4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
24 * Copyright (c) 2012, 2017 by Delphix. All rights reserved.
25 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
26 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
29 #include <sys/zfs_context.h>
32 #include <sys/dmu_send.h>
33 #include <sys/dmu_impl.h>
35 #include <sys/dmu_objset.h>
36 #include <sys/dsl_dataset.h>
37 #include <sys/dsl_dir.h>
38 #include <sys/dmu_tx.h>
41 #include <sys/dmu_zfetch.h>
43 #include <sys/sa_impl.h>
44 #include <sys/zfeature.h>
45 #include <sys/blkptr.h>
46 #include <sys/range_tree.h>
47 #include <sys/trace_dbuf.h>
48 #include <sys/callb.h>
53 typedef struct dbuf_stats {
55 * Various statistics about the size of the dbuf cache.
57 kstat_named_t cache_count;
58 kstat_named_t cache_size_bytes;
59 kstat_named_t cache_size_bytes_max;
61 * Statistics regarding the bounds on the dbuf cache size.
63 kstat_named_t cache_target_bytes;
64 kstat_named_t cache_lowater_bytes;
65 kstat_named_t cache_hiwater_bytes;
67 * Total number of dbuf cache evictions that have occurred.
69 kstat_named_t cache_total_evicts;
71 * The distribution of dbuf levels in the dbuf cache and
72 * the total size of all dbufs at each level.
74 kstat_named_t cache_levels[DN_MAX_LEVELS];
75 kstat_named_t cache_levels_bytes[DN_MAX_LEVELS];
77 * Statistics about the dbuf hash table.
79 kstat_named_t hash_hits;
80 kstat_named_t hash_misses;
81 kstat_named_t hash_collisions;
82 kstat_named_t hash_elements;
83 kstat_named_t hash_elements_max;
85 * Number of sublists containing more than one dbuf in the dbuf
86 * hash table. Keep track of the longest hash chain.
88 kstat_named_t hash_chains;
89 kstat_named_t hash_chain_max;
91 * Number of times a dbuf_create() discovers that a dbuf was
92 * already created and in the dbuf hash table.
94 kstat_named_t hash_insert_race;
97 dbuf_stats_t dbuf_stats = {
98 { "cache_count", KSTAT_DATA_UINT64 },
99 { "cache_size_bytes", KSTAT_DATA_UINT64 },
100 { "cache_size_bytes_max", KSTAT_DATA_UINT64 },
101 { "cache_target_bytes", KSTAT_DATA_UINT64 },
102 { "cache_lowater_bytes", KSTAT_DATA_UINT64 },
103 { "cache_hiwater_bytes", KSTAT_DATA_UINT64 },
104 { "cache_total_evicts", KSTAT_DATA_UINT64 },
105 { { "cache_levels_N", KSTAT_DATA_UINT64 } },
106 { { "cache_levels_bytes_N", KSTAT_DATA_UINT64 } },
107 { "hash_hits", KSTAT_DATA_UINT64 },
108 { "hash_misses", KSTAT_DATA_UINT64 },
109 { "hash_collisions", KSTAT_DATA_UINT64 },
110 { "hash_elements", KSTAT_DATA_UINT64 },
111 { "hash_elements_max", KSTAT_DATA_UINT64 },
112 { "hash_chains", KSTAT_DATA_UINT64 },
113 { "hash_chain_max", KSTAT_DATA_UINT64 },
114 { "hash_insert_race", KSTAT_DATA_UINT64 }
117 #define DBUF_STAT_INCR(stat, val) \
118 atomic_add_64(&dbuf_stats.stat.value.ui64, (val));
119 #define DBUF_STAT_DECR(stat, val) \
120 DBUF_STAT_INCR(stat, -(val));
121 #define DBUF_STAT_BUMP(stat) \
122 DBUF_STAT_INCR(stat, 1);
123 #define DBUF_STAT_BUMPDOWN(stat) \
124 DBUF_STAT_INCR(stat, -1);
125 #define DBUF_STAT_MAX(stat, v) { \
127 while ((v) > (_m = dbuf_stats.stat.value.ui64) && \
128 (_m != atomic_cas_64(&dbuf_stats.stat.value.ui64, _m, (v))))\
132 struct dbuf_hold_impl_data {
133 /* Function arguments */
137 boolean_t dh_fail_sparse;
138 boolean_t dh_fail_uncached;
140 dmu_buf_impl_t **dh_dbp;
141 /* Local variables */
142 dmu_buf_impl_t *dh_db;
143 dmu_buf_impl_t *dh_parent;
146 dbuf_dirty_record_t *dh_dr;
150 static void __dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh,
151 dnode_t *dn, uint8_t level, uint64_t blkid, boolean_t fail_sparse,
152 boolean_t fail_uncached,
153 void *tag, dmu_buf_impl_t **dbp, int depth);
154 static int __dbuf_hold_impl(struct dbuf_hold_impl_data *dh);
156 uint_t zfs_dbuf_evict_key;
158 static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
159 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
161 extern inline void dmu_buf_init_user(dmu_buf_user_t *dbu,
162 dmu_buf_evict_func_t *evict_func_sync,
163 dmu_buf_evict_func_t *evict_func_async,
164 dmu_buf_t **clear_on_evict_dbufp);
167 * Global data structures and functions for the dbuf cache.
169 static kmem_cache_t *dbuf_kmem_cache;
170 static taskq_t *dbu_evict_taskq;
172 static kthread_t *dbuf_cache_evict_thread;
173 static kmutex_t dbuf_evict_lock;
174 static kcondvar_t dbuf_evict_cv;
175 static boolean_t dbuf_evict_thread_exit;
178 * LRU cache of dbufs. The dbuf cache maintains a list of dbufs that
179 * are not currently held but have been recently released. These dbufs
180 * are not eligible for arc eviction until they are aged out of the cache.
181 * Dbufs are added to the dbuf cache once the last hold is released. If a
182 * dbuf is later accessed and still exists in the dbuf cache, then it will
183 * be removed from the cache and later re-added to the head of the cache.
184 * Dbufs that are aged out of the cache will be immediately destroyed and
185 * become eligible for arc eviction.
187 static multilist_t *dbuf_cache;
188 static refcount_t dbuf_cache_size;
189 unsigned long dbuf_cache_max_bytes = 100 * 1024 * 1024;
191 /* Cap the size of the dbuf cache to log2 fraction of arc size. */
192 int dbuf_cache_max_shift = 5;
195 * The dbuf cache uses a three-stage eviction policy:
196 * - A low water marker designates when the dbuf eviction thread
197 * should stop evicting from the dbuf cache.
198 * - When we reach the maximum size (aka mid water mark), we
199 * signal the eviction thread to run.
200 * - The high water mark indicates when the eviction thread
201 * is unable to keep up with the incoming load and eviction must
202 * happen in the context of the calling thread.
206 * low water mid water hi water
207 * +----------------------------------------+----------+----------+
212 * +----------------------------------------+----------+----------+
214 * evicting eviction directly
217 * The high and low water marks indicate the operating range for the eviction
218 * thread. The low water mark is, by default, 90% of the total size of the
219 * cache and the high water mark is at 110% (both of these percentages can be
220 * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct,
221 * respectively). The eviction thread will try to ensure that the cache remains
222 * within this range by waking up every second and checking if the cache is
223 * above the low water mark. The thread can also be woken up by callers adding
224 * elements into the cache if the cache is larger than the mid water (i.e max
225 * cache size). Once the eviction thread is woken up and eviction is required,
226 * it will continue evicting buffers until it's able to reduce the cache size
227 * to the low water mark. If the cache size continues to grow and hits the high
228 * water mark, then callers adding elements to the cache will begin to evict
229 * directly from the cache until the cache is no longer above the high water
234 * The percentage above and below the maximum cache size.
236 uint_t dbuf_cache_hiwater_pct = 10;
237 uint_t dbuf_cache_lowater_pct = 10;
241 dbuf_cons(void *vdb, void *unused, int kmflag)
243 dmu_buf_impl_t *db = vdb;
244 bzero(db, sizeof (dmu_buf_impl_t));
246 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
247 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
248 multilist_link_init(&db->db_cache_link);
249 refcount_create(&db->db_holds);
256 dbuf_dest(void *vdb, void *unused)
258 dmu_buf_impl_t *db = vdb;
259 mutex_destroy(&db->db_mtx);
260 cv_destroy(&db->db_changed);
261 ASSERT(!multilist_link_active(&db->db_cache_link));
262 refcount_destroy(&db->db_holds);
266 * dbuf hash table routines
268 static dbuf_hash_table_t dbuf_hash_table;
270 static uint64_t dbuf_hash_count;
273 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
275 uintptr_t osv = (uintptr_t)os;
276 uint64_t crc = -1ULL;
278 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
279 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF];
280 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF];
281 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF];
282 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF];
283 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF];
284 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF];
286 crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16);
291 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
292 ((dbuf)->db.db_object == (obj) && \
293 (dbuf)->db_objset == (os) && \
294 (dbuf)->db_level == (level) && \
295 (dbuf)->db_blkid == (blkid))
298 dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid)
300 dbuf_hash_table_t *h = &dbuf_hash_table;
305 hv = dbuf_hash(os, obj, level, blkid);
306 idx = hv & h->hash_table_mask;
308 mutex_enter(DBUF_HASH_MUTEX(h, idx));
309 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
310 if (DBUF_EQUAL(db, os, obj, level, blkid)) {
311 mutex_enter(&db->db_mtx);
312 if (db->db_state != DB_EVICTING) {
313 mutex_exit(DBUF_HASH_MUTEX(h, idx));
316 mutex_exit(&db->db_mtx);
319 mutex_exit(DBUF_HASH_MUTEX(h, idx));
323 static dmu_buf_impl_t *
324 dbuf_find_bonus(objset_t *os, uint64_t object)
327 dmu_buf_impl_t *db = NULL;
329 if (dnode_hold(os, object, FTAG, &dn) == 0) {
330 rw_enter(&dn->dn_struct_rwlock, RW_READER);
331 if (dn->dn_bonus != NULL) {
333 mutex_enter(&db->db_mtx);
335 rw_exit(&dn->dn_struct_rwlock);
336 dnode_rele(dn, FTAG);
342 * Insert an entry into the hash table. If there is already an element
343 * equal to elem in the hash table, then the already existing element
344 * will be returned and the new element will not be inserted.
345 * Otherwise returns NULL.
347 static dmu_buf_impl_t *
348 dbuf_hash_insert(dmu_buf_impl_t *db)
350 dbuf_hash_table_t *h = &dbuf_hash_table;
351 objset_t *os = db->db_objset;
352 uint64_t obj = db->db.db_object;
353 int level = db->db_level;
354 uint64_t blkid, hv, idx;
358 blkid = db->db_blkid;
359 hv = dbuf_hash(os, obj, level, blkid);
360 idx = hv & h->hash_table_mask;
362 mutex_enter(DBUF_HASH_MUTEX(h, idx));
363 for (dbf = h->hash_table[idx], i = 0; dbf != NULL;
364 dbf = dbf->db_hash_next, i++) {
365 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
366 mutex_enter(&dbf->db_mtx);
367 if (dbf->db_state != DB_EVICTING) {
368 mutex_exit(DBUF_HASH_MUTEX(h, idx));
371 mutex_exit(&dbf->db_mtx);
376 DBUF_STAT_BUMP(hash_collisions);
378 DBUF_STAT_BUMP(hash_chains);
380 DBUF_STAT_MAX(hash_chain_max, i);
383 mutex_enter(&db->db_mtx);
384 db->db_hash_next = h->hash_table[idx];
385 h->hash_table[idx] = db;
386 mutex_exit(DBUF_HASH_MUTEX(h, idx));
387 atomic_inc_64(&dbuf_hash_count);
388 DBUF_STAT_MAX(hash_elements_max, dbuf_hash_count);
394 * Remove an entry from the hash table. It must be in the EVICTING state.
397 dbuf_hash_remove(dmu_buf_impl_t *db)
399 dbuf_hash_table_t *h = &dbuf_hash_table;
401 dmu_buf_impl_t *dbf, **dbp;
403 hv = dbuf_hash(db->db_objset, db->db.db_object,
404 db->db_level, db->db_blkid);
405 idx = hv & h->hash_table_mask;
408 * We mustn't hold db_mtx to maintain lock ordering:
409 * DBUF_HASH_MUTEX > db_mtx.
411 ASSERT(refcount_is_zero(&db->db_holds));
412 ASSERT(db->db_state == DB_EVICTING);
413 ASSERT(!MUTEX_HELD(&db->db_mtx));
415 mutex_enter(DBUF_HASH_MUTEX(h, idx));
416 dbp = &h->hash_table[idx];
417 while ((dbf = *dbp) != db) {
418 dbp = &dbf->db_hash_next;
421 *dbp = db->db_hash_next;
422 db->db_hash_next = NULL;
423 if (h->hash_table[idx] &&
424 h->hash_table[idx]->db_hash_next == NULL)
425 DBUF_STAT_BUMPDOWN(hash_chains);
426 mutex_exit(DBUF_HASH_MUTEX(h, idx));
427 atomic_dec_64(&dbuf_hash_count);
433 } dbvu_verify_type_t;
436 dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type)
441 if (db->db_user == NULL)
444 /* Only data blocks support the attachment of user data. */
445 ASSERT(db->db_level == 0);
447 /* Clients must resolve a dbuf before attaching user data. */
448 ASSERT(db->db.db_data != NULL);
449 ASSERT3U(db->db_state, ==, DB_CACHED);
451 holds = refcount_count(&db->db_holds);
452 if (verify_type == DBVU_EVICTING) {
454 * Immediate eviction occurs when holds == dirtycnt.
455 * For normal eviction buffers, holds is zero on
456 * eviction, except when dbuf_fix_old_data() calls
457 * dbuf_clear_data(). However, the hold count can grow
458 * during eviction even though db_mtx is held (see
459 * dmu_bonus_hold() for an example), so we can only
460 * test the generic invariant that holds >= dirtycnt.
462 ASSERT3U(holds, >=, db->db_dirtycnt);
464 if (db->db_user_immediate_evict == TRUE)
465 ASSERT3U(holds, >=, db->db_dirtycnt);
467 ASSERT3U(holds, >, 0);
473 dbuf_evict_user(dmu_buf_impl_t *db)
475 dmu_buf_user_t *dbu = db->db_user;
477 ASSERT(MUTEX_HELD(&db->db_mtx));
482 dbuf_verify_user(db, DBVU_EVICTING);
486 if (dbu->dbu_clear_on_evict_dbufp != NULL)
487 *dbu->dbu_clear_on_evict_dbufp = NULL;
491 * There are two eviction callbacks - one that we call synchronously
492 * and one that we invoke via a taskq. The async one is useful for
493 * avoiding lock order reversals and limiting stack depth.
495 * Note that if we have a sync callback but no async callback,
496 * it's likely that the sync callback will free the structure
497 * containing the dbu. In that case we need to take care to not
498 * dereference dbu after calling the sync evict func.
500 boolean_t has_async = (dbu->dbu_evict_func_async != NULL);
502 if (dbu->dbu_evict_func_sync != NULL)
503 dbu->dbu_evict_func_sync(dbu);
506 taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func_async,
507 dbu, 0, &dbu->dbu_tqent);
512 dbuf_is_metadata(dmu_buf_impl_t *db)
515 * Consider indirect blocks and spill blocks to be meta data.
517 if (db->db_level > 0 || db->db_blkid == DMU_SPILL_BLKID) {
520 boolean_t is_metadata;
523 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type);
526 return (is_metadata);
532 * This function *must* return indices evenly distributed between all
533 * sublists of the multilist. This is needed due to how the dbuf eviction
534 * code is laid out; dbuf_evict_thread() assumes dbufs are evenly
535 * distributed between all sublists and uses this assumption when
536 * deciding which sublist to evict from and how much to evict from it.
539 dbuf_cache_multilist_index_func(multilist_t *ml, void *obj)
541 dmu_buf_impl_t *db = obj;
544 * The assumption here, is the hash value for a given
545 * dmu_buf_impl_t will remain constant throughout it's lifetime
546 * (i.e. it's objset, object, level and blkid fields don't change).
547 * Thus, we don't need to store the dbuf's sublist index
548 * on insertion, as this index can be recalculated on removal.
550 * Also, the low order bits of the hash value are thought to be
551 * distributed evenly. Otherwise, in the case that the multilist
552 * has a power of two number of sublists, each sublists' usage
553 * would not be evenly distributed.
555 return (dbuf_hash(db->db_objset, db->db.db_object,
556 db->db_level, db->db_blkid) %
557 multilist_get_num_sublists(ml));
560 static inline unsigned long
561 dbuf_cache_target_bytes(void)
563 return MIN(dbuf_cache_max_bytes,
564 arc_target_bytes() >> dbuf_cache_max_shift);
567 static inline uint64_t
568 dbuf_cache_hiwater_bytes(void)
570 uint64_t dbuf_cache_target = dbuf_cache_target_bytes();
571 return (dbuf_cache_target +
572 (dbuf_cache_target * dbuf_cache_hiwater_pct) / 100);
575 static inline uint64_t
576 dbuf_cache_lowater_bytes(void)
578 uint64_t dbuf_cache_target = dbuf_cache_target_bytes();
579 return (dbuf_cache_target -
580 (dbuf_cache_target * dbuf_cache_lowater_pct) / 100);
583 static inline boolean_t
584 dbuf_cache_above_hiwater(void)
586 return (refcount_count(&dbuf_cache_size) > dbuf_cache_hiwater_bytes());
589 static inline boolean_t
590 dbuf_cache_above_lowater(void)
592 return (refcount_count(&dbuf_cache_size) > dbuf_cache_lowater_bytes());
596 * Evict the oldest eligible dbuf from the dbuf cache.
601 int idx = multilist_get_random_index(dbuf_cache);
602 multilist_sublist_t *mls = multilist_sublist_lock(dbuf_cache, idx);
604 ASSERT(!MUTEX_HELD(&dbuf_evict_lock));
607 * Set the thread's tsd to indicate that it's processing evictions.
608 * Once a thread stops evicting from the dbuf cache it will
609 * reset its tsd to NULL.
611 ASSERT3P(tsd_get(zfs_dbuf_evict_key), ==, NULL);
612 (void) tsd_set(zfs_dbuf_evict_key, (void *)B_TRUE);
614 dmu_buf_impl_t *db = multilist_sublist_tail(mls);
615 while (db != NULL && mutex_tryenter(&db->db_mtx) == 0) {
616 db = multilist_sublist_prev(mls, db);
619 DTRACE_PROBE2(dbuf__evict__one, dmu_buf_impl_t *, db,
620 multilist_sublist_t *, mls);
623 multilist_sublist_remove(mls, db);
624 multilist_sublist_unlock(mls);
625 (void) refcount_remove_many(&dbuf_cache_size,
627 DBUF_STAT_BUMPDOWN(cache_levels[db->db_level]);
628 DBUF_STAT_BUMPDOWN(cache_count);
629 DBUF_STAT_DECR(cache_levels_bytes[db->db_level],
632 DBUF_STAT_MAX(cache_size_bytes_max,
633 refcount_count(&dbuf_cache_size));
634 DBUF_STAT_BUMP(cache_total_evicts);
636 multilist_sublist_unlock(mls);
638 (void) tsd_set(zfs_dbuf_evict_key, NULL);
642 * The dbuf evict thread is responsible for aging out dbufs from the
643 * cache. Once the cache has reached it's maximum size, dbufs are removed
644 * and destroyed. The eviction thread will continue running until the size
645 * of the dbuf cache is at or below the maximum size. Once the dbuf is aged
646 * out of the cache it is destroyed and becomes eligible for arc eviction.
650 dbuf_evict_thread(void *unused)
654 CALLB_CPR_INIT(&cpr, &dbuf_evict_lock, callb_generic_cpr, FTAG);
656 mutex_enter(&dbuf_evict_lock);
657 while (!dbuf_evict_thread_exit) {
658 while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
659 CALLB_CPR_SAFE_BEGIN(&cpr);
660 (void) cv_timedwait_sig_hires(&dbuf_evict_cv,
661 &dbuf_evict_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
662 CALLB_CPR_SAFE_END(&cpr, &dbuf_evict_lock);
664 mutex_exit(&dbuf_evict_lock);
667 * Keep evicting as long as we're above the low water mark
668 * for the cache. We do this without holding the locks to
669 * minimize lock contention.
671 while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
675 mutex_enter(&dbuf_evict_lock);
678 dbuf_evict_thread_exit = B_FALSE;
679 cv_broadcast(&dbuf_evict_cv);
680 CALLB_CPR_EXIT(&cpr); /* drops dbuf_evict_lock */
685 * Wake up the dbuf eviction thread if the dbuf cache is at its max size.
686 * If the dbuf cache is at its high water mark, then evict a dbuf from the
687 * dbuf cache using the callers context.
690 dbuf_evict_notify(void)
694 * We use thread specific data to track when a thread has
695 * started processing evictions. This allows us to avoid deeply
696 * nested stacks that would have a call flow similar to this:
698 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify()
701 * +-----dbuf_destroy()<--dbuf_evict_one()<--------+
703 * The dbuf_eviction_thread will always have its tsd set until
704 * that thread exits. All other threads will only set their tsd
705 * if they are participating in the eviction process. This only
706 * happens if the eviction thread is unable to process evictions
707 * fast enough. To keep the dbuf cache size in check, other threads
708 * can evict from the dbuf cache directly. Those threads will set
709 * their tsd values so that we ensure that they only evict one dbuf
710 * from the dbuf cache.
712 if (tsd_get(zfs_dbuf_evict_key) != NULL)
716 * We check if we should evict without holding the dbuf_evict_lock,
717 * because it's OK to occasionally make the wrong decision here,
718 * and grabbing the lock results in massive lock contention.
720 if (refcount_count(&dbuf_cache_size) > dbuf_cache_target_bytes()) {
721 if (dbuf_cache_above_hiwater())
723 cv_signal(&dbuf_evict_cv);
728 dbuf_kstat_update(kstat_t *ksp, int rw)
730 dbuf_stats_t *ds = ksp->ks_data;
732 if (rw == KSTAT_WRITE) {
733 return (SET_ERROR(EACCES));
735 ds->cache_size_bytes.value.ui64 =
736 refcount_count(&dbuf_cache_size);
737 ds->cache_target_bytes.value.ui64 = dbuf_cache_target_bytes();
738 ds->cache_hiwater_bytes.value.ui64 = dbuf_cache_hiwater_bytes();
739 ds->cache_lowater_bytes.value.ui64 = dbuf_cache_lowater_bytes();
740 ds->hash_elements.value.ui64 = dbuf_hash_count;
749 uint64_t hsize = 1ULL << 16;
750 dbuf_hash_table_t *h = &dbuf_hash_table;
754 * The hash table is big enough to fill all of physical memory
755 * with an average block size of zfs_arc_average_blocksize (default 8K).
756 * By default, the table will take up
757 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
759 while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE)
763 h->hash_table_mask = hsize - 1;
764 #if defined(_KERNEL) && defined(HAVE_SPL)
766 * Large allocations which do not require contiguous pages
767 * should be using vmem_alloc() in the linux kernel
769 h->hash_table = vmem_zalloc(hsize * sizeof (void *), KM_SLEEP);
771 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
773 if (h->hash_table == NULL) {
774 /* XXX - we should really return an error instead of assert */
775 ASSERT(hsize > (1ULL << 10));
780 dbuf_kmem_cache = kmem_cache_create("dmu_buf_impl_t",
781 sizeof (dmu_buf_impl_t),
782 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
784 for (i = 0; i < DBUF_MUTEXES; i++)
785 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
790 * Setup the parameters for the dbuf cache. We cap the size of the
791 * dbuf cache to 1/32nd (default) of the size of the ARC.
793 dbuf_cache_max_bytes = MIN(dbuf_cache_max_bytes,
794 arc_target_bytes() >> dbuf_cache_max_shift);
797 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
798 * configuration is not required.
800 dbu_evict_taskq = taskq_create("dbu_evict", 1, defclsyspri, 0, 0, 0);
802 dbuf_cache = multilist_create(sizeof (dmu_buf_impl_t),
803 offsetof(dmu_buf_impl_t, db_cache_link),
804 dbuf_cache_multilist_index_func);
805 refcount_create(&dbuf_cache_size);
807 tsd_create(&zfs_dbuf_evict_key, NULL);
808 dbuf_evict_thread_exit = B_FALSE;
809 mutex_init(&dbuf_evict_lock, NULL, MUTEX_DEFAULT, NULL);
810 cv_init(&dbuf_evict_cv, NULL, CV_DEFAULT, NULL);
811 dbuf_cache_evict_thread = thread_create(NULL, 0, dbuf_evict_thread,
812 NULL, 0, &p0, TS_RUN, minclsyspri);
814 dbuf_ksp = kstat_create("zfs", 0, "dbufstats", "misc",
815 KSTAT_TYPE_NAMED, sizeof (dbuf_stats) / sizeof (kstat_named_t),
817 if (dbuf_ksp != NULL) {
818 dbuf_ksp->ks_data = &dbuf_stats;
819 dbuf_ksp->ks_update = dbuf_kstat_update;
820 kstat_install(dbuf_ksp);
822 for (i = 0; i < DN_MAX_LEVELS; i++) {
823 snprintf(dbuf_stats.cache_levels[i].name,
824 KSTAT_STRLEN, "cache_level_%d", i);
825 dbuf_stats.cache_levels[i].data_type =
827 snprintf(dbuf_stats.cache_levels_bytes[i].name,
828 KSTAT_STRLEN, "cache_level_%d_bytes", i);
829 dbuf_stats.cache_levels_bytes[i].data_type =
838 dbuf_hash_table_t *h = &dbuf_hash_table;
841 dbuf_stats_destroy();
843 for (i = 0; i < DBUF_MUTEXES; i++)
844 mutex_destroy(&h->hash_mutexes[i]);
845 #if defined(_KERNEL) && defined(HAVE_SPL)
847 * Large allocations which do not require contiguous pages
848 * should be using vmem_free() in the linux kernel
850 vmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
852 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
854 kmem_cache_destroy(dbuf_kmem_cache);
855 taskq_destroy(dbu_evict_taskq);
857 mutex_enter(&dbuf_evict_lock);
858 dbuf_evict_thread_exit = B_TRUE;
859 while (dbuf_evict_thread_exit) {
860 cv_signal(&dbuf_evict_cv);
861 cv_wait(&dbuf_evict_cv, &dbuf_evict_lock);
863 mutex_exit(&dbuf_evict_lock);
864 tsd_destroy(&zfs_dbuf_evict_key);
866 mutex_destroy(&dbuf_evict_lock);
867 cv_destroy(&dbuf_evict_cv);
869 refcount_destroy(&dbuf_cache_size);
870 multilist_destroy(dbuf_cache);
872 if (dbuf_ksp != NULL) {
873 kstat_delete(dbuf_ksp);
884 dbuf_verify(dmu_buf_impl_t *db)
887 dbuf_dirty_record_t *dr;
889 ASSERT(MUTEX_HELD(&db->db_mtx));
891 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
894 ASSERT(db->db_objset != NULL);
898 ASSERT(db->db_parent == NULL);
899 ASSERT(db->db_blkptr == NULL);
901 ASSERT3U(db->db.db_object, ==, dn->dn_object);
902 ASSERT3P(db->db_objset, ==, dn->dn_objset);
903 ASSERT3U(db->db_level, <, dn->dn_nlevels);
904 ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
905 db->db_blkid == DMU_SPILL_BLKID ||
906 !avl_is_empty(&dn->dn_dbufs));
908 if (db->db_blkid == DMU_BONUS_BLKID) {
910 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
911 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
912 } else if (db->db_blkid == DMU_SPILL_BLKID) {
914 ASSERT0(db->db.db_offset);
916 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
919 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
920 ASSERT(dr->dr_dbuf == db);
922 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
923 ASSERT(dr->dr_dbuf == db);
926 * We can't assert that db_size matches dn_datablksz because it
927 * can be momentarily different when another thread is doing
930 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
931 dr = db->db_data_pending;
933 * It should only be modified in syncing context, so
934 * make sure we only have one copy of the data.
936 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
939 /* verify db->db_blkptr */
941 if (db->db_parent == dn->dn_dbuf) {
942 /* db is pointed to by the dnode */
943 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
944 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
945 ASSERT(db->db_parent == NULL);
947 ASSERT(db->db_parent != NULL);
948 if (db->db_blkid != DMU_SPILL_BLKID)
949 ASSERT3P(db->db_blkptr, ==,
950 &dn->dn_phys->dn_blkptr[db->db_blkid]);
952 /* db is pointed to by an indirect block */
953 ASSERTV(int epb = db->db_parent->db.db_size >>
955 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
956 ASSERT3U(db->db_parent->db.db_object, ==,
959 * dnode_grow_indblksz() can make this fail if we don't
960 * have the struct_rwlock. XXX indblksz no longer
961 * grows. safe to do this now?
963 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
964 ASSERT3P(db->db_blkptr, ==,
965 ((blkptr_t *)db->db_parent->db.db_data +
966 db->db_blkid % epb));
970 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
971 (db->db_buf == NULL || db->db_buf->b_data) &&
972 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
973 db->db_state != DB_FILL && !dn->dn_free_txg) {
975 * If the blkptr isn't set but they have nonzero data,
976 * it had better be dirty, otherwise we'll lose that
977 * data when we evict this buffer.
979 * There is an exception to this rule for indirect blocks; in
980 * this case, if the indirect block is a hole, we fill in a few
981 * fields on each of the child blocks (importantly, birth time)
982 * to prevent hole birth times from being lost when you
983 * partially fill in a hole.
985 if (db->db_dirtycnt == 0) {
986 if (db->db_level == 0) {
987 uint64_t *buf = db->db.db_data;
990 for (i = 0; i < db->db.db_size >> 3; i++) {
994 blkptr_t *bps = db->db.db_data;
995 ASSERT3U(1 << DB_DNODE(db)->dn_indblkshift, ==,
998 * We want to verify that all the blkptrs in the
999 * indirect block are holes, but we may have
1000 * automatically set up a few fields for them.
1001 * We iterate through each blkptr and verify
1002 * they only have those fields set.
1005 i < db->db.db_size / sizeof (blkptr_t);
1007 blkptr_t *bp = &bps[i];
1008 ASSERT(ZIO_CHECKSUM_IS_ZERO(
1011 DVA_IS_EMPTY(&bp->blk_dva[0]) &&
1012 DVA_IS_EMPTY(&bp->blk_dva[1]) &&
1013 DVA_IS_EMPTY(&bp->blk_dva[2]));
1014 ASSERT0(bp->blk_fill);
1015 ASSERT0(bp->blk_pad[0]);
1016 ASSERT0(bp->blk_pad[1]);
1017 ASSERT(!BP_IS_EMBEDDED(bp));
1018 ASSERT(BP_IS_HOLE(bp));
1019 ASSERT0(bp->blk_phys_birth);
1029 dbuf_clear_data(dmu_buf_impl_t *db)
1031 ASSERT(MUTEX_HELD(&db->db_mtx));
1032 dbuf_evict_user(db);
1033 ASSERT3P(db->db_buf, ==, NULL);
1034 db->db.db_data = NULL;
1035 if (db->db_state != DB_NOFILL)
1036 db->db_state = DB_UNCACHED;
1040 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
1042 ASSERT(MUTEX_HELD(&db->db_mtx));
1043 ASSERT(buf != NULL);
1046 ASSERT(buf->b_data != NULL);
1047 db->db.db_data = buf->b_data;
1051 * Loan out an arc_buf for read. Return the loaned arc_buf.
1054 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
1058 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1059 mutex_enter(&db->db_mtx);
1060 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
1061 int blksz = db->db.db_size;
1062 spa_t *spa = db->db_objset->os_spa;
1064 mutex_exit(&db->db_mtx);
1065 abuf = arc_loan_buf(spa, B_FALSE, blksz);
1066 bcopy(db->db.db_data, abuf->b_data, blksz);
1069 arc_loan_inuse_buf(abuf, db);
1071 dbuf_clear_data(db);
1072 mutex_exit(&db->db_mtx);
1078 * Calculate which level n block references the data at the level 0 offset
1082 dbuf_whichblock(const dnode_t *dn, const int64_t level, const uint64_t offset)
1084 if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) {
1086 * The level n blkid is equal to the level 0 blkid divided by
1087 * the number of level 0s in a level n block.
1089 * The level 0 blkid is offset >> datablkshift =
1090 * offset / 2^datablkshift.
1092 * The number of level 0s in a level n is the number of block
1093 * pointers in an indirect block, raised to the power of level.
1094 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
1095 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
1097 * Thus, the level n blkid is: offset /
1098 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
1099 * = offset / 2^(datablkshift + level *
1100 * (indblkshift - SPA_BLKPTRSHIFT))
1101 * = offset >> (datablkshift + level *
1102 * (indblkshift - SPA_BLKPTRSHIFT))
1105 const unsigned exp = dn->dn_datablkshift +
1106 level * (dn->dn_indblkshift - SPA_BLKPTRSHIFT);
1108 if (exp >= 8 * sizeof (offset)) {
1109 /* This only happens on the highest indirection level */
1110 ASSERT3U(level, ==, dn->dn_nlevels - 1);
1114 ASSERT3U(exp, <, 8 * sizeof (offset));
1116 return (offset >> exp);
1118 ASSERT3U(offset, <, dn->dn_datablksz);
1124 dbuf_read_done(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
1125 arc_buf_t *buf, void *vdb)
1127 dmu_buf_impl_t *db = vdb;
1129 mutex_enter(&db->db_mtx);
1130 ASSERT3U(db->db_state, ==, DB_READ);
1132 * All reads are synchronous, so we must have a hold on the dbuf
1134 ASSERT(refcount_count(&db->db_holds) > 0);
1135 ASSERT(db->db_buf == NULL);
1136 ASSERT(db->db.db_data == NULL);
1137 if (db->db_level == 0 && db->db_freed_in_flight) {
1138 /* we were freed in flight; disregard any error */
1140 buf = arc_alloc_buf(db->db_objset->os_spa,
1141 db, DBUF_GET_BUFC_TYPE(db), db->db.db_size);
1143 arc_release(buf, db);
1144 bzero(buf->b_data, db->db.db_size);
1145 arc_buf_freeze(buf);
1146 db->db_freed_in_flight = FALSE;
1147 dbuf_set_data(db, buf);
1148 db->db_state = DB_CACHED;
1149 } else if (buf != NULL) {
1150 dbuf_set_data(db, buf);
1151 db->db_state = DB_CACHED;
1153 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1154 ASSERT3P(db->db_buf, ==, NULL);
1155 db->db_state = DB_UNCACHED;
1157 cv_broadcast(&db->db_changed);
1158 dbuf_rele_and_unlock(db, NULL);
1162 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
1165 zbookmark_phys_t zb;
1166 uint32_t aflags = ARC_FLAG_NOWAIT;
1167 int err, zio_flags = 0;
1171 ASSERT(!refcount_is_zero(&db->db_holds));
1172 /* We need the struct_rwlock to prevent db_blkptr from changing. */
1173 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1174 ASSERT(MUTEX_HELD(&db->db_mtx));
1175 ASSERT(db->db_state == DB_UNCACHED);
1176 ASSERT(db->db_buf == NULL);
1178 if (db->db_blkid == DMU_BONUS_BLKID) {
1180 * The bonus length stored in the dnode may be less than
1181 * the maximum available space in the bonus buffer.
1183 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
1184 int max_bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
1185 arc_buf_t *dn_buf = (dn->dn_dbuf != NULL) ?
1186 dn->dn_dbuf->db_buf : NULL;
1188 /* if the underlying dnode block is encrypted, decrypt it */
1189 if (dn_buf != NULL && dn->dn_objset->os_encrypted &&
1190 DMU_OT_IS_ENCRYPTED(dn->dn_bonustype) &&
1191 (flags & DB_RF_NO_DECRYPT) == 0 &&
1192 arc_is_encrypted(dn_buf)) {
1193 err = arc_untransform(dn_buf, dn->dn_objset->os_spa,
1194 dmu_objset_id(dn->dn_objset), B_TRUE);
1197 mutex_exit(&db->db_mtx);
1202 ASSERT3U(bonuslen, <=, db->db.db_size);
1203 db->db.db_data = kmem_alloc(max_bonuslen, KM_SLEEP);
1204 arc_space_consume(max_bonuslen, ARC_SPACE_BONUS);
1205 if (bonuslen < max_bonuslen)
1206 bzero(db->db.db_data, max_bonuslen);
1208 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
1210 db->db_state = DB_CACHED;
1211 mutex_exit(&db->db_mtx);
1216 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
1217 * processes the delete record and clears the bp while we are waiting
1218 * for the dn_mtx (resulting in a "no" from block_freed).
1220 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
1221 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
1222 BP_IS_HOLE(db->db_blkptr)))) {
1223 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1225 dbuf_set_data(db, arc_alloc_buf(db->db_objset->os_spa, db, type,
1227 bzero(db->db.db_data, db->db.db_size);
1229 if (db->db_blkptr != NULL && db->db_level > 0 &&
1230 BP_IS_HOLE(db->db_blkptr) &&
1231 db->db_blkptr->blk_birth != 0) {
1232 blkptr_t *bps = db->db.db_data;
1233 for (int i = 0; i < ((1 <<
1234 DB_DNODE(db)->dn_indblkshift) / sizeof (blkptr_t));
1236 blkptr_t *bp = &bps[i];
1237 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
1238 1 << dn->dn_indblkshift);
1240 BP_GET_LEVEL(db->db_blkptr) == 1 ?
1242 BP_GET_LSIZE(db->db_blkptr));
1243 BP_SET_TYPE(bp, BP_GET_TYPE(db->db_blkptr));
1245 BP_GET_LEVEL(db->db_blkptr) - 1);
1246 BP_SET_BIRTH(bp, db->db_blkptr->blk_birth, 0);
1250 db->db_state = DB_CACHED;
1251 mutex_exit(&db->db_mtx);
1257 db->db_state = DB_READ;
1258 mutex_exit(&db->db_mtx);
1260 if (DBUF_IS_L2CACHEABLE(db))
1261 aflags |= ARC_FLAG_L2CACHE;
1263 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
1264 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
1265 db->db.db_object, db->db_level, db->db_blkid);
1268 * All bps of an encrypted os should have the encryption bit set.
1269 * If this is not true it indicates tampering and we report an error.
1271 if (db->db_objset->os_encrypted && !BP_USES_CRYPT(db->db_blkptr)) {
1272 spa_log_error(db->db_objset->os_spa, &zb);
1273 zfs_panic_recover("unencrypted block in encrypted "
1274 "object set %llu", dmu_objset_id(db->db_objset));
1275 return (SET_ERROR(EIO));
1278 dbuf_add_ref(db, NULL);
1280 zio_flags = (flags & DB_RF_CANFAIL) ?
1281 ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED;
1283 if ((flags & DB_RF_NO_DECRYPT) && BP_IS_PROTECTED(db->db_blkptr))
1284 zio_flags |= ZIO_FLAG_RAW;
1286 err = arc_read(zio, db->db_objset->os_spa, db->db_blkptr,
1287 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ, zio_flags,
1294 * This is our just-in-time copy function. It makes a copy of buffers that
1295 * have been modified in a previous transaction group before we access them in
1296 * the current active group.
1298 * This function is used in three places: when we are dirtying a buffer for the
1299 * first time in a txg, when we are freeing a range in a dnode that includes
1300 * this buffer, and when we are accessing a buffer which was received compressed
1301 * and later referenced in a WRITE_BYREF record.
1303 * Note that when we are called from dbuf_free_range() we do not put a hold on
1304 * the buffer, we just traverse the active dbuf list for the dnode.
1307 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
1309 dbuf_dirty_record_t *dr = db->db_last_dirty;
1311 ASSERT(MUTEX_HELD(&db->db_mtx));
1312 ASSERT(db->db.db_data != NULL);
1313 ASSERT(db->db_level == 0);
1314 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
1317 (dr->dt.dl.dr_data !=
1318 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
1322 * If the last dirty record for this dbuf has not yet synced
1323 * and its referencing the dbuf data, either:
1324 * reset the reference to point to a new copy,
1325 * or (if there a no active holders)
1326 * just null out the current db_data pointer.
1328 ASSERT3U(dr->dr_txg, >=, txg - 2);
1329 if (db->db_blkid == DMU_BONUS_BLKID) {
1330 dnode_t *dn = DB_DNODE(db);
1331 int bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
1332 dr->dt.dl.dr_data = kmem_alloc(bonuslen, KM_SLEEP);
1333 arc_space_consume(bonuslen, ARC_SPACE_BONUS);
1334 bcopy(db->db.db_data, dr->dt.dl.dr_data, bonuslen);
1335 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
1336 dnode_t *dn = DB_DNODE(db);
1337 int size = arc_buf_size(db->db_buf);
1338 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1339 spa_t *spa = db->db_objset->os_spa;
1340 enum zio_compress compress_type =
1341 arc_get_compression(db->db_buf);
1343 if (arc_is_encrypted(db->db_buf)) {
1344 boolean_t byteorder;
1345 uint8_t salt[ZIO_DATA_SALT_LEN];
1346 uint8_t iv[ZIO_DATA_IV_LEN];
1347 uint8_t mac[ZIO_DATA_MAC_LEN];
1349 arc_get_raw_params(db->db_buf, &byteorder, salt,
1351 dr->dt.dl.dr_data = arc_alloc_raw_buf(spa, db,
1352 dmu_objset_id(dn->dn_objset), byteorder, salt, iv,
1353 mac, dn->dn_type, size, arc_buf_lsize(db->db_buf),
1355 } else if (compress_type != ZIO_COMPRESS_OFF) {
1356 ASSERT3U(type, ==, ARC_BUFC_DATA);
1357 dr->dt.dl.dr_data = arc_alloc_compressed_buf(spa, db,
1358 size, arc_buf_lsize(db->db_buf), compress_type);
1360 dr->dt.dl.dr_data = arc_alloc_buf(spa, db, type, size);
1362 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
1365 dbuf_clear_data(db);
1370 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
1377 * We don't have to hold the mutex to check db_state because it
1378 * can't be freed while we have a hold on the buffer.
1380 ASSERT(!refcount_is_zero(&db->db_holds));
1382 if (db->db_state == DB_NOFILL)
1383 return (SET_ERROR(EIO));
1387 if ((flags & DB_RF_HAVESTRUCT) == 0)
1388 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1390 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1391 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
1392 DBUF_IS_CACHEABLE(db);
1394 mutex_enter(&db->db_mtx);
1395 if (db->db_state == DB_CACHED) {
1396 spa_t *spa = dn->dn_objset->os_spa;
1399 * If the arc buf is compressed or encrypted, we need to
1400 * untransform it to read the data. This could happen during
1401 * the "zfs receive" of a stream which is deduplicated and
1402 * either raw or compressed. We do not need to do this if the
1403 * caller wants raw encrypted data.
1405 if (db->db_buf != NULL && (flags & DB_RF_NO_DECRYPT) == 0 &&
1406 (arc_is_encrypted(db->db_buf) ||
1407 arc_get_compression(db->db_buf) != ZIO_COMPRESS_OFF)) {
1408 dbuf_fix_old_data(db, spa_syncing_txg(spa));
1409 err = arc_untransform(db->db_buf, spa,
1410 dmu_objset_id(db->db_objset), B_FALSE);
1411 dbuf_set_data(db, db->db_buf);
1413 mutex_exit(&db->db_mtx);
1415 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1416 if ((flags & DB_RF_HAVESTRUCT) == 0)
1417 rw_exit(&dn->dn_struct_rwlock);
1419 DBUF_STAT_BUMP(hash_hits);
1420 } else if (db->db_state == DB_UNCACHED) {
1421 spa_t *spa = dn->dn_objset->os_spa;
1422 boolean_t need_wait = B_FALSE;
1425 db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) {
1426 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
1429 err = dbuf_read_impl(db, zio, flags);
1431 /* dbuf_read_impl has dropped db_mtx for us */
1433 if (!err && prefetch)
1434 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1436 if ((flags & DB_RF_HAVESTRUCT) == 0)
1437 rw_exit(&dn->dn_struct_rwlock);
1439 DBUF_STAT_BUMP(hash_misses);
1441 if (!err && need_wait)
1442 err = zio_wait(zio);
1445 * Another reader came in while the dbuf was in flight
1446 * between UNCACHED and CACHED. Either a writer will finish
1447 * writing the buffer (sending the dbuf to CACHED) or the
1448 * first reader's request will reach the read_done callback
1449 * and send the dbuf to CACHED. Otherwise, a failure
1450 * occurred and the dbuf went to UNCACHED.
1452 mutex_exit(&db->db_mtx);
1454 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1455 if ((flags & DB_RF_HAVESTRUCT) == 0)
1456 rw_exit(&dn->dn_struct_rwlock);
1458 DBUF_STAT_BUMP(hash_misses);
1460 /* Skip the wait per the caller's request. */
1461 mutex_enter(&db->db_mtx);
1462 if ((flags & DB_RF_NEVERWAIT) == 0) {
1463 while (db->db_state == DB_READ ||
1464 db->db_state == DB_FILL) {
1465 ASSERT(db->db_state == DB_READ ||
1466 (flags & DB_RF_HAVESTRUCT) == 0);
1467 DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *,
1469 cv_wait(&db->db_changed, &db->db_mtx);
1471 if (db->db_state == DB_UNCACHED)
1472 err = SET_ERROR(EIO);
1474 mutex_exit(&db->db_mtx);
1481 dbuf_noread(dmu_buf_impl_t *db)
1483 ASSERT(!refcount_is_zero(&db->db_holds));
1484 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1485 mutex_enter(&db->db_mtx);
1486 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1487 cv_wait(&db->db_changed, &db->db_mtx);
1488 if (db->db_state == DB_UNCACHED) {
1489 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1490 spa_t *spa = db->db_objset->os_spa;
1492 ASSERT(db->db_buf == NULL);
1493 ASSERT(db->db.db_data == NULL);
1494 dbuf_set_data(db, arc_alloc_buf(spa, db, type, db->db.db_size));
1495 db->db_state = DB_FILL;
1496 } else if (db->db_state == DB_NOFILL) {
1497 dbuf_clear_data(db);
1499 ASSERT3U(db->db_state, ==, DB_CACHED);
1501 mutex_exit(&db->db_mtx);
1505 dbuf_unoverride(dbuf_dirty_record_t *dr)
1507 dmu_buf_impl_t *db = dr->dr_dbuf;
1508 blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
1509 uint64_t txg = dr->dr_txg;
1511 ASSERT(MUTEX_HELD(&db->db_mtx));
1513 * This assert is valid because dmu_sync() expects to be called by
1514 * a zilog's get_data while holding a range lock. This call only
1515 * comes from dbuf_dirty() callers who must also hold a range lock.
1517 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
1518 ASSERT(db->db_level == 0);
1520 if (db->db_blkid == DMU_BONUS_BLKID ||
1521 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
1524 ASSERT(db->db_data_pending != dr);
1526 /* free this block */
1527 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite)
1528 zio_free(db->db_objset->os_spa, txg, bp);
1530 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1531 dr->dt.dl.dr_nopwrite = B_FALSE;
1532 dr->dt.dl.dr_raw = B_FALSE;
1535 * Release the already-written buffer, so we leave it in
1536 * a consistent dirty state. Note that all callers are
1537 * modifying the buffer, so they will immediately do
1538 * another (redundant) arc_release(). Therefore, leave
1539 * the buf thawed to save the effort of freezing &
1540 * immediately re-thawing it.
1542 arc_release(dr->dt.dl.dr_data, db);
1546 * Evict (if its unreferenced) or clear (if its referenced) any level-0
1547 * data blocks in the free range, so that any future readers will find
1551 dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
1554 dmu_buf_impl_t *db_search;
1555 dmu_buf_impl_t *db, *db_next;
1556 uint64_t txg = tx->tx_txg;
1559 if (end_blkid > dn->dn_maxblkid &&
1560 !(start_blkid == DMU_SPILL_BLKID || end_blkid == DMU_SPILL_BLKID))
1561 end_blkid = dn->dn_maxblkid;
1562 dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid);
1564 db_search = kmem_alloc(sizeof (dmu_buf_impl_t), KM_SLEEP);
1565 db_search->db_level = 0;
1566 db_search->db_blkid = start_blkid;
1567 db_search->db_state = DB_SEARCH;
1569 mutex_enter(&dn->dn_dbufs_mtx);
1570 db = avl_find(&dn->dn_dbufs, db_search, &where);
1571 ASSERT3P(db, ==, NULL);
1573 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1575 for (; db != NULL; db = db_next) {
1576 db_next = AVL_NEXT(&dn->dn_dbufs, db);
1577 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1579 if (db->db_level != 0 || db->db_blkid > end_blkid) {
1582 ASSERT3U(db->db_blkid, >=, start_blkid);
1584 /* found a level 0 buffer in the range */
1585 mutex_enter(&db->db_mtx);
1586 if (dbuf_undirty(db, tx)) {
1587 /* mutex has been dropped and dbuf destroyed */
1591 if (db->db_state == DB_UNCACHED ||
1592 db->db_state == DB_NOFILL ||
1593 db->db_state == DB_EVICTING) {
1594 ASSERT(db->db.db_data == NULL);
1595 mutex_exit(&db->db_mtx);
1598 if (db->db_state == DB_READ || db->db_state == DB_FILL) {
1599 /* will be handled in dbuf_read_done or dbuf_rele */
1600 db->db_freed_in_flight = TRUE;
1601 mutex_exit(&db->db_mtx);
1604 if (refcount_count(&db->db_holds) == 0) {
1609 /* The dbuf is referenced */
1611 if (db->db_last_dirty != NULL) {
1612 dbuf_dirty_record_t *dr = db->db_last_dirty;
1614 if (dr->dr_txg == txg) {
1616 * This buffer is "in-use", re-adjust the file
1617 * size to reflect that this buffer may
1618 * contain new data when we sync.
1620 if (db->db_blkid != DMU_SPILL_BLKID &&
1621 db->db_blkid > dn->dn_maxblkid)
1622 dn->dn_maxblkid = db->db_blkid;
1623 dbuf_unoverride(dr);
1626 * This dbuf is not dirty in the open context.
1627 * Either uncache it (if its not referenced in
1628 * the open context) or reset its contents to
1631 dbuf_fix_old_data(db, txg);
1634 /* clear the contents if its cached */
1635 if (db->db_state == DB_CACHED) {
1636 ASSERT(db->db.db_data != NULL);
1637 arc_release(db->db_buf, db);
1638 bzero(db->db.db_data, db->db.db_size);
1639 arc_buf_freeze(db->db_buf);
1642 mutex_exit(&db->db_mtx);
1645 kmem_free(db_search, sizeof (dmu_buf_impl_t));
1646 mutex_exit(&dn->dn_dbufs_mtx);
1650 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
1652 arc_buf_t *buf, *obuf;
1653 int osize = db->db.db_size;
1654 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1657 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1662 /* XXX does *this* func really need the lock? */
1663 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1666 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1667 * is OK, because there can be no other references to the db
1668 * when we are changing its size, so no concurrent DB_FILL can
1672 * XXX we should be doing a dbuf_read, checking the return
1673 * value and returning that up to our callers
1675 dmu_buf_will_dirty(&db->db, tx);
1677 /* create the data buffer for the new block */
1678 buf = arc_alloc_buf(dn->dn_objset->os_spa, db, type, size);
1680 /* copy old block data to the new block */
1682 bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
1683 /* zero the remainder */
1685 bzero((uint8_t *)buf->b_data + osize, size - osize);
1687 mutex_enter(&db->db_mtx);
1688 dbuf_set_data(db, buf);
1689 arc_buf_destroy(obuf, db);
1690 db->db.db_size = size;
1692 if (db->db_level == 0) {
1693 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1694 db->db_last_dirty->dt.dl.dr_data = buf;
1696 mutex_exit(&db->db_mtx);
1698 dmu_objset_willuse_space(dn->dn_objset, size - osize, tx);
1703 dbuf_release_bp(dmu_buf_impl_t *db)
1705 ASSERTV(objset_t *os = db->db_objset);
1707 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1708 ASSERT(arc_released(os->os_phys_buf) ||
1709 list_link_active(&os->os_dsl_dataset->ds_synced_link));
1710 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1712 (void) arc_release(db->db_buf, db);
1716 * We already have a dirty record for this TXG, and we are being
1720 dbuf_redirty(dbuf_dirty_record_t *dr)
1722 dmu_buf_impl_t *db = dr->dr_dbuf;
1724 ASSERT(MUTEX_HELD(&db->db_mtx));
1726 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1728 * If this buffer has already been written out,
1729 * we now need to reset its state.
1731 dbuf_unoverride(dr);
1732 if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1733 db->db_state != DB_NOFILL) {
1734 /* Already released on initial dirty, so just thaw. */
1735 ASSERT(arc_released(db->db_buf));
1736 arc_buf_thaw(db->db_buf);
1741 dbuf_dirty_record_t *
1742 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1746 dbuf_dirty_record_t **drp, *dr;
1747 int drop_struct_lock = FALSE;
1748 int txgoff = tx->tx_txg & TXG_MASK;
1750 ASSERT(tx->tx_txg != 0);
1751 ASSERT(!refcount_is_zero(&db->db_holds));
1752 DMU_TX_DIRTY_BUF(tx, db);
1757 * Shouldn't dirty a regular buffer in syncing context. Private
1758 * objects may be dirtied in syncing context, but only if they
1759 * were already pre-dirtied in open context.
1762 if (dn->dn_objset->os_dsl_dataset != NULL) {
1763 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1766 ASSERT(!dmu_tx_is_syncing(tx) ||
1767 BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1768 DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1769 dn->dn_objset->os_dsl_dataset == NULL);
1770 if (dn->dn_objset->os_dsl_dataset != NULL)
1771 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, FTAG);
1774 * We make this assert for private objects as well, but after we
1775 * check if we're already dirty. They are allowed to re-dirty
1776 * in syncing context.
1778 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1779 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1780 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1782 mutex_enter(&db->db_mtx);
1784 * XXX make this true for indirects too? The problem is that
1785 * transactions created with dmu_tx_create_assigned() from
1786 * syncing context don't bother holding ahead.
1788 ASSERT(db->db_level != 0 ||
1789 db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1790 db->db_state == DB_NOFILL);
1792 mutex_enter(&dn->dn_mtx);
1794 * Don't set dirtyctx to SYNC if we're just modifying this as we
1795 * initialize the objset.
1797 if (dn->dn_dirtyctx == DN_UNDIRTIED) {
1798 if (dn->dn_objset->os_dsl_dataset != NULL) {
1799 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1802 if (!BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1803 dn->dn_dirtyctx = (dmu_tx_is_syncing(tx) ?
1804 DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1805 ASSERT(dn->dn_dirtyctx_firstset == NULL);
1806 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1808 if (dn->dn_objset->os_dsl_dataset != NULL) {
1809 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1813 mutex_exit(&dn->dn_mtx);
1815 if (db->db_blkid == DMU_SPILL_BLKID)
1816 dn->dn_have_spill = B_TRUE;
1819 * If this buffer is already dirty, we're done.
1821 drp = &db->db_last_dirty;
1822 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1823 db->db.db_object == DMU_META_DNODE_OBJECT);
1824 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1826 if (dr && dr->dr_txg == tx->tx_txg) {
1830 mutex_exit(&db->db_mtx);
1835 * Only valid if not already dirty.
1837 ASSERT(dn->dn_object == 0 ||
1838 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1839 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1841 ASSERT3U(dn->dn_nlevels, >, db->db_level);
1844 * We should only be dirtying in syncing context if it's the
1845 * mos or we're initializing the os or it's a special object.
1846 * However, we are allowed to dirty in syncing context provided
1847 * we already dirtied it in open context. Hence we must make
1848 * this assertion only if we're not already dirty.
1851 VERIFY3U(tx->tx_txg, <=, spa_final_dirty_txg(os->os_spa));
1853 if (dn->dn_objset->os_dsl_dataset != NULL)
1854 rrw_enter(&os->os_dsl_dataset->ds_bp_rwlock, RW_READER, FTAG);
1855 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1856 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1857 if (dn->dn_objset->os_dsl_dataset != NULL)
1858 rrw_exit(&os->os_dsl_dataset->ds_bp_rwlock, FTAG);
1860 ASSERT(db->db.db_size != 0);
1862 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1864 if (db->db_blkid != DMU_BONUS_BLKID) {
1865 dmu_objset_willuse_space(os, db->db.db_size, tx);
1869 * If this buffer is dirty in an old transaction group we need
1870 * to make a copy of it so that the changes we make in this
1871 * transaction group won't leak out when we sync the older txg.
1873 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1874 list_link_init(&dr->dr_dirty_node);
1875 if (db->db_level == 0) {
1876 void *data_old = db->db_buf;
1878 if (db->db_state != DB_NOFILL) {
1879 if (db->db_blkid == DMU_BONUS_BLKID) {
1880 dbuf_fix_old_data(db, tx->tx_txg);
1881 data_old = db->db.db_data;
1882 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1884 * Release the data buffer from the cache so
1885 * that we can modify it without impacting
1886 * possible other users of this cached data
1887 * block. Note that indirect blocks and
1888 * private objects are not released until the
1889 * syncing state (since they are only modified
1892 arc_release(db->db_buf, db);
1893 dbuf_fix_old_data(db, tx->tx_txg);
1894 data_old = db->db_buf;
1896 ASSERT(data_old != NULL);
1898 dr->dt.dl.dr_data = data_old;
1900 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_NOLOCKDEP, NULL);
1901 list_create(&dr->dt.di.dr_children,
1902 sizeof (dbuf_dirty_record_t),
1903 offsetof(dbuf_dirty_record_t, dr_dirty_node));
1905 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL)
1906 dr->dr_accounted = db->db.db_size;
1908 dr->dr_txg = tx->tx_txg;
1913 * We could have been freed_in_flight between the dbuf_noread
1914 * and dbuf_dirty. We win, as though the dbuf_noread() had
1915 * happened after the free.
1917 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1918 db->db_blkid != DMU_SPILL_BLKID) {
1919 mutex_enter(&dn->dn_mtx);
1920 if (dn->dn_free_ranges[txgoff] != NULL) {
1921 range_tree_clear(dn->dn_free_ranges[txgoff],
1924 mutex_exit(&dn->dn_mtx);
1925 db->db_freed_in_flight = FALSE;
1929 * This buffer is now part of this txg
1931 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1932 db->db_dirtycnt += 1;
1933 ASSERT3U(db->db_dirtycnt, <=, 3);
1935 mutex_exit(&db->db_mtx);
1937 if (db->db_blkid == DMU_BONUS_BLKID ||
1938 db->db_blkid == DMU_SPILL_BLKID) {
1939 mutex_enter(&dn->dn_mtx);
1940 ASSERT(!list_link_active(&dr->dr_dirty_node));
1941 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1942 mutex_exit(&dn->dn_mtx);
1943 dnode_setdirty(dn, tx);
1949 * The dn_struct_rwlock prevents db_blkptr from changing
1950 * due to a write from syncing context completing
1951 * while we are running, so we want to acquire it before
1952 * looking at db_blkptr.
1954 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1955 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1956 drop_struct_lock = TRUE;
1960 * We need to hold the dn_struct_rwlock to make this assertion,
1961 * because it protects dn_phys / dn_next_nlevels from changing.
1963 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1964 dn->dn_phys->dn_nlevels > db->db_level ||
1965 dn->dn_next_nlevels[txgoff] > db->db_level ||
1966 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1967 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1970 * If we are overwriting a dedup BP, then unless it is snapshotted,
1971 * when we get to syncing context we will need to decrement its
1972 * refcount in the DDT. Prefetch the relevant DDT block so that
1973 * syncing context won't have to wait for the i/o.
1975 ddt_prefetch(os->os_spa, db->db_blkptr);
1977 if (db->db_level == 0) {
1978 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1979 ASSERT(dn->dn_maxblkid >= db->db_blkid);
1982 if (db->db_level+1 < dn->dn_nlevels) {
1983 dmu_buf_impl_t *parent = db->db_parent;
1984 dbuf_dirty_record_t *di;
1985 int parent_held = FALSE;
1987 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1988 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1990 parent = dbuf_hold_level(dn, db->db_level+1,
1991 db->db_blkid >> epbs, FTAG);
1992 ASSERT(parent != NULL);
1995 if (drop_struct_lock)
1996 rw_exit(&dn->dn_struct_rwlock);
1997 ASSERT3U(db->db_level+1, ==, parent->db_level);
1998 di = dbuf_dirty(parent, tx);
2000 dbuf_rele(parent, FTAG);
2002 mutex_enter(&db->db_mtx);
2004 * Since we've dropped the mutex, it's possible that
2005 * dbuf_undirty() might have changed this out from under us.
2007 if (db->db_last_dirty == dr ||
2008 dn->dn_object == DMU_META_DNODE_OBJECT) {
2009 mutex_enter(&di->dt.di.dr_mtx);
2010 ASSERT3U(di->dr_txg, ==, tx->tx_txg);
2011 ASSERT(!list_link_active(&dr->dr_dirty_node));
2012 list_insert_tail(&di->dt.di.dr_children, dr);
2013 mutex_exit(&di->dt.di.dr_mtx);
2016 mutex_exit(&db->db_mtx);
2018 ASSERT(db->db_level+1 == dn->dn_nlevels);
2019 ASSERT(db->db_blkid < dn->dn_nblkptr);
2020 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
2021 mutex_enter(&dn->dn_mtx);
2022 ASSERT(!list_link_active(&dr->dr_dirty_node));
2023 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
2024 mutex_exit(&dn->dn_mtx);
2025 if (drop_struct_lock)
2026 rw_exit(&dn->dn_struct_rwlock);
2029 dnode_setdirty(dn, tx);
2035 * Undirty a buffer in the transaction group referenced by the given
2036 * transaction. Return whether this evicted the dbuf.
2039 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
2042 uint64_t txg = tx->tx_txg;
2043 dbuf_dirty_record_t *dr, **drp;
2048 * Due to our use of dn_nlevels below, this can only be called
2049 * in open context, unless we are operating on the MOS.
2050 * From syncing context, dn_nlevels may be different from the
2051 * dn_nlevels used when dbuf was dirtied.
2053 ASSERT(db->db_objset ==
2054 dmu_objset_pool(db->db_objset)->dp_meta_objset ||
2055 txg != spa_syncing_txg(dmu_objset_spa(db->db_objset)));
2056 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2057 ASSERT0(db->db_level);
2058 ASSERT(MUTEX_HELD(&db->db_mtx));
2061 * If this buffer is not dirty, we're done.
2063 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
2064 if (dr->dr_txg <= txg)
2066 if (dr == NULL || dr->dr_txg < txg)
2068 ASSERT(dr->dr_txg == txg);
2069 ASSERT(dr->dr_dbuf == db);
2074 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
2076 ASSERT(db->db.db_size != 0);
2078 dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset),
2079 dr->dr_accounted, txg);
2084 * Note that there are three places in dbuf_dirty()
2085 * where this dirty record may be put on a list.
2086 * Make sure to do a list_remove corresponding to
2087 * every one of those list_insert calls.
2089 if (dr->dr_parent) {
2090 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
2091 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
2092 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
2093 } else if (db->db_blkid == DMU_SPILL_BLKID ||
2094 db->db_level + 1 == dn->dn_nlevels) {
2095 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
2096 mutex_enter(&dn->dn_mtx);
2097 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
2098 mutex_exit(&dn->dn_mtx);
2102 if (db->db_state != DB_NOFILL) {
2103 dbuf_unoverride(dr);
2105 ASSERT(db->db_buf != NULL);
2106 ASSERT(dr->dt.dl.dr_data != NULL);
2107 if (dr->dt.dl.dr_data != db->db_buf)
2108 arc_buf_destroy(dr->dt.dl.dr_data, db);
2111 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2113 ASSERT(db->db_dirtycnt > 0);
2114 db->db_dirtycnt -= 1;
2116 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
2117 ASSERT(db->db_state == DB_NOFILL || arc_released(db->db_buf));
2126 dmu_buf_will_dirty_impl(dmu_buf_t *db_fake, int flags, dmu_tx_t *tx)
2128 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2130 ASSERT(tx->tx_txg != 0);
2131 ASSERT(!refcount_is_zero(&db->db_holds));
2134 * Quick check for dirtyness. For already dirty blocks, this
2135 * reduces runtime of this function by >90%, and overall performance
2136 * by 50% for some workloads (e.g. file deletion with indirect blocks
2139 mutex_enter(&db->db_mtx);
2141 dbuf_dirty_record_t *dr;
2142 for (dr = db->db_last_dirty;
2143 dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) {
2145 * It's possible that it is already dirty but not cached,
2146 * because there are some calls to dbuf_dirty() that don't
2147 * go through dmu_buf_will_dirty().
2149 if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) {
2150 /* This dbuf is already dirty and cached. */
2152 mutex_exit(&db->db_mtx);
2156 mutex_exit(&db->db_mtx);
2159 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
2160 flags |= DB_RF_HAVESTRUCT;
2162 (void) dbuf_read(db, NULL, flags);
2163 (void) dbuf_dirty(db, tx);
2167 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
2169 dmu_buf_will_dirty_impl(db_fake,
2170 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH, tx);
2174 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
2176 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2178 db->db_state = DB_NOFILL;
2180 dmu_buf_will_fill(db_fake, tx);
2184 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
2186 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2188 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2189 ASSERT(tx->tx_txg != 0);
2190 ASSERT(db->db_level == 0);
2191 ASSERT(!refcount_is_zero(&db->db_holds));
2193 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
2194 dmu_tx_private_ok(tx));
2197 (void) dbuf_dirty(db, tx);
2201 * This function is effectively the same as dmu_buf_will_dirty(), but
2202 * indicates the caller expects raw encrypted data in the db. It will
2203 * also set the raw flag on the created dirty record.
2206 dmu_buf_will_change_crypt_params(dmu_buf_t *db_fake, dmu_tx_t *tx)
2208 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2209 dbuf_dirty_record_t *dr;
2211 dmu_buf_will_dirty_impl(db_fake,
2212 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_NO_DECRYPT, tx);
2214 dr = db->db_last_dirty;
2215 while (dr != NULL && dr->dr_txg > tx->tx_txg)
2218 ASSERT3P(dr, !=, NULL);
2219 ASSERT3U(dr->dr_txg, ==, tx->tx_txg);
2220 dr->dt.dl.dr_raw = B_TRUE;
2223 #pragma weak dmu_buf_fill_done = dbuf_fill_done
2226 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
2228 mutex_enter(&db->db_mtx);
2231 if (db->db_state == DB_FILL) {
2232 if (db->db_level == 0 && db->db_freed_in_flight) {
2233 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2234 /* we were freed while filling */
2235 /* XXX dbuf_undirty? */
2236 bzero(db->db.db_data, db->db.db_size);
2237 db->db_freed_in_flight = FALSE;
2239 db->db_state = DB_CACHED;
2240 cv_broadcast(&db->db_changed);
2242 mutex_exit(&db->db_mtx);
2246 dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data,
2247 bp_embedded_type_t etype, enum zio_compress comp,
2248 int uncompressed_size, int compressed_size, int byteorder,
2251 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2252 struct dirty_leaf *dl;
2253 dmu_object_type_t type;
2255 if (etype == BP_EMBEDDED_TYPE_DATA) {
2256 ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset),
2257 SPA_FEATURE_EMBEDDED_DATA));
2261 type = DB_DNODE(db)->dn_type;
2264 ASSERT0(db->db_level);
2265 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2267 dmu_buf_will_not_fill(dbuf, tx);
2269 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
2270 dl = &db->db_last_dirty->dt.dl;
2271 encode_embedded_bp_compressed(&dl->dr_overridden_by,
2272 data, comp, uncompressed_size, compressed_size);
2273 BPE_SET_ETYPE(&dl->dr_overridden_by, etype);
2274 BP_SET_TYPE(&dl->dr_overridden_by, type);
2275 BP_SET_LEVEL(&dl->dr_overridden_by, 0);
2276 BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder);
2278 dl->dr_override_state = DR_OVERRIDDEN;
2279 dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg;
2283 * Directly assign a provided arc buf to a given dbuf if it's not referenced
2284 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
2287 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
2289 ASSERT(!refcount_is_zero(&db->db_holds));
2290 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2291 ASSERT(db->db_level == 0);
2292 ASSERT3U(dbuf_is_metadata(db), ==, arc_is_metadata(buf));
2293 ASSERT(buf != NULL);
2294 ASSERT(arc_buf_lsize(buf) == db->db.db_size);
2295 ASSERT(tx->tx_txg != 0);
2297 arc_return_buf(buf, db);
2298 ASSERT(arc_released(buf));
2300 mutex_enter(&db->db_mtx);
2302 while (db->db_state == DB_READ || db->db_state == DB_FILL)
2303 cv_wait(&db->db_changed, &db->db_mtx);
2305 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
2307 if (db->db_state == DB_CACHED &&
2308 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
2310 * In practice, we will never have a case where we have an
2311 * encrypted arc buffer while additional holds exist on the
2312 * dbuf. We don't handle this here so we simply assert that
2315 ASSERT(!arc_is_encrypted(buf));
2316 mutex_exit(&db->db_mtx);
2317 (void) dbuf_dirty(db, tx);
2318 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
2319 arc_buf_destroy(buf, db);
2320 xuio_stat_wbuf_copied();
2324 xuio_stat_wbuf_nocopy();
2325 if (db->db_state == DB_CACHED) {
2326 dbuf_dirty_record_t *dr = db->db_last_dirty;
2328 ASSERT(db->db_buf != NULL);
2329 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
2330 ASSERT(dr->dt.dl.dr_data == db->db_buf);
2331 IMPLY(arc_is_encrypted(buf), dr->dt.dl.dr_raw);
2333 if (!arc_released(db->db_buf)) {
2334 ASSERT(dr->dt.dl.dr_override_state ==
2336 arc_release(db->db_buf, db);
2338 dr->dt.dl.dr_data = buf;
2339 arc_buf_destroy(db->db_buf, db);
2340 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
2341 arc_release(db->db_buf, db);
2342 arc_buf_destroy(db->db_buf, db);
2346 ASSERT(db->db_buf == NULL);
2347 dbuf_set_data(db, buf);
2348 db->db_state = DB_FILL;
2349 mutex_exit(&db->db_mtx);
2350 (void) dbuf_dirty(db, tx);
2351 dmu_buf_fill_done(&db->db, tx);
2355 dbuf_destroy(dmu_buf_impl_t *db)
2358 dmu_buf_impl_t *parent = db->db_parent;
2359 dmu_buf_impl_t *dndb;
2361 ASSERT(MUTEX_HELD(&db->db_mtx));
2362 ASSERT(refcount_is_zero(&db->db_holds));
2364 if (db->db_buf != NULL) {
2365 arc_buf_destroy(db->db_buf, db);
2369 if (db->db_blkid == DMU_BONUS_BLKID) {
2370 int slots = DB_DNODE(db)->dn_num_slots;
2371 int bonuslen = DN_SLOTS_TO_BONUSLEN(slots);
2372 if (db->db.db_data != NULL) {
2373 kmem_free(db->db.db_data, bonuslen);
2374 arc_space_return(bonuslen, ARC_SPACE_BONUS);
2375 db->db_state = DB_UNCACHED;
2379 dbuf_clear_data(db);
2381 if (multilist_link_active(&db->db_cache_link)) {
2382 multilist_remove(dbuf_cache, db);
2383 (void) refcount_remove_many(&dbuf_cache_size,
2384 db->db.db_size, db);
2385 DBUF_STAT_BUMPDOWN(cache_levels[db->db_level]);
2386 DBUF_STAT_BUMPDOWN(cache_count);
2387 DBUF_STAT_DECR(cache_levels_bytes[db->db_level],
2391 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
2392 ASSERT(db->db_data_pending == NULL);
2394 db->db_state = DB_EVICTING;
2395 db->db_blkptr = NULL;
2398 * Now that db_state is DB_EVICTING, nobody else can find this via
2399 * the hash table. We can now drop db_mtx, which allows us to
2400 * acquire the dn_dbufs_mtx.
2402 mutex_exit(&db->db_mtx);
2407 if (db->db_blkid != DMU_BONUS_BLKID) {
2408 boolean_t needlock = !MUTEX_HELD(&dn->dn_dbufs_mtx);
2410 mutex_enter(&dn->dn_dbufs_mtx);
2411 avl_remove(&dn->dn_dbufs, db);
2412 atomic_dec_32(&dn->dn_dbufs_count);
2416 mutex_exit(&dn->dn_dbufs_mtx);
2418 * Decrementing the dbuf count means that the hold corresponding
2419 * to the removed dbuf is no longer discounted in dnode_move(),
2420 * so the dnode cannot be moved until after we release the hold.
2421 * The membar_producer() ensures visibility of the decremented
2422 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
2426 db->db_dnode_handle = NULL;
2428 dbuf_hash_remove(db);
2433 ASSERT(refcount_is_zero(&db->db_holds));
2435 db->db_parent = NULL;
2437 ASSERT(db->db_buf == NULL);
2438 ASSERT(db->db.db_data == NULL);
2439 ASSERT(db->db_hash_next == NULL);
2440 ASSERT(db->db_blkptr == NULL);
2441 ASSERT(db->db_data_pending == NULL);
2442 ASSERT(!multilist_link_active(&db->db_cache_link));
2444 kmem_cache_free(dbuf_kmem_cache, db);
2445 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
2448 * If this dbuf is referenced from an indirect dbuf,
2449 * decrement the ref count on the indirect dbuf.
2451 if (parent && parent != dndb)
2452 dbuf_rele(parent, db);
2456 * Note: While bpp will always be updated if the function returns success,
2457 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
2458 * this happens when the dnode is the meta-dnode, or a userused or groupused
2461 __attribute__((always_inline))
2463 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
2464 dmu_buf_impl_t **parentp, blkptr_t **bpp, struct dbuf_hold_impl_data *dh)
2469 ASSERT(blkid != DMU_BONUS_BLKID);
2471 if (blkid == DMU_SPILL_BLKID) {
2472 mutex_enter(&dn->dn_mtx);
2473 if (dn->dn_have_spill &&
2474 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
2475 *bpp = DN_SPILL_BLKPTR(dn->dn_phys);
2478 dbuf_add_ref(dn->dn_dbuf, NULL);
2479 *parentp = dn->dn_dbuf;
2480 mutex_exit(&dn->dn_mtx);
2485 (dn->dn_phys->dn_nlevels == 0) ? 1 : dn->dn_phys->dn_nlevels;
2486 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2488 ASSERT3U(level * epbs, <, 64);
2489 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2491 * This assertion shouldn't trip as long as the max indirect block size
2492 * is less than 1M. The reason for this is that up to that point,
2493 * the number of levels required to address an entire object with blocks
2494 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In
2495 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55
2496 * (i.e. we can address the entire object), objects will all use at most
2497 * N-1 levels and the assertion won't overflow. However, once epbs is
2498 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be
2499 * enough to address an entire object, so objects will have 5 levels,
2500 * but then this assertion will overflow.
2502 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we
2503 * need to redo this logic to handle overflows.
2505 ASSERT(level >= nlevels ||
2506 ((nlevels - level - 1) * epbs) +
2507 highbit64(dn->dn_phys->dn_nblkptr) <= 64);
2508 if (level >= nlevels ||
2509 blkid >= ((uint64_t)dn->dn_phys->dn_nblkptr <<
2510 ((nlevels - level - 1) * epbs)) ||
2512 blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
2513 /* the buffer has no parent yet */
2514 return (SET_ERROR(ENOENT));
2515 } else if (level < nlevels-1) {
2516 /* this block is referenced from an indirect block */
2519 err = dbuf_hold_impl(dn, level+1,
2520 blkid >> epbs, fail_sparse, FALSE, NULL, parentp);
2522 __dbuf_hold_impl_init(dh + 1, dn, dh->dh_level + 1,
2523 blkid >> epbs, fail_sparse, FALSE, NULL,
2524 parentp, dh->dh_depth + 1);
2525 err = __dbuf_hold_impl(dh + 1);
2529 err = dbuf_read(*parentp, NULL,
2530 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
2532 dbuf_rele(*parentp, NULL);
2536 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
2537 (blkid & ((1ULL << epbs) - 1));
2538 if (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))
2539 ASSERT(BP_IS_HOLE(*bpp));
2542 /* the block is referenced from the dnode */
2543 ASSERT3U(level, ==, nlevels-1);
2544 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
2545 blkid < dn->dn_phys->dn_nblkptr);
2547 dbuf_add_ref(dn->dn_dbuf, NULL);
2548 *parentp = dn->dn_dbuf;
2550 *bpp = &dn->dn_phys->dn_blkptr[blkid];
2555 static dmu_buf_impl_t *
2556 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
2557 dmu_buf_impl_t *parent, blkptr_t *blkptr)
2559 objset_t *os = dn->dn_objset;
2560 dmu_buf_impl_t *db, *odb;
2562 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2563 ASSERT(dn->dn_type != DMU_OT_NONE);
2565 db = kmem_cache_alloc(dbuf_kmem_cache, KM_SLEEP);
2568 db->db.db_object = dn->dn_object;
2569 db->db_level = level;
2570 db->db_blkid = blkid;
2571 db->db_last_dirty = NULL;
2572 db->db_dirtycnt = 0;
2573 db->db_dnode_handle = dn->dn_handle;
2574 db->db_parent = parent;
2575 db->db_blkptr = blkptr;
2578 db->db_user_immediate_evict = FALSE;
2579 db->db_freed_in_flight = FALSE;
2580 db->db_pending_evict = FALSE;
2582 if (blkid == DMU_BONUS_BLKID) {
2583 ASSERT3P(parent, ==, dn->dn_dbuf);
2584 db->db.db_size = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
2585 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
2586 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
2587 db->db.db_offset = DMU_BONUS_BLKID;
2588 db->db_state = DB_UNCACHED;
2589 /* the bonus dbuf is not placed in the hash table */
2590 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
2592 } else if (blkid == DMU_SPILL_BLKID) {
2593 db->db.db_size = (blkptr != NULL) ?
2594 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
2595 db->db.db_offset = 0;
2598 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
2599 db->db.db_size = blocksize;
2600 db->db.db_offset = db->db_blkid * blocksize;
2604 * Hold the dn_dbufs_mtx while we get the new dbuf
2605 * in the hash table *and* added to the dbufs list.
2606 * This prevents a possible deadlock with someone
2607 * trying to look up this dbuf before its added to the
2610 mutex_enter(&dn->dn_dbufs_mtx);
2611 db->db_state = DB_EVICTING;
2612 if ((odb = dbuf_hash_insert(db)) != NULL) {
2613 /* someone else inserted it first */
2614 kmem_cache_free(dbuf_kmem_cache, db);
2615 mutex_exit(&dn->dn_dbufs_mtx);
2616 DBUF_STAT_BUMP(hash_insert_race);
2619 avl_add(&dn->dn_dbufs, db);
2621 db->db_state = DB_UNCACHED;
2622 mutex_exit(&dn->dn_dbufs_mtx);
2623 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
2625 if (parent && parent != dn->dn_dbuf)
2626 dbuf_add_ref(parent, db);
2628 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
2629 refcount_count(&dn->dn_holds) > 0);
2630 (void) refcount_add(&dn->dn_holds, db);
2631 atomic_inc_32(&dn->dn_dbufs_count);
2633 dprintf_dbuf(db, "db=%p\n", db);
2638 typedef struct dbuf_prefetch_arg {
2639 spa_t *dpa_spa; /* The spa to issue the prefetch in. */
2640 zbookmark_phys_t dpa_zb; /* The target block to prefetch. */
2641 int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */
2642 int dpa_curlevel; /* The current level that we're reading */
2643 dnode_t *dpa_dnode; /* The dnode associated with the prefetch */
2644 zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */
2645 zio_t *dpa_zio; /* The parent zio_t for all prefetches. */
2646 arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */
2647 } dbuf_prefetch_arg_t;
2650 * Actually issue the prefetch read for the block given.
2653 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp)
2655 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
2658 arc_flags_t aflags =
2659 dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
2661 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2662 ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level);
2663 ASSERT(dpa->dpa_zio != NULL);
2664 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL,
2665 dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2666 &aflags, &dpa->dpa_zb);
2670 * Called when an indirect block above our prefetch target is read in. This
2671 * will either read in the next indirect block down the tree or issue the actual
2672 * prefetch if the next block down is our target.
2675 dbuf_prefetch_indirect_done(zio_t *zio, const zbookmark_phys_t *zb,
2676 const blkptr_t *iobp, arc_buf_t *abuf, void *private)
2678 dbuf_prefetch_arg_t *dpa = private;
2680 ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel);
2681 ASSERT3S(dpa->dpa_curlevel, >, 0);
2684 * The dpa_dnode is only valid if we are called with a NULL
2685 * zio. This indicates that the arc_read() returned without
2686 * first calling zio_read() to issue a physical read. Once
2687 * a physical read is made the dpa_dnode must be invalidated
2688 * as the locks guarding it may have been dropped. If the
2689 * dpa_dnode is still valid, then we want to add it to the dbuf
2690 * cache. To do so, we must hold the dbuf associated with the block
2691 * we just prefetched, read its contents so that we associate it
2692 * with an arc_buf_t, and then release it.
2695 ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel);
2696 if (zio->io_flags & ZIO_FLAG_RAW_COMPRESS) {
2697 ASSERT3U(BP_GET_PSIZE(zio->io_bp), ==, zio->io_size);
2699 ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size);
2701 ASSERT3P(zio->io_spa, ==, dpa->dpa_spa);
2703 dpa->dpa_dnode = NULL;
2704 } else if (dpa->dpa_dnode != NULL) {
2705 uint64_t curblkid = dpa->dpa_zb.zb_blkid >>
2706 (dpa->dpa_epbs * (dpa->dpa_curlevel -
2707 dpa->dpa_zb.zb_level));
2708 dmu_buf_impl_t *db = dbuf_hold_level(dpa->dpa_dnode,
2709 dpa->dpa_curlevel, curblkid, FTAG);
2710 (void) dbuf_read(db, NULL,
2711 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_HAVESTRUCT);
2712 dbuf_rele(db, FTAG);
2716 kmem_free(dpa, sizeof (*dpa));
2720 dpa->dpa_curlevel--;
2721 uint64_t nextblkid = dpa->dpa_zb.zb_blkid >>
2722 (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level));
2723 blkptr_t *bp = ((blkptr_t *)abuf->b_data) +
2724 P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs);
2726 if (BP_IS_HOLE(bp)) {
2727 kmem_free(dpa, sizeof (*dpa));
2728 } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) {
2729 ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid);
2730 dbuf_issue_final_prefetch(dpa, bp);
2731 kmem_free(dpa, sizeof (*dpa));
2733 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2734 zbookmark_phys_t zb;
2736 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2737 if (dpa->dpa_aflags & ARC_FLAG_L2CACHE)
2738 iter_aflags |= ARC_FLAG_L2CACHE;
2740 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2742 SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset,
2743 dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid);
2745 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2746 bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio,
2747 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2751 arc_buf_destroy(abuf, private);
2755 * Issue prefetch reads for the given block on the given level. If the indirect
2756 * blocks above that block are not in memory, we will read them in
2757 * asynchronously. As a result, this call never blocks waiting for a read to
2758 * complete. Note that the prefetch might fail if the dataset is encrypted and
2759 * the encryption key is unmapped before the IO completes.
2762 dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio,
2766 int epbs, nlevels, curlevel;
2769 ASSERT(blkid != DMU_BONUS_BLKID);
2770 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2772 if (blkid > dn->dn_maxblkid)
2775 if (dnode_block_freed(dn, blkid))
2779 * This dnode hasn't been written to disk yet, so there's nothing to
2782 nlevels = dn->dn_phys->dn_nlevels;
2783 if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0)
2786 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2787 if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level))
2790 dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object,
2793 mutex_exit(&db->db_mtx);
2795 * This dbuf already exists. It is either CACHED, or
2796 * (we assume) about to be read or filled.
2802 * Find the closest ancestor (indirect block) of the target block
2803 * that is present in the cache. In this indirect block, we will
2804 * find the bp that is at curlevel, curblkid.
2808 while (curlevel < nlevels - 1) {
2809 int parent_level = curlevel + 1;
2810 uint64_t parent_blkid = curblkid >> epbs;
2813 if (dbuf_hold_impl(dn, parent_level, parent_blkid,
2814 FALSE, TRUE, FTAG, &db) == 0) {
2815 blkptr_t *bpp = db->db_buf->b_data;
2816 bp = bpp[P2PHASE(curblkid, 1 << epbs)];
2817 dbuf_rele(db, FTAG);
2821 curlevel = parent_level;
2822 curblkid = parent_blkid;
2825 if (curlevel == nlevels - 1) {
2826 /* No cached indirect blocks found. */
2827 ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr);
2828 bp = dn->dn_phys->dn_blkptr[curblkid];
2830 if (BP_IS_HOLE(&bp))
2833 ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp));
2835 zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL,
2838 dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP);
2839 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
2840 SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2841 dn->dn_object, level, blkid);
2842 dpa->dpa_curlevel = curlevel;
2843 dpa->dpa_prio = prio;
2844 dpa->dpa_aflags = aflags;
2845 dpa->dpa_spa = dn->dn_objset->os_spa;
2846 dpa->dpa_dnode = dn;
2847 dpa->dpa_epbs = epbs;
2850 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2851 if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level))
2852 dpa->dpa_aflags |= ARC_FLAG_L2CACHE;
2855 * If we have the indirect just above us, no need to do the asynchronous
2856 * prefetch chain; we'll just run the last step ourselves. If we're at
2857 * a higher level, though, we want to issue the prefetches for all the
2858 * indirect blocks asynchronously, so we can go on with whatever we were
2861 if (curlevel == level) {
2862 ASSERT3U(curblkid, ==, blkid);
2863 dbuf_issue_final_prefetch(dpa, &bp);
2864 kmem_free(dpa, sizeof (*dpa));
2866 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2867 zbookmark_phys_t zb;
2869 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2870 if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level))
2871 iter_aflags |= ARC_FLAG_L2CACHE;
2873 SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2874 dn->dn_object, curlevel, curblkid);
2875 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2876 &bp, dbuf_prefetch_indirect_done, dpa, prio,
2877 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2881 * We use pio here instead of dpa_zio since it's possible that
2882 * dpa may have already been freed.
2887 #define DBUF_HOLD_IMPL_MAX_DEPTH 20
2890 * Helper function for __dbuf_hold_impl() to copy a buffer. Handles
2891 * the case of encrypted, compressed and uncompressed buffers by
2892 * allocating the new buffer, respectively, with arc_alloc_raw_buf(),
2893 * arc_alloc_compressed_buf() or arc_alloc_buf().*
2895 * NOTE: Declared noinline to avoid stack bloat in __dbuf_hold_impl().
2897 noinline static void
2898 dbuf_hold_copy(struct dbuf_hold_impl_data *dh)
2900 dnode_t *dn = dh->dh_dn;
2901 dmu_buf_impl_t *db = dh->dh_db;
2902 dbuf_dirty_record_t *dr = dh->dh_dr;
2903 arc_buf_t *data = dr->dt.dl.dr_data;
2905 enum zio_compress compress_type = arc_get_compression(data);
2907 if (arc_is_encrypted(data)) {
2908 boolean_t byteorder;
2909 uint8_t salt[ZIO_DATA_SALT_LEN];
2910 uint8_t iv[ZIO_DATA_IV_LEN];
2911 uint8_t mac[ZIO_DATA_MAC_LEN];
2913 arc_get_raw_params(data, &byteorder, salt, iv, mac);
2914 dbuf_set_data(db, arc_alloc_raw_buf(dn->dn_objset->os_spa, db,
2915 dmu_objset_id(dn->dn_objset), byteorder, salt, iv, mac,
2916 dn->dn_type, arc_buf_size(data), arc_buf_lsize(data),
2918 } else if (compress_type != ZIO_COMPRESS_OFF) {
2919 dbuf_set_data(db, arc_alloc_compressed_buf(
2920 dn->dn_objset->os_spa, db, arc_buf_size(data),
2921 arc_buf_lsize(data), compress_type));
2923 dbuf_set_data(db, arc_alloc_buf(dn->dn_objset->os_spa, db,
2924 DBUF_GET_BUFC_TYPE(db), db->db.db_size));
2927 bcopy(data->b_data, db->db.db_data, arc_buf_size(data));
2931 * Returns with db_holds incremented, and db_mtx not held.
2932 * Note: dn_struct_rwlock must be held.
2935 __dbuf_hold_impl(struct dbuf_hold_impl_data *dh)
2937 ASSERT3S(dh->dh_depth, <, DBUF_HOLD_IMPL_MAX_DEPTH);
2938 dh->dh_parent = NULL;
2940 ASSERT(dh->dh_blkid != DMU_BONUS_BLKID);
2941 ASSERT(RW_LOCK_HELD(&dh->dh_dn->dn_struct_rwlock));
2942 ASSERT3U(dh->dh_dn->dn_nlevels, >, dh->dh_level);
2944 *(dh->dh_dbp) = NULL;
2946 /* dbuf_find() returns with db_mtx held */
2947 dh->dh_db = dbuf_find(dh->dh_dn->dn_objset, dh->dh_dn->dn_object,
2948 dh->dh_level, dh->dh_blkid);
2950 if (dh->dh_db == NULL) {
2953 if (dh->dh_fail_uncached)
2954 return (SET_ERROR(ENOENT));
2956 ASSERT3P(dh->dh_parent, ==, NULL);
2957 dh->dh_err = dbuf_findbp(dh->dh_dn, dh->dh_level, dh->dh_blkid,
2958 dh->dh_fail_sparse, &dh->dh_parent, &dh->dh_bp, dh);
2959 if (dh->dh_fail_sparse) {
2960 if (dh->dh_err == 0 &&
2961 dh->dh_bp && BP_IS_HOLE(dh->dh_bp))
2962 dh->dh_err = SET_ERROR(ENOENT);
2965 dbuf_rele(dh->dh_parent, NULL);
2966 return (dh->dh_err);
2969 if (dh->dh_err && dh->dh_err != ENOENT)
2970 return (dh->dh_err);
2971 dh->dh_db = dbuf_create(dh->dh_dn, dh->dh_level, dh->dh_blkid,
2972 dh->dh_parent, dh->dh_bp);
2975 if (dh->dh_fail_uncached && dh->dh_db->db_state != DB_CACHED) {
2976 mutex_exit(&dh->dh_db->db_mtx);
2977 return (SET_ERROR(ENOENT));
2980 if (dh->dh_db->db_buf != NULL) {
2981 arc_buf_access(dh->dh_db->db_buf);
2982 ASSERT3P(dh->dh_db->db.db_data, ==, dh->dh_db->db_buf->b_data);
2985 ASSERT(dh->dh_db->db_buf == NULL || arc_referenced(dh->dh_db->db_buf));
2988 * If this buffer is currently syncing out, and we are are
2989 * still referencing it from db_data, we need to make a copy
2990 * of it in case we decide we want to dirty it again in this txg.
2992 if (dh->dh_db->db_level == 0 &&
2993 dh->dh_db->db_blkid != DMU_BONUS_BLKID &&
2994 dh->dh_dn->dn_object != DMU_META_DNODE_OBJECT &&
2995 dh->dh_db->db_state == DB_CACHED && dh->dh_db->db_data_pending) {
2996 dh->dh_dr = dh->dh_db->db_data_pending;
2997 if (dh->dh_dr->dt.dl.dr_data == dh->dh_db->db_buf)
3001 if (multilist_link_active(&dh->dh_db->db_cache_link)) {
3002 ASSERT(refcount_is_zero(&dh->dh_db->db_holds));
3003 multilist_remove(dbuf_cache, dh->dh_db);
3004 (void) refcount_remove_many(&dbuf_cache_size,
3005 dh->dh_db->db.db_size, dh->dh_db);
3006 DBUF_STAT_BUMPDOWN(cache_levels[dh->dh_db->db_level]);
3007 DBUF_STAT_BUMPDOWN(cache_count);
3008 DBUF_STAT_DECR(cache_levels_bytes[dh->dh_db->db_level],
3009 dh->dh_db->db.db_size);
3011 (void) refcount_add(&dh->dh_db->db_holds, dh->dh_tag);
3012 DBUF_VERIFY(dh->dh_db);
3013 mutex_exit(&dh->dh_db->db_mtx);
3015 /* NOTE: we can't rele the parent until after we drop the db_mtx */
3017 dbuf_rele(dh->dh_parent, NULL);
3019 ASSERT3P(DB_DNODE(dh->dh_db), ==, dh->dh_dn);
3020 ASSERT3U(dh->dh_db->db_blkid, ==, dh->dh_blkid);
3021 ASSERT3U(dh->dh_db->db_level, ==, dh->dh_level);
3022 *(dh->dh_dbp) = dh->dh_db;
3028 * The following code preserves the recursive function dbuf_hold_impl()
3029 * but moves the local variables AND function arguments to the heap to
3030 * minimize the stack frame size. Enough space is initially allocated
3031 * on the stack for 20 levels of recursion.
3034 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid,
3035 boolean_t fail_sparse, boolean_t fail_uncached,
3036 void *tag, dmu_buf_impl_t **dbp)
3038 struct dbuf_hold_impl_data *dh;
3041 dh = kmem_alloc(sizeof (struct dbuf_hold_impl_data) *
3042 DBUF_HOLD_IMPL_MAX_DEPTH, KM_SLEEP);
3043 __dbuf_hold_impl_init(dh, dn, level, blkid, fail_sparse,
3044 fail_uncached, tag, dbp, 0);
3046 error = __dbuf_hold_impl(dh);
3048 kmem_free(dh, sizeof (struct dbuf_hold_impl_data) *
3049 DBUF_HOLD_IMPL_MAX_DEPTH);
3055 __dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh,
3056 dnode_t *dn, uint8_t level, uint64_t blkid,
3057 boolean_t fail_sparse, boolean_t fail_uncached,
3058 void *tag, dmu_buf_impl_t **dbp, int depth)
3061 dh->dh_level = level;
3062 dh->dh_blkid = blkid;
3064 dh->dh_fail_sparse = fail_sparse;
3065 dh->dh_fail_uncached = fail_uncached;
3071 dh->dh_parent = NULL;
3076 dh->dh_depth = depth;
3080 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
3082 return (dbuf_hold_level(dn, 0, blkid, tag));
3086 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
3089 int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db);
3090 return (err ? NULL : db);
3094 dbuf_create_bonus(dnode_t *dn)
3096 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
3098 ASSERT(dn->dn_bonus == NULL);
3099 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
3103 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
3105 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3108 if (db->db_blkid != DMU_SPILL_BLKID)
3109 return (SET_ERROR(ENOTSUP));
3111 blksz = SPA_MINBLOCKSIZE;
3112 ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset)));
3113 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
3117 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3118 dbuf_new_size(db, blksz, tx);
3119 rw_exit(&dn->dn_struct_rwlock);
3126 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
3128 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
3131 #pragma weak dmu_buf_add_ref = dbuf_add_ref
3133 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
3135 int64_t holds = refcount_add(&db->db_holds, tag);
3136 VERIFY3S(holds, >, 1);
3139 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
3141 dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid,
3144 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3145 dmu_buf_impl_t *found_db;
3146 boolean_t result = B_FALSE;
3148 if (blkid == DMU_BONUS_BLKID)
3149 found_db = dbuf_find_bonus(os, obj);
3151 found_db = dbuf_find(os, obj, 0, blkid);
3153 if (found_db != NULL) {
3154 if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) {
3155 (void) refcount_add(&db->db_holds, tag);
3158 mutex_exit(&found_db->db_mtx);
3164 * If you call dbuf_rele() you had better not be referencing the dnode handle
3165 * unless you have some other direct or indirect hold on the dnode. (An indirect
3166 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
3167 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
3168 * dnode's parent dbuf evicting its dnode handles.
3171 dbuf_rele(dmu_buf_impl_t *db, void *tag)
3173 mutex_enter(&db->db_mtx);
3174 dbuf_rele_and_unlock(db, tag);
3178 dmu_buf_rele(dmu_buf_t *db, void *tag)
3180 dbuf_rele((dmu_buf_impl_t *)db, tag);
3184 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
3185 * db_dirtycnt and db_holds to be updated atomically.
3188 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
3192 ASSERT(MUTEX_HELD(&db->db_mtx));
3196 * Remove the reference to the dbuf before removing its hold on the
3197 * dnode so we can guarantee in dnode_move() that a referenced bonus
3198 * buffer has a corresponding dnode hold.
3200 holds = refcount_remove(&db->db_holds, tag);
3204 * We can't freeze indirects if there is a possibility that they
3205 * may be modified in the current syncing context.
3207 if (db->db_buf != NULL &&
3208 holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) {
3209 arc_buf_freeze(db->db_buf);
3212 if (holds == db->db_dirtycnt &&
3213 db->db_level == 0 && db->db_user_immediate_evict)
3214 dbuf_evict_user(db);
3217 if (db->db_blkid == DMU_BONUS_BLKID) {
3219 boolean_t evict_dbuf = db->db_pending_evict;
3222 * If the dnode moves here, we cannot cross this
3223 * barrier until the move completes.
3228 atomic_dec_32(&dn->dn_dbufs_count);
3231 * Decrementing the dbuf count means that the bonus
3232 * buffer's dnode hold is no longer discounted in
3233 * dnode_move(). The dnode cannot move until after
3234 * the dnode_rele() below.
3239 * Do not reference db after its lock is dropped.
3240 * Another thread may evict it.
3242 mutex_exit(&db->db_mtx);
3245 dnode_evict_bonus(dn);
3248 } else if (db->db_buf == NULL) {
3250 * This is a special case: we never associated this
3251 * dbuf with any data allocated from the ARC.
3253 ASSERT(db->db_state == DB_UNCACHED ||
3254 db->db_state == DB_NOFILL);
3256 } else if (arc_released(db->db_buf)) {
3258 * This dbuf has anonymous data associated with it.
3262 boolean_t do_arc_evict = B_FALSE;
3264 spa_t *spa = dmu_objset_spa(db->db_objset);
3266 if (!DBUF_IS_CACHEABLE(db) &&
3267 db->db_blkptr != NULL &&
3268 !BP_IS_HOLE(db->db_blkptr) &&
3269 !BP_IS_EMBEDDED(db->db_blkptr)) {
3270 do_arc_evict = B_TRUE;
3271 bp = *db->db_blkptr;
3274 if (!DBUF_IS_CACHEABLE(db) ||
3275 db->db_pending_evict) {
3277 } else if (!multilist_link_active(&db->db_cache_link)) {
3278 multilist_insert(dbuf_cache, db);
3279 (void) refcount_add_many(&dbuf_cache_size,
3280 db->db.db_size, db);
3281 DBUF_STAT_BUMP(cache_levels[db->db_level]);
3282 DBUF_STAT_BUMP(cache_count);
3283 DBUF_STAT_INCR(cache_levels_bytes[db->db_level],
3285 DBUF_STAT_MAX(cache_size_bytes_max,
3286 refcount_count(&dbuf_cache_size));
3287 mutex_exit(&db->db_mtx);
3289 dbuf_evict_notify();
3293 arc_freed(spa, &bp);
3296 mutex_exit(&db->db_mtx);
3301 #pragma weak dmu_buf_refcount = dbuf_refcount
3303 dbuf_refcount(dmu_buf_impl_t *db)
3305 return (refcount_count(&db->db_holds));
3309 dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user,
3310 dmu_buf_user_t *new_user)
3312 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3314 mutex_enter(&db->db_mtx);
3315 dbuf_verify_user(db, DBVU_NOT_EVICTING);
3316 if (db->db_user == old_user)
3317 db->db_user = new_user;
3319 old_user = db->db_user;
3320 dbuf_verify_user(db, DBVU_NOT_EVICTING);
3321 mutex_exit(&db->db_mtx);
3327 dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
3329 return (dmu_buf_replace_user(db_fake, NULL, user));
3333 dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user)
3335 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3337 db->db_user_immediate_evict = TRUE;
3338 return (dmu_buf_set_user(db_fake, user));
3342 dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
3344 return (dmu_buf_replace_user(db_fake, user, NULL));
3348 dmu_buf_get_user(dmu_buf_t *db_fake)
3350 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3352 dbuf_verify_user(db, DBVU_NOT_EVICTING);
3353 return (db->db_user);
3357 dmu_buf_user_evict_wait()
3359 taskq_wait(dbu_evict_taskq);
3363 dmu_buf_get_blkptr(dmu_buf_t *db)
3365 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3366 return (dbi->db_blkptr);
3370 dmu_buf_get_objset(dmu_buf_t *db)
3372 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3373 return (dbi->db_objset);
3377 dmu_buf_dnode_enter(dmu_buf_t *db)
3379 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3380 DB_DNODE_ENTER(dbi);
3381 return (DB_DNODE(dbi));
3385 dmu_buf_dnode_exit(dmu_buf_t *db)
3387 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3392 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
3394 /* ASSERT(dmu_tx_is_syncing(tx) */
3395 ASSERT(MUTEX_HELD(&db->db_mtx));
3397 if (db->db_blkptr != NULL)
3400 if (db->db_blkid == DMU_SPILL_BLKID) {
3401 db->db_blkptr = DN_SPILL_BLKPTR(dn->dn_phys);
3402 BP_ZERO(db->db_blkptr);
3405 if (db->db_level == dn->dn_phys->dn_nlevels-1) {
3407 * This buffer was allocated at a time when there was
3408 * no available blkptrs from the dnode, or it was
3409 * inappropriate to hook it in (i.e., nlevels mis-match).
3411 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
3412 ASSERT(db->db_parent == NULL);
3413 db->db_parent = dn->dn_dbuf;
3414 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
3417 dmu_buf_impl_t *parent = db->db_parent;
3418 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3420 ASSERT(dn->dn_phys->dn_nlevels > 1);
3421 if (parent == NULL) {
3422 mutex_exit(&db->db_mtx);
3423 rw_enter(&dn->dn_struct_rwlock, RW_READER);
3424 parent = dbuf_hold_level(dn, db->db_level + 1,
3425 db->db_blkid >> epbs, db);
3426 rw_exit(&dn->dn_struct_rwlock);
3427 mutex_enter(&db->db_mtx);
3428 db->db_parent = parent;
3430 db->db_blkptr = (blkptr_t *)parent->db.db_data +
3431 (db->db_blkid & ((1ULL << epbs) - 1));
3437 * Ensure the dbuf's data is untransformed if the associated dirty
3438 * record requires it. This is used by dbuf_sync_leaf() to ensure
3439 * that a dnode block is decrypted before we write new data to it.
3440 * For raw writes we assert that the buffer is already encrypted.
3443 dbuf_check_crypt(dbuf_dirty_record_t *dr)
3446 dmu_buf_impl_t *db = dr->dr_dbuf;
3448 ASSERT(MUTEX_HELD(&db->db_mtx));
3450 if (!dr->dt.dl.dr_raw && arc_is_encrypted(db->db_buf)) {
3452 * Unfortunately, there is currently no mechanism for
3453 * syncing context to handle decryption errors. An error
3454 * here is only possible if an attacker maliciously
3455 * changed a dnode block and updated the associated
3456 * checksums going up the block tree.
3458 err = arc_untransform(db->db_buf, db->db_objset->os_spa,
3459 dmu_objset_id(db->db_objset), B_TRUE);
3461 panic("Invalid dnode block MAC");
3462 } else if (dr->dt.dl.dr_raw) {
3464 * Writing raw encrypted data requires the db's arc buffer
3465 * to be converted to raw by the caller.
3467 ASSERT(arc_is_encrypted(db->db_buf));
3472 * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it
3473 * is critical the we not allow the compiler to inline this function in to
3474 * dbuf_sync_list() thereby drastically bloating the stack usage.
3476 noinline static void
3477 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
3479 dmu_buf_impl_t *db = dr->dr_dbuf;
3483 ASSERT(dmu_tx_is_syncing(tx));
3485 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
3487 mutex_enter(&db->db_mtx);
3489 ASSERT(db->db_level > 0);
3492 /* Read the block if it hasn't been read yet. */
3493 if (db->db_buf == NULL) {
3494 mutex_exit(&db->db_mtx);
3495 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
3496 mutex_enter(&db->db_mtx);
3498 ASSERT3U(db->db_state, ==, DB_CACHED);
3499 ASSERT(db->db_buf != NULL);
3503 /* Indirect block size must match what the dnode thinks it is. */
3504 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3505 dbuf_check_blkptr(dn, db);
3508 /* Provide the pending dirty record to child dbufs */
3509 db->db_data_pending = dr;
3511 mutex_exit(&db->db_mtx);
3512 dbuf_write(dr, db->db_buf, tx);
3515 mutex_enter(&dr->dt.di.dr_mtx);
3516 dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx);
3517 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3518 mutex_exit(&dr->dt.di.dr_mtx);
3523 * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is
3524 * critical the we not allow the compiler to inline this function in to
3525 * dbuf_sync_list() thereby drastically bloating the stack usage.
3527 noinline static void
3528 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
3530 arc_buf_t **datap = &dr->dt.dl.dr_data;
3531 dmu_buf_impl_t *db = dr->dr_dbuf;
3534 uint64_t txg = tx->tx_txg;
3536 ASSERT(dmu_tx_is_syncing(tx));
3538 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
3540 mutex_enter(&db->db_mtx);
3542 * To be synced, we must be dirtied. But we
3543 * might have been freed after the dirty.
3545 if (db->db_state == DB_UNCACHED) {
3546 /* This buffer has been freed since it was dirtied */
3547 ASSERT(db->db.db_data == NULL);
3548 } else if (db->db_state == DB_FILL) {
3549 /* This buffer was freed and is now being re-filled */
3550 ASSERT(db->db.db_data != dr->dt.dl.dr_data);
3552 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
3559 if (db->db_blkid == DMU_SPILL_BLKID) {
3560 mutex_enter(&dn->dn_mtx);
3561 if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
3563 * In the previous transaction group, the bonus buffer
3564 * was entirely used to store the attributes for the
3565 * dnode which overrode the dn_spill field. However,
3566 * when adding more attributes to the file a spill
3567 * block was required to hold the extra attributes.
3569 * Make sure to clear the garbage left in the dn_spill
3570 * field from the previous attributes in the bonus
3571 * buffer. Otherwise, after writing out the spill
3572 * block to the new allocated dva, it will free
3573 * the old block pointed to by the invalid dn_spill.
3575 db->db_blkptr = NULL;
3577 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
3578 mutex_exit(&dn->dn_mtx);
3582 * If this is a bonus buffer, simply copy the bonus data into the
3583 * dnode. It will be written out when the dnode is synced (and it
3584 * will be synced, since it must have been dirty for dbuf_sync to
3587 if (db->db_blkid == DMU_BONUS_BLKID) {
3588 dbuf_dirty_record_t **drp;
3590 ASSERT(*datap != NULL);
3591 ASSERT0(db->db_level);
3592 ASSERT3U(DN_MAX_BONUS_LEN(dn->dn_phys), <=,
3593 DN_SLOTS_TO_BONUSLEN(dn->dn_phys->dn_extra_slots + 1));
3594 bcopy(*datap, DN_BONUS(dn->dn_phys),
3595 DN_MAX_BONUS_LEN(dn->dn_phys));
3598 if (*datap != db->db.db_data) {
3599 int slots = DB_DNODE(db)->dn_num_slots;
3600 int bonuslen = DN_SLOTS_TO_BONUSLEN(slots);
3601 kmem_free(*datap, bonuslen);
3602 arc_space_return(bonuslen, ARC_SPACE_BONUS);
3604 db->db_data_pending = NULL;
3605 drp = &db->db_last_dirty;
3607 drp = &(*drp)->dr_next;
3608 ASSERT(dr->dr_next == NULL);
3609 ASSERT(dr->dr_dbuf == db);
3611 if (dr->dr_dbuf->db_level != 0) {
3612 mutex_destroy(&dr->dt.di.dr_mtx);
3613 list_destroy(&dr->dt.di.dr_children);
3615 kmem_free(dr, sizeof (dbuf_dirty_record_t));
3616 ASSERT(db->db_dirtycnt > 0);
3617 db->db_dirtycnt -= 1;
3618 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
3625 * This function may have dropped the db_mtx lock allowing a dmu_sync
3626 * operation to sneak in. As a result, we need to ensure that we
3627 * don't check the dr_override_state until we have returned from
3628 * dbuf_check_blkptr.
3630 dbuf_check_blkptr(dn, db);
3633 * If this buffer is in the middle of an immediate write,
3634 * wait for the synchronous IO to complete.
3636 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
3637 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
3638 cv_wait(&db->db_changed, &db->db_mtx);
3639 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
3643 * If this is a dnode block, ensure it is appropriately encrypted
3644 * or decrypted, depending on what we are writing to it this txg.
3646 if (os->os_encrypted && dn->dn_object == DMU_META_DNODE_OBJECT)
3647 dbuf_check_crypt(dr);
3649 if (db->db_state != DB_NOFILL &&
3650 dn->dn_object != DMU_META_DNODE_OBJECT &&
3651 refcount_count(&db->db_holds) > 1 &&
3652 dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
3653 *datap == db->db_buf) {
3655 * If this buffer is currently "in use" (i.e., there
3656 * are active holds and db_data still references it),
3657 * then make a copy before we start the write so that
3658 * any modifications from the open txg will not leak
3661 * NOTE: this copy does not need to be made for
3662 * objects only modified in the syncing context (e.g.
3663 * DNONE_DNODE blocks).
3665 int psize = arc_buf_size(*datap);
3666 int lsize = arc_buf_lsize(*datap);
3667 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
3668 enum zio_compress compress_type = arc_get_compression(*datap);
3670 if (arc_is_encrypted(*datap)) {
3671 boolean_t byteorder;
3672 uint8_t salt[ZIO_DATA_SALT_LEN];
3673 uint8_t iv[ZIO_DATA_IV_LEN];
3674 uint8_t mac[ZIO_DATA_MAC_LEN];
3676 arc_get_raw_params(*datap, &byteorder, salt, iv, mac);
3677 *datap = arc_alloc_raw_buf(os->os_spa, db,
3678 dmu_objset_id(os), byteorder, salt, iv, mac,
3679 dn->dn_type, psize, lsize, compress_type);
3680 } else if (compress_type != ZIO_COMPRESS_OFF) {
3681 ASSERT3U(type, ==, ARC_BUFC_DATA);
3682 *datap = arc_alloc_compressed_buf(os->os_spa, db,
3683 psize, lsize, compress_type);
3685 *datap = arc_alloc_buf(os->os_spa, db, type, psize);
3687 bcopy(db->db.db_data, (*datap)->b_data, psize);
3689 db->db_data_pending = dr;
3691 mutex_exit(&db->db_mtx);
3693 dbuf_write(dr, *datap, tx);
3695 ASSERT(!list_link_active(&dr->dr_dirty_node));
3696 if (dn->dn_object == DMU_META_DNODE_OBJECT) {
3697 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
3701 * Although zio_nowait() does not "wait for an IO", it does
3702 * initiate the IO. If this is an empty write it seems plausible
3703 * that the IO could actually be completed before the nowait
3704 * returns. We need to DB_DNODE_EXIT() first in case
3705 * zio_nowait() invalidates the dbuf.
3708 zio_nowait(dr->dr_zio);
3713 dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx)
3715 dbuf_dirty_record_t *dr;
3717 while ((dr = list_head(list))) {
3718 if (dr->dr_zio != NULL) {
3720 * If we find an already initialized zio then we
3721 * are processing the meta-dnode, and we have finished.
3722 * The dbufs for all dnodes are put back on the list
3723 * during processing, so that we can zio_wait()
3724 * these IOs after initiating all child IOs.
3726 ASSERT3U(dr->dr_dbuf->db.db_object, ==,
3727 DMU_META_DNODE_OBJECT);
3730 if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
3731 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
3732 VERIFY3U(dr->dr_dbuf->db_level, ==, level);
3734 list_remove(list, dr);
3735 if (dr->dr_dbuf->db_level > 0)
3736 dbuf_sync_indirect(dr, tx);
3738 dbuf_sync_leaf(dr, tx);
3744 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3746 dmu_buf_impl_t *db = vdb;
3748 blkptr_t *bp = zio->io_bp;
3749 blkptr_t *bp_orig = &zio->io_bp_orig;
3750 spa_t *spa = zio->io_spa;
3755 ASSERT3P(db->db_blkptr, !=, NULL);
3756 ASSERT3P(&db->db_data_pending->dr_bp_copy, ==, bp);
3760 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
3761 dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
3762 zio->io_prev_space_delta = delta;
3764 if (bp->blk_birth != 0) {
3765 ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
3766 BP_GET_TYPE(bp) == dn->dn_type) ||
3767 (db->db_blkid == DMU_SPILL_BLKID &&
3768 BP_GET_TYPE(bp) == dn->dn_bonustype) ||
3769 BP_IS_EMBEDDED(bp));
3770 ASSERT(BP_GET_LEVEL(bp) == db->db_level);
3773 mutex_enter(&db->db_mtx);
3776 if (db->db_blkid == DMU_SPILL_BLKID) {
3777 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3778 ASSERT(!(BP_IS_HOLE(bp)) &&
3779 db->db_blkptr == DN_SPILL_BLKPTR(dn->dn_phys));
3783 if (db->db_level == 0) {
3784 mutex_enter(&dn->dn_mtx);
3785 if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
3786 db->db_blkid != DMU_SPILL_BLKID)
3787 dn->dn_phys->dn_maxblkid = db->db_blkid;
3788 mutex_exit(&dn->dn_mtx);
3790 if (dn->dn_type == DMU_OT_DNODE) {
3792 while (i < db->db.db_size) {
3794 (void *)(((char *)db->db.db_data) + i);
3796 i += DNODE_MIN_SIZE;
3797 if (dnp->dn_type != DMU_OT_NONE) {
3799 i += dnp->dn_extra_slots *
3804 if (BP_IS_HOLE(bp)) {
3811 blkptr_t *ibp = db->db.db_data;
3812 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3813 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
3814 if (BP_IS_HOLE(ibp))
3816 fill += BP_GET_FILL(ibp);
3821 if (!BP_IS_EMBEDDED(bp))
3822 BP_SET_FILL(bp, fill);
3824 mutex_exit(&db->db_mtx);
3826 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3827 *db->db_blkptr = *bp;
3828 rw_exit(&dn->dn_struct_rwlock);
3833 * This function gets called just prior to running through the compression
3834 * stage of the zio pipeline. If we're an indirect block comprised of only
3835 * holes, then we want this indirect to be compressed away to a hole. In
3836 * order to do that we must zero out any information about the holes that
3837 * this indirect points to prior to before we try to compress it.
3840 dbuf_write_children_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3842 dmu_buf_impl_t *db = vdb;
3845 unsigned int epbs, i;
3847 ASSERT3U(db->db_level, >, 0);
3850 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3851 ASSERT3U(epbs, <, 31);
3853 /* Determine if all our children are holes */
3854 for (i = 0, bp = db->db.db_data; i < 1ULL << epbs; i++, bp++) {
3855 if (!BP_IS_HOLE(bp))
3860 * If all the children are holes, then zero them all out so that
3861 * we may get compressed away.
3863 if (i == 1ULL << epbs) {
3865 * We only found holes. Grab the rwlock to prevent
3866 * anybody from reading the blocks we're about to
3869 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3870 bzero(db->db.db_data, db->db.db_size);
3871 rw_exit(&dn->dn_struct_rwlock);
3877 * The SPA will call this callback several times for each zio - once
3878 * for every physical child i/o (zio->io_phys_children times). This
3879 * allows the DMU to monitor the progress of each logical i/o. For example,
3880 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
3881 * block. There may be a long delay before all copies/fragments are completed,
3882 * so this callback allows us to retire dirty space gradually, as the physical
3887 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
3889 dmu_buf_impl_t *db = arg;
3890 objset_t *os = db->db_objset;
3891 dsl_pool_t *dp = dmu_objset_pool(os);
3892 dbuf_dirty_record_t *dr;
3895 dr = db->db_data_pending;
3896 ASSERT3U(dr->dr_txg, ==, zio->io_txg);
3899 * The callback will be called io_phys_children times. Retire one
3900 * portion of our dirty space each time we are called. Any rounding
3901 * error will be cleaned up by dsl_pool_sync()'s call to
3902 * dsl_pool_undirty_space().
3904 delta = dr->dr_accounted / zio->io_phys_children;
3905 dsl_pool_undirty_space(dp, delta, zio->io_txg);
3910 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
3912 dmu_buf_impl_t *db = vdb;
3913 blkptr_t *bp_orig = &zio->io_bp_orig;
3914 blkptr_t *bp = db->db_blkptr;
3915 objset_t *os = db->db_objset;
3916 dmu_tx_t *tx = os->os_synctx;
3917 dbuf_dirty_record_t **drp, *dr;
3919 ASSERT0(zio->io_error);
3920 ASSERT(db->db_blkptr == bp);
3923 * For nopwrites and rewrites we ensure that the bp matches our
3924 * original and bypass all the accounting.
3926 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
3927 ASSERT(BP_EQUAL(bp, bp_orig));
3929 dsl_dataset_t *ds = os->os_dsl_dataset;
3930 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
3931 dsl_dataset_block_born(ds, bp, tx);
3934 mutex_enter(&db->db_mtx);
3938 drp = &db->db_last_dirty;
3939 while ((dr = *drp) != db->db_data_pending)
3941 ASSERT(!list_link_active(&dr->dr_dirty_node));
3942 ASSERT(dr->dr_dbuf == db);
3943 ASSERT(dr->dr_next == NULL);
3947 if (db->db_blkid == DMU_SPILL_BLKID) {
3952 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3953 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
3954 db->db_blkptr == DN_SPILL_BLKPTR(dn->dn_phys));
3959 if (db->db_level == 0) {
3960 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
3961 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
3962 if (db->db_state != DB_NOFILL) {
3963 if (dr->dt.dl.dr_data != db->db_buf)
3964 arc_buf_destroy(dr->dt.dl.dr_data, db);
3971 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3972 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
3973 if (!BP_IS_HOLE(db->db_blkptr)) {
3974 ASSERTV(int epbs = dn->dn_phys->dn_indblkshift -
3976 ASSERT3U(db->db_blkid, <=,
3977 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs));
3978 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
3982 mutex_destroy(&dr->dt.di.dr_mtx);
3983 list_destroy(&dr->dt.di.dr_children);
3985 kmem_free(dr, sizeof (dbuf_dirty_record_t));
3987 cv_broadcast(&db->db_changed);
3988 ASSERT(db->db_dirtycnt > 0);
3989 db->db_dirtycnt -= 1;
3990 db->db_data_pending = NULL;
3991 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg);
3995 dbuf_write_nofill_ready(zio_t *zio)
3997 dbuf_write_ready(zio, NULL, zio->io_private);
4001 dbuf_write_nofill_done(zio_t *zio)
4003 dbuf_write_done(zio, NULL, zio->io_private);
4007 dbuf_write_override_ready(zio_t *zio)
4009 dbuf_dirty_record_t *dr = zio->io_private;
4010 dmu_buf_impl_t *db = dr->dr_dbuf;
4012 dbuf_write_ready(zio, NULL, db);
4016 dbuf_write_override_done(zio_t *zio)
4018 dbuf_dirty_record_t *dr = zio->io_private;
4019 dmu_buf_impl_t *db = dr->dr_dbuf;
4020 blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
4022 mutex_enter(&db->db_mtx);
4023 if (!BP_EQUAL(zio->io_bp, obp)) {
4024 if (!BP_IS_HOLE(obp))
4025 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
4026 arc_release(dr->dt.dl.dr_data, db);
4028 mutex_exit(&db->db_mtx);
4030 dbuf_write_done(zio, NULL, db);
4032 if (zio->io_abd != NULL)
4033 abd_put(zio->io_abd);
4036 /* Issue I/O to commit a dirty buffer to disk. */
4038 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
4040 dmu_buf_impl_t *db = dr->dr_dbuf;
4043 dmu_buf_impl_t *parent = db->db_parent;
4044 uint64_t txg = tx->tx_txg;
4045 zbookmark_phys_t zb;
4050 ASSERT(dmu_tx_is_syncing(tx));
4056 if (db->db_state != DB_NOFILL) {
4057 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
4059 * Private object buffers are released here rather
4060 * than in dbuf_dirty() since they are only modified
4061 * in the syncing context and we don't want the
4062 * overhead of making multiple copies of the data.
4064 if (BP_IS_HOLE(db->db_blkptr)) {
4067 dbuf_release_bp(db);
4072 if (parent != dn->dn_dbuf) {
4073 /* Our parent is an indirect block. */
4074 /* We have a dirty parent that has been scheduled for write. */
4075 ASSERT(parent && parent->db_data_pending);
4076 /* Our parent's buffer is one level closer to the dnode. */
4077 ASSERT(db->db_level == parent->db_level-1);
4079 * We're about to modify our parent's db_data by modifying
4080 * our block pointer, so the parent must be released.
4082 ASSERT(arc_released(parent->db_buf));
4083 zio = parent->db_data_pending->dr_zio;
4085 /* Our parent is the dnode itself. */
4086 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
4087 db->db_blkid != DMU_SPILL_BLKID) ||
4088 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
4089 if (db->db_blkid != DMU_SPILL_BLKID)
4090 ASSERT3P(db->db_blkptr, ==,
4091 &dn->dn_phys->dn_blkptr[db->db_blkid]);
4095 ASSERT(db->db_level == 0 || data == db->db_buf);
4096 ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
4099 SET_BOOKMARK(&zb, os->os_dsl_dataset ?
4100 os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
4101 db->db.db_object, db->db_level, db->db_blkid);
4103 if (db->db_blkid == DMU_SPILL_BLKID)
4105 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
4107 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
4111 * We copy the blkptr now (rather than when we instantiate the dirty
4112 * record), because its value can change between open context and
4113 * syncing context. We do not need to hold dn_struct_rwlock to read
4114 * db_blkptr because we are in syncing context.
4116 dr->dr_bp_copy = *db->db_blkptr;
4118 if (db->db_level == 0 &&
4119 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
4121 * The BP for this block has been provided by open context
4122 * (by dmu_sync() or dmu_buf_write_embedded()).
4124 abd_t *contents = (data != NULL) ?
4125 abd_get_from_buf(data->b_data, arc_buf_size(data)) : NULL;
4127 dr->dr_zio = zio_write(zio, os->os_spa, txg,
4128 &dr->dr_bp_copy, contents, db->db.db_size, db->db.db_size,
4129 &zp, dbuf_write_override_ready, NULL, NULL,
4130 dbuf_write_override_done,
4131 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
4132 mutex_enter(&db->db_mtx);
4133 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
4134 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
4135 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
4136 mutex_exit(&db->db_mtx);
4137 } else if (db->db_state == DB_NOFILL) {
4138 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF ||
4139 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY);
4140 dr->dr_zio = zio_write(zio, os->os_spa, txg,
4141 &dr->dr_bp_copy, NULL, db->db.db_size, db->db.db_size, &zp,
4142 dbuf_write_nofill_ready, NULL, NULL,
4143 dbuf_write_nofill_done, db,
4144 ZIO_PRIORITY_ASYNC_WRITE,
4145 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
4147 ASSERT(arc_released(data));
4150 * For indirect blocks, we want to setup the children
4151 * ready callback so that we can properly handle an indirect
4152 * block that only contains holes.
4154 arc_write_done_func_t *children_ready_cb = NULL;
4155 if (db->db_level != 0)
4156 children_ready_cb = dbuf_write_children_ready;
4158 dr->dr_zio = arc_write(zio, os->os_spa, txg,
4159 &dr->dr_bp_copy, data, DBUF_IS_L2CACHEABLE(db),
4160 &zp, dbuf_write_ready,
4161 children_ready_cb, dbuf_write_physdone,
4162 dbuf_write_done, db, ZIO_PRIORITY_ASYNC_WRITE,
4163 ZIO_FLAG_MUSTSUCCEED, &zb);
4167 #if defined(_KERNEL) && defined(HAVE_SPL)
4168 EXPORT_SYMBOL(dbuf_find);
4169 EXPORT_SYMBOL(dbuf_is_metadata);
4170 EXPORT_SYMBOL(dbuf_destroy);
4171 EXPORT_SYMBOL(dbuf_loan_arcbuf);
4172 EXPORT_SYMBOL(dbuf_whichblock);
4173 EXPORT_SYMBOL(dbuf_read);
4174 EXPORT_SYMBOL(dbuf_unoverride);
4175 EXPORT_SYMBOL(dbuf_free_range);
4176 EXPORT_SYMBOL(dbuf_new_size);
4177 EXPORT_SYMBOL(dbuf_release_bp);
4178 EXPORT_SYMBOL(dbuf_dirty);
4179 EXPORT_SYMBOL(dmu_buf_will_change_crypt_params);
4180 EXPORT_SYMBOL(dmu_buf_will_dirty);
4181 EXPORT_SYMBOL(dmu_buf_will_not_fill);
4182 EXPORT_SYMBOL(dmu_buf_will_fill);
4183 EXPORT_SYMBOL(dmu_buf_fill_done);
4184 EXPORT_SYMBOL(dmu_buf_rele);
4185 EXPORT_SYMBOL(dbuf_assign_arcbuf);
4186 EXPORT_SYMBOL(dbuf_prefetch);
4187 EXPORT_SYMBOL(dbuf_hold_impl);
4188 EXPORT_SYMBOL(dbuf_hold);
4189 EXPORT_SYMBOL(dbuf_hold_level);
4190 EXPORT_SYMBOL(dbuf_create_bonus);
4191 EXPORT_SYMBOL(dbuf_spill_set_blksz);
4192 EXPORT_SYMBOL(dbuf_rm_spill);
4193 EXPORT_SYMBOL(dbuf_add_ref);
4194 EXPORT_SYMBOL(dbuf_rele);
4195 EXPORT_SYMBOL(dbuf_rele_and_unlock);
4196 EXPORT_SYMBOL(dbuf_refcount);
4197 EXPORT_SYMBOL(dbuf_sync_list);
4198 EXPORT_SYMBOL(dmu_buf_set_user);
4199 EXPORT_SYMBOL(dmu_buf_set_user_ie);
4200 EXPORT_SYMBOL(dmu_buf_get_user);
4201 EXPORT_SYMBOL(dmu_buf_get_blkptr);
4204 module_param(dbuf_cache_max_bytes, ulong, 0644);
4205 MODULE_PARM_DESC(dbuf_cache_max_bytes,
4206 "Maximum size in bytes of the dbuf cache.");
4208 module_param(dbuf_cache_hiwater_pct, uint, 0644);
4209 MODULE_PARM_DESC(dbuf_cache_hiwater_pct,
4210 "Percentage over dbuf_cache_max_bytes when dbufs must be evicted "
4213 module_param(dbuf_cache_lowater_pct, uint, 0644);
4214 MODULE_PARM_DESC(dbuf_cache_lowater_pct,
4215 "Percentage below dbuf_cache_max_bytes when the evict thread stops "
4218 module_param(dbuf_cache_max_shift, int, 0644);
4219 MODULE_PARM_DESC(dbuf_cache_max_shift,
4220 "Cap the size of the dbuf cache to a log2 fraction of arc size.");