]> git.proxmox.com Git - mirror_zfs.git/blob - module/zfs/dmu.c
projects / mirror_zfs.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Illumos 5661 - ZFS: "compression = on" should use lz4 if feature is enabled
[mirror_zfs.git] / module / zfs / dmu.c
1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2011, 2014 by Delphix. All rights reserved.
24 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
25 * Copyright (c) 2014, Nexenta Systems, Inc. All rights reserved.
26 */
27
28 #include <sys/dmu.h>
29 #include <sys/dmu_impl.h>
30 #include <sys/dmu_tx.h>
31 #include <sys/dbuf.h>
32 #include <sys/dnode.h>
33 #include <sys/zfs_context.h>
34 #include <sys/dmu_objset.h>
35 #include <sys/dmu_traverse.h>
36 #include <sys/dsl_dataset.h>
37 #include <sys/dsl_dir.h>
38 #include <sys/dsl_pool.h>
39 #include <sys/dsl_synctask.h>
40 #include <sys/dsl_prop.h>
41 #include <sys/dmu_zfetch.h>
42 #include <sys/zfs_ioctl.h>
43 #include <sys/zap.h>
44 #include <sys/zio_checksum.h>
45 #include <sys/zio_compress.h>
46 #include <sys/sa.h>
47 #include <sys/zfeature.h>
48 #ifdef _KERNEL
49 #include <sys/vmsystm.h>
50 #include <sys/zfs_znode.h>
51 #endif
52
53 /*
54 * Enable/disable nopwrite feature.
55 */
56 int zfs_nopwrite_enabled = 1;
57
58 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
59 { DMU_BSWAP_UINT8, TRUE, "unallocated" },
60 { DMU_BSWAP_ZAP, TRUE, "object directory" },
61 { DMU_BSWAP_UINT64, TRUE, "object array" },
62 { DMU_BSWAP_UINT8, TRUE, "packed nvlist" },
63 { DMU_BSWAP_UINT64, TRUE, "packed nvlist size" },
64 { DMU_BSWAP_UINT64, TRUE, "bpobj" },
65 { DMU_BSWAP_UINT64, TRUE, "bpobj header" },
66 { DMU_BSWAP_UINT64, TRUE, "SPA space map header" },
67 { DMU_BSWAP_UINT64, TRUE, "SPA space map" },
68 { DMU_BSWAP_UINT64, TRUE, "ZIL intent log" },
69 { DMU_BSWAP_DNODE, TRUE, "DMU dnode" },
70 { DMU_BSWAP_OBJSET, TRUE, "DMU objset" },
71 { DMU_BSWAP_UINT64, TRUE, "DSL directory" },
72 { DMU_BSWAP_ZAP, TRUE, "DSL directory child map"},
73 { DMU_BSWAP_ZAP, TRUE, "DSL dataset snap map" },
74 { DMU_BSWAP_ZAP, TRUE, "DSL props" },
75 { DMU_BSWAP_UINT64, TRUE, "DSL dataset" },
76 { DMU_BSWAP_ZNODE, TRUE, "ZFS znode" },
77 { DMU_BSWAP_OLDACL, TRUE, "ZFS V0 ACL" },
78 { DMU_BSWAP_UINT8, FALSE, "ZFS plain file" },
79 { DMU_BSWAP_ZAP, TRUE, "ZFS directory" },
80 { DMU_BSWAP_ZAP, TRUE, "ZFS master node" },
81 { DMU_BSWAP_ZAP, TRUE, "ZFS delete queue" },
82 { DMU_BSWAP_UINT8, FALSE, "zvol object" },
83 { DMU_BSWAP_ZAP, TRUE, "zvol prop" },
84 { DMU_BSWAP_UINT8, FALSE, "other uint8[]" },
85 { DMU_BSWAP_UINT64, FALSE, "other uint64[]" },
86 { DMU_BSWAP_ZAP, TRUE, "other ZAP" },
87 { DMU_BSWAP_ZAP, TRUE, "persistent error log" },
88 { DMU_BSWAP_UINT8, TRUE, "SPA history" },
89 { DMU_BSWAP_UINT64, TRUE, "SPA history offsets" },
90 { DMU_BSWAP_ZAP, TRUE, "Pool properties" },
91 { DMU_BSWAP_ZAP, TRUE, "DSL permissions" },
92 { DMU_BSWAP_ACL, TRUE, "ZFS ACL" },
93 { DMU_BSWAP_UINT8, TRUE, "ZFS SYSACL" },
94 { DMU_BSWAP_UINT8, TRUE, "FUID table" },
95 { DMU_BSWAP_UINT64, TRUE, "FUID table size" },
96 { DMU_BSWAP_ZAP, TRUE, "DSL dataset next clones"},
97 { DMU_BSWAP_ZAP, TRUE, "scan work queue" },
98 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group used" },
99 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group quota" },
100 { DMU_BSWAP_ZAP, TRUE, "snapshot refcount tags"},
101 { DMU_BSWAP_ZAP, TRUE, "DDT ZAP algorithm" },
102 { DMU_BSWAP_ZAP, TRUE, "DDT statistics" },
103 { DMU_BSWAP_UINT8, TRUE, "System attributes" },
104 { DMU_BSWAP_ZAP, TRUE, "SA master node" },
105 { DMU_BSWAP_ZAP, TRUE, "SA attr registration" },
106 { DMU_BSWAP_ZAP, TRUE, "SA attr layouts" },
107 { DMU_BSWAP_ZAP, TRUE, "scan translations" },
108 { DMU_BSWAP_UINT8, FALSE, "deduplicated block" },
109 { DMU_BSWAP_ZAP, TRUE, "DSL deadlist map" },
110 { DMU_BSWAP_UINT64, TRUE, "DSL deadlist map hdr" },
111 { DMU_BSWAP_ZAP, TRUE, "DSL dir clones" },
112 { DMU_BSWAP_UINT64, TRUE, "bpobj subobj" }
113 };
114
115 const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
116 { byteswap_uint8_array, "uint8" },
117 { byteswap_uint16_array, "uint16" },
118 { byteswap_uint32_array, "uint32" },
119 { byteswap_uint64_array, "uint64" },
120 { zap_byteswap, "zap" },
121 { dnode_buf_byteswap, "dnode" },
122 { dmu_objset_byteswap, "objset" },
123 { zfs_znode_byteswap, "znode" },
124 { zfs_oldacl_byteswap, "oldacl" },
125 { zfs_acl_byteswap, "acl" }
126 };
127
128 int
129 dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset,
130 void *tag, dmu_buf_t **dbp)
131 {
132 dnode_t *dn;
133 uint64_t blkid;
134 dmu_buf_impl_t *db;
135 int err;
136
137 err = dnode_hold(os, object, FTAG, &dn);
138 if (err)
139 return (err);
140 blkid = dbuf_whichblock(dn, offset);
141 rw_enter(&dn->dn_struct_rwlock, RW_READER);
142 db = dbuf_hold(dn, blkid, tag);
143 rw_exit(&dn->dn_struct_rwlock);
144 dnode_rele(dn, FTAG);
145
146 if (db == NULL) {
147 *dbp = NULL;
148 return (SET_ERROR(EIO));
149 }
150
151 *dbp = &db->db;
152 return (err);
153 }
154
155 int
156 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
157 void *tag, dmu_buf_t **dbp, int flags)
158 {
159 int err;
160 int db_flags = DB_RF_CANFAIL;
161
162 if (flags & DMU_READ_NO_PREFETCH)
163 db_flags |= DB_RF_NOPREFETCH;
164
165 err = dmu_buf_hold_noread(os, object, offset, tag, dbp);
166 if (err == 0) {
167 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
168 err = dbuf_read(db, NULL, db_flags);
169 if (err != 0) {
170 dbuf_rele(db, tag);
171 *dbp = NULL;
172 }
173 }
174
175 return (err);
176 }
177
178 int
179 dmu_bonus_max(void)
180 {
181 return (DN_MAX_BONUSLEN);
182 }
183
184 int
185 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
186 {
187 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
188 dnode_t *dn;
189 int error;
190
191 DB_DNODE_ENTER(db);
192 dn = DB_DNODE(db);
193
194 if (dn->dn_bonus != db) {
195 error = SET_ERROR(EINVAL);
196 } else if (newsize < 0 || newsize > db_fake->db_size) {
197 error = SET_ERROR(EINVAL);
198 } else {
199 dnode_setbonuslen(dn, newsize, tx);
200 error = 0;
201 }
202
203 DB_DNODE_EXIT(db);
204 return (error);
205 }
206
207 int
208 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
209 {
210 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
211 dnode_t *dn;
212 int error;
213
214 DB_DNODE_ENTER(db);
215 dn = DB_DNODE(db);
216
217 if (!DMU_OT_IS_VALID(type)) {
218 error = SET_ERROR(EINVAL);
219 } else if (dn->dn_bonus != db) {
220 error = SET_ERROR(EINVAL);
221 } else {
222 dnode_setbonus_type(dn, type, tx);
223 error = 0;
224 }
225
226 DB_DNODE_EXIT(db);
227 return (error);
228 }
229
230 dmu_object_type_t
231 dmu_get_bonustype(dmu_buf_t *db_fake)
232 {
233 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
234 dnode_t *dn;
235 dmu_object_type_t type;
236
237 DB_DNODE_ENTER(db);
238 dn = DB_DNODE(db);
239 type = dn->dn_bonustype;
240 DB_DNODE_EXIT(db);
241
242 return (type);
243 }
244
245 int
246 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
247 {
248 dnode_t *dn;
249 int error;
250
251 error = dnode_hold(os, object, FTAG, &dn);
252 dbuf_rm_spill(dn, tx);
253 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
254 dnode_rm_spill(dn, tx);
255 rw_exit(&dn->dn_struct_rwlock);
256 dnode_rele(dn, FTAG);
257 return (error);
258 }
259
260 /*
261 * returns ENOENT, EIO, or 0.
262 */
263 int
264 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
265 {
266 dnode_t *dn;
267 dmu_buf_impl_t *db;
268 int error;
269
270 error = dnode_hold(os, object, FTAG, &dn);
271 if (error)
272 return (error);
273
274 rw_enter(&dn->dn_struct_rwlock, RW_READER);
275 if (dn->dn_bonus == NULL) {
276 rw_exit(&dn->dn_struct_rwlock);
277 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
278 if (dn->dn_bonus == NULL)
279 dbuf_create_bonus(dn);
280 }
281 db = dn->dn_bonus;
282
283 /* as long as the bonus buf is held, the dnode will be held */
284 if (refcount_add(&db->db_holds, tag) == 1) {
285 VERIFY(dnode_add_ref(dn, db));
286 atomic_inc_32(&dn->dn_dbufs_count);
287 }
288
289 /*
290 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
291 * hold and incrementing the dbuf count to ensure that dnode_move() sees
292 * a dnode hold for every dbuf.
293 */
294 rw_exit(&dn->dn_struct_rwlock);
295
296 dnode_rele(dn, FTAG);
297
298 VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH));
299
300 *dbp = &db->db;
301 return (0);
302 }
303
304 /*
305 * returns ENOENT, EIO, or 0.
306 *
307 * This interface will allocate a blank spill dbuf when a spill blk
308 * doesn't already exist on the dnode.
309 *
310 * if you only want to find an already existing spill db, then
311 * dmu_spill_hold_existing() should be used.
312 */
313 int
314 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
315 {
316 dmu_buf_impl_t *db = NULL;
317 int err;
318
319 if ((flags & DB_RF_HAVESTRUCT) == 0)
320 rw_enter(&dn->dn_struct_rwlock, RW_READER);
321
322 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
323
324 if ((flags & DB_RF_HAVESTRUCT) == 0)
325 rw_exit(&dn->dn_struct_rwlock);
326
327 ASSERT(db != NULL);
328 err = dbuf_read(db, NULL, flags);
329 if (err == 0)
330 *dbp = &db->db;
331 else
332 dbuf_rele(db, tag);
333 return (err);
334 }
335
336 int
337 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
338 {
339 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
340 dnode_t *dn;
341 int err;
342
343 DB_DNODE_ENTER(db);
344 dn = DB_DNODE(db);
345
346 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
347 err = SET_ERROR(EINVAL);
348 } else {
349 rw_enter(&dn->dn_struct_rwlock, RW_READER);
350
351 if (!dn->dn_have_spill) {
352 err = SET_ERROR(ENOENT);
353 } else {
354 err = dmu_spill_hold_by_dnode(dn,
355 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
356 }
357
358 rw_exit(&dn->dn_struct_rwlock);
359 }
360
361 DB_DNODE_EXIT(db);
362 return (err);
363 }
364
365 int
366 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
367 {
368 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
369 dnode_t *dn;
370 int err;
371
372 DB_DNODE_ENTER(db);
373 dn = DB_DNODE(db);
374 err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp);
375 DB_DNODE_EXIT(db);
376
377 return (err);
378 }
379
380 /*
381 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
382 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
383 * and can induce severe lock contention when writing to several files
384 * whose dnodes are in the same block.
385 */
386 static int
387 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
388 int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
389 {
390 dmu_buf_t **dbp;
391 uint64_t blkid, nblks, i;
392 uint32_t dbuf_flags;
393 int err;
394 zio_t *zio;
395
396 ASSERT(length <= DMU_MAX_ACCESS);
397
398 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT;
399 if (flags & DMU_READ_NO_PREFETCH || length > zfetch_array_rd_sz)
400 dbuf_flags |= DB_RF_NOPREFETCH;
401
402 rw_enter(&dn->dn_struct_rwlock, RW_READER);
403 if (dn->dn_datablkshift) {
404 int blkshift = dn->dn_datablkshift;
405 nblks = (P2ROUNDUP(offset+length, 1ULL<<blkshift) -
406 P2ALIGN(offset, 1ULL<<blkshift)) >> blkshift;
407 } else {
408 if (offset + length > dn->dn_datablksz) {
409 zfs_panic_recover("zfs: accessing past end of object "
410 "%llx/%llx (size=%u access=%llu+%llu)",
411 (longlong_t)dn->dn_objset->
412 os_dsl_dataset->ds_object,
413 (longlong_t)dn->dn_object, dn->dn_datablksz,
414 (longlong_t)offset, (longlong_t)length);
415 rw_exit(&dn->dn_struct_rwlock);
416 return (SET_ERROR(EIO));
417 }
418 nblks = 1;
419 }
420 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
421
422 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
423 blkid = dbuf_whichblock(dn, offset);
424 for (i = 0; i < nblks; i++) {
425 dmu_buf_impl_t *db = dbuf_hold(dn, blkid+i, tag);
426 if (db == NULL) {
427 rw_exit(&dn->dn_struct_rwlock);
428 dmu_buf_rele_array(dbp, nblks, tag);
429 zio_nowait(zio);
430 return (SET_ERROR(EIO));
431 }
432 /* initiate async i/o */
433 if (read) {
434 (void) dbuf_read(db, zio, dbuf_flags);
435 }
436 dbp[i] = &db->db;
437 }
438 rw_exit(&dn->dn_struct_rwlock);
439
440 /* wait for async i/o */
441 err = zio_wait(zio);
442 if (err) {
443 dmu_buf_rele_array(dbp, nblks, tag);
444 return (err);
445 }
446
447 /* wait for other io to complete */
448 if (read) {
449 for (i = 0; i < nblks; i++) {
450 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
451 mutex_enter(&db->db_mtx);
452 while (db->db_state == DB_READ ||
453 db->db_state == DB_FILL)
454 cv_wait(&db->db_changed, &db->db_mtx);
455 if (db->db_state == DB_UNCACHED)
456 err = SET_ERROR(EIO);
457 mutex_exit(&db->db_mtx);
458 if (err) {
459 dmu_buf_rele_array(dbp, nblks, tag);
460 return (err);
461 }
462 }
463 }
464
465 *numbufsp = nblks;
466 *dbpp = dbp;
467 return (0);
468 }
469
470 static int
471 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
472 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
473 {
474 dnode_t *dn;
475 int err;
476
477 err = dnode_hold(os, object, FTAG, &dn);
478 if (err)
479 return (err);
480
481 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
482 numbufsp, dbpp, DMU_READ_PREFETCH);
483
484 dnode_rele(dn, FTAG);
485
486 return (err);
487 }
488
489 int
490 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
491 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
492 {
493 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
494 dnode_t *dn;
495 int err;
496
497 DB_DNODE_ENTER(db);
498 dn = DB_DNODE(db);
499 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
500 numbufsp, dbpp, DMU_READ_PREFETCH);
501 DB_DNODE_EXIT(db);
502
503 return (err);
504 }
505
506 void
507 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
508 {
509 int i;
510 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
511
512 if (numbufs == 0)
513 return;
514
515 for (i = 0; i < numbufs; i++) {
516 if (dbp[i])
517 dbuf_rele(dbp[i], tag);
518 }
519
520 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
521 }
522
523 /*
524 * Issue prefetch i/os for the given blocks.
525 *
526 * Note: The assumption is that we *know* these blocks will be needed
527 * almost immediately. Therefore, the prefetch i/os will be issued at
528 * ZIO_PRIORITY_SYNC_READ
529 *
530 * Note: indirect blocks and other metadata will be read synchronously,
531 * causing this function to block if they are not already cached.
532 */
533 void
534 dmu_prefetch(objset_t *os, uint64_t object, uint64_t offset, uint64_t len)
535 {
536 dnode_t *dn;
537 uint64_t blkid;
538 int nblks, err;
539
540 if (zfs_prefetch_disable)
541 return;
542
543 if (len == 0) { /* they're interested in the bonus buffer */
544 dn = DMU_META_DNODE(os);
545
546 if (object == 0 || object >= DN_MAX_OBJECT)
547 return;
548
549 rw_enter(&dn->dn_struct_rwlock, RW_READER);
550 blkid = dbuf_whichblock(dn, object * sizeof (dnode_phys_t));
551 dbuf_prefetch(dn, blkid, ZIO_PRIORITY_SYNC_READ);
552 rw_exit(&dn->dn_struct_rwlock);
553 return;
554 }
555
556 /*
557 * XXX - Note, if the dnode for the requested object is not
558 * already cached, we will do a *synchronous* read in the
559 * dnode_hold() call. The same is true for any indirects.
560 */
561 err = dnode_hold(os, object, FTAG, &dn);
562 if (err != 0)
563 return;
564
565 rw_enter(&dn->dn_struct_rwlock, RW_READER);
566 if (dn->dn_datablkshift) {
567 int blkshift = dn->dn_datablkshift;
568 nblks = (P2ROUNDUP(offset + len, 1 << blkshift) -
569 P2ALIGN(offset, 1 << blkshift)) >> blkshift;
570 } else {
571 nblks = (offset < dn->dn_datablksz);
572 }
573
574 if (nblks != 0) {
575 int i;
576
577 blkid = dbuf_whichblock(dn, offset);
578 for (i = 0; i < nblks; i++)
579 dbuf_prefetch(dn, blkid + i, ZIO_PRIORITY_SYNC_READ);
580 }
581
582 rw_exit(&dn->dn_struct_rwlock);
583
584 dnode_rele(dn, FTAG);
585 }
586
587 /*
588 * Get the next "chunk" of file data to free. We traverse the file from
589 * the end so that the file gets shorter over time (if we crashes in the
590 * middle, this will leave us in a better state). We find allocated file
591 * data by simply searching the allocated level 1 indirects.
592 *
593 * On input, *start should be the first offset that does not need to be
594 * freed (e.g. "offset + length"). On return, *start will be the first
595 * offset that should be freed.
596 */
597 static int
598 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum)
599 {
600 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
601 /* bytes of data covered by a level-1 indirect block */
602 uint64_t iblkrange =
603 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
604 uint64_t blks;
605
606 ASSERT3U(minimum, <=, *start);
607
608 if (*start - minimum <= iblkrange * maxblks) {
609 *start = minimum;
610 return (0);
611 }
612 ASSERT(ISP2(iblkrange));
613
614 for (blks = 0; *start > minimum && blks < maxblks; blks++) {
615 int err;
616
617 /*
618 * dnode_next_offset(BACKWARDS) will find an allocated L1
619 * indirect block at or before the input offset. We must
620 * decrement *start so that it is at the end of the region
621 * to search.
622 */
623 (*start)--;
624 err = dnode_next_offset(dn,
625 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
626
627 /* if there are no indirect blocks before start, we are done */
628 if (err == ESRCH) {
629 *start = minimum;
630 break;
631 } else if (err != 0) {
632 return (err);
633 }
634
635 /* set start to the beginning of this L1 indirect */
636 *start = P2ALIGN(*start, iblkrange);
637 }
638 if (*start < minimum)
639 *start = minimum;
640 return (0);
641 }
642
643 static int
644 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
645 uint64_t length)
646 {
647 uint64_t object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
648 int err;
649
650 if (offset >= object_size)
651 return (0);
652
653 if (length == DMU_OBJECT_END || offset + length > object_size)
654 length = object_size - offset;
655
656 while (length != 0) {
657 uint64_t chunk_end, chunk_begin;
658 dmu_tx_t *tx;
659
660 chunk_end = chunk_begin = offset + length;
661
662 /* move chunk_begin backwards to the beginning of this chunk */
663 err = get_next_chunk(dn, &chunk_begin, offset);
664 if (err)
665 return (err);
666 ASSERT3U(chunk_begin, >=, offset);
667 ASSERT3U(chunk_begin, <=, chunk_end);
668
669 tx = dmu_tx_create(os);
670 dmu_tx_hold_free(tx, dn->dn_object,
671 chunk_begin, chunk_end - chunk_begin);
672 err = dmu_tx_assign(tx, TXG_WAIT);
673 if (err) {
674 dmu_tx_abort(tx);
675 return (err);
676 }
677 dnode_free_range(dn, chunk_begin, chunk_end - chunk_begin, tx);
678 dmu_tx_commit(tx);
679
680 length -= chunk_end - chunk_begin;
681 }
682 return (0);
683 }
684
685 int
686 dmu_free_long_range(objset_t *os, uint64_t object,
687 uint64_t offset, uint64_t length)
688 {
689 dnode_t *dn;
690 int err;
691
692 err = dnode_hold(os, object, FTAG, &dn);
693 if (err != 0)
694 return (err);
695 err = dmu_free_long_range_impl(os, dn, offset, length);
696
697 /*
698 * It is important to zero out the maxblkid when freeing the entire
699 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
700 * will take the fast path, and (b) dnode_reallocate() can verify
701 * that the entire file has been freed.
702 */
703 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
704 dn->dn_maxblkid = 0;
705
706 dnode_rele(dn, FTAG);
707 return (err);
708 }
709
710 int
711 dmu_free_long_object(objset_t *os, uint64_t object)
712 {
713 dmu_tx_t *tx;
714 int err;
715
716 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
717 if (err != 0)
718 return (err);
719
720 tx = dmu_tx_create(os);
721 dmu_tx_hold_bonus(tx, object);
722 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
723 err = dmu_tx_assign(tx, TXG_WAIT);
724 if (err == 0) {
725 err = dmu_object_free(os, object, tx);
726 dmu_tx_commit(tx);
727 } else {
728 dmu_tx_abort(tx);
729 }
730
731 return (err);
732 }
733
734 int
735 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
736 uint64_t size, dmu_tx_t *tx)
737 {
738 dnode_t *dn;
739 int err = dnode_hold(os, object, FTAG, &dn);
740 if (err)
741 return (err);
742 ASSERT(offset < UINT64_MAX);
743 ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
744 dnode_free_range(dn, offset, size, tx);
745 dnode_rele(dn, FTAG);
746 return (0);
747 }
748
749 int
750 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
751 void *buf, uint32_t flags)
752 {
753 dnode_t *dn;
754 dmu_buf_t **dbp;
755 int numbufs, err;
756
757 err = dnode_hold(os, object, FTAG, &dn);
758 if (err)
759 return (err);
760
761 /*
762 * Deal with odd block sizes, where there can't be data past the first
763 * block. If we ever do the tail block optimization, we will need to
764 * handle that here as well.
765 */
766 if (dn->dn_maxblkid == 0) {
767 uint64_t newsz = offset > dn->dn_datablksz ? 0 :
768 MIN(size, dn->dn_datablksz - offset);
769 bzero((char *)buf + newsz, size - newsz);
770 size = newsz;
771 }
772
773 while (size > 0) {
774 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
775 int i;
776
777 /*
778 * NB: we could do this block-at-a-time, but it's nice
779 * to be reading in parallel.
780 */
781 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
782 TRUE, FTAG, &numbufs, &dbp, flags);
783 if (err)
784 break;
785
786 for (i = 0; i < numbufs; i++) {
787 uint64_t tocpy;
788 int64_t bufoff;
789 dmu_buf_t *db = dbp[i];
790
791 ASSERT(size > 0);
792
793 bufoff = offset - db->db_offset;
794 tocpy = MIN(db->db_size - bufoff, size);
795
796 (void) memcpy(buf, (char *)db->db_data + bufoff, tocpy);
797
798 offset += tocpy;
799 size -= tocpy;
800 buf = (char *)buf + tocpy;
801 }
802 dmu_buf_rele_array(dbp, numbufs, FTAG);
803 }
804 dnode_rele(dn, FTAG);
805 return (err);
806 }
807
808 void
809 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
810 const void *buf, dmu_tx_t *tx)
811 {
812 dmu_buf_t **dbp;
813 int numbufs, i;
814
815 if (size == 0)
816 return;
817
818 VERIFY0(dmu_buf_hold_array(os, object, offset, size,
819 FALSE, FTAG, &numbufs, &dbp));
820
821 for (i = 0; i < numbufs; i++) {
822 uint64_t tocpy;
823 int64_t bufoff;
824 dmu_buf_t *db = dbp[i];
825
826 ASSERT(size > 0);
827
828 bufoff = offset - db->db_offset;
829 tocpy = MIN(db->db_size - bufoff, size);
830
831 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
832
833 if (tocpy == db->db_size)
834 dmu_buf_will_fill(db, tx);
835 else
836 dmu_buf_will_dirty(db, tx);
837
838 (void) memcpy((char *)db->db_data + bufoff, buf, tocpy);
839
840 if (tocpy == db->db_size)
841 dmu_buf_fill_done(db, tx);
842
843 offset += tocpy;
844 size -= tocpy;
845 buf = (char *)buf + tocpy;
846 }
847 dmu_buf_rele_array(dbp, numbufs, FTAG);
848 }
849
850 void
851 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
852 dmu_tx_t *tx)
853 {
854 dmu_buf_t **dbp;
855 int numbufs, i;
856
857 if (size == 0)
858 return;
859
860 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
861 FALSE, FTAG, &numbufs, &dbp));
862
863 for (i = 0; i < numbufs; i++) {
864 dmu_buf_t *db = dbp[i];
865
866 dmu_buf_will_not_fill(db, tx);
867 }
868 dmu_buf_rele_array(dbp, numbufs, FTAG);
869 }
870
871 void
872 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
873 void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
874 int compressed_size, int byteorder, dmu_tx_t *tx)
875 {
876 dmu_buf_t *db;
877
878 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
879 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
880 VERIFY0(dmu_buf_hold_noread(os, object, offset,
881 FTAG, &db));
882
883 dmu_buf_write_embedded(db,
884 data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
885 uncompressed_size, compressed_size, byteorder, tx);
886
887 dmu_buf_rele(db, FTAG);
888 }
889
890 /*
891 * DMU support for xuio
892 */
893 kstat_t *xuio_ksp = NULL;
894
895 typedef struct xuio_stats {
896 /* loaned yet not returned arc_buf */
897 kstat_named_t xuiostat_onloan_rbuf;
898 kstat_named_t xuiostat_onloan_wbuf;
899 /* whether a copy is made when loaning out a read buffer */
900 kstat_named_t xuiostat_rbuf_copied;
901 kstat_named_t xuiostat_rbuf_nocopy;
902 /* whether a copy is made when assigning a write buffer */
903 kstat_named_t xuiostat_wbuf_copied;
904 kstat_named_t xuiostat_wbuf_nocopy;
905 } xuio_stats_t;
906
907 static xuio_stats_t xuio_stats = {
908 { "onloan_read_buf", KSTAT_DATA_UINT64 },
909 { "onloan_write_buf", KSTAT_DATA_UINT64 },
910 { "read_buf_copied", KSTAT_DATA_UINT64 },
911 { "read_buf_nocopy", KSTAT_DATA_UINT64 },
912 { "write_buf_copied", KSTAT_DATA_UINT64 },
913 { "write_buf_nocopy", KSTAT_DATA_UINT64 }
914 };
915
916 #define XUIOSTAT_INCR(stat, val) \
917 atomic_add_64(&xuio_stats.stat.value.ui64, (val))
918 #define XUIOSTAT_BUMP(stat) XUIOSTAT_INCR(stat, 1)
919
920 int
921 dmu_xuio_init(xuio_t *xuio, int nblk)
922 {
923 dmu_xuio_t *priv;
924 uio_t *uio = &xuio->xu_uio;
925
926 uio->uio_iovcnt = nblk;
927 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
928
929 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
930 priv->cnt = nblk;
931 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
932 priv->iovp = uio->uio_iov;
933 XUIO_XUZC_PRIV(xuio) = priv;
934
935 if (XUIO_XUZC_RW(xuio) == UIO_READ)
936 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
937 else
938 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
939
940 return (0);
941 }
942
943 void
944 dmu_xuio_fini(xuio_t *xuio)
945 {
946 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
947 int nblk = priv->cnt;
948
949 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
950 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
951 kmem_free(priv, sizeof (dmu_xuio_t));
952
953 if (XUIO_XUZC_RW(xuio) == UIO_READ)
954 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
955 else
956 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
957 }
958
959 /*
960 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
961 * and increase priv->next by 1.
962 */
963 int
964 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
965 {
966 struct iovec *iov;
967 uio_t *uio = &xuio->xu_uio;
968 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
969 int i = priv->next++;
970
971 ASSERT(i < priv->cnt);
972 ASSERT(off + n <= arc_buf_size(abuf));
973 iov = uio->uio_iov + i;
974 iov->iov_base = (char *)abuf->b_data + off;
975 iov->iov_len = n;
976 priv->bufs[i] = abuf;
977 return (0);
978 }
979
980 int
981 dmu_xuio_cnt(xuio_t *xuio)
982 {
983 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
984 return (priv->cnt);
985 }
986
987 arc_buf_t *
988 dmu_xuio_arcbuf(xuio_t *xuio, int i)
989 {
990 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
991
992 ASSERT(i < priv->cnt);
993 return (priv->bufs[i]);
994 }
995
996 void
997 dmu_xuio_clear(xuio_t *xuio, int i)
998 {
999 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1000
1001 ASSERT(i < priv->cnt);
1002 priv->bufs[i] = NULL;
1003 }
1004
1005 static void
1006 xuio_stat_init(void)
1007 {
1008 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
1009 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
1010 KSTAT_FLAG_VIRTUAL);
1011 if (xuio_ksp != NULL) {
1012 xuio_ksp->ks_data = &xuio_stats;
1013 kstat_install(xuio_ksp);
1014 }
1015 }
1016
1017 static void
1018 xuio_stat_fini(void)
1019 {
1020 if (xuio_ksp != NULL) {
1021 kstat_delete(xuio_ksp);
1022 xuio_ksp = NULL;
1023 }
1024 }
1025
1026 void
1027 xuio_stat_wbuf_copied()
1028 {
1029 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1030 }
1031
1032 void
1033 xuio_stat_wbuf_nocopy()
1034 {
1035 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
1036 }
1037
1038 #ifdef _KERNEL
1039
1040 /*
1041 * Copy up to size bytes between arg_buf and req based on the data direction
1042 * described by the req. If an entire req's data cannot be transfered in one
1043 * pass, you should pass in @req_offset to indicate where to continue. The
1044 * return value is the number of bytes successfully copied to arg_buf.
1045 */
1046 static int
1047 dmu_req_copy(void *arg_buf, int size, struct request *req, size_t req_offset)
1048 {
1049 struct bio_vec bv, *bvp;
1050 struct req_iterator iter;
1051 char *bv_buf;
1052 int tocpy, bv_len, bv_offset;
1053 int offset = 0;
1054
1055 rq_for_each_segment4(bv, bvp, req, iter) {
1056 /*
1057 * Fully consumed the passed arg_buf. We use goto here because
1058 * rq_for_each_segment is a double loop
1059 */
1060 ASSERT3S(offset, <=, size);
1061 if (size == offset)
1062 goto out;
1063
1064 /* Skip already copied bv */
1065 if (req_offset >= bv.bv_len) {
1066 req_offset -= bv.bv_len;
1067 continue;
1068 }
1069
1070 bv_len = bv.bv_len - req_offset;
1071 bv_offset = bv.bv_offset + req_offset;
1072 req_offset = 0;
1073
1074 tocpy = MIN(bv_len, size - offset);
1075 ASSERT3S(tocpy, >=, 0);
1076
1077 bv_buf = page_address(bv.bv_page) + bv_offset;
1078 ASSERT3P(bv_buf, !=, NULL);
1079
1080 if (rq_data_dir(req) == WRITE)
1081 memcpy(arg_buf + offset, bv_buf, tocpy);
1082 else
1083 memcpy(bv_buf, arg_buf + offset, tocpy);
1084
1085 offset += tocpy;
1086 }
1087 out:
1088 return (offset);
1089 }
1090
1091 int
1092 dmu_read_req(objset_t *os, uint64_t object, struct request *req)
1093 {
1094 uint64_t size = blk_rq_bytes(req);
1095 uint64_t offset = blk_rq_pos(req) << 9;
1096 dmu_buf_t **dbp;
1097 int numbufs, i, err;
1098 size_t req_offset;
1099
1100 /*
1101 * NB: we could do this block-at-a-time, but it's nice
1102 * to be reading in parallel.
1103 */
1104 err = dmu_buf_hold_array(os, object, offset, size, TRUE, FTAG,
1105 &numbufs, &dbp);
1106 if (err)
1107 return (err);
1108
1109 req_offset = 0;
1110 for (i = 0; i < numbufs; i++) {
1111 uint64_t tocpy;
1112 int64_t bufoff;
1113 int didcpy;
1114 dmu_buf_t *db = dbp[i];
1115
1116 bufoff = offset - db->db_offset;
1117 ASSERT3S(bufoff, >=, 0);
1118
1119 tocpy = MIN(db->db_size - bufoff, size);
1120 if (tocpy == 0)
1121 break;
1122
1123 didcpy = dmu_req_copy(db->db_data + bufoff, tocpy, req,
1124 req_offset);
1125
1126 if (didcpy < tocpy)
1127 err = EIO;
1128
1129 if (err)
1130 break;
1131
1132 size -= tocpy;
1133 offset += didcpy;
1134 req_offset += didcpy;
1135 err = 0;
1136 }
1137 dmu_buf_rele_array(dbp, numbufs, FTAG);
1138
1139 return (err);
1140 }
1141
1142 int
1143 dmu_write_req(objset_t *os, uint64_t object, struct request *req, dmu_tx_t *tx)
1144 {
1145 uint64_t size = blk_rq_bytes(req);
1146 uint64_t offset = blk_rq_pos(req) << 9;
1147 dmu_buf_t **dbp;
1148 int numbufs, i, err;
1149 size_t req_offset;
1150
1151 if (size == 0)
1152 return (0);
1153
1154 err = dmu_buf_hold_array(os, object, offset, size, FALSE, FTAG,
1155 &numbufs, &dbp);
1156 if (err)
1157 return (err);
1158
1159 req_offset = 0;
1160 for (i = 0; i < numbufs; i++) {
1161 uint64_t tocpy;
1162 int64_t bufoff;
1163 int didcpy;
1164 dmu_buf_t *db = dbp[i];
1165
1166 bufoff = offset - db->db_offset;
1167 ASSERT3S(bufoff, >=, 0);
1168
1169 tocpy = MIN(db->db_size - bufoff, size);
1170 if (tocpy == 0)
1171 break;
1172
1173 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1174
1175 if (tocpy == db->db_size)
1176 dmu_buf_will_fill(db, tx);
1177 else
1178 dmu_buf_will_dirty(db, tx);
1179
1180 didcpy = dmu_req_copy(db->db_data + bufoff, tocpy, req,
1181 req_offset);
1182
1183 if (tocpy == db->db_size)
1184 dmu_buf_fill_done(db, tx);
1185
1186 if (didcpy < tocpy)
1187 err = EIO;
1188
1189 if (err)
1190 break;
1191
1192 size -= tocpy;
1193 offset += didcpy;
1194 req_offset += didcpy;
1195 err = 0;
1196 }
1197
1198 dmu_buf_rele_array(dbp, numbufs, FTAG);
1199 return (err);
1200 }
1201
1202 static int
1203 dmu_read_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size)
1204 {
1205 dmu_buf_t **dbp;
1206 int numbufs, i, err;
1207 xuio_t *xuio = NULL;
1208
1209 /*
1210 * NB: we could do this block-at-a-time, but it's nice
1211 * to be reading in parallel.
1212 */
1213 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1214 TRUE, FTAG, &numbufs, &dbp, 0);
1215 if (err)
1216 return (err);
1217
1218 for (i = 0; i < numbufs; i++) {
1219 uint64_t tocpy;
1220 int64_t bufoff;
1221 dmu_buf_t *db = dbp[i];
1222
1223 ASSERT(size > 0);
1224
1225 bufoff = uio->uio_loffset - db->db_offset;
1226 tocpy = MIN(db->db_size - bufoff, size);
1227
1228 if (xuio) {
1229 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
1230 arc_buf_t *dbuf_abuf = dbi->db_buf;
1231 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
1232 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
1233 if (!err) {
1234 uio->uio_resid -= tocpy;
1235 uio->uio_loffset += tocpy;
1236 }
1237
1238 if (abuf == dbuf_abuf)
1239 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
1240 else
1241 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
1242 } else {
1243 err = uiomove((char *)db->db_data + bufoff, tocpy,
1244 UIO_READ, uio);
1245 }
1246 if (err)
1247 break;
1248
1249 size -= tocpy;
1250 }
1251 dmu_buf_rele_array(dbp, numbufs, FTAG);
1252
1253 return (err);
1254 }
1255
1256 /*
1257 * Read 'size' bytes into the uio buffer.
1258 * From object zdb->db_object.
1259 * Starting at offset uio->uio_loffset.
1260 *
1261 * If the caller already has a dbuf in the target object
1262 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1263 * because we don't have to find the dnode_t for the object.
1264 */
1265 int
1266 dmu_read_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size)
1267 {
1268 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1269 dnode_t *dn;
1270 int err;
1271
1272 if (size == 0)
1273 return (0);
1274
1275 DB_DNODE_ENTER(db);
1276 dn = DB_DNODE(db);
1277 err = dmu_read_uio_dnode(dn, uio, size);
1278 DB_DNODE_EXIT(db);
1279
1280 return (err);
1281 }
1282
1283 /*
1284 * Read 'size' bytes into the uio buffer.
1285 * From the specified object
1286 * Starting at offset uio->uio_loffset.
1287 */
1288 int
1289 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
1290 {
1291 dnode_t *dn;
1292 int err;
1293
1294 if (size == 0)
1295 return (0);
1296
1297 err = dnode_hold(os, object, FTAG, &dn);
1298 if (err)
1299 return (err);
1300
1301 err = dmu_read_uio_dnode(dn, uio, size);
1302
1303 dnode_rele(dn, FTAG);
1304
1305 return (err);
1306 }
1307
1308 static int
1309 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1310 {
1311 dmu_buf_t **dbp;
1312 int numbufs;
1313 int err = 0;
1314 int i;
1315
1316 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1317 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1318 if (err)
1319 return (err);
1320
1321 for (i = 0; i < numbufs; i++) {
1322 uint64_t tocpy;
1323 int64_t bufoff;
1324 dmu_buf_t *db = dbp[i];
1325
1326 ASSERT(size > 0);
1327
1328 bufoff = uio->uio_loffset - db->db_offset;
1329 tocpy = MIN(db->db_size - bufoff, size);
1330
1331 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1332
1333 if (tocpy == db->db_size)
1334 dmu_buf_will_fill(db, tx);
1335 else
1336 dmu_buf_will_dirty(db, tx);
1337
1338 /*
1339 * XXX uiomove could block forever (eg.nfs-backed
1340 * pages). There needs to be a uiolockdown() function
1341 * to lock the pages in memory, so that uiomove won't
1342 * block.
1343 */
1344 err = uiomove((char *)db->db_data + bufoff, tocpy,
1345 UIO_WRITE, uio);
1346
1347 if (tocpy == db->db_size)
1348 dmu_buf_fill_done(db, tx);
1349
1350 if (err)
1351 break;
1352
1353 size -= tocpy;
1354 }
1355
1356 dmu_buf_rele_array(dbp, numbufs, FTAG);
1357 return (err);
1358 }
1359
1360 /*
1361 * Write 'size' bytes from the uio buffer.
1362 * To object zdb->db_object.
1363 * Starting at offset uio->uio_loffset.
1364 *
1365 * If the caller already has a dbuf in the target object
1366 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1367 * because we don't have to find the dnode_t for the object.
1368 */
1369 int
1370 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1371 dmu_tx_t *tx)
1372 {
1373 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1374 dnode_t *dn;
1375 int err;
1376
1377 if (size == 0)
1378 return (0);
1379
1380 DB_DNODE_ENTER(db);
1381 dn = DB_DNODE(db);
1382 err = dmu_write_uio_dnode(dn, uio, size, tx);
1383 DB_DNODE_EXIT(db);
1384
1385 return (err);
1386 }
1387
1388 /*
1389 * Write 'size' bytes from the uio buffer.
1390 * To the specified object.
1391 * Starting at offset uio->uio_loffset.
1392 */
1393 int
1394 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1395 dmu_tx_t *tx)
1396 {
1397 dnode_t *dn;
1398 int err;
1399
1400 if (size == 0)
1401 return (0);
1402
1403 err = dnode_hold(os, object, FTAG, &dn);
1404 if (err)
1405 return (err);
1406
1407 err = dmu_write_uio_dnode(dn, uio, size, tx);
1408
1409 dnode_rele(dn, FTAG);
1410
1411 return (err);
1412 }
1413 #endif /* _KERNEL */
1414
1415 /*
1416 * Allocate a loaned anonymous arc buffer.
1417 */
1418 arc_buf_t *
1419 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1420 {
1421 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1422
1423 return (arc_loan_buf(db->db_objset->os_spa, size));
1424 }
1425
1426 /*
1427 * Free a loaned arc buffer.
1428 */
1429 void
1430 dmu_return_arcbuf(arc_buf_t *buf)
1431 {
1432 arc_return_buf(buf, FTAG);
1433 VERIFY(arc_buf_remove_ref(buf, FTAG));
1434 }
1435
1436 /*
1437 * When possible directly assign passed loaned arc buffer to a dbuf.
1438 * If this is not possible copy the contents of passed arc buf via
1439 * dmu_write().
1440 */
1441 void
1442 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1443 dmu_tx_t *tx)
1444 {
1445 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1446 dnode_t *dn;
1447 dmu_buf_impl_t *db;
1448 uint32_t blksz = (uint32_t)arc_buf_size(buf);
1449 uint64_t blkid;
1450
1451 DB_DNODE_ENTER(dbuf);
1452 dn = DB_DNODE(dbuf);
1453 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1454 blkid = dbuf_whichblock(dn, offset);
1455 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1456 rw_exit(&dn->dn_struct_rwlock);
1457 DB_DNODE_EXIT(dbuf);
1458
1459 /*
1460 * We can only assign if the offset is aligned, the arc buf is the
1461 * same size as the dbuf, and the dbuf is not metadata. It
1462 * can't be metadata because the loaned arc buf comes from the
1463 * user-data kmem area.
1464 */
1465 if (offset == db->db.db_offset && blksz == db->db.db_size &&
1466 DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA) {
1467 dbuf_assign_arcbuf(db, buf, tx);
1468 dbuf_rele(db, FTAG);
1469 } else {
1470 objset_t *os;
1471 uint64_t object;
1472
1473 DB_DNODE_ENTER(dbuf);
1474 dn = DB_DNODE(dbuf);
1475 os = dn->dn_objset;
1476 object = dn->dn_object;
1477 DB_DNODE_EXIT(dbuf);
1478
1479 dbuf_rele(db, FTAG);
1480 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1481 dmu_return_arcbuf(buf);
1482 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1483 }
1484 }
1485
1486 typedef struct {
1487 dbuf_dirty_record_t *dsa_dr;
1488 dmu_sync_cb_t *dsa_done;
1489 zgd_t *dsa_zgd;
1490 dmu_tx_t *dsa_tx;
1491 } dmu_sync_arg_t;
1492
1493 /* ARGSUSED */
1494 static void
1495 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1496 {
1497 dmu_sync_arg_t *dsa = varg;
1498 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1499 blkptr_t *bp = zio->io_bp;
1500
1501 if (zio->io_error == 0) {
1502 if (BP_IS_HOLE(bp)) {
1503 /*
1504 * A block of zeros may compress to a hole, but the
1505 * block size still needs to be known for replay.
1506 */
1507 BP_SET_LSIZE(bp, db->db_size);
1508 } else if (!BP_IS_EMBEDDED(bp)) {
1509 ASSERT(BP_GET_LEVEL(bp) == 0);
1510 bp->blk_fill = 1;
1511 }
1512 }
1513 }
1514
1515 static void
1516 dmu_sync_late_arrival_ready(zio_t *zio)
1517 {
1518 dmu_sync_ready(zio, NULL, zio->io_private);
1519 }
1520
1521 /* ARGSUSED */
1522 static void
1523 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1524 {
1525 dmu_sync_arg_t *dsa = varg;
1526 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1527 dmu_buf_impl_t *db = dr->dr_dbuf;
1528
1529 mutex_enter(&db->db_mtx);
1530 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1531 if (zio->io_error == 0) {
1532 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1533 if (dr->dt.dl.dr_nopwrite) {
1534 ASSERTV(blkptr_t *bp = zio->io_bp);
1535 ASSERTV(blkptr_t *bp_orig = &zio->io_bp_orig);
1536 ASSERTV(uint8_t chksum = BP_GET_CHECKSUM(bp_orig));
1537
1538 ASSERT(BP_EQUAL(bp, bp_orig));
1539 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1540 ASSERT(zio_checksum_table[chksum].ci_dedup);
1541 }
1542 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1543 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1544 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1545
1546 /*
1547 * Old style holes are filled with all zeros, whereas
1548 * new-style holes maintain their lsize, type, level,
1549 * and birth time (see zio_write_compress). While we
1550 * need to reset the BP_SET_LSIZE() call that happened
1551 * in dmu_sync_ready for old style holes, we do *not*
1552 * want to wipe out the information contained in new
1553 * style holes. Thus, only zero out the block pointer if
1554 * it's an old style hole.
1555 */
1556 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) &&
1557 dr->dt.dl.dr_overridden_by.blk_birth == 0)
1558 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1559 } else {
1560 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1561 }
1562 cv_broadcast(&db->db_changed);
1563 mutex_exit(&db->db_mtx);
1564
1565 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1566
1567 kmem_free(dsa, sizeof (*dsa));
1568 }
1569
1570 static void
1571 dmu_sync_late_arrival_done(zio_t *zio)
1572 {
1573 blkptr_t *bp = zio->io_bp;
1574 dmu_sync_arg_t *dsa = zio->io_private;
1575 ASSERTV(blkptr_t *bp_orig = &zio->io_bp_orig);
1576
1577 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1578 /*
1579 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE)
1580 * then there is nothing to do here. Otherwise, free the
1581 * newly allocated block in this txg.
1582 */
1583 if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
1584 ASSERT(BP_EQUAL(bp, bp_orig));
1585 } else {
1586 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1587 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1588 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1589 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1590 }
1591 }
1592
1593 dmu_tx_commit(dsa->dsa_tx);
1594
1595 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1596
1597 kmem_free(dsa, sizeof (*dsa));
1598 }
1599
1600 static int
1601 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1602 zio_prop_t *zp, zbookmark_phys_t *zb)
1603 {
1604 dmu_sync_arg_t *dsa;
1605 dmu_tx_t *tx;
1606
1607 tx = dmu_tx_create(os);
1608 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1609 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1610 dmu_tx_abort(tx);
1611 /* Make zl_get_data do txg_waited_synced() */
1612 return (SET_ERROR(EIO));
1613 }
1614
1615 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1616 dsa->dsa_dr = NULL;
1617 dsa->dsa_done = done;
1618 dsa->dsa_zgd = zgd;
1619 dsa->dsa_tx = tx;
1620
1621 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1622 zgd->zgd_db->db_data, zgd->zgd_db->db_size, zp,
1623 dmu_sync_late_arrival_ready, NULL, dmu_sync_late_arrival_done, dsa,
1624 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL|ZIO_FLAG_FASTWRITE, zb));
1625
1626 return (0);
1627 }
1628
1629 /*
1630 * Intent log support: sync the block associated with db to disk.
1631 * N.B. and XXX: the caller is responsible for making sure that the
1632 * data isn't changing while dmu_sync() is writing it.
1633 *
1634 * Return values:
1635 *
1636 * EEXIST: this txg has already been synced, so there's nothing to do.
1637 * The caller should not log the write.
1638 *
1639 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1640 * The caller should not log the write.
1641 *
1642 * EALREADY: this block is already in the process of being synced.
1643 * The caller should track its progress (somehow).
1644 *
1645 * EIO: could not do the I/O.
1646 * The caller should do a txg_wait_synced().
1647 *
1648 * 0: the I/O has been initiated.
1649 * The caller should log this blkptr in the done callback.
1650 * It is possible that the I/O will fail, in which case
1651 * the error will be reported to the done callback and
1652 * propagated to pio from zio_done().
1653 */
1654 int
1655 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1656 {
1657 blkptr_t *bp = zgd->zgd_bp;
1658 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1659 objset_t *os = db->db_objset;
1660 dsl_dataset_t *ds = os->os_dsl_dataset;
1661 dbuf_dirty_record_t *dr;
1662 dmu_sync_arg_t *dsa;
1663 zbookmark_phys_t zb;
1664 zio_prop_t zp;
1665 dnode_t *dn;
1666
1667 ASSERT(pio != NULL);
1668 ASSERT(txg != 0);
1669
1670 SET_BOOKMARK(&zb, ds->ds_object,
1671 db->db.db_object, db->db_level, db->db_blkid);
1672
1673 DB_DNODE_ENTER(db);
1674 dn = DB_DNODE(db);
1675 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1676 DB_DNODE_EXIT(db);
1677
1678 /*
1679 * If we're frozen (running ziltest), we always need to generate a bp.
1680 */
1681 if (txg > spa_freeze_txg(os->os_spa))
1682 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1683
1684 /*
1685 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1686 * and us. If we determine that this txg is not yet syncing,
1687 * but it begins to sync a moment later, that's OK because the
1688 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1689 */
1690 mutex_enter(&db->db_mtx);
1691
1692 if (txg <= spa_last_synced_txg(os->os_spa)) {
1693 /*
1694 * This txg has already synced. There's nothing to do.
1695 */
1696 mutex_exit(&db->db_mtx);
1697 return (SET_ERROR(EEXIST));
1698 }
1699
1700 if (txg <= spa_syncing_txg(os->os_spa)) {
1701 /*
1702 * This txg is currently syncing, so we can't mess with
1703 * the dirty record anymore; just write a new log block.
1704 */
1705 mutex_exit(&db->db_mtx);
1706 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1707 }
1708
1709 dr = db->db_last_dirty;
1710 while (dr && dr->dr_txg != txg)
1711 dr = dr->dr_next;
1712
1713 if (dr == NULL) {
1714 /*
1715 * There's no dr for this dbuf, so it must have been freed.
1716 * There's no need to log writes to freed blocks, so we're done.
1717 */
1718 mutex_exit(&db->db_mtx);
1719 return (SET_ERROR(ENOENT));
1720 }
1721
1722 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
1723
1724 /*
1725 * Assume the on-disk data is X, the current syncing data (in
1726 * txg - 1) is Y, and the current in-memory data is Z (currently
1727 * in dmu_sync).
1728 *
1729 * We usually want to perform a nopwrite if X and Z are the
1730 * same. However, if Y is different (i.e. the BP is going to
1731 * change before this write takes effect), then a nopwrite will
1732 * be incorrect - we would override with X, which could have
1733 * been freed when Y was written.
1734 *
1735 * (Note that this is not a concern when we are nop-writing from
1736 * syncing context, because X and Y must be identical, because
1737 * all previous txgs have been synced.)
1738 *
1739 * Therefore, we disable nopwrite if the current BP could change
1740 * before this TXG. There are two ways it could change: by
1741 * being dirty (dr_next is non-NULL), or by being freed
1742 * (dnode_block_freed()). This behavior is verified by
1743 * zio_done(), which VERIFYs that the override BP is identical
1744 * to the on-disk BP.
1745 */
1746 DB_DNODE_ENTER(db);
1747 dn = DB_DNODE(db);
1748 if (dr->dr_next != NULL || dnode_block_freed(dn, db->db_blkid))
1749 zp.zp_nopwrite = B_FALSE;
1750 DB_DNODE_EXIT(db);
1751
1752 ASSERT(dr->dr_txg == txg);
1753 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1754 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1755 /*
1756 * We have already issued a sync write for this buffer,
1757 * or this buffer has already been synced. It could not
1758 * have been dirtied since, or we would have cleared the state.
1759 */
1760 mutex_exit(&db->db_mtx);
1761 return (SET_ERROR(EALREADY));
1762 }
1763
1764 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1765 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1766 mutex_exit(&db->db_mtx);
1767
1768 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1769 dsa->dsa_dr = dr;
1770 dsa->dsa_done = done;
1771 dsa->dsa_zgd = zgd;
1772 dsa->dsa_tx = NULL;
1773
1774 zio_nowait(arc_write(pio, os->os_spa, txg,
1775 bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
1776 DBUF_IS_L2COMPRESSIBLE(db), &zp, dmu_sync_ready,
1777 NULL, dmu_sync_done, dsa, ZIO_PRIORITY_SYNC_WRITE,
1778 ZIO_FLAG_CANFAIL, &zb));
1779
1780 return (0);
1781 }
1782
1783 int
1784 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1785 dmu_tx_t *tx)
1786 {
1787 dnode_t *dn;
1788 int err;
1789
1790 err = dnode_hold(os, object, FTAG, &dn);
1791 if (err)
1792 return (err);
1793 err = dnode_set_blksz(dn, size, ibs, tx);
1794 dnode_rele(dn, FTAG);
1795 return (err);
1796 }
1797
1798 void
1799 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1800 dmu_tx_t *tx)
1801 {
1802 dnode_t *dn;
1803
1804 /*
1805 * Send streams include each object's checksum function. This
1806 * check ensures that the receiving system can understand the
1807 * checksum function transmitted.
1808 */
1809 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS);
1810
1811 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1812 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS);
1813 dn->dn_checksum = checksum;
1814 dnode_setdirty(dn, tx);
1815 dnode_rele(dn, FTAG);
1816 }
1817
1818 void
1819 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1820 dmu_tx_t *tx)
1821 {
1822 dnode_t *dn;
1823
1824 /*
1825 * Send streams include each object's compression function. This
1826 * check ensures that the receiving system can understand the
1827 * compression function transmitted.
1828 */
1829 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS);
1830
1831 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1832 dn->dn_compress = compress;
1833 dnode_setdirty(dn, tx);
1834 dnode_rele(dn, FTAG);
1835 }
1836
1837 int zfs_mdcomp_disable = 0;
1838
1839 /*
1840 * When the "redundant_metadata" property is set to "most", only indirect
1841 * blocks of this level and higher will have an additional ditto block.
1842 */
1843 int zfs_redundant_metadata_most_ditto_level = 2;
1844
1845 void
1846 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
1847 {
1848 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
1849 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
1850 (wp & WP_SPILL));
1851 enum zio_checksum checksum = os->os_checksum;
1852 enum zio_compress compress = os->os_compress;
1853 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
1854 boolean_t dedup = B_FALSE;
1855 boolean_t nopwrite = B_FALSE;
1856 boolean_t dedup_verify = os->os_dedup_verify;
1857 int copies = os->os_copies;
1858
1859 /*
1860 * We maintain different write policies for each of the following
1861 * types of data:
1862 * 1. metadata
1863 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1864 * 3. all other level 0 blocks
1865 */
1866 if (ismd) {
1867 if (zfs_mdcomp_disable) {
1868 compress = ZIO_COMPRESS_EMPTY;
1869 } else {
1870 /*
1871 * XXX -- we should design a compression algorithm
1872 * that specializes in arrays of bps.
1873 */
1874 compress = zio_compress_select(os->os_spa,
1875 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON);
1876 }
1877
1878 /*
1879 * Metadata always gets checksummed. If the data
1880 * checksum is multi-bit correctable, and it's not a
1881 * ZBT-style checksum, then it's suitable for metadata
1882 * as well. Otherwise, the metadata checksum defaults
1883 * to fletcher4.
1884 */
1885 if (zio_checksum_table[checksum].ci_correctable < 1 ||
1886 zio_checksum_table[checksum].ci_eck)
1887 checksum = ZIO_CHECKSUM_FLETCHER_4;
1888
1889 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL ||
1890 (os->os_redundant_metadata ==
1891 ZFS_REDUNDANT_METADATA_MOST &&
1892 (level >= zfs_redundant_metadata_most_ditto_level ||
1893 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))))
1894 copies++;
1895 } else if (wp & WP_NOFILL) {
1896 ASSERT(level == 0);
1897
1898 /*
1899 * If we're writing preallocated blocks, we aren't actually
1900 * writing them so don't set any policy properties. These
1901 * blocks are currently only used by an external subsystem
1902 * outside of zfs (i.e. dump) and not written by the zio
1903 * pipeline.
1904 */
1905 compress = ZIO_COMPRESS_OFF;
1906 checksum = ZIO_CHECKSUM_OFF;
1907 } else {
1908 compress = zio_compress_select(os->os_spa, dn->dn_compress,
1909 compress);
1910
1911 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
1912 zio_checksum_select(dn->dn_checksum, checksum) :
1913 dedup_checksum;
1914
1915 /*
1916 * Determine dedup setting. If we are in dmu_sync(),
1917 * we won't actually dedup now because that's all
1918 * done in syncing context; but we do want to use the
1919 * dedup checkum. If the checksum is not strong
1920 * enough to ensure unique signatures, force
1921 * dedup_verify.
1922 */
1923 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
1924 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
1925 if (!zio_checksum_table[checksum].ci_dedup)
1926 dedup_verify = B_TRUE;
1927 }
1928
1929 /*
1930 * Enable nopwrite if we have a cryptographically secure
1931 * checksum that has no known collisions (i.e. SHA-256)
1932 * and compression is enabled. We don't enable nopwrite if
1933 * dedup is enabled as the two features are mutually exclusive.
1934 */
1935 nopwrite = (!dedup && zio_checksum_table[checksum].ci_dedup &&
1936 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
1937 }
1938
1939 zp->zp_checksum = checksum;
1940 zp->zp_compress = compress;
1941 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
1942 zp->zp_level = level;
1943 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
1944 zp->zp_dedup = dedup;
1945 zp->zp_dedup_verify = dedup && dedup_verify;
1946 zp->zp_nopwrite = nopwrite;
1947 }
1948
1949 int
1950 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
1951 {
1952 dnode_t *dn;
1953 int i, err;
1954
1955 err = dnode_hold(os, object, FTAG, &dn);
1956 if (err)
1957 return (err);
1958 /*
1959 * Sync any current changes before
1960 * we go trundling through the block pointers.
1961 */
1962 for (i = 0; i < TXG_SIZE; i++) {
1963 if (list_link_active(&dn->dn_dirty_link[i]))
1964 break;
1965 }
1966 if (i != TXG_SIZE) {
1967 dnode_rele(dn, FTAG);
1968 txg_wait_synced(dmu_objset_pool(os), 0);
1969 err = dnode_hold(os, object, FTAG, &dn);
1970 if (err)
1971 return (err);
1972 }
1973
1974 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
1975 dnode_rele(dn, FTAG);
1976
1977 return (err);
1978 }
1979
1980 void
1981 __dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
1982 {
1983 dnode_phys_t *dnp = dn->dn_phys;
1984 int i;
1985
1986 doi->doi_data_block_size = dn->dn_datablksz;
1987 doi->doi_metadata_block_size = dn->dn_indblkshift ?
1988 1ULL << dn->dn_indblkshift : 0;
1989 doi->doi_type = dn->dn_type;
1990 doi->doi_bonus_type = dn->dn_bonustype;
1991 doi->doi_bonus_size = dn->dn_bonuslen;
1992 doi->doi_indirection = dn->dn_nlevels;
1993 doi->doi_checksum = dn->dn_checksum;
1994 doi->doi_compress = dn->dn_compress;
1995 doi->doi_nblkptr = dn->dn_nblkptr;
1996 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
1997 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
1998 doi->doi_fill_count = 0;
1999 for (i = 0; i < dnp->dn_nblkptr; i++)
2000 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
2001 }
2002
2003 void
2004 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2005 {
2006 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2007 mutex_enter(&dn->dn_mtx);
2008
2009 __dmu_object_info_from_dnode(dn, doi);
2010
2011 mutex_exit(&dn->dn_mtx);
2012 rw_exit(&dn->dn_struct_rwlock);
2013 }
2014
2015 /*
2016 * Get information on a DMU object.
2017 * If doi is NULL, just indicates whether the object exists.
2018 */
2019 int
2020 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
2021 {
2022 dnode_t *dn;
2023 int err = dnode_hold(os, object, FTAG, &dn);
2024
2025 if (err)
2026 return (err);
2027
2028 if (doi != NULL)
2029 dmu_object_info_from_dnode(dn, doi);
2030
2031 dnode_rele(dn, FTAG);
2032 return (0);
2033 }
2034
2035 /*
2036 * As above, but faster; can be used when you have a held dbuf in hand.
2037 */
2038 void
2039 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
2040 {
2041 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2042
2043 DB_DNODE_ENTER(db);
2044 dmu_object_info_from_dnode(DB_DNODE(db), doi);
2045 DB_DNODE_EXIT(db);
2046 }
2047
2048 /*
2049 * Faster still when you only care about the size.
2050 * This is specifically optimized for zfs_getattr().
2051 */
2052 void
2053 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
2054 u_longlong_t *nblk512)
2055 {
2056 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2057 dnode_t *dn;
2058
2059 DB_DNODE_ENTER(db);
2060 dn = DB_DNODE(db);
2061
2062 *blksize = dn->dn_datablksz;
2063 /* add 1 for dnode space */
2064 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
2065 SPA_MINBLOCKSHIFT) + 1;
2066 DB_DNODE_EXIT(db);
2067 }
2068
2069 void
2070 byteswap_uint64_array(void *vbuf, size_t size)
2071 {
2072 uint64_t *buf = vbuf;
2073 size_t count = size >> 3;
2074 int i;
2075
2076 ASSERT((size & 7) == 0);
2077
2078 for (i = 0; i < count; i++)
2079 buf[i] = BSWAP_64(buf[i]);
2080 }
2081
2082 void
2083 byteswap_uint32_array(void *vbuf, size_t size)
2084 {
2085 uint32_t *buf = vbuf;
2086 size_t count = size >> 2;
2087 int i;
2088
2089 ASSERT((size & 3) == 0);
2090
2091 for (i = 0; i < count; i++)
2092 buf[i] = BSWAP_32(buf[i]);
2093 }
2094
2095 void
2096 byteswap_uint16_array(void *vbuf, size_t size)
2097 {
2098 uint16_t *buf = vbuf;
2099 size_t count = size >> 1;
2100 int i;
2101
2102 ASSERT((size & 1) == 0);
2103
2104 for (i = 0; i < count; i++)
2105 buf[i] = BSWAP_16(buf[i]);
2106 }
2107
2108 /* ARGSUSED */
2109 void
2110 byteswap_uint8_array(void *vbuf, size_t size)
2111 {
2112 }
2113
2114 void
2115 dmu_init(void)
2116 {
2117 zfs_dbgmsg_init();
2118 sa_cache_init();
2119 xuio_stat_init();
2120 dmu_objset_init();
2121 dnode_init();
2122 dbuf_init();
2123 zfetch_init();
2124 dmu_tx_init();
2125 l2arc_init();
2126 arc_init();
2127 }
2128
2129 void
2130 dmu_fini(void)
2131 {
2132 arc_fini(); /* arc depends on l2arc, so arc must go first */
2133 l2arc_fini();
2134 dmu_tx_fini();
2135 zfetch_fini();
2136 dbuf_fini();
2137 dnode_fini();
2138 dmu_objset_fini();
2139 xuio_stat_fini();
2140 sa_cache_fini();
2141 zfs_dbgmsg_fini();
2142 }
2143
2144 #if defined(_KERNEL) && defined(HAVE_SPL)
2145 EXPORT_SYMBOL(dmu_bonus_hold);
2146 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus);
2147 EXPORT_SYMBOL(dmu_buf_rele_array);
2148 EXPORT_SYMBOL(dmu_prefetch);
2149 EXPORT_SYMBOL(dmu_free_range);
2150 EXPORT_SYMBOL(dmu_free_long_range);
2151 EXPORT_SYMBOL(dmu_free_long_object);
2152 EXPORT_SYMBOL(dmu_read);
2153 EXPORT_SYMBOL(dmu_write);
2154 EXPORT_SYMBOL(dmu_prealloc);
2155 EXPORT_SYMBOL(dmu_object_info);
2156 EXPORT_SYMBOL(dmu_object_info_from_dnode);
2157 EXPORT_SYMBOL(dmu_object_info_from_db);
2158 EXPORT_SYMBOL(dmu_object_size_from_db);
2159 EXPORT_SYMBOL(dmu_object_set_blocksize);
2160 EXPORT_SYMBOL(dmu_object_set_checksum);
2161 EXPORT_SYMBOL(dmu_object_set_compress);
2162 EXPORT_SYMBOL(dmu_write_policy);
2163 EXPORT_SYMBOL(dmu_sync);
2164 EXPORT_SYMBOL(dmu_request_arcbuf);
2165 EXPORT_SYMBOL(dmu_return_arcbuf);
2166 EXPORT_SYMBOL(dmu_assign_arcbuf);
2167 EXPORT_SYMBOL(dmu_buf_hold);
2168 EXPORT_SYMBOL(dmu_ot);
2169
2170 module_param(zfs_mdcomp_disable, int, 0644);
2171 MODULE_PARM_DESC(zfs_mdcomp_disable, "Disable meta data compression");
2172
2173 module_param(zfs_nopwrite_enabled, int, 0644);
2174 MODULE_PARM_DESC(zfs_nopwrite_enabled, "Enable NOP writes");
2175
2176 #endif
ZFS On Linux mirror for PVE