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