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) 2012, 2014 by Delphix. All rights reserved.
24 */
25 /* Copyright (c) 2013 by Saso Kiselkov. All rights reserved. */
26 /* Copyright (c) 2013, Joyent, Inc. All rights reserved. */
27 /* Copyright (c) 2014, Nexenta Systems, Inc. All rights reserved. */
28
29 #include <sys/dmu.h>
30 #include <sys/dmu_impl.h>
31 #include <sys/dmu_tx.h>
32 #include <sys/dbuf.h>
33 #include <sys/dnode.h>
34 #include <sys/zfs_context.h>
35 #include <sys/dmu_objset.h>
36 #include <sys/dmu_traverse.h>
37 #include <sys/dsl_dataset.h>
38 #include <sys/dsl_dir.h>
39 #include <sys/dsl_pool.h>
40 #include <sys/dsl_synctask.h>
41 #include <sys/dsl_prop.h>
42 #include <sys/dmu_zfetch.h>
43 #include <sys/zfs_ioctl.h>
44 #include <sys/zap.h>
45 #include <sys/zio_checksum.h>
46 #include <sys/zio_compress.h>
47 #include <sys/sa.h>
48 #include <sys/zfeature.h>
49 #ifdef _KERNEL
50 #include <sys/vmsystm.h>
51 #include <sys/zfs_znode.h>
52 #endif
53
54 /*
55 * Enable/disable nopwrite feature.
56 */
57 int zfs_nopwrite_enabled = 1;
58
59 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
60 { DMU_BSWAP_UINT8, TRUE, "unallocated" },
61 { DMU_BSWAP_ZAP, TRUE, "object directory" },
62 { DMU_BSWAP_UINT64, TRUE, "object array" },
63 { DMU_BSWAP_UINT8, TRUE, "packed nvlist" },
64 { DMU_BSWAP_UINT64, TRUE, "packed nvlist size" },
65 { DMU_BSWAP_UINT64, TRUE, "bpobj" },
66 { DMU_BSWAP_UINT64, TRUE, "bpobj header" },
67 { DMU_BSWAP_UINT64, TRUE, "SPA space map header" },
68 { DMU_BSWAP_UINT64, TRUE, "SPA space map" },
69 { DMU_BSWAP_UINT64, TRUE, "ZIL intent log" },
70 { DMU_BSWAP_DNODE, TRUE, "DMU dnode" },
71 { DMU_BSWAP_OBJSET, TRUE, "DMU objset" },
72 { DMU_BSWAP_UINT64, TRUE, "DSL directory" },
73 { DMU_BSWAP_ZAP, TRUE, "DSL directory child map"},
74 { DMU_BSWAP_ZAP, TRUE, "DSL dataset snap map" },
75 { DMU_BSWAP_ZAP, TRUE, "DSL props" },
76 { DMU_BSWAP_UINT64, TRUE, "DSL dataset" },
77 { DMU_BSWAP_ZNODE, TRUE, "ZFS znode" },
78 { DMU_BSWAP_OLDACL, TRUE, "ZFS V0 ACL" },
79 { DMU_BSWAP_UINT8, FALSE, "ZFS plain file" },
80 { DMU_BSWAP_ZAP, TRUE, "ZFS directory" },
81 { DMU_BSWAP_ZAP, TRUE, "ZFS master node" },
82 { DMU_BSWAP_ZAP, TRUE, "ZFS delete queue" },
83 { DMU_BSWAP_UINT8, FALSE, "zvol object" },
84 { DMU_BSWAP_ZAP, TRUE, "zvol prop" },
85 { DMU_BSWAP_UINT8, FALSE, "other uint8[]" },
86 { DMU_BSWAP_UINT64, FALSE, "other uint64[]" },
87 { DMU_BSWAP_ZAP, TRUE, "other ZAP" },
88 { DMU_BSWAP_ZAP, TRUE, "persistent error log" },
89 { DMU_BSWAP_UINT8, TRUE, "SPA history" },
90 { DMU_BSWAP_UINT64, TRUE, "SPA history offsets" },
91 { DMU_BSWAP_ZAP, TRUE, "Pool properties" },
92 { DMU_BSWAP_ZAP, TRUE, "DSL permissions" },
93 { DMU_BSWAP_ACL, TRUE, "ZFS ACL" },
94 { DMU_BSWAP_UINT8, TRUE, "ZFS SYSACL" },
95 { DMU_BSWAP_UINT8, TRUE, "FUID table" },
96 { DMU_BSWAP_UINT64, TRUE, "FUID table size" },
97 { DMU_BSWAP_ZAP, TRUE, "DSL dataset next clones"},
98 { DMU_BSWAP_ZAP, TRUE, "scan work queue" },
99 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group used" },
100 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group quota" },
101 { DMU_BSWAP_ZAP, TRUE, "snapshot refcount tags"},
102 { DMU_BSWAP_ZAP, TRUE, "DDT ZAP algorithm" },
103 { DMU_BSWAP_ZAP, TRUE, "DDT statistics" },
104 { DMU_BSWAP_UINT8, TRUE, "System attributes" },
105 { DMU_BSWAP_ZAP, TRUE, "SA master node" },
106 { DMU_BSWAP_ZAP, TRUE, "SA attr registration" },
107 { DMU_BSWAP_ZAP, TRUE, "SA attr layouts" },
108 { DMU_BSWAP_ZAP, TRUE, "scan translations" },
109 { DMU_BSWAP_UINT8, FALSE, "deduplicated block" },
110 { DMU_BSWAP_ZAP, TRUE, "DSL deadlist map" },
111 { DMU_BSWAP_UINT64, TRUE, "DSL deadlist map hdr" },
112 { DMU_BSWAP_ZAP, TRUE, "DSL dir clones" },
113 { DMU_BSWAP_UINT64, TRUE, "bpobj subobj" }
114 };
115
116 const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
117 { byteswap_uint8_array, "uint8" },
118 { byteswap_uint16_array, "uint16" },
119 { byteswap_uint32_array, "uint32" },
120 { byteswap_uint64_array, "uint64" },
121 { zap_byteswap, "zap" },
122 { dnode_buf_byteswap, "dnode" },
123 { dmu_objset_byteswap, "objset" },
124 { zfs_znode_byteswap, "znode" },
125 { zfs_oldacl_byteswap, "oldacl" },
126 { zfs_acl_byteswap, "acl" }
127 };
128
129 int
130 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
131 void *tag, dmu_buf_t **dbp, int flags)
132 {
133 dnode_t *dn;
134 uint64_t blkid;
135 dmu_buf_impl_t *db;
136 int err;
137 int db_flags = DB_RF_CANFAIL;
138
139 if (flags & DMU_READ_NO_PREFETCH)
140 db_flags |= DB_RF_NOPREFETCH;
141
142 err = dnode_hold(os, object, FTAG, &dn);
143 if (err)
144 return (err);
145 blkid = dbuf_whichblock(dn, offset);
146 rw_enter(&dn->dn_struct_rwlock, RW_READER);
147 db = dbuf_hold(dn, blkid, tag);
148 rw_exit(&dn->dn_struct_rwlock);
149 if (db == NULL) {
150 err = SET_ERROR(EIO);
151 } else {
152 err = dbuf_read(db, NULL, db_flags);
153 if (err) {
154 dbuf_rele(db, tag);
155 db = NULL;
156 }
157 }
158
159 dnode_rele(dn, FTAG);
160 *dbp = &db->db; /* NULL db plus first field offset is NULL */
161 return (err);
162 }
163
164 int
165 dmu_bonus_max(void)
166 {
167 return (DN_MAX_BONUSLEN);
168 }
169
170 int
171 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
172 {
173 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
174 dnode_t *dn;
175 int error;
176
177 DB_DNODE_ENTER(db);
178 dn = DB_DNODE(db);
179
180 if (dn->dn_bonus != db) {
181 error = SET_ERROR(EINVAL);
182 } else if (newsize < 0 || newsize > db_fake->db_size) {
183 error = SET_ERROR(EINVAL);
184 } else {
185 dnode_setbonuslen(dn, newsize, tx);
186 error = 0;
187 }
188
189 DB_DNODE_EXIT(db);
190 return (error);
191 }
192
193 int
194 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
195 {
196 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
197 dnode_t *dn;
198 int error;
199
200 DB_DNODE_ENTER(db);
201 dn = DB_DNODE(db);
202
203 if (!DMU_OT_IS_VALID(type)) {
204 error = SET_ERROR(EINVAL);
205 } else if (dn->dn_bonus != db) {
206 error = SET_ERROR(EINVAL);
207 } else {
208 dnode_setbonus_type(dn, type, tx);
209 error = 0;
210 }
211
212 DB_DNODE_EXIT(db);
213 return (error);
214 }
215
216 dmu_object_type_t
217 dmu_get_bonustype(dmu_buf_t *db_fake)
218 {
219 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
220 dnode_t *dn;
221 dmu_object_type_t type;
222
223 DB_DNODE_ENTER(db);
224 dn = DB_DNODE(db);
225 type = dn->dn_bonustype;
226 DB_DNODE_EXIT(db);
227
228 return (type);
229 }
230
231 int
232 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
233 {
234 dnode_t *dn;
235 int error;
236
237 error = dnode_hold(os, object, FTAG, &dn);
238 dbuf_rm_spill(dn, tx);
239 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
240 dnode_rm_spill(dn, tx);
241 rw_exit(&dn->dn_struct_rwlock);
242 dnode_rele(dn, FTAG);
243 return (error);
244 }
245
246 /*
247 * returns ENOENT, EIO, or 0.
248 */
249 int
250 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
251 {
252 dnode_t *dn;
253 dmu_buf_impl_t *db;
254 int error;
255
256 error = dnode_hold(os, object, FTAG, &dn);
257 if (error)
258 return (error);
259
260 rw_enter(&dn->dn_struct_rwlock, RW_READER);
261 if (dn->dn_bonus == NULL) {
262 rw_exit(&dn->dn_struct_rwlock);
263 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
264 if (dn->dn_bonus == NULL)
265 dbuf_create_bonus(dn);
266 }
267 db = dn->dn_bonus;
268
269 /* as long as the bonus buf is held, the dnode will be held */
270 if (refcount_add(&db->db_holds, tag) == 1) {
271 VERIFY(dnode_add_ref(dn, db));
272 (void) atomic_inc_32_nv(&dn->dn_dbufs_count);
273 }
274
275 /*
276 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
277 * hold and incrementing the dbuf count to ensure that dnode_move() sees
278 * a dnode hold for every dbuf.
279 */
280 rw_exit(&dn->dn_struct_rwlock);
281
282 dnode_rele(dn, FTAG);
283
284 VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH));
285
286 *dbp = &db->db;
287 return (0);
288 }
289
290 /*
291 * returns ENOENT, EIO, or 0.
292 *
293 * This interface will allocate a blank spill dbuf when a spill blk
294 * doesn't already exist on the dnode.
295 *
296 * if you only want to find an already existing spill db, then
297 * dmu_spill_hold_existing() should be used.
298 */
299 int
300 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
301 {
302 dmu_buf_impl_t *db = NULL;
303 int err;
304
305 if ((flags & DB_RF_HAVESTRUCT) == 0)
306 rw_enter(&dn->dn_struct_rwlock, RW_READER);
307
308 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
309
310 if ((flags & DB_RF_HAVESTRUCT) == 0)
311 rw_exit(&dn->dn_struct_rwlock);
312
313 ASSERT(db != NULL);
314 err = dbuf_read(db, NULL, flags);
315 if (err == 0)
316 *dbp = &db->db;
317 else
318 dbuf_rele(db, tag);
319 return (err);
320 }
321
322 int
323 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
324 {
325 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
326 dnode_t *dn;
327 int err;
328
329 DB_DNODE_ENTER(db);
330 dn = DB_DNODE(db);
331
332 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
333 err = SET_ERROR(EINVAL);
334 } else {
335 rw_enter(&dn->dn_struct_rwlock, RW_READER);
336
337 if (!dn->dn_have_spill) {
338 err = SET_ERROR(ENOENT);
339 } else {
340 err = dmu_spill_hold_by_dnode(dn,
341 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
342 }
343
344 rw_exit(&dn->dn_struct_rwlock);
345 }
346
347 DB_DNODE_EXIT(db);
348 return (err);
349 }
350
351 int
352 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
353 {
354 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
355 dnode_t *dn;
356 int err;
357
358 DB_DNODE_ENTER(db);
359 dn = DB_DNODE(db);
360 err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp);
361 DB_DNODE_EXIT(db);
362
363 return (err);
364 }
365
366 /*
367 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
368 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
369 * and can induce severe lock contention when writing to several files
370 * whose dnodes are in the same block.
371 */
372 static int
373 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
374 int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
375 {
376 dmu_buf_t **dbp;
377 uint64_t blkid, nblks, i;
378 uint32_t dbuf_flags;
379 int err;
380 zio_t *zio;
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 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
409 blkid = dbuf_whichblock(dn, offset);
410 for (i = 0; i < nblks; i++) {
411 dmu_buf_impl_t *db = dbuf_hold(dn, blkid+i, tag);
412 if (db == NULL) {
413 rw_exit(&dn->dn_struct_rwlock);
414 dmu_buf_rele_array(dbp, nblks, tag);
415 zio_nowait(zio);
416 return (SET_ERROR(EIO));
417 }
418 /* initiate async i/o */
419 if (read) {
420 (void) dbuf_read(db, zio, dbuf_flags);
421 }
422 dbp[i] = &db->db;
423 }
424 rw_exit(&dn->dn_struct_rwlock);
425
426 /* wait for async i/o */
427 err = zio_wait(zio);
428 if (err) {
429 dmu_buf_rele_array(dbp, nblks, tag);
430 return (err);
431 }
432
433 /* wait for other io to complete */
434 if (read) {
435 for (i = 0; i < nblks; i++) {
436 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
437 mutex_enter(&db->db_mtx);
438 while (db->db_state == DB_READ ||
439 db->db_state == DB_FILL)
440 cv_wait(&db->db_changed, &db->db_mtx);
441 if (db->db_state == DB_UNCACHED)
442 err = SET_ERROR(EIO);
443 mutex_exit(&db->db_mtx);
444 if (err) {
445 dmu_buf_rele_array(dbp, nblks, tag);
446 return (err);
447 }
448 }
449 }
450
451 *numbufsp = nblks;
452 *dbpp = dbp;
453 return (0);
454 }
455
456 static int
457 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
458 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
459 {
460 dnode_t *dn;
461 int err;
462
463 err = dnode_hold(os, object, FTAG, &dn);
464 if (err)
465 return (err);
466
467 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
468 numbufsp, dbpp, DMU_READ_PREFETCH);
469
470 dnode_rele(dn, FTAG);
471
472 return (err);
473 }
474
475 int
476 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
477 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
478 {
479 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
480 dnode_t *dn;
481 int err;
482
483 DB_DNODE_ENTER(db);
484 dn = DB_DNODE(db);
485 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
486 numbufsp, dbpp, DMU_READ_PREFETCH);
487 DB_DNODE_EXIT(db);
488
489 return (err);
490 }
491
492 void
493 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
494 {
495 int i;
496 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
497
498 if (numbufs == 0)
499 return;
500
501 for (i = 0; i < numbufs; i++) {
502 if (dbp[i])
503 dbuf_rele(dbp[i], tag);
504 }
505
506 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
507 }
508
509 /*
510 * Issue prefetch i/os for the given blocks.
511 *
512 * Note: The assumption is that we *know* these blocks will be needed
513 * almost immediately. Therefore, the prefetch i/os will be issued at
514 * ZIO_PRIORITY_SYNC_READ
515 *
516 * Note: indirect blocks and other metadata will be read synchronously,
517 * causing this function to block if they are not already cached.
518 */
519 void
520 dmu_prefetch(objset_t *os, uint64_t object, uint64_t offset, uint64_t len)
521 {
522 dnode_t *dn;
523 uint64_t blkid;
524 int nblks, err;
525
526 if (zfs_prefetch_disable)
527 return;
528
529 if (len == 0) { /* they're interested in the bonus buffer */
530 dn = DMU_META_DNODE(os);
531
532 if (object == 0 || object >= DN_MAX_OBJECT)
533 return;
534
535 rw_enter(&dn->dn_struct_rwlock, RW_READER);
536 blkid = dbuf_whichblock(dn, object * sizeof (dnode_phys_t));
537 dbuf_prefetch(dn, blkid, ZIO_PRIORITY_SYNC_READ);
538 rw_exit(&dn->dn_struct_rwlock);
539 return;
540 }
541
542 /*
543 * XXX - Note, if the dnode for the requested object is not
544 * already cached, we will do a *synchronous* read in the
545 * dnode_hold() call. The same is true for any indirects.
546 */
547 err = dnode_hold(os, object, FTAG, &dn);
548 if (err != 0)
549 return;
550
551 rw_enter(&dn->dn_struct_rwlock, RW_READER);
552 if (dn->dn_datablkshift) {
553 int blkshift = dn->dn_datablkshift;
554 nblks = (P2ROUNDUP(offset + len, 1 << blkshift) -
555 P2ALIGN(offset, 1 << blkshift)) >> blkshift;
556 } else {
557 nblks = (offset < dn->dn_datablksz);
558 }
559
560 if (nblks != 0) {
561 blkid = dbuf_whichblock(dn, offset);
562 for (int i = 0; i < nblks; i++)
563 dbuf_prefetch(dn, blkid + i, ZIO_PRIORITY_SYNC_READ);
564 }
565
566 rw_exit(&dn->dn_struct_rwlock);
567
568 dnode_rele(dn, FTAG);
569 }
570
571 /*
572 * Get the next "chunk" of file data to free. We traverse the file from
573 * the end so that the file gets shorter over time (if we crashes in the
574 * middle, this will leave us in a better state). We find allocated file
575 * data by simply searching the allocated level 1 indirects.
576 *
577 * On input, *start should be the first offset that does not need to be
578 * freed (e.g. "offset + length"). On return, *start will be the first
579 * offset that should be freed.
580 */
581 static int
582 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum)
583 {
584 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
585 /* bytes of data covered by a level-1 indirect block */
586 uint64_t iblkrange =
587 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
588
589 ASSERT3U(minimum, <=, *start);
590
591 if (*start - minimum <= iblkrange * maxblks) {
592 *start = minimum;
593 return (0);
594 }
595 ASSERT(ISP2(iblkrange));
596
597 for (uint64_t blks = 0; *start > minimum && blks < maxblks; blks++) {
598 int err;
599
600 /*
601 * dnode_next_offset(BACKWARDS) will find an allocated L1
602 * indirect block at or before the input offset. We must
603 * decrement *start so that it is at the end of the region
604 * to search.
605 */
606 (*start)--;
607 err = dnode_next_offset(dn,
608 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
609
610 /* if there are no indirect blocks before start, we are done */
611 if (err == ESRCH) {
612 *start = minimum;
613 break;
614 } else if (err != 0) {
615 return (err);
616 }
617
618 /* set start to the beginning of this L1 indirect */
619 *start = P2ALIGN(*start, iblkrange);
620 }
621 if (*start < minimum)
622 *start = minimum;
623 return (0);
624 }
625
626 static int
627 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
628 uint64_t length)
629 {
630 uint64_t object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
631 int err;
632
633 if (offset >= object_size)
634 return (0);
635
636 if (length == DMU_OBJECT_END || offset + length > object_size)
637 length = object_size - offset;
638
639 while (length != 0) {
640 uint64_t chunk_end, chunk_begin;
641
642 chunk_end = chunk_begin = offset + length;
643
644 /* move chunk_begin backwards to the beginning of this chunk */
645 err = get_next_chunk(dn, &chunk_begin, offset);
646 if (err)
647 return (err);
648 ASSERT3U(chunk_begin, >=, offset);
649 ASSERT3U(chunk_begin, <=, chunk_end);
650
651 dmu_tx_t *tx = dmu_tx_create(os);
652 dmu_tx_hold_free(tx, dn->dn_object,
653 chunk_begin, chunk_end - chunk_begin);
654 err = dmu_tx_assign(tx, TXG_WAIT);
655 if (err) {
656 dmu_tx_abort(tx);
657 return (err);
658 }
659 dnode_free_range(dn, chunk_begin, chunk_end - chunk_begin, tx);
660 dmu_tx_commit(tx);
661
662 length -= chunk_end - chunk_begin;
663 }
664 return (0);
665 }
666
667 int
668 dmu_free_long_range(objset_t *os, uint64_t object,
669 uint64_t offset, uint64_t length)
670 {
671 dnode_t *dn;
672 int err;
673
674 err = dnode_hold(os, object, FTAG, &dn);
675 if (err != 0)
676 return (err);
677 err = dmu_free_long_range_impl(os, dn, offset, length);
678
679 /*
680 * It is important to zero out the maxblkid when freeing the entire
681 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
682 * will take the fast path, and (b) dnode_reallocate() can verify
683 * that the entire file has been freed.
684 */
685 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
686 dn->dn_maxblkid = 0;
687
688 dnode_rele(dn, FTAG);
689 return (err);
690 }
691
692 int
693 dmu_free_long_object(objset_t *os, uint64_t object)
694 {
695 dmu_tx_t *tx;
696 int err;
697
698 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
699 if (err != 0)
700 return (err);
701
702 tx = dmu_tx_create(os);
703 dmu_tx_hold_bonus(tx, object);
704 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
705 err = dmu_tx_assign(tx, TXG_WAIT);
706 if (err == 0) {
707 err = dmu_object_free(os, object, tx);
708 dmu_tx_commit(tx);
709 } else {
710 dmu_tx_abort(tx);
711 }
712
713 return (err);
714 }
715
716 int
717 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
718 uint64_t size, dmu_tx_t *tx)
719 {
720 dnode_t *dn;
721 int err = dnode_hold(os, object, FTAG, &dn);
722 if (err)
723 return (err);
724 ASSERT(offset < UINT64_MAX);
725 ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
726 dnode_free_range(dn, offset, size, tx);
727 dnode_rele(dn, FTAG);
728 return (0);
729 }
730
731 int
732 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
733 void *buf, uint32_t flags)
734 {
735 dnode_t *dn;
736 dmu_buf_t **dbp;
737 int numbufs, err;
738
739 err = dnode_hold(os, object, FTAG, &dn);
740 if (err)
741 return (err);
742
743 /*
744 * Deal with odd block sizes, where there can't be data past the first
745 * block. If we ever do the tail block optimization, we will need to
746 * handle that here as well.
747 */
748 if (dn->dn_maxblkid == 0) {
749 int newsz = offset > dn->dn_datablksz ? 0 :
750 MIN(size, dn->dn_datablksz - offset);
751 bzero((char *)buf + newsz, size - newsz);
752 size = newsz;
753 }
754
755 while (size > 0) {
756 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
757 int i;
758
759 /*
760 * NB: we could do this block-at-a-time, but it's nice
761 * to be reading in parallel.
762 */
763 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
764 TRUE, FTAG, &numbufs, &dbp, flags);
765 if (err)
766 break;
767
768 for (i = 0; i < numbufs; i++) {
769 int tocpy;
770 int bufoff;
771 dmu_buf_t *db = dbp[i];
772
773 ASSERT(size > 0);
774
775 bufoff = offset - db->db_offset;
776 tocpy = (int)MIN(db->db_size - bufoff, size);
777
778 bcopy((char *)db->db_data + bufoff, buf, tocpy);
779
780 offset += tocpy;
781 size -= tocpy;
782 buf = (char *)buf + tocpy;
783 }
784 dmu_buf_rele_array(dbp, numbufs, FTAG);
785 }
786 dnode_rele(dn, FTAG);
787 return (err);
788 }
789
790 void
791 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
792 const void *buf, dmu_tx_t *tx)
793 {
794 dmu_buf_t **dbp;
795 int numbufs, i;
796
797 if (size == 0)
798 return;
799
800 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
801 FALSE, FTAG, &numbufs, &dbp));
802
803 for (i = 0; i < numbufs; i++) {
804 int tocpy;
805 int bufoff;
806 dmu_buf_t *db = dbp[i];
807
808 ASSERT(size > 0);
809
810 bufoff = offset - db->db_offset;
811 tocpy = (int)MIN(db->db_size - bufoff, size);
812
813 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
814
815 if (tocpy == db->db_size)
816 dmu_buf_will_fill(db, tx);
817 else
818 dmu_buf_will_dirty(db, tx);
819
820 bcopy(buf, (char *)db->db_data + bufoff, tocpy);
821
822 if (tocpy == db->db_size)
823 dmu_buf_fill_done(db, tx);
824
825 offset += tocpy;
826 size -= tocpy;
827 buf = (char *)buf + tocpy;
828 }
829 dmu_buf_rele_array(dbp, numbufs, FTAG);
830 }
831
832 void
833 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
834 dmu_tx_t *tx)
835 {
836 dmu_buf_t **dbp;
837 int numbufs, i;
838
839 if (size == 0)
840 return;
841
842 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
843 FALSE, FTAG, &numbufs, &dbp));
844
845 for (i = 0; i < numbufs; i++) {
846 dmu_buf_t *db = dbp[i];
847
848 dmu_buf_will_not_fill(db, tx);
849 }
850 dmu_buf_rele_array(dbp, numbufs, FTAG);
851 }
852
853 /*
854 * DMU support for xuio
855 */
856 kstat_t *xuio_ksp = NULL;
857
858 int
859 dmu_xuio_init(xuio_t *xuio, int nblk)
860 {
861 dmu_xuio_t *priv;
862 uio_t *uio = &xuio->xu_uio;
863
864 uio->uio_iovcnt = nblk;
865 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
866
867 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
868 priv->cnt = nblk;
869 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
870 priv->iovp = uio->uio_iov;
871 XUIO_XUZC_PRIV(xuio) = priv;
872
873 if (XUIO_XUZC_RW(xuio) == UIO_READ)
874 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
875 else
876 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
877
878 return (0);
879 }
880
881 void
882 dmu_xuio_fini(xuio_t *xuio)
883 {
884 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
885 int nblk = priv->cnt;
886
887 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
888 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
889 kmem_free(priv, sizeof (dmu_xuio_t));
890
891 if (XUIO_XUZC_RW(xuio) == UIO_READ)
892 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
893 else
894 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
895 }
896
897 /*
898 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
899 * and increase priv->next by 1.
900 */
901 int
902 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
903 {
904 struct iovec *iov;
905 uio_t *uio = &xuio->xu_uio;
906 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
907 int i = priv->next++;
908
909 ASSERT(i < priv->cnt);
910 ASSERT(off + n <= arc_buf_size(abuf));
911 iov = uio->uio_iov + i;
912 iov->iov_base = (char *)abuf->b_data + off;
913 iov->iov_len = n;
914 priv->bufs[i] = abuf;
915 return (0);
916 }
917
918 int
919 dmu_xuio_cnt(xuio_t *xuio)
920 {
921 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
922 return (priv->cnt);
923 }
924
925 arc_buf_t *
926 dmu_xuio_arcbuf(xuio_t *xuio, int i)
927 {
928 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
929
930 ASSERT(i < priv->cnt);
931 return (priv->bufs[i]);
932 }
933
934 void
935 dmu_xuio_clear(xuio_t *xuio, int i)
936 {
937 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
938
939 ASSERT(i < priv->cnt);
940 priv->bufs[i] = NULL;
941 }
942
943 static void
944 xuio_stat_init(void)
945 {
946 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
947 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
948 KSTAT_FLAG_VIRTUAL);
949 if (xuio_ksp != NULL) {
950 xuio_ksp->ks_data = &xuio_stats;
951 kstat_install(xuio_ksp);
952 }
953 }
954
955 static void
956 xuio_stat_fini(void)
957 {
958 if (xuio_ksp != NULL) {
959 kstat_delete(xuio_ksp);
960 xuio_ksp = NULL;
961 }
962 }
963
964 void
965 xuio_stat_wbuf_copied()
966 {
967 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
968 }
969
970 void
971 xuio_stat_wbuf_nocopy()
972 {
973 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
974 }
975
976 #ifdef _KERNEL
977 int
978 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
979 {
980 dmu_buf_t **dbp;
981 int numbufs, i, err;
982 xuio_t *xuio = NULL;
983
984 /*
985 * NB: we could do this block-at-a-time, but it's nice
986 * to be reading in parallel.
987 */
988 err = dmu_buf_hold_array(os, object, uio->uio_loffset, size, TRUE, FTAG,
989 &numbufs, &dbp);
990 if (err)
991 return (err);
992
993 if (uio->uio_extflg == UIO_XUIO)
994 xuio = (xuio_t *)uio;
995
996 for (i = 0; i < numbufs; i++) {
997 int tocpy;
998 int bufoff;
999 dmu_buf_t *db = dbp[i];
1000
1001 ASSERT(size > 0);
1002
1003 bufoff = uio->uio_loffset - db->db_offset;
1004 tocpy = (int)MIN(db->db_size - bufoff, size);
1005
1006 if (xuio) {
1007 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
1008 arc_buf_t *dbuf_abuf = dbi->db_buf;
1009 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
1010 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
1011 if (!err) {
1012 uio->uio_resid -= tocpy;
1013 uio->uio_loffset += tocpy;
1014 }
1015
1016 if (abuf == dbuf_abuf)
1017 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
1018 else
1019 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
1020 } else {
1021 err = uiomove((char *)db->db_data + bufoff, tocpy,
1022 UIO_READ, uio);
1023 }
1024 if (err)
1025 break;
1026
1027 size -= tocpy;
1028 }
1029 dmu_buf_rele_array(dbp, numbufs, FTAG);
1030
1031 return (err);
1032 }
1033
1034 static int
1035 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1036 {
1037 dmu_buf_t **dbp;
1038 int numbufs;
1039 int err = 0;
1040 int i;
1041
1042 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1043 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1044 if (err)
1045 return (err);
1046
1047 for (i = 0; i < numbufs; i++) {
1048 int tocpy;
1049 int bufoff;
1050 dmu_buf_t *db = dbp[i];
1051
1052 ASSERT(size > 0);
1053
1054 bufoff = uio->uio_loffset - db->db_offset;
1055 tocpy = (int)MIN(db->db_size - bufoff, size);
1056
1057 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1058
1059 if (tocpy == db->db_size)
1060 dmu_buf_will_fill(db, tx);
1061 else
1062 dmu_buf_will_dirty(db, tx);
1063
1064 /*
1065 * XXX uiomove could block forever (eg. nfs-backed
1066 * pages). There needs to be a uiolockdown() function
1067 * to lock the pages in memory, so that uiomove won't
1068 * block.
1069 */
1070 err = uiomove((char *)db->db_data + bufoff, tocpy,
1071 UIO_WRITE, uio);
1072
1073 if (tocpy == db->db_size)
1074 dmu_buf_fill_done(db, tx);
1075
1076 if (err)
1077 break;
1078
1079 size -= tocpy;
1080 }
1081
1082 dmu_buf_rele_array(dbp, numbufs, FTAG);
1083 return (err);
1084 }
1085
1086 int
1087 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1088 dmu_tx_t *tx)
1089 {
1090 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1091 dnode_t *dn;
1092 int err;
1093
1094 if (size == 0)
1095 return (0);
1096
1097 DB_DNODE_ENTER(db);
1098 dn = DB_DNODE(db);
1099 err = dmu_write_uio_dnode(dn, uio, size, tx);
1100 DB_DNODE_EXIT(db);
1101
1102 return (err);
1103 }
1104
1105 int
1106 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1107 dmu_tx_t *tx)
1108 {
1109 dnode_t *dn;
1110 int err;
1111
1112 if (size == 0)
1113 return (0);
1114
1115 err = dnode_hold(os, object, FTAG, &dn);
1116 if (err)
1117 return (err);
1118
1119 err = dmu_write_uio_dnode(dn, uio, size, tx);
1120
1121 dnode_rele(dn, FTAG);
1122
1123 return (err);
1124 }
1125
1126 int
1127 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1128 page_t *pp, dmu_tx_t *tx)
1129 {
1130 dmu_buf_t **dbp;
1131 int numbufs, i;
1132 int err;
1133
1134 if (size == 0)
1135 return (0);
1136
1137 err = dmu_buf_hold_array(os, object, offset, size,
1138 FALSE, FTAG, &numbufs, &dbp);
1139 if (err)
1140 return (err);
1141
1142 for (i = 0; i < numbufs; i++) {
1143 int tocpy, copied, thiscpy;
1144 int bufoff;
1145 dmu_buf_t *db = dbp[i];
1146 caddr_t va;
1147
1148 ASSERT(size > 0);
1149 ASSERT3U(db->db_size, >=, PAGESIZE);
1150
1151 bufoff = offset - db->db_offset;
1152 tocpy = (int)MIN(db->db_size - bufoff, size);
1153
1154 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1155
1156 if (tocpy == db->db_size)
1157 dmu_buf_will_fill(db, tx);
1158 else
1159 dmu_buf_will_dirty(db, tx);
1160
1161 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1162 ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff);
1163 thiscpy = MIN(PAGESIZE, tocpy - copied);
1164 va = zfs_map_page(pp, S_READ);
1165 bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1166 zfs_unmap_page(pp, va);
1167 pp = pp->p_next;
1168 bufoff += PAGESIZE;
1169 }
1170
1171 if (tocpy == db->db_size)
1172 dmu_buf_fill_done(db, tx);
1173
1174 offset += tocpy;
1175 size -= tocpy;
1176 }
1177 dmu_buf_rele_array(dbp, numbufs, FTAG);
1178 return (err);
1179 }
1180 #endif
1181
1182 /*
1183 * Allocate a loaned anonymous arc buffer.
1184 */
1185 arc_buf_t *
1186 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1187 {
1188 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1189
1190 return (arc_loan_buf(db->db_objset->os_spa, size));
1191 }
1192
1193 /*
1194 * Free a loaned arc buffer.
1195 */
1196 void
1197 dmu_return_arcbuf(arc_buf_t *buf)
1198 {
1199 arc_return_buf(buf, FTAG);
1200 VERIFY(arc_buf_remove_ref(buf, FTAG));
1201 }
1202
1203 /*
1204 * When possible directly assign passed loaned arc buffer to a dbuf.
1205 * If this is not possible copy the contents of passed arc buf via
1206 * dmu_write().
1207 */
1208 void
1209 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1210 dmu_tx_t *tx)
1211 {
1212 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1213 dnode_t *dn;
1214 dmu_buf_impl_t *db;
1215 uint32_t blksz = (uint32_t)arc_buf_size(buf);
1216 uint64_t blkid;
1217
1218 DB_DNODE_ENTER(dbuf);
1219 dn = DB_DNODE(dbuf);
1220 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1221 blkid = dbuf_whichblock(dn, offset);
1222 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1223 rw_exit(&dn->dn_struct_rwlock);
1224 DB_DNODE_EXIT(dbuf);
1225
1226 if (offset == db->db.db_offset && blksz == db->db.db_size) {
1227 dbuf_assign_arcbuf(db, buf, tx);
1228 dbuf_rele(db, FTAG);
1229 } else {
1230 objset_t *os;
1231 uint64_t object;
1232
1233 DB_DNODE_ENTER(dbuf);
1234 dn = DB_DNODE(dbuf);
1235 os = dn->dn_objset;
1236 object = dn->dn_object;
1237 DB_DNODE_EXIT(dbuf);
1238
1239 dbuf_rele(db, FTAG);
1240 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1241 dmu_return_arcbuf(buf);
1242 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1243 }
1244 }
1245
1246 typedef struct {
1247 dbuf_dirty_record_t *dsa_dr;
1248 dmu_sync_cb_t *dsa_done;
1249 zgd_t *dsa_zgd;
1250 dmu_tx_t *dsa_tx;
1251 } dmu_sync_arg_t;
1252
1253 /* ARGSUSED */
1254 static void
1255 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1256 {
1257 dmu_sync_arg_t *dsa = varg;
1258 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1259 blkptr_t *bp = zio->io_bp;
1260
1261 if (zio->io_error == 0) {
1262 if (BP_IS_HOLE(bp)) {
1263 /*
1264 * A block of zeros may compress to a hole, but the
1265 * block size still needs to be known for replay.
1266 */
1267 BP_SET_LSIZE(bp, db->db_size);
1268 } else {
1269 ASSERT(BP_GET_LEVEL(bp) == 0);
1270 bp->blk_fill = 1;
1271 }
1272 }
1273 }
1274
1275 static void
1276 dmu_sync_late_arrival_ready(zio_t *zio)
1277 {
1278 dmu_sync_ready(zio, NULL, zio->io_private);
1279 }
1280
1281 /* ARGSUSED */
1282 static void
1283 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1284 {
1285 dmu_sync_arg_t *dsa = varg;
1286 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1287 dmu_buf_impl_t *db = dr->dr_dbuf;
1288
1289 mutex_enter(&db->db_mtx);
1290 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1291 if (zio->io_error == 0) {
1292 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1293 if (dr->dt.dl.dr_nopwrite) {
1294 blkptr_t *bp = zio->io_bp;
1295 blkptr_t *bp_orig = &zio->io_bp_orig;
1296 uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1297
1298 ASSERT(BP_EQUAL(bp, bp_orig));
1299 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1300 ASSERT(zio_checksum_table[chksum].ci_dedup);
1301 }
1302 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1303 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1304 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1305 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by))
1306 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1307 } else {
1308 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1309 }
1310 cv_broadcast(&db->db_changed);
1311 mutex_exit(&db->db_mtx);
1312
1313 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1314
1315 kmem_free(dsa, sizeof (*dsa));
1316 }
1317
1318 static void
1319 dmu_sync_late_arrival_done(zio_t *zio)
1320 {
1321 blkptr_t *bp = zio->io_bp;
1322 dmu_sync_arg_t *dsa = zio->io_private;
1323 blkptr_t *bp_orig = &zio->io_bp_orig;
1324
1325 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1326 /*
1327 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE)
1328 * then there is nothing to do here. Otherwise, free the
1329 * newly allocated block in this txg.
1330 */
1331 if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
1332 ASSERT(BP_EQUAL(bp, bp_orig));
1333 } else {
1334 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1335 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1336 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1337 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1338 }
1339 }
1340
1341 dmu_tx_commit(dsa->dsa_tx);
1342
1343 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1344
1345 kmem_free(dsa, sizeof (*dsa));
1346 }
1347
1348 static int
1349 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1350 zio_prop_t *zp, zbookmark_t *zb)
1351 {
1352 dmu_sync_arg_t *dsa;
1353 dmu_tx_t *tx;
1354
1355 tx = dmu_tx_create(os);
1356 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1357 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1358 dmu_tx_abort(tx);
1359 /* Make zl_get_data do txg_waited_synced() */
1360 return (SET_ERROR(EIO));
1361 }
1362
1363 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1364 dsa->dsa_dr = NULL;
1365 dsa->dsa_done = done;
1366 dsa->dsa_zgd = zgd;
1367 dsa->dsa_tx = tx;
1368
1369 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1370 zgd->zgd_db->db_data, zgd->zgd_db->db_size, zp,
1371 dmu_sync_late_arrival_ready, NULL, dmu_sync_late_arrival_done, dsa,
1372 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1373
1374 return (0);
1375 }
1376
1377 /*
1378 * Intent log support: sync the block associated with db to disk.
1379 * N.B. and XXX: the caller is responsible for making sure that the
1380 * data isn't changing while dmu_sync() is writing it.
1381 *
1382 * Return values:
1383 *
1384 * EEXIST: this txg has already been synced, so there's nothing to do.
1385 * The caller should not log the write.
1386 *
1387 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1388 * The caller should not log the write.
1389 *
1390 * EALREADY: this block is already in the process of being synced.
1391 * The caller should track its progress (somehow).
1392 *
1393 * EIO: could not do the I/O.
1394 * The caller should do a txg_wait_synced().
1395 *
1396 * 0: the I/O has been initiated.
1397 * The caller should log this blkptr in the done callback.
1398 * It is possible that the I/O will fail, in which case
1399 * the error will be reported to the done callback and
1400 * propagated to pio from zio_done().
1401 */
1402 int
1403 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1404 {
1405 blkptr_t *bp = zgd->zgd_bp;
1406 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1407 objset_t *os = db->db_objset;
1408 dsl_dataset_t *ds = os->os_dsl_dataset;
1409 dbuf_dirty_record_t *dr;
1410 dmu_sync_arg_t *dsa;
1411 zbookmark_t zb;
1412 zio_prop_t zp;
1413 dnode_t *dn;
1414
1415 ASSERT(pio != NULL);
1416 ASSERT(txg != 0);
1417
1418 SET_BOOKMARK(&zb, ds->ds_object,
1419 db->db.db_object, db->db_level, db->db_blkid);
1420
1421 DB_DNODE_ENTER(db);
1422 dn = DB_DNODE(db);
1423 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1424 DB_DNODE_EXIT(db);
1425
1426 /*
1427 * If we're frozen (running ziltest), we always need to generate a bp.
1428 */
1429 if (txg > spa_freeze_txg(os->os_spa))
1430 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1431
1432 /*
1433 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1434 * and us. If we determine that this txg is not yet syncing,
1435 * but it begins to sync a moment later, that's OK because the
1436 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1437 */
1438 mutex_enter(&db->db_mtx);
1439
1440 if (txg <= spa_last_synced_txg(os->os_spa)) {
1441 /*
1442 * This txg has already synced. There's nothing to do.
1443 */
1444 mutex_exit(&db->db_mtx);
1445 return (SET_ERROR(EEXIST));
1446 }
1447
1448 if (txg <= spa_syncing_txg(os->os_spa)) {
1449 /*
1450 * This txg is currently syncing, so we can't mess with
1451 * the dirty record anymore; just write a new log block.
1452 */
1453 mutex_exit(&db->db_mtx);
1454 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1455 }
1456
1457 dr = db->db_last_dirty;
1458 while (dr && dr->dr_txg != txg)
1459 dr = dr->dr_next;
1460
1461 if (dr == NULL) {
1462 /*
1463 * There's no dr for this dbuf, so it must have been freed.
1464 * There's no need to log writes to freed blocks, so we're done.
1465 */
1466 mutex_exit(&db->db_mtx);
1467 return (SET_ERROR(ENOENT));
1468 }
1469
1470 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
1471
1472 /*
1473 * Assume the on-disk data is X, the current syncing data is Y,
1474 * and the current in-memory data is Z (currently in dmu_sync).
1475 * X and Z are identical but Y is has been modified. Normally,
1476 * when X and Z are the same we will perform a nopwrite but if Y
1477 * is different we must disable nopwrite since the resulting write
1478 * of Y to disk can free the block containing X. If we allowed a
1479 * nopwrite to occur the block pointing to Z would reference a freed
1480 * block. Since this is a rare case we simplify this by disabling
1481 * nopwrite if the current dmu_sync-ing dbuf has been modified in
1482 * a previous transaction.
1483 */
1484 if (dr->dr_next)
1485 zp.zp_nopwrite = B_FALSE;
1486
1487 ASSERT(dr->dr_txg == txg);
1488 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1489 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1490 /*
1491 * We have already issued a sync write for this buffer,
1492 * or this buffer has already been synced. It could not
1493 * have been dirtied since, or we would have cleared the state.
1494 */
1495 mutex_exit(&db->db_mtx);
1496 return (SET_ERROR(EALREADY));
1497 }
1498
1499 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1500 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1501 mutex_exit(&db->db_mtx);
1502
1503 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1504 dsa->dsa_dr = dr;
1505 dsa->dsa_done = done;
1506 dsa->dsa_zgd = zgd;
1507 dsa->dsa_tx = NULL;
1508
1509 zio_nowait(arc_write(pio, os->os_spa, txg,
1510 bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
1511 DBUF_IS_L2COMPRESSIBLE(db), &zp, dmu_sync_ready,
1512 NULL, dmu_sync_done, dsa, ZIO_PRIORITY_SYNC_WRITE,
1513 ZIO_FLAG_CANFAIL, &zb));
1514
1515 return (0);
1516 }
1517
1518 int
1519 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1520 dmu_tx_t *tx)
1521 {
1522 dnode_t *dn;
1523 int err;
1524
1525 err = dnode_hold(os, object, FTAG, &dn);
1526 if (err)
1527 return (err);
1528 err = dnode_set_blksz(dn, size, ibs, tx);
1529 dnode_rele(dn, FTAG);
1530 return (err);
1531 }
1532
1533 void
1534 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1535 dmu_tx_t *tx)
1536 {
1537 dnode_t *dn;
1538
1539 /* XXX assumes dnode_hold will not get an i/o error */
1540 (void) dnode_hold(os, object, FTAG, &dn);
1541 ASSERT(checksum < ZIO_CHECKSUM_FUNCTIONS);
1542 dn->dn_checksum = checksum;
1543 dnode_setdirty(dn, tx);
1544 dnode_rele(dn, FTAG);
1545 }
1546
1547 void
1548 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1549 dmu_tx_t *tx)
1550 {
1551 dnode_t *dn;
1552
1553 /* XXX assumes dnode_hold will not get an i/o error */
1554 (void) dnode_hold(os, object, FTAG, &dn);
1555 ASSERT(compress < ZIO_COMPRESS_FUNCTIONS);
1556 dn->dn_compress = compress;
1557 dnode_setdirty(dn, tx);
1558 dnode_rele(dn, FTAG);
1559 }
1560
1561 int zfs_mdcomp_disable = 0;
1562
1563 /*
1564 * When the "redundant_metadata" property is set to "most", only indirect
1565 * blocks of this level and higher will have an additional ditto block.
1566 */
1567 int zfs_redundant_metadata_most_ditto_level = 2;
1568
1569 void
1570 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
1571 {
1572 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
1573 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
1574 (wp & WP_SPILL));
1575 enum zio_checksum checksum = os->os_checksum;
1576 enum zio_compress compress = os->os_compress;
1577 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
1578 boolean_t dedup = B_FALSE;
1579 boolean_t nopwrite = B_FALSE;
1580 boolean_t dedup_verify = os->os_dedup_verify;
1581 int copies = os->os_copies;
1582
1583 /*
1584 * We maintain different write policies for each of the following
1585 * types of data:
1586 * 1. metadata
1587 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1588 * 3. all other level 0 blocks
1589 */
1590 if (ismd) {
1591 /*
1592 * XXX -- we should design a compression algorithm
1593 * that specializes in arrays of bps.
1594 */
1595 boolean_t lz4_ac = spa_feature_is_active(os->os_spa,
1596 SPA_FEATURE_LZ4_COMPRESS);
1597
1598 if (zfs_mdcomp_disable) {
1599 compress = ZIO_COMPRESS_EMPTY;
1600 } else if (lz4_ac) {
1601 compress = ZIO_COMPRESS_LZ4;
1602 } else {
1603 compress = ZIO_COMPRESS_LZJB;
1604 }
1605
1606 /*
1607 * Metadata always gets checksummed. If the data
1608 * checksum is multi-bit correctable, and it's not a
1609 * ZBT-style checksum, then it's suitable for metadata
1610 * as well. Otherwise, the metadata checksum defaults
1611 * to fletcher4.
1612 */
1613 if (zio_checksum_table[checksum].ci_correctable < 1 ||
1614 zio_checksum_table[checksum].ci_eck)
1615 checksum = ZIO_CHECKSUM_FLETCHER_4;
1616
1617 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL ||
1618 (os->os_redundant_metadata ==
1619 ZFS_REDUNDANT_METADATA_MOST &&
1620 (level >= zfs_redundant_metadata_most_ditto_level ||
1621 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))))
1622 copies++;
1623 } else if (wp & WP_NOFILL) {
1624 ASSERT(level == 0);
1625
1626 /*
1627 * If we're writing preallocated blocks, we aren't actually
1628 * writing them so don't set any policy properties. These
1629 * blocks are currently only used by an external subsystem
1630 * outside of zfs (i.e. dump) and not written by the zio
1631 * pipeline.
1632 */
1633 compress = ZIO_COMPRESS_OFF;
1634 checksum = ZIO_CHECKSUM_NOPARITY;
1635 } else {
1636 compress = zio_compress_select(dn->dn_compress, compress);
1637
1638 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
1639 zio_checksum_select(dn->dn_checksum, checksum) :
1640 dedup_checksum;
1641
1642 /*
1643 * Determine dedup setting. If we are in dmu_sync(),
1644 * we won't actually dedup now because that's all
1645 * done in syncing context; but we do want to use the
1646 * dedup checkum. If the checksum is not strong
1647 * enough to ensure unique signatures, force
1648 * dedup_verify.
1649 */
1650 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
1651 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
1652 if (!zio_checksum_table[checksum].ci_dedup)
1653 dedup_verify = B_TRUE;
1654 }
1655
1656 /*
1657 * Enable nopwrite if we have a cryptographically secure
1658 * checksum that has no known collisions (i.e. SHA-256)
1659 * and compression is enabled. We don't enable nopwrite if
1660 * dedup is enabled as the two features are mutually exclusive.
1661 */
1662 nopwrite = (!dedup && zio_checksum_table[checksum].ci_dedup &&
1663 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
1664 }
1665
1666 zp->zp_checksum = checksum;
1667 zp->zp_compress = compress;
1668 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
1669 zp->zp_level = level;
1670 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
1671 zp->zp_dedup = dedup;
1672 zp->zp_dedup_verify = dedup && dedup_verify;
1673 zp->zp_nopwrite = nopwrite;
1674 }
1675
1676 int
1677 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
1678 {
1679 dnode_t *dn;
1680 int i, err;
1681
1682 err = dnode_hold(os, object, FTAG, &dn);
1683 if (err)
1684 return (err);
1685 /*
1686 * Sync any current changes before
1687 * we go trundling through the block pointers.
1688 */
1689 for (i = 0; i < TXG_SIZE; i++) {
1690 if (list_link_active(&dn->dn_dirty_link[i]))
1691 break;
1692 }
1693 if (i != TXG_SIZE) {
1694 dnode_rele(dn, FTAG);
1695 txg_wait_synced(dmu_objset_pool(os), 0);
1696 err = dnode_hold(os, object, FTAG, &dn);
1697 if (err)
1698 return (err);
1699 }
1700
1701 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
1702 dnode_rele(dn, FTAG);
1703
1704 return (err);
1705 }
1706
1707 void
1708 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
1709 {
1710 dnode_phys_t *dnp;
1711
1712 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1713 mutex_enter(&dn->dn_mtx);
1714
1715 dnp = dn->dn_phys;
1716
1717 doi->doi_data_block_size = dn->dn_datablksz;
1718 doi->doi_metadata_block_size = dn->dn_indblkshift ?
1719 1ULL << dn->dn_indblkshift : 0;
1720 doi->doi_type = dn->dn_type;
1721 doi->doi_bonus_type = dn->dn_bonustype;
1722 doi->doi_bonus_size = dn->dn_bonuslen;
1723 doi->doi_indirection = dn->dn_nlevels;
1724 doi->doi_checksum = dn->dn_checksum;
1725 doi->doi_compress = dn->dn_compress;
1726 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
1727 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
1728 doi->doi_fill_count = 0;
1729 for (int i = 0; i < dnp->dn_nblkptr; i++)
1730 doi->doi_fill_count += dnp->dn_blkptr[i].blk_fill;
1731
1732 mutex_exit(&dn->dn_mtx);
1733 rw_exit(&dn->dn_struct_rwlock);
1734 }
1735
1736 /*
1737 * Get information on a DMU object.
1738 * If doi is NULL, just indicates whether the object exists.
1739 */
1740 int
1741 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
1742 {
1743 dnode_t *dn;
1744 int err = dnode_hold(os, object, FTAG, &dn);
1745
1746 if (err)
1747 return (err);
1748
1749 if (doi != NULL)
1750 dmu_object_info_from_dnode(dn, doi);
1751
1752 dnode_rele(dn, FTAG);
1753 return (0);
1754 }
1755
1756 /*
1757 * As above, but faster; can be used when you have a held dbuf in hand.
1758 */
1759 void
1760 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
1761 {
1762 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1763
1764 DB_DNODE_ENTER(db);
1765 dmu_object_info_from_dnode(DB_DNODE(db), doi);
1766 DB_DNODE_EXIT(db);
1767 }
1768
1769 /*
1770 * Faster still when you only care about the size.
1771 * This is specifically optimized for zfs_getattr().
1772 */
1773 void
1774 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
1775 u_longlong_t *nblk512)
1776 {
1777 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1778 dnode_t *dn;
1779
1780 DB_DNODE_ENTER(db);
1781 dn = DB_DNODE(db);
1782
1783 *blksize = dn->dn_datablksz;
1784 /* add 1 for dnode space */
1785 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
1786 SPA_MINBLOCKSHIFT) + 1;
1787 DB_DNODE_EXIT(db);
1788 }
1789
1790 void
1791 byteswap_uint64_array(void *vbuf, size_t size)
1792 {
1793 uint64_t *buf = vbuf;
1794 size_t count = size >> 3;
1795 int i;
1796
1797 ASSERT((size & 7) == 0);
1798
1799 for (i = 0; i < count; i++)
1800 buf[i] = BSWAP_64(buf[i]);
1801 }
1802
1803 void
1804 byteswap_uint32_array(void *vbuf, size_t size)
1805 {
1806 uint32_t *buf = vbuf;
1807 size_t count = size >> 2;
1808 int i;
1809
1810 ASSERT((size & 3) == 0);
1811
1812 for (i = 0; i < count; i++)
1813 buf[i] = BSWAP_32(buf[i]);
1814 }
1815
1816 void
1817 byteswap_uint16_array(void *vbuf, size_t size)
1818 {
1819 uint16_t *buf = vbuf;
1820 size_t count = size >> 1;
1821 int i;
1822
1823 ASSERT((size & 1) == 0);
1824
1825 for (i = 0; i < count; i++)
1826 buf[i] = BSWAP_16(buf[i]);
1827 }
1828
1829 /* ARGSUSED */
1830 void
1831 byteswap_uint8_array(void *vbuf, size_t size)
1832 {
1833 }
1834
1835 void
1836 dmu_init(void)
1837 {
1838 zfs_dbgmsg_init();
1839 sa_cache_init();
1840 xuio_stat_init();
1841 dmu_objset_init();
1842 dnode_init();
1843 dbuf_init();
1844 zfetch_init();
1845 l2arc_init();
1846 arc_init();
1847 }
1848
1849 void
1850 dmu_fini(void)
1851 {
1852 arc_fini(); /* arc depends on l2arc, so arc must go first */
1853 l2arc_fini();
1854 zfetch_fini();
1855 dbuf_fini();
1856 dnode_fini();
1857 dmu_objset_fini();
1858 xuio_stat_fini();
1859 sa_cache_fini();
1860 zfs_dbgmsg_fini();
1861 }