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.
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.
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]
22 * Copyright 2009 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
26 #include <sys/zfs_context.h>
28 #include <sys/dmu_impl.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>
36 #include <sys/dmu_zfetch.h>
38 static void dbuf_destroy(dmu_buf_impl_t *db);
39 static int dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
40 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
41 static arc_done_func_t dbuf_write_ready;
42 static arc_done_func_t dbuf_write_done;
45 * Global data structures and functions for the dbuf cache.
47 static kmem_cache_t *dbuf_cache;
51 dbuf_cons(void *vdb, void *unused, int kmflag)
53 dmu_buf_impl_t *db = vdb;
54 bzero(db, sizeof (dmu_buf_impl_t));
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);
64 dbuf_dest(void *vdb, void *unused)
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);
73 * dbuf hash table routines
75 static dbuf_hash_table_t dbuf_hash_table;
77 static uint64_t dbuf_hash_count;
80 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
82 uintptr_t osv = (uintptr_t)os;
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];
93 crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16);
98 #define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid);
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))
107 dbuf_find(dnode_t *dn, uint8_t level, uint64_t blkid)
109 dbuf_hash_table_t *h = &dbuf_hash_table;
110 objset_impl_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;
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));
124 mutex_exit(&db->db_mtx);
127 mutex_exit(DBUF_HASH_MUTEX(h, idx));
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.
137 static dmu_buf_impl_t *
138 dbuf_hash_insert(dmu_buf_impl_t *db)
140 dbuf_hash_table_t *h = &dbuf_hash_table;
141 objset_impl_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;
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));
157 mutex_exit(&dbf->db_mtx);
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);
171 * Remove an entry from the hash table. This operation will
172 * fail if there are any existing holds on the db.
175 dbuf_hash_remove(dmu_buf_impl_t *db)
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;
184 * We musn't hold db_mtx to maintin lock ordering:
185 * DBUF_HASH_MUTEX > db_mtx.
187 ASSERT(refcount_is_zero(&db->db_holds));
188 ASSERT(db->db_state == DB_EVICTING);
189 ASSERT(!MUTEX_HELD(&db->db_mtx));
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;
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);
203 static arc_evict_func_t dbuf_do_evict;
206 dbuf_evict_user(dmu_buf_impl_t *db)
208 ASSERT(MUTEX_HELD(&db->db_mtx));
210 if (db->db_level != 0 || db->db_evict_func == NULL)
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;
222 dbuf_evict(dmu_buf_impl_t *db)
224 ASSERT(MUTEX_HELD(&db->db_mtx));
225 ASSERT(db->db_buf == NULL);
226 ASSERT(db->db_data_pending == NULL);
235 uint64_t hsize = 1ULL << 16;
236 dbuf_hash_table_t *h = &dbuf_hash_table;
240 * The hash table is big enough to fill all of physical memory
241 * with an average 4K block size. The table will take up
242 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
244 while (hsize * 4096 < (uint64_t)physmem * PAGESIZE)
248 h->hash_table_mask = hsize - 1;
249 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
250 if (h->hash_table == NULL) {
251 /* XXX - we should really return an error instead of assert */
252 ASSERT(hsize > (1ULL << 10));
257 dbuf_cache = kmem_cache_create("dmu_buf_impl_t",
258 sizeof (dmu_buf_impl_t),
259 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
261 for (i = 0; i < DBUF_MUTEXES; i++)
262 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
268 dbuf_hash_table_t *h = &dbuf_hash_table;
271 for (i = 0; i < DBUF_MUTEXES; i++)
272 mutex_destroy(&h->hash_mutexes[i]);
273 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
274 kmem_cache_destroy(dbuf_cache);
283 dbuf_verify(dmu_buf_impl_t *db)
285 dnode_t *dn = db->db_dnode;
287 ASSERT(MUTEX_HELD(&db->db_mtx));
289 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
292 ASSERT(db->db_objset != NULL);
294 ASSERT(db->db_parent == NULL);
295 ASSERT(db->db_blkptr == NULL);
297 ASSERT3U(db->db.db_object, ==, dn->dn_object);
298 ASSERT3P(db->db_objset, ==, dn->dn_objset);
299 ASSERT3U(db->db_level, <, dn->dn_nlevels);
300 ASSERT(db->db_blkid == DB_BONUS_BLKID ||
301 list_head(&dn->dn_dbufs));
303 if (db->db_blkid == DB_BONUS_BLKID) {
305 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
306 ASSERT3U(db->db.db_offset, ==, DB_BONUS_BLKID);
308 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
312 * We can't assert that db_size matches dn_datablksz because it
313 * can be momentarily different when another thread is doing
316 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
317 dbuf_dirty_record_t *dr = db->db_data_pending;
319 * It should only be modified in syncing context, so
320 * make sure we only have one copy of the data.
322 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
325 /* verify db->db_blkptr */
327 if (db->db_parent == dn->dn_dbuf) {
328 /* db is pointed to by the dnode */
329 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
330 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
331 ASSERT(db->db_parent == NULL);
333 ASSERT(db->db_parent != NULL);
334 ASSERT3P(db->db_blkptr, ==,
335 &dn->dn_phys->dn_blkptr[db->db_blkid]);
337 /* db is pointed to by an indirect block */
338 int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
339 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
340 ASSERT3U(db->db_parent->db.db_object, ==,
343 * dnode_grow_indblksz() can make this fail if we don't
344 * have the struct_rwlock. XXX indblksz no longer
345 * grows. safe to do this now?
347 if (RW_WRITE_HELD(&db->db_dnode->dn_struct_rwlock)) {
348 ASSERT3P(db->db_blkptr, ==,
349 ((blkptr_t *)db->db_parent->db.db_data +
350 db->db_blkid % epb));
354 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
355 db->db.db_data && db->db_blkid != DB_BONUS_BLKID &&
356 db->db_state != DB_FILL && !dn->dn_free_txg) {
358 * If the blkptr isn't set but they have nonzero data,
359 * it had better be dirty, otherwise we'll lose that
360 * data when we evict this buffer.
362 if (db->db_dirtycnt == 0) {
363 uint64_t *buf = db->db.db_data;
366 for (i = 0; i < db->db.db_size >> 3; i++) {
375 dbuf_update_data(dmu_buf_impl_t *db)
377 ASSERT(MUTEX_HELD(&db->db_mtx));
378 if (db->db_level == 0 && db->db_user_data_ptr_ptr) {
379 ASSERT(!refcount_is_zero(&db->db_holds));
380 *db->db_user_data_ptr_ptr = db->db.db_data;
385 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
387 ASSERT(MUTEX_HELD(&db->db_mtx));
388 ASSERT(db->db_buf == NULL || !arc_has_callback(db->db_buf));
391 ASSERT(buf->b_data != NULL);
392 db->db.db_data = buf->b_data;
393 if (!arc_released(buf))
394 arc_set_callback(buf, dbuf_do_evict, db);
395 dbuf_update_data(db);
398 db->db.db_data = NULL;
399 db->db_state = DB_UNCACHED;
404 dbuf_whichblock(dnode_t *dn, uint64_t offset)
406 if (dn->dn_datablkshift) {
407 return (offset >> dn->dn_datablkshift);
409 ASSERT3U(offset, <, dn->dn_datablksz);
415 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
417 dmu_buf_impl_t *db = vdb;
419 mutex_enter(&db->db_mtx);
420 ASSERT3U(db->db_state, ==, DB_READ);
422 * All reads are synchronous, so we must have a hold on the dbuf
424 ASSERT(refcount_count(&db->db_holds) > 0);
425 ASSERT(db->db_buf == NULL);
426 ASSERT(db->db.db_data == NULL);
427 if (db->db_level == 0 && db->db_freed_in_flight) {
428 /* we were freed in flight; disregard any error */
429 arc_release(buf, db);
430 bzero(buf->b_data, db->db.db_size);
432 db->db_freed_in_flight = FALSE;
433 dbuf_set_data(db, buf);
434 db->db_state = DB_CACHED;
435 } else if (zio == NULL || zio->io_error == 0) {
436 dbuf_set_data(db, buf);
437 db->db_state = DB_CACHED;
439 ASSERT(db->db_blkid != DB_BONUS_BLKID);
440 ASSERT3P(db->db_buf, ==, NULL);
441 VERIFY(arc_buf_remove_ref(buf, db) == 1);
442 db->db_state = DB_UNCACHED;
444 cv_broadcast(&db->db_changed);
445 mutex_exit(&db->db_mtx);
450 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t *flags)
452 dnode_t *dn = db->db_dnode;
454 uint32_t aflags = ARC_NOWAIT;
457 ASSERT(!refcount_is_zero(&db->db_holds));
458 /* We need the struct_rwlock to prevent db_blkptr from changing. */
459 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
460 ASSERT(MUTEX_HELD(&db->db_mtx));
461 ASSERT(db->db_state == DB_UNCACHED);
462 ASSERT(db->db_buf == NULL);
464 if (db->db_blkid == DB_BONUS_BLKID) {
465 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
467 ASSERT3U(bonuslen, <=, db->db.db_size);
468 db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN);
469 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
470 if (bonuslen < DN_MAX_BONUSLEN)
471 bzero(db->db.db_data, DN_MAX_BONUSLEN);
473 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
474 dbuf_update_data(db);
475 db->db_state = DB_CACHED;
476 mutex_exit(&db->db_mtx);
481 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
482 * processes the delete record and clears the bp while we are waiting
483 * for the dn_mtx (resulting in a "no" from block_freed).
485 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
486 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
487 BP_IS_HOLE(db->db_blkptr)))) {
488 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
490 dbuf_set_data(db, arc_buf_alloc(dn->dn_objset->os_spa,
491 db->db.db_size, db, type));
492 bzero(db->db.db_data, db->db.db_size);
493 db->db_state = DB_CACHED;
494 *flags |= DB_RF_CACHED;
495 mutex_exit(&db->db_mtx);
499 db->db_state = DB_READ;
500 mutex_exit(&db->db_mtx);
502 if (DBUF_IS_L2CACHEABLE(db))
503 aflags |= ARC_L2CACHE;
505 zb.zb_objset = db->db_objset->os_dsl_dataset ?
506 db->db_objset->os_dsl_dataset->ds_object : 0;
507 zb.zb_object = db->db.db_object;
508 zb.zb_level = db->db_level;
509 zb.zb_blkid = db->db_blkid;
511 dbuf_add_ref(db, NULL);
512 /* ZIO_FLAG_CANFAIL callers have to check the parent zio's error */
515 pbuf = db->db_parent->db_buf;
517 pbuf = db->db_objset->os_phys_buf;
519 (void) arc_read(zio, dn->dn_objset->os_spa, db->db_blkptr, pbuf,
520 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
521 (*flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
523 if (aflags & ARC_CACHED)
524 *flags |= DB_RF_CACHED;
528 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
531 int havepzio = (zio != NULL);
535 * We don't have to hold the mutex to check db_state because it
536 * can't be freed while we have a hold on the buffer.
538 ASSERT(!refcount_is_zero(&db->db_holds));
540 if ((flags & DB_RF_HAVESTRUCT) == 0)
541 rw_enter(&db->db_dnode->dn_struct_rwlock, RW_READER);
543 prefetch = db->db_level == 0 && db->db_blkid != DB_BONUS_BLKID &&
544 (flags & DB_RF_NOPREFETCH) == 0 && db->db_dnode != NULL &&
545 DBUF_IS_CACHEABLE(db);
547 mutex_enter(&db->db_mtx);
548 if (db->db_state == DB_CACHED) {
549 mutex_exit(&db->db_mtx);
551 dmu_zfetch(&db->db_dnode->dn_zfetch, db->db.db_offset,
552 db->db.db_size, TRUE);
553 if ((flags & DB_RF_HAVESTRUCT) == 0)
554 rw_exit(&db->db_dnode->dn_struct_rwlock);
555 } else if (db->db_state == DB_UNCACHED) {
557 zio = zio_root(db->db_dnode->dn_objset->os_spa,
558 NULL, NULL, ZIO_FLAG_CANFAIL);
560 dbuf_read_impl(db, zio, &flags);
562 /* dbuf_read_impl has dropped db_mtx for us */
565 dmu_zfetch(&db->db_dnode->dn_zfetch, db->db.db_offset,
566 db->db.db_size, flags & DB_RF_CACHED);
568 if ((flags & DB_RF_HAVESTRUCT) == 0)
569 rw_exit(&db->db_dnode->dn_struct_rwlock);
574 mutex_exit(&db->db_mtx);
576 dmu_zfetch(&db->db_dnode->dn_zfetch, db->db.db_offset,
577 db->db.db_size, TRUE);
578 if ((flags & DB_RF_HAVESTRUCT) == 0)
579 rw_exit(&db->db_dnode->dn_struct_rwlock);
581 mutex_enter(&db->db_mtx);
582 if ((flags & DB_RF_NEVERWAIT) == 0) {
583 while (db->db_state == DB_READ ||
584 db->db_state == DB_FILL) {
585 ASSERT(db->db_state == DB_READ ||
586 (flags & DB_RF_HAVESTRUCT) == 0);
587 cv_wait(&db->db_changed, &db->db_mtx);
589 if (db->db_state == DB_UNCACHED)
592 mutex_exit(&db->db_mtx);
595 ASSERT(err || havepzio || db->db_state == DB_CACHED);
600 dbuf_noread(dmu_buf_impl_t *db)
602 ASSERT(!refcount_is_zero(&db->db_holds));
603 ASSERT(db->db_blkid != DB_BONUS_BLKID);
604 mutex_enter(&db->db_mtx);
605 while (db->db_state == DB_READ || db->db_state == DB_FILL)
606 cv_wait(&db->db_changed, &db->db_mtx);
607 if (db->db_state == DB_UNCACHED) {
608 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
610 ASSERT(db->db_buf == NULL);
611 ASSERT(db->db.db_data == NULL);
612 dbuf_set_data(db, arc_buf_alloc(db->db_dnode->dn_objset->os_spa,
613 db->db.db_size, db, type));
614 db->db_state = DB_FILL;
616 ASSERT3U(db->db_state, ==, DB_CACHED);
618 mutex_exit(&db->db_mtx);
622 * This is our just-in-time copy function. It makes a copy of
623 * buffers, that have been modified in a previous transaction
624 * group, before we modify them in the current active group.
626 * This function is used in two places: when we are dirtying a
627 * buffer for the first time in a txg, and when we are freeing
628 * a range in a dnode that includes this buffer.
630 * Note that when we are called from dbuf_free_range() we do
631 * not put a hold on the buffer, we just traverse the active
632 * dbuf list for the dnode.
635 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
637 dbuf_dirty_record_t *dr = db->db_last_dirty;
639 ASSERT(MUTEX_HELD(&db->db_mtx));
640 ASSERT(db->db.db_data != NULL);
641 ASSERT(db->db_level == 0);
642 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
645 (dr->dt.dl.dr_data !=
646 ((db->db_blkid == DB_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
650 * If the last dirty record for this dbuf has not yet synced
651 * and its referencing the dbuf data, either:
652 * reset the reference to point to a new copy,
653 * or (if there a no active holders)
654 * just null out the current db_data pointer.
656 ASSERT(dr->dr_txg >= txg - 2);
657 if (db->db_blkid == DB_BONUS_BLKID) {
658 /* Note that the data bufs here are zio_bufs */
659 dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
660 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
661 bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
662 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
663 int size = db->db.db_size;
664 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
665 dr->dt.dl.dr_data = arc_buf_alloc(
666 db->db_dnode->dn_objset->os_spa, size, db, type);
667 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
669 dbuf_set_data(db, NULL);
674 dbuf_unoverride(dbuf_dirty_record_t *dr)
676 dmu_buf_impl_t *db = dr->dr_dbuf;
677 uint64_t txg = dr->dr_txg;
679 ASSERT(MUTEX_HELD(&db->db_mtx));
680 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
681 ASSERT(db->db_level == 0);
683 if (db->db_blkid == DB_BONUS_BLKID ||
684 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
687 /* free this block */
688 if (!BP_IS_HOLE(&dr->dt.dl.dr_overridden_by)) {
689 /* XXX can get silent EIO here */
690 (void) dsl_free(NULL,
691 spa_get_dsl(db->db_dnode->dn_objset->os_spa),
692 txg, &dr->dt.dl.dr_overridden_by, NULL, NULL, ARC_WAIT);
694 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
696 * Release the already-written buffer, so we leave it in
697 * a consistent dirty state. Note that all callers are
698 * modifying the buffer, so they will immediately do
699 * another (redundant) arc_release(). Therefore, leave
700 * the buf thawed to save the effort of freezing &
701 * immediately re-thawing it.
703 arc_release(dr->dt.dl.dr_data, db);
707 * Evict (if its unreferenced) or clear (if its referenced) any level-0
708 * data blocks in the free range, so that any future readers will find
709 * empty blocks. Also, if we happen accross any level-1 dbufs in the
710 * range that have not already been marked dirty, mark them dirty so
711 * they stay in memory.
714 dbuf_free_range(dnode_t *dn, uint64_t start, uint64_t end, dmu_tx_t *tx)
716 dmu_buf_impl_t *db, *db_next;
717 uint64_t txg = tx->tx_txg;
718 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
719 uint64_t first_l1 = start >> epbs;
720 uint64_t last_l1 = end >> epbs;
722 if (end > dn->dn_maxblkid) {
723 end = dn->dn_maxblkid;
724 last_l1 = end >> epbs;
726 dprintf_dnode(dn, "start=%llu end=%llu\n", start, end);
727 mutex_enter(&dn->dn_dbufs_mtx);
728 for (db = list_head(&dn->dn_dbufs); db; db = db_next) {
729 db_next = list_next(&dn->dn_dbufs, db);
730 ASSERT(db->db_blkid != DB_BONUS_BLKID);
732 if (db->db_level == 1 &&
733 db->db_blkid >= first_l1 && db->db_blkid <= last_l1) {
734 mutex_enter(&db->db_mtx);
735 if (db->db_last_dirty &&
736 db->db_last_dirty->dr_txg < txg) {
737 dbuf_add_ref(db, FTAG);
738 mutex_exit(&db->db_mtx);
739 dbuf_will_dirty(db, tx);
742 mutex_exit(&db->db_mtx);
746 if (db->db_level != 0)
748 dprintf_dbuf(db, "found buf %s\n", "");
749 if (db->db_blkid < start || db->db_blkid > end)
752 /* found a level 0 buffer in the range */
753 if (dbuf_undirty(db, tx))
756 mutex_enter(&db->db_mtx);
757 if (db->db_state == DB_UNCACHED ||
758 db->db_state == DB_EVICTING) {
759 ASSERT(db->db.db_data == NULL);
760 mutex_exit(&db->db_mtx);
763 if (db->db_state == DB_READ || db->db_state == DB_FILL) {
764 /* will be handled in dbuf_read_done or dbuf_rele */
765 db->db_freed_in_flight = TRUE;
766 mutex_exit(&db->db_mtx);
769 if (refcount_count(&db->db_holds) == 0) {
774 /* The dbuf is referenced */
776 if (db->db_last_dirty != NULL) {
777 dbuf_dirty_record_t *dr = db->db_last_dirty;
779 if (dr->dr_txg == txg) {
781 * This buffer is "in-use", re-adjust the file
782 * size to reflect that this buffer may
783 * contain new data when we sync.
785 if (db->db_blkid > dn->dn_maxblkid)
786 dn->dn_maxblkid = db->db_blkid;
790 * This dbuf is not dirty in the open context.
791 * Either uncache it (if its not referenced in
792 * the open context) or reset its contents to
795 dbuf_fix_old_data(db, txg);
798 /* clear the contents if its cached */
799 if (db->db_state == DB_CACHED) {
800 ASSERT(db->db.db_data != NULL);
801 arc_release(db->db_buf, db);
802 bzero(db->db.db_data, db->db.db_size);
803 arc_buf_freeze(db->db_buf);
806 mutex_exit(&db->db_mtx);
808 mutex_exit(&dn->dn_dbufs_mtx);
812 dbuf_block_freeable(dmu_buf_impl_t *db)
814 dsl_dataset_t *ds = db->db_objset->os_dsl_dataset;
815 uint64_t birth_txg = 0;
818 * We don't need any locking to protect db_blkptr:
819 * If it's syncing, then db_last_dirty will be set
820 * so we'll ignore db_blkptr.
822 ASSERT(MUTEX_HELD(&db->db_mtx));
823 if (db->db_last_dirty)
824 birth_txg = db->db_last_dirty->dr_txg;
825 else if (db->db_blkptr)
826 birth_txg = db->db_blkptr->blk_birth;
828 /* If we don't exist or are in a snapshot, we can't be freed */
830 return (ds == NULL ||
831 dsl_dataset_block_freeable(ds, birth_txg));
837 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
839 arc_buf_t *buf, *obuf;
840 int osize = db->db.db_size;
841 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
843 ASSERT(db->db_blkid != DB_BONUS_BLKID);
845 /* XXX does *this* func really need the lock? */
846 ASSERT(RW_WRITE_HELD(&db->db_dnode->dn_struct_rwlock));
849 * This call to dbuf_will_dirty() with the dn_struct_rwlock held
850 * is OK, because there can be no other references to the db
851 * when we are changing its size, so no concurrent DB_FILL can
855 * XXX we should be doing a dbuf_read, checking the return
856 * value and returning that up to our callers
858 dbuf_will_dirty(db, tx);
860 /* create the data buffer for the new block */
861 buf = arc_buf_alloc(db->db_dnode->dn_objset->os_spa, size, db, type);
863 /* copy old block data to the new block */
865 bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
866 /* zero the remainder */
868 bzero((uint8_t *)buf->b_data + osize, size - osize);
870 mutex_enter(&db->db_mtx);
871 dbuf_set_data(db, buf);
872 VERIFY(arc_buf_remove_ref(obuf, db) == 1);
873 db->db.db_size = size;
875 if (db->db_level == 0) {
876 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
877 db->db_last_dirty->dt.dl.dr_data = buf;
879 mutex_exit(&db->db_mtx);
881 dnode_willuse_space(db->db_dnode, size-osize, tx);
884 dbuf_dirty_record_t *
885 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
887 dnode_t *dn = db->db_dnode;
888 objset_impl_t *os = dn->dn_objset;
889 dbuf_dirty_record_t **drp, *dr;
890 int drop_struct_lock = FALSE;
891 boolean_t do_free_accounting = B_FALSE;
892 int txgoff = tx->tx_txg & TXG_MASK;
894 ASSERT(tx->tx_txg != 0);
895 ASSERT(!refcount_is_zero(&db->db_holds));
896 DMU_TX_DIRTY_BUF(tx, db);
899 * Shouldn't dirty a regular buffer in syncing context. Private
900 * objects may be dirtied in syncing context, but only if they
901 * were already pre-dirtied in open context.
903 ASSERT(!dmu_tx_is_syncing(tx) ||
904 BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
905 DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
906 dn->dn_objset->os_dsl_dataset == NULL);
908 * We make this assert for private objects as well, but after we
909 * check if we're already dirty. They are allowed to re-dirty
910 * in syncing context.
912 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
913 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
914 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
916 mutex_enter(&db->db_mtx);
918 * XXX make this true for indirects too? The problem is that
919 * transactions created with dmu_tx_create_assigned() from
920 * syncing context don't bother holding ahead.
922 ASSERT(db->db_level != 0 ||
923 db->db_state == DB_CACHED || db->db_state == DB_FILL);
925 mutex_enter(&dn->dn_mtx);
927 * Don't set dirtyctx to SYNC if we're just modifying this as we
928 * initialize the objset.
930 if (dn->dn_dirtyctx == DN_UNDIRTIED &&
931 !BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
933 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN);
934 ASSERT(dn->dn_dirtyctx_firstset == NULL);
935 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
937 mutex_exit(&dn->dn_mtx);
940 * If this buffer is already dirty, we're done.
942 drp = &db->db_last_dirty;
943 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
944 db->db.db_object == DMU_META_DNODE_OBJECT);
945 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
947 if (dr && dr->dr_txg == tx->tx_txg) {
948 if (db->db_level == 0 && db->db_blkid != DB_BONUS_BLKID) {
950 * If this buffer has already been written out,
951 * we now need to reset its state.
954 if (db->db.db_object != DMU_META_DNODE_OBJECT)
955 arc_buf_thaw(db->db_buf);
957 mutex_exit(&db->db_mtx);
962 * Only valid if not already dirty.
964 ASSERT(dn->dn_object == 0 ||
965 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
966 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
968 ASSERT3U(dn->dn_nlevels, >, db->db_level);
969 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
970 dn->dn_phys->dn_nlevels > db->db_level ||
971 dn->dn_next_nlevels[txgoff] > db->db_level ||
972 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
973 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
976 * We should only be dirtying in syncing context if it's the
977 * mos or we're initializing the os or it's a special object.
978 * However, we are allowed to dirty in syncing context provided
979 * we already dirtied it in open context. Hence we must make
980 * this assertion only if we're not already dirty.
982 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
983 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
984 ASSERT(db->db.db_size != 0);
986 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
988 if (db->db_blkid != DB_BONUS_BLKID) {
990 * Update the accounting.
991 * Note: we delay "free accounting" until after we drop
992 * the db_mtx. This keeps us from grabbing other locks
993 * (and possibly deadlocking) in bp_get_dasize() while
994 * also holding the db_mtx.
996 dnode_willuse_space(dn, db->db.db_size, tx);
997 do_free_accounting = dbuf_block_freeable(db);
1001 * If this buffer is dirty in an old transaction group we need
1002 * to make a copy of it so that the changes we make in this
1003 * transaction group won't leak out when we sync the older txg.
1005 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1006 if (db->db_level == 0) {
1007 void *data_old = db->db_buf;
1009 if (db->db_blkid == DB_BONUS_BLKID) {
1010 dbuf_fix_old_data(db, tx->tx_txg);
1011 data_old = db->db.db_data;
1012 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1014 * Release the data buffer from the cache so that we
1015 * can modify it without impacting possible other users
1016 * of this cached data block. Note that indirect
1017 * blocks and private objects are not released until the
1018 * syncing state (since they are only modified then).
1020 arc_release(db->db_buf, db);
1021 dbuf_fix_old_data(db, tx->tx_txg);
1022 data_old = db->db_buf;
1024 ASSERT(data_old != NULL);
1025 dr->dt.dl.dr_data = data_old;
1027 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1028 list_create(&dr->dt.di.dr_children,
1029 sizeof (dbuf_dirty_record_t),
1030 offsetof(dbuf_dirty_record_t, dr_dirty_node));
1033 dr->dr_txg = tx->tx_txg;
1038 * We could have been freed_in_flight between the dbuf_noread
1039 * and dbuf_dirty. We win, as though the dbuf_noread() had
1040 * happened after the free.
1042 if (db->db_level == 0 && db->db_blkid != DB_BONUS_BLKID) {
1043 mutex_enter(&dn->dn_mtx);
1044 dnode_clear_range(dn, db->db_blkid, 1, tx);
1045 mutex_exit(&dn->dn_mtx);
1046 db->db_freed_in_flight = FALSE;
1050 * This buffer is now part of this txg
1052 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1053 db->db_dirtycnt += 1;
1054 ASSERT3U(db->db_dirtycnt, <=, 3);
1056 mutex_exit(&db->db_mtx);
1058 if (db->db_blkid == DB_BONUS_BLKID) {
1059 mutex_enter(&dn->dn_mtx);
1060 ASSERT(!list_link_active(&dr->dr_dirty_node));
1061 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1062 mutex_exit(&dn->dn_mtx);
1063 dnode_setdirty(dn, tx);
1065 } else if (do_free_accounting) {
1066 blkptr_t *bp = db->db_blkptr;
1067 int64_t willfree = (bp && !BP_IS_HOLE(bp)) ?
1068 bp_get_dasize(os->os_spa, bp) : db->db.db_size;
1070 * This is only a guess -- if the dbuf is dirty
1071 * in a previous txg, we don't know how much
1072 * space it will use on disk yet. We should
1073 * really have the struct_rwlock to access
1074 * db_blkptr, but since this is just a guess,
1075 * it's OK if we get an odd answer.
1077 dnode_willuse_space(dn, -willfree, tx);
1080 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1081 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1082 drop_struct_lock = TRUE;
1085 if (db->db_level == 0) {
1086 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1087 ASSERT(dn->dn_maxblkid >= db->db_blkid);
1090 if (db->db_level+1 < dn->dn_nlevels) {
1091 dmu_buf_impl_t *parent = db->db_parent;
1092 dbuf_dirty_record_t *di;
1093 int parent_held = FALSE;
1095 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1096 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1098 parent = dbuf_hold_level(dn, db->db_level+1,
1099 db->db_blkid >> epbs, FTAG);
1102 if (drop_struct_lock)
1103 rw_exit(&dn->dn_struct_rwlock);
1104 ASSERT3U(db->db_level+1, ==, parent->db_level);
1105 di = dbuf_dirty(parent, tx);
1107 dbuf_rele(parent, FTAG);
1109 mutex_enter(&db->db_mtx);
1110 /* possible race with dbuf_undirty() */
1111 if (db->db_last_dirty == dr ||
1112 dn->dn_object == DMU_META_DNODE_OBJECT) {
1113 mutex_enter(&di->dt.di.dr_mtx);
1114 ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1115 ASSERT(!list_link_active(&dr->dr_dirty_node));
1116 list_insert_tail(&di->dt.di.dr_children, dr);
1117 mutex_exit(&di->dt.di.dr_mtx);
1120 mutex_exit(&db->db_mtx);
1122 ASSERT(db->db_level+1 == dn->dn_nlevels);
1123 ASSERT(db->db_blkid < dn->dn_nblkptr);
1124 ASSERT(db->db_parent == NULL ||
1125 db->db_parent == db->db_dnode->dn_dbuf);
1126 mutex_enter(&dn->dn_mtx);
1127 ASSERT(!list_link_active(&dr->dr_dirty_node));
1128 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1129 mutex_exit(&dn->dn_mtx);
1130 if (drop_struct_lock)
1131 rw_exit(&dn->dn_struct_rwlock);
1134 dnode_setdirty(dn, tx);
1139 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1141 dnode_t *dn = db->db_dnode;
1142 uint64_t txg = tx->tx_txg;
1143 dbuf_dirty_record_t *dr, **drp;
1146 ASSERT(db->db_blkid != DB_BONUS_BLKID);
1148 mutex_enter(&db->db_mtx);
1151 * If this buffer is not dirty, we're done.
1153 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1154 if (dr->dr_txg <= txg)
1156 if (dr == NULL || dr->dr_txg < txg) {
1157 mutex_exit(&db->db_mtx);
1160 ASSERT(dr->dr_txg == txg);
1163 * If this buffer is currently held, we cannot undirty
1164 * it, since one of the current holders may be in the
1165 * middle of an update. Note that users of dbuf_undirty()
1166 * should not place a hold on the dbuf before the call.
1168 if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
1169 mutex_exit(&db->db_mtx);
1170 /* Make sure we don't toss this buffer at sync phase */
1171 mutex_enter(&dn->dn_mtx);
1172 dnode_clear_range(dn, db->db_blkid, 1, tx);
1173 mutex_exit(&dn->dn_mtx);
1177 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1179 ASSERT(db->db.db_size != 0);
1181 /* XXX would be nice to fix up dn_towrite_space[] */
1185 if (dr->dr_parent) {
1186 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1187 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1188 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1189 } else if (db->db_level+1 == dn->dn_nlevels) {
1190 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1191 mutex_enter(&dn->dn_mtx);
1192 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1193 mutex_exit(&dn->dn_mtx);
1196 if (db->db_level == 0) {
1197 dbuf_unoverride(dr);
1199 ASSERT(db->db_buf != NULL);
1200 ASSERT(dr->dt.dl.dr_data != NULL);
1201 if (dr->dt.dl.dr_data != db->db_buf)
1202 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db) == 1);
1204 ASSERT(db->db_buf != NULL);
1205 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
1206 mutex_destroy(&dr->dt.di.dr_mtx);
1207 list_destroy(&dr->dt.di.dr_children);
1209 kmem_free(dr, sizeof (dbuf_dirty_record_t));
1211 ASSERT(db->db_dirtycnt > 0);
1212 db->db_dirtycnt -= 1;
1214 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1215 arc_buf_t *buf = db->db_buf;
1217 ASSERT(arc_released(buf));
1218 dbuf_set_data(db, NULL);
1219 VERIFY(arc_buf_remove_ref(buf, db) == 1);
1224 mutex_exit(&db->db_mtx);
1228 #pragma weak dmu_buf_will_dirty = dbuf_will_dirty
1230 dbuf_will_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1232 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1234 ASSERT(tx->tx_txg != 0);
1235 ASSERT(!refcount_is_zero(&db->db_holds));
1237 if (RW_WRITE_HELD(&db->db_dnode->dn_struct_rwlock))
1238 rf |= DB_RF_HAVESTRUCT;
1239 (void) dbuf_read(db, NULL, rf);
1240 (void) dbuf_dirty(db, tx);
1244 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1246 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1248 ASSERT(db->db_blkid != DB_BONUS_BLKID);
1249 ASSERT(tx->tx_txg != 0);
1250 ASSERT(db->db_level == 0);
1251 ASSERT(!refcount_is_zero(&db->db_holds));
1253 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1254 dmu_tx_private_ok(tx));
1257 (void) dbuf_dirty(db, tx);
1260 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1263 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1265 mutex_enter(&db->db_mtx);
1268 if (db->db_state == DB_FILL) {
1269 if (db->db_level == 0 && db->db_freed_in_flight) {
1270 ASSERT(db->db_blkid != DB_BONUS_BLKID);
1271 /* we were freed while filling */
1272 /* XXX dbuf_undirty? */
1273 bzero(db->db.db_data, db->db.db_size);
1274 db->db_freed_in_flight = FALSE;
1276 db->db_state = DB_CACHED;
1277 cv_broadcast(&db->db_changed);
1279 mutex_exit(&db->db_mtx);
1283 * Directly assign a provided arc buf to a given dbuf if it's not referenced
1284 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1287 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
1289 ASSERT(!refcount_is_zero(&db->db_holds));
1290 ASSERT(db->db_dnode->dn_object != DMU_META_DNODE_OBJECT);
1291 ASSERT(db->db_blkid != DB_BONUS_BLKID);
1292 ASSERT(db->db_level == 0);
1293 ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA);
1294 ASSERT(buf != NULL);
1295 ASSERT(arc_buf_size(buf) == db->db.db_size);
1296 ASSERT(tx->tx_txg != 0);
1298 arc_return_buf(buf, db);
1299 ASSERT(arc_released(buf));
1301 mutex_enter(&db->db_mtx);
1303 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1304 cv_wait(&db->db_changed, &db->db_mtx);
1306 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
1308 if (db->db_state == DB_CACHED &&
1309 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
1310 mutex_exit(&db->db_mtx);
1311 (void) dbuf_dirty(db, tx);
1312 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
1313 VERIFY(arc_buf_remove_ref(buf, db) == 1);
1317 if (db->db_state == DB_CACHED) {
1318 dbuf_dirty_record_t *dr = db->db_last_dirty;
1320 ASSERT(db->db_buf != NULL);
1321 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
1322 ASSERT(dr->dt.dl.dr_data == db->db_buf);
1323 if (!arc_released(db->db_buf)) {
1324 ASSERT(dr->dt.dl.dr_override_state ==
1326 arc_release(db->db_buf, db);
1328 dr->dt.dl.dr_data = buf;
1329 VERIFY(arc_buf_remove_ref(db->db_buf, db) == 1);
1330 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
1331 arc_release(db->db_buf, db);
1332 VERIFY(arc_buf_remove_ref(db->db_buf, db) == 1);
1336 ASSERT(db->db_buf == NULL);
1337 dbuf_set_data(db, buf);
1338 db->db_state = DB_FILL;
1339 mutex_exit(&db->db_mtx);
1340 (void) dbuf_dirty(db, tx);
1341 dbuf_fill_done(db, tx);
1345 * "Clear" the contents of this dbuf. This will mark the dbuf
1346 * EVICTING and clear *most* of its references. Unfortunetely,
1347 * when we are not holding the dn_dbufs_mtx, we can't clear the
1348 * entry in the dn_dbufs list. We have to wait until dbuf_destroy()
1349 * in this case. For callers from the DMU we will usually see:
1350 * dbuf_clear()->arc_buf_evict()->dbuf_do_evict()->dbuf_destroy()
1351 * For the arc callback, we will usually see:
1352 * dbuf_do_evict()->dbuf_clear();dbuf_destroy()
1353 * Sometimes, though, we will get a mix of these two:
1354 * DMU: dbuf_clear()->arc_buf_evict()
1355 * ARC: dbuf_do_evict()->dbuf_destroy()
1358 dbuf_clear(dmu_buf_impl_t *db)
1360 dnode_t *dn = db->db_dnode;
1361 dmu_buf_impl_t *parent = db->db_parent;
1362 dmu_buf_impl_t *dndb = dn->dn_dbuf;
1363 int dbuf_gone = FALSE;
1365 ASSERT(MUTEX_HELD(&db->db_mtx));
1366 ASSERT(refcount_is_zero(&db->db_holds));
1368 dbuf_evict_user(db);
1370 if (db->db_state == DB_CACHED) {
1371 ASSERT(db->db.db_data != NULL);
1372 if (db->db_blkid == DB_BONUS_BLKID) {
1373 zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
1374 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
1376 db->db.db_data = NULL;
1377 db->db_state = DB_UNCACHED;
1380 ASSERT3U(db->db_state, ==, DB_UNCACHED);
1381 ASSERT(db->db_data_pending == NULL);
1383 db->db_state = DB_EVICTING;
1384 db->db_blkptr = NULL;
1386 if (db->db_blkid != DB_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) {
1387 list_remove(&dn->dn_dbufs, db);
1389 db->db_dnode = NULL;
1393 dbuf_gone = arc_buf_evict(db->db_buf);
1396 mutex_exit(&db->db_mtx);
1399 * If this dbuf is referened from an indirect dbuf,
1400 * decrement the ref count on the indirect dbuf.
1402 if (parent && parent != dndb)
1403 dbuf_rele(parent, db);
1407 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
1408 dmu_buf_impl_t **parentp, blkptr_t **bpp)
1415 ASSERT(blkid != DB_BONUS_BLKID);
1417 if (dn->dn_phys->dn_nlevels == 0)
1420 nlevels = dn->dn_phys->dn_nlevels;
1422 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1424 ASSERT3U(level * epbs, <, 64);
1425 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1426 if (level >= nlevels ||
1427 (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
1428 /* the buffer has no parent yet */
1430 } else if (level < nlevels-1) {
1431 /* this block is referenced from an indirect block */
1432 int err = dbuf_hold_impl(dn, level+1,
1433 blkid >> epbs, fail_sparse, NULL, parentp);
1436 err = dbuf_read(*parentp, NULL,
1437 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
1439 dbuf_rele(*parentp, NULL);
1443 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
1444 (blkid & ((1ULL << epbs) - 1));
1447 /* the block is referenced from the dnode */
1448 ASSERT3U(level, ==, nlevels-1);
1449 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
1450 blkid < dn->dn_phys->dn_nblkptr);
1452 dbuf_add_ref(dn->dn_dbuf, NULL);
1453 *parentp = dn->dn_dbuf;
1455 *bpp = &dn->dn_phys->dn_blkptr[blkid];
1460 static dmu_buf_impl_t *
1461 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
1462 dmu_buf_impl_t *parent, blkptr_t *blkptr)
1464 objset_impl_t *os = dn->dn_objset;
1465 dmu_buf_impl_t *db, *odb;
1467 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1468 ASSERT(dn->dn_type != DMU_OT_NONE);
1470 db = kmem_cache_alloc(dbuf_cache, KM_SLEEP);
1473 db->db.db_object = dn->dn_object;
1474 db->db_level = level;
1475 db->db_blkid = blkid;
1476 db->db_last_dirty = NULL;
1477 db->db_dirtycnt = 0;
1479 db->db_parent = parent;
1480 db->db_blkptr = blkptr;
1482 db->db_user_ptr = NULL;
1483 db->db_user_data_ptr_ptr = NULL;
1484 db->db_evict_func = NULL;
1485 db->db_immediate_evict = 0;
1486 db->db_freed_in_flight = 0;
1488 if (blkid == DB_BONUS_BLKID) {
1489 ASSERT3P(parent, ==, dn->dn_dbuf);
1490 db->db.db_size = DN_MAX_BONUSLEN -
1491 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
1492 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
1493 db->db.db_offset = DB_BONUS_BLKID;
1494 db->db_state = DB_UNCACHED;
1495 /* the bonus dbuf is not placed in the hash table */
1496 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1500 db->db_level ? 1<<dn->dn_indblkshift : dn->dn_datablksz;
1501 db->db.db_size = blocksize;
1502 db->db.db_offset = db->db_blkid * blocksize;
1506 * Hold the dn_dbufs_mtx while we get the new dbuf
1507 * in the hash table *and* added to the dbufs list.
1508 * This prevents a possible deadlock with someone
1509 * trying to look up this dbuf before its added to the
1512 mutex_enter(&dn->dn_dbufs_mtx);
1513 db->db_state = DB_EVICTING;
1514 if ((odb = dbuf_hash_insert(db)) != NULL) {
1515 /* someone else inserted it first */
1516 kmem_cache_free(dbuf_cache, db);
1517 mutex_exit(&dn->dn_dbufs_mtx);
1520 list_insert_head(&dn->dn_dbufs, db);
1521 db->db_state = DB_UNCACHED;
1522 mutex_exit(&dn->dn_dbufs_mtx);
1523 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1525 if (parent && parent != dn->dn_dbuf)
1526 dbuf_add_ref(parent, db);
1528 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1529 refcount_count(&dn->dn_holds) > 0);
1530 (void) refcount_add(&dn->dn_holds, db);
1532 dprintf_dbuf(db, "db=%p\n", db);
1538 dbuf_do_evict(void *private)
1540 arc_buf_t *buf = private;
1541 dmu_buf_impl_t *db = buf->b_private;
1543 if (!MUTEX_HELD(&db->db_mtx))
1544 mutex_enter(&db->db_mtx);
1546 ASSERT(refcount_is_zero(&db->db_holds));
1548 if (db->db_state != DB_EVICTING) {
1549 ASSERT(db->db_state == DB_CACHED);
1554 mutex_exit(&db->db_mtx);
1561 dbuf_destroy(dmu_buf_impl_t *db)
1563 ASSERT(refcount_is_zero(&db->db_holds));
1565 if (db->db_blkid != DB_BONUS_BLKID) {
1567 * If this dbuf is still on the dn_dbufs list,
1568 * remove it from that list.
1571 dnode_t *dn = db->db_dnode;
1573 mutex_enter(&dn->dn_dbufs_mtx);
1574 list_remove(&dn->dn_dbufs, db);
1575 mutex_exit(&dn->dn_dbufs_mtx);
1578 db->db_dnode = NULL;
1580 dbuf_hash_remove(db);
1582 db->db_parent = NULL;
1585 ASSERT(!list_link_active(&db->db_link));
1586 ASSERT(db->db.db_data == NULL);
1587 ASSERT(db->db_hash_next == NULL);
1588 ASSERT(db->db_blkptr == NULL);
1589 ASSERT(db->db_data_pending == NULL);
1591 kmem_cache_free(dbuf_cache, db);
1592 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1596 dbuf_prefetch(dnode_t *dn, uint64_t blkid)
1598 dmu_buf_impl_t *db = NULL;
1599 blkptr_t *bp = NULL;
1601 ASSERT(blkid != DB_BONUS_BLKID);
1602 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1604 if (dnode_block_freed(dn, blkid))
1607 /* dbuf_find() returns with db_mtx held */
1608 if (db = dbuf_find(dn, 0, blkid)) {
1609 if (refcount_count(&db->db_holds) > 0) {
1611 * This dbuf is active. We assume that it is
1612 * already CACHED, or else about to be either
1615 mutex_exit(&db->db_mtx);
1618 mutex_exit(&db->db_mtx);
1622 if (dbuf_findbp(dn, 0, blkid, TRUE, &db, &bp) == 0) {
1623 if (bp && !BP_IS_HOLE(bp)) {
1625 uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH;
1627 zb.zb_objset = dn->dn_objset->os_dsl_dataset ?
1628 dn->dn_objset->os_dsl_dataset->ds_object : 0;
1629 zb.zb_object = dn->dn_object;
1631 zb.zb_blkid = blkid;
1636 pbuf = dn->dn_objset->os_phys_buf;
1638 (void) arc_read(NULL, dn->dn_objset->os_spa,
1639 bp, pbuf, NULL, NULL, ZIO_PRIORITY_ASYNC_READ,
1640 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1644 dbuf_rele(db, NULL);
1649 * Returns with db_holds incremented, and db_mtx not held.
1650 * Note: dn_struct_rwlock must be held.
1653 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse,
1654 void *tag, dmu_buf_impl_t **dbp)
1656 dmu_buf_impl_t *db, *parent = NULL;
1658 ASSERT(blkid != DB_BONUS_BLKID);
1659 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1660 ASSERT3U(dn->dn_nlevels, >, level);
1664 /* dbuf_find() returns with db_mtx held */
1665 db = dbuf_find(dn, level, blkid);
1668 blkptr_t *bp = NULL;
1671 ASSERT3P(parent, ==, NULL);
1672 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
1674 if (err == 0 && bp && BP_IS_HOLE(bp))
1678 dbuf_rele(parent, NULL);
1682 if (err && err != ENOENT)
1684 db = dbuf_create(dn, level, blkid, parent, bp);
1687 if (db->db_buf && refcount_is_zero(&db->db_holds)) {
1688 arc_buf_add_ref(db->db_buf, db);
1689 if (db->db_buf->b_data == NULL) {
1692 dbuf_rele(parent, NULL);
1697 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
1700 ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
1703 * If this buffer is currently syncing out, and we are are
1704 * still referencing it from db_data, we need to make a copy
1705 * of it in case we decide we want to dirty it again in this txg.
1707 if (db->db_level == 0 && db->db_blkid != DB_BONUS_BLKID &&
1708 dn->dn_object != DMU_META_DNODE_OBJECT &&
1709 db->db_state == DB_CACHED && db->db_data_pending) {
1710 dbuf_dirty_record_t *dr = db->db_data_pending;
1712 if (dr->dt.dl.dr_data == db->db_buf) {
1713 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1716 arc_buf_alloc(db->db_dnode->dn_objset->os_spa,
1717 db->db.db_size, db, type));
1718 bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
1723 (void) refcount_add(&db->db_holds, tag);
1724 dbuf_update_data(db);
1726 mutex_exit(&db->db_mtx);
1728 /* NOTE: we can't rele the parent until after we drop the db_mtx */
1730 dbuf_rele(parent, NULL);
1732 ASSERT3P(db->db_dnode, ==, dn);
1733 ASSERT3U(db->db_blkid, ==, blkid);
1734 ASSERT3U(db->db_level, ==, level);
1741 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
1744 int err = dbuf_hold_impl(dn, 0, blkid, FALSE, tag, &db);
1745 return (err ? NULL : db);
1749 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
1752 int err = dbuf_hold_impl(dn, level, blkid, FALSE, tag, &db);
1753 return (err ? NULL : db);
1757 dbuf_create_bonus(dnode_t *dn)
1759 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1761 ASSERT(dn->dn_bonus == NULL);
1762 dn->dn_bonus = dbuf_create(dn, 0, DB_BONUS_BLKID, dn->dn_dbuf, NULL);
1765 #pragma weak dmu_buf_add_ref = dbuf_add_ref
1767 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
1769 int64_t holds = refcount_add(&db->db_holds, tag);
1773 #pragma weak dmu_buf_rele = dbuf_rele
1775 dbuf_rele(dmu_buf_impl_t *db, void *tag)
1779 mutex_enter(&db->db_mtx);
1782 holds = refcount_remove(&db->db_holds, tag);
1786 * We can't freeze indirects if there is a possibility that they
1787 * may be modified in the current syncing context.
1789 if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0))
1790 arc_buf_freeze(db->db_buf);
1792 if (holds == db->db_dirtycnt &&
1793 db->db_level == 0 && db->db_immediate_evict)
1794 dbuf_evict_user(db);
1797 if (db->db_blkid == DB_BONUS_BLKID) {
1798 mutex_exit(&db->db_mtx);
1799 dnode_rele(db->db_dnode, db);
1800 } else if (db->db_buf == NULL) {
1802 * This is a special case: we never associated this
1803 * dbuf with any data allocated from the ARC.
1805 ASSERT3U(db->db_state, ==, DB_UNCACHED);
1807 } else if (arc_released(db->db_buf)) {
1808 arc_buf_t *buf = db->db_buf;
1810 * This dbuf has anonymous data associated with it.
1812 dbuf_set_data(db, NULL);
1813 VERIFY(arc_buf_remove_ref(buf, db) == 1);
1816 VERIFY(arc_buf_remove_ref(db->db_buf, db) == 0);
1817 if (!DBUF_IS_CACHEABLE(db))
1820 mutex_exit(&db->db_mtx);
1823 mutex_exit(&db->db_mtx);
1827 #pragma weak dmu_buf_refcount = dbuf_refcount
1829 dbuf_refcount(dmu_buf_impl_t *db)
1831 return (refcount_count(&db->db_holds));
1835 dmu_buf_set_user(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr,
1836 dmu_buf_evict_func_t *evict_func)
1838 return (dmu_buf_update_user(db_fake, NULL, user_ptr,
1839 user_data_ptr_ptr, evict_func));
1843 dmu_buf_set_user_ie(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr,
1844 dmu_buf_evict_func_t *evict_func)
1846 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1848 db->db_immediate_evict = TRUE;
1849 return (dmu_buf_update_user(db_fake, NULL, user_ptr,
1850 user_data_ptr_ptr, evict_func));
1854 dmu_buf_update_user(dmu_buf_t *db_fake, void *old_user_ptr, void *user_ptr,
1855 void *user_data_ptr_ptr, dmu_buf_evict_func_t *evict_func)
1857 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1858 ASSERT(db->db_level == 0);
1860 ASSERT((user_ptr == NULL) == (evict_func == NULL));
1862 mutex_enter(&db->db_mtx);
1864 if (db->db_user_ptr == old_user_ptr) {
1865 db->db_user_ptr = user_ptr;
1866 db->db_user_data_ptr_ptr = user_data_ptr_ptr;
1867 db->db_evict_func = evict_func;
1869 dbuf_update_data(db);
1871 old_user_ptr = db->db_user_ptr;
1874 mutex_exit(&db->db_mtx);
1875 return (old_user_ptr);
1879 dmu_buf_get_user(dmu_buf_t *db_fake)
1881 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1882 ASSERT(!refcount_is_zero(&db->db_holds));
1884 return (db->db_user_ptr);
1888 dmu_buf_freeable(dmu_buf_t *dbuf)
1890 boolean_t res = B_FALSE;
1891 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
1894 res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset,
1895 db->db_blkptr->blk_birth);
1901 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
1903 /* ASSERT(dmu_tx_is_syncing(tx) */
1904 ASSERT(MUTEX_HELD(&db->db_mtx));
1906 if (db->db_blkptr != NULL)
1909 if (db->db_level == dn->dn_phys->dn_nlevels-1) {
1911 * This buffer was allocated at a time when there was
1912 * no available blkptrs from the dnode, or it was
1913 * inappropriate to hook it in (i.e., nlevels mis-match).
1915 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
1916 ASSERT(db->db_parent == NULL);
1917 db->db_parent = dn->dn_dbuf;
1918 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
1921 dmu_buf_impl_t *parent = db->db_parent;
1922 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
1924 ASSERT(dn->dn_phys->dn_nlevels > 1);
1925 if (parent == NULL) {
1926 mutex_exit(&db->db_mtx);
1927 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1928 (void) dbuf_hold_impl(dn, db->db_level+1,
1929 db->db_blkid >> epbs, FALSE, db, &parent);
1930 rw_exit(&dn->dn_struct_rwlock);
1931 mutex_enter(&db->db_mtx);
1932 db->db_parent = parent;
1934 db->db_blkptr = (blkptr_t *)parent->db.db_data +
1935 (db->db_blkid & ((1ULL << epbs) - 1));
1941 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
1943 dmu_buf_impl_t *db = dr->dr_dbuf;
1944 dnode_t *dn = db->db_dnode;
1947 ASSERT(dmu_tx_is_syncing(tx));
1949 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
1951 mutex_enter(&db->db_mtx);
1953 ASSERT(db->db_level > 0);
1956 if (db->db_buf == NULL) {
1957 mutex_exit(&db->db_mtx);
1958 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
1959 mutex_enter(&db->db_mtx);
1961 ASSERT3U(db->db_state, ==, DB_CACHED);
1962 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
1963 ASSERT(db->db_buf != NULL);
1965 dbuf_check_blkptr(dn, db);
1967 db->db_data_pending = dr;
1969 mutex_exit(&db->db_mtx);
1970 dbuf_write(dr, db->db_buf, tx);
1973 mutex_enter(&dr->dt.di.dr_mtx);
1974 dbuf_sync_list(&dr->dt.di.dr_children, tx);
1975 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
1976 mutex_exit(&dr->dt.di.dr_mtx);
1981 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
1983 arc_buf_t **datap = &dr->dt.dl.dr_data;
1984 dmu_buf_impl_t *db = dr->dr_dbuf;
1985 dnode_t *dn = db->db_dnode;
1986 objset_impl_t *os = dn->dn_objset;
1987 uint64_t txg = tx->tx_txg;
1989 ASSERT(dmu_tx_is_syncing(tx));
1991 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
1993 mutex_enter(&db->db_mtx);
1995 * To be synced, we must be dirtied. But we
1996 * might have been freed after the dirty.
1998 if (db->db_state == DB_UNCACHED) {
1999 /* This buffer has been freed since it was dirtied */
2000 ASSERT(db->db.db_data == NULL);
2001 } else if (db->db_state == DB_FILL) {
2002 /* This buffer was freed and is now being re-filled */
2003 ASSERT(db->db.db_data != dr->dt.dl.dr_data);
2005 ASSERT3U(db->db_state, ==, DB_CACHED);
2010 * If this is a bonus buffer, simply copy the bonus data into the
2011 * dnode. It will be written out when the dnode is synced (and it
2012 * will be synced, since it must have been dirty for dbuf_sync to
2015 if (db->db_blkid == DB_BONUS_BLKID) {
2016 dbuf_dirty_record_t **drp;
2018 ASSERT(*datap != NULL);
2019 ASSERT3U(db->db_level, ==, 0);
2020 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN);
2021 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen);
2022 if (*datap != db->db.db_data) {
2023 zio_buf_free(*datap, DN_MAX_BONUSLEN);
2024 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
2026 db->db_data_pending = NULL;
2027 drp = &db->db_last_dirty;
2029 drp = &(*drp)->dr_next;
2030 ASSERT(dr->dr_next == NULL);
2032 if (dr->dr_dbuf->db_level != 0) {
2033 list_destroy(&dr->dt.di.dr_children);
2034 mutex_destroy(&dr->dt.di.dr_mtx);
2036 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2037 ASSERT(db->db_dirtycnt > 0);
2038 db->db_dirtycnt -= 1;
2039 mutex_exit(&db->db_mtx);
2040 dbuf_rele(db, (void *)(uintptr_t)txg);
2045 * This function may have dropped the db_mtx lock allowing a dmu_sync
2046 * operation to sneak in. As a result, we need to ensure that we
2047 * don't check the dr_override_state until we have returned from
2048 * dbuf_check_blkptr.
2050 dbuf_check_blkptr(dn, db);
2053 * If this buffer is in the middle of an immdiate write,
2054 * wait for the synchronous IO to complete.
2056 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
2057 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
2058 cv_wait(&db->db_changed, &db->db_mtx);
2059 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
2063 * If this dbuf has already been written out via an immediate write,
2064 * just complete the write by copying over the new block pointer and
2065 * updating the accounting via the write-completion functions.
2067 if (dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
2070 zio_fake.io_private = &db;
2071 zio_fake.io_error = 0;
2072 zio_fake.io_bp = db->db_blkptr;
2073 zio_fake.io_bp_orig = *db->db_blkptr;
2074 zio_fake.io_txg = txg;
2075 zio_fake.io_flags = 0;
2077 *db->db_blkptr = dr->dt.dl.dr_overridden_by;
2078 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
2079 db->db_data_pending = dr;
2080 dr->dr_zio = &zio_fake;
2081 mutex_exit(&db->db_mtx);
2083 ASSERT(!DVA_EQUAL(BP_IDENTITY(zio_fake.io_bp),
2084 BP_IDENTITY(&zio_fake.io_bp_orig)) ||
2085 BP_IS_HOLE(zio_fake.io_bp));
2087 if (BP_IS_OLDER(&zio_fake.io_bp_orig, txg))
2088 (void) dsl_dataset_block_kill(os->os_dsl_dataset,
2089 &zio_fake.io_bp_orig, dn->dn_zio, tx);
2091 dbuf_write_ready(&zio_fake, db->db_buf, db);
2092 dbuf_write_done(&zio_fake, db->db_buf, db);
2097 if (dn->dn_object != DMU_META_DNODE_OBJECT &&
2098 refcount_count(&db->db_holds) > 1 &&
2099 *datap == db->db_buf) {
2101 * If this buffer is currently "in use" (i.e., there
2102 * are active holds and db_data still references it),
2103 * then make a copy before we start the write so that
2104 * any modifications from the open txg will not leak
2107 * NOTE: this copy does not need to be made for
2108 * objects only modified in the syncing context (e.g.
2109 * DNONE_DNODE blocks).
2111 int blksz = arc_buf_size(*datap);
2112 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2113 *datap = arc_buf_alloc(os->os_spa, blksz, db, type);
2114 bcopy(db->db.db_data, (*datap)->b_data, blksz);
2117 ASSERT(*datap != NULL);
2118 db->db_data_pending = dr;
2120 mutex_exit(&db->db_mtx);
2122 dbuf_write(dr, *datap, tx);
2124 ASSERT(!list_link_active(&dr->dr_dirty_node));
2125 if (dn->dn_object == DMU_META_DNODE_OBJECT)
2126 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
2128 zio_nowait(dr->dr_zio);
2132 dbuf_sync_list(list_t *list, dmu_tx_t *tx)
2134 dbuf_dirty_record_t *dr;
2136 while (dr = list_head(list)) {
2137 if (dr->dr_zio != NULL) {
2139 * If we find an already initialized zio then we
2140 * are processing the meta-dnode, and we have finished.
2141 * The dbufs for all dnodes are put back on the list
2142 * during processing, so that we can zio_wait()
2143 * these IOs after initiating all child IOs.
2145 ASSERT3U(dr->dr_dbuf->db.db_object, ==,
2146 DMU_META_DNODE_OBJECT);
2149 list_remove(list, dr);
2150 if (dr->dr_dbuf->db_level > 0)
2151 dbuf_sync_indirect(dr, tx);
2153 dbuf_sync_leaf(dr, tx);
2158 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
2160 dmu_buf_impl_t *db = dr->dr_dbuf;
2161 dnode_t *dn = db->db_dnode;
2162 objset_impl_t *os = dn->dn_objset;
2163 dmu_buf_impl_t *parent = db->db_parent;
2164 uint64_t txg = tx->tx_txg;
2166 writeprops_t wp = { 0 };
2169 if (!BP_IS_HOLE(db->db_blkptr) &&
2170 (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE)) {
2172 * Private object buffers are released here rather
2173 * than in dbuf_dirty() since they are only modified
2174 * in the syncing context and we don't want the
2175 * overhead of making multiple copies of the data.
2177 arc_release(data, db);
2179 ASSERT(arc_released(data));
2180 /* XXX why do we need to thaw here? */
2184 if (parent != dn->dn_dbuf) {
2185 ASSERT(parent && parent->db_data_pending);
2186 ASSERT(db->db_level == parent->db_level-1);
2187 ASSERT(arc_released(parent->db_buf));
2188 zio = parent->db_data_pending->dr_zio;
2190 ASSERT(db->db_level == dn->dn_phys->dn_nlevels-1);
2191 ASSERT3P(db->db_blkptr, ==,
2192 &dn->dn_phys->dn_blkptr[db->db_blkid]);
2196 ASSERT(db->db_level == 0 || data == db->db_buf);
2197 ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
2200 zb.zb_objset = os->os_dsl_dataset ? os->os_dsl_dataset->ds_object : 0;
2201 zb.zb_object = db->db.db_object;
2202 zb.zb_level = db->db_level;
2203 zb.zb_blkid = db->db_blkid;
2205 wp.wp_type = dn->dn_type;
2206 wp.wp_level = db->db_level;
2207 wp.wp_copies = os->os_copies;
2208 wp.wp_dncompress = dn->dn_compress;
2209 wp.wp_oscompress = os->os_compress;
2210 wp.wp_dnchecksum = dn->dn_checksum;
2211 wp.wp_oschecksum = os->os_checksum;
2213 if (BP_IS_OLDER(db->db_blkptr, txg))
2214 (void) dsl_dataset_block_kill(
2215 os->os_dsl_dataset, db->db_blkptr, zio, tx);
2217 dr->dr_zio = arc_write(zio, os->os_spa, &wp,
2218 DBUF_IS_L2CACHEABLE(db), txg, db->db_blkptr,
2219 data, dbuf_write_ready, dbuf_write_done, db,
2220 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
2225 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
2227 dmu_buf_impl_t *db = vdb;
2228 dnode_t *dn = db->db_dnode;
2229 objset_impl_t *os = dn->dn_objset;
2230 blkptr_t *bp = zio->io_bp;
2231 blkptr_t *bp_orig = &zio->io_bp_orig;
2233 int old_size, new_size, i;
2235 ASSERT(db->db_blkptr == bp);
2237 dprintf_dbuf_bp(db, bp_orig, "bp_orig: %s", "");
2239 old_size = bp_get_dasize(os->os_spa, bp_orig);
2240 new_size = bp_get_dasize(os->os_spa, bp);
2242 dnode_diduse_space(dn, new_size - old_size);
2244 if (BP_IS_HOLE(bp)) {
2245 dsl_dataset_t *ds = os->os_dsl_dataset;
2246 dmu_tx_t *tx = os->os_synctx;
2248 if (bp_orig->blk_birth == tx->tx_txg)
2249 (void) dsl_dataset_block_kill(ds, bp_orig, zio, tx);
2250 ASSERT3U(bp->blk_fill, ==, 0);
2254 ASSERT(BP_GET_TYPE(bp) == dn->dn_type);
2255 ASSERT(BP_GET_LEVEL(bp) == db->db_level);
2257 mutex_enter(&db->db_mtx);
2259 if (db->db_level == 0) {
2260 mutex_enter(&dn->dn_mtx);
2261 if (db->db_blkid > dn->dn_phys->dn_maxblkid)
2262 dn->dn_phys->dn_maxblkid = db->db_blkid;
2263 mutex_exit(&dn->dn_mtx);
2265 if (dn->dn_type == DMU_OT_DNODE) {
2266 dnode_phys_t *dnp = db->db.db_data;
2267 for (i = db->db.db_size >> DNODE_SHIFT; i > 0;
2269 if (dnp->dn_type != DMU_OT_NONE)
2276 blkptr_t *ibp = db->db.db_data;
2277 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2278 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
2279 if (BP_IS_HOLE(ibp))
2281 fill += ibp->blk_fill;
2285 bp->blk_fill = fill;
2287 mutex_exit(&db->db_mtx);
2289 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
2290 ASSERT(DVA_EQUAL(BP_IDENTITY(bp), BP_IDENTITY(bp_orig)));
2292 dsl_dataset_t *ds = os->os_dsl_dataset;
2293 dmu_tx_t *tx = os->os_synctx;
2295 if (bp_orig->blk_birth == tx->tx_txg)
2296 (void) dsl_dataset_block_kill(ds, bp_orig, zio, tx);
2297 dsl_dataset_block_born(ds, bp, tx);
2303 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
2305 dmu_buf_impl_t *db = vdb;
2306 uint64_t txg = zio->io_txg;
2307 dbuf_dirty_record_t **drp, *dr;
2309 ASSERT3U(zio->io_error, ==, 0);
2311 mutex_enter(&db->db_mtx);
2313 drp = &db->db_last_dirty;
2314 while ((dr = *drp) != db->db_data_pending)
2316 ASSERT(!list_link_active(&dr->dr_dirty_node));
2317 ASSERT(dr->dr_txg == txg);
2318 ASSERT(dr->dr_next == NULL);
2321 if (db->db_level == 0) {
2322 ASSERT(db->db_blkid != DB_BONUS_BLKID);
2323 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
2325 if (dr->dt.dl.dr_data != db->db_buf)
2326 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db) == 1);
2327 else if (!BP_IS_HOLE(db->db_blkptr))
2328 arc_set_callback(db->db_buf, dbuf_do_evict, db);
2330 ASSERT(arc_released(db->db_buf));
2332 dnode_t *dn = db->db_dnode;
2334 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2335 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2336 if (!BP_IS_HOLE(db->db_blkptr)) {
2338 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2339 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
2341 ASSERT3U(dn->dn_phys->dn_maxblkid
2342 >> (db->db_level * epbs), >=, db->db_blkid);
2343 arc_set_callback(db->db_buf, dbuf_do_evict, db);
2345 mutex_destroy(&dr->dt.di.dr_mtx);
2346 list_destroy(&dr->dt.di.dr_children);
2348 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2350 cv_broadcast(&db->db_changed);
2351 ASSERT(db->db_dirtycnt > 0);
2352 db->db_dirtycnt -= 1;
2353 db->db_data_pending = NULL;
2354 mutex_exit(&db->db_mtx);
2356 dprintf_dbuf_bp(db, zio->io_bp, "bp: %s", "");
2358 dbuf_rele(db, (void *)(uintptr_t)txg);