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 struct dbuf_hold_impl_data {
52 /* Function arguments */
56 boolean_t dh_fail_sparse;
57 boolean_t dh_fail_uncached;
59 dmu_buf_impl_t **dh_dbp;
61 dmu_buf_impl_t *dh_db;
62 dmu_buf_impl_t *dh_parent;
65 dbuf_dirty_record_t *dh_dr;
66 arc_buf_contents_t dh_type;
70 static void __dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh,
71 dnode_t *dn, uint8_t level, uint64_t blkid, boolean_t fail_sparse,
72 boolean_t fail_uncached,
73 void *tag, dmu_buf_impl_t **dbp, int depth);
74 static int __dbuf_hold_impl(struct dbuf_hold_impl_data *dh);
76 uint_t zfs_dbuf_evict_key;
78 static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
79 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
81 extern inline void dmu_buf_init_user(dmu_buf_user_t *dbu,
82 dmu_buf_evict_func_t *evict_func_sync,
83 dmu_buf_evict_func_t *evict_func_async,
84 dmu_buf_t **clear_on_evict_dbufp);
87 * Global data structures and functions for the dbuf cache.
89 static kmem_cache_t *dbuf_kmem_cache;
90 static taskq_t *dbu_evict_taskq;
92 static kthread_t *dbuf_cache_evict_thread;
93 static kmutex_t dbuf_evict_lock;
94 static kcondvar_t dbuf_evict_cv;
95 static boolean_t dbuf_evict_thread_exit;
98 * LRU cache of dbufs. The dbuf cache maintains a list of dbufs that
99 * are not currently held but have been recently released. These dbufs
100 * are not eligible for arc eviction until they are aged out of the cache.
101 * Dbufs are added to the dbuf cache once the last hold is released. If a
102 * dbuf is later accessed and still exists in the dbuf cache, then it will
103 * be removed from the cache and later re-added to the head of the cache.
104 * Dbufs that are aged out of the cache will be immediately destroyed and
105 * become eligible for arc eviction.
107 static multilist_t *dbuf_cache;
108 static refcount_t dbuf_cache_size;
109 unsigned long dbuf_cache_max_bytes = 100 * 1024 * 1024;
111 /* Cap the size of the dbuf cache to log2 fraction of arc size. */
112 int dbuf_cache_max_shift = 5;
115 * The dbuf cache uses a three-stage eviction policy:
116 * - A low water marker designates when the dbuf eviction thread
117 * should stop evicting from the dbuf cache.
118 * - When we reach the maximum size (aka mid water mark), we
119 * signal the eviction thread to run.
120 * - The high water mark indicates when the eviction thread
121 * is unable to keep up with the incoming load and eviction must
122 * happen in the context of the calling thread.
126 * low water mid water hi water
127 * +----------------------------------------+----------+----------+
132 * +----------------------------------------+----------+----------+
134 * evicting eviction directly
137 * The high and low water marks indicate the operating range for the eviction
138 * thread. The low water mark is, by default, 90% of the total size of the
139 * cache and the high water mark is at 110% (both of these percentages can be
140 * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct,
141 * respectively). The eviction thread will try to ensure that the cache remains
142 * within this range by waking up every second and checking if the cache is
143 * above the low water mark. The thread can also be woken up by callers adding
144 * elements into the cache if the cache is larger than the mid water (i.e max
145 * cache size). Once the eviction thread is woken up and eviction is required,
146 * it will continue evicting buffers until it's able to reduce the cache size
147 * to the low water mark. If the cache size continues to grow and hits the high
148 * water mark, then callers adding elements to the cache will begin to evict
149 * directly from the cache until the cache is no longer above the high water
154 * The percentage above and below the maximum cache size.
156 uint_t dbuf_cache_hiwater_pct = 10;
157 uint_t dbuf_cache_lowater_pct = 10;
161 dbuf_cons(void *vdb, void *unused, int kmflag)
163 dmu_buf_impl_t *db = vdb;
164 bzero(db, sizeof (dmu_buf_impl_t));
166 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
167 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
168 multilist_link_init(&db->db_cache_link);
169 refcount_create(&db->db_holds);
176 dbuf_dest(void *vdb, void *unused)
178 dmu_buf_impl_t *db = vdb;
179 mutex_destroy(&db->db_mtx);
180 cv_destroy(&db->db_changed);
181 ASSERT(!multilist_link_active(&db->db_cache_link));
182 refcount_destroy(&db->db_holds);
186 * dbuf hash table routines
188 static dbuf_hash_table_t dbuf_hash_table;
190 static uint64_t dbuf_hash_count;
193 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
195 uintptr_t osv = (uintptr_t)os;
196 uint64_t crc = -1ULL;
198 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
199 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF];
200 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF];
201 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF];
202 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF];
203 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF];
204 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF];
206 crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16);
211 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
212 ((dbuf)->db.db_object == (obj) && \
213 (dbuf)->db_objset == (os) && \
214 (dbuf)->db_level == (level) && \
215 (dbuf)->db_blkid == (blkid))
218 dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid)
220 dbuf_hash_table_t *h = &dbuf_hash_table;
225 hv = dbuf_hash(os, obj, level, blkid);
226 idx = hv & h->hash_table_mask;
228 mutex_enter(DBUF_HASH_MUTEX(h, idx));
229 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
230 if (DBUF_EQUAL(db, os, obj, level, blkid)) {
231 mutex_enter(&db->db_mtx);
232 if (db->db_state != DB_EVICTING) {
233 mutex_exit(DBUF_HASH_MUTEX(h, idx));
236 mutex_exit(&db->db_mtx);
239 mutex_exit(DBUF_HASH_MUTEX(h, idx));
243 static dmu_buf_impl_t *
244 dbuf_find_bonus(objset_t *os, uint64_t object)
247 dmu_buf_impl_t *db = NULL;
249 if (dnode_hold(os, object, FTAG, &dn) == 0) {
250 rw_enter(&dn->dn_struct_rwlock, RW_READER);
251 if (dn->dn_bonus != NULL) {
253 mutex_enter(&db->db_mtx);
255 rw_exit(&dn->dn_struct_rwlock);
256 dnode_rele(dn, FTAG);
262 * Insert an entry into the hash table. If there is already an element
263 * equal to elem in the hash table, then the already existing element
264 * will be returned and the new element will not be inserted.
265 * Otherwise returns NULL.
267 static dmu_buf_impl_t *
268 dbuf_hash_insert(dmu_buf_impl_t *db)
270 dbuf_hash_table_t *h = &dbuf_hash_table;
271 objset_t *os = db->db_objset;
272 uint64_t obj = db->db.db_object;
273 int level = db->db_level;
274 uint64_t blkid, hv, idx;
277 blkid = db->db_blkid;
278 hv = dbuf_hash(os, obj, level, blkid);
279 idx = hv & h->hash_table_mask;
281 mutex_enter(DBUF_HASH_MUTEX(h, idx));
282 for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) {
283 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
284 mutex_enter(&dbf->db_mtx);
285 if (dbf->db_state != DB_EVICTING) {
286 mutex_exit(DBUF_HASH_MUTEX(h, idx));
289 mutex_exit(&dbf->db_mtx);
293 mutex_enter(&db->db_mtx);
294 db->db_hash_next = h->hash_table[idx];
295 h->hash_table[idx] = db;
296 mutex_exit(DBUF_HASH_MUTEX(h, idx));
297 atomic_inc_64(&dbuf_hash_count);
303 * Remove an entry from the hash table. It must be in the EVICTING state.
306 dbuf_hash_remove(dmu_buf_impl_t *db)
308 dbuf_hash_table_t *h = &dbuf_hash_table;
310 dmu_buf_impl_t *dbf, **dbp;
312 hv = dbuf_hash(db->db_objset, db->db.db_object,
313 db->db_level, db->db_blkid);
314 idx = hv & h->hash_table_mask;
317 * We mustn't hold db_mtx to maintain lock ordering:
318 * DBUF_HASH_MUTEX > db_mtx.
320 ASSERT(refcount_is_zero(&db->db_holds));
321 ASSERT(db->db_state == DB_EVICTING);
322 ASSERT(!MUTEX_HELD(&db->db_mtx));
324 mutex_enter(DBUF_HASH_MUTEX(h, idx));
325 dbp = &h->hash_table[idx];
326 while ((dbf = *dbp) != db) {
327 dbp = &dbf->db_hash_next;
330 *dbp = db->db_hash_next;
331 db->db_hash_next = NULL;
332 mutex_exit(DBUF_HASH_MUTEX(h, idx));
333 atomic_dec_64(&dbuf_hash_count);
339 } dbvu_verify_type_t;
342 dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type)
347 if (db->db_user == NULL)
350 /* Only data blocks support the attachment of user data. */
351 ASSERT(db->db_level == 0);
353 /* Clients must resolve a dbuf before attaching user data. */
354 ASSERT(db->db.db_data != NULL);
355 ASSERT3U(db->db_state, ==, DB_CACHED);
357 holds = refcount_count(&db->db_holds);
358 if (verify_type == DBVU_EVICTING) {
360 * Immediate eviction occurs when holds == dirtycnt.
361 * For normal eviction buffers, holds is zero on
362 * eviction, except when dbuf_fix_old_data() calls
363 * dbuf_clear_data(). However, the hold count can grow
364 * during eviction even though db_mtx is held (see
365 * dmu_bonus_hold() for an example), so we can only
366 * test the generic invariant that holds >= dirtycnt.
368 ASSERT3U(holds, >=, db->db_dirtycnt);
370 if (db->db_user_immediate_evict == TRUE)
371 ASSERT3U(holds, >=, db->db_dirtycnt);
373 ASSERT3U(holds, >, 0);
379 dbuf_evict_user(dmu_buf_impl_t *db)
381 dmu_buf_user_t *dbu = db->db_user;
383 ASSERT(MUTEX_HELD(&db->db_mtx));
388 dbuf_verify_user(db, DBVU_EVICTING);
392 if (dbu->dbu_clear_on_evict_dbufp != NULL)
393 *dbu->dbu_clear_on_evict_dbufp = NULL;
397 * There are two eviction callbacks - one that we call synchronously
398 * and one that we invoke via a taskq. The async one is useful for
399 * avoiding lock order reversals and limiting stack depth.
401 * Note that if we have a sync callback but no async callback,
402 * it's likely that the sync callback will free the structure
403 * containing the dbu. In that case we need to take care to not
404 * dereference dbu after calling the sync evict func.
406 boolean_t has_async = (dbu->dbu_evict_func_async != NULL);
408 if (dbu->dbu_evict_func_sync != NULL)
409 dbu->dbu_evict_func_sync(dbu);
412 taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func_async,
413 dbu, 0, &dbu->dbu_tqent);
418 dbuf_is_metadata(dmu_buf_impl_t *db)
421 * Consider indirect blocks and spill blocks to be meta data.
423 if (db->db_level > 0 || db->db_blkid == DMU_SPILL_BLKID) {
426 boolean_t is_metadata;
429 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type);
432 return (is_metadata);
438 * This function *must* return indices evenly distributed between all
439 * sublists of the multilist. This is needed due to how the dbuf eviction
440 * code is laid out; dbuf_evict_thread() assumes dbufs are evenly
441 * distributed between all sublists and uses this assumption when
442 * deciding which sublist to evict from and how much to evict from it.
445 dbuf_cache_multilist_index_func(multilist_t *ml, void *obj)
447 dmu_buf_impl_t *db = obj;
450 * The assumption here, is the hash value for a given
451 * dmu_buf_impl_t will remain constant throughout it's lifetime
452 * (i.e. it's objset, object, level and blkid fields don't change).
453 * Thus, we don't need to store the dbuf's sublist index
454 * on insertion, as this index can be recalculated on removal.
456 * Also, the low order bits of the hash value are thought to be
457 * distributed evenly. Otherwise, in the case that the multilist
458 * has a power of two number of sublists, each sublists' usage
459 * would not be evenly distributed.
461 return (dbuf_hash(db->db_objset, db->db.db_object,
462 db->db_level, db->db_blkid) %
463 multilist_get_num_sublists(ml));
466 static inline unsigned long
467 dbuf_cache_target_bytes(void)
469 return MIN(dbuf_cache_max_bytes,
470 arc_target_bytes() >> dbuf_cache_max_shift);
473 static inline boolean_t
474 dbuf_cache_above_hiwater(void)
476 uint64_t dbuf_cache_target = dbuf_cache_target_bytes();
478 uint64_t dbuf_cache_hiwater_bytes =
479 (dbuf_cache_target * dbuf_cache_hiwater_pct) / 100;
481 return (refcount_count(&dbuf_cache_size) >
482 dbuf_cache_target + dbuf_cache_hiwater_bytes);
485 static inline boolean_t
486 dbuf_cache_above_lowater(void)
488 uint64_t dbuf_cache_target = dbuf_cache_target_bytes();
490 uint64_t dbuf_cache_lowater_bytes =
491 (dbuf_cache_target * dbuf_cache_lowater_pct) / 100;
493 return (refcount_count(&dbuf_cache_size) >
494 dbuf_cache_target - dbuf_cache_lowater_bytes);
498 * Evict the oldest eligible dbuf from the dbuf cache.
503 int idx = multilist_get_random_index(dbuf_cache);
504 multilist_sublist_t *mls = multilist_sublist_lock(dbuf_cache, idx);
506 ASSERT(!MUTEX_HELD(&dbuf_evict_lock));
509 * Set the thread's tsd to indicate that it's processing evictions.
510 * Once a thread stops evicting from the dbuf cache it will
511 * reset its tsd to NULL.
513 ASSERT3P(tsd_get(zfs_dbuf_evict_key), ==, NULL);
514 (void) tsd_set(zfs_dbuf_evict_key, (void *)B_TRUE);
516 db = multilist_sublist_tail(mls);
517 while (db != NULL && mutex_tryenter(&db->db_mtx) == 0) {
518 db = multilist_sublist_prev(mls, db);
521 DTRACE_PROBE2(dbuf__evict__one, dmu_buf_impl_t *, db,
522 multilist_sublist_t *, mls);
525 multilist_sublist_remove(mls, db);
526 multilist_sublist_unlock(mls);
527 (void) refcount_remove_many(&dbuf_cache_size,
531 multilist_sublist_unlock(mls);
533 (void) tsd_set(zfs_dbuf_evict_key, NULL);
537 * The dbuf evict thread is responsible for aging out dbufs from the
538 * cache. Once the cache has reached it's maximum size, dbufs are removed
539 * and destroyed. The eviction thread will continue running until the size
540 * of the dbuf cache is at or below the maximum size. Once the dbuf is aged
541 * out of the cache it is destroyed and becomes eligible for arc eviction.
544 dbuf_evict_thread(void *unused)
548 CALLB_CPR_INIT(&cpr, &dbuf_evict_lock, callb_generic_cpr, FTAG);
550 mutex_enter(&dbuf_evict_lock);
551 while (!dbuf_evict_thread_exit) {
552 while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
553 CALLB_CPR_SAFE_BEGIN(&cpr);
554 (void) cv_timedwait_sig_hires(&dbuf_evict_cv,
555 &dbuf_evict_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
556 CALLB_CPR_SAFE_END(&cpr, &dbuf_evict_lock);
558 mutex_exit(&dbuf_evict_lock);
561 * Keep evicting as long as we're above the low water mark
562 * for the cache. We do this without holding the locks to
563 * minimize lock contention.
565 while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
569 mutex_enter(&dbuf_evict_lock);
572 dbuf_evict_thread_exit = B_FALSE;
573 cv_broadcast(&dbuf_evict_cv);
574 CALLB_CPR_EXIT(&cpr); /* drops dbuf_evict_lock */
579 * Wake up the dbuf eviction thread if the dbuf cache is at its max size.
580 * If the dbuf cache is at its high water mark, then evict a dbuf from the
581 * dbuf cache using the callers context.
584 dbuf_evict_notify(void)
588 * We use thread specific data to track when a thread has
589 * started processing evictions. This allows us to avoid deeply
590 * nested stacks that would have a call flow similar to this:
592 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify()
595 * +-----dbuf_destroy()<--dbuf_evict_one()<--------+
597 * The dbuf_eviction_thread will always have its tsd set until
598 * that thread exits. All other threads will only set their tsd
599 * if they are participating in the eviction process. This only
600 * happens if the eviction thread is unable to process evictions
601 * fast enough. To keep the dbuf cache size in check, other threads
602 * can evict from the dbuf cache directly. Those threads will set
603 * their tsd values so that we ensure that they only evict one dbuf
604 * from the dbuf cache.
606 if (tsd_get(zfs_dbuf_evict_key) != NULL)
610 * We check if we should evict without holding the dbuf_evict_lock,
611 * because it's OK to occasionally make the wrong decision here,
612 * and grabbing the lock results in massive lock contention.
614 if (refcount_count(&dbuf_cache_size) > dbuf_cache_target_bytes()) {
615 if (dbuf_cache_above_hiwater())
617 cv_signal(&dbuf_evict_cv);
626 uint64_t hsize = 1ULL << 16;
627 dbuf_hash_table_t *h = &dbuf_hash_table;
631 * The hash table is big enough to fill all of physical memory
632 * with an average block size of zfs_arc_average_blocksize (default 8K).
633 * By default, the table will take up
634 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
636 while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE)
640 h->hash_table_mask = hsize - 1;
641 #if defined(_KERNEL) && defined(HAVE_SPL)
643 * Large allocations which do not require contiguous pages
644 * should be using vmem_alloc() in the linux kernel
646 h->hash_table = vmem_zalloc(hsize * sizeof (void *), KM_SLEEP);
648 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
650 if (h->hash_table == NULL) {
651 /* XXX - we should really return an error instead of assert */
652 ASSERT(hsize > (1ULL << 10));
657 dbuf_kmem_cache = kmem_cache_create("dmu_buf_impl_t",
658 sizeof (dmu_buf_impl_t),
659 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
661 for (i = 0; i < DBUF_MUTEXES; i++)
662 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
667 * Setup the parameters for the dbuf cache. We cap the size of the
668 * dbuf cache to 1/32nd (default) of the size of the ARC.
670 dbuf_cache_max_bytes = MIN(dbuf_cache_max_bytes,
671 arc_target_bytes() >> dbuf_cache_max_shift);
674 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
675 * configuration is not required.
677 dbu_evict_taskq = taskq_create("dbu_evict", 1, defclsyspri, 0, 0, 0);
679 dbuf_cache = multilist_create(sizeof (dmu_buf_impl_t),
680 offsetof(dmu_buf_impl_t, db_cache_link),
681 dbuf_cache_multilist_index_func);
682 refcount_create(&dbuf_cache_size);
684 tsd_create(&zfs_dbuf_evict_key, NULL);
685 dbuf_evict_thread_exit = B_FALSE;
686 mutex_init(&dbuf_evict_lock, NULL, MUTEX_DEFAULT, NULL);
687 cv_init(&dbuf_evict_cv, NULL, CV_DEFAULT, NULL);
688 dbuf_cache_evict_thread = thread_create(NULL, 0, dbuf_evict_thread,
689 NULL, 0, &p0, TS_RUN, minclsyspri);
695 dbuf_hash_table_t *h = &dbuf_hash_table;
698 dbuf_stats_destroy();
700 for (i = 0; i < DBUF_MUTEXES; i++)
701 mutex_destroy(&h->hash_mutexes[i]);
702 #if defined(_KERNEL) && defined(HAVE_SPL)
704 * Large allocations which do not require contiguous pages
705 * should be using vmem_free() in the linux kernel
707 vmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
709 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
711 kmem_cache_destroy(dbuf_kmem_cache);
712 taskq_destroy(dbu_evict_taskq);
714 mutex_enter(&dbuf_evict_lock);
715 dbuf_evict_thread_exit = B_TRUE;
716 while (dbuf_evict_thread_exit) {
717 cv_signal(&dbuf_evict_cv);
718 cv_wait(&dbuf_evict_cv, &dbuf_evict_lock);
720 mutex_exit(&dbuf_evict_lock);
721 tsd_destroy(&zfs_dbuf_evict_key);
723 mutex_destroy(&dbuf_evict_lock);
724 cv_destroy(&dbuf_evict_cv);
726 refcount_destroy(&dbuf_cache_size);
727 multilist_destroy(dbuf_cache);
736 dbuf_verify(dmu_buf_impl_t *db)
739 dbuf_dirty_record_t *dr;
741 ASSERT(MUTEX_HELD(&db->db_mtx));
743 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
746 ASSERT(db->db_objset != NULL);
750 ASSERT(db->db_parent == NULL);
751 ASSERT(db->db_blkptr == NULL);
753 ASSERT3U(db->db.db_object, ==, dn->dn_object);
754 ASSERT3P(db->db_objset, ==, dn->dn_objset);
755 ASSERT3U(db->db_level, <, dn->dn_nlevels);
756 ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
757 db->db_blkid == DMU_SPILL_BLKID ||
758 !avl_is_empty(&dn->dn_dbufs));
760 if (db->db_blkid == DMU_BONUS_BLKID) {
762 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
763 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
764 } else if (db->db_blkid == DMU_SPILL_BLKID) {
766 ASSERT0(db->db.db_offset);
768 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
771 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
772 ASSERT(dr->dr_dbuf == db);
774 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
775 ASSERT(dr->dr_dbuf == db);
778 * We can't assert that db_size matches dn_datablksz because it
779 * can be momentarily different when another thread is doing
782 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
783 dr = db->db_data_pending;
785 * It should only be modified in syncing context, so
786 * make sure we only have one copy of the data.
788 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
791 /* verify db->db_blkptr */
793 if (db->db_parent == dn->dn_dbuf) {
794 /* db is pointed to by the dnode */
795 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
796 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
797 ASSERT(db->db_parent == NULL);
799 ASSERT(db->db_parent != NULL);
800 if (db->db_blkid != DMU_SPILL_BLKID)
801 ASSERT3P(db->db_blkptr, ==,
802 &dn->dn_phys->dn_blkptr[db->db_blkid]);
804 /* db is pointed to by an indirect block */
805 ASSERTV(int epb = db->db_parent->db.db_size >>
807 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
808 ASSERT3U(db->db_parent->db.db_object, ==,
811 * dnode_grow_indblksz() can make this fail if we don't
812 * have the struct_rwlock. XXX indblksz no longer
813 * grows. safe to do this now?
815 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
816 ASSERT3P(db->db_blkptr, ==,
817 ((blkptr_t *)db->db_parent->db.db_data +
818 db->db_blkid % epb));
822 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
823 (db->db_buf == NULL || db->db_buf->b_data) &&
824 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
825 db->db_state != DB_FILL && !dn->dn_free_txg) {
827 * If the blkptr isn't set but they have nonzero data,
828 * it had better be dirty, otherwise we'll lose that
829 * data when we evict this buffer.
831 * There is an exception to this rule for indirect blocks; in
832 * this case, if the indirect block is a hole, we fill in a few
833 * fields on each of the child blocks (importantly, birth time)
834 * to prevent hole birth times from being lost when you
835 * partially fill in a hole.
837 if (db->db_dirtycnt == 0) {
838 if (db->db_level == 0) {
839 uint64_t *buf = db->db.db_data;
842 for (i = 0; i < db->db.db_size >> 3; i++) {
847 blkptr_t *bps = db->db.db_data;
848 ASSERT3U(1 << DB_DNODE(db)->dn_indblkshift, ==,
851 * We want to verify that all the blkptrs in the
852 * indirect block are holes, but we may have
853 * automatically set up a few fields for them.
854 * We iterate through each blkptr and verify
855 * they only have those fields set.
858 i < db->db.db_size / sizeof (blkptr_t);
860 blkptr_t *bp = &bps[i];
861 ASSERT(ZIO_CHECKSUM_IS_ZERO(
864 DVA_IS_EMPTY(&bp->blk_dva[0]) &&
865 DVA_IS_EMPTY(&bp->blk_dva[1]) &&
866 DVA_IS_EMPTY(&bp->blk_dva[2]));
867 ASSERT0(bp->blk_fill);
868 ASSERT0(bp->blk_pad[0]);
869 ASSERT0(bp->blk_pad[1]);
870 ASSERT(!BP_IS_EMBEDDED(bp));
871 ASSERT(BP_IS_HOLE(bp));
872 ASSERT0(bp->blk_phys_birth);
882 dbuf_clear_data(dmu_buf_impl_t *db)
884 ASSERT(MUTEX_HELD(&db->db_mtx));
886 ASSERT3P(db->db_buf, ==, NULL);
887 db->db.db_data = NULL;
888 if (db->db_state != DB_NOFILL)
889 db->db_state = DB_UNCACHED;
893 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
895 ASSERT(MUTEX_HELD(&db->db_mtx));
899 ASSERT(buf->b_data != NULL);
900 db->db.db_data = buf->b_data;
904 * Loan out an arc_buf for read. Return the loaned arc_buf.
907 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
911 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
912 mutex_enter(&db->db_mtx);
913 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
914 int blksz = db->db.db_size;
915 spa_t *spa = db->db_objset->os_spa;
917 mutex_exit(&db->db_mtx);
918 abuf = arc_loan_buf(spa, B_FALSE, blksz);
919 bcopy(db->db.db_data, abuf->b_data, blksz);
922 arc_loan_inuse_buf(abuf, db);
925 mutex_exit(&db->db_mtx);
931 * Calculate which level n block references the data at the level 0 offset
935 dbuf_whichblock(const dnode_t *dn, const int64_t level, const uint64_t offset)
937 if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) {
939 * The level n blkid is equal to the level 0 blkid divided by
940 * the number of level 0s in a level n block.
942 * The level 0 blkid is offset >> datablkshift =
943 * offset / 2^datablkshift.
945 * The number of level 0s in a level n is the number of block
946 * pointers in an indirect block, raised to the power of level.
947 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
948 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
950 * Thus, the level n blkid is: offset /
951 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
952 * = offset / 2^(datablkshift + level *
953 * (indblkshift - SPA_BLKPTRSHIFT))
954 * = offset >> (datablkshift + level *
955 * (indblkshift - SPA_BLKPTRSHIFT))
958 const unsigned exp = dn->dn_datablkshift +
959 level * (dn->dn_indblkshift - SPA_BLKPTRSHIFT);
961 if (exp >= 8 * sizeof (offset)) {
962 /* This only happens on the highest indirection level */
963 ASSERT3U(level, ==, dn->dn_nlevels - 1);
967 ASSERT3U(exp, <, 8 * sizeof (offset));
969 return (offset >> exp);
971 ASSERT3U(offset, <, dn->dn_datablksz);
977 dbuf_read_done(zio_t *zio, int err, arc_buf_t *buf, void *vdb)
979 dmu_buf_impl_t *db = vdb;
981 mutex_enter(&db->db_mtx);
982 ASSERT3U(db->db_state, ==, DB_READ);
984 * All reads are synchronous, so we must have a hold on the dbuf
986 ASSERT(refcount_count(&db->db_holds) > 0);
987 ASSERT(db->db_buf == NULL);
988 ASSERT(db->db.db_data == NULL);
989 if (db->db_level == 0 && db->db_freed_in_flight) {
990 /* we were freed in flight; disregard any error */
991 arc_release(buf, db);
992 bzero(buf->b_data, db->db.db_size);
994 db->db_freed_in_flight = FALSE;
995 dbuf_set_data(db, buf);
996 db->db_state = DB_CACHED;
997 } else if (err == 0) {
998 dbuf_set_data(db, buf);
999 db->db_state = DB_CACHED;
1001 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1002 ASSERT3P(db->db_buf, ==, NULL);
1003 arc_buf_destroy(buf, db);
1004 db->db_state = DB_UNCACHED;
1006 cv_broadcast(&db->db_changed);
1007 dbuf_rele_and_unlock(db, NULL);
1011 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
1014 zbookmark_phys_t zb;
1015 uint32_t aflags = ARC_FLAG_NOWAIT;
1016 int err, zio_flags = 0;
1020 ASSERT(!refcount_is_zero(&db->db_holds));
1021 /* We need the struct_rwlock to prevent db_blkptr from changing. */
1022 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1023 ASSERT(MUTEX_HELD(&db->db_mtx));
1024 ASSERT(db->db_state == DB_UNCACHED);
1025 ASSERT(db->db_buf == NULL);
1027 if (db->db_blkid == DMU_BONUS_BLKID) {
1029 * The bonus length stored in the dnode may be less than
1030 * the maximum available space in the bonus buffer.
1032 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
1033 int max_bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
1034 arc_buf_t *dn_buf = (dn->dn_dbuf != NULL) ?
1035 dn->dn_dbuf->db_buf : NULL;
1037 /* if the underlying dnode block is encrypted, decrypt it */
1038 if (dn_buf != NULL && dn->dn_objset->os_encrypted &&
1039 DMU_OT_IS_ENCRYPTED(dn->dn_bonustype) &&
1040 (flags & DB_RF_NO_DECRYPT) == 0 &&
1041 arc_is_encrypted(dn_buf)) {
1042 err = arc_untransform(dn_buf, dn->dn_objset->os_spa,
1043 dmu_objset_id(dn->dn_objset), B_TRUE);
1046 mutex_exit(&db->db_mtx);
1051 ASSERT3U(bonuslen, <=, db->db.db_size);
1052 db->db.db_data = kmem_alloc(max_bonuslen, KM_SLEEP);
1053 arc_space_consume(max_bonuslen, ARC_SPACE_BONUS);
1054 if (bonuslen < max_bonuslen)
1055 bzero(db->db.db_data, max_bonuslen);
1057 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
1059 db->db_state = DB_CACHED;
1060 mutex_exit(&db->db_mtx);
1065 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
1066 * processes the delete record and clears the bp while we are waiting
1067 * for the dn_mtx (resulting in a "no" from block_freed).
1069 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
1070 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
1071 BP_IS_HOLE(db->db_blkptr)))) {
1072 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1074 dbuf_set_data(db, arc_alloc_buf(db->db_objset->os_spa, db, type,
1076 bzero(db->db.db_data, db->db.db_size);
1078 if (db->db_blkptr != NULL && db->db_level > 0 &&
1079 BP_IS_HOLE(db->db_blkptr) &&
1080 db->db_blkptr->blk_birth != 0) {
1081 blkptr_t *bps = db->db.db_data;
1083 for (i = 0; i < ((1 <<
1084 DB_DNODE(db)->dn_indblkshift) / sizeof (blkptr_t));
1086 blkptr_t *bp = &bps[i];
1087 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
1088 1 << dn->dn_indblkshift);
1090 BP_GET_LEVEL(db->db_blkptr) == 1 ?
1092 BP_GET_LSIZE(db->db_blkptr));
1093 BP_SET_TYPE(bp, BP_GET_TYPE(db->db_blkptr));
1095 BP_GET_LEVEL(db->db_blkptr) - 1);
1096 BP_SET_BIRTH(bp, db->db_blkptr->blk_birth, 0);
1100 db->db_state = DB_CACHED;
1101 mutex_exit(&db->db_mtx);
1107 db->db_state = DB_READ;
1108 mutex_exit(&db->db_mtx);
1110 if (DBUF_IS_L2CACHEABLE(db))
1111 aflags |= ARC_FLAG_L2CACHE;
1113 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
1114 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
1115 db->db.db_object, db->db_level, db->db_blkid);
1118 * All bps of an encrypted os should have the encryption bit set.
1119 * If this is not true it indicates tampering and we report an error.
1121 if (db->db_objset->os_encrypted && !BP_USES_CRYPT(db->db_blkptr)) {
1122 spa_log_error(db->db_objset->os_spa, &zb);
1123 zfs_panic_recover("unencrypted block in encrypted "
1124 "object set %llu", dmu_objset_id(db->db_objset));
1125 return (SET_ERROR(EIO));
1128 dbuf_add_ref(db, NULL);
1130 zio_flags = (flags & DB_RF_CANFAIL) ?
1131 ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED;
1133 if ((flags & DB_RF_NO_DECRYPT) && BP_IS_PROTECTED(db->db_blkptr))
1134 zio_flags |= ZIO_FLAG_RAW;
1136 err = arc_read(zio, db->db_objset->os_spa, db->db_blkptr,
1137 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ, zio_flags,
1144 * This is our just-in-time copy function. It makes a copy of buffers that
1145 * have been modified in a previous transaction group before we access them in
1146 * the current active group.
1148 * This function is used in three places: when we are dirtying a buffer for the
1149 * first time in a txg, when we are freeing a range in a dnode that includes
1150 * this buffer, and when we are accessing a buffer which was received compressed
1151 * and later referenced in a WRITE_BYREF record.
1153 * Note that when we are called from dbuf_free_range() we do not put a hold on
1154 * the buffer, we just traverse the active dbuf list for the dnode.
1157 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
1159 dbuf_dirty_record_t *dr = db->db_last_dirty;
1161 ASSERT(MUTEX_HELD(&db->db_mtx));
1162 ASSERT(db->db.db_data != NULL);
1163 ASSERT(db->db_level == 0);
1164 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
1167 (dr->dt.dl.dr_data !=
1168 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
1172 * If the last dirty record for this dbuf has not yet synced
1173 * and its referencing the dbuf data, either:
1174 * reset the reference to point to a new copy,
1175 * or (if there a no active holders)
1176 * just null out the current db_data pointer.
1178 ASSERT3U(dr->dr_txg, >=, txg - 2);
1179 if (db->db_blkid == DMU_BONUS_BLKID) {
1180 dnode_t *dn = DB_DNODE(db);
1181 int bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
1182 dr->dt.dl.dr_data = kmem_alloc(bonuslen, KM_SLEEP);
1183 arc_space_consume(bonuslen, ARC_SPACE_BONUS);
1184 bcopy(db->db.db_data, dr->dt.dl.dr_data, bonuslen);
1185 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
1186 dnode_t *dn = DB_DNODE(db);
1187 int size = arc_buf_size(db->db_buf);
1188 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1189 spa_t *spa = db->db_objset->os_spa;
1190 enum zio_compress compress_type =
1191 arc_get_compression(db->db_buf);
1193 if (arc_is_encrypted(db->db_buf)) {
1194 boolean_t byteorder;
1195 uint8_t salt[ZIO_DATA_SALT_LEN];
1196 uint8_t iv[ZIO_DATA_IV_LEN];
1197 uint8_t mac[ZIO_DATA_MAC_LEN];
1199 arc_get_raw_params(db->db_buf, &byteorder, salt,
1201 dr->dt.dl.dr_data = arc_alloc_raw_buf(spa, db,
1202 dmu_objset_id(dn->dn_objset), byteorder, salt, iv,
1203 mac, dn->dn_type, size, arc_buf_lsize(db->db_buf),
1205 } else if (compress_type != ZIO_COMPRESS_OFF) {
1206 ASSERT3U(type, ==, ARC_BUFC_DATA);
1207 dr->dt.dl.dr_data = arc_alloc_compressed_buf(spa, db,
1208 size, arc_buf_lsize(db->db_buf), compress_type);
1210 dr->dt.dl.dr_data = arc_alloc_buf(spa, db, type, size);
1212 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
1215 dbuf_clear_data(db);
1220 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
1227 * We don't have to hold the mutex to check db_state because it
1228 * can't be freed while we have a hold on the buffer.
1230 ASSERT(!refcount_is_zero(&db->db_holds));
1232 if (db->db_state == DB_NOFILL)
1233 return (SET_ERROR(EIO));
1237 if ((flags & DB_RF_HAVESTRUCT) == 0)
1238 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1240 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1241 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
1242 DBUF_IS_CACHEABLE(db);
1244 mutex_enter(&db->db_mtx);
1245 if (db->db_state == DB_CACHED) {
1246 spa_t *spa = dn->dn_objset->os_spa;
1249 * If the arc buf is compressed or encrypted, we need to
1250 * untransform it to read the data. This could happen during
1251 * the "zfs receive" of a stream which is deduplicated and
1252 * either raw or compressed. We do not need to do this if the
1253 * caller wants raw encrypted data.
1255 if (db->db_buf != NULL && (flags & DB_RF_NO_DECRYPT) == 0 &&
1256 (arc_is_encrypted(db->db_buf) ||
1257 arc_get_compression(db->db_buf) != ZIO_COMPRESS_OFF)) {
1258 dbuf_fix_old_data(db, spa_syncing_txg(spa));
1259 err = arc_untransform(db->db_buf, spa,
1260 dmu_objset_id(db->db_objset), B_FALSE);
1261 dbuf_set_data(db, db->db_buf);
1263 mutex_exit(&db->db_mtx);
1265 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1266 if ((flags & DB_RF_HAVESTRUCT) == 0)
1267 rw_exit(&dn->dn_struct_rwlock);
1269 } else if (db->db_state == DB_UNCACHED) {
1270 spa_t *spa = dn->dn_objset->os_spa;
1271 boolean_t need_wait = B_FALSE;
1274 db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) {
1275 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
1278 err = dbuf_read_impl(db, zio, flags);
1280 /* dbuf_read_impl has dropped db_mtx for us */
1282 if (!err && prefetch)
1283 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1285 if ((flags & DB_RF_HAVESTRUCT) == 0)
1286 rw_exit(&dn->dn_struct_rwlock);
1289 if (!err && need_wait)
1290 err = zio_wait(zio);
1293 * Another reader came in while the dbuf was in flight
1294 * between UNCACHED and CACHED. Either a writer will finish
1295 * writing the buffer (sending the dbuf to CACHED) or the
1296 * first reader's request will reach the read_done callback
1297 * and send the dbuf to CACHED. Otherwise, a failure
1298 * occurred and the dbuf went to UNCACHED.
1300 mutex_exit(&db->db_mtx);
1302 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1303 if ((flags & DB_RF_HAVESTRUCT) == 0)
1304 rw_exit(&dn->dn_struct_rwlock);
1307 /* Skip the wait per the caller's request. */
1308 mutex_enter(&db->db_mtx);
1309 if ((flags & DB_RF_NEVERWAIT) == 0) {
1310 while (db->db_state == DB_READ ||
1311 db->db_state == DB_FILL) {
1312 ASSERT(db->db_state == DB_READ ||
1313 (flags & DB_RF_HAVESTRUCT) == 0);
1314 DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *,
1316 cv_wait(&db->db_changed, &db->db_mtx);
1318 if (db->db_state == DB_UNCACHED)
1319 err = SET_ERROR(EIO);
1321 mutex_exit(&db->db_mtx);
1328 dbuf_noread(dmu_buf_impl_t *db)
1330 ASSERT(!refcount_is_zero(&db->db_holds));
1331 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1332 mutex_enter(&db->db_mtx);
1333 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1334 cv_wait(&db->db_changed, &db->db_mtx);
1335 if (db->db_state == DB_UNCACHED) {
1336 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1337 spa_t *spa = db->db_objset->os_spa;
1339 ASSERT(db->db_buf == NULL);
1340 ASSERT(db->db.db_data == NULL);
1341 dbuf_set_data(db, arc_alloc_buf(spa, db, type, db->db.db_size));
1342 db->db_state = DB_FILL;
1343 } else if (db->db_state == DB_NOFILL) {
1344 dbuf_clear_data(db);
1346 ASSERT3U(db->db_state, ==, DB_CACHED);
1348 mutex_exit(&db->db_mtx);
1352 dbuf_unoverride(dbuf_dirty_record_t *dr)
1354 dmu_buf_impl_t *db = dr->dr_dbuf;
1355 blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
1356 uint64_t txg = dr->dr_txg;
1358 ASSERT(MUTEX_HELD(&db->db_mtx));
1360 * This assert is valid because dmu_sync() expects to be called by
1361 * a zilog's get_data while holding a range lock. This call only
1362 * comes from dbuf_dirty() callers who must also hold a range lock.
1364 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
1365 ASSERT(db->db_level == 0);
1367 if (db->db_blkid == DMU_BONUS_BLKID ||
1368 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
1371 ASSERT(db->db_data_pending != dr);
1373 /* free this block */
1374 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite)
1375 zio_free(db->db_objset->os_spa, txg, bp);
1377 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1378 dr->dt.dl.dr_nopwrite = B_FALSE;
1379 dr->dt.dl.dr_raw = B_FALSE;
1382 * Release the already-written buffer, so we leave it in
1383 * a consistent dirty state. Note that all callers are
1384 * modifying the buffer, so they will immediately do
1385 * another (redundant) arc_release(). Therefore, leave
1386 * the buf thawed to save the effort of freezing &
1387 * immediately re-thawing it.
1389 arc_release(dr->dt.dl.dr_data, db);
1393 * Evict (if its unreferenced) or clear (if its referenced) any level-0
1394 * data blocks in the free range, so that any future readers will find
1398 dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
1401 dmu_buf_impl_t *db_search;
1402 dmu_buf_impl_t *db, *db_next;
1403 uint64_t txg = tx->tx_txg;
1406 if (end_blkid > dn->dn_maxblkid &&
1407 !(start_blkid == DMU_SPILL_BLKID || end_blkid == DMU_SPILL_BLKID))
1408 end_blkid = dn->dn_maxblkid;
1409 dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid);
1411 db_search = kmem_alloc(sizeof (dmu_buf_impl_t), KM_SLEEP);
1412 db_search->db_level = 0;
1413 db_search->db_blkid = start_blkid;
1414 db_search->db_state = DB_SEARCH;
1416 mutex_enter(&dn->dn_dbufs_mtx);
1417 db = avl_find(&dn->dn_dbufs, db_search, &where);
1418 ASSERT3P(db, ==, NULL);
1420 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1422 for (; db != NULL; db = db_next) {
1423 db_next = AVL_NEXT(&dn->dn_dbufs, db);
1424 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1426 if (db->db_level != 0 || db->db_blkid > end_blkid) {
1429 ASSERT3U(db->db_blkid, >=, start_blkid);
1431 /* found a level 0 buffer in the range */
1432 mutex_enter(&db->db_mtx);
1433 if (dbuf_undirty(db, tx)) {
1434 /* mutex has been dropped and dbuf destroyed */
1438 if (db->db_state == DB_UNCACHED ||
1439 db->db_state == DB_NOFILL ||
1440 db->db_state == DB_EVICTING) {
1441 ASSERT(db->db.db_data == NULL);
1442 mutex_exit(&db->db_mtx);
1445 if (db->db_state == DB_READ || db->db_state == DB_FILL) {
1446 /* will be handled in dbuf_read_done or dbuf_rele */
1447 db->db_freed_in_flight = TRUE;
1448 mutex_exit(&db->db_mtx);
1451 if (refcount_count(&db->db_holds) == 0) {
1456 /* The dbuf is referenced */
1458 if (db->db_last_dirty != NULL) {
1459 dbuf_dirty_record_t *dr = db->db_last_dirty;
1461 if (dr->dr_txg == txg) {
1463 * This buffer is "in-use", re-adjust the file
1464 * size to reflect that this buffer may
1465 * contain new data when we sync.
1467 if (db->db_blkid != DMU_SPILL_BLKID &&
1468 db->db_blkid > dn->dn_maxblkid)
1469 dn->dn_maxblkid = db->db_blkid;
1470 dbuf_unoverride(dr);
1473 * This dbuf is not dirty in the open context.
1474 * Either uncache it (if its not referenced in
1475 * the open context) or reset its contents to
1478 dbuf_fix_old_data(db, txg);
1481 /* clear the contents if its cached */
1482 if (db->db_state == DB_CACHED) {
1483 ASSERT(db->db.db_data != NULL);
1484 arc_release(db->db_buf, db);
1485 bzero(db->db.db_data, db->db.db_size);
1486 arc_buf_freeze(db->db_buf);
1489 mutex_exit(&db->db_mtx);
1492 kmem_free(db_search, sizeof (dmu_buf_impl_t));
1493 mutex_exit(&dn->dn_dbufs_mtx);
1497 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
1499 arc_buf_t *buf, *obuf;
1500 int osize = db->db.db_size;
1501 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1504 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1509 /* XXX does *this* func really need the lock? */
1510 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1513 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1514 * is OK, because there can be no other references to the db
1515 * when we are changing its size, so no concurrent DB_FILL can
1519 * XXX we should be doing a dbuf_read, checking the return
1520 * value and returning that up to our callers
1522 dmu_buf_will_dirty(&db->db, tx);
1524 /* create the data buffer for the new block */
1525 buf = arc_alloc_buf(dn->dn_objset->os_spa, db, type, size);
1527 /* copy old block data to the new block */
1529 bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
1530 /* zero the remainder */
1532 bzero((uint8_t *)buf->b_data + osize, size - osize);
1534 mutex_enter(&db->db_mtx);
1535 dbuf_set_data(db, buf);
1536 arc_buf_destroy(obuf, db);
1537 db->db.db_size = size;
1539 if (db->db_level == 0) {
1540 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1541 db->db_last_dirty->dt.dl.dr_data = buf;
1543 mutex_exit(&db->db_mtx);
1545 dmu_objset_willuse_space(dn->dn_objset, size - osize, tx);
1550 dbuf_release_bp(dmu_buf_impl_t *db)
1552 ASSERTV(objset_t *os = db->db_objset);
1554 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1555 ASSERT(arc_released(os->os_phys_buf) ||
1556 list_link_active(&os->os_dsl_dataset->ds_synced_link));
1557 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1559 (void) arc_release(db->db_buf, db);
1563 * We already have a dirty record for this TXG, and we are being
1567 dbuf_redirty(dbuf_dirty_record_t *dr)
1569 dmu_buf_impl_t *db = dr->dr_dbuf;
1571 ASSERT(MUTEX_HELD(&db->db_mtx));
1573 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1575 * If this buffer has already been written out,
1576 * we now need to reset its state.
1578 dbuf_unoverride(dr);
1579 if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1580 db->db_state != DB_NOFILL) {
1581 /* Already released on initial dirty, so just thaw. */
1582 ASSERT(arc_released(db->db_buf));
1583 arc_buf_thaw(db->db_buf);
1588 dbuf_dirty_record_t *
1589 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1593 dbuf_dirty_record_t **drp, *dr;
1594 int drop_struct_lock = FALSE;
1595 int txgoff = tx->tx_txg & TXG_MASK;
1597 ASSERT(tx->tx_txg != 0);
1598 ASSERT(!refcount_is_zero(&db->db_holds));
1599 DMU_TX_DIRTY_BUF(tx, db);
1604 * Shouldn't dirty a regular buffer in syncing context. Private
1605 * objects may be dirtied in syncing context, but only if they
1606 * were already pre-dirtied in open context.
1609 if (dn->dn_objset->os_dsl_dataset != NULL) {
1610 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1613 ASSERT(!dmu_tx_is_syncing(tx) ||
1614 BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1615 DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1616 dn->dn_objset->os_dsl_dataset == NULL);
1617 if (dn->dn_objset->os_dsl_dataset != NULL)
1618 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, FTAG);
1621 * We make this assert for private objects as well, but after we
1622 * check if we're already dirty. They are allowed to re-dirty
1623 * in syncing context.
1625 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1626 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1627 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1629 mutex_enter(&db->db_mtx);
1631 * XXX make this true for indirects too? The problem is that
1632 * transactions created with dmu_tx_create_assigned() from
1633 * syncing context don't bother holding ahead.
1635 ASSERT(db->db_level != 0 ||
1636 db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1637 db->db_state == DB_NOFILL);
1639 mutex_enter(&dn->dn_mtx);
1641 * Don't set dirtyctx to SYNC if we're just modifying this as we
1642 * initialize the objset.
1644 if (dn->dn_dirtyctx == DN_UNDIRTIED) {
1645 if (dn->dn_objset->os_dsl_dataset != NULL) {
1646 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1649 if (!BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1650 dn->dn_dirtyctx = (dmu_tx_is_syncing(tx) ?
1651 DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1652 ASSERT(dn->dn_dirtyctx_firstset == NULL);
1653 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1655 if (dn->dn_objset->os_dsl_dataset != NULL) {
1656 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1660 mutex_exit(&dn->dn_mtx);
1662 if (db->db_blkid == DMU_SPILL_BLKID)
1663 dn->dn_have_spill = B_TRUE;
1666 * If this buffer is already dirty, we're done.
1668 drp = &db->db_last_dirty;
1669 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1670 db->db.db_object == DMU_META_DNODE_OBJECT);
1671 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1673 if (dr && dr->dr_txg == tx->tx_txg) {
1677 mutex_exit(&db->db_mtx);
1682 * Only valid if not already dirty.
1684 ASSERT(dn->dn_object == 0 ||
1685 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1686 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1688 ASSERT3U(dn->dn_nlevels, >, db->db_level);
1691 * We should only be dirtying in syncing context if it's the
1692 * mos or we're initializing the os or it's a special object.
1693 * However, we are allowed to dirty in syncing context provided
1694 * we already dirtied it in open context. Hence we must make
1695 * this assertion only if we're not already dirty.
1698 VERIFY3U(tx->tx_txg, <=, spa_final_dirty_txg(os->os_spa));
1700 if (dn->dn_objset->os_dsl_dataset != NULL)
1701 rrw_enter(&os->os_dsl_dataset->ds_bp_rwlock, RW_READER, FTAG);
1702 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1703 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1704 if (dn->dn_objset->os_dsl_dataset != NULL)
1705 rrw_exit(&os->os_dsl_dataset->ds_bp_rwlock, FTAG);
1707 ASSERT(db->db.db_size != 0);
1709 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1711 if (db->db_blkid != DMU_BONUS_BLKID) {
1712 dmu_objset_willuse_space(os, db->db.db_size, tx);
1716 * If this buffer is dirty in an old transaction group we need
1717 * to make a copy of it so that the changes we make in this
1718 * transaction group won't leak out when we sync the older txg.
1720 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1721 list_link_init(&dr->dr_dirty_node);
1722 if (db->db_level == 0) {
1723 void *data_old = db->db_buf;
1725 if (db->db_state != DB_NOFILL) {
1726 if (db->db_blkid == DMU_BONUS_BLKID) {
1727 dbuf_fix_old_data(db, tx->tx_txg);
1728 data_old = db->db.db_data;
1729 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1731 * Release the data buffer from the cache so
1732 * that we can modify it without impacting
1733 * possible other users of this cached data
1734 * block. Note that indirect blocks and
1735 * private objects are not released until the
1736 * syncing state (since they are only modified
1739 arc_release(db->db_buf, db);
1740 dbuf_fix_old_data(db, tx->tx_txg);
1741 data_old = db->db_buf;
1743 ASSERT(data_old != NULL);
1745 dr->dt.dl.dr_data = data_old;
1747 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_NOLOCKDEP, NULL);
1748 list_create(&dr->dt.di.dr_children,
1749 sizeof (dbuf_dirty_record_t),
1750 offsetof(dbuf_dirty_record_t, dr_dirty_node));
1752 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL)
1753 dr->dr_accounted = db->db.db_size;
1755 dr->dr_txg = tx->tx_txg;
1760 * We could have been freed_in_flight between the dbuf_noread
1761 * and dbuf_dirty. We win, as though the dbuf_noread() had
1762 * happened after the free.
1764 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1765 db->db_blkid != DMU_SPILL_BLKID) {
1766 mutex_enter(&dn->dn_mtx);
1767 if (dn->dn_free_ranges[txgoff] != NULL) {
1768 range_tree_clear(dn->dn_free_ranges[txgoff],
1771 mutex_exit(&dn->dn_mtx);
1772 db->db_freed_in_flight = FALSE;
1776 * This buffer is now part of this txg
1778 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1779 db->db_dirtycnt += 1;
1780 ASSERT3U(db->db_dirtycnt, <=, 3);
1782 mutex_exit(&db->db_mtx);
1784 if (db->db_blkid == DMU_BONUS_BLKID ||
1785 db->db_blkid == DMU_SPILL_BLKID) {
1786 mutex_enter(&dn->dn_mtx);
1787 ASSERT(!list_link_active(&dr->dr_dirty_node));
1788 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1789 mutex_exit(&dn->dn_mtx);
1790 dnode_setdirty(dn, tx);
1796 * The dn_struct_rwlock prevents db_blkptr from changing
1797 * due to a write from syncing context completing
1798 * while we are running, so we want to acquire it before
1799 * looking at db_blkptr.
1801 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1802 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1803 drop_struct_lock = TRUE;
1807 * We need to hold the dn_struct_rwlock to make this assertion,
1808 * because it protects dn_phys / dn_next_nlevels from changing.
1810 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1811 dn->dn_phys->dn_nlevels > db->db_level ||
1812 dn->dn_next_nlevels[txgoff] > db->db_level ||
1813 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1814 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1817 * If we are overwriting a dedup BP, then unless it is snapshotted,
1818 * when we get to syncing context we will need to decrement its
1819 * refcount in the DDT. Prefetch the relevant DDT block so that
1820 * syncing context won't have to wait for the i/o.
1822 ddt_prefetch(os->os_spa, db->db_blkptr);
1824 if (db->db_level == 0) {
1825 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1826 ASSERT(dn->dn_maxblkid >= db->db_blkid);
1829 if (db->db_level+1 < dn->dn_nlevels) {
1830 dmu_buf_impl_t *parent = db->db_parent;
1831 dbuf_dirty_record_t *di;
1832 int parent_held = FALSE;
1834 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1835 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1837 parent = dbuf_hold_level(dn, db->db_level+1,
1838 db->db_blkid >> epbs, FTAG);
1839 ASSERT(parent != NULL);
1842 if (drop_struct_lock)
1843 rw_exit(&dn->dn_struct_rwlock);
1844 ASSERT3U(db->db_level+1, ==, parent->db_level);
1845 di = dbuf_dirty(parent, tx);
1847 dbuf_rele(parent, FTAG);
1849 mutex_enter(&db->db_mtx);
1851 * Since we've dropped the mutex, it's possible that
1852 * dbuf_undirty() might have changed this out from under us.
1854 if (db->db_last_dirty == dr ||
1855 dn->dn_object == DMU_META_DNODE_OBJECT) {
1856 mutex_enter(&di->dt.di.dr_mtx);
1857 ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1858 ASSERT(!list_link_active(&dr->dr_dirty_node));
1859 list_insert_tail(&di->dt.di.dr_children, dr);
1860 mutex_exit(&di->dt.di.dr_mtx);
1863 mutex_exit(&db->db_mtx);
1865 ASSERT(db->db_level+1 == dn->dn_nlevels);
1866 ASSERT(db->db_blkid < dn->dn_nblkptr);
1867 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1868 mutex_enter(&dn->dn_mtx);
1869 ASSERT(!list_link_active(&dr->dr_dirty_node));
1870 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1871 mutex_exit(&dn->dn_mtx);
1872 if (drop_struct_lock)
1873 rw_exit(&dn->dn_struct_rwlock);
1876 dnode_setdirty(dn, tx);
1882 * Undirty a buffer in the transaction group referenced by the given
1883 * transaction. Return whether this evicted the dbuf.
1886 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1889 uint64_t txg = tx->tx_txg;
1890 dbuf_dirty_record_t *dr, **drp;
1895 * Due to our use of dn_nlevels below, this can only be called
1896 * in open context, unless we are operating on the MOS.
1897 * From syncing context, dn_nlevels may be different from the
1898 * dn_nlevels used when dbuf was dirtied.
1900 ASSERT(db->db_objset ==
1901 dmu_objset_pool(db->db_objset)->dp_meta_objset ||
1902 txg != spa_syncing_txg(dmu_objset_spa(db->db_objset)));
1903 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1904 ASSERT0(db->db_level);
1905 ASSERT(MUTEX_HELD(&db->db_mtx));
1908 * If this buffer is not dirty, we're done.
1910 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1911 if (dr->dr_txg <= txg)
1913 if (dr == NULL || dr->dr_txg < txg)
1915 ASSERT(dr->dr_txg == txg);
1916 ASSERT(dr->dr_dbuf == db);
1921 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1923 ASSERT(db->db.db_size != 0);
1925 dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset),
1926 dr->dr_accounted, txg);
1931 * Note that there are three places in dbuf_dirty()
1932 * where this dirty record may be put on a list.
1933 * Make sure to do a list_remove corresponding to
1934 * every one of those list_insert calls.
1936 if (dr->dr_parent) {
1937 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1938 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1939 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1940 } else if (db->db_blkid == DMU_SPILL_BLKID ||
1941 db->db_level + 1 == dn->dn_nlevels) {
1942 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1943 mutex_enter(&dn->dn_mtx);
1944 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1945 mutex_exit(&dn->dn_mtx);
1949 if (db->db_state != DB_NOFILL) {
1950 dbuf_unoverride(dr);
1952 ASSERT(db->db_buf != NULL);
1953 ASSERT(dr->dt.dl.dr_data != NULL);
1954 if (dr->dt.dl.dr_data != db->db_buf)
1955 arc_buf_destroy(dr->dt.dl.dr_data, db);
1958 kmem_free(dr, sizeof (dbuf_dirty_record_t));
1960 ASSERT(db->db_dirtycnt > 0);
1961 db->db_dirtycnt -= 1;
1963 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1964 ASSERT(db->db_state == DB_NOFILL || arc_released(db->db_buf));
1973 dmu_buf_will_dirty_impl(dmu_buf_t *db_fake, int flags, dmu_tx_t *tx)
1975 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1976 dbuf_dirty_record_t *dr;
1978 ASSERT(tx->tx_txg != 0);
1979 ASSERT(!refcount_is_zero(&db->db_holds));
1982 * Quick check for dirtyness. For already dirty blocks, this
1983 * reduces runtime of this function by >90%, and overall performance
1984 * by 50% for some workloads (e.g. file deletion with indirect blocks
1987 mutex_enter(&db->db_mtx);
1989 for (dr = db->db_last_dirty;
1990 dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) {
1992 * It's possible that it is already dirty but not cached,
1993 * because there are some calls to dbuf_dirty() that don't
1994 * go through dmu_buf_will_dirty().
1996 if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) {
1997 /* This dbuf is already dirty and cached. */
1999 mutex_exit(&db->db_mtx);
2003 mutex_exit(&db->db_mtx);
2006 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
2007 flags |= DB_RF_HAVESTRUCT;
2009 (void) dbuf_read(db, NULL, flags);
2010 (void) dbuf_dirty(db, tx);
2014 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
2016 dmu_buf_will_dirty_impl(db_fake,
2017 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH, tx);
2021 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
2023 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2025 db->db_state = DB_NOFILL;
2027 dmu_buf_will_fill(db_fake, tx);
2031 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
2033 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2035 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2036 ASSERT(tx->tx_txg != 0);
2037 ASSERT(db->db_level == 0);
2038 ASSERT(!refcount_is_zero(&db->db_holds));
2040 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
2041 dmu_tx_private_ok(tx));
2044 (void) dbuf_dirty(db, tx);
2048 * This function is effectively the same as dmu_buf_will_dirty(), but
2049 * indicates the caller expects raw encrypted data in the db. It will
2050 * also set the raw flag on the created dirty record.
2053 dmu_buf_will_change_crypt_params(dmu_buf_t *db_fake, dmu_tx_t *tx)
2055 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2056 dbuf_dirty_record_t *dr;
2058 dmu_buf_will_dirty_impl(db_fake,
2059 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_NO_DECRYPT, tx);
2061 dr = db->db_last_dirty;
2062 while (dr != NULL && dr->dr_txg > tx->tx_txg)
2065 ASSERT3P(dr, !=, NULL);
2066 ASSERT3U(dr->dr_txg, ==, tx->tx_txg);
2067 dr->dt.dl.dr_raw = B_TRUE;
2070 #pragma weak dmu_buf_fill_done = dbuf_fill_done
2073 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
2075 mutex_enter(&db->db_mtx);
2078 if (db->db_state == DB_FILL) {
2079 if (db->db_level == 0 && db->db_freed_in_flight) {
2080 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2081 /* we were freed while filling */
2082 /* XXX dbuf_undirty? */
2083 bzero(db->db.db_data, db->db.db_size);
2084 db->db_freed_in_flight = FALSE;
2086 db->db_state = DB_CACHED;
2087 cv_broadcast(&db->db_changed);
2089 mutex_exit(&db->db_mtx);
2093 dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data,
2094 bp_embedded_type_t etype, enum zio_compress comp,
2095 int uncompressed_size, int compressed_size, int byteorder,
2098 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2099 struct dirty_leaf *dl;
2100 dmu_object_type_t type;
2102 if (etype == BP_EMBEDDED_TYPE_DATA) {
2103 ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset),
2104 SPA_FEATURE_EMBEDDED_DATA));
2108 type = DB_DNODE(db)->dn_type;
2111 ASSERT0(db->db_level);
2112 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2114 dmu_buf_will_not_fill(dbuf, tx);
2116 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
2117 dl = &db->db_last_dirty->dt.dl;
2118 encode_embedded_bp_compressed(&dl->dr_overridden_by,
2119 data, comp, uncompressed_size, compressed_size);
2120 BPE_SET_ETYPE(&dl->dr_overridden_by, etype);
2121 BP_SET_TYPE(&dl->dr_overridden_by, type);
2122 BP_SET_LEVEL(&dl->dr_overridden_by, 0);
2123 BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder);
2125 dl->dr_override_state = DR_OVERRIDDEN;
2126 dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg;
2130 * Directly assign a provided arc buf to a given dbuf if it's not referenced
2131 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
2134 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
2136 ASSERT(!refcount_is_zero(&db->db_holds));
2137 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2138 ASSERT(db->db_level == 0);
2139 ASSERT3U(dbuf_is_metadata(db), ==, arc_is_metadata(buf));
2140 ASSERT(buf != NULL);
2141 ASSERT(arc_buf_lsize(buf) == db->db.db_size);
2142 ASSERT(tx->tx_txg != 0);
2144 arc_return_buf(buf, db);
2145 ASSERT(arc_released(buf));
2147 mutex_enter(&db->db_mtx);
2149 while (db->db_state == DB_READ || db->db_state == DB_FILL)
2150 cv_wait(&db->db_changed, &db->db_mtx);
2152 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
2154 if (db->db_state == DB_CACHED &&
2155 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
2157 * In practice, we will never have a case where we have an
2158 * encrypted arc buffer while additional holds exist on the
2159 * dbuf. We don't handle this here so we simply assert that
2162 ASSERT(!arc_is_encrypted(buf));
2163 mutex_exit(&db->db_mtx);
2164 (void) dbuf_dirty(db, tx);
2165 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
2166 arc_buf_destroy(buf, db);
2167 xuio_stat_wbuf_copied();
2171 xuio_stat_wbuf_nocopy();
2172 if (db->db_state == DB_CACHED) {
2173 dbuf_dirty_record_t *dr = db->db_last_dirty;
2175 ASSERT(db->db_buf != NULL);
2176 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
2177 ASSERT(dr->dt.dl.dr_data == db->db_buf);
2178 IMPLY(arc_is_encrypted(buf), dr->dt.dl.dr_raw);
2180 if (!arc_released(db->db_buf)) {
2181 ASSERT(dr->dt.dl.dr_override_state ==
2183 arc_release(db->db_buf, db);
2185 dr->dt.dl.dr_data = buf;
2186 arc_buf_destroy(db->db_buf, db);
2187 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
2188 arc_release(db->db_buf, db);
2189 arc_buf_destroy(db->db_buf, db);
2193 ASSERT(db->db_buf == NULL);
2194 dbuf_set_data(db, buf);
2195 db->db_state = DB_FILL;
2196 mutex_exit(&db->db_mtx);
2197 (void) dbuf_dirty(db, tx);
2198 dmu_buf_fill_done(&db->db, tx);
2202 dbuf_destroy(dmu_buf_impl_t *db)
2205 dmu_buf_impl_t *parent = db->db_parent;
2206 dmu_buf_impl_t *dndb;
2208 ASSERT(MUTEX_HELD(&db->db_mtx));
2209 ASSERT(refcount_is_zero(&db->db_holds));
2211 if (db->db_buf != NULL) {
2212 arc_buf_destroy(db->db_buf, db);
2216 if (db->db_blkid == DMU_BONUS_BLKID) {
2217 int slots = DB_DNODE(db)->dn_num_slots;
2218 int bonuslen = DN_SLOTS_TO_BONUSLEN(slots);
2219 if (db->db.db_data != NULL) {
2220 kmem_free(db->db.db_data, bonuslen);
2221 arc_space_return(bonuslen, ARC_SPACE_BONUS);
2222 db->db_state = DB_UNCACHED;
2226 dbuf_clear_data(db);
2228 if (multilist_link_active(&db->db_cache_link)) {
2229 multilist_remove(dbuf_cache, db);
2230 (void) refcount_remove_many(&dbuf_cache_size,
2231 db->db.db_size, db);
2234 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
2235 ASSERT(db->db_data_pending == NULL);
2237 db->db_state = DB_EVICTING;
2238 db->db_blkptr = NULL;
2241 * Now that db_state is DB_EVICTING, nobody else can find this via
2242 * the hash table. We can now drop db_mtx, which allows us to
2243 * acquire the dn_dbufs_mtx.
2245 mutex_exit(&db->db_mtx);
2250 if (db->db_blkid != DMU_BONUS_BLKID) {
2251 boolean_t needlock = !MUTEX_HELD(&dn->dn_dbufs_mtx);
2253 mutex_enter(&dn->dn_dbufs_mtx);
2254 avl_remove(&dn->dn_dbufs, db);
2255 atomic_dec_32(&dn->dn_dbufs_count);
2259 mutex_exit(&dn->dn_dbufs_mtx);
2261 * Decrementing the dbuf count means that the hold corresponding
2262 * to the removed dbuf is no longer discounted in dnode_move(),
2263 * so the dnode cannot be moved until after we release the hold.
2264 * The membar_producer() ensures visibility of the decremented
2265 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
2269 db->db_dnode_handle = NULL;
2271 dbuf_hash_remove(db);
2276 ASSERT(refcount_is_zero(&db->db_holds));
2278 db->db_parent = NULL;
2280 ASSERT(db->db_buf == NULL);
2281 ASSERT(db->db.db_data == NULL);
2282 ASSERT(db->db_hash_next == NULL);
2283 ASSERT(db->db_blkptr == NULL);
2284 ASSERT(db->db_data_pending == NULL);
2285 ASSERT(!multilist_link_active(&db->db_cache_link));
2287 kmem_cache_free(dbuf_kmem_cache, db);
2288 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
2291 * If this dbuf is referenced from an indirect dbuf,
2292 * decrement the ref count on the indirect dbuf.
2294 if (parent && parent != dndb)
2295 dbuf_rele(parent, db);
2299 * Note: While bpp will always be updated if the function returns success,
2300 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
2301 * this happens when the dnode is the meta-dnode, or a userused or groupused
2304 __attribute__((always_inline))
2306 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
2307 dmu_buf_impl_t **parentp, blkptr_t **bpp, struct dbuf_hold_impl_data *dh)
2314 ASSERT(blkid != DMU_BONUS_BLKID);
2316 if (blkid == DMU_SPILL_BLKID) {
2317 mutex_enter(&dn->dn_mtx);
2318 if (dn->dn_have_spill &&
2319 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
2320 *bpp = DN_SPILL_BLKPTR(dn->dn_phys);
2323 dbuf_add_ref(dn->dn_dbuf, NULL);
2324 *parentp = dn->dn_dbuf;
2325 mutex_exit(&dn->dn_mtx);
2330 (dn->dn_phys->dn_nlevels == 0) ? 1 : dn->dn_phys->dn_nlevels;
2331 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2333 ASSERT3U(level * epbs, <, 64);
2334 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2336 * This assertion shouldn't trip as long as the max indirect block size
2337 * is less than 1M. The reason for this is that up to that point,
2338 * the number of levels required to address an entire object with blocks
2339 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In
2340 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55
2341 * (i.e. we can address the entire object), objects will all use at most
2342 * N-1 levels and the assertion won't overflow. However, once epbs is
2343 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be
2344 * enough to address an entire object, so objects will have 5 levels,
2345 * but then this assertion will overflow.
2347 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we
2348 * need to redo this logic to handle overflows.
2350 ASSERT(level >= nlevels ||
2351 ((nlevels - level - 1) * epbs) +
2352 highbit64(dn->dn_phys->dn_nblkptr) <= 64);
2353 if (level >= nlevels ||
2354 blkid >= ((uint64_t)dn->dn_phys->dn_nblkptr <<
2355 ((nlevels - level - 1) * epbs)) ||
2357 blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
2358 /* the buffer has no parent yet */
2359 return (SET_ERROR(ENOENT));
2360 } else if (level < nlevels-1) {
2361 /* this block is referenced from an indirect block */
2364 err = dbuf_hold_impl(dn, level+1,
2365 blkid >> epbs, fail_sparse, FALSE, NULL, parentp);
2367 __dbuf_hold_impl_init(dh + 1, dn, dh->dh_level + 1,
2368 blkid >> epbs, fail_sparse, FALSE, NULL,
2369 parentp, dh->dh_depth + 1);
2370 err = __dbuf_hold_impl(dh + 1);
2374 err = dbuf_read(*parentp, NULL,
2375 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
2377 dbuf_rele(*parentp, NULL);
2381 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
2382 (blkid & ((1ULL << epbs) - 1));
2383 if (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))
2384 ASSERT(BP_IS_HOLE(*bpp));
2387 /* the block is referenced from the dnode */
2388 ASSERT3U(level, ==, nlevels-1);
2389 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
2390 blkid < dn->dn_phys->dn_nblkptr);
2392 dbuf_add_ref(dn->dn_dbuf, NULL);
2393 *parentp = dn->dn_dbuf;
2395 *bpp = &dn->dn_phys->dn_blkptr[blkid];
2400 static dmu_buf_impl_t *
2401 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
2402 dmu_buf_impl_t *parent, blkptr_t *blkptr)
2404 objset_t *os = dn->dn_objset;
2405 dmu_buf_impl_t *db, *odb;
2407 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2408 ASSERT(dn->dn_type != DMU_OT_NONE);
2410 db = kmem_cache_alloc(dbuf_kmem_cache, KM_SLEEP);
2413 db->db.db_object = dn->dn_object;
2414 db->db_level = level;
2415 db->db_blkid = blkid;
2416 db->db_last_dirty = NULL;
2417 db->db_dirtycnt = 0;
2418 db->db_dnode_handle = dn->dn_handle;
2419 db->db_parent = parent;
2420 db->db_blkptr = blkptr;
2423 db->db_user_immediate_evict = FALSE;
2424 db->db_freed_in_flight = FALSE;
2425 db->db_pending_evict = FALSE;
2427 if (blkid == DMU_BONUS_BLKID) {
2428 ASSERT3P(parent, ==, dn->dn_dbuf);
2429 db->db.db_size = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
2430 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
2431 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
2432 db->db.db_offset = DMU_BONUS_BLKID;
2433 db->db_state = DB_UNCACHED;
2434 /* the bonus dbuf is not placed in the hash table */
2435 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
2437 } else if (blkid == DMU_SPILL_BLKID) {
2438 db->db.db_size = (blkptr != NULL) ?
2439 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
2440 db->db.db_offset = 0;
2443 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
2444 db->db.db_size = blocksize;
2445 db->db.db_offset = db->db_blkid * blocksize;
2449 * Hold the dn_dbufs_mtx while we get the new dbuf
2450 * in the hash table *and* added to the dbufs list.
2451 * This prevents a possible deadlock with someone
2452 * trying to look up this dbuf before its added to the
2455 mutex_enter(&dn->dn_dbufs_mtx);
2456 db->db_state = DB_EVICTING;
2457 if ((odb = dbuf_hash_insert(db)) != NULL) {
2458 /* someone else inserted it first */
2459 kmem_cache_free(dbuf_kmem_cache, db);
2460 mutex_exit(&dn->dn_dbufs_mtx);
2463 avl_add(&dn->dn_dbufs, db);
2465 db->db_state = DB_UNCACHED;
2466 mutex_exit(&dn->dn_dbufs_mtx);
2467 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
2469 if (parent && parent != dn->dn_dbuf)
2470 dbuf_add_ref(parent, db);
2472 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
2473 refcount_count(&dn->dn_holds) > 0);
2474 (void) refcount_add(&dn->dn_holds, db);
2475 atomic_inc_32(&dn->dn_dbufs_count);
2477 dprintf_dbuf(db, "db=%p\n", db);
2482 typedef struct dbuf_prefetch_arg {
2483 spa_t *dpa_spa; /* The spa to issue the prefetch in. */
2484 zbookmark_phys_t dpa_zb; /* The target block to prefetch. */
2485 int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */
2486 int dpa_curlevel; /* The current level that we're reading */
2487 dnode_t *dpa_dnode; /* The dnode associated with the prefetch */
2488 zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */
2489 zio_t *dpa_zio; /* The parent zio_t for all prefetches. */
2490 arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */
2491 } dbuf_prefetch_arg_t;
2494 * Actually issue the prefetch read for the block given.
2497 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp)
2500 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
2503 aflags = dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
2505 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2506 ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level);
2507 ASSERT(dpa->dpa_zio != NULL);
2508 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL,
2509 dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2510 &aflags, &dpa->dpa_zb);
2514 * Called when an indirect block above our prefetch target is read in. This
2515 * will either read in the next indirect block down the tree or issue the actual
2516 * prefetch if the next block down is our target.
2519 dbuf_prefetch_indirect_done(zio_t *zio, int err, arc_buf_t *abuf, void *private)
2521 dbuf_prefetch_arg_t *dpa = private;
2525 ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel);
2526 ASSERT3S(dpa->dpa_curlevel, >, 0);
2529 * The dpa_dnode is only valid if we are called with a NULL
2530 * zio. This indicates that the arc_read() returned without
2531 * first calling zio_read() to issue a physical read. Once
2532 * a physical read is made the dpa_dnode must be invalidated
2533 * as the locks guarding it may have been dropped. If the
2534 * dpa_dnode is still valid, then we want to add it to the dbuf
2535 * cache. To do so, we must hold the dbuf associated with the block
2536 * we just prefetched, read its contents so that we associate it
2537 * with an arc_buf_t, and then release it.
2540 ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel);
2541 if (zio->io_flags & ZIO_FLAG_RAW_COMPRESS) {
2542 ASSERT3U(BP_GET_PSIZE(zio->io_bp), ==, zio->io_size);
2544 ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size);
2546 ASSERT3P(zio->io_spa, ==, dpa->dpa_spa);
2548 dpa->dpa_dnode = NULL;
2549 } else if (dpa->dpa_dnode != NULL) {
2550 uint64_t curblkid = dpa->dpa_zb.zb_blkid >>
2551 (dpa->dpa_epbs * (dpa->dpa_curlevel -
2552 dpa->dpa_zb.zb_level));
2553 dmu_buf_impl_t *db = dbuf_hold_level(dpa->dpa_dnode,
2554 dpa->dpa_curlevel, curblkid, FTAG);
2555 (void) dbuf_read(db, NULL,
2556 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_HAVESTRUCT);
2557 dbuf_rele(db, FTAG);
2560 dpa->dpa_curlevel--;
2562 nextblkid = dpa->dpa_zb.zb_blkid >>
2563 (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level));
2564 bp = ((blkptr_t *)abuf->b_data) +
2565 P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs);
2566 if (BP_IS_HOLE(bp) || err != 0) {
2567 kmem_free(dpa, sizeof (*dpa));
2568 } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) {
2569 ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid);
2570 dbuf_issue_final_prefetch(dpa, bp);
2571 kmem_free(dpa, sizeof (*dpa));
2573 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2574 zbookmark_phys_t zb;
2576 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2578 SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset,
2579 dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid);
2581 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2582 bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio,
2583 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2587 arc_buf_destroy(abuf, private);
2591 * Issue prefetch reads for the given block on the given level. If the indirect
2592 * blocks above that block are not in memory, we will read them in
2593 * asynchronously. As a result, this call never blocks waiting for a read to
2594 * complete. Note that the prefetch might fail if the dataset is encrypted and
2595 * the encryption key is unmapped before the IO completes.
2598 dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio,
2602 int epbs, nlevels, curlevel;
2606 dbuf_prefetch_arg_t *dpa;
2609 ASSERT(blkid != DMU_BONUS_BLKID);
2610 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2612 if (blkid > dn->dn_maxblkid)
2615 if (dnode_block_freed(dn, blkid))
2619 * This dnode hasn't been written to disk yet, so there's nothing to
2622 nlevels = dn->dn_phys->dn_nlevels;
2623 if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0)
2626 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2627 if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level))
2630 db = dbuf_find(dn->dn_objset, dn->dn_object,
2633 mutex_exit(&db->db_mtx);
2635 * This dbuf already exists. It is either CACHED, or
2636 * (we assume) about to be read or filled.
2642 * Find the closest ancestor (indirect block) of the target block
2643 * that is present in the cache. In this indirect block, we will
2644 * find the bp that is at curlevel, curblkid.
2648 while (curlevel < nlevels - 1) {
2649 int parent_level = curlevel + 1;
2650 uint64_t parent_blkid = curblkid >> epbs;
2653 if (dbuf_hold_impl(dn, parent_level, parent_blkid,
2654 FALSE, TRUE, FTAG, &db) == 0) {
2655 blkptr_t *bpp = db->db_buf->b_data;
2656 bp = bpp[P2PHASE(curblkid, 1 << epbs)];
2657 dbuf_rele(db, FTAG);
2661 curlevel = parent_level;
2662 curblkid = parent_blkid;
2665 if (curlevel == nlevels - 1) {
2666 /* No cached indirect blocks found. */
2667 ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr);
2668 bp = dn->dn_phys->dn_blkptr[curblkid];
2670 if (BP_IS_HOLE(&bp))
2673 ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp));
2675 pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL,
2678 dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP);
2679 ds = dn->dn_objset->os_dsl_dataset;
2680 SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2681 dn->dn_object, level, blkid);
2682 dpa->dpa_curlevel = curlevel;
2683 dpa->dpa_prio = prio;
2684 dpa->dpa_aflags = aflags;
2685 dpa->dpa_spa = dn->dn_objset->os_spa;
2686 dpa->dpa_dnode = dn;
2687 dpa->dpa_epbs = epbs;
2691 * If we have the indirect just above us, no need to do the asynchronous
2692 * prefetch chain; we'll just run the last step ourselves. If we're at
2693 * a higher level, though, we want to issue the prefetches for all the
2694 * indirect blocks asynchronously, so we can go on with whatever we were
2697 if (curlevel == level) {
2698 ASSERT3U(curblkid, ==, blkid);
2699 dbuf_issue_final_prefetch(dpa, &bp);
2700 kmem_free(dpa, sizeof (*dpa));
2702 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2703 zbookmark_phys_t zb;
2705 SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2706 dn->dn_object, curlevel, curblkid);
2707 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2708 &bp, dbuf_prefetch_indirect_done, dpa, prio,
2709 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2713 * We use pio here instead of dpa_zio since it's possible that
2714 * dpa may have already been freed.
2719 #define DBUF_HOLD_IMPL_MAX_DEPTH 20
2722 * Returns with db_holds incremented, and db_mtx not held.
2723 * Note: dn_struct_rwlock must be held.
2726 __dbuf_hold_impl(struct dbuf_hold_impl_data *dh)
2728 ASSERT3S(dh->dh_depth, <, DBUF_HOLD_IMPL_MAX_DEPTH);
2729 dh->dh_parent = NULL;
2731 ASSERT(dh->dh_blkid != DMU_BONUS_BLKID);
2732 ASSERT(RW_LOCK_HELD(&dh->dh_dn->dn_struct_rwlock));
2733 ASSERT3U(dh->dh_dn->dn_nlevels, >, dh->dh_level);
2735 *(dh->dh_dbp) = NULL;
2737 /* dbuf_find() returns with db_mtx held */
2738 dh->dh_db = dbuf_find(dh->dh_dn->dn_objset, dh->dh_dn->dn_object,
2739 dh->dh_level, dh->dh_blkid);
2741 if (dh->dh_db == NULL) {
2744 if (dh->dh_fail_uncached)
2745 return (SET_ERROR(ENOENT));
2747 ASSERT3P(dh->dh_parent, ==, NULL);
2748 dh->dh_err = dbuf_findbp(dh->dh_dn, dh->dh_level, dh->dh_blkid,
2749 dh->dh_fail_sparse, &dh->dh_parent, &dh->dh_bp, dh);
2750 if (dh->dh_fail_sparse) {
2751 if (dh->dh_err == 0 &&
2752 dh->dh_bp && BP_IS_HOLE(dh->dh_bp))
2753 dh->dh_err = SET_ERROR(ENOENT);
2756 dbuf_rele(dh->dh_parent, NULL);
2757 return (dh->dh_err);
2760 if (dh->dh_err && dh->dh_err != ENOENT)
2761 return (dh->dh_err);
2762 dh->dh_db = dbuf_create(dh->dh_dn, dh->dh_level, dh->dh_blkid,
2763 dh->dh_parent, dh->dh_bp);
2766 if (dh->dh_fail_uncached && dh->dh_db->db_state != DB_CACHED) {
2767 mutex_exit(&dh->dh_db->db_mtx);
2768 return (SET_ERROR(ENOENT));
2771 if (dh->dh_db->db_buf != NULL)
2772 ASSERT3P(dh->dh_db->db.db_data, ==, dh->dh_db->db_buf->b_data);
2774 ASSERT(dh->dh_db->db_buf == NULL || arc_referenced(dh->dh_db->db_buf));
2777 * If this buffer is currently syncing out, and we are are
2778 * still referencing it from db_data, we need to make a copy
2779 * of it in case we decide we want to dirty it again in this txg.
2781 if (dh->dh_db->db_level == 0 &&
2782 dh->dh_db->db_blkid != DMU_BONUS_BLKID &&
2783 dh->dh_dn->dn_object != DMU_META_DNODE_OBJECT &&
2784 dh->dh_db->db_state == DB_CACHED && dh->dh_db->db_data_pending) {
2785 dh->dh_dr = dh->dh_db->db_data_pending;
2787 if (dh->dh_dr->dt.dl.dr_data == dh->dh_db->db_buf) {
2788 dh->dh_type = DBUF_GET_BUFC_TYPE(dh->dh_db);
2790 dbuf_set_data(dh->dh_db,
2791 arc_alloc_buf(dh->dh_dn->dn_objset->os_spa,
2792 dh->dh_db, dh->dh_type, dh->dh_db->db.db_size));
2793 bcopy(dh->dh_dr->dt.dl.dr_data->b_data,
2794 dh->dh_db->db.db_data, dh->dh_db->db.db_size);
2798 if (multilist_link_active(&dh->dh_db->db_cache_link)) {
2799 ASSERT(refcount_is_zero(&dh->dh_db->db_holds));
2800 multilist_remove(dbuf_cache, dh->dh_db);
2801 (void) refcount_remove_many(&dbuf_cache_size,
2802 dh->dh_db->db.db_size, dh->dh_db);
2804 (void) refcount_add(&dh->dh_db->db_holds, dh->dh_tag);
2805 DBUF_VERIFY(dh->dh_db);
2806 mutex_exit(&dh->dh_db->db_mtx);
2808 /* NOTE: we can't rele the parent until after we drop the db_mtx */
2810 dbuf_rele(dh->dh_parent, NULL);
2812 ASSERT3P(DB_DNODE(dh->dh_db), ==, dh->dh_dn);
2813 ASSERT3U(dh->dh_db->db_blkid, ==, dh->dh_blkid);
2814 ASSERT3U(dh->dh_db->db_level, ==, dh->dh_level);
2815 *(dh->dh_dbp) = dh->dh_db;
2821 * The following code preserves the recursive function dbuf_hold_impl()
2822 * but moves the local variables AND function arguments to the heap to
2823 * minimize the stack frame size. Enough space is initially allocated
2824 * on the stack for 20 levels of recursion.
2827 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid,
2828 boolean_t fail_sparse, boolean_t fail_uncached,
2829 void *tag, dmu_buf_impl_t **dbp)
2831 struct dbuf_hold_impl_data *dh;
2834 dh = kmem_alloc(sizeof (struct dbuf_hold_impl_data) *
2835 DBUF_HOLD_IMPL_MAX_DEPTH, KM_SLEEP);
2836 __dbuf_hold_impl_init(dh, dn, level, blkid, fail_sparse,
2837 fail_uncached, tag, dbp, 0);
2839 error = __dbuf_hold_impl(dh);
2841 kmem_free(dh, sizeof (struct dbuf_hold_impl_data) *
2842 DBUF_HOLD_IMPL_MAX_DEPTH);
2848 __dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh,
2849 dnode_t *dn, uint8_t level, uint64_t blkid,
2850 boolean_t fail_sparse, boolean_t fail_uncached,
2851 void *tag, dmu_buf_impl_t **dbp, int depth)
2854 dh->dh_level = level;
2855 dh->dh_blkid = blkid;
2857 dh->dh_fail_sparse = fail_sparse;
2858 dh->dh_fail_uncached = fail_uncached;
2864 dh->dh_parent = NULL;
2870 dh->dh_depth = depth;
2874 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
2876 return (dbuf_hold_level(dn, 0, blkid, tag));
2880 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
2883 int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db);
2884 return (err ? NULL : db);
2888 dbuf_create_bonus(dnode_t *dn)
2890 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
2892 ASSERT(dn->dn_bonus == NULL);
2893 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
2897 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
2899 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2902 if (db->db_blkid != DMU_SPILL_BLKID)
2903 return (SET_ERROR(ENOTSUP));
2905 blksz = SPA_MINBLOCKSIZE;
2906 ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset)));
2907 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
2911 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2912 dbuf_new_size(db, blksz, tx);
2913 rw_exit(&dn->dn_struct_rwlock);
2920 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
2922 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
2925 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2927 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
2929 int64_t holds = refcount_add(&db->db_holds, tag);
2930 VERIFY3S(holds, >, 1);
2933 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
2935 dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid,
2938 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2939 dmu_buf_impl_t *found_db;
2940 boolean_t result = B_FALSE;
2942 if (blkid == DMU_BONUS_BLKID)
2943 found_db = dbuf_find_bonus(os, obj);
2945 found_db = dbuf_find(os, obj, 0, blkid);
2947 if (found_db != NULL) {
2948 if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) {
2949 (void) refcount_add(&db->db_holds, tag);
2952 mutex_exit(&found_db->db_mtx);
2958 * If you call dbuf_rele() you had better not be referencing the dnode handle
2959 * unless you have some other direct or indirect hold on the dnode. (An indirect
2960 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2961 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2962 * dnode's parent dbuf evicting its dnode handles.
2965 dbuf_rele(dmu_buf_impl_t *db, void *tag)
2967 mutex_enter(&db->db_mtx);
2968 dbuf_rele_and_unlock(db, tag);
2972 dmu_buf_rele(dmu_buf_t *db, void *tag)
2974 dbuf_rele((dmu_buf_impl_t *)db, tag);
2978 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2979 * db_dirtycnt and db_holds to be updated atomically.
2982 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2986 ASSERT(MUTEX_HELD(&db->db_mtx));
2990 * Remove the reference to the dbuf before removing its hold on the
2991 * dnode so we can guarantee in dnode_move() that a referenced bonus
2992 * buffer has a corresponding dnode hold.
2994 holds = refcount_remove(&db->db_holds, tag);
2998 * We can't freeze indirects if there is a possibility that they
2999 * may be modified in the current syncing context.
3001 if (db->db_buf != NULL &&
3002 holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) {
3003 arc_buf_freeze(db->db_buf);
3006 if (holds == db->db_dirtycnt &&
3007 db->db_level == 0 && db->db_user_immediate_evict)
3008 dbuf_evict_user(db);
3011 if (db->db_blkid == DMU_BONUS_BLKID) {
3013 boolean_t evict_dbuf = db->db_pending_evict;
3016 * If the dnode moves here, we cannot cross this
3017 * barrier until the move completes.
3022 atomic_dec_32(&dn->dn_dbufs_count);
3025 * Decrementing the dbuf count means that the bonus
3026 * buffer's dnode hold is no longer discounted in
3027 * dnode_move(). The dnode cannot move until after
3028 * the dnode_rele() below.
3033 * Do not reference db after its lock is dropped.
3034 * Another thread may evict it.
3036 mutex_exit(&db->db_mtx);
3039 dnode_evict_bonus(dn);
3042 } else if (db->db_buf == NULL) {
3044 * This is a special case: we never associated this
3045 * dbuf with any data allocated from the ARC.
3047 ASSERT(db->db_state == DB_UNCACHED ||
3048 db->db_state == DB_NOFILL);
3050 } else if (arc_released(db->db_buf)) {
3052 * This dbuf has anonymous data associated with it.
3056 boolean_t do_arc_evict = B_FALSE;
3058 spa_t *spa = dmu_objset_spa(db->db_objset);
3060 if (!DBUF_IS_CACHEABLE(db) &&
3061 db->db_blkptr != NULL &&
3062 !BP_IS_HOLE(db->db_blkptr) &&
3063 !BP_IS_EMBEDDED(db->db_blkptr)) {
3064 do_arc_evict = B_TRUE;
3065 bp = *db->db_blkptr;
3068 if (!DBUF_IS_CACHEABLE(db) ||
3069 db->db_pending_evict) {
3071 } else if (!multilist_link_active(&db->db_cache_link)) {
3072 multilist_insert(dbuf_cache, db);
3073 (void) refcount_add_many(&dbuf_cache_size,
3074 db->db.db_size, db);
3075 mutex_exit(&db->db_mtx);
3077 dbuf_evict_notify();
3081 arc_freed(spa, &bp);
3084 mutex_exit(&db->db_mtx);
3089 #pragma weak dmu_buf_refcount = dbuf_refcount
3091 dbuf_refcount(dmu_buf_impl_t *db)
3093 return (refcount_count(&db->db_holds));
3097 dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user,
3098 dmu_buf_user_t *new_user)
3100 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3102 mutex_enter(&db->db_mtx);
3103 dbuf_verify_user(db, DBVU_NOT_EVICTING);
3104 if (db->db_user == old_user)
3105 db->db_user = new_user;
3107 old_user = db->db_user;
3108 dbuf_verify_user(db, DBVU_NOT_EVICTING);
3109 mutex_exit(&db->db_mtx);
3115 dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
3117 return (dmu_buf_replace_user(db_fake, NULL, user));
3121 dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user)
3123 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3125 db->db_user_immediate_evict = TRUE;
3126 return (dmu_buf_set_user(db_fake, user));
3130 dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
3132 return (dmu_buf_replace_user(db_fake, user, NULL));
3136 dmu_buf_get_user(dmu_buf_t *db_fake)
3138 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3140 dbuf_verify_user(db, DBVU_NOT_EVICTING);
3141 return (db->db_user);
3145 dmu_buf_user_evict_wait()
3147 taskq_wait(dbu_evict_taskq);
3151 dmu_buf_get_blkptr(dmu_buf_t *db)
3153 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3154 return (dbi->db_blkptr);
3158 dmu_buf_get_objset(dmu_buf_t *db)
3160 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3161 return (dbi->db_objset);
3165 dmu_buf_dnode_enter(dmu_buf_t *db)
3167 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3168 DB_DNODE_ENTER(dbi);
3169 return (DB_DNODE(dbi));
3173 dmu_buf_dnode_exit(dmu_buf_t *db)
3175 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3180 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
3182 /* ASSERT(dmu_tx_is_syncing(tx) */
3183 ASSERT(MUTEX_HELD(&db->db_mtx));
3185 if (db->db_blkptr != NULL)
3188 if (db->db_blkid == DMU_SPILL_BLKID) {
3189 db->db_blkptr = DN_SPILL_BLKPTR(dn->dn_phys);
3190 BP_ZERO(db->db_blkptr);
3193 if (db->db_level == dn->dn_phys->dn_nlevels-1) {
3195 * This buffer was allocated at a time when there was
3196 * no available blkptrs from the dnode, or it was
3197 * inappropriate to hook it in (i.e., nlevels mis-match).
3199 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
3200 ASSERT(db->db_parent == NULL);
3201 db->db_parent = dn->dn_dbuf;
3202 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
3205 dmu_buf_impl_t *parent = db->db_parent;
3206 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3208 ASSERT(dn->dn_phys->dn_nlevels > 1);
3209 if (parent == NULL) {
3210 mutex_exit(&db->db_mtx);
3211 rw_enter(&dn->dn_struct_rwlock, RW_READER);
3212 parent = dbuf_hold_level(dn, db->db_level + 1,
3213 db->db_blkid >> epbs, db);
3214 rw_exit(&dn->dn_struct_rwlock);
3215 mutex_enter(&db->db_mtx);
3216 db->db_parent = parent;
3218 db->db_blkptr = (blkptr_t *)parent->db.db_data +
3219 (db->db_blkid & ((1ULL << epbs) - 1));
3225 * Ensure the dbuf's data is untransformed if the associated dirty
3226 * record requires it. This is used by dbuf_sync_leaf() to ensure
3227 * that a dnode block is decrypted before we write new data to it.
3228 * For raw writes we assert that the buffer is already encrypted.
3231 dbuf_check_crypt(dbuf_dirty_record_t *dr)
3234 dmu_buf_impl_t *db = dr->dr_dbuf;
3236 ASSERT(MUTEX_HELD(&db->db_mtx));
3238 if (!dr->dt.dl.dr_raw && arc_is_encrypted(db->db_buf)) {
3240 * Unfortunately, there is currently no mechanism for
3241 * syncing context to handle decryption errors. An error
3242 * here is only possible if an attacker maliciously
3243 * changed a dnode block and updated the associated
3244 * checksums going up the block tree.
3246 err = arc_untransform(db->db_buf, db->db_objset->os_spa,
3247 dmu_objset_id(db->db_objset), B_TRUE);
3249 panic("Invalid dnode block MAC");
3250 } else if (dr->dt.dl.dr_raw) {
3252 * Writing raw encrypted data requires the db's arc buffer
3253 * to be converted to raw by the caller.
3255 ASSERT(arc_is_encrypted(db->db_buf));
3260 * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it
3261 * is critical the we not allow the compiler to inline this function in to
3262 * dbuf_sync_list() thereby drastically bloating the stack usage.
3264 noinline static void
3265 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
3267 dmu_buf_impl_t *db = dr->dr_dbuf;
3271 ASSERT(dmu_tx_is_syncing(tx));
3273 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
3275 mutex_enter(&db->db_mtx);
3277 ASSERT(db->db_level > 0);
3280 /* Read the block if it hasn't been read yet. */
3281 if (db->db_buf == NULL) {
3282 mutex_exit(&db->db_mtx);
3283 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
3284 mutex_enter(&db->db_mtx);
3286 ASSERT3U(db->db_state, ==, DB_CACHED);
3287 ASSERT(db->db_buf != NULL);
3291 /* Indirect block size must match what the dnode thinks it is. */
3292 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3293 dbuf_check_blkptr(dn, db);
3296 /* Provide the pending dirty record to child dbufs */
3297 db->db_data_pending = dr;
3299 mutex_exit(&db->db_mtx);
3300 dbuf_write(dr, db->db_buf, tx);
3303 mutex_enter(&dr->dt.di.dr_mtx);
3304 dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx);
3305 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3306 mutex_exit(&dr->dt.di.dr_mtx);
3311 * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is
3312 * critical the we not allow the compiler to inline this function in to
3313 * dbuf_sync_list() thereby drastically bloating the stack usage.
3315 noinline static void
3316 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
3318 arc_buf_t **datap = &dr->dt.dl.dr_data;
3319 dmu_buf_impl_t *db = dr->dr_dbuf;
3322 uint64_t txg = tx->tx_txg;
3324 ASSERT(dmu_tx_is_syncing(tx));
3326 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
3328 mutex_enter(&db->db_mtx);
3330 * To be synced, we must be dirtied. But we
3331 * might have been freed after the dirty.
3333 if (db->db_state == DB_UNCACHED) {
3334 /* This buffer has been freed since it was dirtied */
3335 ASSERT(db->db.db_data == NULL);
3336 } else if (db->db_state == DB_FILL) {
3337 /* This buffer was freed and is now being re-filled */
3338 ASSERT(db->db.db_data != dr->dt.dl.dr_data);
3340 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
3347 if (db->db_blkid == DMU_SPILL_BLKID) {
3348 mutex_enter(&dn->dn_mtx);
3349 if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
3351 * In the previous transaction group, the bonus buffer
3352 * was entirely used to store the attributes for the
3353 * dnode which overrode the dn_spill field. However,
3354 * when adding more attributes to the file a spill
3355 * block was required to hold the extra attributes.
3357 * Make sure to clear the garbage left in the dn_spill
3358 * field from the previous attributes in the bonus
3359 * buffer. Otherwise, after writing out the spill
3360 * block to the new allocated dva, it will free
3361 * the old block pointed to by the invalid dn_spill.
3363 db->db_blkptr = NULL;
3365 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
3366 mutex_exit(&dn->dn_mtx);
3370 * If this is a bonus buffer, simply copy the bonus data into the
3371 * dnode. It will be written out when the dnode is synced (and it
3372 * will be synced, since it must have been dirty for dbuf_sync to
3375 if (db->db_blkid == DMU_BONUS_BLKID) {
3376 dbuf_dirty_record_t **drp;
3378 ASSERT(*datap != NULL);
3379 ASSERT0(db->db_level);
3380 ASSERT3U(DN_MAX_BONUS_LEN(dn->dn_phys), <=,
3381 DN_SLOTS_TO_BONUSLEN(dn->dn_phys->dn_extra_slots + 1));
3382 bcopy(*datap, DN_BONUS(dn->dn_phys),
3383 DN_MAX_BONUS_LEN(dn->dn_phys));
3386 if (*datap != db->db.db_data) {
3387 int slots = DB_DNODE(db)->dn_num_slots;
3388 int bonuslen = DN_SLOTS_TO_BONUSLEN(slots);
3389 kmem_free(*datap, bonuslen);
3390 arc_space_return(bonuslen, ARC_SPACE_BONUS);
3392 db->db_data_pending = NULL;
3393 drp = &db->db_last_dirty;
3395 drp = &(*drp)->dr_next;
3396 ASSERT(dr->dr_next == NULL);
3397 ASSERT(dr->dr_dbuf == db);
3399 if (dr->dr_dbuf->db_level != 0) {
3400 mutex_destroy(&dr->dt.di.dr_mtx);
3401 list_destroy(&dr->dt.di.dr_children);
3403 kmem_free(dr, sizeof (dbuf_dirty_record_t));
3404 ASSERT(db->db_dirtycnt > 0);
3405 db->db_dirtycnt -= 1;
3406 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
3413 * This function may have dropped the db_mtx lock allowing a dmu_sync
3414 * operation to sneak in. As a result, we need to ensure that we
3415 * don't check the dr_override_state until we have returned from
3416 * dbuf_check_blkptr.
3418 dbuf_check_blkptr(dn, db);
3421 * If this buffer is in the middle of an immediate write,
3422 * wait for the synchronous IO to complete.
3424 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
3425 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
3426 cv_wait(&db->db_changed, &db->db_mtx);
3427 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
3431 * If this is a dnode block, ensure it is appropriately encrypted
3432 * or decrypted, depending on what we are writing to it this txg.
3434 if (os->os_encrypted && dn->dn_object == DMU_META_DNODE_OBJECT)
3435 dbuf_check_crypt(dr);
3437 if (db->db_state != DB_NOFILL &&
3438 dn->dn_object != DMU_META_DNODE_OBJECT &&
3439 refcount_count(&db->db_holds) > 1 &&
3440 dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
3441 *datap == db->db_buf) {
3443 * If this buffer is currently "in use" (i.e., there
3444 * are active holds and db_data still references it),
3445 * then make a copy before we start the write so that
3446 * any modifications from the open txg will not leak
3449 * NOTE: this copy does not need to be made for
3450 * objects only modified in the syncing context (e.g.
3451 * DNONE_DNODE blocks).
3453 int psize = arc_buf_size(*datap);
3454 int lsize = arc_buf_lsize(*datap);
3455 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
3456 enum zio_compress compress_type = arc_get_compression(*datap);
3458 if (arc_is_encrypted(*datap)) {
3459 boolean_t byteorder;
3460 uint8_t salt[ZIO_DATA_SALT_LEN];
3461 uint8_t iv[ZIO_DATA_IV_LEN];
3462 uint8_t mac[ZIO_DATA_MAC_LEN];
3464 arc_get_raw_params(*datap, &byteorder, salt, iv, mac);
3465 *datap = arc_alloc_raw_buf(os->os_spa, db,
3466 dmu_objset_id(os), byteorder, salt, iv, mac,
3467 dn->dn_type, psize, lsize, compress_type);
3468 } else if (compress_type != ZIO_COMPRESS_OFF) {
3469 ASSERT3U(type, ==, ARC_BUFC_DATA);
3470 *datap = arc_alloc_compressed_buf(os->os_spa, db,
3471 psize, lsize, compress_type);
3473 *datap = arc_alloc_buf(os->os_spa, db, type, psize);
3475 bcopy(db->db.db_data, (*datap)->b_data, psize);
3477 db->db_data_pending = dr;
3479 mutex_exit(&db->db_mtx);
3481 dbuf_write(dr, *datap, tx);
3483 ASSERT(!list_link_active(&dr->dr_dirty_node));
3484 if (dn->dn_object == DMU_META_DNODE_OBJECT) {
3485 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
3489 * Although zio_nowait() does not "wait for an IO", it does
3490 * initiate the IO. If this is an empty write it seems plausible
3491 * that the IO could actually be completed before the nowait
3492 * returns. We need to DB_DNODE_EXIT() first in case
3493 * zio_nowait() invalidates the dbuf.
3496 zio_nowait(dr->dr_zio);
3501 dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx)
3503 dbuf_dirty_record_t *dr;
3505 while ((dr = list_head(list))) {
3506 if (dr->dr_zio != NULL) {
3508 * If we find an already initialized zio then we
3509 * are processing the meta-dnode, and we have finished.
3510 * The dbufs for all dnodes are put back on the list
3511 * during processing, so that we can zio_wait()
3512 * these IOs after initiating all child IOs.
3514 ASSERT3U(dr->dr_dbuf->db.db_object, ==,
3515 DMU_META_DNODE_OBJECT);
3518 if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
3519 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
3520 VERIFY3U(dr->dr_dbuf->db_level, ==, level);
3522 list_remove(list, dr);
3523 if (dr->dr_dbuf->db_level > 0)
3524 dbuf_sync_indirect(dr, tx);
3526 dbuf_sync_leaf(dr, tx);
3532 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3534 dmu_buf_impl_t *db = vdb;
3536 blkptr_t *bp = zio->io_bp;
3537 blkptr_t *bp_orig = &zio->io_bp_orig;
3538 spa_t *spa = zio->io_spa;
3543 ASSERT3P(db->db_blkptr, !=, NULL);
3544 ASSERT3P(&db->db_data_pending->dr_bp_copy, ==, bp);
3548 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
3549 dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
3550 zio->io_prev_space_delta = delta;
3552 if (bp->blk_birth != 0) {
3553 ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
3554 BP_GET_TYPE(bp) == dn->dn_type) ||
3555 (db->db_blkid == DMU_SPILL_BLKID &&
3556 BP_GET_TYPE(bp) == dn->dn_bonustype) ||
3557 BP_IS_EMBEDDED(bp));
3558 ASSERT(BP_GET_LEVEL(bp) == db->db_level);
3561 mutex_enter(&db->db_mtx);
3564 if (db->db_blkid == DMU_SPILL_BLKID) {
3565 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3566 ASSERT(!(BP_IS_HOLE(bp)) &&
3567 db->db_blkptr == DN_SPILL_BLKPTR(dn->dn_phys));
3571 if (db->db_level == 0) {
3572 mutex_enter(&dn->dn_mtx);
3573 if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
3574 db->db_blkid != DMU_SPILL_BLKID)
3575 dn->dn_phys->dn_maxblkid = db->db_blkid;
3576 mutex_exit(&dn->dn_mtx);
3578 if (dn->dn_type == DMU_OT_DNODE) {
3580 while (i < db->db.db_size) {
3582 (void *)(((char *)db->db.db_data) + i);
3584 i += DNODE_MIN_SIZE;
3585 if (dnp->dn_type != DMU_OT_NONE) {
3587 i += dnp->dn_extra_slots *
3592 if (BP_IS_HOLE(bp)) {
3599 blkptr_t *ibp = db->db.db_data;
3600 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3601 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
3602 if (BP_IS_HOLE(ibp))
3604 fill += BP_GET_FILL(ibp);
3609 if (!BP_IS_EMBEDDED(bp))
3610 BP_SET_FILL(bp, fill);
3612 mutex_exit(&db->db_mtx);
3614 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3615 *db->db_blkptr = *bp;
3616 rw_exit(&dn->dn_struct_rwlock);
3621 * This function gets called just prior to running through the compression
3622 * stage of the zio pipeline. If we're an indirect block comprised of only
3623 * holes, then we want this indirect to be compressed away to a hole. In
3624 * order to do that we must zero out any information about the holes that
3625 * this indirect points to prior to before we try to compress it.
3628 dbuf_write_children_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3630 dmu_buf_impl_t *db = vdb;
3633 unsigned int epbs, i;
3635 ASSERT3U(db->db_level, >, 0);
3638 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3639 ASSERT3U(epbs, <, 31);
3641 /* Determine if all our children are holes */
3642 for (i = 0, bp = db->db.db_data; i < 1ULL << epbs; i++, bp++) {
3643 if (!BP_IS_HOLE(bp))
3648 * If all the children are holes, then zero them all out so that
3649 * we may get compressed away.
3651 if (i == 1ULL << epbs) {
3653 * We only found holes. Grab the rwlock to prevent
3654 * anybody from reading the blocks we're about to
3657 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3658 bzero(db->db.db_data, db->db.db_size);
3659 rw_exit(&dn->dn_struct_rwlock);
3665 * The SPA will call this callback several times for each zio - once
3666 * for every physical child i/o (zio->io_phys_children times). This
3667 * allows the DMU to monitor the progress of each logical i/o. For example,
3668 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
3669 * block. There may be a long delay before all copies/fragments are completed,
3670 * so this callback allows us to retire dirty space gradually, as the physical
3675 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
3677 dmu_buf_impl_t *db = arg;
3678 objset_t *os = db->db_objset;
3679 dsl_pool_t *dp = dmu_objset_pool(os);
3680 dbuf_dirty_record_t *dr;
3683 dr = db->db_data_pending;
3684 ASSERT3U(dr->dr_txg, ==, zio->io_txg);
3687 * The callback will be called io_phys_children times. Retire one
3688 * portion of our dirty space each time we are called. Any rounding
3689 * error will be cleaned up by dsl_pool_sync()'s call to
3690 * dsl_pool_undirty_space().
3692 delta = dr->dr_accounted / zio->io_phys_children;
3693 dsl_pool_undirty_space(dp, delta, zio->io_txg);
3698 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
3700 dmu_buf_impl_t *db = vdb;
3701 blkptr_t *bp_orig = &zio->io_bp_orig;
3702 blkptr_t *bp = db->db_blkptr;
3703 objset_t *os = db->db_objset;
3704 dmu_tx_t *tx = os->os_synctx;
3705 dbuf_dirty_record_t **drp, *dr;
3707 ASSERT0(zio->io_error);
3708 ASSERT(db->db_blkptr == bp);
3711 * For nopwrites and rewrites we ensure that the bp matches our
3712 * original and bypass all the accounting.
3714 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
3715 ASSERT(BP_EQUAL(bp, bp_orig));
3717 dsl_dataset_t *ds = os->os_dsl_dataset;
3718 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
3719 dsl_dataset_block_born(ds, bp, tx);
3722 mutex_enter(&db->db_mtx);
3726 drp = &db->db_last_dirty;
3727 while ((dr = *drp) != db->db_data_pending)
3729 ASSERT(!list_link_active(&dr->dr_dirty_node));
3730 ASSERT(dr->dr_dbuf == db);
3731 ASSERT(dr->dr_next == NULL);
3735 if (db->db_blkid == DMU_SPILL_BLKID) {
3740 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3741 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
3742 db->db_blkptr == DN_SPILL_BLKPTR(dn->dn_phys));
3747 if (db->db_level == 0) {
3748 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
3749 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
3750 if (db->db_state != DB_NOFILL) {
3751 if (dr->dt.dl.dr_data != db->db_buf)
3752 arc_buf_destroy(dr->dt.dl.dr_data, db);
3759 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3760 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
3761 if (!BP_IS_HOLE(db->db_blkptr)) {
3762 ASSERTV(int epbs = dn->dn_phys->dn_indblkshift -
3764 ASSERT3U(db->db_blkid, <=,
3765 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs));
3766 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
3770 mutex_destroy(&dr->dt.di.dr_mtx);
3771 list_destroy(&dr->dt.di.dr_children);
3773 kmem_free(dr, sizeof (dbuf_dirty_record_t));
3775 cv_broadcast(&db->db_changed);
3776 ASSERT(db->db_dirtycnt > 0);
3777 db->db_dirtycnt -= 1;
3778 db->db_data_pending = NULL;
3779 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg);
3783 dbuf_write_nofill_ready(zio_t *zio)
3785 dbuf_write_ready(zio, NULL, zio->io_private);
3789 dbuf_write_nofill_done(zio_t *zio)
3791 dbuf_write_done(zio, NULL, zio->io_private);
3795 dbuf_write_override_ready(zio_t *zio)
3797 dbuf_dirty_record_t *dr = zio->io_private;
3798 dmu_buf_impl_t *db = dr->dr_dbuf;
3800 dbuf_write_ready(zio, NULL, db);
3804 dbuf_write_override_done(zio_t *zio)
3806 dbuf_dirty_record_t *dr = zio->io_private;
3807 dmu_buf_impl_t *db = dr->dr_dbuf;
3808 blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
3810 mutex_enter(&db->db_mtx);
3811 if (!BP_EQUAL(zio->io_bp, obp)) {
3812 if (!BP_IS_HOLE(obp))
3813 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
3814 arc_release(dr->dt.dl.dr_data, db);
3816 mutex_exit(&db->db_mtx);
3818 dbuf_write_done(zio, NULL, db);
3820 if (zio->io_abd != NULL)
3821 abd_put(zio->io_abd);
3824 /* Issue I/O to commit a dirty buffer to disk. */
3826 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
3828 dmu_buf_impl_t *db = dr->dr_dbuf;
3831 dmu_buf_impl_t *parent = db->db_parent;
3832 uint64_t txg = tx->tx_txg;
3833 zbookmark_phys_t zb;
3838 ASSERT(dmu_tx_is_syncing(tx));
3844 if (db->db_state != DB_NOFILL) {
3845 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
3847 * Private object buffers are released here rather
3848 * than in dbuf_dirty() since they are only modified
3849 * in the syncing context and we don't want the
3850 * overhead of making multiple copies of the data.
3852 if (BP_IS_HOLE(db->db_blkptr)) {
3855 dbuf_release_bp(db);
3860 if (parent != dn->dn_dbuf) {
3861 /* Our parent is an indirect block. */
3862 /* We have a dirty parent that has been scheduled for write. */
3863 ASSERT(parent && parent->db_data_pending);
3864 /* Our parent's buffer is one level closer to the dnode. */
3865 ASSERT(db->db_level == parent->db_level-1);
3867 * We're about to modify our parent's db_data by modifying
3868 * our block pointer, so the parent must be released.
3870 ASSERT(arc_released(parent->db_buf));
3871 zio = parent->db_data_pending->dr_zio;
3873 /* Our parent is the dnode itself. */
3874 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
3875 db->db_blkid != DMU_SPILL_BLKID) ||
3876 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
3877 if (db->db_blkid != DMU_SPILL_BLKID)
3878 ASSERT3P(db->db_blkptr, ==,
3879 &dn->dn_phys->dn_blkptr[db->db_blkid]);
3883 ASSERT(db->db_level == 0 || data == db->db_buf);
3884 ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
3887 SET_BOOKMARK(&zb, os->os_dsl_dataset ?
3888 os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
3889 db->db.db_object, db->db_level, db->db_blkid);
3891 if (db->db_blkid == DMU_SPILL_BLKID)
3893 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
3895 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
3899 * We copy the blkptr now (rather than when we instantiate the dirty
3900 * record), because its value can change between open context and
3901 * syncing context. We do not need to hold dn_struct_rwlock to read
3902 * db_blkptr because we are in syncing context.
3904 dr->dr_bp_copy = *db->db_blkptr;
3906 if (db->db_level == 0 &&
3907 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
3909 * The BP for this block has been provided by open context
3910 * (by dmu_sync() or dmu_buf_write_embedded()).
3912 abd_t *contents = (data != NULL) ?
3913 abd_get_from_buf(data->b_data, arc_buf_size(data)) : NULL;
3915 dr->dr_zio = zio_write(zio, os->os_spa, txg,
3916 &dr->dr_bp_copy, contents, db->db.db_size, db->db.db_size,
3917 &zp, dbuf_write_override_ready, NULL, NULL,
3918 dbuf_write_override_done,
3919 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
3920 mutex_enter(&db->db_mtx);
3921 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
3922 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
3923 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
3924 mutex_exit(&db->db_mtx);
3925 } else if (db->db_state == DB_NOFILL) {
3926 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF ||
3927 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY);
3928 dr->dr_zio = zio_write(zio, os->os_spa, txg,
3929 &dr->dr_bp_copy, NULL, db->db.db_size, db->db.db_size, &zp,
3930 dbuf_write_nofill_ready, NULL, NULL,
3931 dbuf_write_nofill_done, db,
3932 ZIO_PRIORITY_ASYNC_WRITE,
3933 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
3935 arc_write_done_func_t *children_ready_cb = NULL;
3936 ASSERT(arc_released(data));
3939 * For indirect blocks, we want to setup the children
3940 * ready callback so that we can properly handle an indirect
3941 * block that only contains holes.
3943 if (db->db_level != 0)
3944 children_ready_cb = dbuf_write_children_ready;
3946 dr->dr_zio = arc_write(zio, os->os_spa, txg,
3947 &dr->dr_bp_copy, data, DBUF_IS_L2CACHEABLE(db),
3948 &zp, dbuf_write_ready,
3949 children_ready_cb, dbuf_write_physdone,
3950 dbuf_write_done, db, ZIO_PRIORITY_ASYNC_WRITE,
3951 ZIO_FLAG_MUSTSUCCEED, &zb);
3955 #if defined(_KERNEL) && defined(HAVE_SPL)
3956 EXPORT_SYMBOL(dbuf_find);
3957 EXPORT_SYMBOL(dbuf_is_metadata);
3958 EXPORT_SYMBOL(dbuf_destroy);
3959 EXPORT_SYMBOL(dbuf_loan_arcbuf);
3960 EXPORT_SYMBOL(dbuf_whichblock);
3961 EXPORT_SYMBOL(dbuf_read);
3962 EXPORT_SYMBOL(dbuf_unoverride);
3963 EXPORT_SYMBOL(dbuf_free_range);
3964 EXPORT_SYMBOL(dbuf_new_size);
3965 EXPORT_SYMBOL(dbuf_release_bp);
3966 EXPORT_SYMBOL(dbuf_dirty);
3967 EXPORT_SYMBOL(dmu_buf_will_change_crypt_params);
3968 EXPORT_SYMBOL(dmu_buf_will_dirty);
3969 EXPORT_SYMBOL(dmu_buf_will_not_fill);
3970 EXPORT_SYMBOL(dmu_buf_will_fill);
3971 EXPORT_SYMBOL(dmu_buf_fill_done);
3972 EXPORT_SYMBOL(dmu_buf_rele);
3973 EXPORT_SYMBOL(dbuf_assign_arcbuf);
3974 EXPORT_SYMBOL(dbuf_prefetch);
3975 EXPORT_SYMBOL(dbuf_hold_impl);
3976 EXPORT_SYMBOL(dbuf_hold);
3977 EXPORT_SYMBOL(dbuf_hold_level);
3978 EXPORT_SYMBOL(dbuf_create_bonus);
3979 EXPORT_SYMBOL(dbuf_spill_set_blksz);
3980 EXPORT_SYMBOL(dbuf_rm_spill);
3981 EXPORT_SYMBOL(dbuf_add_ref);
3982 EXPORT_SYMBOL(dbuf_rele);
3983 EXPORT_SYMBOL(dbuf_rele_and_unlock);
3984 EXPORT_SYMBOL(dbuf_refcount);
3985 EXPORT_SYMBOL(dbuf_sync_list);
3986 EXPORT_SYMBOL(dmu_buf_set_user);
3987 EXPORT_SYMBOL(dmu_buf_set_user_ie);
3988 EXPORT_SYMBOL(dmu_buf_get_user);
3989 EXPORT_SYMBOL(dmu_buf_get_blkptr);
3992 module_param(dbuf_cache_max_bytes, ulong, 0644);
3993 MODULE_PARM_DESC(dbuf_cache_max_bytes,
3994 "Maximum size in bytes of the dbuf cache.");
3996 module_param(dbuf_cache_hiwater_pct, uint, 0644);
3997 MODULE_PARM_DESC(dbuf_cache_hiwater_pct,
3998 "Percentage over dbuf_cache_max_bytes when dbufs must be evicted "
4001 module_param(dbuf_cache_lowater_pct, uint, 0644);
4002 MODULE_PARM_DESC(dbuf_cache_lowater_pct,
4003 "Percentage below dbuf_cache_max_bytes when the evict thread stops "
4006 module_param(dbuf_cache_max_shift, int, 0644);
4007 MODULE_PARM_DESC(dbuf_cache_max_shift,
4008 "Cap the size of the dbuf cache to a log2 fraction of arc size.");