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 static int
574 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t limit)
575 {
576 uint64_t len = *start - limit;
577 uint64_t blkcnt = 0;
578 uint64_t maxblks = DMU_MAX_ACCESS / (1ULL << (dn->dn_indblkshift + 1));
579 uint64_t iblkrange =
580 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
581
582 ASSERT(limit <= *start);
583
584 if (len <= iblkrange * maxblks) {
585 *start = limit;
586 return (0);
587 }
588 ASSERT(ISP2(iblkrange));
589
590 while (*start > limit && blkcnt < maxblks) {
591 int err;
592
593 /* find next allocated L1 indirect */
594 err = dnode_next_offset(dn,
595 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
596
597 /* if there are no more, then we are done */
598 if (err == ESRCH) {
599 *start = limit;
600 return (0);
601 } else if (err) {
602 return (err);
603 }
604 blkcnt += 1;
605
606 /* reset offset to end of "next" block back */
607 *start = P2ALIGN(*start, iblkrange);
608 if (*start <= limit)
609 *start = limit;
610 else
611 *start -= 1;
612 }
613 return (0);
614 }
615
616 static int
617 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
618 uint64_t length, boolean_t free_dnode)
619 {
620 dmu_tx_t *tx;
621 uint64_t object_size, start, end, len;
622 boolean_t trunc = (length == DMU_OBJECT_END);
623 int align, err;
624
625 align = 1 << dn->dn_datablkshift;
626 ASSERT(align > 0);
627 object_size = align == 1 ? dn->dn_datablksz :
628 (dn->dn_maxblkid + 1) << dn->dn_datablkshift;
629
630 end = offset + length;
631 if (trunc || end > object_size)
632 end = object_size;
633 if (end <= offset)
634 return (0);
635 length = end - offset;
636
637 while (length) {
638 start = end;
639 /* assert(offset <= start) */
640 err = get_next_chunk(dn, &start, offset);
641 if (err)
642 return (err);
643 len = trunc ? DMU_OBJECT_END : end - start;
644
645 tx = dmu_tx_create(os);
646 dmu_tx_hold_free(tx, dn->dn_object, start, len);
647 err = dmu_tx_assign(tx, TXG_WAIT);
648 if (err) {
649 dmu_tx_abort(tx);
650 return (err);
651 }
652
653 dnode_free_range(dn, start, trunc ? -1 : len, tx);
654
655 if (start == 0 && free_dnode) {
656 ASSERT(trunc);
657 dnode_free(dn, tx);
658 }
659
660 length -= end - start;
661
662 dmu_tx_commit(tx);
663 end = start;
664 }
665 return (0);
666 }
667
668 int
669 dmu_free_long_range(objset_t *os, uint64_t object,
670 uint64_t offset, uint64_t length)
671 {
672 dnode_t *dn;
673 int err;
674
675 err = dnode_hold(os, object, FTAG, &dn);
676 if (err != 0)
677 return (err);
678 err = dmu_free_long_range_impl(os, dn, offset, length, FALSE);
679 dnode_rele(dn, FTAG);
680 return (err);
681 }
682
683 int
684 dmu_free_object(objset_t *os, uint64_t object)
685 {
686 dnode_t *dn;
687 dmu_tx_t *tx;
688 int err;
689
690 err = dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED,
691 FTAG, &dn);
692 if (err != 0)
693 return (err);
694 if (dn->dn_nlevels == 1) {
695 tx = dmu_tx_create(os);
696 dmu_tx_hold_bonus(tx, object);
697 dmu_tx_hold_free(tx, dn->dn_object, 0, DMU_OBJECT_END);
698 err = dmu_tx_assign(tx, TXG_WAIT);
699 if (err == 0) {
700 dnode_free_range(dn, 0, DMU_OBJECT_END, tx);
701 dnode_free(dn, tx);
702 dmu_tx_commit(tx);
703 } else {
704 dmu_tx_abort(tx);
705 }
706 } else {
707 err = dmu_free_long_range_impl(os, dn, 0, DMU_OBJECT_END, TRUE);
708 }
709 dnode_rele(dn, FTAG);
710 return (err);
711 }
712
713 int
714 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
715 uint64_t size, dmu_tx_t *tx)
716 {
717 dnode_t *dn;
718 int err = dnode_hold(os, object, FTAG, &dn);
719 if (err)
720 return (err);
721 ASSERT(offset < UINT64_MAX);
722 ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
723 dnode_free_range(dn, offset, size, tx);
724 dnode_rele(dn, FTAG);
725 return (0);
726 }
727
728 int
729 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
730 void *buf, uint32_t flags)
731 {
732 dnode_t *dn;
733 dmu_buf_t **dbp;
734 int numbufs, err;
735
736 err = dnode_hold(os, object, FTAG, &dn);
737 if (err)
738 return (err);
739
740 /*
741 * Deal with odd block sizes, where there can't be data past the first
742 * block. If we ever do the tail block optimization, we will need to
743 * handle that here as well.
744 */
745 if (dn->dn_maxblkid == 0) {
746 int newsz = offset > dn->dn_datablksz ? 0 :
747 MIN(size, dn->dn_datablksz - offset);
748 bzero((char *)buf + newsz, size - newsz);
749 size = newsz;
750 }
751
752 while (size > 0) {
753 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
754 int i;
755
756 /*
757 * NB: we could do this block-at-a-time, but it's nice
758 * to be reading in parallel.
759 */
760 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
761 TRUE, FTAG, &numbufs, &dbp, flags);
762 if (err)
763 break;
764
765 for (i = 0; i < numbufs; i++) {
766 int tocpy;
767 int bufoff;
768 dmu_buf_t *db = dbp[i];
769
770 ASSERT(size > 0);
771
772 bufoff = offset - db->db_offset;
773 tocpy = (int)MIN(db->db_size - bufoff, size);
774
775 bcopy((char *)db->db_data + bufoff, buf, tocpy);
776
777 offset += tocpy;
778 size -= tocpy;
779 buf = (char *)buf + tocpy;
780 }
781 dmu_buf_rele_array(dbp, numbufs, FTAG);
782 }
783 dnode_rele(dn, FTAG);
784 return (err);
785 }
786
787 void
788 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
789 const void *buf, dmu_tx_t *tx)
790 {
791 dmu_buf_t **dbp;
792 int numbufs, i;
793
794 if (size == 0)
795 return;
796
797 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
798 FALSE, FTAG, &numbufs, &dbp));
799
800 for (i = 0; i < numbufs; i++) {
801 int tocpy;
802 int bufoff;
803 dmu_buf_t *db = dbp[i];
804
805 ASSERT(size > 0);
806
807 bufoff = offset - db->db_offset;
808 tocpy = (int)MIN(db->db_size - bufoff, size);
809
810 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
811
812 if (tocpy == db->db_size)
813 dmu_buf_will_fill(db, tx);
814 else
815 dmu_buf_will_dirty(db, tx);
816
817 bcopy(buf, (char *)db->db_data + bufoff, tocpy);
818
819 if (tocpy == db->db_size)
820 dmu_buf_fill_done(db, tx);
821
822 offset += tocpy;
823 size -= tocpy;
824 buf = (char *)buf + tocpy;
825 }
826 dmu_buf_rele_array(dbp, numbufs, FTAG);
827 }
828
829 void
830 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
831 dmu_tx_t *tx)
832 {
833 dmu_buf_t **dbp;
834 int numbufs, i;
835
836 if (size == 0)
837 return;
838
839 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
840 FALSE, FTAG, &numbufs, &dbp));
841
842 for (i = 0; i < numbufs; i++) {
843 dmu_buf_t *db = dbp[i];
844
845 dmu_buf_will_not_fill(db, tx);
846 }
847 dmu_buf_rele_array(dbp, numbufs, FTAG);
848 }
849
850 /*
851 * DMU support for xuio
852 */
853 kstat_t *xuio_ksp = NULL;
854
855 int
856 dmu_xuio_init(xuio_t *xuio, int nblk)
857 {
858 dmu_xuio_t *priv;
859 uio_t *uio = &xuio->xu_uio;
860
861 uio->uio_iovcnt = nblk;
862 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
863
864 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
865 priv->cnt = nblk;
866 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
867 priv->iovp = uio->uio_iov;
868 XUIO_XUZC_PRIV(xuio) = priv;
869
870 if (XUIO_XUZC_RW(xuio) == UIO_READ)
871 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
872 else
873 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
874
875 return (0);
876 }
877
878 void
879 dmu_xuio_fini(xuio_t *xuio)
880 {
881 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
882 int nblk = priv->cnt;
883
884 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
885 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
886 kmem_free(priv, sizeof (dmu_xuio_t));
887
888 if (XUIO_XUZC_RW(xuio) == UIO_READ)
889 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
890 else
891 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
892 }
893
894 /*
895 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
896 * and increase priv->next by 1.
897 */
898 int
899 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
900 {
901 struct iovec *iov;
902 uio_t *uio = &xuio->xu_uio;
903 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
904 int i = priv->next++;
905
906 ASSERT(i < priv->cnt);
907 ASSERT(off + n <= arc_buf_size(abuf));
908 iov = uio->uio_iov + i;
909 iov->iov_base = (char *)abuf->b_data + off;
910 iov->iov_len = n;
911 priv->bufs[i] = abuf;
912 return (0);
913 }
914
915 int
916 dmu_xuio_cnt(xuio_t *xuio)
917 {
918 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
919 return (priv->cnt);
920 }
921
922 arc_buf_t *
923 dmu_xuio_arcbuf(xuio_t *xuio, int i)
924 {
925 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
926
927 ASSERT(i < priv->cnt);
928 return (priv->bufs[i]);
929 }
930
931 void
932 dmu_xuio_clear(xuio_t *xuio, int i)
933 {
934 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
935
936 ASSERT(i < priv->cnt);
937 priv->bufs[i] = NULL;
938 }
939
940 static void
941 xuio_stat_init(void)
942 {
943 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
944 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
945 KSTAT_FLAG_VIRTUAL);
946 if (xuio_ksp != NULL) {
947 xuio_ksp->ks_data = &xuio_stats;
948 kstat_install(xuio_ksp);
949 }
950 }
951
952 static void
953 xuio_stat_fini(void)
954 {
955 if (xuio_ksp != NULL) {
956 kstat_delete(xuio_ksp);
957 xuio_ksp = NULL;
958 }
959 }
960
961 void
962 xuio_stat_wbuf_copied()
963 {
964 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
965 }
966
967 void
968 xuio_stat_wbuf_nocopy()
969 {
970 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
971 }
972
973 #ifdef _KERNEL
974 int
975 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
976 {
977 dmu_buf_t **dbp;
978 int numbufs, i, err;
979 xuio_t *xuio = NULL;
980
981 /*
982 * NB: we could do this block-at-a-time, but it's nice
983 * to be reading in parallel.
984 */
985 err = dmu_buf_hold_array(os, object, uio->uio_loffset, size, TRUE, FTAG,
986 &numbufs, &dbp);
987 if (err)
988 return (err);
989
990 if (uio->uio_extflg == UIO_XUIO)
991 xuio = (xuio_t *)uio;
992
993 for (i = 0; i < numbufs; i++) {
994 int tocpy;
995 int bufoff;
996 dmu_buf_t *db = dbp[i];
997
998 ASSERT(size > 0);
999
1000 bufoff = uio->uio_loffset - db->db_offset;
1001 tocpy = (int)MIN(db->db_size - bufoff, size);
1002
1003 if (xuio) {
1004 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
1005 arc_buf_t *dbuf_abuf = dbi->db_buf;
1006 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
1007 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
1008 if (!err) {
1009 uio->uio_resid -= tocpy;
1010 uio->uio_loffset += tocpy;
1011 }
1012
1013 if (abuf == dbuf_abuf)
1014 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
1015 else
1016 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
1017 } else {
1018 err = uiomove((char *)db->db_data + bufoff, tocpy,
1019 UIO_READ, uio);
1020 }
1021 if (err)
1022 break;
1023
1024 size -= tocpy;
1025 }
1026 dmu_buf_rele_array(dbp, numbufs, FTAG);
1027
1028 return (err);
1029 }
1030
1031 static int
1032 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1033 {
1034 dmu_buf_t **dbp;
1035 int numbufs;
1036 int err = 0;
1037 int i;
1038
1039 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1040 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1041 if (err)
1042 return (err);
1043
1044 for (i = 0; i < numbufs; i++) {
1045 int tocpy;
1046 int bufoff;
1047 dmu_buf_t *db = dbp[i];
1048
1049 ASSERT(size > 0);
1050
1051 bufoff = uio->uio_loffset - db->db_offset;
1052 tocpy = (int)MIN(db->db_size - bufoff, size);
1053
1054 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1055
1056 if (tocpy == db->db_size)
1057 dmu_buf_will_fill(db, tx);
1058 else
1059 dmu_buf_will_dirty(db, tx);
1060
1061 /*
1062 * XXX uiomove could block forever (eg. nfs-backed
1063 * pages). There needs to be a uiolockdown() function
1064 * to lock the pages in memory, so that uiomove won't
1065 * block.
1066 */
1067 err = uiomove((char *)db->db_data + bufoff, tocpy,
1068 UIO_WRITE, uio);
1069
1070 if (tocpy == db->db_size)
1071 dmu_buf_fill_done(db, tx);
1072
1073 if (err)
1074 break;
1075
1076 size -= tocpy;
1077 }
1078
1079 dmu_buf_rele_array(dbp, numbufs, FTAG);
1080 return (err);
1081 }
1082
1083 int
1084 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1085 dmu_tx_t *tx)
1086 {
1087 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1088 dnode_t *dn;
1089 int err;
1090
1091 if (size == 0)
1092 return (0);
1093
1094 DB_DNODE_ENTER(db);
1095 dn = DB_DNODE(db);
1096 err = dmu_write_uio_dnode(dn, uio, size, tx);
1097 DB_DNODE_EXIT(db);
1098
1099 return (err);
1100 }
1101
1102 int
1103 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1104 dmu_tx_t *tx)
1105 {
1106 dnode_t *dn;
1107 int err;
1108
1109 if (size == 0)
1110 return (0);
1111
1112 err = dnode_hold(os, object, FTAG, &dn);
1113 if (err)
1114 return (err);
1115
1116 err = dmu_write_uio_dnode(dn, uio, size, tx);
1117
1118 dnode_rele(dn, FTAG);
1119
1120 return (err);
1121 }
1122
1123 int
1124 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1125 page_t *pp, dmu_tx_t *tx)
1126 {
1127 dmu_buf_t **dbp;
1128 int numbufs, i;
1129 int err;
1130
1131 if (size == 0)
1132 return (0);
1133
1134 err = dmu_buf_hold_array(os, object, offset, size,
1135 FALSE, FTAG, &numbufs, &dbp);
1136 if (err)
1137 return (err);
1138
1139 for (i = 0; i < numbufs; i++) {
1140 int tocpy, copied, thiscpy;
1141 int bufoff;
1142 dmu_buf_t *db = dbp[i];
1143 caddr_t va;
1144
1145 ASSERT(size > 0);
1146 ASSERT3U(db->db_size, >=, PAGESIZE);
1147
1148 bufoff = offset - db->db_offset;
1149 tocpy = (int)MIN(db->db_size - bufoff, size);
1150
1151 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1152
1153 if (tocpy == db->db_size)
1154 dmu_buf_will_fill(db, tx);
1155 else
1156 dmu_buf_will_dirty(db, tx);
1157
1158 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1159 ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff);
1160 thiscpy = MIN(PAGESIZE, tocpy - copied);
1161 va = zfs_map_page(pp, S_READ);
1162 bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1163 zfs_unmap_page(pp, va);
1164 pp = pp->p_next;
1165 bufoff += PAGESIZE;
1166 }
1167
1168 if (tocpy == db->db_size)
1169 dmu_buf_fill_done(db, tx);
1170
1171 offset += tocpy;
1172 size -= tocpy;
1173 }
1174 dmu_buf_rele_array(dbp, numbufs, FTAG);
1175 return (err);
1176 }
1177 #endif
1178
1179 /*
1180 * Allocate a loaned anonymous arc buffer.
1181 */
1182 arc_buf_t *
1183 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1184 {
1185 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1186 spa_t *spa;
1187
1188 DB_GET_SPA(&spa, db);
1189 return (arc_loan_buf(spa, size));
1190 }
1191
1192 /*
1193 * Free a loaned arc buffer.
1194 */
1195 void
1196 dmu_return_arcbuf(arc_buf_t *buf)
1197 {
1198 arc_return_buf(buf, FTAG);
1199 VERIFY(arc_buf_remove_ref(buf, FTAG));
1200 }
1201
1202 /*
1203 * When possible directly assign passed loaned arc buffer to a dbuf.
1204 * If this is not possible copy the contents of passed arc buf via
1205 * dmu_write().
1206 */
1207 void
1208 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1209 dmu_tx_t *tx)
1210 {
1211 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1212 dnode_t *dn;
1213 dmu_buf_impl_t *db;
1214 uint32_t blksz = (uint32_t)arc_buf_size(buf);
1215 uint64_t blkid;
1216
1217 DB_DNODE_ENTER(dbuf);
1218 dn = DB_DNODE(dbuf);
1219 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1220 blkid = dbuf_whichblock(dn, offset);
1221 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1222 rw_exit(&dn->dn_struct_rwlock);
1223 DB_DNODE_EXIT(dbuf);
1224
1225 if (offset == db->db.db_offset && blksz == db->db.db_size) {
1226 dbuf_assign_arcbuf(db, buf, tx);
1227 dbuf_rele(db, FTAG);
1228 } else {
1229 objset_t *os;
1230 uint64_t object;
1231
1232 DB_DNODE_ENTER(dbuf);
1233 dn = DB_DNODE(dbuf);
1234 os = dn->dn_objset;
1235 object = dn->dn_object;
1236 DB_DNODE_EXIT(dbuf);
1237
1238 dbuf_rele(db, FTAG);
1239 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1240 dmu_return_arcbuf(buf);
1241 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1242 }
1243 }
1244
1245 typedef struct {
1246 dbuf_dirty_record_t *dsa_dr;
1247 dmu_sync_cb_t *dsa_done;
1248 zgd_t *dsa_zgd;
1249 dmu_tx_t *dsa_tx;
1250 } dmu_sync_arg_t;
1251
1252 /* ARGSUSED */
1253 static void
1254 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1255 {
1256 dmu_sync_arg_t *dsa = varg;
1257 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1258 blkptr_t *bp = zio->io_bp;
1259
1260 if (zio->io_error == 0) {
1261 if (BP_IS_HOLE(bp)) {
1262 /*
1263 * A block of zeros may compress to a hole, but the
1264 * block size still needs to be known for replay.
1265 */
1266 BP_SET_LSIZE(bp, db->db_size);
1267 } else {
1268 ASSERT(BP_GET_LEVEL(bp) == 0);
1269 bp->blk_fill = 1;
1270 }
1271 }
1272 }
1273
1274 static void
1275 dmu_sync_late_arrival_ready(zio_t *zio)
1276 {
1277 dmu_sync_ready(zio, NULL, zio->io_private);
1278 }
1279
1280 /* ARGSUSED */
1281 static void
1282 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1283 {
1284 dmu_sync_arg_t *dsa = varg;
1285 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1286 dmu_buf_impl_t *db = dr->dr_dbuf;
1287
1288 mutex_enter(&db->db_mtx);
1289 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1290 if (zio->io_error == 0) {
1291 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1292 if (dr->dt.dl.dr_nopwrite) {
1293 blkptr_t *bp = zio->io_bp;
1294 blkptr_t *bp_orig = &zio->io_bp_orig;
1295 uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1296
1297 ASSERT(BP_EQUAL(bp, bp_orig));
1298 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1299 ASSERT(zio_checksum_table[chksum].ci_dedup);
1300 }
1301 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1302 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1303 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1304 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by))
1305 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1306 } else {
1307 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1308 }
1309 cv_broadcast(&db->db_changed);
1310 mutex_exit(&db->db_mtx);
1311
1312 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1313
1314 kmem_free(dsa, sizeof (*dsa));
1315 }
1316
1317 static void
1318 dmu_sync_late_arrival_done(zio_t *zio)
1319 {
1320 blkptr_t *bp = zio->io_bp;
1321 dmu_sync_arg_t *dsa = zio->io_private;
1322 blkptr_t *bp_orig = &zio->io_bp_orig;
1323
1324 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1325 /*
1326 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE)
1327 * then there is nothing to do here. Otherwise, free the
1328 * newly allocated block in this txg.
1329 */
1330 if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
1331 ASSERT(BP_EQUAL(bp, bp_orig));
1332 } else {
1333 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1334 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1335 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1336 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1337 }
1338 }
1339
1340 dmu_tx_commit(dsa->dsa_tx);
1341
1342 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1343
1344 kmem_free(dsa, sizeof (*dsa));
1345 }
1346
1347 static int
1348 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1349 zio_prop_t *zp, zbookmark_t *zb)
1350 {
1351 dmu_sync_arg_t *dsa;
1352 dmu_tx_t *tx;
1353
1354 tx = dmu_tx_create(os);
1355 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1356 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1357 dmu_tx_abort(tx);
1358 /* Make zl_get_data do txg_waited_synced() */
1359 return (SET_ERROR(EIO));
1360 }
1361
1362 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1363 dsa->dsa_dr = NULL;
1364 dsa->dsa_done = done;
1365 dsa->dsa_zgd = zgd;
1366 dsa->dsa_tx = tx;
1367
1368 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1369 zgd->zgd_db->db_data, zgd->zgd_db->db_size, zp,
1370 dmu_sync_late_arrival_ready, dmu_sync_late_arrival_done, dsa,
1371 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1372
1373 return (0);
1374 }
1375
1376 /*
1377 * Intent log support: sync the block associated with db to disk.
1378 * N.B. and XXX: the caller is responsible for making sure that the
1379 * data isn't changing while dmu_sync() is writing it.
1380 *
1381 * Return values:
1382 *
1383 * EEXIST: this txg has already been synced, so there's nothing to do.
1384 * The caller should not log the write.
1385 *
1386 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1387 * The caller should not log the write.
1388 *
1389 * EALREADY: this block is already in the process of being synced.
1390 * The caller should track its progress (somehow).
1391 *
1392 * EIO: could not do the I/O.
1393 * The caller should do a txg_wait_synced().
1394 *
1395 * 0: the I/O has been initiated.
1396 * The caller should log this blkptr in the done callback.
1397 * It is possible that the I/O will fail, in which case
1398 * the error will be reported to the done callback and
1399 * propagated to pio from zio_done().
1400 */
1401 int
1402 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1403 {
1404 blkptr_t *bp = zgd->zgd_bp;
1405 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1406 objset_t *os = db->db_objset;
1407 dsl_dataset_t *ds = os->os_dsl_dataset;
1408 dbuf_dirty_record_t *dr;
1409 dmu_sync_arg_t *dsa;
1410 zbookmark_t zb;
1411 zio_prop_t zp;
1412 dnode_t *dn;
1413
1414 ASSERT(pio != NULL);
1415 ASSERT(txg != 0);
1416
1417 SET_BOOKMARK(&zb, ds->ds_object,
1418 db->db.db_object, db->db_level, db->db_blkid);
1419
1420 DB_DNODE_ENTER(db);
1421 dn = DB_DNODE(db);
1422 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1423 DB_DNODE_EXIT(db);
1424
1425 /*
1426 * If we're frozen (running ziltest), we always need to generate a bp.
1427 */
1428 if (txg > spa_freeze_txg(os->os_spa))
1429 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1430
1431 /*
1432 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1433 * and us. If we determine that this txg is not yet syncing,
1434 * but it begins to sync a moment later, that's OK because the
1435 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1436 */
1437 mutex_enter(&db->db_mtx);
1438
1439 if (txg <= spa_last_synced_txg(os->os_spa)) {
1440 /*
1441 * This txg has already synced. There's nothing to do.
1442 */
1443 mutex_exit(&db->db_mtx);
1444 return (SET_ERROR(EEXIST));
1445 }
1446
1447 if (txg <= spa_syncing_txg(os->os_spa)) {
1448 /*
1449 * This txg is currently syncing, so we can't mess with
1450 * the dirty record anymore; just write a new log block.
1451 */
1452 mutex_exit(&db->db_mtx);
1453 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1454 }
1455
1456 dr = db->db_last_dirty;
1457 while (dr && dr->dr_txg != txg)
1458 dr = dr->dr_next;
1459
1460 if (dr == NULL) {
1461 /*
1462 * There's no dr for this dbuf, so it must have been freed.
1463 * There's no need to log writes to freed blocks, so we're done.
1464 */
1465 mutex_exit(&db->db_mtx);
1466 return (SET_ERROR(ENOENT));
1467 }
1468
1469 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
1470
1471 /*
1472 * Assume the on-disk data is X, the current syncing data is Y,
1473 * and the current in-memory data is Z (currently in dmu_sync).
1474 * X and Z are identical but Y is has been modified. Normally,
1475 * when X and Z are the same we will perform a nopwrite but if Y
1476 * is different we must disable nopwrite since the resulting write
1477 * of Y to disk can free the block containing X. If we allowed a
1478 * nopwrite to occur the block pointing to Z would reference a freed
1479 * block. Since this is a rare case we simplify this by disabling
1480 * nopwrite if the current dmu_sync-ing dbuf has been modified in
1481 * a previous transaction.
1482 */
1483 if (dr->dr_next)
1484 zp.zp_nopwrite = B_FALSE;
1485
1486 ASSERT(dr->dr_txg == txg);
1487 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1488 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1489 /*
1490 * We have already issued a sync write for this buffer,
1491 * or this buffer has already been synced. It could not
1492 * have been dirtied since, or we would have cleared the state.
1493 */
1494 mutex_exit(&db->db_mtx);
1495 return (SET_ERROR(EALREADY));
1496 }
1497
1498 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1499 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1500 mutex_exit(&db->db_mtx);
1501
1502 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1503 dsa->dsa_dr = dr;
1504 dsa->dsa_done = done;
1505 dsa->dsa_zgd = zgd;
1506 dsa->dsa_tx = NULL;
1507
1508 zio_nowait(arc_write(pio, os->os_spa, txg,
1509 bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
1510 DBUF_IS_L2COMPRESSIBLE(db), &zp, dmu_sync_ready, dmu_sync_done,
1511 dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb));
1512
1513 return (0);
1514 }
1515
1516 int
1517 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1518 dmu_tx_t *tx)
1519 {
1520 dnode_t *dn;
1521 int err;
1522
1523 err = dnode_hold(os, object, FTAG, &dn);
1524 if (err)
1525 return (err);
1526 err = dnode_set_blksz(dn, size, ibs, tx);
1527 dnode_rele(dn, FTAG);
1528 return (err);
1529 }
1530
1531 void
1532 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1533 dmu_tx_t *tx)
1534 {
1535 dnode_t *dn;
1536
1537 /* XXX assumes dnode_hold will not get an i/o error */
1538 (void) dnode_hold(os, object, FTAG, &dn);
1539 ASSERT(checksum < ZIO_CHECKSUM_FUNCTIONS);
1540 dn->dn_checksum = checksum;
1541 dnode_setdirty(dn, tx);
1542 dnode_rele(dn, FTAG);
1543 }
1544
1545 void
1546 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1547 dmu_tx_t *tx)
1548 {
1549 dnode_t *dn;
1550
1551 /* XXX assumes dnode_hold will not get an i/o error */
1552 (void) dnode_hold(os, object, FTAG, &dn);
1553 ASSERT(compress < ZIO_COMPRESS_FUNCTIONS);
1554 dn->dn_compress = compress;
1555 dnode_setdirty(dn, tx);
1556 dnode_rele(dn, FTAG);
1557 }
1558
1559 int zfs_mdcomp_disable = 0;
1560
1561 void
1562 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
1563 {
1564 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
1565 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
1566 (wp & WP_SPILL));
1567 enum zio_checksum checksum = os->os_checksum;
1568 enum zio_compress compress = os->os_compress;
1569 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
1570 boolean_t dedup = B_FALSE;
1571 boolean_t nopwrite = B_FALSE;
1572 boolean_t dedup_verify = os->os_dedup_verify;
1573 int copies = os->os_copies;
1574
1575 /*
1576 * We maintain different write policies for each of the following
1577 * types of data:
1578 * 1. metadata
1579 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1580 * 3. all other level 0 blocks
1581 */
1582 if (ismd) {
1583 /*
1584 * XXX -- we should design a compression algorithm
1585 * that specializes in arrays of bps.
1586 */
1587 compress = zfs_mdcomp_disable ? ZIO_COMPRESS_EMPTY :
1588 ZIO_COMPRESS_LZJB;
1589
1590 /*
1591 * Metadata always gets checksummed. If the data
1592 * checksum is multi-bit correctable, and it's not a
1593 * ZBT-style checksum, then it's suitable for metadata
1594 * as well. Otherwise, the metadata checksum defaults
1595 * to fletcher4.
1596 */
1597 if (zio_checksum_table[checksum].ci_correctable < 1 ||
1598 zio_checksum_table[checksum].ci_eck)
1599 checksum = ZIO_CHECKSUM_FLETCHER_4;
1600 } else if (wp & WP_NOFILL) {
1601 ASSERT(level == 0);
1602
1603 /*
1604 * If we're writing preallocated blocks, we aren't actually
1605 * writing them so don't set any policy properties. These
1606 * blocks are currently only used by an external subsystem
1607 * outside of zfs (i.e. dump) and not written by the zio
1608 * pipeline.
1609 */
1610 compress = ZIO_COMPRESS_OFF;
1611 checksum = ZIO_CHECKSUM_OFF;
1612 } else {
1613 compress = zio_compress_select(dn->dn_compress, compress);
1614
1615 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
1616 zio_checksum_select(dn->dn_checksum, checksum) :
1617 dedup_checksum;
1618
1619 /*
1620 * Determine dedup setting. If we are in dmu_sync(),
1621 * we won't actually dedup now because that's all
1622 * done in syncing context; but we do want to use the
1623 * dedup checkum. If the checksum is not strong
1624 * enough to ensure unique signatures, force
1625 * dedup_verify.
1626 */
1627 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
1628 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
1629 if (!zio_checksum_table[checksum].ci_dedup)
1630 dedup_verify = B_TRUE;
1631 }
1632
1633 /*
1634 * Enable nopwrite if we have a cryptographically secure
1635 * checksum that has no known collisions (i.e. SHA-256)
1636 * and compression is enabled. We don't enable nopwrite if
1637 * dedup is enabled as the two features are mutually exclusive.
1638 */
1639 nopwrite = (!dedup && zio_checksum_table[checksum].ci_dedup &&
1640 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
1641 }
1642
1643 zp->zp_checksum = checksum;
1644 zp->zp_compress = compress;
1645 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
1646 zp->zp_level = level;
1647 zp->zp_copies = MIN(copies + ismd, spa_max_replication(os->os_spa));
1648 zp->zp_dedup = dedup;
1649 zp->zp_dedup_verify = dedup && dedup_verify;
1650 zp->zp_nopwrite = nopwrite;
1651 }
1652
1653 int
1654 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
1655 {
1656 dnode_t *dn;
1657 int i, err;
1658
1659 err = dnode_hold(os, object, FTAG, &dn);
1660 if (err)
1661 return (err);
1662 /*
1663 * Sync any current changes before
1664 * we go trundling through the block pointers.
1665 */
1666 for (i = 0; i < TXG_SIZE; i++) {
1667 if (list_link_active(&dn->dn_dirty_link[i]))
1668 break;
1669 }
1670 if (i != TXG_SIZE) {
1671 dnode_rele(dn, FTAG);
1672 txg_wait_synced(dmu_objset_pool(os), 0);
1673 err = dnode_hold(os, object, FTAG, &dn);
1674 if (err)
1675 return (err);
1676 }
1677
1678 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
1679 dnode_rele(dn, FTAG);
1680
1681 return (err);
1682 }
1683
1684 void
1685 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
1686 {
1687 dnode_phys_t *dnp;
1688
1689 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1690 mutex_enter(&dn->dn_mtx);
1691
1692 dnp = dn->dn_phys;
1693
1694 doi->doi_data_block_size = dn->dn_datablksz;
1695 doi->doi_metadata_block_size = dn->dn_indblkshift ?
1696 1ULL << dn->dn_indblkshift : 0;
1697 doi->doi_type = dn->dn_type;
1698 doi->doi_bonus_type = dn->dn_bonustype;
1699 doi->doi_bonus_size = dn->dn_bonuslen;
1700 doi->doi_indirection = dn->dn_nlevels;
1701 doi->doi_checksum = dn->dn_checksum;
1702 doi->doi_compress = dn->dn_compress;
1703 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
1704 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
1705 doi->doi_fill_count = 0;
1706 for (int i = 0; i < dnp->dn_nblkptr; i++)
1707 doi->doi_fill_count += dnp->dn_blkptr[i].blk_fill;
1708
1709 mutex_exit(&dn->dn_mtx);
1710 rw_exit(&dn->dn_struct_rwlock);
1711 }
1712
1713 /*
1714 * Get information on a DMU object.
1715 * If doi is NULL, just indicates whether the object exists.
1716 */
1717 int
1718 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
1719 {
1720 dnode_t *dn;
1721 int err = dnode_hold(os, object, FTAG, &dn);
1722
1723 if (err)
1724 return (err);
1725
1726 if (doi != NULL)
1727 dmu_object_info_from_dnode(dn, doi);
1728
1729 dnode_rele(dn, FTAG);
1730 return (0);
1731 }
1732
1733 /*
1734 * As above, but faster; can be used when you have a held dbuf in hand.
1735 */
1736 void
1737 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
1738 {
1739 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1740
1741 DB_DNODE_ENTER(db);
1742 dmu_object_info_from_dnode(DB_DNODE(db), doi);
1743 DB_DNODE_EXIT(db);
1744 }
1745
1746 /*
1747 * Faster still when you only care about the size.
1748 * This is specifically optimized for zfs_getattr().
1749 */
1750 void
1751 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
1752 u_longlong_t *nblk512)
1753 {
1754 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1755 dnode_t *dn;
1756
1757 DB_DNODE_ENTER(db);
1758 dn = DB_DNODE(db);
1759
1760 *blksize = dn->dn_datablksz;
1761 /* add 1 for dnode space */
1762 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
1763 SPA_MINBLOCKSHIFT) + 1;
1764 DB_DNODE_EXIT(db);
1765 }
1766
1767 void
1768 byteswap_uint64_array(void *vbuf, size_t size)
1769 {
1770 uint64_t *buf = vbuf;
1771 size_t count = size >> 3;
1772 int i;
1773
1774 ASSERT((size & 7) == 0);
1775
1776 for (i = 0; i < count; i++)
1777 buf[i] = BSWAP_64(buf[i]);
1778 }
1779
1780 void
1781 byteswap_uint32_array(void *vbuf, size_t size)
1782 {
1783 uint32_t *buf = vbuf;
1784 size_t count = size >> 2;
1785 int i;
1786
1787 ASSERT((size & 3) == 0);
1788
1789 for (i = 0; i < count; i++)
1790 buf[i] = BSWAP_32(buf[i]);
1791 }
1792
1793 void
1794 byteswap_uint16_array(void *vbuf, size_t size)
1795 {
1796 uint16_t *buf = vbuf;
1797 size_t count = size >> 1;
1798 int i;
1799
1800 ASSERT((size & 1) == 0);
1801
1802 for (i = 0; i < count; i++)
1803 buf[i] = BSWAP_16(buf[i]);
1804 }
1805
1806 /* ARGSUSED */
1807 void
1808 byteswap_uint8_array(void *vbuf, size_t size)
1809 {
1810 }
1811
1812 void
1813 dmu_init(void)
1814 {
1815 zfs_dbgmsg_init();
1816 sa_cache_init();
1817 xuio_stat_init();
1818 dmu_objset_init();
1819 dnode_init();
1820 dbuf_init();
1821 zfetch_init();
1822 l2arc_init();
1823 arc_init();
1824 }
1825
1826 void
1827 dmu_fini(void)
1828 {
1829 arc_fini(); /* arc depends on l2arc, so arc must go first */
1830 l2arc_fini();
1831 zfetch_fini();
1832 dbuf_fini();
1833 dnode_fini();
1834 dmu_objset_fini();
1835 xuio_stat_fini();
1836 sa_cache_fini();
1837 zfs_dbgmsg_fini();
1838 }