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, 2017 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>
51 uint_t zfs_dbuf_evict_key;
53 static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
54 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
57 extern inline void dmu_buf_init_user(dmu_buf_user_t *dbu,
58 dmu_buf_evict_func_t *evict_func_sync,
59 dmu_buf_evict_func_t *evict_func_async,
60 dmu_buf_t **clear_on_evict_dbufp);
64 * Global data structures and functions for the dbuf cache.
66 static kmem_cache_t *dbuf_kmem_cache;
67 static taskq_t *dbu_evict_taskq;
69 static kthread_t *dbuf_cache_evict_thread;
70 static kmutex_t dbuf_evict_lock;
71 static kcondvar_t dbuf_evict_cv;
72 static boolean_t dbuf_evict_thread_exit;
75 * LRU cache of dbufs. The dbuf cache maintains a list of dbufs that
76 * are not currently held but have been recently released. These dbufs
77 * are not eligible for arc eviction until they are aged out of the cache.
78 * Dbufs are added to the dbuf cache once the last hold is released. If a
79 * dbuf is later accessed and still exists in the dbuf cache, then it will
80 * be removed from the cache and later re-added to the head of the cache.
81 * Dbufs that are aged out of the cache will be immediately destroyed and
82 * become eligible for arc eviction.
84 static multilist_t *dbuf_cache;
85 static refcount_t dbuf_cache_size;
86 uint64_t dbuf_cache_max_bytes = 100 * 1024 * 1024;
88 /* Cap the size of the dbuf cache to log2 fraction of arc size. */
89 int dbuf_cache_max_shift = 5;
92 * The dbuf cache uses a three-stage eviction policy:
93 * - A low water marker designates when the dbuf eviction thread
94 * should stop evicting from the dbuf cache.
95 * - When we reach the maximum size (aka mid water mark), we
96 * signal the eviction thread to run.
97 * - The high water mark indicates when the eviction thread
98 * is unable to keep up with the incoming load and eviction must
99 * happen in the context of the calling thread.
103 * low water mid water hi water
104 * +----------------------------------------+----------+----------+
109 * +----------------------------------------+----------+----------+
111 * evicting eviction directly
114 * The high and low water marks indicate the operating range for the eviction
115 * thread. The low water mark is, by default, 90% of the total size of the
116 * cache and the high water mark is at 110% (both of these percentages can be
117 * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct,
118 * respectively). The eviction thread will try to ensure that the cache remains
119 * within this range by waking up every second and checking if the cache is
120 * above the low water mark. The thread can also be woken up by callers adding
121 * elements into the cache if the cache is larger than the mid water (i.e max
122 * cache size). Once the eviction thread is woken up and eviction is required,
123 * it will continue evicting buffers until it's able to reduce the cache size
124 * to the low water mark. If the cache size continues to grow and hits the high
125 * water mark, then callers adding elments to the cache will begin to evict
126 * directly from the cache until the cache is no longer above the high water
131 * The percentage above and below the maximum cache size.
133 uint_t dbuf_cache_hiwater_pct = 10;
134 uint_t dbuf_cache_lowater_pct = 10;
138 dbuf_cons(void *vdb, void *unused, int kmflag)
140 dmu_buf_impl_t *db = vdb;
141 bzero(db, sizeof (dmu_buf_impl_t));
143 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
144 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
145 multilist_link_init(&db->db_cache_link);
146 refcount_create(&db->db_holds);
153 dbuf_dest(void *vdb, void *unused)
155 dmu_buf_impl_t *db = vdb;
156 mutex_destroy(&db->db_mtx);
157 cv_destroy(&db->db_changed);
158 ASSERT(!multilist_link_active(&db->db_cache_link));
159 refcount_destroy(&db->db_holds);
163 * dbuf hash table routines
165 static dbuf_hash_table_t dbuf_hash_table;
167 static uint64_t dbuf_hash_count;
170 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
172 uintptr_t osv = (uintptr_t)os;
173 uint64_t crc = -1ULL;
175 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
176 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF];
177 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF];
178 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF];
179 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF];
180 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF];
181 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF];
183 crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16);
188 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
189 ((dbuf)->db.db_object == (obj) && \
190 (dbuf)->db_objset == (os) && \
191 (dbuf)->db_level == (level) && \
192 (dbuf)->db_blkid == (blkid))
195 dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid)
197 dbuf_hash_table_t *h = &dbuf_hash_table;
198 uint64_t hv = dbuf_hash(os, obj, level, blkid);
199 uint64_t idx = hv & h->hash_table_mask;
202 mutex_enter(DBUF_HASH_MUTEX(h, idx));
203 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
204 if (DBUF_EQUAL(db, os, obj, level, blkid)) {
205 mutex_enter(&db->db_mtx);
206 if (db->db_state != DB_EVICTING) {
207 mutex_exit(DBUF_HASH_MUTEX(h, idx));
210 mutex_exit(&db->db_mtx);
213 mutex_exit(DBUF_HASH_MUTEX(h, idx));
217 static dmu_buf_impl_t *
218 dbuf_find_bonus(objset_t *os, uint64_t object)
221 dmu_buf_impl_t *db = NULL;
223 if (dnode_hold(os, object, FTAG, &dn) == 0) {
224 rw_enter(&dn->dn_struct_rwlock, RW_READER);
225 if (dn->dn_bonus != NULL) {
227 mutex_enter(&db->db_mtx);
229 rw_exit(&dn->dn_struct_rwlock);
230 dnode_rele(dn, FTAG);
236 * Insert an entry into the hash table. If there is already an element
237 * equal to elem in the hash table, then the already existing element
238 * will be returned and the new element will not be inserted.
239 * Otherwise returns NULL.
241 static dmu_buf_impl_t *
242 dbuf_hash_insert(dmu_buf_impl_t *db)
244 dbuf_hash_table_t *h = &dbuf_hash_table;
245 objset_t *os = db->db_objset;
246 uint64_t obj = db->db.db_object;
247 int level = db->db_level;
248 uint64_t blkid = db->db_blkid;
249 uint64_t hv = dbuf_hash(os, obj, level, blkid);
250 uint64_t idx = hv & h->hash_table_mask;
253 mutex_enter(DBUF_HASH_MUTEX(h, idx));
254 for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) {
255 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
256 mutex_enter(&dbf->db_mtx);
257 if (dbf->db_state != DB_EVICTING) {
258 mutex_exit(DBUF_HASH_MUTEX(h, idx));
261 mutex_exit(&dbf->db_mtx);
265 mutex_enter(&db->db_mtx);
266 db->db_hash_next = h->hash_table[idx];
267 h->hash_table[idx] = db;
268 mutex_exit(DBUF_HASH_MUTEX(h, idx));
269 atomic_inc_64(&dbuf_hash_count);
275 * Remove an entry from the hash table. It must be in the EVICTING state.
278 dbuf_hash_remove(dmu_buf_impl_t *db)
280 dbuf_hash_table_t *h = &dbuf_hash_table;
281 uint64_t hv = dbuf_hash(db->db_objset, db->db.db_object,
282 db->db_level, db->db_blkid);
283 uint64_t idx = hv & h->hash_table_mask;
284 dmu_buf_impl_t *dbf, **dbp;
287 * We musn't hold db_mtx to maintain lock ordering:
288 * DBUF_HASH_MUTEX > db_mtx.
290 ASSERT(refcount_is_zero(&db->db_holds));
291 ASSERT(db->db_state == DB_EVICTING);
292 ASSERT(!MUTEX_HELD(&db->db_mtx));
294 mutex_enter(DBUF_HASH_MUTEX(h, idx));
295 dbp = &h->hash_table[idx];
296 while ((dbf = *dbp) != db) {
297 dbp = &dbf->db_hash_next;
300 *dbp = db->db_hash_next;
301 db->db_hash_next = NULL;
302 mutex_exit(DBUF_HASH_MUTEX(h, idx));
303 atomic_dec_64(&dbuf_hash_count);
309 } dbvu_verify_type_t;
312 dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type)
317 if (db->db_user == NULL)
320 /* Only data blocks support the attachment of user data. */
321 ASSERT(db->db_level == 0);
323 /* Clients must resolve a dbuf before attaching user data. */
324 ASSERT(db->db.db_data != NULL);
325 ASSERT3U(db->db_state, ==, DB_CACHED);
327 holds = refcount_count(&db->db_holds);
328 if (verify_type == DBVU_EVICTING) {
330 * Immediate eviction occurs when holds == dirtycnt.
331 * For normal eviction buffers, holds is zero on
332 * eviction, except when dbuf_fix_old_data() calls
333 * dbuf_clear_data(). However, the hold count can grow
334 * during eviction even though db_mtx is held (see
335 * dmu_bonus_hold() for an example), so we can only
336 * test the generic invariant that holds >= dirtycnt.
338 ASSERT3U(holds, >=, db->db_dirtycnt);
340 if (db->db_user_immediate_evict == TRUE)
341 ASSERT3U(holds, >=, db->db_dirtycnt);
343 ASSERT3U(holds, >, 0);
349 dbuf_evict_user(dmu_buf_impl_t *db)
351 dmu_buf_user_t *dbu = db->db_user;
353 ASSERT(MUTEX_HELD(&db->db_mtx));
358 dbuf_verify_user(db, DBVU_EVICTING);
362 if (dbu->dbu_clear_on_evict_dbufp != NULL)
363 *dbu->dbu_clear_on_evict_dbufp = NULL;
367 * There are two eviction callbacks - one that we call synchronously
368 * and one that we invoke via a taskq. The async one is useful for
369 * avoiding lock order reversals and limiting stack depth.
371 * Note that if we have a sync callback but no async callback,
372 * it's likely that the sync callback will free the structure
373 * containing the dbu. In that case we need to take care to not
374 * dereference dbu after calling the sync evict func.
376 boolean_t has_async = (dbu->dbu_evict_func_async != NULL);
378 if (dbu->dbu_evict_func_sync != NULL)
379 dbu->dbu_evict_func_sync(dbu);
382 taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func_async,
383 dbu, 0, &dbu->dbu_tqent);
388 dbuf_is_metadata(dmu_buf_impl_t *db)
390 if (db->db_level > 0) {
393 boolean_t is_metadata;
396 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type);
399 return (is_metadata);
404 * This function *must* return indices evenly distributed between all
405 * sublists of the multilist. This is needed due to how the dbuf eviction
406 * code is laid out; dbuf_evict_thread() assumes dbufs are evenly
407 * distributed between all sublists and uses this assumption when
408 * deciding which sublist to evict from and how much to evict from it.
411 dbuf_cache_multilist_index_func(multilist_t *ml, void *obj)
413 dmu_buf_impl_t *db = obj;
416 * The assumption here, is the hash value for a given
417 * dmu_buf_impl_t will remain constant throughout it's lifetime
418 * (i.e. it's objset, object, level and blkid fields don't change).
419 * Thus, we don't need to store the dbuf's sublist index
420 * on insertion, as this index can be recalculated on removal.
422 * Also, the low order bits of the hash value are thought to be
423 * distributed evenly. Otherwise, in the case that the multilist
424 * has a power of two number of sublists, each sublists' usage
425 * would not be evenly distributed.
427 return (dbuf_hash(db->db_objset, db->db.db_object,
428 db->db_level, db->db_blkid) %
429 multilist_get_num_sublists(ml));
432 static inline boolean_t
433 dbuf_cache_above_hiwater(void)
435 uint64_t dbuf_cache_hiwater_bytes =
436 (dbuf_cache_max_bytes * dbuf_cache_hiwater_pct) / 100;
438 return (refcount_count(&dbuf_cache_size) >
439 dbuf_cache_max_bytes + dbuf_cache_hiwater_bytes);
442 static inline boolean_t
443 dbuf_cache_above_lowater(void)
445 uint64_t dbuf_cache_lowater_bytes =
446 (dbuf_cache_max_bytes * dbuf_cache_lowater_pct) / 100;
448 return (refcount_count(&dbuf_cache_size) >
449 dbuf_cache_max_bytes - dbuf_cache_lowater_bytes);
453 * Evict the oldest eligible dbuf from the dbuf cache.
458 int idx = multilist_get_random_index(dbuf_cache);
459 multilist_sublist_t *mls = multilist_sublist_lock(dbuf_cache, idx);
461 ASSERT(!MUTEX_HELD(&dbuf_evict_lock));
464 * Set the thread's tsd to indicate that it's processing evictions.
465 * Once a thread stops evicting from the dbuf cache it will
466 * reset its tsd to NULL.
468 ASSERT3P(tsd_get(zfs_dbuf_evict_key), ==, NULL);
469 (void) tsd_set(zfs_dbuf_evict_key, (void *)B_TRUE);
471 dmu_buf_impl_t *db = multilist_sublist_tail(mls);
472 while (db != NULL && mutex_tryenter(&db->db_mtx) == 0) {
473 db = multilist_sublist_prev(mls, db);
476 DTRACE_PROBE2(dbuf__evict__one, dmu_buf_impl_t *, db,
477 multilist_sublist_t *, mls);
480 multilist_sublist_remove(mls, db);
481 multilist_sublist_unlock(mls);
482 (void) refcount_remove_many(&dbuf_cache_size,
486 multilist_sublist_unlock(mls);
488 (void) tsd_set(zfs_dbuf_evict_key, NULL);
492 * The dbuf evict thread is responsible for aging out dbufs from the
493 * cache. Once the cache has reached it's maximum size, dbufs are removed
494 * and destroyed. The eviction thread will continue running until the size
495 * of the dbuf cache is at or below the maximum size. Once the dbuf is aged
496 * out of the cache it is destroyed and becomes eligible for arc eviction.
500 dbuf_evict_thread(void *unused)
504 CALLB_CPR_INIT(&cpr, &dbuf_evict_lock, callb_generic_cpr, FTAG);
506 mutex_enter(&dbuf_evict_lock);
507 while (!dbuf_evict_thread_exit) {
508 while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
509 CALLB_CPR_SAFE_BEGIN(&cpr);
510 (void) cv_timedwait_hires(&dbuf_evict_cv,
511 &dbuf_evict_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
512 CALLB_CPR_SAFE_END(&cpr, &dbuf_evict_lock);
514 mutex_exit(&dbuf_evict_lock);
517 * Keep evicting as long as we're above the low water mark
518 * for the cache. We do this without holding the locks to
519 * minimize lock contention.
521 while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
525 mutex_enter(&dbuf_evict_lock);
528 dbuf_evict_thread_exit = B_FALSE;
529 cv_broadcast(&dbuf_evict_cv);
530 CALLB_CPR_EXIT(&cpr); /* drops dbuf_evict_lock */
535 * Wake up the dbuf eviction thread if the dbuf cache is at its max size.
536 * If the dbuf cache is at its high water mark, then evict a dbuf from the
537 * dbuf cache using the callers context.
540 dbuf_evict_notify(void)
544 * We use thread specific data to track when a thread has
545 * started processing evictions. This allows us to avoid deeply
546 * nested stacks that would have a call flow similar to this:
548 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify()
551 * +-----dbuf_destroy()<--dbuf_evict_one()<--------+
553 * The dbuf_eviction_thread will always have its tsd set until
554 * that thread exits. All other threads will only set their tsd
555 * if they are participating in the eviction process. This only
556 * happens if the eviction thread is unable to process evictions
557 * fast enough. To keep the dbuf cache size in check, other threads
558 * can evict from the dbuf cache directly. Those threads will set
559 * their tsd values so that we ensure that they only evict one dbuf
560 * from the dbuf cache.
562 if (tsd_get(zfs_dbuf_evict_key) != NULL)
566 * We check if we should evict without holding the dbuf_evict_lock,
567 * because it's OK to occasionally make the wrong decision here,
568 * and grabbing the lock results in massive lock contention.
570 if (refcount_count(&dbuf_cache_size) > dbuf_cache_max_bytes) {
571 if (dbuf_cache_above_hiwater())
573 cv_signal(&dbuf_evict_cv);
580 uint64_t hsize = 1ULL << 16;
581 dbuf_hash_table_t *h = &dbuf_hash_table;
585 * The hash table is big enough to fill all of physical memory
586 * with an average 4K block size. The table will take up
587 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
589 while (hsize * 4096 < physmem * PAGESIZE)
593 h->hash_table_mask = hsize - 1;
594 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
595 if (h->hash_table == NULL) {
596 /* XXX - we should really return an error instead of assert */
597 ASSERT(hsize > (1ULL << 10));
602 dbuf_kmem_cache = kmem_cache_create("dmu_buf_impl_t",
603 sizeof (dmu_buf_impl_t),
604 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
606 for (i = 0; i < DBUF_MUTEXES; i++)
607 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
610 * Setup the parameters for the dbuf cache. We cap the size of the
611 * dbuf cache to 1/32nd (default) of the size of the ARC.
613 dbuf_cache_max_bytes = MIN(dbuf_cache_max_bytes,
614 arc_max_bytes() >> dbuf_cache_max_shift);
617 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
618 * configuration is not required.
620 dbu_evict_taskq = taskq_create("dbu_evict", 1, minclsyspri, 0, 0, 0);
622 dbuf_cache = multilist_create(sizeof (dmu_buf_impl_t),
623 offsetof(dmu_buf_impl_t, db_cache_link),
624 dbuf_cache_multilist_index_func);
625 refcount_create(&dbuf_cache_size);
627 tsd_create(&zfs_dbuf_evict_key, NULL);
628 dbuf_evict_thread_exit = B_FALSE;
629 mutex_init(&dbuf_evict_lock, NULL, MUTEX_DEFAULT, NULL);
630 cv_init(&dbuf_evict_cv, NULL, CV_DEFAULT, NULL);
631 dbuf_cache_evict_thread = thread_create(NULL, 0, dbuf_evict_thread,
632 NULL, 0, &p0, TS_RUN, minclsyspri);
638 dbuf_hash_table_t *h = &dbuf_hash_table;
641 for (i = 0; i < DBUF_MUTEXES; i++)
642 mutex_destroy(&h->hash_mutexes[i]);
643 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
644 kmem_cache_destroy(dbuf_kmem_cache);
645 taskq_destroy(dbu_evict_taskq);
647 mutex_enter(&dbuf_evict_lock);
648 dbuf_evict_thread_exit = B_TRUE;
649 while (dbuf_evict_thread_exit) {
650 cv_signal(&dbuf_evict_cv);
651 cv_wait(&dbuf_evict_cv, &dbuf_evict_lock);
653 mutex_exit(&dbuf_evict_lock);
654 tsd_destroy(&zfs_dbuf_evict_key);
656 mutex_destroy(&dbuf_evict_lock);
657 cv_destroy(&dbuf_evict_cv);
659 refcount_destroy(&dbuf_cache_size);
660 multilist_destroy(dbuf_cache);
669 dbuf_verify(dmu_buf_impl_t *db)
672 dbuf_dirty_record_t *dr;
674 ASSERT(MUTEX_HELD(&db->db_mtx));
676 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
679 ASSERT(db->db_objset != NULL);
683 ASSERT(db->db_parent == NULL);
684 ASSERT(db->db_blkptr == NULL);
686 ASSERT3U(db->db.db_object, ==, dn->dn_object);
687 ASSERT3P(db->db_objset, ==, dn->dn_objset);
688 ASSERT3U(db->db_level, <, dn->dn_nlevels);
689 ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
690 db->db_blkid == DMU_SPILL_BLKID ||
691 !avl_is_empty(&dn->dn_dbufs));
693 if (db->db_blkid == DMU_BONUS_BLKID) {
695 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
696 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
697 } else if (db->db_blkid == DMU_SPILL_BLKID) {
699 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
700 ASSERT0(db->db.db_offset);
702 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
705 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
706 ASSERT(dr->dr_dbuf == db);
708 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
709 ASSERT(dr->dr_dbuf == db);
712 * We can't assert that db_size matches dn_datablksz because it
713 * can be momentarily different when another thread is doing
716 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
717 dr = db->db_data_pending;
719 * It should only be modified in syncing context, so
720 * make sure we only have one copy of the data.
722 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
725 /* verify db->db_blkptr */
727 if (db->db_parent == dn->dn_dbuf) {
728 /* db is pointed to by the dnode */
729 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
730 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
731 ASSERT(db->db_parent == NULL);
733 ASSERT(db->db_parent != NULL);
734 if (db->db_blkid != DMU_SPILL_BLKID)
735 ASSERT3P(db->db_blkptr, ==,
736 &dn->dn_phys->dn_blkptr[db->db_blkid]);
738 /* db is pointed to by an indirect block */
739 int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
740 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
741 ASSERT3U(db->db_parent->db.db_object, ==,
744 * dnode_grow_indblksz() can make this fail if we don't
745 * have the struct_rwlock. XXX indblksz no longer
746 * grows. safe to do this now?
748 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
749 ASSERT3P(db->db_blkptr, ==,
750 ((blkptr_t *)db->db_parent->db.db_data +
751 db->db_blkid % epb));
755 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
756 (db->db_buf == NULL || db->db_buf->b_data) &&
757 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
758 db->db_state != DB_FILL && !dn->dn_free_txg) {
760 * If the blkptr isn't set but they have nonzero data,
761 * it had better be dirty, otherwise we'll lose that
762 * data when we evict this buffer.
764 * There is an exception to this rule for indirect blocks; in
765 * this case, if the indirect block is a hole, we fill in a few
766 * fields on each of the child blocks (importantly, birth time)
767 * to prevent hole birth times from being lost when you
768 * partially fill in a hole.
770 if (db->db_dirtycnt == 0) {
771 if (db->db_level == 0) {
772 uint64_t *buf = db->db.db_data;
775 for (i = 0; i < db->db.db_size >> 3; i++) {
779 blkptr_t *bps = db->db.db_data;
780 ASSERT3U(1 << DB_DNODE(db)->dn_indblkshift, ==,
783 * We want to verify that all the blkptrs in the
784 * indirect block are holes, but we may have
785 * automatically set up a few fields for them.
786 * We iterate through each blkptr and verify
787 * they only have those fields set.
790 i < db->db.db_size / sizeof (blkptr_t);
792 blkptr_t *bp = &bps[i];
793 ASSERT(ZIO_CHECKSUM_IS_ZERO(
796 DVA_IS_EMPTY(&bp->blk_dva[0]) &&
797 DVA_IS_EMPTY(&bp->blk_dva[1]) &&
798 DVA_IS_EMPTY(&bp->blk_dva[2]));
799 ASSERT0(bp->blk_fill);
800 ASSERT0(bp->blk_pad[0]);
801 ASSERT0(bp->blk_pad[1]);
802 ASSERT(!BP_IS_EMBEDDED(bp));
803 ASSERT(BP_IS_HOLE(bp));
804 ASSERT0(bp->blk_phys_birth);
814 dbuf_clear_data(dmu_buf_impl_t *db)
816 ASSERT(MUTEX_HELD(&db->db_mtx));
818 ASSERT3P(db->db_buf, ==, NULL);
819 db->db.db_data = NULL;
820 if (db->db_state != DB_NOFILL)
821 db->db_state = DB_UNCACHED;
825 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
827 ASSERT(MUTEX_HELD(&db->db_mtx));
831 ASSERT(buf->b_data != NULL);
832 db->db.db_data = buf->b_data;
836 * Loan out an arc_buf for read. Return the loaned arc_buf.
839 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
843 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
844 mutex_enter(&db->db_mtx);
845 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
846 int blksz = db->db.db_size;
847 spa_t *spa = db->db_objset->os_spa;
849 mutex_exit(&db->db_mtx);
850 abuf = arc_loan_buf(spa, B_FALSE, blksz);
851 bcopy(db->db.db_data, abuf->b_data, blksz);
854 arc_loan_inuse_buf(abuf, db);
857 mutex_exit(&db->db_mtx);
863 * Calculate which level n block references the data at the level 0 offset
867 dbuf_whichblock(dnode_t *dn, int64_t level, uint64_t offset)
869 if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) {
871 * The level n blkid is equal to the level 0 blkid divided by
872 * the number of level 0s in a level n block.
874 * The level 0 blkid is offset >> datablkshift =
875 * offset / 2^datablkshift.
877 * The number of level 0s in a level n is the number of block
878 * pointers in an indirect block, raised to the power of level.
879 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
880 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
882 * Thus, the level n blkid is: offset /
883 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
884 * = offset / 2^(datablkshift + level *
885 * (indblkshift - SPA_BLKPTRSHIFT))
886 * = offset >> (datablkshift + level *
887 * (indblkshift - SPA_BLKPTRSHIFT))
889 return (offset >> (dn->dn_datablkshift + level *
890 (dn->dn_indblkshift - SPA_BLKPTRSHIFT)));
892 ASSERT3U(offset, <, dn->dn_datablksz);
898 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
900 dmu_buf_impl_t *db = vdb;
902 mutex_enter(&db->db_mtx);
903 ASSERT3U(db->db_state, ==, DB_READ);
905 * All reads are synchronous, so we must have a hold on the dbuf
907 ASSERT(refcount_count(&db->db_holds) > 0);
908 ASSERT(db->db_buf == NULL);
909 ASSERT(db->db.db_data == NULL);
910 if (db->db_level == 0 && db->db_freed_in_flight) {
911 /* we were freed in flight; disregard any error */
912 arc_release(buf, db);
913 bzero(buf->b_data, db->db.db_size);
915 db->db_freed_in_flight = FALSE;
916 dbuf_set_data(db, buf);
917 db->db_state = DB_CACHED;
918 } else if (zio == NULL || zio->io_error == 0) {
919 dbuf_set_data(db, buf);
920 db->db_state = DB_CACHED;
922 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
923 ASSERT3P(db->db_buf, ==, NULL);
924 arc_buf_destroy(buf, db);
925 db->db_state = DB_UNCACHED;
927 cv_broadcast(&db->db_changed);
928 dbuf_rele_and_unlock(db, NULL);
932 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
936 arc_flags_t aflags = ARC_FLAG_NOWAIT;
940 ASSERT(!refcount_is_zero(&db->db_holds));
941 /* We need the struct_rwlock to prevent db_blkptr from changing. */
942 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
943 ASSERT(MUTEX_HELD(&db->db_mtx));
944 ASSERT(db->db_state == DB_UNCACHED);
945 ASSERT(db->db_buf == NULL);
947 if (db->db_blkid == DMU_BONUS_BLKID) {
948 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
950 ASSERT3U(bonuslen, <=, db->db.db_size);
951 db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN);
952 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
953 if (bonuslen < DN_MAX_BONUSLEN)
954 bzero(db->db.db_data, DN_MAX_BONUSLEN);
956 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
958 db->db_state = DB_CACHED;
959 mutex_exit(&db->db_mtx);
964 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
965 * processes the delete record and clears the bp while we are waiting
966 * for the dn_mtx (resulting in a "no" from block_freed).
968 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
969 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
970 BP_IS_HOLE(db->db_blkptr)))) {
971 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
973 dbuf_set_data(db, arc_alloc_buf(db->db_objset->os_spa, db, type,
975 bzero(db->db.db_data, db->db.db_size);
977 if (db->db_blkptr != NULL && db->db_level > 0 &&
978 BP_IS_HOLE(db->db_blkptr) &&
979 db->db_blkptr->blk_birth != 0) {
980 blkptr_t *bps = db->db.db_data;
981 for (int i = 0; i < ((1 <<
982 DB_DNODE(db)->dn_indblkshift) / sizeof (blkptr_t));
984 blkptr_t *bp = &bps[i];
985 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
986 1 << dn->dn_indblkshift);
988 BP_GET_LEVEL(db->db_blkptr) == 1 ?
990 BP_GET_LSIZE(db->db_blkptr));
991 BP_SET_TYPE(bp, BP_GET_TYPE(db->db_blkptr));
993 BP_GET_LEVEL(db->db_blkptr) - 1);
994 BP_SET_BIRTH(bp, db->db_blkptr->blk_birth, 0);
998 db->db_state = DB_CACHED;
999 mutex_exit(&db->db_mtx);
1005 db->db_state = DB_READ;
1006 mutex_exit(&db->db_mtx);
1008 if (DBUF_IS_L2CACHEABLE(db))
1009 aflags |= ARC_FLAG_L2CACHE;
1011 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
1012 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
1013 db->db.db_object, db->db_level, db->db_blkid);
1015 dbuf_add_ref(db, NULL);
1017 (void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr,
1018 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
1019 (flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
1024 * This is our just-in-time copy function. It makes a copy of buffers that
1025 * have been modified in a previous transaction group before we access them in
1026 * the current active group.
1028 * This function is used in three places: when we are dirtying a buffer for the
1029 * first time in a txg, when we are freeing a range in a dnode that includes
1030 * this buffer, and when we are accessing a buffer which was received compressed
1031 * and later referenced in a WRITE_BYREF record.
1033 * Note that when we are called from dbuf_free_range() we do not put a hold on
1034 * the buffer, we just traverse the active dbuf list for the dnode.
1037 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
1039 dbuf_dirty_record_t *dr = db->db_last_dirty;
1041 ASSERT(MUTEX_HELD(&db->db_mtx));
1042 ASSERT(db->db.db_data != NULL);
1043 ASSERT(db->db_level == 0);
1044 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
1047 (dr->dt.dl.dr_data !=
1048 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
1052 * If the last dirty record for this dbuf has not yet synced
1053 * and its referencing the dbuf data, either:
1054 * reset the reference to point to a new copy,
1055 * or (if there a no active holders)
1056 * just null out the current db_data pointer.
1058 ASSERT(dr->dr_txg >= txg - 2);
1059 if (db->db_blkid == DMU_BONUS_BLKID) {
1060 /* Note that the data bufs here are zio_bufs */
1061 dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
1062 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
1063 bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
1064 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
1065 int size = arc_buf_size(db->db_buf);
1066 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1067 spa_t *spa = db->db_objset->os_spa;
1068 enum zio_compress compress_type =
1069 arc_get_compression(db->db_buf);
1071 if (compress_type == ZIO_COMPRESS_OFF) {
1072 dr->dt.dl.dr_data = arc_alloc_buf(spa, db, type, size);
1074 ASSERT3U(type, ==, ARC_BUFC_DATA);
1075 dr->dt.dl.dr_data = arc_alloc_compressed_buf(spa, db,
1076 size, arc_buf_lsize(db->db_buf), compress_type);
1078 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
1081 dbuf_clear_data(db);
1086 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
1093 * We don't have to hold the mutex to check db_state because it
1094 * can't be freed while we have a hold on the buffer.
1096 ASSERT(!refcount_is_zero(&db->db_holds));
1098 if (db->db_state == DB_NOFILL)
1099 return (SET_ERROR(EIO));
1103 if ((flags & DB_RF_HAVESTRUCT) == 0)
1104 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1106 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1107 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
1108 DBUF_IS_CACHEABLE(db);
1110 mutex_enter(&db->db_mtx);
1111 if (db->db_state == DB_CACHED) {
1113 * If the arc buf is compressed, we need to decompress it to
1114 * read the data. This could happen during the "zfs receive" of
1115 * a stream which is compressed and deduplicated.
1117 if (db->db_buf != NULL &&
1118 arc_get_compression(db->db_buf) != ZIO_COMPRESS_OFF) {
1119 dbuf_fix_old_data(db,
1120 spa_syncing_txg(dmu_objset_spa(db->db_objset)));
1121 err = arc_decompress(db->db_buf);
1122 dbuf_set_data(db, db->db_buf);
1124 mutex_exit(&db->db_mtx);
1126 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1127 if ((flags & DB_RF_HAVESTRUCT) == 0)
1128 rw_exit(&dn->dn_struct_rwlock);
1130 } else if (db->db_state == DB_UNCACHED) {
1131 spa_t *spa = dn->dn_objset->os_spa;
1132 boolean_t need_wait = B_FALSE;
1135 db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) {
1136 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);
1189 dbuf_noread(dmu_buf_impl_t *db)
1191 ASSERT(!refcount_is_zero(&db->db_holds));
1192 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1193 mutex_enter(&db->db_mtx);
1194 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1195 cv_wait(&db->db_changed, &db->db_mtx);
1196 if (db->db_state == DB_UNCACHED) {
1197 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1198 spa_t *spa = db->db_objset->os_spa;
1200 ASSERT(db->db_buf == NULL);
1201 ASSERT(db->db.db_data == NULL);
1202 dbuf_set_data(db, arc_alloc_buf(spa, db, type, db->db.db_size));
1203 db->db_state = DB_FILL;
1204 } else if (db->db_state == DB_NOFILL) {
1205 dbuf_clear_data(db);
1207 ASSERT3U(db->db_state, ==, DB_CACHED);
1209 mutex_exit(&db->db_mtx);
1213 dbuf_unoverride(dbuf_dirty_record_t *dr)
1215 dmu_buf_impl_t *db = dr->dr_dbuf;
1216 blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
1217 uint64_t txg = dr->dr_txg;
1219 ASSERT(MUTEX_HELD(&db->db_mtx));
1221 * This assert is valid because dmu_sync() expects to be called by
1222 * a zilog's get_data while holding a range lock. This call only
1223 * comes from dbuf_dirty() callers who must also hold a range lock.
1225 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
1226 ASSERT(db->db_level == 0);
1228 if (db->db_blkid == DMU_BONUS_BLKID ||
1229 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
1232 ASSERT(db->db_data_pending != dr);
1234 /* free this block */
1235 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite)
1236 zio_free(db->db_objset->os_spa, txg, bp);
1238 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1239 dr->dt.dl.dr_nopwrite = B_FALSE;
1242 * Release the already-written buffer, so we leave it in
1243 * a consistent dirty state. Note that all callers are
1244 * modifying the buffer, so they will immediately do
1245 * another (redundant) arc_release(). Therefore, leave
1246 * the buf thawed to save the effort of freezing &
1247 * immediately re-thawing it.
1249 arc_release(dr->dt.dl.dr_data, db);
1253 * Evict (if its unreferenced) or clear (if its referenced) any level-0
1254 * data blocks in the free range, so that any future readers will find
1258 dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
1261 dmu_buf_impl_t db_search;
1262 dmu_buf_impl_t *db, *db_next;
1263 uint64_t txg = tx->tx_txg;
1266 if (end_blkid > dn->dn_maxblkid &&
1267 !(start_blkid == DMU_SPILL_BLKID || end_blkid == DMU_SPILL_BLKID))
1268 end_blkid = dn->dn_maxblkid;
1269 dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid);
1271 db_search.db_level = 0;
1272 db_search.db_blkid = start_blkid;
1273 db_search.db_state = DB_SEARCH;
1275 mutex_enter(&dn->dn_dbufs_mtx);
1276 db = avl_find(&dn->dn_dbufs, &db_search, &where);
1277 ASSERT3P(db, ==, NULL);
1279 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1281 for (; db != NULL; db = db_next) {
1282 db_next = AVL_NEXT(&dn->dn_dbufs, db);
1283 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1285 if (db->db_level != 0 || db->db_blkid > end_blkid) {
1288 ASSERT3U(db->db_blkid, >=, start_blkid);
1290 /* found a level 0 buffer in the range */
1291 mutex_enter(&db->db_mtx);
1292 if (dbuf_undirty(db, tx)) {
1293 /* mutex has been dropped and dbuf destroyed */
1297 if (db->db_state == DB_UNCACHED ||
1298 db->db_state == DB_NOFILL ||
1299 db->db_state == DB_EVICTING) {
1300 ASSERT(db->db.db_data == NULL);
1301 mutex_exit(&db->db_mtx);
1304 if (db->db_state == DB_READ || db->db_state == DB_FILL) {
1305 /* will be handled in dbuf_read_done or dbuf_rele */
1306 db->db_freed_in_flight = TRUE;
1307 mutex_exit(&db->db_mtx);
1310 if (refcount_count(&db->db_holds) == 0) {
1315 /* The dbuf is referenced */
1317 if (db->db_last_dirty != NULL) {
1318 dbuf_dirty_record_t *dr = db->db_last_dirty;
1320 if (dr->dr_txg == txg) {
1322 * This buffer is "in-use", re-adjust the file
1323 * size to reflect that this buffer may
1324 * contain new data when we sync.
1326 if (db->db_blkid != DMU_SPILL_BLKID &&
1327 db->db_blkid > dn->dn_maxblkid)
1328 dn->dn_maxblkid = db->db_blkid;
1329 dbuf_unoverride(dr);
1332 * This dbuf is not dirty in the open context.
1333 * Either uncache it (if its not referenced in
1334 * the open context) or reset its contents to
1337 dbuf_fix_old_data(db, txg);
1340 /* clear the contents if its cached */
1341 if (db->db_state == DB_CACHED) {
1342 ASSERT(db->db.db_data != NULL);
1343 arc_release(db->db_buf, db);
1344 bzero(db->db.db_data, db->db.db_size);
1345 arc_buf_freeze(db->db_buf);
1348 mutex_exit(&db->db_mtx);
1350 mutex_exit(&dn->dn_dbufs_mtx);
1354 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
1356 arc_buf_t *buf, *obuf;
1357 int osize = db->db.db_size;
1358 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1361 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1366 /* XXX does *this* func really need the lock? */
1367 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1370 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1371 * is OK, because there can be no other references to the db
1372 * when we are changing its size, so no concurrent DB_FILL can
1376 * XXX we should be doing a dbuf_read, checking the return
1377 * value and returning that up to our callers
1379 dmu_buf_will_dirty(&db->db, tx);
1381 /* create the data buffer for the new block */
1382 buf = arc_alloc_buf(dn->dn_objset->os_spa, db, type, size);
1384 /* copy old block data to the new block */
1386 bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
1387 /* zero the remainder */
1389 bzero((uint8_t *)buf->b_data + osize, size - osize);
1391 mutex_enter(&db->db_mtx);
1392 dbuf_set_data(db, buf);
1393 arc_buf_destroy(obuf, db);
1394 db->db.db_size = size;
1396 if (db->db_level == 0) {
1397 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1398 db->db_last_dirty->dt.dl.dr_data = buf;
1400 mutex_exit(&db->db_mtx);
1402 dmu_objset_willuse_space(dn->dn_objset, size - osize, tx);
1407 dbuf_release_bp(dmu_buf_impl_t *db)
1409 objset_t *os = db->db_objset;
1411 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1412 ASSERT(arc_released(os->os_phys_buf) ||
1413 list_link_active(&os->os_dsl_dataset->ds_synced_link));
1414 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1416 (void) arc_release(db->db_buf, db);
1420 * We already have a dirty record for this TXG, and we are being
1424 dbuf_redirty(dbuf_dirty_record_t *dr)
1426 dmu_buf_impl_t *db = dr->dr_dbuf;
1428 ASSERT(MUTEX_HELD(&db->db_mtx));
1430 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1432 * If this buffer has already been written out,
1433 * we now need to reset its state.
1435 dbuf_unoverride(dr);
1436 if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1437 db->db_state != DB_NOFILL) {
1438 /* Already released on initial dirty, so just thaw. */
1439 ASSERT(arc_released(db->db_buf));
1440 arc_buf_thaw(db->db_buf);
1445 dbuf_dirty_record_t *
1446 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1450 dbuf_dirty_record_t **drp, *dr;
1451 int drop_struct_lock = FALSE;
1452 int txgoff = tx->tx_txg & TXG_MASK;
1454 ASSERT(tx->tx_txg != 0);
1455 ASSERT(!refcount_is_zero(&db->db_holds));
1456 DMU_TX_DIRTY_BUF(tx, db);
1461 * Shouldn't dirty a regular buffer in syncing context. Private
1462 * objects may be dirtied in syncing context, but only if they
1463 * were already pre-dirtied in open context.
1466 if (dn->dn_objset->os_dsl_dataset != NULL) {
1467 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1470 ASSERT(!dmu_tx_is_syncing(tx) ||
1471 BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1472 DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1473 dn->dn_objset->os_dsl_dataset == NULL);
1474 if (dn->dn_objset->os_dsl_dataset != NULL)
1475 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, FTAG);
1478 * We make this assert for private objects as well, but after we
1479 * check if we're already dirty. They are allowed to re-dirty
1480 * in syncing context.
1482 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1483 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1484 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1486 mutex_enter(&db->db_mtx);
1488 * XXX make this true for indirects too? The problem is that
1489 * transactions created with dmu_tx_create_assigned() from
1490 * syncing context don't bother holding ahead.
1492 ASSERT(db->db_level != 0 ||
1493 db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1494 db->db_state == DB_NOFILL);
1496 mutex_enter(&dn->dn_mtx);
1498 * Don't set dirtyctx to SYNC if we're just modifying this as we
1499 * initialize the objset.
1501 if (dn->dn_dirtyctx == DN_UNDIRTIED) {
1502 if (dn->dn_objset->os_dsl_dataset != NULL) {
1503 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1506 if (!BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1507 dn->dn_dirtyctx = (dmu_tx_is_syncing(tx) ?
1508 DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1509 ASSERT(dn->dn_dirtyctx_firstset == NULL);
1510 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1512 if (dn->dn_objset->os_dsl_dataset != NULL) {
1513 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1517 mutex_exit(&dn->dn_mtx);
1519 if (db->db_blkid == DMU_SPILL_BLKID)
1520 dn->dn_have_spill = B_TRUE;
1523 * If this buffer is already dirty, we're done.
1525 drp = &db->db_last_dirty;
1526 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1527 db->db.db_object == DMU_META_DNODE_OBJECT);
1528 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1530 if (dr && dr->dr_txg == tx->tx_txg) {
1534 mutex_exit(&db->db_mtx);
1539 * Only valid if not already dirty.
1541 ASSERT(dn->dn_object == 0 ||
1542 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1543 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1545 ASSERT3U(dn->dn_nlevels, >, db->db_level);
1548 * We should only be dirtying in syncing context if it's the
1549 * mos or we're initializing the os or it's a special object.
1550 * However, we are allowed to dirty in syncing context provided
1551 * we already dirtied it in open context. Hence we must make
1552 * this assertion only if we're not already dirty.
1555 VERIFY3U(tx->tx_txg, <=, spa_final_dirty_txg(os->os_spa));
1557 if (dn->dn_objset->os_dsl_dataset != NULL)
1558 rrw_enter(&os->os_dsl_dataset->ds_bp_rwlock, RW_READER, FTAG);
1559 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1560 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1561 if (dn->dn_objset->os_dsl_dataset != NULL)
1562 rrw_exit(&os->os_dsl_dataset->ds_bp_rwlock, FTAG);
1564 ASSERT(db->db.db_size != 0);
1566 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1568 if (db->db_blkid != DMU_BONUS_BLKID) {
1569 dmu_objset_willuse_space(os, db->db.db_size, tx);
1573 * If this buffer is dirty in an old transaction group we need
1574 * to make a copy of it so that the changes we make in this
1575 * transaction group won't leak out when we sync the older txg.
1577 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1578 if (db->db_level == 0) {
1579 void *data_old = db->db_buf;
1581 if (db->db_state != DB_NOFILL) {
1582 if (db->db_blkid == DMU_BONUS_BLKID) {
1583 dbuf_fix_old_data(db, tx->tx_txg);
1584 data_old = db->db.db_data;
1585 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1587 * Release the data buffer from the cache so
1588 * that we can modify it without impacting
1589 * possible other users of this cached data
1590 * block. Note that indirect blocks and
1591 * private objects are not released until the
1592 * syncing state (since they are only modified
1595 arc_release(db->db_buf, db);
1596 dbuf_fix_old_data(db, tx->tx_txg);
1597 data_old = db->db_buf;
1599 ASSERT(data_old != NULL);
1601 dr->dt.dl.dr_data = data_old;
1603 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1604 list_create(&dr->dt.di.dr_children,
1605 sizeof (dbuf_dirty_record_t),
1606 offsetof(dbuf_dirty_record_t, dr_dirty_node));
1608 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL)
1609 dr->dr_accounted = db->db.db_size;
1611 dr->dr_txg = tx->tx_txg;
1616 * We could have been freed_in_flight between the dbuf_noread
1617 * and dbuf_dirty. We win, as though the dbuf_noread() had
1618 * happened after the free.
1620 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1621 db->db_blkid != DMU_SPILL_BLKID) {
1622 mutex_enter(&dn->dn_mtx);
1623 if (dn->dn_free_ranges[txgoff] != NULL) {
1624 range_tree_clear(dn->dn_free_ranges[txgoff],
1627 mutex_exit(&dn->dn_mtx);
1628 db->db_freed_in_flight = FALSE;
1632 * This buffer is now part of this txg
1634 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1635 db->db_dirtycnt += 1;
1636 ASSERT3U(db->db_dirtycnt, <=, 3);
1638 mutex_exit(&db->db_mtx);
1640 if (db->db_blkid == DMU_BONUS_BLKID ||
1641 db->db_blkid == DMU_SPILL_BLKID) {
1642 mutex_enter(&dn->dn_mtx);
1643 ASSERT(!list_link_active(&dr->dr_dirty_node));
1644 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1645 mutex_exit(&dn->dn_mtx);
1646 dnode_setdirty(dn, tx);
1652 * The dn_struct_rwlock prevents db_blkptr from changing
1653 * due to a write from syncing context completing
1654 * while we are running, so we want to acquire it before
1655 * looking at db_blkptr.
1657 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1658 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1659 drop_struct_lock = TRUE;
1663 * We need to hold the dn_struct_rwlock to make this assertion,
1664 * because it protects dn_phys / dn_next_nlevels from changing.
1666 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1667 dn->dn_phys->dn_nlevels > db->db_level ||
1668 dn->dn_next_nlevels[txgoff] > db->db_level ||
1669 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1670 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1673 * If we are overwriting a dedup BP, then unless it is snapshotted,
1674 * when we get to syncing context we will need to decrement its
1675 * refcount in the DDT. Prefetch the relevant DDT block so that
1676 * syncing context won't have to wait for the i/o.
1678 ddt_prefetch(os->os_spa, db->db_blkptr);
1680 if (db->db_level == 0) {
1681 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1682 ASSERT(dn->dn_maxblkid >= db->db_blkid);
1685 if (db->db_level+1 < dn->dn_nlevels) {
1686 dmu_buf_impl_t *parent = db->db_parent;
1687 dbuf_dirty_record_t *di;
1688 int parent_held = FALSE;
1690 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1691 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1693 parent = dbuf_hold_level(dn, db->db_level+1,
1694 db->db_blkid >> epbs, FTAG);
1695 ASSERT(parent != NULL);
1698 if (drop_struct_lock)
1699 rw_exit(&dn->dn_struct_rwlock);
1700 ASSERT3U(db->db_level+1, ==, parent->db_level);
1701 di = dbuf_dirty(parent, tx);
1703 dbuf_rele(parent, FTAG);
1705 mutex_enter(&db->db_mtx);
1707 * Since we've dropped the mutex, it's possible that
1708 * dbuf_undirty() might have changed this out from under us.
1710 if (db->db_last_dirty == dr ||
1711 dn->dn_object == DMU_META_DNODE_OBJECT) {
1712 mutex_enter(&di->dt.di.dr_mtx);
1713 ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1714 ASSERT(!list_link_active(&dr->dr_dirty_node));
1715 list_insert_tail(&di->dt.di.dr_children, dr);
1716 mutex_exit(&di->dt.di.dr_mtx);
1719 mutex_exit(&db->db_mtx);
1721 ASSERT(db->db_level+1 == dn->dn_nlevels);
1722 ASSERT(db->db_blkid < dn->dn_nblkptr);
1723 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1724 mutex_enter(&dn->dn_mtx);
1725 ASSERT(!list_link_active(&dr->dr_dirty_node));
1726 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1727 mutex_exit(&dn->dn_mtx);
1728 if (drop_struct_lock)
1729 rw_exit(&dn->dn_struct_rwlock);
1732 dnode_setdirty(dn, tx);
1738 * Undirty a buffer in the transaction group referenced by the given
1739 * transaction. Return whether this evicted the dbuf.
1742 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1745 uint64_t txg = tx->tx_txg;
1746 dbuf_dirty_record_t *dr, **drp;
1751 * Due to our use of dn_nlevels below, this can only be called
1752 * in open context, unless we are operating on the MOS.
1753 * From syncing context, dn_nlevels may be different from the
1754 * dn_nlevels used when dbuf was dirtied.
1756 ASSERT(db->db_objset ==
1757 dmu_objset_pool(db->db_objset)->dp_meta_objset ||
1758 txg != spa_syncing_txg(dmu_objset_spa(db->db_objset)));
1759 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1760 ASSERT0(db->db_level);
1761 ASSERT(MUTEX_HELD(&db->db_mtx));
1764 * If this buffer is not dirty, we're done.
1766 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1767 if (dr->dr_txg <= txg)
1769 if (dr == NULL || dr->dr_txg < txg)
1771 ASSERT(dr->dr_txg == txg);
1772 ASSERT(dr->dr_dbuf == db);
1777 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1779 ASSERT(db->db.db_size != 0);
1781 dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset),
1782 dr->dr_accounted, txg);
1787 * Note that there are three places in dbuf_dirty()
1788 * where this dirty record may be put on a list.
1789 * Make sure to do a list_remove corresponding to
1790 * every one of those list_insert calls.
1792 if (dr->dr_parent) {
1793 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1794 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1795 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1796 } else if (db->db_blkid == DMU_SPILL_BLKID ||
1797 db->db_level + 1 == dn->dn_nlevels) {
1798 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1799 mutex_enter(&dn->dn_mtx);
1800 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1801 mutex_exit(&dn->dn_mtx);
1805 if (db->db_state != DB_NOFILL) {
1806 dbuf_unoverride(dr);
1808 ASSERT(db->db_buf != NULL);
1809 ASSERT(dr->dt.dl.dr_data != NULL);
1810 if (dr->dt.dl.dr_data != db->db_buf)
1811 arc_buf_destroy(dr->dt.dl.dr_data, db);
1814 kmem_free(dr, sizeof (dbuf_dirty_record_t));
1816 ASSERT(db->db_dirtycnt > 0);
1817 db->db_dirtycnt -= 1;
1819 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1820 ASSERT(db->db_state == DB_NOFILL || arc_released(db->db_buf));
1829 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
1831 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1832 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1834 ASSERT(tx->tx_txg != 0);
1835 ASSERT(!refcount_is_zero(&db->db_holds));
1838 * Quick check for dirtyness. For already dirty blocks, this
1839 * reduces runtime of this function by >90%, and overall performance
1840 * by 50% for some workloads (e.g. file deletion with indirect blocks
1843 mutex_enter(&db->db_mtx);
1844 dbuf_dirty_record_t *dr;
1845 for (dr = db->db_last_dirty;
1846 dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) {
1848 * It's possible that it is already dirty but not cached,
1849 * because there are some calls to dbuf_dirty() that don't
1850 * go through dmu_buf_will_dirty().
1852 if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) {
1853 /* This dbuf is already dirty and cached. */
1855 mutex_exit(&db->db_mtx);
1859 mutex_exit(&db->db_mtx);
1862 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
1863 rf |= DB_RF_HAVESTRUCT;
1865 (void) dbuf_read(db, NULL, rf);
1866 (void) dbuf_dirty(db, tx);
1870 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1872 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1874 db->db_state = DB_NOFILL;
1876 dmu_buf_will_fill(db_fake, tx);
1880 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1882 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1884 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1885 ASSERT(tx->tx_txg != 0);
1886 ASSERT(db->db_level == 0);
1887 ASSERT(!refcount_is_zero(&db->db_holds));
1889 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1890 dmu_tx_private_ok(tx));
1893 (void) dbuf_dirty(db, tx);
1896 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1899 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1901 mutex_enter(&db->db_mtx);
1904 if (db->db_state == DB_FILL) {
1905 if (db->db_level == 0 && db->db_freed_in_flight) {
1906 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1907 /* we were freed while filling */
1908 /* XXX dbuf_undirty? */
1909 bzero(db->db.db_data, db->db.db_size);
1910 db->db_freed_in_flight = FALSE;
1912 db->db_state = DB_CACHED;
1913 cv_broadcast(&db->db_changed);
1915 mutex_exit(&db->db_mtx);
1919 dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data,
1920 bp_embedded_type_t etype, enum zio_compress comp,
1921 int uncompressed_size, int compressed_size, int byteorder,
1924 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
1925 struct dirty_leaf *dl;
1926 dmu_object_type_t type;
1928 if (etype == BP_EMBEDDED_TYPE_DATA) {
1929 ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset),
1930 SPA_FEATURE_EMBEDDED_DATA));
1934 type = DB_DNODE(db)->dn_type;
1937 ASSERT0(db->db_level);
1938 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1940 dmu_buf_will_not_fill(dbuf, tx);
1942 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1943 dl = &db->db_last_dirty->dt.dl;
1944 encode_embedded_bp_compressed(&dl->dr_overridden_by,
1945 data, comp, uncompressed_size, compressed_size);
1946 BPE_SET_ETYPE(&dl->dr_overridden_by, etype);
1947 BP_SET_TYPE(&dl->dr_overridden_by, type);
1948 BP_SET_LEVEL(&dl->dr_overridden_by, 0);
1949 BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder);
1951 dl->dr_override_state = DR_OVERRIDDEN;
1952 dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg;
1956 * Directly assign a provided arc buf to a given dbuf if it's not referenced
1957 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1960 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
1962 ASSERT(!refcount_is_zero(&db->db_holds));
1963 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1964 ASSERT(db->db_level == 0);
1965 ASSERT3U(dbuf_is_metadata(db), ==, arc_is_metadata(buf));
1966 ASSERT(buf != NULL);
1967 ASSERT(arc_buf_lsize(buf) == db->db.db_size);
1968 ASSERT(tx->tx_txg != 0);
1970 arc_return_buf(buf, db);
1971 ASSERT(arc_released(buf));
1973 mutex_enter(&db->db_mtx);
1975 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1976 cv_wait(&db->db_changed, &db->db_mtx);
1978 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
1980 if (db->db_state == DB_CACHED &&
1981 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
1982 mutex_exit(&db->db_mtx);
1983 (void) dbuf_dirty(db, tx);
1984 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
1985 arc_buf_destroy(buf, db);
1986 xuio_stat_wbuf_copied();
1990 xuio_stat_wbuf_nocopy();
1991 if (db->db_state == DB_CACHED) {
1992 dbuf_dirty_record_t *dr = db->db_last_dirty;
1994 ASSERT(db->db_buf != NULL);
1995 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
1996 ASSERT(dr->dt.dl.dr_data == db->db_buf);
1997 if (!arc_released(db->db_buf)) {
1998 ASSERT(dr->dt.dl.dr_override_state ==
2000 arc_release(db->db_buf, db);
2002 dr->dt.dl.dr_data = buf;
2003 arc_buf_destroy(db->db_buf, db);
2004 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
2005 arc_release(db->db_buf, db);
2006 arc_buf_destroy(db->db_buf, db);
2010 ASSERT(db->db_buf == NULL);
2011 dbuf_set_data(db, buf);
2012 db->db_state = DB_FILL;
2013 mutex_exit(&db->db_mtx);
2014 (void) dbuf_dirty(db, tx);
2015 dmu_buf_fill_done(&db->db, tx);
2019 dbuf_destroy(dmu_buf_impl_t *db)
2022 dmu_buf_impl_t *parent = db->db_parent;
2023 dmu_buf_impl_t *dndb;
2025 ASSERT(MUTEX_HELD(&db->db_mtx));
2026 ASSERT(refcount_is_zero(&db->db_holds));
2028 if (db->db_buf != NULL) {
2029 arc_buf_destroy(db->db_buf, db);
2033 if (db->db_blkid == DMU_BONUS_BLKID) {
2034 ASSERT(db->db.db_data != NULL);
2035 zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
2036 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
2037 db->db_state = DB_UNCACHED;
2040 dbuf_clear_data(db);
2042 if (multilist_link_active(&db->db_cache_link)) {
2043 multilist_remove(dbuf_cache, db);
2044 (void) refcount_remove_many(&dbuf_cache_size,
2045 db->db.db_size, db);
2048 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
2049 ASSERT(db->db_data_pending == NULL);
2051 db->db_state = DB_EVICTING;
2052 db->db_blkptr = NULL;
2055 * Now that db_state is DB_EVICTING, nobody else can find this via
2056 * the hash table. We can now drop db_mtx, which allows us to
2057 * acquire the dn_dbufs_mtx.
2059 mutex_exit(&db->db_mtx);
2064 if (db->db_blkid != DMU_BONUS_BLKID) {
2065 boolean_t needlock = !MUTEX_HELD(&dn->dn_dbufs_mtx);
2067 mutex_enter(&dn->dn_dbufs_mtx);
2068 avl_remove(&dn->dn_dbufs, db);
2069 atomic_dec_32(&dn->dn_dbufs_count);
2073 mutex_exit(&dn->dn_dbufs_mtx);
2075 * Decrementing the dbuf count means that the hold corresponding
2076 * to the removed dbuf is no longer discounted in dnode_move(),
2077 * so the dnode cannot be moved until after we release the hold.
2078 * The membar_producer() ensures visibility of the decremented
2079 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
2083 db->db_dnode_handle = NULL;
2085 dbuf_hash_remove(db);
2090 ASSERT(refcount_is_zero(&db->db_holds));
2092 db->db_parent = NULL;
2094 ASSERT(db->db_buf == NULL);
2095 ASSERT(db->db.db_data == NULL);
2096 ASSERT(db->db_hash_next == NULL);
2097 ASSERT(db->db_blkptr == NULL);
2098 ASSERT(db->db_data_pending == NULL);
2099 ASSERT(!multilist_link_active(&db->db_cache_link));
2101 kmem_cache_free(dbuf_kmem_cache, db);
2102 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2105 * If this dbuf is referenced from an indirect dbuf,
2106 * decrement the ref count on the indirect dbuf.
2108 if (parent && parent != dndb)
2109 dbuf_rele(parent, db);
2113 * Note: While bpp will always be updated if the function returns success,
2114 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
2115 * this happens when the dnode is the meta-dnode, or a userused or groupused
2119 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
2120 dmu_buf_impl_t **parentp, blkptr_t **bpp)
2125 ASSERT(blkid != DMU_BONUS_BLKID);
2127 if (blkid == DMU_SPILL_BLKID) {
2128 mutex_enter(&dn->dn_mtx);
2129 if (dn->dn_have_spill &&
2130 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
2131 *bpp = &dn->dn_phys->dn_spill;
2134 dbuf_add_ref(dn->dn_dbuf, NULL);
2135 *parentp = dn->dn_dbuf;
2136 mutex_exit(&dn->dn_mtx);
2141 (dn->dn_phys->dn_nlevels == 0) ? 1 : dn->dn_phys->dn_nlevels;
2142 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2144 ASSERT3U(level * epbs, <, 64);
2145 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2147 * This assertion shouldn't trip as long as the max indirect block size
2148 * is less than 1M. The reason for this is that up to that point,
2149 * the number of levels required to address an entire object with blocks
2150 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In
2151 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55
2152 * (i.e. we can address the entire object), objects will all use at most
2153 * N-1 levels and the assertion won't overflow. However, once epbs is
2154 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be
2155 * enough to address an entire object, so objects will have 5 levels,
2156 * but then this assertion will overflow.
2158 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we
2159 * need to redo this logic to handle overflows.
2161 ASSERT(level >= nlevels ||
2162 ((nlevels - level - 1) * epbs) +
2163 highbit64(dn->dn_phys->dn_nblkptr) <= 64);
2164 if (level >= nlevels ||
2165 blkid >= ((uint64_t)dn->dn_phys->dn_nblkptr <<
2166 ((nlevels - level - 1) * epbs)) ||
2168 blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
2169 /* the buffer has no parent yet */
2170 return (SET_ERROR(ENOENT));
2171 } else if (level < nlevels-1) {
2172 /* this block is referenced from an indirect block */
2173 int err = dbuf_hold_impl(dn, level+1,
2174 blkid >> epbs, fail_sparse, FALSE, NULL, parentp);
2177 err = dbuf_read(*parentp, NULL,
2178 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
2180 dbuf_rele(*parentp, NULL);
2184 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
2185 (blkid & ((1ULL << epbs) - 1));
2186 if (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))
2187 ASSERT(BP_IS_HOLE(*bpp));
2190 /* the block is referenced from the dnode */
2191 ASSERT3U(level, ==, nlevels-1);
2192 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
2193 blkid < dn->dn_phys->dn_nblkptr);
2195 dbuf_add_ref(dn->dn_dbuf, NULL);
2196 *parentp = dn->dn_dbuf;
2198 *bpp = &dn->dn_phys->dn_blkptr[blkid];
2203 static dmu_buf_impl_t *
2204 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
2205 dmu_buf_impl_t *parent, blkptr_t *blkptr)
2207 objset_t *os = dn->dn_objset;
2208 dmu_buf_impl_t *db, *odb;
2210 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2211 ASSERT(dn->dn_type != DMU_OT_NONE);
2213 db = kmem_cache_alloc(dbuf_kmem_cache, KM_SLEEP);
2216 db->db.db_object = dn->dn_object;
2217 db->db_level = level;
2218 db->db_blkid = blkid;
2219 db->db_last_dirty = NULL;
2220 db->db_dirtycnt = 0;
2221 db->db_dnode_handle = dn->dn_handle;
2222 db->db_parent = parent;
2223 db->db_blkptr = blkptr;
2226 db->db_user_immediate_evict = FALSE;
2227 db->db_freed_in_flight = FALSE;
2228 db->db_pending_evict = FALSE;
2230 if (blkid == DMU_BONUS_BLKID) {
2231 ASSERT3P(parent, ==, dn->dn_dbuf);
2232 db->db.db_size = DN_MAX_BONUSLEN -
2233 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
2234 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
2235 db->db.db_offset = DMU_BONUS_BLKID;
2236 db->db_state = DB_UNCACHED;
2237 /* the bonus dbuf is not placed in the hash table */
2238 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2240 } else if (blkid == DMU_SPILL_BLKID) {
2241 db->db.db_size = (blkptr != NULL) ?
2242 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
2243 db->db.db_offset = 0;
2246 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
2247 db->db.db_size = blocksize;
2248 db->db.db_offset = db->db_blkid * blocksize;
2252 * Hold the dn_dbufs_mtx while we get the new dbuf
2253 * in the hash table *and* added to the dbufs list.
2254 * This prevents a possible deadlock with someone
2255 * trying to look up this dbuf before its added to the
2258 mutex_enter(&dn->dn_dbufs_mtx);
2259 db->db_state = DB_EVICTING;
2260 if ((odb = dbuf_hash_insert(db)) != NULL) {
2261 /* someone else inserted it first */
2262 kmem_cache_free(dbuf_kmem_cache, db);
2263 mutex_exit(&dn->dn_dbufs_mtx);
2266 avl_add(&dn->dn_dbufs, db);
2268 db->db_state = DB_UNCACHED;
2269 mutex_exit(&dn->dn_dbufs_mtx);
2270 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2272 if (parent && parent != dn->dn_dbuf)
2273 dbuf_add_ref(parent, db);
2275 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
2276 refcount_count(&dn->dn_holds) > 0);
2277 (void) refcount_add(&dn->dn_holds, db);
2278 atomic_inc_32(&dn->dn_dbufs_count);
2280 dprintf_dbuf(db, "db=%p\n", db);
2285 typedef struct dbuf_prefetch_arg {
2286 spa_t *dpa_spa; /* The spa to issue the prefetch in. */
2287 zbookmark_phys_t dpa_zb; /* The target block to prefetch. */
2288 int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */
2289 int dpa_curlevel; /* The current level that we're reading */
2290 dnode_t *dpa_dnode; /* The dnode associated with the prefetch */
2291 zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */
2292 zio_t *dpa_zio; /* The parent zio_t for all prefetches. */
2293 arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */
2294 } dbuf_prefetch_arg_t;
2297 * Actually issue the prefetch read for the block given.
2300 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp)
2302 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
2305 arc_flags_t aflags =
2306 dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
2308 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2309 ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level);
2310 ASSERT(dpa->dpa_zio != NULL);
2311 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL,
2312 dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2313 &aflags, &dpa->dpa_zb);
2317 * Called when an indirect block above our prefetch target is read in. This
2318 * will either read in the next indirect block down the tree or issue the actual
2319 * prefetch if the next block down is our target.
2322 dbuf_prefetch_indirect_done(zio_t *zio, arc_buf_t *abuf, void *private)
2324 dbuf_prefetch_arg_t *dpa = private;
2326 ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel);
2327 ASSERT3S(dpa->dpa_curlevel, >, 0);
2330 * The dpa_dnode is only valid if we are called with a NULL
2331 * zio. This indicates that the arc_read() returned without
2332 * first calling zio_read() to issue a physical read. Once
2333 * a physical read is made the dpa_dnode must be invalidated
2334 * as the locks guarding it may have been dropped. If the
2335 * dpa_dnode is still valid, then we want to add it to the dbuf
2336 * cache. To do so, we must hold the dbuf associated with the block
2337 * we just prefetched, read its contents so that we associate it
2338 * with an arc_buf_t, and then release it.
2341 ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel);
2342 if (zio->io_flags & ZIO_FLAG_RAW) {
2343 ASSERT3U(BP_GET_PSIZE(zio->io_bp), ==, zio->io_size);
2345 ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size);
2347 ASSERT3P(zio->io_spa, ==, dpa->dpa_spa);
2349 dpa->dpa_dnode = NULL;
2350 } else if (dpa->dpa_dnode != NULL) {
2351 uint64_t curblkid = dpa->dpa_zb.zb_blkid >>
2352 (dpa->dpa_epbs * (dpa->dpa_curlevel -
2353 dpa->dpa_zb.zb_level));
2354 dmu_buf_impl_t *db = dbuf_hold_level(dpa->dpa_dnode,
2355 dpa->dpa_curlevel, curblkid, FTAG);
2356 (void) dbuf_read(db, NULL,
2357 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_HAVESTRUCT);
2358 dbuf_rele(db, FTAG);
2361 dpa->dpa_curlevel--;
2363 uint64_t nextblkid = dpa->dpa_zb.zb_blkid >>
2364 (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level));
2365 blkptr_t *bp = ((blkptr_t *)abuf->b_data) +
2366 P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs);
2367 if (BP_IS_HOLE(bp) || (zio != NULL && zio->io_error != 0)) {
2368 kmem_free(dpa, sizeof (*dpa));
2369 } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) {
2370 ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid);
2371 dbuf_issue_final_prefetch(dpa, bp);
2372 kmem_free(dpa, sizeof (*dpa));
2374 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2375 zbookmark_phys_t zb;
2377 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2378 if (dpa->dpa_aflags & ARC_FLAG_L2CACHE)
2379 iter_aflags |= ARC_FLAG_L2CACHE;
2381 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2383 SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset,
2384 dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid);
2386 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2387 bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio,
2388 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2392 arc_buf_destroy(abuf, private);
2396 * Issue prefetch reads for the given block on the given level. If the indirect
2397 * blocks above that block are not in memory, we will read them in
2398 * asynchronously. As a result, this call never blocks waiting for a read to
2402 dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio,
2406 int epbs, nlevels, curlevel;
2409 ASSERT(blkid != DMU_BONUS_BLKID);
2410 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2412 if (blkid > dn->dn_maxblkid)
2415 if (dnode_block_freed(dn, blkid))
2419 * This dnode hasn't been written to disk yet, so there's nothing to
2422 nlevels = dn->dn_phys->dn_nlevels;
2423 if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0)
2426 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2427 if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level))
2430 dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object,
2433 mutex_exit(&db->db_mtx);
2435 * This dbuf already exists. It is either CACHED, or
2436 * (we assume) about to be read or filled.
2442 * Find the closest ancestor (indirect block) of the target block
2443 * that is present in the cache. In this indirect block, we will
2444 * find the bp that is at curlevel, curblkid.
2448 while (curlevel < nlevels - 1) {
2449 int parent_level = curlevel + 1;
2450 uint64_t parent_blkid = curblkid >> epbs;
2453 if (dbuf_hold_impl(dn, parent_level, parent_blkid,
2454 FALSE, TRUE, FTAG, &db) == 0) {
2455 blkptr_t *bpp = db->db_buf->b_data;
2456 bp = bpp[P2PHASE(curblkid, 1 << epbs)];
2457 dbuf_rele(db, FTAG);
2461 curlevel = parent_level;
2462 curblkid = parent_blkid;
2465 if (curlevel == nlevels - 1) {
2466 /* No cached indirect blocks found. */
2467 ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr);
2468 bp = dn->dn_phys->dn_blkptr[curblkid];
2470 if (BP_IS_HOLE(&bp))
2473 ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp));
2475 zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL,
2478 dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP);
2479 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
2480 SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2481 dn->dn_object, level, blkid);
2482 dpa->dpa_curlevel = curlevel;
2483 dpa->dpa_prio = prio;
2484 dpa->dpa_aflags = aflags;
2485 dpa->dpa_spa = dn->dn_objset->os_spa;
2486 dpa->dpa_dnode = dn;
2487 dpa->dpa_epbs = epbs;
2490 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2491 if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level))
2492 dpa->dpa_aflags |= ARC_FLAG_L2CACHE;
2495 * If we have the indirect just above us, no need to do the asynchronous
2496 * prefetch chain; we'll just run the last step ourselves. If we're at
2497 * a higher level, though, we want to issue the prefetches for all the
2498 * indirect blocks asynchronously, so we can go on with whatever we were
2501 if (curlevel == level) {
2502 ASSERT3U(curblkid, ==, blkid);
2503 dbuf_issue_final_prefetch(dpa, &bp);
2504 kmem_free(dpa, sizeof (*dpa));
2506 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2507 zbookmark_phys_t zb;
2509 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2510 if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level))
2511 iter_aflags |= ARC_FLAG_L2CACHE;
2513 SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2514 dn->dn_object, curlevel, curblkid);
2515 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2516 &bp, dbuf_prefetch_indirect_done, dpa, prio,
2517 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2521 * We use pio here instead of dpa_zio since it's possible that
2522 * dpa may have already been freed.
2528 * Returns with db_holds incremented, and db_mtx not held.
2529 * Note: dn_struct_rwlock must be held.
2532 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid,
2533 boolean_t fail_sparse, boolean_t fail_uncached,
2534 void *tag, dmu_buf_impl_t **dbp)
2536 dmu_buf_impl_t *db, *parent = NULL;
2538 ASSERT(blkid != DMU_BONUS_BLKID);
2539 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2540 ASSERT3U(dn->dn_nlevels, >, level);
2544 /* dbuf_find() returns with db_mtx held */
2545 db = dbuf_find(dn->dn_objset, dn->dn_object, level, blkid);
2548 blkptr_t *bp = NULL;
2552 return (SET_ERROR(ENOENT));
2554 ASSERT3P(parent, ==, NULL);
2555 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
2557 if (err == 0 && bp && BP_IS_HOLE(bp))
2558 err = SET_ERROR(ENOENT);
2561 dbuf_rele(parent, NULL);
2565 if (err && err != ENOENT)
2567 db = dbuf_create(dn, level, blkid, parent, bp);
2570 if (fail_uncached && db->db_state != DB_CACHED) {
2571 mutex_exit(&db->db_mtx);
2572 return (SET_ERROR(ENOENT));
2575 if (db->db_buf != NULL)
2576 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
2578 ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
2581 * If this buffer is currently syncing out, and we are are
2582 * still referencing it from db_data, we need to make a copy
2583 * of it in case we decide we want to dirty it again in this txg.
2585 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
2586 dn->dn_object != DMU_META_DNODE_OBJECT &&
2587 db->db_state == DB_CACHED && db->db_data_pending) {
2588 dbuf_dirty_record_t *dr = db->db_data_pending;
2590 if (dr->dt.dl.dr_data == db->db_buf) {
2591 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2594 arc_alloc_buf(dn->dn_objset->os_spa, db, type,
2596 bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
2601 if (multilist_link_active(&db->db_cache_link)) {
2602 ASSERT(refcount_is_zero(&db->db_holds));
2603 multilist_remove(dbuf_cache, db);
2604 (void) refcount_remove_many(&dbuf_cache_size,
2605 db->db.db_size, db);
2607 (void) refcount_add(&db->db_holds, tag);
2609 mutex_exit(&db->db_mtx);
2611 /* NOTE: we can't rele the parent until after we drop the db_mtx */
2613 dbuf_rele(parent, NULL);
2615 ASSERT3P(DB_DNODE(db), ==, dn);
2616 ASSERT3U(db->db_blkid, ==, blkid);
2617 ASSERT3U(db->db_level, ==, level);
2624 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
2626 return (dbuf_hold_level(dn, 0, blkid, tag));
2630 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
2633 int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db);
2634 return (err ? NULL : db);
2638 dbuf_create_bonus(dnode_t *dn)
2640 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
2642 ASSERT(dn->dn_bonus == NULL);
2643 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
2647 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
2649 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2652 if (db->db_blkid != DMU_SPILL_BLKID)
2653 return (SET_ERROR(ENOTSUP));
2655 blksz = SPA_MINBLOCKSIZE;
2656 ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset)));
2657 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
2661 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2662 dbuf_new_size(db, blksz, tx);
2663 rw_exit(&dn->dn_struct_rwlock);
2670 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
2672 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
2675 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2677 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
2679 int64_t holds = refcount_add(&db->db_holds, tag);
2680 ASSERT3S(holds, >, 1);
2683 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
2685 dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid,
2688 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2689 dmu_buf_impl_t *found_db;
2690 boolean_t result = B_FALSE;
2692 if (db->db_blkid == DMU_BONUS_BLKID)
2693 found_db = dbuf_find_bonus(os, obj);
2695 found_db = dbuf_find(os, obj, 0, blkid);
2697 if (found_db != NULL) {
2698 if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) {
2699 (void) refcount_add(&db->db_holds, tag);
2702 mutex_exit(&db->db_mtx);
2708 * If you call dbuf_rele() you had better not be referencing the dnode handle
2709 * unless you have some other direct or indirect hold on the dnode. (An indirect
2710 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2711 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2712 * dnode's parent dbuf evicting its dnode handles.
2715 dbuf_rele(dmu_buf_impl_t *db, void *tag)
2717 mutex_enter(&db->db_mtx);
2718 dbuf_rele_and_unlock(db, tag);
2722 dmu_buf_rele(dmu_buf_t *db, void *tag)
2724 dbuf_rele((dmu_buf_impl_t *)db, tag);
2728 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2729 * db_dirtycnt and db_holds to be updated atomically.
2732 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2736 ASSERT(MUTEX_HELD(&db->db_mtx));
2740 * Remove the reference to the dbuf before removing its hold on the
2741 * dnode so we can guarantee in dnode_move() that a referenced bonus
2742 * buffer has a corresponding dnode hold.
2744 holds = refcount_remove(&db->db_holds, tag);
2748 * We can't freeze indirects if there is a possibility that they
2749 * may be modified in the current syncing context.
2751 if (db->db_buf != NULL &&
2752 holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) {
2753 arc_buf_freeze(db->db_buf);
2756 if (holds == db->db_dirtycnt &&
2757 db->db_level == 0 && db->db_user_immediate_evict)
2758 dbuf_evict_user(db);
2761 if (db->db_blkid == DMU_BONUS_BLKID) {
2763 boolean_t evict_dbuf = db->db_pending_evict;
2766 * If the dnode moves here, we cannot cross this
2767 * barrier until the move completes.
2772 atomic_dec_32(&dn->dn_dbufs_count);
2775 * Decrementing the dbuf count means that the bonus
2776 * buffer's dnode hold is no longer discounted in
2777 * dnode_move(). The dnode cannot move until after
2778 * the dnode_rele() below.
2783 * Do not reference db after its lock is dropped.
2784 * Another thread may evict it.
2786 mutex_exit(&db->db_mtx);
2789 dnode_evict_bonus(dn);
2792 } else if (db->db_buf == NULL) {
2794 * This is a special case: we never associated this
2795 * dbuf with any data allocated from the ARC.
2797 ASSERT(db->db_state == DB_UNCACHED ||
2798 db->db_state == DB_NOFILL);
2800 } else if (arc_released(db->db_buf)) {
2802 * This dbuf has anonymous data associated with it.
2806 boolean_t do_arc_evict = B_FALSE;
2808 spa_t *spa = dmu_objset_spa(db->db_objset);
2810 if (!DBUF_IS_CACHEABLE(db) &&
2811 db->db_blkptr != NULL &&
2812 !BP_IS_HOLE(db->db_blkptr) &&
2813 !BP_IS_EMBEDDED(db->db_blkptr)) {
2814 do_arc_evict = B_TRUE;
2815 bp = *db->db_blkptr;
2818 if (!DBUF_IS_CACHEABLE(db) ||
2819 db->db_pending_evict) {
2821 } else if (!multilist_link_active(&db->db_cache_link)) {
2822 multilist_insert(dbuf_cache, db);
2823 (void) refcount_add_many(&dbuf_cache_size,
2824 db->db.db_size, db);
2825 mutex_exit(&db->db_mtx);
2827 dbuf_evict_notify();
2831 arc_freed(spa, &bp);
2834 mutex_exit(&db->db_mtx);
2839 #pragma weak dmu_buf_refcount = dbuf_refcount
2841 dbuf_refcount(dmu_buf_impl_t *db)
2843 return (refcount_count(&db->db_holds));
2847 dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user,
2848 dmu_buf_user_t *new_user)
2850 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2852 mutex_enter(&db->db_mtx);
2853 dbuf_verify_user(db, DBVU_NOT_EVICTING);
2854 if (db->db_user == old_user)
2855 db->db_user = new_user;
2857 old_user = db->db_user;
2858 dbuf_verify_user(db, DBVU_NOT_EVICTING);
2859 mutex_exit(&db->db_mtx);
2865 dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2867 return (dmu_buf_replace_user(db_fake, NULL, user));
2871 dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2873 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2875 db->db_user_immediate_evict = TRUE;
2876 return (dmu_buf_set_user(db_fake, user));
2880 dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2882 return (dmu_buf_replace_user(db_fake, user, NULL));
2886 dmu_buf_get_user(dmu_buf_t *db_fake)
2888 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2890 dbuf_verify_user(db, DBVU_NOT_EVICTING);
2891 return (db->db_user);
2895 dmu_buf_user_evict_wait()
2897 taskq_wait(dbu_evict_taskq);
2901 dmu_buf_get_blkptr(dmu_buf_t *db)
2903 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2904 return (dbi->db_blkptr);
2908 dmu_buf_get_objset(dmu_buf_t *db)
2910 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2911 return (dbi->db_objset);
2915 dmu_buf_dnode_enter(dmu_buf_t *db)
2917 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2918 DB_DNODE_ENTER(dbi);
2919 return (DB_DNODE(dbi));
2923 dmu_buf_dnode_exit(dmu_buf_t *db)
2925 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2930 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
2932 /* ASSERT(dmu_tx_is_syncing(tx) */
2933 ASSERT(MUTEX_HELD(&db->db_mtx));
2935 if (db->db_blkptr != NULL)
2938 if (db->db_blkid == DMU_SPILL_BLKID) {
2939 db->db_blkptr = &dn->dn_phys->dn_spill;
2940 BP_ZERO(db->db_blkptr);
2943 if (db->db_level == dn->dn_phys->dn_nlevels-1) {
2945 * This buffer was allocated at a time when there was
2946 * no available blkptrs from the dnode, or it was
2947 * inappropriate to hook it in (i.e., nlevels mis-match).
2949 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
2950 ASSERT(db->db_parent == NULL);
2951 db->db_parent = dn->dn_dbuf;
2952 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
2955 dmu_buf_impl_t *parent = db->db_parent;
2956 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2958 ASSERT(dn->dn_phys->dn_nlevels > 1);
2959 if (parent == NULL) {
2960 mutex_exit(&db->db_mtx);
2961 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2962 parent = dbuf_hold_level(dn, db->db_level + 1,
2963 db->db_blkid >> epbs, db);
2964 rw_exit(&dn->dn_struct_rwlock);
2965 mutex_enter(&db->db_mtx);
2966 db->db_parent = parent;
2968 db->db_blkptr = (blkptr_t *)parent->db.db_data +
2969 (db->db_blkid & ((1ULL << epbs) - 1));
2975 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2977 dmu_buf_impl_t *db = dr->dr_dbuf;
2981 ASSERT(dmu_tx_is_syncing(tx));
2983 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2985 mutex_enter(&db->db_mtx);
2987 ASSERT(db->db_level > 0);
2990 /* Read the block if it hasn't been read yet. */
2991 if (db->db_buf == NULL) {
2992 mutex_exit(&db->db_mtx);
2993 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
2994 mutex_enter(&db->db_mtx);
2996 ASSERT3U(db->db_state, ==, DB_CACHED);
2997 ASSERT(db->db_buf != NULL);
3001 /* Indirect block size must match what the dnode thinks it is. */
3002 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3003 dbuf_check_blkptr(dn, db);
3006 /* Provide the pending dirty record to child dbufs */
3007 db->db_data_pending = dr;
3009 mutex_exit(&db->db_mtx);
3010 dbuf_write(dr, db->db_buf, tx);
3013 mutex_enter(&dr->dt.di.dr_mtx);
3014 dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx);
3015 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3016 mutex_exit(&dr->dt.di.dr_mtx);
3021 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
3023 arc_buf_t **datap = &dr->dt.dl.dr_data;
3024 dmu_buf_impl_t *db = dr->dr_dbuf;
3027 uint64_t txg = tx->tx_txg;
3029 ASSERT(dmu_tx_is_syncing(tx));
3031 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
3033 mutex_enter(&db->db_mtx);
3035 * To be synced, we must be dirtied. But we
3036 * might have been freed after the dirty.
3038 if (db->db_state == DB_UNCACHED) {
3039 /* This buffer has been freed since it was dirtied */
3040 ASSERT(db->db.db_data == NULL);
3041 } else if (db->db_state == DB_FILL) {
3042 /* This buffer was freed and is now being re-filled */
3043 ASSERT(db->db.db_data != dr->dt.dl.dr_data);
3045 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
3052 if (db->db_blkid == DMU_SPILL_BLKID) {
3053 mutex_enter(&dn->dn_mtx);
3054 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
3055 mutex_exit(&dn->dn_mtx);
3059 * If this is a bonus buffer, simply copy the bonus data into the
3060 * dnode. It will be written out when the dnode is synced (and it
3061 * will be synced, since it must have been dirty for dbuf_sync to
3064 if (db->db_blkid == DMU_BONUS_BLKID) {
3065 dbuf_dirty_record_t **drp;
3067 ASSERT(*datap != NULL);
3068 ASSERT0(db->db_level);
3069 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN);
3070 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen);
3073 if (*datap != db->db.db_data) {
3074 zio_buf_free(*datap, DN_MAX_BONUSLEN);
3075 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
3077 db->db_data_pending = NULL;
3078 drp = &db->db_last_dirty;
3080 drp = &(*drp)->dr_next;
3081 ASSERT(dr->dr_next == NULL);
3082 ASSERT(dr->dr_dbuf == db);
3084 kmem_free(dr, sizeof (dbuf_dirty_record_t));
3085 ASSERT(db->db_dirtycnt > 0);
3086 db->db_dirtycnt -= 1;
3087 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
3094 * This function may have dropped the db_mtx lock allowing a dmu_sync
3095 * operation to sneak in. As a result, we need to ensure that we
3096 * don't check the dr_override_state until we have returned from
3097 * dbuf_check_blkptr.
3099 dbuf_check_blkptr(dn, db);
3102 * If this buffer is in the middle of an immediate write,
3103 * wait for the synchronous IO to complete.
3105 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
3106 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
3107 cv_wait(&db->db_changed, &db->db_mtx);
3108 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
3111 if (db->db_state != DB_NOFILL &&
3112 dn->dn_object != DMU_META_DNODE_OBJECT &&
3113 refcount_count(&db->db_holds) > 1 &&
3114 dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
3115 *datap == db->db_buf) {
3117 * If this buffer is currently "in use" (i.e., there
3118 * are active holds and db_data still references it),
3119 * then make a copy before we start the write so that
3120 * any modifications from the open txg will not leak
3123 * NOTE: this copy does not need to be made for
3124 * objects only modified in the syncing context (e.g.
3125 * DNONE_DNODE blocks).
3127 int psize = arc_buf_size(*datap);
3128 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
3129 enum zio_compress compress_type = arc_get_compression(*datap);
3131 if (compress_type == ZIO_COMPRESS_OFF) {
3132 *datap = arc_alloc_buf(os->os_spa, db, type, psize);
3134 ASSERT3U(type, ==, ARC_BUFC_DATA);
3135 int lsize = arc_buf_lsize(*datap);
3136 *datap = arc_alloc_compressed_buf(os->os_spa, db,
3137 psize, lsize, compress_type);
3139 bcopy(db->db.db_data, (*datap)->b_data, psize);
3141 db->db_data_pending = dr;
3143 mutex_exit(&db->db_mtx);
3145 dbuf_write(dr, *datap, tx);
3147 ASSERT(!list_link_active(&dr->dr_dirty_node));
3148 if (dn->dn_object == DMU_META_DNODE_OBJECT) {
3149 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
3153 * Although zio_nowait() does not "wait for an IO", it does
3154 * initiate the IO. If this is an empty write it seems plausible
3155 * that the IO could actually be completed before the nowait
3156 * returns. We need to DB_DNODE_EXIT() first in case
3157 * zio_nowait() invalidates the dbuf.
3160 zio_nowait(dr->dr_zio);
3165 dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx)
3167 dbuf_dirty_record_t *dr;
3169 while (dr = list_head(list)) {
3170 if (dr->dr_zio != NULL) {
3172 * If we find an already initialized zio then we
3173 * are processing the meta-dnode, and we have finished.
3174 * The dbufs for all dnodes are put back on the list
3175 * during processing, so that we can zio_wait()
3176 * these IOs after initiating all child IOs.
3178 ASSERT3U(dr->dr_dbuf->db.db_object, ==,
3179 DMU_META_DNODE_OBJECT);
3182 if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
3183 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
3184 VERIFY3U(dr->dr_dbuf->db_level, ==, level);
3186 list_remove(list, dr);
3187 if (dr->dr_dbuf->db_level > 0)
3188 dbuf_sync_indirect(dr, tx);
3190 dbuf_sync_leaf(dr, tx);
3196 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3198 dmu_buf_impl_t *db = vdb;
3200 blkptr_t *bp = zio->io_bp;
3201 blkptr_t *bp_orig = &zio->io_bp_orig;
3202 spa_t *spa = zio->io_spa;
3207 ASSERT3P(db->db_blkptr, !=, NULL);
3208 ASSERT3P(&db->db_data_pending->dr_bp_copy, ==, bp);
3212 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
3213 dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
3214 zio->io_prev_space_delta = delta;
3216 if (bp->blk_birth != 0) {
3217 ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
3218 BP_GET_TYPE(bp) == dn->dn_type) ||
3219 (db->db_blkid == DMU_SPILL_BLKID &&
3220 BP_GET_TYPE(bp) == dn->dn_bonustype) ||
3221 BP_IS_EMBEDDED(bp));
3222 ASSERT(BP_GET_LEVEL(bp) == db->db_level);
3225 mutex_enter(&db->db_mtx);
3228 if (db->db_blkid == DMU_SPILL_BLKID) {
3229 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3230 ASSERT(!(BP_IS_HOLE(bp)) &&
3231 db->db_blkptr == &dn->dn_phys->dn_spill);
3235 if (db->db_level == 0) {
3236 mutex_enter(&dn->dn_mtx);
3237 if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
3238 db->db_blkid != DMU_SPILL_BLKID)
3239 dn->dn_phys->dn_maxblkid = db->db_blkid;
3240 mutex_exit(&dn->dn_mtx);
3242 if (dn->dn_type == DMU_OT_DNODE) {
3243 dnode_phys_t *dnp = db->db.db_data;
3244 for (i = db->db.db_size >> DNODE_SHIFT; i > 0;
3246 if (dnp->dn_type != DMU_OT_NONE)
3250 if (BP_IS_HOLE(bp)) {
3257 blkptr_t *ibp = db->db.db_data;
3258 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3259 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
3260 if (BP_IS_HOLE(ibp))
3262 fill += BP_GET_FILL(ibp);
3267 if (!BP_IS_EMBEDDED(bp))
3268 bp->blk_fill = fill;
3270 mutex_exit(&db->db_mtx);
3272 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3273 *db->db_blkptr = *bp;
3274 rw_exit(&dn->dn_struct_rwlock);
3279 * This function gets called just prior to running through the compression
3280 * stage of the zio pipeline. If we're an indirect block comprised of only
3281 * holes, then we want this indirect to be compressed away to a hole. In
3282 * order to do that we must zero out any information about the holes that
3283 * this indirect points to prior to before we try to compress it.
3286 dbuf_write_children_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3288 dmu_buf_impl_t *db = vdb;
3291 unsigned int epbs, i;
3293 ASSERT3U(db->db_level, >, 0);
3296 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3297 ASSERT3U(epbs, <, 31);
3299 /* Determine if all our children are holes */
3300 for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++) {
3301 if (!BP_IS_HOLE(bp))
3306 * If all the children are holes, then zero them all out so that
3307 * we may get compressed away.
3309 if (i == 1 << epbs) {
3311 * We only found holes. Grab the rwlock to prevent
3312 * anybody from reading the blocks we're about to
3315 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3316 bzero(db->db.db_data, db->db.db_size);
3317 rw_exit(&dn->dn_struct_rwlock);
3323 * The SPA will call this callback several times for each zio - once
3324 * for every physical child i/o (zio->io_phys_children times). This
3325 * allows the DMU to monitor the progress of each logical i/o. For example,
3326 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
3327 * block. There may be a long delay before all copies/fragments are completed,
3328 * so this callback allows us to retire dirty space gradually, as the physical
3333 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
3335 dmu_buf_impl_t *db = arg;
3336 objset_t *os = db->db_objset;
3337 dsl_pool_t *dp = dmu_objset_pool(os);
3338 dbuf_dirty_record_t *dr;
3341 dr = db->db_data_pending;
3342 ASSERT3U(dr->dr_txg, ==, zio->io_txg);
3345 * The callback will be called io_phys_children times. Retire one
3346 * portion of our dirty space each time we are called. Any rounding
3347 * error will be cleaned up by dsl_pool_sync()'s call to
3348 * dsl_pool_undirty_space().
3350 delta = dr->dr_accounted / zio->io_phys_children;
3351 dsl_pool_undirty_space(dp, delta, zio->io_txg);
3356 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
3358 dmu_buf_impl_t *db = vdb;
3359 blkptr_t *bp_orig = &zio->io_bp_orig;
3360 blkptr_t *bp = db->db_blkptr;
3361 objset_t *os = db->db_objset;
3362 dmu_tx_t *tx = os->os_synctx;
3363 dbuf_dirty_record_t **drp, *dr;
3365 ASSERT0(zio->io_error);
3366 ASSERT(db->db_blkptr == bp);
3369 * For nopwrites and rewrites we ensure that the bp matches our
3370 * original and bypass all the accounting.
3372 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
3373 ASSERT(BP_EQUAL(bp, bp_orig));
3375 dsl_dataset_t *ds = os->os_dsl_dataset;
3376 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
3377 dsl_dataset_block_born(ds, bp, tx);
3380 mutex_enter(&db->db_mtx);
3384 drp = &db->db_last_dirty;
3385 while ((dr = *drp) != db->db_data_pending)
3387 ASSERT(!list_link_active(&dr->dr_dirty_node));
3388 ASSERT(dr->dr_dbuf == db);
3389 ASSERT(dr->dr_next == NULL);
3393 if (db->db_blkid == DMU_SPILL_BLKID) {
3398 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3399 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
3400 db->db_blkptr == &dn->dn_phys->dn_spill);
3405 if (db->db_level == 0) {
3406 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
3407 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
3408 if (db->db_state != DB_NOFILL) {
3409 if (dr->dt.dl.dr_data != db->db_buf)
3410 arc_buf_destroy(dr->dt.dl.dr_data, db);
3417 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3418 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
3419 if (!BP_IS_HOLE(db->db_blkptr)) {
3421 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3422 ASSERT3U(db->db_blkid, <=,
3423 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs));
3424 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
3428 mutex_destroy(&dr->dt.di.dr_mtx);
3429 list_destroy(&dr->dt.di.dr_children);
3431 kmem_free(dr, sizeof (dbuf_dirty_record_t));
3433 cv_broadcast(&db->db_changed);
3434 ASSERT(db->db_dirtycnt > 0);
3435 db->db_dirtycnt -= 1;
3436 db->db_data_pending = NULL;
3437 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg);
3441 dbuf_write_nofill_ready(zio_t *zio)
3443 dbuf_write_ready(zio, NULL, zio->io_private);
3447 dbuf_write_nofill_done(zio_t *zio)
3449 dbuf_write_done(zio, NULL, zio->io_private);
3453 dbuf_write_override_ready(zio_t *zio)
3455 dbuf_dirty_record_t *dr = zio->io_private;
3456 dmu_buf_impl_t *db = dr->dr_dbuf;
3458 dbuf_write_ready(zio, NULL, db);
3462 dbuf_write_override_done(zio_t *zio)
3464 dbuf_dirty_record_t *dr = zio->io_private;
3465 dmu_buf_impl_t *db = dr->dr_dbuf;
3466 blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
3468 mutex_enter(&db->db_mtx);
3469 if (!BP_EQUAL(zio->io_bp, obp)) {
3470 if (!BP_IS_HOLE(obp))
3471 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
3472 arc_release(dr->dt.dl.dr_data, db);
3474 mutex_exit(&db->db_mtx);
3475 dbuf_write_done(zio, NULL, db);
3477 if (zio->io_abd != NULL)
3478 abd_put(zio->io_abd);
3481 /* Issue I/O to commit a dirty buffer to disk. */
3483 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
3485 dmu_buf_impl_t *db = dr->dr_dbuf;
3488 dmu_buf_impl_t *parent = db->db_parent;
3489 uint64_t txg = tx->tx_txg;
3490 zbookmark_phys_t zb;
3495 ASSERT(dmu_tx_is_syncing(tx));
3501 if (db->db_state != DB_NOFILL) {
3502 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
3504 * Private object buffers are released here rather
3505 * than in dbuf_dirty() since they are only modified
3506 * in the syncing context and we don't want the
3507 * overhead of making multiple copies of the data.
3509 if (BP_IS_HOLE(db->db_blkptr)) {
3512 dbuf_release_bp(db);
3517 if (parent != dn->dn_dbuf) {
3518 /* Our parent is an indirect block. */
3519 /* We have a dirty parent that has been scheduled for write. */
3520 ASSERT(parent && parent->db_data_pending);
3521 /* Our parent's buffer is one level closer to the dnode. */
3522 ASSERT(db->db_level == parent->db_level-1);
3524 * We're about to modify our parent's db_data by modifying
3525 * our block pointer, so the parent must be released.
3527 ASSERT(arc_released(parent->db_buf));
3528 zio = parent->db_data_pending->dr_zio;
3530 /* Our parent is the dnode itself. */
3531 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
3532 db->db_blkid != DMU_SPILL_BLKID) ||
3533 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
3534 if (db->db_blkid != DMU_SPILL_BLKID)
3535 ASSERT3P(db->db_blkptr, ==,
3536 &dn->dn_phys->dn_blkptr[db->db_blkid]);
3540 ASSERT(db->db_level == 0 || data == db->db_buf);
3541 ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
3544 SET_BOOKMARK(&zb, os->os_dsl_dataset ?
3545 os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
3546 db->db.db_object, db->db_level, db->db_blkid);
3548 if (db->db_blkid == DMU_SPILL_BLKID)
3550 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
3552 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
3556 * We copy the blkptr now (rather than when we instantiate the dirty
3557 * record), because its value can change between open context and
3558 * syncing context. We do not need to hold dn_struct_rwlock to read
3559 * db_blkptr because we are in syncing context.
3561 dr->dr_bp_copy = *db->db_blkptr;
3563 if (db->db_level == 0 &&
3564 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
3566 * The BP for this block has been provided by open context
3567 * (by dmu_sync() or dmu_buf_write_embedded()).
3569 abd_t *contents = (data != NULL) ?
3570 abd_get_from_buf(data->b_data, arc_buf_size(data)) : NULL;
3572 dr->dr_zio = zio_write(zio, os->os_spa, txg, &dr->dr_bp_copy,
3573 contents, db->db.db_size, db->db.db_size, &zp,
3574 dbuf_write_override_ready, NULL, NULL,
3575 dbuf_write_override_done,
3576 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
3577 mutex_enter(&db->db_mtx);
3578 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
3579 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
3580 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
3581 mutex_exit(&db->db_mtx);
3582 } else if (db->db_state == DB_NOFILL) {
3583 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF ||
3584 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY);
3585 dr->dr_zio = zio_write(zio, os->os_spa, txg,
3586 &dr->dr_bp_copy, NULL, db->db.db_size, db->db.db_size, &zp,
3587 dbuf_write_nofill_ready, NULL, NULL,
3588 dbuf_write_nofill_done, db,
3589 ZIO_PRIORITY_ASYNC_WRITE,
3590 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
3592 ASSERT(arc_released(data));
3595 * For indirect blocks, we want to setup the children
3596 * ready callback so that we can properly handle an indirect
3597 * block that only contains holes.
3599 arc_done_func_t *children_ready_cb = NULL;
3600 if (db->db_level != 0)
3601 children_ready_cb = dbuf_write_children_ready;
3603 dr->dr_zio = arc_write(zio, os->os_spa, txg,
3604 &dr->dr_bp_copy, data, DBUF_IS_L2CACHEABLE(db),
3605 &zp, dbuf_write_ready, children_ready_cb,
3606 dbuf_write_physdone, dbuf_write_done, db,
3607 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);