]> 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 7004 - dmu_tx_hold_zap() does dnode_hold() 7x on same object
[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 && read &&
489 length <= zfetch_array_rd_sz) {
490 dmu_zfetch(&dn->dn_zfetch, blkid, nblks);
491 }
492 rw_exit(&dn->dn_struct_rwlock);
493
494 /* wait for async i/o */
495 err = zio_wait(zio);
496 if (err) {
497 dmu_buf_rele_array(dbp, nblks, tag);
498 return (err);
499 }
500
501 /* wait for other io to complete */
502 if (read) {
503 for (i = 0; i < nblks; i++) {
504 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
505 mutex_enter(&db->db_mtx);
506 while (db->db_state == DB_READ ||
507 db->db_state == DB_FILL)
508 cv_wait(&db->db_changed, &db->db_mtx);
509 if (db->db_state == DB_UNCACHED)
510 err = SET_ERROR(EIO);
511 mutex_exit(&db->db_mtx);
512 if (err) {
513 dmu_buf_rele_array(dbp, nblks, tag);
514 return (err);
515 }
516 }
517 }
518
519 *numbufsp = nblks;
520 *dbpp = dbp;
521 return (0);
522 }
523
524 static int
525 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
526 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
527 {
528 dnode_t *dn;
529 int err;
530
531 err = dnode_hold(os, object, FTAG, &dn);
532 if (err)
533 return (err);
534
535 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
536 numbufsp, dbpp, DMU_READ_PREFETCH);
537
538 dnode_rele(dn, FTAG);
539
540 return (err);
541 }
542
543 int
544 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
545 uint64_t length, boolean_t read, void *tag, int *numbufsp,
546 dmu_buf_t ***dbpp)
547 {
548 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
549 dnode_t *dn;
550 int err;
551
552 DB_DNODE_ENTER(db);
553 dn = DB_DNODE(db);
554 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
555 numbufsp, dbpp, DMU_READ_PREFETCH);
556 DB_DNODE_EXIT(db);
557
558 return (err);
559 }
560
561 void
562 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
563 {
564 int i;
565 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
566
567 if (numbufs == 0)
568 return;
569
570 for (i = 0; i < numbufs; i++) {
571 if (dbp[i])
572 dbuf_rele(dbp[i], tag);
573 }
574
575 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
576 }
577
578 /*
579 * Issue prefetch i/os for the given blocks. If level is greater than 0, the
580 * indirect blocks prefeteched will be those that point to the blocks containing
581 * the data starting at offset, and continuing to offset + len.
582 *
583 * Note that if the indirect blocks above the blocks being prefetched are not in
584 * cache, they will be asychronously read in.
585 */
586 void
587 dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset,
588 uint64_t len, zio_priority_t pri)
589 {
590 dnode_t *dn;
591 uint64_t blkid;
592 int nblks, err;
593
594 if (len == 0) { /* they're interested in the bonus buffer */
595 dn = DMU_META_DNODE(os);
596
597 if (object == 0 || object >= DN_MAX_OBJECT)
598 return;
599
600 rw_enter(&dn->dn_struct_rwlock, RW_READER);
601 blkid = dbuf_whichblock(dn, level,
602 object * sizeof (dnode_phys_t));
603 dbuf_prefetch(dn, level, blkid, pri, 0);
604 rw_exit(&dn->dn_struct_rwlock);
605 return;
606 }
607
608 /*
609 * XXX - Note, if the dnode for the requested object is not
610 * already cached, we will do a *synchronous* read in the
611 * dnode_hold() call. The same is true for any indirects.
612 */
613 err = dnode_hold(os, object, FTAG, &dn);
614 if (err != 0)
615 return;
616
617 rw_enter(&dn->dn_struct_rwlock, RW_READER);
618 /*
619 * offset + len - 1 is the last byte we want to prefetch for, and offset
620 * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the
621 * last block we want to prefetch, and dbuf_whichblock(dn, level,
622 * offset) is the first. Then the number we need to prefetch is the
623 * last - first + 1.
624 */
625 if (level > 0 || dn->dn_datablkshift != 0) {
626 nblks = dbuf_whichblock(dn, level, offset + len - 1) -
627 dbuf_whichblock(dn, level, offset) + 1;
628 } else {
629 nblks = (offset < dn->dn_datablksz);
630 }
631
632 if (nblks != 0) {
633 int i;
634
635 blkid = dbuf_whichblock(dn, level, offset);
636 for (i = 0; i < nblks; i++)
637 dbuf_prefetch(dn, level, blkid + i, pri, 0);
638 }
639
640 rw_exit(&dn->dn_struct_rwlock);
641
642 dnode_rele(dn, FTAG);
643 }
644
645 /*
646 * Get the next "chunk" of file data to free. We traverse the file from
647 * the end so that the file gets shorter over time (if we crashes in the
648 * middle, this will leave us in a better state). We find allocated file
649 * data by simply searching the allocated level 1 indirects.
650 *
651 * On input, *start should be the first offset that does not need to be
652 * freed (e.g. "offset + length"). On return, *start will be the first
653 * offset that should be freed.
654 */
655 static int
656 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum)
657 {
658 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
659 /* bytes of data covered by a level-1 indirect block */
660 uint64_t iblkrange =
661 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
662 uint64_t blks;
663
664 ASSERT3U(minimum, <=, *start);
665
666 if (*start - minimum <= iblkrange * maxblks) {
667 *start = minimum;
668 return (0);
669 }
670 ASSERT(ISP2(iblkrange));
671
672 for (blks = 0; *start > minimum && blks < maxblks; blks++) {
673 int err;
674
675 /*
676 * dnode_next_offset(BACKWARDS) will find an allocated L1
677 * indirect block at or before the input offset. We must
678 * decrement *start so that it is at the end of the region
679 * to search.
680 */
681 (*start)--;
682 err = dnode_next_offset(dn,
683 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
684
685 /* if there are no indirect blocks before start, we are done */
686 if (err == ESRCH) {
687 *start = minimum;
688 break;
689 } else if (err != 0) {
690 return (err);
691 }
692
693 /* set start to the beginning of this L1 indirect */
694 *start = P2ALIGN(*start, iblkrange);
695 }
696 if (*start < minimum)
697 *start = minimum;
698 return (0);
699 }
700
701 static int
702 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
703 uint64_t length)
704 {
705 uint64_t object_size;
706 int err;
707
708 if (dn == NULL)
709 return (SET_ERROR(EINVAL));
710
711 object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
712 if (offset >= object_size)
713 return (0);
714
715 if (length == DMU_OBJECT_END || offset + length > object_size)
716 length = object_size - offset;
717
718 while (length != 0) {
719 uint64_t chunk_end, chunk_begin;
720 dmu_tx_t *tx;
721
722 chunk_end = chunk_begin = offset + length;
723
724 /* move chunk_begin backwards to the beginning of this chunk */
725 err = get_next_chunk(dn, &chunk_begin, offset);
726 if (err)
727 return (err);
728 ASSERT3U(chunk_begin, >=, offset);
729 ASSERT3U(chunk_begin, <=, chunk_end);
730
731 tx = dmu_tx_create(os);
732 dmu_tx_hold_free(tx, dn->dn_object,
733 chunk_begin, chunk_end - chunk_begin);
734
735 /*
736 * Mark this transaction as typically resulting in a net
737 * reduction in space used.
738 */
739 dmu_tx_mark_netfree(tx);
740 err = dmu_tx_assign(tx, TXG_WAIT);
741 if (err) {
742 dmu_tx_abort(tx);
743 return (err);
744 }
745 dnode_free_range(dn, chunk_begin, chunk_end - chunk_begin, tx);
746 dmu_tx_commit(tx);
747
748 length -= chunk_end - chunk_begin;
749 }
750 return (0);
751 }
752
753 int
754 dmu_free_long_range(objset_t *os, uint64_t object,
755 uint64_t offset, uint64_t length)
756 {
757 dnode_t *dn;
758 int err;
759
760 err = dnode_hold(os, object, FTAG, &dn);
761 if (err != 0)
762 return (err);
763 err = dmu_free_long_range_impl(os, dn, offset, length);
764
765 /*
766 * It is important to zero out the maxblkid when freeing the entire
767 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
768 * will take the fast path, and (b) dnode_reallocate() can verify
769 * that the entire file has been freed.
770 */
771 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
772 dn->dn_maxblkid = 0;
773
774 dnode_rele(dn, FTAG);
775 return (err);
776 }
777
778 int
779 dmu_free_long_object(objset_t *os, uint64_t object)
780 {
781 dmu_tx_t *tx;
782 int err;
783
784 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
785 if (err != 0)
786 return (err);
787
788 tx = dmu_tx_create(os);
789 dmu_tx_hold_bonus(tx, object);
790 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
791 dmu_tx_mark_netfree(tx);
792 err = dmu_tx_assign(tx, TXG_WAIT);
793 if (err == 0) {
794 err = dmu_object_free(os, object, tx);
795 dmu_tx_commit(tx);
796 } else {
797 dmu_tx_abort(tx);
798 }
799
800 return (err);
801 }
802
803 int
804 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
805 uint64_t size, dmu_tx_t *tx)
806 {
807 dnode_t *dn;
808 int err = dnode_hold(os, object, FTAG, &dn);
809 if (err)
810 return (err);
811 ASSERT(offset < UINT64_MAX);
812 ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
813 dnode_free_range(dn, offset, size, tx);
814 dnode_rele(dn, FTAG);
815 return (0);
816 }
817
818 int
819 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
820 void *buf, uint32_t flags)
821 {
822 dnode_t *dn;
823 dmu_buf_t **dbp;
824 int numbufs, err;
825
826 err = dnode_hold(os, object, FTAG, &dn);
827 if (err)
828 return (err);
829
830 /*
831 * Deal with odd block sizes, where there can't be data past the first
832 * block. If we ever do the tail block optimization, we will need to
833 * handle that here as well.
834 */
835 if (dn->dn_maxblkid == 0) {
836 uint64_t newsz = offset > dn->dn_datablksz ? 0 :
837 MIN(size, dn->dn_datablksz - offset);
838 bzero((char *)buf + newsz, size - newsz);
839 size = newsz;
840 }
841
842 while (size > 0) {
843 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
844 int i;
845
846 /*
847 * NB: we could do this block-at-a-time, but it's nice
848 * to be reading in parallel.
849 */
850 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
851 TRUE, FTAG, &numbufs, &dbp, flags);
852 if (err)
853 break;
854
855 for (i = 0; i < numbufs; i++) {
856 uint64_t tocpy;
857 int64_t bufoff;
858 dmu_buf_t *db = dbp[i];
859
860 ASSERT(size > 0);
861
862 bufoff = offset - db->db_offset;
863 tocpy = MIN(db->db_size - bufoff, size);
864
865 (void) memcpy(buf, (char *)db->db_data + bufoff, tocpy);
866
867 offset += tocpy;
868 size -= tocpy;
869 buf = (char *)buf + tocpy;
870 }
871 dmu_buf_rele_array(dbp, numbufs, FTAG);
872 }
873 dnode_rele(dn, FTAG);
874 return (err);
875 }
876
877 void
878 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
879 const void *buf, dmu_tx_t *tx)
880 {
881 dmu_buf_t **dbp;
882 int numbufs, i;
883
884 if (size == 0)
885 return;
886
887 VERIFY0(dmu_buf_hold_array(os, object, offset, size,
888 FALSE, FTAG, &numbufs, &dbp));
889
890 for (i = 0; i < numbufs; i++) {
891 uint64_t tocpy;
892 int64_t bufoff;
893 dmu_buf_t *db = dbp[i];
894
895 ASSERT(size > 0);
896
897 bufoff = offset - db->db_offset;
898 tocpy = MIN(db->db_size - bufoff, size);
899
900 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
901
902 if (tocpy == db->db_size)
903 dmu_buf_will_fill(db, tx);
904 else
905 dmu_buf_will_dirty(db, tx);
906
907 (void) memcpy((char *)db->db_data + bufoff, buf, tocpy);
908
909 if (tocpy == db->db_size)
910 dmu_buf_fill_done(db, tx);
911
912 offset += tocpy;
913 size -= tocpy;
914 buf = (char *)buf + tocpy;
915 }
916 dmu_buf_rele_array(dbp, numbufs, FTAG);
917 }
918
919 void
920 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
921 dmu_tx_t *tx)
922 {
923 dmu_buf_t **dbp;
924 int numbufs, i;
925
926 if (size == 0)
927 return;
928
929 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
930 FALSE, FTAG, &numbufs, &dbp));
931
932 for (i = 0; i < numbufs; i++) {
933 dmu_buf_t *db = dbp[i];
934
935 dmu_buf_will_not_fill(db, tx);
936 }
937 dmu_buf_rele_array(dbp, numbufs, FTAG);
938 }
939
940 void
941 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
942 void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
943 int compressed_size, int byteorder, dmu_tx_t *tx)
944 {
945 dmu_buf_t *db;
946
947 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
948 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
949 VERIFY0(dmu_buf_hold_noread(os, object, offset,
950 FTAG, &db));
951
952 dmu_buf_write_embedded(db,
953 data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
954 uncompressed_size, compressed_size, byteorder, tx);
955
956 dmu_buf_rele(db, FTAG);
957 }
958
959 /*
960 * DMU support for xuio
961 */
962 kstat_t *xuio_ksp = NULL;
963
964 typedef struct xuio_stats {
965 /* loaned yet not returned arc_buf */
966 kstat_named_t xuiostat_onloan_rbuf;
967 kstat_named_t xuiostat_onloan_wbuf;
968 /* whether a copy is made when loaning out a read buffer */
969 kstat_named_t xuiostat_rbuf_copied;
970 kstat_named_t xuiostat_rbuf_nocopy;
971 /* whether a copy is made when assigning a write buffer */
972 kstat_named_t xuiostat_wbuf_copied;
973 kstat_named_t xuiostat_wbuf_nocopy;
974 } xuio_stats_t;
975
976 static xuio_stats_t xuio_stats = {
977 { "onloan_read_buf", KSTAT_DATA_UINT64 },
978 { "onloan_write_buf", KSTAT_DATA_UINT64 },
979 { "read_buf_copied", KSTAT_DATA_UINT64 },
980 { "read_buf_nocopy", KSTAT_DATA_UINT64 },
981 { "write_buf_copied", KSTAT_DATA_UINT64 },
982 { "write_buf_nocopy", KSTAT_DATA_UINT64 }
983 };
984
985 #define XUIOSTAT_INCR(stat, val) \
986 atomic_add_64(&xuio_stats.stat.value.ui64, (val))
987 #define XUIOSTAT_BUMP(stat) XUIOSTAT_INCR(stat, 1)
988
989 int
990 dmu_xuio_init(xuio_t *xuio, int nblk)
991 {
992 dmu_xuio_t *priv;
993 uio_t *uio = &xuio->xu_uio;
994
995 uio->uio_iovcnt = nblk;
996 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
997
998 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
999 priv->cnt = nblk;
1000 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
1001 priv->iovp = (iovec_t *)uio->uio_iov;
1002 XUIO_XUZC_PRIV(xuio) = priv;
1003
1004 if (XUIO_XUZC_RW(xuio) == UIO_READ)
1005 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
1006 else
1007 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
1008
1009 return (0);
1010 }
1011
1012 void
1013 dmu_xuio_fini(xuio_t *xuio)
1014 {
1015 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1016 int nblk = priv->cnt;
1017
1018 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
1019 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
1020 kmem_free(priv, sizeof (dmu_xuio_t));
1021
1022 if (XUIO_XUZC_RW(xuio) == UIO_READ)
1023 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
1024 else
1025 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
1026 }
1027
1028 /*
1029 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
1030 * and increase priv->next by 1.
1031 */
1032 int
1033 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
1034 {
1035 struct iovec *iov;
1036 uio_t *uio = &xuio->xu_uio;
1037 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1038 int i = priv->next++;
1039
1040 ASSERT(i < priv->cnt);
1041 ASSERT(off + n <= arc_buf_size(abuf));
1042 iov = (iovec_t *)uio->uio_iov + i;
1043 iov->iov_base = (char *)abuf->b_data + off;
1044 iov->iov_len = n;
1045 priv->bufs[i] = abuf;
1046 return (0);
1047 }
1048
1049 int
1050 dmu_xuio_cnt(xuio_t *xuio)
1051 {
1052 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1053 return (priv->cnt);
1054 }
1055
1056 arc_buf_t *
1057 dmu_xuio_arcbuf(xuio_t *xuio, int i)
1058 {
1059 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1060
1061 ASSERT(i < priv->cnt);
1062 return (priv->bufs[i]);
1063 }
1064
1065 void
1066 dmu_xuio_clear(xuio_t *xuio, int i)
1067 {
1068 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1069
1070 ASSERT(i < priv->cnt);
1071 priv->bufs[i] = NULL;
1072 }
1073
1074 static void
1075 xuio_stat_init(void)
1076 {
1077 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
1078 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
1079 KSTAT_FLAG_VIRTUAL);
1080 if (xuio_ksp != NULL) {
1081 xuio_ksp->ks_data = &xuio_stats;
1082 kstat_install(xuio_ksp);
1083 }
1084 }
1085
1086 static void
1087 xuio_stat_fini(void)
1088 {
1089 if (xuio_ksp != NULL) {
1090 kstat_delete(xuio_ksp);
1091 xuio_ksp = NULL;
1092 }
1093 }
1094
1095 void
1096 xuio_stat_wbuf_copied()
1097 {
1098 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1099 }
1100
1101 void
1102 xuio_stat_wbuf_nocopy()
1103 {
1104 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
1105 }
1106
1107 #ifdef _KERNEL
1108 static int
1109 dmu_read_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size)
1110 {
1111 dmu_buf_t **dbp;
1112 int numbufs, i, err;
1113 xuio_t *xuio = NULL;
1114
1115 /*
1116 * NB: we could do this block-at-a-time, but it's nice
1117 * to be reading in parallel.
1118 */
1119 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1120 TRUE, FTAG, &numbufs, &dbp, 0);
1121 if (err)
1122 return (err);
1123
1124 for (i = 0; i < numbufs; i++) {
1125 uint64_t tocpy;
1126 int64_t bufoff;
1127 dmu_buf_t *db = dbp[i];
1128
1129 ASSERT(size > 0);
1130
1131 bufoff = uio->uio_loffset - db->db_offset;
1132 tocpy = MIN(db->db_size - bufoff, size);
1133
1134 if (xuio) {
1135 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
1136 arc_buf_t *dbuf_abuf = dbi->db_buf;
1137 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
1138 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
1139 if (!err) {
1140 uio->uio_resid -= tocpy;
1141 uio->uio_loffset += tocpy;
1142 }
1143
1144 if (abuf == dbuf_abuf)
1145 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
1146 else
1147 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
1148 } else {
1149 err = uiomove((char *)db->db_data + bufoff, tocpy,
1150 UIO_READ, uio);
1151 }
1152 if (err)
1153 break;
1154
1155 size -= tocpy;
1156 }
1157 dmu_buf_rele_array(dbp, numbufs, FTAG);
1158
1159 return (err);
1160 }
1161
1162 /*
1163 * Read 'size' bytes into the uio buffer.
1164 * From object zdb->db_object.
1165 * Starting at offset uio->uio_loffset.
1166 *
1167 * If the caller already has a dbuf in the target object
1168 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1169 * because we don't have to find the dnode_t for the object.
1170 */
1171 int
1172 dmu_read_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size)
1173 {
1174 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1175 dnode_t *dn;
1176 int err;
1177
1178 if (size == 0)
1179 return (0);
1180
1181 DB_DNODE_ENTER(db);
1182 dn = DB_DNODE(db);
1183 err = dmu_read_uio_dnode(dn, uio, size);
1184 DB_DNODE_EXIT(db);
1185
1186 return (err);
1187 }
1188
1189 /*
1190 * Read 'size' bytes into the uio buffer.
1191 * From the specified object
1192 * Starting at offset uio->uio_loffset.
1193 */
1194 int
1195 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
1196 {
1197 dnode_t *dn;
1198 int err;
1199
1200 if (size == 0)
1201 return (0);
1202
1203 err = dnode_hold(os, object, FTAG, &dn);
1204 if (err)
1205 return (err);
1206
1207 err = dmu_read_uio_dnode(dn, uio, size);
1208
1209 dnode_rele(dn, FTAG);
1210
1211 return (err);
1212 }
1213
1214 static int
1215 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1216 {
1217 dmu_buf_t **dbp;
1218 int numbufs;
1219 int err = 0;
1220 int i;
1221
1222 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1223 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1224 if (err)
1225 return (err);
1226
1227 for (i = 0; i < numbufs; i++) {
1228 uint64_t tocpy;
1229 int64_t bufoff;
1230 dmu_buf_t *db = dbp[i];
1231
1232 ASSERT(size > 0);
1233
1234 bufoff = uio->uio_loffset - db->db_offset;
1235 tocpy = MIN(db->db_size - bufoff, size);
1236
1237 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1238
1239 if (tocpy == db->db_size)
1240 dmu_buf_will_fill(db, tx);
1241 else
1242 dmu_buf_will_dirty(db, tx);
1243
1244 /*
1245 * XXX uiomove could block forever (eg.nfs-backed
1246 * pages). There needs to be a uiolockdown() function
1247 * to lock the pages in memory, so that uiomove won't
1248 * block.
1249 */
1250 err = uiomove((char *)db->db_data + bufoff, tocpy,
1251 UIO_WRITE, uio);
1252
1253 if (tocpy == db->db_size)
1254 dmu_buf_fill_done(db, tx);
1255
1256 if (err)
1257 break;
1258
1259 size -= tocpy;
1260 }
1261
1262 dmu_buf_rele_array(dbp, numbufs, FTAG);
1263 return (err);
1264 }
1265
1266 /*
1267 * Write 'size' bytes from the uio buffer.
1268 * To object zdb->db_object.
1269 * Starting at offset uio->uio_loffset.
1270 *
1271 * If the caller already has a dbuf in the target object
1272 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1273 * because we don't have to find the dnode_t for the object.
1274 */
1275 int
1276 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1277 dmu_tx_t *tx)
1278 {
1279 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1280 dnode_t *dn;
1281 int err;
1282
1283 if (size == 0)
1284 return (0);
1285
1286 DB_DNODE_ENTER(db);
1287 dn = DB_DNODE(db);
1288 err = dmu_write_uio_dnode(dn, uio, size, tx);
1289 DB_DNODE_EXIT(db);
1290
1291 return (err);
1292 }
1293
1294 /*
1295 * Write 'size' bytes from the uio buffer.
1296 * To the specified object.
1297 * Starting at offset uio->uio_loffset.
1298 */
1299 int
1300 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1301 dmu_tx_t *tx)
1302 {
1303 dnode_t *dn;
1304 int err;
1305
1306 if (size == 0)
1307 return (0);
1308
1309 err = dnode_hold(os, object, FTAG, &dn);
1310 if (err)
1311 return (err);
1312
1313 err = dmu_write_uio_dnode(dn, uio, size, tx);
1314
1315 dnode_rele(dn, FTAG);
1316
1317 return (err);
1318 }
1319 #endif /* _KERNEL */
1320
1321 /*
1322 * Allocate a loaned anonymous arc buffer.
1323 */
1324 arc_buf_t *
1325 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1326 {
1327 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1328
1329 return (arc_loan_buf(db->db_objset->os_spa, size));
1330 }
1331
1332 /*
1333 * Free a loaned arc buffer.
1334 */
1335 void
1336 dmu_return_arcbuf(arc_buf_t *buf)
1337 {
1338 arc_return_buf(buf, FTAG);
1339 VERIFY(arc_buf_remove_ref(buf, FTAG));
1340 }
1341
1342 /*
1343 * When possible directly assign passed loaned arc buffer to a dbuf.
1344 * If this is not possible copy the contents of passed arc buf via
1345 * dmu_write().
1346 */
1347 void
1348 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1349 dmu_tx_t *tx)
1350 {
1351 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1352 dnode_t *dn;
1353 dmu_buf_impl_t *db;
1354 uint32_t blksz = (uint32_t)arc_buf_size(buf);
1355 uint64_t blkid;
1356
1357 DB_DNODE_ENTER(dbuf);
1358 dn = DB_DNODE(dbuf);
1359 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1360 blkid = dbuf_whichblock(dn, 0, offset);
1361 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1362 rw_exit(&dn->dn_struct_rwlock);
1363 DB_DNODE_EXIT(dbuf);
1364
1365 /*
1366 * We can only assign if the offset is aligned, the arc buf is the
1367 * same size as the dbuf, and the dbuf is not metadata. It
1368 * can't be metadata because the loaned arc buf comes from the
1369 * user-data kmem area.
1370 */
1371 if (offset == db->db.db_offset && blksz == db->db.db_size &&
1372 DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA) {
1373 dbuf_assign_arcbuf(db, buf, tx);
1374 dbuf_rele(db, FTAG);
1375 } else {
1376 objset_t *os;
1377 uint64_t object;
1378
1379 DB_DNODE_ENTER(dbuf);
1380 dn = DB_DNODE(dbuf);
1381 os = dn->dn_objset;
1382 object = dn->dn_object;
1383 DB_DNODE_EXIT(dbuf);
1384
1385 dbuf_rele(db, FTAG);
1386 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1387 dmu_return_arcbuf(buf);
1388 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1389 }
1390 }
1391
1392 typedef struct {
1393 dbuf_dirty_record_t *dsa_dr;
1394 dmu_sync_cb_t *dsa_done;
1395 zgd_t *dsa_zgd;
1396 dmu_tx_t *dsa_tx;
1397 } dmu_sync_arg_t;
1398
1399 /* ARGSUSED */
1400 static void
1401 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1402 {
1403 dmu_sync_arg_t *dsa = varg;
1404 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1405 blkptr_t *bp = zio->io_bp;
1406
1407 if (zio->io_error == 0) {
1408 if (BP_IS_HOLE(bp)) {
1409 /*
1410 * A block of zeros may compress to a hole, but the
1411 * block size still needs to be known for replay.
1412 */
1413 BP_SET_LSIZE(bp, db->db_size);
1414 } else if (!BP_IS_EMBEDDED(bp)) {
1415 ASSERT(BP_GET_LEVEL(bp) == 0);
1416 bp->blk_fill = 1;
1417 }
1418 }
1419 }
1420
1421 static void
1422 dmu_sync_late_arrival_ready(zio_t *zio)
1423 {
1424 dmu_sync_ready(zio, NULL, zio->io_private);
1425 }
1426
1427 /* ARGSUSED */
1428 static void
1429 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1430 {
1431 dmu_sync_arg_t *dsa = varg;
1432 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1433 dmu_buf_impl_t *db = dr->dr_dbuf;
1434
1435 mutex_enter(&db->db_mtx);
1436 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1437 if (zio->io_error == 0) {
1438 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1439 if (dr->dt.dl.dr_nopwrite) {
1440 ASSERTV(blkptr_t *bp = zio->io_bp);
1441 ASSERTV(blkptr_t *bp_orig = &zio->io_bp_orig);
1442 ASSERTV(uint8_t chksum = BP_GET_CHECKSUM(bp_orig));
1443
1444 ASSERT(BP_EQUAL(bp, bp_orig));
1445 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1446 ASSERT(zio_checksum_table[chksum].ci_dedup);
1447 }
1448 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1449 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1450 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1451
1452 /*
1453 * Old style holes are filled with all zeros, whereas
1454 * new-style holes maintain their lsize, type, level,
1455 * and birth time (see zio_write_compress). While we
1456 * need to reset the BP_SET_LSIZE() call that happened
1457 * in dmu_sync_ready for old style holes, we do *not*
1458 * want to wipe out the information contained in new
1459 * style holes. Thus, only zero out the block pointer if
1460 * it's an old style hole.
1461 */
1462 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) &&
1463 dr->dt.dl.dr_overridden_by.blk_birth == 0)
1464 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1465 } else {
1466 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1467 }
1468 cv_broadcast(&db->db_changed);
1469 mutex_exit(&db->db_mtx);
1470
1471 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1472
1473 kmem_free(dsa, sizeof (*dsa));
1474 }
1475
1476 static void
1477 dmu_sync_late_arrival_done(zio_t *zio)
1478 {
1479 blkptr_t *bp = zio->io_bp;
1480 dmu_sync_arg_t *dsa = zio->io_private;
1481 ASSERTV(blkptr_t *bp_orig = &zio->io_bp_orig);
1482
1483 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1484 /*
1485 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE)
1486 * then there is nothing to do here. Otherwise, free the
1487 * newly allocated block in this txg.
1488 */
1489 if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
1490 ASSERT(BP_EQUAL(bp, bp_orig));
1491 } else {
1492 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1493 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1494 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1495 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1496 }
1497 }
1498
1499 dmu_tx_commit(dsa->dsa_tx);
1500
1501 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1502
1503 kmem_free(dsa, sizeof (*dsa));
1504 }
1505
1506 static int
1507 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1508 zio_prop_t *zp, zbookmark_phys_t *zb)
1509 {
1510 dmu_sync_arg_t *dsa;
1511 dmu_tx_t *tx;
1512
1513 tx = dmu_tx_create(os);
1514 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1515 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1516 dmu_tx_abort(tx);
1517 /* Make zl_get_data do txg_waited_synced() */
1518 return (SET_ERROR(EIO));
1519 }
1520
1521 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1522 dsa->dsa_dr = NULL;
1523 dsa->dsa_done = done;
1524 dsa->dsa_zgd = zgd;
1525 dsa->dsa_tx = tx;
1526
1527 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx),
1528 zgd->zgd_bp, zgd->zgd_db->db_data, zgd->zgd_db->db_size,
1529 zp, dmu_sync_late_arrival_ready, NULL,
1530 NULL, dmu_sync_late_arrival_done, dsa, ZIO_PRIORITY_SYNC_WRITE,
1531 ZIO_FLAG_CANFAIL, zb));
1532
1533 return (0);
1534 }
1535
1536 /*
1537 * Intent log support: sync the block associated with db to disk.
1538 * N.B. and XXX: the caller is responsible for making sure that the
1539 * data isn't changing while dmu_sync() is writing it.
1540 *
1541 * Return values:
1542 *
1543 * EEXIST: this txg has already been synced, so there's nothing to do.
1544 * The caller should not log the write.
1545 *
1546 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1547 * The caller should not log the write.
1548 *
1549 * EALREADY: this block is already in the process of being synced.
1550 * The caller should track its progress (somehow).
1551 *
1552 * EIO: could not do the I/O.
1553 * The caller should do a txg_wait_synced().
1554 *
1555 * 0: the I/O has been initiated.
1556 * The caller should log this blkptr in the done callback.
1557 * It is possible that the I/O will fail, in which case
1558 * the error will be reported to the done callback and
1559 * propagated to pio from zio_done().
1560 */
1561 int
1562 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1563 {
1564 blkptr_t *bp = zgd->zgd_bp;
1565 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1566 objset_t *os = db->db_objset;
1567 dsl_dataset_t *ds = os->os_dsl_dataset;
1568 dbuf_dirty_record_t *dr;
1569 dmu_sync_arg_t *dsa;
1570 zbookmark_phys_t zb;
1571 zio_prop_t zp;
1572 dnode_t *dn;
1573
1574 ASSERT(pio != NULL);
1575 ASSERT(txg != 0);
1576
1577 SET_BOOKMARK(&zb, ds->ds_object,
1578 db->db.db_object, db->db_level, db->db_blkid);
1579
1580 DB_DNODE_ENTER(db);
1581 dn = DB_DNODE(db);
1582 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1583 DB_DNODE_EXIT(db);
1584
1585 /*
1586 * If we're frozen (running ziltest), we always need to generate a bp.
1587 */
1588 if (txg > spa_freeze_txg(os->os_spa))
1589 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1590
1591 /*
1592 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1593 * and us. If we determine that this txg is not yet syncing,
1594 * but it begins to sync a moment later, that's OK because the
1595 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1596 */
1597 mutex_enter(&db->db_mtx);
1598
1599 if (txg <= spa_last_synced_txg(os->os_spa)) {
1600 /*
1601 * This txg has already synced. There's nothing to do.
1602 */
1603 mutex_exit(&db->db_mtx);
1604 return (SET_ERROR(EEXIST));
1605 }
1606
1607 if (txg <= spa_syncing_txg(os->os_spa)) {
1608 /*
1609 * This txg is currently syncing, so we can't mess with
1610 * the dirty record anymore; just write a new log block.
1611 */
1612 mutex_exit(&db->db_mtx);
1613 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1614 }
1615
1616 dr = db->db_last_dirty;
1617 while (dr && dr->dr_txg != txg)
1618 dr = dr->dr_next;
1619
1620 if (dr == NULL) {
1621 /*
1622 * There's no dr for this dbuf, so it must have been freed.
1623 * There's no need to log writes to freed blocks, so we're done.
1624 */
1625 mutex_exit(&db->db_mtx);
1626 return (SET_ERROR(ENOENT));
1627 }
1628
1629 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
1630
1631 /*
1632 * Assume the on-disk data is X, the current syncing data (in
1633 * txg - 1) is Y, and the current in-memory data is Z (currently
1634 * in dmu_sync).
1635 *
1636 * We usually want to perform a nopwrite if X and Z are the
1637 * same. However, if Y is different (i.e. the BP is going to
1638 * change before this write takes effect), then a nopwrite will
1639 * be incorrect - we would override with X, which could have
1640 * been freed when Y was written.
1641 *
1642 * (Note that this is not a concern when we are nop-writing from
1643 * syncing context, because X and Y must be identical, because
1644 * all previous txgs have been synced.)
1645 *
1646 * Therefore, we disable nopwrite if the current BP could change
1647 * before this TXG. There are two ways it could change: by
1648 * being dirty (dr_next is non-NULL), or by being freed
1649 * (dnode_block_freed()). This behavior is verified by
1650 * zio_done(), which VERIFYs that the override BP is identical
1651 * to the on-disk BP.
1652 */
1653 DB_DNODE_ENTER(db);
1654 dn = DB_DNODE(db);
1655 if (dr->dr_next != NULL || dnode_block_freed(dn, db->db_blkid))
1656 zp.zp_nopwrite = B_FALSE;
1657 DB_DNODE_EXIT(db);
1658
1659 ASSERT(dr->dr_txg == txg);
1660 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1661 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1662 /*
1663 * We have already issued a sync write for this buffer,
1664 * or this buffer has already been synced. It could not
1665 * have been dirtied since, or we would have cleared the state.
1666 */
1667 mutex_exit(&db->db_mtx);
1668 return (SET_ERROR(EALREADY));
1669 }
1670
1671 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1672 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1673 mutex_exit(&db->db_mtx);
1674
1675 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1676 dsa->dsa_dr = dr;
1677 dsa->dsa_done = done;
1678 dsa->dsa_zgd = zgd;
1679 dsa->dsa_tx = NULL;
1680
1681 zio_nowait(arc_write(pio, os->os_spa, txg,
1682 bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
1683 DBUF_IS_L2COMPRESSIBLE(db), &zp, dmu_sync_ready,
1684 NULL, NULL, dmu_sync_done, dsa,
1685 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb));
1686
1687 return (0);
1688 }
1689
1690 int
1691 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1692 dmu_tx_t *tx)
1693 {
1694 dnode_t *dn;
1695 int err;
1696
1697 err = dnode_hold(os, object, FTAG, &dn);
1698 if (err)
1699 return (err);
1700 err = dnode_set_blksz(dn, size, ibs, tx);
1701 dnode_rele(dn, FTAG);
1702 return (err);
1703 }
1704
1705 void
1706 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1707 dmu_tx_t *tx)
1708 {
1709 dnode_t *dn;
1710
1711 /*
1712 * Send streams include each object's checksum function. This
1713 * check ensures that the receiving system can understand the
1714 * checksum function transmitted.
1715 */
1716 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS);
1717
1718 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1719 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS);
1720 dn->dn_checksum = checksum;
1721 dnode_setdirty(dn, tx);
1722 dnode_rele(dn, FTAG);
1723 }
1724
1725 void
1726 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1727 dmu_tx_t *tx)
1728 {
1729 dnode_t *dn;
1730
1731 /*
1732 * Send streams include each object's compression function. This
1733 * check ensures that the receiving system can understand the
1734 * compression function transmitted.
1735 */
1736 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS);
1737
1738 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1739 dn->dn_compress = compress;
1740 dnode_setdirty(dn, tx);
1741 dnode_rele(dn, FTAG);
1742 }
1743
1744 int zfs_mdcomp_disable = 0;
1745
1746 /*
1747 * When the "redundant_metadata" property is set to "most", only indirect
1748 * blocks of this level and higher will have an additional ditto block.
1749 */
1750 int zfs_redundant_metadata_most_ditto_level = 2;
1751
1752 void
1753 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
1754 {
1755 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
1756 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
1757 (wp & WP_SPILL));
1758 enum zio_checksum checksum = os->os_checksum;
1759 enum zio_compress compress = os->os_compress;
1760 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
1761 boolean_t dedup = B_FALSE;
1762 boolean_t nopwrite = B_FALSE;
1763 boolean_t dedup_verify = os->os_dedup_verify;
1764 int copies = os->os_copies;
1765
1766 /*
1767 * We maintain different write policies for each of the following
1768 * types of data:
1769 * 1. metadata
1770 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1771 * 3. all other level 0 blocks
1772 */
1773 if (ismd) {
1774 if (zfs_mdcomp_disable) {
1775 compress = ZIO_COMPRESS_EMPTY;
1776 } else {
1777 /*
1778 * XXX -- we should design a compression algorithm
1779 * that specializes in arrays of bps.
1780 */
1781 compress = zio_compress_select(os->os_spa,
1782 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON);
1783 }
1784
1785 /*
1786 * Metadata always gets checksummed. If the data
1787 * checksum is multi-bit correctable, and it's not a
1788 * ZBT-style checksum, then it's suitable for metadata
1789 * as well. Otherwise, the metadata checksum defaults
1790 * to fletcher4.
1791 */
1792 if (zio_checksum_table[checksum].ci_correctable < 1 ||
1793 zio_checksum_table[checksum].ci_eck)
1794 checksum = ZIO_CHECKSUM_FLETCHER_4;
1795
1796 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL ||
1797 (os->os_redundant_metadata ==
1798 ZFS_REDUNDANT_METADATA_MOST &&
1799 (level >= zfs_redundant_metadata_most_ditto_level ||
1800 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))))
1801 copies++;
1802 } else if (wp & WP_NOFILL) {
1803 ASSERT(level == 0);
1804
1805 /*
1806 * If we're writing preallocated blocks, we aren't actually
1807 * writing them so don't set any policy properties. These
1808 * blocks are currently only used by an external subsystem
1809 * outside of zfs (i.e. dump) and not written by the zio
1810 * pipeline.
1811 */
1812 compress = ZIO_COMPRESS_OFF;
1813 checksum = ZIO_CHECKSUM_OFF;
1814 } else {
1815 compress = zio_compress_select(os->os_spa, dn->dn_compress,
1816 compress);
1817
1818 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
1819 zio_checksum_select(dn->dn_checksum, checksum) :
1820 dedup_checksum;
1821
1822 /*
1823 * Determine dedup setting. If we are in dmu_sync(),
1824 * we won't actually dedup now because that's all
1825 * done in syncing context; but we do want to use the
1826 * dedup checkum. If the checksum is not strong
1827 * enough to ensure unique signatures, force
1828 * dedup_verify.
1829 */
1830 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
1831 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
1832 if (!zio_checksum_table[checksum].ci_dedup)
1833 dedup_verify = B_TRUE;
1834 }
1835
1836 /*
1837 * Enable nopwrite if we have a cryptographically secure
1838 * checksum that has no known collisions (i.e. SHA-256)
1839 * and compression is enabled. We don't enable nopwrite if
1840 * dedup is enabled as the two features are mutually exclusive.
1841 */
1842 nopwrite = (!dedup && zio_checksum_table[checksum].ci_dedup &&
1843 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
1844 }
1845
1846 zp->zp_checksum = checksum;
1847 zp->zp_compress = compress;
1848 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
1849 zp->zp_level = level;
1850 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
1851 zp->zp_dedup = dedup;
1852 zp->zp_dedup_verify = dedup && dedup_verify;
1853 zp->zp_nopwrite = nopwrite;
1854 }
1855
1856 int
1857 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
1858 {
1859 dnode_t *dn;
1860 int i, err;
1861
1862 err = dnode_hold(os, object, FTAG, &dn);
1863 if (err)
1864 return (err);
1865 /*
1866 * Sync any current changes before
1867 * we go trundling through the block pointers.
1868 */
1869 for (i = 0; i < TXG_SIZE; i++) {
1870 if (list_link_active(&dn->dn_dirty_link[i]))
1871 break;
1872 }
1873 if (i != TXG_SIZE) {
1874 dnode_rele(dn, FTAG);
1875 txg_wait_synced(dmu_objset_pool(os), 0);
1876 err = dnode_hold(os, object, FTAG, &dn);
1877 if (err)
1878 return (err);
1879 }
1880
1881 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
1882 dnode_rele(dn, FTAG);
1883
1884 return (err);
1885 }
1886
1887 void
1888 __dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
1889 {
1890 dnode_phys_t *dnp = dn->dn_phys;
1891 int i;
1892
1893 doi->doi_data_block_size = dn->dn_datablksz;
1894 doi->doi_metadata_block_size = dn->dn_indblkshift ?
1895 1ULL << dn->dn_indblkshift : 0;
1896 doi->doi_type = dn->dn_type;
1897 doi->doi_bonus_type = dn->dn_bonustype;
1898 doi->doi_bonus_size = dn->dn_bonuslen;
1899 doi->doi_dnodesize = dn->dn_num_slots << DNODE_SHIFT;
1900 doi->doi_indirection = dn->dn_nlevels;
1901 doi->doi_checksum = dn->dn_checksum;
1902 doi->doi_compress = dn->dn_compress;
1903 doi->doi_nblkptr = dn->dn_nblkptr;
1904 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
1905 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
1906 doi->doi_fill_count = 0;
1907 for (i = 0; i < dnp->dn_nblkptr; i++)
1908 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
1909 }
1910
1911 void
1912 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
1913 {
1914 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1915 mutex_enter(&dn->dn_mtx);
1916
1917 __dmu_object_info_from_dnode(dn, doi);
1918
1919 mutex_exit(&dn->dn_mtx);
1920 rw_exit(&dn->dn_struct_rwlock);
1921 }
1922
1923 /*
1924 * Get information on a DMU object.
1925 * If doi is NULL, just indicates whether the object exists.
1926 */
1927 int
1928 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
1929 {
1930 dnode_t *dn;
1931 int err = dnode_hold(os, object, FTAG, &dn);
1932
1933 if (err)
1934 return (err);
1935
1936 if (doi != NULL)
1937 dmu_object_info_from_dnode(dn, doi);
1938
1939 dnode_rele(dn, FTAG);
1940 return (0);
1941 }
1942
1943 /*
1944 * As above, but faster; can be used when you have a held dbuf in hand.
1945 */
1946 void
1947 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
1948 {
1949 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1950
1951 DB_DNODE_ENTER(db);
1952 dmu_object_info_from_dnode(DB_DNODE(db), doi);
1953 DB_DNODE_EXIT(db);
1954 }
1955
1956 /*
1957 * Faster still when you only care about the size.
1958 * This is specifically optimized for zfs_getattr().
1959 */
1960 void
1961 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
1962 u_longlong_t *nblk512)
1963 {
1964 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1965 dnode_t *dn;
1966
1967 DB_DNODE_ENTER(db);
1968 dn = DB_DNODE(db);
1969
1970 *blksize = dn->dn_datablksz;
1971 /* add in number of slots used for the dnode itself */
1972 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
1973 SPA_MINBLOCKSHIFT) + dn->dn_num_slots;
1974 DB_DNODE_EXIT(db);
1975 }
1976
1977 void
1978 dmu_object_dnsize_from_db(dmu_buf_t *db_fake, int *dnsize)
1979 {
1980 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1981 dnode_t *dn;
1982
1983 DB_DNODE_ENTER(db);
1984 dn = DB_DNODE(db);
1985 *dnsize = dn->dn_num_slots << DNODE_SHIFT;
1986 DB_DNODE_EXIT(db);
1987 }
1988
1989 void
1990 byteswap_uint64_array(void *vbuf, size_t size)
1991 {
1992 uint64_t *buf = vbuf;
1993 size_t count = size >> 3;
1994 int i;
1995
1996 ASSERT((size & 7) == 0);
1997
1998 for (i = 0; i < count; i++)
1999 buf[i] = BSWAP_64(buf[i]);
2000 }
2001
2002 void
2003 byteswap_uint32_array(void *vbuf, size_t size)
2004 {
2005 uint32_t *buf = vbuf;
2006 size_t count = size >> 2;
2007 int i;
2008
2009 ASSERT((size & 3) == 0);
2010
2011 for (i = 0; i < count; i++)
2012 buf[i] = BSWAP_32(buf[i]);
2013 }
2014
2015 void
2016 byteswap_uint16_array(void *vbuf, size_t size)
2017 {
2018 uint16_t *buf = vbuf;
2019 size_t count = size >> 1;
2020 int i;
2021
2022 ASSERT((size & 1) == 0);
2023
2024 for (i = 0; i < count; i++)
2025 buf[i] = BSWAP_16(buf[i]);
2026 }
2027
2028 /* ARGSUSED */
2029 void
2030 byteswap_uint8_array(void *vbuf, size_t size)
2031 {
2032 }
2033
2034 void
2035 dmu_init(void)
2036 {
2037 zfs_dbgmsg_init();
2038 sa_cache_init();
2039 xuio_stat_init();
2040 dmu_objset_init();
2041 dnode_init();
2042 dbuf_init();
2043 zfetch_init();
2044 dmu_tx_init();
2045 l2arc_init();
2046 arc_init();
2047 }
2048
2049 void
2050 dmu_fini(void)
2051 {
2052 arc_fini(); /* arc depends on l2arc, so arc must go first */
2053 l2arc_fini();
2054 dmu_tx_fini();
2055 zfetch_fini();
2056 dbuf_fini();
2057 dnode_fini();
2058 dmu_objset_fini();
2059 xuio_stat_fini();
2060 sa_cache_fini();
2061 zfs_dbgmsg_fini();
2062 }
2063
2064 #if defined(_KERNEL) && defined(HAVE_SPL)
2065 EXPORT_SYMBOL(dmu_bonus_hold);
2066 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus);
2067 EXPORT_SYMBOL(dmu_buf_rele_array);
2068 EXPORT_SYMBOL(dmu_prefetch);
2069 EXPORT_SYMBOL(dmu_free_range);
2070 EXPORT_SYMBOL(dmu_free_long_range);
2071 EXPORT_SYMBOL(dmu_free_long_object);
2072 EXPORT_SYMBOL(dmu_read);
2073 EXPORT_SYMBOL(dmu_write);
2074 EXPORT_SYMBOL(dmu_prealloc);
2075 EXPORT_SYMBOL(dmu_object_info);
2076 EXPORT_SYMBOL(dmu_object_info_from_dnode);
2077 EXPORT_SYMBOL(dmu_object_info_from_db);
2078 EXPORT_SYMBOL(dmu_object_size_from_db);
2079 EXPORT_SYMBOL(dmu_object_dnsize_from_db);
2080 EXPORT_SYMBOL(dmu_object_set_blocksize);
2081 EXPORT_SYMBOL(dmu_object_set_checksum);
2082 EXPORT_SYMBOL(dmu_object_set_compress);
2083 EXPORT_SYMBOL(dmu_write_policy);
2084 EXPORT_SYMBOL(dmu_sync);
2085 EXPORT_SYMBOL(dmu_request_arcbuf);
2086 EXPORT_SYMBOL(dmu_return_arcbuf);
2087 EXPORT_SYMBOL(dmu_assign_arcbuf);
2088 EXPORT_SYMBOL(dmu_buf_hold);
2089 EXPORT_SYMBOL(dmu_ot);
2090
2091 module_param(zfs_mdcomp_disable, int, 0644);
2092 MODULE_PARM_DESC(zfs_mdcomp_disable, "Disable meta data compression");
2093
2094 module_param(zfs_nopwrite_enabled, int, 0644);
2095 MODULE_PARM_DESC(zfs_nopwrite_enabled, "Enable NOP writes");
2096
2097 #endif
ZFS On Linux mirror for PVE