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) 2013 by Delphix. All rights reserved.
25 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
28 #include <sys/zfs_context.h>
31 #include <sys/dmu_impl.h>
33 #include <sys/dmu_objset.h>
34 #include <sys/dsl_dataset.h>
35 #include <sys/dsl_dir.h>
36 #include <sys/dmu_tx.h>
39 #include <sys/dmu_zfetch.h>
41 #include <sys/sa_impl.h>
43 struct dbuf_hold_impl_data {
44 /* Function arguments */
50 dmu_buf_impl_t **dh_dbp;
52 dmu_buf_impl_t *dh_db;
53 dmu_buf_impl_t *dh_parent;
56 dbuf_dirty_record_t *dh_dr;
57 arc_buf_contents_t dh_type;
61 static void __dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh,
62 dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse,
63 void *tag, dmu_buf_impl_t **dbp, int depth);
64 static int __dbuf_hold_impl(struct dbuf_hold_impl_data *dh);
66 static void dbuf_destroy(dmu_buf_impl_t *db);
67 static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
68 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
71 * Global data structures and functions for the dbuf cache.
73 static kmem_cache_t *dbuf_cache;
77 dbuf_cons(void *vdb, void *unused, int kmflag)
79 dmu_buf_impl_t *db = vdb;
80 bzero(db, sizeof (dmu_buf_impl_t));
82 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
83 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
84 refcount_create(&db->db_holds);
85 list_link_init(&db->db_link);
91 dbuf_dest(void *vdb, void *unused)
93 dmu_buf_impl_t *db = vdb;
94 mutex_destroy(&db->db_mtx);
95 cv_destroy(&db->db_changed);
96 refcount_destroy(&db->db_holds);
100 * dbuf hash table routines
102 static dbuf_hash_table_t dbuf_hash_table;
104 static uint64_t dbuf_hash_count;
107 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
109 uintptr_t osv = (uintptr_t)os;
110 uint64_t crc = -1ULL;
112 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
113 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF];
114 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF];
115 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF];
116 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF];
117 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF];
118 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF];
120 crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16);
125 #define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid);
127 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
128 ((dbuf)->db.db_object == (obj) && \
129 (dbuf)->db_objset == (os) && \
130 (dbuf)->db_level == (level) && \
131 (dbuf)->db_blkid == (blkid))
134 dbuf_find(dnode_t *dn, uint8_t level, uint64_t blkid)
136 dbuf_hash_table_t *h = &dbuf_hash_table;
137 objset_t *os = dn->dn_objset;
144 hv = DBUF_HASH(os, obj, level, blkid);
145 idx = hv & h->hash_table_mask;
147 mutex_enter(DBUF_HASH_MUTEX(h, idx));
148 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
149 if (DBUF_EQUAL(db, os, obj, level, blkid)) {
150 mutex_enter(&db->db_mtx);
151 if (db->db_state != DB_EVICTING) {
152 mutex_exit(DBUF_HASH_MUTEX(h, idx));
155 mutex_exit(&db->db_mtx);
158 mutex_exit(DBUF_HASH_MUTEX(h, idx));
163 * Insert an entry into the hash table. If there is already an element
164 * equal to elem in the hash table, then the already existing element
165 * will be returned and the new element will not be inserted.
166 * Otherwise returns NULL.
168 static dmu_buf_impl_t *
169 dbuf_hash_insert(dmu_buf_impl_t *db)
171 dbuf_hash_table_t *h = &dbuf_hash_table;
172 objset_t *os = db->db_objset;
173 uint64_t obj = db->db.db_object;
174 int level = db->db_level;
175 uint64_t blkid, hv, idx;
178 blkid = db->db_blkid;
179 hv = DBUF_HASH(os, obj, level, blkid);
180 idx = hv & h->hash_table_mask;
182 mutex_enter(DBUF_HASH_MUTEX(h, idx));
183 for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) {
184 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
185 mutex_enter(&dbf->db_mtx);
186 if (dbf->db_state != DB_EVICTING) {
187 mutex_exit(DBUF_HASH_MUTEX(h, idx));
190 mutex_exit(&dbf->db_mtx);
194 mutex_enter(&db->db_mtx);
195 db->db_hash_next = h->hash_table[idx];
196 h->hash_table[idx] = db;
197 mutex_exit(DBUF_HASH_MUTEX(h, idx));
198 atomic_add_64(&dbuf_hash_count, 1);
204 * Remove an entry from the hash table. This operation will
205 * fail if there are any existing holds on the db.
208 dbuf_hash_remove(dmu_buf_impl_t *db)
210 dbuf_hash_table_t *h = &dbuf_hash_table;
212 dmu_buf_impl_t *dbf, **dbp;
214 hv = DBUF_HASH(db->db_objset, db->db.db_object,
215 db->db_level, db->db_blkid);
216 idx = hv & h->hash_table_mask;
219 * We musn't hold db_mtx to maintin lock ordering:
220 * DBUF_HASH_MUTEX > db_mtx.
222 ASSERT(refcount_is_zero(&db->db_holds));
223 ASSERT(db->db_state == DB_EVICTING);
224 ASSERT(!MUTEX_HELD(&db->db_mtx));
226 mutex_enter(DBUF_HASH_MUTEX(h, idx));
227 dbp = &h->hash_table[idx];
228 while ((dbf = *dbp) != db) {
229 dbp = &dbf->db_hash_next;
232 *dbp = db->db_hash_next;
233 db->db_hash_next = NULL;
234 mutex_exit(DBUF_HASH_MUTEX(h, idx));
235 atomic_add_64(&dbuf_hash_count, -1);
238 static arc_evict_func_t dbuf_do_evict;
241 dbuf_evict_user(dmu_buf_impl_t *db)
243 ASSERT(MUTEX_HELD(&db->db_mtx));
245 if (db->db_level != 0 || db->db_evict_func == NULL)
248 if (db->db_user_data_ptr_ptr)
249 *db->db_user_data_ptr_ptr = db->db.db_data;
250 db->db_evict_func(&db->db, db->db_user_ptr);
251 db->db_user_ptr = NULL;
252 db->db_user_data_ptr_ptr = NULL;
253 db->db_evict_func = NULL;
257 dbuf_is_metadata(dmu_buf_impl_t *db)
259 if (db->db_level > 0) {
262 boolean_t is_metadata;
265 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type);
268 return (is_metadata);
273 dbuf_evict(dmu_buf_impl_t *db)
275 ASSERT(MUTEX_HELD(&db->db_mtx));
276 ASSERT(db->db_buf == NULL);
277 ASSERT(db->db_data_pending == NULL);
286 uint64_t hsize = 1ULL << 16;
287 dbuf_hash_table_t *h = &dbuf_hash_table;
291 * The hash table is big enough to fill all of physical memory
292 * with an average 4K block size. The table will take up
293 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
295 while (hsize * 4096 < physmem * PAGESIZE)
299 h->hash_table_mask = hsize - 1;
300 #if defined(_KERNEL) && defined(HAVE_SPL)
301 /* Large allocations which do not require contiguous pages
302 * should be using vmem_alloc() in the linux kernel */
303 h->hash_table = vmem_zalloc(hsize * sizeof (void *), KM_PUSHPAGE);
305 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
307 if (h->hash_table == NULL) {
308 /* XXX - we should really return an error instead of assert */
309 ASSERT(hsize > (1ULL << 10));
314 dbuf_cache = kmem_cache_create("dmu_buf_impl_t",
315 sizeof (dmu_buf_impl_t),
316 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
318 for (i = 0; i < DBUF_MUTEXES; i++)
319 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
327 dbuf_hash_table_t *h = &dbuf_hash_table;
330 dbuf_stats_destroy();
332 for (i = 0; i < DBUF_MUTEXES; i++)
333 mutex_destroy(&h->hash_mutexes[i]);
334 #if defined(_KERNEL) && defined(HAVE_SPL)
335 /* Large allocations which do not require contiguous pages
336 * should be using vmem_free() in the linux kernel */
337 vmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
339 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
341 kmem_cache_destroy(dbuf_cache);
350 dbuf_verify(dmu_buf_impl_t *db)
353 dbuf_dirty_record_t *dr;
355 ASSERT(MUTEX_HELD(&db->db_mtx));
357 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
360 ASSERT(db->db_objset != NULL);
364 ASSERT(db->db_parent == NULL);
365 ASSERT(db->db_blkptr == NULL);
367 ASSERT3U(db->db.db_object, ==, dn->dn_object);
368 ASSERT3P(db->db_objset, ==, dn->dn_objset);
369 ASSERT3U(db->db_level, <, dn->dn_nlevels);
370 ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
371 db->db_blkid == DMU_SPILL_BLKID ||
372 !list_is_empty(&dn->dn_dbufs));
374 if (db->db_blkid == DMU_BONUS_BLKID) {
376 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
377 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
378 } else if (db->db_blkid == DMU_SPILL_BLKID) {
380 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
381 ASSERT0(db->db.db_offset);
383 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
386 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
387 ASSERT(dr->dr_dbuf == db);
389 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
390 ASSERT(dr->dr_dbuf == db);
393 * We can't assert that db_size matches dn_datablksz because it
394 * can be momentarily different when another thread is doing
397 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
398 dr = db->db_data_pending;
400 * It should only be modified in syncing context, so
401 * make sure we only have one copy of the data.
403 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
406 /* verify db->db_blkptr */
408 if (db->db_parent == dn->dn_dbuf) {
409 /* db is pointed to by the dnode */
410 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
411 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
412 ASSERT(db->db_parent == NULL);
414 ASSERT(db->db_parent != NULL);
415 if (db->db_blkid != DMU_SPILL_BLKID)
416 ASSERT3P(db->db_blkptr, ==,
417 &dn->dn_phys->dn_blkptr[db->db_blkid]);
419 /* db is pointed to by an indirect block */
420 ASSERTV(int epb = db->db_parent->db.db_size >>
422 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
423 ASSERT3U(db->db_parent->db.db_object, ==,
426 * dnode_grow_indblksz() can make this fail if we don't
427 * have the struct_rwlock. XXX indblksz no longer
428 * grows. safe to do this now?
430 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
431 ASSERT3P(db->db_blkptr, ==,
432 ((blkptr_t *)db->db_parent->db.db_data +
433 db->db_blkid % epb));
437 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
438 (db->db_buf == NULL || db->db_buf->b_data) &&
439 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
440 db->db_state != DB_FILL && !dn->dn_free_txg) {
442 * If the blkptr isn't set but they have nonzero data,
443 * it had better be dirty, otherwise we'll lose that
444 * data when we evict this buffer.
446 if (db->db_dirtycnt == 0) {
447 ASSERTV(uint64_t *buf = db->db.db_data);
450 for (i = 0; i < db->db.db_size >> 3; i++) {
460 dbuf_update_data(dmu_buf_impl_t *db)
462 ASSERT(MUTEX_HELD(&db->db_mtx));
463 if (db->db_level == 0 && db->db_user_data_ptr_ptr) {
464 ASSERT(!refcount_is_zero(&db->db_holds));
465 *db->db_user_data_ptr_ptr = db->db.db_data;
470 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
472 ASSERT(MUTEX_HELD(&db->db_mtx));
473 ASSERT(db->db_buf == NULL || !arc_has_callback(db->db_buf));
476 ASSERT(buf->b_data != NULL);
477 db->db.db_data = buf->b_data;
478 if (!arc_released(buf))
479 arc_set_callback(buf, dbuf_do_evict, db);
480 dbuf_update_data(db);
483 db->db.db_data = NULL;
484 if (db->db_state != DB_NOFILL)
485 db->db_state = DB_UNCACHED;
490 * Loan out an arc_buf for read. Return the loaned arc_buf.
493 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
497 mutex_enter(&db->db_mtx);
498 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
499 int blksz = db->db.db_size;
502 mutex_exit(&db->db_mtx);
503 DB_GET_SPA(&spa, db);
504 abuf = arc_loan_buf(spa, blksz);
505 bcopy(db->db.db_data, abuf->b_data, blksz);
508 arc_loan_inuse_buf(abuf, db);
509 dbuf_set_data(db, NULL);
510 mutex_exit(&db->db_mtx);
516 dbuf_whichblock(dnode_t *dn, uint64_t offset)
518 if (dn->dn_datablkshift) {
519 return (offset >> dn->dn_datablkshift);
521 ASSERT3U(offset, <, dn->dn_datablksz);
527 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
529 dmu_buf_impl_t *db = vdb;
531 mutex_enter(&db->db_mtx);
532 ASSERT3U(db->db_state, ==, DB_READ);
534 * All reads are synchronous, so we must have a hold on the dbuf
536 ASSERT(refcount_count(&db->db_holds) > 0);
537 ASSERT(db->db_buf == NULL);
538 ASSERT(db->db.db_data == NULL);
539 if (db->db_level == 0 && db->db_freed_in_flight) {
540 /* we were freed in flight; disregard any error */
541 arc_release(buf, db);
542 bzero(buf->b_data, db->db.db_size);
544 db->db_freed_in_flight = FALSE;
545 dbuf_set_data(db, buf);
546 db->db_state = DB_CACHED;
547 } else if (zio == NULL || zio->io_error == 0) {
548 dbuf_set_data(db, buf);
549 db->db_state = DB_CACHED;
551 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
552 ASSERT3P(db->db_buf, ==, NULL);
553 VERIFY(arc_buf_remove_ref(buf, db));
554 db->db_state = DB_UNCACHED;
556 cv_broadcast(&db->db_changed);
557 dbuf_rele_and_unlock(db, NULL);
561 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t *flags)
566 uint32_t aflags = ARC_NOWAIT;
570 ASSERT(!refcount_is_zero(&db->db_holds));
571 /* We need the struct_rwlock to prevent db_blkptr from changing. */
572 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
573 ASSERT(MUTEX_HELD(&db->db_mtx));
574 ASSERT(db->db_state == DB_UNCACHED);
575 ASSERT(db->db_buf == NULL);
577 if (db->db_blkid == DMU_BONUS_BLKID) {
578 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
580 ASSERT3U(bonuslen, <=, db->db.db_size);
581 db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN);
582 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
583 if (bonuslen < DN_MAX_BONUSLEN)
584 bzero(db->db.db_data, DN_MAX_BONUSLEN);
586 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
588 dbuf_update_data(db);
589 db->db_state = DB_CACHED;
590 mutex_exit(&db->db_mtx);
595 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
596 * processes the delete record and clears the bp while we are waiting
597 * for the dn_mtx (resulting in a "no" from block_freed).
599 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
600 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
601 BP_IS_HOLE(db->db_blkptr)))) {
602 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
604 dbuf_set_data(db, arc_buf_alloc(dn->dn_objset->os_spa,
605 db->db.db_size, db, type));
607 bzero(db->db.db_data, db->db.db_size);
608 db->db_state = DB_CACHED;
609 *flags |= DB_RF_CACHED;
610 mutex_exit(&db->db_mtx);
614 spa = dn->dn_objset->os_spa;
617 db->db_state = DB_READ;
618 mutex_exit(&db->db_mtx);
620 if (DBUF_IS_L2CACHEABLE(db))
621 aflags |= ARC_L2CACHE;
622 if (DBUF_IS_L2COMPRESSIBLE(db))
623 aflags |= ARC_L2COMPRESS;
625 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
626 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
627 db->db.db_object, db->db_level, db->db_blkid);
629 dbuf_add_ref(db, NULL);
631 (void) arc_read(zio, spa, db->db_blkptr,
632 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
633 (*flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
635 if (aflags & ARC_CACHED)
636 *flags |= DB_RF_CACHED;
640 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
643 int havepzio = (zio != NULL);
648 * We don't have to hold the mutex to check db_state because it
649 * can't be freed while we have a hold on the buffer.
651 ASSERT(!refcount_is_zero(&db->db_holds));
653 if (db->db_state == DB_NOFILL)
654 return (SET_ERROR(EIO));
658 if ((flags & DB_RF_HAVESTRUCT) == 0)
659 rw_enter(&dn->dn_struct_rwlock, RW_READER);
661 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
662 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
663 DBUF_IS_CACHEABLE(db);
665 mutex_enter(&db->db_mtx);
666 if (db->db_state == DB_CACHED) {
667 mutex_exit(&db->db_mtx);
669 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
670 db->db.db_size, TRUE);
671 if ((flags & DB_RF_HAVESTRUCT) == 0)
672 rw_exit(&dn->dn_struct_rwlock);
674 } else if (db->db_state == DB_UNCACHED) {
675 spa_t *spa = dn->dn_objset->os_spa;
678 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
679 dbuf_read_impl(db, zio, &flags);
681 /* dbuf_read_impl has dropped db_mtx for us */
684 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
685 db->db.db_size, flags & DB_RF_CACHED);
687 if ((flags & DB_RF_HAVESTRUCT) == 0)
688 rw_exit(&dn->dn_struct_rwlock);
695 * Another reader came in while the dbuf was in flight
696 * between UNCACHED and CACHED. Either a writer will finish
697 * writing the buffer (sending the dbuf to CACHED) or the
698 * first reader's request will reach the read_done callback
699 * and send the dbuf to CACHED. Otherwise, a failure
700 * occurred and the dbuf went to UNCACHED.
702 mutex_exit(&db->db_mtx);
704 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
705 db->db.db_size, TRUE);
706 if ((flags & DB_RF_HAVESTRUCT) == 0)
707 rw_exit(&dn->dn_struct_rwlock);
710 /* Skip the wait per the caller's request. */
711 mutex_enter(&db->db_mtx);
712 if ((flags & DB_RF_NEVERWAIT) == 0) {
713 while (db->db_state == DB_READ ||
714 db->db_state == DB_FILL) {
715 ASSERT(db->db_state == DB_READ ||
716 (flags & DB_RF_HAVESTRUCT) == 0);
717 cv_wait(&db->db_changed, &db->db_mtx);
719 if (db->db_state == DB_UNCACHED)
720 err = SET_ERROR(EIO);
722 mutex_exit(&db->db_mtx);
725 ASSERT(err || havepzio || db->db_state == DB_CACHED);
730 dbuf_noread(dmu_buf_impl_t *db)
732 ASSERT(!refcount_is_zero(&db->db_holds));
733 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
734 mutex_enter(&db->db_mtx);
735 while (db->db_state == DB_READ || db->db_state == DB_FILL)
736 cv_wait(&db->db_changed, &db->db_mtx);
737 if (db->db_state == DB_UNCACHED) {
738 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
741 ASSERT(db->db_buf == NULL);
742 ASSERT(db->db.db_data == NULL);
743 DB_GET_SPA(&spa, db);
744 dbuf_set_data(db, arc_buf_alloc(spa, db->db.db_size, db, type));
745 db->db_state = DB_FILL;
746 } else if (db->db_state == DB_NOFILL) {
747 dbuf_set_data(db, NULL);
749 ASSERT3U(db->db_state, ==, DB_CACHED);
751 mutex_exit(&db->db_mtx);
755 * This is our just-in-time copy function. It makes a copy of
756 * buffers, that have been modified in a previous transaction
757 * group, before we modify them in the current active group.
759 * This function is used in two places: when we are dirtying a
760 * buffer for the first time in a txg, and when we are freeing
761 * a range in a dnode that includes this buffer.
763 * Note that when we are called from dbuf_free_range() we do
764 * not put a hold on the buffer, we just traverse the active
765 * dbuf list for the dnode.
768 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
770 dbuf_dirty_record_t *dr = db->db_last_dirty;
772 ASSERT(MUTEX_HELD(&db->db_mtx));
773 ASSERT(db->db.db_data != NULL);
774 ASSERT(db->db_level == 0);
775 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
778 (dr->dt.dl.dr_data !=
779 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
783 * If the last dirty record for this dbuf has not yet synced
784 * and its referencing the dbuf data, either:
785 * reset the reference to point to a new copy,
786 * or (if there a no active holders)
787 * just null out the current db_data pointer.
789 ASSERT(dr->dr_txg >= txg - 2);
790 if (db->db_blkid == DMU_BONUS_BLKID) {
791 /* Note that the data bufs here are zio_bufs */
792 dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
793 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
794 bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
795 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
796 int size = db->db.db_size;
797 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
800 DB_GET_SPA(&spa, db);
801 dr->dt.dl.dr_data = arc_buf_alloc(spa, size, db, type);
802 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
804 dbuf_set_data(db, NULL);
809 dbuf_unoverride(dbuf_dirty_record_t *dr)
811 dmu_buf_impl_t *db = dr->dr_dbuf;
812 blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
813 uint64_t txg = dr->dr_txg;
815 ASSERT(MUTEX_HELD(&db->db_mtx));
816 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
817 ASSERT(db->db_level == 0);
819 if (db->db_blkid == DMU_BONUS_BLKID ||
820 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
823 ASSERT(db->db_data_pending != dr);
825 /* free this block */
826 if (!BP_IS_HOLE(bp)) {
829 DB_GET_SPA(&spa, db);
830 zio_free(spa, txg, bp);
832 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
834 * Release the already-written buffer, so we leave it in
835 * a consistent dirty state. Note that all callers are
836 * modifying the buffer, so they will immediately do
837 * another (redundant) arc_release(). Therefore, leave
838 * the buf thawed to save the effort of freezing &
839 * immediately re-thawing it.
841 arc_release(dr->dt.dl.dr_data, db);
845 * Evict (if its unreferenced) or clear (if its referenced) any level-0
846 * data blocks in the free range, so that any future readers will find
847 * empty blocks. Also, if we happen accross any level-1 dbufs in the
848 * range that have not already been marked dirty, mark them dirty so
849 * they stay in memory.
852 dbuf_free_range(dnode_t *dn, uint64_t start, uint64_t end, dmu_tx_t *tx)
854 dmu_buf_impl_t *db, *db_next;
855 uint64_t txg = tx->tx_txg;
856 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
857 uint64_t first_l1 = start >> epbs;
858 uint64_t last_l1 = end >> epbs;
860 if (end > dn->dn_maxblkid && (end != DMU_SPILL_BLKID)) {
861 end = dn->dn_maxblkid;
862 last_l1 = end >> epbs;
864 dprintf_dnode(dn, "start=%llu end=%llu\n", start, end);
865 mutex_enter(&dn->dn_dbufs_mtx);
866 for (db = list_head(&dn->dn_dbufs); db; db = db_next) {
867 db_next = list_next(&dn->dn_dbufs, db);
868 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
870 if (db->db_level == 1 &&
871 db->db_blkid >= first_l1 && db->db_blkid <= last_l1) {
872 mutex_enter(&db->db_mtx);
873 if (db->db_last_dirty &&
874 db->db_last_dirty->dr_txg < txg) {
875 dbuf_add_ref(db, FTAG);
876 mutex_exit(&db->db_mtx);
877 dbuf_will_dirty(db, tx);
880 mutex_exit(&db->db_mtx);
884 if (db->db_level != 0)
886 dprintf_dbuf(db, "found buf %s\n", "");
887 if (db->db_blkid < start || db->db_blkid > end)
890 /* found a level 0 buffer in the range */
891 mutex_enter(&db->db_mtx);
892 if (dbuf_undirty(db, tx)) {
893 /* mutex has been dropped and dbuf destroyed */
897 if (db->db_state == DB_UNCACHED ||
898 db->db_state == DB_NOFILL ||
899 db->db_state == DB_EVICTING) {
900 ASSERT(db->db.db_data == NULL);
901 mutex_exit(&db->db_mtx);
904 if (db->db_state == DB_READ || db->db_state == DB_FILL) {
905 /* will be handled in dbuf_read_done or dbuf_rele */
906 db->db_freed_in_flight = TRUE;
907 mutex_exit(&db->db_mtx);
910 if (refcount_count(&db->db_holds) == 0) {
915 /* The dbuf is referenced */
917 if (db->db_last_dirty != NULL) {
918 dbuf_dirty_record_t *dr = db->db_last_dirty;
920 if (dr->dr_txg == txg) {
922 * This buffer is "in-use", re-adjust the file
923 * size to reflect that this buffer may
924 * contain new data when we sync.
926 if (db->db_blkid != DMU_SPILL_BLKID &&
927 db->db_blkid > dn->dn_maxblkid)
928 dn->dn_maxblkid = db->db_blkid;
932 * This dbuf is not dirty in the open context.
933 * Either uncache it (if its not referenced in
934 * the open context) or reset its contents to
937 dbuf_fix_old_data(db, txg);
940 /* clear the contents if its cached */
941 if (db->db_state == DB_CACHED) {
942 ASSERT(db->db.db_data != NULL);
943 arc_release(db->db_buf, db);
944 bzero(db->db.db_data, db->db.db_size);
945 arc_buf_freeze(db->db_buf);
948 mutex_exit(&db->db_mtx);
950 mutex_exit(&dn->dn_dbufs_mtx);
954 dbuf_block_freeable(dmu_buf_impl_t *db)
956 dsl_dataset_t *ds = db->db_objset->os_dsl_dataset;
957 uint64_t birth_txg = 0;
960 * We don't need any locking to protect db_blkptr:
961 * If it's syncing, then db_last_dirty will be set
962 * so we'll ignore db_blkptr.
964 ASSERT(MUTEX_HELD(&db->db_mtx));
965 if (db->db_last_dirty)
966 birth_txg = db->db_last_dirty->dr_txg;
967 else if (db->db_blkptr)
968 birth_txg = db->db_blkptr->blk_birth;
971 * If we don't exist or are in a snapshot, we can't be freed.
972 * Don't pass the bp to dsl_dataset_block_freeable() since we
973 * are holding the db_mtx lock and might deadlock if we are
974 * prefetching a dedup-ed block.
977 return (ds == NULL ||
978 dsl_dataset_block_freeable(ds, NULL, birth_txg));
984 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
986 arc_buf_t *buf, *obuf;
987 int osize = db->db.db_size;
988 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
991 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
996 /* XXX does *this* func really need the lock? */
997 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1000 * This call to dbuf_will_dirty() with the dn_struct_rwlock held
1001 * is OK, because there can be no other references to the db
1002 * when we are changing its size, so no concurrent DB_FILL can
1006 * XXX we should be doing a dbuf_read, checking the return
1007 * value and returning that up to our callers
1009 dbuf_will_dirty(db, tx);
1011 /* create the data buffer for the new block */
1012 buf = arc_buf_alloc(dn->dn_objset->os_spa, size, db, type);
1014 /* copy old block data to the new block */
1016 bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
1017 /* zero the remainder */
1019 bzero((uint8_t *)buf->b_data + osize, size - osize);
1021 mutex_enter(&db->db_mtx);
1022 dbuf_set_data(db, buf);
1023 VERIFY(arc_buf_remove_ref(obuf, db));
1024 db->db.db_size = size;
1026 if (db->db_level == 0) {
1027 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1028 db->db_last_dirty->dt.dl.dr_data = buf;
1030 mutex_exit(&db->db_mtx);
1032 dnode_willuse_space(dn, size-osize, tx);
1037 dbuf_release_bp(dmu_buf_impl_t *db)
1041 DB_GET_OBJSET(&os, db);
1042 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1043 ASSERT(arc_released(os->os_phys_buf) ||
1044 list_link_active(&os->os_dsl_dataset->ds_synced_link));
1045 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1047 (void) arc_release(db->db_buf, db);
1050 dbuf_dirty_record_t *
1051 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1055 dbuf_dirty_record_t **drp, *dr;
1056 int drop_struct_lock = FALSE;
1057 boolean_t do_free_accounting = B_FALSE;
1058 int txgoff = tx->tx_txg & TXG_MASK;
1060 ASSERT(tx->tx_txg != 0);
1061 ASSERT(!refcount_is_zero(&db->db_holds));
1062 DMU_TX_DIRTY_BUF(tx, db);
1067 * Shouldn't dirty a regular buffer in syncing context. Private
1068 * objects may be dirtied in syncing context, but only if they
1069 * were already pre-dirtied in open context.
1071 ASSERT(!dmu_tx_is_syncing(tx) ||
1072 BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1073 DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1074 dn->dn_objset->os_dsl_dataset == NULL);
1076 * We make this assert for private objects as well, but after we
1077 * check if we're already dirty. They are allowed to re-dirty
1078 * in syncing context.
1080 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1081 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1082 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1084 mutex_enter(&db->db_mtx);
1086 * XXX make this true for indirects too? The problem is that
1087 * transactions created with dmu_tx_create_assigned() from
1088 * syncing context don't bother holding ahead.
1090 ASSERT(db->db_level != 0 ||
1091 db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1092 db->db_state == DB_NOFILL);
1094 mutex_enter(&dn->dn_mtx);
1096 * Don't set dirtyctx to SYNC if we're just modifying this as we
1097 * initialize the objset.
1099 if (dn->dn_dirtyctx == DN_UNDIRTIED &&
1100 !BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1102 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1103 ASSERT(dn->dn_dirtyctx_firstset == NULL);
1104 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_PUSHPAGE);
1106 mutex_exit(&dn->dn_mtx);
1108 if (db->db_blkid == DMU_SPILL_BLKID)
1109 dn->dn_have_spill = B_TRUE;
1112 * If this buffer is already dirty, we're done.
1114 drp = &db->db_last_dirty;
1115 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1116 db->db.db_object == DMU_META_DNODE_OBJECT);
1117 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1119 if (dr && dr->dr_txg == tx->tx_txg) {
1122 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1124 * If this buffer has already been written out,
1125 * we now need to reset its state.
1127 dbuf_unoverride(dr);
1128 if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1129 db->db_state != DB_NOFILL)
1130 arc_buf_thaw(db->db_buf);
1132 mutex_exit(&db->db_mtx);
1137 * Only valid if not already dirty.
1139 ASSERT(dn->dn_object == 0 ||
1140 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1141 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1143 ASSERT3U(dn->dn_nlevels, >, db->db_level);
1144 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1145 dn->dn_phys->dn_nlevels > db->db_level ||
1146 dn->dn_next_nlevels[txgoff] > db->db_level ||
1147 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1148 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1151 * We should only be dirtying in syncing context if it's the
1152 * mos or we're initializing the os or it's a special object.
1153 * However, we are allowed to dirty in syncing context provided
1154 * we already dirtied it in open context. Hence we must make
1155 * this assertion only if we're not already dirty.
1158 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1159 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1160 ASSERT(db->db.db_size != 0);
1162 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1164 if (db->db_blkid != DMU_BONUS_BLKID) {
1166 * Update the accounting.
1167 * Note: we delay "free accounting" until after we drop
1168 * the db_mtx. This keeps us from grabbing other locks
1169 * (and possibly deadlocking) in bp_get_dsize() while
1170 * also holding the db_mtx.
1172 dnode_willuse_space(dn, db->db.db_size, tx);
1173 do_free_accounting = dbuf_block_freeable(db);
1177 * If this buffer is dirty in an old transaction group we need
1178 * to make a copy of it so that the changes we make in this
1179 * transaction group won't leak out when we sync the older txg.
1181 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_PUSHPAGE);
1182 list_link_init(&dr->dr_dirty_node);
1183 if (db->db_level == 0) {
1184 void *data_old = db->db_buf;
1186 if (db->db_state != DB_NOFILL) {
1187 if (db->db_blkid == DMU_BONUS_BLKID) {
1188 dbuf_fix_old_data(db, tx->tx_txg);
1189 data_old = db->db.db_data;
1190 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1192 * Release the data buffer from the cache so
1193 * that we can modify it without impacting
1194 * possible other users of this cached data
1195 * block. Note that indirect blocks and
1196 * private objects are not released until the
1197 * syncing state (since they are only modified
1200 arc_release(db->db_buf, db);
1201 dbuf_fix_old_data(db, tx->tx_txg);
1202 data_old = db->db_buf;
1204 ASSERT(data_old != NULL);
1206 dr->dt.dl.dr_data = data_old;
1208 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1209 list_create(&dr->dt.di.dr_children,
1210 sizeof (dbuf_dirty_record_t),
1211 offsetof(dbuf_dirty_record_t, dr_dirty_node));
1214 dr->dr_txg = tx->tx_txg;
1219 * We could have been freed_in_flight between the dbuf_noread
1220 * and dbuf_dirty. We win, as though the dbuf_noread() had
1221 * happened after the free.
1223 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1224 db->db_blkid != DMU_SPILL_BLKID) {
1225 mutex_enter(&dn->dn_mtx);
1226 dnode_clear_range(dn, db->db_blkid, 1, tx);
1227 mutex_exit(&dn->dn_mtx);
1228 db->db_freed_in_flight = FALSE;
1232 * This buffer is now part of this txg
1234 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1235 db->db_dirtycnt += 1;
1236 ASSERT3U(db->db_dirtycnt, <=, 3);
1238 mutex_exit(&db->db_mtx);
1240 if (db->db_blkid == DMU_BONUS_BLKID ||
1241 db->db_blkid == DMU_SPILL_BLKID) {
1242 mutex_enter(&dn->dn_mtx);
1243 ASSERT(!list_link_active(&dr->dr_dirty_node));
1244 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1245 mutex_exit(&dn->dn_mtx);
1246 dnode_setdirty(dn, tx);
1249 } else if (do_free_accounting) {
1250 blkptr_t *bp = db->db_blkptr;
1251 int64_t willfree = (bp && !BP_IS_HOLE(bp)) ?
1252 bp_get_dsize(os->os_spa, bp) : db->db.db_size;
1254 * This is only a guess -- if the dbuf is dirty
1255 * in a previous txg, we don't know how much
1256 * space it will use on disk yet. We should
1257 * really have the struct_rwlock to access
1258 * db_blkptr, but since this is just a guess,
1259 * it's OK if we get an odd answer.
1261 ddt_prefetch(os->os_spa, bp);
1262 dnode_willuse_space(dn, -willfree, tx);
1265 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1266 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1267 drop_struct_lock = TRUE;
1270 if (db->db_level == 0) {
1271 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1272 ASSERT(dn->dn_maxblkid >= db->db_blkid);
1275 if (db->db_level+1 < dn->dn_nlevels) {
1276 dmu_buf_impl_t *parent = db->db_parent;
1277 dbuf_dirty_record_t *di;
1278 int parent_held = FALSE;
1280 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1281 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1283 parent = dbuf_hold_level(dn, db->db_level+1,
1284 db->db_blkid >> epbs, FTAG);
1285 ASSERT(parent != NULL);
1288 if (drop_struct_lock)
1289 rw_exit(&dn->dn_struct_rwlock);
1290 ASSERT3U(db->db_level+1, ==, parent->db_level);
1291 di = dbuf_dirty(parent, tx);
1293 dbuf_rele(parent, FTAG);
1295 mutex_enter(&db->db_mtx);
1296 /* possible race with dbuf_undirty() */
1297 if (db->db_last_dirty == dr ||
1298 dn->dn_object == DMU_META_DNODE_OBJECT) {
1299 mutex_enter(&di->dt.di.dr_mtx);
1300 ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1301 ASSERT(!list_link_active(&dr->dr_dirty_node));
1302 list_insert_tail(&di->dt.di.dr_children, dr);
1303 mutex_exit(&di->dt.di.dr_mtx);
1306 mutex_exit(&db->db_mtx);
1308 ASSERT(db->db_level+1 == dn->dn_nlevels);
1309 ASSERT(db->db_blkid < dn->dn_nblkptr);
1310 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1311 mutex_enter(&dn->dn_mtx);
1312 ASSERT(!list_link_active(&dr->dr_dirty_node));
1313 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1314 mutex_exit(&dn->dn_mtx);
1315 if (drop_struct_lock)
1316 rw_exit(&dn->dn_struct_rwlock);
1319 dnode_setdirty(dn, tx);
1325 * Undirty a buffer in the transaction group referenced by the given
1326 * transaction. Return whether this evicted the dbuf.
1329 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1332 uint64_t txg = tx->tx_txg;
1333 dbuf_dirty_record_t *dr, **drp;
1336 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1337 ASSERT0(db->db_level);
1338 ASSERT(MUTEX_HELD(&db->db_mtx));
1341 * If this buffer is not dirty, we're done.
1343 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1344 if (dr->dr_txg <= txg)
1346 if (dr == NULL || dr->dr_txg < txg)
1348 ASSERT(dr->dr_txg == txg);
1349 ASSERT(dr->dr_dbuf == db);
1355 * Note: This code will probably work even if there are concurrent
1356 * holders, but it is untested in that scenerio, as the ZPL and
1357 * ztest have additional locking (the range locks) that prevents
1358 * that type of concurrent access.
1360 ASSERT3U(refcount_count(&db->db_holds), ==, db->db_dirtycnt);
1362 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1364 ASSERT(db->db.db_size != 0);
1366 /* XXX would be nice to fix up dn_towrite_space[] */
1371 * Note that there are three places in dbuf_dirty()
1372 * where this dirty record may be put on a list.
1373 * Make sure to do a list_remove corresponding to
1374 * every one of those list_insert calls.
1376 if (dr->dr_parent) {
1377 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1378 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1379 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1380 } else if (db->db_blkid == DMU_SPILL_BLKID ||
1381 db->db_level+1 == dn->dn_nlevels) {
1382 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1383 mutex_enter(&dn->dn_mtx);
1384 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1385 mutex_exit(&dn->dn_mtx);
1389 if (db->db_state != DB_NOFILL) {
1390 dbuf_unoverride(dr);
1392 ASSERT(db->db_buf != NULL);
1393 ASSERT(dr->dt.dl.dr_data != NULL);
1394 if (dr->dt.dl.dr_data != db->db_buf)
1395 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db));
1397 kmem_free(dr, sizeof (dbuf_dirty_record_t));
1399 ASSERT(db->db_dirtycnt > 0);
1400 db->db_dirtycnt -= 1;
1402 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1403 arc_buf_t *buf = db->db_buf;
1405 ASSERT(db->db_state == DB_NOFILL || arc_released(buf));
1406 dbuf_set_data(db, NULL);
1407 VERIFY(arc_buf_remove_ref(buf, db));
1415 #pragma weak dmu_buf_will_dirty = dbuf_will_dirty
1417 dbuf_will_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1419 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1421 ASSERT(tx->tx_txg != 0);
1422 ASSERT(!refcount_is_zero(&db->db_holds));
1425 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
1426 rf |= DB_RF_HAVESTRUCT;
1428 (void) dbuf_read(db, NULL, rf);
1429 (void) dbuf_dirty(db, tx);
1433 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1435 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1437 db->db_state = DB_NOFILL;
1439 dmu_buf_will_fill(db_fake, tx);
1443 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1445 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1447 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1448 ASSERT(tx->tx_txg != 0);
1449 ASSERT(db->db_level == 0);
1450 ASSERT(!refcount_is_zero(&db->db_holds));
1452 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1453 dmu_tx_private_ok(tx));
1456 (void) dbuf_dirty(db, tx);
1459 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1462 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1464 mutex_enter(&db->db_mtx);
1467 if (db->db_state == DB_FILL) {
1468 if (db->db_level == 0 && db->db_freed_in_flight) {
1469 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1470 /* we were freed while filling */
1471 /* XXX dbuf_undirty? */
1472 bzero(db->db.db_data, db->db.db_size);
1473 db->db_freed_in_flight = FALSE;
1475 db->db_state = DB_CACHED;
1476 cv_broadcast(&db->db_changed);
1478 mutex_exit(&db->db_mtx);
1482 * Directly assign a provided arc buf to a given dbuf if it's not referenced
1483 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1486 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
1488 ASSERT(!refcount_is_zero(&db->db_holds));
1489 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1490 ASSERT(db->db_level == 0);
1491 ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA);
1492 ASSERT(buf != NULL);
1493 ASSERT(arc_buf_size(buf) == db->db.db_size);
1494 ASSERT(tx->tx_txg != 0);
1496 arc_return_buf(buf, db);
1497 ASSERT(arc_released(buf));
1499 mutex_enter(&db->db_mtx);
1501 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1502 cv_wait(&db->db_changed, &db->db_mtx);
1504 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
1506 if (db->db_state == DB_CACHED &&
1507 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
1508 mutex_exit(&db->db_mtx);
1509 (void) dbuf_dirty(db, tx);
1510 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
1511 VERIFY(arc_buf_remove_ref(buf, db));
1512 xuio_stat_wbuf_copied();
1516 xuio_stat_wbuf_nocopy();
1517 if (db->db_state == DB_CACHED) {
1518 dbuf_dirty_record_t *dr = db->db_last_dirty;
1520 ASSERT(db->db_buf != NULL);
1521 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
1522 ASSERT(dr->dt.dl.dr_data == db->db_buf);
1523 if (!arc_released(db->db_buf)) {
1524 ASSERT(dr->dt.dl.dr_override_state ==
1526 arc_release(db->db_buf, db);
1528 dr->dt.dl.dr_data = buf;
1529 VERIFY(arc_buf_remove_ref(db->db_buf, db));
1530 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
1531 arc_release(db->db_buf, db);
1532 VERIFY(arc_buf_remove_ref(db->db_buf, db));
1536 ASSERT(db->db_buf == NULL);
1537 dbuf_set_data(db, buf);
1538 db->db_state = DB_FILL;
1539 mutex_exit(&db->db_mtx);
1540 (void) dbuf_dirty(db, tx);
1541 dbuf_fill_done(db, tx);
1545 * "Clear" the contents of this dbuf. This will mark the dbuf
1546 * EVICTING and clear *most* of its references. Unfortunetely,
1547 * when we are not holding the dn_dbufs_mtx, we can't clear the
1548 * entry in the dn_dbufs list. We have to wait until dbuf_destroy()
1549 * in this case. For callers from the DMU we will usually see:
1550 * dbuf_clear()->arc_buf_evict()->dbuf_do_evict()->dbuf_destroy()
1551 * For the arc callback, we will usually see:
1552 * dbuf_do_evict()->dbuf_clear();dbuf_destroy()
1553 * Sometimes, though, we will get a mix of these two:
1554 * DMU: dbuf_clear()->arc_buf_evict()
1555 * ARC: dbuf_do_evict()->dbuf_destroy()
1558 dbuf_clear(dmu_buf_impl_t *db)
1561 dmu_buf_impl_t *parent = db->db_parent;
1562 dmu_buf_impl_t *dndb;
1563 int dbuf_gone = FALSE;
1565 ASSERT(MUTEX_HELD(&db->db_mtx));
1566 ASSERT(refcount_is_zero(&db->db_holds));
1568 dbuf_evict_user(db);
1570 if (db->db_state == DB_CACHED) {
1571 ASSERT(db->db.db_data != NULL);
1572 if (db->db_blkid == DMU_BONUS_BLKID) {
1573 zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
1574 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
1576 db->db.db_data = NULL;
1577 db->db_state = DB_UNCACHED;
1580 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
1581 ASSERT(db->db_data_pending == NULL);
1583 db->db_state = DB_EVICTING;
1584 db->db_blkptr = NULL;
1589 if (db->db_blkid != DMU_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) {
1590 list_remove(&dn->dn_dbufs, db);
1591 (void) atomic_dec_32_nv(&dn->dn_dbufs_count);
1595 * Decrementing the dbuf count means that the hold corresponding
1596 * to the removed dbuf is no longer discounted in dnode_move(),
1597 * so the dnode cannot be moved until after we release the hold.
1598 * The membar_producer() ensures visibility of the decremented
1599 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
1603 db->db_dnode_handle = NULL;
1609 dbuf_gone = arc_buf_evict(db->db_buf);
1612 mutex_exit(&db->db_mtx);
1615 * If this dbuf is referenced from an indirect dbuf,
1616 * decrement the ref count on the indirect dbuf.
1618 if (parent && parent != dndb)
1619 dbuf_rele(parent, db);
1622 __attribute__((always_inline))
1624 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
1625 dmu_buf_impl_t **parentp, blkptr_t **bpp, struct dbuf_hold_impl_data *dh)
1632 ASSERT(blkid != DMU_BONUS_BLKID);
1634 if (blkid == DMU_SPILL_BLKID) {
1635 mutex_enter(&dn->dn_mtx);
1636 if (dn->dn_have_spill &&
1637 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
1638 *bpp = &dn->dn_phys->dn_spill;
1641 dbuf_add_ref(dn->dn_dbuf, NULL);
1642 *parentp = dn->dn_dbuf;
1643 mutex_exit(&dn->dn_mtx);
1647 if (dn->dn_phys->dn_nlevels == 0)
1650 nlevels = dn->dn_phys->dn_nlevels;
1652 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1654 ASSERT3U(level * epbs, <, 64);
1655 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1656 if (level >= nlevels ||
1657 (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
1658 /* the buffer has no parent yet */
1659 return (SET_ERROR(ENOENT));
1660 } else if (level < nlevels-1) {
1661 /* this block is referenced from an indirect block */
1664 err = dbuf_hold_impl(dn, level+1, blkid >> epbs,
1665 fail_sparse, NULL, parentp);
1668 __dbuf_hold_impl_init(dh + 1, dn, dh->dh_level + 1,
1669 blkid >> epbs, fail_sparse, NULL,
1670 parentp, dh->dh_depth + 1);
1671 err = __dbuf_hold_impl(dh + 1);
1675 err = dbuf_read(*parentp, NULL,
1676 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
1678 dbuf_rele(*parentp, NULL);
1682 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
1683 (blkid & ((1ULL << epbs) - 1));
1686 /* the block is referenced from the dnode */
1687 ASSERT3U(level, ==, nlevels-1);
1688 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
1689 blkid < dn->dn_phys->dn_nblkptr);
1691 dbuf_add_ref(dn->dn_dbuf, NULL);
1692 *parentp = dn->dn_dbuf;
1694 *bpp = &dn->dn_phys->dn_blkptr[blkid];
1699 static dmu_buf_impl_t *
1700 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
1701 dmu_buf_impl_t *parent, blkptr_t *blkptr)
1703 objset_t *os = dn->dn_objset;
1704 dmu_buf_impl_t *db, *odb;
1706 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1707 ASSERT(dn->dn_type != DMU_OT_NONE);
1709 db = kmem_cache_alloc(dbuf_cache, KM_PUSHPAGE);
1712 db->db.db_object = dn->dn_object;
1713 db->db_level = level;
1714 db->db_blkid = blkid;
1715 db->db_last_dirty = NULL;
1716 db->db_dirtycnt = 0;
1717 db->db_dnode_handle = dn->dn_handle;
1718 db->db_parent = parent;
1719 db->db_blkptr = blkptr;
1721 db->db_user_ptr = NULL;
1722 db->db_user_data_ptr_ptr = NULL;
1723 db->db_evict_func = NULL;
1724 db->db_immediate_evict = 0;
1725 db->db_freed_in_flight = 0;
1727 if (blkid == DMU_BONUS_BLKID) {
1728 ASSERT3P(parent, ==, dn->dn_dbuf);
1729 db->db.db_size = DN_MAX_BONUSLEN -
1730 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
1731 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
1732 db->db.db_offset = DMU_BONUS_BLKID;
1733 db->db_state = DB_UNCACHED;
1734 /* the bonus dbuf is not placed in the hash table */
1735 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1737 } else if (blkid == DMU_SPILL_BLKID) {
1738 db->db.db_size = (blkptr != NULL) ?
1739 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
1740 db->db.db_offset = 0;
1743 db->db_level ? 1<<dn->dn_indblkshift : dn->dn_datablksz;
1744 db->db.db_size = blocksize;
1745 db->db.db_offset = db->db_blkid * blocksize;
1749 * Hold the dn_dbufs_mtx while we get the new dbuf
1750 * in the hash table *and* added to the dbufs list.
1751 * This prevents a possible deadlock with someone
1752 * trying to look up this dbuf before its added to the
1755 mutex_enter(&dn->dn_dbufs_mtx);
1756 db->db_state = DB_EVICTING;
1757 if ((odb = dbuf_hash_insert(db)) != NULL) {
1758 /* someone else inserted it first */
1759 kmem_cache_free(dbuf_cache, db);
1760 mutex_exit(&dn->dn_dbufs_mtx);
1763 list_insert_head(&dn->dn_dbufs, db);
1764 db->db_state = DB_UNCACHED;
1765 mutex_exit(&dn->dn_dbufs_mtx);
1766 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1768 if (parent && parent != dn->dn_dbuf)
1769 dbuf_add_ref(parent, db);
1771 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1772 refcount_count(&dn->dn_holds) > 0);
1773 (void) refcount_add(&dn->dn_holds, db);
1774 (void) atomic_inc_32_nv(&dn->dn_dbufs_count);
1776 dprintf_dbuf(db, "db=%p\n", db);
1782 dbuf_do_evict(void *private)
1784 arc_buf_t *buf = private;
1785 dmu_buf_impl_t *db = buf->b_private;
1787 if (!MUTEX_HELD(&db->db_mtx))
1788 mutex_enter(&db->db_mtx);
1790 ASSERT(refcount_is_zero(&db->db_holds));
1792 if (db->db_state != DB_EVICTING) {
1793 ASSERT(db->db_state == DB_CACHED);
1798 mutex_exit(&db->db_mtx);
1805 dbuf_destroy(dmu_buf_impl_t *db)
1807 ASSERT(refcount_is_zero(&db->db_holds));
1809 if (db->db_blkid != DMU_BONUS_BLKID) {
1811 * If this dbuf is still on the dn_dbufs list,
1812 * remove it from that list.
1814 if (db->db_dnode_handle != NULL) {
1819 mutex_enter(&dn->dn_dbufs_mtx);
1820 list_remove(&dn->dn_dbufs, db);
1821 (void) atomic_dec_32_nv(&dn->dn_dbufs_count);
1822 mutex_exit(&dn->dn_dbufs_mtx);
1825 * Decrementing the dbuf count means that the hold
1826 * corresponding to the removed dbuf is no longer
1827 * discounted in dnode_move(), so the dnode cannot be
1828 * moved until after we release the hold.
1831 db->db_dnode_handle = NULL;
1833 dbuf_hash_remove(db);
1835 db->db_parent = NULL;
1838 ASSERT(!list_link_active(&db->db_link));
1839 ASSERT(db->db.db_data == NULL);
1840 ASSERT(db->db_hash_next == NULL);
1841 ASSERT(db->db_blkptr == NULL);
1842 ASSERT(db->db_data_pending == NULL);
1844 kmem_cache_free(dbuf_cache, db);
1845 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1849 dbuf_prefetch(dnode_t *dn, uint64_t blkid)
1851 dmu_buf_impl_t *db = NULL;
1852 blkptr_t *bp = NULL;
1854 ASSERT(blkid != DMU_BONUS_BLKID);
1855 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1857 if (dnode_block_freed(dn, blkid))
1860 /* dbuf_find() returns with db_mtx held */
1861 if ((db = dbuf_find(dn, 0, blkid))) {
1863 * This dbuf is already in the cache. We assume that
1864 * it is already CACHED, or else about to be either
1867 mutex_exit(&db->db_mtx);
1871 if (dbuf_findbp(dn, 0, blkid, TRUE, &db, &bp, NULL) == 0) {
1872 if (bp && !BP_IS_HOLE(bp)) {
1873 int priority = dn->dn_type == DMU_OT_DDT_ZAP ?
1874 ZIO_PRIORITY_DDT_PREFETCH : ZIO_PRIORITY_ASYNC_READ;
1875 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
1876 uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH;
1879 SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
1880 dn->dn_object, 0, blkid);
1882 (void) arc_read(NULL, dn->dn_objset->os_spa,
1883 bp, NULL, NULL, priority,
1884 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1888 dbuf_rele(db, NULL);
1892 #define DBUF_HOLD_IMPL_MAX_DEPTH 20
1895 * Returns with db_holds incremented, and db_mtx not held.
1896 * Note: dn_struct_rwlock must be held.
1899 __dbuf_hold_impl(struct dbuf_hold_impl_data *dh)
1901 ASSERT3S(dh->dh_depth, <, DBUF_HOLD_IMPL_MAX_DEPTH);
1902 dh->dh_parent = NULL;
1904 ASSERT(dh->dh_blkid != DMU_BONUS_BLKID);
1905 ASSERT(RW_LOCK_HELD(&dh->dh_dn->dn_struct_rwlock));
1906 ASSERT3U(dh->dh_dn->dn_nlevels, >, dh->dh_level);
1908 *(dh->dh_dbp) = NULL;
1910 /* dbuf_find() returns with db_mtx held */
1911 dh->dh_db = dbuf_find(dh->dh_dn, dh->dh_level, dh->dh_blkid);
1913 if (dh->dh_db == NULL) {
1916 ASSERT3P(dh->dh_parent, ==, NULL);
1917 dh->dh_err = dbuf_findbp(dh->dh_dn, dh->dh_level, dh->dh_blkid,
1918 dh->dh_fail_sparse, &dh->dh_parent,
1920 if (dh->dh_fail_sparse) {
1921 if (dh->dh_err == 0 && dh->dh_bp && BP_IS_HOLE(dh->dh_bp))
1922 dh->dh_err = SET_ERROR(ENOENT);
1925 dbuf_rele(dh->dh_parent, NULL);
1926 return (dh->dh_err);
1929 if (dh->dh_err && dh->dh_err != ENOENT)
1930 return (dh->dh_err);
1931 dh->dh_db = dbuf_create(dh->dh_dn, dh->dh_level, dh->dh_blkid,
1932 dh->dh_parent, dh->dh_bp);
1935 if (dh->dh_db->db_buf && refcount_is_zero(&dh->dh_db->db_holds)) {
1936 arc_buf_add_ref(dh->dh_db->db_buf, dh->dh_db);
1937 if (dh->dh_db->db_buf->b_data == NULL) {
1938 dbuf_clear(dh->dh_db);
1939 if (dh->dh_parent) {
1940 dbuf_rele(dh->dh_parent, NULL);
1941 dh->dh_parent = NULL;
1945 ASSERT3P(dh->dh_db->db.db_data, ==, dh->dh_db->db_buf->b_data);
1948 ASSERT(dh->dh_db->db_buf == NULL || arc_referenced(dh->dh_db->db_buf));
1951 * If this buffer is currently syncing out, and we are are
1952 * still referencing it from db_data, we need to make a copy
1953 * of it in case we decide we want to dirty it again in this txg.
1955 if (dh->dh_db->db_level == 0 &&
1956 dh->dh_db->db_blkid != DMU_BONUS_BLKID &&
1957 dh->dh_dn->dn_object != DMU_META_DNODE_OBJECT &&
1958 dh->dh_db->db_state == DB_CACHED && dh->dh_db->db_data_pending) {
1959 dh->dh_dr = dh->dh_db->db_data_pending;
1961 if (dh->dh_dr->dt.dl.dr_data == dh->dh_db->db_buf) {
1962 dh->dh_type = DBUF_GET_BUFC_TYPE(dh->dh_db);
1964 dbuf_set_data(dh->dh_db,
1965 arc_buf_alloc(dh->dh_dn->dn_objset->os_spa,
1966 dh->dh_db->db.db_size, dh->dh_db, dh->dh_type));
1967 bcopy(dh->dh_dr->dt.dl.dr_data->b_data,
1968 dh->dh_db->db.db_data, dh->dh_db->db.db_size);
1972 (void) refcount_add(&dh->dh_db->db_holds, dh->dh_tag);
1973 dbuf_update_data(dh->dh_db);
1974 DBUF_VERIFY(dh->dh_db);
1975 mutex_exit(&dh->dh_db->db_mtx);
1977 /* NOTE: we can't rele the parent until after we drop the db_mtx */
1979 dbuf_rele(dh->dh_parent, NULL);
1981 ASSERT3P(DB_DNODE(dh->dh_db), ==, dh->dh_dn);
1982 ASSERT3U(dh->dh_db->db_blkid, ==, dh->dh_blkid);
1983 ASSERT3U(dh->dh_db->db_level, ==, dh->dh_level);
1984 *(dh->dh_dbp) = dh->dh_db;
1990 * The following code preserves the recursive function dbuf_hold_impl()
1991 * but moves the local variables AND function arguments to the heap to
1992 * minimize the stack frame size. Enough space is initially allocated
1993 * on the stack for 20 levels of recursion.
1996 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse,
1997 void *tag, dmu_buf_impl_t **dbp)
1999 struct dbuf_hold_impl_data *dh;
2002 dh = kmem_zalloc(sizeof(struct dbuf_hold_impl_data) *
2003 DBUF_HOLD_IMPL_MAX_DEPTH, KM_PUSHPAGE);
2004 __dbuf_hold_impl_init(dh, dn, level, blkid, fail_sparse, tag, dbp, 0);
2006 error = __dbuf_hold_impl(dh);
2008 kmem_free(dh, sizeof(struct dbuf_hold_impl_data) *
2009 DBUF_HOLD_IMPL_MAX_DEPTH);
2015 __dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh,
2016 dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse,
2017 void *tag, dmu_buf_impl_t **dbp, int depth)
2020 dh->dh_level = level;
2021 dh->dh_blkid = blkid;
2022 dh->dh_fail_sparse = fail_sparse;
2025 dh->dh_depth = depth;
2029 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
2032 int err = dbuf_hold_impl(dn, 0, blkid, FALSE, tag, &db);
2033 return (err ? NULL : db);
2037 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
2040 int err = dbuf_hold_impl(dn, level, blkid, FALSE, tag, &db);
2041 return (err ? NULL : db);
2045 dbuf_create_bonus(dnode_t *dn)
2047 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
2049 ASSERT(dn->dn_bonus == NULL);
2050 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
2054 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
2056 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2059 if (db->db_blkid != DMU_SPILL_BLKID)
2060 return (SET_ERROR(ENOTSUP));
2062 blksz = SPA_MINBLOCKSIZE;
2063 if (blksz > SPA_MAXBLOCKSIZE)
2064 blksz = SPA_MAXBLOCKSIZE;
2066 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
2070 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2071 dbuf_new_size(db, blksz, tx);
2072 rw_exit(&dn->dn_struct_rwlock);
2079 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
2081 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
2084 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2086 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
2088 VERIFY(refcount_add(&db->db_holds, tag) > 1);
2092 * If you call dbuf_rele() you had better not be referencing the dnode handle
2093 * unless you have some other direct or indirect hold on the dnode. (An indirect
2094 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2095 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2096 * dnode's parent dbuf evicting its dnode handles.
2098 #pragma weak dmu_buf_rele = dbuf_rele
2100 dbuf_rele(dmu_buf_impl_t *db, void *tag)
2102 mutex_enter(&db->db_mtx);
2103 dbuf_rele_and_unlock(db, tag);
2107 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2108 * db_dirtycnt and db_holds to be updated atomically.
2111 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2115 ASSERT(MUTEX_HELD(&db->db_mtx));
2119 * Remove the reference to the dbuf before removing its hold on the
2120 * dnode so we can guarantee in dnode_move() that a referenced bonus
2121 * buffer has a corresponding dnode hold.
2123 holds = refcount_remove(&db->db_holds, tag);
2127 * We can't freeze indirects if there is a possibility that they
2128 * may be modified in the current syncing context.
2130 if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0))
2131 arc_buf_freeze(db->db_buf);
2133 if (holds == db->db_dirtycnt &&
2134 db->db_level == 0 && db->db_immediate_evict)
2135 dbuf_evict_user(db);
2138 if (db->db_blkid == DMU_BONUS_BLKID) {
2139 mutex_exit(&db->db_mtx);
2142 * If the dnode moves here, we cannot cross this barrier
2143 * until the move completes.
2146 (void) atomic_dec_32_nv(&DB_DNODE(db)->dn_dbufs_count);
2149 * The bonus buffer's dnode hold is no longer discounted
2150 * in dnode_move(). The dnode cannot move until after
2153 dnode_rele(DB_DNODE(db), db);
2154 } else if (db->db_buf == NULL) {
2156 * This is a special case: we never associated this
2157 * dbuf with any data allocated from the ARC.
2159 ASSERT(db->db_state == DB_UNCACHED ||
2160 db->db_state == DB_NOFILL);
2162 } else if (arc_released(db->db_buf)) {
2163 arc_buf_t *buf = db->db_buf;
2165 * This dbuf has anonymous data associated with it.
2167 dbuf_set_data(db, NULL);
2168 VERIFY(arc_buf_remove_ref(buf, db));
2171 VERIFY(!arc_buf_remove_ref(db->db_buf, db));
2174 * A dbuf will be eligible for eviction if either the
2175 * 'primarycache' property is set or a duplicate
2176 * copy of this buffer is already cached in the arc.
2178 * In the case of the 'primarycache' a buffer
2179 * is considered for eviction if it matches the
2180 * criteria set in the property.
2182 * To decide if our buffer is considered a
2183 * duplicate, we must call into the arc to determine
2184 * if multiple buffers are referencing the same
2185 * block on-disk. If so, then we simply evict
2188 if (!DBUF_IS_CACHEABLE(db) ||
2189 arc_buf_eviction_needed(db->db_buf))
2192 mutex_exit(&db->db_mtx);
2195 mutex_exit(&db->db_mtx);
2199 #pragma weak dmu_buf_refcount = dbuf_refcount
2201 dbuf_refcount(dmu_buf_impl_t *db)
2203 return (refcount_count(&db->db_holds));
2207 dmu_buf_set_user(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr,
2208 dmu_buf_evict_func_t *evict_func)
2210 return (dmu_buf_update_user(db_fake, NULL, user_ptr,
2211 user_data_ptr_ptr, evict_func));
2215 dmu_buf_set_user_ie(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr,
2216 dmu_buf_evict_func_t *evict_func)
2218 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2220 db->db_immediate_evict = TRUE;
2221 return (dmu_buf_update_user(db_fake, NULL, user_ptr,
2222 user_data_ptr_ptr, evict_func));
2226 dmu_buf_update_user(dmu_buf_t *db_fake, void *old_user_ptr, void *user_ptr,
2227 void *user_data_ptr_ptr, dmu_buf_evict_func_t *evict_func)
2229 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2230 ASSERT(db->db_level == 0);
2232 ASSERT((user_ptr == NULL) == (evict_func == NULL));
2234 mutex_enter(&db->db_mtx);
2236 if (db->db_user_ptr == old_user_ptr) {
2237 db->db_user_ptr = user_ptr;
2238 db->db_user_data_ptr_ptr = user_data_ptr_ptr;
2239 db->db_evict_func = evict_func;
2241 dbuf_update_data(db);
2243 old_user_ptr = db->db_user_ptr;
2246 mutex_exit(&db->db_mtx);
2247 return (old_user_ptr);
2251 dmu_buf_get_user(dmu_buf_t *db_fake)
2253 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2254 ASSERT(!refcount_is_zero(&db->db_holds));
2256 return (db->db_user_ptr);
2260 dmu_buf_freeable(dmu_buf_t *dbuf)
2262 boolean_t res = B_FALSE;
2263 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2266 res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset,
2267 db->db_blkptr, db->db_blkptr->blk_birth);
2273 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
2275 /* ASSERT(dmu_tx_is_syncing(tx) */
2276 ASSERT(MUTEX_HELD(&db->db_mtx));
2278 if (db->db_blkptr != NULL)
2281 if (db->db_blkid == DMU_SPILL_BLKID) {
2282 db->db_blkptr = &dn->dn_phys->dn_spill;
2283 BP_ZERO(db->db_blkptr);
2286 if (db->db_level == dn->dn_phys->dn_nlevels-1) {
2288 * This buffer was allocated at a time when there was
2289 * no available blkptrs from the dnode, or it was
2290 * inappropriate to hook it in (i.e., nlevels mis-match).
2292 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
2293 ASSERT(db->db_parent == NULL);
2294 db->db_parent = dn->dn_dbuf;
2295 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
2298 dmu_buf_impl_t *parent = db->db_parent;
2299 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2301 ASSERT(dn->dn_phys->dn_nlevels > 1);
2302 if (parent == NULL) {
2303 mutex_exit(&db->db_mtx);
2304 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2305 (void) dbuf_hold_impl(dn, db->db_level+1,
2306 db->db_blkid >> epbs, FALSE, db, &parent);
2307 rw_exit(&dn->dn_struct_rwlock);
2308 mutex_enter(&db->db_mtx);
2309 db->db_parent = parent;
2311 db->db_blkptr = (blkptr_t *)parent->db.db_data +
2312 (db->db_blkid & ((1ULL << epbs) - 1));
2317 /* dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it
2318 * is critical the we not allow the compiler to inline this function in to
2319 * dbuf_sync_list() thereby drastically bloating the stack usage.
2321 noinline static void
2322 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2324 dmu_buf_impl_t *db = dr->dr_dbuf;
2328 ASSERT(dmu_tx_is_syncing(tx));
2330 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2332 mutex_enter(&db->db_mtx);
2334 ASSERT(db->db_level > 0);
2337 /* Read the block if it hasn't been read yet. */
2338 if (db->db_buf == NULL) {
2339 mutex_exit(&db->db_mtx);
2340 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
2341 mutex_enter(&db->db_mtx);
2343 ASSERT3U(db->db_state, ==, DB_CACHED);
2344 ASSERT(db->db_buf != NULL);
2348 /* Indirect block size must match what the dnode thinks it is. */
2349 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2350 dbuf_check_blkptr(dn, db);
2353 /* Provide the pending dirty record to child dbufs */
2354 db->db_data_pending = dr;
2356 mutex_exit(&db->db_mtx);
2357 dbuf_write(dr, db->db_buf, tx);
2360 mutex_enter(&dr->dt.di.dr_mtx);
2361 dbuf_sync_list(&dr->dt.di.dr_children, tx);
2362 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2363 mutex_exit(&dr->dt.di.dr_mtx);
2367 /* dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is
2368 * critical the we not allow the compiler to inline this function in to
2369 * dbuf_sync_list() thereby drastically bloating the stack usage.
2371 noinline static void
2372 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2374 arc_buf_t **datap = &dr->dt.dl.dr_data;
2375 dmu_buf_impl_t *db = dr->dr_dbuf;
2378 uint64_t txg = tx->tx_txg;
2380 ASSERT(dmu_tx_is_syncing(tx));
2382 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2384 mutex_enter(&db->db_mtx);
2386 * To be synced, we must be dirtied. But we
2387 * might have been freed after the dirty.
2389 if (db->db_state == DB_UNCACHED) {
2390 /* This buffer has been freed since it was dirtied */
2391 ASSERT(db->db.db_data == NULL);
2392 } else if (db->db_state == DB_FILL) {
2393 /* This buffer was freed and is now being re-filled */
2394 ASSERT(db->db.db_data != dr->dt.dl.dr_data);
2396 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
2403 if (db->db_blkid == DMU_SPILL_BLKID) {
2404 mutex_enter(&dn->dn_mtx);
2405 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
2406 mutex_exit(&dn->dn_mtx);
2410 * If this is a bonus buffer, simply copy the bonus data into the
2411 * dnode. It will be written out when the dnode is synced (and it
2412 * will be synced, since it must have been dirty for dbuf_sync to
2415 if (db->db_blkid == DMU_BONUS_BLKID) {
2416 dbuf_dirty_record_t **drp;
2418 ASSERT(*datap != NULL);
2419 ASSERT0(db->db_level);
2420 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN);
2421 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen);
2424 if (*datap != db->db.db_data) {
2425 zio_buf_free(*datap, DN_MAX_BONUSLEN);
2426 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
2428 db->db_data_pending = NULL;
2429 drp = &db->db_last_dirty;
2431 drp = &(*drp)->dr_next;
2432 ASSERT(dr->dr_next == NULL);
2433 ASSERT(dr->dr_dbuf == db);
2435 if (dr->dr_dbuf->db_level != 0) {
2436 mutex_destroy(&dr->dt.di.dr_mtx);
2437 list_destroy(&dr->dt.di.dr_children);
2439 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2440 ASSERT(db->db_dirtycnt > 0);
2441 db->db_dirtycnt -= 1;
2442 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
2449 * This function may have dropped the db_mtx lock allowing a dmu_sync
2450 * operation to sneak in. As a result, we need to ensure that we
2451 * don't check the dr_override_state until we have returned from
2452 * dbuf_check_blkptr.
2454 dbuf_check_blkptr(dn, db);
2457 * If this buffer is in the middle of an immediate write,
2458 * wait for the synchronous IO to complete.
2460 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
2461 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
2462 cv_wait(&db->db_changed, &db->db_mtx);
2463 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
2466 if (db->db_state != DB_NOFILL &&
2467 dn->dn_object != DMU_META_DNODE_OBJECT &&
2468 refcount_count(&db->db_holds) > 1 &&
2469 dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
2470 *datap == db->db_buf) {
2472 * If this buffer is currently "in use" (i.e., there
2473 * are active holds and db_data still references it),
2474 * then make a copy before we start the write so that
2475 * any modifications from the open txg will not leak
2478 * NOTE: this copy does not need to be made for
2479 * objects only modified in the syncing context (e.g.
2480 * DNONE_DNODE blocks).
2482 int blksz = arc_buf_size(*datap);
2483 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2484 *datap = arc_buf_alloc(os->os_spa, blksz, db, type);
2485 bcopy(db->db.db_data, (*datap)->b_data, blksz);
2487 db->db_data_pending = dr;
2489 mutex_exit(&db->db_mtx);
2491 dbuf_write(dr, *datap, tx);
2493 ASSERT(!list_link_active(&dr->dr_dirty_node));
2494 if (dn->dn_object == DMU_META_DNODE_OBJECT) {
2495 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
2499 * Although zio_nowait() does not "wait for an IO", it does
2500 * initiate the IO. If this is an empty write it seems plausible
2501 * that the IO could actually be completed before the nowait
2502 * returns. We need to DB_DNODE_EXIT() first in case
2503 * zio_nowait() invalidates the dbuf.
2506 zio_nowait(dr->dr_zio);
2511 dbuf_sync_list(list_t *list, dmu_tx_t *tx)
2513 dbuf_dirty_record_t *dr;
2515 while ((dr = list_head(list))) {
2516 if (dr->dr_zio != NULL) {
2518 * If we find an already initialized zio then we
2519 * are processing the meta-dnode, and we have finished.
2520 * The dbufs for all dnodes are put back on the list
2521 * during processing, so that we can zio_wait()
2522 * these IOs after initiating all child IOs.
2524 ASSERT3U(dr->dr_dbuf->db.db_object, ==,
2525 DMU_META_DNODE_OBJECT);
2528 list_remove(list, dr);
2529 if (dr->dr_dbuf->db_level > 0)
2530 dbuf_sync_indirect(dr, tx);
2532 dbuf_sync_leaf(dr, tx);
2538 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
2540 dmu_buf_impl_t *db = vdb;
2542 blkptr_t *bp = zio->io_bp;
2543 blkptr_t *bp_orig = &zio->io_bp_orig;
2544 spa_t *spa = zio->io_spa;
2549 ASSERT(db->db_blkptr == bp);
2553 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
2554 dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
2555 zio->io_prev_space_delta = delta;
2557 if (BP_IS_HOLE(bp)) {
2558 ASSERT(bp->blk_fill == 0);
2563 ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
2564 BP_GET_TYPE(bp) == dn->dn_type) ||
2565 (db->db_blkid == DMU_SPILL_BLKID &&
2566 BP_GET_TYPE(bp) == dn->dn_bonustype));
2567 ASSERT(BP_GET_LEVEL(bp) == db->db_level);
2569 mutex_enter(&db->db_mtx);
2572 if (db->db_blkid == DMU_SPILL_BLKID) {
2573 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2574 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2575 db->db_blkptr == &dn->dn_phys->dn_spill);
2579 if (db->db_level == 0) {
2580 mutex_enter(&dn->dn_mtx);
2581 if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
2582 db->db_blkid != DMU_SPILL_BLKID)
2583 dn->dn_phys->dn_maxblkid = db->db_blkid;
2584 mutex_exit(&dn->dn_mtx);
2586 if (dn->dn_type == DMU_OT_DNODE) {
2587 dnode_phys_t *dnp = db->db.db_data;
2588 for (i = db->db.db_size >> DNODE_SHIFT; i > 0;
2590 if (dnp->dn_type != DMU_OT_NONE)
2597 blkptr_t *ibp = db->db.db_data;
2598 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2599 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
2600 if (BP_IS_HOLE(ibp))
2602 fill += ibp->blk_fill;
2607 bp->blk_fill = fill;
2609 mutex_exit(&db->db_mtx);
2614 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
2616 dmu_buf_impl_t *db = vdb;
2617 blkptr_t *bp = zio->io_bp;
2618 blkptr_t *bp_orig = &zio->io_bp_orig;
2619 uint64_t txg = zio->io_txg;
2620 dbuf_dirty_record_t **drp, *dr;
2622 ASSERT0(zio->io_error);
2623 ASSERT(db->db_blkptr == bp);
2625 if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
2626 ASSERT(BP_EQUAL(bp, bp_orig));
2632 DB_GET_OBJSET(&os, db);
2633 ds = os->os_dsl_dataset;
2636 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
2637 dsl_dataset_block_born(ds, bp, tx);
2640 mutex_enter(&db->db_mtx);
2644 drp = &db->db_last_dirty;
2645 while ((dr = *drp) != db->db_data_pending)
2647 ASSERT(!list_link_active(&dr->dr_dirty_node));
2648 ASSERT(dr->dr_txg == txg);
2649 ASSERT(dr->dr_dbuf == db);
2650 ASSERT(dr->dr_next == NULL);
2654 if (db->db_blkid == DMU_SPILL_BLKID) {
2659 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2660 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2661 db->db_blkptr == &dn->dn_phys->dn_spill);
2666 if (db->db_level == 0) {
2667 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2668 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
2669 if (db->db_state != DB_NOFILL) {
2670 if (dr->dt.dl.dr_data != db->db_buf)
2671 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data,
2673 else if (!arc_released(db->db_buf))
2674 arc_set_callback(db->db_buf, dbuf_do_evict, db);
2681 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2682 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2683 if (!BP_IS_HOLE(db->db_blkptr)) {
2684 ASSERTV(int epbs = dn->dn_phys->dn_indblkshift -
2686 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
2688 ASSERT3U(dn->dn_phys->dn_maxblkid
2689 >> (db->db_level * epbs), >=, db->db_blkid);
2690 arc_set_callback(db->db_buf, dbuf_do_evict, db);
2693 mutex_destroy(&dr->dt.di.dr_mtx);
2694 list_destroy(&dr->dt.di.dr_children);
2696 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2698 cv_broadcast(&db->db_changed);
2699 ASSERT(db->db_dirtycnt > 0);
2700 db->db_dirtycnt -= 1;
2701 db->db_data_pending = NULL;
2702 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
2706 dbuf_write_nofill_ready(zio_t *zio)
2708 dbuf_write_ready(zio, NULL, zio->io_private);
2712 dbuf_write_nofill_done(zio_t *zio)
2714 dbuf_write_done(zio, NULL, zio->io_private);
2718 dbuf_write_override_ready(zio_t *zio)
2720 dbuf_dirty_record_t *dr = zio->io_private;
2721 dmu_buf_impl_t *db = dr->dr_dbuf;
2723 dbuf_write_ready(zio, NULL, db);
2727 dbuf_write_override_done(zio_t *zio)
2729 dbuf_dirty_record_t *dr = zio->io_private;
2730 dmu_buf_impl_t *db = dr->dr_dbuf;
2731 blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
2733 mutex_enter(&db->db_mtx);
2734 if (!BP_EQUAL(zio->io_bp, obp)) {
2735 if (!BP_IS_HOLE(obp))
2736 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
2737 arc_release(dr->dt.dl.dr_data, db);
2739 mutex_exit(&db->db_mtx);
2741 dbuf_write_done(zio, NULL, db);
2744 /* Issue I/O to commit a dirty buffer to disk. */
2746 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
2748 dmu_buf_impl_t *db = dr->dr_dbuf;
2751 dmu_buf_impl_t *parent = db->db_parent;
2752 uint64_t txg = tx->tx_txg;
2762 if (db->db_state != DB_NOFILL) {
2763 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
2765 * Private object buffers are released here rather
2766 * than in dbuf_dirty() since they are only modified
2767 * in the syncing context and we don't want the
2768 * overhead of making multiple copies of the data.
2770 if (BP_IS_HOLE(db->db_blkptr)) {
2773 dbuf_release_bp(db);
2778 if (parent != dn->dn_dbuf) {
2779 /* Our parent is an indirect block. */
2780 /* We have a dirty parent that has been scheduled for write. */
2781 ASSERT(parent && parent->db_data_pending);
2782 /* Our parent's buffer is one level closer to the dnode. */
2783 ASSERT(db->db_level == parent->db_level-1);
2785 * We're about to modify our parent's db_data by modifying
2786 * our block pointer, so the parent must be released.
2788 ASSERT(arc_released(parent->db_buf));
2789 zio = parent->db_data_pending->dr_zio;
2791 /* Our parent is the dnode itself. */
2792 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
2793 db->db_blkid != DMU_SPILL_BLKID) ||
2794 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
2795 if (db->db_blkid != DMU_SPILL_BLKID)
2796 ASSERT3P(db->db_blkptr, ==,
2797 &dn->dn_phys->dn_blkptr[db->db_blkid]);
2801 ASSERT(db->db_level == 0 || data == db->db_buf);
2802 ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
2805 SET_BOOKMARK(&zb, os->os_dsl_dataset ?
2806 os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
2807 db->db.db_object, db->db_level, db->db_blkid);
2809 if (db->db_blkid == DMU_SPILL_BLKID)
2811 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
2813 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
2816 if (db->db_level == 0 && dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
2817 ASSERT(db->db_state != DB_NOFILL);
2818 dr->dr_zio = zio_write(zio, os->os_spa, txg,
2819 db->db_blkptr, data->b_data, arc_buf_size(data), &zp,
2820 dbuf_write_override_ready, dbuf_write_override_done, dr,
2821 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
2822 mutex_enter(&db->db_mtx);
2823 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
2824 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
2825 dr->dt.dl.dr_copies);
2826 mutex_exit(&db->db_mtx);
2827 } else if (db->db_state == DB_NOFILL) {
2828 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF);
2829 dr->dr_zio = zio_write(zio, os->os_spa, txg,
2830 db->db_blkptr, NULL, db->db.db_size, &zp,
2831 dbuf_write_nofill_ready, dbuf_write_nofill_done, db,
2832 ZIO_PRIORITY_ASYNC_WRITE,
2833 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
2835 ASSERT(arc_released(data));
2836 dr->dr_zio = arc_write(zio, os->os_spa, txg,
2837 db->db_blkptr, data, DBUF_IS_L2CACHEABLE(db),
2838 DBUF_IS_L2COMPRESSIBLE(db), &zp, dbuf_write_ready,
2839 dbuf_write_done, db, ZIO_PRIORITY_ASYNC_WRITE,
2840 ZIO_FLAG_MUSTSUCCEED, &zb);
2844 #if defined(_KERNEL) && defined(HAVE_SPL)
2845 EXPORT_SYMBOL(dbuf_find);
2846 EXPORT_SYMBOL(dbuf_is_metadata);
2847 EXPORT_SYMBOL(dbuf_evict);
2848 EXPORT_SYMBOL(dbuf_loan_arcbuf);
2849 EXPORT_SYMBOL(dbuf_whichblock);
2850 EXPORT_SYMBOL(dbuf_read);
2851 EXPORT_SYMBOL(dbuf_unoverride);
2852 EXPORT_SYMBOL(dbuf_free_range);
2853 EXPORT_SYMBOL(dbuf_new_size);
2854 EXPORT_SYMBOL(dbuf_release_bp);
2855 EXPORT_SYMBOL(dbuf_dirty);
2856 EXPORT_SYMBOL(dmu_buf_will_dirty);
2857 EXPORT_SYMBOL(dmu_buf_will_not_fill);
2858 EXPORT_SYMBOL(dmu_buf_will_fill);
2859 EXPORT_SYMBOL(dmu_buf_fill_done);
2860 EXPORT_SYMBOL(dmu_buf_rele);
2861 EXPORT_SYMBOL(dbuf_assign_arcbuf);
2862 EXPORT_SYMBOL(dbuf_clear);
2863 EXPORT_SYMBOL(dbuf_prefetch);
2864 EXPORT_SYMBOL(dbuf_hold_impl);
2865 EXPORT_SYMBOL(dbuf_hold);
2866 EXPORT_SYMBOL(dbuf_hold_level);
2867 EXPORT_SYMBOL(dbuf_create_bonus);
2868 EXPORT_SYMBOL(dbuf_spill_set_blksz);
2869 EXPORT_SYMBOL(dbuf_rm_spill);
2870 EXPORT_SYMBOL(dbuf_add_ref);
2871 EXPORT_SYMBOL(dbuf_rele);
2872 EXPORT_SYMBOL(dbuf_rele_and_unlock);
2873 EXPORT_SYMBOL(dbuf_refcount);
2874 EXPORT_SYMBOL(dbuf_sync_list);
2875 EXPORT_SYMBOL(dmu_buf_set_user);
2876 EXPORT_SYMBOL(dmu_buf_set_user_ie);
2877 EXPORT_SYMBOL(dmu_buf_update_user);
2878 EXPORT_SYMBOL(dmu_buf_get_user);
2879 EXPORT_SYMBOL(dmu_buf_freeable);