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) 2013, Joyent, Inc. All rights reserved.
27 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
28 * Copyright (c) 2014 Integros [integros.com]
31 #include <sys/zfs_context.h>
33 #include <sys/dmu_send.h>
34 #include <sys/dmu_impl.h>
36 #include <sys/dmu_objset.h>
37 #include <sys/dsl_dataset.h>
38 #include <sys/dsl_dir.h>
39 #include <sys/dmu_tx.h>
42 #include <sys/dmu_zfetch.h>
44 #include <sys/sa_impl.h>
45 #include <sys/zfeature.h>
46 #include <sys/blkptr.h>
47 #include <sys/range_tree.h>
48 #include <sys/callb.h>
51 #include <sys/cityhash.h>
53 uint_t zfs_dbuf_evict_key;
55 static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
56 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
59 extern inline void dmu_buf_init_user(dmu_buf_user_t *dbu,
60 dmu_buf_evict_func_t *evict_func_sync,
61 dmu_buf_evict_func_t *evict_func_async,
62 dmu_buf_t **clear_on_evict_dbufp);
66 * Global data structures and functions for the dbuf cache.
68 static kmem_cache_t *dbuf_kmem_cache;
69 static taskq_t *dbu_evict_taskq;
71 static kthread_t *dbuf_cache_evict_thread;
72 static kmutex_t dbuf_evict_lock;
73 static kcondvar_t dbuf_evict_cv;
74 static boolean_t dbuf_evict_thread_exit;
77 * LRU cache of dbufs. The dbuf cache maintains a list of dbufs that
78 * are not currently held but have been recently released. These dbufs
79 * are not eligible for arc eviction until they are aged out of the cache.
80 * Dbufs are added to the dbuf cache once the last hold is released. If a
81 * dbuf is later accessed and still exists in the dbuf cache, then it will
82 * be removed from the cache and later re-added to the head of the cache.
83 * Dbufs that are aged out of the cache will be immediately destroyed and
84 * become eligible for arc eviction.
86 static multilist_t *dbuf_cache;
87 static refcount_t dbuf_cache_size;
88 uint64_t dbuf_cache_max_bytes = 0;
90 /* Set the default size of the dbuf cache to log2 fraction of arc size. */
91 int dbuf_cache_shift = 5;
94 * The dbuf cache uses a three-stage eviction policy:
95 * - A low water marker designates when the dbuf eviction thread
96 * should stop evicting from the dbuf cache.
97 * - When we reach the maximum size (aka mid water mark), we
98 * signal the eviction thread to run.
99 * - The high water mark indicates when the eviction thread
100 * is unable to keep up with the incoming load and eviction must
101 * happen in the context of the calling thread.
105 * low water mid water hi water
106 * +----------------------------------------+----------+----------+
111 * +----------------------------------------+----------+----------+
113 * evicting eviction directly
116 * The high and low water marks indicate the operating range for the eviction
117 * thread. The low water mark is, by default, 90% of the total size of the
118 * cache and the high water mark is at 110% (both of these percentages can be
119 * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct,
120 * respectively). The eviction thread will try to ensure that the cache remains
121 * within this range by waking up every second and checking if the cache is
122 * above the low water mark. The thread can also be woken up by callers adding
123 * elements into the cache if the cache is larger than the mid water (i.e max
124 * cache size). Once the eviction thread is woken up and eviction is required,
125 * it will continue evicting buffers until it's able to reduce the cache size
126 * to the low water mark. If the cache size continues to grow and hits the high
127 * water mark, then callers adding elments to the cache will begin to evict
128 * directly from the cache until the cache is no longer above the high water
133 * The percentage above and below the maximum cache size.
135 uint_t dbuf_cache_hiwater_pct = 10;
136 uint_t dbuf_cache_lowater_pct = 10;
138 SYSCTL_DECL(_vfs_zfs);
139 SYSCTL_QUAD(_vfs_zfs, OID_AUTO, dbuf_cache_max_bytes, CTLFLAG_RWTUN,
140 &dbuf_cache_max_bytes, 0, "dbuf cache size in bytes");
141 SYSCTL_INT(_vfs_zfs, OID_AUTO, dbuf_cache_shift, CTLFLAG_RDTUN,
142 &dbuf_cache_shift, 0, "dbuf cache size as log2 fraction of ARC");
143 SYSCTL_UINT(_vfs_zfs, OID_AUTO, dbuf_cache_hiwater_pct, CTLFLAG_RWTUN,
144 &dbuf_cache_hiwater_pct, 0, "max percents above the dbuf cache size");
145 SYSCTL_UINT(_vfs_zfs, OID_AUTO, dbuf_cache_lowater_pct, CTLFLAG_RWTUN,
146 &dbuf_cache_lowater_pct, 0, "max percents below the dbuf cache size");
150 dbuf_cons(void *vdb, void *unused, int kmflag)
152 dmu_buf_impl_t *db = vdb;
153 bzero(db, sizeof (dmu_buf_impl_t));
155 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
156 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
157 multilist_link_init(&db->db_cache_link);
158 refcount_create(&db->db_holds);
165 dbuf_dest(void *vdb, void *unused)
167 dmu_buf_impl_t *db = vdb;
168 mutex_destroy(&db->db_mtx);
169 cv_destroy(&db->db_changed);
170 ASSERT(!multilist_link_active(&db->db_cache_link));
171 refcount_destroy(&db->db_holds);
175 * dbuf hash table routines
177 static dbuf_hash_table_t dbuf_hash_table;
179 static uint64_t dbuf_hash_count;
182 * We use Cityhash for this. It's fast, and has good hash properties without
183 * requiring any large static buffers.
186 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
188 return (cityhash4((uintptr_t)os, obj, (uint64_t)lvl, blkid));
191 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
192 ((dbuf)->db.db_object == (obj) && \
193 (dbuf)->db_objset == (os) && \
194 (dbuf)->db_level == (level) && \
195 (dbuf)->db_blkid == (blkid))
198 dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid)
200 dbuf_hash_table_t *h = &dbuf_hash_table;
201 uint64_t hv = dbuf_hash(os, obj, level, blkid);
202 uint64_t idx = hv & h->hash_table_mask;
205 mutex_enter(DBUF_HASH_MUTEX(h, idx));
206 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
207 if (DBUF_EQUAL(db, os, obj, level, blkid)) {
208 mutex_enter(&db->db_mtx);
209 if (db->db_state != DB_EVICTING) {
210 mutex_exit(DBUF_HASH_MUTEX(h, idx));
213 mutex_exit(&db->db_mtx);
216 mutex_exit(DBUF_HASH_MUTEX(h, idx));
220 static dmu_buf_impl_t *
221 dbuf_find_bonus(objset_t *os, uint64_t object)
224 dmu_buf_impl_t *db = NULL;
226 if (dnode_hold(os, object, FTAG, &dn) == 0) {
227 rw_enter(&dn->dn_struct_rwlock, RW_READER);
228 if (dn->dn_bonus != NULL) {
230 mutex_enter(&db->db_mtx);
232 rw_exit(&dn->dn_struct_rwlock);
233 dnode_rele(dn, FTAG);
239 * Insert an entry into the hash table. If there is already an element
240 * equal to elem in the hash table, then the already existing element
241 * will be returned and the new element will not be inserted.
242 * Otherwise returns NULL.
244 static dmu_buf_impl_t *
245 dbuf_hash_insert(dmu_buf_impl_t *db)
247 dbuf_hash_table_t *h = &dbuf_hash_table;
248 objset_t *os = db->db_objset;
249 uint64_t obj = db->db.db_object;
250 int level = db->db_level;
251 uint64_t blkid = db->db_blkid;
252 uint64_t hv = dbuf_hash(os, obj, level, blkid);
253 uint64_t idx = hv & h->hash_table_mask;
256 mutex_enter(DBUF_HASH_MUTEX(h, idx));
257 for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) {
258 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
259 mutex_enter(&dbf->db_mtx);
260 if (dbf->db_state != DB_EVICTING) {
261 mutex_exit(DBUF_HASH_MUTEX(h, idx));
264 mutex_exit(&dbf->db_mtx);
268 mutex_enter(&db->db_mtx);
269 db->db_hash_next = h->hash_table[idx];
270 h->hash_table[idx] = db;
271 mutex_exit(DBUF_HASH_MUTEX(h, idx));
272 atomic_inc_64(&dbuf_hash_count);
278 * Remove an entry from the hash table. It must be in the EVICTING state.
281 dbuf_hash_remove(dmu_buf_impl_t *db)
283 dbuf_hash_table_t *h = &dbuf_hash_table;
284 uint64_t hv = dbuf_hash(db->db_objset, db->db.db_object,
285 db->db_level, db->db_blkid);
286 uint64_t idx = hv & h->hash_table_mask;
287 dmu_buf_impl_t *dbf, **dbp;
290 * We musn't hold db_mtx to maintain lock ordering:
291 * DBUF_HASH_MUTEX > db_mtx.
293 ASSERT(refcount_is_zero(&db->db_holds));
294 ASSERT(db->db_state == DB_EVICTING);
295 ASSERT(!MUTEX_HELD(&db->db_mtx));
297 mutex_enter(DBUF_HASH_MUTEX(h, idx));
298 dbp = &h->hash_table[idx];
299 while ((dbf = *dbp) != db) {
300 dbp = &dbf->db_hash_next;
303 *dbp = db->db_hash_next;
304 db->db_hash_next = NULL;
305 mutex_exit(DBUF_HASH_MUTEX(h, idx));
306 atomic_dec_64(&dbuf_hash_count);
312 } dbvu_verify_type_t;
315 dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type)
320 if (db->db_user == NULL)
323 /* Only data blocks support the attachment of user data. */
324 ASSERT(db->db_level == 0);
326 /* Clients must resolve a dbuf before attaching user data. */
327 ASSERT(db->db.db_data != NULL);
328 ASSERT3U(db->db_state, ==, DB_CACHED);
330 holds = refcount_count(&db->db_holds);
331 if (verify_type == DBVU_EVICTING) {
333 * Immediate eviction occurs when holds == dirtycnt.
334 * For normal eviction buffers, holds is zero on
335 * eviction, except when dbuf_fix_old_data() calls
336 * dbuf_clear_data(). However, the hold count can grow
337 * during eviction even though db_mtx is held (see
338 * dmu_bonus_hold() for an example), so we can only
339 * test the generic invariant that holds >= dirtycnt.
341 ASSERT3U(holds, >=, db->db_dirtycnt);
343 if (db->db_user_immediate_evict == TRUE)
344 ASSERT3U(holds, >=, db->db_dirtycnt);
346 ASSERT3U(holds, >, 0);
352 dbuf_evict_user(dmu_buf_impl_t *db)
354 dmu_buf_user_t *dbu = db->db_user;
356 ASSERT(MUTEX_HELD(&db->db_mtx));
361 dbuf_verify_user(db, DBVU_EVICTING);
365 if (dbu->dbu_clear_on_evict_dbufp != NULL)
366 *dbu->dbu_clear_on_evict_dbufp = NULL;
370 * There are two eviction callbacks - one that we call synchronously
371 * and one that we invoke via a taskq. The async one is useful for
372 * avoiding lock order reversals and limiting stack depth.
374 * Note that if we have a sync callback but no async callback,
375 * it's likely that the sync callback will free the structure
376 * containing the dbu. In that case we need to take care to not
377 * dereference dbu after calling the sync evict func.
379 boolean_t has_async = (dbu->dbu_evict_func_async != NULL);
381 if (dbu->dbu_evict_func_sync != NULL)
382 dbu->dbu_evict_func_sync(dbu);
385 taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func_async,
386 dbu, 0, &dbu->dbu_tqent);
391 dbuf_is_metadata(dmu_buf_impl_t *db)
393 if (db->db_level > 0) {
396 boolean_t is_metadata;
399 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type);
402 return (is_metadata);
407 * This function *must* return indices evenly distributed between all
408 * sublists of the multilist. This is needed due to how the dbuf eviction
409 * code is laid out; dbuf_evict_thread() assumes dbufs are evenly
410 * distributed between all sublists and uses this assumption when
411 * deciding which sublist to evict from and how much to evict from it.
414 dbuf_cache_multilist_index_func(multilist_t *ml, void *obj)
416 dmu_buf_impl_t *db = obj;
419 * The assumption here, is the hash value for a given
420 * dmu_buf_impl_t will remain constant throughout it's lifetime
421 * (i.e. it's objset, object, level and blkid fields don't change).
422 * Thus, we don't need to store the dbuf's sublist index
423 * on insertion, as this index can be recalculated on removal.
425 * Also, the low order bits of the hash value are thought to be
426 * distributed evenly. Otherwise, in the case that the multilist
427 * has a power of two number of sublists, each sublists' usage
428 * would not be evenly distributed.
430 return (dbuf_hash(db->db_objset, db->db.db_object,
431 db->db_level, db->db_blkid) %
432 multilist_get_num_sublists(ml));
435 static inline boolean_t
436 dbuf_cache_above_hiwater(void)
438 uint64_t dbuf_cache_hiwater_bytes =
439 (dbuf_cache_max_bytes * dbuf_cache_hiwater_pct) / 100;
441 return (refcount_count(&dbuf_cache_size) >
442 dbuf_cache_max_bytes + dbuf_cache_hiwater_bytes);
445 static inline boolean_t
446 dbuf_cache_above_lowater(void)
448 uint64_t dbuf_cache_lowater_bytes =
449 (dbuf_cache_max_bytes * dbuf_cache_lowater_pct) / 100;
451 return (refcount_count(&dbuf_cache_size) >
452 dbuf_cache_max_bytes - dbuf_cache_lowater_bytes);
456 * Evict the oldest eligible dbuf from the dbuf cache.
461 int idx = multilist_get_random_index(dbuf_cache);
462 multilist_sublist_t *mls = multilist_sublist_lock(dbuf_cache, idx);
464 ASSERT(!MUTEX_HELD(&dbuf_evict_lock));
467 * Set the thread's tsd to indicate that it's processing evictions.
468 * Once a thread stops evicting from the dbuf cache it will
469 * reset its tsd to NULL.
471 ASSERT3P(tsd_get(zfs_dbuf_evict_key), ==, NULL);
472 (void) tsd_set(zfs_dbuf_evict_key, (void *)B_TRUE);
474 dmu_buf_impl_t *db = multilist_sublist_tail(mls);
475 while (db != NULL && mutex_tryenter(&db->db_mtx) == 0) {
476 db = multilist_sublist_prev(mls, db);
479 DTRACE_PROBE2(dbuf__evict__one, dmu_buf_impl_t *, db,
480 multilist_sublist_t *, mls);
483 multilist_sublist_remove(mls, db);
484 multilist_sublist_unlock(mls);
485 (void) refcount_remove_many(&dbuf_cache_size,
489 multilist_sublist_unlock(mls);
491 (void) tsd_set(zfs_dbuf_evict_key, NULL);
495 * The dbuf evict thread is responsible for aging out dbufs from the
496 * cache. Once the cache has reached it's maximum size, dbufs are removed
497 * and destroyed. The eviction thread will continue running until the size
498 * of the dbuf cache is at or below the maximum size. Once the dbuf is aged
499 * out of the cache it is destroyed and becomes eligible for arc eviction.
503 dbuf_evict_thread(void *unused __unused)
507 CALLB_CPR_INIT(&cpr, &dbuf_evict_lock, callb_generic_cpr, FTAG);
509 mutex_enter(&dbuf_evict_lock);
510 while (!dbuf_evict_thread_exit) {
511 while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
512 CALLB_CPR_SAFE_BEGIN(&cpr);
513 (void) cv_timedwait_hires(&dbuf_evict_cv,
514 &dbuf_evict_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
515 CALLB_CPR_SAFE_END(&cpr, &dbuf_evict_lock);
517 mutex_exit(&dbuf_evict_lock);
520 * Keep evicting as long as we're above the low water mark
521 * for the cache. We do this without holding the locks to
522 * minimize lock contention.
524 while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
528 mutex_enter(&dbuf_evict_lock);
531 dbuf_evict_thread_exit = B_FALSE;
532 cv_broadcast(&dbuf_evict_cv);
533 CALLB_CPR_EXIT(&cpr); /* drops dbuf_evict_lock */
538 * Wake up the dbuf eviction thread if the dbuf cache is at its max size.
539 * If the dbuf cache is at its high water mark, then evict a dbuf from the
540 * dbuf cache using the callers context.
543 dbuf_evict_notify(void)
547 * We use thread specific data to track when a thread has
548 * started processing evictions. This allows us to avoid deeply
549 * nested stacks that would have a call flow similar to this:
551 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify()
554 * +-----dbuf_destroy()<--dbuf_evict_one()<--------+
556 * The dbuf_eviction_thread will always have its tsd set until
557 * that thread exits. All other threads will only set their tsd
558 * if they are participating in the eviction process. This only
559 * happens if the eviction thread is unable to process evictions
560 * fast enough. To keep the dbuf cache size in check, other threads
561 * can evict from the dbuf cache directly. Those threads will set
562 * their tsd values so that we ensure that they only evict one dbuf
563 * from the dbuf cache.
565 if (tsd_get(zfs_dbuf_evict_key) != NULL)
569 * We check if we should evict without holding the dbuf_evict_lock,
570 * because it's OK to occasionally make the wrong decision here,
571 * and grabbing the lock results in massive lock contention.
573 if (refcount_count(&dbuf_cache_size) > dbuf_cache_max_bytes) {
574 if (dbuf_cache_above_hiwater())
576 cv_signal(&dbuf_evict_cv);
583 uint64_t hsize = 1ULL << 16;
584 dbuf_hash_table_t *h = &dbuf_hash_table;
588 * The hash table is big enough to fill all of physical memory
589 * with an average 4K block size. The table will take up
590 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
592 while (hsize * 4096 < (uint64_t)physmem * PAGESIZE)
596 h->hash_table_mask = hsize - 1;
597 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
598 if (h->hash_table == NULL) {
599 /* XXX - we should really return an error instead of assert */
600 ASSERT(hsize > (1ULL << 10));
605 dbuf_kmem_cache = kmem_cache_create("dmu_buf_impl_t",
606 sizeof (dmu_buf_impl_t),
607 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
609 for (i = 0; i < DBUF_MUTEXES; i++)
610 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
613 * Setup the parameters for the dbuf cache. We set the size of the
614 * dbuf cache to 1/32nd (default) of the size of the ARC. If the value
615 * has been set in /etc/system and it's not greater than the size of
616 * the ARC, then we honor that value.
618 if (dbuf_cache_max_bytes == 0 ||
619 dbuf_cache_max_bytes >= arc_max_bytes()) {
620 dbuf_cache_max_bytes = arc_max_bytes() >> dbuf_cache_shift;
624 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
625 * configuration is not required.
627 dbu_evict_taskq = taskq_create("dbu_evict", 1, minclsyspri, 0, 0, 0);
629 dbuf_cache = multilist_create(sizeof (dmu_buf_impl_t),
630 offsetof(dmu_buf_impl_t, db_cache_link),
631 dbuf_cache_multilist_index_func);
632 refcount_create(&dbuf_cache_size);
634 tsd_create(&zfs_dbuf_evict_key, NULL);
635 dbuf_evict_thread_exit = B_FALSE;
636 mutex_init(&dbuf_evict_lock, NULL, MUTEX_DEFAULT, NULL);
637 cv_init(&dbuf_evict_cv, NULL, CV_DEFAULT, NULL);
638 dbuf_cache_evict_thread = thread_create(NULL, 0, dbuf_evict_thread,
639 NULL, 0, &p0, TS_RUN, minclsyspri);
645 dbuf_hash_table_t *h = &dbuf_hash_table;
648 for (i = 0; i < DBUF_MUTEXES; i++)
649 mutex_destroy(&h->hash_mutexes[i]);
650 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
651 kmem_cache_destroy(dbuf_kmem_cache);
652 taskq_destroy(dbu_evict_taskq);
654 mutex_enter(&dbuf_evict_lock);
655 dbuf_evict_thread_exit = B_TRUE;
656 while (dbuf_evict_thread_exit) {
657 cv_signal(&dbuf_evict_cv);
658 cv_wait(&dbuf_evict_cv, &dbuf_evict_lock);
660 mutex_exit(&dbuf_evict_lock);
661 tsd_destroy(&zfs_dbuf_evict_key);
663 mutex_destroy(&dbuf_evict_lock);
664 cv_destroy(&dbuf_evict_cv);
666 refcount_destroy(&dbuf_cache_size);
667 multilist_destroy(dbuf_cache);
676 dbuf_verify(dmu_buf_impl_t *db)
679 dbuf_dirty_record_t *dr;
681 ASSERT(MUTEX_HELD(&db->db_mtx));
683 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
686 ASSERT(db->db_objset != NULL);
690 ASSERT(db->db_parent == NULL);
691 ASSERT(db->db_blkptr == NULL);
693 ASSERT3U(db->db.db_object, ==, dn->dn_object);
694 ASSERT3P(db->db_objset, ==, dn->dn_objset);
695 ASSERT3U(db->db_level, <, dn->dn_nlevels);
696 ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
697 db->db_blkid == DMU_SPILL_BLKID ||
698 !avl_is_empty(&dn->dn_dbufs));
700 if (db->db_blkid == DMU_BONUS_BLKID) {
702 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
703 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
704 } else if (db->db_blkid == DMU_SPILL_BLKID) {
706 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
707 ASSERT0(db->db.db_offset);
709 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
712 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
713 ASSERT(dr->dr_dbuf == db);
715 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
716 ASSERT(dr->dr_dbuf == db);
719 * We can't assert that db_size matches dn_datablksz because it
720 * can be momentarily different when another thread is doing
723 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
724 dr = db->db_data_pending;
726 * It should only be modified in syncing context, so
727 * make sure we only have one copy of the data.
729 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
732 /* verify db->db_blkptr */
734 if (db->db_parent == dn->dn_dbuf) {
735 /* db is pointed to by the dnode */
736 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
737 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
738 ASSERT(db->db_parent == NULL);
740 ASSERT(db->db_parent != NULL);
741 if (db->db_blkid != DMU_SPILL_BLKID)
742 ASSERT3P(db->db_blkptr, ==,
743 &dn->dn_phys->dn_blkptr[db->db_blkid]);
745 /* db is pointed to by an indirect block */
746 int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
747 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
748 ASSERT3U(db->db_parent->db.db_object, ==,
751 * dnode_grow_indblksz() can make this fail if we don't
752 * have the struct_rwlock. XXX indblksz no longer
753 * grows. safe to do this now?
755 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
756 ASSERT3P(db->db_blkptr, ==,
757 ((blkptr_t *)db->db_parent->db.db_data +
758 db->db_blkid % epb));
762 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
763 (db->db_buf == NULL || db->db_buf->b_data) &&
764 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
765 db->db_state != DB_FILL && !dn->dn_free_txg) {
767 * If the blkptr isn't set but they have nonzero data,
768 * it had better be dirty, otherwise we'll lose that
769 * data when we evict this buffer.
771 * There is an exception to this rule for indirect blocks; in
772 * this case, if the indirect block is a hole, we fill in a few
773 * fields on each of the child blocks (importantly, birth time)
774 * to prevent hole birth times from being lost when you
775 * partially fill in a hole.
777 if (db->db_dirtycnt == 0) {
778 if (db->db_level == 0) {
779 uint64_t *buf = db->db.db_data;
782 for (i = 0; i < db->db.db_size >> 3; i++) {
786 blkptr_t *bps = db->db.db_data;
787 ASSERT3U(1 << DB_DNODE(db)->dn_indblkshift, ==,
790 * We want to verify that all the blkptrs in the
791 * indirect block are holes, but we may have
792 * automatically set up a few fields for them.
793 * We iterate through each blkptr and verify
794 * they only have those fields set.
797 i < db->db.db_size / sizeof (blkptr_t);
799 blkptr_t *bp = &bps[i];
800 ASSERT(ZIO_CHECKSUM_IS_ZERO(
803 DVA_IS_EMPTY(&bp->blk_dva[0]) &&
804 DVA_IS_EMPTY(&bp->blk_dva[1]) &&
805 DVA_IS_EMPTY(&bp->blk_dva[2]));
806 ASSERT0(bp->blk_fill);
807 ASSERT0(bp->blk_pad[0]);
808 ASSERT0(bp->blk_pad[1]);
809 ASSERT(!BP_IS_EMBEDDED(bp));
810 ASSERT(BP_IS_HOLE(bp));
811 ASSERT0(bp->blk_phys_birth);
821 dbuf_clear_data(dmu_buf_impl_t *db)
823 ASSERT(MUTEX_HELD(&db->db_mtx));
825 ASSERT3P(db->db_buf, ==, NULL);
826 db->db.db_data = NULL;
827 if (db->db_state != DB_NOFILL)
828 db->db_state = DB_UNCACHED;
832 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
834 ASSERT(MUTEX_HELD(&db->db_mtx));
838 ASSERT(buf->b_data != NULL);
839 db->db.db_data = buf->b_data;
843 * Loan out an arc_buf for read. Return the loaned arc_buf.
846 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
850 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
851 mutex_enter(&db->db_mtx);
852 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
853 int blksz = db->db.db_size;
854 spa_t *spa = db->db_objset->os_spa;
856 mutex_exit(&db->db_mtx);
857 abuf = arc_loan_buf(spa, B_FALSE, blksz);
858 bcopy(db->db.db_data, abuf->b_data, blksz);
861 arc_loan_inuse_buf(abuf, db);
864 mutex_exit(&db->db_mtx);
870 * Calculate which level n block references the data at the level 0 offset
874 dbuf_whichblock(dnode_t *dn, int64_t level, uint64_t offset)
876 if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) {
878 * The level n blkid is equal to the level 0 blkid divided by
879 * the number of level 0s in a level n block.
881 * The level 0 blkid is offset >> datablkshift =
882 * offset / 2^datablkshift.
884 * The number of level 0s in a level n is the number of block
885 * pointers in an indirect block, raised to the power of level.
886 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
887 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
889 * Thus, the level n blkid is: offset /
890 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
891 * = offset / 2^(datablkshift + level *
892 * (indblkshift - SPA_BLKPTRSHIFT))
893 * = offset >> (datablkshift + level *
894 * (indblkshift - SPA_BLKPTRSHIFT))
896 return (offset >> (dn->dn_datablkshift + level *
897 (dn->dn_indblkshift - SPA_BLKPTRSHIFT)));
899 ASSERT3U(offset, <, dn->dn_datablksz);
905 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
907 dmu_buf_impl_t *db = vdb;
909 mutex_enter(&db->db_mtx);
910 ASSERT3U(db->db_state, ==, DB_READ);
912 * All reads are synchronous, so we must have a hold on the dbuf
914 ASSERT(refcount_count(&db->db_holds) > 0);
915 ASSERT(db->db_buf == NULL);
916 ASSERT(db->db.db_data == NULL);
917 if (db->db_level == 0 && db->db_freed_in_flight) {
918 /* we were freed in flight; disregard any error */
919 arc_release(buf, db);
920 bzero(buf->b_data, db->db.db_size);
922 db->db_freed_in_flight = FALSE;
923 dbuf_set_data(db, buf);
924 db->db_state = DB_CACHED;
925 } else if (zio == NULL || zio->io_error == 0) {
926 dbuf_set_data(db, buf);
927 db->db_state = DB_CACHED;
929 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
930 ASSERT3P(db->db_buf, ==, NULL);
931 arc_buf_destroy(buf, db);
932 db->db_state = DB_UNCACHED;
934 cv_broadcast(&db->db_changed);
935 dbuf_rele_and_unlock(db, NULL);
939 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
943 arc_flags_t aflags = ARC_FLAG_NOWAIT;
947 ASSERT(!refcount_is_zero(&db->db_holds));
948 /* We need the struct_rwlock to prevent db_blkptr from changing. */
949 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
950 ASSERT(MUTEX_HELD(&db->db_mtx));
951 ASSERT(db->db_state == DB_UNCACHED);
952 ASSERT(db->db_buf == NULL);
954 if (db->db_blkid == DMU_BONUS_BLKID) {
955 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
957 ASSERT3U(bonuslen, <=, db->db.db_size);
958 db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN);
959 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
960 if (bonuslen < DN_MAX_BONUSLEN)
961 bzero(db->db.db_data, DN_MAX_BONUSLEN);
963 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
965 db->db_state = DB_CACHED;
966 mutex_exit(&db->db_mtx);
971 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
972 * processes the delete record and clears the bp while we are waiting
973 * for the dn_mtx (resulting in a "no" from block_freed).
975 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
976 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
977 BP_IS_HOLE(db->db_blkptr)))) {
978 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
980 dbuf_set_data(db, arc_alloc_buf(db->db_objset->os_spa, db, type,
982 bzero(db->db.db_data, db->db.db_size);
984 if (db->db_blkptr != NULL && db->db_level > 0 &&
985 BP_IS_HOLE(db->db_blkptr) &&
986 db->db_blkptr->blk_birth != 0) {
987 blkptr_t *bps = db->db.db_data;
988 for (int i = 0; i < ((1 <<
989 DB_DNODE(db)->dn_indblkshift) / sizeof (blkptr_t));
991 blkptr_t *bp = &bps[i];
992 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
993 1 << dn->dn_indblkshift);
995 BP_GET_LEVEL(db->db_blkptr) == 1 ?
997 BP_GET_LSIZE(db->db_blkptr));
998 BP_SET_TYPE(bp, BP_GET_TYPE(db->db_blkptr));
1000 BP_GET_LEVEL(db->db_blkptr) - 1);
1001 BP_SET_BIRTH(bp, db->db_blkptr->blk_birth, 0);
1005 db->db_state = DB_CACHED;
1006 mutex_exit(&db->db_mtx);
1012 db->db_state = DB_READ;
1013 mutex_exit(&db->db_mtx);
1015 if (DBUF_IS_L2CACHEABLE(db))
1016 aflags |= ARC_FLAG_L2CACHE;
1018 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
1019 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
1020 db->db.db_object, db->db_level, db->db_blkid);
1022 dbuf_add_ref(db, NULL);
1024 (void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr,
1025 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
1026 (flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
1031 * This is our just-in-time copy function. It makes a copy of buffers that
1032 * have been modified in a previous transaction group before we access them in
1033 * the current active group.
1035 * This function is used in three places: when we are dirtying a buffer for the
1036 * first time in a txg, when we are freeing a range in a dnode that includes
1037 * this buffer, and when we are accessing a buffer which was received compressed
1038 * and later referenced in a WRITE_BYREF record.
1040 * Note that when we are called from dbuf_free_range() we do not put a hold on
1041 * the buffer, we just traverse the active dbuf list for the dnode.
1044 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
1046 dbuf_dirty_record_t *dr = db->db_last_dirty;
1048 ASSERT(MUTEX_HELD(&db->db_mtx));
1049 ASSERT(db->db.db_data != NULL);
1050 ASSERT(db->db_level == 0);
1051 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
1054 (dr->dt.dl.dr_data !=
1055 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
1059 * If the last dirty record for this dbuf has not yet synced
1060 * and its referencing the dbuf data, either:
1061 * reset the reference to point to a new copy,
1062 * or (if there a no active holders)
1063 * just null out the current db_data pointer.
1065 ASSERT(dr->dr_txg >= txg - 2);
1066 if (db->db_blkid == DMU_BONUS_BLKID) {
1067 /* Note that the data bufs here are zio_bufs */
1068 dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
1069 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
1070 bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
1071 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
1072 int size = arc_buf_size(db->db_buf);
1073 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1074 spa_t *spa = db->db_objset->os_spa;
1075 enum zio_compress compress_type =
1076 arc_get_compression(db->db_buf);
1078 if (compress_type == ZIO_COMPRESS_OFF) {
1079 dr->dt.dl.dr_data = arc_alloc_buf(spa, db, type, size);
1081 ASSERT3U(type, ==, ARC_BUFC_DATA);
1082 dr->dt.dl.dr_data = arc_alloc_compressed_buf(spa, db,
1083 size, arc_buf_lsize(db->db_buf), compress_type);
1085 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
1088 dbuf_clear_data(db);
1093 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
1100 * We don't have to hold the mutex to check db_state because it
1101 * can't be freed while we have a hold on the buffer.
1103 ASSERT(!refcount_is_zero(&db->db_holds));
1105 if (db->db_state == DB_NOFILL)
1106 return (SET_ERROR(EIO));
1110 if ((flags & DB_RF_HAVESTRUCT) == 0)
1111 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1113 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1114 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
1115 DBUF_IS_CACHEABLE(db);
1117 mutex_enter(&db->db_mtx);
1118 if (db->db_state == DB_CACHED) {
1120 * If the arc buf is compressed, we need to decompress it to
1121 * read the data. This could happen during the "zfs receive" of
1122 * a stream which is compressed and deduplicated.
1124 if (db->db_buf != NULL &&
1125 arc_get_compression(db->db_buf) != ZIO_COMPRESS_OFF) {
1126 dbuf_fix_old_data(db,
1127 spa_syncing_txg(dmu_objset_spa(db->db_objset)));
1128 err = arc_decompress(db->db_buf);
1129 dbuf_set_data(db, db->db_buf);
1131 mutex_exit(&db->db_mtx);
1133 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1134 if ((flags & DB_RF_HAVESTRUCT) == 0)
1135 rw_exit(&dn->dn_struct_rwlock);
1137 } else if (db->db_state == DB_UNCACHED) {
1138 spa_t *spa = dn->dn_objset->os_spa;
1139 boolean_t need_wait = B_FALSE;
1142 db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) {
1143 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
1146 dbuf_read_impl(db, zio, flags);
1148 /* dbuf_read_impl has dropped db_mtx for us */
1151 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1153 if ((flags & DB_RF_HAVESTRUCT) == 0)
1154 rw_exit(&dn->dn_struct_rwlock);
1158 err = zio_wait(zio);
1161 * Another reader came in while the dbuf was in flight
1162 * between UNCACHED and CACHED. Either a writer will finish
1163 * writing the buffer (sending the dbuf to CACHED) or the
1164 * first reader's request will reach the read_done callback
1165 * and send the dbuf to CACHED. Otherwise, a failure
1166 * occurred and the dbuf went to UNCACHED.
1168 mutex_exit(&db->db_mtx);
1170 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1171 if ((flags & DB_RF_HAVESTRUCT) == 0)
1172 rw_exit(&dn->dn_struct_rwlock);
1175 /* Skip the wait per the caller's request. */
1176 mutex_enter(&db->db_mtx);
1177 if ((flags & DB_RF_NEVERWAIT) == 0) {
1178 while (db->db_state == DB_READ ||
1179 db->db_state == DB_FILL) {
1180 ASSERT(db->db_state == DB_READ ||
1181 (flags & DB_RF_HAVESTRUCT) == 0);
1182 DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *,
1184 cv_wait(&db->db_changed, &db->db_mtx);
1186 if (db->db_state == DB_UNCACHED)
1187 err = SET_ERROR(EIO);
1189 mutex_exit(&db->db_mtx);
1196 dbuf_noread(dmu_buf_impl_t *db)
1198 ASSERT(!refcount_is_zero(&db->db_holds));
1199 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1200 mutex_enter(&db->db_mtx);
1201 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1202 cv_wait(&db->db_changed, &db->db_mtx);
1203 if (db->db_state == DB_UNCACHED) {
1204 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1205 spa_t *spa = db->db_objset->os_spa;
1207 ASSERT(db->db_buf == NULL);
1208 ASSERT(db->db.db_data == NULL);
1209 dbuf_set_data(db, arc_alloc_buf(spa, db, type, db->db.db_size));
1210 db->db_state = DB_FILL;
1211 } else if (db->db_state == DB_NOFILL) {
1212 dbuf_clear_data(db);
1214 ASSERT3U(db->db_state, ==, DB_CACHED);
1216 mutex_exit(&db->db_mtx);
1220 dbuf_unoverride(dbuf_dirty_record_t *dr)
1222 dmu_buf_impl_t *db = dr->dr_dbuf;
1223 blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
1224 uint64_t txg = dr->dr_txg;
1226 ASSERT(MUTEX_HELD(&db->db_mtx));
1228 * This assert is valid because dmu_sync() expects to be called by
1229 * a zilog's get_data while holding a range lock. This call only
1230 * comes from dbuf_dirty() callers who must also hold a range lock.
1232 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
1233 ASSERT(db->db_level == 0);
1235 if (db->db_blkid == DMU_BONUS_BLKID ||
1236 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
1239 ASSERT(db->db_data_pending != dr);
1241 /* free this block */
1242 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite)
1243 zio_free(db->db_objset->os_spa, txg, bp);
1245 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1246 dr->dt.dl.dr_nopwrite = B_FALSE;
1249 * Release the already-written buffer, so we leave it in
1250 * a consistent dirty state. Note that all callers are
1251 * modifying the buffer, so they will immediately do
1252 * another (redundant) arc_release(). Therefore, leave
1253 * the buf thawed to save the effort of freezing &
1254 * immediately re-thawing it.
1256 arc_release(dr->dt.dl.dr_data, db);
1260 * Evict (if its unreferenced) or clear (if its referenced) any level-0
1261 * data blocks in the free range, so that any future readers will find
1265 dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
1268 dmu_buf_impl_t db_search;
1269 dmu_buf_impl_t *db, *db_next;
1270 uint64_t txg = tx->tx_txg;
1273 if (end_blkid > dn->dn_maxblkid &&
1274 !(start_blkid == DMU_SPILL_BLKID || end_blkid == DMU_SPILL_BLKID))
1275 end_blkid = dn->dn_maxblkid;
1276 dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid);
1278 db_search.db_level = 0;
1279 db_search.db_blkid = start_blkid;
1280 db_search.db_state = DB_SEARCH;
1282 mutex_enter(&dn->dn_dbufs_mtx);
1283 db = avl_find(&dn->dn_dbufs, &db_search, &where);
1284 ASSERT3P(db, ==, NULL);
1286 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1288 for (; db != NULL; db = db_next) {
1289 db_next = AVL_NEXT(&dn->dn_dbufs, db);
1290 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1292 if (db->db_level != 0 || db->db_blkid > end_blkid) {
1295 ASSERT3U(db->db_blkid, >=, start_blkid);
1297 /* found a level 0 buffer in the range */
1298 mutex_enter(&db->db_mtx);
1299 if (dbuf_undirty(db, tx)) {
1300 /* mutex has been dropped and dbuf destroyed */
1304 if (db->db_state == DB_UNCACHED ||
1305 db->db_state == DB_NOFILL ||
1306 db->db_state == DB_EVICTING) {
1307 ASSERT(db->db.db_data == NULL);
1308 mutex_exit(&db->db_mtx);
1311 if (db->db_state == DB_READ || db->db_state == DB_FILL) {
1312 /* will be handled in dbuf_read_done or dbuf_rele */
1313 db->db_freed_in_flight = TRUE;
1314 mutex_exit(&db->db_mtx);
1317 if (refcount_count(&db->db_holds) == 0) {
1322 /* The dbuf is referenced */
1324 if (db->db_last_dirty != NULL) {
1325 dbuf_dirty_record_t *dr = db->db_last_dirty;
1327 if (dr->dr_txg == txg) {
1329 * This buffer is "in-use", re-adjust the file
1330 * size to reflect that this buffer may
1331 * contain new data when we sync.
1333 if (db->db_blkid != DMU_SPILL_BLKID &&
1334 db->db_blkid > dn->dn_maxblkid)
1335 dn->dn_maxblkid = db->db_blkid;
1336 dbuf_unoverride(dr);
1339 * This dbuf is not dirty in the open context.
1340 * Either uncache it (if its not referenced in
1341 * the open context) or reset its contents to
1344 dbuf_fix_old_data(db, txg);
1347 /* clear the contents if its cached */
1348 if (db->db_state == DB_CACHED) {
1349 ASSERT(db->db.db_data != NULL);
1350 arc_release(db->db_buf, db);
1351 bzero(db->db.db_data, db->db.db_size);
1352 arc_buf_freeze(db->db_buf);
1355 mutex_exit(&db->db_mtx);
1357 mutex_exit(&dn->dn_dbufs_mtx);
1361 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
1363 arc_buf_t *buf, *obuf;
1364 int osize = db->db.db_size;
1365 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1368 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1373 /* XXX does *this* func really need the lock? */
1374 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1377 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1378 * is OK, because there can be no other references to the db
1379 * when we are changing its size, so no concurrent DB_FILL can
1383 * XXX we should be doing a dbuf_read, checking the return
1384 * value and returning that up to our callers
1386 dmu_buf_will_dirty(&db->db, tx);
1388 /* create the data buffer for the new block */
1389 buf = arc_alloc_buf(dn->dn_objset->os_spa, db, type, size);
1391 /* copy old block data to the new block */
1393 bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
1394 /* zero the remainder */
1396 bzero((uint8_t *)buf->b_data + osize, size - osize);
1398 mutex_enter(&db->db_mtx);
1399 dbuf_set_data(db, buf);
1400 arc_buf_destroy(obuf, db);
1401 db->db.db_size = size;
1403 if (db->db_level == 0) {
1404 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1405 db->db_last_dirty->dt.dl.dr_data = buf;
1407 mutex_exit(&db->db_mtx);
1409 dmu_objset_willuse_space(dn->dn_objset, size - osize, tx);
1414 dbuf_release_bp(dmu_buf_impl_t *db)
1416 objset_t *os = db->db_objset;
1418 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1419 ASSERT(arc_released(os->os_phys_buf) ||
1420 list_link_active(&os->os_dsl_dataset->ds_synced_link));
1421 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1423 (void) arc_release(db->db_buf, db);
1427 * We already have a dirty record for this TXG, and we are being
1431 dbuf_redirty(dbuf_dirty_record_t *dr)
1433 dmu_buf_impl_t *db = dr->dr_dbuf;
1435 ASSERT(MUTEX_HELD(&db->db_mtx));
1437 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1439 * If this buffer has already been written out,
1440 * we now need to reset its state.
1442 dbuf_unoverride(dr);
1443 if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1444 db->db_state != DB_NOFILL) {
1445 /* Already released on initial dirty, so just thaw. */
1446 ASSERT(arc_released(db->db_buf));
1447 arc_buf_thaw(db->db_buf);
1452 dbuf_dirty_record_t *
1453 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1457 dbuf_dirty_record_t **drp, *dr;
1458 int drop_struct_lock = FALSE;
1459 int txgoff = tx->tx_txg & TXG_MASK;
1461 ASSERT(tx->tx_txg != 0);
1462 ASSERT(!refcount_is_zero(&db->db_holds));
1463 DMU_TX_DIRTY_BUF(tx, db);
1468 * Shouldn't dirty a regular buffer in syncing context. Private
1469 * objects may be dirtied in syncing context, but only if they
1470 * were already pre-dirtied in open context.
1473 if (dn->dn_objset->os_dsl_dataset != NULL) {
1474 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1477 ASSERT(!dmu_tx_is_syncing(tx) ||
1478 BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1479 DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1480 dn->dn_objset->os_dsl_dataset == NULL);
1481 if (dn->dn_objset->os_dsl_dataset != NULL)
1482 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, FTAG);
1485 * We make this assert for private objects as well, but after we
1486 * check if we're already dirty. They are allowed to re-dirty
1487 * in syncing context.
1489 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1490 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1491 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1493 mutex_enter(&db->db_mtx);
1495 * XXX make this true for indirects too? The problem is that
1496 * transactions created with dmu_tx_create_assigned() from
1497 * syncing context don't bother holding ahead.
1499 ASSERT(db->db_level != 0 ||
1500 db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1501 db->db_state == DB_NOFILL);
1503 mutex_enter(&dn->dn_mtx);
1505 * Don't set dirtyctx to SYNC if we're just modifying this as we
1506 * initialize the objset.
1508 if (dn->dn_dirtyctx == DN_UNDIRTIED) {
1509 if (dn->dn_objset->os_dsl_dataset != NULL) {
1510 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1513 if (!BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1514 dn->dn_dirtyctx = (dmu_tx_is_syncing(tx) ?
1515 DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1516 ASSERT(dn->dn_dirtyctx_firstset == NULL);
1517 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1519 if (dn->dn_objset->os_dsl_dataset != NULL) {
1520 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1524 mutex_exit(&dn->dn_mtx);
1526 if (db->db_blkid == DMU_SPILL_BLKID)
1527 dn->dn_have_spill = B_TRUE;
1530 * If this buffer is already dirty, we're done.
1532 drp = &db->db_last_dirty;
1533 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1534 db->db.db_object == DMU_META_DNODE_OBJECT);
1535 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1537 if (dr && dr->dr_txg == tx->tx_txg) {
1541 mutex_exit(&db->db_mtx);
1546 * Only valid if not already dirty.
1548 ASSERT(dn->dn_object == 0 ||
1549 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1550 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1552 ASSERT3U(dn->dn_nlevels, >, db->db_level);
1555 * We should only be dirtying in syncing context if it's the
1556 * mos or we're initializing the os or it's a special object.
1557 * However, we are allowed to dirty in syncing context provided
1558 * we already dirtied it in open context. Hence we must make
1559 * this assertion only if we're not already dirty.
1562 VERIFY3U(tx->tx_txg, <=, spa_final_dirty_txg(os->os_spa));
1564 if (dn->dn_objset->os_dsl_dataset != NULL)
1565 rrw_enter(&os->os_dsl_dataset->ds_bp_rwlock, RW_READER, FTAG);
1566 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1567 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1568 if (dn->dn_objset->os_dsl_dataset != NULL)
1569 rrw_exit(&os->os_dsl_dataset->ds_bp_rwlock, FTAG);
1571 ASSERT(db->db.db_size != 0);
1573 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1575 if (db->db_blkid != DMU_BONUS_BLKID) {
1576 dmu_objset_willuse_space(os, db->db.db_size, tx);
1580 * If this buffer is dirty in an old transaction group we need
1581 * to make a copy of it so that the changes we make in this
1582 * transaction group won't leak out when we sync the older txg.
1584 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1585 if (db->db_level == 0) {
1586 void *data_old = db->db_buf;
1588 if (db->db_state != DB_NOFILL) {
1589 if (db->db_blkid == DMU_BONUS_BLKID) {
1590 dbuf_fix_old_data(db, tx->tx_txg);
1591 data_old = db->db.db_data;
1592 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1594 * Release the data buffer from the cache so
1595 * that we can modify it without impacting
1596 * possible other users of this cached data
1597 * block. Note that indirect blocks and
1598 * private objects are not released until the
1599 * syncing state (since they are only modified
1602 arc_release(db->db_buf, db);
1603 dbuf_fix_old_data(db, tx->tx_txg);
1604 data_old = db->db_buf;
1606 ASSERT(data_old != NULL);
1608 dr->dt.dl.dr_data = data_old;
1610 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1611 list_create(&dr->dt.di.dr_children,
1612 sizeof (dbuf_dirty_record_t),
1613 offsetof(dbuf_dirty_record_t, dr_dirty_node));
1615 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL)
1616 dr->dr_accounted = db->db.db_size;
1618 dr->dr_txg = tx->tx_txg;
1623 * We could have been freed_in_flight between the dbuf_noread
1624 * and dbuf_dirty. We win, as though the dbuf_noread() had
1625 * happened after the free.
1627 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1628 db->db_blkid != DMU_SPILL_BLKID) {
1629 mutex_enter(&dn->dn_mtx);
1630 if (dn->dn_free_ranges[txgoff] != NULL) {
1631 range_tree_clear(dn->dn_free_ranges[txgoff],
1634 mutex_exit(&dn->dn_mtx);
1635 db->db_freed_in_flight = FALSE;
1639 * This buffer is now part of this txg
1641 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1642 db->db_dirtycnt += 1;
1643 ASSERT3U(db->db_dirtycnt, <=, 3);
1645 mutex_exit(&db->db_mtx);
1647 if (db->db_blkid == DMU_BONUS_BLKID ||
1648 db->db_blkid == DMU_SPILL_BLKID) {
1649 mutex_enter(&dn->dn_mtx);
1650 ASSERT(!list_link_active(&dr->dr_dirty_node));
1651 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1652 mutex_exit(&dn->dn_mtx);
1653 dnode_setdirty(dn, tx);
1659 * The dn_struct_rwlock prevents db_blkptr from changing
1660 * due to a write from syncing context completing
1661 * while we are running, so we want to acquire it before
1662 * looking at db_blkptr.
1664 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1665 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1666 drop_struct_lock = TRUE;
1670 * We need to hold the dn_struct_rwlock to make this assertion,
1671 * because it protects dn_phys / dn_next_nlevels from changing.
1673 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1674 dn->dn_phys->dn_nlevels > db->db_level ||
1675 dn->dn_next_nlevels[txgoff] > db->db_level ||
1676 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1677 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1680 * If we are overwriting a dedup BP, then unless it is snapshotted,
1681 * when we get to syncing context we will need to decrement its
1682 * refcount in the DDT. Prefetch the relevant DDT block so that
1683 * syncing context won't have to wait for the i/o.
1685 ddt_prefetch(os->os_spa, db->db_blkptr);
1687 if (db->db_level == 0) {
1688 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1689 ASSERT(dn->dn_maxblkid >= db->db_blkid);
1692 if (db->db_level+1 < dn->dn_nlevels) {
1693 dmu_buf_impl_t *parent = db->db_parent;
1694 dbuf_dirty_record_t *di;
1695 int parent_held = FALSE;
1697 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1698 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1700 parent = dbuf_hold_level(dn, db->db_level+1,
1701 db->db_blkid >> epbs, FTAG);
1702 ASSERT(parent != NULL);
1705 if (drop_struct_lock)
1706 rw_exit(&dn->dn_struct_rwlock);
1707 ASSERT3U(db->db_level+1, ==, parent->db_level);
1708 di = dbuf_dirty(parent, tx);
1710 dbuf_rele(parent, FTAG);
1712 mutex_enter(&db->db_mtx);
1714 * Since we've dropped the mutex, it's possible that
1715 * dbuf_undirty() might have changed this out from under us.
1717 if (db->db_last_dirty == dr ||
1718 dn->dn_object == DMU_META_DNODE_OBJECT) {
1719 mutex_enter(&di->dt.di.dr_mtx);
1720 ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1721 ASSERT(!list_link_active(&dr->dr_dirty_node));
1722 list_insert_tail(&di->dt.di.dr_children, dr);
1723 mutex_exit(&di->dt.di.dr_mtx);
1726 mutex_exit(&db->db_mtx);
1728 ASSERT(db->db_level+1 == dn->dn_nlevels);
1729 ASSERT(db->db_blkid < dn->dn_nblkptr);
1730 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1731 mutex_enter(&dn->dn_mtx);
1732 ASSERT(!list_link_active(&dr->dr_dirty_node));
1733 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1734 mutex_exit(&dn->dn_mtx);
1735 if (drop_struct_lock)
1736 rw_exit(&dn->dn_struct_rwlock);
1739 dnode_setdirty(dn, tx);
1745 * Undirty a buffer in the transaction group referenced by the given
1746 * transaction. Return whether this evicted the dbuf.
1749 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1752 uint64_t txg = tx->tx_txg;
1753 dbuf_dirty_record_t *dr, **drp;
1758 * Due to our use of dn_nlevels below, this can only be called
1759 * in open context, unless we are operating on the MOS.
1760 * From syncing context, dn_nlevels may be different from the
1761 * dn_nlevels used when dbuf was dirtied.
1763 ASSERT(db->db_objset ==
1764 dmu_objset_pool(db->db_objset)->dp_meta_objset ||
1765 txg != spa_syncing_txg(dmu_objset_spa(db->db_objset)));
1766 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1767 ASSERT0(db->db_level);
1768 ASSERT(MUTEX_HELD(&db->db_mtx));
1771 * If this buffer is not dirty, we're done.
1773 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1774 if (dr->dr_txg <= txg)
1776 if (dr == NULL || dr->dr_txg < txg)
1778 ASSERT(dr->dr_txg == txg);
1779 ASSERT(dr->dr_dbuf == db);
1784 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1786 ASSERT(db->db.db_size != 0);
1788 dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset),
1789 dr->dr_accounted, txg);
1794 * Note that there are three places in dbuf_dirty()
1795 * where this dirty record may be put on a list.
1796 * Make sure to do a list_remove corresponding to
1797 * every one of those list_insert calls.
1799 if (dr->dr_parent) {
1800 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1801 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1802 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1803 } else if (db->db_blkid == DMU_SPILL_BLKID ||
1804 db->db_level + 1 == dn->dn_nlevels) {
1805 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1806 mutex_enter(&dn->dn_mtx);
1807 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1808 mutex_exit(&dn->dn_mtx);
1812 if (db->db_state != DB_NOFILL) {
1813 dbuf_unoverride(dr);
1815 ASSERT(db->db_buf != NULL);
1816 ASSERT(dr->dt.dl.dr_data != NULL);
1817 if (dr->dt.dl.dr_data != db->db_buf)
1818 arc_buf_destroy(dr->dt.dl.dr_data, db);
1821 kmem_free(dr, sizeof (dbuf_dirty_record_t));
1823 ASSERT(db->db_dirtycnt > 0);
1824 db->db_dirtycnt -= 1;
1826 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1827 ASSERT(db->db_state == DB_NOFILL || arc_released(db->db_buf));
1836 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
1838 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1839 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1841 ASSERT(tx->tx_txg != 0);
1842 ASSERT(!refcount_is_zero(&db->db_holds));
1845 * Quick check for dirtyness. For already dirty blocks, this
1846 * reduces runtime of this function by >90%, and overall performance
1847 * by 50% for some workloads (e.g. file deletion with indirect blocks
1850 mutex_enter(&db->db_mtx);
1851 dbuf_dirty_record_t *dr;
1852 for (dr = db->db_last_dirty;
1853 dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) {
1855 * It's possible that it is already dirty but not cached,
1856 * because there are some calls to dbuf_dirty() that don't
1857 * go through dmu_buf_will_dirty().
1859 if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) {
1860 /* This dbuf is already dirty and cached. */
1862 mutex_exit(&db->db_mtx);
1866 mutex_exit(&db->db_mtx);
1869 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
1870 rf |= DB_RF_HAVESTRUCT;
1872 (void) dbuf_read(db, NULL, rf);
1873 (void) dbuf_dirty(db, tx);
1877 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1879 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1881 db->db_state = DB_NOFILL;
1883 dmu_buf_will_fill(db_fake, tx);
1887 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1889 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1891 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1892 ASSERT(tx->tx_txg != 0);
1893 ASSERT(db->db_level == 0);
1894 ASSERT(!refcount_is_zero(&db->db_holds));
1896 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1897 dmu_tx_private_ok(tx));
1900 (void) dbuf_dirty(db, tx);
1903 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1906 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1908 mutex_enter(&db->db_mtx);
1911 if (db->db_state == DB_FILL) {
1912 if (db->db_level == 0 && db->db_freed_in_flight) {
1913 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1914 /* we were freed while filling */
1915 /* XXX dbuf_undirty? */
1916 bzero(db->db.db_data, db->db.db_size);
1917 db->db_freed_in_flight = FALSE;
1919 db->db_state = DB_CACHED;
1920 cv_broadcast(&db->db_changed);
1922 mutex_exit(&db->db_mtx);
1926 dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data,
1927 bp_embedded_type_t etype, enum zio_compress comp,
1928 int uncompressed_size, int compressed_size, int byteorder,
1931 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
1932 struct dirty_leaf *dl;
1933 dmu_object_type_t type;
1935 if (etype == BP_EMBEDDED_TYPE_DATA) {
1936 ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset),
1937 SPA_FEATURE_EMBEDDED_DATA));
1941 type = DB_DNODE(db)->dn_type;
1944 ASSERT0(db->db_level);
1945 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1947 dmu_buf_will_not_fill(dbuf, tx);
1949 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1950 dl = &db->db_last_dirty->dt.dl;
1951 encode_embedded_bp_compressed(&dl->dr_overridden_by,
1952 data, comp, uncompressed_size, compressed_size);
1953 BPE_SET_ETYPE(&dl->dr_overridden_by, etype);
1954 BP_SET_TYPE(&dl->dr_overridden_by, type);
1955 BP_SET_LEVEL(&dl->dr_overridden_by, 0);
1956 BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder);
1958 dl->dr_override_state = DR_OVERRIDDEN;
1959 dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg;
1963 * Directly assign a provided arc buf to a given dbuf if it's not referenced
1964 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1967 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
1969 ASSERT(!refcount_is_zero(&db->db_holds));
1970 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1971 ASSERT(db->db_level == 0);
1972 ASSERT3U(dbuf_is_metadata(db), ==, arc_is_metadata(buf));
1973 ASSERT(buf != NULL);
1974 ASSERT(arc_buf_lsize(buf) == db->db.db_size);
1975 ASSERT(tx->tx_txg != 0);
1977 arc_return_buf(buf, db);
1978 ASSERT(arc_released(buf));
1980 mutex_enter(&db->db_mtx);
1982 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1983 cv_wait(&db->db_changed, &db->db_mtx);
1985 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
1987 if (db->db_state == DB_CACHED &&
1988 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
1989 mutex_exit(&db->db_mtx);
1990 (void) dbuf_dirty(db, tx);
1991 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
1992 arc_buf_destroy(buf, db);
1993 xuio_stat_wbuf_copied();
1997 xuio_stat_wbuf_nocopy();
1998 if (db->db_state == DB_CACHED) {
1999 dbuf_dirty_record_t *dr = db->db_last_dirty;
2001 ASSERT(db->db_buf != NULL);
2002 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
2003 ASSERT(dr->dt.dl.dr_data == db->db_buf);
2004 if (!arc_released(db->db_buf)) {
2005 ASSERT(dr->dt.dl.dr_override_state ==
2007 arc_release(db->db_buf, db);
2009 dr->dt.dl.dr_data = buf;
2010 arc_buf_destroy(db->db_buf, db);
2011 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
2012 arc_release(db->db_buf, db);
2013 arc_buf_destroy(db->db_buf, db);
2017 ASSERT(db->db_buf == NULL);
2018 dbuf_set_data(db, buf);
2019 db->db_state = DB_FILL;
2020 mutex_exit(&db->db_mtx);
2021 (void) dbuf_dirty(db, tx);
2022 dmu_buf_fill_done(&db->db, tx);
2026 dbuf_destroy(dmu_buf_impl_t *db)
2029 dmu_buf_impl_t *parent = db->db_parent;
2030 dmu_buf_impl_t *dndb;
2032 ASSERT(MUTEX_HELD(&db->db_mtx));
2033 ASSERT(refcount_is_zero(&db->db_holds));
2035 if (db->db_buf != NULL) {
2036 arc_buf_destroy(db->db_buf, db);
2040 if (db->db_blkid == DMU_BONUS_BLKID) {
2041 ASSERT(db->db.db_data != NULL);
2042 zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
2043 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
2044 db->db_state = DB_UNCACHED;
2047 dbuf_clear_data(db);
2049 if (multilist_link_active(&db->db_cache_link)) {
2050 multilist_remove(dbuf_cache, db);
2051 (void) refcount_remove_many(&dbuf_cache_size,
2052 db->db.db_size, db);
2055 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
2056 ASSERT(db->db_data_pending == NULL);
2058 db->db_state = DB_EVICTING;
2059 db->db_blkptr = NULL;
2062 * Now that db_state is DB_EVICTING, nobody else can find this via
2063 * the hash table. We can now drop db_mtx, which allows us to
2064 * acquire the dn_dbufs_mtx.
2066 mutex_exit(&db->db_mtx);
2071 if (db->db_blkid != DMU_BONUS_BLKID) {
2072 boolean_t needlock = !MUTEX_HELD(&dn->dn_dbufs_mtx);
2074 mutex_enter(&dn->dn_dbufs_mtx);
2075 avl_remove(&dn->dn_dbufs, db);
2076 atomic_dec_32(&dn->dn_dbufs_count);
2080 mutex_exit(&dn->dn_dbufs_mtx);
2082 * Decrementing the dbuf count means that the hold corresponding
2083 * to the removed dbuf is no longer discounted in dnode_move(),
2084 * so the dnode cannot be moved until after we release the hold.
2085 * The membar_producer() ensures visibility of the decremented
2086 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
2090 db->db_dnode_handle = NULL;
2092 dbuf_hash_remove(db);
2097 ASSERT(refcount_is_zero(&db->db_holds));
2099 db->db_parent = NULL;
2101 ASSERT(db->db_buf == NULL);
2102 ASSERT(db->db.db_data == NULL);
2103 ASSERT(db->db_hash_next == NULL);
2104 ASSERT(db->db_blkptr == NULL);
2105 ASSERT(db->db_data_pending == NULL);
2106 ASSERT(!multilist_link_active(&db->db_cache_link));
2108 kmem_cache_free(dbuf_kmem_cache, db);
2109 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2112 * If this dbuf is referenced from an indirect dbuf,
2113 * decrement the ref count on the indirect dbuf.
2115 if (parent && parent != dndb)
2116 dbuf_rele(parent, db);
2120 * Note: While bpp will always be updated if the function returns success,
2121 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
2122 * this happens when the dnode is the meta-dnode, or a userused or groupused
2126 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
2127 dmu_buf_impl_t **parentp, blkptr_t **bpp)
2132 ASSERT(blkid != DMU_BONUS_BLKID);
2134 if (blkid == DMU_SPILL_BLKID) {
2135 mutex_enter(&dn->dn_mtx);
2136 if (dn->dn_have_spill &&
2137 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
2138 *bpp = &dn->dn_phys->dn_spill;
2141 dbuf_add_ref(dn->dn_dbuf, NULL);
2142 *parentp = dn->dn_dbuf;
2143 mutex_exit(&dn->dn_mtx);
2148 (dn->dn_phys->dn_nlevels == 0) ? 1 : dn->dn_phys->dn_nlevels;
2149 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2151 ASSERT3U(level * epbs, <, 64);
2152 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2154 * This assertion shouldn't trip as long as the max indirect block size
2155 * is less than 1M. The reason for this is that up to that point,
2156 * the number of levels required to address an entire object with blocks
2157 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In
2158 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55
2159 * (i.e. we can address the entire object), objects will all use at most
2160 * N-1 levels and the assertion won't overflow. However, once epbs is
2161 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be
2162 * enough to address an entire object, so objects will have 5 levels,
2163 * but then this assertion will overflow.
2165 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we
2166 * need to redo this logic to handle overflows.
2168 ASSERT(level >= nlevels ||
2169 ((nlevels - level - 1) * epbs) +
2170 highbit64(dn->dn_phys->dn_nblkptr) <= 64);
2171 if (level >= nlevels ||
2172 blkid >= ((uint64_t)dn->dn_phys->dn_nblkptr <<
2173 ((nlevels - level - 1) * epbs)) ||
2175 blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
2176 /* the buffer has no parent yet */
2177 return (SET_ERROR(ENOENT));
2178 } else if (level < nlevels-1) {
2179 /* this block is referenced from an indirect block */
2180 int err = dbuf_hold_impl(dn, level+1,
2181 blkid >> epbs, fail_sparse, FALSE, NULL, parentp);
2184 err = dbuf_read(*parentp, NULL,
2185 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
2187 dbuf_rele(*parentp, NULL);
2191 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
2192 (blkid & ((1ULL << epbs) - 1));
2193 if (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))
2194 ASSERT(BP_IS_HOLE(*bpp));
2197 /* the block is referenced from the dnode */
2198 ASSERT3U(level, ==, nlevels-1);
2199 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
2200 blkid < dn->dn_phys->dn_nblkptr);
2202 dbuf_add_ref(dn->dn_dbuf, NULL);
2203 *parentp = dn->dn_dbuf;
2205 *bpp = &dn->dn_phys->dn_blkptr[blkid];
2210 static dmu_buf_impl_t *
2211 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
2212 dmu_buf_impl_t *parent, blkptr_t *blkptr)
2214 objset_t *os = dn->dn_objset;
2215 dmu_buf_impl_t *db, *odb;
2217 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2218 ASSERT(dn->dn_type != DMU_OT_NONE);
2220 db = kmem_cache_alloc(dbuf_kmem_cache, KM_SLEEP);
2223 db->db.db_object = dn->dn_object;
2224 db->db_level = level;
2225 db->db_blkid = blkid;
2226 db->db_last_dirty = NULL;
2227 db->db_dirtycnt = 0;
2228 db->db_dnode_handle = dn->dn_handle;
2229 db->db_parent = parent;
2230 db->db_blkptr = blkptr;
2233 db->db_user_immediate_evict = FALSE;
2234 db->db_freed_in_flight = FALSE;
2235 db->db_pending_evict = FALSE;
2237 if (blkid == DMU_BONUS_BLKID) {
2238 ASSERT3P(parent, ==, dn->dn_dbuf);
2239 db->db.db_size = DN_MAX_BONUSLEN -
2240 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
2241 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
2242 db->db.db_offset = DMU_BONUS_BLKID;
2243 db->db_state = DB_UNCACHED;
2244 /* the bonus dbuf is not placed in the hash table */
2245 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2247 } else if (blkid == DMU_SPILL_BLKID) {
2248 db->db.db_size = (blkptr != NULL) ?
2249 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
2250 db->db.db_offset = 0;
2253 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
2254 db->db.db_size = blocksize;
2255 db->db.db_offset = db->db_blkid * blocksize;
2259 * Hold the dn_dbufs_mtx while we get the new dbuf
2260 * in the hash table *and* added to the dbufs list.
2261 * This prevents a possible deadlock with someone
2262 * trying to look up this dbuf before its added to the
2265 mutex_enter(&dn->dn_dbufs_mtx);
2266 db->db_state = DB_EVICTING;
2267 if ((odb = dbuf_hash_insert(db)) != NULL) {
2268 /* someone else inserted it first */
2269 kmem_cache_free(dbuf_kmem_cache, db);
2270 mutex_exit(&dn->dn_dbufs_mtx);
2273 avl_add(&dn->dn_dbufs, db);
2275 db->db_state = DB_UNCACHED;
2276 mutex_exit(&dn->dn_dbufs_mtx);
2277 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2279 if (parent && parent != dn->dn_dbuf)
2280 dbuf_add_ref(parent, db);
2282 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
2283 refcount_count(&dn->dn_holds) > 0);
2284 (void) refcount_add(&dn->dn_holds, db);
2285 atomic_inc_32(&dn->dn_dbufs_count);
2287 dprintf_dbuf(db, "db=%p\n", db);
2292 typedef struct dbuf_prefetch_arg {
2293 spa_t *dpa_spa; /* The spa to issue the prefetch in. */
2294 zbookmark_phys_t dpa_zb; /* The target block to prefetch. */
2295 int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */
2296 int dpa_curlevel; /* The current level that we're reading */
2297 dnode_t *dpa_dnode; /* The dnode associated with the prefetch */
2298 zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */
2299 zio_t *dpa_zio; /* The parent zio_t for all prefetches. */
2300 arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */
2301 } dbuf_prefetch_arg_t;
2304 * Actually issue the prefetch read for the block given.
2307 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp)
2309 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
2312 arc_flags_t aflags =
2313 dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
2315 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2316 ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level);
2317 ASSERT(dpa->dpa_zio != NULL);
2318 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL,
2319 dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2320 &aflags, &dpa->dpa_zb);
2324 * Called when an indirect block above our prefetch target is read in. This
2325 * will either read in the next indirect block down the tree or issue the actual
2326 * prefetch if the next block down is our target.
2329 dbuf_prefetch_indirect_done(zio_t *zio, arc_buf_t *abuf, void *private)
2331 dbuf_prefetch_arg_t *dpa = private;
2333 ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel);
2334 ASSERT3S(dpa->dpa_curlevel, >, 0);
2337 * The dpa_dnode is only valid if we are called with a NULL
2338 * zio. This indicates that the arc_read() returned without
2339 * first calling zio_read() to issue a physical read. Once
2340 * a physical read is made the dpa_dnode must be invalidated
2341 * as the locks guarding it may have been dropped. If the
2342 * dpa_dnode is still valid, then we want to add it to the dbuf
2343 * cache. To do so, we must hold the dbuf associated with the block
2344 * we just prefetched, read its contents so that we associate it
2345 * with an arc_buf_t, and then release it.
2348 ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel);
2349 if (zio->io_flags & ZIO_FLAG_RAW) {
2350 ASSERT3U(BP_GET_PSIZE(zio->io_bp), ==, zio->io_size);
2352 ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size);
2354 ASSERT3P(zio->io_spa, ==, dpa->dpa_spa);
2356 dpa->dpa_dnode = NULL;
2357 } else if (dpa->dpa_dnode != NULL) {
2358 uint64_t curblkid = dpa->dpa_zb.zb_blkid >>
2359 (dpa->dpa_epbs * (dpa->dpa_curlevel -
2360 dpa->dpa_zb.zb_level));
2361 dmu_buf_impl_t *db = dbuf_hold_level(dpa->dpa_dnode,
2362 dpa->dpa_curlevel, curblkid, FTAG);
2363 (void) dbuf_read(db, NULL,
2364 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_HAVESTRUCT);
2365 dbuf_rele(db, FTAG);
2368 dpa->dpa_curlevel--;
2370 uint64_t nextblkid = dpa->dpa_zb.zb_blkid >>
2371 (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level));
2372 blkptr_t *bp = ((blkptr_t *)abuf->b_data) +
2373 P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs);
2374 if (BP_IS_HOLE(bp) || (zio != NULL && zio->io_error != 0)) {
2375 kmem_free(dpa, sizeof (*dpa));
2376 } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) {
2377 ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid);
2378 dbuf_issue_final_prefetch(dpa, bp);
2379 kmem_free(dpa, sizeof (*dpa));
2381 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2382 zbookmark_phys_t zb;
2384 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2385 if (dpa->dpa_aflags & ARC_FLAG_L2CACHE)
2386 iter_aflags |= ARC_FLAG_L2CACHE;
2388 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2390 SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset,
2391 dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid);
2393 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2394 bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio,
2395 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2399 arc_buf_destroy(abuf, private);
2403 * Issue prefetch reads for the given block on the given level. If the indirect
2404 * blocks above that block are not in memory, we will read them in
2405 * asynchronously. As a result, this call never blocks waiting for a read to
2409 dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio,
2413 int epbs, nlevels, curlevel;
2416 ASSERT(blkid != DMU_BONUS_BLKID);
2417 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2419 if (blkid > dn->dn_maxblkid)
2422 if (dnode_block_freed(dn, blkid))
2426 * This dnode hasn't been written to disk yet, so there's nothing to
2429 nlevels = dn->dn_phys->dn_nlevels;
2430 if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0)
2433 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2434 if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level))
2437 dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object,
2440 mutex_exit(&db->db_mtx);
2442 * This dbuf already exists. It is either CACHED, or
2443 * (we assume) about to be read or filled.
2449 * Find the closest ancestor (indirect block) of the target block
2450 * that is present in the cache. In this indirect block, we will
2451 * find the bp that is at curlevel, curblkid.
2455 while (curlevel < nlevels - 1) {
2456 int parent_level = curlevel + 1;
2457 uint64_t parent_blkid = curblkid >> epbs;
2460 if (dbuf_hold_impl(dn, parent_level, parent_blkid,
2461 FALSE, TRUE, FTAG, &db) == 0) {
2462 blkptr_t *bpp = db->db_buf->b_data;
2463 bp = bpp[P2PHASE(curblkid, 1 << epbs)];
2464 dbuf_rele(db, FTAG);
2468 curlevel = parent_level;
2469 curblkid = parent_blkid;
2472 if (curlevel == nlevels - 1) {
2473 /* No cached indirect blocks found. */
2474 ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr);
2475 bp = dn->dn_phys->dn_blkptr[curblkid];
2477 if (BP_IS_HOLE(&bp))
2480 ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp));
2482 zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL,
2485 dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP);
2486 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
2487 SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2488 dn->dn_object, level, blkid);
2489 dpa->dpa_curlevel = curlevel;
2490 dpa->dpa_prio = prio;
2491 dpa->dpa_aflags = aflags;
2492 dpa->dpa_spa = dn->dn_objset->os_spa;
2493 dpa->dpa_dnode = dn;
2494 dpa->dpa_epbs = epbs;
2497 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2498 if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level))
2499 dpa->dpa_aflags |= ARC_FLAG_L2CACHE;
2502 * If we have the indirect just above us, no need to do the asynchronous
2503 * prefetch chain; we'll just run the last step ourselves. If we're at
2504 * a higher level, though, we want to issue the prefetches for all the
2505 * indirect blocks asynchronously, so we can go on with whatever we were
2508 if (curlevel == level) {
2509 ASSERT3U(curblkid, ==, blkid);
2510 dbuf_issue_final_prefetch(dpa, &bp);
2511 kmem_free(dpa, sizeof (*dpa));
2513 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2514 zbookmark_phys_t zb;
2516 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2517 if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level))
2518 iter_aflags |= ARC_FLAG_L2CACHE;
2520 SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2521 dn->dn_object, curlevel, curblkid);
2522 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2523 &bp, dbuf_prefetch_indirect_done, dpa, prio,
2524 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2528 * We use pio here instead of dpa_zio since it's possible that
2529 * dpa may have already been freed.
2535 * Returns with db_holds incremented, and db_mtx not held.
2536 * Note: dn_struct_rwlock must be held.
2539 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid,
2540 boolean_t fail_sparse, boolean_t fail_uncached,
2541 void *tag, dmu_buf_impl_t **dbp)
2543 dmu_buf_impl_t *db, *parent = NULL;
2545 ASSERT(blkid != DMU_BONUS_BLKID);
2546 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2547 ASSERT3U(dn->dn_nlevels, >, level);
2551 /* dbuf_find() returns with db_mtx held */
2552 db = dbuf_find(dn->dn_objset, dn->dn_object, level, blkid);
2555 blkptr_t *bp = NULL;
2559 return (SET_ERROR(ENOENT));
2561 ASSERT3P(parent, ==, NULL);
2562 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
2564 if (err == 0 && bp && BP_IS_HOLE(bp))
2565 err = SET_ERROR(ENOENT);
2568 dbuf_rele(parent, NULL);
2572 if (err && err != ENOENT)
2574 db = dbuf_create(dn, level, blkid, parent, bp);
2577 if (fail_uncached && db->db_state != DB_CACHED) {
2578 mutex_exit(&db->db_mtx);
2579 return (SET_ERROR(ENOENT));
2582 if (db->db_buf != NULL) {
2583 arc_buf_access(db->db_buf);
2584 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
2587 ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
2590 * If this buffer is currently syncing out, and we are are
2591 * still referencing it from db_data, we need to make a copy
2592 * of it in case we decide we want to dirty it again in this txg.
2594 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
2595 dn->dn_object != DMU_META_DNODE_OBJECT &&
2596 db->db_state == DB_CACHED && db->db_data_pending) {
2597 dbuf_dirty_record_t *dr = db->db_data_pending;
2599 if (dr->dt.dl.dr_data == db->db_buf) {
2600 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2603 arc_alloc_buf(dn->dn_objset->os_spa, db, type,
2605 bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
2610 if (multilist_link_active(&db->db_cache_link)) {
2611 ASSERT(refcount_is_zero(&db->db_holds));
2612 multilist_remove(dbuf_cache, db);
2613 (void) refcount_remove_many(&dbuf_cache_size,
2614 db->db.db_size, db);
2616 (void) refcount_add(&db->db_holds, tag);
2618 mutex_exit(&db->db_mtx);
2620 /* NOTE: we can't rele the parent until after we drop the db_mtx */
2622 dbuf_rele(parent, NULL);
2624 ASSERT3P(DB_DNODE(db), ==, dn);
2625 ASSERT3U(db->db_blkid, ==, blkid);
2626 ASSERT3U(db->db_level, ==, level);
2633 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
2635 return (dbuf_hold_level(dn, 0, blkid, tag));
2639 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
2642 int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db);
2643 return (err ? NULL : db);
2647 dbuf_create_bonus(dnode_t *dn)
2649 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
2651 ASSERT(dn->dn_bonus == NULL);
2652 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
2656 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
2658 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2661 if (db->db_blkid != DMU_SPILL_BLKID)
2662 return (SET_ERROR(ENOTSUP));
2664 blksz = SPA_MINBLOCKSIZE;
2665 ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset)));
2666 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
2670 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2671 dbuf_new_size(db, blksz, tx);
2672 rw_exit(&dn->dn_struct_rwlock);
2679 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
2681 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
2684 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2686 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
2688 int64_t holds = refcount_add(&db->db_holds, tag);
2689 ASSERT3S(holds, >, 1);
2692 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
2694 dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid,
2697 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2698 dmu_buf_impl_t *found_db;
2699 boolean_t result = B_FALSE;
2701 if (db->db_blkid == DMU_BONUS_BLKID)
2702 found_db = dbuf_find_bonus(os, obj);
2704 found_db = dbuf_find(os, obj, 0, blkid);
2706 if (found_db != NULL) {
2707 if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) {
2708 (void) refcount_add(&db->db_holds, tag);
2711 mutex_exit(&db->db_mtx);
2717 * If you call dbuf_rele() you had better not be referencing the dnode handle
2718 * unless you have some other direct or indirect hold on the dnode. (An indirect
2719 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2720 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2721 * dnode's parent dbuf evicting its dnode handles.
2724 dbuf_rele(dmu_buf_impl_t *db, void *tag)
2726 mutex_enter(&db->db_mtx);
2727 dbuf_rele_and_unlock(db, tag);
2731 dmu_buf_rele(dmu_buf_t *db, void *tag)
2733 dbuf_rele((dmu_buf_impl_t *)db, tag);
2737 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2738 * db_dirtycnt and db_holds to be updated atomically.
2741 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2745 ASSERT(MUTEX_HELD(&db->db_mtx));
2749 * Remove the reference to the dbuf before removing its hold on the
2750 * dnode so we can guarantee in dnode_move() that a referenced bonus
2751 * buffer has a corresponding dnode hold.
2753 holds = refcount_remove(&db->db_holds, tag);
2757 * We can't freeze indirects if there is a possibility that they
2758 * may be modified in the current syncing context.
2760 if (db->db_buf != NULL &&
2761 holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) {
2762 arc_buf_freeze(db->db_buf);
2765 if (holds == db->db_dirtycnt &&
2766 db->db_level == 0 && db->db_user_immediate_evict)
2767 dbuf_evict_user(db);
2770 if (db->db_blkid == DMU_BONUS_BLKID) {
2772 boolean_t evict_dbuf = db->db_pending_evict;
2775 * If the dnode moves here, we cannot cross this
2776 * barrier until the move completes.
2781 atomic_dec_32(&dn->dn_dbufs_count);
2784 * Decrementing the dbuf count means that the bonus
2785 * buffer's dnode hold is no longer discounted in
2786 * dnode_move(). The dnode cannot move until after
2787 * the dnode_rele() below.
2792 * Do not reference db after its lock is dropped.
2793 * Another thread may evict it.
2795 mutex_exit(&db->db_mtx);
2798 dnode_evict_bonus(dn);
2801 } else if (db->db_buf == NULL) {
2803 * This is a special case: we never associated this
2804 * dbuf with any data allocated from the ARC.
2806 ASSERT(db->db_state == DB_UNCACHED ||
2807 db->db_state == DB_NOFILL);
2809 } else if (arc_released(db->db_buf)) {
2811 * This dbuf has anonymous data associated with it.
2815 boolean_t do_arc_evict = B_FALSE;
2817 spa_t *spa = dmu_objset_spa(db->db_objset);
2819 if (!DBUF_IS_CACHEABLE(db) &&
2820 db->db_blkptr != NULL &&
2821 !BP_IS_HOLE(db->db_blkptr) &&
2822 !BP_IS_EMBEDDED(db->db_blkptr)) {
2823 do_arc_evict = B_TRUE;
2824 bp = *db->db_blkptr;
2827 if (!DBUF_IS_CACHEABLE(db) ||
2828 db->db_pending_evict) {
2830 } else if (!multilist_link_active(&db->db_cache_link)) {
2831 multilist_insert(dbuf_cache, db);
2832 (void) refcount_add_many(&dbuf_cache_size,
2833 db->db.db_size, db);
2834 mutex_exit(&db->db_mtx);
2836 dbuf_evict_notify();
2840 arc_freed(spa, &bp);
2843 mutex_exit(&db->db_mtx);
2848 #pragma weak dmu_buf_refcount = dbuf_refcount
2850 dbuf_refcount(dmu_buf_impl_t *db)
2852 return (refcount_count(&db->db_holds));
2856 dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user,
2857 dmu_buf_user_t *new_user)
2859 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2861 mutex_enter(&db->db_mtx);
2862 dbuf_verify_user(db, DBVU_NOT_EVICTING);
2863 if (db->db_user == old_user)
2864 db->db_user = new_user;
2866 old_user = db->db_user;
2867 dbuf_verify_user(db, DBVU_NOT_EVICTING);
2868 mutex_exit(&db->db_mtx);
2874 dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2876 return (dmu_buf_replace_user(db_fake, NULL, user));
2880 dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2882 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2884 db->db_user_immediate_evict = TRUE;
2885 return (dmu_buf_set_user(db_fake, user));
2889 dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2891 return (dmu_buf_replace_user(db_fake, user, NULL));
2895 dmu_buf_get_user(dmu_buf_t *db_fake)
2897 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2899 dbuf_verify_user(db, DBVU_NOT_EVICTING);
2900 return (db->db_user);
2904 dmu_buf_user_evict_wait()
2906 taskq_wait(dbu_evict_taskq);
2910 dmu_buf_get_blkptr(dmu_buf_t *db)
2912 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2913 return (dbi->db_blkptr);
2917 dmu_buf_get_objset(dmu_buf_t *db)
2919 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2920 return (dbi->db_objset);
2924 dmu_buf_dnode_enter(dmu_buf_t *db)
2926 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2927 DB_DNODE_ENTER(dbi);
2928 return (DB_DNODE(dbi));
2932 dmu_buf_dnode_exit(dmu_buf_t *db)
2934 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2939 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
2941 /* ASSERT(dmu_tx_is_syncing(tx) */
2942 ASSERT(MUTEX_HELD(&db->db_mtx));
2944 if (db->db_blkptr != NULL)
2947 if (db->db_blkid == DMU_SPILL_BLKID) {
2948 db->db_blkptr = &dn->dn_phys->dn_spill;
2949 BP_ZERO(db->db_blkptr);
2952 if (db->db_level == dn->dn_phys->dn_nlevels-1) {
2954 * This buffer was allocated at a time when there was
2955 * no available blkptrs from the dnode, or it was
2956 * inappropriate to hook it in (i.e., nlevels mis-match).
2958 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
2959 ASSERT(db->db_parent == NULL);
2960 db->db_parent = dn->dn_dbuf;
2961 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
2964 dmu_buf_impl_t *parent = db->db_parent;
2965 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2967 ASSERT(dn->dn_phys->dn_nlevels > 1);
2968 if (parent == NULL) {
2969 mutex_exit(&db->db_mtx);
2970 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2971 parent = dbuf_hold_level(dn, db->db_level + 1,
2972 db->db_blkid >> epbs, db);
2973 rw_exit(&dn->dn_struct_rwlock);
2974 mutex_enter(&db->db_mtx);
2975 db->db_parent = parent;
2977 db->db_blkptr = (blkptr_t *)parent->db.db_data +
2978 (db->db_blkid & ((1ULL << epbs) - 1));
2984 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2986 dmu_buf_impl_t *db = dr->dr_dbuf;
2990 ASSERT(dmu_tx_is_syncing(tx));
2992 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2994 mutex_enter(&db->db_mtx);
2996 ASSERT(db->db_level > 0);
2999 /* Read the block if it hasn't been read yet. */
3000 if (db->db_buf == NULL) {
3001 mutex_exit(&db->db_mtx);
3002 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
3003 mutex_enter(&db->db_mtx);
3005 ASSERT3U(db->db_state, ==, DB_CACHED);
3006 ASSERT(db->db_buf != NULL);
3010 /* Indirect block size must match what the dnode thinks it is. */
3011 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3012 dbuf_check_blkptr(dn, db);
3015 /* Provide the pending dirty record to child dbufs */
3016 db->db_data_pending = dr;
3018 mutex_exit(&db->db_mtx);
3020 dbuf_write(dr, db->db_buf, tx);
3023 mutex_enter(&dr->dt.di.dr_mtx);
3024 dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx);
3025 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3026 mutex_exit(&dr->dt.di.dr_mtx);
3031 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
3033 arc_buf_t **datap = &dr->dt.dl.dr_data;
3034 dmu_buf_impl_t *db = dr->dr_dbuf;
3037 uint64_t txg = tx->tx_txg;
3039 ASSERT(dmu_tx_is_syncing(tx));
3041 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
3043 mutex_enter(&db->db_mtx);
3045 * To be synced, we must be dirtied. But we
3046 * might have been freed after the dirty.
3048 if (db->db_state == DB_UNCACHED) {
3049 /* This buffer has been freed since it was dirtied */
3050 ASSERT(db->db.db_data == NULL);
3051 } else if (db->db_state == DB_FILL) {
3052 /* This buffer was freed and is now being re-filled */
3053 ASSERT(db->db.db_data != dr->dt.dl.dr_data);
3055 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
3062 if (db->db_blkid == DMU_SPILL_BLKID) {
3063 mutex_enter(&dn->dn_mtx);
3064 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
3065 mutex_exit(&dn->dn_mtx);
3069 * If this is a bonus buffer, simply copy the bonus data into the
3070 * dnode. It will be written out when the dnode is synced (and it
3071 * will be synced, since it must have been dirty for dbuf_sync to
3074 if (db->db_blkid == DMU_BONUS_BLKID) {
3075 dbuf_dirty_record_t **drp;
3077 ASSERT(*datap != NULL);
3078 ASSERT0(db->db_level);
3079 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN);
3080 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen);
3083 if (*datap != db->db.db_data) {
3084 zio_buf_free(*datap, DN_MAX_BONUSLEN);
3085 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
3087 db->db_data_pending = NULL;
3088 drp = &db->db_last_dirty;
3090 drp = &(*drp)->dr_next;
3091 ASSERT(dr->dr_next == NULL);
3092 ASSERT(dr->dr_dbuf == db);
3094 if (dr->dr_dbuf->db_level != 0) {
3095 list_destroy(&dr->dt.di.dr_children);
3096 mutex_destroy(&dr->dt.di.dr_mtx);
3098 kmem_free(dr, sizeof (dbuf_dirty_record_t));
3099 ASSERT(db->db_dirtycnt > 0);
3100 db->db_dirtycnt -= 1;
3101 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
3108 * This function may have dropped the db_mtx lock allowing a dmu_sync
3109 * operation to sneak in. As a result, we need to ensure that we
3110 * don't check the dr_override_state until we have returned from
3111 * dbuf_check_blkptr.
3113 dbuf_check_blkptr(dn, db);
3116 * If this buffer is in the middle of an immediate write,
3117 * wait for the synchronous IO to complete.
3119 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
3120 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
3121 cv_wait(&db->db_changed, &db->db_mtx);
3122 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
3125 if (db->db_state != DB_NOFILL &&
3126 dn->dn_object != DMU_META_DNODE_OBJECT &&
3127 refcount_count(&db->db_holds) > 1 &&
3128 dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
3129 *datap == db->db_buf) {
3131 * If this buffer is currently "in use" (i.e., there
3132 * are active holds and db_data still references it),
3133 * then make a copy before we start the write so that
3134 * any modifications from the open txg will not leak
3137 * NOTE: this copy does not need to be made for
3138 * objects only modified in the syncing context (e.g.
3139 * DNONE_DNODE blocks).
3141 int psize = arc_buf_size(*datap);
3142 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
3143 enum zio_compress compress_type = arc_get_compression(*datap);
3145 if (compress_type == ZIO_COMPRESS_OFF) {
3146 *datap = arc_alloc_buf(os->os_spa, db, type, psize);
3148 ASSERT3U(type, ==, ARC_BUFC_DATA);
3149 int lsize = arc_buf_lsize(*datap);
3150 *datap = arc_alloc_compressed_buf(os->os_spa, db,
3151 psize, lsize, compress_type);
3153 bcopy(db->db.db_data, (*datap)->b_data, psize);
3155 db->db_data_pending = dr;
3157 mutex_exit(&db->db_mtx);
3159 dbuf_write(dr, *datap, tx);
3161 ASSERT(!list_link_active(&dr->dr_dirty_node));
3162 if (dn->dn_object == DMU_META_DNODE_OBJECT) {
3163 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
3167 * Although zio_nowait() does not "wait for an IO", it does
3168 * initiate the IO. If this is an empty write it seems plausible
3169 * that the IO could actually be completed before the nowait
3170 * returns. We need to DB_DNODE_EXIT() first in case
3171 * zio_nowait() invalidates the dbuf.
3174 zio_nowait(dr->dr_zio);
3179 dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx)
3181 dbuf_dirty_record_t *dr;
3183 while (dr = list_head(list)) {
3184 if (dr->dr_zio != NULL) {
3186 * If we find an already initialized zio then we
3187 * are processing the meta-dnode, and we have finished.
3188 * The dbufs for all dnodes are put back on the list
3189 * during processing, so that we can zio_wait()
3190 * these IOs after initiating all child IOs.
3192 ASSERT3U(dr->dr_dbuf->db.db_object, ==,
3193 DMU_META_DNODE_OBJECT);
3196 if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
3197 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
3198 VERIFY3U(dr->dr_dbuf->db_level, ==, level);
3200 list_remove(list, dr);
3201 if (dr->dr_dbuf->db_level > 0)
3202 dbuf_sync_indirect(dr, tx);
3204 dbuf_sync_leaf(dr, tx);
3210 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3212 dmu_buf_impl_t *db = vdb;
3214 blkptr_t *bp = zio->io_bp;
3215 blkptr_t *bp_orig = &zio->io_bp_orig;
3216 spa_t *spa = zio->io_spa;
3221 ASSERT3P(db->db_blkptr, !=, NULL);
3222 ASSERT3P(&db->db_data_pending->dr_bp_copy, ==, bp);
3226 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
3227 dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
3228 zio->io_prev_space_delta = delta;
3230 if (bp->blk_birth != 0) {
3231 ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
3232 BP_GET_TYPE(bp) == dn->dn_type) ||
3233 (db->db_blkid == DMU_SPILL_BLKID &&
3234 BP_GET_TYPE(bp) == dn->dn_bonustype) ||
3235 BP_IS_EMBEDDED(bp));
3236 ASSERT(BP_GET_LEVEL(bp) == db->db_level);
3239 mutex_enter(&db->db_mtx);
3242 if (db->db_blkid == DMU_SPILL_BLKID) {
3243 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3244 ASSERT(!(BP_IS_HOLE(bp)) &&
3245 db->db_blkptr == &dn->dn_phys->dn_spill);
3249 if (db->db_level == 0) {
3250 mutex_enter(&dn->dn_mtx);
3251 if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
3252 db->db_blkid != DMU_SPILL_BLKID)
3253 dn->dn_phys->dn_maxblkid = db->db_blkid;
3254 mutex_exit(&dn->dn_mtx);
3256 if (dn->dn_type == DMU_OT_DNODE) {
3257 dnode_phys_t *dnp = db->db.db_data;
3258 for (i = db->db.db_size >> DNODE_SHIFT; i > 0;
3260 if (dnp->dn_type != DMU_OT_NONE)
3264 if (BP_IS_HOLE(bp)) {
3271 blkptr_t *ibp = db->db.db_data;
3272 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3273 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
3274 if (BP_IS_HOLE(ibp))
3276 fill += BP_GET_FILL(ibp);
3281 if (!BP_IS_EMBEDDED(bp))
3282 bp->blk_fill = fill;
3284 mutex_exit(&db->db_mtx);
3286 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3287 *db->db_blkptr = *bp;
3288 rw_exit(&dn->dn_struct_rwlock);
3293 * This function gets called just prior to running through the compression
3294 * stage of the zio pipeline. If we're an indirect block comprised of only
3295 * holes, then we want this indirect to be compressed away to a hole. In
3296 * order to do that we must zero out any information about the holes that
3297 * this indirect points to prior to before we try to compress it.
3300 dbuf_write_children_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3302 dmu_buf_impl_t *db = vdb;
3305 unsigned int epbs, i;
3307 ASSERT3U(db->db_level, >, 0);
3310 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3311 ASSERT3U(epbs, <, 31);
3313 /* Determine if all our children are holes */
3314 for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++) {
3315 if (!BP_IS_HOLE(bp))
3320 * If all the children are holes, then zero them all out so that
3321 * we may get compressed away.
3323 if (i == 1 << epbs) {
3325 * We only found holes. Grab the rwlock to prevent
3326 * anybody from reading the blocks we're about to
3329 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3330 bzero(db->db.db_data, db->db.db_size);
3331 rw_exit(&dn->dn_struct_rwlock);
3337 * The SPA will call this callback several times for each zio - once
3338 * for every physical child i/o (zio->io_phys_children times). This
3339 * allows the DMU to monitor the progress of each logical i/o. For example,
3340 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
3341 * block. There may be a long delay before all copies/fragments are completed,
3342 * so this callback allows us to retire dirty space gradually, as the physical
3347 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
3349 dmu_buf_impl_t *db = arg;
3350 objset_t *os = db->db_objset;
3351 dsl_pool_t *dp = dmu_objset_pool(os);
3352 dbuf_dirty_record_t *dr;
3355 dr = db->db_data_pending;
3356 ASSERT3U(dr->dr_txg, ==, zio->io_txg);
3359 * The callback will be called io_phys_children times. Retire one
3360 * portion of our dirty space each time we are called. Any rounding
3361 * error will be cleaned up by dsl_pool_sync()'s call to
3362 * dsl_pool_undirty_space().
3364 delta = dr->dr_accounted / zio->io_phys_children;
3365 dsl_pool_undirty_space(dp, delta, zio->io_txg);
3370 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
3372 dmu_buf_impl_t *db = vdb;
3373 blkptr_t *bp_orig = &zio->io_bp_orig;
3374 blkptr_t *bp = db->db_blkptr;
3375 objset_t *os = db->db_objset;
3376 dmu_tx_t *tx = os->os_synctx;
3377 dbuf_dirty_record_t **drp, *dr;
3379 ASSERT0(zio->io_error);
3380 ASSERT(db->db_blkptr == bp);
3383 * For nopwrites and rewrites we ensure that the bp matches our
3384 * original and bypass all the accounting.
3386 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
3387 ASSERT(BP_EQUAL(bp, bp_orig));
3389 dsl_dataset_t *ds = os->os_dsl_dataset;
3390 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
3391 dsl_dataset_block_born(ds, bp, tx);
3394 mutex_enter(&db->db_mtx);
3398 drp = &db->db_last_dirty;
3399 while ((dr = *drp) != db->db_data_pending)
3401 ASSERT(!list_link_active(&dr->dr_dirty_node));
3402 ASSERT(dr->dr_dbuf == db);
3403 ASSERT(dr->dr_next == NULL);
3407 if (db->db_blkid == DMU_SPILL_BLKID) {
3412 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3413 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
3414 db->db_blkptr == &dn->dn_phys->dn_spill);
3419 if (db->db_level == 0) {
3420 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
3421 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
3422 if (db->db_state != DB_NOFILL) {
3423 if (dr->dt.dl.dr_data != db->db_buf)
3424 arc_buf_destroy(dr->dt.dl.dr_data, db);
3431 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3432 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
3433 if (!BP_IS_HOLE(db->db_blkptr)) {
3435 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3436 ASSERT3U(db->db_blkid, <=,
3437 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs));
3438 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
3442 mutex_destroy(&dr->dt.di.dr_mtx);
3443 list_destroy(&dr->dt.di.dr_children);
3445 kmem_free(dr, sizeof (dbuf_dirty_record_t));
3447 cv_broadcast(&db->db_changed);
3448 ASSERT(db->db_dirtycnt > 0);
3449 db->db_dirtycnt -= 1;
3450 db->db_data_pending = NULL;
3451 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg);
3455 dbuf_write_nofill_ready(zio_t *zio)
3457 dbuf_write_ready(zio, NULL, zio->io_private);
3461 dbuf_write_nofill_done(zio_t *zio)
3463 dbuf_write_done(zio, NULL, zio->io_private);
3467 dbuf_write_override_ready(zio_t *zio)
3469 dbuf_dirty_record_t *dr = zio->io_private;
3470 dmu_buf_impl_t *db = dr->dr_dbuf;
3472 dbuf_write_ready(zio, NULL, db);
3476 dbuf_write_override_done(zio_t *zio)
3478 dbuf_dirty_record_t *dr = zio->io_private;
3479 dmu_buf_impl_t *db = dr->dr_dbuf;
3480 blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
3482 mutex_enter(&db->db_mtx);
3483 if (!BP_EQUAL(zio->io_bp, obp)) {
3484 if (!BP_IS_HOLE(obp))
3485 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
3486 arc_release(dr->dt.dl.dr_data, db);
3488 mutex_exit(&db->db_mtx);
3489 dbuf_write_done(zio, NULL, db);
3491 if (zio->io_abd != NULL)
3492 abd_put(zio->io_abd);
3495 typedef struct dbuf_remap_impl_callback_arg {
3497 uint64_t drica_blk_birth;
3499 } dbuf_remap_impl_callback_arg_t;
3502 dbuf_remap_impl_callback(uint64_t vdev, uint64_t offset, uint64_t size,
3505 dbuf_remap_impl_callback_arg_t *drica = arg;
3506 objset_t *os = drica->drica_os;
3507 spa_t *spa = dmu_objset_spa(os);
3508 dmu_tx_t *tx = drica->drica_tx;
3510 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)));
3512 if (os == spa_meta_objset(spa)) {
3513 spa_vdev_indirect_mark_obsolete(spa, vdev, offset, size, tx);
3515 dsl_dataset_block_remapped(dmu_objset_ds(os), vdev, offset,
3516 size, drica->drica_blk_birth, tx);
3521 dbuf_remap_impl(dnode_t *dn, blkptr_t *bp, dmu_tx_t *tx)
3523 blkptr_t bp_copy = *bp;
3524 spa_t *spa = dmu_objset_spa(dn->dn_objset);
3525 dbuf_remap_impl_callback_arg_t drica;
3527 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)));
3529 drica.drica_os = dn->dn_objset;
3530 drica.drica_blk_birth = bp->blk_birth;
3531 drica.drica_tx = tx;
3532 if (spa_remap_blkptr(spa, &bp_copy, dbuf_remap_impl_callback,
3535 * The struct_rwlock prevents dbuf_read_impl() from
3536 * dereferencing the BP while we are changing it. To
3537 * avoid lock contention, only grab it when we are actually
3540 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3542 rw_exit(&dn->dn_struct_rwlock);
3547 * Returns true if a dbuf_remap would modify the dbuf. We do this by attempting
3548 * to remap a copy of every bp in the dbuf.
3551 dbuf_can_remap(const dmu_buf_impl_t *db)
3553 spa_t *spa = dmu_objset_spa(db->db_objset);
3554 blkptr_t *bp = db->db.db_data;
3555 boolean_t ret = B_FALSE;
3557 ASSERT3U(db->db_level, >, 0);
3558 ASSERT3S(db->db_state, ==, DB_CACHED);
3560 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL));
3562 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
3563 for (int i = 0; i < db->db.db_size >> SPA_BLKPTRSHIFT; i++) {
3564 blkptr_t bp_copy = bp[i];
3565 if (spa_remap_blkptr(spa, &bp_copy, NULL, NULL)) {
3570 spa_config_exit(spa, SCL_VDEV, FTAG);
3576 dnode_needs_remap(const dnode_t *dn)
3578 spa_t *spa = dmu_objset_spa(dn->dn_objset);
3579 boolean_t ret = B_FALSE;
3581 if (dn->dn_phys->dn_nlevels == 0) {
3585 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL));
3587 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
3588 for (int j = 0; j < dn->dn_phys->dn_nblkptr; j++) {
3589 blkptr_t bp_copy = dn->dn_phys->dn_blkptr[j];
3590 if (spa_remap_blkptr(spa, &bp_copy, NULL, NULL)) {
3595 spa_config_exit(spa, SCL_VDEV, FTAG);
3601 * Remap any existing BP's to concrete vdevs, if possible.
3604 dbuf_remap(dnode_t *dn, dmu_buf_impl_t *db, dmu_tx_t *tx)
3606 spa_t *spa = dmu_objset_spa(db->db_objset);
3607 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)));
3609 if (!spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL))
3612 if (db->db_level > 0) {
3613 blkptr_t *bp = db->db.db_data;
3614 for (int i = 0; i < db->db.db_size >> SPA_BLKPTRSHIFT; i++) {
3615 dbuf_remap_impl(dn, &bp[i], tx);
3617 } else if (db->db.db_object == DMU_META_DNODE_OBJECT) {
3618 dnode_phys_t *dnp = db->db.db_data;
3619 ASSERT3U(db->db_dnode_handle->dnh_dnode->dn_type, ==,
3621 for (int i = 0; i < db->db.db_size >> DNODE_SHIFT; i++) {
3622 for (int j = 0; j < dnp[i].dn_nblkptr; j++) {
3623 dbuf_remap_impl(dn, &dnp[i].dn_blkptr[j], tx);
3630 /* Issue I/O to commit a dirty buffer to disk. */
3632 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
3634 dmu_buf_impl_t *db = dr->dr_dbuf;
3637 dmu_buf_impl_t *parent = db->db_parent;
3638 uint64_t txg = tx->tx_txg;
3639 zbookmark_phys_t zb;
3644 ASSERT(dmu_tx_is_syncing(tx));
3650 if (db->db_state != DB_NOFILL) {
3651 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
3653 * Private object buffers are released here rather
3654 * than in dbuf_dirty() since they are only modified
3655 * in the syncing context and we don't want the
3656 * overhead of making multiple copies of the data.
3658 if (BP_IS_HOLE(db->db_blkptr)) {
3661 dbuf_release_bp(db);
3663 dbuf_remap(dn, db, tx);
3667 if (parent != dn->dn_dbuf) {
3668 /* Our parent is an indirect block. */
3669 /* We have a dirty parent that has been scheduled for write. */
3670 ASSERT(parent && parent->db_data_pending);
3671 /* Our parent's buffer is one level closer to the dnode. */
3672 ASSERT(db->db_level == parent->db_level-1);
3674 * We're about to modify our parent's db_data by modifying
3675 * our block pointer, so the parent must be released.
3677 ASSERT(arc_released(parent->db_buf));
3678 zio = parent->db_data_pending->dr_zio;
3680 /* Our parent is the dnode itself. */
3681 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
3682 db->db_blkid != DMU_SPILL_BLKID) ||
3683 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
3684 if (db->db_blkid != DMU_SPILL_BLKID)
3685 ASSERT3P(db->db_blkptr, ==,
3686 &dn->dn_phys->dn_blkptr[db->db_blkid]);
3690 ASSERT(db->db_level == 0 || data == db->db_buf);
3691 ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
3694 SET_BOOKMARK(&zb, os->os_dsl_dataset ?
3695 os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
3696 db->db.db_object, db->db_level, db->db_blkid);
3698 if (db->db_blkid == DMU_SPILL_BLKID)
3700 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
3702 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
3706 * We copy the blkptr now (rather than when we instantiate the dirty
3707 * record), because its value can change between open context and
3708 * syncing context. We do not need to hold dn_struct_rwlock to read
3709 * db_blkptr because we are in syncing context.
3711 dr->dr_bp_copy = *db->db_blkptr;
3713 if (db->db_level == 0 &&
3714 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
3716 * The BP for this block has been provided by open context
3717 * (by dmu_sync() or dmu_buf_write_embedded()).
3719 abd_t *contents = (data != NULL) ?
3720 abd_get_from_buf(data->b_data, arc_buf_size(data)) : NULL;
3722 dr->dr_zio = zio_write(zio, os->os_spa, txg, &dr->dr_bp_copy,
3723 contents, db->db.db_size, db->db.db_size, &zp,
3724 dbuf_write_override_ready, NULL, NULL,
3725 dbuf_write_override_done,
3726 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
3727 mutex_enter(&db->db_mtx);
3728 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
3729 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
3730 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
3731 mutex_exit(&db->db_mtx);
3732 } else if (db->db_state == DB_NOFILL) {
3733 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF ||
3734 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY);
3735 dr->dr_zio = zio_write(zio, os->os_spa, txg,
3736 &dr->dr_bp_copy, NULL, db->db.db_size, db->db.db_size, &zp,
3737 dbuf_write_nofill_ready, NULL, NULL,
3738 dbuf_write_nofill_done, db,
3739 ZIO_PRIORITY_ASYNC_WRITE,
3740 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
3742 ASSERT(arc_released(data));
3745 * For indirect blocks, we want to setup the children
3746 * ready callback so that we can properly handle an indirect
3747 * block that only contains holes.
3749 arc_done_func_t *children_ready_cb = NULL;
3750 if (db->db_level != 0)
3751 children_ready_cb = dbuf_write_children_ready;
3753 dr->dr_zio = arc_write(zio, os->os_spa, txg,
3754 &dr->dr_bp_copy, data, DBUF_IS_L2CACHEABLE(db),
3755 &zp, dbuf_write_ready, children_ready_cb,
3756 dbuf_write_physdone, dbuf_write_done, db,
3757 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);