]> 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 error handling incallers of dbuf_hold_level()
[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 /*
77 * This can be used for testing, to ensure that certain actions happen
78 * while in the middle of a remap (which might otherwise complete too
79 * quickly). Used by ztest(8).
80 */
81 int zfs_object_remap_one_indirect_delay_ms = 0;
82
83 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
84 {DMU_BSWAP_UINT8, TRUE, FALSE, FALSE, "unallocated" },
85 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "object directory" },
86 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "object array" },
87 {DMU_BSWAP_UINT8, TRUE, FALSE, FALSE, "packed nvlist" },
88 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "packed nvlist size" },
89 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "bpobj" },
90 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "bpobj header" },
91 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "SPA space map header" },
92 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "SPA space map" },
93 {DMU_BSWAP_UINT64, TRUE, FALSE, TRUE, "ZIL intent log" },
94 {DMU_BSWAP_DNODE, TRUE, FALSE, TRUE, "DMU dnode" },
95 {DMU_BSWAP_OBJSET, TRUE, TRUE, FALSE, "DMU objset" },
96 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "DSL directory" },
97 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL directory child map"},
98 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL dataset snap map" },
99 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL props" },
100 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "DSL dataset" },
101 {DMU_BSWAP_ZNODE, TRUE, FALSE, FALSE, "ZFS znode" },
102 {DMU_BSWAP_OLDACL, TRUE, FALSE, TRUE, "ZFS V0 ACL" },
103 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "ZFS plain file" },
104 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS directory" },
105 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "ZFS master node" },
106 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS delete queue" },
107 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "zvol object" },
108 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "zvol prop" },
109 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "other uint8[]" },
110 {DMU_BSWAP_UINT64, FALSE, FALSE, TRUE, "other uint64[]" },
111 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "other ZAP" },
112 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "persistent error log" },
113 {DMU_BSWAP_UINT8, TRUE, FALSE, FALSE, "SPA history" },
114 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "SPA history offsets" },
115 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "Pool properties" },
116 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL permissions" },
117 {DMU_BSWAP_ACL, TRUE, FALSE, TRUE, "ZFS ACL" },
118 {DMU_BSWAP_UINT8, TRUE, FALSE, TRUE, "ZFS SYSACL" },
119 {DMU_BSWAP_UINT8, TRUE, FALSE, TRUE, "FUID table" },
120 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "FUID table size" },
121 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL dataset next clones"},
122 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "scan work queue" },
123 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS user/group/project used" },
124 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS user/group/project quota"},
125 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "snapshot refcount tags"},
126 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "DDT ZAP algorithm" },
127 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "DDT statistics" },
128 {DMU_BSWAP_UINT8, TRUE, FALSE, TRUE, "System attributes" },
129 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "SA master node" },
130 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "SA attr registration" },
131 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "SA attr layouts" },
132 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "scan translations" },
133 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "deduplicated block" },
134 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL deadlist map" },
135 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "DSL deadlist map hdr" },
136 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL dir clones" },
137 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "bpobj subobj" }
138 };
139
140 const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
141 { byteswap_uint8_array, "uint8" },
142 { byteswap_uint16_array, "uint16" },
143 { byteswap_uint32_array, "uint32" },
144 { byteswap_uint64_array, "uint64" },
145 { zap_byteswap, "zap" },
146 { dnode_buf_byteswap, "dnode" },
147 { dmu_objset_byteswap, "objset" },
148 { zfs_znode_byteswap, "znode" },
149 { zfs_oldacl_byteswap, "oldacl" },
150 { zfs_acl_byteswap, "acl" }
151 };
152
153 int
154 dmu_buf_hold_noread_by_dnode(dnode_t *dn, uint64_t offset,
155 void *tag, dmu_buf_t **dbp)
156 {
157 uint64_t blkid;
158 dmu_buf_impl_t *db;
159
160 blkid = dbuf_whichblock(dn, 0, offset);
161 rw_enter(&dn->dn_struct_rwlock, RW_READER);
162 db = dbuf_hold(dn, blkid, tag);
163 rw_exit(&dn->dn_struct_rwlock);
164
165 if (db == NULL) {
166 *dbp = NULL;
167 return (SET_ERROR(EIO));
168 }
169
170 *dbp = &db->db;
171 return (0);
172 }
173 int
174 dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset,
175 void *tag, dmu_buf_t **dbp)
176 {
177 dnode_t *dn;
178 uint64_t blkid;
179 dmu_buf_impl_t *db;
180 int err;
181
182 err = dnode_hold(os, object, FTAG, &dn);
183 if (err)
184 return (err);
185 blkid = dbuf_whichblock(dn, 0, offset);
186 rw_enter(&dn->dn_struct_rwlock, RW_READER);
187 db = dbuf_hold(dn, blkid, tag);
188 rw_exit(&dn->dn_struct_rwlock);
189 dnode_rele(dn, FTAG);
190
191 if (db == NULL) {
192 *dbp = NULL;
193 return (SET_ERROR(EIO));
194 }
195
196 *dbp = &db->db;
197 return (err);
198 }
199
200 int
201 dmu_buf_hold_by_dnode(dnode_t *dn, uint64_t offset,
202 void *tag, dmu_buf_t **dbp, int flags)
203 {
204 int err;
205 int db_flags = DB_RF_CANFAIL;
206
207 if (flags & DMU_READ_NO_PREFETCH)
208 db_flags |= DB_RF_NOPREFETCH;
209 if (flags & DMU_READ_NO_DECRYPT)
210 db_flags |= DB_RF_NO_DECRYPT;
211
212 err = dmu_buf_hold_noread_by_dnode(dn, offset, tag, dbp);
213 if (err == 0) {
214 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
215 err = dbuf_read(db, NULL, db_flags);
216 if (err != 0) {
217 dbuf_rele(db, tag);
218 *dbp = NULL;
219 }
220 }
221
222 return (err);
223 }
224
225 int
226 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
227 void *tag, dmu_buf_t **dbp, int flags)
228 {
229 int err;
230 int db_flags = DB_RF_CANFAIL;
231
232 if (flags & DMU_READ_NO_PREFETCH)
233 db_flags |= DB_RF_NOPREFETCH;
234 if (flags & DMU_READ_NO_DECRYPT)
235 db_flags |= DB_RF_NO_DECRYPT;
236
237 err = dmu_buf_hold_noread(os, object, offset, tag, dbp);
238 if (err == 0) {
239 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
240 err = dbuf_read(db, NULL, db_flags);
241 if (err != 0) {
242 dbuf_rele(db, tag);
243 *dbp = NULL;
244 }
245 }
246
247 return (err);
248 }
249
250 int
251 dmu_bonus_max(void)
252 {
253 return (DN_OLD_MAX_BONUSLEN);
254 }
255
256 int
257 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
258 {
259 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
260 dnode_t *dn;
261 int error;
262
263 DB_DNODE_ENTER(db);
264 dn = DB_DNODE(db);
265
266 if (dn->dn_bonus != db) {
267 error = SET_ERROR(EINVAL);
268 } else if (newsize < 0 || newsize > db_fake->db_size) {
269 error = SET_ERROR(EINVAL);
270 } else {
271 dnode_setbonuslen(dn, newsize, tx);
272 error = 0;
273 }
274
275 DB_DNODE_EXIT(db);
276 return (error);
277 }
278
279 int
280 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
281 {
282 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
283 dnode_t *dn;
284 int error;
285
286 DB_DNODE_ENTER(db);
287 dn = DB_DNODE(db);
288
289 if (!DMU_OT_IS_VALID(type)) {
290 error = SET_ERROR(EINVAL);
291 } else if (dn->dn_bonus != db) {
292 error = SET_ERROR(EINVAL);
293 } else {
294 dnode_setbonus_type(dn, type, tx);
295 error = 0;
296 }
297
298 DB_DNODE_EXIT(db);
299 return (error);
300 }
301
302 dmu_object_type_t
303 dmu_get_bonustype(dmu_buf_t *db_fake)
304 {
305 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
306 dnode_t *dn;
307 dmu_object_type_t type;
308
309 DB_DNODE_ENTER(db);
310 dn = DB_DNODE(db);
311 type = dn->dn_bonustype;
312 DB_DNODE_EXIT(db);
313
314 return (type);
315 }
316
317 int
318 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
319 {
320 dnode_t *dn;
321 int error;
322
323 error = dnode_hold(os, object, FTAG, &dn);
324 dbuf_rm_spill(dn, tx);
325 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
326 dnode_rm_spill(dn, tx);
327 rw_exit(&dn->dn_struct_rwlock);
328 dnode_rele(dn, FTAG);
329 return (error);
330 }
331
332 /*
333 * Lookup and hold the bonus buffer for the provided dnode. If the dnode
334 * has not yet been allocated a new bonus dbuf a will be allocated.
335 * Returns ENOENT, EIO, or 0.
336 */
337 int dmu_bonus_hold_by_dnode(dnode_t *dn, void *tag, dmu_buf_t **dbp,
338 uint32_t flags)
339 {
340 dmu_buf_impl_t *db;
341 int error;
342 uint32_t db_flags = DB_RF_MUST_SUCCEED;
343
344 if (flags & DMU_READ_NO_PREFETCH)
345 db_flags |= DB_RF_NOPREFETCH;
346 if (flags & DMU_READ_NO_DECRYPT)
347 db_flags |= DB_RF_NO_DECRYPT;
348
349 rw_enter(&dn->dn_struct_rwlock, RW_READER);
350 if (dn->dn_bonus == NULL) {
351 rw_exit(&dn->dn_struct_rwlock);
352 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
353 if (dn->dn_bonus == NULL)
354 dbuf_create_bonus(dn);
355 }
356 db = dn->dn_bonus;
357
358 /* as long as the bonus buf is held, the dnode will be held */
359 if (zfs_refcount_add(&db->db_holds, tag) == 1) {
360 VERIFY(dnode_add_ref(dn, db));
361 atomic_inc_32(&dn->dn_dbufs_count);
362 }
363
364 /*
365 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
366 * hold and incrementing the dbuf count to ensure that dnode_move() sees
367 * a dnode hold for every dbuf.
368 */
369 rw_exit(&dn->dn_struct_rwlock);
370
371 error = dbuf_read(db, NULL, db_flags);
372 if (error) {
373 dnode_evict_bonus(dn);
374 dbuf_rele(db, tag);
375 *dbp = NULL;
376 return (error);
377 }
378
379 *dbp = &db->db;
380 return (0);
381 }
382
383 int
384 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
385 {
386 dnode_t *dn;
387 int error;
388
389 error = dnode_hold(os, object, FTAG, &dn);
390 if (error)
391 return (error);
392
393 error = dmu_bonus_hold_by_dnode(dn, tag, dbp, DMU_READ_NO_PREFETCH);
394 dnode_rele(dn, FTAG);
395
396 return (error);
397 }
398
399 /*
400 * returns ENOENT, EIO, or 0.
401 *
402 * This interface will allocate a blank spill dbuf when a spill blk
403 * doesn't already exist on the dnode.
404 *
405 * if you only want to find an already existing spill db, then
406 * dmu_spill_hold_existing() should be used.
407 */
408 int
409 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
410 {
411 dmu_buf_impl_t *db = NULL;
412 int err;
413
414 if ((flags & DB_RF_HAVESTRUCT) == 0)
415 rw_enter(&dn->dn_struct_rwlock, RW_READER);
416
417 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
418
419 if ((flags & DB_RF_HAVESTRUCT) == 0)
420 rw_exit(&dn->dn_struct_rwlock);
421
422 if (db == NULL) {
423 *dbp = NULL;
424 return (SET_ERROR(EIO));
425 }
426 err = dbuf_read(db, NULL, flags);
427 if (err == 0)
428 *dbp = &db->db;
429 else {
430 dbuf_rele(db, tag);
431 *dbp = NULL;
432 }
433 return (err);
434 }
435
436 int
437 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
438 {
439 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
440 dnode_t *dn;
441 int err;
442
443 DB_DNODE_ENTER(db);
444 dn = DB_DNODE(db);
445
446 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
447 err = SET_ERROR(EINVAL);
448 } else {
449 rw_enter(&dn->dn_struct_rwlock, RW_READER);
450
451 if (!dn->dn_have_spill) {
452 err = SET_ERROR(ENOENT);
453 } else {
454 err = dmu_spill_hold_by_dnode(dn,
455 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
456 }
457
458 rw_exit(&dn->dn_struct_rwlock);
459 }
460
461 DB_DNODE_EXIT(db);
462 return (err);
463 }
464
465 int
466 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, uint32_t flags, void *tag,
467 dmu_buf_t **dbp)
468 {
469 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
470 dnode_t *dn;
471 int err;
472 uint32_t db_flags = DB_RF_CANFAIL;
473
474 if (flags & DMU_READ_NO_DECRYPT)
475 db_flags |= DB_RF_NO_DECRYPT;
476
477 DB_DNODE_ENTER(db);
478 dn = DB_DNODE(db);
479 err = dmu_spill_hold_by_dnode(dn, db_flags, tag, dbp);
480 DB_DNODE_EXIT(db);
481
482 return (err);
483 }
484
485 /*
486 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
487 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
488 * and can induce severe lock contention when writing to several files
489 * whose dnodes are in the same block.
490 */
491 static int
492 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
493 boolean_t read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
494 {
495 dmu_buf_t **dbp;
496 uint64_t blkid, nblks, i;
497 uint32_t dbuf_flags;
498 int err;
499 zio_t *zio;
500
501 ASSERT(length <= DMU_MAX_ACCESS);
502
503 /*
504 * Note: We directly notify the prefetch code of this read, so that
505 * we can tell it about the multi-block read. dbuf_read() only knows
506 * about the one block it is accessing.
507 */
508 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT |
509 DB_RF_NOPREFETCH;
510
511 rw_enter(&dn->dn_struct_rwlock, RW_READER);
512 if (dn->dn_datablkshift) {
513 int blkshift = dn->dn_datablkshift;
514 nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) -
515 P2ALIGN(offset, 1ULL << blkshift)) >> blkshift;
516 } else {
517 if (offset + length > dn->dn_datablksz) {
518 zfs_panic_recover("zfs: accessing past end of object "
519 "%llx/%llx (size=%u access=%llu+%llu)",
520 (longlong_t)dn->dn_objset->
521 os_dsl_dataset->ds_object,
522 (longlong_t)dn->dn_object, dn->dn_datablksz,
523 (longlong_t)offset, (longlong_t)length);
524 rw_exit(&dn->dn_struct_rwlock);
525 return (SET_ERROR(EIO));
526 }
527 nblks = 1;
528 }
529 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
530
531 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
532 blkid = dbuf_whichblock(dn, 0, offset);
533 for (i = 0; i < nblks; i++) {
534 dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag);
535 if (db == NULL) {
536 rw_exit(&dn->dn_struct_rwlock);
537 dmu_buf_rele_array(dbp, nblks, tag);
538 zio_nowait(zio);
539 return (SET_ERROR(EIO));
540 }
541
542 /* initiate async i/o */
543 if (read)
544 (void) dbuf_read(db, zio, dbuf_flags);
545 dbp[i] = &db->db;
546 }
547
548 if ((flags & DMU_READ_NO_PREFETCH) == 0 &&
549 DNODE_META_IS_CACHEABLE(dn) && length <= zfetch_array_rd_sz) {
550 dmu_zfetch(&dn->dn_zfetch, blkid, nblks,
551 read && DNODE_IS_CACHEABLE(dn));
552 }
553 rw_exit(&dn->dn_struct_rwlock);
554
555 /* wait for async i/o */
556 err = zio_wait(zio);
557 if (err) {
558 dmu_buf_rele_array(dbp, nblks, tag);
559 return (err);
560 }
561
562 /* wait for other io to complete */
563 if (read) {
564 for (i = 0; i < nblks; i++) {
565 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
566 mutex_enter(&db->db_mtx);
567 while (db->db_state == DB_READ ||
568 db->db_state == DB_FILL)
569 cv_wait(&db->db_changed, &db->db_mtx);
570 if (db->db_state == DB_UNCACHED)
571 err = SET_ERROR(EIO);
572 mutex_exit(&db->db_mtx);
573 if (err) {
574 dmu_buf_rele_array(dbp, nblks, tag);
575 return (err);
576 }
577 }
578 }
579
580 *numbufsp = nblks;
581 *dbpp = dbp;
582 return (0);
583 }
584
585 static int
586 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
587 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
588 {
589 dnode_t *dn;
590 int err;
591
592 err = dnode_hold(os, object, FTAG, &dn);
593 if (err)
594 return (err);
595
596 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
597 numbufsp, dbpp, DMU_READ_PREFETCH);
598
599 dnode_rele(dn, FTAG);
600
601 return (err);
602 }
603
604 int
605 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
606 uint64_t length, boolean_t read, void *tag, int *numbufsp,
607 dmu_buf_t ***dbpp)
608 {
609 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
610 dnode_t *dn;
611 int err;
612
613 DB_DNODE_ENTER(db);
614 dn = DB_DNODE(db);
615 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
616 numbufsp, dbpp, DMU_READ_PREFETCH);
617 DB_DNODE_EXIT(db);
618
619 return (err);
620 }
621
622 void
623 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
624 {
625 int i;
626 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
627
628 if (numbufs == 0)
629 return;
630
631 for (i = 0; i < numbufs; i++) {
632 if (dbp[i])
633 dbuf_rele(dbp[i], tag);
634 }
635
636 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
637 }
638
639 /*
640 * Issue prefetch i/os for the given blocks. If level is greater than 0, the
641 * indirect blocks prefeteched will be those that point to the blocks containing
642 * the data starting at offset, and continuing to offset + len.
643 *
644 * Note that if the indirect blocks above the blocks being prefetched are not
645 * in cache, they will be asychronously read in.
646 */
647 void
648 dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset,
649 uint64_t len, zio_priority_t pri)
650 {
651 dnode_t *dn;
652 uint64_t blkid;
653 int nblks, err;
654
655 if (len == 0) { /* they're interested in the bonus buffer */
656 dn = DMU_META_DNODE(os);
657
658 if (object == 0 || object >= DN_MAX_OBJECT)
659 return;
660
661 rw_enter(&dn->dn_struct_rwlock, RW_READER);
662 blkid = dbuf_whichblock(dn, level,
663 object * sizeof (dnode_phys_t));
664 dbuf_prefetch(dn, level, blkid, pri, 0);
665 rw_exit(&dn->dn_struct_rwlock);
666 return;
667 }
668
669 /*
670 * XXX - Note, if the dnode for the requested object is not
671 * already cached, we will do a *synchronous* read in the
672 * dnode_hold() call. The same is true for any indirects.
673 */
674 err = dnode_hold(os, object, FTAG, &dn);
675 if (err != 0)
676 return;
677
678 rw_enter(&dn->dn_struct_rwlock, RW_READER);
679 /*
680 * offset + len - 1 is the last byte we want to prefetch for, and offset
681 * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the
682 * last block we want to prefetch, and dbuf_whichblock(dn, level,
683 * offset) is the first. Then the number we need to prefetch is the
684 * last - first + 1.
685 */
686 if (level > 0 || dn->dn_datablkshift != 0) {
687 nblks = dbuf_whichblock(dn, level, offset + len - 1) -
688 dbuf_whichblock(dn, level, offset) + 1;
689 } else {
690 nblks = (offset < dn->dn_datablksz);
691 }
692
693 if (nblks != 0) {
694 blkid = dbuf_whichblock(dn, level, offset);
695 for (int i = 0; i < nblks; i++)
696 dbuf_prefetch(dn, level, blkid + i, pri, 0);
697 }
698
699 rw_exit(&dn->dn_struct_rwlock);
700
701 dnode_rele(dn, FTAG);
702 }
703
704 /*
705 * Get the next "chunk" of file data to free. We traverse the file from
706 * the end so that the file gets shorter over time (if we crashes in the
707 * middle, this will leave us in a better state). We find allocated file
708 * data by simply searching the allocated level 1 indirects.
709 *
710 * On input, *start should be the first offset that does not need to be
711 * freed (e.g. "offset + length"). On return, *start will be the first
712 * offset that should be freed.
713 */
714 static int
715 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum)
716 {
717 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
718 /* bytes of data covered by a level-1 indirect block */
719 uint64_t iblkrange =
720 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
721
722 ASSERT3U(minimum, <=, *start);
723
724 if (*start - minimum <= iblkrange * maxblks) {
725 *start = minimum;
726 return (0);
727 }
728 ASSERT(ISP2(iblkrange));
729
730 for (uint64_t blks = 0; *start > minimum && blks < maxblks; blks++) {
731 int err;
732
733 /*
734 * dnode_next_offset(BACKWARDS) will find an allocated L1
735 * indirect block at or before the input offset. We must
736 * decrement *start so that it is at the end of the region
737 * to search.
738 */
739 (*start)--;
740 err = dnode_next_offset(dn,
741 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
742
743 /* if there are no indirect blocks before start, we are done */
744 if (err == ESRCH) {
745 *start = minimum;
746 break;
747 } else if (err != 0) {
748 return (err);
749 }
750
751 /* set start to the beginning of this L1 indirect */
752 *start = P2ALIGN(*start, iblkrange);
753 }
754 if (*start < minimum)
755 *start = minimum;
756 return (0);
757 }
758
759 /*
760 * If this objset is of type OST_ZFS return true if vfs's unmounted flag is set,
761 * otherwise return false.
762 * Used below in dmu_free_long_range_impl() to enable abort when unmounting
763 */
764 /*ARGSUSED*/
765 static boolean_t
766 dmu_objset_zfs_unmounting(objset_t *os)
767 {
768 #ifdef _KERNEL
769 if (dmu_objset_type(os) == DMU_OST_ZFS)
770 return (zfs_get_vfs_flag_unmounted(os));
771 #endif
772 return (B_FALSE);
773 }
774
775 static int
776 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
777 uint64_t length)
778 {
779 uint64_t object_size;
780 int err;
781 uint64_t dirty_frees_threshold;
782 dsl_pool_t *dp = dmu_objset_pool(os);
783
784 if (dn == NULL)
785 return (SET_ERROR(EINVAL));
786
787 object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
788 if (offset >= object_size)
789 return (0);
790
791 if (zfs_per_txg_dirty_frees_percent <= 100)
792 dirty_frees_threshold =
793 zfs_per_txg_dirty_frees_percent * zfs_dirty_data_max / 100;
794 else
795 dirty_frees_threshold = zfs_dirty_data_max / 4;
796
797 if (length == DMU_OBJECT_END || offset + length > object_size)
798 length = object_size - offset;
799
800 while (length != 0) {
801 uint64_t chunk_end, chunk_begin, chunk_len;
802 uint64_t long_free_dirty_all_txgs = 0;
803 dmu_tx_t *tx;
804
805 if (dmu_objset_zfs_unmounting(dn->dn_objset))
806 return (SET_ERROR(EINTR));
807
808 chunk_end = chunk_begin = offset + length;
809
810 /* move chunk_begin backwards to the beginning of this chunk */
811 err = get_next_chunk(dn, &chunk_begin, offset);
812 if (err)
813 return (err);
814 ASSERT3U(chunk_begin, >=, offset);
815 ASSERT3U(chunk_begin, <=, chunk_end);
816
817 chunk_len = chunk_end - chunk_begin;
818
819 mutex_enter(&dp->dp_lock);
820 for (int t = 0; t < TXG_SIZE; t++) {
821 long_free_dirty_all_txgs +=
822 dp->dp_long_free_dirty_pertxg[t];
823 }
824 mutex_exit(&dp->dp_lock);
825
826 /*
827 * To avoid filling up a TXG with just frees wait for
828 * the next TXG to open before freeing more chunks if
829 * we have reached the threshold of frees
830 */
831 if (dirty_frees_threshold != 0 &&
832 long_free_dirty_all_txgs >= dirty_frees_threshold) {
833 DMU_TX_STAT_BUMP(dmu_tx_dirty_frees_delay);
834 txg_wait_open(dp, 0);
835 continue;
836 }
837
838 tx = dmu_tx_create(os);
839 dmu_tx_hold_free(tx, dn->dn_object, chunk_begin, chunk_len);
840
841 /*
842 * Mark this transaction as typically resulting in a net
843 * reduction in space used.
844 */
845 dmu_tx_mark_netfree(tx);
846 err = dmu_tx_assign(tx, TXG_WAIT);
847 if (err) {
848 dmu_tx_abort(tx);
849 return (err);
850 }
851
852 mutex_enter(&dp->dp_lock);
853 dp->dp_long_free_dirty_pertxg[dmu_tx_get_txg(tx) & TXG_MASK] +=
854 chunk_len;
855 mutex_exit(&dp->dp_lock);
856 DTRACE_PROBE3(free__long__range,
857 uint64_t, long_free_dirty_all_txgs, uint64_t, chunk_len,
858 uint64_t, dmu_tx_get_txg(tx));
859 dnode_free_range(dn, chunk_begin, chunk_len, tx);
860
861 dmu_tx_commit(tx);
862
863 length -= chunk_len;
864 }
865 return (0);
866 }
867
868 int
869 dmu_free_long_range(objset_t *os, uint64_t object,
870 uint64_t offset, uint64_t length)
871 {
872 dnode_t *dn;
873 int err;
874
875 err = dnode_hold(os, object, FTAG, &dn);
876 if (err != 0)
877 return (err);
878 err = dmu_free_long_range_impl(os, dn, offset, length);
879
880 /*
881 * It is important to zero out the maxblkid when freeing the entire
882 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
883 * will take the fast path, and (b) dnode_reallocate() can verify
884 * that the entire file has been freed.
885 */
886 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
887 dn->dn_maxblkid = 0;
888
889 dnode_rele(dn, FTAG);
890 return (err);
891 }
892
893 int
894 dmu_free_long_object(objset_t *os, uint64_t object)
895 {
896 dmu_tx_t *tx;
897 int err;
898
899 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
900 if (err != 0)
901 return (err);
902
903 tx = dmu_tx_create(os);
904 dmu_tx_hold_bonus(tx, object);
905 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
906 dmu_tx_mark_netfree(tx);
907 err = dmu_tx_assign(tx, TXG_WAIT);
908 if (err == 0) {
909 if (err == 0)
910 err = dmu_object_free(os, object, tx);
911
912 dmu_tx_commit(tx);
913 } else {
914 dmu_tx_abort(tx);
915 }
916
917 return (err);
918 }
919
920 int
921 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
922 uint64_t size, dmu_tx_t *tx)
923 {
924 dnode_t *dn;
925 int err = dnode_hold(os, object, FTAG, &dn);
926 if (err)
927 return (err);
928 ASSERT(offset < UINT64_MAX);
929 ASSERT(size == DMU_OBJECT_END || size <= UINT64_MAX - offset);
930 dnode_free_range(dn, offset, size, tx);
931 dnode_rele(dn, FTAG);
932 return (0);
933 }
934
935 static int
936 dmu_read_impl(dnode_t *dn, uint64_t offset, uint64_t size,
937 void *buf, uint32_t flags)
938 {
939 dmu_buf_t **dbp;
940 int numbufs, err = 0;
941
942 /*
943 * Deal with odd block sizes, where there can't be data past the first
944 * block. If we ever do the tail block optimization, we will need to
945 * handle that here as well.
946 */
947 if (dn->dn_maxblkid == 0) {
948 uint64_t newsz = offset > dn->dn_datablksz ? 0 :
949 MIN(size, dn->dn_datablksz - offset);
950 bzero((char *)buf + newsz, size - newsz);
951 size = newsz;
952 }
953
954 while (size > 0) {
955 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
956 int i;
957
958 /*
959 * NB: we could do this block-at-a-time, but it's nice
960 * to be reading in parallel.
961 */
962 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
963 TRUE, FTAG, &numbufs, &dbp, flags);
964 if (err)
965 break;
966
967 for (i = 0; i < numbufs; i++) {
968 uint64_t tocpy;
969 int64_t bufoff;
970 dmu_buf_t *db = dbp[i];
971
972 ASSERT(size > 0);
973
974 bufoff = offset - db->db_offset;
975 tocpy = MIN(db->db_size - bufoff, size);
976
977 (void) memcpy(buf, (char *)db->db_data + bufoff, tocpy);
978
979 offset += tocpy;
980 size -= tocpy;
981 buf = (char *)buf + tocpy;
982 }
983 dmu_buf_rele_array(dbp, numbufs, FTAG);
984 }
985 return (err);
986 }
987
988 int
989 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
990 void *buf, uint32_t flags)
991 {
992 dnode_t *dn;
993 int err;
994
995 err = dnode_hold(os, object, FTAG, &dn);
996 if (err != 0)
997 return (err);
998
999 err = dmu_read_impl(dn, offset, size, buf, flags);
1000 dnode_rele(dn, FTAG);
1001 return (err);
1002 }
1003
1004 int
1005 dmu_read_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, void *buf,
1006 uint32_t flags)
1007 {
1008 return (dmu_read_impl(dn, offset, size, buf, flags));
1009 }
1010
1011 static void
1012 dmu_write_impl(dmu_buf_t **dbp, int numbufs, uint64_t offset, uint64_t size,
1013 const void *buf, dmu_tx_t *tx)
1014 {
1015 int i;
1016
1017 for (i = 0; i < numbufs; i++) {
1018 uint64_t tocpy;
1019 int64_t bufoff;
1020 dmu_buf_t *db = dbp[i];
1021
1022 ASSERT(size > 0);
1023
1024 bufoff = offset - db->db_offset;
1025 tocpy = MIN(db->db_size - bufoff, size);
1026
1027 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1028
1029 if (tocpy == db->db_size)
1030 dmu_buf_will_fill(db, tx);
1031 else
1032 dmu_buf_will_dirty(db, tx);
1033
1034 (void) memcpy((char *)db->db_data + bufoff, buf, tocpy);
1035
1036 if (tocpy == db->db_size)
1037 dmu_buf_fill_done(db, tx);
1038
1039 offset += tocpy;
1040 size -= tocpy;
1041 buf = (char *)buf + tocpy;
1042 }
1043 }
1044
1045 void
1046 dmu_write(objset_t *os, uint64_t object, 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(os, object, offset, size,
1056 FALSE, FTAG, &numbufs, &dbp));
1057 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1058 dmu_buf_rele_array(dbp, numbufs, FTAG);
1059 }
1060
1061 void
1062 dmu_write_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size,
1063 const void *buf, dmu_tx_t *tx)
1064 {
1065 dmu_buf_t **dbp;
1066 int numbufs;
1067
1068 if (size == 0)
1069 return;
1070
1071 VERIFY0(dmu_buf_hold_array_by_dnode(dn, offset, size,
1072 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH));
1073 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1074 dmu_buf_rele_array(dbp, numbufs, FTAG);
1075 }
1076
1077 static int
1078 dmu_object_remap_one_indirect(objset_t *os, dnode_t *dn,
1079 uint64_t last_removal_txg, uint64_t offset)
1080 {
1081 uint64_t l1blkid = dbuf_whichblock(dn, 1, offset);
1082 dnode_t *dn_tx;
1083 int err = 0;
1084
1085 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1086 dmu_buf_impl_t *dbuf = dbuf_hold_level(dn, 1, l1blkid, FTAG);
1087 ASSERT3P(dbuf, !=, NULL);
1088
1089 /*
1090 * If the block hasn't been written yet, this default will ensure
1091 * we don't try to remap it.
1092 */
1093 uint64_t birth = UINT64_MAX;
1094 ASSERT3U(last_removal_txg, !=, UINT64_MAX);
1095 if (dbuf->db_blkptr != NULL)
1096 birth = dbuf->db_blkptr->blk_birth;
1097 rw_exit(&dn->dn_struct_rwlock);
1098
1099 /*
1100 * If this L1 was already written after the last removal, then we've
1101 * already tried to remap it. An additional hold is taken after the
1102 * dmu_tx_assign() to handle the case where the dnode is freed while
1103 * waiting for the next open txg.
1104 */
1105 if (birth <= last_removal_txg &&
1106 dbuf_read(dbuf, NULL, DB_RF_MUST_SUCCEED) == 0 &&
1107 dbuf_can_remap(dbuf)) {
1108 dmu_tx_t *tx = dmu_tx_create(os);
1109 dmu_tx_hold_remap_l1indirect(tx, dn->dn_object);
1110 err = dmu_tx_assign(tx, TXG_WAIT);
1111 if (err == 0) {
1112 err = dnode_hold(os, dn->dn_object, FTAG, &dn_tx);
1113 if (err == 0) {
1114 (void) dbuf_dirty(dbuf, tx);
1115 dnode_rele(dn_tx, FTAG);
1116 }
1117 dmu_tx_commit(tx);
1118 } else {
1119 dmu_tx_abort(tx);
1120 }
1121 }
1122
1123 dbuf_rele(dbuf, FTAG);
1124
1125 delay(MSEC_TO_TICK(zfs_object_remap_one_indirect_delay_ms));
1126
1127 return (err);
1128 }
1129
1130 /*
1131 * Remap all blockpointers in the object, if possible, so that they reference
1132 * only concrete vdevs.
1133 *
1134 * To do this, iterate over the L0 blockpointers and remap any that reference
1135 * an indirect vdev. Note that we only examine L0 blockpointers; since we
1136 * cannot guarantee that we can remap all blockpointer anyways (due to split
1137 * blocks), we do not want to make the code unnecessarily complicated to
1138 * catch the unlikely case that there is an L1 block on an indirect vdev that
1139 * contains no indirect blockpointers.
1140 */
1141 int
1142 dmu_object_remap_indirects(objset_t *os, uint64_t object,
1143 uint64_t last_removal_txg)
1144 {
1145 uint64_t offset, l1span;
1146 int err;
1147 dnode_t *dn, *dn_tx;
1148
1149 err = dnode_hold(os, object, FTAG, &dn);
1150 if (err != 0) {
1151 return (err);
1152 }
1153
1154 if (dn->dn_nlevels <= 1) {
1155 if (issig(JUSTLOOKING) && issig(FORREAL)) {
1156 err = SET_ERROR(EINTR);
1157 }
1158
1159 /*
1160 * If the dnode has no indirect blocks, we cannot dirty them.
1161 * We still want to remap the blkptr(s) in the dnode if
1162 * appropriate, so mark it as dirty. An additional hold is
1163 * taken after the dmu_tx_assign() to handle the case where
1164 * the dnode is freed while waiting for the next open txg.
1165 */
1166 if (err == 0 && dnode_needs_remap(dn)) {
1167 dmu_tx_t *tx = dmu_tx_create(os);
1168 dmu_tx_hold_bonus(tx, object);
1169 err = dmu_tx_assign(tx, TXG_WAIT);
1170 if (err == 0) {
1171 err = dnode_hold(os, object, FTAG, &dn_tx);
1172 if (err == 0) {
1173 dnode_setdirty(dn_tx, tx);
1174 dnode_rele(dn_tx, FTAG);
1175 }
1176 dmu_tx_commit(tx);
1177 } else {
1178 dmu_tx_abort(tx);
1179 }
1180 }
1181
1182 dnode_rele(dn, FTAG);
1183 return (err);
1184 }
1185
1186 offset = 0;
1187 l1span = 1ULL << (dn->dn_indblkshift - SPA_BLKPTRSHIFT +
1188 dn->dn_datablkshift);
1189 /*
1190 * Find the next L1 indirect that is not a hole.
1191 */
1192 while (dnode_next_offset(dn, 0, &offset, 2, 1, 0) == 0) {
1193 if (issig(JUSTLOOKING) && issig(FORREAL)) {
1194 err = SET_ERROR(EINTR);
1195 break;
1196 }
1197 if ((err = dmu_object_remap_one_indirect(os, dn,
1198 last_removal_txg, offset)) != 0) {
1199 break;
1200 }
1201 offset += l1span;
1202 }
1203
1204 dnode_rele(dn, FTAG);
1205 return (err);
1206 }
1207
1208 void
1209 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1210 dmu_tx_t *tx)
1211 {
1212 dmu_buf_t **dbp;
1213 int numbufs, i;
1214
1215 if (size == 0)
1216 return;
1217
1218 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
1219 FALSE, FTAG, &numbufs, &dbp));
1220
1221 for (i = 0; i < numbufs; i++) {
1222 dmu_buf_t *db = dbp[i];
1223
1224 dmu_buf_will_not_fill(db, tx);
1225 }
1226 dmu_buf_rele_array(dbp, numbufs, FTAG);
1227 }
1228
1229 void
1230 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
1231 void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
1232 int compressed_size, int byteorder, dmu_tx_t *tx)
1233 {
1234 dmu_buf_t *db;
1235
1236 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
1237 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
1238 VERIFY0(dmu_buf_hold_noread(os, object, offset,
1239 FTAG, &db));
1240
1241 dmu_buf_write_embedded(db,
1242 data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
1243 uncompressed_size, compressed_size, byteorder, tx);
1244
1245 dmu_buf_rele(db, FTAG);
1246 }
1247
1248 /*
1249 * DMU support for xuio
1250 */
1251 kstat_t *xuio_ksp = NULL;
1252
1253 typedef struct xuio_stats {
1254 /* loaned yet not returned arc_buf */
1255 kstat_named_t xuiostat_onloan_rbuf;
1256 kstat_named_t xuiostat_onloan_wbuf;
1257 /* whether a copy is made when loaning out a read buffer */
1258 kstat_named_t xuiostat_rbuf_copied;
1259 kstat_named_t xuiostat_rbuf_nocopy;
1260 /* whether a copy is made when assigning a write buffer */
1261 kstat_named_t xuiostat_wbuf_copied;
1262 kstat_named_t xuiostat_wbuf_nocopy;
1263 } xuio_stats_t;
1264
1265 static xuio_stats_t xuio_stats = {
1266 { "onloan_read_buf", KSTAT_DATA_UINT64 },
1267 { "onloan_write_buf", KSTAT_DATA_UINT64 },
1268 { "read_buf_copied", KSTAT_DATA_UINT64 },
1269 { "read_buf_nocopy", KSTAT_DATA_UINT64 },
1270 { "write_buf_copied", KSTAT_DATA_UINT64 },
1271 { "write_buf_nocopy", KSTAT_DATA_UINT64 }
1272 };
1273
1274 #define XUIOSTAT_INCR(stat, val) \
1275 atomic_add_64(&xuio_stats.stat.value.ui64, (val))
1276 #define XUIOSTAT_BUMP(stat) XUIOSTAT_INCR(stat, 1)
1277
1278 #ifdef HAVE_UIO_ZEROCOPY
1279 int
1280 dmu_xuio_init(xuio_t *xuio, int nblk)
1281 {
1282 dmu_xuio_t *priv;
1283 uio_t *uio = &xuio->xu_uio;
1284
1285 uio->uio_iovcnt = nblk;
1286 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
1287
1288 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
1289 priv->cnt = nblk;
1290 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
1291 priv->iovp = (iovec_t *)uio->uio_iov;
1292 XUIO_XUZC_PRIV(xuio) = priv;
1293
1294 if (XUIO_XUZC_RW(xuio) == UIO_READ)
1295 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
1296 else
1297 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
1298
1299 return (0);
1300 }
1301
1302 void
1303 dmu_xuio_fini(xuio_t *xuio)
1304 {
1305 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1306 int nblk = priv->cnt;
1307
1308 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
1309 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
1310 kmem_free(priv, sizeof (dmu_xuio_t));
1311
1312 if (XUIO_XUZC_RW(xuio) == UIO_READ)
1313 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
1314 else
1315 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
1316 }
1317
1318 /*
1319 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
1320 * and increase priv->next by 1.
1321 */
1322 int
1323 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
1324 {
1325 struct iovec *iov;
1326 uio_t *uio = &xuio->xu_uio;
1327 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1328 int i = priv->next++;
1329
1330 ASSERT(i < priv->cnt);
1331 ASSERT(off + n <= arc_buf_lsize(abuf));
1332 iov = (iovec_t *)uio->uio_iov + i;
1333 iov->iov_base = (char *)abuf->b_data + off;
1334 iov->iov_len = n;
1335 priv->bufs[i] = abuf;
1336 return (0);
1337 }
1338
1339 int
1340 dmu_xuio_cnt(xuio_t *xuio)
1341 {
1342 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1343 return (priv->cnt);
1344 }
1345
1346 arc_buf_t *
1347 dmu_xuio_arcbuf(xuio_t *xuio, int i)
1348 {
1349 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1350
1351 ASSERT(i < priv->cnt);
1352 return (priv->bufs[i]);
1353 }
1354
1355 void
1356 dmu_xuio_clear(xuio_t *xuio, int i)
1357 {
1358 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1359
1360 ASSERT(i < priv->cnt);
1361 priv->bufs[i] = NULL;
1362 }
1363 #endif /* HAVE_UIO_ZEROCOPY */
1364
1365 static void
1366 xuio_stat_init(void)
1367 {
1368 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
1369 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
1370 KSTAT_FLAG_VIRTUAL);
1371 if (xuio_ksp != NULL) {
1372 xuio_ksp->ks_data = &xuio_stats;
1373 kstat_install(xuio_ksp);
1374 }
1375 }
1376
1377 static void
1378 xuio_stat_fini(void)
1379 {
1380 if (xuio_ksp != NULL) {
1381 kstat_delete(xuio_ksp);
1382 xuio_ksp = NULL;
1383 }
1384 }
1385
1386 void
1387 xuio_stat_wbuf_copied(void)
1388 {
1389 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1390 }
1391
1392 void
1393 xuio_stat_wbuf_nocopy(void)
1394 {
1395 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
1396 }
1397
1398 #ifdef _KERNEL
1399 int
1400 dmu_read_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size)
1401 {
1402 dmu_buf_t **dbp;
1403 int numbufs, i, err;
1404 #ifdef HAVE_UIO_ZEROCOPY
1405 xuio_t *xuio = NULL;
1406 #endif
1407
1408 /*
1409 * NB: we could do this block-at-a-time, but it's nice
1410 * to be reading in parallel.
1411 */
1412 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1413 TRUE, FTAG, &numbufs, &dbp, 0);
1414 if (err)
1415 return (err);
1416
1417 for (i = 0; i < numbufs; i++) {
1418 uint64_t tocpy;
1419 int64_t bufoff;
1420 dmu_buf_t *db = dbp[i];
1421
1422 ASSERT(size > 0);
1423
1424 bufoff = uio->uio_loffset - db->db_offset;
1425 tocpy = MIN(db->db_size - bufoff, size);
1426
1427 #ifdef HAVE_UIO_ZEROCOPY
1428 if (xuio) {
1429 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
1430 arc_buf_t *dbuf_abuf = dbi->db_buf;
1431 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
1432 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
1433 if (!err) {
1434 uio->uio_resid -= tocpy;
1435 uio->uio_loffset += tocpy;
1436 }
1437
1438 if (abuf == dbuf_abuf)
1439 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
1440 else
1441 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
1442 } else
1443 #endif
1444 err = uiomove((char *)db->db_data + bufoff, tocpy,
1445 UIO_READ, uio);
1446 if (err)
1447 break;
1448
1449 size -= tocpy;
1450 }
1451 dmu_buf_rele_array(dbp, numbufs, FTAG);
1452
1453 return (err);
1454 }
1455
1456 /*
1457 * Read 'size' bytes into the uio buffer.
1458 * From object zdb->db_object.
1459 * Starting at offset uio->uio_loffset.
1460 *
1461 * If the caller already has a dbuf in the target object
1462 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1463 * because we don't have to find the dnode_t for the object.
1464 */
1465 int
1466 dmu_read_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size)
1467 {
1468 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1469 dnode_t *dn;
1470 int err;
1471
1472 if (size == 0)
1473 return (0);
1474
1475 DB_DNODE_ENTER(db);
1476 dn = DB_DNODE(db);
1477 err = dmu_read_uio_dnode(dn, uio, size);
1478 DB_DNODE_EXIT(db);
1479
1480 return (err);
1481 }
1482
1483 /*
1484 * Read 'size' bytes into the uio buffer.
1485 * From the specified object
1486 * Starting at offset uio->uio_loffset.
1487 */
1488 int
1489 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
1490 {
1491 dnode_t *dn;
1492 int err;
1493
1494 if (size == 0)
1495 return (0);
1496
1497 err = dnode_hold(os, object, FTAG, &dn);
1498 if (err)
1499 return (err);
1500
1501 err = dmu_read_uio_dnode(dn, uio, size);
1502
1503 dnode_rele(dn, FTAG);
1504
1505 return (err);
1506 }
1507
1508 int
1509 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1510 {
1511 dmu_buf_t **dbp;
1512 int numbufs;
1513 int err = 0;
1514 int i;
1515
1516 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1517 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1518 if (err)
1519 return (err);
1520
1521 for (i = 0; i < numbufs; i++) {
1522 uint64_t tocpy;
1523 int64_t bufoff;
1524 dmu_buf_t *db = dbp[i];
1525
1526 ASSERT(size > 0);
1527
1528 bufoff = uio->uio_loffset - db->db_offset;
1529 tocpy = MIN(db->db_size - bufoff, size);
1530
1531 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1532
1533 if (tocpy == db->db_size)
1534 dmu_buf_will_fill(db, tx);
1535 else
1536 dmu_buf_will_dirty(db, tx);
1537
1538 /*
1539 * XXX uiomove could block forever (eg.nfs-backed
1540 * pages). There needs to be a uiolockdown() function
1541 * to lock the pages in memory, so that uiomove won't
1542 * block.
1543 */
1544 err = uiomove((char *)db->db_data + bufoff, tocpy,
1545 UIO_WRITE, uio);
1546
1547 if (tocpy == db->db_size)
1548 dmu_buf_fill_done(db, tx);
1549
1550 if (err)
1551 break;
1552
1553 size -= tocpy;
1554 }
1555
1556 dmu_buf_rele_array(dbp, numbufs, FTAG);
1557 return (err);
1558 }
1559
1560 /*
1561 * Write 'size' bytes from the uio buffer.
1562 * To object zdb->db_object.
1563 * Starting at offset uio->uio_loffset.
1564 *
1565 * If the caller already has a dbuf in the target object
1566 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1567 * because we don't have to find the dnode_t for the object.
1568 */
1569 int
1570 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1571 dmu_tx_t *tx)
1572 {
1573 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1574 dnode_t *dn;
1575 int err;
1576
1577 if (size == 0)
1578 return (0);
1579
1580 DB_DNODE_ENTER(db);
1581 dn = DB_DNODE(db);
1582 err = dmu_write_uio_dnode(dn, uio, size, tx);
1583 DB_DNODE_EXIT(db);
1584
1585 return (err);
1586 }
1587
1588 /*
1589 * Write 'size' bytes from the uio buffer.
1590 * To the specified object.
1591 * Starting at offset uio->uio_loffset.
1592 */
1593 int
1594 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1595 dmu_tx_t *tx)
1596 {
1597 dnode_t *dn;
1598 int err;
1599
1600 if (size == 0)
1601 return (0);
1602
1603 err = dnode_hold(os, object, FTAG, &dn);
1604 if (err)
1605 return (err);
1606
1607 err = dmu_write_uio_dnode(dn, uio, size, tx);
1608
1609 dnode_rele(dn, FTAG);
1610
1611 return (err);
1612 }
1613 #endif /* _KERNEL */
1614
1615 /*
1616 * Allocate a loaned anonymous arc buffer.
1617 */
1618 arc_buf_t *
1619 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1620 {
1621 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1622
1623 return (arc_loan_buf(db->db_objset->os_spa, B_FALSE, size));
1624 }
1625
1626 /*
1627 * Free a loaned arc buffer.
1628 */
1629 void
1630 dmu_return_arcbuf(arc_buf_t *buf)
1631 {
1632 arc_return_buf(buf, FTAG);
1633 arc_buf_destroy(buf, FTAG);
1634 }
1635
1636 void
1637 dmu_copy_from_buf(objset_t *os, uint64_t object, uint64_t offset,
1638 dmu_buf_t *handle, dmu_tx_t *tx)
1639 {
1640 dmu_buf_t *dst_handle;
1641 dmu_buf_impl_t *dstdb;
1642 dmu_buf_impl_t *srcdb = (dmu_buf_impl_t *)handle;
1643 dmu_object_type_t type;
1644 arc_buf_t *abuf;
1645 uint64_t datalen;
1646 boolean_t byteorder;
1647 uint8_t salt[ZIO_DATA_SALT_LEN];
1648 uint8_t iv[ZIO_DATA_IV_LEN];
1649 uint8_t mac[ZIO_DATA_MAC_LEN];
1650
1651 ASSERT3P(srcdb->db_buf, !=, NULL);
1652
1653 /* hold the db that we want to write to */
1654 VERIFY0(dmu_buf_hold(os, object, offset, FTAG, &dst_handle,
1655 DMU_READ_NO_DECRYPT));
1656 dstdb = (dmu_buf_impl_t *)dst_handle;
1657 datalen = arc_buf_size(srcdb->db_buf);
1658
1659 DB_DNODE_ENTER(dstdb);
1660 type = DB_DNODE(dstdb)->dn_type;
1661 DB_DNODE_EXIT(dstdb);
1662
1663 /* allocated an arc buffer that matches the type of srcdb->db_buf */
1664 if (arc_is_encrypted(srcdb->db_buf)) {
1665 arc_get_raw_params(srcdb->db_buf, &byteorder, salt, iv, mac);
1666 abuf = arc_loan_raw_buf(os->os_spa, dmu_objset_id(os),
1667 byteorder, salt, iv, mac, type,
1668 datalen, arc_buf_lsize(srcdb->db_buf),
1669 arc_get_compression(srcdb->db_buf));
1670 } else {
1671 /* we won't get a compressed db back from dmu_buf_hold() */
1672 ASSERT3U(arc_get_compression(srcdb->db_buf),
1673 ==, ZIO_COMPRESS_OFF);
1674 abuf = arc_loan_buf(os->os_spa,
1675 DMU_OT_IS_METADATA(type), datalen);
1676 }
1677
1678 ASSERT3U(datalen, ==, arc_buf_size(abuf));
1679
1680 /* copy the data to the new buffer and assign it to the dstdb */
1681 bcopy(srcdb->db_buf->b_data, abuf->b_data, datalen);
1682 dbuf_assign_arcbuf(dstdb, abuf, tx);
1683 dmu_buf_rele(dst_handle, FTAG);
1684 }
1685
1686 /*
1687 * When possible directly assign passed loaned arc buffer to a dbuf.
1688 * If this is not possible copy the contents of passed arc buf via
1689 * dmu_write().
1690 */
1691 int
1692 dmu_assign_arcbuf_by_dnode(dnode_t *dn, uint64_t offset, arc_buf_t *buf,
1693 dmu_tx_t *tx)
1694 {
1695 dmu_buf_impl_t *db;
1696 objset_t *os = dn->dn_objset;
1697 uint64_t object = dn->dn_object;
1698 uint32_t blksz = (uint32_t)arc_buf_lsize(buf);
1699 uint64_t blkid;
1700
1701 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1702 blkid = dbuf_whichblock(dn, 0, offset);
1703 db = dbuf_hold(dn, blkid, FTAG);
1704 if (db == NULL)
1705 return (SET_ERROR(EIO));
1706 rw_exit(&dn->dn_struct_rwlock);
1707
1708 /*
1709 * We can only assign if the offset is aligned, the arc buf is the
1710 * same size as the dbuf, and the dbuf is not metadata.
1711 */
1712 if (offset == db->db.db_offset && blksz == db->db.db_size) {
1713 dbuf_assign_arcbuf(db, buf, tx);
1714 dbuf_rele(db, FTAG);
1715 } else {
1716 /* compressed bufs must always be assignable to their dbuf */
1717 ASSERT3U(arc_get_compression(buf), ==, ZIO_COMPRESS_OFF);
1718 ASSERT(!(buf->b_flags & ARC_BUF_FLAG_COMPRESSED));
1719
1720 dbuf_rele(db, FTAG);
1721 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1722 dmu_return_arcbuf(buf);
1723 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1724 }
1725
1726 return (0);
1727 }
1728
1729 int
1730 dmu_assign_arcbuf_by_dbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1731 dmu_tx_t *tx)
1732 {
1733 int err;
1734 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1735
1736 DB_DNODE_ENTER(dbuf);
1737 err = dmu_assign_arcbuf_by_dnode(DB_DNODE(dbuf), offset, buf, tx);
1738 DB_DNODE_EXIT(dbuf);
1739
1740 return (err);
1741 }
1742
1743 typedef struct {
1744 dbuf_dirty_record_t *dsa_dr;
1745 dmu_sync_cb_t *dsa_done;
1746 zgd_t *dsa_zgd;
1747 dmu_tx_t *dsa_tx;
1748 } dmu_sync_arg_t;
1749
1750 /* ARGSUSED */
1751 static void
1752 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1753 {
1754 dmu_sync_arg_t *dsa = varg;
1755 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1756 blkptr_t *bp = zio->io_bp;
1757
1758 if (zio->io_error == 0) {
1759 if (BP_IS_HOLE(bp)) {
1760 /*
1761 * A block of zeros may compress to a hole, but the
1762 * block size still needs to be known for replay.
1763 */
1764 BP_SET_LSIZE(bp, db->db_size);
1765 } else if (!BP_IS_EMBEDDED(bp)) {
1766 ASSERT(BP_GET_LEVEL(bp) == 0);
1767 BP_SET_FILL(bp, 1);
1768 }
1769 }
1770 }
1771
1772 static void
1773 dmu_sync_late_arrival_ready(zio_t *zio)
1774 {
1775 dmu_sync_ready(zio, NULL, zio->io_private);
1776 }
1777
1778 /* ARGSUSED */
1779 static void
1780 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1781 {
1782 dmu_sync_arg_t *dsa = varg;
1783 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1784 dmu_buf_impl_t *db = dr->dr_dbuf;
1785 zgd_t *zgd = dsa->dsa_zgd;
1786
1787 /*
1788 * Record the vdev(s) backing this blkptr so they can be flushed after
1789 * the writes for the lwb have completed.
1790 */
1791 if (zio->io_error == 0) {
1792 zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp);
1793 }
1794
1795 mutex_enter(&db->db_mtx);
1796 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1797 if (zio->io_error == 0) {
1798 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1799 if (dr->dt.dl.dr_nopwrite) {
1800 blkptr_t *bp = zio->io_bp;
1801 blkptr_t *bp_orig = &zio->io_bp_orig;
1802 uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1803
1804 ASSERT(BP_EQUAL(bp, bp_orig));
1805 VERIFY(BP_EQUAL(bp, db->db_blkptr));
1806 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1807 VERIFY(zio_checksum_table[chksum].ci_flags &
1808 ZCHECKSUM_FLAG_NOPWRITE);
1809 }
1810 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1811 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1812 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1813
1814 /*
1815 * Old style holes are filled with all zeros, whereas
1816 * new-style holes maintain their lsize, type, level,
1817 * and birth time (see zio_write_compress). While we
1818 * need to reset the BP_SET_LSIZE() call that happened
1819 * in dmu_sync_ready for old style holes, we do *not*
1820 * want to wipe out the information contained in new
1821 * style holes. Thus, only zero out the block pointer if
1822 * it's an old style hole.
1823 */
1824 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) &&
1825 dr->dt.dl.dr_overridden_by.blk_birth == 0)
1826 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1827 } else {
1828 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1829 }
1830 cv_broadcast(&db->db_changed);
1831 mutex_exit(&db->db_mtx);
1832
1833 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1834
1835 kmem_free(dsa, sizeof (*dsa));
1836 }
1837
1838 static void
1839 dmu_sync_late_arrival_done(zio_t *zio)
1840 {
1841 blkptr_t *bp = zio->io_bp;
1842 dmu_sync_arg_t *dsa = zio->io_private;
1843 zgd_t *zgd = dsa->dsa_zgd;
1844
1845 if (zio->io_error == 0) {
1846 /*
1847 * Record the vdev(s) backing this blkptr so they can be
1848 * flushed after the writes for the lwb have completed.
1849 */
1850 zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp);
1851
1852 if (!BP_IS_HOLE(bp)) {
1853 ASSERTV(blkptr_t *bp_orig = &zio->io_bp_orig);
1854 ASSERT(!(zio->io_flags & ZIO_FLAG_NOPWRITE));
1855 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1856 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1857 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1858 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1859 }
1860 }
1861
1862 dmu_tx_commit(dsa->dsa_tx);
1863
1864 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1865
1866 abd_put(zio->io_abd);
1867 kmem_free(dsa, sizeof (*dsa));
1868 }
1869
1870 static int
1871 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1872 zio_prop_t *zp, zbookmark_phys_t *zb)
1873 {
1874 dmu_sync_arg_t *dsa;
1875 dmu_tx_t *tx;
1876
1877 tx = dmu_tx_create(os);
1878 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1879 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1880 dmu_tx_abort(tx);
1881 /* Make zl_get_data do txg_waited_synced() */
1882 return (SET_ERROR(EIO));
1883 }
1884
1885 /*
1886 * In order to prevent the zgd's lwb from being free'd prior to
1887 * dmu_sync_late_arrival_done() being called, we have to ensure
1888 * the lwb's "max txg" takes this tx's txg into account.
1889 */
1890 zil_lwb_add_txg(zgd->zgd_lwb, dmu_tx_get_txg(tx));
1891
1892 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1893 dsa->dsa_dr = NULL;
1894 dsa->dsa_done = done;
1895 dsa->dsa_zgd = zgd;
1896 dsa->dsa_tx = tx;
1897
1898 /*
1899 * Since we are currently syncing this txg, it's nontrivial to
1900 * determine what BP to nopwrite against, so we disable nopwrite.
1901 *
1902 * When syncing, the db_blkptr is initially the BP of the previous
1903 * txg. We can not nopwrite against it because it will be changed
1904 * (this is similar to the non-late-arrival case where the dbuf is
1905 * dirty in a future txg).
1906 *
1907 * Then dbuf_write_ready() sets bp_blkptr to the location we will write.
1908 * We can not nopwrite against it because although the BP will not
1909 * (typically) be changed, the data has not yet been persisted to this
1910 * location.
1911 *
1912 * Finally, when dbuf_write_done() is called, it is theoretically
1913 * possible to always nopwrite, because the data that was written in
1914 * this txg is the same data that we are trying to write. However we
1915 * would need to check that this dbuf is not dirty in any future
1916 * txg's (as we do in the normal dmu_sync() path). For simplicity, we
1917 * don't nopwrite in this case.
1918 */
1919 zp->zp_nopwrite = B_FALSE;
1920
1921 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1922 abd_get_from_buf(zgd->zgd_db->db_data, zgd->zgd_db->db_size),
1923 zgd->zgd_db->db_size, zgd->zgd_db->db_size, zp,
1924 dmu_sync_late_arrival_ready, NULL, NULL, dmu_sync_late_arrival_done,
1925 dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1926
1927 return (0);
1928 }
1929
1930 /*
1931 * Intent log support: sync the block associated with db to disk.
1932 * N.B. and XXX: the caller is responsible for making sure that the
1933 * data isn't changing while dmu_sync() is writing it.
1934 *
1935 * Return values:
1936 *
1937 * EEXIST: this txg has already been synced, so there's nothing to do.
1938 * The caller should not log the write.
1939 *
1940 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1941 * The caller should not log the write.
1942 *
1943 * EALREADY: this block is already in the process of being synced.
1944 * The caller should track its progress (somehow).
1945 *
1946 * EIO: could not do the I/O.
1947 * The caller should do a txg_wait_synced().
1948 *
1949 * 0: the I/O has been initiated.
1950 * The caller should log this blkptr in the done callback.
1951 * It is possible that the I/O will fail, in which case
1952 * the error will be reported to the done callback and
1953 * propagated to pio from zio_done().
1954 */
1955 int
1956 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1957 {
1958 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1959 objset_t *os = db->db_objset;
1960 dsl_dataset_t *ds = os->os_dsl_dataset;
1961 dbuf_dirty_record_t *dr;
1962 dmu_sync_arg_t *dsa;
1963 zbookmark_phys_t zb;
1964 zio_prop_t zp;
1965 dnode_t *dn;
1966
1967 ASSERT(pio != NULL);
1968 ASSERT(txg != 0);
1969
1970 SET_BOOKMARK(&zb, ds->ds_object,
1971 db->db.db_object, db->db_level, db->db_blkid);
1972
1973 DB_DNODE_ENTER(db);
1974 dn = DB_DNODE(db);
1975 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1976 DB_DNODE_EXIT(db);
1977
1978 /*
1979 * If we're frozen (running ziltest), we always need to generate a bp.
1980 */
1981 if (txg > spa_freeze_txg(os->os_spa))
1982 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1983
1984 /*
1985 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1986 * and us. If we determine that this txg is not yet syncing,
1987 * but it begins to sync a moment later, that's OK because the
1988 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1989 */
1990 mutex_enter(&db->db_mtx);
1991
1992 if (txg <= spa_last_synced_txg(os->os_spa)) {
1993 /*
1994 * This txg has already synced. There's nothing to do.
1995 */
1996 mutex_exit(&db->db_mtx);
1997 return (SET_ERROR(EEXIST));
1998 }
1999
2000 if (txg <= spa_syncing_txg(os->os_spa)) {
2001 /*
2002 * This txg is currently syncing, so we can't mess with
2003 * the dirty record anymore; just write a new log block.
2004 */
2005 mutex_exit(&db->db_mtx);
2006 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
2007 }
2008
2009 dr = db->db_last_dirty;
2010 while (dr && dr->dr_txg != txg)
2011 dr = dr->dr_next;
2012
2013 if (dr == NULL) {
2014 /*
2015 * There's no dr for this dbuf, so it must have been freed.
2016 * There's no need to log writes to freed blocks, so we're done.
2017 */
2018 mutex_exit(&db->db_mtx);
2019 return (SET_ERROR(ENOENT));
2020 }
2021
2022 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
2023
2024 if (db->db_blkptr != NULL) {
2025 /*
2026 * We need to fill in zgd_bp with the current blkptr so that
2027 * the nopwrite code can check if we're writing the same
2028 * data that's already on disk. We can only nopwrite if we
2029 * are sure that after making the copy, db_blkptr will not
2030 * change until our i/o completes. We ensure this by
2031 * holding the db_mtx, and only allowing nopwrite if the
2032 * block is not already dirty (see below). This is verified
2033 * by dmu_sync_done(), which VERIFYs that the db_blkptr has
2034 * not changed.
2035 */
2036 *zgd->zgd_bp = *db->db_blkptr;
2037 }
2038
2039 /*
2040 * Assume the on-disk data is X, the current syncing data (in
2041 * txg - 1) is Y, and the current in-memory data is Z (currently
2042 * in dmu_sync).
2043 *
2044 * We usually want to perform a nopwrite if X and Z are the
2045 * same. However, if Y is different (i.e. the BP is going to
2046 * change before this write takes effect), then a nopwrite will
2047 * be incorrect - we would override with X, which could have
2048 * been freed when Y was written.
2049 *
2050 * (Note that this is not a concern when we are nop-writing from
2051 * syncing context, because X and Y must be identical, because
2052 * all previous txgs have been synced.)
2053 *
2054 * Therefore, we disable nopwrite if the current BP could change
2055 * before this TXG. There are two ways it could change: by
2056 * being dirty (dr_next is non-NULL), or by being freed
2057 * (dnode_block_freed()). This behavior is verified by
2058 * zio_done(), which VERIFYs that the override BP is identical
2059 * to the on-disk BP.
2060 */
2061 DB_DNODE_ENTER(db);
2062 dn = DB_DNODE(db);
2063 if (dr->dr_next != NULL || dnode_block_freed(dn, db->db_blkid))
2064 zp.zp_nopwrite = B_FALSE;
2065 DB_DNODE_EXIT(db);
2066
2067 ASSERT(dr->dr_txg == txg);
2068 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
2069 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
2070 /*
2071 * We have already issued a sync write for this buffer,
2072 * or this buffer has already been synced. It could not
2073 * have been dirtied since, or we would have cleared the state.
2074 */
2075 mutex_exit(&db->db_mtx);
2076 return (SET_ERROR(EALREADY));
2077 }
2078
2079 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
2080 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
2081 mutex_exit(&db->db_mtx);
2082
2083 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
2084 dsa->dsa_dr = dr;
2085 dsa->dsa_done = done;
2086 dsa->dsa_zgd = zgd;
2087 dsa->dsa_tx = NULL;
2088
2089 zio_nowait(arc_write(pio, os->os_spa, txg,
2090 zgd->zgd_bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
2091 &zp, dmu_sync_ready, NULL, NULL, dmu_sync_done, dsa,
2092 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb));
2093
2094 return (0);
2095 }
2096
2097 int
2098 dmu_object_set_nlevels(objset_t *os, uint64_t object, int nlevels, dmu_tx_t *tx)
2099 {
2100 dnode_t *dn;
2101 int err;
2102
2103 err = dnode_hold(os, object, FTAG, &dn);
2104 if (err)
2105 return (err);
2106 err = dnode_set_nlevels(dn, nlevels, tx);
2107 dnode_rele(dn, FTAG);
2108 return (err);
2109 }
2110
2111 int
2112 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
2113 dmu_tx_t *tx)
2114 {
2115 dnode_t *dn;
2116 int err;
2117
2118 err = dnode_hold(os, object, FTAG, &dn);
2119 if (err)
2120 return (err);
2121 err = dnode_set_blksz(dn, size, ibs, tx);
2122 dnode_rele(dn, FTAG);
2123 return (err);
2124 }
2125
2126 int
2127 dmu_object_set_maxblkid(objset_t *os, uint64_t object, uint64_t maxblkid,
2128 dmu_tx_t *tx)
2129 {
2130 dnode_t *dn;
2131 int err;
2132
2133 err = dnode_hold(os, object, FTAG, &dn);
2134 if (err)
2135 return (err);
2136 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2137 dnode_new_blkid(dn, maxblkid, tx, B_FALSE);
2138 rw_exit(&dn->dn_struct_rwlock);
2139 dnode_rele(dn, FTAG);
2140 return (0);
2141 }
2142
2143 void
2144 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
2145 dmu_tx_t *tx)
2146 {
2147 dnode_t *dn;
2148
2149 /*
2150 * Send streams include each object's checksum function. This
2151 * check ensures that the receiving system can understand the
2152 * checksum function transmitted.
2153 */
2154 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS);
2155
2156 VERIFY0(dnode_hold(os, object, FTAG, &dn));
2157 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS);
2158 dn->dn_checksum = checksum;
2159 dnode_setdirty(dn, tx);
2160 dnode_rele(dn, FTAG);
2161 }
2162
2163 void
2164 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
2165 dmu_tx_t *tx)
2166 {
2167 dnode_t *dn;
2168
2169 /*
2170 * Send streams include each object's compression function. This
2171 * check ensures that the receiving system can understand the
2172 * compression function transmitted.
2173 */
2174 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS);
2175
2176 VERIFY0(dnode_hold(os, object, FTAG, &dn));
2177 dn->dn_compress = compress;
2178 dnode_setdirty(dn, tx);
2179 dnode_rele(dn, FTAG);
2180 }
2181
2182 /*
2183 * When the "redundant_metadata" property is set to "most", only indirect
2184 * blocks of this level and higher will have an additional ditto block.
2185 */
2186 int zfs_redundant_metadata_most_ditto_level = 2;
2187
2188 void
2189 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
2190 {
2191 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
2192 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
2193 (wp & WP_SPILL));
2194 enum zio_checksum checksum = os->os_checksum;
2195 enum zio_compress compress = os->os_compress;
2196 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
2197 boolean_t dedup = B_FALSE;
2198 boolean_t nopwrite = B_FALSE;
2199 boolean_t dedup_verify = os->os_dedup_verify;
2200 boolean_t encrypt = B_FALSE;
2201 int copies = os->os_copies;
2202
2203 /*
2204 * We maintain different write policies for each of the following
2205 * types of data:
2206 * 1. metadata
2207 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
2208 * 3. all other level 0 blocks
2209 */
2210 if (ismd) {
2211 /*
2212 * XXX -- we should design a compression algorithm
2213 * that specializes in arrays of bps.
2214 */
2215 compress = zio_compress_select(os->os_spa,
2216 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON);
2217
2218 /*
2219 * Metadata always gets checksummed. If the data
2220 * checksum is multi-bit correctable, and it's not a
2221 * ZBT-style checksum, then it's suitable for metadata
2222 * as well. Otherwise, the metadata checksum defaults
2223 * to fletcher4.
2224 */
2225 if (!(zio_checksum_table[checksum].ci_flags &
2226 ZCHECKSUM_FLAG_METADATA) ||
2227 (zio_checksum_table[checksum].ci_flags &
2228 ZCHECKSUM_FLAG_EMBEDDED))
2229 checksum = ZIO_CHECKSUM_FLETCHER_4;
2230
2231 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL ||
2232 (os->os_redundant_metadata ==
2233 ZFS_REDUNDANT_METADATA_MOST &&
2234 (level >= zfs_redundant_metadata_most_ditto_level ||
2235 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))))
2236 copies++;
2237 } else if (wp & WP_NOFILL) {
2238 ASSERT(level == 0);
2239
2240 /*
2241 * If we're writing preallocated blocks, we aren't actually
2242 * writing them so don't set any policy properties. These
2243 * blocks are currently only used by an external subsystem
2244 * outside of zfs (i.e. dump) and not written by the zio
2245 * pipeline.
2246 */
2247 compress = ZIO_COMPRESS_OFF;
2248 checksum = ZIO_CHECKSUM_OFF;
2249 } else {
2250 compress = zio_compress_select(os->os_spa, dn->dn_compress,
2251 compress);
2252
2253 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
2254 zio_checksum_select(dn->dn_checksum, checksum) :
2255 dedup_checksum;
2256
2257 /*
2258 * Determine dedup setting. If we are in dmu_sync(),
2259 * we won't actually dedup now because that's all
2260 * done in syncing context; but we do want to use the
2261 * dedup checkum. If the checksum is not strong
2262 * enough to ensure unique signatures, force
2263 * dedup_verify.
2264 */
2265 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
2266 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
2267 if (!(zio_checksum_table[checksum].ci_flags &
2268 ZCHECKSUM_FLAG_DEDUP))
2269 dedup_verify = B_TRUE;
2270 }
2271
2272 /*
2273 * Enable nopwrite if we have secure enough checksum
2274 * algorithm (see comment in zio_nop_write) and
2275 * compression is enabled. We don't enable nopwrite if
2276 * dedup is enabled as the two features are mutually
2277 * exclusive.
2278 */
2279 nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags &
2280 ZCHECKSUM_FLAG_NOPWRITE) &&
2281 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
2282 }
2283
2284 /*
2285 * All objects in an encrypted objset are protected from modification
2286 * via a MAC. Encrypted objects store their IV and salt in the last DVA
2287 * in the bp, so we cannot use all copies. Encrypted objects are also
2288 * not subject to nopwrite since writing the same data will still
2289 * result in a new ciphertext. Only encrypted blocks can be dedup'd
2290 * to avoid ambiguity in the dedup code since the DDT does not store
2291 * object types.
2292 */
2293 if (os->os_encrypted && (wp & WP_NOFILL) == 0) {
2294 encrypt = B_TRUE;
2295
2296 if (DMU_OT_IS_ENCRYPTED(type)) {
2297 copies = MIN(copies, SPA_DVAS_PER_BP - 1);
2298 nopwrite = B_FALSE;
2299 } else {
2300 dedup = B_FALSE;
2301 }
2302
2303 if (level <= 0 &&
2304 (type == DMU_OT_DNODE || type == DMU_OT_OBJSET)) {
2305 compress = ZIO_COMPRESS_EMPTY;
2306 }
2307 }
2308
2309 zp->zp_compress = compress;
2310 zp->zp_checksum = checksum;
2311 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
2312 zp->zp_level = level;
2313 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
2314 zp->zp_dedup = dedup;
2315 zp->zp_dedup_verify = dedup && dedup_verify;
2316 zp->zp_nopwrite = nopwrite;
2317 zp->zp_encrypt = encrypt;
2318 zp->zp_byteorder = ZFS_HOST_BYTEORDER;
2319 bzero(zp->zp_salt, ZIO_DATA_SALT_LEN);
2320 bzero(zp->zp_iv, ZIO_DATA_IV_LEN);
2321 bzero(zp->zp_mac, ZIO_DATA_MAC_LEN);
2322 zp->zp_zpl_smallblk = DMU_OT_IS_FILE(zp->zp_type) ?
2323 os->os_zpl_special_smallblock : 0;
2324
2325 ASSERT3U(zp->zp_compress, !=, ZIO_COMPRESS_INHERIT);
2326 }
2327
2328 /*
2329 * This function is only called from zfs_holey_common() for zpl_llseek()
2330 * in order to determine the location of holes. In order to accurately
2331 * report holes all dirty data must be synced to disk. This causes extremely
2332 * poor performance when seeking for holes in a dirty file. As a compromise,
2333 * only provide hole data when the dnode is clean. When a dnode is dirty
2334 * report the dnode as having no holes which is always a safe thing to do.
2335 */
2336 int
2337 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
2338 {
2339 dnode_t *dn;
2340 int i, err;
2341 boolean_t clean = B_TRUE;
2342
2343 err = dnode_hold(os, object, FTAG, &dn);
2344 if (err)
2345 return (err);
2346
2347 /*
2348 * Check if dnode is dirty
2349 */
2350 for (i = 0; i < TXG_SIZE; i++) {
2351 if (multilist_link_active(&dn->dn_dirty_link[i])) {
2352 clean = B_FALSE;
2353 break;
2354 }
2355 }
2356
2357 /*
2358 * If compatibility option is on, sync any current changes before
2359 * we go trundling through the block pointers.
2360 */
2361 if (!clean && zfs_dmu_offset_next_sync) {
2362 clean = B_TRUE;
2363 dnode_rele(dn, FTAG);
2364 txg_wait_synced(dmu_objset_pool(os), 0);
2365 err = dnode_hold(os, object, FTAG, &dn);
2366 if (err)
2367 return (err);
2368 }
2369
2370 if (clean)
2371 err = dnode_next_offset(dn,
2372 (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
2373 else
2374 err = SET_ERROR(EBUSY);
2375
2376 dnode_rele(dn, FTAG);
2377
2378 return (err);
2379 }
2380
2381 void
2382 __dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2383 {
2384 dnode_phys_t *dnp = dn->dn_phys;
2385
2386 doi->doi_data_block_size = dn->dn_datablksz;
2387 doi->doi_metadata_block_size = dn->dn_indblkshift ?
2388 1ULL << dn->dn_indblkshift : 0;
2389 doi->doi_type = dn->dn_type;
2390 doi->doi_bonus_type = dn->dn_bonustype;
2391 doi->doi_bonus_size = dn->dn_bonuslen;
2392 doi->doi_dnodesize = dn->dn_num_slots << DNODE_SHIFT;
2393 doi->doi_indirection = dn->dn_nlevels;
2394 doi->doi_checksum = dn->dn_checksum;
2395 doi->doi_compress = dn->dn_compress;
2396 doi->doi_nblkptr = dn->dn_nblkptr;
2397 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
2398 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
2399 doi->doi_fill_count = 0;
2400 for (int i = 0; i < dnp->dn_nblkptr; i++)
2401 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
2402 }
2403
2404 void
2405 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2406 {
2407 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2408 mutex_enter(&dn->dn_mtx);
2409
2410 __dmu_object_info_from_dnode(dn, doi);
2411
2412 mutex_exit(&dn->dn_mtx);
2413 rw_exit(&dn->dn_struct_rwlock);
2414 }
2415
2416 /*
2417 * Get information on a DMU object.
2418 * If doi is NULL, just indicates whether the object exists.
2419 */
2420 int
2421 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
2422 {
2423 dnode_t *dn;
2424 int err = dnode_hold(os, object, FTAG, &dn);
2425
2426 if (err)
2427 return (err);
2428
2429 if (doi != NULL)
2430 dmu_object_info_from_dnode(dn, doi);
2431
2432 dnode_rele(dn, FTAG);
2433 return (0);
2434 }
2435
2436 /*
2437 * As above, but faster; can be used when you have a held dbuf in hand.
2438 */
2439 void
2440 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
2441 {
2442 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2443
2444 DB_DNODE_ENTER(db);
2445 dmu_object_info_from_dnode(DB_DNODE(db), doi);
2446 DB_DNODE_EXIT(db);
2447 }
2448
2449 /*
2450 * Faster still when you only care about the size.
2451 * This is specifically optimized for zfs_getattr().
2452 */
2453 void
2454 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
2455 u_longlong_t *nblk512)
2456 {
2457 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2458 dnode_t *dn;
2459
2460 DB_DNODE_ENTER(db);
2461 dn = DB_DNODE(db);
2462
2463 *blksize = dn->dn_datablksz;
2464 /* add in number of slots used for the dnode itself */
2465 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
2466 SPA_MINBLOCKSHIFT) + dn->dn_num_slots;
2467 DB_DNODE_EXIT(db);
2468 }
2469
2470 void
2471 dmu_object_dnsize_from_db(dmu_buf_t *db_fake, int *dnsize)
2472 {
2473 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2474 dnode_t *dn;
2475
2476 DB_DNODE_ENTER(db);
2477 dn = DB_DNODE(db);
2478 *dnsize = dn->dn_num_slots << DNODE_SHIFT;
2479 DB_DNODE_EXIT(db);
2480 }
2481
2482 void
2483 byteswap_uint64_array(void *vbuf, size_t size)
2484 {
2485 uint64_t *buf = vbuf;
2486 size_t count = size >> 3;
2487 int i;
2488
2489 ASSERT((size & 7) == 0);
2490
2491 for (i = 0; i < count; i++)
2492 buf[i] = BSWAP_64(buf[i]);
2493 }
2494
2495 void
2496 byteswap_uint32_array(void *vbuf, size_t size)
2497 {
2498 uint32_t *buf = vbuf;
2499 size_t count = size >> 2;
2500 int i;
2501
2502 ASSERT((size & 3) == 0);
2503
2504 for (i = 0; i < count; i++)
2505 buf[i] = BSWAP_32(buf[i]);
2506 }
2507
2508 void
2509 byteswap_uint16_array(void *vbuf, size_t size)
2510 {
2511 uint16_t *buf = vbuf;
2512 size_t count = size >> 1;
2513 int i;
2514
2515 ASSERT((size & 1) == 0);
2516
2517 for (i = 0; i < count; i++)
2518 buf[i] = BSWAP_16(buf[i]);
2519 }
2520
2521 /* ARGSUSED */
2522 void
2523 byteswap_uint8_array(void *vbuf, size_t size)
2524 {
2525 }
2526
2527 void
2528 dmu_init(void)
2529 {
2530 abd_init();
2531 zfs_dbgmsg_init();
2532 sa_cache_init();
2533 xuio_stat_init();
2534 dmu_objset_init();
2535 dnode_init();
2536 zfetch_init();
2537 dmu_tx_init();
2538 l2arc_init();
2539 arc_init();
2540 dbuf_init();
2541 }
2542
2543 void
2544 dmu_fini(void)
2545 {
2546 arc_fini(); /* arc depends on l2arc, so arc must go first */
2547 l2arc_fini();
2548 dmu_tx_fini();
2549 zfetch_fini();
2550 dbuf_fini();
2551 dnode_fini();
2552 dmu_objset_fini();
2553 xuio_stat_fini();
2554 sa_cache_fini();
2555 zfs_dbgmsg_fini();
2556 abd_fini();
2557 }
2558
2559 #if defined(_KERNEL)
2560 EXPORT_SYMBOL(dmu_bonus_hold);
2561 EXPORT_SYMBOL(dmu_bonus_hold_by_dnode);
2562 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus);
2563 EXPORT_SYMBOL(dmu_buf_rele_array);
2564 EXPORT_SYMBOL(dmu_prefetch);
2565 EXPORT_SYMBOL(dmu_free_range);
2566 EXPORT_SYMBOL(dmu_free_long_range);
2567 EXPORT_SYMBOL(dmu_free_long_object);
2568 EXPORT_SYMBOL(dmu_read);
2569 EXPORT_SYMBOL(dmu_read_by_dnode);
2570 EXPORT_SYMBOL(dmu_write);
2571 EXPORT_SYMBOL(dmu_write_by_dnode);
2572 EXPORT_SYMBOL(dmu_prealloc);
2573 EXPORT_SYMBOL(dmu_object_info);
2574 EXPORT_SYMBOL(dmu_object_info_from_dnode);
2575 EXPORT_SYMBOL(dmu_object_info_from_db);
2576 EXPORT_SYMBOL(dmu_object_size_from_db);
2577 EXPORT_SYMBOL(dmu_object_dnsize_from_db);
2578 EXPORT_SYMBOL(dmu_object_set_nlevels);
2579 EXPORT_SYMBOL(dmu_object_set_blocksize);
2580 EXPORT_SYMBOL(dmu_object_set_maxblkid);
2581 EXPORT_SYMBOL(dmu_object_set_checksum);
2582 EXPORT_SYMBOL(dmu_object_set_compress);
2583 EXPORT_SYMBOL(dmu_write_policy);
2584 EXPORT_SYMBOL(dmu_sync);
2585 EXPORT_SYMBOL(dmu_request_arcbuf);
2586 EXPORT_SYMBOL(dmu_return_arcbuf);
2587 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dnode);
2588 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dbuf);
2589 EXPORT_SYMBOL(dmu_buf_hold);
2590 EXPORT_SYMBOL(dmu_ot);
2591
2592 /* BEGIN CSTYLED */
2593 module_param(zfs_nopwrite_enabled, int, 0644);
2594 MODULE_PARM_DESC(zfs_nopwrite_enabled, "Enable NOP writes");
2595
2596 module_param(zfs_per_txg_dirty_frees_percent, ulong, 0644);
2597 MODULE_PARM_DESC(zfs_per_txg_dirty_frees_percent,
2598 "percentage of dirtied blocks from frees in one TXG");
2599
2600 module_param(zfs_dmu_offset_next_sync, int, 0644);
2601 MODULE_PARM_DESC(zfs_dmu_offset_next_sync,
2602 "Enable forcing txg sync to find holes");
2603
2604 /* END CSTYLED */
2605
2606 #endif
ZFS On Linux mirror for PVE