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) 2013 by Delphix. All rights reserved.
24 */
25
26 /* Copyright (c) 2013 by Saso Kiselkov. All rights reserved. */
27
28 #include <sys/dmu.h>
29 #include <sys/dmu_impl.h>
30 #include <sys/dmu_tx.h>
31 #include <sys/dbuf.h>
32 #include <sys/dnode.h>
33 #include <sys/zfs_context.h>
34 #include <sys/dmu_objset.h>
35 #include <sys/dmu_traverse.h>
36 #include <sys/dsl_dataset.h>
37 #include <sys/dsl_dir.h>
38 #include <sys/dsl_pool.h>
39 #include <sys/dsl_synctask.h>
40 #include <sys/dsl_prop.h>
41 #include <sys/dmu_zfetch.h>
42 #include <sys/zfs_ioctl.h>
43 #include <sys/zap.h>
44 #include <sys/zio_checksum.h>
45 #include <sys/zio_compress.h>
46 #include <sys/sa.h>
47 #ifdef _KERNEL
48 #include <sys/vmsystm.h>
49 #include <sys/zfs_znode.h>
50 #endif
51
52 /*
53 * Enable/disable nopwrite feature.
54 */
55 int zfs_nopwrite_enabled = 1;
56
57 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
58 { DMU_BSWAP_UINT8, TRUE, "unallocated" },
59 { DMU_BSWAP_ZAP, TRUE, "object directory" },
60 { DMU_BSWAP_UINT64, TRUE, "object array" },
61 { DMU_BSWAP_UINT8, TRUE, "packed nvlist" },
62 { DMU_BSWAP_UINT64, TRUE, "packed nvlist size" },
63 { DMU_BSWAP_UINT64, TRUE, "bpobj" },
64 { DMU_BSWAP_UINT64, TRUE, "bpobj header" },
65 { DMU_BSWAP_UINT64, TRUE, "SPA space map header" },
66 { DMU_BSWAP_UINT64, TRUE, "SPA space map" },
67 { DMU_BSWAP_UINT64, TRUE, "ZIL intent log" },
68 { DMU_BSWAP_DNODE, TRUE, "DMU dnode" },
69 { DMU_BSWAP_OBJSET, TRUE, "DMU objset" },
70 { DMU_BSWAP_UINT64, TRUE, "DSL directory" },
71 { DMU_BSWAP_ZAP, TRUE, "DSL directory child map"},
72 { DMU_BSWAP_ZAP, TRUE, "DSL dataset snap map" },
73 { DMU_BSWAP_ZAP, TRUE, "DSL props" },
74 { DMU_BSWAP_UINT64, TRUE, "DSL dataset" },
75 { DMU_BSWAP_ZNODE, TRUE, "ZFS znode" },
76 { DMU_BSWAP_OLDACL, TRUE, "ZFS V0 ACL" },
77 { DMU_BSWAP_UINT8, FALSE, "ZFS plain file" },
78 { DMU_BSWAP_ZAP, TRUE, "ZFS directory" },
79 { DMU_BSWAP_ZAP, TRUE, "ZFS master node" },
80 { DMU_BSWAP_ZAP, TRUE, "ZFS delete queue" },
81 { DMU_BSWAP_UINT8, FALSE, "zvol object" },
82 { DMU_BSWAP_ZAP, TRUE, "zvol prop" },
83 { DMU_BSWAP_UINT8, FALSE, "other uint8[]" },
84 { DMU_BSWAP_UINT64, FALSE, "other uint64[]" },
85 { DMU_BSWAP_ZAP, TRUE, "other ZAP" },
86 { DMU_BSWAP_ZAP, TRUE, "persistent error log" },
87 { DMU_BSWAP_UINT8, TRUE, "SPA history" },
88 { DMU_BSWAP_UINT64, TRUE, "SPA history offsets" },
89 { DMU_BSWAP_ZAP, TRUE, "Pool properties" },
90 { DMU_BSWAP_ZAP, TRUE, "DSL permissions" },
91 { DMU_BSWAP_ACL, TRUE, "ZFS ACL" },
92 { DMU_BSWAP_UINT8, TRUE, "ZFS SYSACL" },
93 { DMU_BSWAP_UINT8, TRUE, "FUID table" },
94 { DMU_BSWAP_UINT64, TRUE, "FUID table size" },
95 { DMU_BSWAP_ZAP, TRUE, "DSL dataset next clones"},
96 { DMU_BSWAP_ZAP, TRUE, "scan work queue" },
97 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group used" },
98 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group quota" },
99 { DMU_BSWAP_ZAP, TRUE, "snapshot refcount tags"},
100 { DMU_BSWAP_ZAP, TRUE, "DDT ZAP algorithm" },
101 { DMU_BSWAP_ZAP, TRUE, "DDT statistics" },
102 { DMU_BSWAP_UINT8, TRUE, "System attributes" },
103 { DMU_BSWAP_ZAP, TRUE, "SA master node" },
104 { DMU_BSWAP_ZAP, TRUE, "SA attr registration" },
105 { DMU_BSWAP_ZAP, TRUE, "SA attr layouts" },
106 { DMU_BSWAP_ZAP, TRUE, "scan translations" },
107 { DMU_BSWAP_UINT8, FALSE, "deduplicated block" },
108 { DMU_BSWAP_ZAP, TRUE, "DSL deadlist map" },
109 { DMU_BSWAP_UINT64, TRUE, "DSL deadlist map hdr" },
110 { DMU_BSWAP_ZAP, TRUE, "DSL dir clones" },
111 { DMU_BSWAP_UINT64, TRUE, "bpobj subobj" }
112 };
113
114 const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
115 { byteswap_uint8_array, "uint8" },
116 { byteswap_uint16_array, "uint16" },
117 { byteswap_uint32_array, "uint32" },
118 { byteswap_uint64_array, "uint64" },
119 { zap_byteswap, "zap" },
120 { dnode_buf_byteswap, "dnode" },
121 { dmu_objset_byteswap, "objset" },
122 { zfs_znode_byteswap, "znode" },
123 { zfs_oldacl_byteswap, "oldacl" },
124 { zfs_acl_byteswap, "acl" }
125 };
126
127 int
128 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
129 void *tag, dmu_buf_t **dbp, int flags)
130 {
131 dnode_t *dn;
132 uint64_t blkid;
133 dmu_buf_impl_t *db;
134 int err;
135 int db_flags = DB_RF_CANFAIL;
136
137 if (flags & DMU_READ_NO_PREFETCH)
138 db_flags |= DB_RF_NOPREFETCH;
139
140 err = dnode_hold(os, object, FTAG, &dn);
141 if (err)
142 return (err);
143 blkid = dbuf_whichblock(dn, offset);
144 rw_enter(&dn->dn_struct_rwlock, RW_READER);
145 db = dbuf_hold(dn, blkid, tag);
146 rw_exit(&dn->dn_struct_rwlock);
147 if (db == NULL) {
148 err = SET_ERROR(EIO);
149 } else {
150 err = dbuf_read(db, NULL, db_flags);
151 if (err) {
152 dbuf_rele(db, tag);
153 db = NULL;
154 }
155 }
156
157 dnode_rele(dn, FTAG);
158 *dbp = &db->db; /* NULL db plus first field offset is NULL */
159 return (err);
160 }
161
162 int
163 dmu_bonus_max(void)
164 {
165 return (DN_MAX_BONUSLEN);
166 }
167
168 int
169 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
170 {
171 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
172 dnode_t *dn;
173 int error;
174
175 DB_DNODE_ENTER(db);
176 dn = DB_DNODE(db);
177
178 if (dn->dn_bonus != db) {
179 error = SET_ERROR(EINVAL);
180 } else if (newsize < 0 || newsize > db_fake->db_size) {
181 error = SET_ERROR(EINVAL);
182 } else {
183 dnode_setbonuslen(dn, newsize, tx);
184 error = 0;
185 }
186
187 DB_DNODE_EXIT(db);
188 return (error);
189 }
190
191 int
192 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
193 {
194 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
195 dnode_t *dn;
196 int error;
197
198 DB_DNODE_ENTER(db);
199 dn = DB_DNODE(db);
200
201 if (!DMU_OT_IS_VALID(type)) {
202 error = SET_ERROR(EINVAL);
203 } else if (dn->dn_bonus != db) {
204 error = SET_ERROR(EINVAL);
205 } else {
206 dnode_setbonus_type(dn, type, tx);
207 error = 0;
208 }
209
210 DB_DNODE_EXIT(db);
211 return (error);
212 }
213
214 dmu_object_type_t
215 dmu_get_bonustype(dmu_buf_t *db_fake)
216 {
217 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
218 dnode_t *dn;
219 dmu_object_type_t type;
220
221 DB_DNODE_ENTER(db);
222 dn = DB_DNODE(db);
223 type = dn->dn_bonustype;
224 DB_DNODE_EXIT(db);
225
226 return (type);
227 }
228
229 int
230 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
231 {
232 dnode_t *dn;
233 int error;
234
235 error = dnode_hold(os, object, FTAG, &dn);
236 dbuf_rm_spill(dn, tx);
237 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
238 dnode_rm_spill(dn, tx);
239 rw_exit(&dn->dn_struct_rwlock);
240 dnode_rele(dn, FTAG);
241 return (error);
242 }
243
244 /*
245 * returns ENOENT, EIO, or 0.
246 */
247 int
248 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
249 {
250 dnode_t *dn;
251 dmu_buf_impl_t *db;
252 int error;
253
254 error = dnode_hold(os, object, FTAG, &dn);
255 if (error)
256 return (error);
257
258 rw_enter(&dn->dn_struct_rwlock, RW_READER);
259 if (dn->dn_bonus == NULL) {
260 rw_exit(&dn->dn_struct_rwlock);
261 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
262 if (dn->dn_bonus == NULL)
263 dbuf_create_bonus(dn);
264 }
265 db = dn->dn_bonus;
266
267 /* as long as the bonus buf is held, the dnode will be held */
268 if (refcount_add(&db->db_holds, tag) == 1) {
269 VERIFY(dnode_add_ref(dn, db));
270 (void) atomic_inc_32_nv(&dn->dn_dbufs_count);
271 }
272
273 /*
274 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
275 * hold and incrementing the dbuf count to ensure that dnode_move() sees
276 * a dnode hold for every dbuf.
277 */
278 rw_exit(&dn->dn_struct_rwlock);
279
280 dnode_rele(dn, FTAG);
281
282 VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH));
283
284 *dbp = &db->db;
285 return (0);
286 }
287
288 /*
289 * returns ENOENT, EIO, or 0.
290 *
291 * This interface will allocate a blank spill dbuf when a spill blk
292 * doesn't already exist on the dnode.
293 *
294 * if you only want to find an already existing spill db, then
295 * dmu_spill_hold_existing() should be used.
296 */
297 int
298 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
299 {
300 dmu_buf_impl_t *db = NULL;
301 int err;
302
303 if ((flags & DB_RF_HAVESTRUCT) == 0)
304 rw_enter(&dn->dn_struct_rwlock, RW_READER);
305
306 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
307
308 if ((flags & DB_RF_HAVESTRUCT) == 0)
309 rw_exit(&dn->dn_struct_rwlock);
310
311 ASSERT(db != NULL);
312 err = dbuf_read(db, NULL, flags);
313 if (err == 0)
314 *dbp = &db->db;
315 else
316 dbuf_rele(db, tag);
317 return (err);
318 }
319
320 int
321 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
322 {
323 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
324 dnode_t *dn;
325 int err;
326
327 DB_DNODE_ENTER(db);
328 dn = DB_DNODE(db);
329
330 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
331 err = SET_ERROR(EINVAL);
332 } else {
333 rw_enter(&dn->dn_struct_rwlock, RW_READER);
334
335 if (!dn->dn_have_spill) {
336 err = SET_ERROR(ENOENT);
337 } else {
338 err = dmu_spill_hold_by_dnode(dn,
339 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
340 }
341
342 rw_exit(&dn->dn_struct_rwlock);
343 }
344
345 DB_DNODE_EXIT(db);
346 return (err);
347 }
348
349 int
350 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
351 {
352 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
353 dnode_t *dn;
354 int err;
355
356 DB_DNODE_ENTER(db);
357 dn = DB_DNODE(db);
358 err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp);
359 DB_DNODE_EXIT(db);
360
361 return (err);
362 }
363
364 /*
365 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
366 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
367 * and can induce severe lock contention when writing to several files
368 * whose dnodes are in the same block.
369 */
370 static int
371 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
372 int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
373 {
374 dsl_pool_t *dp = NULL;
375 dmu_buf_t **dbp;
376 uint64_t blkid, nblks, i;
377 uint32_t dbuf_flags;
378 int err;
379 zio_t *zio;
380 hrtime_t start;
381
382 ASSERT(length <= DMU_MAX_ACCESS);
383
384 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT;
385 if (flags & DMU_READ_NO_PREFETCH || length > zfetch_array_rd_sz)
386 dbuf_flags |= DB_RF_NOPREFETCH;
387
388 rw_enter(&dn->dn_struct_rwlock, RW_READER);
389 if (dn->dn_datablkshift) {
390 int blkshift = dn->dn_datablkshift;
391 nblks = (P2ROUNDUP(offset+length, 1ULL<<blkshift) -
392 P2ALIGN(offset, 1ULL<<blkshift)) >> blkshift;
393 } else {
394 if (offset + length > dn->dn_datablksz) {
395 zfs_panic_recover("zfs: accessing past end of object "
396 "%llx/%llx (size=%u access=%llu+%llu)",
397 (longlong_t)dn->dn_objset->
398 os_dsl_dataset->ds_object,
399 (longlong_t)dn->dn_object, dn->dn_datablksz,
400 (longlong_t)offset, (longlong_t)length);
401 rw_exit(&dn->dn_struct_rwlock);
402 return (SET_ERROR(EIO));
403 }
404 nblks = 1;
405 }
406 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
407
408 if (dn->dn_objset->os_dsl_dataset)
409 dp = dn->dn_objset->os_dsl_dataset->ds_dir->dd_pool;
410 start = gethrtime();
411 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
412 blkid = dbuf_whichblock(dn, offset);
413 for (i = 0; i < nblks; i++) {
414 dmu_buf_impl_t *db = dbuf_hold(dn, blkid+i, tag);
415 if (db == NULL) {
416 rw_exit(&dn->dn_struct_rwlock);
417 dmu_buf_rele_array(dbp, nblks, tag);
418 zio_nowait(zio);
419 return (SET_ERROR(EIO));
420 }
421 /* initiate async i/o */
422 if (read) {
423 (void) dbuf_read(db, zio, dbuf_flags);
424 }
425 dbp[i] = &db->db;
426 }
427 rw_exit(&dn->dn_struct_rwlock);
428
429 /* wait for async i/o */
430 err = zio_wait(zio);
431 /* track read overhead when we are in sync context */
432 if (dp && dsl_pool_sync_context(dp))
433 dp->dp_read_overhead += gethrtime() - start;
434 if (err) {
435 dmu_buf_rele_array(dbp, nblks, tag);
436 return (err);
437 }
438
439 /* wait for other io to complete */
440 if (read) {
441 for (i = 0; i < nblks; i++) {
442 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
443 mutex_enter(&db->db_mtx);
444 while (db->db_state == DB_READ ||
445 db->db_state == DB_FILL)
446 cv_wait(&db->db_changed, &db->db_mtx);
447 if (db->db_state == DB_UNCACHED)
448 err = SET_ERROR(EIO);
449 mutex_exit(&db->db_mtx);
450 if (err) {
451 dmu_buf_rele_array(dbp, nblks, tag);
452 return (err);
453 }
454 }
455 }
456
457 *numbufsp = nblks;
458 *dbpp = dbp;
459 return (0);
460 }
461
462 static int
463 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
464 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
465 {
466 dnode_t *dn;
467 int err;
468
469 err = dnode_hold(os, object, FTAG, &dn);
470 if (err)
471 return (err);
472
473 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
474 numbufsp, dbpp, DMU_READ_PREFETCH);
475
476 dnode_rele(dn, FTAG);
477
478 return (err);
479 }
480
481 int
482 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
483 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
484 {
485 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
486 dnode_t *dn;
487 int err;
488
489 DB_DNODE_ENTER(db);
490 dn = DB_DNODE(db);
491 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
492 numbufsp, dbpp, DMU_READ_PREFETCH);
493 DB_DNODE_EXIT(db);
494
495 return (err);
496 }
497
498 void
499 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
500 {
501 int i;
502 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
503
504 if (numbufs == 0)
505 return;
506
507 for (i = 0; i < numbufs; i++) {
508 if (dbp[i])
509 dbuf_rele(dbp[i], tag);
510 }
511
512 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
513 }
514
515 void
516 dmu_prefetch(objset_t *os, uint64_t object, uint64_t offset, uint64_t len)
517 {
518 dnode_t *dn;
519 uint64_t blkid;
520 int nblks, i, err;
521
522 if (zfs_prefetch_disable)
523 return;
524
525 if (len == 0) { /* they're interested in the bonus buffer */
526 dn = DMU_META_DNODE(os);
527
528 if (object == 0 || object >= DN_MAX_OBJECT)
529 return;
530
531 rw_enter(&dn->dn_struct_rwlock, RW_READER);
532 blkid = dbuf_whichblock(dn, object * sizeof (dnode_phys_t));
533 dbuf_prefetch(dn, blkid);
534 rw_exit(&dn->dn_struct_rwlock);
535 return;
536 }
537
538 /*
539 * XXX - Note, if the dnode for the requested object is not
540 * already cached, we will do a *synchronous* read in the
541 * dnode_hold() call. The same is true for any indirects.
542 */
543 err = dnode_hold(os, object, FTAG, &dn);
544 if (err != 0)
545 return;
546
547 rw_enter(&dn->dn_struct_rwlock, RW_READER);
548 if (dn->dn_datablkshift) {
549 int blkshift = dn->dn_datablkshift;
550 nblks = (P2ROUNDUP(offset+len, 1<<blkshift) -
551 P2ALIGN(offset, 1<<blkshift)) >> blkshift;
552 } else {
553 nblks = (offset < dn->dn_datablksz);
554 }
555
556 if (nblks != 0) {
557 blkid = dbuf_whichblock(dn, offset);
558 for (i = 0; i < nblks; i++)
559 dbuf_prefetch(dn, blkid+i);
560 }
561
562 rw_exit(&dn->dn_struct_rwlock);
563
564 dnode_rele(dn, FTAG);
565 }
566
567 /*
568 * Get the next "chunk" of file data to free. We traverse the file from
569 * the end so that the file gets shorter over time (if we crashes in the
570 * middle, this will leave us in a better state). We find allocated file
571 * data by simply searching the allocated level 1 indirects.
572 *
573 * On input, *start should be the first offset that does not need to be
574 * freed (e.g. "offset + length"). On return, *start will be the first
575 * offset that should be freed.
576 */
577 static int
578 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum)
579 {
580 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
581 /* bytes of data covered by a level-1 indirect block */
582 uint64_t iblkrange =
583 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
584
585 ASSERT3U(minimum, <=, *start);
586
587 if (*start - minimum <= iblkrange * maxblks) {
588 *start = minimum;
589 return (0);
590 }
591 ASSERT(ISP2(iblkrange));
592
593 for (uint64_t blks = 0; *start > minimum && blks < maxblks; blks++) {
594 int err;
595
596 /*
597 * dnode_next_offset(BACKWARDS) will find an allocated L1
598 * indirect block at or before the input offset. We must
599 * decrement *start so that it is at the end of the region
600 * to search.
601 */
602 (*start)--;
603 err = dnode_next_offset(dn,
604 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
605
606 /* if there are no indirect blocks before start, we are done */
607 if (err == ESRCH) {
608 *start = minimum;
609 break;
610 } else if (err != 0) {
611 return (err);
612 }
613
614 /* set start to the beginning of this L1 indirect */
615 *start = P2ALIGN(*start, iblkrange);
616 }
617 if (*start < minimum)
618 *start = minimum;
619 return (0);
620 }
621
622 static int
623 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
624 uint64_t length)
625 {
626 uint64_t object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
627 int err;
628
629 if (offset >= object_size)
630 return (0);
631
632 if (length == DMU_OBJECT_END || offset + length > object_size)
633 length = object_size - offset;
634
635 while (length != 0) {
636 uint64_t chunk_end, chunk_begin;
637
638 chunk_end = chunk_begin = offset + length;
639
640 /* move chunk_begin backwards to the beginning of this chunk */
641 err = get_next_chunk(dn, &chunk_begin, offset);
642 if (err)
643 return (err);
644 ASSERT3U(chunk_begin, >=, offset);
645 ASSERT3U(chunk_begin, <=, chunk_end);
646
647 dmu_tx_t *tx = dmu_tx_create(os);
648 dmu_tx_hold_free(tx, dn->dn_object,
649 chunk_begin, chunk_end - chunk_begin);
650 err = dmu_tx_assign(tx, TXG_WAIT);
651 if (err) {
652 dmu_tx_abort(tx);
653 return (err);
654 }
655 dnode_free_range(dn, chunk_begin, chunk_end - chunk_begin, tx);
656 dmu_tx_commit(tx);
657
658 length -= chunk_end - chunk_begin;
659 }
660 return (0);
661 }
662
663 int
664 dmu_free_long_range(objset_t *os, uint64_t object,
665 uint64_t offset, uint64_t length)
666 {
667 dnode_t *dn;
668 int err;
669
670 err = dnode_hold(os, object, FTAG, &dn);
671 if (err != 0)
672 return (err);
673 err = dmu_free_long_range_impl(os, dn, offset, length);
674 dnode_rele(dn, FTAG);
675 return (err);
676 }
677
678 int
679 dmu_free_long_object(objset_t *os, uint64_t object)
680 {
681 dmu_tx_t *tx;
682 int err;
683
684 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
685 if (err != 0)
686 return (err);
687
688 tx = dmu_tx_create(os);
689 dmu_tx_hold_bonus(tx, object);
690 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
691 err = dmu_tx_assign(tx, TXG_WAIT);
692 if (err == 0) {
693 err = dmu_object_free(os, object, tx);
694 dmu_tx_commit(tx);
695 } else {
696 dmu_tx_abort(tx);
697 }
698
699 return (err);
700 }
701
702 int
703 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
704 uint64_t size, dmu_tx_t *tx)
705 {
706 dnode_t *dn;
707 int err = dnode_hold(os, object, FTAG, &dn);
708 if (err)
709 return (err);
710 ASSERT(offset < UINT64_MAX);
711 ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
712 dnode_free_range(dn, offset, size, tx);
713 dnode_rele(dn, FTAG);
714 return (0);
715 }
716
717 int
718 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
719 void *buf, uint32_t flags)
720 {
721 dnode_t *dn;
722 dmu_buf_t **dbp;
723 int numbufs, err;
724
725 err = dnode_hold(os, object, FTAG, &dn);
726 if (err)
727 return (err);
728
729 /*
730 * Deal with odd block sizes, where there can't be data past the first
731 * block. If we ever do the tail block optimization, we will need to
732 * handle that here as well.
733 */
734 if (dn->dn_maxblkid == 0) {
735 int newsz = offset > dn->dn_datablksz ? 0 :
736 MIN(size, dn->dn_datablksz - offset);
737 bzero((char *)buf + newsz, size - newsz);
738 size = newsz;
739 }
740
741 while (size > 0) {
742 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
743 int i;
744
745 /*
746 * NB: we could do this block-at-a-time, but it's nice
747 * to be reading in parallel.
748 */
749 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
750 TRUE, FTAG, &numbufs, &dbp, flags);
751 if (err)
752 break;
753
754 for (i = 0; i < numbufs; i++) {
755 int tocpy;
756 int bufoff;
757 dmu_buf_t *db = dbp[i];
758
759 ASSERT(size > 0);
760
761 bufoff = offset - db->db_offset;
762 tocpy = (int)MIN(db->db_size - bufoff, size);
763
764 bcopy((char *)db->db_data + bufoff, buf, tocpy);
765
766 offset += tocpy;
767 size -= tocpy;
768 buf = (char *)buf + tocpy;
769 }
770 dmu_buf_rele_array(dbp, numbufs, FTAG);
771 }
772 dnode_rele(dn, FTAG);
773 return (err);
774 }
775
776 void
777 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
778 const void *buf, dmu_tx_t *tx)
779 {
780 dmu_buf_t **dbp;
781 int numbufs, i;
782
783 if (size == 0)
784 return;
785
786 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
787 FALSE, FTAG, &numbufs, &dbp));
788
789 for (i = 0; i < numbufs; i++) {
790 int tocpy;
791 int bufoff;
792 dmu_buf_t *db = dbp[i];
793
794 ASSERT(size > 0);
795
796 bufoff = offset - db->db_offset;
797 tocpy = (int)MIN(db->db_size - bufoff, size);
798
799 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
800
801 if (tocpy == db->db_size)
802 dmu_buf_will_fill(db, tx);
803 else
804 dmu_buf_will_dirty(db, tx);
805
806 bcopy(buf, (char *)db->db_data + bufoff, tocpy);
807
808 if (tocpy == db->db_size)
809 dmu_buf_fill_done(db, tx);
810
811 offset += tocpy;
812 size -= tocpy;
813 buf = (char *)buf + tocpy;
814 }
815 dmu_buf_rele_array(dbp, numbufs, FTAG);
816 }
817
818 void
819 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
820 dmu_tx_t *tx)
821 {
822 dmu_buf_t **dbp;
823 int numbufs, i;
824
825 if (size == 0)
826 return;
827
828 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
829 FALSE, FTAG, &numbufs, &dbp));
830
831 for (i = 0; i < numbufs; i++) {
832 dmu_buf_t *db = dbp[i];
833
834 dmu_buf_will_not_fill(db, tx);
835 }
836 dmu_buf_rele_array(dbp, numbufs, FTAG);
837 }
838
839 /*
840 * DMU support for xuio
841 */
842 kstat_t *xuio_ksp = NULL;
843
844 int
845 dmu_xuio_init(xuio_t *xuio, int nblk)
846 {
847 dmu_xuio_t *priv;
848 uio_t *uio = &xuio->xu_uio;
849
850 uio->uio_iovcnt = nblk;
851 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
852
853 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
854 priv->cnt = nblk;
855 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
856 priv->iovp = uio->uio_iov;
857 XUIO_XUZC_PRIV(xuio) = priv;
858
859 if (XUIO_XUZC_RW(xuio) == UIO_READ)
860 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
861 else
862 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
863
864 return (0);
865 }
866
867 void
868 dmu_xuio_fini(xuio_t *xuio)
869 {
870 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
871 int nblk = priv->cnt;
872
873 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
874 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
875 kmem_free(priv, sizeof (dmu_xuio_t));
876
877 if (XUIO_XUZC_RW(xuio) == UIO_READ)
878 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
879 else
880 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
881 }
882
883 /*
884 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
885 * and increase priv->next by 1.
886 */
887 int
888 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
889 {
890 struct iovec *iov;
891 uio_t *uio = &xuio->xu_uio;
892 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
893 int i = priv->next++;
894
895 ASSERT(i < priv->cnt);
896 ASSERT(off + n <= arc_buf_size(abuf));
897 iov = uio->uio_iov + i;
898 iov->iov_base = (char *)abuf->b_data + off;
899 iov->iov_len = n;
900 priv->bufs[i] = abuf;
901 return (0);
902 }
903
904 int
905 dmu_xuio_cnt(xuio_t *xuio)
906 {
907 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
908 return (priv->cnt);
909 }
910
911 arc_buf_t *
912 dmu_xuio_arcbuf(xuio_t *xuio, int i)
913 {
914 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
915
916 ASSERT(i < priv->cnt);
917 return (priv->bufs[i]);
918 }
919
920 void
921 dmu_xuio_clear(xuio_t *xuio, int i)
922 {
923 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
924
925 ASSERT(i < priv->cnt);
926 priv->bufs[i] = NULL;
927 }
928
929 static void
930 xuio_stat_init(void)
931 {
932 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
933 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
934 KSTAT_FLAG_VIRTUAL);
935 if (xuio_ksp != NULL) {
936 xuio_ksp->ks_data = &xuio_stats;
937 kstat_install(xuio_ksp);
938 }
939 }
940
941 static void
942 xuio_stat_fini(void)
943 {
944 if (xuio_ksp != NULL) {
945 kstat_delete(xuio_ksp);
946 xuio_ksp = NULL;
947 }
948 }
949
950 void
951 xuio_stat_wbuf_copied()
952 {
953 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
954 }
955
956 void
957 xuio_stat_wbuf_nocopy()
958 {
959 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
960 }
961
962 #ifdef _KERNEL
963 int
964 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
965 {
966 dmu_buf_t **dbp;
967 int numbufs, i, err;
968 xuio_t *xuio = NULL;
969
970 /*
971 * NB: we could do this block-at-a-time, but it's nice
972 * to be reading in parallel.
973 */
974 err = dmu_buf_hold_array(os, object, uio->uio_loffset, size, TRUE, FTAG,
975 &numbufs, &dbp);
976 if (err)
977 return (err);
978
979 if (uio->uio_extflg == UIO_XUIO)
980 xuio = (xuio_t *)uio;
981
982 for (i = 0; i < numbufs; i++) {
983 int tocpy;
984 int bufoff;
985 dmu_buf_t *db = dbp[i];
986
987 ASSERT(size > 0);
988
989 bufoff = uio->uio_loffset - db->db_offset;
990 tocpy = (int)MIN(db->db_size - bufoff, size);
991
992 if (xuio) {
993 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
994 arc_buf_t *dbuf_abuf = dbi->db_buf;
995 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
996 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
997 if (!err) {
998 uio->uio_resid -= tocpy;
999 uio->uio_loffset += tocpy;
1000 }
1001
1002 if (abuf == dbuf_abuf)
1003 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
1004 else
1005 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
1006 } else {
1007 err = uiomove((char *)db->db_data + bufoff, tocpy,
1008 UIO_READ, uio);
1009 }
1010 if (err)
1011 break;
1012
1013 size -= tocpy;
1014 }
1015 dmu_buf_rele_array(dbp, numbufs, FTAG);
1016
1017 return (err);
1018 }
1019
1020 static int
1021 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1022 {
1023 dmu_buf_t **dbp;
1024 int numbufs;
1025 int err = 0;
1026 int i;
1027
1028 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1029 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1030 if (err)
1031 return (err);
1032
1033 for (i = 0; i < numbufs; i++) {
1034 int tocpy;
1035 int bufoff;
1036 dmu_buf_t *db = dbp[i];
1037
1038 ASSERT(size > 0);
1039
1040 bufoff = uio->uio_loffset - db->db_offset;
1041 tocpy = (int)MIN(db->db_size - bufoff, size);
1042
1043 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1044
1045 if (tocpy == db->db_size)
1046 dmu_buf_will_fill(db, tx);
1047 else
1048 dmu_buf_will_dirty(db, tx);
1049
1050 /*
1051 * XXX uiomove could block forever (eg. nfs-backed
1052 * pages). There needs to be a uiolockdown() function
1053 * to lock the pages in memory, so that uiomove won't
1054 * block.
1055 */
1056 err = uiomove((char *)db->db_data + bufoff, tocpy,
1057 UIO_WRITE, uio);
1058
1059 if (tocpy == db->db_size)
1060 dmu_buf_fill_done(db, tx);
1061
1062 if (err)
1063 break;
1064
1065 size -= tocpy;
1066 }
1067
1068 dmu_buf_rele_array(dbp, numbufs, FTAG);
1069 return (err);
1070 }
1071
1072 int
1073 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1074 dmu_tx_t *tx)
1075 {
1076 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1077 dnode_t *dn;
1078 int err;
1079
1080 if (size == 0)
1081 return (0);
1082
1083 DB_DNODE_ENTER(db);
1084 dn = DB_DNODE(db);
1085 err = dmu_write_uio_dnode(dn, uio, size, tx);
1086 DB_DNODE_EXIT(db);
1087
1088 return (err);
1089 }
1090
1091 int
1092 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1093 dmu_tx_t *tx)
1094 {
1095 dnode_t *dn;
1096 int err;
1097
1098 if (size == 0)
1099 return (0);
1100
1101 err = dnode_hold(os, object, FTAG, &dn);
1102 if (err)
1103 return (err);
1104
1105 err = dmu_write_uio_dnode(dn, uio, size, tx);
1106
1107 dnode_rele(dn, FTAG);
1108
1109 return (err);
1110 }
1111
1112 int
1113 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1114 page_t *pp, dmu_tx_t *tx)
1115 {
1116 dmu_buf_t **dbp;
1117 int numbufs, i;
1118 int err;
1119
1120 if (size == 0)
1121 return (0);
1122
1123 err = dmu_buf_hold_array(os, object, offset, size,
1124 FALSE, FTAG, &numbufs, &dbp);
1125 if (err)
1126 return (err);
1127
1128 for (i = 0; i < numbufs; i++) {
1129 int tocpy, copied, thiscpy;
1130 int bufoff;
1131 dmu_buf_t *db = dbp[i];
1132 caddr_t va;
1133
1134 ASSERT(size > 0);
1135 ASSERT3U(db->db_size, >=, PAGESIZE);
1136
1137 bufoff = offset - db->db_offset;
1138 tocpy = (int)MIN(db->db_size - bufoff, size);
1139
1140 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1141
1142 if (tocpy == db->db_size)
1143 dmu_buf_will_fill(db, tx);
1144 else
1145 dmu_buf_will_dirty(db, tx);
1146
1147 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1148 ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff);
1149 thiscpy = MIN(PAGESIZE, tocpy - copied);
1150 va = zfs_map_page(pp, S_READ);
1151 bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1152 zfs_unmap_page(pp, va);
1153 pp = pp->p_next;
1154 bufoff += PAGESIZE;
1155 }
1156
1157 if (tocpy == db->db_size)
1158 dmu_buf_fill_done(db, tx);
1159
1160 offset += tocpy;
1161 size -= tocpy;
1162 }
1163 dmu_buf_rele_array(dbp, numbufs, FTAG);
1164 return (err);
1165 }
1166 #endif
1167
1168 /*
1169 * Allocate a loaned anonymous arc buffer.
1170 */
1171 arc_buf_t *
1172 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1173 {
1174 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1175 spa_t *spa;
1176
1177 DB_GET_SPA(&spa, db);
1178 return (arc_loan_buf(spa, size));
1179 }
1180
1181 /*
1182 * Free a loaned arc buffer.
1183 */
1184 void
1185 dmu_return_arcbuf(arc_buf_t *buf)
1186 {
1187 arc_return_buf(buf, FTAG);
1188 VERIFY(arc_buf_remove_ref(buf, FTAG));
1189 }
1190
1191 /*
1192 * When possible directly assign passed loaned arc buffer to a dbuf.
1193 * If this is not possible copy the contents of passed arc buf via
1194 * dmu_write().
1195 */
1196 void
1197 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1198 dmu_tx_t *tx)
1199 {
1200 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1201 dnode_t *dn;
1202 dmu_buf_impl_t *db;
1203 uint32_t blksz = (uint32_t)arc_buf_size(buf);
1204 uint64_t blkid;
1205
1206 DB_DNODE_ENTER(dbuf);
1207 dn = DB_DNODE(dbuf);
1208 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1209 blkid = dbuf_whichblock(dn, offset);
1210 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1211 rw_exit(&dn->dn_struct_rwlock);
1212 DB_DNODE_EXIT(dbuf);
1213
1214 if (offset == db->db.db_offset && blksz == db->db.db_size) {
1215 dbuf_assign_arcbuf(db, buf, tx);
1216 dbuf_rele(db, FTAG);
1217 } else {
1218 objset_t *os;
1219 uint64_t object;
1220
1221 DB_DNODE_ENTER(dbuf);
1222 dn = DB_DNODE(dbuf);
1223 os = dn->dn_objset;
1224 object = dn->dn_object;
1225 DB_DNODE_EXIT(dbuf);
1226
1227 dbuf_rele(db, FTAG);
1228 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1229 dmu_return_arcbuf(buf);
1230 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1231 }
1232 }
1233
1234 typedef struct {
1235 dbuf_dirty_record_t *dsa_dr;
1236 dmu_sync_cb_t *dsa_done;
1237 zgd_t *dsa_zgd;
1238 dmu_tx_t *dsa_tx;
1239 } dmu_sync_arg_t;
1240
1241 /* ARGSUSED */
1242 static void
1243 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1244 {
1245 dmu_sync_arg_t *dsa = varg;
1246 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1247 blkptr_t *bp = zio->io_bp;
1248
1249 if (zio->io_error == 0) {
1250 if (BP_IS_HOLE(bp)) {
1251 /*
1252 * A block of zeros may compress to a hole, but the
1253 * block size still needs to be known for replay.
1254 */
1255 BP_SET_LSIZE(bp, db->db_size);
1256 } else {
1257 ASSERT(BP_GET_LEVEL(bp) == 0);
1258 bp->blk_fill = 1;
1259 }
1260 }
1261 }
1262
1263 static void
1264 dmu_sync_late_arrival_ready(zio_t *zio)
1265 {
1266 dmu_sync_ready(zio, NULL, zio->io_private);
1267 }
1268
1269 /* ARGSUSED */
1270 static void
1271 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1272 {
1273 dmu_sync_arg_t *dsa = varg;
1274 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1275 dmu_buf_impl_t *db = dr->dr_dbuf;
1276
1277 mutex_enter(&db->db_mtx);
1278 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1279 if (zio->io_error == 0) {
1280 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1281 if (dr->dt.dl.dr_nopwrite) {
1282 blkptr_t *bp = zio->io_bp;
1283 blkptr_t *bp_orig = &zio->io_bp_orig;
1284 uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1285
1286 ASSERT(BP_EQUAL(bp, bp_orig));
1287 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1288 ASSERT(zio_checksum_table[chksum].ci_dedup);
1289 }
1290 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1291 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1292 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1293 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by))
1294 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1295 } else {
1296 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1297 }
1298 cv_broadcast(&db->db_changed);
1299 mutex_exit(&db->db_mtx);
1300
1301 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1302
1303 kmem_free(dsa, sizeof (*dsa));
1304 }
1305
1306 static void
1307 dmu_sync_late_arrival_done(zio_t *zio)
1308 {
1309 blkptr_t *bp = zio->io_bp;
1310 dmu_sync_arg_t *dsa = zio->io_private;
1311 blkptr_t *bp_orig = &zio->io_bp_orig;
1312
1313 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1314 /*
1315 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE)
1316 * then there is nothing to do here. Otherwise, free the
1317 * newly allocated block in this txg.
1318 */
1319 if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
1320 ASSERT(BP_EQUAL(bp, bp_orig));
1321 } else {
1322 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1323 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1324 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1325 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1326 }
1327 }
1328
1329 dmu_tx_commit(dsa->dsa_tx);
1330
1331 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1332
1333 kmem_free(dsa, sizeof (*dsa));
1334 }
1335
1336 static int
1337 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1338 zio_prop_t *zp, zbookmark_t *zb)
1339 {
1340 dmu_sync_arg_t *dsa;
1341 dmu_tx_t *tx;
1342
1343 tx = dmu_tx_create(os);
1344 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1345 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1346 dmu_tx_abort(tx);
1347 /* Make zl_get_data do txg_waited_synced() */
1348 return (SET_ERROR(EIO));
1349 }
1350
1351 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1352 dsa->dsa_dr = NULL;
1353 dsa->dsa_done = done;
1354 dsa->dsa_zgd = zgd;
1355 dsa->dsa_tx = tx;
1356
1357 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1358 zgd->zgd_db->db_data, zgd->zgd_db->db_size, zp,
1359 dmu_sync_late_arrival_ready, dmu_sync_late_arrival_done, dsa,
1360 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1361
1362 return (0);
1363 }
1364
1365 /*
1366 * Intent log support: sync the block associated with db to disk.
1367 * N.B. and XXX: the caller is responsible for making sure that the
1368 * data isn't changing while dmu_sync() is writing it.
1369 *
1370 * Return values:
1371 *
1372 * EEXIST: this txg has already been synced, so there's nothing to do.
1373 * The caller should not log the write.
1374 *
1375 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1376 * The caller should not log the write.
1377 *
1378 * EALREADY: this block is already in the process of being synced.
1379 * The caller should track its progress (somehow).
1380 *
1381 * EIO: could not do the I/O.
1382 * The caller should do a txg_wait_synced().
1383 *
1384 * 0: the I/O has been initiated.
1385 * The caller should log this blkptr in the done callback.
1386 * It is possible that the I/O will fail, in which case
1387 * the error will be reported to the done callback and
1388 * propagated to pio from zio_done().
1389 */
1390 int
1391 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1392 {
1393 blkptr_t *bp = zgd->zgd_bp;
1394 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1395 objset_t *os = db->db_objset;
1396 dsl_dataset_t *ds = os->os_dsl_dataset;
1397 dbuf_dirty_record_t *dr;
1398 dmu_sync_arg_t *dsa;
1399 zbookmark_t zb;
1400 zio_prop_t zp;
1401 dnode_t *dn;
1402
1403 ASSERT(pio != NULL);
1404 ASSERT(txg != 0);
1405
1406 SET_BOOKMARK(&zb, ds->ds_object,
1407 db->db.db_object, db->db_level, db->db_blkid);
1408
1409 DB_DNODE_ENTER(db);
1410 dn = DB_DNODE(db);
1411 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1412 DB_DNODE_EXIT(db);
1413
1414 /*
1415 * If we're frozen (running ziltest), we always need to generate a bp.
1416 */
1417 if (txg > spa_freeze_txg(os->os_spa))
1418 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1419
1420 /*
1421 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1422 * and us. If we determine that this txg is not yet syncing,
1423 * but it begins to sync a moment later, that's OK because the
1424 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1425 */
1426 mutex_enter(&db->db_mtx);
1427
1428 if (txg <= spa_last_synced_txg(os->os_spa)) {
1429 /*
1430 * This txg has already synced. There's nothing to do.
1431 */
1432 mutex_exit(&db->db_mtx);
1433 return (SET_ERROR(EEXIST));
1434 }
1435
1436 if (txg <= spa_syncing_txg(os->os_spa)) {
1437 /*
1438 * This txg is currently syncing, so we can't mess with
1439 * the dirty record anymore; just write a new log block.
1440 */
1441 mutex_exit(&db->db_mtx);
1442 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1443 }
1444
1445 dr = db->db_last_dirty;
1446 while (dr && dr->dr_txg != txg)
1447 dr = dr->dr_next;
1448
1449 if (dr == NULL) {
1450 /*
1451 * There's no dr for this dbuf, so it must have been freed.
1452 * There's no need to log writes to freed blocks, so we're done.
1453 */
1454 mutex_exit(&db->db_mtx);
1455 return (SET_ERROR(ENOENT));
1456 }
1457
1458 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
1459
1460 /*
1461 * Assume the on-disk data is X, the current syncing data is Y,
1462 * and the current in-memory data is Z (currently in dmu_sync).
1463 * X and Z are identical but Y is has been modified. Normally,
1464 * when X and Z are the same we will perform a nopwrite but if Y
1465 * is different we must disable nopwrite since the resulting write
1466 * of Y to disk can free the block containing X. If we allowed a
1467 * nopwrite to occur the block pointing to Z would reference a freed
1468 * block. Since this is a rare case we simplify this by disabling
1469 * nopwrite if the current dmu_sync-ing dbuf has been modified in
1470 * a previous transaction.
1471 */
1472 if (dr->dr_next)
1473 zp.zp_nopwrite = B_FALSE;
1474
1475 ASSERT(dr->dr_txg == txg);
1476 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1477 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1478 /*
1479 * We have already issued a sync write for this buffer,
1480 * or this buffer has already been synced. It could not
1481 * have been dirtied since, or we would have cleared the state.
1482 */
1483 mutex_exit(&db->db_mtx);
1484 return (SET_ERROR(EALREADY));
1485 }
1486
1487 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1488 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1489 mutex_exit(&db->db_mtx);
1490
1491 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1492 dsa->dsa_dr = dr;
1493 dsa->dsa_done = done;
1494 dsa->dsa_zgd = zgd;
1495 dsa->dsa_tx = NULL;
1496
1497 zio_nowait(arc_write(pio, os->os_spa, txg,
1498 bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
1499 DBUF_IS_L2COMPRESSIBLE(db), &zp, dmu_sync_ready, dmu_sync_done,
1500 dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb));
1501
1502 return (0);
1503 }
1504
1505 int
1506 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1507 dmu_tx_t *tx)
1508 {
1509 dnode_t *dn;
1510 int err;
1511
1512 err = dnode_hold(os, object, FTAG, &dn);
1513 if (err)
1514 return (err);
1515 err = dnode_set_blksz(dn, size, ibs, tx);
1516 dnode_rele(dn, FTAG);
1517 return (err);
1518 }
1519
1520 void
1521 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1522 dmu_tx_t *tx)
1523 {
1524 dnode_t *dn;
1525
1526 /* XXX assumes dnode_hold will not get an i/o error */
1527 (void) dnode_hold(os, object, FTAG, &dn);
1528 ASSERT(checksum < ZIO_CHECKSUM_FUNCTIONS);
1529 dn->dn_checksum = checksum;
1530 dnode_setdirty(dn, tx);
1531 dnode_rele(dn, FTAG);
1532 }
1533
1534 void
1535 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1536 dmu_tx_t *tx)
1537 {
1538 dnode_t *dn;
1539
1540 /* XXX assumes dnode_hold will not get an i/o error */
1541 (void) dnode_hold(os, object, FTAG, &dn);
1542 ASSERT(compress < ZIO_COMPRESS_FUNCTIONS);
1543 dn->dn_compress = compress;
1544 dnode_setdirty(dn, tx);
1545 dnode_rele(dn, FTAG);
1546 }
1547
1548 int zfs_mdcomp_disable = 0;
1549
1550 void
1551 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
1552 {
1553 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
1554 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
1555 (wp & WP_SPILL));
1556 enum zio_checksum checksum = os->os_checksum;
1557 enum zio_compress compress = os->os_compress;
1558 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
1559 boolean_t dedup = B_FALSE;
1560 boolean_t nopwrite = B_FALSE;
1561 boolean_t dedup_verify = os->os_dedup_verify;
1562 int copies = os->os_copies;
1563
1564 /*
1565 * We maintain different write policies for each of the following
1566 * types of data:
1567 * 1. metadata
1568 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1569 * 3. all other level 0 blocks
1570 */
1571 if (ismd) {
1572 /*
1573 * XXX -- we should design a compression algorithm
1574 * that specializes in arrays of bps.
1575 */
1576 compress = zfs_mdcomp_disable ? ZIO_COMPRESS_EMPTY :
1577 ZIO_COMPRESS_LZJB;
1578
1579 /*
1580 * Metadata always gets checksummed. If the data
1581 * checksum is multi-bit correctable, and it's not a
1582 * ZBT-style checksum, then it's suitable for metadata
1583 * as well. Otherwise, the metadata checksum defaults
1584 * to fletcher4.
1585 */
1586 if (zio_checksum_table[checksum].ci_correctable < 1 ||
1587 zio_checksum_table[checksum].ci_eck)
1588 checksum = ZIO_CHECKSUM_FLETCHER_4;
1589 } else if (wp & WP_NOFILL) {
1590 ASSERT(level == 0);
1591
1592 /*
1593 * If we're writing preallocated blocks, we aren't actually
1594 * writing them so don't set any policy properties. These
1595 * blocks are currently only used by an external subsystem
1596 * outside of zfs (i.e. dump) and not written by the zio
1597 * pipeline.
1598 */
1599 compress = ZIO_COMPRESS_OFF;
1600 checksum = ZIO_CHECKSUM_OFF;
1601 } else {
1602 compress = zio_compress_select(dn->dn_compress, compress);
1603
1604 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
1605 zio_checksum_select(dn->dn_checksum, checksum) :
1606 dedup_checksum;
1607
1608 /*
1609 * Determine dedup setting. If we are in dmu_sync(),
1610 * we won't actually dedup now because that's all
1611 * done in syncing context; but we do want to use the
1612 * dedup checkum. If the checksum is not strong
1613 * enough to ensure unique signatures, force
1614 * dedup_verify.
1615 */
1616 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
1617 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
1618 if (!zio_checksum_table[checksum].ci_dedup)
1619 dedup_verify = B_TRUE;
1620 }
1621
1622 /*
1623 * Enable nopwrite if we have a cryptographically secure
1624 * checksum that has no known collisions (i.e. SHA-256)
1625 * and compression is enabled. We don't enable nopwrite if
1626 * dedup is enabled as the two features are mutually exclusive.
1627 */
1628 nopwrite = (!dedup && zio_checksum_table[checksum].ci_dedup &&
1629 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
1630 }
1631
1632 zp->zp_checksum = checksum;
1633 zp->zp_compress = compress;
1634 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
1635 zp->zp_level = level;
1636 zp->zp_copies = MIN(copies + ismd, spa_max_replication(os->os_spa));
1637 zp->zp_dedup = dedup;
1638 zp->zp_dedup_verify = dedup && dedup_verify;
1639 zp->zp_nopwrite = nopwrite;
1640 }
1641
1642 int
1643 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
1644 {
1645 dnode_t *dn;
1646 int i, err;
1647
1648 err = dnode_hold(os, object, FTAG, &dn);
1649 if (err)
1650 return (err);
1651 /*
1652 * Sync any current changes before
1653 * we go trundling through the block pointers.
1654 */
1655 for (i = 0; i < TXG_SIZE; i++) {
1656 if (list_link_active(&dn->dn_dirty_link[i]))
1657 break;
1658 }
1659 if (i != TXG_SIZE) {
1660 dnode_rele(dn, FTAG);
1661 txg_wait_synced(dmu_objset_pool(os), 0);
1662 err = dnode_hold(os, object, FTAG, &dn);
1663 if (err)
1664 return (err);
1665 }
1666
1667 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
1668 dnode_rele(dn, FTAG);
1669
1670 return (err);
1671 }
1672
1673 void
1674 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
1675 {
1676 dnode_phys_t *dnp;
1677
1678 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1679 mutex_enter(&dn->dn_mtx);
1680
1681 dnp = dn->dn_phys;
1682
1683 doi->doi_data_block_size = dn->dn_datablksz;
1684 doi->doi_metadata_block_size = dn->dn_indblkshift ?
1685 1ULL << dn->dn_indblkshift : 0;
1686 doi->doi_type = dn->dn_type;
1687 doi->doi_bonus_type = dn->dn_bonustype;
1688 doi->doi_bonus_size = dn->dn_bonuslen;
1689 doi->doi_indirection = dn->dn_nlevels;
1690 doi->doi_checksum = dn->dn_checksum;
1691 doi->doi_compress = dn->dn_compress;
1692 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
1693 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
1694 doi->doi_fill_count = 0;
1695 for (int i = 0; i < dnp->dn_nblkptr; i++)
1696 doi->doi_fill_count += dnp->dn_blkptr[i].blk_fill;
1697
1698 mutex_exit(&dn->dn_mtx);
1699 rw_exit(&dn->dn_struct_rwlock);
1700 }
1701
1702 /*
1703 * Get information on a DMU object.
1704 * If doi is NULL, just indicates whether the object exists.
1705 */
1706 int
1707 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
1708 {
1709 dnode_t *dn;
1710 int err = dnode_hold(os, object, FTAG, &dn);
1711
1712 if (err)
1713 return (err);
1714
1715 if (doi != NULL)
1716 dmu_object_info_from_dnode(dn, doi);
1717
1718 dnode_rele(dn, FTAG);
1719 return (0);
1720 }
1721
1722 /*
1723 * As above, but faster; can be used when you have a held dbuf in hand.
1724 */
1725 void
1726 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
1727 {
1728 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1729
1730 DB_DNODE_ENTER(db);
1731 dmu_object_info_from_dnode(DB_DNODE(db), doi);
1732 DB_DNODE_EXIT(db);
1733 }
1734
1735 /*
1736 * Faster still when you only care about the size.
1737 * This is specifically optimized for zfs_getattr().
1738 */
1739 void
1740 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
1741 u_longlong_t *nblk512)
1742 {
1743 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1744 dnode_t *dn;
1745
1746 DB_DNODE_ENTER(db);
1747 dn = DB_DNODE(db);
1748
1749 *blksize = dn->dn_datablksz;
1750 /* add 1 for dnode space */
1751 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
1752 SPA_MINBLOCKSHIFT) + 1;
1753 DB_DNODE_EXIT(db);
1754 }
1755
1756 void
1757 byteswap_uint64_array(void *vbuf, size_t size)
1758 {
1759 uint64_t *buf = vbuf;
1760 size_t count = size >> 3;
1761 int i;
1762
1763 ASSERT((size & 7) == 0);
1764
1765 for (i = 0; i < count; i++)
1766 buf[i] = BSWAP_64(buf[i]);
1767 }
1768
1769 void
1770 byteswap_uint32_array(void *vbuf, size_t size)
1771 {
1772 uint32_t *buf = vbuf;
1773 size_t count = size >> 2;
1774 int i;
1775
1776 ASSERT((size & 3) == 0);
1777
1778 for (i = 0; i < count; i++)
1779 buf[i] = BSWAP_32(buf[i]);
1780 }
1781
1782 void
1783 byteswap_uint16_array(void *vbuf, size_t size)
1784 {
1785 uint16_t *buf = vbuf;
1786 size_t count = size >> 1;
1787 int i;
1788
1789 ASSERT((size & 1) == 0);
1790
1791 for (i = 0; i < count; i++)
1792 buf[i] = BSWAP_16(buf[i]);
1793 }
1794
1795 /* ARGSUSED */
1796 void
1797 byteswap_uint8_array(void *vbuf, size_t size)
1798 {
1799 }
1800
1801 void
1802 dmu_init(void)
1803 {
1804 zfs_dbgmsg_init();
1805 sa_cache_init();
1806 xuio_stat_init();
1807 dmu_objset_init();
1808 dnode_init();
1809 dbuf_init();
1810 zfetch_init();
1811 l2arc_init();
1812 arc_init();
1813 }
1814
1815 void
1816 dmu_fini(void)
1817 {
1818 arc_fini(); /* arc depends on l2arc, so arc must go first */
1819 l2arc_fini();
1820 zfetch_fini();
1821 dbuf_fini();
1822 dnode_fini();
1823 dmu_objset_fini();
1824 xuio_stat_fini();
1825 sa_cache_fini();
1826 zfs_dbgmsg_fini();
1827 }