]> 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
Make zvol operations use _by_dnode routines
[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, 2017 by Delphix. All rights reserved.
24 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
25 * Copyright (c) 2013, Joyent, Inc. All rights reserved.
26 * Copyright (c) 2016, Nexenta Systems, Inc. All rights reserved.
27 * Copyright (c) 2015 by Chunwei Chen. All rights reserved.
28 */
29
30 #include <sys/dmu.h>
31 #include <sys/dmu_impl.h>
32 #include <sys/dmu_tx.h>
33 #include <sys/dbuf.h>
34 #include <sys/dnode.h>
35 #include <sys/zfs_context.h>
36 #include <sys/dmu_objset.h>
37 #include <sys/dmu_traverse.h>
38 #include <sys/dsl_dataset.h>
39 #include <sys/dsl_dir.h>
40 #include <sys/dsl_pool.h>
41 #include <sys/dsl_synctask.h>
42 #include <sys/dsl_prop.h>
43 #include <sys/dmu_zfetch.h>
44 #include <sys/zfs_ioctl.h>
45 #include <sys/zap.h>
46 #include <sys/zio_checksum.h>
47 #include <sys/zio_compress.h>
48 #include <sys/sa.h>
49 #include <sys/zfeature.h>
50 #include <sys/abd.h>
51 #include <sys/trace_dmu.h>
52 #ifdef _KERNEL
53 #include <sys/vmsystm.h>
54 #include <sys/zfs_znode.h>
55 #endif
56
57 /*
58 * Enable/disable nopwrite feature.
59 */
60 int zfs_nopwrite_enabled = 1;
61
62 /*
63 * Tunable to control percentage of dirtied blocks from frees in one TXG.
64 * After this threshold is crossed, additional dirty blocks from frees
65 * wait until the next TXG.
66 * A value of zero will disable this throttle.
67 */
68 unsigned long zfs_per_txg_dirty_frees_percent = 30;
69
70 /*
71 * Enable/disable forcing txg sync when dirty in dmu_offset_next.
72 */
73 int zfs_dmu_offset_next_sync = 0;
74
75 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
76 { DMU_BSWAP_UINT8, TRUE, "unallocated" },
77 { DMU_BSWAP_ZAP, TRUE, "object directory" },
78 { DMU_BSWAP_UINT64, TRUE, "object array" },
79 { DMU_BSWAP_UINT8, TRUE, "packed nvlist" },
80 { DMU_BSWAP_UINT64, TRUE, "packed nvlist size" },
81 { DMU_BSWAP_UINT64, TRUE, "bpobj" },
82 { DMU_BSWAP_UINT64, TRUE, "bpobj header" },
83 { DMU_BSWAP_UINT64, TRUE, "SPA space map header" },
84 { DMU_BSWAP_UINT64, TRUE, "SPA space map" },
85 { DMU_BSWAP_UINT64, TRUE, "ZIL intent log" },
86 { DMU_BSWAP_DNODE, TRUE, "DMU dnode" },
87 { DMU_BSWAP_OBJSET, TRUE, "DMU objset" },
88 { DMU_BSWAP_UINT64, TRUE, "DSL directory" },
89 { DMU_BSWAP_ZAP, TRUE, "DSL directory child map"},
90 { DMU_BSWAP_ZAP, TRUE, "DSL dataset snap map" },
91 { DMU_BSWAP_ZAP, TRUE, "DSL props" },
92 { DMU_BSWAP_UINT64, TRUE, "DSL dataset" },
93 { DMU_BSWAP_ZNODE, TRUE, "ZFS znode" },
94 { DMU_BSWAP_OLDACL, TRUE, "ZFS V0 ACL" },
95 { DMU_BSWAP_UINT8, FALSE, "ZFS plain file" },
96 { DMU_BSWAP_ZAP, TRUE, "ZFS directory" },
97 { DMU_BSWAP_ZAP, TRUE, "ZFS master node" },
98 { DMU_BSWAP_ZAP, TRUE, "ZFS delete queue" },
99 { DMU_BSWAP_UINT8, FALSE, "zvol object" },
100 { DMU_BSWAP_ZAP, TRUE, "zvol prop" },
101 { DMU_BSWAP_UINT8, FALSE, "other uint8[]" },
102 { DMU_BSWAP_UINT64, FALSE, "other uint64[]" },
103 { DMU_BSWAP_ZAP, TRUE, "other ZAP" },
104 { DMU_BSWAP_ZAP, TRUE, "persistent error log" },
105 { DMU_BSWAP_UINT8, TRUE, "SPA history" },
106 { DMU_BSWAP_UINT64, TRUE, "SPA history offsets" },
107 { DMU_BSWAP_ZAP, TRUE, "Pool properties" },
108 { DMU_BSWAP_ZAP, TRUE, "DSL permissions" },
109 { DMU_BSWAP_ACL, TRUE, "ZFS ACL" },
110 { DMU_BSWAP_UINT8, TRUE, "ZFS SYSACL" },
111 { DMU_BSWAP_UINT8, TRUE, "FUID table" },
112 { DMU_BSWAP_UINT64, TRUE, "FUID table size" },
113 { DMU_BSWAP_ZAP, TRUE, "DSL dataset next clones"},
114 { DMU_BSWAP_ZAP, TRUE, "scan work queue" },
115 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group used" },
116 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group quota" },
117 { DMU_BSWAP_ZAP, TRUE, "snapshot refcount tags"},
118 { DMU_BSWAP_ZAP, TRUE, "DDT ZAP algorithm" },
119 { DMU_BSWAP_ZAP, TRUE, "DDT statistics" },
120 { DMU_BSWAP_UINT8, TRUE, "System attributes" },
121 { DMU_BSWAP_ZAP, TRUE, "SA master node" },
122 { DMU_BSWAP_ZAP, TRUE, "SA attr registration" },
123 { DMU_BSWAP_ZAP, TRUE, "SA attr layouts" },
124 { DMU_BSWAP_ZAP, TRUE, "scan translations" },
125 { DMU_BSWAP_UINT8, FALSE, "deduplicated block" },
126 { DMU_BSWAP_ZAP, TRUE, "DSL deadlist map" },
127 { DMU_BSWAP_UINT64, TRUE, "DSL deadlist map hdr" },
128 { DMU_BSWAP_ZAP, TRUE, "DSL dir clones" },
129 { DMU_BSWAP_UINT64, TRUE, "bpobj subobj" }
130 };
131
132 const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
133 { byteswap_uint8_array, "uint8" },
134 { byteswap_uint16_array, "uint16" },
135 { byteswap_uint32_array, "uint32" },
136 { byteswap_uint64_array, "uint64" },
137 { zap_byteswap, "zap" },
138 { dnode_buf_byteswap, "dnode" },
139 { dmu_objset_byteswap, "objset" },
140 { zfs_znode_byteswap, "znode" },
141 { zfs_oldacl_byteswap, "oldacl" },
142 { zfs_acl_byteswap, "acl" }
143 };
144
145 int
146 dmu_buf_hold_noread_by_dnode(dnode_t *dn, uint64_t offset,
147 void *tag, dmu_buf_t **dbp)
148 {
149 uint64_t blkid;
150 dmu_buf_impl_t *db;
151
152 blkid = dbuf_whichblock(dn, 0, offset);
153 rw_enter(&dn->dn_struct_rwlock, RW_READER);
154 db = dbuf_hold(dn, blkid, tag);
155 rw_exit(&dn->dn_struct_rwlock);
156
157 if (db == NULL) {
158 *dbp = NULL;
159 return (SET_ERROR(EIO));
160 }
161
162 *dbp = &db->db;
163 return (0);
164 }
165 int
166 dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset,
167 void *tag, dmu_buf_t **dbp)
168 {
169 dnode_t *dn;
170 uint64_t blkid;
171 dmu_buf_impl_t *db;
172 int err;
173
174 err = dnode_hold(os, object, FTAG, &dn);
175 if (err)
176 return (err);
177 blkid = dbuf_whichblock(dn, 0, offset);
178 rw_enter(&dn->dn_struct_rwlock, RW_READER);
179 db = dbuf_hold(dn, blkid, tag);
180 rw_exit(&dn->dn_struct_rwlock);
181 dnode_rele(dn, FTAG);
182
183 if (db == NULL) {
184 *dbp = NULL;
185 return (SET_ERROR(EIO));
186 }
187
188 *dbp = &db->db;
189 return (err);
190 }
191
192 int
193 dmu_buf_hold_by_dnode(dnode_t *dn, uint64_t offset,
194 void *tag, dmu_buf_t **dbp, int flags)
195 {
196 int err;
197 int db_flags = DB_RF_CANFAIL;
198
199 if (flags & DMU_READ_NO_PREFETCH)
200 db_flags |= DB_RF_NOPREFETCH;
201
202 err = dmu_buf_hold_noread_by_dnode(dn, offset, tag, dbp);
203 if (err == 0) {
204 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
205 err = dbuf_read(db, NULL, db_flags);
206 if (err != 0) {
207 dbuf_rele(db, tag);
208 *dbp = NULL;
209 }
210 }
211
212 return (err);
213 }
214
215 int
216 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
217 void *tag, dmu_buf_t **dbp, int flags)
218 {
219 int err;
220 int db_flags = DB_RF_CANFAIL;
221
222 if (flags & DMU_READ_NO_PREFETCH)
223 db_flags |= DB_RF_NOPREFETCH;
224
225 err = dmu_buf_hold_noread(os, object, offset, tag, dbp);
226 if (err == 0) {
227 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
228 err = dbuf_read(db, NULL, db_flags);
229 if (err != 0) {
230 dbuf_rele(db, tag);
231 *dbp = NULL;
232 }
233 }
234
235 return (err);
236 }
237
238 int
239 dmu_bonus_max(void)
240 {
241 return (DN_OLD_MAX_BONUSLEN);
242 }
243
244 int
245 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
246 {
247 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
248 dnode_t *dn;
249 int error;
250
251 DB_DNODE_ENTER(db);
252 dn = DB_DNODE(db);
253
254 if (dn->dn_bonus != db) {
255 error = SET_ERROR(EINVAL);
256 } else if (newsize < 0 || newsize > db_fake->db_size) {
257 error = SET_ERROR(EINVAL);
258 } else {
259 dnode_setbonuslen(dn, newsize, tx);
260 error = 0;
261 }
262
263 DB_DNODE_EXIT(db);
264 return (error);
265 }
266
267 int
268 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
269 {
270 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
271 dnode_t *dn;
272 int error;
273
274 DB_DNODE_ENTER(db);
275 dn = DB_DNODE(db);
276
277 if (!DMU_OT_IS_VALID(type)) {
278 error = SET_ERROR(EINVAL);
279 } else if (dn->dn_bonus != db) {
280 error = SET_ERROR(EINVAL);
281 } else {
282 dnode_setbonus_type(dn, type, tx);
283 error = 0;
284 }
285
286 DB_DNODE_EXIT(db);
287 return (error);
288 }
289
290 dmu_object_type_t
291 dmu_get_bonustype(dmu_buf_t *db_fake)
292 {
293 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
294 dnode_t *dn;
295 dmu_object_type_t type;
296
297 DB_DNODE_ENTER(db);
298 dn = DB_DNODE(db);
299 type = dn->dn_bonustype;
300 DB_DNODE_EXIT(db);
301
302 return (type);
303 }
304
305 int
306 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
307 {
308 dnode_t *dn;
309 int error;
310
311 error = dnode_hold(os, object, FTAG, &dn);
312 dbuf_rm_spill(dn, tx);
313 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
314 dnode_rm_spill(dn, tx);
315 rw_exit(&dn->dn_struct_rwlock);
316 dnode_rele(dn, FTAG);
317 return (error);
318 }
319
320 /*
321 * returns ENOENT, EIO, or 0.
322 */
323 int
324 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
325 {
326 dnode_t *dn;
327 dmu_buf_impl_t *db;
328 int error;
329
330 error = dnode_hold(os, object, FTAG, &dn);
331 if (error)
332 return (error);
333
334 rw_enter(&dn->dn_struct_rwlock, RW_READER);
335 if (dn->dn_bonus == NULL) {
336 rw_exit(&dn->dn_struct_rwlock);
337 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
338 if (dn->dn_bonus == NULL)
339 dbuf_create_bonus(dn);
340 }
341 db = dn->dn_bonus;
342
343 /* as long as the bonus buf is held, the dnode will be held */
344 if (refcount_add(&db->db_holds, tag) == 1) {
345 VERIFY(dnode_add_ref(dn, db));
346 atomic_inc_32(&dn->dn_dbufs_count);
347 }
348
349 /*
350 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
351 * hold and incrementing the dbuf count to ensure that dnode_move() sees
352 * a dnode hold for every dbuf.
353 */
354 rw_exit(&dn->dn_struct_rwlock);
355
356 dnode_rele(dn, FTAG);
357
358 VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH));
359
360 *dbp = &db->db;
361 return (0);
362 }
363
364 /*
365 * returns ENOENT, EIO, or 0.
366 *
367 * This interface will allocate a blank spill dbuf when a spill blk
368 * doesn't already exist on the dnode.
369 *
370 * if you only want to find an already existing spill db, then
371 * dmu_spill_hold_existing() should be used.
372 */
373 int
374 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
375 {
376 dmu_buf_impl_t *db = NULL;
377 int err;
378
379 if ((flags & DB_RF_HAVESTRUCT) == 0)
380 rw_enter(&dn->dn_struct_rwlock, RW_READER);
381
382 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
383
384 if ((flags & DB_RF_HAVESTRUCT) == 0)
385 rw_exit(&dn->dn_struct_rwlock);
386
387 if (db == NULL) {
388 *dbp = NULL;
389 return (SET_ERROR(EIO));
390 }
391 err = dbuf_read(db, NULL, flags);
392 if (err == 0)
393 *dbp = &db->db;
394 else {
395 dbuf_rele(db, tag);
396 *dbp = NULL;
397 }
398 return (err);
399 }
400
401 int
402 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
403 {
404 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
405 dnode_t *dn;
406 int err;
407
408 DB_DNODE_ENTER(db);
409 dn = DB_DNODE(db);
410
411 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
412 err = SET_ERROR(EINVAL);
413 } else {
414 rw_enter(&dn->dn_struct_rwlock, RW_READER);
415
416 if (!dn->dn_have_spill) {
417 err = SET_ERROR(ENOENT);
418 } else {
419 err = dmu_spill_hold_by_dnode(dn,
420 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
421 }
422
423 rw_exit(&dn->dn_struct_rwlock);
424 }
425
426 DB_DNODE_EXIT(db);
427 return (err);
428 }
429
430 int
431 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
432 {
433 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
434 dnode_t *dn;
435 int err;
436
437 DB_DNODE_ENTER(db);
438 dn = DB_DNODE(db);
439 err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp);
440 DB_DNODE_EXIT(db);
441
442 return (err);
443 }
444
445 /*
446 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
447 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
448 * and can induce severe lock contention when writing to several files
449 * whose dnodes are in the same block.
450 */
451 static int
452 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
453 boolean_t read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
454 {
455 dmu_buf_t **dbp;
456 uint64_t blkid, nblks, i;
457 uint32_t dbuf_flags;
458 int err;
459 zio_t *zio;
460
461 ASSERT(length <= DMU_MAX_ACCESS);
462
463 /*
464 * Note: We directly notify the prefetch code of this read, so that
465 * we can tell it about the multi-block read. dbuf_read() only knows
466 * about the one block it is accessing.
467 */
468 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT |
469 DB_RF_NOPREFETCH;
470
471 rw_enter(&dn->dn_struct_rwlock, RW_READER);
472 if (dn->dn_datablkshift) {
473 int blkshift = dn->dn_datablkshift;
474 nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) -
475 P2ALIGN(offset, 1ULL << blkshift)) >> blkshift;
476 } else {
477 if (offset + length > dn->dn_datablksz) {
478 zfs_panic_recover("zfs: accessing past end of object "
479 "%llx/%llx (size=%u access=%llu+%llu)",
480 (longlong_t)dn->dn_objset->
481 os_dsl_dataset->ds_object,
482 (longlong_t)dn->dn_object, dn->dn_datablksz,
483 (longlong_t)offset, (longlong_t)length);
484 rw_exit(&dn->dn_struct_rwlock);
485 return (SET_ERROR(EIO));
486 }
487 nblks = 1;
488 }
489 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
490
491 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
492 blkid = dbuf_whichblock(dn, 0, offset);
493 for (i = 0; i < nblks; i++) {
494 dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag);
495 if (db == NULL) {
496 rw_exit(&dn->dn_struct_rwlock);
497 dmu_buf_rele_array(dbp, nblks, tag);
498 zio_nowait(zio);
499 return (SET_ERROR(EIO));
500 }
501
502 /* initiate async i/o */
503 if (read)
504 (void) dbuf_read(db, zio, dbuf_flags);
505 dbp[i] = &db->db;
506 }
507
508 if ((flags & DMU_READ_NO_PREFETCH) == 0 &&
509 DNODE_META_IS_CACHEABLE(dn) && length <= zfetch_array_rd_sz) {
510 dmu_zfetch(&dn->dn_zfetch, blkid, nblks,
511 read && DNODE_IS_CACHEABLE(dn));
512 }
513 rw_exit(&dn->dn_struct_rwlock);
514
515 /* wait for async i/o */
516 err = zio_wait(zio);
517 if (err) {
518 dmu_buf_rele_array(dbp, nblks, tag);
519 return (err);
520 }
521
522 /* wait for other io to complete */
523 if (read) {
524 for (i = 0; i < nblks; i++) {
525 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
526 mutex_enter(&db->db_mtx);
527 while (db->db_state == DB_READ ||
528 db->db_state == DB_FILL)
529 cv_wait(&db->db_changed, &db->db_mtx);
530 if (db->db_state == DB_UNCACHED)
531 err = SET_ERROR(EIO);
532 mutex_exit(&db->db_mtx);
533 if (err) {
534 dmu_buf_rele_array(dbp, nblks, tag);
535 return (err);
536 }
537 }
538 }
539
540 *numbufsp = nblks;
541 *dbpp = dbp;
542 return (0);
543 }
544
545 static int
546 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
547 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
548 {
549 dnode_t *dn;
550 int err;
551
552 err = dnode_hold(os, object, FTAG, &dn);
553 if (err)
554 return (err);
555
556 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
557 numbufsp, dbpp, DMU_READ_PREFETCH);
558
559 dnode_rele(dn, FTAG);
560
561 return (err);
562 }
563
564 int
565 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
566 uint64_t length, boolean_t read, void *tag, int *numbufsp,
567 dmu_buf_t ***dbpp)
568 {
569 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
570 dnode_t *dn;
571 int err;
572
573 DB_DNODE_ENTER(db);
574 dn = DB_DNODE(db);
575 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
576 numbufsp, dbpp, DMU_READ_PREFETCH);
577 DB_DNODE_EXIT(db);
578
579 return (err);
580 }
581
582 void
583 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
584 {
585 int i;
586 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
587
588 if (numbufs == 0)
589 return;
590
591 for (i = 0; i < numbufs; i++) {
592 if (dbp[i])
593 dbuf_rele(dbp[i], tag);
594 }
595
596 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
597 }
598
599 /*
600 * Issue prefetch i/os for the given blocks. If level is greater than 0, the
601 * indirect blocks prefeteched will be those that point to the blocks containing
602 * the data starting at offset, and continuing to offset + len.
603 *
604 * Note that if the indirect blocks above the blocks being prefetched are not in
605 * cache, they will be asychronously read in.
606 */
607 void
608 dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset,
609 uint64_t len, zio_priority_t pri)
610 {
611 dnode_t *dn;
612 uint64_t blkid;
613 int nblks, err;
614
615 if (len == 0) { /* they're interested in the bonus buffer */
616 dn = DMU_META_DNODE(os);
617
618 if (object == 0 || object >= DN_MAX_OBJECT)
619 return;
620
621 rw_enter(&dn->dn_struct_rwlock, RW_READER);
622 blkid = dbuf_whichblock(dn, level,
623 object * sizeof (dnode_phys_t));
624 dbuf_prefetch(dn, level, blkid, pri, 0);
625 rw_exit(&dn->dn_struct_rwlock);
626 return;
627 }
628
629 /*
630 * XXX - Note, if the dnode for the requested object is not
631 * already cached, we will do a *synchronous* read in the
632 * dnode_hold() call. The same is true for any indirects.
633 */
634 err = dnode_hold(os, object, FTAG, &dn);
635 if (err != 0)
636 return;
637
638 rw_enter(&dn->dn_struct_rwlock, RW_READER);
639 /*
640 * offset + len - 1 is the last byte we want to prefetch for, and offset
641 * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the
642 * last block we want to prefetch, and dbuf_whichblock(dn, level,
643 * offset) is the first. Then the number we need to prefetch is the
644 * last - first + 1.
645 */
646 if (level > 0 || dn->dn_datablkshift != 0) {
647 nblks = dbuf_whichblock(dn, level, offset + len - 1) -
648 dbuf_whichblock(dn, level, offset) + 1;
649 } else {
650 nblks = (offset < dn->dn_datablksz);
651 }
652
653 if (nblks != 0) {
654 int i;
655
656 blkid = dbuf_whichblock(dn, level, offset);
657 for (i = 0; i < nblks; i++)
658 dbuf_prefetch(dn, level, blkid + i, pri, 0);
659 }
660
661 rw_exit(&dn->dn_struct_rwlock);
662
663 dnode_rele(dn, FTAG);
664 }
665
666 /*
667 * Get the next "chunk" of file data to free. We traverse the file from
668 * the end so that the file gets shorter over time (if we crashes in the
669 * middle, this will leave us in a better state). We find allocated file
670 * data by simply searching the allocated level 1 indirects.
671 *
672 * On input, *start should be the first offset that does not need to be
673 * freed (e.g. "offset + length"). On return, *start will be the first
674 * offset that should be freed.
675 */
676 static int
677 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum)
678 {
679 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
680 /* bytes of data covered by a level-1 indirect block */
681 uint64_t iblkrange =
682 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
683 uint64_t blks;
684
685 ASSERT3U(minimum, <=, *start);
686
687 if (*start - minimum <= iblkrange * maxblks) {
688 *start = minimum;
689 return (0);
690 }
691 ASSERT(ISP2(iblkrange));
692
693 for (blks = 0; *start > minimum && blks < maxblks; blks++) {
694 int err;
695
696 /*
697 * dnode_next_offset(BACKWARDS) will find an allocated L1
698 * indirect block at or before the input offset. We must
699 * decrement *start so that it is at the end of the region
700 * to search.
701 */
702 (*start)--;
703 err = dnode_next_offset(dn,
704 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
705
706 /* if there are no indirect blocks before start, we are done */
707 if (err == ESRCH) {
708 *start = minimum;
709 break;
710 } else if (err != 0) {
711 return (err);
712 }
713
714 /* set start to the beginning of this L1 indirect */
715 *start = P2ALIGN(*start, iblkrange);
716 }
717 if (*start < minimum)
718 *start = minimum;
719 return (0);
720 }
721
722 /*
723 * If this objset is of type OST_ZFS return true if vfs's unmounted flag is set,
724 * otherwise return false.
725 * Used below in dmu_free_long_range_impl() to enable abort when unmounting
726 */
727 /*ARGSUSED*/
728 static boolean_t
729 dmu_objset_zfs_unmounting(objset_t *os)
730 {
731 #ifdef _KERNEL
732 if (dmu_objset_type(os) == DMU_OST_ZFS)
733 return (zfs_get_vfs_flag_unmounted(os));
734 #endif
735 return (B_FALSE);
736 }
737
738 static int
739 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
740 uint64_t length)
741 {
742 uint64_t object_size;
743 int err;
744 uint64_t dirty_frees_threshold;
745 dsl_pool_t *dp = dmu_objset_pool(os);
746 int t;
747
748 if (dn == NULL)
749 return (SET_ERROR(EINVAL));
750
751 object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
752 if (offset >= object_size)
753 return (0);
754
755 if (zfs_per_txg_dirty_frees_percent <= 100)
756 dirty_frees_threshold =
757 zfs_per_txg_dirty_frees_percent * zfs_dirty_data_max / 100;
758 else
759 dirty_frees_threshold = zfs_dirty_data_max / 4;
760
761 if (length == DMU_OBJECT_END || offset + length > object_size)
762 length = object_size - offset;
763
764 while (length != 0) {
765 uint64_t chunk_end, chunk_begin, chunk_len;
766 uint64_t long_free_dirty_all_txgs = 0;
767 dmu_tx_t *tx;
768
769 if (dmu_objset_zfs_unmounting(dn->dn_objset))
770 return (SET_ERROR(EINTR));
771
772 chunk_end = chunk_begin = offset + length;
773
774 /* move chunk_begin backwards to the beginning of this chunk */
775 err = get_next_chunk(dn, &chunk_begin, offset);
776 if (err)
777 return (err);
778 ASSERT3U(chunk_begin, >=, offset);
779 ASSERT3U(chunk_begin, <=, chunk_end);
780
781 chunk_len = chunk_end - chunk_begin;
782
783 mutex_enter(&dp->dp_lock);
784 for (t = 0; t < TXG_SIZE; t++) {
785 long_free_dirty_all_txgs +=
786 dp->dp_long_free_dirty_pertxg[t];
787 }
788 mutex_exit(&dp->dp_lock);
789
790 /*
791 * To avoid filling up a TXG with just frees wait for
792 * the next TXG to open before freeing more chunks if
793 * we have reached the threshold of frees
794 */
795 if (dirty_frees_threshold != 0 &&
796 long_free_dirty_all_txgs >= dirty_frees_threshold) {
797 txg_wait_open(dp, 0);
798 continue;
799 }
800
801 tx = dmu_tx_create(os);
802 dmu_tx_hold_free(tx, dn->dn_object, chunk_begin, chunk_len);
803
804 /*
805 * Mark this transaction as typically resulting in a net
806 * reduction in space used.
807 */
808 dmu_tx_mark_netfree(tx);
809 err = dmu_tx_assign(tx, TXG_WAIT);
810 if (err) {
811 dmu_tx_abort(tx);
812 return (err);
813 }
814
815 mutex_enter(&dp->dp_lock);
816 dp->dp_long_free_dirty_pertxg[dmu_tx_get_txg(tx) & TXG_MASK] +=
817 chunk_len;
818 mutex_exit(&dp->dp_lock);
819 DTRACE_PROBE3(free__long__range,
820 uint64_t, long_free_dirty_all_txgs, uint64_t, chunk_len,
821 uint64_t, dmu_tx_get_txg(tx));
822 dnode_free_range(dn, chunk_begin, chunk_len, tx);
823 dmu_tx_commit(tx);
824
825 length -= chunk_len;
826 }
827 return (0);
828 }
829
830 int
831 dmu_free_long_range(objset_t *os, uint64_t object,
832 uint64_t offset, uint64_t length)
833 {
834 dnode_t *dn;
835 int err;
836
837 err = dnode_hold(os, object, FTAG, &dn);
838 if (err != 0)
839 return (err);
840 err = dmu_free_long_range_impl(os, dn, offset, length);
841
842 /*
843 * It is important to zero out the maxblkid when freeing the entire
844 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
845 * will take the fast path, and (b) dnode_reallocate() can verify
846 * that the entire file has been freed.
847 */
848 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
849 dn->dn_maxblkid = 0;
850
851 dnode_rele(dn, FTAG);
852 return (err);
853 }
854
855 int
856 dmu_free_long_object(objset_t *os, uint64_t object)
857 {
858 dmu_tx_t *tx;
859 int err;
860
861 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
862 if (err != 0)
863 return (err);
864
865 tx = dmu_tx_create(os);
866 dmu_tx_hold_bonus(tx, object);
867 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
868 dmu_tx_mark_netfree(tx);
869 err = dmu_tx_assign(tx, TXG_WAIT);
870 if (err == 0) {
871 err = dmu_object_free(os, object, tx);
872 dmu_tx_commit(tx);
873 } else {
874 dmu_tx_abort(tx);
875 }
876
877 return (err);
878 }
879
880 int
881 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
882 uint64_t size, dmu_tx_t *tx)
883 {
884 dnode_t *dn;
885 int err = dnode_hold(os, object, FTAG, &dn);
886 if (err)
887 return (err);
888 ASSERT(offset < UINT64_MAX);
889 ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
890 dnode_free_range(dn, offset, size, tx);
891 dnode_rele(dn, FTAG);
892 return (0);
893 }
894
895 static int
896 dmu_read_impl(dnode_t *dn, uint64_t offset, uint64_t size,
897 void *buf, uint32_t flags)
898 {
899 dmu_buf_t **dbp;
900 int numbufs, err = 0;
901
902 /*
903 * Deal with odd block sizes, where there can't be data past the first
904 * block. If we ever do the tail block optimization, we will need to
905 * handle that here as well.
906 */
907 if (dn->dn_maxblkid == 0) {
908 uint64_t newsz = offset > dn->dn_datablksz ? 0 :
909 MIN(size, dn->dn_datablksz - offset);
910 bzero((char *)buf + newsz, size - newsz);
911 size = newsz;
912 }
913
914 while (size > 0) {
915 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
916 int i;
917
918 /*
919 * NB: we could do this block-at-a-time, but it's nice
920 * to be reading in parallel.
921 */
922 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
923 TRUE, FTAG, &numbufs, &dbp, flags);
924 if (err)
925 break;
926
927 for (i = 0; i < numbufs; i++) {
928 uint64_t tocpy;
929 int64_t bufoff;
930 dmu_buf_t *db = dbp[i];
931
932 ASSERT(size > 0);
933
934 bufoff = offset - db->db_offset;
935 tocpy = MIN(db->db_size - bufoff, size);
936
937 (void) memcpy(buf, (char *)db->db_data + bufoff, tocpy);
938
939 offset += tocpy;
940 size -= tocpy;
941 buf = (char *)buf + tocpy;
942 }
943 dmu_buf_rele_array(dbp, numbufs, FTAG);
944 }
945 return (err);
946 }
947
948 int
949 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
950 void *buf, uint32_t flags)
951 {
952 dnode_t *dn;
953 int err;
954
955 err = dnode_hold(os, object, FTAG, &dn);
956 if (err != 0)
957 return (err);
958
959 err = dmu_read_impl(dn, offset, size, buf, flags);
960 dnode_rele(dn, FTAG);
961 return (err);
962 }
963
964 int
965 dmu_read_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, void *buf,
966 uint32_t flags)
967 {
968 return (dmu_read_impl(dn, offset, size, buf, flags));
969 }
970
971 static void
972 dmu_write_impl(dmu_buf_t **dbp, int numbufs, uint64_t offset, uint64_t size,
973 const void *buf, dmu_tx_t *tx)
974 {
975 int i;
976
977 for (i = 0; i < numbufs; i++) {
978 uint64_t tocpy;
979 int64_t bufoff;
980 dmu_buf_t *db = dbp[i];
981
982 ASSERT(size > 0);
983
984 bufoff = offset - db->db_offset;
985 tocpy = MIN(db->db_size - bufoff, size);
986
987 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
988
989 if (tocpy == db->db_size)
990 dmu_buf_will_fill(db, tx);
991 else
992 dmu_buf_will_dirty(db, tx);
993
994 (void) memcpy((char *)db->db_data + bufoff, buf, tocpy);
995
996 if (tocpy == db->db_size)
997 dmu_buf_fill_done(db, tx);
998
999 offset += tocpy;
1000 size -= tocpy;
1001 buf = (char *)buf + tocpy;
1002 }
1003 }
1004
1005 void
1006 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1007 const void *buf, dmu_tx_t *tx)
1008 {
1009 dmu_buf_t **dbp;
1010 int numbufs;
1011
1012 if (size == 0)
1013 return;
1014
1015 VERIFY0(dmu_buf_hold_array(os, object, offset, size,
1016 FALSE, FTAG, &numbufs, &dbp));
1017 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1018 dmu_buf_rele_array(dbp, numbufs, FTAG);
1019 }
1020
1021 void
1022 dmu_write_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size,
1023 const void *buf, dmu_tx_t *tx)
1024 {
1025 dmu_buf_t **dbp;
1026 int numbufs;
1027
1028 if (size == 0)
1029 return;
1030
1031 VERIFY0(dmu_buf_hold_array_by_dnode(dn, offset, size,
1032 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH));
1033 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1034 dmu_buf_rele_array(dbp, numbufs, FTAG);
1035 }
1036
1037 void
1038 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1039 dmu_tx_t *tx)
1040 {
1041 dmu_buf_t **dbp;
1042 int numbufs, i;
1043
1044 if (size == 0)
1045 return;
1046
1047 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
1048 FALSE, FTAG, &numbufs, &dbp));
1049
1050 for (i = 0; i < numbufs; i++) {
1051 dmu_buf_t *db = dbp[i];
1052
1053 dmu_buf_will_not_fill(db, tx);
1054 }
1055 dmu_buf_rele_array(dbp, numbufs, FTAG);
1056 }
1057
1058 void
1059 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
1060 void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
1061 int compressed_size, int byteorder, dmu_tx_t *tx)
1062 {
1063 dmu_buf_t *db;
1064
1065 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
1066 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
1067 VERIFY0(dmu_buf_hold_noread(os, object, offset,
1068 FTAG, &db));
1069
1070 dmu_buf_write_embedded(db,
1071 data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
1072 uncompressed_size, compressed_size, byteorder, tx);
1073
1074 dmu_buf_rele(db, FTAG);
1075 }
1076
1077 /*
1078 * DMU support for xuio
1079 */
1080 kstat_t *xuio_ksp = NULL;
1081
1082 typedef struct xuio_stats {
1083 /* loaned yet not returned arc_buf */
1084 kstat_named_t xuiostat_onloan_rbuf;
1085 kstat_named_t xuiostat_onloan_wbuf;
1086 /* whether a copy is made when loaning out a read buffer */
1087 kstat_named_t xuiostat_rbuf_copied;
1088 kstat_named_t xuiostat_rbuf_nocopy;
1089 /* whether a copy is made when assigning a write buffer */
1090 kstat_named_t xuiostat_wbuf_copied;
1091 kstat_named_t xuiostat_wbuf_nocopy;
1092 } xuio_stats_t;
1093
1094 static xuio_stats_t xuio_stats = {
1095 { "onloan_read_buf", KSTAT_DATA_UINT64 },
1096 { "onloan_write_buf", KSTAT_DATA_UINT64 },
1097 { "read_buf_copied", KSTAT_DATA_UINT64 },
1098 { "read_buf_nocopy", KSTAT_DATA_UINT64 },
1099 { "write_buf_copied", KSTAT_DATA_UINT64 },
1100 { "write_buf_nocopy", KSTAT_DATA_UINT64 }
1101 };
1102
1103 #define XUIOSTAT_INCR(stat, val) \
1104 atomic_add_64(&xuio_stats.stat.value.ui64, (val))
1105 #define XUIOSTAT_BUMP(stat) XUIOSTAT_INCR(stat, 1)
1106
1107 #ifdef HAVE_UIO_ZEROCOPY
1108 int
1109 dmu_xuio_init(xuio_t *xuio, int nblk)
1110 {
1111 dmu_xuio_t *priv;
1112 uio_t *uio = &xuio->xu_uio;
1113
1114 uio->uio_iovcnt = nblk;
1115 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
1116
1117 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
1118 priv->cnt = nblk;
1119 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
1120 priv->iovp = (iovec_t *)uio->uio_iov;
1121 XUIO_XUZC_PRIV(xuio) = priv;
1122
1123 if (XUIO_XUZC_RW(xuio) == UIO_READ)
1124 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
1125 else
1126 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
1127
1128 return (0);
1129 }
1130
1131 void
1132 dmu_xuio_fini(xuio_t *xuio)
1133 {
1134 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1135 int nblk = priv->cnt;
1136
1137 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
1138 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
1139 kmem_free(priv, sizeof (dmu_xuio_t));
1140
1141 if (XUIO_XUZC_RW(xuio) == UIO_READ)
1142 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
1143 else
1144 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
1145 }
1146
1147 /*
1148 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
1149 * and increase priv->next by 1.
1150 */
1151 int
1152 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
1153 {
1154 struct iovec *iov;
1155 uio_t *uio = &xuio->xu_uio;
1156 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1157 int i = priv->next++;
1158
1159 ASSERT(i < priv->cnt);
1160 ASSERT(off + n <= arc_buf_lsize(abuf));
1161 iov = (iovec_t *)uio->uio_iov + i;
1162 iov->iov_base = (char *)abuf->b_data + off;
1163 iov->iov_len = n;
1164 priv->bufs[i] = abuf;
1165 return (0);
1166 }
1167
1168 int
1169 dmu_xuio_cnt(xuio_t *xuio)
1170 {
1171 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1172 return (priv->cnt);
1173 }
1174
1175 arc_buf_t *
1176 dmu_xuio_arcbuf(xuio_t *xuio, int i)
1177 {
1178 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1179
1180 ASSERT(i < priv->cnt);
1181 return (priv->bufs[i]);
1182 }
1183
1184 void
1185 dmu_xuio_clear(xuio_t *xuio, int i)
1186 {
1187 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1188
1189 ASSERT(i < priv->cnt);
1190 priv->bufs[i] = NULL;
1191 }
1192 #endif /* HAVE_UIO_ZEROCOPY */
1193
1194 static void
1195 xuio_stat_init(void)
1196 {
1197 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
1198 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
1199 KSTAT_FLAG_VIRTUAL);
1200 if (xuio_ksp != NULL) {
1201 xuio_ksp->ks_data = &xuio_stats;
1202 kstat_install(xuio_ksp);
1203 }
1204 }
1205
1206 static void
1207 xuio_stat_fini(void)
1208 {
1209 if (xuio_ksp != NULL) {
1210 kstat_delete(xuio_ksp);
1211 xuio_ksp = NULL;
1212 }
1213 }
1214
1215 void
1216 xuio_stat_wbuf_copied(void)
1217 {
1218 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1219 }
1220
1221 void
1222 xuio_stat_wbuf_nocopy(void)
1223 {
1224 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
1225 }
1226
1227 #ifdef _KERNEL
1228 int
1229 dmu_read_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size)
1230 {
1231 dmu_buf_t **dbp;
1232 int numbufs, i, err;
1233 #ifdef HAVE_UIO_ZEROCOPY
1234 xuio_t *xuio = NULL;
1235 #endif
1236
1237 /*
1238 * NB: we could do this block-at-a-time, but it's nice
1239 * to be reading in parallel.
1240 */
1241 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1242 TRUE, FTAG, &numbufs, &dbp, 0);
1243 if (err)
1244 return (err);
1245
1246 for (i = 0; i < numbufs; i++) {
1247 uint64_t tocpy;
1248 int64_t bufoff;
1249 dmu_buf_t *db = dbp[i];
1250
1251 ASSERT(size > 0);
1252
1253 bufoff = uio->uio_loffset - db->db_offset;
1254 tocpy = MIN(db->db_size - bufoff, size);
1255
1256 #ifdef HAVE_UIO_ZEROCOPY
1257 if (xuio) {
1258 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
1259 arc_buf_t *dbuf_abuf = dbi->db_buf;
1260 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
1261 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
1262 if (!err) {
1263 uio->uio_resid -= tocpy;
1264 uio->uio_loffset += tocpy;
1265 }
1266
1267 if (abuf == dbuf_abuf)
1268 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
1269 else
1270 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
1271 } else
1272 #endif
1273 err = uiomove((char *)db->db_data + bufoff, tocpy,
1274 UIO_READ, uio);
1275 if (err)
1276 break;
1277
1278 size -= tocpy;
1279 }
1280 dmu_buf_rele_array(dbp, numbufs, FTAG);
1281
1282 return (err);
1283 }
1284
1285 /*
1286 * Read 'size' bytes into the uio buffer.
1287 * From object zdb->db_object.
1288 * Starting at offset uio->uio_loffset.
1289 *
1290 * If the caller already has a dbuf in the target object
1291 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1292 * because we don't have to find the dnode_t for the object.
1293 */
1294 int
1295 dmu_read_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size)
1296 {
1297 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1298 dnode_t *dn;
1299 int err;
1300
1301 if (size == 0)
1302 return (0);
1303
1304 DB_DNODE_ENTER(db);
1305 dn = DB_DNODE(db);
1306 err = dmu_read_uio_dnode(dn, uio, size);
1307 DB_DNODE_EXIT(db);
1308
1309 return (err);
1310 }
1311
1312 /*
1313 * Read 'size' bytes into the uio buffer.
1314 * From the specified object
1315 * Starting at offset uio->uio_loffset.
1316 */
1317 int
1318 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
1319 {
1320 dnode_t *dn;
1321 int err;
1322
1323 if (size == 0)
1324 return (0);
1325
1326 err = dnode_hold(os, object, FTAG, &dn);
1327 if (err)
1328 return (err);
1329
1330 err = dmu_read_uio_dnode(dn, uio, size);
1331
1332 dnode_rele(dn, FTAG);
1333
1334 return (err);
1335 }
1336
1337 int
1338 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1339 {
1340 dmu_buf_t **dbp;
1341 int numbufs;
1342 int err = 0;
1343 int i;
1344
1345 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1346 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1347 if (err)
1348 return (err);
1349
1350 for (i = 0; i < numbufs; i++) {
1351 uint64_t tocpy;
1352 int64_t bufoff;
1353 dmu_buf_t *db = dbp[i];
1354
1355 ASSERT(size > 0);
1356
1357 bufoff = uio->uio_loffset - db->db_offset;
1358 tocpy = MIN(db->db_size - bufoff, size);
1359
1360 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1361
1362 if (tocpy == db->db_size)
1363 dmu_buf_will_fill(db, tx);
1364 else
1365 dmu_buf_will_dirty(db, tx);
1366
1367 /*
1368 * XXX uiomove could block forever (eg.nfs-backed
1369 * pages). There needs to be a uiolockdown() function
1370 * to lock the pages in memory, so that uiomove won't
1371 * block.
1372 */
1373 err = uiomove((char *)db->db_data + bufoff, tocpy,
1374 UIO_WRITE, uio);
1375
1376 if (tocpy == db->db_size)
1377 dmu_buf_fill_done(db, tx);
1378
1379 if (err)
1380 break;
1381
1382 size -= tocpy;
1383 }
1384
1385 dmu_buf_rele_array(dbp, numbufs, FTAG);
1386 return (err);
1387 }
1388
1389 /*
1390 * Write 'size' bytes from the uio buffer.
1391 * To object zdb->db_object.
1392 * Starting at offset uio->uio_loffset.
1393 *
1394 * If the caller already has a dbuf in the target object
1395 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1396 * because we don't have to find the dnode_t for the object.
1397 */
1398 int
1399 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1400 dmu_tx_t *tx)
1401 {
1402 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1403 dnode_t *dn;
1404 int err;
1405
1406 if (size == 0)
1407 return (0);
1408
1409 DB_DNODE_ENTER(db);
1410 dn = DB_DNODE(db);
1411 err = dmu_write_uio_dnode(dn, uio, size, tx);
1412 DB_DNODE_EXIT(db);
1413
1414 return (err);
1415 }
1416
1417 /*
1418 * Write 'size' bytes from the uio buffer.
1419 * To the specified object.
1420 * Starting at offset uio->uio_loffset.
1421 */
1422 int
1423 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1424 dmu_tx_t *tx)
1425 {
1426 dnode_t *dn;
1427 int err;
1428
1429 if (size == 0)
1430 return (0);
1431
1432 err = dnode_hold(os, object, FTAG, &dn);
1433 if (err)
1434 return (err);
1435
1436 err = dmu_write_uio_dnode(dn, uio, size, tx);
1437
1438 dnode_rele(dn, FTAG);
1439
1440 return (err);
1441 }
1442 #endif /* _KERNEL */
1443
1444 /*
1445 * Allocate a loaned anonymous arc buffer.
1446 */
1447 arc_buf_t *
1448 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1449 {
1450 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1451
1452 return (arc_loan_buf(db->db_objset->os_spa, B_FALSE, size));
1453 }
1454
1455 /*
1456 * Free a loaned arc buffer.
1457 */
1458 void
1459 dmu_return_arcbuf(arc_buf_t *buf)
1460 {
1461 arc_return_buf(buf, FTAG);
1462 arc_buf_destroy(buf, FTAG);
1463 }
1464
1465 /*
1466 * When possible directly assign passed loaned arc buffer to a dbuf.
1467 * If this is not possible copy the contents of passed arc buf via
1468 * dmu_write().
1469 */
1470 void
1471 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1472 dmu_tx_t *tx)
1473 {
1474 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1475 dnode_t *dn;
1476 dmu_buf_impl_t *db;
1477 uint32_t blksz = (uint32_t)arc_buf_lsize(buf);
1478 uint64_t blkid;
1479
1480 DB_DNODE_ENTER(dbuf);
1481 dn = DB_DNODE(dbuf);
1482 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1483 blkid = dbuf_whichblock(dn, 0, offset);
1484 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1485 rw_exit(&dn->dn_struct_rwlock);
1486 DB_DNODE_EXIT(dbuf);
1487
1488 /*
1489 * We can only assign if the offset is aligned, the arc buf is the
1490 * same size as the dbuf, and the dbuf is not metadata.
1491 */
1492 if (offset == db->db.db_offset && blksz == db->db.db_size) {
1493 dbuf_assign_arcbuf(db, buf, tx);
1494 dbuf_rele(db, FTAG);
1495 } else {
1496 objset_t *os;
1497 uint64_t object;
1498
1499 /* compressed bufs must always be assignable to their dbuf */
1500 ASSERT3U(arc_get_compression(buf), ==, ZIO_COMPRESS_OFF);
1501 ASSERT(!(buf->b_flags & ARC_BUF_FLAG_COMPRESSED));
1502
1503 DB_DNODE_ENTER(dbuf);
1504 dn = DB_DNODE(dbuf);
1505 os = dn->dn_objset;
1506 object = dn->dn_object;
1507 DB_DNODE_EXIT(dbuf);
1508
1509 dbuf_rele(db, FTAG);
1510 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1511 dmu_return_arcbuf(buf);
1512 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1513 }
1514 }
1515
1516 typedef struct {
1517 dbuf_dirty_record_t *dsa_dr;
1518 dmu_sync_cb_t *dsa_done;
1519 zgd_t *dsa_zgd;
1520 dmu_tx_t *dsa_tx;
1521 } dmu_sync_arg_t;
1522
1523 /* ARGSUSED */
1524 static void
1525 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1526 {
1527 dmu_sync_arg_t *dsa = varg;
1528 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1529 blkptr_t *bp = zio->io_bp;
1530
1531 if (zio->io_error == 0) {
1532 if (BP_IS_HOLE(bp)) {
1533 /*
1534 * A block of zeros may compress to a hole, but the
1535 * block size still needs to be known for replay.
1536 */
1537 BP_SET_LSIZE(bp, db->db_size);
1538 } else if (!BP_IS_EMBEDDED(bp)) {
1539 ASSERT(BP_GET_LEVEL(bp) == 0);
1540 bp->blk_fill = 1;
1541 }
1542 }
1543 }
1544
1545 static void
1546 dmu_sync_late_arrival_ready(zio_t *zio)
1547 {
1548 dmu_sync_ready(zio, NULL, zio->io_private);
1549 }
1550
1551 /* ARGSUSED */
1552 static void
1553 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1554 {
1555 dmu_sync_arg_t *dsa = varg;
1556 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1557 dmu_buf_impl_t *db = dr->dr_dbuf;
1558
1559 mutex_enter(&db->db_mtx);
1560 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1561 if (zio->io_error == 0) {
1562 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1563 if (dr->dt.dl.dr_nopwrite) {
1564 ASSERTV(blkptr_t *bp = zio->io_bp);
1565 ASSERTV(blkptr_t *bp_orig = &zio->io_bp_orig);
1566 ASSERTV(uint8_t chksum = BP_GET_CHECKSUM(bp_orig));
1567
1568 ASSERT(BP_EQUAL(bp, bp_orig));
1569 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1570 ASSERT(zio_checksum_table[chksum].ci_flags &
1571 ZCHECKSUM_FLAG_NOPWRITE);
1572 }
1573 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1574 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1575 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1576
1577 /*
1578 * Old style holes are filled with all zeros, whereas
1579 * new-style holes maintain their lsize, type, level,
1580 * and birth time (see zio_write_compress). While we
1581 * need to reset the BP_SET_LSIZE() call that happened
1582 * in dmu_sync_ready for old style holes, we do *not*
1583 * want to wipe out the information contained in new
1584 * style holes. Thus, only zero out the block pointer if
1585 * it's an old style hole.
1586 */
1587 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) &&
1588 dr->dt.dl.dr_overridden_by.blk_birth == 0)
1589 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1590 } else {
1591 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1592 }
1593 cv_broadcast(&db->db_changed);
1594 mutex_exit(&db->db_mtx);
1595
1596 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1597
1598 kmem_free(dsa, sizeof (*dsa));
1599 }
1600
1601 static void
1602 dmu_sync_late_arrival_done(zio_t *zio)
1603 {
1604 blkptr_t *bp = zio->io_bp;
1605 dmu_sync_arg_t *dsa = zio->io_private;
1606 ASSERTV(blkptr_t *bp_orig = &zio->io_bp_orig);
1607
1608 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1609 /*
1610 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE)
1611 * then there is nothing to do here. Otherwise, free the
1612 * newly allocated block in this txg.
1613 */
1614 if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
1615 ASSERT(BP_EQUAL(bp, bp_orig));
1616 } else {
1617 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1618 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1619 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1620 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1621 }
1622 }
1623
1624 dmu_tx_commit(dsa->dsa_tx);
1625
1626 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1627
1628 abd_put(zio->io_abd);
1629 kmem_free(dsa, sizeof (*dsa));
1630 }
1631
1632 static int
1633 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1634 zio_prop_t *zp, zbookmark_phys_t *zb)
1635 {
1636 dmu_sync_arg_t *dsa;
1637 dmu_tx_t *tx;
1638
1639 tx = dmu_tx_create(os);
1640 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1641 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1642 dmu_tx_abort(tx);
1643 /* Make zl_get_data do txg_waited_synced() */
1644 return (SET_ERROR(EIO));
1645 }
1646
1647 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1648 dsa->dsa_dr = NULL;
1649 dsa->dsa_done = done;
1650 dsa->dsa_zgd = zgd;
1651 dsa->dsa_tx = tx;
1652
1653 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1654 abd_get_from_buf(zgd->zgd_db->db_data, zgd->zgd_db->db_size),
1655 zgd->zgd_db->db_size, zgd->zgd_db->db_size, zp,
1656 dmu_sync_late_arrival_ready, NULL, NULL, dmu_sync_late_arrival_done,
1657 dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1658
1659 return (0);
1660 }
1661
1662 /*
1663 * Intent log support: sync the block associated with db to disk.
1664 * N.B. and XXX: the caller is responsible for making sure that the
1665 * data isn't changing while dmu_sync() is writing it.
1666 *
1667 * Return values:
1668 *
1669 * EEXIST: this txg has already been synced, so there's nothing to do.
1670 * The caller should not log the write.
1671 *
1672 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1673 * The caller should not log the write.
1674 *
1675 * EALREADY: this block is already in the process of being synced.
1676 * The caller should track its progress (somehow).
1677 *
1678 * EIO: could not do the I/O.
1679 * The caller should do a txg_wait_synced().
1680 *
1681 * 0: the I/O has been initiated.
1682 * The caller should log this blkptr in the done callback.
1683 * It is possible that the I/O will fail, in which case
1684 * the error will be reported to the done callback and
1685 * propagated to pio from zio_done().
1686 */
1687 int
1688 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1689 {
1690 blkptr_t *bp = zgd->zgd_bp;
1691 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1692 objset_t *os = db->db_objset;
1693 dsl_dataset_t *ds = os->os_dsl_dataset;
1694 dbuf_dirty_record_t *dr;
1695 dmu_sync_arg_t *dsa;
1696 zbookmark_phys_t zb;
1697 zio_prop_t zp;
1698 dnode_t *dn;
1699
1700 ASSERT(pio != NULL);
1701 ASSERT(txg != 0);
1702
1703 SET_BOOKMARK(&zb, ds->ds_object,
1704 db->db.db_object, db->db_level, db->db_blkid);
1705
1706 DB_DNODE_ENTER(db);
1707 dn = DB_DNODE(db);
1708 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1709 DB_DNODE_EXIT(db);
1710
1711 /*
1712 * If we're frozen (running ziltest), we always need to generate a bp.
1713 */
1714 if (txg > spa_freeze_txg(os->os_spa))
1715 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1716
1717 /*
1718 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1719 * and us. If we determine that this txg is not yet syncing,
1720 * but it begins to sync a moment later, that's OK because the
1721 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1722 */
1723 mutex_enter(&db->db_mtx);
1724
1725 if (txg <= spa_last_synced_txg(os->os_spa)) {
1726 /*
1727 * This txg has already synced. There's nothing to do.
1728 */
1729 mutex_exit(&db->db_mtx);
1730 return (SET_ERROR(EEXIST));
1731 }
1732
1733 if (txg <= spa_syncing_txg(os->os_spa)) {
1734 /*
1735 * This txg is currently syncing, so we can't mess with
1736 * the dirty record anymore; just write a new log block.
1737 */
1738 mutex_exit(&db->db_mtx);
1739 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1740 }
1741
1742 dr = db->db_last_dirty;
1743 while (dr && dr->dr_txg != txg)
1744 dr = dr->dr_next;
1745
1746 if (dr == NULL) {
1747 /*
1748 * There's no dr for this dbuf, so it must have been freed.
1749 * There's no need to log writes to freed blocks, so we're done.
1750 */
1751 mutex_exit(&db->db_mtx);
1752 return (SET_ERROR(ENOENT));
1753 }
1754
1755 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
1756
1757 /*
1758 * Assume the on-disk data is X, the current syncing data (in
1759 * txg - 1) is Y, and the current in-memory data is Z (currently
1760 * in dmu_sync).
1761 *
1762 * We usually want to perform a nopwrite if X and Z are the
1763 * same. However, if Y is different (i.e. the BP is going to
1764 * change before this write takes effect), then a nopwrite will
1765 * be incorrect - we would override with X, which could have
1766 * been freed when Y was written.
1767 *
1768 * (Note that this is not a concern when we are nop-writing from
1769 * syncing context, because X and Y must be identical, because
1770 * all previous txgs have been synced.)
1771 *
1772 * Therefore, we disable nopwrite if the current BP could change
1773 * before this TXG. There are two ways it could change: by
1774 * being dirty (dr_next is non-NULL), or by being freed
1775 * (dnode_block_freed()). This behavior is verified by
1776 * zio_done(), which VERIFYs that the override BP is identical
1777 * to the on-disk BP.
1778 */
1779 DB_DNODE_ENTER(db);
1780 dn = DB_DNODE(db);
1781 if (dr->dr_next != NULL || dnode_block_freed(dn, db->db_blkid))
1782 zp.zp_nopwrite = B_FALSE;
1783 DB_DNODE_EXIT(db);
1784
1785 ASSERT(dr->dr_txg == txg);
1786 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1787 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1788 /*
1789 * We have already issued a sync write for this buffer,
1790 * or this buffer has already been synced. It could not
1791 * have been dirtied since, or we would have cleared the state.
1792 */
1793 mutex_exit(&db->db_mtx);
1794 return (SET_ERROR(EALREADY));
1795 }
1796
1797 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1798 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1799 mutex_exit(&db->db_mtx);
1800
1801 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1802 dsa->dsa_dr = dr;
1803 dsa->dsa_done = done;
1804 dsa->dsa_zgd = zgd;
1805 dsa->dsa_tx = NULL;
1806
1807 zio_nowait(arc_write(pio, os->os_spa, txg,
1808 bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
1809 &zp, dmu_sync_ready, NULL, NULL, dmu_sync_done, dsa,
1810 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb));
1811
1812 return (0);
1813 }
1814
1815 int
1816 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1817 dmu_tx_t *tx)
1818 {
1819 dnode_t *dn;
1820 int err;
1821
1822 err = dnode_hold(os, object, FTAG, &dn);
1823 if (err)
1824 return (err);
1825 err = dnode_set_blksz(dn, size, ibs, tx);
1826 dnode_rele(dn, FTAG);
1827 return (err);
1828 }
1829
1830 void
1831 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1832 dmu_tx_t *tx)
1833 {
1834 dnode_t *dn;
1835
1836 /*
1837 * Send streams include each object's checksum function. This
1838 * check ensures that the receiving system can understand the
1839 * checksum function transmitted.
1840 */
1841 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS);
1842
1843 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1844 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS);
1845 dn->dn_checksum = checksum;
1846 dnode_setdirty(dn, tx);
1847 dnode_rele(dn, FTAG);
1848 }
1849
1850 void
1851 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1852 dmu_tx_t *tx)
1853 {
1854 dnode_t *dn;
1855
1856 /*
1857 * Send streams include each object's compression function. This
1858 * check ensures that the receiving system can understand the
1859 * compression function transmitted.
1860 */
1861 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS);
1862
1863 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1864 dn->dn_compress = compress;
1865 dnode_setdirty(dn, tx);
1866 dnode_rele(dn, FTAG);
1867 }
1868
1869 int zfs_mdcomp_disable = 0;
1870
1871 /*
1872 * When the "redundant_metadata" property is set to "most", only indirect
1873 * blocks of this level and higher will have an additional ditto block.
1874 */
1875 int zfs_redundant_metadata_most_ditto_level = 2;
1876
1877 void
1878 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
1879 {
1880 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
1881 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
1882 (wp & WP_SPILL));
1883 enum zio_checksum checksum = os->os_checksum;
1884 enum zio_compress compress = os->os_compress;
1885 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
1886 boolean_t dedup = B_FALSE;
1887 boolean_t nopwrite = B_FALSE;
1888 boolean_t dedup_verify = os->os_dedup_verify;
1889 int copies = os->os_copies;
1890
1891 /*
1892 * We maintain different write policies for each of the following
1893 * types of data:
1894 * 1. metadata
1895 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1896 * 3. all other level 0 blocks
1897 */
1898 if (ismd) {
1899 if (zfs_mdcomp_disable) {
1900 compress = ZIO_COMPRESS_EMPTY;
1901 } else {
1902 /*
1903 * XXX -- we should design a compression algorithm
1904 * that specializes in arrays of bps.
1905 */
1906 compress = zio_compress_select(os->os_spa,
1907 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON);
1908 }
1909
1910 /*
1911 * Metadata always gets checksummed. If the data
1912 * checksum is multi-bit correctable, and it's not a
1913 * ZBT-style checksum, then it's suitable for metadata
1914 * as well. Otherwise, the metadata checksum defaults
1915 * to fletcher4.
1916 */
1917 if (!(zio_checksum_table[checksum].ci_flags &
1918 ZCHECKSUM_FLAG_METADATA) ||
1919 (zio_checksum_table[checksum].ci_flags &
1920 ZCHECKSUM_FLAG_EMBEDDED))
1921 checksum = ZIO_CHECKSUM_FLETCHER_4;
1922
1923 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL ||
1924 (os->os_redundant_metadata ==
1925 ZFS_REDUNDANT_METADATA_MOST &&
1926 (level >= zfs_redundant_metadata_most_ditto_level ||
1927 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))))
1928 copies++;
1929 } else if (wp & WP_NOFILL) {
1930 ASSERT(level == 0);
1931
1932 /*
1933 * If we're writing preallocated blocks, we aren't actually
1934 * writing them so don't set any policy properties. These
1935 * blocks are currently only used by an external subsystem
1936 * outside of zfs (i.e. dump) and not written by the zio
1937 * pipeline.
1938 */
1939 compress = ZIO_COMPRESS_OFF;
1940 checksum = ZIO_CHECKSUM_OFF;
1941 } else {
1942 compress = zio_compress_select(os->os_spa, dn->dn_compress,
1943 compress);
1944
1945 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
1946 zio_checksum_select(dn->dn_checksum, checksum) :
1947 dedup_checksum;
1948
1949 /*
1950 * Determine dedup setting. If we are in dmu_sync(),
1951 * we won't actually dedup now because that's all
1952 * done in syncing context; but we do want to use the
1953 * dedup checkum. If the checksum is not strong
1954 * enough to ensure unique signatures, force
1955 * dedup_verify.
1956 */
1957 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
1958 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
1959 if (!(zio_checksum_table[checksum].ci_flags &
1960 ZCHECKSUM_FLAG_DEDUP))
1961 dedup_verify = B_TRUE;
1962 }
1963
1964 /*
1965 * Enable nopwrite if we have secure enough checksum
1966 * algorithm (see comment in zio_nop_write) and
1967 * compression is enabled. We don't enable nopwrite if
1968 * dedup is enabled as the two features are mutually
1969 * exclusive.
1970 */
1971 nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags &
1972 ZCHECKSUM_FLAG_NOPWRITE) &&
1973 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
1974 }
1975
1976 zp->zp_checksum = checksum;
1977 zp->zp_compress = compress;
1978 ASSERT3U(zp->zp_compress, !=, ZIO_COMPRESS_INHERIT);
1979
1980 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
1981 zp->zp_level = level;
1982 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
1983 zp->zp_dedup = dedup;
1984 zp->zp_dedup_verify = dedup && dedup_verify;
1985 zp->zp_nopwrite = nopwrite;
1986 }
1987
1988 /*
1989 * This function is only called from zfs_holey_common() for zpl_llseek()
1990 * in order to determine the location of holes. In order to accurately
1991 * report holes all dirty data must be synced to disk. This causes extremely
1992 * poor performance when seeking for holes in a dirty file. As a compromise,
1993 * only provide hole data when the dnode is clean. When a dnode is dirty
1994 * report the dnode as having no holes which is always a safe thing to do.
1995 */
1996 int
1997 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
1998 {
1999 dnode_t *dn;
2000 int i, err;
2001 boolean_t clean = B_TRUE;
2002
2003 err = dnode_hold(os, object, FTAG, &dn);
2004 if (err)
2005 return (err);
2006
2007 /*
2008 * Check if dnode is dirty
2009 */
2010 if (dn->dn_dirtyctx != DN_UNDIRTIED) {
2011 for (i = 0; i < TXG_SIZE; i++) {
2012 if (!list_is_empty(&dn->dn_dirty_records[i])) {
2013 clean = B_FALSE;
2014 break;
2015 }
2016 }
2017 }
2018
2019 /*
2020 * If compatibility option is on, sync any current changes before
2021 * we go trundling through the block pointers.
2022 */
2023 if (!clean && zfs_dmu_offset_next_sync) {
2024 clean = B_TRUE;
2025 dnode_rele(dn, FTAG);
2026 txg_wait_synced(dmu_objset_pool(os), 0);
2027 err = dnode_hold(os, object, FTAG, &dn);
2028 if (err)
2029 return (err);
2030 }
2031
2032 if (clean)
2033 err = dnode_next_offset(dn,
2034 (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
2035 else
2036 err = SET_ERROR(EBUSY);
2037
2038 dnode_rele(dn, FTAG);
2039
2040 return (err);
2041 }
2042
2043 void
2044 __dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2045 {
2046 dnode_phys_t *dnp = dn->dn_phys;
2047 int i;
2048
2049 doi->doi_data_block_size = dn->dn_datablksz;
2050 doi->doi_metadata_block_size = dn->dn_indblkshift ?
2051 1ULL << dn->dn_indblkshift : 0;
2052 doi->doi_type = dn->dn_type;
2053 doi->doi_bonus_type = dn->dn_bonustype;
2054 doi->doi_bonus_size = dn->dn_bonuslen;
2055 doi->doi_dnodesize = dn->dn_num_slots << DNODE_SHIFT;
2056 doi->doi_indirection = dn->dn_nlevels;
2057 doi->doi_checksum = dn->dn_checksum;
2058 doi->doi_compress = dn->dn_compress;
2059 doi->doi_nblkptr = dn->dn_nblkptr;
2060 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
2061 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
2062 doi->doi_fill_count = 0;
2063 for (i = 0; i < dnp->dn_nblkptr; i++)
2064 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
2065 }
2066
2067 void
2068 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2069 {
2070 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2071 mutex_enter(&dn->dn_mtx);
2072
2073 __dmu_object_info_from_dnode(dn, doi);
2074
2075 mutex_exit(&dn->dn_mtx);
2076 rw_exit(&dn->dn_struct_rwlock);
2077 }
2078
2079 /*
2080 * Get information on a DMU object.
2081 * If doi is NULL, just indicates whether the object exists.
2082 */
2083 int
2084 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
2085 {
2086 dnode_t *dn;
2087 int err = dnode_hold(os, object, FTAG, &dn);
2088
2089 if (err)
2090 return (err);
2091
2092 if (doi != NULL)
2093 dmu_object_info_from_dnode(dn, doi);
2094
2095 dnode_rele(dn, FTAG);
2096 return (0);
2097 }
2098
2099 /*
2100 * As above, but faster; can be used when you have a held dbuf in hand.
2101 */
2102 void
2103 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
2104 {
2105 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2106
2107 DB_DNODE_ENTER(db);
2108 dmu_object_info_from_dnode(DB_DNODE(db), doi);
2109 DB_DNODE_EXIT(db);
2110 }
2111
2112 /*
2113 * Faster still when you only care about the size.
2114 * This is specifically optimized for zfs_getattr().
2115 */
2116 void
2117 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
2118 u_longlong_t *nblk512)
2119 {
2120 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2121 dnode_t *dn;
2122
2123 DB_DNODE_ENTER(db);
2124 dn = DB_DNODE(db);
2125
2126 *blksize = dn->dn_datablksz;
2127 /* add in number of slots used for the dnode itself */
2128 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
2129 SPA_MINBLOCKSHIFT) + dn->dn_num_slots;
2130 DB_DNODE_EXIT(db);
2131 }
2132
2133 void
2134 dmu_object_dnsize_from_db(dmu_buf_t *db_fake, int *dnsize)
2135 {
2136 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2137 dnode_t *dn;
2138
2139 DB_DNODE_ENTER(db);
2140 dn = DB_DNODE(db);
2141 *dnsize = dn->dn_num_slots << DNODE_SHIFT;
2142 DB_DNODE_EXIT(db);
2143 }
2144
2145 void
2146 byteswap_uint64_array(void *vbuf, size_t size)
2147 {
2148 uint64_t *buf = vbuf;
2149 size_t count = size >> 3;
2150 int i;
2151
2152 ASSERT((size & 7) == 0);
2153
2154 for (i = 0; i < count; i++)
2155 buf[i] = BSWAP_64(buf[i]);
2156 }
2157
2158 void
2159 byteswap_uint32_array(void *vbuf, size_t size)
2160 {
2161 uint32_t *buf = vbuf;
2162 size_t count = size >> 2;
2163 int i;
2164
2165 ASSERT((size & 3) == 0);
2166
2167 for (i = 0; i < count; i++)
2168 buf[i] = BSWAP_32(buf[i]);
2169 }
2170
2171 void
2172 byteswap_uint16_array(void *vbuf, size_t size)
2173 {
2174 uint16_t *buf = vbuf;
2175 size_t count = size >> 1;
2176 int i;
2177
2178 ASSERT((size & 1) == 0);
2179
2180 for (i = 0; i < count; i++)
2181 buf[i] = BSWAP_16(buf[i]);
2182 }
2183
2184 /* ARGSUSED */
2185 void
2186 byteswap_uint8_array(void *vbuf, size_t size)
2187 {
2188 }
2189
2190 void
2191 dmu_init(void)
2192 {
2193 abd_init();
2194 zfs_dbgmsg_init();
2195 sa_cache_init();
2196 xuio_stat_init();
2197 dmu_objset_init();
2198 dnode_init();
2199 zfetch_init();
2200 dmu_tx_init();
2201 l2arc_init();
2202 arc_init();
2203 dbuf_init();
2204 }
2205
2206 void
2207 dmu_fini(void)
2208 {
2209 arc_fini(); /* arc depends on l2arc, so arc must go first */
2210 l2arc_fini();
2211 dmu_tx_fini();
2212 zfetch_fini();
2213 dbuf_fini();
2214 dnode_fini();
2215 dmu_objset_fini();
2216 xuio_stat_fini();
2217 sa_cache_fini();
2218 zfs_dbgmsg_fini();
2219 abd_fini();
2220 }
2221
2222 #if defined(_KERNEL) && defined(HAVE_SPL)
2223 EXPORT_SYMBOL(dmu_bonus_hold);
2224 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus);
2225 EXPORT_SYMBOL(dmu_buf_rele_array);
2226 EXPORT_SYMBOL(dmu_prefetch);
2227 EXPORT_SYMBOL(dmu_free_range);
2228 EXPORT_SYMBOL(dmu_free_long_range);
2229 EXPORT_SYMBOL(dmu_free_long_object);
2230 EXPORT_SYMBOL(dmu_read);
2231 EXPORT_SYMBOL(dmu_read_by_dnode);
2232 EXPORT_SYMBOL(dmu_write);
2233 EXPORT_SYMBOL(dmu_write_by_dnode);
2234 EXPORT_SYMBOL(dmu_prealloc);
2235 EXPORT_SYMBOL(dmu_object_info);
2236 EXPORT_SYMBOL(dmu_object_info_from_dnode);
2237 EXPORT_SYMBOL(dmu_object_info_from_db);
2238 EXPORT_SYMBOL(dmu_object_size_from_db);
2239 EXPORT_SYMBOL(dmu_object_dnsize_from_db);
2240 EXPORT_SYMBOL(dmu_object_set_blocksize);
2241 EXPORT_SYMBOL(dmu_object_set_checksum);
2242 EXPORT_SYMBOL(dmu_object_set_compress);
2243 EXPORT_SYMBOL(dmu_write_policy);
2244 EXPORT_SYMBOL(dmu_sync);
2245 EXPORT_SYMBOL(dmu_request_arcbuf);
2246 EXPORT_SYMBOL(dmu_return_arcbuf);
2247 EXPORT_SYMBOL(dmu_assign_arcbuf);
2248 EXPORT_SYMBOL(dmu_buf_hold);
2249 EXPORT_SYMBOL(dmu_ot);
2250
2251 /* BEGIN CSTYLED */
2252 module_param(zfs_mdcomp_disable, int, 0644);
2253 MODULE_PARM_DESC(zfs_mdcomp_disable, "Disable meta data compression");
2254
2255 module_param(zfs_nopwrite_enabled, int, 0644);
2256 MODULE_PARM_DESC(zfs_nopwrite_enabled, "Enable NOP writes");
2257
2258 module_param(zfs_per_txg_dirty_frees_percent, ulong, 0644);
2259 MODULE_PARM_DESC(zfs_per_txg_dirty_frees_percent,
2260 "percentage of dirtied blocks from frees in one TXG");
2261
2262 module_param(zfs_dmu_offset_next_sync, int, 0644);
2263 MODULE_PARM_DESC(zfs_dmu_offset_next_sync,
2264 "Enable forcing txg sync to find holes");
2265
2266 /* END CSTYLED */
2267
2268 #endif
ZFS On Linux mirror for PVE