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, 2016 by Delphix. All rights reserved.
25 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
26 * Copyright (c) 2013, Joyent, Inc. All rights reserved.
27 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
28 * Copyright (c) 2014 Integros [integros.com]
31 #include <sys/zfs_context.h>
33 #include <sys/dmu_send.h>
34 #include <sys/dmu_impl.h>
36 #include <sys/dmu_objset.h>
37 #include <sys/dsl_dataset.h>
38 #include <sys/dsl_dir.h>
39 #include <sys/dmu_tx.h>
42 #include <sys/dmu_zfetch.h>
44 #include <sys/sa_impl.h>
45 #include <sys/zfeature.h>
46 #include <sys/blkptr.h>
47 #include <sys/range_tree.h>
48 #include <sys/callb.h>
50 uint_t zfs_dbuf_evict_key;
52 static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
53 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
56 extern inline void dmu_buf_init_user(dmu_buf_user_t *dbu,
57 dmu_buf_evict_func_t *evict_func_sync,
58 dmu_buf_evict_func_t *evict_func_async,
59 dmu_buf_t **clear_on_evict_dbufp);
63 * Global data structures and functions for the dbuf cache.
65 static kmem_cache_t *dbuf_kmem_cache;
66 static taskq_t *dbu_evict_taskq;
68 static kthread_t *dbuf_cache_evict_thread;
69 static kmutex_t dbuf_evict_lock;
70 static kcondvar_t dbuf_evict_cv;
71 static boolean_t dbuf_evict_thread_exit;
74 * LRU cache of dbufs. The dbuf cache maintains a list of dbufs that
75 * are not currently held but have been recently released. These dbufs
76 * are not eligible for arc eviction until they are aged out of the cache.
77 * Dbufs are added to the dbuf cache once the last hold is released. If a
78 * dbuf is later accessed and still exists in the dbuf cache, then it will
79 * be removed from the cache and later re-added to the head of the cache.
80 * Dbufs that are aged out of the cache will be immediately destroyed and
81 * become eligible for arc eviction.
83 static multilist_t dbuf_cache;
84 static refcount_t dbuf_cache_size;
85 uint64_t dbuf_cache_max_bytes = 100 * 1024 * 1024;
87 /* Cap the size of the dbuf cache to log2 fraction of arc size. */
88 int dbuf_cache_max_shift = 5;
91 * The dbuf cache uses a three-stage eviction policy:
92 * - A low water marker designates when the dbuf eviction thread
93 * should stop evicting from the dbuf cache.
94 * - When we reach the maximum size (aka mid water mark), we
95 * signal the eviction thread to run.
96 * - The high water mark indicates when the eviction thread
97 * is unable to keep up with the incoming load and eviction must
98 * happen in the context of the calling thread.
102 * low water mid water hi water
103 * +----------------------------------------+----------+----------+
108 * +----------------------------------------+----------+----------+
110 * evicting eviction directly
113 * The high and low water marks indicate the operating range for the eviction
114 * thread. The low water mark is, by default, 90% of the total size of the
115 * cache and the high water mark is at 110% (both of these percentages can be
116 * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct,
117 * respectively). The eviction thread will try to ensure that the cache remains
118 * within this range by waking up every second and checking if the cache is
119 * above the low water mark. The thread can also be woken up by callers adding
120 * elements into the cache if the cache is larger than the mid water (i.e max
121 * cache size). Once the eviction thread is woken up and eviction is required,
122 * it will continue evicting buffers until it's able to reduce the cache size
123 * to the low water mark. If the cache size continues to grow and hits the high
124 * water mark, then callers adding elments to the cache will begin to evict
125 * directly from the cache until the cache is no longer above the high water
130 * The percentage above and below the maximum cache size.
132 uint_t dbuf_cache_hiwater_pct = 10;
133 uint_t dbuf_cache_lowater_pct = 10;
137 dbuf_cons(void *vdb, void *unused, int kmflag)
139 dmu_buf_impl_t *db = vdb;
140 bzero(db, sizeof (dmu_buf_impl_t));
142 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
143 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
144 multilist_link_init(&db->db_cache_link);
145 refcount_create(&db->db_holds);
152 dbuf_dest(void *vdb, void *unused)
154 dmu_buf_impl_t *db = vdb;
155 mutex_destroy(&db->db_mtx);
156 cv_destroy(&db->db_changed);
157 ASSERT(!multilist_link_active(&db->db_cache_link));
158 refcount_destroy(&db->db_holds);
162 * dbuf hash table routines
164 static dbuf_hash_table_t dbuf_hash_table;
166 static uint64_t dbuf_hash_count;
169 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
171 uintptr_t osv = (uintptr_t)os;
172 uint64_t crc = -1ULL;
174 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
175 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF];
176 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF];
177 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF];
178 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF];
179 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF];
180 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF];
182 crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16);
187 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
188 ((dbuf)->db.db_object == (obj) && \
189 (dbuf)->db_objset == (os) && \
190 (dbuf)->db_level == (level) && \
191 (dbuf)->db_blkid == (blkid))
194 dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid)
196 dbuf_hash_table_t *h = &dbuf_hash_table;
197 uint64_t hv = dbuf_hash(os, obj, level, blkid);
198 uint64_t idx = hv & h->hash_table_mask;
201 mutex_enter(DBUF_HASH_MUTEX(h, idx));
202 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
203 if (DBUF_EQUAL(db, os, obj, level, blkid)) {
204 mutex_enter(&db->db_mtx);
205 if (db->db_state != DB_EVICTING) {
206 mutex_exit(DBUF_HASH_MUTEX(h, idx));
209 mutex_exit(&db->db_mtx);
212 mutex_exit(DBUF_HASH_MUTEX(h, idx));
216 static dmu_buf_impl_t *
217 dbuf_find_bonus(objset_t *os, uint64_t object)
220 dmu_buf_impl_t *db = NULL;
222 if (dnode_hold(os, object, FTAG, &dn) == 0) {
223 rw_enter(&dn->dn_struct_rwlock, RW_READER);
224 if (dn->dn_bonus != NULL) {
226 mutex_enter(&db->db_mtx);
228 rw_exit(&dn->dn_struct_rwlock);
229 dnode_rele(dn, FTAG);
235 * Insert an entry into the hash table. If there is already an element
236 * equal to elem in the hash table, then the already existing element
237 * will be returned and the new element will not be inserted.
238 * Otherwise returns NULL.
240 static dmu_buf_impl_t *
241 dbuf_hash_insert(dmu_buf_impl_t *db)
243 dbuf_hash_table_t *h = &dbuf_hash_table;
244 objset_t *os = db->db_objset;
245 uint64_t obj = db->db.db_object;
246 int level = db->db_level;
247 uint64_t blkid = db->db_blkid;
248 uint64_t hv = dbuf_hash(os, obj, level, blkid);
249 uint64_t idx = hv & h->hash_table_mask;
252 mutex_enter(DBUF_HASH_MUTEX(h, idx));
253 for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) {
254 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
255 mutex_enter(&dbf->db_mtx);
256 if (dbf->db_state != DB_EVICTING) {
257 mutex_exit(DBUF_HASH_MUTEX(h, idx));
260 mutex_exit(&dbf->db_mtx);
264 mutex_enter(&db->db_mtx);
265 db->db_hash_next = h->hash_table[idx];
266 h->hash_table[idx] = db;
267 mutex_exit(DBUF_HASH_MUTEX(h, idx));
268 atomic_inc_64(&dbuf_hash_count);
274 * Remove an entry from the hash table. It must be in the EVICTING state.
277 dbuf_hash_remove(dmu_buf_impl_t *db)
279 dbuf_hash_table_t *h = &dbuf_hash_table;
280 uint64_t hv = dbuf_hash(db->db_objset, db->db.db_object,
281 db->db_level, db->db_blkid);
282 uint64_t idx = hv & h->hash_table_mask;
283 dmu_buf_impl_t *dbf, **dbp;
286 * We musn't hold db_mtx to maintain lock ordering:
287 * DBUF_HASH_MUTEX > db_mtx.
289 ASSERT(refcount_is_zero(&db->db_holds));
290 ASSERT(db->db_state == DB_EVICTING);
291 ASSERT(!MUTEX_HELD(&db->db_mtx));
293 mutex_enter(DBUF_HASH_MUTEX(h, idx));
294 dbp = &h->hash_table[idx];
295 while ((dbf = *dbp) != db) {
296 dbp = &dbf->db_hash_next;
299 *dbp = db->db_hash_next;
300 db->db_hash_next = NULL;
301 mutex_exit(DBUF_HASH_MUTEX(h, idx));
302 atomic_dec_64(&dbuf_hash_count);
308 } dbvu_verify_type_t;
311 dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type)
316 if (db->db_user == NULL)
319 /* Only data blocks support the attachment of user data. */
320 ASSERT(db->db_level == 0);
322 /* Clients must resolve a dbuf before attaching user data. */
323 ASSERT(db->db.db_data != NULL);
324 ASSERT3U(db->db_state, ==, DB_CACHED);
326 holds = refcount_count(&db->db_holds);
327 if (verify_type == DBVU_EVICTING) {
329 * Immediate eviction occurs when holds == dirtycnt.
330 * For normal eviction buffers, holds is zero on
331 * eviction, except when dbuf_fix_old_data() calls
332 * dbuf_clear_data(). However, the hold count can grow
333 * during eviction even though db_mtx is held (see
334 * dmu_bonus_hold() for an example), so we can only
335 * test the generic invariant that holds >= dirtycnt.
337 ASSERT3U(holds, >=, db->db_dirtycnt);
339 if (db->db_user_immediate_evict == TRUE)
340 ASSERT3U(holds, >=, db->db_dirtycnt);
342 ASSERT3U(holds, >, 0);
348 dbuf_evict_user(dmu_buf_impl_t *db)
350 dmu_buf_user_t *dbu = db->db_user;
352 ASSERT(MUTEX_HELD(&db->db_mtx));
357 dbuf_verify_user(db, DBVU_EVICTING);
361 if (dbu->dbu_clear_on_evict_dbufp != NULL)
362 *dbu->dbu_clear_on_evict_dbufp = NULL;
366 * There are two eviction callbacks - one that we call synchronously
367 * and one that we invoke via a taskq. The async one is useful for
368 * avoiding lock order reversals and limiting stack depth.
370 * Note that if we have a sync callback but no async callback,
371 * it's likely that the sync callback will free the structure
372 * containing the dbu. In that case we need to take care to not
373 * dereference dbu after calling the sync evict func.
375 boolean_t has_async = (dbu->dbu_evict_func_async != NULL);
377 if (dbu->dbu_evict_func_sync != NULL)
378 dbu->dbu_evict_func_sync(dbu);
381 taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func_async,
382 dbu, 0, &dbu->dbu_tqent);
387 dbuf_is_metadata(dmu_buf_impl_t *db)
389 if (db->db_level > 0) {
392 boolean_t is_metadata;
395 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type);
398 return (is_metadata);
403 * This function *must* return indices evenly distributed between all
404 * sublists of the multilist. This is needed due to how the dbuf eviction
405 * code is laid out; dbuf_evict_thread() assumes dbufs are evenly
406 * distributed between all sublists and uses this assumption when
407 * deciding which sublist to evict from and how much to evict from it.
410 dbuf_cache_multilist_index_func(multilist_t *ml, void *obj)
412 dmu_buf_impl_t *db = obj;
415 * The assumption here, is the hash value for a given
416 * dmu_buf_impl_t will remain constant throughout it's lifetime
417 * (i.e. it's objset, object, level and blkid fields don't change).
418 * Thus, we don't need to store the dbuf's sublist index
419 * on insertion, as this index can be recalculated on removal.
421 * Also, the low order bits of the hash value are thought to be
422 * distributed evenly. Otherwise, in the case that the multilist
423 * has a power of two number of sublists, each sublists' usage
424 * would not be evenly distributed.
426 return (dbuf_hash(db->db_objset, db->db.db_object,
427 db->db_level, db->db_blkid) %
428 multilist_get_num_sublists(ml));
431 static inline boolean_t
432 dbuf_cache_above_hiwater(void)
434 uint64_t dbuf_cache_hiwater_bytes =
435 (dbuf_cache_max_bytes * dbuf_cache_hiwater_pct) / 100;
437 return (refcount_count(&dbuf_cache_size) >
438 dbuf_cache_max_bytes + dbuf_cache_hiwater_bytes);
441 static inline boolean_t
442 dbuf_cache_above_lowater(void)
444 uint64_t dbuf_cache_lowater_bytes =
445 (dbuf_cache_max_bytes * dbuf_cache_lowater_pct) / 100;
447 return (refcount_count(&dbuf_cache_size) >
448 dbuf_cache_max_bytes - dbuf_cache_lowater_bytes);
452 * Evict the oldest eligible dbuf from the dbuf cache.
457 int idx = multilist_get_random_index(&dbuf_cache);
458 multilist_sublist_t *mls = multilist_sublist_lock(&dbuf_cache, idx);
460 ASSERT(!MUTEX_HELD(&dbuf_evict_lock));
463 * Set the thread's tsd to indicate that it's processing evictions.
464 * Once a thread stops evicting from the dbuf cache it will
465 * reset its tsd to NULL.
467 ASSERT3P(tsd_get(zfs_dbuf_evict_key), ==, NULL);
468 (void) tsd_set(zfs_dbuf_evict_key, (void *)B_TRUE);
470 dmu_buf_impl_t *db = multilist_sublist_tail(mls);
471 while (db != NULL && mutex_tryenter(&db->db_mtx) == 0) {
472 db = multilist_sublist_prev(mls, db);
475 DTRACE_PROBE2(dbuf__evict__one, dmu_buf_impl_t *, db,
476 multilist_sublist_t *, mls);
479 multilist_sublist_remove(mls, db);
480 multilist_sublist_unlock(mls);
481 (void) refcount_remove_many(&dbuf_cache_size,
485 multilist_sublist_unlock(mls);
487 (void) tsd_set(zfs_dbuf_evict_key, NULL);
491 * The dbuf evict thread is responsible for aging out dbufs from the
492 * cache. Once the cache has reached it's maximum size, dbufs are removed
493 * and destroyed. The eviction thread will continue running until the size
494 * of the dbuf cache is at or below the maximum size. Once the dbuf is aged
495 * out of the cache it is destroyed and becomes eligible for arc eviction.
498 dbuf_evict_thread(void *dummy __unused)
502 CALLB_CPR_INIT(&cpr, &dbuf_evict_lock, callb_generic_cpr, FTAG);
504 mutex_enter(&dbuf_evict_lock);
505 while (!dbuf_evict_thread_exit) {
506 while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
507 CALLB_CPR_SAFE_BEGIN(&cpr);
508 (void) cv_timedwait_hires(&dbuf_evict_cv,
509 &dbuf_evict_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
510 CALLB_CPR_SAFE_END(&cpr, &dbuf_evict_lock);
512 mutex_exit(&dbuf_evict_lock);
515 * Keep evicting as long as we're above the low water mark
516 * for the cache. We do this without holding the locks to
517 * minimize lock contention.
519 while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
523 mutex_enter(&dbuf_evict_lock);
526 dbuf_evict_thread_exit = B_FALSE;
527 cv_broadcast(&dbuf_evict_cv);
528 CALLB_CPR_EXIT(&cpr); /* drops dbuf_evict_lock */
533 * Wake up the dbuf eviction thread if the dbuf cache is at its max size.
534 * If the dbuf cache is at its high water mark, then evict a dbuf from the
535 * dbuf cache using the callers context.
538 dbuf_evict_notify(void)
542 * We use thread specific data to track when a thread has
543 * started processing evictions. This allows us to avoid deeply
544 * nested stacks that would have a call flow similar to this:
546 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify()
549 * +-----dbuf_destroy()<--dbuf_evict_one()<--------+
551 * The dbuf_eviction_thread will always have its tsd set until
552 * that thread exits. All other threads will only set their tsd
553 * if they are participating in the eviction process. This only
554 * happens if the eviction thread is unable to process evictions
555 * fast enough. To keep the dbuf cache size in check, other threads
556 * can evict from the dbuf cache directly. Those threads will set
557 * their tsd values so that we ensure that they only evict one dbuf
558 * from the dbuf cache.
560 if (tsd_get(zfs_dbuf_evict_key) != NULL)
563 if (refcount_count(&dbuf_cache_size) > dbuf_cache_max_bytes) {
564 boolean_t evict_now = B_FALSE;
566 mutex_enter(&dbuf_evict_lock);
567 if (refcount_count(&dbuf_cache_size) > dbuf_cache_max_bytes) {
568 evict_now = dbuf_cache_above_hiwater();
569 cv_signal(&dbuf_evict_cv);
571 mutex_exit(&dbuf_evict_lock);
582 uint64_t hsize = 1ULL << 16;
583 dbuf_hash_table_t *h = &dbuf_hash_table;
587 * The hash table is big enough to fill all of physical memory
588 * with an average 4K block size. The table will take up
589 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
591 while (hsize * 4096 < (uint64_t)physmem * PAGESIZE)
595 h->hash_table_mask = hsize - 1;
596 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
597 if (h->hash_table == NULL) {
598 /* XXX - we should really return an error instead of assert */
599 ASSERT(hsize > (1ULL << 10));
604 dbuf_kmem_cache = kmem_cache_create("dmu_buf_impl_t",
605 sizeof (dmu_buf_impl_t),
606 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
608 for (i = 0; i < DBUF_MUTEXES; i++)
609 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
612 * Setup the parameters for the dbuf cache. We cap the size of the
613 * dbuf cache to 1/32nd (default) of the size of the ARC.
615 dbuf_cache_max_bytes = MIN(dbuf_cache_max_bytes,
616 arc_max_bytes() >> dbuf_cache_max_shift);
619 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
620 * configuration is not required.
622 dbu_evict_taskq = taskq_create("dbu_evict", 1, minclsyspri, 0, 0, 0);
624 multilist_create(&dbuf_cache, sizeof (dmu_buf_impl_t),
625 offsetof(dmu_buf_impl_t, db_cache_link),
626 zfs_arc_num_sublists_per_state,
627 dbuf_cache_multilist_index_func);
628 refcount_create(&dbuf_cache_size);
630 tsd_create(&zfs_dbuf_evict_key, NULL);
631 dbuf_evict_thread_exit = B_FALSE;
632 mutex_init(&dbuf_evict_lock, NULL, MUTEX_DEFAULT, NULL);
633 cv_init(&dbuf_evict_cv, NULL, CV_DEFAULT, NULL);
634 dbuf_cache_evict_thread = thread_create(NULL, 0, dbuf_evict_thread,
635 NULL, 0, &p0, TS_RUN, minclsyspri);
641 dbuf_hash_table_t *h = &dbuf_hash_table;
644 for (i = 0; i < DBUF_MUTEXES; i++)
645 mutex_destroy(&h->hash_mutexes[i]);
646 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
647 kmem_cache_destroy(dbuf_kmem_cache);
648 taskq_destroy(dbu_evict_taskq);
650 mutex_enter(&dbuf_evict_lock);
651 dbuf_evict_thread_exit = B_TRUE;
652 while (dbuf_evict_thread_exit) {
653 cv_signal(&dbuf_evict_cv);
654 cv_wait(&dbuf_evict_cv, &dbuf_evict_lock);
656 mutex_exit(&dbuf_evict_lock);
657 tsd_destroy(&zfs_dbuf_evict_key);
659 mutex_destroy(&dbuf_evict_lock);
660 cv_destroy(&dbuf_evict_cv);
662 refcount_destroy(&dbuf_cache_size);
663 multilist_destroy(&dbuf_cache);
672 dbuf_verify(dmu_buf_impl_t *db)
675 dbuf_dirty_record_t *dr;
677 ASSERT(MUTEX_HELD(&db->db_mtx));
679 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
682 ASSERT(db->db_objset != NULL);
686 ASSERT(db->db_parent == NULL);
687 ASSERT(db->db_blkptr == NULL);
689 ASSERT3U(db->db.db_object, ==, dn->dn_object);
690 ASSERT3P(db->db_objset, ==, dn->dn_objset);
691 ASSERT3U(db->db_level, <, dn->dn_nlevels);
692 ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
693 db->db_blkid == DMU_SPILL_BLKID ||
694 !avl_is_empty(&dn->dn_dbufs));
696 if (db->db_blkid == DMU_BONUS_BLKID) {
698 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
699 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
700 } else if (db->db_blkid == DMU_SPILL_BLKID) {
702 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
703 ASSERT0(db->db.db_offset);
705 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
708 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
709 ASSERT(dr->dr_dbuf == db);
711 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
712 ASSERT(dr->dr_dbuf == db);
715 * We can't assert that db_size matches dn_datablksz because it
716 * can be momentarily different when another thread is doing
719 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
720 dr = db->db_data_pending;
722 * It should only be modified in syncing context, so
723 * make sure we only have one copy of the data.
725 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
728 /* verify db->db_blkptr */
730 if (db->db_parent == dn->dn_dbuf) {
731 /* db is pointed to by the dnode */
732 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
733 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
734 ASSERT(db->db_parent == NULL);
736 ASSERT(db->db_parent != NULL);
737 if (db->db_blkid != DMU_SPILL_BLKID)
738 ASSERT3P(db->db_blkptr, ==,
739 &dn->dn_phys->dn_blkptr[db->db_blkid]);
741 /* db is pointed to by an indirect block */
742 int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
743 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
744 ASSERT3U(db->db_parent->db.db_object, ==,
747 * dnode_grow_indblksz() can make this fail if we don't
748 * have the struct_rwlock. XXX indblksz no longer
749 * grows. safe to do this now?
751 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
752 ASSERT3P(db->db_blkptr, ==,
753 ((blkptr_t *)db->db_parent->db.db_data +
754 db->db_blkid % epb));
758 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
759 (db->db_buf == NULL || db->db_buf->b_data) &&
760 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
761 db->db_state != DB_FILL && !dn->dn_free_txg) {
763 * If the blkptr isn't set but they have nonzero data,
764 * it had better be dirty, otherwise we'll lose that
765 * data when we evict this buffer.
767 * There is an exception to this rule for indirect blocks; in
768 * this case, if the indirect block is a hole, we fill in a few
769 * fields on each of the child blocks (importantly, birth time)
770 * to prevent hole birth times from being lost when you
771 * partially fill in a hole.
773 if (db->db_dirtycnt == 0) {
774 if (db->db_level == 0) {
775 uint64_t *buf = db->db.db_data;
778 for (i = 0; i < db->db.db_size >> 3; i++) {
782 blkptr_t *bps = db->db.db_data;
783 ASSERT3U(1 << DB_DNODE(db)->dn_indblkshift, ==,
786 * We want to verify that all the blkptrs in the
787 * indirect block are holes, but we may have
788 * automatically set up a few fields for them.
789 * We iterate through each blkptr and verify
790 * they only have those fields set.
793 i < db->db.db_size / sizeof (blkptr_t);
795 blkptr_t *bp = &bps[i];
796 ASSERT(ZIO_CHECKSUM_IS_ZERO(
799 DVA_IS_EMPTY(&bp->blk_dva[0]) &&
800 DVA_IS_EMPTY(&bp->blk_dva[1]) &&
801 DVA_IS_EMPTY(&bp->blk_dva[2]));
802 ASSERT0(bp->blk_fill);
803 ASSERT0(bp->blk_pad[0]);
804 ASSERT0(bp->blk_pad[1]);
805 ASSERT(!BP_IS_EMBEDDED(bp));
806 ASSERT(BP_IS_HOLE(bp));
807 ASSERT0(bp->blk_phys_birth);
817 dbuf_clear_data(dmu_buf_impl_t *db)
819 ASSERT(MUTEX_HELD(&db->db_mtx));
821 ASSERT3P(db->db_buf, ==, NULL);
822 db->db.db_data = NULL;
823 if (db->db_state != DB_NOFILL)
824 db->db_state = DB_UNCACHED;
828 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
830 ASSERT(MUTEX_HELD(&db->db_mtx));
834 ASSERT(buf->b_data != NULL);
835 db->db.db_data = buf->b_data;
839 * Loan out an arc_buf for read. Return the loaned arc_buf.
842 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
846 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
847 mutex_enter(&db->db_mtx);
848 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
849 int blksz = db->db.db_size;
850 spa_t *spa = db->db_objset->os_spa;
852 mutex_exit(&db->db_mtx);
853 abuf = arc_loan_buf(spa, blksz);
854 bcopy(db->db.db_data, abuf->b_data, blksz);
857 arc_loan_inuse_buf(abuf, db);
860 mutex_exit(&db->db_mtx);
866 * Calculate which level n block references the data at the level 0 offset
870 dbuf_whichblock(dnode_t *dn, int64_t level, uint64_t offset)
872 if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) {
874 * The level n blkid is equal to the level 0 blkid divided by
875 * the number of level 0s in a level n block.
877 * The level 0 blkid is offset >> datablkshift =
878 * offset / 2^datablkshift.
880 * The number of level 0s in a level n is the number of block
881 * pointers in an indirect block, raised to the power of level.
882 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
883 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
885 * Thus, the level n blkid is: offset /
886 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
887 * = offset / 2^(datablkshift + level *
888 * (indblkshift - SPA_BLKPTRSHIFT))
889 * = offset >> (datablkshift + level *
890 * (indblkshift - SPA_BLKPTRSHIFT))
892 return (offset >> (dn->dn_datablkshift + level *
893 (dn->dn_indblkshift - SPA_BLKPTRSHIFT)));
895 ASSERT3U(offset, <, dn->dn_datablksz);
901 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
903 dmu_buf_impl_t *db = vdb;
905 mutex_enter(&db->db_mtx);
906 ASSERT3U(db->db_state, ==, DB_READ);
908 * All reads are synchronous, so we must have a hold on the dbuf
910 ASSERT(refcount_count(&db->db_holds) > 0);
911 ASSERT(db->db_buf == NULL);
912 ASSERT(db->db.db_data == NULL);
913 if (db->db_level == 0 && db->db_freed_in_flight) {
914 /* we were freed in flight; disregard any error */
915 arc_release(buf, db);
916 bzero(buf->b_data, db->db.db_size);
918 db->db_freed_in_flight = FALSE;
919 dbuf_set_data(db, buf);
920 db->db_state = DB_CACHED;
921 } else if (zio == NULL || zio->io_error == 0) {
922 dbuf_set_data(db, buf);
923 db->db_state = DB_CACHED;
925 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
926 ASSERT3P(db->db_buf, ==, NULL);
927 arc_buf_destroy(buf, db);
928 db->db_state = DB_UNCACHED;
930 cv_broadcast(&db->db_changed);
931 dbuf_rele_and_unlock(db, NULL);
935 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
939 arc_flags_t aflags = ARC_FLAG_NOWAIT;
943 ASSERT(!refcount_is_zero(&db->db_holds));
944 /* We need the struct_rwlock to prevent db_blkptr from changing. */
945 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
946 ASSERT(MUTEX_HELD(&db->db_mtx));
947 ASSERT(db->db_state == DB_UNCACHED);
948 ASSERT(db->db_buf == NULL);
950 if (db->db_blkid == DMU_BONUS_BLKID) {
951 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
953 ASSERT3U(bonuslen, <=, db->db.db_size);
954 db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN);
955 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
956 if (bonuslen < DN_MAX_BONUSLEN)
957 bzero(db->db.db_data, DN_MAX_BONUSLEN);
959 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
961 db->db_state = DB_CACHED;
962 mutex_exit(&db->db_mtx);
967 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
968 * processes the delete record and clears the bp while we are waiting
969 * for the dn_mtx (resulting in a "no" from block_freed).
971 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
972 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
973 BP_IS_HOLE(db->db_blkptr)))) {
974 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
976 dbuf_set_data(db, arc_alloc_buf(db->db_objset->os_spa,
977 db->db.db_size, db, type));
978 bzero(db->db.db_data, db->db.db_size);
980 if (db->db_blkptr != NULL && db->db_level > 0 &&
981 BP_IS_HOLE(db->db_blkptr) &&
982 db->db_blkptr->blk_birth != 0) {
983 blkptr_t *bps = db->db.db_data;
984 for (int i = 0; i < ((1 <<
985 DB_DNODE(db)->dn_indblkshift) / sizeof (blkptr_t));
987 blkptr_t *bp = &bps[i];
988 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
989 1 << dn->dn_indblkshift);
991 BP_GET_LEVEL(db->db_blkptr) == 1 ?
993 BP_GET_LSIZE(db->db_blkptr));
994 BP_SET_TYPE(bp, BP_GET_TYPE(db->db_blkptr));
996 BP_GET_LEVEL(db->db_blkptr) - 1);
997 BP_SET_BIRTH(bp, db->db_blkptr->blk_birth, 0);
1001 db->db_state = DB_CACHED;
1002 mutex_exit(&db->db_mtx);
1008 db->db_state = DB_READ;
1009 mutex_exit(&db->db_mtx);
1011 if (DBUF_IS_L2CACHEABLE(db))
1012 aflags |= ARC_FLAG_L2CACHE;
1014 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
1015 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
1016 db->db.db_object, db->db_level, db->db_blkid);
1018 dbuf_add_ref(db, NULL);
1020 (void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr,
1021 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
1022 (flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
1027 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
1030 boolean_t havepzio = (zio != NULL);
1035 * We don't have to hold the mutex to check db_state because it
1036 * can't be freed while we have a hold on the buffer.
1038 ASSERT(!refcount_is_zero(&db->db_holds));
1040 if (db->db_state == DB_NOFILL)
1041 return (SET_ERROR(EIO));
1045 if ((flags & DB_RF_HAVESTRUCT) == 0)
1046 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1048 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1049 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
1050 DBUF_IS_CACHEABLE(db);
1052 mutex_enter(&db->db_mtx);
1053 if (db->db_state == DB_CACHED) {
1054 mutex_exit(&db->db_mtx);
1056 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1057 if ((flags & DB_RF_HAVESTRUCT) == 0)
1058 rw_exit(&dn->dn_struct_rwlock);
1060 } else if (db->db_state == DB_UNCACHED) {
1061 spa_t *spa = dn->dn_objset->os_spa;
1064 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
1065 dbuf_read_impl(db, zio, flags);
1067 /* dbuf_read_impl has dropped db_mtx for us */
1070 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1072 if ((flags & DB_RF_HAVESTRUCT) == 0)
1073 rw_exit(&dn->dn_struct_rwlock);
1077 err = zio_wait(zio);
1080 * Another reader came in while the dbuf was in flight
1081 * between UNCACHED and CACHED. Either a writer will finish
1082 * writing the buffer (sending the dbuf to CACHED) or the
1083 * first reader's request will reach the read_done callback
1084 * and send the dbuf to CACHED. Otherwise, a failure
1085 * occurred and the dbuf went to UNCACHED.
1087 mutex_exit(&db->db_mtx);
1089 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1090 if ((flags & DB_RF_HAVESTRUCT) == 0)
1091 rw_exit(&dn->dn_struct_rwlock);
1094 /* Skip the wait per the caller's request. */
1095 mutex_enter(&db->db_mtx);
1096 if ((flags & DB_RF_NEVERWAIT) == 0) {
1097 while (db->db_state == DB_READ ||
1098 db->db_state == DB_FILL) {
1099 ASSERT(db->db_state == DB_READ ||
1100 (flags & DB_RF_HAVESTRUCT) == 0);
1101 DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *,
1103 cv_wait(&db->db_changed, &db->db_mtx);
1105 if (db->db_state == DB_UNCACHED)
1106 err = SET_ERROR(EIO);
1108 mutex_exit(&db->db_mtx);
1111 ASSERT(err || havepzio || db->db_state == DB_CACHED);
1116 dbuf_noread(dmu_buf_impl_t *db)
1118 ASSERT(!refcount_is_zero(&db->db_holds));
1119 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1120 mutex_enter(&db->db_mtx);
1121 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1122 cv_wait(&db->db_changed, &db->db_mtx);
1123 if (db->db_state == DB_UNCACHED) {
1124 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1125 spa_t *spa = db->db_objset->os_spa;
1127 ASSERT(db->db_buf == NULL);
1128 ASSERT(db->db.db_data == NULL);
1129 dbuf_set_data(db, arc_alloc_buf(spa, db->db.db_size, db, type));
1130 db->db_state = DB_FILL;
1131 } else if (db->db_state == DB_NOFILL) {
1132 dbuf_clear_data(db);
1134 ASSERT3U(db->db_state, ==, DB_CACHED);
1136 mutex_exit(&db->db_mtx);
1140 * This is our just-in-time copy function. It makes a copy of
1141 * buffers, that have been modified in a previous transaction
1142 * group, before we modify them in the current active group.
1144 * This function is used in two places: when we are dirtying a
1145 * buffer for the first time in a txg, and when we are freeing
1146 * a range in a dnode that includes this buffer.
1148 * Note that when we are called from dbuf_free_range() we do
1149 * not put a hold on the buffer, we just traverse the active
1150 * dbuf list for the dnode.
1153 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
1155 dbuf_dirty_record_t *dr = db->db_last_dirty;
1157 ASSERT(MUTEX_HELD(&db->db_mtx));
1158 ASSERT(db->db.db_data != NULL);
1159 ASSERT(db->db_level == 0);
1160 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
1163 (dr->dt.dl.dr_data !=
1164 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
1168 * If the last dirty record for this dbuf has not yet synced
1169 * and its referencing the dbuf data, either:
1170 * reset the reference to point to a new copy,
1171 * or (if there a no active holders)
1172 * just null out the current db_data pointer.
1174 ASSERT(dr->dr_txg >= txg - 2);
1175 if (db->db_blkid == DMU_BONUS_BLKID) {
1176 /* Note that the data bufs here are zio_bufs */
1177 dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
1178 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
1179 bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
1180 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
1181 int size = db->db.db_size;
1182 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1183 spa_t *spa = db->db_objset->os_spa;
1185 dr->dt.dl.dr_data = arc_alloc_buf(spa, size, db, type);
1186 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
1189 dbuf_clear_data(db);
1194 dbuf_unoverride(dbuf_dirty_record_t *dr)
1196 dmu_buf_impl_t *db = dr->dr_dbuf;
1197 blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
1198 uint64_t txg = dr->dr_txg;
1200 ASSERT(MUTEX_HELD(&db->db_mtx));
1201 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
1202 ASSERT(db->db_level == 0);
1204 if (db->db_blkid == DMU_BONUS_BLKID ||
1205 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
1208 ASSERT(db->db_data_pending != dr);
1210 /* free this block */
1211 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite)
1212 zio_free(db->db_objset->os_spa, txg, bp);
1214 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1215 dr->dt.dl.dr_nopwrite = B_FALSE;
1218 * Release the already-written buffer, so we leave it in
1219 * a consistent dirty state. Note that all callers are
1220 * modifying the buffer, so they will immediately do
1221 * another (redundant) arc_release(). Therefore, leave
1222 * the buf thawed to save the effort of freezing &
1223 * immediately re-thawing it.
1225 arc_release(dr->dt.dl.dr_data, db);
1229 * Evict (if its unreferenced) or clear (if its referenced) any level-0
1230 * data blocks in the free range, so that any future readers will find
1234 dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
1237 dmu_buf_impl_t db_search;
1238 dmu_buf_impl_t *db, *db_next;
1239 uint64_t txg = tx->tx_txg;
1242 if (end_blkid > dn->dn_maxblkid &&
1243 !(start_blkid == DMU_SPILL_BLKID || end_blkid == DMU_SPILL_BLKID))
1244 end_blkid = dn->dn_maxblkid;
1245 dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid);
1247 db_search.db_level = 0;
1248 db_search.db_blkid = start_blkid;
1249 db_search.db_state = DB_SEARCH;
1251 mutex_enter(&dn->dn_dbufs_mtx);
1252 db = avl_find(&dn->dn_dbufs, &db_search, &where);
1253 ASSERT3P(db, ==, NULL);
1255 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1257 for (; db != NULL; db = db_next) {
1258 db_next = AVL_NEXT(&dn->dn_dbufs, db);
1259 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1261 if (db->db_level != 0 || db->db_blkid > end_blkid) {
1264 ASSERT3U(db->db_blkid, >=, start_blkid);
1266 /* found a level 0 buffer in the range */
1267 mutex_enter(&db->db_mtx);
1268 if (dbuf_undirty(db, tx)) {
1269 /* mutex has been dropped and dbuf destroyed */
1273 if (db->db_state == DB_UNCACHED ||
1274 db->db_state == DB_NOFILL ||
1275 db->db_state == DB_EVICTING) {
1276 ASSERT(db->db.db_data == NULL);
1277 mutex_exit(&db->db_mtx);
1280 if (db->db_state == DB_READ || db->db_state == DB_FILL) {
1281 /* will be handled in dbuf_read_done or dbuf_rele */
1282 db->db_freed_in_flight = TRUE;
1283 mutex_exit(&db->db_mtx);
1286 if (refcount_count(&db->db_holds) == 0) {
1291 /* The dbuf is referenced */
1293 if (db->db_last_dirty != NULL) {
1294 dbuf_dirty_record_t *dr = db->db_last_dirty;
1296 if (dr->dr_txg == txg) {
1298 * This buffer is "in-use", re-adjust the file
1299 * size to reflect that this buffer may
1300 * contain new data when we sync.
1302 if (db->db_blkid != DMU_SPILL_BLKID &&
1303 db->db_blkid > dn->dn_maxblkid)
1304 dn->dn_maxblkid = db->db_blkid;
1305 dbuf_unoverride(dr);
1308 * This dbuf is not dirty in the open context.
1309 * Either uncache it (if its not referenced in
1310 * the open context) or reset its contents to
1313 dbuf_fix_old_data(db, txg);
1316 /* clear the contents if its cached */
1317 if (db->db_state == DB_CACHED) {
1318 ASSERT(db->db.db_data != NULL);
1319 arc_release(db->db_buf, db);
1320 bzero(db->db.db_data, db->db.db_size);
1321 arc_buf_freeze(db->db_buf);
1324 mutex_exit(&db->db_mtx);
1326 mutex_exit(&dn->dn_dbufs_mtx);
1330 dbuf_block_freeable(dmu_buf_impl_t *db)
1332 dsl_dataset_t *ds = db->db_objset->os_dsl_dataset;
1333 uint64_t birth_txg = 0;
1336 * We don't need any locking to protect db_blkptr:
1337 * If it's syncing, then db_last_dirty will be set
1338 * so we'll ignore db_blkptr.
1340 * This logic ensures that only block births for
1341 * filled blocks are considered.
1343 ASSERT(MUTEX_HELD(&db->db_mtx));
1344 if (db->db_last_dirty && (db->db_blkptr == NULL ||
1345 !BP_IS_HOLE(db->db_blkptr))) {
1346 birth_txg = db->db_last_dirty->dr_txg;
1347 } else if (db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) {
1348 birth_txg = db->db_blkptr->blk_birth;
1352 * If this block don't exist or is in a snapshot, it can't be freed.
1353 * Don't pass the bp to dsl_dataset_block_freeable() since we
1354 * are holding the db_mtx lock and might deadlock if we are
1355 * prefetching a dedup-ed block.
1358 return (ds == NULL ||
1359 dsl_dataset_block_freeable(ds, NULL, birth_txg));
1365 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
1367 arc_buf_t *buf, *obuf;
1368 int osize = db->db.db_size;
1369 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1372 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1377 /* XXX does *this* func really need the lock? */
1378 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1381 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1382 * is OK, because there can be no other references to the db
1383 * when we are changing its size, so no concurrent DB_FILL can
1387 * XXX we should be doing a dbuf_read, checking the return
1388 * value and returning that up to our callers
1390 dmu_buf_will_dirty(&db->db, tx);
1392 /* create the data buffer for the new block */
1393 buf = arc_alloc_buf(dn->dn_objset->os_spa, size, db, type);
1395 /* copy old block data to the new block */
1397 bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
1398 /* zero the remainder */
1400 bzero((uint8_t *)buf->b_data + osize, size - osize);
1402 mutex_enter(&db->db_mtx);
1403 dbuf_set_data(db, buf);
1404 arc_buf_destroy(obuf, db);
1405 db->db.db_size = size;
1407 if (db->db_level == 0) {
1408 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1409 db->db_last_dirty->dt.dl.dr_data = buf;
1411 mutex_exit(&db->db_mtx);
1413 dnode_willuse_space(dn, size-osize, tx);
1418 dbuf_release_bp(dmu_buf_impl_t *db)
1420 objset_t *os = db->db_objset;
1422 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1423 ASSERT(arc_released(os->os_phys_buf) ||
1424 list_link_active(&os->os_dsl_dataset->ds_synced_link));
1425 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1427 (void) arc_release(db->db_buf, db);
1431 * We already have a dirty record for this TXG, and we are being
1435 dbuf_redirty(dbuf_dirty_record_t *dr)
1437 dmu_buf_impl_t *db = dr->dr_dbuf;
1439 ASSERT(MUTEX_HELD(&db->db_mtx));
1441 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1443 * If this buffer has already been written out,
1444 * we now need to reset its state.
1446 dbuf_unoverride(dr);
1447 if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1448 db->db_state != DB_NOFILL) {
1449 /* Already released on initial dirty, so just thaw. */
1450 ASSERT(arc_released(db->db_buf));
1451 arc_buf_thaw(db->db_buf);
1456 dbuf_dirty_record_t *
1457 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1461 dbuf_dirty_record_t **drp, *dr;
1462 int drop_struct_lock = FALSE;
1463 boolean_t do_free_accounting = B_FALSE;
1464 int txgoff = tx->tx_txg & TXG_MASK;
1466 ASSERT(tx->tx_txg != 0);
1467 ASSERT(!refcount_is_zero(&db->db_holds));
1468 DMU_TX_DIRTY_BUF(tx, db);
1473 * Shouldn't dirty a regular buffer in syncing context. Private
1474 * objects may be dirtied in syncing context, but only if they
1475 * were already pre-dirtied in open context.
1478 if (dn->dn_objset->os_dsl_dataset != NULL) {
1479 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1482 ASSERT(!dmu_tx_is_syncing(tx) ||
1483 BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1484 DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1485 dn->dn_objset->os_dsl_dataset == NULL);
1486 if (dn->dn_objset->os_dsl_dataset != NULL)
1487 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, FTAG);
1490 * We make this assert for private objects as well, but after we
1491 * check if we're already dirty. They are allowed to re-dirty
1492 * in syncing context.
1494 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1495 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1496 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1498 mutex_enter(&db->db_mtx);
1500 * XXX make this true for indirects too? The problem is that
1501 * transactions created with dmu_tx_create_assigned() from
1502 * syncing context don't bother holding ahead.
1504 ASSERT(db->db_level != 0 ||
1505 db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1506 db->db_state == DB_NOFILL);
1508 mutex_enter(&dn->dn_mtx);
1510 * Don't set dirtyctx to SYNC if we're just modifying this as we
1511 * initialize the objset.
1513 if (dn->dn_dirtyctx == DN_UNDIRTIED) {
1514 if (dn->dn_objset->os_dsl_dataset != NULL) {
1515 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1518 if (!BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1519 dn->dn_dirtyctx = (dmu_tx_is_syncing(tx) ?
1520 DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1521 ASSERT(dn->dn_dirtyctx_firstset == NULL);
1522 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1524 if (dn->dn_objset->os_dsl_dataset != NULL) {
1525 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1529 mutex_exit(&dn->dn_mtx);
1531 if (db->db_blkid == DMU_SPILL_BLKID)
1532 dn->dn_have_spill = B_TRUE;
1535 * If this buffer is already dirty, we're done.
1537 drp = &db->db_last_dirty;
1538 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1539 db->db.db_object == DMU_META_DNODE_OBJECT);
1540 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1542 if (dr && dr->dr_txg == tx->tx_txg) {
1546 mutex_exit(&db->db_mtx);
1551 * Only valid if not already dirty.
1553 ASSERT(dn->dn_object == 0 ||
1554 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1555 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1557 ASSERT3U(dn->dn_nlevels, >, db->db_level);
1558 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1559 dn->dn_phys->dn_nlevels > db->db_level ||
1560 dn->dn_next_nlevels[txgoff] > db->db_level ||
1561 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1562 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1565 * We should only be dirtying in syncing context if it's the
1566 * mos or we're initializing the os or it's a special object.
1567 * However, we are allowed to dirty in syncing context provided
1568 * we already dirtied it in open context. Hence we must make
1569 * this assertion only if we're not already dirty.
1573 if (dn->dn_objset->os_dsl_dataset != NULL)
1574 rrw_enter(&os->os_dsl_dataset->ds_bp_rwlock, RW_READER, FTAG);
1575 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1576 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1577 if (dn->dn_objset->os_dsl_dataset != NULL)
1578 rrw_exit(&os->os_dsl_dataset->ds_bp_rwlock, FTAG);
1580 ASSERT(db->db.db_size != 0);
1582 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1584 if (db->db_blkid != DMU_BONUS_BLKID) {
1586 * Update the accounting.
1587 * Note: we delay "free accounting" until after we drop
1588 * the db_mtx. This keeps us from grabbing other locks
1589 * (and possibly deadlocking) in bp_get_dsize() while
1590 * also holding the db_mtx.
1592 dnode_willuse_space(dn, db->db.db_size, tx);
1593 do_free_accounting = dbuf_block_freeable(db);
1597 * If this buffer is dirty in an old transaction group we need
1598 * to make a copy of it so that the changes we make in this
1599 * transaction group won't leak out when we sync the older txg.
1601 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1602 if (db->db_level == 0) {
1603 void *data_old = db->db_buf;
1605 if (db->db_state != DB_NOFILL) {
1606 if (db->db_blkid == DMU_BONUS_BLKID) {
1607 dbuf_fix_old_data(db, tx->tx_txg);
1608 data_old = db->db.db_data;
1609 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1611 * Release the data buffer from the cache so
1612 * that we can modify it without impacting
1613 * possible other users of this cached data
1614 * block. Note that indirect blocks and
1615 * private objects are not released until the
1616 * syncing state (since they are only modified
1619 arc_release(db->db_buf, db);
1620 dbuf_fix_old_data(db, tx->tx_txg);
1621 data_old = db->db_buf;
1623 ASSERT(data_old != NULL);
1625 dr->dt.dl.dr_data = data_old;
1627 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1628 list_create(&dr->dt.di.dr_children,
1629 sizeof (dbuf_dirty_record_t),
1630 offsetof(dbuf_dirty_record_t, dr_dirty_node));
1632 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL)
1633 dr->dr_accounted = db->db.db_size;
1635 dr->dr_txg = tx->tx_txg;
1640 * We could have been freed_in_flight between the dbuf_noread
1641 * and dbuf_dirty. We win, as though the dbuf_noread() had
1642 * happened after the free.
1644 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1645 db->db_blkid != DMU_SPILL_BLKID) {
1646 mutex_enter(&dn->dn_mtx);
1647 if (dn->dn_free_ranges[txgoff] != NULL) {
1648 range_tree_clear(dn->dn_free_ranges[txgoff],
1651 mutex_exit(&dn->dn_mtx);
1652 db->db_freed_in_flight = FALSE;
1656 * This buffer is now part of this txg
1658 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1659 db->db_dirtycnt += 1;
1660 ASSERT3U(db->db_dirtycnt, <=, 3);
1662 mutex_exit(&db->db_mtx);
1664 if (db->db_blkid == DMU_BONUS_BLKID ||
1665 db->db_blkid == DMU_SPILL_BLKID) {
1666 mutex_enter(&dn->dn_mtx);
1667 ASSERT(!list_link_active(&dr->dr_dirty_node));
1668 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1669 mutex_exit(&dn->dn_mtx);
1670 dnode_setdirty(dn, tx);
1676 * The dn_struct_rwlock prevents db_blkptr from changing
1677 * due to a write from syncing context completing
1678 * while we are running, so we want to acquire it before
1679 * looking at db_blkptr.
1681 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1682 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1683 drop_struct_lock = TRUE;
1686 if (do_free_accounting) {
1687 blkptr_t *bp = db->db_blkptr;
1688 int64_t willfree = (bp && !BP_IS_HOLE(bp)) ?
1689 bp_get_dsize(os->os_spa, bp) : db->db.db_size;
1691 * This is only a guess -- if the dbuf is dirty
1692 * in a previous txg, we don't know how much
1693 * space it will use on disk yet. We should
1694 * really have the struct_rwlock to access
1695 * db_blkptr, but since this is just a guess,
1696 * it's OK if we get an odd answer.
1698 ddt_prefetch(os->os_spa, bp);
1699 dnode_willuse_space(dn, -willfree, tx);
1702 if (db->db_level == 0) {
1703 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1704 ASSERT(dn->dn_maxblkid >= db->db_blkid);
1707 if (db->db_level+1 < dn->dn_nlevels) {
1708 dmu_buf_impl_t *parent = db->db_parent;
1709 dbuf_dirty_record_t *di;
1710 int parent_held = FALSE;
1712 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1713 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1715 parent = dbuf_hold_level(dn, db->db_level+1,
1716 db->db_blkid >> epbs, FTAG);
1717 ASSERT(parent != NULL);
1720 if (drop_struct_lock)
1721 rw_exit(&dn->dn_struct_rwlock);
1722 ASSERT3U(db->db_level+1, ==, parent->db_level);
1723 di = dbuf_dirty(parent, tx);
1725 dbuf_rele(parent, FTAG);
1727 mutex_enter(&db->db_mtx);
1729 * Since we've dropped the mutex, it's possible that
1730 * dbuf_undirty() might have changed this out from under us.
1732 if (db->db_last_dirty == dr ||
1733 dn->dn_object == DMU_META_DNODE_OBJECT) {
1734 mutex_enter(&di->dt.di.dr_mtx);
1735 ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1736 ASSERT(!list_link_active(&dr->dr_dirty_node));
1737 list_insert_tail(&di->dt.di.dr_children, dr);
1738 mutex_exit(&di->dt.di.dr_mtx);
1741 mutex_exit(&db->db_mtx);
1743 ASSERT(db->db_level+1 == dn->dn_nlevels);
1744 ASSERT(db->db_blkid < dn->dn_nblkptr);
1745 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1746 mutex_enter(&dn->dn_mtx);
1747 ASSERT(!list_link_active(&dr->dr_dirty_node));
1748 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1749 mutex_exit(&dn->dn_mtx);
1750 if (drop_struct_lock)
1751 rw_exit(&dn->dn_struct_rwlock);
1754 dnode_setdirty(dn, tx);
1760 * Undirty a buffer in the transaction group referenced by the given
1761 * transaction. Return whether this evicted the dbuf.
1764 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1767 uint64_t txg = tx->tx_txg;
1768 dbuf_dirty_record_t *dr, **drp;
1773 * Due to our use of dn_nlevels below, this can only be called
1774 * in open context, unless we are operating on the MOS.
1775 * From syncing context, dn_nlevels may be different from the
1776 * dn_nlevels used when dbuf was dirtied.
1778 ASSERT(db->db_objset ==
1779 dmu_objset_pool(db->db_objset)->dp_meta_objset ||
1780 txg != spa_syncing_txg(dmu_objset_spa(db->db_objset)));
1781 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1782 ASSERT0(db->db_level);
1783 ASSERT(MUTEX_HELD(&db->db_mtx));
1786 * If this buffer is not dirty, we're done.
1788 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1789 if (dr->dr_txg <= txg)
1791 if (dr == NULL || dr->dr_txg < txg)
1793 ASSERT(dr->dr_txg == txg);
1794 ASSERT(dr->dr_dbuf == db);
1799 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1801 ASSERT(db->db.db_size != 0);
1803 dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset),
1804 dr->dr_accounted, txg);
1809 * Note that there are three places in dbuf_dirty()
1810 * where this dirty record may be put on a list.
1811 * Make sure to do a list_remove corresponding to
1812 * every one of those list_insert calls.
1814 if (dr->dr_parent) {
1815 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1816 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1817 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1818 } else if (db->db_blkid == DMU_SPILL_BLKID ||
1819 db->db_level + 1 == dn->dn_nlevels) {
1820 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1821 mutex_enter(&dn->dn_mtx);
1822 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1823 mutex_exit(&dn->dn_mtx);
1827 if (db->db_state != DB_NOFILL) {
1828 dbuf_unoverride(dr);
1830 ASSERT(db->db_buf != NULL);
1831 ASSERT(dr->dt.dl.dr_data != NULL);
1832 if (dr->dt.dl.dr_data != db->db_buf)
1833 arc_buf_destroy(dr->dt.dl.dr_data, db);
1836 kmem_free(dr, sizeof (dbuf_dirty_record_t));
1838 ASSERT(db->db_dirtycnt > 0);
1839 db->db_dirtycnt -= 1;
1841 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1842 ASSERT(db->db_state == DB_NOFILL || arc_released(db->db_buf));
1851 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
1853 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1854 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1856 ASSERT(tx->tx_txg != 0);
1857 ASSERT(!refcount_is_zero(&db->db_holds));
1860 * Quick check for dirtyness. For already dirty blocks, this
1861 * reduces runtime of this function by >90%, and overall performance
1862 * by 50% for some workloads (e.g. file deletion with indirect blocks
1865 mutex_enter(&db->db_mtx);
1866 dbuf_dirty_record_t *dr;
1867 for (dr = db->db_last_dirty;
1868 dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) {
1870 * It's possible that it is already dirty but not cached,
1871 * because there are some calls to dbuf_dirty() that don't
1872 * go through dmu_buf_will_dirty().
1874 if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) {
1875 /* This dbuf is already dirty and cached. */
1877 mutex_exit(&db->db_mtx);
1881 mutex_exit(&db->db_mtx);
1884 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
1885 rf |= DB_RF_HAVESTRUCT;
1887 (void) dbuf_read(db, NULL, rf);
1888 (void) dbuf_dirty(db, tx);
1892 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1894 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1896 db->db_state = DB_NOFILL;
1898 dmu_buf_will_fill(db_fake, tx);
1902 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1904 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1906 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1907 ASSERT(tx->tx_txg != 0);
1908 ASSERT(db->db_level == 0);
1909 ASSERT(!refcount_is_zero(&db->db_holds));
1911 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1912 dmu_tx_private_ok(tx));
1915 (void) dbuf_dirty(db, tx);
1918 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1921 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1923 mutex_enter(&db->db_mtx);
1926 if (db->db_state == DB_FILL) {
1927 if (db->db_level == 0 && db->db_freed_in_flight) {
1928 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1929 /* we were freed while filling */
1930 /* XXX dbuf_undirty? */
1931 bzero(db->db.db_data, db->db.db_size);
1932 db->db_freed_in_flight = FALSE;
1934 db->db_state = DB_CACHED;
1935 cv_broadcast(&db->db_changed);
1937 mutex_exit(&db->db_mtx);
1941 dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data,
1942 bp_embedded_type_t etype, enum zio_compress comp,
1943 int uncompressed_size, int compressed_size, int byteorder,
1946 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
1947 struct dirty_leaf *dl;
1948 dmu_object_type_t type;
1950 if (etype == BP_EMBEDDED_TYPE_DATA) {
1951 ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset),
1952 SPA_FEATURE_EMBEDDED_DATA));
1956 type = DB_DNODE(db)->dn_type;
1959 ASSERT0(db->db_level);
1960 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1962 dmu_buf_will_not_fill(dbuf, tx);
1964 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1965 dl = &db->db_last_dirty->dt.dl;
1966 encode_embedded_bp_compressed(&dl->dr_overridden_by,
1967 data, comp, uncompressed_size, compressed_size);
1968 BPE_SET_ETYPE(&dl->dr_overridden_by, etype);
1969 BP_SET_TYPE(&dl->dr_overridden_by, type);
1970 BP_SET_LEVEL(&dl->dr_overridden_by, 0);
1971 BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder);
1973 dl->dr_override_state = DR_OVERRIDDEN;
1974 dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg;
1978 * Directly assign a provided arc buf to a given dbuf if it's not referenced
1979 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1982 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
1984 ASSERT(!refcount_is_zero(&db->db_holds));
1985 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1986 ASSERT(db->db_level == 0);
1987 ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA);
1988 ASSERT(buf != NULL);
1989 ASSERT(arc_buf_size(buf) == db->db.db_size);
1990 ASSERT(tx->tx_txg != 0);
1992 arc_return_buf(buf, db);
1993 ASSERT(arc_released(buf));
1995 mutex_enter(&db->db_mtx);
1997 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1998 cv_wait(&db->db_changed, &db->db_mtx);
2000 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
2002 if (db->db_state == DB_CACHED &&
2003 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
2004 mutex_exit(&db->db_mtx);
2005 (void) dbuf_dirty(db, tx);
2006 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
2007 arc_buf_destroy(buf, db);
2008 xuio_stat_wbuf_copied();
2012 xuio_stat_wbuf_nocopy();
2013 if (db->db_state == DB_CACHED) {
2014 dbuf_dirty_record_t *dr = db->db_last_dirty;
2016 ASSERT(db->db_buf != NULL);
2017 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
2018 ASSERT(dr->dt.dl.dr_data == db->db_buf);
2019 if (!arc_released(db->db_buf)) {
2020 ASSERT(dr->dt.dl.dr_override_state ==
2022 arc_release(db->db_buf, db);
2024 dr->dt.dl.dr_data = buf;
2025 arc_buf_destroy(db->db_buf, db);
2026 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
2027 arc_release(db->db_buf, db);
2028 arc_buf_destroy(db->db_buf, db);
2032 ASSERT(db->db_buf == NULL);
2033 dbuf_set_data(db, buf);
2034 db->db_state = DB_FILL;
2035 mutex_exit(&db->db_mtx);
2036 (void) dbuf_dirty(db, tx);
2037 dmu_buf_fill_done(&db->db, tx);
2041 dbuf_destroy(dmu_buf_impl_t *db)
2044 dmu_buf_impl_t *parent = db->db_parent;
2045 dmu_buf_impl_t *dndb;
2047 ASSERT(MUTEX_HELD(&db->db_mtx));
2048 ASSERT(refcount_is_zero(&db->db_holds));
2050 if (db->db_buf != NULL) {
2051 arc_buf_destroy(db->db_buf, db);
2055 if (db->db_blkid == DMU_BONUS_BLKID) {
2056 ASSERT(db->db.db_data != NULL);
2057 zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
2058 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
2059 db->db_state = DB_UNCACHED;
2062 dbuf_clear_data(db);
2064 if (multilist_link_active(&db->db_cache_link)) {
2065 multilist_remove(&dbuf_cache, db);
2066 (void) refcount_remove_many(&dbuf_cache_size,
2067 db->db.db_size, db);
2070 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
2071 ASSERT(db->db_data_pending == NULL);
2073 db->db_state = DB_EVICTING;
2074 db->db_blkptr = NULL;
2077 * Now that db_state is DB_EVICTING, nobody else can find this via
2078 * the hash table. We can now drop db_mtx, which allows us to
2079 * acquire the dn_dbufs_mtx.
2081 mutex_exit(&db->db_mtx);
2086 if (db->db_blkid != DMU_BONUS_BLKID) {
2087 boolean_t needlock = !MUTEX_HELD(&dn->dn_dbufs_mtx);
2089 mutex_enter(&dn->dn_dbufs_mtx);
2090 avl_remove(&dn->dn_dbufs, db);
2091 atomic_dec_32(&dn->dn_dbufs_count);
2095 mutex_exit(&dn->dn_dbufs_mtx);
2097 * Decrementing the dbuf count means that the hold corresponding
2098 * to the removed dbuf is no longer discounted in dnode_move(),
2099 * so the dnode cannot be moved until after we release the hold.
2100 * The membar_producer() ensures visibility of the decremented
2101 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
2105 db->db_dnode_handle = NULL;
2107 dbuf_hash_remove(db);
2112 ASSERT(refcount_is_zero(&db->db_holds));
2114 db->db_parent = NULL;
2116 ASSERT(db->db_buf == NULL);
2117 ASSERT(db->db.db_data == NULL);
2118 ASSERT(db->db_hash_next == NULL);
2119 ASSERT(db->db_blkptr == NULL);
2120 ASSERT(db->db_data_pending == NULL);
2121 ASSERT(!multilist_link_active(&db->db_cache_link));
2123 kmem_cache_free(dbuf_kmem_cache, db);
2124 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2127 * If this dbuf is referenced from an indirect dbuf,
2128 * decrement the ref count on the indirect dbuf.
2130 if (parent && parent != dndb)
2131 dbuf_rele(parent, db);
2135 * Note: While bpp will always be updated if the function returns success,
2136 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
2137 * this happens when the dnode is the meta-dnode, or a userused or groupused
2141 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
2142 dmu_buf_impl_t **parentp, blkptr_t **bpp)
2149 ASSERT(blkid != DMU_BONUS_BLKID);
2151 if (blkid == DMU_SPILL_BLKID) {
2152 mutex_enter(&dn->dn_mtx);
2153 if (dn->dn_have_spill &&
2154 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
2155 *bpp = &dn->dn_phys->dn_spill;
2158 dbuf_add_ref(dn->dn_dbuf, NULL);
2159 *parentp = dn->dn_dbuf;
2160 mutex_exit(&dn->dn_mtx);
2164 if (dn->dn_phys->dn_nlevels == 0)
2167 nlevels = dn->dn_phys->dn_nlevels;
2169 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2171 ASSERT3U(level * epbs, <, 64);
2172 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2173 if (level >= nlevels ||
2174 (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
2175 /* the buffer has no parent yet */
2176 return (SET_ERROR(ENOENT));
2177 } else if (level < nlevels-1) {
2178 /* this block is referenced from an indirect block */
2179 int err = dbuf_hold_impl(dn, level+1,
2180 blkid >> epbs, fail_sparse, FALSE, NULL, parentp);
2183 err = dbuf_read(*parentp, NULL,
2184 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
2186 dbuf_rele(*parentp, NULL);
2190 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
2191 (blkid & ((1ULL << epbs) - 1));
2194 /* the block is referenced from the dnode */
2195 ASSERT3U(level, ==, nlevels-1);
2196 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
2197 blkid < dn->dn_phys->dn_nblkptr);
2199 dbuf_add_ref(dn->dn_dbuf, NULL);
2200 *parentp = dn->dn_dbuf;
2202 *bpp = &dn->dn_phys->dn_blkptr[blkid];
2207 static dmu_buf_impl_t *
2208 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
2209 dmu_buf_impl_t *parent, blkptr_t *blkptr)
2211 objset_t *os = dn->dn_objset;
2212 dmu_buf_impl_t *db, *odb;
2214 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2215 ASSERT(dn->dn_type != DMU_OT_NONE);
2217 db = kmem_cache_alloc(dbuf_kmem_cache, KM_SLEEP);
2220 db->db.db_object = dn->dn_object;
2221 db->db_level = level;
2222 db->db_blkid = blkid;
2223 db->db_last_dirty = NULL;
2224 db->db_dirtycnt = 0;
2225 db->db_dnode_handle = dn->dn_handle;
2226 db->db_parent = parent;
2227 db->db_blkptr = blkptr;
2230 db->db_user_immediate_evict = FALSE;
2231 db->db_freed_in_flight = FALSE;
2232 db->db_pending_evict = FALSE;
2234 if (blkid == DMU_BONUS_BLKID) {
2235 ASSERT3P(parent, ==, dn->dn_dbuf);
2236 db->db.db_size = DN_MAX_BONUSLEN -
2237 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
2238 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
2239 db->db.db_offset = DMU_BONUS_BLKID;
2240 db->db_state = DB_UNCACHED;
2241 /* the bonus dbuf is not placed in the hash table */
2242 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2244 } else if (blkid == DMU_SPILL_BLKID) {
2245 db->db.db_size = (blkptr != NULL) ?
2246 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
2247 db->db.db_offset = 0;
2250 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
2251 db->db.db_size = blocksize;
2252 db->db.db_offset = db->db_blkid * blocksize;
2256 * Hold the dn_dbufs_mtx while we get the new dbuf
2257 * in the hash table *and* added to the dbufs list.
2258 * This prevents a possible deadlock with someone
2259 * trying to look up this dbuf before its added to the
2262 mutex_enter(&dn->dn_dbufs_mtx);
2263 db->db_state = DB_EVICTING;
2264 if ((odb = dbuf_hash_insert(db)) != NULL) {
2265 /* someone else inserted it first */
2266 kmem_cache_free(dbuf_kmem_cache, db);
2267 mutex_exit(&dn->dn_dbufs_mtx);
2270 avl_add(&dn->dn_dbufs, db);
2272 db->db_state = DB_UNCACHED;
2273 mutex_exit(&dn->dn_dbufs_mtx);
2274 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2276 if (parent && parent != dn->dn_dbuf)
2277 dbuf_add_ref(parent, db);
2279 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
2280 refcount_count(&dn->dn_holds) > 0);
2281 (void) refcount_add(&dn->dn_holds, db);
2282 atomic_inc_32(&dn->dn_dbufs_count);
2284 dprintf_dbuf(db, "db=%p\n", db);
2289 typedef struct dbuf_prefetch_arg {
2290 spa_t *dpa_spa; /* The spa to issue the prefetch in. */
2291 zbookmark_phys_t dpa_zb; /* The target block to prefetch. */
2292 int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */
2293 int dpa_curlevel; /* The current level that we're reading */
2294 dnode_t *dpa_dnode; /* The dnode associated with the prefetch */
2295 zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */
2296 zio_t *dpa_zio; /* The parent zio_t for all prefetches. */
2297 arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */
2298 } dbuf_prefetch_arg_t;
2301 * Actually issue the prefetch read for the block given.
2304 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp)
2306 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
2309 arc_flags_t aflags =
2310 dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
2312 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2313 ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level);
2314 ASSERT(dpa->dpa_zio != NULL);
2315 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL,
2316 dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2317 &aflags, &dpa->dpa_zb);
2321 * Called when an indirect block above our prefetch target is read in. This
2322 * will either read in the next indirect block down the tree or issue the actual
2323 * prefetch if the next block down is our target.
2326 dbuf_prefetch_indirect_done(zio_t *zio, arc_buf_t *abuf, void *private)
2328 dbuf_prefetch_arg_t *dpa = private;
2330 ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel);
2331 ASSERT3S(dpa->dpa_curlevel, >, 0);
2334 * The dpa_dnode is only valid if we are called with a NULL
2335 * zio. This indicates that the arc_read() returned without
2336 * first calling zio_read() to issue a physical read. Once
2337 * a physical read is made the dpa_dnode must be invalidated
2338 * as the locks guarding it may have been dropped. If the
2339 * dpa_dnode is still valid, then we want to add it to the dbuf
2340 * cache. To do so, we must hold the dbuf associated with the block
2341 * we just prefetched, read its contents so that we associate it
2342 * with an arc_buf_t, and then release it.
2345 ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel);
2346 if (zio->io_flags & ZIO_FLAG_RAW) {
2347 ASSERT3U(BP_GET_PSIZE(zio->io_bp), ==, zio->io_size);
2349 ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size);
2351 ASSERT3P(zio->io_spa, ==, dpa->dpa_spa);
2353 dpa->dpa_dnode = NULL;
2354 } else if (dpa->dpa_dnode != NULL) {
2355 uint64_t curblkid = dpa->dpa_zb.zb_blkid >>
2356 (dpa->dpa_epbs * (dpa->dpa_curlevel -
2357 dpa->dpa_zb.zb_level));
2358 dmu_buf_impl_t *db = dbuf_hold_level(dpa->dpa_dnode,
2359 dpa->dpa_curlevel, curblkid, FTAG);
2360 (void) dbuf_read(db, NULL,
2361 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_HAVESTRUCT);
2362 dbuf_rele(db, FTAG);
2365 dpa->dpa_curlevel--;
2367 uint64_t nextblkid = dpa->dpa_zb.zb_blkid >>
2368 (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level));
2369 blkptr_t *bp = ((blkptr_t *)abuf->b_data) +
2370 P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs);
2371 if (BP_IS_HOLE(bp) || (zio != NULL && zio->io_error != 0)) {
2372 kmem_free(dpa, sizeof (*dpa));
2373 } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) {
2374 ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid);
2375 dbuf_issue_final_prefetch(dpa, bp);
2376 kmem_free(dpa, sizeof (*dpa));
2378 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2379 zbookmark_phys_t zb;
2381 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2383 SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset,
2384 dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid);
2386 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2387 bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio,
2388 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2392 arc_buf_destroy(abuf, private);
2396 * Issue prefetch reads for the given block on the given level. If the indirect
2397 * blocks above that block are not in memory, we will read them in
2398 * asynchronously. As a result, this call never blocks waiting for a read to
2402 dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio,
2406 int epbs, nlevels, curlevel;
2409 ASSERT(blkid != DMU_BONUS_BLKID);
2410 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2412 if (blkid > dn->dn_maxblkid)
2415 if (dnode_block_freed(dn, blkid))
2419 * This dnode hasn't been written to disk yet, so there's nothing to
2422 nlevels = dn->dn_phys->dn_nlevels;
2423 if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0)
2426 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2427 if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level))
2430 dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object,
2433 mutex_exit(&db->db_mtx);
2435 * This dbuf already exists. It is either CACHED, or
2436 * (we assume) about to be read or filled.
2442 * Find the closest ancestor (indirect block) of the target block
2443 * that is present in the cache. In this indirect block, we will
2444 * find the bp that is at curlevel, curblkid.
2448 while (curlevel < nlevels - 1) {
2449 int parent_level = curlevel + 1;
2450 uint64_t parent_blkid = curblkid >> epbs;
2453 if (dbuf_hold_impl(dn, parent_level, parent_blkid,
2454 FALSE, TRUE, FTAG, &db) == 0) {
2455 blkptr_t *bpp = db->db_buf->b_data;
2456 bp = bpp[P2PHASE(curblkid, 1 << epbs)];
2457 dbuf_rele(db, FTAG);
2461 curlevel = parent_level;
2462 curblkid = parent_blkid;
2465 if (curlevel == nlevels - 1) {
2466 /* No cached indirect blocks found. */
2467 ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr);
2468 bp = dn->dn_phys->dn_blkptr[curblkid];
2470 if (BP_IS_HOLE(&bp))
2473 ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp));
2475 zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL,
2478 dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP);
2479 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
2480 SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2481 dn->dn_object, level, blkid);
2482 dpa->dpa_curlevel = curlevel;
2483 dpa->dpa_prio = prio;
2484 dpa->dpa_aflags = aflags;
2485 dpa->dpa_spa = dn->dn_objset->os_spa;
2486 dpa->dpa_dnode = dn;
2487 dpa->dpa_epbs = epbs;
2491 * If we have the indirect just above us, no need to do the asynchronous
2492 * prefetch chain; we'll just run the last step ourselves. If we're at
2493 * a higher level, though, we want to issue the prefetches for all the
2494 * indirect blocks asynchronously, so we can go on with whatever we were
2497 if (curlevel == level) {
2498 ASSERT3U(curblkid, ==, blkid);
2499 dbuf_issue_final_prefetch(dpa, &bp);
2500 kmem_free(dpa, sizeof (*dpa));
2502 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2503 zbookmark_phys_t zb;
2505 SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2506 dn->dn_object, curlevel, curblkid);
2507 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2508 &bp, dbuf_prefetch_indirect_done, dpa, prio,
2509 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2513 * We use pio here instead of dpa_zio since it's possible that
2514 * dpa may have already been freed.
2520 * Returns with db_holds incremented, and db_mtx not held.
2521 * Note: dn_struct_rwlock must be held.
2524 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid,
2525 boolean_t fail_sparse, boolean_t fail_uncached,
2526 void *tag, dmu_buf_impl_t **dbp)
2528 dmu_buf_impl_t *db, *parent = NULL;
2530 ASSERT(blkid != DMU_BONUS_BLKID);
2531 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2532 ASSERT3U(dn->dn_nlevels, >, level);
2536 /* dbuf_find() returns with db_mtx held */
2537 db = dbuf_find(dn->dn_objset, dn->dn_object, level, blkid);
2540 blkptr_t *bp = NULL;
2544 return (SET_ERROR(ENOENT));
2546 ASSERT3P(parent, ==, NULL);
2547 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
2549 if (err == 0 && bp && BP_IS_HOLE(bp))
2550 err = SET_ERROR(ENOENT);
2553 dbuf_rele(parent, NULL);
2557 if (err && err != ENOENT)
2559 db = dbuf_create(dn, level, blkid, parent, bp);
2562 if (fail_uncached && db->db_state != DB_CACHED) {
2563 mutex_exit(&db->db_mtx);
2564 return (SET_ERROR(ENOENT));
2567 if (db->db_buf != NULL)
2568 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
2570 ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
2573 * If this buffer is currently syncing out, and we are are
2574 * still referencing it from db_data, we need to make a copy
2575 * of it in case we decide we want to dirty it again in this txg.
2577 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
2578 dn->dn_object != DMU_META_DNODE_OBJECT &&
2579 db->db_state == DB_CACHED && db->db_data_pending) {
2580 dbuf_dirty_record_t *dr = db->db_data_pending;
2582 if (dr->dt.dl.dr_data == db->db_buf) {
2583 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2586 arc_alloc_buf(dn->dn_objset->os_spa,
2587 db->db.db_size, db, type));
2588 bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
2593 if (multilist_link_active(&db->db_cache_link)) {
2594 ASSERT(refcount_is_zero(&db->db_holds));
2595 multilist_remove(&dbuf_cache, db);
2596 (void) refcount_remove_many(&dbuf_cache_size,
2597 db->db.db_size, db);
2599 (void) refcount_add(&db->db_holds, tag);
2601 mutex_exit(&db->db_mtx);
2603 /* NOTE: we can't rele the parent until after we drop the db_mtx */
2605 dbuf_rele(parent, NULL);
2607 ASSERT3P(DB_DNODE(db), ==, dn);
2608 ASSERT3U(db->db_blkid, ==, blkid);
2609 ASSERT3U(db->db_level, ==, level);
2616 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
2618 return (dbuf_hold_level(dn, 0, blkid, tag));
2622 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
2625 int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db);
2626 return (err ? NULL : db);
2630 dbuf_create_bonus(dnode_t *dn)
2632 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
2634 ASSERT(dn->dn_bonus == NULL);
2635 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
2639 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
2641 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2644 if (db->db_blkid != DMU_SPILL_BLKID)
2645 return (SET_ERROR(ENOTSUP));
2647 blksz = SPA_MINBLOCKSIZE;
2648 ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset)));
2649 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
2653 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2654 dbuf_new_size(db, blksz, tx);
2655 rw_exit(&dn->dn_struct_rwlock);
2662 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
2664 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
2667 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2669 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
2671 int64_t holds = refcount_add(&db->db_holds, tag);
2672 ASSERT3S(holds, >, 1);
2675 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
2677 dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid,
2680 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2681 dmu_buf_impl_t *found_db;
2682 boolean_t result = B_FALSE;
2684 if (db->db_blkid == DMU_BONUS_BLKID)
2685 found_db = dbuf_find_bonus(os, obj);
2687 found_db = dbuf_find(os, obj, 0, blkid);
2689 if (found_db != NULL) {
2690 if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) {
2691 (void) refcount_add(&db->db_holds, tag);
2694 mutex_exit(&db->db_mtx);
2700 * If you call dbuf_rele() you had better not be referencing the dnode handle
2701 * unless you have some other direct or indirect hold on the dnode. (An indirect
2702 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2703 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2704 * dnode's parent dbuf evicting its dnode handles.
2707 dbuf_rele(dmu_buf_impl_t *db, void *tag)
2709 mutex_enter(&db->db_mtx);
2710 dbuf_rele_and_unlock(db, tag);
2714 dmu_buf_rele(dmu_buf_t *db, void *tag)
2716 dbuf_rele((dmu_buf_impl_t *)db, tag);
2720 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2721 * db_dirtycnt and db_holds to be updated atomically.
2724 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2728 ASSERT(MUTEX_HELD(&db->db_mtx));
2732 * Remove the reference to the dbuf before removing its hold on the
2733 * dnode so we can guarantee in dnode_move() that a referenced bonus
2734 * buffer has a corresponding dnode hold.
2736 holds = refcount_remove(&db->db_holds, tag);
2740 * We can't freeze indirects if there is a possibility that they
2741 * may be modified in the current syncing context.
2743 if (db->db_buf != NULL &&
2744 holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) {
2745 arc_buf_freeze(db->db_buf);
2748 if (holds == db->db_dirtycnt &&
2749 db->db_level == 0 && db->db_user_immediate_evict)
2750 dbuf_evict_user(db);
2753 if (db->db_blkid == DMU_BONUS_BLKID) {
2755 boolean_t evict_dbuf = db->db_pending_evict;
2758 * If the dnode moves here, we cannot cross this
2759 * barrier until the move completes.
2764 atomic_dec_32(&dn->dn_dbufs_count);
2767 * Decrementing the dbuf count means that the bonus
2768 * buffer's dnode hold is no longer discounted in
2769 * dnode_move(). The dnode cannot move until after
2770 * the dnode_rele() below.
2775 * Do not reference db after its lock is dropped.
2776 * Another thread may evict it.
2778 mutex_exit(&db->db_mtx);
2781 dnode_evict_bonus(dn);
2784 } else if (db->db_buf == NULL) {
2786 * This is a special case: we never associated this
2787 * dbuf with any data allocated from the ARC.
2789 ASSERT(db->db_state == DB_UNCACHED ||
2790 db->db_state == DB_NOFILL);
2792 } else if (arc_released(db->db_buf)) {
2794 * This dbuf has anonymous data associated with it.
2798 boolean_t do_arc_evict = B_FALSE;
2800 spa_t *spa = dmu_objset_spa(db->db_objset);
2802 if (!DBUF_IS_CACHEABLE(db) &&
2803 db->db_blkptr != NULL &&
2804 !BP_IS_HOLE(db->db_blkptr) &&
2805 !BP_IS_EMBEDDED(db->db_blkptr)) {
2806 do_arc_evict = B_TRUE;
2807 bp = *db->db_blkptr;
2810 if (!DBUF_IS_CACHEABLE(db) ||
2811 db->db_pending_evict) {
2813 } else if (!multilist_link_active(&db->db_cache_link)) {
2814 multilist_insert(&dbuf_cache, db);
2815 (void) refcount_add_many(&dbuf_cache_size,
2816 db->db.db_size, db);
2817 mutex_exit(&db->db_mtx);
2819 dbuf_evict_notify();
2823 arc_freed(spa, &bp);
2826 mutex_exit(&db->db_mtx);
2831 #pragma weak dmu_buf_refcount = dbuf_refcount
2833 dbuf_refcount(dmu_buf_impl_t *db)
2835 return (refcount_count(&db->db_holds));
2839 dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user,
2840 dmu_buf_user_t *new_user)
2842 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2844 mutex_enter(&db->db_mtx);
2845 dbuf_verify_user(db, DBVU_NOT_EVICTING);
2846 if (db->db_user == old_user)
2847 db->db_user = new_user;
2849 old_user = db->db_user;
2850 dbuf_verify_user(db, DBVU_NOT_EVICTING);
2851 mutex_exit(&db->db_mtx);
2857 dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2859 return (dmu_buf_replace_user(db_fake, NULL, user));
2863 dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2865 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2867 db->db_user_immediate_evict = TRUE;
2868 return (dmu_buf_set_user(db_fake, user));
2872 dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2874 return (dmu_buf_replace_user(db_fake, user, NULL));
2878 dmu_buf_get_user(dmu_buf_t *db_fake)
2880 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2882 dbuf_verify_user(db, DBVU_NOT_EVICTING);
2883 return (db->db_user);
2887 dmu_buf_user_evict_wait()
2889 taskq_wait(dbu_evict_taskq);
2893 dmu_buf_freeable(dmu_buf_t *dbuf)
2895 boolean_t res = B_FALSE;
2896 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2899 res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset,
2900 db->db_blkptr, db->db_blkptr->blk_birth);
2906 dmu_buf_get_blkptr(dmu_buf_t *db)
2908 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2909 return (dbi->db_blkptr);
2913 dmu_buf_get_objset(dmu_buf_t *db)
2915 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2916 return (dbi->db_objset);
2920 dmu_buf_dnode_enter(dmu_buf_t *db)
2922 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2923 DB_DNODE_ENTER(dbi);
2924 return (DB_DNODE(dbi));
2928 dmu_buf_dnode_exit(dmu_buf_t *db)
2930 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2935 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
2937 /* ASSERT(dmu_tx_is_syncing(tx) */
2938 ASSERT(MUTEX_HELD(&db->db_mtx));
2940 if (db->db_blkptr != NULL)
2943 if (db->db_blkid == DMU_SPILL_BLKID) {
2944 db->db_blkptr = &dn->dn_phys->dn_spill;
2945 BP_ZERO(db->db_blkptr);
2948 if (db->db_level == dn->dn_phys->dn_nlevels-1) {
2950 * This buffer was allocated at a time when there was
2951 * no available blkptrs from the dnode, or it was
2952 * inappropriate to hook it in (i.e., nlevels mis-match).
2954 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
2955 ASSERT(db->db_parent == NULL);
2956 db->db_parent = dn->dn_dbuf;
2957 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
2960 dmu_buf_impl_t *parent = db->db_parent;
2961 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2963 ASSERT(dn->dn_phys->dn_nlevels > 1);
2964 if (parent == NULL) {
2965 mutex_exit(&db->db_mtx);
2966 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2967 parent = dbuf_hold_level(dn, db->db_level + 1,
2968 db->db_blkid >> epbs, db);
2969 rw_exit(&dn->dn_struct_rwlock);
2970 mutex_enter(&db->db_mtx);
2971 db->db_parent = parent;
2973 db->db_blkptr = (blkptr_t *)parent->db.db_data +
2974 (db->db_blkid & ((1ULL << epbs) - 1));
2980 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2982 dmu_buf_impl_t *db = dr->dr_dbuf;
2986 ASSERT(dmu_tx_is_syncing(tx));
2988 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2990 mutex_enter(&db->db_mtx);
2992 ASSERT(db->db_level > 0);
2995 /* Read the block if it hasn't been read yet. */
2996 if (db->db_buf == NULL) {
2997 mutex_exit(&db->db_mtx);
2998 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
2999 mutex_enter(&db->db_mtx);
3001 ASSERT3U(db->db_state, ==, DB_CACHED);
3002 ASSERT(db->db_buf != NULL);
3006 /* Indirect block size must match what the dnode thinks it is. */
3007 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3008 dbuf_check_blkptr(dn, db);
3011 /* Provide the pending dirty record to child dbufs */
3012 db->db_data_pending = dr;
3014 mutex_exit(&db->db_mtx);
3015 dbuf_write(dr, db->db_buf, tx);
3018 mutex_enter(&dr->dt.di.dr_mtx);
3019 dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx);
3020 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3021 mutex_exit(&dr->dt.di.dr_mtx);
3026 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
3028 arc_buf_t **datap = &dr->dt.dl.dr_data;
3029 dmu_buf_impl_t *db = dr->dr_dbuf;
3032 uint64_t txg = tx->tx_txg;
3034 ASSERT(dmu_tx_is_syncing(tx));
3036 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
3038 mutex_enter(&db->db_mtx);
3040 * To be synced, we must be dirtied. But we
3041 * might have been freed after the dirty.
3043 if (db->db_state == DB_UNCACHED) {
3044 /* This buffer has been freed since it was dirtied */
3045 ASSERT(db->db.db_data == NULL);
3046 } else if (db->db_state == DB_FILL) {
3047 /* This buffer was freed and is now being re-filled */
3048 ASSERT(db->db.db_data != dr->dt.dl.dr_data);
3050 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
3057 if (db->db_blkid == DMU_SPILL_BLKID) {
3058 mutex_enter(&dn->dn_mtx);
3059 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
3060 mutex_exit(&dn->dn_mtx);
3064 * If this is a bonus buffer, simply copy the bonus data into the
3065 * dnode. It will be written out when the dnode is synced (and it
3066 * will be synced, since it must have been dirty for dbuf_sync to
3069 if (db->db_blkid == DMU_BONUS_BLKID) {
3070 dbuf_dirty_record_t **drp;
3072 ASSERT(*datap != NULL);
3073 ASSERT0(db->db_level);
3074 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN);
3075 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen);
3078 if (*datap != db->db.db_data) {
3079 zio_buf_free(*datap, DN_MAX_BONUSLEN);
3080 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
3082 db->db_data_pending = NULL;
3083 drp = &db->db_last_dirty;
3085 drp = &(*drp)->dr_next;
3086 ASSERT(dr->dr_next == NULL);
3087 ASSERT(dr->dr_dbuf == db);
3089 if (dr->dr_dbuf->db_level != 0) {
3090 list_destroy(&dr->dt.di.dr_children);
3091 mutex_destroy(&dr->dt.di.dr_mtx);
3093 kmem_free(dr, sizeof (dbuf_dirty_record_t));
3094 ASSERT(db->db_dirtycnt > 0);
3095 db->db_dirtycnt -= 1;
3096 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
3103 * This function may have dropped the db_mtx lock allowing a dmu_sync
3104 * operation to sneak in. As a result, we need to ensure that we
3105 * don't check the dr_override_state until we have returned from
3106 * dbuf_check_blkptr.
3108 dbuf_check_blkptr(dn, db);
3111 * If this buffer is in the middle of an immediate write,
3112 * wait for the synchronous IO to complete.
3114 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
3115 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
3116 cv_wait(&db->db_changed, &db->db_mtx);
3117 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
3120 if (db->db_state != DB_NOFILL &&
3121 dn->dn_object != DMU_META_DNODE_OBJECT &&
3122 refcount_count(&db->db_holds) > 1 &&
3123 dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
3124 *datap == db->db_buf) {
3126 * If this buffer is currently "in use" (i.e., there
3127 * are active holds and db_data still references it),
3128 * then make a copy before we start the write so that
3129 * any modifications from the open txg will not leak
3132 * NOTE: this copy does not need to be made for
3133 * objects only modified in the syncing context (e.g.
3134 * DNONE_DNODE blocks).
3136 int blksz = arc_buf_size(*datap);
3137 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
3138 *datap = arc_alloc_buf(os->os_spa, blksz, db, type);
3139 bcopy(db->db.db_data, (*datap)->b_data, blksz);
3141 db->db_data_pending = dr;
3143 mutex_exit(&db->db_mtx);
3145 dbuf_write(dr, *datap, tx);
3147 ASSERT(!list_link_active(&dr->dr_dirty_node));
3148 if (dn->dn_object == DMU_META_DNODE_OBJECT) {
3149 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
3153 * Although zio_nowait() does not "wait for an IO", it does
3154 * initiate the IO. If this is an empty write it seems plausible
3155 * that the IO could actually be completed before the nowait
3156 * returns. We need to DB_DNODE_EXIT() first in case
3157 * zio_nowait() invalidates the dbuf.
3160 zio_nowait(dr->dr_zio);
3165 dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx)
3167 dbuf_dirty_record_t *dr;
3169 while (dr = list_head(list)) {
3170 if (dr->dr_zio != NULL) {
3172 * If we find an already initialized zio then we
3173 * are processing the meta-dnode, and we have finished.
3174 * The dbufs for all dnodes are put back on the list
3175 * during processing, so that we can zio_wait()
3176 * these IOs after initiating all child IOs.
3178 ASSERT3U(dr->dr_dbuf->db.db_object, ==,
3179 DMU_META_DNODE_OBJECT);
3182 if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
3183 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
3184 VERIFY3U(dr->dr_dbuf->db_level, ==, level);
3186 list_remove(list, dr);
3187 if (dr->dr_dbuf->db_level > 0)
3188 dbuf_sync_indirect(dr, tx);
3190 dbuf_sync_leaf(dr, tx);
3196 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3198 dmu_buf_impl_t *db = vdb;
3200 blkptr_t *bp = zio->io_bp;
3201 blkptr_t *bp_orig = &zio->io_bp_orig;
3202 spa_t *spa = zio->io_spa;
3207 ASSERT3P(db->db_blkptr, !=, NULL);
3208 ASSERT3P(&db->db_data_pending->dr_bp_copy, ==, bp);
3212 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
3213 dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
3214 zio->io_prev_space_delta = delta;
3216 if (bp->blk_birth != 0) {
3217 ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
3218 BP_GET_TYPE(bp) == dn->dn_type) ||
3219 (db->db_blkid == DMU_SPILL_BLKID &&
3220 BP_GET_TYPE(bp) == dn->dn_bonustype) ||
3221 BP_IS_EMBEDDED(bp));
3222 ASSERT(BP_GET_LEVEL(bp) == db->db_level);
3225 mutex_enter(&db->db_mtx);
3228 if (db->db_blkid == DMU_SPILL_BLKID) {
3229 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3230 ASSERT(!(BP_IS_HOLE(bp)) &&
3231 db->db_blkptr == &dn->dn_phys->dn_spill);
3235 if (db->db_level == 0) {
3236 mutex_enter(&dn->dn_mtx);
3237 if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
3238 db->db_blkid != DMU_SPILL_BLKID)
3239 dn->dn_phys->dn_maxblkid = db->db_blkid;
3240 mutex_exit(&dn->dn_mtx);
3242 if (dn->dn_type == DMU_OT_DNODE) {
3243 dnode_phys_t *dnp = db->db.db_data;
3244 for (i = db->db.db_size >> DNODE_SHIFT; i > 0;
3246 if (dnp->dn_type != DMU_OT_NONE)
3250 if (BP_IS_HOLE(bp)) {
3257 blkptr_t *ibp = db->db.db_data;
3258 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3259 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
3260 if (BP_IS_HOLE(ibp))
3262 fill += BP_GET_FILL(ibp);
3267 if (!BP_IS_EMBEDDED(bp))
3268 bp->blk_fill = fill;
3270 mutex_exit(&db->db_mtx);
3272 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3273 *db->db_blkptr = *bp;
3274 rw_exit(&dn->dn_struct_rwlock);
3279 * This function gets called just prior to running through the compression
3280 * stage of the zio pipeline. If we're an indirect block comprised of only
3281 * holes, then we want this indirect to be compressed away to a hole. In
3282 * order to do that we must zero out any information about the holes that
3283 * this indirect points to prior to before we try to compress it.
3286 dbuf_write_children_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3288 dmu_buf_impl_t *db = vdb;
3294 ASSERT3U(db->db_level, >, 0);
3297 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3299 /* Determine if all our children are holes */
3300 for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++) {
3301 if (!BP_IS_HOLE(bp))
3306 * If all the children are holes, then zero them all out so that
3307 * we may get compressed away.
3309 if (i == 1 << epbs) {
3310 /* didn't find any non-holes */
3311 bzero(db->db.db_data, db->db.db_size);
3317 * The SPA will call this callback several times for each zio - once
3318 * for every physical child i/o (zio->io_phys_children times). This
3319 * allows the DMU to monitor the progress of each logical i/o. For example,
3320 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
3321 * block. There may be a long delay before all copies/fragments are completed,
3322 * so this callback allows us to retire dirty space gradually, as the physical
3327 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
3329 dmu_buf_impl_t *db = arg;
3330 objset_t *os = db->db_objset;
3331 dsl_pool_t *dp = dmu_objset_pool(os);
3332 dbuf_dirty_record_t *dr;
3335 dr = db->db_data_pending;
3336 ASSERT3U(dr->dr_txg, ==, zio->io_txg);
3339 * The callback will be called io_phys_children times. Retire one
3340 * portion of our dirty space each time we are called. Any rounding
3341 * error will be cleaned up by dsl_pool_sync()'s call to
3342 * dsl_pool_undirty_space().
3344 delta = dr->dr_accounted / zio->io_phys_children;
3345 dsl_pool_undirty_space(dp, delta, zio->io_txg);
3350 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
3352 dmu_buf_impl_t *db = vdb;
3353 blkptr_t *bp_orig = &zio->io_bp_orig;
3354 blkptr_t *bp = db->db_blkptr;
3355 objset_t *os = db->db_objset;
3356 dmu_tx_t *tx = os->os_synctx;
3357 dbuf_dirty_record_t **drp, *dr;
3359 ASSERT0(zio->io_error);
3360 ASSERT(db->db_blkptr == bp);
3363 * For nopwrites and rewrites we ensure that the bp matches our
3364 * original and bypass all the accounting.
3366 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
3367 ASSERT(BP_EQUAL(bp, bp_orig));
3369 dsl_dataset_t *ds = os->os_dsl_dataset;
3370 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
3371 dsl_dataset_block_born(ds, bp, tx);
3374 mutex_enter(&db->db_mtx);
3378 drp = &db->db_last_dirty;
3379 while ((dr = *drp) != db->db_data_pending)
3381 ASSERT(!list_link_active(&dr->dr_dirty_node));
3382 ASSERT(dr->dr_dbuf == db);
3383 ASSERT(dr->dr_next == NULL);
3387 if (db->db_blkid == DMU_SPILL_BLKID) {
3392 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3393 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
3394 db->db_blkptr == &dn->dn_phys->dn_spill);
3399 if (db->db_level == 0) {
3400 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
3401 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
3402 if (db->db_state != DB_NOFILL) {
3403 if (dr->dt.dl.dr_data != db->db_buf)
3404 arc_buf_destroy(dr->dt.dl.dr_data, db);
3411 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3412 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
3413 if (!BP_IS_HOLE(db->db_blkptr)) {
3415 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3416 ASSERT3U(db->db_blkid, <=,
3417 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs));
3418 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
3422 mutex_destroy(&dr->dt.di.dr_mtx);
3423 list_destroy(&dr->dt.di.dr_children);
3425 kmem_free(dr, sizeof (dbuf_dirty_record_t));
3427 cv_broadcast(&db->db_changed);
3428 ASSERT(db->db_dirtycnt > 0);
3429 db->db_dirtycnt -= 1;
3430 db->db_data_pending = NULL;
3431 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg);
3435 dbuf_write_nofill_ready(zio_t *zio)
3437 dbuf_write_ready(zio, NULL, zio->io_private);
3441 dbuf_write_nofill_done(zio_t *zio)
3443 dbuf_write_done(zio, NULL, zio->io_private);
3447 dbuf_write_override_ready(zio_t *zio)
3449 dbuf_dirty_record_t *dr = zio->io_private;
3450 dmu_buf_impl_t *db = dr->dr_dbuf;
3452 dbuf_write_ready(zio, NULL, db);
3456 dbuf_write_override_done(zio_t *zio)
3458 dbuf_dirty_record_t *dr = zio->io_private;
3459 dmu_buf_impl_t *db = dr->dr_dbuf;
3460 blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
3462 mutex_enter(&db->db_mtx);
3463 if (!BP_EQUAL(zio->io_bp, obp)) {
3464 if (!BP_IS_HOLE(obp))
3465 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
3466 arc_release(dr->dt.dl.dr_data, db);
3468 mutex_exit(&db->db_mtx);
3470 dbuf_write_done(zio, NULL, db);
3473 /* Issue I/O to commit a dirty buffer to disk. */
3475 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
3477 dmu_buf_impl_t *db = dr->dr_dbuf;
3480 dmu_buf_impl_t *parent = db->db_parent;
3481 uint64_t txg = tx->tx_txg;
3482 zbookmark_phys_t zb;
3487 ASSERT(dmu_tx_is_syncing(tx));
3493 if (db->db_state != DB_NOFILL) {
3494 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
3496 * Private object buffers are released here rather
3497 * than in dbuf_dirty() since they are only modified
3498 * in the syncing context and we don't want the
3499 * overhead of making multiple copies of the data.
3501 if (BP_IS_HOLE(db->db_blkptr)) {
3504 dbuf_release_bp(db);
3509 if (parent != dn->dn_dbuf) {
3510 /* Our parent is an indirect block. */
3511 /* We have a dirty parent that has been scheduled for write. */
3512 ASSERT(parent && parent->db_data_pending);
3513 /* Our parent's buffer is one level closer to the dnode. */
3514 ASSERT(db->db_level == parent->db_level-1);
3516 * We're about to modify our parent's db_data by modifying
3517 * our block pointer, so the parent must be released.
3519 ASSERT(arc_released(parent->db_buf));
3520 zio = parent->db_data_pending->dr_zio;
3522 /* Our parent is the dnode itself. */
3523 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
3524 db->db_blkid != DMU_SPILL_BLKID) ||
3525 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
3526 if (db->db_blkid != DMU_SPILL_BLKID)
3527 ASSERT3P(db->db_blkptr, ==,
3528 &dn->dn_phys->dn_blkptr[db->db_blkid]);
3532 ASSERT(db->db_level == 0 || data == db->db_buf);
3533 ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
3536 SET_BOOKMARK(&zb, os->os_dsl_dataset ?
3537 os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
3538 db->db.db_object, db->db_level, db->db_blkid);
3540 if (db->db_blkid == DMU_SPILL_BLKID)
3542 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
3544 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
3548 * We copy the blkptr now (rather than when we instantiate the dirty
3549 * record), because its value can change between open context and
3550 * syncing context. We do not need to hold dn_struct_rwlock to read
3551 * db_blkptr because we are in syncing context.
3553 dr->dr_bp_copy = *db->db_blkptr;
3555 if (db->db_level == 0 &&
3556 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
3558 * The BP for this block has been provided by open context
3559 * (by dmu_sync() or dmu_buf_write_embedded()).
3561 void *contents = (data != NULL) ? data->b_data : NULL;
3563 dr->dr_zio = zio_write(zio, os->os_spa, txg,
3564 &dr->dr_bp_copy, contents, db->db.db_size, &zp,
3565 dbuf_write_override_ready, NULL, NULL,
3566 dbuf_write_override_done,
3567 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
3568 mutex_enter(&db->db_mtx);
3569 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
3570 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
3571 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
3572 mutex_exit(&db->db_mtx);
3573 } else if (db->db_state == DB_NOFILL) {
3574 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF ||
3575 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY);
3576 dr->dr_zio = zio_write(zio, os->os_spa, txg,
3577 &dr->dr_bp_copy, NULL, db->db.db_size, &zp,
3578 dbuf_write_nofill_ready, NULL, NULL,
3579 dbuf_write_nofill_done, db,
3580 ZIO_PRIORITY_ASYNC_WRITE,
3581 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
3583 ASSERT(arc_released(data));
3586 * For indirect blocks, we want to setup the children
3587 * ready callback so that we can properly handle an indirect
3588 * block that only contains holes.
3590 arc_done_func_t *children_ready_cb = NULL;
3591 if (db->db_level != 0)
3592 children_ready_cb = dbuf_write_children_ready;
3594 dr->dr_zio = arc_write(zio, os->os_spa, txg,
3595 &dr->dr_bp_copy, data, DBUF_IS_L2CACHEABLE(db),
3596 &zp, dbuf_write_ready, children_ready_cb,
3597 dbuf_write_physdone, dbuf_write_done, db,
3598 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);