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 boolean_t
467 dbuf_cache_above_hiwater(void)
469 uint64_t dbuf_cache_hiwater_bytes =
470 (dbuf_cache_max_bytes * dbuf_cache_hiwater_pct) / 100;
472 return (refcount_count(&dbuf_cache_size) >
473 dbuf_cache_max_bytes + dbuf_cache_hiwater_bytes);
476 static inline boolean_t
477 dbuf_cache_above_lowater(void)
479 uint64_t dbuf_cache_lowater_bytes =
480 (dbuf_cache_max_bytes * dbuf_cache_lowater_pct) / 100;
482 return (refcount_count(&dbuf_cache_size) >
483 dbuf_cache_max_bytes - dbuf_cache_lowater_bytes);
487 * Evict the oldest eligible dbuf from the dbuf cache.
492 int idx = multilist_get_random_index(dbuf_cache);
493 multilist_sublist_t *mls = multilist_sublist_lock(dbuf_cache, idx);
495 ASSERT(!MUTEX_HELD(&dbuf_evict_lock));
498 * Set the thread's tsd to indicate that it's processing evictions.
499 * Once a thread stops evicting from the dbuf cache it will
500 * reset its tsd to NULL.
502 ASSERT3P(tsd_get(zfs_dbuf_evict_key), ==, NULL);
503 (void) tsd_set(zfs_dbuf_evict_key, (void *)B_TRUE);
505 db = multilist_sublist_tail(mls);
506 while (db != NULL && mutex_tryenter(&db->db_mtx) == 0) {
507 db = multilist_sublist_prev(mls, db);
510 DTRACE_PROBE2(dbuf__evict__one, dmu_buf_impl_t *, db,
511 multilist_sublist_t *, mls);
514 multilist_sublist_remove(mls, db);
515 multilist_sublist_unlock(mls);
516 (void) refcount_remove_many(&dbuf_cache_size,
520 multilist_sublist_unlock(mls);
522 (void) tsd_set(zfs_dbuf_evict_key, NULL);
526 * The dbuf evict thread is responsible for aging out dbufs from the
527 * cache. Once the cache has reached it's maximum size, dbufs are removed
528 * and destroyed. The eviction thread will continue running until the size
529 * of the dbuf cache is at or below the maximum size. Once the dbuf is aged
530 * out of the cache it is destroyed and becomes eligible for arc eviction.
533 dbuf_evict_thread(void *unused)
537 CALLB_CPR_INIT(&cpr, &dbuf_evict_lock, callb_generic_cpr, FTAG);
539 mutex_enter(&dbuf_evict_lock);
540 while (!dbuf_evict_thread_exit) {
541 while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
542 CALLB_CPR_SAFE_BEGIN(&cpr);
543 (void) cv_timedwait_sig_hires(&dbuf_evict_cv,
544 &dbuf_evict_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
545 CALLB_CPR_SAFE_END(&cpr, &dbuf_evict_lock);
547 mutex_exit(&dbuf_evict_lock);
550 * Keep evicting as long as we're above the low water mark
551 * for the cache. We do this without holding the locks to
552 * minimize lock contention.
554 while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
558 mutex_enter(&dbuf_evict_lock);
561 dbuf_evict_thread_exit = B_FALSE;
562 cv_broadcast(&dbuf_evict_cv);
563 CALLB_CPR_EXIT(&cpr); /* drops dbuf_evict_lock */
568 * Wake up the dbuf eviction thread if the dbuf cache is at its max size.
569 * If the dbuf cache is at its high water mark, then evict a dbuf from the
570 * dbuf cache using the callers context.
573 dbuf_evict_notify(void)
577 * We use thread specific data to track when a thread has
578 * started processing evictions. This allows us to avoid deeply
579 * nested stacks that would have a call flow similar to this:
581 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify()
584 * +-----dbuf_destroy()<--dbuf_evict_one()<--------+
586 * The dbuf_eviction_thread will always have its tsd set until
587 * that thread exits. All other threads will only set their tsd
588 * if they are participating in the eviction process. This only
589 * happens if the eviction thread is unable to process evictions
590 * fast enough. To keep the dbuf cache size in check, other threads
591 * can evict from the dbuf cache directly. Those threads will set
592 * their tsd values so that we ensure that they only evict one dbuf
593 * from the dbuf cache.
595 if (tsd_get(zfs_dbuf_evict_key) != NULL)
599 * We check if we should evict without holding the dbuf_evict_lock,
600 * because it's OK to occasionally make the wrong decision here,
601 * and grabbing the lock results in massive lock contention.
603 if (refcount_count(&dbuf_cache_size) > dbuf_cache_max_bytes) {
604 if (dbuf_cache_above_hiwater())
606 cv_signal(&dbuf_evict_cv);
615 uint64_t hsize = 1ULL << 16;
616 dbuf_hash_table_t *h = &dbuf_hash_table;
620 * The hash table is big enough to fill all of physical memory
621 * with an average block size of zfs_arc_average_blocksize (default 8K).
622 * By default, the table will take up
623 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
625 while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE)
629 h->hash_table_mask = hsize - 1;
630 #if defined(_KERNEL) && defined(HAVE_SPL)
632 * Large allocations which do not require contiguous pages
633 * should be using vmem_alloc() in the linux kernel
635 h->hash_table = vmem_zalloc(hsize * sizeof (void *), KM_SLEEP);
637 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
639 if (h->hash_table == NULL) {
640 /* XXX - we should really return an error instead of assert */
641 ASSERT(hsize > (1ULL << 10));
646 dbuf_kmem_cache = kmem_cache_create("dmu_buf_impl_t",
647 sizeof (dmu_buf_impl_t),
648 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
650 for (i = 0; i < DBUF_MUTEXES; i++)
651 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
656 * Setup the parameters for the dbuf cache. We cap the size of the
657 * dbuf cache to 1/32nd (default) of the size of the ARC.
659 dbuf_cache_max_bytes = MIN(dbuf_cache_max_bytes,
660 arc_max_bytes() >> dbuf_cache_max_shift);
663 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
664 * configuration is not required.
666 dbu_evict_taskq = taskq_create("dbu_evict", 1, defclsyspri, 0, 0, 0);
668 dbuf_cache = multilist_create(sizeof (dmu_buf_impl_t),
669 offsetof(dmu_buf_impl_t, db_cache_link),
670 dbuf_cache_multilist_index_func);
671 refcount_create(&dbuf_cache_size);
673 tsd_create(&zfs_dbuf_evict_key, NULL);
674 dbuf_evict_thread_exit = B_FALSE;
675 mutex_init(&dbuf_evict_lock, NULL, MUTEX_DEFAULT, NULL);
676 cv_init(&dbuf_evict_cv, NULL, CV_DEFAULT, NULL);
677 dbuf_cache_evict_thread = thread_create(NULL, 0, dbuf_evict_thread,
678 NULL, 0, &p0, TS_RUN, minclsyspri);
684 dbuf_hash_table_t *h = &dbuf_hash_table;
687 dbuf_stats_destroy();
689 for (i = 0; i < DBUF_MUTEXES; i++)
690 mutex_destroy(&h->hash_mutexes[i]);
691 #if defined(_KERNEL) && defined(HAVE_SPL)
693 * Large allocations which do not require contiguous pages
694 * should be using vmem_free() in the linux kernel
696 vmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
698 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
700 kmem_cache_destroy(dbuf_kmem_cache);
701 taskq_destroy(dbu_evict_taskq);
703 mutex_enter(&dbuf_evict_lock);
704 dbuf_evict_thread_exit = B_TRUE;
705 while (dbuf_evict_thread_exit) {
706 cv_signal(&dbuf_evict_cv);
707 cv_wait(&dbuf_evict_cv, &dbuf_evict_lock);
709 mutex_exit(&dbuf_evict_lock);
710 tsd_destroy(&zfs_dbuf_evict_key);
712 mutex_destroy(&dbuf_evict_lock);
713 cv_destroy(&dbuf_evict_cv);
715 refcount_destroy(&dbuf_cache_size);
716 multilist_destroy(dbuf_cache);
725 dbuf_verify(dmu_buf_impl_t *db)
728 dbuf_dirty_record_t *dr;
730 ASSERT(MUTEX_HELD(&db->db_mtx));
732 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
735 ASSERT(db->db_objset != NULL);
739 ASSERT(db->db_parent == NULL);
740 ASSERT(db->db_blkptr == NULL);
742 ASSERT3U(db->db.db_object, ==, dn->dn_object);
743 ASSERT3P(db->db_objset, ==, dn->dn_objset);
744 ASSERT3U(db->db_level, <, dn->dn_nlevels);
745 ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
746 db->db_blkid == DMU_SPILL_BLKID ||
747 !avl_is_empty(&dn->dn_dbufs));
749 if (db->db_blkid == DMU_BONUS_BLKID) {
751 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
752 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
753 } else if (db->db_blkid == DMU_SPILL_BLKID) {
755 ASSERT0(db->db.db_offset);
757 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
760 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
761 ASSERT(dr->dr_dbuf == db);
763 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
764 ASSERT(dr->dr_dbuf == db);
767 * We can't assert that db_size matches dn_datablksz because it
768 * can be momentarily different when another thread is doing
771 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
772 dr = db->db_data_pending;
774 * It should only be modified in syncing context, so
775 * make sure we only have one copy of the data.
777 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
780 /* verify db->db_blkptr */
782 if (db->db_parent == dn->dn_dbuf) {
783 /* db is pointed to by the dnode */
784 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
785 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
786 ASSERT(db->db_parent == NULL);
788 ASSERT(db->db_parent != NULL);
789 if (db->db_blkid != DMU_SPILL_BLKID)
790 ASSERT3P(db->db_blkptr, ==,
791 &dn->dn_phys->dn_blkptr[db->db_blkid]);
793 /* db is pointed to by an indirect block */
794 ASSERTV(int epb = db->db_parent->db.db_size >>
796 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
797 ASSERT3U(db->db_parent->db.db_object, ==,
800 * dnode_grow_indblksz() can make this fail if we don't
801 * have the struct_rwlock. XXX indblksz no longer
802 * grows. safe to do this now?
804 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
805 ASSERT3P(db->db_blkptr, ==,
806 ((blkptr_t *)db->db_parent->db.db_data +
807 db->db_blkid % epb));
811 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
812 (db->db_buf == NULL || db->db_buf->b_data) &&
813 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
814 db->db_state != DB_FILL && !dn->dn_free_txg) {
816 * If the blkptr isn't set but they have nonzero data,
817 * it had better be dirty, otherwise we'll lose that
818 * data when we evict this buffer.
820 * There is an exception to this rule for indirect blocks; in
821 * this case, if the indirect block is a hole, we fill in a few
822 * fields on each of the child blocks (importantly, birth time)
823 * to prevent hole birth times from being lost when you
824 * partially fill in a hole.
826 if (db->db_dirtycnt == 0) {
827 if (db->db_level == 0) {
828 uint64_t *buf = db->db.db_data;
831 for (i = 0; i < db->db.db_size >> 3; i++) {
836 blkptr_t *bps = db->db.db_data;
837 ASSERT3U(1 << DB_DNODE(db)->dn_indblkshift, ==,
840 * We want to verify that all the blkptrs in the
841 * indirect block are holes, but we may have
842 * automatically set up a few fields for them.
843 * We iterate through each blkptr and verify
844 * they only have those fields set.
847 i < db->db.db_size / sizeof (blkptr_t);
849 blkptr_t *bp = &bps[i];
850 ASSERT(ZIO_CHECKSUM_IS_ZERO(
853 DVA_IS_EMPTY(&bp->blk_dva[0]) &&
854 DVA_IS_EMPTY(&bp->blk_dva[1]) &&
855 DVA_IS_EMPTY(&bp->blk_dva[2]));
856 ASSERT0(bp->blk_fill);
857 ASSERT0(bp->blk_pad[0]);
858 ASSERT0(bp->blk_pad[1]);
859 ASSERT(!BP_IS_EMBEDDED(bp));
860 ASSERT(BP_IS_HOLE(bp));
861 ASSERT0(bp->blk_phys_birth);
871 dbuf_clear_data(dmu_buf_impl_t *db)
873 ASSERT(MUTEX_HELD(&db->db_mtx));
875 ASSERT3P(db->db_buf, ==, NULL);
876 db->db.db_data = NULL;
877 if (db->db_state != DB_NOFILL)
878 db->db_state = DB_UNCACHED;
882 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
884 ASSERT(MUTEX_HELD(&db->db_mtx));
888 ASSERT(buf->b_data != NULL);
889 db->db.db_data = buf->b_data;
893 * Loan out an arc_buf for read. Return the loaned arc_buf.
896 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
900 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
901 mutex_enter(&db->db_mtx);
902 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
903 int blksz = db->db.db_size;
904 spa_t *spa = db->db_objset->os_spa;
906 mutex_exit(&db->db_mtx);
907 abuf = arc_loan_buf(spa, B_FALSE, blksz);
908 bcopy(db->db.db_data, abuf->b_data, blksz);
911 arc_loan_inuse_buf(abuf, db);
914 mutex_exit(&db->db_mtx);
920 * Calculate which level n block references the data at the level 0 offset
924 dbuf_whichblock(const dnode_t *dn, const int64_t level, const uint64_t offset)
926 if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) {
928 * The level n blkid is equal to the level 0 blkid divided by
929 * the number of level 0s in a level n block.
931 * The level 0 blkid is offset >> datablkshift =
932 * offset / 2^datablkshift.
934 * The number of level 0s in a level n is the number of block
935 * pointers in an indirect block, raised to the power of level.
936 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
937 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
939 * Thus, the level n blkid is: offset /
940 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
941 * = offset / 2^(datablkshift + level *
942 * (indblkshift - SPA_BLKPTRSHIFT))
943 * = offset >> (datablkshift + level *
944 * (indblkshift - SPA_BLKPTRSHIFT))
947 const unsigned exp = dn->dn_datablkshift +
948 level * (dn->dn_indblkshift - SPA_BLKPTRSHIFT);
950 if (exp >= 8 * sizeof (offset)) {
951 /* This only happens on the highest indirection level */
952 ASSERT3U(level, ==, dn->dn_nlevels - 1);
956 ASSERT3U(exp, <, 8 * sizeof (offset));
958 return (offset >> exp);
960 ASSERT3U(offset, <, dn->dn_datablksz);
966 dbuf_read_done(zio_t *zio, int err, arc_buf_t *buf, void *vdb)
968 dmu_buf_impl_t *db = vdb;
970 mutex_enter(&db->db_mtx);
971 ASSERT3U(db->db_state, ==, DB_READ);
973 * All reads are synchronous, so we must have a hold on the dbuf
975 ASSERT(refcount_count(&db->db_holds) > 0);
976 ASSERT(db->db_buf == NULL);
977 ASSERT(db->db.db_data == NULL);
978 if (db->db_level == 0 && db->db_freed_in_flight) {
979 /* we were freed in flight; disregard any error */
980 arc_release(buf, db);
981 bzero(buf->b_data, db->db.db_size);
983 db->db_freed_in_flight = FALSE;
984 dbuf_set_data(db, buf);
985 db->db_state = DB_CACHED;
986 } else if (err == 0) {
987 dbuf_set_data(db, buf);
988 db->db_state = DB_CACHED;
990 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
991 ASSERT3P(db->db_buf, ==, NULL);
992 arc_buf_destroy(buf, db);
993 db->db_state = DB_UNCACHED;
995 cv_broadcast(&db->db_changed);
996 dbuf_rele_and_unlock(db, NULL);
1000 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
1003 zbookmark_phys_t zb;
1004 uint32_t aflags = ARC_FLAG_NOWAIT;
1005 int err, zio_flags = 0;
1009 ASSERT(!refcount_is_zero(&db->db_holds));
1010 /* We need the struct_rwlock to prevent db_blkptr from changing. */
1011 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1012 ASSERT(MUTEX_HELD(&db->db_mtx));
1013 ASSERT(db->db_state == DB_UNCACHED);
1014 ASSERT(db->db_buf == NULL);
1016 if (db->db_blkid == DMU_BONUS_BLKID) {
1018 * The bonus length stored in the dnode may be less than
1019 * the maximum available space in the bonus buffer.
1021 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
1022 int max_bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
1023 arc_buf_t *dn_buf = (dn->dn_dbuf != NULL) ?
1024 dn->dn_dbuf->db_buf : NULL;
1026 /* if the underlying dnode block is encrypted, decrypt it */
1027 if (dn_buf != NULL && dn->dn_objset->os_encrypted &&
1028 DMU_OT_IS_ENCRYPTED(dn->dn_bonustype) &&
1029 (flags & DB_RF_NO_DECRYPT) == 0 &&
1030 arc_is_encrypted(dn_buf)) {
1031 err = arc_untransform(dn_buf, dn->dn_objset->os_spa,
1032 dmu_objset_id(dn->dn_objset), B_TRUE);
1035 mutex_exit(&db->db_mtx);
1040 ASSERT3U(bonuslen, <=, db->db.db_size);
1041 db->db.db_data = kmem_alloc(max_bonuslen, KM_SLEEP);
1042 arc_space_consume(max_bonuslen, ARC_SPACE_BONUS);
1043 if (bonuslen < max_bonuslen)
1044 bzero(db->db.db_data, max_bonuslen);
1046 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
1048 db->db_state = DB_CACHED;
1049 mutex_exit(&db->db_mtx);
1054 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
1055 * processes the delete record and clears the bp while we are waiting
1056 * for the dn_mtx (resulting in a "no" from block_freed).
1058 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
1059 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
1060 BP_IS_HOLE(db->db_blkptr)))) {
1061 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1063 dbuf_set_data(db, arc_alloc_buf(db->db_objset->os_spa, db, type,
1065 bzero(db->db.db_data, db->db.db_size);
1067 if (db->db_blkptr != NULL && db->db_level > 0 &&
1068 BP_IS_HOLE(db->db_blkptr) &&
1069 db->db_blkptr->blk_birth != 0) {
1070 blkptr_t *bps = db->db.db_data;
1072 for (i = 0; i < ((1 <<
1073 DB_DNODE(db)->dn_indblkshift) / sizeof (blkptr_t));
1075 blkptr_t *bp = &bps[i];
1076 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
1077 1 << dn->dn_indblkshift);
1079 BP_GET_LEVEL(db->db_blkptr) == 1 ?
1081 BP_GET_LSIZE(db->db_blkptr));
1082 BP_SET_TYPE(bp, BP_GET_TYPE(db->db_blkptr));
1084 BP_GET_LEVEL(db->db_blkptr) - 1);
1085 BP_SET_BIRTH(bp, db->db_blkptr->blk_birth, 0);
1089 db->db_state = DB_CACHED;
1090 mutex_exit(&db->db_mtx);
1096 db->db_state = DB_READ;
1097 mutex_exit(&db->db_mtx);
1099 if (DBUF_IS_L2CACHEABLE(db))
1100 aflags |= ARC_FLAG_L2CACHE;
1102 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
1103 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
1104 db->db.db_object, db->db_level, db->db_blkid);
1107 * All bps of an encrypted os should have the encryption bit set.
1108 * If this is not true it indicates tampering and we report an error.
1110 if (db->db_objset->os_encrypted && !BP_USES_CRYPT(db->db_blkptr)) {
1111 spa_log_error(db->db_objset->os_spa, &zb);
1112 zfs_panic_recover("unencrypted block in encrypted "
1113 "object set %llu", dmu_objset_id(db->db_objset));
1114 return (SET_ERROR(EIO));
1117 dbuf_add_ref(db, NULL);
1119 zio_flags = (flags & DB_RF_CANFAIL) ?
1120 ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED;
1122 if ((flags & DB_RF_NO_DECRYPT) && BP_IS_PROTECTED(db->db_blkptr))
1123 zio_flags |= ZIO_FLAG_RAW;
1125 err = arc_read(zio, db->db_objset->os_spa, db->db_blkptr,
1126 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ, zio_flags,
1133 * This is our just-in-time copy function. It makes a copy of buffers that
1134 * have been modified in a previous transaction group before we access them in
1135 * the current active group.
1137 * This function is used in three places: when we are dirtying a buffer for the
1138 * first time in a txg, when we are freeing a range in a dnode that includes
1139 * this buffer, and when we are accessing a buffer which was received compressed
1140 * and later referenced in a WRITE_BYREF record.
1142 * Note that when we are called from dbuf_free_range() we do not put a hold on
1143 * the buffer, we just traverse the active dbuf list for the dnode.
1146 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
1148 dbuf_dirty_record_t *dr = db->db_last_dirty;
1150 ASSERT(MUTEX_HELD(&db->db_mtx));
1151 ASSERT(db->db.db_data != NULL);
1152 ASSERT(db->db_level == 0);
1153 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
1156 (dr->dt.dl.dr_data !=
1157 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
1161 * If the last dirty record for this dbuf has not yet synced
1162 * and its referencing the dbuf data, either:
1163 * reset the reference to point to a new copy,
1164 * or (if there a no active holders)
1165 * just null out the current db_data pointer.
1167 ASSERT(dr->dr_txg >= txg - 2);
1168 if (db->db_blkid == DMU_BONUS_BLKID) {
1169 dnode_t *dn = DB_DNODE(db);
1170 int bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
1171 dr->dt.dl.dr_data = kmem_alloc(bonuslen, KM_SLEEP);
1172 arc_space_consume(bonuslen, ARC_SPACE_BONUS);
1173 bcopy(db->db.db_data, dr->dt.dl.dr_data, bonuslen);
1174 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
1175 dnode_t *dn = DB_DNODE(db);
1176 int size = arc_buf_size(db->db_buf);
1177 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1178 spa_t *spa = db->db_objset->os_spa;
1179 enum zio_compress compress_type =
1180 arc_get_compression(db->db_buf);
1182 if (arc_is_encrypted(db->db_buf)) {
1183 boolean_t byteorder;
1184 uint8_t salt[ZIO_DATA_SALT_LEN];
1185 uint8_t iv[ZIO_DATA_IV_LEN];
1186 uint8_t mac[ZIO_DATA_MAC_LEN];
1188 arc_get_raw_params(db->db_buf, &byteorder, salt,
1190 dr->dt.dl.dr_data = arc_alloc_raw_buf(spa, db,
1191 dmu_objset_id(dn->dn_objset), byteorder, salt, iv,
1192 mac, dn->dn_type, size, arc_buf_lsize(db->db_buf),
1194 } else if (compress_type != ZIO_COMPRESS_OFF) {
1195 ASSERT3U(type, ==, ARC_BUFC_DATA);
1196 dr->dt.dl.dr_data = arc_alloc_compressed_buf(spa, db,
1197 size, arc_buf_lsize(db->db_buf), compress_type);
1199 dr->dt.dl.dr_data = arc_alloc_buf(spa, db, type, size);
1201 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
1204 dbuf_clear_data(db);
1209 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
1216 * We don't have to hold the mutex to check db_state because it
1217 * can't be freed while we have a hold on the buffer.
1219 ASSERT(!refcount_is_zero(&db->db_holds));
1221 if (db->db_state == DB_NOFILL)
1222 return (SET_ERROR(EIO));
1226 if ((flags & DB_RF_HAVESTRUCT) == 0)
1227 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1229 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1230 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
1231 DBUF_IS_CACHEABLE(db);
1233 mutex_enter(&db->db_mtx);
1234 if (db->db_state == DB_CACHED) {
1235 spa_t *spa = dn->dn_objset->os_spa;
1238 * If the arc buf is compressed or encrypted, we need to
1239 * untransform it to read the data. This could happen during
1240 * the "zfs receive" of a stream which is deduplicated and
1241 * either raw or compressed. We do not need to do this if the
1242 * caller wants raw encrypted data.
1244 if (db->db_buf != NULL && (flags & DB_RF_NO_DECRYPT) == 0 &&
1245 (arc_is_encrypted(db->db_buf) ||
1246 arc_get_compression(db->db_buf) != ZIO_COMPRESS_OFF)) {
1247 dbuf_fix_old_data(db, spa_syncing_txg(spa));
1248 err = arc_untransform(db->db_buf, spa,
1249 dmu_objset_id(db->db_objset), B_FALSE);
1250 dbuf_set_data(db, db->db_buf);
1252 mutex_exit(&db->db_mtx);
1254 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1255 if ((flags & DB_RF_HAVESTRUCT) == 0)
1256 rw_exit(&dn->dn_struct_rwlock);
1258 } else if (db->db_state == DB_UNCACHED) {
1259 spa_t *spa = dn->dn_objset->os_spa;
1260 boolean_t need_wait = B_FALSE;
1263 db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) {
1264 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
1267 err = dbuf_read_impl(db, zio, flags);
1269 /* dbuf_read_impl has dropped db_mtx for us */
1271 if (!err && prefetch)
1272 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1274 if ((flags & DB_RF_HAVESTRUCT) == 0)
1275 rw_exit(&dn->dn_struct_rwlock);
1278 if (!err && need_wait)
1279 err = zio_wait(zio);
1282 * Another reader came in while the dbuf was in flight
1283 * between UNCACHED and CACHED. Either a writer will finish
1284 * writing the buffer (sending the dbuf to CACHED) or the
1285 * first reader's request will reach the read_done callback
1286 * and send the dbuf to CACHED. Otherwise, a failure
1287 * occurred and the dbuf went to UNCACHED.
1289 mutex_exit(&db->db_mtx);
1291 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1292 if ((flags & DB_RF_HAVESTRUCT) == 0)
1293 rw_exit(&dn->dn_struct_rwlock);
1296 /* Skip the wait per the caller's request. */
1297 mutex_enter(&db->db_mtx);
1298 if ((flags & DB_RF_NEVERWAIT) == 0) {
1299 while (db->db_state == DB_READ ||
1300 db->db_state == DB_FILL) {
1301 ASSERT(db->db_state == DB_READ ||
1302 (flags & DB_RF_HAVESTRUCT) == 0);
1303 DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *,
1305 cv_wait(&db->db_changed, &db->db_mtx);
1307 if (db->db_state == DB_UNCACHED)
1308 err = SET_ERROR(EIO);
1310 mutex_exit(&db->db_mtx);
1317 dbuf_noread(dmu_buf_impl_t *db)
1319 ASSERT(!refcount_is_zero(&db->db_holds));
1320 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1321 mutex_enter(&db->db_mtx);
1322 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1323 cv_wait(&db->db_changed, &db->db_mtx);
1324 if (db->db_state == DB_UNCACHED) {
1325 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1326 spa_t *spa = db->db_objset->os_spa;
1328 ASSERT(db->db_buf == NULL);
1329 ASSERT(db->db.db_data == NULL);
1330 dbuf_set_data(db, arc_alloc_buf(spa, db, type, db->db.db_size));
1331 db->db_state = DB_FILL;
1332 } else if (db->db_state == DB_NOFILL) {
1333 dbuf_clear_data(db);
1335 ASSERT3U(db->db_state, ==, DB_CACHED);
1337 mutex_exit(&db->db_mtx);
1341 dbuf_unoverride(dbuf_dirty_record_t *dr)
1343 dmu_buf_impl_t *db = dr->dr_dbuf;
1344 blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
1345 uint64_t txg = dr->dr_txg;
1347 ASSERT(MUTEX_HELD(&db->db_mtx));
1349 * This assert is valid because dmu_sync() expects to be called by
1350 * a zilog's get_data while holding a range lock. This call only
1351 * comes from dbuf_dirty() callers who must also hold a range lock.
1353 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
1354 ASSERT(db->db_level == 0);
1356 if (db->db_blkid == DMU_BONUS_BLKID ||
1357 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
1360 ASSERT(db->db_data_pending != dr);
1362 /* free this block */
1363 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite)
1364 zio_free(db->db_objset->os_spa, txg, bp);
1366 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1367 dr->dt.dl.dr_nopwrite = B_FALSE;
1368 dr->dt.dl.dr_raw = B_FALSE;
1371 * Release the already-written buffer, so we leave it in
1372 * a consistent dirty state. Note that all callers are
1373 * modifying the buffer, so they will immediately do
1374 * another (redundant) arc_release(). Therefore, leave
1375 * the buf thawed to save the effort of freezing &
1376 * immediately re-thawing it.
1378 arc_release(dr->dt.dl.dr_data, db);
1382 * Evict (if its unreferenced) or clear (if its referenced) any level-0
1383 * data blocks in the free range, so that any future readers will find
1387 dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
1390 dmu_buf_impl_t *db_search;
1391 dmu_buf_impl_t *db, *db_next;
1392 uint64_t txg = tx->tx_txg;
1395 if (end_blkid > dn->dn_maxblkid &&
1396 !(start_blkid == DMU_SPILL_BLKID || end_blkid == DMU_SPILL_BLKID))
1397 end_blkid = dn->dn_maxblkid;
1398 dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid);
1400 db_search = kmem_alloc(sizeof (dmu_buf_impl_t), KM_SLEEP);
1401 db_search->db_level = 0;
1402 db_search->db_blkid = start_blkid;
1403 db_search->db_state = DB_SEARCH;
1405 mutex_enter(&dn->dn_dbufs_mtx);
1406 db = avl_find(&dn->dn_dbufs, db_search, &where);
1407 ASSERT3P(db, ==, NULL);
1409 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1411 for (; db != NULL; db = db_next) {
1412 db_next = AVL_NEXT(&dn->dn_dbufs, db);
1413 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1415 if (db->db_level != 0 || db->db_blkid > end_blkid) {
1418 ASSERT3U(db->db_blkid, >=, start_blkid);
1420 /* found a level 0 buffer in the range */
1421 mutex_enter(&db->db_mtx);
1422 if (dbuf_undirty(db, tx)) {
1423 /* mutex has been dropped and dbuf destroyed */
1427 if (db->db_state == DB_UNCACHED ||
1428 db->db_state == DB_NOFILL ||
1429 db->db_state == DB_EVICTING) {
1430 ASSERT(db->db.db_data == NULL);
1431 mutex_exit(&db->db_mtx);
1434 if (db->db_state == DB_READ || db->db_state == DB_FILL) {
1435 /* will be handled in dbuf_read_done or dbuf_rele */
1436 db->db_freed_in_flight = TRUE;
1437 mutex_exit(&db->db_mtx);
1440 if (refcount_count(&db->db_holds) == 0) {
1445 /* The dbuf is referenced */
1447 if (db->db_last_dirty != NULL) {
1448 dbuf_dirty_record_t *dr = db->db_last_dirty;
1450 if (dr->dr_txg == txg) {
1452 * This buffer is "in-use", re-adjust the file
1453 * size to reflect that this buffer may
1454 * contain new data when we sync.
1456 if (db->db_blkid != DMU_SPILL_BLKID &&
1457 db->db_blkid > dn->dn_maxblkid)
1458 dn->dn_maxblkid = db->db_blkid;
1459 dbuf_unoverride(dr);
1462 * This dbuf is not dirty in the open context.
1463 * Either uncache it (if its not referenced in
1464 * the open context) or reset its contents to
1467 dbuf_fix_old_data(db, txg);
1470 /* clear the contents if its cached */
1471 if (db->db_state == DB_CACHED) {
1472 ASSERT(db->db.db_data != NULL);
1473 arc_release(db->db_buf, db);
1474 bzero(db->db.db_data, db->db.db_size);
1475 arc_buf_freeze(db->db_buf);
1478 mutex_exit(&db->db_mtx);
1481 kmem_free(db_search, sizeof (dmu_buf_impl_t));
1482 mutex_exit(&dn->dn_dbufs_mtx);
1486 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
1488 arc_buf_t *buf, *obuf;
1489 int osize = db->db.db_size;
1490 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1493 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1498 /* XXX does *this* func really need the lock? */
1499 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1502 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1503 * is OK, because there can be no other references to the db
1504 * when we are changing its size, so no concurrent DB_FILL can
1508 * XXX we should be doing a dbuf_read, checking the return
1509 * value and returning that up to our callers
1511 dmu_buf_will_dirty(&db->db, tx);
1513 /* create the data buffer for the new block */
1514 buf = arc_alloc_buf(dn->dn_objset->os_spa, db, type, size);
1516 /* copy old block data to the new block */
1518 bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
1519 /* zero the remainder */
1521 bzero((uint8_t *)buf->b_data + osize, size - osize);
1523 mutex_enter(&db->db_mtx);
1524 dbuf_set_data(db, buf);
1525 arc_buf_destroy(obuf, db);
1526 db->db.db_size = size;
1528 if (db->db_level == 0) {
1529 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1530 db->db_last_dirty->dt.dl.dr_data = buf;
1532 mutex_exit(&db->db_mtx);
1534 dmu_objset_willuse_space(dn->dn_objset, size - osize, tx);
1539 dbuf_release_bp(dmu_buf_impl_t *db)
1541 ASSERTV(objset_t *os = db->db_objset);
1543 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1544 ASSERT(arc_released(os->os_phys_buf) ||
1545 list_link_active(&os->os_dsl_dataset->ds_synced_link));
1546 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1548 (void) arc_release(db->db_buf, db);
1552 * We already have a dirty record for this TXG, and we are being
1556 dbuf_redirty(dbuf_dirty_record_t *dr)
1558 dmu_buf_impl_t *db = dr->dr_dbuf;
1560 ASSERT(MUTEX_HELD(&db->db_mtx));
1562 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1564 * If this buffer has already been written out,
1565 * we now need to reset its state.
1567 dbuf_unoverride(dr);
1568 if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1569 db->db_state != DB_NOFILL) {
1570 /* Already released on initial dirty, so just thaw. */
1571 ASSERT(arc_released(db->db_buf));
1572 arc_buf_thaw(db->db_buf);
1577 dbuf_dirty_record_t *
1578 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1582 dbuf_dirty_record_t **drp, *dr;
1583 int drop_struct_lock = FALSE;
1584 int txgoff = tx->tx_txg & TXG_MASK;
1586 ASSERT(tx->tx_txg != 0);
1587 ASSERT(!refcount_is_zero(&db->db_holds));
1588 DMU_TX_DIRTY_BUF(tx, db);
1593 * Shouldn't dirty a regular buffer in syncing context. Private
1594 * objects may be dirtied in syncing context, but only if they
1595 * were already pre-dirtied in open context.
1598 if (dn->dn_objset->os_dsl_dataset != NULL) {
1599 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1602 ASSERT(!dmu_tx_is_syncing(tx) ||
1603 BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1604 DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1605 dn->dn_objset->os_dsl_dataset == NULL);
1606 if (dn->dn_objset->os_dsl_dataset != NULL)
1607 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, FTAG);
1610 * We make this assert for private objects as well, but after we
1611 * check if we're already dirty. They are allowed to re-dirty
1612 * in syncing context.
1614 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1615 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1616 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1618 mutex_enter(&db->db_mtx);
1620 * XXX make this true for indirects too? The problem is that
1621 * transactions created with dmu_tx_create_assigned() from
1622 * syncing context don't bother holding ahead.
1624 ASSERT(db->db_level != 0 ||
1625 db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1626 db->db_state == DB_NOFILL);
1628 mutex_enter(&dn->dn_mtx);
1630 * Don't set dirtyctx to SYNC if we're just modifying this as we
1631 * initialize the objset.
1633 if (dn->dn_dirtyctx == DN_UNDIRTIED) {
1634 if (dn->dn_objset->os_dsl_dataset != NULL) {
1635 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1638 if (!BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1639 dn->dn_dirtyctx = (dmu_tx_is_syncing(tx) ?
1640 DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1641 ASSERT(dn->dn_dirtyctx_firstset == NULL);
1642 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1644 if (dn->dn_objset->os_dsl_dataset != NULL) {
1645 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1649 mutex_exit(&dn->dn_mtx);
1651 if (db->db_blkid == DMU_SPILL_BLKID)
1652 dn->dn_have_spill = B_TRUE;
1655 * If this buffer is already dirty, we're done.
1657 drp = &db->db_last_dirty;
1658 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1659 db->db.db_object == DMU_META_DNODE_OBJECT);
1660 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1662 if (dr && dr->dr_txg == tx->tx_txg) {
1666 mutex_exit(&db->db_mtx);
1671 * Only valid if not already dirty.
1673 ASSERT(dn->dn_object == 0 ||
1674 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1675 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1677 ASSERT3U(dn->dn_nlevels, >, db->db_level);
1680 * We should only be dirtying in syncing context if it's the
1681 * mos or we're initializing the os or it's a special object.
1682 * However, we are allowed to dirty in syncing context provided
1683 * we already dirtied it in open context. Hence we must make
1684 * this assertion only if we're not already dirty.
1687 VERIFY3U(tx->tx_txg, <=, spa_final_dirty_txg(os->os_spa));
1689 if (dn->dn_objset->os_dsl_dataset != NULL)
1690 rrw_enter(&os->os_dsl_dataset->ds_bp_rwlock, RW_READER, FTAG);
1691 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1692 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1693 if (dn->dn_objset->os_dsl_dataset != NULL)
1694 rrw_exit(&os->os_dsl_dataset->ds_bp_rwlock, FTAG);
1696 ASSERT(db->db.db_size != 0);
1698 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1700 if (db->db_blkid != DMU_BONUS_BLKID) {
1701 dmu_objset_willuse_space(os, db->db.db_size, tx);
1705 * If this buffer is dirty in an old transaction group we need
1706 * to make a copy of it so that the changes we make in this
1707 * transaction group won't leak out when we sync the older txg.
1709 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1710 list_link_init(&dr->dr_dirty_node);
1711 if (db->db_level == 0) {
1712 void *data_old = db->db_buf;
1714 if (db->db_state != DB_NOFILL) {
1715 if (db->db_blkid == DMU_BONUS_BLKID) {
1716 dbuf_fix_old_data(db, tx->tx_txg);
1717 data_old = db->db.db_data;
1718 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1720 * Release the data buffer from the cache so
1721 * that we can modify it without impacting
1722 * possible other users of this cached data
1723 * block. Note that indirect blocks and
1724 * private objects are not released until the
1725 * syncing state (since they are only modified
1728 arc_release(db->db_buf, db);
1729 dbuf_fix_old_data(db, tx->tx_txg);
1730 data_old = db->db_buf;
1732 ASSERT(data_old != NULL);
1734 dr->dt.dl.dr_data = data_old;
1736 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_NOLOCKDEP, NULL);
1737 list_create(&dr->dt.di.dr_children,
1738 sizeof (dbuf_dirty_record_t),
1739 offsetof(dbuf_dirty_record_t, dr_dirty_node));
1741 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL)
1742 dr->dr_accounted = db->db.db_size;
1744 dr->dr_txg = tx->tx_txg;
1749 * We could have been freed_in_flight between the dbuf_noread
1750 * and dbuf_dirty. We win, as though the dbuf_noread() had
1751 * happened after the free.
1753 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1754 db->db_blkid != DMU_SPILL_BLKID) {
1755 mutex_enter(&dn->dn_mtx);
1756 if (dn->dn_free_ranges[txgoff] != NULL) {
1757 range_tree_clear(dn->dn_free_ranges[txgoff],
1760 mutex_exit(&dn->dn_mtx);
1761 db->db_freed_in_flight = FALSE;
1765 * This buffer is now part of this txg
1767 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1768 db->db_dirtycnt += 1;
1769 ASSERT3U(db->db_dirtycnt, <=, 3);
1771 mutex_exit(&db->db_mtx);
1773 if (db->db_blkid == DMU_BONUS_BLKID ||
1774 db->db_blkid == DMU_SPILL_BLKID) {
1775 mutex_enter(&dn->dn_mtx);
1776 ASSERT(!list_link_active(&dr->dr_dirty_node));
1777 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1778 mutex_exit(&dn->dn_mtx);
1779 dnode_setdirty(dn, tx);
1785 * The dn_struct_rwlock prevents db_blkptr from changing
1786 * due to a write from syncing context completing
1787 * while we are running, so we want to acquire it before
1788 * looking at db_blkptr.
1790 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1791 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1792 drop_struct_lock = TRUE;
1796 * We need to hold the dn_struct_rwlock to make this assertion,
1797 * because it protects dn_phys / dn_next_nlevels from changing.
1799 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1800 dn->dn_phys->dn_nlevels > db->db_level ||
1801 dn->dn_next_nlevels[txgoff] > db->db_level ||
1802 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1803 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1806 * If we are overwriting a dedup BP, then unless it is snapshotted,
1807 * when we get to syncing context we will need to decrement its
1808 * refcount in the DDT. Prefetch the relevant DDT block so that
1809 * syncing context won't have to wait for the i/o.
1811 ddt_prefetch(os->os_spa, db->db_blkptr);
1813 if (db->db_level == 0) {
1814 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1815 ASSERT(dn->dn_maxblkid >= db->db_blkid);
1818 if (db->db_level+1 < dn->dn_nlevels) {
1819 dmu_buf_impl_t *parent = db->db_parent;
1820 dbuf_dirty_record_t *di;
1821 int parent_held = FALSE;
1823 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1824 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1826 parent = dbuf_hold_level(dn, db->db_level+1,
1827 db->db_blkid >> epbs, FTAG);
1828 ASSERT(parent != NULL);
1831 if (drop_struct_lock)
1832 rw_exit(&dn->dn_struct_rwlock);
1833 ASSERT3U(db->db_level+1, ==, parent->db_level);
1834 di = dbuf_dirty(parent, tx);
1836 dbuf_rele(parent, FTAG);
1838 mutex_enter(&db->db_mtx);
1840 * Since we've dropped the mutex, it's possible that
1841 * dbuf_undirty() might have changed this out from under us.
1843 if (db->db_last_dirty == dr ||
1844 dn->dn_object == DMU_META_DNODE_OBJECT) {
1845 mutex_enter(&di->dt.di.dr_mtx);
1846 ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1847 ASSERT(!list_link_active(&dr->dr_dirty_node));
1848 list_insert_tail(&di->dt.di.dr_children, dr);
1849 mutex_exit(&di->dt.di.dr_mtx);
1852 mutex_exit(&db->db_mtx);
1854 ASSERT(db->db_level+1 == dn->dn_nlevels);
1855 ASSERT(db->db_blkid < dn->dn_nblkptr);
1856 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1857 mutex_enter(&dn->dn_mtx);
1858 ASSERT(!list_link_active(&dr->dr_dirty_node));
1859 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1860 mutex_exit(&dn->dn_mtx);
1861 if (drop_struct_lock)
1862 rw_exit(&dn->dn_struct_rwlock);
1865 dnode_setdirty(dn, tx);
1871 * Undirty a buffer in the transaction group referenced by the given
1872 * transaction. Return whether this evicted the dbuf.
1875 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1878 uint64_t txg = tx->tx_txg;
1879 dbuf_dirty_record_t *dr, **drp;
1884 * Due to our use of dn_nlevels below, this can only be called
1885 * in open context, unless we are operating on the MOS.
1886 * From syncing context, dn_nlevels may be different from the
1887 * dn_nlevels used when dbuf was dirtied.
1889 ASSERT(db->db_objset ==
1890 dmu_objset_pool(db->db_objset)->dp_meta_objset ||
1891 txg != spa_syncing_txg(dmu_objset_spa(db->db_objset)));
1892 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1893 ASSERT0(db->db_level);
1894 ASSERT(MUTEX_HELD(&db->db_mtx));
1897 * If this buffer is not dirty, we're done.
1899 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1900 if (dr->dr_txg <= txg)
1902 if (dr == NULL || dr->dr_txg < txg)
1904 ASSERT(dr->dr_txg == txg);
1905 ASSERT(dr->dr_dbuf == db);
1910 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1912 ASSERT(db->db.db_size != 0);
1914 dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset),
1915 dr->dr_accounted, txg);
1920 * Note that there are three places in dbuf_dirty()
1921 * where this dirty record may be put on a list.
1922 * Make sure to do a list_remove corresponding to
1923 * every one of those list_insert calls.
1925 if (dr->dr_parent) {
1926 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1927 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1928 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1929 } else if (db->db_blkid == DMU_SPILL_BLKID ||
1930 db->db_level + 1 == dn->dn_nlevels) {
1931 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1932 mutex_enter(&dn->dn_mtx);
1933 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1934 mutex_exit(&dn->dn_mtx);
1938 if (db->db_state != DB_NOFILL) {
1939 dbuf_unoverride(dr);
1941 ASSERT(db->db_buf != NULL);
1942 ASSERT(dr->dt.dl.dr_data != NULL);
1943 if (dr->dt.dl.dr_data != db->db_buf)
1944 arc_buf_destroy(dr->dt.dl.dr_data, db);
1947 kmem_free(dr, sizeof (dbuf_dirty_record_t));
1949 ASSERT(db->db_dirtycnt > 0);
1950 db->db_dirtycnt -= 1;
1952 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1953 ASSERT(db->db_state == DB_NOFILL || arc_released(db->db_buf));
1962 dmu_buf_will_dirty_impl(dmu_buf_t *db_fake, int flags, dmu_tx_t *tx)
1964 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1965 dbuf_dirty_record_t *dr;
1967 ASSERT(tx->tx_txg != 0);
1968 ASSERT(!refcount_is_zero(&db->db_holds));
1971 * Quick check for dirtyness. For already dirty blocks, this
1972 * reduces runtime of this function by >90%, and overall performance
1973 * by 50% for some workloads (e.g. file deletion with indirect blocks
1976 mutex_enter(&db->db_mtx);
1978 for (dr = db->db_last_dirty;
1979 dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) {
1981 * It's possible that it is already dirty but not cached,
1982 * because there are some calls to dbuf_dirty() that don't
1983 * go through dmu_buf_will_dirty().
1985 if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) {
1986 /* This dbuf is already dirty and cached. */
1988 mutex_exit(&db->db_mtx);
1992 mutex_exit(&db->db_mtx);
1995 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
1996 flags |= DB_RF_HAVESTRUCT;
1998 (void) dbuf_read(db, NULL, flags);
1999 (void) dbuf_dirty(db, tx);
2003 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
2005 dmu_buf_will_dirty_impl(db_fake,
2006 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH, tx);
2010 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
2012 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2014 db->db_state = DB_NOFILL;
2016 dmu_buf_will_fill(db_fake, tx);
2020 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
2022 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2024 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2025 ASSERT(tx->tx_txg != 0);
2026 ASSERT(db->db_level == 0);
2027 ASSERT(!refcount_is_zero(&db->db_holds));
2029 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
2030 dmu_tx_private_ok(tx));
2033 (void) dbuf_dirty(db, tx);
2037 * This function is effectively the same as dmu_buf_will_dirty(), but
2038 * indicates the caller expects raw encrypted data in the db. It will
2039 * also set the raw flag on the created dirty record.
2042 dmu_buf_will_change_crypt_params(dmu_buf_t *db_fake, dmu_tx_t *tx)
2044 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2045 dbuf_dirty_record_t *dr;
2047 dmu_buf_will_dirty_impl(db_fake,
2048 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_NO_DECRYPT, tx);
2050 dr = db->db_last_dirty;
2051 while (dr != NULL && dr->dr_txg > tx->tx_txg)
2054 ASSERT3P(dr, !=, NULL);
2055 ASSERT3U(dr->dr_txg, ==, tx->tx_txg);
2056 dr->dt.dl.dr_raw = B_TRUE;
2059 #pragma weak dmu_buf_fill_done = dbuf_fill_done
2062 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
2064 mutex_enter(&db->db_mtx);
2067 if (db->db_state == DB_FILL) {
2068 if (db->db_level == 0 && db->db_freed_in_flight) {
2069 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2070 /* we were freed while filling */
2071 /* XXX dbuf_undirty? */
2072 bzero(db->db.db_data, db->db.db_size);
2073 db->db_freed_in_flight = FALSE;
2075 db->db_state = DB_CACHED;
2076 cv_broadcast(&db->db_changed);
2078 mutex_exit(&db->db_mtx);
2082 dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data,
2083 bp_embedded_type_t etype, enum zio_compress comp,
2084 int uncompressed_size, int compressed_size, int byteorder,
2087 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2088 struct dirty_leaf *dl;
2089 dmu_object_type_t type;
2091 if (etype == BP_EMBEDDED_TYPE_DATA) {
2092 ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset),
2093 SPA_FEATURE_EMBEDDED_DATA));
2097 type = DB_DNODE(db)->dn_type;
2100 ASSERT0(db->db_level);
2101 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2103 dmu_buf_will_not_fill(dbuf, tx);
2105 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
2106 dl = &db->db_last_dirty->dt.dl;
2107 encode_embedded_bp_compressed(&dl->dr_overridden_by,
2108 data, comp, uncompressed_size, compressed_size);
2109 BPE_SET_ETYPE(&dl->dr_overridden_by, etype);
2110 BP_SET_TYPE(&dl->dr_overridden_by, type);
2111 BP_SET_LEVEL(&dl->dr_overridden_by, 0);
2112 BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder);
2114 dl->dr_override_state = DR_OVERRIDDEN;
2115 dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg;
2119 * Directly assign a provided arc buf to a given dbuf if it's not referenced
2120 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
2123 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
2125 ASSERT(!refcount_is_zero(&db->db_holds));
2126 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2127 ASSERT(db->db_level == 0);
2128 ASSERT3U(dbuf_is_metadata(db), ==, arc_is_metadata(buf));
2129 ASSERT(buf != NULL);
2130 ASSERT(arc_buf_lsize(buf) == db->db.db_size);
2131 ASSERT(tx->tx_txg != 0);
2133 arc_return_buf(buf, db);
2134 ASSERT(arc_released(buf));
2136 mutex_enter(&db->db_mtx);
2138 while (db->db_state == DB_READ || db->db_state == DB_FILL)
2139 cv_wait(&db->db_changed, &db->db_mtx);
2141 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
2143 if (db->db_state == DB_CACHED &&
2144 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
2145 mutex_exit(&db->db_mtx);
2146 (void) dbuf_dirty(db, tx);
2147 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
2148 arc_buf_destroy(buf, db);
2149 xuio_stat_wbuf_copied();
2153 xuio_stat_wbuf_nocopy();
2154 if (db->db_state == DB_CACHED) {
2155 dbuf_dirty_record_t *dr = db->db_last_dirty;
2157 ASSERT(db->db_buf != NULL);
2158 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
2159 ASSERT(dr->dt.dl.dr_data == db->db_buf);
2160 if (!arc_released(db->db_buf)) {
2161 ASSERT(dr->dt.dl.dr_override_state ==
2163 arc_release(db->db_buf, db);
2165 dr->dt.dl.dr_data = buf;
2166 arc_buf_destroy(db->db_buf, db);
2167 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
2168 arc_release(db->db_buf, db);
2169 arc_buf_destroy(db->db_buf, db);
2173 ASSERT(db->db_buf == NULL);
2174 dbuf_set_data(db, buf);
2175 db->db_state = DB_FILL;
2176 mutex_exit(&db->db_mtx);
2177 (void) dbuf_dirty(db, tx);
2178 dmu_buf_fill_done(&db->db, tx);
2182 dbuf_destroy(dmu_buf_impl_t *db)
2185 dmu_buf_impl_t *parent = db->db_parent;
2186 dmu_buf_impl_t *dndb;
2188 ASSERT(MUTEX_HELD(&db->db_mtx));
2189 ASSERT(refcount_is_zero(&db->db_holds));
2191 if (db->db_buf != NULL) {
2192 arc_buf_destroy(db->db_buf, db);
2196 if (db->db_blkid == DMU_BONUS_BLKID) {
2197 int slots = DB_DNODE(db)->dn_num_slots;
2198 int bonuslen = DN_SLOTS_TO_BONUSLEN(slots);
2199 if (db->db.db_data != NULL) {
2200 kmem_free(db->db.db_data, bonuslen);
2201 arc_space_return(bonuslen, ARC_SPACE_BONUS);
2202 db->db_state = DB_UNCACHED;
2206 dbuf_clear_data(db);
2208 if (multilist_link_active(&db->db_cache_link)) {
2209 multilist_remove(dbuf_cache, db);
2210 (void) refcount_remove_many(&dbuf_cache_size,
2211 db->db.db_size, db);
2214 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
2215 ASSERT(db->db_data_pending == NULL);
2217 db->db_state = DB_EVICTING;
2218 db->db_blkptr = NULL;
2221 * Now that db_state is DB_EVICTING, nobody else can find this via
2222 * the hash table. We can now drop db_mtx, which allows us to
2223 * acquire the dn_dbufs_mtx.
2225 mutex_exit(&db->db_mtx);
2230 if (db->db_blkid != DMU_BONUS_BLKID) {
2231 boolean_t needlock = !MUTEX_HELD(&dn->dn_dbufs_mtx);
2233 mutex_enter(&dn->dn_dbufs_mtx);
2234 avl_remove(&dn->dn_dbufs, db);
2235 atomic_dec_32(&dn->dn_dbufs_count);
2239 mutex_exit(&dn->dn_dbufs_mtx);
2241 * Decrementing the dbuf count means that the hold corresponding
2242 * to the removed dbuf is no longer discounted in dnode_move(),
2243 * so the dnode cannot be moved until after we release the hold.
2244 * The membar_producer() ensures visibility of the decremented
2245 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
2249 db->db_dnode_handle = NULL;
2251 dbuf_hash_remove(db);
2256 ASSERT(refcount_is_zero(&db->db_holds));
2258 db->db_parent = NULL;
2260 ASSERT(db->db_buf == NULL);
2261 ASSERT(db->db.db_data == NULL);
2262 ASSERT(db->db_hash_next == NULL);
2263 ASSERT(db->db_blkptr == NULL);
2264 ASSERT(db->db_data_pending == NULL);
2265 ASSERT(!multilist_link_active(&db->db_cache_link));
2267 kmem_cache_free(dbuf_kmem_cache, db);
2268 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
2271 * If this dbuf is referenced from an indirect dbuf,
2272 * decrement the ref count on the indirect dbuf.
2274 if (parent && parent != dndb)
2275 dbuf_rele(parent, db);
2279 * Note: While bpp will always be updated if the function returns success,
2280 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
2281 * this happens when the dnode is the meta-dnode, or a userused or groupused
2284 __attribute__((always_inline))
2286 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
2287 dmu_buf_impl_t **parentp, blkptr_t **bpp, struct dbuf_hold_impl_data *dh)
2294 ASSERT(blkid != DMU_BONUS_BLKID);
2296 if (blkid == DMU_SPILL_BLKID) {
2297 mutex_enter(&dn->dn_mtx);
2298 if (dn->dn_have_spill &&
2299 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
2300 *bpp = DN_SPILL_BLKPTR(dn->dn_phys);
2303 dbuf_add_ref(dn->dn_dbuf, NULL);
2304 *parentp = dn->dn_dbuf;
2305 mutex_exit(&dn->dn_mtx);
2310 (dn->dn_phys->dn_nlevels == 0) ? 1 : dn->dn_phys->dn_nlevels;
2311 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2313 ASSERT3U(level * epbs, <, 64);
2314 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2316 * This assertion shouldn't trip as long as the max indirect block size
2317 * is less than 1M. The reason for this is that up to that point,
2318 * the number of levels required to address an entire object with blocks
2319 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In
2320 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55
2321 * (i.e. we can address the entire object), objects will all use at most
2322 * N-1 levels and the assertion won't overflow. However, once epbs is
2323 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be
2324 * enough to address an entire object, so objects will have 5 levels,
2325 * but then this assertion will overflow.
2327 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we
2328 * need to redo this logic to handle overflows.
2330 ASSERT(level >= nlevels ||
2331 ((nlevels - level - 1) * epbs) +
2332 highbit64(dn->dn_phys->dn_nblkptr) <= 64);
2333 if (level >= nlevels ||
2334 blkid >= ((uint64_t)dn->dn_phys->dn_nblkptr <<
2335 ((nlevels - level - 1) * epbs)) ||
2337 blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
2338 /* the buffer has no parent yet */
2339 return (SET_ERROR(ENOENT));
2340 } else if (level < nlevels-1) {
2341 /* this block is referenced from an indirect block */
2344 err = dbuf_hold_impl(dn, level+1,
2345 blkid >> epbs, fail_sparse, FALSE, NULL, parentp);
2347 __dbuf_hold_impl_init(dh + 1, dn, dh->dh_level + 1,
2348 blkid >> epbs, fail_sparse, FALSE, NULL,
2349 parentp, dh->dh_depth + 1);
2350 err = __dbuf_hold_impl(dh + 1);
2354 err = dbuf_read(*parentp, NULL,
2355 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
2357 dbuf_rele(*parentp, NULL);
2361 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
2362 (blkid & ((1ULL << epbs) - 1));
2363 if (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))
2364 ASSERT(BP_IS_HOLE(*bpp));
2367 /* the block is referenced from the dnode */
2368 ASSERT3U(level, ==, nlevels-1);
2369 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
2370 blkid < dn->dn_phys->dn_nblkptr);
2372 dbuf_add_ref(dn->dn_dbuf, NULL);
2373 *parentp = dn->dn_dbuf;
2375 *bpp = &dn->dn_phys->dn_blkptr[blkid];
2380 static dmu_buf_impl_t *
2381 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
2382 dmu_buf_impl_t *parent, blkptr_t *blkptr)
2384 objset_t *os = dn->dn_objset;
2385 dmu_buf_impl_t *db, *odb;
2387 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2388 ASSERT(dn->dn_type != DMU_OT_NONE);
2390 db = kmem_cache_alloc(dbuf_kmem_cache, KM_SLEEP);
2393 db->db.db_object = dn->dn_object;
2394 db->db_level = level;
2395 db->db_blkid = blkid;
2396 db->db_last_dirty = NULL;
2397 db->db_dirtycnt = 0;
2398 db->db_dnode_handle = dn->dn_handle;
2399 db->db_parent = parent;
2400 db->db_blkptr = blkptr;
2403 db->db_user_immediate_evict = FALSE;
2404 db->db_freed_in_flight = FALSE;
2405 db->db_pending_evict = FALSE;
2407 if (blkid == DMU_BONUS_BLKID) {
2408 ASSERT3P(parent, ==, dn->dn_dbuf);
2409 db->db.db_size = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
2410 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
2411 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
2412 db->db.db_offset = DMU_BONUS_BLKID;
2413 db->db_state = DB_UNCACHED;
2414 /* the bonus dbuf is not placed in the hash table */
2415 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
2417 } else if (blkid == DMU_SPILL_BLKID) {
2418 db->db.db_size = (blkptr != NULL) ?
2419 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
2420 db->db.db_offset = 0;
2423 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
2424 db->db.db_size = blocksize;
2425 db->db.db_offset = db->db_blkid * blocksize;
2429 * Hold the dn_dbufs_mtx while we get the new dbuf
2430 * in the hash table *and* added to the dbufs list.
2431 * This prevents a possible deadlock with someone
2432 * trying to look up this dbuf before its added to the
2435 mutex_enter(&dn->dn_dbufs_mtx);
2436 db->db_state = DB_EVICTING;
2437 if ((odb = dbuf_hash_insert(db)) != NULL) {
2438 /* someone else inserted it first */
2439 kmem_cache_free(dbuf_kmem_cache, db);
2440 mutex_exit(&dn->dn_dbufs_mtx);
2443 avl_add(&dn->dn_dbufs, db);
2445 db->db_state = DB_UNCACHED;
2446 mutex_exit(&dn->dn_dbufs_mtx);
2447 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
2449 if (parent && parent != dn->dn_dbuf)
2450 dbuf_add_ref(parent, db);
2452 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
2453 refcount_count(&dn->dn_holds) > 0);
2454 (void) refcount_add(&dn->dn_holds, db);
2455 atomic_inc_32(&dn->dn_dbufs_count);
2457 dprintf_dbuf(db, "db=%p\n", db);
2462 typedef struct dbuf_prefetch_arg {
2463 spa_t *dpa_spa; /* The spa to issue the prefetch in. */
2464 zbookmark_phys_t dpa_zb; /* The target block to prefetch. */
2465 int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */
2466 int dpa_curlevel; /* The current level that we're reading */
2467 dnode_t *dpa_dnode; /* The dnode associated with the prefetch */
2468 zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */
2469 zio_t *dpa_zio; /* The parent zio_t for all prefetches. */
2470 arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */
2471 } dbuf_prefetch_arg_t;
2474 * Actually issue the prefetch read for the block given.
2477 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp)
2480 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
2483 aflags = dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
2485 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2486 ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level);
2487 ASSERT(dpa->dpa_zio != NULL);
2488 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL,
2489 dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2490 &aflags, &dpa->dpa_zb);
2494 * Called when an indirect block above our prefetch target is read in. This
2495 * will either read in the next indirect block down the tree or issue the actual
2496 * prefetch if the next block down is our target.
2499 dbuf_prefetch_indirect_done(zio_t *zio, int err, arc_buf_t *abuf, void *private)
2501 dbuf_prefetch_arg_t *dpa = private;
2505 ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel);
2506 ASSERT3S(dpa->dpa_curlevel, >, 0);
2509 * The dpa_dnode is only valid if we are called with a NULL
2510 * zio. This indicates that the arc_read() returned without
2511 * first calling zio_read() to issue a physical read. Once
2512 * a physical read is made the dpa_dnode must be invalidated
2513 * as the locks guarding it may have been dropped. If the
2514 * dpa_dnode is still valid, then we want to add it to the dbuf
2515 * cache. To do so, we must hold the dbuf associated with the block
2516 * we just prefetched, read its contents so that we associate it
2517 * with an arc_buf_t, and then release it.
2520 ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel);
2521 if (zio->io_flags & ZIO_FLAG_RAW_COMPRESS) {
2522 ASSERT3U(BP_GET_PSIZE(zio->io_bp), ==, zio->io_size);
2524 ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size);
2526 ASSERT3P(zio->io_spa, ==, dpa->dpa_spa);
2528 dpa->dpa_dnode = NULL;
2529 } else if (dpa->dpa_dnode != NULL) {
2530 uint64_t curblkid = dpa->dpa_zb.zb_blkid >>
2531 (dpa->dpa_epbs * (dpa->dpa_curlevel -
2532 dpa->dpa_zb.zb_level));
2533 dmu_buf_impl_t *db = dbuf_hold_level(dpa->dpa_dnode,
2534 dpa->dpa_curlevel, curblkid, FTAG);
2535 (void) dbuf_read(db, NULL,
2536 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_HAVESTRUCT);
2537 dbuf_rele(db, FTAG);
2540 dpa->dpa_curlevel--;
2542 nextblkid = dpa->dpa_zb.zb_blkid >>
2543 (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level));
2544 bp = ((blkptr_t *)abuf->b_data) +
2545 P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs);
2546 if (BP_IS_HOLE(bp) || err != 0) {
2547 kmem_free(dpa, sizeof (*dpa));
2548 } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) {
2549 ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid);
2550 dbuf_issue_final_prefetch(dpa, bp);
2551 kmem_free(dpa, sizeof (*dpa));
2553 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2554 zbookmark_phys_t zb;
2556 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2558 SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset,
2559 dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid);
2561 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2562 bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio,
2563 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2567 arc_buf_destroy(abuf, private);
2571 * Issue prefetch reads for the given block on the given level. If the indirect
2572 * blocks above that block are not in memory, we will read them in
2573 * asynchronously. As a result, this call never blocks waiting for a read to
2574 * complete. Note that the prefetch might fail if the dataset is encrypted and
2575 * the encryption key is unmapped before the IO completes.
2578 dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio,
2582 int epbs, nlevels, curlevel;
2586 dbuf_prefetch_arg_t *dpa;
2589 ASSERT(blkid != DMU_BONUS_BLKID);
2590 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2592 if (blkid > dn->dn_maxblkid)
2595 if (dnode_block_freed(dn, blkid))
2599 * This dnode hasn't been written to disk yet, so there's nothing to
2602 nlevels = dn->dn_phys->dn_nlevels;
2603 if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0)
2606 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2607 if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level))
2610 db = dbuf_find(dn->dn_objset, dn->dn_object,
2613 mutex_exit(&db->db_mtx);
2615 * This dbuf already exists. It is either CACHED, or
2616 * (we assume) about to be read or filled.
2622 * Find the closest ancestor (indirect block) of the target block
2623 * that is present in the cache. In this indirect block, we will
2624 * find the bp that is at curlevel, curblkid.
2628 while (curlevel < nlevels - 1) {
2629 int parent_level = curlevel + 1;
2630 uint64_t parent_blkid = curblkid >> epbs;
2633 if (dbuf_hold_impl(dn, parent_level, parent_blkid,
2634 FALSE, TRUE, FTAG, &db) == 0) {
2635 blkptr_t *bpp = db->db_buf->b_data;
2636 bp = bpp[P2PHASE(curblkid, 1 << epbs)];
2637 dbuf_rele(db, FTAG);
2641 curlevel = parent_level;
2642 curblkid = parent_blkid;
2645 if (curlevel == nlevels - 1) {
2646 /* No cached indirect blocks found. */
2647 ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr);
2648 bp = dn->dn_phys->dn_blkptr[curblkid];
2650 if (BP_IS_HOLE(&bp))
2653 ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp));
2655 pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL,
2658 dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP);
2659 ds = dn->dn_objset->os_dsl_dataset;
2660 SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2661 dn->dn_object, level, blkid);
2662 dpa->dpa_curlevel = curlevel;
2663 dpa->dpa_prio = prio;
2664 dpa->dpa_aflags = aflags;
2665 dpa->dpa_spa = dn->dn_objset->os_spa;
2666 dpa->dpa_dnode = dn;
2667 dpa->dpa_epbs = epbs;
2671 * If we have the indirect just above us, no need to do the asynchronous
2672 * prefetch chain; we'll just run the last step ourselves. If we're at
2673 * a higher level, though, we want to issue the prefetches for all the
2674 * indirect blocks asynchronously, so we can go on with whatever we were
2677 if (curlevel == level) {
2678 ASSERT3U(curblkid, ==, blkid);
2679 dbuf_issue_final_prefetch(dpa, &bp);
2680 kmem_free(dpa, sizeof (*dpa));
2682 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2683 zbookmark_phys_t zb;
2685 SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2686 dn->dn_object, curlevel, curblkid);
2687 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2688 &bp, dbuf_prefetch_indirect_done, dpa, prio,
2689 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2693 * We use pio here instead of dpa_zio since it's possible that
2694 * dpa may have already been freed.
2699 #define DBUF_HOLD_IMPL_MAX_DEPTH 20
2702 * Returns with db_holds incremented, and db_mtx not held.
2703 * Note: dn_struct_rwlock must be held.
2706 __dbuf_hold_impl(struct dbuf_hold_impl_data *dh)
2708 ASSERT3S(dh->dh_depth, <, DBUF_HOLD_IMPL_MAX_DEPTH);
2709 dh->dh_parent = NULL;
2711 ASSERT(dh->dh_blkid != DMU_BONUS_BLKID);
2712 ASSERT(RW_LOCK_HELD(&dh->dh_dn->dn_struct_rwlock));
2713 ASSERT3U(dh->dh_dn->dn_nlevels, >, dh->dh_level);
2715 *(dh->dh_dbp) = NULL;
2717 /* dbuf_find() returns with db_mtx held */
2718 dh->dh_db = dbuf_find(dh->dh_dn->dn_objset, dh->dh_dn->dn_object,
2719 dh->dh_level, dh->dh_blkid);
2721 if (dh->dh_db == NULL) {
2724 if (dh->dh_fail_uncached)
2725 return (SET_ERROR(ENOENT));
2727 ASSERT3P(dh->dh_parent, ==, NULL);
2728 dh->dh_err = dbuf_findbp(dh->dh_dn, dh->dh_level, dh->dh_blkid,
2729 dh->dh_fail_sparse, &dh->dh_parent, &dh->dh_bp, dh);
2730 if (dh->dh_fail_sparse) {
2731 if (dh->dh_err == 0 &&
2732 dh->dh_bp && BP_IS_HOLE(dh->dh_bp))
2733 dh->dh_err = SET_ERROR(ENOENT);
2736 dbuf_rele(dh->dh_parent, NULL);
2737 return (dh->dh_err);
2740 if (dh->dh_err && dh->dh_err != ENOENT)
2741 return (dh->dh_err);
2742 dh->dh_db = dbuf_create(dh->dh_dn, dh->dh_level, dh->dh_blkid,
2743 dh->dh_parent, dh->dh_bp);
2746 if (dh->dh_fail_uncached && dh->dh_db->db_state != DB_CACHED) {
2747 mutex_exit(&dh->dh_db->db_mtx);
2748 return (SET_ERROR(ENOENT));
2751 if (dh->dh_db->db_buf != NULL)
2752 ASSERT3P(dh->dh_db->db.db_data, ==, dh->dh_db->db_buf->b_data);
2754 ASSERT(dh->dh_db->db_buf == NULL || arc_referenced(dh->dh_db->db_buf));
2757 * If this buffer is currently syncing out, and we are are
2758 * still referencing it from db_data, we need to make a copy
2759 * of it in case we decide we want to dirty it again in this txg.
2761 if (dh->dh_db->db_level == 0 &&
2762 dh->dh_db->db_blkid != DMU_BONUS_BLKID &&
2763 dh->dh_dn->dn_object != DMU_META_DNODE_OBJECT &&
2764 dh->dh_db->db_state == DB_CACHED && dh->dh_db->db_data_pending) {
2765 dh->dh_dr = dh->dh_db->db_data_pending;
2767 if (dh->dh_dr->dt.dl.dr_data == dh->dh_db->db_buf) {
2768 dh->dh_type = DBUF_GET_BUFC_TYPE(dh->dh_db);
2770 dbuf_set_data(dh->dh_db,
2771 arc_alloc_buf(dh->dh_dn->dn_objset->os_spa,
2772 dh->dh_db, dh->dh_type, dh->dh_db->db.db_size));
2773 bcopy(dh->dh_dr->dt.dl.dr_data->b_data,
2774 dh->dh_db->db.db_data, dh->dh_db->db.db_size);
2778 if (multilist_link_active(&dh->dh_db->db_cache_link)) {
2779 ASSERT(refcount_is_zero(&dh->dh_db->db_holds));
2780 multilist_remove(dbuf_cache, dh->dh_db);
2781 (void) refcount_remove_many(&dbuf_cache_size,
2782 dh->dh_db->db.db_size, dh->dh_db);
2784 (void) refcount_add(&dh->dh_db->db_holds, dh->dh_tag);
2785 DBUF_VERIFY(dh->dh_db);
2786 mutex_exit(&dh->dh_db->db_mtx);
2788 /* NOTE: we can't rele the parent until after we drop the db_mtx */
2790 dbuf_rele(dh->dh_parent, NULL);
2792 ASSERT3P(DB_DNODE(dh->dh_db), ==, dh->dh_dn);
2793 ASSERT3U(dh->dh_db->db_blkid, ==, dh->dh_blkid);
2794 ASSERT3U(dh->dh_db->db_level, ==, dh->dh_level);
2795 *(dh->dh_dbp) = dh->dh_db;
2801 * The following code preserves the recursive function dbuf_hold_impl()
2802 * but moves the local variables AND function arguments to the heap to
2803 * minimize the stack frame size. Enough space is initially allocated
2804 * on the stack for 20 levels of recursion.
2807 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid,
2808 boolean_t fail_sparse, boolean_t fail_uncached,
2809 void *tag, dmu_buf_impl_t **dbp)
2811 struct dbuf_hold_impl_data *dh;
2814 dh = kmem_alloc(sizeof (struct dbuf_hold_impl_data) *
2815 DBUF_HOLD_IMPL_MAX_DEPTH, KM_SLEEP);
2816 __dbuf_hold_impl_init(dh, dn, level, blkid, fail_sparse,
2817 fail_uncached, tag, dbp, 0);
2819 error = __dbuf_hold_impl(dh);
2821 kmem_free(dh, sizeof (struct dbuf_hold_impl_data) *
2822 DBUF_HOLD_IMPL_MAX_DEPTH);
2828 __dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh,
2829 dnode_t *dn, uint8_t level, uint64_t blkid,
2830 boolean_t fail_sparse, boolean_t fail_uncached,
2831 void *tag, dmu_buf_impl_t **dbp, int depth)
2834 dh->dh_level = level;
2835 dh->dh_blkid = blkid;
2837 dh->dh_fail_sparse = fail_sparse;
2838 dh->dh_fail_uncached = fail_uncached;
2844 dh->dh_parent = NULL;
2850 dh->dh_depth = depth;
2854 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
2856 return (dbuf_hold_level(dn, 0, blkid, tag));
2860 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
2863 int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db);
2864 return (err ? NULL : db);
2868 dbuf_create_bonus(dnode_t *dn)
2870 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
2872 ASSERT(dn->dn_bonus == NULL);
2873 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
2877 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
2879 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2882 if (db->db_blkid != DMU_SPILL_BLKID)
2883 return (SET_ERROR(ENOTSUP));
2885 blksz = SPA_MINBLOCKSIZE;
2886 ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset)));
2887 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
2891 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2892 dbuf_new_size(db, blksz, tx);
2893 rw_exit(&dn->dn_struct_rwlock);
2900 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
2902 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
2905 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2907 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
2909 int64_t holds = refcount_add(&db->db_holds, tag);
2910 VERIFY3S(holds, >, 1);
2913 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
2915 dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid,
2918 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2919 dmu_buf_impl_t *found_db;
2920 boolean_t result = B_FALSE;
2922 if (blkid == DMU_BONUS_BLKID)
2923 found_db = dbuf_find_bonus(os, obj);
2925 found_db = dbuf_find(os, obj, 0, blkid);
2927 if (found_db != NULL) {
2928 if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) {
2929 (void) refcount_add(&db->db_holds, tag);
2932 mutex_exit(&found_db->db_mtx);
2938 * If you call dbuf_rele() you had better not be referencing the dnode handle
2939 * unless you have some other direct or indirect hold on the dnode. (An indirect
2940 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2941 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2942 * dnode's parent dbuf evicting its dnode handles.
2945 dbuf_rele(dmu_buf_impl_t *db, void *tag)
2947 mutex_enter(&db->db_mtx);
2948 dbuf_rele_and_unlock(db, tag);
2952 dmu_buf_rele(dmu_buf_t *db, void *tag)
2954 dbuf_rele((dmu_buf_impl_t *)db, tag);
2958 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2959 * db_dirtycnt and db_holds to be updated atomically.
2962 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2966 ASSERT(MUTEX_HELD(&db->db_mtx));
2970 * Remove the reference to the dbuf before removing its hold on the
2971 * dnode so we can guarantee in dnode_move() that a referenced bonus
2972 * buffer has a corresponding dnode hold.
2974 holds = refcount_remove(&db->db_holds, tag);
2978 * We can't freeze indirects if there is a possibility that they
2979 * may be modified in the current syncing context.
2981 if (db->db_buf != NULL &&
2982 holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) {
2983 arc_buf_freeze(db->db_buf);
2986 if (holds == db->db_dirtycnt &&
2987 db->db_level == 0 && db->db_user_immediate_evict)
2988 dbuf_evict_user(db);
2991 if (db->db_blkid == DMU_BONUS_BLKID) {
2993 boolean_t evict_dbuf = db->db_pending_evict;
2996 * If the dnode moves here, we cannot cross this
2997 * barrier until the move completes.
3002 atomic_dec_32(&dn->dn_dbufs_count);
3005 * Decrementing the dbuf count means that the bonus
3006 * buffer's dnode hold is no longer discounted in
3007 * dnode_move(). The dnode cannot move until after
3008 * the dnode_rele() below.
3013 * Do not reference db after its lock is dropped.
3014 * Another thread may evict it.
3016 mutex_exit(&db->db_mtx);
3019 dnode_evict_bonus(dn);
3022 } else if (db->db_buf == NULL) {
3024 * This is a special case: we never associated this
3025 * dbuf with any data allocated from the ARC.
3027 ASSERT(db->db_state == DB_UNCACHED ||
3028 db->db_state == DB_NOFILL);
3030 } else if (arc_released(db->db_buf)) {
3032 * This dbuf has anonymous data associated with it.
3036 boolean_t do_arc_evict = B_FALSE;
3038 spa_t *spa = dmu_objset_spa(db->db_objset);
3040 if (!DBUF_IS_CACHEABLE(db) &&
3041 db->db_blkptr != NULL &&
3042 !BP_IS_HOLE(db->db_blkptr) &&
3043 !BP_IS_EMBEDDED(db->db_blkptr)) {
3044 do_arc_evict = B_TRUE;
3045 bp = *db->db_blkptr;
3048 if (!DBUF_IS_CACHEABLE(db) ||
3049 db->db_pending_evict) {
3051 } else if (!multilist_link_active(&db->db_cache_link)) {
3052 multilist_insert(dbuf_cache, db);
3053 (void) refcount_add_many(&dbuf_cache_size,
3054 db->db.db_size, db);
3055 mutex_exit(&db->db_mtx);
3057 dbuf_evict_notify();
3061 arc_freed(spa, &bp);
3064 mutex_exit(&db->db_mtx);
3069 #pragma weak dmu_buf_refcount = dbuf_refcount
3071 dbuf_refcount(dmu_buf_impl_t *db)
3073 return (refcount_count(&db->db_holds));
3077 dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user,
3078 dmu_buf_user_t *new_user)
3080 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3082 mutex_enter(&db->db_mtx);
3083 dbuf_verify_user(db, DBVU_NOT_EVICTING);
3084 if (db->db_user == old_user)
3085 db->db_user = new_user;
3087 old_user = db->db_user;
3088 dbuf_verify_user(db, DBVU_NOT_EVICTING);
3089 mutex_exit(&db->db_mtx);
3095 dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
3097 return (dmu_buf_replace_user(db_fake, NULL, user));
3101 dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user)
3103 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3105 db->db_user_immediate_evict = TRUE;
3106 return (dmu_buf_set_user(db_fake, user));
3110 dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
3112 return (dmu_buf_replace_user(db_fake, user, NULL));
3116 dmu_buf_get_user(dmu_buf_t *db_fake)
3118 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3120 dbuf_verify_user(db, DBVU_NOT_EVICTING);
3121 return (db->db_user);
3125 dmu_buf_user_evict_wait()
3127 taskq_wait(dbu_evict_taskq);
3131 dmu_buf_get_blkptr(dmu_buf_t *db)
3133 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3134 return (dbi->db_blkptr);
3138 dmu_buf_get_objset(dmu_buf_t *db)
3140 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3141 return (dbi->db_objset);
3145 dmu_buf_dnode_enter(dmu_buf_t *db)
3147 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3148 DB_DNODE_ENTER(dbi);
3149 return (DB_DNODE(dbi));
3153 dmu_buf_dnode_exit(dmu_buf_t *db)
3155 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3160 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
3162 /* ASSERT(dmu_tx_is_syncing(tx) */
3163 ASSERT(MUTEX_HELD(&db->db_mtx));
3165 if (db->db_blkptr != NULL)
3168 if (db->db_blkid == DMU_SPILL_BLKID) {
3169 db->db_blkptr = DN_SPILL_BLKPTR(dn->dn_phys);
3170 BP_ZERO(db->db_blkptr);
3173 if (db->db_level == dn->dn_phys->dn_nlevels-1) {
3175 * This buffer was allocated at a time when there was
3176 * no available blkptrs from the dnode, or it was
3177 * inappropriate to hook it in (i.e., nlevels mis-match).
3179 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
3180 ASSERT(db->db_parent == NULL);
3181 db->db_parent = dn->dn_dbuf;
3182 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
3185 dmu_buf_impl_t *parent = db->db_parent;
3186 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3188 ASSERT(dn->dn_phys->dn_nlevels > 1);
3189 if (parent == NULL) {
3190 mutex_exit(&db->db_mtx);
3191 rw_enter(&dn->dn_struct_rwlock, RW_READER);
3192 parent = dbuf_hold_level(dn, db->db_level + 1,
3193 db->db_blkid >> epbs, db);
3194 rw_exit(&dn->dn_struct_rwlock);
3195 mutex_enter(&db->db_mtx);
3196 db->db_parent = parent;
3198 db->db_blkptr = (blkptr_t *)parent->db.db_data +
3199 (db->db_blkid & ((1ULL << epbs) - 1));
3205 * Ensure the dbuf's data is untransformed if the associated dirty
3206 * record requires it. This is used by dbuf_sync_leaf() to ensure
3207 * that a dnode block is decrypted before we write new data to it.
3208 * For raw writes we assert that the buffer is already encrypted.
3211 dbuf_check_crypt(dbuf_dirty_record_t *dr)
3214 dmu_buf_impl_t *db = dr->dr_dbuf;
3216 ASSERT(MUTEX_HELD(&db->db_mtx));
3218 if (!dr->dt.dl.dr_raw && arc_is_encrypted(db->db_buf)) {
3220 * Unfortunately, there is currently no mechanism for
3221 * syncing context to handle decryption errors. An error
3222 * here is only possible if an attacker maliciously
3223 * changed a dnode block and updated the associated
3224 * checksums going up the block tree.
3226 err = arc_untransform(db->db_buf, db->db_objset->os_spa,
3227 dmu_objset_id(db->db_objset), B_TRUE);
3229 panic("Invalid dnode block MAC");
3230 } else if (dr->dt.dl.dr_raw) {
3232 * Writing raw encrypted data requires the db's arc buffer
3233 * to be converted to raw by the caller.
3235 ASSERT(arc_is_encrypted(db->db_buf));
3240 * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it
3241 * is critical the we not allow the compiler to inline this function in to
3242 * dbuf_sync_list() thereby drastically bloating the stack usage.
3244 noinline static void
3245 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
3247 dmu_buf_impl_t *db = dr->dr_dbuf;
3251 ASSERT(dmu_tx_is_syncing(tx));
3253 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
3255 mutex_enter(&db->db_mtx);
3257 ASSERT(db->db_level > 0);
3260 /* Read the block if it hasn't been read yet. */
3261 if (db->db_buf == NULL) {
3262 mutex_exit(&db->db_mtx);
3263 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
3264 mutex_enter(&db->db_mtx);
3266 ASSERT3U(db->db_state, ==, DB_CACHED);
3267 ASSERT(db->db_buf != NULL);
3271 /* Indirect block size must match what the dnode thinks it is. */
3272 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3273 dbuf_check_blkptr(dn, db);
3276 /* Provide the pending dirty record to child dbufs */
3277 db->db_data_pending = dr;
3279 mutex_exit(&db->db_mtx);
3280 dbuf_write(dr, db->db_buf, tx);
3283 mutex_enter(&dr->dt.di.dr_mtx);
3284 dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx);
3285 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3286 mutex_exit(&dr->dt.di.dr_mtx);
3291 * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is
3292 * critical the we not allow the compiler to inline this function in to
3293 * dbuf_sync_list() thereby drastically bloating the stack usage.
3295 noinline static void
3296 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
3298 arc_buf_t **datap = &dr->dt.dl.dr_data;
3299 dmu_buf_impl_t *db = dr->dr_dbuf;
3302 uint64_t txg = tx->tx_txg;
3304 ASSERT(dmu_tx_is_syncing(tx));
3306 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
3308 mutex_enter(&db->db_mtx);
3310 * To be synced, we must be dirtied. But we
3311 * might have been freed after the dirty.
3313 if (db->db_state == DB_UNCACHED) {
3314 /* This buffer has been freed since it was dirtied */
3315 ASSERT(db->db.db_data == NULL);
3316 } else if (db->db_state == DB_FILL) {
3317 /* This buffer was freed and is now being re-filled */
3318 ASSERT(db->db.db_data != dr->dt.dl.dr_data);
3320 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
3327 if (db->db_blkid == DMU_SPILL_BLKID) {
3328 mutex_enter(&dn->dn_mtx);
3329 if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
3331 * In the previous transaction group, the bonus buffer
3332 * was entirely used to store the attributes for the
3333 * dnode which overrode the dn_spill field. However,
3334 * when adding more attributes to the file a spill
3335 * block was required to hold the extra attributes.
3337 * Make sure to clear the garbage left in the dn_spill
3338 * field from the previous attributes in the bonus
3339 * buffer. Otherwise, after writing out the spill
3340 * block to the new allocated dva, it will free
3341 * the old block pointed to by the invalid dn_spill.
3343 db->db_blkptr = NULL;
3345 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
3346 mutex_exit(&dn->dn_mtx);
3350 * If this is a bonus buffer, simply copy the bonus data into the
3351 * dnode. It will be written out when the dnode is synced (and it
3352 * will be synced, since it must have been dirty for dbuf_sync to
3355 if (db->db_blkid == DMU_BONUS_BLKID) {
3356 dbuf_dirty_record_t **drp;
3358 ASSERT(*datap != NULL);
3359 ASSERT0(db->db_level);
3360 ASSERT3U(DN_MAX_BONUS_LEN(dn->dn_phys), <=,
3361 DN_SLOTS_TO_BONUSLEN(dn->dn_phys->dn_extra_slots + 1));
3362 bcopy(*datap, DN_BONUS(dn->dn_phys),
3363 DN_MAX_BONUS_LEN(dn->dn_phys));
3366 if (*datap != db->db.db_data) {
3367 int slots = DB_DNODE(db)->dn_num_slots;
3368 int bonuslen = DN_SLOTS_TO_BONUSLEN(slots);
3369 kmem_free(*datap, bonuslen);
3370 arc_space_return(bonuslen, ARC_SPACE_BONUS);
3372 db->db_data_pending = NULL;
3373 drp = &db->db_last_dirty;
3375 drp = &(*drp)->dr_next;
3376 ASSERT(dr->dr_next == NULL);
3377 ASSERT(dr->dr_dbuf == db);
3379 if (dr->dr_dbuf->db_level != 0) {
3380 mutex_destroy(&dr->dt.di.dr_mtx);
3381 list_destroy(&dr->dt.di.dr_children);
3383 kmem_free(dr, sizeof (dbuf_dirty_record_t));
3384 ASSERT(db->db_dirtycnt > 0);
3385 db->db_dirtycnt -= 1;
3386 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
3393 * This function may have dropped the db_mtx lock allowing a dmu_sync
3394 * operation to sneak in. As a result, we need to ensure that we
3395 * don't check the dr_override_state until we have returned from
3396 * dbuf_check_blkptr.
3398 dbuf_check_blkptr(dn, db);
3401 * If this buffer is in the middle of an immediate write,
3402 * wait for the synchronous IO to complete.
3404 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
3405 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
3406 cv_wait(&db->db_changed, &db->db_mtx);
3407 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
3411 * If this is a dnode block, ensure it is appropriately encrypted
3412 * or decrypted, depending on what we are writing to it this txg.
3414 if (os->os_encrypted && dn->dn_object == DMU_META_DNODE_OBJECT)
3415 dbuf_check_crypt(dr);
3417 if (db->db_state != DB_NOFILL &&
3418 dn->dn_object != DMU_META_DNODE_OBJECT &&
3419 refcount_count(&db->db_holds) > 1 &&
3420 dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
3421 *datap == db->db_buf) {
3423 * If this buffer is currently "in use" (i.e., there
3424 * are active holds and db_data still references it),
3425 * then make a copy before we start the write so that
3426 * any modifications from the open txg will not leak
3429 * NOTE: this copy does not need to be made for
3430 * objects only modified in the syncing context (e.g.
3431 * DNONE_DNODE blocks).
3433 int psize = arc_buf_size(*datap);
3434 int lsize = arc_buf_lsize(*datap);
3435 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
3436 enum zio_compress compress_type = arc_get_compression(*datap);
3438 if (arc_is_encrypted(*datap)) {
3439 boolean_t byteorder;
3440 uint8_t salt[ZIO_DATA_SALT_LEN];
3441 uint8_t iv[ZIO_DATA_IV_LEN];
3442 uint8_t mac[ZIO_DATA_MAC_LEN];
3444 arc_get_raw_params(*datap, &byteorder, salt, iv, mac);
3445 *datap = arc_alloc_raw_buf(os->os_spa, db,
3446 dmu_objset_id(os), byteorder, salt, iv, mac,
3447 dn->dn_type, psize, lsize, compress_type);
3448 } else if (compress_type != ZIO_COMPRESS_OFF) {
3449 ASSERT3U(type, ==, ARC_BUFC_DATA);
3450 int lsize = arc_buf_lsize(*datap);
3451 *datap = arc_alloc_compressed_buf(os->os_spa, db,
3452 psize, lsize, compress_type);
3454 *datap = arc_alloc_buf(os->os_spa, db, type, psize);
3456 bcopy(db->db.db_data, (*datap)->b_data, psize);
3458 db->db_data_pending = dr;
3460 mutex_exit(&db->db_mtx);
3462 dbuf_write(dr, *datap, tx);
3464 ASSERT(!list_link_active(&dr->dr_dirty_node));
3465 if (dn->dn_object == DMU_META_DNODE_OBJECT) {
3466 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
3470 * Although zio_nowait() does not "wait for an IO", it does
3471 * initiate the IO. If this is an empty write it seems plausible
3472 * that the IO could actually be completed before the nowait
3473 * returns. We need to DB_DNODE_EXIT() first in case
3474 * zio_nowait() invalidates the dbuf.
3477 zio_nowait(dr->dr_zio);
3482 dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx)
3484 dbuf_dirty_record_t *dr;
3486 while ((dr = list_head(list))) {
3487 if (dr->dr_zio != NULL) {
3489 * If we find an already initialized zio then we
3490 * are processing the meta-dnode, and we have finished.
3491 * The dbufs for all dnodes are put back on the list
3492 * during processing, so that we can zio_wait()
3493 * these IOs after initiating all child IOs.
3495 ASSERT3U(dr->dr_dbuf->db.db_object, ==,
3496 DMU_META_DNODE_OBJECT);
3499 if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
3500 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
3501 VERIFY3U(dr->dr_dbuf->db_level, ==, level);
3503 list_remove(list, dr);
3504 if (dr->dr_dbuf->db_level > 0)
3505 dbuf_sync_indirect(dr, tx);
3507 dbuf_sync_leaf(dr, tx);
3513 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3515 dmu_buf_impl_t *db = vdb;
3517 blkptr_t *bp = zio->io_bp;
3518 blkptr_t *bp_orig = &zio->io_bp_orig;
3519 spa_t *spa = zio->io_spa;
3524 ASSERT3P(db->db_blkptr, !=, NULL);
3525 ASSERT3P(&db->db_data_pending->dr_bp_copy, ==, bp);
3529 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
3530 dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
3531 zio->io_prev_space_delta = delta;
3533 if (bp->blk_birth != 0) {
3534 ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
3535 BP_GET_TYPE(bp) == dn->dn_type) ||
3536 (db->db_blkid == DMU_SPILL_BLKID &&
3537 BP_GET_TYPE(bp) == dn->dn_bonustype) ||
3538 BP_IS_EMBEDDED(bp));
3539 ASSERT(BP_GET_LEVEL(bp) == db->db_level);
3542 mutex_enter(&db->db_mtx);
3545 if (db->db_blkid == DMU_SPILL_BLKID) {
3546 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3547 ASSERT(!(BP_IS_HOLE(bp)) &&
3548 db->db_blkptr == DN_SPILL_BLKPTR(dn->dn_phys));
3552 if (db->db_level == 0) {
3553 mutex_enter(&dn->dn_mtx);
3554 if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
3555 db->db_blkid != DMU_SPILL_BLKID)
3556 dn->dn_phys->dn_maxblkid = db->db_blkid;
3557 mutex_exit(&dn->dn_mtx);
3559 if (dn->dn_type == DMU_OT_DNODE) {
3561 while (i < db->db.db_size) {
3563 (void *)(((char *)db->db.db_data) + i);
3565 i += DNODE_MIN_SIZE;
3566 if (dnp->dn_type != DMU_OT_NONE) {
3568 i += dnp->dn_extra_slots *
3573 if (BP_IS_HOLE(bp)) {
3580 blkptr_t *ibp = db->db.db_data;
3581 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3582 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
3583 if (BP_IS_HOLE(ibp))
3585 fill += BP_GET_FILL(ibp);
3590 if (!BP_IS_EMBEDDED(bp))
3591 BP_SET_FILL(bp, fill);
3593 mutex_exit(&db->db_mtx);
3595 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3596 *db->db_blkptr = *bp;
3597 rw_exit(&dn->dn_struct_rwlock);
3602 * This function gets called just prior to running through the compression
3603 * stage of the zio pipeline. If we're an indirect block comprised of only
3604 * holes, then we want this indirect to be compressed away to a hole. In
3605 * order to do that we must zero out any information about the holes that
3606 * this indirect points to prior to before we try to compress it.
3609 dbuf_write_children_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3611 dmu_buf_impl_t *db = vdb;
3614 unsigned int epbs, i;
3616 ASSERT3U(db->db_level, >, 0);
3619 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3620 ASSERT3U(epbs, <, 31);
3622 /* Determine if all our children are holes */
3623 for (i = 0, bp = db->db.db_data; i < 1ULL << epbs; i++, bp++) {
3624 if (!BP_IS_HOLE(bp))
3629 * If all the children are holes, then zero them all out so that
3630 * we may get compressed away.
3632 if (i == 1ULL << epbs) {
3634 * We only found holes. Grab the rwlock to prevent
3635 * anybody from reading the blocks we're about to
3638 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3639 bzero(db->db.db_data, db->db.db_size);
3640 rw_exit(&dn->dn_struct_rwlock);
3646 * The SPA will call this callback several times for each zio - once
3647 * for every physical child i/o (zio->io_phys_children times). This
3648 * allows the DMU to monitor the progress of each logical i/o. For example,
3649 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
3650 * block. There may be a long delay before all copies/fragments are completed,
3651 * so this callback allows us to retire dirty space gradually, as the physical
3656 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
3658 dmu_buf_impl_t *db = arg;
3659 objset_t *os = db->db_objset;
3660 dsl_pool_t *dp = dmu_objset_pool(os);
3661 dbuf_dirty_record_t *dr;
3664 dr = db->db_data_pending;
3665 ASSERT3U(dr->dr_txg, ==, zio->io_txg);
3668 * The callback will be called io_phys_children times. Retire one
3669 * portion of our dirty space each time we are called. Any rounding
3670 * error will be cleaned up by dsl_pool_sync()'s call to
3671 * dsl_pool_undirty_space().
3673 delta = dr->dr_accounted / zio->io_phys_children;
3674 dsl_pool_undirty_space(dp, delta, zio->io_txg);
3679 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
3681 dmu_buf_impl_t *db = vdb;
3682 blkptr_t *bp_orig = &zio->io_bp_orig;
3683 blkptr_t *bp = db->db_blkptr;
3684 objset_t *os = db->db_objset;
3685 dmu_tx_t *tx = os->os_synctx;
3686 dbuf_dirty_record_t **drp, *dr;
3688 ASSERT0(zio->io_error);
3689 ASSERT(db->db_blkptr == bp);
3692 * For nopwrites and rewrites we ensure that the bp matches our
3693 * original and bypass all the accounting.
3695 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
3696 ASSERT(BP_EQUAL(bp, bp_orig));
3698 dsl_dataset_t *ds = os->os_dsl_dataset;
3699 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
3700 dsl_dataset_block_born(ds, bp, tx);
3703 mutex_enter(&db->db_mtx);
3707 drp = &db->db_last_dirty;
3708 while ((dr = *drp) != db->db_data_pending)
3710 ASSERT(!list_link_active(&dr->dr_dirty_node));
3711 ASSERT(dr->dr_dbuf == db);
3712 ASSERT(dr->dr_next == NULL);
3716 if (db->db_blkid == DMU_SPILL_BLKID) {
3721 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3722 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
3723 db->db_blkptr == DN_SPILL_BLKPTR(dn->dn_phys));
3728 if (db->db_level == 0) {
3729 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
3730 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
3731 if (db->db_state != DB_NOFILL) {
3732 if (dr->dt.dl.dr_data != db->db_buf)
3733 arc_buf_destroy(dr->dt.dl.dr_data, db);
3740 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3741 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
3742 if (!BP_IS_HOLE(db->db_blkptr)) {
3743 ASSERTV(int epbs = dn->dn_phys->dn_indblkshift -
3745 ASSERT3U(db->db_blkid, <=,
3746 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs));
3747 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
3751 mutex_destroy(&dr->dt.di.dr_mtx);
3752 list_destroy(&dr->dt.di.dr_children);
3754 kmem_free(dr, sizeof (dbuf_dirty_record_t));
3756 cv_broadcast(&db->db_changed);
3757 ASSERT(db->db_dirtycnt > 0);
3758 db->db_dirtycnt -= 1;
3759 db->db_data_pending = NULL;
3760 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg);
3764 dbuf_write_nofill_ready(zio_t *zio)
3766 dbuf_write_ready(zio, NULL, zio->io_private);
3770 dbuf_write_nofill_done(zio_t *zio)
3772 dbuf_write_done(zio, NULL, zio->io_private);
3776 dbuf_write_override_ready(zio_t *zio)
3778 dbuf_dirty_record_t *dr = zio->io_private;
3779 dmu_buf_impl_t *db = dr->dr_dbuf;
3781 dbuf_write_ready(zio, NULL, db);
3785 dbuf_write_override_done(zio_t *zio)
3787 dbuf_dirty_record_t *dr = zio->io_private;
3788 dmu_buf_impl_t *db = dr->dr_dbuf;
3789 blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
3791 mutex_enter(&db->db_mtx);
3792 if (!BP_EQUAL(zio->io_bp, obp)) {
3793 if (!BP_IS_HOLE(obp))
3794 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
3795 arc_release(dr->dt.dl.dr_data, db);
3797 mutex_exit(&db->db_mtx);
3799 dbuf_write_done(zio, NULL, db);
3801 if (zio->io_abd != NULL)
3802 abd_put(zio->io_abd);
3805 /* Issue I/O to commit a dirty buffer to disk. */
3807 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
3809 dmu_buf_impl_t *db = dr->dr_dbuf;
3812 dmu_buf_impl_t *parent = db->db_parent;
3813 uint64_t txg = tx->tx_txg;
3814 zbookmark_phys_t zb;
3819 ASSERT(dmu_tx_is_syncing(tx));
3825 if (db->db_state != DB_NOFILL) {
3826 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
3828 * Private object buffers are released here rather
3829 * than in dbuf_dirty() since they are only modified
3830 * in the syncing context and we don't want the
3831 * overhead of making multiple copies of the data.
3833 if (BP_IS_HOLE(db->db_blkptr)) {
3836 dbuf_release_bp(db);
3841 if (parent != dn->dn_dbuf) {
3842 /* Our parent is an indirect block. */
3843 /* We have a dirty parent that has been scheduled for write. */
3844 ASSERT(parent && parent->db_data_pending);
3845 /* Our parent's buffer is one level closer to the dnode. */
3846 ASSERT(db->db_level == parent->db_level-1);
3848 * We're about to modify our parent's db_data by modifying
3849 * our block pointer, so the parent must be released.
3851 ASSERT(arc_released(parent->db_buf));
3852 zio = parent->db_data_pending->dr_zio;
3854 /* Our parent is the dnode itself. */
3855 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
3856 db->db_blkid != DMU_SPILL_BLKID) ||
3857 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
3858 if (db->db_blkid != DMU_SPILL_BLKID)
3859 ASSERT3P(db->db_blkptr, ==,
3860 &dn->dn_phys->dn_blkptr[db->db_blkid]);
3864 ASSERT(db->db_level == 0 || data == db->db_buf);
3865 ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
3868 SET_BOOKMARK(&zb, os->os_dsl_dataset ?
3869 os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
3870 db->db.db_object, db->db_level, db->db_blkid);
3872 if (db->db_blkid == DMU_SPILL_BLKID)
3874 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
3876 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
3880 * We copy the blkptr now (rather than when we instantiate the dirty
3881 * record), because its value can change between open context and
3882 * syncing context. We do not need to hold dn_struct_rwlock to read
3883 * db_blkptr because we are in syncing context.
3885 dr->dr_bp_copy = *db->db_blkptr;
3887 if (db->db_level == 0 &&
3888 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
3890 * The BP for this block has been provided by open context
3891 * (by dmu_sync() or dmu_buf_write_embedded()).
3893 abd_t *contents = (data != NULL) ?
3894 abd_get_from_buf(data->b_data, arc_buf_size(data)) : NULL;
3896 dr->dr_zio = zio_write(zio, os->os_spa, txg,
3897 &dr->dr_bp_copy, contents, db->db.db_size, db->db.db_size,
3898 &zp, dbuf_write_override_ready, NULL, NULL,
3899 dbuf_write_override_done,
3900 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
3901 mutex_enter(&db->db_mtx);
3902 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
3903 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
3904 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
3905 mutex_exit(&db->db_mtx);
3906 } else if (db->db_state == DB_NOFILL) {
3907 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF ||
3908 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY);
3909 dr->dr_zio = zio_write(zio, os->os_spa, txg,
3910 &dr->dr_bp_copy, NULL, db->db.db_size, db->db.db_size, &zp,
3911 dbuf_write_nofill_ready, NULL, NULL,
3912 dbuf_write_nofill_done, db,
3913 ZIO_PRIORITY_ASYNC_WRITE,
3914 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
3916 arc_write_done_func_t *children_ready_cb = NULL;
3917 ASSERT(arc_released(data));
3920 * For indirect blocks, we want to setup the children
3921 * ready callback so that we can properly handle an indirect
3922 * block that only contains holes.
3924 if (db->db_level != 0)
3925 children_ready_cb = dbuf_write_children_ready;
3927 dr->dr_zio = arc_write(zio, os->os_spa, txg,
3928 &dr->dr_bp_copy, data, DBUF_IS_L2CACHEABLE(db),
3929 &zp, dbuf_write_ready,
3930 children_ready_cb, dbuf_write_physdone,
3931 dbuf_write_done, db, ZIO_PRIORITY_ASYNC_WRITE,
3932 ZIO_FLAG_MUSTSUCCEED, &zb);
3936 #if defined(_KERNEL) && defined(HAVE_SPL)
3937 EXPORT_SYMBOL(dbuf_find);
3938 EXPORT_SYMBOL(dbuf_is_metadata);
3939 EXPORT_SYMBOL(dbuf_destroy);
3940 EXPORT_SYMBOL(dbuf_loan_arcbuf);
3941 EXPORT_SYMBOL(dbuf_whichblock);
3942 EXPORT_SYMBOL(dbuf_read);
3943 EXPORT_SYMBOL(dbuf_unoverride);
3944 EXPORT_SYMBOL(dbuf_free_range);
3945 EXPORT_SYMBOL(dbuf_new_size);
3946 EXPORT_SYMBOL(dbuf_release_bp);
3947 EXPORT_SYMBOL(dbuf_dirty);
3948 EXPORT_SYMBOL(dmu_buf_will_change_crypt_params);
3949 EXPORT_SYMBOL(dmu_buf_will_dirty);
3950 EXPORT_SYMBOL(dmu_buf_will_not_fill);
3951 EXPORT_SYMBOL(dmu_buf_will_fill);
3952 EXPORT_SYMBOL(dmu_buf_fill_done);
3953 EXPORT_SYMBOL(dmu_buf_rele);
3954 EXPORT_SYMBOL(dbuf_assign_arcbuf);
3955 EXPORT_SYMBOL(dbuf_prefetch);
3956 EXPORT_SYMBOL(dbuf_hold_impl);
3957 EXPORT_SYMBOL(dbuf_hold);
3958 EXPORT_SYMBOL(dbuf_hold_level);
3959 EXPORT_SYMBOL(dbuf_create_bonus);
3960 EXPORT_SYMBOL(dbuf_spill_set_blksz);
3961 EXPORT_SYMBOL(dbuf_rm_spill);
3962 EXPORT_SYMBOL(dbuf_add_ref);
3963 EXPORT_SYMBOL(dbuf_rele);
3964 EXPORT_SYMBOL(dbuf_rele_and_unlock);
3965 EXPORT_SYMBOL(dbuf_refcount);
3966 EXPORT_SYMBOL(dbuf_sync_list);
3967 EXPORT_SYMBOL(dmu_buf_set_user);
3968 EXPORT_SYMBOL(dmu_buf_set_user_ie);
3969 EXPORT_SYMBOL(dmu_buf_get_user);
3970 EXPORT_SYMBOL(dmu_buf_get_blkptr);
3973 module_param(dbuf_cache_max_bytes, ulong, 0644);
3974 MODULE_PARM_DESC(dbuf_cache_max_bytes,
3975 "Maximum size in bytes of the dbuf cache.");
3977 module_param(dbuf_cache_hiwater_pct, uint, 0644);
3978 MODULE_PARM_DESC(dbuf_cache_hiwater_pct,
3979 "Percentage over dbuf_cache_max_bytes when dbufs must be evicted "
3982 module_param(dbuf_cache_lowater_pct, uint, 0644);
3983 MODULE_PARM_DESC(dbuf_cache_lowater_pct,
3984 "Percentage below dbuf_cache_max_bytes when the evict thread stops "
3987 module_param(dbuf_cache_max_shift, int, 0644);
3988 MODULE_PARM_DESC(dbuf_cache_max_shift,
3989 "Cap the size of the dbuf cache to a log2 fraction of arc size.");