?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Allow disabling of unmapped I/O on FreeBSD
[zfs.git] / module / zfs / dmu.c
blob1c47430953b1a609ee4ad747011e3b009e19da54
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, 2020 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 * Copyright (c) 2019 Datto Inc.
29 * Copyright (c) 2019, Klara Inc.
30 * Copyright (c) 2019, Allan Jude
31 */
32
33 #include <sys/dmu.h>
34 #include <sys/dmu_impl.h>
35 #include <sys/dmu_tx.h>
36 #include <sys/dbuf.h>
37 #include <sys/dnode.h>
38 #include <sys/zfs_context.h>
39 #include <sys/dmu_objset.h>
40 #include <sys/dmu_traverse.h>
41 #include <sys/dsl_dataset.h>
42 #include <sys/dsl_dir.h>
43 #include <sys/dsl_pool.h>
44 #include <sys/dsl_synctask.h>
45 #include <sys/dsl_prop.h>
46 #include <sys/dmu_zfetch.h>
47 #include <sys/zfs_ioctl.h>
48 #include <sys/zap.h>
49 #include <sys/zio_checksum.h>
50 #include <sys/zio_compress.h>
51 #include <sys/sa.h>
52 #include <sys/zfeature.h>
53 #include <sys/abd.h>
54 #include <sys/trace_zfs.h>
55 #include <sys/zfs_racct.h>
56 #include <sys/zfs_rlock.h>
57 #ifdef _KERNEL
58 #include <sys/vmsystm.h>
59 #include <sys/zfs_znode.h>
60 #endif
61
62 /*
63 * Enable/disable nopwrite feature.
64 */
65 int zfs_nopwrite_enabled = 1;
66
67 /*
68 * Tunable to control percentage of dirtied L1 blocks from frees allowed into
69 * one TXG. After this threshold is crossed, additional dirty blocks from frees
70 * will wait until the next TXG.
71 * A value of zero will disable this throttle.
72 */
73 unsigned long zfs_per_txg_dirty_frees_percent = 5;
74
75 /*
76 * Enable/disable forcing txg sync when dirty in dmu_offset_next.
77 */
78 int zfs_dmu_offset_next_sync = 0;
79
80 /*
81 * Limit the amount we can prefetch with one call to this amount. This
82 * helps to limit the amount of memory that can be used by prefetching.
83 * Larger objects should be prefetched a bit at a time.
84 */
85 int dmu_prefetch_max = 8 * SPA_MAXBLOCKSIZE;
86
87 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
88 {DMU_BSWAP_UINT8, TRUE, FALSE, FALSE, "unallocated" },
89 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "object directory" },
90 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "object array" },
91 {DMU_BSWAP_UINT8, TRUE, FALSE, FALSE, "packed nvlist" },
92 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "packed nvlist size" },
93 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "bpobj" },
94 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "bpobj header" },
95 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "SPA space map header" },
96 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "SPA space map" },
97 {DMU_BSWAP_UINT64, TRUE, FALSE, TRUE, "ZIL intent log" },
98 {DMU_BSWAP_DNODE, TRUE, FALSE, TRUE, "DMU dnode" },
99 {DMU_BSWAP_OBJSET, TRUE, TRUE, FALSE, "DMU objset" },
100 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "DSL directory" },
101 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL directory child map"},
102 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL dataset snap map" },
103 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL props" },
104 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "DSL dataset" },
105 {DMU_BSWAP_ZNODE, TRUE, FALSE, FALSE, "ZFS znode" },
106 {DMU_BSWAP_OLDACL, TRUE, FALSE, TRUE, "ZFS V0 ACL" },
107 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "ZFS plain file" },
108 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS directory" },
109 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "ZFS master node" },
110 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS delete queue" },
111 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "zvol object" },
112 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "zvol prop" },
113 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "other uint8[]" },
114 {DMU_BSWAP_UINT64, FALSE, FALSE, TRUE, "other uint64[]" },
115 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "other ZAP" },
116 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "persistent error log" },
117 {DMU_BSWAP_UINT8, TRUE, FALSE, FALSE, "SPA history" },
118 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "SPA history offsets" },
119 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "Pool properties" },
120 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL permissions" },
121 {DMU_BSWAP_ACL, TRUE, FALSE, TRUE, "ZFS ACL" },
122 {DMU_BSWAP_UINT8, TRUE, FALSE, TRUE, "ZFS SYSACL" },
123 {DMU_BSWAP_UINT8, TRUE, FALSE, TRUE, "FUID table" },
124 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "FUID table size" },
125 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL dataset next clones"},
126 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "scan work queue" },
127 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS user/group/project used" },
128 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS user/group/project quota"},
129 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "snapshot refcount tags"},
130 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "DDT ZAP algorithm" },
131 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "DDT statistics" },
132 {DMU_BSWAP_UINT8, TRUE, FALSE, TRUE, "System attributes" },
133 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "SA master node" },
134 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "SA attr registration" },
135 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "SA attr layouts" },
136 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "scan translations" },
137 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "deduplicated block" },
138 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL deadlist map" },
139 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "DSL deadlist map hdr" },
140 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL dir clones" },
141 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "bpobj subobj" }
142 };
143
144 const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
145 { byteswap_uint8_array, "uint8" },
146 { byteswap_uint16_array, "uint16" },
147 { byteswap_uint32_array, "uint32" },
148 { byteswap_uint64_array, "uint64" },
149 { zap_byteswap, "zap" },
150 { dnode_buf_byteswap, "dnode" },
151 { dmu_objset_byteswap, "objset" },
152 { zfs_znode_byteswap, "znode" },
153 { zfs_oldacl_byteswap, "oldacl" },
154 { zfs_acl_byteswap, "acl" }
155 };
156
157 static int
158 dmu_buf_hold_noread_by_dnode(dnode_t *dn, uint64_t offset,
159 void *tag, dmu_buf_t **dbp)
160 {
161 uint64_t blkid;
162 dmu_buf_impl_t *db;
163
164 rw_enter(&dn->dn_struct_rwlock, RW_READER);
165 blkid = dbuf_whichblock(dn, 0, offset);
166 db = dbuf_hold(dn, blkid, tag);
167 rw_exit(&dn->dn_struct_rwlock);
168
169 if (db == NULL) {
170 *dbp = NULL;
171 return (SET_ERROR(EIO));
172 }
173
174 *dbp = &db->db;
175 return (0);
176 }
177 int
178 dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset,
179 void *tag, dmu_buf_t **dbp)
180 {
181 dnode_t *dn;
182 uint64_t blkid;
183 dmu_buf_impl_t *db;
184 int err;
185
186 err = dnode_hold(os, object, FTAG, &dn);
187 if (err)
188 return (err);
189 rw_enter(&dn->dn_struct_rwlock, RW_READER);
190 blkid = dbuf_whichblock(dn, 0, offset);
191 db = dbuf_hold(dn, blkid, tag);
192 rw_exit(&dn->dn_struct_rwlock);
193 dnode_rele(dn, FTAG);
194
195 if (db == NULL) {
196 *dbp = NULL;
197 return (SET_ERROR(EIO));
198 }
199
200 *dbp = &db->db;
201 return (err);
202 }
203
204 int
205 dmu_buf_hold_by_dnode(dnode_t *dn, uint64_t offset,
206 void *tag, dmu_buf_t **dbp, int flags)
207 {
208 int err;
209 int db_flags = DB_RF_CANFAIL;
210
211 if (flags & DMU_READ_NO_PREFETCH)
212 db_flags |= DB_RF_NOPREFETCH;
213 if (flags & DMU_READ_NO_DECRYPT)
214 db_flags |= DB_RF_NO_DECRYPT;
215
216 err = dmu_buf_hold_noread_by_dnode(dn, offset, tag, dbp);
217 if (err == 0) {
218 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
219 err = dbuf_read(db, NULL, db_flags);
220 if (err != 0) {
221 dbuf_rele(db, tag);
222 *dbp = NULL;
223 }
224 }
225
226 return (err);
227 }
228
229 int
230 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
231 void *tag, dmu_buf_t **dbp, int flags)
232 {
233 int err;
234 int db_flags = DB_RF_CANFAIL;
235
236 if (flags & DMU_READ_NO_PREFETCH)
237 db_flags |= DB_RF_NOPREFETCH;
238 if (flags & DMU_READ_NO_DECRYPT)
239 db_flags |= DB_RF_NO_DECRYPT;
240
241 err = dmu_buf_hold_noread(os, object, offset, tag, dbp);
242 if (err == 0) {
243 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
244 err = dbuf_read(db, NULL, db_flags);
245 if (err != 0) {
246 dbuf_rele(db, tag);
247 *dbp = NULL;
248 }
249 }
250
251 return (err);
252 }
253
254 int
255 dmu_bonus_max(void)
256 {
257 return (DN_OLD_MAX_BONUSLEN);
258 }
259
260 int
261 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
262 {
263 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
264 dnode_t *dn;
265 int error;
266
267 DB_DNODE_ENTER(db);
268 dn = DB_DNODE(db);
269
270 if (dn->dn_bonus != db) {
271 error = SET_ERROR(EINVAL);
272 } else if (newsize < 0 || newsize > db_fake->db_size) {
273 error = SET_ERROR(EINVAL);
274 } else {
275 dnode_setbonuslen(dn, newsize, tx);
276 error = 0;
277 }
278
279 DB_DNODE_EXIT(db);
280 return (error);
281 }
282
283 int
284 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
285 {
286 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
287 dnode_t *dn;
288 int error;
289
290 DB_DNODE_ENTER(db);
291 dn = DB_DNODE(db);
292
293 if (!DMU_OT_IS_VALID(type)) {
294 error = SET_ERROR(EINVAL);
295 } else if (dn->dn_bonus != db) {
296 error = SET_ERROR(EINVAL);
297 } else {
298 dnode_setbonus_type(dn, type, tx);
299 error = 0;
300 }
301
302 DB_DNODE_EXIT(db);
303 return (error);
304 }
305
306 dmu_object_type_t
307 dmu_get_bonustype(dmu_buf_t *db_fake)
308 {
309 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
310 dnode_t *dn;
311 dmu_object_type_t type;
312
313 DB_DNODE_ENTER(db);
314 dn = DB_DNODE(db);
315 type = dn->dn_bonustype;
316 DB_DNODE_EXIT(db);
317
318 return (type);
319 }
320
321 int
322 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
323 {
324 dnode_t *dn;
325 int error;
326
327 error = dnode_hold(os, object, FTAG, &dn);
328 dbuf_rm_spill(dn, tx);
329 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
330 dnode_rm_spill(dn, tx);
331 rw_exit(&dn->dn_struct_rwlock);
332 dnode_rele(dn, FTAG);
333 return (error);
334 }
335
336 /*
337 * Lookup and hold the bonus buffer for the provided dnode. If the dnode
338 * has not yet been allocated a new bonus dbuf a will be allocated.
339 * Returns ENOENT, EIO, or 0.
340 */
341 int dmu_bonus_hold_by_dnode(dnode_t *dn, void *tag, dmu_buf_t **dbp,
342 uint32_t flags)
343 {
344 dmu_buf_impl_t *db;
345 int error;
346 uint32_t db_flags = DB_RF_MUST_SUCCEED;
347
348 if (flags & DMU_READ_NO_PREFETCH)
349 db_flags |= DB_RF_NOPREFETCH;
350 if (flags & DMU_READ_NO_DECRYPT)
351 db_flags |= DB_RF_NO_DECRYPT;
352
353 rw_enter(&dn->dn_struct_rwlock, RW_READER);
354 if (dn->dn_bonus == NULL) {
355 rw_exit(&dn->dn_struct_rwlock);
356 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
357 if (dn->dn_bonus == NULL)
358 dbuf_create_bonus(dn);
359 }
360 db = dn->dn_bonus;
361
362 /* as long as the bonus buf is held, the dnode will be held */
363 if (zfs_refcount_add(&db->db_holds, tag) == 1) {
364 VERIFY(dnode_add_ref(dn, db));
365 atomic_inc_32(&dn->dn_dbufs_count);
366 }
367
368 /*
369 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
370 * hold and incrementing the dbuf count to ensure that dnode_move() sees
371 * a dnode hold for every dbuf.
372 */
373 rw_exit(&dn->dn_struct_rwlock);
374
375 error = dbuf_read(db, NULL, db_flags);
376 if (error) {
377 dnode_evict_bonus(dn);
378 dbuf_rele(db, tag);
379 *dbp = NULL;
380 return (error);
381 }
382
383 *dbp = &db->db;
384 return (0);
385 }
386
387 int
388 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
389 {
390 dnode_t *dn;
391 int error;
392
393 error = dnode_hold(os, object, FTAG, &dn);
394 if (error)
395 return (error);
396
397 error = dmu_bonus_hold_by_dnode(dn, tag, dbp, DMU_READ_NO_PREFETCH);
398 dnode_rele(dn, FTAG);
399
400 return (error);
401 }
402
403 /*
404 * returns ENOENT, EIO, or 0.
405 *
406 * This interface will allocate a blank spill dbuf when a spill blk
407 * doesn't already exist on the dnode.
408 *
409 * if you only want to find an already existing spill db, then
410 * dmu_spill_hold_existing() should be used.
411 */
412 int
413 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
414 {
415 dmu_buf_impl_t *db = NULL;
416 int err;
417
418 if ((flags & DB_RF_HAVESTRUCT) == 0)
419 rw_enter(&dn->dn_struct_rwlock, RW_READER);
420
421 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
422
423 if ((flags & DB_RF_HAVESTRUCT) == 0)
424 rw_exit(&dn->dn_struct_rwlock);
425
426 if (db == NULL) {
427 *dbp = NULL;
428 return (SET_ERROR(EIO));
429 }
430 err = dbuf_read(db, NULL, flags);
431 if (err == 0)
432 *dbp = &db->db;
433 else {
434 dbuf_rele(db, tag);
435 *dbp = NULL;
436 }
437 return (err);
438 }
439
440 int
441 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
442 {
443 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
444 dnode_t *dn;
445 int err;
446
447 DB_DNODE_ENTER(db);
448 dn = DB_DNODE(db);
449
450 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
451 err = SET_ERROR(EINVAL);
452 } else {
453 rw_enter(&dn->dn_struct_rwlock, RW_READER);
454
455 if (!dn->dn_have_spill) {
456 err = SET_ERROR(ENOENT);
457 } else {
458 err = dmu_spill_hold_by_dnode(dn,
459 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
460 }
461
462 rw_exit(&dn->dn_struct_rwlock);
463 }
464
465 DB_DNODE_EXIT(db);
466 return (err);
467 }
468
469 int
470 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, uint32_t flags, void *tag,
471 dmu_buf_t **dbp)
472 {
473 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
474 dnode_t *dn;
475 int err;
476 uint32_t db_flags = DB_RF_CANFAIL;
477
478 if (flags & DMU_READ_NO_DECRYPT)
479 db_flags |= DB_RF_NO_DECRYPT;
480
481 DB_DNODE_ENTER(db);
482 dn = DB_DNODE(db);
483 err = dmu_spill_hold_by_dnode(dn, db_flags, tag, dbp);
484 DB_DNODE_EXIT(db);
485
486 return (err);
487 }
488
489 /*
490 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
491 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
492 * and can induce severe lock contention when writing to several files
493 * whose dnodes are in the same block.
494 */
495 int
496 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
497 boolean_t read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
498 {
499 dmu_buf_t **dbp;
500 zstream_t *zs = NULL;
501 uint64_t blkid, nblks, i;
502 uint32_t dbuf_flags;
503 int err;
504 zio_t *zio = NULL;
505 boolean_t missed = B_FALSE;
506
507 ASSERT(length <= DMU_MAX_ACCESS);
508
509 /*
510 * Note: We directly notify the prefetch code of this read, so that
511 * we can tell it about the multi-block read. dbuf_read() only knows
512 * about the one block it is accessing.
513 */
514 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT |
515 DB_RF_NOPREFETCH;
516
517 rw_enter(&dn->dn_struct_rwlock, RW_READER);
518 if (dn->dn_datablkshift) {
519 int blkshift = dn->dn_datablkshift;
520 nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) -
521 P2ALIGN(offset, 1ULL << blkshift)) >> blkshift;
522 } else {
523 if (offset + length > dn->dn_datablksz) {
524 zfs_panic_recover("zfs: accessing past end of object "
525 "%llx/%llx (size=%u access=%llu+%llu)",
526 (longlong_t)dn->dn_objset->
527 os_dsl_dataset->ds_object,
528 (longlong_t)dn->dn_object, dn->dn_datablksz,
529 (longlong_t)offset, (longlong_t)length);
530 rw_exit(&dn->dn_struct_rwlock);
531 return (SET_ERROR(EIO));
532 }
533 nblks = 1;
534 }
535 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
536
537 if (read)
538 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL,
539 ZIO_FLAG_CANFAIL);
540 blkid = dbuf_whichblock(dn, 0, offset);
541 if ((flags & DMU_READ_NO_PREFETCH) == 0 &&
542 DNODE_META_IS_CACHEABLE(dn) && length <= zfetch_array_rd_sz) {
543 /*
544 * Prepare the zfetch before initiating the demand reads, so
545 * that if multiple threads block on same indirect block, we
546 * base predictions on the original less racy request order.
547 */
548 zs = dmu_zfetch_prepare(&dn->dn_zfetch, blkid, nblks,
549 read && DNODE_IS_CACHEABLE(dn), B_TRUE);
550 }
551 for (i = 0; i < nblks; i++) {
552 dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag);
553 if (db == NULL) {
554 if (zs)
555 dmu_zfetch_run(zs, missed, B_TRUE);
556 rw_exit(&dn->dn_struct_rwlock);
557 dmu_buf_rele_array(dbp, nblks, tag);
558 if (read)
559 zio_nowait(zio);
560 return (SET_ERROR(EIO));
561 }
562
563 /*
564 * Initiate async demand data read.
565 * We check the db_state after calling dbuf_read() because
566 * (1) dbuf_read() may change the state to CACHED due to a
567 * hit in the ARC, and (2) on a cache miss, a child will
568 * have been added to "zio" but not yet completed, so the
569 * state will not yet be CACHED.
570 */
571 if (read) {
572 (void) dbuf_read(db, zio, dbuf_flags);
573 if (db->db_state != DB_CACHED)
574 missed = B_TRUE;
575 }
576 dbp[i] = &db->db;
577 }
578
579 if (!read)
580 zfs_racct_write(length, nblks);
581
582 if (zs)
583 dmu_zfetch_run(zs, missed, B_TRUE);
584 rw_exit(&dn->dn_struct_rwlock);
585
586 if (read) {
587 /* wait for async read i/o */
588 err = zio_wait(zio);
589 if (err) {
590 dmu_buf_rele_array(dbp, nblks, tag);
591 return (err);
592 }
593
594 /* wait for other io to complete */
595 for (i = 0; i < nblks; i++) {
596 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
597 mutex_enter(&db->db_mtx);
598 while (db->db_state == DB_READ ||
599 db->db_state == DB_FILL)
600 cv_wait(&db->db_changed, &db->db_mtx);
601 if (db->db_state == DB_UNCACHED)
602 err = SET_ERROR(EIO);
603 mutex_exit(&db->db_mtx);
604 if (err) {
605 dmu_buf_rele_array(dbp, nblks, tag);
606 return (err);
607 }
608 }
609 }
610
611 *numbufsp = nblks;
612 *dbpp = dbp;
613 return (0);
614 }
615
616 static int
617 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
618 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
619 {
620 dnode_t *dn;
621 int err;
622
623 err = dnode_hold(os, object, FTAG, &dn);
624 if (err)
625 return (err);
626
627 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
628 numbufsp, dbpp, DMU_READ_PREFETCH);
629
630 dnode_rele(dn, FTAG);
631
632 return (err);
633 }
634
635 int
636 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
637 uint64_t length, boolean_t read, void *tag, int *numbufsp,
638 dmu_buf_t ***dbpp)
639 {
640 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
641 dnode_t *dn;
642 int err;
643
644 DB_DNODE_ENTER(db);
645 dn = DB_DNODE(db);
646 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
647 numbufsp, dbpp, DMU_READ_PREFETCH);
648 DB_DNODE_EXIT(db);
649
650 return (err);
651 }
652
653 void
654 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
655 {
656 int i;
657 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
658
659 if (numbufs == 0)
660 return;
661
662 for (i = 0; i < numbufs; i++) {
663 if (dbp[i])
664 dbuf_rele(dbp[i], tag);
665 }
666
667 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
668 }
669
670 /*
671 * Issue prefetch i/os for the given blocks. If level is greater than 0, the
672 * indirect blocks prefetched will be those that point to the blocks containing
673 * the data starting at offset, and continuing to offset + len.
674 *
675 * Note that if the indirect blocks above the blocks being prefetched are not
676 * in cache, they will be asynchronously read in.
677 */
678 void
679 dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset,
680 uint64_t len, zio_priority_t pri)
681 {
682 dnode_t *dn;
683 uint64_t blkid;
684 int nblks, err;
685
686 if (len == 0) { /* they're interested in the bonus buffer */
687 dn = DMU_META_DNODE(os);
688
689 if (object == 0 || object >= DN_MAX_OBJECT)
690 return;
691
692 rw_enter(&dn->dn_struct_rwlock, RW_READER);
693 blkid = dbuf_whichblock(dn, level,
694 object * sizeof (dnode_phys_t));
695 dbuf_prefetch(dn, level, blkid, pri, 0);
696 rw_exit(&dn->dn_struct_rwlock);
697 return;
698 }
699
700 /*
701 * See comment before the definition of dmu_prefetch_max.
702 */
703 len = MIN(len, dmu_prefetch_max);
704
705 /*
706 * XXX - Note, if the dnode for the requested object is not
707 * already cached, we will do a *synchronous* read in the
708 * dnode_hold() call. The same is true for any indirects.
709 */
710 err = dnode_hold(os, object, FTAG, &dn);
711 if (err != 0)
712 return;
713
714 /*
715 * offset + len - 1 is the last byte we want to prefetch for, and offset
716 * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the
717 * last block we want to prefetch, and dbuf_whichblock(dn, level,
718 * offset) is the first. Then the number we need to prefetch is the
719 * last - first + 1.
720 */
721 rw_enter(&dn->dn_struct_rwlock, RW_READER);
722 if (level > 0 || dn->dn_datablkshift != 0) {
723 nblks = dbuf_whichblock(dn, level, offset + len - 1) -
724 dbuf_whichblock(dn, level, offset) + 1;
725 } else {
726 nblks = (offset < dn->dn_datablksz);
727 }
728
729 if (nblks != 0) {
730 blkid = dbuf_whichblock(dn, level, offset);
731 for (int i = 0; i < nblks; i++)
732 dbuf_prefetch(dn, level, blkid + i, pri, 0);
733 }
734 rw_exit(&dn->dn_struct_rwlock);
735
736 dnode_rele(dn, FTAG);
737 }
738
739 /*
740 * Get the next "chunk" of file data to free. We traverse the file from
741 * the end so that the file gets shorter over time (if we crashes in the
742 * middle, this will leave us in a better state). We find allocated file
743 * data by simply searching the allocated level 1 indirects.
744 *
745 * On input, *start should be the first offset that does not need to be
746 * freed (e.g. "offset + length"). On return, *start will be the first
747 * offset that should be freed and l1blks is set to the number of level 1
748 * indirect blocks found within the chunk.
749 */
750 static int
751 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum, uint64_t *l1blks)
752 {
753 uint64_t blks;
754 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
755 /* bytes of data covered by a level-1 indirect block */
756 uint64_t iblkrange = (uint64_t)dn->dn_datablksz *
757 EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
758
759 ASSERT3U(minimum, <=, *start);
760
761 /*
762 * Check if we can free the entire range assuming that all of the
763 * L1 blocks in this range have data. If we can, we use this
764 * worst case value as an estimate so we can avoid having to look
765 * at the object's actual data.
766 */
767 uint64_t total_l1blks =
768 (roundup(*start, iblkrange) - (minimum / iblkrange * iblkrange)) /
769 iblkrange;
770 if (total_l1blks <= maxblks) {
771 *l1blks = total_l1blks;
772 *start = minimum;
773 return (0);
774 }
775 ASSERT(ISP2(iblkrange));
776
777 for (blks = 0; *start > minimum && blks < maxblks; blks++) {
778 int err;
779
780 /*
781 * dnode_next_offset(BACKWARDS) will find an allocated L1
782 * indirect block at or before the input offset. We must
783 * decrement *start so that it is at the end of the region
784 * to search.
785 */
786 (*start)--;
787
788 err = dnode_next_offset(dn,
789 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
790
791 /* if there are no indirect blocks before start, we are done */
792 if (err == ESRCH) {
793 *start = minimum;
794 break;
795 } else if (err != 0) {
796 *l1blks = blks;
797 return (err);
798 }
799
800 /* set start to the beginning of this L1 indirect */
801 *start = P2ALIGN(*start, iblkrange);
802 }
803 if (*start < minimum)
804 *start = minimum;
805 *l1blks = blks;
806
807 return (0);
808 }
809
810 /*
811 * If this objset is of type OST_ZFS return true if vfs's unmounted flag is set,
812 * otherwise return false.
813 * Used below in dmu_free_long_range_impl() to enable abort when unmounting
814 */
815 /*ARGSUSED*/
816 static boolean_t
817 dmu_objset_zfs_unmounting(objset_t *os)
818 {
819 #ifdef _KERNEL
820 if (dmu_objset_type(os) == DMU_OST_ZFS)
821 return (zfs_get_vfs_flag_unmounted(os));
822 #endif
823 return (B_FALSE);
824 }
825
826 static int
827 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
828 uint64_t length)
829 {
830 uint64_t object_size;
831 int err;
832 uint64_t dirty_frees_threshold;
833 dsl_pool_t *dp = dmu_objset_pool(os);
834
835 if (dn == NULL)
836 return (SET_ERROR(EINVAL));
837
838 object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
839 if (offset >= object_size)
840 return (0);
841
842 if (zfs_per_txg_dirty_frees_percent <= 100)
843 dirty_frees_threshold =
844 zfs_per_txg_dirty_frees_percent * zfs_dirty_data_max / 100;
845 else
846 dirty_frees_threshold = zfs_dirty_data_max / 20;
847
848 if (length == DMU_OBJECT_END || offset + length > object_size)
849 length = object_size - offset;
850
851 while (length != 0) {
852 uint64_t chunk_end, chunk_begin, chunk_len;
853 uint64_t l1blks;
854 dmu_tx_t *tx;
855
856 if (dmu_objset_zfs_unmounting(dn->dn_objset))
857 return (SET_ERROR(EINTR));
858
859 chunk_end = chunk_begin = offset + length;
860
861 /* move chunk_begin backwards to the beginning of this chunk */
862 err = get_next_chunk(dn, &chunk_begin, offset, &l1blks);
863 if (err)
864 return (err);
865 ASSERT3U(chunk_begin, >=, offset);
866 ASSERT3U(chunk_begin, <=, chunk_end);
867
868 chunk_len = chunk_end - chunk_begin;
869
870 tx = dmu_tx_create(os);
871 dmu_tx_hold_free(tx, dn->dn_object, chunk_begin, chunk_len);
872
873 /*
874 * Mark this transaction as typically resulting in a net
875 * reduction in space used.
876 */
877 dmu_tx_mark_netfree(tx);
878 err = dmu_tx_assign(tx, TXG_WAIT);
879 if (err) {
880 dmu_tx_abort(tx);
881 return (err);
882 }
883
884 uint64_t txg = dmu_tx_get_txg(tx);
885
886 mutex_enter(&dp->dp_lock);
887 uint64_t long_free_dirty =
888 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK];
889 mutex_exit(&dp->dp_lock);
890
891 /*
892 * To avoid filling up a TXG with just frees, wait for
893 * the next TXG to open before freeing more chunks if
894 * we have reached the threshold of frees.
895 */
896 if (dirty_frees_threshold != 0 &&
897 long_free_dirty >= dirty_frees_threshold) {
898 DMU_TX_STAT_BUMP(dmu_tx_dirty_frees_delay);
899 dmu_tx_commit(tx);
900 txg_wait_open(dp, 0, B_TRUE);
901 continue;
902 }
903
904 /*
905 * In order to prevent unnecessary write throttling, for each
906 * TXG, we track the cumulative size of L1 blocks being dirtied
907 * in dnode_free_range() below. We compare this number to a
908 * tunable threshold, past which we prevent new L1 dirty freeing
909 * blocks from being added into the open TXG. See
910 * dmu_free_long_range_impl() for details. The threshold
911 * prevents write throttle activation due to dirty freeing L1
912 * blocks taking up a large percentage of zfs_dirty_data_max.
913 */
914 mutex_enter(&dp->dp_lock);
915 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] +=
916 l1blks << dn->dn_indblkshift;
917 mutex_exit(&dp->dp_lock);
918 DTRACE_PROBE3(free__long__range,
919 uint64_t, long_free_dirty, uint64_t, chunk_len,
920 uint64_t, txg);
921 dnode_free_range(dn, chunk_begin, chunk_len, tx);
922
923 dmu_tx_commit(tx);
924
925 length -= chunk_len;
926 }
927 return (0);
928 }
929
930 int
931 dmu_free_long_range(objset_t *os, uint64_t object,
932 uint64_t offset, uint64_t length)
933 {
934 dnode_t *dn;
935 int err;
936
937 err = dnode_hold(os, object, FTAG, &dn);
938 if (err != 0)
939 return (err);
940 err = dmu_free_long_range_impl(os, dn, offset, length);
941
942 /*
943 * It is important to zero out the maxblkid when freeing the entire
944 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
945 * will take the fast path, and (b) dnode_reallocate() can verify
946 * that the entire file has been freed.
947 */
948 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
949 dn->dn_maxblkid = 0;
950
951 dnode_rele(dn, FTAG);
952 return (err);
953 }
954
955 int
956 dmu_free_long_object(objset_t *os, uint64_t object)
957 {
958 dmu_tx_t *tx;
959 int err;
960
961 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
962 if (err != 0)
963 return (err);
964
965 tx = dmu_tx_create(os);
966 dmu_tx_hold_bonus(tx, object);
967 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
968 dmu_tx_mark_netfree(tx);
969 err = dmu_tx_assign(tx, TXG_WAIT);
970 if (err == 0) {
971 if (err == 0)
972 err = dmu_object_free(os, object, tx);
973
974 dmu_tx_commit(tx);
975 } else {
976 dmu_tx_abort(tx);
977 }
978
979 return (err);
980 }
981
982 int
983 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
984 uint64_t size, dmu_tx_t *tx)
985 {
986 dnode_t *dn;
987 int err = dnode_hold(os, object, FTAG, &dn);
988 if (err)
989 return (err);
990 ASSERT(offset < UINT64_MAX);
991 ASSERT(size == DMU_OBJECT_END || size <= UINT64_MAX - offset);
992 dnode_free_range(dn, offset, size, tx);
993 dnode_rele(dn, FTAG);
994 return (0);
995 }
996
997 static int
998 dmu_read_impl(dnode_t *dn, uint64_t offset, uint64_t size,
999 void *buf, uint32_t flags)
1000 {
1001 dmu_buf_t **dbp;
1002 int numbufs, err = 0;
1003
1004 /*
1005 * Deal with odd block sizes, where there can't be data past the first
1006 * block. If we ever do the tail block optimization, we will need to
1007 * handle that here as well.
1008 */
1009 if (dn->dn_maxblkid == 0) {
1010 uint64_t newsz = offset > dn->dn_datablksz ? 0 :
1011 MIN(size, dn->dn_datablksz - offset);
1012 bzero((char *)buf + newsz, size - newsz);
1013 size = newsz;
1014 }
1015
1016 while (size > 0) {
1017 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
1018 int i;
1019
1020 /*
1021 * NB: we could do this block-at-a-time, but it's nice
1022 * to be reading in parallel.
1023 */
1024 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
1025 TRUE, FTAG, &numbufs, &dbp, flags);
1026 if (err)
1027 break;
1028
1029 for (i = 0; i < numbufs; i++) {
1030 uint64_t tocpy;
1031 int64_t bufoff;
1032 dmu_buf_t *db = dbp[i];
1033
1034 ASSERT(size > 0);
1035
1036 bufoff = offset - db->db_offset;
1037 tocpy = MIN(db->db_size - bufoff, size);
1038
1039 (void) memcpy(buf, (char *)db->db_data + bufoff, tocpy);
1040
1041 offset += tocpy;
1042 size -= tocpy;
1043 buf = (char *)buf + tocpy;
1044 }
1045 dmu_buf_rele_array(dbp, numbufs, FTAG);
1046 }
1047 return (err);
1048 }
1049
1050 int
1051 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1052 void *buf, uint32_t flags)
1053 {
1054 dnode_t *dn;
1055 int err;
1056
1057 err = dnode_hold(os, object, FTAG, &dn);
1058 if (err != 0)
1059 return (err);
1060
1061 err = dmu_read_impl(dn, offset, size, buf, flags);
1062 dnode_rele(dn, FTAG);
1063 return (err);
1064 }
1065
1066 int
1067 dmu_read_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, void *buf,
1068 uint32_t flags)
1069 {
1070 return (dmu_read_impl(dn, offset, size, buf, flags));
1071 }
1072
1073 static void
1074 dmu_write_impl(dmu_buf_t **dbp, int numbufs, uint64_t offset, uint64_t size,
1075 const void *buf, dmu_tx_t *tx)
1076 {
1077 int i;
1078
1079 for (i = 0; i < numbufs; i++) {
1080 uint64_t tocpy;
1081 int64_t bufoff;
1082 dmu_buf_t *db = dbp[i];
1083
1084 ASSERT(size > 0);
1085
1086 bufoff = offset - db->db_offset;
1087 tocpy = MIN(db->db_size - bufoff, size);
1088
1089 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1090
1091 if (tocpy == db->db_size)
1092 dmu_buf_will_fill(db, tx);
1093 else
1094 dmu_buf_will_dirty(db, tx);
1095
1096 (void) memcpy((char *)db->db_data + bufoff, buf, tocpy);
1097
1098 if (tocpy == db->db_size)
1099 dmu_buf_fill_done(db, tx);
1100
1101 offset += tocpy;
1102 size -= tocpy;
1103 buf = (char *)buf + tocpy;
1104 }
1105 }
1106
1107 void
1108 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1109 const void *buf, dmu_tx_t *tx)
1110 {
1111 dmu_buf_t **dbp;
1112 int numbufs;
1113
1114 if (size == 0)
1115 return;
1116
1117 VERIFY0(dmu_buf_hold_array(os, object, offset, size,
1118 FALSE, FTAG, &numbufs, &dbp));
1119 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1120 dmu_buf_rele_array(dbp, numbufs, FTAG);
1121 }
1122
1123 /*
1124 * Note: Lustre is an external consumer of this interface.
1125 */
1126 void
1127 dmu_write_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size,
1128 const void *buf, dmu_tx_t *tx)
1129 {
1130 dmu_buf_t **dbp;
1131 int numbufs;
1132
1133 if (size == 0)
1134 return;
1135
1136 VERIFY0(dmu_buf_hold_array_by_dnode(dn, offset, size,
1137 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH));
1138 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1139 dmu_buf_rele_array(dbp, numbufs, FTAG);
1140 }
1141
1142 void
1143 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1144 dmu_tx_t *tx)
1145 {
1146 dmu_buf_t **dbp;
1147 int numbufs, i;
1148
1149 if (size == 0)
1150 return;
1151
1152 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
1153 FALSE, FTAG, &numbufs, &dbp));
1154
1155 for (i = 0; i < numbufs; i++) {
1156 dmu_buf_t *db = dbp[i];
1157
1158 dmu_buf_will_not_fill(db, tx);
1159 }
1160 dmu_buf_rele_array(dbp, numbufs, FTAG);
1161 }
1162
1163 void
1164 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
1165 void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
1166 int compressed_size, int byteorder, dmu_tx_t *tx)
1167 {
1168 dmu_buf_t *db;
1169
1170 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
1171 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
1172 VERIFY0(dmu_buf_hold_noread(os, object, offset,
1173 FTAG, &db));
1174
1175 dmu_buf_write_embedded(db,
1176 data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
1177 uncompressed_size, compressed_size, byteorder, tx);
1178
1179 dmu_buf_rele(db, FTAG);
1180 }
1181
1182 void
1183 dmu_redact(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1184 dmu_tx_t *tx)
1185 {
1186 int numbufs, i;
1187 dmu_buf_t **dbp;
1188
1189 VERIFY0(dmu_buf_hold_array(os, object, offset, size, FALSE, FTAG,
1190 &numbufs, &dbp));
1191 for (i = 0; i < numbufs; i++)
1192 dmu_buf_redact(dbp[i], tx);
1193 dmu_buf_rele_array(dbp, numbufs, FTAG);
1194 }
1195
1196 #ifdef _KERNEL
1197 int
1198 dmu_read_uio_dnode(dnode_t *dn, zfs_uio_t *uio, uint64_t size)
1199 {
1200 dmu_buf_t **dbp;
1201 int numbufs, i, err;
1202
1203 /*
1204 * NB: we could do this block-at-a-time, but it's nice
1205 * to be reading in parallel.
1206 */
1207 err = dmu_buf_hold_array_by_dnode(dn, zfs_uio_offset(uio), size,
1208 TRUE, FTAG, &numbufs, &dbp, 0);
1209 if (err)
1210 return (err);
1211
1212 for (i = 0; i < numbufs; i++) {
1213 uint64_t tocpy;
1214 int64_t bufoff;
1215 dmu_buf_t *db = dbp[i];
1216
1217 ASSERT(size > 0);
1218
1219 bufoff = zfs_uio_offset(uio) - db->db_offset;
1220 tocpy = MIN(db->db_size - bufoff, size);
1221
1222 err = zfs_uio_fault_move((char *)db->db_data + bufoff, tocpy,
1223 UIO_READ, uio);
1224
1225 if (err)
1226 break;
1227
1228 size -= tocpy;
1229 }
1230 dmu_buf_rele_array(dbp, numbufs, FTAG);
1231
1232 return (err);
1233 }
1234
1235 /*
1236 * Read 'size' bytes into the uio buffer.
1237 * From object zdb->db_object.
1238 * Starting at zfs_uio_offset(uio).
1239 *
1240 * If the caller already has a dbuf in the target object
1241 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1242 * because we don't have to find the dnode_t for the object.
1243 */
1244 int
1245 dmu_read_uio_dbuf(dmu_buf_t *zdb, zfs_uio_t *uio, uint64_t size)
1246 {
1247 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1248 dnode_t *dn;
1249 int err;
1250
1251 if (size == 0)
1252 return (0);
1253
1254 DB_DNODE_ENTER(db);
1255 dn = DB_DNODE(db);
1256 err = dmu_read_uio_dnode(dn, uio, size);
1257 DB_DNODE_EXIT(db);
1258
1259 return (err);
1260 }
1261
1262 /*
1263 * Read 'size' bytes into the uio buffer.
1264 * From the specified object
1265 * Starting at offset zfs_uio_offset(uio).
1266 */
1267 int
1268 dmu_read_uio(objset_t *os, uint64_t object, zfs_uio_t *uio, uint64_t size)
1269 {
1270 dnode_t *dn;
1271 int err;
1272
1273 if (size == 0)
1274 return (0);
1275
1276 err = dnode_hold(os, object, FTAG, &dn);
1277 if (err)
1278 return (err);
1279
1280 err = dmu_read_uio_dnode(dn, uio, size);
1281
1282 dnode_rele(dn, FTAG);
1283
1284 return (err);
1285 }
1286
1287 int
1288 dmu_write_uio_dnode(dnode_t *dn, zfs_uio_t *uio, uint64_t size, dmu_tx_t *tx)
1289 {
1290 dmu_buf_t **dbp;
1291 int numbufs;
1292 int err = 0;
1293 int i;
1294
1295 err = dmu_buf_hold_array_by_dnode(dn, zfs_uio_offset(uio), size,
1296 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1297 if (err)
1298 return (err);
1299
1300 for (i = 0; i < numbufs; i++) {
1301 uint64_t tocpy;
1302 int64_t bufoff;
1303 dmu_buf_t *db = dbp[i];
1304
1305 ASSERT(size > 0);
1306
1307 bufoff = zfs_uio_offset(uio) - db->db_offset;
1308 tocpy = MIN(db->db_size - bufoff, size);
1309
1310 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1311
1312 if (tocpy == db->db_size)
1313 dmu_buf_will_fill(db, tx);
1314 else
1315 dmu_buf_will_dirty(db, tx);
1316
1317 /*
1318 * XXX zfs_uiomove could block forever (eg.nfs-backed
1319 * pages). There needs to be a uiolockdown() function
1320 * to lock the pages in memory, so that zfs_uiomove won't
1321 * block.
1322 */
1323 err = zfs_uio_fault_move((char *)db->db_data + bufoff,
1324 tocpy, UIO_WRITE, uio);
1325
1326 if (tocpy == db->db_size)
1327 dmu_buf_fill_done(db, tx);
1328
1329 if (err)
1330 break;
1331
1332 size -= tocpy;
1333 }
1334
1335 dmu_buf_rele_array(dbp, numbufs, FTAG);
1336 return (err);
1337 }
1338
1339 /*
1340 * Write 'size' bytes from the uio buffer.
1341 * To object zdb->db_object.
1342 * Starting at offset zfs_uio_offset(uio).
1343 *
1344 * If the caller already has a dbuf in the target object
1345 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1346 * because we don't have to find the dnode_t for the object.
1347 */
1348 int
1349 dmu_write_uio_dbuf(dmu_buf_t *zdb, zfs_uio_t *uio, uint64_t size,
1350 dmu_tx_t *tx)
1351 {
1352 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1353 dnode_t *dn;
1354 int err;
1355
1356 if (size == 0)
1357 return (0);
1358
1359 DB_DNODE_ENTER(db);
1360 dn = DB_DNODE(db);
1361 err = dmu_write_uio_dnode(dn, uio, size, tx);
1362 DB_DNODE_EXIT(db);
1363
1364 return (err);
1365 }
1366
1367 /*
1368 * Write 'size' bytes from the uio buffer.
1369 * To the specified object.
1370 * Starting at offset zfs_uio_offset(uio).
1371 */
1372 int
1373 dmu_write_uio(objset_t *os, uint64_t object, zfs_uio_t *uio, uint64_t size,
1374 dmu_tx_t *tx)
1375 {
1376 dnode_t *dn;
1377 int err;
1378
1379 if (size == 0)
1380 return (0);
1381
1382 err = dnode_hold(os, object, FTAG, &dn);
1383 if (err)
1384 return (err);
1385
1386 err = dmu_write_uio_dnode(dn, uio, size, tx);
1387
1388 dnode_rele(dn, FTAG);
1389
1390 return (err);
1391 }
1392 #endif /* _KERNEL */
1393
1394 /*
1395 * Allocate a loaned anonymous arc buffer.
1396 */
1397 arc_buf_t *
1398 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1399 {
1400 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1401
1402 return (arc_loan_buf(db->db_objset->os_spa, B_FALSE, size));
1403 }
1404
1405 /*
1406 * Free a loaned arc buffer.
1407 */
1408 void
1409 dmu_return_arcbuf(arc_buf_t *buf)
1410 {
1411 arc_return_buf(buf, FTAG);
1412 arc_buf_destroy(buf, FTAG);
1413 }
1414
1415 /*
1416 * A "lightweight" write is faster than a regular write (e.g.
1417 * dmu_write_by_dnode() or dmu_assign_arcbuf_by_dnode()), because it avoids the
1418 * CPU cost of creating a dmu_buf_impl_t and arc_buf_[hdr_]_t. However, the
1419 * data can not be read or overwritten until the transaction's txg has been
1420 * synced. This makes it appropriate for workloads that are known to be
1421 * (temporarily) write-only, like "zfs receive".
1422 *
1423 * A single block is written, starting at the specified offset in bytes. If
1424 * the call is successful, it returns 0 and the provided abd has been
1425 * consumed (the caller should not free it).
1426 */
1427 int
1428 dmu_lightweight_write_by_dnode(dnode_t *dn, uint64_t offset, abd_t *abd,
1429 const zio_prop_t *zp, enum zio_flag flags, dmu_tx_t *tx)
1430 {
1431 dbuf_dirty_record_t *dr =
1432 dbuf_dirty_lightweight(dn, dbuf_whichblock(dn, 0, offset), tx);
1433 if (dr == NULL)
1434 return (SET_ERROR(EIO));
1435 dr->dt.dll.dr_abd = abd;
1436 dr->dt.dll.dr_props = *zp;
1437 dr->dt.dll.dr_flags = flags;
1438 return (0);
1439 }
1440
1441 /*
1442 * When possible directly assign passed loaned arc buffer to a dbuf.
1443 * If this is not possible copy the contents of passed arc buf via
1444 * dmu_write().
1445 */
1446 int
1447 dmu_assign_arcbuf_by_dnode(dnode_t *dn, uint64_t offset, arc_buf_t *buf,
1448 dmu_tx_t *tx)
1449 {
1450 dmu_buf_impl_t *db;
1451 objset_t *os = dn->dn_objset;
1452 uint64_t object = dn->dn_object;
1453 uint32_t blksz = (uint32_t)arc_buf_lsize(buf);
1454 uint64_t blkid;
1455
1456 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1457 blkid = dbuf_whichblock(dn, 0, offset);
1458 db = dbuf_hold(dn, blkid, FTAG);
1459 if (db == NULL)
1460 return (SET_ERROR(EIO));
1461 rw_exit(&dn->dn_struct_rwlock);
1462
1463 /*
1464 * We can only assign if the offset is aligned and the arc buf is the
1465 * same size as the dbuf.
1466 */
1467 if (offset == db->db.db_offset && blksz == db->db.db_size) {
1468 zfs_racct_write(blksz, 1);
1469 dbuf_assign_arcbuf(db, buf, tx);
1470 dbuf_rele(db, FTAG);
1471 } else {
1472 /* compressed bufs must always be assignable to their dbuf */
1473 ASSERT3U(arc_get_compression(buf), ==, ZIO_COMPRESS_OFF);
1474 ASSERT(!(buf->b_flags & ARC_BUF_FLAG_COMPRESSED));
1475
1476 dbuf_rele(db, FTAG);
1477 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1478 dmu_return_arcbuf(buf);
1479 }
1480
1481 return (0);
1482 }
1483
1484 int
1485 dmu_assign_arcbuf_by_dbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1486 dmu_tx_t *tx)
1487 {
1488 int err;
1489 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1490
1491 DB_DNODE_ENTER(dbuf);
1492 err = dmu_assign_arcbuf_by_dnode(DB_DNODE(dbuf), offset, buf, tx);
1493 DB_DNODE_EXIT(dbuf);
1494
1495 return (err);
1496 }
1497
1498 typedef struct {
1499 dbuf_dirty_record_t *dsa_dr;
1500 dmu_sync_cb_t *dsa_done;
1501 zgd_t *dsa_zgd;
1502 dmu_tx_t *dsa_tx;
1503 } dmu_sync_arg_t;
1504
1505 /* ARGSUSED */
1506 static void
1507 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1508 {
1509 dmu_sync_arg_t *dsa = varg;
1510 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1511 blkptr_t *bp = zio->io_bp;
1512
1513 if (zio->io_error == 0) {
1514 if (BP_IS_HOLE(bp)) {
1515 /*
1516 * A block of zeros may compress to a hole, but the
1517 * block size still needs to be known for replay.
1518 */
1519 BP_SET_LSIZE(bp, db->db_size);
1520 } else if (!BP_IS_EMBEDDED(bp)) {
1521 ASSERT(BP_GET_LEVEL(bp) == 0);
1522 BP_SET_FILL(bp, 1);
1523 }
1524 }
1525 }
1526
1527 static void
1528 dmu_sync_late_arrival_ready(zio_t *zio)
1529 {
1530 dmu_sync_ready(zio, NULL, zio->io_private);
1531 }
1532
1533 /* ARGSUSED */
1534 static void
1535 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1536 {
1537 dmu_sync_arg_t *dsa = varg;
1538 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1539 dmu_buf_impl_t *db = dr->dr_dbuf;
1540 zgd_t *zgd = dsa->dsa_zgd;
1541
1542 /*
1543 * Record the vdev(s) backing this blkptr so they can be flushed after
1544 * the writes for the lwb have completed.
1545 */
1546 if (zio->io_error == 0) {
1547 zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp);
1548 }
1549
1550 mutex_enter(&db->db_mtx);
1551 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1552 if (zio->io_error == 0) {
1553 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1554 if (dr->dt.dl.dr_nopwrite) {
1555 blkptr_t *bp = zio->io_bp;
1556 blkptr_t *bp_orig = &zio->io_bp_orig;
1557 uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1558
1559 ASSERT(BP_EQUAL(bp, bp_orig));
1560 VERIFY(BP_EQUAL(bp, db->db_blkptr));
1561 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1562 VERIFY(zio_checksum_table[chksum].ci_flags &
1563 ZCHECKSUM_FLAG_NOPWRITE);
1564 }
1565 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1566 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1567 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1568
1569 /*
1570 * Old style holes are filled with all zeros, whereas
1571 * new-style holes maintain their lsize, type, level,
1572 * and birth time (see zio_write_compress). While we
1573 * need to reset the BP_SET_LSIZE() call that happened
1574 * in dmu_sync_ready for old style holes, we do *not*
1575 * want to wipe out the information contained in new
1576 * style holes. Thus, only zero out the block pointer if
1577 * it's an old style hole.
1578 */
1579 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) &&
1580 dr->dt.dl.dr_overridden_by.blk_birth == 0)
1581 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1582 } else {
1583 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1584 }
1585 cv_broadcast(&db->db_changed);
1586 mutex_exit(&db->db_mtx);
1587
1588 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1589
1590 kmem_free(dsa, sizeof (*dsa));
1591 }
1592
1593 static void
1594 dmu_sync_late_arrival_done(zio_t *zio)
1595 {
1596 blkptr_t *bp = zio->io_bp;
1597 dmu_sync_arg_t *dsa = zio->io_private;
1598 zgd_t *zgd = dsa->dsa_zgd;
1599
1600 if (zio->io_error == 0) {
1601 /*
1602 * Record the vdev(s) backing this blkptr so they can be
1603 * flushed after the writes for the lwb have completed.
1604 */
1605 zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp);
1606
1607 if (!BP_IS_HOLE(bp)) {
1608 blkptr_t *bp_orig __maybe_unused = &zio->io_bp_orig;
1609 ASSERT(!(zio->io_flags & ZIO_FLAG_NOPWRITE));
1610 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1611 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1612 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1613 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1614 }
1615 }
1616
1617 dmu_tx_commit(dsa->dsa_tx);
1618
1619 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1620
1621 abd_free(zio->io_abd);
1622 kmem_free(dsa, sizeof (*dsa));
1623 }
1624
1625 static int
1626 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1627 zio_prop_t *zp, zbookmark_phys_t *zb)
1628 {
1629 dmu_sync_arg_t *dsa;
1630 dmu_tx_t *tx;
1631
1632 tx = dmu_tx_create(os);
1633 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1634 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1635 dmu_tx_abort(tx);
1636 /* Make zl_get_data do txg_waited_synced() */
1637 return (SET_ERROR(EIO));
1638 }
1639
1640 /*
1641 * In order to prevent the zgd's lwb from being free'd prior to
1642 * dmu_sync_late_arrival_done() being called, we have to ensure
1643 * the lwb's "max txg" takes this tx's txg into account.
1644 */
1645 zil_lwb_add_txg(zgd->zgd_lwb, dmu_tx_get_txg(tx));
1646
1647 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1648 dsa->dsa_dr = NULL;
1649 dsa->dsa_done = done;
1650 dsa->dsa_zgd = zgd;
1651 dsa->dsa_tx = tx;
1652
1653 /*
1654 * Since we are currently syncing this txg, it's nontrivial to
1655 * determine what BP to nopwrite against, so we disable nopwrite.
1656 *
1657 * When syncing, the db_blkptr is initially the BP of the previous
1658 * txg. We can not nopwrite against it because it will be changed
1659 * (this is similar to the non-late-arrival case where the dbuf is
1660 * dirty in a future txg).
1661 *
1662 * Then dbuf_write_ready() sets bp_blkptr to the location we will write.
1663 * We can not nopwrite against it because although the BP will not
1664 * (typically) be changed, the data has not yet been persisted to this
1665 * location.
1666 *
1667 * Finally, when dbuf_write_done() is called, it is theoretically
1668 * possible to always nopwrite, because the data that was written in
1669 * this txg is the same data that we are trying to write. However we
1670 * would need to check that this dbuf is not dirty in any future
1671 * txg's (as we do in the normal dmu_sync() path). For simplicity, we
1672 * don't nopwrite in this case.
1673 */
1674 zp->zp_nopwrite = B_FALSE;
1675
1676 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1677 abd_get_from_buf(zgd->zgd_db->db_data, zgd->zgd_db->db_size),
1678 zgd->zgd_db->db_size, zgd->zgd_db->db_size, zp,
1679 dmu_sync_late_arrival_ready, NULL, NULL, dmu_sync_late_arrival_done,
1680 dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1681
1682 return (0);
1683 }
1684
1685 /*
1686 * Intent log support: sync the block associated with db to disk.
1687 * N.B. and XXX: the caller is responsible for making sure that the
1688 * data isn't changing while dmu_sync() is writing it.
1689 *
1690 * Return values:
1691 *
1692 * EEXIST: this txg has already been synced, so there's nothing to do.
1693 * The caller should not log the write.
1694 *
1695 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1696 * The caller should not log the write.
1697 *
1698 * EALREADY: this block is already in the process of being synced.
1699 * The caller should track its progress (somehow).
1700 *
1701 * EIO: could not do the I/O.
1702 * The caller should do a txg_wait_synced().
1703 *
1704 * 0: the I/O has been initiated.
1705 * The caller should log this blkptr in the done callback.
1706 * It is possible that the I/O will fail, in which case
1707 * the error will be reported to the done callback and
1708 * propagated to pio from zio_done().
1709 */
1710 int
1711 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1712 {
1713 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1714 objset_t *os = db->db_objset;
1715 dsl_dataset_t *ds = os->os_dsl_dataset;
1716 dbuf_dirty_record_t *dr, *dr_next;
1717 dmu_sync_arg_t *dsa;
1718 zbookmark_phys_t zb;
1719 zio_prop_t zp;
1720 dnode_t *dn;
1721
1722 ASSERT(pio != NULL);
1723 ASSERT(txg != 0);
1724
1725 SET_BOOKMARK(&zb, ds->ds_object,
1726 db->db.db_object, db->db_level, db->db_blkid);
1727
1728 DB_DNODE_ENTER(db);
1729 dn = DB_DNODE(db);
1730 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1731 DB_DNODE_EXIT(db);
1732
1733 /*
1734 * If we're frozen (running ziltest), we always need to generate a bp.
1735 */
1736 if (txg > spa_freeze_txg(os->os_spa))
1737 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1738
1739 /*
1740 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1741 * and us. If we determine that this txg is not yet syncing,
1742 * but it begins to sync a moment later, that's OK because the
1743 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1744 */
1745 mutex_enter(&db->db_mtx);
1746
1747 if (txg <= spa_last_synced_txg(os->os_spa)) {
1748 /*
1749 * This txg has already synced. There's nothing to do.
1750 */
1751 mutex_exit(&db->db_mtx);
1752 return (SET_ERROR(EEXIST));
1753 }
1754
1755 if (txg <= spa_syncing_txg(os->os_spa)) {
1756 /*
1757 * This txg is currently syncing, so we can't mess with
1758 * the dirty record anymore; just write a new log block.
1759 */
1760 mutex_exit(&db->db_mtx);
1761 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1762 }
1763
1764 dr = dbuf_find_dirty_eq(db, txg);
1765
1766 if (dr == NULL) {
1767 /*
1768 * There's no dr for this dbuf, so it must have been freed.
1769 * There's no need to log writes to freed blocks, so we're done.
1770 */
1771 mutex_exit(&db->db_mtx);
1772 return (SET_ERROR(ENOENT));
1773 }
1774
1775 dr_next = list_next(&db->db_dirty_records, dr);
1776 ASSERT(dr_next == NULL || dr_next->dr_txg < txg);
1777
1778 if (db->db_blkptr != NULL) {
1779 /*
1780 * We need to fill in zgd_bp with the current blkptr so that
1781 * the nopwrite code can check if we're writing the same
1782 * data that's already on disk. We can only nopwrite if we
1783 * are sure that after making the copy, db_blkptr will not
1784 * change until our i/o completes. We ensure this by
1785 * holding the db_mtx, and only allowing nopwrite if the
1786 * block is not already dirty (see below). This is verified
1787 * by dmu_sync_done(), which VERIFYs that the db_blkptr has
1788 * not changed.
1789 */
1790 *zgd->zgd_bp = *db->db_blkptr;
1791 }
1792
1793 /*
1794 * Assume the on-disk data is X, the current syncing data (in
1795 * txg - 1) is Y, and the current in-memory data is Z (currently
1796 * in dmu_sync).
1797 *
1798 * We usually want to perform a nopwrite if X and Z are the
1799 * same. However, if Y is different (i.e. the BP is going to
1800 * change before this write takes effect), then a nopwrite will
1801 * be incorrect - we would override with X, which could have
1802 * been freed when Y was written.
1803 *
1804 * (Note that this is not a concern when we are nop-writing from
1805 * syncing context, because X and Y must be identical, because
1806 * all previous txgs have been synced.)
1807 *
1808 * Therefore, we disable nopwrite if the current BP could change
1809 * before this TXG. There are two ways it could change: by
1810 * being dirty (dr_next is non-NULL), or by being freed
1811 * (dnode_block_freed()). This behavior is verified by
1812 * zio_done(), which VERIFYs that the override BP is identical
1813 * to the on-disk BP.
1814 */
1815 DB_DNODE_ENTER(db);
1816 dn = DB_DNODE(db);
1817 if (dr_next != NULL || dnode_block_freed(dn, db->db_blkid))
1818 zp.zp_nopwrite = B_FALSE;
1819 DB_DNODE_EXIT(db);
1820
1821 ASSERT(dr->dr_txg == txg);
1822 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1823 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1824 /*
1825 * We have already issued a sync write for this buffer,
1826 * or this buffer has already been synced. It could not
1827 * have been dirtied since, or we would have cleared the state.
1828 */
1829 mutex_exit(&db->db_mtx);
1830 return (SET_ERROR(EALREADY));
1831 }
1832
1833 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1834 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1835 mutex_exit(&db->db_mtx);
1836
1837 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1838 dsa->dsa_dr = dr;
1839 dsa->dsa_done = done;
1840 dsa->dsa_zgd = zgd;
1841 dsa->dsa_tx = NULL;
1842
1843 zio_nowait(arc_write(pio, os->os_spa, txg,
1844 zgd->zgd_bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
1845 &zp, dmu_sync_ready, NULL, NULL, dmu_sync_done, dsa,
1846 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb));
1847
1848 return (0);
1849 }
1850
1851 int
1852 dmu_object_set_nlevels(objset_t *os, uint64_t object, int nlevels, dmu_tx_t *tx)
1853 {
1854 dnode_t *dn;
1855 int err;
1856
1857 err = dnode_hold(os, object, FTAG, &dn);
1858 if (err)
1859 return (err);
1860 err = dnode_set_nlevels(dn, nlevels, tx);
1861 dnode_rele(dn, FTAG);
1862 return (err);
1863 }
1864
1865 int
1866 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1867 dmu_tx_t *tx)
1868 {
1869 dnode_t *dn;
1870 int err;
1871
1872 err = dnode_hold(os, object, FTAG, &dn);
1873 if (err)
1874 return (err);
1875 err = dnode_set_blksz(dn, size, ibs, tx);
1876 dnode_rele(dn, FTAG);
1877 return (err);
1878 }
1879
1880 int
1881 dmu_object_set_maxblkid(objset_t *os, uint64_t object, uint64_t maxblkid,
1882 dmu_tx_t *tx)
1883 {
1884 dnode_t *dn;
1885 int err;
1886
1887 err = dnode_hold(os, object, FTAG, &dn);
1888 if (err)
1889 return (err);
1890 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1891 dnode_new_blkid(dn, maxblkid, tx, B_FALSE, B_TRUE);
1892 rw_exit(&dn->dn_struct_rwlock);
1893 dnode_rele(dn, FTAG);
1894 return (0);
1895 }
1896
1897 void
1898 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1899 dmu_tx_t *tx)
1900 {
1901 dnode_t *dn;
1902
1903 /*
1904 * Send streams include each object's checksum function. This
1905 * check ensures that the receiving system can understand the
1906 * checksum function transmitted.
1907 */
1908 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS);
1909
1910 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1911 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS);
1912 dn->dn_checksum = checksum;
1913 dnode_setdirty(dn, tx);
1914 dnode_rele(dn, FTAG);
1915 }
1916
1917 void
1918 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1919 dmu_tx_t *tx)
1920 {
1921 dnode_t *dn;
1922
1923 /*
1924 * Send streams include each object's compression function. This
1925 * check ensures that the receiving system can understand the
1926 * compression function transmitted.
1927 */
1928 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS);
1929
1930 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1931 dn->dn_compress = compress;
1932 dnode_setdirty(dn, tx);
1933 dnode_rele(dn, FTAG);
1934 }
1935
1936 /*
1937 * When the "redundant_metadata" property is set to "most", only indirect
1938 * blocks of this level and higher will have an additional ditto block.
1939 */
1940 int zfs_redundant_metadata_most_ditto_level = 2;
1941
1942 void
1943 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
1944 {
1945 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
1946 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
1947 (wp & WP_SPILL));
1948 enum zio_checksum checksum = os->os_checksum;
1949 enum zio_compress compress = os->os_compress;
1950 uint8_t complevel = os->os_complevel;
1951 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
1952 boolean_t dedup = B_FALSE;
1953 boolean_t nopwrite = B_FALSE;
1954 boolean_t dedup_verify = os->os_dedup_verify;
1955 boolean_t encrypt = B_FALSE;
1956 int copies = os->os_copies;
1957
1958 /*
1959 * We maintain different write policies for each of the following
1960 * types of data:
1961 * 1. metadata
1962 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1963 * 3. all other level 0 blocks
1964 */
1965 if (ismd) {
1966 /*
1967 * XXX -- we should design a compression algorithm
1968 * that specializes in arrays of bps.
1969 */
1970 compress = zio_compress_select(os->os_spa,
1971 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON);
1972
1973 /*
1974 * Metadata always gets checksummed. If the data
1975 * checksum is multi-bit correctable, and it's not a
1976 * ZBT-style checksum, then it's suitable for metadata
1977 * as well. Otherwise, the metadata checksum defaults
1978 * to fletcher4.
1979 */
1980 if (!(zio_checksum_table[checksum].ci_flags &
1981 ZCHECKSUM_FLAG_METADATA) ||
1982 (zio_checksum_table[checksum].ci_flags &
1983 ZCHECKSUM_FLAG_EMBEDDED))
1984 checksum = ZIO_CHECKSUM_FLETCHER_4;
1985
1986 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL ||
1987 (os->os_redundant_metadata ==
1988 ZFS_REDUNDANT_METADATA_MOST &&
1989 (level >= zfs_redundant_metadata_most_ditto_level ||
1990 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))))
1991 copies++;
1992 } else if (wp & WP_NOFILL) {
1993 ASSERT(level == 0);
1994
1995 /*
1996 * If we're writing preallocated blocks, we aren't actually
1997 * writing them so don't set any policy properties. These
1998 * blocks are currently only used by an external subsystem
1999 * outside of zfs (i.e. dump) and not written by the zio
2000 * pipeline.
2001 */
2002 compress = ZIO_COMPRESS_OFF;
2003 checksum = ZIO_CHECKSUM_OFF;
2004 } else {
2005 compress = zio_compress_select(os->os_spa, dn->dn_compress,
2006 compress);
2007 complevel = zio_complevel_select(os->os_spa, compress,
2008 complevel, complevel);
2009
2010 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
2011 zio_checksum_select(dn->dn_checksum, checksum) :
2012 dedup_checksum;
2013
2014 /*
2015 * Determine dedup setting. If we are in dmu_sync(),
2016 * we won't actually dedup now because that's all
2017 * done in syncing context; but we do want to use the
2018 * dedup checksum. If the checksum is not strong
2019 * enough to ensure unique signatures, force
2020 * dedup_verify.
2021 */
2022 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
2023 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
2024 if (!(zio_checksum_table[checksum].ci_flags &
2025 ZCHECKSUM_FLAG_DEDUP))
2026 dedup_verify = B_TRUE;
2027 }
2028
2029 /*
2030 * Enable nopwrite if we have secure enough checksum
2031 * algorithm (see comment in zio_nop_write) and
2032 * compression is enabled. We don't enable nopwrite if
2033 * dedup is enabled as the two features are mutually
2034 * exclusive.
2035 */
2036 nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags &
2037 ZCHECKSUM_FLAG_NOPWRITE) &&
2038 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
2039 }
2040
2041 /*
2042 * All objects in an encrypted objset are protected from modification
2043 * via a MAC. Encrypted objects store their IV and salt in the last DVA
2044 * in the bp, so we cannot use all copies. Encrypted objects are also
2045 * not subject to nopwrite since writing the same data will still
2046 * result in a new ciphertext. Only encrypted blocks can be dedup'd
2047 * to avoid ambiguity in the dedup code since the DDT does not store
2048 * object types.
2049 */
2050 if (os->os_encrypted && (wp & WP_NOFILL) == 0) {
2051 encrypt = B_TRUE;
2052
2053 if (DMU_OT_IS_ENCRYPTED(type)) {
2054 copies = MIN(copies, SPA_DVAS_PER_BP - 1);
2055 nopwrite = B_FALSE;
2056 } else {
2057 dedup = B_FALSE;
2058 }
2059
2060 if (level <= 0 &&
2061 (type == DMU_OT_DNODE || type == DMU_OT_OBJSET)) {
2062 compress = ZIO_COMPRESS_EMPTY;
2063 }
2064 }
2065
2066 zp->zp_compress = compress;
2067 zp->zp_complevel = complevel;
2068 zp->zp_checksum = checksum;
2069 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
2070 zp->zp_level = level;
2071 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
2072 zp->zp_dedup = dedup;
2073 zp->zp_dedup_verify = dedup && dedup_verify;
2074 zp->zp_nopwrite = nopwrite;
2075 zp->zp_encrypt = encrypt;
2076 zp->zp_byteorder = ZFS_HOST_BYTEORDER;
2077 bzero(zp->zp_salt, ZIO_DATA_SALT_LEN);
2078 bzero(zp->zp_iv, ZIO_DATA_IV_LEN);
2079 bzero(zp->zp_mac, ZIO_DATA_MAC_LEN);
2080 zp->zp_zpl_smallblk = DMU_OT_IS_FILE(zp->zp_type) ?
2081 os->os_zpl_special_smallblock : 0;
2082
2083 ASSERT3U(zp->zp_compress, !=, ZIO_COMPRESS_INHERIT);
2084 }
2085
2086 /*
2087 * This function is only called from zfs_holey_common() for zpl_llseek()
2088 * in order to determine the location of holes. In order to accurately
2089 * report holes all dirty data must be synced to disk. This causes extremely
2090 * poor performance when seeking for holes in a dirty file. As a compromise,
2091 * only provide hole data when the dnode is clean. When a dnode is dirty
2092 * report the dnode as having no holes which is always a safe thing to do.
2093 */
2094 int
2095 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
2096 {
2097 dnode_t *dn;
2098 int i, err;
2099 boolean_t clean = B_TRUE;
2100
2101 err = dnode_hold(os, object, FTAG, &dn);
2102 if (err)
2103 return (err);
2104
2105 /*
2106 * Check if dnode is dirty
2107 */
2108 for (i = 0; i < TXG_SIZE; i++) {
2109 if (multilist_link_active(&dn->dn_dirty_link[i])) {
2110 clean = B_FALSE;
2111 break;
2112 }
2113 }
2114
2115 /*
2116 * If compatibility option is on, sync any current changes before
2117 * we go trundling through the block pointers.
2118 */
2119 if (!clean && zfs_dmu_offset_next_sync) {
2120 clean = B_TRUE;
2121 dnode_rele(dn, FTAG);
2122 txg_wait_synced(dmu_objset_pool(os), 0);
2123 err = dnode_hold(os, object, FTAG, &dn);
2124 if (err)
2125 return (err);
2126 }
2127
2128 if (clean)
2129 err = dnode_next_offset(dn,
2130 (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
2131 else
2132 err = SET_ERROR(EBUSY);
2133
2134 dnode_rele(dn, FTAG);
2135
2136 return (err);
2137 }
2138
2139 void
2140 __dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2141 {
2142 dnode_phys_t *dnp = dn->dn_phys;
2143
2144 doi->doi_data_block_size = dn->dn_datablksz;
2145 doi->doi_metadata_block_size = dn->dn_indblkshift ?
2146 1ULL << dn->dn_indblkshift : 0;
2147 doi->doi_type = dn->dn_type;
2148 doi->doi_bonus_type = dn->dn_bonustype;
2149 doi->doi_bonus_size = dn->dn_bonuslen;
2150 doi->doi_dnodesize = dn->dn_num_slots << DNODE_SHIFT;
2151 doi->doi_indirection = dn->dn_nlevels;
2152 doi->doi_checksum = dn->dn_checksum;
2153 doi->doi_compress = dn->dn_compress;
2154 doi->doi_nblkptr = dn->dn_nblkptr;
2155 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
2156 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
2157 doi->doi_fill_count = 0;
2158 for (int i = 0; i < dnp->dn_nblkptr; i++)
2159 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
2160 }
2161
2162 void
2163 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2164 {
2165 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2166 mutex_enter(&dn->dn_mtx);
2167
2168 __dmu_object_info_from_dnode(dn, doi);
2169
2170 mutex_exit(&dn->dn_mtx);
2171 rw_exit(&dn->dn_struct_rwlock);
2172 }
2173
2174 /*
2175 * Get information on a DMU object.
2176 * If doi is NULL, just indicates whether the object exists.
2177 */
2178 int
2179 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
2180 {
2181 dnode_t *dn;
2182 int err = dnode_hold(os, object, FTAG, &dn);
2183
2184 if (err)
2185 return (err);
2186
2187 if (doi != NULL)
2188 dmu_object_info_from_dnode(dn, doi);
2189
2190 dnode_rele(dn, FTAG);
2191 return (0);
2192 }
2193
2194 /*
2195 * As above, but faster; can be used when you have a held dbuf in hand.
2196 */
2197 void
2198 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
2199 {
2200 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2201
2202 DB_DNODE_ENTER(db);
2203 dmu_object_info_from_dnode(DB_DNODE(db), doi);
2204 DB_DNODE_EXIT(db);
2205 }
2206
2207 /*
2208 * Faster still when you only care about the size.
2209 */
2210 void
2211 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
2212 u_longlong_t *nblk512)
2213 {
2214 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2215 dnode_t *dn;
2216
2217 DB_DNODE_ENTER(db);
2218 dn = DB_DNODE(db);
2219
2220 *blksize = dn->dn_datablksz;
2221 /* add in number of slots used for the dnode itself */
2222 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
2223 SPA_MINBLOCKSHIFT) + dn->dn_num_slots;
2224 DB_DNODE_EXIT(db);
2225 }
2226
2227 void
2228 dmu_object_dnsize_from_db(dmu_buf_t *db_fake, int *dnsize)
2229 {
2230 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2231 dnode_t *dn;
2232
2233 DB_DNODE_ENTER(db);
2234 dn = DB_DNODE(db);
2235 *dnsize = dn->dn_num_slots << DNODE_SHIFT;
2236 DB_DNODE_EXIT(db);
2237 }
2238
2239 void
2240 byteswap_uint64_array(void *vbuf, size_t size)
2241 {
2242 uint64_t *buf = vbuf;
2243 size_t count = size >> 3;
2244 int i;
2245
2246 ASSERT((size & 7) == 0);
2247
2248 for (i = 0; i < count; i++)
2249 buf[i] = BSWAP_64(buf[i]);
2250 }
2251
2252 void
2253 byteswap_uint32_array(void *vbuf, size_t size)
2254 {
2255 uint32_t *buf = vbuf;
2256 size_t count = size >> 2;
2257 int i;
2258
2259 ASSERT((size & 3) == 0);
2260
2261 for (i = 0; i < count; i++)
2262 buf[i] = BSWAP_32(buf[i]);
2263 }
2264
2265 void
2266 byteswap_uint16_array(void *vbuf, size_t size)
2267 {
2268 uint16_t *buf = vbuf;
2269 size_t count = size >> 1;
2270 int i;
2271
2272 ASSERT((size & 1) == 0);
2273
2274 for (i = 0; i < count; i++)
2275 buf[i] = BSWAP_16(buf[i]);
2276 }
2277
2278 /* ARGSUSED */
2279 void
2280 byteswap_uint8_array(void *vbuf, size_t size)
2281 {
2282 }
2283
2284 void
2285 dmu_init(void)
2286 {
2287 abd_init();
2288 zfs_dbgmsg_init();
2289 sa_cache_init();
2290 dmu_objset_init();
2291 dnode_init();
2292 zfetch_init();
2293 dmu_tx_init();
2294 l2arc_init();
2295 arc_init();
2296 dbuf_init();
2297 }
2298
2299 void
2300 dmu_fini(void)
2301 {
2302 arc_fini(); /* arc depends on l2arc, so arc must go first */
2303 l2arc_fini();
2304 dmu_tx_fini();
2305 zfetch_fini();
2306 dbuf_fini();
2307 dnode_fini();
2308 dmu_objset_fini();
2309 sa_cache_fini();
2310 zfs_dbgmsg_fini();
2311 abd_fini();
2312 }
2313
2314 EXPORT_SYMBOL(dmu_bonus_hold);
2315 EXPORT_SYMBOL(dmu_bonus_hold_by_dnode);
2316 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus);
2317 EXPORT_SYMBOL(dmu_buf_rele_array);
2318 EXPORT_SYMBOL(dmu_prefetch);
2319 EXPORT_SYMBOL(dmu_free_range);
2320 EXPORT_SYMBOL(dmu_free_long_range);
2321 EXPORT_SYMBOL(dmu_free_long_object);
2322 EXPORT_SYMBOL(dmu_read);
2323 EXPORT_SYMBOL(dmu_read_by_dnode);
2324 EXPORT_SYMBOL(dmu_write);
2325 EXPORT_SYMBOL(dmu_write_by_dnode);
2326 EXPORT_SYMBOL(dmu_prealloc);
2327 EXPORT_SYMBOL(dmu_object_info);
2328 EXPORT_SYMBOL(dmu_object_info_from_dnode);
2329 EXPORT_SYMBOL(dmu_object_info_from_db);
2330 EXPORT_SYMBOL(dmu_object_size_from_db);
2331 EXPORT_SYMBOL(dmu_object_dnsize_from_db);
2332 EXPORT_SYMBOL(dmu_object_set_nlevels);
2333 EXPORT_SYMBOL(dmu_object_set_blocksize);
2334 EXPORT_SYMBOL(dmu_object_set_maxblkid);
2335 EXPORT_SYMBOL(dmu_object_set_checksum);
2336 EXPORT_SYMBOL(dmu_object_set_compress);
2337 EXPORT_SYMBOL(dmu_offset_next);
2338 EXPORT_SYMBOL(dmu_write_policy);
2339 EXPORT_SYMBOL(dmu_sync);
2340 EXPORT_SYMBOL(dmu_request_arcbuf);
2341 EXPORT_SYMBOL(dmu_return_arcbuf);
2342 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dnode);
2343 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dbuf);
2344 EXPORT_SYMBOL(dmu_buf_hold);
2345 EXPORT_SYMBOL(dmu_ot);
2346
2347 /* BEGIN CSTYLED */
2348 ZFS_MODULE_PARAM(zfs, zfs_, nopwrite_enabled, INT, ZMOD_RW,
2349 "Enable NOP writes");
2350
2351 ZFS_MODULE_PARAM(zfs, zfs_, per_txg_dirty_frees_percent, ULONG, ZMOD_RW,
2352 "Percentage of dirtied blocks from frees in one TXG");
2353
2354 ZFS_MODULE_PARAM(zfs, zfs_, dmu_offset_next_sync, INT, ZMOD_RW,
2355 "Enable forcing txg sync to find holes");
2356
2357 ZFS_MODULE_PARAM(zfs, , dmu_prefetch_max, INT, ZMOD_RW,
2358 "Limit one prefetch call to this size");
2359 /* END CSTYLED */
OpenZFS on Linux and FreeBSD