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, B_FALSE, 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, 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 * This is our just-in-time copy function. It makes a copy of buffers that
1028 * have been modified in a previous transaction group before we access them in
1029 * the current active group.
1031 * This function is used in three places: when we are dirtying a buffer for the
1032 * first time in a txg, when we are freeing a range in a dnode that includes
1033 * this buffer, and when we are accessing a buffer which was received compressed
1034 * and later referenced in a WRITE_BYREF record.
1036 * Note that when we are called from dbuf_free_range() we do not put a hold on
1037 * the buffer, we just traverse the active dbuf list for the dnode.
1040 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
1042 dbuf_dirty_record_t *dr = db->db_last_dirty;
1044 ASSERT(MUTEX_HELD(&db->db_mtx));
1045 ASSERT(db->db.db_data != NULL);
1046 ASSERT(db->db_level == 0);
1047 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
1050 (dr->dt.dl.dr_data !=
1051 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
1055 * If the last dirty record for this dbuf has not yet synced
1056 * and its referencing the dbuf data, either:
1057 * reset the reference to point to a new copy,
1058 * or (if there a no active holders)
1059 * just null out the current db_data pointer.
1061 ASSERT(dr->dr_txg >= txg - 2);
1062 if (db->db_blkid == DMU_BONUS_BLKID) {
1063 /* Note that the data bufs here are zio_bufs */
1064 dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
1065 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
1066 bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
1067 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
1068 int size = arc_buf_size(db->db_buf);
1069 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1070 spa_t *spa = db->db_objset->os_spa;
1071 enum zio_compress compress_type =
1072 arc_get_compression(db->db_buf);
1074 if (compress_type == ZIO_COMPRESS_OFF) {
1075 dr->dt.dl.dr_data = arc_alloc_buf(spa, db, type, size);
1077 ASSERT3U(type, ==, ARC_BUFC_DATA);
1078 dr->dt.dl.dr_data = arc_alloc_compressed_buf(spa, db,
1079 size, arc_buf_lsize(db->db_buf), compress_type);
1081 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
1084 dbuf_clear_data(db);
1089 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
1092 boolean_t havepzio = (zio != NULL);
1097 * We don't have to hold the mutex to check db_state because it
1098 * can't be freed while we have a hold on the buffer.
1100 ASSERT(!refcount_is_zero(&db->db_holds));
1102 if (db->db_state == DB_NOFILL)
1103 return (SET_ERROR(EIO));
1107 if ((flags & DB_RF_HAVESTRUCT) == 0)
1108 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1110 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1111 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
1112 DBUF_IS_CACHEABLE(db);
1114 mutex_enter(&db->db_mtx);
1115 if (db->db_state == DB_CACHED) {
1117 * If the arc buf is compressed, we need to decompress it to
1118 * read the data. This could happen during the "zfs receive" of
1119 * a stream which is compressed and deduplicated.
1121 if (db->db_buf != NULL &&
1122 arc_get_compression(db->db_buf) != ZIO_COMPRESS_OFF) {
1123 dbuf_fix_old_data(db,
1124 spa_syncing_txg(dmu_objset_spa(db->db_objset)));
1125 err = arc_decompress(db->db_buf);
1126 dbuf_set_data(db, db->db_buf);
1128 mutex_exit(&db->db_mtx);
1130 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1131 if ((flags & DB_RF_HAVESTRUCT) == 0)
1132 rw_exit(&dn->dn_struct_rwlock);
1134 } else if (db->db_state == DB_UNCACHED) {
1135 spa_t *spa = dn->dn_objset->os_spa;
1138 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
1139 dbuf_read_impl(db, zio, flags);
1141 /* dbuf_read_impl has dropped db_mtx for us */
1144 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1146 if ((flags & DB_RF_HAVESTRUCT) == 0)
1147 rw_exit(&dn->dn_struct_rwlock);
1151 err = zio_wait(zio);
1154 * Another reader came in while the dbuf was in flight
1155 * between UNCACHED and CACHED. Either a writer will finish
1156 * writing the buffer (sending the dbuf to CACHED) or the
1157 * first reader's request will reach the read_done callback
1158 * and send the dbuf to CACHED. Otherwise, a failure
1159 * occurred and the dbuf went to UNCACHED.
1161 mutex_exit(&db->db_mtx);
1163 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1164 if ((flags & DB_RF_HAVESTRUCT) == 0)
1165 rw_exit(&dn->dn_struct_rwlock);
1168 /* Skip the wait per the caller's request. */
1169 mutex_enter(&db->db_mtx);
1170 if ((flags & DB_RF_NEVERWAIT) == 0) {
1171 while (db->db_state == DB_READ ||
1172 db->db_state == DB_FILL) {
1173 ASSERT(db->db_state == DB_READ ||
1174 (flags & DB_RF_HAVESTRUCT) == 0);
1175 DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *,
1177 cv_wait(&db->db_changed, &db->db_mtx);
1179 if (db->db_state == DB_UNCACHED)
1180 err = SET_ERROR(EIO);
1182 mutex_exit(&db->db_mtx);
1185 ASSERT(err || havepzio || db->db_state == DB_CACHED);
1190 dbuf_noread(dmu_buf_impl_t *db)
1192 ASSERT(!refcount_is_zero(&db->db_holds));
1193 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1194 mutex_enter(&db->db_mtx);
1195 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1196 cv_wait(&db->db_changed, &db->db_mtx);
1197 if (db->db_state == DB_UNCACHED) {
1198 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1199 spa_t *spa = db->db_objset->os_spa;
1201 ASSERT(db->db_buf == NULL);
1202 ASSERT(db->db.db_data == NULL);
1203 dbuf_set_data(db, arc_alloc_buf(spa, db, type, db->db.db_size));
1204 db->db_state = DB_FILL;
1205 } else if (db->db_state == DB_NOFILL) {
1206 dbuf_clear_data(db);
1208 ASSERT3U(db->db_state, ==, DB_CACHED);
1210 mutex_exit(&db->db_mtx);
1214 dbuf_unoverride(dbuf_dirty_record_t *dr)
1216 dmu_buf_impl_t *db = dr->dr_dbuf;
1217 blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
1218 uint64_t txg = dr->dr_txg;
1220 ASSERT(MUTEX_HELD(&db->db_mtx));
1221 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
1222 ASSERT(db->db_level == 0);
1224 if (db->db_blkid == DMU_BONUS_BLKID ||
1225 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
1228 ASSERT(db->db_data_pending != dr);
1230 /* free this block */
1231 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite)
1232 zio_free(db->db_objset->os_spa, txg, bp);
1234 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1235 dr->dt.dl.dr_nopwrite = B_FALSE;
1238 * Release the already-written buffer, so we leave it in
1239 * a consistent dirty state. Note that all callers are
1240 * modifying the buffer, so they will immediately do
1241 * another (redundant) arc_release(). Therefore, leave
1242 * the buf thawed to save the effort of freezing &
1243 * immediately re-thawing it.
1245 arc_release(dr->dt.dl.dr_data, db);
1249 * Evict (if its unreferenced) or clear (if its referenced) any level-0
1250 * data blocks in the free range, so that any future readers will find
1254 dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
1257 dmu_buf_impl_t db_search;
1258 dmu_buf_impl_t *db, *db_next;
1259 uint64_t txg = tx->tx_txg;
1262 if (end_blkid > dn->dn_maxblkid &&
1263 !(start_blkid == DMU_SPILL_BLKID || end_blkid == DMU_SPILL_BLKID))
1264 end_blkid = dn->dn_maxblkid;
1265 dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid);
1267 db_search.db_level = 0;
1268 db_search.db_blkid = start_blkid;
1269 db_search.db_state = DB_SEARCH;
1271 mutex_enter(&dn->dn_dbufs_mtx);
1272 db = avl_find(&dn->dn_dbufs, &db_search, &where);
1273 ASSERT3P(db, ==, NULL);
1275 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1277 for (; db != NULL; db = db_next) {
1278 db_next = AVL_NEXT(&dn->dn_dbufs, db);
1279 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1281 if (db->db_level != 0 || db->db_blkid > end_blkid) {
1284 ASSERT3U(db->db_blkid, >=, start_blkid);
1286 /* found a level 0 buffer in the range */
1287 mutex_enter(&db->db_mtx);
1288 if (dbuf_undirty(db, tx)) {
1289 /* mutex has been dropped and dbuf destroyed */
1293 if (db->db_state == DB_UNCACHED ||
1294 db->db_state == DB_NOFILL ||
1295 db->db_state == DB_EVICTING) {
1296 ASSERT(db->db.db_data == NULL);
1297 mutex_exit(&db->db_mtx);
1300 if (db->db_state == DB_READ || db->db_state == DB_FILL) {
1301 /* will be handled in dbuf_read_done or dbuf_rele */
1302 db->db_freed_in_flight = TRUE;
1303 mutex_exit(&db->db_mtx);
1306 if (refcount_count(&db->db_holds) == 0) {
1311 /* The dbuf is referenced */
1313 if (db->db_last_dirty != NULL) {
1314 dbuf_dirty_record_t *dr = db->db_last_dirty;
1316 if (dr->dr_txg == txg) {
1318 * This buffer is "in-use", re-adjust the file
1319 * size to reflect that this buffer may
1320 * contain new data when we sync.
1322 if (db->db_blkid != DMU_SPILL_BLKID &&
1323 db->db_blkid > dn->dn_maxblkid)
1324 dn->dn_maxblkid = db->db_blkid;
1325 dbuf_unoverride(dr);
1328 * This dbuf is not dirty in the open context.
1329 * Either uncache it (if its not referenced in
1330 * the open context) or reset its contents to
1333 dbuf_fix_old_data(db, txg);
1336 /* clear the contents if its cached */
1337 if (db->db_state == DB_CACHED) {
1338 ASSERT(db->db.db_data != NULL);
1339 arc_release(db->db_buf, db);
1340 bzero(db->db.db_data, db->db.db_size);
1341 arc_buf_freeze(db->db_buf);
1344 mutex_exit(&db->db_mtx);
1346 mutex_exit(&dn->dn_dbufs_mtx);
1350 dbuf_block_freeable(dmu_buf_impl_t *db)
1352 dsl_dataset_t *ds = db->db_objset->os_dsl_dataset;
1353 uint64_t birth_txg = 0;
1356 * We don't need any locking to protect db_blkptr:
1357 * If it's syncing, then db_last_dirty will be set
1358 * so we'll ignore db_blkptr.
1360 * This logic ensures that only block births for
1361 * filled blocks are considered.
1363 ASSERT(MUTEX_HELD(&db->db_mtx));
1364 if (db->db_last_dirty && (db->db_blkptr == NULL ||
1365 !BP_IS_HOLE(db->db_blkptr))) {
1366 birth_txg = db->db_last_dirty->dr_txg;
1367 } else if (db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) {
1368 birth_txg = db->db_blkptr->blk_birth;
1372 * If this block don't exist or is in a snapshot, it can't be freed.
1373 * Don't pass the bp to dsl_dataset_block_freeable() since we
1374 * are holding the db_mtx lock and might deadlock if we are
1375 * prefetching a dedup-ed block.
1378 return (ds == NULL ||
1379 dsl_dataset_block_freeable(ds, NULL, birth_txg));
1385 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
1387 arc_buf_t *buf, *obuf;
1388 int osize = db->db.db_size;
1389 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1392 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1397 /* XXX does *this* func really need the lock? */
1398 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1401 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1402 * is OK, because there can be no other references to the db
1403 * when we are changing its size, so no concurrent DB_FILL can
1407 * XXX we should be doing a dbuf_read, checking the return
1408 * value and returning that up to our callers
1410 dmu_buf_will_dirty(&db->db, tx);
1412 /* create the data buffer for the new block */
1413 buf = arc_alloc_buf(dn->dn_objset->os_spa, db, type, size);
1415 /* copy old block data to the new block */
1417 bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
1418 /* zero the remainder */
1420 bzero((uint8_t *)buf->b_data + osize, size - osize);
1422 mutex_enter(&db->db_mtx);
1423 dbuf_set_data(db, buf);
1424 arc_buf_destroy(obuf, db);
1425 db->db.db_size = size;
1427 if (db->db_level == 0) {
1428 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1429 db->db_last_dirty->dt.dl.dr_data = buf;
1431 mutex_exit(&db->db_mtx);
1433 dnode_willuse_space(dn, size-osize, tx);
1438 dbuf_release_bp(dmu_buf_impl_t *db)
1440 objset_t *os = db->db_objset;
1442 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1443 ASSERT(arc_released(os->os_phys_buf) ||
1444 list_link_active(&os->os_dsl_dataset->ds_synced_link));
1445 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1447 (void) arc_release(db->db_buf, db);
1451 * We already have a dirty record for this TXG, and we are being
1455 dbuf_redirty(dbuf_dirty_record_t *dr)
1457 dmu_buf_impl_t *db = dr->dr_dbuf;
1459 ASSERT(MUTEX_HELD(&db->db_mtx));
1461 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1463 * If this buffer has already been written out,
1464 * we now need to reset its state.
1466 dbuf_unoverride(dr);
1467 if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1468 db->db_state != DB_NOFILL) {
1469 /* Already released on initial dirty, so just thaw. */
1470 ASSERT(arc_released(db->db_buf));
1471 arc_buf_thaw(db->db_buf);
1476 dbuf_dirty_record_t *
1477 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1481 dbuf_dirty_record_t **drp, *dr;
1482 int drop_struct_lock = FALSE;
1483 boolean_t do_free_accounting = B_FALSE;
1484 int txgoff = tx->tx_txg & TXG_MASK;
1486 ASSERT(tx->tx_txg != 0);
1487 ASSERT(!refcount_is_zero(&db->db_holds));
1488 DMU_TX_DIRTY_BUF(tx, db);
1493 * Shouldn't dirty a regular buffer in syncing context. Private
1494 * objects may be dirtied in syncing context, but only if they
1495 * were already pre-dirtied in open context.
1498 if (dn->dn_objset->os_dsl_dataset != NULL) {
1499 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1502 ASSERT(!dmu_tx_is_syncing(tx) ||
1503 BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1504 DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1505 dn->dn_objset->os_dsl_dataset == NULL);
1506 if (dn->dn_objset->os_dsl_dataset != NULL)
1507 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, FTAG);
1510 * We make this assert for private objects as well, but after we
1511 * check if we're already dirty. They are allowed to re-dirty
1512 * in syncing context.
1514 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1515 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1516 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1518 mutex_enter(&db->db_mtx);
1520 * XXX make this true for indirects too? The problem is that
1521 * transactions created with dmu_tx_create_assigned() from
1522 * syncing context don't bother holding ahead.
1524 ASSERT(db->db_level != 0 ||
1525 db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1526 db->db_state == DB_NOFILL);
1528 mutex_enter(&dn->dn_mtx);
1530 * Don't set dirtyctx to SYNC if we're just modifying this as we
1531 * initialize the objset.
1533 if (dn->dn_dirtyctx == DN_UNDIRTIED) {
1534 if (dn->dn_objset->os_dsl_dataset != NULL) {
1535 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1538 if (!BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1539 dn->dn_dirtyctx = (dmu_tx_is_syncing(tx) ?
1540 DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1541 ASSERT(dn->dn_dirtyctx_firstset == NULL);
1542 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1544 if (dn->dn_objset->os_dsl_dataset != NULL) {
1545 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1549 mutex_exit(&dn->dn_mtx);
1551 if (db->db_blkid == DMU_SPILL_BLKID)
1552 dn->dn_have_spill = B_TRUE;
1555 * If this buffer is already dirty, we're done.
1557 drp = &db->db_last_dirty;
1558 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1559 db->db.db_object == DMU_META_DNODE_OBJECT);
1560 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1562 if (dr && dr->dr_txg == tx->tx_txg) {
1566 mutex_exit(&db->db_mtx);
1571 * Only valid if not already dirty.
1573 ASSERT(dn->dn_object == 0 ||
1574 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1575 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1577 ASSERT3U(dn->dn_nlevels, >, db->db_level);
1578 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1579 dn->dn_phys->dn_nlevels > db->db_level ||
1580 dn->dn_next_nlevels[txgoff] > db->db_level ||
1581 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1582 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1585 * We should only be dirtying in syncing context if it's the
1586 * mos or we're initializing the os or it's a special object.
1587 * However, we are allowed to dirty in syncing context provided
1588 * we already dirtied it in open context. Hence we must make
1589 * this assertion only if we're not already dirty.
1593 if (dn->dn_objset->os_dsl_dataset != NULL)
1594 rrw_enter(&os->os_dsl_dataset->ds_bp_rwlock, RW_READER, FTAG);
1595 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1596 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1597 if (dn->dn_objset->os_dsl_dataset != NULL)
1598 rrw_exit(&os->os_dsl_dataset->ds_bp_rwlock, FTAG);
1600 ASSERT(db->db.db_size != 0);
1602 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1604 if (db->db_blkid != DMU_BONUS_BLKID) {
1606 * Update the accounting.
1607 * Note: we delay "free accounting" until after we drop
1608 * the db_mtx. This keeps us from grabbing other locks
1609 * (and possibly deadlocking) in bp_get_dsize() while
1610 * also holding the db_mtx.
1612 dnode_willuse_space(dn, db->db.db_size, tx);
1613 do_free_accounting = dbuf_block_freeable(db);
1617 * If this buffer is dirty in an old transaction group we need
1618 * to make a copy of it so that the changes we make in this
1619 * transaction group won't leak out when we sync the older txg.
1621 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1622 if (db->db_level == 0) {
1623 void *data_old = db->db_buf;
1625 if (db->db_state != DB_NOFILL) {
1626 if (db->db_blkid == DMU_BONUS_BLKID) {
1627 dbuf_fix_old_data(db, tx->tx_txg);
1628 data_old = db->db.db_data;
1629 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1631 * Release the data buffer from the cache so
1632 * that we can modify it without impacting
1633 * possible other users of this cached data
1634 * block. Note that indirect blocks and
1635 * private objects are not released until the
1636 * syncing state (since they are only modified
1639 arc_release(db->db_buf, db);
1640 dbuf_fix_old_data(db, tx->tx_txg);
1641 data_old = db->db_buf;
1643 ASSERT(data_old != NULL);
1645 dr->dt.dl.dr_data = data_old;
1647 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1648 list_create(&dr->dt.di.dr_children,
1649 sizeof (dbuf_dirty_record_t),
1650 offsetof(dbuf_dirty_record_t, dr_dirty_node));
1652 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL)
1653 dr->dr_accounted = db->db.db_size;
1655 dr->dr_txg = tx->tx_txg;
1660 * We could have been freed_in_flight between the dbuf_noread
1661 * and dbuf_dirty. We win, as though the dbuf_noread() had
1662 * happened after the free.
1664 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1665 db->db_blkid != DMU_SPILL_BLKID) {
1666 mutex_enter(&dn->dn_mtx);
1667 if (dn->dn_free_ranges[txgoff] != NULL) {
1668 range_tree_clear(dn->dn_free_ranges[txgoff],
1671 mutex_exit(&dn->dn_mtx);
1672 db->db_freed_in_flight = FALSE;
1676 * This buffer is now part of this txg
1678 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1679 db->db_dirtycnt += 1;
1680 ASSERT3U(db->db_dirtycnt, <=, 3);
1682 mutex_exit(&db->db_mtx);
1684 if (db->db_blkid == DMU_BONUS_BLKID ||
1685 db->db_blkid == DMU_SPILL_BLKID) {
1686 mutex_enter(&dn->dn_mtx);
1687 ASSERT(!list_link_active(&dr->dr_dirty_node));
1688 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1689 mutex_exit(&dn->dn_mtx);
1690 dnode_setdirty(dn, tx);
1696 * The dn_struct_rwlock prevents db_blkptr from changing
1697 * due to a write from syncing context completing
1698 * while we are running, so we want to acquire it before
1699 * looking at db_blkptr.
1701 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1702 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1703 drop_struct_lock = TRUE;
1706 if (do_free_accounting) {
1707 blkptr_t *bp = db->db_blkptr;
1708 int64_t willfree = (bp && !BP_IS_HOLE(bp)) ?
1709 bp_get_dsize(os->os_spa, bp) : db->db.db_size;
1711 * This is only a guess -- if the dbuf is dirty
1712 * in a previous txg, we don't know how much
1713 * space it will use on disk yet. We should
1714 * really have the struct_rwlock to access
1715 * db_blkptr, but since this is just a guess,
1716 * it's OK if we get an odd answer.
1718 ddt_prefetch(os->os_spa, bp);
1719 dnode_willuse_space(dn, -willfree, tx);
1722 if (db->db_level == 0) {
1723 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1724 ASSERT(dn->dn_maxblkid >= db->db_blkid);
1727 if (db->db_level+1 < dn->dn_nlevels) {
1728 dmu_buf_impl_t *parent = db->db_parent;
1729 dbuf_dirty_record_t *di;
1730 int parent_held = FALSE;
1732 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1733 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1735 parent = dbuf_hold_level(dn, db->db_level+1,
1736 db->db_blkid >> epbs, FTAG);
1737 ASSERT(parent != NULL);
1740 if (drop_struct_lock)
1741 rw_exit(&dn->dn_struct_rwlock);
1742 ASSERT3U(db->db_level+1, ==, parent->db_level);
1743 di = dbuf_dirty(parent, tx);
1745 dbuf_rele(parent, FTAG);
1747 mutex_enter(&db->db_mtx);
1749 * Since we've dropped the mutex, it's possible that
1750 * dbuf_undirty() might have changed this out from under us.
1752 if (db->db_last_dirty == dr ||
1753 dn->dn_object == DMU_META_DNODE_OBJECT) {
1754 mutex_enter(&di->dt.di.dr_mtx);
1755 ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1756 ASSERT(!list_link_active(&dr->dr_dirty_node));
1757 list_insert_tail(&di->dt.di.dr_children, dr);
1758 mutex_exit(&di->dt.di.dr_mtx);
1761 mutex_exit(&db->db_mtx);
1763 ASSERT(db->db_level+1 == dn->dn_nlevels);
1764 ASSERT(db->db_blkid < dn->dn_nblkptr);
1765 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1766 mutex_enter(&dn->dn_mtx);
1767 ASSERT(!list_link_active(&dr->dr_dirty_node));
1768 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1769 mutex_exit(&dn->dn_mtx);
1770 if (drop_struct_lock)
1771 rw_exit(&dn->dn_struct_rwlock);
1774 dnode_setdirty(dn, tx);
1780 * Undirty a buffer in the transaction group referenced by the given
1781 * transaction. Return whether this evicted the dbuf.
1784 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1787 uint64_t txg = tx->tx_txg;
1788 dbuf_dirty_record_t *dr, **drp;
1793 * Due to our use of dn_nlevels below, this can only be called
1794 * in open context, unless we are operating on the MOS.
1795 * From syncing context, dn_nlevels may be different from the
1796 * dn_nlevels used when dbuf was dirtied.
1798 ASSERT(db->db_objset ==
1799 dmu_objset_pool(db->db_objset)->dp_meta_objset ||
1800 txg != spa_syncing_txg(dmu_objset_spa(db->db_objset)));
1801 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1802 ASSERT0(db->db_level);
1803 ASSERT(MUTEX_HELD(&db->db_mtx));
1806 * If this buffer is not dirty, we're done.
1808 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1809 if (dr->dr_txg <= txg)
1811 if (dr == NULL || dr->dr_txg < txg)
1813 ASSERT(dr->dr_txg == txg);
1814 ASSERT(dr->dr_dbuf == db);
1819 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1821 ASSERT(db->db.db_size != 0);
1823 dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset),
1824 dr->dr_accounted, txg);
1829 * Note that there are three places in dbuf_dirty()
1830 * where this dirty record may be put on a list.
1831 * Make sure to do a list_remove corresponding to
1832 * every one of those list_insert calls.
1834 if (dr->dr_parent) {
1835 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1836 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1837 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1838 } else if (db->db_blkid == DMU_SPILL_BLKID ||
1839 db->db_level + 1 == dn->dn_nlevels) {
1840 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1841 mutex_enter(&dn->dn_mtx);
1842 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1843 mutex_exit(&dn->dn_mtx);
1847 if (db->db_state != DB_NOFILL) {
1848 dbuf_unoverride(dr);
1850 ASSERT(db->db_buf != NULL);
1851 ASSERT(dr->dt.dl.dr_data != NULL);
1852 if (dr->dt.dl.dr_data != db->db_buf)
1853 arc_buf_destroy(dr->dt.dl.dr_data, db);
1856 kmem_free(dr, sizeof (dbuf_dirty_record_t));
1858 ASSERT(db->db_dirtycnt > 0);
1859 db->db_dirtycnt -= 1;
1861 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1862 ASSERT(db->db_state == DB_NOFILL || arc_released(db->db_buf));
1871 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
1873 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1874 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1876 ASSERT(tx->tx_txg != 0);
1877 ASSERT(!refcount_is_zero(&db->db_holds));
1880 * Quick check for dirtyness. For already dirty blocks, this
1881 * reduces runtime of this function by >90%, and overall performance
1882 * by 50% for some workloads (e.g. file deletion with indirect blocks
1885 mutex_enter(&db->db_mtx);
1886 dbuf_dirty_record_t *dr;
1887 for (dr = db->db_last_dirty;
1888 dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) {
1890 * It's possible that it is already dirty but not cached,
1891 * because there are some calls to dbuf_dirty() that don't
1892 * go through dmu_buf_will_dirty().
1894 if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) {
1895 /* This dbuf is already dirty and cached. */
1897 mutex_exit(&db->db_mtx);
1901 mutex_exit(&db->db_mtx);
1904 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
1905 rf |= DB_RF_HAVESTRUCT;
1907 (void) dbuf_read(db, NULL, rf);
1908 (void) dbuf_dirty(db, tx);
1912 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1914 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1916 db->db_state = DB_NOFILL;
1918 dmu_buf_will_fill(db_fake, tx);
1922 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1924 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1926 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1927 ASSERT(tx->tx_txg != 0);
1928 ASSERT(db->db_level == 0);
1929 ASSERT(!refcount_is_zero(&db->db_holds));
1931 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1932 dmu_tx_private_ok(tx));
1935 (void) dbuf_dirty(db, tx);
1938 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1941 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1943 mutex_enter(&db->db_mtx);
1946 if (db->db_state == DB_FILL) {
1947 if (db->db_level == 0 && db->db_freed_in_flight) {
1948 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1949 /* we were freed while filling */
1950 /* XXX dbuf_undirty? */
1951 bzero(db->db.db_data, db->db.db_size);
1952 db->db_freed_in_flight = FALSE;
1954 db->db_state = DB_CACHED;
1955 cv_broadcast(&db->db_changed);
1957 mutex_exit(&db->db_mtx);
1961 dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data,
1962 bp_embedded_type_t etype, enum zio_compress comp,
1963 int uncompressed_size, int compressed_size, int byteorder,
1966 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
1967 struct dirty_leaf *dl;
1968 dmu_object_type_t type;
1970 if (etype == BP_EMBEDDED_TYPE_DATA) {
1971 ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset),
1972 SPA_FEATURE_EMBEDDED_DATA));
1976 type = DB_DNODE(db)->dn_type;
1979 ASSERT0(db->db_level);
1980 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1982 dmu_buf_will_not_fill(dbuf, tx);
1984 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1985 dl = &db->db_last_dirty->dt.dl;
1986 encode_embedded_bp_compressed(&dl->dr_overridden_by,
1987 data, comp, uncompressed_size, compressed_size);
1988 BPE_SET_ETYPE(&dl->dr_overridden_by, etype);
1989 BP_SET_TYPE(&dl->dr_overridden_by, type);
1990 BP_SET_LEVEL(&dl->dr_overridden_by, 0);
1991 BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder);
1993 dl->dr_override_state = DR_OVERRIDDEN;
1994 dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg;
1998 * Directly assign a provided arc buf to a given dbuf if it's not referenced
1999 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
2002 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
2004 ASSERT(!refcount_is_zero(&db->db_holds));
2005 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2006 ASSERT(db->db_level == 0);
2007 ASSERT3U(dbuf_is_metadata(db), ==, arc_is_metadata(buf));
2008 ASSERT(buf != NULL);
2009 ASSERT(arc_buf_lsize(buf) == db->db.db_size);
2010 ASSERT(tx->tx_txg != 0);
2012 arc_return_buf(buf, db);
2013 ASSERT(arc_released(buf));
2015 mutex_enter(&db->db_mtx);
2017 while (db->db_state == DB_READ || db->db_state == DB_FILL)
2018 cv_wait(&db->db_changed, &db->db_mtx);
2020 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
2022 if (db->db_state == DB_CACHED &&
2023 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
2024 mutex_exit(&db->db_mtx);
2025 (void) dbuf_dirty(db, tx);
2026 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
2027 arc_buf_destroy(buf, db);
2028 xuio_stat_wbuf_copied();
2032 xuio_stat_wbuf_nocopy();
2033 if (db->db_state == DB_CACHED) {
2034 dbuf_dirty_record_t *dr = db->db_last_dirty;
2036 ASSERT(db->db_buf != NULL);
2037 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
2038 ASSERT(dr->dt.dl.dr_data == db->db_buf);
2039 if (!arc_released(db->db_buf)) {
2040 ASSERT(dr->dt.dl.dr_override_state ==
2042 arc_release(db->db_buf, db);
2044 dr->dt.dl.dr_data = buf;
2045 arc_buf_destroy(db->db_buf, db);
2046 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
2047 arc_release(db->db_buf, db);
2048 arc_buf_destroy(db->db_buf, db);
2052 ASSERT(db->db_buf == NULL);
2053 dbuf_set_data(db, buf);
2054 db->db_state = DB_FILL;
2055 mutex_exit(&db->db_mtx);
2056 (void) dbuf_dirty(db, tx);
2057 dmu_buf_fill_done(&db->db, tx);
2061 dbuf_destroy(dmu_buf_impl_t *db)
2064 dmu_buf_impl_t *parent = db->db_parent;
2065 dmu_buf_impl_t *dndb;
2067 ASSERT(MUTEX_HELD(&db->db_mtx));
2068 ASSERT(refcount_is_zero(&db->db_holds));
2070 if (db->db_buf != NULL) {
2071 arc_buf_destroy(db->db_buf, db);
2075 if (db->db_blkid == DMU_BONUS_BLKID) {
2076 ASSERT(db->db.db_data != NULL);
2077 zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
2078 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
2079 db->db_state = DB_UNCACHED;
2082 dbuf_clear_data(db);
2084 if (multilist_link_active(&db->db_cache_link)) {
2085 multilist_remove(&dbuf_cache, db);
2086 (void) refcount_remove_many(&dbuf_cache_size,
2087 db->db.db_size, db);
2090 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
2091 ASSERT(db->db_data_pending == NULL);
2093 db->db_state = DB_EVICTING;
2094 db->db_blkptr = NULL;
2097 * Now that db_state is DB_EVICTING, nobody else can find this via
2098 * the hash table. We can now drop db_mtx, which allows us to
2099 * acquire the dn_dbufs_mtx.
2101 mutex_exit(&db->db_mtx);
2106 if (db->db_blkid != DMU_BONUS_BLKID) {
2107 boolean_t needlock = !MUTEX_HELD(&dn->dn_dbufs_mtx);
2109 mutex_enter(&dn->dn_dbufs_mtx);
2110 avl_remove(&dn->dn_dbufs, db);
2111 atomic_dec_32(&dn->dn_dbufs_count);
2115 mutex_exit(&dn->dn_dbufs_mtx);
2117 * Decrementing the dbuf count means that the hold corresponding
2118 * to the removed dbuf is no longer discounted in dnode_move(),
2119 * so the dnode cannot be moved until after we release the hold.
2120 * The membar_producer() ensures visibility of the decremented
2121 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
2125 db->db_dnode_handle = NULL;
2127 dbuf_hash_remove(db);
2132 ASSERT(refcount_is_zero(&db->db_holds));
2134 db->db_parent = NULL;
2136 ASSERT(db->db_buf == NULL);
2137 ASSERT(db->db.db_data == NULL);
2138 ASSERT(db->db_hash_next == NULL);
2139 ASSERT(db->db_blkptr == NULL);
2140 ASSERT(db->db_data_pending == NULL);
2141 ASSERT(!multilist_link_active(&db->db_cache_link));
2143 kmem_cache_free(dbuf_kmem_cache, db);
2144 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2147 * If this dbuf is referenced from an indirect dbuf,
2148 * decrement the ref count on the indirect dbuf.
2150 if (parent && parent != dndb)
2151 dbuf_rele(parent, db);
2155 * Note: While bpp will always be updated if the function returns success,
2156 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
2157 * this happens when the dnode is the meta-dnode, or a userused or groupused
2161 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
2162 dmu_buf_impl_t **parentp, blkptr_t **bpp)
2167 ASSERT(blkid != DMU_BONUS_BLKID);
2169 if (blkid == DMU_SPILL_BLKID) {
2170 mutex_enter(&dn->dn_mtx);
2171 if (dn->dn_have_spill &&
2172 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
2173 *bpp = &dn->dn_phys->dn_spill;
2176 dbuf_add_ref(dn->dn_dbuf, NULL);
2177 *parentp = dn->dn_dbuf;
2178 mutex_exit(&dn->dn_mtx);
2183 (dn->dn_phys->dn_nlevels == 0) ? 1 : dn->dn_phys->dn_nlevels;
2184 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2186 ASSERT3U(level * epbs, <, 64);
2187 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2189 * This assertion shouldn't trip as long as the max indirect block size
2190 * is less than 1M. The reason for this is that up to that point,
2191 * the number of levels required to address an entire object with blocks
2192 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In
2193 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55
2194 * (i.e. we can address the entire object), objects will all use at most
2195 * N-1 levels and the assertion won't overflow. However, once epbs is
2196 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be
2197 * enough to address an entire object, so objects will have 5 levels,
2198 * but then this assertion will overflow.
2200 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we
2201 * need to redo this logic to handle overflows.
2203 ASSERT(level >= nlevels ||
2204 ((nlevels - level - 1) * epbs) +
2205 highbit64(dn->dn_phys->dn_nblkptr) <= 64);
2206 if (level >= nlevels ||
2207 blkid >= ((uint64_t)dn->dn_phys->dn_nblkptr <<
2208 ((nlevels - level - 1) * epbs)) ||
2210 blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
2211 /* the buffer has no parent yet */
2212 return (SET_ERROR(ENOENT));
2213 } else if (level < nlevels-1) {
2214 /* this block is referenced from an indirect block */
2215 int err = dbuf_hold_impl(dn, level+1,
2216 blkid >> epbs, fail_sparse, FALSE, NULL, parentp);
2219 err = dbuf_read(*parentp, NULL,
2220 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
2222 dbuf_rele(*parentp, NULL);
2226 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
2227 (blkid & ((1ULL << epbs) - 1));
2228 if (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))
2229 ASSERT(BP_IS_HOLE(*bpp));
2232 /* the block is referenced from the dnode */
2233 ASSERT3U(level, ==, nlevels-1);
2234 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
2235 blkid < dn->dn_phys->dn_nblkptr);
2237 dbuf_add_ref(dn->dn_dbuf, NULL);
2238 *parentp = dn->dn_dbuf;
2240 *bpp = &dn->dn_phys->dn_blkptr[blkid];
2245 static dmu_buf_impl_t *
2246 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
2247 dmu_buf_impl_t *parent, blkptr_t *blkptr)
2249 objset_t *os = dn->dn_objset;
2250 dmu_buf_impl_t *db, *odb;
2252 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2253 ASSERT(dn->dn_type != DMU_OT_NONE);
2255 db = kmem_cache_alloc(dbuf_kmem_cache, KM_SLEEP);
2258 db->db.db_object = dn->dn_object;
2259 db->db_level = level;
2260 db->db_blkid = blkid;
2261 db->db_last_dirty = NULL;
2262 db->db_dirtycnt = 0;
2263 db->db_dnode_handle = dn->dn_handle;
2264 db->db_parent = parent;
2265 db->db_blkptr = blkptr;
2268 db->db_user_immediate_evict = FALSE;
2269 db->db_freed_in_flight = FALSE;
2270 db->db_pending_evict = FALSE;
2272 if (blkid == DMU_BONUS_BLKID) {
2273 ASSERT3P(parent, ==, dn->dn_dbuf);
2274 db->db.db_size = DN_MAX_BONUSLEN -
2275 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
2276 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
2277 db->db.db_offset = DMU_BONUS_BLKID;
2278 db->db_state = DB_UNCACHED;
2279 /* the bonus dbuf is not placed in the hash table */
2280 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2282 } else if (blkid == DMU_SPILL_BLKID) {
2283 db->db.db_size = (blkptr != NULL) ?
2284 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
2285 db->db.db_offset = 0;
2288 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
2289 db->db.db_size = blocksize;
2290 db->db.db_offset = db->db_blkid * blocksize;
2294 * Hold the dn_dbufs_mtx while we get the new dbuf
2295 * in the hash table *and* added to the dbufs list.
2296 * This prevents a possible deadlock with someone
2297 * trying to look up this dbuf before its added to the
2300 mutex_enter(&dn->dn_dbufs_mtx);
2301 db->db_state = DB_EVICTING;
2302 if ((odb = dbuf_hash_insert(db)) != NULL) {
2303 /* someone else inserted it first */
2304 kmem_cache_free(dbuf_kmem_cache, db);
2305 mutex_exit(&dn->dn_dbufs_mtx);
2308 avl_add(&dn->dn_dbufs, db);
2310 db->db_state = DB_UNCACHED;
2311 mutex_exit(&dn->dn_dbufs_mtx);
2312 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2314 if (parent && parent != dn->dn_dbuf)
2315 dbuf_add_ref(parent, db);
2317 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
2318 refcount_count(&dn->dn_holds) > 0);
2319 (void) refcount_add(&dn->dn_holds, db);
2320 atomic_inc_32(&dn->dn_dbufs_count);
2322 dprintf_dbuf(db, "db=%p\n", db);
2327 typedef struct dbuf_prefetch_arg {
2328 spa_t *dpa_spa; /* The spa to issue the prefetch in. */
2329 zbookmark_phys_t dpa_zb; /* The target block to prefetch. */
2330 int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */
2331 int dpa_curlevel; /* The current level that we're reading */
2332 dnode_t *dpa_dnode; /* The dnode associated with the prefetch */
2333 zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */
2334 zio_t *dpa_zio; /* The parent zio_t for all prefetches. */
2335 arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */
2336 } dbuf_prefetch_arg_t;
2339 * Actually issue the prefetch read for the block given.
2342 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp)
2344 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
2347 arc_flags_t aflags =
2348 dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
2350 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2351 ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level);
2352 ASSERT(dpa->dpa_zio != NULL);
2353 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL,
2354 dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2355 &aflags, &dpa->dpa_zb);
2359 * Called when an indirect block above our prefetch target is read in. This
2360 * will either read in the next indirect block down the tree or issue the actual
2361 * prefetch if the next block down is our target.
2364 dbuf_prefetch_indirect_done(zio_t *zio, arc_buf_t *abuf, void *private)
2366 dbuf_prefetch_arg_t *dpa = private;
2368 ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel);
2369 ASSERT3S(dpa->dpa_curlevel, >, 0);
2372 * The dpa_dnode is only valid if we are called with a NULL
2373 * zio. This indicates that the arc_read() returned without
2374 * first calling zio_read() to issue a physical read. Once
2375 * a physical read is made the dpa_dnode must be invalidated
2376 * as the locks guarding it may have been dropped. If the
2377 * dpa_dnode is still valid, then we want to add it to the dbuf
2378 * cache. To do so, we must hold the dbuf associated with the block
2379 * we just prefetched, read its contents so that we associate it
2380 * with an arc_buf_t, and then release it.
2383 ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel);
2384 if (zio->io_flags & ZIO_FLAG_RAW) {
2385 ASSERT3U(BP_GET_PSIZE(zio->io_bp), ==, zio->io_size);
2387 ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size);
2389 ASSERT3P(zio->io_spa, ==, dpa->dpa_spa);
2391 dpa->dpa_dnode = NULL;
2392 } else if (dpa->dpa_dnode != NULL) {
2393 uint64_t curblkid = dpa->dpa_zb.zb_blkid >>
2394 (dpa->dpa_epbs * (dpa->dpa_curlevel -
2395 dpa->dpa_zb.zb_level));
2396 dmu_buf_impl_t *db = dbuf_hold_level(dpa->dpa_dnode,
2397 dpa->dpa_curlevel, curblkid, FTAG);
2398 (void) dbuf_read(db, NULL,
2399 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_HAVESTRUCT);
2400 dbuf_rele(db, FTAG);
2403 dpa->dpa_curlevel--;
2405 uint64_t nextblkid = dpa->dpa_zb.zb_blkid >>
2406 (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level));
2407 blkptr_t *bp = ((blkptr_t *)abuf->b_data) +
2408 P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs);
2409 if (BP_IS_HOLE(bp) || (zio != NULL && zio->io_error != 0)) {
2410 kmem_free(dpa, sizeof (*dpa));
2411 } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) {
2412 ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid);
2413 dbuf_issue_final_prefetch(dpa, bp);
2414 kmem_free(dpa, sizeof (*dpa));
2416 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2417 zbookmark_phys_t zb;
2419 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2421 SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset,
2422 dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid);
2424 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2425 bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio,
2426 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2430 arc_buf_destroy(abuf, private);
2434 * Issue prefetch reads for the given block on the given level. If the indirect
2435 * blocks above that block are not in memory, we will read them in
2436 * asynchronously. As a result, this call never blocks waiting for a read to
2440 dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio,
2444 int epbs, nlevels, curlevel;
2447 ASSERT(blkid != DMU_BONUS_BLKID);
2448 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2450 if (blkid > dn->dn_maxblkid)
2453 if (dnode_block_freed(dn, blkid))
2457 * This dnode hasn't been written to disk yet, so there's nothing to
2460 nlevels = dn->dn_phys->dn_nlevels;
2461 if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0)
2464 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2465 if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level))
2468 dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object,
2471 mutex_exit(&db->db_mtx);
2473 * This dbuf already exists. It is either CACHED, or
2474 * (we assume) about to be read or filled.
2480 * Find the closest ancestor (indirect block) of the target block
2481 * that is present in the cache. In this indirect block, we will
2482 * find the bp that is at curlevel, curblkid.
2486 while (curlevel < nlevels - 1) {
2487 int parent_level = curlevel + 1;
2488 uint64_t parent_blkid = curblkid >> epbs;
2491 if (dbuf_hold_impl(dn, parent_level, parent_blkid,
2492 FALSE, TRUE, FTAG, &db) == 0) {
2493 blkptr_t *bpp = db->db_buf->b_data;
2494 bp = bpp[P2PHASE(curblkid, 1 << epbs)];
2495 dbuf_rele(db, FTAG);
2499 curlevel = parent_level;
2500 curblkid = parent_blkid;
2503 if (curlevel == nlevels - 1) {
2504 /* No cached indirect blocks found. */
2505 ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr);
2506 bp = dn->dn_phys->dn_blkptr[curblkid];
2508 if (BP_IS_HOLE(&bp))
2511 ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp));
2513 zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL,
2516 dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP);
2517 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
2518 SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2519 dn->dn_object, level, blkid);
2520 dpa->dpa_curlevel = curlevel;
2521 dpa->dpa_prio = prio;
2522 dpa->dpa_aflags = aflags;
2523 dpa->dpa_spa = dn->dn_objset->os_spa;
2524 dpa->dpa_dnode = dn;
2525 dpa->dpa_epbs = epbs;
2529 * If we have the indirect just above us, no need to do the asynchronous
2530 * prefetch chain; we'll just run the last step ourselves. If we're at
2531 * a higher level, though, we want to issue the prefetches for all the
2532 * indirect blocks asynchronously, so we can go on with whatever we were
2535 if (curlevel == level) {
2536 ASSERT3U(curblkid, ==, blkid);
2537 dbuf_issue_final_prefetch(dpa, &bp);
2538 kmem_free(dpa, sizeof (*dpa));
2540 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2541 zbookmark_phys_t zb;
2543 SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2544 dn->dn_object, curlevel, curblkid);
2545 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2546 &bp, dbuf_prefetch_indirect_done, dpa, prio,
2547 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2551 * We use pio here instead of dpa_zio since it's possible that
2552 * dpa may have already been freed.
2558 * Returns with db_holds incremented, and db_mtx not held.
2559 * Note: dn_struct_rwlock must be held.
2562 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid,
2563 boolean_t fail_sparse, boolean_t fail_uncached,
2564 void *tag, dmu_buf_impl_t **dbp)
2566 dmu_buf_impl_t *db, *parent = NULL;
2568 ASSERT(blkid != DMU_BONUS_BLKID);
2569 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2570 ASSERT3U(dn->dn_nlevels, >, level);
2574 /* dbuf_find() returns with db_mtx held */
2575 db = dbuf_find(dn->dn_objset, dn->dn_object, level, blkid);
2578 blkptr_t *bp = NULL;
2582 return (SET_ERROR(ENOENT));
2584 ASSERT3P(parent, ==, NULL);
2585 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
2587 if (err == 0 && bp && BP_IS_HOLE(bp))
2588 err = SET_ERROR(ENOENT);
2591 dbuf_rele(parent, NULL);
2595 if (err && err != ENOENT)
2597 db = dbuf_create(dn, level, blkid, parent, bp);
2600 if (fail_uncached && db->db_state != DB_CACHED) {
2601 mutex_exit(&db->db_mtx);
2602 return (SET_ERROR(ENOENT));
2605 if (db->db_buf != NULL)
2606 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
2608 ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
2611 * If this buffer is currently syncing out, and we are are
2612 * still referencing it from db_data, we need to make a copy
2613 * of it in case we decide we want to dirty it again in this txg.
2615 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
2616 dn->dn_object != DMU_META_DNODE_OBJECT &&
2617 db->db_state == DB_CACHED && db->db_data_pending) {
2618 dbuf_dirty_record_t *dr = db->db_data_pending;
2620 if (dr->dt.dl.dr_data == db->db_buf) {
2621 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2624 arc_alloc_buf(dn->dn_objset->os_spa, db, type,
2626 bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
2631 if (multilist_link_active(&db->db_cache_link)) {
2632 ASSERT(refcount_is_zero(&db->db_holds));
2633 multilist_remove(&dbuf_cache, db);
2634 (void) refcount_remove_many(&dbuf_cache_size,
2635 db->db.db_size, db);
2637 (void) refcount_add(&db->db_holds, tag);
2639 mutex_exit(&db->db_mtx);
2641 /* NOTE: we can't rele the parent until after we drop the db_mtx */
2643 dbuf_rele(parent, NULL);
2645 ASSERT3P(DB_DNODE(db), ==, dn);
2646 ASSERT3U(db->db_blkid, ==, blkid);
2647 ASSERT3U(db->db_level, ==, level);
2654 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
2656 return (dbuf_hold_level(dn, 0, blkid, tag));
2660 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
2663 int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db);
2664 return (err ? NULL : db);
2668 dbuf_create_bonus(dnode_t *dn)
2670 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
2672 ASSERT(dn->dn_bonus == NULL);
2673 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
2677 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
2679 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2682 if (db->db_blkid != DMU_SPILL_BLKID)
2683 return (SET_ERROR(ENOTSUP));
2685 blksz = SPA_MINBLOCKSIZE;
2686 ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset)));
2687 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
2691 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2692 dbuf_new_size(db, blksz, tx);
2693 rw_exit(&dn->dn_struct_rwlock);
2700 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
2702 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
2705 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2707 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
2709 int64_t holds = refcount_add(&db->db_holds, tag);
2710 ASSERT3S(holds, >, 1);
2713 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
2715 dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid,
2718 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2719 dmu_buf_impl_t *found_db;
2720 boolean_t result = B_FALSE;
2722 if (db->db_blkid == DMU_BONUS_BLKID)
2723 found_db = dbuf_find_bonus(os, obj);
2725 found_db = dbuf_find(os, obj, 0, blkid);
2727 if (found_db != NULL) {
2728 if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) {
2729 (void) refcount_add(&db->db_holds, tag);
2732 mutex_exit(&db->db_mtx);
2738 * If you call dbuf_rele() you had better not be referencing the dnode handle
2739 * unless you have some other direct or indirect hold on the dnode. (An indirect
2740 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2741 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2742 * dnode's parent dbuf evicting its dnode handles.
2745 dbuf_rele(dmu_buf_impl_t *db, void *tag)
2747 mutex_enter(&db->db_mtx);
2748 dbuf_rele_and_unlock(db, tag);
2752 dmu_buf_rele(dmu_buf_t *db, void *tag)
2754 dbuf_rele((dmu_buf_impl_t *)db, tag);
2758 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2759 * db_dirtycnt and db_holds to be updated atomically.
2762 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2766 ASSERT(MUTEX_HELD(&db->db_mtx));
2770 * Remove the reference to the dbuf before removing its hold on the
2771 * dnode so we can guarantee in dnode_move() that a referenced bonus
2772 * buffer has a corresponding dnode hold.
2774 holds = refcount_remove(&db->db_holds, tag);
2778 * We can't freeze indirects if there is a possibility that they
2779 * may be modified in the current syncing context.
2781 if (db->db_buf != NULL &&
2782 holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) {
2783 arc_buf_freeze(db->db_buf);
2786 if (holds == db->db_dirtycnt &&
2787 db->db_level == 0 && db->db_user_immediate_evict)
2788 dbuf_evict_user(db);
2791 if (db->db_blkid == DMU_BONUS_BLKID) {
2793 boolean_t evict_dbuf = db->db_pending_evict;
2796 * If the dnode moves here, we cannot cross this
2797 * barrier until the move completes.
2802 atomic_dec_32(&dn->dn_dbufs_count);
2805 * Decrementing the dbuf count means that the bonus
2806 * buffer's dnode hold is no longer discounted in
2807 * dnode_move(). The dnode cannot move until after
2808 * the dnode_rele() below.
2813 * Do not reference db after its lock is dropped.
2814 * Another thread may evict it.
2816 mutex_exit(&db->db_mtx);
2819 dnode_evict_bonus(dn);
2822 } else if (db->db_buf == NULL) {
2824 * This is a special case: we never associated this
2825 * dbuf with any data allocated from the ARC.
2827 ASSERT(db->db_state == DB_UNCACHED ||
2828 db->db_state == DB_NOFILL);
2830 } else if (arc_released(db->db_buf)) {
2832 * This dbuf has anonymous data associated with it.
2836 boolean_t do_arc_evict = B_FALSE;
2838 spa_t *spa = dmu_objset_spa(db->db_objset);
2840 if (!DBUF_IS_CACHEABLE(db) &&
2841 db->db_blkptr != NULL &&
2842 !BP_IS_HOLE(db->db_blkptr) &&
2843 !BP_IS_EMBEDDED(db->db_blkptr)) {
2844 do_arc_evict = B_TRUE;
2845 bp = *db->db_blkptr;
2848 if (!DBUF_IS_CACHEABLE(db) ||
2849 db->db_pending_evict) {
2851 } else if (!multilist_link_active(&db->db_cache_link)) {
2852 multilist_insert(&dbuf_cache, db);
2853 (void) refcount_add_many(&dbuf_cache_size,
2854 db->db.db_size, db);
2855 mutex_exit(&db->db_mtx);
2857 dbuf_evict_notify();
2861 arc_freed(spa, &bp);
2864 mutex_exit(&db->db_mtx);
2869 #pragma weak dmu_buf_refcount = dbuf_refcount
2871 dbuf_refcount(dmu_buf_impl_t *db)
2873 return (refcount_count(&db->db_holds));
2877 dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user,
2878 dmu_buf_user_t *new_user)
2880 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2882 mutex_enter(&db->db_mtx);
2883 dbuf_verify_user(db, DBVU_NOT_EVICTING);
2884 if (db->db_user == old_user)
2885 db->db_user = new_user;
2887 old_user = db->db_user;
2888 dbuf_verify_user(db, DBVU_NOT_EVICTING);
2889 mutex_exit(&db->db_mtx);
2895 dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2897 return (dmu_buf_replace_user(db_fake, NULL, user));
2901 dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2903 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2905 db->db_user_immediate_evict = TRUE;
2906 return (dmu_buf_set_user(db_fake, user));
2910 dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2912 return (dmu_buf_replace_user(db_fake, user, NULL));
2916 dmu_buf_get_user(dmu_buf_t *db_fake)
2918 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2920 dbuf_verify_user(db, DBVU_NOT_EVICTING);
2921 return (db->db_user);
2925 dmu_buf_user_evict_wait()
2927 taskq_wait(dbu_evict_taskq);
2931 dmu_buf_freeable(dmu_buf_t *dbuf)
2933 boolean_t res = B_FALSE;
2934 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2937 res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset,
2938 db->db_blkptr, db->db_blkptr->blk_birth);
2944 dmu_buf_get_blkptr(dmu_buf_t *db)
2946 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2947 return (dbi->db_blkptr);
2951 dmu_buf_get_objset(dmu_buf_t *db)
2953 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2954 return (dbi->db_objset);
2958 dmu_buf_dnode_enter(dmu_buf_t *db)
2960 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2961 DB_DNODE_ENTER(dbi);
2962 return (DB_DNODE(dbi));
2966 dmu_buf_dnode_exit(dmu_buf_t *db)
2968 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2973 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
2975 /* ASSERT(dmu_tx_is_syncing(tx) */
2976 ASSERT(MUTEX_HELD(&db->db_mtx));
2978 if (db->db_blkptr != NULL)
2981 if (db->db_blkid == DMU_SPILL_BLKID) {
2982 db->db_blkptr = &dn->dn_phys->dn_spill;
2983 BP_ZERO(db->db_blkptr);
2986 if (db->db_level == dn->dn_phys->dn_nlevels-1) {
2988 * This buffer was allocated at a time when there was
2989 * no available blkptrs from the dnode, or it was
2990 * inappropriate to hook it in (i.e., nlevels mis-match).
2992 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
2993 ASSERT(db->db_parent == NULL);
2994 db->db_parent = dn->dn_dbuf;
2995 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
2998 dmu_buf_impl_t *parent = db->db_parent;
2999 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3001 ASSERT(dn->dn_phys->dn_nlevels > 1);
3002 if (parent == NULL) {
3003 mutex_exit(&db->db_mtx);
3004 rw_enter(&dn->dn_struct_rwlock, RW_READER);
3005 parent = dbuf_hold_level(dn, db->db_level + 1,
3006 db->db_blkid >> epbs, db);
3007 rw_exit(&dn->dn_struct_rwlock);
3008 mutex_enter(&db->db_mtx);
3009 db->db_parent = parent;
3011 db->db_blkptr = (blkptr_t *)parent->db.db_data +
3012 (db->db_blkid & ((1ULL << epbs) - 1));
3018 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
3020 dmu_buf_impl_t *db = dr->dr_dbuf;
3024 ASSERT(dmu_tx_is_syncing(tx));
3026 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
3028 mutex_enter(&db->db_mtx);
3030 ASSERT(db->db_level > 0);
3033 /* Read the block if it hasn't been read yet. */
3034 if (db->db_buf == NULL) {
3035 mutex_exit(&db->db_mtx);
3036 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
3037 mutex_enter(&db->db_mtx);
3039 ASSERT3U(db->db_state, ==, DB_CACHED);
3040 ASSERT(db->db_buf != NULL);
3044 /* Indirect block size must match what the dnode thinks it is. */
3045 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3046 dbuf_check_blkptr(dn, db);
3049 /* Provide the pending dirty record to child dbufs */
3050 db->db_data_pending = dr;
3052 mutex_exit(&db->db_mtx);
3053 dbuf_write(dr, db->db_buf, tx);
3056 mutex_enter(&dr->dt.di.dr_mtx);
3057 dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx);
3058 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3059 mutex_exit(&dr->dt.di.dr_mtx);
3064 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
3066 arc_buf_t **datap = &dr->dt.dl.dr_data;
3067 dmu_buf_impl_t *db = dr->dr_dbuf;
3070 uint64_t txg = tx->tx_txg;
3072 ASSERT(dmu_tx_is_syncing(tx));
3074 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
3076 mutex_enter(&db->db_mtx);
3078 * To be synced, we must be dirtied. But we
3079 * might have been freed after the dirty.
3081 if (db->db_state == DB_UNCACHED) {
3082 /* This buffer has been freed since it was dirtied */
3083 ASSERT(db->db.db_data == NULL);
3084 } else if (db->db_state == DB_FILL) {
3085 /* This buffer was freed and is now being re-filled */
3086 ASSERT(db->db.db_data != dr->dt.dl.dr_data);
3088 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
3095 if (db->db_blkid == DMU_SPILL_BLKID) {
3096 mutex_enter(&dn->dn_mtx);
3097 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
3098 mutex_exit(&dn->dn_mtx);
3102 * If this is a bonus buffer, simply copy the bonus data into the
3103 * dnode. It will be written out when the dnode is synced (and it
3104 * will be synced, since it must have been dirty for dbuf_sync to
3107 if (db->db_blkid == DMU_BONUS_BLKID) {
3108 dbuf_dirty_record_t **drp;
3110 ASSERT(*datap != NULL);
3111 ASSERT0(db->db_level);
3112 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN);
3113 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen);
3116 if (*datap != db->db.db_data) {
3117 zio_buf_free(*datap, DN_MAX_BONUSLEN);
3118 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
3120 db->db_data_pending = NULL;
3121 drp = &db->db_last_dirty;
3123 drp = &(*drp)->dr_next;
3124 ASSERT(dr->dr_next == NULL);
3125 ASSERT(dr->dr_dbuf == db);
3127 if (dr->dr_dbuf->db_level != 0) {
3128 list_destroy(&dr->dt.di.dr_children);
3129 mutex_destroy(&dr->dt.di.dr_mtx);
3131 kmem_free(dr, sizeof (dbuf_dirty_record_t));
3132 ASSERT(db->db_dirtycnt > 0);
3133 db->db_dirtycnt -= 1;
3134 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
3141 * This function may have dropped the db_mtx lock allowing a dmu_sync
3142 * operation to sneak in. As a result, we need to ensure that we
3143 * don't check the dr_override_state until we have returned from
3144 * dbuf_check_blkptr.
3146 dbuf_check_blkptr(dn, db);
3149 * If this buffer is in the middle of an immediate write,
3150 * wait for the synchronous IO to complete.
3152 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
3153 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
3154 cv_wait(&db->db_changed, &db->db_mtx);
3155 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
3158 if (db->db_state != DB_NOFILL &&
3159 dn->dn_object != DMU_META_DNODE_OBJECT &&
3160 refcount_count(&db->db_holds) > 1 &&
3161 dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
3162 *datap == db->db_buf) {
3164 * If this buffer is currently "in use" (i.e., there
3165 * are active holds and db_data still references it),
3166 * then make a copy before we start the write so that
3167 * any modifications from the open txg will not leak
3170 * NOTE: this copy does not need to be made for
3171 * objects only modified in the syncing context (e.g.
3172 * DNONE_DNODE blocks).
3174 int psize = arc_buf_size(*datap);
3175 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
3176 enum zio_compress compress_type = arc_get_compression(*datap);
3178 if (compress_type == ZIO_COMPRESS_OFF) {
3179 *datap = arc_alloc_buf(os->os_spa, db, type, psize);
3181 ASSERT3U(type, ==, ARC_BUFC_DATA);
3182 int lsize = arc_buf_lsize(*datap);
3183 *datap = arc_alloc_compressed_buf(os->os_spa, db,
3184 psize, lsize, compress_type);
3186 bcopy(db->db.db_data, (*datap)->b_data, psize);
3188 db->db_data_pending = dr;
3190 mutex_exit(&db->db_mtx);
3192 dbuf_write(dr, *datap, tx);
3194 ASSERT(!list_link_active(&dr->dr_dirty_node));
3195 if (dn->dn_object == DMU_META_DNODE_OBJECT) {
3196 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
3200 * Although zio_nowait() does not "wait for an IO", it does
3201 * initiate the IO. If this is an empty write it seems plausible
3202 * that the IO could actually be completed before the nowait
3203 * returns. We need to DB_DNODE_EXIT() first in case
3204 * zio_nowait() invalidates the dbuf.
3207 zio_nowait(dr->dr_zio);
3212 dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx)
3214 dbuf_dirty_record_t *dr;
3216 while (dr = list_head(list)) {
3217 if (dr->dr_zio != NULL) {
3219 * If we find an already initialized zio then we
3220 * are processing the meta-dnode, and we have finished.
3221 * The dbufs for all dnodes are put back on the list
3222 * during processing, so that we can zio_wait()
3223 * these IOs after initiating all child IOs.
3225 ASSERT3U(dr->dr_dbuf->db.db_object, ==,
3226 DMU_META_DNODE_OBJECT);
3229 if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
3230 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
3231 VERIFY3U(dr->dr_dbuf->db_level, ==, level);
3233 list_remove(list, dr);
3234 if (dr->dr_dbuf->db_level > 0)
3235 dbuf_sync_indirect(dr, tx);
3237 dbuf_sync_leaf(dr, tx);
3243 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3245 dmu_buf_impl_t *db = vdb;
3247 blkptr_t *bp = zio->io_bp;
3248 blkptr_t *bp_orig = &zio->io_bp_orig;
3249 spa_t *spa = zio->io_spa;
3254 ASSERT3P(db->db_blkptr, !=, NULL);
3255 ASSERT3P(&db->db_data_pending->dr_bp_copy, ==, bp);
3259 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
3260 dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
3261 zio->io_prev_space_delta = delta;
3263 if (bp->blk_birth != 0) {
3264 ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
3265 BP_GET_TYPE(bp) == dn->dn_type) ||
3266 (db->db_blkid == DMU_SPILL_BLKID &&
3267 BP_GET_TYPE(bp) == dn->dn_bonustype) ||
3268 BP_IS_EMBEDDED(bp));
3269 ASSERT(BP_GET_LEVEL(bp) == db->db_level);
3272 mutex_enter(&db->db_mtx);
3275 if (db->db_blkid == DMU_SPILL_BLKID) {
3276 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3277 ASSERT(!(BP_IS_HOLE(bp)) &&
3278 db->db_blkptr == &dn->dn_phys->dn_spill);
3282 if (db->db_level == 0) {
3283 mutex_enter(&dn->dn_mtx);
3284 if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
3285 db->db_blkid != DMU_SPILL_BLKID)
3286 dn->dn_phys->dn_maxblkid = db->db_blkid;
3287 mutex_exit(&dn->dn_mtx);
3289 if (dn->dn_type == DMU_OT_DNODE) {
3290 dnode_phys_t *dnp = db->db.db_data;
3291 for (i = db->db.db_size >> DNODE_SHIFT; i > 0;
3293 if (dnp->dn_type != DMU_OT_NONE)
3297 if (BP_IS_HOLE(bp)) {
3304 blkptr_t *ibp = db->db.db_data;
3305 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3306 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
3307 if (BP_IS_HOLE(ibp))
3309 fill += BP_GET_FILL(ibp);
3314 if (!BP_IS_EMBEDDED(bp))
3315 bp->blk_fill = fill;
3317 mutex_exit(&db->db_mtx);
3319 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3320 *db->db_blkptr = *bp;
3321 rw_exit(&dn->dn_struct_rwlock);
3326 * This function gets called just prior to running through the compression
3327 * stage of the zio pipeline. If we're an indirect block comprised of only
3328 * holes, then we want this indirect to be compressed away to a hole. In
3329 * order to do that we must zero out any information about the holes that
3330 * this indirect points to prior to before we try to compress it.
3333 dbuf_write_children_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3335 dmu_buf_impl_t *db = vdb;
3338 unsigned int epbs, i;
3340 ASSERT3U(db->db_level, >, 0);
3343 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3344 ASSERT3U(epbs, <, 31);
3346 /* Determine if all our children are holes */
3347 for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++) {
3348 if (!BP_IS_HOLE(bp))
3353 * If all the children are holes, then zero them all out so that
3354 * we may get compressed away.
3356 if (i == 1 << epbs) {
3358 * We only found holes. Grab the rwlock to prevent
3359 * anybody from reading the blocks we're about to
3362 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3363 bzero(db->db.db_data, db->db.db_size);
3364 rw_exit(&dn->dn_struct_rwlock);
3370 * The SPA will call this callback several times for each zio - once
3371 * for every physical child i/o (zio->io_phys_children times). This
3372 * allows the DMU to monitor the progress of each logical i/o. For example,
3373 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
3374 * block. There may be a long delay before all copies/fragments are completed,
3375 * so this callback allows us to retire dirty space gradually, as the physical
3380 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
3382 dmu_buf_impl_t *db = arg;
3383 objset_t *os = db->db_objset;
3384 dsl_pool_t *dp = dmu_objset_pool(os);
3385 dbuf_dirty_record_t *dr;
3388 dr = db->db_data_pending;
3389 ASSERT3U(dr->dr_txg, ==, zio->io_txg);
3392 * The callback will be called io_phys_children times. Retire one
3393 * portion of our dirty space each time we are called. Any rounding
3394 * error will be cleaned up by dsl_pool_sync()'s call to
3395 * dsl_pool_undirty_space().
3397 delta = dr->dr_accounted / zio->io_phys_children;
3398 dsl_pool_undirty_space(dp, delta, zio->io_txg);
3403 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
3405 dmu_buf_impl_t *db = vdb;
3406 blkptr_t *bp_orig = &zio->io_bp_orig;
3407 blkptr_t *bp = db->db_blkptr;
3408 objset_t *os = db->db_objset;
3409 dmu_tx_t *tx = os->os_synctx;
3410 dbuf_dirty_record_t **drp, *dr;
3412 ASSERT0(zio->io_error);
3413 ASSERT(db->db_blkptr == bp);
3416 * For nopwrites and rewrites we ensure that the bp matches our
3417 * original and bypass all the accounting.
3419 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
3420 ASSERT(BP_EQUAL(bp, bp_orig));
3422 dsl_dataset_t *ds = os->os_dsl_dataset;
3423 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
3424 dsl_dataset_block_born(ds, bp, tx);
3427 mutex_enter(&db->db_mtx);
3431 drp = &db->db_last_dirty;
3432 while ((dr = *drp) != db->db_data_pending)
3434 ASSERT(!list_link_active(&dr->dr_dirty_node));
3435 ASSERT(dr->dr_dbuf == db);
3436 ASSERT(dr->dr_next == NULL);
3440 if (db->db_blkid == DMU_SPILL_BLKID) {
3445 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3446 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
3447 db->db_blkptr == &dn->dn_phys->dn_spill);
3452 if (db->db_level == 0) {
3453 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
3454 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
3455 if (db->db_state != DB_NOFILL) {
3456 if (dr->dt.dl.dr_data != db->db_buf)
3457 arc_buf_destroy(dr->dt.dl.dr_data, db);
3464 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3465 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
3466 if (!BP_IS_HOLE(db->db_blkptr)) {
3468 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3469 ASSERT3U(db->db_blkid, <=,
3470 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs));
3471 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
3475 mutex_destroy(&dr->dt.di.dr_mtx);
3476 list_destroy(&dr->dt.di.dr_children);
3478 kmem_free(dr, sizeof (dbuf_dirty_record_t));
3480 cv_broadcast(&db->db_changed);
3481 ASSERT(db->db_dirtycnt > 0);
3482 db->db_dirtycnt -= 1;
3483 db->db_data_pending = NULL;
3484 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg);
3488 dbuf_write_nofill_ready(zio_t *zio)
3490 dbuf_write_ready(zio, NULL, zio->io_private);
3494 dbuf_write_nofill_done(zio_t *zio)
3496 dbuf_write_done(zio, NULL, zio->io_private);
3500 dbuf_write_override_ready(zio_t *zio)
3502 dbuf_dirty_record_t *dr = zio->io_private;
3503 dmu_buf_impl_t *db = dr->dr_dbuf;
3505 dbuf_write_ready(zio, NULL, db);
3509 dbuf_write_override_done(zio_t *zio)
3511 dbuf_dirty_record_t *dr = zio->io_private;
3512 dmu_buf_impl_t *db = dr->dr_dbuf;
3513 blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
3515 mutex_enter(&db->db_mtx);
3516 if (!BP_EQUAL(zio->io_bp, obp)) {
3517 if (!BP_IS_HOLE(obp))
3518 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
3519 arc_release(dr->dt.dl.dr_data, db);
3521 mutex_exit(&db->db_mtx);
3523 dbuf_write_done(zio, NULL, db);
3526 /* Issue I/O to commit a dirty buffer to disk. */
3528 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
3530 dmu_buf_impl_t *db = dr->dr_dbuf;
3533 dmu_buf_impl_t *parent = db->db_parent;
3534 uint64_t txg = tx->tx_txg;
3535 zbookmark_phys_t zb;
3540 ASSERT(dmu_tx_is_syncing(tx));
3546 if (db->db_state != DB_NOFILL) {
3547 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
3549 * Private object buffers are released here rather
3550 * than in dbuf_dirty() since they are only modified
3551 * in the syncing context and we don't want the
3552 * overhead of making multiple copies of the data.
3554 if (BP_IS_HOLE(db->db_blkptr)) {
3557 dbuf_release_bp(db);
3562 if (parent != dn->dn_dbuf) {
3563 /* Our parent is an indirect block. */
3564 /* We have a dirty parent that has been scheduled for write. */
3565 ASSERT(parent && parent->db_data_pending);
3566 /* Our parent's buffer is one level closer to the dnode. */
3567 ASSERT(db->db_level == parent->db_level-1);
3569 * We're about to modify our parent's db_data by modifying
3570 * our block pointer, so the parent must be released.
3572 ASSERT(arc_released(parent->db_buf));
3573 zio = parent->db_data_pending->dr_zio;
3575 /* Our parent is the dnode itself. */
3576 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
3577 db->db_blkid != DMU_SPILL_BLKID) ||
3578 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
3579 if (db->db_blkid != DMU_SPILL_BLKID)
3580 ASSERT3P(db->db_blkptr, ==,
3581 &dn->dn_phys->dn_blkptr[db->db_blkid]);
3585 ASSERT(db->db_level == 0 || data == db->db_buf);
3586 ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
3589 SET_BOOKMARK(&zb, os->os_dsl_dataset ?
3590 os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
3591 db->db.db_object, db->db_level, db->db_blkid);
3593 if (db->db_blkid == DMU_SPILL_BLKID)
3595 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
3597 dmu_write_policy(os, dn, db->db_level, wp_flag,
3598 (data != NULL && arc_get_compression(data) != ZIO_COMPRESS_OFF) ?
3599 arc_get_compression(data) : ZIO_COMPRESS_INHERIT, &zp);
3603 * We copy the blkptr now (rather than when we instantiate the dirty
3604 * record), because its value can change between open context and
3605 * syncing context. We do not need to hold dn_struct_rwlock to read
3606 * db_blkptr because we are in syncing context.
3608 dr->dr_bp_copy = *db->db_blkptr;
3610 if (db->db_level == 0 &&
3611 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
3613 * The BP for this block has been provided by open context
3614 * (by dmu_sync() or dmu_buf_write_embedded()).
3616 void *contents = (data != NULL) ? data->b_data : NULL;
3618 dr->dr_zio = zio_write(zio, os->os_spa, txg, &dr->dr_bp_copy,
3619 contents, db->db.db_size, db->db.db_size, &zp,
3620 dbuf_write_override_ready, NULL, NULL,
3621 dbuf_write_override_done,
3622 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
3623 mutex_enter(&db->db_mtx);
3624 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
3625 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
3626 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
3627 mutex_exit(&db->db_mtx);
3628 } else if (db->db_state == DB_NOFILL) {
3629 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF ||
3630 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY);
3631 dr->dr_zio = zio_write(zio, os->os_spa, txg,
3632 &dr->dr_bp_copy, NULL, db->db.db_size, db->db.db_size, &zp,
3633 dbuf_write_nofill_ready, NULL, NULL,
3634 dbuf_write_nofill_done, db,
3635 ZIO_PRIORITY_ASYNC_WRITE,
3636 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
3638 ASSERT(arc_released(data));
3641 * For indirect blocks, we want to setup the children
3642 * ready callback so that we can properly handle an indirect
3643 * block that only contains holes.
3645 arc_done_func_t *children_ready_cb = NULL;
3646 if (db->db_level != 0)
3647 children_ready_cb = dbuf_write_children_ready;
3649 dr->dr_zio = arc_write(zio, os->os_spa, txg,
3650 &dr->dr_bp_copy, data, DBUF_IS_L2CACHEABLE(db),
3651 &zp, dbuf_write_ready, children_ready_cb,
3652 dbuf_write_physdone, dbuf_write_done, db,
3653 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);