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 */
25 /* Copyright (c) 2013 by Saso Kiselkov. All rights reserved. */
26 /* Copyright (c) 2013, Joyent, Inc. All rights reserved. */
27 /* Copyright 2016 Nexenta Systems, Inc. All rights reserved. */
28
29 #include <sys/dmu.h>
30 #include <sys/dmu_impl.h>
31 #include <sys/dmu_tx.h>
32 #include <sys/dbuf.h>
33 #include <sys/dnode.h>
34 #include <sys/zfs_context.h>
35 #include <sys/dmu_objset.h>
36 #include <sys/dmu_traverse.h>
37 #include <sys/dsl_dataset.h>
38 #include <sys/dsl_dir.h>
39 #include <sys/dsl_pool.h>
40 #include <sys/dsl_synctask.h>
41 #include <sys/dsl_prop.h>
42 #include <sys/dmu_zfetch.h>
43 #include <sys/zfs_ioctl.h>
44 #include <sys/zap.h>
45 #include <sys/zio_checksum.h>
46 #include <sys/zio_compress.h>
47 #include <sys/sa.h>
48 #include <sys/zfeature.h>
49 #include <sys/abd.h>
50 #ifdef _KERNEL
51 #include <sys/vmsystm.h>
52 #include <sys/zfs_znode.h>
53 #endif
54
55 /*
56 * Enable/disable nopwrite feature.
57 */
58 int zfs_nopwrite_enabled = 1;
59
60 /*
61 * Tunable to control percentage of dirtied blocks from frees in one TXG.
62 * After this threshold is crossed, additional dirty blocks from frees
63 * wait until the next TXG.
64 * A value of zero will disable this throttle.
65 */
66 uint32_t zfs_per_txg_dirty_frees_percent = 30;
67
68 /*
69 * This can be used for testing, to ensure that certain actions happen
70 * while in the middle of a remap (which might otherwise complete too
71 * quickly).
72 */
73 int zfs_object_remap_one_indirect_delay_ticks = 0;
74
75 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
76 { DMU_BSWAP_UINT8, TRUE, "unallocated" },
77 { DMU_BSWAP_ZAP, TRUE, "object directory" },
78 { DMU_BSWAP_UINT64, TRUE, "object array" },
79 { DMU_BSWAP_UINT8, TRUE, "packed nvlist" },
80 { DMU_BSWAP_UINT64, TRUE, "packed nvlist size" },
81 { DMU_BSWAP_UINT64, TRUE, "bpobj" },
82 { DMU_BSWAP_UINT64, TRUE, "bpobj header" },
83 { DMU_BSWAP_UINT64, TRUE, "SPA space map header" },
84 { DMU_BSWAP_UINT64, TRUE, "SPA space map" },
85 { DMU_BSWAP_UINT64, TRUE, "ZIL intent log" },
86 { DMU_BSWAP_DNODE, TRUE, "DMU dnode" },
87 { DMU_BSWAP_OBJSET, TRUE, "DMU objset" },
88 { DMU_BSWAP_UINT64, TRUE, "DSL directory" },
89 { DMU_BSWAP_ZAP, TRUE, "DSL directory child map"},
90 { DMU_BSWAP_ZAP, TRUE, "DSL dataset snap map" },
91 { DMU_BSWAP_ZAP, TRUE, "DSL props" },
92 { DMU_BSWAP_UINT64, TRUE, "DSL dataset" },
93 { DMU_BSWAP_ZNODE, TRUE, "ZFS znode" },
94 { DMU_BSWAP_OLDACL, TRUE, "ZFS V0 ACL" },
95 { DMU_BSWAP_UINT8, FALSE, "ZFS plain file" },
96 { DMU_BSWAP_ZAP, TRUE, "ZFS directory" },
97 { DMU_BSWAP_ZAP, TRUE, "ZFS master node" },
98 { DMU_BSWAP_ZAP, TRUE, "ZFS delete queue" },
99 { DMU_BSWAP_UINT8, FALSE, "zvol object" },
100 { DMU_BSWAP_ZAP, TRUE, "zvol prop" },
101 { DMU_BSWAP_UINT8, FALSE, "other uint8[]" },
102 { DMU_BSWAP_UINT64, FALSE, "other uint64[]" },
103 { DMU_BSWAP_ZAP, TRUE, "other ZAP" },
104 { DMU_BSWAP_ZAP, TRUE, "persistent error log" },
105 { DMU_BSWAP_UINT8, TRUE, "SPA history" },
106 { DMU_BSWAP_UINT64, TRUE, "SPA history offsets" },
107 { DMU_BSWAP_ZAP, TRUE, "Pool properties" },
108 { DMU_BSWAP_ZAP, TRUE, "DSL permissions" },
109 { DMU_BSWAP_ACL, TRUE, "ZFS ACL" },
110 { DMU_BSWAP_UINT8, TRUE, "ZFS SYSACL" },
111 { DMU_BSWAP_UINT8, TRUE, "FUID table" },
112 { DMU_BSWAP_UINT64, TRUE, "FUID table size" },
113 { DMU_BSWAP_ZAP, TRUE, "DSL dataset next clones"},
114 { DMU_BSWAP_ZAP, TRUE, "scan work queue" },
115 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group used" },
116 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group quota" },
117 { DMU_BSWAP_ZAP, TRUE, "snapshot refcount tags"},
118 { DMU_BSWAP_ZAP, TRUE, "DDT ZAP algorithm" },
119 { DMU_BSWAP_ZAP, TRUE, "DDT statistics" },
120 { DMU_BSWAP_UINT8, TRUE, "System attributes" },
121 { DMU_BSWAP_ZAP, TRUE, "SA master node" },
122 { DMU_BSWAP_ZAP, TRUE, "SA attr registration" },
123 { DMU_BSWAP_ZAP, TRUE, "SA attr layouts" },
124 { DMU_BSWAP_ZAP, TRUE, "scan translations" },
125 { DMU_BSWAP_UINT8, FALSE, "deduplicated block" },
126 { DMU_BSWAP_ZAP, TRUE, "DSL deadlist map" },
127 { DMU_BSWAP_UINT64, TRUE, "DSL deadlist map hdr" },
128 { DMU_BSWAP_ZAP, TRUE, "DSL dir clones" },
129 { DMU_BSWAP_UINT64, TRUE, "bpobj subobj" }
130 };
131
132 const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
133 { byteswap_uint8_array, "uint8" },
134 { byteswap_uint16_array, "uint16" },
135 { byteswap_uint32_array, "uint32" },
136 { byteswap_uint64_array, "uint64" },
137 { zap_byteswap, "zap" },
138 { dnode_buf_byteswap, "dnode" },
139 { dmu_objset_byteswap, "objset" },
140 { zfs_znode_byteswap, "znode" },
141 { zfs_oldacl_byteswap, "oldacl" },
142 { zfs_acl_byteswap, "acl" }
143 };
144
145 int
146 dmu_buf_hold_noread_by_dnode(dnode_t *dn, uint64_t offset,
147 void *tag, dmu_buf_t **dbp)
148 {
149 uint64_t blkid;
150 dmu_buf_impl_t *db;
151
152 blkid = dbuf_whichblock(dn, 0, offset);
153 rw_enter(&dn->dn_struct_rwlock, RW_READER);
154 db = dbuf_hold(dn, blkid, tag);
155 rw_exit(&dn->dn_struct_rwlock);
156
157 if (db == NULL) {
158 *dbp = NULL;
159 return (SET_ERROR(EIO));
160 }
161
162 *dbp = &db->db;
163 return (0);
164 }
165 int
166 dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset,
167 void *tag, dmu_buf_t **dbp)
168 {
169 dnode_t *dn;
170 uint64_t blkid;
171 dmu_buf_impl_t *db;
172 int err;
173
174 err = dnode_hold(os, object, FTAG, &dn);
175 if (err)
176 return (err);
177 blkid = dbuf_whichblock(dn, 0, offset);
178 rw_enter(&dn->dn_struct_rwlock, RW_READER);
179 db = dbuf_hold(dn, blkid, tag);
180 rw_exit(&dn->dn_struct_rwlock);
181 dnode_rele(dn, FTAG);
182
183 if (db == NULL) {
184 *dbp = NULL;
185 return (SET_ERROR(EIO));
186 }
187
188 *dbp = &db->db;
189 return (err);
190 }
191
192 int
193 dmu_buf_hold_by_dnode(dnode_t *dn, uint64_t offset,
194 void *tag, dmu_buf_t **dbp, int flags)
195 {
196 int err;
197 int db_flags = DB_RF_CANFAIL;
198
199 if (flags & DMU_READ_NO_PREFETCH)
200 db_flags |= DB_RF_NOPREFETCH;
201
202 err = dmu_buf_hold_noread_by_dnode(dn, offset, tag, dbp);
203 if (err == 0) {
204 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
205 err = dbuf_read(db, NULL, db_flags);
206 if (err != 0) {
207 dbuf_rele(db, tag);
208 *dbp = NULL;
209 }
210 }
211
212 return (err);
213 }
214
215 int
216 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
217 void *tag, dmu_buf_t **dbp, int flags)
218 {
219 int err;
220 int db_flags = DB_RF_CANFAIL;
221
222 if (flags & DMU_READ_NO_PREFETCH)
223 db_flags |= DB_RF_NOPREFETCH;
224
225 err = dmu_buf_hold_noread(os, object, offset, tag, dbp);
226 if (err == 0) {
227 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
228 err = dbuf_read(db, NULL, db_flags);
229 if (err != 0) {
230 dbuf_rele(db, tag);
231 *dbp = NULL;
232 }
233 }
234
235 return (err);
236 }
237
238 int
239 dmu_bonus_max(void)
240 {
241 return (DN_MAX_BONUSLEN);
242 }
243
244 int
245 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
246 {
247 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
248 dnode_t *dn;
249 int error;
250
251 DB_DNODE_ENTER(db);
252 dn = DB_DNODE(db);
253
254 if (dn->dn_bonus != db) {
255 error = SET_ERROR(EINVAL);
256 } else if (newsize < 0 || newsize > db_fake->db_size) {
257 error = SET_ERROR(EINVAL);
258 } else {
259 dnode_setbonuslen(dn, newsize, tx);
260 error = 0;
261 }
262
263 DB_DNODE_EXIT(db);
264 return (error);
265 }
266
267 int
268 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
269 {
270 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
271 dnode_t *dn;
272 int error;
273
274 DB_DNODE_ENTER(db);
275 dn = DB_DNODE(db);
276
277 if (!DMU_OT_IS_VALID(type)) {
278 error = SET_ERROR(EINVAL);
279 } else if (dn->dn_bonus != db) {
280 error = SET_ERROR(EINVAL);
281 } else {
282 dnode_setbonus_type(dn, type, tx);
283 error = 0;
284 }
285
286 DB_DNODE_EXIT(db);
287 return (error);
288 }
289
290 dmu_object_type_t
291 dmu_get_bonustype(dmu_buf_t *db_fake)
292 {
293 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
294 dnode_t *dn;
295 dmu_object_type_t type;
296
297 DB_DNODE_ENTER(db);
298 dn = DB_DNODE(db);
299 type = dn->dn_bonustype;
300 DB_DNODE_EXIT(db);
301
302 return (type);
303 }
304
305 int
306 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
307 {
308 dnode_t *dn;
309 int error;
310
311 error = dnode_hold(os, object, FTAG, &dn);
312 dbuf_rm_spill(dn, tx);
313 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
314 dnode_rm_spill(dn, tx);
315 rw_exit(&dn->dn_struct_rwlock);
316 dnode_rele(dn, FTAG);
317 return (error);
318 }
319
320 /*
321 * returns ENOENT, EIO, or 0.
322 */
323 int
324 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
325 {
326 dnode_t *dn;
327 dmu_buf_impl_t *db;
328 int error;
329
330 error = dnode_hold(os, object, FTAG, &dn);
331 if (error)
332 return (error);
333
334 rw_enter(&dn->dn_struct_rwlock, RW_READER);
335 if (dn->dn_bonus == NULL) {
336 rw_exit(&dn->dn_struct_rwlock);
337 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
338 if (dn->dn_bonus == NULL)
339 dbuf_create_bonus(dn);
340 }
341 db = dn->dn_bonus;
342
343 /* as long as the bonus buf is held, the dnode will be held */
344 if (refcount_add(&db->db_holds, tag) == 1) {
345 VERIFY(dnode_add_ref(dn, db));
346 atomic_inc_32(&dn->dn_dbufs_count);
347 }
348
349 /*
350 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
351 * hold and incrementing the dbuf count to ensure that dnode_move() sees
352 * a dnode hold for every dbuf.
353 */
354 rw_exit(&dn->dn_struct_rwlock);
355
356 dnode_rele(dn, FTAG);
357
358 VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH));
359
360 *dbp = &db->db;
361 return (0);
362 }
363
364 /*
365 * returns ENOENT, EIO, or 0.
366 *
367 * This interface will allocate a blank spill dbuf when a spill blk
368 * doesn't already exist on the dnode.
369 *
370 * if you only want to find an already existing spill db, then
371 * dmu_spill_hold_existing() should be used.
372 */
373 int
374 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
375 {
376 dmu_buf_impl_t *db = NULL;
377 int err;
378
379 if ((flags & DB_RF_HAVESTRUCT) == 0)
380 rw_enter(&dn->dn_struct_rwlock, RW_READER);
381
382 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
383
384 if ((flags & DB_RF_HAVESTRUCT) == 0)
385 rw_exit(&dn->dn_struct_rwlock);
386
387 ASSERT(db != NULL);
388 err = dbuf_read(db, NULL, flags);
389 if (err == 0)
390 *dbp = &db->db;
391 else
392 dbuf_rele(db, tag);
393 return (err);
394 }
395
396 int
397 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
398 {
399 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
400 dnode_t *dn;
401 int err;
402
403 DB_DNODE_ENTER(db);
404 dn = DB_DNODE(db);
405
406 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
407 err = SET_ERROR(EINVAL);
408 } else {
409 rw_enter(&dn->dn_struct_rwlock, RW_READER);
410
411 if (!dn->dn_have_spill) {
412 err = SET_ERROR(ENOENT);
413 } else {
414 err = dmu_spill_hold_by_dnode(dn,
415 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
416 }
417
418 rw_exit(&dn->dn_struct_rwlock);
419 }
420
421 DB_DNODE_EXIT(db);
422 return (err);
423 }
424
425 int
426 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
427 {
428 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
429 dnode_t *dn;
430 int err;
431
432 DB_DNODE_ENTER(db);
433 dn = DB_DNODE(db);
434 err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp);
435 DB_DNODE_EXIT(db);
436
437 return (err);
438 }
439
440 /*
441 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
442 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
443 * and can induce severe lock contention when writing to several files
444 * whose dnodes are in the same block.
445 */
446 static int
447 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
448 boolean_t read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
449 {
450 dmu_buf_t **dbp;
451 uint64_t blkid, nblks, i;
452 uint32_t dbuf_flags;
453 int err;
454 zio_t *zio;
455
456 ASSERT(length <= DMU_MAX_ACCESS);
457
458 /*
459 * Note: We directly notify the prefetch code of this read, so that
460 * we can tell it about the multi-block read. dbuf_read() only knows
461 * about the one block it is accessing.
462 */
463 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT |
464 DB_RF_NOPREFETCH;
465
466 rw_enter(&dn->dn_struct_rwlock, RW_READER);
467 if (dn->dn_datablkshift) {
468 int blkshift = dn->dn_datablkshift;
469 nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) -
470 P2ALIGN(offset, 1ULL << blkshift)) >> blkshift;
471 } else {
472 if (offset + length > dn->dn_datablksz) {
473 zfs_panic_recover("zfs: accessing past end of object "
474 "%llx/%llx (size=%u access=%llu+%llu)",
475 (longlong_t)dn->dn_objset->
476 os_dsl_dataset->ds_object,
477 (longlong_t)dn->dn_object, dn->dn_datablksz,
478 (longlong_t)offset, (longlong_t)length);
479 rw_exit(&dn->dn_struct_rwlock);
480 return (SET_ERROR(EIO));
481 }
482 nblks = 1;
483 }
484 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
485
486 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
487 blkid = dbuf_whichblock(dn, 0, offset);
488 for (i = 0; i < nblks; i++) {
489 dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag);
490 if (db == NULL) {
491 rw_exit(&dn->dn_struct_rwlock);
492 dmu_buf_rele_array(dbp, nblks, tag);
493 zio_nowait(zio);
494 return (SET_ERROR(EIO));
495 }
496
497 /* initiate async i/o */
498 if (read)
499 (void) dbuf_read(db, zio, dbuf_flags);
500 dbp[i] = &db->db;
501 }
502
503 if ((flags & DMU_READ_NO_PREFETCH) == 0 &&
504 DNODE_META_IS_CACHEABLE(dn) && length <= zfetch_array_rd_sz) {
505 dmu_zfetch(&dn->dn_zfetch, blkid, nblks,
506 read && DNODE_IS_CACHEABLE(dn));
507 }
508 rw_exit(&dn->dn_struct_rwlock);
509
510 /* wait for async i/o */
511 err = zio_wait(zio);
512 if (err) {
513 dmu_buf_rele_array(dbp, nblks, tag);
514 return (err);
515 }
516
517 /* wait for other io to complete */
518 if (read) {
519 for (i = 0; i < nblks; i++) {
520 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
521 mutex_enter(&db->db_mtx);
522 while (db->db_state == DB_READ ||
523 db->db_state == DB_FILL)
524 cv_wait(&db->db_changed, &db->db_mtx);
525 if (db->db_state == DB_UNCACHED)
526 err = SET_ERROR(EIO);
527 mutex_exit(&db->db_mtx);
528 if (err) {
529 dmu_buf_rele_array(dbp, nblks, tag);
530 return (err);
531 }
532 }
533 }
534
535 *numbufsp = nblks;
536 *dbpp = dbp;
537 return (0);
538 }
539
540 static int
541 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
542 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
543 {
544 dnode_t *dn;
545 int err;
546
547 err = dnode_hold(os, object, FTAG, &dn);
548 if (err)
549 return (err);
550
551 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
552 numbufsp, dbpp, DMU_READ_PREFETCH);
553
554 dnode_rele(dn, FTAG);
555
556 return (err);
557 }
558
559 int
560 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
561 uint64_t length, boolean_t read, void *tag, int *numbufsp,
562 dmu_buf_t ***dbpp)
563 {
564 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
565 dnode_t *dn;
566 int err;
567
568 DB_DNODE_ENTER(db);
569 dn = DB_DNODE(db);
570 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
571 numbufsp, dbpp, DMU_READ_PREFETCH);
572 DB_DNODE_EXIT(db);
573
574 return (err);
575 }
576
577 void
578 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
579 {
580 int i;
581 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
582
583 if (numbufs == 0)
584 return;
585
586 for (i = 0; i < numbufs; i++) {
587 if (dbp[i])
588 dbuf_rele(dbp[i], tag);
589 }
590
591 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
592 }
593
594 /*
595 * Issue prefetch i/os for the given blocks. If level is greater than 0, the
596 * indirect blocks prefeteched will be those that point to the blocks containing
597 * the data starting at offset, and continuing to offset + len.
598 *
599 * Note that if the indirect blocks above the blocks being prefetched are not in
600 * cache, they will be asychronously read in.
601 */
602 void
603 dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset,
604 uint64_t len, zio_priority_t pri)
605 {
606 dnode_t *dn;
607 uint64_t blkid;
608 int nblks, err;
609
610 if (len == 0) { /* they're interested in the bonus buffer */
611 dn = DMU_META_DNODE(os);
612
613 if (object == 0 || object >= DN_MAX_OBJECT)
614 return;
615
616 rw_enter(&dn->dn_struct_rwlock, RW_READER);
617 blkid = dbuf_whichblock(dn, level,
618 object * sizeof (dnode_phys_t));
619 dbuf_prefetch(dn, level, blkid, pri, 0);
620 rw_exit(&dn->dn_struct_rwlock);
621 return;
622 }
623
624 /*
625 * XXX - Note, if the dnode for the requested object is not
626 * already cached, we will do a *synchronous* read in the
627 * dnode_hold() call. The same is true for any indirects.
628 */
629 err = dnode_hold(os, object, FTAG, &dn);
630 if (err != 0)
631 return;
632
633 rw_enter(&dn->dn_struct_rwlock, RW_READER);
634 /*
635 * offset + len - 1 is the last byte we want to prefetch for, and offset
636 * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the
637 * last block we want to prefetch, and dbuf_whichblock(dn, level,
638 * offset) is the first. Then the number we need to prefetch is the
639 * last - first + 1.
640 */
641 if (level > 0 || dn->dn_datablkshift != 0) {
642 nblks = dbuf_whichblock(dn, level, offset + len - 1) -
643 dbuf_whichblock(dn, level, offset) + 1;
644 } else {
645 nblks = (offset < dn->dn_datablksz);
646 }
647
648 if (nblks != 0) {
649 blkid = dbuf_whichblock(dn, level, offset);
650 for (int i = 0; i < nblks; i++)
651 dbuf_prefetch(dn, level, blkid + i, pri, 0);
652 }
653
654 rw_exit(&dn->dn_struct_rwlock);
655
656 dnode_rele(dn, FTAG);
657 }
658
659 /*
660 * Get the next "chunk" of file data to free. We traverse the file from
661 * the end so that the file gets shorter over time (if we crashes in the
662 * middle, this will leave us in a better state). We find allocated file
663 * data by simply searching the allocated level 1 indirects.
664 *
665 * On input, *start should be the first offset that does not need to be
666 * freed (e.g. "offset + length"). On return, *start will be the first
667 * offset that should be freed.
668 */
669 static int
670 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum)
671 {
672 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
673 /* bytes of data covered by a level-1 indirect block */
674 uint64_t iblkrange =
675 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
676
677 ASSERT3U(minimum, <=, *start);
678
679 if (*start - minimum <= iblkrange * maxblks) {
680 *start = minimum;
681 return (0);
682 }
683 ASSERT(ISP2(iblkrange));
684
685 for (uint64_t blks = 0; *start > minimum && blks < maxblks; blks++) {
686 int err;
687
688 /*
689 * dnode_next_offset(BACKWARDS) will find an allocated L1
690 * indirect block at or before the input offset. We must
691 * decrement *start so that it is at the end of the region
692 * to search.
693 */
694 (*start)--;
695 err = dnode_next_offset(dn,
696 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
697
698 /* if there are no indirect blocks before start, we are done */
699 if (err == ESRCH) {
700 *start = minimum;
701 break;
702 } else if (err != 0) {
703 return (err);
704 }
705
706 /* set start to the beginning of this L1 indirect */
707 *start = P2ALIGN(*start, iblkrange);
708 }
709 if (*start < minimum)
710 *start = minimum;
711 return (0);
712 }
713
714 /*
715 * If this objset is of type OST_ZFS return true if vfs's unmounted flag is set,
716 * otherwise return false.
717 * Used below in dmu_free_long_range_impl() to enable abort when unmounting
718 */
719 /*ARGSUSED*/
720 static boolean_t
721 dmu_objset_zfs_unmounting(objset_t *os)
722 {
723 #ifdef _KERNEL
724 if (dmu_objset_type(os) == DMU_OST_ZFS)
725 return (zfs_get_vfs_flag_unmounted(os));
726 #endif
727 return (B_FALSE);
728 }
729
730 static int
731 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
732 uint64_t length)
733 {
734 uint64_t object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
735 int err;
736 uint64_t dirty_frees_threshold;
737 dsl_pool_t *dp = dmu_objset_pool(os);
738
739 if (offset >= object_size)
740 return (0);
741
742 if (zfs_per_txg_dirty_frees_percent <= 100)
743 dirty_frees_threshold =
744 zfs_per_txg_dirty_frees_percent * zfs_dirty_data_max / 100;
745 else
746 dirty_frees_threshold = zfs_dirty_data_max / 4;
747
748 if (length == DMU_OBJECT_END || offset + length > object_size)
749 length = object_size - offset;
750
751 while (length != 0) {
752 uint64_t chunk_end, chunk_begin, chunk_len;
753 uint64_t long_free_dirty_all_txgs = 0;
754 dmu_tx_t *tx;
755
756 if (dmu_objset_zfs_unmounting(dn->dn_objset))
757 return (SET_ERROR(EINTR));
758
759 chunk_end = chunk_begin = offset + length;
760
761 /* move chunk_begin backwards to the beginning of this chunk */
762 err = get_next_chunk(dn, &chunk_begin, offset);
763 if (err)
764 return (err);
765 ASSERT3U(chunk_begin, >=, offset);
766 ASSERT3U(chunk_begin, <=, chunk_end);
767
768 chunk_len = chunk_end - chunk_begin;
769
770 mutex_enter(&dp->dp_lock);
771 for (int t = 0; t < TXG_SIZE; t++) {
772 long_free_dirty_all_txgs +=
773 dp->dp_long_free_dirty_pertxg[t];
774 }
775 mutex_exit(&dp->dp_lock);
776
777 /*
778 * To avoid filling up a TXG with just frees wait for
779 * the next TXG to open before freeing more chunks if
780 * we have reached the threshold of frees
781 */
782 if (dirty_frees_threshold != 0 &&
783 long_free_dirty_all_txgs >= dirty_frees_threshold) {
784 txg_wait_open(dp, 0);
785 continue;
786 }
787
788 tx = dmu_tx_create(os);
789 dmu_tx_hold_free(tx, dn->dn_object, chunk_begin, chunk_len);
790
791 /*
792 * Mark this transaction as typically resulting in a net
793 * reduction in space used.
794 */
795 dmu_tx_mark_netfree(tx);
796 err = dmu_tx_assign(tx, TXG_WAIT);
797 if (err) {
798 dmu_tx_abort(tx);
799 return (err);
800 }
801
802 mutex_enter(&dp->dp_lock);
803 dp->dp_long_free_dirty_pertxg[dmu_tx_get_txg(tx) & TXG_MASK] +=
804 chunk_len;
805 mutex_exit(&dp->dp_lock);
806 DTRACE_PROBE3(free__long__range,
807 uint64_t, long_free_dirty_all_txgs, uint64_t, chunk_len,
808 uint64_t, dmu_tx_get_txg(tx));
809 dnode_free_range(dn, chunk_begin, chunk_len, tx);
810 dmu_tx_commit(tx);
811
812 length -= chunk_len;
813 }
814 return (0);
815 }
816
817 int
818 dmu_free_long_range(objset_t *os, uint64_t object,
819 uint64_t offset, uint64_t length)
820 {
821 dnode_t *dn;
822 int err;
823
824 err = dnode_hold(os, object, FTAG, &dn);
825 if (err != 0)
826 return (err);
827 err = dmu_free_long_range_impl(os, dn, offset, length);
828
829 /*
830 * It is important to zero out the maxblkid when freeing the entire
831 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
832 * will take the fast path, and (b) dnode_reallocate() can verify
833 * that the entire file has been freed.
834 */
835 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
836 dn->dn_maxblkid = 0;
837
838 dnode_rele(dn, FTAG);
839 return (err);
840 }
841
842 int
843 dmu_free_long_object(objset_t *os, uint64_t object)
844 {
845 dmu_tx_t *tx;
846 int err;
847
848 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
849 if (err != 0)
850 return (err);
851
852 tx = dmu_tx_create(os);
853 dmu_tx_hold_bonus(tx, object);
854 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
855 dmu_tx_mark_netfree(tx);
856 err = dmu_tx_assign(tx, TXG_WAIT);
857 if (err == 0) {
858 err = dmu_object_free(os, object, tx);
859 dmu_tx_commit(tx);
860 } else {
861 dmu_tx_abort(tx);
862 }
863
864 return (err);
865 }
866
867 int
868 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
869 uint64_t size, dmu_tx_t *tx)
870 {
871 dnode_t *dn;
872 int err = dnode_hold(os, object, FTAG, &dn);
873 if (err)
874 return (err);
875 ASSERT(offset < UINT64_MAX);
876 ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
877 dnode_free_range(dn, offset, size, tx);
878 dnode_rele(dn, FTAG);
879 return (0);
880 }
881
882 static int
883 dmu_read_impl(dnode_t *dn, uint64_t offset, uint64_t size,
884 void *buf, uint32_t flags)
885 {
886 dmu_buf_t **dbp;
887 int numbufs, err = 0;
888
889 /*
890 * Deal with odd block sizes, where there can't be data past the first
891 * block. If we ever do the tail block optimization, we will need to
892 * handle that here as well.
893 */
894 if (dn->dn_maxblkid == 0) {
895 int newsz = offset > dn->dn_datablksz ? 0 :
896 MIN(size, dn->dn_datablksz - offset);
897 bzero((char *)buf + newsz, size - newsz);
898 size = newsz;
899 }
900
901 while (size > 0) {
902 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
903 int i;
904
905 /*
906 * NB: we could do this block-at-a-time, but it's nice
907 * to be reading in parallel.
908 */
909 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
910 TRUE, FTAG, &numbufs, &dbp, flags);
911 if (err)
912 break;
913
914 for (i = 0; i < numbufs; i++) {
915 int tocpy;
916 int bufoff;
917 dmu_buf_t *db = dbp[i];
918
919 ASSERT(size > 0);
920
921 bufoff = offset - db->db_offset;
922 tocpy = (int)MIN(db->db_size - bufoff, size);
923
924 bcopy((char *)db->db_data + bufoff, buf, tocpy);
925
926 offset += tocpy;
927 size -= tocpy;
928 buf = (char *)buf + tocpy;
929 }
930 dmu_buf_rele_array(dbp, numbufs, FTAG);
931 }
932 return (err);
933 }
934
935 int
936 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
937 void *buf, uint32_t flags)
938 {
939 dnode_t *dn;
940 int err;
941
942 err = dnode_hold(os, object, FTAG, &dn);
943 if (err != 0)
944 return (err);
945
946 err = dmu_read_impl(dn, offset, size, buf, flags);
947 dnode_rele(dn, FTAG);
948 return (err);
949 }
950
951 int
952 dmu_read_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, void *buf,
953 uint32_t flags)
954 {
955 return (dmu_read_impl(dn, offset, size, buf, flags));
956 }
957
958 static void
959 dmu_write_impl(dmu_buf_t **dbp, int numbufs, uint64_t offset, uint64_t size,
960 const void *buf, dmu_tx_t *tx)
961 {
962 int i;
963
964 for (i = 0; i < numbufs; i++) {
965 int tocpy;
966 int bufoff;
967 dmu_buf_t *db = dbp[i];
968
969 ASSERT(size > 0);
970
971 bufoff = offset - db->db_offset;
972 tocpy = (int)MIN(db->db_size - bufoff, size);
973
974 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
975
976 if (tocpy == db->db_size)
977 dmu_buf_will_fill(db, tx);
978 else
979 dmu_buf_will_dirty(db, tx);
980
981 bcopy(buf, (char *)db->db_data + bufoff, tocpy);
982
983 if (tocpy == db->db_size)
984 dmu_buf_fill_done(db, tx);
985
986 offset += tocpy;
987 size -= tocpy;
988 buf = (char *)buf + tocpy;
989 }
990 }
991
992 void
993 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
994 const void *buf, dmu_tx_t *tx)
995 {
996 dmu_buf_t **dbp;
997 int numbufs;
998
999 if (size == 0)
1000 return;
1001
1002 VERIFY0(dmu_buf_hold_array(os, object, offset, size,
1003 FALSE, FTAG, &numbufs, &dbp));
1004 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1005 dmu_buf_rele_array(dbp, numbufs, FTAG);
1006 }
1007
1008 void
1009 dmu_write_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size,
1010 const void *buf, dmu_tx_t *tx)
1011 {
1012 dmu_buf_t **dbp;
1013 int numbufs;
1014
1015 if (size == 0)
1016 return;
1017
1018 VERIFY0(dmu_buf_hold_array_by_dnode(dn, offset, size,
1019 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH));
1020 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1021 dmu_buf_rele_array(dbp, numbufs, FTAG);
1022 }
1023
1024 static int
1025 dmu_object_remap_one_indirect(objset_t *os, dnode_t *dn,
1026 uint64_t last_removal_txg, uint64_t offset)
1027 {
1028 uint64_t l1blkid = dbuf_whichblock(dn, 1, offset);
1029 int err = 0;
1030
1031 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1032 dmu_buf_impl_t *dbuf = dbuf_hold_level(dn, 1, l1blkid, FTAG);
1033 ASSERT3P(dbuf, !=, NULL);
1034
1035 /*
1036 * If the block hasn't been written yet, this default will ensure
1037 * we don't try to remap it.
1038 */
1039 uint64_t birth = UINT64_MAX;
1040 ASSERT3U(last_removal_txg, !=, UINT64_MAX);
1041 if (dbuf->db_blkptr != NULL)
1042 birth = dbuf->db_blkptr->blk_birth;
1043 rw_exit(&dn->dn_struct_rwlock);
1044
1045 /*
1046 * If this L1 was already written after the last removal, then we've
1047 * already tried to remap it.
1048 */
1049 if (birth <= last_removal_txg &&
1050 dbuf_read(dbuf, NULL, DB_RF_MUST_SUCCEED) == 0 &&
1051 dbuf_can_remap(dbuf)) {
1052 dmu_tx_t *tx = dmu_tx_create(os);
1053 dmu_tx_hold_remap_l1indirect(tx, dn->dn_object);
1054 err = dmu_tx_assign(tx, TXG_WAIT);
1055 if (err == 0) {
1056 (void) dbuf_dirty(dbuf, tx);
1057 dmu_tx_commit(tx);
1058 } else {
1059 dmu_tx_abort(tx);
1060 }
1061 }
1062
1063 dbuf_rele(dbuf, FTAG);
1064
1065 delay(zfs_object_remap_one_indirect_delay_ticks);
1066
1067 return (err);
1068 }
1069
1070 /*
1071 * Remap all blockpointers in the object, if possible, so that they reference
1072 * only concrete vdevs.
1073 *
1074 * To do this, iterate over the L0 blockpointers and remap any that reference
1075 * an indirect vdev. Note that we only examine L0 blockpointers; since we
1076 * cannot guarantee that we can remap all blockpointer anyways (due to split
1077 * blocks), we do not want to make the code unnecessarily complicated to
1078 * catch the unlikely case that there is an L1 block on an indirect vdev that
1079 * contains no indirect blockpointers.
1080 */
1081 int
1082 dmu_object_remap_indirects(objset_t *os, uint64_t object,
1083 uint64_t last_removal_txg)
1084 {
1085 uint64_t offset, l1span;
1086 int err;
1087 dnode_t *dn;
1088
1089 err = dnode_hold(os, object, FTAG, &dn);
1090 if (err != 0) {
1091 return (err);
1092 }
1093
1094 if (dn->dn_nlevels <= 1) {
1095 if (issig(JUSTLOOKING) && issig(FORREAL)) {
1096 err = SET_ERROR(EINTR);
1097 }
1098
1099 /*
1100 * If the dnode has no indirect blocks, we cannot dirty them.
1101 * We still want to remap the blkptr(s) in the dnode if
1102 * appropriate, so mark it as dirty.
1103 */
1104 if (err == 0 && dnode_needs_remap(dn)) {
1105 dmu_tx_t *tx = dmu_tx_create(os);
1106 dmu_tx_hold_bonus(tx, dn->dn_object);
1107 if ((err = dmu_tx_assign(tx, TXG_WAIT)) == 0) {
1108 dnode_setdirty(dn, tx);
1109 dmu_tx_commit(tx);
1110 } else {
1111 dmu_tx_abort(tx);
1112 }
1113 }
1114
1115 dnode_rele(dn, FTAG);
1116 return (err);
1117 }
1118
1119 offset = 0;
1120 l1span = 1ULL << (dn->dn_indblkshift - SPA_BLKPTRSHIFT +
1121 dn->dn_datablkshift);
1122 /*
1123 * Find the next L1 indirect that is not a hole.
1124 */
1125 while (dnode_next_offset(dn, 0, &offset, 2, 1, 0) == 0) {
1126 if (issig(JUSTLOOKING) && issig(FORREAL)) {
1127 err = SET_ERROR(EINTR);
1128 break;
1129 }
1130 if ((err = dmu_object_remap_one_indirect(os, dn,
1131 last_removal_txg, offset)) != 0) {
1132 break;
1133 }
1134 offset += l1span;
1135 }
1136
1137 dnode_rele(dn, FTAG);
1138 return (err);
1139 }
1140
1141 void
1142 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1143 dmu_tx_t *tx)
1144 {
1145 dmu_buf_t **dbp;
1146 int numbufs, i;
1147
1148 if (size == 0)
1149 return;
1150
1151 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
1152 FALSE, FTAG, &numbufs, &dbp));
1153
1154 for (i = 0; i < numbufs; i++) {
1155 dmu_buf_t *db = dbp[i];
1156
1157 dmu_buf_will_not_fill(db, tx);
1158 }
1159 dmu_buf_rele_array(dbp, numbufs, FTAG);
1160 }
1161
1162 void
1163 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
1164 void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
1165 int compressed_size, int byteorder, dmu_tx_t *tx)
1166 {
1167 dmu_buf_t *db;
1168
1169 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
1170 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
1171 VERIFY0(dmu_buf_hold_noread(os, object, offset,
1172 FTAG, &db));
1173
1174 dmu_buf_write_embedded(db,
1175 data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
1176 uncompressed_size, compressed_size, byteorder, tx);
1177
1178 dmu_buf_rele(db, FTAG);
1179 }
1180
1181 /*
1182 * DMU support for xuio
1183 */
1184 kstat_t *xuio_ksp = NULL;
1185
1186 int
1187 dmu_xuio_init(xuio_t *xuio, int nblk)
1188 {
1189 dmu_xuio_t *priv;
1190 uio_t *uio = &xuio->xu_uio;
1191
1192 uio->uio_iovcnt = nblk;
1193 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
1194
1195 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
1196 priv->cnt = nblk;
1197 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
1198 priv->iovp = uio->uio_iov;
1199 XUIO_XUZC_PRIV(xuio) = priv;
1200
1201 if (XUIO_XUZC_RW(xuio) == UIO_READ)
1202 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
1203 else
1204 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
1205
1206 return (0);
1207 }
1208
1209 void
1210 dmu_xuio_fini(xuio_t *xuio)
1211 {
1212 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1213 int nblk = priv->cnt;
1214
1215 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
1216 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
1217 kmem_free(priv, sizeof (dmu_xuio_t));
1218
1219 if (XUIO_XUZC_RW(xuio) == UIO_READ)
1220 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
1221 else
1222 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
1223 }
1224
1225 /*
1226 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
1227 * and increase priv->next by 1.
1228 */
1229 int
1230 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
1231 {
1232 struct iovec *iov;
1233 uio_t *uio = &xuio->xu_uio;
1234 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1235 int i = priv->next++;
1236
1237 ASSERT(i < priv->cnt);
1238 ASSERT(off + n <= arc_buf_lsize(abuf));
1239 iov = uio->uio_iov + i;
1240 iov->iov_base = (char *)abuf->b_data + off;
1241 iov->iov_len = n;
1242 priv->bufs[i] = abuf;
1243 return (0);
1244 }
1245
1246 int
1247 dmu_xuio_cnt(xuio_t *xuio)
1248 {
1249 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1250 return (priv->cnt);
1251 }
1252
1253 arc_buf_t *
1254 dmu_xuio_arcbuf(xuio_t *xuio, int i)
1255 {
1256 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1257
1258 ASSERT(i < priv->cnt);
1259 return (priv->bufs[i]);
1260 }
1261
1262 void
1263 dmu_xuio_clear(xuio_t *xuio, int i)
1264 {
1265 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1266
1267 ASSERT(i < priv->cnt);
1268 priv->bufs[i] = NULL;
1269 }
1270
1271 static void
1272 xuio_stat_init(void)
1273 {
1274 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
1275 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
1276 KSTAT_FLAG_VIRTUAL);
1277 if (xuio_ksp != NULL) {
1278 xuio_ksp->ks_data = &xuio_stats;
1279 kstat_install(xuio_ksp);
1280 }
1281 }
1282
1283 static void
1284 xuio_stat_fini(void)
1285 {
1286 if (xuio_ksp != NULL) {
1287 kstat_delete(xuio_ksp);
1288 xuio_ksp = NULL;
1289 }
1290 }
1291
1292 void
1293 xuio_stat_wbuf_copied(void)
1294 {
1295 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1296 }
1297
1298 void
1299 xuio_stat_wbuf_nocopy(void)
1300 {
1301 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
1302 }
1303
1304 #ifdef _KERNEL
1305 static int
1306 dmu_read_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size)
1307 {
1308 dmu_buf_t **dbp;
1309 int numbufs, i, err;
1310 xuio_t *xuio = NULL;
1311
1312 /*
1313 * NB: we could do this block-at-a-time, but it's nice
1314 * to be reading in parallel.
1315 */
1316 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1317 TRUE, FTAG, &numbufs, &dbp, 0);
1318 if (err)
1319 return (err);
1320
1321 if (uio->uio_extflg == UIO_XUIO)
1322 xuio = (xuio_t *)uio;
1323
1324 for (i = 0; i < numbufs; i++) {
1325 int tocpy;
1326 int bufoff;
1327 dmu_buf_t *db = dbp[i];
1328
1329 ASSERT(size > 0);
1330
1331 bufoff = uio->uio_loffset - db->db_offset;
1332 tocpy = (int)MIN(db->db_size - bufoff, size);
1333
1334 if (xuio) {
1335 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
1336 arc_buf_t *dbuf_abuf = dbi->db_buf;
1337 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
1338 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
1339 if (!err) {
1340 uio->uio_resid -= tocpy;
1341 uio->uio_loffset += tocpy;
1342 }
1343
1344 if (abuf == dbuf_abuf)
1345 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
1346 else
1347 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
1348 } else {
1349 err = uiomove((char *)db->db_data + bufoff, tocpy,
1350 UIO_READ, uio);
1351 }
1352 if (err)
1353 break;
1354
1355 size -= tocpy;
1356 }
1357 dmu_buf_rele_array(dbp, numbufs, FTAG);
1358
1359 return (err);
1360 }
1361
1362 /*
1363 * Read 'size' bytes into the uio buffer.
1364 * From object zdb->db_object.
1365 * Starting at offset uio->uio_loffset.
1366 *
1367 * If the caller already has a dbuf in the target object
1368 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1369 * because we don't have to find the dnode_t for the object.
1370 */
1371 int
1372 dmu_read_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size)
1373 {
1374 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1375 dnode_t *dn;
1376 int err;
1377
1378 if (size == 0)
1379 return (0);
1380
1381 DB_DNODE_ENTER(db);
1382 dn = DB_DNODE(db);
1383 err = dmu_read_uio_dnode(dn, uio, size);
1384 DB_DNODE_EXIT(db);
1385
1386 return (err);
1387 }
1388
1389 /*
1390 * Read 'size' bytes into the uio buffer.
1391 * From the specified object
1392 * Starting at offset uio->uio_loffset.
1393 */
1394 int
1395 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
1396 {
1397 dnode_t *dn;
1398 int err;
1399
1400 if (size == 0)
1401 return (0);
1402
1403 err = dnode_hold(os, object, FTAG, &dn);
1404 if (err)
1405 return (err);
1406
1407 err = dmu_read_uio_dnode(dn, uio, size);
1408
1409 dnode_rele(dn, FTAG);
1410
1411 return (err);
1412 }
1413
1414 static int
1415 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1416 {
1417 dmu_buf_t **dbp;
1418 int numbufs;
1419 int err = 0;
1420 int i;
1421
1422 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1423 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1424 if (err)
1425 return (err);
1426
1427 for (i = 0; i < numbufs; i++) {
1428 int tocpy;
1429 int bufoff;
1430 dmu_buf_t *db = dbp[i];
1431
1432 ASSERT(size > 0);
1433
1434 bufoff = uio->uio_loffset - db->db_offset;
1435 tocpy = (int)MIN(db->db_size - bufoff, size);
1436
1437 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1438
1439 if (tocpy == db->db_size)
1440 dmu_buf_will_fill(db, tx);
1441 else
1442 dmu_buf_will_dirty(db, tx);
1443
1444 /*
1445 * XXX uiomove could block forever (eg. nfs-backed
1446 * pages). There needs to be a uiolockdown() function
1447 * to lock the pages in memory, so that uiomove won't
1448 * block.
1449 */
1450 err = uiomove((char *)db->db_data + bufoff, tocpy,
1451 UIO_WRITE, uio);
1452
1453 if (tocpy == db->db_size)
1454 dmu_buf_fill_done(db, tx);
1455
1456 if (err)
1457 break;
1458
1459 size -= tocpy;
1460 }
1461
1462 dmu_buf_rele_array(dbp, numbufs, FTAG);
1463 return (err);
1464 }
1465
1466 /*
1467 * Write 'size' bytes from the uio buffer.
1468 * To object zdb->db_object.
1469 * Starting at offset uio->uio_loffset.
1470 *
1471 * If the caller already has a dbuf in the target object
1472 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1473 * because we don't have to find the dnode_t for the object.
1474 */
1475 int
1476 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1477 dmu_tx_t *tx)
1478 {
1479 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1480 dnode_t *dn;
1481 int err;
1482
1483 if (size == 0)
1484 return (0);
1485
1486 DB_DNODE_ENTER(db);
1487 dn = DB_DNODE(db);
1488 err = dmu_write_uio_dnode(dn, uio, size, tx);
1489 DB_DNODE_EXIT(db);
1490
1491 return (err);
1492 }
1493
1494 /*
1495 * Write 'size' bytes from the uio buffer.
1496 * To the specified object.
1497 * Starting at offset uio->uio_loffset.
1498 */
1499 int
1500 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1501 dmu_tx_t *tx)
1502 {
1503 dnode_t *dn;
1504 int err;
1505
1506 if (size == 0)
1507 return (0);
1508
1509 err = dnode_hold(os, object, FTAG, &dn);
1510 if (err)
1511 return (err);
1512
1513 err = dmu_write_uio_dnode(dn, uio, size, tx);
1514
1515 dnode_rele(dn, FTAG);
1516
1517 return (err);
1518 }
1519
1520 int
1521 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1522 page_t *pp, dmu_tx_t *tx)
1523 {
1524 dmu_buf_t **dbp;
1525 int numbufs, i;
1526 int err;
1527
1528 if (size == 0)
1529 return (0);
1530
1531 err = dmu_buf_hold_array(os, object, offset, size,
1532 FALSE, FTAG, &numbufs, &dbp);
1533 if (err)
1534 return (err);
1535
1536 for (i = 0; i < numbufs; i++) {
1537 int tocpy, copied, thiscpy;
1538 int bufoff;
1539 dmu_buf_t *db = dbp[i];
1540 caddr_t va;
1541
1542 ASSERT(size > 0);
1543 ASSERT3U(db->db_size, >=, PAGESIZE);
1544
1545 bufoff = offset - db->db_offset;
1546 tocpy = (int)MIN(db->db_size - bufoff, size);
1547
1548 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1549
1550 if (tocpy == db->db_size)
1551 dmu_buf_will_fill(db, tx);
1552 else
1553 dmu_buf_will_dirty(db, tx);
1554
1555 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1556 ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff);
1557 thiscpy = MIN(PAGESIZE, tocpy - copied);
1558 va = zfs_map_page(pp, S_READ);
1559 bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1560 zfs_unmap_page(pp, va);
1561 pp = pp->p_next;
1562 bufoff += PAGESIZE;
1563 }
1564
1565 if (tocpy == db->db_size)
1566 dmu_buf_fill_done(db, tx);
1567
1568 offset += tocpy;
1569 size -= tocpy;
1570 }
1571 dmu_buf_rele_array(dbp, numbufs, FTAG);
1572 return (err);
1573 }
1574 #endif
1575
1576 /*
1577 * Allocate a loaned anonymous arc buffer.
1578 */
1579 arc_buf_t *
1580 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1581 {
1582 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1583
1584 return (arc_loan_buf(db->db_objset->os_spa, B_FALSE, size));
1585 }
1586
1587 /*
1588 * Free a loaned arc buffer.
1589 */
1590 void
1591 dmu_return_arcbuf(arc_buf_t *buf)
1592 {
1593 arc_return_buf(buf, FTAG);
1594 arc_buf_destroy(buf, FTAG);
1595 }
1596
1597 /*
1598 * When possible directly assign passed loaned arc buffer to a dbuf.
1599 * If this is not possible copy the contents of passed arc buf via
1600 * dmu_write().
1601 */
1602 void
1603 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1604 dmu_tx_t *tx)
1605 {
1606 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1607 dnode_t *dn;
1608 dmu_buf_impl_t *db;
1609 uint32_t blksz = (uint32_t)arc_buf_lsize(buf);
1610 uint64_t blkid;
1611
1612 DB_DNODE_ENTER(dbuf);
1613 dn = DB_DNODE(dbuf);
1614 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1615 blkid = dbuf_whichblock(dn, 0, offset);
1616 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1617 rw_exit(&dn->dn_struct_rwlock);
1618 DB_DNODE_EXIT(dbuf);
1619
1620 /*
1621 * We can only assign if the offset is aligned, the arc buf is the
1622 * same size as the dbuf, and the dbuf is not metadata.
1623 */
1624 if (offset == db->db.db_offset && blksz == db->db.db_size) {
1625 dbuf_assign_arcbuf(db, buf, tx);
1626 dbuf_rele(db, FTAG);
1627 } else {
1628 objset_t *os;
1629 uint64_t object;
1630
1631 /* compressed bufs must always be assignable to their dbuf */
1632 ASSERT3U(arc_get_compression(buf), ==, ZIO_COMPRESS_OFF);
1633 ASSERT(!(buf->b_flags & ARC_BUF_FLAG_COMPRESSED));
1634
1635 DB_DNODE_ENTER(dbuf);
1636 dn = DB_DNODE(dbuf);
1637 os = dn->dn_objset;
1638 object = dn->dn_object;
1639 DB_DNODE_EXIT(dbuf);
1640
1641 dbuf_rele(db, FTAG);
1642 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1643 dmu_return_arcbuf(buf);
1644 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1645 }
1646 }
1647
1648 typedef struct {
1649 dbuf_dirty_record_t *dsa_dr;
1650 dmu_sync_cb_t *dsa_done;
1651 zgd_t *dsa_zgd;
1652 dmu_tx_t *dsa_tx;
1653 } dmu_sync_arg_t;
1654
1655 /* ARGSUSED */
1656 static void
1657 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1658 {
1659 dmu_sync_arg_t *dsa = varg;
1660 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1661 blkptr_t *bp = zio->io_bp;
1662
1663 if (zio->io_error == 0) {
1664 if (BP_IS_HOLE(bp)) {
1665 /*
1666 * A block of zeros may compress to a hole, but the
1667 * block size still needs to be known for replay.
1668 */
1669 BP_SET_LSIZE(bp, db->db_size);
1670 } else if (!BP_IS_EMBEDDED(bp)) {
1671 ASSERT(BP_GET_LEVEL(bp) == 0);
1672 bp->blk_fill = 1;
1673 }
1674 }
1675 }
1676
1677 static void
1678 dmu_sync_late_arrival_ready(zio_t *zio)
1679 {
1680 dmu_sync_ready(zio, NULL, zio->io_private);
1681 }
1682
1683 /* ARGSUSED */
1684 static void
1685 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1686 {
1687 dmu_sync_arg_t *dsa = varg;
1688 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1689 dmu_buf_impl_t *db = dr->dr_dbuf;
1690
1691 mutex_enter(&db->db_mtx);
1692 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1693 if (zio->io_error == 0) {
1694 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1695 if (dr->dt.dl.dr_nopwrite) {
1696 blkptr_t *bp = zio->io_bp;
1697 blkptr_t *bp_orig = &zio->io_bp_orig;
1698 uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1699
1700 ASSERT(BP_EQUAL(bp, bp_orig));
1701 VERIFY(BP_EQUAL(bp, db->db_blkptr));
1702 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1703 ASSERT(zio_checksum_table[chksum].ci_flags &
1704 ZCHECKSUM_FLAG_NOPWRITE);
1705 }
1706 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1707 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1708 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1709
1710 /*
1711 * Old style holes are filled with all zeros, whereas
1712 * new-style holes maintain their lsize, type, level,
1713 * and birth time (see zio_write_compress). While we
1714 * need to reset the BP_SET_LSIZE() call that happened
1715 * in dmu_sync_ready for old style holes, we do *not*
1716 * want to wipe out the information contained in new
1717 * style holes. Thus, only zero out the block pointer if
1718 * it's an old style hole.
1719 */
1720 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) &&
1721 dr->dt.dl.dr_overridden_by.blk_birth == 0)
1722 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1723 } else {
1724 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1725 }
1726 cv_broadcast(&db->db_changed);
1727 mutex_exit(&db->db_mtx);
1728
1729 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1730
1731 kmem_free(dsa, sizeof (*dsa));
1732 }
1733
1734 static void
1735 dmu_sync_late_arrival_done(zio_t *zio)
1736 {
1737 blkptr_t *bp = zio->io_bp;
1738 dmu_sync_arg_t *dsa = zio->io_private;
1739 blkptr_t *bp_orig = &zio->io_bp_orig;
1740
1741 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1742 ASSERT(!(zio->io_flags & ZIO_FLAG_NOPWRITE));
1743 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1744 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1745 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1746 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1747 }
1748
1749 dmu_tx_commit(dsa->dsa_tx);
1750
1751 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1752
1753 abd_put(zio->io_abd);
1754 kmem_free(dsa, sizeof (*dsa));
1755 }
1756
1757 static int
1758 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1759 zio_prop_t *zp, zbookmark_phys_t *zb)
1760 {
1761 dmu_sync_arg_t *dsa;
1762 dmu_tx_t *tx;
1763
1764 tx = dmu_tx_create(os);
1765 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1766 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1767 dmu_tx_abort(tx);
1768 /* Make zl_get_data do txg_waited_synced() */
1769 return (SET_ERROR(EIO));
1770 }
1771
1772 /*
1773 * In order to prevent the zgd's lwb from being free'd prior to
1774 * dmu_sync_late_arrival_done() being called, we have to ensure
1775 * the lwb's "max txg" takes this tx's txg into account.
1776 */
1777 zil_lwb_add_txg(zgd->zgd_lwb, dmu_tx_get_txg(tx));
1778
1779 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1780 dsa->dsa_dr = NULL;
1781 dsa->dsa_done = done;
1782 dsa->dsa_zgd = zgd;
1783 dsa->dsa_tx = tx;
1784
1785 /*
1786 * Since we are currently syncing this txg, it's nontrivial to
1787 * determine what BP to nopwrite against, so we disable nopwrite.
1788 *
1789 * When syncing, the db_blkptr is initially the BP of the previous
1790 * txg. We can not nopwrite against it because it will be changed
1791 * (this is similar to the non-late-arrival case where the dbuf is
1792 * dirty in a future txg).
1793 *
1794 * Then dbuf_write_ready() sets bp_blkptr to the location we will write.
1795 * We can not nopwrite against it because although the BP will not
1796 * (typically) be changed, the data has not yet been persisted to this
1797 * location.
1798 *
1799 * Finally, when dbuf_write_done() is called, it is theoretically
1800 * possible to always nopwrite, because the data that was written in
1801 * this txg is the same data that we are trying to write. However we
1802 * would need to check that this dbuf is not dirty in any future
1803 * txg's (as we do in the normal dmu_sync() path). For simplicity, we
1804 * don't nopwrite in this case.
1805 */
1806 zp->zp_nopwrite = B_FALSE;
1807
1808 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1809 abd_get_from_buf(zgd->zgd_db->db_data, zgd->zgd_db->db_size),
1810 zgd->zgd_db->db_size, zgd->zgd_db->db_size, zp,
1811 dmu_sync_late_arrival_ready, NULL, NULL, dmu_sync_late_arrival_done,
1812 dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1813
1814 return (0);
1815 }
1816
1817 /*
1818 * Intent log support: sync the block associated with db to disk.
1819 * N.B. and XXX: the caller is responsible for making sure that the
1820 * data isn't changing while dmu_sync() is writing it.
1821 *
1822 * Return values:
1823 *
1824 * EEXIST: this txg has already been synced, so there's nothing to do.
1825 * The caller should not log the write.
1826 *
1827 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1828 * The caller should not log the write.
1829 *
1830 * EALREADY: this block is already in the process of being synced.
1831 * The caller should track its progress (somehow).
1832 *
1833 * EIO: could not do the I/O.
1834 * The caller should do a txg_wait_synced().
1835 *
1836 * 0: the I/O has been initiated.
1837 * The caller should log this blkptr in the done callback.
1838 * It is possible that the I/O will fail, in which case
1839 * the error will be reported to the done callback and
1840 * propagated to pio from zio_done().
1841 */
1842 int
1843 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1844 {
1845 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1846 objset_t *os = db->db_objset;
1847 dsl_dataset_t *ds = os->os_dsl_dataset;
1848 dbuf_dirty_record_t *dr;
1849 dmu_sync_arg_t *dsa;
1850 zbookmark_phys_t zb;
1851 zio_prop_t zp;
1852 dnode_t *dn;
1853
1854 ASSERT(pio != NULL);
1855 ASSERT(txg != 0);
1856
1857 SET_BOOKMARK(&zb, ds->ds_object,
1858 db->db.db_object, db->db_level, db->db_blkid);
1859
1860 DB_DNODE_ENTER(db);
1861 dn = DB_DNODE(db);
1862 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1863 DB_DNODE_EXIT(db);
1864
1865 /*
1866 * If we're frozen (running ziltest), we always need to generate a bp.
1867 */
1868 if (txg > spa_freeze_txg(os->os_spa))
1869 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1870
1871 /*
1872 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1873 * and us. If we determine that this txg is not yet syncing,
1874 * but it begins to sync a moment later, that's OK because the
1875 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1876 */
1877 mutex_enter(&db->db_mtx);
1878
1879 if (txg <= spa_last_synced_txg(os->os_spa)) {
1880 /*
1881 * This txg has already synced. There's nothing to do.
1882 */
1883 mutex_exit(&db->db_mtx);
1884 return (SET_ERROR(EEXIST));
1885 }
1886
1887 if (txg <= spa_syncing_txg(os->os_spa)) {
1888 /*
1889 * This txg is currently syncing, so we can't mess with
1890 * the dirty record anymore; just write a new log block.
1891 */
1892 mutex_exit(&db->db_mtx);
1893 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1894 }
1895
1896 dr = db->db_last_dirty;
1897 while (dr && dr->dr_txg != txg)
1898 dr = dr->dr_next;
1899
1900 if (dr == NULL) {
1901 /*
1902 * There's no dr for this dbuf, so it must have been freed.
1903 * There's no need to log writes to freed blocks, so we're done.
1904 */
1905 mutex_exit(&db->db_mtx);
1906 return (SET_ERROR(ENOENT));
1907 }
1908
1909 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
1910
1911 if (db->db_blkptr != NULL) {
1912 /*
1913 * We need to fill in zgd_bp with the current blkptr so that
1914 * the nopwrite code can check if we're writing the same
1915 * data that's already on disk. We can only nopwrite if we
1916 * are sure that after making the copy, db_blkptr will not
1917 * change until our i/o completes. We ensure this by
1918 * holding the db_mtx, and only allowing nopwrite if the
1919 * block is not already dirty (see below). This is verified
1920 * by dmu_sync_done(), which VERIFYs that the db_blkptr has
1921 * not changed.
1922 */
1923 *zgd->zgd_bp = *db->db_blkptr;
1924 }
1925
1926 /*
1927 * Assume the on-disk data is X, the current syncing data (in
1928 * txg - 1) is Y, and the current in-memory data is Z (currently
1929 * in dmu_sync).
1930 *
1931 * We usually want to perform a nopwrite if X and Z are the
1932 * same. However, if Y is different (i.e. the BP is going to
1933 * change before this write takes effect), then a nopwrite will
1934 * be incorrect - we would override with X, which could have
1935 * been freed when Y was written.
1936 *
1937 * (Note that this is not a concern when we are nop-writing from
1938 * syncing context, because X and Y must be identical, because
1939 * all previous txgs have been synced.)
1940 *
1941 * Therefore, we disable nopwrite if the current BP could change
1942 * before this TXG. There are two ways it could change: by
1943 * being dirty (dr_next is non-NULL), or by being freed
1944 * (dnode_block_freed()). This behavior is verified by
1945 * zio_done(), which VERIFYs that the override BP is identical
1946 * to the on-disk BP.
1947 */
1948 DB_DNODE_ENTER(db);
1949 dn = DB_DNODE(db);
1950 if (dr->dr_next != NULL || dnode_block_freed(dn, db->db_blkid))
1951 zp.zp_nopwrite = B_FALSE;
1952 DB_DNODE_EXIT(db);
1953
1954 ASSERT(dr->dr_txg == txg);
1955 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1956 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1957 /*
1958 * We have already issued a sync write for this buffer,
1959 * or this buffer has already been synced. It could not
1960 * have been dirtied since, or we would have cleared the state.
1961 */
1962 mutex_exit(&db->db_mtx);
1963 return (SET_ERROR(EALREADY));
1964 }
1965
1966 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1967 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1968 mutex_exit(&db->db_mtx);
1969
1970 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1971 dsa->dsa_dr = dr;
1972 dsa->dsa_done = done;
1973 dsa->dsa_zgd = zgd;
1974 dsa->dsa_tx = NULL;
1975
1976 zio_nowait(arc_write(pio, os->os_spa, txg,
1977 zgd->zgd_bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
1978 &zp, dmu_sync_ready, NULL, NULL, dmu_sync_done, dsa,
1979 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb));
1980
1981 return (0);
1982 }
1983
1984 int
1985 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1986 dmu_tx_t *tx)
1987 {
1988 dnode_t *dn;
1989 int err;
1990
1991 err = dnode_hold(os, object, FTAG, &dn);
1992 if (err)
1993 return (err);
1994 err = dnode_set_blksz(dn, size, ibs, tx);
1995 dnode_rele(dn, FTAG);
1996 return (err);
1997 }
1998
1999 void
2000 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
2001 dmu_tx_t *tx)
2002 {
2003 dnode_t *dn;
2004
2005 /*
2006 * Send streams include each object's checksum function. This
2007 * check ensures that the receiving system can understand the
2008 * checksum function transmitted.
2009 */
2010 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS);
2011
2012 VERIFY0(dnode_hold(os, object, FTAG, &dn));
2013 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS);
2014 dn->dn_checksum = checksum;
2015 dnode_setdirty(dn, tx);
2016 dnode_rele(dn, FTAG);
2017 }
2018
2019 void
2020 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
2021 dmu_tx_t *tx)
2022 {
2023 dnode_t *dn;
2024
2025 /*
2026 * Send streams include each object's compression function. This
2027 * check ensures that the receiving system can understand the
2028 * compression function transmitted.
2029 */
2030 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS);
2031
2032 VERIFY0(dnode_hold(os, object, FTAG, &dn));
2033 dn->dn_compress = compress;
2034 dnode_setdirty(dn, tx);
2035 dnode_rele(dn, FTAG);
2036 }
2037
2038 int zfs_mdcomp_disable = 0;
2039
2040 /*
2041 * When the "redundant_metadata" property is set to "most", only indirect
2042 * blocks of this level and higher will have an additional ditto block.
2043 */
2044 int zfs_redundant_metadata_most_ditto_level = 2;
2045
2046 void
2047 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
2048 {
2049 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
2050 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
2051 (wp & WP_SPILL));
2052 enum zio_checksum checksum = os->os_checksum;
2053 enum zio_compress compress = os->os_compress;
2054 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
2055 boolean_t dedup = B_FALSE;
2056 boolean_t nopwrite = B_FALSE;
2057 boolean_t dedup_verify = os->os_dedup_verify;
2058 int copies = os->os_copies;
2059
2060 /*
2061 * We maintain different write policies for each of the following
2062 * types of data:
2063 * 1. metadata
2064 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
2065 * 3. all other level 0 blocks
2066 */
2067 if (ismd) {
2068 if (zfs_mdcomp_disable) {
2069 compress = ZIO_COMPRESS_EMPTY;
2070 } else {
2071 /*
2072 * XXX -- we should design a compression algorithm
2073 * that specializes in arrays of bps.
2074 */
2075 compress = zio_compress_select(os->os_spa,
2076 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON);
2077 }
2078
2079 /*
2080 * Metadata always gets checksummed. If the data
2081 * checksum is multi-bit correctable, and it's not a
2082 * ZBT-style checksum, then it's suitable for metadata
2083 * as well. Otherwise, the metadata checksum defaults
2084 * to fletcher4.
2085 */
2086 if (!(zio_checksum_table[checksum].ci_flags &
2087 ZCHECKSUM_FLAG_METADATA) ||
2088 (zio_checksum_table[checksum].ci_flags &
2089 ZCHECKSUM_FLAG_EMBEDDED))
2090 checksum = ZIO_CHECKSUM_FLETCHER_4;
2091
2092 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL ||
2093 (os->os_redundant_metadata ==
2094 ZFS_REDUNDANT_METADATA_MOST &&
2095 (level >= zfs_redundant_metadata_most_ditto_level ||
2096 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))))
2097 copies++;
2098 } else if (wp & WP_NOFILL) {
2099 ASSERT(level == 0);
2100
2101 /*
2102 * If we're writing preallocated blocks, we aren't actually
2103 * writing them so don't set any policy properties. These
2104 * blocks are currently only used by an external subsystem
2105 * outside of zfs (i.e. dump) and not written by the zio
2106 * pipeline.
2107 */
2108 compress = ZIO_COMPRESS_OFF;
2109 checksum = ZIO_CHECKSUM_NOPARITY;
2110 } else {
2111 compress = zio_compress_select(os->os_spa, dn->dn_compress,
2112 compress);
2113
2114 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
2115 zio_checksum_select(dn->dn_checksum, checksum) :
2116 dedup_checksum;
2117
2118 /*
2119 * Determine dedup setting. If we are in dmu_sync(),
2120 * we won't actually dedup now because that's all
2121 * done in syncing context; but we do want to use the
2122 * dedup checkum. If the checksum is not strong
2123 * enough to ensure unique signatures, force
2124 * dedup_verify.
2125 */
2126 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
2127 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
2128 if (!(zio_checksum_table[checksum].ci_flags &
2129 ZCHECKSUM_FLAG_DEDUP))
2130 dedup_verify = B_TRUE;
2131 }
2132
2133 /*
2134 * Enable nopwrite if we have secure enough checksum
2135 * algorithm (see comment in zio_nop_write) and
2136 * compression is enabled. We don't enable nopwrite if
2137 * dedup is enabled as the two features are mutually
2138 * exclusive.
2139 */
2140 nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags &
2141 ZCHECKSUM_FLAG_NOPWRITE) &&
2142 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
2143 }
2144
2145 zp->zp_checksum = checksum;
2146 zp->zp_compress = compress;
2147 ASSERT3U(zp->zp_compress, !=, ZIO_COMPRESS_INHERIT);
2148
2149 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
2150 zp->zp_level = level;
2151 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
2152 zp->zp_dedup = dedup;
2153 zp->zp_dedup_verify = dedup && dedup_verify;
2154 zp->zp_nopwrite = nopwrite;
2155 }
2156
2157 int
2158 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
2159 {
2160 dnode_t *dn;
2161 int err;
2162
2163 /*
2164 * Sync any current changes before
2165 * we go trundling through the block pointers.
2166 */
2167 err = dmu_object_wait_synced(os, object);
2168 if (err) {
2169 return (err);
2170 }
2171
2172 err = dnode_hold(os, object, FTAG, &dn);
2173 if (err) {
2174 return (err);
2175 }
2176
2177 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
2178 dnode_rele(dn, FTAG);
2179
2180 return (err);
2181 }
2182
2183 /*
2184 * Given the ZFS object, if it contains any dirty nodes
2185 * this function flushes all dirty blocks to disk. This
2186 * ensures the DMU object info is updated. A more efficient
2187 * future version might just find the TXG with the maximum
2188 * ID and wait for that to be synced.
2189 */
2190 int
2191 dmu_object_wait_synced(objset_t *os, uint64_t object)
2192 {
2193 dnode_t *dn;
2194 int error, i;
2195
2196 error = dnode_hold(os, object, FTAG, &dn);
2197 if (error) {
2198 return (error);
2199 }
2200
2201 for (i = 0; i < TXG_SIZE; i++) {
2202 if (list_link_active(&dn->dn_dirty_link[i])) {
2203 break;
2204 }
2205 }
2206 dnode_rele(dn, FTAG);
2207 if (i != TXG_SIZE) {
2208 txg_wait_synced(dmu_objset_pool(os), 0);
2209 }
2210
2211 return (0);
2212 }
2213
2214 void
2215 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2216 {
2217 dnode_phys_t *dnp;
2218
2219 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2220 mutex_enter(&dn->dn_mtx);
2221
2222 dnp = dn->dn_phys;
2223
2224 doi->doi_data_block_size = dn->dn_datablksz;
2225 doi->doi_metadata_block_size = dn->dn_indblkshift ?
2226 1ULL << dn->dn_indblkshift : 0;
2227 doi->doi_type = dn->dn_type;
2228 doi->doi_bonus_type = dn->dn_bonustype;
2229 doi->doi_bonus_size = dn->dn_bonuslen;
2230 doi->doi_indirection = dn->dn_nlevels;
2231 doi->doi_checksum = dn->dn_checksum;
2232 doi->doi_compress = dn->dn_compress;
2233 doi->doi_nblkptr = dn->dn_nblkptr;
2234 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
2235 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
2236 doi->doi_fill_count = 0;
2237 for (int i = 0; i < dnp->dn_nblkptr; i++)
2238 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
2239
2240 mutex_exit(&dn->dn_mtx);
2241 rw_exit(&dn->dn_struct_rwlock);
2242 }
2243
2244 /*
2245 * Get information on a DMU object.
2246 * If doi is NULL, just indicates whether the object exists.
2247 */
2248 int
2249 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
2250 {
2251 dnode_t *dn;
2252 int err = dnode_hold(os, object, FTAG, &dn);
2253
2254 if (err)
2255 return (err);
2256
2257 if (doi != NULL)
2258 dmu_object_info_from_dnode(dn, doi);
2259
2260 dnode_rele(dn, FTAG);
2261 return (0);
2262 }
2263
2264 /*
2265 * As above, but faster; can be used when you have a held dbuf in hand.
2266 */
2267 void
2268 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
2269 {
2270 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2271
2272 DB_DNODE_ENTER(db);
2273 dmu_object_info_from_dnode(DB_DNODE(db), doi);
2274 DB_DNODE_EXIT(db);
2275 }
2276
2277 /*
2278 * Faster still when you only care about the size.
2279 * This is specifically optimized for zfs_getattr().
2280 */
2281 void
2282 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
2283 u_longlong_t *nblk512)
2284 {
2285 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2286 dnode_t *dn;
2287
2288 DB_DNODE_ENTER(db);
2289 dn = DB_DNODE(db);
2290
2291 *blksize = dn->dn_datablksz;
2292 /* add 1 for dnode space */
2293 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
2294 SPA_MINBLOCKSHIFT) + 1;
2295 DB_DNODE_EXIT(db);
2296 }
2297
2298 void
2299 byteswap_uint64_array(void *vbuf, size_t size)
2300 {
2301 uint64_t *buf = vbuf;
2302 size_t count = size >> 3;
2303 int i;
2304
2305 ASSERT((size & 7) == 0);
2306
2307 for (i = 0; i < count; i++)
2308 buf[i] = BSWAP_64(buf[i]);
2309 }
2310
2311 void
2312 byteswap_uint32_array(void *vbuf, size_t size)
2313 {
2314 uint32_t *buf = vbuf;
2315 size_t count = size >> 2;
2316 int i;
2317
2318 ASSERT((size & 3) == 0);
2319
2320 for (i = 0; i < count; i++)
2321 buf[i] = BSWAP_32(buf[i]);
2322 }
2323
2324 void
2325 byteswap_uint16_array(void *vbuf, size_t size)
2326 {
2327 uint16_t *buf = vbuf;
2328 size_t count = size >> 1;
2329 int i;
2330
2331 ASSERT((size & 1) == 0);
2332
2333 for (i = 0; i < count; i++)
2334 buf[i] = BSWAP_16(buf[i]);
2335 }
2336
2337 /* ARGSUSED */
2338 void
2339 byteswap_uint8_array(void *vbuf, size_t size)
2340 {
2341 }
2342
2343 void
2344 dmu_init(void)
2345 {
2346 abd_init();
2347 zfs_dbgmsg_init();
2348 sa_cache_init();
2349 xuio_stat_init();
2350 dmu_objset_init();
2351 dnode_init();
2352 zfetch_init();
2353 l2arc_init();
2354 arc_init();
2355 dbuf_init();
2356 }
2357
2358 void
2359 dmu_fini(void)
2360 {
2361 arc_fini(); /* arc depends on l2arc, so arc must go first */
2362 l2arc_fini();
2363 zfetch_fini();
2364 dbuf_fini();
2365 dnode_fini();
2366 dmu_objset_fini();
2367 xuio_stat_fini();
2368 sa_cache_fini();
2369 zfs_dbgmsg_fini();
2370 abd_fini();
2371 }