]> 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
Fix range locking in ZIL commit codepath
[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 #include <sys/zfs_rlock.h>
53 #ifdef _KERNEL
54 #include <sys/vmsystm.h>
55 #include <sys/zfs_znode.h>
56 #endif
57
58 /*
59 * Enable/disable nopwrite feature.
60 */
61 int zfs_nopwrite_enabled = 1;
62
63 /*
64 * Tunable to control percentage of dirtied blocks from frees in one TXG.
65 * After this threshold is crossed, additional dirty blocks from frees
66 * wait until the next TXG.
67 * A value of zero will disable this throttle.
68 */
69 unsigned long zfs_per_txg_dirty_frees_percent = 30;
70
71 /*
72 * Enable/disable forcing txg sync when dirty in dmu_offset_next.
73 */
74 int zfs_dmu_offset_next_sync = 0;
75
76 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
77 { DMU_BSWAP_UINT8, TRUE, FALSE, "unallocated" },
78 { DMU_BSWAP_ZAP, TRUE, FALSE, "object directory" },
79 { DMU_BSWAP_UINT64, TRUE, FALSE, "object array" },
80 { DMU_BSWAP_UINT8, TRUE, FALSE, "packed nvlist" },
81 { DMU_BSWAP_UINT64, TRUE, FALSE, "packed nvlist size" },
82 { DMU_BSWAP_UINT64, TRUE, FALSE, "bpobj" },
83 { DMU_BSWAP_UINT64, TRUE, FALSE, "bpobj header" },
84 { DMU_BSWAP_UINT64, TRUE, FALSE, "SPA space map header" },
85 { DMU_BSWAP_UINT64, TRUE, FALSE, "SPA space map" },
86 { DMU_BSWAP_UINT64, TRUE, TRUE, "ZIL intent log" },
87 { DMU_BSWAP_DNODE, TRUE, TRUE, "DMU dnode" },
88 { DMU_BSWAP_OBJSET, TRUE, FALSE, "DMU objset" },
89 { DMU_BSWAP_UINT64, TRUE, FALSE, "DSL directory" },
90 { DMU_BSWAP_ZAP, TRUE, FALSE, "DSL directory child map"},
91 { DMU_BSWAP_ZAP, TRUE, FALSE, "DSL dataset snap map" },
92 { DMU_BSWAP_ZAP, TRUE, FALSE, "DSL props" },
93 { DMU_BSWAP_UINT64, TRUE, FALSE, "DSL dataset" },
94 { DMU_BSWAP_ZNODE, TRUE, FALSE, "ZFS znode" },
95 { DMU_BSWAP_OLDACL, TRUE, TRUE, "ZFS V0 ACL" },
96 { DMU_BSWAP_UINT8, FALSE, TRUE, "ZFS plain file" },
97 { DMU_BSWAP_ZAP, TRUE, TRUE, "ZFS directory" },
98 { DMU_BSWAP_ZAP, TRUE, FALSE, "ZFS master node" },
99 { DMU_BSWAP_ZAP, TRUE, TRUE, "ZFS delete queue" },
100 { DMU_BSWAP_UINT8, FALSE, TRUE, "zvol object" },
101 { DMU_BSWAP_ZAP, TRUE, FALSE, "zvol prop" },
102 { DMU_BSWAP_UINT8, FALSE, TRUE, "other uint8[]" },
103 { DMU_BSWAP_UINT64, FALSE, TRUE, "other uint64[]" },
104 { DMU_BSWAP_ZAP, TRUE, FALSE, "other ZAP" },
105 { DMU_BSWAP_ZAP, TRUE, FALSE, "persistent error log" },
106 { DMU_BSWAP_UINT8, TRUE, FALSE, "SPA history" },
107 { DMU_BSWAP_UINT64, TRUE, FALSE, "SPA history offsets" },
108 { DMU_BSWAP_ZAP, TRUE, FALSE, "Pool properties" },
109 { DMU_BSWAP_ZAP, TRUE, FALSE, "DSL permissions" },
110 { DMU_BSWAP_ACL, TRUE, TRUE, "ZFS ACL" },
111 { DMU_BSWAP_UINT8, TRUE, TRUE, "ZFS SYSACL" },
112 { DMU_BSWAP_UINT8, TRUE, TRUE, "FUID table" },
113 { DMU_BSWAP_UINT64, TRUE, FALSE, "FUID table size" },
114 { DMU_BSWAP_ZAP, TRUE, FALSE, "DSL dataset next clones"},
115 { DMU_BSWAP_ZAP, TRUE, FALSE, "scan work queue" },
116 { DMU_BSWAP_ZAP, TRUE, TRUE, "ZFS user/group used" },
117 { DMU_BSWAP_ZAP, TRUE, TRUE, "ZFS user/group quota" },
118 { DMU_BSWAP_ZAP, TRUE, FALSE, "snapshot refcount tags"},
119 { DMU_BSWAP_ZAP, TRUE, FALSE, "DDT ZAP algorithm" },
120 { DMU_BSWAP_ZAP, TRUE, FALSE, "DDT statistics" },
121 { DMU_BSWAP_UINT8, TRUE, TRUE, "System attributes" },
122 { DMU_BSWAP_ZAP, TRUE, TRUE, "SA master node" },
123 { DMU_BSWAP_ZAP, TRUE, TRUE, "SA attr registration" },
124 { DMU_BSWAP_ZAP, TRUE, TRUE, "SA attr layouts" },
125 { DMU_BSWAP_ZAP, TRUE, FALSE, "scan translations" },
126 { DMU_BSWAP_UINT8, FALSE, TRUE, "deduplicated block" },
127 { DMU_BSWAP_ZAP, TRUE, FALSE, "DSL deadlist map" },
128 { DMU_BSWAP_UINT64, TRUE, FALSE, "DSL deadlist map hdr" },
129 { DMU_BSWAP_ZAP, TRUE, FALSE, "DSL dir clones" },
130 { DMU_BSWAP_UINT64, TRUE, FALSE, "bpobj subobj" }
131 };
132
133 const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
134 { byteswap_uint8_array, "uint8" },
135 { byteswap_uint16_array, "uint16" },
136 { byteswap_uint32_array, "uint32" },
137 { byteswap_uint64_array, "uint64" },
138 { zap_byteswap, "zap" },
139 { dnode_buf_byteswap, "dnode" },
140 { dmu_objset_byteswap, "objset" },
141 { zfs_znode_byteswap, "znode" },
142 { zfs_oldacl_byteswap, "oldacl" },
143 { zfs_acl_byteswap, "acl" }
144 };
145
146 int
147 dmu_buf_hold_noread_by_dnode(dnode_t *dn, uint64_t offset,
148 void *tag, dmu_buf_t **dbp)
149 {
150 uint64_t blkid;
151 dmu_buf_impl_t *db;
152
153 blkid = dbuf_whichblock(dn, 0, offset);
154 rw_enter(&dn->dn_struct_rwlock, RW_READER);
155 db = dbuf_hold(dn, blkid, tag);
156 rw_exit(&dn->dn_struct_rwlock);
157
158 if (db == NULL) {
159 *dbp = NULL;
160 return (SET_ERROR(EIO));
161 }
162
163 *dbp = &db->db;
164 return (0);
165 }
166 int
167 dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset,
168 void *tag, dmu_buf_t **dbp)
169 {
170 dnode_t *dn;
171 uint64_t blkid;
172 dmu_buf_impl_t *db;
173 int err;
174
175 err = dnode_hold(os, object, FTAG, &dn);
176 if (err)
177 return (err);
178 blkid = dbuf_whichblock(dn, 0, offset);
179 rw_enter(&dn->dn_struct_rwlock, RW_READER);
180 db = dbuf_hold(dn, blkid, tag);
181 rw_exit(&dn->dn_struct_rwlock);
182 dnode_rele(dn, FTAG);
183
184 if (db == NULL) {
185 *dbp = NULL;
186 return (SET_ERROR(EIO));
187 }
188
189 *dbp = &db->db;
190 return (err);
191 }
192
193 int
194 dmu_buf_hold_by_dnode(dnode_t *dn, uint64_t offset,
195 void *tag, dmu_buf_t **dbp, int flags)
196 {
197 int err;
198 int db_flags = DB_RF_CANFAIL;
199
200 if (flags & DMU_READ_NO_PREFETCH)
201 db_flags |= DB_RF_NOPREFETCH;
202 if (flags & DMU_READ_NO_DECRYPT)
203 db_flags |= DB_RF_NO_DECRYPT;
204
205 err = dmu_buf_hold_noread_by_dnode(dn, offset, tag, dbp);
206 if (err == 0) {
207 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
208 err = dbuf_read(db, NULL, db_flags);
209 if (err != 0) {
210 dbuf_rele(db, tag);
211 *dbp = NULL;
212 }
213 }
214
215 return (err);
216 }
217
218 int
219 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
220 void *tag, dmu_buf_t **dbp, int flags)
221 {
222 int err;
223 int db_flags = DB_RF_CANFAIL;
224
225 if (flags & DMU_READ_NO_PREFETCH)
226 db_flags |= DB_RF_NOPREFETCH;
227 if (flags & DMU_READ_NO_DECRYPT)
228 db_flags |= DB_RF_NO_DECRYPT;
229
230 err = dmu_buf_hold_noread(os, object, offset, tag, dbp);
231 if (err == 0) {
232 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
233 err = dbuf_read(db, NULL, db_flags);
234 if (err != 0) {
235 dbuf_rele(db, tag);
236 *dbp = NULL;
237 }
238 }
239
240 return (err);
241 }
242
243 int
244 dmu_bonus_max(void)
245 {
246 return (DN_OLD_MAX_BONUSLEN);
247 }
248
249 int
250 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
251 {
252 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
253 dnode_t *dn;
254 int error;
255
256 DB_DNODE_ENTER(db);
257 dn = DB_DNODE(db);
258
259 if (dn->dn_bonus != db) {
260 error = SET_ERROR(EINVAL);
261 } else if (newsize < 0 || newsize > db_fake->db_size) {
262 error = SET_ERROR(EINVAL);
263 } else {
264 dnode_setbonuslen(dn, newsize, tx);
265 error = 0;
266 }
267
268 DB_DNODE_EXIT(db);
269 return (error);
270 }
271
272 int
273 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
274 {
275 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
276 dnode_t *dn;
277 int error;
278
279 DB_DNODE_ENTER(db);
280 dn = DB_DNODE(db);
281
282 if (!DMU_OT_IS_VALID(type)) {
283 error = SET_ERROR(EINVAL);
284 } else if (dn->dn_bonus != db) {
285 error = SET_ERROR(EINVAL);
286 } else {
287 dnode_setbonus_type(dn, type, tx);
288 error = 0;
289 }
290
291 DB_DNODE_EXIT(db);
292 return (error);
293 }
294
295 dmu_object_type_t
296 dmu_get_bonustype(dmu_buf_t *db_fake)
297 {
298 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
299 dnode_t *dn;
300 dmu_object_type_t type;
301
302 DB_DNODE_ENTER(db);
303 dn = DB_DNODE(db);
304 type = dn->dn_bonustype;
305 DB_DNODE_EXIT(db);
306
307 return (type);
308 }
309
310 int
311 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
312 {
313 dnode_t *dn;
314 int error;
315
316 error = dnode_hold(os, object, FTAG, &dn);
317 dbuf_rm_spill(dn, tx);
318 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
319 dnode_rm_spill(dn, tx);
320 rw_exit(&dn->dn_struct_rwlock);
321 dnode_rele(dn, FTAG);
322 return (error);
323 }
324
325 /*
326 * returns ENOENT, EIO, or 0.
327 */
328 int
329 dmu_bonus_hold_impl(objset_t *os, uint64_t object, void *tag, uint32_t flags,
330 dmu_buf_t **dbp)
331 {
332 dnode_t *dn;
333 dmu_buf_impl_t *db;
334 int error;
335 uint32_t db_flags = DB_RF_MUST_SUCCEED;
336
337 if (flags & DMU_READ_NO_PREFETCH)
338 db_flags |= DB_RF_NOPREFETCH;
339 if (flags & DMU_READ_NO_DECRYPT)
340 db_flags |= DB_RF_NO_DECRYPT;
341
342 error = dnode_hold(os, object, FTAG, &dn);
343 if (error)
344 return (error);
345
346 rw_enter(&dn->dn_struct_rwlock, RW_READER);
347 if (dn->dn_bonus == NULL) {
348 rw_exit(&dn->dn_struct_rwlock);
349 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
350 if (dn->dn_bonus == NULL)
351 dbuf_create_bonus(dn);
352 }
353 db = dn->dn_bonus;
354
355 /* as long as the bonus buf is held, the dnode will be held */
356 if (refcount_add(&db->db_holds, tag) == 1) {
357 VERIFY(dnode_add_ref(dn, db));
358 atomic_inc_32(&dn->dn_dbufs_count);
359 }
360
361 /*
362 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
363 * hold and incrementing the dbuf count to ensure that dnode_move() sees
364 * a dnode hold for every dbuf.
365 */
366 rw_exit(&dn->dn_struct_rwlock);
367
368 dnode_rele(dn, FTAG);
369
370 error = dbuf_read(db, NULL, db_flags);
371 if (error) {
372 dnode_evict_bonus(dn);
373 dbuf_rele(db, tag);
374 *dbp = NULL;
375 return (error);
376 }
377
378 *dbp = &db->db;
379 return (0);
380 }
381
382 int
383 dmu_bonus_hold(objset_t *os, uint64_t obj, void *tag, dmu_buf_t **dbp)
384 {
385 return (dmu_bonus_hold_impl(os, obj, tag, DMU_READ_NO_PREFETCH, dbp));
386 }
387
388 /*
389 * returns ENOENT, EIO, or 0.
390 *
391 * This interface will allocate a blank spill dbuf when a spill blk
392 * doesn't already exist on the dnode.
393 *
394 * if you only want to find an already existing spill db, then
395 * dmu_spill_hold_existing() should be used.
396 */
397 int
398 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
399 {
400 dmu_buf_impl_t *db = NULL;
401 int err;
402
403 if ((flags & DB_RF_HAVESTRUCT) == 0)
404 rw_enter(&dn->dn_struct_rwlock, RW_READER);
405
406 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
407
408 if ((flags & DB_RF_HAVESTRUCT) == 0)
409 rw_exit(&dn->dn_struct_rwlock);
410
411 if (db == NULL) {
412 *dbp = NULL;
413 return (SET_ERROR(EIO));
414 }
415 err = dbuf_read(db, NULL, flags);
416 if (err == 0)
417 *dbp = &db->db;
418 else {
419 dbuf_rele(db, tag);
420 *dbp = NULL;
421 }
422 return (err);
423 }
424
425 int
426 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
427 {
428 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
429 dnode_t *dn;
430 int err;
431
432 DB_DNODE_ENTER(db);
433 dn = DB_DNODE(db);
434
435 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
436 err = SET_ERROR(EINVAL);
437 } else {
438 rw_enter(&dn->dn_struct_rwlock, RW_READER);
439
440 if (!dn->dn_have_spill) {
441 err = SET_ERROR(ENOENT);
442 } else {
443 err = dmu_spill_hold_by_dnode(dn,
444 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
445 }
446
447 rw_exit(&dn->dn_struct_rwlock);
448 }
449
450 DB_DNODE_EXIT(db);
451 return (err);
452 }
453
454 int
455 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
456 {
457 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
458 dnode_t *dn;
459 int err;
460
461 DB_DNODE_ENTER(db);
462 dn = DB_DNODE(db);
463 err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp);
464 DB_DNODE_EXIT(db);
465
466 return (err);
467 }
468
469 /*
470 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
471 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
472 * and can induce severe lock contention when writing to several files
473 * whose dnodes are in the same block.
474 */
475 static int
476 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
477 boolean_t read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
478 {
479 dmu_buf_t **dbp;
480 uint64_t blkid, nblks, i;
481 uint32_t dbuf_flags;
482 int err;
483 zio_t *zio;
484
485 ASSERT(length <= DMU_MAX_ACCESS);
486
487 /*
488 * Note: We directly notify the prefetch code of this read, so that
489 * we can tell it about the multi-block read. dbuf_read() only knows
490 * about the one block it is accessing.
491 */
492 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT |
493 DB_RF_NOPREFETCH;
494
495 rw_enter(&dn->dn_struct_rwlock, RW_READER);
496 if (dn->dn_datablkshift) {
497 int blkshift = dn->dn_datablkshift;
498 nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) -
499 P2ALIGN(offset, 1ULL << blkshift)) >> blkshift;
500 } else {
501 if (offset + length > dn->dn_datablksz) {
502 zfs_panic_recover("zfs: accessing past end of object "
503 "%llx/%llx (size=%u access=%llu+%llu)",
504 (longlong_t)dn->dn_objset->
505 os_dsl_dataset->ds_object,
506 (longlong_t)dn->dn_object, dn->dn_datablksz,
507 (longlong_t)offset, (longlong_t)length);
508 rw_exit(&dn->dn_struct_rwlock);
509 return (SET_ERROR(EIO));
510 }
511 nblks = 1;
512 }
513 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
514
515 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
516 blkid = dbuf_whichblock(dn, 0, offset);
517 for (i = 0; i < nblks; i++) {
518 dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag);
519 if (db == NULL) {
520 rw_exit(&dn->dn_struct_rwlock);
521 dmu_buf_rele_array(dbp, nblks, tag);
522 zio_nowait(zio);
523 return (SET_ERROR(EIO));
524 }
525
526 /* initiate async i/o */
527 if (read)
528 (void) dbuf_read(db, zio, dbuf_flags);
529 dbp[i] = &db->db;
530 }
531
532 if ((flags & DMU_READ_NO_PREFETCH) == 0 &&
533 DNODE_META_IS_CACHEABLE(dn) && length <= zfetch_array_rd_sz) {
534 dmu_zfetch(&dn->dn_zfetch, blkid, nblks,
535 read && DNODE_IS_CACHEABLE(dn));
536 }
537 rw_exit(&dn->dn_struct_rwlock);
538
539 /* wait for async i/o */
540 err = zio_wait(zio);
541 if (err) {
542 dmu_buf_rele_array(dbp, nblks, tag);
543 return (err);
544 }
545
546 /* wait for other io to complete */
547 if (read) {
548 for (i = 0; i < nblks; i++) {
549 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
550 mutex_enter(&db->db_mtx);
551 while (db->db_state == DB_READ ||
552 db->db_state == DB_FILL)
553 cv_wait(&db->db_changed, &db->db_mtx);
554 if (db->db_state == DB_UNCACHED)
555 err = SET_ERROR(EIO);
556 mutex_exit(&db->db_mtx);
557 if (err) {
558 dmu_buf_rele_array(dbp, nblks, tag);
559 return (err);
560 }
561 }
562 }
563
564 *numbufsp = nblks;
565 *dbpp = dbp;
566 return (0);
567 }
568
569 static int
570 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
571 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
572 {
573 dnode_t *dn;
574 int err;
575
576 err = dnode_hold(os, object, FTAG, &dn);
577 if (err)
578 return (err);
579
580 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
581 numbufsp, dbpp, DMU_READ_PREFETCH);
582
583 dnode_rele(dn, FTAG);
584
585 return (err);
586 }
587
588 int
589 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
590 uint64_t length, boolean_t read, void *tag, int *numbufsp,
591 dmu_buf_t ***dbpp)
592 {
593 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
594 dnode_t *dn;
595 int err;
596
597 DB_DNODE_ENTER(db);
598 dn = DB_DNODE(db);
599 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
600 numbufsp, dbpp, DMU_READ_PREFETCH);
601 DB_DNODE_EXIT(db);
602
603 return (err);
604 }
605
606 void
607 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
608 {
609 int i;
610 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
611
612 if (numbufs == 0)
613 return;
614
615 for (i = 0; i < numbufs; i++) {
616 if (dbp[i])
617 dbuf_rele(dbp[i], tag);
618 }
619
620 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
621 }
622
623 /*
624 * Issue prefetch i/os for the given blocks. If level is greater than 0, the
625 * indirect blocks prefeteched will be those that point to the blocks containing
626 * the data starting at offset, and continuing to offset + len.
627 *
628 * Note that if the indirect blocks above the blocks being prefetched are not
629 * in cache, they will be asychronously read in.
630 */
631 void
632 dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset,
633 uint64_t len, zio_priority_t pri)
634 {
635 dnode_t *dn;
636 uint64_t blkid;
637 int nblks, err;
638
639 if (len == 0) { /* they're interested in the bonus buffer */
640 dn = DMU_META_DNODE(os);
641
642 if (object == 0 || object >= DN_MAX_OBJECT)
643 return;
644
645 rw_enter(&dn->dn_struct_rwlock, RW_READER);
646 blkid = dbuf_whichblock(dn, level,
647 object * sizeof (dnode_phys_t));
648 dbuf_prefetch(dn, level, blkid, pri, 0);
649 rw_exit(&dn->dn_struct_rwlock);
650 return;
651 }
652
653 /*
654 * XXX - Note, if the dnode for the requested object is not
655 * already cached, we will do a *synchronous* read in the
656 * dnode_hold() call. The same is true for any indirects.
657 */
658 err = dnode_hold(os, object, FTAG, &dn);
659 if (err != 0)
660 return;
661
662 rw_enter(&dn->dn_struct_rwlock, RW_READER);
663 /*
664 * offset + len - 1 is the last byte we want to prefetch for, and offset
665 * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the
666 * last block we want to prefetch, and dbuf_whichblock(dn, level,
667 * offset) is the first. Then the number we need to prefetch is the
668 * last - first + 1.
669 */
670 if (level > 0 || dn->dn_datablkshift != 0) {
671 nblks = dbuf_whichblock(dn, level, offset + len - 1) -
672 dbuf_whichblock(dn, level, offset) + 1;
673 } else {
674 nblks = (offset < dn->dn_datablksz);
675 }
676
677 if (nblks != 0) {
678 int i;
679
680 blkid = dbuf_whichblock(dn, level, offset);
681 for (i = 0; i < nblks; i++)
682 dbuf_prefetch(dn, level, blkid + i, pri, 0);
683 }
684
685 rw_exit(&dn->dn_struct_rwlock);
686
687 dnode_rele(dn, FTAG);
688 }
689
690 /*
691 * Get the next "chunk" of file data to free. We traverse the file from
692 * the end so that the file gets shorter over time (if we crashes in the
693 * middle, this will leave us in a better state). We find allocated file
694 * data by simply searching the allocated level 1 indirects.
695 *
696 * On input, *start should be the first offset that does not need to be
697 * freed (e.g. "offset + length"). On return, *start will be the first
698 * offset that should be freed.
699 */
700 static int
701 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum)
702 {
703 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
704 /* bytes of data covered by a level-1 indirect block */
705 uint64_t iblkrange =
706 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
707 uint64_t blks;
708
709 ASSERT3U(minimum, <=, *start);
710
711 if (*start - minimum <= iblkrange * maxblks) {
712 *start = minimum;
713 return (0);
714 }
715 ASSERT(ISP2(iblkrange));
716
717 for (blks = 0; *start > minimum && blks < maxblks; blks++) {
718 int err;
719
720 /*
721 * dnode_next_offset(BACKWARDS) will find an allocated L1
722 * indirect block at or before the input offset. We must
723 * decrement *start so that it is at the end of the region
724 * to search.
725 */
726 (*start)--;
727 err = dnode_next_offset(dn,
728 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
729
730 /* if there are no indirect blocks before start, we are done */
731 if (err == ESRCH) {
732 *start = minimum;
733 break;
734 } else if (err != 0) {
735 return (err);
736 }
737
738 /* set start to the beginning of this L1 indirect */
739 *start = P2ALIGN(*start, iblkrange);
740 }
741 if (*start < minimum)
742 *start = minimum;
743 return (0);
744 }
745
746 /*
747 * If this objset is of type OST_ZFS return true if vfs's unmounted flag is set,
748 * otherwise return false.
749 * Used below in dmu_free_long_range_impl() to enable abort when unmounting
750 */
751 /*ARGSUSED*/
752 static boolean_t
753 dmu_objset_zfs_unmounting(objset_t *os)
754 {
755 #ifdef _KERNEL
756 if (dmu_objset_type(os) == DMU_OST_ZFS)
757 return (zfs_get_vfs_flag_unmounted(os));
758 #endif
759 return (B_FALSE);
760 }
761
762 static int
763 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
764 uint64_t length)
765 {
766 uint64_t object_size;
767 int err;
768 uint64_t dirty_frees_threshold;
769 dsl_pool_t *dp = dmu_objset_pool(os);
770 int t;
771
772 if (dn == NULL)
773 return (SET_ERROR(EINVAL));
774
775 object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
776 if (offset >= object_size)
777 return (0);
778
779 if (zfs_per_txg_dirty_frees_percent <= 100)
780 dirty_frees_threshold =
781 zfs_per_txg_dirty_frees_percent * zfs_dirty_data_max / 100;
782 else
783 dirty_frees_threshold = zfs_dirty_data_max / 4;
784
785 if (length == DMU_OBJECT_END || offset + length > object_size)
786 length = object_size - offset;
787
788 while (length != 0) {
789 uint64_t chunk_end, chunk_begin, chunk_len;
790 uint64_t long_free_dirty_all_txgs = 0;
791 dmu_tx_t *tx;
792
793 if (dmu_objset_zfs_unmounting(dn->dn_objset))
794 return (SET_ERROR(EINTR));
795
796 chunk_end = chunk_begin = offset + length;
797
798 /* move chunk_begin backwards to the beginning of this chunk */
799 err = get_next_chunk(dn, &chunk_begin, offset);
800 if (err)
801 return (err);
802 ASSERT3U(chunk_begin, >=, offset);
803 ASSERT3U(chunk_begin, <=, chunk_end);
804
805 chunk_len = chunk_end - chunk_begin;
806
807 mutex_enter(&dp->dp_lock);
808 for (t = 0; t < TXG_SIZE; t++) {
809 long_free_dirty_all_txgs +=
810 dp->dp_long_free_dirty_pertxg[t];
811 }
812 mutex_exit(&dp->dp_lock);
813
814 /*
815 * To avoid filling up a TXG with just frees wait for
816 * the next TXG to open before freeing more chunks if
817 * we have reached the threshold of frees
818 */
819 if (dirty_frees_threshold != 0 &&
820 long_free_dirty_all_txgs >= dirty_frees_threshold) {
821 txg_wait_open(dp, 0);
822 continue;
823 }
824
825 tx = dmu_tx_create(os);
826 dmu_tx_hold_free(tx, dn->dn_object, chunk_begin, chunk_len);
827
828 /*
829 * Mark this transaction as typically resulting in a net
830 * reduction in space used.
831 */
832 dmu_tx_mark_netfree(tx);
833 err = dmu_tx_assign(tx, TXG_WAIT);
834 if (err) {
835 dmu_tx_abort(tx);
836 return (err);
837 }
838
839 mutex_enter(&dp->dp_lock);
840 dp->dp_long_free_dirty_pertxg[dmu_tx_get_txg(tx) & TXG_MASK] +=
841 chunk_len;
842 mutex_exit(&dp->dp_lock);
843 DTRACE_PROBE3(free__long__range,
844 uint64_t, long_free_dirty_all_txgs, uint64_t, chunk_len,
845 uint64_t, dmu_tx_get_txg(tx));
846 dnode_free_range(dn, chunk_begin, chunk_len, tx);
847 dmu_tx_commit(tx);
848
849 length -= chunk_len;
850 }
851 return (0);
852 }
853
854 int
855 dmu_free_long_range(objset_t *os, uint64_t object,
856 uint64_t offset, uint64_t length)
857 {
858 dnode_t *dn;
859 int err;
860
861 err = dnode_hold(os, object, FTAG, &dn);
862 if (err != 0)
863 return (err);
864 err = dmu_free_long_range_impl(os, dn, offset, length);
865
866 /*
867 * It is important to zero out the maxblkid when freeing the entire
868 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
869 * will take the fast path, and (b) dnode_reallocate() can verify
870 * that the entire file has been freed.
871 */
872 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
873 dn->dn_maxblkid = 0;
874
875 dnode_rele(dn, FTAG);
876 return (err);
877 }
878
879 int
880 dmu_free_long_object(objset_t *os, uint64_t object)
881 {
882 dmu_tx_t *tx;
883 int err;
884
885 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
886 if (err != 0)
887 return (err);
888
889 tx = dmu_tx_create(os);
890 dmu_tx_hold_bonus(tx, object);
891 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
892 dmu_tx_mark_netfree(tx);
893 err = dmu_tx_assign(tx, TXG_WAIT);
894 if (err == 0) {
895 err = dmu_object_free(os, object, tx);
896 dmu_tx_commit(tx);
897 } else {
898 dmu_tx_abort(tx);
899 }
900
901 return (err);
902 }
903
904 int
905 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
906 uint64_t size, dmu_tx_t *tx)
907 {
908 dnode_t *dn;
909 int err = dnode_hold(os, object, FTAG, &dn);
910 if (err)
911 return (err);
912 ASSERT(offset < UINT64_MAX);
913 ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
914 dnode_free_range(dn, offset, size, tx);
915 dnode_rele(dn, FTAG);
916 return (0);
917 }
918
919 static int
920 dmu_read_impl(dnode_t *dn, uint64_t offset, uint64_t size,
921 void *buf, uint32_t flags)
922 {
923 dmu_buf_t **dbp;
924 int numbufs, err = 0;
925
926 /*
927 * Deal with odd block sizes, where there can't be data past the first
928 * block. If we ever do the tail block optimization, we will need to
929 * handle that here as well.
930 */
931 if (dn->dn_maxblkid == 0) {
932 uint64_t newsz = offset > dn->dn_datablksz ? 0 :
933 MIN(size, dn->dn_datablksz - offset);
934 bzero((char *)buf + newsz, size - newsz);
935 size = newsz;
936 }
937
938 while (size > 0) {
939 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
940 int i;
941
942 /*
943 * NB: we could do this block-at-a-time, but it's nice
944 * to be reading in parallel.
945 */
946 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
947 TRUE, FTAG, &numbufs, &dbp, flags);
948 if (err)
949 break;
950
951 for (i = 0; i < numbufs; i++) {
952 uint64_t tocpy;
953 int64_t bufoff;
954 dmu_buf_t *db = dbp[i];
955
956 ASSERT(size > 0);
957
958 bufoff = offset - db->db_offset;
959 tocpy = MIN(db->db_size - bufoff, size);
960
961 (void) memcpy(buf, (char *)db->db_data + bufoff, tocpy);
962
963 offset += tocpy;
964 size -= tocpy;
965 buf = (char *)buf + tocpy;
966 }
967 dmu_buf_rele_array(dbp, numbufs, FTAG);
968 }
969 return (err);
970 }
971
972 int
973 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
974 void *buf, uint32_t flags)
975 {
976 dnode_t *dn;
977 int err;
978
979 err = dnode_hold(os, object, FTAG, &dn);
980 if (err != 0)
981 return (err);
982
983 err = dmu_read_impl(dn, offset, size, buf, flags);
984 dnode_rele(dn, FTAG);
985 return (err);
986 }
987
988 int
989 dmu_read_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, void *buf,
990 uint32_t flags)
991 {
992 return (dmu_read_impl(dn, offset, size, buf, flags));
993 }
994
995 static void
996 dmu_write_impl(dmu_buf_t **dbp, int numbufs, uint64_t offset, uint64_t size,
997 const void *buf, dmu_tx_t *tx)
998 {
999 int i;
1000
1001 for (i = 0; i < numbufs; i++) {
1002 uint64_t tocpy;
1003 int64_t bufoff;
1004 dmu_buf_t *db = dbp[i];
1005
1006 ASSERT(size > 0);
1007
1008 bufoff = offset - db->db_offset;
1009 tocpy = MIN(db->db_size - bufoff, size);
1010
1011 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1012
1013 if (tocpy == db->db_size)
1014 dmu_buf_will_fill(db, tx);
1015 else
1016 dmu_buf_will_dirty(db, tx);
1017
1018 (void) memcpy((char *)db->db_data + bufoff, buf, tocpy);
1019
1020 if (tocpy == db->db_size)
1021 dmu_buf_fill_done(db, tx);
1022
1023 offset += tocpy;
1024 size -= tocpy;
1025 buf = (char *)buf + tocpy;
1026 }
1027 }
1028
1029 void
1030 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1031 const void *buf, dmu_tx_t *tx)
1032 {
1033 dmu_buf_t **dbp;
1034 int numbufs;
1035
1036 if (size == 0)
1037 return;
1038
1039 VERIFY0(dmu_buf_hold_array(os, object, offset, size,
1040 FALSE, FTAG, &numbufs, &dbp));
1041 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1042 dmu_buf_rele_array(dbp, numbufs, FTAG);
1043 }
1044
1045 void
1046 dmu_write_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size,
1047 const void *buf, dmu_tx_t *tx)
1048 {
1049 dmu_buf_t **dbp;
1050 int numbufs;
1051
1052 if (size == 0)
1053 return;
1054
1055 VERIFY0(dmu_buf_hold_array_by_dnode(dn, offset, size,
1056 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH));
1057 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1058 dmu_buf_rele_array(dbp, numbufs, FTAG);
1059 }
1060
1061 void
1062 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1063 dmu_tx_t *tx)
1064 {
1065 dmu_buf_t **dbp;
1066 int numbufs, i;
1067
1068 if (size == 0)
1069 return;
1070
1071 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
1072 FALSE, FTAG, &numbufs, &dbp));
1073
1074 for (i = 0; i < numbufs; i++) {
1075 dmu_buf_t *db = dbp[i];
1076
1077 dmu_buf_will_not_fill(db, tx);
1078 }
1079 dmu_buf_rele_array(dbp, numbufs, FTAG);
1080 }
1081
1082 void
1083 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
1084 void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
1085 int compressed_size, int byteorder, dmu_tx_t *tx)
1086 {
1087 dmu_buf_t *db;
1088
1089 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
1090 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
1091 VERIFY0(dmu_buf_hold_noread(os, object, offset,
1092 FTAG, &db));
1093
1094 dmu_buf_write_embedded(db,
1095 data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
1096 uncompressed_size, compressed_size, byteorder, tx);
1097
1098 dmu_buf_rele(db, FTAG);
1099 }
1100
1101 /*
1102 * DMU support for xuio
1103 */
1104 kstat_t *xuio_ksp = NULL;
1105
1106 typedef struct xuio_stats {
1107 /* loaned yet not returned arc_buf */
1108 kstat_named_t xuiostat_onloan_rbuf;
1109 kstat_named_t xuiostat_onloan_wbuf;
1110 /* whether a copy is made when loaning out a read buffer */
1111 kstat_named_t xuiostat_rbuf_copied;
1112 kstat_named_t xuiostat_rbuf_nocopy;
1113 /* whether a copy is made when assigning a write buffer */
1114 kstat_named_t xuiostat_wbuf_copied;
1115 kstat_named_t xuiostat_wbuf_nocopy;
1116 } xuio_stats_t;
1117
1118 static xuio_stats_t xuio_stats = {
1119 { "onloan_read_buf", KSTAT_DATA_UINT64 },
1120 { "onloan_write_buf", KSTAT_DATA_UINT64 },
1121 { "read_buf_copied", KSTAT_DATA_UINT64 },
1122 { "read_buf_nocopy", KSTAT_DATA_UINT64 },
1123 { "write_buf_copied", KSTAT_DATA_UINT64 },
1124 { "write_buf_nocopy", KSTAT_DATA_UINT64 }
1125 };
1126
1127 #define XUIOSTAT_INCR(stat, val) \
1128 atomic_add_64(&xuio_stats.stat.value.ui64, (val))
1129 #define XUIOSTAT_BUMP(stat) XUIOSTAT_INCR(stat, 1)
1130
1131 #ifdef HAVE_UIO_ZEROCOPY
1132 int
1133 dmu_xuio_init(xuio_t *xuio, int nblk)
1134 {
1135 dmu_xuio_t *priv;
1136 uio_t *uio = &xuio->xu_uio;
1137
1138 uio->uio_iovcnt = nblk;
1139 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
1140
1141 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
1142 priv->cnt = nblk;
1143 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
1144 priv->iovp = (iovec_t *)uio->uio_iov;
1145 XUIO_XUZC_PRIV(xuio) = priv;
1146
1147 if (XUIO_XUZC_RW(xuio) == UIO_READ)
1148 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
1149 else
1150 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
1151
1152 return (0);
1153 }
1154
1155 void
1156 dmu_xuio_fini(xuio_t *xuio)
1157 {
1158 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1159 int nblk = priv->cnt;
1160
1161 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
1162 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
1163 kmem_free(priv, sizeof (dmu_xuio_t));
1164
1165 if (XUIO_XUZC_RW(xuio) == UIO_READ)
1166 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
1167 else
1168 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
1169 }
1170
1171 /*
1172 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
1173 * and increase priv->next by 1.
1174 */
1175 int
1176 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
1177 {
1178 struct iovec *iov;
1179 uio_t *uio = &xuio->xu_uio;
1180 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1181 int i = priv->next++;
1182
1183 ASSERT(i < priv->cnt);
1184 ASSERT(off + n <= arc_buf_lsize(abuf));
1185 iov = (iovec_t *)uio->uio_iov + i;
1186 iov->iov_base = (char *)abuf->b_data + off;
1187 iov->iov_len = n;
1188 priv->bufs[i] = abuf;
1189 return (0);
1190 }
1191
1192 int
1193 dmu_xuio_cnt(xuio_t *xuio)
1194 {
1195 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1196 return (priv->cnt);
1197 }
1198
1199 arc_buf_t *
1200 dmu_xuio_arcbuf(xuio_t *xuio, int i)
1201 {
1202 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1203
1204 ASSERT(i < priv->cnt);
1205 return (priv->bufs[i]);
1206 }
1207
1208 void
1209 dmu_xuio_clear(xuio_t *xuio, int i)
1210 {
1211 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1212
1213 ASSERT(i < priv->cnt);
1214 priv->bufs[i] = NULL;
1215 }
1216 #endif /* HAVE_UIO_ZEROCOPY */
1217
1218 static void
1219 xuio_stat_init(void)
1220 {
1221 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
1222 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
1223 KSTAT_FLAG_VIRTUAL);
1224 if (xuio_ksp != NULL) {
1225 xuio_ksp->ks_data = &xuio_stats;
1226 kstat_install(xuio_ksp);
1227 }
1228 }
1229
1230 static void
1231 xuio_stat_fini(void)
1232 {
1233 if (xuio_ksp != NULL) {
1234 kstat_delete(xuio_ksp);
1235 xuio_ksp = NULL;
1236 }
1237 }
1238
1239 void
1240 xuio_stat_wbuf_copied(void)
1241 {
1242 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1243 }
1244
1245 void
1246 xuio_stat_wbuf_nocopy(void)
1247 {
1248 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
1249 }
1250
1251 #ifdef _KERNEL
1252 int
1253 dmu_read_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size)
1254 {
1255 dmu_buf_t **dbp;
1256 int numbufs, i, err;
1257 #ifdef HAVE_UIO_ZEROCOPY
1258 xuio_t *xuio = NULL;
1259 #endif
1260
1261 /*
1262 * NB: we could do this block-at-a-time, but it's nice
1263 * to be reading in parallel.
1264 */
1265 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1266 TRUE, FTAG, &numbufs, &dbp, 0);
1267 if (err)
1268 return (err);
1269
1270 for (i = 0; i < numbufs; i++) {
1271 uint64_t tocpy;
1272 int64_t bufoff;
1273 dmu_buf_t *db = dbp[i];
1274
1275 ASSERT(size > 0);
1276
1277 bufoff = uio->uio_loffset - db->db_offset;
1278 tocpy = MIN(db->db_size - bufoff, size);
1279
1280 #ifdef HAVE_UIO_ZEROCOPY
1281 if (xuio) {
1282 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
1283 arc_buf_t *dbuf_abuf = dbi->db_buf;
1284 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
1285 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
1286 if (!err) {
1287 uio->uio_resid -= tocpy;
1288 uio->uio_loffset += tocpy;
1289 }
1290
1291 if (abuf == dbuf_abuf)
1292 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
1293 else
1294 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
1295 } else
1296 #endif
1297 err = uiomove((char *)db->db_data + bufoff, tocpy,
1298 UIO_READ, uio);
1299 if (err)
1300 break;
1301
1302 size -= tocpy;
1303 }
1304 dmu_buf_rele_array(dbp, numbufs, FTAG);
1305
1306 return (err);
1307 }
1308
1309 /*
1310 * Read 'size' bytes into the uio buffer.
1311 * From object zdb->db_object.
1312 * Starting at offset uio->uio_loffset.
1313 *
1314 * If the caller already has a dbuf in the target object
1315 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1316 * because we don't have to find the dnode_t for the object.
1317 */
1318 int
1319 dmu_read_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size)
1320 {
1321 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1322 dnode_t *dn;
1323 int err;
1324
1325 if (size == 0)
1326 return (0);
1327
1328 DB_DNODE_ENTER(db);
1329 dn = DB_DNODE(db);
1330 err = dmu_read_uio_dnode(dn, uio, size);
1331 DB_DNODE_EXIT(db);
1332
1333 return (err);
1334 }
1335
1336 /*
1337 * Read 'size' bytes into the uio buffer.
1338 * From the specified object
1339 * Starting at offset uio->uio_loffset.
1340 */
1341 int
1342 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
1343 {
1344 dnode_t *dn;
1345 int err;
1346
1347 if (size == 0)
1348 return (0);
1349
1350 err = dnode_hold(os, object, FTAG, &dn);
1351 if (err)
1352 return (err);
1353
1354 err = dmu_read_uio_dnode(dn, uio, size);
1355
1356 dnode_rele(dn, FTAG);
1357
1358 return (err);
1359 }
1360
1361 int
1362 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1363 {
1364 dmu_buf_t **dbp;
1365 int numbufs;
1366 int err = 0;
1367 int i;
1368
1369 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1370 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1371 if (err)
1372 return (err);
1373
1374 for (i = 0; i < numbufs; i++) {
1375 uint64_t tocpy;
1376 int64_t bufoff;
1377 dmu_buf_t *db = dbp[i];
1378
1379 ASSERT(size > 0);
1380
1381 bufoff = uio->uio_loffset - db->db_offset;
1382 tocpy = MIN(db->db_size - bufoff, size);
1383
1384 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1385
1386 if (tocpy == db->db_size)
1387 dmu_buf_will_fill(db, tx);
1388 else
1389 dmu_buf_will_dirty(db, tx);
1390
1391 /*
1392 * XXX uiomove could block forever (eg.nfs-backed
1393 * pages). There needs to be a uiolockdown() function
1394 * to lock the pages in memory, so that uiomove won't
1395 * block.
1396 */
1397 err = uiomove((char *)db->db_data + bufoff, tocpy,
1398 UIO_WRITE, uio);
1399
1400 if (tocpy == db->db_size)
1401 dmu_buf_fill_done(db, tx);
1402
1403 if (err)
1404 break;
1405
1406 size -= tocpy;
1407 }
1408
1409 dmu_buf_rele_array(dbp, numbufs, FTAG);
1410 return (err);
1411 }
1412
1413 /*
1414 * Write 'size' bytes from the uio buffer.
1415 * To object zdb->db_object.
1416 * Starting at offset uio->uio_loffset.
1417 *
1418 * If the caller already has a dbuf in the target object
1419 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1420 * because we don't have to find the dnode_t for the object.
1421 */
1422 int
1423 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1424 dmu_tx_t *tx)
1425 {
1426 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1427 dnode_t *dn;
1428 int err;
1429
1430 if (size == 0)
1431 return (0);
1432
1433 DB_DNODE_ENTER(db);
1434 dn = DB_DNODE(db);
1435 err = dmu_write_uio_dnode(dn, uio, size, tx);
1436 DB_DNODE_EXIT(db);
1437
1438 return (err);
1439 }
1440
1441 /*
1442 * Write 'size' bytes from the uio buffer.
1443 * To the specified object.
1444 * Starting at offset uio->uio_loffset.
1445 */
1446 int
1447 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1448 dmu_tx_t *tx)
1449 {
1450 dnode_t *dn;
1451 int err;
1452
1453 if (size == 0)
1454 return (0);
1455
1456 err = dnode_hold(os, object, FTAG, &dn);
1457 if (err)
1458 return (err);
1459
1460 err = dmu_write_uio_dnode(dn, uio, size, tx);
1461
1462 dnode_rele(dn, FTAG);
1463
1464 return (err);
1465 }
1466 #endif /* _KERNEL */
1467
1468 /*
1469 * Allocate a loaned anonymous arc buffer.
1470 */
1471 arc_buf_t *
1472 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1473 {
1474 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1475
1476 return (arc_loan_buf(db->db_objset->os_spa, B_FALSE, size));
1477 }
1478
1479 /*
1480 * Free a loaned arc buffer.
1481 */
1482 void
1483 dmu_return_arcbuf(arc_buf_t *buf)
1484 {
1485 arc_return_buf(buf, FTAG);
1486 arc_buf_destroy(buf, FTAG);
1487 }
1488
1489 void
1490 dmu_assign_arcbuf_impl(dmu_buf_t *handle, arc_buf_t *buf, dmu_tx_t *tx)
1491 {
1492 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1493 dbuf_assign_arcbuf(db, buf, tx);
1494 }
1495
1496 void
1497 dmu_convert_to_raw(dmu_buf_t *handle, boolean_t byteorder, const uint8_t *salt,
1498 const uint8_t *iv, const uint8_t *mac, dmu_tx_t *tx)
1499 {
1500 dmu_object_type_t type;
1501 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1502 uint64_t dsobj = dmu_objset_id(db->db_objset);
1503
1504 ASSERT3P(db->db_buf, !=, NULL);
1505 ASSERT3U(dsobj, !=, 0);
1506
1507 dmu_buf_will_change_crypt_params(handle, tx);
1508
1509 DB_DNODE_ENTER(db);
1510 type = DB_DNODE(db)->dn_type;
1511 DB_DNODE_EXIT(db);
1512
1513 /*
1514 * This technically violates the assumption the dmu code makes
1515 * that dnode blocks are only released in syncing context.
1516 */
1517 (void) arc_release(db->db_buf, db);
1518 arc_convert_to_raw(db->db_buf, dsobj, byteorder, type, salt, iv, mac);
1519 }
1520
1521 void
1522 dmu_copy_from_buf(objset_t *os, uint64_t object, uint64_t offset,
1523 dmu_buf_t *handle, dmu_tx_t *tx)
1524 {
1525 dmu_buf_t *dst_handle;
1526 dmu_buf_impl_t *dstdb;
1527 dmu_buf_impl_t *srcdb = (dmu_buf_impl_t *)handle;
1528 arc_buf_t *abuf;
1529 uint64_t datalen;
1530 boolean_t byteorder;
1531 uint8_t salt[ZIO_DATA_SALT_LEN];
1532 uint8_t iv[ZIO_DATA_IV_LEN];
1533 uint8_t mac[ZIO_DATA_MAC_LEN];
1534
1535 ASSERT3P(srcdb->db_buf, !=, NULL);
1536
1537 /* hold the db that we want to write to */
1538 VERIFY0(dmu_buf_hold(os, object, offset, FTAG, &dst_handle,
1539 DMU_READ_NO_DECRYPT));
1540 dstdb = (dmu_buf_impl_t *)dst_handle;
1541 datalen = arc_buf_size(srcdb->db_buf);
1542
1543 /* allocated an arc buffer that matches the type of srcdb->db_buf */
1544 if (arc_is_encrypted(srcdb->db_buf)) {
1545 arc_get_raw_params(srcdb->db_buf, &byteorder, salt, iv, mac);
1546 abuf = arc_loan_raw_buf(os->os_spa, dmu_objset_id(os),
1547 byteorder, salt, iv, mac, DB_DNODE(dstdb)->dn_type,
1548 datalen, arc_buf_lsize(srcdb->db_buf),
1549 arc_get_compression(srcdb->db_buf));
1550 } else {
1551 /* we won't get a compressed db back from dmu_buf_hold() */
1552 ASSERT3U(arc_get_compression(srcdb->db_buf),
1553 ==, ZIO_COMPRESS_OFF);
1554 abuf = arc_loan_buf(os->os_spa,
1555 DMU_OT_IS_METADATA(DB_DNODE(dstdb)->dn_type), datalen);
1556 }
1557
1558 ASSERT3U(datalen, ==, arc_buf_size(abuf));
1559
1560 /* copy the data to the new buffer and assign it to the dstdb */
1561 bcopy(srcdb->db_buf->b_data, abuf->b_data, datalen);
1562 dbuf_assign_arcbuf(dstdb, abuf, tx);
1563 dmu_buf_rele(dst_handle, FTAG);
1564 }
1565
1566 /*
1567 * When possible directly assign passed loaned arc buffer to a dbuf.
1568 * If this is not possible copy the contents of passed arc buf via
1569 * dmu_write().
1570 */
1571 void
1572 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1573 dmu_tx_t *tx)
1574 {
1575 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1576 dnode_t *dn;
1577 dmu_buf_impl_t *db;
1578 uint32_t blksz = (uint32_t)arc_buf_lsize(buf);
1579 uint64_t blkid;
1580
1581 DB_DNODE_ENTER(dbuf);
1582 dn = DB_DNODE(dbuf);
1583 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1584 blkid = dbuf_whichblock(dn, 0, offset);
1585 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1586 rw_exit(&dn->dn_struct_rwlock);
1587 DB_DNODE_EXIT(dbuf);
1588
1589 /*
1590 * We can only assign if the offset is aligned, the arc buf is the
1591 * same size as the dbuf, and the dbuf is not metadata.
1592 */
1593 if (offset == db->db.db_offset && blksz == db->db.db_size) {
1594 dbuf_assign_arcbuf(db, buf, tx);
1595 dbuf_rele(db, FTAG);
1596 } else {
1597 objset_t *os;
1598 uint64_t object;
1599
1600 /* compressed bufs must always be assignable to their dbuf */
1601 ASSERT3U(arc_get_compression(buf), ==, ZIO_COMPRESS_OFF);
1602 ASSERT(!(buf->b_flags & ARC_BUF_FLAG_COMPRESSED));
1603
1604 DB_DNODE_ENTER(dbuf);
1605 dn = DB_DNODE(dbuf);
1606 os = dn->dn_objset;
1607 object = dn->dn_object;
1608 DB_DNODE_EXIT(dbuf);
1609
1610 dbuf_rele(db, FTAG);
1611 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1612 dmu_return_arcbuf(buf);
1613 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1614 }
1615 }
1616
1617 typedef struct {
1618 dbuf_dirty_record_t *dsa_dr;
1619 dmu_sync_cb_t *dsa_done;
1620 zgd_t *dsa_zgd;
1621 dmu_tx_t *dsa_tx;
1622 } dmu_sync_arg_t;
1623
1624 /* ARGSUSED */
1625 static void
1626 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1627 {
1628 dmu_sync_arg_t *dsa = varg;
1629 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1630 blkptr_t *bp = zio->io_bp;
1631
1632 if (zio->io_error == 0) {
1633 if (BP_IS_HOLE(bp)) {
1634 /*
1635 * A block of zeros may compress to a hole, but the
1636 * block size still needs to be known for replay.
1637 */
1638 BP_SET_LSIZE(bp, db->db_size);
1639 } else if (!BP_IS_EMBEDDED(bp)) {
1640 ASSERT(BP_GET_LEVEL(bp) == 0);
1641 BP_SET_FILL(bp, 1);
1642 }
1643 }
1644 }
1645
1646 static void
1647 dmu_sync_late_arrival_ready(zio_t *zio)
1648 {
1649 dmu_sync_ready(zio, NULL, zio->io_private);
1650 }
1651
1652 /* ARGSUSED */
1653 static void
1654 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1655 {
1656 dmu_sync_arg_t *dsa = varg;
1657 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1658 dmu_buf_impl_t *db = dr->dr_dbuf;
1659
1660 mutex_enter(&db->db_mtx);
1661 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1662 if (zio->io_error == 0) {
1663 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1664 if (dr->dt.dl.dr_nopwrite) {
1665 blkptr_t *bp = zio->io_bp;
1666 blkptr_t *bp_orig = &zio->io_bp_orig;
1667 uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1668
1669 ASSERT(BP_EQUAL(bp, bp_orig));
1670 VERIFY(BP_EQUAL(bp, db->db_blkptr));
1671 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1672 VERIFY(zio_checksum_table[chksum].ci_flags &
1673 ZCHECKSUM_FLAG_NOPWRITE);
1674 }
1675 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1676 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1677 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1678
1679 /*
1680 * Old style holes are filled with all zeros, whereas
1681 * new-style holes maintain their lsize, type, level,
1682 * and birth time (see zio_write_compress). While we
1683 * need to reset the BP_SET_LSIZE() call that happened
1684 * in dmu_sync_ready for old style holes, we do *not*
1685 * want to wipe out the information contained in new
1686 * style holes. Thus, only zero out the block pointer if
1687 * it's an old style hole.
1688 */
1689 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) &&
1690 dr->dt.dl.dr_overridden_by.blk_birth == 0)
1691 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1692 } else {
1693 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1694 }
1695 cv_broadcast(&db->db_changed);
1696 mutex_exit(&db->db_mtx);
1697
1698 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1699
1700 kmem_free(dsa, sizeof (*dsa));
1701 }
1702
1703 static void
1704 dmu_sync_late_arrival_done(zio_t *zio)
1705 {
1706 blkptr_t *bp = zio->io_bp;
1707 dmu_sync_arg_t *dsa = zio->io_private;
1708 ASSERTV(blkptr_t *bp_orig = &zio->io_bp_orig);
1709
1710 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1711 ASSERT(!(zio->io_flags & ZIO_FLAG_NOPWRITE));
1712 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1713 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1714 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1715 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1716 }
1717
1718 dmu_tx_commit(dsa->dsa_tx);
1719
1720 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1721
1722 abd_put(zio->io_abd);
1723 kmem_free(dsa, sizeof (*dsa));
1724 }
1725
1726 static int
1727 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1728 zio_prop_t *zp, zbookmark_phys_t *zb)
1729 {
1730 dmu_sync_arg_t *dsa;
1731 dmu_tx_t *tx;
1732
1733 tx = dmu_tx_create(os);
1734 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1735 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1736 dmu_tx_abort(tx);
1737 /* Make zl_get_data do txg_waited_synced() */
1738 return (SET_ERROR(EIO));
1739 }
1740
1741 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1742 dsa->dsa_dr = NULL;
1743 dsa->dsa_done = done;
1744 dsa->dsa_zgd = zgd;
1745 dsa->dsa_tx = tx;
1746
1747 /*
1748 * Since we are currently syncing this txg, it's nontrivial to
1749 * determine what BP to nopwrite against, so we disable nopwrite.
1750 *
1751 * When syncing, the db_blkptr is initially the BP of the previous
1752 * txg. We can not nopwrite against it because it will be changed
1753 * (this is similar to the non-late-arrival case where the dbuf is
1754 * dirty in a future txg).
1755 *
1756 * Then dbuf_write_ready() sets bp_blkptr to the location we will write.
1757 * We can not nopwrite against it because although the BP will not
1758 * (typically) be changed, the data has not yet been persisted to this
1759 * location.
1760 *
1761 * Finally, when dbuf_write_done() is called, it is theoretically
1762 * possible to always nopwrite, because the data that was written in
1763 * this txg is the same data that we are trying to write. However we
1764 * would need to check that this dbuf is not dirty in any future
1765 * txg's (as we do in the normal dmu_sync() path). For simplicity, we
1766 * don't nopwrite in this case.
1767 */
1768 zp->zp_nopwrite = B_FALSE;
1769
1770 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1771 abd_get_from_buf(zgd->zgd_db->db_data, zgd->zgd_db->db_size),
1772 zgd->zgd_db->db_size, zgd->zgd_db->db_size, zp,
1773 dmu_sync_late_arrival_ready, NULL, NULL, dmu_sync_late_arrival_done,
1774 dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1775
1776 return (0);
1777 }
1778
1779 /*
1780 * Intent log support: sync the block associated with db to disk.
1781 * N.B. and XXX: the caller is responsible for making sure that the
1782 * data isn't changing while dmu_sync() is writing it.
1783 *
1784 * Return values:
1785 *
1786 * EEXIST: this txg has already been synced, so there's nothing to do.
1787 * The caller should not log the write.
1788 *
1789 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1790 * The caller should not log the write.
1791 *
1792 * EALREADY: this block is already in the process of being synced.
1793 * The caller should track its progress (somehow).
1794 *
1795 * EIO: could not do the I/O.
1796 * The caller should do a txg_wait_synced().
1797 *
1798 * 0: the I/O has been initiated.
1799 * The caller should log this blkptr in the done callback.
1800 * It is possible that the I/O will fail, in which case
1801 * the error will be reported to the done callback and
1802 * propagated to pio from zio_done().
1803 */
1804 int
1805 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1806 {
1807 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1808 objset_t *os = db->db_objset;
1809 dsl_dataset_t *ds = os->os_dsl_dataset;
1810 dbuf_dirty_record_t *dr;
1811 dmu_sync_arg_t *dsa;
1812 zbookmark_phys_t zb;
1813 zio_prop_t zp;
1814 dnode_t *dn;
1815
1816 ASSERT(pio != NULL);
1817 ASSERT(txg != 0);
1818
1819 /* dbuf is within the locked range */
1820 ASSERT3U(db->db.db_offset, >=, zgd->zgd_rl->r_off);
1821 ASSERT3U(db->db.db_offset + db->db.db_size, <=,
1822 zgd->zgd_rl->r_off + zgd->zgd_rl->r_len);
1823
1824 SET_BOOKMARK(&zb, ds->ds_object,
1825 db->db.db_object, db->db_level, db->db_blkid);
1826
1827 DB_DNODE_ENTER(db);
1828 dn = DB_DNODE(db);
1829 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1830 DB_DNODE_EXIT(db);
1831
1832 /*
1833 * If we're frozen (running ziltest), we always need to generate a bp.
1834 */
1835 if (txg > spa_freeze_txg(os->os_spa))
1836 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1837
1838 /*
1839 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1840 * and us. If we determine that this txg is not yet syncing,
1841 * but it begins to sync a moment later, that's OK because the
1842 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1843 */
1844 mutex_enter(&db->db_mtx);
1845
1846 if (txg <= spa_last_synced_txg(os->os_spa)) {
1847 /*
1848 * This txg has already synced. There's nothing to do.
1849 */
1850 mutex_exit(&db->db_mtx);
1851 return (SET_ERROR(EEXIST));
1852 }
1853
1854 if (txg <= spa_syncing_txg(os->os_spa)) {
1855 /*
1856 * This txg is currently syncing, so we can't mess with
1857 * the dirty record anymore; just write a new log block.
1858 */
1859 mutex_exit(&db->db_mtx);
1860 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1861 }
1862
1863 dr = db->db_last_dirty;
1864 while (dr && dr->dr_txg != txg)
1865 dr = dr->dr_next;
1866
1867 if (dr == NULL) {
1868 /*
1869 * There's no dr for this dbuf, so it must have been freed.
1870 * There's no need to log writes to freed blocks, so we're done.
1871 */
1872 mutex_exit(&db->db_mtx);
1873 return (SET_ERROR(ENOENT));
1874 }
1875
1876 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
1877
1878 if (db->db_blkptr != NULL) {
1879 /*
1880 * We need to fill in zgd_bp with the current blkptr so that
1881 * the nopwrite code can check if we're writing the same
1882 * data that's already on disk. We can only nopwrite if we
1883 * are sure that after making the copy, db_blkptr will not
1884 * change until our i/o completes. We ensure this by
1885 * holding the db_mtx, and only allowing nopwrite if the
1886 * block is not already dirty (see below). This is verified
1887 * by dmu_sync_done(), which VERIFYs that the db_blkptr has
1888 * not changed.
1889 */
1890 *zgd->zgd_bp = *db->db_blkptr;
1891 }
1892
1893 /*
1894 * Assume the on-disk data is X, the current syncing data (in
1895 * txg - 1) is Y, and the current in-memory data is Z (currently
1896 * in dmu_sync).
1897 *
1898 * We usually want to perform a nopwrite if X and Z are the
1899 * same. However, if Y is different (i.e. the BP is going to
1900 * change before this write takes effect), then a nopwrite will
1901 * be incorrect - we would override with X, which could have
1902 * been freed when Y was written.
1903 *
1904 * (Note that this is not a concern when we are nop-writing from
1905 * syncing context, because X and Y must be identical, because
1906 * all previous txgs have been synced.)
1907 *
1908 * Therefore, we disable nopwrite if the current BP could change
1909 * before this TXG. There are two ways it could change: by
1910 * being dirty (dr_next is non-NULL), or by being freed
1911 * (dnode_block_freed()). This behavior is verified by
1912 * zio_done(), which VERIFYs that the override BP is identical
1913 * to the on-disk BP.
1914 */
1915 DB_DNODE_ENTER(db);
1916 dn = DB_DNODE(db);
1917 if (dr->dr_next != NULL || dnode_block_freed(dn, db->db_blkid))
1918 zp.zp_nopwrite = B_FALSE;
1919 DB_DNODE_EXIT(db);
1920
1921 ASSERT(dr->dr_txg == txg);
1922 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1923 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1924 /*
1925 * We have already issued a sync write for this buffer,
1926 * or this buffer has already been synced. It could not
1927 * have been dirtied since, or we would have cleared the state.
1928 */
1929 mutex_exit(&db->db_mtx);
1930 return (SET_ERROR(EALREADY));
1931 }
1932
1933 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1934 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1935 mutex_exit(&db->db_mtx);
1936
1937 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1938 dsa->dsa_dr = dr;
1939 dsa->dsa_done = done;
1940 dsa->dsa_zgd = zgd;
1941 dsa->dsa_tx = NULL;
1942
1943 zio_nowait(arc_write(pio, os->os_spa, txg,
1944 zgd->zgd_bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
1945 &zp, dmu_sync_ready, NULL, NULL, dmu_sync_done, dsa,
1946 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb));
1947
1948 return (0);
1949 }
1950
1951 int
1952 dmu_object_set_nlevels(objset_t *os, uint64_t object, int nlevels, dmu_tx_t *tx)
1953 {
1954 dnode_t *dn;
1955 int err;
1956
1957 err = dnode_hold(os, object, FTAG, &dn);
1958 if (err)
1959 return (err);
1960 err = dnode_set_nlevels(dn, nlevels, tx);
1961 dnode_rele(dn, FTAG);
1962 return (err);
1963 }
1964
1965 int
1966 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1967 dmu_tx_t *tx)
1968 {
1969 dnode_t *dn;
1970 int err;
1971
1972 err = dnode_hold(os, object, FTAG, &dn);
1973 if (err)
1974 return (err);
1975 err = dnode_set_blksz(dn, size, ibs, tx);
1976 dnode_rele(dn, FTAG);
1977 return (err);
1978 }
1979
1980 void
1981 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1982 dmu_tx_t *tx)
1983 {
1984 dnode_t *dn;
1985
1986 /*
1987 * Send streams include each object's checksum function. This
1988 * check ensures that the receiving system can understand the
1989 * checksum function transmitted.
1990 */
1991 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS);
1992
1993 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1994 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS);
1995 dn->dn_checksum = checksum;
1996 dnode_setdirty(dn, tx);
1997 dnode_rele(dn, FTAG);
1998 }
1999
2000 void
2001 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
2002 dmu_tx_t *tx)
2003 {
2004 dnode_t *dn;
2005
2006 /*
2007 * Send streams include each object's compression function. This
2008 * check ensures that the receiving system can understand the
2009 * compression function transmitted.
2010 */
2011 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS);
2012
2013 VERIFY0(dnode_hold(os, object, FTAG, &dn));
2014 dn->dn_compress = compress;
2015 dnode_setdirty(dn, tx);
2016 dnode_rele(dn, FTAG);
2017 }
2018
2019 int zfs_mdcomp_disable = 0;
2020
2021 /*
2022 * When the "redundant_metadata" property is set to "most", only indirect
2023 * blocks of this level and higher will have an additional ditto block.
2024 */
2025 int zfs_redundant_metadata_most_ditto_level = 2;
2026
2027 void
2028 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
2029 {
2030 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
2031 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
2032 (wp & WP_SPILL));
2033 enum zio_checksum checksum = os->os_checksum;
2034 enum zio_compress compress = os->os_compress;
2035 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
2036 boolean_t dedup = B_FALSE;
2037 boolean_t nopwrite = B_FALSE;
2038 boolean_t dedup_verify = os->os_dedup_verify;
2039 boolean_t encrypt = B_FALSE;
2040 int copies = os->os_copies;
2041
2042 /*
2043 * We maintain different write policies for each of the following
2044 * types of data:
2045 * 1. metadata
2046 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
2047 * 3. all other level 0 blocks
2048 */
2049 if (ismd) {
2050 if (zfs_mdcomp_disable) {
2051 compress = ZIO_COMPRESS_EMPTY;
2052 } else {
2053 /*
2054 * XXX -- we should design a compression algorithm
2055 * that specializes in arrays of bps.
2056 */
2057 compress = zio_compress_select(os->os_spa,
2058 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON);
2059 }
2060
2061 /*
2062 * Metadata always gets checksummed. If the data
2063 * checksum is multi-bit correctable, and it's not a
2064 * ZBT-style checksum, then it's suitable for metadata
2065 * as well. Otherwise, the metadata checksum defaults
2066 * to fletcher4.
2067 */
2068 if (!(zio_checksum_table[checksum].ci_flags &
2069 ZCHECKSUM_FLAG_METADATA) ||
2070 (zio_checksum_table[checksum].ci_flags &
2071 ZCHECKSUM_FLAG_EMBEDDED))
2072 checksum = ZIO_CHECKSUM_FLETCHER_4;
2073
2074 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL ||
2075 (os->os_redundant_metadata ==
2076 ZFS_REDUNDANT_METADATA_MOST &&
2077 (level >= zfs_redundant_metadata_most_ditto_level ||
2078 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))))
2079 copies++;
2080 } else if (wp & WP_NOFILL) {
2081 ASSERT(level == 0);
2082
2083 /*
2084 * If we're writing preallocated blocks, we aren't actually
2085 * writing them so don't set any policy properties. These
2086 * blocks are currently only used by an external subsystem
2087 * outside of zfs (i.e. dump) and not written by the zio
2088 * pipeline.
2089 */
2090 compress = ZIO_COMPRESS_OFF;
2091 checksum = ZIO_CHECKSUM_OFF;
2092 } else {
2093 compress = zio_compress_select(os->os_spa, dn->dn_compress,
2094 compress);
2095
2096 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
2097 zio_checksum_select(dn->dn_checksum, checksum) :
2098 dedup_checksum;
2099
2100 /*
2101 * Determine dedup setting. If we are in dmu_sync(),
2102 * we won't actually dedup now because that's all
2103 * done in syncing context; but we do want to use the
2104 * dedup checkum. If the checksum is not strong
2105 * enough to ensure unique signatures, force
2106 * dedup_verify.
2107 */
2108 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
2109 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
2110 if (!(zio_checksum_table[checksum].ci_flags &
2111 ZCHECKSUM_FLAG_DEDUP))
2112 dedup_verify = B_TRUE;
2113 }
2114
2115 /*
2116 * Enable nopwrite if we have secure enough checksum
2117 * algorithm (see comment in zio_nop_write) and
2118 * compression is enabled. We don't enable nopwrite if
2119 * dedup is enabled as the two features are mutually
2120 * exclusive.
2121 */
2122 nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags &
2123 ZCHECKSUM_FLAG_NOPWRITE) &&
2124 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
2125 }
2126
2127 /*
2128 * All objects in an encrypted objset are protected from modification
2129 * via a MAC. Encrypted objects store their IV and salt in the last DVA
2130 * in the bp, so we cannot use all copies. Encrypted objects are also
2131 * not subject to nopwrite since writing the same data will still
2132 * result in a new ciphertext. Only encrypted blocks can be dedup'd
2133 * to avoid ambiguity in the dedup code since the DDT does not store
2134 * object types.
2135 */
2136 if (os->os_encrypted && (wp & WP_NOFILL) == 0) {
2137 encrypt = B_TRUE;
2138
2139 if (DMU_OT_IS_ENCRYPTED(type)) {
2140 copies = MIN(copies, SPA_DVAS_PER_BP - 1);
2141 nopwrite = B_FALSE;
2142 } else {
2143 dedup = B_FALSE;
2144 }
2145
2146 if (type == DMU_OT_DNODE || type == DMU_OT_OBJSET)
2147 compress = ZIO_COMPRESS_EMPTY;
2148 }
2149
2150 zp->zp_compress = compress;
2151 zp->zp_checksum = checksum;
2152 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
2153 zp->zp_level = level;
2154 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
2155 zp->zp_dedup = dedup;
2156 zp->zp_dedup_verify = dedup && dedup_verify;
2157 zp->zp_nopwrite = nopwrite;
2158 zp->zp_encrypt = encrypt;
2159 zp->zp_byteorder = ZFS_HOST_BYTEORDER;
2160 bzero(zp->zp_salt, ZIO_DATA_SALT_LEN);
2161 bzero(zp->zp_iv, ZIO_DATA_IV_LEN);
2162 bzero(zp->zp_mac, ZIO_DATA_MAC_LEN);
2163
2164 ASSERT3U(zp->zp_compress, !=, ZIO_COMPRESS_INHERIT);
2165 }
2166
2167 /*
2168 * This function is only called from zfs_holey_common() for zpl_llseek()
2169 * in order to determine the location of holes. In order to accurately
2170 * report holes all dirty data must be synced to disk. This causes extremely
2171 * poor performance when seeking for holes in a dirty file. As a compromise,
2172 * only provide hole data when the dnode is clean. When a dnode is dirty
2173 * report the dnode as having no holes which is always a safe thing to do.
2174 */
2175 int
2176 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
2177 {
2178 dnode_t *dn;
2179 int i, err;
2180 boolean_t clean = B_TRUE;
2181
2182 err = dnode_hold(os, object, FTAG, &dn);
2183 if (err)
2184 return (err);
2185
2186 /*
2187 * Check if dnode is dirty
2188 */
2189 if (dn->dn_dirtyctx != DN_UNDIRTIED) {
2190 for (i = 0; i < TXG_SIZE; i++) {
2191 if (!list_is_empty(&dn->dn_dirty_records[i])) {
2192 clean = B_FALSE;
2193 break;
2194 }
2195 }
2196 }
2197
2198 /*
2199 * If compatibility option is on, sync any current changes before
2200 * we go trundling through the block pointers.
2201 */
2202 if (!clean && zfs_dmu_offset_next_sync) {
2203 clean = B_TRUE;
2204 dnode_rele(dn, FTAG);
2205 txg_wait_synced(dmu_objset_pool(os), 0);
2206 err = dnode_hold(os, object, FTAG, &dn);
2207 if (err)
2208 return (err);
2209 }
2210
2211 if (clean)
2212 err = dnode_next_offset(dn,
2213 (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
2214 else
2215 err = SET_ERROR(EBUSY);
2216
2217 dnode_rele(dn, FTAG);
2218
2219 return (err);
2220 }
2221
2222 void
2223 __dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2224 {
2225 dnode_phys_t *dnp = dn->dn_phys;
2226 int i;
2227
2228 doi->doi_data_block_size = dn->dn_datablksz;
2229 doi->doi_metadata_block_size = dn->dn_indblkshift ?
2230 1ULL << dn->dn_indblkshift : 0;
2231 doi->doi_type = dn->dn_type;
2232 doi->doi_bonus_type = dn->dn_bonustype;
2233 doi->doi_bonus_size = dn->dn_bonuslen;
2234 doi->doi_dnodesize = dn->dn_num_slots << DNODE_SHIFT;
2235 doi->doi_indirection = dn->dn_nlevels;
2236 doi->doi_checksum = dn->dn_checksum;
2237 doi->doi_compress = dn->dn_compress;
2238 doi->doi_nblkptr = dn->dn_nblkptr;
2239 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
2240 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
2241 doi->doi_fill_count = 0;
2242 for (i = 0; i < dnp->dn_nblkptr; i++)
2243 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
2244 }
2245
2246 void
2247 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2248 {
2249 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2250 mutex_enter(&dn->dn_mtx);
2251
2252 __dmu_object_info_from_dnode(dn, doi);
2253
2254 mutex_exit(&dn->dn_mtx);
2255 rw_exit(&dn->dn_struct_rwlock);
2256 }
2257
2258 /*
2259 * Get information on a DMU object.
2260 * If doi is NULL, just indicates whether the object exists.
2261 */
2262 int
2263 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
2264 {
2265 dnode_t *dn;
2266 int err = dnode_hold(os, object, FTAG, &dn);
2267
2268 if (err)
2269 return (err);
2270
2271 if (doi != NULL)
2272 dmu_object_info_from_dnode(dn, doi);
2273
2274 dnode_rele(dn, FTAG);
2275 return (0);
2276 }
2277
2278 /*
2279 * As above, but faster; can be used when you have a held dbuf in hand.
2280 */
2281 void
2282 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
2283 {
2284 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2285
2286 DB_DNODE_ENTER(db);
2287 dmu_object_info_from_dnode(DB_DNODE(db), doi);
2288 DB_DNODE_EXIT(db);
2289 }
2290
2291 /*
2292 * Faster still when you only care about the size.
2293 * This is specifically optimized for zfs_getattr().
2294 */
2295 void
2296 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
2297 u_longlong_t *nblk512)
2298 {
2299 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2300 dnode_t *dn;
2301
2302 DB_DNODE_ENTER(db);
2303 dn = DB_DNODE(db);
2304
2305 *blksize = dn->dn_datablksz;
2306 /* add in number of slots used for the dnode itself */
2307 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
2308 SPA_MINBLOCKSHIFT) + dn->dn_num_slots;
2309 DB_DNODE_EXIT(db);
2310 }
2311
2312 void
2313 dmu_object_dnsize_from_db(dmu_buf_t *db_fake, int *dnsize)
2314 {
2315 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2316 dnode_t *dn;
2317
2318 DB_DNODE_ENTER(db);
2319 dn = DB_DNODE(db);
2320 *dnsize = dn->dn_num_slots << DNODE_SHIFT;
2321 DB_DNODE_EXIT(db);
2322 }
2323
2324 void
2325 byteswap_uint64_array(void *vbuf, size_t size)
2326 {
2327 uint64_t *buf = vbuf;
2328 size_t count = size >> 3;
2329 int i;
2330
2331 ASSERT((size & 7) == 0);
2332
2333 for (i = 0; i < count; i++)
2334 buf[i] = BSWAP_64(buf[i]);
2335 }
2336
2337 void
2338 byteswap_uint32_array(void *vbuf, size_t size)
2339 {
2340 uint32_t *buf = vbuf;
2341 size_t count = size >> 2;
2342 int i;
2343
2344 ASSERT((size & 3) == 0);
2345
2346 for (i = 0; i < count; i++)
2347 buf[i] = BSWAP_32(buf[i]);
2348 }
2349
2350 void
2351 byteswap_uint16_array(void *vbuf, size_t size)
2352 {
2353 uint16_t *buf = vbuf;
2354 size_t count = size >> 1;
2355 int i;
2356
2357 ASSERT((size & 1) == 0);
2358
2359 for (i = 0; i < count; i++)
2360 buf[i] = BSWAP_16(buf[i]);
2361 }
2362
2363 /* ARGSUSED */
2364 void
2365 byteswap_uint8_array(void *vbuf, size_t size)
2366 {
2367 }
2368
2369 void
2370 dmu_init(void)
2371 {
2372 abd_init();
2373 zfs_dbgmsg_init();
2374 sa_cache_init();
2375 xuio_stat_init();
2376 dmu_objset_init();
2377 dnode_init();
2378 zfetch_init();
2379 dmu_tx_init();
2380 l2arc_init();
2381 arc_init();
2382 dbuf_init();
2383 }
2384
2385 void
2386 dmu_fini(void)
2387 {
2388 arc_fini(); /* arc depends on l2arc, so arc must go first */
2389 l2arc_fini();
2390 dmu_tx_fini();
2391 zfetch_fini();
2392 dbuf_fini();
2393 dnode_fini();
2394 dmu_objset_fini();
2395 xuio_stat_fini();
2396 sa_cache_fini();
2397 zfs_dbgmsg_fini();
2398 abd_fini();
2399 }
2400
2401 #if defined(_KERNEL) && defined(HAVE_SPL)
2402 EXPORT_SYMBOL(dmu_bonus_hold);
2403 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus);
2404 EXPORT_SYMBOL(dmu_buf_rele_array);
2405 EXPORT_SYMBOL(dmu_prefetch);
2406 EXPORT_SYMBOL(dmu_free_range);
2407 EXPORT_SYMBOL(dmu_free_long_range);
2408 EXPORT_SYMBOL(dmu_free_long_object);
2409 EXPORT_SYMBOL(dmu_read);
2410 EXPORT_SYMBOL(dmu_read_by_dnode);
2411 EXPORT_SYMBOL(dmu_write);
2412 EXPORT_SYMBOL(dmu_write_by_dnode);
2413 EXPORT_SYMBOL(dmu_prealloc);
2414 EXPORT_SYMBOL(dmu_object_info);
2415 EXPORT_SYMBOL(dmu_object_info_from_dnode);
2416 EXPORT_SYMBOL(dmu_object_info_from_db);
2417 EXPORT_SYMBOL(dmu_object_size_from_db);
2418 EXPORT_SYMBOL(dmu_object_dnsize_from_db);
2419 EXPORT_SYMBOL(dmu_object_set_nlevels);
2420 EXPORT_SYMBOL(dmu_object_set_blocksize);
2421 EXPORT_SYMBOL(dmu_object_set_checksum);
2422 EXPORT_SYMBOL(dmu_object_set_compress);
2423 EXPORT_SYMBOL(dmu_write_policy);
2424 EXPORT_SYMBOL(dmu_sync);
2425 EXPORT_SYMBOL(dmu_request_arcbuf);
2426 EXPORT_SYMBOL(dmu_return_arcbuf);
2427 EXPORT_SYMBOL(dmu_assign_arcbuf);
2428 EXPORT_SYMBOL(dmu_buf_hold);
2429 EXPORT_SYMBOL(dmu_ot);
2430
2431 /* BEGIN CSTYLED */
2432 module_param(zfs_mdcomp_disable, int, 0644);
2433 MODULE_PARM_DESC(zfs_mdcomp_disable, "Disable meta data compression");
2434
2435 module_param(zfs_nopwrite_enabled, int, 0644);
2436 MODULE_PARM_DESC(zfs_nopwrite_enabled, "Enable NOP writes");
2437
2438 module_param(zfs_per_txg_dirty_frees_percent, ulong, 0644);
2439 MODULE_PARM_DESC(zfs_per_txg_dirty_frees_percent,
2440 "percentage of dirtied blocks from frees in one TXG");
2441
2442 module_param(zfs_dmu_offset_next_sync, int, 0644);
2443 MODULE_PARM_DESC(zfs_dmu_offset_next_sync,
2444 "Enable forcing txg sync to find holes");
2445
2446 /* END CSTYLED */
2447
2448 #endif
ZFS On Linux mirror for PVE