already been synced, so there's nothing to do.
1545 * The caller should not log the write.
1546 *
1547 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1548 * The caller should not log the write.
1549 *
1550 * EALREADY: this block is already in the process of being synced.
1551 * The caller should track its progress (somehow).
1552 *
1553 * EIO: could not do the I/O.
1554 * The caller should do a txg_wait_synced().
1555 *
1556 * 0: the I/O has been initiated.
1557 * The caller should log this blkptr in the done callback.
1558 * It is possible that the I/O will fail, in which case
1559 * the error will be reported to the done callback and
1560 * propagated to pio from zio_done().
1561 */
1562 int
1563 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1564 {
1565 blkptr_t *bp = zgd->zgd_bp;
1566 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1567 objset_t *os = db->db_objset;
1568 dsl_dataset_t *ds = os->os_dsl_dataset;
1569 dbuf_dirty_record_t *dr;
1570 dmu_sync_arg_t *dsa;
1571 zbookmark_phys_t zb;
1572 zio_prop_t zp;
1573 dnode_t *dn;
1574
1575 ASSERT(pio != NULL);
1576 ASSERT(txg != 0);
1577
1578 SET_BOOKMARK(&zb, ds->ds_object,
1579 db->db.db_object, db->db_level, db->db_blkid);
1580
1581 DB_DNODE_ENTER(db);
1582 dn = DB_DNODE(db);
1583 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1584 DB_DNODE_EXIT(db);
1585
1586 /*
1587 * If we're frozen (running ziltest), we always need to generate a bp.
1588 */
1589 if (txg > spa_freeze_txg(os->os_spa))
1590 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1591
1592 /*
1593 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1594 * and us. If we determine that this txg is not yet syncing,
1595 * but it begins to sync a moment later, that's OK because the
1596 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1597 */
1598 mutex_enter(&db->db_mtx);
1599
1600 if (txg <= spa_last_synced_txg(os->os_spa)) {
1601 /*
1602 * This txg has already synced. There's nothing to do.
1603 */
1604 mutex_exit(&db->db_mtx);
1605 return (SET_ERROR(EEXIST));
1606 }
1607
1608 if (txg <= spa_syncing_txg(os->os_spa)) {
1609 /*
1610 * This txg is currently syncing, so we can't mess with
1611 * the dirty record anymore; just write a new log block.
1612 */
1613 mutex_exit(&db->db_mtx);
1614 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1615 }
1616
1617 dr = db->db_last_dirty;
1618 while (dr && dr->dr_txg != txg)
1619 dr = dr->dr_next;
1620
1621 if (dr == NULL) {
1622 /*
1623 * There's no dr for this dbuf, so it must have been freed.
1624 * There's no need to log writes to freed blocks, so we're done.
1625 */
1626 mutex_exit(&db->db_mtx);
1627 return (SET_ERROR(ENOENT));
1628 }
1629
1630 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
1631
1632 /*
1633 * Assume the on-disk data is X, the current syncing data (in
1634 * txg - 1) is Y, and the current in-memory data is Z (currently
1635 * in dmu_sync).
1636 *
1637 * We usually want to perform a nopwrite if X and Z are the
1638 * same. However, if Y is different (i.e. the BP is going to
1639 * change before this write takes effect), then a nopwrite will
1640 * be incorrect - we would override with X, which could have
1641 * been freed when Y was written.
1642 *
1643 * (Note that this is not a concern when we are nop-writing from
1644 * syncing context, because X and Y must be identical, because
1645 * all previous txgs have been synced.)
1646 *
1647 * Therefore, we disable nopwrite if the current BP could change
1648 * before this TXG. There are two ways it could change: by
1649 * being dirty (dr_next is non-NULL), or by being freed
1650 * (dnode_block_freed()). This behavior is verified by
1651 * zio_done(), which VERIFYs that the override BP is identical
1652 * to the on-disk BP.
1653 */
1654 DB_DNODE_ENTER(db);
1655 dn = DB_DNODE(db);
1656 if (dr->dr_next != NULL || dnode_block_freed(dn, db->db_blkid))
1657 zp.zp_nopwrite = B_FALSE;
1658 DB_DNODE_EXIT(db);
1659
1660 ASSERT(dr->dr_txg == txg);
1661 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1662 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1663 /*
1664 * We have already issued a sync write for this buffer,
1665 * or this buffer has already been synced. It could not
1666 * have been dirtied since, or we would have cleared the state.
1667 */
1668 mutex_exit(&db->db_mtx);
1669 return (SET_ERROR(EALREADY));
1670 }
1671
1672 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1673 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1674 mutex_exit(&db->db_mtx);
1675
1676 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1677 dsa->dsa_dr = dr;
1678 dsa->dsa_done = done;
1679 dsa->dsa_zgd = zgd;
1680 dsa->dsa_tx = NULL;
1681
1682 zio_nowait(arc_write(pio, os->os_spa, txg,
1683 bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
1684 DBUF_IS_L2COMPRESSIBLE(db), &zp, dmu_sync_ready,
1685 NULL, NULL, dmu_sync_done, dsa,
1686 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb));
1687
1688 return (0);
1689 }
1690
1691 int
1692 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1693 dmu_tx_t *tx)
1694 {
1695 dnode_t *dn;
1696 int err;
1697
1698 err = dnode_hold(os, object, FTAG, &dn);
1699 if (err)
1700 return (err);
1701 err = dnode_set_blksz(dn, size, ibs, tx);
1702 dnode_rele(dn, FTAG);
1703 return (err);
1704 }
1705
1706 void
1707 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1708 dmu_tx_t *tx)
1709 {
1710 dnode_t *dn;
1711
1712 /*
1713 * Send streams include each object's checksum function. This
1714 * check ensures that the receiving system can understand the
1715 * checksum function transmitted.
1716 */
1717 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS);
1718
1719 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1720 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS);
1721 dn->dn_checksum = checksum;
1722 dnode_setdirty(dn, tx);
1723 dnode_rele(dn, FTAG);
1724 }
1725
1726 void
1727 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1728 dmu_tx_t *tx)
1729 {
1730 dnode_t *dn;
1731
1732 /*
1733 * Send streams include each object's compression function. This
1734 * check ensures that the receiving system can understand the
1735 * compression function transmitted.
1736 */
1737 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS);
1738
1739 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1740 dn->dn_compress = compress;
1741 dnode_setdirty(dn, tx);
1742 dnode_rele(dn, FTAG);
1743 }
1744
1745 int zfs_mdcomp_disable = 0;
1746
1747 /*
1748 * When the "redundant_metadata" property is set to "most", only indirect
1749 * blocks of this level and higher will have an additional ditto block.
1750 */
1751 int zfs_redundant_metadata_most_ditto_level = 2;
1752
1753 void
1754 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
1755 {
1756 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
1757 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
1758 (wp & WP_SPILL));
1759 enum zio_checksum checksum = os->os_checksum;
1760 enum zio_compress compress = os->os_compress;
1761 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
1762 boolean_t dedup = B_FALSE;
1763 boolean_t nopwrite = B_FALSE;
1764 boolean_t dedup_verify = os->os_dedup_verify;
1765 int copies = os->os_copies;
1766
1767 /*
1768 * We maintain different write policies for each of the following
1769 * types of data:
1770 * 1. metadata
1771 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1772 * 3. all other level 0 blocks
1773 */
1774 if (ismd) {
1775 if (zfs_mdcomp_disable) {
1776 compress = ZIO_COMPRESS_EMPTY;
1777 } else {
1778 /*
1779 * XXX -- we should design a compression algorithm
1780 * that specializes in arrays of bps.
1781 */
1782 compress = zio_compress_select(os->os_spa,
1783 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON);
1784 }
1785
1786 /*
1787 * Metadata always gets checksummed. If the data
1788 * checksum is multi-bit correctable, and it's not a
1789 * ZBT-style checksum, then it's suitable for metadata
1790 * as well. Otherwise, the metadata checksum defaults
1791 * to fletcher4.
1792 */
1793 if (zio_checksum_table[checksum].ci_correctable < 1 ||
1794 zio_checksum_table[checksum].ci_eck)
1795 checksum = ZIO_CHECKSUM_FLETCHER_4;
1796
1797 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL ||
1798 (os->os_redundant_metadata ==
1799 ZFS_REDUNDANT_METADATA_MOST &&
1800 (level >= zfs_redundant_metadata_most_ditto_level ||
1801 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))))
1802 copies++;
1803 } else if (wp & WP_NOFILL) {
1804 ASSERT(level == 0);
1805
1806 /*
1807 * If we're writing preallocated blocks, we aren't actually
1808 * writing them so don't set any policy properties. These
1809 * blocks are currently only used by an external subsystem
1810 * outside of zfs (i.e. dump) and not written by the zio
1811 * pipeline.
1812 */
1813 compress = ZIO_COMPRESS_OFF;
1814 checksum = ZIO_CHECKSUM_OFF;
1815 } else {
1816 compress = zio_compress_select(os->os_spa, dn->dn_compress,
1817 compress);
1818
1819 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
1820 zio_checksum_select(dn->dn_checksum, checksum) :
1821 dedup_checksum;
1822
1823 /*
1824 * Determine dedup setting. If we are in dmu_sync(),
1825 * we won't actually dedup now because that's all
1826 * done in syncing context; but we do want to use the
1827 * dedup checkum. If the checksum is not strong
1828 * enough to ensure unique signatures, force
1829 * dedup_verify.
1830 */
1831 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
1832 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
1833 if (!zio_checksum_table[checksum].ci_dedup)
1834 dedup_verify = B_TRUE;
1835 }
1836
1837 /*
1838 * Enable nopwrite if we have a cryptographically secure
1839 * checksum that has no known collisions (i.e. SHA-256)
1840 * and compression is enabled. We don't enable nopwrite if
1841 * dedup is enabled as the two features are mutually exclusive.
1842 */
1843 nopwrite = (!dedup && zio_checksum_table[checksum].ci_dedup &&
1844 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
1845 }
1846
1847 zp->zp_checksum = checksum;
1848 zp->zp_compress = compress;
1849 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
1850 zp->zp_level = level;
1851 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
1852 zp->zp_dedup = dedup;
1853 zp->zp_dedup_verify = dedup && dedup_verify;
1854 zp->zp_nopwrite = nopwrite;
1855 }
1856
1857 int
1858 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
1859 {
1860 dnode_t *dn;
1861 int i, err;
1862
1863 err = dnode_hold(os, object, FTAG, &dn);
1864 if (err)
1865 return (err);
1866 /*
1867 * Sync any current changes before
1868 * we go trundling through the block pointers.
1869 */
1870 for (i = 0; i < TXG_SIZE; i++) {
1871 if (list_link_active(&dn->dn_dirty_link[i]))
1872 break;
1873 }
1874 if (i != TXG_SIZE) {
1875 dnode_rele(dn, FTAG);
1876 txg_wait_synced(dmu_objset_pool(os), 0);
1877 err = dnode_hold(os, object, FTAG, &dn);
1878 if (err)
1879 return (err);
1880 }
1881
1882 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
1883 dnode_rele(dn, FTAG);
1884
1885 return (err);
1886 }
1887
1888 void
1889 __dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
1890 {
1891 dnode_phys_t *dnp = dn->dn_phys;
1892 int i;
1893
1894 doi->doi_data_block_size = dn->dn_datablksz;
1895 doi->doi_metadata_block_size = dn->dn_indblkshift ?
1896 1ULL << dn->dn_indblkshift : 0;
1897 doi->doi_type = dn->dn_type;
1898 doi->doi_bonus_type = dn->dn_bonustype;
1899 doi->doi_bonus_size = dn->dn_bonuslen;
1900 doi->doi_dnodesize = dn->dn_num_slots << DNODE_SHIFT;
1901 doi->doi_indirection = dn->dn_nlevels;
1902 doi->doi_checksum = dn->dn_checksum;
1903 doi->doi_compress = dn->dn_compress;
1904 doi->doi_nblkptr = dn->dn_nblkptr;
1905 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
1906 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
1907 doi->doi_fill_count = 0;
1908 for (i = 0; i < dnp->dn_nblkptr; i++)
1909 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
1910 }
1911
1912 void
1913 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
1914 {
1915 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1916 mutex_enter(&dn->dn_mtx);
1917
1918 __dmu_object_info_from_dnode(dn, doi);
1919
1920 mutex_exit(&dn->dn_mtx);
1921 rw_exit(&dn->dn_struct_rwlock);
1922 }
1923
1924 /*
1925 * Get information on a DMU object.
1926 * If doi is NULL, just indicates whether the object exists.
1927 */
1928 int
1929 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
1930 {
1931 dnode_t *dn;
1932 int err = dnode_hold(os, object, FTAG, &dn);
1933
1934 if (err)
1935 return (err);
1936
1937 if (doi != NULL)
1938 dmu_object_info_from_dnode(dn, doi);
1939
1940 dnode_rele(dn, FTAG);
1941 return (0);
1942 }
1943
1944 /*
1945 * As above, but faster; can be used when you have a held dbuf in hand.
1946 */
1947 void
1948 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
1949 {
1950 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1951
1952 DB_DNODE_ENTER(db);
1953 dmu_object_info_from_dnode(DB_DNODE(db), doi);
1954 DB_DNODE_EXIT(db);
1955 }
1956
1957 /*
1958 * Faster still when you only care about the size.
1959 * This is specifically optimized for zfs_getattr().
1960 */
1961 void
1962 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
1963 u_longlong_t *nblk512)
1964 {
1965 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1966 dnode_t *dn;
1967
1968 DB_DNODE_ENTER(db);
1969 dn = DB_DNODE(db);
1970
1971 *blksize = dn->dn_datablksz;
1972 /* add in number of slots used for the dnode itself */
1973 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
1974 SPA_MINBLOCKSHIFT) + dn->dn_num_slots;
1975 DB_DNODE_EXIT(db);
1976 }
1977
1978 void
1979 dmu_object_dnsize_from_db(dmu_buf_t *db_fake, int *dnsize)
1980 {
1981 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1982 dnode_t *dn;
1983
1984 DB_DNODE_ENTER(db);
1985 dn = DB_DNODE(db);
1986 *dnsize = dn->dn_num_slots << DNODE_SHIFT;
1987 DB_DNODE_EXIT(db);
1988 }
1989
1990 void
1991 byteswap_uint64_array(void *vbuf, size_t size)
1992 {
1993 uint64_t *buf = vbuf;
1994 size_t count = size >> 3;
1995 int i;
1996
1997 ASSERT((size & 7) == 0);
1998
1999 for (i = 0; i < count; i++)
2000 buf[i] = BSWAP_64(buf[i]);
2001 }
2002
2003 void
2004 byteswap_uint32_array(void *vbuf, size_t size)
2005 {
2006 uint32_t *buf = vbuf;
2007 size_t count = size >> 2;
2008 int i;
2009
2010 ASSERT((size & 3) == 0);
2011
2012 for (i = 0; i < count; i++)
2013 buf[i] = BSWAP_32(buf[i]);
2014 }
2015
2016 void
2017 byteswap_uint16_array(void *vbuf, size_t size)
2018 {
2019 uint16_t *buf = vbuf;
2020 size_t count = size >> 1;
2021 int i;
2022
2023 ASSERT((size & 1) == 0);
2024
2025 for (i = 0; i < count; i++)
2026 buf[i] = BSWAP_16(buf[i]);
2027 }
2028
2029 /* ARGSUSED */
2030 void
2031 byteswap_uint8_array(void *vbuf, size_t size)
2032 {
2033 }
2034
2035 void
2036 dmu_init(void)
2037 {
2038 zfs_dbgmsg_init();
2039 sa_cache_init();
2040 xuio_stat_init();
2041 dmu_objset_init();
2042 dnode_init();
2043 dbuf_init();
2044 zfetch_init();
2045 dmu_tx_init();
2046 l2arc_init();
2047 arc_init();
2048 }
2049
2050 void
2051 dmu_fini(void)
2052 {
2053 arc_fini(); /* arc depends on l2arc, so arc must go first */
2054 l2arc_fini();
2055 dmu_tx_fini();
2056 zfetch_fini();
2057 dbuf_fini();
2058 dnode_fini();
2059 dmu_objset_fini();
2060 xuio_stat_fini();
2061 sa_cache_fini();
2062 zfs_dbgmsg_fini();
2063 }
2064
2065 #if defined(_KERNEL) && defined(HAVE_SPL)
2066 EXPORT_SYMBOL(dmu_bonus_hold);
2067 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus);
2068 EXPORT_SYMBOL(dmu_buf_rele_array);
2069 EXPORT_SYMBOL(dmu_prefetch);
2070 EXPORT_SYMBOL(dmu_free_range);
2071 EXPORT_SYMBOL(dmu_free_long_range);
2072 EXPORT_SYMBOL(dmu_free_long_object);
2073 EXPORT_SYMBOL(dmu_read);
2074 EXPORT_SYMBOL(dmu_write);
2075 EXPORT_SYMBOL(dmu_prealloc);
2076 EXPORT_SYMBOL(dmu_object_info);
2077 EXPORT_SYMBOL(dmu_object_info_from_dnode);
2078 EXPORT_SYMBOL(dmu_object_info_from_db);
2079 EXPORT_SYMBOL(dmu_object_size_from_db);
2080 EXPORT_SYMBOL(dmu_object_dnsize_from_db);
2081 EXPORT_SYMBOL(dmu_object_set_blocksize);
2082 EXPORT_SYMBOL(dmu_object_set_checksum);
2083 EXPORT_SYMBOL(dmu_object_set_compress);
2084 EXPORT_SYMBOL(dmu_write_policy);
2085 EXPORT_SYMBOL(dmu_sync);
2086 EXPORT_SYMBOL(dmu_request_arcbuf);
2087 EXPORT_SYMBOL(dmu_return_arcbuf);
2088 EXPORT_SYMBOL(dmu_assign_arcbuf);
2089 EXPORT_SYMBOL(dmu_buf_hold);
2090 EXPORT_SYMBOL(dmu_ot);
2091
2092 module_param(zfs_mdcomp_disable, int, 0644);
2093 MODULE_PARM_DESC(zfs_mdcomp_disable, "Disable meta data compression");
2094
2095 module_param(zfs_nopwrite_enabled, int, 0644);
2096 MODULE_PARM_DESC(zfs_nopwrite_enabled, "Enable NOP writes");
2097
2098 #endif
ZFS On Linux mirror for PVE