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>
51 #include <sys/cityhash.h>
55 typedef struct dbuf_stats {
57 * Various statistics about the size of the dbuf cache.
59 kstat_named_t cache_count;
60 kstat_named_t cache_size_bytes;
61 kstat_named_t cache_size_bytes_max;
63 * Statistics regarding the bounds on the dbuf cache size.
65 kstat_named_t cache_target_bytes;
66 kstat_named_t cache_lowater_bytes;
67 kstat_named_t cache_hiwater_bytes;
69 * Total number of dbuf cache evictions that have occurred.
71 kstat_named_t cache_total_evicts;
73 * The distribution of dbuf levels in the dbuf cache and
74 * the total size of all dbufs at each level.
76 kstat_named_t cache_levels[DN_MAX_LEVELS];
77 kstat_named_t cache_levels_bytes[DN_MAX_LEVELS];
79 * Statistics about the dbuf hash table.
81 kstat_named_t hash_hits;
82 kstat_named_t hash_misses;
83 kstat_named_t hash_collisions;
84 kstat_named_t hash_elements;
85 kstat_named_t hash_elements_max;
87 * Number of sublists containing more than one dbuf in the dbuf
88 * hash table. Keep track of the longest hash chain.
90 kstat_named_t hash_chains;
91 kstat_named_t hash_chain_max;
93 * Number of times a dbuf_create() discovers that a dbuf was
94 * already created and in the dbuf hash table.
96 kstat_named_t hash_insert_race;
99 dbuf_stats_t dbuf_stats = {
100 { "cache_count", KSTAT_DATA_UINT64 },
101 { "cache_size_bytes", KSTAT_DATA_UINT64 },
102 { "cache_size_bytes_max", KSTAT_DATA_UINT64 },
103 { "cache_target_bytes", KSTAT_DATA_UINT64 },
104 { "cache_lowater_bytes", KSTAT_DATA_UINT64 },
105 { "cache_hiwater_bytes", KSTAT_DATA_UINT64 },
106 { "cache_total_evicts", KSTAT_DATA_UINT64 },
107 { { "cache_levels_N", KSTAT_DATA_UINT64 } },
108 { { "cache_levels_bytes_N", KSTAT_DATA_UINT64 } },
109 { "hash_hits", KSTAT_DATA_UINT64 },
110 { "hash_misses", KSTAT_DATA_UINT64 },
111 { "hash_collisions", KSTAT_DATA_UINT64 },
112 { "hash_elements", KSTAT_DATA_UINT64 },
113 { "hash_elements_max", KSTAT_DATA_UINT64 },
114 { "hash_chains", KSTAT_DATA_UINT64 },
115 { "hash_chain_max", KSTAT_DATA_UINT64 },
116 { "hash_insert_race", KSTAT_DATA_UINT64 }
119 #define DBUF_STAT_INCR(stat, val) \
120 atomic_add_64(&dbuf_stats.stat.value.ui64, (val));
121 #define DBUF_STAT_DECR(stat, val) \
122 DBUF_STAT_INCR(stat, -(val));
123 #define DBUF_STAT_BUMP(stat) \
124 DBUF_STAT_INCR(stat, 1);
125 #define DBUF_STAT_BUMPDOWN(stat) \
126 DBUF_STAT_INCR(stat, -1);
127 #define DBUF_STAT_MAX(stat, v) { \
129 while ((v) > (_m = dbuf_stats.stat.value.ui64) && \
130 (_m != atomic_cas_64(&dbuf_stats.stat.value.ui64, _m, (v))))\
134 struct dbuf_hold_impl_data {
135 /* Function arguments */
139 boolean_t dh_fail_sparse;
140 boolean_t dh_fail_uncached;
142 dmu_buf_impl_t **dh_dbp;
143 /* Local variables */
144 dmu_buf_impl_t *dh_db;
145 dmu_buf_impl_t *dh_parent;
148 dbuf_dirty_record_t *dh_dr;
152 static void __dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh,
153 dnode_t *dn, uint8_t level, uint64_t blkid, boolean_t fail_sparse,
154 boolean_t fail_uncached,
155 void *tag, dmu_buf_impl_t **dbp, int depth);
156 static int __dbuf_hold_impl(struct dbuf_hold_impl_data *dh);
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 * We use Cityhash for this. It's fast, and has good hash properties without
274 * requiring any large static buffers.
277 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
279 return (cityhash4((uintptr_t)os, obj, (uint64_t)lvl, blkid));
282 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
283 ((dbuf)->db.db_object == (obj) && \
284 (dbuf)->db_objset == (os) && \
285 (dbuf)->db_level == (level) && \
286 (dbuf)->db_blkid == (blkid))
289 dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid)
291 dbuf_hash_table_t *h = &dbuf_hash_table;
296 hv = dbuf_hash(os, obj, level, blkid);
297 idx = hv & h->hash_table_mask;
299 mutex_enter(DBUF_HASH_MUTEX(h, idx));
300 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
301 if (DBUF_EQUAL(db, os, obj, level, blkid)) {
302 mutex_enter(&db->db_mtx);
303 if (db->db_state != DB_EVICTING) {
304 mutex_exit(DBUF_HASH_MUTEX(h, idx));
307 mutex_exit(&db->db_mtx);
310 mutex_exit(DBUF_HASH_MUTEX(h, idx));
314 static dmu_buf_impl_t *
315 dbuf_find_bonus(objset_t *os, uint64_t object)
318 dmu_buf_impl_t *db = NULL;
320 if (dnode_hold(os, object, FTAG, &dn) == 0) {
321 rw_enter(&dn->dn_struct_rwlock, RW_READER);
322 if (dn->dn_bonus != NULL) {
324 mutex_enter(&db->db_mtx);
326 rw_exit(&dn->dn_struct_rwlock);
327 dnode_rele(dn, FTAG);
333 * Insert an entry into the hash table. If there is already an element
334 * equal to elem in the hash table, then the already existing element
335 * will be returned and the new element will not be inserted.
336 * Otherwise returns NULL.
338 static dmu_buf_impl_t *
339 dbuf_hash_insert(dmu_buf_impl_t *db)
341 dbuf_hash_table_t *h = &dbuf_hash_table;
342 objset_t *os = db->db_objset;
343 uint64_t obj = db->db.db_object;
344 int level = db->db_level;
345 uint64_t blkid, hv, idx;
349 blkid = db->db_blkid;
350 hv = dbuf_hash(os, obj, level, blkid);
351 idx = hv & h->hash_table_mask;
353 mutex_enter(DBUF_HASH_MUTEX(h, idx));
354 for (dbf = h->hash_table[idx], i = 0; dbf != NULL;
355 dbf = dbf->db_hash_next, i++) {
356 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
357 mutex_enter(&dbf->db_mtx);
358 if (dbf->db_state != DB_EVICTING) {
359 mutex_exit(DBUF_HASH_MUTEX(h, idx));
362 mutex_exit(&dbf->db_mtx);
367 DBUF_STAT_BUMP(hash_collisions);
369 DBUF_STAT_BUMP(hash_chains);
371 DBUF_STAT_MAX(hash_chain_max, i);
374 mutex_enter(&db->db_mtx);
375 db->db_hash_next = h->hash_table[idx];
376 h->hash_table[idx] = db;
377 mutex_exit(DBUF_HASH_MUTEX(h, idx));
378 atomic_inc_64(&dbuf_hash_count);
379 DBUF_STAT_MAX(hash_elements_max, dbuf_hash_count);
385 * Remove an entry from the hash table. It must be in the EVICTING state.
388 dbuf_hash_remove(dmu_buf_impl_t *db)
390 dbuf_hash_table_t *h = &dbuf_hash_table;
392 dmu_buf_impl_t *dbf, **dbp;
394 hv = dbuf_hash(db->db_objset, db->db.db_object,
395 db->db_level, db->db_blkid);
396 idx = hv & h->hash_table_mask;
399 * We mustn't hold db_mtx to maintain lock ordering:
400 * DBUF_HASH_MUTEX > db_mtx.
402 ASSERT(refcount_is_zero(&db->db_holds));
403 ASSERT(db->db_state == DB_EVICTING);
404 ASSERT(!MUTEX_HELD(&db->db_mtx));
406 mutex_enter(DBUF_HASH_MUTEX(h, idx));
407 dbp = &h->hash_table[idx];
408 while ((dbf = *dbp) != db) {
409 dbp = &dbf->db_hash_next;
412 *dbp = db->db_hash_next;
413 db->db_hash_next = NULL;
414 if (h->hash_table[idx] &&
415 h->hash_table[idx]->db_hash_next == NULL)
416 DBUF_STAT_BUMPDOWN(hash_chains);
417 mutex_exit(DBUF_HASH_MUTEX(h, idx));
418 atomic_dec_64(&dbuf_hash_count);
424 } dbvu_verify_type_t;
427 dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type)
432 if (db->db_user == NULL)
435 /* Only data blocks support the attachment of user data. */
436 ASSERT(db->db_level == 0);
438 /* Clients must resolve a dbuf before attaching user data. */
439 ASSERT(db->db.db_data != NULL);
440 ASSERT3U(db->db_state, ==, DB_CACHED);
442 holds = refcount_count(&db->db_holds);
443 if (verify_type == DBVU_EVICTING) {
445 * Immediate eviction occurs when holds == dirtycnt.
446 * For normal eviction buffers, holds is zero on
447 * eviction, except when dbuf_fix_old_data() calls
448 * dbuf_clear_data(). However, the hold count can grow
449 * during eviction even though db_mtx is held (see
450 * dmu_bonus_hold() for an example), so we can only
451 * test the generic invariant that holds >= dirtycnt.
453 ASSERT3U(holds, >=, db->db_dirtycnt);
455 if (db->db_user_immediate_evict == TRUE)
456 ASSERT3U(holds, >=, db->db_dirtycnt);
458 ASSERT3U(holds, >, 0);
464 dbuf_evict_user(dmu_buf_impl_t *db)
466 dmu_buf_user_t *dbu = db->db_user;
468 ASSERT(MUTEX_HELD(&db->db_mtx));
473 dbuf_verify_user(db, DBVU_EVICTING);
477 if (dbu->dbu_clear_on_evict_dbufp != NULL)
478 *dbu->dbu_clear_on_evict_dbufp = NULL;
482 * There are two eviction callbacks - one that we call synchronously
483 * and one that we invoke via a taskq. The async one is useful for
484 * avoiding lock order reversals and limiting stack depth.
486 * Note that if we have a sync callback but no async callback,
487 * it's likely that the sync callback will free the structure
488 * containing the dbu. In that case we need to take care to not
489 * dereference dbu after calling the sync evict func.
491 boolean_t has_async = (dbu->dbu_evict_func_async != NULL);
493 if (dbu->dbu_evict_func_sync != NULL)
494 dbu->dbu_evict_func_sync(dbu);
497 taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func_async,
498 dbu, 0, &dbu->dbu_tqent);
503 dbuf_is_metadata(dmu_buf_impl_t *db)
506 * Consider indirect blocks and spill blocks to be meta data.
508 if (db->db_level > 0 || db->db_blkid == DMU_SPILL_BLKID) {
511 boolean_t is_metadata;
514 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type);
517 return (is_metadata);
523 * This function *must* return indices evenly distributed between all
524 * sublists of the multilist. This is needed due to how the dbuf eviction
525 * code is laid out; dbuf_evict_thread() assumes dbufs are evenly
526 * distributed between all sublists and uses this assumption when
527 * deciding which sublist to evict from and how much to evict from it.
530 dbuf_cache_multilist_index_func(multilist_t *ml, void *obj)
532 dmu_buf_impl_t *db = obj;
535 * The assumption here, is the hash value for a given
536 * dmu_buf_impl_t will remain constant throughout it's lifetime
537 * (i.e. it's objset, object, level and blkid fields don't change).
538 * Thus, we don't need to store the dbuf's sublist index
539 * on insertion, as this index can be recalculated on removal.
541 * Also, the low order bits of the hash value are thought to be
542 * distributed evenly. Otherwise, in the case that the multilist
543 * has a power of two number of sublists, each sublists' usage
544 * would not be evenly distributed.
546 return (dbuf_hash(db->db_objset, db->db.db_object,
547 db->db_level, db->db_blkid) %
548 multilist_get_num_sublists(ml));
551 static inline unsigned long
552 dbuf_cache_target_bytes(void)
554 return MIN(dbuf_cache_max_bytes,
555 arc_target_bytes() >> dbuf_cache_shift);
558 static inline uint64_t
559 dbuf_cache_hiwater_bytes(void)
561 uint64_t dbuf_cache_target = dbuf_cache_target_bytes();
562 return (dbuf_cache_target +
563 (dbuf_cache_target * dbuf_cache_hiwater_pct) / 100);
566 static inline uint64_t
567 dbuf_cache_lowater_bytes(void)
569 uint64_t dbuf_cache_target = dbuf_cache_target_bytes();
570 return (dbuf_cache_target -
571 (dbuf_cache_target * dbuf_cache_lowater_pct) / 100);
574 static inline boolean_t
575 dbuf_cache_above_hiwater(void)
577 return (refcount_count(&dbuf_cache_size) > dbuf_cache_hiwater_bytes());
580 static inline boolean_t
581 dbuf_cache_above_lowater(void)
583 return (refcount_count(&dbuf_cache_size) > dbuf_cache_lowater_bytes());
587 * Evict the oldest eligible dbuf from the dbuf cache.
592 int idx = multilist_get_random_index(dbuf_cache);
593 multilist_sublist_t *mls = multilist_sublist_lock(dbuf_cache, idx);
595 ASSERT(!MUTEX_HELD(&dbuf_evict_lock));
597 dmu_buf_impl_t *db = multilist_sublist_tail(mls);
598 while (db != NULL && mutex_tryenter(&db->db_mtx) == 0) {
599 db = multilist_sublist_prev(mls, db);
602 DTRACE_PROBE2(dbuf__evict__one, dmu_buf_impl_t *, db,
603 multilist_sublist_t *, mls);
606 multilist_sublist_remove(mls, db);
607 multilist_sublist_unlock(mls);
608 (void) refcount_remove_many(&dbuf_cache_size,
610 DBUF_STAT_BUMPDOWN(cache_levels[db->db_level]);
611 DBUF_STAT_BUMPDOWN(cache_count);
612 DBUF_STAT_DECR(cache_levels_bytes[db->db_level],
615 DBUF_STAT_MAX(cache_size_bytes_max,
616 refcount_count(&dbuf_cache_size));
617 DBUF_STAT_BUMP(cache_total_evicts);
619 multilist_sublist_unlock(mls);
624 * The dbuf evict thread is responsible for aging out dbufs from the
625 * cache. Once the cache has reached it's maximum size, dbufs are removed
626 * and destroyed. The eviction thread will continue running until the size
627 * of the dbuf cache is at or below the maximum size. Once the dbuf is aged
628 * out of the cache it is destroyed and becomes eligible for arc eviction.
632 dbuf_evict_thread(void *unused)
636 CALLB_CPR_INIT(&cpr, &dbuf_evict_lock, callb_generic_cpr, FTAG);
638 mutex_enter(&dbuf_evict_lock);
639 while (!dbuf_evict_thread_exit) {
640 while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
641 CALLB_CPR_SAFE_BEGIN(&cpr);
642 (void) cv_timedwait_sig_hires(&dbuf_evict_cv,
643 &dbuf_evict_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
644 CALLB_CPR_SAFE_END(&cpr, &dbuf_evict_lock);
646 mutex_exit(&dbuf_evict_lock);
649 * Keep evicting as long as we're above the low water mark
650 * for the cache. We do this without holding the locks to
651 * minimize lock contention.
653 while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
657 mutex_enter(&dbuf_evict_lock);
660 dbuf_evict_thread_exit = B_FALSE;
661 cv_broadcast(&dbuf_evict_cv);
662 CALLB_CPR_EXIT(&cpr); /* drops dbuf_evict_lock */
667 * Wake up the dbuf eviction thread if the dbuf cache is at its max size.
668 * If the dbuf cache is at its high water mark, then evict a dbuf from the
669 * dbuf cache using the callers context.
672 dbuf_evict_notify(void)
675 * We check if we should evict without holding the dbuf_evict_lock,
676 * because it's OK to occasionally make the wrong decision here,
677 * and grabbing the lock results in massive lock contention.
679 if (refcount_count(&dbuf_cache_size) > dbuf_cache_target_bytes()) {
680 if (dbuf_cache_above_hiwater())
682 cv_signal(&dbuf_evict_cv);
687 dbuf_kstat_update(kstat_t *ksp, int rw)
689 dbuf_stats_t *ds = ksp->ks_data;
691 if (rw == KSTAT_WRITE) {
692 return (SET_ERROR(EACCES));
694 ds->cache_size_bytes.value.ui64 =
695 refcount_count(&dbuf_cache_size);
696 ds->cache_target_bytes.value.ui64 = dbuf_cache_target_bytes();
697 ds->cache_hiwater_bytes.value.ui64 = dbuf_cache_hiwater_bytes();
698 ds->cache_lowater_bytes.value.ui64 = dbuf_cache_lowater_bytes();
699 ds->hash_elements.value.ui64 = dbuf_hash_count;
708 uint64_t hsize = 1ULL << 16;
709 dbuf_hash_table_t *h = &dbuf_hash_table;
713 * The hash table is big enough to fill all of physical memory
714 * with an average block size of zfs_arc_average_blocksize (default 8K).
715 * By default, the table will take up
716 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
718 while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE)
722 h->hash_table_mask = hsize - 1;
725 * Large allocations which do not require contiguous pages
726 * should be using vmem_alloc() in the linux kernel
728 h->hash_table = vmem_zalloc(hsize * sizeof (void *), KM_SLEEP);
730 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
732 if (h->hash_table == NULL) {
733 /* XXX - we should really return an error instead of assert */
734 ASSERT(hsize > (1ULL << 10));
739 dbuf_kmem_cache = kmem_cache_create("dmu_buf_impl_t",
740 sizeof (dmu_buf_impl_t),
741 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
743 for (i = 0; i < DBUF_MUTEXES; i++)
744 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
749 * Setup the parameters for the dbuf cache. We set the size of the
750 * dbuf cache to 1/32nd (default) of the target size of the ARC. If
751 * the value has been specified as a module option and it's not
752 * greater than the target size of the ARC, then we honor that value.
754 if (dbuf_cache_max_bytes == 0 ||
755 dbuf_cache_max_bytes >= arc_target_bytes()) {
756 dbuf_cache_max_bytes = arc_target_bytes() >> dbuf_cache_shift;
760 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
761 * configuration is not required.
763 dbu_evict_taskq = taskq_create("dbu_evict", 1, defclsyspri, 0, 0, 0);
765 dbuf_cache = multilist_create(sizeof (dmu_buf_impl_t),
766 offsetof(dmu_buf_impl_t, db_cache_link),
767 dbuf_cache_multilist_index_func);
768 refcount_create(&dbuf_cache_size);
770 dbuf_evict_thread_exit = B_FALSE;
771 mutex_init(&dbuf_evict_lock, NULL, MUTEX_DEFAULT, NULL);
772 cv_init(&dbuf_evict_cv, NULL, CV_DEFAULT, NULL);
773 dbuf_cache_evict_thread = thread_create(NULL, 0, dbuf_evict_thread,
774 NULL, 0, &p0, TS_RUN, minclsyspri);
776 dbuf_ksp = kstat_create("zfs", 0, "dbufstats", "misc",
777 KSTAT_TYPE_NAMED, sizeof (dbuf_stats) / sizeof (kstat_named_t),
779 if (dbuf_ksp != NULL) {
780 dbuf_ksp->ks_data = &dbuf_stats;
781 dbuf_ksp->ks_update = dbuf_kstat_update;
782 kstat_install(dbuf_ksp);
784 for (i = 0; i < DN_MAX_LEVELS; i++) {
785 snprintf(dbuf_stats.cache_levels[i].name,
786 KSTAT_STRLEN, "cache_level_%d", i);
787 dbuf_stats.cache_levels[i].data_type =
789 snprintf(dbuf_stats.cache_levels_bytes[i].name,
790 KSTAT_STRLEN, "cache_level_%d_bytes", i);
791 dbuf_stats.cache_levels_bytes[i].data_type =
800 dbuf_hash_table_t *h = &dbuf_hash_table;
803 dbuf_stats_destroy();
805 for (i = 0; i < DBUF_MUTEXES; i++)
806 mutex_destroy(&h->hash_mutexes[i]);
809 * Large allocations which do not require contiguous pages
810 * should be using vmem_free() in the linux kernel
812 vmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
814 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
816 kmem_cache_destroy(dbuf_kmem_cache);
817 taskq_destroy(dbu_evict_taskq);
819 mutex_enter(&dbuf_evict_lock);
820 dbuf_evict_thread_exit = B_TRUE;
821 while (dbuf_evict_thread_exit) {
822 cv_signal(&dbuf_evict_cv);
823 cv_wait(&dbuf_evict_cv, &dbuf_evict_lock);
825 mutex_exit(&dbuf_evict_lock);
827 mutex_destroy(&dbuf_evict_lock);
828 cv_destroy(&dbuf_evict_cv);
830 refcount_destroy(&dbuf_cache_size);
831 multilist_destroy(dbuf_cache);
833 if (dbuf_ksp != NULL) {
834 kstat_delete(dbuf_ksp);
845 dbuf_verify(dmu_buf_impl_t *db)
848 dbuf_dirty_record_t *dr;
850 ASSERT(MUTEX_HELD(&db->db_mtx));
852 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
855 ASSERT(db->db_objset != NULL);
859 ASSERT(db->db_parent == NULL);
860 ASSERT(db->db_blkptr == NULL);
862 ASSERT3U(db->db.db_object, ==, dn->dn_object);
863 ASSERT3P(db->db_objset, ==, dn->dn_objset);
864 ASSERT3U(db->db_level, <, dn->dn_nlevels);
865 ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
866 db->db_blkid == DMU_SPILL_BLKID ||
867 !avl_is_empty(&dn->dn_dbufs));
869 if (db->db_blkid == DMU_BONUS_BLKID) {
871 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
872 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
873 } else if (db->db_blkid == DMU_SPILL_BLKID) {
875 ASSERT0(db->db.db_offset);
877 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
880 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
881 ASSERT(dr->dr_dbuf == db);
883 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
884 ASSERT(dr->dr_dbuf == db);
887 * We can't assert that db_size matches dn_datablksz because it
888 * can be momentarily different when another thread is doing
891 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
892 dr = db->db_data_pending;
894 * It should only be modified in syncing context, so
895 * make sure we only have one copy of the data.
897 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
900 /* verify db->db_blkptr */
902 if (db->db_parent == dn->dn_dbuf) {
903 /* db is pointed to by the dnode */
904 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
905 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
906 ASSERT(db->db_parent == NULL);
908 ASSERT(db->db_parent != NULL);
909 if (db->db_blkid != DMU_SPILL_BLKID)
910 ASSERT3P(db->db_blkptr, ==,
911 &dn->dn_phys->dn_blkptr[db->db_blkid]);
913 /* db is pointed to by an indirect block */
914 ASSERTV(int epb = db->db_parent->db.db_size >>
916 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
917 ASSERT3U(db->db_parent->db.db_object, ==,
920 * dnode_grow_indblksz() can make this fail if we don't
921 * have the struct_rwlock. XXX indblksz no longer
922 * grows. safe to do this now?
924 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
925 ASSERT3P(db->db_blkptr, ==,
926 ((blkptr_t *)db->db_parent->db.db_data +
927 db->db_blkid % epb));
931 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
932 (db->db_buf == NULL || db->db_buf->b_data) &&
933 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
934 db->db_state != DB_FILL && !dn->dn_free_txg) {
936 * If the blkptr isn't set but they have nonzero data,
937 * it had better be dirty, otherwise we'll lose that
938 * data when we evict this buffer.
940 * There is an exception to this rule for indirect blocks; in
941 * this case, if the indirect block is a hole, we fill in a few
942 * fields on each of the child blocks (importantly, birth time)
943 * to prevent hole birth times from being lost when you
944 * partially fill in a hole.
946 if (db->db_dirtycnt == 0) {
947 if (db->db_level == 0) {
948 uint64_t *buf = db->db.db_data;
951 for (i = 0; i < db->db.db_size >> 3; i++) {
955 blkptr_t *bps = db->db.db_data;
956 ASSERT3U(1 << DB_DNODE(db)->dn_indblkshift, ==,
959 * We want to verify that all the blkptrs in the
960 * indirect block are holes, but we may have
961 * automatically set up a few fields for them.
962 * We iterate through each blkptr and verify
963 * they only have those fields set.
966 i < db->db.db_size / sizeof (blkptr_t);
968 blkptr_t *bp = &bps[i];
969 ASSERT(ZIO_CHECKSUM_IS_ZERO(
972 DVA_IS_EMPTY(&bp->blk_dva[0]) &&
973 DVA_IS_EMPTY(&bp->blk_dva[1]) &&
974 DVA_IS_EMPTY(&bp->blk_dva[2]));
975 ASSERT0(bp->blk_fill);
976 ASSERT0(bp->blk_pad[0]);
977 ASSERT0(bp->blk_pad[1]);
978 ASSERT(!BP_IS_EMBEDDED(bp));
979 ASSERT(BP_IS_HOLE(bp));
980 ASSERT0(bp->blk_phys_birth);
990 dbuf_clear_data(dmu_buf_impl_t *db)
992 ASSERT(MUTEX_HELD(&db->db_mtx));
994 ASSERT3P(db->db_buf, ==, NULL);
995 db->db.db_data = NULL;
996 if (db->db_state != DB_NOFILL)
997 db->db_state = DB_UNCACHED;
1001 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
1003 ASSERT(MUTEX_HELD(&db->db_mtx));
1004 ASSERT(buf != NULL);
1007 ASSERT(buf->b_data != NULL);
1008 db->db.db_data = buf->b_data;
1012 * Loan out an arc_buf for read. Return the loaned arc_buf.
1015 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
1019 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1020 mutex_enter(&db->db_mtx);
1021 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
1022 int blksz = db->db.db_size;
1023 spa_t *spa = db->db_objset->os_spa;
1025 mutex_exit(&db->db_mtx);
1026 abuf = arc_loan_buf(spa, B_FALSE, blksz);
1027 bcopy(db->db.db_data, abuf->b_data, blksz);
1030 arc_loan_inuse_buf(abuf, db);
1032 dbuf_clear_data(db);
1033 mutex_exit(&db->db_mtx);
1039 * Calculate which level n block references the data at the level 0 offset
1043 dbuf_whichblock(const dnode_t *dn, const int64_t level, const uint64_t offset)
1045 if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) {
1047 * The level n blkid is equal to the level 0 blkid divided by
1048 * the number of level 0s in a level n block.
1050 * The level 0 blkid is offset >> datablkshift =
1051 * offset / 2^datablkshift.
1053 * The number of level 0s in a level n is the number of block
1054 * pointers in an indirect block, raised to the power of level.
1055 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
1056 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
1058 * Thus, the level n blkid is: offset /
1059 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
1060 * = offset / 2^(datablkshift + level *
1061 * (indblkshift - SPA_BLKPTRSHIFT))
1062 * = offset >> (datablkshift + level *
1063 * (indblkshift - SPA_BLKPTRSHIFT))
1066 const unsigned exp = dn->dn_datablkshift +
1067 level * (dn->dn_indblkshift - SPA_BLKPTRSHIFT);
1069 if (exp >= 8 * sizeof (offset)) {
1070 /* This only happens on the highest indirection level */
1071 ASSERT3U(level, ==, dn->dn_nlevels - 1);
1075 ASSERT3U(exp, <, 8 * sizeof (offset));
1077 return (offset >> exp);
1079 ASSERT3U(offset, <, dn->dn_datablksz);
1085 dbuf_read_done(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp,
1086 arc_buf_t *buf, void *vdb)
1088 dmu_buf_impl_t *db = vdb;
1090 mutex_enter(&db->db_mtx);
1091 ASSERT3U(db->db_state, ==, DB_READ);
1093 * All reads are synchronous, so we must have a hold on the dbuf
1095 ASSERT(refcount_count(&db->db_holds) > 0);
1096 ASSERT(db->db_buf == NULL);
1097 ASSERT(db->db.db_data == NULL);
1098 if (db->db_level == 0 && db->db_freed_in_flight) {
1099 /* we were freed in flight; disregard any error */
1101 buf = arc_alloc_buf(db->db_objset->os_spa,
1102 db, DBUF_GET_BUFC_TYPE(db), db->db.db_size);
1104 arc_release(buf, db);
1105 bzero(buf->b_data, db->db.db_size);
1106 arc_buf_freeze(buf);
1107 db->db_freed_in_flight = FALSE;
1108 dbuf_set_data(db, buf);
1109 db->db_state = DB_CACHED;
1110 } else if (buf != NULL) {
1111 dbuf_set_data(db, buf);
1112 db->db_state = DB_CACHED;
1114 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1115 ASSERT3P(db->db_buf, ==, NULL);
1116 db->db_state = DB_UNCACHED;
1118 cv_broadcast(&db->db_changed);
1119 dbuf_rele_and_unlock(db, NULL, B_FALSE);
1123 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
1126 zbookmark_phys_t zb;
1127 uint32_t aflags = ARC_FLAG_NOWAIT;
1128 int err, zio_flags = 0;
1132 ASSERT(!refcount_is_zero(&db->db_holds));
1133 /* We need the struct_rwlock to prevent db_blkptr from changing. */
1134 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1135 ASSERT(MUTEX_HELD(&db->db_mtx));
1136 ASSERT(db->db_state == DB_UNCACHED);
1137 ASSERT(db->db_buf == NULL);
1139 if (db->db_blkid == DMU_BONUS_BLKID) {
1141 * The bonus length stored in the dnode may be less than
1142 * the maximum available space in the bonus buffer.
1144 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
1145 int max_bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
1146 arc_buf_t *dn_buf = (dn->dn_dbuf != NULL) ?
1147 dn->dn_dbuf->db_buf : NULL;
1149 /* if the underlying dnode block is encrypted, decrypt it */
1150 if (dn_buf != NULL && dn->dn_objset->os_encrypted &&
1151 DMU_OT_IS_ENCRYPTED(dn->dn_bonustype) &&
1152 (flags & DB_RF_NO_DECRYPT) == 0 &&
1153 arc_is_encrypted(dn_buf)) {
1154 SET_BOOKMARK(&zb, dmu_objset_id(db->db_objset),
1155 DMU_META_DNODE_OBJECT, 0, dn->dn_dbuf->db_blkid);
1156 err = arc_untransform(dn_buf, dn->dn_objset->os_spa,
1160 mutex_exit(&db->db_mtx);
1165 ASSERT3U(bonuslen, <=, db->db.db_size);
1166 db->db.db_data = kmem_alloc(max_bonuslen, KM_SLEEP);
1167 arc_space_consume(max_bonuslen, ARC_SPACE_BONUS);
1168 if (bonuslen < max_bonuslen)
1169 bzero(db->db.db_data, max_bonuslen);
1171 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
1173 db->db_state = DB_CACHED;
1174 mutex_exit(&db->db_mtx);
1179 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
1180 * processes the delete record and clears the bp while we are waiting
1181 * for the dn_mtx (resulting in a "no" from block_freed).
1183 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
1184 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
1185 BP_IS_HOLE(db->db_blkptr)))) {
1186 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1188 dbuf_set_data(db, arc_alloc_buf(db->db_objset->os_spa, db, type,
1190 bzero(db->db.db_data, db->db.db_size);
1192 if (db->db_blkptr != NULL && db->db_level > 0 &&
1193 BP_IS_HOLE(db->db_blkptr) &&
1194 db->db_blkptr->blk_birth != 0) {
1195 blkptr_t *bps = db->db.db_data;
1196 for (int i = 0; i < ((1 <<
1197 DB_DNODE(db)->dn_indblkshift) / sizeof (blkptr_t));
1199 blkptr_t *bp = &bps[i];
1200 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
1201 1 << dn->dn_indblkshift);
1203 BP_GET_LEVEL(db->db_blkptr) == 1 ?
1205 BP_GET_LSIZE(db->db_blkptr));
1206 BP_SET_TYPE(bp, BP_GET_TYPE(db->db_blkptr));
1208 BP_GET_LEVEL(db->db_blkptr) - 1);
1209 BP_SET_BIRTH(bp, db->db_blkptr->blk_birth, 0);
1213 db->db_state = DB_CACHED;
1214 mutex_exit(&db->db_mtx);
1220 db->db_state = DB_READ;
1221 mutex_exit(&db->db_mtx);
1223 if (DBUF_IS_L2CACHEABLE(db))
1224 aflags |= ARC_FLAG_L2CACHE;
1226 SET_BOOKMARK(&zb, dmu_objset_id(db->db_objset),
1227 db->db.db_object, db->db_level, db->db_blkid);
1230 * All bps of an encrypted os should have the encryption bit set.
1231 * If this is not true it indicates tampering and we report an error.
1233 if (db->db_objset->os_encrypted && !BP_USES_CRYPT(db->db_blkptr)) {
1234 spa_log_error(db->db_objset->os_spa, &zb);
1235 zfs_panic_recover("unencrypted block in encrypted "
1236 "object set %llu", dmu_objset_id(db->db_objset));
1237 return (SET_ERROR(EIO));
1240 dbuf_add_ref(db, NULL);
1242 zio_flags = (flags & DB_RF_CANFAIL) ?
1243 ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED;
1245 if ((flags & DB_RF_NO_DECRYPT) && BP_IS_PROTECTED(db->db_blkptr))
1246 zio_flags |= ZIO_FLAG_RAW;
1248 err = arc_read(zio, db->db_objset->os_spa, db->db_blkptr,
1249 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ, zio_flags,
1256 * This is our just-in-time copy function. It makes a copy of buffers that
1257 * have been modified in a previous transaction group before we access them in
1258 * the current active group.
1260 * This function is used in three places: when we are dirtying a buffer for the
1261 * first time in a txg, when we are freeing a range in a dnode that includes
1262 * this buffer, and when we are accessing a buffer which was received compressed
1263 * and later referenced in a WRITE_BYREF record.
1265 * Note that when we are called from dbuf_free_range() we do not put a hold on
1266 * the buffer, we just traverse the active dbuf list for the dnode.
1269 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
1271 dbuf_dirty_record_t *dr = db->db_last_dirty;
1273 ASSERT(MUTEX_HELD(&db->db_mtx));
1274 ASSERT(db->db.db_data != NULL);
1275 ASSERT(db->db_level == 0);
1276 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
1279 (dr->dt.dl.dr_data !=
1280 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
1284 * If the last dirty record for this dbuf has not yet synced
1285 * and its referencing the dbuf data, either:
1286 * reset the reference to point to a new copy,
1287 * or (if there a no active holders)
1288 * just null out the current db_data pointer.
1290 ASSERT3U(dr->dr_txg, >=, txg - 2);
1291 if (db->db_blkid == DMU_BONUS_BLKID) {
1292 dnode_t *dn = DB_DNODE(db);
1293 int bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
1294 dr->dt.dl.dr_data = kmem_alloc(bonuslen, KM_SLEEP);
1295 arc_space_consume(bonuslen, ARC_SPACE_BONUS);
1296 bcopy(db->db.db_data, dr->dt.dl.dr_data, bonuslen);
1297 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
1298 dnode_t *dn = DB_DNODE(db);
1299 int size = arc_buf_size(db->db_buf);
1300 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1301 spa_t *spa = db->db_objset->os_spa;
1302 enum zio_compress compress_type =
1303 arc_get_compression(db->db_buf);
1305 if (arc_is_encrypted(db->db_buf)) {
1306 boolean_t byteorder;
1307 uint8_t salt[ZIO_DATA_SALT_LEN];
1308 uint8_t iv[ZIO_DATA_IV_LEN];
1309 uint8_t mac[ZIO_DATA_MAC_LEN];
1311 arc_get_raw_params(db->db_buf, &byteorder, salt,
1313 dr->dt.dl.dr_data = arc_alloc_raw_buf(spa, db,
1314 dmu_objset_id(dn->dn_objset), byteorder, salt, iv,
1315 mac, dn->dn_type, size, arc_buf_lsize(db->db_buf),
1317 } else if (compress_type != ZIO_COMPRESS_OFF) {
1318 ASSERT3U(type, ==, ARC_BUFC_DATA);
1319 dr->dt.dl.dr_data = arc_alloc_compressed_buf(spa, db,
1320 size, arc_buf_lsize(db->db_buf), compress_type);
1322 dr->dt.dl.dr_data = arc_alloc_buf(spa, db, type, size);
1324 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
1327 dbuf_clear_data(db);
1332 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
1339 * We don't have to hold the mutex to check db_state because it
1340 * can't be freed while we have a hold on the buffer.
1342 ASSERT(!refcount_is_zero(&db->db_holds));
1344 if (db->db_state == DB_NOFILL)
1345 return (SET_ERROR(EIO));
1349 if ((flags & DB_RF_HAVESTRUCT) == 0)
1350 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1352 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1353 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
1354 DBUF_IS_CACHEABLE(db);
1356 mutex_enter(&db->db_mtx);
1357 if (db->db_state == DB_CACHED) {
1358 spa_t *spa = dn->dn_objset->os_spa;
1361 * If the arc buf is compressed or encrypted, we need to
1362 * untransform it to read the data. This could happen during
1363 * the "zfs receive" of a stream which is deduplicated and
1364 * either raw or compressed. We do not need to do this if the
1365 * caller wants raw encrypted data.
1367 if (db->db_buf != NULL && (flags & DB_RF_NO_DECRYPT) == 0 &&
1368 (arc_is_encrypted(db->db_buf) ||
1369 arc_get_compression(db->db_buf) != ZIO_COMPRESS_OFF)) {
1370 zbookmark_phys_t zb;
1372 SET_BOOKMARK(&zb, dmu_objset_id(db->db_objset),
1373 db->db.db_object, db->db_level, db->db_blkid);
1374 dbuf_fix_old_data(db, spa_syncing_txg(spa));
1375 err = arc_untransform(db->db_buf, spa, &zb, B_FALSE);
1376 dbuf_set_data(db, db->db_buf);
1378 mutex_exit(&db->db_mtx);
1380 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1381 if ((flags & DB_RF_HAVESTRUCT) == 0)
1382 rw_exit(&dn->dn_struct_rwlock);
1384 DBUF_STAT_BUMP(hash_hits);
1385 } else if (db->db_state == DB_UNCACHED) {
1386 spa_t *spa = dn->dn_objset->os_spa;
1387 boolean_t need_wait = B_FALSE;
1390 db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) {
1391 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
1394 err = dbuf_read_impl(db, zio, flags);
1396 /* dbuf_read_impl has dropped db_mtx for us */
1398 if (!err && prefetch)
1399 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1401 if ((flags & DB_RF_HAVESTRUCT) == 0)
1402 rw_exit(&dn->dn_struct_rwlock);
1404 DBUF_STAT_BUMP(hash_misses);
1406 if (!err && need_wait)
1407 err = zio_wait(zio);
1410 * Another reader came in while the dbuf was in flight
1411 * between UNCACHED and CACHED. Either a writer will finish
1412 * writing the buffer (sending the dbuf to CACHED) or the
1413 * first reader's request will reach the read_done callback
1414 * and send the dbuf to CACHED. Otherwise, a failure
1415 * occurred and the dbuf went to UNCACHED.
1417 mutex_exit(&db->db_mtx);
1419 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1420 if ((flags & DB_RF_HAVESTRUCT) == 0)
1421 rw_exit(&dn->dn_struct_rwlock);
1423 DBUF_STAT_BUMP(hash_misses);
1425 /* Skip the wait per the caller's request. */
1426 mutex_enter(&db->db_mtx);
1427 if ((flags & DB_RF_NEVERWAIT) == 0) {
1428 while (db->db_state == DB_READ ||
1429 db->db_state == DB_FILL) {
1430 ASSERT(db->db_state == DB_READ ||
1431 (flags & DB_RF_HAVESTRUCT) == 0);
1432 DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *,
1434 cv_wait(&db->db_changed, &db->db_mtx);
1436 if (db->db_state == DB_UNCACHED)
1437 err = SET_ERROR(EIO);
1439 mutex_exit(&db->db_mtx);
1446 dbuf_noread(dmu_buf_impl_t *db)
1448 ASSERT(!refcount_is_zero(&db->db_holds));
1449 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1450 mutex_enter(&db->db_mtx);
1451 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1452 cv_wait(&db->db_changed, &db->db_mtx);
1453 if (db->db_state == DB_UNCACHED) {
1454 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1455 spa_t *spa = db->db_objset->os_spa;
1457 ASSERT(db->db_buf == NULL);
1458 ASSERT(db->db.db_data == NULL);
1459 dbuf_set_data(db, arc_alloc_buf(spa, db, type, db->db.db_size));
1460 db->db_state = DB_FILL;
1461 } else if (db->db_state == DB_NOFILL) {
1462 dbuf_clear_data(db);
1464 ASSERT3U(db->db_state, ==, DB_CACHED);
1466 mutex_exit(&db->db_mtx);
1470 dbuf_unoverride(dbuf_dirty_record_t *dr)
1472 dmu_buf_impl_t *db = dr->dr_dbuf;
1473 blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
1474 uint64_t txg = dr->dr_txg;
1476 ASSERT(MUTEX_HELD(&db->db_mtx));
1478 * This assert is valid because dmu_sync() expects to be called by
1479 * a zilog's get_data while holding a range lock. This call only
1480 * comes from dbuf_dirty() callers who must also hold a range lock.
1482 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
1483 ASSERT(db->db_level == 0);
1485 if (db->db_blkid == DMU_BONUS_BLKID ||
1486 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
1489 ASSERT(db->db_data_pending != dr);
1491 /* free this block */
1492 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite)
1493 zio_free(db->db_objset->os_spa, txg, bp);
1495 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1496 dr->dt.dl.dr_nopwrite = B_FALSE;
1497 dr->dt.dl.dr_has_raw_params = B_FALSE;
1500 * Release the already-written buffer, so we leave it in
1501 * a consistent dirty state. Note that all callers are
1502 * modifying the buffer, so they will immediately do
1503 * another (redundant) arc_release(). Therefore, leave
1504 * the buf thawed to save the effort of freezing &
1505 * immediately re-thawing it.
1507 arc_release(dr->dt.dl.dr_data, db);
1511 * Evict (if its unreferenced) or clear (if its referenced) any level-0
1512 * data blocks in the free range, so that any future readers will find
1516 dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
1519 dmu_buf_impl_t *db_search;
1520 dmu_buf_impl_t *db, *db_next;
1521 uint64_t txg = tx->tx_txg;
1524 if (end_blkid > dn->dn_maxblkid &&
1525 !(start_blkid == DMU_SPILL_BLKID || end_blkid == DMU_SPILL_BLKID))
1526 end_blkid = dn->dn_maxblkid;
1527 dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid);
1529 db_search = kmem_alloc(sizeof (dmu_buf_impl_t), KM_SLEEP);
1530 db_search->db_level = 0;
1531 db_search->db_blkid = start_blkid;
1532 db_search->db_state = DB_SEARCH;
1534 mutex_enter(&dn->dn_dbufs_mtx);
1535 db = avl_find(&dn->dn_dbufs, db_search, &where);
1536 ASSERT3P(db, ==, NULL);
1538 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1540 for (; db != NULL; db = db_next) {
1541 db_next = AVL_NEXT(&dn->dn_dbufs, db);
1542 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1544 if (db->db_level != 0 || db->db_blkid > end_blkid) {
1547 ASSERT3U(db->db_blkid, >=, start_blkid);
1549 /* found a level 0 buffer in the range */
1550 mutex_enter(&db->db_mtx);
1551 if (dbuf_undirty(db, tx)) {
1552 /* mutex has been dropped and dbuf destroyed */
1556 if (db->db_state == DB_UNCACHED ||
1557 db->db_state == DB_NOFILL ||
1558 db->db_state == DB_EVICTING) {
1559 ASSERT(db->db.db_data == NULL);
1560 mutex_exit(&db->db_mtx);
1563 if (db->db_state == DB_READ || db->db_state == DB_FILL) {
1564 /* will be handled in dbuf_read_done or dbuf_rele */
1565 db->db_freed_in_flight = TRUE;
1566 mutex_exit(&db->db_mtx);
1569 if (refcount_count(&db->db_holds) == 0) {
1574 /* The dbuf is referenced */
1576 if (db->db_last_dirty != NULL) {
1577 dbuf_dirty_record_t *dr = db->db_last_dirty;
1579 if (dr->dr_txg == txg) {
1581 * This buffer is "in-use", re-adjust the file
1582 * size to reflect that this buffer may
1583 * contain new data when we sync.
1585 if (db->db_blkid != DMU_SPILL_BLKID &&
1586 db->db_blkid > dn->dn_maxblkid)
1587 dn->dn_maxblkid = db->db_blkid;
1588 dbuf_unoverride(dr);
1591 * This dbuf is not dirty in the open context.
1592 * Either uncache it (if its not referenced in
1593 * the open context) or reset its contents to
1596 dbuf_fix_old_data(db, txg);
1599 /* clear the contents if its cached */
1600 if (db->db_state == DB_CACHED) {
1601 ASSERT(db->db.db_data != NULL);
1602 arc_release(db->db_buf, db);
1603 bzero(db->db.db_data, db->db.db_size);
1604 arc_buf_freeze(db->db_buf);
1607 mutex_exit(&db->db_mtx);
1610 kmem_free(db_search, sizeof (dmu_buf_impl_t));
1611 mutex_exit(&dn->dn_dbufs_mtx);
1615 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
1617 arc_buf_t *buf, *obuf;
1618 int osize = db->db.db_size;
1619 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1622 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1627 /* XXX does *this* func really need the lock? */
1628 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1631 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1632 * is OK, because there can be no other references to the db
1633 * when we are changing its size, so no concurrent DB_FILL can
1637 * XXX we should be doing a dbuf_read, checking the return
1638 * value and returning that up to our callers
1640 dmu_buf_will_dirty(&db->db, tx);
1642 /* create the data buffer for the new block */
1643 buf = arc_alloc_buf(dn->dn_objset->os_spa, db, type, size);
1645 /* copy old block data to the new block */
1647 bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
1648 /* zero the remainder */
1650 bzero((uint8_t *)buf->b_data + osize, size - osize);
1652 mutex_enter(&db->db_mtx);
1653 dbuf_set_data(db, buf);
1654 arc_buf_destroy(obuf, db);
1655 db->db.db_size = size;
1657 if (db->db_level == 0) {
1658 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1659 db->db_last_dirty->dt.dl.dr_data = buf;
1661 mutex_exit(&db->db_mtx);
1663 dmu_objset_willuse_space(dn->dn_objset, size - osize, tx);
1668 dbuf_release_bp(dmu_buf_impl_t *db)
1670 ASSERTV(objset_t *os = db->db_objset);
1672 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1673 ASSERT(arc_released(os->os_phys_buf) ||
1674 list_link_active(&os->os_dsl_dataset->ds_synced_link));
1675 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1677 (void) arc_release(db->db_buf, db);
1681 * We already have a dirty record for this TXG, and we are being
1685 dbuf_redirty(dbuf_dirty_record_t *dr)
1687 dmu_buf_impl_t *db = dr->dr_dbuf;
1689 ASSERT(MUTEX_HELD(&db->db_mtx));
1691 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1693 * If this buffer has already been written out,
1694 * we now need to reset its state.
1696 dbuf_unoverride(dr);
1697 if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1698 db->db_state != DB_NOFILL) {
1699 /* Already released on initial dirty, so just thaw. */
1700 ASSERT(arc_released(db->db_buf));
1701 arc_buf_thaw(db->db_buf);
1706 dbuf_dirty_record_t *
1707 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1711 dbuf_dirty_record_t **drp, *dr;
1712 int drop_struct_lock = FALSE;
1713 int txgoff = tx->tx_txg & TXG_MASK;
1715 ASSERT(tx->tx_txg != 0);
1716 ASSERT(!refcount_is_zero(&db->db_holds));
1717 DMU_TX_DIRTY_BUF(tx, db);
1722 * Shouldn't dirty a regular buffer in syncing context. Private
1723 * objects may be dirtied in syncing context, but only if they
1724 * were already pre-dirtied in open context.
1727 if (dn->dn_objset->os_dsl_dataset != NULL) {
1728 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1731 ASSERT(!dmu_tx_is_syncing(tx) ||
1732 BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1733 DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1734 dn->dn_objset->os_dsl_dataset == NULL);
1735 if (dn->dn_objset->os_dsl_dataset != NULL)
1736 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, FTAG);
1739 * We make this assert for private objects as well, but after we
1740 * check if we're already dirty. They are allowed to re-dirty
1741 * in syncing context.
1743 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1744 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1745 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1747 mutex_enter(&db->db_mtx);
1749 * XXX make this true for indirects too? The problem is that
1750 * transactions created with dmu_tx_create_assigned() from
1751 * syncing context don't bother holding ahead.
1753 ASSERT(db->db_level != 0 ||
1754 db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1755 db->db_state == DB_NOFILL);
1757 mutex_enter(&dn->dn_mtx);
1759 * Don't set dirtyctx to SYNC if we're just modifying this as we
1760 * initialize the objset.
1762 if (dn->dn_dirtyctx == DN_UNDIRTIED) {
1763 if (dn->dn_objset->os_dsl_dataset != NULL) {
1764 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1767 if (!BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1768 dn->dn_dirtyctx = (dmu_tx_is_syncing(tx) ?
1769 DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1770 ASSERT(dn->dn_dirtyctx_firstset == NULL);
1771 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1773 if (dn->dn_objset->os_dsl_dataset != NULL) {
1774 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1779 if (tx->tx_txg > dn->dn_dirty_txg)
1780 dn->dn_dirty_txg = tx->tx_txg;
1781 mutex_exit(&dn->dn_mtx);
1783 if (db->db_blkid == DMU_SPILL_BLKID)
1784 dn->dn_have_spill = B_TRUE;
1787 * If this buffer is already dirty, we're done.
1789 drp = &db->db_last_dirty;
1790 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1791 db->db.db_object == DMU_META_DNODE_OBJECT);
1792 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1794 if (dr && dr->dr_txg == tx->tx_txg) {
1798 mutex_exit(&db->db_mtx);
1803 * Only valid if not already dirty.
1805 ASSERT(dn->dn_object == 0 ||
1806 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1807 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1809 ASSERT3U(dn->dn_nlevels, >, db->db_level);
1812 * We should only be dirtying in syncing context if it's the
1813 * mos or we're initializing the os or it's a special object.
1814 * However, we are allowed to dirty in syncing context provided
1815 * we already dirtied it in open context. Hence we must make
1816 * this assertion only if we're not already dirty.
1819 VERIFY3U(tx->tx_txg, <=, spa_final_dirty_txg(os->os_spa));
1821 if (dn->dn_objset->os_dsl_dataset != NULL)
1822 rrw_enter(&os->os_dsl_dataset->ds_bp_rwlock, RW_READER, FTAG);
1823 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1824 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1825 if (dn->dn_objset->os_dsl_dataset != NULL)
1826 rrw_exit(&os->os_dsl_dataset->ds_bp_rwlock, FTAG);
1828 ASSERT(db->db.db_size != 0);
1830 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1832 if (db->db_blkid != DMU_BONUS_BLKID) {
1833 dmu_objset_willuse_space(os, db->db.db_size, tx);
1837 * If this buffer is dirty in an old transaction group we need
1838 * to make a copy of it so that the changes we make in this
1839 * transaction group won't leak out when we sync the older txg.
1841 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1842 list_link_init(&dr->dr_dirty_node);
1843 if (db->db_level == 0) {
1844 void *data_old = db->db_buf;
1846 if (db->db_state != DB_NOFILL) {
1847 if (db->db_blkid == DMU_BONUS_BLKID) {
1848 dbuf_fix_old_data(db, tx->tx_txg);
1849 data_old = db->db.db_data;
1850 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1852 * Release the data buffer from the cache so
1853 * that we can modify it without impacting
1854 * possible other users of this cached data
1855 * block. Note that indirect blocks and
1856 * private objects are not released until the
1857 * syncing state (since they are only modified
1860 arc_release(db->db_buf, db);
1861 dbuf_fix_old_data(db, tx->tx_txg);
1862 data_old = db->db_buf;
1864 ASSERT(data_old != NULL);
1866 dr->dt.dl.dr_data = data_old;
1868 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_NOLOCKDEP, NULL);
1869 list_create(&dr->dt.di.dr_children,
1870 sizeof (dbuf_dirty_record_t),
1871 offsetof(dbuf_dirty_record_t, dr_dirty_node));
1873 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL)
1874 dr->dr_accounted = db->db.db_size;
1876 dr->dr_txg = tx->tx_txg;
1881 * We could have been freed_in_flight between the dbuf_noread
1882 * and dbuf_dirty. We win, as though the dbuf_noread() had
1883 * happened after the free.
1885 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1886 db->db_blkid != DMU_SPILL_BLKID) {
1887 mutex_enter(&dn->dn_mtx);
1888 if (dn->dn_free_ranges[txgoff] != NULL) {
1889 range_tree_clear(dn->dn_free_ranges[txgoff],
1892 mutex_exit(&dn->dn_mtx);
1893 db->db_freed_in_flight = FALSE;
1897 * This buffer is now part of this txg
1899 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1900 db->db_dirtycnt += 1;
1901 ASSERT3U(db->db_dirtycnt, <=, 3);
1903 mutex_exit(&db->db_mtx);
1905 if (db->db_blkid == DMU_BONUS_BLKID ||
1906 db->db_blkid == DMU_SPILL_BLKID) {
1907 mutex_enter(&dn->dn_mtx);
1908 ASSERT(!list_link_active(&dr->dr_dirty_node));
1909 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1910 mutex_exit(&dn->dn_mtx);
1911 dnode_setdirty(dn, tx);
1917 * The dn_struct_rwlock prevents db_blkptr from changing
1918 * due to a write from syncing context completing
1919 * while we are running, so we want to acquire it before
1920 * looking at db_blkptr.
1922 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1923 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1924 drop_struct_lock = TRUE;
1928 * We need to hold the dn_struct_rwlock to make this assertion,
1929 * because it protects dn_phys / dn_next_nlevels from changing.
1931 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1932 dn->dn_phys->dn_nlevels > db->db_level ||
1933 dn->dn_next_nlevels[txgoff] > db->db_level ||
1934 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1935 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1938 * If we are overwriting a dedup BP, then unless it is snapshotted,
1939 * when we get to syncing context we will need to decrement its
1940 * refcount in the DDT. Prefetch the relevant DDT block so that
1941 * syncing context won't have to wait for the i/o.
1943 ddt_prefetch(os->os_spa, db->db_blkptr);
1945 if (db->db_level == 0) {
1946 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1947 ASSERT(dn->dn_maxblkid >= db->db_blkid);
1950 if (db->db_level+1 < dn->dn_nlevels) {
1951 dmu_buf_impl_t *parent = db->db_parent;
1952 dbuf_dirty_record_t *di;
1953 int parent_held = FALSE;
1955 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1956 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1958 parent = dbuf_hold_level(dn, db->db_level+1,
1959 db->db_blkid >> epbs, FTAG);
1960 ASSERT(parent != NULL);
1963 if (drop_struct_lock)
1964 rw_exit(&dn->dn_struct_rwlock);
1965 ASSERT3U(db->db_level+1, ==, parent->db_level);
1966 di = dbuf_dirty(parent, tx);
1968 dbuf_rele(parent, FTAG);
1970 mutex_enter(&db->db_mtx);
1972 * Since we've dropped the mutex, it's possible that
1973 * dbuf_undirty() might have changed this out from under us.
1975 if (db->db_last_dirty == dr ||
1976 dn->dn_object == DMU_META_DNODE_OBJECT) {
1977 mutex_enter(&di->dt.di.dr_mtx);
1978 ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1979 ASSERT(!list_link_active(&dr->dr_dirty_node));
1980 list_insert_tail(&di->dt.di.dr_children, dr);
1981 mutex_exit(&di->dt.di.dr_mtx);
1984 mutex_exit(&db->db_mtx);
1986 ASSERT(db->db_level+1 == dn->dn_nlevels);
1987 ASSERT(db->db_blkid < dn->dn_nblkptr);
1988 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1989 mutex_enter(&dn->dn_mtx);
1990 ASSERT(!list_link_active(&dr->dr_dirty_node));
1991 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1992 mutex_exit(&dn->dn_mtx);
1993 if (drop_struct_lock)
1994 rw_exit(&dn->dn_struct_rwlock);
1997 dnode_setdirty(dn, tx);
2003 * Undirty a buffer in the transaction group referenced by the given
2004 * transaction. Return whether this evicted the dbuf.
2007 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
2010 uint64_t txg = tx->tx_txg;
2011 dbuf_dirty_record_t *dr, **drp;
2016 * Due to our use of dn_nlevels below, this can only be called
2017 * in open context, unless we are operating on the MOS.
2018 * From syncing context, dn_nlevels may be different from the
2019 * dn_nlevels used when dbuf was dirtied.
2021 ASSERT(db->db_objset ==
2022 dmu_objset_pool(db->db_objset)->dp_meta_objset ||
2023 txg != spa_syncing_txg(dmu_objset_spa(db->db_objset)));
2024 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2025 ASSERT0(db->db_level);
2026 ASSERT(MUTEX_HELD(&db->db_mtx));
2029 * If this buffer is not dirty, we're done.
2031 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
2032 if (dr->dr_txg <= txg)
2034 if (dr == NULL || dr->dr_txg < txg)
2036 ASSERT(dr->dr_txg == txg);
2037 ASSERT(dr->dr_dbuf == db);
2042 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
2044 ASSERT(db->db.db_size != 0);
2046 dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset),
2047 dr->dr_accounted, txg);
2052 * Note that there are three places in dbuf_dirty()
2053 * where this dirty record may be put on a list.
2054 * Make sure to do a list_remove corresponding to
2055 * every one of those list_insert calls.
2057 if (dr->dr_parent) {
2058 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
2059 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
2060 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
2061 } else if (db->db_blkid == DMU_SPILL_BLKID ||
2062 db->db_level + 1 == dn->dn_nlevels) {
2063 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
2064 mutex_enter(&dn->dn_mtx);
2065 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
2066 mutex_exit(&dn->dn_mtx);
2070 if (db->db_state != DB_NOFILL) {
2071 dbuf_unoverride(dr);
2073 ASSERT(db->db_buf != NULL);
2074 ASSERT(dr->dt.dl.dr_data != NULL);
2075 if (dr->dt.dl.dr_data != db->db_buf)
2076 arc_buf_destroy(dr->dt.dl.dr_data, db);
2079 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2081 ASSERT(db->db_dirtycnt > 0);
2082 db->db_dirtycnt -= 1;
2084 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
2085 ASSERT(db->db_state == DB_NOFILL || arc_released(db->db_buf));
2094 dmu_buf_will_dirty_impl(dmu_buf_t *db_fake, int flags, dmu_tx_t *tx)
2096 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2098 ASSERT(tx->tx_txg != 0);
2099 ASSERT(!refcount_is_zero(&db->db_holds));
2102 * Quick check for dirtyness. For already dirty blocks, this
2103 * reduces runtime of this function by >90%, and overall performance
2104 * by 50% for some workloads (e.g. file deletion with indirect blocks
2107 mutex_enter(&db->db_mtx);
2109 dbuf_dirty_record_t *dr;
2110 for (dr = db->db_last_dirty;
2111 dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) {
2113 * It's possible that it is already dirty but not cached,
2114 * because there are some calls to dbuf_dirty() that don't
2115 * go through dmu_buf_will_dirty().
2117 if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) {
2118 /* This dbuf is already dirty and cached. */
2120 mutex_exit(&db->db_mtx);
2124 mutex_exit(&db->db_mtx);
2127 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
2128 flags |= DB_RF_HAVESTRUCT;
2130 (void) dbuf_read(db, NULL, flags);
2131 (void) dbuf_dirty(db, tx);
2135 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
2137 dmu_buf_will_dirty_impl(db_fake,
2138 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH, tx);
2142 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
2144 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2146 db->db_state = DB_NOFILL;
2148 dmu_buf_will_fill(db_fake, tx);
2152 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
2154 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2156 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2157 ASSERT(tx->tx_txg != 0);
2158 ASSERT(db->db_level == 0);
2159 ASSERT(!refcount_is_zero(&db->db_holds));
2161 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
2162 dmu_tx_private_ok(tx));
2165 (void) dbuf_dirty(db, tx);
2169 * This function is effectively the same as dmu_buf_will_dirty(), but
2170 * indicates the caller expects raw encrypted data in the db, and provides
2171 * the crypt params (byteorder, salt, iv, mac) which should be stored in the
2172 * blkptr_t when this dbuf is written. This is only used for blocks of
2173 * dnodes, during raw receive.
2176 dmu_buf_set_crypt_params(dmu_buf_t *db_fake, boolean_t byteorder,
2177 const uint8_t *salt, const uint8_t *iv, const uint8_t *mac, dmu_tx_t *tx)
2179 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2180 dbuf_dirty_record_t *dr;
2183 * dr_has_raw_params is only processed for blocks of dnodes
2184 * (see dbuf_sync_dnode_leaf_crypt()).
2186 ASSERT3U(db->db.db_object, ==, DMU_META_DNODE_OBJECT);
2187 ASSERT3U(db->db_level, ==, 0);
2188 ASSERT(db->db_objset->os_raw_receive);
2190 dmu_buf_will_dirty_impl(db_fake,
2191 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_NO_DECRYPT, tx);
2193 dr = db->db_last_dirty;
2194 while (dr != NULL && dr->dr_txg > tx->tx_txg)
2197 ASSERT3P(dr, !=, NULL);
2198 ASSERT3U(dr->dr_txg, ==, tx->tx_txg);
2200 dr->dt.dl.dr_has_raw_params = B_TRUE;
2201 dr->dt.dl.dr_byteorder = byteorder;
2202 bcopy(salt, dr->dt.dl.dr_salt, ZIO_DATA_SALT_LEN);
2203 bcopy(iv, dr->dt.dl.dr_iv, ZIO_DATA_IV_LEN);
2204 bcopy(mac, dr->dt.dl.dr_mac, ZIO_DATA_MAC_LEN);
2207 #pragma weak dmu_buf_fill_done = dbuf_fill_done
2210 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
2212 mutex_enter(&db->db_mtx);
2215 if (db->db_state == DB_FILL) {
2216 if (db->db_level == 0 && db->db_freed_in_flight) {
2217 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2218 /* we were freed while filling */
2219 /* XXX dbuf_undirty? */
2220 bzero(db->db.db_data, db->db.db_size);
2221 db->db_freed_in_flight = FALSE;
2223 db->db_state = DB_CACHED;
2224 cv_broadcast(&db->db_changed);
2226 mutex_exit(&db->db_mtx);
2230 dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data,
2231 bp_embedded_type_t etype, enum zio_compress comp,
2232 int uncompressed_size, int compressed_size, int byteorder,
2235 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2236 struct dirty_leaf *dl;
2237 dmu_object_type_t type;
2239 if (etype == BP_EMBEDDED_TYPE_DATA) {
2240 ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset),
2241 SPA_FEATURE_EMBEDDED_DATA));
2245 type = DB_DNODE(db)->dn_type;
2248 ASSERT0(db->db_level);
2249 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2251 dmu_buf_will_not_fill(dbuf, tx);
2253 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
2254 dl = &db->db_last_dirty->dt.dl;
2255 encode_embedded_bp_compressed(&dl->dr_overridden_by,
2256 data, comp, uncompressed_size, compressed_size);
2257 BPE_SET_ETYPE(&dl->dr_overridden_by, etype);
2258 BP_SET_TYPE(&dl->dr_overridden_by, type);
2259 BP_SET_LEVEL(&dl->dr_overridden_by, 0);
2260 BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder);
2262 dl->dr_override_state = DR_OVERRIDDEN;
2263 dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg;
2267 * Directly assign a provided arc buf to a given dbuf if it's not referenced
2268 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
2271 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
2273 ASSERT(!refcount_is_zero(&db->db_holds));
2274 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2275 ASSERT(db->db_level == 0);
2276 ASSERT3U(dbuf_is_metadata(db), ==, arc_is_metadata(buf));
2277 ASSERT(buf != NULL);
2278 ASSERT(arc_buf_lsize(buf) == db->db.db_size);
2279 ASSERT(tx->tx_txg != 0);
2281 arc_return_buf(buf, db);
2282 ASSERT(arc_released(buf));
2284 mutex_enter(&db->db_mtx);
2286 while (db->db_state == DB_READ || db->db_state == DB_FILL)
2287 cv_wait(&db->db_changed, &db->db_mtx);
2289 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
2291 if (db->db_state == DB_CACHED &&
2292 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
2294 * In practice, we will never have a case where we have an
2295 * encrypted arc buffer while additional holds exist on the
2296 * dbuf. We don't handle this here so we simply assert that
2299 ASSERT(!arc_is_encrypted(buf));
2300 mutex_exit(&db->db_mtx);
2301 (void) dbuf_dirty(db, tx);
2302 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
2303 arc_buf_destroy(buf, db);
2304 xuio_stat_wbuf_copied();
2308 xuio_stat_wbuf_nocopy();
2309 if (db->db_state == DB_CACHED) {
2310 dbuf_dirty_record_t *dr = db->db_last_dirty;
2312 ASSERT(db->db_buf != NULL);
2313 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
2314 ASSERT(dr->dt.dl.dr_data == db->db_buf);
2316 if (!arc_released(db->db_buf)) {
2317 ASSERT(dr->dt.dl.dr_override_state ==
2319 arc_release(db->db_buf, db);
2321 dr->dt.dl.dr_data = buf;
2322 arc_buf_destroy(db->db_buf, db);
2323 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
2324 arc_release(db->db_buf, db);
2325 arc_buf_destroy(db->db_buf, db);
2329 ASSERT(db->db_buf == NULL);
2330 dbuf_set_data(db, buf);
2331 db->db_state = DB_FILL;
2332 mutex_exit(&db->db_mtx);
2333 (void) dbuf_dirty(db, tx);
2334 dmu_buf_fill_done(&db->db, tx);
2338 dbuf_destroy(dmu_buf_impl_t *db)
2341 dmu_buf_impl_t *parent = db->db_parent;
2342 dmu_buf_impl_t *dndb;
2344 ASSERT(MUTEX_HELD(&db->db_mtx));
2345 ASSERT(refcount_is_zero(&db->db_holds));
2347 if (db->db_buf != NULL) {
2348 arc_buf_destroy(db->db_buf, db);
2352 if (db->db_blkid == DMU_BONUS_BLKID) {
2353 int slots = DB_DNODE(db)->dn_num_slots;
2354 int bonuslen = DN_SLOTS_TO_BONUSLEN(slots);
2355 if (db->db.db_data != NULL) {
2356 kmem_free(db->db.db_data, bonuslen);
2357 arc_space_return(bonuslen, ARC_SPACE_BONUS);
2358 db->db_state = DB_UNCACHED;
2362 dbuf_clear_data(db);
2364 if (multilist_link_active(&db->db_cache_link)) {
2365 multilist_remove(dbuf_cache, db);
2366 (void) refcount_remove_many(&dbuf_cache_size,
2367 db->db.db_size, db);
2368 DBUF_STAT_BUMPDOWN(cache_levels[db->db_level]);
2369 DBUF_STAT_BUMPDOWN(cache_count);
2370 DBUF_STAT_DECR(cache_levels_bytes[db->db_level],
2374 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
2375 ASSERT(db->db_data_pending == NULL);
2377 db->db_state = DB_EVICTING;
2378 db->db_blkptr = NULL;
2381 * Now that db_state is DB_EVICTING, nobody else can find this via
2382 * the hash table. We can now drop db_mtx, which allows us to
2383 * acquire the dn_dbufs_mtx.
2385 mutex_exit(&db->db_mtx);
2390 if (db->db_blkid != DMU_BONUS_BLKID) {
2391 boolean_t needlock = !MUTEX_HELD(&dn->dn_dbufs_mtx);
2393 mutex_enter(&dn->dn_dbufs_mtx);
2394 avl_remove(&dn->dn_dbufs, db);
2395 atomic_dec_32(&dn->dn_dbufs_count);
2399 mutex_exit(&dn->dn_dbufs_mtx);
2401 * Decrementing the dbuf count means that the hold corresponding
2402 * to the removed dbuf is no longer discounted in dnode_move(),
2403 * so the dnode cannot be moved until after we release the hold.
2404 * The membar_producer() ensures visibility of the decremented
2405 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
2408 mutex_enter(&dn->dn_mtx);
2409 dnode_rele_and_unlock(dn, db, B_TRUE);
2410 db->db_dnode_handle = NULL;
2412 dbuf_hash_remove(db);
2417 ASSERT(refcount_is_zero(&db->db_holds));
2419 db->db_parent = NULL;
2421 ASSERT(db->db_buf == NULL);
2422 ASSERT(db->db.db_data == NULL);
2423 ASSERT(db->db_hash_next == NULL);
2424 ASSERT(db->db_blkptr == NULL);
2425 ASSERT(db->db_data_pending == NULL);
2426 ASSERT(!multilist_link_active(&db->db_cache_link));
2428 kmem_cache_free(dbuf_kmem_cache, db);
2429 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
2432 * If this dbuf is referenced from an indirect dbuf,
2433 * decrement the ref count on the indirect dbuf.
2435 if (parent && parent != dndb) {
2436 mutex_enter(&parent->db_mtx);
2437 dbuf_rele_and_unlock(parent, db, B_TRUE);
2442 * Note: While bpp will always be updated if the function returns success,
2443 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
2444 * this happens when the dnode is the meta-dnode, or {user|group|project}used
2447 __attribute__((always_inline))
2449 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
2450 dmu_buf_impl_t **parentp, blkptr_t **bpp, struct dbuf_hold_impl_data *dh)
2455 ASSERT(blkid != DMU_BONUS_BLKID);
2457 if (blkid == DMU_SPILL_BLKID) {
2458 mutex_enter(&dn->dn_mtx);
2459 if (dn->dn_have_spill &&
2460 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
2461 *bpp = DN_SPILL_BLKPTR(dn->dn_phys);
2464 dbuf_add_ref(dn->dn_dbuf, NULL);
2465 *parentp = dn->dn_dbuf;
2466 mutex_exit(&dn->dn_mtx);
2471 (dn->dn_phys->dn_nlevels == 0) ? 1 : dn->dn_phys->dn_nlevels;
2472 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2474 ASSERT3U(level * epbs, <, 64);
2475 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2477 * This assertion shouldn't trip as long as the max indirect block size
2478 * is less than 1M. The reason for this is that up to that point,
2479 * the number of levels required to address an entire object with blocks
2480 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In
2481 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55
2482 * (i.e. we can address the entire object), objects will all use at most
2483 * N-1 levels and the assertion won't overflow. However, once epbs is
2484 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be
2485 * enough to address an entire object, so objects will have 5 levels,
2486 * but then this assertion will overflow.
2488 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we
2489 * need to redo this logic to handle overflows.
2491 ASSERT(level >= nlevels ||
2492 ((nlevels - level - 1) * epbs) +
2493 highbit64(dn->dn_phys->dn_nblkptr) <= 64);
2494 if (level >= nlevels ||
2495 blkid >= ((uint64_t)dn->dn_phys->dn_nblkptr <<
2496 ((nlevels - level - 1) * epbs)) ||
2498 blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
2499 /* the buffer has no parent yet */
2500 return (SET_ERROR(ENOENT));
2501 } else if (level < nlevels-1) {
2502 /* this block is referenced from an indirect block */
2505 err = dbuf_hold_impl(dn, level+1,
2506 blkid >> epbs, fail_sparse, FALSE, NULL, parentp);
2508 __dbuf_hold_impl_init(dh + 1, dn, dh->dh_level + 1,
2509 blkid >> epbs, fail_sparse, FALSE, NULL,
2510 parentp, dh->dh_depth + 1);
2511 err = __dbuf_hold_impl(dh + 1);
2515 err = dbuf_read(*parentp, NULL,
2516 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
2518 dbuf_rele(*parentp, NULL);
2522 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
2523 (blkid & ((1ULL << epbs) - 1));
2524 if (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))
2525 ASSERT(BP_IS_HOLE(*bpp));
2528 /* the block is referenced from the dnode */
2529 ASSERT3U(level, ==, nlevels-1);
2530 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
2531 blkid < dn->dn_phys->dn_nblkptr);
2533 dbuf_add_ref(dn->dn_dbuf, NULL);
2534 *parentp = dn->dn_dbuf;
2536 *bpp = &dn->dn_phys->dn_blkptr[blkid];
2541 static dmu_buf_impl_t *
2542 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
2543 dmu_buf_impl_t *parent, blkptr_t *blkptr)
2545 objset_t *os = dn->dn_objset;
2546 dmu_buf_impl_t *db, *odb;
2548 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2549 ASSERT(dn->dn_type != DMU_OT_NONE);
2551 db = kmem_cache_alloc(dbuf_kmem_cache, KM_SLEEP);
2554 db->db.db_object = dn->dn_object;
2555 db->db_level = level;
2556 db->db_blkid = blkid;
2557 db->db_last_dirty = NULL;
2558 db->db_dirtycnt = 0;
2559 db->db_dnode_handle = dn->dn_handle;
2560 db->db_parent = parent;
2561 db->db_blkptr = blkptr;
2564 db->db_user_immediate_evict = FALSE;
2565 db->db_freed_in_flight = FALSE;
2566 db->db_pending_evict = FALSE;
2568 if (blkid == DMU_BONUS_BLKID) {
2569 ASSERT3P(parent, ==, dn->dn_dbuf);
2570 db->db.db_size = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
2571 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
2572 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
2573 db->db.db_offset = DMU_BONUS_BLKID;
2574 db->db_state = DB_UNCACHED;
2575 /* the bonus dbuf is not placed in the hash table */
2576 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
2578 } else if (blkid == DMU_SPILL_BLKID) {
2579 db->db.db_size = (blkptr != NULL) ?
2580 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
2581 db->db.db_offset = 0;
2584 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
2585 db->db.db_size = blocksize;
2586 db->db.db_offset = db->db_blkid * blocksize;
2590 * Hold the dn_dbufs_mtx while we get the new dbuf
2591 * in the hash table *and* added to the dbufs list.
2592 * This prevents a possible deadlock with someone
2593 * trying to look up this dbuf before its added to the
2596 mutex_enter(&dn->dn_dbufs_mtx);
2597 db->db_state = DB_EVICTING;
2598 if ((odb = dbuf_hash_insert(db)) != NULL) {
2599 /* someone else inserted it first */
2600 kmem_cache_free(dbuf_kmem_cache, db);
2601 mutex_exit(&dn->dn_dbufs_mtx);
2602 DBUF_STAT_BUMP(hash_insert_race);
2605 avl_add(&dn->dn_dbufs, db);
2607 db->db_state = DB_UNCACHED;
2608 mutex_exit(&dn->dn_dbufs_mtx);
2609 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
2611 if (parent && parent != dn->dn_dbuf)
2612 dbuf_add_ref(parent, db);
2614 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
2615 refcount_count(&dn->dn_holds) > 0);
2616 (void) refcount_add(&dn->dn_holds, db);
2617 atomic_inc_32(&dn->dn_dbufs_count);
2619 dprintf_dbuf(db, "db=%p\n", db);
2624 typedef struct dbuf_prefetch_arg {
2625 spa_t *dpa_spa; /* The spa to issue the prefetch in. */
2626 zbookmark_phys_t dpa_zb; /* The target block to prefetch. */
2627 int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */
2628 int dpa_curlevel; /* The current level that we're reading */
2629 dnode_t *dpa_dnode; /* The dnode associated with the prefetch */
2630 zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */
2631 zio_t *dpa_zio; /* The parent zio_t for all prefetches. */
2632 arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */
2633 } dbuf_prefetch_arg_t;
2636 * Actually issue the prefetch read for the block given.
2639 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp)
2641 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
2644 int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE;
2645 arc_flags_t aflags =
2646 dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
2648 /* dnodes are always read as raw and then converted later */
2649 if (BP_GET_TYPE(bp) == DMU_OT_DNODE && BP_IS_PROTECTED(bp) &&
2650 dpa->dpa_curlevel == 0)
2651 zio_flags |= ZIO_FLAG_RAW;
2653 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2654 ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level);
2655 ASSERT(dpa->dpa_zio != NULL);
2656 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL,
2657 dpa->dpa_prio, zio_flags, &aflags, &dpa->dpa_zb);
2661 * Called when an indirect block above our prefetch target is read in. This
2662 * will either read in the next indirect block down the tree or issue the actual
2663 * prefetch if the next block down is our target.
2666 dbuf_prefetch_indirect_done(zio_t *zio, const zbookmark_phys_t *zb,
2667 const blkptr_t *iobp, arc_buf_t *abuf, void *private)
2669 dbuf_prefetch_arg_t *dpa = private;
2671 ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel);
2672 ASSERT3S(dpa->dpa_curlevel, >, 0);
2675 * The dpa_dnode is only valid if we are called with a NULL
2676 * zio. This indicates that the arc_read() returned without
2677 * first calling zio_read() to issue a physical read. Once
2678 * a physical read is made the dpa_dnode must be invalidated
2679 * as the locks guarding it may have been dropped. If the
2680 * dpa_dnode is still valid, then we want to add it to the dbuf
2681 * cache. To do so, we must hold the dbuf associated with the block
2682 * we just prefetched, read its contents so that we associate it
2683 * with an arc_buf_t, and then release it.
2686 ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel);
2687 if (zio->io_flags & ZIO_FLAG_RAW_COMPRESS) {
2688 ASSERT3U(BP_GET_PSIZE(zio->io_bp), ==, zio->io_size);
2690 ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size);
2692 ASSERT3P(zio->io_spa, ==, dpa->dpa_spa);
2694 dpa->dpa_dnode = NULL;
2695 } else if (dpa->dpa_dnode != NULL) {
2696 uint64_t curblkid = dpa->dpa_zb.zb_blkid >>
2697 (dpa->dpa_epbs * (dpa->dpa_curlevel -
2698 dpa->dpa_zb.zb_level));
2699 dmu_buf_impl_t *db = dbuf_hold_level(dpa->dpa_dnode,
2700 dpa->dpa_curlevel, curblkid, FTAG);
2701 (void) dbuf_read(db, NULL,
2702 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_HAVESTRUCT);
2703 dbuf_rele(db, FTAG);
2707 kmem_free(dpa, sizeof (*dpa));
2711 dpa->dpa_curlevel--;
2712 uint64_t nextblkid = dpa->dpa_zb.zb_blkid >>
2713 (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level));
2714 blkptr_t *bp = ((blkptr_t *)abuf->b_data) +
2715 P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs);
2717 if (BP_IS_HOLE(bp)) {
2718 kmem_free(dpa, sizeof (*dpa));
2719 } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) {
2720 ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid);
2721 dbuf_issue_final_prefetch(dpa, bp);
2722 kmem_free(dpa, sizeof (*dpa));
2724 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2725 zbookmark_phys_t zb;
2727 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2728 if (dpa->dpa_aflags & ARC_FLAG_L2CACHE)
2729 iter_aflags |= ARC_FLAG_L2CACHE;
2731 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2733 SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset,
2734 dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid);
2736 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2737 bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio,
2738 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2742 arc_buf_destroy(abuf, private);
2746 * Issue prefetch reads for the given block on the given level. If the indirect
2747 * blocks above that block are not in memory, we will read them in
2748 * asynchronously. As a result, this call never blocks waiting for a read to
2749 * complete. Note that the prefetch might fail if the dataset is encrypted and
2750 * the encryption key is unmapped before the IO completes.
2753 dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio,
2757 int epbs, nlevels, curlevel;
2760 ASSERT(blkid != DMU_BONUS_BLKID);
2761 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2763 if (blkid > dn->dn_maxblkid)
2766 if (dnode_block_freed(dn, blkid))
2770 * This dnode hasn't been written to disk yet, so there's nothing to
2773 nlevels = dn->dn_phys->dn_nlevels;
2774 if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0)
2777 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2778 if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level))
2781 dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object,
2784 mutex_exit(&db->db_mtx);
2786 * This dbuf already exists. It is either CACHED, or
2787 * (we assume) about to be read or filled.
2793 * Find the closest ancestor (indirect block) of the target block
2794 * that is present in the cache. In this indirect block, we will
2795 * find the bp that is at curlevel, curblkid.
2799 while (curlevel < nlevels - 1) {
2800 int parent_level = curlevel + 1;
2801 uint64_t parent_blkid = curblkid >> epbs;
2804 if (dbuf_hold_impl(dn, parent_level, parent_blkid,
2805 FALSE, TRUE, FTAG, &db) == 0) {
2806 blkptr_t *bpp = db->db_buf->b_data;
2807 bp = bpp[P2PHASE(curblkid, 1 << epbs)];
2808 dbuf_rele(db, FTAG);
2812 curlevel = parent_level;
2813 curblkid = parent_blkid;
2816 if (curlevel == nlevels - 1) {
2817 /* No cached indirect blocks found. */
2818 ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr);
2819 bp = dn->dn_phys->dn_blkptr[curblkid];
2821 if (BP_IS_HOLE(&bp))
2824 ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp));
2826 zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL,
2829 dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP);
2830 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
2831 SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2832 dn->dn_object, level, blkid);
2833 dpa->dpa_curlevel = curlevel;
2834 dpa->dpa_prio = prio;
2835 dpa->dpa_aflags = aflags;
2836 dpa->dpa_spa = dn->dn_objset->os_spa;
2837 dpa->dpa_dnode = dn;
2838 dpa->dpa_epbs = epbs;
2841 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2842 if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level))
2843 dpa->dpa_aflags |= ARC_FLAG_L2CACHE;
2846 * If we have the indirect just above us, no need to do the asynchronous
2847 * prefetch chain; we'll just run the last step ourselves. If we're at
2848 * a higher level, though, we want to issue the prefetches for all the
2849 * indirect blocks asynchronously, so we can go on with whatever we were
2852 if (curlevel == level) {
2853 ASSERT3U(curblkid, ==, blkid);
2854 dbuf_issue_final_prefetch(dpa, &bp);
2855 kmem_free(dpa, sizeof (*dpa));
2857 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2858 zbookmark_phys_t zb;
2860 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2861 if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level))
2862 iter_aflags |= ARC_FLAG_L2CACHE;
2864 SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2865 dn->dn_object, curlevel, curblkid);
2866 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2867 &bp, dbuf_prefetch_indirect_done, dpa, prio,
2868 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2872 * We use pio here instead of dpa_zio since it's possible that
2873 * dpa may have already been freed.
2878 #define DBUF_HOLD_IMPL_MAX_DEPTH 20
2881 * Helper function for __dbuf_hold_impl() to copy a buffer. Handles
2882 * the case of encrypted, compressed and uncompressed buffers by
2883 * allocating the new buffer, respectively, with arc_alloc_raw_buf(),
2884 * arc_alloc_compressed_buf() or arc_alloc_buf().*
2886 * NOTE: Declared noinline to avoid stack bloat in __dbuf_hold_impl().
2888 noinline static void
2889 dbuf_hold_copy(struct dbuf_hold_impl_data *dh)
2891 dnode_t *dn = dh->dh_dn;
2892 dmu_buf_impl_t *db = dh->dh_db;
2893 dbuf_dirty_record_t *dr = dh->dh_dr;
2894 arc_buf_t *data = dr->dt.dl.dr_data;
2896 enum zio_compress compress_type = arc_get_compression(data);
2898 if (arc_is_encrypted(data)) {
2899 boolean_t byteorder;
2900 uint8_t salt[ZIO_DATA_SALT_LEN];
2901 uint8_t iv[ZIO_DATA_IV_LEN];
2902 uint8_t mac[ZIO_DATA_MAC_LEN];
2904 arc_get_raw_params(data, &byteorder, salt, iv, mac);
2905 dbuf_set_data(db, arc_alloc_raw_buf(dn->dn_objset->os_spa, db,
2906 dmu_objset_id(dn->dn_objset), byteorder, salt, iv, mac,
2907 dn->dn_type, arc_buf_size(data), arc_buf_lsize(data),
2909 } else if (compress_type != ZIO_COMPRESS_OFF) {
2910 dbuf_set_data(db, arc_alloc_compressed_buf(
2911 dn->dn_objset->os_spa, db, arc_buf_size(data),
2912 arc_buf_lsize(data), compress_type));
2914 dbuf_set_data(db, arc_alloc_buf(dn->dn_objset->os_spa, db,
2915 DBUF_GET_BUFC_TYPE(db), db->db.db_size));
2918 bcopy(data->b_data, db->db.db_data, arc_buf_size(data));
2922 * Returns with db_holds incremented, and db_mtx not held.
2923 * Note: dn_struct_rwlock must be held.
2926 __dbuf_hold_impl(struct dbuf_hold_impl_data *dh)
2928 ASSERT3S(dh->dh_depth, <, DBUF_HOLD_IMPL_MAX_DEPTH);
2929 dh->dh_parent = NULL;
2931 ASSERT(dh->dh_blkid != DMU_BONUS_BLKID);
2932 ASSERT(RW_LOCK_HELD(&dh->dh_dn->dn_struct_rwlock));
2933 ASSERT3U(dh->dh_dn->dn_nlevels, >, dh->dh_level);
2935 *(dh->dh_dbp) = NULL;
2937 /* dbuf_find() returns with db_mtx held */
2938 dh->dh_db = dbuf_find(dh->dh_dn->dn_objset, dh->dh_dn->dn_object,
2939 dh->dh_level, dh->dh_blkid);
2941 if (dh->dh_db == NULL) {
2944 if (dh->dh_fail_uncached)
2945 return (SET_ERROR(ENOENT));
2947 ASSERT3P(dh->dh_parent, ==, NULL);
2948 dh->dh_err = dbuf_findbp(dh->dh_dn, dh->dh_level, dh->dh_blkid,
2949 dh->dh_fail_sparse, &dh->dh_parent, &dh->dh_bp, dh);
2950 if (dh->dh_fail_sparse) {
2951 if (dh->dh_err == 0 &&
2952 dh->dh_bp && BP_IS_HOLE(dh->dh_bp))
2953 dh->dh_err = SET_ERROR(ENOENT);
2956 dbuf_rele(dh->dh_parent, NULL);
2957 return (dh->dh_err);
2960 if (dh->dh_err && dh->dh_err != ENOENT)
2961 return (dh->dh_err);
2962 dh->dh_db = dbuf_create(dh->dh_dn, dh->dh_level, dh->dh_blkid,
2963 dh->dh_parent, dh->dh_bp);
2966 if (dh->dh_fail_uncached && dh->dh_db->db_state != DB_CACHED) {
2967 mutex_exit(&dh->dh_db->db_mtx);
2968 return (SET_ERROR(ENOENT));
2971 if (dh->dh_db->db_buf != NULL) {
2972 arc_buf_access(dh->dh_db->db_buf);
2973 ASSERT3P(dh->dh_db->db.db_data, ==, dh->dh_db->db_buf->b_data);
2976 ASSERT(dh->dh_db->db_buf == NULL || arc_referenced(dh->dh_db->db_buf));
2979 * If this buffer is currently syncing out, and we are are
2980 * still referencing it from db_data, we need to make a copy
2981 * of it in case we decide we want to dirty it again in this txg.
2983 if (dh->dh_db->db_level == 0 &&
2984 dh->dh_db->db_blkid != DMU_BONUS_BLKID &&
2985 dh->dh_dn->dn_object != DMU_META_DNODE_OBJECT &&
2986 dh->dh_db->db_state == DB_CACHED && dh->dh_db->db_data_pending) {
2987 dh->dh_dr = dh->dh_db->db_data_pending;
2988 if (dh->dh_dr->dt.dl.dr_data == dh->dh_db->db_buf)
2992 if (multilist_link_active(&dh->dh_db->db_cache_link)) {
2993 ASSERT(refcount_is_zero(&dh->dh_db->db_holds));
2994 multilist_remove(dbuf_cache, dh->dh_db);
2995 (void) refcount_remove_many(&dbuf_cache_size,
2996 dh->dh_db->db.db_size, dh->dh_db);
2997 DBUF_STAT_BUMPDOWN(cache_levels[dh->dh_db->db_level]);
2998 DBUF_STAT_BUMPDOWN(cache_count);
2999 DBUF_STAT_DECR(cache_levels_bytes[dh->dh_db->db_level],
3000 dh->dh_db->db.db_size);
3002 (void) refcount_add(&dh->dh_db->db_holds, dh->dh_tag);
3003 DBUF_VERIFY(dh->dh_db);
3004 mutex_exit(&dh->dh_db->db_mtx);
3006 /* NOTE: we can't rele the parent until after we drop the db_mtx */
3008 dbuf_rele(dh->dh_parent, NULL);
3010 ASSERT3P(DB_DNODE(dh->dh_db), ==, dh->dh_dn);
3011 ASSERT3U(dh->dh_db->db_blkid, ==, dh->dh_blkid);
3012 ASSERT3U(dh->dh_db->db_level, ==, dh->dh_level);
3013 *(dh->dh_dbp) = dh->dh_db;
3019 * The following code preserves the recursive function dbuf_hold_impl()
3020 * but moves the local variables AND function arguments to the heap to
3021 * minimize the stack frame size. Enough space is initially allocated
3022 * on the stack for 20 levels of recursion.
3025 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid,
3026 boolean_t fail_sparse, boolean_t fail_uncached,
3027 void *tag, dmu_buf_impl_t **dbp)
3029 struct dbuf_hold_impl_data *dh;
3032 dh = kmem_alloc(sizeof (struct dbuf_hold_impl_data) *
3033 DBUF_HOLD_IMPL_MAX_DEPTH, KM_SLEEP);
3034 __dbuf_hold_impl_init(dh, dn, level, blkid, fail_sparse,
3035 fail_uncached, tag, dbp, 0);
3037 error = __dbuf_hold_impl(dh);
3039 kmem_free(dh, sizeof (struct dbuf_hold_impl_data) *
3040 DBUF_HOLD_IMPL_MAX_DEPTH);
3046 __dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh,
3047 dnode_t *dn, uint8_t level, uint64_t blkid,
3048 boolean_t fail_sparse, boolean_t fail_uncached,
3049 void *tag, dmu_buf_impl_t **dbp, int depth)
3052 dh->dh_level = level;
3053 dh->dh_blkid = blkid;
3055 dh->dh_fail_sparse = fail_sparse;
3056 dh->dh_fail_uncached = fail_uncached;
3062 dh->dh_parent = NULL;
3067 dh->dh_depth = depth;
3071 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
3073 return (dbuf_hold_level(dn, 0, blkid, tag));
3077 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
3080 int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db);
3081 return (err ? NULL : db);
3085 dbuf_create_bonus(dnode_t *dn)
3087 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
3089 ASSERT(dn->dn_bonus == NULL);
3090 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
3094 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
3096 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3099 if (db->db_blkid != DMU_SPILL_BLKID)
3100 return (SET_ERROR(ENOTSUP));
3102 blksz = SPA_MINBLOCKSIZE;
3103 ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset)));
3104 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
3108 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3109 dbuf_new_size(db, blksz, tx);
3110 rw_exit(&dn->dn_struct_rwlock);
3117 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
3119 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
3122 #pragma weak dmu_buf_add_ref = dbuf_add_ref
3124 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
3126 int64_t holds = refcount_add(&db->db_holds, tag);
3127 VERIFY3S(holds, >, 1);
3130 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
3132 dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid,
3135 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3136 dmu_buf_impl_t *found_db;
3137 boolean_t result = B_FALSE;
3139 if (blkid == DMU_BONUS_BLKID)
3140 found_db = dbuf_find_bonus(os, obj);
3142 found_db = dbuf_find(os, obj, 0, blkid);
3144 if (found_db != NULL) {
3145 if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) {
3146 (void) refcount_add(&db->db_holds, tag);
3149 mutex_exit(&found_db->db_mtx);
3155 * If you call dbuf_rele() you had better not be referencing the dnode handle
3156 * unless you have some other direct or indirect hold on the dnode. (An indirect
3157 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
3158 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
3159 * dnode's parent dbuf evicting its dnode handles.
3162 dbuf_rele(dmu_buf_impl_t *db, void *tag)
3164 mutex_enter(&db->db_mtx);
3165 dbuf_rele_and_unlock(db, tag, B_FALSE);
3169 dmu_buf_rele(dmu_buf_t *db, void *tag)
3171 dbuf_rele((dmu_buf_impl_t *)db, tag);
3175 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
3176 * db_dirtycnt and db_holds to be updated atomically. The 'evicting'
3177 * argument should be set if we are already in the dbuf-evicting code
3178 * path, in which case we don't want to recursively evict. This allows us to
3179 * avoid deeply nested stacks that would have a call flow similar to this:
3181 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify()
3184 * +-----dbuf_destroy()<--dbuf_evict_one()<--------+
3188 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag, boolean_t evicting)
3192 ASSERT(MUTEX_HELD(&db->db_mtx));
3196 * Remove the reference to the dbuf before removing its hold on the
3197 * dnode so we can guarantee in dnode_move() that a referenced bonus
3198 * buffer has a corresponding dnode hold.
3200 holds = refcount_remove(&db->db_holds, tag);
3204 * We can't freeze indirects if there is a possibility that they
3205 * may be modified in the current syncing context.
3207 if (db->db_buf != NULL &&
3208 holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) {
3209 arc_buf_freeze(db->db_buf);
3212 if (holds == db->db_dirtycnt &&
3213 db->db_level == 0 && db->db_user_immediate_evict)
3214 dbuf_evict_user(db);
3217 if (db->db_blkid == DMU_BONUS_BLKID) {
3219 boolean_t evict_dbuf = db->db_pending_evict;
3222 * If the dnode moves here, we cannot cross this
3223 * barrier until the move completes.
3228 atomic_dec_32(&dn->dn_dbufs_count);
3231 * Decrementing the dbuf count means that the bonus
3232 * buffer's dnode hold is no longer discounted in
3233 * dnode_move(). The dnode cannot move until after
3234 * the dnode_rele() below.
3239 * Do not reference db after its lock is dropped.
3240 * Another thread may evict it.
3242 mutex_exit(&db->db_mtx);
3245 dnode_evict_bonus(dn);
3248 } else if (db->db_buf == NULL) {
3250 * This is a special case: we never associated this
3251 * dbuf with any data allocated from the ARC.
3253 ASSERT(db->db_state == DB_UNCACHED ||
3254 db->db_state == DB_NOFILL);
3256 } else if (arc_released(db->db_buf)) {
3258 * This dbuf has anonymous data associated with it.
3262 boolean_t do_arc_evict = B_FALSE;
3264 spa_t *spa = dmu_objset_spa(db->db_objset);
3266 if (!DBUF_IS_CACHEABLE(db) &&
3267 db->db_blkptr != NULL &&
3268 !BP_IS_HOLE(db->db_blkptr) &&
3269 !BP_IS_EMBEDDED(db->db_blkptr)) {
3270 do_arc_evict = B_TRUE;
3271 bp = *db->db_blkptr;
3274 if (!DBUF_IS_CACHEABLE(db) ||
3275 db->db_pending_evict) {
3277 } else if (!multilist_link_active(&db->db_cache_link)) {
3278 multilist_insert(dbuf_cache, db);
3279 (void) refcount_add_many(&dbuf_cache_size,
3280 db->db.db_size, db);
3281 DBUF_STAT_BUMP(cache_levels[db->db_level]);
3282 DBUF_STAT_BUMP(cache_count);
3283 DBUF_STAT_INCR(cache_levels_bytes[db->db_level],
3285 DBUF_STAT_MAX(cache_size_bytes_max,
3286 refcount_count(&dbuf_cache_size));
3287 mutex_exit(&db->db_mtx);
3290 dbuf_evict_notify();
3294 arc_freed(spa, &bp);
3297 mutex_exit(&db->db_mtx);
3302 #pragma weak dmu_buf_refcount = dbuf_refcount
3304 dbuf_refcount(dmu_buf_impl_t *db)
3306 return (refcount_count(&db->db_holds));
3310 dmu_buf_user_refcount(dmu_buf_t *db_fake)
3313 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3315 mutex_enter(&db->db_mtx);
3316 ASSERT3U(refcount_count(&db->db_holds), >=, db->db_dirtycnt);
3317 holds = refcount_count(&db->db_holds) - db->db_dirtycnt;
3318 mutex_exit(&db->db_mtx);
3324 dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user,
3325 dmu_buf_user_t *new_user)
3327 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3329 mutex_enter(&db->db_mtx);
3330 dbuf_verify_user(db, DBVU_NOT_EVICTING);
3331 if (db->db_user == old_user)
3332 db->db_user = new_user;
3334 old_user = db->db_user;
3335 dbuf_verify_user(db, DBVU_NOT_EVICTING);
3336 mutex_exit(&db->db_mtx);
3342 dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
3344 return (dmu_buf_replace_user(db_fake, NULL, user));
3348 dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user)
3350 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3352 db->db_user_immediate_evict = TRUE;
3353 return (dmu_buf_set_user(db_fake, user));
3357 dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
3359 return (dmu_buf_replace_user(db_fake, user, NULL));
3363 dmu_buf_get_user(dmu_buf_t *db_fake)
3365 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3367 dbuf_verify_user(db, DBVU_NOT_EVICTING);
3368 return (db->db_user);
3372 dmu_buf_user_evict_wait()
3374 taskq_wait(dbu_evict_taskq);
3378 dmu_buf_get_blkptr(dmu_buf_t *db)
3380 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3381 return (dbi->db_blkptr);
3385 dmu_buf_get_objset(dmu_buf_t *db)
3387 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3388 return (dbi->db_objset);
3392 dmu_buf_dnode_enter(dmu_buf_t *db)
3394 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3395 DB_DNODE_ENTER(dbi);
3396 return (DB_DNODE(dbi));
3400 dmu_buf_dnode_exit(dmu_buf_t *db)
3402 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3407 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
3409 /* ASSERT(dmu_tx_is_syncing(tx) */
3410 ASSERT(MUTEX_HELD(&db->db_mtx));
3412 if (db->db_blkptr != NULL)
3415 if (db->db_blkid == DMU_SPILL_BLKID) {
3416 db->db_blkptr = DN_SPILL_BLKPTR(dn->dn_phys);
3417 BP_ZERO(db->db_blkptr);
3420 if (db->db_level == dn->dn_phys->dn_nlevels-1) {
3422 * This buffer was allocated at a time when there was
3423 * no available blkptrs from the dnode, or it was
3424 * inappropriate to hook it in (i.e., nlevels mis-match).
3426 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
3427 ASSERT(db->db_parent == NULL);
3428 db->db_parent = dn->dn_dbuf;
3429 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
3432 dmu_buf_impl_t *parent = db->db_parent;
3433 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3435 ASSERT(dn->dn_phys->dn_nlevels > 1);
3436 if (parent == NULL) {
3437 mutex_exit(&db->db_mtx);
3438 rw_enter(&dn->dn_struct_rwlock, RW_READER);
3439 parent = dbuf_hold_level(dn, db->db_level + 1,
3440 db->db_blkid >> epbs, db);
3441 rw_exit(&dn->dn_struct_rwlock);
3442 mutex_enter(&db->db_mtx);
3443 db->db_parent = parent;
3445 db->db_blkptr = (blkptr_t *)parent->db.db_data +
3446 (db->db_blkid & ((1ULL << epbs) - 1));
3452 * When syncing out a blocks of dnodes, adjust the block to deal with
3453 * encryption. Normally, we make sure the block is decrypted before writing
3454 * it. If we have crypt params, then we are writing a raw (encrypted) block,
3455 * from a raw receive. In this case, set the ARC buf's crypt params so
3456 * that the BP will be filled with the correct byteorder, salt, iv, and mac.
3459 dbuf_prepare_encrypted_dnode_leaf(dbuf_dirty_record_t *dr)
3462 dmu_buf_impl_t *db = dr->dr_dbuf;
3464 ASSERT(MUTEX_HELD(&db->db_mtx));
3465 ASSERT3U(db->db.db_object, ==, DMU_META_DNODE_OBJECT);
3466 ASSERT3U(db->db_level, ==, 0);
3468 if (!db->db_objset->os_raw_receive && arc_is_encrypted(db->db_buf)) {
3469 zbookmark_phys_t zb;
3472 * Unfortunately, there is currently no mechanism for
3473 * syncing context to handle decryption errors. An error
3474 * here is only possible if an attacker maliciously
3475 * changed a dnode block and updated the associated
3476 * checksums going up the block tree.
3478 SET_BOOKMARK(&zb, dmu_objset_id(db->db_objset),
3479 db->db.db_object, db->db_level, db->db_blkid);
3480 err = arc_untransform(db->db_buf, db->db_objset->os_spa,
3483 panic("Invalid dnode block MAC");
3484 } else if (dr->dt.dl.dr_has_raw_params) {
3485 (void) arc_release(dr->dt.dl.dr_data, db);
3486 arc_convert_to_raw(dr->dt.dl.dr_data,
3487 dmu_objset_id(db->db_objset),
3488 dr->dt.dl.dr_byteorder, DMU_OT_DNODE,
3489 dr->dt.dl.dr_salt, dr->dt.dl.dr_iv, dr->dt.dl.dr_mac);
3494 * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it
3495 * is critical the we not allow the compiler to inline this function in to
3496 * dbuf_sync_list() thereby drastically bloating the stack usage.
3498 noinline static void
3499 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
3501 dmu_buf_impl_t *db = dr->dr_dbuf;
3505 ASSERT(dmu_tx_is_syncing(tx));
3507 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
3509 mutex_enter(&db->db_mtx);
3511 ASSERT(db->db_level > 0);
3514 /* Read the block if it hasn't been read yet. */
3515 if (db->db_buf == NULL) {
3516 mutex_exit(&db->db_mtx);
3517 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
3518 mutex_enter(&db->db_mtx);
3520 ASSERT3U(db->db_state, ==, DB_CACHED);
3521 ASSERT(db->db_buf != NULL);
3525 /* Indirect block size must match what the dnode thinks it is. */
3526 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3527 dbuf_check_blkptr(dn, db);
3530 /* Provide the pending dirty record to child dbufs */
3531 db->db_data_pending = dr;
3533 mutex_exit(&db->db_mtx);
3535 dbuf_write(dr, db->db_buf, tx);
3538 mutex_enter(&dr->dt.di.dr_mtx);
3539 dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx);
3540 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3541 mutex_exit(&dr->dt.di.dr_mtx);
3546 * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is
3547 * critical the we not allow the compiler to inline this function in to
3548 * dbuf_sync_list() thereby drastically bloating the stack usage.
3550 noinline static void
3551 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
3553 arc_buf_t **datap = &dr->dt.dl.dr_data;
3554 dmu_buf_impl_t *db = dr->dr_dbuf;
3557 uint64_t txg = tx->tx_txg;
3559 ASSERT(dmu_tx_is_syncing(tx));
3561 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
3563 mutex_enter(&db->db_mtx);
3565 * To be synced, we must be dirtied. But we
3566 * might have been freed after the dirty.
3568 if (db->db_state == DB_UNCACHED) {
3569 /* This buffer has been freed since it was dirtied */
3570 ASSERT(db->db.db_data == NULL);
3571 } else if (db->db_state == DB_FILL) {
3572 /* This buffer was freed and is now being re-filled */
3573 ASSERT(db->db.db_data != dr->dt.dl.dr_data);
3575 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
3582 if (db->db_blkid == DMU_SPILL_BLKID) {
3583 mutex_enter(&dn->dn_mtx);
3584 if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
3586 * In the previous transaction group, the bonus buffer
3587 * was entirely used to store the attributes for the
3588 * dnode which overrode the dn_spill field. However,
3589 * when adding more attributes to the file a spill
3590 * block was required to hold the extra attributes.
3592 * Make sure to clear the garbage left in the dn_spill
3593 * field from the previous attributes in the bonus
3594 * buffer. Otherwise, after writing out the spill
3595 * block to the new allocated dva, it will free
3596 * the old block pointed to by the invalid dn_spill.
3598 db->db_blkptr = NULL;
3600 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
3601 mutex_exit(&dn->dn_mtx);
3605 * If this is a bonus buffer, simply copy the bonus data into the
3606 * dnode. It will be written out when the dnode is synced (and it
3607 * will be synced, since it must have been dirty for dbuf_sync to
3610 if (db->db_blkid == DMU_BONUS_BLKID) {
3611 dbuf_dirty_record_t **drp;
3613 ASSERT(*datap != NULL);
3614 ASSERT0(db->db_level);
3615 ASSERT3U(DN_MAX_BONUS_LEN(dn->dn_phys), <=,
3616 DN_SLOTS_TO_BONUSLEN(dn->dn_phys->dn_extra_slots + 1));
3617 bcopy(*datap, DN_BONUS(dn->dn_phys),
3618 DN_MAX_BONUS_LEN(dn->dn_phys));
3621 if (*datap != db->db.db_data) {
3622 int slots = DB_DNODE(db)->dn_num_slots;
3623 int bonuslen = DN_SLOTS_TO_BONUSLEN(slots);
3624 kmem_free(*datap, bonuslen);
3625 arc_space_return(bonuslen, ARC_SPACE_BONUS);
3627 db->db_data_pending = NULL;
3628 drp = &db->db_last_dirty;
3630 drp = &(*drp)->dr_next;
3631 ASSERT(dr->dr_next == NULL);
3632 ASSERT(dr->dr_dbuf == db);
3634 if (dr->dr_dbuf->db_level != 0) {
3635 mutex_destroy(&dr->dt.di.dr_mtx);
3636 list_destroy(&dr->dt.di.dr_children);
3638 kmem_free(dr, sizeof (dbuf_dirty_record_t));
3639 ASSERT(db->db_dirtycnt > 0);
3640 db->db_dirtycnt -= 1;
3641 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg, B_FALSE);
3648 * This function may have dropped the db_mtx lock allowing a dmu_sync
3649 * operation to sneak in. As a result, we need to ensure that we
3650 * don't check the dr_override_state until we have returned from
3651 * dbuf_check_blkptr.
3653 dbuf_check_blkptr(dn, db);
3656 * If this buffer is in the middle of an immediate write,
3657 * wait for the synchronous IO to complete.
3659 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
3660 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
3661 cv_wait(&db->db_changed, &db->db_mtx);
3662 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
3666 * If this is a dnode block, ensure it is appropriately encrypted
3667 * or decrypted, depending on what we are writing to it this txg.
3669 if (os->os_encrypted && dn->dn_object == DMU_META_DNODE_OBJECT)
3670 dbuf_prepare_encrypted_dnode_leaf(dr);
3672 if (db->db_state != DB_NOFILL &&
3673 dn->dn_object != DMU_META_DNODE_OBJECT &&
3674 refcount_count(&db->db_holds) > 1 &&
3675 dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
3676 *datap == db->db_buf) {
3678 * If this buffer is currently "in use" (i.e., there
3679 * are active holds and db_data still references it),
3680 * then make a copy before we start the write so that
3681 * any modifications from the open txg will not leak
3684 * NOTE: this copy does not need to be made for
3685 * objects only modified in the syncing context (e.g.
3686 * DNONE_DNODE blocks).
3688 int psize = arc_buf_size(*datap);
3689 int lsize = arc_buf_lsize(*datap);
3690 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
3691 enum zio_compress compress_type = arc_get_compression(*datap);
3693 if (arc_is_encrypted(*datap)) {
3694 boolean_t byteorder;
3695 uint8_t salt[ZIO_DATA_SALT_LEN];
3696 uint8_t iv[ZIO_DATA_IV_LEN];
3697 uint8_t mac[ZIO_DATA_MAC_LEN];
3699 arc_get_raw_params(*datap, &byteorder, salt, iv, mac);
3700 *datap = arc_alloc_raw_buf(os->os_spa, db,
3701 dmu_objset_id(os), byteorder, salt, iv, mac,
3702 dn->dn_type, psize, lsize, compress_type);
3703 } else if (compress_type != ZIO_COMPRESS_OFF) {
3704 ASSERT3U(type, ==, ARC_BUFC_DATA);
3705 *datap = arc_alloc_compressed_buf(os->os_spa, db,
3706 psize, lsize, compress_type);
3708 *datap = arc_alloc_buf(os->os_spa, db, type, psize);
3710 bcopy(db->db.db_data, (*datap)->b_data, psize);
3712 db->db_data_pending = dr;
3714 mutex_exit(&db->db_mtx);
3716 dbuf_write(dr, *datap, tx);
3718 ASSERT(!list_link_active(&dr->dr_dirty_node));
3719 if (dn->dn_object == DMU_META_DNODE_OBJECT) {
3720 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
3724 * Although zio_nowait() does not "wait for an IO", it does
3725 * initiate the IO. If this is an empty write it seems plausible
3726 * that the IO could actually be completed before the nowait
3727 * returns. We need to DB_DNODE_EXIT() first in case
3728 * zio_nowait() invalidates the dbuf.
3731 zio_nowait(dr->dr_zio);
3736 dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx)
3738 dbuf_dirty_record_t *dr;
3740 while ((dr = list_head(list))) {
3741 if (dr->dr_zio != NULL) {
3743 * If we find an already initialized zio then we
3744 * are processing the meta-dnode, and we have finished.
3745 * The dbufs for all dnodes are put back on the list
3746 * during processing, so that we can zio_wait()
3747 * these IOs after initiating all child IOs.
3749 ASSERT3U(dr->dr_dbuf->db.db_object, ==,
3750 DMU_META_DNODE_OBJECT);
3753 if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
3754 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
3755 VERIFY3U(dr->dr_dbuf->db_level, ==, level);
3757 list_remove(list, dr);
3758 if (dr->dr_dbuf->db_level > 0)
3759 dbuf_sync_indirect(dr, tx);
3761 dbuf_sync_leaf(dr, tx);
3767 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3769 dmu_buf_impl_t *db = vdb;
3771 blkptr_t *bp = zio->io_bp;
3772 blkptr_t *bp_orig = &zio->io_bp_orig;
3773 spa_t *spa = zio->io_spa;
3778 ASSERT3P(db->db_blkptr, !=, NULL);
3779 ASSERT3P(&db->db_data_pending->dr_bp_copy, ==, bp);
3783 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
3784 dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
3785 zio->io_prev_space_delta = delta;
3787 if (bp->blk_birth != 0) {
3788 ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
3789 BP_GET_TYPE(bp) == dn->dn_type) ||
3790 (db->db_blkid == DMU_SPILL_BLKID &&
3791 BP_GET_TYPE(bp) == dn->dn_bonustype) ||
3792 BP_IS_EMBEDDED(bp));
3793 ASSERT(BP_GET_LEVEL(bp) == db->db_level);
3796 mutex_enter(&db->db_mtx);
3799 if (db->db_blkid == DMU_SPILL_BLKID) {
3800 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3801 ASSERT(!(BP_IS_HOLE(bp)) &&
3802 db->db_blkptr == DN_SPILL_BLKPTR(dn->dn_phys));
3806 if (db->db_level == 0) {
3807 mutex_enter(&dn->dn_mtx);
3808 if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
3809 db->db_blkid != DMU_SPILL_BLKID)
3810 dn->dn_phys->dn_maxblkid = db->db_blkid;
3811 mutex_exit(&dn->dn_mtx);
3813 if (dn->dn_type == DMU_OT_DNODE) {
3815 while (i < db->db.db_size) {
3817 (void *)(((char *)db->db.db_data) + i);
3819 i += DNODE_MIN_SIZE;
3820 if (dnp->dn_type != DMU_OT_NONE) {
3822 i += dnp->dn_extra_slots *
3827 if (BP_IS_HOLE(bp)) {
3834 blkptr_t *ibp = db->db.db_data;
3835 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3836 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
3837 if (BP_IS_HOLE(ibp))
3839 fill += BP_GET_FILL(ibp);
3844 if (!BP_IS_EMBEDDED(bp))
3845 BP_SET_FILL(bp, fill);
3847 mutex_exit(&db->db_mtx);
3849 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3850 *db->db_blkptr = *bp;
3851 rw_exit(&dn->dn_struct_rwlock);
3856 * This function gets called just prior to running through the compression
3857 * stage of the zio pipeline. If we're an indirect block comprised of only
3858 * holes, then we want this indirect to be compressed away to a hole. In
3859 * order to do that we must zero out any information about the holes that
3860 * this indirect points to prior to before we try to compress it.
3863 dbuf_write_children_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3865 dmu_buf_impl_t *db = vdb;
3868 unsigned int epbs, i;
3870 ASSERT3U(db->db_level, >, 0);
3873 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3874 ASSERT3U(epbs, <, 31);
3876 /* Determine if all our children are holes */
3877 for (i = 0, bp = db->db.db_data; i < 1ULL << epbs; i++, bp++) {
3878 if (!BP_IS_HOLE(bp))
3883 * If all the children are holes, then zero them all out so that
3884 * we may get compressed away.
3886 if (i == 1ULL << epbs) {
3888 * We only found holes. Grab the rwlock to prevent
3889 * anybody from reading the blocks we're about to
3892 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3893 bzero(db->db.db_data, db->db.db_size);
3894 rw_exit(&dn->dn_struct_rwlock);
3900 * The SPA will call this callback several times for each zio - once
3901 * for every physical child i/o (zio->io_phys_children times). This
3902 * allows the DMU to monitor the progress of each logical i/o. For example,
3903 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
3904 * block. There may be a long delay before all copies/fragments are completed,
3905 * so this callback allows us to retire dirty space gradually, as the physical
3910 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
3912 dmu_buf_impl_t *db = arg;
3913 objset_t *os = db->db_objset;
3914 dsl_pool_t *dp = dmu_objset_pool(os);
3915 dbuf_dirty_record_t *dr;
3918 dr = db->db_data_pending;
3919 ASSERT3U(dr->dr_txg, ==, zio->io_txg);
3922 * The callback will be called io_phys_children times. Retire one
3923 * portion of our dirty space each time we are called. Any rounding
3924 * error will be cleaned up by dsl_pool_sync()'s call to
3925 * dsl_pool_undirty_space().
3927 delta = dr->dr_accounted / zio->io_phys_children;
3928 dsl_pool_undirty_space(dp, delta, zio->io_txg);
3933 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
3935 dmu_buf_impl_t *db = vdb;
3936 blkptr_t *bp_orig = &zio->io_bp_orig;
3937 blkptr_t *bp = db->db_blkptr;
3938 objset_t *os = db->db_objset;
3939 dmu_tx_t *tx = os->os_synctx;
3940 dbuf_dirty_record_t **drp, *dr;
3942 ASSERT0(zio->io_error);
3943 ASSERT(db->db_blkptr == bp);
3946 * For nopwrites and rewrites we ensure that the bp matches our
3947 * original and bypass all the accounting.
3949 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
3950 ASSERT(BP_EQUAL(bp, bp_orig));
3952 dsl_dataset_t *ds = os->os_dsl_dataset;
3953 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
3954 dsl_dataset_block_born(ds, bp, tx);
3957 mutex_enter(&db->db_mtx);
3961 drp = &db->db_last_dirty;
3962 while ((dr = *drp) != db->db_data_pending)
3964 ASSERT(!list_link_active(&dr->dr_dirty_node));
3965 ASSERT(dr->dr_dbuf == db);
3966 ASSERT(dr->dr_next == NULL);
3970 if (db->db_blkid == DMU_SPILL_BLKID) {
3975 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3976 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
3977 db->db_blkptr == DN_SPILL_BLKPTR(dn->dn_phys));
3982 if (db->db_level == 0) {
3983 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
3984 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
3985 if (db->db_state != DB_NOFILL) {
3986 if (dr->dt.dl.dr_data != db->db_buf)
3987 arc_buf_destroy(dr->dt.dl.dr_data, db);
3994 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3995 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
3996 if (!BP_IS_HOLE(db->db_blkptr)) {
3997 ASSERTV(int epbs = dn->dn_phys->dn_indblkshift -
3999 ASSERT3U(db->db_blkid, <=,
4000 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs));
4001 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
4005 mutex_destroy(&dr->dt.di.dr_mtx);
4006 list_destroy(&dr->dt.di.dr_children);
4008 kmem_free(dr, sizeof (dbuf_dirty_record_t));
4010 cv_broadcast(&db->db_changed);
4011 ASSERT(db->db_dirtycnt > 0);
4012 db->db_dirtycnt -= 1;
4013 db->db_data_pending = NULL;
4014 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg, B_FALSE);
4018 dbuf_write_nofill_ready(zio_t *zio)
4020 dbuf_write_ready(zio, NULL, zio->io_private);
4024 dbuf_write_nofill_done(zio_t *zio)
4026 dbuf_write_done(zio, NULL, zio->io_private);
4030 dbuf_write_override_ready(zio_t *zio)
4032 dbuf_dirty_record_t *dr = zio->io_private;
4033 dmu_buf_impl_t *db = dr->dr_dbuf;
4035 dbuf_write_ready(zio, NULL, db);
4039 dbuf_write_override_done(zio_t *zio)
4041 dbuf_dirty_record_t *dr = zio->io_private;
4042 dmu_buf_impl_t *db = dr->dr_dbuf;
4043 blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
4045 mutex_enter(&db->db_mtx);
4046 if (!BP_EQUAL(zio->io_bp, obp)) {
4047 if (!BP_IS_HOLE(obp))
4048 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
4049 arc_release(dr->dt.dl.dr_data, db);
4051 mutex_exit(&db->db_mtx);
4053 dbuf_write_done(zio, NULL, db);
4055 if (zio->io_abd != NULL)
4056 abd_put(zio->io_abd);
4059 typedef struct dbuf_remap_impl_callback_arg {
4061 uint64_t drica_blk_birth;
4063 } dbuf_remap_impl_callback_arg_t;
4066 dbuf_remap_impl_callback(uint64_t vdev, uint64_t offset, uint64_t size,
4069 dbuf_remap_impl_callback_arg_t *drica = arg;
4070 objset_t *os = drica->drica_os;
4071 spa_t *spa = dmu_objset_spa(os);
4072 dmu_tx_t *tx = drica->drica_tx;
4074 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)));
4076 if (os == spa_meta_objset(spa)) {
4077 spa_vdev_indirect_mark_obsolete(spa, vdev, offset, size, tx);
4079 dsl_dataset_block_remapped(dmu_objset_ds(os), vdev, offset,
4080 size, drica->drica_blk_birth, tx);
4085 dbuf_remap_impl(dnode_t *dn, blkptr_t *bp, dmu_tx_t *tx)
4087 blkptr_t bp_copy = *bp;
4088 spa_t *spa = dmu_objset_spa(dn->dn_objset);
4089 dbuf_remap_impl_callback_arg_t drica;
4091 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)));
4093 drica.drica_os = dn->dn_objset;
4094 drica.drica_blk_birth = bp->blk_birth;
4095 drica.drica_tx = tx;
4096 if (spa_remap_blkptr(spa, &bp_copy, dbuf_remap_impl_callback,
4099 * The struct_rwlock prevents dbuf_read_impl() from
4100 * dereferencing the BP while we are changing it. To
4101 * avoid lock contention, only grab it when we are actually
4104 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
4106 rw_exit(&dn->dn_struct_rwlock);
4111 * Returns true if a dbuf_remap would modify the dbuf. We do this by attempting
4112 * to remap a copy of every bp in the dbuf.
4115 dbuf_can_remap(const dmu_buf_impl_t *db)
4117 spa_t *spa = dmu_objset_spa(db->db_objset);
4118 blkptr_t *bp = db->db.db_data;
4119 boolean_t ret = B_FALSE;
4121 ASSERT3U(db->db_level, >, 0);
4122 ASSERT3S(db->db_state, ==, DB_CACHED);
4124 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL));
4126 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
4127 for (int i = 0; i < db->db.db_size >> SPA_BLKPTRSHIFT; i++) {
4128 blkptr_t bp_copy = bp[i];
4129 if (spa_remap_blkptr(spa, &bp_copy, NULL, NULL)) {
4134 spa_config_exit(spa, SCL_VDEV, FTAG);
4140 dnode_needs_remap(const dnode_t *dn)
4142 spa_t *spa = dmu_objset_spa(dn->dn_objset);
4143 boolean_t ret = B_FALSE;
4145 if (dn->dn_phys->dn_nlevels == 0) {
4149 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL));
4151 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
4152 for (int j = 0; j < dn->dn_phys->dn_nblkptr; j++) {
4153 blkptr_t bp_copy = dn->dn_phys->dn_blkptr[j];
4154 if (spa_remap_blkptr(spa, &bp_copy, NULL, NULL)) {
4159 spa_config_exit(spa, SCL_VDEV, FTAG);
4165 * Remap any existing BP's to concrete vdevs, if possible.
4168 dbuf_remap(dnode_t *dn, dmu_buf_impl_t *db, dmu_tx_t *tx)
4170 spa_t *spa = dmu_objset_spa(db->db_objset);
4171 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)));
4173 if (!spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL))
4176 if (db->db_level > 0) {
4177 blkptr_t *bp = db->db.db_data;
4178 for (int i = 0; i < db->db.db_size >> SPA_BLKPTRSHIFT; i++) {
4179 dbuf_remap_impl(dn, &bp[i], tx);
4181 } else if (db->db.db_object == DMU_META_DNODE_OBJECT) {
4182 dnode_phys_t *dnp = db->db.db_data;
4183 ASSERT3U(db->db_dnode_handle->dnh_dnode->dn_type, ==,
4185 for (int i = 0; i < db->db.db_size >> DNODE_SHIFT;
4186 i += dnp[i].dn_extra_slots + 1) {
4187 for (int j = 0; j < dnp[i].dn_nblkptr; j++) {
4188 dbuf_remap_impl(dn, &dnp[i].dn_blkptr[j], tx);
4195 /* Issue I/O to commit a dirty buffer to disk. */
4197 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
4199 dmu_buf_impl_t *db = dr->dr_dbuf;
4202 dmu_buf_impl_t *parent = db->db_parent;
4203 uint64_t txg = tx->tx_txg;
4204 zbookmark_phys_t zb;
4209 ASSERT(dmu_tx_is_syncing(tx));
4215 if (db->db_state != DB_NOFILL) {
4216 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
4218 * Private object buffers are released here rather
4219 * than in dbuf_dirty() since they are only modified
4220 * in the syncing context and we don't want the
4221 * overhead of making multiple copies of the data.
4223 if (BP_IS_HOLE(db->db_blkptr)) {
4226 dbuf_release_bp(db);
4228 dbuf_remap(dn, db, tx);
4232 if (parent != dn->dn_dbuf) {
4233 /* Our parent is an indirect block. */
4234 /* We have a dirty parent that has been scheduled for write. */
4235 ASSERT(parent && parent->db_data_pending);
4236 /* Our parent's buffer is one level closer to the dnode. */
4237 ASSERT(db->db_level == parent->db_level-1);
4239 * We're about to modify our parent's db_data by modifying
4240 * our block pointer, so the parent must be released.
4242 ASSERT(arc_released(parent->db_buf));
4243 zio = parent->db_data_pending->dr_zio;
4245 /* Our parent is the dnode itself. */
4246 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
4247 db->db_blkid != DMU_SPILL_BLKID) ||
4248 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
4249 if (db->db_blkid != DMU_SPILL_BLKID)
4250 ASSERT3P(db->db_blkptr, ==,
4251 &dn->dn_phys->dn_blkptr[db->db_blkid]);
4255 ASSERT(db->db_level == 0 || data == db->db_buf);
4256 ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
4259 SET_BOOKMARK(&zb, os->os_dsl_dataset ?
4260 os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
4261 db->db.db_object, db->db_level, db->db_blkid);
4263 if (db->db_blkid == DMU_SPILL_BLKID)
4265 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
4267 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
4271 * We copy the blkptr now (rather than when we instantiate the dirty
4272 * record), because its value can change between open context and
4273 * syncing context. We do not need to hold dn_struct_rwlock to read
4274 * db_blkptr because we are in syncing context.
4276 dr->dr_bp_copy = *db->db_blkptr;
4278 if (db->db_level == 0 &&
4279 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
4281 * The BP for this block has been provided by open context
4282 * (by dmu_sync() or dmu_buf_write_embedded()).
4284 abd_t *contents = (data != NULL) ?
4285 abd_get_from_buf(data->b_data, arc_buf_size(data)) : NULL;
4287 dr->dr_zio = zio_write(zio, os->os_spa, txg,
4288 &dr->dr_bp_copy, contents, db->db.db_size, db->db.db_size,
4289 &zp, dbuf_write_override_ready, NULL, NULL,
4290 dbuf_write_override_done,
4291 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
4292 mutex_enter(&db->db_mtx);
4293 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
4294 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
4295 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
4296 mutex_exit(&db->db_mtx);
4297 } else if (db->db_state == DB_NOFILL) {
4298 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF ||
4299 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY);
4300 dr->dr_zio = zio_write(zio, os->os_spa, txg,
4301 &dr->dr_bp_copy, NULL, db->db.db_size, db->db.db_size, &zp,
4302 dbuf_write_nofill_ready, NULL, NULL,
4303 dbuf_write_nofill_done, db,
4304 ZIO_PRIORITY_ASYNC_WRITE,
4305 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
4307 ASSERT(arc_released(data));
4310 * For indirect blocks, we want to setup the children
4311 * ready callback so that we can properly handle an indirect
4312 * block that only contains holes.
4314 arc_write_done_func_t *children_ready_cb = NULL;
4315 if (db->db_level != 0)
4316 children_ready_cb = dbuf_write_children_ready;
4318 dr->dr_zio = arc_write(zio, os->os_spa, txg,
4319 &dr->dr_bp_copy, data, DBUF_IS_L2CACHEABLE(db),
4320 &zp, dbuf_write_ready,
4321 children_ready_cb, dbuf_write_physdone,
4322 dbuf_write_done, db, ZIO_PRIORITY_ASYNC_WRITE,
4323 ZIO_FLAG_MUSTSUCCEED, &zb);
4327 #if defined(_KERNEL)
4328 EXPORT_SYMBOL(dbuf_find);
4329 EXPORT_SYMBOL(dbuf_is_metadata);
4330 EXPORT_SYMBOL(dbuf_destroy);
4331 EXPORT_SYMBOL(dbuf_loan_arcbuf);
4332 EXPORT_SYMBOL(dbuf_whichblock);
4333 EXPORT_SYMBOL(dbuf_read);
4334 EXPORT_SYMBOL(dbuf_unoverride);
4335 EXPORT_SYMBOL(dbuf_free_range);
4336 EXPORT_SYMBOL(dbuf_new_size);
4337 EXPORT_SYMBOL(dbuf_release_bp);
4338 EXPORT_SYMBOL(dbuf_dirty);
4339 EXPORT_SYMBOL(dmu_buf_set_crypt_params);
4340 EXPORT_SYMBOL(dmu_buf_will_dirty);
4341 EXPORT_SYMBOL(dmu_buf_will_not_fill);
4342 EXPORT_SYMBOL(dmu_buf_will_fill);
4343 EXPORT_SYMBOL(dmu_buf_fill_done);
4344 EXPORT_SYMBOL(dmu_buf_rele);
4345 EXPORT_SYMBOL(dbuf_assign_arcbuf);
4346 EXPORT_SYMBOL(dbuf_prefetch);
4347 EXPORT_SYMBOL(dbuf_hold_impl);
4348 EXPORT_SYMBOL(dbuf_hold);
4349 EXPORT_SYMBOL(dbuf_hold_level);
4350 EXPORT_SYMBOL(dbuf_create_bonus);
4351 EXPORT_SYMBOL(dbuf_spill_set_blksz);
4352 EXPORT_SYMBOL(dbuf_rm_spill);
4353 EXPORT_SYMBOL(dbuf_add_ref);
4354 EXPORT_SYMBOL(dbuf_rele);
4355 EXPORT_SYMBOL(dbuf_rele_and_unlock);
4356 EXPORT_SYMBOL(dbuf_refcount);
4357 EXPORT_SYMBOL(dbuf_sync_list);
4358 EXPORT_SYMBOL(dmu_buf_set_user);
4359 EXPORT_SYMBOL(dmu_buf_set_user_ie);
4360 EXPORT_SYMBOL(dmu_buf_get_user);
4361 EXPORT_SYMBOL(dmu_buf_get_blkptr);
4364 module_param(dbuf_cache_max_bytes, ulong, 0644);
4365 MODULE_PARM_DESC(dbuf_cache_max_bytes,
4366 "Maximum size in bytes of the dbuf cache.");
4368 module_param(dbuf_cache_hiwater_pct, uint, 0644);
4369 MODULE_PARM_DESC(dbuf_cache_hiwater_pct,
4370 "Percentage over dbuf_cache_max_bytes when dbufs must be evicted "
4373 module_param(dbuf_cache_lowater_pct, uint, 0644);
4374 MODULE_PARM_DESC(dbuf_cache_lowater_pct,
4375 "Percentage below dbuf_cache_max_bytes when the evict thread stops "
4378 module_param(dbuf_cache_shift, int, 0644);
4379 MODULE_PARM_DESC(dbuf_cache_shift,
4380 "Set the size of the dbuf cache to a log2 fraction of arc size.");