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
= 0;
191 /* Set the default size of the dbuf cache to log2 fraction of arc size. */
192 int dbuf_cache_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_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 set the size of the
791 * dbuf cache to 1/32nd (default) of the target size of the ARC. If
792 * the value has been specified as a module option and it's not
793 * greater than the target size of the ARC, then we honor that value.
795 if (dbuf_cache_max_bytes
== 0 ||
796 dbuf_cache_max_bytes
>= arc_target_bytes()) {
797 dbuf_cache_max_bytes
= arc_target_bytes() >> dbuf_cache_shift
;
801 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
802 * configuration is not required.
804 dbu_evict_taskq
= taskq_create("dbu_evict", 1, defclsyspri
, 0, 0, 0);
806 dbuf_cache
= multilist_create(sizeof (dmu_buf_impl_t
),
807 offsetof(dmu_buf_impl_t
, db_cache_link
),
808 dbuf_cache_multilist_index_func
);
809 refcount_create(&dbuf_cache_size
);
811 tsd_create(&zfs_dbuf_evict_key
, NULL
);
812 dbuf_evict_thread_exit
= B_FALSE
;
813 mutex_init(&dbuf_evict_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
814 cv_init(&dbuf_evict_cv
, NULL
, CV_DEFAULT
, NULL
);
815 dbuf_cache_evict_thread
= thread_create(NULL
, 0, dbuf_evict_thread
,
816 NULL
, 0, &p0
, TS_RUN
, minclsyspri
);
818 dbuf_ksp
= kstat_create("zfs", 0, "dbufstats", "misc",
819 KSTAT_TYPE_NAMED
, sizeof (dbuf_stats
) / sizeof (kstat_named_t
),
821 if (dbuf_ksp
!= NULL
) {
822 dbuf_ksp
->ks_data
= &dbuf_stats
;
823 dbuf_ksp
->ks_update
= dbuf_kstat_update
;
824 kstat_install(dbuf_ksp
);
826 for (i
= 0; i
< DN_MAX_LEVELS
; i
++) {
827 snprintf(dbuf_stats
.cache_levels
[i
].name
,
828 KSTAT_STRLEN
, "cache_level_%d", i
);
829 dbuf_stats
.cache_levels
[i
].data_type
=
831 snprintf(dbuf_stats
.cache_levels_bytes
[i
].name
,
832 KSTAT_STRLEN
, "cache_level_%d_bytes", i
);
833 dbuf_stats
.cache_levels_bytes
[i
].data_type
=
842 dbuf_hash_table_t
*h
= &dbuf_hash_table
;
845 dbuf_stats_destroy();
847 for (i
= 0; i
< DBUF_MUTEXES
; i
++)
848 mutex_destroy(&h
->hash_mutexes
[i
]);
849 #if defined(_KERNEL) && defined(HAVE_SPL)
851 * Large allocations which do not require contiguous pages
852 * should be using vmem_free() in the linux kernel
854 vmem_free(h
->hash_table
, (h
->hash_table_mask
+ 1) * sizeof (void *));
856 kmem_free(h
->hash_table
, (h
->hash_table_mask
+ 1) * sizeof (void *));
858 kmem_cache_destroy(dbuf_kmem_cache
);
859 taskq_destroy(dbu_evict_taskq
);
861 mutex_enter(&dbuf_evict_lock
);
862 dbuf_evict_thread_exit
= B_TRUE
;
863 while (dbuf_evict_thread_exit
) {
864 cv_signal(&dbuf_evict_cv
);
865 cv_wait(&dbuf_evict_cv
, &dbuf_evict_lock
);
867 mutex_exit(&dbuf_evict_lock
);
868 tsd_destroy(&zfs_dbuf_evict_key
);
870 mutex_destroy(&dbuf_evict_lock
);
871 cv_destroy(&dbuf_evict_cv
);
873 refcount_destroy(&dbuf_cache_size
);
874 multilist_destroy(dbuf_cache
);
876 if (dbuf_ksp
!= NULL
) {
877 kstat_delete(dbuf_ksp
);
888 dbuf_verify(dmu_buf_impl_t
*db
)
891 dbuf_dirty_record_t
*dr
;
893 ASSERT(MUTEX_HELD(&db
->db_mtx
));
895 if (!(zfs_flags
& ZFS_DEBUG_DBUF_VERIFY
))
898 ASSERT(db
->db_objset
!= NULL
);
902 ASSERT(db
->db_parent
== NULL
);
903 ASSERT(db
->db_blkptr
== NULL
);
905 ASSERT3U(db
->db
.db_object
, ==, dn
->dn_object
);
906 ASSERT3P(db
->db_objset
, ==, dn
->dn_objset
);
907 ASSERT3U(db
->db_level
, <, dn
->dn_nlevels
);
908 ASSERT(db
->db_blkid
== DMU_BONUS_BLKID
||
909 db
->db_blkid
== DMU_SPILL_BLKID
||
910 !avl_is_empty(&dn
->dn_dbufs
));
912 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
914 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
915 ASSERT3U(db
->db
.db_offset
, ==, DMU_BONUS_BLKID
);
916 } else if (db
->db_blkid
== DMU_SPILL_BLKID
) {
918 ASSERT0(db
->db
.db_offset
);
920 ASSERT3U(db
->db
.db_offset
, ==, db
->db_blkid
* db
->db
.db_size
);
923 for (dr
= db
->db_data_pending
; dr
!= NULL
; dr
= dr
->dr_next
)
924 ASSERT(dr
->dr_dbuf
== db
);
926 for (dr
= db
->db_last_dirty
; dr
!= NULL
; dr
= dr
->dr_next
)
927 ASSERT(dr
->dr_dbuf
== db
);
930 * We can't assert that db_size matches dn_datablksz because it
931 * can be momentarily different when another thread is doing
934 if (db
->db_level
== 0 && db
->db
.db_object
== DMU_META_DNODE_OBJECT
) {
935 dr
= db
->db_data_pending
;
937 * It should only be modified in syncing context, so
938 * make sure we only have one copy of the data.
940 ASSERT(dr
== NULL
|| dr
->dt
.dl
.dr_data
== db
->db_buf
);
943 /* verify db->db_blkptr */
945 if (db
->db_parent
== dn
->dn_dbuf
) {
946 /* db is pointed to by the dnode */
947 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
948 if (DMU_OBJECT_IS_SPECIAL(db
->db
.db_object
))
949 ASSERT(db
->db_parent
== NULL
);
951 ASSERT(db
->db_parent
!= NULL
);
952 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
953 ASSERT3P(db
->db_blkptr
, ==,
954 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
956 /* db is pointed to by an indirect block */
957 ASSERTV(int epb
= db
->db_parent
->db
.db_size
>>
959 ASSERT3U(db
->db_parent
->db_level
, ==, db
->db_level
+1);
960 ASSERT3U(db
->db_parent
->db
.db_object
, ==,
963 * dnode_grow_indblksz() can make this fail if we don't
964 * have the struct_rwlock. XXX indblksz no longer
965 * grows. safe to do this now?
967 if (RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
968 ASSERT3P(db
->db_blkptr
, ==,
969 ((blkptr_t
*)db
->db_parent
->db
.db_data
+
970 db
->db_blkid
% epb
));
974 if ((db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
)) &&
975 (db
->db_buf
== NULL
|| db
->db_buf
->b_data
) &&
976 db
->db
.db_data
&& db
->db_blkid
!= DMU_BONUS_BLKID
&&
977 db
->db_state
!= DB_FILL
&& !dn
->dn_free_txg
) {
979 * If the blkptr isn't set but they have nonzero data,
980 * it had better be dirty, otherwise we'll lose that
981 * data when we evict this buffer.
983 * There is an exception to this rule for indirect blocks; in
984 * this case, if the indirect block is a hole, we fill in a few
985 * fields on each of the child blocks (importantly, birth time)
986 * to prevent hole birth times from being lost when you
987 * partially fill in a hole.
989 if (db
->db_dirtycnt
== 0) {
990 if (db
->db_level
== 0) {
991 uint64_t *buf
= db
->db
.db_data
;
994 for (i
= 0; i
< db
->db
.db_size
>> 3; i
++) {
998 blkptr_t
*bps
= db
->db
.db_data
;
999 ASSERT3U(1 << DB_DNODE(db
)->dn_indblkshift
, ==,
1002 * We want to verify that all the blkptrs in the
1003 * indirect block are holes, but we may have
1004 * automatically set up a few fields for them.
1005 * We iterate through each blkptr and verify
1006 * they only have those fields set.
1009 i
< db
->db
.db_size
/ sizeof (blkptr_t
);
1011 blkptr_t
*bp
= &bps
[i
];
1012 ASSERT(ZIO_CHECKSUM_IS_ZERO(
1015 DVA_IS_EMPTY(&bp
->blk_dva
[0]) &&
1016 DVA_IS_EMPTY(&bp
->blk_dva
[1]) &&
1017 DVA_IS_EMPTY(&bp
->blk_dva
[2]));
1018 ASSERT0(bp
->blk_fill
);
1019 ASSERT0(bp
->blk_pad
[0]);
1020 ASSERT0(bp
->blk_pad
[1]);
1021 ASSERT(!BP_IS_EMBEDDED(bp
));
1022 ASSERT(BP_IS_HOLE(bp
));
1023 ASSERT0(bp
->blk_phys_birth
);
1033 dbuf_clear_data(dmu_buf_impl_t
*db
)
1035 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1036 dbuf_evict_user(db
);
1037 ASSERT3P(db
->db_buf
, ==, NULL
);
1038 db
->db
.db_data
= NULL
;
1039 if (db
->db_state
!= DB_NOFILL
)
1040 db
->db_state
= DB_UNCACHED
;
1044 dbuf_set_data(dmu_buf_impl_t
*db
, arc_buf_t
*buf
)
1046 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1047 ASSERT(buf
!= NULL
);
1050 ASSERT(buf
->b_data
!= NULL
);
1051 db
->db
.db_data
= buf
->b_data
;
1055 * Loan out an arc_buf for read. Return the loaned arc_buf.
1058 dbuf_loan_arcbuf(dmu_buf_impl_t
*db
)
1062 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1063 mutex_enter(&db
->db_mtx
);
1064 if (arc_released(db
->db_buf
) || refcount_count(&db
->db_holds
) > 1) {
1065 int blksz
= db
->db
.db_size
;
1066 spa_t
*spa
= db
->db_objset
->os_spa
;
1068 mutex_exit(&db
->db_mtx
);
1069 abuf
= arc_loan_buf(spa
, B_FALSE
, blksz
);
1070 bcopy(db
->db
.db_data
, abuf
->b_data
, blksz
);
1073 arc_loan_inuse_buf(abuf
, db
);
1075 dbuf_clear_data(db
);
1076 mutex_exit(&db
->db_mtx
);
1082 * Calculate which level n block references the data at the level 0 offset
1086 dbuf_whichblock(const dnode_t
*dn
, const int64_t level
, const uint64_t offset
)
1088 if (dn
->dn_datablkshift
!= 0 && dn
->dn_indblkshift
!= 0) {
1090 * The level n blkid is equal to the level 0 blkid divided by
1091 * the number of level 0s in a level n block.
1093 * The level 0 blkid is offset >> datablkshift =
1094 * offset / 2^datablkshift.
1096 * The number of level 0s in a level n is the number of block
1097 * pointers in an indirect block, raised to the power of level.
1098 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
1099 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
1101 * Thus, the level n blkid is: offset /
1102 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
1103 * = offset / 2^(datablkshift + level *
1104 * (indblkshift - SPA_BLKPTRSHIFT))
1105 * = offset >> (datablkshift + level *
1106 * (indblkshift - SPA_BLKPTRSHIFT))
1109 const unsigned exp
= dn
->dn_datablkshift
+
1110 level
* (dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
);
1112 if (exp
>= 8 * sizeof (offset
)) {
1113 /* This only happens on the highest indirection level */
1114 ASSERT3U(level
, ==, dn
->dn_nlevels
- 1);
1118 ASSERT3U(exp
, <, 8 * sizeof (offset
));
1120 return (offset
>> exp
);
1122 ASSERT3U(offset
, <, dn
->dn_datablksz
);
1128 dbuf_read_done(zio_t
*zio
, const zbookmark_phys_t
*zb
, const blkptr_t
*bp
,
1129 arc_buf_t
*buf
, void *vdb
)
1131 dmu_buf_impl_t
*db
= vdb
;
1133 mutex_enter(&db
->db_mtx
);
1134 ASSERT3U(db
->db_state
, ==, DB_READ
);
1136 * All reads are synchronous, so we must have a hold on the dbuf
1138 ASSERT(refcount_count(&db
->db_holds
) > 0);
1139 ASSERT(db
->db_buf
== NULL
);
1140 ASSERT(db
->db
.db_data
== NULL
);
1141 if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
1142 /* we were freed in flight; disregard any error */
1144 buf
= arc_alloc_buf(db
->db_objset
->os_spa
,
1145 db
, DBUF_GET_BUFC_TYPE(db
), db
->db
.db_size
);
1147 arc_release(buf
, db
);
1148 bzero(buf
->b_data
, db
->db
.db_size
);
1149 arc_buf_freeze(buf
);
1150 db
->db_freed_in_flight
= FALSE
;
1151 dbuf_set_data(db
, buf
);
1152 db
->db_state
= DB_CACHED
;
1153 } else if (buf
!= NULL
) {
1154 dbuf_set_data(db
, buf
);
1155 db
->db_state
= DB_CACHED
;
1157 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1158 ASSERT3P(db
->db_buf
, ==, NULL
);
1159 db
->db_state
= DB_UNCACHED
;
1161 cv_broadcast(&db
->db_changed
);
1162 dbuf_rele_and_unlock(db
, NULL
);
1166 dbuf_read_impl(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
)
1169 zbookmark_phys_t zb
;
1170 uint32_t aflags
= ARC_FLAG_NOWAIT
;
1171 int err
, zio_flags
= 0;
1175 ASSERT(!refcount_is_zero(&db
->db_holds
));
1176 /* We need the struct_rwlock to prevent db_blkptr from changing. */
1177 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
1178 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1179 ASSERT(db
->db_state
== DB_UNCACHED
);
1180 ASSERT(db
->db_buf
== NULL
);
1182 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1184 * The bonus length stored in the dnode may be less than
1185 * the maximum available space in the bonus buffer.
1187 int bonuslen
= MIN(dn
->dn_bonuslen
, dn
->dn_phys
->dn_bonuslen
);
1188 int max_bonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
1189 arc_buf_t
*dn_buf
= (dn
->dn_dbuf
!= NULL
) ?
1190 dn
->dn_dbuf
->db_buf
: NULL
;
1192 /* if the underlying dnode block is encrypted, decrypt it */
1193 if (dn_buf
!= NULL
&& dn
->dn_objset
->os_encrypted
&&
1194 DMU_OT_IS_ENCRYPTED(dn
->dn_bonustype
) &&
1195 (flags
& DB_RF_NO_DECRYPT
) == 0 &&
1196 arc_is_encrypted(dn_buf
)) {
1197 SET_BOOKMARK(&zb
, dmu_objset_id(db
->db_objset
),
1198 DMU_META_DNODE_OBJECT
, 0, dn
->dn_dbuf
->db_blkid
);
1199 err
= arc_untransform(dn_buf
, dn
->dn_objset
->os_spa
,
1203 mutex_exit(&db
->db_mtx
);
1208 ASSERT3U(bonuslen
, <=, db
->db
.db_size
);
1209 db
->db
.db_data
= kmem_alloc(max_bonuslen
, KM_SLEEP
);
1210 arc_space_consume(max_bonuslen
, ARC_SPACE_BONUS
);
1211 if (bonuslen
< max_bonuslen
)
1212 bzero(db
->db
.db_data
, max_bonuslen
);
1214 bcopy(DN_BONUS(dn
->dn_phys
), db
->db
.db_data
, bonuslen
);
1216 db
->db_state
= DB_CACHED
;
1217 mutex_exit(&db
->db_mtx
);
1222 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
1223 * processes the delete record and clears the bp while we are waiting
1224 * for the dn_mtx (resulting in a "no" from block_freed).
1226 if (db
->db_blkptr
== NULL
|| BP_IS_HOLE(db
->db_blkptr
) ||
1227 (db
->db_level
== 0 && (dnode_block_freed(dn
, db
->db_blkid
) ||
1228 BP_IS_HOLE(db
->db_blkptr
)))) {
1229 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1231 dbuf_set_data(db
, arc_alloc_buf(db
->db_objset
->os_spa
, db
, type
,
1233 bzero(db
->db
.db_data
, db
->db
.db_size
);
1235 if (db
->db_blkptr
!= NULL
&& db
->db_level
> 0 &&
1236 BP_IS_HOLE(db
->db_blkptr
) &&
1237 db
->db_blkptr
->blk_birth
!= 0) {
1238 blkptr_t
*bps
= db
->db
.db_data
;
1239 for (int i
= 0; i
< ((1 <<
1240 DB_DNODE(db
)->dn_indblkshift
) / sizeof (blkptr_t
));
1242 blkptr_t
*bp
= &bps
[i
];
1243 ASSERT3U(BP_GET_LSIZE(db
->db_blkptr
), ==,
1244 1 << dn
->dn_indblkshift
);
1246 BP_GET_LEVEL(db
->db_blkptr
) == 1 ?
1248 BP_GET_LSIZE(db
->db_blkptr
));
1249 BP_SET_TYPE(bp
, BP_GET_TYPE(db
->db_blkptr
));
1251 BP_GET_LEVEL(db
->db_blkptr
) - 1);
1252 BP_SET_BIRTH(bp
, db
->db_blkptr
->blk_birth
, 0);
1256 db
->db_state
= DB_CACHED
;
1257 mutex_exit(&db
->db_mtx
);
1263 db
->db_state
= DB_READ
;
1264 mutex_exit(&db
->db_mtx
);
1266 if (DBUF_IS_L2CACHEABLE(db
))
1267 aflags
|= ARC_FLAG_L2CACHE
;
1269 SET_BOOKMARK(&zb
, dmu_objset_id(db
->db_objset
),
1270 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1273 * All bps of an encrypted os should have the encryption bit set.
1274 * If this is not true it indicates tampering and we report an error.
1276 if (db
->db_objset
->os_encrypted
&& !BP_USES_CRYPT(db
->db_blkptr
)) {
1277 spa_log_error(db
->db_objset
->os_spa
, &zb
);
1278 zfs_panic_recover("unencrypted block in encrypted "
1279 "object set %llu", dmu_objset_id(db
->db_objset
));
1280 return (SET_ERROR(EIO
));
1283 dbuf_add_ref(db
, NULL
);
1285 zio_flags
= (flags
& DB_RF_CANFAIL
) ?
1286 ZIO_FLAG_CANFAIL
: ZIO_FLAG_MUSTSUCCEED
;
1288 if ((flags
& DB_RF_NO_DECRYPT
) && BP_IS_PROTECTED(db
->db_blkptr
))
1289 zio_flags
|= ZIO_FLAG_RAW
;
1291 err
= arc_read(zio
, db
->db_objset
->os_spa
, db
->db_blkptr
,
1292 dbuf_read_done
, db
, ZIO_PRIORITY_SYNC_READ
, zio_flags
,
1299 * This is our just-in-time copy function. It makes a copy of buffers that
1300 * have been modified in a previous transaction group before we access them in
1301 * the current active group.
1303 * This function is used in three places: when we are dirtying a buffer for the
1304 * first time in a txg, when we are freeing a range in a dnode that includes
1305 * this buffer, and when we are accessing a buffer which was received compressed
1306 * and later referenced in a WRITE_BYREF record.
1308 * Note that when we are called from dbuf_free_range() we do not put a hold on
1309 * the buffer, we just traverse the active dbuf list for the dnode.
1312 dbuf_fix_old_data(dmu_buf_impl_t
*db
, uint64_t txg
)
1314 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
1316 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1317 ASSERT(db
->db
.db_data
!= NULL
);
1318 ASSERT(db
->db_level
== 0);
1319 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
);
1322 (dr
->dt
.dl
.dr_data
!=
1323 ((db
->db_blkid
== DMU_BONUS_BLKID
) ? db
->db
.db_data
: db
->db_buf
)))
1327 * If the last dirty record for this dbuf has not yet synced
1328 * and its referencing the dbuf data, either:
1329 * reset the reference to point to a new copy,
1330 * or (if there a no active holders)
1331 * just null out the current db_data pointer.
1333 ASSERT3U(dr
->dr_txg
, >=, txg
- 2);
1334 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1335 dnode_t
*dn
= DB_DNODE(db
);
1336 int bonuslen
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
);
1337 dr
->dt
.dl
.dr_data
= kmem_alloc(bonuslen
, KM_SLEEP
);
1338 arc_space_consume(bonuslen
, ARC_SPACE_BONUS
);
1339 bcopy(db
->db
.db_data
, dr
->dt
.dl
.dr_data
, bonuslen
);
1340 } else if (refcount_count(&db
->db_holds
) > db
->db_dirtycnt
) {
1341 dnode_t
*dn
= DB_DNODE(db
);
1342 int size
= arc_buf_size(db
->db_buf
);
1343 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1344 spa_t
*spa
= db
->db_objset
->os_spa
;
1345 enum zio_compress compress_type
=
1346 arc_get_compression(db
->db_buf
);
1348 if (arc_is_encrypted(db
->db_buf
)) {
1349 boolean_t byteorder
;
1350 uint8_t salt
[ZIO_DATA_SALT_LEN
];
1351 uint8_t iv
[ZIO_DATA_IV_LEN
];
1352 uint8_t mac
[ZIO_DATA_MAC_LEN
];
1354 arc_get_raw_params(db
->db_buf
, &byteorder
, salt
,
1356 dr
->dt
.dl
.dr_data
= arc_alloc_raw_buf(spa
, db
,
1357 dmu_objset_id(dn
->dn_objset
), byteorder
, salt
, iv
,
1358 mac
, dn
->dn_type
, size
, arc_buf_lsize(db
->db_buf
),
1360 } else if (compress_type
!= ZIO_COMPRESS_OFF
) {
1361 ASSERT3U(type
, ==, ARC_BUFC_DATA
);
1362 dr
->dt
.dl
.dr_data
= arc_alloc_compressed_buf(spa
, db
,
1363 size
, arc_buf_lsize(db
->db_buf
), compress_type
);
1365 dr
->dt
.dl
.dr_data
= arc_alloc_buf(spa
, db
, type
, size
);
1367 bcopy(db
->db
.db_data
, dr
->dt
.dl
.dr_data
->b_data
, size
);
1370 dbuf_clear_data(db
);
1375 dbuf_read(dmu_buf_impl_t
*db
, zio_t
*zio
, uint32_t flags
)
1382 * We don't have to hold the mutex to check db_state because it
1383 * can't be freed while we have a hold on the buffer.
1385 ASSERT(!refcount_is_zero(&db
->db_holds
));
1387 if (db
->db_state
== DB_NOFILL
)
1388 return (SET_ERROR(EIO
));
1392 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
1393 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1395 prefetch
= db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
1396 (flags
& DB_RF_NOPREFETCH
) == 0 && dn
!= NULL
&&
1397 DBUF_IS_CACHEABLE(db
);
1399 mutex_enter(&db
->db_mtx
);
1400 if (db
->db_state
== DB_CACHED
) {
1401 spa_t
*spa
= dn
->dn_objset
->os_spa
;
1404 * If the arc buf is compressed or encrypted, we need to
1405 * untransform it to read the data. This could happen during
1406 * the "zfs receive" of a stream which is deduplicated and
1407 * either raw or compressed. We do not need to do this if the
1408 * caller wants raw encrypted data.
1410 if (db
->db_buf
!= NULL
&& (flags
& DB_RF_NO_DECRYPT
) == 0 &&
1411 (arc_is_encrypted(db
->db_buf
) ||
1412 arc_get_compression(db
->db_buf
) != ZIO_COMPRESS_OFF
)) {
1413 zbookmark_phys_t zb
;
1415 SET_BOOKMARK(&zb
, dmu_objset_id(db
->db_objset
),
1416 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
1417 dbuf_fix_old_data(db
, spa_syncing_txg(spa
));
1418 err
= arc_untransform(db
->db_buf
, spa
, &zb
, B_FALSE
);
1419 dbuf_set_data(db
, db
->db_buf
);
1421 mutex_exit(&db
->db_mtx
);
1423 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
);
1424 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
1425 rw_exit(&dn
->dn_struct_rwlock
);
1427 DBUF_STAT_BUMP(hash_hits
);
1428 } else if (db
->db_state
== DB_UNCACHED
) {
1429 spa_t
*spa
= dn
->dn_objset
->os_spa
;
1430 boolean_t need_wait
= B_FALSE
;
1433 db
->db_blkptr
!= NULL
&& !BP_IS_HOLE(db
->db_blkptr
)) {
1434 zio
= zio_root(spa
, NULL
, NULL
, ZIO_FLAG_CANFAIL
);
1437 err
= dbuf_read_impl(db
, zio
, flags
);
1439 /* dbuf_read_impl has dropped db_mtx for us */
1441 if (!err
&& prefetch
)
1442 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
);
1444 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
1445 rw_exit(&dn
->dn_struct_rwlock
);
1447 DBUF_STAT_BUMP(hash_misses
);
1449 if (!err
&& need_wait
)
1450 err
= zio_wait(zio
);
1453 * Another reader came in while the dbuf was in flight
1454 * between UNCACHED and CACHED. Either a writer will finish
1455 * writing the buffer (sending the dbuf to CACHED) or the
1456 * first reader's request will reach the read_done callback
1457 * and send the dbuf to CACHED. Otherwise, a failure
1458 * occurred and the dbuf went to UNCACHED.
1460 mutex_exit(&db
->db_mtx
);
1462 dmu_zfetch(&dn
->dn_zfetch
, db
->db_blkid
, 1, B_TRUE
);
1463 if ((flags
& DB_RF_HAVESTRUCT
) == 0)
1464 rw_exit(&dn
->dn_struct_rwlock
);
1466 DBUF_STAT_BUMP(hash_misses
);
1468 /* Skip the wait per the caller's request. */
1469 mutex_enter(&db
->db_mtx
);
1470 if ((flags
& DB_RF_NEVERWAIT
) == 0) {
1471 while (db
->db_state
== DB_READ
||
1472 db
->db_state
== DB_FILL
) {
1473 ASSERT(db
->db_state
== DB_READ
||
1474 (flags
& DB_RF_HAVESTRUCT
) == 0);
1475 DTRACE_PROBE2(blocked__read
, dmu_buf_impl_t
*,
1477 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1479 if (db
->db_state
== DB_UNCACHED
)
1480 err
= SET_ERROR(EIO
);
1482 mutex_exit(&db
->db_mtx
);
1489 dbuf_noread(dmu_buf_impl_t
*db
)
1491 ASSERT(!refcount_is_zero(&db
->db_holds
));
1492 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1493 mutex_enter(&db
->db_mtx
);
1494 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
1495 cv_wait(&db
->db_changed
, &db
->db_mtx
);
1496 if (db
->db_state
== DB_UNCACHED
) {
1497 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1498 spa_t
*spa
= db
->db_objset
->os_spa
;
1500 ASSERT(db
->db_buf
== NULL
);
1501 ASSERT(db
->db
.db_data
== NULL
);
1502 dbuf_set_data(db
, arc_alloc_buf(spa
, db
, type
, db
->db
.db_size
));
1503 db
->db_state
= DB_FILL
;
1504 } else if (db
->db_state
== DB_NOFILL
) {
1505 dbuf_clear_data(db
);
1507 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
1509 mutex_exit(&db
->db_mtx
);
1513 dbuf_unoverride(dbuf_dirty_record_t
*dr
)
1515 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1516 blkptr_t
*bp
= &dr
->dt
.dl
.dr_overridden_by
;
1517 uint64_t txg
= dr
->dr_txg
;
1519 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1521 * This assert is valid because dmu_sync() expects to be called by
1522 * a zilog's get_data while holding a range lock. This call only
1523 * comes from dbuf_dirty() callers who must also hold a range lock.
1525 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_IN_DMU_SYNC
);
1526 ASSERT(db
->db_level
== 0);
1528 if (db
->db_blkid
== DMU_BONUS_BLKID
||
1529 dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
)
1532 ASSERT(db
->db_data_pending
!= dr
);
1534 /* free this block */
1535 if (!BP_IS_HOLE(bp
) && !dr
->dt
.dl
.dr_nopwrite
)
1536 zio_free(db
->db_objset
->os_spa
, txg
, bp
);
1538 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
1539 dr
->dt
.dl
.dr_nopwrite
= B_FALSE
;
1540 dr
->dt
.dl
.dr_raw
= B_FALSE
;
1543 * Release the already-written buffer, so we leave it in
1544 * a consistent dirty state. Note that all callers are
1545 * modifying the buffer, so they will immediately do
1546 * another (redundant) arc_release(). Therefore, leave
1547 * the buf thawed to save the effort of freezing &
1548 * immediately re-thawing it.
1550 arc_release(dr
->dt
.dl
.dr_data
, db
);
1554 * Evict (if its unreferenced) or clear (if its referenced) any level-0
1555 * data blocks in the free range, so that any future readers will find
1559 dbuf_free_range(dnode_t
*dn
, uint64_t start_blkid
, uint64_t end_blkid
,
1562 dmu_buf_impl_t
*db_search
;
1563 dmu_buf_impl_t
*db
, *db_next
;
1564 uint64_t txg
= tx
->tx_txg
;
1567 if (end_blkid
> dn
->dn_maxblkid
&&
1568 !(start_blkid
== DMU_SPILL_BLKID
|| end_blkid
== DMU_SPILL_BLKID
))
1569 end_blkid
= dn
->dn_maxblkid
;
1570 dprintf_dnode(dn
, "start=%llu end=%llu\n", start_blkid
, end_blkid
);
1572 db_search
= kmem_alloc(sizeof (dmu_buf_impl_t
), KM_SLEEP
);
1573 db_search
->db_level
= 0;
1574 db_search
->db_blkid
= start_blkid
;
1575 db_search
->db_state
= DB_SEARCH
;
1577 mutex_enter(&dn
->dn_dbufs_mtx
);
1578 db
= avl_find(&dn
->dn_dbufs
, db_search
, &where
);
1579 ASSERT3P(db
, ==, NULL
);
1581 db
= avl_nearest(&dn
->dn_dbufs
, where
, AVL_AFTER
);
1583 for (; db
!= NULL
; db
= db_next
) {
1584 db_next
= AVL_NEXT(&dn
->dn_dbufs
, db
);
1585 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1587 if (db
->db_level
!= 0 || db
->db_blkid
> end_blkid
) {
1590 ASSERT3U(db
->db_blkid
, >=, start_blkid
);
1592 /* found a level 0 buffer in the range */
1593 mutex_enter(&db
->db_mtx
);
1594 if (dbuf_undirty(db
, tx
)) {
1595 /* mutex has been dropped and dbuf destroyed */
1599 if (db
->db_state
== DB_UNCACHED
||
1600 db
->db_state
== DB_NOFILL
||
1601 db
->db_state
== DB_EVICTING
) {
1602 ASSERT(db
->db
.db_data
== NULL
);
1603 mutex_exit(&db
->db_mtx
);
1606 if (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
) {
1607 /* will be handled in dbuf_read_done or dbuf_rele */
1608 db
->db_freed_in_flight
= TRUE
;
1609 mutex_exit(&db
->db_mtx
);
1612 if (refcount_count(&db
->db_holds
) == 0) {
1617 /* The dbuf is referenced */
1619 if (db
->db_last_dirty
!= NULL
) {
1620 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
1622 if (dr
->dr_txg
== txg
) {
1624 * This buffer is "in-use", re-adjust the file
1625 * size to reflect that this buffer may
1626 * contain new data when we sync.
1628 if (db
->db_blkid
!= DMU_SPILL_BLKID
&&
1629 db
->db_blkid
> dn
->dn_maxblkid
)
1630 dn
->dn_maxblkid
= db
->db_blkid
;
1631 dbuf_unoverride(dr
);
1634 * This dbuf is not dirty in the open context.
1635 * Either uncache it (if its not referenced in
1636 * the open context) or reset its contents to
1639 dbuf_fix_old_data(db
, txg
);
1642 /* clear the contents if its cached */
1643 if (db
->db_state
== DB_CACHED
) {
1644 ASSERT(db
->db
.db_data
!= NULL
);
1645 arc_release(db
->db_buf
, db
);
1646 bzero(db
->db
.db_data
, db
->db
.db_size
);
1647 arc_buf_freeze(db
->db_buf
);
1650 mutex_exit(&db
->db_mtx
);
1653 kmem_free(db_search
, sizeof (dmu_buf_impl_t
));
1654 mutex_exit(&dn
->dn_dbufs_mtx
);
1658 dbuf_new_size(dmu_buf_impl_t
*db
, int size
, dmu_tx_t
*tx
)
1660 arc_buf_t
*buf
, *obuf
;
1661 int osize
= db
->db
.db_size
;
1662 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
1665 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
1670 /* XXX does *this* func really need the lock? */
1671 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
1674 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1675 * is OK, because there can be no other references to the db
1676 * when we are changing its size, so no concurrent DB_FILL can
1680 * XXX we should be doing a dbuf_read, checking the return
1681 * value and returning that up to our callers
1683 dmu_buf_will_dirty(&db
->db
, tx
);
1685 /* create the data buffer for the new block */
1686 buf
= arc_alloc_buf(dn
->dn_objset
->os_spa
, db
, type
, size
);
1688 /* copy old block data to the new block */
1690 bcopy(obuf
->b_data
, buf
->b_data
, MIN(osize
, size
));
1691 /* zero the remainder */
1693 bzero((uint8_t *)buf
->b_data
+ osize
, size
- osize
);
1695 mutex_enter(&db
->db_mtx
);
1696 dbuf_set_data(db
, buf
);
1697 arc_buf_destroy(obuf
, db
);
1698 db
->db
.db_size
= size
;
1700 if (db
->db_level
== 0) {
1701 ASSERT3U(db
->db_last_dirty
->dr_txg
, ==, tx
->tx_txg
);
1702 db
->db_last_dirty
->dt
.dl
.dr_data
= buf
;
1704 mutex_exit(&db
->db_mtx
);
1706 dmu_objset_willuse_space(dn
->dn_objset
, size
- osize
, tx
);
1711 dbuf_release_bp(dmu_buf_impl_t
*db
)
1713 ASSERTV(objset_t
*os
= db
->db_objset
);
1715 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os
)));
1716 ASSERT(arc_released(os
->os_phys_buf
) ||
1717 list_link_active(&os
->os_dsl_dataset
->ds_synced_link
));
1718 ASSERT(db
->db_parent
== NULL
|| arc_released(db
->db_parent
->db_buf
));
1720 (void) arc_release(db
->db_buf
, db
);
1724 * We already have a dirty record for this TXG, and we are being
1728 dbuf_redirty(dbuf_dirty_record_t
*dr
)
1730 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
1732 ASSERT(MUTEX_HELD(&db
->db_mtx
));
1734 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
) {
1736 * If this buffer has already been written out,
1737 * we now need to reset its state.
1739 dbuf_unoverride(dr
);
1740 if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
&&
1741 db
->db_state
!= DB_NOFILL
) {
1742 /* Already released on initial dirty, so just thaw. */
1743 ASSERT(arc_released(db
->db_buf
));
1744 arc_buf_thaw(db
->db_buf
);
1749 dbuf_dirty_record_t
*
1750 dbuf_dirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
1754 dbuf_dirty_record_t
**drp
, *dr
;
1755 int drop_struct_lock
= FALSE
;
1756 int txgoff
= tx
->tx_txg
& TXG_MASK
;
1758 ASSERT(tx
->tx_txg
!= 0);
1759 ASSERT(!refcount_is_zero(&db
->db_holds
));
1760 DMU_TX_DIRTY_BUF(tx
, db
);
1765 * Shouldn't dirty a regular buffer in syncing context. Private
1766 * objects may be dirtied in syncing context, but only if they
1767 * were already pre-dirtied in open context.
1770 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
) {
1771 rrw_enter(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
,
1774 ASSERT(!dmu_tx_is_syncing(tx
) ||
1775 BP_IS_HOLE(dn
->dn_objset
->os_rootbp
) ||
1776 DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
1777 dn
->dn_objset
->os_dsl_dataset
== NULL
);
1778 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
1779 rrw_exit(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
, FTAG
);
1782 * We make this assert for private objects as well, but after we
1783 * check if we're already dirty. They are allowed to re-dirty
1784 * in syncing context.
1786 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
1787 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
1788 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
1790 mutex_enter(&db
->db_mtx
);
1792 * XXX make this true for indirects too? The problem is that
1793 * transactions created with dmu_tx_create_assigned() from
1794 * syncing context don't bother holding ahead.
1796 ASSERT(db
->db_level
!= 0 ||
1797 db
->db_state
== DB_CACHED
|| db
->db_state
== DB_FILL
||
1798 db
->db_state
== DB_NOFILL
);
1800 mutex_enter(&dn
->dn_mtx
);
1802 * Don't set dirtyctx to SYNC if we're just modifying this as we
1803 * initialize the objset.
1805 if (dn
->dn_dirtyctx
== DN_UNDIRTIED
) {
1806 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
) {
1807 rrw_enter(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
,
1810 if (!BP_IS_HOLE(dn
->dn_objset
->os_rootbp
)) {
1811 dn
->dn_dirtyctx
= (dmu_tx_is_syncing(tx
) ?
1812 DN_DIRTY_SYNC
: DN_DIRTY_OPEN
);
1813 ASSERT(dn
->dn_dirtyctx_firstset
== NULL
);
1814 dn
->dn_dirtyctx_firstset
= kmem_alloc(1, KM_SLEEP
);
1816 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
) {
1817 rrw_exit(&dn
->dn_objset
->os_dsl_dataset
->ds_bp_rwlock
,
1822 if (tx
->tx_txg
> dn
->dn_dirty_txg
)
1823 dn
->dn_dirty_txg
= tx
->tx_txg
;
1824 mutex_exit(&dn
->dn_mtx
);
1826 if (db
->db_blkid
== DMU_SPILL_BLKID
)
1827 dn
->dn_have_spill
= B_TRUE
;
1830 * If this buffer is already dirty, we're done.
1832 drp
= &db
->db_last_dirty
;
1833 ASSERT(*drp
== NULL
|| (*drp
)->dr_txg
<= tx
->tx_txg
||
1834 db
->db
.db_object
== DMU_META_DNODE_OBJECT
);
1835 while ((dr
= *drp
) != NULL
&& dr
->dr_txg
> tx
->tx_txg
)
1837 if (dr
&& dr
->dr_txg
== tx
->tx_txg
) {
1841 mutex_exit(&db
->db_mtx
);
1846 * Only valid if not already dirty.
1848 ASSERT(dn
->dn_object
== 0 ||
1849 dn
->dn_dirtyctx
== DN_UNDIRTIED
|| dn
->dn_dirtyctx
==
1850 (dmu_tx_is_syncing(tx
) ? DN_DIRTY_SYNC
: DN_DIRTY_OPEN
));
1852 ASSERT3U(dn
->dn_nlevels
, >, db
->db_level
);
1855 * We should only be dirtying in syncing context if it's the
1856 * mos or we're initializing the os or it's a special object.
1857 * However, we are allowed to dirty in syncing context provided
1858 * we already dirtied it in open context. Hence we must make
1859 * this assertion only if we're not already dirty.
1862 VERIFY3U(tx
->tx_txg
, <=, spa_final_dirty_txg(os
->os_spa
));
1864 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
1865 rrw_enter(&os
->os_dsl_dataset
->ds_bp_rwlock
, RW_READER
, FTAG
);
1866 ASSERT(!dmu_tx_is_syncing(tx
) || DMU_OBJECT_IS_SPECIAL(dn
->dn_object
) ||
1867 os
->os_dsl_dataset
== NULL
|| BP_IS_HOLE(os
->os_rootbp
));
1868 if (dn
->dn_objset
->os_dsl_dataset
!= NULL
)
1869 rrw_exit(&os
->os_dsl_dataset
->ds_bp_rwlock
, FTAG
);
1871 ASSERT(db
->db
.db_size
!= 0);
1873 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
1875 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
1876 dmu_objset_willuse_space(os
, db
->db
.db_size
, tx
);
1880 * If this buffer is dirty in an old transaction group we need
1881 * to make a copy of it so that the changes we make in this
1882 * transaction group won't leak out when we sync the older txg.
1884 dr
= kmem_zalloc(sizeof (dbuf_dirty_record_t
), KM_SLEEP
);
1885 list_link_init(&dr
->dr_dirty_node
);
1886 if (db
->db_level
== 0) {
1887 void *data_old
= db
->db_buf
;
1889 if (db
->db_state
!= DB_NOFILL
) {
1890 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
1891 dbuf_fix_old_data(db
, tx
->tx_txg
);
1892 data_old
= db
->db
.db_data
;
1893 } else if (db
->db
.db_object
!= DMU_META_DNODE_OBJECT
) {
1895 * Release the data buffer from the cache so
1896 * that we can modify it without impacting
1897 * possible other users of this cached data
1898 * block. Note that indirect blocks and
1899 * private objects are not released until the
1900 * syncing state (since they are only modified
1903 arc_release(db
->db_buf
, db
);
1904 dbuf_fix_old_data(db
, tx
->tx_txg
);
1905 data_old
= db
->db_buf
;
1907 ASSERT(data_old
!= NULL
);
1909 dr
->dt
.dl
.dr_data
= data_old
;
1911 mutex_init(&dr
->dt
.di
.dr_mtx
, NULL
, MUTEX_NOLOCKDEP
, NULL
);
1912 list_create(&dr
->dt
.di
.dr_children
,
1913 sizeof (dbuf_dirty_record_t
),
1914 offsetof(dbuf_dirty_record_t
, dr_dirty_node
));
1916 if (db
->db_blkid
!= DMU_BONUS_BLKID
&& os
->os_dsl_dataset
!= NULL
)
1917 dr
->dr_accounted
= db
->db
.db_size
;
1919 dr
->dr_txg
= tx
->tx_txg
;
1924 * We could have been freed_in_flight between the dbuf_noread
1925 * and dbuf_dirty. We win, as though the dbuf_noread() had
1926 * happened after the free.
1928 if (db
->db_level
== 0 && db
->db_blkid
!= DMU_BONUS_BLKID
&&
1929 db
->db_blkid
!= DMU_SPILL_BLKID
) {
1930 mutex_enter(&dn
->dn_mtx
);
1931 if (dn
->dn_free_ranges
[txgoff
] != NULL
) {
1932 range_tree_clear(dn
->dn_free_ranges
[txgoff
],
1935 mutex_exit(&dn
->dn_mtx
);
1936 db
->db_freed_in_flight
= FALSE
;
1940 * This buffer is now part of this txg
1942 dbuf_add_ref(db
, (void *)(uintptr_t)tx
->tx_txg
);
1943 db
->db_dirtycnt
+= 1;
1944 ASSERT3U(db
->db_dirtycnt
, <=, 3);
1946 mutex_exit(&db
->db_mtx
);
1948 if (db
->db_blkid
== DMU_BONUS_BLKID
||
1949 db
->db_blkid
== DMU_SPILL_BLKID
) {
1950 mutex_enter(&dn
->dn_mtx
);
1951 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
1952 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
1953 mutex_exit(&dn
->dn_mtx
);
1954 dnode_setdirty(dn
, tx
);
1960 * The dn_struct_rwlock prevents db_blkptr from changing
1961 * due to a write from syncing context completing
1962 * while we are running, so we want to acquire it before
1963 * looking at db_blkptr.
1965 if (!RW_WRITE_HELD(&dn
->dn_struct_rwlock
)) {
1966 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
1967 drop_struct_lock
= TRUE
;
1971 * We need to hold the dn_struct_rwlock to make this assertion,
1972 * because it protects dn_phys / dn_next_nlevels from changing.
1974 ASSERT((dn
->dn_phys
->dn_nlevels
== 0 && db
->db_level
== 0) ||
1975 dn
->dn_phys
->dn_nlevels
> db
->db_level
||
1976 dn
->dn_next_nlevels
[txgoff
] > db
->db_level
||
1977 dn
->dn_next_nlevels
[(tx
->tx_txg
-1) & TXG_MASK
] > db
->db_level
||
1978 dn
->dn_next_nlevels
[(tx
->tx_txg
-2) & TXG_MASK
] > db
->db_level
);
1981 * If we are overwriting a dedup BP, then unless it is snapshotted,
1982 * when we get to syncing context we will need to decrement its
1983 * refcount in the DDT. Prefetch the relevant DDT block so that
1984 * syncing context won't have to wait for the i/o.
1986 ddt_prefetch(os
->os_spa
, db
->db_blkptr
);
1988 if (db
->db_level
== 0) {
1989 dnode_new_blkid(dn
, db
->db_blkid
, tx
, drop_struct_lock
);
1990 ASSERT(dn
->dn_maxblkid
>= db
->db_blkid
);
1993 if (db
->db_level
+1 < dn
->dn_nlevels
) {
1994 dmu_buf_impl_t
*parent
= db
->db_parent
;
1995 dbuf_dirty_record_t
*di
;
1996 int parent_held
= FALSE
;
1998 if (db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
) {
1999 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2001 parent
= dbuf_hold_level(dn
, db
->db_level
+1,
2002 db
->db_blkid
>> epbs
, FTAG
);
2003 ASSERT(parent
!= NULL
);
2006 if (drop_struct_lock
)
2007 rw_exit(&dn
->dn_struct_rwlock
);
2008 ASSERT3U(db
->db_level
+1, ==, parent
->db_level
);
2009 di
= dbuf_dirty(parent
, tx
);
2011 dbuf_rele(parent
, FTAG
);
2013 mutex_enter(&db
->db_mtx
);
2015 * Since we've dropped the mutex, it's possible that
2016 * dbuf_undirty() might have changed this out from under us.
2018 if (db
->db_last_dirty
== dr
||
2019 dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
2020 mutex_enter(&di
->dt
.di
.dr_mtx
);
2021 ASSERT3U(di
->dr_txg
, ==, tx
->tx_txg
);
2022 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
2023 list_insert_tail(&di
->dt
.di
.dr_children
, dr
);
2024 mutex_exit(&di
->dt
.di
.dr_mtx
);
2027 mutex_exit(&db
->db_mtx
);
2029 ASSERT(db
->db_level
+1 == dn
->dn_nlevels
);
2030 ASSERT(db
->db_blkid
< dn
->dn_nblkptr
);
2031 ASSERT(db
->db_parent
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
2032 mutex_enter(&dn
->dn_mtx
);
2033 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
2034 list_insert_tail(&dn
->dn_dirty_records
[txgoff
], dr
);
2035 mutex_exit(&dn
->dn_mtx
);
2036 if (drop_struct_lock
)
2037 rw_exit(&dn
->dn_struct_rwlock
);
2040 dnode_setdirty(dn
, tx
);
2046 * Undirty a buffer in the transaction group referenced by the given
2047 * transaction. Return whether this evicted the dbuf.
2050 dbuf_undirty(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
2053 uint64_t txg
= tx
->tx_txg
;
2054 dbuf_dirty_record_t
*dr
, **drp
;
2059 * Due to our use of dn_nlevels below, this can only be called
2060 * in open context, unless we are operating on the MOS.
2061 * From syncing context, dn_nlevels may be different from the
2062 * dn_nlevels used when dbuf was dirtied.
2064 ASSERT(db
->db_objset
==
2065 dmu_objset_pool(db
->db_objset
)->dp_meta_objset
||
2066 txg
!= spa_syncing_txg(dmu_objset_spa(db
->db_objset
)));
2067 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2068 ASSERT0(db
->db_level
);
2069 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2072 * If this buffer is not dirty, we're done.
2074 for (drp
= &db
->db_last_dirty
; (dr
= *drp
) != NULL
; drp
= &dr
->dr_next
)
2075 if (dr
->dr_txg
<= txg
)
2077 if (dr
== NULL
|| dr
->dr_txg
< txg
)
2079 ASSERT(dr
->dr_txg
== txg
);
2080 ASSERT(dr
->dr_dbuf
== db
);
2085 dprintf_dbuf(db
, "size=%llx\n", (u_longlong_t
)db
->db
.db_size
);
2087 ASSERT(db
->db
.db_size
!= 0);
2089 dsl_pool_undirty_space(dmu_objset_pool(dn
->dn_objset
),
2090 dr
->dr_accounted
, txg
);
2095 * Note that there are three places in dbuf_dirty()
2096 * where this dirty record may be put on a list.
2097 * Make sure to do a list_remove corresponding to
2098 * every one of those list_insert calls.
2100 if (dr
->dr_parent
) {
2101 mutex_enter(&dr
->dr_parent
->dt
.di
.dr_mtx
);
2102 list_remove(&dr
->dr_parent
->dt
.di
.dr_children
, dr
);
2103 mutex_exit(&dr
->dr_parent
->dt
.di
.dr_mtx
);
2104 } else if (db
->db_blkid
== DMU_SPILL_BLKID
||
2105 db
->db_level
+ 1 == dn
->dn_nlevels
) {
2106 ASSERT(db
->db_blkptr
== NULL
|| db
->db_parent
== dn
->dn_dbuf
);
2107 mutex_enter(&dn
->dn_mtx
);
2108 list_remove(&dn
->dn_dirty_records
[txg
& TXG_MASK
], dr
);
2109 mutex_exit(&dn
->dn_mtx
);
2113 if (db
->db_state
!= DB_NOFILL
) {
2114 dbuf_unoverride(dr
);
2116 ASSERT(db
->db_buf
!= NULL
);
2117 ASSERT(dr
->dt
.dl
.dr_data
!= NULL
);
2118 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
2119 arc_buf_destroy(dr
->dt
.dl
.dr_data
, db
);
2122 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
2124 ASSERT(db
->db_dirtycnt
> 0);
2125 db
->db_dirtycnt
-= 1;
2127 if (refcount_remove(&db
->db_holds
, (void *)(uintptr_t)txg
) == 0) {
2128 ASSERT(db
->db_state
== DB_NOFILL
|| arc_released(db
->db_buf
));
2137 dmu_buf_will_dirty_impl(dmu_buf_t
*db_fake
, int flags
, dmu_tx_t
*tx
)
2139 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2141 ASSERT(tx
->tx_txg
!= 0);
2142 ASSERT(!refcount_is_zero(&db
->db_holds
));
2145 * Quick check for dirtyness. For already dirty blocks, this
2146 * reduces runtime of this function by >90%, and overall performance
2147 * by 50% for some workloads (e.g. file deletion with indirect blocks
2150 mutex_enter(&db
->db_mtx
);
2152 dbuf_dirty_record_t
*dr
;
2153 for (dr
= db
->db_last_dirty
;
2154 dr
!= NULL
&& dr
->dr_txg
>= tx
->tx_txg
; dr
= dr
->dr_next
) {
2156 * It's possible that it is already dirty but not cached,
2157 * because there are some calls to dbuf_dirty() that don't
2158 * go through dmu_buf_will_dirty().
2160 if (dr
->dr_txg
== tx
->tx_txg
&& db
->db_state
== DB_CACHED
) {
2161 /* This dbuf is already dirty and cached. */
2163 mutex_exit(&db
->db_mtx
);
2167 mutex_exit(&db
->db_mtx
);
2170 if (RW_WRITE_HELD(&DB_DNODE(db
)->dn_struct_rwlock
))
2171 flags
|= DB_RF_HAVESTRUCT
;
2173 (void) dbuf_read(db
, NULL
, flags
);
2174 (void) dbuf_dirty(db
, tx
);
2178 dmu_buf_will_dirty(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2180 dmu_buf_will_dirty_impl(db_fake
,
2181 DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
, tx
);
2185 dmu_buf_will_not_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2187 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2189 db
->db_state
= DB_NOFILL
;
2191 dmu_buf_will_fill(db_fake
, tx
);
2195 dmu_buf_will_fill(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2197 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2199 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2200 ASSERT(tx
->tx_txg
!= 0);
2201 ASSERT(db
->db_level
== 0);
2202 ASSERT(!refcount_is_zero(&db
->db_holds
));
2204 ASSERT(db
->db
.db_object
!= DMU_META_DNODE_OBJECT
||
2205 dmu_tx_private_ok(tx
));
2208 (void) dbuf_dirty(db
, tx
);
2212 * This function is effectively the same as dmu_buf_will_dirty(), but
2213 * indicates the caller expects raw encrypted data in the db. It will
2214 * also set the raw flag on the created dirty record.
2217 dmu_buf_will_change_crypt_params(dmu_buf_t
*db_fake
, dmu_tx_t
*tx
)
2219 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
2220 dbuf_dirty_record_t
*dr
;
2222 dmu_buf_will_dirty_impl(db_fake
,
2223 DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
| DB_RF_NO_DECRYPT
, tx
);
2225 dr
= db
->db_last_dirty
;
2226 while (dr
!= NULL
&& dr
->dr_txg
> tx
->tx_txg
)
2229 ASSERT3P(dr
, !=, NULL
);
2230 ASSERT3U(dr
->dr_txg
, ==, tx
->tx_txg
);
2231 dr
->dt
.dl
.dr_raw
= B_TRUE
;
2232 db
->db_objset
->os_next_write_raw
[tx
->tx_txg
& TXG_MASK
] = B_TRUE
;
2235 #pragma weak dmu_buf_fill_done = dbuf_fill_done
2238 dbuf_fill_done(dmu_buf_impl_t
*db
, dmu_tx_t
*tx
)
2240 mutex_enter(&db
->db_mtx
);
2243 if (db
->db_state
== DB_FILL
) {
2244 if (db
->db_level
== 0 && db
->db_freed_in_flight
) {
2245 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2246 /* we were freed while filling */
2247 /* XXX dbuf_undirty? */
2248 bzero(db
->db
.db_data
, db
->db
.db_size
);
2249 db
->db_freed_in_flight
= FALSE
;
2251 db
->db_state
= DB_CACHED
;
2252 cv_broadcast(&db
->db_changed
);
2254 mutex_exit(&db
->db_mtx
);
2258 dmu_buf_write_embedded(dmu_buf_t
*dbuf
, void *data
,
2259 bp_embedded_type_t etype
, enum zio_compress comp
,
2260 int uncompressed_size
, int compressed_size
, int byteorder
,
2263 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)dbuf
;
2264 struct dirty_leaf
*dl
;
2265 dmu_object_type_t type
;
2267 if (etype
== BP_EMBEDDED_TYPE_DATA
) {
2268 ASSERT(spa_feature_is_active(dmu_objset_spa(db
->db_objset
),
2269 SPA_FEATURE_EMBEDDED_DATA
));
2273 type
= DB_DNODE(db
)->dn_type
;
2276 ASSERT0(db
->db_level
);
2277 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2279 dmu_buf_will_not_fill(dbuf
, tx
);
2281 ASSERT3U(db
->db_last_dirty
->dr_txg
, ==, tx
->tx_txg
);
2282 dl
= &db
->db_last_dirty
->dt
.dl
;
2283 encode_embedded_bp_compressed(&dl
->dr_overridden_by
,
2284 data
, comp
, uncompressed_size
, compressed_size
);
2285 BPE_SET_ETYPE(&dl
->dr_overridden_by
, etype
);
2286 BP_SET_TYPE(&dl
->dr_overridden_by
, type
);
2287 BP_SET_LEVEL(&dl
->dr_overridden_by
, 0);
2288 BP_SET_BYTEORDER(&dl
->dr_overridden_by
, byteorder
);
2290 dl
->dr_override_state
= DR_OVERRIDDEN
;
2291 dl
->dr_overridden_by
.blk_birth
= db
->db_last_dirty
->dr_txg
;
2295 * Directly assign a provided arc buf to a given dbuf if it's not referenced
2296 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
2299 dbuf_assign_arcbuf(dmu_buf_impl_t
*db
, arc_buf_t
*buf
, dmu_tx_t
*tx
)
2301 ASSERT(!refcount_is_zero(&db
->db_holds
));
2302 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
2303 ASSERT(db
->db_level
== 0);
2304 ASSERT3U(dbuf_is_metadata(db
), ==, arc_is_metadata(buf
));
2305 ASSERT(buf
!= NULL
);
2306 ASSERT(arc_buf_lsize(buf
) == db
->db
.db_size
);
2307 ASSERT(tx
->tx_txg
!= 0);
2309 arc_return_buf(buf
, db
);
2310 ASSERT(arc_released(buf
));
2312 mutex_enter(&db
->db_mtx
);
2314 while (db
->db_state
== DB_READ
|| db
->db_state
== DB_FILL
)
2315 cv_wait(&db
->db_changed
, &db
->db_mtx
);
2317 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_UNCACHED
);
2319 if (db
->db_state
== DB_CACHED
&&
2320 refcount_count(&db
->db_holds
) - 1 > db
->db_dirtycnt
) {
2322 * In practice, we will never have a case where we have an
2323 * encrypted arc buffer while additional holds exist on the
2324 * dbuf. We don't handle this here so we simply assert that
2327 ASSERT(!arc_is_encrypted(buf
));
2328 mutex_exit(&db
->db_mtx
);
2329 (void) dbuf_dirty(db
, tx
);
2330 bcopy(buf
->b_data
, db
->db
.db_data
, db
->db
.db_size
);
2331 arc_buf_destroy(buf
, db
);
2332 xuio_stat_wbuf_copied();
2336 xuio_stat_wbuf_nocopy();
2337 if (db
->db_state
== DB_CACHED
) {
2338 dbuf_dirty_record_t
*dr
= db
->db_last_dirty
;
2340 ASSERT(db
->db_buf
!= NULL
);
2341 if (dr
!= NULL
&& dr
->dr_txg
== tx
->tx_txg
) {
2342 ASSERT(dr
->dt
.dl
.dr_data
== db
->db_buf
);
2343 IMPLY(arc_is_encrypted(buf
), dr
->dt
.dl
.dr_raw
);
2345 if (!arc_released(db
->db_buf
)) {
2346 ASSERT(dr
->dt
.dl
.dr_override_state
==
2348 arc_release(db
->db_buf
, db
);
2350 dr
->dt
.dl
.dr_data
= buf
;
2351 arc_buf_destroy(db
->db_buf
, db
);
2352 } else if (dr
== NULL
|| dr
->dt
.dl
.dr_data
!= db
->db_buf
) {
2353 arc_release(db
->db_buf
, db
);
2354 arc_buf_destroy(db
->db_buf
, db
);
2358 ASSERT(db
->db_buf
== NULL
);
2359 dbuf_set_data(db
, buf
);
2360 db
->db_state
= DB_FILL
;
2361 mutex_exit(&db
->db_mtx
);
2362 (void) dbuf_dirty(db
, tx
);
2363 dmu_buf_fill_done(&db
->db
, tx
);
2367 dbuf_destroy(dmu_buf_impl_t
*db
)
2370 dmu_buf_impl_t
*parent
= db
->db_parent
;
2371 dmu_buf_impl_t
*dndb
;
2373 ASSERT(MUTEX_HELD(&db
->db_mtx
));
2374 ASSERT(refcount_is_zero(&db
->db_holds
));
2376 if (db
->db_buf
!= NULL
) {
2377 arc_buf_destroy(db
->db_buf
, db
);
2381 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
2382 int slots
= DB_DNODE(db
)->dn_num_slots
;
2383 int bonuslen
= DN_SLOTS_TO_BONUSLEN(slots
);
2384 if (db
->db
.db_data
!= NULL
) {
2385 kmem_free(db
->db
.db_data
, bonuslen
);
2386 arc_space_return(bonuslen
, ARC_SPACE_BONUS
);
2387 db
->db_state
= DB_UNCACHED
;
2391 dbuf_clear_data(db
);
2393 if (multilist_link_active(&db
->db_cache_link
)) {
2394 multilist_remove(dbuf_cache
, db
);
2395 (void) refcount_remove_many(&dbuf_cache_size
,
2396 db
->db
.db_size
, db
);
2397 DBUF_STAT_BUMPDOWN(cache_levels
[db
->db_level
]);
2398 DBUF_STAT_BUMPDOWN(cache_count
);
2399 DBUF_STAT_DECR(cache_levels_bytes
[db
->db_level
],
2403 ASSERT(db
->db_state
== DB_UNCACHED
|| db
->db_state
== DB_NOFILL
);
2404 ASSERT(db
->db_data_pending
== NULL
);
2406 db
->db_state
= DB_EVICTING
;
2407 db
->db_blkptr
= NULL
;
2410 * Now that db_state is DB_EVICTING, nobody else can find this via
2411 * the hash table. We can now drop db_mtx, which allows us to
2412 * acquire the dn_dbufs_mtx.
2414 mutex_exit(&db
->db_mtx
);
2419 if (db
->db_blkid
!= DMU_BONUS_BLKID
) {
2420 boolean_t needlock
= !MUTEX_HELD(&dn
->dn_dbufs_mtx
);
2422 mutex_enter(&dn
->dn_dbufs_mtx
);
2423 avl_remove(&dn
->dn_dbufs
, db
);
2424 atomic_dec_32(&dn
->dn_dbufs_count
);
2428 mutex_exit(&dn
->dn_dbufs_mtx
);
2430 * Decrementing the dbuf count means that the hold corresponding
2431 * to the removed dbuf is no longer discounted in dnode_move(),
2432 * so the dnode cannot be moved until after we release the hold.
2433 * The membar_producer() ensures visibility of the decremented
2434 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
2438 db
->db_dnode_handle
= NULL
;
2440 dbuf_hash_remove(db
);
2445 ASSERT(refcount_is_zero(&db
->db_holds
));
2447 db
->db_parent
= NULL
;
2449 ASSERT(db
->db_buf
== NULL
);
2450 ASSERT(db
->db
.db_data
== NULL
);
2451 ASSERT(db
->db_hash_next
== NULL
);
2452 ASSERT(db
->db_blkptr
== NULL
);
2453 ASSERT(db
->db_data_pending
== NULL
);
2454 ASSERT(!multilist_link_active(&db
->db_cache_link
));
2456 kmem_cache_free(dbuf_kmem_cache
, db
);
2457 arc_space_return(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
2460 * If this dbuf is referenced from an indirect dbuf,
2461 * decrement the ref count on the indirect dbuf.
2463 if (parent
&& parent
!= dndb
)
2464 dbuf_rele(parent
, db
);
2468 * Note: While bpp will always be updated if the function returns success,
2469 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
2470 * this happens when the dnode is the meta-dnode, or {user|group|project}used
2473 __attribute__((always_inline
))
2475 dbuf_findbp(dnode_t
*dn
, int level
, uint64_t blkid
, int fail_sparse
,
2476 dmu_buf_impl_t
**parentp
, blkptr_t
**bpp
, struct dbuf_hold_impl_data
*dh
)
2481 ASSERT(blkid
!= DMU_BONUS_BLKID
);
2483 if (blkid
== DMU_SPILL_BLKID
) {
2484 mutex_enter(&dn
->dn_mtx
);
2485 if (dn
->dn_have_spill
&&
2486 (dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
))
2487 *bpp
= DN_SPILL_BLKPTR(dn
->dn_phys
);
2490 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
2491 *parentp
= dn
->dn_dbuf
;
2492 mutex_exit(&dn
->dn_mtx
);
2497 (dn
->dn_phys
->dn_nlevels
== 0) ? 1 : dn
->dn_phys
->dn_nlevels
;
2498 int epbs
= dn
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2500 ASSERT3U(level
* epbs
, <, 64);
2501 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2503 * This assertion shouldn't trip as long as the max indirect block size
2504 * is less than 1M. The reason for this is that up to that point,
2505 * the number of levels required to address an entire object with blocks
2506 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In
2507 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55
2508 * (i.e. we can address the entire object), objects will all use at most
2509 * N-1 levels and the assertion won't overflow. However, once epbs is
2510 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be
2511 * enough to address an entire object, so objects will have 5 levels,
2512 * but then this assertion will overflow.
2514 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we
2515 * need to redo this logic to handle overflows.
2517 ASSERT(level
>= nlevels
||
2518 ((nlevels
- level
- 1) * epbs
) +
2519 highbit64(dn
->dn_phys
->dn_nblkptr
) <= 64);
2520 if (level
>= nlevels
||
2521 blkid
>= ((uint64_t)dn
->dn_phys
->dn_nblkptr
<<
2522 ((nlevels
- level
- 1) * epbs
)) ||
2524 blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))) {
2525 /* the buffer has no parent yet */
2526 return (SET_ERROR(ENOENT
));
2527 } else if (level
< nlevels
-1) {
2528 /* this block is referenced from an indirect block */
2531 err
= dbuf_hold_impl(dn
, level
+1,
2532 blkid
>> epbs
, fail_sparse
, FALSE
, NULL
, parentp
);
2534 __dbuf_hold_impl_init(dh
+ 1, dn
, dh
->dh_level
+ 1,
2535 blkid
>> epbs
, fail_sparse
, FALSE
, NULL
,
2536 parentp
, dh
->dh_depth
+ 1);
2537 err
= __dbuf_hold_impl(dh
+ 1);
2541 err
= dbuf_read(*parentp
, NULL
,
2542 (DB_RF_HAVESTRUCT
| DB_RF_NOPREFETCH
| DB_RF_CANFAIL
));
2544 dbuf_rele(*parentp
, NULL
);
2548 *bpp
= ((blkptr_t
*)(*parentp
)->db
.db_data
) +
2549 (blkid
& ((1ULL << epbs
) - 1));
2550 if (blkid
> (dn
->dn_phys
->dn_maxblkid
>> (level
* epbs
)))
2551 ASSERT(BP_IS_HOLE(*bpp
));
2554 /* the block is referenced from the dnode */
2555 ASSERT3U(level
, ==, nlevels
-1);
2556 ASSERT(dn
->dn_phys
->dn_nblkptr
== 0 ||
2557 blkid
< dn
->dn_phys
->dn_nblkptr
);
2559 dbuf_add_ref(dn
->dn_dbuf
, NULL
);
2560 *parentp
= dn
->dn_dbuf
;
2562 *bpp
= &dn
->dn_phys
->dn_blkptr
[blkid
];
2567 static dmu_buf_impl_t
*
2568 dbuf_create(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
2569 dmu_buf_impl_t
*parent
, blkptr_t
*blkptr
)
2571 objset_t
*os
= dn
->dn_objset
;
2572 dmu_buf_impl_t
*db
, *odb
;
2574 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2575 ASSERT(dn
->dn_type
!= DMU_OT_NONE
);
2577 db
= kmem_cache_alloc(dbuf_kmem_cache
, KM_SLEEP
);
2580 db
->db
.db_object
= dn
->dn_object
;
2581 db
->db_level
= level
;
2582 db
->db_blkid
= blkid
;
2583 db
->db_last_dirty
= NULL
;
2584 db
->db_dirtycnt
= 0;
2585 db
->db_dnode_handle
= dn
->dn_handle
;
2586 db
->db_parent
= parent
;
2587 db
->db_blkptr
= blkptr
;
2590 db
->db_user_immediate_evict
= FALSE
;
2591 db
->db_freed_in_flight
= FALSE
;
2592 db
->db_pending_evict
= FALSE
;
2594 if (blkid
== DMU_BONUS_BLKID
) {
2595 ASSERT3P(parent
, ==, dn
->dn_dbuf
);
2596 db
->db
.db_size
= DN_SLOTS_TO_BONUSLEN(dn
->dn_num_slots
) -
2597 (dn
->dn_nblkptr
-1) * sizeof (blkptr_t
);
2598 ASSERT3U(db
->db
.db_size
, >=, dn
->dn_bonuslen
);
2599 db
->db
.db_offset
= DMU_BONUS_BLKID
;
2600 db
->db_state
= DB_UNCACHED
;
2601 /* the bonus dbuf is not placed in the hash table */
2602 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
2604 } else if (blkid
== DMU_SPILL_BLKID
) {
2605 db
->db
.db_size
= (blkptr
!= NULL
) ?
2606 BP_GET_LSIZE(blkptr
) : SPA_MINBLOCKSIZE
;
2607 db
->db
.db_offset
= 0;
2610 db
->db_level
? 1 << dn
->dn_indblkshift
: dn
->dn_datablksz
;
2611 db
->db
.db_size
= blocksize
;
2612 db
->db
.db_offset
= db
->db_blkid
* blocksize
;
2616 * Hold the dn_dbufs_mtx while we get the new dbuf
2617 * in the hash table *and* added to the dbufs list.
2618 * This prevents a possible deadlock with someone
2619 * trying to look up this dbuf before its added to the
2622 mutex_enter(&dn
->dn_dbufs_mtx
);
2623 db
->db_state
= DB_EVICTING
;
2624 if ((odb
= dbuf_hash_insert(db
)) != NULL
) {
2625 /* someone else inserted it first */
2626 kmem_cache_free(dbuf_kmem_cache
, db
);
2627 mutex_exit(&dn
->dn_dbufs_mtx
);
2628 DBUF_STAT_BUMP(hash_insert_race
);
2631 avl_add(&dn
->dn_dbufs
, db
);
2633 db
->db_state
= DB_UNCACHED
;
2634 mutex_exit(&dn
->dn_dbufs_mtx
);
2635 arc_space_consume(sizeof (dmu_buf_impl_t
), ARC_SPACE_DBUF
);
2637 if (parent
&& parent
!= dn
->dn_dbuf
)
2638 dbuf_add_ref(parent
, db
);
2640 ASSERT(dn
->dn_object
== DMU_META_DNODE_OBJECT
||
2641 refcount_count(&dn
->dn_holds
) > 0);
2642 (void) refcount_add(&dn
->dn_holds
, db
);
2643 atomic_inc_32(&dn
->dn_dbufs_count
);
2645 dprintf_dbuf(db
, "db=%p\n", db
);
2650 typedef struct dbuf_prefetch_arg
{
2651 spa_t
*dpa_spa
; /* The spa to issue the prefetch in. */
2652 zbookmark_phys_t dpa_zb
; /* The target block to prefetch. */
2653 int dpa_epbs
; /* Entries (blkptr_t's) Per Block Shift. */
2654 int dpa_curlevel
; /* The current level that we're reading */
2655 dnode_t
*dpa_dnode
; /* The dnode associated with the prefetch */
2656 zio_priority_t dpa_prio
; /* The priority I/Os should be issued at. */
2657 zio_t
*dpa_zio
; /* The parent zio_t for all prefetches. */
2658 arc_flags_t dpa_aflags
; /* Flags to pass to the final prefetch. */
2659 } dbuf_prefetch_arg_t
;
2662 * Actually issue the prefetch read for the block given.
2665 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t
*dpa
, blkptr_t
*bp
)
2667 if (BP_IS_HOLE(bp
) || BP_IS_EMBEDDED(bp
))
2670 int zio_flags
= ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
;
2671 arc_flags_t aflags
=
2672 dpa
->dpa_aflags
| ARC_FLAG_NOWAIT
| ARC_FLAG_PREFETCH
;
2674 /* dnodes are always read as raw and then converted later */
2675 if (BP_GET_TYPE(bp
) == DMU_OT_DNODE
&& BP_IS_PROTECTED(bp
) &&
2676 dpa
->dpa_curlevel
== 0)
2677 zio_flags
|= ZIO_FLAG_RAW
;
2679 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
2680 ASSERT3U(dpa
->dpa_curlevel
, ==, dpa
->dpa_zb
.zb_level
);
2681 ASSERT(dpa
->dpa_zio
!= NULL
);
2682 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
, bp
, NULL
, NULL
,
2683 dpa
->dpa_prio
, zio_flags
, &aflags
, &dpa
->dpa_zb
);
2687 * Called when an indirect block above our prefetch target is read in. This
2688 * will either read in the next indirect block down the tree or issue the actual
2689 * prefetch if the next block down is our target.
2692 dbuf_prefetch_indirect_done(zio_t
*zio
, const zbookmark_phys_t
*zb
,
2693 const blkptr_t
*iobp
, arc_buf_t
*abuf
, void *private)
2695 dbuf_prefetch_arg_t
*dpa
= private;
2697 ASSERT3S(dpa
->dpa_zb
.zb_level
, <, dpa
->dpa_curlevel
);
2698 ASSERT3S(dpa
->dpa_curlevel
, >, 0);
2701 * The dpa_dnode is only valid if we are called with a NULL
2702 * zio. This indicates that the arc_read() returned without
2703 * first calling zio_read() to issue a physical read. Once
2704 * a physical read is made the dpa_dnode must be invalidated
2705 * as the locks guarding it may have been dropped. If the
2706 * dpa_dnode is still valid, then we want to add it to the dbuf
2707 * cache. To do so, we must hold the dbuf associated with the block
2708 * we just prefetched, read its contents so that we associate it
2709 * with an arc_buf_t, and then release it.
2712 ASSERT3S(BP_GET_LEVEL(zio
->io_bp
), ==, dpa
->dpa_curlevel
);
2713 if (zio
->io_flags
& ZIO_FLAG_RAW_COMPRESS
) {
2714 ASSERT3U(BP_GET_PSIZE(zio
->io_bp
), ==, zio
->io_size
);
2716 ASSERT3U(BP_GET_LSIZE(zio
->io_bp
), ==, zio
->io_size
);
2718 ASSERT3P(zio
->io_spa
, ==, dpa
->dpa_spa
);
2720 dpa
->dpa_dnode
= NULL
;
2721 } else if (dpa
->dpa_dnode
!= NULL
) {
2722 uint64_t curblkid
= dpa
->dpa_zb
.zb_blkid
>>
2723 (dpa
->dpa_epbs
* (dpa
->dpa_curlevel
-
2724 dpa
->dpa_zb
.zb_level
));
2725 dmu_buf_impl_t
*db
= dbuf_hold_level(dpa
->dpa_dnode
,
2726 dpa
->dpa_curlevel
, curblkid
, FTAG
);
2727 (void) dbuf_read(db
, NULL
,
2728 DB_RF_MUST_SUCCEED
| DB_RF_NOPREFETCH
| DB_RF_HAVESTRUCT
);
2729 dbuf_rele(db
, FTAG
);
2733 kmem_free(dpa
, sizeof (*dpa
));
2737 dpa
->dpa_curlevel
--;
2738 uint64_t nextblkid
= dpa
->dpa_zb
.zb_blkid
>>
2739 (dpa
->dpa_epbs
* (dpa
->dpa_curlevel
- dpa
->dpa_zb
.zb_level
));
2740 blkptr_t
*bp
= ((blkptr_t
*)abuf
->b_data
) +
2741 P2PHASE(nextblkid
, 1ULL << dpa
->dpa_epbs
);
2743 if (BP_IS_HOLE(bp
)) {
2744 kmem_free(dpa
, sizeof (*dpa
));
2745 } else if (dpa
->dpa_curlevel
== dpa
->dpa_zb
.zb_level
) {
2746 ASSERT3U(nextblkid
, ==, dpa
->dpa_zb
.zb_blkid
);
2747 dbuf_issue_final_prefetch(dpa
, bp
);
2748 kmem_free(dpa
, sizeof (*dpa
));
2750 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
2751 zbookmark_phys_t zb
;
2753 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2754 if (dpa
->dpa_aflags
& ARC_FLAG_L2CACHE
)
2755 iter_aflags
|= ARC_FLAG_L2CACHE
;
2757 ASSERT3U(dpa
->dpa_curlevel
, ==, BP_GET_LEVEL(bp
));
2759 SET_BOOKMARK(&zb
, dpa
->dpa_zb
.zb_objset
,
2760 dpa
->dpa_zb
.zb_object
, dpa
->dpa_curlevel
, nextblkid
);
2762 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
2763 bp
, dbuf_prefetch_indirect_done
, dpa
, dpa
->dpa_prio
,
2764 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2768 arc_buf_destroy(abuf
, private);
2772 * Issue prefetch reads for the given block on the given level. If the indirect
2773 * blocks above that block are not in memory, we will read them in
2774 * asynchronously. As a result, this call never blocks waiting for a read to
2775 * complete. Note that the prefetch might fail if the dataset is encrypted and
2776 * the encryption key is unmapped before the IO completes.
2779 dbuf_prefetch(dnode_t
*dn
, int64_t level
, uint64_t blkid
, zio_priority_t prio
,
2783 int epbs
, nlevels
, curlevel
;
2786 ASSERT(blkid
!= DMU_BONUS_BLKID
);
2787 ASSERT(RW_LOCK_HELD(&dn
->dn_struct_rwlock
));
2789 if (blkid
> dn
->dn_maxblkid
)
2792 if (dnode_block_freed(dn
, blkid
))
2796 * This dnode hasn't been written to disk yet, so there's nothing to
2799 nlevels
= dn
->dn_phys
->dn_nlevels
;
2800 if (level
>= nlevels
|| dn
->dn_phys
->dn_nblkptr
== 0)
2803 epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
2804 if (dn
->dn_phys
->dn_maxblkid
< blkid
<< (epbs
* level
))
2807 dmu_buf_impl_t
*db
= dbuf_find(dn
->dn_objset
, dn
->dn_object
,
2810 mutex_exit(&db
->db_mtx
);
2812 * This dbuf already exists. It is either CACHED, or
2813 * (we assume) about to be read or filled.
2819 * Find the closest ancestor (indirect block) of the target block
2820 * that is present in the cache. In this indirect block, we will
2821 * find the bp that is at curlevel, curblkid.
2825 while (curlevel
< nlevels
- 1) {
2826 int parent_level
= curlevel
+ 1;
2827 uint64_t parent_blkid
= curblkid
>> epbs
;
2830 if (dbuf_hold_impl(dn
, parent_level
, parent_blkid
,
2831 FALSE
, TRUE
, FTAG
, &db
) == 0) {
2832 blkptr_t
*bpp
= db
->db_buf
->b_data
;
2833 bp
= bpp
[P2PHASE(curblkid
, 1 << epbs
)];
2834 dbuf_rele(db
, FTAG
);
2838 curlevel
= parent_level
;
2839 curblkid
= parent_blkid
;
2842 if (curlevel
== nlevels
- 1) {
2843 /* No cached indirect blocks found. */
2844 ASSERT3U(curblkid
, <, dn
->dn_phys
->dn_nblkptr
);
2845 bp
= dn
->dn_phys
->dn_blkptr
[curblkid
];
2847 if (BP_IS_HOLE(&bp
))
2850 ASSERT3U(curlevel
, ==, BP_GET_LEVEL(&bp
));
2852 zio_t
*pio
= zio_root(dmu_objset_spa(dn
->dn_objset
), NULL
, NULL
,
2855 dbuf_prefetch_arg_t
*dpa
= kmem_zalloc(sizeof (*dpa
), KM_SLEEP
);
2856 dsl_dataset_t
*ds
= dn
->dn_objset
->os_dsl_dataset
;
2857 SET_BOOKMARK(&dpa
->dpa_zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
2858 dn
->dn_object
, level
, blkid
);
2859 dpa
->dpa_curlevel
= curlevel
;
2860 dpa
->dpa_prio
= prio
;
2861 dpa
->dpa_aflags
= aflags
;
2862 dpa
->dpa_spa
= dn
->dn_objset
->os_spa
;
2863 dpa
->dpa_dnode
= dn
;
2864 dpa
->dpa_epbs
= epbs
;
2867 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2868 if (DNODE_LEVEL_IS_L2CACHEABLE(dn
, level
))
2869 dpa
->dpa_aflags
|= ARC_FLAG_L2CACHE
;
2872 * If we have the indirect just above us, no need to do the asynchronous
2873 * prefetch chain; we'll just run the last step ourselves. If we're at
2874 * a higher level, though, we want to issue the prefetches for all the
2875 * indirect blocks asynchronously, so we can go on with whatever we were
2878 if (curlevel
== level
) {
2879 ASSERT3U(curblkid
, ==, blkid
);
2880 dbuf_issue_final_prefetch(dpa
, &bp
);
2881 kmem_free(dpa
, sizeof (*dpa
));
2883 arc_flags_t iter_aflags
= ARC_FLAG_NOWAIT
;
2884 zbookmark_phys_t zb
;
2886 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2887 if (DNODE_LEVEL_IS_L2CACHEABLE(dn
, level
))
2888 iter_aflags
|= ARC_FLAG_L2CACHE
;
2890 SET_BOOKMARK(&zb
, ds
!= NULL
? ds
->ds_object
: DMU_META_OBJSET
,
2891 dn
->dn_object
, curlevel
, curblkid
);
2892 (void) arc_read(dpa
->dpa_zio
, dpa
->dpa_spa
,
2893 &bp
, dbuf_prefetch_indirect_done
, dpa
, prio
,
2894 ZIO_FLAG_CANFAIL
| ZIO_FLAG_SPECULATIVE
,
2898 * We use pio here instead of dpa_zio since it's possible that
2899 * dpa may have already been freed.
2904 #define DBUF_HOLD_IMPL_MAX_DEPTH 20
2907 * Helper function for __dbuf_hold_impl() to copy a buffer. Handles
2908 * the case of encrypted, compressed and uncompressed buffers by
2909 * allocating the new buffer, respectively, with arc_alloc_raw_buf(),
2910 * arc_alloc_compressed_buf() or arc_alloc_buf().*
2912 * NOTE: Declared noinline to avoid stack bloat in __dbuf_hold_impl().
2914 noinline
static void
2915 dbuf_hold_copy(struct dbuf_hold_impl_data
*dh
)
2917 dnode_t
*dn
= dh
->dh_dn
;
2918 dmu_buf_impl_t
*db
= dh
->dh_db
;
2919 dbuf_dirty_record_t
*dr
= dh
->dh_dr
;
2920 arc_buf_t
*data
= dr
->dt
.dl
.dr_data
;
2922 enum zio_compress compress_type
= arc_get_compression(data
);
2924 if (arc_is_encrypted(data
)) {
2925 boolean_t byteorder
;
2926 uint8_t salt
[ZIO_DATA_SALT_LEN
];
2927 uint8_t iv
[ZIO_DATA_IV_LEN
];
2928 uint8_t mac
[ZIO_DATA_MAC_LEN
];
2930 arc_get_raw_params(data
, &byteorder
, salt
, iv
, mac
);
2931 dbuf_set_data(db
, arc_alloc_raw_buf(dn
->dn_objset
->os_spa
, db
,
2932 dmu_objset_id(dn
->dn_objset
), byteorder
, salt
, iv
, mac
,
2933 dn
->dn_type
, arc_buf_size(data
), arc_buf_lsize(data
),
2935 } else if (compress_type
!= ZIO_COMPRESS_OFF
) {
2936 dbuf_set_data(db
, arc_alloc_compressed_buf(
2937 dn
->dn_objset
->os_spa
, db
, arc_buf_size(data
),
2938 arc_buf_lsize(data
), compress_type
));
2940 dbuf_set_data(db
, arc_alloc_buf(dn
->dn_objset
->os_spa
, db
,
2941 DBUF_GET_BUFC_TYPE(db
), db
->db
.db_size
));
2944 bcopy(data
->b_data
, db
->db
.db_data
, arc_buf_size(data
));
2948 * Returns with db_holds incremented, and db_mtx not held.
2949 * Note: dn_struct_rwlock must be held.
2952 __dbuf_hold_impl(struct dbuf_hold_impl_data
*dh
)
2954 ASSERT3S(dh
->dh_depth
, <, DBUF_HOLD_IMPL_MAX_DEPTH
);
2955 dh
->dh_parent
= NULL
;
2957 ASSERT(dh
->dh_blkid
!= DMU_BONUS_BLKID
);
2958 ASSERT(RW_LOCK_HELD(&dh
->dh_dn
->dn_struct_rwlock
));
2959 ASSERT3U(dh
->dh_dn
->dn_nlevels
, >, dh
->dh_level
);
2961 *(dh
->dh_dbp
) = NULL
;
2963 /* dbuf_find() returns with db_mtx held */
2964 dh
->dh_db
= dbuf_find(dh
->dh_dn
->dn_objset
, dh
->dh_dn
->dn_object
,
2965 dh
->dh_level
, dh
->dh_blkid
);
2967 if (dh
->dh_db
== NULL
) {
2970 if (dh
->dh_fail_uncached
)
2971 return (SET_ERROR(ENOENT
));
2973 ASSERT3P(dh
->dh_parent
, ==, NULL
);
2974 dh
->dh_err
= dbuf_findbp(dh
->dh_dn
, dh
->dh_level
, dh
->dh_blkid
,
2975 dh
->dh_fail_sparse
, &dh
->dh_parent
, &dh
->dh_bp
, dh
);
2976 if (dh
->dh_fail_sparse
) {
2977 if (dh
->dh_err
== 0 &&
2978 dh
->dh_bp
&& BP_IS_HOLE(dh
->dh_bp
))
2979 dh
->dh_err
= SET_ERROR(ENOENT
);
2982 dbuf_rele(dh
->dh_parent
, NULL
);
2983 return (dh
->dh_err
);
2986 if (dh
->dh_err
&& dh
->dh_err
!= ENOENT
)
2987 return (dh
->dh_err
);
2988 dh
->dh_db
= dbuf_create(dh
->dh_dn
, dh
->dh_level
, dh
->dh_blkid
,
2989 dh
->dh_parent
, dh
->dh_bp
);
2992 if (dh
->dh_fail_uncached
&& dh
->dh_db
->db_state
!= DB_CACHED
) {
2993 mutex_exit(&dh
->dh_db
->db_mtx
);
2994 return (SET_ERROR(ENOENT
));
2997 if (dh
->dh_db
->db_buf
!= NULL
) {
2998 arc_buf_access(dh
->dh_db
->db_buf
);
2999 ASSERT3P(dh
->dh_db
->db
.db_data
, ==, dh
->dh_db
->db_buf
->b_data
);
3002 ASSERT(dh
->dh_db
->db_buf
== NULL
|| arc_referenced(dh
->dh_db
->db_buf
));
3005 * If this buffer is currently syncing out, and we are are
3006 * still referencing it from db_data, we need to make a copy
3007 * of it in case we decide we want to dirty it again in this txg.
3009 if (dh
->dh_db
->db_level
== 0 &&
3010 dh
->dh_db
->db_blkid
!= DMU_BONUS_BLKID
&&
3011 dh
->dh_dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
3012 dh
->dh_db
->db_state
== DB_CACHED
&& dh
->dh_db
->db_data_pending
) {
3013 dh
->dh_dr
= dh
->dh_db
->db_data_pending
;
3014 if (dh
->dh_dr
->dt
.dl
.dr_data
== dh
->dh_db
->db_buf
)
3018 if (multilist_link_active(&dh
->dh_db
->db_cache_link
)) {
3019 ASSERT(refcount_is_zero(&dh
->dh_db
->db_holds
));
3020 multilist_remove(dbuf_cache
, dh
->dh_db
);
3021 (void) refcount_remove_many(&dbuf_cache_size
,
3022 dh
->dh_db
->db
.db_size
, dh
->dh_db
);
3023 DBUF_STAT_BUMPDOWN(cache_levels
[dh
->dh_db
->db_level
]);
3024 DBUF_STAT_BUMPDOWN(cache_count
);
3025 DBUF_STAT_DECR(cache_levels_bytes
[dh
->dh_db
->db_level
],
3026 dh
->dh_db
->db
.db_size
);
3028 (void) refcount_add(&dh
->dh_db
->db_holds
, dh
->dh_tag
);
3029 DBUF_VERIFY(dh
->dh_db
);
3030 mutex_exit(&dh
->dh_db
->db_mtx
);
3032 /* NOTE: we can't rele the parent until after we drop the db_mtx */
3034 dbuf_rele(dh
->dh_parent
, NULL
);
3036 ASSERT3P(DB_DNODE(dh
->dh_db
), ==, dh
->dh_dn
);
3037 ASSERT3U(dh
->dh_db
->db_blkid
, ==, dh
->dh_blkid
);
3038 ASSERT3U(dh
->dh_db
->db_level
, ==, dh
->dh_level
);
3039 *(dh
->dh_dbp
) = dh
->dh_db
;
3045 * The following code preserves the recursive function dbuf_hold_impl()
3046 * but moves the local variables AND function arguments to the heap to
3047 * minimize the stack frame size. Enough space is initially allocated
3048 * on the stack for 20 levels of recursion.
3051 dbuf_hold_impl(dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
3052 boolean_t fail_sparse
, boolean_t fail_uncached
,
3053 void *tag
, dmu_buf_impl_t
**dbp
)
3055 struct dbuf_hold_impl_data
*dh
;
3058 dh
= kmem_alloc(sizeof (struct dbuf_hold_impl_data
) *
3059 DBUF_HOLD_IMPL_MAX_DEPTH
, KM_SLEEP
);
3060 __dbuf_hold_impl_init(dh
, dn
, level
, blkid
, fail_sparse
,
3061 fail_uncached
, tag
, dbp
, 0);
3063 error
= __dbuf_hold_impl(dh
);
3065 kmem_free(dh
, sizeof (struct dbuf_hold_impl_data
) *
3066 DBUF_HOLD_IMPL_MAX_DEPTH
);
3072 __dbuf_hold_impl_init(struct dbuf_hold_impl_data
*dh
,
3073 dnode_t
*dn
, uint8_t level
, uint64_t blkid
,
3074 boolean_t fail_sparse
, boolean_t fail_uncached
,
3075 void *tag
, dmu_buf_impl_t
**dbp
, int depth
)
3078 dh
->dh_level
= level
;
3079 dh
->dh_blkid
= blkid
;
3081 dh
->dh_fail_sparse
= fail_sparse
;
3082 dh
->dh_fail_uncached
= fail_uncached
;
3088 dh
->dh_parent
= NULL
;
3093 dh
->dh_depth
= depth
;
3097 dbuf_hold(dnode_t
*dn
, uint64_t blkid
, void *tag
)
3099 return (dbuf_hold_level(dn
, 0, blkid
, tag
));
3103 dbuf_hold_level(dnode_t
*dn
, int level
, uint64_t blkid
, void *tag
)
3106 int err
= dbuf_hold_impl(dn
, level
, blkid
, FALSE
, FALSE
, tag
, &db
);
3107 return (err
? NULL
: db
);
3111 dbuf_create_bonus(dnode_t
*dn
)
3113 ASSERT(RW_WRITE_HELD(&dn
->dn_struct_rwlock
));
3115 ASSERT(dn
->dn_bonus
== NULL
);
3116 dn
->dn_bonus
= dbuf_create(dn
, 0, DMU_BONUS_BLKID
, dn
->dn_dbuf
, NULL
);
3120 dbuf_spill_set_blksz(dmu_buf_t
*db_fake
, uint64_t blksz
, dmu_tx_t
*tx
)
3122 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3125 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
3126 return (SET_ERROR(ENOTSUP
));
3128 blksz
= SPA_MINBLOCKSIZE
;
3129 ASSERT3U(blksz
, <=, spa_maxblocksize(dmu_objset_spa(db
->db_objset
)));
3130 blksz
= P2ROUNDUP(blksz
, SPA_MINBLOCKSIZE
);
3134 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
3135 dbuf_new_size(db
, blksz
, tx
);
3136 rw_exit(&dn
->dn_struct_rwlock
);
3143 dbuf_rm_spill(dnode_t
*dn
, dmu_tx_t
*tx
)
3145 dbuf_free_range(dn
, DMU_SPILL_BLKID
, DMU_SPILL_BLKID
, tx
);
3148 #pragma weak dmu_buf_add_ref = dbuf_add_ref
3150 dbuf_add_ref(dmu_buf_impl_t
*db
, void *tag
)
3152 int64_t holds
= refcount_add(&db
->db_holds
, tag
);
3153 VERIFY3S(holds
, >, 1);
3156 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
3158 dbuf_try_add_ref(dmu_buf_t
*db_fake
, objset_t
*os
, uint64_t obj
, uint64_t blkid
,
3161 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3162 dmu_buf_impl_t
*found_db
;
3163 boolean_t result
= B_FALSE
;
3165 if (blkid
== DMU_BONUS_BLKID
)
3166 found_db
= dbuf_find_bonus(os
, obj
);
3168 found_db
= dbuf_find(os
, obj
, 0, blkid
);
3170 if (found_db
!= NULL
) {
3171 if (db
== found_db
&& dbuf_refcount(db
) > db
->db_dirtycnt
) {
3172 (void) refcount_add(&db
->db_holds
, tag
);
3175 mutex_exit(&found_db
->db_mtx
);
3181 * If you call dbuf_rele() you had better not be referencing the dnode handle
3182 * unless you have some other direct or indirect hold on the dnode. (An indirect
3183 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
3184 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
3185 * dnode's parent dbuf evicting its dnode handles.
3188 dbuf_rele(dmu_buf_impl_t
*db
, void *tag
)
3190 mutex_enter(&db
->db_mtx
);
3191 dbuf_rele_and_unlock(db
, tag
);
3195 dmu_buf_rele(dmu_buf_t
*db
, void *tag
)
3197 dbuf_rele((dmu_buf_impl_t
*)db
, tag
);
3201 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
3202 * db_dirtycnt and db_holds to be updated atomically.
3205 dbuf_rele_and_unlock(dmu_buf_impl_t
*db
, void *tag
)
3209 ASSERT(MUTEX_HELD(&db
->db_mtx
));
3213 * Remove the reference to the dbuf before removing its hold on the
3214 * dnode so we can guarantee in dnode_move() that a referenced bonus
3215 * buffer has a corresponding dnode hold.
3217 holds
= refcount_remove(&db
->db_holds
, tag
);
3221 * We can't freeze indirects if there is a possibility that they
3222 * may be modified in the current syncing context.
3224 if (db
->db_buf
!= NULL
&&
3225 holds
== (db
->db_level
== 0 ? db
->db_dirtycnt
: 0)) {
3226 arc_buf_freeze(db
->db_buf
);
3229 if (holds
== db
->db_dirtycnt
&&
3230 db
->db_level
== 0 && db
->db_user_immediate_evict
)
3231 dbuf_evict_user(db
);
3234 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
3236 boolean_t evict_dbuf
= db
->db_pending_evict
;
3239 * If the dnode moves here, we cannot cross this
3240 * barrier until the move completes.
3245 atomic_dec_32(&dn
->dn_dbufs_count
);
3248 * Decrementing the dbuf count means that the bonus
3249 * buffer's dnode hold is no longer discounted in
3250 * dnode_move(). The dnode cannot move until after
3251 * the dnode_rele() below.
3256 * Do not reference db after its lock is dropped.
3257 * Another thread may evict it.
3259 mutex_exit(&db
->db_mtx
);
3262 dnode_evict_bonus(dn
);
3265 } else if (db
->db_buf
== NULL
) {
3267 * This is a special case: we never associated this
3268 * dbuf with any data allocated from the ARC.
3270 ASSERT(db
->db_state
== DB_UNCACHED
||
3271 db
->db_state
== DB_NOFILL
);
3273 } else if (arc_released(db
->db_buf
)) {
3275 * This dbuf has anonymous data associated with it.
3279 boolean_t do_arc_evict
= B_FALSE
;
3281 spa_t
*spa
= dmu_objset_spa(db
->db_objset
);
3283 if (!DBUF_IS_CACHEABLE(db
) &&
3284 db
->db_blkptr
!= NULL
&&
3285 !BP_IS_HOLE(db
->db_blkptr
) &&
3286 !BP_IS_EMBEDDED(db
->db_blkptr
)) {
3287 do_arc_evict
= B_TRUE
;
3288 bp
= *db
->db_blkptr
;
3291 if (!DBUF_IS_CACHEABLE(db
) ||
3292 db
->db_pending_evict
) {
3294 } else if (!multilist_link_active(&db
->db_cache_link
)) {
3295 multilist_insert(dbuf_cache
, db
);
3296 (void) refcount_add_many(&dbuf_cache_size
,
3297 db
->db
.db_size
, db
);
3298 DBUF_STAT_BUMP(cache_levels
[db
->db_level
]);
3299 DBUF_STAT_BUMP(cache_count
);
3300 DBUF_STAT_INCR(cache_levels_bytes
[db
->db_level
],
3302 DBUF_STAT_MAX(cache_size_bytes_max
,
3303 refcount_count(&dbuf_cache_size
));
3304 mutex_exit(&db
->db_mtx
);
3306 dbuf_evict_notify();
3310 arc_freed(spa
, &bp
);
3313 mutex_exit(&db
->db_mtx
);
3318 #pragma weak dmu_buf_refcount = dbuf_refcount
3320 dbuf_refcount(dmu_buf_impl_t
*db
)
3322 return (refcount_count(&db
->db_holds
));
3326 dmu_buf_replace_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*old_user
,
3327 dmu_buf_user_t
*new_user
)
3329 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3331 mutex_enter(&db
->db_mtx
);
3332 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
3333 if (db
->db_user
== old_user
)
3334 db
->db_user
= new_user
;
3336 old_user
= db
->db_user
;
3337 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
3338 mutex_exit(&db
->db_mtx
);
3344 dmu_buf_set_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
3346 return (dmu_buf_replace_user(db_fake
, NULL
, user
));
3350 dmu_buf_set_user_ie(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
3352 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3354 db
->db_user_immediate_evict
= TRUE
;
3355 return (dmu_buf_set_user(db_fake
, user
));
3359 dmu_buf_remove_user(dmu_buf_t
*db_fake
, dmu_buf_user_t
*user
)
3361 return (dmu_buf_replace_user(db_fake
, user
, NULL
));
3365 dmu_buf_get_user(dmu_buf_t
*db_fake
)
3367 dmu_buf_impl_t
*db
= (dmu_buf_impl_t
*)db_fake
;
3369 dbuf_verify_user(db
, DBVU_NOT_EVICTING
);
3370 return (db
->db_user
);
3374 dmu_buf_user_evict_wait()
3376 taskq_wait(dbu_evict_taskq
);
3380 dmu_buf_get_blkptr(dmu_buf_t
*db
)
3382 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3383 return (dbi
->db_blkptr
);
3387 dmu_buf_get_objset(dmu_buf_t
*db
)
3389 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3390 return (dbi
->db_objset
);
3394 dmu_buf_dnode_enter(dmu_buf_t
*db
)
3396 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3397 DB_DNODE_ENTER(dbi
);
3398 return (DB_DNODE(dbi
));
3402 dmu_buf_dnode_exit(dmu_buf_t
*db
)
3404 dmu_buf_impl_t
*dbi
= (dmu_buf_impl_t
*)db
;
3409 dbuf_check_blkptr(dnode_t
*dn
, dmu_buf_impl_t
*db
)
3411 /* ASSERT(dmu_tx_is_syncing(tx) */
3412 ASSERT(MUTEX_HELD(&db
->db_mtx
));
3414 if (db
->db_blkptr
!= NULL
)
3417 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3418 db
->db_blkptr
= DN_SPILL_BLKPTR(dn
->dn_phys
);
3419 BP_ZERO(db
->db_blkptr
);
3422 if (db
->db_level
== dn
->dn_phys
->dn_nlevels
-1) {
3424 * This buffer was allocated at a time when there was
3425 * no available blkptrs from the dnode, or it was
3426 * inappropriate to hook it in (i.e., nlevels mis-match).
3428 ASSERT(db
->db_blkid
< dn
->dn_phys
->dn_nblkptr
);
3429 ASSERT(db
->db_parent
== NULL
);
3430 db
->db_parent
= dn
->dn_dbuf
;
3431 db
->db_blkptr
= &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
];
3434 dmu_buf_impl_t
*parent
= db
->db_parent
;
3435 int epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
3437 ASSERT(dn
->dn_phys
->dn_nlevels
> 1);
3438 if (parent
== NULL
) {
3439 mutex_exit(&db
->db_mtx
);
3440 rw_enter(&dn
->dn_struct_rwlock
, RW_READER
);
3441 parent
= dbuf_hold_level(dn
, db
->db_level
+ 1,
3442 db
->db_blkid
>> epbs
, db
);
3443 rw_exit(&dn
->dn_struct_rwlock
);
3444 mutex_enter(&db
->db_mtx
);
3445 db
->db_parent
= parent
;
3447 db
->db_blkptr
= (blkptr_t
*)parent
->db
.db_data
+
3448 (db
->db_blkid
& ((1ULL << epbs
) - 1));
3454 * Ensure the dbuf's data is untransformed if the associated dirty
3455 * record requires it. This is used by dbuf_sync_leaf() to ensure
3456 * that a dnode block is decrypted before we write new data to it.
3457 * For raw writes we assert that the buffer is already encrypted.
3460 dbuf_check_crypt(dbuf_dirty_record_t
*dr
)
3463 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3465 ASSERT(MUTEX_HELD(&db
->db_mtx
));
3467 if (!dr
->dt
.dl
.dr_raw
&& arc_is_encrypted(db
->db_buf
)) {
3468 zbookmark_phys_t zb
;
3471 * Unfortunately, there is currently no mechanism for
3472 * syncing context to handle decryption errors. An error
3473 * here is only possible if an attacker maliciously
3474 * changed a dnode block and updated the associated
3475 * checksums going up the block tree.
3477 SET_BOOKMARK(&zb
, dmu_objset_id(db
->db_objset
),
3478 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
3479 err
= arc_untransform(db
->db_buf
, db
->db_objset
->os_spa
,
3482 panic("Invalid dnode block MAC");
3483 } else if (dr
->dt
.dl
.dr_raw
) {
3485 * Writing raw encrypted data requires the db's arc buffer
3486 * to be converted to raw by the caller.
3488 ASSERT(arc_is_encrypted(db
->db_buf
));
3493 * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it
3494 * is critical the we not allow the compiler to inline this function in to
3495 * dbuf_sync_list() thereby drastically bloating the stack usage.
3497 noinline
static void
3498 dbuf_sync_indirect(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
3500 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3504 ASSERT(dmu_tx_is_syncing(tx
));
3506 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
3508 mutex_enter(&db
->db_mtx
);
3510 ASSERT(db
->db_level
> 0);
3513 /* Read the block if it hasn't been read yet. */
3514 if (db
->db_buf
== NULL
) {
3515 mutex_exit(&db
->db_mtx
);
3516 (void) dbuf_read(db
, NULL
, DB_RF_MUST_SUCCEED
);
3517 mutex_enter(&db
->db_mtx
);
3519 ASSERT3U(db
->db_state
, ==, DB_CACHED
);
3520 ASSERT(db
->db_buf
!= NULL
);
3524 /* Indirect block size must match what the dnode thinks it is. */
3525 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
3526 dbuf_check_blkptr(dn
, db
);
3529 /* Provide the pending dirty record to child dbufs */
3530 db
->db_data_pending
= dr
;
3532 mutex_exit(&db
->db_mtx
);
3533 dbuf_write(dr
, db
->db_buf
, tx
);
3536 mutex_enter(&dr
->dt
.di
.dr_mtx
);
3537 dbuf_sync_list(&dr
->dt
.di
.dr_children
, db
->db_level
- 1, tx
);
3538 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
3539 mutex_exit(&dr
->dt
.di
.dr_mtx
);
3544 * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is
3545 * critical the we not allow the compiler to inline this function in to
3546 * dbuf_sync_list() thereby drastically bloating the stack usage.
3548 noinline
static void
3549 dbuf_sync_leaf(dbuf_dirty_record_t
*dr
, dmu_tx_t
*tx
)
3551 arc_buf_t
**datap
= &dr
->dt
.dl
.dr_data
;
3552 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
3555 uint64_t txg
= tx
->tx_txg
;
3557 ASSERT(dmu_tx_is_syncing(tx
));
3559 dprintf_dbuf_bp(db
, db
->db_blkptr
, "blkptr=%p", db
->db_blkptr
);
3561 mutex_enter(&db
->db_mtx
);
3563 * To be synced, we must be dirtied. But we
3564 * might have been freed after the dirty.
3566 if (db
->db_state
== DB_UNCACHED
) {
3567 /* This buffer has been freed since it was dirtied */
3568 ASSERT(db
->db
.db_data
== NULL
);
3569 } else if (db
->db_state
== DB_FILL
) {
3570 /* This buffer was freed and is now being re-filled */
3571 ASSERT(db
->db
.db_data
!= dr
->dt
.dl
.dr_data
);
3573 ASSERT(db
->db_state
== DB_CACHED
|| db
->db_state
== DB_NOFILL
);
3580 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3581 mutex_enter(&dn
->dn_mtx
);
3582 if (!(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
)) {
3584 * In the previous transaction group, the bonus buffer
3585 * was entirely used to store the attributes for the
3586 * dnode which overrode the dn_spill field. However,
3587 * when adding more attributes to the file a spill
3588 * block was required to hold the extra attributes.
3590 * Make sure to clear the garbage left in the dn_spill
3591 * field from the previous attributes in the bonus
3592 * buffer. Otherwise, after writing out the spill
3593 * block to the new allocated dva, it will free
3594 * the old block pointed to by the invalid dn_spill.
3596 db
->db_blkptr
= NULL
;
3598 dn
->dn_phys
->dn_flags
|= DNODE_FLAG_SPILL_BLKPTR
;
3599 mutex_exit(&dn
->dn_mtx
);
3603 * If this is a bonus buffer, simply copy the bonus data into the
3604 * dnode. It will be written out when the dnode is synced (and it
3605 * will be synced, since it must have been dirty for dbuf_sync to
3608 if (db
->db_blkid
== DMU_BONUS_BLKID
) {
3609 dbuf_dirty_record_t
**drp
;
3611 ASSERT(*datap
!= NULL
);
3612 ASSERT0(db
->db_level
);
3613 ASSERT3U(DN_MAX_BONUS_LEN(dn
->dn_phys
), <=,
3614 DN_SLOTS_TO_BONUSLEN(dn
->dn_phys
->dn_extra_slots
+ 1));
3615 bcopy(*datap
, DN_BONUS(dn
->dn_phys
),
3616 DN_MAX_BONUS_LEN(dn
->dn_phys
));
3619 if (*datap
!= db
->db
.db_data
) {
3620 int slots
= DB_DNODE(db
)->dn_num_slots
;
3621 int bonuslen
= DN_SLOTS_TO_BONUSLEN(slots
);
3622 kmem_free(*datap
, bonuslen
);
3623 arc_space_return(bonuslen
, ARC_SPACE_BONUS
);
3625 db
->db_data_pending
= NULL
;
3626 drp
= &db
->db_last_dirty
;
3628 drp
= &(*drp
)->dr_next
;
3629 ASSERT(dr
->dr_next
== NULL
);
3630 ASSERT(dr
->dr_dbuf
== db
);
3632 if (dr
->dr_dbuf
->db_level
!= 0) {
3633 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
3634 list_destroy(&dr
->dt
.di
.dr_children
);
3636 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
3637 ASSERT(db
->db_dirtycnt
> 0);
3638 db
->db_dirtycnt
-= 1;
3639 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)txg
);
3646 * This function may have dropped the db_mtx lock allowing a dmu_sync
3647 * operation to sneak in. As a result, we need to ensure that we
3648 * don't check the dr_override_state until we have returned from
3649 * dbuf_check_blkptr.
3651 dbuf_check_blkptr(dn
, db
);
3654 * If this buffer is in the middle of an immediate write,
3655 * wait for the synchronous IO to complete.
3657 while (dr
->dt
.dl
.dr_override_state
== DR_IN_DMU_SYNC
) {
3658 ASSERT(dn
->dn_object
!= DMU_META_DNODE_OBJECT
);
3659 cv_wait(&db
->db_changed
, &db
->db_mtx
);
3660 ASSERT(dr
->dt
.dl
.dr_override_state
!= DR_NOT_OVERRIDDEN
);
3664 * If this is a dnode block, ensure it is appropriately encrypted
3665 * or decrypted, depending on what we are writing to it this txg.
3667 if (os
->os_encrypted
&& dn
->dn_object
== DMU_META_DNODE_OBJECT
)
3668 dbuf_check_crypt(dr
);
3670 if (db
->db_state
!= DB_NOFILL
&&
3671 dn
->dn_object
!= DMU_META_DNODE_OBJECT
&&
3672 refcount_count(&db
->db_holds
) > 1 &&
3673 dr
->dt
.dl
.dr_override_state
!= DR_OVERRIDDEN
&&
3674 *datap
== db
->db_buf
) {
3676 * If this buffer is currently "in use" (i.e., there
3677 * are active holds and db_data still references it),
3678 * then make a copy before we start the write so that
3679 * any modifications from the open txg will not leak
3682 * NOTE: this copy does not need to be made for
3683 * objects only modified in the syncing context (e.g.
3684 * DNONE_DNODE blocks).
3686 int psize
= arc_buf_size(*datap
);
3687 int lsize
= arc_buf_lsize(*datap
);
3688 arc_buf_contents_t type
= DBUF_GET_BUFC_TYPE(db
);
3689 enum zio_compress compress_type
= arc_get_compression(*datap
);
3691 if (arc_is_encrypted(*datap
)) {
3692 boolean_t byteorder
;
3693 uint8_t salt
[ZIO_DATA_SALT_LEN
];
3694 uint8_t iv
[ZIO_DATA_IV_LEN
];
3695 uint8_t mac
[ZIO_DATA_MAC_LEN
];
3697 arc_get_raw_params(*datap
, &byteorder
, salt
, iv
, mac
);
3698 *datap
= arc_alloc_raw_buf(os
->os_spa
, db
,
3699 dmu_objset_id(os
), byteorder
, salt
, iv
, mac
,
3700 dn
->dn_type
, psize
, lsize
, compress_type
);
3701 } else if (compress_type
!= ZIO_COMPRESS_OFF
) {
3702 ASSERT3U(type
, ==, ARC_BUFC_DATA
);
3703 *datap
= arc_alloc_compressed_buf(os
->os_spa
, db
,
3704 psize
, lsize
, compress_type
);
3706 *datap
= arc_alloc_buf(os
->os_spa
, db
, type
, psize
);
3708 bcopy(db
->db
.db_data
, (*datap
)->b_data
, psize
);
3710 db
->db_data_pending
= dr
;
3712 mutex_exit(&db
->db_mtx
);
3714 dbuf_write(dr
, *datap
, tx
);
3716 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
3717 if (dn
->dn_object
== DMU_META_DNODE_OBJECT
) {
3718 list_insert_tail(&dn
->dn_dirty_records
[txg
&TXG_MASK
], dr
);
3722 * Although zio_nowait() does not "wait for an IO", it does
3723 * initiate the IO. If this is an empty write it seems plausible
3724 * that the IO could actually be completed before the nowait
3725 * returns. We need to DB_DNODE_EXIT() first in case
3726 * zio_nowait() invalidates the dbuf.
3729 zio_nowait(dr
->dr_zio
);
3734 dbuf_sync_list(list_t
*list
, int level
, dmu_tx_t
*tx
)
3736 dbuf_dirty_record_t
*dr
;
3738 while ((dr
= list_head(list
))) {
3739 if (dr
->dr_zio
!= NULL
) {
3741 * If we find an already initialized zio then we
3742 * are processing the meta-dnode, and we have finished.
3743 * The dbufs for all dnodes are put back on the list
3744 * during processing, so that we can zio_wait()
3745 * these IOs after initiating all child IOs.
3747 ASSERT3U(dr
->dr_dbuf
->db
.db_object
, ==,
3748 DMU_META_DNODE_OBJECT
);
3751 if (dr
->dr_dbuf
->db_blkid
!= DMU_BONUS_BLKID
&&
3752 dr
->dr_dbuf
->db_blkid
!= DMU_SPILL_BLKID
) {
3753 VERIFY3U(dr
->dr_dbuf
->db_level
, ==, level
);
3755 list_remove(list
, dr
);
3756 if (dr
->dr_dbuf
->db_level
> 0)
3757 dbuf_sync_indirect(dr
, tx
);
3759 dbuf_sync_leaf(dr
, tx
);
3765 dbuf_write_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
3767 dmu_buf_impl_t
*db
= vdb
;
3769 blkptr_t
*bp
= zio
->io_bp
;
3770 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
3771 spa_t
*spa
= zio
->io_spa
;
3776 ASSERT3P(db
->db_blkptr
, !=, NULL
);
3777 ASSERT3P(&db
->db_data_pending
->dr_bp_copy
, ==, bp
);
3781 delta
= bp_get_dsize_sync(spa
, bp
) - bp_get_dsize_sync(spa
, bp_orig
);
3782 dnode_diduse_space(dn
, delta
- zio
->io_prev_space_delta
);
3783 zio
->io_prev_space_delta
= delta
;
3785 if (bp
->blk_birth
!= 0) {
3786 ASSERT((db
->db_blkid
!= DMU_SPILL_BLKID
&&
3787 BP_GET_TYPE(bp
) == dn
->dn_type
) ||
3788 (db
->db_blkid
== DMU_SPILL_BLKID
&&
3789 BP_GET_TYPE(bp
) == dn
->dn_bonustype
) ||
3790 BP_IS_EMBEDDED(bp
));
3791 ASSERT(BP_GET_LEVEL(bp
) == db
->db_level
);
3794 mutex_enter(&db
->db_mtx
);
3797 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3798 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
3799 ASSERT(!(BP_IS_HOLE(bp
)) &&
3800 db
->db_blkptr
== DN_SPILL_BLKPTR(dn
->dn_phys
));
3804 if (db
->db_level
== 0) {
3805 mutex_enter(&dn
->dn_mtx
);
3806 if (db
->db_blkid
> dn
->dn_phys
->dn_maxblkid
&&
3807 db
->db_blkid
!= DMU_SPILL_BLKID
)
3808 dn
->dn_phys
->dn_maxblkid
= db
->db_blkid
;
3809 mutex_exit(&dn
->dn_mtx
);
3811 if (dn
->dn_type
== DMU_OT_DNODE
) {
3813 while (i
< db
->db
.db_size
) {
3815 (void *)(((char *)db
->db
.db_data
) + i
);
3817 i
+= DNODE_MIN_SIZE
;
3818 if (dnp
->dn_type
!= DMU_OT_NONE
) {
3820 i
+= dnp
->dn_extra_slots
*
3825 if (BP_IS_HOLE(bp
)) {
3832 blkptr_t
*ibp
= db
->db
.db_data
;
3833 ASSERT3U(db
->db
.db_size
, ==, 1<<dn
->dn_phys
->dn_indblkshift
);
3834 for (i
= db
->db
.db_size
>> SPA_BLKPTRSHIFT
; i
> 0; i
--, ibp
++) {
3835 if (BP_IS_HOLE(ibp
))
3837 fill
+= BP_GET_FILL(ibp
);
3842 if (!BP_IS_EMBEDDED(bp
))
3843 BP_SET_FILL(bp
, fill
);
3845 mutex_exit(&db
->db_mtx
);
3847 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
3848 *db
->db_blkptr
= *bp
;
3849 rw_exit(&dn
->dn_struct_rwlock
);
3854 * This function gets called just prior to running through the compression
3855 * stage of the zio pipeline. If we're an indirect block comprised of only
3856 * holes, then we want this indirect to be compressed away to a hole. In
3857 * order to do that we must zero out any information about the holes that
3858 * this indirect points to prior to before we try to compress it.
3861 dbuf_write_children_ready(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
3863 dmu_buf_impl_t
*db
= vdb
;
3866 unsigned int epbs
, i
;
3868 ASSERT3U(db
->db_level
, >, 0);
3871 epbs
= dn
->dn_phys
->dn_indblkshift
- SPA_BLKPTRSHIFT
;
3872 ASSERT3U(epbs
, <, 31);
3874 /* Determine if all our children are holes */
3875 for (i
= 0, bp
= db
->db
.db_data
; i
< 1ULL << epbs
; i
++, bp
++) {
3876 if (!BP_IS_HOLE(bp
))
3881 * If all the children are holes, then zero them all out so that
3882 * we may get compressed away.
3884 if (i
== 1ULL << epbs
) {
3886 * We only found holes. Grab the rwlock to prevent
3887 * anybody from reading the blocks we're about to
3890 rw_enter(&dn
->dn_struct_rwlock
, RW_WRITER
);
3891 bzero(db
->db
.db_data
, db
->db
.db_size
);
3892 rw_exit(&dn
->dn_struct_rwlock
);
3898 * The SPA will call this callback several times for each zio - once
3899 * for every physical child i/o (zio->io_phys_children times). This
3900 * allows the DMU to monitor the progress of each logical i/o. For example,
3901 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
3902 * block. There may be a long delay before all copies/fragments are completed,
3903 * so this callback allows us to retire dirty space gradually, as the physical
3908 dbuf_write_physdone(zio_t
*zio
, arc_buf_t
*buf
, void *arg
)
3910 dmu_buf_impl_t
*db
= arg
;
3911 objset_t
*os
= db
->db_objset
;
3912 dsl_pool_t
*dp
= dmu_objset_pool(os
);
3913 dbuf_dirty_record_t
*dr
;
3916 dr
= db
->db_data_pending
;
3917 ASSERT3U(dr
->dr_txg
, ==, zio
->io_txg
);
3920 * The callback will be called io_phys_children times. Retire one
3921 * portion of our dirty space each time we are called. Any rounding
3922 * error will be cleaned up by dsl_pool_sync()'s call to
3923 * dsl_pool_undirty_space().
3925 delta
= dr
->dr_accounted
/ zio
->io_phys_children
;
3926 dsl_pool_undirty_space(dp
, delta
, zio
->io_txg
);
3931 dbuf_write_done(zio_t
*zio
, arc_buf_t
*buf
, void *vdb
)
3933 dmu_buf_impl_t
*db
= vdb
;
3934 blkptr_t
*bp_orig
= &zio
->io_bp_orig
;
3935 blkptr_t
*bp
= db
->db_blkptr
;
3936 objset_t
*os
= db
->db_objset
;
3937 dmu_tx_t
*tx
= os
->os_synctx
;
3938 dbuf_dirty_record_t
**drp
, *dr
;
3940 ASSERT0(zio
->io_error
);
3941 ASSERT(db
->db_blkptr
== bp
);
3944 * For nopwrites and rewrites we ensure that the bp matches our
3945 * original and bypass all the accounting.
3947 if (zio
->io_flags
& (ZIO_FLAG_IO_REWRITE
| ZIO_FLAG_NOPWRITE
)) {
3948 ASSERT(BP_EQUAL(bp
, bp_orig
));
3950 dsl_dataset_t
*ds
= os
->os_dsl_dataset
;
3951 (void) dsl_dataset_block_kill(ds
, bp_orig
, tx
, B_TRUE
);
3952 dsl_dataset_block_born(ds
, bp
, tx
);
3955 mutex_enter(&db
->db_mtx
);
3959 drp
= &db
->db_last_dirty
;
3960 while ((dr
= *drp
) != db
->db_data_pending
)
3962 ASSERT(!list_link_active(&dr
->dr_dirty_node
));
3963 ASSERT(dr
->dr_dbuf
== db
);
3964 ASSERT(dr
->dr_next
== NULL
);
3968 if (db
->db_blkid
== DMU_SPILL_BLKID
) {
3973 ASSERT(dn
->dn_phys
->dn_flags
& DNODE_FLAG_SPILL_BLKPTR
);
3974 ASSERT(!(BP_IS_HOLE(db
->db_blkptr
)) &&
3975 db
->db_blkptr
== DN_SPILL_BLKPTR(dn
->dn_phys
));
3980 if (db
->db_level
== 0) {
3981 ASSERT(db
->db_blkid
!= DMU_BONUS_BLKID
);
3982 ASSERT(dr
->dt
.dl
.dr_override_state
== DR_NOT_OVERRIDDEN
);
3983 if (db
->db_state
!= DB_NOFILL
) {
3984 if (dr
->dt
.dl
.dr_data
!= db
->db_buf
)
3985 arc_buf_destroy(dr
->dt
.dl
.dr_data
, db
);
3992 ASSERT(list_head(&dr
->dt
.di
.dr_children
) == NULL
);
3993 ASSERT3U(db
->db
.db_size
, ==, 1 << dn
->dn_phys
->dn_indblkshift
);
3994 if (!BP_IS_HOLE(db
->db_blkptr
)) {
3995 ASSERTV(int epbs
= dn
->dn_phys
->dn_indblkshift
-
3997 ASSERT3U(db
->db_blkid
, <=,
3998 dn
->dn_phys
->dn_maxblkid
>> (db
->db_level
* epbs
));
3999 ASSERT3U(BP_GET_LSIZE(db
->db_blkptr
), ==,
4003 mutex_destroy(&dr
->dt
.di
.dr_mtx
);
4004 list_destroy(&dr
->dt
.di
.dr_children
);
4006 kmem_free(dr
, sizeof (dbuf_dirty_record_t
));
4008 cv_broadcast(&db
->db_changed
);
4009 ASSERT(db
->db_dirtycnt
> 0);
4010 db
->db_dirtycnt
-= 1;
4011 db
->db_data_pending
= NULL
;
4012 dbuf_rele_and_unlock(db
, (void *)(uintptr_t)tx
->tx_txg
);
4016 dbuf_write_nofill_ready(zio_t
*zio
)
4018 dbuf_write_ready(zio
, NULL
, zio
->io_private
);
4022 dbuf_write_nofill_done(zio_t
*zio
)
4024 dbuf_write_done(zio
, NULL
, zio
->io_private
);
4028 dbuf_write_override_ready(zio_t
*zio
)
4030 dbuf_dirty_record_t
*dr
= zio
->io_private
;
4031 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4033 dbuf_write_ready(zio
, NULL
, db
);
4037 dbuf_write_override_done(zio_t
*zio
)
4039 dbuf_dirty_record_t
*dr
= zio
->io_private
;
4040 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4041 blkptr_t
*obp
= &dr
->dt
.dl
.dr_overridden_by
;
4043 mutex_enter(&db
->db_mtx
);
4044 if (!BP_EQUAL(zio
->io_bp
, obp
)) {
4045 if (!BP_IS_HOLE(obp
))
4046 dsl_free(spa_get_dsl(zio
->io_spa
), zio
->io_txg
, obp
);
4047 arc_release(dr
->dt
.dl
.dr_data
, db
);
4049 mutex_exit(&db
->db_mtx
);
4051 dbuf_write_done(zio
, NULL
, db
);
4053 if (zio
->io_abd
!= NULL
)
4054 abd_put(zio
->io_abd
);
4057 /* Issue I/O to commit a dirty buffer to disk. */
4059 dbuf_write(dbuf_dirty_record_t
*dr
, arc_buf_t
*data
, dmu_tx_t
*tx
)
4061 dmu_buf_impl_t
*db
= dr
->dr_dbuf
;
4064 dmu_buf_impl_t
*parent
= db
->db_parent
;
4065 uint64_t txg
= tx
->tx_txg
;
4066 zbookmark_phys_t zb
;
4071 ASSERT(dmu_tx_is_syncing(tx
));
4077 if (db
->db_state
!= DB_NOFILL
) {
4078 if (db
->db_level
> 0 || dn
->dn_type
== DMU_OT_DNODE
) {
4080 * Private object buffers are released here rather
4081 * than in dbuf_dirty() since they are only modified
4082 * in the syncing context and we don't want the
4083 * overhead of making multiple copies of the data.
4085 if (BP_IS_HOLE(db
->db_blkptr
)) {
4088 dbuf_release_bp(db
);
4093 if (parent
!= dn
->dn_dbuf
) {
4094 /* Our parent is an indirect block. */
4095 /* We have a dirty parent that has been scheduled for write. */
4096 ASSERT(parent
&& parent
->db_data_pending
);
4097 /* Our parent's buffer is one level closer to the dnode. */
4098 ASSERT(db
->db_level
== parent
->db_level
-1);
4100 * We're about to modify our parent's db_data by modifying
4101 * our block pointer, so the parent must be released.
4103 ASSERT(arc_released(parent
->db_buf
));
4104 zio
= parent
->db_data_pending
->dr_zio
;
4106 /* Our parent is the dnode itself. */
4107 ASSERT((db
->db_level
== dn
->dn_phys
->dn_nlevels
-1 &&
4108 db
->db_blkid
!= DMU_SPILL_BLKID
) ||
4109 (db
->db_blkid
== DMU_SPILL_BLKID
&& db
->db_level
== 0));
4110 if (db
->db_blkid
!= DMU_SPILL_BLKID
)
4111 ASSERT3P(db
->db_blkptr
, ==,
4112 &dn
->dn_phys
->dn_blkptr
[db
->db_blkid
]);
4116 ASSERT(db
->db_level
== 0 || data
== db
->db_buf
);
4117 ASSERT3U(db
->db_blkptr
->blk_birth
, <=, txg
);
4120 SET_BOOKMARK(&zb
, os
->os_dsl_dataset
?
4121 os
->os_dsl_dataset
->ds_object
: DMU_META_OBJSET
,
4122 db
->db
.db_object
, db
->db_level
, db
->db_blkid
);
4124 if (db
->db_blkid
== DMU_SPILL_BLKID
)
4126 wp_flag
|= (db
->db_state
== DB_NOFILL
) ? WP_NOFILL
: 0;
4128 dmu_write_policy(os
, dn
, db
->db_level
, wp_flag
, &zp
);
4132 * We copy the blkptr now (rather than when we instantiate the dirty
4133 * record), because its value can change between open context and
4134 * syncing context. We do not need to hold dn_struct_rwlock to read
4135 * db_blkptr because we are in syncing context.
4137 dr
->dr_bp_copy
= *db
->db_blkptr
;
4139 if (db
->db_level
== 0 &&
4140 dr
->dt
.dl
.dr_override_state
== DR_OVERRIDDEN
) {
4142 * The BP for this block has been provided by open context
4143 * (by dmu_sync() or dmu_buf_write_embedded()).
4145 abd_t
*contents
= (data
!= NULL
) ?
4146 abd_get_from_buf(data
->b_data
, arc_buf_size(data
)) : NULL
;
4148 dr
->dr_zio
= zio_write(zio
, os
->os_spa
, txg
,
4149 &dr
->dr_bp_copy
, contents
, db
->db
.db_size
, db
->db
.db_size
,
4150 &zp
, dbuf_write_override_ready
, NULL
, NULL
,
4151 dbuf_write_override_done
,
4152 dr
, ZIO_PRIORITY_ASYNC_WRITE
, ZIO_FLAG_MUSTSUCCEED
, &zb
);
4153 mutex_enter(&db
->db_mtx
);
4154 dr
->dt
.dl
.dr_override_state
= DR_NOT_OVERRIDDEN
;
4155 zio_write_override(dr
->dr_zio
, &dr
->dt
.dl
.dr_overridden_by
,
4156 dr
->dt
.dl
.dr_copies
, dr
->dt
.dl
.dr_nopwrite
);
4157 mutex_exit(&db
->db_mtx
);
4158 } else if (db
->db_state
== DB_NOFILL
) {
4159 ASSERT(zp
.zp_checksum
== ZIO_CHECKSUM_OFF
||
4160 zp
.zp_checksum
== ZIO_CHECKSUM_NOPARITY
);
4161 dr
->dr_zio
= zio_write(zio
, os
->os_spa
, txg
,
4162 &dr
->dr_bp_copy
, NULL
, db
->db
.db_size
, db
->db
.db_size
, &zp
,
4163 dbuf_write_nofill_ready
, NULL
, NULL
,
4164 dbuf_write_nofill_done
, db
,
4165 ZIO_PRIORITY_ASYNC_WRITE
,
4166 ZIO_FLAG_MUSTSUCCEED
| ZIO_FLAG_NODATA
, &zb
);
4168 ASSERT(arc_released(data
));
4171 * For indirect blocks, we want to setup the children
4172 * ready callback so that we can properly handle an indirect
4173 * block that only contains holes.
4175 arc_write_done_func_t
*children_ready_cb
= NULL
;
4176 if (db
->db_level
!= 0)
4177 children_ready_cb
= dbuf_write_children_ready
;
4179 dr
->dr_zio
= arc_write(zio
, os
->os_spa
, txg
,
4180 &dr
->dr_bp_copy
, data
, DBUF_IS_L2CACHEABLE(db
),
4181 &zp
, dbuf_write_ready
,
4182 children_ready_cb
, dbuf_write_physdone
,
4183 dbuf_write_done
, db
, ZIO_PRIORITY_ASYNC_WRITE
,
4184 ZIO_FLAG_MUSTSUCCEED
, &zb
);
4188 #if defined(_KERNEL) && defined(HAVE_SPL)
4189 EXPORT_SYMBOL(dbuf_find
);
4190 EXPORT_SYMBOL(dbuf_is_metadata
);
4191 EXPORT_SYMBOL(dbuf_destroy
);
4192 EXPORT_SYMBOL(dbuf_loan_arcbuf
);
4193 EXPORT_SYMBOL(dbuf_whichblock
);
4194 EXPORT_SYMBOL(dbuf_read
);
4195 EXPORT_SYMBOL(dbuf_unoverride
);
4196 EXPORT_SYMBOL(dbuf_free_range
);
4197 EXPORT_SYMBOL(dbuf_new_size
);
4198 EXPORT_SYMBOL(dbuf_release_bp
);
4199 EXPORT_SYMBOL(dbuf_dirty
);
4200 EXPORT_SYMBOL(dmu_buf_will_change_crypt_params
);
4201 EXPORT_SYMBOL(dmu_buf_will_dirty
);
4202 EXPORT_SYMBOL(dmu_buf_will_not_fill
);
4203 EXPORT_SYMBOL(dmu_buf_will_fill
);
4204 EXPORT_SYMBOL(dmu_buf_fill_done
);
4205 EXPORT_SYMBOL(dmu_buf_rele
);
4206 EXPORT_SYMBOL(dbuf_assign_arcbuf
);
4207 EXPORT_SYMBOL(dbuf_prefetch
);
4208 EXPORT_SYMBOL(dbuf_hold_impl
);
4209 EXPORT_SYMBOL(dbuf_hold
);
4210 EXPORT_SYMBOL(dbuf_hold_level
);
4211 EXPORT_SYMBOL(dbuf_create_bonus
);
4212 EXPORT_SYMBOL(dbuf_spill_set_blksz
);
4213 EXPORT_SYMBOL(dbuf_rm_spill
);
4214 EXPORT_SYMBOL(dbuf_add_ref
);
4215 EXPORT_SYMBOL(dbuf_rele
);
4216 EXPORT_SYMBOL(dbuf_rele_and_unlock
);
4217 EXPORT_SYMBOL(dbuf_refcount
);
4218 EXPORT_SYMBOL(dbuf_sync_list
);
4219 EXPORT_SYMBOL(dmu_buf_set_user
);
4220 EXPORT_SYMBOL(dmu_buf_set_user_ie
);
4221 EXPORT_SYMBOL(dmu_buf_get_user
);
4222 EXPORT_SYMBOL(dmu_buf_get_blkptr
);
4225 module_param(dbuf_cache_max_bytes
, ulong
, 0644);
4226 MODULE_PARM_DESC(dbuf_cache_max_bytes
,
4227 "Maximum size in bytes of the dbuf cache.");
4229 module_param(dbuf_cache_hiwater_pct
, uint
, 0644);
4230 MODULE_PARM_DESC(dbuf_cache_hiwater_pct
,
4231 "Percentage over dbuf_cache_max_bytes when dbufs must be evicted "
4234 module_param(dbuf_cache_lowater_pct
, uint
, 0644);
4235 MODULE_PARM_DESC(dbuf_cache_lowater_pct
,
4236 "Percentage below dbuf_cache_max_bytes when the evict thread stops "
4239 module_param(dbuf_cache_shift
, int, 0644);
4240 MODULE_PARM_DESC(dbuf_cache_shift
,
4241 "Set the size of the dbuf cache to a log2 fraction of arc size.");