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 /* Copyright (c) 2013 by Saso Kiselkov. All rights reserved. */
26 /* Copyright (c) 2013, Joyent, Inc. 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 dmu_buf_t **dbp;
375 uint64_t blkid, nblks, i;
376 uint32_t dbuf_flags;
377 int err;
378 zio_t *zio;
379
380 ASSERT(length <= DMU_MAX_ACCESS);
381
382 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT;
383 if (flags & DMU_READ_NO_PREFETCH || length > zfetch_array_rd_sz)
384 dbuf_flags |= DB_RF_NOPREFETCH;
385
386 rw_enter(&dn->dn_struct_rwlock, RW_READER);
387 if (dn->dn_datablkshift) {
388 int blkshift = dn->dn_datablkshift;
389 nblks = (P2ROUNDUP(offset+length, 1ULL<<blkshift) -
390 P2ALIGN(offset, 1ULL<<blkshift)) >> blkshift;
391 } else {
392 if (offset + length > dn->dn_datablksz) {
393 zfs_panic_recover("zfs: accessing past end of object "
394 "%llx/%llx (size=%u access=%llu+%llu)",
395 (longlong_t)dn->dn_objset->
396 os_dsl_dataset->ds_object,
397 (longlong_t)dn->dn_object, dn->dn_datablksz,
398 (longlong_t)offset, (longlong_t)length);
399 rw_exit(&dn->dn_struct_rwlock);
400 return (SET_ERROR(EIO));
401 }
402 nblks = 1;
403 }
404 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
405
406 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
407 blkid = dbuf_whichblock(dn, offset);
408 for (i = 0; i < nblks; i++) {
409 dmu_buf_impl_t *db = dbuf_hold(dn, blkid+i, tag);
410 if (db == NULL) {
411 rw_exit(&dn->dn_struct_rwlock);
412 dmu_buf_rele_array(dbp, nblks, tag);
413 zio_nowait(zio);
414 return (SET_ERROR(EIO));
415 }
416 /* initiate async i/o */
417 if (read) {
418 (void) dbuf_read(db, zio, dbuf_flags);
419 }
420 dbp[i] = &db->db;
421 }
422 rw_exit(&dn->dn_struct_rwlock);
423
424 /* wait for async i/o */
425 err = zio_wait(zio);
426 if (err) {
427 dmu_buf_rele_array(dbp, nblks, tag);
428 return (err);
429 }
430
431 /* wait for other io to complete */
432 if (read) {
433 for (i = 0; i < nblks; i++) {
434 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
435 mutex_enter(&db->db_mtx);
436 while (db->db_state == DB_READ ||
437 db->db_state == DB_FILL)
438 cv_wait(&db->db_changed, &db->db_mtx);
439 if (db->db_state == DB_UNCACHED)
440 err = SET_ERROR(EIO);
441 mutex_exit(&db->db_mtx);
442 if (err) {
443 dmu_buf_rele_array(dbp, nblks, tag);
444 return (err);
445 }
446 }
447 }
448
449 *numbufsp = nblks;
450 *dbpp = dbp;
451 return (0);
452 }
453
454 static int
455 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
456 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
457 {
458 dnode_t *dn;
459 int err;
460
461 err = dnode_hold(os, object, FTAG, &dn);
462 if (err)
463 return (err);
464
465 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
466 numbufsp, dbpp, DMU_READ_PREFETCH);
467
468 dnode_rele(dn, FTAG);
469
470 return (err);
471 }
472
473 int
474 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
475 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
476 {
477 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
478 dnode_t *dn;
479 int err;
480
481 DB_DNODE_ENTER(db);
482 dn = DB_DNODE(db);
483 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
484 numbufsp, dbpp, DMU_READ_PREFETCH);
485 DB_DNODE_EXIT(db);
486
487 return (err);
488 }
489
490 void
491 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
492 {
493 int i;
494 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
495
496 if (numbufs == 0)
497 return;
498
499 for (i = 0; i < numbufs; i++) {
500 if (dbp[i])
501 dbuf_rele(dbp[i], tag);
502 }
503
504 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
505 }
506
507 /*
508 * Issue prefetch i/os for the given blocks.
509 *
510 * Note: The assumption is that we *know* these blocks will be needed
511 * almost immediately. Therefore, the prefetch i/os will be issued at
512 * ZIO_PRIORITY_SYNC_READ
513 *
514 * Note: indirect blocks and other metadata will be read synchronously,
515 * causing this function to block if they are not already cached.
516 */
517 void
518 dmu_prefetch(objset_t *os, uint64_t object, uint64_t offset, uint64_t len)
519 {
520 dnode_t *dn;
521 uint64_t blkid;
522 int nblks, err;
523
524 if (zfs_prefetch_disable)
525 return;
526
527 if (len == 0) { /* they're interested in the bonus buffer */
528 dn = DMU_META_DNODE(os);
529
530 if (object == 0 || object >= DN_MAX_OBJECT)
531 return;
532
533 rw_enter(&dn->dn_struct_rwlock, RW_READER);
534 blkid = dbuf_whichblock(dn, object * sizeof (dnode_phys_t));
535 dbuf_prefetch(dn, blkid, ZIO_PRIORITY_SYNC_READ);
536 rw_exit(&dn->dn_struct_rwlock);
537 return;
538 }
539
540 /*
541 * XXX - Note, if the dnode for the requested object is not
542 * already cached, we will do a *synchronous* read in the
543 * dnode_hold() call. The same is true for any indirects.
544 */
545 err = dnode_hold(os, object, FTAG, &dn);
546 if (err != 0)
547 return;
548
549 rw_enter(&dn->dn_struct_rwlock, RW_READER);
550 if (dn->dn_datablkshift) {
551 int blkshift = dn->dn_datablkshift;
552 nblks = (P2ROUNDUP(offset + len, 1 << blkshift) -
553 P2ALIGN(offset, 1 << blkshift)) >> blkshift;
554 } else {
555 nblks = (offset < dn->dn_datablksz);
556 }
557
558 if (nblks != 0) {
559 blkid = dbuf_whichblock(dn, offset);
560 for (int i = 0; i < nblks; i++)
561 dbuf_prefetch(dn, blkid + i, ZIO_PRIORITY_SYNC_READ);
562 }
563
564 rw_exit(&dn->dn_struct_rwlock);
565
566 dnode_rele(dn, FTAG);
567 }
568
569 /*
570 * Get the next "chunk" of file data to free. We traverse the file from
571 * the end so that the file gets shorter over time (if we crashes in the
572 * middle, this will leave us in a better state). We find allocated file
573 * data by simply searching the allocated level 1 indirects.
574 *
575 * On input, *start should be the first offset that does not need to be
576 * freed (e.g. "offset + length"). On return, *start will be the first
577 * offset that should be freed.
578 */
579 static int
580 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum)
581 {
582 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
583 /* bytes of data covered by a level-1 indirect block */
584 uint64_t iblkrange =
585 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
586
587 ASSERT3U(minimum, <=, *start);
588
589 if (*start - minimum <= iblkrange * maxblks) {
590 *start = minimum;
591 return (0);
592 }
593 ASSERT(ISP2(iblkrange));
594
595 for (uint64_t blks = 0; *start > minimum && blks < maxblks; blks++) {
596 int err;
597
598 /*
599 * dnode_next_offset(BACKWARDS) will find an allocated L1
600 * indirect block at or before the input offset. We must
601 * decrement *start so that it is at the end of the region
602 * to search.
603 */
604 (*start)--;
605 err = dnode_next_offset(dn,
606 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
607
608 /* if there are no indirect blocks before start, we are done */
609 if (err == ESRCH) {
610 *start = minimum;
611 break;
612 } else if (err != 0) {
613 return (err);
614 }
615
616 /* set start to the beginning of this L1 indirect */
617 *start = P2ALIGN(*start, iblkrange);
618 }
619 if (*start < minimum)
620 *start = minimum;
621 return (0);
622 }
623
624 static int
625 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
626 uint64_t length)
627 {
628 uint64_t object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
629 int err;
630
631 if (offset >= object_size)
632 return (0);
633
634 if (length == DMU_OBJECT_END || offset + length > object_size)
635 length = object_size - offset;
636
637 while (length != 0) {
638 uint64_t chunk_end, chunk_begin;
639
640 chunk_end = chunk_begin = offset + length;
641
642 /* move chunk_begin backwards to the beginning of this chunk */
643 err = get_next_chunk(dn, &chunk_begin, offset);
644 if (err)
645 return (err);
646 ASSERT3U(chunk_begin, >=, offset);
647 ASSERT3U(chunk_begin, <=, chunk_end);
648
649 dmu_tx_t *tx = dmu_tx_create(os);
650 dmu_tx_hold_free(tx, dn->dn_object,
651 chunk_begin, chunk_end - chunk_begin);
652 err = dmu_tx_assign(tx, TXG_WAIT);
653 if (err) {
654 dmu_tx_abort(tx);
655 return (err);
656 }
657 dnode_free_range(dn, chunk_begin, chunk_end - chunk_begin, tx);
658 dmu_tx_commit(tx);
659
660 length -= chunk_end - chunk_begin;
661 }
662 return (0);
663 }
664
665 int
666 dmu_free_long_range(objset_t *os, uint64_t object,
667 uint64_t offset, uint64_t length)
668 {
669 dnode_t *dn;
670 int err;
671
672 err = dnode_hold(os, object, FTAG, &dn);
673 if (err != 0)
674 return (err);
675 err = dmu_free_long_range_impl(os, dn, offset, length);
676 dnode_rele(dn, FTAG);
677 return (err);
678 }
679
680 int
681 dmu_free_long_object(objset_t *os, uint64_t object)
682 {
683 dmu_tx_t *tx;
684 int err;
685
686 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
687 if (err != 0)
688 return (err);
689
690 tx = dmu_tx_create(os);
691 dmu_tx_hold_bonus(tx, object);
692 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
693 err = dmu_tx_assign(tx, TXG_WAIT);
694 if (err == 0) {
695 err = dmu_object_free(os, object, tx);
696 dmu_tx_commit(tx);
697 } else {
698 dmu_tx_abort(tx);
699 }
700
701 return (err);
702 }
703
704 int
705 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
706 uint64_t size, dmu_tx_t *tx)
707 {
708 dnode_t *dn;
709 int err = dnode_hold(os, object, FTAG, &dn);
710 if (err)
711 return (err);
712 ASSERT(offset < UINT64_MAX);
713 ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
714 dnode_free_range(dn, offset, size, tx);
715 dnode_rele(dn, FTAG);
716 return (0);
717 }
718
719 int
720 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
721 void *buf, uint32_t flags)
722 {
723 dnode_t *dn;
724 dmu_buf_t **dbp;
725 int numbufs, err;
726
727 err = dnode_hold(os, object, FTAG, &dn);
728 if (err)
729 return (err);
730
731 /*
732 * Deal with odd block sizes, where there can't be data past the first
733 * block. If we ever do the tail block optimization, we will need to
734 * handle that here as well.
735 */
736 if (dn->dn_maxblkid == 0) {
737 int newsz = offset > dn->dn_datablksz ? 0 :
738 MIN(size, dn->dn_datablksz - offset);
739 bzero((char *)buf + newsz, size - newsz);
740 size = newsz;
741 }
742
743 while (size > 0) {
744 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
745 int i;
746
747 /*
748 * NB: we could do this block-at-a-time, but it's nice
749 * to be reading in parallel.
750 */
751 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
752 TRUE, FTAG, &numbufs, &dbp, flags);
753 if (err)
754 break;
755
756 for (i = 0; i < numbufs; i++) {
757 int tocpy;
758 int bufoff;
759 dmu_buf_t *db = dbp[i];
760
761 ASSERT(size > 0);
762
763 bufoff = offset - db->db_offset;
764 tocpy = (int)MIN(db->db_size - bufoff, size);
765
766 bcopy((char *)db->db_data + bufoff, buf, tocpy);
767
768 offset += tocpy;
769 size -= tocpy;
770 buf = (char *)buf + tocpy;
771 }
772 dmu_buf_rele_array(dbp, numbufs, FTAG);
773 }
774 dnode_rele(dn, FTAG);
775 return (err);
776 }
777
778 void
779 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
780 const void *buf, dmu_tx_t *tx)
781 {
782 dmu_buf_t **dbp;
783 int numbufs, i;
784
785 if (size == 0)
786 return;
787
788 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
789 FALSE, FTAG, &numbufs, &dbp));
790
791 for (i = 0; i < numbufs; i++) {
792 int tocpy;
793 int bufoff;
794 dmu_buf_t *db = dbp[i];
795
796 ASSERT(size > 0);
797
798 bufoff = offset - db->db_offset;
799 tocpy = (int)MIN(db->db_size - bufoff, size);
800
801 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
802
803 if (tocpy == db->db_size)
804 dmu_buf_will_fill(db, tx);
805 else
806 dmu_buf_will_dirty(db, tx);
807
808 bcopy(buf, (char *)db->db_data + bufoff, tocpy);
809
810 if (tocpy == db->db_size)
811 dmu_buf_fill_done(db, tx);
812
813 offset += tocpy;
814 size -= tocpy;
815 buf = (char *)buf + tocpy;
816 }
817 dmu_buf_rele_array(dbp, numbufs, FTAG);
818 }
819
820 void
821 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
822 dmu_tx_t *tx)
823 {
824 dmu_buf_t **dbp;
825 int numbufs, i;
826
827 if (size == 0)
828 return;
829
830 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
831 FALSE, FTAG, &numbufs, &dbp));
832
833 for (i = 0; i < numbufs; i++) {
834 dmu_buf_t *db = dbp[i];
835
836 dmu_buf_will_not_fill(db, tx);
837 }
838 dmu_buf_rele_array(dbp, numbufs, FTAG);
839 }
840
841 /*
842 * DMU support for xuio
843 */
844 kstat_t *xuio_ksp = NULL;
845
846 int
847 dmu_xuio_init(xuio_t *xuio, int nblk)
848 {
849 dmu_xuio_t *priv;
850 uio_t *uio = &xuio->xu_uio;
851
852 uio->uio_iovcnt = nblk;
853 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
854
855 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
856 priv->cnt = nblk;
857 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
858 priv->iovp = uio->uio_iov;
859 XUIO_XUZC_PRIV(xuio) = priv;
860
861 if (XUIO_XUZC_RW(xuio) == UIO_READ)
862 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
863 else
864 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
865
866 return (0);
867 }
868
869 void
870 dmu_xuio_fini(xuio_t *xuio)
871 {
872 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
873 int nblk = priv->cnt;
874
875 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
876 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
877 kmem_free(priv, sizeof (dmu_xuio_t));
878
879 if (XUIO_XUZC_RW(xuio) == UIO_READ)
880 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
881 else
882 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
883 }
884
885 /*
886 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
887 * and increase priv->next by 1.
888 */
889 int
890 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
891 {
892 struct iovec *iov;
893 uio_t *uio = &xuio->xu_uio;
894 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
895 int i = priv->next++;
896
897 ASSERT(i < priv->cnt);
898 ASSERT(off + n <= arc_buf_size(abuf));
899 iov = uio->uio_iov + i;
900 iov->iov_base = (char *)abuf->b_data + off;
901 iov->iov_len = n;
902 priv->bufs[i] = abuf;
903 return (0);
904 }
905
906 int
907 dmu_xuio_cnt(xuio_t *xuio)
908 {
909 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
910 return (priv->cnt);
911 }
912
913 arc_buf_t *
914 dmu_xuio_arcbuf(xuio_t *xuio, int i)
915 {
916 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
917
918 ASSERT(i < priv->cnt);
919 return (priv->bufs[i]);
920 }
921
922 void
923 dmu_xuio_clear(xuio_t *xuio, int i)
924 {
925 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
926
927 ASSERT(i < priv->cnt);
928 priv->bufs[i] = NULL;
929 }
930
931 static void
932 xuio_stat_init(void)
933 {
934 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
935 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
936 KSTAT_FLAG_VIRTUAL);
937 if (xuio_ksp != NULL) {
938 xuio_ksp->ks_data = &xuio_stats;
939 kstat_install(xuio_ksp);
940 }
941 }
942
943 static void
944 xuio_stat_fini(void)
945 {
946 if (xuio_ksp != NULL) {
947 kstat_delete(xuio_ksp);
948 xuio_ksp = NULL;
949 }
950 }
951
952 void
953 xuio_stat_wbuf_copied()
954 {
955 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
956 }
957
958 void
959 xuio_stat_wbuf_nocopy()
960 {
961 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
962 }
963
964 #ifdef _KERNEL
965 int
966 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
967 {
968 dmu_buf_t **dbp;
969 int numbufs, i, err;
970 xuio_t *xuio = NULL;
971
972 /*
973 * NB: we could do this block-at-a-time, but it's nice
974 * to be reading in parallel.
975 */
976 err = dmu_buf_hold_array(os, object, uio->uio_loffset, size, TRUE, FTAG,
977 &numbufs, &dbp);
978 if (err)
979 return (err);
980
981 if (uio->uio_extflg == UIO_XUIO)
982 xuio = (xuio_t *)uio;
983
984 for (i = 0; i < numbufs; i++) {
985 int tocpy;
986 int bufoff;
987 dmu_buf_t *db = dbp[i];
988
989 ASSERT(size > 0);
990
991 bufoff = uio->uio_loffset - db->db_offset;
992 tocpy = (int)MIN(db->db_size - bufoff, size);
993
994 if (xuio) {
995 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
996 arc_buf_t *dbuf_abuf = dbi->db_buf;
997 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
998 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
999 if (!err) {
1000 uio->uio_resid -= tocpy;
1001 uio->uio_loffset += tocpy;
1002 }
1003
1004 if (abuf == dbuf_abuf)
1005 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
1006 else
1007 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
1008 } else {
1009 err = uiomove((char *)db->db_data + bufoff, tocpy,
1010 UIO_READ, uio);
1011 }
1012 if (err)
1013 break;
1014
1015 size -= tocpy;
1016 }
1017 dmu_buf_rele_array(dbp, numbufs, FTAG);
1018
1019 return (err);
1020 }
1021
1022 static int
1023 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1024 {
1025 dmu_buf_t **dbp;
1026 int numbufs;
1027 int err = 0;
1028 int i;
1029
1030 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1031 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1032 if (err)
1033 return (err);
1034
1035 for (i = 0; i < numbufs; i++) {
1036 int tocpy;
1037 int bufoff;
1038 dmu_buf_t *db = dbp[i];
1039
1040 ASSERT(size > 0);
1041
1042 bufoff = uio->uio_loffset - db->db_offset;
1043 tocpy = (int)MIN(db->db_size - bufoff, size);
1044
1045 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1046
1047 if (tocpy == db->db_size)
1048 dmu_buf_will_fill(db, tx);
1049 else
1050 dmu_buf_will_dirty(db, tx);
1051
1052 /*
1053 * XXX uiomove could block forever (eg. nfs-backed
1054 * pages). There needs to be a uiolockdown() function
1055 * to lock the pages in memory, so that uiomove won't
1056 * block.
1057 */
1058 err = uiomove((char *)db->db_data + bufoff, tocpy,
1059 UIO_WRITE, uio);
1060
1061 if (tocpy == db->db_size)
1062 dmu_buf_fill_done(db, tx);
1063
1064 if (err)
1065 break;
1066
1067 size -= tocpy;
1068 }
1069
1070 dmu_buf_rele_array(dbp, numbufs, FTAG);
1071 return (err);
1072 }
1073
1074 int
1075 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1076 dmu_tx_t *tx)
1077 {
1078 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1079 dnode_t *dn;
1080 int err;
1081
1082 if (size == 0)
1083 return (0);
1084
1085 DB_DNODE_ENTER(db);
1086 dn = DB_DNODE(db);
1087 err = dmu_write_uio_dnode(dn, uio, size, tx);
1088 DB_DNODE_EXIT(db);
1089
1090 return (err);
1091 }
1092
1093 int
1094 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1095 dmu_tx_t *tx)
1096 {
1097 dnode_t *dn;
1098 int err;
1099
1100 if (size == 0)
1101 return (0);
1102
1103 err = dnode_hold(os, object, FTAG, &dn);
1104 if (err)
1105 return (err);
1106
1107 err = dmu_write_uio_dnode(dn, uio, size, tx);
1108
1109 dnode_rele(dn, FTAG);
1110
1111 return (err);
1112 }
1113
1114 int
1115 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1116 page_t *pp, dmu_tx_t *tx)
1117 {
1118 dmu_buf_t **dbp;
1119 int numbufs, i;
1120 int err;
1121
1122 if (size == 0)
1123 return (0);
1124
1125 err = dmu_buf_hold_array(os, object, offset, size,
1126 FALSE, FTAG, &numbufs, &dbp);
1127 if (err)
1128 return (err);
1129
1130 for (i = 0; i < numbufs; i++) {
1131 int tocpy, copied, thiscpy;
1132 int bufoff;
1133 dmu_buf_t *db = dbp[i];
1134 caddr_t va;
1135
1136 ASSERT(size > 0);
1137 ASSERT3U(db->db_size, >=, PAGESIZE);
1138
1139 bufoff = offset - db->db_offset;
1140 tocpy = (int)MIN(db->db_size - bufoff, size);
1141
1142 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1143
1144 if (tocpy == db->db_size)
1145 dmu_buf_will_fill(db, tx);
1146 else
1147 dmu_buf_will_dirty(db, tx);
1148
1149 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1150 ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff);
1151 thiscpy = MIN(PAGESIZE, tocpy - copied);
1152 va = zfs_map_page(pp, S_READ);
1153 bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1154 zfs_unmap_page(pp, va);
1155 pp = pp->p_next;
1156 bufoff += PAGESIZE;
1157 }
1158
1159 if (tocpy == db->db_size)
1160 dmu_buf_fill_done(db, tx);
1161
1162 offset += tocpy;
1163 size -= tocpy;
1164 }
1165 dmu_buf_rele_array(dbp, numbufs, FTAG);
1166 return (err);
1167 }
1168 #endif
1169
1170 /*
1171 * Allocate a loaned anonymous arc buffer.
1172 */
1173 arc_buf_t *
1174 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1175 {
1176 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1177 spa_t *spa;
1178
1179 DB_GET_SPA(&spa, db);
1180 return (arc_loan_buf(spa, size));
1181 }
1182
1183 /*
1184 * Free a loaned arc buffer.
1185 */
1186 void
1187 dmu_return_arcbuf(arc_buf_t *buf)
1188 {
1189 arc_return_buf(buf, FTAG);
1190 VERIFY(arc_buf_remove_ref(buf, FTAG));
1191 }
1192
1193 /*
1194 * When possible directly assign passed loaned arc buffer to a dbuf.
1195 * If this is not possible copy the contents of passed arc buf via
1196 * dmu_write().
1197 */
1198 void
1199 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1200 dmu_tx_t *tx)
1201 {
1202 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1203 dnode_t *dn;
1204 dmu_buf_impl_t *db;
1205 uint32_t blksz = (uint32_t)arc_buf_size(buf);
1206 uint64_t blkid;
1207
1208 DB_DNODE_ENTER(dbuf);
1209 dn = DB_DNODE(dbuf);
1210 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1211 blkid = dbuf_whichblock(dn, offset);
1212 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1213 rw_exit(&dn->dn_struct_rwlock);
1214 DB_DNODE_EXIT(dbuf);
1215
1216 if (offset == db->db.db_offset && blksz == db->db.db_size) {
1217 dbuf_assign_arcbuf(db, buf, tx);
1218 dbuf_rele(db, FTAG);
1219 } else {
1220 objset_t *os;
1221 uint64_t object;
1222
1223 DB_DNODE_ENTER(dbuf);
1224 dn = DB_DNODE(dbuf);
1225 os = dn->dn_objset;
1226 object = dn->dn_object;
1227 DB_DNODE_EXIT(dbuf);
1228
1229 dbuf_rele(db, FTAG);
1230 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1231 dmu_return_arcbuf(buf);
1232 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1233 }
1234 }
1235
1236 typedef struct {
1237 dbuf_dirty_record_t *dsa_dr;
1238 dmu_sync_cb_t *dsa_done;
1239 zgd_t *dsa_zgd;
1240 dmu_tx_t *dsa_tx;
1241 } dmu_sync_arg_t;
1242
1243 /* ARGSUSED */
1244 static void
1245 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1246 {
1247 dmu_sync_arg_t *dsa = varg;
1248 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1249 blkptr_t *bp = zio->io_bp;
1250
1251 if (zio->io_error == 0) {
1252 if (BP_IS_HOLE(bp)) {
1253 /*
1254 * A block of zeros may compress to a hole, but the
1255 * block size still needs to be known for replay.
1256 */
1257 BP_SET_LSIZE(bp, db->db_size);
1258 } else {
1259 ASSERT(BP_GET_LEVEL(bp) == 0);
1260 bp->blk_fill = 1;
1261 }
1262 }
1263 }
1264
1265 static void
1266 dmu_sync_late_arrival_ready(zio_t *zio)
1267 {
1268 dmu_sync_ready(zio, NULL, zio->io_private);
1269 }
1270
1271 /* ARGSUSED */
1272 static void
1273 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1274 {
1275 dmu_sync_arg_t *dsa = varg;
1276 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1277 dmu_buf_impl_t *db = dr->dr_dbuf;
1278
1279 mutex_enter(&db->db_mtx);
1280 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1281 if (zio->io_error == 0) {
1282 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1283 if (dr->dt.dl.dr_nopwrite) {
1284 blkptr_t *bp = zio->io_bp;
1285 blkptr_t *bp_orig = &zio->io_bp_orig;
1286 uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1287
1288 ASSERT(BP_EQUAL(bp, bp_orig));
1289 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1290 ASSERT(zio_checksum_table[chksum].ci_dedup);
1291 }
1292 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1293 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1294 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1295 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by))
1296 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1297 } else {
1298 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1299 }
1300 cv_broadcast(&db->db_changed);
1301 mutex_exit(&db->db_mtx);
1302
1303 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1304
1305 kmem_free(dsa, sizeof (*dsa));
1306 }
1307
1308 static void
1309 dmu_sync_late_arrival_done(zio_t *zio)
1310 {
1311 blkptr_t *bp = zio->io_bp;
1312 dmu_sync_arg_t *dsa = zio->io_private;
1313 blkptr_t *bp_orig = &zio->io_bp_orig;
1314
1315 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1316 /*
1317 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE)
1318 * then there is nothing to do here. Otherwise, free the
1319 * newly allocated block in this txg.
1320 */
1321 if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
1322 ASSERT(BP_EQUAL(bp, bp_orig));
1323 } else {
1324 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1325 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1326 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1327 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1328 }
1329 }
1330
1331 dmu_tx_commit(dsa->dsa_tx);
1332
1333 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1334
1335 kmem_free(dsa, sizeof (*dsa));
1336 }
1337
1338 static int
1339 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1340 zio_prop_t *zp, zbookmark_t *zb)
1341 {
1342 dmu_sync_arg_t *dsa;
1343 dmu_tx_t *tx;
1344
1345 tx = dmu_tx_create(os);
1346 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1347 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1348 dmu_tx_abort(tx);
1349 /* Make zl_get_data do txg_waited_synced() */
1350 return (SET_ERROR(EIO));
1351 }
1352
1353 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1354 dsa->dsa_dr = NULL;
1355 dsa->dsa_done = done;
1356 dsa->dsa_zgd = zgd;
1357 dsa->dsa_tx = tx;
1358
1359 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1360 zgd->zgd_db->db_data, zgd->zgd_db->db_size, zp,
1361 dmu_sync_late_arrival_ready, NULL, dmu_sync_late_arrival_done, dsa,
1362 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1363
1364 return (0);
1365 }
1366
1367 /*
1368 * Intent log support: sync the block associated with db to disk.
1369 * N.B. and XXX: the caller is responsible for making sure that the
1370 * data isn't changing while dmu_sync() is writing it.
1371 *
1372 * Return values:
1373 *
1374 * EEXIST: this txg has already been synced, so there's nothing to do.
1375 * The caller should not log the write.
1376 *
1377 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1378 * The caller should not log the write.
1379 *
1380 * EALREADY: this block is already in the process of being synced.
1381 * The caller should track its progress (somehow).
1382 *
1383 * EIO: could not do the I/O.
1384 * The caller should do a txg_wait_synced().
1385 *
1386 * 0: the I/O has been initiated.
1387 * The caller should log this blkptr in the done callback.
1388 * It is possible that the I/O will fail, in which case
1389 * the error will be reported to the done callback and
1390 * propagated to pio from zio_done().
1391 */
1392 int
1393 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1394 {
1395 blkptr_t *bp = zgd->zgd_bp;
1396 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1397 objset_t *os = db->db_objset;
1398 dsl_dataset_t *ds = os->os_dsl_dataset;
1399 dbuf_dirty_record_t *dr;
1400 dmu_sync_arg_t *dsa;
1401 zbookmark_t zb;
1402 zio_prop_t zp;
1403 dnode_t *dn;
1404
1405 ASSERT(pio != NULL);
1406 ASSERT(txg != 0);
1407
1408 SET_BOOKMARK(&zb, ds->ds_object,
1409 db->db.db_object, db->db_level, db->db_blkid);
1410
1411 DB_DNODE_ENTER(db);
1412 dn = DB_DNODE(db);
1413 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1414 DB_DNODE_EXIT(db);
1415
1416 /*
1417 * If we're frozen (running ziltest), we always need to generate a bp.
1418 */
1419 if (txg > spa_freeze_txg(os->os_spa))
1420 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1421
1422 /*
1423 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1424 * and us. If we determine that this txg is not yet syncing,
1425 * but it begins to sync a moment later, that's OK because the
1426 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1427 */
1428 mutex_enter(&db->db_mtx);
1429
1430 if (txg <= spa_last_synced_txg(os->os_spa)) {
1431 /*
1432 * This txg has already synced. There's nothing to do.
1433 */
1434 mutex_exit(&db->db_mtx);
1435 return (SET_ERROR(EEXIST));
1436 }
1437
1438 if (txg <= spa_syncing_txg(os->os_spa)) {
1439 /*
1440 * This txg is currently syncing, so we can't mess with
1441 * the dirty record anymore; just write a new log block.
1442 */
1443 mutex_exit(&db->db_mtx);
1444 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1445 }
1446
1447 dr = db->db_last_dirty;
1448 while (dr && dr->dr_txg != txg)
1449 dr = dr->dr_next;
1450
1451 if (dr == NULL) {
1452 /*
1453 * There's no dr for this dbuf, so it must have been freed.
1454 * There's no need to log writes to freed blocks, so we're done.
1455 */
1456 mutex_exit(&db->db_mtx);
1457 return (SET_ERROR(ENOENT));
1458 }
1459
1460 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
1461
1462 /*
1463 * Assume the on-disk data is X, the current syncing data is Y,
1464 * and the current in-memory data is Z (currently in dmu_sync).
1465 * X and Z are identical but Y is has been modified. Normally,
1466 * when X and Z are the same we will perform a nopwrite but if Y
1467 * is different we must disable nopwrite since the resulting write
1468 * of Y to disk can free the block containing X. If we allowed a
1469 * nopwrite to occur the block pointing to Z would reference a freed
1470 * block. Since this is a rare case we simplify this by disabling
1471 * nopwrite if the current dmu_sync-ing dbuf has been modified in
1472 * a previous transaction.
1473 */
1474 if (dr->dr_next)
1475 zp.zp_nopwrite = B_FALSE;
1476
1477 ASSERT(dr->dr_txg == txg);
1478 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1479 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1480 /*
1481 * We have already issued a sync write for this buffer,
1482 * or this buffer has already been synced. It could not
1483 * have been dirtied since, or we would have cleared the state.
1484 */
1485 mutex_exit(&db->db_mtx);
1486 return (SET_ERROR(EALREADY));
1487 }
1488
1489 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1490 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1491 mutex_exit(&db->db_mtx);
1492
1493 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1494 dsa->dsa_dr = dr;
1495 dsa->dsa_done = done;
1496 dsa->dsa_zgd = zgd;
1497 dsa->dsa_tx = NULL;
1498
1499 zio_nowait(arc_write(pio, os->os_spa, txg,
1500 bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
1501 DBUF_IS_L2COMPRESSIBLE(db), &zp, dmu_sync_ready,
1502 NULL, dmu_sync_done, dsa, ZIO_PRIORITY_SYNC_WRITE,
1503 ZIO_FLAG_CANFAIL, &zb));
1504
1505 return (0);
1506 }
1507
1508 int
1509 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1510 dmu_tx_t *tx)
1511 {
1512 dnode_t *dn;
1513 int err;
1514
1515 err = dnode_hold(os, object, FTAG, &dn);
1516 if (err)
1517 return (err);
1518 err = dnode_set_blksz(dn, size, ibs, tx);
1519 dnode_rele(dn, FTAG);
1520 return (err);
1521 }
1522
1523 void
1524 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1525 dmu_tx_t *tx)
1526 {
1527 dnode_t *dn;
1528
1529 /* XXX assumes dnode_hold will not get an i/o error */
1530 (void) dnode_hold(os, object, FTAG, &dn);
1531 ASSERT(checksum < ZIO_CHECKSUM_FUNCTIONS);
1532 dn->dn_checksum = checksum;
1533 dnode_setdirty(dn, tx);
1534 dnode_rele(dn, FTAG);
1535 }
1536
1537 void
1538 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1539 dmu_tx_t *tx)
1540 {
1541 dnode_t *dn;
1542
1543 /* XXX assumes dnode_hold will not get an i/o error */
1544 (void) dnode_hold(os, object, FTAG, &dn);
1545 ASSERT(compress < ZIO_COMPRESS_FUNCTIONS);
1546 dn->dn_compress = compress;
1547 dnode_setdirty(dn, tx);
1548 dnode_rele(dn, FTAG);
1549 }
1550
1551 int zfs_mdcomp_disable = 0;
1552
1553 void
1554 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
1555 {
1556 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
1557 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
1558 (wp & WP_SPILL));
1559 enum zio_checksum checksum = os->os_checksum;
1560 enum zio_compress compress = os->os_compress;
1561 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
1562 boolean_t dedup = B_FALSE;
1563 boolean_t nopwrite = B_FALSE;
1564 boolean_t dedup_verify = os->os_dedup_verify;
1565 int copies = os->os_copies;
1566
1567 /*
1568 * We maintain different write policies for each of the following
1569 * types of data:
1570 * 1. metadata
1571 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1572 * 3. all other level 0 blocks
1573 */
1574 if (ismd) {
1575 /*
1576 * XXX -- we should design a compression algorithm
1577 * that specializes in arrays of bps.
1578 */
1579 compress = zfs_mdcomp_disable ? ZIO_COMPRESS_EMPTY :
1580 ZIO_COMPRESS_LZJB;
1581
1582 /*
1583 * Metadata always gets checksummed. If the data
1584 * checksum is multi-bit correctable, and it's not a
1585 * ZBT-style checksum, then it's suitable for metadata
1586 * as well. Otherwise, the metadata checksum defaults
1587 * to fletcher4.
1588 */
1589 if (zio_checksum_table[checksum].ci_correctable < 1 ||
1590 zio_checksum_table[checksum].ci_eck)
1591 checksum = ZIO_CHECKSUM_FLETCHER_4;
1592 } else if (wp & WP_NOFILL) {
1593 ASSERT(level == 0);
1594
1595 /*
1596 * If we're writing preallocated blocks, we aren't actually
1597 * writing them so don't set any policy properties. These
1598 * blocks are currently only used by an external subsystem
1599 * outside of zfs (i.e. dump) and not written by the zio
1600 * pipeline.
1601 */
1602 compress = ZIO_COMPRESS_OFF;
1603 checksum = ZIO_CHECKSUM_NOPARITY;
1604 } else {
1605 compress = zio_compress_select(dn->dn_compress, compress);
1606
1607 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
1608 zio_checksum_select(dn->dn_checksum, checksum) :
1609 dedup_checksum;
1610
1611 /*
1612 * Determine dedup setting. If we are in dmu_sync(),
1613 * we won't actually dedup now because that's all
1614 * done in syncing context; but we do want to use the
1615 * dedup checkum. If the checksum is not strong
1616 * enough to ensure unique signatures, force
1617 * dedup_verify.
1618 */
1619 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
1620 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
1621 if (!zio_checksum_table[checksum].ci_dedup)
1622 dedup_verify = B_TRUE;
1623 }
1624
1625 /*
1626 * Enable nopwrite if we have a cryptographically secure
1627 * checksum that has no known collisions (i.e. SHA-256)
1628 * and compression is enabled. We don't enable nopwrite if
1629 * dedup is enabled as the two features are mutually exclusive.
1630 */
1631 nopwrite = (!dedup && zio_checksum_table[checksum].ci_dedup &&
1632 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
1633 }
1634
1635 zp->zp_checksum = checksum;
1636 zp->zp_compress = compress;
1637 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
1638 zp->zp_level = level;
1639 zp->zp_copies = MIN(copies + ismd, spa_max_replication(os->os_spa));
1640 zp->zp_dedup = dedup;
1641 zp->zp_dedup_verify = dedup && dedup_verify;
1642 zp->zp_nopwrite = nopwrite;
1643 }
1644
1645 int
1646 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
1647 {
1648 dnode_t *dn;
1649 int i, err;
1650
1651 err = dnode_hold(os, object, FTAG, &dn);
1652 if (err)
1653 return (err);
1654 /*
1655 * Sync any current changes before
1656 * we go trundling through the block pointers.
1657 */
1658 for (i = 0; i < TXG_SIZE; i++) {
1659 if (list_link_active(&dn->dn_dirty_link[i]))
1660 break;
1661 }
1662 if (i != TXG_SIZE) {
1663 dnode_rele(dn, FTAG);
1664 txg_wait_synced(dmu_objset_pool(os), 0);
1665 err = dnode_hold(os, object, FTAG, &dn);
1666 if (err)
1667 return (err);
1668 }
1669
1670 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
1671 dnode_rele(dn, FTAG);
1672
1673 return (err);
1674 }
1675
1676 void
1677 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
1678 {
1679 dnode_phys_t *dnp;
1680
1681 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1682 mutex_enter(&dn->dn_mtx);
1683
1684 dnp = dn->dn_phys;
1685
1686 doi->doi_data_block_size = dn->dn_datablksz;
1687 doi->doi_metadata_block_size = dn->dn_indblkshift ?
1688 1ULL << dn->dn_indblkshift : 0;
1689 doi->doi_type = dn->dn_type;
1690 doi->doi_bonus_type = dn->dn_bonustype;
1691 doi->doi_bonus_size = dn->dn_bonuslen;
1692 doi->doi_indirection = dn->dn_nlevels;
1693 doi->doi_checksum = dn->dn_checksum;
1694 doi->doi_compress = dn->dn_compress;
1695 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
1696 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
1697 doi->doi_fill_count = 0;
1698 for (int i = 0; i < dnp->dn_nblkptr; i++)
1699 doi->doi_fill_count += dnp->dn_blkptr[i].blk_fill;
1700
1701 mutex_exit(&dn->dn_mtx);
1702 rw_exit(&dn->dn_struct_rwlock);
1703 }
1704
1705 /*
1706 * Get information on a DMU object.
1707 * If doi is NULL, just indicates whether the object exists.
1708 */
1709 int
1710 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
1711 {
1712 dnode_t *dn;
1713 int err = dnode_hold(os, object, FTAG, &dn);
1714
1715 if (err)
1716 return (err);
1717
1718 if (doi != NULL)
1719 dmu_object_info_from_dnode(dn, doi);
1720
1721 dnode_rele(dn, FTAG);
1722 return (0);
1723 }
1724
1725 /*
1726 * As above, but faster; can be used when you have a held dbuf in hand.
1727 */
1728 void
1729 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
1730 {
1731 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1732
1733 DB_DNODE_ENTER(db);
1734 dmu_object_info_from_dnode(DB_DNODE(db), doi);
1735 DB_DNODE_EXIT(db);
1736 }
1737
1738 /*
1739 * Faster still when you only care about the size.
1740 * This is specifically optimized for zfs_getattr().
1741 */
1742 void
1743 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
1744 u_longlong_t *nblk512)
1745 {
1746 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1747 dnode_t *dn;
1748
1749 DB_DNODE_ENTER(db);
1750 dn = DB_DNODE(db);
1751
1752 *blksize = dn->dn_datablksz;
1753 /* add 1 for dnode space */
1754 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
1755 SPA_MINBLOCKSHIFT) + 1;
1756 DB_DNODE_EXIT(db);
1757 }
1758
1759 void
1760 byteswap_uint64_array(void *vbuf, size_t size)
1761 {
1762 uint64_t *buf = vbuf;
1763 size_t count = size >> 3;
1764 int i;
1765
1766 ASSERT((size & 7) == 0);
1767
1768 for (i = 0; i < count; i++)
1769 buf[i] = BSWAP_64(buf[i]);
1770 }
1771
1772 void
1773 byteswap_uint32_array(void *vbuf, size_t size)
1774 {
1775 uint32_t *buf = vbuf;
1776 size_t count = size >> 2;
1777 int i;
1778
1779 ASSERT((size & 3) == 0);
1780
1781 for (i = 0; i < count; i++)
1782 buf[i] = BSWAP_32(buf[i]);
1783 }
1784
1785 void
1786 byteswap_uint16_array(void *vbuf, size_t size)
1787 {
1788 uint16_t *buf = vbuf;
1789 size_t count = size >> 1;
1790 int i;
1791
1792 ASSERT((size & 1) == 0);
1793
1794 for (i = 0; i < count; i++)
1795 buf[i] = BSWAP_16(buf[i]);
1796 }
1797
1798 /* ARGSUSED */
1799 void
1800 byteswap_uint8_array(void *vbuf, size_t size)
1801 {
1802 }
1803
1804 void
1805 dmu_init(void)
1806 {
1807 zfs_dbgmsg_init();
1808 sa_cache_init();
1809 xuio_stat_init();
1810 dmu_objset_init();
1811 dnode_init();
1812 dbuf_init();
1813 zfetch_init();
1814 l2arc_init();
1815 arc_init();
1816 }
1817
1818 void
1819 dmu_fini(void)
1820 {
1821 arc_fini(); /* arc depends on l2arc, so arc must go first */
1822 l2arc_fini();
1823 zfetch_fini();
1824 dbuf_fini();
1825 dnode_fini();
1826 dmu_objset_fini();
1827 xuio_stat_fini();
1828 sa_cache_fini();
1829 zfs_dbgmsg_fini();
1830 }