]> git.proxmox.com Git - mirror_zfs.git/blob - module/zfs/dmu.c
projects / mirror_zfs.git / blob
? search:


5b87c81c639af00528e1db63967985b678d36b75
[mirror_zfs.git] / module / zfs / dmu.c
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;
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, int newsize, dmu_tx_t *tx)
148 {
149 dnode_t *dn = ((dmu_buf_impl_t *)db)->db_dnode;
150
151 if (dn->dn_bonus != (dmu_buf_impl_t *)db)
152 return (EINVAL);
153 if (newsize < 0 || newsize > db->db_size)
154 return (EINVAL);
155 dnode_setbonuslen(dn, newsize, tx);
156 return (0);
157 }
158
159 int
160 dmu_set_bonustype(dmu_buf_t *db, dmu_object_type_t type, dmu_tx_t *tx)
161 {
162 dnode_t *dn = ((dmu_buf_impl_t *)db)->db_dnode;
163
164 if (type > DMU_OT_NUMTYPES)
165 return (EINVAL);
166
167 if (dn->dn_bonus != (dmu_buf_impl_t *)db)
168 return (EINVAL);
169
170 dnode_setbonus_type(dn, type, tx);
171 return (0);
172 }
173
174 int
175 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
176 {
177 dnode_t *dn;
178 int error;
179
180 error = dnode_hold(os, object, FTAG, &dn);
181 dbuf_rm_spill(dn, tx);
182 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
183 dnode_rm_spill(dn, tx);
184 rw_exit(&dn->dn_struct_rwlock);
185 dnode_rele(dn, FTAG);
186 return (error);
187 }
188
189 /*
190 * returns ENOENT, EIO, or 0.
191 */
192 int
193 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
194 {
195 dnode_t *dn;
196 dmu_buf_impl_t *db;
197 int error;
198
199 error = dnode_hold(os, object, FTAG, &dn);
200 if (error)
201 return (error);
202
203 rw_enter(&dn->dn_struct_rwlock, RW_READER);
204 if (dn->dn_bonus == NULL) {
205 rw_exit(&dn->dn_struct_rwlock);
206 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
207 if (dn->dn_bonus == NULL)
208 dbuf_create_bonus(dn);
209 }
210 db = dn->dn_bonus;
211 rw_exit(&dn->dn_struct_rwlock);
212
213 /* as long as the bonus buf is held, the dnode will be held */
214 if (refcount_add(&db->db_holds, tag) == 1)
215 VERIFY(dnode_add_ref(dn, db));
216
217 dnode_rele(dn, FTAG);
218
219 VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH));
220
221 *dbp = &db->db;
222 return (0);
223 }
224
225 /*
226 * returns ENOENT, EIO, or 0.
227 *
228 * This interface will allocate a blank spill dbuf when a spill blk
229 * doesn't already exist on the dnode.
230 *
231 * if you only want to find an already existing spill db, then
232 * dmu_spill_hold_existing() should be used.
233 */
234 int
235 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
236 {
237 dmu_buf_impl_t *db = NULL;
238 int err;
239
240 if ((flags & DB_RF_HAVESTRUCT) == 0)
241 rw_enter(&dn->dn_struct_rwlock, RW_READER);
242
243 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
244
245 if ((flags & DB_RF_HAVESTRUCT) == 0)
246 rw_exit(&dn->dn_struct_rwlock);
247
248 ASSERT(db != NULL);
249 err = dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | flags);
250 *dbp = &db->db;
251 return (err);
252 }
253
254 int
255 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
256 {
257 dnode_t *dn = ((dmu_buf_impl_t *)bonus)->db_dnode;
258 int err;
259
260 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA)
261 return (EINVAL);
262 rw_enter(&dn->dn_struct_rwlock, RW_READER);
263
264 if (!dn->dn_have_spill) {
265 rw_exit(&dn->dn_struct_rwlock);
266 return (ENOENT);
267 }
268 err = dmu_spill_hold_by_dnode(dn, DB_RF_HAVESTRUCT, tag, dbp);
269 rw_exit(&dn->dn_struct_rwlock);
270 return (err);
271 }
272
273 int
274 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
275 {
276 return (dmu_spill_hold_by_dnode(((dmu_buf_impl_t *)bonus)->db_dnode,
277 0, tag, dbp));
278 }
279
280 /*
281 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
282 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
283 * and can induce severe lock contention when writing to several files
284 * whose dnodes are in the same block.
285 */
286 static int
287 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
288 int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
289 {
290 dsl_pool_t *dp = NULL;
291 dmu_buf_t **dbp;
292 uint64_t blkid, nblks, i;
293 uint32_t dbuf_flags;
294 int err;
295 zio_t *zio;
296 hrtime_t start;
297
298 ASSERT(length <= DMU_MAX_ACCESS);
299
300 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT;
301 if (flags & DMU_READ_NO_PREFETCH || length > zfetch_array_rd_sz)
302 dbuf_flags |= DB_RF_NOPREFETCH;
303
304 rw_enter(&dn->dn_struct_rwlock, RW_READER);
305 if (dn->dn_datablkshift) {
306 int blkshift = dn->dn_datablkshift;
307 nblks = (P2ROUNDUP(offset+length, 1ULL<<blkshift) -
308 P2ALIGN(offset, 1ULL<<blkshift)) >> blkshift;
309 } else {
310 if (offset + length > dn->dn_datablksz) {
311 zfs_panic_recover("zfs: accessing past end of object "
312 "%llx/%llx (size=%u access=%llu+%llu)",
313 (longlong_t)dn->dn_objset->
314 os_dsl_dataset->ds_object,
315 (longlong_t)dn->dn_object, dn->dn_datablksz,
316 (longlong_t)offset, (longlong_t)length);
317 rw_exit(&dn->dn_struct_rwlock);
318 return (EIO);
319 }
320 nblks = 1;
321 }
322 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
323
324 if (dn->dn_objset->os_dsl_dataset)
325 dp = dn->dn_objset->os_dsl_dataset->ds_dir->dd_pool;
326 if (dp && dsl_pool_sync_context(dp))
327 start = gethrtime();
328 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
329 blkid = dbuf_whichblock(dn, offset);
330 for (i = 0; i < nblks; i++) {
331 dmu_buf_impl_t *db = dbuf_hold(dn, blkid+i, tag);
332 if (db == NULL) {
333 rw_exit(&dn->dn_struct_rwlock);
334 dmu_buf_rele_array(dbp, nblks, tag);
335 zio_nowait(zio);
336 return (EIO);
337 }
338 /* initiate async i/o */
339 if (read) {
340 (void) dbuf_read(db, zio, dbuf_flags);
341 }
342 dbp[i] = &db->db;
343 }
344 rw_exit(&dn->dn_struct_rwlock);
345
346 /* wait for async i/o */
347 err = zio_wait(zio);
348 /* track read overhead when we are in sync context */
349 if (dp && dsl_pool_sync_context(dp))
350 dp->dp_read_overhead += gethrtime() - start;
351 if (err) {
352 dmu_buf_rele_array(dbp, nblks, tag);
353 return (err);
354 }
355
356 /* wait for other io to complete */
357 if (read) {
358 for (i = 0; i < nblks; i++) {
359 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
360 mutex_enter(&db->db_mtx);
361 while (db->db_state == DB_READ ||
362 db->db_state == DB_FILL)
363 cv_wait(&db->db_changed, &db->db_mtx);
364 if (db->db_state == DB_UNCACHED)
365 err = EIO;
366 mutex_exit(&db->db_mtx);
367 if (err) {
368 dmu_buf_rele_array(dbp, nblks, tag);
369 return (err);
370 }
371 }
372 }
373
374 *numbufsp = nblks;
375 *dbpp = dbp;
376 return (0);
377 }
378
379 static int
380 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
381 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
382 {
383 dnode_t *dn;
384 int err;
385
386 err = dnode_hold(os, object, FTAG, &dn);
387 if (err)
388 return (err);
389
390 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
391 numbufsp, dbpp, DMU_READ_PREFETCH);
392
393 dnode_rele(dn, FTAG);
394
395 return (err);
396 }
397
398 int
399 dmu_buf_hold_array_by_bonus(dmu_buf_t *db, uint64_t offset,
400 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
401 {
402 dnode_t *dn = ((dmu_buf_impl_t *)db)->db_dnode;
403 int err;
404
405 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
406 numbufsp, dbpp, DMU_READ_PREFETCH);
407
408 return (err);
409 }
410
411 void
412 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
413 {
414 int i;
415 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
416
417 if (numbufs == 0)
418 return;
419
420 for (i = 0; i < numbufs; i++) {
421 if (dbp[i])
422 dbuf_rele(dbp[i], tag);
423 }
424
425 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
426 }
427
428 void
429 dmu_prefetch(objset_t *os, uint64_t object, uint64_t offset, uint64_t len)
430 {
431 dnode_t *dn;
432 uint64_t blkid;
433 int nblks, i, err;
434
435 if (zfs_prefetch_disable)
436 return;
437
438 if (len == 0) { /* they're interested in the bonus buffer */
439 dn = os->os_meta_dnode;
440
441 if (object == 0 || object >= DN_MAX_OBJECT)
442 return;
443
444 rw_enter(&dn->dn_struct_rwlock, RW_READER);
445 blkid = dbuf_whichblock(dn, object * sizeof (dnode_phys_t));
446 dbuf_prefetch(dn, blkid);
447 rw_exit(&dn->dn_struct_rwlock);
448 return;
449 }
450
451 /*
452 * XXX - Note, if the dnode for the requested object is not
453 * already cached, we will do a *synchronous* read in the
454 * dnode_hold() call. The same is true for any indirects.
455 */
456 err = dnode_hold(os, object, FTAG, &dn);
457 if (err != 0)
458 return;
459
460 rw_enter(&dn->dn_struct_rwlock, RW_READER);
461 if (dn->dn_datablkshift) {
462 int blkshift = dn->dn_datablkshift;
463 nblks = (P2ROUNDUP(offset+len, 1<<blkshift) -
464 P2ALIGN(offset, 1<<blkshift)) >> blkshift;
465 } else {
466 nblks = (offset < dn->dn_datablksz);
467 }
468
469 if (nblks != 0) {
470 blkid = dbuf_whichblock(dn, offset);
471 for (i = 0; i < nblks; i++)
472 dbuf_prefetch(dn, blkid+i);
473 }
474
475 rw_exit(&dn->dn_struct_rwlock);
476
477 dnode_rele(dn, FTAG);
478 }
479
480 /*
481 * Get the next "chunk" of file data to free. We traverse the file from
482 * the end so that the file gets shorter over time (if we crashes in the
483 * middle, this will leave us in a better state). We find allocated file
484 * data by simply searching the allocated level 1 indirects.
485 */
486 static int
487 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t limit)
488 {
489 uint64_t len = *start - limit;
490 uint64_t blkcnt = 0;
491 uint64_t maxblks = DMU_MAX_ACCESS / (1ULL << (dn->dn_indblkshift + 1));
492 uint64_t iblkrange =
493 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
494
495 ASSERT(limit <= *start);
496
497 if (len <= iblkrange * maxblks) {
498 *start = limit;
499 return (0);
500 }
501 ASSERT(ISP2(iblkrange));
502
503 while (*start > limit && blkcnt < maxblks) {
504 int err;
505
506 /* find next allocated L1 indirect */
507 err = dnode_next_offset(dn,
508 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
509
510 /* if there are no more, then we are done */
511 if (err == ESRCH) {
512 *start = limit;
513 return (0);
514 } else if (err) {
515 return (err);
516 }
517 blkcnt += 1;
518
519 /* reset offset to end of "next" block back */
520 *start = P2ALIGN(*start, iblkrange);
521 if (*start <= limit)
522 *start = limit;
523 else
524 *start -= 1;
525 }
526 return (0);
527 }
528
529 static int
530 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
531 uint64_t length, boolean_t free_dnode)
532 {
533 dmu_tx_t *tx;
534 uint64_t object_size, start, end, len;
535 boolean_t trunc = (length == DMU_OBJECT_END);
536 int align, err;
537
538 align = 1 << dn->dn_datablkshift;
539 ASSERT(align > 0);
540 object_size = align == 1 ? dn->dn_datablksz :
541 (dn->dn_maxblkid + 1) << dn->dn_datablkshift;
542
543 end = offset + length;
544 if (trunc || end > object_size)
545 end = object_size;
546 if (end <= offset)
547 return (0);
548 length = end - offset;
549
550 while (length) {
551 start = end;
552 /* assert(offset <= start) */
553 err = get_next_chunk(dn, &start, offset);
554 if (err)
555 return (err);
556 len = trunc ? DMU_OBJECT_END : end - start;
557
558 tx = dmu_tx_create(os);
559 dmu_tx_hold_free(tx, dn->dn_object, start, len);
560 err = dmu_tx_assign(tx, TXG_WAIT);
561 if (err) {
562 dmu_tx_abort(tx);
563 return (err);
564 }
565
566 dnode_free_range(dn, start, trunc ? -1 : len, tx);
567
568 if (start == 0 && free_dnode) {
569 ASSERT(trunc);
570 dnode_free(dn, tx);
571 }
572
573 length -= end - start;
574
575 dmu_tx_commit(tx);
576 end = start;
577 }
578 return (0);
579 }
580
581 int
582 dmu_free_long_range(objset_t *os, uint64_t object,
583 uint64_t offset, uint64_t length)
584 {
585 dnode_t *dn;
586 int err;
587
588 err = dnode_hold(os, object, FTAG, &dn);
589 if (err != 0)
590 return (err);
591 err = dmu_free_long_range_impl(os, dn, offset, length, FALSE);
592 dnode_rele(dn, FTAG);
593 return (err);
594 }
595
596 int
597 dmu_free_object(objset_t *os, uint64_t object)
598 {
599 dnode_t *dn;
600 dmu_tx_t *tx;
601 int err;
602
603 err = dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED,
604 FTAG, &dn);
605 if (err != 0)
606 return (err);
607 if (dn->dn_nlevels == 1) {
608 tx = dmu_tx_create(os);
609 dmu_tx_hold_bonus(tx, object);
610 dmu_tx_hold_free(tx, dn->dn_object, 0, DMU_OBJECT_END);
611 err = dmu_tx_assign(tx, TXG_WAIT);
612 if (err == 0) {
613 dnode_free_range(dn, 0, DMU_OBJECT_END, tx);
614 dnode_free(dn, tx);
615 dmu_tx_commit(tx);
616 } else {
617 dmu_tx_abort(tx);
618 }
619 } else {
620 err = dmu_free_long_range_impl(os, dn, 0, DMU_OBJECT_END, TRUE);
621 }
622 dnode_rele(dn, FTAG);
623 return (err);
624 }
625
626 int
627 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
628 uint64_t size, dmu_tx_t *tx)
629 {
630 dnode_t *dn;
631 int err = dnode_hold(os, object, FTAG, &dn);
632 if (err)
633 return (err);
634 ASSERT(offset < UINT64_MAX);
635 ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
636 dnode_free_range(dn, offset, size, tx);
637 dnode_rele(dn, FTAG);
638 return (0);
639 }
640
641 int
642 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
643 void *buf, uint32_t flags)
644 {
645 dnode_t *dn;
646 dmu_buf_t **dbp;
647 int numbufs, err;
648
649 err = dnode_hold(os, object, FTAG, &dn);
650 if (err)
651 return (err);
652
653 /*
654 * Deal with odd block sizes, where there can't be data past the first
655 * block. If we ever do the tail block optimization, we will need to
656 * handle that here as well.
657 */
658 if (dn->dn_maxblkid == 0) {
659 int newsz = offset > dn->dn_datablksz ? 0 :
660 MIN(size, dn->dn_datablksz - offset);
661 bzero((char *)buf + newsz, size - newsz);
662 size = newsz;
663 }
664
665 while (size > 0) {
666 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
667 int i;
668
669 /*
670 * NB: we could do this block-at-a-time, but it's nice
671 * to be reading in parallel.
672 */
673 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
674 TRUE, FTAG, &numbufs, &dbp, flags);
675 if (err)
676 break;
677
678 for (i = 0; i < numbufs; i++) {
679 int tocpy;
680 int bufoff;
681 dmu_buf_t *db = dbp[i];
682
683 ASSERT(size > 0);
684
685 bufoff = offset - db->db_offset;
686 tocpy = (int)MIN(db->db_size - bufoff, size);
687
688 bcopy((char *)db->db_data + bufoff, buf, tocpy);
689
690 offset += tocpy;
691 size -= tocpy;
692 buf = (char *)buf + tocpy;
693 }
694 dmu_buf_rele_array(dbp, numbufs, FTAG);
695 }
696 dnode_rele(dn, FTAG);
697 return (err);
698 }
699
700 void
701 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
702 const void *buf, dmu_tx_t *tx)
703 {
704 dmu_buf_t **dbp;
705 int numbufs, i;
706
707 if (size == 0)
708 return;
709
710 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
711 FALSE, FTAG, &numbufs, &dbp));
712
713 for (i = 0; i < numbufs; i++) {
714 int tocpy;
715 int bufoff;
716 dmu_buf_t *db = dbp[i];
717
718 ASSERT(size > 0);
719
720 bufoff = offset - db->db_offset;
721 tocpy = (int)MIN(db->db_size - bufoff, size);
722
723 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
724
725 if (tocpy == db->db_size)
726 dmu_buf_will_fill(db, tx);
727 else
728 dmu_buf_will_dirty(db, tx);
729
730 bcopy(buf, (char *)db->db_data + bufoff, tocpy);
731
732 if (tocpy == db->db_size)
733 dmu_buf_fill_done(db, tx);
734
735 offset += tocpy;
736 size -= tocpy;
737 buf = (char *)buf + tocpy;
738 }
739 dmu_buf_rele_array(dbp, numbufs, FTAG);
740 }
741
742 void
743 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
744 dmu_tx_t *tx)
745 {
746 dmu_buf_t **dbp;
747 int numbufs, i;
748
749 if (size == 0)
750 return;
751
752 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
753 FALSE, FTAG, &numbufs, &dbp));
754
755 for (i = 0; i < numbufs; i++) {
756 dmu_buf_t *db = dbp[i];
757
758 dmu_buf_will_not_fill(db, tx);
759 }
760 dmu_buf_rele_array(dbp, numbufs, FTAG);
761 }
762
763 /*
764 * DMU support for xuio
765 */
766 kstat_t *xuio_ksp = NULL;
767
768 int
769 dmu_xuio_init(xuio_t *xuio, int nblk)
770 {
771 dmu_xuio_t *priv;
772 uio_t *uio = &xuio->xu_uio;
773
774 uio->uio_iovcnt = nblk;
775 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
776
777 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
778 priv->cnt = nblk;
779 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
780 priv->iovp = uio->uio_iov;
781 XUIO_XUZC_PRIV(xuio) = priv;
782
783 if (XUIO_XUZC_RW(xuio) == UIO_READ)
784 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
785 else
786 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
787
788 return (0);
789 }
790
791 void
792 dmu_xuio_fini(xuio_t *xuio)
793 {
794 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
795 int nblk = priv->cnt;
796
797 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
798 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
799 kmem_free(priv, sizeof (dmu_xuio_t));
800
801 if (XUIO_XUZC_RW(xuio) == UIO_READ)
802 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
803 else
804 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
805 }
806
807 /*
808 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
809 * and increase priv->next by 1.
810 */
811 int
812 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
813 {
814 struct iovec *iov;
815 uio_t *uio = &xuio->xu_uio;
816 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
817 int i = priv->next++;
818
819 ASSERT(i < priv->cnt);
820 ASSERT(off + n <= arc_buf_size(abuf));
821 iov = uio->uio_iov + i;
822 iov->iov_base = (char *)abuf->b_data + off;
823 iov->iov_len = n;
824 priv->bufs[i] = abuf;
825 return (0);
826 }
827
828 int
829 dmu_xuio_cnt(xuio_t *xuio)
830 {
831 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
832 return (priv->cnt);
833 }
834
835 arc_buf_t *
836 dmu_xuio_arcbuf(xuio_t *xuio, int i)
837 {
838 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
839
840 ASSERT(i < priv->cnt);
841 return (priv->bufs[i]);
842 }
843
844 void
845 dmu_xuio_clear(xuio_t *xuio, int i)
846 {
847 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
848
849 ASSERT(i < priv->cnt);
850 priv->bufs[i] = NULL;
851 }
852
853 static void
854 xuio_stat_init(void)
855 {
856 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
857 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
858 KSTAT_FLAG_VIRTUAL);
859 if (xuio_ksp != NULL) {
860 xuio_ksp->ks_data = &xuio_stats;
861 kstat_install(xuio_ksp);
862 }
863 }
864
865 static void
866 xuio_stat_fini(void)
867 {
868 if (xuio_ksp != NULL) {
869 kstat_delete(xuio_ksp);
870 xuio_ksp = NULL;
871 }
872 }
873
874 void
875 xuio_stat_wbuf_copied()
876 {
877 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
878 }
879
880 void
881 xuio_stat_wbuf_nocopy()
882 {
883 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
884 }
885
886 #ifdef _KERNEL
887 int
888 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
889 {
890 dmu_buf_t **dbp;
891 int numbufs, i, err;
892 xuio_t *xuio = NULL;
893
894 /*
895 * NB: we could do this block-at-a-time, but it's nice
896 * to be reading in parallel.
897 */
898 err = dmu_buf_hold_array(os, object, uio->uio_loffset, size, TRUE, FTAG,
899 &numbufs, &dbp);
900 if (err)
901 return (err);
902
903 if (uio->uio_extflg == UIO_XUIO)
904 xuio = (xuio_t *)uio;
905
906 for (i = 0; i < numbufs; i++) {
907 int tocpy;
908 int bufoff;
909 dmu_buf_t *db = dbp[i];
910
911 ASSERT(size > 0);
912
913 bufoff = uio->uio_loffset - db->db_offset;
914 tocpy = (int)MIN(db->db_size - bufoff, size);
915
916 if (xuio) {
917 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
918 arc_buf_t *dbuf_abuf = dbi->db_buf;
919 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
920 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
921 if (!err) {
922 uio->uio_resid -= tocpy;
923 uio->uio_loffset += tocpy;
924 }
925
926 if (abuf == dbuf_abuf)
927 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
928 else
929 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
930 } else {
931 err = uiomove((char *)db->db_data + bufoff, tocpy,
932 UIO_READ, uio);
933 }
934 if (err)
935 break;
936
937 size -= tocpy;
938 }
939 dmu_buf_rele_array(dbp, numbufs, FTAG);
940
941 return (err);
942 }
943
944 static int
945 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
946 {
947 dmu_buf_t **dbp;
948 int numbufs;
949 int err = 0;
950 int i;
951
952 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
953 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
954 if (err)
955 return (err);
956
957 for (i = 0; i < numbufs; i++) {
958 int tocpy;
959 int bufoff;
960 dmu_buf_t *db = dbp[i];
961
962 ASSERT(size > 0);
963
964 bufoff = uio->uio_loffset - db->db_offset;
965 tocpy = (int)MIN(db->db_size - bufoff, size);
966
967 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
968
969 if (tocpy == db->db_size)
970 dmu_buf_will_fill(db, tx);
971 else
972 dmu_buf_will_dirty(db, tx);
973
974 /*
975 * XXX uiomove could block forever (eg. nfs-backed
976 * pages). There needs to be a uiolockdown() function
977 * to lock the pages in memory, so that uiomove won't
978 * block.
979 */
980 err = uiomove((char *)db->db_data + bufoff, tocpy,
981 UIO_WRITE, uio);
982
983 if (tocpy == db->db_size)
984 dmu_buf_fill_done(db, tx);
985
986 if (err)
987 break;
988
989 size -= tocpy;
990 }
991
992 dmu_buf_rele_array(dbp, numbufs, FTAG);
993 return (err);
994 }
995
996 int
997 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
998 dmu_tx_t *tx)
999 {
1000 if (size == 0)
1001 return (0);
1002
1003 return (dmu_write_uio_dnode(((dmu_buf_impl_t *)zdb)->db_dnode,
1004 uio, size, tx));
1005 }
1006
1007 int
1008 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1009 dmu_tx_t *tx)
1010 {
1011 dnode_t *dn;
1012 int err;
1013
1014 if (size == 0)
1015 return (0);
1016
1017 err = dnode_hold(os, object, FTAG, &dn);
1018 if (err)
1019 return (err);
1020
1021 err = dmu_write_uio_dnode(dn, uio, size, tx);
1022
1023 dnode_rele(dn, FTAG);
1024
1025 return (err);
1026 }
1027
1028 int
1029 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1030 page_t *pp, dmu_tx_t *tx)
1031 {
1032 dmu_buf_t **dbp;
1033 int numbufs, i;
1034 int err;
1035
1036 if (size == 0)
1037 return (0);
1038
1039 err = dmu_buf_hold_array(os, object, offset, size,
1040 FALSE, FTAG, &numbufs, &dbp);
1041 if (err)
1042 return (err);
1043
1044 for (i = 0; i < numbufs; i++) {
1045 int tocpy, copied, thiscpy;
1046 int bufoff;
1047 dmu_buf_t *db = dbp[i];
1048 caddr_t va;
1049
1050 ASSERT(size > 0);
1051 ASSERT3U(db->db_size, >=, PAGESIZE);
1052
1053 bufoff = offset - db->db_offset;
1054 tocpy = (int)MIN(db->db_size - bufoff, size);
1055
1056 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1057
1058 if (tocpy == db->db_size)
1059 dmu_buf_will_fill(db, tx);
1060 else
1061 dmu_buf_will_dirty(db, tx);
1062
1063 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1064 ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff);
1065 thiscpy = MIN(PAGESIZE, tocpy - copied);
1066 va = zfs_map_page(pp, S_READ);
1067 bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1068 zfs_unmap_page(pp, va);
1069 pp = pp->p_next;
1070 bufoff += PAGESIZE;
1071 }
1072
1073 if (tocpy == db->db_size)
1074 dmu_buf_fill_done(db, tx);
1075
1076 offset += tocpy;
1077 size -= tocpy;
1078 }
1079 dmu_buf_rele_array(dbp, numbufs, FTAG);
1080 return (err);
1081 }
1082 #endif
1083
1084 /*
1085 * Allocate a loaned anonymous arc buffer.
1086 */
1087 arc_buf_t *
1088 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1089 {
1090 dnode_t *dn = ((dmu_buf_impl_t *)handle)->db_dnode;
1091
1092 return (arc_loan_buf(dn->dn_objset->os_spa, size));
1093 }
1094
1095 /*
1096 * Free a loaned arc buffer.
1097 */
1098 void
1099 dmu_return_arcbuf(arc_buf_t *buf)
1100 {
1101 arc_return_buf(buf, FTAG);
1102 VERIFY(arc_buf_remove_ref(buf, FTAG) == 1);
1103 }
1104
1105 /*
1106 * When possible directly assign passed loaned arc buffer to a dbuf.
1107 * If this is not possible copy the contents of passed arc buf via
1108 * dmu_write().
1109 */
1110 void
1111 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1112 dmu_tx_t *tx)
1113 {
1114 dnode_t *dn = ((dmu_buf_impl_t *)handle)->db_dnode;
1115 dmu_buf_impl_t *db;
1116 uint32_t blksz = (uint32_t)arc_buf_size(buf);
1117 uint64_t blkid;
1118
1119 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1120 blkid = dbuf_whichblock(dn, offset);
1121 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1122 rw_exit(&dn->dn_struct_rwlock);
1123
1124 if (offset == db->db.db_offset && blksz == db->db.db_size) {
1125 dbuf_assign_arcbuf(db, buf, tx);
1126 dbuf_rele(db, FTAG);
1127 } else {
1128 dbuf_rele(db, FTAG);
1129 dmu_write(dn->dn_objset, dn->dn_object, offset, blksz,
1130 buf->b_data, tx);
1131 dmu_return_arcbuf(buf);
1132 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1133 }
1134 }
1135
1136 typedef struct {
1137 dbuf_dirty_record_t *dsa_dr;
1138 dmu_sync_cb_t *dsa_done;
1139 zgd_t *dsa_zgd;
1140 dmu_tx_t *dsa_tx;
1141 } dmu_sync_arg_t;
1142
1143 /* ARGSUSED */
1144 static void
1145 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1146 {
1147 dmu_sync_arg_t *dsa = varg;
1148 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1149 dnode_t *dn = ((dmu_buf_impl_t *)db)->db_dnode;
1150 blkptr_t *bp = zio->io_bp;
1151
1152 if (zio->io_error == 0) {
1153 if (BP_IS_HOLE(bp)) {
1154 /*
1155 * A block of zeros may compress to a hole, but the
1156 * block size still needs to be known for replay.
1157 */
1158 BP_SET_LSIZE(bp, db->db_size);
1159 } else {
1160 ASSERT(BP_GET_TYPE(bp) == dn->dn_type);
1161 ASSERT(BP_GET_LEVEL(bp) == 0);
1162 bp->blk_fill = 1;
1163 }
1164 }
1165 }
1166
1167 static void
1168 dmu_sync_late_arrival_ready(zio_t *zio)
1169 {
1170 dmu_sync_ready(zio, NULL, zio->io_private);
1171 }
1172
1173 /* ARGSUSED */
1174 static void
1175 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1176 {
1177 dmu_sync_arg_t *dsa = varg;
1178 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1179 dmu_buf_impl_t *db = dr->dr_dbuf;
1180
1181 mutex_enter(&db->db_mtx);
1182 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1183 if (zio->io_error == 0) {
1184 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1185 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1186 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1187 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by))
1188 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1189 } else {
1190 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1191 }
1192 cv_broadcast(&db->db_changed);
1193 mutex_exit(&db->db_mtx);
1194
1195 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1196
1197 kmem_free(dsa, sizeof (*dsa));
1198 }
1199
1200 static void
1201 dmu_sync_late_arrival_done(zio_t *zio)
1202 {
1203 blkptr_t *bp = zio->io_bp;
1204 dmu_sync_arg_t *dsa = zio->io_private;
1205
1206 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1207 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1208 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1209 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1210 }
1211
1212 dmu_tx_commit(dsa->dsa_tx);
1213
1214 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1215
1216 kmem_free(dsa, sizeof (*dsa));
1217 }
1218
1219 static int
1220 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1221 zio_prop_t *zp, zbookmark_t *zb)
1222 {
1223 dmu_sync_arg_t *dsa;
1224 dmu_tx_t *tx;
1225
1226 tx = dmu_tx_create(os);
1227 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1228 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1229 dmu_tx_abort(tx);
1230 return (EIO); /* Make zl_get_data do txg_waited_synced() */
1231 }
1232
1233 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1234 dsa->dsa_dr = NULL;
1235 dsa->dsa_done = done;
1236 dsa->dsa_zgd = zgd;
1237 dsa->dsa_tx = tx;
1238
1239 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1240 zgd->zgd_db->db_data, zgd->zgd_db->db_size, zp,
1241 dmu_sync_late_arrival_ready, dmu_sync_late_arrival_done, dsa,
1242 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1243
1244 return (0);
1245 }
1246
1247 /*
1248 * Intent log support: sync the block associated with db to disk.
1249 * N.B. and XXX: the caller is responsible for making sure that the
1250 * data isn't changing while dmu_sync() is writing it.
1251 *
1252 * Return values:
1253 *
1254 * EEXIST: this txg has already been synced, so there's nothing to to.
1255 * The caller should not log the write.
1256 *
1257 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1258 * The caller should not log the write.
1259 *
1260 * EALREADY: this block is already in the process of being synced.
1261 * The caller should track its progress (somehow).
1262 *
1263 * EIO: could not do the I/O.
1264 * The caller should do a txg_wait_synced().
1265 *
1266 * 0: the I/O has been initiated.
1267 * The caller should log this blkptr in the done callback.
1268 * It is possible that the I/O will fail, in which case
1269 * the error will be reported to the done callback and
1270 * propagated to pio from zio_done().
1271 */
1272 int
1273 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1274 {
1275 blkptr_t *bp = zgd->zgd_bp;
1276 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1277 objset_t *os = db->db_objset;
1278 dsl_dataset_t *ds = os->os_dsl_dataset;
1279 dbuf_dirty_record_t *dr;
1280 dmu_sync_arg_t *dsa;
1281 zbookmark_t zb;
1282 zio_prop_t zp;
1283
1284 ASSERT(pio != NULL);
1285 ASSERT(BP_IS_HOLE(bp));
1286 ASSERT(txg != 0);
1287
1288 SET_BOOKMARK(&zb, ds->ds_object,
1289 db->db.db_object, db->db_level, db->db_blkid);
1290
1291 dmu_write_policy(os, db->db_dnode, db->db_level, WP_DMU_SYNC, &zp);
1292
1293 /*
1294 * If we're frozen (running ziltest), we always need to generate a bp.
1295 */
1296 if (txg > spa_freeze_txg(os->os_spa))
1297 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1298
1299 /*
1300 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1301 * and us. If we determine that this txg is not yet syncing,
1302 * but it begins to sync a moment later, that's OK because the
1303 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1304 */
1305 mutex_enter(&db->db_mtx);
1306
1307 if (txg <= spa_last_synced_txg(os->os_spa)) {
1308 /*
1309 * This txg has already synced. There's nothing to do.
1310 */
1311 mutex_exit(&db->db_mtx);
1312 return (EEXIST);
1313 }
1314
1315 if (txg <= spa_syncing_txg(os->os_spa)) {
1316 /*
1317 * This txg is currently syncing, so we can't mess with
1318 * the dirty record anymore; just write a new log block.
1319 */
1320 mutex_exit(&db->db_mtx);
1321 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1322 }
1323
1324 dr = db->db_last_dirty;
1325 while (dr && dr->dr_txg != txg)
1326 dr = dr->dr_next;
1327
1328 if (dr == NULL) {
1329 /*
1330 * There's no dr for this dbuf, so it must have been freed.
1331 * There's no need to log writes to freed blocks, so we're done.
1332 */
1333 mutex_exit(&db->db_mtx);
1334 return (ENOENT);
1335 }
1336
1337 ASSERT(dr->dr_txg == txg);
1338 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1339 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1340 /*
1341 * We have already issued a sync write for this buffer,
1342 * or this buffer has already been synced. It could not
1343 * have been dirtied since, or we would have cleared the state.
1344 */
1345 mutex_exit(&db->db_mtx);
1346 return (EALREADY);
1347 }
1348
1349 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1350 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1351 mutex_exit(&db->db_mtx);
1352
1353 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1354 dsa->dsa_dr = dr;
1355 dsa->dsa_done = done;
1356 dsa->dsa_zgd = zgd;
1357 dsa->dsa_tx = NULL;
1358
1359 zio_nowait(arc_write(pio, os->os_spa, txg,
1360 bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db), &zp,
1361 dmu_sync_ready, dmu_sync_done, dsa,
1362 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb));
1363
1364 return (0);
1365 }
1366
1367 int
1368 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1369 dmu_tx_t *tx)
1370 {
1371 dnode_t *dn;
1372 int err;
1373
1374 err = dnode_hold(os, object, FTAG, &dn);
1375 if (err)
1376 return (err);
1377 err = dnode_set_blksz(dn, size, ibs, tx);
1378 dnode_rele(dn, FTAG);
1379 return (err);
1380 }
1381
1382 void
1383 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1384 dmu_tx_t *tx)
1385 {
1386 dnode_t *dn;
1387
1388 /* XXX assumes dnode_hold will not get an i/o error */
1389 (void) dnode_hold(os, object, FTAG, &dn);
1390 ASSERT(checksum < ZIO_CHECKSUM_FUNCTIONS);
1391 dn->dn_checksum = checksum;
1392 dnode_setdirty(dn, tx);
1393 dnode_rele(dn, FTAG);
1394 }
1395
1396 void
1397 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1398 dmu_tx_t *tx)
1399 {
1400 dnode_t *dn;
1401
1402 /* XXX assumes dnode_hold will not get an i/o error */
1403 (void) dnode_hold(os, object, FTAG, &dn);
1404 ASSERT(compress < ZIO_COMPRESS_FUNCTIONS);
1405 dn->dn_compress = compress;
1406 dnode_setdirty(dn, tx);
1407 dnode_rele(dn, FTAG);
1408 }
1409
1410 int zfs_mdcomp_disable = 0;
1411
1412 void
1413 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
1414 {
1415 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
1416 boolean_t ismd = (level > 0 || dmu_ot[type].ot_metadata);
1417 enum zio_checksum checksum = os->os_checksum;
1418 enum zio_compress compress = os->os_compress;
1419 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
1420 boolean_t dedup;
1421 boolean_t dedup_verify = os->os_dedup_verify;
1422 int copies = os->os_copies;
1423
1424 /*
1425 * Determine checksum setting.
1426 */
1427 if (ismd) {
1428 /*
1429 * Metadata always gets checksummed. If the data
1430 * checksum is multi-bit correctable, and it's not a
1431 * ZBT-style checksum, then it's suitable for metadata
1432 * as well. Otherwise, the metadata checksum defaults
1433 * to fletcher4.
1434 */
1435 if (zio_checksum_table[checksum].ci_correctable < 1 ||
1436 zio_checksum_table[checksum].ci_eck)
1437 checksum = ZIO_CHECKSUM_FLETCHER_4;
1438 } else {
1439 checksum = zio_checksum_select(dn->dn_checksum, checksum);
1440 }
1441
1442 /*
1443 * Determine compression setting.
1444 */
1445 if (ismd) {
1446 /*
1447 * XXX -- we should design a compression algorithm
1448 * that specializes in arrays of bps.
1449 */
1450 compress = zfs_mdcomp_disable ? ZIO_COMPRESS_EMPTY :
1451 ZIO_COMPRESS_LZJB;
1452 } else {
1453 compress = zio_compress_select(dn->dn_compress, compress);
1454 }
1455
1456 /*
1457 * Determine dedup setting. If we are in dmu_sync(), we won't
1458 * actually dedup now because that's all done in syncing context;
1459 * but we do want to use the dedup checkum. If the checksum is not
1460 * strong enough to ensure unique signatures, force dedup_verify.
1461 */
1462 dedup = (!ismd && dedup_checksum != ZIO_CHECKSUM_OFF);
1463 if (dedup) {
1464 checksum = dedup_checksum;
1465 if (!zio_checksum_table[checksum].ci_dedup)
1466 dedup_verify = 1;
1467 }
1468
1469 if (wp & WP_DMU_SYNC)
1470 dedup = 0;
1471
1472 if (wp & WP_NOFILL) {
1473 ASSERT(!ismd && level == 0);
1474 checksum = ZIO_CHECKSUM_OFF;
1475 compress = ZIO_COMPRESS_OFF;
1476 dedup = B_FALSE;
1477 }
1478
1479 zp->zp_checksum = checksum;
1480 zp->zp_compress = compress;
1481 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
1482 zp->zp_level = level;
1483 zp->zp_copies = MIN(copies + ismd, spa_max_replication(os->os_spa));
1484 zp->zp_dedup = dedup;
1485 zp->zp_dedup_verify = dedup && dedup_verify;
1486 }
1487
1488 int
1489 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
1490 {
1491 dnode_t *dn;
1492 int i, err;
1493
1494 err = dnode_hold(os, object, FTAG, &dn);
1495 if (err)
1496 return (err);
1497 /*
1498 * Sync any current changes before
1499 * we go trundling through the block pointers.
1500 */
1501 for (i = 0; i < TXG_SIZE; i++) {
1502 if (list_link_active(&dn->dn_dirty_link[i]))
1503 break;
1504 }
1505 if (i != TXG_SIZE) {
1506 dnode_rele(dn, FTAG);
1507 txg_wait_synced(dmu_objset_pool(os), 0);
1508 err = dnode_hold(os, object, FTAG, &dn);
1509 if (err)
1510 return (err);
1511 }
1512
1513 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
1514 dnode_rele(dn, FTAG);
1515
1516 return (err);
1517 }
1518
1519 void
1520 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
1521 {
1522 dnode_phys_t *dnp;
1523
1524 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1525 mutex_enter(&dn->dn_mtx);
1526
1527 dnp = dn->dn_phys;
1528
1529 doi->doi_data_block_size = dn->dn_datablksz;
1530 doi->doi_metadata_block_size = dn->dn_indblkshift ?
1531 1ULL << dn->dn_indblkshift : 0;
1532 doi->doi_type = dn->dn_type;
1533 doi->doi_bonus_type = dn->dn_bonustype;
1534 doi->doi_bonus_size = dn->dn_bonuslen;
1535 doi->doi_indirection = dn->dn_nlevels;
1536 doi->doi_checksum = dn->dn_checksum;
1537 doi->doi_compress = dn->dn_compress;
1538 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
1539 doi->doi_max_offset = (dnp->dn_maxblkid + 1) * dn->dn_datablksz;
1540 doi->doi_fill_count = 0;
1541 for (int i = 0; i < dnp->dn_nblkptr; i++)
1542 doi->doi_fill_count += dnp->dn_blkptr[i].blk_fill;
1543
1544 mutex_exit(&dn->dn_mtx);
1545 rw_exit(&dn->dn_struct_rwlock);
1546 }
1547
1548 /*
1549 * Get information on a DMU object.
1550 * If doi is NULL, just indicates whether the object exists.
1551 */
1552 int
1553 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
1554 {
1555 dnode_t *dn;
1556 int err = dnode_hold(os, object, FTAG, &dn);
1557
1558 if (err)
1559 return (err);
1560
1561 if (doi != NULL)
1562 dmu_object_info_from_dnode(dn, doi);
1563
1564 dnode_rele(dn, FTAG);
1565 return (0);
1566 }
1567
1568 /*
1569 * As above, but faster; can be used when you have a held dbuf in hand.
1570 */
1571 void
1572 dmu_object_info_from_db(dmu_buf_t *db, dmu_object_info_t *doi)
1573 {
1574 dmu_object_info_from_dnode(((dmu_buf_impl_t *)db)->db_dnode, doi);
1575 }
1576
1577 /*
1578 * Faster still when you only care about the size.
1579 * This is specifically optimized for zfs_getattr().
1580 */
1581 void
1582 dmu_object_size_from_db(dmu_buf_t *db, uint32_t *blksize, u_longlong_t *nblk512)
1583 {
1584 dnode_t *dn = ((dmu_buf_impl_t *)db)->db_dnode;
1585
1586 *blksize = dn->dn_datablksz;
1587 /* add 1 for dnode space */
1588 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
1589 SPA_MINBLOCKSHIFT) + 1;
1590 }
1591
1592 void
1593 byteswap_uint64_array(void *vbuf, size_t size)
1594 {
1595 uint64_t *buf = vbuf;
1596 size_t count = size >> 3;
1597 int i;
1598
1599 ASSERT((size & 7) == 0);
1600
1601 for (i = 0; i < count; i++)
1602 buf[i] = BSWAP_64(buf[i]);
1603 }
1604
1605 void
1606 byteswap_uint32_array(void *vbuf, size_t size)
1607 {
1608 uint32_t *buf = vbuf;
1609 size_t count = size >> 2;
1610 int i;
1611
1612 ASSERT((size & 3) == 0);
1613
1614 for (i = 0; i < count; i++)
1615 buf[i] = BSWAP_32(buf[i]);
1616 }
1617
1618 void
1619 byteswap_uint16_array(void *vbuf, size_t size)
1620 {
1621 uint16_t *buf = vbuf;
1622 size_t count = size >> 1;
1623 int i;
1624
1625 ASSERT((size & 1) == 0);
1626
1627 for (i = 0; i < count; i++)
1628 buf[i] = BSWAP_16(buf[i]);
1629 }
1630
1631 /* ARGSUSED */
1632 void
1633 byteswap_uint8_array(void *vbuf, size_t size)
1634 {
1635 }
1636
1637 void
1638 dmu_init(void)
1639 {
1640 zfs_dbgmsg_init();
1641 dbuf_init();
1642 dnode_init();
1643 zfetch_init();
1644 arc_init();
1645 l2arc_init();
1646 xuio_stat_init();
1647 sa_cache_init();
1648 }
1649
1650 void
1651 dmu_fini(void)
1652 {
1653 arc_fini();
1654 zfetch_fini();
1655 dnode_fini();
1656 dbuf_fini();
1657 l2arc_fini();
1658 xuio_stat_fini();
1659 sa_cache_fini();
1660 zfs_dbgmsg_fini();
1661 }
ZFS On Linux mirror for PVE