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, 2014 by Delphix. All rights reserved.
25 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
26 * Copyright (c) 2013, Joyent, Inc. All rights reserved.
29 #include <sys/zfs_context.h>
31 #include <sys/dmu_send.h>
32 #include <sys/dmu_impl.h>
34 #include <sys/dmu_objset.h>
35 #include <sys/dsl_dataset.h>
36 #include <sys/dsl_dir.h>
37 #include <sys/dmu_tx.h>
40 #include <sys/dmu_zfetch.h>
42 #include <sys/sa_impl.h>
43 #include <sys/zfeature.h>
44 #include <sys/blkptr.h>
45 #include <sys/range_tree.h>
48 * Number of times that zfs_free_range() took the slow path while doing
49 * a zfs receive. A nonzero value indicates a potential performance problem.
51 uint64_t zfs_free_range_recv_miss;
53 static void dbuf_destroy(dmu_buf_impl_t *db);
54 static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
55 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
58 * Global data structures and functions for the dbuf cache.
60 static kmem_cache_t *dbuf_cache;
64 dbuf_cons(void *vdb, void *unused, int kmflag)
66 dmu_buf_impl_t *db = vdb;
67 bzero(db, sizeof (dmu_buf_impl_t));
69 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
70 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
71 refcount_create(&db->db_holds);
78 dbuf_dest(void *vdb, void *unused)
80 dmu_buf_impl_t *db = vdb;
81 mutex_destroy(&db->db_mtx);
82 cv_destroy(&db->db_changed);
83 refcount_destroy(&db->db_holds);
87 * dbuf hash table routines
89 static dbuf_hash_table_t dbuf_hash_table;
91 static uint64_t dbuf_hash_count;
94 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
96 uintptr_t osv = (uintptr_t)os;
99 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
100 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF];
101 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF];
102 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF];
103 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF];
104 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF];
105 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF];
107 crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16);
112 #define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid);
114 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
115 ((dbuf)->db.db_object == (obj) && \
116 (dbuf)->db_objset == (os) && \
117 (dbuf)->db_level == (level) && \
118 (dbuf)->db_blkid == (blkid))
121 dbuf_find(dnode_t *dn, uint8_t level, uint64_t blkid)
123 dbuf_hash_table_t *h = &dbuf_hash_table;
124 objset_t *os = dn->dn_objset;
125 uint64_t obj = dn->dn_object;
126 uint64_t hv = DBUF_HASH(os, obj, level, blkid);
127 uint64_t idx = hv & h->hash_table_mask;
130 mutex_enter(DBUF_HASH_MUTEX(h, idx));
131 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
132 if (DBUF_EQUAL(db, os, obj, level, blkid)) {
133 mutex_enter(&db->db_mtx);
134 if (db->db_state != DB_EVICTING) {
135 mutex_exit(DBUF_HASH_MUTEX(h, idx));
138 mutex_exit(&db->db_mtx);
141 mutex_exit(DBUF_HASH_MUTEX(h, idx));
146 * Insert an entry into the hash table. If there is already an element
147 * equal to elem in the hash table, then the already existing element
148 * will be returned and the new element will not be inserted.
149 * Otherwise returns NULL.
151 static dmu_buf_impl_t *
152 dbuf_hash_insert(dmu_buf_impl_t *db)
154 dbuf_hash_table_t *h = &dbuf_hash_table;
155 objset_t *os = db->db_objset;
156 uint64_t obj = db->db.db_object;
157 int level = db->db_level;
158 uint64_t blkid = db->db_blkid;
159 uint64_t hv = DBUF_HASH(os, obj, level, blkid);
160 uint64_t idx = hv & h->hash_table_mask;
163 mutex_enter(DBUF_HASH_MUTEX(h, idx));
164 for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) {
165 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
166 mutex_enter(&dbf->db_mtx);
167 if (dbf->db_state != DB_EVICTING) {
168 mutex_exit(DBUF_HASH_MUTEX(h, idx));
171 mutex_exit(&dbf->db_mtx);
175 mutex_enter(&db->db_mtx);
176 db->db_hash_next = h->hash_table[idx];
177 h->hash_table[idx] = db;
178 mutex_exit(DBUF_HASH_MUTEX(h, idx));
179 atomic_inc_64(&dbuf_hash_count);
185 * Remove an entry from the hash table. It must be in the EVICTING state.
188 dbuf_hash_remove(dmu_buf_impl_t *db)
190 dbuf_hash_table_t *h = &dbuf_hash_table;
191 uint64_t hv = DBUF_HASH(db->db_objset, db->db.db_object,
192 db->db_level, db->db_blkid);
193 uint64_t idx = hv & h->hash_table_mask;
194 dmu_buf_impl_t *dbf, **dbp;
197 * We musn't hold db_mtx to maintain lock ordering:
198 * DBUF_HASH_MUTEX > db_mtx.
200 ASSERT(refcount_is_zero(&db->db_holds));
201 ASSERT(db->db_state == DB_EVICTING);
202 ASSERT(!MUTEX_HELD(&db->db_mtx));
204 mutex_enter(DBUF_HASH_MUTEX(h, idx));
205 dbp = &h->hash_table[idx];
206 while ((dbf = *dbp) != db) {
207 dbp = &dbf->db_hash_next;
210 *dbp = db->db_hash_next;
211 db->db_hash_next = NULL;
212 mutex_exit(DBUF_HASH_MUTEX(h, idx));
213 atomic_dec_64(&dbuf_hash_count);
216 static arc_evict_func_t dbuf_do_evict;
219 dbuf_evict_user(dmu_buf_impl_t *db)
221 ASSERT(MUTEX_HELD(&db->db_mtx));
223 if (db->db_level != 0 || db->db_evict_func == NULL)
226 if (db->db_user_data_ptr_ptr)
227 *db->db_user_data_ptr_ptr = db->db.db_data;
228 db->db_evict_func(&db->db, db->db_user_ptr);
229 db->db_user_ptr = NULL;
230 db->db_user_data_ptr_ptr = NULL;
231 db->db_evict_func = NULL;
235 dbuf_is_metadata(dmu_buf_impl_t *db)
237 if (db->db_level > 0) {
240 boolean_t is_metadata;
243 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type);
246 return (is_metadata);
251 dbuf_evict(dmu_buf_impl_t *db)
253 ASSERT(MUTEX_HELD(&db->db_mtx));
254 ASSERT(db->db_buf == NULL);
255 ASSERT(db->db_data_pending == NULL);
264 uint64_t hsize = 1ULL << 16;
265 dbuf_hash_table_t *h = &dbuf_hash_table;
269 * The hash table is big enough to fill all of physical memory
270 * with an average 4K block size. The table will take up
271 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
273 while (hsize * 4096 < (uint64_t)physmem * PAGESIZE)
277 h->hash_table_mask = hsize - 1;
278 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
279 if (h->hash_table == NULL) {
280 /* XXX - we should really return an error instead of assert */
281 ASSERT(hsize > (1ULL << 10));
286 dbuf_cache = kmem_cache_create("dmu_buf_impl_t",
287 sizeof (dmu_buf_impl_t),
288 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
290 for (i = 0; i < DBUF_MUTEXES; i++)
291 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
297 dbuf_hash_table_t *h = &dbuf_hash_table;
300 for (i = 0; i < DBUF_MUTEXES; i++)
301 mutex_destroy(&h->hash_mutexes[i]);
302 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
303 kmem_cache_destroy(dbuf_cache);
312 dbuf_verify(dmu_buf_impl_t *db)
315 dbuf_dirty_record_t *dr;
317 ASSERT(MUTEX_HELD(&db->db_mtx));
319 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
322 ASSERT(db->db_objset != NULL);
326 ASSERT(db->db_parent == NULL);
327 ASSERT(db->db_blkptr == NULL);
329 ASSERT3U(db->db.db_object, ==, dn->dn_object);
330 ASSERT3P(db->db_objset, ==, dn->dn_objset);
331 ASSERT3U(db->db_level, <, dn->dn_nlevels);
332 ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
333 db->db_blkid == DMU_SPILL_BLKID ||
334 !avl_is_empty(&dn->dn_dbufs));
336 if (db->db_blkid == DMU_BONUS_BLKID) {
338 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
339 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
340 } else if (db->db_blkid == DMU_SPILL_BLKID) {
342 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
343 ASSERT0(db->db.db_offset);
345 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
348 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
349 ASSERT(dr->dr_dbuf == db);
351 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
352 ASSERT(dr->dr_dbuf == db);
355 * We can't assert that db_size matches dn_datablksz because it
356 * can be momentarily different when another thread is doing
359 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
360 dr = db->db_data_pending;
362 * It should only be modified in syncing context, so
363 * make sure we only have one copy of the data.
365 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
368 /* verify db->db_blkptr */
370 if (db->db_parent == dn->dn_dbuf) {
371 /* db is pointed to by the dnode */
372 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
373 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
374 ASSERT(db->db_parent == NULL);
376 ASSERT(db->db_parent != NULL);
377 if (db->db_blkid != DMU_SPILL_BLKID)
378 ASSERT3P(db->db_blkptr, ==,
379 &dn->dn_phys->dn_blkptr[db->db_blkid]);
381 /* db is pointed to by an indirect block */
382 int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
383 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
384 ASSERT3U(db->db_parent->db.db_object, ==,
387 * dnode_grow_indblksz() can make this fail if we don't
388 * have the struct_rwlock. XXX indblksz no longer
389 * grows. safe to do this now?
391 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
392 ASSERT3P(db->db_blkptr, ==,
393 ((blkptr_t *)db->db_parent->db.db_data +
394 db->db_blkid % epb));
398 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
399 (db->db_buf == NULL || db->db_buf->b_data) &&
400 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
401 db->db_state != DB_FILL && !dn->dn_free_txg) {
403 * If the blkptr isn't set but they have nonzero data,
404 * it had better be dirty, otherwise we'll lose that
405 * data when we evict this buffer.
407 if (db->db_dirtycnt == 0) {
408 uint64_t *buf = db->db.db_data;
411 for (i = 0; i < db->db.db_size >> 3; i++) {
421 dbuf_update_data(dmu_buf_impl_t *db)
423 ASSERT(MUTEX_HELD(&db->db_mtx));
424 if (db->db_level == 0 && db->db_user_data_ptr_ptr) {
425 ASSERT(!refcount_is_zero(&db->db_holds));
426 *db->db_user_data_ptr_ptr = db->db.db_data;
431 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
433 ASSERT(MUTEX_HELD(&db->db_mtx));
436 ASSERT(buf->b_data != NULL);
437 db->db.db_data = buf->b_data;
438 if (!arc_released(buf))
439 arc_set_callback(buf, dbuf_do_evict, db);
440 dbuf_update_data(db);
443 db->db.db_data = NULL;
444 if (db->db_state != DB_NOFILL)
445 db->db_state = DB_UNCACHED;
450 * Loan out an arc_buf for read. Return the loaned arc_buf.
453 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
457 mutex_enter(&db->db_mtx);
458 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
459 int blksz = db->db.db_size;
460 spa_t *spa = db->db_objset->os_spa;
462 mutex_exit(&db->db_mtx);
463 abuf = arc_loan_buf(spa, blksz);
464 bcopy(db->db.db_data, abuf->b_data, blksz);
467 arc_loan_inuse_buf(abuf, db);
468 dbuf_set_data(db, NULL);
469 mutex_exit(&db->db_mtx);
475 dbuf_whichblock(dnode_t *dn, uint64_t offset)
477 if (dn->dn_datablkshift) {
478 return (offset >> dn->dn_datablkshift);
480 ASSERT3U(offset, <, dn->dn_datablksz);
486 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
488 dmu_buf_impl_t *db = vdb;
490 mutex_enter(&db->db_mtx);
491 ASSERT3U(db->db_state, ==, DB_READ);
493 * All reads are synchronous, so we must have a hold on the dbuf
495 ASSERT(refcount_count(&db->db_holds) > 0);
496 ASSERT(db->db_buf == NULL);
497 ASSERT(db->db.db_data == NULL);
498 if (db->db_level == 0 && db->db_freed_in_flight) {
499 /* we were freed in flight; disregard any error */
500 arc_release(buf, db);
501 bzero(buf->b_data, db->db.db_size);
503 db->db_freed_in_flight = FALSE;
504 dbuf_set_data(db, buf);
505 db->db_state = DB_CACHED;
506 } else if (zio == NULL || zio->io_error == 0) {
507 dbuf_set_data(db, buf);
508 db->db_state = DB_CACHED;
510 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
511 ASSERT3P(db->db_buf, ==, NULL);
512 VERIFY(arc_buf_remove_ref(buf, db));
513 db->db_state = DB_UNCACHED;
515 cv_broadcast(&db->db_changed);
516 dbuf_rele_and_unlock(db, NULL);
520 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t *flags)
524 uint32_t aflags = ARC_NOWAIT;
528 ASSERT(!refcount_is_zero(&db->db_holds));
529 /* We need the struct_rwlock to prevent db_blkptr from changing. */
530 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
531 ASSERT(MUTEX_HELD(&db->db_mtx));
532 ASSERT(db->db_state == DB_UNCACHED);
533 ASSERT(db->db_buf == NULL);
535 if (db->db_blkid == DMU_BONUS_BLKID) {
536 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
538 ASSERT3U(bonuslen, <=, db->db.db_size);
539 db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN);
540 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
541 if (bonuslen < DN_MAX_BONUSLEN)
542 bzero(db->db.db_data, DN_MAX_BONUSLEN);
544 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
546 dbuf_update_data(db);
547 db->db_state = DB_CACHED;
548 mutex_exit(&db->db_mtx);
553 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
554 * processes the delete record and clears the bp while we are waiting
555 * for the dn_mtx (resulting in a "no" from block_freed).
557 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
558 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
559 BP_IS_HOLE(db->db_blkptr)))) {
560 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
563 dbuf_set_data(db, arc_buf_alloc(db->db_objset->os_spa,
564 db->db.db_size, db, type));
565 bzero(db->db.db_data, db->db.db_size);
566 db->db_state = DB_CACHED;
567 *flags |= DB_RF_CACHED;
568 mutex_exit(&db->db_mtx);
574 db->db_state = DB_READ;
575 mutex_exit(&db->db_mtx);
577 if (DBUF_IS_L2CACHEABLE(db))
578 aflags |= ARC_L2CACHE;
579 if (DBUF_IS_L2COMPRESSIBLE(db))
580 aflags |= ARC_L2COMPRESS;
582 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
583 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
584 db->db.db_object, db->db_level, db->db_blkid);
586 dbuf_add_ref(db, NULL);
588 (void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr,
589 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
590 (*flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
592 if (aflags & ARC_CACHED)
593 *flags |= DB_RF_CACHED;
597 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
600 boolean_t havepzio = (zio != NULL);
605 * We don't have to hold the mutex to check db_state because it
606 * can't be freed while we have a hold on the buffer.
608 ASSERT(!refcount_is_zero(&db->db_holds));
610 if (db->db_state == DB_NOFILL)
611 return (SET_ERROR(EIO));
615 if ((flags & DB_RF_HAVESTRUCT) == 0)
616 rw_enter(&dn->dn_struct_rwlock, RW_READER);
618 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
619 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
620 DBUF_IS_CACHEABLE(db);
622 mutex_enter(&db->db_mtx);
623 if (db->db_state == DB_CACHED) {
624 mutex_exit(&db->db_mtx);
626 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
627 db->db.db_size, TRUE);
628 if ((flags & DB_RF_HAVESTRUCT) == 0)
629 rw_exit(&dn->dn_struct_rwlock);
631 } else if (db->db_state == DB_UNCACHED) {
632 spa_t *spa = dn->dn_objset->os_spa;
635 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
636 dbuf_read_impl(db, zio, &flags);
638 /* dbuf_read_impl has dropped db_mtx for us */
641 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
642 db->db.db_size, flags & DB_RF_CACHED);
644 if ((flags & DB_RF_HAVESTRUCT) == 0)
645 rw_exit(&dn->dn_struct_rwlock);
652 * Another reader came in while the dbuf was in flight
653 * between UNCACHED and CACHED. Either a writer will finish
654 * writing the buffer (sending the dbuf to CACHED) or the
655 * first reader's request will reach the read_done callback
656 * and send the dbuf to CACHED. Otherwise, a failure
657 * occurred and the dbuf went to UNCACHED.
659 mutex_exit(&db->db_mtx);
661 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
662 db->db.db_size, TRUE);
663 if ((flags & DB_RF_HAVESTRUCT) == 0)
664 rw_exit(&dn->dn_struct_rwlock);
667 /* Skip the wait per the caller's request. */
668 mutex_enter(&db->db_mtx);
669 if ((flags & DB_RF_NEVERWAIT) == 0) {
670 while (db->db_state == DB_READ ||
671 db->db_state == DB_FILL) {
672 ASSERT(db->db_state == DB_READ ||
673 (flags & DB_RF_HAVESTRUCT) == 0);
674 cv_wait(&db->db_changed, &db->db_mtx);
676 if (db->db_state == DB_UNCACHED)
677 err = SET_ERROR(EIO);
679 mutex_exit(&db->db_mtx);
682 ASSERT(err || havepzio || db->db_state == DB_CACHED);
687 dbuf_noread(dmu_buf_impl_t *db)
689 ASSERT(!refcount_is_zero(&db->db_holds));
690 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
691 mutex_enter(&db->db_mtx);
692 while (db->db_state == DB_READ || db->db_state == DB_FILL)
693 cv_wait(&db->db_changed, &db->db_mtx);
694 if (db->db_state == DB_UNCACHED) {
695 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
696 spa_t *spa = db->db_objset->os_spa;
698 ASSERT(db->db_buf == NULL);
699 ASSERT(db->db.db_data == NULL);
700 dbuf_set_data(db, arc_buf_alloc(spa, db->db.db_size, db, type));
701 db->db_state = DB_FILL;
702 } else if (db->db_state == DB_NOFILL) {
703 dbuf_set_data(db, NULL);
705 ASSERT3U(db->db_state, ==, DB_CACHED);
707 mutex_exit(&db->db_mtx);
711 * This is our just-in-time copy function. It makes a copy of
712 * buffers, that have been modified in a previous transaction
713 * group, before we modify them in the current active group.
715 * This function is used in two places: when we are dirtying a
716 * buffer for the first time in a txg, and when we are freeing
717 * a range in a dnode that includes this buffer.
719 * Note that when we are called from dbuf_free_range() we do
720 * not put a hold on the buffer, we just traverse the active
721 * dbuf list for the dnode.
724 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
726 dbuf_dirty_record_t *dr = db->db_last_dirty;
728 ASSERT(MUTEX_HELD(&db->db_mtx));
729 ASSERT(db->db.db_data != NULL);
730 ASSERT(db->db_level == 0);
731 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
734 (dr->dt.dl.dr_data !=
735 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
739 * If the last dirty record for this dbuf has not yet synced
740 * and its referencing the dbuf data, either:
741 * reset the reference to point to a new copy,
742 * or (if there a no active holders)
743 * just null out the current db_data pointer.
745 ASSERT(dr->dr_txg >= txg - 2);
746 if (db->db_blkid == DMU_BONUS_BLKID) {
747 /* Note that the data bufs here are zio_bufs */
748 dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
749 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
750 bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
751 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
752 int size = db->db.db_size;
753 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
754 spa_t *spa = db->db_objset->os_spa;
756 dr->dt.dl.dr_data = arc_buf_alloc(spa, size, db, type);
757 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
759 dbuf_set_data(db, NULL);
764 dbuf_unoverride(dbuf_dirty_record_t *dr)
766 dmu_buf_impl_t *db = dr->dr_dbuf;
767 blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
768 uint64_t txg = dr->dr_txg;
770 ASSERT(MUTEX_HELD(&db->db_mtx));
771 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
772 ASSERT(db->db_level == 0);
774 if (db->db_blkid == DMU_BONUS_BLKID ||
775 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
778 ASSERT(db->db_data_pending != dr);
780 /* free this block */
781 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite)
782 zio_free(db->db_objset->os_spa, txg, bp);
784 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
785 dr->dt.dl.dr_nopwrite = B_FALSE;
788 * Release the already-written buffer, so we leave it in
789 * a consistent dirty state. Note that all callers are
790 * modifying the buffer, so they will immediately do
791 * another (redundant) arc_release(). Therefore, leave
792 * the buf thawed to save the effort of freezing &
793 * immediately re-thawing it.
795 arc_release(dr->dt.dl.dr_data, db);
799 * Evict (if its unreferenced) or clear (if its referenced) any level-0
800 * data blocks in the free range, so that any future readers will find
803 * This is a no-op if the dataset is in the middle of an incremental
804 * receive; see comment below for details.
807 dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
810 dmu_buf_impl_t *db, *db_next, db_search;
811 uint64_t txg = tx->tx_txg;
814 if (end_blkid > dn->dn_maxblkid && (end_blkid != DMU_SPILL_BLKID))
815 end_blkid = dn->dn_maxblkid;
816 dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid);
818 db_search.db_level = 0;
819 db_search.db_blkid = start_blkid;
820 db_search.db_state = DB_SEARCH;
822 mutex_enter(&dn->dn_dbufs_mtx);
823 if (start_blkid >= dn->dn_unlisted_l0_blkid) {
824 /* There can't be any dbufs in this range; no need to search. */
826 db = avl_find(&dn->dn_dbufs, &db_search, &where);
827 ASSERT3P(db, ==, NULL);
828 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
829 ASSERT(db == NULL || db->db_level > 0);
831 mutex_exit(&dn->dn_dbufs_mtx);
833 } else if (dmu_objset_is_receiving(dn->dn_objset)) {
835 * If we are receiving, we expect there to be no dbufs in
836 * the range to be freed, because receive modifies each
837 * block at most once, and in offset order. If this is
838 * not the case, it can lead to performance problems,
839 * so note that we unexpectedly took the slow path.
841 atomic_inc_64(&zfs_free_range_recv_miss);
844 db = avl_find(&dn->dn_dbufs, &db_search, &where);
845 ASSERT3P(db, ==, NULL);
846 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
848 for (; db != NULL; db = db_next) {
849 db_next = AVL_NEXT(&dn->dn_dbufs, db);
850 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
852 if (db->db_level != 0 || db->db_blkid > end_blkid) {
855 ASSERT3U(db->db_blkid, >=, start_blkid);
857 /* found a level 0 buffer in the range */
858 mutex_enter(&db->db_mtx);
859 if (dbuf_undirty(db, tx)) {
860 /* mutex has been dropped and dbuf destroyed */
864 if (db->db_state == DB_UNCACHED ||
865 db->db_state == DB_NOFILL ||
866 db->db_state == DB_EVICTING) {
867 ASSERT(db->db.db_data == NULL);
868 mutex_exit(&db->db_mtx);
871 if (db->db_state == DB_READ || db->db_state == DB_FILL) {
872 /* will be handled in dbuf_read_done or dbuf_rele */
873 db->db_freed_in_flight = TRUE;
874 mutex_exit(&db->db_mtx);
877 if (refcount_count(&db->db_holds) == 0) {
882 /* The dbuf is referenced */
884 if (db->db_last_dirty != NULL) {
885 dbuf_dirty_record_t *dr = db->db_last_dirty;
887 if (dr->dr_txg == txg) {
889 * This buffer is "in-use", re-adjust the file
890 * size to reflect that this buffer may
891 * contain new data when we sync.
893 if (db->db_blkid != DMU_SPILL_BLKID &&
894 db->db_blkid > dn->dn_maxblkid)
895 dn->dn_maxblkid = db->db_blkid;
899 * This dbuf is not dirty in the open context.
900 * Either uncache it (if its not referenced in
901 * the open context) or reset its contents to
904 dbuf_fix_old_data(db, txg);
907 /* clear the contents if its cached */
908 if (db->db_state == DB_CACHED) {
909 ASSERT(db->db.db_data != NULL);
910 arc_release(db->db_buf, db);
911 bzero(db->db.db_data, db->db.db_size);
912 arc_buf_freeze(db->db_buf);
915 mutex_exit(&db->db_mtx);
917 mutex_exit(&dn->dn_dbufs_mtx);
921 dbuf_block_freeable(dmu_buf_impl_t *db)
923 dsl_dataset_t *ds = db->db_objset->os_dsl_dataset;
924 uint64_t birth_txg = 0;
927 * We don't need any locking to protect db_blkptr:
928 * If it's syncing, then db_last_dirty will be set
929 * so we'll ignore db_blkptr.
931 * This logic ensures that only block births for
932 * filled blocks are considered.
934 ASSERT(MUTEX_HELD(&db->db_mtx));
935 if (db->db_last_dirty && (db->db_blkptr == NULL ||
936 !BP_IS_HOLE(db->db_blkptr))) {
937 birth_txg = db->db_last_dirty->dr_txg;
938 } else if (db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) {
939 birth_txg = db->db_blkptr->blk_birth;
943 * If this block don't exist or is in a snapshot, it can't be freed.
944 * Don't pass the bp to dsl_dataset_block_freeable() since we
945 * are holding the db_mtx lock and might deadlock if we are
946 * prefetching a dedup-ed block.
949 return (ds == NULL ||
950 dsl_dataset_block_freeable(ds, NULL, birth_txg));
956 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
958 arc_buf_t *buf, *obuf;
959 int osize = db->db.db_size;
960 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
963 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
968 /* XXX does *this* func really need the lock? */
969 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
972 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
973 * is OK, because there can be no other references to the db
974 * when we are changing its size, so no concurrent DB_FILL can
978 * XXX we should be doing a dbuf_read, checking the return
979 * value and returning that up to our callers
981 dmu_buf_will_dirty(&db->db, tx);
983 /* create the data buffer for the new block */
984 buf = arc_buf_alloc(dn->dn_objset->os_spa, size, db, type);
986 /* copy old block data to the new block */
988 bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
989 /* zero the remainder */
991 bzero((uint8_t *)buf->b_data + osize, size - osize);
993 mutex_enter(&db->db_mtx);
994 dbuf_set_data(db, buf);
995 VERIFY(arc_buf_remove_ref(obuf, db));
996 db->db.db_size = size;
998 if (db->db_level == 0) {
999 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1000 db->db_last_dirty->dt.dl.dr_data = buf;
1002 mutex_exit(&db->db_mtx);
1004 dnode_willuse_space(dn, size-osize, tx);
1009 dbuf_release_bp(dmu_buf_impl_t *db)
1011 objset_t *os = db->db_objset;
1013 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1014 ASSERT(arc_released(os->os_phys_buf) ||
1015 list_link_active(&os->os_dsl_dataset->ds_synced_link));
1016 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1018 (void) arc_release(db->db_buf, db);
1021 dbuf_dirty_record_t *
1022 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1026 dbuf_dirty_record_t **drp, *dr;
1027 int drop_struct_lock = FALSE;
1028 boolean_t do_free_accounting = B_FALSE;
1029 int txgoff = tx->tx_txg & TXG_MASK;
1031 ASSERT(tx->tx_txg != 0);
1032 ASSERT(!refcount_is_zero(&db->db_holds));
1033 DMU_TX_DIRTY_BUF(tx, db);
1038 * Shouldn't dirty a regular buffer in syncing context. Private
1039 * objects may be dirtied in syncing context, but only if they
1040 * were already pre-dirtied in open context.
1042 ASSERT(!dmu_tx_is_syncing(tx) ||
1043 BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1044 DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1045 dn->dn_objset->os_dsl_dataset == NULL);
1047 * We make this assert for private objects as well, but after we
1048 * check if we're already dirty. They are allowed to re-dirty
1049 * in syncing context.
1051 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1052 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1053 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1055 mutex_enter(&db->db_mtx);
1057 * XXX make this true for indirects too? The problem is that
1058 * transactions created with dmu_tx_create_assigned() from
1059 * syncing context don't bother holding ahead.
1061 ASSERT(db->db_level != 0 ||
1062 db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1063 db->db_state == DB_NOFILL);
1065 mutex_enter(&dn->dn_mtx);
1067 * Don't set dirtyctx to SYNC if we're just modifying this as we
1068 * initialize the objset.
1070 if (dn->dn_dirtyctx == DN_UNDIRTIED &&
1071 !BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1073 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1074 ASSERT(dn->dn_dirtyctx_firstset == NULL);
1075 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1077 mutex_exit(&dn->dn_mtx);
1079 if (db->db_blkid == DMU_SPILL_BLKID)
1080 dn->dn_have_spill = B_TRUE;
1083 * If this buffer is already dirty, we're done.
1085 drp = &db->db_last_dirty;
1086 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1087 db->db.db_object == DMU_META_DNODE_OBJECT);
1088 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1090 if (dr && dr->dr_txg == tx->tx_txg) {
1093 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1095 * If this buffer has already been written out,
1096 * we now need to reset its state.
1098 dbuf_unoverride(dr);
1099 if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1100 db->db_state != DB_NOFILL)
1101 arc_buf_thaw(db->db_buf);
1103 mutex_exit(&db->db_mtx);
1108 * Only valid if not already dirty.
1110 ASSERT(dn->dn_object == 0 ||
1111 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1112 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1114 ASSERT3U(dn->dn_nlevels, >, db->db_level);
1115 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1116 dn->dn_phys->dn_nlevels > db->db_level ||
1117 dn->dn_next_nlevels[txgoff] > db->db_level ||
1118 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1119 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1122 * We should only be dirtying in syncing context if it's the
1123 * mos or we're initializing the os or it's a special object.
1124 * However, we are allowed to dirty in syncing context provided
1125 * we already dirtied it in open context. Hence we must make
1126 * this assertion only if we're not already dirty.
1129 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1130 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1131 ASSERT(db->db.db_size != 0);
1133 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1135 if (db->db_blkid != DMU_BONUS_BLKID) {
1137 * Update the accounting.
1138 * Note: we delay "free accounting" until after we drop
1139 * the db_mtx. This keeps us from grabbing other locks
1140 * (and possibly deadlocking) in bp_get_dsize() while
1141 * also holding the db_mtx.
1143 dnode_willuse_space(dn, db->db.db_size, tx);
1144 do_free_accounting = dbuf_block_freeable(db);
1148 * If this buffer is dirty in an old transaction group we need
1149 * to make a copy of it so that the changes we make in this
1150 * transaction group won't leak out when we sync the older txg.
1152 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1153 if (db->db_level == 0) {
1154 void *data_old = db->db_buf;
1156 if (db->db_state != DB_NOFILL) {
1157 if (db->db_blkid == DMU_BONUS_BLKID) {
1158 dbuf_fix_old_data(db, tx->tx_txg);
1159 data_old = db->db.db_data;
1160 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1162 * Release the data buffer from the cache so
1163 * that we can modify it without impacting
1164 * possible other users of this cached data
1165 * block. Note that indirect blocks and
1166 * private objects are not released until the
1167 * syncing state (since they are only modified
1170 arc_release(db->db_buf, db);
1171 dbuf_fix_old_data(db, tx->tx_txg);
1172 data_old = db->db_buf;
1174 ASSERT(data_old != NULL);
1176 dr->dt.dl.dr_data = data_old;
1178 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1179 list_create(&dr->dt.di.dr_children,
1180 sizeof (dbuf_dirty_record_t),
1181 offsetof(dbuf_dirty_record_t, dr_dirty_node));
1183 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL)
1184 dr->dr_accounted = db->db.db_size;
1186 dr->dr_txg = tx->tx_txg;
1191 * We could have been freed_in_flight between the dbuf_noread
1192 * and dbuf_dirty. We win, as though the dbuf_noread() had
1193 * happened after the free.
1195 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1196 db->db_blkid != DMU_SPILL_BLKID) {
1197 mutex_enter(&dn->dn_mtx);
1198 if (dn->dn_free_ranges[txgoff] != NULL) {
1199 range_tree_clear(dn->dn_free_ranges[txgoff],
1202 mutex_exit(&dn->dn_mtx);
1203 db->db_freed_in_flight = FALSE;
1207 * This buffer is now part of this txg
1209 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1210 db->db_dirtycnt += 1;
1211 ASSERT3U(db->db_dirtycnt, <=, 3);
1213 mutex_exit(&db->db_mtx);
1215 if (db->db_blkid == DMU_BONUS_BLKID ||
1216 db->db_blkid == DMU_SPILL_BLKID) {
1217 mutex_enter(&dn->dn_mtx);
1218 ASSERT(!list_link_active(&dr->dr_dirty_node));
1219 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1220 mutex_exit(&dn->dn_mtx);
1221 dnode_setdirty(dn, tx);
1224 } else if (do_free_accounting) {
1225 blkptr_t *bp = db->db_blkptr;
1226 int64_t willfree = (bp && !BP_IS_HOLE(bp)) ?
1227 bp_get_dsize(os->os_spa, bp) : db->db.db_size;
1229 * This is only a guess -- if the dbuf is dirty
1230 * in a previous txg, we don't know how much
1231 * space it will use on disk yet. We should
1232 * really have the struct_rwlock to access
1233 * db_blkptr, but since this is just a guess,
1234 * it's OK if we get an odd answer.
1236 ddt_prefetch(os->os_spa, bp);
1237 dnode_willuse_space(dn, -willfree, tx);
1240 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1241 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1242 drop_struct_lock = TRUE;
1245 if (db->db_level == 0) {
1246 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1247 ASSERT(dn->dn_maxblkid >= db->db_blkid);
1250 if (db->db_level+1 < dn->dn_nlevels) {
1251 dmu_buf_impl_t *parent = db->db_parent;
1252 dbuf_dirty_record_t *di;
1253 int parent_held = FALSE;
1255 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1256 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1258 parent = dbuf_hold_level(dn, db->db_level+1,
1259 db->db_blkid >> epbs, FTAG);
1260 ASSERT(parent != NULL);
1263 if (drop_struct_lock)
1264 rw_exit(&dn->dn_struct_rwlock);
1265 ASSERT3U(db->db_level+1, ==, parent->db_level);
1266 di = dbuf_dirty(parent, tx);
1268 dbuf_rele(parent, FTAG);
1270 mutex_enter(&db->db_mtx);
1272 * Since we've dropped the mutex, it's possible that
1273 * dbuf_undirty() might have changed this out from under us.
1275 if (db->db_last_dirty == dr ||
1276 dn->dn_object == DMU_META_DNODE_OBJECT) {
1277 mutex_enter(&di->dt.di.dr_mtx);
1278 ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1279 ASSERT(!list_link_active(&dr->dr_dirty_node));
1280 list_insert_tail(&di->dt.di.dr_children, dr);
1281 mutex_exit(&di->dt.di.dr_mtx);
1284 mutex_exit(&db->db_mtx);
1286 ASSERT(db->db_level+1 == dn->dn_nlevels);
1287 ASSERT(db->db_blkid < dn->dn_nblkptr);
1288 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1289 mutex_enter(&dn->dn_mtx);
1290 ASSERT(!list_link_active(&dr->dr_dirty_node));
1291 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1292 mutex_exit(&dn->dn_mtx);
1293 if (drop_struct_lock)
1294 rw_exit(&dn->dn_struct_rwlock);
1297 dnode_setdirty(dn, tx);
1303 * Undirty a buffer in the transaction group referenced by the given
1304 * transaction. Return whether this evicted the dbuf.
1307 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1310 uint64_t txg = tx->tx_txg;
1311 dbuf_dirty_record_t *dr, **drp;
1314 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1315 ASSERT0(db->db_level);
1316 ASSERT(MUTEX_HELD(&db->db_mtx));
1319 * If this buffer is not dirty, we're done.
1321 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1322 if (dr->dr_txg <= txg)
1324 if (dr == NULL || dr->dr_txg < txg)
1326 ASSERT(dr->dr_txg == txg);
1327 ASSERT(dr->dr_dbuf == db);
1332 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1334 ASSERT(db->db.db_size != 0);
1337 * Any space we accounted for in dp_dirty_* will be cleaned up by
1338 * dsl_pool_sync(). This is relatively rare so the discrepancy
1339 * is not a big deal.
1345 * Note that there are three places in dbuf_dirty()
1346 * where this dirty record may be put on a list.
1347 * Make sure to do a list_remove corresponding to
1348 * every one of those list_insert calls.
1350 if (dr->dr_parent) {
1351 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1352 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1353 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1354 } else if (db->db_blkid == DMU_SPILL_BLKID ||
1355 db->db_level+1 == dn->dn_nlevels) {
1356 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1357 mutex_enter(&dn->dn_mtx);
1358 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1359 mutex_exit(&dn->dn_mtx);
1363 if (db->db_state != DB_NOFILL) {
1364 dbuf_unoverride(dr);
1366 ASSERT(db->db_buf != NULL);
1367 ASSERT(dr->dt.dl.dr_data != NULL);
1368 if (dr->dt.dl.dr_data != db->db_buf)
1369 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db));
1372 if (db->db_level != 0) {
1373 mutex_destroy(&dr->dt.di.dr_mtx);
1374 list_destroy(&dr->dt.di.dr_children);
1377 kmem_free(dr, sizeof (dbuf_dirty_record_t));
1379 ASSERT(db->db_dirtycnt > 0);
1380 db->db_dirtycnt -= 1;
1382 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1383 arc_buf_t *buf = db->db_buf;
1385 ASSERT(db->db_state == DB_NOFILL || arc_released(buf));
1386 dbuf_set_data(db, NULL);
1387 VERIFY(arc_buf_remove_ref(buf, db));
1396 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
1398 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1399 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1401 ASSERT(tx->tx_txg != 0);
1402 ASSERT(!refcount_is_zero(&db->db_holds));
1405 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
1406 rf |= DB_RF_HAVESTRUCT;
1408 (void) dbuf_read(db, NULL, rf);
1409 (void) dbuf_dirty(db, tx);
1413 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1415 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1417 db->db_state = DB_NOFILL;
1419 dmu_buf_will_fill(db_fake, tx);
1423 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1425 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1427 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1428 ASSERT(tx->tx_txg != 0);
1429 ASSERT(db->db_level == 0);
1430 ASSERT(!refcount_is_zero(&db->db_holds));
1432 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1433 dmu_tx_private_ok(tx));
1436 (void) dbuf_dirty(db, tx);
1439 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1442 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1444 mutex_enter(&db->db_mtx);
1447 if (db->db_state == DB_FILL) {
1448 if (db->db_level == 0 && db->db_freed_in_flight) {
1449 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1450 /* we were freed while filling */
1451 /* XXX dbuf_undirty? */
1452 bzero(db->db.db_data, db->db.db_size);
1453 db->db_freed_in_flight = FALSE;
1455 db->db_state = DB_CACHED;
1456 cv_broadcast(&db->db_changed);
1458 mutex_exit(&db->db_mtx);
1462 dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data,
1463 bp_embedded_type_t etype, enum zio_compress comp,
1464 int uncompressed_size, int compressed_size, int byteorder,
1467 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
1468 struct dirty_leaf *dl;
1469 dmu_object_type_t type;
1472 type = DB_DNODE(db)->dn_type;
1475 ASSERT0(db->db_level);
1476 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1478 dmu_buf_will_not_fill(dbuf, tx);
1480 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1481 dl = &db->db_last_dirty->dt.dl;
1482 encode_embedded_bp_compressed(&dl->dr_overridden_by,
1483 data, comp, uncompressed_size, compressed_size);
1484 BPE_SET_ETYPE(&dl->dr_overridden_by, etype);
1485 BP_SET_TYPE(&dl->dr_overridden_by, type);
1486 BP_SET_LEVEL(&dl->dr_overridden_by, 0);
1487 BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder);
1489 dl->dr_override_state = DR_OVERRIDDEN;
1490 dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg;
1494 * Directly assign a provided arc buf to a given dbuf if it's not referenced
1495 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1498 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
1500 ASSERT(!refcount_is_zero(&db->db_holds));
1501 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1502 ASSERT(db->db_level == 0);
1503 ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA);
1504 ASSERT(buf != NULL);
1505 ASSERT(arc_buf_size(buf) == db->db.db_size);
1506 ASSERT(tx->tx_txg != 0);
1508 arc_return_buf(buf, db);
1509 ASSERT(arc_released(buf));
1511 mutex_enter(&db->db_mtx);
1513 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1514 cv_wait(&db->db_changed, &db->db_mtx);
1516 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
1518 if (db->db_state == DB_CACHED &&
1519 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
1520 mutex_exit(&db->db_mtx);
1521 (void) dbuf_dirty(db, tx);
1522 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
1523 VERIFY(arc_buf_remove_ref(buf, db));
1524 xuio_stat_wbuf_copied();
1528 xuio_stat_wbuf_nocopy();
1529 if (db->db_state == DB_CACHED) {
1530 dbuf_dirty_record_t *dr = db->db_last_dirty;
1532 ASSERT(db->db_buf != NULL);
1533 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
1534 ASSERT(dr->dt.dl.dr_data == db->db_buf);
1535 if (!arc_released(db->db_buf)) {
1536 ASSERT(dr->dt.dl.dr_override_state ==
1538 arc_release(db->db_buf, db);
1540 dr->dt.dl.dr_data = buf;
1541 VERIFY(arc_buf_remove_ref(db->db_buf, db));
1542 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
1543 arc_release(db->db_buf, db);
1544 VERIFY(arc_buf_remove_ref(db->db_buf, db));
1548 ASSERT(db->db_buf == NULL);
1549 dbuf_set_data(db, buf);
1550 db->db_state = DB_FILL;
1551 mutex_exit(&db->db_mtx);
1552 (void) dbuf_dirty(db, tx);
1553 dmu_buf_fill_done(&db->db, tx);
1557 * "Clear" the contents of this dbuf. This will mark the dbuf
1558 * EVICTING and clear *most* of its references. Unfortunately,
1559 * when we are not holding the dn_dbufs_mtx, we can't clear the
1560 * entry in the dn_dbufs list. We have to wait until dbuf_destroy()
1561 * in this case. For callers from the DMU we will usually see:
1562 * dbuf_clear()->arc_clear_callback()->dbuf_do_evict()->dbuf_destroy()
1563 * For the arc callback, we will usually see:
1564 * dbuf_do_evict()->dbuf_clear();dbuf_destroy()
1565 * Sometimes, though, we will get a mix of these two:
1566 * DMU: dbuf_clear()->arc_clear_callback()
1567 * ARC: dbuf_do_evict()->dbuf_destroy()
1569 * This routine will dissociate the dbuf from the arc, by calling
1570 * arc_clear_callback(), but will not evict the data from the ARC.
1573 dbuf_clear(dmu_buf_impl_t *db)
1576 dmu_buf_impl_t *parent = db->db_parent;
1577 dmu_buf_impl_t *dndb;
1578 boolean_t dbuf_gone = B_FALSE;
1580 ASSERT(MUTEX_HELD(&db->db_mtx));
1581 ASSERT(refcount_is_zero(&db->db_holds));
1583 dbuf_evict_user(db);
1585 if (db->db_state == DB_CACHED) {
1586 ASSERT(db->db.db_data != NULL);
1587 if (db->db_blkid == DMU_BONUS_BLKID) {
1588 zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
1589 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
1591 db->db.db_data = NULL;
1592 db->db_state = DB_UNCACHED;
1595 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
1596 ASSERT(db->db_data_pending == NULL);
1598 db->db_state = DB_EVICTING;
1599 db->db_blkptr = NULL;
1604 if (db->db_blkid != DMU_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) {
1605 avl_remove(&dn->dn_dbufs, db);
1606 (void) atomic_dec_32_nv(&dn->dn_dbufs_count);
1610 * Decrementing the dbuf count means that the hold corresponding
1611 * to the removed dbuf is no longer discounted in dnode_move(),
1612 * so the dnode cannot be moved until after we release the hold.
1613 * The membar_producer() ensures visibility of the decremented
1614 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
1618 db->db_dnode_handle = NULL;
1624 dbuf_gone = arc_clear_callback(db->db_buf);
1627 mutex_exit(&db->db_mtx);
1630 * If this dbuf is referenced from an indirect dbuf,
1631 * decrement the ref count on the indirect dbuf.
1633 if (parent && parent != dndb)
1634 dbuf_rele(parent, db);
1638 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
1639 dmu_buf_impl_t **parentp, blkptr_t **bpp)
1646 ASSERT(blkid != DMU_BONUS_BLKID);
1648 if (blkid == DMU_SPILL_BLKID) {
1649 mutex_enter(&dn->dn_mtx);
1650 if (dn->dn_have_spill &&
1651 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
1652 *bpp = &dn->dn_phys->dn_spill;
1655 dbuf_add_ref(dn->dn_dbuf, NULL);
1656 *parentp = dn->dn_dbuf;
1657 mutex_exit(&dn->dn_mtx);
1661 if (dn->dn_phys->dn_nlevels == 0)
1664 nlevels = dn->dn_phys->dn_nlevels;
1666 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1668 ASSERT3U(level * epbs, <, 64);
1669 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1670 if (level >= nlevels ||
1671 (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
1672 /* the buffer has no parent yet */
1673 return (SET_ERROR(ENOENT));
1674 } else if (level < nlevels-1) {
1675 /* this block is referenced from an indirect block */
1676 int err = dbuf_hold_impl(dn, level+1,
1677 blkid >> epbs, fail_sparse, NULL, parentp);
1680 err = dbuf_read(*parentp, NULL,
1681 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
1683 dbuf_rele(*parentp, NULL);
1687 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
1688 (blkid & ((1ULL << epbs) - 1));
1691 /* the block is referenced from the dnode */
1692 ASSERT3U(level, ==, nlevels-1);
1693 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
1694 blkid < dn->dn_phys->dn_nblkptr);
1696 dbuf_add_ref(dn->dn_dbuf, NULL);
1697 *parentp = dn->dn_dbuf;
1699 *bpp = &dn->dn_phys->dn_blkptr[blkid];
1704 static dmu_buf_impl_t *
1705 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
1706 dmu_buf_impl_t *parent, blkptr_t *blkptr)
1708 objset_t *os = dn->dn_objset;
1709 dmu_buf_impl_t *db, *odb;
1711 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1712 ASSERT(dn->dn_type != DMU_OT_NONE);
1714 db = kmem_cache_alloc(dbuf_cache, KM_SLEEP);
1717 db->db.db_object = dn->dn_object;
1718 db->db_level = level;
1719 db->db_blkid = blkid;
1720 db->db_last_dirty = NULL;
1721 db->db_dirtycnt = 0;
1722 db->db_dnode_handle = dn->dn_handle;
1723 db->db_parent = parent;
1724 db->db_blkptr = blkptr;
1726 db->db_user_ptr = NULL;
1727 db->db_user_data_ptr_ptr = NULL;
1728 db->db_evict_func = NULL;
1729 db->db_immediate_evict = 0;
1730 db->db_freed_in_flight = 0;
1732 if (blkid == DMU_BONUS_BLKID) {
1733 ASSERT3P(parent, ==, dn->dn_dbuf);
1734 db->db.db_size = DN_MAX_BONUSLEN -
1735 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
1736 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
1737 db->db.db_offset = DMU_BONUS_BLKID;
1738 db->db_state = DB_UNCACHED;
1739 /* the bonus dbuf is not placed in the hash table */
1740 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1742 } else if (blkid == DMU_SPILL_BLKID) {
1743 db->db.db_size = (blkptr != NULL) ?
1744 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
1745 db->db.db_offset = 0;
1748 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
1749 db->db.db_size = blocksize;
1750 db->db.db_offset = db->db_blkid * blocksize;
1754 * Hold the dn_dbufs_mtx while we get the new dbuf
1755 * in the hash table *and* added to the dbufs list.
1756 * This prevents a possible deadlock with someone
1757 * trying to look up this dbuf before its added to the
1760 mutex_enter(&dn->dn_dbufs_mtx);
1761 db->db_state = DB_EVICTING;
1762 if ((odb = dbuf_hash_insert(db)) != NULL) {
1763 /* someone else inserted it first */
1764 kmem_cache_free(dbuf_cache, db);
1765 mutex_exit(&dn->dn_dbufs_mtx);
1768 avl_add(&dn->dn_dbufs, db);
1769 if (db->db_level == 0 && db->db_blkid >=
1770 dn->dn_unlisted_l0_blkid)
1771 dn->dn_unlisted_l0_blkid = db->db_blkid + 1;
1772 db->db_state = DB_UNCACHED;
1773 mutex_exit(&dn->dn_dbufs_mtx);
1774 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1776 if (parent && parent != dn->dn_dbuf)
1777 dbuf_add_ref(parent, db);
1779 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1780 refcount_count(&dn->dn_holds) > 0);
1781 (void) refcount_add(&dn->dn_holds, db);
1782 (void) atomic_inc_32_nv(&dn->dn_dbufs_count);
1784 dprintf_dbuf(db, "db=%p\n", db);
1790 dbuf_do_evict(void *private)
1792 dmu_buf_impl_t *db = private;
1794 if (!MUTEX_HELD(&db->db_mtx))
1795 mutex_enter(&db->db_mtx);
1797 ASSERT(refcount_is_zero(&db->db_holds));
1799 if (db->db_state != DB_EVICTING) {
1800 ASSERT(db->db_state == DB_CACHED);
1805 mutex_exit(&db->db_mtx);
1812 dbuf_destroy(dmu_buf_impl_t *db)
1814 ASSERT(refcount_is_zero(&db->db_holds));
1816 if (db->db_blkid != DMU_BONUS_BLKID) {
1818 * If this dbuf is still on the dn_dbufs list,
1819 * remove it from that list.
1821 if (db->db_dnode_handle != NULL) {
1826 mutex_enter(&dn->dn_dbufs_mtx);
1827 avl_remove(&dn->dn_dbufs, db);
1828 (void) atomic_dec_32_nv(&dn->dn_dbufs_count);
1829 mutex_exit(&dn->dn_dbufs_mtx);
1832 * Decrementing the dbuf count means that the hold
1833 * corresponding to the removed dbuf is no longer
1834 * discounted in dnode_move(), so the dnode cannot be
1835 * moved until after we release the hold.
1838 db->db_dnode_handle = NULL;
1840 dbuf_hash_remove(db);
1842 db->db_parent = NULL;
1845 ASSERT(db->db.db_data == NULL);
1846 ASSERT(db->db_hash_next == NULL);
1847 ASSERT(db->db_blkptr == NULL);
1848 ASSERT(db->db_data_pending == NULL);
1850 kmem_cache_free(dbuf_cache, db);
1851 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1855 dbuf_prefetch(dnode_t *dn, uint64_t blkid, zio_priority_t prio)
1857 dmu_buf_impl_t *db = NULL;
1858 blkptr_t *bp = NULL;
1860 ASSERT(blkid != DMU_BONUS_BLKID);
1861 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1863 if (dnode_block_freed(dn, blkid))
1866 /* dbuf_find() returns with db_mtx held */
1867 if (db = dbuf_find(dn, 0, blkid)) {
1869 * This dbuf is already in the cache. We assume that
1870 * it is already CACHED, or else about to be either
1873 mutex_exit(&db->db_mtx);
1877 if (dbuf_findbp(dn, 0, blkid, TRUE, &db, &bp) == 0) {
1878 if (bp && !BP_IS_HOLE(bp) && !BP_IS_EMBEDDED(bp)) {
1879 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
1880 uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH;
1881 zbookmark_phys_t zb;
1883 SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
1884 dn->dn_object, 0, blkid);
1886 (void) arc_read(NULL, dn->dn_objset->os_spa,
1887 bp, NULL, NULL, prio,
1888 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1892 dbuf_rele(db, NULL);
1897 * Returns with db_holds incremented, and db_mtx not held.
1898 * Note: dn_struct_rwlock must be held.
1901 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse,
1902 void *tag, dmu_buf_impl_t **dbp)
1904 dmu_buf_impl_t *db, *parent = NULL;
1906 ASSERT(blkid != DMU_BONUS_BLKID);
1907 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1908 ASSERT3U(dn->dn_nlevels, >, level);
1912 /* dbuf_find() returns with db_mtx held */
1913 db = dbuf_find(dn, level, blkid);
1916 blkptr_t *bp = NULL;
1919 ASSERT3P(parent, ==, NULL);
1920 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
1922 if (err == 0 && bp && BP_IS_HOLE(bp))
1923 err = SET_ERROR(ENOENT);
1926 dbuf_rele(parent, NULL);
1930 if (err && err != ENOENT)
1932 db = dbuf_create(dn, level, blkid, parent, bp);
1935 if (db->db_buf && refcount_is_zero(&db->db_holds)) {
1936 arc_buf_add_ref(db->db_buf, db);
1937 if (db->db_buf->b_data == NULL) {
1940 dbuf_rele(parent, NULL);
1945 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
1948 ASSERT(db->db_buf == NULL || arc_referenced(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 (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1956 dn->dn_object != DMU_META_DNODE_OBJECT &&
1957 db->db_state == DB_CACHED && db->db_data_pending) {
1958 dbuf_dirty_record_t *dr = db->db_data_pending;
1960 if (dr->dt.dl.dr_data == db->db_buf) {
1961 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1964 arc_buf_alloc(dn->dn_objset->os_spa,
1965 db->db.db_size, db, type));
1966 bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
1971 (void) refcount_add(&db->db_holds, tag);
1972 dbuf_update_data(db);
1974 mutex_exit(&db->db_mtx);
1976 /* NOTE: we can't rele the parent until after we drop the db_mtx */
1978 dbuf_rele(parent, NULL);
1980 ASSERT3P(DB_DNODE(db), ==, dn);
1981 ASSERT3U(db->db_blkid, ==, blkid);
1982 ASSERT3U(db->db_level, ==, level);
1989 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
1992 int err = dbuf_hold_impl(dn, 0, blkid, FALSE, tag, &db);
1993 return (err ? NULL : db);
1997 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
2000 int err = dbuf_hold_impl(dn, level, blkid, FALSE, tag, &db);
2001 return (err ? NULL : db);
2005 dbuf_create_bonus(dnode_t *dn)
2007 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
2009 ASSERT(dn->dn_bonus == NULL);
2010 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
2014 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
2016 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2019 if (db->db_blkid != DMU_SPILL_BLKID)
2020 return (SET_ERROR(ENOTSUP));
2022 blksz = SPA_MINBLOCKSIZE;
2023 if (blksz > SPA_MAXBLOCKSIZE)
2024 blksz = SPA_MAXBLOCKSIZE;
2026 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
2030 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2031 dbuf_new_size(db, blksz, tx);
2032 rw_exit(&dn->dn_struct_rwlock);
2039 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
2041 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
2044 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2046 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
2048 int64_t holds = refcount_add(&db->db_holds, tag);
2053 * If you call dbuf_rele() you had better not be referencing the dnode handle
2054 * unless you have some other direct or indirect hold on the dnode. (An indirect
2055 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2056 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2057 * dnode's parent dbuf evicting its dnode handles.
2060 dbuf_rele(dmu_buf_impl_t *db, void *tag)
2062 mutex_enter(&db->db_mtx);
2063 dbuf_rele_and_unlock(db, tag);
2067 dmu_buf_rele(dmu_buf_t *db, void *tag)
2069 dbuf_rele((dmu_buf_impl_t *)db, tag);
2073 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2074 * db_dirtycnt and db_holds to be updated atomically.
2077 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2081 ASSERT(MUTEX_HELD(&db->db_mtx));
2085 * Remove the reference to the dbuf before removing its hold on the
2086 * dnode so we can guarantee in dnode_move() that a referenced bonus
2087 * buffer has a corresponding dnode hold.
2089 holds = refcount_remove(&db->db_holds, tag);
2093 * We can't freeze indirects if there is a possibility that they
2094 * may be modified in the current syncing context.
2096 if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0))
2097 arc_buf_freeze(db->db_buf);
2099 if (holds == db->db_dirtycnt &&
2100 db->db_level == 0 && db->db_immediate_evict)
2101 dbuf_evict_user(db);
2104 if (db->db_blkid == DMU_BONUS_BLKID) {
2105 mutex_exit(&db->db_mtx);
2108 * If the dnode moves here, we cannot cross this barrier
2109 * until the move completes.
2112 (void) atomic_dec_32_nv(&DB_DNODE(db)->dn_dbufs_count);
2115 * The bonus buffer's dnode hold is no longer discounted
2116 * in dnode_move(). The dnode cannot move until after
2119 dnode_rele(DB_DNODE(db), db);
2120 } else if (db->db_buf == NULL) {
2122 * This is a special case: we never associated this
2123 * dbuf with any data allocated from the ARC.
2125 ASSERT(db->db_state == DB_UNCACHED ||
2126 db->db_state == DB_NOFILL);
2128 } else if (arc_released(db->db_buf)) {
2129 arc_buf_t *buf = db->db_buf;
2131 * This dbuf has anonymous data associated with it.
2133 dbuf_set_data(db, NULL);
2134 VERIFY(arc_buf_remove_ref(buf, db));
2137 VERIFY(!arc_buf_remove_ref(db->db_buf, db));
2140 * A dbuf will be eligible for eviction if either the
2141 * 'primarycache' property is set or a duplicate
2142 * copy of this buffer is already cached in the arc.
2144 * In the case of the 'primarycache' a buffer
2145 * is considered for eviction if it matches the
2146 * criteria set in the property.
2148 * To decide if our buffer is considered a
2149 * duplicate, we must call into the arc to determine
2150 * if multiple buffers are referencing the same
2151 * block on-disk. If so, then we simply evict
2154 if (!DBUF_IS_CACHEABLE(db)) {
2155 if (db->db_blkptr != NULL &&
2156 !BP_IS_HOLE(db->db_blkptr) &&
2157 !BP_IS_EMBEDDED(db->db_blkptr)) {
2159 dmu_objset_spa(db->db_objset);
2160 blkptr_t bp = *db->db_blkptr;
2162 arc_freed(spa, &bp);
2166 } else if (arc_buf_eviction_needed(db->db_buf)) {
2169 mutex_exit(&db->db_mtx);
2173 mutex_exit(&db->db_mtx);
2177 #pragma weak dmu_buf_refcount = dbuf_refcount
2179 dbuf_refcount(dmu_buf_impl_t *db)
2181 return (refcount_count(&db->db_holds));
2185 dmu_buf_set_user(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr,
2186 dmu_buf_evict_func_t *evict_func)
2188 return (dmu_buf_update_user(db_fake, NULL, user_ptr,
2189 user_data_ptr_ptr, evict_func));
2193 dmu_buf_set_user_ie(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr,
2194 dmu_buf_evict_func_t *evict_func)
2196 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2198 db->db_immediate_evict = TRUE;
2199 return (dmu_buf_update_user(db_fake, NULL, user_ptr,
2200 user_data_ptr_ptr, evict_func));
2204 dmu_buf_update_user(dmu_buf_t *db_fake, void *old_user_ptr, void *user_ptr,
2205 void *user_data_ptr_ptr, dmu_buf_evict_func_t *evict_func)
2207 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2208 ASSERT(db->db_level == 0);
2210 ASSERT((user_ptr == NULL) == (evict_func == NULL));
2212 mutex_enter(&db->db_mtx);
2214 if (db->db_user_ptr == old_user_ptr) {
2215 db->db_user_ptr = user_ptr;
2216 db->db_user_data_ptr_ptr = user_data_ptr_ptr;
2217 db->db_evict_func = evict_func;
2219 dbuf_update_data(db);
2221 old_user_ptr = db->db_user_ptr;
2224 mutex_exit(&db->db_mtx);
2225 return (old_user_ptr);
2229 dmu_buf_get_user(dmu_buf_t *db_fake)
2231 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2232 ASSERT(!refcount_is_zero(&db->db_holds));
2234 return (db->db_user_ptr);
2238 dmu_buf_freeable(dmu_buf_t *dbuf)
2240 boolean_t res = B_FALSE;
2241 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2244 res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset,
2245 db->db_blkptr, db->db_blkptr->blk_birth);
2251 dmu_buf_get_blkptr(dmu_buf_t *db)
2253 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2254 return (dbi->db_blkptr);
2258 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
2260 /* ASSERT(dmu_tx_is_syncing(tx) */
2261 ASSERT(MUTEX_HELD(&db->db_mtx));
2263 if (db->db_blkptr != NULL)
2266 if (db->db_blkid == DMU_SPILL_BLKID) {
2267 db->db_blkptr = &dn->dn_phys->dn_spill;
2268 BP_ZERO(db->db_blkptr);
2271 if (db->db_level == dn->dn_phys->dn_nlevels-1) {
2273 * This buffer was allocated at a time when there was
2274 * no available blkptrs from the dnode, or it was
2275 * inappropriate to hook it in (i.e., nlevels mis-match).
2277 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
2278 ASSERT(db->db_parent == NULL);
2279 db->db_parent = dn->dn_dbuf;
2280 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
2283 dmu_buf_impl_t *parent = db->db_parent;
2284 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2286 ASSERT(dn->dn_phys->dn_nlevels > 1);
2287 if (parent == NULL) {
2288 mutex_exit(&db->db_mtx);
2289 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2290 (void) dbuf_hold_impl(dn, db->db_level+1,
2291 db->db_blkid >> epbs, FALSE, db, &parent);
2292 rw_exit(&dn->dn_struct_rwlock);
2293 mutex_enter(&db->db_mtx);
2294 db->db_parent = parent;
2296 db->db_blkptr = (blkptr_t *)parent->db.db_data +
2297 (db->db_blkid & ((1ULL << epbs) - 1));
2303 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2305 dmu_buf_impl_t *db = dr->dr_dbuf;
2309 ASSERT(dmu_tx_is_syncing(tx));
2311 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2313 mutex_enter(&db->db_mtx);
2315 ASSERT(db->db_level > 0);
2318 /* Read the block if it hasn't been read yet. */
2319 if (db->db_buf == NULL) {
2320 mutex_exit(&db->db_mtx);
2321 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
2322 mutex_enter(&db->db_mtx);
2324 ASSERT3U(db->db_state, ==, DB_CACHED);
2325 ASSERT(db->db_buf != NULL);
2329 /* Indirect block size must match what the dnode thinks it is. */
2330 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2331 dbuf_check_blkptr(dn, db);
2334 /* Provide the pending dirty record to child dbufs */
2335 db->db_data_pending = dr;
2337 mutex_exit(&db->db_mtx);
2338 dbuf_write(dr, db->db_buf, tx);
2341 mutex_enter(&dr->dt.di.dr_mtx);
2342 dbuf_sync_list(&dr->dt.di.dr_children, tx);
2343 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2344 mutex_exit(&dr->dt.di.dr_mtx);
2349 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2351 arc_buf_t **datap = &dr->dt.dl.dr_data;
2352 dmu_buf_impl_t *db = dr->dr_dbuf;
2355 uint64_t txg = tx->tx_txg;
2357 ASSERT(dmu_tx_is_syncing(tx));
2359 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2361 mutex_enter(&db->db_mtx);
2363 * To be synced, we must be dirtied. But we
2364 * might have been freed after the dirty.
2366 if (db->db_state == DB_UNCACHED) {
2367 /* This buffer has been freed since it was dirtied */
2368 ASSERT(db->db.db_data == NULL);
2369 } else if (db->db_state == DB_FILL) {
2370 /* This buffer was freed and is now being re-filled */
2371 ASSERT(db->db.db_data != dr->dt.dl.dr_data);
2373 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
2380 if (db->db_blkid == DMU_SPILL_BLKID) {
2381 mutex_enter(&dn->dn_mtx);
2382 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
2383 mutex_exit(&dn->dn_mtx);
2387 * If this is a bonus buffer, simply copy the bonus data into the
2388 * dnode. It will be written out when the dnode is synced (and it
2389 * will be synced, since it must have been dirty for dbuf_sync to
2392 if (db->db_blkid == DMU_BONUS_BLKID) {
2393 dbuf_dirty_record_t **drp;
2395 ASSERT(*datap != NULL);
2396 ASSERT0(db->db_level);
2397 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN);
2398 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen);
2401 if (*datap != db->db.db_data) {
2402 zio_buf_free(*datap, DN_MAX_BONUSLEN);
2403 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
2405 db->db_data_pending = NULL;
2406 drp = &db->db_last_dirty;
2408 drp = &(*drp)->dr_next;
2409 ASSERT(dr->dr_next == NULL);
2410 ASSERT(dr->dr_dbuf == db);
2412 if (dr->dr_dbuf->db_level != 0) {
2413 list_destroy(&dr->dt.di.dr_children);
2414 mutex_destroy(&dr->dt.di.dr_mtx);
2416 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2417 ASSERT(db->db_dirtycnt > 0);
2418 db->db_dirtycnt -= 1;
2419 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
2426 * This function may have dropped the db_mtx lock allowing a dmu_sync
2427 * operation to sneak in. As a result, we need to ensure that we
2428 * don't check the dr_override_state until we have returned from
2429 * dbuf_check_blkptr.
2431 dbuf_check_blkptr(dn, db);
2434 * If this buffer is in the middle of an immediate write,
2435 * wait for the synchronous IO to complete.
2437 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
2438 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
2439 cv_wait(&db->db_changed, &db->db_mtx);
2440 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
2443 if (db->db_state != DB_NOFILL &&
2444 dn->dn_object != DMU_META_DNODE_OBJECT &&
2445 refcount_count(&db->db_holds) > 1 &&
2446 dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
2447 *datap == db->db_buf) {
2449 * If this buffer is currently "in use" (i.e., there
2450 * are active holds and db_data still references it),
2451 * then make a copy before we start the write so that
2452 * any modifications from the open txg will not leak
2455 * NOTE: this copy does not need to be made for
2456 * objects only modified in the syncing context (e.g.
2457 * DNONE_DNODE blocks).
2459 int blksz = arc_buf_size(*datap);
2460 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2461 *datap = arc_buf_alloc(os->os_spa, blksz, db, type);
2462 bcopy(db->db.db_data, (*datap)->b_data, blksz);
2464 db->db_data_pending = dr;
2466 mutex_exit(&db->db_mtx);
2468 dbuf_write(dr, *datap, tx);
2470 ASSERT(!list_link_active(&dr->dr_dirty_node));
2471 if (dn->dn_object == DMU_META_DNODE_OBJECT) {
2472 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
2476 * Although zio_nowait() does not "wait for an IO", it does
2477 * initiate the IO. If this is an empty write it seems plausible
2478 * that the IO could actually be completed before the nowait
2479 * returns. We need to DB_DNODE_EXIT() first in case
2480 * zio_nowait() invalidates the dbuf.
2483 zio_nowait(dr->dr_zio);
2488 dbuf_sync_list(list_t *list, dmu_tx_t *tx)
2490 dbuf_dirty_record_t *dr;
2492 while (dr = list_head(list)) {
2493 if (dr->dr_zio != NULL) {
2495 * If we find an already initialized zio then we
2496 * are processing the meta-dnode, and we have finished.
2497 * The dbufs for all dnodes are put back on the list
2498 * during processing, so that we can zio_wait()
2499 * these IOs after initiating all child IOs.
2501 ASSERT3U(dr->dr_dbuf->db.db_object, ==,
2502 DMU_META_DNODE_OBJECT);
2505 list_remove(list, dr);
2506 if (dr->dr_dbuf->db_level > 0)
2507 dbuf_sync_indirect(dr, tx);
2509 dbuf_sync_leaf(dr, tx);
2515 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
2517 dmu_buf_impl_t *db = vdb;
2519 blkptr_t *bp = zio->io_bp;
2520 blkptr_t *bp_orig = &zio->io_bp_orig;
2521 spa_t *spa = zio->io_spa;
2526 ASSERT3P(db->db_blkptr, ==, bp);
2530 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
2531 dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
2532 zio->io_prev_space_delta = delta;
2534 if (bp->blk_birth != 0) {
2535 ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
2536 BP_GET_TYPE(bp) == dn->dn_type) ||
2537 (db->db_blkid == DMU_SPILL_BLKID &&
2538 BP_GET_TYPE(bp) == dn->dn_bonustype) ||
2539 BP_IS_EMBEDDED(bp));
2540 ASSERT(BP_GET_LEVEL(bp) == db->db_level);
2543 mutex_enter(&db->db_mtx);
2546 if (db->db_blkid == DMU_SPILL_BLKID) {
2547 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2548 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2549 db->db_blkptr == &dn->dn_phys->dn_spill);
2553 if (db->db_level == 0) {
2554 mutex_enter(&dn->dn_mtx);
2555 if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
2556 db->db_blkid != DMU_SPILL_BLKID)
2557 dn->dn_phys->dn_maxblkid = db->db_blkid;
2558 mutex_exit(&dn->dn_mtx);
2560 if (dn->dn_type == DMU_OT_DNODE) {
2561 dnode_phys_t *dnp = db->db.db_data;
2562 for (i = db->db.db_size >> DNODE_SHIFT; i > 0;
2564 if (dnp->dn_type != DMU_OT_NONE)
2568 if (BP_IS_HOLE(bp)) {
2575 blkptr_t *ibp = db->db.db_data;
2576 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2577 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
2578 if (BP_IS_HOLE(ibp))
2580 fill += BP_GET_FILL(ibp);
2585 if (!BP_IS_EMBEDDED(bp))
2586 bp->blk_fill = fill;
2588 mutex_exit(&db->db_mtx);
2592 * The SPA will call this callback several times for each zio - once
2593 * for every physical child i/o (zio->io_phys_children times). This
2594 * allows the DMU to monitor the progress of each logical i/o. For example,
2595 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
2596 * block. There may be a long delay before all copies/fragments are completed,
2597 * so this callback allows us to retire dirty space gradually, as the physical
2602 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
2604 dmu_buf_impl_t *db = arg;
2605 objset_t *os = db->db_objset;
2606 dsl_pool_t *dp = dmu_objset_pool(os);
2607 dbuf_dirty_record_t *dr;
2610 dr = db->db_data_pending;
2611 ASSERT3U(dr->dr_txg, ==, zio->io_txg);
2614 * The callback will be called io_phys_children times. Retire one
2615 * portion of our dirty space each time we are called. Any rounding
2616 * error will be cleaned up by dsl_pool_sync()'s call to
2617 * dsl_pool_undirty_space().
2619 delta = dr->dr_accounted / zio->io_phys_children;
2620 dsl_pool_undirty_space(dp, delta, zio->io_txg);
2625 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
2627 dmu_buf_impl_t *db = vdb;
2628 blkptr_t *bp_orig = &zio->io_bp_orig;
2629 blkptr_t *bp = db->db_blkptr;
2630 objset_t *os = db->db_objset;
2631 dmu_tx_t *tx = os->os_synctx;
2632 dbuf_dirty_record_t **drp, *dr;
2634 ASSERT0(zio->io_error);
2635 ASSERT(db->db_blkptr == bp);
2638 * For nopwrites and rewrites we ensure that the bp matches our
2639 * original and bypass all the accounting.
2641 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
2642 ASSERT(BP_EQUAL(bp, bp_orig));
2644 dsl_dataset_t *ds = os->os_dsl_dataset;
2645 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
2646 dsl_dataset_block_born(ds, bp, tx);
2649 mutex_enter(&db->db_mtx);
2653 drp = &db->db_last_dirty;
2654 while ((dr = *drp) != db->db_data_pending)
2656 ASSERT(!list_link_active(&dr->dr_dirty_node));
2657 ASSERT(dr->dr_dbuf == db);
2658 ASSERT(dr->dr_next == NULL);
2662 if (db->db_blkid == DMU_SPILL_BLKID) {
2667 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2668 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2669 db->db_blkptr == &dn->dn_phys->dn_spill);
2674 if (db->db_level == 0) {
2675 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2676 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
2677 if (db->db_state != DB_NOFILL) {
2678 if (dr->dt.dl.dr_data != db->db_buf)
2679 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data,
2681 else if (!arc_released(db->db_buf))
2682 arc_set_callback(db->db_buf, dbuf_do_evict, db);
2689 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2690 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
2691 if (!BP_IS_HOLE(db->db_blkptr)) {
2693 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2694 ASSERT3U(db->db_blkid, <=,
2695 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs));
2696 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
2698 if (!arc_released(db->db_buf))
2699 arc_set_callback(db->db_buf, dbuf_do_evict, db);
2702 mutex_destroy(&dr->dt.di.dr_mtx);
2703 list_destroy(&dr->dt.di.dr_children);
2705 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2707 cv_broadcast(&db->db_changed);
2708 ASSERT(db->db_dirtycnt > 0);
2709 db->db_dirtycnt -= 1;
2710 db->db_data_pending = NULL;
2711 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg);
2715 dbuf_write_nofill_ready(zio_t *zio)
2717 dbuf_write_ready(zio, NULL, zio->io_private);
2721 dbuf_write_nofill_done(zio_t *zio)
2723 dbuf_write_done(zio, NULL, zio->io_private);
2727 dbuf_write_override_ready(zio_t *zio)
2729 dbuf_dirty_record_t *dr = zio->io_private;
2730 dmu_buf_impl_t *db = dr->dr_dbuf;
2732 dbuf_write_ready(zio, NULL, db);
2736 dbuf_write_override_done(zio_t *zio)
2738 dbuf_dirty_record_t *dr = zio->io_private;
2739 dmu_buf_impl_t *db = dr->dr_dbuf;
2740 blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
2742 mutex_enter(&db->db_mtx);
2743 if (!BP_EQUAL(zio->io_bp, obp)) {
2744 if (!BP_IS_HOLE(obp))
2745 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
2746 arc_release(dr->dt.dl.dr_data, db);
2748 mutex_exit(&db->db_mtx);
2750 dbuf_write_done(zio, NULL, db);
2753 /* Issue I/O to commit a dirty buffer to disk. */
2755 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
2757 dmu_buf_impl_t *db = dr->dr_dbuf;
2760 dmu_buf_impl_t *parent = db->db_parent;
2761 uint64_t txg = tx->tx_txg;
2762 zbookmark_phys_t zb;
2771 if (db->db_state != DB_NOFILL) {
2772 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
2774 * Private object buffers are released here rather
2775 * than in dbuf_dirty() since they are only modified
2776 * in the syncing context and we don't want the
2777 * overhead of making multiple copies of the data.
2779 if (BP_IS_HOLE(db->db_blkptr)) {
2782 dbuf_release_bp(db);
2787 if (parent != dn->dn_dbuf) {
2788 /* Our parent is an indirect block. */
2789 /* We have a dirty parent that has been scheduled for write. */
2790 ASSERT(parent && parent->db_data_pending);
2791 /* Our parent's buffer is one level closer to the dnode. */
2792 ASSERT(db->db_level == parent->db_level-1);
2794 * We're about to modify our parent's db_data by modifying
2795 * our block pointer, so the parent must be released.
2797 ASSERT(arc_released(parent->db_buf));
2798 zio = parent->db_data_pending->dr_zio;
2800 /* Our parent is the dnode itself. */
2801 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
2802 db->db_blkid != DMU_SPILL_BLKID) ||
2803 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
2804 if (db->db_blkid != DMU_SPILL_BLKID)
2805 ASSERT3P(db->db_blkptr, ==,
2806 &dn->dn_phys->dn_blkptr[db->db_blkid]);
2810 ASSERT(db->db_level == 0 || data == db->db_buf);
2811 ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
2814 SET_BOOKMARK(&zb, os->os_dsl_dataset ?
2815 os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
2816 db->db.db_object, db->db_level, db->db_blkid);
2818 if (db->db_blkid == DMU_SPILL_BLKID)
2820 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
2822 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
2825 if (db->db_level == 0 &&
2826 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
2828 * The BP for this block has been provided by open context
2829 * (by dmu_sync() or dmu_buf_write_embedded()).
2831 void *contents = (data != NULL) ? data->b_data : NULL;
2833 dr->dr_zio = zio_write(zio, os->os_spa, txg,
2834 db->db_blkptr, contents, db->db.db_size, &zp,
2835 dbuf_write_override_ready, NULL, dbuf_write_override_done,
2836 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
2837 mutex_enter(&db->db_mtx);
2838 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
2839 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
2840 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
2841 mutex_exit(&db->db_mtx);
2842 } else if (db->db_state == DB_NOFILL) {
2843 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF ||
2844 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY);
2845 dr->dr_zio = zio_write(zio, os->os_spa, txg,
2846 db->db_blkptr, NULL, db->db.db_size, &zp,
2847 dbuf_write_nofill_ready, NULL, dbuf_write_nofill_done, db,
2848 ZIO_PRIORITY_ASYNC_WRITE,
2849 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
2851 ASSERT(arc_released(data));
2852 dr->dr_zio = arc_write(zio, os->os_spa, txg,
2853 db->db_blkptr, data, DBUF_IS_L2CACHEABLE(db),
2854 DBUF_IS_L2COMPRESSIBLE(db), &zp, dbuf_write_ready,
2855 dbuf_write_physdone, dbuf_write_done, db,
2856 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);