xref: /illumos-gate/usr/src/uts/common/fs/zfs/dbuf.c (revision dcbf3bd6)
1fa9e4066Sahrens /*
2fa9e4066Sahrens  * CDDL HEADER START
3fa9e4066Sahrens  *
4fa9e4066Sahrens  * The contents of this file are subject to the terms of the
5f65e61c0Sahrens  * Common Development and Distribution License (the "License").
6f65e61c0Sahrens  * You may not use this file except in compliance with the License.
7fa9e4066Sahrens  *
8fa9e4066Sahrens  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9fa9e4066Sahrens  * or http://www.opensolaris.org/os/licensing.
10fa9e4066Sahrens  * See the License for the specific language governing permissions
11fa9e4066Sahrens  * and limitations under the License.
12fa9e4066Sahrens  *
13fa9e4066Sahrens  * When distributing Covered Code, include this CDDL HEADER in each
14fa9e4066Sahrens  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15fa9e4066Sahrens  * If applicable, add the following below this CDDL HEADER, with the
16fa9e4066Sahrens  * fields enclosed by brackets "[]" replaced with your own identifying
17fa9e4066Sahrens  * information: Portions Copyright [yyyy] [name of copyright owner]
18fa9e4066Sahrens  *
19fa9e4066Sahrens  * CDDL HEADER END
20fa9e4066Sahrens  */
21fa9e4066Sahrens /*
2206e0070dSMark Shellenbaum  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
233f2366c2SGordon Ross  * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
2446e1baa6SMatthew Ahrens  * Copyright (c) 2012, 2015 by Delphix. All rights reserved.
25aad02571SSaso Kiselkov  * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
26810e43b2SBill Pijewski  * Copyright (c) 2013, Joyent, Inc. All rights reserved.
27bc9014e6SJustin Gibbs  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
28c3d26abcSMatthew Ahrens  * Copyright (c) 2014 Integros [integros.com]
29fa9e4066Sahrens  */
30fa9e4066Sahrens 
31fa9e4066Sahrens #include <sys/zfs_context.h>
32fa9e4066Sahrens #include <sys/dmu.h>
332f3d8780SMatthew Ahrens #include <sys/dmu_send.h>
34fa9e4066Sahrens #include <sys/dmu_impl.h>
35fa9e4066Sahrens #include <sys/dbuf.h>
36fa9e4066Sahrens #include <sys/dmu_objset.h>
37fa9e4066Sahrens #include <sys/dsl_dataset.h>
38fa9e4066Sahrens #include <sys/dsl_dir.h>
39fa9e4066Sahrens #include <sys/dmu_tx.h>
40fa9e4066Sahrens #include <sys/spa.h>
41fa9e4066Sahrens #include <sys/zio.h>
42fa9e4066Sahrens #include <sys/dmu_zfetch.h>
430a586ceaSMark Shellenbaum #include <sys/sa.h>
440a586ceaSMark Shellenbaum #include <sys/sa_impl.h>
455d7b4d43SMatthew Ahrens #include <sys/zfeature.h>
465d7b4d43SMatthew Ahrens #include <sys/blkptr.h>
47bf16b11eSMatthew Ahrens #include <sys/range_tree.h>
48*dcbf3bd6SGeorge Wilson #include <sys/callb.h>
49*dcbf3bd6SGeorge Wilson 
50*dcbf3bd6SGeorge Wilson uint_t zfs_dbuf_evict_key;
51fa9e4066Sahrens 
52713d6c20SMatthew Ahrens /*
53713d6c20SMatthew Ahrens  * Number of times that zfs_free_range() took the slow path while doing
54713d6c20SMatthew Ahrens  * a zfs receive.  A nonzero value indicates a potential performance problem.
55713d6c20SMatthew Ahrens  */
56713d6c20SMatthew Ahrens uint64_t zfs_free_range_recv_miss;
57713d6c20SMatthew Ahrens 
583b2aab18SMatthew Ahrens static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
59088f3894Sahrens static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
60fa9e4066Sahrens 
61bc9014e6SJustin Gibbs #ifndef __lint
62bc9014e6SJustin Gibbs extern inline void dmu_buf_init_user(dmu_buf_user_t *dbu,
63bc9014e6SJustin Gibbs     dmu_buf_evict_func_t *evict_func, dmu_buf_t **clear_on_evict_dbufp);
64bc9014e6SJustin Gibbs #endif /* ! __lint */
65bc9014e6SJustin Gibbs 
66fa9e4066Sahrens /*
67fa9e4066Sahrens  * Global data structures and functions for the dbuf cache.
68fa9e4066Sahrens  */
69*dcbf3bd6SGeorge Wilson static kmem_cache_t *dbuf_kmem_cache;
70bc9014e6SJustin Gibbs static taskq_t *dbu_evict_taskq;
71fa9e4066Sahrens 
72*dcbf3bd6SGeorge Wilson static kthread_t *dbuf_cache_evict_thread;
73*dcbf3bd6SGeorge Wilson static kmutex_t dbuf_evict_lock;
74*dcbf3bd6SGeorge Wilson static kcondvar_t dbuf_evict_cv;
75*dcbf3bd6SGeorge Wilson static boolean_t dbuf_evict_thread_exit;
76*dcbf3bd6SGeorge Wilson 
77*dcbf3bd6SGeorge Wilson /*
78*dcbf3bd6SGeorge Wilson  * LRU cache of dbufs. The dbuf cache maintains a list of dbufs that
79*dcbf3bd6SGeorge Wilson  * are not currently held but have been recently released. These dbufs
80*dcbf3bd6SGeorge Wilson  * are not eligible for arc eviction until they are aged out of the cache.
81*dcbf3bd6SGeorge Wilson  * Dbufs are added to the dbuf cache once the last hold is released. If a
82*dcbf3bd6SGeorge Wilson  * dbuf is later accessed and still exists in the dbuf cache, then it will
83*dcbf3bd6SGeorge Wilson  * be removed from the cache and later re-added to the head of the cache.
84*dcbf3bd6SGeorge Wilson  * Dbufs that are aged out of the cache will be immediately destroyed and
85*dcbf3bd6SGeorge Wilson  * become eligible for arc eviction.
86*dcbf3bd6SGeorge Wilson  */
87*dcbf3bd6SGeorge Wilson static multilist_t dbuf_cache;
88*dcbf3bd6SGeorge Wilson static refcount_t dbuf_cache_size;
89*dcbf3bd6SGeorge Wilson uint64_t dbuf_cache_max_bytes = 100 * 1024 * 1024;
90*dcbf3bd6SGeorge Wilson 
91*dcbf3bd6SGeorge Wilson /* Cap the size of the dbuf cache to log2 fraction of arc size. */
92*dcbf3bd6SGeorge Wilson int dbuf_cache_max_shift = 5;
93*dcbf3bd6SGeorge Wilson 
94*dcbf3bd6SGeorge Wilson /*
95*dcbf3bd6SGeorge Wilson  * The dbuf cache uses a three-stage eviction policy:
96*dcbf3bd6SGeorge Wilson  *	- A low water marker designates when the dbuf eviction thread
97*dcbf3bd6SGeorge Wilson  *	should stop evicting from the dbuf cache.
98*dcbf3bd6SGeorge Wilson  *	- When we reach the maximum size (aka mid water mark), we
99*dcbf3bd6SGeorge Wilson  *	signal the eviction thread to run.
100*dcbf3bd6SGeorge Wilson  *	- The high water mark indicates when the eviction thread
101*dcbf3bd6SGeorge Wilson  *	is unable to keep up with the incoming load and eviction must
102*dcbf3bd6SGeorge Wilson  *	happen in the context of the calling thread.
103*dcbf3bd6SGeorge Wilson  *
104*dcbf3bd6SGeorge Wilson  * The dbuf cache:
105*dcbf3bd6SGeorge Wilson  *                                                 (max size)
106*dcbf3bd6SGeorge Wilson  *                                      low water   mid water   hi water
107*dcbf3bd6SGeorge Wilson  * +----------------------------------------+----------+----------+
108*dcbf3bd6SGeorge Wilson  * |                                        |          |          |
109*dcbf3bd6SGeorge Wilson  * |                                        |          |          |
110*dcbf3bd6SGeorge Wilson  * |                                        |          |          |
111*dcbf3bd6SGeorge Wilson  * |                                        |          |          |
112*dcbf3bd6SGeorge Wilson  * +----------------------------------------+----------+----------+
113*dcbf3bd6SGeorge Wilson  *                                        stop        signal     evict
114*dcbf3bd6SGeorge Wilson  *                                      evicting     eviction   directly
115*dcbf3bd6SGeorge Wilson  *                                                    thread
116*dcbf3bd6SGeorge Wilson  *
117*dcbf3bd6SGeorge Wilson  * The high and low water marks indicate the operating range for the eviction
118*dcbf3bd6SGeorge Wilson  * thread. The low water mark is, by default, 90% of the total size of the
119*dcbf3bd6SGeorge Wilson  * cache and the high water mark is at 110% (both of these percentages can be
120*dcbf3bd6SGeorge Wilson  * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct,
121*dcbf3bd6SGeorge Wilson  * respectively). The eviction thread will try to ensure that the cache remains
122*dcbf3bd6SGeorge Wilson  * within this range by waking up every second and checking if the cache is
123*dcbf3bd6SGeorge Wilson  * above the low water mark. The thread can also be woken up by callers adding
124*dcbf3bd6SGeorge Wilson  * elements into the cache if the cache is larger than the mid water (i.e max
125*dcbf3bd6SGeorge Wilson  * cache size). Once the eviction thread is woken up and eviction is required,
126*dcbf3bd6SGeorge Wilson  * it will continue evicting buffers until it's able to reduce the cache size
127*dcbf3bd6SGeorge Wilson  * to the low water mark. If the cache size continues to grow and hits the high
128*dcbf3bd6SGeorge Wilson  * water mark, then callers adding elments to the cache will begin to evict
129*dcbf3bd6SGeorge Wilson  * directly from the cache until the cache is no longer above the high water
130*dcbf3bd6SGeorge Wilson  * mark.
131*dcbf3bd6SGeorge Wilson  */
132*dcbf3bd6SGeorge Wilson 
133*dcbf3bd6SGeorge Wilson /*
134*dcbf3bd6SGeorge Wilson  * The percentage above and below the maximum cache size.
135*dcbf3bd6SGeorge Wilson  */
136*dcbf3bd6SGeorge Wilson uint_t dbuf_cache_hiwater_pct = 10;
137*dcbf3bd6SGeorge Wilson uint_t dbuf_cache_lowater_pct = 10;
138*dcbf3bd6SGeorge Wilson 
139fa9e4066Sahrens /* ARGSUSED */
140fa9e4066Sahrens static int
141fa9e4066Sahrens dbuf_cons(void *vdb, void *unused, int kmflag)
142fa9e4066Sahrens {
143fa9e4066Sahrens 	dmu_buf_impl_t *db = vdb;
144fa9e4066Sahrens 	bzero(db, sizeof (dmu_buf_impl_t));
145fa9e4066Sahrens 
146fa9e4066Sahrens 	mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
147fa9e4066Sahrens 	cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
148*dcbf3bd6SGeorge Wilson 	multilist_link_init(&db->db_cache_link);
149fa9e4066Sahrens 	refcount_create(&db->db_holds);
1500f6d88adSAlex Reece 
151fa9e4066Sahrens 	return (0);
152fa9e4066Sahrens }
153fa9e4066Sahrens 
154fa9e4066Sahrens /* ARGSUSED */
155fa9e4066Sahrens static void
156fa9e4066Sahrens dbuf_dest(void *vdb, void *unused)
157fa9e4066Sahrens {
158fa9e4066Sahrens 	dmu_buf_impl_t *db = vdb;
159fa9e4066Sahrens 	mutex_destroy(&db->db_mtx);
160fa9e4066Sahrens 	cv_destroy(&db->db_changed);
161*dcbf3bd6SGeorge Wilson 	ASSERT(!multilist_link_active(&db->db_cache_link));
162fa9e4066Sahrens 	refcount_destroy(&db->db_holds);
163fa9e4066Sahrens }
164fa9e4066Sahrens 
165fa9e4066Sahrens /*
166fa9e4066Sahrens  * dbuf hash table routines
167fa9e4066Sahrens  */
168fa9e4066Sahrens static dbuf_hash_table_t dbuf_hash_table;
169fa9e4066Sahrens 
170fa9e4066Sahrens static uint64_t dbuf_hash_count;
171fa9e4066Sahrens 
172fa9e4066Sahrens static uint64_t
173fa9e4066Sahrens dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
174fa9e4066Sahrens {
175fa9e4066Sahrens 	uintptr_t osv = (uintptr_t)os;
176fa9e4066Sahrens 	uint64_t crc = -1ULL;
177fa9e4066Sahrens 
178fa9e4066Sahrens 	ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
179fa9e4066Sahrens 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF];
180fa9e4066Sahrens 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF];
181fa9e4066Sahrens 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF];
182fa9e4066Sahrens 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF];
183fa9e4066Sahrens 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF];
184fa9e4066Sahrens 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF];
185fa9e4066Sahrens 
186fa9e4066Sahrens 	crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16);
187fa9e4066Sahrens 
188fa9e4066Sahrens 	return (crc);
189fa9e4066Sahrens }
190fa9e4066Sahrens 
191fa9e4066Sahrens #define	DBUF_EQUAL(dbuf, os, obj, level, blkid)		\
192fa9e4066Sahrens 	((dbuf)->db.db_object == (obj) &&		\
193fa9e4066Sahrens 	(dbuf)->db_objset == (os) &&			\
194fa9e4066Sahrens 	(dbuf)->db_level == (level) &&			\
195fa9e4066Sahrens 	(dbuf)->db_blkid == (blkid))
196fa9e4066Sahrens 
197fa9e4066Sahrens dmu_buf_impl_t *
198e57a022bSJustin T. Gibbs dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid)
199fa9e4066Sahrens {
200fa9e4066Sahrens 	dbuf_hash_table_t *h = &dbuf_hash_table;
201*dcbf3bd6SGeorge Wilson 	uint64_t hv = dbuf_hash(os, obj, level, blkid);
202fa9e4066Sahrens 	uint64_t idx = hv & h->hash_table_mask;
203fa9e4066Sahrens 	dmu_buf_impl_t *db;
204fa9e4066Sahrens 
205fa9e4066Sahrens 	mutex_enter(DBUF_HASH_MUTEX(h, idx));
206fa9e4066Sahrens 	for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
207fa9e4066Sahrens 		if (DBUF_EQUAL(db, os, obj, level, blkid)) {
208fa9e4066Sahrens 			mutex_enter(&db->db_mtx);
209ea8dc4b6Seschrock 			if (db->db_state != DB_EVICTING) {
210fa9e4066Sahrens 				mutex_exit(DBUF_HASH_MUTEX(h, idx));
211fa9e4066Sahrens 				return (db);
212fa9e4066Sahrens 			}
213fa9e4066Sahrens 			mutex_exit(&db->db_mtx);
214fa9e4066Sahrens 		}
215fa9e4066Sahrens 	}
216fa9e4066Sahrens 	mutex_exit(DBUF_HASH_MUTEX(h, idx));
217fa9e4066Sahrens 	return (NULL);
218fa9e4066Sahrens }
219fa9e4066Sahrens 
220e57a022bSJustin T. Gibbs static dmu_buf_impl_t *
221e57a022bSJustin T. Gibbs dbuf_find_bonus(objset_t *os, uint64_t object)
222e57a022bSJustin T. Gibbs {
223e57a022bSJustin T. Gibbs 	dnode_t *dn;
224e57a022bSJustin T. Gibbs 	dmu_buf_impl_t *db = NULL;
225e57a022bSJustin T. Gibbs 
226e57a022bSJustin T. Gibbs 	if (dnode_hold(os, object, FTAG, &dn) == 0) {
227e57a022bSJustin T. Gibbs 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
228e57a022bSJustin T. Gibbs 		if (dn->dn_bonus != NULL) {
229e57a022bSJustin T. Gibbs 			db = dn->dn_bonus;
230e57a022bSJustin T. Gibbs 			mutex_enter(&db->db_mtx);
231e57a022bSJustin T. Gibbs 		}
232e57a022bSJustin T. Gibbs 		rw_exit(&dn->dn_struct_rwlock);
233e57a022bSJustin T. Gibbs 		dnode_rele(dn, FTAG);
234e57a022bSJustin T. Gibbs 	}
235e57a022bSJustin T. Gibbs 	return (db);
236e57a022bSJustin T. Gibbs }
237e57a022bSJustin T. Gibbs 
238fa9e4066Sahrens /*
239fa9e4066Sahrens  * Insert an entry into the hash table.  If there is already an element
240fa9e4066Sahrens  * equal to elem in the hash table, then the already existing element
241fa9e4066Sahrens  * will be returned and the new element will not be inserted.
242fa9e4066Sahrens  * Otherwise returns NULL.
243fa9e4066Sahrens  */
244fa9e4066Sahrens static dmu_buf_impl_t *
245fa9e4066Sahrens dbuf_hash_insert(dmu_buf_impl_t *db)
246fa9e4066Sahrens {
247fa9e4066Sahrens 	dbuf_hash_table_t *h = &dbuf_hash_table;
248503ad85cSMatthew Ahrens 	objset_t *os = db->db_objset;
249fa9e4066Sahrens 	uint64_t obj = db->db.db_object;
250fa9e4066Sahrens 	int level = db->db_level;
251fa9e4066Sahrens 	uint64_t blkid = db->db_blkid;
252*dcbf3bd6SGeorge Wilson 	uint64_t hv = dbuf_hash(os, obj, level, blkid);
253fa9e4066Sahrens 	uint64_t idx = hv & h->hash_table_mask;
254fa9e4066Sahrens 	dmu_buf_impl_t *dbf;
255fa9e4066Sahrens 
256fa9e4066Sahrens 	mutex_enter(DBUF_HASH_MUTEX(h, idx));
257fa9e4066Sahrens 	for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) {
258fa9e4066Sahrens 		if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
259fa9e4066Sahrens 			mutex_enter(&dbf->db_mtx);
260ea8dc4b6Seschrock 			if (dbf->db_state != DB_EVICTING) {
261fa9e4066Sahrens 				mutex_exit(DBUF_HASH_MUTEX(h, idx));
262fa9e4066Sahrens 				return (dbf);
263fa9e4066Sahrens 			}
264fa9e4066Sahrens 			mutex_exit(&dbf->db_mtx);
265fa9e4066Sahrens 		}
266fa9e4066Sahrens 	}
267fa9e4066Sahrens 
268fa9e4066Sahrens 	mutex_enter(&db->db_mtx);
269fa9e4066Sahrens 	db->db_hash_next = h->hash_table[idx];
270fa9e4066Sahrens 	h->hash_table[idx] = db;
271fa9e4066Sahrens 	mutex_exit(DBUF_HASH_MUTEX(h, idx));
2721a5e258fSJosef 'Jeff' Sipek 	atomic_inc_64(&dbuf_hash_count);
273fa9e4066Sahrens 
274fa9e4066Sahrens 	return (NULL);
275fa9e4066Sahrens }
276fa9e4066Sahrens 
277fa9e4066Sahrens /*
278bbfa8ea8SMatthew Ahrens  * Remove an entry from the hash table.  It must be in the EVICTING state.
279fa9e4066Sahrens  */
280fa9e4066Sahrens static void
281fa9e4066Sahrens dbuf_hash_remove(dmu_buf_impl_t *db)
282fa9e4066Sahrens {
283fa9e4066Sahrens 	dbuf_hash_table_t *h = &dbuf_hash_table;
284*dcbf3bd6SGeorge Wilson 	uint64_t hv = dbuf_hash(db->db_objset, db->db.db_object,
285fa9e4066Sahrens 	    db->db_level, db->db_blkid);
286fa9e4066Sahrens 	uint64_t idx = hv & h->hash_table_mask;
287fa9e4066Sahrens 	dmu_buf_impl_t *dbf, **dbp;
288fa9e4066Sahrens 
289fa9e4066Sahrens 	/*
290bbfa8ea8SMatthew Ahrens 	 * We musn't hold db_mtx to maintain lock ordering:
291fa9e4066Sahrens 	 * DBUF_HASH_MUTEX > db_mtx.
292fa9e4066Sahrens 	 */
293fa9e4066Sahrens 	ASSERT(refcount_is_zero(&db->db_holds));
294ea8dc4b6Seschrock 	ASSERT(db->db_state == DB_EVICTING);
295fa9e4066Sahrens 	ASSERT(!MUTEX_HELD(&db->db_mtx));
296fa9e4066Sahrens 
297fa9e4066Sahrens 	mutex_enter(DBUF_HASH_MUTEX(h, idx));
298fa9e4066Sahrens 	dbp = &h->hash_table[idx];
299fa9e4066Sahrens 	while ((dbf = *dbp) != db) {
300fa9e4066Sahrens 		dbp = &dbf->db_hash_next;
301fa9e4066Sahrens 		ASSERT(dbf != NULL);
302fa9e4066Sahrens 	}
303fa9e4066Sahrens 	*dbp = db->db_hash_next;
304fa9e4066Sahrens 	db->db_hash_next = NULL;
305fa9e4066Sahrens 	mutex_exit(DBUF_HASH_MUTEX(h, idx));
3061a5e258fSJosef 'Jeff' Sipek 	atomic_dec_64(&dbuf_hash_count);
307fa9e4066Sahrens }
308fa9e4066Sahrens 
309bc9014e6SJustin Gibbs typedef enum {
310bc9014e6SJustin Gibbs 	DBVU_EVICTING,
311bc9014e6SJustin Gibbs 	DBVU_NOT_EVICTING
312bc9014e6SJustin Gibbs } dbvu_verify_type_t;
313bc9014e6SJustin Gibbs 
314bc9014e6SJustin Gibbs static void
315bc9014e6SJustin Gibbs dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type)
316bc9014e6SJustin Gibbs {
317bc9014e6SJustin Gibbs #ifdef ZFS_DEBUG
318bc9014e6SJustin Gibbs 	int64_t holds;
319bc9014e6SJustin Gibbs 
320bc9014e6SJustin Gibbs 	if (db->db_user == NULL)
321bc9014e6SJustin Gibbs 		return;
322bc9014e6SJustin Gibbs 
323bc9014e6SJustin Gibbs 	/* Only data blocks support the attachment of user data. */
324bc9014e6SJustin Gibbs 	ASSERT(db->db_level == 0);
325bc9014e6SJustin Gibbs 
326bc9014e6SJustin Gibbs 	/* Clients must resolve a dbuf before attaching user data. */
327bc9014e6SJustin Gibbs 	ASSERT(db->db.db_data != NULL);
328bc9014e6SJustin Gibbs 	ASSERT3U(db->db_state, ==, DB_CACHED);
329bc9014e6SJustin Gibbs 
330bc9014e6SJustin Gibbs 	holds = refcount_count(&db->db_holds);
331bc9014e6SJustin Gibbs 	if (verify_type == DBVU_EVICTING) {
332bc9014e6SJustin Gibbs 		/*
333bc9014e6SJustin Gibbs 		 * Immediate eviction occurs when holds == dirtycnt.
334bc9014e6SJustin Gibbs 		 * For normal eviction buffers, holds is zero on
335bc9014e6SJustin Gibbs 		 * eviction, except when dbuf_fix_old_data() calls
336bc9014e6SJustin Gibbs 		 * dbuf_clear_data().  However, the hold count can grow
337bc9014e6SJustin Gibbs 		 * during eviction even though db_mtx is held (see
338bc9014e6SJustin Gibbs 		 * dmu_bonus_hold() for an example), so we can only
339bc9014e6SJustin Gibbs 		 * test the generic invariant that holds >= dirtycnt.
340bc9014e6SJustin Gibbs 		 */
341bc9014e6SJustin Gibbs 		ASSERT3U(holds, >=, db->db_dirtycnt);
342bc9014e6SJustin Gibbs 	} else {
343d2058105SJustin T. Gibbs 		if (db->db_user_immediate_evict == TRUE)
344bc9014e6SJustin Gibbs 			ASSERT3U(holds, >=, db->db_dirtycnt);
345bc9014e6SJustin Gibbs 		else
346bc9014e6SJustin Gibbs 			ASSERT3U(holds, >, 0);
347bc9014e6SJustin Gibbs 	}
348bc9014e6SJustin Gibbs #endif
349bc9014e6SJustin Gibbs }
350bc9014e6SJustin Gibbs 
351fa9e4066Sahrens static void
352fa9e4066Sahrens dbuf_evict_user(dmu_buf_impl_t *db)
353fa9e4066Sahrens {
354bc9014e6SJustin Gibbs 	dmu_buf_user_t *dbu = db->db_user;
355bc9014e6SJustin Gibbs 
356fa9e4066Sahrens 	ASSERT(MUTEX_HELD(&db->db_mtx));
357fa9e4066Sahrens 
358bc9014e6SJustin Gibbs 	if (dbu == NULL)
359fa9e4066Sahrens 		return;
360fa9e4066Sahrens 
361bc9014e6SJustin Gibbs 	dbuf_verify_user(db, DBVU_EVICTING);
362bc9014e6SJustin Gibbs 	db->db_user = NULL;
363bc9014e6SJustin Gibbs 
364bc9014e6SJustin Gibbs #ifdef ZFS_DEBUG
365bc9014e6SJustin Gibbs 	if (dbu->dbu_clear_on_evict_dbufp != NULL)
366bc9014e6SJustin Gibbs 		*dbu->dbu_clear_on_evict_dbufp = NULL;
367bc9014e6SJustin Gibbs #endif
368bc9014e6SJustin Gibbs 
369bc9014e6SJustin Gibbs 	/*
370bc9014e6SJustin Gibbs 	 * Invoke the callback from a taskq to avoid lock order reversals
371bc9014e6SJustin Gibbs 	 * and limit stack depth.
372bc9014e6SJustin Gibbs 	 */
373bc9014e6SJustin Gibbs 	taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func, dbu, 0,
374bc9014e6SJustin Gibbs 	    &dbu->dbu_tqent);
375fa9e4066Sahrens }
376fa9e4066Sahrens 
377744947dcSTom Erickson boolean_t
378744947dcSTom Erickson dbuf_is_metadata(dmu_buf_impl_t *db)
379744947dcSTom Erickson {
380744947dcSTom Erickson 	if (db->db_level > 0) {
381744947dcSTom Erickson 		return (B_TRUE);
382744947dcSTom Erickson 	} else {
383744947dcSTom Erickson 		boolean_t is_metadata;
384744947dcSTom Erickson 
385744947dcSTom Erickson 		DB_DNODE_ENTER(db);
386ad135b5dSChristopher Siden 		is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type);
387744947dcSTom Erickson 		DB_DNODE_EXIT(db);
388744947dcSTom Erickson 
389744947dcSTom Erickson 		return (is_metadata);
390744947dcSTom Erickson 	}
391744947dcSTom Erickson }
392744947dcSTom Erickson 
393*dcbf3bd6SGeorge Wilson /*
394*dcbf3bd6SGeorge Wilson  * This function *must* return indices evenly distributed between all
395*dcbf3bd6SGeorge Wilson  * sublists of the multilist. This is needed due to how the dbuf eviction
396*dcbf3bd6SGeorge Wilson  * code is laid out; dbuf_evict_thread() assumes dbufs are evenly
397*dcbf3bd6SGeorge Wilson  * distributed between all sublists and uses this assumption when
398*dcbf3bd6SGeorge Wilson  * deciding which sublist to evict from and how much to evict from it.
399*dcbf3bd6SGeorge Wilson  */
400*dcbf3bd6SGeorge Wilson unsigned int
401*dcbf3bd6SGeorge Wilson dbuf_cache_multilist_index_func(multilist_t *ml, void *obj)
402ea8dc4b6Seschrock {
403*dcbf3bd6SGeorge Wilson 	dmu_buf_impl_t *db = obj;
404*dcbf3bd6SGeorge Wilson 
405*dcbf3bd6SGeorge Wilson 	/*
406*dcbf3bd6SGeorge Wilson 	 * The assumption here, is the hash value for a given
407*dcbf3bd6SGeorge Wilson 	 * dmu_buf_impl_t will remain constant throughout it's lifetime
408*dcbf3bd6SGeorge Wilson 	 * (i.e. it's objset, object, level and blkid fields don't change).
409*dcbf3bd6SGeorge Wilson 	 * Thus, we don't need to store the dbuf's sublist index
410*dcbf3bd6SGeorge Wilson 	 * on insertion, as this index can be recalculated on removal.
411*dcbf3bd6SGeorge Wilson 	 *
412*dcbf3bd6SGeorge Wilson 	 * Also, the low order bits of the hash value are thought to be
413*dcbf3bd6SGeorge Wilson 	 * distributed evenly. Otherwise, in the case that the multilist
414*dcbf3bd6SGeorge Wilson 	 * has a power of two number of sublists, each sublists' usage
415*dcbf3bd6SGeorge Wilson 	 * would not be evenly distributed.
416*dcbf3bd6SGeorge Wilson 	 */
417*dcbf3bd6SGeorge Wilson 	return (dbuf_hash(db->db_objset, db->db.db_object,
418*dcbf3bd6SGeorge Wilson 	    db->db_level, db->db_blkid) %
419*dcbf3bd6SGeorge Wilson 	    multilist_get_num_sublists(ml));
420*dcbf3bd6SGeorge Wilson }
421*dcbf3bd6SGeorge Wilson 
422*dcbf3bd6SGeorge Wilson static inline boolean_t
423*dcbf3bd6SGeorge Wilson dbuf_cache_above_hiwater(void)
424*dcbf3bd6SGeorge Wilson {
425*dcbf3bd6SGeorge Wilson 	uint64_t dbuf_cache_hiwater_bytes =
426*dcbf3bd6SGeorge Wilson 	    (dbuf_cache_max_bytes * dbuf_cache_hiwater_pct) / 100;
427*dcbf3bd6SGeorge Wilson 
428*dcbf3bd6SGeorge Wilson 	return (refcount_count(&dbuf_cache_size) >
429*dcbf3bd6SGeorge Wilson 	    dbuf_cache_max_bytes + dbuf_cache_hiwater_bytes);
430*dcbf3bd6SGeorge Wilson }
431*dcbf3bd6SGeorge Wilson 
432*dcbf3bd6SGeorge Wilson static inline boolean_t
433*dcbf3bd6SGeorge Wilson dbuf_cache_above_lowater(void)
434*dcbf3bd6SGeorge Wilson {
435*dcbf3bd6SGeorge Wilson 	uint64_t dbuf_cache_lowater_bytes =
436*dcbf3bd6SGeorge Wilson 	    (dbuf_cache_max_bytes * dbuf_cache_lowater_pct) / 100;
437*dcbf3bd6SGeorge Wilson 
438*dcbf3bd6SGeorge Wilson 	return (refcount_count(&dbuf_cache_size) >
439*dcbf3bd6SGeorge Wilson 	    dbuf_cache_max_bytes - dbuf_cache_lowater_bytes);
440*dcbf3bd6SGeorge Wilson }
441*dcbf3bd6SGeorge Wilson 
442*dcbf3bd6SGeorge Wilson /*
443*dcbf3bd6SGeorge Wilson  * Evict the oldest eligible dbuf from the dbuf cache.
444*dcbf3bd6SGeorge Wilson  */
445*dcbf3bd6SGeorge Wilson static void
446*dcbf3bd6SGeorge Wilson dbuf_evict_one(void)
447*dcbf3bd6SGeorge Wilson {
448*dcbf3bd6SGeorge Wilson 	int idx = multilist_get_random_index(&dbuf_cache);
449*dcbf3bd6SGeorge Wilson 	multilist_sublist_t *mls = multilist_sublist_lock(&dbuf_cache, idx);
450*dcbf3bd6SGeorge Wilson 
451*dcbf3bd6SGeorge Wilson 	ASSERT(!MUTEX_HELD(&dbuf_evict_lock));
452*dcbf3bd6SGeorge Wilson 
453*dcbf3bd6SGeorge Wilson 	/*
454*dcbf3bd6SGeorge Wilson 	 * Set the thread's tsd to indicate that it's processing evictions.
455*dcbf3bd6SGeorge Wilson 	 * Once a thread stops evicting from the dbuf cache it will
456*dcbf3bd6SGeorge Wilson 	 * reset its tsd to NULL.
457*dcbf3bd6SGeorge Wilson 	 */
458*dcbf3bd6SGeorge Wilson 	ASSERT3P(tsd_get(zfs_dbuf_evict_key), ==, NULL);
459*dcbf3bd6SGeorge Wilson 	(void) tsd_set(zfs_dbuf_evict_key, (void *)B_TRUE);
460*dcbf3bd6SGeorge Wilson 
461*dcbf3bd6SGeorge Wilson 	dmu_buf_impl_t *db = multilist_sublist_tail(mls);
462*dcbf3bd6SGeorge Wilson 	while (db != NULL && mutex_tryenter(&db->db_mtx) == 0) {
463*dcbf3bd6SGeorge Wilson 		db = multilist_sublist_prev(mls, db);
464*dcbf3bd6SGeorge Wilson 	}
465*dcbf3bd6SGeorge Wilson 
466*dcbf3bd6SGeorge Wilson 	DTRACE_PROBE2(dbuf__evict__one, dmu_buf_impl_t *, db,
467*dcbf3bd6SGeorge Wilson 	    multilist_sublist_t *, mls);
468*dcbf3bd6SGeorge Wilson 
469*dcbf3bd6SGeorge Wilson 	if (db != NULL) {
470*dcbf3bd6SGeorge Wilson 		multilist_sublist_remove(mls, db);
471*dcbf3bd6SGeorge Wilson 		multilist_sublist_unlock(mls);
472*dcbf3bd6SGeorge Wilson 		(void) refcount_remove_many(&dbuf_cache_size,
473*dcbf3bd6SGeorge Wilson 		    db->db.db_size, db);
474*dcbf3bd6SGeorge Wilson 		dbuf_destroy(db);
475*dcbf3bd6SGeorge Wilson 	} else {
476*dcbf3bd6SGeorge Wilson 		multilist_sublist_unlock(mls);
477*dcbf3bd6SGeorge Wilson 	}
478*dcbf3bd6SGeorge Wilson 	(void) tsd_set(zfs_dbuf_evict_key, NULL);
479*dcbf3bd6SGeorge Wilson }
480*dcbf3bd6SGeorge Wilson 
481*dcbf3bd6SGeorge Wilson /*
482*dcbf3bd6SGeorge Wilson  * The dbuf evict thread is responsible for aging out dbufs from the
483*dcbf3bd6SGeorge Wilson  * cache. Once the cache has reached it's maximum size, dbufs are removed
484*dcbf3bd6SGeorge Wilson  * and destroyed. The eviction thread will continue running until the size
485*dcbf3bd6SGeorge Wilson  * of the dbuf cache is at or below the maximum size. Once the dbuf is aged
486*dcbf3bd6SGeorge Wilson  * out of the cache it is destroyed and becomes eligible for arc eviction.
487*dcbf3bd6SGeorge Wilson  */
488*dcbf3bd6SGeorge Wilson static void
489*dcbf3bd6SGeorge Wilson dbuf_evict_thread(void)
490*dcbf3bd6SGeorge Wilson {
491*dcbf3bd6SGeorge Wilson 	callb_cpr_t cpr;
492*dcbf3bd6SGeorge Wilson 
493*dcbf3bd6SGeorge Wilson 	CALLB_CPR_INIT(&cpr, &dbuf_evict_lock, callb_generic_cpr, FTAG);
494*dcbf3bd6SGeorge Wilson 
495*dcbf3bd6SGeorge Wilson 	mutex_enter(&dbuf_evict_lock);
496*dcbf3bd6SGeorge Wilson 	while (!dbuf_evict_thread_exit) {
497*dcbf3bd6SGeorge Wilson 		while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
498*dcbf3bd6SGeorge Wilson 			CALLB_CPR_SAFE_BEGIN(&cpr);
499*dcbf3bd6SGeorge Wilson 			(void) cv_timedwait_hires(&dbuf_evict_cv,
500*dcbf3bd6SGeorge Wilson 			    &dbuf_evict_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
501*dcbf3bd6SGeorge Wilson 			CALLB_CPR_SAFE_END(&cpr, &dbuf_evict_lock);
502*dcbf3bd6SGeorge Wilson 		}
503*dcbf3bd6SGeorge Wilson 		mutex_exit(&dbuf_evict_lock);
504*dcbf3bd6SGeorge Wilson 
505*dcbf3bd6SGeorge Wilson 		/*
506*dcbf3bd6SGeorge Wilson 		 * Keep evicting as long as we're above the low water mark
507*dcbf3bd6SGeorge Wilson 		 * for the cache. We do this without holding the locks to
508*dcbf3bd6SGeorge Wilson 		 * minimize lock contention.
509*dcbf3bd6SGeorge Wilson 		 */
510*dcbf3bd6SGeorge Wilson 		while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
511*dcbf3bd6SGeorge Wilson 			dbuf_evict_one();
512*dcbf3bd6SGeorge Wilson 		}
513*dcbf3bd6SGeorge Wilson 
514*dcbf3bd6SGeorge Wilson 		mutex_enter(&dbuf_evict_lock);
515*dcbf3bd6SGeorge Wilson 	}
516ea8dc4b6Seschrock 
517*dcbf3bd6SGeorge Wilson 	dbuf_evict_thread_exit = B_FALSE;
518*dcbf3bd6SGeorge Wilson 	cv_broadcast(&dbuf_evict_cv);
519*dcbf3bd6SGeorge Wilson 	CALLB_CPR_EXIT(&cpr);	/* drops dbuf_evict_lock */
520*dcbf3bd6SGeorge Wilson 	thread_exit();
521*dcbf3bd6SGeorge Wilson }
522*dcbf3bd6SGeorge Wilson 
523*dcbf3bd6SGeorge Wilson /*
524*dcbf3bd6SGeorge Wilson  * Wake up the dbuf eviction thread if the dbuf cache is at its max size.
525*dcbf3bd6SGeorge Wilson  * If the dbuf cache is at its high water mark, then evict a dbuf from the
526*dcbf3bd6SGeorge Wilson  * dbuf cache using the callers context.
527*dcbf3bd6SGeorge Wilson  */
528*dcbf3bd6SGeorge Wilson static void
529*dcbf3bd6SGeorge Wilson dbuf_evict_notify(void)
530*dcbf3bd6SGeorge Wilson {
531*dcbf3bd6SGeorge Wilson 
532*dcbf3bd6SGeorge Wilson 	/*
533*dcbf3bd6SGeorge Wilson 	 * We use thread specific data to track when a thread has
534*dcbf3bd6SGeorge Wilson 	 * started processing evictions. This allows us to avoid deeply
535*dcbf3bd6SGeorge Wilson 	 * nested stacks that would have a call flow similar to this:
536*dcbf3bd6SGeorge Wilson 	 *
537*dcbf3bd6SGeorge Wilson 	 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify()
538*dcbf3bd6SGeorge Wilson 	 *	^						|
539*dcbf3bd6SGeorge Wilson 	 *	|						|
540*dcbf3bd6SGeorge Wilson 	 *	+-----dbuf_destroy()<--dbuf_evict_one()<--------+
541*dcbf3bd6SGeorge Wilson 	 *
542*dcbf3bd6SGeorge Wilson 	 * The dbuf_eviction_thread will always have its tsd set until
543*dcbf3bd6SGeorge Wilson 	 * that thread exits. All other threads will only set their tsd
544*dcbf3bd6SGeorge Wilson 	 * if they are participating in the eviction process. This only
545*dcbf3bd6SGeorge Wilson 	 * happens if the eviction thread is unable to process evictions
546*dcbf3bd6SGeorge Wilson 	 * fast enough. To keep the dbuf cache size in check, other threads
547*dcbf3bd6SGeorge Wilson 	 * can evict from the dbuf cache directly. Those threads will set
548*dcbf3bd6SGeorge Wilson 	 * their tsd values so that we ensure that they only evict one dbuf
549*dcbf3bd6SGeorge Wilson 	 * from the dbuf cache.
550*dcbf3bd6SGeorge Wilson 	 */
551*dcbf3bd6SGeorge Wilson 	if (tsd_get(zfs_dbuf_evict_key) != NULL)
552*dcbf3bd6SGeorge Wilson 		return;
553*dcbf3bd6SGeorge Wilson 
554*dcbf3bd6SGeorge Wilson 	if (refcount_count(&dbuf_cache_size) > dbuf_cache_max_bytes) {
555*dcbf3bd6SGeorge Wilson 		boolean_t evict_now = B_FALSE;
556*dcbf3bd6SGeorge Wilson 
557*dcbf3bd6SGeorge Wilson 		mutex_enter(&dbuf_evict_lock);
558*dcbf3bd6SGeorge Wilson 		if (refcount_count(&dbuf_cache_size) > dbuf_cache_max_bytes) {
559*dcbf3bd6SGeorge Wilson 			evict_now = dbuf_cache_above_hiwater();
560*dcbf3bd6SGeorge Wilson 			cv_signal(&dbuf_evict_cv);
561*dcbf3bd6SGeorge Wilson 		}
562*dcbf3bd6SGeorge Wilson 		mutex_exit(&dbuf_evict_lock);
563*dcbf3bd6SGeorge Wilson 
564*dcbf3bd6SGeorge Wilson 		if (evict_now) {
565*dcbf3bd6SGeorge Wilson 			dbuf_evict_one();
566*dcbf3bd6SGeorge Wilson 		}
567*dcbf3bd6SGeorge Wilson 	}
568ea8dc4b6Seschrock }
569ea8dc4b6Seschrock 
570fa9e4066Sahrens void
571fa9e4066Sahrens dbuf_init(void)
572fa9e4066Sahrens {
573ea8dc4b6Seschrock 	uint64_t hsize = 1ULL << 16;
574fa9e4066Sahrens 	dbuf_hash_table_t *h = &dbuf_hash_table;
575fa9e4066Sahrens 	int i;
576fa9e4066Sahrens 
577fa9e4066Sahrens 	/*
578fa9e4066Sahrens 	 * The hash table is big enough to fill all of physical memory
579ea8dc4b6Seschrock 	 * with an average 4K block size.  The table will take up
580ea8dc4b6Seschrock 	 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
581fa9e4066Sahrens 	 */
582ea8dc4b6Seschrock 	while (hsize * 4096 < physmem * PAGESIZE)
583fa9e4066Sahrens 		hsize <<= 1;
584fa9e4066Sahrens 
585ea8dc4b6Seschrock retry:
586fa9e4066Sahrens 	h->hash_table_mask = hsize - 1;
587ea8dc4b6Seschrock 	h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
588ea8dc4b6Seschrock 	if (h->hash_table == NULL) {
589ea8dc4b6Seschrock 		/* XXX - we should really return an error instead of assert */
590ea8dc4b6Seschrock 		ASSERT(hsize > (1ULL << 10));
591ea8dc4b6Seschrock 		hsize >>= 1;
592ea8dc4b6Seschrock 		goto retry;
593ea8dc4b6Seschrock 	}
594fa9e4066Sahrens 
595*dcbf3bd6SGeorge Wilson 	dbuf_kmem_cache = kmem_cache_create("dmu_buf_impl_t",
596fa9e4066Sahrens 	    sizeof (dmu_buf_impl_t),
597fa9e4066Sahrens 	    0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
598fa9e4066Sahrens 
599fa9e4066Sahrens 	for (i = 0; i < DBUF_MUTEXES; i++)
600fa9e4066Sahrens 		mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
601bc9014e6SJustin Gibbs 
602*dcbf3bd6SGeorge Wilson 	/*
603*dcbf3bd6SGeorge Wilson 	 * Setup the parameters for the dbuf cache. We cap the size of the
604*dcbf3bd6SGeorge Wilson 	 * dbuf cache to 1/32nd (default) of the size of the ARC.
605*dcbf3bd6SGeorge Wilson 	 */
606*dcbf3bd6SGeorge Wilson 	dbuf_cache_max_bytes = MIN(dbuf_cache_max_bytes,
607*dcbf3bd6SGeorge Wilson 	    arc_max_bytes() >> dbuf_cache_max_shift);
608*dcbf3bd6SGeorge Wilson 
609bc9014e6SJustin Gibbs 	/*
610bc9014e6SJustin Gibbs 	 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
611bc9014e6SJustin Gibbs 	 * configuration is not required.
612bc9014e6SJustin Gibbs 	 */
613bc9014e6SJustin Gibbs 	dbu_evict_taskq = taskq_create("dbu_evict", 1, minclsyspri, 0, 0, 0);
614*dcbf3bd6SGeorge Wilson 
615*dcbf3bd6SGeorge Wilson 	multilist_create(&dbuf_cache, sizeof (dmu_buf_impl_t),
616*dcbf3bd6SGeorge Wilson 	    offsetof(dmu_buf_impl_t, db_cache_link),
617*dcbf3bd6SGeorge Wilson 	    zfs_arc_num_sublists_per_state,
618*dcbf3bd6SGeorge Wilson 	    dbuf_cache_multilist_index_func);
619*dcbf3bd6SGeorge Wilson 	refcount_create(&dbuf_cache_size);
620*dcbf3bd6SGeorge Wilson 
621*dcbf3bd6SGeorge Wilson 	tsd_create(&zfs_dbuf_evict_key, NULL);
622*dcbf3bd6SGeorge Wilson 	dbuf_evict_thread_exit = B_FALSE;
623*dcbf3bd6SGeorge Wilson 	mutex_init(&dbuf_evict_lock, NULL, MUTEX_DEFAULT, NULL);
624*dcbf3bd6SGeorge Wilson 	cv_init(&dbuf_evict_cv, NULL, CV_DEFAULT, NULL);
625*dcbf3bd6SGeorge Wilson 	dbuf_cache_evict_thread = thread_create(NULL, 0, dbuf_evict_thread,
626*dcbf3bd6SGeorge Wilson 	    NULL, 0, &p0, TS_RUN, minclsyspri);
627fa9e4066Sahrens }
628fa9e4066Sahrens 
629fa9e4066Sahrens void
630fa9e4066Sahrens dbuf_fini(void)
631fa9e4066Sahrens {
632fa9e4066Sahrens 	dbuf_hash_table_t *h = &dbuf_hash_table;
633fa9e4066Sahrens 	int i;
634fa9e4066Sahrens 
635fa9e4066Sahrens 	for (i = 0; i < DBUF_MUTEXES; i++)
636fa9e4066Sahrens 		mutex_destroy(&h->hash_mutexes[i]);
637fa9e4066Sahrens 	kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
638*dcbf3bd6SGeorge Wilson 	kmem_cache_destroy(dbuf_kmem_cache);
639bc9014e6SJustin Gibbs 	taskq_destroy(dbu_evict_taskq);
640*dcbf3bd6SGeorge Wilson 
641*dcbf3bd6SGeorge Wilson 	mutex_enter(&dbuf_evict_lock);
642*dcbf3bd6SGeorge Wilson 	dbuf_evict_thread_exit = B_TRUE;
643*dcbf3bd6SGeorge Wilson 	while (dbuf_evict_thread_exit) {
644*dcbf3bd6SGeorge Wilson 		cv_signal(&dbuf_evict_cv);
645*dcbf3bd6SGeorge Wilson 		cv_wait(&dbuf_evict_cv, &dbuf_evict_lock);
646*dcbf3bd6SGeorge Wilson 	}
647*dcbf3bd6SGeorge Wilson 	mutex_exit(&dbuf_evict_lock);
648*dcbf3bd6SGeorge Wilson 	tsd_destroy(&zfs_dbuf_evict_key);
649*dcbf3bd6SGeorge Wilson 
650*dcbf3bd6SGeorge Wilson 	mutex_destroy(&dbuf_evict_lock);
651*dcbf3bd6SGeorge Wilson 	cv_destroy(&dbuf_evict_cv);
652*dcbf3bd6SGeorge Wilson 
653*dcbf3bd6SGeorge Wilson 	refcount_destroy(&dbuf_cache_size);
654*dcbf3bd6SGeorge Wilson 	multilist_destroy(&dbuf_cache);
655fa9e4066Sahrens }
656fa9e4066Sahrens 
657fa9e4066Sahrens /*
658fa9e4066Sahrens  * Other stuff.
659fa9e4066Sahrens  */
660fa9e4066Sahrens 
6619c9dc39aSek #ifdef ZFS_DEBUG
662fa9e4066Sahrens static void
663fa9e4066Sahrens dbuf_verify(dmu_buf_impl_t *db)
664fa9e4066Sahrens {
665744947dcSTom Erickson 	dnode_t *dn;
666b24ab676SJeff Bonwick 	dbuf_dirty_record_t *dr;
667fa9e4066Sahrens 
668fa9e4066Sahrens 	ASSERT(MUTEX_HELD(&db->db_mtx));
669fa9e4066Sahrens 
670fa9e4066Sahrens 	if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
671fa9e4066Sahrens 		return;
672fa9e4066Sahrens 
673fa9e4066Sahrens 	ASSERT(db->db_objset != NULL);
674744947dcSTom Erickson 	DB_DNODE_ENTER(db);
675744947dcSTom Erickson 	dn = DB_DNODE(db);
676fa9e4066Sahrens 	if (dn == NULL) {
677fa9e4066Sahrens 		ASSERT(db->db_parent == NULL);
678fa9e4066Sahrens 		ASSERT(db->db_blkptr == NULL);
679fa9e4066Sahrens 	} else {
680fa9e4066Sahrens 		ASSERT3U(db->db.db_object, ==, dn->dn_object);
681fa9e4066Sahrens 		ASSERT3P(db->db_objset, ==, dn->dn_objset);
682fa9e4066Sahrens 		ASSERT3U(db->db_level, <, dn->dn_nlevels);
683744947dcSTom Erickson 		ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
684744947dcSTom Erickson 		    db->db_blkid == DMU_SPILL_BLKID ||
6850f6d88adSAlex Reece 		    !avl_is_empty(&dn->dn_dbufs));
686fa9e4066Sahrens 	}
6870a586ceaSMark Shellenbaum 	if (db->db_blkid == DMU_BONUS_BLKID) {
688fa9e4066Sahrens 		ASSERT(dn != NULL);
6891934e92fSmaybee 		ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
6900a586ceaSMark Shellenbaum 		ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
6910a586ceaSMark Shellenbaum 	} else if (db->db_blkid == DMU_SPILL_BLKID) {
6920a586ceaSMark Shellenbaum 		ASSERT(dn != NULL);
6930a586ceaSMark Shellenbaum 		ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
694fb09f5aaSMadhav Suresh 		ASSERT0(db->db.db_offset);
695fa9e4066Sahrens 	} else {
696fa9e4066Sahrens 		ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
697fa9e4066Sahrens 	}
698fa9e4066Sahrens 
699b24ab676SJeff Bonwick 	for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
700b24ab676SJeff Bonwick 		ASSERT(dr->dr_dbuf == db);
701b24ab676SJeff Bonwick 
702b24ab676SJeff Bonwick 	for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
703b24ab676SJeff Bonwick 		ASSERT(dr->dr_dbuf == db);
704b24ab676SJeff Bonwick 
70588b7b0f2SMatthew Ahrens 	/*
70688b7b0f2SMatthew Ahrens 	 * We can't assert that db_size matches dn_datablksz because it
70788b7b0f2SMatthew Ahrens 	 * can be momentarily different when another thread is doing
70888b7b0f2SMatthew Ahrens 	 * dnode_set_blksz().
70988b7b0f2SMatthew Ahrens 	 */
71088b7b0f2SMatthew Ahrens 	if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
711b24ab676SJeff Bonwick 		dr = db->db_data_pending;
71288b7b0f2SMatthew Ahrens 		/*
71388b7b0f2SMatthew Ahrens 		 * It should only be modified in syncing context, so
71488b7b0f2SMatthew Ahrens 		 * make sure we only have one copy of the data.
71588b7b0f2SMatthew Ahrens 		 */
71688b7b0f2SMatthew Ahrens 		ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
717fa9e4066Sahrens 	}
718fa9e4066Sahrens 
719fa9e4066Sahrens 	/* verify db->db_blkptr */
720fa9e4066Sahrens 	if (db->db_blkptr) {
721fa9e4066Sahrens 		if (db->db_parent == dn->dn_dbuf) {
722fa9e4066Sahrens 			/* db is pointed to by the dnode */
723fa9e4066Sahrens 			/* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
72414843421SMatthew Ahrens 			if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
725fa9e4066Sahrens 				ASSERT(db->db_parent == NULL);
726fa9e4066Sahrens 			else
727fa9e4066Sahrens 				ASSERT(db->db_parent != NULL);
7280a586ceaSMark Shellenbaum 			if (db->db_blkid != DMU_SPILL_BLKID)
7290a586ceaSMark Shellenbaum 				ASSERT3P(db->db_blkptr, ==,
7300a586ceaSMark Shellenbaum 				    &dn->dn_phys->dn_blkptr[db->db_blkid]);
731fa9e4066Sahrens 		} else {
732fa9e4066Sahrens 			/* db is pointed to by an indirect block */
733fa9e4066Sahrens 			int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
734fa9e4066Sahrens 			ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
735fa9e4066Sahrens 			ASSERT3U(db->db_parent->db.db_object, ==,
736fa9e4066Sahrens 			    db->db.db_object);
737fa9e4066Sahrens 			/*
738fa9e4066Sahrens 			 * dnode_grow_indblksz() can make this fail if we don't
739fa9e4066Sahrens 			 * have the struct_rwlock.  XXX indblksz no longer
740fa9e4066Sahrens 			 * grows.  safe to do this now?
741fa9e4066Sahrens 			 */
742744947dcSTom Erickson 			if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
743fa9e4066Sahrens 				ASSERT3P(db->db_blkptr, ==,
744fa9e4066Sahrens 				    ((blkptr_t *)db->db_parent->db.db_data +
745fa9e4066Sahrens 				    db->db_blkid % epb));
746fa9e4066Sahrens 			}
747fa9e4066Sahrens 		}
748fa9e4066Sahrens 	}
749fa9e4066Sahrens 	if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
7503f9d6ad7SLin Ling 	    (db->db_buf == NULL || db->db_buf->b_data) &&
7510a586ceaSMark Shellenbaum 	    db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
752fa9e4066Sahrens 	    db->db_state != DB_FILL && !dn->dn_free_txg) {
753fa9e4066Sahrens 		/*
754fa9e4066Sahrens 		 * If the blkptr isn't set but they have nonzero data,
755fa9e4066Sahrens 		 * it had better be dirty, otherwise we'll lose that
756fa9e4066Sahrens 		 * data when we evict this buffer.
7578df0bcf0SPaul Dagnelie 		 *
7588df0bcf0SPaul Dagnelie 		 * There is an exception to this rule for indirect blocks; in
7598df0bcf0SPaul Dagnelie 		 * this case, if the indirect block is a hole, we fill in a few
7608df0bcf0SPaul Dagnelie 		 * fields on each of the child blocks (importantly, birth time)
7618df0bcf0SPaul Dagnelie 		 * to prevent hole birth times from being lost when you
7628df0bcf0SPaul Dagnelie 		 * partially fill in a hole.
763fa9e4066Sahrens 		 */
764fa9e4066Sahrens 		if (db->db_dirtycnt == 0) {
7658df0bcf0SPaul Dagnelie 			if (db->db_level == 0) {
7668df0bcf0SPaul Dagnelie 				uint64_t *buf = db->db.db_data;
7678df0bcf0SPaul Dagnelie 				int i;
768fa9e4066Sahrens 
7698df0bcf0SPaul Dagnelie 				for (i = 0; i < db->db.db_size >> 3; i++) {
7708df0bcf0SPaul Dagnelie 					ASSERT(buf[i] == 0);
7718df0bcf0SPaul Dagnelie 				}
7728df0bcf0SPaul Dagnelie 			} else {
7738df0bcf0SPaul Dagnelie 				blkptr_t *bps = db->db.db_data;
7748df0bcf0SPaul Dagnelie 				ASSERT3U(1 << DB_DNODE(db)->dn_indblkshift, ==,
7758df0bcf0SPaul Dagnelie 				    db->db.db_size);
7768df0bcf0SPaul Dagnelie 				/*
7778df0bcf0SPaul Dagnelie 				 * We want to verify that all the blkptrs in the
7788df0bcf0SPaul Dagnelie 				 * indirect block are holes, but we may have
7798df0bcf0SPaul Dagnelie 				 * automatically set up a few fields for them.
7808df0bcf0SPaul Dagnelie 				 * We iterate through each blkptr and verify
7818df0bcf0SPaul Dagnelie 				 * they only have those fields set.
7828df0bcf0SPaul Dagnelie 				 */
7838df0bcf0SPaul Dagnelie 				for (int i = 0;
7848df0bcf0SPaul Dagnelie 				    i < db->db.db_size / sizeof (blkptr_t);
7858df0bcf0SPaul Dagnelie 				    i++) {
7868df0bcf0SPaul Dagnelie 					blkptr_t *bp = &bps[i];
7878df0bcf0SPaul Dagnelie 					ASSERT(ZIO_CHECKSUM_IS_ZERO(
7888df0bcf0SPaul Dagnelie 					    &bp->blk_cksum));
7898df0bcf0SPaul Dagnelie 					ASSERT(
7908df0bcf0SPaul Dagnelie 					    DVA_IS_EMPTY(&bp->blk_dva[0]) &&
7918df0bcf0SPaul Dagnelie 					    DVA_IS_EMPTY(&bp->blk_dva[1]) &&
7928df0bcf0SPaul Dagnelie 					    DVA_IS_EMPTY(&bp->blk_dva[2]));
7938df0bcf0SPaul Dagnelie 					ASSERT0(bp->blk_fill);
7948df0bcf0SPaul Dagnelie 					ASSERT0(bp->blk_pad[0]);
7958df0bcf0SPaul Dagnelie 					ASSERT0(bp->blk_pad[1]);
7968df0bcf0SPaul Dagnelie 					ASSERT(!BP_IS_EMBEDDED(bp));
7978df0bcf0SPaul Dagnelie 					ASSERT(BP_IS_HOLE(bp));
7988df0bcf0SPaul Dagnelie 					ASSERT0(bp->blk_phys_birth);
7998df0bcf0SPaul Dagnelie 				}
800fa9e4066Sahrens 			}
801fa9e4066Sahrens 		}
802fa9e4066Sahrens 	}
803744947dcSTom Erickson 	DB_DNODE_EXIT(db);
804fa9e4066Sahrens }
8059c9dc39aSek #endif
806fa9e4066Sahrens 
807bc9014e6SJustin Gibbs static void
808bc9014e6SJustin Gibbs dbuf_clear_data(dmu_buf_impl_t *db)
809bc9014e6SJustin Gibbs {
810bc9014e6SJustin Gibbs 	ASSERT(MUTEX_HELD(&db->db_mtx));
811bc9014e6SJustin Gibbs 	dbuf_evict_user(db);
812*dcbf3bd6SGeorge Wilson 	ASSERT3P(db->db_buf, ==, NULL);
813bc9014e6SJustin Gibbs 	db->db.db_data = NULL;
814bc9014e6SJustin Gibbs 	if (db->db_state != DB_NOFILL)
815bc9014e6SJustin Gibbs 		db->db_state = DB_UNCACHED;
816bc9014e6SJustin Gibbs }
817bc9014e6SJustin Gibbs 
818fa9e4066Sahrens static void
819fa9e4066Sahrens dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
820fa9e4066Sahrens {
821fa9e4066Sahrens 	ASSERT(MUTEX_HELD(&db->db_mtx));
822bc9014e6SJustin Gibbs 	ASSERT(buf != NULL);
823bc9014e6SJustin Gibbs 
824fa9e4066Sahrens 	db->db_buf = buf;
825bc9014e6SJustin Gibbs 	ASSERT(buf->b_data != NULL);
826bc9014e6SJustin Gibbs 	db->db.db_data = buf->b_data;
827fa9e4066Sahrens }
828fa9e4066Sahrens 
829c242f9a0Schunli zhang - Sun Microsystems - Irvine United States /*
830c242f9a0Schunli zhang - Sun Microsystems - Irvine United States  * Loan out an arc_buf for read.  Return the loaned arc_buf.
831c242f9a0Schunli zhang - Sun Microsystems - Irvine United States  */
832c242f9a0Schunli zhang - Sun Microsystems - Irvine United States arc_buf_t *
833c242f9a0Schunli zhang - Sun Microsystems - Irvine United States dbuf_loan_arcbuf(dmu_buf_impl_t *db)
834c242f9a0Schunli zhang - Sun Microsystems - Irvine United States {
835c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 	arc_buf_t *abuf;
836c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 
837*dcbf3bd6SGeorge Wilson 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
838c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 	mutex_enter(&db->db_mtx);
839c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 	if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
840c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 		int blksz = db->db.db_size;
84143466aaeSMax Grossman 		spa_t *spa = db->db_objset->os_spa;
842744947dcSTom Erickson 
843c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 		mutex_exit(&db->db_mtx);
844744947dcSTom Erickson 		abuf = arc_loan_buf(spa, blksz);
845c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 		bcopy(db->db.db_data, abuf->b_data, blksz);
846c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 	} else {
847c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 		abuf = db->db_buf;
848c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 		arc_loan_inuse_buf(abuf, db);
849*dcbf3bd6SGeorge Wilson 		db->db_buf = NULL;
850bc9014e6SJustin Gibbs 		dbuf_clear_data(db);
851c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 		mutex_exit(&db->db_mtx);
852c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 	}
853c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 	return (abuf);
854c242f9a0Schunli zhang - Sun Microsystems - Irvine United States }
855c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 
856a2cdcdd2SPaul Dagnelie /*
857a2cdcdd2SPaul Dagnelie  * Calculate which level n block references the data at the level 0 offset
858a2cdcdd2SPaul Dagnelie  * provided.
859a2cdcdd2SPaul Dagnelie  */
860fa9e4066Sahrens uint64_t
861a2cdcdd2SPaul Dagnelie dbuf_whichblock(dnode_t *dn, int64_t level, uint64_t offset)
862fa9e4066Sahrens {
863a2cdcdd2SPaul Dagnelie 	if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) {
864a2cdcdd2SPaul Dagnelie 		/*
865a2cdcdd2SPaul Dagnelie 		 * The level n blkid is equal to the level 0 blkid divided by
866a2cdcdd2SPaul Dagnelie 		 * the number of level 0s in a level n block.
867a2cdcdd2SPaul Dagnelie 		 *
868a2cdcdd2SPaul Dagnelie 		 * The level 0 blkid is offset >> datablkshift =
869a2cdcdd2SPaul Dagnelie 		 * offset / 2^datablkshift.
870a2cdcdd2SPaul Dagnelie 		 *
871a2cdcdd2SPaul Dagnelie 		 * The number of level 0s in a level n is the number of block
872a2cdcdd2SPaul Dagnelie 		 * pointers in an indirect block, raised to the power of level.
873a2cdcdd2SPaul Dagnelie 		 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
874a2cdcdd2SPaul Dagnelie 		 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
875a2cdcdd2SPaul Dagnelie 		 *
876a2cdcdd2SPaul Dagnelie 		 * Thus, the level n blkid is: offset /
877a2cdcdd2SPaul Dagnelie 		 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
878a2cdcdd2SPaul Dagnelie 		 * = offset / 2^(datablkshift + level *
879a2cdcdd2SPaul Dagnelie 		 *   (indblkshift - SPA_BLKPTRSHIFT))
880a2cdcdd2SPaul Dagnelie 		 * = offset >> (datablkshift + level *
881a2cdcdd2SPaul Dagnelie 		 *   (indblkshift - SPA_BLKPTRSHIFT))
882a2cdcdd2SPaul Dagnelie 		 */
883a2cdcdd2SPaul Dagnelie 		return (offset >> (dn->dn_datablkshift + level *
884a2cdcdd2SPaul Dagnelie 		    (dn->dn_indblkshift - SPA_BLKPTRSHIFT)));
885fa9e4066Sahrens 	} else {
886fa9e4066Sahrens 		ASSERT3U(offset, <, dn->dn_datablksz);
887fa9e4066Sahrens 		return (0);
888fa9e4066Sahrens 	}
889fa9e4066Sahrens }
890fa9e4066Sahrens 
891fa9e4066Sahrens static void
892fa9e4066Sahrens dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
893fa9e4066Sahrens {
894fa9e4066Sahrens 	dmu_buf_impl_t *db = vdb;
895fa9e4066Sahrens 
896fa9e4066Sahrens 	mutex_enter(&db->db_mtx);
897fa9e4066Sahrens 	ASSERT3U(db->db_state, ==, DB_READ);
898fa9e4066Sahrens 	/*
899fa9e4066Sahrens 	 * All reads are synchronous, so we must have a hold on the dbuf
900fa9e4066Sahrens 	 */
901fa9e4066Sahrens 	ASSERT(refcount_count(&db->db_holds) > 0);
902ea8dc4b6Seschrock 	ASSERT(db->db_buf == NULL);
903fa9e4066Sahrens 	ASSERT(db->db.db_data == NULL);
904c717a561Smaybee 	if (db->db_level == 0 && db->db_freed_in_flight) {
905fa9e4066Sahrens 		/* we were freed in flight; disregard any error */
906fa9e4066Sahrens 		arc_release(buf, db);
907fa9e4066Sahrens 		bzero(buf->b_data, db->db.db_size);
9086b4acc8bSahrens 		arc_buf_freeze(buf);
909c717a561Smaybee 		db->db_freed_in_flight = FALSE;
910fa9e4066Sahrens 		dbuf_set_data(db, buf);
911fa9e4066Sahrens 		db->db_state = DB_CACHED;
912fa9e4066Sahrens 	} else if (zio == NULL || zio->io_error == 0) {
913fa9e4066Sahrens 		dbuf_set_data(db, buf);
914fa9e4066Sahrens 		db->db_state = DB_CACHED;
915fa9e4066Sahrens 	} else {
9160a586ceaSMark Shellenbaum 		ASSERT(db->db_blkid != DMU_BONUS_BLKID);
917fa9e4066Sahrens 		ASSERT3P(db->db_buf, ==, NULL);
918*dcbf3bd6SGeorge Wilson 		arc_buf_destroy(buf, db);
919ea8dc4b6Seschrock 		db->db_state = DB_UNCACHED;
920fa9e4066Sahrens 	}
921fa9e4066Sahrens 	cv_broadcast(&db->db_changed);
9223f9d6ad7SLin Ling 	dbuf_rele_and_unlock(db, NULL);
923fa9e4066Sahrens }
924fa9e4066Sahrens 
925ea8dc4b6Seschrock static void
926cf6106c8SMatthew Ahrens dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
927fa9e4066Sahrens {
928744947dcSTom Erickson 	dnode_t *dn;
9297802d7bfSMatthew Ahrens 	zbookmark_phys_t zb;
9307adb730bSGeorge Wilson 	arc_flags_t aflags = ARC_FLAG_NOWAIT;
931fa9e4066Sahrens 
932744947dcSTom Erickson 	DB_DNODE_ENTER(db);
933744947dcSTom Erickson 	dn = DB_DNODE(db);
934fa9e4066Sahrens 	ASSERT(!refcount_is_zero(&db->db_holds));
935fa9e4066Sahrens 	/* We need the struct_rwlock to prevent db_blkptr from changing. */
936088f3894Sahrens 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
937ea8dc4b6Seschrock 	ASSERT(MUTEX_HELD(&db->db_mtx));
938ea8dc4b6Seschrock 	ASSERT(db->db_state == DB_UNCACHED);
939ea8dc4b6Seschrock 	ASSERT(db->db_buf == NULL);
940fa9e4066Sahrens 
9410a586ceaSMark Shellenbaum 	if (db->db_blkid == DMU_BONUS_BLKID) {
942cf04dda1SMark Maybee 		int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
9431934e92fSmaybee 
9441934e92fSmaybee 		ASSERT3U(bonuslen, <=, db->db.db_size);
945ea8dc4b6Seschrock 		db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN);
9465a98e54bSBrendan Gregg - Sun Microsystems 		arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
9471934e92fSmaybee 		if (bonuslen < DN_MAX_BONUSLEN)
948ea8dc4b6Seschrock 			bzero(db->db.db_data, DN_MAX_BONUSLEN);
949cf04dda1SMark Maybee 		if (bonuslen)
950cf04dda1SMark Maybee 			bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
951744947dcSTom Erickson 		DB_DNODE_EXIT(db);
952fa9e4066Sahrens 		db->db_state = DB_CACHED;
953fa9e4066Sahrens 		mutex_exit(&db->db_mtx);
954fa9e4066Sahrens 		return;
955fa9e4066Sahrens 	}
956fa9e4066Sahrens 
9571c8564a7SMark Maybee 	/*
9581c8564a7SMark Maybee 	 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
9591c8564a7SMark Maybee 	 * processes the delete record and clears the bp while we are waiting
9601c8564a7SMark Maybee 	 * for the dn_mtx (resulting in a "no" from block_freed).
9611c8564a7SMark Maybee 	 */
962088f3894Sahrens 	if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
9631c8564a7SMark Maybee 	    (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
9641c8564a7SMark Maybee 	    BP_IS_HOLE(db->db_blkptr)))) {
965ad23a2dbSjohansen 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
966ad23a2dbSjohansen 
967*dcbf3bd6SGeorge Wilson 		dbuf_set_data(db, arc_alloc_buf(db->db_objset->os_spa,
96843466aaeSMax Grossman 		    db->db.db_size, db, type));
969fa9e4066Sahrens 		bzero(db->db.db_data, db->db.db_size);
9708df0bcf0SPaul Dagnelie 
9718df0bcf0SPaul Dagnelie 		if (db->db_blkptr != NULL && db->db_level > 0 &&
9728df0bcf0SPaul Dagnelie 		    BP_IS_HOLE(db->db_blkptr) &&
9738df0bcf0SPaul Dagnelie 		    db->db_blkptr->blk_birth != 0) {
9748df0bcf0SPaul Dagnelie 			blkptr_t *bps = db->db.db_data;
9758df0bcf0SPaul Dagnelie 			for (int i = 0; i < ((1 <<
9768df0bcf0SPaul Dagnelie 			    DB_DNODE(db)->dn_indblkshift) / sizeof (blkptr_t));
9778df0bcf0SPaul Dagnelie 			    i++) {
9788df0bcf0SPaul Dagnelie 				blkptr_t *bp = &bps[i];
9798df0bcf0SPaul Dagnelie 				ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
9808df0bcf0SPaul Dagnelie 				    1 << dn->dn_indblkshift);
9818df0bcf0SPaul Dagnelie 				BP_SET_LSIZE(bp,
9828df0bcf0SPaul Dagnelie 				    BP_GET_LEVEL(db->db_blkptr) == 1 ?
9838df0bcf0SPaul Dagnelie 				    dn->dn_datablksz :
9848df0bcf0SPaul Dagnelie 				    BP_GET_LSIZE(db->db_blkptr));
9858df0bcf0SPaul Dagnelie 				BP_SET_TYPE(bp, BP_GET_TYPE(db->db_blkptr));
9868df0bcf0SPaul Dagnelie 				BP_SET_LEVEL(bp,
9878df0bcf0SPaul Dagnelie 				    BP_GET_LEVEL(db->db_blkptr) - 1);
9888df0bcf0SPaul Dagnelie 				BP_SET_BIRTH(bp, db->db_blkptr->blk_birth, 0);
9898df0bcf0SPaul Dagnelie 			}
9908df0bcf0SPaul Dagnelie 		}
9918df0bcf0SPaul Dagnelie 		DB_DNODE_EXIT(db);
992fa9e4066Sahrens 		db->db_state = DB_CACHED;
993fa9e4066Sahrens 		mutex_exit(&db->db_mtx);
994fa9e4066Sahrens 		return;
995fa9e4066Sahrens 	}
996fa9e4066Sahrens 
997744947dcSTom Erickson 	DB_DNODE_EXIT(db);
998744947dcSTom Erickson 
999fa9e4066Sahrens 	db->db_state = DB_READ;
1000fa9e4066Sahrens 	mutex_exit(&db->db_mtx);
1001fa9e4066Sahrens 
10023baa08fcSek 	if (DBUF_IS_L2CACHEABLE(db))
10037adb730bSGeorge Wilson 		aflags |= ARC_FLAG_L2CACHE;
10043baa08fcSek 
1005b24ab676SJeff Bonwick 	SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
1006b24ab676SJeff Bonwick 	    db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
1007b24ab676SJeff Bonwick 	    db->db.db_object, db->db_level, db->db_blkid);
1008ea8dc4b6Seschrock 
1009ea8dc4b6Seschrock 	dbuf_add_ref(db, NULL);
1010088f3894Sahrens 
101143466aaeSMax Grossman 	(void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr,
1012fa9e4066Sahrens 	    dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
1013cf6106c8SMatthew Ahrens 	    (flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
101413506d1eSmaybee 	    &aflags, &zb);
1015fa9e4066Sahrens }
1016fa9e4066Sahrens 
1017ea8dc4b6Seschrock int
1018ea8dc4b6Seschrock dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
1019fa9e4066Sahrens {
1020ea8dc4b6Seschrock 	int err = 0;
102143466aaeSMax Grossman 	boolean_t havepzio = (zio != NULL);
102243466aaeSMax Grossman 	boolean_t prefetch;
1023744947dcSTom Erickson 	dnode_t *dn;
1024fa9e4066Sahrens 
1025fa9e4066Sahrens 	/*
1026fa9e4066Sahrens 	 * We don't have to hold the mutex to check db_state because it
1027fa9e4066Sahrens 	 * can't be freed while we have a hold on the buffer.
1028fa9e4066Sahrens 	 */
1029fa9e4066Sahrens 	ASSERT(!refcount_is_zero(&db->db_holds));
1030fa9e4066Sahrens 
103182c9918fSTim Haley 	if (db->db_state == DB_NOFILL)
1032be6fd75aSMatthew Ahrens 		return (SET_ERROR(EIO));
103382c9918fSTim Haley 
1034744947dcSTom Erickson 	DB_DNODE_ENTER(db);
1035744947dcSTom Erickson 	dn = DB_DNODE(db);
1036ea8dc4b6Seschrock 	if ((flags & DB_RF_HAVESTRUCT) == 0)
1037744947dcSTom Erickson 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
1038ea8dc4b6Seschrock 
10390a586ceaSMark Shellenbaum 	prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1040744947dcSTom Erickson 	    (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
10413baa08fcSek 	    DBUF_IS_CACHEABLE(db);
104213506d1eSmaybee 
1043ea8dc4b6Seschrock 	mutex_enter(&db->db_mtx);
1044ea8dc4b6Seschrock 	if (db->db_state == DB_CACHED) {
1045ea8dc4b6Seschrock 		mutex_exit(&db->db_mtx);
104613506d1eSmaybee 		if (prefetch)
1047cb92f413SAlexander Motin 			dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1048fa9e4066Sahrens 		if ((flags & DB_RF_HAVESTRUCT) == 0)
1049744947dcSTom Erickson 			rw_exit(&dn->dn_struct_rwlock);
1050744947dcSTom Erickson 		DB_DNODE_EXIT(db);
1051ea8dc4b6Seschrock 	} else if (db->db_state == DB_UNCACHED) {
1052744947dcSTom Erickson 		spa_t *spa = dn->dn_objset->os_spa;
1053744947dcSTom Erickson 
1054744947dcSTom Erickson 		if (zio == NULL)
1055744947dcSTom Erickson 			zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
1056cf6106c8SMatthew Ahrens 		dbuf_read_impl(db, zio, flags);
105713506d1eSmaybee 
1058ea8dc4b6Seschrock 		/* dbuf_read_impl has dropped db_mtx for us */
1059ea8dc4b6Seschrock 
106013506d1eSmaybee 		if (prefetch)
1061cb92f413SAlexander Motin 			dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1062ea8dc4b6Seschrock 
1063fa9e4066Sahrens 		if ((flags & DB_RF_HAVESTRUCT) == 0)
1064744947dcSTom Erickson 			rw_exit(&dn->dn_struct_rwlock);
1065744947dcSTom Erickson 		DB_DNODE_EXIT(db);
1066fa9e4066Sahrens 
1067ea8dc4b6Seschrock 		if (!havepzio)
1068ea8dc4b6Seschrock 			err = zio_wait(zio);
1069ea8dc4b6Seschrock 	} else {
10703e30c24aSWill Andrews 		/*
10713e30c24aSWill Andrews 		 * Another reader came in while the dbuf was in flight
10723e30c24aSWill Andrews 		 * between UNCACHED and CACHED.  Either a writer will finish
10733e30c24aSWill Andrews 		 * writing the buffer (sending the dbuf to CACHED) or the
10743e30c24aSWill Andrews 		 * first reader's request will reach the read_done callback
10753e30c24aSWill Andrews 		 * and send the dbuf to CACHED.  Otherwise, a failure
10763e30c24aSWill Andrews 		 * occurred and the dbuf went to UNCACHED.
10773e30c24aSWill Andrews 		 */
107813506d1eSmaybee 		mutex_exit(&db->db_mtx);
107913506d1eSmaybee 		if (prefetch)
1080cb92f413SAlexander Motin 			dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1081ea8dc4b6Seschrock 		if ((flags & DB_RF_HAVESTRUCT) == 0)
1082744947dcSTom Erickson 			rw_exit(&dn->dn_struct_rwlock);
1083744947dcSTom Erickson 		DB_DNODE_EXIT(db);
108413506d1eSmaybee 
10853e30c24aSWill Andrews 		/* Skip the wait per the caller's request. */
108613506d1eSmaybee 		mutex_enter(&db->db_mtx);
1087ea8dc4b6Seschrock 		if ((flags & DB_RF_NEVERWAIT) == 0) {
1088ea8dc4b6Seschrock 			while (db->db_state == DB_READ ||
1089ea8dc4b6Seschrock 			    db->db_state == DB_FILL) {
1090ea8dc4b6Seschrock 				ASSERT(db->db_state == DB_READ ||
1091ea8dc4b6Seschrock 				    (flags & DB_RF_HAVESTRUCT) == 0);
1092f6164ad6SAdam H. Leventhal 				DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *,
1093f6164ad6SAdam H. Leventhal 				    db, zio_t *, zio);
1094ea8dc4b6Seschrock 				cv_wait(&db->db_changed, &db->db_mtx);
1095ea8dc4b6Seschrock 			}
1096ea8dc4b6Seschrock 			if (db->db_state == DB_UNCACHED)
1097be6fd75aSMatthew Ahrens 				err = SET_ERROR(EIO);
1098ea8dc4b6Seschrock 		}
1099ea8dc4b6Seschrock 		mutex_exit(&db->db_mtx);
1100fa9e4066Sahrens 	}
1101fa9e4066Sahrens 
1102ea8dc4b6Seschrock 	ASSERT(err || havepzio || db->db_state == DB_CACHED);
1103ea8dc4b6Seschrock 	return (err);
1104fa9e4066Sahrens }
1105fa9e4066Sahrens 
1106fa9e4066Sahrens static void
1107fa9e4066Sahrens dbuf_noread(dmu_buf_impl_t *db)
1108fa9e4066Sahrens {
1109fa9e4066Sahrens 	ASSERT(!refcount_is_zero(&db->db_holds));
11100a586ceaSMark Shellenbaum 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1111fa9e4066Sahrens 	mutex_enter(&db->db_mtx);
1112fa9e4066Sahrens 	while (db->db_state == DB_READ || db->db_state == DB_FILL)
1113fa9e4066Sahrens 		cv_wait(&db->db_changed, &db->db_mtx);
1114fa9e4066Sahrens 	if (db->db_state == DB_UNCACHED) {
1115ad23a2dbSjohansen 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
111643466aaeSMax Grossman 		spa_t *spa = db->db_objset->os_spa;
1117ad23a2dbSjohansen 
1118ea8dc4b6Seschrock 		ASSERT(db->db_buf == NULL);
1119fa9e4066Sahrens 		ASSERT(db->db.db_data == NULL);
1120*dcbf3bd6SGeorge Wilson 		dbuf_set_data(db, arc_alloc_buf(spa, db->db.db_size, db, type));
1121fa9e4066Sahrens 		db->db_state = DB_FILL;
112282c9918fSTim Haley 	} else if (db->db_state == DB_NOFILL) {
1123bc9014e6SJustin Gibbs 		dbuf_clear_data(db);
1124fa9e4066Sahrens 	} else {
1125fa9e4066Sahrens 		ASSERT3U(db->db_state, ==, DB_CACHED);
1126fa9e4066Sahrens 	}
1127fa9e4066Sahrens 	mutex_exit(&db->db_mtx);
1128fa9e4066Sahrens }
1129fa9e4066Sahrens 
1130fa9e4066Sahrens /*
1131fa9e4066Sahrens  * This is our just-in-time copy function.  It makes a copy of
1132fa9e4066Sahrens  * buffers, that have been modified in a previous transaction
1133fa9e4066Sahrens  * group, before we modify them in the current active group.
1134fa9e4066Sahrens  *
1135fa9e4066Sahrens  * This function is used in two places: when we are dirtying a
1136fa9e4066Sahrens  * buffer for the first time in a txg, and when we are freeing
1137fa9e4066Sahrens  * a range in a dnode that includes this buffer.
1138fa9e4066Sahrens  *
1139fa9e4066Sahrens  * Note that when we are called from dbuf_free_range() we do
1140fa9e4066Sahrens  * not put a hold on the buffer, we just traverse the active
1141fa9e4066Sahrens  * dbuf list for the dnode.
1142fa9e4066Sahrens  */
1143fa9e4066Sahrens static void
1144fa9e4066Sahrens dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
1145fa9e4066Sahrens {
1146c717a561Smaybee 	dbuf_dirty_record_t *dr = db->db_last_dirty;
1147fa9e4066Sahrens 
1148fa9e4066Sahrens 	ASSERT(MUTEX_HELD(&db->db_mtx));
1149fa9e4066Sahrens 	ASSERT(db->db.db_data != NULL);
1150c717a561Smaybee 	ASSERT(db->db_level == 0);
1151c717a561Smaybee 	ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
1152fa9e4066Sahrens 
11534d31c452Smaybee 	if (dr == NULL ||
11544d31c452Smaybee 	    (dr->dt.dl.dr_data !=
11550a586ceaSMark Shellenbaum 	    ((db->db_blkid  == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
1156fa9e4066Sahrens 		return;
1157fa9e4066Sahrens 
1158fa9e4066Sahrens 	/*
1159c717a561Smaybee 	 * If the last dirty record for this dbuf has not yet synced
1160c717a561Smaybee 	 * and its referencing the dbuf data, either:
1161744947dcSTom Erickson 	 *	reset the reference to point to a new copy,
1162c717a561Smaybee 	 * or (if there a no active holders)
1163c717a561Smaybee 	 *	just null out the current db_data pointer.
1164fa9e4066Sahrens 	 */
1165c717a561Smaybee 	ASSERT(dr->dr_txg >= txg - 2);
11660a586ceaSMark Shellenbaum 	if (db->db_blkid == DMU_BONUS_BLKID) {
1167c717a561Smaybee 		/* Note that the data bufs here are zio_bufs */
1168c717a561Smaybee 		dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
11695a98e54bSBrendan Gregg - Sun Microsystems 		arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
1170c717a561Smaybee 		bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
1171c717a561Smaybee 	} else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
1172c717a561Smaybee 		int size = db->db.db_size;
1173c717a561Smaybee 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
117443466aaeSMax Grossman 		spa_t *spa = db->db_objset->os_spa;
1175744947dcSTom Erickson 
1176*dcbf3bd6SGeorge Wilson 		dr->dt.dl.dr_data = arc_alloc_buf(spa, size, db, type);
1177c717a561Smaybee 		bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
1178c717a561Smaybee 	} else {
1179*dcbf3bd6SGeorge Wilson 		db->db_buf = NULL;
1180bc9014e6SJustin Gibbs 		dbuf_clear_data(db);
1181fa9e4066Sahrens 	}
1182fa9e4066Sahrens }
1183fa9e4066Sahrens 
1184c717a561Smaybee void
1185c717a561Smaybee dbuf_unoverride(dbuf_dirty_record_t *dr)
1186ea8dc4b6Seschrock {
1187c717a561Smaybee 	dmu_buf_impl_t *db = dr->dr_dbuf;
1188b24ab676SJeff Bonwick 	blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
1189c717a561Smaybee 	uint64_t txg = dr->dr_txg;
1190ea8dc4b6Seschrock 
1191ea8dc4b6Seschrock 	ASSERT(MUTEX_HELD(&db->db_mtx));
1192c717a561Smaybee 	ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
1193c717a561Smaybee 	ASSERT(db->db_level == 0);
1194ea8dc4b6Seschrock 
11950a586ceaSMark Shellenbaum 	if (db->db_blkid == DMU_BONUS_BLKID ||
1196c717a561Smaybee 	    dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
1197c717a561Smaybee 		return;
1198ea8dc4b6Seschrock 
1199b24ab676SJeff Bonwick 	ASSERT(db->db_data_pending != dr);
1200b24ab676SJeff Bonwick 
1201c717a561Smaybee 	/* free this block */
120243466aaeSMax Grossman 	if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite)
120343466aaeSMax Grossman 		zio_free(db->db_objset->os_spa, txg, bp);
1204b24ab676SJeff Bonwick 
1205c717a561Smaybee 	dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
120680901aeaSGeorge Wilson 	dr->dt.dl.dr_nopwrite = B_FALSE;
120780901aeaSGeorge Wilson 
1208c717a561Smaybee 	/*
1209c717a561Smaybee 	 * Release the already-written buffer, so we leave it in
1210c717a561Smaybee 	 * a consistent dirty state.  Note that all callers are
1211c717a561Smaybee 	 * modifying the buffer, so they will immediately do
1212c717a561Smaybee 	 * another (redundant) arc_release().  Therefore, leave
1213c717a561Smaybee 	 * the buf thawed to save the effort of freezing &
1214c717a561Smaybee 	 * immediately re-thawing it.
1215c717a561Smaybee 	 */
1216c717a561Smaybee 	arc_release(dr->dt.dl.dr_data, db);
1217fa9e4066Sahrens }
1218fa9e4066Sahrens 
1219cdb0ab79Smaybee /*
1220cdb0ab79Smaybee  * Evict (if its unreferenced) or clear (if its referenced) any level-0
1221cdb0ab79Smaybee  * data blocks in the free range, so that any future readers will find
122243466aaeSMax Grossman  * empty blocks.
12232f3d8780SMatthew Ahrens  *
12242f3d8780SMatthew Ahrens  * This is a no-op if the dataset is in the middle of an incremental
12252f3d8780SMatthew Ahrens  * receive; see comment below for details.
1226cdb0ab79Smaybee  */
1227fa9e4066Sahrens void
12280f6d88adSAlex Reece dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
12290f6d88adSAlex Reece     dmu_tx_t *tx)
1230fa9e4066Sahrens {
1231bc9014e6SJustin Gibbs 	dmu_buf_impl_t db_search;
1232bc9014e6SJustin Gibbs 	dmu_buf_impl_t *db, *db_next;
1233fa9e4066Sahrens 	uint64_t txg = tx->tx_txg;
12340f6d88adSAlex Reece 	avl_index_t where;
1235445e6780STim Chase 	boolean_t freespill =
1236445e6780STim Chase 	    (start_blkid == DMU_SPILL_BLKID || end_blkid == DMU_SPILL_BLKID);
1237fa9e4066Sahrens 
1238445e6780STim Chase 	if (end_blkid > dn->dn_maxblkid && !freespill)
12390f6d88adSAlex Reece 		end_blkid = dn->dn_maxblkid;
12400f6d88adSAlex Reece 	dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid);
12410f6d88adSAlex Reece 
12420f6d88adSAlex Reece 	db_search.db_level = 0;
12430f6d88adSAlex Reece 	db_search.db_blkid = start_blkid;
124486bb58aeSAlex Reece 	db_search.db_state = DB_SEARCH;
12452f3d8780SMatthew Ahrens 
1246713d6c20SMatthew Ahrens 	mutex_enter(&dn->dn_dbufs_mtx);
1247445e6780STim Chase 	if (start_blkid >= dn->dn_unlisted_l0_blkid && !freespill) {
1248713d6c20SMatthew Ahrens 		/* There can't be any dbufs in this range; no need to search. */
12490f6d88adSAlex Reece #ifdef DEBUG
12500f6d88adSAlex Reece 		db = avl_find(&dn->dn_dbufs, &db_search, &where);
12510f6d88adSAlex Reece 		ASSERT3P(db, ==, NULL);
12520f6d88adSAlex Reece 		db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
12530f6d88adSAlex Reece 		ASSERT(db == NULL || db->db_level > 0);
12540f6d88adSAlex Reece #endif
1255713d6c20SMatthew Ahrens 		mutex_exit(&dn->dn_dbufs_mtx);
1256713d6c20SMatthew Ahrens 		return;
1257713d6c20SMatthew Ahrens 	} else if (dmu_objset_is_receiving(dn->dn_objset)) {
12582f3d8780SMatthew Ahrens 		/*
1259713d6c20SMatthew Ahrens 		 * If we are receiving, we expect there to be no dbufs in
1260713d6c20SMatthew Ahrens 		 * the range to be freed, because receive modifies each
1261713d6c20SMatthew Ahrens 		 * block at most once, and in offset order.  If this is
1262713d6c20SMatthew Ahrens 		 * not the case, it can lead to performance problems,
1263713d6c20SMatthew Ahrens 		 * so note that we unexpectedly took the slow path.
12642f3d8780SMatthew Ahrens 		 */
1265713d6c20SMatthew Ahrens 		atomic_inc_64(&zfs_free_range_recv_miss);
12662f3d8780SMatthew Ahrens 	}
12672f3d8780SMatthew Ahrens 
12680f6d88adSAlex Reece 	db = avl_find(&dn->dn_dbufs, &db_search, &where);
12690f6d88adSAlex Reece 	ASSERT3P(db, ==, NULL);
12700f6d88adSAlex Reece 	db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
12710f6d88adSAlex Reece 
12720f6d88adSAlex Reece 	for (; db != NULL; db = db_next) {
12730f6d88adSAlex Reece 		db_next = AVL_NEXT(&dn->dn_dbufs, db);
12740a586ceaSMark Shellenbaum 		ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1275cdb0ab79Smaybee 
12760f6d88adSAlex Reece 		if (db->db_level != 0 || db->db_blkid > end_blkid) {
12770f6d88adSAlex Reece 			break;
12780f6d88adSAlex Reece 		}
12790f6d88adSAlex Reece 		ASSERT3U(db->db_blkid, >=, start_blkid);
1280fa9e4066Sahrens 
1281fa9e4066Sahrens 		/* found a level 0 buffer in the range */
12823b2aab18SMatthew Ahrens 		mutex_enter(&db->db_mtx);
12833b2aab18SMatthew Ahrens 		if (dbuf_undirty(db, tx)) {
12843b2aab18SMatthew Ahrens 			/* mutex has been dropped and dbuf destroyed */
1285fa9e4066Sahrens 			continue;
12863b2aab18SMatthew Ahrens 		}
1287fa9e4066Sahrens 
1288ea8dc4b6Seschrock 		if (db->db_state == DB_UNCACHED ||
128982c9918fSTim Haley 		    db->db_state == DB_NOFILL ||
1290ea8dc4b6Seschrock 		    db->db_state == DB_EVICTING) {
1291fa9e4066Sahrens 			ASSERT(db->db.db_data == NULL);
1292fa9e4066Sahrens 			mutex_exit(&db->db_mtx);
1293fa9e4066Sahrens 			continue;
1294fa9e4066Sahrens 		}
1295c543ec06Sahrens 		if (db->db_state == DB_READ || db->db_state == DB_FILL) {
1296c543ec06Sahrens 			/* will be handled in dbuf_read_done or dbuf_rele */
1297c717a561Smaybee 			db->db_freed_in_flight = TRUE;
1298fa9e4066Sahrens 			mutex_exit(&db->db_mtx);
1299fa9e4066Sahrens 			continue;
1300fa9e4066Sahrens 		}
1301ea8dc4b6Seschrock 		if (refcount_count(&db->db_holds) == 0) {
1302ea8dc4b6Seschrock 			ASSERT(db->db_buf);
1303*dcbf3bd6SGeorge Wilson 			dbuf_destroy(db);
1304ea8dc4b6Seschrock 			continue;
1305ea8dc4b6Seschrock 		}
1306c717a561Smaybee 		/* The dbuf is referenced */
1307fa9e4066Sahrens 
1308c717a561Smaybee 		if (db->db_last_dirty != NULL) {
1309c717a561Smaybee 			dbuf_dirty_record_t *dr = db->db_last_dirty;
1310c717a561Smaybee 
1311c717a561Smaybee 			if (dr->dr_txg == txg) {
131244eda4d7Smaybee 				/*
1313c717a561Smaybee 				 * This buffer is "in-use", re-adjust the file
1314c717a561Smaybee 				 * size to reflect that this buffer may
1315c717a561Smaybee 				 * contain new data when we sync.
131644eda4d7Smaybee 				 */
131706e0070dSMark Shellenbaum 				if (db->db_blkid != DMU_SPILL_BLKID &&
131806e0070dSMark Shellenbaum 				    db->db_blkid > dn->dn_maxblkid)
1319c717a561Smaybee 					dn->dn_maxblkid = db->db_blkid;
1320c717a561Smaybee 				dbuf_unoverride(dr);
1321c717a561Smaybee 			} else {
1322c717a561Smaybee 				/*
1323c717a561Smaybee 				 * This dbuf is not dirty in the open context.
1324c717a561Smaybee 				 * Either uncache it (if its not referenced in
1325c717a561Smaybee 				 * the open context) or reset its contents to
1326c717a561Smaybee 				 * empty.
1327c717a561Smaybee 				 */
1328c717a561Smaybee 				dbuf_fix_old_data(db, txg);
132944eda4d7Smaybee 			}
1330ea8dc4b6Seschrock 		}
1331c717a561Smaybee 		/* clear the contents if its cached */
1332ea8dc4b6Seschrock 		if (db->db_state == DB_CACHED) {
1333ea8dc4b6Seschrock 			ASSERT(db->db.db_data != NULL);
1334fa9e4066Sahrens 			arc_release(db->db_buf, db);
1335fa9e4066Sahrens 			bzero(db->db.db_data, db->db.db_size);
13366b4acc8bSahrens 			arc_buf_freeze(db->db_buf);
1337fa9e4066Sahrens 		}
1338ea8dc4b6Seschrock 
1339fa9e4066Sahrens 		mutex_exit(&db->db_mtx);
1340fa9e4066Sahrens 	}
1341fa9e4066Sahrens 	mutex_exit(&dn->dn_dbufs_mtx);
1342fa9e4066Sahrens }
1343fa9e4066Sahrens 
1344fa9e4066Sahrens static int
13451934e92fSmaybee dbuf_block_freeable(dmu_buf_impl_t *db)
1346fa9e4066Sahrens {
1347fa9e4066Sahrens 	dsl_dataset_t *ds = db->db_objset->os_dsl_dataset;
1348fa9e4066Sahrens 	uint64_t birth_txg = 0;
1349fa9e4066Sahrens 
1350fa9e4066Sahrens 	/*
1351fa9e4066Sahrens 	 * We don't need any locking to protect db_blkptr:
1352c717a561Smaybee 	 * If it's syncing, then db_last_dirty will be set
1353c717a561Smaybee 	 * so we'll ignore db_blkptr.
135443466aaeSMax Grossman 	 *
135543466aaeSMax Grossman 	 * This logic ensures that only block births for
135643466aaeSMax Grossman 	 * filled blocks are considered.
1357fa9e4066Sahrens 	 */
1358c717a561Smaybee 	ASSERT(MUTEX_HELD(&db->db_mtx));
135943466aaeSMax Grossman 	if (db->db_last_dirty && (db->db_blkptr == NULL ||
136043466aaeSMax Grossman 	    !BP_IS_HOLE(db->db_blkptr))) {
1361c717a561Smaybee 		birth_txg = db->db_last_dirty->dr_txg;
136243466aaeSMax Grossman 	} else if (db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) {
1363fa9e4066Sahrens 		birth_txg = db->db_blkptr->blk_birth;
136443466aaeSMax Grossman 	}
1365fa9e4066Sahrens 
1366837b568bSGeorge Wilson 	/*
136743466aaeSMax Grossman 	 * If this block don't exist or is in a snapshot, it can't be freed.
1368837b568bSGeorge Wilson 	 * Don't pass the bp to dsl_dataset_block_freeable() since we
1369837b568bSGeorge Wilson 	 * are holding the db_mtx lock and might deadlock if we are
1370837b568bSGeorge Wilson 	 * prefetching a dedup-ed block.
1371837b568bSGeorge Wilson 	 */
137243466aaeSMax Grossman 	if (birth_txg != 0)
13731934e92fSmaybee 		return (ds == NULL ||
1374837b568bSGeorge Wilson 		    dsl_dataset_block_freeable(ds, NULL, birth_txg));
1375fa9e4066Sahrens 	else
137643466aaeSMax Grossman 		return (B_FALSE);
1377fa9e4066Sahrens }
1378fa9e4066Sahrens 
1379fa9e4066Sahrens void
1380fa9e4066Sahrens dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
1381fa9e4066Sahrens {
1382fa9e4066Sahrens 	arc_buf_t *buf, *obuf;
1383fa9e4066Sahrens 	int osize = db->db.db_size;
1384ad23a2dbSjohansen 	arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1385744947dcSTom Erickson 	dnode_t *dn;
1386fa9e4066Sahrens 
13870a586ceaSMark Shellenbaum 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1388ea8dc4b6Seschrock 
1389744947dcSTom Erickson 	DB_DNODE_ENTER(db);
1390744947dcSTom Erickson 	dn = DB_DNODE(db);
1391744947dcSTom Erickson 
1392fa9e4066Sahrens 	/* XXX does *this* func really need the lock? */
1393744947dcSTom Erickson 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1394fa9e4066Sahrens 
1395fa9e4066Sahrens 	/*
139643466aaeSMax Grossman 	 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1397fa9e4066Sahrens 	 * is OK, because there can be no other references to the db
1398fa9e4066Sahrens 	 * when we are changing its size, so no concurrent DB_FILL can
1399fa9e4066Sahrens 	 * be happening.
1400fa9e4066Sahrens 	 */
1401ea8dc4b6Seschrock 	/*
1402ea8dc4b6Seschrock 	 * XXX we should be doing a dbuf_read, checking the return
1403ea8dc4b6Seschrock 	 * value and returning that up to our callers
1404ea8dc4b6Seschrock 	 */
140543466aaeSMax Grossman 	dmu_buf_will_dirty(&db->db, tx);
1406fa9e4066Sahrens 
1407fa9e4066Sahrens 	/* create the data buffer for the new block */
1408*dcbf3bd6SGeorge Wilson 	buf = arc_alloc_buf(dn->dn_objset->os_spa, size, db, type);
1409fa9e4066Sahrens 
1410fa9e4066Sahrens 	/* copy old block data to the new block */
1411fa9e4066Sahrens 	obuf = db->db_buf;
1412f65e61c0Sahrens 	bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
1413fa9e4066Sahrens 	/* zero the remainder */
1414f65e61c0Sahrens 	if (size > osize)
1415f65e61c0Sahrens 		bzero((uint8_t *)buf->b_data + osize, size - osize);
1416fa9e4066Sahrens 
1417fa9e4066Sahrens 	mutex_enter(&db->db_mtx);
1418fa9e4066Sahrens 	dbuf_set_data(db, buf);
1419*dcbf3bd6SGeorge Wilson 	arc_buf_destroy(obuf, db);
1420fa9e4066Sahrens 	db->db.db_size = size;
1421fa9e4066Sahrens 
1422c717a561Smaybee 	if (db->db_level == 0) {
1423c717a561Smaybee 		ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1424c717a561Smaybee 		db->db_last_dirty->dt.dl.dr_data = buf;
1425c717a561Smaybee 	}
1426fa9e4066Sahrens 	mutex_exit(&db->db_mtx);
1427fa9e4066Sahrens 
1428744947dcSTom Erickson 	dnode_willuse_space(dn, size-osize, tx);
1429744947dcSTom Erickson 	DB_DNODE_EXIT(db);
1430fa9e4066Sahrens }
1431fa9e4066Sahrens 
14323f9d6ad7SLin Ling void
14333f9d6ad7SLin Ling dbuf_release_bp(dmu_buf_impl_t *db)
14343f9d6ad7SLin Ling {
143543466aaeSMax Grossman 	objset_t *os = db->db_objset;
14363f9d6ad7SLin Ling 
14373f9d6ad7SLin Ling 	ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
14383f9d6ad7SLin Ling 	ASSERT(arc_released(os->os_phys_buf) ||
14393f9d6ad7SLin Ling 	    list_link_active(&os->os_dsl_dataset->ds_synced_link));
14403f9d6ad7SLin Ling 	ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
14413f9d6ad7SLin Ling 
14421b912ec7SGeorge Wilson 	(void) arc_release(db->db_buf, db);
14433f9d6ad7SLin Ling }
14443f9d6ad7SLin Ling 
14450f2e7d03SMatthew Ahrens /*
14460f2e7d03SMatthew Ahrens  * We already have a dirty record for this TXG, and we are being
14470f2e7d03SMatthew Ahrens  * dirtied again.
14480f2e7d03SMatthew Ahrens  */
14490f2e7d03SMatthew Ahrens static void
14500f2e7d03SMatthew Ahrens dbuf_redirty(dbuf_dirty_record_t *dr)
14510f2e7d03SMatthew Ahrens {
14520f2e7d03SMatthew Ahrens 	dmu_buf_impl_t *db = dr->dr_dbuf;
14530f2e7d03SMatthew Ahrens 
14540f2e7d03SMatthew Ahrens 	ASSERT(MUTEX_HELD(&db->db_mtx));
14550f2e7d03SMatthew Ahrens 
14560f2e7d03SMatthew Ahrens 	if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
14570f2e7d03SMatthew Ahrens 		/*
14580f2e7d03SMatthew Ahrens 		 * If this buffer has already been written out,
14590f2e7d03SMatthew Ahrens 		 * we now need to reset its state.
14600f2e7d03SMatthew Ahrens 		 */
14610f2e7d03SMatthew Ahrens 		dbuf_unoverride(dr);
14620f2e7d03SMatthew Ahrens 		if (db->db.db_object != DMU_META_DNODE_OBJECT &&
14630f2e7d03SMatthew Ahrens 		    db->db_state != DB_NOFILL) {
14640f2e7d03SMatthew Ahrens 			/* Already released on initial dirty, so just thaw. */
14650f2e7d03SMatthew Ahrens 			ASSERT(arc_released(db->db_buf));
14660f2e7d03SMatthew Ahrens 			arc_buf_thaw(db->db_buf);
14670f2e7d03SMatthew Ahrens 		}
14680f2e7d03SMatthew Ahrens 	}
14690f2e7d03SMatthew Ahrens }
14700f2e7d03SMatthew Ahrens 
1471c717a561Smaybee dbuf_dirty_record_t *
1472fa9e4066Sahrens dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1473fa9e4066Sahrens {
1474744947dcSTom Erickson 	dnode_t *dn;
1475744947dcSTom Erickson 	objset_t *os;
1476c717a561Smaybee 	dbuf_dirty_record_t **drp, *dr;
1477fa9e4066Sahrens 	int drop_struct_lock = FALSE;
1478d3469faaSMark Maybee 	boolean_t do_free_accounting = B_FALSE;
1479fa9e4066Sahrens 	int txgoff = tx->tx_txg & TXG_MASK;
1480fa9e4066Sahrens 
1481fa9e4066Sahrens 	ASSERT(tx->tx_txg != 0);
1482fa9e4066Sahrens 	ASSERT(!refcount_is_zero(&db->db_holds));
14839c9dc39aSek 	DMU_TX_DIRTY_BUF(tx, db);
1484fa9e4066Sahrens 
1485744947dcSTom Erickson 	DB_DNODE_ENTER(db);
1486744947dcSTom Erickson 	dn = DB_DNODE(db);
1487fa9e4066Sahrens 	/*
1488fa9e4066Sahrens 	 * Shouldn't dirty a regular buffer in syncing context.  Private
1489fa9e4066Sahrens 	 * objects may be dirtied in syncing context, but only if they
1490fa9e4066Sahrens 	 * were already pre-dirtied in open context.
1491fa9e4066Sahrens 	 */
1492c717a561Smaybee 	ASSERT(!dmu_tx_is_syncing(tx) ||
1493c717a561Smaybee 	    BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
149414843421SMatthew Ahrens 	    DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
149514843421SMatthew Ahrens 	    dn->dn_objset->os_dsl_dataset == NULL);
1496fa9e4066Sahrens 	/*
1497fa9e4066Sahrens 	 * We make this assert for private objects as well, but after we
1498fa9e4066Sahrens 	 * check if we're already dirty.  They are allowed to re-dirty
1499fa9e4066Sahrens 	 * in syncing context.
1500fa9e4066Sahrens 	 */
1501ea8dc4b6Seschrock 	ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1502c717a561Smaybee 	    dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1503fa9e4066Sahrens 	    (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1504fa9e4066Sahrens 
1505fa9e4066Sahrens 	mutex_enter(&db->db_mtx);
1506fa9e4066Sahrens 	/*
1507c717a561Smaybee 	 * XXX make this true for indirects too?  The problem is that
1508c717a561Smaybee 	 * transactions created with dmu_tx_create_assigned() from
1509c717a561Smaybee 	 * syncing context don't bother holding ahead.
1510fa9e4066Sahrens 	 */
1511c717a561Smaybee 	ASSERT(db->db_level != 0 ||
151282c9918fSTim Haley 	    db->db_state == DB_CACHED || db->db_state == DB_FILL ||
151382c9918fSTim Haley 	    db->db_state == DB_NOFILL);
1514fa9e4066Sahrens 
1515fa9e4066Sahrens 	mutex_enter(&dn->dn_mtx);
1516fa9e4066Sahrens 	/*
1517fa9e4066Sahrens 	 * Don't set dirtyctx to SYNC if we're just modifying this as we
1518fa9e4066Sahrens 	 * initialize the objset.
1519fa9e4066Sahrens 	 */
1520fa9e4066Sahrens 	if (dn->dn_dirtyctx == DN_UNDIRTIED &&
1521c717a561Smaybee 	    !BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1522fa9e4066Sahrens 		dn->dn_dirtyctx =
1523fa9e4066Sahrens 		    (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1524fa9e4066Sahrens 		ASSERT(dn->dn_dirtyctx_firstset == NULL);
1525fa9e4066Sahrens 		dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1526fa9e4066Sahrens 	}
1527fa9e4066Sahrens 	mutex_exit(&dn->dn_mtx);
1528fa9e4066Sahrens 
15290a586ceaSMark Shellenbaum 	if (db->db_blkid == DMU_SPILL_BLKID)
15300a586ceaSMark Shellenbaum 		dn->dn_have_spill = B_TRUE;
15310a586ceaSMark Shellenbaum 
1532fa9e4066Sahrens 	/*
1533fa9e4066Sahrens 	 * If this buffer is already dirty, we're done.
1534fa9e4066Sahrens 	 */
1535c717a561Smaybee 	drp = &db->db_last_dirty;
1536c717a561Smaybee 	ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1537c717a561Smaybee 	    db->db.db_object == DMU_META_DNODE_OBJECT);
15387e2186e3Sbonwick 	while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
15397e2186e3Sbonwick 		drp = &dr->dr_next;
15407e2186e3Sbonwick 	if (dr && dr->dr_txg == tx->tx_txg) {
1541744947dcSTom Erickson 		DB_DNODE_EXIT(db);
1542744947dcSTom Erickson 
15430f2e7d03SMatthew Ahrens 		dbuf_redirty(dr);
1544fa9e4066Sahrens 		mutex_exit(&db->db_mtx);
15457e2186e3Sbonwick 		return (dr);
1546fa9e4066Sahrens 	}
1547fa9e4066Sahrens 
1548fa9e4066Sahrens 	/*
1549fa9e4066Sahrens 	 * Only valid if not already dirty.
1550fa9e4066Sahrens 	 */
155114843421SMatthew Ahrens 	ASSERT(dn->dn_object == 0 ||
155214843421SMatthew Ahrens 	    dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1553fa9e4066Sahrens 	    (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1554fa9e4066Sahrens 
1555fa9e4066Sahrens 	ASSERT3U(dn->dn_nlevels, >, db->db_level);
1556fa9e4066Sahrens 	ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1557fa9e4066Sahrens 	    dn->dn_phys->dn_nlevels > db->db_level ||
1558fa9e4066Sahrens 	    dn->dn_next_nlevels[txgoff] > db->db_level ||
1559fa9e4066Sahrens 	    dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1560fa9e4066Sahrens 	    dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1561fa9e4066Sahrens 
1562fa9e4066Sahrens 	/*
1563fa9e4066Sahrens 	 * We should only be dirtying in syncing context if it's the
156414843421SMatthew Ahrens 	 * mos or we're initializing the os or it's a special object.
156514843421SMatthew Ahrens 	 * However, we are allowed to dirty in syncing context provided
156614843421SMatthew Ahrens 	 * we already dirtied it in open context.  Hence we must make
156714843421SMatthew Ahrens 	 * this assertion only if we're not already dirty.
1568fa9e4066Sahrens 	 */
1569744947dcSTom Erickson 	os = dn->dn_objset;
157014843421SMatthew Ahrens 	ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
157114843421SMatthew Ahrens 	    os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1572fa9e4066Sahrens 	ASSERT(db->db.db_size != 0);
1573fa9e4066Sahrens 
1574fa9e4066Sahrens 	dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1575fa9e4066Sahrens 
15760a586ceaSMark Shellenbaum 	if (db->db_blkid != DMU_BONUS_BLKID) {
15771934e92fSmaybee 		/*
15781934e92fSmaybee 		 * Update the accounting.
1579d3469faaSMark Maybee 		 * Note: we delay "free accounting" until after we drop
1580d3469faaSMark Maybee 		 * the db_mtx.  This keeps us from grabbing other locks
1581b24ab676SJeff Bonwick 		 * (and possibly deadlocking) in bp_get_dsize() while
1582d3469faaSMark Maybee 		 * also holding the db_mtx.
15831934e92fSmaybee 		 */
15841934e92fSmaybee 		dnode_willuse_space(dn, db->db.db_size, tx);
1585d3469faaSMark Maybee 		do_free_accounting = dbuf_block_freeable(db);
15861934e92fSmaybee 	}
15871934e92fSmaybee 
1588ea8dc4b6Seschrock 	/*
1589ea8dc4b6Seschrock 	 * If this buffer is dirty in an old transaction group we need
1590ea8dc4b6Seschrock 	 * to make a copy of it so that the changes we make in this
1591ea8dc4b6Seschrock 	 * transaction group won't leak out when we sync the older txg.
1592ea8dc4b6Seschrock 	 */
1593c717a561Smaybee 	dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1594c717a561Smaybee 	if (db->db_level == 0) {
1595c717a561Smaybee 		void *data_old = db->db_buf;
1596c717a561Smaybee 
159782c9918fSTim Haley 		if (db->db_state != DB_NOFILL) {
15980a586ceaSMark Shellenbaum 			if (db->db_blkid == DMU_BONUS_BLKID) {
159982c9918fSTim Haley 				dbuf_fix_old_data(db, tx->tx_txg);
160082c9918fSTim Haley 				data_old = db->db.db_data;
160182c9918fSTim Haley 			} else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
160282c9918fSTim Haley 				/*
160382c9918fSTim Haley 				 * Release the data buffer from the cache so
160482c9918fSTim Haley 				 * that we can modify it without impacting
160582c9918fSTim Haley 				 * possible other users of this cached data
160682c9918fSTim Haley 				 * block.  Note that indirect blocks and
160782c9918fSTim Haley 				 * private objects are not released until the
160882c9918fSTim Haley 				 * syncing state (since they are only modified
160982c9918fSTim Haley 				 * then).
161082c9918fSTim Haley 				 */
161182c9918fSTim Haley 				arc_release(db->db_buf, db);
161282c9918fSTim Haley 				dbuf_fix_old_data(db, tx->tx_txg);
161382c9918fSTim Haley 				data_old = db->db_buf;
161482c9918fSTim Haley 			}
161582c9918fSTim Haley 			ASSERT(data_old != NULL);
1616fa9e4066Sahrens 		}
1617c717a561Smaybee 		dr->dt.dl.dr_data = data_old;
1618c717a561Smaybee 	} else {
1619c717a561Smaybee 		mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1620c717a561Smaybee 		list_create(&dr->dt.di.dr_children,
1621c717a561Smaybee 		    sizeof (dbuf_dirty_record_t),
1622c717a561Smaybee 		    offsetof(dbuf_dirty_record_t, dr_dirty_node));
1623fa9e4066Sahrens 	}
162469962b56SMatthew Ahrens 	if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL)
162569962b56SMatthew Ahrens 		dr->dr_accounted = db->db.db_size;
1626c717a561Smaybee 	dr->dr_dbuf = db;
1627c717a561Smaybee 	dr->dr_txg = tx->tx_txg;
1628c717a561Smaybee 	dr->dr_next = *drp;
1629c717a561Smaybee 	*drp = dr;
1630fa9e4066Sahrens 
1631fa9e4066Sahrens 	/*
1632fa9e4066Sahrens 	 * We could have been freed_in_flight between the dbuf_noread
1633fa9e4066Sahrens 	 * and dbuf_dirty.  We win, as though the dbuf_noread() had
1634fa9e4066Sahrens 	 * happened after the free.
1635fa9e4066Sahrens 	 */
16360a586ceaSMark Shellenbaum 	if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
16370a586ceaSMark Shellenbaum 	    db->db_blkid != DMU_SPILL_BLKID) {
1638c717a561Smaybee 		mutex_enter(&dn->dn_mtx);
1639bf16b11eSMatthew Ahrens 		if (dn->dn_free_ranges[txgoff] != NULL) {
1640bf16b11eSMatthew Ahrens 			range_tree_clear(dn->dn_free_ranges[txgoff],
1641bf16b11eSMatthew Ahrens 			    db->db_blkid, 1);
1642bf16b11eSMatthew Ahrens 		}
1643c717a561Smaybee 		mutex_exit(&dn->dn_mtx);
1644c717a561Smaybee 		db->db_freed_in_flight = FALSE;
1645fa9e4066Sahrens 	}
1646fa9e4066Sahrens 
1647fa9e4066Sahrens 	/*
1648fa9e4066Sahrens 	 * This buffer is now part of this txg
1649fa9e4066Sahrens 	 */
1650fa9e4066Sahrens 	dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1651fa9e4066Sahrens 	db->db_dirtycnt += 1;
1652fa9e4066Sahrens 	ASSERT3U(db->db_dirtycnt, <=, 3);
1653fa9e4066Sahrens 
1654fa9e4066Sahrens 	mutex_exit(&db->db_mtx);
1655fa9e4066Sahrens 
16560a586ceaSMark Shellenbaum 	if (db->db_blkid == DMU_BONUS_BLKID ||
16570a586ceaSMark Shellenbaum 	    db->db_blkid == DMU_SPILL_BLKID) {
1658c717a561Smaybee 		mutex_enter(&dn->dn_mtx);
1659c717a561Smaybee 		ASSERT(!list_link_active(&dr->dr_dirty_node));
1660c717a561Smaybee 		list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1661c717a561Smaybee 		mutex_exit(&dn->dn_mtx);
1662fa9e4066Sahrens 		dnode_setdirty(dn, tx);
1663744947dcSTom Erickson 		DB_DNODE_EXIT(db);
1664c717a561Smaybee 		return (dr);
1665d3469faaSMark Maybee 	} else if (do_free_accounting) {
1666d3469faaSMark Maybee 		blkptr_t *bp = db->db_blkptr;
1667d3469faaSMark Maybee 		int64_t willfree = (bp && !BP_IS_HOLE(bp)) ?
1668b24ab676SJeff Bonwick 		    bp_get_dsize(os->os_spa, bp) : db->db.db_size;
1669d3469faaSMark Maybee 		/*
1670d3469faaSMark Maybee 		 * This is only a guess -- if the dbuf is dirty
1671d3469faaSMark Maybee 		 * in a previous txg, we don't know how much
1672d3469faaSMark Maybee 		 * space it will use on disk yet.  We should
1673d3469faaSMark Maybee 		 * really have the struct_rwlock to access
1674d3469faaSMark Maybee 		 * db_blkptr, but since this is just a guess,
1675d3469faaSMark Maybee 		 * it's OK if we get an odd answer.
1676d3469faaSMark Maybee 		 */
1677837b568bSGeorge Wilson 		ddt_prefetch(os->os_spa, bp);
1678d3469faaSMark Maybee 		dnode_willuse_space(dn, -willfree, tx);
1679fa9e4066Sahrens 	}
1680fa9e4066Sahrens 
1681fa9e4066Sahrens 	if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1682fa9e4066Sahrens 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
1683fa9e4066Sahrens 		drop_struct_lock = TRUE;
1684fa9e4066Sahrens 	}
1685fa9e4066Sahrens 
16868346f03fSJonathan W Adams 	if (db->db_level == 0) {
16878346f03fSJonathan W Adams 		dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
16888346f03fSJonathan W Adams 		ASSERT(dn->dn_maxblkid >= db->db_blkid);
16898346f03fSJonathan W Adams 	}
16908346f03fSJonathan W Adams 
169144eda4d7Smaybee 	if (db->db_level+1 < dn->dn_nlevels) {
1692c717a561Smaybee 		dmu_buf_impl_t *parent = db->db_parent;
1693c717a561Smaybee 		dbuf_dirty_record_t *di;
1694c717a561Smaybee 		int parent_held = FALSE;
1695c717a561Smaybee 
1696c717a561Smaybee 		if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1697c717a561Smaybee 			int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1698c717a561Smaybee 
1699c717a561Smaybee 			parent = dbuf_hold_level(dn, db->db_level+1,
1700c717a561Smaybee 			    db->db_blkid >> epbs, FTAG);
170101025c89SJohn Harres 			ASSERT(parent != NULL);
1702c717a561Smaybee 			parent_held = TRUE;
1703c717a561Smaybee 		}
1704fa9e4066Sahrens 		if (drop_struct_lock)
1705fa9e4066Sahrens 			rw_exit(&dn->dn_struct_rwlock);
1706c717a561Smaybee 		ASSERT3U(db->db_level+1, ==, parent->db_level);
1707c717a561Smaybee 		di = dbuf_dirty(parent, tx);
1708c717a561Smaybee 		if (parent_held)
1709c717a561Smaybee 			dbuf_rele(parent, FTAG);
1710c717a561Smaybee 
1711c717a561Smaybee 		mutex_enter(&db->db_mtx);
171269962b56SMatthew Ahrens 		/*
171369962b56SMatthew Ahrens 		 * Since we've dropped the mutex, it's possible that
171469962b56SMatthew Ahrens 		 * dbuf_undirty() might have changed this out from under us.
171569962b56SMatthew Ahrens 		 */
1716c717a561Smaybee 		if (db->db_last_dirty == dr ||
1717c717a561Smaybee 		    dn->dn_object == DMU_META_DNODE_OBJECT) {
1718c717a561Smaybee 			mutex_enter(&di->dt.di.dr_mtx);
1719c717a561Smaybee 			ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1720c717a561Smaybee 			ASSERT(!list_link_active(&dr->dr_dirty_node));
1721c717a561Smaybee 			list_insert_tail(&di->dt.di.dr_children, dr);
1722c717a561Smaybee 			mutex_exit(&di->dt.di.dr_mtx);
1723c717a561Smaybee 			dr->dr_parent = di;
1724c717a561Smaybee 		}
1725c717a561Smaybee 		mutex_exit(&db->db_mtx);
1726fa9e4066Sahrens 	} else {
1727c717a561Smaybee 		ASSERT(db->db_level+1 == dn->dn_nlevels);
1728c717a561Smaybee 		ASSERT(db->db_blkid < dn->dn_nblkptr);
1729744947dcSTom Erickson 		ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1730c717a561Smaybee 		mutex_enter(&dn->dn_mtx);
1731c717a561Smaybee 		ASSERT(!list_link_active(&dr->dr_dirty_node));
1732c717a561Smaybee 		list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1733c717a561Smaybee 		mutex_exit(&dn->dn_mtx);
1734fa9e4066Sahrens 		if (drop_struct_lock)
1735fa9e4066Sahrens 			rw_exit(&dn->dn_struct_rwlock);
1736fa9e4066Sahrens 	}
1737fa9e4066Sahrens 
1738fa9e4066Sahrens 	dnode_setdirty(dn, tx);
1739744947dcSTom Erickson 	DB_DNODE_EXIT(db);
1740c717a561Smaybee 	return (dr);
1741fa9e4066Sahrens }
1742fa9e4066Sahrens 
17433b2aab18SMatthew Ahrens /*
17443e30c24aSWill Andrews  * Undirty a buffer in the transaction group referenced by the given
17453e30c24aSWill Andrews  * transaction.  Return whether this evicted the dbuf.
17463b2aab18SMatthew Ahrens  */
17473b2aab18SMatthew Ahrens static boolean_t
1748fa9e4066Sahrens dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1749fa9e4066Sahrens {
1750744947dcSTom Erickson 	dnode_t *dn;
1751c717a561Smaybee 	uint64_t txg = tx->tx_txg;
175217f17c2dSbonwick 	dbuf_dirty_record_t *dr, **drp;
1753fa9e4066Sahrens 
1754c717a561Smaybee 	ASSERT(txg != 0);
175546e1baa6SMatthew Ahrens 
175646e1baa6SMatthew Ahrens 	/*
175746e1baa6SMatthew Ahrens 	 * Due to our use of dn_nlevels below, this can only be called
175846e1baa6SMatthew Ahrens 	 * in open context, unless we are operating on the MOS.
175946e1baa6SMatthew Ahrens 	 * From syncing context, dn_nlevels may be different from the
176046e1baa6SMatthew Ahrens 	 * dn_nlevels used when dbuf was dirtied.
176146e1baa6SMatthew Ahrens 	 */
176246e1baa6SMatthew Ahrens 	ASSERT(db->db_objset ==
176346e1baa6SMatthew Ahrens 	    dmu_objset_pool(db->db_objset)->dp_meta_objset ||
176446e1baa6SMatthew Ahrens 	    txg != spa_syncing_txg(dmu_objset_spa(db->db_objset)));
17650a586ceaSMark Shellenbaum 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
17663b2aab18SMatthew Ahrens 	ASSERT0(db->db_level);
17673b2aab18SMatthew Ahrens 	ASSERT(MUTEX_HELD(&db->db_mtx));
1768fa9e4066Sahrens 
1769fa9e4066Sahrens 	/*
1770fa9e4066Sahrens 	 * If this buffer is not dirty, we're done.
1771fa9e4066Sahrens 	 */
177217f17c2dSbonwick 	for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1773c717a561Smaybee 		if (dr->dr_txg <= txg)
1774c717a561Smaybee 			break;
17753b2aab18SMatthew Ahrens 	if (dr == NULL || dr->dr_txg < txg)
17763b2aab18SMatthew Ahrens 		return (B_FALSE);
1777c717a561Smaybee 	ASSERT(dr->dr_txg == txg);
1778b24ab676SJeff Bonwick 	ASSERT(dr->dr_dbuf == db);
1779fa9e4066Sahrens 
1780744947dcSTom Erickson 	DB_DNODE_ENTER(db);
1781744947dcSTom Erickson 	dn = DB_DNODE(db);
1782744947dcSTom Erickson 
1783fa9e4066Sahrens 	dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1784fa9e4066Sahrens 
1785fa9e4066Sahrens 	ASSERT(db->db.db_size != 0);
1786fa9e4066Sahrens 
178746e1baa6SMatthew Ahrens 	dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset),
178846e1baa6SMatthew Ahrens 	    dr->dr_accounted, txg);
1789fa9e4066Sahrens 
179017f17c2dSbonwick 	*drp = dr->dr_next;
1791c717a561Smaybee 
17923f2366c2SGordon Ross 	/*
17933f2366c2SGordon Ross 	 * Note that there are three places in dbuf_dirty()
17943f2366c2SGordon Ross 	 * where this dirty record may be put on a list.
17953f2366c2SGordon Ross 	 * Make sure to do a list_remove corresponding to
17963f2366c2SGordon Ross 	 * every one of those list_insert calls.
17973f2366c2SGordon Ross 	 */
1798c717a561Smaybee 	if (dr->dr_parent) {
1799c717a561Smaybee 		mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1800c717a561Smaybee 		list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1801c717a561Smaybee 		mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
18023f2366c2SGordon Ross 	} else if (db->db_blkid == DMU_SPILL_BLKID ||
180346e1baa6SMatthew Ahrens 	    db->db_level + 1 == dn->dn_nlevels) {
1804cdb0ab79Smaybee 		ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1805c717a561Smaybee 		mutex_enter(&dn->dn_mtx);
1806c717a561Smaybee 		list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1807c717a561Smaybee 		mutex_exit(&dn->dn_mtx);
1808c717a561Smaybee 	}
1809744947dcSTom Erickson 	DB_DNODE_EXIT(db);
1810c717a561Smaybee 
18113b2aab18SMatthew Ahrens 	if (db->db_state != DB_NOFILL) {
18123b2aab18SMatthew Ahrens 		dbuf_unoverride(dr);
1813c717a561Smaybee 
1814c717a561Smaybee 		ASSERT(db->db_buf != NULL);
18153b2aab18SMatthew Ahrens 		ASSERT(dr->dt.dl.dr_data != NULL);
18163b2aab18SMatthew Ahrens 		if (dr->dt.dl.dr_data != db->db_buf)
1817*dcbf3bd6SGeorge Wilson 			arc_buf_destroy(dr->dt.dl.dr_data, db);
1818c717a561Smaybee 	}
1819d2b3cbbdSJorgen Lundman 
1820c717a561Smaybee 	kmem_free(dr, sizeof (dbuf_dirty_record_t));
1821fa9e4066Sahrens 
1822fa9e4066Sahrens 	ASSERT(db->db_dirtycnt > 0);
1823fa9e4066Sahrens 	db->db_dirtycnt -= 1;
1824fa9e4066Sahrens 
1825c717a561Smaybee 	if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1826*dcbf3bd6SGeorge Wilson 		ASSERT(db->db_state == DB_NOFILL || arc_released(db->db_buf));
1827*dcbf3bd6SGeorge Wilson 		dbuf_destroy(db);
18283b2aab18SMatthew Ahrens 		return (B_TRUE);
1829fa9e4066Sahrens 	}
1830fa9e4066Sahrens 
18313b2aab18SMatthew Ahrens 	return (B_FALSE);
1832fa9e4066Sahrens }
1833fa9e4066Sahrens 
1834fa9e4066Sahrens void
183543466aaeSMax Grossman dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
1836fa9e4066Sahrens {
183743466aaeSMax Grossman 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
18381ab7f2deSmaybee 	int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1839fa9e4066Sahrens 
1840fa9e4066Sahrens 	ASSERT(tx->tx_txg != 0);
1841fa9e4066Sahrens 	ASSERT(!refcount_is_zero(&db->db_holds));
1842fa9e4066Sahrens 
18430f2e7d03SMatthew Ahrens 	/*
18440f2e7d03SMatthew Ahrens 	 * Quick check for dirtyness.  For already dirty blocks, this
18450f2e7d03SMatthew Ahrens 	 * reduces runtime of this function by >90%, and overall performance
18460f2e7d03SMatthew Ahrens 	 * by 50% for some workloads (e.g. file deletion with indirect blocks
18470f2e7d03SMatthew Ahrens 	 * cached).
18480f2e7d03SMatthew Ahrens 	 */
18490f2e7d03SMatthew Ahrens 	mutex_enter(&db->db_mtx);
18500f2e7d03SMatthew Ahrens 	dbuf_dirty_record_t *dr;
18510f2e7d03SMatthew Ahrens 	for (dr = db->db_last_dirty;
18520f2e7d03SMatthew Ahrens 	    dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) {
18530f2e7d03SMatthew Ahrens 		/*
18540f2e7d03SMatthew Ahrens 		 * It's possible that it is already dirty but not cached,
18550f2e7d03SMatthew Ahrens 		 * because there are some calls to dbuf_dirty() that don't
18560f2e7d03SMatthew Ahrens 		 * go through dmu_buf_will_dirty().
18570f2e7d03SMatthew Ahrens 		 */
18580f2e7d03SMatthew Ahrens 		if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) {
18590f2e7d03SMatthew Ahrens 			/* This dbuf is already dirty and cached. */
18600f2e7d03SMatthew Ahrens 			dbuf_redirty(dr);
18610f2e7d03SMatthew Ahrens 			mutex_exit(&db->db_mtx);
18620f2e7d03SMatthew Ahrens 			return;
18630f2e7d03SMatthew Ahrens 		}
18640f2e7d03SMatthew Ahrens 	}
18650f2e7d03SMatthew Ahrens 	mutex_exit(&db->db_mtx);
18660f2e7d03SMatthew Ahrens 
1867744947dcSTom Erickson 	DB_DNODE_ENTER(db);
1868744947dcSTom Erickson 	if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
1869fa9e4066Sahrens 		rf |= DB_RF_HAVESTRUCT;
1870744947dcSTom Erickson 	DB_DNODE_EXIT(db);
1871ea8dc4b6Seschrock 	(void) dbuf_read(db, NULL, rf);
1872c717a561Smaybee 	(void) dbuf_dirty(db, tx);
1873fa9e4066Sahrens }
1874fa9e4066Sahrens 
187582c9918fSTim Haley void
187682c9918fSTim Haley dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
187782c9918fSTim Haley {
187882c9918fSTim Haley 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
187982c9918fSTim Haley 
188082c9918fSTim Haley 	db->db_state = DB_NOFILL;
188182c9918fSTim Haley 
188282c9918fSTim Haley 	dmu_buf_will_fill(db_fake, tx);
188382c9918fSTim Haley }
188482c9918fSTim Haley 
1885fa9e4066Sahrens void
1886ea8dc4b6Seschrock dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1887fa9e4066Sahrens {
1888ea8dc4b6Seschrock 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1889ea8dc4b6Seschrock 
18900a586ceaSMark Shellenbaum 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1891fa9e4066Sahrens 	ASSERT(tx->tx_txg != 0);
1892fa9e4066Sahrens 	ASSERT(db->db_level == 0);
1893fa9e4066Sahrens 	ASSERT(!refcount_is_zero(&db->db_holds));
1894fa9e4066Sahrens 
1895ea8dc4b6Seschrock 	ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1896fa9e4066Sahrens 	    dmu_tx_private_ok(tx));
1897fa9e4066Sahrens 
1898fa9e4066Sahrens 	dbuf_noread(db);
1899c717a561Smaybee 	(void) dbuf_dirty(db, tx);
1900fa9e4066Sahrens }
1901fa9e4066Sahrens 
1902fa9e4066Sahrens #pragma weak dmu_buf_fill_done = dbuf_fill_done
1903fa9e4066Sahrens /* ARGSUSED */
1904fa9e4066Sahrens void
1905fa9e4066Sahrens dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1906fa9e4066Sahrens {
1907fa9e4066Sahrens 	mutex_enter(&db->db_mtx);
19089c9dc39aSek 	DBUF_VERIFY(db);
1909fa9e4066Sahrens 
1910fa9e4066Sahrens 	if (db->db_state == DB_FILL) {
1911c717a561Smaybee 		if (db->db_level == 0 && db->db_freed_in_flight) {
19120a586ceaSMark Shellenbaum 			ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1913fa9e4066Sahrens 			/* we were freed while filling */
1914fa9e4066Sahrens 			/* XXX dbuf_undirty? */
1915fa9e4066Sahrens 			bzero(db->db.db_data, db->db.db_size);
1916c717a561Smaybee 			db->db_freed_in_flight = FALSE;
1917fa9e4066Sahrens 		}
1918fa9e4066Sahrens 		db->db_state = DB_CACHED;
1919fa9e4066Sahrens 		cv_broadcast(&db->db_changed);
1920fa9e4066Sahrens 	}
1921fa9e4066Sahrens 	mutex_exit(&db->db_mtx);
1922fa9e4066Sahrens }
1923fa9e4066Sahrens 
19245d7b4d43SMatthew Ahrens void
19255d7b4d43SMatthew Ahrens dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data,
19265d7b4d43SMatthew Ahrens     bp_embedded_type_t etype, enum zio_compress comp,
19275d7b4d43SMatthew Ahrens     int uncompressed_size, int compressed_size, int byteorder,
19285d7b4d43SMatthew Ahrens     dmu_tx_t *tx)
19295d7b4d43SMatthew Ahrens {
19305d7b4d43SMatthew Ahrens 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
19315d7b4d43SMatthew Ahrens 	struct dirty_leaf *dl;
19325d7b4d43SMatthew Ahrens 	dmu_object_type_t type;
19335d7b4d43SMatthew Ahrens 
1934ca0cc391SMatthew Ahrens 	if (etype == BP_EMBEDDED_TYPE_DATA) {
1935ca0cc391SMatthew Ahrens 		ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset),
1936ca0cc391SMatthew Ahrens 		    SPA_FEATURE_EMBEDDED_DATA));
1937ca0cc391SMatthew Ahrens 	}
1938ca0cc391SMatthew Ahrens 
19395d7b4d43SMatthew Ahrens 	DB_DNODE_ENTER(db);
19405d7b4d43SMatthew Ahrens 	type = DB_DNODE(db)->dn_type;
19415d7b4d43SMatthew Ahrens 	DB_DNODE_EXIT(db);
19425d7b4d43SMatthew Ahrens 
19435d7b4d43SMatthew Ahrens 	ASSERT0(db->db_level);
19445d7b4d43SMatthew Ahrens 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
19455d7b4d43SMatthew Ahrens 
19465d7b4d43SMatthew Ahrens 	dmu_buf_will_not_fill(dbuf, tx);
19475d7b4d43SMatthew Ahrens 
19485d7b4d43SMatthew Ahrens 	ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
19495d7b4d43SMatthew Ahrens 	dl = &db->db_last_dirty->dt.dl;
19505d7b4d43SMatthew Ahrens 	encode_embedded_bp_compressed(&dl->dr_overridden_by,
19515d7b4d43SMatthew Ahrens 	    data, comp, uncompressed_size, compressed_size);
19525d7b4d43SMatthew Ahrens 	BPE_SET_ETYPE(&dl->dr_overridden_by, etype);
19535d7b4d43SMatthew Ahrens 	BP_SET_TYPE(&dl->dr_overridden_by, type);
19545d7b4d43SMatthew Ahrens 	BP_SET_LEVEL(&dl->dr_overridden_by, 0);
19555d7b4d43SMatthew Ahrens 	BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder);
19565d7b4d43SMatthew Ahrens 
19575d7b4d43SMatthew Ahrens 	dl->dr_override_state = DR_OVERRIDDEN;
19585d7b4d43SMatthew Ahrens 	dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg;
19595d7b4d43SMatthew Ahrens }
19605d7b4d43SMatthew Ahrens 
19612fdbea25SAleksandr Guzovskiy /*
19622fdbea25SAleksandr Guzovskiy  * Directly assign a provided arc buf to a given dbuf if it's not referenced
19632fdbea25SAleksandr Guzovskiy  * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
19642fdbea25SAleksandr Guzovskiy  */
19652fdbea25SAleksandr Guzovskiy void
19662fdbea25SAleksandr Guzovskiy dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
19672fdbea25SAleksandr Guzovskiy {
19682fdbea25SAleksandr Guzovskiy 	ASSERT(!refcount_is_zero(&db->db_holds));
19690a586ceaSMark Shellenbaum 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
19702fdbea25SAleksandr Guzovskiy 	ASSERT(db->db_level == 0);
19712fdbea25SAleksandr Guzovskiy 	ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA);
19722fdbea25SAleksandr Guzovskiy 	ASSERT(buf != NULL);
19732fdbea25SAleksandr Guzovskiy 	ASSERT(arc_buf_size(buf) == db->db.db_size);
19742fdbea25SAleksandr Guzovskiy 	ASSERT(tx->tx_txg != 0);
19752fdbea25SAleksandr Guzovskiy 
19762fdbea25SAleksandr Guzovskiy 	arc_return_buf(buf, db);
19772fdbea25SAleksandr Guzovskiy 	ASSERT(arc_released(buf));
19782fdbea25SAleksandr Guzovskiy 
19792fdbea25SAleksandr Guzovskiy 	mutex_enter(&db->db_mtx);
19802fdbea25SAleksandr Guzovskiy 
19812fdbea25SAleksandr Guzovskiy 	while (db->db_state == DB_READ || db->db_state == DB_FILL)
19822fdbea25SAleksandr Guzovskiy 		cv_wait(&db->db_changed, &db->db_mtx);
19832fdbea25SAleksandr Guzovskiy 
19842fdbea25SAleksandr Guzovskiy 	ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
19852fdbea25SAleksandr Guzovskiy 
19862fdbea25SAleksandr Guzovskiy 	if (db->db_state == DB_CACHED &&
19872fdbea25SAleksandr Guzovskiy 	    refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
19882fdbea25SAleksandr Guzovskiy 		mutex_exit(&db->db_mtx);
19892fdbea25SAleksandr Guzovskiy 		(void) dbuf_dirty(db, tx);
19902fdbea25SAleksandr Guzovskiy 		bcopy(buf->b_data, db->db.db_data, db->db.db_size);
1991*dcbf3bd6SGeorge Wilson 		arc_buf_destroy(buf, db);
1992c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 		xuio_stat_wbuf_copied();
19932fdbea25SAleksandr Guzovskiy 		return;
19942fdbea25SAleksandr Guzovskiy 	}
19952fdbea25SAleksandr Guzovskiy 
1996c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 	xuio_stat_wbuf_nocopy();
19972fdbea25SAleksandr Guzovskiy 	if (db->db_state == DB_CACHED) {
19982fdbea25SAleksandr Guzovskiy 		dbuf_dirty_record_t *dr = db->db_last_dirty;
19992fdbea25SAleksandr Guzovskiy 
20002fdbea25SAleksandr Guzovskiy 		ASSERT(db->db_buf != NULL);
20012fdbea25SAleksandr Guzovskiy 		if (dr != NULL && dr->dr_txg == tx->tx_txg) {
20022fdbea25SAleksandr Guzovskiy 			ASSERT(dr->dt.dl.dr_data == db->db_buf);
20032fdbea25SAleksandr Guzovskiy 			if (!arc_released(db->db_buf)) {
20042fdbea25SAleksandr Guzovskiy 				ASSERT(dr->dt.dl.dr_override_state ==
20052fdbea25SAleksandr Guzovskiy 				    DR_OVERRIDDEN);
20062fdbea25SAleksandr Guzovskiy 				arc_release(db->db_buf, db);
20072fdbea25SAleksandr Guzovskiy 			}
20082fdbea25SAleksandr Guzovskiy 			dr->dt.dl.dr_data = buf;
2009*dcbf3bd6SGeorge Wilson 			arc_buf_destroy(db->db_buf, db);
20102fdbea25SAleksandr Guzovskiy 		} else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
20112fdbea25SAleksandr Guzovskiy 			arc_release(db->db_buf, db);
2012*dcbf3bd6SGeorge Wilson 			arc_buf_destroy(db->db_buf, db);
20132fdbea25SAleksandr Guzovskiy 		}
20142fdbea25SAleksandr Guzovskiy 		db->db_buf = NULL;
20152fdbea25SAleksandr Guzovskiy 	}
20162fdbea25SAleksandr Guzovskiy 	ASSERT(db->db_buf == NULL);
20172fdbea25SAleksandr Guzovskiy 	dbuf_set_data(db, buf);
20182fdbea25SAleksandr Guzovskiy 	db->db_state = DB_FILL;
20192fdbea25SAleksandr Guzovskiy 	mutex_exit(&db->db_mtx);
20202fdbea25SAleksandr Guzovskiy 	(void) dbuf_dirty(db, tx);
202143466aaeSMax Grossman 	dmu_buf_fill_done(&db->db, tx);
20222fdbea25SAleksandr Guzovskiy }
20232fdbea25SAleksandr Guzovskiy 
2024ea8dc4b6Seschrock void
2025*dcbf3bd6SGeorge Wilson dbuf_destroy(dmu_buf_impl_t *db)
2026fa9e4066Sahrens {
2027744947dcSTom Erickson 	dnode_t *dn;
2028ea8dc4b6Seschrock 	dmu_buf_impl_t *parent = db->db_parent;
2029744947dcSTom Erickson 	dmu_buf_impl_t *dndb;
2030fa9e4066Sahrens 
2031fa9e4066Sahrens 	ASSERT(MUTEX_HELD(&db->db_mtx));
2032fa9e4066Sahrens 	ASSERT(refcount_is_zero(&db->db_holds));
2033fa9e4066Sahrens 
2034*dcbf3bd6SGeorge Wilson 	if (db->db_buf != NULL) {
2035*dcbf3bd6SGeorge Wilson 		arc_buf_destroy(db->db_buf, db);
2036*dcbf3bd6SGeorge Wilson 		db->db_buf = NULL;
2037*dcbf3bd6SGeorge Wilson 	}
2038ea8dc4b6Seschrock 
2039*dcbf3bd6SGeorge Wilson 	if (db->db_blkid == DMU_BONUS_BLKID) {
2040ea8dc4b6Seschrock 		ASSERT(db->db.db_data != NULL);
2041*dcbf3bd6SGeorge Wilson 		zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
2042*dcbf3bd6SGeorge Wilson 		arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
2043fa9e4066Sahrens 		db->db_state = DB_UNCACHED;
2044fa9e4066Sahrens 	}
2045fa9e4066Sahrens 
2046*dcbf3bd6SGeorge Wilson 	dbuf_clear_data(db);
2047*dcbf3bd6SGeorge Wilson 
2048*dcbf3bd6SGeorge Wilson 	if (multilist_link_active(&db->db_cache_link)) {
2049*dcbf3bd6SGeorge Wilson 		multilist_remove(&dbuf_cache, db);
2050*dcbf3bd6SGeorge Wilson 		(void) refcount_remove_many(&dbuf_cache_size,
2051*dcbf3bd6SGeorge Wilson 		    db->db.db_size, db);
2052*dcbf3bd6SGeorge Wilson 	}
2053*dcbf3bd6SGeorge Wilson 
205482c9918fSTim Haley 	ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
2055fa9e4066Sahrens 	ASSERT(db->db_data_pending == NULL);
2056fa9e4066Sahrens 
2057ea8dc4b6Seschrock 	db->db_state = DB_EVICTING;
2058ea8dc4b6Seschrock 	db->db_blkptr = NULL;
2059ea8dc4b6Seschrock 
2060*dcbf3bd6SGeorge Wilson 	/*
2061*dcbf3bd6SGeorge Wilson 	 * Now that db_state is DB_EVICTING, nobody else can find this via
2062*dcbf3bd6SGeorge Wilson 	 * the hash table.  We can now drop db_mtx, which allows us to
2063*dcbf3bd6SGeorge Wilson 	 * acquire the dn_dbufs_mtx.
2064*dcbf3bd6SGeorge Wilson 	 */
2065*dcbf3bd6SGeorge Wilson 	mutex_exit(&db->db_mtx);
2066*dcbf3bd6SGeorge Wilson 
2067744947dcSTom Erickson 	DB_DNODE_ENTER(db);
2068744947dcSTom Erickson 	dn = DB_DNODE(db);
2069744947dcSTom Erickson 	dndb = dn->dn_dbuf;
2070*dcbf3bd6SGeorge Wilson 	if (db->db_blkid != DMU_BONUS_BLKID) {
2071*dcbf3bd6SGeorge Wilson 		boolean_t needlock = !MUTEX_HELD(&dn->dn_dbufs_mtx);
2072*dcbf3bd6SGeorge Wilson 		if (needlock)
2073*dcbf3bd6SGeorge Wilson 			mutex_enter(&dn->dn_dbufs_mtx);
20740f6d88adSAlex Reece 		avl_remove(&dn->dn_dbufs, db);
2075640c1670SJosef 'Jeff' Sipek 		atomic_dec_32(&dn->dn_dbufs_count);
2076744947dcSTom Erickson 		membar_producer();
2077744947dcSTom Erickson 		DB_DNODE_EXIT(db);
2078*dcbf3bd6SGeorge Wilson 		if (needlock)
2079*dcbf3bd6SGeorge Wilson 			mutex_exit(&dn->dn_dbufs_mtx);
2080744947dcSTom Erickson 		/*
2081744947dcSTom Erickson 		 * Decrementing the dbuf count means that the hold corresponding
2082744947dcSTom Erickson 		 * to the removed dbuf is no longer discounted in dnode_move(),
2083744947dcSTom Erickson 		 * so the dnode cannot be moved until after we release the hold.
2084744947dcSTom Erickson 		 * The membar_producer() ensures visibility of the decremented
2085744947dcSTom Erickson 		 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
2086744947dcSTom Erickson 		 * release any lock.
2087744947dcSTom Erickson 		 */
2088ea8dc4b6Seschrock 		dnode_rele(dn, db);
2089744947dcSTom Erickson 		db->db_dnode_handle = NULL;
2090*dcbf3bd6SGeorge Wilson 
2091*dcbf3bd6SGeorge Wilson 		dbuf_hash_remove(db);
2092744947dcSTom Erickson 	} else {
2093744947dcSTom Erickson 		DB_DNODE_EXIT(db);
2094ea8dc4b6Seschrock 	}
2095ea8dc4b6Seschrock 
2096*dcbf3bd6SGeorge Wilson 	ASSERT(refcount_is_zero(&db->db_holds));
2097ea8dc4b6Seschrock 
2098*dcbf3bd6SGeorge Wilson 	db->db_parent = NULL;
2099*dcbf3bd6SGeorge Wilson 
2100*dcbf3bd6SGeorge Wilson 	ASSERT(db->db_buf == NULL);
2101*dcbf3bd6SGeorge Wilson 	ASSERT(db->db.db_data == NULL);
2102*dcbf3bd6SGeorge Wilson 	ASSERT(db->db_hash_next == NULL);
2103*dcbf3bd6SGeorge Wilson 	ASSERT(db->db_blkptr == NULL);
2104*dcbf3bd6SGeorge Wilson 	ASSERT(db->db_data_pending == NULL);
2105*dcbf3bd6SGeorge Wilson 	ASSERT(!multilist_link_active(&db->db_cache_link));
2106*dcbf3bd6SGeorge Wilson 
2107*dcbf3bd6SGeorge Wilson 	kmem_cache_free(dbuf_kmem_cache, db);
2108*dcbf3bd6SGeorge Wilson 	arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2109fa9e4066Sahrens 
2110fa9e4066Sahrens 	/*
2111744947dcSTom Erickson 	 * If this dbuf is referenced from an indirect dbuf,
2112fa9e4066Sahrens 	 * decrement the ref count on the indirect dbuf.
2113fa9e4066Sahrens 	 */
2114c543ec06Sahrens 	if (parent && parent != dndb)
2115ea8dc4b6Seschrock 		dbuf_rele(parent, db);
2116fa9e4066Sahrens }
2117fa9e4066Sahrens 
2118a2cdcdd2SPaul Dagnelie /*
2119a2cdcdd2SPaul Dagnelie  * Note: While bpp will always be updated if the function returns success,
2120a2cdcdd2SPaul Dagnelie  * parentp will not be updated if the dnode does not have dn_dbuf filled in;
2121a2cdcdd2SPaul Dagnelie  * this happens when the dnode is the meta-dnode, or a userused or groupused
2122a2cdcdd2SPaul Dagnelie  * object.
2123a2cdcdd2SPaul Dagnelie  */
2124fa9e4066Sahrens static int
2125fa9e4066Sahrens dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
2126fa9e4066Sahrens     dmu_buf_impl_t **parentp, blkptr_t **bpp)
2127fa9e4066Sahrens {
2128fa9e4066Sahrens 	int nlevels, epbs;
2129fa9e4066Sahrens 
21300b69c2f0Sahrens 	*parentp = NULL;
21310b69c2f0Sahrens 	*bpp = NULL;
21320b69c2f0Sahrens 
21330a586ceaSMark Shellenbaum 	ASSERT(blkid != DMU_BONUS_BLKID);
21340a586ceaSMark Shellenbaum 
21350a586ceaSMark Shellenbaum 	if (blkid == DMU_SPILL_BLKID) {
21360a586ceaSMark Shellenbaum 		mutex_enter(&dn->dn_mtx);
213706e0070dSMark Shellenbaum 		if (dn->dn_have_spill &&
213806e0070dSMark Shellenbaum 		    (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
21390a586ceaSMark Shellenbaum 			*bpp = &dn->dn_phys->dn_spill;
21400a586ceaSMark Shellenbaum 		else
21410a586ceaSMark Shellenbaum 			*bpp = NULL;
21420a586ceaSMark Shellenbaum 		dbuf_add_ref(dn->dn_dbuf, NULL);
21430a586ceaSMark Shellenbaum 		*parentp = dn->dn_dbuf;
21440a586ceaSMark Shellenbaum 		mutex_exit(&dn->dn_mtx);
21450a586ceaSMark Shellenbaum 		return (0);
21460a586ceaSMark Shellenbaum 	}
2147ea8dc4b6Seschrock 
2148fa9e4066Sahrens 	if (dn->dn_phys->dn_nlevels == 0)
2149fa9e4066Sahrens 		nlevels = 1;
2150fa9e4066Sahrens 	else
2151fa9e4066Sahrens 		nlevels = dn->dn_phys->dn_nlevels;
2152fa9e4066Sahrens 
2153fa9e4066Sahrens 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2154fa9e4066Sahrens 
2155fa9e4066Sahrens 	ASSERT3U(level * epbs, <, 64);
2156fa9e4066Sahrens 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2157ea8dc4b6Seschrock 	if (level >= nlevels ||
2158fa9e4066Sahrens 	    (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
2159fa9e4066Sahrens 		/* the buffer has no parent yet */
2160be6fd75aSMatthew Ahrens 		return (SET_ERROR(ENOENT));
2161fa9e4066Sahrens 	} else if (level < nlevels-1) {
2162fa9e4066Sahrens 		/* this block is referenced from an indirect block */
2163fa9e4066Sahrens 		int err = dbuf_hold_impl(dn, level+1,
2164a2cdcdd2SPaul Dagnelie 		    blkid >> epbs, fail_sparse, FALSE, NULL, parentp);
2165fa9e4066Sahrens 		if (err)
2166fa9e4066Sahrens 			return (err);
2167ea8dc4b6Seschrock 		err = dbuf_read(*parentp, NULL,
2168ea8dc4b6Seschrock 		    (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
2169c543ec06Sahrens 		if (err) {
2170c543ec06Sahrens 			dbuf_rele(*parentp, NULL);
2171c543ec06Sahrens 			*parentp = NULL;
2172c543ec06Sahrens 			return (err);
2173ea8dc4b6Seschrock 		}
2174c543ec06Sahrens 		*bpp = ((blkptr_t *)(*parentp)->db.db_data) +
2175c543ec06Sahrens 		    (blkid & ((1ULL << epbs) - 1));
2176c543ec06Sahrens 		return (0);
2177fa9e4066Sahrens 	} else {
2178fa9e4066Sahrens 		/* the block is referenced from the dnode */
2179fa9e4066Sahrens 		ASSERT3U(level, ==, nlevels-1);
2180fa9e4066Sahrens 		ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
2181fa9e4066Sahrens 		    blkid < dn->dn_phys->dn_nblkptr);
2182c543ec06Sahrens 		if (dn->dn_dbuf) {
2183c543ec06Sahrens 			dbuf_add_ref(dn->dn_dbuf, NULL);
2184c543ec06Sahrens 			*parentp = dn->dn_dbuf;
2185c543ec06Sahrens 		}
2186fa9e4066Sahrens 		*bpp = &dn->dn_phys->dn_blkptr[blkid];
2187fa9e4066Sahrens 		return (0);
2188fa9e4066Sahrens 	}
2189fa9e4066Sahrens }
2190fa9e4066Sahrens 
2191fa9e4066Sahrens static dmu_buf_impl_t *
2192fa9e4066Sahrens dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
2193fa9e4066Sahrens     dmu_buf_impl_t *parent, blkptr_t *blkptr)
2194fa9e4066Sahrens {
2195503ad85cSMatthew Ahrens 	objset_t *os = dn->dn_objset;
2196fa9e4066Sahrens 	dmu_buf_impl_t *db, *odb;
2197fa9e4066Sahrens 
2198fa9e4066Sahrens 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2199fa9e4066Sahrens 	ASSERT(dn->dn_type != DMU_OT_NONE);
2200fa9e4066Sahrens 
2201*dcbf3bd6SGeorge Wilson 	db = kmem_cache_alloc(dbuf_kmem_cache, KM_SLEEP);
2202fa9e4066Sahrens 
2203fa9e4066Sahrens 	db->db_objset = os;
2204fa9e4066Sahrens 	db->db.db_object = dn->dn_object;
2205fa9e4066Sahrens 	db->db_level = level;
2206fa9e4066Sahrens 	db->db_blkid = blkid;
2207c717a561Smaybee 	db->db_last_dirty = NULL;
2208ea8dc4b6Seschrock 	db->db_dirtycnt = 0;
2209744947dcSTom Erickson 	db->db_dnode_handle = dn->dn_handle;
2210ea8dc4b6Seschrock 	db->db_parent = parent;
2211ea8dc4b6Seschrock 	db->db_blkptr = blkptr;
2212fa9e4066Sahrens 
2213bc9014e6SJustin Gibbs 	db->db_user = NULL;
2214d2058105SJustin T. Gibbs 	db->db_user_immediate_evict = FALSE;
2215d2058105SJustin T. Gibbs 	db->db_freed_in_flight = FALSE;
2216d2058105SJustin T. Gibbs 	db->db_pending_evict = FALSE;
2217ea8dc4b6Seschrock 
22180a586ceaSMark Shellenbaum 	if (blkid == DMU_BONUS_BLKID) {
2219ea8dc4b6Seschrock 		ASSERT3P(parent, ==, dn->dn_dbuf);
22201934e92fSmaybee 		db->db.db_size = DN_MAX_BONUSLEN -
22211934e92fSmaybee 		    (dn->dn_nblkptr-1) * sizeof (blkptr_t);
22221934e92fSmaybee 		ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
22230a586ceaSMark Shellenbaum 		db->db.db_offset = DMU_BONUS_BLKID;
2224ea8dc4b6Seschrock 		db->db_state = DB_UNCACHED;
2225ea8dc4b6Seschrock 		/* the bonus dbuf is not placed in the hash table */
22265a98e54bSBrendan Gregg - Sun Microsystems 		arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2227ea8dc4b6Seschrock 		return (db);
22280a586ceaSMark Shellenbaum 	} else if (blkid == DMU_SPILL_BLKID) {
22290a586ceaSMark Shellenbaum 		db->db.db_size = (blkptr != NULL) ?
22300a586ceaSMark Shellenbaum 		    BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
22310a586ceaSMark Shellenbaum 		db->db.db_offset = 0;
2232fa9e4066Sahrens 	} else {
2233fa9e4066Sahrens 		int blocksize =
223469962b56SMatthew Ahrens 		    db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
2235fa9e4066Sahrens 		db->db.db_size = blocksize;
2236fa9e4066Sahrens 		db->db.db_offset = db->db_blkid * blocksize;
2237fa9e4066Sahrens 	}
2238fa9e4066Sahrens 
2239fa9e4066Sahrens 	/*
2240fa9e4066Sahrens 	 * Hold the dn_dbufs_mtx while we get the new dbuf
2241fa9e4066Sahrens 	 * in the hash table *and* added to the dbufs list.
2242fa9e4066Sahrens 	 * This prevents a possible deadlock with someone
2243fa9e4066Sahrens 	 * trying to look up this dbuf before its added to the
2244fa9e4066Sahrens 	 * dn_dbufs list.
2245fa9e4066Sahrens 	 */
2246fa9e4066Sahrens 	mutex_enter(&dn->dn_dbufs_mtx);
2247ea8dc4b6Seschrock 	db->db_state = DB_EVICTING;
2248fa9e4066Sahrens 	if ((odb = dbuf_hash_insert(db)) != NULL) {
2249fa9e4066Sahrens 		/* someone else inserted it first */
2250*dcbf3bd6SGeorge Wilson 		kmem_cache_free(dbuf_kmem_cache, db);
2251fa9e4066Sahrens 		mutex_exit(&dn->dn_dbufs_mtx);
2252fa9e4066Sahrens 		return (odb);
2253fa9e4066Sahrens 	}
22540f6d88adSAlex Reece 	avl_add(&dn->dn_dbufs, db);
2255713d6c20SMatthew Ahrens 	if (db->db_level == 0 && db->db_blkid >=
2256713d6c20SMatthew Ahrens 	    dn->dn_unlisted_l0_blkid)
2257713d6c20SMatthew Ahrens 		dn->dn_unlisted_l0_blkid = db->db_blkid + 1;
2258ea8dc4b6Seschrock 	db->db_state = DB_UNCACHED;
2259fa9e4066Sahrens 	mutex_exit(&dn->dn_dbufs_mtx);
22605a98e54bSBrendan Gregg - Sun Microsystems 	arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2261fa9e4066Sahrens 
2262fa9e4066Sahrens 	if (parent && parent != dn->dn_dbuf)
2263fa9e4066Sahrens 		dbuf_add_ref(parent, db);
2264fa9e4066Sahrens 
2265ea8dc4b6Seschrock 	ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
2266ea8dc4b6Seschrock 	    refcount_count(&dn->dn_holds) > 0);
2267fa9e4066Sahrens 	(void) refcount_add(&dn->dn_holds, db);
2268640c1670SJosef 'Jeff' Sipek 	atomic_inc_32(&dn->dn_dbufs_count);
2269fa9e4066Sahrens 
2270fa9e4066Sahrens 	dprintf_dbuf(db, "db=%p\n", db);
2271fa9e4066Sahrens 
2272fa9e4066Sahrens 	return (db);
2273fa9e4066Sahrens }
2274fa9e4066Sahrens 
2275a2cdcdd2SPaul Dagnelie typedef struct dbuf_prefetch_arg {
2276a2cdcdd2SPaul Dagnelie 	spa_t *dpa_spa;	/* The spa to issue the prefetch in. */
2277a2cdcdd2SPaul Dagnelie 	zbookmark_phys_t dpa_zb; /* The target block to prefetch. */
2278a2cdcdd2SPaul Dagnelie 	int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */
2279a2cdcdd2SPaul Dagnelie 	int dpa_curlevel; /* The current level that we're reading */
2280*dcbf3bd6SGeorge Wilson 	dnode_t *dpa_dnode; /* The dnode associated with the prefetch */
2281a2cdcdd2SPaul Dagnelie 	zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */
2282a2cdcdd2SPaul Dagnelie 	zio_t *dpa_zio; /* The parent zio_t for all prefetches. */
2283a2cdcdd2SPaul Dagnelie 	arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */
2284a2cdcdd2SPaul Dagnelie } dbuf_prefetch_arg_t;
2285a2cdcdd2SPaul Dagnelie 
2286a2cdcdd2SPaul Dagnelie /*
2287a2cdcdd2SPaul Dagnelie  * Actually issue the prefetch read for the block given.
2288a2cdcdd2SPaul Dagnelie  */
2289a2cdcdd2SPaul Dagnelie static void
2290a2cdcdd2SPaul Dagnelie dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp)
2291a2cdcdd2SPaul Dagnelie {
2292a2cdcdd2SPaul Dagnelie 	if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
2293a2cdcdd2SPaul Dagnelie 		return;
2294a2cdcdd2SPaul Dagnelie 
2295a2cdcdd2SPaul Dagnelie 	arc_flags_t aflags =
2296a2cdcdd2SPaul Dagnelie 	    dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
2297a2cdcdd2SPaul Dagnelie 
2298a2cdcdd2SPaul Dagnelie 	ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2299a2cdcdd2SPaul Dagnelie 	ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level);
2300a2cdcdd2SPaul Dagnelie 	ASSERT(dpa->dpa_zio != NULL);
2301a2cdcdd2SPaul Dagnelie 	(void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL,
2302a2cdcdd2SPaul Dagnelie 	    dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2303a2cdcdd2SPaul Dagnelie 	    &aflags, &dpa->dpa_zb);
2304a2cdcdd2SPaul Dagnelie }
2305a2cdcdd2SPaul Dagnelie 
2306a2cdcdd2SPaul Dagnelie /*
2307a2cdcdd2SPaul Dagnelie  * Called when an indirect block above our prefetch target is read in.  This
2308a2cdcdd2SPaul Dagnelie  * will either read in the next indirect block down the tree or issue the actual
2309a2cdcdd2SPaul Dagnelie  * prefetch if the next block down is our target.
2310a2cdcdd2SPaul Dagnelie  */
2311a2cdcdd2SPaul Dagnelie static void
2312a2cdcdd2SPaul Dagnelie dbuf_prefetch_indirect_done(zio_t *zio, arc_buf_t *abuf, void *private)
2313a2cdcdd2SPaul Dagnelie {
2314a2cdcdd2SPaul Dagnelie 	dbuf_prefetch_arg_t *dpa = private;
2315a2cdcdd2SPaul Dagnelie 
2316a2cdcdd2SPaul Dagnelie 	ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel);
2317a2cdcdd2SPaul Dagnelie 	ASSERT3S(dpa->dpa_curlevel, >, 0);
2318*dcbf3bd6SGeorge Wilson 
2319*dcbf3bd6SGeorge Wilson 	/*
2320*dcbf3bd6SGeorge Wilson 	 * The dpa_dnode is only valid if we are called with a NULL
2321*dcbf3bd6SGeorge Wilson 	 * zio. This indicates that the arc_read() returned without
2322*dcbf3bd6SGeorge Wilson 	 * first calling zio_read() to issue a physical read. Once
2323*dcbf3bd6SGeorge Wilson 	 * a physical read is made the dpa_dnode must be invalidated
2324*dcbf3bd6SGeorge Wilson 	 * as the locks guarding it may have been dropped. If the
2325*dcbf3bd6SGeorge Wilson 	 * dpa_dnode is still valid, then we want to add it to the dbuf
2326*dcbf3bd6SGeorge Wilson 	 * cache. To do so, we must hold the dbuf associated with the block
2327*dcbf3bd6SGeorge Wilson 	 * we just prefetched, read its contents so that we associate it
2328*dcbf3bd6SGeorge Wilson 	 * with an arc_buf_t, and then release it.
2329*dcbf3bd6SGeorge Wilson 	 */
2330a2cdcdd2SPaul Dagnelie 	if (zio != NULL) {
2331a2cdcdd2SPaul Dagnelie 		ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel);
2332*dcbf3bd6SGeorge Wilson 		if (zio->io_flags & ZIO_FLAG_RAW) {
2333*dcbf3bd6SGeorge Wilson 			ASSERT3U(BP_GET_PSIZE(zio->io_bp), ==, zio->io_size);
2334*dcbf3bd6SGeorge Wilson 		} else {
2335*dcbf3bd6SGeorge Wilson 			ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size);
2336*dcbf3bd6SGeorge Wilson 		}
2337a2cdcdd2SPaul Dagnelie 		ASSERT3P(zio->io_spa, ==, dpa->dpa_spa);
2338*dcbf3bd6SGeorge Wilson 
2339*dcbf3bd6SGeorge Wilson 		dpa->dpa_dnode = NULL;
2340*dcbf3bd6SGeorge Wilson 	} else if (dpa->dpa_dnode != NULL) {
2341*dcbf3bd6SGeorge Wilson 		uint64_t curblkid = dpa->dpa_zb.zb_blkid >>
2342*dcbf3bd6SGeorge Wilson 		    (dpa->dpa_epbs * (dpa->dpa_curlevel -
2343*dcbf3bd6SGeorge Wilson 		    dpa->dpa_zb.zb_level));
2344*dcbf3bd6SGeorge Wilson 		dmu_buf_impl_t *db = dbuf_hold_level(dpa->dpa_dnode,
2345*dcbf3bd6SGeorge Wilson 		    dpa->dpa_curlevel, curblkid, FTAG);
2346*dcbf3bd6SGeorge Wilson 		(void) dbuf_read(db, NULL,
2347*dcbf3bd6SGeorge Wilson 		    DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_HAVESTRUCT);
2348*dcbf3bd6SGeorge Wilson 		dbuf_rele(db, FTAG);
2349a2cdcdd2SPaul Dagnelie 	}
2350a2cdcdd2SPaul Dagnelie 
2351a2cdcdd2SPaul Dagnelie 	dpa->dpa_curlevel--;
2352a2cdcdd2SPaul Dagnelie 
2353a2cdcdd2SPaul Dagnelie 	uint64_t nextblkid = dpa->dpa_zb.zb_blkid >>
2354a2cdcdd2SPaul Dagnelie 	    (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level));
2355a2cdcdd2SPaul Dagnelie 	blkptr_t *bp = ((blkptr_t *)abuf->b_data) +
2356a2cdcdd2SPaul Dagnelie 	    P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs);
2357a2cdcdd2SPaul Dagnelie 	if (BP_IS_HOLE(bp) || (zio != NULL && zio->io_error != 0)) {
2358a2cdcdd2SPaul Dagnelie 		kmem_free(dpa, sizeof (*dpa));
2359a2cdcdd2SPaul Dagnelie 	} else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) {
2360a2cdcdd2SPaul Dagnelie 		ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid);
2361a2cdcdd2SPaul Dagnelie 		dbuf_issue_final_prefetch(dpa, bp);
2362a2cdcdd2SPaul Dagnelie 		kmem_free(dpa, sizeof (*dpa));
2363a2cdcdd2SPaul Dagnelie 	} else {
2364a2cdcdd2SPaul Dagnelie 		arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2365a2cdcdd2SPaul Dagnelie 		zbookmark_phys_t zb;
2366a2cdcdd2SPaul Dagnelie 
2367a2cdcdd2SPaul Dagnelie 		ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2368a2cdcdd2SPaul Dagnelie 
2369a2cdcdd2SPaul Dagnelie 		SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset,
2370a2cdcdd2SPaul Dagnelie 		    dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid);
2371a2cdcdd2SPaul Dagnelie 
2372a2cdcdd2SPaul Dagnelie 		(void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2373a2cdcdd2SPaul Dagnelie 		    bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio,
2374a2cdcdd2SPaul Dagnelie 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2375a2cdcdd2SPaul Dagnelie 		    &iter_aflags, &zb);
2376a2cdcdd2SPaul Dagnelie 	}
2377*dcbf3bd6SGeorge Wilson 
2378*dcbf3bd6SGeorge Wilson 	arc_buf_destroy(abuf, private);
2379a2cdcdd2SPaul Dagnelie }
2380a2cdcdd2SPaul Dagnelie 
2381a2cdcdd2SPaul Dagnelie /*
2382a2cdcdd2SPaul Dagnelie  * Issue prefetch reads for the given block on the given level.  If the indirect
2383a2cdcdd2SPaul Dagnelie  * blocks above that block are not in memory, we will read them in
2384a2cdcdd2SPaul Dagnelie  * asynchronously.  As a result, this call never blocks waiting for a read to
2385a2cdcdd2SPaul Dagnelie  * complete.
2386a2cdcdd2SPaul Dagnelie  */
2387fa9e4066Sahrens void
2388a2cdcdd2SPaul Dagnelie dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio,
2389a2cdcdd2SPaul Dagnelie     arc_flags_t aflags)
2390fa9e4066Sahrens {
2391a2cdcdd2SPaul Dagnelie 	blkptr_t bp;
2392a2cdcdd2SPaul Dagnelie 	int epbs, nlevels, curlevel;
2393a2cdcdd2SPaul Dagnelie 	uint64_t curblkid;
2394fa9e4066Sahrens 
23950a586ceaSMark Shellenbaum 	ASSERT(blkid != DMU_BONUS_BLKID);
2396fa9e4066Sahrens 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2397fa9e4066Sahrens 
2398cf6106c8SMatthew Ahrens 	if (blkid > dn->dn_maxblkid)
2399cf6106c8SMatthew Ahrens 		return;
2400cf6106c8SMatthew Ahrens 
2401fa9e4066Sahrens 	if (dnode_block_freed(dn, blkid))
2402fa9e4066Sahrens 		return;
2403fa9e4066Sahrens 
2404a2cdcdd2SPaul Dagnelie 	/*
2405a2cdcdd2SPaul Dagnelie 	 * This dnode hasn't been written to disk yet, so there's nothing to
2406a2cdcdd2SPaul Dagnelie 	 * prefetch.
2407a2cdcdd2SPaul Dagnelie 	 */
2408a2cdcdd2SPaul Dagnelie 	nlevels = dn->dn_phys->dn_nlevels;
2409a2cdcdd2SPaul Dagnelie 	if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0)
2410a2cdcdd2SPaul Dagnelie 		return;
2411a2cdcdd2SPaul Dagnelie 
2412a2cdcdd2SPaul Dagnelie 	epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2413a2cdcdd2SPaul Dagnelie 	if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level))
2414a2cdcdd2SPaul Dagnelie 		return;
2415a2cdcdd2SPaul Dagnelie 
2416a2cdcdd2SPaul Dagnelie 	dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object,
2417a2cdcdd2SPaul Dagnelie 	    level, blkid);
2418a2cdcdd2SPaul Dagnelie 	if (db != NULL) {
2419a2cdcdd2SPaul Dagnelie 		mutex_exit(&db->db_mtx);
24209e9c486fSGeorge Wilson 		/*
2421a2cdcdd2SPaul Dagnelie 		 * This dbuf already exists.  It is either CACHED, or
2422a2cdcdd2SPaul Dagnelie 		 * (we assume) about to be read or filled.
24239e9c486fSGeorge Wilson 		 */
24249e9c486fSGeorge Wilson 		return;
2425fa9e4066Sahrens 	}
2426fa9e4066Sahrens 
2427a2cdcdd2SPaul Dagnelie 	/*
2428a2cdcdd2SPaul Dagnelie 	 * Find the closest ancestor (indirect block) of the target block
2429a2cdcdd2SPaul Dagnelie 	 * that is present in the cache.  In this indirect block, we will
2430a2cdcdd2SPaul Dagnelie 	 * find the bp that is at curlevel, curblkid.
2431a2cdcdd2SPaul Dagnelie 	 */
2432a2cdcdd2SPaul Dagnelie 	curlevel = level;
2433a2cdcdd2SPaul Dagnelie 	curblkid = blkid;
2434a2cdcdd2SPaul Dagnelie 	while (curlevel < nlevels - 1) {
2435a2cdcdd2SPaul Dagnelie 		int parent_level = curlevel + 1;
2436a2cdcdd2SPaul Dagnelie 		uint64_t parent_blkid = curblkid >> epbs;
2437a2cdcdd2SPaul Dagnelie 		dmu_buf_impl_t *db;
2438a2cdcdd2SPaul Dagnelie 
2439a2cdcdd2SPaul Dagnelie 		if (dbuf_hold_impl(dn, parent_level, parent_blkid,
2440a2cdcdd2SPaul Dagnelie 		    FALSE, TRUE, FTAG, &db) == 0) {
2441a2cdcdd2SPaul Dagnelie 			blkptr_t *bpp = db->db_buf->b_data;
2442a2cdcdd2SPaul Dagnelie 			bp = bpp[P2PHASE(curblkid, 1 << epbs)];
2443a2cdcdd2SPaul Dagnelie 			dbuf_rele(db, FTAG);
2444a2cdcdd2SPaul Dagnelie 			break;
2445a2cdcdd2SPaul Dagnelie 		}
2446b24ab676SJeff Bonwick 
2447a2cdcdd2SPaul Dagnelie 		curlevel = parent_level;
2448a2cdcdd2SPaul Dagnelie 		curblkid = parent_blkid;
2449a2cdcdd2SPaul Dagnelie 	}
2450ea8dc4b6Seschrock 
2451a2cdcdd2SPaul Dagnelie 	if (curlevel == nlevels - 1) {
2452a2cdcdd2SPaul Dagnelie 		/* No cached indirect blocks found. */
2453a2cdcdd2SPaul Dagnelie 		ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr);
2454a2cdcdd2SPaul Dagnelie 		bp = dn->dn_phys->dn_blkptr[curblkid];
2455fa9e4066Sahrens 	}
2456a2cdcdd2SPaul Dagnelie 	if (BP_IS_HOLE(&bp))
2457a2cdcdd2SPaul Dagnelie 		return;
2458a2cdcdd2SPaul Dagnelie 
2459a2cdcdd2SPaul Dagnelie 	ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp));
2460a2cdcdd2SPaul Dagnelie 
2461a2cdcdd2SPaul Dagnelie 	zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL,
2462a2cdcdd2SPaul Dagnelie 	    ZIO_FLAG_CANFAIL);
2463a2cdcdd2SPaul Dagnelie 
2464a2cdcdd2SPaul Dagnelie 	dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP);
2465a2cdcdd2SPaul Dagnelie 	dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
2466a2cdcdd2SPaul Dagnelie 	SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2467a2cdcdd2SPaul Dagnelie 	    dn->dn_object, level, blkid);
2468a2cdcdd2SPaul Dagnelie 	dpa->dpa_curlevel = curlevel;
2469a2cdcdd2SPaul Dagnelie 	dpa->dpa_prio = prio;
2470a2cdcdd2SPaul Dagnelie 	dpa->dpa_aflags = aflags;
2471a2cdcdd2SPaul Dagnelie 	dpa->dpa_spa = dn->dn_objset->os_spa;
2472*dcbf3bd6SGeorge Wilson 	dpa->dpa_dnode = dn;
2473a2cdcdd2SPaul Dagnelie 	dpa->dpa_epbs = epbs;
2474a2cdcdd2SPaul Dagnelie 	dpa->dpa_zio = pio;
2475a2cdcdd2SPaul Dagnelie 
2476a2cdcdd2SPaul Dagnelie 	/*
2477a2cdcdd2SPaul Dagnelie 	 * If we have the indirect just above us, no need to do the asynchronous
2478a2cdcdd2SPaul Dagnelie 	 * prefetch chain; we'll just run the last step ourselves.  If we're at
2479a2cdcdd2SPaul Dagnelie 	 * a higher level, though, we want to issue the prefetches for all the
2480a2cdcdd2SPaul Dagnelie 	 * indirect blocks asynchronously, so we can go on with whatever we were
2481a2cdcdd2SPaul Dagnelie 	 * doing.
2482a2cdcdd2SPaul Dagnelie 	 */
2483a2cdcdd2SPaul Dagnelie 	if (curlevel == level) {
2484a2cdcdd2SPaul Dagnelie 		ASSERT3U(curblkid, ==, blkid);
2485a2cdcdd2SPaul Dagnelie 		dbuf_issue_final_prefetch(dpa, &bp);
2486a2cdcdd2SPaul Dagnelie 		kmem_free(dpa, sizeof (*dpa));
2487a2cdcdd2SPaul Dagnelie 	} else {
2488a2cdcdd2SPaul Dagnelie 		arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2489a2cdcdd2SPaul Dagnelie 		zbookmark_phys_t zb;
2490a2cdcdd2SPaul Dagnelie 
2491a2cdcdd2SPaul Dagnelie 		SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2492a2cdcdd2SPaul Dagnelie 		    dn->dn_object, curlevel, curblkid);
2493a2cdcdd2SPaul Dagnelie 		(void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2494a2cdcdd2SPaul Dagnelie 		    &bp, dbuf_prefetch_indirect_done, dpa, prio,
2495a2cdcdd2SPaul Dagnelie 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2496a2cdcdd2SPaul Dagnelie 		    &iter_aflags, &zb);
2497a2cdcdd2SPaul Dagnelie 	}
2498a2cdcdd2SPaul Dagnelie 	/*
2499a2cdcdd2SPaul Dagnelie 	 * We use pio here instead of dpa_zio since it's possible that
2500a2cdcdd2SPaul Dagnelie 	 * dpa may have already been freed.
2501a2cdcdd2SPaul Dagnelie 	 */
2502a2cdcdd2SPaul Dagnelie 	zio_nowait(pio);
2503fa9e4066Sahrens }
2504fa9e4066Sahrens 
2505fa9e4066Sahrens /*
2506fa9e4066Sahrens  * Returns with db_holds incremented, and db_mtx not held.
2507fa9e4066Sahrens  * Note: dn_struct_rwlock must be held.
2508fa9e4066Sahrens  */
2509fa9e4066Sahrens int
2510a2cdcdd2SPaul Dagnelie dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid,
2511a2cdcdd2SPaul Dagnelie     boolean_t fail_sparse, boolean_t fail_uncached,
2512fa9e4066Sahrens     void *tag, dmu_buf_impl_t **dbp)
2513fa9e4066Sahrens {
2514fa9e4066Sahrens 	dmu_buf_impl_t *db, *parent = NULL;
2515fa9e4066Sahrens 
25160a586ceaSMark Shellenbaum 	ASSERT(blkid != DMU_BONUS_BLKID);
2517fa9e4066Sahrens 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2518fa9e4066Sahrens 	ASSERT3U(dn->dn_nlevels, >, level);
2519fa9e4066Sahrens 
2520fa9e4066Sahrens 	*dbp = NULL;
2521ea8dc4b6Seschrock top:
2522fa9e4066Sahrens 	/* dbuf_find() returns with db_mtx held */
2523e57a022bSJustin T. Gibbs 	db = dbuf_find(dn->dn_objset, dn->dn_object, level, blkid);
2524fa9e4066Sahrens 
2525fa9e4066Sahrens 	if (db == NULL) {
2526fa9e4066Sahrens 		blkptr_t *bp = NULL;
2527fa9e4066Sahrens 		int err;
2528fa9e4066Sahrens 
2529a2cdcdd2SPaul Dagnelie 		if (fail_uncached)
2530a2cdcdd2SPaul Dagnelie 			return (SET_ERROR(ENOENT));
2531a2cdcdd2SPaul Dagnelie 
2532c543ec06Sahrens 		ASSERT3P(parent, ==, NULL);
2533fa9e4066Sahrens 		err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
2534fa9e4066Sahrens 		if (fail_sparse) {
2535fa9e4066Sahrens 			if (err == 0 && bp && BP_IS_HOLE(bp))
2536be6fd75aSMatthew Ahrens 				err = SET_ERROR(ENOENT);
2537fa9e4066Sahrens 			if (err) {
2538c543ec06Sahrens 				if (parent)
2539ea8dc4b6Seschrock 					dbuf_rele(parent, NULL);
2540fa9e4066Sahrens 				return (err);
2541fa9e4066Sahrens 			}
2542fa9e4066Sahrens 		}
2543ea8dc4b6Seschrock 		if (err && err != ENOENT)
2544ea8dc4b6Seschrock 			return (err);
2545fa9e4066Sahrens 		db = dbuf_create(dn, level, blkid, parent, bp);
2546fa9e4066Sahrens 	}
2547fa9e4066Sahrens 
2548a2cdcdd2SPaul Dagnelie 	if (fail_uncached && db->db_state != DB_CACHED) {
2549a2cdcdd2SPaul Dagnelie 		mutex_exit(&db->db_mtx);
2550a2cdcdd2SPaul Dagnelie 		return (SET_ERROR(ENOENT));
2551a2cdcdd2SPaul Dagnelie 	}
2552a2cdcdd2SPaul Dagnelie 
2553*dcbf3bd6SGeorge Wilson 	if (db->db_buf != NULL)
2554ea8dc4b6Seschrock 		ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
2555ea8dc4b6Seschrock 
2556ea8dc4b6Seschrock 	ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
2557ea8dc4b6Seschrock 
2558fa9e4066Sahrens 	/*
2559c717a561Smaybee 	 * If this buffer is currently syncing out, and we are are
2560c717a561Smaybee 	 * still referencing it from db_data, we need to make a copy
2561c717a561Smaybee 	 * of it in case we decide we want to dirty it again in this txg.
2562fa9e4066Sahrens 	 */
25630a586ceaSMark Shellenbaum 	if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
2564ea8dc4b6Seschrock 	    dn->dn_object != DMU_META_DNODE_OBJECT &&
2565c717a561Smaybee 	    db->db_state == DB_CACHED && db->db_data_pending) {
2566c717a561Smaybee 		dbuf_dirty_record_t *dr = db->db_data_pending;
2567fa9e4066Sahrens 
2568c717a561Smaybee 		if (dr->dt.dl.dr_data == db->db_buf) {
2569c717a561Smaybee 			arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2570c717a561Smaybee 
2571c717a561Smaybee 			dbuf_set_data(db,
2572*dcbf3bd6SGeorge Wilson 			    arc_alloc_buf(dn->dn_objset->os_spa,
2573c717a561Smaybee 			    db->db.db_size, db, type));
2574c717a561Smaybee 			bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
2575c717a561Smaybee 			    db->db.db_size);
2576c717a561Smaybee 		}
2577fa9e4066Sahrens 	}
2578fa9e4066Sahrens 
2579*dcbf3bd6SGeorge Wilson 	if (multilist_link_active(&db->db_cache_link)) {
2580*dcbf3bd6SGeorge Wilson 		ASSERT(refcount_is_zero(&db->db_holds));
2581*dcbf3bd6SGeorge Wilson 		multilist_remove(&dbuf_cache, db);
2582*dcbf3bd6SGeorge Wilson 		(void) refcount_remove_many(&dbuf_cache_size,
2583*dcbf3bd6SGeorge Wilson 		    db->db.db_size, db);
2584*dcbf3bd6SGeorge Wilson 	}
2585ea8dc4b6Seschrock 	(void) refcount_add(&db->db_holds, tag);
25869c9dc39aSek 	DBUF_VERIFY(db);
2587fa9e4066Sahrens 	mutex_exit(&db->db_mtx);
2588fa9e4066Sahrens 
2589fa9e4066Sahrens 	/* NOTE: we can't rele the parent until after we drop the db_mtx */
2590c543ec06Sahrens 	if (parent)
2591ea8dc4b6Seschrock 		dbuf_rele(parent, NULL);
2592fa9e4066Sahrens 
2593744947dcSTom Erickson 	ASSERT3P(DB_DNODE(db), ==, dn);
2594fa9e4066Sahrens 	ASSERT3U(db->db_blkid, ==, blkid);
2595fa9e4066Sahrens 	ASSERT3U(db->db_level, ==, level);
2596fa9e4066Sahrens 	*dbp = db;
2597fa9e4066Sahrens 
2598fa9e4066Sahrens 	return (0);
2599fa9e4066Sahrens }
2600fa9e4066Sahrens 
2601fa9e4066Sahrens dmu_buf_impl_t *
2602ea8dc4b6Seschrock dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
2603fa9e4066Sahrens {
2604a2cdcdd2SPaul Dagnelie 	return (dbuf_hold_level(dn, 0, blkid, tag));
2605fa9e4066Sahrens }
2606fa9e4066Sahrens 
2607fa9e4066Sahrens dmu_buf_impl_t *
2608fa9e4066Sahrens dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
2609fa9e4066Sahrens {
2610fa9e4066Sahrens 	dmu_buf_impl_t *db;
2611a2cdcdd2SPaul Dagnelie 	int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db);
2612ea8dc4b6Seschrock 	return (err ? NULL : db);
2613fa9e4066Sahrens }
2614fa9e4066Sahrens 
26151934e92fSmaybee void
2616ea8dc4b6Seschrock dbuf_create_bonus(dnode_t *dn)
2617fa9e4066Sahrens {
2618ea8dc4b6Seschrock 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
2619ea8dc4b6Seschrock 
2620ea8dc4b6Seschrock 	ASSERT(dn->dn_bonus == NULL);
26210a586ceaSMark Shellenbaum 	dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
26220a586ceaSMark Shellenbaum }
26230a586ceaSMark Shellenbaum 
26240a586ceaSMark Shellenbaum int
26250a586ceaSMark Shellenbaum dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
26260a586ceaSMark Shellenbaum {
26270a586ceaSMark Shellenbaum 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2628744947dcSTom Erickson 	dnode_t *dn;
2629744947dcSTom Erickson 
26300a586ceaSMark Shellenbaum 	if (db->db_blkid != DMU_SPILL_BLKID)
2631be6fd75aSMatthew Ahrens 		return (SET_ERROR(ENOTSUP));
26320a586ceaSMark Shellenbaum 	if (blksz == 0)
26330a586ceaSMark Shellenbaum 		blksz = SPA_MINBLOCKSIZE;
2634b5152584SMatthew Ahrens 	ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset)));
2635b5152584SMatthew Ahrens 	blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
26360a586ceaSMark Shellenbaum 
2637744947dcSTom Erickson 	DB_DNODE_ENTER(db);
2638744947dcSTom Erickson 	dn = DB_DNODE(db);
2639744947dcSTom Erickson 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
26400a586ceaSMark Shellenbaum 	dbuf_new_size(db, blksz, tx);
2641744947dcSTom Erickson 	rw_exit(&dn->dn_struct_rwlock);
2642744947dcSTom Erickson 	DB_DNODE_EXIT(db);
26430a586ceaSMark Shellenbaum 
26440a586ceaSMark Shellenbaum 	return (0);
26450a586ceaSMark Shellenbaum }
26460a586ceaSMark Shellenbaum 
26470a586ceaSMark Shellenbaum void
26480a586ceaSMark Shellenbaum dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
26490a586ceaSMark Shellenbaum {
26500a586ceaSMark Shellenbaum 	dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
2651fa9e4066Sahrens }
2652fa9e4066Sahrens 
2653ea8dc4b6Seschrock #pragma weak dmu_buf_add_ref = dbuf_add_ref
2654fa9e4066Sahrens void
2655fa9e4066Sahrens dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
2656fa9e4066Sahrens {
2657ea8dc4b6Seschrock 	int64_t holds = refcount_add(&db->db_holds, tag);
2658*dcbf3bd6SGeorge Wilson 	ASSERT3S(holds, >, 1);
2659fa9e4066Sahrens }
2660fa9e4066Sahrens 
2661e57a022bSJustin T. Gibbs #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
2662e57a022bSJustin T. Gibbs boolean_t
2663e57a022bSJustin T. Gibbs dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid,
2664e57a022bSJustin T. Gibbs     void *tag)
2665e57a022bSJustin T. Gibbs {
2666e57a022bSJustin T. Gibbs 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2667e57a022bSJustin T. Gibbs 	dmu_buf_impl_t *found_db;
2668e57a022bSJustin T. Gibbs 	boolean_t result = B_FALSE;
2669e57a022bSJustin T. Gibbs 
2670e57a022bSJustin T. Gibbs 	if (db->db_blkid == DMU_BONUS_BLKID)
2671e57a022bSJustin T. Gibbs 		found_db = dbuf_find_bonus(os, obj);
2672e57a022bSJustin T. Gibbs 	else
2673e57a022bSJustin T. Gibbs 		found_db = dbuf_find(os, obj, 0, blkid);
2674e57a022bSJustin T. Gibbs 
2675e57a022bSJustin T. Gibbs 	if (found_db != NULL) {
2676e57a022bSJustin T. Gibbs 		if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) {
2677e57a022bSJustin T. Gibbs 			(void) refcount_add(&db->db_holds, tag);
2678e57a022bSJustin T. Gibbs 			result = B_TRUE;
2679e57a022bSJustin T. Gibbs 		}
2680e57a022bSJustin T. Gibbs 		mutex_exit(&db->db_mtx);
2681e57a022bSJustin T. Gibbs 	}
2682e57a022bSJustin T. Gibbs 	return (result);
2683e57a022bSJustin T. Gibbs }
2684e57a022bSJustin T. Gibbs 
2685744947dcSTom Erickson /*
2686744947dcSTom Erickson  * If you call dbuf_rele() you had better not be referencing the dnode handle
2687744947dcSTom Erickson  * unless you have some other direct or indirect hold on the dnode. (An indirect
2688744947dcSTom Erickson  * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2689744947dcSTom Erickson  * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2690744947dcSTom Erickson  * dnode's parent dbuf evicting its dnode handles.
2691744947dcSTom Erickson  */
2692fa9e4066Sahrens void
2693ea8dc4b6Seschrock dbuf_rele(dmu_buf_impl_t *db, void *tag)
2694b24ab676SJeff Bonwick {
2695b24ab676SJeff Bonwick 	mutex_enter(&db->db_mtx);
2696b24ab676SJeff Bonwick 	dbuf_rele_and_unlock(db, tag);
2697b24ab676SJeff Bonwick }
2698b24ab676SJeff Bonwick 
269943466aaeSMax Grossman void
270043466aaeSMax Grossman dmu_buf_rele(dmu_buf_t *db, void *tag)
270143466aaeSMax Grossman {
270243466aaeSMax Grossman 	dbuf_rele((dmu_buf_impl_t *)db, tag);
270343466aaeSMax Grossman }
270443466aaeSMax Grossman 
2705b24ab676SJeff Bonwick /*
2706b24ab676SJeff Bonwick  * dbuf_rele() for an already-locked dbuf.  This is necessary to allow
2707b24ab676SJeff Bonwick  * db_dirtycnt and db_holds to be updated atomically.
2708b24ab676SJeff Bonwick  */
2709b24ab676SJeff Bonwick void
2710b24ab676SJeff Bonwick dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2711fa9e4066Sahrens {
2712fa9e4066Sahrens 	int64_t holds;
2713fa9e4066Sahrens 
2714b24ab676SJeff Bonwick 	ASSERT(MUTEX_HELD(&db->db_mtx));
27159c9dc39aSek 	DBUF_VERIFY(db);
2716fa9e4066Sahrens 
2717744947dcSTom Erickson 	/*
2718744947dcSTom Erickson 	 * Remove the reference to the dbuf before removing its hold on the
2719744947dcSTom Erickson 	 * dnode so we can guarantee in dnode_move() that a referenced bonus
2720744947dcSTom Erickson 	 * buffer has a corresponding dnode hold.
2721744947dcSTom Erickson 	 */
2722fa9e4066Sahrens 	holds = refcount_remove(&db->db_holds, tag);
2723ea8dc4b6Seschrock 	ASSERT(holds >= 0);
2724ea8dc4b6Seschrock 
2725c717a561Smaybee 	/*
2726c717a561Smaybee 	 * We can't freeze indirects if there is a possibility that they
2727c717a561Smaybee 	 * may be modified in the current syncing context.
2728c717a561Smaybee 	 */
2729*dcbf3bd6SGeorge Wilson 	if (db->db_buf != NULL &&
2730*dcbf3bd6SGeorge Wilson 	    holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) {
27316b4acc8bSahrens 		arc_buf_freeze(db->db_buf);
2732*dcbf3bd6SGeorge Wilson 	}
27336b4acc8bSahrens 
2734ea8dc4b6Seschrock 	if (holds == db->db_dirtycnt &&
2735d2058105SJustin T. Gibbs 	    db->db_level == 0 && db->db_user_immediate_evict)
2736ea8dc4b6Seschrock 		dbuf_evict_user(db);
2737fa9e4066Sahrens 
2738fa9e4066Sahrens 	if (holds == 0) {
27390a586ceaSMark Shellenbaum 		if (db->db_blkid == DMU_BONUS_BLKID) {
2740cd485b49SJustin T. Gibbs 			dnode_t *dn;
2741d2058105SJustin T. Gibbs 			boolean_t evict_dbuf = db->db_pending_evict;
2742744947dcSTom Erickson 
2743744947dcSTom Erickson 			/*
2744cd485b49SJustin T. Gibbs 			 * If the dnode moves here, we cannot cross this
2745cd485b49SJustin T. Gibbs 			 * barrier until the move completes.
2746744947dcSTom Erickson 			 */
2747744947dcSTom Erickson 			DB_DNODE_ENTER(db);
2748cd485b49SJustin T. Gibbs 
2749cd485b49SJustin T. Gibbs 			dn = DB_DNODE(db);
2750cd485b49SJustin T. Gibbs 			atomic_dec_32(&dn->dn_dbufs_count);
2751cd485b49SJustin T. Gibbs 
2752cd485b49SJustin T. Gibbs 			/*
2753cd485b49SJustin T. Gibbs 			 * Decrementing the dbuf count means that the bonus
2754cd485b49SJustin T. Gibbs 			 * buffer's dnode hold is no longer discounted in
2755cd485b49SJustin T. Gibbs 			 * dnode_move(). The dnode cannot move until after
2756d2058105SJustin T. Gibbs 			 * the dnode_rele() below.
2757cd485b49SJustin T. Gibbs 			 */
2758744947dcSTom Erickson 			DB_DNODE_EXIT(db);
2759cd485b49SJustin T. Gibbs 
2760cd485b49SJustin T. Gibbs 			/*
2761cd485b49SJustin T. Gibbs 			 * Do not reference db after its lock is dropped.
2762cd485b49SJustin T. Gibbs 			 * Another thread may evict it.
2763cd485b49SJustin T. Gibbs 			 */
2764cd485b49SJustin T. Gibbs 			mutex_exit(&db->db_mtx);
2765cd485b49SJustin T. Gibbs 
2766d2058105SJustin T. Gibbs 			if (evict_dbuf)
2767cd485b49SJustin T. Gibbs 				dnode_evict_bonus(dn);
2768d2058105SJustin T. Gibbs 
2769d2058105SJustin T. Gibbs 			dnode_rele(dn, db);
2770ea8dc4b6Seschrock 		} else if (db->db_buf == NULL) {
2771ea8dc4b6Seschrock 			/*
2772ea8dc4b6Seschrock 			 * This is a special case: we never associated this
2773ea8dc4b6Seschrock 			 * dbuf with any data allocated from the ARC.
2774ea8dc4b6Seschrock 			 */
277582c9918fSTim Haley 			ASSERT(db->db_state == DB_UNCACHED ||
277682c9918fSTim Haley 			    db->db_state == DB_NOFILL);
2777*dcbf3bd6SGeorge Wilson 			dbuf_destroy(db);
27786b4acc8bSahrens 		} else if (arc_released(db->db_buf)) {
2779ea8dc4b6Seschrock 			/*
2780ea8dc4b6Seschrock 			 * This dbuf has anonymous data associated with it.
2781ea8dc4b6Seschrock 			 */
2782*dcbf3bd6SGeorge Wilson 			dbuf_destroy(db);
2783ea8dc4b6Seschrock 		} else {
2784*dcbf3bd6SGeorge Wilson 			boolean_t do_arc_evict = B_FALSE;
2785*dcbf3bd6SGeorge Wilson 			blkptr_t bp;
2786*dcbf3bd6SGeorge Wilson 			spa_t *spa = dmu_objset_spa(db->db_objset);
2787*dcbf3bd6SGeorge Wilson 
2788*dcbf3bd6SGeorge Wilson 			if (!DBUF_IS_CACHEABLE(db) &&
2789*dcbf3bd6SGeorge Wilson 			    db->db_blkptr != NULL &&
2790*dcbf3bd6SGeorge Wilson 			    !BP_IS_HOLE(db->db_blkptr) &&
2791*dcbf3bd6SGeorge Wilson 			    !BP_IS_EMBEDDED(db->db_blkptr)) {
2792*dcbf3bd6SGeorge Wilson 				do_arc_evict = B_TRUE;
2793*dcbf3bd6SGeorge Wilson 				bp = *db->db_blkptr;
2794*dcbf3bd6SGeorge Wilson 			}
27959253d63dSGeorge Wilson 
2796*dcbf3bd6SGeorge Wilson 			if (!DBUF_IS_CACHEABLE(db) ||
2797*dcbf3bd6SGeorge Wilson 			    db->db_pending_evict) {
2798*dcbf3bd6SGeorge Wilson 				dbuf_destroy(db);
2799*dcbf3bd6SGeorge Wilson 			} else if (!multilist_link_active(&db->db_cache_link)) {
2800*dcbf3bd6SGeorge Wilson 				multilist_insert(&dbuf_cache, db);
2801*dcbf3bd6SGeorge Wilson 				(void) refcount_add_many(&dbuf_cache_size,
2802*dcbf3bd6SGeorge Wilson 				    db->db.db_size, db);
28033baa08fcSek 				mutex_exit(&db->db_mtx);
2804*dcbf3bd6SGeorge Wilson 
2805*dcbf3bd6SGeorge Wilson 				dbuf_evict_notify();
2806bbfa8ea8SMatthew Ahrens 			}
2807*dcbf3bd6SGeorge Wilson 
2808*dcbf3bd6SGeorge Wilson 			if (do_arc_evict)
2809*dcbf3bd6SGeorge Wilson 				arc_freed(spa, &bp);
2810ea8dc4b6Seschrock 		}
2811fa9e4066Sahrens 	} else {
2812fa9e4066Sahrens 		mutex_exit(&db->db_mtx);
2813fa9e4066Sahrens 	}
2814*dcbf3bd6SGeorge Wilson 
2815fa9e4066Sahrens }
2816fa9e4066Sahrens 
2817fa9e4066Sahrens #pragma weak dmu_buf_refcount = dbuf_refcount
2818fa9e4066Sahrens uint64_t
2819fa9e4066Sahrens dbuf_refcount(dmu_buf_impl_t *db)
2820fa9e4066Sahrens {
2821fa9e4066Sahrens 	return (refcount_count(&db->db_holds));
2822fa9e4066Sahrens }
2823fa9e4066Sahrens 
2824fa9e4066Sahrens void *
2825bc9014e6SJustin Gibbs dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user,
2826bc9014e6SJustin Gibbs     dmu_buf_user_t *new_user)
2827fa9e4066Sahrens {
2828bc9014e6SJustin Gibbs 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2829bc9014e6SJustin Gibbs 
2830bc9014e6SJustin Gibbs 	mutex_enter(&db->db_mtx);
2831bc9014e6SJustin Gibbs 	dbuf_verify_user(db, DBVU_NOT_EVICTING);
2832bc9014e6SJustin Gibbs 	if (db->db_user == old_user)
2833bc9014e6SJustin Gibbs 		db->db_user = new_user;
2834bc9014e6SJustin Gibbs 	else
2835bc9014e6SJustin Gibbs 		old_user = db->db_user;
2836bc9014e6SJustin Gibbs 	dbuf_verify_user(db, DBVU_NOT_EVICTING);
2837bc9014e6SJustin Gibbs 	mutex_exit(&db->db_mtx);
2838bc9014e6SJustin Gibbs 
2839bc9014e6SJustin Gibbs 	return (old_user);
2840fa9e4066Sahrens }
2841fa9e4066Sahrens 
2842fa9e4066Sahrens void *
2843bc9014e6SJustin Gibbs dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2844fa9e4066Sahrens {
2845bc9014e6SJustin Gibbs 	return (dmu_buf_replace_user(db_fake, NULL, user));
2846fa9e4066Sahrens }
2847fa9e4066Sahrens 
2848fa9e4066Sahrens void *
2849bc9014e6SJustin Gibbs dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2850fa9e4066Sahrens {
2851fa9e4066Sahrens 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2852fa9e4066Sahrens 
2853d2058105SJustin T. Gibbs 	db->db_user_immediate_evict = TRUE;
2854bc9014e6SJustin Gibbs 	return (dmu_buf_set_user(db_fake, user));
2855bc9014e6SJustin Gibbs }
2856fa9e4066Sahrens 
2857bc9014e6SJustin Gibbs void *
2858bc9014e6SJustin Gibbs dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2859bc9014e6SJustin Gibbs {
2860bc9014e6SJustin Gibbs 	return (dmu_buf_replace_user(db_fake, user, NULL));
2861fa9e4066Sahrens }
2862fa9e4066Sahrens 
2863fa9e4066Sahrens void *
2864fa9e4066Sahrens dmu_buf_get_user(dmu_buf_t *db_fake)
2865fa9e4066Sahrens {
2866fa9e4066Sahrens 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2867fa9e4066Sahrens 
2868bc9014e6SJustin Gibbs 	dbuf_verify_user(db, DBVU_NOT_EVICTING);
2869bc9014e6SJustin Gibbs 	return (db->db_user);
2870bc9014e6SJustin Gibbs }
2871bc9014e6SJustin Gibbs 
2872bc9014e6SJustin Gibbs void
2873bc9014e6SJustin Gibbs dmu_buf_user_evict_wait()
2874bc9014e6SJustin Gibbs {
2875bc9014e6SJustin Gibbs 	taskq_wait(dbu_evict_taskq);
2876fa9e4066Sahrens }
2877fa9e4066Sahrens 
28783d692628SSanjeev Bagewadi boolean_t
28793d692628SSanjeev Bagewadi dmu_buf_freeable(dmu_buf_t *dbuf)
28803d692628SSanjeev Bagewadi {
28813d692628SSanjeev Bagewadi 	boolean_t res = B_FALSE;
28823d692628SSanjeev Bagewadi 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
28833d692628SSanjeev Bagewadi 
28843d692628SSanjeev Bagewadi 	if (db->db_blkptr)
28853d692628SSanjeev Bagewadi 		res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset,
2886c7cd2421SGeorge Wilson 		    db->db_blkptr, db->db_blkptr->blk_birth);
28873d692628SSanjeev Bagewadi 
28883d692628SSanjeev Bagewadi 	return (res);
28893d692628SSanjeev Bagewadi }
28903d692628SSanjeev Bagewadi 
289180901aeaSGeorge Wilson blkptr_t *
289280901aeaSGeorge Wilson dmu_buf_get_blkptr(dmu_buf_t *db)
289380901aeaSGeorge Wilson {
289480901aeaSGeorge Wilson 	dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
289580901aeaSGeorge Wilson 	return (dbi->db_blkptr);
289680901aeaSGeorge Wilson }
289780901aeaSGeorge Wilson 
2898c717a561Smaybee static void
2899c717a561Smaybee dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
2900fa9e4066Sahrens {
2901c717a561Smaybee 	/* ASSERT(dmu_tx_is_syncing(tx) */
2902c717a561Smaybee 	ASSERT(MUTEX_HELD(&db->db_mtx));
2903c717a561Smaybee 
2904c717a561Smaybee 	if (db->db_blkptr != NULL)
2905c717a561Smaybee 		return;
2906c717a561Smaybee 
29070a586ceaSMark Shellenbaum 	if (db->db_blkid == DMU_SPILL_BLKID) {
29080a586ceaSMark Shellenbaum 		db->db_blkptr = &dn->dn_phys->dn_spill;
29090a586ceaSMark Shellenbaum 		BP_ZERO(db->db_blkptr);
29100a586ceaSMark Shellenbaum 		return;
29110a586ceaSMark Shellenbaum 	}
2912c717a561Smaybee 	if (db->db_level == dn->dn_phys->dn_nlevels-1) {
2913c717a561Smaybee 		/*
2914c717a561Smaybee 		 * This buffer was allocated at a time when there was
2915c717a561Smaybee 		 * no available blkptrs from the dnode, or it was
2916c717a561Smaybee 		 * inappropriate to hook it in (i.e., nlevels mis-match).
2917c717a561Smaybee 		 */
2918c717a561Smaybee 		ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
2919c717a561Smaybee 		ASSERT(db->db_parent == NULL);
2920c717a561Smaybee 		db->db_parent = dn->dn_dbuf;
2921c717a561Smaybee 		db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
2922c717a561Smaybee 		DBUF_VERIFY(db);
2923c717a561Smaybee 	} else {
2924c717a561Smaybee 		dmu_buf_impl_t *parent = db->db_parent;
2925c717a561Smaybee 		int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2926c717a561Smaybee 
2927c717a561Smaybee 		ASSERT(dn->dn_phys->dn_nlevels > 1);
2928c717a561Smaybee 		if (parent == NULL) {
2929c717a561Smaybee 			mutex_exit(&db->db_mtx);
2930c717a561Smaybee 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
2931a2cdcdd2SPaul Dagnelie 			parent = dbuf_hold_level(dn, db->db_level + 1,
2932a2cdcdd2SPaul Dagnelie 			    db->db_blkid >> epbs, db);
2933c717a561Smaybee 			rw_exit(&dn->dn_struct_rwlock);
2934c717a561Smaybee 			mutex_enter(&db->db_mtx);
2935c717a561Smaybee 			db->db_parent = parent;
2936c717a561Smaybee 		}
2937c717a561Smaybee 		db->db_blkptr = (blkptr_t *)parent->db.db_data +
2938c717a561Smaybee 		    (db->db_blkid & ((1ULL << epbs) - 1));
2939c717a561Smaybee 		DBUF_VERIFY(db);
2940c717a561Smaybee 	}
2941c717a561Smaybee }
2942c717a561Smaybee 
2943c717a561Smaybee static void
2944c717a561Smaybee dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2945c717a561Smaybee {
2946c717a561Smaybee 	dmu_buf_impl_t *db = dr->dr_dbuf;
2947744947dcSTom Erickson 	dnode_t *dn;
2948c717a561Smaybee 	zio_t *zio;
2949c717a561Smaybee 
2950c717a561Smaybee 	ASSERT(dmu_tx_is_syncing(tx));
2951c717a561Smaybee 
2952c717a561Smaybee 	dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2953c717a561Smaybee 
2954c717a561Smaybee 	mutex_enter(&db->db_mtx);
2955c717a561Smaybee 
2956c717a561Smaybee 	ASSERT(db->db_level > 0);
2957c717a561Smaybee 	DBUF_VERIFY(db);
2958c717a561Smaybee 
29593e30c24aSWill Andrews 	/* Read the block if it hasn't been read yet. */
2960c717a561Smaybee 	if (db->db_buf == NULL) {
2961c717a561Smaybee 		mutex_exit(&db->db_mtx);
2962c717a561Smaybee 		(void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
2963c717a561Smaybee 		mutex_enter(&db->db_mtx);
2964c717a561Smaybee 	}
2965c717a561Smaybee 	ASSERT3U(db->db_state, ==, DB_CACHED);
2966c717a561Smaybee 	ASSERT(db->db_buf != NULL);
2967c717a561Smaybee 
2968744947dcSTom Erickson 	DB_DNODE_ENTER(db);
2969744947dcSTom Erickson 	dn = DB_DNODE(db);
29703e30c24aSWill Andrews 	/* Indirect block size must match what the dnode thinks it is. */
2971744947dcSTom Erickson 	ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2972c717a561Smaybee 	dbuf_check_blkptr(dn, db);
2973744947dcSTom Erickson 	DB_DNODE_EXIT(db);
2974c717a561Smaybee 
29753e30c24aSWill Andrews 	/* Provide the pending dirty record to child dbufs */
2976c717a561Smaybee 	db->db_data_pending = dr;
2977c717a561Smaybee 
2978af2c4821Smaybee 	mutex_exit(&db->db_mtx);
2979088f3894Sahrens 	dbuf_write(dr, db->db_buf, tx);
2980c717a561Smaybee 
2981c717a561Smaybee 	zio = dr->dr_zio;
2982c717a561Smaybee 	mutex_enter(&dr->dt.di.dr_mtx);
298346e1baa6SMatthew Ahrens 	dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx);
2984c717a561Smaybee 	ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2985c717a561Smaybee 	mutex_exit(&dr->dt.di.dr_mtx);
2986c717a561Smaybee 	zio_nowait(zio);
2987c717a561Smaybee }
2988c717a561Smaybee 
2989c717a561Smaybee static void
2990c717a561Smaybee dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2991c717a561Smaybee {
2992c717a561Smaybee 	arc_buf_t **datap = &dr->dt.dl.dr_data;
2993c717a561Smaybee 	dmu_buf_impl_t *db = dr->dr_dbuf;
2994744947dcSTom Erickson 	dnode_t *dn;
2995744947dcSTom Erickson 	objset_t *os;
2996c717a561Smaybee 	uint64_t txg = tx->tx_txg;
2997fa9e4066Sahrens 
2998fa9e4066Sahrens 	ASSERT(dmu_tx_is_syncing(tx));
2999fa9e4066Sahrens 
3000fa9e4066Sahrens 	dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
3001fa9e4066Sahrens 
3002fa9e4066Sahrens 	mutex_enter(&db->db_mtx);
3003fa9e4066Sahrens 	/*
3004fa9e4066Sahrens 	 * To be synced, we must be dirtied.  But we
3005fa9e4066Sahrens 	 * might have been freed after the dirty.
3006fa9e4066Sahrens 	 */
3007fa9e4066Sahrens 	if (db->db_state == DB_UNCACHED) {
3008fa9e4066Sahrens 		/* This buffer has been freed since it was dirtied */
3009fa9e4066Sahrens 		ASSERT(db->db.db_data == NULL);
3010fa9e4066Sahrens 	} else if (db->db_state == DB_FILL) {
3011fa9e4066Sahrens 		/* This buffer was freed and is now being re-filled */
3012c717a561Smaybee 		ASSERT(db->db.db_data != dr->dt.dl.dr_data);
3013fa9e4066Sahrens 	} else {
301482c9918fSTim Haley 		ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
3015fa9e4066Sahrens 	}
30169c9dc39aSek 	DBUF_VERIFY(db);
3017fa9e4066Sahrens 
3018744947dcSTom Erickson 	DB_DNODE_ENTER(db);
3019744947dcSTom Erickson 	dn = DB_DNODE(db);
3020744947dcSTom Erickson 
30210a586ceaSMark Shellenbaum 	if (db->db_blkid == DMU_SPILL_BLKID) {
30220a586ceaSMark Shellenbaum 		mutex_enter(&dn->dn_mtx);
30230a586ceaSMark Shellenbaum 		dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
30240a586ceaSMark Shellenbaum 		mutex_exit(&dn->dn_mtx);
30250a586ceaSMark Shellenbaum 	}
30260a586ceaSMark Shellenbaum 
3027fa9e4066Sahrens 	/*
3028c717a561Smaybee 	 * If this is a bonus buffer, simply copy the bonus data into the
3029c717a561Smaybee 	 * dnode.  It will be written out when the dnode is synced (and it
3030c717a561Smaybee 	 * will be synced, since it must have been dirty for dbuf_sync to
3031c717a561Smaybee 	 * be called).
3032fa9e4066Sahrens 	 */
30330a586ceaSMark Shellenbaum 	if (db->db_blkid == DMU_BONUS_BLKID) {
3034c717a561Smaybee 		dbuf_dirty_record_t **drp;
30351934e92fSmaybee 
3036ea8dc4b6Seschrock 		ASSERT(*datap != NULL);
3037fb09f5aaSMadhav Suresh 		ASSERT0(db->db_level);
3038ea8dc4b6Seschrock 		ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN);
3039ea8dc4b6Seschrock 		bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen);
3040744947dcSTom Erickson 		DB_DNODE_EXIT(db);
3041744947dcSTom Erickson 
30420e8c6158Smaybee 		if (*datap != db->db.db_data) {
3043ea8dc4b6Seschrock 			zio_buf_free(*datap, DN_MAX_BONUSLEN);
30445a98e54bSBrendan Gregg - Sun Microsystems 			arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
30450e8c6158Smaybee 		}
3046ea8dc4b6Seschrock 		db->db_data_pending = NULL;
3047c717a561Smaybee 		drp = &db->db_last_dirty;
3048c717a561Smaybee 		while (*drp != dr)
3049c717a561Smaybee 			drp = &(*drp)->dr_next;
305017f17c2dSbonwick 		ASSERT(dr->dr_next == NULL);
3051b24ab676SJeff Bonwick 		ASSERT(dr->dr_dbuf == db);
305217f17c2dSbonwick 		*drp = dr->dr_next;
3053c717a561Smaybee 		kmem_free(dr, sizeof (dbuf_dirty_record_t));
3054ea8dc4b6Seschrock 		ASSERT(db->db_dirtycnt > 0);
3055ea8dc4b6Seschrock 		db->db_dirtycnt -= 1;
3056b24ab676SJeff Bonwick 		dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
3057ea8dc4b6Seschrock 		return;
3058ea8dc4b6Seschrock 	}
3059ea8dc4b6Seschrock 
3060744947dcSTom Erickson 	os = dn->dn_objset;
3061744947dcSTom Erickson 
3062f82bfe17Sgw 	/*
3063f82bfe17Sgw 	 * This function may have dropped the db_mtx lock allowing a dmu_sync
3064f82bfe17Sgw 	 * operation to sneak in. As a result, we need to ensure that we
3065f82bfe17Sgw 	 * don't check the dr_override_state until we have returned from
3066f82bfe17Sgw 	 * dbuf_check_blkptr.
3067f82bfe17Sgw 	 */
3068f82bfe17Sgw 	dbuf_check_blkptr(dn, db);
3069f82bfe17Sgw 
3070c717a561Smaybee 	/*
3071744947dcSTom Erickson 	 * If this buffer is in the middle of an immediate write,
3072c717a561Smaybee 	 * wait for the synchronous IO to complete.
3073c717a561Smaybee 	 */
3074c717a561Smaybee 	while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
3075c717a561Smaybee 		ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
3076c717a561Smaybee 		cv_wait(&db->db_changed, &db->db_mtx);
3077c717a561Smaybee 		ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
3078c717a561Smaybee 	}
3079c5c6ffa0Smaybee 
3080ab69d62fSMatthew Ahrens 	if (db->db_state != DB_NOFILL &&
3081ab69d62fSMatthew Ahrens 	    dn->dn_object != DMU_META_DNODE_OBJECT &&
3082ab69d62fSMatthew Ahrens 	    refcount_count(&db->db_holds) > 1 &&
3083b24ab676SJeff Bonwick 	    dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
3084ab69d62fSMatthew Ahrens 	    *datap == db->db_buf) {
3085ab69d62fSMatthew Ahrens 		/*
3086ab69d62fSMatthew Ahrens 		 * If this buffer is currently "in use" (i.e., there
3087ab69d62fSMatthew Ahrens 		 * are active holds and db_data still references it),
3088ab69d62fSMatthew Ahrens 		 * then make a copy before we start the write so that
3089ab69d62fSMatthew Ahrens 		 * any modifications from the open txg will not leak
3090ab69d62fSMatthew Ahrens 		 * into this write.
3091ab69d62fSMatthew Ahrens 		 *
3092ab69d62fSMatthew Ahrens 		 * NOTE: this copy does not need to be made for
3093ab69d62fSMatthew Ahrens 		 * objects only modified in the syncing context (e.g.
3094ab69d62fSMatthew Ahrens 		 * DNONE_DNODE blocks).
3095ab69d62fSMatthew Ahrens 		 */
3096ab69d62fSMatthew Ahrens 		int blksz = arc_buf_size(*datap);
3097ab69d62fSMatthew Ahrens 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
3098*dcbf3bd6SGeorge Wilson 		*datap = arc_alloc_buf(os->os_spa, blksz, db, type);
3099ab69d62fSMatthew Ahrens 		bcopy(db->db.db_data, (*datap)->b_data, blksz);
310082c9918fSTim Haley 	}
3101c717a561Smaybee 	db->db_data_pending = dr;
3102fa9e4066Sahrens 
3103c717a561Smaybee 	mutex_exit(&db->db_mtx);
3104fa9e4066Sahrens 
3105088f3894Sahrens 	dbuf_write(dr, *datap, tx);
3106fa9e4066Sahrens 
3107c717a561Smaybee 	ASSERT(!list_link_active(&dr->dr_dirty_node));
3108744947dcSTom Erickson 	if (dn->dn_object == DMU_META_DNODE_OBJECT) {
3109c717a561Smaybee 		list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
3110744947dcSTom Erickson 		DB_DNODE_EXIT(db);
3111744947dcSTom Erickson 	} else {
3112744947dcSTom Erickson 		/*
3113744947dcSTom Erickson 		 * Although zio_nowait() does not "wait for an IO", it does
3114744947dcSTom Erickson 		 * initiate the IO. If this is an empty write it seems plausible
3115744947dcSTom Erickson 		 * that the IO could actually be completed before the nowait
3116744947dcSTom Erickson 		 * returns. We need to DB_DNODE_EXIT() first in case
3117744947dcSTom Erickson 		 * zio_nowait() invalidates the dbuf.
3118744947dcSTom Erickson 		 */
3119744947dcSTom Erickson 		DB_DNODE_EXIT(db);
3120c717a561Smaybee 		zio_nowait(dr->dr_zio);
3121744947dcSTom Erickson 	}
3122c717a561Smaybee }
312323b11526Smaybee 
3124c717a561Smaybee void
312546e1baa6SMatthew Ahrens dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx)
3126c717a561Smaybee {
3127c717a561Smaybee 	dbuf_dirty_record_t *dr;
3128c717a561Smaybee 
3129c717a561Smaybee 	while (dr = list_head(list)) {
3130c717a561Smaybee 		if (dr->dr_zio != NULL) {
3131c717a561Smaybee 			/*
3132c717a561Smaybee 			 * If we find an already initialized zio then we
3133c717a561Smaybee 			 * are processing the meta-dnode, and we have finished.
3134c717a561Smaybee 			 * The dbufs for all dnodes are put back on the list
3135c717a561Smaybee 			 * during processing, so that we can zio_wait()
3136c717a561Smaybee 			 * these IOs after initiating all child IOs.
3137c717a561Smaybee 			 */
3138c717a561Smaybee 			ASSERT3U(dr->dr_dbuf->db.db_object, ==,
3139c717a561Smaybee 			    DMU_META_DNODE_OBJECT);
3140c717a561Smaybee 			break;
314123b11526Smaybee 		}
314246e1baa6SMatthew Ahrens 		if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
314346e1baa6SMatthew Ahrens 		    dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
314446e1baa6SMatthew Ahrens 			VERIFY3U(dr->dr_dbuf->db_level, ==, level);
314546e1baa6SMatthew Ahrens 		}
3146c717a561Smaybee 		list_remove(list, dr);
3147c717a561Smaybee 		if (dr->dr_dbuf->db_level > 0)
3148c717a561Smaybee 			dbuf_sync_indirect(dr, tx);
3149c717a561Smaybee 		else
3150c717a561Smaybee 			dbuf_sync_leaf(dr, tx);
315123b11526Smaybee 	}
3152c717a561Smaybee }
315323b11526Smaybee 
3154fa9e4066Sahrens /* ARGSUSED */
3155fa9e4066Sahrens static void
3156c717a561Smaybee dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3157fa9e4066Sahrens {
3158fa9e4066Sahrens 	dmu_buf_impl_t *db = vdb;
3159744947dcSTom Erickson 	dnode_t *dn;
3160e14bb325SJeff Bonwick 	blkptr_t *bp = zio->io_bp;
3161c717a561Smaybee 	blkptr_t *bp_orig = &zio->io_bp_orig;
3162b24ab676SJeff Bonwick 	spa_t *spa = zio->io_spa;
3163b24ab676SJeff Bonwick 	int64_t delta;
3164fa9e4066Sahrens 	uint64_t fill = 0;
3165b24ab676SJeff Bonwick 	int i;
3166fa9e4066Sahrens 
316711ceac77SAlex Reece 	ASSERT3P(db->db_blkptr, !=, NULL);
316811ceac77SAlex Reece 	ASSERT3P(&db->db_data_pending->dr_bp_copy, ==, bp);
3169e14bb325SJeff Bonwick 
3170744947dcSTom Erickson 	DB_DNODE_ENTER(db);
3171744947dcSTom Erickson 	dn = DB_DNODE(db);
3172b24ab676SJeff Bonwick 	delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
3173b24ab676SJeff Bonwick 	dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
3174b24ab676SJeff Bonwick 	zio->io_prev_space_delta = delta;
3175fa9e4066Sahrens 
317643466aaeSMax Grossman 	if (bp->blk_birth != 0) {
317743466aaeSMax Grossman 		ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
317843466aaeSMax Grossman 		    BP_GET_TYPE(bp) == dn->dn_type) ||
317943466aaeSMax Grossman 		    (db->db_blkid == DMU_SPILL_BLKID &&
31805d7b4d43SMatthew Ahrens 		    BP_GET_TYPE(bp) == dn->dn_bonustype) ||
31815d7b4d43SMatthew Ahrens 		    BP_IS_EMBEDDED(bp));
318243466aaeSMax Grossman 		ASSERT(BP_GET_LEVEL(bp) == db->db_level);
3183c717a561Smaybee 	}
3184c5c6ffa0Smaybee 
3185c717a561Smaybee 	mutex_enter(&db->db_mtx);
3186fa9e4066Sahrens 
31870a586ceaSMark Shellenbaum #ifdef ZFS_DEBUG
31880a586ceaSMark Shellenbaum 	if (db->db_blkid == DMU_SPILL_BLKID) {
31890a586ceaSMark Shellenbaum 		ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
319011ceac77SAlex Reece 		ASSERT(!(BP_IS_HOLE(bp)) &&
31910a586ceaSMark Shellenbaum 		    db->db_blkptr == &dn->dn_phys->dn_spill);
31920a586ceaSMark Shellenbaum 	}
31930a586ceaSMark Shellenbaum #endif
31940a586ceaSMark Shellenbaum 
3195fa9e4066Sahrens 	if (db->db_level == 0) {
3196fa9e4066Sahrens 		mutex_enter(&dn->dn_mtx);
31970a586ceaSMark Shellenbaum 		if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
31980a586ceaSMark Shellenbaum 		    db->db_blkid != DMU_SPILL_BLKID)
3199fa9e4066Sahrens 			dn->dn_phys->dn_maxblkid = db->db_blkid;
3200fa9e4066Sahrens 		mutex_exit(&dn->dn_mtx);
3201fa9e4066Sahrens 
3202fa9e4066Sahrens 		if (dn->dn_type == DMU_OT_DNODE) {
3203fa9e4066Sahrens 			dnode_phys_t *dnp = db->db.db_data;
3204fa9e4066Sahrens 			for (i = db->db.db_size >> DNODE_SHIFT; i > 0;
3205fa9e4066Sahrens 			    i--, dnp++) {
3206fa9e4066Sahrens 				if (dnp->dn_type != DMU_OT_NONE)
3207fa9e4066Sahrens 					fill++;
3208fa9e4066Sahrens 			}
3209fa9e4066Sahrens 		} else {
321043466aaeSMax Grossman 			if (BP_IS_HOLE(bp)) {
321143466aaeSMax Grossman 				fill = 0;
321243466aaeSMax Grossman 			} else {
321343466aaeSMax Grossman 				fill = 1;
321443466aaeSMax Grossman 			}
3215fa9e4066Sahrens 		}
3216fa9e4066Sahrens 	} else {
3217e14bb325SJeff Bonwick 		blkptr_t *ibp = db->db.db_data;
3218fa9e4066Sahrens 		ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3219e14bb325SJeff Bonwick 		for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
3220e14bb325SJeff Bonwick 			if (BP_IS_HOLE(ibp))
3221fa9e4066Sahrens 				continue;
32225d7b4d43SMatthew Ahrens 			fill += BP_GET_FILL(ibp);
3223fa9e4066Sahrens 		}
3224fa9e4066Sahrens 	}
3225744947dcSTom Erickson 	DB_DNODE_EXIT(db);
3226fa9e4066Sahrens 
32275d7b4d43SMatthew Ahrens 	if (!BP_IS_EMBEDDED(bp))
32285d7b4d43SMatthew Ahrens 		bp->blk_fill = fill;
3229c717a561Smaybee 
3230c717a561Smaybee 	mutex_exit(&db->db_mtx);
323111ceac77SAlex Reece 
323211ceac77SAlex Reece 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
323311ceac77SAlex Reece 	*db->db_blkptr = *bp;
323411ceac77SAlex Reece 	rw_exit(&dn->dn_struct_rwlock);
3235c717a561Smaybee }
3236fa9e4066Sahrens 
32378df0bcf0SPaul Dagnelie /* ARGSUSED */
32388df0bcf0SPaul Dagnelie /*
32398df0bcf0SPaul Dagnelie  * This function gets called just prior to running through the compression
32408df0bcf0SPaul Dagnelie  * stage of the zio pipeline. If we're an indirect block comprised of only
32418df0bcf0SPaul Dagnelie  * holes, then we want this indirect to be compressed away to a hole. In
32428df0bcf0SPaul Dagnelie  * order to do that we must zero out any information about the holes that
32438df0bcf0SPaul Dagnelie  * this indirect points to prior to before we try to compress it.
32448df0bcf0SPaul Dagnelie  */
32458df0bcf0SPaul Dagnelie static void
32468df0bcf0SPaul Dagnelie dbuf_write_children_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
32478df0bcf0SPaul Dagnelie {
32488df0bcf0SPaul Dagnelie 	dmu_buf_impl_t *db = vdb;
32498df0bcf0SPaul Dagnelie 	dnode_t *dn;
32508df0bcf0SPaul Dagnelie 	blkptr_t *bp;
32518df0bcf0SPaul Dagnelie 	uint64_t i;
32528df0bcf0SPaul Dagnelie 	int epbs;
32538df0bcf0SPaul Dagnelie 
32548df0bcf0SPaul Dagnelie 	ASSERT3U(db->db_level, >, 0);
32558df0bcf0SPaul Dagnelie 	DB_DNODE_ENTER(db);
32568df0bcf0SPaul Dagnelie 	dn = DB_DNODE(db);
32578df0bcf0SPaul Dagnelie 	epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
32588df0bcf0SPaul Dagnelie 
32598df0bcf0SPaul Dagnelie 	/* Determine if all our children are holes */
32608df0bcf0SPaul Dagnelie 	for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++) {
32618df0bcf0SPaul Dagnelie 		if (!BP_IS_HOLE(bp))
32628df0bcf0SPaul Dagnelie 			break;
32638df0bcf0SPaul Dagnelie 	}
32648df0bcf0SPaul Dagnelie 
32658df0bcf0SPaul Dagnelie 	/*
32668df0bcf0SPaul Dagnelie 	 * If all the children are holes, then zero them all out so that
32678df0bcf0SPaul Dagnelie 	 * we may get compressed away.
32688df0bcf0SPaul Dagnelie 	 */
32698df0bcf0SPaul Dagnelie 	if (i == 1 << epbs) {
32708df0bcf0SPaul Dagnelie 		/* didn't find any non-holes */
32718df0bcf0SPaul Dagnelie 		bzero(db->db.db_data, db->db.db_size);
32728df0bcf0SPaul Dagnelie 	}
32738df0bcf0SPaul Dagnelie 	DB_DNODE_EXIT(db);
32748df0bcf0SPaul Dagnelie }
32758df0bcf0SPaul Dagnelie 
327669962b56SMatthew Ahrens /*
327769962b56SMatthew Ahrens  * The SPA will call this callback several times for each zio - once
327869962b56SMatthew Ahrens  * for every physical child i/o (zio->io_phys_children times).  This
327969962b56SMatthew Ahrens  * allows the DMU to monitor the progress of each logical i/o.  For example,
328069962b56SMatthew Ahrens  * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
328169962b56SMatthew Ahrens  * block.  There may be a long delay before all copies/fragments are completed,
328269962b56SMatthew Ahrens  * so this callback allows us to retire dirty space gradually, as the physical
328369962b56SMatthew Ahrens  * i/os complete.
328469962b56SMatthew Ahrens  */
328569962b56SMatthew Ahrens /* ARGSUSED */
328669962b56SMatthew Ahrens static void
328769962b56SMatthew Ahrens dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
328869962b56SMatthew Ahrens {
328969962b56SMatthew Ahrens 	dmu_buf_impl_t *db = arg;
329069962b56SMatthew Ahrens 	objset_t *os = db->db_objset;
329169962b56SMatthew Ahrens 	dsl_pool_t *dp = dmu_objset_pool(os);
329269962b56SMatthew Ahrens 	dbuf_dirty_record_t *dr;
329369962b56SMatthew Ahrens 	int delta = 0;
329469962b56SMatthew Ahrens 
329569962b56SMatthew Ahrens 	dr = db->db_data_pending;
329669962b56SMatthew Ahrens 	ASSERT3U(dr->dr_txg, ==, zio->io_txg);
329769962b56SMatthew Ahrens 
329869962b56SMatthew Ahrens 	/*
329969962b56SMatthew Ahrens 	 * The callback will be called io_phys_children times.  Retire one
330069962b56SMatthew Ahrens 	 * portion of our dirty space each time we are called.  Any rounding
330169962b56SMatthew Ahrens 	 * error will be cleaned up by dsl_pool_sync()'s call to
330269962b56SMatthew Ahrens 	 * dsl_pool_undirty_space().
330369962b56SMatthew Ahrens 	 */
330469962b56SMatthew Ahrens 	delta = dr->dr_accounted / zio->io_phys_children;
330569962b56SMatthew Ahrens 	dsl_pool_undirty_space(dp, delta, zio->io_txg);
330669962b56SMatthew Ahrens }
330769962b56SMatthew Ahrens 
3308c717a561Smaybee /* ARGSUSED */
3309c717a561Smaybee static void
3310c717a561Smaybee dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
3311c717a561Smaybee {
3312c717a561Smaybee 	dmu_buf_impl_t *db = vdb;
3313b24ab676SJeff Bonwick 	blkptr_t *bp_orig = &zio->io_bp_orig;
331443466aaeSMax Grossman 	blkptr_t *bp = db->db_blkptr;
331543466aaeSMax Grossman 	objset_t *os = db->db_objset;
331643466aaeSMax Grossman 	dmu_tx_t *tx = os->os_synctx;
3317c717a561Smaybee 	dbuf_dirty_record_t **drp, *dr;
3318c717a561Smaybee 
3319fb09f5aaSMadhav Suresh 	ASSERT0(zio->io_error);
3320b24ab676SJeff Bonwick 	ASSERT(db->db_blkptr == bp);
3321b24ab676SJeff Bonwick 
332280901aeaSGeorge Wilson 	/*
332380901aeaSGeorge Wilson 	 * For nopwrites and rewrites we ensure that the bp matches our
332480901aeaSGeorge Wilson 	 * original and bypass all the accounting.
332580901aeaSGeorge Wilson 	 */
332680901aeaSGeorge Wilson 	if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
3327b24ab676SJeff Bonwick 		ASSERT(BP_EQUAL(bp, bp_orig));
3328b24ab676SJeff Bonwick 	} else {
332943466aaeSMax Grossman 		dsl_dataset_t *ds = os->os_dsl_dataset;
3330b24ab676SJeff Bonwick 		(void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
3331b24ab676SJeff Bonwick 		dsl_dataset_block_born(ds, bp, tx);
3332b24ab676SJeff Bonwick 	}
3333c717a561Smaybee 
3334c717a561Smaybee 	mutex_enter(&db->db_mtx);
3335c717a561Smaybee 
3336b24ab676SJeff Bonwick 	DBUF_VERIFY(db);
3337b24ab676SJeff Bonwick 
3338c717a561Smaybee 	drp = &db->db_last_dirty;
333917f17c2dSbonwick 	while ((dr = *drp) != db->db_data_pending)
334017f17c2dSbonwick 		drp = &dr->dr_next;
334117f17c2dSbonwick 	ASSERT(!list_link_active(&dr->dr_dirty_node));
3342b24ab676SJeff Bonwick 	ASSERT(dr->dr_dbuf == db);
334317f17c2dSbonwick 	ASSERT(dr->dr_next == NULL);
334417f17c2dSbonwick 	*drp = dr->dr_next;
3345c717a561Smaybee 
33460a586ceaSMark Shellenbaum #ifdef ZFS_DEBUG
33470a586ceaSMark Shellenbaum 	if (db->db_blkid == DMU_SPILL_BLKID) {
3348744947dcSTom Erickson 		dnode_t *dn;
3349744947dcSTom Erickson 
3350744947dcSTom Erickson 		DB_DNODE_ENTER(db);
3351744947dcSTom Erickson 		dn = DB_DNODE(db);
33520a586ceaSMark Shellenbaum 		ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
33530a586ceaSMark Shellenbaum 		ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
33540a586ceaSMark Shellenbaum 		    db->db_blkptr == &dn->dn_phys->dn_spill);
3355744947dcSTom Erickson 		DB_DNODE_EXIT(db);
33560a586ceaSMark Shellenbaum 	}
33570a586ceaSMark Shellenbaum #endif
33580a586ceaSMark Shellenbaum 
3359c717a561Smaybee 	if (db->db_level == 0) {
33600a586ceaSMark Shellenbaum 		ASSERT(db->db_blkid != DMU_BONUS_BLKID);
3361c717a561Smaybee 		ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
336282c9918fSTim Haley 		if (db->db_state != DB_NOFILL) {
336382c9918fSTim Haley 			if (dr->dt.dl.dr_data != db->db_buf)
3364*dcbf3bd6SGeorge Wilson 				arc_buf_destroy(dr->dt.dl.dr_data, db);
336582c9918fSTim Haley 		}
3366c717a561Smaybee 	} else {
3367744947dcSTom Erickson 		dnode_t *dn;
3368744947dcSTom Erickson 
3369744947dcSTom Erickson 		DB_DNODE_ENTER(db);
3370744947dcSTom Erickson 		dn = DB_DNODE(db);
3371c717a561Smaybee 		ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
337243466aaeSMax Grossman 		ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
3373c717a561Smaybee 		if (!BP_IS_HOLE(db->db_blkptr)) {
3374c717a561Smaybee 			int epbs =
3375c717a561Smaybee 			    dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
337643466aaeSMax Grossman 			ASSERT3U(db->db_blkid, <=,
337743466aaeSMax Grossman 			    dn->dn_phys->dn_maxblkid >> (db->db_level * epbs));
3378c717a561Smaybee 			ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
3379c717a561Smaybee 			    db->db.db_size);
3380c717a561Smaybee 		}
3381744947dcSTom Erickson 		DB_DNODE_EXIT(db);
3382c25056deSgw 		mutex_destroy(&dr->dt.di.dr_mtx);
3383c25056deSgw 		list_destroy(&dr->dt.di.dr_children);
3384c717a561Smaybee 	}
3385c717a561Smaybee 	kmem_free(dr, sizeof (dbuf_dirty_record_t));
3386fa9e4066Sahrens 
3387fa9e4066Sahrens 	cv_broadcast(&db->db_changed);
3388fa9e4066Sahrens 	ASSERT(db->db_dirtycnt > 0);
3389fa9e4066Sahrens 	db->db_dirtycnt -= 1;
3390c717a561Smaybee 	db->db_data_pending = NULL;
339143466aaeSMax Grossman 	dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg);
3392b24ab676SJeff Bonwick }
3393b24ab676SJeff Bonwick 
3394b24ab676SJeff Bonwick static void
3395b24ab676SJeff Bonwick dbuf_write_nofill_ready(zio_t *zio)
3396b24ab676SJeff Bonwick {
3397b24ab676SJeff Bonwick 	dbuf_write_ready(zio, NULL, zio->io_private);
3398b24ab676SJeff Bonwick }
3399b24ab676SJeff Bonwick 
3400b24ab676SJeff Bonwick static void
3401b24ab676SJeff Bonwick dbuf_write_nofill_done(zio_t *zio)
3402b24ab676SJeff Bonwick {
3403b24ab676SJeff Bonwick 	dbuf_write_done(zio, NULL, zio->io_private);
3404b24ab676SJeff Bonwick }
3405b24ab676SJeff Bonwick 
3406b24ab676SJeff Bonwick static void
3407b24ab676SJeff Bonwick dbuf_write_override_ready(zio_t *zio)
3408b24ab676SJeff Bonwick {
3409b24ab676SJeff Bonwick 	dbuf_dirty_record_t *dr = zio->io_private;
3410b24ab676SJeff Bonwick 	dmu_buf_impl_t *db = dr->dr_dbuf;
3411b24ab676SJeff Bonwick 
3412b24ab676SJeff Bonwick 	dbuf_write_ready(zio, NULL, db);
3413b24ab676SJeff Bonwick }
3414b24ab676SJeff Bonwick 
3415b24ab676SJeff Bonwick static void
3416b24ab676SJeff Bonwick dbuf_write_override_done(zio_t *zio)
3417b24ab676SJeff Bonwick {
3418b24ab676SJeff Bonwick 	dbuf_dirty_record_t *dr = zio->io_private;
3419b24ab676SJeff Bonwick 	dmu_buf_impl_t *db = dr->dr_dbuf;
3420b24ab676SJeff Bonwick 	blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
3421b24ab676SJeff Bonwick 
3422b24ab676SJeff Bonwick 	mutex_enter(&db->db_mtx);
3423b24ab676SJeff Bonwick 	if (!BP_EQUAL(zio->io_bp, obp)) {
3424b24ab676SJeff Bonwick 		if (!BP_IS_HOLE(obp))
3425b24ab676SJeff Bonwick 			dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
3426b24ab676SJeff Bonwick 		arc_release(dr->dt.dl.dr_data, db);
3427b24ab676SJeff Bonwick 	}
3428fa9e4066Sahrens 	mutex_exit(&db->db_mtx);
3429fa9e4066Sahrens 
3430b24ab676SJeff Bonwick 	dbuf_write_done(zio, NULL, db);
3431b24ab676SJeff Bonwick }
3432b24ab676SJeff Bonwick 
34333e30c24aSWill Andrews /* Issue I/O to commit a dirty buffer to disk. */
3434b24ab676SJeff Bonwick static void
3435b24ab676SJeff Bonwick dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
3436b24ab676SJeff Bonwick {
3437b24ab676SJeff Bonwick 	dmu_buf_impl_t *db = dr->dr_dbuf;
3438744947dcSTom Erickson 	dnode_t *dn;
3439744947dcSTom Erickson 	objset_t *os;
3440b24ab676SJeff Bonwick 	dmu_buf_impl_t *parent = db->db_parent;
3441b24ab676SJeff Bonwick 	uint64_t txg = tx->tx_txg;
34427802d7bfSMatthew Ahrens 	zbookmark_phys_t zb;
3443b24ab676SJeff Bonwick 	zio_prop_t zp;
3444b24ab676SJeff Bonwick 	zio_t *zio;
34450a586ceaSMark Shellenbaum 	int wp_flag = 0;
3446b24ab676SJeff Bonwick 
344711ceac77SAlex Reece 	ASSERT(dmu_tx_is_syncing(tx));
344811ceac77SAlex Reece 
3449744947dcSTom Erickson 	DB_DNODE_ENTER(db);
3450744947dcSTom Erickson 	dn = DB_DNODE(db);
3451744947dcSTom Erickson 	os = dn->dn_objset;
3452744947dcSTom Erickson 
3453b24ab676SJeff Bonwick 	if (db->db_state != DB_NOFILL) {
3454b24ab676SJeff Bonwick 		if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
3455b24ab676SJeff Bonwick 			/*
3456b24ab676SJeff Bonwick 			 * Private object buffers are released here rather
3457b24ab676SJeff Bonwick 			 * than in dbuf_dirty() since they are only modified
3458b24ab676SJeff Bonwick 			 * in the syncing context and we don't want the
3459b24ab676SJeff Bonwick 			 * overhead of making multiple copies of the data.
3460b24ab676SJeff Bonwick 			 */
3461b24ab676SJeff Bonwick 			if (BP_IS_HOLE(db->db_blkptr)) {
3462b24ab676SJeff Bonwick 				arc_buf_thaw(data);
3463b24ab676SJeff Bonwick 			} else {
34643f9d6ad7SLin Ling 				dbuf_release_bp(db);
3465b24ab676SJeff Bonwick 			}
3466b24ab676SJeff Bonwick 		}
3467b24ab676SJeff Bonwick 	}
3468b24ab676SJeff Bonwick 
3469b24ab676SJeff Bonwick 	if (parent != dn->dn_dbuf) {
34703e30c24aSWill Andrews 		/* Our parent is an indirect block. */
34713e30c24aSWill Andrews 		/* We have a dirty parent that has been scheduled for write. */
3472b24ab676SJeff Bonwick 		ASSERT(parent && parent->db_data_pending);
34733e30c24aSWill Andrews 		/* Our parent's buffer is one level closer to the dnode. */
3474b24ab676SJeff Bonwick 		ASSERT(db->db_level == parent->db_level-1);
34753e30c24aSWill Andrews 		/*
34763e30c24aSWill Andrews 		 * We're about to modify our parent's db_data by modifying
34773e30c24aSWill Andrews 		 * our block pointer, so the parent must be released.
34783e30c24aSWill Andrews 		 */
3479b24ab676SJeff Bonwick 		ASSERT(arc_released(parent->db_buf));
3480b24ab676SJeff Bonwick 		zio = parent->db_data_pending->dr_zio;
3481b24ab676SJeff Bonwick 	} else {
34823e30c24aSWill Andrews 		/* Our parent is the dnode itself. */
34830a586ceaSMark Shellenbaum 		ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
34840a586ceaSMark Shellenbaum 		    db->db_blkid != DMU_SPILL_BLKID) ||
34850a586ceaSMark Shellenbaum 		    (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
34860a586ceaSMark Shellenbaum 		if (db->db_blkid != DMU_SPILL_BLKID)
34870a586ceaSMark Shellenbaum 			ASSERT3P(db->db_blkptr, ==,
34880a586ceaSMark Shellenbaum 			    &dn->dn_phys->dn_blkptr[db->db_blkid]);
3489b24ab676SJeff Bonwick 		zio = dn->dn_zio;
3490b24ab676SJeff Bonwick 	}
3491b24ab676SJeff Bonwick 
3492b24ab676SJeff Bonwick 	ASSERT(db->db_level == 0 || data == db->db_buf);
3493b24ab676SJeff Bonwick 	ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
3494b24ab676SJeff Bonwick 	ASSERT(zio);
3495fa9e4066Sahrens 
3496b24ab676SJeff Bonwick 	SET_BOOKMARK(&zb, os->os_dsl_dataset ?
3497b24ab676SJeff Bonwick 	    os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
3498b24ab676SJeff Bonwick 	    db->db.db_object, db->db_level, db->db_blkid);
3499b24ab676SJeff Bonwick 
35000a586ceaSMark Shellenbaum 	if (db->db_blkid == DMU_SPILL_BLKID)
35010a586ceaSMark Shellenbaum 		wp_flag = WP_SPILL;
35020a586ceaSMark Shellenbaum 	wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
35030a586ceaSMark Shellenbaum 
35040a586ceaSMark Shellenbaum 	dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
3505744947dcSTom Erickson 	DB_DNODE_EXIT(db);
3506b24ab676SJeff Bonwick 
350711ceac77SAlex Reece 	/*
350811ceac77SAlex Reece 	 * We copy the blkptr now (rather than when we instantiate the dirty
350911ceac77SAlex Reece 	 * record), because its value can change between open context and
351011ceac77SAlex Reece 	 * syncing context. We do not need to hold dn_struct_rwlock to read
351111ceac77SAlex Reece 	 * db_blkptr because we are in syncing context.
351211ceac77SAlex Reece 	 */
351311ceac77SAlex Reece 	dr->dr_bp_copy = *db->db_blkptr;
351411ceac77SAlex Reece 
35155d7b4d43SMatthew Ahrens 	if (db->db_level == 0 &&
35165d7b4d43SMatthew Ahrens 	    dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
35175d7b4d43SMatthew Ahrens 		/*
35185d7b4d43SMatthew Ahrens 		 * The BP for this block has been provided by open context
35195d7b4d43SMatthew Ahrens 		 * (by dmu_sync() or dmu_buf_write_embedded()).
35205d7b4d43SMatthew Ahrens 		 */
35215d7b4d43SMatthew Ahrens 		void *contents = (data != NULL) ? data->b_data : NULL;
35225d7b4d43SMatthew Ahrens 
3523b24ab676SJeff Bonwick 		dr->dr_zio = zio_write(zio, os->os_spa, txg,
352411ceac77SAlex Reece 		    &dr->dr_bp_copy, contents, db->db.db_size, &zp,
35258df0bcf0SPaul Dagnelie 		    dbuf_write_override_ready, NULL, NULL,
35268df0bcf0SPaul Dagnelie 		    dbuf_write_override_done,
352769962b56SMatthew Ahrens 		    dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
3528b24ab676SJeff Bonwick 		mutex_enter(&db->db_mtx);
3529b24ab676SJeff Bonwick 		dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
3530b24ab676SJeff Bonwick 		zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
353180901aeaSGeorge Wilson 		    dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
3532b24ab676SJeff Bonwick 		mutex_exit(&db->db_mtx);
3533b24ab676SJeff Bonwick 	} else if (db->db_state == DB_NOFILL) {
3534810e43b2SBill Pijewski 		ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF ||
3535810e43b2SBill Pijewski 		    zp.zp_checksum == ZIO_CHECKSUM_NOPARITY);
3536b24ab676SJeff Bonwick 		dr->dr_zio = zio_write(zio, os->os_spa, txg,
353711ceac77SAlex Reece 		    &dr->dr_bp_copy, NULL, db->db.db_size, &zp,
35388df0bcf0SPaul Dagnelie 		    dbuf_write_nofill_ready, NULL, NULL,
35398df0bcf0SPaul Dagnelie 		    dbuf_write_nofill_done, db,
3540b24ab676SJeff Bonwick 		    ZIO_PRIORITY_ASYNC_WRITE,
3541b24ab676SJeff Bonwick 		    ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
3542b24ab676SJeff Bonwick 	} else {
3543b24ab676SJeff Bonwick 		ASSERT(arc_released(data));
35448df0bcf0SPaul Dagnelie 
35458df0bcf0SPaul Dagnelie 		/*
35468df0bcf0SPaul Dagnelie 		 * For indirect blocks, we want to setup the children
35478df0bcf0SPaul Dagnelie 		 * ready callback so that we can properly handle an indirect
35488df0bcf0SPaul Dagnelie 		 * block that only contains holes.
35498df0bcf0SPaul Dagnelie 		 */
35508df0bcf0SPaul Dagnelie 		arc_done_func_t *children_ready_cb = NULL;
35518df0bcf0SPaul Dagnelie 		if (db->db_level != 0)
35528df0bcf0SPaul Dagnelie 			children_ready_cb = dbuf_write_children_ready;
35538df0bcf0SPaul Dagnelie 
3554b24ab676SJeff Bonwick 		dr->dr_zio = arc_write(zio, os->os_spa, txg,
355511ceac77SAlex Reece 		    &dr->dr_bp_copy, data, DBUF_IS_L2CACHEABLE(db),
3556*dcbf3bd6SGeorge Wilson 		    &zp, dbuf_write_ready, children_ready_cb,
355769962b56SMatthew Ahrens 		    dbuf_write_physdone, dbuf_write_done, db,
355869962b56SMatthew Ahrens 		    ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
3559b24ab676SJeff Bonwick 	}
3560fa9e4066Sahrens }
3561