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
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
677 /*
678 * It is important to zero out the maxblkid when freeing the entire
679 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
680 * will take the fast path, and (b) dnode_reallocate() can verify
681 * that the entire file has been freed.
682 */
683 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
684 dn->dn_maxblkid = 0;
685
686 dnode_rele(dn, FTAG);
687 return (err);
688 }
689
690 int
691 dmu_free_long_object(objset_t *os, uint64_t object)
692 {
693 dmu_tx_t *tx;
694 int err;
695
696 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
697 if (err != 0)
698 return (err);
699
700 tx = dmu_tx_create(os);
701 dmu_tx_hold_bonus(tx, object);
702 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
703 err = dmu_tx_assign(tx, TXG_WAIT);
704 if (err == 0) {
705 err = dmu_object_free(os, object, tx);
706 dmu_tx_commit(tx);
707 } else {
708 dmu_tx_abort(tx);
709 }
710
711 return (err);
712 }
713
714 int
715 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
716 uint64_t size, dmu_tx_t *tx)
717 {
718 dnode_t *dn;
719 int err = dnode_hold(os, object, FTAG, &dn);
720 if (err)
721 return (err);
722 ASSERT(offset < UINT64_MAX);
723 ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
724 dnode_free_range(dn, offset, size, tx);
725 dnode_rele(dn, FTAG);
726 return (0);
727 }
728
729 int
730 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
731 void *buf, uint32_t flags)
732 {
733 dnode_t *dn;
734 dmu_buf_t **dbp;
735 int numbufs, err;
736
737 err = dnode_hold(os, object, FTAG, &dn);
738 if (err)
739 return (err);
740
741 /*
742 * Deal with odd block sizes, where there can't be data past the first
743 * block. If we ever do the tail block optimization, we will need to
744 * handle that here as well.
745 */
746 if (dn->dn_maxblkid == 0) {
747 int newsz = offset > dn->dn_datablksz ? 0 :
748 MIN(size, dn->dn_datablksz - offset);
749 bzero((char *)buf + newsz, size - newsz);
750 size = newsz;
751 }
752
753 while (size > 0) {
754 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
755 int i;
756
757 /*
758 * NB: we could do this block-at-a-time, but it's nice
759 * to be reading in parallel.
760 */
761 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
762 TRUE, FTAG, &numbufs, &dbp, flags);
763 if (err)
764 break;
765
766 for (i = 0; i < numbufs; i++) {
767 int tocpy;
768 int bufoff;
769 dmu_buf_t *db = dbp[i];
770
771 ASSERT(size > 0);
772
773 bufoff = offset - db->db_offset;
774 tocpy = (int)MIN(db->db_size - bufoff, size);
775
776 bcopy((char *)db->db_data + bufoff, buf, tocpy);
777
778 offset += tocpy;
779 size -= tocpy;
780 buf = (char *)buf + tocpy;
781 }
782 dmu_buf_rele_array(dbp, numbufs, FTAG);
783 }
784 dnode_rele(dn, FTAG);
785 return (err);
786 }
787
788 void
789 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
790 const void *buf, dmu_tx_t *tx)
791 {
792 dmu_buf_t **dbp;
793 int numbufs, i;
794
795 if (size == 0)
796 return;
797
798 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
799 FALSE, FTAG, &numbufs, &dbp));
800
801 for (i = 0; i < numbufs; i++) {
802 int tocpy;
803 int bufoff;
804 dmu_buf_t *db = dbp[i];
805
806 ASSERT(size > 0);
807
808 bufoff = offset - db->db_offset;
809 tocpy = (int)MIN(db->db_size - bufoff, size);
810
811 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
812
813 if (tocpy == db->db_size)
814 dmu_buf_will_fill(db, tx);
815 else
816 dmu_buf_will_dirty(db, tx);
817
818 bcopy(buf, (char *)db->db_data + bufoff, tocpy);
819
820 if (tocpy == db->db_size)
821 dmu_buf_fill_done(db, tx);
822
823 offset += tocpy;
824 size -= tocpy;
825 buf = (char *)buf + tocpy;
826 }
827 dmu_buf_rele_array(dbp, numbufs, FTAG);
828 }
829
830 void
831 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
832 dmu_tx_t *tx)
833 {
834 dmu_buf_t **dbp;
835 int numbufs, i;
836
837 if (size == 0)
838 return;
839
840 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
841 FALSE, FTAG, &numbufs, &dbp));
842
843 for (i = 0; i < numbufs; i++) {
844 dmu_buf_t *db = dbp[i];
845
846 dmu_buf_will_not_fill(db, tx);
847 }
848 dmu_buf_rele_array(dbp, numbufs, FTAG);
849 }
850
851 /*
852 * DMU support for xuio
853 */
854 kstat_t *xuio_ksp = NULL;
855
856 int
857 dmu_xuio_init(xuio_t *xuio, int nblk)
858 {
859 dmu_xuio_t *priv;
860 uio_t *uio = &xuio->xu_uio;
861
862 uio->uio_iovcnt = nblk;
863 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
864
865 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
866 priv->cnt = nblk;
867 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
868 priv->iovp = uio->uio_iov;
869 XUIO_XUZC_PRIV(xuio) = priv;
870
871 if (XUIO_XUZC_RW(xuio) == UIO_READ)
872 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
873 else
874 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
875
876 return (0);
877 }
878
879 void
880 dmu_xuio_fini(xuio_t *xuio)
881 {
882 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
883 int nblk = priv->cnt;
884
885 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
886 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
887 kmem_free(priv, sizeof (dmu_xuio_t));
888
889 if (XUIO_XUZC_RW(xuio) == UIO_READ)
890 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
891 else
892 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
893 }
894
895 /*
896 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
897 * and increase priv->next by 1.
898 */
899 int
900 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
901 {
902 struct iovec *iov;
903 uio_t *uio = &xuio->xu_uio;
904 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
905 int i = priv->next++;
906
907 ASSERT(i < priv->cnt);
908 ASSERT(off + n <= arc_buf_size(abuf));
909 iov = uio->uio_iov + i;
910 iov->iov_base = (char *)abuf->b_data + off;
911 iov->iov_len = n;
912 priv->bufs[i] = abuf;
913 return (0);
914 }
915
916 int
917 dmu_xuio_cnt(xuio_t *xuio)
918 {
919 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
920 return (priv->cnt);
921 }
922
923 arc_buf_t *
924 dmu_xuio_arcbuf(xuio_t *xuio, int i)
925 {
926 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
927
928 ASSERT(i < priv->cnt);
929 return (priv->bufs[i]);
930 }
931
932 void
933 dmu_xuio_clear(xuio_t *xuio, int i)
934 {
935 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
936
937 ASSERT(i < priv->cnt);
938 priv->bufs[i] = NULL;
939 }
940
941 static void
942 xuio_stat_init(void)
943 {
944 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
945 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
946 KSTAT_FLAG_VIRTUAL);
947 if (xuio_ksp != NULL) {
948 xuio_ksp->ks_data = &xuio_stats;
949 kstat_install(xuio_ksp);
950 }
951 }
952
953 static void
954 xuio_stat_fini(void)
955 {
956 if (xuio_ksp != NULL) {
957 kstat_delete(xuio_ksp);
958 xuio_ksp = NULL;
959 }
960 }
961
962 void
963 xuio_stat_wbuf_copied()
964 {
965 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
966 }
967
968 void
969 xuio_stat_wbuf_nocopy()
970 {
971 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
972 }
973
974 #ifdef _KERNEL
975 int
976 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
977 {
978 dmu_buf_t **dbp;
979 int numbufs, i, err;
980 xuio_t *xuio = NULL;
981
982 /*
983 * NB: we could do this block-at-a-time, but it's nice
984 * to be reading in parallel.
985 */
986 err = dmu_buf_hold_array(os, object, uio->uio_loffset, size, TRUE, FTAG,
987 &numbufs, &dbp);
988 if (err)
989 return (err);
990
991 if (uio->uio_extflg == UIO_XUIO)
992 xuio = (xuio_t *)uio;
993
994 for (i = 0; i < numbufs; i++) {
995 int tocpy;
996 int bufoff;
997 dmu_buf_t *db = dbp[i];
998
999 ASSERT(size > 0);
1000
1001 bufoff = uio->uio_loffset - db->db_offset;
1002 tocpy = (int)MIN(db->db_size - bufoff, size);
1003
1004 if (xuio) {
1005 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
1006 arc_buf_t *dbuf_abuf = dbi->db_buf;
1007 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
1008 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
1009 if (!err) {
1010 uio->uio_resid -= tocpy;
1011 uio->uio_loffset += tocpy;
1012 }
1013
1014 if (abuf == dbuf_abuf)
1015 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
1016 else
1017 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
1018 } else {
1019 err = uiomove((char *)db->db_data + bufoff, tocpy,
1020 UIO_READ, uio);
1021 }
1022 if (err)
1023 break;
1024
1025 size -= tocpy;
1026 }
1027 dmu_buf_rele_array(dbp, numbufs, FTAG);
1028
1029 return (err);
1030 }
1031
1032 static int
1033 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1034 {
1035 dmu_buf_t **dbp;
1036 int numbufs;
1037 int err = 0;
1038 int i;
1039
1040 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1041 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1042 if (err)
1043 return (err);
1044
1045 for (i = 0; i < numbufs; i++) {
1046 int tocpy;
1047 int bufoff;
1048 dmu_buf_t *db = dbp[i];
1049
1050 ASSERT(size > 0);
1051
1052 bufoff = uio->uio_loffset - db->db_offset;
1053 tocpy = (int)MIN(db->db_size - bufoff, size);
1054
1055 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1056
1057 if (tocpy == db->db_size)
1058 dmu_buf_will_fill(db, tx);
1059 else
1060 dmu_buf_will_dirty(db, tx);
1061
1062 /*
1063 * XXX uiomove could block forever (eg. nfs-backed
1064 * pages). There needs to be a uiolockdown() function
1065 * to lock the pages in memory, so that uiomove won't
1066 * block.
1067 */
1068 err = uiomove((char *)db->db_data + bufoff, tocpy,
1069 UIO_WRITE, uio);
1070
1071 if (tocpy == db->db_size)
1072 dmu_buf_fill_done(db, tx);
1073
1074 if (err)
1075 break;
1076
1077 size -= tocpy;
1078 }
1079
1080 dmu_buf_rele_array(dbp, numbufs, FTAG);
1081 return (err);
1082 }
1083
1084 int
1085 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1086 dmu_tx_t *tx)
1087 {
1088 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1089 dnode_t *dn;
1090 int err;
1091
1092 if (size == 0)
1093 return (0);
1094
1095 DB_DNODE_ENTER(db);
1096 dn = DB_DNODE(db);
1097 err = dmu_write_uio_dnode(dn, uio, size, tx);
1098 DB_DNODE_EXIT(db);
1099
1100 return (err);
1101 }
1102
1103 int
1104 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1105 dmu_tx_t *tx)
1106 {
1107 dnode_t *dn;
1108 int err;
1109
1110 if (size == 0)
1111 return (0);
1112
1113 err = dnode_hold(os, object, FTAG, &dn);
1114 if (err)
1115 return (err);
1116
1117 err = dmu_write_uio_dnode(dn, uio, size, tx);
1118
1119 dnode_rele(dn, FTAG);
1120
1121 return (err);
1122 }
1123
1124 int
1125 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1126 page_t *pp, dmu_tx_t *tx)
1127 {
1128 dmu_buf_t **dbp;
1129 int numbufs, i;
1130 int err;
1131
1132 if (size == 0)
1133 return (0);
1134
1135 err = dmu_buf_hold_array(os, object, offset, size,
1136 FALSE, FTAG, &numbufs, &dbp);
1137 if (err)
1138 return (err);
1139
1140 for (i = 0; i < numbufs; i++) {
1141 int tocpy, copied, thiscpy;
1142 int bufoff;
1143 dmu_buf_t *db = dbp[i];
1144 caddr_t va;
1145
1146 ASSERT(size > 0);
1147 ASSERT3U(db->db_size, >=, PAGESIZE);
1148
1149 bufoff = offset - db->db_offset;
1150 tocpy = (int)MIN(db->db_size - bufoff, size);
1151
1152 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1153
1154 if (tocpy == db->db_size)
1155 dmu_buf_will_fill(db, tx);
1156 else
1157 dmu_buf_will_dirty(db, tx);
1158
1159 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1160 ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff);
1161 thiscpy = MIN(PAGESIZE, tocpy - copied);
1162 va = zfs_map_page(pp, S_READ);
1163 bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1164 zfs_unmap_page(pp, va);
1165 pp = pp->p_next;
1166 bufoff += PAGESIZE;
1167 }
1168
1169 if (tocpy == db->db_size)
1170 dmu_buf_fill_done(db, tx);
1171
1172 offset += tocpy;
1173 size -= tocpy;
1174 }
1175 dmu_buf_rele_array(dbp, numbufs, FTAG);
1176 return (err);
1177 }
1178 #endif
1179
1180 /*
1181 * Allocate a loaned anonymous arc buffer.
1182 */
1183 arc_buf_t *
1184 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1185 {
1186 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1187
1188 return (arc_loan_buf(db->db_objset->os_spa, size));
1189 }
1190
1191 /*
1192 * Free a loaned arc buffer.
1193 */
1194 void
1195 dmu_return_arcbuf(arc_buf_t *buf)
1196 {
1197 arc_return_buf(buf, FTAG);
1198 VERIFY(arc_buf_remove_ref(buf, FTAG));
1199 }
1200
1201 /*
1202 * When possible directly assign passed loaned arc buffer to a dbuf.
1203 * If this is not possible copy the contents of passed arc buf via
1204 * dmu_write().
1205 */
1206 void
1207 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1208 dmu_tx_t *tx)
1209 {
1210 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1211 dnode_t *dn;
1212 dmu_buf_impl_t *db;
1213 uint32_t blksz = (uint32_t)arc_buf_size(buf);
1214 uint64_t blkid;
1215
1216 DB_DNODE_ENTER(dbuf);
1217 dn = DB_DNODE(dbuf);
1218 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1219 blkid = dbuf_whichblock(dn, offset);
1220 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1221 rw_exit(&dn->dn_struct_rwlock);
1222 DB_DNODE_EXIT(dbuf);
1223
1224 if (offset == db->db.db_offset && blksz == db->db.db_size) {
1225 dbuf_assign_arcbuf(db, buf, tx);
1226 dbuf_rele(db, FTAG);
1227 } else {
1228 objset_t *os;
1229 uint64_t object;
1230
1231 DB_DNODE_ENTER(dbuf);
1232 dn = DB_DNODE(dbuf);
1233 os = dn->dn_objset;
1234 object = dn->dn_object;
1235 DB_DNODE_EXIT(dbuf);
1236
1237 dbuf_rele(db, FTAG);
1238 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1239 dmu_return_arcbuf(buf);
1240 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1241 }
1242 }
1243
1244 typedef struct {
1245 dbuf_dirty_record_t *dsa_dr;
1246 dmu_sync_cb_t *dsa_done;
1247 zgd_t *dsa_zgd;
1248 dmu_tx_t *dsa_tx;
1249 } dmu_sync_arg_t;
1250
1251 /* ARGSUSED */
1252 static void
1253 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1254 {
1255 dmu_sync_arg_t *dsa = varg;
1256 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1257 blkptr_t *bp = zio->io_bp;
1258
1259 if (zio->io_error == 0) {
1260 if (BP_IS_HOLE(bp)) {
1261 /*
1262 * A block of zeros may compress to a hole, but the
1263 * block size still needs to be known for replay.
1264 */
1265 BP_SET_LSIZE(bp, db->db_size);
1266 } else {
1267 ASSERT(BP_GET_LEVEL(bp) == 0);
1268 bp->blk_fill = 1;
1269 }
1270 }
1271 }
1272
1273 static void
1274 dmu_sync_late_arrival_ready(zio_t *zio)
1275 {
1276 dmu_sync_ready(zio, NULL, zio->io_private);
1277 }
1278
1279 /* ARGSUSED */
1280 static void
1281 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1282 {
1283 dmu_sync_arg_t *dsa = varg;
1284 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1285 dmu_buf_impl_t *db = dr->dr_dbuf;
1286
1287 mutex_enter(&db->db_mtx);
1288 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1289 if (zio->io_error == 0) {
1290 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1291 if (dr->dt.dl.dr_nopwrite) {
1292 blkptr_t *bp = zio->io_bp;
1293 blkptr_t *bp_orig = &zio->io_bp_orig;
1294 uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1295
1296 ASSERT(BP_EQUAL(bp, bp_orig));
1297 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1298 ASSERT(zio_checksum_table[chksum].ci_dedup);
1299 }
1300 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1301 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1302 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1303 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by))
1304 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1305 } else {
1306 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1307 }
1308 cv_broadcast(&db->db_changed);
1309 mutex_exit(&db->db_mtx);
1310
1311 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1312
1313 kmem_free(dsa, sizeof (*dsa));
1314 }
1315
1316 static void
1317 dmu_sync_late_arrival_done(zio_t *zio)
1318 {
1319 blkptr_t *bp = zio->io_bp;
1320 dmu_sync_arg_t *dsa = zio->io_private;
1321 blkptr_t *bp_orig = &zio->io_bp_orig;
1322
1323 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1324 /*
1325 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE)
1326 * then there is nothing to do here. Otherwise, free the
1327 * newly allocated block in this txg.
1328 */
1329 if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
1330 ASSERT(BP_EQUAL(bp, bp_orig));
1331 } else {
1332 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1333 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1334 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1335 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1336 }
1337 }
1338
1339 dmu_tx_commit(dsa->dsa_tx);
1340
1341 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1342
1343 kmem_free(dsa, sizeof (*dsa));
1344 }
1345
1346 static int
1347 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1348 zio_prop_t *zp, zbookmark_t *zb)
1349 {
1350 dmu_sync_arg_t *dsa;
1351 dmu_tx_t *tx;
1352
1353 tx = dmu_tx_create(os);
1354 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1355 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1356 dmu_tx_abort(tx);
1357 /* Make zl_get_data do txg_waited_synced() */
1358 return (SET_ERROR(EIO));
1359 }
1360
1361 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1362 dsa->dsa_dr = NULL;
1363 dsa->dsa_done = done;
1364 dsa->dsa_zgd = zgd;
1365 dsa->dsa_tx = tx;
1366
1367 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1368 zgd->zgd_db->db_data, zgd->zgd_db->db_size, zp,
1369 dmu_sync_late_arrival_ready, NULL, dmu_sync_late_arrival_done, dsa,
1370 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1371
1372 return (0);
1373 }
1374
1375 /*
1376 * Intent log support: sync the block associated with db to disk.
1377 * N.B. and XXX: the caller is responsible for making sure that the
1378 * data isn't changing while dmu_sync() is writing it.
1379 *
1380 * Return values:
1381 *
1382 * EEXIST: this txg has already been synced, so there's nothing to do.
1383 * The caller should not log the write.
1384 *
1385 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1386 * The caller should not log the write.
1387 *
1388 * EALREADY: this block is already in the process of being synced.
1389 * The caller should track its progress (somehow).
1390 *
1391 * EIO: could not do the I/O.
1392 * The caller should do a txg_wait_synced().
1393 *
1394 * 0: the I/O has been initiated.
1395 * The caller should log this blkptr in the done callback.
1396 * It is possible that the I/O will fail, in which case
1397 * the error will be reported to the done callback and
1398 * propagated to pio from zio_done().
1399 */
1400 int
1401 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1402 {
1403 blkptr_t *bp = zgd->zgd_bp;
1404 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1405 objset_t *os = db->db_objset;
1406 dsl_dataset_t *ds = os->os_dsl_dataset;
1407 dbuf_dirty_record_t *dr;
1408 dmu_sync_arg_t *dsa;
1409 zbookmark_t zb;
1410 zio_prop_t zp;
1411 dnode_t *dn;
1412
1413 ASSERT(pio != NULL);
1414 ASSERT(txg != 0);
1415
1416 SET_BOOKMARK(&zb, ds->ds_object,
1417 db->db.db_object, db->db_level, db->db_blkid);
1418
1419 DB_DNODE_ENTER(db);
1420 dn = DB_DNODE(db);
1421 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1422 DB_DNODE_EXIT(db);
1423
1424 /*
1425 * If we're frozen (running ziltest), we always need to generate a bp.
1426 */
1427 if (txg > spa_freeze_txg(os->os_spa))
1428 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1429
1430 /*
1431 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1432 * and us. If we determine that this txg is not yet syncing,
1433 * but it begins to sync a moment later, that's OK because the
1434 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1435 */
1436 mutex_enter(&db->db_mtx);
1437
1438 if (txg <= spa_last_synced_txg(os->os_spa)) {
1439 /*
1440 * This txg has already synced. There's nothing to do.
1441 */
1442 mutex_exit(&db->db_mtx);
1443 return (SET_ERROR(EEXIST));
1444 }
1445
1446 if (txg <= spa_syncing_txg(os->os_spa)) {
1447 /*
1448 * This txg is currently syncing, so we can't mess with
1449 * the dirty record anymore; just write a new log block.
1450 */
1451 mutex_exit(&db->db_mtx);
1452 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1453 }
1454
1455 dr = db->db_last_dirty;
1456 while (dr && dr->dr_txg != txg)
1457 dr = dr->dr_next;
1458
1459 if (dr == NULL) {
1460 /*
1461 * There's no dr for this dbuf, so it must have been freed.
1462 * There's no need to log writes to freed blocks, so we're done.
1463 */
1464 mutex_exit(&db->db_mtx);
1465 return (SET_ERROR(ENOENT));
1466 }
1467
1468 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
1469
1470 /*
1471 * Assume the on-disk data is X, the current syncing data is Y,
1472 * and the current in-memory data is Z (currently in dmu_sync).
1473 * X and Z are identical but Y is has been modified. Normally,
1474 * when X and Z are the same we will perform a nopwrite but if Y
1475 * is different we must disable nopwrite since the resulting write
1476 * of Y to disk can free the block containing X. If we allowed a
1477 * nopwrite to occur the block pointing to Z would reference a freed
1478 * block. Since this is a rare case we simplify this by disabling
1479 * nopwrite if the current dmu_sync-ing dbuf has been modified in
1480 * a previous transaction.
1481 */
1482 if (dr->dr_next)
1483 zp.zp_nopwrite = B_FALSE;
1484
1485 ASSERT(dr->dr_txg == txg);
1486 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1487 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1488 /*
1489 * We have already issued a sync write for this buffer,
1490 * or this buffer has already been synced. It could not
1491 * have been dirtied since, or we would have cleared the state.
1492 */
1493 mutex_exit(&db->db_mtx);
1494 return (SET_ERROR(EALREADY));
1495 }
1496
1497 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1498 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1499 mutex_exit(&db->db_mtx);
1500
1501 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1502 dsa->dsa_dr = dr;
1503 dsa->dsa_done = done;
1504 dsa->dsa_zgd = zgd;
1505 dsa->dsa_tx = NULL;
1506
1507 zio_nowait(arc_write(pio, os->os_spa, txg,
1508 bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
1509 DBUF_IS_L2COMPRESSIBLE(db), &zp, dmu_sync_ready,
1510 NULL, dmu_sync_done, dsa, ZIO_PRIORITY_SYNC_WRITE,
1511 ZIO_FLAG_CANFAIL, &zb));
1512
1513 return (0);
1514 }
1515
1516 int
1517 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1518 dmu_tx_t *tx)
1519 {
1520 dnode_t *dn;
1521 int err;
1522
1523 err = dnode_hold(os, object, FTAG, &dn);
1524 if (err)
1525 return (err);
1526 err = dnode_set_blksz(dn, size, ibs, tx);
1527 dnode_rele(dn, FTAG);
1528 return (err);
1529 }
1530
1531 void
1532 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1533 dmu_tx_t *tx)
1534 {
1535 dnode_t *dn;
1536
1537 /* XXX assumes dnode_hold will not get an i/o error */
1538 (void) dnode_hold(os, object, FTAG, &dn);
1539 ASSERT(checksum < ZIO_CHECKSUM_FUNCTIONS);
1540 dn->dn_checksum = checksum;
1541 dnode_setdirty(dn, tx);
1542 dnode_rele(dn, FTAG);
1543 }
1544
1545 void
1546 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1547 dmu_tx_t *tx)
1548 {
1549 dnode_t *dn;
1550
1551 /* XXX assumes dnode_hold will not get an i/o error */
1552 (void) dnode_hold(os, object, FTAG, &dn);
1553 ASSERT(compress < ZIO_COMPRESS_FUNCTIONS);
1554 dn->dn_compress = compress;
1555 dnode_setdirty(dn, tx);
1556 dnode_rele(dn, FTAG);
1557 }
1558
1559 int zfs_mdcomp_disable = 0;
1560
1561 /*
1562 * When the "redundant_metadata" property is set to "most", only indirect
1563 * blocks of this level and higher will have an additional ditto block.
1564 */
1565 int zfs_redundant_metadata_most_ditto_level = 2;
1566
1567 void
1568 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
1569 {
1570 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
1571 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
1572 (wp & WP_SPILL));
1573 enum zio_checksum checksum = os->os_checksum;
1574 enum zio_compress compress = os->os_compress;
1575 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
1576 boolean_t dedup = B_FALSE;
1577 boolean_t nopwrite = B_FALSE;
1578 boolean_t dedup_verify = os->os_dedup_verify;
1579 int copies = os->os_copies;
1580
1581 /*
1582 * We maintain different write policies for each of the following
1583 * types of data:
1584 * 1. metadata
1585 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1586 * 3. all other level 0 blocks
1587 */
1588 if (ismd) {
1589 /*
1590 * XXX -- we should design a compression algorithm
1591 * that specializes in arrays of bps.
1592 */
1593 compress = zfs_mdcomp_disable ? ZIO_COMPRESS_EMPTY :
1594 ZIO_COMPRESS_LZJB;
1595
1596 /*
1597 * Metadata always gets checksummed. If the data
1598 * checksum is multi-bit correctable, and it's not a
1599 * ZBT-style checksum, then it's suitable for metadata
1600 * as well. Otherwise, the metadata checksum defaults
1601 * to fletcher4.
1602 */
1603 if (zio_checksum_table[checksum].ci_correctable < 1 ||
1604 zio_checksum_table[checksum].ci_eck)
1605 checksum = ZIO_CHECKSUM_FLETCHER_4;
1606
1607 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL ||
1608 (os->os_redundant_metadata ==
1609 ZFS_REDUNDANT_METADATA_MOST &&
1610 (level >= zfs_redundant_metadata_most_ditto_level ||
1611 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))))
1612 copies++;
1613 } else if (wp & WP_NOFILL) {
1614 ASSERT(level == 0);
1615
1616 /*
1617 * If we're writing preallocated blocks, we aren't actually
1618 * writing them so don't set any policy properties. These
1619 * blocks are currently only used by an external subsystem
1620 * outside of zfs (i.e. dump) and not written by the zio
1621 * pipeline.
1622 */
1623 compress = ZIO_COMPRESS_OFF;
1624 checksum = ZIO_CHECKSUM_NOPARITY;
1625 } else {
1626 compress = zio_compress_select(dn->dn_compress, compress);
1627
1628 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
1629 zio_checksum_select(dn->dn_checksum, checksum) :
1630 dedup_checksum;
1631
1632 /*
1633 * Determine dedup setting. If we are in dmu_sync(),
1634 * we won't actually dedup now because that's all
1635 * done in syncing context; but we do want to use the
1636 * dedup checkum. If the checksum is not strong
1637 * enough to ensure unique signatures, force
1638 * dedup_verify.
1639 */
1640 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
1641 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
1642 if (!zio_checksum_table[checksum].ci_dedup)
1643 dedup_verify = B_TRUE;
1644 }
1645
1646 /*
1647 * Enable nopwrite if we have a cryptographically secure
1648 * checksum that has no known collisions (i.e. SHA-256)
1649 * and compression is enabled. We don't enable nopwrite if
1650 * dedup is enabled as the two features are mutually exclusive.
1651 */
1652 nopwrite = (!dedup && zio_checksum_table[checksum].ci_dedup &&
1653 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
1654 }
1655
1656 zp->zp_checksum = checksum;
1657 zp->zp_compress = compress;
1658 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
1659 zp->zp_level = level;
1660 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
1661 zp->zp_dedup = dedup;
1662 zp->zp_dedup_verify = dedup && dedup_verify;
1663 zp->zp_nopwrite = nopwrite;
1664 }
1665
1666 int
1667 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
1668 {
1669 dnode_t *dn;
1670 int i, err;
1671
1672 err = dnode_hold(os, object, FTAG, &dn);
1673 if (err)
1674 return (err);
1675 /*
1676 * Sync any current changes before
1677 * we go trundling through the block pointers.
1678 */
1679 for (i = 0; i < TXG_SIZE; i++) {
1680 if (list_link_active(&dn->dn_dirty_link[i]))
1681 break;
1682 }
1683 if (i != TXG_SIZE) {
1684 dnode_rele(dn, FTAG);
1685 txg_wait_synced(dmu_objset_pool(os), 0);
1686 err = dnode_hold(os, object, FTAG, &dn);
1687 if (err)
1688 return (err);
1689 }
1690
1691 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
1692 dnode_rele(dn, FTAG);
1693
1694 return (err);
1695 }
1696
1697 void
1698 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
1699 {
1700 dnode_phys_t *dnp;
1701
1702 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1703 mutex_enter(&dn->dn_mtx);
1704
1705 dnp = dn->dn_phys;
1706
1707 doi->doi_data_block_size = dn->dn_datablksz;
1708 doi->doi_metadata_block_size = dn->dn_indblkshift ?
1709 1ULL << dn->dn_indblkshift : 0;
1710 doi->doi_type = dn->dn_type;
1711 doi->doi_bonus_type = dn->dn_bonustype;
1712 doi->doi_bonus_size = dn->dn_bonuslen;
1713 doi->doi_indirection = dn->dn_nlevels;
1714 doi->doi_checksum = dn->dn_checksum;
1715 doi->doi_compress = dn->dn_compress;
1716 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
1717 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
1718 doi->doi_fill_count = 0;
1719 for (int i = 0; i < dnp->dn_nblkptr; i++)
1720 doi->doi_fill_count += dnp->dn_blkptr[i].blk_fill;
1721
1722 mutex_exit(&dn->dn_mtx);
1723 rw_exit(&dn->dn_struct_rwlock);
1724 }
1725
1726 /*
1727 * Get information on a DMU object.
1728 * If doi is NULL, just indicates whether the object exists.
1729 */
1730 int
1731 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
1732 {
1733 dnode_t *dn;
1734 int err = dnode_hold(os, object, FTAG, &dn);
1735
1736 if (err)
1737 return (err);
1738
1739 if (doi != NULL)
1740 dmu_object_info_from_dnode(dn, doi);
1741
1742 dnode_rele(dn, FTAG);
1743 return (0);
1744 }
1745
1746 /*
1747 * As above, but faster; can be used when you have a held dbuf in hand.
1748 */
1749 void
1750 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
1751 {
1752 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1753
1754 DB_DNODE_ENTER(db);
1755 dmu_object_info_from_dnode(DB_DNODE(db), doi);
1756 DB_DNODE_EXIT(db);
1757 }
1758
1759 /*
1760 * Faster still when you only care about the size.
1761 * This is specifically optimized for zfs_getattr().
1762 */
1763 void
1764 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
1765 u_longlong_t *nblk512)
1766 {
1767 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1768 dnode_t *dn;
1769
1770 DB_DNODE_ENTER(db);
1771 dn = DB_DNODE(db);
1772
1773 *blksize = dn->dn_datablksz;
1774 /* add 1 for dnode space */
1775 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
1776 SPA_MINBLOCKSHIFT) + 1;
1777 DB_DNODE_EXIT(db);
1778 }
1779
1780 void
1781 byteswap_uint64_array(void *vbuf, size_t size)
1782 {
1783 uint64_t *buf = vbuf;
1784 size_t count = size >> 3;
1785 int i;
1786
1787 ASSERT((size & 7) == 0);
1788
1789 for (i = 0; i < count; i++)
1790 buf[i] = BSWAP_64(buf[i]);
1791 }
1792
1793 void
1794 byteswap_uint32_array(void *vbuf, size_t size)
1795 {
1796 uint32_t *buf = vbuf;
1797 size_t count = size >> 2;
1798 int i;
1799
1800 ASSERT((size & 3) == 0);
1801
1802 for (i = 0; i < count; i++)
1803 buf[i] = BSWAP_32(buf[i]);
1804 }
1805
1806 void
1807 byteswap_uint16_array(void *vbuf, size_t size)
1808 {
1809 uint16_t *buf = vbuf;
1810 size_t count = size >> 1;
1811 int i;
1812
1813 ASSERT((size & 1) == 0);
1814
1815 for (i = 0; i < count; i++)
1816 buf[i] = BSWAP_16(buf[i]);
1817 }
1818
1819 /* ARGSUSED */
1820 void
1821 byteswap_uint8_array(void *vbuf, size_t size)
1822 {
1823 }
1824
1825 void
1826 dmu_init(void)
1827 {
1828 zfs_dbgmsg_init();
1829 sa_cache_init();
1830 xuio_stat_init();
1831 dmu_objset_init();
1832 dnode_init();
1833 dbuf_init();
1834 zfetch_init();
1835 l2arc_init();
1836 arc_init();
1837 }
1838
1839 void
1840 dmu_fini(void)
1841 {
1842 arc_fini(); /* arc depends on l2arc, so arc must go first */
1843 l2arc_fini();
1844 zfetch_fini();
1845 dbuf_fini();
1846 dnode_fini();
1847 dmu_objset_fini();
1848 xuio_stat_fini();
1849 sa_cache_fini();
1850 zfs_dbgmsg_fini();
1851 }