1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
24 */
25
26 #include <sys/zfs_context.h>
27 #include <sys/dmu.h>
28 #include <sys/dmu_impl.h>
29 #include <sys/dbuf.h>
30 #include <sys/dmu_objset.h>
31 #include <sys/dsl_dataset.h>
32 #include <sys/dsl_dir.h>
33 #include <sys/dmu_tx.h>
34 #include <sys/spa.h>
35 #include <sys/zio.h>
36 #include <sys/dmu_zfetch.h>
37 #include <sys/sa.h>
38 #include <sys/sa_impl.h>
39
40 static void dbuf_destroy(dmu_buf_impl_t *db);
41 static int dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
42 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
43
44 /*
45 * Global data structures and functions for the dbuf cache.
46 */
47 static kmem_cache_t *dbuf_cache;
48
49 /* ARGSUSED */
50 static int
51 dbuf_cons(void *vdb, void *unused, int kmflag)
52 {
53 dmu_buf_impl_t *db = vdb;
54 bzero(db, sizeof (dmu_buf_impl_t));
55
56 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
57 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
58 refcount_create(&db->db_holds);
59 return (0);
60 }
61
62 /* ARGSUSED */
63 static void
64 dbuf_dest(void *vdb, void *unused)
65 {
66 dmu_buf_impl_t *db = vdb;
67 mutex_destroy(&db->db_mtx);
68 cv_destroy(&db->db_changed);
69 refcount_destroy(&db->db_holds);
70 }
71
72 /*
73 * dbuf hash table routines
74 */
75 static dbuf_hash_table_t dbuf_hash_table;
76
77 static uint64_t dbuf_hash_count;
78
79 static uint64_t
80 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
81 {
82 uintptr_t osv = (uintptr_t)os;
83 uint64_t crc = -1ULL;
84
85 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
86 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF];
87 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF];
88 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF];
89 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF];
90 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF];
91 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF];
92
93 crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16);
94
95 return (crc);
96 }
97
98 #define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid);
99
100 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
101 ((dbuf)->db.db_object == (obj) && \
102 (dbuf)->db_objset == (os) && \
103 (dbuf)->db_level == (level) && \
104 (dbuf)->db_blkid == (blkid))
105
106 dmu_buf_impl_t *
107 dbuf_find(dnode_t *dn, uint8_t level, uint64_t blkid)
108 {
109 dbuf_hash_table_t *h = &dbuf_hash_table;
110 objset_t *os = dn->dn_objset;
111 uint64_t obj = dn->dn_object;
112 uint64_t hv = DBUF_HASH(os, obj, level, blkid);
113 uint64_t idx = hv & h->hash_table_mask;
114 dmu_buf_impl_t *db;
115
116 mutex_enter(DBUF_HASH_MUTEX(h, idx));
117 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
118 if (DBUF_EQUAL(db, os, obj, level, blkid)) {
119 mutex_enter(&db->db_mtx);
120 if (db->db_state != DB_EVICTING) {
121 mutex_exit(DBUF_HASH_MUTEX(h, idx));
122 return (db);
123 }
124 mutex_exit(&db->db_mtx);
125 }
126 }
127 mutex_exit(DBUF_HASH_MUTEX(h, idx));
128 return (NULL);
129 }
130
131 /*
132 * Insert an entry into the hash table. If there is already an element
133 * equal to elem in the hash table, then the already existing element
134 * will be returned and the new element will not be inserted.
135 * Otherwise returns NULL.
136 */
137 static dmu_buf_impl_t *
138 dbuf_hash_insert(dmu_buf_impl_t *db)
139 {
140 dbuf_hash_table_t *h = &dbuf_hash_table;
141 objset_t *os = db->db_objset;
142 uint64_t obj = db->db.db_object;
143 int level = db->db_level;
144 uint64_t blkid = db->db_blkid;
145 uint64_t hv = DBUF_HASH(os, obj, level, blkid);
146 uint64_t idx = hv & h->hash_table_mask;
147 dmu_buf_impl_t *dbf;
148
149 mutex_enter(DBUF_HASH_MUTEX(h, idx));
150 for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) {
151 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
152 mutex_enter(&dbf->db_mtx);
153 if (dbf->db_state != DB_EVICTING) {
154 mutex_exit(DBUF_HASH_MUTEX(h, idx));
155 return (dbf);
156 }
157 mutex_exit(&dbf->db_mtx);
158 }
159 }
160
161 mutex_enter(&db->db_mtx);
162 db->db_hash_next = h->hash_table[idx];
163 h->hash_table[idx] = db;
164 mutex_exit(DBUF_HASH_MUTEX(h, idx));
165 atomic_add_64(&dbuf_hash_count, 1);
166
167 return (NULL);
168 }
169
170 /*
171 * Remove an entry from the hash table. This operation will
172 * fail if there are any existing holds on the db.
173 */
174 static void
175 dbuf_hash_remove(dmu_buf_impl_t *db)
176 {
177 dbuf_hash_table_t *h = &dbuf_hash_table;
178 uint64_t hv = DBUF_HASH(db->db_objset, db->db.db_object,
179 db->db_level, db->db_blkid);
180 uint64_t idx = hv & h->hash_table_mask;
181 dmu_buf_impl_t *dbf, **dbp;
182
183 /*
184 * We musn't hold db_mtx to maintin lock ordering:
185 * DBUF_HASH_MUTEX > db_mtx.
186 */
187 ASSERT(refcount_is_zero(&db->db_holds));
188 ASSERT(db->db_state == DB_EVICTING);
189 ASSERT(!MUTEX_HELD(&db->db_mtx));
190
191 mutex_enter(DBUF_HASH_MUTEX(h, idx));
192 dbp = &h->hash_table[idx];
193 while ((dbf = *dbp) != db) {
194 dbp = &dbf->db_hash_next;
195 ASSERT(dbf != NULL);
196 }
197 *dbp = db->db_hash_next;
198 db->db_hash_next = NULL;
199 mutex_exit(DBUF_HASH_MUTEX(h, idx));
200 atomic_add_64(&dbuf_hash_count, -1);
201 }
202
203 static arc_evict_func_t dbuf_do_evict;
204
205 static void
206 dbuf_evict_user(dmu_buf_impl_t *db)
207 {
208 ASSERT(MUTEX_HELD(&db->db_mtx));
209
210 if (db->db_level != 0 || db->db_evict_func == NULL)
211 return;
212
213 if (db->db_user_data_ptr_ptr)
214 *db->db_user_data_ptr_ptr = db->db.db_data;
215 db->db_evict_func(&db->db, db->db_user_ptr);
216 db->db_user_ptr = NULL;
217 db->db_user_data_ptr_ptr = NULL;
218 db->db_evict_func = NULL;
219 }
220
221 boolean_t
222 dbuf_is_metadata(dmu_buf_impl_t *db)
223 {
224 if (db->db_level > 0) {
225 return (B_TRUE);
226 } else {
227 boolean_t is_metadata;
228
229 DB_DNODE_ENTER(db);
230 is_metadata = dmu_ot[DB_DNODE(db)->dn_type].ot_metadata;
231 DB_DNODE_EXIT(db);
232
233 return (is_metadata);
234 }
235 }
236
237 void
238 dbuf_evict(dmu_buf_impl_t *db)
239 {
240 ASSERT(MUTEX_HELD(&db->db_mtx));
241 ASSERT(db->db_buf == NULL);
242 ASSERT(db->db_data_pending == NULL);
243
244 dbuf_clear(db);
245 dbuf_destroy(db);
246 }
247
248 void
249 dbuf_init(void)
250 {
251 uint64_t hsize = 1ULL << 16;
252 dbuf_hash_table_t *h = &dbuf_hash_table;
253 int i;
254
255 /*
256 * The hash table is big enough to fill all of physical memory
257 * with an average 4K block size. The table will take up
258 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
259 */
260 while (hsize * 4096 < physmem * PAGESIZE)
261 hsize <<= 1;
262
263 retry:
264 h->hash_table_mask = hsize - 1;
265 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
266 if (h->hash_table == NULL) {
267 /* XXX - we should really return an error instead of assert */
268 ASSERT(hsize > (1ULL << 10));
269 hsize >>= 1;
270 goto retry;
271 }
272
273 dbuf_cache = kmem_cache_create("dmu_buf_impl_t",
274 sizeof (dmu_buf_impl_t),
275 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
276
277 for (i = 0; i < DBUF_MUTEXES; i++)
278 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
279 }
280
281 void
282 dbuf_fini(void)
283 {
284 dbuf_hash_table_t *h = &dbuf_hash_table;
285 int i;
286
287 for (i = 0; i < DBUF_MUTEXES; i++)
288 mutex_destroy(&h->hash_mutexes[i]);
289 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
290 kmem_cache_destroy(dbuf_cache);
291 }
292
293 /*
294 * Other stuff.
295 */
296
297 #ifdef ZFS_DEBUG
298 static void
299 dbuf_verify(dmu_buf_impl_t *db)
300 {
301 dnode_t *dn;
302 dbuf_dirty_record_t *dr;
303
304 ASSERT(MUTEX_HELD(&db->db_mtx));
305
306 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
307 return;
308
309 ASSERT(db->db_objset != NULL);
310 DB_DNODE_ENTER(db);
311 dn = DB_DNODE(db);
312 if (dn == NULL) {
313 ASSERT(db->db_parent == NULL);
314 ASSERT(db->db_blkptr == NULL);
315 } else {
316 ASSERT3U(db->db.db_object, ==, dn->dn_object);
317 ASSERT3P(db->db_objset, ==, dn->dn_objset);
318 ASSERT3U(db->db_level, <, dn->dn_nlevels);
319 ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
320 db->db_blkid == DMU_SPILL_BLKID ||
321 !list_is_empty(&dn->dn_dbufs));
322 }
323 if (db->db_blkid == DMU_BONUS_BLKID) {
324 ASSERT(dn != NULL);
325 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
326 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
327 } else if (db->db_blkid == DMU_SPILL_BLKID) {
328 ASSERT(dn != NULL);
329 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
330 ASSERT3U(db->db.db_offset, ==, 0);
331 } else {
332 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
333 }
334
335 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
336 ASSERT(dr->dr_dbuf == db);
337
338 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
339 ASSERT(dr->dr_dbuf == db);
340
341 /*
342 * We can't assert that db_size matches dn_datablksz because it
343 * can be momentarily different when another thread is doing
344 * dnode_set_blksz().
345 */
346 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
347 dr = db->db_data_pending;
348 /*
349 * It should only be modified in syncing context, so
350 * make sure we only have one copy of the data.
351 */
352 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
353 }
354
355 /* verify db->db_blkptr */
356 if (db->db_blkptr) {
357 if (db->db_parent == dn->dn_dbuf) {
358 /* db is pointed to by the dnode */
359 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
360 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
361 ASSERT(db->db_parent == NULL);
362 else
363 ASSERT(db->db_parent != NULL);
364 if (db->db_blkid != DMU_SPILL_BLKID)
365 ASSERT3P(db->db_blkptr, ==,
366 &dn->dn_phys->dn_blkptr[db->db_blkid]);
367 } else {
368 /* db is pointed to by an indirect block */
369 int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
370 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
371 ASSERT3U(db->db_parent->db.db_object, ==,
372 db->db.db_object);
373 /*
374 * dnode_grow_indblksz() can make this fail if we don't
375 * have the struct_rwlock. XXX indblksz no longer
376 * grows. safe to do this now?
377 */
378 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
379 ASSERT3P(db->db_blkptr, ==,
380 ((blkptr_t *)db->db_parent->db.db_data +
381 db->db_blkid % epb));
382 }
383 }
384 }
385 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
386 (db->db_buf == NULL || db->db_buf->b_data) &&
387 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
388 db->db_state != DB_FILL && !dn->dn_free_txg) {
389 /*
390 * If the blkptr isn't set but they have nonzero data,
391 * it had better be dirty, otherwise we'll lose that
392 * data when we evict this buffer.
393 */
394 if (db->db_dirtycnt == 0) {
395 uint64_t *buf = db->db.db_data;
396 int i;
397
398 for (i = 0; i < db->db.db_size >> 3; i++) {
399 ASSERT(buf[i] == 0);
400 }
401 }
402 }
403 DB_DNODE_EXIT(db);
404 }
405 #endif
406
407 static void
408 dbuf_update_data(dmu_buf_impl_t *db)
409 {
410 ASSERT(MUTEX_HELD(&db->db_mtx));
411 if (db->db_level == 0 && db->db_user_data_ptr_ptr) {
412 ASSERT(!refcount_is_zero(&db->db_holds));
413 *db->db_user_data_ptr_ptr = db->db.db_data;
414 }
415 }
416
417 static void
418 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
419 {
420 ASSERT(MUTEX_HELD(&db->db_mtx));
421 ASSERT(db->db_buf == NULL || !arc_has_callback(db->db_buf));
422 db->db_buf = buf;
423 if (buf != NULL) {
424 ASSERT(buf->b_data != NULL);
425 db->db.db_data = buf->b_data;
426 if (!arc_released(buf))
427 arc_set_callback(buf, dbuf_do_evict, db);
428 dbuf_update_data(db);
429 } else {
430 dbuf_evict_user(db);
431 db->db.db_data = NULL;
432 if (db->db_state != DB_NOFILL)
433 db->db_state = DB_UNCACHED;
434 }
435 }
436
437 /*
438 * Loan out an arc_buf for read. Return the loaned arc_buf.
439 */
440 arc_buf_t *
441 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
442 {
443 arc_buf_t *abuf;
444
445 mutex_enter(&db->db_mtx);
446 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
447 int blksz = db->db.db_size;
448 spa_t *spa;
449
450 mutex_exit(&db->db_mtx);
451 DB_GET_SPA(&spa, db);
452 abuf = arc_loan_buf(spa, blksz);
453 bcopy(db->db.db_data, abuf->b_data, blksz);
454 } else {
455 abuf = db->db_buf;
456 arc_loan_inuse_buf(abuf, db);
457 dbuf_set_data(db, NULL);
458 mutex_exit(&db->db_mtx);
459 }
460 return (abuf);
461 }
462
463 uint64_t
464 dbuf_whichblock(dnode_t *dn, uint64_t offset)
465 {
466 if (dn->dn_datablkshift) {
467 return (offset >> dn->dn_datablkshift);
468 } else {
469 ASSERT3U(offset, <, dn->dn_datablksz);
470 return (0);
471 }
472 }
473
474 static void
475 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
476 {
477 dmu_buf_impl_t *db = vdb;
478
479 mutex_enter(&db->db_mtx);
480 ASSERT3U(db->db_state, ==, DB_READ);
481 /*
482 * All reads are synchronous, so we must have a hold on the dbuf
483 */
484 ASSERT(refcount_count(&db->db_holds) > 0);
485 ASSERT(db->db_buf == NULL);
486 ASSERT(db->db.db_data == NULL);
487 if (db->db_level == 0 && db->db_freed_in_flight) {
488 /* we were freed in flight; disregard any error */
489 arc_release(buf, db);
490 bzero(buf->b_data, db->db.db_size);
491 arc_buf_freeze(buf);
492 db->db_freed_in_flight = FALSE;
493 dbuf_set_data(db, buf);
494 db->db_state = DB_CACHED;
495 } else if (zio == NULL || zio->io_error == 0) {
496 dbuf_set_data(db, buf);
497 db->db_state = DB_CACHED;
498 } else {
499 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
500 ASSERT3P(db->db_buf, ==, NULL);
501 VERIFY(arc_buf_remove_ref(buf, db) == 1);
502 db->db_state = DB_UNCACHED;
503 }
504 cv_broadcast(&db->db_changed);
505 dbuf_rele_and_unlock(db, NULL);
506 }
507
508 static void
509 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t *flags)
510 {
511 dnode_t *dn;
512 spa_t *spa;
513 zbookmark_t zb;
514 uint32_t aflags = ARC_NOWAIT;
515 arc_buf_t *pbuf;
516
517 DB_DNODE_ENTER(db);
518 dn = DB_DNODE(db);
519 ASSERT(!refcount_is_zero(&db->db_holds));
520 /* We need the struct_rwlock to prevent db_blkptr from changing. */
521 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
522 ASSERT(MUTEX_HELD(&db->db_mtx));
523 ASSERT(db->db_state == DB_UNCACHED);
524 ASSERT(db->db_buf == NULL);
525
526 if (db->db_blkid == DMU_BONUS_BLKID) {
527 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
528
529 ASSERT3U(bonuslen, <=, db->db.db_size);
530 db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN);
531 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
532 if (bonuslen < DN_MAX_BONUSLEN)
533 bzero(db->db.db_data, DN_MAX_BONUSLEN);
534 if (bonuslen)
535 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
536 DB_DNODE_EXIT(db);
537 dbuf_update_data(db);
538 db->db_state = DB_CACHED;
539 mutex_exit(&db->db_mtx);
540 return;
541 }
542
543 /*
544 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
545 * processes the delete record and clears the bp while we are waiting
546 * for the dn_mtx (resulting in a "no" from block_freed).
547 */
548 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
549 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
550 BP_IS_HOLE(db->db_blkptr)))) {
551 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
552
553 dbuf_set_data(db, arc_buf_alloc(dn->dn_objset->os_spa,
554 db->db.db_size, db, type));
555 DB_DNODE_EXIT(db);
556 bzero(db->db.db_data, db->db.db_size);
557 db->db_state = DB_CACHED;
558 *flags |= DB_RF_CACHED;
559 mutex_exit(&db->db_mtx);
560 return;
561 }
562
563 spa = dn->dn_objset->os_spa;
564 DB_DNODE_EXIT(db);
565
566 db->db_state = DB_READ;
567 mutex_exit(&db->db_mtx);
568
569 if (DBUF_IS_L2CACHEABLE(db))
570 aflags |= ARC_L2CACHE;
571
572 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
573 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
574 db->db.db_object, db->db_level, db->db_blkid);
575
576 dbuf_add_ref(db, NULL);
577 /* ZIO_FLAG_CANFAIL callers have to check the parent zio's error */
578
579 if (db->db_parent)
580 pbuf = db->db_parent->db_buf;
581 else
582 pbuf = db->db_objset->os_phys_buf;
583
584 (void) dsl_read(zio, spa, db->db_blkptr, pbuf,
585 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
586 (*flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
587 &aflags, &zb);
588 if (aflags & ARC_CACHED)
589 *flags |= DB_RF_CACHED;
590 }
591
592 int
593 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
594 {
595 int err = 0;
596 int havepzio = (zio != NULL);
597 int prefetch;
598 dnode_t *dn;
599
600 /*
601 * We don't have to hold the mutex to check db_state because it
602 * can't be freed while we have a hold on the buffer.
603 */
604 ASSERT(!refcount_is_zero(&db->db_holds));
605
606 if (db->db_state == DB_NOFILL)
607 return (EIO);
608
609 DB_DNODE_ENTER(db);
610 dn = DB_DNODE(db);
611 if ((flags & DB_RF_HAVESTRUCT) == 0)
612 rw_enter(&dn->dn_struct_rwlock, RW_READER);
613
614 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
615 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
616 DBUF_IS_CACHEABLE(db);
617
618 mutex_enter(&db->db_mtx);
619 if (db->db_state == DB_CACHED) {
620 mutex_exit(&db->db_mtx);
621 if (prefetch)
622 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
623 db->db.db_size, TRUE);
624 if ((flags & DB_RF_HAVESTRUCT) == 0)
625 rw_exit(&dn->dn_struct_rwlock);
626 DB_DNODE_EXIT(db);
627 } else if (db->db_state == DB_UNCACHED) {
628 spa_t *spa = dn->dn_objset->os_spa;
629
630 if (zio == NULL)
631 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
632 dbuf_read_impl(db, zio, &flags);
633
634 /* dbuf_read_impl has dropped db_mtx for us */
635
636 if (prefetch)
637 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
638 db->db.db_size, flags & DB_RF_CACHED);
639
640 if ((flags & DB_RF_HAVESTRUCT) == 0)
641 rw_exit(&dn->dn_struct_rwlock);
642 DB_DNODE_EXIT(db);
643
644 if (!havepzio)
645 err = zio_wait(zio);
646 } else {
647 mutex_exit(&db->db_mtx);
648 if (prefetch)
649 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
650 db->db.db_size, TRUE);
651 if ((flags & DB_RF_HAVESTRUCT) == 0)
652 rw_exit(&dn->dn_struct_rwlock);
653 DB_DNODE_EXIT(db);
654
655 mutex_enter(&db->db_mtx);
656 if ((flags & DB_RF_NEVERWAIT) == 0) {
657 while (db->db_state == DB_READ ||
658 db->db_state == DB_FILL) {
659 ASSERT(db->db_state == DB_READ ||
660 (flags & DB_RF_HAVESTRUCT) == 0);
661 cv_wait(&db->db_changed, &db->db_mtx);
662 }
663 if (db->db_state == DB_UNCACHED)
664 err = EIO;
665 }
666 mutex_exit(&db->db_mtx);
667 }
668
669 ASSERT(err || havepzio || db->db_state == DB_CACHED);
670 return (err);
671 }
672
673 static void
674 dbuf_noread(dmu_buf_impl_t *db)
675 {
676 ASSERT(!refcount_is_zero(&db->db_holds));
677 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
678 mutex_enter(&db->db_mtx);
679 while (db->db_state == DB_READ || db->db_state == DB_FILL)
680 cv_wait(&db->db_changed, &db->db_mtx);
681 if (db->db_state == DB_UNCACHED) {
682 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
683 spa_t *spa;
684
685 ASSERT(db->db_buf == NULL);
686 ASSERT(db->db.db_data == NULL);
687 DB_GET_SPA(&spa, db);
688 dbuf_set_data(db, arc_buf_alloc(spa, db->db.db_size, db, type));
689 db->db_state = DB_FILL;
690 } else if (db->db_state == DB_NOFILL) {
691 dbuf_set_data(db, NULL);
692 } else {
693 ASSERT3U(db->db_state, ==, DB_CACHED);
694 }
695 mutex_exit(&db->db_mtx);
696 }
697
698 /*
699 * This is our just-in-time copy function. It makes a copy of
700 * buffers, that have been modified in a previous transaction
701 * group, before we modify them in the current active group.
702 *
703 * This function is used in two places: when we are dirtying a
704 * buffer for the first time in a txg, and when we are freeing
705 * a range in a dnode that includes this buffer.
706 *
707 * Note that when we are called from dbuf_free_range() we do
708 * not put a hold on the buffer, we just traverse the active
709 * dbuf list for the dnode.
710 */
711 static void
712 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
713 {
714 dbuf_dirty_record_t *dr = db->db_last_dirty;
715
716 ASSERT(MUTEX_HELD(&db->db_mtx));
717 ASSERT(db->db.db_data != NULL);
718 ASSERT(db->db_level == 0);
719 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
720
721 if (dr == NULL ||
722 (dr->dt.dl.dr_data !=
723 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
724 return;
725
726 /*
727 * If the last dirty record for this dbuf has not yet synced
728 * and its referencing the dbuf data, either:
729 * reset the reference to point to a new copy,
730 * or (if there a no active holders)
731 * just null out the current db_data pointer.
732 */
733 ASSERT(dr->dr_txg >= txg - 2);
734 if (db->db_blkid == DMU_BONUS_BLKID) {
735 /* Note that the data bufs here are zio_bufs */
736 dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
737 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
738 bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
739 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
740 int size = db->db.db_size;
741 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
742 spa_t *spa;
743
744 DB_GET_SPA(&spa, db);
745 dr->dt.dl.dr_data = arc_buf_alloc(spa, size, db, type);
746 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
747 } else {
748 dbuf_set_data(db, NULL);
749 }
750 }
751
752 void
753 dbuf_unoverride(dbuf_dirty_record_t *dr)
754 {
755 dmu_buf_impl_t *db = dr->dr_dbuf;
756 blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
757 uint64_t txg = dr->dr_txg;
758
759 ASSERT(MUTEX_HELD(&db->db_mtx));
760 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
761 ASSERT(db->db_level == 0);
762
763 if (db->db_blkid == DMU_BONUS_BLKID ||
764 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
765 return;
766
767 ASSERT(db->db_data_pending != dr);
768
769 /* free this block */
770 if (!BP_IS_HOLE(bp)) {
771 spa_t *spa;
772
773 DB_GET_SPA(&spa, db);
774 zio_free(spa, txg, bp);
775 }
776 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
777 /*
778 * Release the already-written buffer, so we leave it in
779 * a consistent dirty state. Note that all callers are
780 * modifying the buffer, so they will immediately do
781 * another (redundant) arc_release(). Therefore, leave
782 * the buf thawed to save the effort of freezing &
783 * immediately re-thawing it.
784 */
785 arc_release(dr->dt.dl.dr_data, db);
786 }
787
788 /*
789 * Evict (if its unreferenced) or clear (if its referenced) any level-0
790 * data blocks in the free range, so that any future readers will find
791 * empty blocks. Also, if we happen accross any level-1 dbufs in the
792 * range that have not already been marked dirty, mark them dirty so
793 * they stay in memory.
794 */
795 void
796 dbuf_free_range(dnode_t *dn, uint64_t start, uint64_t end, dmu_tx_t *tx)
797 {
798 dmu_buf_impl_t *db, *db_next;
799 uint64_t txg = tx->tx_txg;
800 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
801 uint64_t first_l1 = start >> epbs;
802 uint64_t last_l1 = end >> epbs;
803
804 if (end > dn->dn_maxblkid && (end != DMU_SPILL_BLKID)) {
805 end = dn->dn_maxblkid;
806 last_l1 = end >> epbs;
807 }
808 dprintf_dnode(dn, "start=%llu end=%llu\n", start, end);
809 mutex_enter(&dn->dn_dbufs_mtx);
810 for (db = list_head(&dn->dn_dbufs); db; db = db_next) {
811 db_next = list_next(&dn->dn_dbufs, db);
812 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
813
814 if (db->db_level == 1 &&
815 db->db_blkid >= first_l1 && db->db_blkid <= last_l1) {
816 mutex_enter(&db->db_mtx);
817 if (db->db_last_dirty &&
818 db->db_last_dirty->dr_txg < txg) {
819 dbuf_add_ref(db, FTAG);
820 mutex_exit(&db->db_mtx);
821 dbuf_will_dirty(db, tx);
822 dbuf_rele(db, FTAG);
823 } else {
824 mutex_exit(&db->db_mtx);
825 }
826 }
827
828 if (db->db_level != 0)
829 continue;
830 dprintf_dbuf(db, "found buf %s\n", "");
831 if (db->db_blkid < start || db->db_blkid > end)
832 continue;
833
834 /* found a level 0 buffer in the range */
835 if (dbuf_undirty(db, tx))
836 continue;
837
838 mutex_enter(&db->db_mtx);
839 if (db->db_state == DB_UNCACHED ||
840 db->db_state == DB_NOFILL ||
841 db->db_state == DB_EVICTING) {
842 ASSERT(db->db.db_data == NULL);
843 mutex_exit(&db->db_mtx);
844 continue;
845 }
846 if (db->db_state == DB_READ || db->db_state == DB_FILL) {
847 /* will be handled in dbuf_read_done or dbuf_rele */
848 db->db_freed_in_flight = TRUE;
849 mutex_exit(&db->db_mtx);
850 continue;
851 }
852 if (refcount_count(&db->db_holds) == 0) {
853 ASSERT(db->db_buf);
854 dbuf_clear(db);
855 continue;
856 }
857 /* The dbuf is referenced */
858
859 if (db->db_last_dirty != NULL) {
860 dbuf_dirty_record_t *dr = db->db_last_dirty;
861
862 if (dr->dr_txg == txg) {
863 /*
864 * This buffer is "in-use", re-adjust the file
865 * size to reflect that this buffer may
866 * contain new data when we sync.
867 */
868 if (db->db_blkid != DMU_SPILL_BLKID &&
869 db->db_blkid > dn->dn_maxblkid)
870 dn->dn_maxblkid = db->db_blkid;
871 dbuf_unoverride(dr);
872 } else {
873 /*
874 * This dbuf is not dirty in the open context.
875 * Either uncache it (if its not referenced in
876 * the open context) or reset its contents to
877 * empty.
878 */
879 dbuf_fix_old_data(db, txg);
880 }
881 }
882 /* clear the contents if its cached */
883 if (db->db_state == DB_CACHED) {
884 ASSERT(db->db.db_data != NULL);
885 arc_release(db->db_buf, db);
886 bzero(db->db.db_data, db->db.db_size);
887 arc_buf_freeze(db->db_buf);
888 }
889
890 mutex_exit(&db->db_mtx);
891 }
892 mutex_exit(&dn->dn_dbufs_mtx);
893 }
894
895 static int
896 dbuf_block_freeable(dmu_buf_impl_t *db)
897 {
898 dsl_dataset_t *ds = db->db_objset->os_dsl_dataset;
899 uint64_t birth_txg = 0;
900
901 /*
902 * We don't need any locking to protect db_blkptr:
903 * If it's syncing, then db_last_dirty will be set
904 * so we'll ignore db_blkptr.
905 */
906 ASSERT(MUTEX_HELD(&db->db_mtx));
907 if (db->db_last_dirty)
908 birth_txg = db->db_last_dirty->dr_txg;
909 else if (db->db_blkptr)
910 birth_txg = db->db_blkptr->blk_birth;
911
912 /*
913 * If we don't exist or are in a snapshot, we can't be freed.
914 * Don't pass the bp to dsl_dataset_block_freeable() since we
915 * are holding the db_mtx lock and might deadlock if we are
916 * prefetching a dedup-ed block.
917 */
918 if (birth_txg)
919 return (ds == NULL ||
920 dsl_dataset_block_freeable(ds, NULL, birth_txg));
921 else
922 return (FALSE);
923 }
924
925 void
926 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
927 {
928 arc_buf_t *buf, *obuf;
929 int osize = db->db.db_size;
930 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
931 dnode_t *dn;
932
933 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
934
935 DB_DNODE_ENTER(db);
936 dn = DB_DNODE(db);
937
938 /* XXX does *this* func really need the lock? */
939 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
940
941 /*
942 * This call to dbuf_will_dirty() with the dn_struct_rwlock held
943 * is OK, because there can be no other references to the db
944 * when we are changing its size, so no concurrent DB_FILL can
945 * be happening.
946 */
947 /*
948 * XXX we should be doing a dbuf_read, checking the return
949 * value and returning that up to our callers
950 */
951 dbuf_will_dirty(db, tx);
952
953 /* create the data buffer for the new block */
954 buf = arc_buf_alloc(dn->dn_objset->os_spa, size, db, type);
955
956 /* copy old block data to the new block */
957 obuf = db->db_buf;
958 bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
959 /* zero the remainder */
960 if (size > osize)
961 bzero((uint8_t *)buf->b_data + osize, size - osize);
962
963 mutex_enter(&db->db_mtx);
964 dbuf_set_data(db, buf);
965 VERIFY(arc_buf_remove_ref(obuf, db) == 1);
966 db->db.db_size = size;
967
968 if (db->db_level == 0) {
969 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
970 db->db_last_dirty->dt.dl.dr_data = buf;
971 }
972 mutex_exit(&db->db_mtx);
973
974 dnode_willuse_space(dn, size-osize, tx);
975 DB_DNODE_EXIT(db);
976 }
977
978 void
979 dbuf_release_bp(dmu_buf_impl_t *db)
980 {
981 objset_t *os;
982 zbookmark_t zb;
983
984 DB_GET_OBJSET(&os, db);
985 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
986 ASSERT(arc_released(os->os_phys_buf) ||
987 list_link_active(&os->os_dsl_dataset->ds_synced_link));
988 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
989
990 zb.zb_objset = os->os_dsl_dataset ?
991 os->os_dsl_dataset->ds_object : 0;
992 zb.zb_object = db->db.db_object;
993 zb.zb_level = db->db_level;
994 zb.zb_blkid = db->db_blkid;
995 (void) arc_release_bp(db->db_buf, db,
996 db->db_blkptr, os->os_spa, &zb);
997 }
998
999 dbuf_dirty_record_t *
1000 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1001 {
1002 dnode_t *dn;
1003 objset_t *os;
1004 dbuf_dirty_record_t **drp, *dr;
1005 int drop_struct_lock = FALSE;
1006 boolean_t do_free_accounting = B_FALSE;
1007 int txgoff = tx->tx_txg & TXG_MASK;
1008
1009 ASSERT(tx->tx_txg != 0);
1010 ASSERT(!refcount_is_zero(&db->db_holds));
1011 DMU_TX_DIRTY_BUF(tx, db);
1012
1013 DB_DNODE_ENTER(db);
1014 dn = DB_DNODE(db);
1015 /*
1016 * Shouldn't dirty a regular buffer in syncing context. Private
1017 * objects may be dirtied in syncing context, but only if they
1018 * were already pre-dirtied in open context.
1019 */
1020 ASSERT(!dmu_tx_is_syncing(tx) ||
1021 BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1022 DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1023 dn->dn_objset->os_dsl_dataset == NULL);
1024 /*
1025 * We make this assert for private objects as well, but after we
1026 * check if we're already dirty. They are allowed to re-dirty
1027 * in syncing context.
1028 */
1029 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1030 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1031 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1032
1033 mutex_enter(&db->db_mtx);
1034 /*
1035 * XXX make this true for indirects too? The problem is that
1036 * transactions created with dmu_tx_create_assigned() from
1037 * syncing context don't bother holding ahead.
1038 */
1039 ASSERT(db->db_level != 0 ||
1040 db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1041 db->db_state == DB_NOFILL);
1042
1043 mutex_enter(&dn->dn_mtx);
1044 /*
1045 * Don't set dirtyctx to SYNC if we're just modifying this as we
1046 * initialize the objset.
1047 */
1048 if (dn->dn_dirtyctx == DN_UNDIRTIED &&
1049 !BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1050 dn->dn_dirtyctx =
1051 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1052 ASSERT(dn->dn_dirtyctx_firstset == NULL);
1053 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1054 }
1055 mutex_exit(&dn->dn_mtx);
1056
1057 if (db->db_blkid == DMU_SPILL_BLKID)
1058 dn->dn_have_spill = B_TRUE;
1059
1060 /*
1061 * If this buffer is already dirty, we're done.
1062 */
1063 drp = &db->db_last_dirty;
1064 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1065 db->db.db_object == DMU_META_DNODE_OBJECT);
1066 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1067 drp = &dr->dr_next;
1068 if (dr && dr->dr_txg == tx->tx_txg) {
1069 DB_DNODE_EXIT(db);
1070
1071 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1072 /*
1073 * If this buffer has already been written out,
1074 * we now need to reset its state.
1075 */
1076 dbuf_unoverride(dr);
1077 if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1078 db->db_state != DB_NOFILL)
1079 arc_buf_thaw(db->db_buf);
1080 }
1081 mutex_exit(&db->db_mtx);
1082 return (dr);
1083 }
1084
1085 /*
1086 * Only valid if not already dirty.
1087 */
1088 ASSERT(dn->dn_object == 0 ||
1089 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1090 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1091
1092 ASSERT3U(dn->dn_nlevels, >, db->db_level);
1093 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1094 dn->dn_phys->dn_nlevels > db->db_level ||
1095 dn->dn_next_nlevels[txgoff] > db->db_level ||
1096 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1097 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1098
1099 /*
1100 * We should only be dirtying in syncing context if it's the
1101 * mos or we're initializing the os or it's a special object.
1102 * However, we are allowed to dirty in syncing context provided
1103 * we already dirtied it in open context. Hence we must make
1104 * this assertion only if we're not already dirty.
1105 */
1106 os = dn->dn_objset;
1107 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1108 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1109 ASSERT(db->db.db_size != 0);
1110
1111 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1112
1113 if (db->db_blkid != DMU_BONUS_BLKID) {
1114 /*
1115 * Update the accounting.
1116 * Note: we delay "free accounting" until after we drop
1117 * the db_mtx. This keeps us from grabbing other locks
1118 * (and possibly deadlocking) in bp_get_dsize() while
1119 * also holding the db_mtx.
1120 */
1121 dnode_willuse_space(dn, db->db.db_size, tx);
1122 do_free_accounting = dbuf_block_freeable(db);
1123 }
1124
1125 /*
1126 * If this buffer is dirty in an old transaction group we need
1127 * to make a copy of it so that the changes we make in this
1128 * transaction group won't leak out when we sync the older txg.
1129 */
1130 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1131 if (db->db_level == 0) {
1132 void *data_old = db->db_buf;
1133
1134 if (db->db_state != DB_NOFILL) {
1135 if (db->db_blkid == DMU_BONUS_BLKID) {
1136 dbuf_fix_old_data(db, tx->tx_txg);
1137 data_old = db->db.db_data;
1138 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1139 /*
1140 * Release the data buffer from the cache so
1141 * that we can modify it without impacting
1142 * possible other users of this cached data
1143 * block. Note that indirect blocks and
1144 * private objects are not released until the
1145 * syncing state (since they are only modified
1146 * then).
1147 */
1148 arc_release(db->db_buf, db);
1149 dbuf_fix_old_data(db, tx->tx_txg);
1150 data_old = db->db_buf;
1151 }
1152 ASSERT(data_old != NULL);
1153 }
1154 dr->dt.dl.dr_data = data_old;
1155 } else {
1156 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1157 list_create(&dr->dt.di.dr_children,
1158 sizeof (dbuf_dirty_record_t),
1159 offsetof(dbuf_dirty_record_t, dr_dirty_node));
1160 }
1161 dr->dr_dbuf = db;
1162 dr->dr_txg = tx->tx_txg;
1163 dr->dr_next = *drp;
1164 *drp = dr;
1165
1166 /*
1167 * We could have been freed_in_flight between the dbuf_noread
1168 * and dbuf_dirty. We win, as though the dbuf_noread() had
1169 * happened after the free.
1170 */
1171 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1172 db->db_blkid != DMU_SPILL_BLKID) {
1173 mutex_enter(&dn->dn_mtx);
1174 dnode_clear_range(dn, db->db_blkid, 1, tx);
1175 mutex_exit(&dn->dn_mtx);
1176 db->db_freed_in_flight = FALSE;
1177 }
1178
1179 /*
1180 * This buffer is now part of this txg
1181 */
1182 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1183 db->db_dirtycnt += 1;
1184 ASSERT3U(db->db_dirtycnt, <=, 3);
1185
1186 mutex_exit(&db->db_mtx);
1187
1188 if (db->db_blkid == DMU_BONUS_BLKID ||
1189 db->db_blkid == DMU_SPILL_BLKID) {
1190 mutex_enter(&dn->dn_mtx);
1191 ASSERT(!list_link_active(&dr->dr_dirty_node));
1192 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1193 mutex_exit(&dn->dn_mtx);
1194 dnode_setdirty(dn, tx);
1195 DB_DNODE_EXIT(db);
1196 return (dr);
1197 } else if (do_free_accounting) {
1198 blkptr_t *bp = db->db_blkptr;
1199 int64_t willfree = (bp && !BP_IS_HOLE(bp)) ?
1200 bp_get_dsize(os->os_spa, bp) : db->db.db_size;
1201 /*
1202 * This is only a guess -- if the dbuf is dirty
1203 * in a previous txg, we don't know how much
1204 * space it will use on disk yet. We should
1205 * really have the struct_rwlock to access
1206 * db_blkptr, but since this is just a guess,
1207 * it's OK if we get an odd answer.
1208 */
1209 ddt_prefetch(os->os_spa, bp);
1210 dnode_willuse_space(dn, -willfree, tx);
1211 }
1212
1213 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1214 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1215 drop_struct_lock = TRUE;
1216 }
1217
1218 if (db->db_level == 0) {
1219 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1220 ASSERT(dn->dn_maxblkid >= db->db_blkid);
1221 }
1222
1223 if (db->db_level+1 < dn->dn_nlevels) {
1224 dmu_buf_impl_t *parent = db->db_parent;
1225 dbuf_dirty_record_t *di;
1226 int parent_held = FALSE;
1227
1228 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1229 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1230
1231 parent = dbuf_hold_level(dn, db->db_level+1,
1232 db->db_blkid >> epbs, FTAG);
1233 ASSERT(parent != NULL);
1234 parent_held = TRUE;
1235 }
1236 if (drop_struct_lock)
1237 rw_exit(&dn->dn_struct_rwlock);
1238 ASSERT3U(db->db_level+1, ==, parent->db_level);
1239 di = dbuf_dirty(parent, tx);
1240 if (parent_held)
1241 dbuf_rele(parent, FTAG);
1242
1243 mutex_enter(&db->db_mtx);
1244 /* possible race with dbuf_undirty() */
1245 if (db->db_last_dirty == dr ||
1246 dn->dn_object == DMU_META_DNODE_OBJECT) {
1247 mutex_enter(&di->dt.di.dr_mtx);
1248 ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1249 ASSERT(!list_link_active(&dr->dr_dirty_node));
1250 list_insert_tail(&di->dt.di.dr_children, dr);
1251 mutex_exit(&di->dt.di.dr_mtx);
1252 dr->dr_parent = di;
1253 }
1254 mutex_exit(&db->db_mtx);
1255 } else {
1256 ASSERT(db->db_level+1 == dn->dn_nlevels);
1257 ASSERT(db->db_blkid < dn->dn_nblkptr);
1258 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1259 mutex_enter(&dn->dn_mtx);
1260 ASSERT(!list_link_active(&dr->dr_dirty_node));
1261 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1262 mutex_exit(&dn->dn_mtx);
1263 if (drop_struct_lock)
1264 rw_exit(&dn->dn_struct_rwlock);
1265 }
1266
1267 dnode_setdirty(dn, tx);
1268 DB_DNODE_EXIT(db);
1269 return (dr);
1270 }
1271
1272 static int
1273 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1274 {
1275 dnode_t *dn;
1276 uint64_t txg = tx->tx_txg;
1277 dbuf_dirty_record_t *dr, **drp;
1278
1279 ASSERT(txg != 0);
1280 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1281
1282 mutex_enter(&db->db_mtx);
1283 /*
1284 * If this buffer is not dirty, we're done.
1285 */
1286 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1287 if (dr->dr_txg <= txg)
1288 break;
1289 if (dr == NULL || dr->dr_txg < txg) {
1290 mutex_exit(&db->db_mtx);
1291 return (0);
1292 }
1293 ASSERT(dr->dr_txg == txg);
1294 ASSERT(dr->dr_dbuf == db);
1295
1296 DB_DNODE_ENTER(db);
1297 dn = DB_DNODE(db);
1298
1299 /*
1300 * If this buffer is currently held, we cannot undirty
1301 * it, since one of the current holders may be in the
1302 * middle of an update. Note that users of dbuf_undirty()
1303 * should not place a hold on the dbuf before the call.
1304 * Also note: we can get here with a spill block, so
1305 * test for that similar to how dbuf_dirty does.
1306 */
1307 if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
1308 mutex_exit(&db->db_mtx);
1309 /* Make sure we don't toss this buffer at sync phase */
1310 if (db->db_blkid != DMU_SPILL_BLKID) {
1311 mutex_enter(&dn->dn_mtx);
1312 dnode_clear_range(dn, db->db_blkid, 1, tx);
1313 mutex_exit(&dn->dn_mtx);
1314 }
1315 DB_DNODE_EXIT(db);
1316 return (0);
1317 }
1318
1319 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1320
1321 ASSERT(db->db.db_size != 0);
1322
1323 /* XXX would be nice to fix up dn_towrite_space[] */
1324
1325 *drp = dr->dr_next;
1326
1327 /*
1328 * Note that there are three places in dbuf_dirty()
1329 * where this dirty record may be put on a list.
1330 * Make sure to do a list_remove corresponding to
1331 * every one of those list_insert calls.
1332 */
1333 if (dr->dr_parent) {
1334 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1335 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1336 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1337 } else if (db->db_blkid == DMU_SPILL_BLKID ||
1338 db->db_level+1 == dn->dn_nlevels) {
1339 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1340 mutex_enter(&dn->dn_mtx);
1341 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1342 mutex_exit(&dn->dn_mtx);
1343 }
1344 DB_DNODE_EXIT(db);
1345
1346 if (db->db_level == 0) {
1347 if (db->db_state != DB_NOFILL) {
1348 dbuf_unoverride(dr);
1349
1350 ASSERT(db->db_buf != NULL);
1351 ASSERT(dr->dt.dl.dr_data != NULL);
1352 if (dr->dt.dl.dr_data != db->db_buf)
1353 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data,
1354 db) == 1);
1355 }
1356 } else {
1357 ASSERT(db->db_buf != NULL);
1358 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
1359 mutex_destroy(&dr->dt.di.dr_mtx);
1360 list_destroy(&dr->dt.di.dr_children);
1361 }
1362 kmem_free(dr, sizeof (dbuf_dirty_record_t));
1363
1364 ASSERT(db->db_dirtycnt > 0);
1365 db->db_dirtycnt -= 1;
1366
1367 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1368 arc_buf_t *buf = db->db_buf;
1369
1370 ASSERT(db->db_state == DB_NOFILL || arc_released(buf));
1371 dbuf_set_data(db, NULL);
1372 VERIFY(arc_buf_remove_ref(buf, db) == 1);
1373 dbuf_evict(db);
1374 return (1);
1375 }
1376
1377 mutex_exit(&db->db_mtx);
1378 return (0);
1379 }
1380
1381 #pragma weak dmu_buf_will_dirty = dbuf_will_dirty
1382 void
1383 dbuf_will_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1384 {
1385 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1386
1387 ASSERT(tx->tx_txg != 0);
1388 ASSERT(!refcount_is_zero(&db->db_holds));
1389
1390 DB_DNODE_ENTER(db);
1391 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
1392 rf |= DB_RF_HAVESTRUCT;
1393 DB_DNODE_EXIT(db);
1394 (void) dbuf_read(db, NULL, rf);
1395 (void) dbuf_dirty(db, tx);
1396 }
1397
1398 void
1399 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1400 {
1401 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1402
1403 db->db_state = DB_NOFILL;
1404
1405 dmu_buf_will_fill(db_fake, tx);
1406 }
1407
1408 void
1409 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1410 {
1411 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1412
1413 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1414 ASSERT(tx->tx_txg != 0);
1415 ASSERT(db->db_level == 0);
1416 ASSERT(!refcount_is_zero(&db->db_holds));
1417
1418 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1419 dmu_tx_private_ok(tx));
1420
1421 dbuf_noread(db);
1422 (void) dbuf_dirty(db, tx);
1423 }
1424
1425 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1426 /* ARGSUSED */
1427 void
1428 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1429 {
1430 mutex_enter(&db->db_mtx);
1431 DBUF_VERIFY(db);
1432
1433 if (db->db_state == DB_FILL) {
1434 if (db->db_level == 0 && db->db_freed_in_flight) {
1435 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1436 /* we were freed while filling */
1437 /* XXX dbuf_undirty? */
1438 bzero(db->db.db_data, db->db.db_size);
1439 db->db_freed_in_flight = FALSE;
1440 }
1441 db->db_state = DB_CACHED;
1442 cv_broadcast(&db->db_changed);
1443 }
1444 mutex_exit(&db->db_mtx);
1445 }
1446
1447 /*
1448 * Directly assign a provided arc buf to a given dbuf if it's not referenced
1449 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1450 */
1451 void
1452 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
1453 {
1454 ASSERT(!refcount_is_zero(&db->db_holds));
1455 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1456 ASSERT(db->db_level == 0);
1457 ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA);
1458 ASSERT(buf != NULL);
1459 ASSERT(arc_buf_size(buf) == db->db.db_size);
1460 ASSERT(tx->tx_txg != 0);
1461
1462 arc_return_buf(buf, db);
1463 ASSERT(arc_released(buf));
1464
1465 mutex_enter(&db->db_mtx);
1466
1467 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1468 cv_wait(&db->db_changed, &db->db_mtx);
1469
1470 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
1471
1472 if (db->db_state == DB_CACHED &&
1473 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
1474 mutex_exit(&db->db_mtx);
1475 (void) dbuf_dirty(db, tx);
1476 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
1477 VERIFY(arc_buf_remove_ref(buf, db) == 1);
1478 xuio_stat_wbuf_copied();
1479 return;
1480 }
1481
1482 xuio_stat_wbuf_nocopy();
1483 if (db->db_state == DB_CACHED) {
1484 dbuf_dirty_record_t *dr = db->db_last_dirty;
1485
1486 ASSERT(db->db_buf != NULL);
1487 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
1488 ASSERT(dr->dt.dl.dr_data == db->db_buf);
1489 if (!arc_released(db->db_buf)) {
1490 ASSERT(dr->dt.dl.dr_override_state ==
1491 DR_OVERRIDDEN);
1492 arc_release(db->db_buf, db);
1493 }
1494 dr->dt.dl.dr_data = buf;
1495 VERIFY(arc_buf_remove_ref(db->db_buf, db) == 1);
1496 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
1497 arc_release(db->db_buf, db);
1498 VERIFY(arc_buf_remove_ref(db->db_buf, db) == 1);
1499 }
1500 db->db_buf = NULL;
1501 }
1502 ASSERT(db->db_buf == NULL);
1503 dbuf_set_data(db, buf);
1504 db->db_state = DB_FILL;
1505 mutex_exit(&db->db_mtx);
1506 (void) dbuf_dirty(db, tx);
1507 dbuf_fill_done(db, tx);
1508 }
1509
1510 /*
1511 * "Clear" the contents of this dbuf. This will mark the dbuf
1512 * EVICTING and clear *most* of its references. Unfortunetely,
1513 * when we are not holding the dn_dbufs_mtx, we can't clear the
1514 * entry in the dn_dbufs list. We have to wait until dbuf_destroy()
1515 * in this case. For callers from the DMU we will usually see:
1516 * dbuf_clear()->arc_buf_evict()->dbuf_do_evict()->dbuf_destroy()
1517 * For the arc callback, we will usually see:
1518 * dbuf_do_evict()->dbuf_clear();dbuf_destroy()
1519 * Sometimes, though, we will get a mix of these two:
1520 * DMU: dbuf_clear()->arc_buf_evict()
1521 * ARC: dbuf_do_evict()->dbuf_destroy()
1522 */
1523 void
1524 dbuf_clear(dmu_buf_impl_t *db)
1525 {
1526 dnode_t *dn;
1527 dmu_buf_impl_t *parent = db->db_parent;
1528 dmu_buf_impl_t *dndb;
1529 int dbuf_gone = FALSE;
1530
1531 ASSERT(MUTEX_HELD(&db->db_mtx));
1532 ASSERT(refcount_is_zero(&db->db_holds));
1533
1534 dbuf_evict_user(db);
1535
1536 if (db->db_state == DB_CACHED) {
1537 ASSERT(db->db.db_data != NULL);
1538 if (db->db_blkid == DMU_BONUS_BLKID) {
1539 zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
1540 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
1541 }
1542 db->db.db_data = NULL;
1543 db->db_state = DB_UNCACHED;
1544 }
1545
1546 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
1547 ASSERT(db->db_data_pending == NULL);
1548
1549 db->db_state = DB_EVICTING;
1550 db->db_blkptr = NULL;
1551
1552 DB_DNODE_ENTER(db);
1553 dn = DB_DNODE(db);
1554 dndb = dn->dn_dbuf;
1555 if (db->db_blkid != DMU_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) {
1556 list_remove(&dn->dn_dbufs, db);
1557 (void) atomic_dec_32_nv(&dn->dn_dbufs_count);
1558 membar_producer();
1559 DB_DNODE_EXIT(db);
1560 /*
1561 * Decrementing the dbuf count means that the hold corresponding
1562 * to the removed dbuf is no longer discounted in dnode_move(),
1563 * so the dnode cannot be moved until after we release the hold.
1564 * The membar_producer() ensures visibility of the decremented
1565 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
1566 * release any lock.
1567 */
1568 dnode_rele(dn, db);
1569 db->db_dnode_handle = NULL;
1570 } else {
1571 DB_DNODE_EXIT(db);
1572 }
1573
1574 if (db->db_buf)
1575 dbuf_gone = arc_buf_evict(db->db_buf);
1576
1577 if (!dbuf_gone)
1578 mutex_exit(&db->db_mtx);
1579
1580 /*
1581 * If this dbuf is referenced from an indirect dbuf,
1582 * decrement the ref count on the indirect dbuf.
1583 */
1584 if (parent && parent != dndb)
1585 dbuf_rele(parent, db);
1586 }
1587
1588 static int
1589 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
1590 dmu_buf_impl_t **parentp, blkptr_t **bpp)
1591 {
1592 int nlevels, epbs;
1593
1594 *parentp = NULL;
1595 *bpp = NULL;
1596
1597 ASSERT(blkid != DMU_BONUS_BLKID);
1598
1599 if (blkid == DMU_SPILL_BLKID) {
1600 mutex_enter(&dn->dn_mtx);
1601 if (dn->dn_have_spill &&
1602 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
1603 *bpp = &dn->dn_phys->dn_spill;
1604 else
1605 *bpp = NULL;
1606 dbuf_add_ref(dn->dn_dbuf, NULL);
1607 *parentp = dn->dn_dbuf;
1608 mutex_exit(&dn->dn_mtx);
1609 return (0);
1610 }
1611
1612 if (dn->dn_phys->dn_nlevels == 0)
1613 nlevels = 1;
1614 else
1615 nlevels = dn->dn_phys->dn_nlevels;
1616
1617 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1618
1619 ASSERT3U(level * epbs, <, 64);
1620 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1621 if (level >= nlevels ||
1622 (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
1623 /* the buffer has no parent yet */
1624 return (ENOENT);
1625 } else if (level < nlevels-1) {
1626 /* this block is referenced from an indirect block */
1627 int err = dbuf_hold_impl(dn, level+1,
1628 blkid >> epbs, fail_sparse, NULL, parentp);
1629 if (err)
1630 return (err);
1631 err = dbuf_read(*parentp, NULL,
1632 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
1633 if (err) {
1634 dbuf_rele(*parentp, NULL);
1635 *parentp = NULL;
1636 return (err);
1637 }
1638 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
1639 (blkid & ((1ULL << epbs) - 1));
1640 return (0);
1641 } else {
1642 /* the block is referenced from the dnode */
1643 ASSERT3U(level, ==, nlevels-1);
1644 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
1645 blkid < dn->dn_phys->dn_nblkptr);
1646 if (dn->dn_dbuf) {
1647 dbuf_add_ref(dn->dn_dbuf, NULL);
1648 *parentp = dn->dn_dbuf;
1649 }
1650 *bpp = &dn->dn_phys->dn_blkptr[blkid];
1651 return (0);
1652 }
1653 }
1654
1655 static dmu_buf_impl_t *
1656 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
1657 dmu_buf_impl_t *parent, blkptr_t *blkptr)
1658 {
1659 objset_t *os = dn->dn_objset;
1660 dmu_buf_impl_t *db, *odb;
1661
1662 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1663 ASSERT(dn->dn_type != DMU_OT_NONE);
1664
1665 db = kmem_cache_alloc(dbuf_cache, KM_SLEEP);
1666
1667 db->db_objset = os;
1668 db->db.db_object = dn->dn_object;
1669 db->db_level = level;
1670 db->db_blkid = blkid;
1671 db->db_last_dirty = NULL;
1672 db->db_dirtycnt = 0;
1673 db->db_dnode_handle = dn->dn_handle;
1674 db->db_parent = parent;
1675 db->db_blkptr = blkptr;
1676
1677 db->db_user_ptr = NULL;
1678 db->db_user_data_ptr_ptr = NULL;
1679 db->db_evict_func = NULL;
1680 db->db_immediate_evict = 0;
1681 db->db_freed_in_flight = 0;
1682
1683 if (blkid == DMU_BONUS_BLKID) {
1684 ASSERT3P(parent, ==, dn->dn_dbuf);
1685 db->db.db_size = DN_MAX_BONUSLEN -
1686 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
1687 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
1688 db->db.db_offset = DMU_BONUS_BLKID;
1689 db->db_state = DB_UNCACHED;
1690 /* the bonus dbuf is not placed in the hash table */
1691 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1692 return (db);
1693 } else if (blkid == DMU_SPILL_BLKID) {
1694 db->db.db_size = (blkptr != NULL) ?
1695 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
1696 db->db.db_offset = 0;
1697 } else {
1698 int blocksize =
1699 db->db_level ? 1<<dn->dn_indblkshift : dn->dn_datablksz;
1700 db->db.db_size = blocksize;
1701 db->db.db_offset = db->db_blkid * blocksize;
1702 }
1703
1704 /*
1705 * Hold the dn_dbufs_mtx while we get the new dbuf
1706 * in the hash table *and* added to the dbufs list.
1707 * This prevents a possible deadlock with someone
1708 * trying to look up this dbuf before its added to the
1709 * dn_dbufs list.
1710 */
1711 mutex_enter(&dn->dn_dbufs_mtx);
1712 db->db_state = DB_EVICTING;
1713 if ((odb = dbuf_hash_insert(db)) != NULL) {
1714 /* someone else inserted it first */
1715 kmem_cache_free(dbuf_cache, db);
1716 mutex_exit(&dn->dn_dbufs_mtx);
1717 return (odb);
1718 }
1719 list_insert_head(&dn->dn_dbufs, db);
1720 db->db_state = DB_UNCACHED;
1721 mutex_exit(&dn->dn_dbufs_mtx);
1722 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1723
1724 if (parent && parent != dn->dn_dbuf)
1725 dbuf_add_ref(parent, db);
1726
1727 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1728 refcount_count(&dn->dn_holds) > 0);
1729 (void) refcount_add(&dn->dn_holds, db);
1730 (void) atomic_inc_32_nv(&dn->dn_dbufs_count);
1731
1732 dprintf_dbuf(db, "db=%p\n", db);
1733
1734 return (db);
1735 }
1736
1737 static int
1738 dbuf_do_evict(void *private)
1739 {
1740 arc_buf_t *buf = private;
1741 dmu_buf_impl_t *db = buf->b_private;
1742
1743 if (!MUTEX_HELD(&db->db_mtx))
1744 mutex_enter(&db->db_mtx);
1745
1746 ASSERT(refcount_is_zero(&db->db_holds));
1747
1748 if (db->db_state != DB_EVICTING) {
1749 ASSERT(db->db_state == DB_CACHED);
1750 DBUF_VERIFY(db);
1751 db->db_buf = NULL;
1752 dbuf_evict(db);
1753 } else {
1754 mutex_exit(&db->db_mtx);
1755 dbuf_destroy(db);
1756 }
1757 return (0);
1758 }
1759
1760 static void
1761 dbuf_destroy(dmu_buf_impl_t *db)
1762 {
1763 ASSERT(refcount_is_zero(&db->db_holds));
1764
1765 if (db->db_blkid != DMU_BONUS_BLKID) {
1766 /*
1767 * If this dbuf is still on the dn_dbufs list,
1768 * remove it from that list.
1769 */
1770 if (db->db_dnode_handle != NULL) {
1771 dnode_t *dn;
1772
1773 DB_DNODE_ENTER(db);
1774 dn = DB_DNODE(db);
1775 mutex_enter(&dn->dn_dbufs_mtx);
1776 list_remove(&dn->dn_dbufs, db);
1777 (void) atomic_dec_32_nv(&dn->dn_dbufs_count);
1778 mutex_exit(&dn->dn_dbufs_mtx);
1779 DB_DNODE_EXIT(db);
1780 /*
1781 * Decrementing the dbuf count means that the hold
1782 * corresponding to the removed dbuf is no longer
1783 * discounted in dnode_move(), so the dnode cannot be
1784 * moved until after we release the hold.
1785 */
1786 dnode_rele(dn, db);
1787 db->db_dnode_handle = NULL;
1788 }
1789 dbuf_hash_remove(db);
1790 }
1791 db->db_parent = NULL;
1792 db->db_buf = NULL;
1793
1794 ASSERT(!list_link_active(&db->db_link));
1795 ASSERT(db->db.db_data == NULL);
1796 ASSERT(db->db_hash_next == NULL);
1797 ASSERT(db->db_blkptr == NULL);
1798 ASSERT(db->db_data_pending == NULL);
1799
1800 kmem_cache_free(dbuf_cache, db);
1801 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1802 }
1803
1804 void
1805 dbuf_prefetch(dnode_t *dn, uint64_t blkid)
1806 {
1807 dmu_buf_impl_t *db = NULL;
1808 blkptr_t *bp = NULL;
1809
1810 ASSERT(blkid != DMU_BONUS_BLKID);
1811 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1812
1813 if (dnode_block_freed(dn, blkid))
1814 return;
1815
1816 /* dbuf_find() returns with db_mtx held */
1817 if (db = dbuf_find(dn, 0, blkid)) {
1818 /*
1819 * This dbuf is already in the cache. We assume that
1820 * it is already CACHED, or else about to be either
1821 * read or filled.
1822 */
1823 mutex_exit(&db->db_mtx);
1824 return;
1825 }
1826
1827 if (dbuf_findbp(dn, 0, blkid, TRUE, &db, &bp) == 0) {
1828 if (bp && !BP_IS_HOLE(bp)) {
1829 int priority = dn->dn_type == DMU_OT_DDT_ZAP ?
1830 ZIO_PRIORITY_DDT_PREFETCH : ZIO_PRIORITY_ASYNC_READ;
1831 arc_buf_t *pbuf;
1832 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
1833 uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH;
1834 zbookmark_t zb;
1835
1836 SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
1837 dn->dn_object, 0, blkid);
1838
1839 if (db)
1840 pbuf = db->db_buf;
1841 else
1842 pbuf = dn->dn_objset->os_phys_buf;
1843
1844 (void) dsl_read(NULL, dn->dn_objset->os_spa,
1845 bp, pbuf, NULL, NULL, priority,
1846 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1847 &aflags, &zb);
1848 }
1849 if (db)
1850 dbuf_rele(db, NULL);
1851 }
1852 }
1853
1854 /*
1855 * Returns with db_holds incremented, and db_mtx not held.
1856 * Note: dn_struct_rwlock must be held.
1857 */
1858 int
1859 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse,
1860 void *tag, dmu_buf_impl_t **dbp)
1861 {
1862 dmu_buf_impl_t *db, *parent = NULL;
1863
1864 ASSERT(blkid != DMU_BONUS_BLKID);
1865 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1866 ASSERT3U(dn->dn_nlevels, >, level);
1867
1868 *dbp = NULL;
1869 top:
1870 /* dbuf_find() returns with db_mtx held */
1871 db = dbuf_find(dn, level, blkid);
1872
1873 if (db == NULL) {
1874 blkptr_t *bp = NULL;
1875 int err;
1876
1877 ASSERT3P(parent, ==, NULL);
1878 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
1879 if (fail_sparse) {
1880 if (err == 0 && bp && BP_IS_HOLE(bp))
1881 err = ENOENT;
1882 if (err) {
1883 if (parent)
1884 dbuf_rele(parent, NULL);
1885 return (err);
1886 }
1887 }
1888 if (err && err != ENOENT)
1889 return (err);
1890 db = dbuf_create(dn, level, blkid, parent, bp);
1891 }
1892
1893 if (db->db_buf && refcount_is_zero(&db->db_holds)) {
1894 arc_buf_add_ref(db->db_buf, db);
1895 if (db->db_buf->b_data == NULL) {
1896 dbuf_clear(db);
1897 if (parent) {
1898 dbuf_rele(parent, NULL);
1899 parent = NULL;
1900 }
1901 goto top;
1902 }
1903 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
1904 }
1905
1906 ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
1907
1908 /*
1909 * If this buffer is currently syncing out, and we are are
1910 * still referencing it from db_data, we need to make a copy
1911 * of it in case we decide we want to dirty it again in this txg.
1912 */
1913 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1914 dn->dn_object != DMU_META_DNODE_OBJECT &&
1915 db->db_state == DB_CACHED && db->db_data_pending) {
1916 dbuf_dirty_record_t *dr = db->db_data_pending;
1917
1918 if (dr->dt.dl.dr_data == db->db_buf) {
1919 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1920
1921 dbuf_set_data(db,
1922 arc_buf_alloc(dn->dn_objset->os_spa,
1923 db->db.db_size, db, type));
1924 bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
1925 db->db.db_size);
1926 }
1927 }
1928
1929 (void) refcount_add(&db->db_holds, tag);
1930 dbuf_update_data(db);
1931 DBUF_VERIFY(db);
1932 mutex_exit(&db->db_mtx);
1933
1934 /* NOTE: we can't rele the parent until after we drop the db_mtx */
1935 if (parent)
1936 dbuf_rele(parent, NULL);
1937
1938 ASSERT3P(DB_DNODE(db), ==, dn);
1939 ASSERT3U(db->db_blkid, ==, blkid);
1940 ASSERT3U(db->db_level, ==, level);
1941 *dbp = db;
1942
1943 return (0);
1944 }
1945
1946 dmu_buf_impl_t *
1947 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
1948 {
1949 dmu_buf_impl_t *db;
1950 int err = dbuf_hold_impl(dn, 0, blkid, FALSE, tag, &db);
1951 return (err ? NULL : db);
1952 }
1953
1954 dmu_buf_impl_t *
1955 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
1956 {
1957 dmu_buf_impl_t *db;
1958 int err = dbuf_hold_impl(dn, level, blkid, FALSE, tag, &db);
1959 return (err ? NULL : db);
1960 }
1961
1962 void
1963 dbuf_create_bonus(dnode_t *dn)
1964 {
1965 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1966
1967 ASSERT(dn->dn_bonus == NULL);
1968 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
1969 }
1970
1971 int
1972 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
1973 {
1974 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1975 dnode_t *dn;
1976
1977 if (db->db_blkid != DMU_SPILL_BLKID)
1978 return (ENOTSUP);
1979 if (blksz == 0)
1980 blksz = SPA_MINBLOCKSIZE;
1981 if (blksz > SPA_MAXBLOCKSIZE)
1982 blksz = SPA_MAXBLOCKSIZE;
1983 else
1984 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
1985
1986 DB_DNODE_ENTER(db);
1987 dn = DB_DNODE(db);
1988 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1989 dbuf_new_size(db, blksz, tx);
1990 rw_exit(&dn->dn_struct_rwlock);
1991 DB_DNODE_EXIT(db);
1992
1993 return (0);
1994 }
1995
1996 void
1997 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
1998 {
1999 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
2000 }
2001
2002 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2003 void
2004 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
2005 {
2006 int64_t holds = refcount_add(&db->db_holds, tag);
2007 ASSERT(holds > 1);
2008 }
2009
2010 /*
2011 * If you call dbuf_rele() you had better not be referencing the dnode handle
2012 * unless you have some other direct or indirect hold on the dnode. (An indirect
2013 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2014 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2015 * dnode's parent dbuf evicting its dnode handles.
2016 */
2017 #pragma weak dmu_buf_rele = dbuf_rele
2018 void
2019 dbuf_rele(dmu_buf_impl_t *db, void *tag)
2020 {
2021 mutex_enter(&db->db_mtx);
2022 dbuf_rele_and_unlock(db, tag);
2023 }
2024
2025 /*
2026 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2027 * db_dirtycnt and db_holds to be updated atomically.
2028 */
2029 void
2030 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2031 {
2032 int64_t holds;
2033
2034 ASSERT(MUTEX_HELD(&db->db_mtx));
2035 DBUF_VERIFY(db);
2036
2037 /*
2038 * Remove the reference to the dbuf before removing its hold on the
2039 * dnode so we can guarantee in dnode_move() that a referenced bonus
2040 * buffer has a corresponding dnode hold.
2041 */
2042 holds = refcount_remove(&db->db_holds, tag);
2043 ASSERT(holds >= 0);
2044
2045 /*
2046 * We can't freeze indirects if there is a possibility that they
2047 * may be modified in the current syncing context.
2048 */
2049 if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0))
2050 arc_buf_freeze(db->db_buf);
2051
2052 if (holds == db->db_dirtycnt &&
2053 db->db_level == 0 && db->db_immediate_evict)
2054 dbuf_evict_user(db);
2055
2056 if (holds == 0) {
2057 if (db->db_blkid == DMU_BONUS_BLKID) {
2058 mutex_exit(&db->db_mtx);
2059
2060 /*
2061 * If the dnode moves here, we cannot cross this barrier
2062 * until the move completes.
2063 */
2064 DB_DNODE_ENTER(db);
2065 (void) atomic_dec_32_nv(&DB_DNODE(db)->dn_dbufs_count);
2066 DB_DNODE_EXIT(db);
2067 /*
2068 * The bonus buffer's dnode hold is no longer discounted
2069 * in dnode_move(). The dnode cannot move until after
2070 * the dnode_rele().
2071 */
2072 dnode_rele(DB_DNODE(db), db);
2073 } else if (db->db_buf == NULL) {
2074 /*
2075 * This is a special case: we never associated this
2076 * dbuf with any data allocated from the ARC.
2077 */
2078 ASSERT(db->db_state == DB_UNCACHED ||
2079 db->db_state == DB_NOFILL);
2080 dbuf_evict(db);
2081 } else if (arc_released(db->db_buf)) {
2082 arc_buf_t *buf = db->db_buf;
2083 /*
2084 * This dbuf has anonymous data associated with it.
2085 */
2086 dbuf_set_data(db, NULL);
2087 VERIFY(arc_buf_remove_ref(buf, db) == 1);
2088 dbuf_evict(db);
2089 } else {
2090 VERIFY(arc_buf_remove_ref(db->db_buf, db) == 0);
2091 if (!DBUF_IS_CACHEABLE(db))
2092 dbuf_clear(db);
2093 else
2094 mutex_exit(&db->db_mtx);
2095 }
2096 } else {
2097 mutex_exit(&db->db_mtx);
2098 }
2099 }
2100
2101 #pragma weak dmu_buf_refcount = dbuf_refcount
2102 uint64_t
2103 dbuf_refcount(dmu_buf_impl_t *db)
2104 {
2105 return (refcount_count(&db->db_holds));
2106 }
2107
2108 void *
2109 dmu_buf_set_user(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr,
2110 dmu_buf_evict_func_t *evict_func)
2111 {
2112 return (dmu_buf_update_user(db_fake, NULL, user_ptr,
2113 user_data_ptr_ptr, evict_func));
2114 }
2115
2116 void *
2117 dmu_buf_set_user_ie(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr,
2118 dmu_buf_evict_func_t *evict_func)
2119 {
2120 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2121
2122 db->db_immediate_evict = TRUE;
2123 return (dmu_buf_update_user(db_fake, NULL, user_ptr,
2124 user_data_ptr_ptr, evict_func));
2125 }
2126
2127 void *
2128 dmu_buf_update_user(dmu_buf_t *db_fake, void *old_user_ptr, void *user_ptr,
2129 void *user_data_ptr_ptr, dmu_buf_evict_func_t *evict_func)
2130 {
2131 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2132 ASSERT(db->db_level == 0);
2133
2134 ASSERT((user_ptr == NULL) == (evict_func == NULL));
2135
2136 mutex_enter(&db->db_mtx);
2137
2138 if (db->db_user_ptr == old_user_ptr) {
2139 db->db_user_ptr = user_ptr;
2140 db->db_user_data_ptr_ptr = user_data_ptr_ptr;
2141 db->db_evict_func = evict_func;
2142
2143 dbuf_update_data(db);
2144 } else {
2145 old_user_ptr = db->db_user_ptr;
2146 }
2147
2148 mutex_exit(&db->db_mtx);
2149 return (old_user_ptr);
2150 }
2151
2152 void *
2153 dmu_buf_get_user(dmu_buf_t *db_fake)
2154 {
2155 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2156 ASSERT(!refcount_is_zero(&db->db_holds));
2157
2158 return (db->db_user_ptr);
2159 }
2160
2161 boolean_t
2162 dmu_buf_freeable(dmu_buf_t *dbuf)
2163 {
2164 boolean_t res = B_FALSE;
2165 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2166
2167 if (db->db_blkptr)
2168 res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset,
2169 db->db_blkptr, db->db_blkptr->blk_birth);
2170
2171 return (res);
2172 }
2173
2174 static void
2175 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
2176 {
2177 /* ASSERT(dmu_tx_is_syncing(tx) */
2178 ASSERT(MUTEX_HELD(&db->db_mtx));
2179
2180 if (db->db_blkptr != NULL)
2181 return;
2182
2183 if (db->db_blkid == DMU_SPILL_BLKID) {
2184 db->db_blkptr = &dn->dn_phys->dn_spill;
2185 BP_ZERO(db->db_blkptr);
2186 return;
2187 }
2188 if (db->db_level == dn->dn_phys->dn_nlevels-1) {
2189 /*
2190 * This buffer was allocated at a time when there was
2191 * no available blkptrs from the dnode, or it was
2192 * inappropriate to hook it in (i.e., nlevels mis-match).
2193 */
2194 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
2195 ASSERT(db->db_parent == NULL);
2196 db->db_parent = dn->dn_dbuf;
2197 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
2198 DBUF_VERIFY(db);
2199 } else {
2200 dmu_buf_impl_t *parent = db->db_parent;
2201 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2202
2203 ASSERT(dn->dn_phys->dn_nlevels > 1);
2204 if (parent == NULL) {
2205 mutex_exit(&db->db_mtx);
2206 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2207 (void) dbuf_hold_impl(dn, db->db_level+1,
2208 db->db_blkid >> epbs, FALSE, db, &parent);
2209 rw_exit(&dn->dn_struct_rwlock);
2210 mutex_enter(&db->db_mtx);
2211 db->db_parent = parent;
2212 }
2213 db->db_blkptr = (blkptr_t *)parent->db.db_data +
2214 (db->db_blkid & ((1ULL << epbs) - 1));
2215 DBUF_VERIFY(db);
2216 }
2217 }
2218
2219 static void
2220 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2221 {
2222 dmu_buf_impl_t *db = dr->dr_dbuf;
2223 dnode_t *dn;
2224 zio_t *zio;
2225
2226 ASSERT(dmu_tx_is_syncing(tx));
2227
2228 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2229
2230 mutex_enter(&db->db_mtx);
2231
2232 ASSERT(db->db_level > 0);
2233 DBUF_VERIFY(db);
2234
2235 if (db->db_buf == NULL) {
2236 mutex_exit(&db->db_mtx);
2237 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
2238 mutex_enter(&db->db_mtx);
2239 }
2240 ASSERT3U(db->db_state, ==, DB_CACHED);
2241 ASSERT(db->db_buf != NULL);
2242
2243 DB_DNODE_ENTER(db);
2244 dn = DB_DNODE(db);
2245 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2246 dbuf_check_blkptr(dn, db);
2247 DB_DNODE_EXIT(db);
2248
2249 db->db_data_pending = dr;
2250
2251 mutex_exit(&db->db_mtx);
2252 dbuf_write(dr, db->db_buf, tx);
2253
2254 zio = dr->dr_zio;
2255 mutex_enter(&dr->dt.di.dr_mtx);
2256 dbuf_sync_list(&dr->dt.di.dr_children, tx);
2257 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2258 mutex_exit(&dr->dt.di.dr_mtx);
2259 zio_nowait(zio);
2260 }
2261
2262 static void
2263 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2264 {
2265 arc_buf_t **datap = &dr->dt.dl.dr_data;
2266 dmu_buf_impl_t *db = dr->dr_dbuf;
2267 dnode_t *dn;
2268 objset_t *os;
2269 uint64_t txg = tx->tx_txg;
2270
2271 ASSERT(dmu_tx_is_syncing(tx));
2272
2273 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2274
2275 mutex_enter(&db->db_mtx);
2276 /*
2277 * To be synced, we must be dirtied. But we
2278 * might have been freed after the dirty.
2279 */
2280 if (db->db_state == DB_UNCACHED) {
2281 /* This buffer has been freed since it was dirtied */
2282 ASSERT(db->db.db_data == NULL);
2283 } else if (db->db_state == DB_FILL) {
2284 /* This buffer was freed and is now being re-filled */
2285 ASSERT(db->db.db_data != dr->dt.dl.dr_data);
2286 } else {
2287 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
2288 }
2289 DBUF_VERIFY(db);
2290
2291 DB_DNODE_ENTER(db);
2292 dn = DB_DNODE(db);
2293
2294 if (db->db_blkid == DMU_SPILL_BLKID) {
2295 mutex_enter(&dn->dn_mtx);
2296 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
2297 mutex_exit(&dn->dn_mtx);
2298 }
2299
2300 /*
2301 * If this is a bonus buffer, simply copy the bonus data into the
2302 * dnode. It will be written out when the dnode is synced (and it
2303 * will be synced, since it must have been dirty for dbuf_sync to
2304 * be called).
2305 */
2306 if (db->db_blkid == DMU_BONUS_BLKID) {
2307 dbuf_dirty_record_t **drp;
2308
2309 ASSERT(*datap != NULL);
2310 ASSERT3U(db->db_level, ==, 0);
2311 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN);
2312 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen);
2313 DB_DNODE_EXIT(db);
2314
2315 if (*datap != db->db.db_data) {
2316 zio_buf_free(*datap, DN_MAX_BONUSLEN);
2317 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
2318 }
2319 db->db_data_pending = NULL;
2320 drp = &db->db_last_dirty;
2321 while (*drp != dr)
2322 drp = &(*drp)->dr_next;
2323 ASSERT(dr->dr_next == NULL);
2324 ASSERT(dr->dr_dbuf == db);
2325 *drp = dr->dr_next;
2326 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2327 ASSERT(db->db_dirtycnt > 0);
2328 db->db_dirtycnt -= 1;
2329 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
2330 return;
2331 }
2332
2333 os = dn->dn_objset;
2334
2335 /*
2336 * This function may have dropped the db_mtx lock allowing a dmu_sync
2337 * operation to sneak in. As a result, we need to ensure that we
2338 * don't check the dr_override_state until we have returned from
2339 * dbuf_check_blkptr.
2340 */
2341 dbuf_check_blkptr(dn, db);
2342
2343 /*
2344 * If this buffer is in the middle of an immediate write,
2345 * wait for the synchronous IO to complete.
2346 */
2347 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
2348 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
2349 cv_wait(&db->db_changed, &db->db_mtx);
2350 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
2351 }
2352
2353 if (db->db_state != DB_NOFILL &&
2354 dn->dn_object != DMU_META_DNODE_OBJECT &&
2355 refcount_count(&db->db_holds) > 1 &&
2356 dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
2357 *datap == db->db_buf) {
2358 /*
2359 * If this buffer is currently "in use" (i.e., there
2360 * are active holds and db_data still references it),
2361 * then make a copy before we start the write so that
2362 * any modifications from the open txg will not leak
2363 * into this write.
2364 *
2365 * NOTE: this copy does not need to be made for
2366 * objects only modified in the syncing context (e.g.
2367 * DNONE_DNODE blocks).
2368 */
2369 int blksz = arc_buf_size(*datap);
2370 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2371 *datap = arc_buf_alloc(os->os_spa, blksz, db, type);
2372 bcopy(db->db.db_data, (*datap)->b_data, blksz);
2373 }
2374 db->db_data_pending = dr;
2375
2376 mutex_exit(&db->db_mtx);
2377
2378 dbuf_write(dr, *datap, tx);
2379
2380 ASSERT(!list_link_active(&dr->dr_dirty_node));
2381 if (dn->dn_object == DMU_META_DNODE_OBJECT) {
2382 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
2383 DB_DNODE_EXIT(db);
2384 } else {
2385 /*
2386 * Although zio_nowait() does not "wait for an IO", it does
2387 * initiate the IO. If this is an empty write it seems plausible
2388 * that the IO could actually be completed before the nowait
2389 * returns. We need to DB_DNODE_EXIT() first in case
2390 * zio_nowait() invalidates the dbuf.
2391 */
2392 DB_DNODE_EXIT(db);
2393 zio_nowait(dr->dr_zio);
2394 }
2395 }
2396
2397 void
2398 dbuf_sync_list(list_t *list, dmu_tx_t *tx)
2399 {
2400 dbuf_dirty_record_t *dr;
2401
2402 while (dr = list_head(list)) {
2403 if (dr->dr_zio != NULL) {
2404 /*
2405 * If we find an already initialized zio then we
2406 * are processing the meta-dnode, and we have finished.
2407 * The dbufs for all dnodes are put back on the list
2408 * during processing, so that we can zio_wait()
2409 * these IOs after initiating all child IOs.
2410 */
2411 ASSERT3U(dr->dr_dbuf->db.db_object, ==,
2412 DMU_META_DNODE_OBJECT);
2413 break;
2414 }
2415 list_remove(list, dr);
2416 if (dr->dr_dbuf->db_level > 0)
2417 dbuf_sync_indirect(dr, tx);
2418 else
2419 dbuf_sync_leaf(dr, tx);
2420 }
2421 }
2422
2423 /* ARGSUSED */
2424 static void
2425 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
2426 {
2427 dmu_buf_impl_t *db = vdb;
2428 dnode_t *dn;
2429 blkptr_t *bp = zio->io_bp;
2430 blkptr_t *bp_orig = &zio->io_bp_orig;
2431 spa_t *spa = zio->io_spa;
2432 int64_t delta;
2433 uint64_t fill = 0;
2434 int i;
2435
2436 ASSERT(db->db_blkptr == bp);
2437
2438 DB_DNODE_ENTER(db);
2439 dn = DB_DNODE(db);
2440 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
2441 dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
2442 zio->io_prev_space_delta = delta;
2443
2444 if (BP_IS_HOLE(bp)) {
2445 ASSERT(bp->blk_fill == 0);
2446 DB_DNODE_EXIT(db);
2447 return;
2448 }
2449
2450 ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
2451 BP_GET_TYPE(bp) == dn->dn_type) ||
2452 (db->db_blkid == DMU_SPILL_BLKID &&
2453 BP_GET_TYPE(bp) == dn->dn_bonustype));
2454 ASSERT(BP_GET_LEVEL(bp) == db->db_level);
2455
2456 mutex_enter(&db->db_mtx);
2457
2458 #ifdef ZFS_DEBUG
2459 if (db->db_blkid == DMU_SPILL_BLKID) {
2460 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2461 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2462 db->db_blkptr == &dn->dn_phys->dn_spill);
2463 }
2464 #endif
2465
2466 if (db->db_level == 0) {
2467 mutex_enter(&dn->dn_mtx);
2468 if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
2469 db->db_blkid != DMU_SPILL_BLKID)
2470 dn->dn_phys->dn_maxblkid = db->db_blkid;
2471 mutex_exit(&dn->dn_mtx);
2472
2473 if (dn->dn_type == DMU_OT_DNODE) {
2474 dnode_phys_t *dnp = db->db.db_data;
2475 for (i = db->db.db_size >> DNODE_SHIFT; i > 0;
2476 i--, dnp++) {
2477 if (dnp->dn_type != DMU_OT_NONE)
2478 fill++;
2479 }
2480 } else {
2481 fill = 1;
2482 }
2483 } else {
2484 blkptr_t *ibp = db->db.db_data;
2485 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2486 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
2487 if (BP_IS_HOLE(ibp))
2488 continue;
2489 fill += ibp->blk_fill;
2490 }
2491 }
2492 DB_DNODE_EXIT(db);
2493
2494 bp->blk_fill = fill;
2495
2496 mutex_exit(&db->db_mtx);
2497 }
2498
2499 /* ARGSUSED */
2500 static void
2501 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
2502 {
2503 dmu_buf_impl_t *db = vdb;
2504 blkptr_t *bp = zio->io_bp;
2505 blkptr_t *bp_orig = &zio->io_bp_orig;
2506 uint64_t txg = zio->io_txg;
2507 dbuf_dirty_record_t **drp, *dr;
2508
2509 ASSERT3U(zio->io_error, ==, 0);
2510 ASSERT(db->db_blkptr == bp);
2511
2512 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
2513 ASSERT(BP_EQUAL(bp, bp_orig));
2514 } else {
2515 objset_t *os;
2516 dsl_dataset_t *ds;
2517 dmu_tx_t *tx;
2518
2519 DB_GET_OBJSET(&os, db);
2520 ds = os->os_dsl_dataset;
2521 tx = os->os_synctx;
2522
2523 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
2524 dsl_dataset_block_born(ds, bp, tx);
2525 }
2526
2527 mutex_enter(&db->db_mtx);
2528
2529 DBUF_VERIFY(db);
2530
2531 drp = &db->db_last_dirty;
2532 while ((dr = *drp) != db->db_data_pending)
2533 drp = &dr->dr_next;
2534 ASSERT(!list_link_active(&dr->dr_dirty_node));
2535 ASSERT(dr->dr_txg == txg);
2536 ASSERT(dr->dr_dbuf == db);
2537 ASSERT(dr->dr_next == NULL);
2538 *drp = dr->dr_next;
2539
2540 #ifdef ZFS_DEBUG
2541 if (db->db_blkid == DMU_SPILL_BLKID) {
2542 dnode_t *dn;
2543
2544 DB_DNODE_ENTER(db);
2545 dn = DB_DNODE(db);
2546 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2547 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2548 db->db_blkptr == &dn->dn_phys->dn_spill);
2549 DB_DNODE_EXIT(db);
2550 }
2551 #endif
2552
2553 if (db->db_level == 0) {
2554 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2555 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
2556 if (db->db_state != DB_NOFILL) {
2557 if (dr->dt.dl.dr_data != db->db_buf)
2558 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data,
2559 db) == 1);
2560 else if (!arc_released(db->db_buf))
2561 arc_set_callback(db->db_buf, dbuf_do_evict, db);
2562 }
2563 } else {
2564 dnode_t *dn;
2565
2566 DB_DNODE_ENTER(db);
2567 dn = DB_DNODE(db);
2568 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2569 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2570 if (!BP_IS_HOLE(db->db_blkptr)) {
2571 int epbs =
2572 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2573 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
2574 db->db.db_size);
2575 ASSERT3U(dn->dn_phys->dn_maxblkid
2576 >> (db->db_level * epbs), >=, db->db_blkid);
2577 arc_set_callback(db->db_buf, dbuf_do_evict, db);
2578 }
2579 DB_DNODE_EXIT(db);
2580 mutex_destroy(&dr->dt.di.dr_mtx);
2581 list_destroy(&dr->dt.di.dr_children);
2582 }
2583 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2584
2585 cv_broadcast(&db->db_changed);
2586 ASSERT(db->db_dirtycnt > 0);
2587 db->db_dirtycnt -= 1;
2588 db->db_data_pending = NULL;
2589 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
2590 }
2591
2592 static void
2593 dbuf_write_nofill_ready(zio_t *zio)
2594 {
2595 dbuf_write_ready(zio, NULL, zio->io_private);
2596 }
2597
2598 static void
2599 dbuf_write_nofill_done(zio_t *zio)
2600 {
2601 dbuf_write_done(zio, NULL, zio->io_private);
2602 }
2603
2604 static void
2605 dbuf_write_override_ready(zio_t *zio)
2606 {
2607 dbuf_dirty_record_t *dr = zio->io_private;
2608 dmu_buf_impl_t *db = dr->dr_dbuf;
2609
2610 dbuf_write_ready(zio, NULL, db);
2611 }
2612
2613 static void
2614 dbuf_write_override_done(zio_t *zio)
2615 {
2616 dbuf_dirty_record_t *dr = zio->io_private;
2617 dmu_buf_impl_t *db = dr->dr_dbuf;
2618 blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
2619
2620 mutex_enter(&db->db_mtx);
2621 if (!BP_EQUAL(zio->io_bp, obp)) {
2622 if (!BP_IS_HOLE(obp))
2623 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
2624 arc_release(dr->dt.dl.dr_data, db);
2625 }
2626 mutex_exit(&db->db_mtx);
2627
2628 dbuf_write_done(zio, NULL, db);
2629 }
2630
2631 static void
2632 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
2633 {
2634 dmu_buf_impl_t *db = dr->dr_dbuf;
2635 dnode_t *dn;
2636 objset_t *os;
2637 dmu_buf_impl_t *parent = db->db_parent;
2638 uint64_t txg = tx->tx_txg;
2639 zbookmark_t zb;
2640 zio_prop_t zp;
2641 zio_t *zio;
2642 int wp_flag = 0;
2643
2644 DB_DNODE_ENTER(db);
2645 dn = DB_DNODE(db);
2646 os = dn->dn_objset;
2647
2648 if (db->db_state != DB_NOFILL) {
2649 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
2650 /*
2651 * Private object buffers are released here rather
2652 * than in dbuf_dirty() since they are only modified
2653 * in the syncing context and we don't want the
2654 * overhead of making multiple copies of the data.
2655 */
2656 if (BP_IS_HOLE(db->db_blkptr)) {
2657 arc_buf_thaw(data);
2658 } else {
2659 dbuf_release_bp(db);
2660 }
2661 }
2662 }
2663
2664 if (parent != dn->dn_dbuf) {
2665 ASSERT(parent && parent->db_data_pending);
2666 ASSERT(db->db_level == parent->db_level-1);
2667 ASSERT(arc_released(parent->db_buf));
2668 zio = parent->db_data_pending->dr_zio;
2669 } else {
2670 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
2671 db->db_blkid != DMU_SPILL_BLKID) ||
2672 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
2673 if (db->db_blkid != DMU_SPILL_BLKID)
2674 ASSERT3P(db->db_blkptr, ==,
2675 &dn->dn_phys->dn_blkptr[db->db_blkid]);
2676 zio = dn->dn_zio;
2677 }
2678
2679 ASSERT(db->db_level == 0 || data == db->db_buf);
2680 ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
2681 ASSERT(zio);
2682
2683 SET_BOOKMARK(&zb, os->os_dsl_dataset ?
2684 os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
2685 db->db.db_object, db->db_level, db->db_blkid);
2686
2687 if (db->db_blkid == DMU_SPILL_BLKID)
2688 wp_flag = WP_SPILL;
2689 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
2690
2691 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
2692 DB_DNODE_EXIT(db);
2693
2694 if (db->db_level == 0 && dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
2695 ASSERT(db->db_state != DB_NOFILL);
2696 dr->dr_zio = zio_write(zio, os->os_spa, txg,
2697 db->db_blkptr, data->b_data, arc_buf_size(data), &zp,
2698 dbuf_write_override_ready, dbuf_write_override_done, dr,
2699 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
2700 mutex_enter(&db->db_mtx);
2701 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
2702 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
2703 dr->dt.dl.dr_copies);
2704 mutex_exit(&db->db_mtx);
2705 } else if (db->db_state == DB_NOFILL) {
2706 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF);
2707 dr->dr_zio = zio_write(zio, os->os_spa, txg,
2708 db->db_blkptr, NULL, db->db.db_size, &zp,
2709 dbuf_write_nofill_ready, dbuf_write_nofill_done, db,
2710 ZIO_PRIORITY_ASYNC_WRITE,
2711 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
2712 } else {
2713 ASSERT(arc_released(data));
2714 dr->dr_zio = arc_write(zio, os->os_spa, txg,
2715 db->db_blkptr, data, DBUF_IS_L2CACHEABLE(db), &zp,
2716 dbuf_write_ready, dbuf_write_done, db,
2717 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
2718 }
2719 }