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


00a7a07f40c1cea7d0cdf14484101e5f4f70cf7e
[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; /* 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 = 0;
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_PUSHPAGE | KM_NODEBUG);
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 (void) memcpy((char *)db->db_data + bufoff, buf, 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 typedef struct xuio_stats {
835 /* loaned yet not returned arc_buf */
836 kstat_named_t xuiostat_onloan_rbuf;
837 kstat_named_t xuiostat_onloan_wbuf;
838 /* whether a copy is made when loaning out a read buffer */
839 kstat_named_t xuiostat_rbuf_copied;
840 kstat_named_t xuiostat_rbuf_nocopy;
841 /* whether a copy is made when assigning a write buffer */
842 kstat_named_t xuiostat_wbuf_copied;
843 kstat_named_t xuiostat_wbuf_nocopy;
844 } xuio_stats_t;
845
846 static xuio_stats_t xuio_stats = {
847 { "onloan_read_buf", KSTAT_DATA_UINT64 },
848 { "onloan_write_buf", KSTAT_DATA_UINT64 },
849 { "read_buf_copied", KSTAT_DATA_UINT64 },
850 { "read_buf_nocopy", KSTAT_DATA_UINT64 },
851 { "write_buf_copied", KSTAT_DATA_UINT64 },
852 { "write_buf_nocopy", KSTAT_DATA_UINT64 }
853 };
854
855 #define XUIOSTAT_INCR(stat, val) \
856 atomic_add_64(&xuio_stats.stat.value.ui64, (val))
857 #define XUIOSTAT_BUMP(stat) XUIOSTAT_INCR(stat, 1)
858
859 int
860 dmu_xuio_init(xuio_t *xuio, int nblk)
861 {
862 dmu_xuio_t *priv;
863 uio_t *uio = &xuio->xu_uio;
864
865 uio->uio_iovcnt = nblk;
866 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_PUSHPAGE);
867
868 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_PUSHPAGE);
869 priv->cnt = nblk;
870 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_PUSHPAGE);
871 priv->iovp = uio->uio_iov;
872 XUIO_XUZC_PRIV(xuio) = priv;
873
874 if (XUIO_XUZC_RW(xuio) == UIO_READ)
875 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
876 else
877 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
878
879 return (0);
880 }
881
882 void
883 dmu_xuio_fini(xuio_t *xuio)
884 {
885 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
886 int nblk = priv->cnt;
887
888 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
889 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
890 kmem_free(priv, sizeof (dmu_xuio_t));
891
892 if (XUIO_XUZC_RW(xuio) == UIO_READ)
893 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
894 else
895 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
896 }
897
898 /*
899 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
900 * and increase priv->next by 1.
901 */
902 int
903 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
904 {
905 struct iovec *iov;
906 uio_t *uio = &xuio->xu_uio;
907 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
908 int i = priv->next++;
909
910 ASSERT(i < priv->cnt);
911 ASSERT(off + n <= arc_buf_size(abuf));
912 iov = uio->uio_iov + i;
913 iov->iov_base = (char *)abuf->b_data + off;
914 iov->iov_len = n;
915 priv->bufs[i] = abuf;
916 return (0);
917 }
918
919 int
920 dmu_xuio_cnt(xuio_t *xuio)
921 {
922 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
923 return (priv->cnt);
924 }
925
926 arc_buf_t *
927 dmu_xuio_arcbuf(xuio_t *xuio, int i)
928 {
929 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
930
931 ASSERT(i < priv->cnt);
932 return (priv->bufs[i]);
933 }
934
935 void
936 dmu_xuio_clear(xuio_t *xuio, int i)
937 {
938 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
939
940 ASSERT(i < priv->cnt);
941 priv->bufs[i] = NULL;
942 }
943
944 static void
945 xuio_stat_init(void)
946 {
947 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
948 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
949 KSTAT_FLAG_VIRTUAL);
950 if (xuio_ksp != NULL) {
951 xuio_ksp->ks_data = &xuio_stats;
952 kstat_install(xuio_ksp);
953 }
954 }
955
956 static void
957 xuio_stat_fini(void)
958 {
959 if (xuio_ksp != NULL) {
960 kstat_delete(xuio_ksp);
961 xuio_ksp = NULL;
962 }
963 }
964
965 void
966 xuio_stat_wbuf_copied()
967 {
968 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
969 }
970
971 void
972 xuio_stat_wbuf_nocopy()
973 {
974 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
975 }
976
977 #ifdef _KERNEL
978
979 /*
980 * Copy up to size bytes between arg_buf and req based on the data direction
981 * described by the req. If an entire req's data cannot be transfered the
982 * req's is updated such that it's current index and bv offsets correctly
983 * reference any residual data which could not be copied. The return value
984 * is the number of bytes successfully copied to arg_buf.
985 */
986 static int
987 dmu_req_copy(void *arg_buf, int size, int *offset, struct request *req)
988 {
989 struct bio_vec *bv;
990 struct req_iterator iter;
991 char *bv_buf;
992 int tocpy;
993
994 *offset = 0;
995 rq_for_each_segment(bv, req, iter) {
996
997 /* Fully consumed the passed arg_buf */
998 ASSERT3S(*offset, <=, size);
999 if (size == *offset)
1000 break;
1001
1002 /* Skip fully consumed bv's */
1003 if (bv->bv_len == 0)
1004 continue;
1005
1006 tocpy = MIN(bv->bv_len, size - *offset);
1007 ASSERT3S(tocpy, >=, 0);
1008
1009 bv_buf = page_address(bv->bv_page) + bv->bv_offset;
1010 ASSERT3P(bv_buf, !=, NULL);
1011
1012 if (rq_data_dir(req) == WRITE)
1013 memcpy(arg_buf + *offset, bv_buf, tocpy);
1014 else
1015 memcpy(bv_buf, arg_buf + *offset, tocpy);
1016
1017 *offset += tocpy;
1018 bv->bv_offset += tocpy;
1019 bv->bv_len -= tocpy;
1020 }
1021
1022 return 0;
1023 }
1024
1025 static void
1026 dmu_bio_put(struct bio *bio)
1027 {
1028 struct bio *bio_next;
1029
1030 while (bio) {
1031 bio_next = bio->bi_next;
1032 bio_put(bio);
1033 bio = bio_next;
1034 }
1035 }
1036
1037 static int
1038 dmu_bio_clone(struct bio *bio, struct bio **bio_copy)
1039 {
1040 struct bio *bio_root = NULL;
1041 struct bio *bio_last = NULL;
1042 struct bio *bio_new;
1043
1044 if (bio == NULL)
1045 return EINVAL;
1046
1047 while (bio) {
1048 bio_new = bio_clone(bio, GFP_NOIO);
1049 if (bio_new == NULL) {
1050 dmu_bio_put(bio_root);
1051 return ENOMEM;
1052 }
1053
1054 if (bio_last) {
1055 bio_last->bi_next = bio_new;
1056 bio_last = bio_new;
1057 } else {
1058 bio_root = bio_new;
1059 bio_last = bio_new;
1060 }
1061
1062 bio = bio->bi_next;
1063 }
1064
1065 *bio_copy = bio_root;
1066
1067 return 0;
1068 }
1069
1070 int
1071 dmu_read_req(objset_t *os, uint64_t object, struct request *req)
1072 {
1073 uint64_t size = blk_rq_bytes(req);
1074 uint64_t offset = blk_rq_pos(req) << 9;
1075 struct bio *bio_saved = req->bio;
1076 dmu_buf_t **dbp;
1077 int numbufs, i, err;
1078
1079 /*
1080 * NB: we could do this block-at-a-time, but it's nice
1081 * to be reading in parallel.
1082 */
1083 err = dmu_buf_hold_array(os, object, offset, size, TRUE, FTAG,
1084 &numbufs, &dbp);
1085 if (err)
1086 return (err);
1087
1088 /*
1089 * Clone the bio list so the bv->bv_offset and bv->bv_len members
1090 * can be safely modified. The original bio list is relinked in to
1091 * the request when the function exits. This is required because
1092 * some file systems blindly assume that these values will remain
1093 * constant between bio_submit() and the IO completion callback.
1094 */
1095 err = dmu_bio_clone(bio_saved, &req->bio);
1096 if (err)
1097 goto error;
1098
1099 for (i = 0; i < numbufs; i++) {
1100 int tocpy, didcpy, bufoff;
1101 dmu_buf_t *db = dbp[i];
1102
1103 bufoff = offset - db->db_offset;
1104 ASSERT3S(bufoff, >=, 0);
1105
1106 tocpy = (int)MIN(db->db_size - bufoff, size);
1107 if (tocpy == 0)
1108 break;
1109
1110 err = dmu_req_copy(db->db_data + bufoff, tocpy, &didcpy, req);
1111
1112 if (didcpy < tocpy)
1113 err = EIO;
1114
1115 if (err)
1116 break;
1117
1118 size -= tocpy;
1119 offset += didcpy;
1120 err = 0;
1121 }
1122
1123 dmu_bio_put(req->bio);
1124 req->bio = bio_saved;
1125 error:
1126 dmu_buf_rele_array(dbp, numbufs, FTAG);
1127
1128 return (err);
1129 }
1130
1131 int
1132 dmu_write_req(objset_t *os, uint64_t object, struct request *req, dmu_tx_t *tx)
1133 {
1134 uint64_t size = blk_rq_bytes(req);
1135 uint64_t offset = blk_rq_pos(req) << 9;
1136 struct bio *bio_saved = req->bio;
1137 dmu_buf_t **dbp;
1138 int numbufs;
1139 int err = 0;
1140 int i;
1141
1142 if (size == 0)
1143 return (0);
1144
1145 err = dmu_buf_hold_array(os, object, offset, size, FALSE, FTAG,
1146 &numbufs, &dbp);
1147 if (err)
1148 return (err);
1149
1150 /*
1151 * Clone the bio list so the bv->bv_offset and bv->bv_len members
1152 * can be safely modified. The original bio list is relinked in to
1153 * the request when the function exits. This is required because
1154 * some file systems blindly assume that these values will remain
1155 * constant between bio_submit() and the IO completion callback.
1156 */
1157 err = dmu_bio_clone(bio_saved, &req->bio);
1158 if (err)
1159 goto error;
1160
1161 for (i = 0; i < numbufs; i++) {
1162 int tocpy, didcpy, bufoff;
1163 dmu_buf_t *db = dbp[i];
1164
1165 bufoff = offset - db->db_offset;
1166 ASSERT3S(bufoff, >=, 0);
1167
1168 tocpy = (int)MIN(db->db_size - bufoff, size);
1169 if (tocpy == 0)
1170 break;
1171
1172 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1173
1174 if (tocpy == db->db_size)
1175 dmu_buf_will_fill(db, tx);
1176 else
1177 dmu_buf_will_dirty(db, tx);
1178
1179 err = dmu_req_copy(db->db_data + bufoff, tocpy, &didcpy, req);
1180
1181 if (tocpy == db->db_size)
1182 dmu_buf_fill_done(db, tx);
1183
1184 if (didcpy < tocpy)
1185 err = EIO;
1186
1187 if (err)
1188 break;
1189
1190 size -= tocpy;
1191 offset += didcpy;
1192 err = 0;
1193 }
1194
1195 dmu_bio_put(req->bio);
1196 req->bio = bio_saved;
1197 error:
1198 dmu_buf_rele_array(dbp, numbufs, FTAG);
1199
1200 return (err);
1201 }
1202
1203 int
1204 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
1205 {
1206 dmu_buf_t **dbp;
1207 int numbufs, i, err;
1208 xuio_t *xuio = NULL;
1209
1210 /*
1211 * NB: we could do this block-at-a-time, but it's nice
1212 * to be reading in parallel.
1213 */
1214 err = dmu_buf_hold_array(os, object, uio->uio_loffset, size, TRUE, FTAG,
1215 &numbufs, &dbp);
1216 if (err)
1217 return (err);
1218
1219 for (i = 0; i < numbufs; i++) {
1220 int tocpy;
1221 int bufoff;
1222 dmu_buf_t *db = dbp[i];
1223
1224 ASSERT(size > 0);
1225
1226 bufoff = uio->uio_loffset - db->db_offset;
1227 tocpy = (int)MIN(db->db_size - bufoff, size);
1228
1229 if (xuio) {
1230 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
1231 arc_buf_t *dbuf_abuf = dbi->db_buf;
1232 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
1233 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
1234 if (!err) {
1235 uio->uio_resid -= tocpy;
1236 uio->uio_loffset += tocpy;
1237 }
1238
1239 if (abuf == dbuf_abuf)
1240 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
1241 else
1242 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
1243 } else {
1244 err = uiomove((char *)db->db_data + bufoff, tocpy,
1245 UIO_READ, uio);
1246 }
1247 if (err)
1248 break;
1249
1250 size -= tocpy;
1251 }
1252 dmu_buf_rele_array(dbp, numbufs, FTAG);
1253
1254 return (err);
1255 }
1256
1257 static int
1258 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1259 {
1260 dmu_buf_t **dbp;
1261 int numbufs;
1262 int err = 0;
1263 int i;
1264
1265 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1266 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1267 if (err)
1268 return (err);
1269
1270 for (i = 0; i < numbufs; i++) {
1271 int tocpy;
1272 int bufoff;
1273 dmu_buf_t *db = dbp[i];
1274
1275 ASSERT(size > 0);
1276
1277 bufoff = uio->uio_loffset - db->db_offset;
1278 tocpy = (int)MIN(db->db_size - bufoff, size);
1279
1280 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1281
1282 if (tocpy == db->db_size)
1283 dmu_buf_will_fill(db, tx);
1284 else
1285 dmu_buf_will_dirty(db, tx);
1286
1287 /*
1288 * XXX uiomove could block forever (eg.nfs-backed
1289 * pages). There needs to be a uiolockdown() function
1290 * to lock the pages in memory, so that uiomove won't
1291 * block.
1292 */
1293 err = uiomove((char *)db->db_data + bufoff, tocpy,
1294 UIO_WRITE, uio);
1295
1296 if (tocpy == db->db_size)
1297 dmu_buf_fill_done(db, tx);
1298
1299 if (err)
1300 break;
1301
1302 size -= tocpy;
1303 }
1304
1305 dmu_buf_rele_array(dbp, numbufs, FTAG);
1306 return (err);
1307 }
1308
1309 int
1310 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1311 dmu_tx_t *tx)
1312 {
1313 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1314 dnode_t *dn;
1315 int err;
1316
1317 if (size == 0)
1318 return (0);
1319
1320 DB_DNODE_ENTER(db);
1321 dn = DB_DNODE(db);
1322 err = dmu_write_uio_dnode(dn, uio, size, tx);
1323 DB_DNODE_EXIT(db);
1324
1325 return (err);
1326 }
1327
1328 int
1329 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1330 dmu_tx_t *tx)
1331 {
1332 dnode_t *dn;
1333 int err;
1334
1335 if (size == 0)
1336 return (0);
1337
1338 err = dnode_hold(os, object, FTAG, &dn);
1339 if (err)
1340 return (err);
1341
1342 err = dmu_write_uio_dnode(dn, uio, size, tx);
1343
1344 dnode_rele(dn, FTAG);
1345
1346 return (err);
1347 }
1348 #endif /* _KERNEL */
1349
1350 /*
1351 * Allocate a loaned anonymous arc buffer.
1352 */
1353 arc_buf_t *
1354 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1355 {
1356 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1357 spa_t *spa;
1358
1359 DB_GET_SPA(&spa, db);
1360 return (arc_loan_buf(spa, size));
1361 }
1362
1363 /*
1364 * Free a loaned arc buffer.
1365 */
1366 void
1367 dmu_return_arcbuf(arc_buf_t *buf)
1368 {
1369 arc_return_buf(buf, FTAG);
1370 VERIFY(arc_buf_remove_ref(buf, FTAG) == 1);
1371 }
1372
1373 /*
1374 * When possible directly assign passed loaned arc buffer to a dbuf.
1375 * If this is not possible copy the contents of passed arc buf via
1376 * dmu_write().
1377 */
1378 void
1379 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1380 dmu_tx_t *tx)
1381 {
1382 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1383 dnode_t *dn;
1384 dmu_buf_impl_t *db;
1385 uint32_t blksz = (uint32_t)arc_buf_size(buf);
1386 uint64_t blkid;
1387
1388 DB_DNODE_ENTER(dbuf);
1389 dn = DB_DNODE(dbuf);
1390 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1391 blkid = dbuf_whichblock(dn, offset);
1392 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1393 rw_exit(&dn->dn_struct_rwlock);
1394 DB_DNODE_EXIT(dbuf);
1395
1396 if (offset == db->db.db_offset && blksz == db->db.db_size) {
1397 dbuf_assign_arcbuf(db, buf, tx);
1398 dbuf_rele(db, FTAG);
1399 } else {
1400 objset_t *os;
1401 uint64_t object;
1402
1403 DB_DNODE_ENTER(dbuf);
1404 dn = DB_DNODE(dbuf);
1405 os = dn->dn_objset;
1406 object = dn->dn_object;
1407 DB_DNODE_EXIT(dbuf);
1408
1409 dbuf_rele(db, FTAG);
1410 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1411 dmu_return_arcbuf(buf);
1412 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1413 }
1414 }
1415
1416 typedef struct {
1417 dbuf_dirty_record_t *dsa_dr;
1418 dmu_sync_cb_t *dsa_done;
1419 zgd_t *dsa_zgd;
1420 dmu_tx_t *dsa_tx;
1421 } dmu_sync_arg_t;
1422
1423 /* ARGSUSED */
1424 static void
1425 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1426 {
1427 dmu_sync_arg_t *dsa = varg;
1428 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1429 blkptr_t *bp = zio->io_bp;
1430
1431 if (zio->io_error == 0) {
1432 if (BP_IS_HOLE(bp)) {
1433 /*
1434 * A block of zeros may compress to a hole, but the
1435 * block size still needs to be known for replay.
1436 */
1437 BP_SET_LSIZE(bp, db->db_size);
1438 } else {
1439 ASSERT(BP_GET_LEVEL(bp) == 0);
1440 bp->blk_fill = 1;
1441 }
1442 }
1443 }
1444
1445 static void
1446 dmu_sync_late_arrival_ready(zio_t *zio)
1447 {
1448 dmu_sync_ready(zio, NULL, zio->io_private);
1449 }
1450
1451 /* ARGSUSED */
1452 static void
1453 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1454 {
1455 dmu_sync_arg_t *dsa = varg;
1456 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1457 dmu_buf_impl_t *db = dr->dr_dbuf;
1458
1459 mutex_enter(&db->db_mtx);
1460 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1461 if (zio->io_error == 0) {
1462 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1463 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1464 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1465 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by))
1466 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1467 } else {
1468 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1469 }
1470 cv_broadcast(&db->db_changed);
1471 mutex_exit(&db->db_mtx);
1472
1473 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1474
1475 kmem_free(dsa, sizeof (*dsa));
1476 }
1477
1478 static void
1479 dmu_sync_late_arrival_done(zio_t *zio)
1480 {
1481 blkptr_t *bp = zio->io_bp;
1482 dmu_sync_arg_t *dsa = zio->io_private;
1483
1484 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1485 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1486 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1487 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1488 }
1489
1490 dmu_tx_commit(dsa->dsa_tx);
1491
1492 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1493
1494 kmem_free(dsa, sizeof (*dsa));
1495 }
1496
1497 static int
1498 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1499 zio_prop_t *zp, zbookmark_t *zb)
1500 {
1501 dmu_sync_arg_t *dsa;
1502 dmu_tx_t *tx;
1503
1504 tx = dmu_tx_create(os);
1505 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1506 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1507 dmu_tx_abort(tx);
1508 return (EIO); /* Make zl_get_data do txg_waited_synced() */
1509 }
1510
1511 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_PUSHPAGE);
1512 dsa->dsa_dr = NULL;
1513 dsa->dsa_done = done;
1514 dsa->dsa_zgd = zgd;
1515 dsa->dsa_tx = tx;
1516
1517 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1518 zgd->zgd_db->db_data, zgd->zgd_db->db_size, zp,
1519 dmu_sync_late_arrival_ready, dmu_sync_late_arrival_done, dsa,
1520 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL | ZIO_FLAG_FASTWRITE, zb));
1521
1522 return (0);
1523 }
1524
1525 /*
1526 * Intent log support: sync the block associated with db to disk.
1527 * N.B. and XXX: the caller is responsible for making sure that the
1528 * data isn't changing while dmu_sync() is writing it.
1529 *
1530 * Return values:
1531 *
1532 * EEXIST: this txg has already been synced, so there's nothing to to.
1533 * The caller should not log the write.
1534 *
1535 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1536 * The caller should not log the write.
1537 *
1538 * EALREADY: this block is already in the process of being synced.
1539 * The caller should track its progress (somehow).
1540 *
1541 * EIO: could not do the I/O.
1542 * The caller should do a txg_wait_synced().
1543 *
1544 * 0: the I/O has been initiated.
1545 * The caller should log this blkptr in the done callback.
1546 * It is possible that the I/O will fail, in which case
1547 * the error will be reported to the done callback and
1548 * propagated to pio from zio_done().
1549 */
1550 int
1551 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1552 {
1553 blkptr_t *bp = zgd->zgd_bp;
1554 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1555 objset_t *os = db->db_objset;
1556 dsl_dataset_t *ds = os->os_dsl_dataset;
1557 dbuf_dirty_record_t *dr;
1558 dmu_sync_arg_t *dsa;
1559 zbookmark_t zb;
1560 zio_prop_t zp;
1561 dnode_t *dn;
1562
1563 ASSERT(pio != NULL);
1564 ASSERT(BP_IS_HOLE(bp));
1565 ASSERT(txg != 0);
1566
1567 SET_BOOKMARK(&zb, ds->ds_object,
1568 db->db.db_object, db->db_level, db->db_blkid);
1569
1570 DB_DNODE_ENTER(db);
1571 dn = DB_DNODE(db);
1572 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1573 DB_DNODE_EXIT(db);
1574
1575 /*
1576 * If we're frozen (running ziltest), we always need to generate a bp.
1577 */
1578 if (txg > spa_freeze_txg(os->os_spa))
1579 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1580
1581 /*
1582 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1583 * and us. If we determine that this txg is not yet syncing,
1584 * but it begins to sync a moment later, that's OK because the
1585 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1586 */
1587 mutex_enter(&db->db_mtx);
1588
1589 if (txg <= spa_last_synced_txg(os->os_spa)) {
1590 /*
1591 * This txg has already synced. There's nothing to do.
1592 */
1593 mutex_exit(&db->db_mtx);
1594 return (EEXIST);
1595 }
1596
1597 if (txg <= spa_syncing_txg(os->os_spa)) {
1598 /*
1599 * This txg is currently syncing, so we can't mess with
1600 * the dirty record anymore; just write a new log block.
1601 */
1602 mutex_exit(&db->db_mtx);
1603 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1604 }
1605
1606 dr = db->db_last_dirty;
1607 while (dr && dr->dr_txg != txg)
1608 dr = dr->dr_next;
1609
1610 if (dr == NULL) {
1611 /*
1612 * There's no dr for this dbuf, so it must have been freed.
1613 * There's no need to log writes to freed blocks, so we're done.
1614 */
1615 mutex_exit(&db->db_mtx);
1616 return (ENOENT);
1617 }
1618
1619 ASSERT(dr->dr_txg == txg);
1620 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1621 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1622 /*
1623 * We have already issued a sync write for this buffer,
1624 * or this buffer has already been synced. It could not
1625 * have been dirtied since, or we would have cleared the state.
1626 */
1627 mutex_exit(&db->db_mtx);
1628 return (EALREADY);
1629 }
1630
1631 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1632 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1633 mutex_exit(&db->db_mtx);
1634
1635 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_PUSHPAGE);
1636 dsa->dsa_dr = dr;
1637 dsa->dsa_done = done;
1638 dsa->dsa_zgd = zgd;
1639 dsa->dsa_tx = NULL;
1640
1641 zio_nowait(arc_write(pio, os->os_spa, txg,
1642 bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db), &zp,
1643 dmu_sync_ready, dmu_sync_done, dsa,
1644 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL | ZIO_FLAG_FASTWRITE, &zb));
1645
1646 return (0);
1647 }
1648
1649 int
1650 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1651 dmu_tx_t *tx)
1652 {
1653 dnode_t *dn;
1654 int err;
1655
1656 err = dnode_hold(os, object, FTAG, &dn);
1657 if (err)
1658 return (err);
1659 err = dnode_set_blksz(dn, size, ibs, tx);
1660 dnode_rele(dn, FTAG);
1661 return (err);
1662 }
1663
1664 void
1665 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1666 dmu_tx_t *tx)
1667 {
1668 dnode_t *dn;
1669
1670 /* XXX assumes dnode_hold will not get an i/o error */
1671 (void) dnode_hold(os, object, FTAG, &dn);
1672 ASSERT(checksum < ZIO_CHECKSUM_FUNCTIONS);
1673 dn->dn_checksum = checksum;
1674 dnode_setdirty(dn, tx);
1675 dnode_rele(dn, FTAG);
1676 }
1677
1678 void
1679 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1680 dmu_tx_t *tx)
1681 {
1682 dnode_t *dn;
1683
1684 /* XXX assumes dnode_hold will not get an i/o error */
1685 (void) dnode_hold(os, object, FTAG, &dn);
1686 ASSERT(compress < ZIO_COMPRESS_FUNCTIONS);
1687 dn->dn_compress = compress;
1688 dnode_setdirty(dn, tx);
1689 dnode_rele(dn, FTAG);
1690 }
1691
1692 int zfs_mdcomp_disable = 0;
1693
1694 void
1695 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
1696 {
1697 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
1698 boolean_t ismd = (level > 0 || dmu_ot[type].ot_metadata ||
1699 (wp & WP_SPILL));
1700 enum zio_checksum checksum = os->os_checksum;
1701 enum zio_compress compress = os->os_compress;
1702 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
1703 boolean_t dedup;
1704 boolean_t dedup_verify = os->os_dedup_verify;
1705 int copies = os->os_copies;
1706
1707 /*
1708 * Determine checksum setting.
1709 */
1710 if (ismd) {
1711 /*
1712 * Metadata always gets checksummed. If the data
1713 * checksum is multi-bit correctable, and it's not a
1714 * ZBT-style checksum, then it's suitable for metadata
1715 * as well. Otherwise, the metadata checksum defaults
1716 * to fletcher4.
1717 */
1718 if (zio_checksum_table[checksum].ci_correctable < 1 ||
1719 zio_checksum_table[checksum].ci_eck)
1720 checksum = ZIO_CHECKSUM_FLETCHER_4;
1721 } else {
1722 checksum = zio_checksum_select(dn->dn_checksum, checksum);
1723 }
1724
1725 /*
1726 * Determine compression setting.
1727 */
1728 if (ismd) {
1729 /*
1730 * XXX -- we should design a compression algorithm
1731 * that specializes in arrays of bps.
1732 */
1733 compress = zfs_mdcomp_disable ? ZIO_COMPRESS_EMPTY :
1734 ZIO_COMPRESS_LZJB;
1735 } else {
1736 compress = zio_compress_select(dn->dn_compress, compress);
1737 }
1738
1739 /*
1740 * Determine dedup setting. If we are in dmu_sync(), we won't
1741 * actually dedup now because that's all done in syncing context;
1742 * but we do want to use the dedup checkum. If the checksum is not
1743 * strong enough to ensure unique signatures, force dedup_verify.
1744 */
1745 dedup = (!ismd && dedup_checksum != ZIO_CHECKSUM_OFF);
1746 if (dedup) {
1747 checksum = dedup_checksum;
1748 if (!zio_checksum_table[checksum].ci_dedup)
1749 dedup_verify = 1;
1750 }
1751
1752 if (wp & WP_DMU_SYNC)
1753 dedup = 0;
1754
1755 if (wp & WP_NOFILL) {
1756 ASSERT(!ismd && level == 0);
1757 checksum = ZIO_CHECKSUM_OFF;
1758 compress = ZIO_COMPRESS_OFF;
1759 dedup = B_FALSE;
1760 }
1761
1762 zp->zp_checksum = checksum;
1763 zp->zp_compress = compress;
1764 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
1765 zp->zp_level = level;
1766 zp->zp_copies = MIN(copies + ismd, spa_max_replication(os->os_spa));
1767 zp->zp_dedup = dedup;
1768 zp->zp_dedup_verify = dedup && dedup_verify;
1769 }
1770
1771 int
1772 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
1773 {
1774 dnode_t *dn;
1775 int i, err;
1776
1777 err = dnode_hold(os, object, FTAG, &dn);
1778 if (err)
1779 return (err);
1780 /*
1781 * Sync any current changes before
1782 * we go trundling through the block pointers.
1783 */
1784 for (i = 0; i < TXG_SIZE; i++) {
1785 if (list_link_active(&dn->dn_dirty_link[i]))
1786 break;
1787 }
1788 if (i != TXG_SIZE) {
1789 dnode_rele(dn, FTAG);
1790 txg_wait_synced(dmu_objset_pool(os), 0);
1791 err = dnode_hold(os, object, FTAG, &dn);
1792 if (err)
1793 return (err);
1794 }
1795
1796 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
1797 dnode_rele(dn, FTAG);
1798
1799 return (err);
1800 }
1801
1802 void
1803 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
1804 {
1805 dnode_phys_t *dnp;
1806 int i;
1807
1808 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1809 mutex_enter(&dn->dn_mtx);
1810
1811 dnp = dn->dn_phys;
1812
1813 doi->doi_data_block_size = dn->dn_datablksz;
1814 doi->doi_metadata_block_size = dn->dn_indblkshift ?
1815 1ULL << dn->dn_indblkshift : 0;
1816 doi->doi_type = dn->dn_type;
1817 doi->doi_bonus_type = dn->dn_bonustype;
1818 doi->doi_bonus_size = dn->dn_bonuslen;
1819 doi->doi_indirection = dn->dn_nlevels;
1820 doi->doi_checksum = dn->dn_checksum;
1821 doi->doi_compress = dn->dn_compress;
1822 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
1823 doi->doi_max_offset = (dnp->dn_maxblkid + 1) * dn->dn_datablksz;
1824 doi->doi_fill_count = 0;
1825 for (i = 0; i < dnp->dn_nblkptr; i++)
1826 doi->doi_fill_count += dnp->dn_blkptr[i].blk_fill;
1827
1828 mutex_exit(&dn->dn_mtx);
1829 rw_exit(&dn->dn_struct_rwlock);
1830 }
1831
1832 /*
1833 * Get information on a DMU object.
1834 * If doi is NULL, just indicates whether the object exists.
1835 */
1836 int
1837 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
1838 {
1839 dnode_t *dn;
1840 int err = dnode_hold(os, object, FTAG, &dn);
1841
1842 if (err)
1843 return (err);
1844
1845 if (doi != NULL)
1846 dmu_object_info_from_dnode(dn, doi);
1847
1848 dnode_rele(dn, FTAG);
1849 return (0);
1850 }
1851
1852 /*
1853 * As above, but faster; can be used when you have a held dbuf in hand.
1854 */
1855 void
1856 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
1857 {
1858 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1859
1860 DB_DNODE_ENTER(db);
1861 dmu_object_info_from_dnode(DB_DNODE(db), doi);
1862 DB_DNODE_EXIT(db);
1863 }
1864
1865 /*
1866 * Faster still when you only care about the size.
1867 * This is specifically optimized for zfs_getattr().
1868 */
1869 void
1870 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
1871 u_longlong_t *nblk512)
1872 {
1873 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1874 dnode_t *dn;
1875
1876 DB_DNODE_ENTER(db);
1877 dn = DB_DNODE(db);
1878
1879 *blksize = dn->dn_datablksz;
1880 /* add 1 for dnode space */
1881 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
1882 SPA_MINBLOCKSHIFT) + 1;
1883 DB_DNODE_EXIT(db);
1884 }
1885
1886 void
1887 byteswap_uint64_array(void *vbuf, size_t size)
1888 {
1889 uint64_t *buf = vbuf;
1890 size_t count = size >> 3;
1891 int i;
1892
1893 ASSERT((size & 7) == 0);
1894
1895 for (i = 0; i < count; i++)
1896 buf[i] = BSWAP_64(buf[i]);
1897 }
1898
1899 void
1900 byteswap_uint32_array(void *vbuf, size_t size)
1901 {
1902 uint32_t *buf = vbuf;
1903 size_t count = size >> 2;
1904 int i;
1905
1906 ASSERT((size & 3) == 0);
1907
1908 for (i = 0; i < count; i++)
1909 buf[i] = BSWAP_32(buf[i]);
1910 }
1911
1912 void
1913 byteswap_uint16_array(void *vbuf, size_t size)
1914 {
1915 uint16_t *buf = vbuf;
1916 size_t count = size >> 1;
1917 int i;
1918
1919 ASSERT((size & 1) == 0);
1920
1921 for (i = 0; i < count; i++)
1922 buf[i] = BSWAP_16(buf[i]);
1923 }
1924
1925 /* ARGSUSED */
1926 void
1927 byteswap_uint8_array(void *vbuf, size_t size)
1928 {
1929 }
1930
1931 void
1932 dmu_init(void)
1933 {
1934 zfs_dbgmsg_init();
1935 sa_cache_init();
1936 xuio_stat_init();
1937 dmu_objset_init();
1938 dnode_init();
1939 dbuf_init();
1940 zfetch_init();
1941 dmu_tx_init();
1942 arc_init();
1943 l2arc_init();
1944 }
1945
1946 void
1947 dmu_fini(void)
1948 {
1949 l2arc_fini();
1950 arc_fini();
1951 dmu_tx_fini();
1952 zfetch_fini();
1953 dbuf_fini();
1954 dnode_fini();
1955 dmu_objset_fini();
1956 xuio_stat_fini();
1957 sa_cache_fini();
1958 zfs_dbgmsg_fini();
1959 }
1960
1961 #if defined(_KERNEL) && defined(HAVE_SPL)
1962 EXPORT_SYMBOL(dmu_bonus_hold);
1963 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus);
1964 EXPORT_SYMBOL(dmu_buf_rele_array);
1965 EXPORT_SYMBOL(dmu_free_range);
1966 EXPORT_SYMBOL(dmu_read);
1967 EXPORT_SYMBOL(dmu_write);
1968 EXPORT_SYMBOL(dmu_object_info);
1969 EXPORT_SYMBOL(dmu_object_info_from_dnode);
1970 EXPORT_SYMBOL(dmu_object_info_from_db);
1971 EXPORT_SYMBOL(dmu_object_size_from_db);
1972 EXPORT_SYMBOL(dmu_object_set_blocksize);
1973 EXPORT_SYMBOL(dmu_object_set_checksum);
1974 EXPORT_SYMBOL(dmu_object_set_compress);
1975 EXPORT_SYMBOL(dmu_request_arcbuf);
1976 EXPORT_SYMBOL(dmu_return_arcbuf);
1977 EXPORT_SYMBOL(dmu_assign_arcbuf);
1978 EXPORT_SYMBOL(dmu_buf_hold);
1979 EXPORT_SYMBOL(dmu_ot);
1980
1981 module_param(zfs_mdcomp_disable, int, 0644);
1982 MODULE_PARM_DESC(zfs_mdcomp_disable, "Disable meta data compression");
1983 #endif
ZFS On Linux mirror for PVE