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) 2013 by Delphix. All rights reserved.
25 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
26 * Copyright (c) 2013, Joyent, Inc. All rights reserved.
29 #include <sys/zfs_context.h>
31 #include <sys/dmu_send.h>
32 #include <sys/dmu_impl.h>
34 #include <sys/dmu_objset.h>
35 #include <sys/dsl_dataset.h>
36 #include <sys/dsl_dir.h>
37 #include <sys/dmu_tx.h>
40 #include <sys/dmu_zfetch.h>
42 #include <sys/sa_impl.h>
45 * Number of times that zfs_free_range() took the slow path while doing
46 * a zfs receive. A nonzero value indicates a potential performance problem.
48 uint64_t zfs_free_range_recv_miss;
50 static void dbuf_destroy(dmu_buf_impl_t *db);
51 static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
52 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
55 * Global data structures and functions for the dbuf cache.
57 static kmem_cache_t *dbuf_cache;
61 dbuf_cons(void *vdb, void *unused, int kmflag)
63 dmu_buf_impl_t *db = vdb;
64 bzero(db, sizeof (dmu_buf_impl_t));
66 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
67 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
68 refcount_create(&db->db_holds);
74 dbuf_dest(void *vdb, void *unused)
76 dmu_buf_impl_t *db = vdb;
77 mutex_destroy(&db->db_mtx);
78 cv_destroy(&db->db_changed);
79 refcount_destroy(&db->db_holds);
83 * dbuf hash table routines
85 static dbuf_hash_table_t dbuf_hash_table;
87 static uint64_t dbuf_hash_count;
90 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
92 uintptr_t osv = (uintptr_t)os;
95 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
96 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF];
97 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF];
98 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF];
99 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF];
100 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF];
101 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF];
103 crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16);
108 #define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid);
110 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
111 ((dbuf)->db.db_object == (obj) && \
112 (dbuf)->db_objset == (os) && \
113 (dbuf)->db_level == (level) && \
114 (dbuf)->db_blkid == (blkid))
117 dbuf_find(dnode_t *dn, uint8_t level, uint64_t blkid)
119 dbuf_hash_table_t *h = &dbuf_hash_table;
120 objset_t *os = dn->dn_objset;
121 uint64_t obj = dn->dn_object;
122 uint64_t hv = DBUF_HASH(os, obj, level, blkid);
123 uint64_t idx = hv & h->hash_table_mask;
126 mutex_enter(DBUF_HASH_MUTEX(h, idx));
127 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
128 if (DBUF_EQUAL(db, os, obj, level, blkid)) {
129 mutex_enter(&db->db_mtx);
130 if (db->db_state != DB_EVICTING) {
131 mutex_exit(DBUF_HASH_MUTEX(h, idx));
134 mutex_exit(&db->db_mtx);
137 mutex_exit(DBUF_HASH_MUTEX(h, idx));
142 * Insert an entry into the hash table. If there is already an element
143 * equal to elem in the hash table, then the already existing element
144 * will be returned and the new element will not be inserted.
145 * Otherwise returns NULL.
147 static dmu_buf_impl_t *
148 dbuf_hash_insert(dmu_buf_impl_t *db)
150 dbuf_hash_table_t *h = &dbuf_hash_table;
151 objset_t *os = db->db_objset;
152 uint64_t obj = db->db.db_object;
153 int level = db->db_level;
154 uint64_t blkid = db->db_blkid;
155 uint64_t hv = DBUF_HASH(os, obj, level, blkid);
156 uint64_t idx = hv & h->hash_table_mask;
159 mutex_enter(DBUF_HASH_MUTEX(h, idx));
160 for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) {
161 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
162 mutex_enter(&dbf->db_mtx);
163 if (dbf->db_state != DB_EVICTING) {
164 mutex_exit(DBUF_HASH_MUTEX(h, idx));
167 mutex_exit(&dbf->db_mtx);
171 mutex_enter(&db->db_mtx);
172 db->db_hash_next = h->hash_table[idx];
173 h->hash_table[idx] = db;
174 mutex_exit(DBUF_HASH_MUTEX(h, idx));
175 atomic_add_64(&dbuf_hash_count, 1);
181 * Remove an entry from the hash table. This operation will
182 * fail if there are any existing holds on the db.
185 dbuf_hash_remove(dmu_buf_impl_t *db)
187 dbuf_hash_table_t *h = &dbuf_hash_table;
188 uint64_t hv = DBUF_HASH(db->db_objset, db->db.db_object,
189 db->db_level, db->db_blkid);
190 uint64_t idx = hv & h->hash_table_mask;
191 dmu_buf_impl_t *dbf, **dbp;
194 * We musn't hold db_mtx to maintin lock ordering:
195 * DBUF_HASH_MUTEX > db_mtx.
197 ASSERT(refcount_is_zero(&db->db_holds));
198 ASSERT(db->db_state == DB_EVICTING);
199 ASSERT(!MUTEX_HELD(&db->db_mtx));
201 mutex_enter(DBUF_HASH_MUTEX(h, idx));
202 dbp = &h->hash_table[idx];
203 while ((dbf = *dbp) != db) {
204 dbp = &dbf->db_hash_next;
207 *dbp = db->db_hash_next;
208 db->db_hash_next = NULL;
209 mutex_exit(DBUF_HASH_MUTEX(h, idx));
210 atomic_add_64(&dbuf_hash_count, -1);
213 static arc_evict_func_t dbuf_do_evict;
216 dbuf_evict_user(dmu_buf_impl_t *db)
218 ASSERT(MUTEX_HELD(&db->db_mtx));
220 if (db->db_level != 0 || db->db_evict_func == NULL)
223 if (db->db_user_data_ptr_ptr)
224 *db->db_user_data_ptr_ptr = db->db.db_data;
225 db->db_evict_func(&db->db, db->db_user_ptr);
226 db->db_user_ptr = NULL;
227 db->db_user_data_ptr_ptr = NULL;
228 db->db_evict_func = NULL;
232 dbuf_is_metadata(dmu_buf_impl_t *db)
234 if (db->db_level > 0) {
237 boolean_t is_metadata;
240 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type);
243 return (is_metadata);
248 dbuf_evict(dmu_buf_impl_t *db)
250 ASSERT(MUTEX_HELD(&db->db_mtx));
251 ASSERT(db->db_buf == NULL);
252 ASSERT(db->db_data_pending == NULL);
261 uint64_t hsize = 1ULL << 16;
262 dbuf_hash_table_t *h = &dbuf_hash_table;
266 * The hash table is big enough to fill all of physical memory
267 * with an average 4K block size. The table will take up
268 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
270 while (hsize * 4096 < (uint64_t)physmem * PAGESIZE)
274 h->hash_table_mask = hsize - 1;
275 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
276 if (h->hash_table == NULL) {
277 /* XXX - we should really return an error instead of assert */
278 ASSERT(hsize > (1ULL << 10));
283 dbuf_cache = kmem_cache_create("dmu_buf_impl_t",
284 sizeof (dmu_buf_impl_t),
285 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
287 for (i = 0; i < DBUF_MUTEXES; i++)
288 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
294 dbuf_hash_table_t *h = &dbuf_hash_table;
297 for (i = 0; i < DBUF_MUTEXES; i++)
298 mutex_destroy(&h->hash_mutexes[i]);
299 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
300 kmem_cache_destroy(dbuf_cache);
309 dbuf_verify(dmu_buf_impl_t *db)
312 dbuf_dirty_record_t *dr;
314 ASSERT(MUTEX_HELD(&db->db_mtx));
316 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
319 ASSERT(db->db_objset != NULL);
323 ASSERT(db->db_parent == NULL);
324 ASSERT(db->db_blkptr == NULL);
326 ASSERT3U(db->db.db_object, ==, dn->dn_object);
327 ASSERT3P(db->db_objset, ==, dn->dn_objset);
328 ASSERT3U(db->db_level, <, dn->dn_nlevels);
329 ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
330 db->db_blkid == DMU_SPILL_BLKID ||
331 !list_is_empty(&dn->dn_dbufs));
333 if (db->db_blkid == DMU_BONUS_BLKID) {
335 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
336 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
337 } else if (db->db_blkid == DMU_SPILL_BLKID) {
339 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
340 ASSERT0(db->db.db_offset);
342 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
345 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
346 ASSERT(dr->dr_dbuf == db);
348 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
349 ASSERT(dr->dr_dbuf == db);
352 * We can't assert that db_size matches dn_datablksz because it
353 * can be momentarily different when another thread is doing
356 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
357 dr = db->db_data_pending;
359 * It should only be modified in syncing context, so
360 * make sure we only have one copy of the data.
362 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
365 /* verify db->db_blkptr */
367 if (db->db_parent == dn->dn_dbuf) {
368 /* db is pointed to by the dnode */
369 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
370 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
371 ASSERT(db->db_parent == NULL);
373 ASSERT(db->db_parent != NULL);
374 if (db->db_blkid != DMU_SPILL_BLKID)
375 ASSERT3P(db->db_blkptr, ==,
376 &dn->dn_phys->dn_blkptr[db->db_blkid]);
378 /* db is pointed to by an indirect block */
379 int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
380 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
381 ASSERT3U(db->db_parent->db.db_object, ==,
384 * dnode_grow_indblksz() can make this fail if we don't
385 * have the struct_rwlock. XXX indblksz no longer
386 * grows. safe to do this now?
388 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
389 ASSERT3P(db->db_blkptr, ==,
390 ((blkptr_t *)db->db_parent->db.db_data +
391 db->db_blkid % epb));
395 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
396 (db->db_buf == NULL || db->db_buf->b_data) &&
397 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
398 db->db_state != DB_FILL && !dn->dn_free_txg) {
400 * If the blkptr isn't set but they have nonzero data,
401 * it had better be dirty, otherwise we'll lose that
402 * data when we evict this buffer.
404 if (db->db_dirtycnt == 0) {
405 uint64_t *buf = db->db.db_data;
408 for (i = 0; i < db->db.db_size >> 3; i++) {
418 dbuf_update_data(dmu_buf_impl_t *db)
420 ASSERT(MUTEX_HELD(&db->db_mtx));
421 if (db->db_level == 0 && db->db_user_data_ptr_ptr) {
422 ASSERT(!refcount_is_zero(&db->db_holds));
423 *db->db_user_data_ptr_ptr = db->db.db_data;
428 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
430 ASSERT(MUTEX_HELD(&db->db_mtx));
431 ASSERT(db->db_buf == NULL || !arc_has_callback(db->db_buf));
434 ASSERT(buf->b_data != NULL);
435 db->db.db_data = buf->b_data;
436 if (!arc_released(buf))
437 arc_set_callback(buf, dbuf_do_evict, db);
438 dbuf_update_data(db);
441 db->db.db_data = NULL;
442 if (db->db_state != DB_NOFILL)
443 db->db_state = DB_UNCACHED;
448 * Loan out an arc_buf for read. Return the loaned arc_buf.
451 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
455 mutex_enter(&db->db_mtx);
456 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
457 int blksz = db->db.db_size;
460 mutex_exit(&db->db_mtx);
461 DB_GET_SPA(&spa, db);
462 abuf = arc_loan_buf(spa, blksz);
463 bcopy(db->db.db_data, abuf->b_data, blksz);
466 arc_loan_inuse_buf(abuf, db);
467 dbuf_set_data(db, NULL);
468 mutex_exit(&db->db_mtx);
474 dbuf_whichblock(dnode_t *dn, uint64_t offset)
476 if (dn->dn_datablkshift) {
477 return (offset >> dn->dn_datablkshift);
479 ASSERT3U(offset, <, dn->dn_datablksz);
485 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
487 dmu_buf_impl_t *db = vdb;
489 mutex_enter(&db->db_mtx);
490 ASSERT3U(db->db_state, ==, DB_READ);
492 * All reads are synchronous, so we must have a hold on the dbuf
494 ASSERT(refcount_count(&db->db_holds) > 0);
495 ASSERT(db->db_buf == NULL);
496 ASSERT(db->db.db_data == NULL);
497 if (db->db_level == 0 && db->db_freed_in_flight) {
498 /* we were freed in flight; disregard any error */
499 arc_release(buf, db);
500 bzero(buf->b_data, db->db.db_size);
502 db->db_freed_in_flight = FALSE;
503 dbuf_set_data(db, buf);
504 db->db_state = DB_CACHED;
505 } else if (zio == NULL || zio->io_error == 0) {
506 dbuf_set_data(db, buf);
507 db->db_state = DB_CACHED;
509 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
510 ASSERT3P(db->db_buf, ==, NULL);
511 VERIFY(arc_buf_remove_ref(buf, db));
512 db->db_state = DB_UNCACHED;
514 cv_broadcast(&db->db_changed);
515 dbuf_rele_and_unlock(db, NULL);
519 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t *flags)
524 uint32_t aflags = ARC_NOWAIT;
528 ASSERT(!refcount_is_zero(&db->db_holds));
529 /* We need the struct_rwlock to prevent db_blkptr from changing. */
530 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
531 ASSERT(MUTEX_HELD(&db->db_mtx));
532 ASSERT(db->db_state == DB_UNCACHED);
533 ASSERT(db->db_buf == NULL);
535 if (db->db_blkid == DMU_BONUS_BLKID) {
536 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
538 ASSERT3U(bonuslen, <=, db->db.db_size);
539 db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN);
540 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
541 if (bonuslen < DN_MAX_BONUSLEN)
542 bzero(db->db.db_data, DN_MAX_BONUSLEN);
544 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
546 dbuf_update_data(db);
547 db->db_state = DB_CACHED;
548 mutex_exit(&db->db_mtx);
553 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
554 * processes the delete record and clears the bp while we are waiting
555 * for the dn_mtx (resulting in a "no" from block_freed).
557 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
558 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
559 BP_IS_HOLE(db->db_blkptr)))) {
560 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
562 dbuf_set_data(db, arc_buf_alloc(dn->dn_objset->os_spa,
563 db->db.db_size, db, type));
565 bzero(db->db.db_data, db->db.db_size);
566 db->db_state = DB_CACHED;
567 *flags |= DB_RF_CACHED;
568 mutex_exit(&db->db_mtx);
572 spa = dn->dn_objset->os_spa;
575 db->db_state = DB_READ;
576 mutex_exit(&db->db_mtx);
578 if (DBUF_IS_L2CACHEABLE(db))
579 aflags |= ARC_L2CACHE;
580 if (DBUF_IS_L2COMPRESSIBLE(db))
581 aflags |= ARC_L2COMPRESS;
583 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
584 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
585 db->db.db_object, db->db_level, db->db_blkid);
587 dbuf_add_ref(db, NULL);
589 (void) arc_read(zio, spa, db->db_blkptr,
590 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
591 (*flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
593 if (aflags & ARC_CACHED)
594 *flags |= DB_RF_CACHED;
598 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
601 int havepzio = (zio != NULL);
606 * We don't have to hold the mutex to check db_state because it
607 * can't be freed while we have a hold on the buffer.
609 ASSERT(!refcount_is_zero(&db->db_holds));
611 if (db->db_state == DB_NOFILL)
612 return (SET_ERROR(EIO));
616 if ((flags & DB_RF_HAVESTRUCT) == 0)
617 rw_enter(&dn->dn_struct_rwlock, RW_READER);
619 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
620 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
621 DBUF_IS_CACHEABLE(db);
623 mutex_enter(&db->db_mtx);
624 if (db->db_state == DB_CACHED) {
625 mutex_exit(&db->db_mtx);
627 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
628 db->db.db_size, TRUE);
629 if ((flags & DB_RF_HAVESTRUCT) == 0)
630 rw_exit(&dn->dn_struct_rwlock);
632 } else if (db->db_state == DB_UNCACHED) {
633 spa_t *spa = dn->dn_objset->os_spa;
636 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
637 dbuf_read_impl(db, zio, &flags);
639 /* dbuf_read_impl has dropped db_mtx for us */
642 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
643 db->db.db_size, flags & DB_RF_CACHED);
645 if ((flags & DB_RF_HAVESTRUCT) == 0)
646 rw_exit(&dn->dn_struct_rwlock);
653 * Another reader came in while the dbuf was in flight
654 * between UNCACHED and CACHED. Either a writer will finish
655 * writing the buffer (sending the dbuf to CACHED) or the
656 * first reader's request will reach the read_done callback
657 * and send the dbuf to CACHED. Otherwise, a failure
658 * occurred and the dbuf went to UNCACHED.
660 mutex_exit(&db->db_mtx);
662 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
663 db->db.db_size, TRUE);
664 if ((flags & DB_RF_HAVESTRUCT) == 0)
665 rw_exit(&dn->dn_struct_rwlock);
668 /* Skip the wait per the caller's request. */
669 mutex_enter(&db->db_mtx);
670 if ((flags & DB_RF_NEVERWAIT) == 0) {
671 while (db->db_state == DB_READ ||
672 db->db_state == DB_FILL) {
673 ASSERT(db->db_state == DB_READ ||
674 (flags & DB_RF_HAVESTRUCT) == 0);
675 cv_wait(&db->db_changed, &db->db_mtx);
677 if (db->db_state == DB_UNCACHED)
678 err = SET_ERROR(EIO);
680 mutex_exit(&db->db_mtx);
683 ASSERT(err || havepzio || db->db_state == DB_CACHED);
688 dbuf_noread(dmu_buf_impl_t *db)
690 ASSERT(!refcount_is_zero(&db->db_holds));
691 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
692 mutex_enter(&db->db_mtx);
693 while (db->db_state == DB_READ || db->db_state == DB_FILL)
694 cv_wait(&db->db_changed, &db->db_mtx);
695 if (db->db_state == DB_UNCACHED) {
696 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
699 ASSERT(db->db_buf == NULL);
700 ASSERT(db->db.db_data == NULL);
701 DB_GET_SPA(&spa, db);
702 dbuf_set_data(db, arc_buf_alloc(spa, db->db.db_size, db, type));
703 db->db_state = DB_FILL;
704 } else if (db->db_state == DB_NOFILL) {
705 dbuf_set_data(db, NULL);
707 ASSERT3U(db->db_state, ==, DB_CACHED);
709 mutex_exit(&db->db_mtx);
713 * This is our just-in-time copy function. It makes a copy of
714 * buffers, that have been modified in a previous transaction
715 * group, before we modify them in the current active group.
717 * This function is used in two places: when we are dirtying a
718 * buffer for the first time in a txg, and when we are freeing
719 * a range in a dnode that includes this buffer.
721 * Note that when we are called from dbuf_free_range() we do
722 * not put a hold on the buffer, we just traverse the active
723 * dbuf list for the dnode.
726 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
728 dbuf_dirty_record_t *dr = db->db_last_dirty;
730 ASSERT(MUTEX_HELD(&db->db_mtx));
731 ASSERT(db->db.db_data != NULL);
732 ASSERT(db->db_level == 0);
733 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
736 (dr->dt.dl.dr_data !=
737 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
741 * If the last dirty record for this dbuf has not yet synced
742 * and its referencing the dbuf data, either:
743 * reset the reference to point to a new copy,
744 * or (if there a no active holders)
745 * just null out the current db_data pointer.
747 ASSERT(dr->dr_txg >= txg - 2);
748 if (db->db_blkid == DMU_BONUS_BLKID) {
749 /* Note that the data bufs here are zio_bufs */
750 dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
751 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
752 bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
753 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
754 int size = db->db.db_size;
755 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
758 DB_GET_SPA(&spa, db);
759 dr->dt.dl.dr_data = arc_buf_alloc(spa, size, db, type);
760 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
762 dbuf_set_data(db, NULL);
767 dbuf_unoverride(dbuf_dirty_record_t *dr)
769 dmu_buf_impl_t *db = dr->dr_dbuf;
770 blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
771 uint64_t txg = dr->dr_txg;
773 ASSERT(MUTEX_HELD(&db->db_mtx));
774 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
775 ASSERT(db->db_level == 0);
777 if (db->db_blkid == DMU_BONUS_BLKID ||
778 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
781 ASSERT(db->db_data_pending != dr);
783 /* free this block */
784 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite) {
787 DB_GET_SPA(&spa, db);
788 zio_free(spa, txg, bp);
790 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
791 dr->dt.dl.dr_nopwrite = B_FALSE;
794 * Release the already-written buffer, so we leave it in
795 * a consistent dirty state. Note that all callers are
796 * modifying the buffer, so they will immediately do
797 * another (redundant) arc_release(). Therefore, leave
798 * the buf thawed to save the effort of freezing &
799 * immediately re-thawing it.
801 arc_release(dr->dt.dl.dr_data, db);
805 * Evict (if its unreferenced) or clear (if its referenced) any level-0
806 * data blocks in the free range, so that any future readers will find
807 * empty blocks. Also, if we happen across any level-1 dbufs in the
808 * range that have not already been marked dirty, mark them dirty so
809 * they stay in memory.
811 * This is a no-op if the dataset is in the middle of an incremental
812 * receive; see comment below for details.
815 dbuf_free_range(dnode_t *dn, uint64_t start, uint64_t end, dmu_tx_t *tx)
817 dmu_buf_impl_t *db, *db_next;
818 uint64_t txg = tx->tx_txg;
819 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
820 uint64_t first_l1 = start >> epbs;
821 uint64_t last_l1 = end >> epbs;
823 if (end > dn->dn_maxblkid && (end != DMU_SPILL_BLKID)) {
824 end = dn->dn_maxblkid;
825 last_l1 = end >> epbs;
827 dprintf_dnode(dn, "start=%llu end=%llu\n", start, end);
829 mutex_enter(&dn->dn_dbufs_mtx);
830 if (start >= dn->dn_unlisted_l0_blkid * dn->dn_datablksz) {
831 /* There can't be any dbufs in this range; no need to search. */
832 mutex_exit(&dn->dn_dbufs_mtx);
834 } else if (dmu_objset_is_receiving(dn->dn_objset)) {
836 * If we are receiving, we expect there to be no dbufs in
837 * the range to be freed, because receive modifies each
838 * block at most once, and in offset order. If this is
839 * not the case, it can lead to performance problems,
840 * so note that we unexpectedly took the slow path.
842 atomic_inc_64(&zfs_free_range_recv_miss);
845 for (db = list_head(&dn->dn_dbufs); db; db = db_next) {
846 db_next = list_next(&dn->dn_dbufs, db);
847 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
849 if (db->db_level == 1 &&
850 db->db_blkid >= first_l1 && db->db_blkid <= last_l1) {
851 mutex_enter(&db->db_mtx);
852 if (db->db_last_dirty &&
853 db->db_last_dirty->dr_txg < txg) {
854 dbuf_add_ref(db, FTAG);
855 mutex_exit(&db->db_mtx);
856 dbuf_will_dirty(db, tx);
859 mutex_exit(&db->db_mtx);
863 if (db->db_level != 0)
865 dprintf_dbuf(db, "found buf %s\n", "");
866 if (db->db_blkid < start || db->db_blkid > end)
869 /* found a level 0 buffer in the range */
870 mutex_enter(&db->db_mtx);
871 if (dbuf_undirty(db, tx)) {
872 /* mutex has been dropped and dbuf destroyed */
876 if (db->db_state == DB_UNCACHED ||
877 db->db_state == DB_NOFILL ||
878 db->db_state == DB_EVICTING) {
879 ASSERT(db->db.db_data == NULL);
880 mutex_exit(&db->db_mtx);
883 if (db->db_state == DB_READ || db->db_state == DB_FILL) {
884 /* will be handled in dbuf_read_done or dbuf_rele */
885 db->db_freed_in_flight = TRUE;
886 mutex_exit(&db->db_mtx);
889 if (refcount_count(&db->db_holds) == 0) {
894 /* The dbuf is referenced */
896 if (db->db_last_dirty != NULL) {
897 dbuf_dirty_record_t *dr = db->db_last_dirty;
899 if (dr->dr_txg == txg) {
901 * This buffer is "in-use", re-adjust the file
902 * size to reflect that this buffer may
903 * contain new data when we sync.
905 if (db->db_blkid != DMU_SPILL_BLKID &&
906 db->db_blkid > dn->dn_maxblkid)
907 dn->dn_maxblkid = db->db_blkid;
911 * This dbuf is not dirty in the open context.
912 * Either uncache it (if its not referenced in
913 * the open context) or reset its contents to
916 dbuf_fix_old_data(db, txg);
919 /* clear the contents if its cached */
920 if (db->db_state == DB_CACHED) {
921 ASSERT(db->db.db_data != NULL);
922 arc_release(db->db_buf, db);
923 bzero(db->db.db_data, db->db.db_size);
924 arc_buf_freeze(db->db_buf);
927 mutex_exit(&db->db_mtx);
929 mutex_exit(&dn->dn_dbufs_mtx);
933 dbuf_block_freeable(dmu_buf_impl_t *db)
935 dsl_dataset_t *ds = db->db_objset->os_dsl_dataset;
936 uint64_t birth_txg = 0;
939 * We don't need any locking to protect db_blkptr:
940 * If it's syncing, then db_last_dirty will be set
941 * so we'll ignore db_blkptr.
943 ASSERT(MUTEX_HELD(&db->db_mtx));
944 if (db->db_last_dirty)
945 birth_txg = db->db_last_dirty->dr_txg;
946 else if (db->db_blkptr)
947 birth_txg = db->db_blkptr->blk_birth;
950 * If we don't exist or are in a snapshot, we can't be freed.
951 * Don't pass the bp to dsl_dataset_block_freeable() since we
952 * are holding the db_mtx lock and might deadlock if we are
953 * prefetching a dedup-ed block.
956 return (ds == NULL ||
957 dsl_dataset_block_freeable(ds, NULL, birth_txg));
963 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
965 arc_buf_t *buf, *obuf;
966 int osize = db->db.db_size;
967 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
970 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
975 /* XXX does *this* func really need the lock? */
976 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
979 * This call to dbuf_will_dirty() with the dn_struct_rwlock held
980 * is OK, because there can be no other references to the db
981 * when we are changing its size, so no concurrent DB_FILL can
985 * XXX we should be doing a dbuf_read, checking the return
986 * value and returning that up to our callers
988 dbuf_will_dirty(db, tx);
990 /* create the data buffer for the new block */
991 buf = arc_buf_alloc(dn->dn_objset->os_spa, size, db, type);
993 /* copy old block data to the new block */
995 bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
996 /* zero the remainder */
998 bzero((uint8_t *)buf->b_data + osize, size - osize);
1000 mutex_enter(&db->db_mtx);
1001 dbuf_set_data(db, buf);
1002 VERIFY(arc_buf_remove_ref(obuf, db));
1003 db->db.db_size = size;
1005 if (db->db_level == 0) {
1006 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1007 db->db_last_dirty->dt.dl.dr_data = buf;
1009 mutex_exit(&db->db_mtx);
1011 dnode_willuse_space(dn, size-osize, tx);
1016 dbuf_release_bp(dmu_buf_impl_t *db)
1020 DB_GET_OBJSET(&os, db);
1021 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1022 ASSERT(arc_released(os->os_phys_buf) ||
1023 list_link_active(&os->os_dsl_dataset->ds_synced_link));
1024 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1026 (void) arc_release(db->db_buf, db);
1029 dbuf_dirty_record_t *
1030 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1034 dbuf_dirty_record_t **drp, *dr;
1035 int drop_struct_lock = FALSE;
1036 boolean_t do_free_accounting = B_FALSE;
1037 int txgoff = tx->tx_txg & TXG_MASK;
1039 ASSERT(tx->tx_txg != 0);
1040 ASSERT(!refcount_is_zero(&db->db_holds));
1041 DMU_TX_DIRTY_BUF(tx, db);
1046 * Shouldn't dirty a regular buffer in syncing context. Private
1047 * objects may be dirtied in syncing context, but only if they
1048 * were already pre-dirtied in open context.
1050 ASSERT(!dmu_tx_is_syncing(tx) ||
1051 BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1052 DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1053 dn->dn_objset->os_dsl_dataset == NULL);
1055 * We make this assert for private objects as well, but after we
1056 * check if we're already dirty. They are allowed to re-dirty
1057 * in syncing context.
1059 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1060 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1061 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1063 mutex_enter(&db->db_mtx);
1065 * XXX make this true for indirects too? The problem is that
1066 * transactions created with dmu_tx_create_assigned() from
1067 * syncing context don't bother holding ahead.
1069 ASSERT(db->db_level != 0 ||
1070 db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1071 db->db_state == DB_NOFILL);
1073 mutex_enter(&dn->dn_mtx);
1075 * Don't set dirtyctx to SYNC if we're just modifying this as we
1076 * initialize the objset.
1078 if (dn->dn_dirtyctx == DN_UNDIRTIED &&
1079 !BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1081 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1082 ASSERT(dn->dn_dirtyctx_firstset == NULL);
1083 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1085 mutex_exit(&dn->dn_mtx);
1087 if (db->db_blkid == DMU_SPILL_BLKID)
1088 dn->dn_have_spill = B_TRUE;
1091 * If this buffer is already dirty, we're done.
1093 drp = &db->db_last_dirty;
1094 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1095 db->db.db_object == DMU_META_DNODE_OBJECT);
1096 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1098 if (dr && dr->dr_txg == tx->tx_txg) {
1101 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1103 * If this buffer has already been written out,
1104 * we now need to reset its state.
1106 dbuf_unoverride(dr);
1107 if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1108 db->db_state != DB_NOFILL)
1109 arc_buf_thaw(db->db_buf);
1111 mutex_exit(&db->db_mtx);
1116 * Only valid if not already dirty.
1118 ASSERT(dn->dn_object == 0 ||
1119 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1120 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1122 ASSERT3U(dn->dn_nlevels, >, db->db_level);
1123 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1124 dn->dn_phys->dn_nlevels > db->db_level ||
1125 dn->dn_next_nlevels[txgoff] > db->db_level ||
1126 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1127 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1130 * We should only be dirtying in syncing context if it's the
1131 * mos or we're initializing the os or it's a special object.
1132 * However, we are allowed to dirty in syncing context provided
1133 * we already dirtied it in open context. Hence we must make
1134 * this assertion only if we're not already dirty.
1137 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1138 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1139 ASSERT(db->db.db_size != 0);
1141 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1143 if (db->db_blkid != DMU_BONUS_BLKID) {
1145 * Update the accounting.
1146 * Note: we delay "free accounting" until after we drop
1147 * the db_mtx. This keeps us from grabbing other locks
1148 * (and possibly deadlocking) in bp_get_dsize() while
1149 * also holding the db_mtx.
1151 dnode_willuse_space(dn, db->db.db_size, tx);
1152 do_free_accounting = dbuf_block_freeable(db);
1156 * If this buffer is dirty in an old transaction group we need
1157 * to make a copy of it so that the changes we make in this
1158 * transaction group won't leak out when we sync the older txg.
1160 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1161 if (db->db_level == 0) {
1162 void *data_old = db->db_buf;
1164 if (db->db_state != DB_NOFILL) {
1165 if (db->db_blkid == DMU_BONUS_BLKID) {
1166 dbuf_fix_old_data(db, tx->tx_txg);
1167 data_old = db->db.db_data;
1168 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1170 * Release the data buffer from the cache so
1171 * that we can modify it without impacting
1172 * possible other users of this cached data
1173 * block. Note that indirect blocks and
1174 * private objects are not released until the
1175 * syncing state (since they are only modified
1178 arc_release(db->db_buf, db);
1179 dbuf_fix_old_data(db, tx->tx_txg);
1180 data_old = db->db_buf;
1182 ASSERT(data_old != NULL);
1184 dr->dt.dl.dr_data = data_old;
1186 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1187 list_create(&dr->dt.di.dr_children,
1188 sizeof (dbuf_dirty_record_t),
1189 offsetof(dbuf_dirty_record_t, dr_dirty_node));
1192 dr->dr_txg = tx->tx_txg;
1197 * We could have been freed_in_flight between the dbuf_noread
1198 * and dbuf_dirty. We win, as though the dbuf_noread() had
1199 * happened after the free.
1201 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1202 db->db_blkid != DMU_SPILL_BLKID) {
1203 mutex_enter(&dn->dn_mtx);
1204 dnode_clear_range(dn, db->db_blkid, 1, tx);
1205 mutex_exit(&dn->dn_mtx);
1206 db->db_freed_in_flight = FALSE;
1210 * This buffer is now part of this txg
1212 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1213 db->db_dirtycnt += 1;
1214 ASSERT3U(db->db_dirtycnt, <=, 3);
1216 mutex_exit(&db->db_mtx);
1218 if (db->db_blkid == DMU_BONUS_BLKID ||
1219 db->db_blkid == DMU_SPILL_BLKID) {
1220 mutex_enter(&dn->dn_mtx);
1221 ASSERT(!list_link_active(&dr->dr_dirty_node));
1222 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1223 mutex_exit(&dn->dn_mtx);
1224 dnode_setdirty(dn, tx);
1227 } else if (do_free_accounting) {
1228 blkptr_t *bp = db->db_blkptr;
1229 int64_t willfree = (bp && !BP_IS_HOLE(bp)) ?
1230 bp_get_dsize(os->os_spa, bp) : db->db.db_size;
1232 * This is only a guess -- if the dbuf is dirty
1233 * in a previous txg, we don't know how much
1234 * space it will use on disk yet. We should
1235 * really have the struct_rwlock to access
1236 * db_blkptr, but since this is just a guess,
1237 * it's OK if we get an odd answer.
1239 ddt_prefetch(os->os_spa, bp);
1240 dnode_willuse_space(dn, -willfree, tx);
1243 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1244 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1245 drop_struct_lock = TRUE;
1248 if (db->db_level == 0) {
1249 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1250 ASSERT(dn->dn_maxblkid >= db->db_blkid);
1253 if (db->db_level+1 < dn->dn_nlevels) {
1254 dmu_buf_impl_t *parent = db->db_parent;
1255 dbuf_dirty_record_t *di;
1256 int parent_held = FALSE;
1258 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1259 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1261 parent = dbuf_hold_level(dn, db->db_level+1,
1262 db->db_blkid >> epbs, FTAG);
1263 ASSERT(parent != NULL);
1266 if (drop_struct_lock)
1267 rw_exit(&dn->dn_struct_rwlock);
1268 ASSERT3U(db->db_level+1, ==, parent->db_level);
1269 di = dbuf_dirty(parent, tx);
1271 dbuf_rele(parent, FTAG);
1273 mutex_enter(&db->db_mtx);
1274 /* possible race with dbuf_undirty() */
1275 if (db->db_last_dirty == dr ||
1276 dn->dn_object == DMU_META_DNODE_OBJECT) {
1277 mutex_enter(&di->dt.di.dr_mtx);
1278 ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1279 ASSERT(!list_link_active(&dr->dr_dirty_node));
1280 list_insert_tail(&di->dt.di.dr_children, dr);
1281 mutex_exit(&di->dt.di.dr_mtx);
1284 mutex_exit(&db->db_mtx);
1286 ASSERT(db->db_level+1 == dn->dn_nlevels);
1287 ASSERT(db->db_blkid < dn->dn_nblkptr);
1288 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1289 mutex_enter(&dn->dn_mtx);
1290 ASSERT(!list_link_active(&dr->dr_dirty_node));
1291 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1292 mutex_exit(&dn->dn_mtx);
1293 if (drop_struct_lock)
1294 rw_exit(&dn->dn_struct_rwlock);
1297 dnode_setdirty(dn, tx);
1303 * Undirty a buffer in the transaction group referenced by the given
1304 * transaction. Return whether this evicted the dbuf.
1307 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1310 uint64_t txg = tx->tx_txg;
1311 dbuf_dirty_record_t *dr, **drp;
1314 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1315 ASSERT0(db->db_level);
1316 ASSERT(MUTEX_HELD(&db->db_mtx));
1319 * If this buffer is not dirty, we're done.
1321 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1322 if (dr->dr_txg <= txg)
1324 if (dr == NULL || dr->dr_txg < txg)
1326 ASSERT(dr->dr_txg == txg);
1327 ASSERT(dr->dr_dbuf == db);
1333 * Note: This code will probably work even if there are concurrent
1334 * holders, but it is untested in that scenerio, as the ZPL and
1335 * ztest have additional locking (the range locks) that prevents
1336 * that type of concurrent access.
1338 ASSERT3U(refcount_count(&db->db_holds), ==, db->db_dirtycnt);
1340 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1342 ASSERT(db->db.db_size != 0);
1344 /* XXX would be nice to fix up dn_towrite_space[] */
1349 * Note that there are three places in dbuf_dirty()
1350 * where this dirty record may be put on a list.
1351 * Make sure to do a list_remove corresponding to
1352 * every one of those list_insert calls.
1354 if (dr->dr_parent) {
1355 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1356 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1357 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1358 } else if (db->db_blkid == DMU_SPILL_BLKID ||
1359 db->db_level+1 == dn->dn_nlevels) {
1360 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1361 mutex_enter(&dn->dn_mtx);
1362 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1363 mutex_exit(&dn->dn_mtx);
1367 if (db->db_state != DB_NOFILL) {
1368 dbuf_unoverride(dr);
1370 ASSERT(db->db_buf != NULL);
1371 ASSERT(dr->dt.dl.dr_data != NULL);
1372 if (dr->dt.dl.dr_data != db->db_buf)
1373 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db));
1375 kmem_free(dr, sizeof (dbuf_dirty_record_t));
1377 ASSERT(db->db_dirtycnt > 0);
1378 db->db_dirtycnt -= 1;
1380 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1381 arc_buf_t *buf = db->db_buf;
1383 ASSERT(db->db_state == DB_NOFILL || arc_released(buf));
1384 dbuf_set_data(db, NULL);
1385 VERIFY(arc_buf_remove_ref(buf, db));
1393 #pragma weak dmu_buf_will_dirty = dbuf_will_dirty
1395 dbuf_will_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1397 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1399 ASSERT(tx->tx_txg != 0);
1400 ASSERT(!refcount_is_zero(&db->db_holds));
1403 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
1404 rf |= DB_RF_HAVESTRUCT;
1406 (void) dbuf_read(db, NULL, rf);
1407 (void) dbuf_dirty(db, tx);
1411 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1413 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1415 db->db_state = DB_NOFILL;
1417 dmu_buf_will_fill(db_fake, tx);
1421 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1423 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1425 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1426 ASSERT(tx->tx_txg != 0);
1427 ASSERT(db->db_level == 0);
1428 ASSERT(!refcount_is_zero(&db->db_holds));
1430 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1431 dmu_tx_private_ok(tx));
1434 (void) dbuf_dirty(db, tx);
1437 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1440 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1442 mutex_enter(&db->db_mtx);
1445 if (db->db_state == DB_FILL) {
1446 if (db->db_level == 0 && db->db_freed_in_flight) {
1447 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1448 /* we were freed while filling */
1449 /* XXX dbuf_undirty? */
1450 bzero(db->db.db_data, db->db.db_size);
1451 db->db_freed_in_flight = FALSE;
1453 db->db_state = DB_CACHED;
1454 cv_broadcast(&db->db_changed);
1456 mutex_exit(&db->db_mtx);
1460 * Directly assign a provided arc buf to a given dbuf if it's not referenced
1461 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1464 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
1466 ASSERT(!refcount_is_zero(&db->db_holds));
1467 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1468 ASSERT(db->db_level == 0);
1469 ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA);
1470 ASSERT(buf != NULL);
1471 ASSERT(arc_buf_size(buf) == db->db.db_size);
1472 ASSERT(tx->tx_txg != 0);
1474 arc_return_buf(buf, db);
1475 ASSERT(arc_released(buf));
1477 mutex_enter(&db->db_mtx);
1479 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1480 cv_wait(&db->db_changed, &db->db_mtx);
1482 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
1484 if (db->db_state == DB_CACHED &&
1485 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
1486 mutex_exit(&db->db_mtx);
1487 (void) dbuf_dirty(db, tx);
1488 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
1489 VERIFY(arc_buf_remove_ref(buf, db));
1490 xuio_stat_wbuf_copied();
1494 xuio_stat_wbuf_nocopy();
1495 if (db->db_state == DB_CACHED) {
1496 dbuf_dirty_record_t *dr = db->db_last_dirty;
1498 ASSERT(db->db_buf != NULL);
1499 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
1500 ASSERT(dr->dt.dl.dr_data == db->db_buf);
1501 if (!arc_released(db->db_buf)) {
1502 ASSERT(dr->dt.dl.dr_override_state ==
1504 arc_release(db->db_buf, db);
1506 dr->dt.dl.dr_data = buf;
1507 VERIFY(arc_buf_remove_ref(db->db_buf, db));
1508 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
1509 arc_release(db->db_buf, db);
1510 VERIFY(arc_buf_remove_ref(db->db_buf, db));
1514 ASSERT(db->db_buf == NULL);
1515 dbuf_set_data(db, buf);
1516 db->db_state = DB_FILL;
1517 mutex_exit(&db->db_mtx);
1518 (void) dbuf_dirty(db, tx);
1519 dbuf_fill_done(db, tx);
1523 * "Clear" the contents of this dbuf. This will mark the dbuf
1524 * EVICTING and clear *most* of its references. Unfortunetely,
1525 * when we are not holding the dn_dbufs_mtx, we can't clear the
1526 * entry in the dn_dbufs list. We have to wait until dbuf_destroy()
1527 * in this case. For callers from the DMU we will usually see:
1528 * dbuf_clear()->arc_buf_evict()->dbuf_do_evict()->dbuf_destroy()
1529 * For the arc callback, we will usually see:
1530 * dbuf_do_evict()->dbuf_clear();dbuf_destroy()
1531 * Sometimes, though, we will get a mix of these two:
1532 * DMU: dbuf_clear()->arc_buf_evict()
1533 * ARC: dbuf_do_evict()->dbuf_destroy()
1536 dbuf_clear(dmu_buf_impl_t *db)
1539 dmu_buf_impl_t *parent = db->db_parent;
1540 dmu_buf_impl_t *dndb;
1541 int dbuf_gone = FALSE;
1543 ASSERT(MUTEX_HELD(&db->db_mtx));
1544 ASSERT(refcount_is_zero(&db->db_holds));
1546 dbuf_evict_user(db);
1548 if (db->db_state == DB_CACHED) {
1549 ASSERT(db->db.db_data != NULL);
1550 if (db->db_blkid == DMU_BONUS_BLKID) {
1551 zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
1552 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
1554 db->db.db_data = NULL;
1555 db->db_state = DB_UNCACHED;
1558 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
1559 ASSERT(db->db_data_pending == NULL);
1561 db->db_state = DB_EVICTING;
1562 db->db_blkptr = NULL;
1567 if (db->db_blkid != DMU_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) {
1568 list_remove(&dn->dn_dbufs, db);
1569 (void) atomic_dec_32_nv(&dn->dn_dbufs_count);
1573 * Decrementing the dbuf count means that the hold corresponding
1574 * to the removed dbuf is no longer discounted in dnode_move(),
1575 * so the dnode cannot be moved until after we release the hold.
1576 * The membar_producer() ensures visibility of the decremented
1577 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
1581 db->db_dnode_handle = NULL;
1587 dbuf_gone = arc_buf_evict(db->db_buf);
1590 mutex_exit(&db->db_mtx);
1593 * If this dbuf is referenced from an indirect dbuf,
1594 * decrement the ref count on the indirect dbuf.
1596 if (parent && parent != dndb)
1597 dbuf_rele(parent, db);
1601 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
1602 dmu_buf_impl_t **parentp, blkptr_t **bpp)
1609 ASSERT(blkid != DMU_BONUS_BLKID);
1611 if (blkid == DMU_SPILL_BLKID) {
1612 mutex_enter(&dn->dn_mtx);
1613 if (dn->dn_have_spill &&
1614 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
1615 *bpp = &dn->dn_phys->dn_spill;
1618 dbuf_add_ref(dn->dn_dbuf, NULL);
1619 *parentp = dn->dn_dbuf;
1620 mutex_exit(&dn->dn_mtx);
1624 if (dn->dn_phys->dn_nlevels == 0)
1627 nlevels = dn->dn_phys->dn_nlevels;
1629 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1631 ASSERT3U(level * epbs, <, 64);
1632 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1633 if (level >= nlevels ||
1634 (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
1635 /* the buffer has no parent yet */
1636 return (SET_ERROR(ENOENT));
1637 } else if (level < nlevels-1) {
1638 /* this block is referenced from an indirect block */
1639 int err = dbuf_hold_impl(dn, level+1,
1640 blkid >> epbs, fail_sparse, NULL, parentp);
1643 err = dbuf_read(*parentp, NULL,
1644 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
1646 dbuf_rele(*parentp, NULL);
1650 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
1651 (blkid & ((1ULL << epbs) - 1));
1654 /* the block is referenced from the dnode */
1655 ASSERT3U(level, ==, nlevels-1);
1656 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
1657 blkid < dn->dn_phys->dn_nblkptr);
1659 dbuf_add_ref(dn->dn_dbuf, NULL);
1660 *parentp = dn->dn_dbuf;
1662 *bpp = &dn->dn_phys->dn_blkptr[blkid];
1667 static dmu_buf_impl_t *
1668 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
1669 dmu_buf_impl_t *parent, blkptr_t *blkptr)
1671 objset_t *os = dn->dn_objset;
1672 dmu_buf_impl_t *db, *odb;
1674 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1675 ASSERT(dn->dn_type != DMU_OT_NONE);
1677 db = kmem_cache_alloc(dbuf_cache, KM_SLEEP);
1680 db->db.db_object = dn->dn_object;
1681 db->db_level = level;
1682 db->db_blkid = blkid;
1683 db->db_last_dirty = NULL;
1684 db->db_dirtycnt = 0;
1685 db->db_dnode_handle = dn->dn_handle;
1686 db->db_parent = parent;
1687 db->db_blkptr = blkptr;
1689 db->db_user_ptr = NULL;
1690 db->db_user_data_ptr_ptr = NULL;
1691 db->db_evict_func = NULL;
1692 db->db_immediate_evict = 0;
1693 db->db_freed_in_flight = 0;
1695 if (blkid == DMU_BONUS_BLKID) {
1696 ASSERT3P(parent, ==, dn->dn_dbuf);
1697 db->db.db_size = DN_MAX_BONUSLEN -
1698 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
1699 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
1700 db->db.db_offset = DMU_BONUS_BLKID;
1701 db->db_state = DB_UNCACHED;
1702 /* the bonus dbuf is not placed in the hash table */
1703 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1705 } else if (blkid == DMU_SPILL_BLKID) {
1706 db->db.db_size = (blkptr != NULL) ?
1707 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
1708 db->db.db_offset = 0;
1711 db->db_level ? 1<<dn->dn_indblkshift : dn->dn_datablksz;
1712 db->db.db_size = blocksize;
1713 db->db.db_offset = db->db_blkid * blocksize;
1717 * Hold the dn_dbufs_mtx while we get the new dbuf
1718 * in the hash table *and* added to the dbufs list.
1719 * This prevents a possible deadlock with someone
1720 * trying to look up this dbuf before its added to the
1723 mutex_enter(&dn->dn_dbufs_mtx);
1724 db->db_state = DB_EVICTING;
1725 if ((odb = dbuf_hash_insert(db)) != NULL) {
1726 /* someone else inserted it first */
1727 kmem_cache_free(dbuf_cache, db);
1728 mutex_exit(&dn->dn_dbufs_mtx);
1731 list_insert_head(&dn->dn_dbufs, db);
1732 if (db->db_level == 0 && db->db_blkid >=
1733 dn->dn_unlisted_l0_blkid)
1734 dn->dn_unlisted_l0_blkid = db->db_blkid + 1;
1735 db->db_state = DB_UNCACHED;
1736 mutex_exit(&dn->dn_dbufs_mtx);
1737 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1739 if (parent && parent != dn->dn_dbuf)
1740 dbuf_add_ref(parent, db);
1742 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1743 refcount_count(&dn->dn_holds) > 0);
1744 (void) refcount_add(&dn->dn_holds, db);
1745 (void) atomic_inc_32_nv(&dn->dn_dbufs_count);
1747 dprintf_dbuf(db, "db=%p\n", db);
1753 dbuf_do_evict(void *private)
1755 arc_buf_t *buf = private;
1756 dmu_buf_impl_t *db = buf->b_private;
1758 if (!MUTEX_HELD(&db->db_mtx))
1759 mutex_enter(&db->db_mtx);
1761 ASSERT(refcount_is_zero(&db->db_holds));
1763 if (db->db_state != DB_EVICTING) {
1764 ASSERT(db->db_state == DB_CACHED);
1769 mutex_exit(&db->db_mtx);
1776 dbuf_destroy(dmu_buf_impl_t *db)
1778 ASSERT(refcount_is_zero(&db->db_holds));
1780 if (db->db_blkid != DMU_BONUS_BLKID) {
1782 * If this dbuf is still on the dn_dbufs list,
1783 * remove it from that list.
1785 if (db->db_dnode_handle != NULL) {
1790 mutex_enter(&dn->dn_dbufs_mtx);
1791 list_remove(&dn->dn_dbufs, db);
1792 (void) atomic_dec_32_nv(&dn->dn_dbufs_count);
1793 mutex_exit(&dn->dn_dbufs_mtx);
1796 * Decrementing the dbuf count means that the hold
1797 * corresponding to the removed dbuf is no longer
1798 * discounted in dnode_move(), so the dnode cannot be
1799 * moved until after we release the hold.
1802 db->db_dnode_handle = NULL;
1804 dbuf_hash_remove(db);
1806 db->db_parent = NULL;
1809 ASSERT(!list_link_active(&db->db_link));
1810 ASSERT(db->db.db_data == NULL);
1811 ASSERT(db->db_hash_next == NULL);
1812 ASSERT(db->db_blkptr == NULL);
1813 ASSERT(db->db_data_pending == NULL);
1815 kmem_cache_free(dbuf_cache, db);
1816 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1820 dbuf_prefetch(dnode_t *dn, uint64_t blkid)
1822 dmu_buf_impl_t *db = NULL;
1823 blkptr_t *bp = NULL;
1825 ASSERT(blkid != DMU_BONUS_BLKID);
1826 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1828 if (dnode_block_freed(dn, blkid))
1831 /* dbuf_find() returns with db_mtx held */
1832 if (db = dbuf_find(dn, 0, blkid)) {
1834 * This dbuf is already in the cache. We assume that
1835 * it is already CACHED, or else about to be either
1838 mutex_exit(&db->db_mtx);
1842 if (dbuf_findbp(dn, 0, blkid, TRUE, &db, &bp) == 0) {
1843 if (bp && !BP_IS_HOLE(bp)) {
1844 int priority = dn->dn_type == DMU_OT_DDT_ZAP ?
1845 ZIO_PRIORITY_DDT_PREFETCH : ZIO_PRIORITY_ASYNC_READ;
1846 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
1847 uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH;
1850 SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
1851 dn->dn_object, 0, blkid);
1853 (void) arc_read(NULL, dn->dn_objset->os_spa,
1854 bp, NULL, NULL, priority,
1855 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1859 dbuf_rele(db, NULL);
1864 * Returns with db_holds incremented, and db_mtx not held.
1865 * Note: dn_struct_rwlock must be held.
1868 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse,
1869 void *tag, dmu_buf_impl_t **dbp)
1871 dmu_buf_impl_t *db, *parent = NULL;
1873 ASSERT(blkid != DMU_BONUS_BLKID);
1874 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1875 ASSERT3U(dn->dn_nlevels, >, level);
1879 /* dbuf_find() returns with db_mtx held */
1880 db = dbuf_find(dn, level, blkid);
1883 blkptr_t *bp = NULL;
1886 ASSERT3P(parent, ==, NULL);
1887 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
1889 if (err == 0 && bp && BP_IS_HOLE(bp))
1890 err = SET_ERROR(ENOENT);
1893 dbuf_rele(parent, NULL);
1897 if (err && err != ENOENT)
1899 db = dbuf_create(dn, level, blkid, parent, bp);
1902 if (db->db_buf && refcount_is_zero(&db->db_holds)) {
1903 arc_buf_add_ref(db->db_buf, db);
1904 if (db->db_buf->b_data == NULL) {
1907 dbuf_rele(parent, NULL);
1912 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
1915 ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
1918 * If this buffer is currently syncing out, and we are are
1919 * still referencing it from db_data, we need to make a copy
1920 * of it in case we decide we want to dirty it again in this txg.
1922 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1923 dn->dn_object != DMU_META_DNODE_OBJECT &&
1924 db->db_state == DB_CACHED && db->db_data_pending) {
1925 dbuf_dirty_record_t *dr = db->db_data_pending;
1927 if (dr->dt.dl.dr_data == db->db_buf) {
1928 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1931 arc_buf_alloc(dn->dn_objset->os_spa,
1932 db->db.db_size, db, type));
1933 bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
1938 (void) refcount_add(&db->db_holds, tag);
1939 dbuf_update_data(db);
1941 mutex_exit(&db->db_mtx);
1943 /* NOTE: we can't rele the parent until after we drop the db_mtx */
1945 dbuf_rele(parent, NULL);
1947 ASSERT3P(DB_DNODE(db), ==, dn);
1948 ASSERT3U(db->db_blkid, ==, blkid);
1949 ASSERT3U(db->db_level, ==, level);
1956 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
1959 int err = dbuf_hold_impl(dn, 0, blkid, FALSE, tag, &db);
1960 return (err ? NULL : db);
1964 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
1967 int err = dbuf_hold_impl(dn, level, blkid, FALSE, tag, &db);
1968 return (err ? NULL : db);
1972 dbuf_create_bonus(dnode_t *dn)
1974 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1976 ASSERT(dn->dn_bonus == NULL);
1977 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
1981 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
1983 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1986 if (db->db_blkid != DMU_SPILL_BLKID)
1987 return (SET_ERROR(ENOTSUP));
1989 blksz = SPA_MINBLOCKSIZE;
1990 if (blksz > SPA_MAXBLOCKSIZE)
1991 blksz = SPA_MAXBLOCKSIZE;
1993 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
1997 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1998 dbuf_new_size(db, blksz, tx);
1999 rw_exit(&dn->dn_struct_rwlock);
2006 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
2008 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
2011 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2013 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
2015 int64_t holds = refcount_add(&db->db_holds, tag);
2020 * If you call dbuf_rele() you had better not be referencing the dnode handle
2021 * unless you have some other direct or indirect hold on the dnode. (An indirect
2022 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2023 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2024 * dnode's parent dbuf evicting its dnode handles.
2026 #pragma weak dmu_buf_rele = dbuf_rele
2028 dbuf_rele(dmu_buf_impl_t *db, void *tag)
2030 mutex_enter(&db->db_mtx);
2031 dbuf_rele_and_unlock(db, tag);
2035 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2036 * db_dirtycnt and db_holds to be updated atomically.
2039 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2043 ASSERT(MUTEX_HELD(&db->db_mtx));
2047 * Remove the reference to the dbuf before removing its hold on the
2048 * dnode so we can guarantee in dnode_move() that a referenced bonus
2049 * buffer has a corresponding dnode hold.
2051 holds = refcount_remove(&db->db_holds, tag);
2055 * We can't freeze indirects if there is a possibility that they
2056 * may be modified in the current syncing context.
2058 if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0))
2059 arc_buf_freeze(db->db_buf);
2061 if (holds == db->db_dirtycnt &&
2062 db->db_level == 0 && db->db_immediate_evict)
2063 dbuf_evict_user(db);
2066 if (db->db_blkid == DMU_BONUS_BLKID) {
2067 mutex_exit(&db->db_mtx);
2070 * If the dnode moves here, we cannot cross this barrier
2071 * until the move completes.
2074 (void) atomic_dec_32_nv(&DB_DNODE(db)->dn_dbufs_count);
2077 * The bonus buffer's dnode hold is no longer discounted
2078 * in dnode_move(). The dnode cannot move until after
2081 dnode_rele(DB_DNODE(db), db);
2082 } else if (db->db_buf == NULL) {
2084 * This is a special case: we never associated this
2085 * dbuf with any data allocated from the ARC.
2087 ASSERT(db->db_state == DB_UNCACHED ||
2088 db->db_state == DB_NOFILL);
2090 } else if (arc_released(db->db_buf)) {
2091 arc_buf_t *buf = db->db_buf;
2093 * This dbuf has anonymous data associated with it.
2095 dbuf_set_data(db, NULL);
2096 VERIFY(arc_buf_remove_ref(buf, db));
2099 VERIFY(!arc_buf_remove_ref(db->db_buf, db));
2102 * A dbuf will be eligible for eviction if either the
2103 * 'primarycache' property is set or a duplicate
2104 * copy of this buffer is already cached in the arc.
2106 * In the case of the 'primarycache' a buffer
2107 * is considered for eviction if it matches the
2108 * criteria set in the property.
2110 * To decide if our buffer is considered a
2111 * duplicate, we must call into the arc to determine
2112 * if multiple buffers are referencing the same
2113 * block on-disk. If so, then we simply evict
2116 if (!DBUF_IS_CACHEABLE(db) ||
2117 arc_buf_eviction_needed(db->db_buf))
2120 mutex_exit(&db->db_mtx);
2123 mutex_exit(&db->db_mtx);
2127 #pragma weak dmu_buf_refcount = dbuf_refcount
2129 dbuf_refcount(dmu_buf_impl_t *db)
2131 return (refcount_count(&db->db_holds));
2135 dmu_buf_set_user(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr,
2136 dmu_buf_evict_func_t *evict_func)
2138 return (dmu_buf_update_user(db_fake, NULL, user_ptr,
2139 user_data_ptr_ptr, evict_func));
2143 dmu_buf_set_user_ie(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr,
2144 dmu_buf_evict_func_t *evict_func)
2146 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2148 db->db_immediate_evict = TRUE;
2149 return (dmu_buf_update_user(db_fake, NULL, user_ptr,
2150 user_data_ptr_ptr, evict_func));
2154 dmu_buf_update_user(dmu_buf_t *db_fake, void *old_user_ptr, void *user_ptr,
2155 void *user_data_ptr_ptr, dmu_buf_evict_func_t *evict_func)
2157 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2158 ASSERT(db->db_level == 0);
2160 ASSERT((user_ptr == NULL) == (evict_func == NULL));
2162 mutex_enter(&db->db_mtx);
2164 if (db->db_user_ptr == old_user_ptr) {
2165 db->db_user_ptr = user_ptr;
2166 db->db_user_data_ptr_ptr = user_data_ptr_ptr;
2167 db->db_evict_func = evict_func;
2169 dbuf_update_data(db);
2171 old_user_ptr = db->db_user_ptr;
2174 mutex_exit(&db->db_mtx);
2175 return (old_user_ptr);
2179 dmu_buf_get_user(dmu_buf_t *db_fake)
2181 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2182 ASSERT(!refcount_is_zero(&db->db_holds));
2184 return (db->db_user_ptr);
2188 dmu_buf_freeable(dmu_buf_t *dbuf)
2190 boolean_t res = B_FALSE;
2191 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2194 res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset,
2195 db->db_blkptr, db->db_blkptr->blk_birth);
2201 dmu_buf_get_blkptr(dmu_buf_t *db)
2203 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2204 return (dbi->db_blkptr);
2208 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
2210 /* ASSERT(dmu_tx_is_syncing(tx) */
2211 ASSERT(MUTEX_HELD(&db->db_mtx));
2213 if (db->db_blkptr != NULL)
2216 if (db->db_blkid == DMU_SPILL_BLKID) {
2217 db->db_blkptr = &dn->dn_phys->dn_spill;
2218 BP_ZERO(db->db_blkptr);
2221 if (db->db_level == dn->dn_phys->dn_nlevels-1) {
2223 * This buffer was allocated at a time when there was
2224 * no available blkptrs from the dnode, or it was
2225 * inappropriate to hook it in (i.e., nlevels mis-match).
2227 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
2228 ASSERT(db->db_parent == NULL);
2229 db->db_parent = dn->dn_dbuf;
2230 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
2233 dmu_buf_impl_t *parent = db->db_parent;
2234 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2236 ASSERT(dn->dn_phys->dn_nlevels > 1);
2237 if (parent == NULL) {
2238 mutex_exit(&db->db_mtx);
2239 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2240 (void) dbuf_hold_impl(dn, db->db_level+1,
2241 db->db_blkid >> epbs, FALSE, db, &parent);
2242 rw_exit(&dn->dn_struct_rwlock);
2243 mutex_enter(&db->db_mtx);
2244 db->db_parent = parent;
2246 db->db_blkptr = (blkptr_t *)parent->db.db_data +
2247 (db->db_blkid & ((1ULL << epbs) - 1));
2253 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2255 dmu_buf_impl_t *db = dr->dr_dbuf;
2259 ASSERT(dmu_tx_is_syncing(tx));
2261 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2263 mutex_enter(&db->db_mtx);
2265 ASSERT(db->db_level > 0);
2268 /* Read the block if it hasn't been read yet. */
2269 if (db->db_buf == NULL) {
2270 mutex_exit(&db->db_mtx);
2271 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
2272 mutex_enter(&db->db_mtx);
2274 ASSERT3U(db->db_state, ==, DB_CACHED);
2275 ASSERT(db->db_buf != NULL);
2279 /* Indirect block size must match what the dnode thinks it is. */
2280 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2281 dbuf_check_blkptr(dn, db);
2284 /* Provide the pending dirty record to child dbufs */
2285 db->db_data_pending = dr;
2287 mutex_exit(&db->db_mtx);
2288 dbuf_write(dr, db->db_buf, tx);
2291 mutex_enter(&dr->dt.di.dr_mtx);
2292 dbuf_sync_list(&dr->dt.di.dr_children, tx);
2293 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2294 mutex_exit(&dr->dt.di.dr_mtx);
2299 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2301 arc_buf_t **datap = &dr->dt.dl.dr_data;
2302 dmu_buf_impl_t *db = dr->dr_dbuf;
2305 uint64_t txg = tx->tx_txg;
2307 ASSERT(dmu_tx_is_syncing(tx));
2309 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2311 mutex_enter(&db->db_mtx);
2313 * To be synced, we must be dirtied. But we
2314 * might have been freed after the dirty.
2316 if (db->db_state == DB_UNCACHED) {
2317 /* This buffer has been freed since it was dirtied */
2318 ASSERT(db->db.db_data == NULL);
2319 } else if (db->db_state == DB_FILL) {
2320 /* This buffer was freed and is now being re-filled */
2321 ASSERT(db->db.db_data != dr->dt.dl.dr_data);
2323 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
2330 if (db->db_blkid == DMU_SPILL_BLKID) {
2331 mutex_enter(&dn->dn_mtx);
2332 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
2333 mutex_exit(&dn->dn_mtx);
2337 * If this is a bonus buffer, simply copy the bonus data into the
2338 * dnode. It will be written out when the dnode is synced (and it
2339 * will be synced, since it must have been dirty for dbuf_sync to
2342 if (db->db_blkid == DMU_BONUS_BLKID) {
2343 dbuf_dirty_record_t **drp;
2345 ASSERT(*datap != NULL);
2346 ASSERT0(db->db_level);
2347 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN);
2348 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen);
2351 if (*datap != db->db.db_data) {
2352 zio_buf_free(*datap, DN_MAX_BONUSLEN);
2353 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
2355 db->db_data_pending = NULL;
2356 drp = &db->db_last_dirty;
2358 drp = &(*drp)->dr_next;
2359 ASSERT(dr->dr_next == NULL);
2360 ASSERT(dr->dr_dbuf == db);
2362 if (dr->dr_dbuf->db_level != 0) {
2363 list_destroy(&dr->dt.di.dr_children);
2364 mutex_destroy(&dr->dt.di.dr_mtx);
2366 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2367 ASSERT(db->db_dirtycnt > 0);
2368 db->db_dirtycnt -= 1;
2369 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
2376 * This function may have dropped the db_mtx lock allowing a dmu_sync
2377 * operation to sneak in. As a result, we need to ensure that we
2378 * don't check the dr_override_state until we have returned from
2379 * dbuf_check_blkptr.
2381 dbuf_check_blkptr(dn, db);
2384 * If this buffer is in the middle of an immediate write,
2385 * wait for the synchronous IO to complete.
2387 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
2388 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
2389 cv_wait(&db->db_changed, &db->db_mtx);
2390 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
2393 if (db->db_state != DB_NOFILL &&
2394 dn->dn_object != DMU_META_DNODE_OBJECT &&
2395 refcount_count(&db->db_holds) > 1 &&
2396 dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
2397 *datap == db->db_buf) {
2399 * If this buffer is currently "in use" (i.e., there
2400 * are active holds and db_data still references it),
2401 * then make a copy before we start the write so that
2402 * any modifications from the open txg will not leak
2405 * NOTE: this copy does not need to be made for
2406 * objects only modified in the syncing context (e.g.
2407 * DNONE_DNODE blocks).
2409 int blksz = arc_buf_size(*datap);
2410 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2411 *datap = arc_buf_alloc(os->os_spa, blksz, db, type);
2412 bcopy(db->db.db_data, (*datap)->b_data, blksz);
2414 db->db_data_pending = dr;
2416 mutex_exit(&db->db_mtx);
2418 dbuf_write(dr, *datap, tx);
2420 ASSERT(!list_link_active(&dr->dr_dirty_node));
2421 if (dn->dn_object == DMU_META_DNODE_OBJECT) {
2422 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
2426 * Although zio_nowait() does not "wait for an IO", it does
2427 * initiate the IO. If this is an empty write it seems plausible
2428 * that the IO could actually be completed before the nowait
2429 * returns. We need to DB_DNODE_EXIT() first in case
2430 * zio_nowait() invalidates the dbuf.
2433 zio_nowait(dr->dr_zio);
2438 dbuf_sync_list(list_t *list, dmu_tx_t *tx)
2440 dbuf_dirty_record_t *dr;
2442 while (dr = list_head(list)) {
2443 if (dr->dr_zio != NULL) {
2445 * If we find an already initialized zio then we
2446 * are processing the meta-dnode, and we have finished.
2447 * The dbufs for all dnodes are put back on the list
2448 * during processing, so that we can zio_wait()
2449 * these IOs after initiating all child IOs.
2451 ASSERT3U(dr->dr_dbuf->db.db_object, ==,
2452 DMU_META_DNODE_OBJECT);
2455 list_remove(list, dr);
2456 if (dr->dr_dbuf->db_level > 0)
2457 dbuf_sync_indirect(dr, tx);
2459 dbuf_sync_leaf(dr, tx);
2465 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
2467 dmu_buf_impl_t *db = vdb;
2469 blkptr_t *bp = zio->io_bp;
2470 blkptr_t *bp_orig = &zio->io_bp_orig;
2471 spa_t *spa = zio->io_spa;
2476 ASSERT(db->db_blkptr == bp);
2480 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
2481 dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
2482 zio->io_prev_space_delta = delta;
2484 if (BP_IS_HOLE(bp)) {
2485 ASSERT(bp->blk_fill == 0);
2490 ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
2491 BP_GET_TYPE(bp) == dn->dn_type) ||
2492 (db->db_blkid == DMU_SPILL_BLKID &&
2493 BP_GET_TYPE(bp) == dn->dn_bonustype));
2494 ASSERT(BP_GET_LEVEL(bp) == db->db_level);
2496 mutex_enter(&db->db_mtx);
2499 if (db->db_blkid == DMU_SPILL_BLKID) {
2500 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2501 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2502 db->db_blkptr == &dn->dn_phys->dn_spill);
2506 if (db->db_level == 0) {
2507 mutex_enter(&dn->dn_mtx);
2508 if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
2509 db->db_blkid != DMU_SPILL_BLKID)
2510 dn->dn_phys->dn_maxblkid = db->db_blkid;
2511 mutex_exit(&dn->dn_mtx);
2513 if (dn->dn_type == DMU_OT_DNODE) {
2514 dnode_phys_t *dnp = db->db.db_data;
2515 for (i = db->db.db_size >> DNODE_SHIFT; i > 0;
2517 if (dnp->dn_type != DMU_OT_NONE)
2524 blkptr_t *ibp = db->db.db_data;
2525 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2526 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
2527 if (BP_IS_HOLE(ibp))
2529 fill += ibp->blk_fill;
2534 bp->blk_fill = fill;
2536 mutex_exit(&db->db_mtx);
2541 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
2543 dmu_buf_impl_t *db = vdb;
2544 blkptr_t *bp = zio->io_bp;
2545 blkptr_t *bp_orig = &zio->io_bp_orig;
2546 uint64_t txg = zio->io_txg;
2547 dbuf_dirty_record_t **drp, *dr;
2549 ASSERT0(zio->io_error);
2550 ASSERT(db->db_blkptr == bp);
2553 * For nopwrites and rewrites we ensure that the bp matches our
2554 * original and bypass all the accounting.
2556 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
2557 ASSERT(BP_EQUAL(bp, bp_orig));
2563 DB_GET_OBJSET(&os, db);
2564 ds = os->os_dsl_dataset;
2567 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
2568 dsl_dataset_block_born(ds, bp, tx);
2571 mutex_enter(&db->db_mtx);
2575 drp = &db->db_last_dirty;
2576 while ((dr = *drp) != db->db_data_pending)
2578 ASSERT(!list_link_active(&dr->dr_dirty_node));
2579 ASSERT(dr->dr_txg == txg);
2580 ASSERT(dr->dr_dbuf == db);
2581 ASSERT(dr->dr_next == NULL);
2585 if (db->db_blkid == DMU_SPILL_BLKID) {
2590 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2591 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2592 db->db_blkptr == &dn->dn_phys->dn_spill);
2597 if (db->db_level == 0) {
2598 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2599 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
2600 if (db->db_state != DB_NOFILL) {
2601 if (dr->dt.dl.dr_data != db->db_buf)
2602 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data,
2604 else if (!arc_released(db->db_buf))
2605 arc_set_callback(db->db_buf, dbuf_do_evict, db);
2612 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2613 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2614 if (!BP_IS_HOLE(db->db_blkptr)) {
2616 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2617 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
2619 ASSERT3U(dn->dn_phys->dn_maxblkid
2620 >> (db->db_level * epbs), >=, db->db_blkid);
2621 arc_set_callback(db->db_buf, dbuf_do_evict, db);
2624 mutex_destroy(&dr->dt.di.dr_mtx);
2625 list_destroy(&dr->dt.di.dr_children);
2627 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2629 cv_broadcast(&db->db_changed);
2630 ASSERT(db->db_dirtycnt > 0);
2631 db->db_dirtycnt -= 1;
2632 db->db_data_pending = NULL;
2633 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
2637 dbuf_write_nofill_ready(zio_t *zio)
2639 dbuf_write_ready(zio, NULL, zio->io_private);
2643 dbuf_write_nofill_done(zio_t *zio)
2645 dbuf_write_done(zio, NULL, zio->io_private);
2649 dbuf_write_override_ready(zio_t *zio)
2651 dbuf_dirty_record_t *dr = zio->io_private;
2652 dmu_buf_impl_t *db = dr->dr_dbuf;
2654 dbuf_write_ready(zio, NULL, db);
2658 dbuf_write_override_done(zio_t *zio)
2660 dbuf_dirty_record_t *dr = zio->io_private;
2661 dmu_buf_impl_t *db = dr->dr_dbuf;
2662 blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
2664 mutex_enter(&db->db_mtx);
2665 if (!BP_EQUAL(zio->io_bp, obp)) {
2666 if (!BP_IS_HOLE(obp))
2667 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
2668 arc_release(dr->dt.dl.dr_data, db);
2670 mutex_exit(&db->db_mtx);
2672 dbuf_write_done(zio, NULL, db);
2675 /* Issue I/O to commit a dirty buffer to disk. */
2677 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
2679 dmu_buf_impl_t *db = dr->dr_dbuf;
2682 dmu_buf_impl_t *parent = db->db_parent;
2683 uint64_t txg = tx->tx_txg;
2693 if (db->db_state != DB_NOFILL) {
2694 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
2696 * Private object buffers are released here rather
2697 * than in dbuf_dirty() since they are only modified
2698 * in the syncing context and we don't want the
2699 * overhead of making multiple copies of the data.
2701 if (BP_IS_HOLE(db->db_blkptr)) {
2704 dbuf_release_bp(db);
2709 if (parent != dn->dn_dbuf) {
2710 /* Our parent is an indirect block. */
2711 /* We have a dirty parent that has been scheduled for write. */
2712 ASSERT(parent && parent->db_data_pending);
2713 /* Our parent's buffer is one level closer to the dnode. */
2714 ASSERT(db->db_level == parent->db_level-1);
2716 * We're about to modify our parent's db_data by modifying
2717 * our block pointer, so the parent must be released.
2719 ASSERT(arc_released(parent->db_buf));
2720 zio = parent->db_data_pending->dr_zio;
2722 /* Our parent is the dnode itself. */
2723 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
2724 db->db_blkid != DMU_SPILL_BLKID) ||
2725 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
2726 if (db->db_blkid != DMU_SPILL_BLKID)
2727 ASSERT3P(db->db_blkptr, ==,
2728 &dn->dn_phys->dn_blkptr[db->db_blkid]);
2732 ASSERT(db->db_level == 0 || data == db->db_buf);
2733 ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
2736 SET_BOOKMARK(&zb, os->os_dsl_dataset ?
2737 os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
2738 db->db.db_object, db->db_level, db->db_blkid);
2740 if (db->db_blkid == DMU_SPILL_BLKID)
2742 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
2744 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
2747 if (db->db_level == 0 && dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
2748 ASSERT(db->db_state != DB_NOFILL);
2749 dr->dr_zio = zio_write(zio, os->os_spa, txg,
2750 db->db_blkptr, data->b_data, arc_buf_size(data), &zp,
2751 dbuf_write_override_ready, dbuf_write_override_done, dr,
2752 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
2753 mutex_enter(&db->db_mtx);
2754 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
2755 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
2756 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
2757 mutex_exit(&db->db_mtx);
2758 } else if (db->db_state == DB_NOFILL) {
2759 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF ||
2760 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY);
2761 dr->dr_zio = zio_write(zio, os->os_spa, txg,
2762 db->db_blkptr, NULL, db->db.db_size, &zp,
2763 dbuf_write_nofill_ready, dbuf_write_nofill_done, db,
2764 ZIO_PRIORITY_ASYNC_WRITE,
2765 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
2767 ASSERT(arc_released(data));
2768 dr->dr_zio = arc_write(zio, os->os_spa, txg,
2769 db->db_blkptr, data, DBUF_IS_L2CACHEABLE(db),
2770 DBUF_IS_L2COMPRESSIBLE(db), &zp, dbuf_write_ready,
2771 dbuf_write_done, db, ZIO_PRIORITY_ASYNC_WRITE,
2772 ZIO_FLAG_MUSTSUCCEED, &zb);