]> git.proxmox.com Git - mirror_zfs.git/blob - module/zfs/dmu.c
projects / mirror_zfs.git / blob
? search:


ade13b9f08d5f30b6c3d4e792fd21910ecf91e2d
[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) 2013 by Delphix. All rights reserved.
24 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
25 */
26
27 #include <sys/dmu.h>
28 #include <sys/dmu_impl.h>
29 #include <sys/dmu_tx.h>
30 #include <sys/dbuf.h>
31 #include <sys/dnode.h>
32 #include <sys/zfs_context.h>
33 #include <sys/dmu_objset.h>
34 #include <sys/dmu_traverse.h>
35 #include <sys/dsl_dataset.h>
36 #include <sys/dsl_dir.h>
37 #include <sys/dsl_pool.h>
38 #include <sys/dsl_synctask.h>
39 #include <sys/dsl_prop.h>
40 #include <sys/dmu_zfetch.h>
41 #include <sys/zfs_ioctl.h>
42 #include <sys/zap.h>
43 #include <sys/zio_checksum.h>
44 #include <sys/zio_compress.h>
45 #include <sys/sa.h>
46 #ifdef _KERNEL
47 #include <sys/vmsystm.h>
48 #include <sys/zfs_znode.h>
49 #endif
50
51 /*
52 * Enable/disable nopwrite feature.
53 */
54 int zfs_nopwrite_enabled = 1;
55
56 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
57 { DMU_BSWAP_UINT8, TRUE, "unallocated" },
58 { DMU_BSWAP_ZAP, TRUE, "object directory" },
59 { DMU_BSWAP_UINT64, TRUE, "object array" },
60 { DMU_BSWAP_UINT8, TRUE, "packed nvlist" },
61 { DMU_BSWAP_UINT64, TRUE, "packed nvlist size" },
62 { DMU_BSWAP_UINT64, TRUE, "bpobj" },
63 { DMU_BSWAP_UINT64, TRUE, "bpobj header" },
64 { DMU_BSWAP_UINT64, TRUE, "SPA space map header" },
65 { DMU_BSWAP_UINT64, TRUE, "SPA space map" },
66 { DMU_BSWAP_UINT64, TRUE, "ZIL intent log" },
67 { DMU_BSWAP_DNODE, TRUE, "DMU dnode" },
68 { DMU_BSWAP_OBJSET, TRUE, "DMU objset" },
69 { DMU_BSWAP_UINT64, TRUE, "DSL directory" },
70 { DMU_BSWAP_ZAP, TRUE, "DSL directory child map"},
71 { DMU_BSWAP_ZAP, TRUE, "DSL dataset snap map" },
72 { DMU_BSWAP_ZAP, TRUE, "DSL props" },
73 { DMU_BSWAP_UINT64, TRUE, "DSL dataset" },
74 { DMU_BSWAP_ZNODE, TRUE, "ZFS znode" },
75 { DMU_BSWAP_OLDACL, TRUE, "ZFS V0 ACL" },
76 { DMU_BSWAP_UINT8, FALSE, "ZFS plain file" },
77 { DMU_BSWAP_ZAP, TRUE, "ZFS directory" },
78 { DMU_BSWAP_ZAP, TRUE, "ZFS master node" },
79 { DMU_BSWAP_ZAP, TRUE, "ZFS delete queue" },
80 { DMU_BSWAP_UINT8, FALSE, "zvol object" },
81 { DMU_BSWAP_ZAP, TRUE, "zvol prop" },
82 { DMU_BSWAP_UINT8, FALSE, "other uint8[]" },
83 { DMU_BSWAP_UINT64, FALSE, "other uint64[]" },
84 { DMU_BSWAP_ZAP, TRUE, "other ZAP" },
85 { DMU_BSWAP_ZAP, TRUE, "persistent error log" },
86 { DMU_BSWAP_UINT8, TRUE, "SPA history" },
87 { DMU_BSWAP_UINT64, TRUE, "SPA history offsets" },
88 { DMU_BSWAP_ZAP, TRUE, "Pool properties" },
89 { DMU_BSWAP_ZAP, TRUE, "DSL permissions" },
90 { DMU_BSWAP_ACL, TRUE, "ZFS ACL" },
91 { DMU_BSWAP_UINT8, TRUE, "ZFS SYSACL" },
92 { DMU_BSWAP_UINT8, TRUE, "FUID table" },
93 { DMU_BSWAP_UINT64, TRUE, "FUID table size" },
94 { DMU_BSWAP_ZAP, TRUE, "DSL dataset next clones"},
95 { DMU_BSWAP_ZAP, TRUE, "scan work queue" },
96 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group used" },
97 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group quota" },
98 { DMU_BSWAP_ZAP, TRUE, "snapshot refcount tags"},
99 { DMU_BSWAP_ZAP, TRUE, "DDT ZAP algorithm" },
100 { DMU_BSWAP_ZAP, TRUE, "DDT statistics" },
101 { DMU_BSWAP_UINT8, TRUE, "System attributes" },
102 { DMU_BSWAP_ZAP, TRUE, "SA master node" },
103 { DMU_BSWAP_ZAP, TRUE, "SA attr registration" },
104 { DMU_BSWAP_ZAP, TRUE, "SA attr layouts" },
105 { DMU_BSWAP_ZAP, TRUE, "scan translations" },
106 { DMU_BSWAP_UINT8, FALSE, "deduplicated block" },
107 { DMU_BSWAP_ZAP, TRUE, "DSL deadlist map" },
108 { DMU_BSWAP_UINT64, TRUE, "DSL deadlist map hdr" },
109 { DMU_BSWAP_ZAP, TRUE, "DSL dir clones" },
110 { DMU_BSWAP_UINT64, TRUE, "bpobj subobj" }
111 };
112
113 const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
114 { byteswap_uint8_array, "uint8" },
115 { byteswap_uint16_array, "uint16" },
116 { byteswap_uint32_array, "uint32" },
117 { byteswap_uint64_array, "uint64" },
118 { zap_byteswap, "zap" },
119 { dnode_buf_byteswap, "dnode" },
120 { dmu_objset_byteswap, "objset" },
121 { zfs_znode_byteswap, "znode" },
122 { zfs_oldacl_byteswap, "oldacl" },
123 { zfs_acl_byteswap, "acl" }
124 };
125
126 int
127 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
128 void *tag, dmu_buf_t **dbp, int flags)
129 {
130 dnode_t *dn;
131 uint64_t blkid;
132 dmu_buf_impl_t *db;
133 int err;
134 int db_flags = DB_RF_CANFAIL;
135
136 if (flags & DMU_READ_NO_PREFETCH)
137 db_flags |= DB_RF_NOPREFETCH;
138
139 err = dnode_hold(os, object, FTAG, &dn);
140 if (err)
141 return (err);
142 blkid = dbuf_whichblock(dn, offset);
143 rw_enter(&dn->dn_struct_rwlock, RW_READER);
144 db = dbuf_hold(dn, blkid, tag);
145 rw_exit(&dn->dn_struct_rwlock);
146 if (db == NULL) {
147 err = SET_ERROR(EIO);
148 } else {
149 err = dbuf_read(db, NULL, db_flags);
150 if (err) {
151 dbuf_rele(db, tag);
152 db = NULL;
153 }
154 }
155
156 dnode_rele(dn, FTAG);
157 *dbp = &db->db; /* NULL db plus first field offset is NULL */
158 return (err);
159 }
160
161 int
162 dmu_bonus_max(void)
163 {
164 return (DN_MAX_BONUSLEN);
165 }
166
167 int
168 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
169 {
170 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
171 dnode_t *dn;
172 int error;
173
174 DB_DNODE_ENTER(db);
175 dn = DB_DNODE(db);
176
177 if (dn->dn_bonus != db) {
178 error = SET_ERROR(EINVAL);
179 } else if (newsize < 0 || newsize > db_fake->db_size) {
180 error = SET_ERROR(EINVAL);
181 } else {
182 dnode_setbonuslen(dn, newsize, tx);
183 error = 0;
184 }
185
186 DB_DNODE_EXIT(db);
187 return (error);
188 }
189
190 int
191 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
192 {
193 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
194 dnode_t *dn;
195 int error;
196
197 DB_DNODE_ENTER(db);
198 dn = DB_DNODE(db);
199
200 if (!DMU_OT_IS_VALID(type)) {
201 error = SET_ERROR(EINVAL);
202 } else if (dn->dn_bonus != db) {
203 error = SET_ERROR(EINVAL);
204 } else {
205 dnode_setbonus_type(dn, type, tx);
206 error = 0;
207 }
208
209 DB_DNODE_EXIT(db);
210 return (error);
211 }
212
213 dmu_object_type_t
214 dmu_get_bonustype(dmu_buf_t *db_fake)
215 {
216 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
217 dnode_t *dn;
218 dmu_object_type_t type;
219
220 DB_DNODE_ENTER(db);
221 dn = DB_DNODE(db);
222 type = dn->dn_bonustype;
223 DB_DNODE_EXIT(db);
224
225 return (type);
226 }
227
228 int
229 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
230 {
231 dnode_t *dn;
232 int error;
233
234 error = dnode_hold(os, object, FTAG, &dn);
235 dbuf_rm_spill(dn, tx);
236 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
237 dnode_rm_spill(dn, tx);
238 rw_exit(&dn->dn_struct_rwlock);
239 dnode_rele(dn, FTAG);
240 return (error);
241 }
242
243 /*
244 * returns ENOENT, EIO, or 0.
245 */
246 int
247 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
248 {
249 dnode_t *dn;
250 dmu_buf_impl_t *db;
251 int error;
252
253 error = dnode_hold(os, object, FTAG, &dn);
254 if (error)
255 return (error);
256
257 rw_enter(&dn->dn_struct_rwlock, RW_READER);
258 if (dn->dn_bonus == NULL) {
259 rw_exit(&dn->dn_struct_rwlock);
260 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
261 if (dn->dn_bonus == NULL)
262 dbuf_create_bonus(dn);
263 }
264 db = dn->dn_bonus;
265
266 /* as long as the bonus buf is held, the dnode will be held */
267 if (refcount_add(&db->db_holds, tag) == 1) {
268 VERIFY(dnode_add_ref(dn, db));
269 (void) atomic_inc_32_nv(&dn->dn_dbufs_count);
270 }
271
272 /*
273 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
274 * hold and incrementing the dbuf count to ensure that dnode_move() sees
275 * a dnode hold for every dbuf.
276 */
277 rw_exit(&dn->dn_struct_rwlock);
278
279 dnode_rele(dn, FTAG);
280
281 VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH));
282
283 *dbp = &db->db;
284 return (0);
285 }
286
287 /*
288 * returns ENOENT, EIO, or 0.
289 *
290 * This interface will allocate a blank spill dbuf when a spill blk
291 * doesn't already exist on the dnode.
292 *
293 * if you only want to find an already existing spill db, then
294 * dmu_spill_hold_existing() should be used.
295 */
296 int
297 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
298 {
299 dmu_buf_impl_t *db = NULL;
300 int err;
301
302 if ((flags & DB_RF_HAVESTRUCT) == 0)
303 rw_enter(&dn->dn_struct_rwlock, RW_READER);
304
305 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
306
307 if ((flags & DB_RF_HAVESTRUCT) == 0)
308 rw_exit(&dn->dn_struct_rwlock);
309
310 ASSERT(db != NULL);
311 err = dbuf_read(db, NULL, flags);
312 if (err == 0)
313 *dbp = &db->db;
314 else
315 dbuf_rele(db, tag);
316 return (err);
317 }
318
319 int
320 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
321 {
322 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
323 dnode_t *dn;
324 int err;
325
326 DB_DNODE_ENTER(db);
327 dn = DB_DNODE(db);
328
329 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
330 err = SET_ERROR(EINVAL);
331 } else {
332 rw_enter(&dn->dn_struct_rwlock, RW_READER);
333
334 if (!dn->dn_have_spill) {
335 err = SET_ERROR(ENOENT);
336 } else {
337 err = dmu_spill_hold_by_dnode(dn,
338 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
339 }
340
341 rw_exit(&dn->dn_struct_rwlock);
342 }
343
344 DB_DNODE_EXIT(db);
345 return (err);
346 }
347
348 int
349 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
350 {
351 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
352 dnode_t *dn;
353 int err;
354
355 DB_DNODE_ENTER(db);
356 dn = DB_DNODE(db);
357 err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp);
358 DB_DNODE_EXIT(db);
359
360 return (err);
361 }
362
363 /*
364 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
365 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
366 * and can induce severe lock contention when writing to several files
367 * whose dnodes are in the same block.
368 */
369 static int
370 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
371 int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
372 {
373 dmu_buf_t **dbp;
374 uint64_t blkid, nblks, i;
375 uint32_t dbuf_flags;
376 int err;
377 zio_t *zio;
378
379 ASSERT(length <= DMU_MAX_ACCESS);
380
381 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT;
382 if (flags & DMU_READ_NO_PREFETCH || length > zfetch_array_rd_sz)
383 dbuf_flags |= DB_RF_NOPREFETCH;
384
385 rw_enter(&dn->dn_struct_rwlock, RW_READER);
386 if (dn->dn_datablkshift) {
387 int blkshift = dn->dn_datablkshift;
388 nblks = (P2ROUNDUP(offset+length, 1ULL<<blkshift) -
389 P2ALIGN(offset, 1ULL<<blkshift)) >> blkshift;
390 } else {
391 if (offset + length > dn->dn_datablksz) {
392 zfs_panic_recover("zfs: accessing past end of object "
393 "%llx/%llx (size=%u access=%llu+%llu)",
394 (longlong_t)dn->dn_objset->
395 os_dsl_dataset->ds_object,
396 (longlong_t)dn->dn_object, dn->dn_datablksz,
397 (longlong_t)offset, (longlong_t)length);
398 rw_exit(&dn->dn_struct_rwlock);
399 return (SET_ERROR(EIO));
400 }
401 nblks = 1;
402 }
403 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_PUSHPAGE | KM_NODEBUG);
404
405 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
406 blkid = dbuf_whichblock(dn, offset);
407 for (i = 0; i < nblks; i++) {
408 dmu_buf_impl_t *db = dbuf_hold(dn, blkid+i, tag);
409 if (db == NULL) {
410 rw_exit(&dn->dn_struct_rwlock);
411 dmu_buf_rele_array(dbp, nblks, tag);
412 zio_nowait(zio);
413 return (SET_ERROR(EIO));
414 }
415 /* initiate async i/o */
416 if (read) {
417 (void) dbuf_read(db, zio, dbuf_flags);
418 }
419 dbp[i] = &db->db;
420 }
421 rw_exit(&dn->dn_struct_rwlock);
422
423 /* wait for async i/o */
424 err = zio_wait(zio);
425 if (err) {
426 dmu_buf_rele_array(dbp, nblks, tag);
427 return (err);
428 }
429
430 /* wait for other io to complete */
431 if (read) {
432 for (i = 0; i < nblks; i++) {
433 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
434 mutex_enter(&db->db_mtx);
435 while (db->db_state == DB_READ ||
436 db->db_state == DB_FILL)
437 cv_wait(&db->db_changed, &db->db_mtx);
438 if (db->db_state == DB_UNCACHED)
439 err = SET_ERROR(EIO);
440 mutex_exit(&db->db_mtx);
441 if (err) {
442 dmu_buf_rele_array(dbp, nblks, tag);
443 return (err);
444 }
445 }
446 }
447
448 *numbufsp = nblks;
449 *dbpp = dbp;
450 return (0);
451 }
452
453 static int
454 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
455 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
456 {
457 dnode_t *dn;
458 int err;
459
460 err = dnode_hold(os, object, FTAG, &dn);
461 if (err)
462 return (err);
463
464 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
465 numbufsp, dbpp, DMU_READ_PREFETCH);
466
467 dnode_rele(dn, FTAG);
468
469 return (err);
470 }
471
472 int
473 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
474 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
475 {
476 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
477 dnode_t *dn;
478 int err;
479
480 DB_DNODE_ENTER(db);
481 dn = DB_DNODE(db);
482 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
483 numbufsp, dbpp, DMU_READ_PREFETCH);
484 DB_DNODE_EXIT(db);
485
486 return (err);
487 }
488
489 void
490 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
491 {
492 int i;
493 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
494
495 if (numbufs == 0)
496 return;
497
498 for (i = 0; i < numbufs; i++) {
499 if (dbp[i])
500 dbuf_rele(dbp[i], tag);
501 }
502
503 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
504 }
505
506 /*
507 * Issue prefetch i/os for the given blocks.
508 *
509 * Note: The assumption is that we *know* these blocks will be needed
510 * almost immediately. Therefore, the prefetch i/os will be issued at
511 * ZIO_PRIORITY_SYNC_READ
512 *
513 * Note: indirect blocks and other metadata will be read synchronously,
514 * causing this function to block if they are not already cached.
515 */
516 void
517 dmu_prefetch(objset_t *os, uint64_t object, uint64_t offset, uint64_t len)
518 {
519 dnode_t *dn;
520 uint64_t blkid;
521 int nblks, err;
522
523 if (zfs_prefetch_disable)
524 return;
525
526 if (len == 0) { /* they're interested in the bonus buffer */
527 dn = DMU_META_DNODE(os);
528
529 if (object == 0 || object >= DN_MAX_OBJECT)
530 return;
531
532 rw_enter(&dn->dn_struct_rwlock, RW_READER);
533 blkid = dbuf_whichblock(dn, object * sizeof (dnode_phys_t));
534 dbuf_prefetch(dn, blkid, ZIO_PRIORITY_SYNC_READ);
535 rw_exit(&dn->dn_struct_rwlock);
536 return;
537 }
538
539 /*
540 * XXX - Note, if the dnode for the requested object is not
541 * already cached, we will do a *synchronous* read in the
542 * dnode_hold() call. The same is true for any indirects.
543 */
544 err = dnode_hold(os, object, FTAG, &dn);
545 if (err != 0)
546 return;
547
548 rw_enter(&dn->dn_struct_rwlock, RW_READER);
549 if (dn->dn_datablkshift) {
550 int blkshift = dn->dn_datablkshift;
551 nblks = (P2ROUNDUP(offset + len, 1 << blkshift) -
552 P2ALIGN(offset, 1 << blkshift)) >> blkshift;
553 } else {
554 nblks = (offset < dn->dn_datablksz);
555 }
556
557 if (nblks != 0) {
558 int i;
559
560 blkid = dbuf_whichblock(dn, offset);
561 for (i = 0; i < nblks; i++)
562 dbuf_prefetch(dn, blkid + i, ZIO_PRIORITY_SYNC_READ);
563 }
564
565 rw_exit(&dn->dn_struct_rwlock);
566
567 dnode_rele(dn, FTAG);
568 }
569
570 /*
571 * Get the next "chunk" of file data to free. We traverse the file from
572 * the end so that the file gets shorter over time (if we crashes in the
573 * middle, this will leave us in a better state). We find allocated file
574 * data by simply searching the allocated level 1 indirects.
575 *
576 * On input, *start should be the first offset that does not need to be
577 * freed (e.g. "offset + length"). On return, *start will be the first
578 * offset that should be freed.
579 */
580 static int
581 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum)
582 {
583 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
584 /* bytes of data covered by a level-1 indirect block */
585 uint64_t iblkrange =
586 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
587 uint64_t blks;
588
589 ASSERT3U(minimum, <=, *start);
590
591 if (*start - minimum <= iblkrange * maxblks) {
592 *start = minimum;
593 return (0);
594 }
595 ASSERT(ISP2(iblkrange));
596
597 for (blks = 0; *start > minimum && blks < maxblks; blks++) {
598 int err;
599
600 /*
601 * dnode_next_offset(BACKWARDS) will find an allocated L1
602 * indirect block at or before the input offset. We must
603 * decrement *start so that it is at the end of the region
604 * to search.
605 */
606 (*start)--;
607 err = dnode_next_offset(dn,
608 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
609
610 /* if there are no indirect blocks before start, we are done */
611 if (err == ESRCH) {
612 *start = minimum;
613 break;
614 } else if (err != 0) {
615 return (err);
616 }
617
618 /* set start to the beginning of this L1 indirect */
619 *start = P2ALIGN(*start, iblkrange);
620 }
621 if (*start < minimum)
622 *start = minimum;
623 return (0);
624 }
625
626 static int
627 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
628 uint64_t length)
629 {
630 uint64_t object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
631 int err;
632
633 if (offset >= object_size)
634 return (0);
635
636 if (length == DMU_OBJECT_END || offset + length > object_size)
637 length = object_size - offset;
638
639 while (length != 0) {
640 uint64_t chunk_end, chunk_begin;
641 dmu_tx_t *tx;
642
643 chunk_end = chunk_begin = offset + length;
644
645 /* move chunk_begin backwards to the beginning of this chunk */
646 err = get_next_chunk(dn, &chunk_begin, offset);
647 if (err)
648 return (err);
649 ASSERT3U(chunk_begin, >=, offset);
650 ASSERT3U(chunk_begin, <=, chunk_end);
651
652 tx = dmu_tx_create(os);
653 dmu_tx_hold_free(tx, dn->dn_object,
654 chunk_begin, chunk_end - chunk_begin);
655 err = dmu_tx_assign(tx, TXG_WAIT);
656 if (err) {
657 dmu_tx_abort(tx);
658 return (err);
659 }
660 dnode_free_range(dn, chunk_begin, chunk_end - chunk_begin, tx);
661 dmu_tx_commit(tx);
662
663 length -= chunk_end - chunk_begin;
664 }
665 return (0);
666 }
667
668 int
669 dmu_free_long_range(objset_t *os, uint64_t object,
670 uint64_t offset, uint64_t length)
671 {
672 dnode_t *dn;
673 int err;
674
675 err = dnode_hold(os, object, FTAG, &dn);
676 if (err != 0)
677 return (err);
678 err = dmu_free_long_range_impl(os, dn, offset, length);
679
680 /*
681 * It is important to zero out the maxblkid when freeing the entire
682 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
683 * will take the fast path, and (b) dnode_reallocate() can verify
684 * that the entire file has been freed.
685 */
686 if (offset == 0 && length == DMU_OBJECT_END)
687 dn->dn_maxblkid = 0;
688
689 dnode_rele(dn, FTAG);
690 return (err);
691 }
692
693 int
694 dmu_free_long_object(objset_t *os, uint64_t object)
695 {
696 dmu_tx_t *tx;
697 int err;
698
699 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
700 if (err != 0)
701 return (err);
702
703 tx = dmu_tx_create(os);
704 dmu_tx_hold_bonus(tx, object);
705 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
706 err = dmu_tx_assign(tx, TXG_WAIT);
707 if (err == 0) {
708 err = dmu_object_free(os, object, tx);
709 dmu_tx_commit(tx);
710 } else {
711 dmu_tx_abort(tx);
712 }
713
714 return (err);
715 }
716
717 int
718 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
719 uint64_t size, dmu_tx_t *tx)
720 {
721 dnode_t *dn;
722 int err = dnode_hold(os, object, FTAG, &dn);
723 if (err)
724 return (err);
725 ASSERT(offset < UINT64_MAX);
726 ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
727 dnode_free_range(dn, offset, size, tx);
728 dnode_rele(dn, FTAG);
729 return (0);
730 }
731
732 int
733 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
734 void *buf, uint32_t flags)
735 {
736 dnode_t *dn;
737 dmu_buf_t **dbp;
738 int numbufs, err;
739
740 err = dnode_hold(os, object, FTAG, &dn);
741 if (err)
742 return (err);
743
744 /*
745 * Deal with odd block sizes, where there can't be data past the first
746 * block. If we ever do the tail block optimization, we will need to
747 * handle that here as well.
748 */
749 if (dn->dn_maxblkid == 0) {
750 int newsz = offset > dn->dn_datablksz ? 0 :
751 MIN(size, dn->dn_datablksz - offset);
752 bzero((char *)buf + newsz, size - newsz);
753 size = newsz;
754 }
755
756 while (size > 0) {
757 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
758 int i;
759
760 /*
761 * NB: we could do this block-at-a-time, but it's nice
762 * to be reading in parallel.
763 */
764 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
765 TRUE, FTAG, &numbufs, &dbp, flags);
766 if (err)
767 break;
768
769 for (i = 0; i < numbufs; i++) {
770 int tocpy;
771 int bufoff;
772 dmu_buf_t *db = dbp[i];
773
774 ASSERT(size > 0);
775
776 bufoff = offset - db->db_offset;
777 tocpy = (int)MIN(db->db_size - bufoff, size);
778
779 bcopy((char *)db->db_data + bufoff, buf, tocpy);
780
781 offset += tocpy;
782 size -= tocpy;
783 buf = (char *)buf + tocpy;
784 }
785 dmu_buf_rele_array(dbp, numbufs, FTAG);
786 }
787 dnode_rele(dn, FTAG);
788 return (err);
789 }
790
791 void
792 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
793 const void *buf, dmu_tx_t *tx)
794 {
795 dmu_buf_t **dbp;
796 int numbufs, i;
797
798 if (size == 0)
799 return;
800
801 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
802 FALSE, FTAG, &numbufs, &dbp));
803
804 for (i = 0; i < numbufs; i++) {
805 int tocpy;
806 int bufoff;
807 dmu_buf_t *db = dbp[i];
808
809 ASSERT(size > 0);
810
811 bufoff = offset - db->db_offset;
812 tocpy = (int)MIN(db->db_size - bufoff, size);
813
814 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
815
816 if (tocpy == db->db_size)
817 dmu_buf_will_fill(db, tx);
818 else
819 dmu_buf_will_dirty(db, tx);
820
821 (void) memcpy((char *)db->db_data + bufoff, buf, tocpy);
822
823 if (tocpy == db->db_size)
824 dmu_buf_fill_done(db, tx);
825
826 offset += tocpy;
827 size -= tocpy;
828 buf = (char *)buf + tocpy;
829 }
830 dmu_buf_rele_array(dbp, numbufs, FTAG);
831 }
832
833 void
834 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
835 dmu_tx_t *tx)
836 {
837 dmu_buf_t **dbp;
838 int numbufs, i;
839
840 if (size == 0)
841 return;
842
843 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
844 FALSE, FTAG, &numbufs, &dbp));
845
846 for (i = 0; i < numbufs; i++) {
847 dmu_buf_t *db = dbp[i];
848
849 dmu_buf_will_not_fill(db, tx);
850 }
851 dmu_buf_rele_array(dbp, numbufs, FTAG);
852 }
853
854 /*
855 * DMU support for xuio
856 */
857 kstat_t *xuio_ksp = NULL;
858
859 typedef struct xuio_stats {
860 /* loaned yet not returned arc_buf */
861 kstat_named_t xuiostat_onloan_rbuf;
862 kstat_named_t xuiostat_onloan_wbuf;
863 /* whether a copy is made when loaning out a read buffer */
864 kstat_named_t xuiostat_rbuf_copied;
865 kstat_named_t xuiostat_rbuf_nocopy;
866 /* whether a copy is made when assigning a write buffer */
867 kstat_named_t xuiostat_wbuf_copied;
868 kstat_named_t xuiostat_wbuf_nocopy;
869 } xuio_stats_t;
870
871 static xuio_stats_t xuio_stats = {
872 { "onloan_read_buf", KSTAT_DATA_UINT64 },
873 { "onloan_write_buf", KSTAT_DATA_UINT64 },
874 { "read_buf_copied", KSTAT_DATA_UINT64 },
875 { "read_buf_nocopy", KSTAT_DATA_UINT64 },
876 { "write_buf_copied", KSTAT_DATA_UINT64 },
877 { "write_buf_nocopy", KSTAT_DATA_UINT64 }
878 };
879
880 #define XUIOSTAT_INCR(stat, val) \
881 atomic_add_64(&xuio_stats.stat.value.ui64, (val))
882 #define XUIOSTAT_BUMP(stat) XUIOSTAT_INCR(stat, 1)
883
884 int
885 dmu_xuio_init(xuio_t *xuio, int nblk)
886 {
887 dmu_xuio_t *priv;
888 uio_t *uio = &xuio->xu_uio;
889
890 uio->uio_iovcnt = nblk;
891 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_PUSHPAGE);
892
893 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_PUSHPAGE);
894 priv->cnt = nblk;
895 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_PUSHPAGE);
896 priv->iovp = uio->uio_iov;
897 XUIO_XUZC_PRIV(xuio) = priv;
898
899 if (XUIO_XUZC_RW(xuio) == UIO_READ)
900 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
901 else
902 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
903
904 return (0);
905 }
906
907 void
908 dmu_xuio_fini(xuio_t *xuio)
909 {
910 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
911 int nblk = priv->cnt;
912
913 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
914 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
915 kmem_free(priv, sizeof (dmu_xuio_t));
916
917 if (XUIO_XUZC_RW(xuio) == UIO_READ)
918 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
919 else
920 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
921 }
922
923 /*
924 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
925 * and increase priv->next by 1.
926 */
927 int
928 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
929 {
930 struct iovec *iov;
931 uio_t *uio = &xuio->xu_uio;
932 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
933 int i = priv->next++;
934
935 ASSERT(i < priv->cnt);
936 ASSERT(off + n <= arc_buf_size(abuf));
937 iov = uio->uio_iov + i;
938 iov->iov_base = (char *)abuf->b_data + off;
939 iov->iov_len = n;
940 priv->bufs[i] = abuf;
941 return (0);
942 }
943
944 int
945 dmu_xuio_cnt(xuio_t *xuio)
946 {
947 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
948 return (priv->cnt);
949 }
950
951 arc_buf_t *
952 dmu_xuio_arcbuf(xuio_t *xuio, int i)
953 {
954 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
955
956 ASSERT(i < priv->cnt);
957 return (priv->bufs[i]);
958 }
959
960 void
961 dmu_xuio_clear(xuio_t *xuio, int i)
962 {
963 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
964
965 ASSERT(i < priv->cnt);
966 priv->bufs[i] = NULL;
967 }
968
969 static void
970 xuio_stat_init(void)
971 {
972 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
973 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
974 KSTAT_FLAG_VIRTUAL);
975 if (xuio_ksp != NULL) {
976 xuio_ksp->ks_data = &xuio_stats;
977 kstat_install(xuio_ksp);
978 }
979 }
980
981 static void
982 xuio_stat_fini(void)
983 {
984 if (xuio_ksp != NULL) {
985 kstat_delete(xuio_ksp);
986 xuio_ksp = NULL;
987 }
988 }
989
990 void
991 xuio_stat_wbuf_copied()
992 {
993 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
994 }
995
996 void
997 xuio_stat_wbuf_nocopy()
998 {
999 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
1000 }
1001
1002 #ifdef _KERNEL
1003
1004 /*
1005 * Copy up to size bytes between arg_buf and req based on the data direction
1006 * described by the req. If an entire req's data cannot be transfered the
1007 * req's is updated such that it's current index and bv offsets correctly
1008 * reference any residual data which could not be copied. The return value
1009 * is the number of bytes successfully copied to arg_buf.
1010 */
1011 static int
1012 dmu_req_copy(void *arg_buf, int size, int *offset, struct request *req)
1013 {
1014 struct bio_vec *bv;
1015 struct req_iterator iter;
1016 char *bv_buf;
1017 int tocpy;
1018
1019 *offset = 0;
1020 rq_for_each_segment(bv, req, iter) {
1021
1022 /* Fully consumed the passed arg_buf */
1023 ASSERT3S(*offset, <=, size);
1024 if (size == *offset)
1025 break;
1026
1027 /* Skip fully consumed bv's */
1028 if (bv->bv_len == 0)
1029 continue;
1030
1031 tocpy = MIN(bv->bv_len, size - *offset);
1032 ASSERT3S(tocpy, >=, 0);
1033
1034 bv_buf = page_address(bv->bv_page) + bv->bv_offset;
1035 ASSERT3P(bv_buf, !=, NULL);
1036
1037 if (rq_data_dir(req) == WRITE)
1038 memcpy(arg_buf + *offset, bv_buf, tocpy);
1039 else
1040 memcpy(bv_buf, arg_buf + *offset, tocpy);
1041
1042 *offset += tocpy;
1043 bv->bv_offset += tocpy;
1044 bv->bv_len -= tocpy;
1045 }
1046
1047 return 0;
1048 }
1049
1050 static void
1051 dmu_bio_put(struct bio *bio)
1052 {
1053 struct bio *bio_next;
1054
1055 while (bio) {
1056 bio_next = bio->bi_next;
1057 bio_put(bio);
1058 bio = bio_next;
1059 }
1060 }
1061
1062 static int
1063 dmu_bio_clone(struct bio *bio, struct bio **bio_copy)
1064 {
1065 struct bio *bio_root = NULL;
1066 struct bio *bio_last = NULL;
1067 struct bio *bio_new;
1068
1069 if (bio == NULL)
1070 return EINVAL;
1071
1072 while (bio) {
1073 bio_new = bio_clone(bio, GFP_NOIO);
1074 if (bio_new == NULL) {
1075 dmu_bio_put(bio_root);
1076 return ENOMEM;
1077 }
1078
1079 if (bio_last) {
1080 bio_last->bi_next = bio_new;
1081 bio_last = bio_new;
1082 } else {
1083 bio_root = bio_new;
1084 bio_last = bio_new;
1085 }
1086
1087 bio = bio->bi_next;
1088 }
1089
1090 *bio_copy = bio_root;
1091
1092 return 0;
1093 }
1094
1095 int
1096 dmu_read_req(objset_t *os, uint64_t object, struct request *req)
1097 {
1098 uint64_t size = blk_rq_bytes(req);
1099 uint64_t offset = blk_rq_pos(req) << 9;
1100 struct bio *bio_saved = req->bio;
1101 dmu_buf_t **dbp;
1102 int numbufs, i, err;
1103
1104 /*
1105 * NB: we could do this block-at-a-time, but it's nice
1106 * to be reading in parallel.
1107 */
1108 err = dmu_buf_hold_array(os, object, offset, size, TRUE, FTAG,
1109 &numbufs, &dbp);
1110 if (err)
1111 return (err);
1112
1113 /*
1114 * Clone the bio list so the bv->bv_offset and bv->bv_len members
1115 * can be safely modified. The original bio list is relinked in to
1116 * the request when the function exits. This is required because
1117 * some file systems blindly assume that these values will remain
1118 * constant between bio_submit() and the IO completion callback.
1119 */
1120 err = dmu_bio_clone(bio_saved, &req->bio);
1121 if (err)
1122 goto error;
1123
1124 for (i = 0; i < numbufs; i++) {
1125 int tocpy, didcpy, bufoff;
1126 dmu_buf_t *db = dbp[i];
1127
1128 bufoff = offset - db->db_offset;
1129 ASSERT3S(bufoff, >=, 0);
1130
1131 tocpy = (int)MIN(db->db_size - bufoff, size);
1132 if (tocpy == 0)
1133 break;
1134
1135 err = dmu_req_copy(db->db_data + bufoff, tocpy, &didcpy, req);
1136
1137 if (didcpy < tocpy)
1138 err = EIO;
1139
1140 if (err)
1141 break;
1142
1143 size -= tocpy;
1144 offset += didcpy;
1145 err = 0;
1146 }
1147
1148 dmu_bio_put(req->bio);
1149 req->bio = bio_saved;
1150 error:
1151 dmu_buf_rele_array(dbp, numbufs, FTAG);
1152
1153 return (err);
1154 }
1155
1156 int
1157 dmu_write_req(objset_t *os, uint64_t object, struct request *req, dmu_tx_t *tx)
1158 {
1159 uint64_t size = blk_rq_bytes(req);
1160 uint64_t offset = blk_rq_pos(req) << 9;
1161 struct bio *bio_saved = req->bio;
1162 dmu_buf_t **dbp;
1163 int numbufs;
1164 int err = 0;
1165 int i;
1166
1167 if (size == 0)
1168 return (0);
1169
1170 err = dmu_buf_hold_array(os, object, offset, size, FALSE, FTAG,
1171 &numbufs, &dbp);
1172 if (err)
1173 return (err);
1174
1175 /*
1176 * Clone the bio list so the bv->bv_offset and bv->bv_len members
1177 * can be safely modified. The original bio list is relinked in to
1178 * the request when the function exits. This is required because
1179 * some file systems blindly assume that these values will remain
1180 * constant between bio_submit() and the IO completion callback.
1181 */
1182 err = dmu_bio_clone(bio_saved, &req->bio);
1183 if (err)
1184 goto error;
1185
1186 for (i = 0; i < numbufs; i++) {
1187 int tocpy, didcpy, bufoff;
1188 dmu_buf_t *db = dbp[i];
1189
1190 bufoff = offset - db->db_offset;
1191 ASSERT3S(bufoff, >=, 0);
1192
1193 tocpy = (int)MIN(db->db_size - bufoff, size);
1194 if (tocpy == 0)
1195 break;
1196
1197 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1198
1199 if (tocpy == db->db_size)
1200 dmu_buf_will_fill(db, tx);
1201 else
1202 dmu_buf_will_dirty(db, tx);
1203
1204 err = dmu_req_copy(db->db_data + bufoff, tocpy, &didcpy, req);
1205
1206 if (tocpy == db->db_size)
1207 dmu_buf_fill_done(db, tx);
1208
1209 if (didcpy < tocpy)
1210 err = EIO;
1211
1212 if (err)
1213 break;
1214
1215 size -= tocpy;
1216 offset += didcpy;
1217 err = 0;
1218 }
1219
1220 dmu_bio_put(req->bio);
1221 req->bio = bio_saved;
1222 error:
1223 dmu_buf_rele_array(dbp, numbufs, FTAG);
1224
1225 return (err);
1226 }
1227
1228 int
1229 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
1230 {
1231 dmu_buf_t **dbp;
1232 int numbufs, i, err;
1233 xuio_t *xuio = NULL;
1234
1235 /*
1236 * NB: we could do this block-at-a-time, but it's nice
1237 * to be reading in parallel.
1238 */
1239 err = dmu_buf_hold_array(os, object, uio->uio_loffset, size, TRUE, FTAG,
1240 &numbufs, &dbp);
1241 if (err)
1242 return (err);
1243
1244 for (i = 0; i < numbufs; i++) {
1245 int tocpy;
1246 int bufoff;
1247 dmu_buf_t *db = dbp[i];
1248
1249 ASSERT(size > 0);
1250
1251 bufoff = uio->uio_loffset - db->db_offset;
1252 tocpy = (int)MIN(db->db_size - bufoff, size);
1253
1254 if (xuio) {
1255 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
1256 arc_buf_t *dbuf_abuf = dbi->db_buf;
1257 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
1258 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
1259 if (!err) {
1260 uio->uio_resid -= tocpy;
1261 uio->uio_loffset += tocpy;
1262 }
1263
1264 if (abuf == dbuf_abuf)
1265 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
1266 else
1267 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
1268 } else {
1269 err = uiomove((char *)db->db_data + bufoff, tocpy,
1270 UIO_READ, uio);
1271 }
1272 if (err)
1273 break;
1274
1275 size -= tocpy;
1276 }
1277 dmu_buf_rele_array(dbp, numbufs, FTAG);
1278
1279 return (err);
1280 }
1281
1282 static int
1283 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1284 {
1285 dmu_buf_t **dbp;
1286 int numbufs;
1287 int err = 0;
1288 int i;
1289
1290 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1291 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1292 if (err)
1293 return (err);
1294
1295 for (i = 0; i < numbufs; i++) {
1296 int tocpy;
1297 int bufoff;
1298 dmu_buf_t *db = dbp[i];
1299
1300 ASSERT(size > 0);
1301
1302 bufoff = uio->uio_loffset - db->db_offset;
1303 tocpy = (int)MIN(db->db_size - bufoff, size);
1304
1305 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1306
1307 if (tocpy == db->db_size)
1308 dmu_buf_will_fill(db, tx);
1309 else
1310 dmu_buf_will_dirty(db, tx);
1311
1312 /*
1313 * XXX uiomove could block forever (eg.nfs-backed
1314 * pages). There needs to be a uiolockdown() function
1315 * to lock the pages in memory, so that uiomove won't
1316 * block.
1317 */
1318 err = uiomove((char *)db->db_data + bufoff, tocpy,
1319 UIO_WRITE, uio);
1320
1321 if (tocpy == db->db_size)
1322 dmu_buf_fill_done(db, tx);
1323
1324 if (err)
1325 break;
1326
1327 size -= tocpy;
1328 }
1329
1330 dmu_buf_rele_array(dbp, numbufs, FTAG);
1331 return (err);
1332 }
1333
1334 int
1335 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1336 dmu_tx_t *tx)
1337 {
1338 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1339 dnode_t *dn;
1340 int err;
1341
1342 if (size == 0)
1343 return (0);
1344
1345 DB_DNODE_ENTER(db);
1346 dn = DB_DNODE(db);
1347 err = dmu_write_uio_dnode(dn, uio, size, tx);
1348 DB_DNODE_EXIT(db);
1349
1350 return (err);
1351 }
1352
1353 int
1354 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1355 dmu_tx_t *tx)
1356 {
1357 dnode_t *dn;
1358 int err;
1359
1360 if (size == 0)
1361 return (0);
1362
1363 err = dnode_hold(os, object, FTAG, &dn);
1364 if (err)
1365 return (err);
1366
1367 err = dmu_write_uio_dnode(dn, uio, size, tx);
1368
1369 dnode_rele(dn, FTAG);
1370
1371 return (err);
1372 }
1373 #endif /* _KERNEL */
1374
1375 /*
1376 * Allocate a loaned anonymous arc buffer.
1377 */
1378 arc_buf_t *
1379 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1380 {
1381 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1382 spa_t *spa;
1383
1384 DB_GET_SPA(&spa, db);
1385 return (arc_loan_buf(spa, size));
1386 }
1387
1388 /*
1389 * Free a loaned arc buffer.
1390 */
1391 void
1392 dmu_return_arcbuf(arc_buf_t *buf)
1393 {
1394 arc_return_buf(buf, FTAG);
1395 VERIFY(arc_buf_remove_ref(buf, FTAG));
1396 }
1397
1398 /*
1399 * When possible directly assign passed loaned arc buffer to a dbuf.
1400 * If this is not possible copy the contents of passed arc buf via
1401 * dmu_write().
1402 */
1403 void
1404 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1405 dmu_tx_t *tx)
1406 {
1407 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1408 dnode_t *dn;
1409 dmu_buf_impl_t *db;
1410 uint32_t blksz = (uint32_t)arc_buf_size(buf);
1411 uint64_t blkid;
1412
1413 DB_DNODE_ENTER(dbuf);
1414 dn = DB_DNODE(dbuf);
1415 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1416 blkid = dbuf_whichblock(dn, offset);
1417 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1418 rw_exit(&dn->dn_struct_rwlock);
1419 DB_DNODE_EXIT(dbuf);
1420
1421 if (offset == db->db.db_offset && blksz == db->db.db_size) {
1422 dbuf_assign_arcbuf(db, buf, tx);
1423 dbuf_rele(db, FTAG);
1424 } else {
1425 objset_t *os;
1426 uint64_t object;
1427
1428 DB_DNODE_ENTER(dbuf);
1429 dn = DB_DNODE(dbuf);
1430 os = dn->dn_objset;
1431 object = dn->dn_object;
1432 DB_DNODE_EXIT(dbuf);
1433
1434 dbuf_rele(db, FTAG);
1435 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1436 dmu_return_arcbuf(buf);
1437 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1438 }
1439 }
1440
1441 typedef struct {
1442 dbuf_dirty_record_t *dsa_dr;
1443 dmu_sync_cb_t *dsa_done;
1444 zgd_t *dsa_zgd;
1445 dmu_tx_t *dsa_tx;
1446 } dmu_sync_arg_t;
1447
1448 /* ARGSUSED */
1449 static void
1450 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1451 {
1452 dmu_sync_arg_t *dsa = varg;
1453 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1454 blkptr_t *bp = zio->io_bp;
1455
1456 if (zio->io_error == 0) {
1457 if (BP_IS_HOLE(bp)) {
1458 /*
1459 * A block of zeros may compress to a hole, but the
1460 * block size still needs to be known for replay.
1461 */
1462 BP_SET_LSIZE(bp, db->db_size);
1463 } else {
1464 ASSERT(BP_GET_LEVEL(bp) == 0);
1465 bp->blk_fill = 1;
1466 }
1467 }
1468 }
1469
1470 static void
1471 dmu_sync_late_arrival_ready(zio_t *zio)
1472 {
1473 dmu_sync_ready(zio, NULL, zio->io_private);
1474 }
1475
1476 /* ARGSUSED */
1477 static void
1478 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1479 {
1480 dmu_sync_arg_t *dsa = varg;
1481 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1482 dmu_buf_impl_t *db = dr->dr_dbuf;
1483
1484 mutex_enter(&db->db_mtx);
1485 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1486 if (zio->io_error == 0) {
1487 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1488 if (dr->dt.dl.dr_nopwrite) {
1489 ASSERTV(blkptr_t *bp = zio->io_bp);
1490 ASSERTV(blkptr_t *bp_orig = &zio->io_bp_orig);
1491 ASSERTV(uint8_t chksum = BP_GET_CHECKSUM(bp_orig));
1492
1493 ASSERT(BP_EQUAL(bp, bp_orig));
1494 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1495 ASSERT(zio_checksum_table[chksum].ci_dedup);
1496 }
1497 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1498 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1499 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1500 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by))
1501 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1502 } else {
1503 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1504 }
1505 cv_broadcast(&db->db_changed);
1506 mutex_exit(&db->db_mtx);
1507
1508 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1509
1510 kmem_free(dsa, sizeof (*dsa));
1511 }
1512
1513 static void
1514 dmu_sync_late_arrival_done(zio_t *zio)
1515 {
1516 blkptr_t *bp = zio->io_bp;
1517 dmu_sync_arg_t *dsa = zio->io_private;
1518 ASSERTV(blkptr_t *bp_orig = &zio->io_bp_orig);
1519
1520 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1521 /*
1522 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE)
1523 * then there is nothing to do here. Otherwise, free the
1524 * newly allocated block in this txg.
1525 */
1526 if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
1527 ASSERT(BP_EQUAL(bp, bp_orig));
1528 } else {
1529 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1530 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1531 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1532 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1533 }
1534 }
1535
1536 dmu_tx_commit(dsa->dsa_tx);
1537
1538 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1539
1540 kmem_free(dsa, sizeof (*dsa));
1541 }
1542
1543 static int
1544 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1545 zio_prop_t *zp, zbookmark_t *zb)
1546 {
1547 dmu_sync_arg_t *dsa;
1548 dmu_tx_t *tx;
1549
1550 tx = dmu_tx_create(os);
1551 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1552 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1553 dmu_tx_abort(tx);
1554 /* Make zl_get_data do txg_waited_synced() */
1555 return (SET_ERROR(EIO));
1556 }
1557
1558 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_PUSHPAGE);
1559 dsa->dsa_dr = NULL;
1560 dsa->dsa_done = done;
1561 dsa->dsa_zgd = zgd;
1562 dsa->dsa_tx = tx;
1563
1564 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1565 zgd->zgd_db->db_data, zgd->zgd_db->db_size, zp,
1566 dmu_sync_late_arrival_ready, NULL, dmu_sync_late_arrival_done, dsa,
1567 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL | ZIO_FLAG_FASTWRITE, zb));
1568
1569 return (0);
1570 }
1571
1572 /*
1573 * Intent log support: sync the block associated with db to disk.
1574 * N.B. and XXX: the caller is responsible for making sure that the
1575 * data isn't changing while dmu_sync() is writing it.
1576 *
1577 * Return values:
1578 *
1579 * EEXIST: this txg has already been synced, so there's nothing to do.
1580 * The caller should not log the write.
1581 *
1582 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1583 * The caller should not log the write.
1584 *
1585 * EALREADY: this block is already in the process of being synced.
1586 * The caller should track its progress (somehow).
1587 *
1588 * EIO: could not do the I/O.
1589 * The caller should do a txg_wait_synced().
1590 *
1591 * 0: the I/O has been initiated.
1592 * The caller should log this blkptr in the done callback.
1593 * It is possible that the I/O will fail, in which case
1594 * the error will be reported to the done callback and
1595 * propagated to pio from zio_done().
1596 */
1597 int
1598 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1599 {
1600 blkptr_t *bp = zgd->zgd_bp;
1601 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1602 objset_t *os = db->db_objset;
1603 dsl_dataset_t *ds = os->os_dsl_dataset;
1604 dbuf_dirty_record_t *dr;
1605 dmu_sync_arg_t *dsa;
1606 zbookmark_t zb;
1607 zio_prop_t zp;
1608 dnode_t *dn;
1609
1610 ASSERT(pio != NULL);
1611 ASSERT(txg != 0);
1612
1613 SET_BOOKMARK(&zb, ds->ds_object,
1614 db->db.db_object, db->db_level, db->db_blkid);
1615
1616 DB_DNODE_ENTER(db);
1617 dn = DB_DNODE(db);
1618 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1619 DB_DNODE_EXIT(db);
1620
1621 /*
1622 * If we're frozen (running ziltest), we always need to generate a bp.
1623 */
1624 if (txg > spa_freeze_txg(os->os_spa))
1625 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1626
1627 /*
1628 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1629 * and us. If we determine that this txg is not yet syncing,
1630 * but it begins to sync a moment later, that's OK because the
1631 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1632 */
1633 mutex_enter(&db->db_mtx);
1634
1635 if (txg <= spa_last_synced_txg(os->os_spa)) {
1636 /*
1637 * This txg has already synced. There's nothing to do.
1638 */
1639 mutex_exit(&db->db_mtx);
1640 return (SET_ERROR(EEXIST));
1641 }
1642
1643 if (txg <= spa_syncing_txg(os->os_spa)) {
1644 /*
1645 * This txg is currently syncing, so we can't mess with
1646 * the dirty record anymore; just write a new log block.
1647 */
1648 mutex_exit(&db->db_mtx);
1649 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1650 }
1651
1652 dr = db->db_last_dirty;
1653 while (dr && dr->dr_txg != txg)
1654 dr = dr->dr_next;
1655
1656 if (dr == NULL) {
1657 /*
1658 * There's no dr for this dbuf, so it must have been freed.
1659 * There's no need to log writes to freed blocks, so we're done.
1660 */
1661 mutex_exit(&db->db_mtx);
1662 return (SET_ERROR(ENOENT));
1663 }
1664
1665 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
1666
1667 /*
1668 * Assume the on-disk data is X, the current syncing data is Y,
1669 * and the current in-memory data is Z (currently in dmu_sync).
1670 * X and Z are identical but Y is has been modified. Normally,
1671 * when X and Z are the same we will perform a nopwrite but if Y
1672 * is different we must disable nopwrite since the resulting write
1673 * of Y to disk can free the block containing X. If we allowed a
1674 * nopwrite to occur the block pointing to Z would reference a freed
1675 * block. Since this is a rare case we simplify this by disabling
1676 * nopwrite if the current dmu_sync-ing dbuf has been modified in
1677 * a previous transaction.
1678 */
1679 if (dr->dr_next)
1680 zp.zp_nopwrite = B_FALSE;
1681
1682 ASSERT(dr->dr_txg == txg);
1683 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1684 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1685 /*
1686 * We have already issued a sync write for this buffer,
1687 * or this buffer has already been synced. It could not
1688 * have been dirtied since, or we would have cleared the state.
1689 */
1690 mutex_exit(&db->db_mtx);
1691 return (SET_ERROR(EALREADY));
1692 }
1693
1694 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1695 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1696 mutex_exit(&db->db_mtx);
1697
1698 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_PUSHPAGE);
1699 dsa->dsa_dr = dr;
1700 dsa->dsa_done = done;
1701 dsa->dsa_zgd = zgd;
1702 dsa->dsa_tx = NULL;
1703
1704 zio_nowait(arc_write(pio, os->os_spa, txg,
1705 bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
1706 DBUF_IS_L2COMPRESSIBLE(db), &zp, dmu_sync_ready,
1707 NULL, dmu_sync_done, dsa, ZIO_PRIORITY_SYNC_WRITE,
1708 ZIO_FLAG_CANFAIL, &zb));
1709
1710 return (0);
1711 }
1712
1713 int
1714 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1715 dmu_tx_t *tx)
1716 {
1717 dnode_t *dn;
1718 int err;
1719
1720 err = dnode_hold(os, object, FTAG, &dn);
1721 if (err)
1722 return (err);
1723 err = dnode_set_blksz(dn, size, ibs, tx);
1724 dnode_rele(dn, FTAG);
1725 return (err);
1726 }
1727
1728 void
1729 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1730 dmu_tx_t *tx)
1731 {
1732 dnode_t *dn;
1733
1734 /* XXX assumes dnode_hold will not get an i/o error */
1735 (void) dnode_hold(os, object, FTAG, &dn);
1736 ASSERT(checksum < ZIO_CHECKSUM_FUNCTIONS);
1737 dn->dn_checksum = checksum;
1738 dnode_setdirty(dn, tx);
1739 dnode_rele(dn, FTAG);
1740 }
1741
1742 void
1743 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1744 dmu_tx_t *tx)
1745 {
1746 dnode_t *dn;
1747
1748 /* XXX assumes dnode_hold will not get an i/o error */
1749 (void) dnode_hold(os, object, FTAG, &dn);
1750 ASSERT(compress < ZIO_COMPRESS_FUNCTIONS);
1751 dn->dn_compress = compress;
1752 dnode_setdirty(dn, tx);
1753 dnode_rele(dn, FTAG);
1754 }
1755
1756 int zfs_mdcomp_disable = 0;
1757
1758 void
1759 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
1760 {
1761 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
1762 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
1763 (wp & WP_SPILL));
1764 enum zio_checksum checksum = os->os_checksum;
1765 enum zio_compress compress = os->os_compress;
1766 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
1767 boolean_t dedup = B_FALSE;
1768 boolean_t nopwrite = B_FALSE;
1769 boolean_t dedup_verify = os->os_dedup_verify;
1770 int copies = os->os_copies;
1771
1772 /*
1773 * We maintain different write policies for each of the following
1774 * types of data:
1775 * 1. metadata
1776 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1777 * 3. all other level 0 blocks
1778 */
1779 if (ismd) {
1780 /*
1781 * XXX -- we should design a compression algorithm
1782 * that specializes in arrays of bps.
1783 */
1784 compress = zfs_mdcomp_disable ? ZIO_COMPRESS_EMPTY :
1785 ZIO_COMPRESS_LZJB;
1786
1787 /*
1788 * Metadata always gets checksummed. If the data
1789 * checksum is multi-bit correctable, and it's not a
1790 * ZBT-style checksum, then it's suitable for metadata
1791 * as well. Otherwise, the metadata checksum defaults
1792 * to fletcher4.
1793 */
1794 if (zio_checksum_table[checksum].ci_correctable < 1 ||
1795 zio_checksum_table[checksum].ci_eck)
1796 checksum = ZIO_CHECKSUM_FLETCHER_4;
1797 } else if (wp & WP_NOFILL) {
1798 ASSERT(level == 0);
1799
1800 /*
1801 * If we're writing preallocated blocks, we aren't actually
1802 * writing them so don't set any policy properties. These
1803 * blocks are currently only used by an external subsystem
1804 * outside of zfs (i.e. dump) and not written by the zio
1805 * pipeline.
1806 */
1807 compress = ZIO_COMPRESS_OFF;
1808 checksum = ZIO_CHECKSUM_OFF;
1809 } else {
1810 compress = zio_compress_select(dn->dn_compress, compress);
1811
1812 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
1813 zio_checksum_select(dn->dn_checksum, checksum) :
1814 dedup_checksum;
1815
1816 /*
1817 * Determine dedup setting. If we are in dmu_sync(),
1818 * we won't actually dedup now because that's all
1819 * done in syncing context; but we do want to use the
1820 * dedup checkum. If the checksum is not strong
1821 * enough to ensure unique signatures, force
1822 * dedup_verify.
1823 */
1824 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
1825 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
1826 if (!zio_checksum_table[checksum].ci_dedup)
1827 dedup_verify = B_TRUE;
1828 }
1829
1830 /*
1831 * Enable nopwrite if we have a cryptographically secure
1832 * checksum that has no known collisions (i.e. SHA-256)
1833 * and compression is enabled. We don't enable nopwrite if
1834 * dedup is enabled as the two features are mutually exclusive.
1835 */
1836 nopwrite = (!dedup && zio_checksum_table[checksum].ci_dedup &&
1837 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
1838 }
1839
1840 zp->zp_checksum = checksum;
1841 zp->zp_compress = compress;
1842 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
1843 zp->zp_level = level;
1844 zp->zp_copies = MIN(copies + ismd, spa_max_replication(os->os_spa));
1845 zp->zp_dedup = dedup;
1846 zp->zp_dedup_verify = dedup && dedup_verify;
1847 zp->zp_nopwrite = nopwrite;
1848 }
1849
1850 int
1851 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
1852 {
1853 dnode_t *dn;
1854 int i, err;
1855
1856 err = dnode_hold(os, object, FTAG, &dn);
1857 if (err)
1858 return (err);
1859 /*
1860 * Sync any current changes before
1861 * we go trundling through the block pointers.
1862 */
1863 for (i = 0; i < TXG_SIZE; i++) {
1864 if (list_link_active(&dn->dn_dirty_link[i]))
1865 break;
1866 }
1867 if (i != TXG_SIZE) {
1868 dnode_rele(dn, FTAG);
1869 txg_wait_synced(dmu_objset_pool(os), 0);
1870 err = dnode_hold(os, object, FTAG, &dn);
1871 if (err)
1872 return (err);
1873 }
1874
1875 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
1876 dnode_rele(dn, FTAG);
1877
1878 return (err);
1879 }
1880
1881 void
1882 __dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
1883 {
1884 dnode_phys_t *dnp = dn->dn_phys;
1885 int i;
1886
1887 doi->doi_data_block_size = dn->dn_datablksz;
1888 doi->doi_metadata_block_size = dn->dn_indblkshift ?
1889 1ULL << dn->dn_indblkshift : 0;
1890 doi->doi_type = dn->dn_type;
1891 doi->doi_bonus_type = dn->dn_bonustype;
1892 doi->doi_bonus_size = dn->dn_bonuslen;
1893 doi->doi_indirection = dn->dn_nlevels;
1894 doi->doi_checksum = dn->dn_checksum;
1895 doi->doi_compress = dn->dn_compress;
1896 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
1897 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
1898 doi->doi_fill_count = 0;
1899 for (i = 0; i < dnp->dn_nblkptr; i++)
1900 doi->doi_fill_count += dnp->dn_blkptr[i].blk_fill;
1901 }
1902
1903 void
1904 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
1905 {
1906 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1907 mutex_enter(&dn->dn_mtx);
1908
1909 __dmu_object_info_from_dnode(dn, doi);
1910
1911 mutex_exit(&dn->dn_mtx);
1912 rw_exit(&dn->dn_struct_rwlock);
1913 }
1914
1915 /*
1916 * Get information on a DMU object.
1917 * If doi is NULL, just indicates whether the object exists.
1918 */
1919 int
1920 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
1921 {
1922 dnode_t *dn;
1923 int err = dnode_hold(os, object, FTAG, &dn);
1924
1925 if (err)
1926 return (err);
1927
1928 if (doi != NULL)
1929 dmu_object_info_from_dnode(dn, doi);
1930
1931 dnode_rele(dn, FTAG);
1932 return (0);
1933 }
1934
1935 /*
1936 * As above, but faster; can be used when you have a held dbuf in hand.
1937 */
1938 void
1939 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
1940 {
1941 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1942
1943 DB_DNODE_ENTER(db);
1944 dmu_object_info_from_dnode(DB_DNODE(db), doi);
1945 DB_DNODE_EXIT(db);
1946 }
1947
1948 /*
1949 * Faster still when you only care about the size.
1950 * This is specifically optimized for zfs_getattr().
1951 */
1952 void
1953 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
1954 u_longlong_t *nblk512)
1955 {
1956 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1957 dnode_t *dn;
1958
1959 DB_DNODE_ENTER(db);
1960 dn = DB_DNODE(db);
1961
1962 *blksize = dn->dn_datablksz;
1963 /* add 1 for dnode space */
1964 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
1965 SPA_MINBLOCKSHIFT) + 1;
1966 DB_DNODE_EXIT(db);
1967 }
1968
1969 void
1970 byteswap_uint64_array(void *vbuf, size_t size)
1971 {
1972 uint64_t *buf = vbuf;
1973 size_t count = size >> 3;
1974 int i;
1975
1976 ASSERT((size & 7) == 0);
1977
1978 for (i = 0; i < count; i++)
1979 buf[i] = BSWAP_64(buf[i]);
1980 }
1981
1982 void
1983 byteswap_uint32_array(void *vbuf, size_t size)
1984 {
1985 uint32_t *buf = vbuf;
1986 size_t count = size >> 2;
1987 int i;
1988
1989 ASSERT((size & 3) == 0);
1990
1991 for (i = 0; i < count; i++)
1992 buf[i] = BSWAP_32(buf[i]);
1993 }
1994
1995 void
1996 byteswap_uint16_array(void *vbuf, size_t size)
1997 {
1998 uint16_t *buf = vbuf;
1999 size_t count = size >> 1;
2000 int i;
2001
2002 ASSERT((size & 1) == 0);
2003
2004 for (i = 0; i < count; i++)
2005 buf[i] = BSWAP_16(buf[i]);
2006 }
2007
2008 /* ARGSUSED */
2009 void
2010 byteswap_uint8_array(void *vbuf, size_t size)
2011 {
2012 }
2013
2014 void
2015 dmu_init(void)
2016 {
2017 zfs_dbgmsg_init();
2018 sa_cache_init();
2019 xuio_stat_init();
2020 dmu_objset_init();
2021 dnode_init();
2022 dbuf_init();
2023 zfetch_init();
2024 dmu_tx_init();
2025 l2arc_init();
2026 arc_init();
2027 }
2028
2029 void
2030 dmu_fini(void)
2031 {
2032 arc_fini(); /* arc depends on l2arc, so arc must go first */
2033 l2arc_fini();
2034 dmu_tx_fini();
2035 zfetch_fini();
2036 dbuf_fini();
2037 dnode_fini();
2038 dmu_objset_fini();
2039 xuio_stat_fini();
2040 sa_cache_fini();
2041 zfs_dbgmsg_fini();
2042 }
2043
2044 #if defined(_KERNEL) && defined(HAVE_SPL)
2045 EXPORT_SYMBOL(dmu_bonus_hold);
2046 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus);
2047 EXPORT_SYMBOL(dmu_buf_rele_array);
2048 EXPORT_SYMBOL(dmu_prefetch);
2049 EXPORT_SYMBOL(dmu_free_range);
2050 EXPORT_SYMBOL(dmu_free_long_range);
2051 EXPORT_SYMBOL(dmu_free_long_object);
2052 EXPORT_SYMBOL(dmu_read);
2053 EXPORT_SYMBOL(dmu_write);
2054 EXPORT_SYMBOL(dmu_prealloc);
2055 EXPORT_SYMBOL(dmu_object_info);
2056 EXPORT_SYMBOL(dmu_object_info_from_dnode);
2057 EXPORT_SYMBOL(dmu_object_info_from_db);
2058 EXPORT_SYMBOL(dmu_object_size_from_db);
2059 EXPORT_SYMBOL(dmu_object_set_blocksize);
2060 EXPORT_SYMBOL(dmu_object_set_checksum);
2061 EXPORT_SYMBOL(dmu_object_set_compress);
2062 EXPORT_SYMBOL(dmu_write_policy);
2063 EXPORT_SYMBOL(dmu_sync);
2064 EXPORT_SYMBOL(dmu_request_arcbuf);
2065 EXPORT_SYMBOL(dmu_return_arcbuf);
2066 EXPORT_SYMBOL(dmu_assign_arcbuf);
2067 EXPORT_SYMBOL(dmu_buf_hold);
2068 EXPORT_SYMBOL(dmu_ot);
2069
2070 module_param(zfs_mdcomp_disable, int, 0644);
2071 MODULE_PARM_DESC(zfs_mdcomp_disable, "Disable meta data compression");
2072
2073 module_param(zfs_nopwrite_enabled, int, 0644);
2074 MODULE_PARM_DESC(zfs_nopwrite_enabled, "Enable NOP writes");
2075
2076 #endif
ZFS On Linux mirror for PVE