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 (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
24 * Copyright (c) 2012, 2015 by Delphix. All rights reserved.
25 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
26 * Copyright (c) 2013, Joyent, Inc. All rights reserved.
27 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
30 #include <sys/zfs_context.h>
32 #include <sys/dmu_send.h>
33 #include <sys/dmu_impl.h>
35 #include <sys/dmu_objset.h>
36 #include <sys/dsl_dataset.h>
37 #include <sys/dsl_dir.h>
38 #include <sys/dmu_tx.h>
41 #include <sys/dmu_zfetch.h>
43 #include <sys/sa_impl.h>
44 #include <sys/zfeature.h>
45 #include <sys/blkptr.h>
46 #include <sys/range_tree.h>
49 * Number of times that zfs_free_range() took the slow path while doing
50 * a zfs receive. A nonzero value indicates a potential performance problem.
52 uint64_t zfs_free_range_recv_miss;
54 static void dbuf_destroy(dmu_buf_impl_t *db);
55 static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
56 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
59 * Global data structures and functions for the dbuf cache.
61 static kmem_cache_t *dbuf_cache;
62 static taskq_t *dbu_evict_taskq;
66 dbuf_cons(void *vdb, void *unused, int kmflag)
68 dmu_buf_impl_t *db = vdb;
69 bzero(db, sizeof (dmu_buf_impl_t));
71 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
72 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
73 refcount_create(&db->db_holds);
80 dbuf_dest(void *vdb, void *unused)
82 dmu_buf_impl_t *db = vdb;
83 mutex_destroy(&db->db_mtx);
84 cv_destroy(&db->db_changed);
85 refcount_destroy(&db->db_holds);
89 * dbuf hash table routines
91 static dbuf_hash_table_t dbuf_hash_table;
93 static uint64_t dbuf_hash_count;
96 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
98 uintptr_t osv = (uintptr_t)os;
101 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
102 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF];
103 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF];
104 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF];
105 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF];
106 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF];
107 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF];
109 crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16);
114 #define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid);
116 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
117 ((dbuf)->db.db_object == (obj) && \
118 (dbuf)->db_objset == (os) && \
119 (dbuf)->db_level == (level) && \
120 (dbuf)->db_blkid == (blkid))
123 dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid)
125 dbuf_hash_table_t *h = &dbuf_hash_table;
126 uint64_t hv = DBUF_HASH(os, obj, level, blkid);
127 uint64_t idx = hv & h->hash_table_mask;
130 mutex_enter(DBUF_HASH_MUTEX(h, idx));
131 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
132 if (DBUF_EQUAL(db, os, obj, level, blkid)) {
133 mutex_enter(&db->db_mtx);
134 if (db->db_state != DB_EVICTING) {
135 mutex_exit(DBUF_HASH_MUTEX(h, idx));
138 mutex_exit(&db->db_mtx);
141 mutex_exit(DBUF_HASH_MUTEX(h, idx));
145 static dmu_buf_impl_t *
146 dbuf_find_bonus(objset_t *os, uint64_t object)
149 dmu_buf_impl_t *db = NULL;
151 if (dnode_hold(os, object, FTAG, &dn) == 0) {
152 rw_enter(&dn->dn_struct_rwlock, RW_READER);
153 if (dn->dn_bonus != NULL) {
155 mutex_enter(&db->db_mtx);
157 rw_exit(&dn->dn_struct_rwlock);
158 dnode_rele(dn, FTAG);
164 * Insert an entry into the hash table. If there is already an element
165 * equal to elem in the hash table, then the already existing element
166 * will be returned and the new element will not be inserted.
167 * Otherwise returns NULL.
169 static dmu_buf_impl_t *
170 dbuf_hash_insert(dmu_buf_impl_t *db)
172 dbuf_hash_table_t *h = &dbuf_hash_table;
173 objset_t *os = db->db_objset;
174 uint64_t obj = db->db.db_object;
175 int level = db->db_level;
176 uint64_t blkid = db->db_blkid;
177 uint64_t hv = DBUF_HASH(os, obj, level, blkid);
178 uint64_t idx = hv & h->hash_table_mask;
181 mutex_enter(DBUF_HASH_MUTEX(h, idx));
182 for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) {
183 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
184 mutex_enter(&dbf->db_mtx);
185 if (dbf->db_state != DB_EVICTING) {
186 mutex_exit(DBUF_HASH_MUTEX(h, idx));
189 mutex_exit(&dbf->db_mtx);
193 mutex_enter(&db->db_mtx);
194 db->db_hash_next = h->hash_table[idx];
195 h->hash_table[idx] = db;
196 mutex_exit(DBUF_HASH_MUTEX(h, idx));
197 atomic_inc_64(&dbuf_hash_count);
203 * Remove an entry from the hash table. It must be in the EVICTING state.
206 dbuf_hash_remove(dmu_buf_impl_t *db)
208 dbuf_hash_table_t *h = &dbuf_hash_table;
209 uint64_t hv = DBUF_HASH(db->db_objset, db->db.db_object,
210 db->db_level, db->db_blkid);
211 uint64_t idx = hv & h->hash_table_mask;
212 dmu_buf_impl_t *dbf, **dbp;
215 * We musn't hold db_mtx to maintain lock ordering:
216 * DBUF_HASH_MUTEX > db_mtx.
218 ASSERT(refcount_is_zero(&db->db_holds));
219 ASSERT(db->db_state == DB_EVICTING);
220 ASSERT(!MUTEX_HELD(&db->db_mtx));
222 mutex_enter(DBUF_HASH_MUTEX(h, idx));
223 dbp = &h->hash_table[idx];
224 while ((dbf = *dbp) != db) {
225 dbp = &dbf->db_hash_next;
228 *dbp = db->db_hash_next;
229 db->db_hash_next = NULL;
230 mutex_exit(DBUF_HASH_MUTEX(h, idx));
231 atomic_dec_64(&dbuf_hash_count);
234 static arc_evict_func_t dbuf_do_evict;
239 } dbvu_verify_type_t;
242 dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type)
247 if (db->db_user == NULL)
250 /* Only data blocks support the attachment of user data. */
251 ASSERT(db->db_level == 0);
253 /* Clients must resolve a dbuf before attaching user data. */
254 ASSERT(db->db.db_data != NULL);
255 ASSERT3U(db->db_state, ==, DB_CACHED);
257 holds = refcount_count(&db->db_holds);
258 if (verify_type == DBVU_EVICTING) {
260 * Immediate eviction occurs when holds == dirtycnt.
261 * For normal eviction buffers, holds is zero on
262 * eviction, except when dbuf_fix_old_data() calls
263 * dbuf_clear_data(). However, the hold count can grow
264 * during eviction even though db_mtx is held (see
265 * dmu_bonus_hold() for an example), so we can only
266 * test the generic invariant that holds >= dirtycnt.
268 ASSERT3U(holds, >=, db->db_dirtycnt);
270 if (db->db_user_immediate_evict == TRUE)
271 ASSERT3U(holds, >=, db->db_dirtycnt);
273 ASSERT3U(holds, >, 0);
279 dbuf_evict_user(dmu_buf_impl_t *db)
281 dmu_buf_user_t *dbu = db->db_user;
283 ASSERT(MUTEX_HELD(&db->db_mtx));
288 dbuf_verify_user(db, DBVU_EVICTING);
292 if (dbu->dbu_clear_on_evict_dbufp != NULL)
293 *dbu->dbu_clear_on_evict_dbufp = NULL;
297 * Invoke the callback from a taskq to avoid lock order reversals
298 * and limit stack depth.
300 taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func, dbu, 0,
305 dbuf_is_metadata(dmu_buf_impl_t *db)
307 if (db->db_level > 0) {
310 boolean_t is_metadata;
313 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type);
316 return (is_metadata);
321 dbuf_evict(dmu_buf_impl_t *db)
323 ASSERT(MUTEX_HELD(&db->db_mtx));
324 ASSERT(db->db_buf == NULL);
325 ASSERT(db->db_data_pending == NULL);
334 uint64_t hsize = 1ULL << 16;
335 dbuf_hash_table_t *h = &dbuf_hash_table;
339 * The hash table is big enough to fill all of physical memory
340 * with an average 4K block size. The table will take up
341 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
343 while (hsize * 4096 < (uint64_t)physmem * PAGESIZE)
347 h->hash_table_mask = hsize - 1;
348 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
349 if (h->hash_table == NULL) {
350 /* XXX - we should really return an error instead of assert */
351 ASSERT(hsize > (1ULL << 10));
356 dbuf_cache = kmem_cache_create("dmu_buf_impl_t",
357 sizeof (dmu_buf_impl_t),
358 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
360 for (i = 0; i < DBUF_MUTEXES; i++)
361 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
364 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
365 * configuration is not required.
367 dbu_evict_taskq = taskq_create("dbu_evict", 1, minclsyspri, 0, 0, 0);
373 dbuf_hash_table_t *h = &dbuf_hash_table;
376 for (i = 0; i < DBUF_MUTEXES; i++)
377 mutex_destroy(&h->hash_mutexes[i]);
378 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
379 kmem_cache_destroy(dbuf_cache);
380 taskq_destroy(dbu_evict_taskq);
389 dbuf_verify(dmu_buf_impl_t *db)
392 dbuf_dirty_record_t *dr;
394 ASSERT(MUTEX_HELD(&db->db_mtx));
396 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
399 ASSERT(db->db_objset != NULL);
403 ASSERT(db->db_parent == NULL);
404 ASSERT(db->db_blkptr == NULL);
406 ASSERT3U(db->db.db_object, ==, dn->dn_object);
407 ASSERT3P(db->db_objset, ==, dn->dn_objset);
408 ASSERT3U(db->db_level, <, dn->dn_nlevels);
409 ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
410 db->db_blkid == DMU_SPILL_BLKID ||
411 !avl_is_empty(&dn->dn_dbufs));
413 if (db->db_blkid == DMU_BONUS_BLKID) {
415 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
416 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
417 } else if (db->db_blkid == DMU_SPILL_BLKID) {
419 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
420 ASSERT0(db->db.db_offset);
422 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
425 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
426 ASSERT(dr->dr_dbuf == db);
428 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
429 ASSERT(dr->dr_dbuf == db);
432 * We can't assert that db_size matches dn_datablksz because it
433 * can be momentarily different when another thread is doing
436 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
437 dr = db->db_data_pending;
439 * It should only be modified in syncing context, so
440 * make sure we only have one copy of the data.
442 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
445 /* verify db->db_blkptr */
447 if (db->db_parent == dn->dn_dbuf) {
448 /* db is pointed to by the dnode */
449 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
450 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
451 ASSERT(db->db_parent == NULL);
453 ASSERT(db->db_parent != NULL);
454 if (db->db_blkid != DMU_SPILL_BLKID)
455 ASSERT3P(db->db_blkptr, ==,
456 &dn->dn_phys->dn_blkptr[db->db_blkid]);
458 /* db is pointed to by an indirect block */
459 int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
460 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
461 ASSERT3U(db->db_parent->db.db_object, ==,
464 * dnode_grow_indblksz() can make this fail if we don't
465 * have the struct_rwlock. XXX indblksz no longer
466 * grows. safe to do this now?
468 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
469 ASSERT3P(db->db_blkptr, ==,
470 ((blkptr_t *)db->db_parent->db.db_data +
471 db->db_blkid % epb));
475 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
476 (db->db_buf == NULL || db->db_buf->b_data) &&
477 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
478 db->db_state != DB_FILL && !dn->dn_free_txg) {
480 * If the blkptr isn't set but they have nonzero data,
481 * it had better be dirty, otherwise we'll lose that
482 * data when we evict this buffer.
484 if (db->db_dirtycnt == 0) {
485 uint64_t *buf = db->db.db_data;
488 for (i = 0; i < db->db.db_size >> 3; i++) {
498 dbuf_clear_data(dmu_buf_impl_t *db)
500 ASSERT(MUTEX_HELD(&db->db_mtx));
503 db->db.db_data = NULL;
504 if (db->db_state != DB_NOFILL)
505 db->db_state = DB_UNCACHED;
509 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
511 ASSERT(MUTEX_HELD(&db->db_mtx));
515 ASSERT(buf->b_data != NULL);
516 db->db.db_data = buf->b_data;
517 if (!arc_released(buf))
518 arc_set_callback(buf, dbuf_do_evict, db);
522 * Loan out an arc_buf for read. Return the loaned arc_buf.
525 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
529 mutex_enter(&db->db_mtx);
530 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
531 int blksz = db->db.db_size;
532 spa_t *spa = db->db_objset->os_spa;
534 mutex_exit(&db->db_mtx);
535 abuf = arc_loan_buf(spa, blksz);
536 bcopy(db->db.db_data, abuf->b_data, blksz);
539 arc_loan_inuse_buf(abuf, db);
541 mutex_exit(&db->db_mtx);
547 * Calculate which level n block references the data at the level 0 offset
551 dbuf_whichblock(dnode_t *dn, int64_t level, uint64_t offset)
553 if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) {
555 * The level n blkid is equal to the level 0 blkid divided by
556 * the number of level 0s in a level n block.
558 * The level 0 blkid is offset >> datablkshift =
559 * offset / 2^datablkshift.
561 * The number of level 0s in a level n is the number of block
562 * pointers in an indirect block, raised to the power of level.
563 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
564 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
566 * Thus, the level n blkid is: offset /
567 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
568 * = offset / 2^(datablkshift + level *
569 * (indblkshift - SPA_BLKPTRSHIFT))
570 * = offset >> (datablkshift + level *
571 * (indblkshift - SPA_BLKPTRSHIFT))
573 return (offset >> (dn->dn_datablkshift + level *
574 (dn->dn_indblkshift - SPA_BLKPTRSHIFT)));
576 ASSERT3U(offset, <, dn->dn_datablksz);
582 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
584 dmu_buf_impl_t *db = vdb;
586 mutex_enter(&db->db_mtx);
587 ASSERT3U(db->db_state, ==, DB_READ);
589 * All reads are synchronous, so we must have a hold on the dbuf
591 ASSERT(refcount_count(&db->db_holds) > 0);
592 ASSERT(db->db_buf == NULL);
593 ASSERT(db->db.db_data == NULL);
594 if (db->db_level == 0 && db->db_freed_in_flight) {
595 /* we were freed in flight; disregard any error */
596 arc_release(buf, db);
597 bzero(buf->b_data, db->db.db_size);
599 db->db_freed_in_flight = FALSE;
600 dbuf_set_data(db, buf);
601 db->db_state = DB_CACHED;
602 } else if (zio == NULL || zio->io_error == 0) {
603 dbuf_set_data(db, buf);
604 db->db_state = DB_CACHED;
606 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
607 ASSERT3P(db->db_buf, ==, NULL);
608 VERIFY(arc_buf_remove_ref(buf, db));
609 db->db_state = DB_UNCACHED;
611 cv_broadcast(&db->db_changed);
612 dbuf_rele_and_unlock(db, NULL);
616 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
620 arc_flags_t aflags = ARC_FLAG_NOWAIT;
624 ASSERT(!refcount_is_zero(&db->db_holds));
625 /* We need the struct_rwlock to prevent db_blkptr from changing. */
626 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
627 ASSERT(MUTEX_HELD(&db->db_mtx));
628 ASSERT(db->db_state == DB_UNCACHED);
629 ASSERT(db->db_buf == NULL);
631 if (db->db_blkid == DMU_BONUS_BLKID) {
632 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
634 ASSERT3U(bonuslen, <=, db->db.db_size);
635 db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN);
636 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
637 if (bonuslen < DN_MAX_BONUSLEN)
638 bzero(db->db.db_data, DN_MAX_BONUSLEN);
640 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
642 db->db_state = DB_CACHED;
643 mutex_exit(&db->db_mtx);
648 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
649 * processes the delete record and clears the bp while we are waiting
650 * for the dn_mtx (resulting in a "no" from block_freed).
652 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
653 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
654 BP_IS_HOLE(db->db_blkptr)))) {
655 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
658 dbuf_set_data(db, arc_buf_alloc(db->db_objset->os_spa,
659 db->db.db_size, db, type));
660 bzero(db->db.db_data, db->db.db_size);
661 db->db_state = DB_CACHED;
662 mutex_exit(&db->db_mtx);
668 db->db_state = DB_READ;
669 mutex_exit(&db->db_mtx);
671 if (DBUF_IS_L2CACHEABLE(db))
672 aflags |= ARC_FLAG_L2CACHE;
673 if (DBUF_IS_L2COMPRESSIBLE(db))
674 aflags |= ARC_FLAG_L2COMPRESS;
676 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
677 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
678 db->db.db_object, db->db_level, db->db_blkid);
680 dbuf_add_ref(db, NULL);
682 (void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr,
683 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
684 (flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
689 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
692 boolean_t havepzio = (zio != NULL);
697 * We don't have to hold the mutex to check db_state because it
698 * can't be freed while we have a hold on the buffer.
700 ASSERT(!refcount_is_zero(&db->db_holds));
702 if (db->db_state == DB_NOFILL)
703 return (SET_ERROR(EIO));
707 if ((flags & DB_RF_HAVESTRUCT) == 0)
708 rw_enter(&dn->dn_struct_rwlock, RW_READER);
710 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
711 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
712 DBUF_IS_CACHEABLE(db);
714 mutex_enter(&db->db_mtx);
715 if (db->db_state == DB_CACHED) {
716 mutex_exit(&db->db_mtx);
718 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1);
719 if ((flags & DB_RF_HAVESTRUCT) == 0)
720 rw_exit(&dn->dn_struct_rwlock);
722 } else if (db->db_state == DB_UNCACHED) {
723 spa_t *spa = dn->dn_objset->os_spa;
726 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
727 dbuf_read_impl(db, zio, flags);
729 /* dbuf_read_impl has dropped db_mtx for us */
732 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1);
734 if ((flags & DB_RF_HAVESTRUCT) == 0)
735 rw_exit(&dn->dn_struct_rwlock);
742 * Another reader came in while the dbuf was in flight
743 * between UNCACHED and CACHED. Either a writer will finish
744 * writing the buffer (sending the dbuf to CACHED) or the
745 * first reader's request will reach the read_done callback
746 * and send the dbuf to CACHED. Otherwise, a failure
747 * occurred and the dbuf went to UNCACHED.
749 mutex_exit(&db->db_mtx);
751 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1);
752 if ((flags & DB_RF_HAVESTRUCT) == 0)
753 rw_exit(&dn->dn_struct_rwlock);
756 /* Skip the wait per the caller's request. */
757 mutex_enter(&db->db_mtx);
758 if ((flags & DB_RF_NEVERWAIT) == 0) {
759 while (db->db_state == DB_READ ||
760 db->db_state == DB_FILL) {
761 ASSERT(db->db_state == DB_READ ||
762 (flags & DB_RF_HAVESTRUCT) == 0);
763 DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *,
765 cv_wait(&db->db_changed, &db->db_mtx);
767 if (db->db_state == DB_UNCACHED)
768 err = SET_ERROR(EIO);
770 mutex_exit(&db->db_mtx);
773 ASSERT(err || havepzio || db->db_state == DB_CACHED);
778 dbuf_noread(dmu_buf_impl_t *db)
780 ASSERT(!refcount_is_zero(&db->db_holds));
781 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
782 mutex_enter(&db->db_mtx);
783 while (db->db_state == DB_READ || db->db_state == DB_FILL)
784 cv_wait(&db->db_changed, &db->db_mtx);
785 if (db->db_state == DB_UNCACHED) {
786 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
787 spa_t *spa = db->db_objset->os_spa;
789 ASSERT(db->db_buf == NULL);
790 ASSERT(db->db.db_data == NULL);
791 dbuf_set_data(db, arc_buf_alloc(spa, db->db.db_size, db, type));
792 db->db_state = DB_FILL;
793 } else if (db->db_state == DB_NOFILL) {
796 ASSERT3U(db->db_state, ==, DB_CACHED);
798 mutex_exit(&db->db_mtx);
802 * This is our just-in-time copy function. It makes a copy of
803 * buffers, that have been modified in a previous transaction
804 * group, before we modify them in the current active group.
806 * This function is used in two places: when we are dirtying a
807 * buffer for the first time in a txg, and when we are freeing
808 * a range in a dnode that includes this buffer.
810 * Note that when we are called from dbuf_free_range() we do
811 * not put a hold on the buffer, we just traverse the active
812 * dbuf list for the dnode.
815 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
817 dbuf_dirty_record_t *dr = db->db_last_dirty;
819 ASSERT(MUTEX_HELD(&db->db_mtx));
820 ASSERT(db->db.db_data != NULL);
821 ASSERT(db->db_level == 0);
822 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
825 (dr->dt.dl.dr_data !=
826 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
830 * If the last dirty record for this dbuf has not yet synced
831 * and its referencing the dbuf data, either:
832 * reset the reference to point to a new copy,
833 * or (if there a no active holders)
834 * just null out the current db_data pointer.
836 ASSERT(dr->dr_txg >= txg - 2);
837 if (db->db_blkid == DMU_BONUS_BLKID) {
838 /* Note that the data bufs here are zio_bufs */
839 dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
840 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
841 bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
842 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
843 int size = db->db.db_size;
844 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
845 spa_t *spa = db->db_objset->os_spa;
847 dr->dt.dl.dr_data = arc_buf_alloc(spa, size, db, type);
848 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
855 dbuf_unoverride(dbuf_dirty_record_t *dr)
857 dmu_buf_impl_t *db = dr->dr_dbuf;
858 blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
859 uint64_t txg = dr->dr_txg;
861 ASSERT(MUTEX_HELD(&db->db_mtx));
862 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
863 ASSERT(db->db_level == 0);
865 if (db->db_blkid == DMU_BONUS_BLKID ||
866 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
869 ASSERT(db->db_data_pending != dr);
871 /* free this block */
872 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite)
873 zio_free(db->db_objset->os_spa, txg, bp);
875 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
876 dr->dt.dl.dr_nopwrite = B_FALSE;
879 * Release the already-written buffer, so we leave it in
880 * a consistent dirty state. Note that all callers are
881 * modifying the buffer, so they will immediately do
882 * another (redundant) arc_release(). Therefore, leave
883 * the buf thawed to save the effort of freezing &
884 * immediately re-thawing it.
886 arc_release(dr->dt.dl.dr_data, db);
890 * Evict (if its unreferenced) or clear (if its referenced) any level-0
891 * data blocks in the free range, so that any future readers will find
894 * This is a no-op if the dataset is in the middle of an incremental
895 * receive; see comment below for details.
898 dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
901 dmu_buf_impl_t db_search;
902 dmu_buf_impl_t *db, *db_next;
903 uint64_t txg = tx->tx_txg;
906 if (end_blkid > dn->dn_maxblkid && (end_blkid != DMU_SPILL_BLKID))
907 end_blkid = dn->dn_maxblkid;
908 dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid);
910 db_search.db_level = 0;
911 db_search.db_blkid = start_blkid;
912 db_search.db_state = DB_SEARCH;
914 mutex_enter(&dn->dn_dbufs_mtx);
915 if (start_blkid >= dn->dn_unlisted_l0_blkid) {
916 /* There can't be any dbufs in this range; no need to search. */
918 db = avl_find(&dn->dn_dbufs, &db_search, &where);
919 ASSERT3P(db, ==, NULL);
920 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
921 ASSERT(db == NULL || db->db_level > 0);
923 mutex_exit(&dn->dn_dbufs_mtx);
925 } else if (dmu_objset_is_receiving(dn->dn_objset)) {
927 * If we are receiving, we expect there to be no dbufs in
928 * the range to be freed, because receive modifies each
929 * block at most once, and in offset order. If this is
930 * not the case, it can lead to performance problems,
931 * so note that we unexpectedly took the slow path.
933 atomic_inc_64(&zfs_free_range_recv_miss);
936 db = avl_find(&dn->dn_dbufs, &db_search, &where);
937 ASSERT3P(db, ==, NULL);
938 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
940 for (; db != NULL; db = db_next) {
941 db_next = AVL_NEXT(&dn->dn_dbufs, db);
942 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
944 if (db->db_level != 0 || db->db_blkid > end_blkid) {
947 ASSERT3U(db->db_blkid, >=, start_blkid);
949 /* found a level 0 buffer in the range */
950 mutex_enter(&db->db_mtx);
951 if (dbuf_undirty(db, tx)) {
952 /* mutex has been dropped and dbuf destroyed */
956 if (db->db_state == DB_UNCACHED ||
957 db->db_state == DB_NOFILL ||
958 db->db_state == DB_EVICTING) {
959 ASSERT(db->db.db_data == NULL);
960 mutex_exit(&db->db_mtx);
963 if (db->db_state == DB_READ || db->db_state == DB_FILL) {
964 /* will be handled in dbuf_read_done or dbuf_rele */
965 db->db_freed_in_flight = TRUE;
966 mutex_exit(&db->db_mtx);
969 if (refcount_count(&db->db_holds) == 0) {
974 /* The dbuf is referenced */
976 if (db->db_last_dirty != NULL) {
977 dbuf_dirty_record_t *dr = db->db_last_dirty;
979 if (dr->dr_txg == txg) {
981 * This buffer is "in-use", re-adjust the file
982 * size to reflect that this buffer may
983 * contain new data when we sync.
985 if (db->db_blkid != DMU_SPILL_BLKID &&
986 db->db_blkid > dn->dn_maxblkid)
987 dn->dn_maxblkid = db->db_blkid;
991 * This dbuf is not dirty in the open context.
992 * Either uncache it (if its not referenced in
993 * the open context) or reset its contents to
996 dbuf_fix_old_data(db, txg);
999 /* clear the contents if its cached */
1000 if (db->db_state == DB_CACHED) {
1001 ASSERT(db->db.db_data != NULL);
1002 arc_release(db->db_buf, db);
1003 bzero(db->db.db_data, db->db.db_size);
1004 arc_buf_freeze(db->db_buf);
1007 mutex_exit(&db->db_mtx);
1009 mutex_exit(&dn->dn_dbufs_mtx);
1013 dbuf_block_freeable(dmu_buf_impl_t *db)
1015 dsl_dataset_t *ds = db->db_objset->os_dsl_dataset;
1016 uint64_t birth_txg = 0;
1019 * We don't need any locking to protect db_blkptr:
1020 * If it's syncing, then db_last_dirty will be set
1021 * so we'll ignore db_blkptr.
1023 * This logic ensures that only block births for
1024 * filled blocks are considered.
1026 ASSERT(MUTEX_HELD(&db->db_mtx));
1027 if (db->db_last_dirty && (db->db_blkptr == NULL ||
1028 !BP_IS_HOLE(db->db_blkptr))) {
1029 birth_txg = db->db_last_dirty->dr_txg;
1030 } else if (db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) {
1031 birth_txg = db->db_blkptr->blk_birth;
1035 * If this block don't exist or is in a snapshot, it can't be freed.
1036 * Don't pass the bp to dsl_dataset_block_freeable() since we
1037 * are holding the db_mtx lock and might deadlock if we are
1038 * prefetching a dedup-ed block.
1041 return (ds == NULL ||
1042 dsl_dataset_block_freeable(ds, NULL, birth_txg));
1048 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
1050 arc_buf_t *buf, *obuf;
1051 int osize = db->db.db_size;
1052 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1055 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1060 /* XXX does *this* func really need the lock? */
1061 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1064 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1065 * is OK, because there can be no other references to the db
1066 * when we are changing its size, so no concurrent DB_FILL can
1070 * XXX we should be doing a dbuf_read, checking the return
1071 * value and returning that up to our callers
1073 dmu_buf_will_dirty(&db->db, tx);
1075 /* create the data buffer for the new block */
1076 buf = arc_buf_alloc(dn->dn_objset->os_spa, size, db, type);
1078 /* copy old block data to the new block */
1080 bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
1081 /* zero the remainder */
1083 bzero((uint8_t *)buf->b_data + osize, size - osize);
1085 mutex_enter(&db->db_mtx);
1086 dbuf_set_data(db, buf);
1087 VERIFY(arc_buf_remove_ref(obuf, db));
1088 db->db.db_size = size;
1090 if (db->db_level == 0) {
1091 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1092 db->db_last_dirty->dt.dl.dr_data = buf;
1094 mutex_exit(&db->db_mtx);
1096 dnode_willuse_space(dn, size-osize, tx);
1101 dbuf_release_bp(dmu_buf_impl_t *db)
1103 objset_t *os = db->db_objset;
1105 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1106 ASSERT(arc_released(os->os_phys_buf) ||
1107 list_link_active(&os->os_dsl_dataset->ds_synced_link));
1108 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1110 (void) arc_release(db->db_buf, db);
1114 * We already have a dirty record for this TXG, and we are being
1118 dbuf_redirty(dbuf_dirty_record_t *dr)
1120 dmu_buf_impl_t *db = dr->dr_dbuf;
1122 ASSERT(MUTEX_HELD(&db->db_mtx));
1124 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1126 * If this buffer has already been written out,
1127 * we now need to reset its state.
1129 dbuf_unoverride(dr);
1130 if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1131 db->db_state != DB_NOFILL) {
1132 /* Already released on initial dirty, so just thaw. */
1133 ASSERT(arc_released(db->db_buf));
1134 arc_buf_thaw(db->db_buf);
1139 dbuf_dirty_record_t *
1140 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1144 dbuf_dirty_record_t **drp, *dr;
1145 int drop_struct_lock = FALSE;
1146 boolean_t do_free_accounting = B_FALSE;
1147 int txgoff = tx->tx_txg & TXG_MASK;
1149 ASSERT(tx->tx_txg != 0);
1150 ASSERT(!refcount_is_zero(&db->db_holds));
1151 DMU_TX_DIRTY_BUF(tx, db);
1156 * Shouldn't dirty a regular buffer in syncing context. Private
1157 * objects may be dirtied in syncing context, but only if they
1158 * were already pre-dirtied in open context.
1160 ASSERT(!dmu_tx_is_syncing(tx) ||
1161 BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1162 DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1163 dn->dn_objset->os_dsl_dataset == NULL);
1165 * We make this assert for private objects as well, but after we
1166 * check if we're already dirty. They are allowed to re-dirty
1167 * in syncing context.
1169 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1170 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1171 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1173 mutex_enter(&db->db_mtx);
1175 * XXX make this true for indirects too? The problem is that
1176 * transactions created with dmu_tx_create_assigned() from
1177 * syncing context don't bother holding ahead.
1179 ASSERT(db->db_level != 0 ||
1180 db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1181 db->db_state == DB_NOFILL);
1183 mutex_enter(&dn->dn_mtx);
1185 * Don't set dirtyctx to SYNC if we're just modifying this as we
1186 * initialize the objset.
1188 if (dn->dn_dirtyctx == DN_UNDIRTIED &&
1189 !BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1191 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1192 ASSERT(dn->dn_dirtyctx_firstset == NULL);
1193 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1195 mutex_exit(&dn->dn_mtx);
1197 if (db->db_blkid == DMU_SPILL_BLKID)
1198 dn->dn_have_spill = B_TRUE;
1201 * If this buffer is already dirty, we're done.
1203 drp = &db->db_last_dirty;
1204 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1205 db->db.db_object == DMU_META_DNODE_OBJECT);
1206 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1208 if (dr && dr->dr_txg == tx->tx_txg) {
1212 mutex_exit(&db->db_mtx);
1217 * Only valid if not already dirty.
1219 ASSERT(dn->dn_object == 0 ||
1220 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1221 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1223 ASSERT3U(dn->dn_nlevels, >, db->db_level);
1224 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1225 dn->dn_phys->dn_nlevels > db->db_level ||
1226 dn->dn_next_nlevels[txgoff] > db->db_level ||
1227 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1228 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1231 * We should only be dirtying in syncing context if it's the
1232 * mos or we're initializing the os or it's a special object.
1233 * However, we are allowed to dirty in syncing context provided
1234 * we already dirtied it in open context. Hence we must make
1235 * this assertion only if we're not already dirty.
1238 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1239 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1240 ASSERT(db->db.db_size != 0);
1242 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1244 if (db->db_blkid != DMU_BONUS_BLKID) {
1246 * Update the accounting.
1247 * Note: we delay "free accounting" until after we drop
1248 * the db_mtx. This keeps us from grabbing other locks
1249 * (and possibly deadlocking) in bp_get_dsize() while
1250 * also holding the db_mtx.
1252 dnode_willuse_space(dn, db->db.db_size, tx);
1253 do_free_accounting = dbuf_block_freeable(db);
1257 * If this buffer is dirty in an old transaction group we need
1258 * to make a copy of it so that the changes we make in this
1259 * transaction group won't leak out when we sync the older txg.
1261 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1262 if (db->db_level == 0) {
1263 void *data_old = db->db_buf;
1265 if (db->db_state != DB_NOFILL) {
1266 if (db->db_blkid == DMU_BONUS_BLKID) {
1267 dbuf_fix_old_data(db, tx->tx_txg);
1268 data_old = db->db.db_data;
1269 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1271 * Release the data buffer from the cache so
1272 * that we can modify it without impacting
1273 * possible other users of this cached data
1274 * block. Note that indirect blocks and
1275 * private objects are not released until the
1276 * syncing state (since they are only modified
1279 arc_release(db->db_buf, db);
1280 dbuf_fix_old_data(db, tx->tx_txg);
1281 data_old = db->db_buf;
1283 ASSERT(data_old != NULL);
1285 dr->dt.dl.dr_data = data_old;
1287 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1288 list_create(&dr->dt.di.dr_children,
1289 sizeof (dbuf_dirty_record_t),
1290 offsetof(dbuf_dirty_record_t, dr_dirty_node));
1292 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL)
1293 dr->dr_accounted = db->db.db_size;
1295 dr->dr_txg = tx->tx_txg;
1300 * We could have been freed_in_flight between the dbuf_noread
1301 * and dbuf_dirty. We win, as though the dbuf_noread() had
1302 * happened after the free.
1304 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1305 db->db_blkid != DMU_SPILL_BLKID) {
1306 mutex_enter(&dn->dn_mtx);
1307 if (dn->dn_free_ranges[txgoff] != NULL) {
1308 range_tree_clear(dn->dn_free_ranges[txgoff],
1311 mutex_exit(&dn->dn_mtx);
1312 db->db_freed_in_flight = FALSE;
1316 * This buffer is now part of this txg
1318 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1319 db->db_dirtycnt += 1;
1320 ASSERT3U(db->db_dirtycnt, <=, 3);
1322 mutex_exit(&db->db_mtx);
1324 if (db->db_blkid == DMU_BONUS_BLKID ||
1325 db->db_blkid == DMU_SPILL_BLKID) {
1326 mutex_enter(&dn->dn_mtx);
1327 ASSERT(!list_link_active(&dr->dr_dirty_node));
1328 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1329 mutex_exit(&dn->dn_mtx);
1330 dnode_setdirty(dn, tx);
1333 } else if (do_free_accounting) {
1334 blkptr_t *bp = db->db_blkptr;
1335 int64_t willfree = (bp && !BP_IS_HOLE(bp)) ?
1336 bp_get_dsize(os->os_spa, bp) : db->db.db_size;
1338 * This is only a guess -- if the dbuf is dirty
1339 * in a previous txg, we don't know how much
1340 * space it will use on disk yet. We should
1341 * really have the struct_rwlock to access
1342 * db_blkptr, but since this is just a guess,
1343 * it's OK if we get an odd answer.
1345 ddt_prefetch(os->os_spa, bp);
1346 dnode_willuse_space(dn, -willfree, tx);
1349 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1350 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1351 drop_struct_lock = TRUE;
1354 if (db->db_level == 0) {
1355 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1356 ASSERT(dn->dn_maxblkid >= db->db_blkid);
1359 if (db->db_level+1 < dn->dn_nlevels) {
1360 dmu_buf_impl_t *parent = db->db_parent;
1361 dbuf_dirty_record_t *di;
1362 int parent_held = FALSE;
1364 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1365 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1367 parent = dbuf_hold_level(dn, db->db_level+1,
1368 db->db_blkid >> epbs, FTAG);
1369 ASSERT(parent != NULL);
1372 if (drop_struct_lock)
1373 rw_exit(&dn->dn_struct_rwlock);
1374 ASSERT3U(db->db_level+1, ==, parent->db_level);
1375 di = dbuf_dirty(parent, tx);
1377 dbuf_rele(parent, FTAG);
1379 mutex_enter(&db->db_mtx);
1381 * Since we've dropped the mutex, it's possible that
1382 * dbuf_undirty() might have changed this out from under us.
1384 if (db->db_last_dirty == dr ||
1385 dn->dn_object == DMU_META_DNODE_OBJECT) {
1386 mutex_enter(&di->dt.di.dr_mtx);
1387 ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1388 ASSERT(!list_link_active(&dr->dr_dirty_node));
1389 list_insert_tail(&di->dt.di.dr_children, dr);
1390 mutex_exit(&di->dt.di.dr_mtx);
1393 mutex_exit(&db->db_mtx);
1395 ASSERT(db->db_level+1 == dn->dn_nlevels);
1396 ASSERT(db->db_blkid < dn->dn_nblkptr);
1397 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1398 mutex_enter(&dn->dn_mtx);
1399 ASSERT(!list_link_active(&dr->dr_dirty_node));
1400 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1401 mutex_exit(&dn->dn_mtx);
1402 if (drop_struct_lock)
1403 rw_exit(&dn->dn_struct_rwlock);
1406 dnode_setdirty(dn, tx);
1412 * Undirty a buffer in the transaction group referenced by the given
1413 * transaction. Return whether this evicted the dbuf.
1416 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1419 uint64_t txg = tx->tx_txg;
1420 dbuf_dirty_record_t *dr, **drp;
1425 * Due to our use of dn_nlevels below, this can only be called
1426 * in open context, unless we are operating on the MOS.
1427 * From syncing context, dn_nlevels may be different from the
1428 * dn_nlevels used when dbuf was dirtied.
1430 ASSERT(db->db_objset ==
1431 dmu_objset_pool(db->db_objset)->dp_meta_objset ||
1432 txg != spa_syncing_txg(dmu_objset_spa(db->db_objset)));
1433 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1434 ASSERT0(db->db_level);
1435 ASSERT(MUTEX_HELD(&db->db_mtx));
1438 * If this buffer is not dirty, we're done.
1440 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1441 if (dr->dr_txg <= txg)
1443 if (dr == NULL || dr->dr_txg < txg)
1445 ASSERT(dr->dr_txg == txg);
1446 ASSERT(dr->dr_dbuf == db);
1451 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1453 ASSERT(db->db.db_size != 0);
1455 dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset),
1456 dr->dr_accounted, txg);
1461 * Note that there are three places in dbuf_dirty()
1462 * where this dirty record may be put on a list.
1463 * Make sure to do a list_remove corresponding to
1464 * every one of those list_insert calls.
1466 if (dr->dr_parent) {
1467 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1468 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1469 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1470 } else if (db->db_blkid == DMU_SPILL_BLKID ||
1471 db->db_level + 1 == dn->dn_nlevels) {
1472 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1473 mutex_enter(&dn->dn_mtx);
1474 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1475 mutex_exit(&dn->dn_mtx);
1479 if (db->db_state != DB_NOFILL) {
1480 dbuf_unoverride(dr);
1482 ASSERT(db->db_buf != NULL);
1483 ASSERT(dr->dt.dl.dr_data != NULL);
1484 if (dr->dt.dl.dr_data != db->db_buf)
1485 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db));
1488 kmem_free(dr, sizeof (dbuf_dirty_record_t));
1490 ASSERT(db->db_dirtycnt > 0);
1491 db->db_dirtycnt -= 1;
1493 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1494 arc_buf_t *buf = db->db_buf;
1496 ASSERT(db->db_state == DB_NOFILL || arc_released(buf));
1497 dbuf_clear_data(db);
1498 VERIFY(arc_buf_remove_ref(buf, db));
1507 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
1509 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1510 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1512 ASSERT(tx->tx_txg != 0);
1513 ASSERT(!refcount_is_zero(&db->db_holds));
1516 * Quick check for dirtyness. For already dirty blocks, this
1517 * reduces runtime of this function by >90%, and overall performance
1518 * by 50% for some workloads (e.g. file deletion with indirect blocks
1521 mutex_enter(&db->db_mtx);
1522 dbuf_dirty_record_t *dr;
1523 for (dr = db->db_last_dirty;
1524 dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) {
1526 * It's possible that it is already dirty but not cached,
1527 * because there are some calls to dbuf_dirty() that don't
1528 * go through dmu_buf_will_dirty().
1530 if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) {
1531 /* This dbuf is already dirty and cached. */
1533 mutex_exit(&db->db_mtx);
1537 mutex_exit(&db->db_mtx);
1540 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
1541 rf |= DB_RF_HAVESTRUCT;
1543 (void) dbuf_read(db, NULL, rf);
1544 (void) dbuf_dirty(db, tx);
1548 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1550 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1552 db->db_state = DB_NOFILL;
1554 dmu_buf_will_fill(db_fake, tx);
1558 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1560 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1562 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1563 ASSERT(tx->tx_txg != 0);
1564 ASSERT(db->db_level == 0);
1565 ASSERT(!refcount_is_zero(&db->db_holds));
1567 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1568 dmu_tx_private_ok(tx));
1571 (void) dbuf_dirty(db, tx);
1574 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1577 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1579 mutex_enter(&db->db_mtx);
1582 if (db->db_state == DB_FILL) {
1583 if (db->db_level == 0 && db->db_freed_in_flight) {
1584 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1585 /* we were freed while filling */
1586 /* XXX dbuf_undirty? */
1587 bzero(db->db.db_data, db->db.db_size);
1588 db->db_freed_in_flight = FALSE;
1590 db->db_state = DB_CACHED;
1591 cv_broadcast(&db->db_changed);
1593 mutex_exit(&db->db_mtx);
1597 dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data,
1598 bp_embedded_type_t etype, enum zio_compress comp,
1599 int uncompressed_size, int compressed_size, int byteorder,
1602 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
1603 struct dirty_leaf *dl;
1604 dmu_object_type_t type;
1606 if (etype == BP_EMBEDDED_TYPE_DATA) {
1607 ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset),
1608 SPA_FEATURE_EMBEDDED_DATA));
1612 type = DB_DNODE(db)->dn_type;
1615 ASSERT0(db->db_level);
1616 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1618 dmu_buf_will_not_fill(dbuf, tx);
1620 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1621 dl = &db->db_last_dirty->dt.dl;
1622 encode_embedded_bp_compressed(&dl->dr_overridden_by,
1623 data, comp, uncompressed_size, compressed_size);
1624 BPE_SET_ETYPE(&dl->dr_overridden_by, etype);
1625 BP_SET_TYPE(&dl->dr_overridden_by, type);
1626 BP_SET_LEVEL(&dl->dr_overridden_by, 0);
1627 BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder);
1629 dl->dr_override_state = DR_OVERRIDDEN;
1630 dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg;
1634 * Directly assign a provided arc buf to a given dbuf if it's not referenced
1635 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1638 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
1640 ASSERT(!refcount_is_zero(&db->db_holds));
1641 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1642 ASSERT(db->db_level == 0);
1643 ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA);
1644 ASSERT(buf != NULL);
1645 ASSERT(arc_buf_size(buf) == db->db.db_size);
1646 ASSERT(tx->tx_txg != 0);
1648 arc_return_buf(buf, db);
1649 ASSERT(arc_released(buf));
1651 mutex_enter(&db->db_mtx);
1653 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1654 cv_wait(&db->db_changed, &db->db_mtx);
1656 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
1658 if (db->db_state == DB_CACHED &&
1659 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
1660 mutex_exit(&db->db_mtx);
1661 (void) dbuf_dirty(db, tx);
1662 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
1663 VERIFY(arc_buf_remove_ref(buf, db));
1664 xuio_stat_wbuf_copied();
1668 xuio_stat_wbuf_nocopy();
1669 if (db->db_state == DB_CACHED) {
1670 dbuf_dirty_record_t *dr = db->db_last_dirty;
1672 ASSERT(db->db_buf != NULL);
1673 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
1674 ASSERT(dr->dt.dl.dr_data == db->db_buf);
1675 if (!arc_released(db->db_buf)) {
1676 ASSERT(dr->dt.dl.dr_override_state ==
1678 arc_release(db->db_buf, db);
1680 dr->dt.dl.dr_data = buf;
1681 VERIFY(arc_buf_remove_ref(db->db_buf, db));
1682 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
1683 arc_release(db->db_buf, db);
1684 VERIFY(arc_buf_remove_ref(db->db_buf, db));
1688 ASSERT(db->db_buf == NULL);
1689 dbuf_set_data(db, buf);
1690 db->db_state = DB_FILL;
1691 mutex_exit(&db->db_mtx);
1692 (void) dbuf_dirty(db, tx);
1693 dmu_buf_fill_done(&db->db, tx);
1697 * "Clear" the contents of this dbuf. This will mark the dbuf
1698 * EVICTING and clear *most* of its references. Unfortunately,
1699 * when we are not holding the dn_dbufs_mtx, we can't clear the
1700 * entry in the dn_dbufs list. We have to wait until dbuf_destroy()
1701 * in this case. For callers from the DMU we will usually see:
1702 * dbuf_clear()->arc_clear_callback()->dbuf_do_evict()->dbuf_destroy()
1703 * For the arc callback, we will usually see:
1704 * dbuf_do_evict()->dbuf_clear();dbuf_destroy()
1705 * Sometimes, though, we will get a mix of these two:
1706 * DMU: dbuf_clear()->arc_clear_callback()
1707 * ARC: dbuf_do_evict()->dbuf_destroy()
1709 * This routine will dissociate the dbuf from the arc, by calling
1710 * arc_clear_callback(), but will not evict the data from the ARC.
1713 dbuf_clear(dmu_buf_impl_t *db)
1716 dmu_buf_impl_t *parent = db->db_parent;
1717 dmu_buf_impl_t *dndb;
1718 boolean_t dbuf_gone = B_FALSE;
1720 ASSERT(MUTEX_HELD(&db->db_mtx));
1721 ASSERT(refcount_is_zero(&db->db_holds));
1723 dbuf_evict_user(db);
1725 if (db->db_state == DB_CACHED) {
1726 ASSERT(db->db.db_data != NULL);
1727 if (db->db_blkid == DMU_BONUS_BLKID) {
1728 zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
1729 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
1731 db->db.db_data = NULL;
1732 db->db_state = DB_UNCACHED;
1735 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
1736 ASSERT(db->db_data_pending == NULL);
1738 db->db_state = DB_EVICTING;
1739 db->db_blkptr = NULL;
1744 if (db->db_blkid != DMU_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) {
1745 avl_remove(&dn->dn_dbufs, db);
1746 atomic_dec_32(&dn->dn_dbufs_count);
1750 * Decrementing the dbuf count means that the hold corresponding
1751 * to the removed dbuf is no longer discounted in dnode_move(),
1752 * so the dnode cannot be moved until after we release the hold.
1753 * The membar_producer() ensures visibility of the decremented
1754 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
1758 db->db_dnode_handle = NULL;
1764 dbuf_gone = arc_clear_callback(db->db_buf);
1767 mutex_exit(&db->db_mtx);
1770 * If this dbuf is referenced from an indirect dbuf,
1771 * decrement the ref count on the indirect dbuf.
1773 if (parent && parent != dndb)
1774 dbuf_rele(parent, db);
1778 * Note: While bpp will always be updated if the function returns success,
1779 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
1780 * this happens when the dnode is the meta-dnode, or a userused or groupused
1784 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
1785 dmu_buf_impl_t **parentp, blkptr_t **bpp)
1792 ASSERT(blkid != DMU_BONUS_BLKID);
1794 if (blkid == DMU_SPILL_BLKID) {
1795 mutex_enter(&dn->dn_mtx);
1796 if (dn->dn_have_spill &&
1797 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
1798 *bpp = &dn->dn_phys->dn_spill;
1801 dbuf_add_ref(dn->dn_dbuf, NULL);
1802 *parentp = dn->dn_dbuf;
1803 mutex_exit(&dn->dn_mtx);
1807 if (dn->dn_phys->dn_nlevels == 0)
1810 nlevels = dn->dn_phys->dn_nlevels;
1812 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1814 ASSERT3U(level * epbs, <, 64);
1815 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1816 if (level >= nlevels ||
1817 (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
1818 /* the buffer has no parent yet */
1819 return (SET_ERROR(ENOENT));
1820 } else if (level < nlevels-1) {
1821 /* this block is referenced from an indirect block */
1822 int err = dbuf_hold_impl(dn, level+1,
1823 blkid >> epbs, fail_sparse, FALSE, NULL, parentp);
1826 err = dbuf_read(*parentp, NULL,
1827 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
1829 dbuf_rele(*parentp, NULL);
1833 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
1834 (blkid & ((1ULL << epbs) - 1));
1837 /* the block is referenced from the dnode */
1838 ASSERT3U(level, ==, nlevels-1);
1839 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
1840 blkid < dn->dn_phys->dn_nblkptr);
1842 dbuf_add_ref(dn->dn_dbuf, NULL);
1843 *parentp = dn->dn_dbuf;
1845 *bpp = &dn->dn_phys->dn_blkptr[blkid];
1850 static dmu_buf_impl_t *
1851 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
1852 dmu_buf_impl_t *parent, blkptr_t *blkptr)
1854 objset_t *os = dn->dn_objset;
1855 dmu_buf_impl_t *db, *odb;
1857 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1858 ASSERT(dn->dn_type != DMU_OT_NONE);
1860 db = kmem_cache_alloc(dbuf_cache, KM_SLEEP);
1863 db->db.db_object = dn->dn_object;
1864 db->db_level = level;
1865 db->db_blkid = blkid;
1866 db->db_last_dirty = NULL;
1867 db->db_dirtycnt = 0;
1868 db->db_dnode_handle = dn->dn_handle;
1869 db->db_parent = parent;
1870 db->db_blkptr = blkptr;
1873 db->db_user_immediate_evict = FALSE;
1874 db->db_freed_in_flight = FALSE;
1875 db->db_pending_evict = FALSE;
1877 if (blkid == DMU_BONUS_BLKID) {
1878 ASSERT3P(parent, ==, dn->dn_dbuf);
1879 db->db.db_size = DN_MAX_BONUSLEN -
1880 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
1881 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
1882 db->db.db_offset = DMU_BONUS_BLKID;
1883 db->db_state = DB_UNCACHED;
1884 /* the bonus dbuf is not placed in the hash table */
1885 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1887 } else if (blkid == DMU_SPILL_BLKID) {
1888 db->db.db_size = (blkptr != NULL) ?
1889 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
1890 db->db.db_offset = 0;
1893 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
1894 db->db.db_size = blocksize;
1895 db->db.db_offset = db->db_blkid * blocksize;
1899 * Hold the dn_dbufs_mtx while we get the new dbuf
1900 * in the hash table *and* added to the dbufs list.
1901 * This prevents a possible deadlock with someone
1902 * trying to look up this dbuf before its added to the
1905 mutex_enter(&dn->dn_dbufs_mtx);
1906 db->db_state = DB_EVICTING;
1907 if ((odb = dbuf_hash_insert(db)) != NULL) {
1908 /* someone else inserted it first */
1909 kmem_cache_free(dbuf_cache, db);
1910 mutex_exit(&dn->dn_dbufs_mtx);
1913 avl_add(&dn->dn_dbufs, db);
1914 if (db->db_level == 0 && db->db_blkid >=
1915 dn->dn_unlisted_l0_blkid)
1916 dn->dn_unlisted_l0_blkid = db->db_blkid + 1;
1917 db->db_state = DB_UNCACHED;
1918 mutex_exit(&dn->dn_dbufs_mtx);
1919 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1921 if (parent && parent != dn->dn_dbuf)
1922 dbuf_add_ref(parent, db);
1924 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1925 refcount_count(&dn->dn_holds) > 0);
1926 (void) refcount_add(&dn->dn_holds, db);
1927 atomic_inc_32(&dn->dn_dbufs_count);
1929 dprintf_dbuf(db, "db=%p\n", db);
1935 dbuf_do_evict(void *private)
1937 dmu_buf_impl_t *db = private;
1939 if (!MUTEX_HELD(&db->db_mtx))
1940 mutex_enter(&db->db_mtx);
1942 ASSERT(refcount_is_zero(&db->db_holds));
1944 if (db->db_state != DB_EVICTING) {
1945 ASSERT(db->db_state == DB_CACHED);
1950 mutex_exit(&db->db_mtx);
1957 dbuf_destroy(dmu_buf_impl_t *db)
1959 ASSERT(refcount_is_zero(&db->db_holds));
1961 if (db->db_blkid != DMU_BONUS_BLKID) {
1963 * If this dbuf is still on the dn_dbufs list,
1964 * remove it from that list.
1966 if (db->db_dnode_handle != NULL) {
1971 mutex_enter(&dn->dn_dbufs_mtx);
1972 avl_remove(&dn->dn_dbufs, db);
1973 atomic_dec_32(&dn->dn_dbufs_count);
1974 mutex_exit(&dn->dn_dbufs_mtx);
1977 * Decrementing the dbuf count means that the hold
1978 * corresponding to the removed dbuf is no longer
1979 * discounted in dnode_move(), so the dnode cannot be
1980 * moved until after we release the hold.
1983 db->db_dnode_handle = NULL;
1985 dbuf_hash_remove(db);
1987 db->db_parent = NULL;
1990 ASSERT(db->db.db_data == NULL);
1991 ASSERT(db->db_hash_next == NULL);
1992 ASSERT(db->db_blkptr == NULL);
1993 ASSERT(db->db_data_pending == NULL);
1995 kmem_cache_free(dbuf_cache, db);
1996 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1999 typedef struct dbuf_prefetch_arg {
2000 spa_t *dpa_spa; /* The spa to issue the prefetch in. */
2001 zbookmark_phys_t dpa_zb; /* The target block to prefetch. */
2002 int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */
2003 int dpa_curlevel; /* The current level that we're reading */
2004 zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */
2005 zio_t *dpa_zio; /* The parent zio_t for all prefetches. */
2006 arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */
2007 } dbuf_prefetch_arg_t;
2010 * Actually issue the prefetch read for the block given.
2013 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp)
2015 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
2018 arc_flags_t aflags =
2019 dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
2021 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2022 ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level);
2023 ASSERT(dpa->dpa_zio != NULL);
2024 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL,
2025 dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2026 &aflags, &dpa->dpa_zb);
2030 * Called when an indirect block above our prefetch target is read in. This
2031 * will either read in the next indirect block down the tree or issue the actual
2032 * prefetch if the next block down is our target.
2035 dbuf_prefetch_indirect_done(zio_t *zio, arc_buf_t *abuf, void *private)
2037 dbuf_prefetch_arg_t *dpa = private;
2039 ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel);
2040 ASSERT3S(dpa->dpa_curlevel, >, 0);
2042 ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel);
2043 ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size);
2044 ASSERT3P(zio->io_spa, ==, dpa->dpa_spa);
2047 dpa->dpa_curlevel--;
2049 uint64_t nextblkid = dpa->dpa_zb.zb_blkid >>
2050 (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level));
2051 blkptr_t *bp = ((blkptr_t *)abuf->b_data) +
2052 P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs);
2053 if (BP_IS_HOLE(bp) || (zio != NULL && zio->io_error != 0)) {
2054 kmem_free(dpa, sizeof (*dpa));
2055 } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) {
2056 ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid);
2057 dbuf_issue_final_prefetch(dpa, bp);
2058 kmem_free(dpa, sizeof (*dpa));
2060 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2061 zbookmark_phys_t zb;
2063 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2065 SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset,
2066 dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid);
2068 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2069 bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio,
2070 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2073 (void) arc_buf_remove_ref(abuf, private);
2077 * Issue prefetch reads for the given block on the given level. If the indirect
2078 * blocks above that block are not in memory, we will read them in
2079 * asynchronously. As a result, this call never blocks waiting for a read to
2083 dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio,
2087 int epbs, nlevels, curlevel;
2090 ASSERT(blkid != DMU_BONUS_BLKID);
2091 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2093 if (blkid > dn->dn_maxblkid)
2096 if (dnode_block_freed(dn, blkid))
2100 * This dnode hasn't been written to disk yet, so there's nothing to
2103 nlevels = dn->dn_phys->dn_nlevels;
2104 if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0)
2107 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2108 if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level))
2111 dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object,
2114 mutex_exit(&db->db_mtx);
2116 * This dbuf already exists. It is either CACHED, or
2117 * (we assume) about to be read or filled.
2123 * Find the closest ancestor (indirect block) of the target block
2124 * that is present in the cache. In this indirect block, we will
2125 * find the bp that is at curlevel, curblkid.
2129 while (curlevel < nlevels - 1) {
2130 int parent_level = curlevel + 1;
2131 uint64_t parent_blkid = curblkid >> epbs;
2134 if (dbuf_hold_impl(dn, parent_level, parent_blkid,
2135 FALSE, TRUE, FTAG, &db) == 0) {
2136 blkptr_t *bpp = db->db_buf->b_data;
2137 bp = bpp[P2PHASE(curblkid, 1 << epbs)];
2138 dbuf_rele(db, FTAG);
2142 curlevel = parent_level;
2143 curblkid = parent_blkid;
2146 if (curlevel == nlevels - 1) {
2147 /* No cached indirect blocks found. */
2148 ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr);
2149 bp = dn->dn_phys->dn_blkptr[curblkid];
2151 if (BP_IS_HOLE(&bp))
2154 ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp));
2156 zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL,
2159 dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP);
2160 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
2161 SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2162 dn->dn_object, level, blkid);
2163 dpa->dpa_curlevel = curlevel;
2164 dpa->dpa_prio = prio;
2165 dpa->dpa_aflags = aflags;
2166 dpa->dpa_spa = dn->dn_objset->os_spa;
2167 dpa->dpa_epbs = epbs;
2171 * If we have the indirect just above us, no need to do the asynchronous
2172 * prefetch chain; we'll just run the last step ourselves. If we're at
2173 * a higher level, though, we want to issue the prefetches for all the
2174 * indirect blocks asynchronously, so we can go on with whatever we were
2177 if (curlevel == level) {
2178 ASSERT3U(curblkid, ==, blkid);
2179 dbuf_issue_final_prefetch(dpa, &bp);
2180 kmem_free(dpa, sizeof (*dpa));
2182 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2183 zbookmark_phys_t zb;
2185 SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2186 dn->dn_object, curlevel, curblkid);
2187 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2188 &bp, dbuf_prefetch_indirect_done, dpa, prio,
2189 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2193 * We use pio here instead of dpa_zio since it's possible that
2194 * dpa may have already been freed.
2200 * Returns with db_holds incremented, and db_mtx not held.
2201 * Note: dn_struct_rwlock must be held.
2204 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid,
2205 boolean_t fail_sparse, boolean_t fail_uncached,
2206 void *tag, dmu_buf_impl_t **dbp)
2208 dmu_buf_impl_t *db, *parent = NULL;
2210 ASSERT(blkid != DMU_BONUS_BLKID);
2211 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2212 ASSERT3U(dn->dn_nlevels, >, level);
2216 /* dbuf_find() returns with db_mtx held */
2217 db = dbuf_find(dn->dn_objset, dn->dn_object, level, blkid);
2220 blkptr_t *bp = NULL;
2224 return (SET_ERROR(ENOENT));
2226 ASSERT3P(parent, ==, NULL);
2227 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
2229 if (err == 0 && bp && BP_IS_HOLE(bp))
2230 err = SET_ERROR(ENOENT);
2233 dbuf_rele(parent, NULL);
2237 if (err && err != ENOENT)
2239 db = dbuf_create(dn, level, blkid, parent, bp);
2242 if (fail_uncached && db->db_state != DB_CACHED) {
2243 mutex_exit(&db->db_mtx);
2244 return (SET_ERROR(ENOENT));
2247 if (db->db_buf && refcount_is_zero(&db->db_holds)) {
2248 arc_buf_add_ref(db->db_buf, db);
2249 if (db->db_buf->b_data == NULL) {
2252 dbuf_rele(parent, NULL);
2257 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
2260 ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
2263 * If this buffer is currently syncing out, and we are are
2264 * still referencing it from db_data, we need to make a copy
2265 * of it in case we decide we want to dirty it again in this txg.
2267 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
2268 dn->dn_object != DMU_META_DNODE_OBJECT &&
2269 db->db_state == DB_CACHED && db->db_data_pending) {
2270 dbuf_dirty_record_t *dr = db->db_data_pending;
2272 if (dr->dt.dl.dr_data == db->db_buf) {
2273 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2276 arc_buf_alloc(dn->dn_objset->os_spa,
2277 db->db.db_size, db, type));
2278 bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
2283 (void) refcount_add(&db->db_holds, tag);
2285 mutex_exit(&db->db_mtx);
2287 /* NOTE: we can't rele the parent until after we drop the db_mtx */
2289 dbuf_rele(parent, NULL);
2291 ASSERT3P(DB_DNODE(db), ==, dn);
2292 ASSERT3U(db->db_blkid, ==, blkid);
2293 ASSERT3U(db->db_level, ==, level);
2300 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
2302 return (dbuf_hold_level(dn, 0, blkid, tag));
2306 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
2309 int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db);
2310 return (err ? NULL : db);
2314 dbuf_create_bonus(dnode_t *dn)
2316 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
2318 ASSERT(dn->dn_bonus == NULL);
2319 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
2323 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
2325 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2328 if (db->db_blkid != DMU_SPILL_BLKID)
2329 return (SET_ERROR(ENOTSUP));
2331 blksz = SPA_MINBLOCKSIZE;
2332 ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset)));
2333 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
2337 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2338 dbuf_new_size(db, blksz, tx);
2339 rw_exit(&dn->dn_struct_rwlock);
2346 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
2348 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
2351 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2353 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
2355 int64_t holds = refcount_add(&db->db_holds, tag);
2359 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
2361 dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid,
2364 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2365 dmu_buf_impl_t *found_db;
2366 boolean_t result = B_FALSE;
2368 if (db->db_blkid == DMU_BONUS_BLKID)
2369 found_db = dbuf_find_bonus(os, obj);
2371 found_db = dbuf_find(os, obj, 0, blkid);
2373 if (found_db != NULL) {
2374 if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) {
2375 (void) refcount_add(&db->db_holds, tag);
2378 mutex_exit(&db->db_mtx);
2384 * If you call dbuf_rele() you had better not be referencing the dnode handle
2385 * unless you have some other direct or indirect hold on the dnode. (An indirect
2386 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2387 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2388 * dnode's parent dbuf evicting its dnode handles.
2391 dbuf_rele(dmu_buf_impl_t *db, void *tag)
2393 mutex_enter(&db->db_mtx);
2394 dbuf_rele_and_unlock(db, tag);
2398 dmu_buf_rele(dmu_buf_t *db, void *tag)
2400 dbuf_rele((dmu_buf_impl_t *)db, tag);
2404 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2405 * db_dirtycnt and db_holds to be updated atomically.
2408 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2412 ASSERT(MUTEX_HELD(&db->db_mtx));
2416 * Remove the reference to the dbuf before removing its hold on the
2417 * dnode so we can guarantee in dnode_move() that a referenced bonus
2418 * buffer has a corresponding dnode hold.
2420 holds = refcount_remove(&db->db_holds, tag);
2424 * We can't freeze indirects if there is a possibility that they
2425 * may be modified in the current syncing context.
2427 if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0))
2428 arc_buf_freeze(db->db_buf);
2430 if (holds == db->db_dirtycnt &&
2431 db->db_level == 0 && db->db_user_immediate_evict)
2432 dbuf_evict_user(db);
2435 if (db->db_blkid == DMU_BONUS_BLKID) {
2437 boolean_t evict_dbuf = db->db_pending_evict;
2440 * If the dnode moves here, we cannot cross this
2441 * barrier until the move completes.
2446 atomic_dec_32(&dn->dn_dbufs_count);
2449 * Decrementing the dbuf count means that the bonus
2450 * buffer's dnode hold is no longer discounted in
2451 * dnode_move(). The dnode cannot move until after
2452 * the dnode_rele() below.
2457 * Do not reference db after its lock is dropped.
2458 * Another thread may evict it.
2460 mutex_exit(&db->db_mtx);
2463 dnode_evict_bonus(dn);
2466 } else if (db->db_buf == NULL) {
2468 * This is a special case: we never associated this
2469 * dbuf with any data allocated from the ARC.
2471 ASSERT(db->db_state == DB_UNCACHED ||
2472 db->db_state == DB_NOFILL);
2474 } else if (arc_released(db->db_buf)) {
2475 arc_buf_t *buf = db->db_buf;
2477 * This dbuf has anonymous data associated with it.
2479 dbuf_clear_data(db);
2480 VERIFY(arc_buf_remove_ref(buf, db));
2483 VERIFY(!arc_buf_remove_ref(db->db_buf, db));
2486 * A dbuf will be eligible for eviction if either the
2487 * 'primarycache' property is set or a duplicate
2488 * copy of this buffer is already cached in the arc.
2490 * In the case of the 'primarycache' a buffer
2491 * is considered for eviction if it matches the
2492 * criteria set in the property.
2494 * To decide if our buffer is considered a
2495 * duplicate, we must call into the arc to determine
2496 * if multiple buffers are referencing the same
2497 * block on-disk. If so, then we simply evict
2500 if (!DBUF_IS_CACHEABLE(db)) {
2501 if (db->db_blkptr != NULL &&
2502 !BP_IS_HOLE(db->db_blkptr) &&
2503 !BP_IS_EMBEDDED(db->db_blkptr)) {
2505 dmu_objset_spa(db->db_objset);
2506 blkptr_t bp = *db->db_blkptr;
2508 arc_freed(spa, &bp);
2512 } else if (db->db_pending_evict ||
2513 arc_buf_eviction_needed(db->db_buf)) {
2516 mutex_exit(&db->db_mtx);
2520 mutex_exit(&db->db_mtx);
2524 #pragma weak dmu_buf_refcount = dbuf_refcount
2526 dbuf_refcount(dmu_buf_impl_t *db)
2528 return (refcount_count(&db->db_holds));
2532 dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user,
2533 dmu_buf_user_t *new_user)
2535 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2537 mutex_enter(&db->db_mtx);
2538 dbuf_verify_user(db, DBVU_NOT_EVICTING);
2539 if (db->db_user == old_user)
2540 db->db_user = new_user;
2542 old_user = db->db_user;
2543 dbuf_verify_user(db, DBVU_NOT_EVICTING);
2544 mutex_exit(&db->db_mtx);
2550 dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2552 return (dmu_buf_replace_user(db_fake, NULL, user));
2556 dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2558 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2560 db->db_user_immediate_evict = TRUE;
2561 return (dmu_buf_set_user(db_fake, user));
2565 dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2567 return (dmu_buf_replace_user(db_fake, user, NULL));
2571 dmu_buf_get_user(dmu_buf_t *db_fake)
2573 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2575 dbuf_verify_user(db, DBVU_NOT_EVICTING);
2576 return (db->db_user);
2580 dmu_buf_user_evict_wait()
2582 taskq_wait(dbu_evict_taskq);
2586 dmu_buf_freeable(dmu_buf_t *dbuf)
2588 boolean_t res = B_FALSE;
2589 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2592 res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset,
2593 db->db_blkptr, db->db_blkptr->blk_birth);
2599 dmu_buf_get_blkptr(dmu_buf_t *db)
2601 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2602 return (dbi->db_blkptr);
2606 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
2608 /* ASSERT(dmu_tx_is_syncing(tx) */
2609 ASSERT(MUTEX_HELD(&db->db_mtx));
2611 if (db->db_blkptr != NULL)
2614 if (db->db_blkid == DMU_SPILL_BLKID) {
2615 db->db_blkptr = &dn->dn_phys->dn_spill;
2616 BP_ZERO(db->db_blkptr);
2619 if (db->db_level == dn->dn_phys->dn_nlevels-1) {
2621 * This buffer was allocated at a time when there was
2622 * no available blkptrs from the dnode, or it was
2623 * inappropriate to hook it in (i.e., nlevels mis-match).
2625 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
2626 ASSERT(db->db_parent == NULL);
2627 db->db_parent = dn->dn_dbuf;
2628 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
2631 dmu_buf_impl_t *parent = db->db_parent;
2632 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2634 ASSERT(dn->dn_phys->dn_nlevels > 1);
2635 if (parent == NULL) {
2636 mutex_exit(&db->db_mtx);
2637 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2638 parent = dbuf_hold_level(dn, db->db_level + 1,
2639 db->db_blkid >> epbs, db);
2640 rw_exit(&dn->dn_struct_rwlock);
2641 mutex_enter(&db->db_mtx);
2642 db->db_parent = parent;
2644 db->db_blkptr = (blkptr_t *)parent->db.db_data +
2645 (db->db_blkid & ((1ULL << epbs) - 1));
2651 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2653 dmu_buf_impl_t *db = dr->dr_dbuf;
2657 ASSERT(dmu_tx_is_syncing(tx));
2659 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2661 mutex_enter(&db->db_mtx);
2663 ASSERT(db->db_level > 0);
2666 /* Read the block if it hasn't been read yet. */
2667 if (db->db_buf == NULL) {
2668 mutex_exit(&db->db_mtx);
2669 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
2670 mutex_enter(&db->db_mtx);
2672 ASSERT3U(db->db_state, ==, DB_CACHED);
2673 ASSERT(db->db_buf != NULL);
2677 /* Indirect block size must match what the dnode thinks it is. */
2678 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2679 dbuf_check_blkptr(dn, db);
2682 /* Provide the pending dirty record to child dbufs */
2683 db->db_data_pending = dr;
2685 mutex_exit(&db->db_mtx);
2686 dbuf_write(dr, db->db_buf, tx);
2689 mutex_enter(&dr->dt.di.dr_mtx);
2690 dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx);
2691 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2692 mutex_exit(&dr->dt.di.dr_mtx);
2697 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2699 arc_buf_t **datap = &dr->dt.dl.dr_data;
2700 dmu_buf_impl_t *db = dr->dr_dbuf;
2703 uint64_t txg = tx->tx_txg;
2705 ASSERT(dmu_tx_is_syncing(tx));
2707 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2709 mutex_enter(&db->db_mtx);
2711 * To be synced, we must be dirtied. But we
2712 * might have been freed after the dirty.
2714 if (db->db_state == DB_UNCACHED) {
2715 /* This buffer has been freed since it was dirtied */
2716 ASSERT(db->db.db_data == NULL);
2717 } else if (db->db_state == DB_FILL) {
2718 /* This buffer was freed and is now being re-filled */
2719 ASSERT(db->db.db_data != dr->dt.dl.dr_data);
2721 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
2728 if (db->db_blkid == DMU_SPILL_BLKID) {
2729 mutex_enter(&dn->dn_mtx);
2730 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
2731 mutex_exit(&dn->dn_mtx);
2735 * If this is a bonus buffer, simply copy the bonus data into the
2736 * dnode. It will be written out when the dnode is synced (and it
2737 * will be synced, since it must have been dirty for dbuf_sync to
2740 if (db->db_blkid == DMU_BONUS_BLKID) {
2741 dbuf_dirty_record_t **drp;
2743 ASSERT(*datap != NULL);
2744 ASSERT0(db->db_level);
2745 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN);
2746 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen);
2749 if (*datap != db->db.db_data) {
2750 zio_buf_free(*datap, DN_MAX_BONUSLEN);
2751 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
2753 db->db_data_pending = NULL;
2754 drp = &db->db_last_dirty;
2756 drp = &(*drp)->dr_next;
2757 ASSERT(dr->dr_next == NULL);
2758 ASSERT(dr->dr_dbuf == db);
2760 if (dr->dr_dbuf->db_level != 0) {
2761 list_destroy(&dr->dt.di.dr_children);
2762 mutex_destroy(&dr->dt.di.dr_mtx);
2764 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2765 ASSERT(db->db_dirtycnt > 0);
2766 db->db_dirtycnt -= 1;
2767 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
2774 * This function may have dropped the db_mtx lock allowing a dmu_sync
2775 * operation to sneak in. As a result, we need to ensure that we
2776 * don't check the dr_override_state until we have returned from
2777 * dbuf_check_blkptr.
2779 dbuf_check_blkptr(dn, db);
2782 * If this buffer is in the middle of an immediate write,
2783 * wait for the synchronous IO to complete.
2785 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
2786 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
2787 cv_wait(&db->db_changed, &db->db_mtx);
2788 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
2791 if (db->db_state != DB_NOFILL &&
2792 dn->dn_object != DMU_META_DNODE_OBJECT &&
2793 refcount_count(&db->db_holds) > 1 &&
2794 dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
2795 *datap == db->db_buf) {
2797 * If this buffer is currently "in use" (i.e., there
2798 * are active holds and db_data still references it),
2799 * then make a copy before we start the write so that
2800 * any modifications from the open txg will not leak
2803 * NOTE: this copy does not need to be made for
2804 * objects only modified in the syncing context (e.g.
2805 * DNONE_DNODE blocks).
2807 int blksz = arc_buf_size(*datap);
2808 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2809 *datap = arc_buf_alloc(os->os_spa, blksz, db, type);
2810 bcopy(db->db.db_data, (*datap)->b_data, blksz);
2812 db->db_data_pending = dr;
2814 mutex_exit(&db->db_mtx);
2816 dbuf_write(dr, *datap, tx);
2818 ASSERT(!list_link_active(&dr->dr_dirty_node));
2819 if (dn->dn_object == DMU_META_DNODE_OBJECT) {
2820 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
2824 * Although zio_nowait() does not "wait for an IO", it does
2825 * initiate the IO. If this is an empty write it seems plausible
2826 * that the IO could actually be completed before the nowait
2827 * returns. We need to DB_DNODE_EXIT() first in case
2828 * zio_nowait() invalidates the dbuf.
2831 zio_nowait(dr->dr_zio);
2836 dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx)
2838 dbuf_dirty_record_t *dr;
2840 while (dr = list_head(list)) {
2841 if (dr->dr_zio != NULL) {
2843 * If we find an already initialized zio then we
2844 * are processing the meta-dnode, and we have finished.
2845 * The dbufs for all dnodes are put back on the list
2846 * during processing, so that we can zio_wait()
2847 * these IOs after initiating all child IOs.
2849 ASSERT3U(dr->dr_dbuf->db.db_object, ==,
2850 DMU_META_DNODE_OBJECT);
2853 if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
2854 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
2855 VERIFY3U(dr->dr_dbuf->db_level, ==, level);
2857 list_remove(list, dr);
2858 if (dr->dr_dbuf->db_level > 0)
2859 dbuf_sync_indirect(dr, tx);
2861 dbuf_sync_leaf(dr, tx);
2867 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
2869 dmu_buf_impl_t *db = vdb;
2871 blkptr_t *bp = zio->io_bp;
2872 blkptr_t *bp_orig = &zio->io_bp_orig;
2873 spa_t *spa = zio->io_spa;
2878 ASSERT3P(db->db_blkptr, ==, bp);
2882 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
2883 dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
2884 zio->io_prev_space_delta = delta;
2886 if (bp->blk_birth != 0) {
2887 ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
2888 BP_GET_TYPE(bp) == dn->dn_type) ||
2889 (db->db_blkid == DMU_SPILL_BLKID &&
2890 BP_GET_TYPE(bp) == dn->dn_bonustype) ||
2891 BP_IS_EMBEDDED(bp));
2892 ASSERT(BP_GET_LEVEL(bp) == db->db_level);
2895 mutex_enter(&db->db_mtx);
2898 if (db->db_blkid == DMU_SPILL_BLKID) {
2899 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2900 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2901 db->db_blkptr == &dn->dn_phys->dn_spill);
2905 if (db->db_level == 0) {
2906 mutex_enter(&dn->dn_mtx);
2907 if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
2908 db->db_blkid != DMU_SPILL_BLKID)
2909 dn->dn_phys->dn_maxblkid = db->db_blkid;
2910 mutex_exit(&dn->dn_mtx);
2912 if (dn->dn_type == DMU_OT_DNODE) {
2913 dnode_phys_t *dnp = db->db.db_data;
2914 for (i = db->db.db_size >> DNODE_SHIFT; i > 0;
2916 if (dnp->dn_type != DMU_OT_NONE)
2920 if (BP_IS_HOLE(bp)) {
2927 blkptr_t *ibp = db->db.db_data;
2928 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2929 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
2930 if (BP_IS_HOLE(ibp))
2932 fill += BP_GET_FILL(ibp);
2937 if (!BP_IS_EMBEDDED(bp))
2938 bp->blk_fill = fill;
2940 mutex_exit(&db->db_mtx);
2944 * The SPA will call this callback several times for each zio - once
2945 * for every physical child i/o (zio->io_phys_children times). This
2946 * allows the DMU to monitor the progress of each logical i/o. For example,
2947 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
2948 * block. There may be a long delay before all copies/fragments are completed,
2949 * so this callback allows us to retire dirty space gradually, as the physical
2954 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
2956 dmu_buf_impl_t *db = arg;
2957 objset_t *os = db->db_objset;
2958 dsl_pool_t *dp = dmu_objset_pool(os);
2959 dbuf_dirty_record_t *dr;
2962 dr = db->db_data_pending;
2963 ASSERT3U(dr->dr_txg, ==, zio->io_txg);
2966 * The callback will be called io_phys_children times. Retire one
2967 * portion of our dirty space each time we are called. Any rounding
2968 * error will be cleaned up by dsl_pool_sync()'s call to
2969 * dsl_pool_undirty_space().
2971 delta = dr->dr_accounted / zio->io_phys_children;
2972 dsl_pool_undirty_space(dp, delta, zio->io_txg);
2977 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
2979 dmu_buf_impl_t *db = vdb;
2980 blkptr_t *bp_orig = &zio->io_bp_orig;
2981 blkptr_t *bp = db->db_blkptr;
2982 objset_t *os = db->db_objset;
2983 dmu_tx_t *tx = os->os_synctx;
2984 dbuf_dirty_record_t **drp, *dr;
2986 ASSERT0(zio->io_error);
2987 ASSERT(db->db_blkptr == bp);
2990 * For nopwrites and rewrites we ensure that the bp matches our
2991 * original and bypass all the accounting.
2993 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
2994 ASSERT(BP_EQUAL(bp, bp_orig));
2996 dsl_dataset_t *ds = os->os_dsl_dataset;
2997 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
2998 dsl_dataset_block_born(ds, bp, tx);
3001 mutex_enter(&db->db_mtx);
3005 drp = &db->db_last_dirty;
3006 while ((dr = *drp) != db->db_data_pending)
3008 ASSERT(!list_link_active(&dr->dr_dirty_node));
3009 ASSERT(dr->dr_dbuf == db);
3010 ASSERT(dr->dr_next == NULL);
3014 if (db->db_blkid == DMU_SPILL_BLKID) {
3019 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3020 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
3021 db->db_blkptr == &dn->dn_phys->dn_spill);
3026 if (db->db_level == 0) {
3027 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
3028 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
3029 if (db->db_state != DB_NOFILL) {
3030 if (dr->dt.dl.dr_data != db->db_buf)
3031 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data,
3033 else if (!arc_released(db->db_buf))
3034 arc_set_callback(db->db_buf, dbuf_do_evict, db);
3041 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3042 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
3043 if (!BP_IS_HOLE(db->db_blkptr)) {
3045 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3046 ASSERT3U(db->db_blkid, <=,
3047 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs));
3048 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
3050 if (!arc_released(db->db_buf))
3051 arc_set_callback(db->db_buf, dbuf_do_evict, db);
3054 mutex_destroy(&dr->dt.di.dr_mtx);
3055 list_destroy(&dr->dt.di.dr_children);
3057 kmem_free(dr, sizeof (dbuf_dirty_record_t));
3059 cv_broadcast(&db->db_changed);
3060 ASSERT(db->db_dirtycnt > 0);
3061 db->db_dirtycnt -= 1;
3062 db->db_data_pending = NULL;
3063 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg);
3067 dbuf_write_nofill_ready(zio_t *zio)
3069 dbuf_write_ready(zio, NULL, zio->io_private);
3073 dbuf_write_nofill_done(zio_t *zio)
3075 dbuf_write_done(zio, NULL, zio->io_private);
3079 dbuf_write_override_ready(zio_t *zio)
3081 dbuf_dirty_record_t *dr = zio->io_private;
3082 dmu_buf_impl_t *db = dr->dr_dbuf;
3084 dbuf_write_ready(zio, NULL, db);
3088 dbuf_write_override_done(zio_t *zio)
3090 dbuf_dirty_record_t *dr = zio->io_private;
3091 dmu_buf_impl_t *db = dr->dr_dbuf;
3092 blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
3094 mutex_enter(&db->db_mtx);
3095 if (!BP_EQUAL(zio->io_bp, obp)) {
3096 if (!BP_IS_HOLE(obp))
3097 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
3098 arc_release(dr->dt.dl.dr_data, db);
3100 mutex_exit(&db->db_mtx);
3102 dbuf_write_done(zio, NULL, db);
3105 /* Issue I/O to commit a dirty buffer to disk. */
3107 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
3109 dmu_buf_impl_t *db = dr->dr_dbuf;
3112 dmu_buf_impl_t *parent = db->db_parent;
3113 uint64_t txg = tx->tx_txg;
3114 zbookmark_phys_t zb;
3123 if (db->db_state != DB_NOFILL) {
3124 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
3126 * Private object buffers are released here rather
3127 * than in dbuf_dirty() since they are only modified
3128 * in the syncing context and we don't want the
3129 * overhead of making multiple copies of the data.
3131 if (BP_IS_HOLE(db->db_blkptr)) {
3134 dbuf_release_bp(db);
3139 if (parent != dn->dn_dbuf) {
3140 /* Our parent is an indirect block. */
3141 /* We have a dirty parent that has been scheduled for write. */
3142 ASSERT(parent && parent->db_data_pending);
3143 /* Our parent's buffer is one level closer to the dnode. */
3144 ASSERT(db->db_level == parent->db_level-1);
3146 * We're about to modify our parent's db_data by modifying
3147 * our block pointer, so the parent must be released.
3149 ASSERT(arc_released(parent->db_buf));
3150 zio = parent->db_data_pending->dr_zio;
3152 /* Our parent is the dnode itself. */
3153 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
3154 db->db_blkid != DMU_SPILL_BLKID) ||
3155 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
3156 if (db->db_blkid != DMU_SPILL_BLKID)
3157 ASSERT3P(db->db_blkptr, ==,
3158 &dn->dn_phys->dn_blkptr[db->db_blkid]);
3162 ASSERT(db->db_level == 0 || data == db->db_buf);
3163 ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
3166 SET_BOOKMARK(&zb, os->os_dsl_dataset ?
3167 os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
3168 db->db.db_object, db->db_level, db->db_blkid);
3170 if (db->db_blkid == DMU_SPILL_BLKID)
3172 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
3174 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
3177 if (db->db_level == 0 &&
3178 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
3180 * The BP for this block has been provided by open context
3181 * (by dmu_sync() or dmu_buf_write_embedded()).
3183 void *contents = (data != NULL) ? data->b_data : NULL;
3185 dr->dr_zio = zio_write(zio, os->os_spa, txg,
3186 db->db_blkptr, contents, db->db.db_size, &zp,
3187 dbuf_write_override_ready, NULL, dbuf_write_override_done,
3188 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
3189 mutex_enter(&db->db_mtx);
3190 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
3191 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
3192 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
3193 mutex_exit(&db->db_mtx);
3194 } else if (db->db_state == DB_NOFILL) {
3195 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF ||
3196 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY);
3197 dr->dr_zio = zio_write(zio, os->os_spa, txg,
3198 db->db_blkptr, NULL, db->db.db_size, &zp,
3199 dbuf_write_nofill_ready, NULL, dbuf_write_nofill_done, db,
3200 ZIO_PRIORITY_ASYNC_WRITE,
3201 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
3203 ASSERT(arc_released(data));
3204 dr->dr_zio = arc_write(zio, os->os_spa, txg,
3205 db->db_blkptr, data, DBUF_IS_L2CACHEABLE(db),
3206 DBUF_IS_L2COMPRESSIBLE(db), &zp, dbuf_write_ready,
3207 dbuf_write_physdone, dbuf_write_done, db,
3208 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);