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, 2014 by Delphix. All rights reserved.
24 */
25 /* Copyright (c) 2013 by Saso Kiselkov. All rights reserved. */
26 /* Copyright (c) 2013, Joyent, Inc. All rights reserved. */
27 /* Copyright (c) 2014, Nexenta Systems, Inc. All rights reserved. */
28
29 #include <sys/dmu.h>
30 #include <sys/dmu_impl.h>
31 #include <sys/dmu_tx.h>
32 #include <sys/dbuf.h>
33 #include <sys/dnode.h>
34 #include <sys/zfs_context.h>
35 #include <sys/dmu_objset.h>
36 #include <sys/dmu_traverse.h>
37 #include <sys/dsl_dataset.h>
38 #include <sys/dsl_dir.h>
39 #include <sys/dsl_pool.h>
40 #include <sys/dsl_synctask.h>
41 #include <sys/dsl_prop.h>
42 #include <sys/dmu_zfetch.h>
43 #include <sys/zfs_ioctl.h>
44 #include <sys/zap.h>
45 #include <sys/zio_checksum.h>
46 #include <sys/zio_compress.h>
47 #include <sys/sa.h>
48 #include <sys/zfeature.h>
49 #ifdef _KERNEL
50 #include <sys/vm.h>
51 #include <sys/zfs_znode.h>
52 #endif
53
54 /*
55 * Enable/disable nopwrite feature.
56 */
57 int zfs_nopwrite_enabled = 1;
58 SYSCTL_DECL(_vfs_zfs);
59 TUNABLE_INT("vfs.zfs.nopwrite_enabled", &zfs_nopwrite_enabled);
60 SYSCTL_INT(_vfs_zfs, OID_AUTO, nopwrite_enabled, CTLFLAG_RDTUN,
61 &zfs_nopwrite_enabled, 0, "Enable nopwrite feature");
62
63 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
64 { DMU_BSWAP_UINT8, TRUE, "unallocated" },
65 { DMU_BSWAP_ZAP, TRUE, "object directory" },
66 { DMU_BSWAP_UINT64, TRUE, "object array" },
67 { DMU_BSWAP_UINT8, TRUE, "packed nvlist" },
68 { DMU_BSWAP_UINT64, TRUE, "packed nvlist size" },
69 { DMU_BSWAP_UINT64, TRUE, "bpobj" },
70 { DMU_BSWAP_UINT64, TRUE, "bpobj header" },
71 { DMU_BSWAP_UINT64, TRUE, "SPA space map header" },
72 { DMU_BSWAP_UINT64, TRUE, "SPA space map" },
73 { DMU_BSWAP_UINT64, TRUE, "ZIL intent log" },
74 { DMU_BSWAP_DNODE, TRUE, "DMU dnode" },
75 { DMU_BSWAP_OBJSET, TRUE, "DMU objset" },
76 { DMU_BSWAP_UINT64, TRUE, "DSL directory" },
77 { DMU_BSWAP_ZAP, TRUE, "DSL directory child map"},
78 { DMU_BSWAP_ZAP, TRUE, "DSL dataset snap map" },
79 { DMU_BSWAP_ZAP, TRUE, "DSL props" },
80 { DMU_BSWAP_UINT64, TRUE, "DSL dataset" },
81 { DMU_BSWAP_ZNODE, TRUE, "ZFS znode" },
82 { DMU_BSWAP_OLDACL, TRUE, "ZFS V0 ACL" },
83 { DMU_BSWAP_UINT8, FALSE, "ZFS plain file" },
84 { DMU_BSWAP_ZAP, TRUE, "ZFS directory" },
85 { DMU_BSWAP_ZAP, TRUE, "ZFS master node" },
86 { DMU_BSWAP_ZAP, TRUE, "ZFS delete queue" },
87 { DMU_BSWAP_UINT8, FALSE, "zvol object" },
88 { DMU_BSWAP_ZAP, TRUE, "zvol prop" },
89 { DMU_BSWAP_UINT8, FALSE, "other uint8[]" },
90 { DMU_BSWAP_UINT64, FALSE, "other uint64[]" },
91 { DMU_BSWAP_ZAP, TRUE, "other ZAP" },
92 { DMU_BSWAP_ZAP, TRUE, "persistent error log" },
93 { DMU_BSWAP_UINT8, TRUE, "SPA history" },
94 { DMU_BSWAP_UINT64, TRUE, "SPA history offsets" },
95 { DMU_BSWAP_ZAP, TRUE, "Pool properties" },
96 { DMU_BSWAP_ZAP, TRUE, "DSL permissions" },
97 { DMU_BSWAP_ACL, TRUE, "ZFS ACL" },
98 { DMU_BSWAP_UINT8, TRUE, "ZFS SYSACL" },
99 { DMU_BSWAP_UINT8, TRUE, "FUID table" },
100 { DMU_BSWAP_UINT64, TRUE, "FUID table size" },
101 { DMU_BSWAP_ZAP, TRUE, "DSL dataset next clones"},
102 { DMU_BSWAP_ZAP, TRUE, "scan work queue" },
103 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group used" },
104 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group quota" },
105 { DMU_BSWAP_ZAP, TRUE, "snapshot refcount tags"},
106 { DMU_BSWAP_ZAP, TRUE, "DDT ZAP algorithm" },
107 { DMU_BSWAP_ZAP, TRUE, "DDT statistics" },
108 { DMU_BSWAP_UINT8, TRUE, "System attributes" },
109 { DMU_BSWAP_ZAP, TRUE, "SA master node" },
110 { DMU_BSWAP_ZAP, TRUE, "SA attr registration" },
111 { DMU_BSWAP_ZAP, TRUE, "SA attr layouts" },
112 { DMU_BSWAP_ZAP, TRUE, "scan translations" },
113 { DMU_BSWAP_UINT8, FALSE, "deduplicated block" },
114 { DMU_BSWAP_ZAP, TRUE, "DSL deadlist map" },
115 { DMU_BSWAP_UINT64, TRUE, "DSL deadlist map hdr" },
116 { DMU_BSWAP_ZAP, TRUE, "DSL dir clones" },
117 { DMU_BSWAP_UINT64, TRUE, "bpobj subobj" }
118 };
119
120 const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
121 { byteswap_uint8_array, "uint8" },
122 { byteswap_uint16_array, "uint16" },
123 { byteswap_uint32_array, "uint32" },
124 { byteswap_uint64_array, "uint64" },
125 { zap_byteswap, "zap" },
126 { dnode_buf_byteswap, "dnode" },
127 { dmu_objset_byteswap, "objset" },
128 { zfs_znode_byteswap, "znode" },
129 { zfs_oldacl_byteswap, "oldacl" },
130 { zfs_acl_byteswap, "acl" }
131 };
132
133 int
dmu_buf_hold_noread(objset_t * os,uint64_t object,uint64_t offset,void * tag,dmu_buf_t ** dbp)134 dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset,
135 void *tag, dmu_buf_t **dbp)
136 {
137 dnode_t *dn;
138 uint64_t blkid;
139 dmu_buf_impl_t *db;
140 int err;
141
142 err = dnode_hold(os, object, FTAG, &dn);
143 if (err)
144 return (err);
145 blkid = dbuf_whichblock(dn, offset);
146 rw_enter(&dn->dn_struct_rwlock, RW_READER);
147 db = dbuf_hold(dn, blkid, tag);
148 rw_exit(&dn->dn_struct_rwlock);
149 dnode_rele(dn, FTAG);
150
151 if (db == NULL) {
152 *dbp = NULL;
153 return (SET_ERROR(EIO));
154 }
155
156 *dbp = &db->db;
157 return (err);
158 }
159
160 int
dmu_buf_hold(objset_t * os,uint64_t object,uint64_t offset,void * tag,dmu_buf_t ** dbp,int flags)161 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
162 void *tag, dmu_buf_t **dbp, int flags)
163 {
164 int err;
165 int db_flags = DB_RF_CANFAIL;
166
167 if (flags & DMU_READ_NO_PREFETCH)
168 db_flags |= DB_RF_NOPREFETCH;
169
170 err = dmu_buf_hold_noread(os, object, offset, tag, dbp);
171 if (err == 0) {
172 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
173 err = dbuf_read(db, NULL, db_flags);
174 if (err != 0) {
175 dbuf_rele(db, tag);
176 *dbp = NULL;
177 }
178 }
179
180 return (err);
181 }
182
183 int
dmu_bonus_max(void)184 dmu_bonus_max(void)
185 {
186 return (DN_MAX_BONUSLEN);
187 }
188
189 int
dmu_set_bonus(dmu_buf_t * db_fake,int newsize,dmu_tx_t * tx)190 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
191 {
192 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
193 dnode_t *dn;
194 int error;
195
196 DB_DNODE_ENTER(db);
197 dn = DB_DNODE(db);
198
199 if (dn->dn_bonus != db) {
200 error = SET_ERROR(EINVAL);
201 } else if (newsize < 0 || newsize > db_fake->db_size) {
202 error = SET_ERROR(EINVAL);
203 } else {
204 dnode_setbonuslen(dn, newsize, tx);
205 error = 0;
206 }
207
208 DB_DNODE_EXIT(db);
209 return (error);
210 }
211
212 int
dmu_set_bonustype(dmu_buf_t * db_fake,dmu_object_type_t type,dmu_tx_t * tx)213 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
214 {
215 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
216 dnode_t *dn;
217 int error;
218
219 DB_DNODE_ENTER(db);
220 dn = DB_DNODE(db);
221
222 if (!DMU_OT_IS_VALID(type)) {
223 error = SET_ERROR(EINVAL);
224 } else if (dn->dn_bonus != db) {
225 error = SET_ERROR(EINVAL);
226 } else {
227 dnode_setbonus_type(dn, type, tx);
228 error = 0;
229 }
230
231 DB_DNODE_EXIT(db);
232 return (error);
233 }
234
235 dmu_object_type_t
dmu_get_bonustype(dmu_buf_t * db_fake)236 dmu_get_bonustype(dmu_buf_t *db_fake)
237 {
238 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
239 dnode_t *dn;
240 dmu_object_type_t type;
241
242 DB_DNODE_ENTER(db);
243 dn = DB_DNODE(db);
244 type = dn->dn_bonustype;
245 DB_DNODE_EXIT(db);
246
247 return (type);
248 }
249
250 int
dmu_rm_spill(objset_t * os,uint64_t object,dmu_tx_t * tx)251 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
252 {
253 dnode_t *dn;
254 int error;
255
256 error = dnode_hold(os, object, FTAG, &dn);
257 dbuf_rm_spill(dn, tx);
258 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
259 dnode_rm_spill(dn, tx);
260 rw_exit(&dn->dn_struct_rwlock);
261 dnode_rele(dn, FTAG);
262 return (error);
263 }
264
265 /*
266 * returns ENOENT, EIO, or 0.
267 */
268 int
dmu_bonus_hold(objset_t * os,uint64_t object,void * tag,dmu_buf_t ** dbp)269 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
270 {
271 dnode_t *dn;
272 dmu_buf_impl_t *db;
273 int error;
274
275 error = dnode_hold(os, object, FTAG, &dn);
276 if (error)
277 return (error);
278
279 rw_enter(&dn->dn_struct_rwlock, RW_READER);
280 if (dn->dn_bonus == NULL) {
281 rw_exit(&dn->dn_struct_rwlock);
282 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
283 if (dn->dn_bonus == NULL)
284 dbuf_create_bonus(dn);
285 }
286 db = dn->dn_bonus;
287
288 /* as long as the bonus buf is held, the dnode will be held */
289 if (refcount_add(&db->db_holds, tag) == 1) {
290 VERIFY(dnode_add_ref(dn, db));
291 atomic_inc_32(&dn->dn_dbufs_count);
292 }
293
294 /*
295 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
296 * hold and incrementing the dbuf count to ensure that dnode_move() sees
297 * a dnode hold for every dbuf.
298 */
299 rw_exit(&dn->dn_struct_rwlock);
300
301 dnode_rele(dn, FTAG);
302
303 VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH));
304
305 *dbp = &db->db;
306 return (0);
307 }
308
309 /*
310 * returns ENOENT, EIO, or 0.
311 *
312 * This interface will allocate a blank spill dbuf when a spill blk
313 * doesn't already exist on the dnode.
314 *
315 * if you only want to find an already existing spill db, then
316 * dmu_spill_hold_existing() should be used.
317 */
318 int
dmu_spill_hold_by_dnode(dnode_t * dn,uint32_t flags,void * tag,dmu_buf_t ** dbp)319 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
320 {
321 dmu_buf_impl_t *db = NULL;
322 int err;
323
324 if ((flags & DB_RF_HAVESTRUCT) == 0)
325 rw_enter(&dn->dn_struct_rwlock, RW_READER);
326
327 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
328
329 if ((flags & DB_RF_HAVESTRUCT) == 0)
330 rw_exit(&dn->dn_struct_rwlock);
331
332 ASSERT(db != NULL);
333 err = dbuf_read(db, NULL, flags);
334 if (err == 0)
335 *dbp = &db->db;
336 else
337 dbuf_rele(db, tag);
338 return (err);
339 }
340
341 int
dmu_spill_hold_existing(dmu_buf_t * bonus,void * tag,dmu_buf_t ** dbp)342 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
343 {
344 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
345 dnode_t *dn;
346 int err;
347
348 DB_DNODE_ENTER(db);
349 dn = DB_DNODE(db);
350
351 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
352 err = SET_ERROR(EINVAL);
353 } else {
354 rw_enter(&dn->dn_struct_rwlock, RW_READER);
355
356 if (!dn->dn_have_spill) {
357 err = SET_ERROR(ENOENT);
358 } else {
359 err = dmu_spill_hold_by_dnode(dn,
360 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
361 }
362
363 rw_exit(&dn->dn_struct_rwlock);
364 }
365
366 DB_DNODE_EXIT(db);
367 return (err);
368 }
369
370 int
dmu_spill_hold_by_bonus(dmu_buf_t * bonus,void * tag,dmu_buf_t ** dbp)371 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
372 {
373 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
374 dnode_t *dn;
375 int err;
376
377 DB_DNODE_ENTER(db);
378 dn = DB_DNODE(db);
379 err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp);
380 DB_DNODE_EXIT(db);
381
382 return (err);
383 }
384
385 /*
386 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
387 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
388 * and can induce severe lock contention when writing to several files
389 * whose dnodes are in the same block.
390 */
391 static int
dmu_buf_hold_array_by_dnode(dnode_t * dn,uint64_t offset,uint64_t length,int read,void * tag,int * numbufsp,dmu_buf_t *** dbpp,uint32_t flags)392 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
393 int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
394 {
395 dmu_buf_t **dbp;
396 uint64_t blkid, nblks, i;
397 uint32_t dbuf_flags;
398 int err;
399 zio_t *zio;
400
401 ASSERT(length <= DMU_MAX_ACCESS);
402
403 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT;
404 if (flags & DMU_READ_NO_PREFETCH || length > zfetch_array_rd_sz)
405 dbuf_flags |= DB_RF_NOPREFETCH;
406
407 rw_enter(&dn->dn_struct_rwlock, RW_READER);
408 if (dn->dn_datablkshift) {
409 int blkshift = dn->dn_datablkshift;
410 nblks = (P2ROUNDUP(offset+length, 1ULL<<blkshift) -
411 P2ALIGN(offset, 1ULL<<blkshift)) >> blkshift;
412 } else {
413 if (offset + length > dn->dn_datablksz) {
414 zfs_panic_recover("zfs: accessing past end of object "
415 "%llx/%llx (size=%u access=%llu+%llu)",
416 (longlong_t)dn->dn_objset->
417 os_dsl_dataset->ds_object,
418 (longlong_t)dn->dn_object, dn->dn_datablksz,
419 (longlong_t)offset, (longlong_t)length);
420 rw_exit(&dn->dn_struct_rwlock);
421 return (SET_ERROR(EIO));
422 }
423 nblks = 1;
424 }
425 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
426
427 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
428 blkid = dbuf_whichblock(dn, offset);
429 for (i = 0; i < nblks; i++) {
430 dmu_buf_impl_t *db = dbuf_hold(dn, blkid+i, tag);
431 if (db == NULL) {
432 rw_exit(&dn->dn_struct_rwlock);
433 dmu_buf_rele_array(dbp, nblks, tag);
434 zio_nowait(zio);
435 return (SET_ERROR(EIO));
436 }
437 /* initiate async i/o */
438 if (read)
439 (void) dbuf_read(db, zio, dbuf_flags);
440 #ifdef _KERNEL
441 else
442 curthread->td_ru.ru_oublock++;
443 #endif
444 dbp[i] = &db->db;
445 }
446 rw_exit(&dn->dn_struct_rwlock);
447
448 /* wait for async i/o */
449 err = zio_wait(zio);
450 if (err) {
451 dmu_buf_rele_array(dbp, nblks, tag);
452 return (err);
453 }
454
455 /* wait for other io to complete */
456 if (read) {
457 for (i = 0; i < nblks; i++) {
458 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
459 mutex_enter(&db->db_mtx);
460 while (db->db_state == DB_READ ||
461 db->db_state == DB_FILL)
462 cv_wait(&db->db_changed, &db->db_mtx);
463 if (db->db_state == DB_UNCACHED)
464 err = SET_ERROR(EIO);
465 mutex_exit(&db->db_mtx);
466 if (err) {
467 dmu_buf_rele_array(dbp, nblks, tag);
468 return (err);
469 }
470 }
471 }
472
473 *numbufsp = nblks;
474 *dbpp = dbp;
475 return (0);
476 }
477
478 static int
dmu_buf_hold_array(objset_t * os,uint64_t object,uint64_t offset,uint64_t length,int read,void * tag,int * numbufsp,dmu_buf_t *** dbpp)479 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
480 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
481 {
482 dnode_t *dn;
483 int err;
484
485 err = dnode_hold(os, object, FTAG, &dn);
486 if (err)
487 return (err);
488
489 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
490 numbufsp, dbpp, DMU_READ_PREFETCH);
491
492 dnode_rele(dn, FTAG);
493
494 return (err);
495 }
496
497 int
dmu_buf_hold_array_by_bonus(dmu_buf_t * db_fake,uint64_t offset,uint64_t length,int read,void * tag,int * numbufsp,dmu_buf_t *** dbpp)498 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
499 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
500 {
501 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
502 dnode_t *dn;
503 int err;
504
505 DB_DNODE_ENTER(db);
506 dn = DB_DNODE(db);
507 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
508 numbufsp, dbpp, DMU_READ_PREFETCH);
509 DB_DNODE_EXIT(db);
510
511 return (err);
512 }
513
514 void
dmu_buf_rele_array(dmu_buf_t ** dbp_fake,int numbufs,void * tag)515 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
516 {
517 int i;
518 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
519
520 if (numbufs == 0)
521 return;
522
523 for (i = 0; i < numbufs; i++) {
524 if (dbp[i])
525 dbuf_rele(dbp[i], tag);
526 }
527
528 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
529 }
530
531 /*
532 * Issue prefetch i/os for the given blocks.
533 *
534 * Note: The assumption is that we *know* these blocks will be needed
535 * almost immediately. Therefore, the prefetch i/os will be issued at
536 * ZIO_PRIORITY_SYNC_READ
537 *
538 * Note: indirect blocks and other metadata will be read synchronously,
539 * causing this function to block if they are not already cached.
540 */
541 void
dmu_prefetch(objset_t * os,uint64_t object,uint64_t offset,uint64_t len)542 dmu_prefetch(objset_t *os, uint64_t object, uint64_t offset, uint64_t len)
543 {
544 dnode_t *dn;
545 uint64_t blkid;
546 int nblks, err;
547
548 if (zfs_prefetch_disable)
549 return;
550
551 if (len == 0) { /* they're interested in the bonus buffer */
552 dn = DMU_META_DNODE(os);
553
554 if (object == 0 || object >= DN_MAX_OBJECT)
555 return;
556
557 rw_enter(&dn->dn_struct_rwlock, RW_READER);
558 blkid = dbuf_whichblock(dn, object * sizeof (dnode_phys_t));
559 dbuf_prefetch(dn, blkid, ZIO_PRIORITY_SYNC_READ);
560 rw_exit(&dn->dn_struct_rwlock);
561 return;
562 }
563
564 /*
565 * XXX - Note, if the dnode for the requested object is not
566 * already cached, we will do a *synchronous* read in the
567 * dnode_hold() call. The same is true for any indirects.
568 */
569 err = dnode_hold(os, object, FTAG, &dn);
570 if (err != 0)
571 return;
572
573 rw_enter(&dn->dn_struct_rwlock, RW_READER);
574 if (dn->dn_datablkshift) {
575 int blkshift = dn->dn_datablkshift;
576 nblks = (P2ROUNDUP(offset + len, 1 << blkshift) -
577 P2ALIGN(offset, 1 << blkshift)) >> blkshift;
578 } else {
579 nblks = (offset < dn->dn_datablksz);
580 }
581
582 if (nblks != 0) {
583 blkid = dbuf_whichblock(dn, offset);
584 for (int i = 0; i < nblks; i++)
585 dbuf_prefetch(dn, blkid + i, ZIO_PRIORITY_SYNC_READ);
586 }
587
588 rw_exit(&dn->dn_struct_rwlock);
589
590 dnode_rele(dn, FTAG);
591 }
592
593 /*
594 * Get the next "chunk" of file data to free. We traverse the file from
595 * the end so that the file gets shorter over time (if we crashes in the
596 * middle, this will leave us in a better state). We find allocated file
597 * data by simply searching the allocated level 1 indirects.
598 *
599 * On input, *start should be the first offset that does not need to be
600 * freed (e.g. "offset + length"). On return, *start will be the first
601 * offset that should be freed.
602 */
603 static int
get_next_chunk(dnode_t * dn,uint64_t * start,uint64_t minimum)604 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum)
605 {
606 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
607 /* bytes of data covered by a level-1 indirect block */
608 uint64_t iblkrange =
609 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
610
611 ASSERT3U(minimum, <=, *start);
612
613 if (*start - minimum <= iblkrange * maxblks) {
614 *start = minimum;
615 return (0);
616 }
617 ASSERT(ISP2(iblkrange));
618
619 for (uint64_t blks = 0; *start > minimum && blks < maxblks; blks++) {
620 int err;
621
622 /*
623 * dnode_next_offset(BACKWARDS) will find an allocated L1
624 * indirect block at or before the input offset. We must
625 * decrement *start so that it is at the end of the region
626 * to search.
627 */
628 (*start)--;
629 err = dnode_next_offset(dn,
630 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
631
632 /* if there are no indirect blocks before start, we are done */
633 if (err == ESRCH) {
634 *start = minimum;
635 break;
636 } else if (err != 0) {
637 return (err);
638 }
639
640 /* set start to the beginning of this L1 indirect */
641 *start = P2ALIGN(*start, iblkrange);
642 }
643 if (*start < minimum)
644 *start = minimum;
645 return (0);
646 }
647
648 static int
dmu_free_long_range_impl(objset_t * os,dnode_t * dn,uint64_t offset,uint64_t length)649 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
650 uint64_t length)
651 {
652 uint64_t object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
653 int err;
654
655 if (offset >= object_size)
656 return (0);
657
658 if (length == DMU_OBJECT_END || offset + length > object_size)
659 length = object_size - offset;
660
661 while (length != 0) {
662 uint64_t chunk_end, chunk_begin;
663
664 chunk_end = chunk_begin = offset + length;
665
666 /* move chunk_begin backwards to the beginning of this chunk */
667 err = get_next_chunk(dn, &chunk_begin, offset);
668 if (err)
669 return (err);
670 ASSERT3U(chunk_begin, >=, offset);
671 ASSERT3U(chunk_begin, <=, chunk_end);
672
673 dmu_tx_t *tx = dmu_tx_create(os);
674 dmu_tx_hold_free(tx, dn->dn_object,
675 chunk_begin, chunk_end - chunk_begin);
676
677 /*
678 * Mark this transaction as typically resulting in a net
679 * reduction in space used.
680 */
681 dmu_tx_mark_netfree(tx);
682 err = dmu_tx_assign(tx, TXG_WAIT);
683 if (err) {
684 dmu_tx_abort(tx);
685 return (err);
686 }
687 dnode_free_range(dn, chunk_begin, chunk_end - chunk_begin, tx);
688 dmu_tx_commit(tx);
689
690 length -= chunk_end - chunk_begin;
691 }
692 return (0);
693 }
694
695 int
dmu_free_long_range(objset_t * os,uint64_t object,uint64_t offset,uint64_t length)696 dmu_free_long_range(objset_t *os, uint64_t object,
697 uint64_t offset, uint64_t length)
698 {
699 dnode_t *dn;
700 int err;
701
702 err = dnode_hold(os, object, FTAG, &dn);
703 if (err != 0)
704 return (err);
705 err = dmu_free_long_range_impl(os, dn, offset, length);
706
707 /*
708 * It is important to zero out the maxblkid when freeing the entire
709 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
710 * will take the fast path, and (b) dnode_reallocate() can verify
711 * that the entire file has been freed.
712 */
713 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
714 dn->dn_maxblkid = 0;
715
716 dnode_rele(dn, FTAG);
717 return (err);
718 }
719
720 int
dmu_free_long_object(objset_t * os,uint64_t object)721 dmu_free_long_object(objset_t *os, uint64_t object)
722 {
723 dmu_tx_t *tx;
724 int err;
725
726 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
727 if (err != 0)
728 return (err);
729
730 tx = dmu_tx_create(os);
731 dmu_tx_hold_bonus(tx, object);
732 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
733 dmu_tx_mark_netfree(tx);
734 err = dmu_tx_assign(tx, TXG_WAIT);
735 if (err == 0) {
736 err = dmu_object_free(os, object, tx);
737 dmu_tx_commit(tx);
738 } else {
739 dmu_tx_abort(tx);
740 }
741
742 return (err);
743 }
744
745 int
dmu_free_range(objset_t * os,uint64_t object,uint64_t offset,uint64_t size,dmu_tx_t * tx)746 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
747 uint64_t size, dmu_tx_t *tx)
748 {
749 dnode_t *dn;
750 int err = dnode_hold(os, object, FTAG, &dn);
751 if (err)
752 return (err);
753 ASSERT(offset < UINT64_MAX);
754 ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
755 dnode_free_range(dn, offset, size, tx);
756 dnode_rele(dn, FTAG);
757 return (0);
758 }
759
760 int
dmu_read(objset_t * os,uint64_t object,uint64_t offset,uint64_t size,void * buf,uint32_t flags)761 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
762 void *buf, uint32_t flags)
763 {
764 dnode_t *dn;
765 dmu_buf_t **dbp;
766 int numbufs, err;
767
768 err = dnode_hold(os, object, FTAG, &dn);
769 if (err)
770 return (err);
771
772 /*
773 * Deal with odd block sizes, where there can't be data past the first
774 * block. If we ever do the tail block optimization, we will need to
775 * handle that here as well.
776 */
777 if (dn->dn_maxblkid == 0) {
778 int newsz = offset > dn->dn_datablksz ? 0 :
779 MIN(size, dn->dn_datablksz - offset);
780 bzero((char *)buf + newsz, size - newsz);
781 size = newsz;
782 }
783
784 while (size > 0) {
785 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
786 int i;
787
788 /*
789 * NB: we could do this block-at-a-time, but it's nice
790 * to be reading in parallel.
791 */
792 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
793 TRUE, FTAG, &numbufs, &dbp, flags);
794 if (err)
795 break;
796
797 for (i = 0; i < numbufs; i++) {
798 int tocpy;
799 int bufoff;
800 dmu_buf_t *db = dbp[i];
801
802 ASSERT(size > 0);
803
804 bufoff = offset - db->db_offset;
805 tocpy = (int)MIN(db->db_size - bufoff, size);
806
807 bcopy((char *)db->db_data + bufoff, buf, tocpy);
808
809 offset += tocpy;
810 size -= tocpy;
811 buf = (char *)buf + tocpy;
812 }
813 dmu_buf_rele_array(dbp, numbufs, FTAG);
814 }
815 dnode_rele(dn, FTAG);
816 return (err);
817 }
818
819 void
dmu_write(objset_t * os,uint64_t object,uint64_t offset,uint64_t size,const void * buf,dmu_tx_t * tx)820 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
821 const void *buf, dmu_tx_t *tx)
822 {
823 dmu_buf_t **dbp;
824 int numbufs, i;
825
826 if (size == 0)
827 return;
828
829 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
830 FALSE, FTAG, &numbufs, &dbp));
831
832 for (i = 0; i < numbufs; i++) {
833 int tocpy;
834 int bufoff;
835 dmu_buf_t *db = dbp[i];
836
837 ASSERT(size > 0);
838
839 bufoff = offset - db->db_offset;
840 tocpy = (int)MIN(db->db_size - bufoff, size);
841
842 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
843
844 if (tocpy == db->db_size)
845 dmu_buf_will_fill(db, tx);
846 else
847 dmu_buf_will_dirty(db, tx);
848
849 bcopy(buf, (char *)db->db_data + bufoff, tocpy);
850
851 if (tocpy == db->db_size)
852 dmu_buf_fill_done(db, tx);
853
854 offset += tocpy;
855 size -= tocpy;
856 buf = (char *)buf + tocpy;
857 }
858 dmu_buf_rele_array(dbp, numbufs, FTAG);
859 }
860
861 void
dmu_prealloc(objset_t * os,uint64_t object,uint64_t offset,uint64_t size,dmu_tx_t * tx)862 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
863 dmu_tx_t *tx)
864 {
865 dmu_buf_t **dbp;
866 int numbufs, i;
867
868 if (size == 0)
869 return;
870
871 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
872 FALSE, FTAG, &numbufs, &dbp));
873
874 for (i = 0; i < numbufs; i++) {
875 dmu_buf_t *db = dbp[i];
876
877 dmu_buf_will_not_fill(db, tx);
878 }
879 dmu_buf_rele_array(dbp, numbufs, FTAG);
880 }
881
882 void
dmu_write_embedded(objset_t * os,uint64_t object,uint64_t offset,void * data,uint8_t etype,uint8_t comp,int uncompressed_size,int compressed_size,int byteorder,dmu_tx_t * tx)883 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
884 void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
885 int compressed_size, int byteorder, dmu_tx_t *tx)
886 {
887 dmu_buf_t *db;
888
889 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
890 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
891 VERIFY0(dmu_buf_hold_noread(os, object, offset,
892 FTAG, &db));
893
894 dmu_buf_write_embedded(db,
895 data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
896 uncompressed_size, compressed_size, byteorder, tx);
897
898 dmu_buf_rele(db, FTAG);
899 }
900
901 /*
902 * DMU support for xuio
903 */
904 kstat_t *xuio_ksp = NULL;
905
906 int
dmu_xuio_init(xuio_t * xuio,int nblk)907 dmu_xuio_init(xuio_t *xuio, int nblk)
908 {
909 dmu_xuio_t *priv;
910 uio_t *uio = &xuio->xu_uio;
911
912 uio->uio_iovcnt = nblk;
913 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
914
915 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
916 priv->cnt = nblk;
917 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
918 priv->iovp = uio->uio_iov;
919 XUIO_XUZC_PRIV(xuio) = priv;
920
921 if (XUIO_XUZC_RW(xuio) == UIO_READ)
922 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
923 else
924 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
925
926 return (0);
927 }
928
929 void
dmu_xuio_fini(xuio_t * xuio)930 dmu_xuio_fini(xuio_t *xuio)
931 {
932 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
933 int nblk = priv->cnt;
934
935 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
936 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
937 kmem_free(priv, sizeof (dmu_xuio_t));
938
939 if (XUIO_XUZC_RW(xuio) == UIO_READ)
940 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
941 else
942 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
943 }
944
945 /*
946 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
947 * and increase priv->next by 1.
948 */
949 int
dmu_xuio_add(xuio_t * xuio,arc_buf_t * abuf,offset_t off,size_t n)950 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
951 {
952 struct iovec *iov;
953 uio_t *uio = &xuio->xu_uio;
954 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
955 int i = priv->next++;
956
957 ASSERT(i < priv->cnt);
958 ASSERT(off + n <= arc_buf_size(abuf));
959 iov = uio->uio_iov + i;
960 iov->iov_base = (char *)abuf->b_data + off;
961 iov->iov_len = n;
962 priv->bufs[i] = abuf;
963 return (0);
964 }
965
966 int
dmu_xuio_cnt(xuio_t * xuio)967 dmu_xuio_cnt(xuio_t *xuio)
968 {
969 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
970 return (priv->cnt);
971 }
972
973 arc_buf_t *
dmu_xuio_arcbuf(xuio_t * xuio,int i)974 dmu_xuio_arcbuf(xuio_t *xuio, int i)
975 {
976 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
977
978 ASSERT(i < priv->cnt);
979 return (priv->bufs[i]);
980 }
981
982 void
dmu_xuio_clear(xuio_t * xuio,int i)983 dmu_xuio_clear(xuio_t *xuio, int i)
984 {
985 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
986
987 ASSERT(i < priv->cnt);
988 priv->bufs[i] = NULL;
989 }
990
991 static void
xuio_stat_init(void)992 xuio_stat_init(void)
993 {
994 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
995 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
996 KSTAT_FLAG_VIRTUAL);
997 if (xuio_ksp != NULL) {
998 xuio_ksp->ks_data = &xuio_stats;
999 kstat_install(xuio_ksp);
1000 }
1001 }
1002
1003 static void
xuio_stat_fini(void)1004 xuio_stat_fini(void)
1005 {
1006 if (xuio_ksp != NULL) {
1007 kstat_delete(xuio_ksp);
1008 xuio_ksp = NULL;
1009 }
1010 }
1011
1012 void
xuio_stat_wbuf_copied()1013 xuio_stat_wbuf_copied()
1014 {
1015 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1016 }
1017
1018 void
xuio_stat_wbuf_nocopy()1019 xuio_stat_wbuf_nocopy()
1020 {
1021 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
1022 }
1023
1024 #ifdef _KERNEL
1025 static int
dmu_read_uio_dnode(dnode_t * dn,uio_t * uio,uint64_t size)1026 dmu_read_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size)
1027 {
1028 dmu_buf_t **dbp;
1029 int numbufs, i, err;
1030 xuio_t *xuio = NULL;
1031
1032 /*
1033 * NB: we could do this block-at-a-time, but it's nice
1034 * to be reading in parallel.
1035 */
1036 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1037 TRUE, FTAG, &numbufs, &dbp, 0);
1038 if (err)
1039 return (err);
1040
1041 #ifdef UIO_XUIO
1042 if (uio->uio_extflg == UIO_XUIO)
1043 xuio = (xuio_t *)uio;
1044 #endif
1045
1046 for (i = 0; i < numbufs; i++) {
1047 int tocpy;
1048 int bufoff;
1049 dmu_buf_t *db = dbp[i];
1050
1051 ASSERT(size > 0);
1052
1053 bufoff = uio->uio_loffset - db->db_offset;
1054 tocpy = (int)MIN(db->db_size - bufoff, size);
1055
1056 if (xuio) {
1057 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
1058 arc_buf_t *dbuf_abuf = dbi->db_buf;
1059 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
1060 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
1061 if (!err) {
1062 uio->uio_resid -= tocpy;
1063 uio->uio_loffset += tocpy;
1064 }
1065
1066 if (abuf == dbuf_abuf)
1067 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
1068 else
1069 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
1070 } else {
1071 err = uiomove((char *)db->db_data + bufoff, tocpy,
1072 UIO_READ, uio);
1073 }
1074 if (err)
1075 break;
1076
1077 size -= tocpy;
1078 }
1079 dmu_buf_rele_array(dbp, numbufs, FTAG);
1080
1081 return (err);
1082 }
1083
1084 /*
1085 * Read 'size' bytes into the uio buffer.
1086 * From object zdb->db_object.
1087 * Starting at offset uio->uio_loffset.
1088 *
1089 * If the caller already has a dbuf in the target object
1090 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1091 * because we don't have to find the dnode_t for the object.
1092 */
1093 int
dmu_read_uio_dbuf(dmu_buf_t * zdb,uio_t * uio,uint64_t size)1094 dmu_read_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size)
1095 {
1096 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1097 dnode_t *dn;
1098 int err;
1099
1100 if (size == 0)
1101 return (0);
1102
1103 DB_DNODE_ENTER(db);
1104 dn = DB_DNODE(db);
1105 err = dmu_read_uio_dnode(dn, uio, size);
1106 DB_DNODE_EXIT(db);
1107
1108 return (err);
1109 }
1110
1111 /*
1112 * Read 'size' bytes into the uio buffer.
1113 * From the specified object
1114 * Starting at offset uio->uio_loffset.
1115 */
1116 int
dmu_read_uio(objset_t * os,uint64_t object,uio_t * uio,uint64_t size)1117 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
1118 {
1119 dnode_t *dn;
1120 int err;
1121
1122 if (size == 0)
1123 return (0);
1124
1125 err = dnode_hold(os, object, FTAG, &dn);
1126 if (err)
1127 return (err);
1128
1129 err = dmu_read_uio_dnode(dn, uio, size);
1130
1131 dnode_rele(dn, FTAG);
1132
1133 return (err);
1134 }
1135
1136 static int
dmu_write_uio_dnode(dnode_t * dn,uio_t * uio,uint64_t size,dmu_tx_t * tx)1137 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1138 {
1139 dmu_buf_t **dbp;
1140 int numbufs;
1141 int err = 0;
1142 int i;
1143
1144 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1145 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1146 if (err)
1147 return (err);
1148
1149 for (i = 0; i < numbufs; i++) {
1150 int tocpy;
1151 int bufoff;
1152 dmu_buf_t *db = dbp[i];
1153
1154 ASSERT(size > 0);
1155
1156 bufoff = uio->uio_loffset - db->db_offset;
1157 tocpy = (int)MIN(db->db_size - bufoff, size);
1158
1159 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1160
1161 if (tocpy == db->db_size)
1162 dmu_buf_will_fill(db, tx);
1163 else
1164 dmu_buf_will_dirty(db, tx);
1165
1166 /*
1167 * XXX uiomove could block forever (eg. nfs-backed
1168 * pages). There needs to be a uiolockdown() function
1169 * to lock the pages in memory, so that uiomove won't
1170 * block.
1171 */
1172 err = uiomove((char *)db->db_data + bufoff, tocpy,
1173 UIO_WRITE, uio);
1174
1175 if (tocpy == db->db_size)
1176 dmu_buf_fill_done(db, tx);
1177
1178 if (err)
1179 break;
1180
1181 size -= tocpy;
1182 }
1183
1184 dmu_buf_rele_array(dbp, numbufs, FTAG);
1185 return (err);
1186 }
1187
1188 /*
1189 * Write 'size' bytes from the uio buffer.
1190 * To object zdb->db_object.
1191 * Starting at offset uio->uio_loffset.
1192 *
1193 * If the caller already has a dbuf in the target object
1194 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1195 * because we don't have to find the dnode_t for the object.
1196 */
1197 int
dmu_write_uio_dbuf(dmu_buf_t * zdb,uio_t * uio,uint64_t size,dmu_tx_t * tx)1198 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1199 dmu_tx_t *tx)
1200 {
1201 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1202 dnode_t *dn;
1203 int err;
1204
1205 if (size == 0)
1206 return (0);
1207
1208 DB_DNODE_ENTER(db);
1209 dn = DB_DNODE(db);
1210 err = dmu_write_uio_dnode(dn, uio, size, tx);
1211 DB_DNODE_EXIT(db);
1212
1213 return (err);
1214 }
1215
1216 /*
1217 * Write 'size' bytes from the uio buffer.
1218 * To the specified object.
1219 * Starting at offset uio->uio_loffset.
1220 */
1221 int
dmu_write_uio(objset_t * os,uint64_t object,uio_t * uio,uint64_t size,dmu_tx_t * tx)1222 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1223 dmu_tx_t *tx)
1224 {
1225 dnode_t *dn;
1226 int err;
1227
1228 if (size == 0)
1229 return (0);
1230
1231 err = dnode_hold(os, object, FTAG, &dn);
1232 if (err)
1233 return (err);
1234
1235 err = dmu_write_uio_dnode(dn, uio, size, tx);
1236
1237 dnode_rele(dn, FTAG);
1238
1239 return (err);
1240 }
1241
1242 #ifdef sun
1243 int
dmu_write_pages(objset_t * os,uint64_t object,uint64_t offset,uint64_t size,page_t * pp,dmu_tx_t * tx)1244 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1245 page_t *pp, dmu_tx_t *tx)
1246 {
1247 dmu_buf_t **dbp;
1248 int numbufs, i;
1249 int err;
1250
1251 if (size == 0)
1252 return (0);
1253
1254 err = dmu_buf_hold_array(os, object, offset, size,
1255 FALSE, FTAG, &numbufs, &dbp);
1256 if (err)
1257 return (err);
1258
1259 for (i = 0; i < numbufs; i++) {
1260 int tocpy, copied, thiscpy;
1261 int bufoff;
1262 dmu_buf_t *db = dbp[i];
1263 caddr_t va;
1264
1265 ASSERT(size > 0);
1266 ASSERT3U(db->db_size, >=, PAGESIZE);
1267
1268 bufoff = offset - db->db_offset;
1269 tocpy = (int)MIN(db->db_size - bufoff, size);
1270
1271 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1272
1273 if (tocpy == db->db_size)
1274 dmu_buf_will_fill(db, tx);
1275 else
1276 dmu_buf_will_dirty(db, tx);
1277
1278 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1279 ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff);
1280 thiscpy = MIN(PAGESIZE, tocpy - copied);
1281 va = zfs_map_page(pp, S_READ);
1282 bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1283 zfs_unmap_page(pp, va);
1284 pp = pp->p_next;
1285 bufoff += PAGESIZE;
1286 }
1287
1288 if (tocpy == db->db_size)
1289 dmu_buf_fill_done(db, tx);
1290
1291 offset += tocpy;
1292 size -= tocpy;
1293 }
1294 dmu_buf_rele_array(dbp, numbufs, FTAG);
1295 return (err);
1296 }
1297
1298 #else
1299
1300 int
dmu_write_pages(objset_t * os,uint64_t object,uint64_t offset,uint64_t size,vm_page_t * ma,dmu_tx_t * tx)1301 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1302 vm_page_t *ma, dmu_tx_t *tx)
1303 {
1304 dmu_buf_t **dbp;
1305 struct sf_buf *sf;
1306 int numbufs, i;
1307 int err;
1308
1309 if (size == 0)
1310 return (0);
1311
1312 err = dmu_buf_hold_array(os, object, offset, size,
1313 FALSE, FTAG, &numbufs, &dbp);
1314 if (err)
1315 return (err);
1316
1317 for (i = 0; i < numbufs; i++) {
1318 int tocpy, copied, thiscpy;
1319 int bufoff;
1320 dmu_buf_t *db = dbp[i];
1321 caddr_t va;
1322
1323 ASSERT(size > 0);
1324 ASSERT3U(db->db_size, >=, PAGESIZE);
1325
1326 bufoff = offset - db->db_offset;
1327 tocpy = (int)MIN(db->db_size - bufoff, size);
1328
1329 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1330
1331 if (tocpy == db->db_size)
1332 dmu_buf_will_fill(db, tx);
1333 else
1334 dmu_buf_will_dirty(db, tx);
1335
1336 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1337 ASSERT3U(ptoa((*ma)->pindex), ==, db->db_offset + bufoff);
1338 thiscpy = MIN(PAGESIZE, tocpy - copied);
1339 va = zfs_map_page(*ma, &sf);
1340 bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1341 zfs_unmap_page(sf);
1342 ma += 1;
1343 bufoff += PAGESIZE;
1344 }
1345
1346 if (tocpy == db->db_size)
1347 dmu_buf_fill_done(db, tx);
1348
1349 offset += tocpy;
1350 size -= tocpy;
1351 }
1352 dmu_buf_rele_array(dbp, numbufs, FTAG);
1353 return (err);
1354 }
1355 #endif /* sun */
1356 #endif
1357
1358 /*
1359 * Allocate a loaned anonymous arc buffer.
1360 */
1361 arc_buf_t *
dmu_request_arcbuf(dmu_buf_t * handle,int size)1362 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1363 {
1364 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1365
1366 return (arc_loan_buf(db->db_objset->os_spa, size));
1367 }
1368
1369 /*
1370 * Free a loaned arc buffer.
1371 */
1372 void
dmu_return_arcbuf(arc_buf_t * buf)1373 dmu_return_arcbuf(arc_buf_t *buf)
1374 {
1375 arc_return_buf(buf, FTAG);
1376 VERIFY(arc_buf_remove_ref(buf, FTAG));
1377 }
1378
1379 /*
1380 * When possible directly assign passed loaned arc buffer to a dbuf.
1381 * If this is not possible copy the contents of passed arc buf via
1382 * dmu_write().
1383 */
1384 void
dmu_assign_arcbuf(dmu_buf_t * handle,uint64_t offset,arc_buf_t * buf,dmu_tx_t * tx)1385 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1386 dmu_tx_t *tx)
1387 {
1388 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1389 dnode_t *dn;
1390 dmu_buf_impl_t *db;
1391 uint32_t blksz = (uint32_t)arc_buf_size(buf);
1392 uint64_t blkid;
1393
1394 DB_DNODE_ENTER(dbuf);
1395 dn = DB_DNODE(dbuf);
1396 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1397 blkid = dbuf_whichblock(dn, offset);
1398 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1399 rw_exit(&dn->dn_struct_rwlock);
1400 DB_DNODE_EXIT(dbuf);
1401
1402 /*
1403 * We can only assign if the offset is aligned, the arc buf is the
1404 * same size as the dbuf, and the dbuf is not metadata. It
1405 * can't be metadata because the loaned arc buf comes from the
1406 * user-data kmem arena.
1407 */
1408 if (offset == db->db.db_offset && blksz == db->db.db_size &&
1409 DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA) {
1410 dbuf_assign_arcbuf(db, buf, tx);
1411 dbuf_rele(db, FTAG);
1412 } else {
1413 objset_t *os;
1414 uint64_t object;
1415
1416 DB_DNODE_ENTER(dbuf);
1417 dn = DB_DNODE(dbuf);
1418 os = dn->dn_objset;
1419 object = dn->dn_object;
1420 DB_DNODE_EXIT(dbuf);
1421
1422 dbuf_rele(db, FTAG);
1423 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1424 dmu_return_arcbuf(buf);
1425 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1426 }
1427 }
1428
1429 typedef struct {
1430 dbuf_dirty_record_t *dsa_dr;
1431 dmu_sync_cb_t *dsa_done;
1432 zgd_t *dsa_zgd;
1433 dmu_tx_t *dsa_tx;
1434 } dmu_sync_arg_t;
1435
1436 /* ARGSUSED */
1437 static void
dmu_sync_ready(zio_t * zio,arc_buf_t * buf,void * varg)1438 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1439 {
1440 dmu_sync_arg_t *dsa = varg;
1441 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1442 blkptr_t *bp = zio->io_bp;
1443
1444 if (zio->io_error == 0) {
1445 if (BP_IS_HOLE(bp)) {
1446 /*
1447 * A block of zeros may compress to a hole, but the
1448 * block size still needs to be known for replay.
1449 */
1450 BP_SET_LSIZE(bp, db->db_size);
1451 } else if (!BP_IS_EMBEDDED(bp)) {
1452 ASSERT(BP_GET_LEVEL(bp) == 0);
1453 bp->blk_fill = 1;
1454 }
1455 }
1456 }
1457
1458 static void
dmu_sync_late_arrival_ready(zio_t * zio)1459 dmu_sync_late_arrival_ready(zio_t *zio)
1460 {
1461 dmu_sync_ready(zio, NULL, zio->io_private);
1462 }
1463
1464 /* ARGSUSED */
1465 static void
dmu_sync_done(zio_t * zio,arc_buf_t * buf,void * varg)1466 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1467 {
1468 dmu_sync_arg_t *dsa = varg;
1469 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1470 dmu_buf_impl_t *db = dr->dr_dbuf;
1471
1472 mutex_enter(&db->db_mtx);
1473 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1474 if (zio->io_error == 0) {
1475 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1476 if (dr->dt.dl.dr_nopwrite) {
1477 blkptr_t *bp = zio->io_bp;
1478 blkptr_t *bp_orig = &zio->io_bp_orig;
1479 uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1480
1481 ASSERT(BP_EQUAL(bp, bp_orig));
1482 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1483 ASSERT(zio_checksum_table[chksum].ci_dedup);
1484 }
1485 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1486 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1487 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1488 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by))
1489 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1490 } else {
1491 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1492 }
1493 cv_broadcast(&db->db_changed);
1494 mutex_exit(&db->db_mtx);
1495
1496 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1497
1498 kmem_free(dsa, sizeof (*dsa));
1499 }
1500
1501 static void
dmu_sync_late_arrival_done(zio_t * zio)1502 dmu_sync_late_arrival_done(zio_t *zio)
1503 {
1504 blkptr_t *bp = zio->io_bp;
1505 dmu_sync_arg_t *dsa = zio->io_private;
1506 blkptr_t *bp_orig = &zio->io_bp_orig;
1507
1508 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1509 /*
1510 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE)
1511 * then there is nothing to do here. Otherwise, free the
1512 * newly allocated block in this txg.
1513 */
1514 if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
1515 ASSERT(BP_EQUAL(bp, bp_orig));
1516 } else {
1517 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1518 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1519 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1520 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1521 }
1522 }
1523
1524 dmu_tx_commit(dsa->dsa_tx);
1525
1526 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1527
1528 kmem_free(dsa, sizeof (*dsa));
1529 }
1530
1531 static int
dmu_sync_late_arrival(zio_t * pio,objset_t * os,dmu_sync_cb_t * done,zgd_t * zgd,zio_prop_t * zp,zbookmark_phys_t * zb)1532 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1533 zio_prop_t *zp, zbookmark_phys_t *zb)
1534 {
1535 dmu_sync_arg_t *dsa;
1536 dmu_tx_t *tx;
1537
1538 tx = dmu_tx_create(os);
1539 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1540 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1541 dmu_tx_abort(tx);
1542 /* Make zl_get_data do txg_waited_synced() */
1543 return (SET_ERROR(EIO));
1544 }
1545
1546 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1547 dsa->dsa_dr = NULL;
1548 dsa->dsa_done = done;
1549 dsa->dsa_zgd = zgd;
1550 dsa->dsa_tx = tx;
1551
1552 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1553 zgd->zgd_db->db_data, zgd->zgd_db->db_size, zp,
1554 dmu_sync_late_arrival_ready, NULL, dmu_sync_late_arrival_done, dsa,
1555 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1556
1557 return (0);
1558 }
1559
1560 /*
1561 * Intent log support: sync the block associated with db to disk.
1562 * N.B. and XXX: the caller is responsible for making sure that the
1563 * data isn't changing while dmu_sync() is writing it.
1564 *
1565 * Return values:
1566 *
1567 * EEXIST: this txg has already been synced, so there's nothing to do.
1568 * The caller should not log the write.
1569 *
1570 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1571 * The caller should not log the write.
1572 *
1573 * EALREADY: this block is already in the process of being synced.
1574 * The caller should track its progress (somehow).
1575 *
1576 * EIO: could not do the I/O.
1577 * The caller should do a txg_wait_synced().
1578 *
1579 * 0: the I/O has been initiated.
1580 * The caller should log this blkptr in the done callback.
1581 * It is possible that the I/O will fail, in which case
1582 * the error will be reported to the done callback and
1583 * propagated to pio from zio_done().
1584 */
1585 int
dmu_sync(zio_t * pio,uint64_t txg,dmu_sync_cb_t * done,zgd_t * zgd)1586 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1587 {
1588 blkptr_t *bp = zgd->zgd_bp;
1589 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1590 objset_t *os = db->db_objset;
1591 dsl_dataset_t *ds = os->os_dsl_dataset;
1592 dbuf_dirty_record_t *dr;
1593 dmu_sync_arg_t *dsa;
1594 zbookmark_phys_t zb;
1595 zio_prop_t zp;
1596 dnode_t *dn;
1597
1598 ASSERT(pio != NULL);
1599 ASSERT(txg != 0);
1600
1601 SET_BOOKMARK(&zb, ds->ds_object,
1602 db->db.db_object, db->db_level, db->db_blkid);
1603
1604 DB_DNODE_ENTER(db);
1605 dn = DB_DNODE(db);
1606 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1607 DB_DNODE_EXIT(db);
1608
1609 /*
1610 * If we're frozen (running ziltest), we always need to generate a bp.
1611 */
1612 if (txg > spa_freeze_txg(os->os_spa))
1613 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1614
1615 /*
1616 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1617 * and us. If we determine that this txg is not yet syncing,
1618 * but it begins to sync a moment later, that's OK because the
1619 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1620 */
1621 mutex_enter(&db->db_mtx);
1622
1623 if (txg <= spa_last_synced_txg(os->os_spa)) {
1624 /*
1625 * This txg has already synced. There's nothing to do.
1626 */
1627 mutex_exit(&db->db_mtx);
1628 return (SET_ERROR(EEXIST));
1629 }
1630
1631 if (txg <= spa_syncing_txg(os->os_spa)) {
1632 /*
1633 * This txg is currently syncing, so we can't mess with
1634 * the dirty record anymore; just write a new log block.
1635 */
1636 mutex_exit(&db->db_mtx);
1637 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1638 }
1639
1640 dr = db->db_last_dirty;
1641 while (dr && dr->dr_txg != txg)
1642 dr = dr->dr_next;
1643
1644 if (dr == NULL) {
1645 /*
1646 * There's no dr for this dbuf, so it must have been freed.
1647 * There's no need to log writes to freed blocks, so we're done.
1648 */
1649 mutex_exit(&db->db_mtx);
1650 return (SET_ERROR(ENOENT));
1651 }
1652
1653 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
1654
1655 /*
1656 * Assume the on-disk data is X, the current syncing data is Y,
1657 * and the current in-memory data is Z (currently in dmu_sync).
1658 * X and Z are identical but Y is has been modified. Normally,
1659 * when X and Z are the same we will perform a nopwrite but if Y
1660 * is different we must disable nopwrite since the resulting write
1661 * of Y to disk can free the block containing X. If we allowed a
1662 * nopwrite to occur the block pointing to Z would reference a freed
1663 * block. Since this is a rare case we simplify this by disabling
1664 * nopwrite if the current dmu_sync-ing dbuf has been modified in
1665 * a previous transaction.
1666 */
1667 if (dr->dr_next)
1668 zp.zp_nopwrite = B_FALSE;
1669
1670 ASSERT(dr->dr_txg == txg);
1671 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1672 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1673 /*
1674 * We have already issued a sync write for this buffer,
1675 * or this buffer has already been synced. It could not
1676 * have been dirtied since, or we would have cleared the state.
1677 */
1678 mutex_exit(&db->db_mtx);
1679 return (SET_ERROR(EALREADY));
1680 }
1681
1682 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1683 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1684 mutex_exit(&db->db_mtx);
1685
1686 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1687 dsa->dsa_dr = dr;
1688 dsa->dsa_done = done;
1689 dsa->dsa_zgd = zgd;
1690 dsa->dsa_tx = NULL;
1691
1692 zio_nowait(arc_write(pio, os->os_spa, txg,
1693 bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
1694 DBUF_IS_L2COMPRESSIBLE(db), &zp, dmu_sync_ready,
1695 NULL, dmu_sync_done, dsa, ZIO_PRIORITY_SYNC_WRITE,
1696 ZIO_FLAG_CANFAIL, &zb));
1697
1698 return (0);
1699 }
1700
1701 int
dmu_object_set_blocksize(objset_t * os,uint64_t object,uint64_t size,int ibs,dmu_tx_t * tx)1702 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1703 dmu_tx_t *tx)
1704 {
1705 dnode_t *dn;
1706 int err;
1707
1708 err = dnode_hold(os, object, FTAG, &dn);
1709 if (err)
1710 return (err);
1711 err = dnode_set_blksz(dn, size, ibs, tx);
1712 dnode_rele(dn, FTAG);
1713 return (err);
1714 }
1715
1716 void
dmu_object_set_checksum(objset_t * os,uint64_t object,uint8_t checksum,dmu_tx_t * tx)1717 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1718 dmu_tx_t *tx)
1719 {
1720 dnode_t *dn;
1721
1722 /*
1723 * Send streams include each object's checksum function. This
1724 * check ensures that the receiving system can understand the
1725 * checksum function transmitted.
1726 */
1727 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS);
1728
1729 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1730 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS);
1731 dn->dn_checksum = checksum;
1732 dnode_setdirty(dn, tx);
1733 dnode_rele(dn, FTAG);
1734 }
1735
1736 void
dmu_object_set_compress(objset_t * os,uint64_t object,uint8_t compress,dmu_tx_t * tx)1737 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1738 dmu_tx_t *tx)
1739 {
1740 dnode_t *dn;
1741
1742 /*
1743 * Send streams include each object's compression function. This
1744 * check ensures that the receiving system can understand the
1745 * compression function transmitted.
1746 */
1747 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS);
1748
1749 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1750 dn->dn_compress = compress;
1751 dnode_setdirty(dn, tx);
1752 dnode_rele(dn, FTAG);
1753 }
1754
1755 int zfs_mdcomp_disable = 0;
1756 TUNABLE_INT("vfs.zfs.mdcomp_disable", &zfs_mdcomp_disable);
1757 SYSCTL_INT(_vfs_zfs, OID_AUTO, mdcomp_disable, CTLFLAG_RW,
1758 &zfs_mdcomp_disable, 0, "Disable metadata compression");
1759
1760 /*
1761 * When the "redundant_metadata" property is set to "most", only indirect
1762 * blocks of this level and higher will have an additional ditto block.
1763 */
1764 int zfs_redundant_metadata_most_ditto_level = 2;
1765
1766 void
dmu_write_policy(objset_t * os,dnode_t * dn,int level,int wp,zio_prop_t * zp)1767 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
1768 {
1769 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
1770 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
1771 (wp & WP_SPILL));
1772 enum zio_checksum checksum = os->os_checksum;
1773 enum zio_compress compress = os->os_compress;
1774 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
1775 boolean_t dedup = B_FALSE;
1776 boolean_t nopwrite = B_FALSE;
1777 boolean_t dedup_verify = os->os_dedup_verify;
1778 int copies = os->os_copies;
1779
1780 /*
1781 * We maintain different write policies for each of the following
1782 * types of data:
1783 * 1. metadata
1784 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1785 * 3. all other level 0 blocks
1786 */
1787 if (ismd) {
1788 /*
1789 * XXX -- we should design a compression algorithm
1790 * that specializes in arrays of bps.
1791 */
1792 boolean_t lz4_ac = spa_feature_is_active(os->os_spa,
1793 SPA_FEATURE_LZ4_COMPRESS);
1794
1795 if (zfs_mdcomp_disable) {
1796 compress = ZIO_COMPRESS_EMPTY;
1797 } else if (lz4_ac) {
1798 compress = ZIO_COMPRESS_LZ4;
1799 } else {
1800 compress = ZIO_COMPRESS_LZJB;
1801 }
1802
1803 /*
1804 * Metadata always gets checksummed. If the data
1805 * checksum is multi-bit correctable, and it's not a
1806 * ZBT-style checksum, then it's suitable for metadata
1807 * as well. Otherwise, the metadata checksum defaults
1808 * to fletcher4.
1809 */
1810 if (zio_checksum_table[checksum].ci_correctable < 1 ||
1811 zio_checksum_table[checksum].ci_eck)
1812 checksum = ZIO_CHECKSUM_FLETCHER_4;
1813
1814 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL ||
1815 (os->os_redundant_metadata ==
1816 ZFS_REDUNDANT_METADATA_MOST &&
1817 (level >= zfs_redundant_metadata_most_ditto_level ||
1818 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))))
1819 copies++;
1820 } else if (wp & WP_NOFILL) {
1821 ASSERT(level == 0);
1822
1823 /*
1824 * If we're writing preallocated blocks, we aren't actually
1825 * writing them so don't set any policy properties. These
1826 * blocks are currently only used by an external subsystem
1827 * outside of zfs (i.e. dump) and not written by the zio
1828 * pipeline.
1829 */
1830 compress = ZIO_COMPRESS_OFF;
1831 checksum = ZIO_CHECKSUM_NOPARITY;
1832 } else {
1833 compress = zio_compress_select(dn->dn_compress, compress);
1834
1835 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
1836 zio_checksum_select(dn->dn_checksum, checksum) :
1837 dedup_checksum;
1838
1839 /*
1840 * Determine dedup setting. If we are in dmu_sync(),
1841 * we won't actually dedup now because that's all
1842 * done in syncing context; but we do want to use the
1843 * dedup checkum. If the checksum is not strong
1844 * enough to ensure unique signatures, force
1845 * dedup_verify.
1846 */
1847 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
1848 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
1849 if (!zio_checksum_table[checksum].ci_dedup)
1850 dedup_verify = B_TRUE;
1851 }
1852
1853 /*
1854 * Enable nopwrite if we have a cryptographically secure
1855 * checksum that has no known collisions (i.e. SHA-256)
1856 * and compression is enabled. We don't enable nopwrite if
1857 * dedup is enabled as the two features are mutually exclusive.
1858 */
1859 nopwrite = (!dedup && zio_checksum_table[checksum].ci_dedup &&
1860 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
1861 }
1862
1863 zp->zp_checksum = checksum;
1864 zp->zp_compress = compress;
1865 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
1866 zp->zp_level = level;
1867 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
1868 zp->zp_dedup = dedup;
1869 zp->zp_dedup_verify = dedup && dedup_verify;
1870 zp->zp_nopwrite = nopwrite;
1871 }
1872
1873 int
dmu_offset_next(objset_t * os,uint64_t object,boolean_t hole,uint64_t * off)1874 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
1875 {
1876 dnode_t *dn;
1877 int i, err;
1878
1879 err = dnode_hold(os, object, FTAG, &dn);
1880 if (err)
1881 return (err);
1882 /*
1883 * Sync any current changes before
1884 * we go trundling through the block pointers.
1885 */
1886 for (i = 0; i < TXG_SIZE; i++) {
1887 if (list_link_active(&dn->dn_dirty_link[i]))
1888 break;
1889 }
1890 if (i != TXG_SIZE) {
1891 dnode_rele(dn, FTAG);
1892 txg_wait_synced(dmu_objset_pool(os), 0);
1893 err = dnode_hold(os, object, FTAG, &dn);
1894 if (err)
1895 return (err);
1896 }
1897
1898 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
1899 dnode_rele(dn, FTAG);
1900
1901 return (err);
1902 }
1903
1904 void
dmu_object_info_from_dnode(dnode_t * dn,dmu_object_info_t * doi)1905 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
1906 {
1907 dnode_phys_t *dnp;
1908
1909 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1910 mutex_enter(&dn->dn_mtx);
1911
1912 dnp = dn->dn_phys;
1913
1914 doi->doi_data_block_size = dn->dn_datablksz;
1915 doi->doi_metadata_block_size = dn->dn_indblkshift ?
1916 1ULL << dn->dn_indblkshift : 0;
1917 doi->doi_type = dn->dn_type;
1918 doi->doi_bonus_type = dn->dn_bonustype;
1919 doi->doi_bonus_size = dn->dn_bonuslen;
1920 doi->doi_indirection = dn->dn_nlevels;
1921 doi->doi_checksum = dn->dn_checksum;
1922 doi->doi_compress = dn->dn_compress;
1923 doi->doi_nblkptr = dn->dn_nblkptr;
1924 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
1925 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
1926 doi->doi_fill_count = 0;
1927 for (int i = 0; i < dnp->dn_nblkptr; i++)
1928 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
1929
1930 mutex_exit(&dn->dn_mtx);
1931 rw_exit(&dn->dn_struct_rwlock);
1932 }
1933
1934 /*
1935 * Get information on a DMU object.
1936 * If doi is NULL, just indicates whether the object exists.
1937 */
1938 int
dmu_object_info(objset_t * os,uint64_t object,dmu_object_info_t * doi)1939 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
1940 {
1941 dnode_t *dn;
1942 int err = dnode_hold(os, object, FTAG, &dn);
1943
1944 if (err)
1945 return (err);
1946
1947 if (doi != NULL)
1948 dmu_object_info_from_dnode(dn, doi);
1949
1950 dnode_rele(dn, FTAG);
1951 return (0);
1952 }
1953
1954 /*
1955 * As above, but faster; can be used when you have a held dbuf in hand.
1956 */
1957 void
dmu_object_info_from_db(dmu_buf_t * db_fake,dmu_object_info_t * doi)1958 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
1959 {
1960 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1961
1962 DB_DNODE_ENTER(db);
1963 dmu_object_info_from_dnode(DB_DNODE(db), doi);
1964 DB_DNODE_EXIT(db);
1965 }
1966
1967 /*
1968 * Faster still when you only care about the size.
1969 * This is specifically optimized for zfs_getattr().
1970 */
1971 void
dmu_object_size_from_db(dmu_buf_t * db_fake,uint32_t * blksize,u_longlong_t * nblk512)1972 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
1973 u_longlong_t *nblk512)
1974 {
1975 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1976 dnode_t *dn;
1977
1978 DB_DNODE_ENTER(db);
1979 dn = DB_DNODE(db);
1980
1981 *blksize = dn->dn_datablksz;
1982 /* add 1 for dnode space */
1983 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
1984 SPA_MINBLOCKSHIFT) + 1;
1985 DB_DNODE_EXIT(db);
1986 }
1987
1988 void
byteswap_uint64_array(void * vbuf,size_t size)1989 byteswap_uint64_array(void *vbuf, size_t size)
1990 {
1991 uint64_t *buf = vbuf;
1992 size_t count = size >> 3;
1993 int i;
1994
1995 ASSERT((size & 7) == 0);
1996
1997 for (i = 0; i < count; i++)
1998 buf[i] = BSWAP_64(buf[i]);
1999 }
2000
2001 void
byteswap_uint32_array(void * vbuf,size_t size)2002 byteswap_uint32_array(void *vbuf, size_t size)
2003 {
2004 uint32_t *buf = vbuf;
2005 size_t count = size >> 2;
2006 int i;
2007
2008 ASSERT((size & 3) == 0);
2009
2010 for (i = 0; i < count; i++)
2011 buf[i] = BSWAP_32(buf[i]);
2012 }
2013
2014 void
byteswap_uint16_array(void * vbuf,size_t size)2015 byteswap_uint16_array(void *vbuf, size_t size)
2016 {
2017 uint16_t *buf = vbuf;
2018 size_t count = size >> 1;
2019 int i;
2020
2021 ASSERT((size & 1) == 0);
2022
2023 for (i = 0; i < count; i++)
2024 buf[i] = BSWAP_16(buf[i]);
2025 }
2026
2027 /* ARGSUSED */
2028 void
byteswap_uint8_array(void * vbuf,size_t size)2029 byteswap_uint8_array(void *vbuf, size_t size)
2030 {
2031 }
2032
2033 void
dmu_init(void)2034 dmu_init(void)
2035 {
2036 zfs_dbgmsg_init();
2037 sa_cache_init();
2038 xuio_stat_init();
2039 dmu_objset_init();
2040 dnode_init();
2041 dbuf_init();
2042 zfetch_init();
2043 zio_compress_init();
2044 l2arc_init();
2045 arc_init();
2046 }
2047
2048 void
dmu_fini(void)2049 dmu_fini(void)
2050 {
2051 arc_fini(); /* arc depends on l2arc, so arc must go first */
2052 l2arc_fini();
2053 zfetch_fini();
2054 zio_compress_fini();
2055 dbuf_fini();
2056 dnode_fini();
2057 dmu_objset_fini();
2058 xuio_stat_fini();
2059 sa_cache_fini();
2060 zfs_dbgmsg_fini();
2061 }
2062