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 = 100 * 1024 * 1024;
90 /* Cap the size of the dbuf cache to log2 fraction of arc size. */
91 int dbuf_cache_max_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_max_shift, CTLFLAG_RDTUN,
142 &dbuf_cache_max_shift, 0, "dbuf 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 cap the size of the
614 * dbuf cache to 1/32nd (default) of the size of the ARC.
616 dbuf_cache_max_bytes = MIN(dbuf_cache_max_bytes,
617 arc_max_bytes() >> dbuf_cache_max_shift);
620 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
621 * configuration is not required.
623 dbu_evict_taskq = taskq_create("dbu_evict", 1, minclsyspri, 0, 0, 0);
625 dbuf_cache = multilist_create(sizeof (dmu_buf_impl_t),
626 offsetof(dmu_buf_impl_t, db_cache_link),
627 dbuf_cache_multilist_index_func);
628 refcount_create(&dbuf_cache_size);
630 tsd_create(&zfs_dbuf_evict_key, NULL);
631 dbuf_evict_thread_exit = B_FALSE;
632 mutex_init(&dbuf_evict_lock, NULL, MUTEX_DEFAULT, NULL);
633 cv_init(&dbuf_evict_cv, NULL, CV_DEFAULT, NULL);
634 dbuf_cache_evict_thread = thread_create(NULL, 0, dbuf_evict_thread,
635 NULL, 0, &p0, TS_RUN, minclsyspri);
641 dbuf_hash_table_t *h = &dbuf_hash_table;
644 for (i = 0; i < DBUF_MUTEXES; i++)
645 mutex_destroy(&h->hash_mutexes[i]);
646 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
647 kmem_cache_destroy(dbuf_kmem_cache);
648 taskq_destroy(dbu_evict_taskq);
650 mutex_enter(&dbuf_evict_lock);
651 dbuf_evict_thread_exit = B_TRUE;
652 while (dbuf_evict_thread_exit) {
653 cv_signal(&dbuf_evict_cv);
654 cv_wait(&dbuf_evict_cv, &dbuf_evict_lock);
656 mutex_exit(&dbuf_evict_lock);
657 tsd_destroy(&zfs_dbuf_evict_key);
659 mutex_destroy(&dbuf_evict_lock);
660 cv_destroy(&dbuf_evict_cv);
662 refcount_destroy(&dbuf_cache_size);
663 multilist_destroy(dbuf_cache);
672 dbuf_verify(dmu_buf_impl_t *db)
675 dbuf_dirty_record_t *dr;
677 ASSERT(MUTEX_HELD(&db->db_mtx));
679 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
682 ASSERT(db->db_objset != NULL);
686 ASSERT(db->db_parent == NULL);
687 ASSERT(db->db_blkptr == NULL);
689 ASSERT3U(db->db.db_object, ==, dn->dn_object);
690 ASSERT3P(db->db_objset, ==, dn->dn_objset);
691 ASSERT3U(db->db_level, <, dn->dn_nlevels);
692 ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
693 db->db_blkid == DMU_SPILL_BLKID ||
694 !avl_is_empty(&dn->dn_dbufs));
696 if (db->db_blkid == DMU_BONUS_BLKID) {
698 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
699 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
700 } else if (db->db_blkid == DMU_SPILL_BLKID) {
702 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
703 ASSERT0(db->db.db_offset);
705 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
708 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
709 ASSERT(dr->dr_dbuf == db);
711 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
712 ASSERT(dr->dr_dbuf == db);
715 * We can't assert that db_size matches dn_datablksz because it
716 * can be momentarily different when another thread is doing
719 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
720 dr = db->db_data_pending;
722 * It should only be modified in syncing context, so
723 * make sure we only have one copy of the data.
725 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
728 /* verify db->db_blkptr */
730 if (db->db_parent == dn->dn_dbuf) {
731 /* db is pointed to by the dnode */
732 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
733 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
734 ASSERT(db->db_parent == NULL);
736 ASSERT(db->db_parent != NULL);
737 if (db->db_blkid != DMU_SPILL_BLKID)
738 ASSERT3P(db->db_blkptr, ==,
739 &dn->dn_phys->dn_blkptr[db->db_blkid]);
741 /* db is pointed to by an indirect block */
742 int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
743 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
744 ASSERT3U(db->db_parent->db.db_object, ==,
747 * dnode_grow_indblksz() can make this fail if we don't
748 * have the struct_rwlock. XXX indblksz no longer
749 * grows. safe to do this now?
751 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
752 ASSERT3P(db->db_blkptr, ==,
753 ((blkptr_t *)db->db_parent->db.db_data +
754 db->db_blkid % epb));
758 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
759 (db->db_buf == NULL || db->db_buf->b_data) &&
760 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
761 db->db_state != DB_FILL && !dn->dn_free_txg) {
763 * If the blkptr isn't set but they have nonzero data,
764 * it had better be dirty, otherwise we'll lose that
765 * data when we evict this buffer.
767 * There is an exception to this rule for indirect blocks; in
768 * this case, if the indirect block is a hole, we fill in a few
769 * fields on each of the child blocks (importantly, birth time)
770 * to prevent hole birth times from being lost when you
771 * partially fill in a hole.
773 if (db->db_dirtycnt == 0) {
774 if (db->db_level == 0) {
775 uint64_t *buf = db->db.db_data;
778 for (i = 0; i < db->db.db_size >> 3; i++) {
782 blkptr_t *bps = db->db.db_data;
783 ASSERT3U(1 << DB_DNODE(db)->dn_indblkshift, ==,
786 * We want to verify that all the blkptrs in the
787 * indirect block are holes, but we may have
788 * automatically set up a few fields for them.
789 * We iterate through each blkptr and verify
790 * they only have those fields set.
793 i < db->db.db_size / sizeof (blkptr_t);
795 blkptr_t *bp = &bps[i];
796 ASSERT(ZIO_CHECKSUM_IS_ZERO(
799 DVA_IS_EMPTY(&bp->blk_dva[0]) &&
800 DVA_IS_EMPTY(&bp->blk_dva[1]) &&
801 DVA_IS_EMPTY(&bp->blk_dva[2]));
802 ASSERT0(bp->blk_fill);
803 ASSERT0(bp->blk_pad[0]);
804 ASSERT0(bp->blk_pad[1]);
805 ASSERT(!BP_IS_EMBEDDED(bp));
806 ASSERT(BP_IS_HOLE(bp));
807 ASSERT0(bp->blk_phys_birth);
817 dbuf_clear_data(dmu_buf_impl_t *db)
819 ASSERT(MUTEX_HELD(&db->db_mtx));
821 ASSERT3P(db->db_buf, ==, NULL);
822 db->db.db_data = NULL;
823 if (db->db_state != DB_NOFILL)
824 db->db_state = DB_UNCACHED;
828 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
830 ASSERT(MUTEX_HELD(&db->db_mtx));
834 ASSERT(buf->b_data != NULL);
835 db->db.db_data = buf->b_data;
839 * Loan out an arc_buf for read. Return the loaned arc_buf.
842 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
846 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
847 mutex_enter(&db->db_mtx);
848 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
849 int blksz = db->db.db_size;
850 spa_t *spa = db->db_objset->os_spa;
852 mutex_exit(&db->db_mtx);
853 abuf = arc_loan_buf(spa, B_FALSE, blksz);
854 bcopy(db->db.db_data, abuf->b_data, blksz);
857 arc_loan_inuse_buf(abuf, db);
860 mutex_exit(&db->db_mtx);
866 * Calculate which level n block references the data at the level 0 offset
870 dbuf_whichblock(dnode_t *dn, int64_t level, uint64_t offset)
872 if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) {
874 * The level n blkid is equal to the level 0 blkid divided by
875 * the number of level 0s in a level n block.
877 * The level 0 blkid is offset >> datablkshift =
878 * offset / 2^datablkshift.
880 * The number of level 0s in a level n is the number of block
881 * pointers in an indirect block, raised to the power of level.
882 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
883 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
885 * Thus, the level n blkid is: offset /
886 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
887 * = offset / 2^(datablkshift + level *
888 * (indblkshift - SPA_BLKPTRSHIFT))
889 * = offset >> (datablkshift + level *
890 * (indblkshift - SPA_BLKPTRSHIFT))
892 return (offset >> (dn->dn_datablkshift + level *
893 (dn->dn_indblkshift - SPA_BLKPTRSHIFT)));
895 ASSERT3U(offset, <, dn->dn_datablksz);
901 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
903 dmu_buf_impl_t *db = vdb;
905 mutex_enter(&db->db_mtx);
906 ASSERT3U(db->db_state, ==, DB_READ);
908 * All reads are synchronous, so we must have a hold on the dbuf
910 ASSERT(refcount_count(&db->db_holds) > 0);
911 ASSERT(db->db_buf == NULL);
912 ASSERT(db->db.db_data == NULL);
913 if (db->db_level == 0 && db->db_freed_in_flight) {
914 /* we were freed in flight; disregard any error */
915 arc_release(buf, db);
916 bzero(buf->b_data, db->db.db_size);
918 db->db_freed_in_flight = FALSE;
919 dbuf_set_data(db, buf);
920 db->db_state = DB_CACHED;
921 } else if (zio == NULL || zio->io_error == 0) {
922 dbuf_set_data(db, buf);
923 db->db_state = DB_CACHED;
925 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
926 ASSERT3P(db->db_buf, ==, NULL);
927 arc_buf_destroy(buf, db);
928 db->db_state = DB_UNCACHED;
930 cv_broadcast(&db->db_changed);
931 dbuf_rele_and_unlock(db, NULL);
935 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
939 arc_flags_t aflags = ARC_FLAG_NOWAIT;
943 ASSERT(!refcount_is_zero(&db->db_holds));
944 /* We need the struct_rwlock to prevent db_blkptr from changing. */
945 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
946 ASSERT(MUTEX_HELD(&db->db_mtx));
947 ASSERT(db->db_state == DB_UNCACHED);
948 ASSERT(db->db_buf == NULL);
950 if (db->db_blkid == DMU_BONUS_BLKID) {
951 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
953 ASSERT3U(bonuslen, <=, db->db.db_size);
954 db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN);
955 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
956 if (bonuslen < DN_MAX_BONUSLEN)
957 bzero(db->db.db_data, DN_MAX_BONUSLEN);
959 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
961 db->db_state = DB_CACHED;
962 mutex_exit(&db->db_mtx);
967 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
968 * processes the delete record and clears the bp while we are waiting
969 * for the dn_mtx (resulting in a "no" from block_freed).
971 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
972 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
973 BP_IS_HOLE(db->db_blkptr)))) {
974 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
976 dbuf_set_data(db, arc_alloc_buf(db->db_objset->os_spa, db, type,
978 bzero(db->db.db_data, db->db.db_size);
980 if (db->db_blkptr != NULL && db->db_level > 0 &&
981 BP_IS_HOLE(db->db_blkptr) &&
982 db->db_blkptr->blk_birth != 0) {
983 blkptr_t *bps = db->db.db_data;
984 for (int i = 0; i < ((1 <<
985 DB_DNODE(db)->dn_indblkshift) / sizeof (blkptr_t));
987 blkptr_t *bp = &bps[i];
988 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
989 1 << dn->dn_indblkshift);
991 BP_GET_LEVEL(db->db_blkptr) == 1 ?
993 BP_GET_LSIZE(db->db_blkptr));
994 BP_SET_TYPE(bp, BP_GET_TYPE(db->db_blkptr));
996 BP_GET_LEVEL(db->db_blkptr) - 1);
997 BP_SET_BIRTH(bp, db->db_blkptr->blk_birth, 0);
1001 db->db_state = DB_CACHED;
1002 mutex_exit(&db->db_mtx);
1008 db->db_state = DB_READ;
1009 mutex_exit(&db->db_mtx);
1011 if (DBUF_IS_L2CACHEABLE(db))
1012 aflags |= ARC_FLAG_L2CACHE;
1014 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
1015 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
1016 db->db.db_object, db->db_level, db->db_blkid);
1018 dbuf_add_ref(db, NULL);
1020 (void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr,
1021 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
1022 (flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
1027 * This is our just-in-time copy function. It makes a copy of buffers that
1028 * have been modified in a previous transaction group before we access them in
1029 * the current active group.
1031 * This function is used in three places: when we are dirtying a buffer for the
1032 * first time in a txg, when we are freeing a range in a dnode that includes
1033 * this buffer, and when we are accessing a buffer which was received compressed
1034 * and later referenced in a WRITE_BYREF record.
1036 * Note that when we are called from dbuf_free_range() we do not put a hold on
1037 * the buffer, we just traverse the active dbuf list for the dnode.
1040 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
1042 dbuf_dirty_record_t *dr = db->db_last_dirty;
1044 ASSERT(MUTEX_HELD(&db->db_mtx));
1045 ASSERT(db->db.db_data != NULL);
1046 ASSERT(db->db_level == 0);
1047 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
1050 (dr->dt.dl.dr_data !=
1051 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
1055 * If the last dirty record for this dbuf has not yet synced
1056 * and its referencing the dbuf data, either:
1057 * reset the reference to point to a new copy,
1058 * or (if there a no active holders)
1059 * just null out the current db_data pointer.
1061 ASSERT(dr->dr_txg >= txg - 2);
1062 if (db->db_blkid == DMU_BONUS_BLKID) {
1063 /* Note that the data bufs here are zio_bufs */
1064 dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
1065 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
1066 bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
1067 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
1068 int size = arc_buf_size(db->db_buf);
1069 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1070 spa_t *spa = db->db_objset->os_spa;
1071 enum zio_compress compress_type =
1072 arc_get_compression(db->db_buf);
1074 if (compress_type == ZIO_COMPRESS_OFF) {
1075 dr->dt.dl.dr_data = arc_alloc_buf(spa, db, type, size);
1077 ASSERT3U(type, ==, ARC_BUFC_DATA);
1078 dr->dt.dl.dr_data = arc_alloc_compressed_buf(spa, db,
1079 size, arc_buf_lsize(db->db_buf), compress_type);
1081 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
1084 dbuf_clear_data(db);
1089 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
1096 * We don't have to hold the mutex to check db_state because it
1097 * can't be freed while we have a hold on the buffer.
1099 ASSERT(!refcount_is_zero(&db->db_holds));
1101 if (db->db_state == DB_NOFILL)
1102 return (SET_ERROR(EIO));
1106 if ((flags & DB_RF_HAVESTRUCT) == 0)
1107 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1109 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1110 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
1111 DBUF_IS_CACHEABLE(db);
1113 mutex_enter(&db->db_mtx);
1114 if (db->db_state == DB_CACHED) {
1116 * If the arc buf is compressed, we need to decompress it to
1117 * read the data. This could happen during the "zfs receive" of
1118 * a stream which is compressed and deduplicated.
1120 if (db->db_buf != NULL &&
1121 arc_get_compression(db->db_buf) != ZIO_COMPRESS_OFF) {
1122 dbuf_fix_old_data(db,
1123 spa_syncing_txg(dmu_objset_spa(db->db_objset)));
1124 err = arc_decompress(db->db_buf);
1125 dbuf_set_data(db, db->db_buf);
1127 mutex_exit(&db->db_mtx);
1129 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1130 if ((flags & DB_RF_HAVESTRUCT) == 0)
1131 rw_exit(&dn->dn_struct_rwlock);
1133 } else if (db->db_state == DB_UNCACHED) {
1134 spa_t *spa = dn->dn_objset->os_spa;
1135 boolean_t need_wait = B_FALSE;
1138 db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) {
1139 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
1142 dbuf_read_impl(db, zio, flags);
1144 /* dbuf_read_impl has dropped db_mtx for us */
1147 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1149 if ((flags & DB_RF_HAVESTRUCT) == 0)
1150 rw_exit(&dn->dn_struct_rwlock);
1154 err = zio_wait(zio);
1157 * Another reader came in while the dbuf was in flight
1158 * between UNCACHED and CACHED. Either a writer will finish
1159 * writing the buffer (sending the dbuf to CACHED) or the
1160 * first reader's request will reach the read_done callback
1161 * and send the dbuf to CACHED. Otherwise, a failure
1162 * occurred and the dbuf went to UNCACHED.
1164 mutex_exit(&db->db_mtx);
1166 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1167 if ((flags & DB_RF_HAVESTRUCT) == 0)
1168 rw_exit(&dn->dn_struct_rwlock);
1171 /* Skip the wait per the caller's request. */
1172 mutex_enter(&db->db_mtx);
1173 if ((flags & DB_RF_NEVERWAIT) == 0) {
1174 while (db->db_state == DB_READ ||
1175 db->db_state == DB_FILL) {
1176 ASSERT(db->db_state == DB_READ ||
1177 (flags & DB_RF_HAVESTRUCT) == 0);
1178 DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *,
1180 cv_wait(&db->db_changed, &db->db_mtx);
1182 if (db->db_state == DB_UNCACHED)
1183 err = SET_ERROR(EIO);
1185 mutex_exit(&db->db_mtx);
1192 dbuf_noread(dmu_buf_impl_t *db)
1194 ASSERT(!refcount_is_zero(&db->db_holds));
1195 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1196 mutex_enter(&db->db_mtx);
1197 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1198 cv_wait(&db->db_changed, &db->db_mtx);
1199 if (db->db_state == DB_UNCACHED) {
1200 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1201 spa_t *spa = db->db_objset->os_spa;
1203 ASSERT(db->db_buf == NULL);
1204 ASSERT(db->db.db_data == NULL);
1205 dbuf_set_data(db, arc_alloc_buf(spa, db, type, db->db.db_size));
1206 db->db_state = DB_FILL;
1207 } else if (db->db_state == DB_NOFILL) {
1208 dbuf_clear_data(db);
1210 ASSERT3U(db->db_state, ==, DB_CACHED);
1212 mutex_exit(&db->db_mtx);
1216 dbuf_unoverride(dbuf_dirty_record_t *dr)
1218 dmu_buf_impl_t *db = dr->dr_dbuf;
1219 blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
1220 uint64_t txg = dr->dr_txg;
1222 ASSERT(MUTEX_HELD(&db->db_mtx));
1224 * This assert is valid because dmu_sync() expects to be called by
1225 * a zilog's get_data while holding a range lock. This call only
1226 * comes from dbuf_dirty() callers who must also hold a range lock.
1228 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
1229 ASSERT(db->db_level == 0);
1231 if (db->db_blkid == DMU_BONUS_BLKID ||
1232 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
1235 ASSERT(db->db_data_pending != dr);
1237 /* free this block */
1238 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite)
1239 zio_free(db->db_objset->os_spa, txg, bp);
1241 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1242 dr->dt.dl.dr_nopwrite = B_FALSE;
1245 * Release the already-written buffer, so we leave it in
1246 * a consistent dirty state. Note that all callers are
1247 * modifying the buffer, so they will immediately do
1248 * another (redundant) arc_release(). Therefore, leave
1249 * the buf thawed to save the effort of freezing &
1250 * immediately re-thawing it.
1252 arc_release(dr->dt.dl.dr_data, db);
1256 * Evict (if its unreferenced) or clear (if its referenced) any level-0
1257 * data blocks in the free range, so that any future readers will find
1261 dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
1264 dmu_buf_impl_t db_search;
1265 dmu_buf_impl_t *db, *db_next;
1266 uint64_t txg = tx->tx_txg;
1269 if (end_blkid > dn->dn_maxblkid &&
1270 !(start_blkid == DMU_SPILL_BLKID || end_blkid == DMU_SPILL_BLKID))
1271 end_blkid = dn->dn_maxblkid;
1272 dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid);
1274 db_search.db_level = 0;
1275 db_search.db_blkid = start_blkid;
1276 db_search.db_state = DB_SEARCH;
1278 mutex_enter(&dn->dn_dbufs_mtx);
1279 db = avl_find(&dn->dn_dbufs, &db_search, &where);
1280 ASSERT3P(db, ==, NULL);
1282 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1284 for (; db != NULL; db = db_next) {
1285 db_next = AVL_NEXT(&dn->dn_dbufs, db);
1286 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1288 if (db->db_level != 0 || db->db_blkid > end_blkid) {
1291 ASSERT3U(db->db_blkid, >=, start_blkid);
1293 /* found a level 0 buffer in the range */
1294 mutex_enter(&db->db_mtx);
1295 if (dbuf_undirty(db, tx)) {
1296 /* mutex has been dropped and dbuf destroyed */
1300 if (db->db_state == DB_UNCACHED ||
1301 db->db_state == DB_NOFILL ||
1302 db->db_state == DB_EVICTING) {
1303 ASSERT(db->db.db_data == NULL);
1304 mutex_exit(&db->db_mtx);
1307 if (db->db_state == DB_READ || db->db_state == DB_FILL) {
1308 /* will be handled in dbuf_read_done or dbuf_rele */
1309 db->db_freed_in_flight = TRUE;
1310 mutex_exit(&db->db_mtx);
1313 if (refcount_count(&db->db_holds) == 0) {
1318 /* The dbuf is referenced */
1320 if (db->db_last_dirty != NULL) {
1321 dbuf_dirty_record_t *dr = db->db_last_dirty;
1323 if (dr->dr_txg == txg) {
1325 * This buffer is "in-use", re-adjust the file
1326 * size to reflect that this buffer may
1327 * contain new data when we sync.
1329 if (db->db_blkid != DMU_SPILL_BLKID &&
1330 db->db_blkid > dn->dn_maxblkid)
1331 dn->dn_maxblkid = db->db_blkid;
1332 dbuf_unoverride(dr);
1335 * This dbuf is not dirty in the open context.
1336 * Either uncache it (if its not referenced in
1337 * the open context) or reset its contents to
1340 dbuf_fix_old_data(db, txg);
1343 /* clear the contents if its cached */
1344 if (db->db_state == DB_CACHED) {
1345 ASSERT(db->db.db_data != NULL);
1346 arc_release(db->db_buf, db);
1347 bzero(db->db.db_data, db->db.db_size);
1348 arc_buf_freeze(db->db_buf);
1351 mutex_exit(&db->db_mtx);
1353 mutex_exit(&dn->dn_dbufs_mtx);
1357 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
1359 arc_buf_t *buf, *obuf;
1360 int osize = db->db.db_size;
1361 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1364 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1369 /* XXX does *this* func really need the lock? */
1370 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1373 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1374 * is OK, because there can be no other references to the db
1375 * when we are changing its size, so no concurrent DB_FILL can
1379 * XXX we should be doing a dbuf_read, checking the return
1380 * value and returning that up to our callers
1382 dmu_buf_will_dirty(&db->db, tx);
1384 /* create the data buffer for the new block */
1385 buf = arc_alloc_buf(dn->dn_objset->os_spa, db, type, size);
1387 /* copy old block data to the new block */
1389 bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
1390 /* zero the remainder */
1392 bzero((uint8_t *)buf->b_data + osize, size - osize);
1394 mutex_enter(&db->db_mtx);
1395 dbuf_set_data(db, buf);
1396 arc_buf_destroy(obuf, db);
1397 db->db.db_size = size;
1399 if (db->db_level == 0) {
1400 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1401 db->db_last_dirty->dt.dl.dr_data = buf;
1403 mutex_exit(&db->db_mtx);
1405 dmu_objset_willuse_space(dn->dn_objset, size - osize, tx);
1410 dbuf_release_bp(dmu_buf_impl_t *db)
1412 objset_t *os = db->db_objset;
1414 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1415 ASSERT(arc_released(os->os_phys_buf) ||
1416 list_link_active(&os->os_dsl_dataset->ds_synced_link));
1417 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1419 (void) arc_release(db->db_buf, db);
1423 * We already have a dirty record for this TXG, and we are being
1427 dbuf_redirty(dbuf_dirty_record_t *dr)
1429 dmu_buf_impl_t *db = dr->dr_dbuf;
1431 ASSERT(MUTEX_HELD(&db->db_mtx));
1433 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1435 * If this buffer has already been written out,
1436 * we now need to reset its state.
1438 dbuf_unoverride(dr);
1439 if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1440 db->db_state != DB_NOFILL) {
1441 /* Already released on initial dirty, so just thaw. */
1442 ASSERT(arc_released(db->db_buf));
1443 arc_buf_thaw(db->db_buf);
1448 dbuf_dirty_record_t *
1449 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1453 dbuf_dirty_record_t **drp, *dr;
1454 int drop_struct_lock = FALSE;
1455 int txgoff = tx->tx_txg & TXG_MASK;
1457 ASSERT(tx->tx_txg != 0);
1458 ASSERT(!refcount_is_zero(&db->db_holds));
1459 DMU_TX_DIRTY_BUF(tx, db);
1464 * Shouldn't dirty a regular buffer in syncing context. Private
1465 * objects may be dirtied in syncing context, but only if they
1466 * were already pre-dirtied in open context.
1469 if (dn->dn_objset->os_dsl_dataset != NULL) {
1470 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1473 ASSERT(!dmu_tx_is_syncing(tx) ||
1474 BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1475 DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1476 dn->dn_objset->os_dsl_dataset == NULL);
1477 if (dn->dn_objset->os_dsl_dataset != NULL)
1478 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, FTAG);
1481 * We make this assert for private objects as well, but after we
1482 * check if we're already dirty. They are allowed to re-dirty
1483 * in syncing context.
1485 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1486 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1487 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1489 mutex_enter(&db->db_mtx);
1491 * XXX make this true for indirects too? The problem is that
1492 * transactions created with dmu_tx_create_assigned() from
1493 * syncing context don't bother holding ahead.
1495 ASSERT(db->db_level != 0 ||
1496 db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1497 db->db_state == DB_NOFILL);
1499 mutex_enter(&dn->dn_mtx);
1501 * Don't set dirtyctx to SYNC if we're just modifying this as we
1502 * initialize the objset.
1504 if (dn->dn_dirtyctx == DN_UNDIRTIED) {
1505 if (dn->dn_objset->os_dsl_dataset != NULL) {
1506 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1509 if (!BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1510 dn->dn_dirtyctx = (dmu_tx_is_syncing(tx) ?
1511 DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1512 ASSERT(dn->dn_dirtyctx_firstset == NULL);
1513 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1515 if (dn->dn_objset->os_dsl_dataset != NULL) {
1516 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1520 mutex_exit(&dn->dn_mtx);
1522 if (db->db_blkid == DMU_SPILL_BLKID)
1523 dn->dn_have_spill = B_TRUE;
1526 * If this buffer is already dirty, we're done.
1528 drp = &db->db_last_dirty;
1529 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1530 db->db.db_object == DMU_META_DNODE_OBJECT);
1531 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1533 if (dr && dr->dr_txg == tx->tx_txg) {
1537 mutex_exit(&db->db_mtx);
1542 * Only valid if not already dirty.
1544 ASSERT(dn->dn_object == 0 ||
1545 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1546 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1548 ASSERT3U(dn->dn_nlevels, >, db->db_level);
1551 * We should only be dirtying in syncing context if it's the
1552 * mos or we're initializing the os or it's a special object.
1553 * However, we are allowed to dirty in syncing context provided
1554 * we already dirtied it in open context. Hence we must make
1555 * this assertion only if we're not already dirty.
1558 VERIFY3U(tx->tx_txg, <=, spa_final_dirty_txg(os->os_spa));
1560 if (dn->dn_objset->os_dsl_dataset != NULL)
1561 rrw_enter(&os->os_dsl_dataset->ds_bp_rwlock, RW_READER, FTAG);
1562 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1563 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1564 if (dn->dn_objset->os_dsl_dataset != NULL)
1565 rrw_exit(&os->os_dsl_dataset->ds_bp_rwlock, FTAG);
1567 ASSERT(db->db.db_size != 0);
1569 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1571 if (db->db_blkid != DMU_BONUS_BLKID) {
1572 dmu_objset_willuse_space(os, db->db.db_size, tx);
1576 * If this buffer is dirty in an old transaction group we need
1577 * to make a copy of it so that the changes we make in this
1578 * transaction group won't leak out when we sync the older txg.
1580 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1581 if (db->db_level == 0) {
1582 void *data_old = db->db_buf;
1584 if (db->db_state != DB_NOFILL) {
1585 if (db->db_blkid == DMU_BONUS_BLKID) {
1586 dbuf_fix_old_data(db, tx->tx_txg);
1587 data_old = db->db.db_data;
1588 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1590 * Release the data buffer from the cache so
1591 * that we can modify it without impacting
1592 * possible other users of this cached data
1593 * block. Note that indirect blocks and
1594 * private objects are not released until the
1595 * syncing state (since they are only modified
1598 arc_release(db->db_buf, db);
1599 dbuf_fix_old_data(db, tx->tx_txg);
1600 data_old = db->db_buf;
1602 ASSERT(data_old != NULL);
1604 dr->dt.dl.dr_data = data_old;
1606 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1607 list_create(&dr->dt.di.dr_children,
1608 sizeof (dbuf_dirty_record_t),
1609 offsetof(dbuf_dirty_record_t, dr_dirty_node));
1611 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL)
1612 dr->dr_accounted = db->db.db_size;
1614 dr->dr_txg = tx->tx_txg;
1619 * We could have been freed_in_flight between the dbuf_noread
1620 * and dbuf_dirty. We win, as though the dbuf_noread() had
1621 * happened after the free.
1623 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1624 db->db_blkid != DMU_SPILL_BLKID) {
1625 mutex_enter(&dn->dn_mtx);
1626 if (dn->dn_free_ranges[txgoff] != NULL) {
1627 range_tree_clear(dn->dn_free_ranges[txgoff],
1630 mutex_exit(&dn->dn_mtx);
1631 db->db_freed_in_flight = FALSE;
1635 * This buffer is now part of this txg
1637 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1638 db->db_dirtycnt += 1;
1639 ASSERT3U(db->db_dirtycnt, <=, 3);
1641 mutex_exit(&db->db_mtx);
1643 if (db->db_blkid == DMU_BONUS_BLKID ||
1644 db->db_blkid == DMU_SPILL_BLKID) {
1645 mutex_enter(&dn->dn_mtx);
1646 ASSERT(!list_link_active(&dr->dr_dirty_node));
1647 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1648 mutex_exit(&dn->dn_mtx);
1649 dnode_setdirty(dn, tx);
1655 * The dn_struct_rwlock prevents db_blkptr from changing
1656 * due to a write from syncing context completing
1657 * while we are running, so we want to acquire it before
1658 * looking at db_blkptr.
1660 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1661 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1662 drop_struct_lock = TRUE;
1666 * We need to hold the dn_struct_rwlock to make this assertion,
1667 * because it protects dn_phys / dn_next_nlevels from changing.
1669 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1670 dn->dn_phys->dn_nlevels > db->db_level ||
1671 dn->dn_next_nlevels[txgoff] > db->db_level ||
1672 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1673 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1676 * If we are overwriting a dedup BP, then unless it is snapshotted,
1677 * when we get to syncing context we will need to decrement its
1678 * refcount in the DDT. Prefetch the relevant DDT block so that
1679 * syncing context won't have to wait for the i/o.
1681 ddt_prefetch(os->os_spa, db->db_blkptr);
1683 if (db->db_level == 0) {
1684 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1685 ASSERT(dn->dn_maxblkid >= db->db_blkid);
1688 if (db->db_level+1 < dn->dn_nlevels) {
1689 dmu_buf_impl_t *parent = db->db_parent;
1690 dbuf_dirty_record_t *di;
1691 int parent_held = FALSE;
1693 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1694 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1696 parent = dbuf_hold_level(dn, db->db_level+1,
1697 db->db_blkid >> epbs, FTAG);
1698 ASSERT(parent != NULL);
1701 if (drop_struct_lock)
1702 rw_exit(&dn->dn_struct_rwlock);
1703 ASSERT3U(db->db_level+1, ==, parent->db_level);
1704 di = dbuf_dirty(parent, tx);
1706 dbuf_rele(parent, FTAG);
1708 mutex_enter(&db->db_mtx);
1710 * Since we've dropped the mutex, it's possible that
1711 * dbuf_undirty() might have changed this out from under us.
1713 if (db->db_last_dirty == dr ||
1714 dn->dn_object == DMU_META_DNODE_OBJECT) {
1715 mutex_enter(&di->dt.di.dr_mtx);
1716 ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1717 ASSERT(!list_link_active(&dr->dr_dirty_node));
1718 list_insert_tail(&di->dt.di.dr_children, dr);
1719 mutex_exit(&di->dt.di.dr_mtx);
1722 mutex_exit(&db->db_mtx);
1724 ASSERT(db->db_level+1 == dn->dn_nlevels);
1725 ASSERT(db->db_blkid < dn->dn_nblkptr);
1726 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1727 mutex_enter(&dn->dn_mtx);
1728 ASSERT(!list_link_active(&dr->dr_dirty_node));
1729 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1730 mutex_exit(&dn->dn_mtx);
1731 if (drop_struct_lock)
1732 rw_exit(&dn->dn_struct_rwlock);
1735 dnode_setdirty(dn, tx);
1741 * Undirty a buffer in the transaction group referenced by the given
1742 * transaction. Return whether this evicted the dbuf.
1745 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1748 uint64_t txg = tx->tx_txg;
1749 dbuf_dirty_record_t *dr, **drp;
1754 * Due to our use of dn_nlevels below, this can only be called
1755 * in open context, unless we are operating on the MOS.
1756 * From syncing context, dn_nlevels may be different from the
1757 * dn_nlevels used when dbuf was dirtied.
1759 ASSERT(db->db_objset ==
1760 dmu_objset_pool(db->db_objset)->dp_meta_objset ||
1761 txg != spa_syncing_txg(dmu_objset_spa(db->db_objset)));
1762 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1763 ASSERT0(db->db_level);
1764 ASSERT(MUTEX_HELD(&db->db_mtx));
1767 * If this buffer is not dirty, we're done.
1769 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1770 if (dr->dr_txg <= txg)
1772 if (dr == NULL || dr->dr_txg < txg)
1774 ASSERT(dr->dr_txg == txg);
1775 ASSERT(dr->dr_dbuf == db);
1780 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1782 ASSERT(db->db.db_size != 0);
1784 dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset),
1785 dr->dr_accounted, txg);
1790 * Note that there are three places in dbuf_dirty()
1791 * where this dirty record may be put on a list.
1792 * Make sure to do a list_remove corresponding to
1793 * every one of those list_insert calls.
1795 if (dr->dr_parent) {
1796 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1797 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1798 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1799 } else if (db->db_blkid == DMU_SPILL_BLKID ||
1800 db->db_level + 1 == dn->dn_nlevels) {
1801 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1802 mutex_enter(&dn->dn_mtx);
1803 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1804 mutex_exit(&dn->dn_mtx);
1808 if (db->db_state != DB_NOFILL) {
1809 dbuf_unoverride(dr);
1811 ASSERT(db->db_buf != NULL);
1812 ASSERT(dr->dt.dl.dr_data != NULL);
1813 if (dr->dt.dl.dr_data != db->db_buf)
1814 arc_buf_destroy(dr->dt.dl.dr_data, db);
1817 kmem_free(dr, sizeof (dbuf_dirty_record_t));
1819 ASSERT(db->db_dirtycnt > 0);
1820 db->db_dirtycnt -= 1;
1822 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1823 ASSERT(db->db_state == DB_NOFILL || arc_released(db->db_buf));
1832 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
1834 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1835 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1837 ASSERT(tx->tx_txg != 0);
1838 ASSERT(!refcount_is_zero(&db->db_holds));
1841 * Quick check for dirtyness. For already dirty blocks, this
1842 * reduces runtime of this function by >90%, and overall performance
1843 * by 50% for some workloads (e.g. file deletion with indirect blocks
1846 mutex_enter(&db->db_mtx);
1847 dbuf_dirty_record_t *dr;
1848 for (dr = db->db_last_dirty;
1849 dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) {
1851 * It's possible that it is already dirty but not cached,
1852 * because there are some calls to dbuf_dirty() that don't
1853 * go through dmu_buf_will_dirty().
1855 if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) {
1856 /* This dbuf is already dirty and cached. */
1858 mutex_exit(&db->db_mtx);
1862 mutex_exit(&db->db_mtx);
1865 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
1866 rf |= DB_RF_HAVESTRUCT;
1868 (void) dbuf_read(db, NULL, rf);
1869 (void) dbuf_dirty(db, tx);
1873 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1875 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1877 db->db_state = DB_NOFILL;
1879 dmu_buf_will_fill(db_fake, tx);
1883 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1885 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1887 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1888 ASSERT(tx->tx_txg != 0);
1889 ASSERT(db->db_level == 0);
1890 ASSERT(!refcount_is_zero(&db->db_holds));
1892 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1893 dmu_tx_private_ok(tx));
1896 (void) dbuf_dirty(db, tx);
1899 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1902 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1904 mutex_enter(&db->db_mtx);
1907 if (db->db_state == DB_FILL) {
1908 if (db->db_level == 0 && db->db_freed_in_flight) {
1909 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1910 /* we were freed while filling */
1911 /* XXX dbuf_undirty? */
1912 bzero(db->db.db_data, db->db.db_size);
1913 db->db_freed_in_flight = FALSE;
1915 db->db_state = DB_CACHED;
1916 cv_broadcast(&db->db_changed);
1918 mutex_exit(&db->db_mtx);
1922 dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data,
1923 bp_embedded_type_t etype, enum zio_compress comp,
1924 int uncompressed_size, int compressed_size, int byteorder,
1927 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
1928 struct dirty_leaf *dl;
1929 dmu_object_type_t type;
1931 if (etype == BP_EMBEDDED_TYPE_DATA) {
1932 ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset),
1933 SPA_FEATURE_EMBEDDED_DATA));
1937 type = DB_DNODE(db)->dn_type;
1940 ASSERT0(db->db_level);
1941 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1943 dmu_buf_will_not_fill(dbuf, tx);
1945 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1946 dl = &db->db_last_dirty->dt.dl;
1947 encode_embedded_bp_compressed(&dl->dr_overridden_by,
1948 data, comp, uncompressed_size, compressed_size);
1949 BPE_SET_ETYPE(&dl->dr_overridden_by, etype);
1950 BP_SET_TYPE(&dl->dr_overridden_by, type);
1951 BP_SET_LEVEL(&dl->dr_overridden_by, 0);
1952 BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder);
1954 dl->dr_override_state = DR_OVERRIDDEN;
1955 dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg;
1959 * Directly assign a provided arc buf to a given dbuf if it's not referenced
1960 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1963 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
1965 ASSERT(!refcount_is_zero(&db->db_holds));
1966 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1967 ASSERT(db->db_level == 0);
1968 ASSERT3U(dbuf_is_metadata(db), ==, arc_is_metadata(buf));
1969 ASSERT(buf != NULL);
1970 ASSERT(arc_buf_lsize(buf) == db->db.db_size);
1971 ASSERT(tx->tx_txg != 0);
1973 arc_return_buf(buf, db);
1974 ASSERT(arc_released(buf));
1976 mutex_enter(&db->db_mtx);
1978 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1979 cv_wait(&db->db_changed, &db->db_mtx);
1981 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
1983 if (db->db_state == DB_CACHED &&
1984 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
1985 mutex_exit(&db->db_mtx);
1986 (void) dbuf_dirty(db, tx);
1987 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
1988 arc_buf_destroy(buf, db);
1989 xuio_stat_wbuf_copied();
1993 xuio_stat_wbuf_nocopy();
1994 if (db->db_state == DB_CACHED) {
1995 dbuf_dirty_record_t *dr = db->db_last_dirty;
1997 ASSERT(db->db_buf != NULL);
1998 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
1999 ASSERT(dr->dt.dl.dr_data == db->db_buf);
2000 if (!arc_released(db->db_buf)) {
2001 ASSERT(dr->dt.dl.dr_override_state ==
2003 arc_release(db->db_buf, db);
2005 dr->dt.dl.dr_data = buf;
2006 arc_buf_destroy(db->db_buf, db);
2007 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
2008 arc_release(db->db_buf, db);
2009 arc_buf_destroy(db->db_buf, db);
2013 ASSERT(db->db_buf == NULL);
2014 dbuf_set_data(db, buf);
2015 db->db_state = DB_FILL;
2016 mutex_exit(&db->db_mtx);
2017 (void) dbuf_dirty(db, tx);
2018 dmu_buf_fill_done(&db->db, tx);
2022 dbuf_destroy(dmu_buf_impl_t *db)
2025 dmu_buf_impl_t *parent = db->db_parent;
2026 dmu_buf_impl_t *dndb;
2028 ASSERT(MUTEX_HELD(&db->db_mtx));
2029 ASSERT(refcount_is_zero(&db->db_holds));
2031 if (db->db_buf != NULL) {
2032 arc_buf_destroy(db->db_buf, db);
2036 if (db->db_blkid == DMU_BONUS_BLKID) {
2037 ASSERT(db->db.db_data != NULL);
2038 zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
2039 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
2040 db->db_state = DB_UNCACHED;
2043 dbuf_clear_data(db);
2045 if (multilist_link_active(&db->db_cache_link)) {
2046 multilist_remove(dbuf_cache, db);
2047 (void) refcount_remove_many(&dbuf_cache_size,
2048 db->db.db_size, db);
2051 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
2052 ASSERT(db->db_data_pending == NULL);
2054 db->db_state = DB_EVICTING;
2055 db->db_blkptr = NULL;
2058 * Now that db_state is DB_EVICTING, nobody else can find this via
2059 * the hash table. We can now drop db_mtx, which allows us to
2060 * acquire the dn_dbufs_mtx.
2062 mutex_exit(&db->db_mtx);
2067 if (db->db_blkid != DMU_BONUS_BLKID) {
2068 boolean_t needlock = !MUTEX_HELD(&dn->dn_dbufs_mtx);
2070 mutex_enter(&dn->dn_dbufs_mtx);
2071 avl_remove(&dn->dn_dbufs, db);
2072 atomic_dec_32(&dn->dn_dbufs_count);
2076 mutex_exit(&dn->dn_dbufs_mtx);
2078 * Decrementing the dbuf count means that the hold corresponding
2079 * to the removed dbuf is no longer discounted in dnode_move(),
2080 * so the dnode cannot be moved until after we release the hold.
2081 * The membar_producer() ensures visibility of the decremented
2082 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
2086 db->db_dnode_handle = NULL;
2088 dbuf_hash_remove(db);
2093 ASSERT(refcount_is_zero(&db->db_holds));
2095 db->db_parent = NULL;
2097 ASSERT(db->db_buf == NULL);
2098 ASSERT(db->db.db_data == NULL);
2099 ASSERT(db->db_hash_next == NULL);
2100 ASSERT(db->db_blkptr == NULL);
2101 ASSERT(db->db_data_pending == NULL);
2102 ASSERT(!multilist_link_active(&db->db_cache_link));
2104 kmem_cache_free(dbuf_kmem_cache, db);
2105 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2108 * If this dbuf is referenced from an indirect dbuf,
2109 * decrement the ref count on the indirect dbuf.
2111 if (parent && parent != dndb)
2112 dbuf_rele(parent, db);
2116 * Note: While bpp will always be updated if the function returns success,
2117 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
2118 * this happens when the dnode is the meta-dnode, or a userused or groupused
2122 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
2123 dmu_buf_impl_t **parentp, blkptr_t **bpp)
2128 ASSERT(blkid != DMU_BONUS_BLKID);
2130 if (blkid == DMU_SPILL_BLKID) {
2131 mutex_enter(&dn->dn_mtx);
2132 if (dn->dn_have_spill &&
2133 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
2134 *bpp = &dn->dn_phys->dn_spill;
2137 dbuf_add_ref(dn->dn_dbuf, NULL);
2138 *parentp = dn->dn_dbuf;
2139 mutex_exit(&dn->dn_mtx);
2144 (dn->dn_phys->dn_nlevels == 0) ? 1 : dn->dn_phys->dn_nlevels;
2145 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2147 ASSERT3U(level * epbs, <, 64);
2148 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2150 * This assertion shouldn't trip as long as the max indirect block size
2151 * is less than 1M. The reason for this is that up to that point,
2152 * the number of levels required to address an entire object with blocks
2153 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In
2154 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55
2155 * (i.e. we can address the entire object), objects will all use at most
2156 * N-1 levels and the assertion won't overflow. However, once epbs is
2157 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be
2158 * enough to address an entire object, so objects will have 5 levels,
2159 * but then this assertion will overflow.
2161 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we
2162 * need to redo this logic to handle overflows.
2164 ASSERT(level >= nlevels ||
2165 ((nlevels - level - 1) * epbs) +
2166 highbit64(dn->dn_phys->dn_nblkptr) <= 64);
2167 if (level >= nlevels ||
2168 blkid >= ((uint64_t)dn->dn_phys->dn_nblkptr <<
2169 ((nlevels - level - 1) * epbs)) ||
2171 blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
2172 /* the buffer has no parent yet */
2173 return (SET_ERROR(ENOENT));
2174 } else if (level < nlevels-1) {
2175 /* this block is referenced from an indirect block */
2176 int err = dbuf_hold_impl(dn, level+1,
2177 blkid >> epbs, fail_sparse, FALSE, NULL, parentp);
2180 err = dbuf_read(*parentp, NULL,
2181 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
2183 dbuf_rele(*parentp, NULL);
2187 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
2188 (blkid & ((1ULL << epbs) - 1));
2189 if (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))
2190 ASSERT(BP_IS_HOLE(*bpp));
2193 /* the block is referenced from the dnode */
2194 ASSERT3U(level, ==, nlevels-1);
2195 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
2196 blkid < dn->dn_phys->dn_nblkptr);
2198 dbuf_add_ref(dn->dn_dbuf, NULL);
2199 *parentp = dn->dn_dbuf;
2201 *bpp = &dn->dn_phys->dn_blkptr[blkid];
2206 static dmu_buf_impl_t *
2207 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
2208 dmu_buf_impl_t *parent, blkptr_t *blkptr)
2210 objset_t *os = dn->dn_objset;
2211 dmu_buf_impl_t *db, *odb;
2213 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2214 ASSERT(dn->dn_type != DMU_OT_NONE);
2216 db = kmem_cache_alloc(dbuf_kmem_cache, KM_SLEEP);
2219 db->db.db_object = dn->dn_object;
2220 db->db_level = level;
2221 db->db_blkid = blkid;
2222 db->db_last_dirty = NULL;
2223 db->db_dirtycnt = 0;
2224 db->db_dnode_handle = dn->dn_handle;
2225 db->db_parent = parent;
2226 db->db_blkptr = blkptr;
2229 db->db_user_immediate_evict = FALSE;
2230 db->db_freed_in_flight = FALSE;
2231 db->db_pending_evict = FALSE;
2233 if (blkid == DMU_BONUS_BLKID) {
2234 ASSERT3P(parent, ==, dn->dn_dbuf);
2235 db->db.db_size = DN_MAX_BONUSLEN -
2236 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
2237 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
2238 db->db.db_offset = DMU_BONUS_BLKID;
2239 db->db_state = DB_UNCACHED;
2240 /* the bonus dbuf is not placed in the hash table */
2241 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2243 } else if (blkid == DMU_SPILL_BLKID) {
2244 db->db.db_size = (blkptr != NULL) ?
2245 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
2246 db->db.db_offset = 0;
2249 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
2250 db->db.db_size = blocksize;
2251 db->db.db_offset = db->db_blkid * blocksize;
2255 * Hold the dn_dbufs_mtx while we get the new dbuf
2256 * in the hash table *and* added to the dbufs list.
2257 * This prevents a possible deadlock with someone
2258 * trying to look up this dbuf before its added to the
2261 mutex_enter(&dn->dn_dbufs_mtx);
2262 db->db_state = DB_EVICTING;
2263 if ((odb = dbuf_hash_insert(db)) != NULL) {
2264 /* someone else inserted it first */
2265 kmem_cache_free(dbuf_kmem_cache, db);
2266 mutex_exit(&dn->dn_dbufs_mtx);
2269 avl_add(&dn->dn_dbufs, db);
2271 db->db_state = DB_UNCACHED;
2272 mutex_exit(&dn->dn_dbufs_mtx);
2273 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2275 if (parent && parent != dn->dn_dbuf)
2276 dbuf_add_ref(parent, db);
2278 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
2279 refcount_count(&dn->dn_holds) > 0);
2280 (void) refcount_add(&dn->dn_holds, db);
2281 atomic_inc_32(&dn->dn_dbufs_count);
2283 dprintf_dbuf(db, "db=%p\n", db);
2288 typedef struct dbuf_prefetch_arg {
2289 spa_t *dpa_spa; /* The spa to issue the prefetch in. */
2290 zbookmark_phys_t dpa_zb; /* The target block to prefetch. */
2291 int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */
2292 int dpa_curlevel; /* The current level that we're reading */
2293 dnode_t *dpa_dnode; /* The dnode associated with the prefetch */
2294 zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */
2295 zio_t *dpa_zio; /* The parent zio_t for all prefetches. */
2296 arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */
2297 } dbuf_prefetch_arg_t;
2300 * Actually issue the prefetch read for the block given.
2303 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp)
2305 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
2308 arc_flags_t aflags =
2309 dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
2311 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2312 ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level);
2313 ASSERT(dpa->dpa_zio != NULL);
2314 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL,
2315 dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2316 &aflags, &dpa->dpa_zb);
2320 * Called when an indirect block above our prefetch target is read in. This
2321 * will either read in the next indirect block down the tree or issue the actual
2322 * prefetch if the next block down is our target.
2325 dbuf_prefetch_indirect_done(zio_t *zio, arc_buf_t *abuf, void *private)
2327 dbuf_prefetch_arg_t *dpa = private;
2329 ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel);
2330 ASSERT3S(dpa->dpa_curlevel, >, 0);
2333 * The dpa_dnode is only valid if we are called with a NULL
2334 * zio. This indicates that the arc_read() returned without
2335 * first calling zio_read() to issue a physical read. Once
2336 * a physical read is made the dpa_dnode must be invalidated
2337 * as the locks guarding it may have been dropped. If the
2338 * dpa_dnode is still valid, then we want to add it to the dbuf
2339 * cache. To do so, we must hold the dbuf associated with the block
2340 * we just prefetched, read its contents so that we associate it
2341 * with an arc_buf_t, and then release it.
2344 ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel);
2345 if (zio->io_flags & ZIO_FLAG_RAW) {
2346 ASSERT3U(BP_GET_PSIZE(zio->io_bp), ==, zio->io_size);
2348 ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size);
2350 ASSERT3P(zio->io_spa, ==, dpa->dpa_spa);
2352 dpa->dpa_dnode = NULL;
2353 } else if (dpa->dpa_dnode != NULL) {
2354 uint64_t curblkid = dpa->dpa_zb.zb_blkid >>
2355 (dpa->dpa_epbs * (dpa->dpa_curlevel -
2356 dpa->dpa_zb.zb_level));
2357 dmu_buf_impl_t *db = dbuf_hold_level(dpa->dpa_dnode,
2358 dpa->dpa_curlevel, curblkid, FTAG);
2359 (void) dbuf_read(db, NULL,
2360 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_HAVESTRUCT);
2361 dbuf_rele(db, FTAG);
2364 dpa->dpa_curlevel--;
2366 uint64_t nextblkid = dpa->dpa_zb.zb_blkid >>
2367 (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level));
2368 blkptr_t *bp = ((blkptr_t *)abuf->b_data) +
2369 P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs);
2370 if (BP_IS_HOLE(bp) || (zio != NULL && zio->io_error != 0)) {
2371 kmem_free(dpa, sizeof (*dpa));
2372 } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) {
2373 ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid);
2374 dbuf_issue_final_prefetch(dpa, bp);
2375 kmem_free(dpa, sizeof (*dpa));
2377 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2378 zbookmark_phys_t zb;
2380 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2381 if (dpa->dpa_aflags & ARC_FLAG_L2CACHE)
2382 iter_aflags |= ARC_FLAG_L2CACHE;
2384 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2386 SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset,
2387 dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid);
2389 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2390 bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio,
2391 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2395 arc_buf_destroy(abuf, private);
2399 * Issue prefetch reads for the given block on the given level. If the indirect
2400 * blocks above that block are not in memory, we will read them in
2401 * asynchronously. As a result, this call never blocks waiting for a read to
2405 dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio,
2409 int epbs, nlevels, curlevel;
2412 ASSERT(blkid != DMU_BONUS_BLKID);
2413 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2415 if (blkid > dn->dn_maxblkid)
2418 if (dnode_block_freed(dn, blkid))
2422 * This dnode hasn't been written to disk yet, so there's nothing to
2425 nlevels = dn->dn_phys->dn_nlevels;
2426 if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0)
2429 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2430 if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level))
2433 dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object,
2436 mutex_exit(&db->db_mtx);
2438 * This dbuf already exists. It is either CACHED, or
2439 * (we assume) about to be read or filled.
2445 * Find the closest ancestor (indirect block) of the target block
2446 * that is present in the cache. In this indirect block, we will
2447 * find the bp that is at curlevel, curblkid.
2451 while (curlevel < nlevels - 1) {
2452 int parent_level = curlevel + 1;
2453 uint64_t parent_blkid = curblkid >> epbs;
2456 if (dbuf_hold_impl(dn, parent_level, parent_blkid,
2457 FALSE, TRUE, FTAG, &db) == 0) {
2458 blkptr_t *bpp = db->db_buf->b_data;
2459 bp = bpp[P2PHASE(curblkid, 1 << epbs)];
2460 dbuf_rele(db, FTAG);
2464 curlevel = parent_level;
2465 curblkid = parent_blkid;
2468 if (curlevel == nlevels - 1) {
2469 /* No cached indirect blocks found. */
2470 ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr);
2471 bp = dn->dn_phys->dn_blkptr[curblkid];
2473 if (BP_IS_HOLE(&bp))
2476 ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp));
2478 zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL,
2481 dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP);
2482 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
2483 SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2484 dn->dn_object, level, blkid);
2485 dpa->dpa_curlevel = curlevel;
2486 dpa->dpa_prio = prio;
2487 dpa->dpa_aflags = aflags;
2488 dpa->dpa_spa = dn->dn_objset->os_spa;
2489 dpa->dpa_dnode = dn;
2490 dpa->dpa_epbs = epbs;
2493 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2494 if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level))
2495 dpa->dpa_aflags |= ARC_FLAG_L2CACHE;
2498 * If we have the indirect just above us, no need to do the asynchronous
2499 * prefetch chain; we'll just run the last step ourselves. If we're at
2500 * a higher level, though, we want to issue the prefetches for all the
2501 * indirect blocks asynchronously, so we can go on with whatever we were
2504 if (curlevel == level) {
2505 ASSERT3U(curblkid, ==, blkid);
2506 dbuf_issue_final_prefetch(dpa, &bp);
2507 kmem_free(dpa, sizeof (*dpa));
2509 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2510 zbookmark_phys_t zb;
2512 /* flag if L2ARC eligible, l2arc_noprefetch then decides */
2513 if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level))
2514 iter_aflags |= ARC_FLAG_L2CACHE;
2516 SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2517 dn->dn_object, curlevel, curblkid);
2518 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2519 &bp, dbuf_prefetch_indirect_done, dpa, prio,
2520 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2524 * We use pio here instead of dpa_zio since it's possible that
2525 * dpa may have already been freed.
2531 * Returns with db_holds incremented, and db_mtx not held.
2532 * Note: dn_struct_rwlock must be held.
2535 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid,
2536 boolean_t fail_sparse, boolean_t fail_uncached,
2537 void *tag, dmu_buf_impl_t **dbp)
2539 dmu_buf_impl_t *db, *parent = NULL;
2541 ASSERT(blkid != DMU_BONUS_BLKID);
2542 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2543 ASSERT3U(dn->dn_nlevels, >, level);
2547 /* dbuf_find() returns with db_mtx held */
2548 db = dbuf_find(dn->dn_objset, dn->dn_object, level, blkid);
2551 blkptr_t *bp = NULL;
2555 return (SET_ERROR(ENOENT));
2557 ASSERT3P(parent, ==, NULL);
2558 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
2560 if (err == 0 && bp && BP_IS_HOLE(bp))
2561 err = SET_ERROR(ENOENT);
2564 dbuf_rele(parent, NULL);
2568 if (err && err != ENOENT)
2570 db = dbuf_create(dn, level, blkid, parent, bp);
2573 if (fail_uncached && db->db_state != DB_CACHED) {
2574 mutex_exit(&db->db_mtx);
2575 return (SET_ERROR(ENOENT));
2578 if (db->db_buf != NULL)
2579 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
2581 ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
2584 * If this buffer is currently syncing out, and we are are
2585 * still referencing it from db_data, we need to make a copy
2586 * of it in case we decide we want to dirty it again in this txg.
2588 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
2589 dn->dn_object != DMU_META_DNODE_OBJECT &&
2590 db->db_state == DB_CACHED && db->db_data_pending) {
2591 dbuf_dirty_record_t *dr = db->db_data_pending;
2593 if (dr->dt.dl.dr_data == db->db_buf) {
2594 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2597 arc_alloc_buf(dn->dn_objset->os_spa, db, type,
2599 bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
2604 if (multilist_link_active(&db->db_cache_link)) {
2605 ASSERT(refcount_is_zero(&db->db_holds));
2606 multilist_remove(dbuf_cache, db);
2607 (void) refcount_remove_many(&dbuf_cache_size,
2608 db->db.db_size, db);
2610 (void) refcount_add(&db->db_holds, tag);
2612 mutex_exit(&db->db_mtx);
2614 /* NOTE: we can't rele the parent until after we drop the db_mtx */
2616 dbuf_rele(parent, NULL);
2618 ASSERT3P(DB_DNODE(db), ==, dn);
2619 ASSERT3U(db->db_blkid, ==, blkid);
2620 ASSERT3U(db->db_level, ==, level);
2627 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
2629 return (dbuf_hold_level(dn, 0, blkid, tag));
2633 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
2636 int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db);
2637 return (err ? NULL : db);
2641 dbuf_create_bonus(dnode_t *dn)
2643 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
2645 ASSERT(dn->dn_bonus == NULL);
2646 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
2650 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
2652 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2655 if (db->db_blkid != DMU_SPILL_BLKID)
2656 return (SET_ERROR(ENOTSUP));
2658 blksz = SPA_MINBLOCKSIZE;
2659 ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset)));
2660 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
2664 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2665 dbuf_new_size(db, blksz, tx);
2666 rw_exit(&dn->dn_struct_rwlock);
2673 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
2675 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
2678 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2680 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
2682 int64_t holds = refcount_add(&db->db_holds, tag);
2683 ASSERT3S(holds, >, 1);
2686 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
2688 dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid,
2691 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2692 dmu_buf_impl_t *found_db;
2693 boolean_t result = B_FALSE;
2695 if (db->db_blkid == DMU_BONUS_BLKID)
2696 found_db = dbuf_find_bonus(os, obj);
2698 found_db = dbuf_find(os, obj, 0, blkid);
2700 if (found_db != NULL) {
2701 if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) {
2702 (void) refcount_add(&db->db_holds, tag);
2705 mutex_exit(&db->db_mtx);
2711 * If you call dbuf_rele() you had better not be referencing the dnode handle
2712 * unless you have some other direct or indirect hold on the dnode. (An indirect
2713 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2714 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2715 * dnode's parent dbuf evicting its dnode handles.
2718 dbuf_rele(dmu_buf_impl_t *db, void *tag)
2720 mutex_enter(&db->db_mtx);
2721 dbuf_rele_and_unlock(db, tag);
2725 dmu_buf_rele(dmu_buf_t *db, void *tag)
2727 dbuf_rele((dmu_buf_impl_t *)db, tag);
2731 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2732 * db_dirtycnt and db_holds to be updated atomically.
2735 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2739 ASSERT(MUTEX_HELD(&db->db_mtx));
2743 * Remove the reference to the dbuf before removing its hold on the
2744 * dnode so we can guarantee in dnode_move() that a referenced bonus
2745 * buffer has a corresponding dnode hold.
2747 holds = refcount_remove(&db->db_holds, tag);
2751 * We can't freeze indirects if there is a possibility that they
2752 * may be modified in the current syncing context.
2754 if (db->db_buf != NULL &&
2755 holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) {
2756 arc_buf_freeze(db->db_buf);
2759 if (holds == db->db_dirtycnt &&
2760 db->db_level == 0 && db->db_user_immediate_evict)
2761 dbuf_evict_user(db);
2764 if (db->db_blkid == DMU_BONUS_BLKID) {
2766 boolean_t evict_dbuf = db->db_pending_evict;
2769 * If the dnode moves here, we cannot cross this
2770 * barrier until the move completes.
2775 atomic_dec_32(&dn->dn_dbufs_count);
2778 * Decrementing the dbuf count means that the bonus
2779 * buffer's dnode hold is no longer discounted in
2780 * dnode_move(). The dnode cannot move until after
2781 * the dnode_rele() below.
2786 * Do not reference db after its lock is dropped.
2787 * Another thread may evict it.
2789 mutex_exit(&db->db_mtx);
2792 dnode_evict_bonus(dn);
2795 } else if (db->db_buf == NULL) {
2797 * This is a special case: we never associated this
2798 * dbuf with any data allocated from the ARC.
2800 ASSERT(db->db_state == DB_UNCACHED ||
2801 db->db_state == DB_NOFILL);
2803 } else if (arc_released(db->db_buf)) {
2805 * This dbuf has anonymous data associated with it.
2809 boolean_t do_arc_evict = B_FALSE;
2811 spa_t *spa = dmu_objset_spa(db->db_objset);
2813 if (!DBUF_IS_CACHEABLE(db) &&
2814 db->db_blkptr != NULL &&
2815 !BP_IS_HOLE(db->db_blkptr) &&
2816 !BP_IS_EMBEDDED(db->db_blkptr)) {
2817 do_arc_evict = B_TRUE;
2818 bp = *db->db_blkptr;
2821 if (!DBUF_IS_CACHEABLE(db) ||
2822 db->db_pending_evict) {
2824 } else if (!multilist_link_active(&db->db_cache_link)) {
2825 multilist_insert(dbuf_cache, db);
2826 (void) refcount_add_many(&dbuf_cache_size,
2827 db->db.db_size, db);
2828 mutex_exit(&db->db_mtx);
2830 dbuf_evict_notify();
2834 arc_freed(spa, &bp);
2837 mutex_exit(&db->db_mtx);
2842 #pragma weak dmu_buf_refcount = dbuf_refcount
2844 dbuf_refcount(dmu_buf_impl_t *db)
2846 return (refcount_count(&db->db_holds));
2850 dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user,
2851 dmu_buf_user_t *new_user)
2853 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2855 mutex_enter(&db->db_mtx);
2856 dbuf_verify_user(db, DBVU_NOT_EVICTING);
2857 if (db->db_user == old_user)
2858 db->db_user = new_user;
2860 old_user = db->db_user;
2861 dbuf_verify_user(db, DBVU_NOT_EVICTING);
2862 mutex_exit(&db->db_mtx);
2868 dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2870 return (dmu_buf_replace_user(db_fake, NULL, user));
2874 dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2876 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2878 db->db_user_immediate_evict = TRUE;
2879 return (dmu_buf_set_user(db_fake, user));
2883 dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2885 return (dmu_buf_replace_user(db_fake, user, NULL));
2889 dmu_buf_get_user(dmu_buf_t *db_fake)
2891 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2893 dbuf_verify_user(db, DBVU_NOT_EVICTING);
2894 return (db->db_user);
2898 dmu_buf_user_evict_wait()
2900 taskq_wait(dbu_evict_taskq);
2904 dmu_buf_get_blkptr(dmu_buf_t *db)
2906 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2907 return (dbi->db_blkptr);
2911 dmu_buf_get_objset(dmu_buf_t *db)
2913 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2914 return (dbi->db_objset);
2918 dmu_buf_dnode_enter(dmu_buf_t *db)
2920 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2921 DB_DNODE_ENTER(dbi);
2922 return (DB_DNODE(dbi));
2926 dmu_buf_dnode_exit(dmu_buf_t *db)
2928 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2933 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
2935 /* ASSERT(dmu_tx_is_syncing(tx) */
2936 ASSERT(MUTEX_HELD(&db->db_mtx));
2938 if (db->db_blkptr != NULL)
2941 if (db->db_blkid == DMU_SPILL_BLKID) {
2942 db->db_blkptr = &dn->dn_phys->dn_spill;
2943 BP_ZERO(db->db_blkptr);
2946 if (db->db_level == dn->dn_phys->dn_nlevels-1) {
2948 * This buffer was allocated at a time when there was
2949 * no available blkptrs from the dnode, or it was
2950 * inappropriate to hook it in (i.e., nlevels mis-match).
2952 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
2953 ASSERT(db->db_parent == NULL);
2954 db->db_parent = dn->dn_dbuf;
2955 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
2958 dmu_buf_impl_t *parent = db->db_parent;
2959 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2961 ASSERT(dn->dn_phys->dn_nlevels > 1);
2962 if (parent == NULL) {
2963 mutex_exit(&db->db_mtx);
2964 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2965 parent = dbuf_hold_level(dn, db->db_level + 1,
2966 db->db_blkid >> epbs, db);
2967 rw_exit(&dn->dn_struct_rwlock);
2968 mutex_enter(&db->db_mtx);
2969 db->db_parent = parent;
2971 db->db_blkptr = (blkptr_t *)parent->db.db_data +
2972 (db->db_blkid & ((1ULL << epbs) - 1));
2978 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2980 dmu_buf_impl_t *db = dr->dr_dbuf;
2984 ASSERT(dmu_tx_is_syncing(tx));
2986 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2988 mutex_enter(&db->db_mtx);
2990 ASSERT(db->db_level > 0);
2993 /* Read the block if it hasn't been read yet. */
2994 if (db->db_buf == NULL) {
2995 mutex_exit(&db->db_mtx);
2996 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
2997 mutex_enter(&db->db_mtx);
2999 ASSERT3U(db->db_state, ==, DB_CACHED);
3000 ASSERT(db->db_buf != NULL);
3004 /* Indirect block size must match what the dnode thinks it is. */
3005 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3006 dbuf_check_blkptr(dn, db);
3009 /* Provide the pending dirty record to child dbufs */
3010 db->db_data_pending = dr;
3012 mutex_exit(&db->db_mtx);
3014 dbuf_write(dr, db->db_buf, tx);
3017 mutex_enter(&dr->dt.di.dr_mtx);
3018 dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx);
3019 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3020 mutex_exit(&dr->dt.di.dr_mtx);
3025 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
3027 arc_buf_t **datap = &dr->dt.dl.dr_data;
3028 dmu_buf_impl_t *db = dr->dr_dbuf;
3031 uint64_t txg = tx->tx_txg;
3033 ASSERT(dmu_tx_is_syncing(tx));
3035 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
3037 mutex_enter(&db->db_mtx);
3039 * To be synced, we must be dirtied. But we
3040 * might have been freed after the dirty.
3042 if (db->db_state == DB_UNCACHED) {
3043 /* This buffer has been freed since it was dirtied */
3044 ASSERT(db->db.db_data == NULL);
3045 } else if (db->db_state == DB_FILL) {
3046 /* This buffer was freed and is now being re-filled */
3047 ASSERT(db->db.db_data != dr->dt.dl.dr_data);
3049 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
3056 if (db->db_blkid == DMU_SPILL_BLKID) {
3057 mutex_enter(&dn->dn_mtx);
3058 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
3059 mutex_exit(&dn->dn_mtx);
3063 * If this is a bonus buffer, simply copy the bonus data into the
3064 * dnode. It will be written out when the dnode is synced (and it
3065 * will be synced, since it must have been dirty for dbuf_sync to
3068 if (db->db_blkid == DMU_BONUS_BLKID) {
3069 dbuf_dirty_record_t **drp;
3071 ASSERT(*datap != NULL);
3072 ASSERT0(db->db_level);
3073 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN);
3074 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen);
3077 if (*datap != db->db.db_data) {
3078 zio_buf_free(*datap, DN_MAX_BONUSLEN);
3079 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
3081 db->db_data_pending = NULL;
3082 drp = &db->db_last_dirty;
3084 drp = &(*drp)->dr_next;
3085 ASSERT(dr->dr_next == NULL);
3086 ASSERT(dr->dr_dbuf == db);
3088 if (dr->dr_dbuf->db_level != 0) {
3089 list_destroy(&dr->dt.di.dr_children);
3090 mutex_destroy(&dr->dt.di.dr_mtx);
3092 kmem_free(dr, sizeof (dbuf_dirty_record_t));
3093 ASSERT(db->db_dirtycnt > 0);
3094 db->db_dirtycnt -= 1;
3095 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
3102 * This function may have dropped the db_mtx lock allowing a dmu_sync
3103 * operation to sneak in. As a result, we need to ensure that we
3104 * don't check the dr_override_state until we have returned from
3105 * dbuf_check_blkptr.
3107 dbuf_check_blkptr(dn, db);
3110 * If this buffer is in the middle of an immediate write,
3111 * wait for the synchronous IO to complete.
3113 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
3114 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
3115 cv_wait(&db->db_changed, &db->db_mtx);
3116 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
3119 if (db->db_state != DB_NOFILL &&
3120 dn->dn_object != DMU_META_DNODE_OBJECT &&
3121 refcount_count(&db->db_holds) > 1 &&
3122 dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
3123 *datap == db->db_buf) {
3125 * If this buffer is currently "in use" (i.e., there
3126 * are active holds and db_data still references it),
3127 * then make a copy before we start the write so that
3128 * any modifications from the open txg will not leak
3131 * NOTE: this copy does not need to be made for
3132 * objects only modified in the syncing context (e.g.
3133 * DNONE_DNODE blocks).
3135 int psize = arc_buf_size(*datap);
3136 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
3137 enum zio_compress compress_type = arc_get_compression(*datap);
3139 if (compress_type == ZIO_COMPRESS_OFF) {
3140 *datap = arc_alloc_buf(os->os_spa, db, type, psize);
3142 ASSERT3U(type, ==, ARC_BUFC_DATA);
3143 int lsize = arc_buf_lsize(*datap);
3144 *datap = arc_alloc_compressed_buf(os->os_spa, db,
3145 psize, lsize, compress_type);
3147 bcopy(db->db.db_data, (*datap)->b_data, psize);
3149 db->db_data_pending = dr;
3151 mutex_exit(&db->db_mtx);
3153 dbuf_write(dr, *datap, tx);
3155 ASSERT(!list_link_active(&dr->dr_dirty_node));
3156 if (dn->dn_object == DMU_META_DNODE_OBJECT) {
3157 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
3161 * Although zio_nowait() does not "wait for an IO", it does
3162 * initiate the IO. If this is an empty write it seems plausible
3163 * that the IO could actually be completed before the nowait
3164 * returns. We need to DB_DNODE_EXIT() first in case
3165 * zio_nowait() invalidates the dbuf.
3168 zio_nowait(dr->dr_zio);
3173 dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx)
3175 dbuf_dirty_record_t *dr;
3177 while (dr = list_head(list)) {
3178 if (dr->dr_zio != NULL) {
3180 * If we find an already initialized zio then we
3181 * are processing the meta-dnode, and we have finished.
3182 * The dbufs for all dnodes are put back on the list
3183 * during processing, so that we can zio_wait()
3184 * these IOs after initiating all child IOs.
3186 ASSERT3U(dr->dr_dbuf->db.db_object, ==,
3187 DMU_META_DNODE_OBJECT);
3190 if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
3191 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
3192 VERIFY3U(dr->dr_dbuf->db_level, ==, level);
3194 list_remove(list, dr);
3195 if (dr->dr_dbuf->db_level > 0)
3196 dbuf_sync_indirect(dr, tx);
3198 dbuf_sync_leaf(dr, tx);
3204 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3206 dmu_buf_impl_t *db = vdb;
3208 blkptr_t *bp = zio->io_bp;
3209 blkptr_t *bp_orig = &zio->io_bp_orig;
3210 spa_t *spa = zio->io_spa;
3215 ASSERT3P(db->db_blkptr, !=, NULL);
3216 ASSERT3P(&db->db_data_pending->dr_bp_copy, ==, bp);
3220 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
3221 dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
3222 zio->io_prev_space_delta = delta;
3224 if (bp->blk_birth != 0) {
3225 ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
3226 BP_GET_TYPE(bp) == dn->dn_type) ||
3227 (db->db_blkid == DMU_SPILL_BLKID &&
3228 BP_GET_TYPE(bp) == dn->dn_bonustype) ||
3229 BP_IS_EMBEDDED(bp));
3230 ASSERT(BP_GET_LEVEL(bp) == db->db_level);
3233 mutex_enter(&db->db_mtx);
3236 if (db->db_blkid == DMU_SPILL_BLKID) {
3237 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3238 ASSERT(!(BP_IS_HOLE(bp)) &&
3239 db->db_blkptr == &dn->dn_phys->dn_spill);
3243 if (db->db_level == 0) {
3244 mutex_enter(&dn->dn_mtx);
3245 if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
3246 db->db_blkid != DMU_SPILL_BLKID)
3247 dn->dn_phys->dn_maxblkid = db->db_blkid;
3248 mutex_exit(&dn->dn_mtx);
3250 if (dn->dn_type == DMU_OT_DNODE) {
3251 dnode_phys_t *dnp = db->db.db_data;
3252 for (i = db->db.db_size >> DNODE_SHIFT; i > 0;
3254 if (dnp->dn_type != DMU_OT_NONE)
3258 if (BP_IS_HOLE(bp)) {
3265 blkptr_t *ibp = db->db.db_data;
3266 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3267 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
3268 if (BP_IS_HOLE(ibp))
3270 fill += BP_GET_FILL(ibp);
3275 if (!BP_IS_EMBEDDED(bp))
3276 bp->blk_fill = fill;
3278 mutex_exit(&db->db_mtx);
3280 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3281 *db->db_blkptr = *bp;
3282 rw_exit(&dn->dn_struct_rwlock);
3287 * This function gets called just prior to running through the compression
3288 * stage of the zio pipeline. If we're an indirect block comprised of only
3289 * holes, then we want this indirect to be compressed away to a hole. In
3290 * order to do that we must zero out any information about the holes that
3291 * this indirect points to prior to before we try to compress it.
3294 dbuf_write_children_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3296 dmu_buf_impl_t *db = vdb;
3299 unsigned int epbs, i;
3301 ASSERT3U(db->db_level, >, 0);
3304 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3305 ASSERT3U(epbs, <, 31);
3307 /* Determine if all our children are holes */
3308 for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++) {
3309 if (!BP_IS_HOLE(bp))
3314 * If all the children are holes, then zero them all out so that
3315 * we may get compressed away.
3317 if (i == 1 << epbs) {
3319 * We only found holes. Grab the rwlock to prevent
3320 * anybody from reading the blocks we're about to
3323 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3324 bzero(db->db.db_data, db->db.db_size);
3325 rw_exit(&dn->dn_struct_rwlock);
3331 * The SPA will call this callback several times for each zio - once
3332 * for every physical child i/o (zio->io_phys_children times). This
3333 * allows the DMU to monitor the progress of each logical i/o. For example,
3334 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
3335 * block. There may be a long delay before all copies/fragments are completed,
3336 * so this callback allows us to retire dirty space gradually, as the physical
3341 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
3343 dmu_buf_impl_t *db = arg;
3344 objset_t *os = db->db_objset;
3345 dsl_pool_t *dp = dmu_objset_pool(os);
3346 dbuf_dirty_record_t *dr;
3349 dr = db->db_data_pending;
3350 ASSERT3U(dr->dr_txg, ==, zio->io_txg);
3353 * The callback will be called io_phys_children times. Retire one
3354 * portion of our dirty space each time we are called. Any rounding
3355 * error will be cleaned up by dsl_pool_sync()'s call to
3356 * dsl_pool_undirty_space().
3358 delta = dr->dr_accounted / zio->io_phys_children;
3359 dsl_pool_undirty_space(dp, delta, zio->io_txg);
3364 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
3366 dmu_buf_impl_t *db = vdb;
3367 blkptr_t *bp_orig = &zio->io_bp_orig;
3368 blkptr_t *bp = db->db_blkptr;
3369 objset_t *os = db->db_objset;
3370 dmu_tx_t *tx = os->os_synctx;
3371 dbuf_dirty_record_t **drp, *dr;
3373 ASSERT0(zio->io_error);
3374 ASSERT(db->db_blkptr == bp);
3377 * For nopwrites and rewrites we ensure that the bp matches our
3378 * original and bypass all the accounting.
3380 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
3381 ASSERT(BP_EQUAL(bp, bp_orig));
3383 dsl_dataset_t *ds = os->os_dsl_dataset;
3384 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
3385 dsl_dataset_block_born(ds, bp, tx);
3388 mutex_enter(&db->db_mtx);
3392 drp = &db->db_last_dirty;
3393 while ((dr = *drp) != db->db_data_pending)
3395 ASSERT(!list_link_active(&dr->dr_dirty_node));
3396 ASSERT(dr->dr_dbuf == db);
3397 ASSERT(dr->dr_next == NULL);
3401 if (db->db_blkid == DMU_SPILL_BLKID) {
3406 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3407 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
3408 db->db_blkptr == &dn->dn_phys->dn_spill);
3413 if (db->db_level == 0) {
3414 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
3415 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
3416 if (db->db_state != DB_NOFILL) {
3417 if (dr->dt.dl.dr_data != db->db_buf)
3418 arc_buf_destroy(dr->dt.dl.dr_data, db);
3425 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3426 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
3427 if (!BP_IS_HOLE(db->db_blkptr)) {
3429 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3430 ASSERT3U(db->db_blkid, <=,
3431 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs));
3432 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
3436 mutex_destroy(&dr->dt.di.dr_mtx);
3437 list_destroy(&dr->dt.di.dr_children);
3439 kmem_free(dr, sizeof (dbuf_dirty_record_t));
3441 cv_broadcast(&db->db_changed);
3442 ASSERT(db->db_dirtycnt > 0);
3443 db->db_dirtycnt -= 1;
3444 db->db_data_pending = NULL;
3445 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg);
3449 dbuf_write_nofill_ready(zio_t *zio)
3451 dbuf_write_ready(zio, NULL, zio->io_private);
3455 dbuf_write_nofill_done(zio_t *zio)
3457 dbuf_write_done(zio, NULL, zio->io_private);
3461 dbuf_write_override_ready(zio_t *zio)
3463 dbuf_dirty_record_t *dr = zio->io_private;
3464 dmu_buf_impl_t *db = dr->dr_dbuf;
3466 dbuf_write_ready(zio, NULL, db);
3470 dbuf_write_override_done(zio_t *zio)
3472 dbuf_dirty_record_t *dr = zio->io_private;
3473 dmu_buf_impl_t *db = dr->dr_dbuf;
3474 blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
3476 mutex_enter(&db->db_mtx);
3477 if (!BP_EQUAL(zio->io_bp, obp)) {
3478 if (!BP_IS_HOLE(obp))
3479 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
3480 arc_release(dr->dt.dl.dr_data, db);
3482 mutex_exit(&db->db_mtx);
3483 dbuf_write_done(zio, NULL, db);
3485 if (zio->io_abd != NULL)
3486 abd_put(zio->io_abd);
3489 typedef struct dbuf_remap_impl_callback_arg {
3491 uint64_t drica_blk_birth;
3493 } dbuf_remap_impl_callback_arg_t;
3496 dbuf_remap_impl_callback(uint64_t vdev, uint64_t offset, uint64_t size,
3499 dbuf_remap_impl_callback_arg_t *drica = arg;
3500 objset_t *os = drica->drica_os;
3501 spa_t *spa = dmu_objset_spa(os);
3502 dmu_tx_t *tx = drica->drica_tx;
3504 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)));
3506 if (os == spa_meta_objset(spa)) {
3507 spa_vdev_indirect_mark_obsolete(spa, vdev, offset, size, tx);
3509 dsl_dataset_block_remapped(dmu_objset_ds(os), vdev, offset,
3510 size, drica->drica_blk_birth, tx);
3515 dbuf_remap_impl(dnode_t *dn, blkptr_t *bp, dmu_tx_t *tx)
3517 blkptr_t bp_copy = *bp;
3518 spa_t *spa = dmu_objset_spa(dn->dn_objset);
3519 dbuf_remap_impl_callback_arg_t drica;
3521 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)));
3523 drica.drica_os = dn->dn_objset;
3524 drica.drica_blk_birth = bp->blk_birth;
3525 drica.drica_tx = tx;
3526 if (spa_remap_blkptr(spa, &bp_copy, dbuf_remap_impl_callback,
3529 * The struct_rwlock prevents dbuf_read_impl() from
3530 * dereferencing the BP while we are changing it. To
3531 * avoid lock contention, only grab it when we are actually
3534 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3536 rw_exit(&dn->dn_struct_rwlock);
3541 * Returns true if a dbuf_remap would modify the dbuf. We do this by attempting
3542 * to remap a copy of every bp in the dbuf.
3545 dbuf_can_remap(const dmu_buf_impl_t *db)
3547 spa_t *spa = dmu_objset_spa(db->db_objset);
3548 blkptr_t *bp = db->db.db_data;
3549 boolean_t ret = B_FALSE;
3551 ASSERT3U(db->db_level, >, 0);
3552 ASSERT3S(db->db_state, ==, DB_CACHED);
3554 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL));
3556 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
3557 for (int i = 0; i < db->db.db_size >> SPA_BLKPTRSHIFT; i++) {
3558 blkptr_t bp_copy = bp[i];
3559 if (spa_remap_blkptr(spa, &bp_copy, NULL, NULL)) {
3564 spa_config_exit(spa, SCL_VDEV, FTAG);
3570 dnode_needs_remap(const dnode_t *dn)
3572 spa_t *spa = dmu_objset_spa(dn->dn_objset);
3573 boolean_t ret = B_FALSE;
3575 if (dn->dn_phys->dn_nlevels == 0) {
3579 ASSERT(spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL));
3581 spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
3582 for (int j = 0; j < dn->dn_phys->dn_nblkptr; j++) {
3583 blkptr_t bp_copy = dn->dn_phys->dn_blkptr[j];
3584 if (spa_remap_blkptr(spa, &bp_copy, NULL, NULL)) {
3589 spa_config_exit(spa, SCL_VDEV, FTAG);
3595 * Remap any existing BP's to concrete vdevs, if possible.
3598 dbuf_remap(dnode_t *dn, dmu_buf_impl_t *db, dmu_tx_t *tx)
3600 spa_t *spa = dmu_objset_spa(db->db_objset);
3601 ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)));
3603 if (!spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL))
3606 if (db->db_level > 0) {
3607 blkptr_t *bp = db->db.db_data;
3608 for (int i = 0; i < db->db.db_size >> SPA_BLKPTRSHIFT; i++) {
3609 dbuf_remap_impl(dn, &bp[i], tx);
3611 } else if (db->db.db_object == DMU_META_DNODE_OBJECT) {
3612 dnode_phys_t *dnp = db->db.db_data;
3613 ASSERT3U(db->db_dnode_handle->dnh_dnode->dn_type, ==,
3615 for (int i = 0; i < db->db.db_size >> DNODE_SHIFT; i++) {
3616 for (int j = 0; j < dnp[i].dn_nblkptr; j++) {
3617 dbuf_remap_impl(dn, &dnp[i].dn_blkptr[j], tx);
3624 /* Issue I/O to commit a dirty buffer to disk. */
3626 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
3628 dmu_buf_impl_t *db = dr->dr_dbuf;
3631 dmu_buf_impl_t *parent = db->db_parent;
3632 uint64_t txg = tx->tx_txg;
3633 zbookmark_phys_t zb;
3638 ASSERT(dmu_tx_is_syncing(tx));
3644 if (db->db_state != DB_NOFILL) {
3645 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
3647 * Private object buffers are released here rather
3648 * than in dbuf_dirty() since they are only modified
3649 * in the syncing context and we don't want the
3650 * overhead of making multiple copies of the data.
3652 if (BP_IS_HOLE(db->db_blkptr)) {
3655 dbuf_release_bp(db);
3657 dbuf_remap(dn, db, tx);
3661 if (parent != dn->dn_dbuf) {
3662 /* Our parent is an indirect block. */
3663 /* We have a dirty parent that has been scheduled for write. */
3664 ASSERT(parent && parent->db_data_pending);
3665 /* Our parent's buffer is one level closer to the dnode. */
3666 ASSERT(db->db_level == parent->db_level-1);
3668 * We're about to modify our parent's db_data by modifying
3669 * our block pointer, so the parent must be released.
3671 ASSERT(arc_released(parent->db_buf));
3672 zio = parent->db_data_pending->dr_zio;
3674 /* Our parent is the dnode itself. */
3675 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
3676 db->db_blkid != DMU_SPILL_BLKID) ||
3677 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
3678 if (db->db_blkid != DMU_SPILL_BLKID)
3679 ASSERT3P(db->db_blkptr, ==,
3680 &dn->dn_phys->dn_blkptr[db->db_blkid]);
3684 ASSERT(db->db_level == 0 || data == db->db_buf);
3685 ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
3688 SET_BOOKMARK(&zb, os->os_dsl_dataset ?
3689 os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
3690 db->db.db_object, db->db_level, db->db_blkid);
3692 if (db->db_blkid == DMU_SPILL_BLKID)
3694 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
3696 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
3700 * We copy the blkptr now (rather than when we instantiate the dirty
3701 * record), because its value can change between open context and
3702 * syncing context. We do not need to hold dn_struct_rwlock to read
3703 * db_blkptr because we are in syncing context.
3705 dr->dr_bp_copy = *db->db_blkptr;
3707 if (db->db_level == 0 &&
3708 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
3710 * The BP for this block has been provided by open context
3711 * (by dmu_sync() or dmu_buf_write_embedded()).
3713 abd_t *contents = (data != NULL) ?
3714 abd_get_from_buf(data->b_data, arc_buf_size(data)) : NULL;
3716 dr->dr_zio = zio_write(zio, os->os_spa, txg, &dr->dr_bp_copy,
3717 contents, db->db.db_size, db->db.db_size, &zp,
3718 dbuf_write_override_ready, NULL, NULL,
3719 dbuf_write_override_done,
3720 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
3721 mutex_enter(&db->db_mtx);
3722 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
3723 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
3724 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
3725 mutex_exit(&db->db_mtx);
3726 } else if (db->db_state == DB_NOFILL) {
3727 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF ||
3728 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY);
3729 dr->dr_zio = zio_write(zio, os->os_spa, txg,
3730 &dr->dr_bp_copy, NULL, db->db.db_size, db->db.db_size, &zp,
3731 dbuf_write_nofill_ready, NULL, NULL,
3732 dbuf_write_nofill_done, db,
3733 ZIO_PRIORITY_ASYNC_WRITE,
3734 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
3736 ASSERT(arc_released(data));
3739 * For indirect blocks, we want to setup the children
3740 * ready callback so that we can properly handle an indirect
3741 * block that only contains holes.
3743 arc_done_func_t *children_ready_cb = NULL;
3744 if (db->db_level != 0)
3745 children_ready_cb = dbuf_write_children_ready;
3747 dr->dr_zio = arc_write(zio, os->os_spa, txg,
3748 &dr->dr_bp_copy, data, DBUF_IS_L2CACHEABLE(db),
3749 &zp, dbuf_write_ready, children_ready_cb,
3750 dbuf_write_physdone, dbuf_write_done, db,
3751 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);