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 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 int 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);
325 dbuf_hash_table_t *h = &dbuf_hash_table;
328 for (i = 0; i < DBUF_MUTEXES; i++)
329 mutex_destroy(&h->hash_mutexes[i]);
330 #if defined(_KERNEL) && defined(HAVE_SPL)
331 /* Large allocations which do not require contiguous pages
332 * should be using vmem_free() in the linux kernel */
333 vmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
335 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
337 kmem_cache_destroy(dbuf_cache);
346 dbuf_verify(dmu_buf_impl_t *db)
349 dbuf_dirty_record_t *dr;
351 ASSERT(MUTEX_HELD(&db->db_mtx));
353 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
356 ASSERT(db->db_objset != NULL);
360 ASSERT(db->db_parent == NULL);
361 ASSERT(db->db_blkptr == NULL);
363 ASSERT3U(db->db.db_object, ==, dn->dn_object);
364 ASSERT3P(db->db_objset, ==, dn->dn_objset);
365 ASSERT3U(db->db_level, <, dn->dn_nlevels);
366 ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
367 db->db_blkid == DMU_SPILL_BLKID ||
368 !list_is_empty(&dn->dn_dbufs));
370 if (db->db_blkid == DMU_BONUS_BLKID) {
372 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
373 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
374 } else if (db->db_blkid == DMU_SPILL_BLKID) {
376 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
377 ASSERT0(db->db.db_offset);
379 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
382 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
383 ASSERT(dr->dr_dbuf == db);
385 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
386 ASSERT(dr->dr_dbuf == db);
389 * We can't assert that db_size matches dn_datablksz because it
390 * can be momentarily different when another thread is doing
393 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
394 dr = db->db_data_pending;
396 * It should only be modified in syncing context, so
397 * make sure we only have one copy of the data.
399 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
402 /* verify db->db_blkptr */
404 if (db->db_parent == dn->dn_dbuf) {
405 /* db is pointed to by the dnode */
406 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
407 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
408 ASSERT(db->db_parent == NULL);
410 ASSERT(db->db_parent != NULL);
411 if (db->db_blkid != DMU_SPILL_BLKID)
412 ASSERT3P(db->db_blkptr, ==,
413 &dn->dn_phys->dn_blkptr[db->db_blkid]);
415 /* db is pointed to by an indirect block */
416 ASSERTV(int epb = db->db_parent->db.db_size >>
418 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
419 ASSERT3U(db->db_parent->db.db_object, ==,
422 * dnode_grow_indblksz() can make this fail if we don't
423 * have the struct_rwlock. XXX indblksz no longer
424 * grows. safe to do this now?
426 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
427 ASSERT3P(db->db_blkptr, ==,
428 ((blkptr_t *)db->db_parent->db.db_data +
429 db->db_blkid % epb));
433 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
434 (db->db_buf == NULL || db->db_buf->b_data) &&
435 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
436 db->db_state != DB_FILL && !dn->dn_free_txg) {
438 * If the blkptr isn't set but they have nonzero data,
439 * it had better be dirty, otherwise we'll lose that
440 * data when we evict this buffer.
442 if (db->db_dirtycnt == 0) {
443 ASSERTV(uint64_t *buf = db->db.db_data);
446 for (i = 0; i < db->db.db_size >> 3; i++) {
456 dbuf_update_data(dmu_buf_impl_t *db)
458 ASSERT(MUTEX_HELD(&db->db_mtx));
459 if (db->db_level == 0 && db->db_user_data_ptr_ptr) {
460 ASSERT(!refcount_is_zero(&db->db_holds));
461 *db->db_user_data_ptr_ptr = db->db.db_data;
466 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
468 ASSERT(MUTEX_HELD(&db->db_mtx));
469 ASSERT(db->db_buf == NULL || !arc_has_callback(db->db_buf));
472 ASSERT(buf->b_data != NULL);
473 db->db.db_data = buf->b_data;
474 if (!arc_released(buf))
475 arc_set_callback(buf, dbuf_do_evict, db);
476 dbuf_update_data(db);
479 db->db.db_data = NULL;
480 if (db->db_state != DB_NOFILL)
481 db->db_state = DB_UNCACHED;
486 * Loan out an arc_buf for read. Return the loaned arc_buf.
489 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
493 mutex_enter(&db->db_mtx);
494 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
495 int blksz = db->db.db_size;
498 mutex_exit(&db->db_mtx);
499 DB_GET_SPA(&spa, db);
500 abuf = arc_loan_buf(spa, blksz);
501 bcopy(db->db.db_data, abuf->b_data, blksz);
504 arc_loan_inuse_buf(abuf, db);
505 dbuf_set_data(db, NULL);
506 mutex_exit(&db->db_mtx);
512 dbuf_whichblock(dnode_t *dn, uint64_t offset)
514 if (dn->dn_datablkshift) {
515 return (offset >> dn->dn_datablkshift);
517 ASSERT3U(offset, <, dn->dn_datablksz);
523 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
525 dmu_buf_impl_t *db = vdb;
527 mutex_enter(&db->db_mtx);
528 ASSERT3U(db->db_state, ==, DB_READ);
530 * All reads are synchronous, so we must have a hold on the dbuf
532 ASSERT(refcount_count(&db->db_holds) > 0);
533 ASSERT(db->db_buf == NULL);
534 ASSERT(db->db.db_data == NULL);
535 if (db->db_level == 0 && db->db_freed_in_flight) {
536 /* we were freed in flight; disregard any error */
537 arc_release(buf, db);
538 bzero(buf->b_data, db->db.db_size);
540 db->db_freed_in_flight = FALSE;
541 dbuf_set_data(db, buf);
542 db->db_state = DB_CACHED;
543 } else if (zio == NULL || zio->io_error == 0) {
544 dbuf_set_data(db, buf);
545 db->db_state = DB_CACHED;
547 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
548 ASSERT3P(db->db_buf, ==, NULL);
549 VERIFY(arc_buf_remove_ref(buf, db) == 1);
550 db->db_state = DB_UNCACHED;
552 cv_broadcast(&db->db_changed);
553 dbuf_rele_and_unlock(db, NULL);
557 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t *flags)
562 uint32_t aflags = ARC_NOWAIT;
566 ASSERT(!refcount_is_zero(&db->db_holds));
567 /* We need the struct_rwlock to prevent db_blkptr from changing. */
568 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
569 ASSERT(MUTEX_HELD(&db->db_mtx));
570 ASSERT(db->db_state == DB_UNCACHED);
571 ASSERT(db->db_buf == NULL);
573 if (db->db_blkid == DMU_BONUS_BLKID) {
574 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
576 ASSERT3U(bonuslen, <=, db->db.db_size);
577 db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN);
578 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
579 if (bonuslen < DN_MAX_BONUSLEN)
580 bzero(db->db.db_data, DN_MAX_BONUSLEN);
582 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
584 dbuf_update_data(db);
585 db->db_state = DB_CACHED;
586 mutex_exit(&db->db_mtx);
591 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
592 * processes the delete record and clears the bp while we are waiting
593 * for the dn_mtx (resulting in a "no" from block_freed).
595 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
596 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
597 BP_IS_HOLE(db->db_blkptr)))) {
598 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
600 dbuf_set_data(db, arc_buf_alloc(dn->dn_objset->os_spa,
601 db->db.db_size, db, type));
603 bzero(db->db.db_data, db->db.db_size);
604 db->db_state = DB_CACHED;
605 *flags |= DB_RF_CACHED;
606 mutex_exit(&db->db_mtx);
610 spa = dn->dn_objset->os_spa;
613 db->db_state = DB_READ;
614 mutex_exit(&db->db_mtx);
616 if (DBUF_IS_L2CACHEABLE(db))
617 aflags |= ARC_L2CACHE;
618 if (DBUF_IS_L2COMPRESSIBLE(db))
619 aflags |= ARC_L2COMPRESS;
621 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
622 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
623 db->db.db_object, db->db_level, db->db_blkid);
625 dbuf_add_ref(db, NULL);
627 (void) arc_read(zio, spa, db->db_blkptr,
628 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
629 (*flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
631 if (aflags & ARC_CACHED)
632 *flags |= DB_RF_CACHED;
636 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
639 int havepzio = (zio != NULL);
644 * We don't have to hold the mutex to check db_state because it
645 * can't be freed while we have a hold on the buffer.
647 ASSERT(!refcount_is_zero(&db->db_holds));
649 if (db->db_state == DB_NOFILL)
654 if ((flags & DB_RF_HAVESTRUCT) == 0)
655 rw_enter(&dn->dn_struct_rwlock, RW_READER);
657 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
658 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
659 DBUF_IS_CACHEABLE(db);
661 mutex_enter(&db->db_mtx);
662 if (db->db_state == DB_CACHED) {
663 mutex_exit(&db->db_mtx);
665 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
666 db->db.db_size, TRUE);
667 if ((flags & DB_RF_HAVESTRUCT) == 0)
668 rw_exit(&dn->dn_struct_rwlock);
670 } else if (db->db_state == DB_UNCACHED) {
671 spa_t *spa = dn->dn_objset->os_spa;
674 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
675 dbuf_read_impl(db, zio, &flags);
677 /* dbuf_read_impl has dropped db_mtx for us */
680 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
681 db->db.db_size, flags & DB_RF_CACHED);
683 if ((flags & DB_RF_HAVESTRUCT) == 0)
684 rw_exit(&dn->dn_struct_rwlock);
690 mutex_exit(&db->db_mtx);
692 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
693 db->db.db_size, TRUE);
694 if ((flags & DB_RF_HAVESTRUCT) == 0)
695 rw_exit(&dn->dn_struct_rwlock);
698 mutex_enter(&db->db_mtx);
699 if ((flags & DB_RF_NEVERWAIT) == 0) {
700 while (db->db_state == DB_READ ||
701 db->db_state == DB_FILL) {
702 ASSERT(db->db_state == DB_READ ||
703 (flags & DB_RF_HAVESTRUCT) == 0);
704 cv_wait(&db->db_changed, &db->db_mtx);
706 if (db->db_state == DB_UNCACHED)
709 mutex_exit(&db->db_mtx);
712 ASSERT(err || havepzio || db->db_state == DB_CACHED);
717 dbuf_noread(dmu_buf_impl_t *db)
719 ASSERT(!refcount_is_zero(&db->db_holds));
720 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
721 mutex_enter(&db->db_mtx);
722 while (db->db_state == DB_READ || db->db_state == DB_FILL)
723 cv_wait(&db->db_changed, &db->db_mtx);
724 if (db->db_state == DB_UNCACHED) {
725 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
728 ASSERT(db->db_buf == NULL);
729 ASSERT(db->db.db_data == NULL);
730 DB_GET_SPA(&spa, db);
731 dbuf_set_data(db, arc_buf_alloc(spa, db->db.db_size, db, type));
732 db->db_state = DB_FILL;
733 } else if (db->db_state == DB_NOFILL) {
734 dbuf_set_data(db, NULL);
736 ASSERT3U(db->db_state, ==, DB_CACHED);
738 mutex_exit(&db->db_mtx);
742 * This is our just-in-time copy function. It makes a copy of
743 * buffers, that have been modified in a previous transaction
744 * group, before we modify them in the current active group.
746 * This function is used in two places: when we are dirtying a
747 * buffer for the first time in a txg, and when we are freeing
748 * a range in a dnode that includes this buffer.
750 * Note that when we are called from dbuf_free_range() we do
751 * not put a hold on the buffer, we just traverse the active
752 * dbuf list for the dnode.
755 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
757 dbuf_dirty_record_t *dr = db->db_last_dirty;
759 ASSERT(MUTEX_HELD(&db->db_mtx));
760 ASSERT(db->db.db_data != NULL);
761 ASSERT(db->db_level == 0);
762 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
765 (dr->dt.dl.dr_data !=
766 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
770 * If the last dirty record for this dbuf has not yet synced
771 * and its referencing the dbuf data, either:
772 * reset the reference to point to a new copy,
773 * or (if there a no active holders)
774 * just null out the current db_data pointer.
776 ASSERT(dr->dr_txg >= txg - 2);
777 if (db->db_blkid == DMU_BONUS_BLKID) {
778 /* Note that the data bufs here are zio_bufs */
779 dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
780 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
781 bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
782 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
783 int size = db->db.db_size;
784 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
787 DB_GET_SPA(&spa, db);
788 dr->dt.dl.dr_data = arc_buf_alloc(spa, size, db, type);
789 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
791 dbuf_set_data(db, NULL);
796 dbuf_unoverride(dbuf_dirty_record_t *dr)
798 dmu_buf_impl_t *db = dr->dr_dbuf;
799 blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
800 uint64_t txg = dr->dr_txg;
802 ASSERT(MUTEX_HELD(&db->db_mtx));
803 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
804 ASSERT(db->db_level == 0);
806 if (db->db_blkid == DMU_BONUS_BLKID ||
807 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
810 ASSERT(db->db_data_pending != dr);
812 /* free this block */
813 if (!BP_IS_HOLE(bp)) {
816 DB_GET_SPA(&spa, db);
817 zio_free(spa, txg, bp);
819 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
821 * Release the already-written buffer, so we leave it in
822 * a consistent dirty state. Note that all callers are
823 * modifying the buffer, so they will immediately do
824 * another (redundant) arc_release(). Therefore, leave
825 * the buf thawed to save the effort of freezing &
826 * immediately re-thawing it.
828 arc_release(dr->dt.dl.dr_data, db);
832 * Evict (if its unreferenced) or clear (if its referenced) any level-0
833 * data blocks in the free range, so that any future readers will find
834 * empty blocks. Also, if we happen accross any level-1 dbufs in the
835 * range that have not already been marked dirty, mark them dirty so
836 * they stay in memory.
839 dbuf_free_range(dnode_t *dn, uint64_t start, uint64_t end, dmu_tx_t *tx)
841 dmu_buf_impl_t *db, *db_next;
842 uint64_t txg = tx->tx_txg;
843 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
844 uint64_t first_l1 = start >> epbs;
845 uint64_t last_l1 = end >> epbs;
847 if (end > dn->dn_maxblkid && (end != DMU_SPILL_BLKID)) {
848 end = dn->dn_maxblkid;
849 last_l1 = end >> epbs;
851 dprintf_dnode(dn, "start=%llu end=%llu\n", start, end);
852 mutex_enter(&dn->dn_dbufs_mtx);
853 for (db = list_head(&dn->dn_dbufs); db; db = db_next) {
854 db_next = list_next(&dn->dn_dbufs, db);
855 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
857 if (db->db_level == 1 &&
858 db->db_blkid >= first_l1 && db->db_blkid <= last_l1) {
859 mutex_enter(&db->db_mtx);
860 if (db->db_last_dirty &&
861 db->db_last_dirty->dr_txg < txg) {
862 dbuf_add_ref(db, FTAG);
863 mutex_exit(&db->db_mtx);
864 dbuf_will_dirty(db, tx);
867 mutex_exit(&db->db_mtx);
871 if (db->db_level != 0)
873 dprintf_dbuf(db, "found buf %s\n", "");
874 if (db->db_blkid < start || db->db_blkid > end)
877 /* found a level 0 buffer in the range */
878 if (dbuf_undirty(db, tx))
881 mutex_enter(&db->db_mtx);
882 if (db->db_state == DB_UNCACHED ||
883 db->db_state == DB_NOFILL ||
884 db->db_state == DB_EVICTING) {
885 ASSERT(db->db.db_data == NULL);
886 mutex_exit(&db->db_mtx);
889 if (db->db_state == DB_READ || db->db_state == DB_FILL) {
890 /* will be handled in dbuf_read_done or dbuf_rele */
891 db->db_freed_in_flight = TRUE;
892 mutex_exit(&db->db_mtx);
895 if (refcount_count(&db->db_holds) == 0) {
900 /* The dbuf is referenced */
902 if (db->db_last_dirty != NULL) {
903 dbuf_dirty_record_t *dr = db->db_last_dirty;
905 if (dr->dr_txg == txg) {
907 * This buffer is "in-use", re-adjust the file
908 * size to reflect that this buffer may
909 * contain new data when we sync.
911 if (db->db_blkid != DMU_SPILL_BLKID &&
912 db->db_blkid > dn->dn_maxblkid)
913 dn->dn_maxblkid = db->db_blkid;
917 * This dbuf is not dirty in the open context.
918 * Either uncache it (if its not referenced in
919 * the open context) or reset its contents to
922 dbuf_fix_old_data(db, txg);
925 /* clear the contents if its cached */
926 if (db->db_state == DB_CACHED) {
927 ASSERT(db->db.db_data != NULL);
928 arc_release(db->db_buf, db);
929 bzero(db->db.db_data, db->db.db_size);
930 arc_buf_freeze(db->db_buf);
933 mutex_exit(&db->db_mtx);
935 mutex_exit(&dn->dn_dbufs_mtx);
939 dbuf_block_freeable(dmu_buf_impl_t *db)
941 dsl_dataset_t *ds = db->db_objset->os_dsl_dataset;
942 uint64_t birth_txg = 0;
945 * We don't need any locking to protect db_blkptr:
946 * If it's syncing, then db_last_dirty will be set
947 * so we'll ignore db_blkptr.
949 ASSERT(MUTEX_HELD(&db->db_mtx));
950 if (db->db_last_dirty)
951 birth_txg = db->db_last_dirty->dr_txg;
952 else if (db->db_blkptr)
953 birth_txg = db->db_blkptr->blk_birth;
956 * If we don't exist or are in a snapshot, we can't be freed.
957 * Don't pass the bp to dsl_dataset_block_freeable() since we
958 * are holding the db_mtx lock and might deadlock if we are
959 * prefetching a dedup-ed block.
962 return (ds == NULL ||
963 dsl_dataset_block_freeable(ds, NULL, birth_txg));
969 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
971 arc_buf_t *buf, *obuf;
972 int osize = db->db.db_size;
973 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
976 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
981 /* XXX does *this* func really need the lock? */
982 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
985 * This call to dbuf_will_dirty() with the dn_struct_rwlock held
986 * is OK, because there can be no other references to the db
987 * when we are changing its size, so no concurrent DB_FILL can
991 * XXX we should be doing a dbuf_read, checking the return
992 * value and returning that up to our callers
994 dbuf_will_dirty(db, tx);
996 /* create the data buffer for the new block */
997 buf = arc_buf_alloc(dn->dn_objset->os_spa, size, db, type);
999 /* copy old block data to the new block */
1001 bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
1002 /* zero the remainder */
1004 bzero((uint8_t *)buf->b_data + osize, size - osize);
1006 mutex_enter(&db->db_mtx);
1007 dbuf_set_data(db, buf);
1008 VERIFY(arc_buf_remove_ref(obuf, db) == 1);
1009 db->db.db_size = size;
1011 if (db->db_level == 0) {
1012 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1013 db->db_last_dirty->dt.dl.dr_data = buf;
1015 mutex_exit(&db->db_mtx);
1017 dnode_willuse_space(dn, size-osize, tx);
1022 dbuf_release_bp(dmu_buf_impl_t *db)
1026 DB_GET_OBJSET(&os, db);
1027 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1028 ASSERT(arc_released(os->os_phys_buf) ||
1029 list_link_active(&os->os_dsl_dataset->ds_synced_link));
1030 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1032 (void) arc_release(db->db_buf, db);
1035 dbuf_dirty_record_t *
1036 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1040 dbuf_dirty_record_t **drp, *dr;
1041 int drop_struct_lock = FALSE;
1042 boolean_t do_free_accounting = B_FALSE;
1043 int txgoff = tx->tx_txg & TXG_MASK;
1045 ASSERT(tx->tx_txg != 0);
1046 ASSERT(!refcount_is_zero(&db->db_holds));
1047 DMU_TX_DIRTY_BUF(tx, db);
1052 * Shouldn't dirty a regular buffer in syncing context. Private
1053 * objects may be dirtied in syncing context, but only if they
1054 * were already pre-dirtied in open context.
1056 ASSERT(!dmu_tx_is_syncing(tx) ||
1057 BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1058 DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1059 dn->dn_objset->os_dsl_dataset == NULL);
1061 * We make this assert for private objects as well, but after we
1062 * check if we're already dirty. They are allowed to re-dirty
1063 * in syncing context.
1065 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1066 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1067 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1069 mutex_enter(&db->db_mtx);
1071 * XXX make this true for indirects too? The problem is that
1072 * transactions created with dmu_tx_create_assigned() from
1073 * syncing context don't bother holding ahead.
1075 ASSERT(db->db_level != 0 ||
1076 db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1077 db->db_state == DB_NOFILL);
1079 mutex_enter(&dn->dn_mtx);
1081 * Don't set dirtyctx to SYNC if we're just modifying this as we
1082 * initialize the objset.
1084 if (dn->dn_dirtyctx == DN_UNDIRTIED &&
1085 !BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1087 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1088 ASSERT(dn->dn_dirtyctx_firstset == NULL);
1089 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_PUSHPAGE);
1091 mutex_exit(&dn->dn_mtx);
1093 if (db->db_blkid == DMU_SPILL_BLKID)
1094 dn->dn_have_spill = B_TRUE;
1097 * If this buffer is already dirty, we're done.
1099 drp = &db->db_last_dirty;
1100 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1101 db->db.db_object == DMU_META_DNODE_OBJECT);
1102 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1104 if (dr && dr->dr_txg == tx->tx_txg) {
1107 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1109 * If this buffer has already been written out,
1110 * we now need to reset its state.
1112 dbuf_unoverride(dr);
1113 if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1114 db->db_state != DB_NOFILL)
1115 arc_buf_thaw(db->db_buf);
1117 mutex_exit(&db->db_mtx);
1122 * Only valid if not already dirty.
1124 ASSERT(dn->dn_object == 0 ||
1125 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1126 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1128 ASSERT3U(dn->dn_nlevels, >, db->db_level);
1129 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1130 dn->dn_phys->dn_nlevels > db->db_level ||
1131 dn->dn_next_nlevels[txgoff] > db->db_level ||
1132 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1133 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1136 * We should only be dirtying in syncing context if it's the
1137 * mos or we're initializing the os or it's a special object.
1138 * However, we are allowed to dirty in syncing context provided
1139 * we already dirtied it in open context. Hence we must make
1140 * this assertion only if we're not already dirty.
1143 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1144 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1145 ASSERT(db->db.db_size != 0);
1147 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1149 if (db->db_blkid != DMU_BONUS_BLKID) {
1151 * Update the accounting.
1152 * Note: we delay "free accounting" until after we drop
1153 * the db_mtx. This keeps us from grabbing other locks
1154 * (and possibly deadlocking) in bp_get_dsize() while
1155 * also holding the db_mtx.
1157 dnode_willuse_space(dn, db->db.db_size, tx);
1158 do_free_accounting = dbuf_block_freeable(db);
1162 * If this buffer is dirty in an old transaction group we need
1163 * to make a copy of it so that the changes we make in this
1164 * transaction group won't leak out when we sync the older txg.
1166 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_PUSHPAGE);
1167 list_link_init(&dr->dr_dirty_node);
1168 if (db->db_level == 0) {
1169 void *data_old = db->db_buf;
1171 if (db->db_state != DB_NOFILL) {
1172 if (db->db_blkid == DMU_BONUS_BLKID) {
1173 dbuf_fix_old_data(db, tx->tx_txg);
1174 data_old = db->db.db_data;
1175 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1177 * Release the data buffer from the cache so
1178 * that we can modify it without impacting
1179 * possible other users of this cached data
1180 * block. Note that indirect blocks and
1181 * private objects are not released until the
1182 * syncing state (since they are only modified
1185 arc_release(db->db_buf, db);
1186 dbuf_fix_old_data(db, tx->tx_txg);
1187 data_old = db->db_buf;
1189 ASSERT(data_old != NULL);
1191 dr->dt.dl.dr_data = data_old;
1193 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1194 list_create(&dr->dt.di.dr_children,
1195 sizeof (dbuf_dirty_record_t),
1196 offsetof(dbuf_dirty_record_t, dr_dirty_node));
1199 dr->dr_txg = tx->tx_txg;
1204 * We could have been freed_in_flight between the dbuf_noread
1205 * and dbuf_dirty. We win, as though the dbuf_noread() had
1206 * happened after the free.
1208 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1209 db->db_blkid != DMU_SPILL_BLKID) {
1210 mutex_enter(&dn->dn_mtx);
1211 dnode_clear_range(dn, db->db_blkid, 1, tx);
1212 mutex_exit(&dn->dn_mtx);
1213 db->db_freed_in_flight = FALSE;
1217 * This buffer is now part of this txg
1219 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1220 db->db_dirtycnt += 1;
1221 ASSERT3U(db->db_dirtycnt, <=, 3);
1223 mutex_exit(&db->db_mtx);
1225 if (db->db_blkid == DMU_BONUS_BLKID ||
1226 db->db_blkid == DMU_SPILL_BLKID) {
1227 mutex_enter(&dn->dn_mtx);
1228 ASSERT(!list_link_active(&dr->dr_dirty_node));
1229 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1230 mutex_exit(&dn->dn_mtx);
1231 dnode_setdirty(dn, tx);
1234 } else if (do_free_accounting) {
1235 blkptr_t *bp = db->db_blkptr;
1236 int64_t willfree = (bp && !BP_IS_HOLE(bp)) ?
1237 bp_get_dsize(os->os_spa, bp) : db->db.db_size;
1239 * This is only a guess -- if the dbuf is dirty
1240 * in a previous txg, we don't know how much
1241 * space it will use on disk yet. We should
1242 * really have the struct_rwlock to access
1243 * db_blkptr, but since this is just a guess,
1244 * it's OK if we get an odd answer.
1246 ddt_prefetch(os->os_spa, bp);
1247 dnode_willuse_space(dn, -willfree, tx);
1250 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1251 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1252 drop_struct_lock = TRUE;
1255 if (db->db_level == 0) {
1256 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1257 ASSERT(dn->dn_maxblkid >= db->db_blkid);
1260 if (db->db_level+1 < dn->dn_nlevels) {
1261 dmu_buf_impl_t *parent = db->db_parent;
1262 dbuf_dirty_record_t *di;
1263 int parent_held = FALSE;
1265 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1266 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1268 parent = dbuf_hold_level(dn, db->db_level+1,
1269 db->db_blkid >> epbs, FTAG);
1270 ASSERT(parent != NULL);
1273 if (drop_struct_lock)
1274 rw_exit(&dn->dn_struct_rwlock);
1275 ASSERT3U(db->db_level+1, ==, parent->db_level);
1276 di = dbuf_dirty(parent, tx);
1278 dbuf_rele(parent, FTAG);
1280 mutex_enter(&db->db_mtx);
1281 /* possible race with dbuf_undirty() */
1282 if (db->db_last_dirty == dr ||
1283 dn->dn_object == DMU_META_DNODE_OBJECT) {
1284 mutex_enter(&di->dt.di.dr_mtx);
1285 ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1286 ASSERT(!list_link_active(&dr->dr_dirty_node));
1287 list_insert_tail(&di->dt.di.dr_children, dr);
1288 mutex_exit(&di->dt.di.dr_mtx);
1291 mutex_exit(&db->db_mtx);
1293 ASSERT(db->db_level+1 == dn->dn_nlevels);
1294 ASSERT(db->db_blkid < dn->dn_nblkptr);
1295 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1296 mutex_enter(&dn->dn_mtx);
1297 ASSERT(!list_link_active(&dr->dr_dirty_node));
1298 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1299 mutex_exit(&dn->dn_mtx);
1300 if (drop_struct_lock)
1301 rw_exit(&dn->dn_struct_rwlock);
1304 dnode_setdirty(dn, tx);
1310 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1313 uint64_t txg = tx->tx_txg;
1314 dbuf_dirty_record_t *dr, **drp;
1317 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1319 mutex_enter(&db->db_mtx);
1321 * If this buffer is not dirty, we're done.
1323 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1324 if (dr->dr_txg <= txg)
1326 if (dr == NULL || dr->dr_txg < txg) {
1327 mutex_exit(&db->db_mtx);
1330 ASSERT(dr->dr_txg == txg);
1331 ASSERT(dr->dr_dbuf == db);
1337 * If this buffer is currently held, we cannot undirty
1338 * it, since one of the current holders may be in the
1339 * middle of an update. Note that users of dbuf_undirty()
1340 * should not place a hold on the dbuf before the call.
1341 * Also note: we can get here with a spill block, so
1342 * test for that similar to how dbuf_dirty does.
1344 if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
1345 mutex_exit(&db->db_mtx);
1346 /* Make sure we don't toss this buffer at sync phase */
1347 if (db->db_blkid != DMU_SPILL_BLKID) {
1348 mutex_enter(&dn->dn_mtx);
1349 dnode_clear_range(dn, db->db_blkid, 1, tx);
1350 mutex_exit(&dn->dn_mtx);
1356 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1358 ASSERT(db->db.db_size != 0);
1360 /* XXX would be nice to fix up dn_towrite_space[] */
1365 * Note that there are three places in dbuf_dirty()
1366 * where this dirty record may be put on a list.
1367 * Make sure to do a list_remove corresponding to
1368 * every one of those list_insert calls.
1370 if (dr->dr_parent) {
1371 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1372 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1373 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1374 } else if (db->db_blkid == DMU_SPILL_BLKID ||
1375 db->db_level+1 == dn->dn_nlevels) {
1376 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1377 mutex_enter(&dn->dn_mtx);
1378 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1379 mutex_exit(&dn->dn_mtx);
1383 if (db->db_level == 0) {
1384 if (db->db_state != DB_NOFILL) {
1385 dbuf_unoverride(dr);
1387 ASSERT(db->db_buf != NULL);
1388 ASSERT(dr->dt.dl.dr_data != NULL);
1389 if (dr->dt.dl.dr_data != db->db_buf)
1390 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data,
1394 ASSERT(db->db_buf != NULL);
1395 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
1396 mutex_destroy(&dr->dt.di.dr_mtx);
1397 list_destroy(&dr->dt.di.dr_children);
1399 kmem_free(dr, sizeof (dbuf_dirty_record_t));
1401 ASSERT(db->db_dirtycnt > 0);
1402 db->db_dirtycnt -= 1;
1404 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1405 arc_buf_t *buf = db->db_buf;
1407 ASSERT(db->db_state == DB_NOFILL || arc_released(buf));
1408 dbuf_set_data(db, NULL);
1409 VERIFY(arc_buf_remove_ref(buf, db) == 1);
1414 mutex_exit(&db->db_mtx);
1418 #pragma weak dmu_buf_will_dirty = dbuf_will_dirty
1420 dbuf_will_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1422 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1424 ASSERT(tx->tx_txg != 0);
1425 ASSERT(!refcount_is_zero(&db->db_holds));
1428 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
1429 rf |= DB_RF_HAVESTRUCT;
1431 (void) dbuf_read(db, NULL, rf);
1432 (void) dbuf_dirty(db, tx);
1436 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1438 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1440 db->db_state = DB_NOFILL;
1442 dmu_buf_will_fill(db_fake, tx);
1446 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1448 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1450 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1451 ASSERT(tx->tx_txg != 0);
1452 ASSERT(db->db_level == 0);
1453 ASSERT(!refcount_is_zero(&db->db_holds));
1455 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1456 dmu_tx_private_ok(tx));
1459 (void) dbuf_dirty(db, tx);
1462 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1465 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1467 mutex_enter(&db->db_mtx);
1470 if (db->db_state == DB_FILL) {
1471 if (db->db_level == 0 && db->db_freed_in_flight) {
1472 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1473 /* we were freed while filling */
1474 /* XXX dbuf_undirty? */
1475 bzero(db->db.db_data, db->db.db_size);
1476 db->db_freed_in_flight = FALSE;
1478 db->db_state = DB_CACHED;
1479 cv_broadcast(&db->db_changed);
1481 mutex_exit(&db->db_mtx);
1485 * Directly assign a provided arc buf to a given dbuf if it's not referenced
1486 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1489 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
1491 ASSERT(!refcount_is_zero(&db->db_holds));
1492 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1493 ASSERT(db->db_level == 0);
1494 ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA);
1495 ASSERT(buf != NULL);
1496 ASSERT(arc_buf_size(buf) == db->db.db_size);
1497 ASSERT(tx->tx_txg != 0);
1499 arc_return_buf(buf, db);
1500 ASSERT(arc_released(buf));
1502 mutex_enter(&db->db_mtx);
1504 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1505 cv_wait(&db->db_changed, &db->db_mtx);
1507 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
1509 if (db->db_state == DB_CACHED &&
1510 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
1511 mutex_exit(&db->db_mtx);
1512 (void) dbuf_dirty(db, tx);
1513 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
1514 VERIFY(arc_buf_remove_ref(buf, db) == 1);
1515 xuio_stat_wbuf_copied();
1519 xuio_stat_wbuf_nocopy();
1520 if (db->db_state == DB_CACHED) {
1521 dbuf_dirty_record_t *dr = db->db_last_dirty;
1523 ASSERT(db->db_buf != NULL);
1524 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
1525 ASSERT(dr->dt.dl.dr_data == db->db_buf);
1526 if (!arc_released(db->db_buf)) {
1527 ASSERT(dr->dt.dl.dr_override_state ==
1529 arc_release(db->db_buf, db);
1531 dr->dt.dl.dr_data = buf;
1532 VERIFY(arc_buf_remove_ref(db->db_buf, db) == 1);
1533 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
1534 arc_release(db->db_buf, db);
1535 VERIFY(arc_buf_remove_ref(db->db_buf, db) == 1);
1539 ASSERT(db->db_buf == NULL);
1540 dbuf_set_data(db, buf);
1541 db->db_state = DB_FILL;
1542 mutex_exit(&db->db_mtx);
1543 (void) dbuf_dirty(db, tx);
1544 dbuf_fill_done(db, tx);
1548 * "Clear" the contents of this dbuf. This will mark the dbuf
1549 * EVICTING and clear *most* of its references. Unfortunetely,
1550 * when we are not holding the dn_dbufs_mtx, we can't clear the
1551 * entry in the dn_dbufs list. We have to wait until dbuf_destroy()
1552 * in this case. For callers from the DMU we will usually see:
1553 * dbuf_clear()->arc_buf_evict()->dbuf_do_evict()->dbuf_destroy()
1554 * For the arc callback, we will usually see:
1555 * dbuf_do_evict()->dbuf_clear();dbuf_destroy()
1556 * Sometimes, though, we will get a mix of these two:
1557 * DMU: dbuf_clear()->arc_buf_evict()
1558 * ARC: dbuf_do_evict()->dbuf_destroy()
1561 dbuf_clear(dmu_buf_impl_t *db)
1564 dmu_buf_impl_t *parent = db->db_parent;
1565 dmu_buf_impl_t *dndb;
1566 int dbuf_gone = FALSE;
1568 ASSERT(MUTEX_HELD(&db->db_mtx));
1569 ASSERT(refcount_is_zero(&db->db_holds));
1571 dbuf_evict_user(db);
1573 if (db->db_state == DB_CACHED) {
1574 ASSERT(db->db.db_data != NULL);
1575 if (db->db_blkid == DMU_BONUS_BLKID) {
1576 zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
1577 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
1579 db->db.db_data = NULL;
1580 db->db_state = DB_UNCACHED;
1583 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
1584 ASSERT(db->db_data_pending == NULL);
1586 db->db_state = DB_EVICTING;
1587 db->db_blkptr = NULL;
1592 if (db->db_blkid != DMU_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) {
1593 list_remove(&dn->dn_dbufs, db);
1594 (void) atomic_dec_32_nv(&dn->dn_dbufs_count);
1598 * Decrementing the dbuf count means that the hold corresponding
1599 * to the removed dbuf is no longer discounted in dnode_move(),
1600 * so the dnode cannot be moved until after we release the hold.
1601 * The membar_producer() ensures visibility of the decremented
1602 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
1606 db->db_dnode_handle = NULL;
1612 dbuf_gone = arc_buf_evict(db->db_buf);
1615 mutex_exit(&db->db_mtx);
1618 * If this dbuf is referenced from an indirect dbuf,
1619 * decrement the ref count on the indirect dbuf.
1621 if (parent && parent != dndb)
1622 dbuf_rele(parent, db);
1625 __attribute__((always_inline))
1627 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
1628 dmu_buf_impl_t **parentp, blkptr_t **bpp, struct dbuf_hold_impl_data *dh)
1635 ASSERT(blkid != DMU_BONUS_BLKID);
1637 if (blkid == DMU_SPILL_BLKID) {
1638 mutex_enter(&dn->dn_mtx);
1639 if (dn->dn_have_spill &&
1640 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
1641 *bpp = &dn->dn_phys->dn_spill;
1644 dbuf_add_ref(dn->dn_dbuf, NULL);
1645 *parentp = dn->dn_dbuf;
1646 mutex_exit(&dn->dn_mtx);
1650 if (dn->dn_phys->dn_nlevels == 0)
1653 nlevels = dn->dn_phys->dn_nlevels;
1655 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1657 ASSERT3U(level * epbs, <, 64);
1658 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1659 if (level >= nlevels ||
1660 (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
1661 /* the buffer has no parent yet */
1663 } else if (level < nlevels-1) {
1664 /* this block is referenced from an indirect block */
1667 err = dbuf_hold_impl(dn, level+1, blkid >> epbs,
1668 fail_sparse, NULL, parentp);
1671 __dbuf_hold_impl_init(dh + 1, dn, dh->dh_level + 1,
1672 blkid >> epbs, fail_sparse, NULL,
1673 parentp, dh->dh_depth + 1);
1674 err = __dbuf_hold_impl(dh + 1);
1678 err = dbuf_read(*parentp, NULL,
1679 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
1681 dbuf_rele(*parentp, NULL);
1685 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
1686 (blkid & ((1ULL << epbs) - 1));
1689 /* the block is referenced from the dnode */
1690 ASSERT3U(level, ==, nlevels-1);
1691 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
1692 blkid < dn->dn_phys->dn_nblkptr);
1694 dbuf_add_ref(dn->dn_dbuf, NULL);
1695 *parentp = dn->dn_dbuf;
1697 *bpp = &dn->dn_phys->dn_blkptr[blkid];
1702 static dmu_buf_impl_t *
1703 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
1704 dmu_buf_impl_t *parent, blkptr_t *blkptr)
1706 objset_t *os = dn->dn_objset;
1707 dmu_buf_impl_t *db, *odb;
1709 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1710 ASSERT(dn->dn_type != DMU_OT_NONE);
1712 db = kmem_cache_alloc(dbuf_cache, KM_PUSHPAGE);
1715 db->db.db_object = dn->dn_object;
1716 db->db_level = level;
1717 db->db_blkid = blkid;
1718 db->db_last_dirty = NULL;
1719 db->db_dirtycnt = 0;
1720 db->db_dnode_handle = dn->dn_handle;
1721 db->db_parent = parent;
1722 db->db_blkptr = blkptr;
1724 db->db_user_ptr = NULL;
1725 db->db_user_data_ptr_ptr = NULL;
1726 db->db_evict_func = NULL;
1727 db->db_immediate_evict = 0;
1728 db->db_freed_in_flight = 0;
1730 if (blkid == DMU_BONUS_BLKID) {
1731 ASSERT3P(parent, ==, dn->dn_dbuf);
1732 db->db.db_size = DN_MAX_BONUSLEN -
1733 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
1734 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
1735 db->db.db_offset = DMU_BONUS_BLKID;
1736 db->db_state = DB_UNCACHED;
1737 /* the bonus dbuf is not placed in the hash table */
1738 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1740 } else if (blkid == DMU_SPILL_BLKID) {
1741 db->db.db_size = (blkptr != NULL) ?
1742 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
1743 db->db.db_offset = 0;
1746 db->db_level ? 1<<dn->dn_indblkshift : dn->dn_datablksz;
1747 db->db.db_size = blocksize;
1748 db->db.db_offset = db->db_blkid * blocksize;
1752 * Hold the dn_dbufs_mtx while we get the new dbuf
1753 * in the hash table *and* added to the dbufs list.
1754 * This prevents a possible deadlock with someone
1755 * trying to look up this dbuf before its added to the
1758 mutex_enter(&dn->dn_dbufs_mtx);
1759 db->db_state = DB_EVICTING;
1760 if ((odb = dbuf_hash_insert(db)) != NULL) {
1761 /* someone else inserted it first */
1762 kmem_cache_free(dbuf_cache, db);
1763 mutex_exit(&dn->dn_dbufs_mtx);
1766 list_insert_head(&dn->dn_dbufs, db);
1767 db->db_state = DB_UNCACHED;
1768 mutex_exit(&dn->dn_dbufs_mtx);
1769 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1771 if (parent && parent != dn->dn_dbuf)
1772 dbuf_add_ref(parent, db);
1774 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1775 refcount_count(&dn->dn_holds) > 0);
1776 (void) refcount_add(&dn->dn_holds, db);
1777 (void) atomic_inc_32_nv(&dn->dn_dbufs_count);
1779 dprintf_dbuf(db, "db=%p\n", db);
1785 dbuf_do_evict(void *private)
1787 arc_buf_t *buf = private;
1788 dmu_buf_impl_t *db = buf->b_private;
1790 if (!MUTEX_HELD(&db->db_mtx))
1791 mutex_enter(&db->db_mtx);
1793 ASSERT(refcount_is_zero(&db->db_holds));
1795 if (db->db_state != DB_EVICTING) {
1796 ASSERT(db->db_state == DB_CACHED);
1801 mutex_exit(&db->db_mtx);
1808 dbuf_destroy(dmu_buf_impl_t *db)
1810 ASSERT(refcount_is_zero(&db->db_holds));
1812 if (db->db_blkid != DMU_BONUS_BLKID) {
1814 * If this dbuf is still on the dn_dbufs list,
1815 * remove it from that list.
1817 if (db->db_dnode_handle != NULL) {
1822 mutex_enter(&dn->dn_dbufs_mtx);
1823 list_remove(&dn->dn_dbufs, db);
1824 (void) atomic_dec_32_nv(&dn->dn_dbufs_count);
1825 mutex_exit(&dn->dn_dbufs_mtx);
1828 * Decrementing the dbuf count means that the hold
1829 * corresponding to the removed dbuf is no longer
1830 * discounted in dnode_move(), so the dnode cannot be
1831 * moved until after we release the hold.
1834 db->db_dnode_handle = NULL;
1836 dbuf_hash_remove(db);
1838 db->db_parent = NULL;
1841 ASSERT(!list_link_active(&db->db_link));
1842 ASSERT(db->db.db_data == NULL);
1843 ASSERT(db->db_hash_next == NULL);
1844 ASSERT(db->db_blkptr == NULL);
1845 ASSERT(db->db_data_pending == NULL);
1847 kmem_cache_free(dbuf_cache, db);
1848 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1852 dbuf_prefetch(dnode_t *dn, uint64_t blkid)
1854 dmu_buf_impl_t *db = NULL;
1855 blkptr_t *bp = NULL;
1857 ASSERT(blkid != DMU_BONUS_BLKID);
1858 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1860 if (dnode_block_freed(dn, blkid))
1863 /* dbuf_find() returns with db_mtx held */
1864 if ((db = dbuf_find(dn, 0, blkid))) {
1866 * This dbuf is already in the cache. We assume that
1867 * it is already CACHED, or else about to be either
1870 mutex_exit(&db->db_mtx);
1874 if (dbuf_findbp(dn, 0, blkid, TRUE, &db, &bp, NULL) == 0) {
1875 if (bp && !BP_IS_HOLE(bp)) {
1876 int priority = dn->dn_type == DMU_OT_DDT_ZAP ?
1877 ZIO_PRIORITY_DDT_PREFETCH : ZIO_PRIORITY_ASYNC_READ;
1878 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
1879 uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH;
1882 SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
1883 dn->dn_object, 0, blkid);
1885 (void) arc_read(NULL, dn->dn_objset->os_spa,
1886 bp, NULL, NULL, priority,
1887 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1891 dbuf_rele(db, NULL);
1895 #define DBUF_HOLD_IMPL_MAX_DEPTH 20
1898 * Returns with db_holds incremented, and db_mtx not held.
1899 * Note: dn_struct_rwlock must be held.
1902 __dbuf_hold_impl(struct dbuf_hold_impl_data *dh)
1904 ASSERT3S(dh->dh_depth, <, DBUF_HOLD_IMPL_MAX_DEPTH);
1905 dh->dh_parent = NULL;
1907 ASSERT(dh->dh_blkid != DMU_BONUS_BLKID);
1908 ASSERT(RW_LOCK_HELD(&dh->dh_dn->dn_struct_rwlock));
1909 ASSERT3U(dh->dh_dn->dn_nlevels, >, dh->dh_level);
1911 *(dh->dh_dbp) = NULL;
1913 /* dbuf_find() returns with db_mtx held */
1914 dh->dh_db = dbuf_find(dh->dh_dn, dh->dh_level, dh->dh_blkid);
1916 if (dh->dh_db == NULL) {
1919 ASSERT3P(dh->dh_parent, ==, NULL);
1920 dh->dh_err = dbuf_findbp(dh->dh_dn, dh->dh_level, dh->dh_blkid,
1921 dh->dh_fail_sparse, &dh->dh_parent,
1923 if (dh->dh_fail_sparse) {
1924 if (dh->dh_err == 0 && dh->dh_bp && BP_IS_HOLE(dh->dh_bp))
1925 dh->dh_err = ENOENT;
1928 dbuf_rele(dh->dh_parent, NULL);
1929 return (dh->dh_err);
1932 if (dh->dh_err && dh->dh_err != ENOENT)
1933 return (dh->dh_err);
1934 dh->dh_db = dbuf_create(dh->dh_dn, dh->dh_level, dh->dh_blkid,
1935 dh->dh_parent, dh->dh_bp);
1938 if (dh->dh_db->db_buf && refcount_is_zero(&dh->dh_db->db_holds)) {
1939 arc_buf_add_ref(dh->dh_db->db_buf, dh->dh_db);
1940 if (dh->dh_db->db_buf->b_data == NULL) {
1941 dbuf_clear(dh->dh_db);
1942 if (dh->dh_parent) {
1943 dbuf_rele(dh->dh_parent, NULL);
1944 dh->dh_parent = NULL;
1948 ASSERT3P(dh->dh_db->db.db_data, ==, dh->dh_db->db_buf->b_data);
1951 ASSERT(dh->dh_db->db_buf == NULL || arc_referenced(dh->dh_db->db_buf));
1954 * If this buffer is currently syncing out, and we are are
1955 * still referencing it from db_data, we need to make a copy
1956 * of it in case we decide we want to dirty it again in this txg.
1958 if (dh->dh_db->db_level == 0 &&
1959 dh->dh_db->db_blkid != DMU_BONUS_BLKID &&
1960 dh->dh_dn->dn_object != DMU_META_DNODE_OBJECT &&
1961 dh->dh_db->db_state == DB_CACHED && dh->dh_db->db_data_pending) {
1962 dh->dh_dr = dh->dh_db->db_data_pending;
1964 if (dh->dh_dr->dt.dl.dr_data == dh->dh_db->db_buf) {
1965 dh->dh_type = DBUF_GET_BUFC_TYPE(dh->dh_db);
1967 dbuf_set_data(dh->dh_db,
1968 arc_buf_alloc(dh->dh_dn->dn_objset->os_spa,
1969 dh->dh_db->db.db_size, dh->dh_db, dh->dh_type));
1970 bcopy(dh->dh_dr->dt.dl.dr_data->b_data,
1971 dh->dh_db->db.db_data, dh->dh_db->db.db_size);
1975 (void) refcount_add(&dh->dh_db->db_holds, dh->dh_tag);
1976 dbuf_update_data(dh->dh_db);
1977 DBUF_VERIFY(dh->dh_db);
1978 mutex_exit(&dh->dh_db->db_mtx);
1980 /* NOTE: we can't rele the parent until after we drop the db_mtx */
1982 dbuf_rele(dh->dh_parent, NULL);
1984 ASSERT3P(DB_DNODE(dh->dh_db), ==, dh->dh_dn);
1985 ASSERT3U(dh->dh_db->db_blkid, ==, dh->dh_blkid);
1986 ASSERT3U(dh->dh_db->db_level, ==, dh->dh_level);
1987 *(dh->dh_dbp) = dh->dh_db;
1993 * The following code preserves the recursive function dbuf_hold_impl()
1994 * but moves the local variables AND function arguments to the heap to
1995 * minimize the stack frame size. Enough space is initially allocated
1996 * on the stack for 20 levels of recursion.
1999 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse,
2000 void *tag, dmu_buf_impl_t **dbp)
2002 struct dbuf_hold_impl_data *dh;
2005 dh = kmem_zalloc(sizeof(struct dbuf_hold_impl_data) *
2006 DBUF_HOLD_IMPL_MAX_DEPTH, KM_PUSHPAGE);
2007 __dbuf_hold_impl_init(dh, dn, level, blkid, fail_sparse, tag, dbp, 0);
2009 error = __dbuf_hold_impl(dh);
2011 kmem_free(dh, sizeof(struct dbuf_hold_impl_data) *
2012 DBUF_HOLD_IMPL_MAX_DEPTH);
2018 __dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh,
2019 dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse,
2020 void *tag, dmu_buf_impl_t **dbp, int depth)
2023 dh->dh_level = level;
2024 dh->dh_blkid = blkid;
2025 dh->dh_fail_sparse = fail_sparse;
2028 dh->dh_depth = depth;
2032 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
2035 int err = dbuf_hold_impl(dn, 0, blkid, FALSE, tag, &db);
2036 return (err ? NULL : db);
2040 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
2043 int err = dbuf_hold_impl(dn, level, blkid, FALSE, tag, &db);
2044 return (err ? NULL : db);
2048 dbuf_create_bonus(dnode_t *dn)
2050 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
2052 ASSERT(dn->dn_bonus == NULL);
2053 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
2057 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
2059 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2062 if (db->db_blkid != DMU_SPILL_BLKID)
2065 blksz = SPA_MINBLOCKSIZE;
2066 if (blksz > SPA_MAXBLOCKSIZE)
2067 blksz = SPA_MAXBLOCKSIZE;
2069 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
2073 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2074 dbuf_new_size(db, blksz, tx);
2075 rw_exit(&dn->dn_struct_rwlock);
2082 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
2084 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
2087 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2089 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
2091 VERIFY(refcount_add(&db->db_holds, tag) > 1);
2095 * If you call dbuf_rele() you had better not be referencing the dnode handle
2096 * unless you have some other direct or indirect hold on the dnode. (An indirect
2097 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2098 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2099 * dnode's parent dbuf evicting its dnode handles.
2101 #pragma weak dmu_buf_rele = dbuf_rele
2103 dbuf_rele(dmu_buf_impl_t *db, void *tag)
2105 mutex_enter(&db->db_mtx);
2106 dbuf_rele_and_unlock(db, tag);
2110 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2111 * db_dirtycnt and db_holds to be updated atomically.
2114 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2118 ASSERT(MUTEX_HELD(&db->db_mtx));
2122 * Remove the reference to the dbuf before removing its hold on the
2123 * dnode so we can guarantee in dnode_move() that a referenced bonus
2124 * buffer has a corresponding dnode hold.
2126 holds = refcount_remove(&db->db_holds, tag);
2130 * We can't freeze indirects if there is a possibility that they
2131 * may be modified in the current syncing context.
2133 if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0))
2134 arc_buf_freeze(db->db_buf);
2136 if (holds == db->db_dirtycnt &&
2137 db->db_level == 0 && db->db_immediate_evict)
2138 dbuf_evict_user(db);
2141 if (db->db_blkid == DMU_BONUS_BLKID) {
2142 mutex_exit(&db->db_mtx);
2145 * If the dnode moves here, we cannot cross this barrier
2146 * until the move completes.
2149 (void) atomic_dec_32_nv(&DB_DNODE(db)->dn_dbufs_count);
2152 * The bonus buffer's dnode hold is no longer discounted
2153 * in dnode_move(). The dnode cannot move until after
2156 dnode_rele(DB_DNODE(db), db);
2157 } else if (db->db_buf == NULL) {
2159 * This is a special case: we never associated this
2160 * dbuf with any data allocated from the ARC.
2162 ASSERT(db->db_state == DB_UNCACHED ||
2163 db->db_state == DB_NOFILL);
2165 } else if (arc_released(db->db_buf)) {
2166 arc_buf_t *buf = db->db_buf;
2168 * This dbuf has anonymous data associated with it.
2170 dbuf_set_data(db, NULL);
2171 VERIFY(arc_buf_remove_ref(buf, db) == 1);
2174 VERIFY(arc_buf_remove_ref(db->db_buf, db) == 0);
2177 * A dbuf will be eligible for eviction if either the
2178 * 'primarycache' property is set or a duplicate
2179 * copy of this buffer is already cached in the arc.
2181 * In the case of the 'primarycache' a buffer
2182 * is considered for eviction if it matches the
2183 * criteria set in the property.
2185 * To decide if our buffer is considered a
2186 * duplicate, we must call into the arc to determine
2187 * if multiple buffers are referencing the same
2188 * block on-disk. If so, then we simply evict
2191 if (!DBUF_IS_CACHEABLE(db) ||
2192 arc_buf_eviction_needed(db->db_buf))
2195 mutex_exit(&db->db_mtx);
2198 mutex_exit(&db->db_mtx);
2202 #pragma weak dmu_buf_refcount = dbuf_refcount
2204 dbuf_refcount(dmu_buf_impl_t *db)
2206 return (refcount_count(&db->db_holds));
2210 dmu_buf_set_user(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr,
2211 dmu_buf_evict_func_t *evict_func)
2213 return (dmu_buf_update_user(db_fake, NULL, user_ptr,
2214 user_data_ptr_ptr, evict_func));
2218 dmu_buf_set_user_ie(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr,
2219 dmu_buf_evict_func_t *evict_func)
2221 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2223 db->db_immediate_evict = TRUE;
2224 return (dmu_buf_update_user(db_fake, NULL, user_ptr,
2225 user_data_ptr_ptr, evict_func));
2229 dmu_buf_update_user(dmu_buf_t *db_fake, void *old_user_ptr, void *user_ptr,
2230 void *user_data_ptr_ptr, dmu_buf_evict_func_t *evict_func)
2232 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2233 ASSERT(db->db_level == 0);
2235 ASSERT((user_ptr == NULL) == (evict_func == NULL));
2237 mutex_enter(&db->db_mtx);
2239 if (db->db_user_ptr == old_user_ptr) {
2240 db->db_user_ptr = user_ptr;
2241 db->db_user_data_ptr_ptr = user_data_ptr_ptr;
2242 db->db_evict_func = evict_func;
2244 dbuf_update_data(db);
2246 old_user_ptr = db->db_user_ptr;
2249 mutex_exit(&db->db_mtx);
2250 return (old_user_ptr);
2254 dmu_buf_get_user(dmu_buf_t *db_fake)
2256 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2257 ASSERT(!refcount_is_zero(&db->db_holds));
2259 return (db->db_user_ptr);
2263 dmu_buf_freeable(dmu_buf_t *dbuf)
2265 boolean_t res = B_FALSE;
2266 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2269 res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset,
2270 db->db_blkptr, db->db_blkptr->blk_birth);
2276 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
2278 /* ASSERT(dmu_tx_is_syncing(tx) */
2279 ASSERT(MUTEX_HELD(&db->db_mtx));
2281 if (db->db_blkptr != NULL)
2284 if (db->db_blkid == DMU_SPILL_BLKID) {
2285 db->db_blkptr = &dn->dn_phys->dn_spill;
2286 BP_ZERO(db->db_blkptr);
2289 if (db->db_level == dn->dn_phys->dn_nlevels-1) {
2291 * This buffer was allocated at a time when there was
2292 * no available blkptrs from the dnode, or it was
2293 * inappropriate to hook it in (i.e., nlevels mis-match).
2295 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
2296 ASSERT(db->db_parent == NULL);
2297 db->db_parent = dn->dn_dbuf;
2298 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
2301 dmu_buf_impl_t *parent = db->db_parent;
2302 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2304 ASSERT(dn->dn_phys->dn_nlevels > 1);
2305 if (parent == NULL) {
2306 mutex_exit(&db->db_mtx);
2307 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2308 (void) dbuf_hold_impl(dn, db->db_level+1,
2309 db->db_blkid >> epbs, FALSE, db, &parent);
2310 rw_exit(&dn->dn_struct_rwlock);
2311 mutex_enter(&db->db_mtx);
2312 db->db_parent = parent;
2314 db->db_blkptr = (blkptr_t *)parent->db.db_data +
2315 (db->db_blkid & ((1ULL << epbs) - 1));
2320 /* dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it
2321 * is critical the we not allow the compiler to inline this function in to
2322 * dbuf_sync_list() thereby drastically bloating the stack usage.
2324 noinline static void
2325 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2327 dmu_buf_impl_t *db = dr->dr_dbuf;
2331 ASSERT(dmu_tx_is_syncing(tx));
2333 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2335 mutex_enter(&db->db_mtx);
2337 ASSERT(db->db_level > 0);
2340 if (db->db_buf == NULL) {
2341 mutex_exit(&db->db_mtx);
2342 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
2343 mutex_enter(&db->db_mtx);
2345 ASSERT3U(db->db_state, ==, DB_CACHED);
2346 ASSERT(db->db_buf != NULL);
2350 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2351 dbuf_check_blkptr(dn, db);
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);
2745 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
2747 dmu_buf_impl_t *db = dr->dr_dbuf;
2750 dmu_buf_impl_t *parent = db->db_parent;
2751 uint64_t txg = tx->tx_txg;
2761 if (db->db_state != DB_NOFILL) {
2762 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
2764 * Private object buffers are released here rather
2765 * than in dbuf_dirty() since they are only modified
2766 * in the syncing context and we don't want the
2767 * overhead of making multiple copies of the data.
2769 if (BP_IS_HOLE(db->db_blkptr)) {
2772 dbuf_release_bp(db);
2777 if (parent != dn->dn_dbuf) {
2778 ASSERT(parent && parent->db_data_pending);
2779 ASSERT(db->db_level == parent->db_level-1);
2780 ASSERT(arc_released(parent->db_buf));
2781 zio = parent->db_data_pending->dr_zio;
2783 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
2784 db->db_blkid != DMU_SPILL_BLKID) ||
2785 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
2786 if (db->db_blkid != DMU_SPILL_BLKID)
2787 ASSERT3P(db->db_blkptr, ==,
2788 &dn->dn_phys->dn_blkptr[db->db_blkid]);
2792 ASSERT(db->db_level == 0 || data == db->db_buf);
2793 ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
2796 SET_BOOKMARK(&zb, os->os_dsl_dataset ?
2797 os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
2798 db->db.db_object, db->db_level, db->db_blkid);
2800 if (db->db_blkid == DMU_SPILL_BLKID)
2802 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
2804 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
2807 if (db->db_level == 0 && dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
2808 ASSERT(db->db_state != DB_NOFILL);
2809 dr->dr_zio = zio_write(zio, os->os_spa, txg,
2810 db->db_blkptr, data->b_data, arc_buf_size(data), &zp,
2811 dbuf_write_override_ready, dbuf_write_override_done, dr,
2812 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
2813 mutex_enter(&db->db_mtx);
2814 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
2815 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
2816 dr->dt.dl.dr_copies);
2817 mutex_exit(&db->db_mtx);
2818 } else if (db->db_state == DB_NOFILL) {
2819 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF);
2820 dr->dr_zio = zio_write(zio, os->os_spa, txg,
2821 db->db_blkptr, NULL, db->db.db_size, &zp,
2822 dbuf_write_nofill_ready, dbuf_write_nofill_done, db,
2823 ZIO_PRIORITY_ASYNC_WRITE,
2824 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
2826 ASSERT(arc_released(data));
2827 dr->dr_zio = arc_write(zio, os->os_spa, txg,
2828 db->db_blkptr, data, DBUF_IS_L2CACHEABLE(db),
2829 DBUF_IS_L2COMPRESSIBLE(db), &zp, dbuf_write_ready,
2830 dbuf_write_done, db, ZIO_PRIORITY_ASYNC_WRITE,
2831 ZIO_FLAG_MUSTSUCCEED, &zb);
2835 #if defined(_KERNEL) && defined(HAVE_SPL)
2836 EXPORT_SYMBOL(dbuf_find);
2837 EXPORT_SYMBOL(dbuf_is_metadata);
2838 EXPORT_SYMBOL(dbuf_evict);
2839 EXPORT_SYMBOL(dbuf_loan_arcbuf);
2840 EXPORT_SYMBOL(dbuf_whichblock);
2841 EXPORT_SYMBOL(dbuf_read);
2842 EXPORT_SYMBOL(dbuf_unoverride);
2843 EXPORT_SYMBOL(dbuf_free_range);
2844 EXPORT_SYMBOL(dbuf_new_size);
2845 EXPORT_SYMBOL(dbuf_release_bp);
2846 EXPORT_SYMBOL(dbuf_dirty);
2847 EXPORT_SYMBOL(dmu_buf_will_dirty);
2848 EXPORT_SYMBOL(dmu_buf_will_not_fill);
2849 EXPORT_SYMBOL(dmu_buf_will_fill);
2850 EXPORT_SYMBOL(dmu_buf_fill_done);
2851 EXPORT_SYMBOL(dmu_buf_rele);
2852 EXPORT_SYMBOL(dbuf_assign_arcbuf);
2853 EXPORT_SYMBOL(dbuf_clear);
2854 EXPORT_SYMBOL(dbuf_prefetch);
2855 EXPORT_SYMBOL(dbuf_hold_impl);
2856 EXPORT_SYMBOL(dbuf_hold);
2857 EXPORT_SYMBOL(dbuf_hold_level);
2858 EXPORT_SYMBOL(dbuf_create_bonus);
2859 EXPORT_SYMBOL(dbuf_spill_set_blksz);
2860 EXPORT_SYMBOL(dbuf_rm_spill);
2861 EXPORT_SYMBOL(dbuf_add_ref);
2862 EXPORT_SYMBOL(dbuf_rele);
2863 EXPORT_SYMBOL(dbuf_rele_and_unlock);
2864 EXPORT_SYMBOL(dbuf_refcount);
2865 EXPORT_SYMBOL(dbuf_sync_list);
2866 EXPORT_SYMBOL(dmu_buf_set_user);
2867 EXPORT_SYMBOL(dmu_buf_set_user_ie);
2868 EXPORT_SYMBOL(dmu_buf_update_user);
2869 EXPORT_SYMBOL(dmu_buf_get_user);
2870 EXPORT_SYMBOL(dmu_buf_freeable);