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, 2015 by Delphix. All rights reserved.
25 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
26 * Copyright (c) 2013, Joyent, Inc. All rights reserved.
27 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
28 * Copyright (c) 2014 Integros [integros.com]
31 #include <sys/zfs_context.h>
33 #include <sys/dmu_send.h>
34 #include <sys/dmu_impl.h>
36 #include <sys/dmu_objset.h>
37 #include <sys/dsl_dataset.h>
38 #include <sys/dsl_dir.h>
39 #include <sys/dmu_tx.h>
42 #include <sys/dmu_zfetch.h>
44 #include <sys/sa_impl.h>
45 #include <sys/zfeature.h>
46 #include <sys/blkptr.h>
47 #include <sys/range_tree.h>
48 #include <sys/callb.h>
50 uint_t zfs_dbuf_evict_key;
53 * Number of times that zfs_free_range() took the slow path while doing
54 * a zfs receive. A nonzero value indicates a potential performance problem.
56 uint64_t zfs_free_range_recv_miss;
58 static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
59 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
62 * Global data structures and functions for the dbuf cache.
64 static kmem_cache_t *dbuf_kmem_cache;
65 static taskq_t *dbu_evict_taskq;
67 static kthread_t *dbuf_cache_evict_thread;
68 static kmutex_t dbuf_evict_lock;
69 static kcondvar_t dbuf_evict_cv;
70 static boolean_t dbuf_evict_thread_exit;
73 * LRU cache of dbufs. The dbuf cache maintains a list of dbufs that
74 * are not currently held but have been recently released. These dbufs
75 * are not eligible for arc eviction until they are aged out of the cache.
76 * Dbufs are added to the dbuf cache once the last hold is released. If a
77 * dbuf is later accessed and still exists in the dbuf cache, then it will
78 * be removed from the cache and later re-added to the head of the cache.
79 * Dbufs that are aged out of the cache will be immediately destroyed and
80 * become eligible for arc eviction.
82 static multilist_t dbuf_cache;
83 static refcount_t dbuf_cache_size;
84 uint64_t dbuf_cache_max_bytes = 100 * 1024 * 1024;
86 /* Cap the size of the dbuf cache to log2 fraction of arc size. */
87 int dbuf_cache_max_shift = 5;
90 * The dbuf cache uses a three-stage eviction policy:
91 * - A low water marker designates when the dbuf eviction thread
92 * should stop evicting from the dbuf cache.
93 * - When we reach the maximum size (aka mid water mark), we
94 * signal the eviction thread to run.
95 * - The high water mark indicates when the eviction thread
96 * is unable to keep up with the incoming load and eviction must
97 * happen in the context of the calling thread.
101 * low water mid water hi water
102 * +----------------------------------------+----------+----------+
107 * +----------------------------------------+----------+----------+
109 * evicting eviction directly
112 * The high and low water marks indicate the operating range for the eviction
113 * thread. The low water mark is, by default, 90% of the total size of the
114 * cache and the high water mark is at 110% (both of these percentages can be
115 * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct,
116 * respectively). The eviction thread will try to ensure that the cache remains
117 * within this range by waking up every second and checking if the cache is
118 * above the low water mark. The thread can also be woken up by callers adding
119 * elements into the cache if the cache is larger than the mid water (i.e max
120 * cache size). Once the eviction thread is woken up and eviction is required,
121 * it will continue evicting buffers until it's able to reduce the cache size
122 * to the low water mark. If the cache size continues to grow and hits the high
123 * water mark, then callers adding elments to the cache will begin to evict
124 * directly from the cache until the cache is no longer above the high water
129 * The percentage above and below the maximum cache size.
131 uint_t dbuf_cache_hiwater_pct = 10;
132 uint_t dbuf_cache_lowater_pct = 10;
136 dbuf_cons(void *vdb, void *unused, int kmflag)
138 dmu_buf_impl_t *db = vdb;
139 bzero(db, sizeof (dmu_buf_impl_t));
141 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
142 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
143 multilist_link_init(&db->db_cache_link);
144 refcount_create(&db->db_holds);
151 dbuf_dest(void *vdb, void *unused)
153 dmu_buf_impl_t *db = vdb;
154 mutex_destroy(&db->db_mtx);
155 cv_destroy(&db->db_changed);
156 ASSERT(!multilist_link_active(&db->db_cache_link));
157 refcount_destroy(&db->db_holds);
161 * dbuf hash table routines
163 static dbuf_hash_table_t dbuf_hash_table;
165 static uint64_t dbuf_hash_count;
168 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
170 uintptr_t osv = (uintptr_t)os;
171 uint64_t crc = -1ULL;
173 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
174 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF];
175 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF];
176 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF];
177 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF];
178 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF];
179 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF];
181 crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16);
186 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
187 ((dbuf)->db.db_object == (obj) && \
188 (dbuf)->db_objset == (os) && \
189 (dbuf)->db_level == (level) && \
190 (dbuf)->db_blkid == (blkid))
193 dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid)
195 dbuf_hash_table_t *h = &dbuf_hash_table;
196 uint64_t hv = dbuf_hash(os, obj, level, blkid);
197 uint64_t idx = hv & h->hash_table_mask;
200 mutex_enter(DBUF_HASH_MUTEX(h, idx));
201 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
202 if (DBUF_EQUAL(db, os, obj, level, blkid)) {
203 mutex_enter(&db->db_mtx);
204 if (db->db_state != DB_EVICTING) {
205 mutex_exit(DBUF_HASH_MUTEX(h, idx));
208 mutex_exit(&db->db_mtx);
211 mutex_exit(DBUF_HASH_MUTEX(h, idx));
215 static dmu_buf_impl_t *
216 dbuf_find_bonus(objset_t *os, uint64_t object)
219 dmu_buf_impl_t *db = NULL;
221 if (dnode_hold(os, object, FTAG, &dn) == 0) {
222 rw_enter(&dn->dn_struct_rwlock, RW_READER);
223 if (dn->dn_bonus != NULL) {
225 mutex_enter(&db->db_mtx);
227 rw_exit(&dn->dn_struct_rwlock);
228 dnode_rele(dn, FTAG);
234 * Insert an entry into the hash table. If there is already an element
235 * equal to elem in the hash table, then the already existing element
236 * will be returned and the new element will not be inserted.
237 * Otherwise returns NULL.
239 static dmu_buf_impl_t *
240 dbuf_hash_insert(dmu_buf_impl_t *db)
242 dbuf_hash_table_t *h = &dbuf_hash_table;
243 objset_t *os = db->db_objset;
244 uint64_t obj = db->db.db_object;
245 int level = db->db_level;
246 uint64_t blkid = db->db_blkid;
247 uint64_t hv = dbuf_hash(os, obj, level, blkid);
248 uint64_t idx = hv & h->hash_table_mask;
251 mutex_enter(DBUF_HASH_MUTEX(h, idx));
252 for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) {
253 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
254 mutex_enter(&dbf->db_mtx);
255 if (dbf->db_state != DB_EVICTING) {
256 mutex_exit(DBUF_HASH_MUTEX(h, idx));
259 mutex_exit(&dbf->db_mtx);
263 mutex_enter(&db->db_mtx);
264 db->db_hash_next = h->hash_table[idx];
265 h->hash_table[idx] = db;
266 mutex_exit(DBUF_HASH_MUTEX(h, idx));
267 atomic_inc_64(&dbuf_hash_count);
273 * Remove an entry from the hash table. It must be in the EVICTING state.
276 dbuf_hash_remove(dmu_buf_impl_t *db)
278 dbuf_hash_table_t *h = &dbuf_hash_table;
279 uint64_t hv = dbuf_hash(db->db_objset, db->db.db_object,
280 db->db_level, db->db_blkid);
281 uint64_t idx = hv & h->hash_table_mask;
282 dmu_buf_impl_t *dbf, **dbp;
285 * We musn't hold db_mtx to maintain lock ordering:
286 * DBUF_HASH_MUTEX > db_mtx.
288 ASSERT(refcount_is_zero(&db->db_holds));
289 ASSERT(db->db_state == DB_EVICTING);
290 ASSERT(!MUTEX_HELD(&db->db_mtx));
292 mutex_enter(DBUF_HASH_MUTEX(h, idx));
293 dbp = &h->hash_table[idx];
294 while ((dbf = *dbp) != db) {
295 dbp = &dbf->db_hash_next;
298 *dbp = db->db_hash_next;
299 db->db_hash_next = NULL;
300 mutex_exit(DBUF_HASH_MUTEX(h, idx));
301 atomic_dec_64(&dbuf_hash_count);
307 } dbvu_verify_type_t;
310 dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type)
315 if (db->db_user == NULL)
318 /* Only data blocks support the attachment of user data. */
319 ASSERT(db->db_level == 0);
321 /* Clients must resolve a dbuf before attaching user data. */
322 ASSERT(db->db.db_data != NULL);
323 ASSERT3U(db->db_state, ==, DB_CACHED);
325 holds = refcount_count(&db->db_holds);
326 if (verify_type == DBVU_EVICTING) {
328 * Immediate eviction occurs when holds == dirtycnt.
329 * For normal eviction buffers, holds is zero on
330 * eviction, except when dbuf_fix_old_data() calls
331 * dbuf_clear_data(). However, the hold count can grow
332 * during eviction even though db_mtx is held (see
333 * dmu_bonus_hold() for an example), so we can only
334 * test the generic invariant that holds >= dirtycnt.
336 ASSERT3U(holds, >=, db->db_dirtycnt);
338 if (db->db_user_immediate_evict == TRUE)
339 ASSERT3U(holds, >=, db->db_dirtycnt);
341 ASSERT3U(holds, >, 0);
347 dbuf_evict_user(dmu_buf_impl_t *db)
349 dmu_buf_user_t *dbu = db->db_user;
351 ASSERT(MUTEX_HELD(&db->db_mtx));
356 dbuf_verify_user(db, DBVU_EVICTING);
360 if (dbu->dbu_clear_on_evict_dbufp != NULL)
361 *dbu->dbu_clear_on_evict_dbufp = NULL;
365 * Invoke the callback from a taskq to avoid lock order reversals
366 * and limit stack depth.
368 taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func, dbu, 0,
373 dbuf_is_metadata(dmu_buf_impl_t *db)
375 if (db->db_level > 0) {
378 boolean_t is_metadata;
381 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type);
384 return (is_metadata);
389 * This function *must* return indices evenly distributed between all
390 * sublists of the multilist. This is needed due to how the dbuf eviction
391 * code is laid out; dbuf_evict_thread() assumes dbufs are evenly
392 * distributed between all sublists and uses this assumption when
393 * deciding which sublist to evict from and how much to evict from it.
396 dbuf_cache_multilist_index_func(multilist_t *ml, void *obj)
398 dmu_buf_impl_t *db = obj;
401 * The assumption here, is the hash value for a given
402 * dmu_buf_impl_t will remain constant throughout it's lifetime
403 * (i.e. it's objset, object, level and blkid fields don't change).
404 * Thus, we don't need to store the dbuf's sublist index
405 * on insertion, as this index can be recalculated on removal.
407 * Also, the low order bits of the hash value are thought to be
408 * distributed evenly. Otherwise, in the case that the multilist
409 * has a power of two number of sublists, each sublists' usage
410 * would not be evenly distributed.
412 return (dbuf_hash(db->db_objset, db->db.db_object,
413 db->db_level, db->db_blkid) %
414 multilist_get_num_sublists(ml));
417 static inline boolean_t
418 dbuf_cache_above_hiwater(void)
420 uint64_t dbuf_cache_hiwater_bytes =
421 (dbuf_cache_max_bytes * dbuf_cache_hiwater_pct) / 100;
423 return (refcount_count(&dbuf_cache_size) >
424 dbuf_cache_max_bytes + dbuf_cache_hiwater_bytes);
427 static inline boolean_t
428 dbuf_cache_above_lowater(void)
430 uint64_t dbuf_cache_lowater_bytes =
431 (dbuf_cache_max_bytes * dbuf_cache_lowater_pct) / 100;
433 return (refcount_count(&dbuf_cache_size) >
434 dbuf_cache_max_bytes - dbuf_cache_lowater_bytes);
438 * Evict the oldest eligible dbuf from the dbuf cache.
443 int idx = multilist_get_random_index(&dbuf_cache);
444 multilist_sublist_t *mls = multilist_sublist_lock(&dbuf_cache, idx);
446 ASSERT(!MUTEX_HELD(&dbuf_evict_lock));
449 * Set the thread's tsd to indicate that it's processing evictions.
450 * Once a thread stops evicting from the dbuf cache it will
451 * reset its tsd to NULL.
453 ASSERT3P(tsd_get(zfs_dbuf_evict_key), ==, NULL);
454 (void) tsd_set(zfs_dbuf_evict_key, (void *)B_TRUE);
456 dmu_buf_impl_t *db = multilist_sublist_tail(mls);
457 while (db != NULL && mutex_tryenter(&db->db_mtx) == 0) {
458 db = multilist_sublist_prev(mls, db);
461 DTRACE_PROBE2(dbuf__evict__one, dmu_buf_impl_t *, db,
462 multilist_sublist_t *, mls);
465 multilist_sublist_remove(mls, db);
466 multilist_sublist_unlock(mls);
467 (void) refcount_remove_many(&dbuf_cache_size,
471 multilist_sublist_unlock(mls);
473 (void) tsd_set(zfs_dbuf_evict_key, NULL);
477 * The dbuf evict thread is responsible for aging out dbufs from the
478 * cache. Once the cache has reached it's maximum size, dbufs are removed
479 * and destroyed. The eviction thread will continue running until the size
480 * of the dbuf cache is at or below the maximum size. Once the dbuf is aged
481 * out of the cache it is destroyed and becomes eligible for arc eviction.
484 dbuf_evict_thread(void *dummy __unused)
488 CALLB_CPR_INIT(&cpr, &dbuf_evict_lock, callb_generic_cpr, FTAG);
490 mutex_enter(&dbuf_evict_lock);
491 while (!dbuf_evict_thread_exit) {
492 while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
493 CALLB_CPR_SAFE_BEGIN(&cpr);
494 (void) cv_timedwait_hires(&dbuf_evict_cv,
495 &dbuf_evict_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
496 CALLB_CPR_SAFE_END(&cpr, &dbuf_evict_lock);
498 mutex_exit(&dbuf_evict_lock);
501 * Keep evicting as long as we're above the low water mark
502 * for the cache. We do this without holding the locks to
503 * minimize lock contention.
505 while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
509 mutex_enter(&dbuf_evict_lock);
512 dbuf_evict_thread_exit = B_FALSE;
513 cv_broadcast(&dbuf_evict_cv);
514 CALLB_CPR_EXIT(&cpr); /* drops dbuf_evict_lock */
519 * Wake up the dbuf eviction thread if the dbuf cache is at its max size.
520 * If the dbuf cache is at its high water mark, then evict a dbuf from the
521 * dbuf cache using the callers context.
524 dbuf_evict_notify(void)
528 * We use thread specific data to track when a thread has
529 * started processing evictions. This allows us to avoid deeply
530 * nested stacks that would have a call flow similar to this:
532 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify()
535 * +-----dbuf_destroy()<--dbuf_evict_one()<--------+
537 * The dbuf_eviction_thread will always have its tsd set until
538 * that thread exits. All other threads will only set their tsd
539 * if they are participating in the eviction process. This only
540 * happens if the eviction thread is unable to process evictions
541 * fast enough. To keep the dbuf cache size in check, other threads
542 * can evict from the dbuf cache directly. Those threads will set
543 * their tsd values so that we ensure that they only evict one dbuf
544 * from the dbuf cache.
546 if (tsd_get(zfs_dbuf_evict_key) != NULL)
549 if (refcount_count(&dbuf_cache_size) > dbuf_cache_max_bytes) {
550 boolean_t evict_now = B_FALSE;
552 mutex_enter(&dbuf_evict_lock);
553 if (refcount_count(&dbuf_cache_size) > dbuf_cache_max_bytes) {
554 evict_now = dbuf_cache_above_hiwater();
555 cv_signal(&dbuf_evict_cv);
557 mutex_exit(&dbuf_evict_lock);
568 uint64_t hsize = 1ULL << 16;
569 dbuf_hash_table_t *h = &dbuf_hash_table;
573 * The hash table is big enough to fill all of physical memory
574 * with an average 4K block size. The table will take up
575 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
577 while (hsize * 4096 < (uint64_t)physmem * PAGESIZE)
581 h->hash_table_mask = hsize - 1;
582 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
583 if (h->hash_table == NULL) {
584 /* XXX - we should really return an error instead of assert */
585 ASSERT(hsize > (1ULL << 10));
590 dbuf_kmem_cache = kmem_cache_create("dmu_buf_impl_t",
591 sizeof (dmu_buf_impl_t),
592 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
594 for (i = 0; i < DBUF_MUTEXES; i++)
595 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
598 * Setup the parameters for the dbuf cache. We cap the size of the
599 * dbuf cache to 1/32nd (default) of the size of the ARC.
601 dbuf_cache_max_bytes = MIN(dbuf_cache_max_bytes,
602 arc_max_bytes() >> dbuf_cache_max_shift);
605 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
606 * configuration is not required.
608 dbu_evict_taskq = taskq_create("dbu_evict", 1, minclsyspri, 0, 0, 0);
610 multilist_create(&dbuf_cache, sizeof (dmu_buf_impl_t),
611 offsetof(dmu_buf_impl_t, db_cache_link),
612 zfs_arc_num_sublists_per_state,
613 dbuf_cache_multilist_index_func);
614 refcount_create(&dbuf_cache_size);
616 tsd_create(&zfs_dbuf_evict_key, NULL);
617 dbuf_evict_thread_exit = B_FALSE;
618 mutex_init(&dbuf_evict_lock, NULL, MUTEX_DEFAULT, NULL);
619 cv_init(&dbuf_evict_cv, NULL, CV_DEFAULT, NULL);
620 dbuf_cache_evict_thread = thread_create(NULL, 0, dbuf_evict_thread,
621 NULL, 0, &p0, TS_RUN, minclsyspri);
627 dbuf_hash_table_t *h = &dbuf_hash_table;
630 for (i = 0; i < DBUF_MUTEXES; i++)
631 mutex_destroy(&h->hash_mutexes[i]);
632 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
633 kmem_cache_destroy(dbuf_kmem_cache);
634 taskq_destroy(dbu_evict_taskq);
636 mutex_enter(&dbuf_evict_lock);
637 dbuf_evict_thread_exit = B_TRUE;
638 while (dbuf_evict_thread_exit) {
639 cv_signal(&dbuf_evict_cv);
640 cv_wait(&dbuf_evict_cv, &dbuf_evict_lock);
642 mutex_exit(&dbuf_evict_lock);
643 tsd_destroy(&zfs_dbuf_evict_key);
645 mutex_destroy(&dbuf_evict_lock);
646 cv_destroy(&dbuf_evict_cv);
648 refcount_destroy(&dbuf_cache_size);
649 multilist_destroy(&dbuf_cache);
658 dbuf_verify(dmu_buf_impl_t *db)
661 dbuf_dirty_record_t *dr;
663 ASSERT(MUTEX_HELD(&db->db_mtx));
665 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
668 ASSERT(db->db_objset != NULL);
672 ASSERT(db->db_parent == NULL);
673 ASSERT(db->db_blkptr == NULL);
675 ASSERT3U(db->db.db_object, ==, dn->dn_object);
676 ASSERT3P(db->db_objset, ==, dn->dn_objset);
677 ASSERT3U(db->db_level, <, dn->dn_nlevels);
678 ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
679 db->db_blkid == DMU_SPILL_BLKID ||
680 !avl_is_empty(&dn->dn_dbufs));
682 if (db->db_blkid == DMU_BONUS_BLKID) {
684 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
685 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
686 } else if (db->db_blkid == DMU_SPILL_BLKID) {
688 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
689 ASSERT0(db->db.db_offset);
691 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
694 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
695 ASSERT(dr->dr_dbuf == db);
697 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
698 ASSERT(dr->dr_dbuf == db);
701 * We can't assert that db_size matches dn_datablksz because it
702 * can be momentarily different when another thread is doing
705 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
706 dr = db->db_data_pending;
708 * It should only be modified in syncing context, so
709 * make sure we only have one copy of the data.
711 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
714 /* verify db->db_blkptr */
716 if (db->db_parent == dn->dn_dbuf) {
717 /* db is pointed to by the dnode */
718 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
719 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
720 ASSERT(db->db_parent == NULL);
722 ASSERT(db->db_parent != NULL);
723 if (db->db_blkid != DMU_SPILL_BLKID)
724 ASSERT3P(db->db_blkptr, ==,
725 &dn->dn_phys->dn_blkptr[db->db_blkid]);
727 /* db is pointed to by an indirect block */
728 int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
729 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
730 ASSERT3U(db->db_parent->db.db_object, ==,
733 * dnode_grow_indblksz() can make this fail if we don't
734 * have the struct_rwlock. XXX indblksz no longer
735 * grows. safe to do this now?
737 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
738 ASSERT3P(db->db_blkptr, ==,
739 ((blkptr_t *)db->db_parent->db.db_data +
740 db->db_blkid % epb));
744 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
745 (db->db_buf == NULL || db->db_buf->b_data) &&
746 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
747 db->db_state != DB_FILL && !dn->dn_free_txg) {
749 * If the blkptr isn't set but they have nonzero data,
750 * it had better be dirty, otherwise we'll lose that
751 * data when we evict this buffer.
753 * There is an exception to this rule for indirect blocks; in
754 * this case, if the indirect block is a hole, we fill in a few
755 * fields on each of the child blocks (importantly, birth time)
756 * to prevent hole birth times from being lost when you
757 * partially fill in a hole.
759 if (db->db_dirtycnt == 0) {
760 if (db->db_level == 0) {
761 uint64_t *buf = db->db.db_data;
764 for (i = 0; i < db->db.db_size >> 3; i++) {
768 blkptr_t *bps = db->db.db_data;
769 ASSERT3U(1 << DB_DNODE(db)->dn_indblkshift, ==,
772 * We want to verify that all the blkptrs in the
773 * indirect block are holes, but we may have
774 * automatically set up a few fields for them.
775 * We iterate through each blkptr and verify
776 * they only have those fields set.
779 i < db->db.db_size / sizeof (blkptr_t);
781 blkptr_t *bp = &bps[i];
782 ASSERT(ZIO_CHECKSUM_IS_ZERO(
785 DVA_IS_EMPTY(&bp->blk_dva[0]) &&
786 DVA_IS_EMPTY(&bp->blk_dva[1]) &&
787 DVA_IS_EMPTY(&bp->blk_dva[2]));
788 ASSERT0(bp->blk_fill);
789 ASSERT0(bp->blk_pad[0]);
790 ASSERT0(bp->blk_pad[1]);
791 ASSERT(!BP_IS_EMBEDDED(bp));
792 ASSERT(BP_IS_HOLE(bp));
793 ASSERT0(bp->blk_phys_birth);
803 dbuf_clear_data(dmu_buf_impl_t *db)
805 ASSERT(MUTEX_HELD(&db->db_mtx));
807 ASSERT3P(db->db_buf, ==, NULL);
808 db->db.db_data = NULL;
809 if (db->db_state != DB_NOFILL)
810 db->db_state = DB_UNCACHED;
814 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
816 ASSERT(MUTEX_HELD(&db->db_mtx));
820 ASSERT(buf->b_data != NULL);
821 db->db.db_data = buf->b_data;
825 * Loan out an arc_buf for read. Return the loaned arc_buf.
828 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
832 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
833 mutex_enter(&db->db_mtx);
834 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
835 int blksz = db->db.db_size;
836 spa_t *spa = db->db_objset->os_spa;
838 mutex_exit(&db->db_mtx);
839 abuf = arc_loan_buf(spa, blksz);
840 bcopy(db->db.db_data, abuf->b_data, blksz);
843 arc_loan_inuse_buf(abuf, db);
846 mutex_exit(&db->db_mtx);
852 * Calculate which level n block references the data at the level 0 offset
856 dbuf_whichblock(dnode_t *dn, int64_t level, uint64_t offset)
858 if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) {
860 * The level n blkid is equal to the level 0 blkid divided by
861 * the number of level 0s in a level n block.
863 * The level 0 blkid is offset >> datablkshift =
864 * offset / 2^datablkshift.
866 * The number of level 0s in a level n is the number of block
867 * pointers in an indirect block, raised to the power of level.
868 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
869 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
871 * Thus, the level n blkid is: offset /
872 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
873 * = offset / 2^(datablkshift + level *
874 * (indblkshift - SPA_BLKPTRSHIFT))
875 * = offset >> (datablkshift + level *
876 * (indblkshift - SPA_BLKPTRSHIFT))
878 return (offset >> (dn->dn_datablkshift + level *
879 (dn->dn_indblkshift - SPA_BLKPTRSHIFT)));
881 ASSERT3U(offset, <, dn->dn_datablksz);
887 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
889 dmu_buf_impl_t *db = vdb;
891 mutex_enter(&db->db_mtx);
892 ASSERT3U(db->db_state, ==, DB_READ);
894 * All reads are synchronous, so we must have a hold on the dbuf
896 ASSERT(refcount_count(&db->db_holds) > 0);
897 ASSERT(db->db_buf == NULL);
898 ASSERT(db->db.db_data == NULL);
899 if (db->db_level == 0 && db->db_freed_in_flight) {
900 /* we were freed in flight; disregard any error */
901 arc_release(buf, db);
902 bzero(buf->b_data, db->db.db_size);
904 db->db_freed_in_flight = FALSE;
905 dbuf_set_data(db, buf);
906 db->db_state = DB_CACHED;
907 } else if (zio == NULL || zio->io_error == 0) {
908 dbuf_set_data(db, buf);
909 db->db_state = DB_CACHED;
911 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
912 ASSERT3P(db->db_buf, ==, NULL);
913 arc_buf_destroy(buf, db);
914 db->db_state = DB_UNCACHED;
916 cv_broadcast(&db->db_changed);
917 dbuf_rele_and_unlock(db, NULL);
921 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
925 arc_flags_t aflags = ARC_FLAG_NOWAIT;
929 ASSERT(!refcount_is_zero(&db->db_holds));
930 /* We need the struct_rwlock to prevent db_blkptr from changing. */
931 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
932 ASSERT(MUTEX_HELD(&db->db_mtx));
933 ASSERT(db->db_state == DB_UNCACHED);
934 ASSERT(db->db_buf == NULL);
936 if (db->db_blkid == DMU_BONUS_BLKID) {
937 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
939 ASSERT3U(bonuslen, <=, db->db.db_size);
940 db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN);
941 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
942 if (bonuslen < DN_MAX_BONUSLEN)
943 bzero(db->db.db_data, DN_MAX_BONUSLEN);
945 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
947 db->db_state = DB_CACHED;
948 mutex_exit(&db->db_mtx);
953 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
954 * processes the delete record and clears the bp while we are waiting
955 * for the dn_mtx (resulting in a "no" from block_freed).
957 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
958 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
959 BP_IS_HOLE(db->db_blkptr)))) {
960 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
962 dbuf_set_data(db, arc_alloc_buf(db->db_objset->os_spa,
963 db->db.db_size, db, type));
964 bzero(db->db.db_data, db->db.db_size);
966 if (db->db_blkptr != NULL && db->db_level > 0 &&
967 BP_IS_HOLE(db->db_blkptr) &&
968 db->db_blkptr->blk_birth != 0) {
969 blkptr_t *bps = db->db.db_data;
970 for (int i = 0; i < ((1 <<
971 DB_DNODE(db)->dn_indblkshift) / sizeof (blkptr_t));
973 blkptr_t *bp = &bps[i];
974 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
975 1 << dn->dn_indblkshift);
977 BP_GET_LEVEL(db->db_blkptr) == 1 ?
979 BP_GET_LSIZE(db->db_blkptr));
980 BP_SET_TYPE(bp, BP_GET_TYPE(db->db_blkptr));
982 BP_GET_LEVEL(db->db_blkptr) - 1);
983 BP_SET_BIRTH(bp, db->db_blkptr->blk_birth, 0);
987 db->db_state = DB_CACHED;
988 mutex_exit(&db->db_mtx);
994 db->db_state = DB_READ;
995 mutex_exit(&db->db_mtx);
997 if (DBUF_IS_L2CACHEABLE(db))
998 aflags |= ARC_FLAG_L2CACHE;
1000 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
1001 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
1002 db->db.db_object, db->db_level, db->db_blkid);
1004 dbuf_add_ref(db, NULL);
1006 (void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr,
1007 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
1008 (flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
1013 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
1016 boolean_t havepzio = (zio != NULL);
1021 * We don't have to hold the mutex to check db_state because it
1022 * can't be freed while we have a hold on the buffer.
1024 ASSERT(!refcount_is_zero(&db->db_holds));
1026 if (db->db_state == DB_NOFILL)
1027 return (SET_ERROR(EIO));
1031 if ((flags & DB_RF_HAVESTRUCT) == 0)
1032 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1034 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1035 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
1036 DBUF_IS_CACHEABLE(db);
1038 mutex_enter(&db->db_mtx);
1039 if (db->db_state == DB_CACHED) {
1040 mutex_exit(&db->db_mtx);
1042 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1043 if ((flags & DB_RF_HAVESTRUCT) == 0)
1044 rw_exit(&dn->dn_struct_rwlock);
1046 } else if (db->db_state == DB_UNCACHED) {
1047 spa_t *spa = dn->dn_objset->os_spa;
1050 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
1051 dbuf_read_impl(db, zio, flags);
1053 /* dbuf_read_impl has dropped db_mtx for us */
1056 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1058 if ((flags & DB_RF_HAVESTRUCT) == 0)
1059 rw_exit(&dn->dn_struct_rwlock);
1063 err = zio_wait(zio);
1066 * Another reader came in while the dbuf was in flight
1067 * between UNCACHED and CACHED. Either a writer will finish
1068 * writing the buffer (sending the dbuf to CACHED) or the
1069 * first reader's request will reach the read_done callback
1070 * and send the dbuf to CACHED. Otherwise, a failure
1071 * occurred and the dbuf went to UNCACHED.
1073 mutex_exit(&db->db_mtx);
1075 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1076 if ((flags & DB_RF_HAVESTRUCT) == 0)
1077 rw_exit(&dn->dn_struct_rwlock);
1080 /* Skip the wait per the caller's request. */
1081 mutex_enter(&db->db_mtx);
1082 if ((flags & DB_RF_NEVERWAIT) == 0) {
1083 while (db->db_state == DB_READ ||
1084 db->db_state == DB_FILL) {
1085 ASSERT(db->db_state == DB_READ ||
1086 (flags & DB_RF_HAVESTRUCT) == 0);
1087 DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *,
1089 cv_wait(&db->db_changed, &db->db_mtx);
1091 if (db->db_state == DB_UNCACHED)
1092 err = SET_ERROR(EIO);
1094 mutex_exit(&db->db_mtx);
1097 ASSERT(err || havepzio || db->db_state == DB_CACHED);
1102 dbuf_noread(dmu_buf_impl_t *db)
1104 ASSERT(!refcount_is_zero(&db->db_holds));
1105 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1106 mutex_enter(&db->db_mtx);
1107 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1108 cv_wait(&db->db_changed, &db->db_mtx);
1109 if (db->db_state == DB_UNCACHED) {
1110 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1111 spa_t *spa = db->db_objset->os_spa;
1113 ASSERT(db->db_buf == NULL);
1114 ASSERT(db->db.db_data == NULL);
1115 dbuf_set_data(db, arc_alloc_buf(spa, db->db.db_size, db, type));
1116 db->db_state = DB_FILL;
1117 } else if (db->db_state == DB_NOFILL) {
1118 dbuf_clear_data(db);
1120 ASSERT3U(db->db_state, ==, DB_CACHED);
1122 mutex_exit(&db->db_mtx);
1126 * This is our just-in-time copy function. It makes a copy of
1127 * buffers, that have been modified in a previous transaction
1128 * group, before we modify them in the current active group.
1130 * This function is used in two places: when we are dirtying a
1131 * buffer for the first time in a txg, and when we are freeing
1132 * a range in a dnode that includes this buffer.
1134 * Note that when we are called from dbuf_free_range() we do
1135 * not put a hold on the buffer, we just traverse the active
1136 * dbuf list for the dnode.
1139 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
1141 dbuf_dirty_record_t *dr = db->db_last_dirty;
1143 ASSERT(MUTEX_HELD(&db->db_mtx));
1144 ASSERT(db->db.db_data != NULL);
1145 ASSERT(db->db_level == 0);
1146 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
1149 (dr->dt.dl.dr_data !=
1150 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
1154 * If the last dirty record for this dbuf has not yet synced
1155 * and its referencing the dbuf data, either:
1156 * reset the reference to point to a new copy,
1157 * or (if there a no active holders)
1158 * just null out the current db_data pointer.
1160 ASSERT(dr->dr_txg >= txg - 2);
1161 if (db->db_blkid == DMU_BONUS_BLKID) {
1162 /* Note that the data bufs here are zio_bufs */
1163 dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
1164 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
1165 bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
1166 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
1167 int size = db->db.db_size;
1168 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1169 spa_t *spa = db->db_objset->os_spa;
1171 dr->dt.dl.dr_data = arc_alloc_buf(spa, size, db, type);
1172 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
1175 dbuf_clear_data(db);
1180 dbuf_unoverride(dbuf_dirty_record_t *dr)
1182 dmu_buf_impl_t *db = dr->dr_dbuf;
1183 blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
1184 uint64_t txg = dr->dr_txg;
1186 ASSERT(MUTEX_HELD(&db->db_mtx));
1187 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
1188 ASSERT(db->db_level == 0);
1190 if (db->db_blkid == DMU_BONUS_BLKID ||
1191 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
1194 ASSERT(db->db_data_pending != dr);
1196 /* free this block */
1197 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite)
1198 zio_free(db->db_objset->os_spa, txg, bp);
1200 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1201 dr->dt.dl.dr_nopwrite = B_FALSE;
1204 * Release the already-written buffer, so we leave it in
1205 * a consistent dirty state. Note that all callers are
1206 * modifying the buffer, so they will immediately do
1207 * another (redundant) arc_release(). Therefore, leave
1208 * the buf thawed to save the effort of freezing &
1209 * immediately re-thawing it.
1211 arc_release(dr->dt.dl.dr_data, db);
1215 * Evict (if its unreferenced) or clear (if its referenced) any level-0
1216 * data blocks in the free range, so that any future readers will find
1219 * This is a no-op if the dataset is in the middle of an incremental
1220 * receive; see comment below for details.
1223 dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
1226 dmu_buf_impl_t db_search;
1227 dmu_buf_impl_t *db, *db_next;
1228 uint64_t txg = tx->tx_txg;
1231 if (end_blkid > dn->dn_maxblkid && (end_blkid != DMU_SPILL_BLKID))
1232 end_blkid = dn->dn_maxblkid;
1233 dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid);
1235 db_search.db_level = 0;
1236 db_search.db_blkid = start_blkid;
1237 db_search.db_state = DB_SEARCH;
1239 mutex_enter(&dn->dn_dbufs_mtx);
1240 if (start_blkid >= dn->dn_unlisted_l0_blkid) {
1241 /* There can't be any dbufs in this range; no need to search. */
1243 db = avl_find(&dn->dn_dbufs, &db_search, &where);
1244 ASSERT3P(db, ==, NULL);
1245 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1246 ASSERT(db == NULL || db->db_level > 0);
1248 mutex_exit(&dn->dn_dbufs_mtx);
1250 } else if (dmu_objset_is_receiving(dn->dn_objset)) {
1252 * If we are receiving, we expect there to be no dbufs in
1253 * the range to be freed, because receive modifies each
1254 * block at most once, and in offset order. If this is
1255 * not the case, it can lead to performance problems,
1256 * so note that we unexpectedly took the slow path.
1258 atomic_inc_64(&zfs_free_range_recv_miss);
1261 db = avl_find(&dn->dn_dbufs, &db_search, &where);
1262 ASSERT3P(db, ==, NULL);
1263 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1265 for (; db != NULL; db = db_next) {
1266 db_next = AVL_NEXT(&dn->dn_dbufs, db);
1267 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1269 if (db->db_level != 0 || db->db_blkid > end_blkid) {
1272 ASSERT3U(db->db_blkid, >=, start_blkid);
1274 /* found a level 0 buffer in the range */
1275 mutex_enter(&db->db_mtx);
1276 if (dbuf_undirty(db, tx)) {
1277 /* mutex has been dropped and dbuf destroyed */
1281 if (db->db_state == DB_UNCACHED ||
1282 db->db_state == DB_NOFILL ||
1283 db->db_state == DB_EVICTING) {
1284 ASSERT(db->db.db_data == NULL);
1285 mutex_exit(&db->db_mtx);
1288 if (db->db_state == DB_READ || db->db_state == DB_FILL) {
1289 /* will be handled in dbuf_read_done or dbuf_rele */
1290 db->db_freed_in_flight = TRUE;
1291 mutex_exit(&db->db_mtx);
1294 if (refcount_count(&db->db_holds) == 0) {
1299 /* The dbuf is referenced */
1301 if (db->db_last_dirty != NULL) {
1302 dbuf_dirty_record_t *dr = db->db_last_dirty;
1304 if (dr->dr_txg == txg) {
1306 * This buffer is "in-use", re-adjust the file
1307 * size to reflect that this buffer may
1308 * contain new data when we sync.
1310 if (db->db_blkid != DMU_SPILL_BLKID &&
1311 db->db_blkid > dn->dn_maxblkid)
1312 dn->dn_maxblkid = db->db_blkid;
1313 dbuf_unoverride(dr);
1316 * This dbuf is not dirty in the open context.
1317 * Either uncache it (if its not referenced in
1318 * the open context) or reset its contents to
1321 dbuf_fix_old_data(db, txg);
1324 /* clear the contents if its cached */
1325 if (db->db_state == DB_CACHED) {
1326 ASSERT(db->db.db_data != NULL);
1327 arc_release(db->db_buf, db);
1328 bzero(db->db.db_data, db->db.db_size);
1329 arc_buf_freeze(db->db_buf);
1332 mutex_exit(&db->db_mtx);
1334 mutex_exit(&dn->dn_dbufs_mtx);
1338 dbuf_block_freeable(dmu_buf_impl_t *db)
1340 dsl_dataset_t *ds = db->db_objset->os_dsl_dataset;
1341 uint64_t birth_txg = 0;
1344 * We don't need any locking to protect db_blkptr:
1345 * If it's syncing, then db_last_dirty will be set
1346 * so we'll ignore db_blkptr.
1348 * This logic ensures that only block births for
1349 * filled blocks are considered.
1351 ASSERT(MUTEX_HELD(&db->db_mtx));
1352 if (db->db_last_dirty && (db->db_blkptr == NULL ||
1353 !BP_IS_HOLE(db->db_blkptr))) {
1354 birth_txg = db->db_last_dirty->dr_txg;
1355 } else if (db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) {
1356 birth_txg = db->db_blkptr->blk_birth;
1360 * If this block don't exist or is in a snapshot, it can't be freed.
1361 * Don't pass the bp to dsl_dataset_block_freeable() since we
1362 * are holding the db_mtx lock and might deadlock if we are
1363 * prefetching a dedup-ed block.
1366 return (ds == NULL ||
1367 dsl_dataset_block_freeable(ds, NULL, birth_txg));
1373 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
1375 arc_buf_t *buf, *obuf;
1376 int osize = db->db.db_size;
1377 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1380 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1385 /* XXX does *this* func really need the lock? */
1386 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1389 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1390 * is OK, because there can be no other references to the db
1391 * when we are changing its size, so no concurrent DB_FILL can
1395 * XXX we should be doing a dbuf_read, checking the return
1396 * value and returning that up to our callers
1398 dmu_buf_will_dirty(&db->db, tx);
1400 /* create the data buffer for the new block */
1401 buf = arc_alloc_buf(dn->dn_objset->os_spa, size, db, type);
1403 /* copy old block data to the new block */
1405 bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
1406 /* zero the remainder */
1408 bzero((uint8_t *)buf->b_data + osize, size - osize);
1410 mutex_enter(&db->db_mtx);
1411 dbuf_set_data(db, buf);
1412 arc_buf_destroy(obuf, db);
1413 db->db.db_size = size;
1415 if (db->db_level == 0) {
1416 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1417 db->db_last_dirty->dt.dl.dr_data = buf;
1419 mutex_exit(&db->db_mtx);
1421 dnode_willuse_space(dn, size-osize, tx);
1426 dbuf_release_bp(dmu_buf_impl_t *db)
1428 objset_t *os = db->db_objset;
1430 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1431 ASSERT(arc_released(os->os_phys_buf) ||
1432 list_link_active(&os->os_dsl_dataset->ds_synced_link));
1433 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1435 (void) arc_release(db->db_buf, db);
1439 * We already have a dirty record for this TXG, and we are being
1443 dbuf_redirty(dbuf_dirty_record_t *dr)
1445 dmu_buf_impl_t *db = dr->dr_dbuf;
1447 ASSERT(MUTEX_HELD(&db->db_mtx));
1449 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1451 * If this buffer has already been written out,
1452 * we now need to reset its state.
1454 dbuf_unoverride(dr);
1455 if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1456 db->db_state != DB_NOFILL) {
1457 /* Already released on initial dirty, so just thaw. */
1458 ASSERT(arc_released(db->db_buf));
1459 arc_buf_thaw(db->db_buf);
1464 dbuf_dirty_record_t *
1465 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1469 dbuf_dirty_record_t **drp, *dr;
1470 int drop_struct_lock = FALSE;
1471 boolean_t do_free_accounting = B_FALSE;
1472 int txgoff = tx->tx_txg & TXG_MASK;
1474 ASSERT(tx->tx_txg != 0);
1475 ASSERT(!refcount_is_zero(&db->db_holds));
1476 DMU_TX_DIRTY_BUF(tx, db);
1481 * Shouldn't dirty a regular buffer in syncing context. Private
1482 * objects may be dirtied in syncing context, but only if they
1483 * were already pre-dirtied in open context.
1485 ASSERT(!dmu_tx_is_syncing(tx) ||
1486 BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1487 DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1488 dn->dn_objset->os_dsl_dataset == NULL);
1490 * We make this assert for private objects as well, but after we
1491 * check if we're already dirty. They are allowed to re-dirty
1492 * in syncing context.
1494 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1495 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1496 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1498 mutex_enter(&db->db_mtx);
1500 * XXX make this true for indirects too? The problem is that
1501 * transactions created with dmu_tx_create_assigned() from
1502 * syncing context don't bother holding ahead.
1504 ASSERT(db->db_level != 0 ||
1505 db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1506 db->db_state == DB_NOFILL);
1508 mutex_enter(&dn->dn_mtx);
1510 * Don't set dirtyctx to SYNC if we're just modifying this as we
1511 * initialize the objset.
1513 if (dn->dn_dirtyctx == DN_UNDIRTIED &&
1514 !BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1516 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1517 ASSERT(dn->dn_dirtyctx_firstset == NULL);
1518 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1520 mutex_exit(&dn->dn_mtx);
1522 if (db->db_blkid == DMU_SPILL_BLKID)
1523 dn->dn_have_spill = B_TRUE;
1526 * If this buffer is already dirty, we're done.
1528 drp = &db->db_last_dirty;
1529 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1530 db->db.db_object == DMU_META_DNODE_OBJECT);
1531 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1533 if (dr && dr->dr_txg == tx->tx_txg) {
1537 mutex_exit(&db->db_mtx);
1542 * Only valid if not already dirty.
1544 ASSERT(dn->dn_object == 0 ||
1545 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1546 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1548 ASSERT3U(dn->dn_nlevels, >, db->db_level);
1549 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1550 dn->dn_phys->dn_nlevels > db->db_level ||
1551 dn->dn_next_nlevels[txgoff] > db->db_level ||
1552 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1553 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1556 * We should only be dirtying in syncing context if it's the
1557 * mos or we're initializing the os or it's a special object.
1558 * However, we are allowed to dirty in syncing context provided
1559 * we already dirtied it in open context. Hence we must make
1560 * this assertion only if we're not already dirty.
1563 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1564 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1565 ASSERT(db->db.db_size != 0);
1567 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1569 if (db->db_blkid != DMU_BONUS_BLKID) {
1571 * Update the accounting.
1572 * Note: we delay "free accounting" until after we drop
1573 * the db_mtx. This keeps us from grabbing other locks
1574 * (and possibly deadlocking) in bp_get_dsize() while
1575 * also holding the db_mtx.
1577 dnode_willuse_space(dn, db->db.db_size, tx);
1578 do_free_accounting = dbuf_block_freeable(db);
1582 * If this buffer is dirty in an old transaction group we need
1583 * to make a copy of it so that the changes we make in this
1584 * transaction group won't leak out when we sync the older txg.
1586 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1587 if (db->db_level == 0) {
1588 void *data_old = db->db_buf;
1590 if (db->db_state != DB_NOFILL) {
1591 if (db->db_blkid == DMU_BONUS_BLKID) {
1592 dbuf_fix_old_data(db, tx->tx_txg);
1593 data_old = db->db.db_data;
1594 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1596 * Release the data buffer from the cache so
1597 * that we can modify it without impacting
1598 * possible other users of this cached data
1599 * block. Note that indirect blocks and
1600 * private objects are not released until the
1601 * syncing state (since they are only modified
1604 arc_release(db->db_buf, db);
1605 dbuf_fix_old_data(db, tx->tx_txg);
1606 data_old = db->db_buf;
1608 ASSERT(data_old != NULL);
1610 dr->dt.dl.dr_data = data_old;
1612 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1613 list_create(&dr->dt.di.dr_children,
1614 sizeof (dbuf_dirty_record_t),
1615 offsetof(dbuf_dirty_record_t, dr_dirty_node));
1617 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL)
1618 dr->dr_accounted = db->db.db_size;
1620 dr->dr_txg = tx->tx_txg;
1625 * We could have been freed_in_flight between the dbuf_noread
1626 * and dbuf_dirty. We win, as though the dbuf_noread() had
1627 * happened after the free.
1629 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1630 db->db_blkid != DMU_SPILL_BLKID) {
1631 mutex_enter(&dn->dn_mtx);
1632 if (dn->dn_free_ranges[txgoff] != NULL) {
1633 range_tree_clear(dn->dn_free_ranges[txgoff],
1636 mutex_exit(&dn->dn_mtx);
1637 db->db_freed_in_flight = FALSE;
1641 * This buffer is now part of this txg
1643 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1644 db->db_dirtycnt += 1;
1645 ASSERT3U(db->db_dirtycnt, <=, 3);
1647 mutex_exit(&db->db_mtx);
1649 if (db->db_blkid == DMU_BONUS_BLKID ||
1650 db->db_blkid == DMU_SPILL_BLKID) {
1651 mutex_enter(&dn->dn_mtx);
1652 ASSERT(!list_link_active(&dr->dr_dirty_node));
1653 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1654 mutex_exit(&dn->dn_mtx);
1655 dnode_setdirty(dn, tx);
1658 } else if (do_free_accounting) {
1659 blkptr_t *bp = db->db_blkptr;
1660 int64_t willfree = (bp && !BP_IS_HOLE(bp)) ?
1661 bp_get_dsize(os->os_spa, bp) : db->db.db_size;
1663 * This is only a guess -- if the dbuf is dirty
1664 * in a previous txg, we don't know how much
1665 * space it will use on disk yet. We should
1666 * really have the struct_rwlock to access
1667 * db_blkptr, but since this is just a guess,
1668 * it's OK if we get an odd answer.
1670 ddt_prefetch(os->os_spa, bp);
1671 dnode_willuse_space(dn, -willfree, tx);
1674 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1675 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1676 drop_struct_lock = TRUE;
1679 if (db->db_level == 0) {
1680 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1681 ASSERT(dn->dn_maxblkid >= db->db_blkid);
1684 if (db->db_level+1 < dn->dn_nlevels) {
1685 dmu_buf_impl_t *parent = db->db_parent;
1686 dbuf_dirty_record_t *di;
1687 int parent_held = FALSE;
1689 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1690 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1692 parent = dbuf_hold_level(dn, db->db_level+1,
1693 db->db_blkid >> epbs, FTAG);
1694 ASSERT(parent != NULL);
1697 if (drop_struct_lock)
1698 rw_exit(&dn->dn_struct_rwlock);
1699 ASSERT3U(db->db_level+1, ==, parent->db_level);
1700 di = dbuf_dirty(parent, tx);
1702 dbuf_rele(parent, FTAG);
1704 mutex_enter(&db->db_mtx);
1706 * Since we've dropped the mutex, it's possible that
1707 * dbuf_undirty() might have changed this out from under us.
1709 if (db->db_last_dirty == dr ||
1710 dn->dn_object == DMU_META_DNODE_OBJECT) {
1711 mutex_enter(&di->dt.di.dr_mtx);
1712 ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1713 ASSERT(!list_link_active(&dr->dr_dirty_node));
1714 list_insert_tail(&di->dt.di.dr_children, dr);
1715 mutex_exit(&di->dt.di.dr_mtx);
1718 mutex_exit(&db->db_mtx);
1720 ASSERT(db->db_level+1 == dn->dn_nlevels);
1721 ASSERT(db->db_blkid < dn->dn_nblkptr);
1722 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1723 mutex_enter(&dn->dn_mtx);
1724 ASSERT(!list_link_active(&dr->dr_dirty_node));
1725 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1726 mutex_exit(&dn->dn_mtx);
1727 if (drop_struct_lock)
1728 rw_exit(&dn->dn_struct_rwlock);
1731 dnode_setdirty(dn, tx);
1737 * Undirty a buffer in the transaction group referenced by the given
1738 * transaction. Return whether this evicted the dbuf.
1741 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1744 uint64_t txg = tx->tx_txg;
1745 dbuf_dirty_record_t *dr, **drp;
1750 * Due to our use of dn_nlevels below, this can only be called
1751 * in open context, unless we are operating on the MOS.
1752 * From syncing context, dn_nlevels may be different from the
1753 * dn_nlevels used when dbuf was dirtied.
1755 ASSERT(db->db_objset ==
1756 dmu_objset_pool(db->db_objset)->dp_meta_objset ||
1757 txg != spa_syncing_txg(dmu_objset_spa(db->db_objset)));
1758 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1759 ASSERT0(db->db_level);
1760 ASSERT(MUTEX_HELD(&db->db_mtx));
1763 * If this buffer is not dirty, we're done.
1765 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1766 if (dr->dr_txg <= txg)
1768 if (dr == NULL || dr->dr_txg < txg)
1770 ASSERT(dr->dr_txg == txg);
1771 ASSERT(dr->dr_dbuf == db);
1776 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1778 ASSERT(db->db.db_size != 0);
1780 dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset),
1781 dr->dr_accounted, txg);
1786 * Note that there are three places in dbuf_dirty()
1787 * where this dirty record may be put on a list.
1788 * Make sure to do a list_remove corresponding to
1789 * every one of those list_insert calls.
1791 if (dr->dr_parent) {
1792 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1793 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1794 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1795 } else if (db->db_blkid == DMU_SPILL_BLKID ||
1796 db->db_level + 1 == dn->dn_nlevels) {
1797 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1798 mutex_enter(&dn->dn_mtx);
1799 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1800 mutex_exit(&dn->dn_mtx);
1804 if (db->db_state != DB_NOFILL) {
1805 dbuf_unoverride(dr);
1807 ASSERT(db->db_buf != NULL);
1808 ASSERT(dr->dt.dl.dr_data != NULL);
1809 if (dr->dt.dl.dr_data != db->db_buf)
1810 arc_buf_destroy(dr->dt.dl.dr_data, db);
1813 kmem_free(dr, sizeof (dbuf_dirty_record_t));
1815 ASSERT(db->db_dirtycnt > 0);
1816 db->db_dirtycnt -= 1;
1818 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1819 ASSERT(db->db_state == DB_NOFILL || arc_released(db->db_buf));
1828 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
1830 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1831 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1833 ASSERT(tx->tx_txg != 0);
1834 ASSERT(!refcount_is_zero(&db->db_holds));
1837 * Quick check for dirtyness. For already dirty blocks, this
1838 * reduces runtime of this function by >90%, and overall performance
1839 * by 50% for some workloads (e.g. file deletion with indirect blocks
1842 mutex_enter(&db->db_mtx);
1843 dbuf_dirty_record_t *dr;
1844 for (dr = db->db_last_dirty;
1845 dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) {
1847 * It's possible that it is already dirty but not cached,
1848 * because there are some calls to dbuf_dirty() that don't
1849 * go through dmu_buf_will_dirty().
1851 if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) {
1852 /* This dbuf is already dirty and cached. */
1854 mutex_exit(&db->db_mtx);
1858 mutex_exit(&db->db_mtx);
1861 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
1862 rf |= DB_RF_HAVESTRUCT;
1864 (void) dbuf_read(db, NULL, rf);
1865 (void) dbuf_dirty(db, tx);
1869 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1871 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1873 db->db_state = DB_NOFILL;
1875 dmu_buf_will_fill(db_fake, tx);
1879 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1881 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1883 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1884 ASSERT(tx->tx_txg != 0);
1885 ASSERT(db->db_level == 0);
1886 ASSERT(!refcount_is_zero(&db->db_holds));
1888 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1889 dmu_tx_private_ok(tx));
1892 (void) dbuf_dirty(db, tx);
1895 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1898 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1900 mutex_enter(&db->db_mtx);
1903 if (db->db_state == DB_FILL) {
1904 if (db->db_level == 0 && db->db_freed_in_flight) {
1905 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1906 /* we were freed while filling */
1907 /* XXX dbuf_undirty? */
1908 bzero(db->db.db_data, db->db.db_size);
1909 db->db_freed_in_flight = FALSE;
1911 db->db_state = DB_CACHED;
1912 cv_broadcast(&db->db_changed);
1914 mutex_exit(&db->db_mtx);
1918 dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data,
1919 bp_embedded_type_t etype, enum zio_compress comp,
1920 int uncompressed_size, int compressed_size, int byteorder,
1923 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
1924 struct dirty_leaf *dl;
1925 dmu_object_type_t type;
1927 if (etype == BP_EMBEDDED_TYPE_DATA) {
1928 ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset),
1929 SPA_FEATURE_EMBEDDED_DATA));
1933 type = DB_DNODE(db)->dn_type;
1936 ASSERT0(db->db_level);
1937 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1939 dmu_buf_will_not_fill(dbuf, tx);
1941 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1942 dl = &db->db_last_dirty->dt.dl;
1943 encode_embedded_bp_compressed(&dl->dr_overridden_by,
1944 data, comp, uncompressed_size, compressed_size);
1945 BPE_SET_ETYPE(&dl->dr_overridden_by, etype);
1946 BP_SET_TYPE(&dl->dr_overridden_by, type);
1947 BP_SET_LEVEL(&dl->dr_overridden_by, 0);
1948 BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder);
1950 dl->dr_override_state = DR_OVERRIDDEN;
1951 dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg;
1955 * Directly assign a provided arc buf to a given dbuf if it's not referenced
1956 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1959 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
1961 ASSERT(!refcount_is_zero(&db->db_holds));
1962 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1963 ASSERT(db->db_level == 0);
1964 ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA);
1965 ASSERT(buf != NULL);
1966 ASSERT(arc_buf_size(buf) == db->db.db_size);
1967 ASSERT(tx->tx_txg != 0);
1969 arc_return_buf(buf, db);
1970 ASSERT(arc_released(buf));
1972 mutex_enter(&db->db_mtx);
1974 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1975 cv_wait(&db->db_changed, &db->db_mtx);
1977 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
1979 if (db->db_state == DB_CACHED &&
1980 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
1981 mutex_exit(&db->db_mtx);
1982 (void) dbuf_dirty(db, tx);
1983 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
1984 arc_buf_destroy(buf, db);
1985 xuio_stat_wbuf_copied();
1989 xuio_stat_wbuf_nocopy();
1990 if (db->db_state == DB_CACHED) {
1991 dbuf_dirty_record_t *dr = db->db_last_dirty;
1993 ASSERT(db->db_buf != NULL);
1994 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
1995 ASSERT(dr->dt.dl.dr_data == db->db_buf);
1996 if (!arc_released(db->db_buf)) {
1997 ASSERT(dr->dt.dl.dr_override_state ==
1999 arc_release(db->db_buf, db);
2001 dr->dt.dl.dr_data = buf;
2002 arc_buf_destroy(db->db_buf, db);
2003 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
2004 arc_release(db->db_buf, db);
2005 arc_buf_destroy(db->db_buf, db);
2009 ASSERT(db->db_buf == NULL);
2010 dbuf_set_data(db, buf);
2011 db->db_state = DB_FILL;
2012 mutex_exit(&db->db_mtx);
2013 (void) dbuf_dirty(db, tx);
2014 dmu_buf_fill_done(&db->db, tx);
2018 dbuf_destroy(dmu_buf_impl_t *db)
2021 dmu_buf_impl_t *parent = db->db_parent;
2022 dmu_buf_impl_t *dndb;
2024 ASSERT(MUTEX_HELD(&db->db_mtx));
2025 ASSERT(refcount_is_zero(&db->db_holds));
2027 if (db->db_buf != NULL) {
2028 arc_buf_destroy(db->db_buf, db);
2032 if (db->db_blkid == DMU_BONUS_BLKID) {
2033 ASSERT(db->db.db_data != NULL);
2034 zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
2035 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
2036 db->db_state = DB_UNCACHED;
2039 dbuf_clear_data(db);
2041 if (multilist_link_active(&db->db_cache_link)) {
2042 multilist_remove(&dbuf_cache, db);
2043 (void) refcount_remove_many(&dbuf_cache_size,
2044 db->db.db_size, db);
2047 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
2048 ASSERT(db->db_data_pending == NULL);
2050 db->db_state = DB_EVICTING;
2051 db->db_blkptr = NULL;
2054 * Now that db_state is DB_EVICTING, nobody else can find this via
2055 * the hash table. We can now drop db_mtx, which allows us to
2056 * acquire the dn_dbufs_mtx.
2058 mutex_exit(&db->db_mtx);
2063 if (db->db_blkid != DMU_BONUS_BLKID) {
2064 boolean_t needlock = !MUTEX_HELD(&dn->dn_dbufs_mtx);
2066 mutex_enter(&dn->dn_dbufs_mtx);
2067 avl_remove(&dn->dn_dbufs, db);
2068 atomic_dec_32(&dn->dn_dbufs_count);
2072 mutex_exit(&dn->dn_dbufs_mtx);
2074 * Decrementing the dbuf count means that the hold corresponding
2075 * to the removed dbuf is no longer discounted in dnode_move(),
2076 * so the dnode cannot be moved until after we release the hold.
2077 * The membar_producer() ensures visibility of the decremented
2078 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
2082 db->db_dnode_handle = NULL;
2084 dbuf_hash_remove(db);
2089 ASSERT(refcount_is_zero(&db->db_holds));
2091 db->db_parent = NULL;
2093 ASSERT(db->db_buf == NULL);
2094 ASSERT(db->db.db_data == NULL);
2095 ASSERT(db->db_hash_next == NULL);
2096 ASSERT(db->db_blkptr == NULL);
2097 ASSERT(db->db_data_pending == NULL);
2098 ASSERT(!multilist_link_active(&db->db_cache_link));
2100 kmem_cache_free(dbuf_kmem_cache, db);
2101 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2104 * If this dbuf is referenced from an indirect dbuf,
2105 * decrement the ref count on the indirect dbuf.
2107 if (parent && parent != dndb)
2108 dbuf_rele(parent, db);
2112 * Note: While bpp will always be updated if the function returns success,
2113 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
2114 * this happens when the dnode is the meta-dnode, or a userused or groupused
2118 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
2119 dmu_buf_impl_t **parentp, blkptr_t **bpp)
2126 ASSERT(blkid != DMU_BONUS_BLKID);
2128 if (blkid == DMU_SPILL_BLKID) {
2129 mutex_enter(&dn->dn_mtx);
2130 if (dn->dn_have_spill &&
2131 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
2132 *bpp = &dn->dn_phys->dn_spill;
2135 dbuf_add_ref(dn->dn_dbuf, NULL);
2136 *parentp = dn->dn_dbuf;
2137 mutex_exit(&dn->dn_mtx);
2141 if (dn->dn_phys->dn_nlevels == 0)
2144 nlevels = dn->dn_phys->dn_nlevels;
2146 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2148 ASSERT3U(level * epbs, <, 64);
2149 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2150 if (level >= nlevels ||
2151 (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
2152 /* the buffer has no parent yet */
2153 return (SET_ERROR(ENOENT));
2154 } else if (level < nlevels-1) {
2155 /* this block is referenced from an indirect block */
2156 int err = dbuf_hold_impl(dn, level+1,
2157 blkid >> epbs, fail_sparse, FALSE, NULL, parentp);
2160 err = dbuf_read(*parentp, NULL,
2161 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
2163 dbuf_rele(*parentp, NULL);
2167 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
2168 (blkid & ((1ULL << epbs) - 1));
2171 /* the block is referenced from the dnode */
2172 ASSERT3U(level, ==, nlevels-1);
2173 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
2174 blkid < dn->dn_phys->dn_nblkptr);
2176 dbuf_add_ref(dn->dn_dbuf, NULL);
2177 *parentp = dn->dn_dbuf;
2179 *bpp = &dn->dn_phys->dn_blkptr[blkid];
2184 static dmu_buf_impl_t *
2185 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
2186 dmu_buf_impl_t *parent, blkptr_t *blkptr)
2188 objset_t *os = dn->dn_objset;
2189 dmu_buf_impl_t *db, *odb;
2191 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2192 ASSERT(dn->dn_type != DMU_OT_NONE);
2194 db = kmem_cache_alloc(dbuf_kmem_cache, KM_SLEEP);
2197 db->db.db_object = dn->dn_object;
2198 db->db_level = level;
2199 db->db_blkid = blkid;
2200 db->db_last_dirty = NULL;
2201 db->db_dirtycnt = 0;
2202 db->db_dnode_handle = dn->dn_handle;
2203 db->db_parent = parent;
2204 db->db_blkptr = blkptr;
2207 db->db_user_immediate_evict = FALSE;
2208 db->db_freed_in_flight = FALSE;
2209 db->db_pending_evict = FALSE;
2211 if (blkid == DMU_BONUS_BLKID) {
2212 ASSERT3P(parent, ==, dn->dn_dbuf);
2213 db->db.db_size = DN_MAX_BONUSLEN -
2214 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
2215 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
2216 db->db.db_offset = DMU_BONUS_BLKID;
2217 db->db_state = DB_UNCACHED;
2218 /* the bonus dbuf is not placed in the hash table */
2219 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2221 } else if (blkid == DMU_SPILL_BLKID) {
2222 db->db.db_size = (blkptr != NULL) ?
2223 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
2224 db->db.db_offset = 0;
2227 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
2228 db->db.db_size = blocksize;
2229 db->db.db_offset = db->db_blkid * blocksize;
2233 * Hold the dn_dbufs_mtx while we get the new dbuf
2234 * in the hash table *and* added to the dbufs list.
2235 * This prevents a possible deadlock with someone
2236 * trying to look up this dbuf before its added to the
2239 mutex_enter(&dn->dn_dbufs_mtx);
2240 db->db_state = DB_EVICTING;
2241 if ((odb = dbuf_hash_insert(db)) != NULL) {
2242 /* someone else inserted it first */
2243 kmem_cache_free(dbuf_kmem_cache, db);
2244 mutex_exit(&dn->dn_dbufs_mtx);
2247 avl_add(&dn->dn_dbufs, db);
2248 if (db->db_level == 0 && db->db_blkid >=
2249 dn->dn_unlisted_l0_blkid)
2250 dn->dn_unlisted_l0_blkid = db->db_blkid + 1;
2251 db->db_state = DB_UNCACHED;
2252 mutex_exit(&dn->dn_dbufs_mtx);
2253 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2255 if (parent && parent != dn->dn_dbuf)
2256 dbuf_add_ref(parent, db);
2258 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
2259 refcount_count(&dn->dn_holds) > 0);
2260 (void) refcount_add(&dn->dn_holds, db);
2261 atomic_inc_32(&dn->dn_dbufs_count);
2263 dprintf_dbuf(db, "db=%p\n", db);
2268 typedef struct dbuf_prefetch_arg {
2269 spa_t *dpa_spa; /* The spa to issue the prefetch in. */
2270 zbookmark_phys_t dpa_zb; /* The target block to prefetch. */
2271 int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */
2272 int dpa_curlevel; /* The current level that we're reading */
2273 dnode_t *dpa_dnode; /* The dnode associated with the prefetch */
2274 zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */
2275 zio_t *dpa_zio; /* The parent zio_t for all prefetches. */
2276 arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */
2277 } dbuf_prefetch_arg_t;
2280 * Actually issue the prefetch read for the block given.
2283 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp)
2285 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
2288 arc_flags_t aflags =
2289 dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
2291 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2292 ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level);
2293 ASSERT(dpa->dpa_zio != NULL);
2294 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL,
2295 dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2296 &aflags, &dpa->dpa_zb);
2300 * Called when an indirect block above our prefetch target is read in. This
2301 * will either read in the next indirect block down the tree or issue the actual
2302 * prefetch if the next block down is our target.
2305 dbuf_prefetch_indirect_done(zio_t *zio, arc_buf_t *abuf, void *private)
2307 dbuf_prefetch_arg_t *dpa = private;
2309 ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel);
2310 ASSERT3S(dpa->dpa_curlevel, >, 0);
2313 * The dpa_dnode is only valid if we are called with a NULL
2314 * zio. This indicates that the arc_read() returned without
2315 * first calling zio_read() to issue a physical read. Once
2316 * a physical read is made the dpa_dnode must be invalidated
2317 * as the locks guarding it may have been dropped. If the
2318 * dpa_dnode is still valid, then we want to add it to the dbuf
2319 * cache. To do so, we must hold the dbuf associated with the block
2320 * we just prefetched, read its contents so that we associate it
2321 * with an arc_buf_t, and then release it.
2324 ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel);
2325 if (zio->io_flags & ZIO_FLAG_RAW) {
2326 ASSERT3U(BP_GET_PSIZE(zio->io_bp), ==, zio->io_size);
2328 ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size);
2330 ASSERT3P(zio->io_spa, ==, dpa->dpa_spa);
2332 dpa->dpa_dnode = NULL;
2333 } else if (dpa->dpa_dnode != NULL) {
2334 uint64_t curblkid = dpa->dpa_zb.zb_blkid >>
2335 (dpa->dpa_epbs * (dpa->dpa_curlevel -
2336 dpa->dpa_zb.zb_level));
2337 dmu_buf_impl_t *db = dbuf_hold_level(dpa->dpa_dnode,
2338 dpa->dpa_curlevel, curblkid, FTAG);
2339 (void) dbuf_read(db, NULL,
2340 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_HAVESTRUCT);
2341 dbuf_rele(db, FTAG);
2344 dpa->dpa_curlevel--;
2346 uint64_t nextblkid = dpa->dpa_zb.zb_blkid >>
2347 (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level));
2348 blkptr_t *bp = ((blkptr_t *)abuf->b_data) +
2349 P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs);
2350 if (BP_IS_HOLE(bp) || (zio != NULL && zio->io_error != 0)) {
2351 kmem_free(dpa, sizeof (*dpa));
2352 } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) {
2353 ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid);
2354 dbuf_issue_final_prefetch(dpa, bp);
2355 kmem_free(dpa, sizeof (*dpa));
2357 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2358 zbookmark_phys_t zb;
2360 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2362 SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset,
2363 dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid);
2365 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2366 bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio,
2367 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2371 arc_buf_destroy(abuf, private);
2375 * Issue prefetch reads for the given block on the given level. If the indirect
2376 * blocks above that block are not in memory, we will read them in
2377 * asynchronously. As a result, this call never blocks waiting for a read to
2381 dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio,
2385 int epbs, nlevels, curlevel;
2388 ASSERT(blkid != DMU_BONUS_BLKID);
2389 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2391 if (blkid > dn->dn_maxblkid)
2394 if (dnode_block_freed(dn, blkid))
2398 * This dnode hasn't been written to disk yet, so there's nothing to
2401 nlevels = dn->dn_phys->dn_nlevels;
2402 if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0)
2405 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2406 if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level))
2409 dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object,
2412 mutex_exit(&db->db_mtx);
2414 * This dbuf already exists. It is either CACHED, or
2415 * (we assume) about to be read or filled.
2421 * Find the closest ancestor (indirect block) of the target block
2422 * that is present in the cache. In this indirect block, we will
2423 * find the bp that is at curlevel, curblkid.
2427 while (curlevel < nlevels - 1) {
2428 int parent_level = curlevel + 1;
2429 uint64_t parent_blkid = curblkid >> epbs;
2432 if (dbuf_hold_impl(dn, parent_level, parent_blkid,
2433 FALSE, TRUE, FTAG, &db) == 0) {
2434 blkptr_t *bpp = db->db_buf->b_data;
2435 bp = bpp[P2PHASE(curblkid, 1 << epbs)];
2436 dbuf_rele(db, FTAG);
2440 curlevel = parent_level;
2441 curblkid = parent_blkid;
2444 if (curlevel == nlevels - 1) {
2445 /* No cached indirect blocks found. */
2446 ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr);
2447 bp = dn->dn_phys->dn_blkptr[curblkid];
2449 if (BP_IS_HOLE(&bp))
2452 ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp));
2454 zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL,
2457 dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP);
2458 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
2459 SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2460 dn->dn_object, level, blkid);
2461 dpa->dpa_curlevel = curlevel;
2462 dpa->dpa_prio = prio;
2463 dpa->dpa_aflags = aflags;
2464 dpa->dpa_spa = dn->dn_objset->os_spa;
2465 dpa->dpa_dnode = dn;
2466 dpa->dpa_epbs = epbs;
2470 * If we have the indirect just above us, no need to do the asynchronous
2471 * prefetch chain; we'll just run the last step ourselves. If we're at
2472 * a higher level, though, we want to issue the prefetches for all the
2473 * indirect blocks asynchronously, so we can go on with whatever we were
2476 if (curlevel == level) {
2477 ASSERT3U(curblkid, ==, blkid);
2478 dbuf_issue_final_prefetch(dpa, &bp);
2479 kmem_free(dpa, sizeof (*dpa));
2481 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2482 zbookmark_phys_t zb;
2484 SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2485 dn->dn_object, curlevel, curblkid);
2486 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2487 &bp, dbuf_prefetch_indirect_done, dpa, prio,
2488 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2492 * We use pio here instead of dpa_zio since it's possible that
2493 * dpa may have already been freed.
2499 * Returns with db_holds incremented, and db_mtx not held.
2500 * Note: dn_struct_rwlock must be held.
2503 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid,
2504 boolean_t fail_sparse, boolean_t fail_uncached,
2505 void *tag, dmu_buf_impl_t **dbp)
2507 dmu_buf_impl_t *db, *parent = NULL;
2509 ASSERT(blkid != DMU_BONUS_BLKID);
2510 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2511 ASSERT3U(dn->dn_nlevels, >, level);
2515 /* dbuf_find() returns with db_mtx held */
2516 db = dbuf_find(dn->dn_objset, dn->dn_object, level, blkid);
2519 blkptr_t *bp = NULL;
2523 return (SET_ERROR(ENOENT));
2525 ASSERT3P(parent, ==, NULL);
2526 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
2528 if (err == 0 && bp && BP_IS_HOLE(bp))
2529 err = SET_ERROR(ENOENT);
2532 dbuf_rele(parent, NULL);
2536 if (err && err != ENOENT)
2538 db = dbuf_create(dn, level, blkid, parent, bp);
2541 if (fail_uncached && db->db_state != DB_CACHED) {
2542 mutex_exit(&db->db_mtx);
2543 return (SET_ERROR(ENOENT));
2546 if (db->db_buf != NULL)
2547 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
2549 ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
2552 * If this buffer is currently syncing out, and we are are
2553 * still referencing it from db_data, we need to make a copy
2554 * of it in case we decide we want to dirty it again in this txg.
2556 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
2557 dn->dn_object != DMU_META_DNODE_OBJECT &&
2558 db->db_state == DB_CACHED && db->db_data_pending) {
2559 dbuf_dirty_record_t *dr = db->db_data_pending;
2561 if (dr->dt.dl.dr_data == db->db_buf) {
2562 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2565 arc_alloc_buf(dn->dn_objset->os_spa,
2566 db->db.db_size, db, type));
2567 bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
2572 if (multilist_link_active(&db->db_cache_link)) {
2573 ASSERT(refcount_is_zero(&db->db_holds));
2574 multilist_remove(&dbuf_cache, db);
2575 (void) refcount_remove_many(&dbuf_cache_size,
2576 db->db.db_size, db);
2578 (void) refcount_add(&db->db_holds, tag);
2580 mutex_exit(&db->db_mtx);
2582 /* NOTE: we can't rele the parent until after we drop the db_mtx */
2584 dbuf_rele(parent, NULL);
2586 ASSERT3P(DB_DNODE(db), ==, dn);
2587 ASSERT3U(db->db_blkid, ==, blkid);
2588 ASSERT3U(db->db_level, ==, level);
2595 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
2597 return (dbuf_hold_level(dn, 0, blkid, tag));
2601 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
2604 int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db);
2605 return (err ? NULL : db);
2609 dbuf_create_bonus(dnode_t *dn)
2611 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
2613 ASSERT(dn->dn_bonus == NULL);
2614 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
2618 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
2620 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2623 if (db->db_blkid != DMU_SPILL_BLKID)
2624 return (SET_ERROR(ENOTSUP));
2626 blksz = SPA_MINBLOCKSIZE;
2627 ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset)));
2628 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
2632 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2633 dbuf_new_size(db, blksz, tx);
2634 rw_exit(&dn->dn_struct_rwlock);
2641 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
2643 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
2646 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2648 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
2650 int64_t holds = refcount_add(&db->db_holds, tag);
2651 ASSERT3S(holds, >, 1);
2654 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
2656 dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid,
2659 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2660 dmu_buf_impl_t *found_db;
2661 boolean_t result = B_FALSE;
2663 if (db->db_blkid == DMU_BONUS_BLKID)
2664 found_db = dbuf_find_bonus(os, obj);
2666 found_db = dbuf_find(os, obj, 0, blkid);
2668 if (found_db != NULL) {
2669 if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) {
2670 (void) refcount_add(&db->db_holds, tag);
2673 mutex_exit(&db->db_mtx);
2679 * If you call dbuf_rele() you had better not be referencing the dnode handle
2680 * unless you have some other direct or indirect hold on the dnode. (An indirect
2681 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2682 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2683 * dnode's parent dbuf evicting its dnode handles.
2686 dbuf_rele(dmu_buf_impl_t *db, void *tag)
2688 mutex_enter(&db->db_mtx);
2689 dbuf_rele_and_unlock(db, tag);
2693 dmu_buf_rele(dmu_buf_t *db, void *tag)
2695 dbuf_rele((dmu_buf_impl_t *)db, tag);
2699 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2700 * db_dirtycnt and db_holds to be updated atomically.
2703 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2707 ASSERT(MUTEX_HELD(&db->db_mtx));
2711 * Remove the reference to the dbuf before removing its hold on the
2712 * dnode so we can guarantee in dnode_move() that a referenced bonus
2713 * buffer has a corresponding dnode hold.
2715 holds = refcount_remove(&db->db_holds, tag);
2719 * We can't freeze indirects if there is a possibility that they
2720 * may be modified in the current syncing context.
2722 if (db->db_buf != NULL &&
2723 holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) {
2724 arc_buf_freeze(db->db_buf);
2727 if (holds == db->db_dirtycnt &&
2728 db->db_level == 0 && db->db_user_immediate_evict)
2729 dbuf_evict_user(db);
2732 if (db->db_blkid == DMU_BONUS_BLKID) {
2734 boolean_t evict_dbuf = db->db_pending_evict;
2737 * If the dnode moves here, we cannot cross this
2738 * barrier until the move completes.
2743 atomic_dec_32(&dn->dn_dbufs_count);
2746 * Decrementing the dbuf count means that the bonus
2747 * buffer's dnode hold is no longer discounted in
2748 * dnode_move(). The dnode cannot move until after
2749 * the dnode_rele() below.
2754 * Do not reference db after its lock is dropped.
2755 * Another thread may evict it.
2757 mutex_exit(&db->db_mtx);
2760 dnode_evict_bonus(dn);
2763 } else if (db->db_buf == NULL) {
2765 * This is a special case: we never associated this
2766 * dbuf with any data allocated from the ARC.
2768 ASSERT(db->db_state == DB_UNCACHED ||
2769 db->db_state == DB_NOFILL);
2771 } else if (arc_released(db->db_buf)) {
2773 * This dbuf has anonymous data associated with it.
2777 boolean_t do_arc_evict = B_FALSE;
2779 spa_t *spa = dmu_objset_spa(db->db_objset);
2781 if (!DBUF_IS_CACHEABLE(db) &&
2782 db->db_blkptr != NULL &&
2783 !BP_IS_HOLE(db->db_blkptr) &&
2784 !BP_IS_EMBEDDED(db->db_blkptr)) {
2785 do_arc_evict = B_TRUE;
2786 bp = *db->db_blkptr;
2789 if (!DBUF_IS_CACHEABLE(db) ||
2790 db->db_pending_evict) {
2792 } else if (!multilist_link_active(&db->db_cache_link)) {
2793 multilist_insert(&dbuf_cache, db);
2794 (void) refcount_add_many(&dbuf_cache_size,
2795 db->db.db_size, db);
2796 mutex_exit(&db->db_mtx);
2798 dbuf_evict_notify();
2802 arc_freed(spa, &bp);
2805 mutex_exit(&db->db_mtx);
2810 #pragma weak dmu_buf_refcount = dbuf_refcount
2812 dbuf_refcount(dmu_buf_impl_t *db)
2814 return (refcount_count(&db->db_holds));
2818 dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user,
2819 dmu_buf_user_t *new_user)
2821 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2823 mutex_enter(&db->db_mtx);
2824 dbuf_verify_user(db, DBVU_NOT_EVICTING);
2825 if (db->db_user == old_user)
2826 db->db_user = new_user;
2828 old_user = db->db_user;
2829 dbuf_verify_user(db, DBVU_NOT_EVICTING);
2830 mutex_exit(&db->db_mtx);
2836 dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2838 return (dmu_buf_replace_user(db_fake, NULL, user));
2842 dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2844 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2846 db->db_user_immediate_evict = TRUE;
2847 return (dmu_buf_set_user(db_fake, user));
2851 dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2853 return (dmu_buf_replace_user(db_fake, user, NULL));
2857 dmu_buf_get_user(dmu_buf_t *db_fake)
2859 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2861 dbuf_verify_user(db, DBVU_NOT_EVICTING);
2862 return (db->db_user);
2866 dmu_buf_user_evict_wait()
2868 taskq_wait(dbu_evict_taskq);
2872 dmu_buf_freeable(dmu_buf_t *dbuf)
2874 boolean_t res = B_FALSE;
2875 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2878 res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset,
2879 db->db_blkptr, db->db_blkptr->blk_birth);
2885 dmu_buf_get_blkptr(dmu_buf_t *db)
2887 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2888 return (dbi->db_blkptr);
2892 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
2894 /* ASSERT(dmu_tx_is_syncing(tx) */
2895 ASSERT(MUTEX_HELD(&db->db_mtx));
2897 if (db->db_blkptr != NULL)
2900 if (db->db_blkid == DMU_SPILL_BLKID) {
2901 db->db_blkptr = &dn->dn_phys->dn_spill;
2902 BP_ZERO(db->db_blkptr);
2905 if (db->db_level == dn->dn_phys->dn_nlevels-1) {
2907 * This buffer was allocated at a time when there was
2908 * no available blkptrs from the dnode, or it was
2909 * inappropriate to hook it in (i.e., nlevels mis-match).
2911 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
2912 ASSERT(db->db_parent == NULL);
2913 db->db_parent = dn->dn_dbuf;
2914 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
2917 dmu_buf_impl_t *parent = db->db_parent;
2918 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2920 ASSERT(dn->dn_phys->dn_nlevels > 1);
2921 if (parent == NULL) {
2922 mutex_exit(&db->db_mtx);
2923 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2924 parent = dbuf_hold_level(dn, db->db_level + 1,
2925 db->db_blkid >> epbs, db);
2926 rw_exit(&dn->dn_struct_rwlock);
2927 mutex_enter(&db->db_mtx);
2928 db->db_parent = parent;
2930 db->db_blkptr = (blkptr_t *)parent->db.db_data +
2931 (db->db_blkid & ((1ULL << epbs) - 1));
2937 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2939 dmu_buf_impl_t *db = dr->dr_dbuf;
2943 ASSERT(dmu_tx_is_syncing(tx));
2945 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2947 mutex_enter(&db->db_mtx);
2949 ASSERT(db->db_level > 0);
2952 /* Read the block if it hasn't been read yet. */
2953 if (db->db_buf == NULL) {
2954 mutex_exit(&db->db_mtx);
2955 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
2956 mutex_enter(&db->db_mtx);
2958 ASSERT3U(db->db_state, ==, DB_CACHED);
2959 ASSERT(db->db_buf != NULL);
2963 /* Indirect block size must match what the dnode thinks it is. */
2964 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2965 dbuf_check_blkptr(dn, db);
2968 /* Provide the pending dirty record to child dbufs */
2969 db->db_data_pending = dr;
2971 mutex_exit(&db->db_mtx);
2972 dbuf_write(dr, db->db_buf, tx);
2975 mutex_enter(&dr->dt.di.dr_mtx);
2976 dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx);
2977 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2978 mutex_exit(&dr->dt.di.dr_mtx);
2983 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2985 arc_buf_t **datap = &dr->dt.dl.dr_data;
2986 dmu_buf_impl_t *db = dr->dr_dbuf;
2989 uint64_t txg = tx->tx_txg;
2991 ASSERT(dmu_tx_is_syncing(tx));
2993 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2995 mutex_enter(&db->db_mtx);
2997 * To be synced, we must be dirtied. But we
2998 * might have been freed after the dirty.
3000 if (db->db_state == DB_UNCACHED) {
3001 /* This buffer has been freed since it was dirtied */
3002 ASSERT(db->db.db_data == NULL);
3003 } else if (db->db_state == DB_FILL) {
3004 /* This buffer was freed and is now being re-filled */
3005 ASSERT(db->db.db_data != dr->dt.dl.dr_data);
3007 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
3014 if (db->db_blkid == DMU_SPILL_BLKID) {
3015 mutex_enter(&dn->dn_mtx);
3016 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
3017 mutex_exit(&dn->dn_mtx);
3021 * If this is a bonus buffer, simply copy the bonus data into the
3022 * dnode. It will be written out when the dnode is synced (and it
3023 * will be synced, since it must have been dirty for dbuf_sync to
3026 if (db->db_blkid == DMU_BONUS_BLKID) {
3027 dbuf_dirty_record_t **drp;
3029 ASSERT(*datap != NULL);
3030 ASSERT0(db->db_level);
3031 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN);
3032 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen);
3035 if (*datap != db->db.db_data) {
3036 zio_buf_free(*datap, DN_MAX_BONUSLEN);
3037 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
3039 db->db_data_pending = NULL;
3040 drp = &db->db_last_dirty;
3042 drp = &(*drp)->dr_next;
3043 ASSERT(dr->dr_next == NULL);
3044 ASSERT(dr->dr_dbuf == db);
3046 if (dr->dr_dbuf->db_level != 0) {
3047 list_destroy(&dr->dt.di.dr_children);
3048 mutex_destroy(&dr->dt.di.dr_mtx);
3050 kmem_free(dr, sizeof (dbuf_dirty_record_t));
3051 ASSERT(db->db_dirtycnt > 0);
3052 db->db_dirtycnt -= 1;
3053 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
3060 * This function may have dropped the db_mtx lock allowing a dmu_sync
3061 * operation to sneak in. As a result, we need to ensure that we
3062 * don't check the dr_override_state until we have returned from
3063 * dbuf_check_blkptr.
3065 dbuf_check_blkptr(dn, db);
3068 * If this buffer is in the middle of an immediate write,
3069 * wait for the synchronous IO to complete.
3071 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
3072 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
3073 cv_wait(&db->db_changed, &db->db_mtx);
3074 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
3077 if (db->db_state != DB_NOFILL &&
3078 dn->dn_object != DMU_META_DNODE_OBJECT &&
3079 refcount_count(&db->db_holds) > 1 &&
3080 dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
3081 *datap == db->db_buf) {
3083 * If this buffer is currently "in use" (i.e., there
3084 * are active holds and db_data still references it),
3085 * then make a copy before we start the write so that
3086 * any modifications from the open txg will not leak
3089 * NOTE: this copy does not need to be made for
3090 * objects only modified in the syncing context (e.g.
3091 * DNONE_DNODE blocks).
3093 int blksz = arc_buf_size(*datap);
3094 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
3095 *datap = arc_alloc_buf(os->os_spa, blksz, db, type);
3096 bcopy(db->db.db_data, (*datap)->b_data, blksz);
3098 db->db_data_pending = dr;
3100 mutex_exit(&db->db_mtx);
3102 dbuf_write(dr, *datap, tx);
3104 ASSERT(!list_link_active(&dr->dr_dirty_node));
3105 if (dn->dn_object == DMU_META_DNODE_OBJECT) {
3106 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
3110 * Although zio_nowait() does not "wait for an IO", it does
3111 * initiate the IO. If this is an empty write it seems plausible
3112 * that the IO could actually be completed before the nowait
3113 * returns. We need to DB_DNODE_EXIT() first in case
3114 * zio_nowait() invalidates the dbuf.
3117 zio_nowait(dr->dr_zio);
3122 dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx)
3124 dbuf_dirty_record_t *dr;
3126 while (dr = list_head(list)) {
3127 if (dr->dr_zio != NULL) {
3129 * If we find an already initialized zio then we
3130 * are processing the meta-dnode, and we have finished.
3131 * The dbufs for all dnodes are put back on the list
3132 * during processing, so that we can zio_wait()
3133 * these IOs after initiating all child IOs.
3135 ASSERT3U(dr->dr_dbuf->db.db_object, ==,
3136 DMU_META_DNODE_OBJECT);
3139 if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
3140 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
3141 VERIFY3U(dr->dr_dbuf->db_level, ==, level);
3143 list_remove(list, dr);
3144 if (dr->dr_dbuf->db_level > 0)
3145 dbuf_sync_indirect(dr, tx);
3147 dbuf_sync_leaf(dr, tx);
3153 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3155 dmu_buf_impl_t *db = vdb;
3157 blkptr_t *bp = zio->io_bp;
3158 blkptr_t *bp_orig = &zio->io_bp_orig;
3159 spa_t *spa = zio->io_spa;
3164 ASSERT3P(db->db_blkptr, !=, NULL);
3165 ASSERT3P(&db->db_data_pending->dr_bp_copy, ==, bp);
3169 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
3170 dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
3171 zio->io_prev_space_delta = delta;
3173 if (bp->blk_birth != 0) {
3174 ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
3175 BP_GET_TYPE(bp) == dn->dn_type) ||
3176 (db->db_blkid == DMU_SPILL_BLKID &&
3177 BP_GET_TYPE(bp) == dn->dn_bonustype) ||
3178 BP_IS_EMBEDDED(bp));
3179 ASSERT(BP_GET_LEVEL(bp) == db->db_level);
3182 mutex_enter(&db->db_mtx);
3185 if (db->db_blkid == DMU_SPILL_BLKID) {
3186 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3187 ASSERT(!(BP_IS_HOLE(bp)) &&
3188 db->db_blkptr == &dn->dn_phys->dn_spill);
3192 if (db->db_level == 0) {
3193 mutex_enter(&dn->dn_mtx);
3194 if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
3195 db->db_blkid != DMU_SPILL_BLKID)
3196 dn->dn_phys->dn_maxblkid = db->db_blkid;
3197 mutex_exit(&dn->dn_mtx);
3199 if (dn->dn_type == DMU_OT_DNODE) {
3200 dnode_phys_t *dnp = db->db.db_data;
3201 for (i = db->db.db_size >> DNODE_SHIFT; i > 0;
3203 if (dnp->dn_type != DMU_OT_NONE)
3207 if (BP_IS_HOLE(bp)) {
3214 blkptr_t *ibp = db->db.db_data;
3215 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3216 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
3217 if (BP_IS_HOLE(ibp))
3219 fill += BP_GET_FILL(ibp);
3224 if (!BP_IS_EMBEDDED(bp))
3225 bp->blk_fill = fill;
3227 mutex_exit(&db->db_mtx);
3229 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3230 *db->db_blkptr = *bp;
3231 rw_exit(&dn->dn_struct_rwlock);
3236 * This function gets called just prior to running through the compression
3237 * stage of the zio pipeline. If we're an indirect block comprised of only
3238 * holes, then we want this indirect to be compressed away to a hole. In
3239 * order to do that we must zero out any information about the holes that
3240 * this indirect points to prior to before we try to compress it.
3243 dbuf_write_children_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3245 dmu_buf_impl_t *db = vdb;
3251 ASSERT3U(db->db_level, >, 0);
3254 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3256 /* Determine if all our children are holes */
3257 for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++) {
3258 if (!BP_IS_HOLE(bp))
3263 * If all the children are holes, then zero them all out so that
3264 * we may get compressed away.
3266 if (i == 1 << epbs) {
3267 /* didn't find any non-holes */
3268 bzero(db->db.db_data, db->db.db_size);
3274 * The SPA will call this callback several times for each zio - once
3275 * for every physical child i/o (zio->io_phys_children times). This
3276 * allows the DMU to monitor the progress of each logical i/o. For example,
3277 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
3278 * block. There may be a long delay before all copies/fragments are completed,
3279 * so this callback allows us to retire dirty space gradually, as the physical
3284 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
3286 dmu_buf_impl_t *db = arg;
3287 objset_t *os = db->db_objset;
3288 dsl_pool_t *dp = dmu_objset_pool(os);
3289 dbuf_dirty_record_t *dr;
3292 dr = db->db_data_pending;
3293 ASSERT3U(dr->dr_txg, ==, zio->io_txg);
3296 * The callback will be called io_phys_children times. Retire one
3297 * portion of our dirty space each time we are called. Any rounding
3298 * error will be cleaned up by dsl_pool_sync()'s call to
3299 * dsl_pool_undirty_space().
3301 delta = dr->dr_accounted / zio->io_phys_children;
3302 dsl_pool_undirty_space(dp, delta, zio->io_txg);
3307 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
3309 dmu_buf_impl_t *db = vdb;
3310 blkptr_t *bp_orig = &zio->io_bp_orig;
3311 blkptr_t *bp = db->db_blkptr;
3312 objset_t *os = db->db_objset;
3313 dmu_tx_t *tx = os->os_synctx;
3314 dbuf_dirty_record_t **drp, *dr;
3316 ASSERT0(zio->io_error);
3317 ASSERT(db->db_blkptr == bp);
3320 * For nopwrites and rewrites we ensure that the bp matches our
3321 * original and bypass all the accounting.
3323 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
3324 ASSERT(BP_EQUAL(bp, bp_orig));
3326 dsl_dataset_t *ds = os->os_dsl_dataset;
3327 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
3328 dsl_dataset_block_born(ds, bp, tx);
3331 mutex_enter(&db->db_mtx);
3335 drp = &db->db_last_dirty;
3336 while ((dr = *drp) != db->db_data_pending)
3338 ASSERT(!list_link_active(&dr->dr_dirty_node));
3339 ASSERT(dr->dr_dbuf == db);
3340 ASSERT(dr->dr_next == NULL);
3344 if (db->db_blkid == DMU_SPILL_BLKID) {
3349 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3350 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
3351 db->db_blkptr == &dn->dn_phys->dn_spill);
3356 if (db->db_level == 0) {
3357 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
3358 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
3359 if (db->db_state != DB_NOFILL) {
3360 if (dr->dt.dl.dr_data != db->db_buf)
3361 arc_buf_destroy(dr->dt.dl.dr_data, db);
3368 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3369 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
3370 if (!BP_IS_HOLE(db->db_blkptr)) {
3372 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3373 ASSERT3U(db->db_blkid, <=,
3374 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs));
3375 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
3379 mutex_destroy(&dr->dt.di.dr_mtx);
3380 list_destroy(&dr->dt.di.dr_children);
3382 kmem_free(dr, sizeof (dbuf_dirty_record_t));
3384 cv_broadcast(&db->db_changed);
3385 ASSERT(db->db_dirtycnt > 0);
3386 db->db_dirtycnt -= 1;
3387 db->db_data_pending = NULL;
3388 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg);
3392 dbuf_write_nofill_ready(zio_t *zio)
3394 dbuf_write_ready(zio, NULL, zio->io_private);
3398 dbuf_write_nofill_done(zio_t *zio)
3400 dbuf_write_done(zio, NULL, zio->io_private);
3404 dbuf_write_override_ready(zio_t *zio)
3406 dbuf_dirty_record_t *dr = zio->io_private;
3407 dmu_buf_impl_t *db = dr->dr_dbuf;
3409 dbuf_write_ready(zio, NULL, db);
3413 dbuf_write_override_done(zio_t *zio)
3415 dbuf_dirty_record_t *dr = zio->io_private;
3416 dmu_buf_impl_t *db = dr->dr_dbuf;
3417 blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
3419 mutex_enter(&db->db_mtx);
3420 if (!BP_EQUAL(zio->io_bp, obp)) {
3421 if (!BP_IS_HOLE(obp))
3422 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
3423 arc_release(dr->dt.dl.dr_data, db);
3425 mutex_exit(&db->db_mtx);
3427 dbuf_write_done(zio, NULL, db);
3430 /* Issue I/O to commit a dirty buffer to disk. */
3432 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
3434 dmu_buf_impl_t *db = dr->dr_dbuf;
3437 dmu_buf_impl_t *parent = db->db_parent;
3438 uint64_t txg = tx->tx_txg;
3439 zbookmark_phys_t zb;
3444 ASSERT(dmu_tx_is_syncing(tx));
3450 if (db->db_state != DB_NOFILL) {
3451 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
3453 * Private object buffers are released here rather
3454 * than in dbuf_dirty() since they are only modified
3455 * in the syncing context and we don't want the
3456 * overhead of making multiple copies of the data.
3458 if (BP_IS_HOLE(db->db_blkptr)) {
3461 dbuf_release_bp(db);
3466 if (parent != dn->dn_dbuf) {
3467 /* Our parent is an indirect block. */
3468 /* We have a dirty parent that has been scheduled for write. */
3469 ASSERT(parent && parent->db_data_pending);
3470 /* Our parent's buffer is one level closer to the dnode. */
3471 ASSERT(db->db_level == parent->db_level-1);
3473 * We're about to modify our parent's db_data by modifying
3474 * our block pointer, so the parent must be released.
3476 ASSERT(arc_released(parent->db_buf));
3477 zio = parent->db_data_pending->dr_zio;
3479 /* Our parent is the dnode itself. */
3480 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
3481 db->db_blkid != DMU_SPILL_BLKID) ||
3482 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
3483 if (db->db_blkid != DMU_SPILL_BLKID)
3484 ASSERT3P(db->db_blkptr, ==,
3485 &dn->dn_phys->dn_blkptr[db->db_blkid]);
3489 ASSERT(db->db_level == 0 || data == db->db_buf);
3490 ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
3493 SET_BOOKMARK(&zb, os->os_dsl_dataset ?
3494 os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
3495 db->db.db_object, db->db_level, db->db_blkid);
3497 if (db->db_blkid == DMU_SPILL_BLKID)
3499 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
3501 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
3505 * We copy the blkptr now (rather than when we instantiate the dirty
3506 * record), because its value can change between open context and
3507 * syncing context. We do not need to hold dn_struct_rwlock to read
3508 * db_blkptr because we are in syncing context.
3510 dr->dr_bp_copy = *db->db_blkptr;
3512 if (db->db_level == 0 &&
3513 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
3515 * The BP for this block has been provided by open context
3516 * (by dmu_sync() or dmu_buf_write_embedded()).
3518 void *contents = (data != NULL) ? data->b_data : NULL;
3520 dr->dr_zio = zio_write(zio, os->os_spa, txg,
3521 &dr->dr_bp_copy, contents, db->db.db_size, &zp,
3522 dbuf_write_override_ready, NULL, NULL,
3523 dbuf_write_override_done,
3524 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
3525 mutex_enter(&db->db_mtx);
3526 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
3527 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
3528 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
3529 mutex_exit(&db->db_mtx);
3530 } else if (db->db_state == DB_NOFILL) {
3531 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF ||
3532 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY);
3533 dr->dr_zio = zio_write(zio, os->os_spa, txg,
3534 &dr->dr_bp_copy, NULL, db->db.db_size, &zp,
3535 dbuf_write_nofill_ready, NULL, NULL,
3536 dbuf_write_nofill_done, db,
3537 ZIO_PRIORITY_ASYNC_WRITE,
3538 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
3540 ASSERT(arc_released(data));
3543 * For indirect blocks, we want to setup the children
3544 * ready callback so that we can properly handle an indirect
3545 * block that only contains holes.
3547 arc_done_func_t *children_ready_cb = NULL;
3548 if (db->db_level != 0)
3549 children_ready_cb = dbuf_write_children_ready;
3551 dr->dr_zio = arc_write(zio, os->os_spa, txg,
3552 &dr->dr_bp_copy, data, DBUF_IS_L2CACHEABLE(db),
3553 &zp, dbuf_write_ready, children_ready_cb,
3554 dbuf_write_physdone, dbuf_write_done, db,
3555 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);