xref: /illumos-gate/usr/src/uts/common/fs/zfs/dbuf.c (revision 5cabbc6b)
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.
2494c2d0ebSMatthew Ahrens  * Copyright (c) 2012, 2017 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>
48dcbf3bd6SGeorge Wilson #include <sys/callb.h>
49770499e1SDan Kimmel #include <sys/abd.h>
50*5cabbc6bSPrashanth Sreenivasa #include <sys/vdev.h>
51dcbf3bd6SGeorge Wilson 
52dcbf3bd6SGeorge Wilson uint_t zfs_dbuf_evict_key;
53fa9e4066Sahrens 
543b2aab18SMatthew Ahrens static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
55088f3894Sahrens static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
56fa9e4066Sahrens 
57bc9014e6SJustin Gibbs #ifndef __lint
58bc9014e6SJustin Gibbs extern inline void dmu_buf_init_user(dmu_buf_user_t *dbu,
5940510e8eSJosef 'Jeff' Sipek     dmu_buf_evict_func_t *evict_func_sync,
6040510e8eSJosef 'Jeff' Sipek     dmu_buf_evict_func_t *evict_func_async,
6140510e8eSJosef 'Jeff' Sipek     dmu_buf_t **clear_on_evict_dbufp);
62bc9014e6SJustin Gibbs #endif /* ! __lint */
63bc9014e6SJustin Gibbs 
64fa9e4066Sahrens /*
65fa9e4066Sahrens  * Global data structures and functions for the dbuf cache.
66fa9e4066Sahrens  */
67dcbf3bd6SGeorge Wilson static kmem_cache_t *dbuf_kmem_cache;
68bc9014e6SJustin Gibbs static taskq_t *dbu_evict_taskq;
69fa9e4066Sahrens 
70dcbf3bd6SGeorge Wilson static kthread_t *dbuf_cache_evict_thread;
71dcbf3bd6SGeorge Wilson static kmutex_t dbuf_evict_lock;
72dcbf3bd6SGeorge Wilson static kcondvar_t dbuf_evict_cv;
73dcbf3bd6SGeorge Wilson static boolean_t dbuf_evict_thread_exit;
74dcbf3bd6SGeorge Wilson 
75dcbf3bd6SGeorge Wilson /*
76dcbf3bd6SGeorge Wilson  * LRU cache of dbufs. The dbuf cache maintains a list of dbufs that
77dcbf3bd6SGeorge Wilson  * are not currently held but have been recently released. These dbufs
78dcbf3bd6SGeorge Wilson  * are not eligible for arc eviction until they are aged out of the cache.
79dcbf3bd6SGeorge Wilson  * Dbufs are added to the dbuf cache once the last hold is released. If a
80dcbf3bd6SGeorge Wilson  * dbuf is later accessed and still exists in the dbuf cache, then it will
81dcbf3bd6SGeorge Wilson  * be removed from the cache and later re-added to the head of the cache.
82dcbf3bd6SGeorge Wilson  * Dbufs that are aged out of the cache will be immediately destroyed and
83dcbf3bd6SGeorge Wilson  * become eligible for arc eviction.
84dcbf3bd6SGeorge Wilson  */
8594c2d0ebSMatthew Ahrens static multilist_t *dbuf_cache;
86dcbf3bd6SGeorge Wilson static refcount_t dbuf_cache_size;
87dcbf3bd6SGeorge Wilson uint64_t dbuf_cache_max_bytes = 100 * 1024 * 1024;
88dcbf3bd6SGeorge Wilson 
89dcbf3bd6SGeorge Wilson /* Cap the size of the dbuf cache to log2 fraction of arc size. */
90dcbf3bd6SGeorge Wilson int dbuf_cache_max_shift = 5;
91dcbf3bd6SGeorge Wilson 
92dcbf3bd6SGeorge Wilson /*
93dcbf3bd6SGeorge Wilson  * The dbuf cache uses a three-stage eviction policy:
94dcbf3bd6SGeorge Wilson  *	- A low water marker designates when the dbuf eviction thread
95dcbf3bd6SGeorge Wilson  *	should stop evicting from the dbuf cache.
96dcbf3bd6SGeorge Wilson  *	- When we reach the maximum size (aka mid water mark), we
97dcbf3bd6SGeorge Wilson  *	signal the eviction thread to run.
98dcbf3bd6SGeorge Wilson  *	- The high water mark indicates when the eviction thread
99dcbf3bd6SGeorge Wilson  *	is unable to keep up with the incoming load and eviction must
100dcbf3bd6SGeorge Wilson  *	happen in the context of the calling thread.
101dcbf3bd6SGeorge Wilson  *
102dcbf3bd6SGeorge Wilson  * The dbuf cache:
103dcbf3bd6SGeorge Wilson  *                                                 (max size)
104dcbf3bd6SGeorge Wilson  *                                      low water   mid water   hi water
105dcbf3bd6SGeorge Wilson  * +----------------------------------------+----------+----------+
106dcbf3bd6SGeorge Wilson  * |                                        |          |          |
107dcbf3bd6SGeorge Wilson  * |                                        |          |          |
108dcbf3bd6SGeorge Wilson  * |                                        |          |          |
109dcbf3bd6SGeorge Wilson  * |                                        |          |          |
110dcbf3bd6SGeorge Wilson  * +----------------------------------------+----------+----------+
111dcbf3bd6SGeorge Wilson  *                                        stop        signal     evict
112dcbf3bd6SGeorge Wilson  *                                      evicting     eviction   directly
113dcbf3bd6SGeorge Wilson  *                                                    thread
114dcbf3bd6SGeorge Wilson  *
115dcbf3bd6SGeorge Wilson  * The high and low water marks indicate the operating range for the eviction
116dcbf3bd6SGeorge Wilson  * thread. The low water mark is, by default, 90% of the total size of the
117dcbf3bd6SGeorge Wilson  * cache and the high water mark is at 110% (both of these percentages can be
118dcbf3bd6SGeorge Wilson  * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct,
119dcbf3bd6SGeorge Wilson  * respectively). The eviction thread will try to ensure that the cache remains
120dcbf3bd6SGeorge Wilson  * within this range by waking up every second and checking if the cache is
121dcbf3bd6SGeorge Wilson  * above the low water mark. The thread can also be woken up by callers adding
122dcbf3bd6SGeorge Wilson  * elements into the cache if the cache is larger than the mid water (i.e max
123dcbf3bd6SGeorge Wilson  * cache size). Once the eviction thread is woken up and eviction is required,
124dcbf3bd6SGeorge Wilson  * it will continue evicting buffers until it's able to reduce the cache size
125dcbf3bd6SGeorge Wilson  * to the low water mark. If the cache size continues to grow and hits the high
126dcbf3bd6SGeorge Wilson  * water mark, then callers adding elments to the cache will begin to evict
127dcbf3bd6SGeorge Wilson  * directly from the cache until the cache is no longer above the high water
128dcbf3bd6SGeorge Wilson  * mark.
129dcbf3bd6SGeorge Wilson  */
130dcbf3bd6SGeorge Wilson 
131dcbf3bd6SGeorge Wilson /*
132dcbf3bd6SGeorge Wilson  * The percentage above and below the maximum cache size.
133dcbf3bd6SGeorge Wilson  */
134dcbf3bd6SGeorge Wilson uint_t dbuf_cache_hiwater_pct = 10;
135dcbf3bd6SGeorge Wilson uint_t dbuf_cache_lowater_pct = 10;
136dcbf3bd6SGeorge Wilson 
137fa9e4066Sahrens /* ARGSUSED */
138fa9e4066Sahrens static int
139fa9e4066Sahrens dbuf_cons(void *vdb, void *unused, int kmflag)
140fa9e4066Sahrens {
141fa9e4066Sahrens 	dmu_buf_impl_t *db = vdb;
142fa9e4066Sahrens 	bzero(db, sizeof (dmu_buf_impl_t));
143fa9e4066Sahrens 
144fa9e4066Sahrens 	mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
145fa9e4066Sahrens 	cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
146dcbf3bd6SGeorge Wilson 	multilist_link_init(&db->db_cache_link);
147fa9e4066Sahrens 	refcount_create(&db->db_holds);
1480f6d88adSAlex Reece 
149fa9e4066Sahrens 	return (0);
150fa9e4066Sahrens }
151fa9e4066Sahrens 
152fa9e4066Sahrens /* ARGSUSED */
153fa9e4066Sahrens static void
154fa9e4066Sahrens dbuf_dest(void *vdb, void *unused)
155fa9e4066Sahrens {
156fa9e4066Sahrens 	dmu_buf_impl_t *db = vdb;
157fa9e4066Sahrens 	mutex_destroy(&db->db_mtx);
158fa9e4066Sahrens 	cv_destroy(&db->db_changed);
159dcbf3bd6SGeorge Wilson 	ASSERT(!multilist_link_active(&db->db_cache_link));
160fa9e4066Sahrens 	refcount_destroy(&db->db_holds);
161fa9e4066Sahrens }
162fa9e4066Sahrens 
163fa9e4066Sahrens /*
164fa9e4066Sahrens  * dbuf hash table routines
165fa9e4066Sahrens  */
166fa9e4066Sahrens static dbuf_hash_table_t dbuf_hash_table;
167fa9e4066Sahrens 
168fa9e4066Sahrens static uint64_t dbuf_hash_count;
169fa9e4066Sahrens 
170fa9e4066Sahrens static uint64_t
171fa9e4066Sahrens dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
172fa9e4066Sahrens {
173fa9e4066Sahrens 	uintptr_t osv = (uintptr_t)os;
174fa9e4066Sahrens 	uint64_t crc = -1ULL;
175fa9e4066Sahrens 
176fa9e4066Sahrens 	ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
177fa9e4066Sahrens 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF];
178fa9e4066Sahrens 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF];
179fa9e4066Sahrens 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF];
180fa9e4066Sahrens 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF];
181fa9e4066Sahrens 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF];
182fa9e4066Sahrens 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF];
183fa9e4066Sahrens 
184fa9e4066Sahrens 	crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16);
185fa9e4066Sahrens 
186fa9e4066Sahrens 	return (crc);
187fa9e4066Sahrens }
188fa9e4066Sahrens 
189fa9e4066Sahrens #define	DBUF_EQUAL(dbuf, os, obj, level, blkid)		\
190fa9e4066Sahrens 	((dbuf)->db.db_object == (obj) &&		\
191fa9e4066Sahrens 	(dbuf)->db_objset == (os) &&			\
192fa9e4066Sahrens 	(dbuf)->db_level == (level) &&			\
193fa9e4066Sahrens 	(dbuf)->db_blkid == (blkid))
194fa9e4066Sahrens 
195fa9e4066Sahrens dmu_buf_impl_t *
196e57a022bSJustin T. Gibbs dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid)
197fa9e4066Sahrens {
198fa9e4066Sahrens 	dbuf_hash_table_t *h = &dbuf_hash_table;
199dcbf3bd6SGeorge Wilson 	uint64_t hv = dbuf_hash(os, obj, level, blkid);
200fa9e4066Sahrens 	uint64_t idx = hv & h->hash_table_mask;
201fa9e4066Sahrens 	dmu_buf_impl_t *db;
202fa9e4066Sahrens 
203fa9e4066Sahrens 	mutex_enter(DBUF_HASH_MUTEX(h, idx));
204fa9e4066Sahrens 	for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
205fa9e4066Sahrens 		if (DBUF_EQUAL(db, os, obj, level, blkid)) {
206fa9e4066Sahrens 			mutex_enter(&db->db_mtx);
207ea8dc4b6Seschrock 			if (db->db_state != DB_EVICTING) {
208fa9e4066Sahrens 				mutex_exit(DBUF_HASH_MUTEX(h, idx));
209fa9e4066Sahrens 				return (db);
210fa9e4066Sahrens 			}
211fa9e4066Sahrens 			mutex_exit(&db->db_mtx);
212fa9e4066Sahrens 		}
213fa9e4066Sahrens 	}
214fa9e4066Sahrens 	mutex_exit(DBUF_HASH_MUTEX(h, idx));
215fa9e4066Sahrens 	return (NULL);
216fa9e4066Sahrens }
217fa9e4066Sahrens 
218e57a022bSJustin T. Gibbs static dmu_buf_impl_t *
219e57a022bSJustin T. Gibbs dbuf_find_bonus(objset_t *os, uint64_t object)
220e57a022bSJustin T. Gibbs {
221e57a022bSJustin T. Gibbs 	dnode_t *dn;
222e57a022bSJustin T. Gibbs 	dmu_buf_impl_t *db = NULL;
223e57a022bSJustin T. Gibbs 
224e57a022bSJustin T. Gibbs 	if (dnode_hold(os, object, FTAG, &dn) == 0) {
225e57a022bSJustin T. Gibbs 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
226e57a022bSJustin T. Gibbs 		if (dn->dn_bonus != NULL) {
227e57a022bSJustin T. Gibbs 			db = dn->dn_bonus;
228e57a022bSJustin T. Gibbs 			mutex_enter(&db->db_mtx);
229e57a022bSJustin T. Gibbs 		}
230e57a022bSJustin T. Gibbs 		rw_exit(&dn->dn_struct_rwlock);
231e57a022bSJustin T. Gibbs 		dnode_rele(dn, FTAG);
232e57a022bSJustin T. Gibbs 	}
233e57a022bSJustin T. Gibbs 	return (db);
234e57a022bSJustin T. Gibbs }
235e57a022bSJustin T. Gibbs 
236fa9e4066Sahrens /*
237fa9e4066Sahrens  * Insert an entry into the hash table.  If there is already an element
238fa9e4066Sahrens  * equal to elem in the hash table, then the already existing element
239fa9e4066Sahrens  * will be returned and the new element will not be inserted.
240fa9e4066Sahrens  * Otherwise returns NULL.
241fa9e4066Sahrens  */
242fa9e4066Sahrens static dmu_buf_impl_t *
243fa9e4066Sahrens dbuf_hash_insert(dmu_buf_impl_t *db)
244fa9e4066Sahrens {
245fa9e4066Sahrens 	dbuf_hash_table_t *h = &dbuf_hash_table;
246503ad85cSMatthew Ahrens 	objset_t *os = db->db_objset;
247fa9e4066Sahrens 	uint64_t obj = db->db.db_object;
248fa9e4066Sahrens 	int level = db->db_level;
249fa9e4066Sahrens 	uint64_t blkid = db->db_blkid;
250dcbf3bd6SGeorge Wilson 	uint64_t hv = dbuf_hash(os, obj, level, blkid);
251fa9e4066Sahrens 	uint64_t idx = hv & h->hash_table_mask;
252fa9e4066Sahrens 	dmu_buf_impl_t *dbf;
253fa9e4066Sahrens 
254fa9e4066Sahrens 	mutex_enter(DBUF_HASH_MUTEX(h, idx));
255fa9e4066Sahrens 	for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) {
256fa9e4066Sahrens 		if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
257fa9e4066Sahrens 			mutex_enter(&dbf->db_mtx);
258ea8dc4b6Seschrock 			if (dbf->db_state != DB_EVICTING) {
259fa9e4066Sahrens 				mutex_exit(DBUF_HASH_MUTEX(h, idx));
260fa9e4066Sahrens 				return (dbf);
261fa9e4066Sahrens 			}
262fa9e4066Sahrens 			mutex_exit(&dbf->db_mtx);
263fa9e4066Sahrens 		}
264fa9e4066Sahrens 	}
265fa9e4066Sahrens 
266fa9e4066Sahrens 	mutex_enter(&db->db_mtx);
267fa9e4066Sahrens 	db->db_hash_next = h->hash_table[idx];
268fa9e4066Sahrens 	h->hash_table[idx] = db;
269fa9e4066Sahrens 	mutex_exit(DBUF_HASH_MUTEX(h, idx));
2701a5e258fSJosef 'Jeff' Sipek 	atomic_inc_64(&dbuf_hash_count);
271fa9e4066Sahrens 
272fa9e4066Sahrens 	return (NULL);
273fa9e4066Sahrens }
274fa9e4066Sahrens 
275fa9e4066Sahrens /*
276bbfa8ea8SMatthew Ahrens  * Remove an entry from the hash table.  It must be in the EVICTING state.
277fa9e4066Sahrens  */
278fa9e4066Sahrens static void
279fa9e4066Sahrens dbuf_hash_remove(dmu_buf_impl_t *db)
280fa9e4066Sahrens {
281fa9e4066Sahrens 	dbuf_hash_table_t *h = &dbuf_hash_table;
282dcbf3bd6SGeorge Wilson 	uint64_t hv = dbuf_hash(db->db_objset, db->db.db_object,
283fa9e4066Sahrens 	    db->db_level, db->db_blkid);
284fa9e4066Sahrens 	uint64_t idx = hv & h->hash_table_mask;
285fa9e4066Sahrens 	dmu_buf_impl_t *dbf, **dbp;
286fa9e4066Sahrens 
287fa9e4066Sahrens 	/*
288bbfa8ea8SMatthew Ahrens 	 * We musn't hold db_mtx to maintain lock ordering:
289fa9e4066Sahrens 	 * DBUF_HASH_MUTEX > db_mtx.
290fa9e4066Sahrens 	 */
291fa9e4066Sahrens 	ASSERT(refcount_is_zero(&db->db_holds));
292ea8dc4b6Seschrock 	ASSERT(db->db_state == DB_EVICTING);
293fa9e4066Sahrens 	ASSERT(!MUTEX_HELD(&db->db_mtx));
294fa9e4066Sahrens 
295fa9e4066Sahrens 	mutex_enter(DBUF_HASH_MUTEX(h, idx));
296fa9e4066Sahrens 	dbp = &h->hash_table[idx];
297fa9e4066Sahrens 	while ((dbf = *dbp) != db) {
298fa9e4066Sahrens 		dbp = &dbf->db_hash_next;
299fa9e4066Sahrens 		ASSERT(dbf != NULL);
300fa9e4066Sahrens 	}
301fa9e4066Sahrens 	*dbp = db->db_hash_next;
302fa9e4066Sahrens 	db->db_hash_next = NULL;
303fa9e4066Sahrens 	mutex_exit(DBUF_HASH_MUTEX(h, idx));
3041a5e258fSJosef 'Jeff' Sipek 	atomic_dec_64(&dbuf_hash_count);
305fa9e4066Sahrens }
306fa9e4066Sahrens 
307bc9014e6SJustin Gibbs typedef enum {
308bc9014e6SJustin Gibbs 	DBVU_EVICTING,
309bc9014e6SJustin Gibbs 	DBVU_NOT_EVICTING
310bc9014e6SJustin Gibbs } dbvu_verify_type_t;
311bc9014e6SJustin Gibbs 
312bc9014e6SJustin Gibbs static void
313bc9014e6SJustin Gibbs dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type)
314bc9014e6SJustin Gibbs {
315bc9014e6SJustin Gibbs #ifdef ZFS_DEBUG
316bc9014e6SJustin Gibbs 	int64_t holds;
317bc9014e6SJustin Gibbs 
318bc9014e6SJustin Gibbs 	if (db->db_user == NULL)
319bc9014e6SJustin Gibbs 		return;
320bc9014e6SJustin Gibbs 
321bc9014e6SJustin Gibbs 	/* Only data blocks support the attachment of user data. */
322bc9014e6SJustin Gibbs 	ASSERT(db->db_level == 0);
323bc9014e6SJustin Gibbs 
324bc9014e6SJustin Gibbs 	/* Clients must resolve a dbuf before attaching user data. */
325bc9014e6SJustin Gibbs 	ASSERT(db->db.db_data != NULL);
326bc9014e6SJustin Gibbs 	ASSERT3U(db->db_state, ==, DB_CACHED);
327bc9014e6SJustin Gibbs 
328bc9014e6SJustin Gibbs 	holds = refcount_count(&db->db_holds);
329bc9014e6SJustin Gibbs 	if (verify_type == DBVU_EVICTING) {
330bc9014e6SJustin Gibbs 		/*
331bc9014e6SJustin Gibbs 		 * Immediate eviction occurs when holds == dirtycnt.
332bc9014e6SJustin Gibbs 		 * For normal eviction buffers, holds is zero on
333bc9014e6SJustin Gibbs 		 * eviction, except when dbuf_fix_old_data() calls
334bc9014e6SJustin Gibbs 		 * dbuf_clear_data().  However, the hold count can grow
335bc9014e6SJustin Gibbs 		 * during eviction even though db_mtx is held (see
336bc9014e6SJustin Gibbs 		 * dmu_bonus_hold() for an example), so we can only
337bc9014e6SJustin Gibbs 		 * test the generic invariant that holds >= dirtycnt.
338bc9014e6SJustin Gibbs 		 */
339bc9014e6SJustin Gibbs 		ASSERT3U(holds, >=, db->db_dirtycnt);
340bc9014e6SJustin Gibbs 	} else {
341d2058105SJustin T. Gibbs 		if (db->db_user_immediate_evict == TRUE)
342bc9014e6SJustin Gibbs 			ASSERT3U(holds, >=, db->db_dirtycnt);
343bc9014e6SJustin Gibbs 		else
344bc9014e6SJustin Gibbs 			ASSERT3U(holds, >, 0);
345bc9014e6SJustin Gibbs 	}
346bc9014e6SJustin Gibbs #endif
347bc9014e6SJustin Gibbs }
348bc9014e6SJustin Gibbs 
349fa9e4066Sahrens static void
350fa9e4066Sahrens dbuf_evict_user(dmu_buf_impl_t *db)
351fa9e4066Sahrens {
352bc9014e6SJustin Gibbs 	dmu_buf_user_t *dbu = db->db_user;
353bc9014e6SJustin Gibbs 
354fa9e4066Sahrens 	ASSERT(MUTEX_HELD(&db->db_mtx));
355fa9e4066Sahrens 
356bc9014e6SJustin Gibbs 	if (dbu == NULL)
357fa9e4066Sahrens 		return;
358fa9e4066Sahrens 
359bc9014e6SJustin Gibbs 	dbuf_verify_user(db, DBVU_EVICTING);
360bc9014e6SJustin Gibbs 	db->db_user = NULL;
361bc9014e6SJustin Gibbs 
362bc9014e6SJustin Gibbs #ifdef ZFS_DEBUG
363bc9014e6SJustin Gibbs 	if (dbu->dbu_clear_on_evict_dbufp != NULL)
364bc9014e6SJustin Gibbs 		*dbu->dbu_clear_on_evict_dbufp = NULL;
365bc9014e6SJustin Gibbs #endif
366bc9014e6SJustin Gibbs 
367bc9014e6SJustin Gibbs 	/*
36840510e8eSJosef 'Jeff' Sipek 	 * There are two eviction callbacks - one that we call synchronously
36940510e8eSJosef 'Jeff' Sipek 	 * and one that we invoke via a taskq.  The async one is useful for
37040510e8eSJosef 'Jeff' Sipek 	 * avoiding lock order reversals and limiting stack depth.
37140510e8eSJosef 'Jeff' Sipek 	 *
37240510e8eSJosef 'Jeff' Sipek 	 * Note that if we have a sync callback but no async callback,
37340510e8eSJosef 'Jeff' Sipek 	 * it's likely that the sync callback will free the structure
37440510e8eSJosef 'Jeff' Sipek 	 * containing the dbu.  In that case we need to take care to not
37540510e8eSJosef 'Jeff' Sipek 	 * dereference dbu after calling the sync evict func.
376bc9014e6SJustin Gibbs 	 */
37740510e8eSJosef 'Jeff' Sipek 	boolean_t has_async = (dbu->dbu_evict_func_async != NULL);
37840510e8eSJosef 'Jeff' Sipek 
37940510e8eSJosef 'Jeff' Sipek 	if (dbu->dbu_evict_func_sync != NULL)
38040510e8eSJosef 'Jeff' Sipek 		dbu->dbu_evict_func_sync(dbu);
38140510e8eSJosef 'Jeff' Sipek 
38240510e8eSJosef 'Jeff' Sipek 	if (has_async) {
38340510e8eSJosef 'Jeff' Sipek 		taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func_async,
38440510e8eSJosef 'Jeff' Sipek 		    dbu, 0, &dbu->dbu_tqent);
38540510e8eSJosef 'Jeff' Sipek 	}
386fa9e4066Sahrens }
387fa9e4066Sahrens 
388744947dcSTom Erickson boolean_t
389744947dcSTom Erickson dbuf_is_metadata(dmu_buf_impl_t *db)
390744947dcSTom Erickson {
391744947dcSTom Erickson 	if (db->db_level > 0) {
392744947dcSTom Erickson 		return (B_TRUE);
393744947dcSTom Erickson 	} else {
394744947dcSTom Erickson 		boolean_t is_metadata;
395744947dcSTom Erickson 
396744947dcSTom Erickson 		DB_DNODE_ENTER(db);
397ad135b5dSChristopher Siden 		is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type);
398744947dcSTom Erickson 		DB_DNODE_EXIT(db);
399744947dcSTom Erickson 
400744947dcSTom Erickson 		return (is_metadata);
401744947dcSTom Erickson 	}
402744947dcSTom Erickson }
403744947dcSTom Erickson 
404dcbf3bd6SGeorge Wilson /*
405dcbf3bd6SGeorge Wilson  * This function *must* return indices evenly distributed between all
406dcbf3bd6SGeorge Wilson  * sublists of the multilist. This is needed due to how the dbuf eviction
407dcbf3bd6SGeorge Wilson  * code is laid out; dbuf_evict_thread() assumes dbufs are evenly
408dcbf3bd6SGeorge Wilson  * distributed between all sublists and uses this assumption when
409dcbf3bd6SGeorge Wilson  * deciding which sublist to evict from and how much to evict from it.
410dcbf3bd6SGeorge Wilson  */
411dcbf3bd6SGeorge Wilson unsigned int
412dcbf3bd6SGeorge Wilson dbuf_cache_multilist_index_func(multilist_t *ml, void *obj)
413ea8dc4b6Seschrock {
414dcbf3bd6SGeorge Wilson 	dmu_buf_impl_t *db = obj;
415dcbf3bd6SGeorge Wilson 
416dcbf3bd6SGeorge Wilson 	/*
417dcbf3bd6SGeorge Wilson 	 * The assumption here, is the hash value for a given
418dcbf3bd6SGeorge Wilson 	 * dmu_buf_impl_t will remain constant throughout it's lifetime
419dcbf3bd6SGeorge Wilson 	 * (i.e. it's objset, object, level and blkid fields don't change).
420dcbf3bd6SGeorge Wilson 	 * Thus, we don't need to store the dbuf's sublist index
421dcbf3bd6SGeorge Wilson 	 * on insertion, as this index can be recalculated on removal.
422dcbf3bd6SGeorge Wilson 	 *
423dcbf3bd6SGeorge Wilson 	 * Also, the low order bits of the hash value are thought to be
424dcbf3bd6SGeorge Wilson 	 * distributed evenly. Otherwise, in the case that the multilist
425dcbf3bd6SGeorge Wilson 	 * has a power of two number of sublists, each sublists' usage
426dcbf3bd6SGeorge Wilson 	 * would not be evenly distributed.
427dcbf3bd6SGeorge Wilson 	 */
428dcbf3bd6SGeorge Wilson 	return (dbuf_hash(db->db_objset, db->db.db_object,
429dcbf3bd6SGeorge Wilson 	    db->db_level, db->db_blkid) %
430dcbf3bd6SGeorge Wilson 	    multilist_get_num_sublists(ml));
431dcbf3bd6SGeorge Wilson }
432dcbf3bd6SGeorge Wilson 
433dcbf3bd6SGeorge Wilson static inline boolean_t
434dcbf3bd6SGeorge Wilson dbuf_cache_above_hiwater(void)
435dcbf3bd6SGeorge Wilson {
436dcbf3bd6SGeorge Wilson 	uint64_t dbuf_cache_hiwater_bytes =
437dcbf3bd6SGeorge Wilson 	    (dbuf_cache_max_bytes * dbuf_cache_hiwater_pct) / 100;
438dcbf3bd6SGeorge Wilson 
439dcbf3bd6SGeorge Wilson 	return (refcount_count(&dbuf_cache_size) >
440dcbf3bd6SGeorge Wilson 	    dbuf_cache_max_bytes + dbuf_cache_hiwater_bytes);
441dcbf3bd6SGeorge Wilson }
442dcbf3bd6SGeorge Wilson 
443dcbf3bd6SGeorge Wilson static inline boolean_t
444dcbf3bd6SGeorge Wilson dbuf_cache_above_lowater(void)
445dcbf3bd6SGeorge Wilson {
446dcbf3bd6SGeorge Wilson 	uint64_t dbuf_cache_lowater_bytes =
447dcbf3bd6SGeorge Wilson 	    (dbuf_cache_max_bytes * dbuf_cache_lowater_pct) / 100;
448dcbf3bd6SGeorge Wilson 
449dcbf3bd6SGeorge Wilson 	return (refcount_count(&dbuf_cache_size) >
450dcbf3bd6SGeorge Wilson 	    dbuf_cache_max_bytes - dbuf_cache_lowater_bytes);
451dcbf3bd6SGeorge Wilson }
452dcbf3bd6SGeorge Wilson 
453dcbf3bd6SGeorge Wilson /*
454dcbf3bd6SGeorge Wilson  * Evict the oldest eligible dbuf from the dbuf cache.
455dcbf3bd6SGeorge Wilson  */
456dcbf3bd6SGeorge Wilson static void
457dcbf3bd6SGeorge Wilson dbuf_evict_one(void)
458dcbf3bd6SGeorge Wilson {
45994c2d0ebSMatthew Ahrens 	int idx = multilist_get_random_index(dbuf_cache);
46094c2d0ebSMatthew Ahrens 	multilist_sublist_t *mls = multilist_sublist_lock(dbuf_cache, idx);
461dcbf3bd6SGeorge Wilson 
462dcbf3bd6SGeorge Wilson 	ASSERT(!MUTEX_HELD(&dbuf_evict_lock));
463dcbf3bd6SGeorge Wilson 
464dcbf3bd6SGeorge Wilson 	/*
465dcbf3bd6SGeorge Wilson 	 * Set the thread's tsd to indicate that it's processing evictions.
466dcbf3bd6SGeorge Wilson 	 * Once a thread stops evicting from the dbuf cache it will
467dcbf3bd6SGeorge Wilson 	 * reset its tsd to NULL.
468dcbf3bd6SGeorge Wilson 	 */
469dcbf3bd6SGeorge Wilson 	ASSERT3P(tsd_get(zfs_dbuf_evict_key), ==, NULL);
470dcbf3bd6SGeorge Wilson 	(void) tsd_set(zfs_dbuf_evict_key, (void *)B_TRUE);
471dcbf3bd6SGeorge Wilson 
472dcbf3bd6SGeorge Wilson 	dmu_buf_impl_t *db = multilist_sublist_tail(mls);
473dcbf3bd6SGeorge Wilson 	while (db != NULL && mutex_tryenter(&db->db_mtx) == 0) {
474dcbf3bd6SGeorge Wilson 		db = multilist_sublist_prev(mls, db);
475dcbf3bd6SGeorge Wilson 	}
476dcbf3bd6SGeorge Wilson 
477dcbf3bd6SGeorge Wilson 	DTRACE_PROBE2(dbuf__evict__one, dmu_buf_impl_t *, db,
478dcbf3bd6SGeorge Wilson 	    multilist_sublist_t *, mls);
479dcbf3bd6SGeorge Wilson 
480dcbf3bd6SGeorge Wilson 	if (db != NULL) {
481dcbf3bd6SGeorge Wilson 		multilist_sublist_remove(mls, db);
482dcbf3bd6SGeorge Wilson 		multilist_sublist_unlock(mls);
483dcbf3bd6SGeorge Wilson 		(void) refcount_remove_many(&dbuf_cache_size,
484dcbf3bd6SGeorge Wilson 		    db->db.db_size, db);
485dcbf3bd6SGeorge Wilson 		dbuf_destroy(db);
486dcbf3bd6SGeorge Wilson 	} else {
487dcbf3bd6SGeorge Wilson 		multilist_sublist_unlock(mls);
488dcbf3bd6SGeorge Wilson 	}
489dcbf3bd6SGeorge Wilson 	(void) tsd_set(zfs_dbuf_evict_key, NULL);
490dcbf3bd6SGeorge Wilson }
491dcbf3bd6SGeorge Wilson 
492dcbf3bd6SGeorge Wilson /*
493dcbf3bd6SGeorge Wilson  * The dbuf evict thread is responsible for aging out dbufs from the
494dcbf3bd6SGeorge Wilson  * cache. Once the cache has reached it's maximum size, dbufs are removed
495dcbf3bd6SGeorge Wilson  * and destroyed. The eviction thread will continue running until the size
496dcbf3bd6SGeorge Wilson  * of the dbuf cache is at or below the maximum size. Once the dbuf is aged
497dcbf3bd6SGeorge Wilson  * out of the cache it is destroyed and becomes eligible for arc eviction.
498dcbf3bd6SGeorge Wilson  */
4993f7978d0SAlan Somers /* ARGSUSED */
500dcbf3bd6SGeorge Wilson static void
5013f7978d0SAlan Somers dbuf_evict_thread(void *unused)
502dcbf3bd6SGeorge Wilson {
503dcbf3bd6SGeorge Wilson 	callb_cpr_t cpr;
504dcbf3bd6SGeorge Wilson 
505dcbf3bd6SGeorge Wilson 	CALLB_CPR_INIT(&cpr, &dbuf_evict_lock, callb_generic_cpr, FTAG);
506dcbf3bd6SGeorge Wilson 
507dcbf3bd6SGeorge Wilson 	mutex_enter(&dbuf_evict_lock);
508dcbf3bd6SGeorge Wilson 	while (!dbuf_evict_thread_exit) {
509dcbf3bd6SGeorge Wilson 		while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
510dcbf3bd6SGeorge Wilson 			CALLB_CPR_SAFE_BEGIN(&cpr);
511dcbf3bd6SGeorge Wilson 			(void) cv_timedwait_hires(&dbuf_evict_cv,
512dcbf3bd6SGeorge Wilson 			    &dbuf_evict_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
513dcbf3bd6SGeorge Wilson 			CALLB_CPR_SAFE_END(&cpr, &dbuf_evict_lock);
514dcbf3bd6SGeorge Wilson 		}
515dcbf3bd6SGeorge Wilson 		mutex_exit(&dbuf_evict_lock);
516dcbf3bd6SGeorge Wilson 
517dcbf3bd6SGeorge Wilson 		/*
518dcbf3bd6SGeorge Wilson 		 * Keep evicting as long as we're above the low water mark
519dcbf3bd6SGeorge Wilson 		 * for the cache. We do this without holding the locks to
520dcbf3bd6SGeorge Wilson 		 * minimize lock contention.
521dcbf3bd6SGeorge Wilson 		 */
522dcbf3bd6SGeorge Wilson 		while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
523dcbf3bd6SGeorge Wilson 			dbuf_evict_one();
524dcbf3bd6SGeorge Wilson 		}
525dcbf3bd6SGeorge Wilson 
526dcbf3bd6SGeorge Wilson 		mutex_enter(&dbuf_evict_lock);
527dcbf3bd6SGeorge Wilson 	}
528ea8dc4b6Seschrock 
529dcbf3bd6SGeorge Wilson 	dbuf_evict_thread_exit = B_FALSE;
530dcbf3bd6SGeorge Wilson 	cv_broadcast(&dbuf_evict_cv);
531dcbf3bd6SGeorge Wilson 	CALLB_CPR_EXIT(&cpr);	/* drops dbuf_evict_lock */
532dcbf3bd6SGeorge Wilson 	thread_exit();
533dcbf3bd6SGeorge Wilson }
534dcbf3bd6SGeorge Wilson 
535dcbf3bd6SGeorge Wilson /*
536dcbf3bd6SGeorge Wilson  * Wake up the dbuf eviction thread if the dbuf cache is at its max size.
537dcbf3bd6SGeorge Wilson  * If the dbuf cache is at its high water mark, then evict a dbuf from the
538dcbf3bd6SGeorge Wilson  * dbuf cache using the callers context.
539dcbf3bd6SGeorge Wilson  */
540dcbf3bd6SGeorge Wilson static void
541dcbf3bd6SGeorge Wilson dbuf_evict_notify(void)
542dcbf3bd6SGeorge Wilson {
543dcbf3bd6SGeorge Wilson 
544dcbf3bd6SGeorge Wilson 	/*
545dcbf3bd6SGeorge Wilson 	 * We use thread specific data to track when a thread has
546dcbf3bd6SGeorge Wilson 	 * started processing evictions. This allows us to avoid deeply
547dcbf3bd6SGeorge Wilson 	 * nested stacks that would have a call flow similar to this:
548dcbf3bd6SGeorge Wilson 	 *
549dcbf3bd6SGeorge Wilson 	 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify()
550dcbf3bd6SGeorge Wilson 	 *	^						|
551dcbf3bd6SGeorge Wilson 	 *	|						|
552dcbf3bd6SGeorge Wilson 	 *	+-----dbuf_destroy()<--dbuf_evict_one()<--------+
553dcbf3bd6SGeorge Wilson 	 *
554dcbf3bd6SGeorge Wilson 	 * The dbuf_eviction_thread will always have its tsd set until
555dcbf3bd6SGeorge Wilson 	 * that thread exits. All other threads will only set their tsd
556dcbf3bd6SGeorge Wilson 	 * if they are participating in the eviction process. This only
557dcbf3bd6SGeorge Wilson 	 * happens if the eviction thread is unable to process evictions
558dcbf3bd6SGeorge Wilson 	 * fast enough. To keep the dbuf cache size in check, other threads
559dcbf3bd6SGeorge Wilson 	 * can evict from the dbuf cache directly. Those threads will set
560dcbf3bd6SGeorge Wilson 	 * their tsd values so that we ensure that they only evict one dbuf
561dcbf3bd6SGeorge Wilson 	 * from the dbuf cache.
562dcbf3bd6SGeorge Wilson 	 */
563dcbf3bd6SGeorge Wilson 	if (tsd_get(zfs_dbuf_evict_key) != NULL)
564dcbf3bd6SGeorge Wilson 		return;
565dcbf3bd6SGeorge Wilson 
566dbfd9f93SMatthew Ahrens 	/*
567dbfd9f93SMatthew Ahrens 	 * We check if we should evict without holding the dbuf_evict_lock,
568dbfd9f93SMatthew Ahrens 	 * because it's OK to occasionally make the wrong decision here,
569dbfd9f93SMatthew Ahrens 	 * and grabbing the lock results in massive lock contention.
570dbfd9f93SMatthew Ahrens 	 */
571dcbf3bd6SGeorge Wilson 	if (refcount_count(&dbuf_cache_size) > dbuf_cache_max_bytes) {
572dbfd9f93SMatthew Ahrens 		if (dbuf_cache_above_hiwater())
573dcbf3bd6SGeorge Wilson 			dbuf_evict_one();
574dbfd9f93SMatthew Ahrens 		cv_signal(&dbuf_evict_cv);
575dcbf3bd6SGeorge Wilson 	}
576ea8dc4b6Seschrock }
577ea8dc4b6Seschrock 
578fa9e4066Sahrens void
579fa9e4066Sahrens dbuf_init(void)
580fa9e4066Sahrens {
581ea8dc4b6Seschrock 	uint64_t hsize = 1ULL << 16;
582fa9e4066Sahrens 	dbuf_hash_table_t *h = &dbuf_hash_table;
583fa9e4066Sahrens 	int i;
584fa9e4066Sahrens 
585fa9e4066Sahrens 	/*
586fa9e4066Sahrens 	 * The hash table is big enough to fill all of physical memory
587ea8dc4b6Seschrock 	 * with an average 4K block size.  The table will take up
588ea8dc4b6Seschrock 	 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
589fa9e4066Sahrens 	 */
590ea8dc4b6Seschrock 	while (hsize * 4096 < physmem * PAGESIZE)
591fa9e4066Sahrens 		hsize <<= 1;
592fa9e4066Sahrens 
593ea8dc4b6Seschrock retry:
594fa9e4066Sahrens 	h->hash_table_mask = hsize - 1;
595ea8dc4b6Seschrock 	h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
596ea8dc4b6Seschrock 	if (h->hash_table == NULL) {
597ea8dc4b6Seschrock 		/* XXX - we should really return an error instead of assert */
598ea8dc4b6Seschrock 		ASSERT(hsize > (1ULL << 10));
599ea8dc4b6Seschrock 		hsize >>= 1;
600ea8dc4b6Seschrock 		goto retry;
601ea8dc4b6Seschrock 	}
602fa9e4066Sahrens 
603dcbf3bd6SGeorge Wilson 	dbuf_kmem_cache = kmem_cache_create("dmu_buf_impl_t",
604fa9e4066Sahrens 	    sizeof (dmu_buf_impl_t),
605fa9e4066Sahrens 	    0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
606fa9e4066Sahrens 
607fa9e4066Sahrens 	for (i = 0; i < DBUF_MUTEXES; i++)
608fa9e4066Sahrens 		mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
609bc9014e6SJustin Gibbs 
610dcbf3bd6SGeorge Wilson 	/*
611dcbf3bd6SGeorge Wilson 	 * Setup the parameters for the dbuf cache. We cap the size of the
612dcbf3bd6SGeorge Wilson 	 * dbuf cache to 1/32nd (default) of the size of the ARC.
613dcbf3bd6SGeorge Wilson 	 */
614dcbf3bd6SGeorge Wilson 	dbuf_cache_max_bytes = MIN(dbuf_cache_max_bytes,
615dcbf3bd6SGeorge Wilson 	    arc_max_bytes() >> dbuf_cache_max_shift);
616dcbf3bd6SGeorge Wilson 
617bc9014e6SJustin Gibbs 	/*
618bc9014e6SJustin Gibbs 	 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
619bc9014e6SJustin Gibbs 	 * configuration is not required.
620bc9014e6SJustin Gibbs 	 */
621bc9014e6SJustin Gibbs 	dbu_evict_taskq = taskq_create("dbu_evict", 1, minclsyspri, 0, 0, 0);
622dcbf3bd6SGeorge Wilson 
62394c2d0ebSMatthew Ahrens 	dbuf_cache = multilist_create(sizeof (dmu_buf_impl_t),
624dcbf3bd6SGeorge Wilson 	    offsetof(dmu_buf_impl_t, db_cache_link),
625dcbf3bd6SGeorge Wilson 	    dbuf_cache_multilist_index_func);
626dcbf3bd6SGeorge Wilson 	refcount_create(&dbuf_cache_size);
627dcbf3bd6SGeorge Wilson 
628dcbf3bd6SGeorge Wilson 	tsd_create(&zfs_dbuf_evict_key, NULL);
629dcbf3bd6SGeorge Wilson 	dbuf_evict_thread_exit = B_FALSE;
630dcbf3bd6SGeorge Wilson 	mutex_init(&dbuf_evict_lock, NULL, MUTEX_DEFAULT, NULL);
631dcbf3bd6SGeorge Wilson 	cv_init(&dbuf_evict_cv, NULL, CV_DEFAULT, NULL);
632dcbf3bd6SGeorge Wilson 	dbuf_cache_evict_thread = thread_create(NULL, 0, dbuf_evict_thread,
633dcbf3bd6SGeorge Wilson 	    NULL, 0, &p0, TS_RUN, minclsyspri);
634fa9e4066Sahrens }
635fa9e4066Sahrens 
636fa9e4066Sahrens void
637fa9e4066Sahrens dbuf_fini(void)
638fa9e4066Sahrens {
639fa9e4066Sahrens 	dbuf_hash_table_t *h = &dbuf_hash_table;
640fa9e4066Sahrens 	int i;
641fa9e4066Sahrens 
642fa9e4066Sahrens 	for (i = 0; i < DBUF_MUTEXES; i++)
643fa9e4066Sahrens 		mutex_destroy(&h->hash_mutexes[i]);
644fa9e4066Sahrens 	kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
645dcbf3bd6SGeorge Wilson 	kmem_cache_destroy(dbuf_kmem_cache);
646bc9014e6SJustin Gibbs 	taskq_destroy(dbu_evict_taskq);
647dcbf3bd6SGeorge Wilson 
648dcbf3bd6SGeorge Wilson 	mutex_enter(&dbuf_evict_lock);
649dcbf3bd6SGeorge Wilson 	dbuf_evict_thread_exit = B_TRUE;
650dcbf3bd6SGeorge Wilson 	while (dbuf_evict_thread_exit) {
651dcbf3bd6SGeorge Wilson 		cv_signal(&dbuf_evict_cv);
652dcbf3bd6SGeorge Wilson 		cv_wait(&dbuf_evict_cv, &dbuf_evict_lock);
653dcbf3bd6SGeorge Wilson 	}
654dcbf3bd6SGeorge Wilson 	mutex_exit(&dbuf_evict_lock);
655dcbf3bd6SGeorge Wilson 	tsd_destroy(&zfs_dbuf_evict_key);
656dcbf3bd6SGeorge Wilson 
657dcbf3bd6SGeorge Wilson 	mutex_destroy(&dbuf_evict_lock);
658dcbf3bd6SGeorge Wilson 	cv_destroy(&dbuf_evict_cv);
659dcbf3bd6SGeorge Wilson 
660dcbf3bd6SGeorge Wilson 	refcount_destroy(&dbuf_cache_size);
66194c2d0ebSMatthew Ahrens 	multilist_destroy(dbuf_cache);
662fa9e4066Sahrens }
663fa9e4066Sahrens 
664fa9e4066Sahrens /*
665fa9e4066Sahrens  * Other stuff.
666fa9e4066Sahrens  */
667fa9e4066Sahrens 
6689c9dc39aSek #ifdef ZFS_DEBUG
669fa9e4066Sahrens static void
670fa9e4066Sahrens dbuf_verify(dmu_buf_impl_t *db)
671fa9e4066Sahrens {
672744947dcSTom Erickson 	dnode_t *dn;
673b24ab676SJeff Bonwick 	dbuf_dirty_record_t *dr;
674fa9e4066Sahrens 
675fa9e4066Sahrens 	ASSERT(MUTEX_HELD(&db->db_mtx));
676fa9e4066Sahrens 
677fa9e4066Sahrens 	if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
678fa9e4066Sahrens 		return;
679fa9e4066Sahrens 
680fa9e4066Sahrens 	ASSERT(db->db_objset != NULL);
681744947dcSTom Erickson 	DB_DNODE_ENTER(db);
682744947dcSTom Erickson 	dn = DB_DNODE(db);
683fa9e4066Sahrens 	if (dn == NULL) {
684fa9e4066Sahrens 		ASSERT(db->db_parent == NULL);
685fa9e4066Sahrens 		ASSERT(db->db_blkptr == NULL);
686fa9e4066Sahrens 	} else {
687fa9e4066Sahrens 		ASSERT3U(db->db.db_object, ==, dn->dn_object);
688fa9e4066Sahrens 		ASSERT3P(db->db_objset, ==, dn->dn_objset);
689fa9e4066Sahrens 		ASSERT3U(db->db_level, <, dn->dn_nlevels);
690744947dcSTom Erickson 		ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
691744947dcSTom Erickson 		    db->db_blkid == DMU_SPILL_BLKID ||
6920f6d88adSAlex Reece 		    !avl_is_empty(&dn->dn_dbufs));
693fa9e4066Sahrens 	}
6940a586ceaSMark Shellenbaum 	if (db->db_blkid == DMU_BONUS_BLKID) {
695fa9e4066Sahrens 		ASSERT(dn != NULL);
6961934e92fSmaybee 		ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
6970a586ceaSMark Shellenbaum 		ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
6980a586ceaSMark Shellenbaum 	} else if (db->db_blkid == DMU_SPILL_BLKID) {
6990a586ceaSMark Shellenbaum 		ASSERT(dn != NULL);
7000a586ceaSMark Shellenbaum 		ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
701fb09f5aaSMadhav Suresh 		ASSERT0(db->db.db_offset);
702fa9e4066Sahrens 	} else {
703fa9e4066Sahrens 		ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
704fa9e4066Sahrens 	}
705fa9e4066Sahrens 
706b24ab676SJeff Bonwick 	for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
707b24ab676SJeff Bonwick 		ASSERT(dr->dr_dbuf == db);
708b24ab676SJeff Bonwick 
709b24ab676SJeff Bonwick 	for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
710b24ab676SJeff Bonwick 		ASSERT(dr->dr_dbuf == db);
711b24ab676SJeff Bonwick 
71288b7b0f2SMatthew Ahrens 	/*
71388b7b0f2SMatthew Ahrens 	 * We can't assert that db_size matches dn_datablksz because it
71488b7b0f2SMatthew Ahrens 	 * can be momentarily different when another thread is doing
71588b7b0f2SMatthew Ahrens 	 * dnode_set_blksz().
71688b7b0f2SMatthew Ahrens 	 */
71788b7b0f2SMatthew Ahrens 	if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
718b24ab676SJeff Bonwick 		dr = db->db_data_pending;
71988b7b0f2SMatthew Ahrens 		/*
72088b7b0f2SMatthew Ahrens 		 * It should only be modified in syncing context, so
72188b7b0f2SMatthew Ahrens 		 * make sure we only have one copy of the data.
72288b7b0f2SMatthew Ahrens 		 */
72388b7b0f2SMatthew Ahrens 		ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
724fa9e4066Sahrens 	}
725fa9e4066Sahrens 
726fa9e4066Sahrens 	/* verify db->db_blkptr */
727fa9e4066Sahrens 	if (db->db_blkptr) {
728fa9e4066Sahrens 		if (db->db_parent == dn->dn_dbuf) {
729fa9e4066Sahrens 			/* db is pointed to by the dnode */
730fa9e4066Sahrens 			/* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
73114843421SMatthew Ahrens 			if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
732fa9e4066Sahrens 				ASSERT(db->db_parent == NULL);
733fa9e4066Sahrens 			else
734fa9e4066Sahrens 				ASSERT(db->db_parent != NULL);
7350a586ceaSMark Shellenbaum 			if (db->db_blkid != DMU_SPILL_BLKID)
7360a586ceaSMark Shellenbaum 				ASSERT3P(db->db_blkptr, ==,
7370a586ceaSMark Shellenbaum 				    &dn->dn_phys->dn_blkptr[db->db_blkid]);
738fa9e4066Sahrens 		} else {
739fa9e4066Sahrens 			/* db is pointed to by an indirect block */
740fa9e4066Sahrens 			int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
741fa9e4066Sahrens 			ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
742fa9e4066Sahrens 			ASSERT3U(db->db_parent->db.db_object, ==,
743fa9e4066Sahrens 			    db->db.db_object);
744fa9e4066Sahrens 			/*
745fa9e4066Sahrens 			 * dnode_grow_indblksz() can make this fail if we don't
746fa9e4066Sahrens 			 * have the struct_rwlock.  XXX indblksz no longer
747fa9e4066Sahrens 			 * grows.  safe to do this now?
748fa9e4066Sahrens 			 */
749744947dcSTom Erickson 			if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
750fa9e4066Sahrens 				ASSERT3P(db->db_blkptr, ==,
751fa9e4066Sahrens 				    ((blkptr_t *)db->db_parent->db.db_data +
752fa9e4066Sahrens 				    db->db_blkid % epb));
753fa9e4066Sahrens 			}
754fa9e4066Sahrens 		}
755fa9e4066Sahrens 	}
756fa9e4066Sahrens 	if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
7573f9d6ad7SLin Ling 	    (db->db_buf == NULL || db->db_buf->b_data) &&
7580a586ceaSMark Shellenbaum 	    db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
759fa9e4066Sahrens 	    db->db_state != DB_FILL && !dn->dn_free_txg) {
760fa9e4066Sahrens 		/*
761fa9e4066Sahrens 		 * If the blkptr isn't set but they have nonzero data,
762fa9e4066Sahrens 		 * it had better be dirty, otherwise we'll lose that
763fa9e4066Sahrens 		 * data when we evict this buffer.
7648df0bcf0SPaul Dagnelie 		 *
7658df0bcf0SPaul Dagnelie 		 * There is an exception to this rule for indirect blocks; in
7668df0bcf0SPaul Dagnelie 		 * this case, if the indirect block is a hole, we fill in a few
7678df0bcf0SPaul Dagnelie 		 * fields on each of the child blocks (importantly, birth time)
7688df0bcf0SPaul Dagnelie 		 * to prevent hole birth times from being lost when you
7698df0bcf0SPaul Dagnelie 		 * partially fill in a hole.
770fa9e4066Sahrens 		 */
771fa9e4066Sahrens 		if (db->db_dirtycnt == 0) {
7728df0bcf0SPaul Dagnelie 			if (db->db_level == 0) {
7738df0bcf0SPaul Dagnelie 				uint64_t *buf = db->db.db_data;
7748df0bcf0SPaul Dagnelie 				int i;
775fa9e4066Sahrens 
7768df0bcf0SPaul Dagnelie 				for (i = 0; i < db->db.db_size >> 3; i++) {
7778df0bcf0SPaul Dagnelie 					ASSERT(buf[i] == 0);
7788df0bcf0SPaul Dagnelie 				}
7798df0bcf0SPaul Dagnelie 			} else {
7808df0bcf0SPaul Dagnelie 				blkptr_t *bps = db->db.db_data;
7818df0bcf0SPaul Dagnelie 				ASSERT3U(1 << DB_DNODE(db)->dn_indblkshift, ==,
7828df0bcf0SPaul Dagnelie 				    db->db.db_size);
7838df0bcf0SPaul Dagnelie 				/*
7848df0bcf0SPaul Dagnelie 				 * We want to verify that all the blkptrs in the
7858df0bcf0SPaul Dagnelie 				 * indirect block are holes, but we may have
7868df0bcf0SPaul Dagnelie 				 * automatically set up a few fields for them.
7878df0bcf0SPaul Dagnelie 				 * We iterate through each blkptr and verify
7888df0bcf0SPaul Dagnelie 				 * they only have those fields set.
7898df0bcf0SPaul Dagnelie 				 */
7908df0bcf0SPaul Dagnelie 				for (int i = 0;
7918df0bcf0SPaul Dagnelie 				    i < db->db.db_size / sizeof (blkptr_t);
7928df0bcf0SPaul Dagnelie 				    i++) {
7938df0bcf0SPaul Dagnelie 					blkptr_t *bp = &bps[i];
7948df0bcf0SPaul Dagnelie 					ASSERT(ZIO_CHECKSUM_IS_ZERO(
7958df0bcf0SPaul Dagnelie 					    &bp->blk_cksum));
7968df0bcf0SPaul Dagnelie 					ASSERT(
7978df0bcf0SPaul Dagnelie 					    DVA_IS_EMPTY(&bp->blk_dva[0]) &&
7988df0bcf0SPaul Dagnelie 					    DVA_IS_EMPTY(&bp->blk_dva[1]) &&
7998df0bcf0SPaul Dagnelie 					    DVA_IS_EMPTY(&bp->blk_dva[2]));
8008df0bcf0SPaul Dagnelie 					ASSERT0(bp->blk_fill);
8018df0bcf0SPaul Dagnelie 					ASSERT0(bp->blk_pad[0]);
8028df0bcf0SPaul Dagnelie 					ASSERT0(bp->blk_pad[1]);
8038df0bcf0SPaul Dagnelie 					ASSERT(!BP_IS_EMBEDDED(bp));
8048df0bcf0SPaul Dagnelie 					ASSERT(BP_IS_HOLE(bp));
8058df0bcf0SPaul Dagnelie 					ASSERT0(bp->blk_phys_birth);
8068df0bcf0SPaul Dagnelie 				}
807fa9e4066Sahrens 			}
808fa9e4066Sahrens 		}
809fa9e4066Sahrens 	}
810744947dcSTom Erickson 	DB_DNODE_EXIT(db);
811fa9e4066Sahrens }
8129c9dc39aSek #endif
813fa9e4066Sahrens 
814bc9014e6SJustin Gibbs static void
815bc9014e6SJustin Gibbs dbuf_clear_data(dmu_buf_impl_t *db)
816bc9014e6SJustin Gibbs {
817bc9014e6SJustin Gibbs 	ASSERT(MUTEX_HELD(&db->db_mtx));
818bc9014e6SJustin Gibbs 	dbuf_evict_user(db);
819dcbf3bd6SGeorge Wilson 	ASSERT3P(db->db_buf, ==, NULL);
820bc9014e6SJustin Gibbs 	db->db.db_data = NULL;
821bc9014e6SJustin Gibbs 	if (db->db_state != DB_NOFILL)
822bc9014e6SJustin Gibbs 		db->db_state = DB_UNCACHED;
823bc9014e6SJustin Gibbs }
824bc9014e6SJustin Gibbs 
825fa9e4066Sahrens static void
826fa9e4066Sahrens dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
827fa9e4066Sahrens {
828fa9e4066Sahrens 	ASSERT(MUTEX_HELD(&db->db_mtx));
829bc9014e6SJustin Gibbs 	ASSERT(buf != NULL);
830bc9014e6SJustin Gibbs 
831fa9e4066Sahrens 	db->db_buf = buf;
832bc9014e6SJustin Gibbs 	ASSERT(buf->b_data != NULL);
833bc9014e6SJustin Gibbs 	db->db.db_data = buf->b_data;
834fa9e4066Sahrens }
835fa9e4066Sahrens 
836c242f9a0Schunli zhang - Sun Microsystems - Irvine United States /*
837c242f9a0Schunli zhang - Sun Microsystems - Irvine United States  * Loan out an arc_buf for read.  Return the loaned arc_buf.
838c242f9a0Schunli zhang - Sun Microsystems - Irvine United States  */
839c242f9a0Schunli zhang - Sun Microsystems - Irvine United States arc_buf_t *
840c242f9a0Schunli zhang - Sun Microsystems - Irvine United States dbuf_loan_arcbuf(dmu_buf_impl_t *db)
841c242f9a0Schunli zhang - Sun Microsystems - Irvine United States {
842c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 	arc_buf_t *abuf;
843c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 
844dcbf3bd6SGeorge Wilson 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
845c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 	mutex_enter(&db->db_mtx);
846c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 	if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
847c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 		int blksz = db->db.db_size;
84843466aaeSMax Grossman 		spa_t *spa = db->db_objset->os_spa;
849744947dcSTom Erickson 
850c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 		mutex_exit(&db->db_mtx);
8515602294fSDan Kimmel 		abuf = arc_loan_buf(spa, B_FALSE, blksz);
852c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 		bcopy(db->db.db_data, abuf->b_data, blksz);
853c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 	} else {
854c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 		abuf = db->db_buf;
855c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 		arc_loan_inuse_buf(abuf, db);
856dcbf3bd6SGeorge Wilson 		db->db_buf = NULL;
857bc9014e6SJustin Gibbs 		dbuf_clear_data(db);
858c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 		mutex_exit(&db->db_mtx);
859c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 	}
860c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 	return (abuf);
861c242f9a0Schunli zhang - Sun Microsystems - Irvine United States }
862c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 
863a2cdcdd2SPaul Dagnelie /*
864a2cdcdd2SPaul Dagnelie  * Calculate which level n block references the data at the level 0 offset
865a2cdcdd2SPaul Dagnelie  * provided.
866a2cdcdd2SPaul Dagnelie  */
867fa9e4066Sahrens uint64_t
868a2cdcdd2SPaul Dagnelie dbuf_whichblock(dnode_t *dn, int64_t level, uint64_t offset)
869fa9e4066Sahrens {
870a2cdcdd2SPaul Dagnelie 	if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) {
871a2cdcdd2SPaul Dagnelie 		/*
872a2cdcdd2SPaul Dagnelie 		 * The level n blkid is equal to the level 0 blkid divided by
873a2cdcdd2SPaul Dagnelie 		 * the number of level 0s in a level n block.
874a2cdcdd2SPaul Dagnelie 		 *
875a2cdcdd2SPaul Dagnelie 		 * The level 0 blkid is offset >> datablkshift =
876a2cdcdd2SPaul Dagnelie 		 * offset / 2^datablkshift.
877a2cdcdd2SPaul Dagnelie 		 *
878a2cdcdd2SPaul Dagnelie 		 * The number of level 0s in a level n is the number of block
879a2cdcdd2SPaul Dagnelie 		 * pointers in an indirect block, raised to the power of level.
880a2cdcdd2SPaul Dagnelie 		 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
881a2cdcdd2SPaul Dagnelie 		 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
882a2cdcdd2SPaul Dagnelie 		 *
883a2cdcdd2SPaul Dagnelie 		 * Thus, the level n blkid is: offset /
884a2cdcdd2SPaul Dagnelie 		 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
885a2cdcdd2SPaul Dagnelie 		 * = offset / 2^(datablkshift + level *
886a2cdcdd2SPaul Dagnelie 		 *   (indblkshift - SPA_BLKPTRSHIFT))
887a2cdcdd2SPaul Dagnelie 		 * = offset >> (datablkshift + level *
888a2cdcdd2SPaul Dagnelie 		 *   (indblkshift - SPA_BLKPTRSHIFT))
889a2cdcdd2SPaul Dagnelie 		 */
890a2cdcdd2SPaul Dagnelie 		return (offset >> (dn->dn_datablkshift + level *
891a2cdcdd2SPaul Dagnelie 		    (dn->dn_indblkshift - SPA_BLKPTRSHIFT)));
892fa9e4066Sahrens 	} else {
893fa9e4066Sahrens 		ASSERT3U(offset, <, dn->dn_datablksz);
894fa9e4066Sahrens 		return (0);
895fa9e4066Sahrens 	}
896fa9e4066Sahrens }
897fa9e4066Sahrens 
898fa9e4066Sahrens static void
899fa9e4066Sahrens dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
900fa9e4066Sahrens {
901fa9e4066Sahrens 	dmu_buf_impl_t *db = vdb;
902fa9e4066Sahrens 
903fa9e4066Sahrens 	mutex_enter(&db->db_mtx);
904fa9e4066Sahrens 	ASSERT3U(db->db_state, ==, DB_READ);
905fa9e4066Sahrens 	/*
906fa9e4066Sahrens 	 * All reads are synchronous, so we must have a hold on the dbuf
907fa9e4066Sahrens 	 */
908fa9e4066Sahrens 	ASSERT(refcount_count(&db->db_holds) > 0);
909ea8dc4b6Seschrock 	ASSERT(db->db_buf == NULL);
910fa9e4066Sahrens 	ASSERT(db->db.db_data == NULL);
911c717a561Smaybee 	if (db->db_level == 0 && db->db_freed_in_flight) {
912fa9e4066Sahrens 		/* we were freed in flight; disregard any error */
913fa9e4066Sahrens 		arc_release(buf, db);
914fa9e4066Sahrens 		bzero(buf->b_data, db->db.db_size);
9156b4acc8bSahrens 		arc_buf_freeze(buf);
916c717a561Smaybee 		db->db_freed_in_flight = FALSE;
917fa9e4066Sahrens 		dbuf_set_data(db, buf);
918fa9e4066Sahrens 		db->db_state = DB_CACHED;
919fa9e4066Sahrens 	} else if (zio == NULL || zio->io_error == 0) {
920fa9e4066Sahrens 		dbuf_set_data(db, buf);
921fa9e4066Sahrens 		db->db_state = DB_CACHED;
922fa9e4066Sahrens 	} else {
9230a586ceaSMark Shellenbaum 		ASSERT(db->db_blkid != DMU_BONUS_BLKID);
924fa9e4066Sahrens 		ASSERT3P(db->db_buf, ==, NULL);
925dcbf3bd6SGeorge Wilson 		arc_buf_destroy(buf, db);
926ea8dc4b6Seschrock 		db->db_state = DB_UNCACHED;
927fa9e4066Sahrens 	}
928fa9e4066Sahrens 	cv_broadcast(&db->db_changed);
9293f9d6ad7SLin Ling 	dbuf_rele_and_unlock(db, NULL);
930fa9e4066Sahrens }
931fa9e4066Sahrens 
932ea8dc4b6Seschrock static void
933cf6106c8SMatthew Ahrens dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
934fa9e4066Sahrens {
935744947dcSTom Erickson 	dnode_t *dn;
9367802d7bfSMatthew Ahrens 	zbookmark_phys_t zb;
9377adb730bSGeorge Wilson 	arc_flags_t aflags = ARC_FLAG_NOWAIT;
938fa9e4066Sahrens 
939744947dcSTom Erickson 	DB_DNODE_ENTER(db);
940744947dcSTom Erickson 	dn = DB_DNODE(db);
941fa9e4066Sahrens 	ASSERT(!refcount_is_zero(&db->db_holds));
942fa9e4066Sahrens 	/* We need the struct_rwlock to prevent db_blkptr from changing. */
943088f3894Sahrens 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
944ea8dc4b6Seschrock 	ASSERT(MUTEX_HELD(&db->db_mtx));
945ea8dc4b6Seschrock 	ASSERT(db->db_state == DB_UNCACHED);
946ea8dc4b6Seschrock 	ASSERT(db->db_buf == NULL);
947fa9e4066Sahrens 
9480a586ceaSMark Shellenbaum 	if (db->db_blkid == DMU_BONUS_BLKID) {
949cf04dda1SMark Maybee 		int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
9501934e92fSmaybee 
9511934e92fSmaybee 		ASSERT3U(bonuslen, <=, db->db.db_size);
952ea8dc4b6Seschrock 		db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN);
9535a98e54bSBrendan Gregg - Sun Microsystems 		arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
9541934e92fSmaybee 		if (bonuslen < DN_MAX_BONUSLEN)
955ea8dc4b6Seschrock 			bzero(db->db.db_data, DN_MAX_BONUSLEN);
956cf04dda1SMark Maybee 		if (bonuslen)
957cf04dda1SMark Maybee 			bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
958744947dcSTom Erickson 		DB_DNODE_EXIT(db);
959fa9e4066Sahrens 		db->db_state = DB_CACHED;
960fa9e4066Sahrens 		mutex_exit(&db->db_mtx);
961fa9e4066Sahrens 		return;
962fa9e4066Sahrens 	}
963fa9e4066Sahrens 
9641c8564a7SMark Maybee 	/*
9651c8564a7SMark Maybee 	 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
9661c8564a7SMark Maybee 	 * processes the delete record and clears the bp while we are waiting
9671c8564a7SMark Maybee 	 * for the dn_mtx (resulting in a "no" from block_freed).
9681c8564a7SMark Maybee 	 */
969088f3894Sahrens 	if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
9701c8564a7SMark Maybee 	    (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
9711c8564a7SMark Maybee 	    BP_IS_HOLE(db->db_blkptr)))) {
972ad23a2dbSjohansen 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
973ad23a2dbSjohansen 
9745602294fSDan Kimmel 		dbuf_set_data(db, arc_alloc_buf(db->db_objset->os_spa, db, type,
9755602294fSDan Kimmel 		    db->db.db_size));
976fa9e4066Sahrens 		bzero(db->db.db_data, db->db.db_size);
9778df0bcf0SPaul Dagnelie 
9788df0bcf0SPaul Dagnelie 		if (db->db_blkptr != NULL && db->db_level > 0 &&
9798df0bcf0SPaul Dagnelie 		    BP_IS_HOLE(db->db_blkptr) &&
9808df0bcf0SPaul Dagnelie 		    db->db_blkptr->blk_birth != 0) {
9818df0bcf0SPaul Dagnelie 			blkptr_t *bps = db->db.db_data;
9828df0bcf0SPaul Dagnelie 			for (int i = 0; i < ((1 <<
9838df0bcf0SPaul Dagnelie 			    DB_DNODE(db)->dn_indblkshift) / sizeof (blkptr_t));
9848df0bcf0SPaul Dagnelie 			    i++) {
9858df0bcf0SPaul Dagnelie 				blkptr_t *bp = &bps[i];
9868df0bcf0SPaul Dagnelie 				ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
9878df0bcf0SPaul Dagnelie 				    1 << dn->dn_indblkshift);
9888df0bcf0SPaul Dagnelie 				BP_SET_LSIZE(bp,
9898df0bcf0SPaul Dagnelie 				    BP_GET_LEVEL(db->db_blkptr) == 1 ?
9908df0bcf0SPaul Dagnelie 				    dn->dn_datablksz :
9918df0bcf0SPaul Dagnelie 				    BP_GET_LSIZE(db->db_blkptr));
9928df0bcf0SPaul Dagnelie 				BP_SET_TYPE(bp, BP_GET_TYPE(db->db_blkptr));
9938df0bcf0SPaul Dagnelie 				BP_SET_LEVEL(bp,
9948df0bcf0SPaul Dagnelie 				    BP_GET_LEVEL(db->db_blkptr) - 1);
9958df0bcf0SPaul Dagnelie 				BP_SET_BIRTH(bp, db->db_blkptr->blk_birth, 0);
9968df0bcf0SPaul Dagnelie 			}
9978df0bcf0SPaul Dagnelie 		}
9988df0bcf0SPaul Dagnelie 		DB_DNODE_EXIT(db);
999fa9e4066Sahrens 		db->db_state = DB_CACHED;
1000fa9e4066Sahrens 		mutex_exit(&db->db_mtx);
1001fa9e4066Sahrens 		return;
1002fa9e4066Sahrens 	}
1003fa9e4066Sahrens 
1004744947dcSTom Erickson 	DB_DNODE_EXIT(db);
1005744947dcSTom Erickson 
1006fa9e4066Sahrens 	db->db_state = DB_READ;
1007fa9e4066Sahrens 	mutex_exit(&db->db_mtx);
1008fa9e4066Sahrens 
10093baa08fcSek 	if (DBUF_IS_L2CACHEABLE(db))
10107adb730bSGeorge Wilson 		aflags |= ARC_FLAG_L2CACHE;
10113baa08fcSek 
1012b24ab676SJeff Bonwick 	SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
1013b24ab676SJeff Bonwick 	    db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
1014b24ab676SJeff Bonwick 	    db->db.db_object, db->db_level, db->db_blkid);
1015ea8dc4b6Seschrock 
1016ea8dc4b6Seschrock 	dbuf_add_ref(db, NULL);
1017088f3894Sahrens 
101843466aaeSMax Grossman 	(void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr,
1019fa9e4066Sahrens 	    dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
1020cf6106c8SMatthew Ahrens 	    (flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
102113506d1eSmaybee 	    &aflags, &zb);
1022fa9e4066Sahrens }
1023fa9e4066Sahrens 
10245602294fSDan Kimmel /*
10255602294fSDan Kimmel  * This is our just-in-time copy function.  It makes a copy of buffers that
10265602294fSDan Kimmel  * have been modified in a previous transaction group before we access them in
10275602294fSDan Kimmel  * the current active group.
10285602294fSDan Kimmel  *
10295602294fSDan Kimmel  * This function is used in three places: when we are dirtying a buffer for the
10305602294fSDan Kimmel  * first time in a txg, when we are freeing a range in a dnode that includes
10315602294fSDan Kimmel  * this buffer, and when we are accessing a buffer which was received compressed
10325602294fSDan Kimmel  * and later referenced in a WRITE_BYREF record.
10335602294fSDan Kimmel  *
10345602294fSDan Kimmel  * Note that when we are called from dbuf_free_range() we do not put a hold on
10355602294fSDan Kimmel  * the buffer, we just traverse the active dbuf list for the dnode.
10365602294fSDan Kimmel  */
10375602294fSDan Kimmel static void
10385602294fSDan Kimmel dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
10395602294fSDan Kimmel {
10405602294fSDan Kimmel 	dbuf_dirty_record_t *dr = db->db_last_dirty;
10415602294fSDan Kimmel 
10425602294fSDan Kimmel 	ASSERT(MUTEX_HELD(&db->db_mtx));
10435602294fSDan Kimmel 	ASSERT(db->db.db_data != NULL);
10445602294fSDan Kimmel 	ASSERT(db->db_level == 0);
10455602294fSDan Kimmel 	ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
10465602294fSDan Kimmel 
10475602294fSDan Kimmel 	if (dr == NULL ||
10485602294fSDan Kimmel 	    (dr->dt.dl.dr_data !=
10495602294fSDan Kimmel 	    ((db->db_blkid  == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
10505602294fSDan Kimmel 		return;
10515602294fSDan Kimmel 
10525602294fSDan Kimmel 	/*
10535602294fSDan Kimmel 	 * If the last dirty record for this dbuf has not yet synced
10545602294fSDan Kimmel 	 * and its referencing the dbuf data, either:
10555602294fSDan Kimmel 	 *	reset the reference to point to a new copy,
10565602294fSDan Kimmel 	 * or (if there a no active holders)
10575602294fSDan Kimmel 	 *	just null out the current db_data pointer.
10585602294fSDan Kimmel 	 */
10595602294fSDan Kimmel 	ASSERT(dr->dr_txg >= txg - 2);
10605602294fSDan Kimmel 	if (db->db_blkid == DMU_BONUS_BLKID) {
10615602294fSDan Kimmel 		/* Note that the data bufs here are zio_bufs */
10625602294fSDan Kimmel 		dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
10635602294fSDan Kimmel 		arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
10645602294fSDan Kimmel 		bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
10655602294fSDan Kimmel 	} else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
10665602294fSDan Kimmel 		int size = arc_buf_size(db->db_buf);
10675602294fSDan Kimmel 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
10685602294fSDan Kimmel 		spa_t *spa = db->db_objset->os_spa;
10695602294fSDan Kimmel 		enum zio_compress compress_type =
10705602294fSDan Kimmel 		    arc_get_compression(db->db_buf);
10715602294fSDan Kimmel 
10725602294fSDan Kimmel 		if (compress_type == ZIO_COMPRESS_OFF) {
10735602294fSDan Kimmel 			dr->dt.dl.dr_data = arc_alloc_buf(spa, db, type, size);
10745602294fSDan Kimmel 		} else {
10755602294fSDan Kimmel 			ASSERT3U(type, ==, ARC_BUFC_DATA);
10765602294fSDan Kimmel 			dr->dt.dl.dr_data = arc_alloc_compressed_buf(spa, db,
10775602294fSDan Kimmel 			    size, arc_buf_lsize(db->db_buf), compress_type);
10785602294fSDan Kimmel 		}
10795602294fSDan Kimmel 		bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
10805602294fSDan Kimmel 	} else {
10815602294fSDan Kimmel 		db->db_buf = NULL;
10825602294fSDan Kimmel 		dbuf_clear_data(db);
10835602294fSDan Kimmel 	}
10845602294fSDan Kimmel }
10855602294fSDan Kimmel 
1086ea8dc4b6Seschrock int
1087ea8dc4b6Seschrock dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
1088fa9e4066Sahrens {
1089ea8dc4b6Seschrock 	int err = 0;
109043466aaeSMax Grossman 	boolean_t prefetch;
1091744947dcSTom Erickson 	dnode_t *dn;
1092fa9e4066Sahrens 
1093fa9e4066Sahrens 	/*
1094fa9e4066Sahrens 	 * We don't have to hold the mutex to check db_state because it
1095fa9e4066Sahrens 	 * can't be freed while we have a hold on the buffer.
1096fa9e4066Sahrens 	 */
1097fa9e4066Sahrens 	ASSERT(!refcount_is_zero(&db->db_holds));
1098fa9e4066Sahrens 
109982c9918fSTim Haley 	if (db->db_state == DB_NOFILL)
1100be6fd75aSMatthew Ahrens 		return (SET_ERROR(EIO));
110182c9918fSTim Haley 
1102744947dcSTom Erickson 	DB_DNODE_ENTER(db);
1103744947dcSTom Erickson 	dn = DB_DNODE(db);
1104ea8dc4b6Seschrock 	if ((flags & DB_RF_HAVESTRUCT) == 0)
1105744947dcSTom Erickson 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
1106ea8dc4b6Seschrock 
11070a586ceaSMark Shellenbaum 	prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1108744947dcSTom Erickson 	    (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
11093baa08fcSek 	    DBUF_IS_CACHEABLE(db);
111013506d1eSmaybee 
1111ea8dc4b6Seschrock 	mutex_enter(&db->db_mtx);
1112ea8dc4b6Seschrock 	if (db->db_state == DB_CACHED) {
11135602294fSDan Kimmel 		/*
11145602294fSDan Kimmel 		 * If the arc buf is compressed, we need to decompress it to
11155602294fSDan Kimmel 		 * read the data. This could happen during the "zfs receive" of
11165602294fSDan Kimmel 		 * a stream which is compressed and deduplicated.
11175602294fSDan Kimmel 		 */
11185602294fSDan Kimmel 		if (db->db_buf != NULL &&
11195602294fSDan Kimmel 		    arc_get_compression(db->db_buf) != ZIO_COMPRESS_OFF) {
11205602294fSDan Kimmel 			dbuf_fix_old_data(db,
11215602294fSDan Kimmel 			    spa_syncing_txg(dmu_objset_spa(db->db_objset)));
11225602294fSDan Kimmel 			err = arc_decompress(db->db_buf);
11235602294fSDan Kimmel 			dbuf_set_data(db, db->db_buf);
11245602294fSDan Kimmel 		}
1125ea8dc4b6Seschrock 		mutex_exit(&db->db_mtx);
112613506d1eSmaybee 		if (prefetch)
1127cb92f413SAlexander Motin 			dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1128fa9e4066Sahrens 		if ((flags & DB_RF_HAVESTRUCT) == 0)
1129744947dcSTom Erickson 			rw_exit(&dn->dn_struct_rwlock);
1130744947dcSTom Erickson 		DB_DNODE_EXIT(db);
1131ea8dc4b6Seschrock 	} else if (db->db_state == DB_UNCACHED) {
1132744947dcSTom Erickson 		spa_t *spa = dn->dn_objset->os_spa;
1133def4fac5SMatthew Ahrens 		boolean_t need_wait = B_FALSE;
1134744947dcSTom Erickson 
1135def4fac5SMatthew Ahrens 		if (zio == NULL &&
1136def4fac5SMatthew Ahrens 		    db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) {
1137744947dcSTom Erickson 			zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
1138def4fac5SMatthew Ahrens 			need_wait = B_TRUE;
1139def4fac5SMatthew Ahrens 		}
1140cf6106c8SMatthew Ahrens 		dbuf_read_impl(db, zio, flags);
114113506d1eSmaybee 
1142ea8dc4b6Seschrock 		/* dbuf_read_impl has dropped db_mtx for us */
1143ea8dc4b6Seschrock 
114413506d1eSmaybee 		if (prefetch)
1145cb92f413SAlexander Motin 			dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1146ea8dc4b6Seschrock 
1147fa9e4066Sahrens 		if ((flags & DB_RF_HAVESTRUCT) == 0)
1148744947dcSTom Erickson 			rw_exit(&dn->dn_struct_rwlock);
1149744947dcSTom Erickson 		DB_DNODE_EXIT(db);
1150fa9e4066Sahrens 
1151def4fac5SMatthew Ahrens 		if (need_wait)
1152ea8dc4b6Seschrock 			err = zio_wait(zio);
1153ea8dc4b6Seschrock 	} else {
11543e30c24aSWill Andrews 		/*
11553e30c24aSWill Andrews 		 * Another reader came in while the dbuf was in flight
11563e30c24aSWill Andrews 		 * between UNCACHED and CACHED.  Either a writer will finish
11573e30c24aSWill Andrews 		 * writing the buffer (sending the dbuf to CACHED) or the
11583e30c24aSWill Andrews 		 * first reader's request will reach the read_done callback
11593e30c24aSWill Andrews 		 * and send the dbuf to CACHED.  Otherwise, a failure
11603e30c24aSWill Andrews 		 * occurred and the dbuf went to UNCACHED.
11613e30c24aSWill Andrews 		 */
116213506d1eSmaybee 		mutex_exit(&db->db_mtx);
116313506d1eSmaybee 		if (prefetch)
1164cb92f413SAlexander Motin 			dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1165ea8dc4b6Seschrock 		if ((flags & DB_RF_HAVESTRUCT) == 0)
1166744947dcSTom Erickson 			rw_exit(&dn->dn_struct_rwlock);
1167744947dcSTom Erickson 		DB_DNODE_EXIT(db);
116813506d1eSmaybee 
11693e30c24aSWill Andrews 		/* Skip the wait per the caller's request. */
117013506d1eSmaybee 		mutex_enter(&db->db_mtx);
1171ea8dc4b6Seschrock 		if ((flags & DB_RF_NEVERWAIT) == 0) {
1172ea8dc4b6Seschrock 			while (db->db_state == DB_READ ||
1173ea8dc4b6Seschrock 			    db->db_state == DB_FILL) {
1174ea8dc4b6Seschrock 				ASSERT(db->db_state == DB_READ ||
1175ea8dc4b6Seschrock 				    (flags & DB_RF_HAVESTRUCT) == 0);
1176f6164ad6SAdam H. Leventhal 				DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *,
1177f6164ad6SAdam H. Leventhal 				    db, zio_t *, zio);
1178ea8dc4b6Seschrock 				cv_wait(&db->db_changed, &db->db_mtx);
1179ea8dc4b6Seschrock 			}
1180ea8dc4b6Seschrock 			if (db->db_state == DB_UNCACHED)
1181be6fd75aSMatthew Ahrens 				err = SET_ERROR(EIO);
1182ea8dc4b6Seschrock 		}
1183ea8dc4b6Seschrock 		mutex_exit(&db->db_mtx);
1184fa9e4066Sahrens 	}
1185fa9e4066Sahrens 
1186ea8dc4b6Seschrock 	return (err);
1187fa9e4066Sahrens }
1188fa9e4066Sahrens 
1189fa9e4066Sahrens static void
1190fa9e4066Sahrens dbuf_noread(dmu_buf_impl_t *db)
1191fa9e4066Sahrens {
1192fa9e4066Sahrens 	ASSERT(!refcount_is_zero(&db->db_holds));
11930a586ceaSMark Shellenbaum 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1194fa9e4066Sahrens 	mutex_enter(&db->db_mtx);
1195fa9e4066Sahrens 	while (db->db_state == DB_READ || db->db_state == DB_FILL)
1196fa9e4066Sahrens 		cv_wait(&db->db_changed, &db->db_mtx);
1197fa9e4066Sahrens 	if (db->db_state == DB_UNCACHED) {
1198ad23a2dbSjohansen 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
119943466aaeSMax Grossman 		spa_t *spa = db->db_objset->os_spa;
1200ad23a2dbSjohansen 
1201ea8dc4b6Seschrock 		ASSERT(db->db_buf == NULL);
1202fa9e4066Sahrens 		ASSERT(db->db.db_data == NULL);
12035602294fSDan Kimmel 		dbuf_set_data(db, arc_alloc_buf(spa, db, type, db->db.db_size));
1204fa9e4066Sahrens 		db->db_state = DB_FILL;
120582c9918fSTim Haley 	} else if (db->db_state == DB_NOFILL) {
1206bc9014e6SJustin Gibbs 		dbuf_clear_data(db);
1207fa9e4066Sahrens 	} else {
1208fa9e4066Sahrens 		ASSERT3U(db->db_state, ==, DB_CACHED);
1209fa9e4066Sahrens 	}
1210fa9e4066Sahrens 	mutex_exit(&db->db_mtx);
1211fa9e4066Sahrens }
1212fa9e4066Sahrens 
1213c717a561Smaybee void
1214c717a561Smaybee dbuf_unoverride(dbuf_dirty_record_t *dr)
1215ea8dc4b6Seschrock {
1216c717a561Smaybee 	dmu_buf_impl_t *db = dr->dr_dbuf;
1217b24ab676SJeff Bonwick 	blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
1218c717a561Smaybee 	uint64_t txg = dr->dr_txg;
1219ea8dc4b6Seschrock 
1220ea8dc4b6Seschrock 	ASSERT(MUTEX_HELD(&db->db_mtx));
122140713f2bSAlan Somers 	/*
122240713f2bSAlan Somers 	 * This assert is valid because dmu_sync() expects to be called by
122340713f2bSAlan Somers 	 * a zilog's get_data while holding a range lock.  This call only
122440713f2bSAlan Somers 	 * comes from dbuf_dirty() callers who must also hold a range lock.
122540713f2bSAlan Somers 	 */
1226c717a561Smaybee 	ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
1227c717a561Smaybee 	ASSERT(db->db_level == 0);
1228ea8dc4b6Seschrock 
12290a586ceaSMark Shellenbaum 	if (db->db_blkid == DMU_BONUS_BLKID ||
1230c717a561Smaybee 	    dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
1231c717a561Smaybee 		return;
1232ea8dc4b6Seschrock 
1233b24ab676SJeff Bonwick 	ASSERT(db->db_data_pending != dr);
1234b24ab676SJeff Bonwick 
1235c717a561Smaybee 	/* free this block */
123643466aaeSMax Grossman 	if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite)
123743466aaeSMax Grossman 		zio_free(db->db_objset->os_spa, txg, bp);
1238b24ab676SJeff Bonwick 
1239c717a561Smaybee 	dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
124080901aeaSGeorge Wilson 	dr->dt.dl.dr_nopwrite = B_FALSE;
124180901aeaSGeorge Wilson 
1242c717a561Smaybee 	/*
1243c717a561Smaybee 	 * Release the already-written buffer, so we leave it in
1244c717a561Smaybee 	 * a consistent dirty state.  Note that all callers are
1245c717a561Smaybee 	 * modifying the buffer, so they will immediately do
1246c717a561Smaybee 	 * another (redundant) arc_release().  Therefore, leave
1247c717a561Smaybee 	 * the buf thawed to save the effort of freezing &
1248c717a561Smaybee 	 * immediately re-thawing it.
1249c717a561Smaybee 	 */
1250c717a561Smaybee 	arc_release(dr->dt.dl.dr_data, db);
1251fa9e4066Sahrens }
1252fa9e4066Sahrens 
1253cdb0ab79Smaybee /*
1254cdb0ab79Smaybee  * Evict (if its unreferenced) or clear (if its referenced) any level-0
1255cdb0ab79Smaybee  * data blocks in the free range, so that any future readers will find
125643466aaeSMax Grossman  * empty blocks.
1257cdb0ab79Smaybee  */
1258fa9e4066Sahrens void
12590f6d88adSAlex Reece dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
12600f6d88adSAlex Reece     dmu_tx_t *tx)
1261fa9e4066Sahrens {
1262bc9014e6SJustin Gibbs 	dmu_buf_impl_t db_search;
1263bc9014e6SJustin Gibbs 	dmu_buf_impl_t *db, *db_next;
1264fa9e4066Sahrens 	uint64_t txg = tx->tx_txg;
12650f6d88adSAlex Reece 	avl_index_t where;
1266fa9e4066Sahrens 
1267653af1b8SStephen Blinick 	if (end_blkid > dn->dn_maxblkid &&
1268653af1b8SStephen Blinick 	    !(start_blkid == DMU_SPILL_BLKID || end_blkid == DMU_SPILL_BLKID))
12690f6d88adSAlex Reece 		end_blkid = dn->dn_maxblkid;
12700f6d88adSAlex Reece 	dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid);
12710f6d88adSAlex Reece 
12720f6d88adSAlex Reece 	db_search.db_level = 0;
12730f6d88adSAlex Reece 	db_search.db_blkid = start_blkid;
127486bb58aeSAlex Reece 	db_search.db_state = DB_SEARCH;
12752f3d8780SMatthew Ahrens 
1276713d6c20SMatthew Ahrens 	mutex_enter(&dn->dn_dbufs_mtx);
12770f6d88adSAlex Reece 	db = avl_find(&dn->dn_dbufs, &db_search, &where);
12780f6d88adSAlex Reece 	ASSERT3P(db, ==, NULL);
1279653af1b8SStephen Blinick 
12800f6d88adSAlex Reece 	db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
12810f6d88adSAlex Reece 
12820f6d88adSAlex Reece 	for (; db != NULL; db = db_next) {
12830f6d88adSAlex Reece 		db_next = AVL_NEXT(&dn->dn_dbufs, db);
12840a586ceaSMark Shellenbaum 		ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1285cdb0ab79Smaybee 
12860f6d88adSAlex Reece 		if (db->db_level != 0 || db->db_blkid > end_blkid) {
12870f6d88adSAlex Reece 			break;
12880f6d88adSAlex Reece 		}
12890f6d88adSAlex Reece 		ASSERT3U(db->db_blkid, >=, start_blkid);
1290fa9e4066Sahrens 
1291fa9e4066Sahrens 		/* found a level 0 buffer in the range */
12923b2aab18SMatthew Ahrens 		mutex_enter(&db->db_mtx);
12933b2aab18SMatthew Ahrens 		if (dbuf_undirty(db, tx)) {
12943b2aab18SMatthew Ahrens 			/* mutex has been dropped and dbuf destroyed */
1295fa9e4066Sahrens 			continue;
12963b2aab18SMatthew Ahrens 		}
1297fa9e4066Sahrens 
1298ea8dc4b6Seschrock 		if (db->db_state == DB_UNCACHED ||
129982c9918fSTim Haley 		    db->db_state == DB_NOFILL ||
1300ea8dc4b6Seschrock 		    db->db_state == DB_EVICTING) {
1301fa9e4066Sahrens 			ASSERT(db->db.db_data == NULL);
1302fa9e4066Sahrens 			mutex_exit(&db->db_mtx);
1303fa9e4066Sahrens 			continue;
1304fa9e4066Sahrens 		}
1305c543ec06Sahrens 		if (db->db_state == DB_READ || db->db_state == DB_FILL) {
1306c543ec06Sahrens 			/* will be handled in dbuf_read_done or dbuf_rele */
1307c717a561Smaybee 			db->db_freed_in_flight = TRUE;
1308fa9e4066Sahrens 			mutex_exit(&db->db_mtx);
1309fa9e4066Sahrens 			continue;
1310fa9e4066Sahrens 		}
1311ea8dc4b6Seschrock 		if (refcount_count(&db->db_holds) == 0) {
1312ea8dc4b6Seschrock 			ASSERT(db->db_buf);
1313dcbf3bd6SGeorge Wilson 			dbuf_destroy(db);
1314ea8dc4b6Seschrock 			continue;
1315ea8dc4b6Seschrock 		}
1316c717a561Smaybee 		/* The dbuf is referenced */
1317fa9e4066Sahrens 
1318c717a561Smaybee 		if (db->db_last_dirty != NULL) {
1319c717a561Smaybee 			dbuf_dirty_record_t *dr = db->db_last_dirty;
1320c717a561Smaybee 
1321c717a561Smaybee 			if (dr->dr_txg == txg) {
132244eda4d7Smaybee 				/*
1323c717a561Smaybee 				 * This buffer is "in-use", re-adjust the file
1324c717a561Smaybee 				 * size to reflect that this buffer may
1325c717a561Smaybee 				 * contain new data when we sync.
132644eda4d7Smaybee 				 */
132706e0070dSMark Shellenbaum 				if (db->db_blkid != DMU_SPILL_BLKID &&
132806e0070dSMark Shellenbaum 				    db->db_blkid > dn->dn_maxblkid)
1329c717a561Smaybee 					dn->dn_maxblkid = db->db_blkid;
1330c717a561Smaybee 				dbuf_unoverride(dr);
1331c717a561Smaybee 			} else {
1332c717a561Smaybee 				/*
1333c717a561Smaybee 				 * This dbuf is not dirty in the open context.
1334c717a561Smaybee 				 * Either uncache it (if its not referenced in
1335c717a561Smaybee 				 * the open context) or reset its contents to
1336c717a561Smaybee 				 * empty.
1337c717a561Smaybee 				 */
1338c717a561Smaybee 				dbuf_fix_old_data(db, txg);
133944eda4d7Smaybee 			}
1340ea8dc4b6Seschrock 		}
1341c717a561Smaybee 		/* clear the contents if its cached */
1342ea8dc4b6Seschrock 		if (db->db_state == DB_CACHED) {
1343ea8dc4b6Seschrock 			ASSERT(db->db.db_data != NULL);
1344fa9e4066Sahrens 			arc_release(db->db_buf, db);
1345fa9e4066Sahrens 			bzero(db->db.db_data, db->db.db_size);
13466b4acc8bSahrens 			arc_buf_freeze(db->db_buf);
1347fa9e4066Sahrens 		}
1348ea8dc4b6Seschrock 
1349fa9e4066Sahrens 		mutex_exit(&db->db_mtx);
1350fa9e4066Sahrens 	}
1351fa9e4066Sahrens 	mutex_exit(&dn->dn_dbufs_mtx);
1352fa9e4066Sahrens }
1353fa9e4066Sahrens 
1354fa9e4066Sahrens void
1355fa9e4066Sahrens dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
1356fa9e4066Sahrens {
1357fa9e4066Sahrens 	arc_buf_t *buf, *obuf;
1358fa9e4066Sahrens 	int osize = db->db.db_size;
1359ad23a2dbSjohansen 	arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1360744947dcSTom Erickson 	dnode_t *dn;
1361fa9e4066Sahrens 
13620a586ceaSMark Shellenbaum 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1363ea8dc4b6Seschrock 
1364744947dcSTom Erickson 	DB_DNODE_ENTER(db);
1365744947dcSTom Erickson 	dn = DB_DNODE(db);
1366744947dcSTom Erickson 
1367fa9e4066Sahrens 	/* XXX does *this* func really need the lock? */
1368744947dcSTom Erickson 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1369fa9e4066Sahrens 
1370fa9e4066Sahrens 	/*
137143466aaeSMax Grossman 	 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1372fa9e4066Sahrens 	 * is OK, because there can be no other references to the db
1373fa9e4066Sahrens 	 * when we are changing its size, so no concurrent DB_FILL can
1374fa9e4066Sahrens 	 * be happening.
1375fa9e4066Sahrens 	 */
1376ea8dc4b6Seschrock 	/*
1377ea8dc4b6Seschrock 	 * XXX we should be doing a dbuf_read, checking the return
1378ea8dc4b6Seschrock 	 * value and returning that up to our callers
1379ea8dc4b6Seschrock 	 */
138043466aaeSMax Grossman 	dmu_buf_will_dirty(&db->db, tx);
1381fa9e4066Sahrens 
1382fa9e4066Sahrens 	/* create the data buffer for the new block */
13835602294fSDan Kimmel 	buf = arc_alloc_buf(dn->dn_objset->os_spa, db, type, size);
1384fa9e4066Sahrens 
1385fa9e4066Sahrens 	/* copy old block data to the new block */
1386fa9e4066Sahrens 	obuf = db->db_buf;
1387f65e61c0Sahrens 	bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
1388fa9e4066Sahrens 	/* zero the remainder */
1389f65e61c0Sahrens 	if (size > osize)
1390f65e61c0Sahrens 		bzero((uint8_t *)buf->b_data + osize, size - osize);
1391fa9e4066Sahrens 
1392fa9e4066Sahrens 	mutex_enter(&db->db_mtx);
1393fa9e4066Sahrens 	dbuf_set_data(db, buf);
1394dcbf3bd6SGeorge Wilson 	arc_buf_destroy(obuf, db);
1395fa9e4066Sahrens 	db->db.db_size = size;
1396fa9e4066Sahrens 
1397c717a561Smaybee 	if (db->db_level == 0) {
1398c717a561Smaybee 		ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1399c717a561Smaybee 		db->db_last_dirty->dt.dl.dr_data = buf;
1400c717a561Smaybee 	}
1401fa9e4066Sahrens 	mutex_exit(&db->db_mtx);
1402fa9e4066Sahrens 
140361e255ceSMatthew Ahrens 	dmu_objset_willuse_space(dn->dn_objset, size - osize, tx);
1404744947dcSTom Erickson 	DB_DNODE_EXIT(db);
1405fa9e4066Sahrens }
1406fa9e4066Sahrens 
14073f9d6ad7SLin Ling void
14083f9d6ad7SLin Ling dbuf_release_bp(dmu_buf_impl_t *db)
14093f9d6ad7SLin Ling {
141043466aaeSMax Grossman 	objset_t *os = db->db_objset;
14113f9d6ad7SLin Ling 
14123f9d6ad7SLin Ling 	ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
14133f9d6ad7SLin Ling 	ASSERT(arc_released(os->os_phys_buf) ||
14143f9d6ad7SLin Ling 	    list_link_active(&os->os_dsl_dataset->ds_synced_link));
14153f9d6ad7SLin Ling 	ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
14163f9d6ad7SLin Ling 
14171b912ec7SGeorge Wilson 	(void) arc_release(db->db_buf, db);
14183f9d6ad7SLin Ling }
14193f9d6ad7SLin Ling 
14200f2e7d03SMatthew Ahrens /*
14210f2e7d03SMatthew Ahrens  * We already have a dirty record for this TXG, and we are being
14220f2e7d03SMatthew Ahrens  * dirtied again.
14230f2e7d03SMatthew Ahrens  */
14240f2e7d03SMatthew Ahrens static void
14250f2e7d03SMatthew Ahrens dbuf_redirty(dbuf_dirty_record_t *dr)
14260f2e7d03SMatthew Ahrens {
14270f2e7d03SMatthew Ahrens 	dmu_buf_impl_t *db = dr->dr_dbuf;
14280f2e7d03SMatthew Ahrens 
14290f2e7d03SMatthew Ahrens 	ASSERT(MUTEX_HELD(&db->db_mtx));
14300f2e7d03SMatthew Ahrens 
14310f2e7d03SMatthew Ahrens 	if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
14320f2e7d03SMatthew Ahrens 		/*
14330f2e7d03SMatthew Ahrens 		 * If this buffer has already been written out,
14340f2e7d03SMatthew Ahrens 		 * we now need to reset its state.
14350f2e7d03SMatthew Ahrens 		 */
14360f2e7d03SMatthew Ahrens 		dbuf_unoverride(dr);
14370f2e7d03SMatthew Ahrens 		if (db->db.db_object != DMU_META_DNODE_OBJECT &&
14380f2e7d03SMatthew Ahrens 		    db->db_state != DB_NOFILL) {
14390f2e7d03SMatthew Ahrens 			/* Already released on initial dirty, so just thaw. */
14400f2e7d03SMatthew Ahrens 			ASSERT(arc_released(db->db_buf));
14410f2e7d03SMatthew Ahrens 			arc_buf_thaw(db->db_buf);
14420f2e7d03SMatthew Ahrens 		}
14430f2e7d03SMatthew Ahrens 	}
14440f2e7d03SMatthew Ahrens }
14450f2e7d03SMatthew Ahrens 
1446c717a561Smaybee dbuf_dirty_record_t *
1447fa9e4066Sahrens dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1448fa9e4066Sahrens {
1449744947dcSTom Erickson 	dnode_t *dn;
1450744947dcSTom Erickson 	objset_t *os;
1451c717a561Smaybee 	dbuf_dirty_record_t **drp, *dr;
1452fa9e4066Sahrens 	int drop_struct_lock = FALSE;
1453fa9e4066Sahrens 	int txgoff = tx->tx_txg & TXG_MASK;
1454fa9e4066Sahrens 
1455fa9e4066Sahrens 	ASSERT(tx->tx_txg != 0);
1456fa9e4066Sahrens 	ASSERT(!refcount_is_zero(&db->db_holds));
14579c9dc39aSek 	DMU_TX_DIRTY_BUF(tx, db);
1458fa9e4066Sahrens 
1459744947dcSTom Erickson 	DB_DNODE_ENTER(db);
1460744947dcSTom Erickson 	dn = DB_DNODE(db);
1461fa9e4066Sahrens 	/*
1462fa9e4066Sahrens 	 * Shouldn't dirty a regular buffer in syncing context.  Private
1463fa9e4066Sahrens 	 * objects may be dirtied in syncing context, but only if they
1464fa9e4066Sahrens 	 * were already pre-dirtied in open context.
1465fa9e4066Sahrens 	 */
1466c166b69dSPaul Dagnelie #ifdef DEBUG
1467c166b69dSPaul Dagnelie 	if (dn->dn_objset->os_dsl_dataset != NULL) {
1468c166b69dSPaul Dagnelie 		rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1469c166b69dSPaul Dagnelie 		    RW_READER, FTAG);
1470c166b69dSPaul Dagnelie 	}
1471c717a561Smaybee 	ASSERT(!dmu_tx_is_syncing(tx) ||
1472c717a561Smaybee 	    BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
147314843421SMatthew Ahrens 	    DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
147414843421SMatthew Ahrens 	    dn->dn_objset->os_dsl_dataset == NULL);
1475c166b69dSPaul Dagnelie 	if (dn->dn_objset->os_dsl_dataset != NULL)
1476c166b69dSPaul Dagnelie 		rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, FTAG);
1477c166b69dSPaul Dagnelie #endif
1478fa9e4066Sahrens 	/*
1479fa9e4066Sahrens 	 * We make this assert for private objects as well, but after we
1480fa9e4066Sahrens 	 * check if we're already dirty.  They are allowed to re-dirty
1481fa9e4066Sahrens 	 * in syncing context.
1482fa9e4066Sahrens 	 */
1483ea8dc4b6Seschrock 	ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1484c717a561Smaybee 	    dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1485fa9e4066Sahrens 	    (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1486fa9e4066Sahrens 
1487fa9e4066Sahrens 	mutex_enter(&db->db_mtx);
1488fa9e4066Sahrens 	/*
1489c717a561Smaybee 	 * XXX make this true for indirects too?  The problem is that
1490c717a561Smaybee 	 * transactions created with dmu_tx_create_assigned() from
1491c717a561Smaybee 	 * syncing context don't bother holding ahead.
1492fa9e4066Sahrens 	 */
1493c717a561Smaybee 	ASSERT(db->db_level != 0 ||
149482c9918fSTim Haley 	    db->db_state == DB_CACHED || db->db_state == DB_FILL ||
149582c9918fSTim Haley 	    db->db_state == DB_NOFILL);
1496fa9e4066Sahrens 
1497fa9e4066Sahrens 	mutex_enter(&dn->dn_mtx);
1498fa9e4066Sahrens 	/*
1499fa9e4066Sahrens 	 * Don't set dirtyctx to SYNC if we're just modifying this as we
1500fa9e4066Sahrens 	 * initialize the objset.
1501fa9e4066Sahrens 	 */
1502c166b69dSPaul Dagnelie 	if (dn->dn_dirtyctx == DN_UNDIRTIED) {
1503c166b69dSPaul Dagnelie 		if (dn->dn_objset->os_dsl_dataset != NULL) {
1504c166b69dSPaul Dagnelie 			rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1505c166b69dSPaul Dagnelie 			    RW_READER, FTAG);
1506c166b69dSPaul Dagnelie 		}
1507c166b69dSPaul Dagnelie 		if (!BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1508c166b69dSPaul Dagnelie 			dn->dn_dirtyctx = (dmu_tx_is_syncing(tx) ?
1509c166b69dSPaul Dagnelie 			    DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1510c166b69dSPaul Dagnelie 			ASSERT(dn->dn_dirtyctx_firstset == NULL);
1511c166b69dSPaul Dagnelie 			dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1512c166b69dSPaul Dagnelie 		}
1513c166b69dSPaul Dagnelie 		if (dn->dn_objset->os_dsl_dataset != NULL) {
1514c166b69dSPaul Dagnelie 			rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1515c166b69dSPaul Dagnelie 			    FTAG);
1516c166b69dSPaul Dagnelie 		}
1517fa9e4066Sahrens 	}
1518fa9e4066Sahrens 	mutex_exit(&dn->dn_mtx);
1519fa9e4066Sahrens 
15200a586ceaSMark Shellenbaum 	if (db->db_blkid == DMU_SPILL_BLKID)
15210a586ceaSMark Shellenbaum 		dn->dn_have_spill = B_TRUE;
15220a586ceaSMark Shellenbaum 
1523fa9e4066Sahrens 	/*
1524fa9e4066Sahrens 	 * If this buffer is already dirty, we're done.
1525fa9e4066Sahrens 	 */
1526c717a561Smaybee 	drp = &db->db_last_dirty;
1527c717a561Smaybee 	ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1528c717a561Smaybee 	    db->db.db_object == DMU_META_DNODE_OBJECT);
15297e2186e3Sbonwick 	while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
15307e2186e3Sbonwick 		drp = &dr->dr_next;
15317e2186e3Sbonwick 	if (dr && dr->dr_txg == tx->tx_txg) {
1532744947dcSTom Erickson 		DB_DNODE_EXIT(db);
1533744947dcSTom Erickson 
15340f2e7d03SMatthew Ahrens 		dbuf_redirty(dr);
1535fa9e4066Sahrens 		mutex_exit(&db->db_mtx);
15367e2186e3Sbonwick 		return (dr);
1537fa9e4066Sahrens 	}
1538fa9e4066Sahrens 
1539fa9e4066Sahrens 	/*
1540fa9e4066Sahrens 	 * Only valid if not already dirty.
1541fa9e4066Sahrens 	 */
154214843421SMatthew Ahrens 	ASSERT(dn->dn_object == 0 ||
154314843421SMatthew Ahrens 	    dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1544fa9e4066Sahrens 	    (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1545fa9e4066Sahrens 
1546fa9e4066Sahrens 	ASSERT3U(dn->dn_nlevels, >, db->db_level);
1547fa9e4066Sahrens 
1548fa9e4066Sahrens 	/*
1549fa9e4066Sahrens 	 * We should only be dirtying in syncing context if it's the
155014843421SMatthew Ahrens 	 * mos or we're initializing the os or it's a special object.
155114843421SMatthew Ahrens 	 * However, we are allowed to dirty in syncing context provided
155214843421SMatthew Ahrens 	 * we already dirtied it in open context.  Hence we must make
155314843421SMatthew Ahrens 	 * this assertion only if we're not already dirty.
1554fa9e4066Sahrens 	 */
1555744947dcSTom Erickson 	os = dn->dn_objset;
15563991b535SGeorge Wilson 	VERIFY3U(tx->tx_txg, <=, spa_final_dirty_txg(os->os_spa));
1557c166b69dSPaul Dagnelie #ifdef DEBUG
1558c166b69dSPaul Dagnelie 	if (dn->dn_objset->os_dsl_dataset != NULL)
1559c166b69dSPaul Dagnelie 		rrw_enter(&os->os_dsl_dataset->ds_bp_rwlock, RW_READER, FTAG);
156014843421SMatthew Ahrens 	ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
156114843421SMatthew Ahrens 	    os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1562c166b69dSPaul Dagnelie 	if (dn->dn_objset->os_dsl_dataset != NULL)
1563c166b69dSPaul Dagnelie 		rrw_exit(&os->os_dsl_dataset->ds_bp_rwlock, FTAG);
1564c166b69dSPaul Dagnelie #endif
1565fa9e4066Sahrens 	ASSERT(db->db.db_size != 0);
1566fa9e4066Sahrens 
1567fa9e4066Sahrens 	dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1568fa9e4066Sahrens 
15690a586ceaSMark Shellenbaum 	if (db->db_blkid != DMU_BONUS_BLKID) {
157061e255ceSMatthew Ahrens 		dmu_objset_willuse_space(os, db->db.db_size, tx);
15711934e92fSmaybee 	}
15721934e92fSmaybee 
1573ea8dc4b6Seschrock 	/*
1574ea8dc4b6Seschrock 	 * If this buffer is dirty in an old transaction group we need
1575ea8dc4b6Seschrock 	 * to make a copy of it so that the changes we make in this
1576ea8dc4b6Seschrock 	 * transaction group won't leak out when we sync the older txg.
1577ea8dc4b6Seschrock 	 */
1578c717a561Smaybee 	dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1579c717a561Smaybee 	if (db->db_level == 0) {
1580c717a561Smaybee 		void *data_old = db->db_buf;
1581c717a561Smaybee 
158282c9918fSTim Haley 		if (db->db_state != DB_NOFILL) {
15830a586ceaSMark Shellenbaum 			if (db->db_blkid == DMU_BONUS_BLKID) {
158482c9918fSTim Haley 				dbuf_fix_old_data(db, tx->tx_txg);
158582c9918fSTim Haley 				data_old = db->db.db_data;
158682c9918fSTim Haley 			} else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
158782c9918fSTim Haley 				/*
158882c9918fSTim Haley 				 * Release the data buffer from the cache so
158982c9918fSTim Haley 				 * that we can modify it without impacting
159082c9918fSTim Haley 				 * possible other users of this cached data
159182c9918fSTim Haley 				 * block.  Note that indirect blocks and
159282c9918fSTim Haley 				 * private objects are not released until the
159382c9918fSTim Haley 				 * syncing state (since they are only modified
159482c9918fSTim Haley 				 * then).
159582c9918fSTim Haley 				 */
159682c9918fSTim Haley 				arc_release(db->db_buf, db);
159782c9918fSTim Haley 				dbuf_fix_old_data(db, tx->tx_txg);
159882c9918fSTim Haley 				data_old = db->db_buf;
159982c9918fSTim Haley 			}
160082c9918fSTim Haley 			ASSERT(data_old != NULL);
1601fa9e4066Sahrens 		}
1602c717a561Smaybee 		dr->dt.dl.dr_data = data_old;
1603c717a561Smaybee 	} else {
1604c717a561Smaybee 		mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1605c717a561Smaybee 		list_create(&dr->dt.di.dr_children,
1606c717a561Smaybee 		    sizeof (dbuf_dirty_record_t),
1607c717a561Smaybee 		    offsetof(dbuf_dirty_record_t, dr_dirty_node));
1608fa9e4066Sahrens 	}
160969962b56SMatthew Ahrens 	if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL)
161069962b56SMatthew Ahrens 		dr->dr_accounted = db->db.db_size;
1611c717a561Smaybee 	dr->dr_dbuf = db;
1612c717a561Smaybee 	dr->dr_txg = tx->tx_txg;
1613c717a561Smaybee 	dr->dr_next = *drp;
1614c717a561Smaybee 	*drp = dr;
1615fa9e4066Sahrens 
1616fa9e4066Sahrens 	/*
1617fa9e4066Sahrens 	 * We could have been freed_in_flight between the dbuf_noread
1618fa9e4066Sahrens 	 * and dbuf_dirty.  We win, as though the dbuf_noread() had
1619fa9e4066Sahrens 	 * happened after the free.
1620fa9e4066Sahrens 	 */
16210a586ceaSMark Shellenbaum 	if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
16220a586ceaSMark Shellenbaum 	    db->db_blkid != DMU_SPILL_BLKID) {
1623c717a561Smaybee 		mutex_enter(&dn->dn_mtx);
1624bf16b11eSMatthew Ahrens 		if (dn->dn_free_ranges[txgoff] != NULL) {
1625bf16b11eSMatthew Ahrens 			range_tree_clear(dn->dn_free_ranges[txgoff],
1626bf16b11eSMatthew Ahrens 			    db->db_blkid, 1);
1627bf16b11eSMatthew Ahrens 		}
1628c717a561Smaybee 		mutex_exit(&dn->dn_mtx);
1629c717a561Smaybee 		db->db_freed_in_flight = FALSE;
1630fa9e4066Sahrens 	}
1631fa9e4066Sahrens 
1632fa9e4066Sahrens 	/*
1633fa9e4066Sahrens 	 * This buffer is now part of this txg
1634fa9e4066Sahrens 	 */
1635fa9e4066Sahrens 	dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1636fa9e4066Sahrens 	db->db_dirtycnt += 1;
1637fa9e4066Sahrens 	ASSERT3U(db->db_dirtycnt, <=, 3);
1638fa9e4066Sahrens 
1639fa9e4066Sahrens 	mutex_exit(&db->db_mtx);
1640fa9e4066Sahrens 
16410a586ceaSMark Shellenbaum 	if (db->db_blkid == DMU_BONUS_BLKID ||
16420a586ceaSMark Shellenbaum 	    db->db_blkid == DMU_SPILL_BLKID) {
1643c717a561Smaybee 		mutex_enter(&dn->dn_mtx);
1644c717a561Smaybee 		ASSERT(!list_link_active(&dr->dr_dirty_node));
1645c717a561Smaybee 		list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1646c717a561Smaybee 		mutex_exit(&dn->dn_mtx);
1647fa9e4066Sahrens 		dnode_setdirty(dn, tx);
1648744947dcSTom Erickson 		DB_DNODE_EXIT(db);
1649c717a561Smaybee 		return (dr);
165092654925SMatthew Ahrens 	}
165192654925SMatthew Ahrens 
165292654925SMatthew Ahrens 	/*
165392654925SMatthew Ahrens 	 * The dn_struct_rwlock prevents db_blkptr from changing
165492654925SMatthew Ahrens 	 * due to a write from syncing context completing
165592654925SMatthew Ahrens 	 * while we are running, so we want to acquire it before
165692654925SMatthew Ahrens 	 * looking at db_blkptr.
165792654925SMatthew Ahrens 	 */
165892654925SMatthew Ahrens 	if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
165992654925SMatthew Ahrens 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
166092654925SMatthew Ahrens 		drop_struct_lock = TRUE;
166192654925SMatthew Ahrens 	}
166292654925SMatthew Ahrens 
1663dcb6872cSMatthew Ahrens 	/*
1664dcb6872cSMatthew Ahrens 	 * We need to hold the dn_struct_rwlock to make this assertion,
1665dcb6872cSMatthew Ahrens 	 * because it protects dn_phys / dn_next_nlevels from changing.
1666dcb6872cSMatthew Ahrens 	 */
1667dcb6872cSMatthew Ahrens 	ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1668dcb6872cSMatthew Ahrens 	    dn->dn_phys->dn_nlevels > db->db_level ||
1669dcb6872cSMatthew Ahrens 	    dn->dn_next_nlevels[txgoff] > db->db_level ||
1670dcb6872cSMatthew Ahrens 	    dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1671dcb6872cSMatthew Ahrens 	    dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1672dcb6872cSMatthew Ahrens 
167361e255ceSMatthew Ahrens 	/*
167461e255ceSMatthew Ahrens 	 * If we are overwriting a dedup BP, then unless it is snapshotted,
167561e255ceSMatthew Ahrens 	 * when we get to syncing context we will need to decrement its
167661e255ceSMatthew Ahrens 	 * refcount in the DDT.  Prefetch the relevant DDT block so that
167761e255ceSMatthew Ahrens 	 * syncing context won't have to wait for the i/o.
167861e255ceSMatthew Ahrens 	 */
167961e255ceSMatthew Ahrens 	ddt_prefetch(os->os_spa, db->db_blkptr);
1680fa9e4066Sahrens 
16818346f03fSJonathan W Adams 	if (db->db_level == 0) {
16828346f03fSJonathan W Adams 		dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
16838346f03fSJonathan W Adams 		ASSERT(dn->dn_maxblkid >= db->db_blkid);
16848346f03fSJonathan W Adams 	}
16858346f03fSJonathan W Adams 
168644eda4d7Smaybee 	if (db->db_level+1 < dn->dn_nlevels) {
1687c717a561Smaybee 		dmu_buf_impl_t *parent = db->db_parent;
1688c717a561Smaybee 		dbuf_dirty_record_t *di;
1689c717a561Smaybee 		int parent_held = FALSE;
1690c717a561Smaybee 
1691c717a561Smaybee 		if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1692c717a561Smaybee 			int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1693c717a561Smaybee 
1694c717a561Smaybee 			parent = dbuf_hold_level(dn, db->db_level+1,
1695c717a561Smaybee 			    db->db_blkid >> epbs, FTAG);
169601025c89SJohn Harres 			ASSERT(parent != NULL);
1697c717a561Smaybee 			parent_held = TRUE;
1698c717a561Smaybee 		}
1699fa9e4066Sahrens 		if (drop_struct_lock)
1700fa9e4066Sahrens 			rw_exit(&dn->dn_struct_rwlock);
1701c717a561Smaybee 		ASSERT3U(db->db_level+1, ==, parent->db_level);
1702c717a561Smaybee 		di = dbuf_dirty(parent, tx);
1703c717a561Smaybee 		if (parent_held)
1704c717a561Smaybee 			dbuf_rele(parent, FTAG);
1705c717a561Smaybee 
1706c717a561Smaybee 		mutex_enter(&db->db_mtx);
170769962b56SMatthew Ahrens 		/*
170869962b56SMatthew Ahrens 		 * Since we've dropped the mutex, it's possible that
170969962b56SMatthew Ahrens 		 * dbuf_undirty() might have changed this out from under us.
171069962b56SMatthew Ahrens 		 */
1711c717a561Smaybee 		if (db->db_last_dirty == dr ||
1712c717a561Smaybee 		    dn->dn_object == DMU_META_DNODE_OBJECT) {
1713c717a561Smaybee 			mutex_enter(&di->dt.di.dr_mtx);
1714c717a561Smaybee 			ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1715c717a561Smaybee 			ASSERT(!list_link_active(&dr->dr_dirty_node));
1716c717a561Smaybee 			list_insert_tail(&di->dt.di.dr_children, dr);
1717c717a561Smaybee 			mutex_exit(&di->dt.di.dr_mtx);
1718c717a561Smaybee 			dr->dr_parent = di;
1719c717a561Smaybee 		}
1720c717a561Smaybee 		mutex_exit(&db->db_mtx);
1721fa9e4066Sahrens 	} else {
1722c717a561Smaybee 		ASSERT(db->db_level+1 == dn->dn_nlevels);
1723c717a561Smaybee 		ASSERT(db->db_blkid < dn->dn_nblkptr);
1724744947dcSTom Erickson 		ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1725c717a561Smaybee 		mutex_enter(&dn->dn_mtx);
1726c717a561Smaybee 		ASSERT(!list_link_active(&dr->dr_dirty_node));
1727c717a561Smaybee 		list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1728c717a561Smaybee 		mutex_exit(&dn->dn_mtx);
1729fa9e4066Sahrens 		if (drop_struct_lock)
1730fa9e4066Sahrens 			rw_exit(&dn->dn_struct_rwlock);
1731fa9e4066Sahrens 	}
1732fa9e4066Sahrens 
1733fa9e4066Sahrens 	dnode_setdirty(dn, tx);
1734744947dcSTom Erickson 	DB_DNODE_EXIT(db);
1735c717a561Smaybee 	return (dr);
1736fa9e4066Sahrens }
1737fa9e4066Sahrens 
17383b2aab18SMatthew Ahrens /*
17393e30c24aSWill Andrews  * Undirty a buffer in the transaction group referenced by the given
17403e30c24aSWill Andrews  * transaction.  Return whether this evicted the dbuf.
17413b2aab18SMatthew Ahrens  */
17423b2aab18SMatthew Ahrens static boolean_t
1743fa9e4066Sahrens dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1744fa9e4066Sahrens {
1745744947dcSTom Erickson 	dnode_t *dn;
1746c717a561Smaybee 	uint64_t txg = tx->tx_txg;
174717f17c2dSbonwick 	dbuf_dirty_record_t *dr, **drp;
1748fa9e4066Sahrens 
1749c717a561Smaybee 	ASSERT(txg != 0);
175046e1baa6SMatthew Ahrens 
175146e1baa6SMatthew Ahrens 	/*
175246e1baa6SMatthew Ahrens 	 * Due to our use of dn_nlevels below, this can only be called
175346e1baa6SMatthew Ahrens 	 * in open context, unless we are operating on the MOS.
175446e1baa6SMatthew Ahrens 	 * From syncing context, dn_nlevels may be different from the
175546e1baa6SMatthew Ahrens 	 * dn_nlevels used when dbuf was dirtied.
175646e1baa6SMatthew Ahrens 	 */
175746e1baa6SMatthew Ahrens 	ASSERT(db->db_objset ==
175846e1baa6SMatthew Ahrens 	    dmu_objset_pool(db->db_objset)->dp_meta_objset ||
175946e1baa6SMatthew Ahrens 	    txg != spa_syncing_txg(dmu_objset_spa(db->db_objset)));
17600a586ceaSMark Shellenbaum 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
17613b2aab18SMatthew Ahrens 	ASSERT0(db->db_level);
17623b2aab18SMatthew Ahrens 	ASSERT(MUTEX_HELD(&db->db_mtx));
1763fa9e4066Sahrens 
1764fa9e4066Sahrens 	/*
1765fa9e4066Sahrens 	 * If this buffer is not dirty, we're done.
1766fa9e4066Sahrens 	 */
176717f17c2dSbonwick 	for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1768c717a561Smaybee 		if (dr->dr_txg <= txg)
1769c717a561Smaybee 			break;
17703b2aab18SMatthew Ahrens 	if (dr == NULL || dr->dr_txg < txg)
17713b2aab18SMatthew Ahrens 		return (B_FALSE);
1772c717a561Smaybee 	ASSERT(dr->dr_txg == txg);
1773b24ab676SJeff Bonwick 	ASSERT(dr->dr_dbuf == db);
1774fa9e4066Sahrens 
1775744947dcSTom Erickson 	DB_DNODE_ENTER(db);
1776744947dcSTom Erickson 	dn = DB_DNODE(db);
1777744947dcSTom Erickson 
1778fa9e4066Sahrens 	dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1779fa9e4066Sahrens 
1780fa9e4066Sahrens 	ASSERT(db->db.db_size != 0);
1781fa9e4066Sahrens 
178246e1baa6SMatthew Ahrens 	dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset),
178346e1baa6SMatthew Ahrens 	    dr->dr_accounted, txg);
1784fa9e4066Sahrens 
178517f17c2dSbonwick 	*drp = dr->dr_next;
1786c717a561Smaybee 
17873f2366c2SGordon Ross 	/*
17883f2366c2SGordon Ross 	 * Note that there are three places in dbuf_dirty()
17893f2366c2SGordon Ross 	 * where this dirty record may be put on a list.
17903f2366c2SGordon Ross 	 * Make sure to do a list_remove corresponding to
17913f2366c2SGordon Ross 	 * every one of those list_insert calls.
17923f2366c2SGordon Ross 	 */
1793c717a561Smaybee 	if (dr->dr_parent) {
1794c717a561Smaybee 		mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1795c717a561Smaybee 		list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1796c717a561Smaybee 		mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
17973f2366c2SGordon Ross 	} else if (db->db_blkid == DMU_SPILL_BLKID ||
179846e1baa6SMatthew Ahrens 	    db->db_level + 1 == dn->dn_nlevels) {
1799cdb0ab79Smaybee 		ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1800c717a561Smaybee 		mutex_enter(&dn->dn_mtx);
1801c717a561Smaybee 		list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1802c717a561Smaybee 		mutex_exit(&dn->dn_mtx);
1803c717a561Smaybee 	}
1804744947dcSTom Erickson 	DB_DNODE_EXIT(db);
1805c717a561Smaybee 
18063b2aab18SMatthew Ahrens 	if (db->db_state != DB_NOFILL) {
18073b2aab18SMatthew Ahrens 		dbuf_unoverride(dr);
1808c717a561Smaybee 
1809c717a561Smaybee 		ASSERT(db->db_buf != NULL);
18103b2aab18SMatthew Ahrens 		ASSERT(dr->dt.dl.dr_data != NULL);
18113b2aab18SMatthew Ahrens 		if (dr->dt.dl.dr_data != db->db_buf)
1812dcbf3bd6SGeorge Wilson 			arc_buf_destroy(dr->dt.dl.dr_data, db);
1813c717a561Smaybee 	}
1814d2b3cbbdSJorgen Lundman 
1815c717a561Smaybee 	kmem_free(dr, sizeof (dbuf_dirty_record_t));
1816fa9e4066Sahrens 
1817fa9e4066Sahrens 	ASSERT(db->db_dirtycnt > 0);
1818fa9e4066Sahrens 	db->db_dirtycnt -= 1;
1819fa9e4066Sahrens 
1820c717a561Smaybee 	if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1821dcbf3bd6SGeorge Wilson 		ASSERT(db->db_state == DB_NOFILL || arc_released(db->db_buf));
1822dcbf3bd6SGeorge Wilson 		dbuf_destroy(db);
18233b2aab18SMatthew Ahrens 		return (B_TRUE);
1824fa9e4066Sahrens 	}
1825fa9e4066Sahrens 
18263b2aab18SMatthew Ahrens 	return (B_FALSE);
1827fa9e4066Sahrens }
1828fa9e4066Sahrens 
1829fa9e4066Sahrens void
183043466aaeSMax Grossman dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
1831fa9e4066Sahrens {
183243466aaeSMax Grossman 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
18331ab7f2deSmaybee 	int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1834fa9e4066Sahrens 
1835fa9e4066Sahrens 	ASSERT(tx->tx_txg != 0);
1836fa9e4066Sahrens 	ASSERT(!refcount_is_zero(&db->db_holds));
1837fa9e4066Sahrens 
18380f2e7d03SMatthew Ahrens 	/*
18390f2e7d03SMatthew Ahrens 	 * Quick check for dirtyness.  For already dirty blocks, this
18400f2e7d03SMatthew Ahrens 	 * reduces runtime of this function by >90%, and overall performance
18410f2e7d03SMatthew Ahrens 	 * by 50% for some workloads (e.g. file deletion with indirect blocks
18420f2e7d03SMatthew Ahrens 	 * cached).
18430f2e7d03SMatthew Ahrens 	 */
18440f2e7d03SMatthew Ahrens 	mutex_enter(&db->db_mtx);
18450f2e7d03SMatthew Ahrens 	dbuf_dirty_record_t *dr;
18460f2e7d03SMatthew Ahrens 	for (dr = db->db_last_dirty;
18470f2e7d03SMatthew Ahrens 	    dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) {
18480f2e7d03SMatthew Ahrens 		/*
18490f2e7d03SMatthew Ahrens 		 * It's possible that it is already dirty but not cached,
18500f2e7d03SMatthew Ahrens 		 * because there are some calls to dbuf_dirty() that don't
18510f2e7d03SMatthew Ahrens 		 * go through dmu_buf_will_dirty().
18520f2e7d03SMatthew Ahrens 		 */
18530f2e7d03SMatthew Ahrens 		if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) {
18540f2e7d03SMatthew Ahrens 			/* This dbuf is already dirty and cached. */
18550f2e7d03SMatthew Ahrens 			dbuf_redirty(dr);
18560f2e7d03SMatthew Ahrens 			mutex_exit(&db->db_mtx);
18570f2e7d03SMatthew Ahrens 			return;
18580f2e7d03SMatthew Ahrens 		}
18590f2e7d03SMatthew Ahrens 	}
18600f2e7d03SMatthew Ahrens 	mutex_exit(&db->db_mtx);
18610f2e7d03SMatthew Ahrens 
1862744947dcSTom Erickson 	DB_DNODE_ENTER(db);
1863744947dcSTom Erickson 	if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
1864fa9e4066Sahrens 		rf |= DB_RF_HAVESTRUCT;
1865744947dcSTom Erickson 	DB_DNODE_EXIT(db);
1866ea8dc4b6Seschrock 	(void) dbuf_read(db, NULL, rf);
1867c717a561Smaybee 	(void) dbuf_dirty(db, tx);
1868fa9e4066Sahrens }
1869fa9e4066Sahrens 
187082c9918fSTim Haley void
187182c9918fSTim Haley dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
187282c9918fSTim Haley {
187382c9918fSTim Haley 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
187482c9918fSTim Haley 
187582c9918fSTim Haley 	db->db_state = DB_NOFILL;
187682c9918fSTim Haley 
187782c9918fSTim Haley 	dmu_buf_will_fill(db_fake, tx);
187882c9918fSTim Haley }
187982c9918fSTim Haley 
1880fa9e4066Sahrens void
1881ea8dc4b6Seschrock dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1882fa9e4066Sahrens {
1883ea8dc4b6Seschrock 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1884ea8dc4b6Seschrock 
18850a586ceaSMark Shellenbaum 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1886fa9e4066Sahrens 	ASSERT(tx->tx_txg != 0);
1887fa9e4066Sahrens 	ASSERT(db->db_level == 0);
1888fa9e4066Sahrens 	ASSERT(!refcount_is_zero(&db->db_holds));
1889fa9e4066Sahrens 
1890ea8dc4b6Seschrock 	ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1891fa9e4066Sahrens 	    dmu_tx_private_ok(tx));
1892fa9e4066Sahrens 
1893fa9e4066Sahrens 	dbuf_noread(db);
1894c717a561Smaybee 	(void) dbuf_dirty(db, tx);
1895fa9e4066Sahrens }
1896fa9e4066Sahrens 
1897fa9e4066Sahrens #pragma weak dmu_buf_fill_done = dbuf_fill_done
1898fa9e4066Sahrens /* ARGSUSED */
1899fa9e4066Sahrens void
1900fa9e4066Sahrens dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1901fa9e4066Sahrens {
1902fa9e4066Sahrens 	mutex_enter(&db->db_mtx);
19039c9dc39aSek 	DBUF_VERIFY(db);
1904fa9e4066Sahrens 
1905fa9e4066Sahrens 	if (db->db_state == DB_FILL) {
1906c717a561Smaybee 		if (db->db_level == 0 && db->db_freed_in_flight) {
19070a586ceaSMark Shellenbaum 			ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1908fa9e4066Sahrens 			/* we were freed while filling */
1909fa9e4066Sahrens 			/* XXX dbuf_undirty? */
1910fa9e4066Sahrens 			bzero(db->db.db_data, db->db.db_size);
1911c717a561Smaybee 			db->db_freed_in_flight = FALSE;
1912fa9e4066Sahrens 		}
1913fa9e4066Sahrens 		db->db_state = DB_CACHED;
1914fa9e4066Sahrens 		cv_broadcast(&db->db_changed);
1915fa9e4066Sahrens 	}
1916fa9e4066Sahrens 	mutex_exit(&db->db_mtx);
1917fa9e4066Sahrens }
1918fa9e4066Sahrens 
19195d7b4d43SMatthew Ahrens void
19205d7b4d43SMatthew Ahrens dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data,
19215d7b4d43SMatthew Ahrens     bp_embedded_type_t etype, enum zio_compress comp,
19225d7b4d43SMatthew Ahrens     int uncompressed_size, int compressed_size, int byteorder,
19235d7b4d43SMatthew Ahrens     dmu_tx_t *tx)
19245d7b4d43SMatthew Ahrens {
19255d7b4d43SMatthew Ahrens 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
19265d7b4d43SMatthew Ahrens 	struct dirty_leaf *dl;
19275d7b4d43SMatthew Ahrens 	dmu_object_type_t type;
19285d7b4d43SMatthew Ahrens 
1929ca0cc391SMatthew Ahrens 	if (etype == BP_EMBEDDED_TYPE_DATA) {
1930ca0cc391SMatthew Ahrens 		ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset),
1931ca0cc391SMatthew Ahrens 		    SPA_FEATURE_EMBEDDED_DATA));
1932ca0cc391SMatthew Ahrens 	}
1933ca0cc391SMatthew Ahrens 
19345d7b4d43SMatthew Ahrens 	DB_DNODE_ENTER(db);
19355d7b4d43SMatthew Ahrens 	type = DB_DNODE(db)->dn_type;
19365d7b4d43SMatthew Ahrens 	DB_DNODE_EXIT(db);
19375d7b4d43SMatthew Ahrens 
19385d7b4d43SMatthew Ahrens 	ASSERT0(db->db_level);
19395d7b4d43SMatthew Ahrens 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
19405d7b4d43SMatthew Ahrens 
19415d7b4d43SMatthew Ahrens 	dmu_buf_will_not_fill(dbuf, tx);
19425d7b4d43SMatthew Ahrens 
19435d7b4d43SMatthew Ahrens 	ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
19445d7b4d43SMatthew Ahrens 	dl = &db->db_last_dirty->dt.dl;
19455d7b4d43SMatthew Ahrens 	encode_embedded_bp_compressed(&dl->dr_overridden_by,
19465d7b4d43SMatthew Ahrens 	    data, comp, uncompressed_size, compressed_size);
19475d7b4d43SMatthew Ahrens 	BPE_SET_ETYPE(&dl->dr_overridden_by, etype);
19485d7b4d43SMatthew Ahrens 	BP_SET_TYPE(&dl->dr_overridden_by, type);
19495d7b4d43SMatthew Ahrens 	BP_SET_LEVEL(&dl->dr_overridden_by, 0);
19505d7b4d43SMatthew Ahrens 	BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder);
19515d7b4d43SMatthew Ahrens 
19525d7b4d43SMatthew Ahrens 	dl->dr_override_state = DR_OVERRIDDEN;
19535d7b4d43SMatthew Ahrens 	dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg;
19545d7b4d43SMatthew Ahrens }
19555d7b4d43SMatthew Ahrens 
19562fdbea25SAleksandr Guzovskiy /*
19572fdbea25SAleksandr Guzovskiy  * Directly assign a provided arc buf to a given dbuf if it's not referenced
19582fdbea25SAleksandr Guzovskiy  * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
19592fdbea25SAleksandr Guzovskiy  */
19602fdbea25SAleksandr Guzovskiy void
19612fdbea25SAleksandr Guzovskiy dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
19622fdbea25SAleksandr Guzovskiy {
19632fdbea25SAleksandr Guzovskiy 	ASSERT(!refcount_is_zero(&db->db_holds));
19640a586ceaSMark Shellenbaum 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
19652fdbea25SAleksandr Guzovskiy 	ASSERT(db->db_level == 0);
19665602294fSDan Kimmel 	ASSERT3U(dbuf_is_metadata(db), ==, arc_is_metadata(buf));
19672fdbea25SAleksandr Guzovskiy 	ASSERT(buf != NULL);
19685602294fSDan Kimmel 	ASSERT(arc_buf_lsize(buf) == db->db.db_size);
19692fdbea25SAleksandr Guzovskiy 	ASSERT(tx->tx_txg != 0);
19702fdbea25SAleksandr Guzovskiy 
19712fdbea25SAleksandr Guzovskiy 	arc_return_buf(buf, db);
19722fdbea25SAleksandr Guzovskiy 	ASSERT(arc_released(buf));
19732fdbea25SAleksandr Guzovskiy 
19742fdbea25SAleksandr Guzovskiy 	mutex_enter(&db->db_mtx);
19752fdbea25SAleksandr Guzovskiy 
19762fdbea25SAleksandr Guzovskiy 	while (db->db_state == DB_READ || db->db_state == DB_FILL)
19772fdbea25SAleksandr Guzovskiy 		cv_wait(&db->db_changed, &db->db_mtx);
19782fdbea25SAleksandr Guzovskiy 
19792fdbea25SAleksandr Guzovskiy 	ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
19802fdbea25SAleksandr Guzovskiy 
19812fdbea25SAleksandr Guzovskiy 	if (db->db_state == DB_CACHED &&
19822fdbea25SAleksandr Guzovskiy 	    refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
19832fdbea25SAleksandr Guzovskiy 		mutex_exit(&db->db_mtx);
19842fdbea25SAleksandr Guzovskiy 		(void) dbuf_dirty(db, tx);
19852fdbea25SAleksandr Guzovskiy 		bcopy(buf->b_data, db->db.db_data, db->db.db_size);
1986dcbf3bd6SGeorge Wilson 		arc_buf_destroy(buf, db);
1987c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 		xuio_stat_wbuf_copied();
19882fdbea25SAleksandr Guzovskiy 		return;
19892fdbea25SAleksandr Guzovskiy 	}
19902fdbea25SAleksandr Guzovskiy 
1991c242f9a0Schunli zhang - Sun Microsystems - Irvine United States 	xuio_stat_wbuf_nocopy();
19922fdbea25SAleksandr Guzovskiy 	if (db->db_state == DB_CACHED) {
19932fdbea25SAleksandr Guzovskiy 		dbuf_dirty_record_t *dr = db->db_last_dirty;
19942fdbea25SAleksandr Guzovskiy 
19952fdbea25SAleksandr Guzovskiy 		ASSERT(db->db_buf != NULL);
19962fdbea25SAleksandr Guzovskiy 		if (dr != NULL && dr->dr_txg == tx->tx_txg) {
19972fdbea25SAleksandr Guzovskiy 			ASSERT(dr->dt.dl.dr_data == db->db_buf);
19982fdbea25SAleksandr Guzovskiy 			if (!arc_released(db->db_buf)) {
19992fdbea25SAleksandr Guzovskiy 				ASSERT(dr->dt.dl.dr_override_state ==
20002fdbea25SAleksandr Guzovskiy 				    DR_OVERRIDDEN);
20012fdbea25SAleksandr Guzovskiy 				arc_release(db->db_buf, db);
20022fdbea25SAleksandr Guzovskiy 			}
20032fdbea25SAleksandr Guzovskiy 			dr->dt.dl.dr_data = buf;
2004dcbf3bd6SGeorge Wilson 			arc_buf_destroy(db->db_buf, db);
20052fdbea25SAleksandr Guzovskiy 		} else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
20062fdbea25SAleksandr Guzovskiy 			arc_release(db->db_buf, db);
2007dcbf3bd6SGeorge Wilson 			arc_buf_destroy(db->db_buf, db);
20082fdbea25SAleksandr Guzovskiy 		}
20092fdbea25SAleksandr Guzovskiy 		db->db_buf = NULL;
20102fdbea25SAleksandr Guzovskiy 	}
20112fdbea25SAleksandr Guzovskiy 	ASSERT(db->db_buf == NULL);
20122fdbea25SAleksandr Guzovskiy 	dbuf_set_data(db, buf);
20132fdbea25SAleksandr Guzovskiy 	db->db_state = DB_FILL;
20142fdbea25SAleksandr Guzovskiy 	mutex_exit(&db->db_mtx);
20152fdbea25SAleksandr Guzovskiy 	(void) dbuf_dirty(db, tx);
201643466aaeSMax Grossman 	dmu_buf_fill_done(&db->db, tx);
20172fdbea25SAleksandr Guzovskiy }
20182fdbea25SAleksandr Guzovskiy 
2019ea8dc4b6Seschrock void
2020dcbf3bd6SGeorge Wilson dbuf_destroy(dmu_buf_impl_t *db)
2021fa9e4066Sahrens {
2022744947dcSTom Erickson 	dnode_t *dn;
2023ea8dc4b6Seschrock 	dmu_buf_impl_t *parent = db->db_parent;
2024744947dcSTom Erickson 	dmu_buf_impl_t *dndb;
2025fa9e4066Sahrens 
2026fa9e4066Sahrens 	ASSERT(MUTEX_HELD(&db->db_mtx));
2027fa9e4066Sahrens 	ASSERT(refcount_is_zero(&db->db_holds));
2028fa9e4066Sahrens 
2029dcbf3bd6SGeorge Wilson 	if (db->db_buf != NULL) {
2030dcbf3bd6SGeorge Wilson 		arc_buf_destroy(db->db_buf, db);
2031dcbf3bd6SGeorge Wilson 		db->db_buf = NULL;
2032dcbf3bd6SGeorge Wilson 	}
2033ea8dc4b6Seschrock 
2034dcbf3bd6SGeorge Wilson 	if (db->db_blkid == DMU_BONUS_BLKID) {
2035ea8dc4b6Seschrock 		ASSERT(db->db.db_data != NULL);
2036dcbf3bd6SGeorge Wilson 		zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
2037dcbf3bd6SGeorge Wilson 		arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
2038fa9e4066Sahrens 		db->db_state = DB_UNCACHED;
2039fa9e4066Sahrens 	}
2040fa9e4066Sahrens 
2041dcbf3bd6SGeorge Wilson 	dbuf_clear_data(db);
2042dcbf3bd6SGeorge Wilson 
2043dcbf3bd6SGeorge Wilson 	if (multilist_link_active(&db->db_cache_link)) {
204494c2d0ebSMatthew Ahrens 		multilist_remove(dbuf_cache, db);
2045dcbf3bd6SGeorge Wilson 		(void) refcount_remove_many(&dbuf_cache_size,
2046dcbf3bd6SGeorge Wilson 		    db->db.db_size, db);
2047dcbf3bd6SGeorge Wilson 	}
2048dcbf3bd6SGeorge Wilson 
204982c9918fSTim Haley 	ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
2050fa9e4066Sahrens 	ASSERT(db->db_data_pending == NULL);
2051fa9e4066Sahrens 
2052ea8dc4b6Seschrock 	db->db_state = DB_EVICTING;
2053ea8dc4b6Seschrock 	db->db_blkptr = NULL;
2054ea8dc4b6Seschrock 
2055dcbf3bd6SGeorge Wilson 	/*
2056dcbf3bd6SGeorge Wilson 	 * Now that db_state is DB_EVICTING, nobody else can find this via
2057dcbf3bd6SGeorge Wilson 	 * the hash table.  We can now drop db_mtx, which allows us to
2058dcbf3bd6SGeorge Wilson 	 * acquire the dn_dbufs_mtx.
2059dcbf3bd6SGeorge Wilson 	 */
2060dcbf3bd6SGeorge Wilson 	mutex_exit(&db->db_mtx);
2061dcbf3bd6SGeorge Wilson 
2062744947dcSTom Erickson 	DB_DNODE_ENTER(db);
2063744947dcSTom Erickson 	dn = DB_DNODE(db);
2064744947dcSTom Erickson 	dndb = dn->dn_dbuf;
2065dcbf3bd6SGeorge Wilson 	if (db->db_blkid != DMU_BONUS_BLKID) {
2066dcbf3bd6SGeorge Wilson 		boolean_t needlock = !MUTEX_HELD(&dn->dn_dbufs_mtx);
2067dcbf3bd6SGeorge Wilson 		if (needlock)
2068dcbf3bd6SGeorge Wilson 			mutex_enter(&dn->dn_dbufs_mtx);
20690f6d88adSAlex Reece 		avl_remove(&dn->dn_dbufs, db);
2070640c1670SJosef 'Jeff' Sipek 		atomic_dec_32(&dn->dn_dbufs_count);
2071744947dcSTom Erickson 		membar_producer();
2072744947dcSTom Erickson 		DB_DNODE_EXIT(db);
2073dcbf3bd6SGeorge Wilson 		if (needlock)
2074dcbf3bd6SGeorge Wilson 			mutex_exit(&dn->dn_dbufs_mtx);
2075744947dcSTom Erickson 		/*
2076744947dcSTom Erickson 		 * Decrementing the dbuf count means that the hold corresponding
2077744947dcSTom Erickson 		 * to the removed dbuf is no longer discounted in dnode_move(),
2078744947dcSTom Erickson 		 * so the dnode cannot be moved until after we release the hold.
2079744947dcSTom Erickson 		 * The membar_producer() ensures visibility of the decremented
2080744947dcSTom Erickson 		 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
2081744947dcSTom Erickson 		 * release any lock.
2082744947dcSTom Erickson 		 */
2083ea8dc4b6Seschrock 		dnode_rele(dn, db);
2084744947dcSTom Erickson 		db->db_dnode_handle = NULL;
2085dcbf3bd6SGeorge Wilson 
2086dcbf3bd6SGeorge Wilson 		dbuf_hash_remove(db);
2087744947dcSTom Erickson 	} else {
2088744947dcSTom Erickson 		DB_DNODE_EXIT(db);
2089ea8dc4b6Seschrock 	}
2090ea8dc4b6Seschrock 
2091dcbf3bd6SGeorge Wilson 	ASSERT(refcount_is_zero(&db->db_holds));
2092ea8dc4b6Seschrock 
2093dcbf3bd6SGeorge Wilson 	db->db_parent = NULL;
2094dcbf3bd6SGeorge Wilson 
2095dcbf3bd6SGeorge Wilson 	ASSERT(db->db_buf == NULL);
2096dcbf3bd6SGeorge Wilson 	ASSERT(db->db.db_data == NULL);
2097dcbf3bd6SGeorge Wilson 	ASSERT(db->db_hash_next == NULL);
2098dcbf3bd6SGeorge Wilson 	ASSERT(db->db_blkptr == NULL);
2099dcbf3bd6SGeorge Wilson 	ASSERT(db->db_data_pending == NULL);
2100dcbf3bd6SGeorge Wilson 	ASSERT(!multilist_link_active(&db->db_cache_link));
2101dcbf3bd6SGeorge Wilson 
2102dcbf3bd6SGeorge Wilson 	kmem_cache_free(dbuf_kmem_cache, db);
2103dcbf3bd6SGeorge Wilson 	arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2104fa9e4066Sahrens 
2105fa9e4066Sahrens 	/*
2106744947dcSTom Erickson 	 * If this dbuf is referenced from an indirect dbuf,
2107fa9e4066Sahrens 	 * decrement the ref count on the indirect dbuf.
2108fa9e4066Sahrens 	 */
2109c543ec06Sahrens 	if (parent && parent != dndb)
2110ea8dc4b6Seschrock 		dbuf_rele(parent, db);
2111fa9e4066Sahrens }
2112fa9e4066Sahrens 
2113a2cdcdd2SPaul Dagnelie /*
2114a2cdcdd2SPaul Dagnelie  * Note: While bpp will always be updated if the function returns success,
2115a2cdcdd2SPaul Dagnelie  * parentp will not be updated if the dnode does not have dn_dbuf filled in;
2116a2cdcdd2SPaul Dagnelie  * this happens when the dnode is the meta-dnode, or a userused or groupused
2117a2cdcdd2SPaul Dagnelie  * object.
2118a2cdcdd2SPaul Dagnelie  */
2119fa9e4066Sahrens static int
2120fa9e4066Sahrens dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
2121fa9e4066Sahrens     dmu_buf_impl_t **parentp, blkptr_t **bpp)
2122fa9e4066Sahrens {
21230b69c2f0Sahrens 	*parentp = NULL;
21240b69c2f0Sahrens 	*bpp = NULL;
21250b69c2f0Sahrens 
21260a586ceaSMark Shellenbaum 	ASSERT(blkid != DMU_BONUS_BLKID);
21270a586ceaSMark Shellenbaum 
21280a586ceaSMark Shellenbaum 	if (blkid == DMU_SPILL_BLKID) {
21290a586ceaSMark Shellenbaum 		mutex_enter(&dn->dn_mtx);
213006e0070dSMark Shellenbaum 		if (dn->dn_have_spill &&
213106e0070dSMark Shellenbaum 		    (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
21320a586ceaSMark Shellenbaum 			*bpp = &dn->dn_phys->dn_spill;
21330a586ceaSMark Shellenbaum 		else
21340a586ceaSMark Shellenbaum 			*bpp = NULL;
21350a586ceaSMark Shellenbaum 		dbuf_add_ref(dn->dn_dbuf, NULL);
21360a586ceaSMark Shellenbaum 		*parentp = dn->dn_dbuf;
21370a586ceaSMark Shellenbaum 		mutex_exit(&dn->dn_mtx);
21380a586ceaSMark Shellenbaum 		return (0);
21390a586ceaSMark Shellenbaum 	}
2140ea8dc4b6Seschrock 
21417de35a3eSPaul Dagnelie 	int nlevels =
21427de35a3eSPaul Dagnelie 	    (dn->dn_phys->dn_nlevels == 0) ? 1 : dn->dn_phys->dn_nlevels;
21437de35a3eSPaul Dagnelie 	int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2144fa9e4066Sahrens 
2145fa9e4066Sahrens 	ASSERT3U(level * epbs, <, 64);
2146fa9e4066Sahrens 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
21477de35a3eSPaul Dagnelie 	/*
21487de35a3eSPaul Dagnelie 	 * This assertion shouldn't trip as long as the max indirect block size
21497de35a3eSPaul Dagnelie 	 * is less than 1M.  The reason for this is that up to that point,
21507de35a3eSPaul Dagnelie 	 * the number of levels required to address an entire object with blocks
21517de35a3eSPaul Dagnelie 	 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64.  In
21527de35a3eSPaul Dagnelie 	 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55
21537de35a3eSPaul Dagnelie 	 * (i.e. we can address the entire object), objects will all use at most
21547de35a3eSPaul Dagnelie 	 * N-1 levels and the assertion won't overflow.  However, once epbs is
21557de35a3eSPaul Dagnelie 	 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66.  Then, 4 levels will not be
21567de35a3eSPaul Dagnelie 	 * enough to address an entire object, so objects will have 5 levels,
21577de35a3eSPaul Dagnelie 	 * but then this assertion will overflow.
21587de35a3eSPaul Dagnelie 	 *
21597de35a3eSPaul Dagnelie 	 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we
21607de35a3eSPaul Dagnelie 	 * need to redo this logic to handle overflows.
21617de35a3eSPaul Dagnelie 	 */
21627de35a3eSPaul Dagnelie 	ASSERT(level >= nlevels ||
21637de35a3eSPaul Dagnelie 	    ((nlevels - level - 1) * epbs) +
21647de35a3eSPaul Dagnelie 	    highbit64(dn->dn_phys->dn_nblkptr) <= 64);
2165ea8dc4b6Seschrock 	if (level >= nlevels ||
21667de35a3eSPaul Dagnelie 	    blkid >= ((uint64_t)dn->dn_phys->dn_nblkptr <<
21677de35a3eSPaul Dagnelie 	    ((nlevels - level - 1) * epbs)) ||
21687de35a3eSPaul Dagnelie 	    (fail_sparse &&
21697de35a3eSPaul Dagnelie 	    blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
2170fa9e4066Sahrens 		/* the buffer has no parent yet */
2171be6fd75aSMatthew Ahrens 		return (SET_ERROR(ENOENT));
2172fa9e4066Sahrens 	} else if (level < nlevels-1) {
2173fa9e4066Sahrens 		/* this block is referenced from an indirect block */
2174fa9e4066Sahrens 		int err = dbuf_hold_impl(dn, level+1,
2175a2cdcdd2SPaul Dagnelie 		    blkid >> epbs, fail_sparse, FALSE, NULL, parentp);
2176fa9e4066Sahrens 		if (err)
2177fa9e4066Sahrens 			return (err);
2178ea8dc4b6Seschrock 		err = dbuf_read(*parentp, NULL,
2179ea8dc4b6Seschrock 		    (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
2180c543ec06Sahrens 		if (err) {
2181c543ec06Sahrens 			dbuf_rele(*parentp, NULL);
2182c543ec06Sahrens 			*parentp = NULL;
2183c543ec06Sahrens 			return (err);
2184ea8dc4b6Seschrock 		}
2185c543ec06Sahrens 		*bpp = ((blkptr_t *)(*parentp)->db.db_data) +
2186c543ec06Sahrens 		    (blkid & ((1ULL << epbs) - 1));
21877de35a3eSPaul Dagnelie 		if (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))
21887de35a3eSPaul Dagnelie 			ASSERT(BP_IS_HOLE(*bpp));
2189c543ec06Sahrens 		return (0);
2190fa9e4066Sahrens 	} else {
2191fa9e4066Sahrens 		/* the block is referenced from the dnode */
2192fa9e4066Sahrens 		ASSERT3U(level, ==, nlevels-1);
2193fa9e4066Sahrens 		ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
2194fa9e4066Sahrens 		    blkid < dn->dn_phys->dn_nblkptr);
2195c543ec06Sahrens 		if (dn->dn_dbuf) {
2196c543ec06Sahrens 			dbuf_add_ref(dn->dn_dbuf, NULL);
2197c543ec06Sahrens 			*parentp = dn->dn_dbuf;
2198c543ec06Sahrens 		}
2199fa9e4066Sahrens 		*bpp = &dn->dn_phys->dn_blkptr[blkid];
2200fa9e4066Sahrens 		return (0);
2201fa9e4066Sahrens 	}
2202fa9e4066Sahrens }
2203fa9e4066Sahrens 
2204fa9e4066Sahrens static dmu_buf_impl_t *
2205fa9e4066Sahrens dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
2206fa9e4066Sahrens     dmu_buf_impl_t *parent, blkptr_t *blkptr)
2207fa9e4066Sahrens {
2208503ad85cSMatthew Ahrens 	objset_t *os = dn->dn_objset;
2209fa9e4066Sahrens 	dmu_buf_impl_t *db, *odb;
2210fa9e4066Sahrens 
2211fa9e4066Sahrens 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2212fa9e4066Sahrens 	ASSERT(dn->dn_type != DMU_OT_NONE);
2213fa9e4066Sahrens 
2214dcbf3bd6SGeorge Wilson 	db = kmem_cache_alloc(dbuf_kmem_cache, KM_SLEEP);
2215fa9e4066Sahrens 
2216fa9e4066Sahrens 	db->db_objset = os;
2217fa9e4066Sahrens 	db->db.db_object = dn->dn_object;
2218fa9e4066Sahrens 	db->db_level = level;
2219fa9e4066Sahrens 	db->db_blkid = blkid;
2220c717a561Smaybee 	db->db_last_dirty = NULL;
2221ea8dc4b6Seschrock 	db->db_dirtycnt = 0;
2222744947dcSTom Erickson 	db->db_dnode_handle = dn->dn_handle;
2223ea8dc4b6Seschrock 	db->db_parent = parent;
2224ea8dc4b6Seschrock 	db->db_blkptr = blkptr;
2225fa9e4066Sahrens 
2226bc9014e6SJustin Gibbs 	db->db_user = NULL;
2227d2058105SJustin T. Gibbs 	db->db_user_immediate_evict = FALSE;
2228d2058105SJustin T. Gibbs 	db->db_freed_in_flight = FALSE;
2229d2058105SJustin T. Gibbs 	db->db_pending_evict = FALSE;
2230ea8dc4b6Seschrock 
22310a586ceaSMark Shellenbaum 	if (blkid == DMU_BONUS_BLKID) {
2232ea8dc4b6Seschrock 		ASSERT3P(parent, ==, dn->dn_dbuf);
22331934e92fSmaybee 		db->db.db_size = DN_MAX_BONUSLEN -
22341934e92fSmaybee 		    (dn->dn_nblkptr-1) * sizeof (blkptr_t);
22351934e92fSmaybee 		ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
22360a586ceaSMark Shellenbaum 		db->db.db_offset = DMU_BONUS_BLKID;
2237ea8dc4b6Seschrock 		db->db_state = DB_UNCACHED;
2238ea8dc4b6Seschrock 		/* the bonus dbuf is not placed in the hash table */
22395a98e54bSBrendan Gregg - Sun Microsystems 		arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2240ea8dc4b6Seschrock 		return (db);
22410a586ceaSMark Shellenbaum 	} else if (blkid == DMU_SPILL_BLKID) {
22420a586ceaSMark Shellenbaum 		db->db.db_size = (blkptr != NULL) ?
22430a586ceaSMark Shellenbaum 		    BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
22440a586ceaSMark Shellenbaum 		db->db.db_offset = 0;
2245fa9e4066Sahrens 	} else {
2246fa9e4066Sahrens 		int blocksize =
224769962b56SMatthew Ahrens 		    db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
2248fa9e4066Sahrens 		db->db.db_size = blocksize;
2249fa9e4066Sahrens 		db->db.db_offset = db->db_blkid * blocksize;
2250fa9e4066Sahrens 	}
2251fa9e4066Sahrens 
2252fa9e4066Sahrens 	/*
2253fa9e4066Sahrens 	 * Hold the dn_dbufs_mtx while we get the new dbuf
2254fa9e4066Sahrens 	 * in the hash table *and* added to the dbufs list.
2255fa9e4066Sahrens 	 * This prevents a possible deadlock with someone
2256fa9e4066Sahrens 	 * trying to look up this dbuf before its added to the
2257fa9e4066Sahrens 	 * dn_dbufs list.
2258fa9e4066Sahrens 	 */
2259fa9e4066Sahrens 	mutex_enter(&dn->dn_dbufs_mtx);
2260ea8dc4b6Seschrock 	db->db_state = DB_EVICTING;
2261fa9e4066Sahrens 	if ((odb = dbuf_hash_insert(db)) != NULL) {
2262fa9e4066Sahrens 		/* someone else inserted it first */
2263dcbf3bd6SGeorge Wilson 		kmem_cache_free(dbuf_kmem_cache, db);
2264fa9e4066Sahrens 		mutex_exit(&dn->dn_dbufs_mtx);
2265fa9e4066Sahrens 		return (odb);
2266fa9e4066Sahrens 	}
22670f6d88adSAlex Reece 	avl_add(&dn->dn_dbufs, db);
2268653af1b8SStephen Blinick 
2269ea8dc4b6Seschrock 	db->db_state = DB_UNCACHED;
2270fa9e4066Sahrens 	mutex_exit(&dn->dn_dbufs_mtx);
22715a98e54bSBrendan Gregg - Sun Microsystems 	arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2272fa9e4066Sahrens 
2273fa9e4066Sahrens 	if (parent && parent != dn->dn_dbuf)
2274fa9e4066Sahrens 		dbuf_add_ref(parent, db);
2275fa9e4066Sahrens 
2276ea8dc4b6Seschrock 	ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
2277ea8dc4b6Seschrock 	    refcount_count(&dn->dn_holds) > 0);
2278fa9e4066Sahrens 	(void) refcount_add(&dn->dn_holds, db);
2279640c1670SJosef 'Jeff' Sipek 	atomic_inc_32(&dn->dn_dbufs_count);
2280fa9e4066Sahrens 
2281fa9e4066Sahrens 	dprintf_dbuf(db, "db=%p\n", db);
2282fa9e4066Sahrens 
2283fa9e4066Sahrens 	return (db);
2284fa9e4066Sahrens }
2285fa9e4066Sahrens 
2286a2cdcdd2SPaul Dagnelie typedef struct dbuf_prefetch_arg {
2287a2cdcdd2SPaul Dagnelie 	spa_t *dpa_spa;	/* The spa to issue the prefetch in. */
2288a2cdcdd2SPaul Dagnelie 	zbookmark_phys_t dpa_zb; /* The target block to prefetch. */
2289a2cdcdd2SPaul Dagnelie 	int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */
2290a2cdcdd2SPaul Dagnelie 	int dpa_curlevel; /* The current level that we're reading */
2291dcbf3bd6SGeorge Wilson 	dnode_t *dpa_dnode; /* The dnode associated with the prefetch */
2292a2cdcdd2SPaul Dagnelie 	zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */
2293a2cdcdd2SPaul Dagnelie 	zio_t *dpa_zio; /* The parent zio_t for all prefetches. */
2294a2cdcdd2SPaul Dagnelie 	arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */
2295a2cdcdd2SPaul Dagnelie } dbuf_prefetch_arg_t;
2296a2cdcdd2SPaul Dagnelie 
2297a2cdcdd2SPaul Dagnelie /*
2298a2cdcdd2SPaul Dagnelie  * Actually issue the prefetch read for the block given.
2299a2cdcdd2SPaul Dagnelie  */
2300a2cdcdd2SPaul Dagnelie static void
2301a2cdcdd2SPaul Dagnelie dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp)
2302a2cdcdd2SPaul Dagnelie {
2303a2cdcdd2SPaul Dagnelie 	if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
2304a2cdcdd2SPaul Dagnelie 		return;
2305a2cdcdd2SPaul Dagnelie 
2306a2cdcdd2SPaul Dagnelie 	arc_flags_t aflags =
2307a2cdcdd2SPaul Dagnelie 	    dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
2308a2cdcdd2SPaul Dagnelie 
2309a2cdcdd2SPaul Dagnelie 	ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2310a2cdcdd2SPaul Dagnelie 	ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level);
2311a2cdcdd2SPaul Dagnelie 	ASSERT(dpa->dpa_zio != NULL);
2312a2cdcdd2SPaul Dagnelie 	(void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL,
2313a2cdcdd2SPaul Dagnelie 	    dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2314a2cdcdd2SPaul Dagnelie 	    &aflags, &dpa->dpa_zb);
2315a2cdcdd2SPaul Dagnelie }
2316a2cdcdd2SPaul Dagnelie 
2317a2cdcdd2SPaul Dagnelie /*
2318a2cdcdd2SPaul Dagnelie  * Called when an indirect block above our prefetch target is read in.  This
2319a2cdcdd2SPaul Dagnelie  * will either read in the next indirect block down the tree or issue the actual
2320a2cdcdd2SPaul Dagnelie  * prefetch if the next block down is our target.
2321a2cdcdd2SPaul Dagnelie  */
2322a2cdcdd2SPaul Dagnelie static void
2323a2cdcdd2SPaul Dagnelie dbuf_prefetch_indirect_done(zio_t *zio, arc_buf_t *abuf, void *private)
2324a2cdcdd2SPaul Dagnelie {
2325a2cdcdd2SPaul Dagnelie 	dbuf_prefetch_arg_t *dpa = private;
2326a2cdcdd2SPaul Dagnelie 
2327a2cdcdd2SPaul Dagnelie 	ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel);
2328a2cdcdd2SPaul Dagnelie 	ASSERT3S(dpa->dpa_curlevel, >, 0);
2329dcbf3bd6SGeorge Wilson 
2330dcbf3bd6SGeorge Wilson 	/*
2331dcbf3bd6SGeorge Wilson 	 * The dpa_dnode is only valid if we are called with a NULL
2332dcbf3bd6SGeorge Wilson 	 * zio. This indicates that the arc_read() returned without
2333dcbf3bd6SGeorge Wilson 	 * first calling zio_read() to issue a physical read. Once
2334dcbf3bd6SGeorge Wilson 	 * a physical read is made the dpa_dnode must be invalidated
2335dcbf3bd6SGeorge Wilson 	 * as the locks guarding it may have been dropped. If the
2336dcbf3bd6SGeorge Wilson 	 * dpa_dnode is still valid, then we want to add it to the dbuf
2337dcbf3bd6SGeorge Wilson 	 * cache. To do so, we must hold the dbuf associated with the block
2338dcbf3bd6SGeorge Wilson 	 * we just prefetched, read its contents so that we associate it
2339dcbf3bd6SGeorge Wilson 	 * with an arc_buf_t, and then release it.
2340dcbf3bd6SGeorge Wilson 	 */
2341a2cdcdd2SPaul Dagnelie 	if (zio != NULL) {
2342a2cdcdd2SPaul Dagnelie 		ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel);
2343dcbf3bd6SGeorge Wilson 		if (zio->io_flags & ZIO_FLAG_RAW) {
2344dcbf3bd6SGeorge Wilson 			ASSERT3U(BP_GET_PSIZE(zio->io_bp), ==, zio->io_size);
2345dcbf3bd6SGeorge Wilson 		} else {
2346dcbf3bd6SGeorge Wilson 			ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size);
2347dcbf3bd6SGeorge Wilson 		}
2348a2cdcdd2SPaul Dagnelie 		ASSERT3P(zio->io_spa, ==, dpa->dpa_spa);
2349dcbf3bd6SGeorge Wilson 
2350dcbf3bd6SGeorge Wilson 		dpa->dpa_dnode = NULL;
2351dcbf3bd6SGeorge Wilson 	} else if (dpa->dpa_dnode != NULL) {
2352dcbf3bd6SGeorge Wilson 		uint64_t curblkid = dpa->dpa_zb.zb_blkid >>
2353dcbf3bd6SGeorge Wilson 		    (dpa->dpa_epbs * (dpa->dpa_curlevel -
2354dcbf3bd6SGeorge Wilson 		    dpa->dpa_zb.zb_level));
2355dcbf3bd6SGeorge Wilson 		dmu_buf_impl_t *db = dbuf_hold_level(dpa->dpa_dnode,
2356dcbf3bd6SGeorge Wilson 		    dpa->dpa_curlevel, curblkid, FTAG);
2357dcbf3bd6SGeorge Wilson 		(void) dbuf_read(db, NULL,
2358dcbf3bd6SGeorge Wilson 		    DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_HAVESTRUCT);
2359dcbf3bd6SGeorge Wilson 		dbuf_rele(db, FTAG);
2360a2cdcdd2SPaul Dagnelie 	}
2361a2cdcdd2SPaul Dagnelie 
2362a2cdcdd2SPaul Dagnelie 	dpa->dpa_curlevel--;
2363a2cdcdd2SPaul Dagnelie 
2364a2cdcdd2SPaul Dagnelie 	uint64_t nextblkid = dpa->dpa_zb.zb_blkid >>
2365a2cdcdd2SPaul Dagnelie 	    (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level));
2366a2cdcdd2SPaul Dagnelie 	blkptr_t *bp = ((blkptr_t *)abuf->b_data) +
2367a2cdcdd2SPaul Dagnelie 	    P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs);
2368a2cdcdd2SPaul Dagnelie 	if (BP_IS_HOLE(bp) || (zio != NULL && zio->io_error != 0)) {
2369a2cdcdd2SPaul Dagnelie 		kmem_free(dpa, sizeof (*dpa));
2370a2cdcdd2SPaul Dagnelie 	} else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) {
2371a2cdcdd2SPaul Dagnelie 		ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid);
2372a2cdcdd2SPaul Dagnelie 		dbuf_issue_final_prefetch(dpa, bp);
2373a2cdcdd2SPaul Dagnelie 		kmem_free(dpa, sizeof (*dpa));
2374a2cdcdd2SPaul Dagnelie 	} else {
2375a2cdcdd2SPaul Dagnelie 		arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2376a2cdcdd2SPaul Dagnelie 		zbookmark_phys_t zb;
2377a2cdcdd2SPaul Dagnelie 
237827295216Sbenrubson 		/* flag if L2ARC eligible, l2arc_noprefetch then decides */
237927295216Sbenrubson 		if (dpa->dpa_aflags & ARC_FLAG_L2CACHE)
238027295216Sbenrubson 			iter_aflags |= ARC_FLAG_L2CACHE;
238127295216Sbenrubson 
2382a2cdcdd2SPaul Dagnelie 		ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2383a2cdcdd2SPaul Dagnelie 
2384a2cdcdd2SPaul Dagnelie 		SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset,
2385a2cdcdd2SPaul Dagnelie 		    dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid);
2386a2cdcdd2SPaul Dagnelie 
2387a2cdcdd2SPaul Dagnelie 		(void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2388a2cdcdd2SPaul Dagnelie 		    bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio,
2389a2cdcdd2SPaul Dagnelie 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2390a2cdcdd2SPaul Dagnelie 		    &iter_aflags, &zb);
2391a2cdcdd2SPaul Dagnelie 	}
2392dcbf3bd6SGeorge Wilson 
2393dcbf3bd6SGeorge Wilson 	arc_buf_destroy(abuf, private);
2394a2cdcdd2SPaul Dagnelie }
2395a2cdcdd2SPaul Dagnelie 
2396a2cdcdd2SPaul Dagnelie /*
2397a2cdcdd2SPaul Dagnelie  * Issue prefetch reads for the given block on the given level.  If the indirect
2398a2cdcdd2SPaul Dagnelie  * blocks above that block are not in memory, we will read them in
2399a2cdcdd2SPaul Dagnelie  * asynchronously.  As a result, this call never blocks waiting for a read to
2400a2cdcdd2SPaul Dagnelie  * complete.
2401a2cdcdd2SPaul Dagnelie  */
2402fa9e4066Sahrens void
2403a2cdcdd2SPaul Dagnelie dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio,
2404a2cdcdd2SPaul Dagnelie     arc_flags_t aflags)
2405fa9e4066Sahrens {
2406a2cdcdd2SPaul Dagnelie 	blkptr_t bp;
2407a2cdcdd2SPaul Dagnelie 	int epbs, nlevels, curlevel;
2408a2cdcdd2SPaul Dagnelie 	uint64_t curblkid;
2409fa9e4066Sahrens 
24100a586ceaSMark Shellenbaum 	ASSERT(blkid != DMU_BONUS_BLKID);
2411fa9e4066Sahrens 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2412fa9e4066Sahrens 
2413cf6106c8SMatthew Ahrens 	if (blkid > dn->dn_maxblkid)
2414cf6106c8SMatthew Ahrens 		return;
2415cf6106c8SMatthew Ahrens 
2416fa9e4066Sahrens 	if (dnode_block_freed(dn, blkid))
2417fa9e4066Sahrens 		return;
2418fa9e4066Sahrens 
2419a2cdcdd2SPaul Dagnelie 	/*
2420a2cdcdd2SPaul Dagnelie 	 * This dnode hasn't been written to disk yet, so there's nothing to
2421a2cdcdd2SPaul Dagnelie 	 * prefetch.
2422a2cdcdd2SPaul Dagnelie 	 */
2423a2cdcdd2SPaul Dagnelie 	nlevels = dn->dn_phys->dn_nlevels;
2424a2cdcdd2SPaul Dagnelie 	if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0)
2425a2cdcdd2SPaul Dagnelie 		return;
2426a2cdcdd2SPaul Dagnelie 
2427a2cdcdd2SPaul Dagnelie 	epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2428a2cdcdd2SPaul Dagnelie 	if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level))
2429a2cdcdd2SPaul Dagnelie 		return;
2430a2cdcdd2SPaul Dagnelie 
2431a2cdcdd2SPaul Dagnelie 	dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object,
2432a2cdcdd2SPaul Dagnelie 	    level, blkid);
2433a2cdcdd2SPaul Dagnelie 	if (db != NULL) {
2434a2cdcdd2SPaul Dagnelie 		mutex_exit(&db->db_mtx);
24359e9c486fSGeorge Wilson 		/*
2436a2cdcdd2SPaul Dagnelie 		 * This dbuf already exists.  It is either CACHED, or
2437a2cdcdd2SPaul Dagnelie 		 * (we assume) about to be read or filled.
24389e9c486fSGeorge Wilson 		 */
24399e9c486fSGeorge Wilson 		return;
2440fa9e4066Sahrens 	}
2441fa9e4066Sahrens 
2442a2cdcdd2SPaul Dagnelie 	/*
2443a2cdcdd2SPaul Dagnelie 	 * Find the closest ancestor (indirect block) of the target block
2444a2cdcdd2SPaul Dagnelie 	 * that is present in the cache.  In this indirect block, we will
2445a2cdcdd2SPaul Dagnelie 	 * find the bp that is at curlevel, curblkid.
2446a2cdcdd2SPaul Dagnelie 	 */
2447a2cdcdd2SPaul Dagnelie 	curlevel = level;
2448a2cdcdd2SPaul Dagnelie 	curblkid = blkid;
2449a2cdcdd2SPaul Dagnelie 	while (curlevel < nlevels - 1) {
2450a2cdcdd2SPaul Dagnelie 		int parent_level = curlevel + 1;
2451a2cdcdd2SPaul Dagnelie 		uint64_t parent_blkid = curblkid >> epbs;
2452a2cdcdd2SPaul Dagnelie 		dmu_buf_impl_t *db;
2453a2cdcdd2SPaul Dagnelie 
2454a2cdcdd2SPaul Dagnelie 		if (dbuf_hold_impl(dn, parent_level, parent_blkid,
2455a2cdcdd2SPaul Dagnelie 		    FALSE, TRUE, FTAG, &db) == 0) {
2456a2cdcdd2SPaul Dagnelie 			blkptr_t *bpp = db->db_buf->b_data;
2457a2cdcdd2SPaul Dagnelie 			bp = bpp[P2PHASE(curblkid, 1 << epbs)];
2458a2cdcdd2SPaul Dagnelie 			dbuf_rele(db, FTAG);
2459a2cdcdd2SPaul Dagnelie 			break;
2460a2cdcdd2SPaul Dagnelie 		}
2461b24ab676SJeff Bonwick 
2462a2cdcdd2SPaul Dagnelie 		curlevel = parent_level;
2463a2cdcdd2SPaul Dagnelie 		curblkid = parent_blkid;
2464a2cdcdd2SPaul Dagnelie 	}
2465ea8dc4b6Seschrock 
2466a2cdcdd2SPaul Dagnelie 	if (curlevel == nlevels - 1) {
2467a2cdcdd2SPaul Dagnelie 		/* No cached indirect blocks found. */
2468a2cdcdd2SPaul Dagnelie 		ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr);
2469a2cdcdd2SPaul Dagnelie 		bp = dn->dn_phys->dn_blkptr[curblkid];
2470fa9e4066Sahrens 	}
2471a2cdcdd2SPaul Dagnelie 	if (BP_IS_HOLE(&bp))
2472a2cdcdd2SPaul Dagnelie 		return;
2473a2cdcdd2SPaul Dagnelie 
2474a2cdcdd2SPaul Dagnelie 	ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp));
2475a2cdcdd2SPaul Dagnelie 
2476a2cdcdd2SPaul Dagnelie 	zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL,
2477a2cdcdd2SPaul Dagnelie 	    ZIO_FLAG_CANFAIL);
2478a2cdcdd2SPaul Dagnelie 
2479a2cdcdd2SPaul Dagnelie 	dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP);
2480a2cdcdd2SPaul Dagnelie 	dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
2481a2cdcdd2SPaul Dagnelie 	SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2482a2cdcdd2SPaul Dagnelie 	    dn->dn_object, level, blkid);
2483a2cdcdd2SPaul Dagnelie 	dpa->dpa_curlevel = curlevel;
2484a2cdcdd2SPaul Dagnelie 	dpa->dpa_prio = prio;
2485a2cdcdd2SPaul Dagnelie 	dpa->dpa_aflags = aflags;
2486a2cdcdd2SPaul Dagnelie 	dpa->dpa_spa = dn->dn_objset->os_spa;
2487dcbf3bd6SGeorge Wilson 	dpa->dpa_dnode = dn;
2488a2cdcdd2SPaul Dagnelie 	dpa->dpa_epbs = epbs;
2489a2cdcdd2SPaul Dagnelie 	dpa->dpa_zio = pio;
2490a2cdcdd2SPaul Dagnelie 
249127295216Sbenrubson 	/* flag if L2ARC eligible, l2arc_noprefetch then decides */
249227295216Sbenrubson 	if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level))
249327295216Sbenrubson 		dpa->dpa_aflags |= ARC_FLAG_L2CACHE;
249427295216Sbenrubson 
2495a2cdcdd2SPaul Dagnelie 	/*
2496a2cdcdd2SPaul Dagnelie 	 * If we have the indirect just above us, no need to do the asynchronous
2497a2cdcdd2SPaul Dagnelie 	 * prefetch chain; we'll just run the last step ourselves.  If we're at
2498a2cdcdd2SPaul Dagnelie 	 * a higher level, though, we want to issue the prefetches for all the
2499a2cdcdd2SPaul Dagnelie 	 * indirect blocks asynchronously, so we can go on with whatever we were
2500a2cdcdd2SPaul Dagnelie 	 * doing.
2501a2cdcdd2SPaul Dagnelie 	 */
2502a2cdcdd2SPaul Dagnelie 	if (curlevel == level) {
2503a2cdcdd2SPaul Dagnelie 		ASSERT3U(curblkid, ==, blkid);
2504a2cdcdd2SPaul Dagnelie 		dbuf_issue_final_prefetch(dpa, &bp);
2505a2cdcdd2SPaul Dagnelie 		kmem_free(dpa, sizeof (*dpa));
2506a2cdcdd2SPaul Dagnelie 	} else {
2507a2cdcdd2SPaul Dagnelie 		arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2508a2cdcdd2SPaul Dagnelie 		zbookmark_phys_t zb;
2509a2cdcdd2SPaul Dagnelie 
251027295216Sbenrubson 		/* flag if L2ARC eligible, l2arc_noprefetch then decides */
251127295216Sbenrubson 		if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level))
251227295216Sbenrubson 			iter_aflags |= ARC_FLAG_L2CACHE;
251327295216Sbenrubson 
2514a2cdcdd2SPaul Dagnelie 		SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2515a2cdcdd2SPaul Dagnelie 		    dn->dn_object, curlevel, curblkid);
2516a2cdcdd2SPaul Dagnelie 		(void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2517a2cdcdd2SPaul Dagnelie 		    &bp, dbuf_prefetch_indirect_done, dpa, prio,
2518a2cdcdd2SPaul Dagnelie 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2519a2cdcdd2SPaul Dagnelie 		    &iter_aflags, &zb);
2520a2cdcdd2SPaul Dagnelie 	}
2521a2cdcdd2SPaul Dagnelie 	/*
2522a2cdcdd2SPaul Dagnelie 	 * We use pio here instead of dpa_zio since it's possible that
2523a2cdcdd2SPaul Dagnelie 	 * dpa may have already been freed.
2524a2cdcdd2SPaul Dagnelie 	 */
2525a2cdcdd2SPaul Dagnelie 	zio_nowait(pio);
2526fa9e4066Sahrens }
2527fa9e4066Sahrens 
2528fa9e4066Sahrens /*
2529fa9e4066Sahrens  * Returns with db_holds incremented, and db_mtx not held.
2530fa9e4066Sahrens  * Note: dn_struct_rwlock must be held.
2531fa9e4066Sahrens  */
2532fa9e4066Sahrens int
2533a2cdcdd2SPaul Dagnelie dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid,
2534a2cdcdd2SPaul Dagnelie     boolean_t fail_sparse, boolean_t fail_uncached,
2535fa9e4066Sahrens     void *tag, dmu_buf_impl_t **dbp)
2536fa9e4066Sahrens {
2537fa9e4066Sahrens 	dmu_buf_impl_t *db, *parent = NULL;
2538fa9e4066Sahrens 
25390a586ceaSMark Shellenbaum 	ASSERT(blkid != DMU_BONUS_BLKID);
2540fa9e4066Sahrens 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2541fa9e4066Sahrens 	ASSERT3U(dn->dn_nlevels, >, level);
2542fa9e4066Sahrens 
2543fa9e4066Sahrens 	*dbp = NULL;
2544ea8dc4b6Seschrock top:
2545fa9e4066Sahrens 	/* dbuf_find() returns with db_mtx held */
2546e57a022bSJustin T. Gibbs 	db = dbuf_find(dn->dn_objset, dn->dn_object, level, blkid);
2547fa9e4066Sahrens 
2548fa9e4066Sahrens 	if (db == NULL) {
2549fa9e4066Sahrens 		blkptr_t *bp = NULL;
2550fa9e4066Sahrens 		int err;
2551fa9e4066Sahrens 
2552a2cdcdd2SPaul Dagnelie 		if (fail_uncached)
2553a2cdcdd2SPaul Dagnelie 			return (SET_ERROR(ENOENT));
2554a2cdcdd2SPaul Dagnelie 
2555c543ec06Sahrens 		ASSERT3P(parent, ==, NULL);
2556fa9e4066Sahrens 		err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
2557fa9e4066Sahrens 		if (fail_sparse) {
2558fa9e4066Sahrens 			if (err == 0 && bp && BP_IS_HOLE(bp))
2559be6fd75aSMatthew Ahrens 				err = SET_ERROR(ENOENT);
2560fa9e4066Sahrens 			if (err) {
2561c543ec06Sahrens 				if (parent)
2562ea8dc4b6Seschrock 					dbuf_rele(parent, NULL);
2563fa9e4066Sahrens 				return (err);
2564fa9e4066Sahrens 			}
2565fa9e4066Sahrens 		}
2566ea8dc4b6Seschrock 		if (err && err != ENOENT)
2567ea8dc4b6Seschrock 			return (err);
2568fa9e4066Sahrens 		db = dbuf_create(dn, level, blkid, parent, bp);
2569fa9e4066Sahrens 	}
2570fa9e4066Sahrens 
2571a2cdcdd2SPaul Dagnelie 	if (fail_uncached && db->db_state != DB_CACHED) {
2572a2cdcdd2SPaul Dagnelie 		mutex_exit(&db->db_mtx);
2573a2cdcdd2SPaul Dagnelie 		return (SET_ERROR(ENOENT));
2574a2cdcdd2SPaul Dagnelie 	}
2575a2cdcdd2SPaul Dagnelie 
2576dcbf3bd6SGeorge Wilson 	if (db->db_buf != NULL)
2577ea8dc4b6Seschrock 		ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
2578ea8dc4b6Seschrock 
2579ea8dc4b6Seschrock 	ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
2580ea8dc4b6Seschrock 
2581fa9e4066Sahrens 	/*
2582c717a561Smaybee 	 * If this buffer is currently syncing out, and we are are
2583c717a561Smaybee 	 * still referencing it from db_data, we need to make a copy
2584c717a561Smaybee 	 * of it in case we decide we want to dirty it again in this txg.
2585fa9e4066Sahrens 	 */
25860a586ceaSMark Shellenbaum 	if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
2587ea8dc4b6Seschrock 	    dn->dn_object != DMU_META_DNODE_OBJECT &&
2588c717a561Smaybee 	    db->db_state == DB_CACHED && db->db_data_pending) {
2589c717a561Smaybee 		dbuf_dirty_record_t *dr = db->db_data_pending;
2590fa9e4066Sahrens 
2591c717a561Smaybee 		if (dr->dt.dl.dr_data == db->db_buf) {
2592c717a561Smaybee 			arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2593c717a561Smaybee 
2594c717a561Smaybee 			dbuf_set_data(db,
25955602294fSDan Kimmel 			    arc_alloc_buf(dn->dn_objset->os_spa, db, type,
25965602294fSDan Kimmel 			    db->db.db_size));
2597c717a561Smaybee 			bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
2598c717a561Smaybee 			    db->db.db_size);
2599c717a561Smaybee 		}
2600fa9e4066Sahrens 	}
2601fa9e4066Sahrens 
2602dcbf3bd6SGeorge Wilson 	if (multilist_link_active(&db->db_cache_link)) {
2603dcbf3bd6SGeorge Wilson 		ASSERT(refcount_is_zero(&db->db_holds));
260494c2d0ebSMatthew Ahrens 		multilist_remove(dbuf_cache, db);
2605dcbf3bd6SGeorge Wilson 		(void) refcount_remove_many(&dbuf_cache_size,
2606dcbf3bd6SGeorge Wilson 		    db->db.db_size, db);
2607dcbf3bd6SGeorge Wilson 	}
2608ea8dc4b6Seschrock 	(void) refcount_add(&db->db_holds, tag);
26099c9dc39aSek 	DBUF_VERIFY(db);
2610fa9e4066Sahrens 	mutex_exit(&db->db_mtx);
2611fa9e4066Sahrens 
2612fa9e4066Sahrens 	/* NOTE: we can't rele the parent until after we drop the db_mtx */
2613c543ec06Sahrens 	if (parent)
2614ea8dc4b6Seschrock 		dbuf_rele(parent, NULL);
2615fa9e4066Sahrens 
2616744947dcSTom Erickson 	ASSERT3P(DB_DNODE(db), ==, dn);
2617fa9e4066Sahrens 	ASSERT3U(db->db_blkid, ==, blkid);
2618fa9e4066Sahrens 	ASSERT3U(db->db_level, ==, level);
2619fa9e4066Sahrens 	*dbp = db;
2620fa9e4066Sahrens 
2621fa9e4066Sahrens 	return (0);
2622fa9e4066Sahrens }
2623fa9e4066Sahrens 
2624fa9e4066Sahrens dmu_buf_impl_t *
2625ea8dc4b6Seschrock dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
2626fa9e4066Sahrens {
2627a2cdcdd2SPaul Dagnelie 	return (dbuf_hold_level(dn, 0, blkid, tag));
2628fa9e4066Sahrens }
2629fa9e4066Sahrens 
2630fa9e4066Sahrens dmu_buf_impl_t *
2631fa9e4066Sahrens dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
2632fa9e4066Sahrens {
2633fa9e4066Sahrens 	dmu_buf_impl_t *db;
2634a2cdcdd2SPaul Dagnelie 	int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db);
2635ea8dc4b6Seschrock 	return (err ? NULL : db);
2636fa9e4066Sahrens }
2637fa9e4066Sahrens 
26381934e92fSmaybee void
2639ea8dc4b6Seschrock dbuf_create_bonus(dnode_t *dn)
2640fa9e4066Sahrens {
2641ea8dc4b6Seschrock 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
2642ea8dc4b6Seschrock 
2643ea8dc4b6Seschrock 	ASSERT(dn->dn_bonus == NULL);
26440a586ceaSMark Shellenbaum 	dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
26450a586ceaSMark Shellenbaum }
26460a586ceaSMark Shellenbaum 
26470a586ceaSMark Shellenbaum int
26480a586ceaSMark Shellenbaum dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
26490a586ceaSMark Shellenbaum {
26500a586ceaSMark Shellenbaum 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2651744947dcSTom Erickson 	dnode_t *dn;
2652744947dcSTom Erickson 
26530a586ceaSMark Shellenbaum 	if (db->db_blkid != DMU_SPILL_BLKID)
2654be6fd75aSMatthew Ahrens 		return (SET_ERROR(ENOTSUP));
26550a586ceaSMark Shellenbaum 	if (blksz == 0)
26560a586ceaSMark Shellenbaum 		blksz = SPA_MINBLOCKSIZE;
2657b5152584SMatthew Ahrens 	ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset)));
2658b5152584SMatthew Ahrens 	blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
26590a586ceaSMark Shellenbaum 
2660744947dcSTom Erickson 	DB_DNODE_ENTER(db);
2661744947dcSTom Erickson 	dn = DB_DNODE(db);
2662744947dcSTom Erickson 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
26630a586ceaSMark Shellenbaum 	dbuf_new_size(db, blksz, tx);
2664744947dcSTom Erickson 	rw_exit(&dn->dn_struct_rwlock);
2665744947dcSTom Erickson 	DB_DNODE_EXIT(db);
26660a586ceaSMark Shellenbaum 
26670a586ceaSMark Shellenbaum 	return (0);
26680a586ceaSMark Shellenbaum }
26690a586ceaSMark Shellenbaum 
26700a586ceaSMark Shellenbaum void
26710a586ceaSMark Shellenbaum dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
26720a586ceaSMark Shellenbaum {
26730a586ceaSMark Shellenbaum 	dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
2674fa9e4066Sahrens }
2675fa9e4066Sahrens 
2676ea8dc4b6Seschrock #pragma weak dmu_buf_add_ref = dbuf_add_ref
2677fa9e4066Sahrens void
2678fa9e4066Sahrens dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
2679fa9e4066Sahrens {
2680ea8dc4b6Seschrock 	int64_t holds = refcount_add(&db->db_holds, tag);
2681dcbf3bd6SGeorge Wilson 	ASSERT3S(holds, >, 1);
2682fa9e4066Sahrens }
2683fa9e4066Sahrens 
2684e57a022bSJustin T. Gibbs #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
2685e57a022bSJustin T. Gibbs boolean_t
2686e57a022bSJustin T. Gibbs dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid,
2687e57a022bSJustin T. Gibbs     void *tag)
2688e57a022bSJustin T. Gibbs {
2689e57a022bSJustin T. Gibbs 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2690e57a022bSJustin T. Gibbs 	dmu_buf_impl_t *found_db;
2691e57a022bSJustin T. Gibbs 	boolean_t result = B_FALSE;
2692e57a022bSJustin T. Gibbs 
2693e57a022bSJustin T. Gibbs 	if (db->db_blkid == DMU_BONUS_BLKID)
2694e57a022bSJustin T. Gibbs 		found_db = dbuf_find_bonus(os, obj);
2695e57a022bSJustin T. Gibbs 	else
2696e57a022bSJustin T. Gibbs 		found_db = dbuf_find(os, obj, 0, blkid);
2697e57a022bSJustin T. Gibbs 
2698e57a022bSJustin T. Gibbs 	if (found_db != NULL) {
2699e57a022bSJustin T. Gibbs 		if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) {
2700e57a022bSJustin T. Gibbs 			(void) refcount_add(&db->db_holds, tag);
2701e57a022bSJustin T. Gibbs 			result = B_TRUE;
2702e57a022bSJustin T. Gibbs 		}
2703e57a022bSJustin T. Gibbs 		mutex_exit(&db->db_mtx);
2704e57a022bSJustin T. Gibbs 	}
2705e57a022bSJustin T. Gibbs 	return (result);
2706e57a022bSJustin T. Gibbs }
2707e57a022bSJustin T. Gibbs 
2708744947dcSTom Erickson /*
2709744947dcSTom Erickson  * If you call dbuf_rele() you had better not be referencing the dnode handle
2710744947dcSTom Erickson  * unless you have some other direct or indirect hold on the dnode. (An indirect
2711744947dcSTom Erickson  * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2712744947dcSTom Erickson  * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2713744947dcSTom Erickson  * dnode's parent dbuf evicting its dnode handles.
2714744947dcSTom Erickson  */
2715fa9e4066Sahrens void
2716ea8dc4b6Seschrock dbuf_rele(dmu_buf_impl_t *db, void *tag)
2717b24ab676SJeff Bonwick {
2718b24ab676SJeff Bonwick 	mutex_enter(&db->db_mtx);
2719b24ab676SJeff Bonwick 	dbuf_rele_and_unlock(db, tag);
2720b24ab676SJeff Bonwick }
2721b24ab676SJeff Bonwick 
272243466aaeSMax Grossman void
272343466aaeSMax Grossman dmu_buf_rele(dmu_buf_t *db, void *tag)
272443466aaeSMax Grossman {
272543466aaeSMax Grossman 	dbuf_rele((dmu_buf_impl_t *)db, tag);
272643466aaeSMax Grossman }
272743466aaeSMax Grossman 
2728b24ab676SJeff Bonwick /*
2729b24ab676SJeff Bonwick  * dbuf_rele() for an already-locked dbuf.  This is necessary to allow
2730b24ab676SJeff Bonwick  * db_dirtycnt and db_holds to be updated atomically.
2731b24ab676SJeff Bonwick  */
2732b24ab676SJeff Bonwick void
2733b24ab676SJeff Bonwick dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2734fa9e4066Sahrens {
2735fa9e4066Sahrens 	int64_t holds;
2736fa9e4066Sahrens 
2737b24ab676SJeff Bonwick 	ASSERT(MUTEX_HELD(&db->db_mtx));
27389c9dc39aSek 	DBUF_VERIFY(db);
2739fa9e4066Sahrens 
2740744947dcSTom Erickson 	/*
2741744947dcSTom Erickson 	 * Remove the reference to the dbuf before removing its hold on the
2742744947dcSTom Erickson 	 * dnode so we can guarantee in dnode_move() that a referenced bonus
2743744947dcSTom Erickson 	 * buffer has a corresponding dnode hold.
2744744947dcSTom Erickson 	 */
2745fa9e4066Sahrens 	holds = refcount_remove(&db->db_holds, tag);
2746ea8dc4b6Seschrock 	ASSERT(holds >= 0);
2747ea8dc4b6Seschrock 
2748c717a561Smaybee 	/*
2749c717a561Smaybee 	 * We can't freeze indirects if there is a possibility that they
2750c717a561Smaybee 	 * may be modified in the current syncing context.
2751c717a561Smaybee 	 */
2752dcbf3bd6SGeorge Wilson 	if (db->db_buf != NULL &&
2753dcbf3bd6SGeorge Wilson 	    holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) {
27546b4acc8bSahrens 		arc_buf_freeze(db->db_buf);
2755dcbf3bd6SGeorge Wilson 	}
27566b4acc8bSahrens 
2757ea8dc4b6Seschrock 	if (holds == db->db_dirtycnt &&
2758d2058105SJustin T. Gibbs 	    db->db_level == 0 && db->db_user_immediate_evict)
2759ea8dc4b6Seschrock 		dbuf_evict_user(db);
2760fa9e4066Sahrens 
2761fa9e4066Sahrens 	if (holds == 0) {
27620a586ceaSMark Shellenbaum 		if (db->db_blkid == DMU_BONUS_BLKID) {
2763cd485b49SJustin T. Gibbs 			dnode_t *dn;
2764d2058105SJustin T. Gibbs 			boolean_t evict_dbuf = db->db_pending_evict;
2765744947dcSTom Erickson 
2766744947dcSTom Erickson 			/*
2767cd485b49SJustin T. Gibbs 			 * If the dnode moves here, we cannot cross this
2768cd485b49SJustin T. Gibbs 			 * barrier until the move completes.
2769744947dcSTom Erickson 			 */
2770744947dcSTom Erickson 			DB_DNODE_ENTER(db);
2771cd485b49SJustin T. Gibbs 
2772cd485b49SJustin T. Gibbs 			dn = DB_DNODE(db);
2773cd485b49SJustin T. Gibbs 			atomic_dec_32(&dn->dn_dbufs_count);
2774cd485b49SJustin T. Gibbs 
2775cd485b49SJustin T. Gibbs 			/*
2776cd485b49SJustin T. Gibbs 			 * Decrementing the dbuf count means that the bonus
2777cd485b49SJustin T. Gibbs 			 * buffer's dnode hold is no longer discounted in
2778cd485b49SJustin T. Gibbs 			 * dnode_move(). The dnode cannot move until after
2779d2058105SJustin T. Gibbs 			 * the dnode_rele() below.
2780cd485b49SJustin T. Gibbs 			 */
2781744947dcSTom Erickson 			DB_DNODE_EXIT(db);
2782cd485b49SJustin T. Gibbs 
2783cd485b49SJustin T. Gibbs 			/*
2784cd485b49SJustin T. Gibbs 			 * Do not reference db after its lock is dropped.
2785cd485b49SJustin T. Gibbs 			 * Another thread may evict it.
2786cd485b49SJustin T. Gibbs 			 */
2787cd485b49SJustin T. Gibbs 			mutex_exit(&db->db_mtx);
2788cd485b49SJustin T. Gibbs 
2789d2058105SJustin T. Gibbs 			if (evict_dbuf)
2790cd485b49SJustin T. Gibbs 				dnode_evict_bonus(dn);
2791d2058105SJustin T. Gibbs 
2792d2058105SJustin T. Gibbs 			dnode_rele(dn, db);
2793ea8dc4b6Seschrock 		} else if (db->db_buf == NULL) {
2794ea8dc4b6Seschrock 			/*
2795ea8dc4b6Seschrock 			 * This is a special case: we never associated this
2796ea8dc4b6Seschrock 			 * dbuf with any data allocated from the ARC.
2797ea8dc4b6Seschrock 			 */
279882c9918fSTim Haley 			ASSERT(db->db_state == DB_UNCACHED ||
279982c9918fSTim Haley 			    db->db_state == DB_NOFILL);
2800dcbf3bd6SGeorge Wilson 			dbuf_destroy(db);
28016b4acc8bSahrens 		} else if (arc_released(db->db_buf)) {
2802ea8dc4b6Seschrock 			/*
2803ea8dc4b6Seschrock 			 * This dbuf has anonymous data associated with it.
2804ea8dc4b6Seschrock 			 */
2805dcbf3bd6SGeorge Wilson 			dbuf_destroy(db);
2806ea8dc4b6Seschrock 		} else {
2807dcbf3bd6SGeorge Wilson 			boolean_t do_arc_evict = B_FALSE;
2808dcbf3bd6SGeorge Wilson 			blkptr_t bp;
2809dcbf3bd6SGeorge Wilson 			spa_t *spa = dmu_objset_spa(db->db_objset);
2810dcbf3bd6SGeorge Wilson 
2811dcbf3bd6SGeorge Wilson 			if (!DBUF_IS_CACHEABLE(db) &&
2812dcbf3bd6SGeorge Wilson 			    db->db_blkptr != NULL &&
2813dcbf3bd6SGeorge Wilson 			    !BP_IS_HOLE(db->db_blkptr) &&
2814dcbf3bd6SGeorge Wilson 			    !BP_IS_EMBEDDED(db->db_blkptr)) {
2815dcbf3bd6SGeorge Wilson 				do_arc_evict = B_TRUE;
2816dcbf3bd6SGeorge Wilson 				bp = *db->db_blkptr;
2817dcbf3bd6SGeorge Wilson 			}
28189253d63dSGeorge Wilson 
2819dcbf3bd6SGeorge Wilson 			if (!DBUF_IS_CACHEABLE(db) ||
2820dcbf3bd6SGeorge Wilson 			    db->db_pending_evict) {
2821dcbf3bd6SGeorge Wilson 				dbuf_destroy(db);
2822dcbf3bd6SGeorge Wilson 			} else if (!multilist_link_active(&db->db_cache_link)) {
282394c2d0ebSMatthew Ahrens 				multilist_insert(dbuf_cache, db);
2824dcbf3bd6SGeorge Wilson 				(void) refcount_add_many(&dbuf_cache_size,
2825dcbf3bd6SGeorge Wilson 				    db->db.db_size, db);
28263baa08fcSek 				mutex_exit(&db->db_mtx);
2827dcbf3bd6SGeorge Wilson 
2828dcbf3bd6SGeorge Wilson 				dbuf_evict_notify();
2829bbfa8ea8SMatthew Ahrens 			}
2830dcbf3bd6SGeorge Wilson 
2831dcbf3bd6SGeorge Wilson 			if (do_arc_evict)
2832dcbf3bd6SGeorge Wilson 				arc_freed(spa, &bp);
2833ea8dc4b6Seschrock 		}
2834fa9e4066Sahrens 	} else {
2835fa9e4066Sahrens 		mutex_exit(&db->db_mtx);
2836fa9e4066Sahrens 	}
2837dcbf3bd6SGeorge Wilson 
2838fa9e4066Sahrens }
2839fa9e4066Sahrens 
2840fa9e4066Sahrens #pragma weak dmu_buf_refcount = dbuf_refcount
2841fa9e4066Sahrens uint64_t
2842fa9e4066Sahrens dbuf_refcount(dmu_buf_impl_t *db)
2843fa9e4066Sahrens {
2844fa9e4066Sahrens 	return (refcount_count(&db->db_holds));
2845fa9e4066Sahrens }
2846fa9e4066Sahrens 
2847fa9e4066Sahrens void *
2848bc9014e6SJustin Gibbs dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user,
2849bc9014e6SJustin Gibbs     dmu_buf_user_t *new_user)
2850fa9e4066Sahrens {
2851bc9014e6SJustin Gibbs 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2852bc9014e6SJustin Gibbs 
2853bc9014e6SJustin Gibbs 	mutex_enter(&db->db_mtx);
2854bc9014e6SJustin Gibbs 	dbuf_verify_user(db, DBVU_NOT_EVICTING);
2855bc9014e6SJustin Gibbs 	if (db->db_user == old_user)
2856bc9014e6SJustin Gibbs 		db->db_user = new_user;
2857bc9014e6SJustin Gibbs 	else
2858bc9014e6SJustin Gibbs 		old_user = db->db_user;
2859bc9014e6SJustin Gibbs 	dbuf_verify_user(db, DBVU_NOT_EVICTING);
2860bc9014e6SJustin Gibbs 	mutex_exit(&db->db_mtx);
2861bc9014e6SJustin Gibbs 
2862bc9014e6SJustin Gibbs 	return (old_user);
2863fa9e4066Sahrens }
2864fa9e4066Sahrens 
2865fa9e4066Sahrens void *
2866bc9014e6SJustin Gibbs dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2867fa9e4066Sahrens {
2868bc9014e6SJustin Gibbs 	return (dmu_buf_replace_user(db_fake, NULL, user));
2869fa9e4066Sahrens }
2870fa9e4066Sahrens 
2871fa9e4066Sahrens void *
2872bc9014e6SJustin Gibbs dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2873fa9e4066Sahrens {
2874fa9e4066Sahrens 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2875fa9e4066Sahrens 
2876d2058105SJustin T. Gibbs 	db->db_user_immediate_evict = TRUE;
2877bc9014e6SJustin Gibbs 	return (dmu_buf_set_user(db_fake, user));
2878bc9014e6SJustin Gibbs }
2879fa9e4066Sahrens 
2880bc9014e6SJustin Gibbs void *
2881bc9014e6SJustin Gibbs dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2882bc9014e6SJustin Gibbs {
2883bc9014e6SJustin Gibbs 	return (dmu_buf_replace_user(db_fake, user, NULL));
2884fa9e4066Sahrens }
2885fa9e4066Sahrens 
2886fa9e4066Sahrens void *
2887fa9e4066Sahrens dmu_buf_get_user(dmu_buf_t *db_fake)
2888fa9e4066Sahrens {
2889fa9e4066Sahrens 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2890fa9e4066Sahrens 
2891bc9014e6SJustin Gibbs 	dbuf_verify_user(db, DBVU_NOT_EVICTING);
2892bc9014e6SJustin Gibbs 	return (db->db_user);
2893bc9014e6SJustin Gibbs }
2894bc9014e6SJustin Gibbs 
2895bc9014e6SJustin Gibbs void
2896bc9014e6SJustin Gibbs dmu_buf_user_evict_wait()
2897bc9014e6SJustin Gibbs {
2898bc9014e6SJustin Gibbs 	taskq_wait(dbu_evict_taskq);
2899fa9e4066Sahrens }
2900fa9e4066Sahrens 
290180901aeaSGeorge Wilson blkptr_t *
290280901aeaSGeorge Wilson dmu_buf_get_blkptr(dmu_buf_t *db)
290380901aeaSGeorge Wilson {
290480901aeaSGeorge Wilson 	dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
290580901aeaSGeorge Wilson 	return (dbi->db_blkptr);
290680901aeaSGeorge Wilson }
290780901aeaSGeorge Wilson 
2908ae972795SMatthew Ahrens objset_t *
2909ae972795SMatthew Ahrens dmu_buf_get_objset(dmu_buf_t *db)
2910ae972795SMatthew Ahrens {
2911ae972795SMatthew Ahrens 	dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2912ae972795SMatthew Ahrens 	return (dbi->db_objset);
2913ae972795SMatthew Ahrens }
2914ae972795SMatthew Ahrens 
291579d72832SMatthew Ahrens dnode_t *
291679d72832SMatthew Ahrens dmu_buf_dnode_enter(dmu_buf_t *db)
291779d72832SMatthew Ahrens {
291879d72832SMatthew Ahrens 	dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
291979d72832SMatthew Ahrens 	DB_DNODE_ENTER(dbi);
292079d72832SMatthew Ahrens 	return (DB_DNODE(dbi));
292179d72832SMatthew Ahrens }
292279d72832SMatthew Ahrens 
292379d72832SMatthew Ahrens void
292479d72832SMatthew Ahrens dmu_buf_dnode_exit(dmu_buf_t *db)
292579d72832SMatthew Ahrens {
292679d72832SMatthew Ahrens 	dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
292779d72832SMatthew Ahrens 	DB_DNODE_EXIT(dbi);
292879d72832SMatthew Ahrens }
292979d72832SMatthew Ahrens 
2930c717a561Smaybee static void
2931c717a561Smaybee dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
2932fa9e4066Sahrens {
2933c717a561Smaybee 	/* ASSERT(dmu_tx_is_syncing(tx) */
2934c717a561Smaybee 	ASSERT(MUTEX_HELD(&db->db_mtx));
2935c717a561Smaybee 
2936c717a561Smaybee 	if (db->db_blkptr != NULL)
2937c717a561Smaybee 		return;
2938c717a561Smaybee 
29390a586ceaSMark Shellenbaum 	if (db->db_blkid == DMU_SPILL_BLKID) {
29400a586ceaSMark Shellenbaum 		db->db_blkptr = &dn->dn_phys->dn_spill;
29410a586ceaSMark Shellenbaum 		BP_ZERO(db->db_blkptr);
29420a586ceaSMark Shellenbaum 		return;
29430a586ceaSMark Shellenbaum 	}
2944c717a561Smaybee 	if (db->db_level == dn->dn_phys->dn_nlevels-1) {
2945c717a561Smaybee 		/*
2946c717a561Smaybee 		 * This buffer was allocated at a time when there was
2947c717a561Smaybee 		 * no available blkptrs from the dnode, or it was
2948c717a561Smaybee 		 * inappropriate to hook it in (i.e., nlevels mis-match).
2949c717a561Smaybee 		 */
2950c717a561Smaybee 		ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
2951c717a561Smaybee 		ASSERT(db->db_parent == NULL);
2952c717a561Smaybee 		db->db_parent = dn->dn_dbuf;
2953c717a561Smaybee 		db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
2954c717a561Smaybee 		DBUF_VERIFY(db);
2955c717a561Smaybee 	} else {
2956c717a561Smaybee 		dmu_buf_impl_t *parent = db->db_parent;
2957c717a561Smaybee 		int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2958c717a561Smaybee 
2959c717a561Smaybee 		ASSERT(dn->dn_phys->dn_nlevels > 1);
2960c717a561Smaybee 		if (parent == NULL) {
2961c717a561Smaybee 			mutex_exit(&db->db_mtx);
2962c717a561Smaybee 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
2963a2cdcdd2SPaul Dagnelie 			parent = dbuf_hold_level(dn, db->db_level + 1,
2964a2cdcdd2SPaul Dagnelie 			    db->db_blkid >> epbs, db);
2965c717a561Smaybee 			rw_exit(&dn->dn_struct_rwlock);
2966c717a561Smaybee 			mutex_enter(&db->db_mtx);
2967c717a561Smaybee 			db->db_parent = parent;
2968c717a561Smaybee 		}
2969c717a561Smaybee 		db->db_blkptr = (blkptr_t *)parent->db.db_data +
2970c717a561Smaybee 		    (db->db_blkid & ((1ULL << epbs) - 1));
2971c717a561Smaybee 		DBUF_VERIFY(db);
2972c717a561Smaybee 	}
2973c717a561Smaybee }
2974c717a561Smaybee 
2975c717a561Smaybee static void
2976c717a561Smaybee dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2977c717a561Smaybee {
2978c717a561Smaybee 	dmu_buf_impl_t *db = dr->dr_dbuf;
2979744947dcSTom Erickson 	dnode_t *dn;
2980c717a561Smaybee 	zio_t *zio;
2981c717a561Smaybee 
2982c717a561Smaybee 	ASSERT(dmu_tx_is_syncing(tx));
2983c717a561Smaybee 
2984c717a561Smaybee 	dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2985c717a561Smaybee 
2986c717a561Smaybee 	mutex_enter(&db->db_mtx);
2987c717a561Smaybee 
2988c717a561Smaybee 	ASSERT(db->db_level > 0);
2989c717a561Smaybee 	DBUF_VERIFY(db);
2990c717a561Smaybee 
29913e30c24aSWill Andrews 	/* Read the block if it hasn't been read yet. */
2992c717a561Smaybee 	if (db->db_buf == NULL) {
2993c717a561Smaybee 		mutex_exit(&db->db_mtx);
2994c717a561Smaybee 		(void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
2995c717a561Smaybee 		mutex_enter(&db->db_mtx);
2996c717a561Smaybee 	}
2997c717a561Smaybee 	ASSERT3U(db->db_state, ==, DB_CACHED);
2998c717a561Smaybee 	ASSERT(db->db_buf != NULL);
2999c717a561Smaybee 
3000744947dcSTom Erickson 	DB_DNODE_ENTER(db);
3001744947dcSTom Erickson 	dn = DB_DNODE(db);
30023e30c24aSWill Andrews 	/* Indirect block size must match what the dnode thinks it is. */
3003744947dcSTom Erickson 	ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3004c717a561Smaybee 	dbuf_check_blkptr(dn, db);
3005744947dcSTom Erickson 	DB_DNODE_EXIT(db);
3006c717a561Smaybee 
30073e30c24aSWill Andrews 	/* Provide the pending dirty record to child dbufs */
3008c717a561Smaybee 	db->db_data_pending = dr;
3009c717a561Smaybee 
3010af2c4821Smaybee 	mutex_exit(&db->db_mtx);
3011*5cabbc6bSPrashanth Sreenivasa 
3012088f3894Sahrens 	dbuf_write(dr, db->db_buf, tx);
3013c717a561Smaybee 
3014c717a561Smaybee 	zio = dr->dr_zio;
3015c717a561Smaybee 	mutex_enter(&dr->dt.di.dr_mtx);
301646e1baa6SMatthew Ahrens 	dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx);
3017c717a561Smaybee 	ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3018c717a561Smaybee 	mutex_exit(&dr->dt.di.dr_mtx);
3019c717a561Smaybee 	zio_nowait(zio);
3020c717a561Smaybee }
3021c717a561Smaybee 
3022c717a561Smaybee static void
3023c717a561Smaybee dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
3024c717a561Smaybee {
3025c717a561Smaybee 	arc_buf_t **datap = &dr->dt.dl.dr_data;
3026c717a561Smaybee 	dmu_buf_impl_t *db = dr->dr_dbuf;
3027744947dcSTom Erickson 	dnode_t *dn;
3028744947dcSTom Erickson 	objset_t *os;
3029c717a561Smaybee 	uint64_t txg = tx->tx_txg;
3030fa9e4066Sahrens 
3031fa9e4066Sahrens 	ASSERT(dmu_tx_is_syncing(tx));
3032fa9e4066Sahrens 
3033fa9e4066Sahrens 	dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
3034fa9e4066Sahrens 
3035fa9e4066Sahrens 	mutex_enter(&db->db_mtx);
3036fa9e4066Sahrens 	/*
3037fa9e4066Sahrens 	 * To be synced, we must be dirtied.  But we
3038fa9e4066Sahrens 	 * might have been freed after the dirty.
3039fa9e4066Sahrens 	 */
3040fa9e4066Sahrens 	if (db->db_state == DB_UNCACHED) {
3041fa9e4066Sahrens 		/* This buffer has been freed since it was dirtied */
3042fa9e4066Sahrens 		ASSERT(db->db.db_data == NULL);
3043fa9e4066Sahrens 	} else if (db->db_state == DB_FILL) {
3044fa9e4066Sahrens 		/* This buffer was freed and is now being re-filled */
3045c717a561Smaybee 		ASSERT(db->db.db_data != dr->dt.dl.dr_data);
3046fa9e4066Sahrens 	} else {
304782c9918fSTim Haley 		ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
3048fa9e4066Sahrens 	}
30499c9dc39aSek 	DBUF_VERIFY(db);
3050fa9e4066Sahrens 
3051744947dcSTom Erickson 	DB_DNODE_ENTER(db);
3052744947dcSTom Erickson 	dn = DB_DNODE(db);
3053744947dcSTom Erickson 
30540a586ceaSMark Shellenbaum 	if (db->db_blkid == DMU_SPILL_BLKID) {
30550a586ceaSMark Shellenbaum 		mutex_enter(&dn->dn_mtx);
30560a586ceaSMark Shellenbaum 		dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
30570a586ceaSMark Shellenbaum 		mutex_exit(&dn->dn_mtx);
30580a586ceaSMark Shellenbaum 	}
30590a586ceaSMark Shellenbaum 
3060fa9e4066Sahrens 	/*
3061c717a561Smaybee 	 * If this is a bonus buffer, simply copy the bonus data into the
3062c717a561Smaybee 	 * dnode.  It will be written out when the dnode is synced (and it
3063c717a561Smaybee 	 * will be synced, since it must have been dirty for dbuf_sync to
3064c717a561Smaybee 	 * be called).
3065fa9e4066Sahrens 	 */
30660a586ceaSMark Shellenbaum 	if (db->db_blkid == DMU_BONUS_BLKID) {
3067c717a561Smaybee 		dbuf_dirty_record_t **drp;
30681934e92fSmaybee 
3069ea8dc4b6Seschrock 		ASSERT(*datap != NULL);
3070fb09f5aaSMadhav Suresh 		ASSERT0(db->db_level);
3071ea8dc4b6Seschrock 		ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN);
3072ea8dc4b6Seschrock 		bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen);
3073744947dcSTom Erickson 		DB_DNODE_EXIT(db);
3074744947dcSTom Erickson 
30750e8c6158Smaybee 		if (*datap != db->db.db_data) {
3076ea8dc4b6Seschrock 			zio_buf_free(*datap, DN_MAX_BONUSLEN);
30775a98e54bSBrendan Gregg - Sun Microsystems 			arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
30780e8c6158Smaybee 		}
3079ea8dc4b6Seschrock 		db->db_data_pending = NULL;
3080c717a561Smaybee 		drp = &db->db_last_dirty;
3081c717a561Smaybee 		while (*drp != dr)
3082c717a561Smaybee 			drp = &(*drp)->dr_next;
308317f17c2dSbonwick 		ASSERT(dr->dr_next == NULL);
3084b24ab676SJeff Bonwick 		ASSERT(dr->dr_dbuf == db);
308517f17c2dSbonwick 		*drp = dr->dr_next;
3086c717a561Smaybee 		kmem_free(dr, sizeof (dbuf_dirty_record_t));
3087ea8dc4b6Seschrock 		ASSERT(db->db_dirtycnt > 0);
3088ea8dc4b6Seschrock 		db->db_dirtycnt -= 1;
3089b24ab676SJeff Bonwick 		dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
3090ea8dc4b6Seschrock 		return;
3091ea8dc4b6Seschrock 	}
3092ea8dc4b6Seschrock 
3093744947dcSTom Erickson 	os = dn->dn_objset;
3094744947dcSTom Erickson 
3095f82bfe17Sgw 	/*
3096f82bfe17Sgw 	 * This function may have dropped the db_mtx lock allowing a dmu_sync
3097f82bfe17Sgw 	 * operation to sneak in. As a result, we need to ensure that we
3098f82bfe17Sgw 	 * don't check the dr_override_state until we have returned from
3099f82bfe17Sgw 	 * dbuf_check_blkptr.
3100f82bfe17Sgw 	 */
3101f82bfe17Sgw 	dbuf_check_blkptr(dn, db);
3102f82bfe17Sgw 
3103c717a561Smaybee 	/*
3104744947dcSTom Erickson 	 * If this buffer is in the middle of an immediate write,
3105c717a561Smaybee 	 * wait for the synchronous IO to complete.
3106c717a561Smaybee 	 */
3107c717a561Smaybee 	while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
3108c717a561Smaybee 		ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
3109c717a561Smaybee 		cv_wait(&db->db_changed, &db->db_mtx);
3110c717a561Smaybee 		ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
3111c717a561Smaybee 	}
3112c5c6ffa0Smaybee 
3113ab69d62fSMatthew Ahrens 	if (db->db_state != DB_NOFILL &&
3114ab69d62fSMatthew Ahrens 	    dn->dn_object != DMU_META_DNODE_OBJECT &&
3115ab69d62fSMatthew Ahrens 	    refcount_count(&db->db_holds) > 1 &&
3116b24ab676SJeff Bonwick 	    dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
3117ab69d62fSMatthew Ahrens 	    *datap == db->db_buf) {
3118ab69d62fSMatthew Ahrens 		/*
3119ab69d62fSMatthew Ahrens 		 * If this buffer is currently "in use" (i.e., there
3120ab69d62fSMatthew Ahrens 		 * are active holds and db_data still references it),
3121ab69d62fSMatthew Ahrens 		 * then make a copy before we start the write so that
3122ab69d62fSMatthew Ahrens 		 * any modifications from the open txg will not leak
3123ab69d62fSMatthew Ahrens 		 * into this write.
3124ab69d62fSMatthew Ahrens 		 *
3125ab69d62fSMatthew Ahrens 		 * NOTE: this copy does not need to be made for
3126ab69d62fSMatthew Ahrens 		 * objects only modified in the syncing context (e.g.
3127ab69d62fSMatthew Ahrens 		 * DNONE_DNODE blocks).
3128ab69d62fSMatthew Ahrens 		 */
31295602294fSDan Kimmel 		int psize = arc_buf_size(*datap);
3130ab69d62fSMatthew Ahrens 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
31315602294fSDan Kimmel 		enum zio_compress compress_type = arc_get_compression(*datap);
31325602294fSDan Kimmel 
31335602294fSDan Kimmel 		if (compress_type == ZIO_COMPRESS_OFF) {
31345602294fSDan Kimmel 			*datap = arc_alloc_buf(os->os_spa, db, type, psize);
31355602294fSDan Kimmel 		} else {
31365602294fSDan Kimmel 			ASSERT3U(type, ==, ARC_BUFC_DATA);
31375602294fSDan Kimmel 			int lsize = arc_buf_lsize(*datap);
31385602294fSDan Kimmel 			*datap = arc_alloc_compressed_buf(os->os_spa, db,
31395602294fSDan Kimmel 			    psize, lsize, compress_type);
31405602294fSDan Kimmel 		}
31415602294fSDan Kimmel 		bcopy(db->db.db_data, (*datap)->b_data, psize);
314282c9918fSTim Haley 	}
3143c717a561Smaybee 	db->db_data_pending = dr;
3144fa9e4066Sahrens 
3145c717a561Smaybee 	mutex_exit(&db->db_mtx);
3146fa9e4066Sahrens 
3147088f3894Sahrens 	dbuf_write(dr, *datap, tx);
3148fa9e4066Sahrens 
3149c717a561Smaybee 	ASSERT(!list_link_active(&dr->dr_dirty_node));
3150744947dcSTom Erickson 	if (dn->dn_object == DMU_META_DNODE_OBJECT) {
3151c717a561Smaybee 		list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
3152744947dcSTom Erickson 		DB_DNODE_EXIT(db);
3153744947dcSTom Erickson 	} else {
3154744947dcSTom Erickson 		/*
3155744947dcSTom Erickson 		 * Although zio_nowait() does not "wait for an IO", it does
3156744947dcSTom Erickson 		 * initiate the IO. If this is an empty write it seems plausible
3157744947dcSTom Erickson 		 * that the IO could actually be completed before the nowait
3158744947dcSTom Erickson 		 * returns. We need to DB_DNODE_EXIT() first in case
3159744947dcSTom Erickson 		 * zio_nowait() invalidates the dbuf.
3160744947dcSTom Erickson 		 */
3161744947dcSTom Erickson 		DB_DNODE_EXIT(db);
3162c717a561Smaybee 		zio_nowait(dr->dr_zio);
3163744947dcSTom Erickson 	}
3164c717a561Smaybee }
316523b11526Smaybee 
3166c717a561Smaybee void
316746e1baa6SMatthew Ahrens dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx)
3168c717a561Smaybee {
3169c717a561Smaybee 	dbuf_dirty_record_t *dr;
3170c717a561Smaybee 
3171c717a561Smaybee 	while (dr = list_head(list)) {
3172c717a561Smaybee 		if (dr->dr_zio != NULL) {
3173c717a561Smaybee 			/*
3174c717a561Smaybee 			 * If we find an already initialized zio then we
3175c717a561Smaybee 			 * are processing the meta-dnode, and we have finished.
3176c717a561Smaybee 			 * The dbufs for all dnodes are put back on the list
3177c717a561Smaybee 			 * during processing, so that we can zio_wait()
3178c717a561Smaybee 			 * these IOs after initiating all child IOs.
3179c717a561Smaybee 			 */
3180c717a561Smaybee 			ASSERT3U(dr->dr_dbuf->db.db_object, ==,
3181c717a561Smaybee 			    DMU_META_DNODE_OBJECT);
3182c717a561Smaybee 			break;
318323b11526Smaybee 		}
318446e1baa6SMatthew Ahrens 		if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
318546e1baa6SMatthew Ahrens 		    dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
318646e1baa6SMatthew Ahrens 			VERIFY3U(dr->dr_dbuf->db_level, ==, level);
318746e1baa6SMatthew Ahrens 		}
3188c717a561Smaybee 		list_remove(list, dr);
3189c717a561Smaybee 		if (dr->dr_dbuf->db_level > 0)
3190c717a561Smaybee 			dbuf_sync_indirect(dr, tx);
3191c717a561Smaybee 		else
3192c717a561Smaybee 			dbuf_sync_leaf(dr, tx);
319323b11526Smaybee 	}
3194c717a561Smaybee }
319523b11526Smaybee 
3196fa9e4066Sahrens /* ARGSUSED */
3197fa9e4066Sahrens static void
3198c717a561Smaybee dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3199fa9e4066Sahrens {
3200fa9e4066Sahrens 	dmu_buf_impl_t *db = vdb;
3201744947dcSTom Erickson 	dnode_t *dn;
3202e14bb325SJeff Bonwick 	blkptr_t *bp = zio->io_bp;
3203c717a561Smaybee 	blkptr_t *bp_orig = &zio->io_bp_orig;
3204b24ab676SJeff Bonwick 	spa_t *spa = zio->io_spa;
3205b24ab676SJeff Bonwick 	int64_t delta;
3206fa9e4066Sahrens 	uint64_t fill = 0;
3207b24ab676SJeff Bonwick 	int i;
3208fa9e4066Sahrens 
320911ceac77SAlex Reece 	ASSERT3P(db->db_blkptr, !=, NULL);
321011ceac77SAlex Reece 	ASSERT3P(&db->db_data_pending->dr_bp_copy, ==, bp);
3211e14bb325SJeff Bonwick 
3212744947dcSTom Erickson 	DB_DNODE_ENTER(db);
3213744947dcSTom Erickson 	dn = DB_DNODE(db);
3214b24ab676SJeff Bonwick 	delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
3215b24ab676SJeff Bonwick 	dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
3216b24ab676SJeff Bonwick 	zio->io_prev_space_delta = delta;
3217fa9e4066Sahrens 
321843466aaeSMax Grossman 	if (bp->blk_birth != 0) {
321943466aaeSMax Grossman 		ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
322043466aaeSMax Grossman 		    BP_GET_TYPE(bp) == dn->dn_type) ||
322143466aaeSMax Grossman 		    (db->db_blkid == DMU_SPILL_BLKID &&
32225d7b4d43SMatthew Ahrens 		    BP_GET_TYPE(bp) == dn->dn_bonustype) ||
32235d7b4d43SMatthew Ahrens 		    BP_IS_EMBEDDED(bp));
322443466aaeSMax Grossman 		ASSERT(BP_GET_LEVEL(bp) == db->db_level);
3225c717a561Smaybee 	}
3226c5c6ffa0Smaybee 
3227c717a561Smaybee 	mutex_enter(&db->db_mtx);
3228fa9e4066Sahrens 
32290a586ceaSMark Shellenbaum #ifdef ZFS_DEBUG
32300a586ceaSMark Shellenbaum 	if (db->db_blkid == DMU_SPILL_BLKID) {
32310a586ceaSMark Shellenbaum 		ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
323211ceac77SAlex Reece 		ASSERT(!(BP_IS_HOLE(bp)) &&
32330a586ceaSMark Shellenbaum 		    db->db_blkptr == &dn->dn_phys->dn_spill);
32340a586ceaSMark Shellenbaum 	}
32350a586ceaSMark Shellenbaum #endif
32360a586ceaSMark Shellenbaum 
3237fa9e4066Sahrens 	if (db->db_level == 0) {
3238fa9e4066Sahrens 		mutex_enter(&dn->dn_mtx);
32390a586ceaSMark Shellenbaum 		if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
32400a586ceaSMark Shellenbaum 		    db->db_blkid != DMU_SPILL_BLKID)
3241fa9e4066Sahrens 			dn->dn_phys->dn_maxblkid = db->db_blkid;
3242fa9e4066Sahrens 		mutex_exit(&dn->dn_mtx);
3243fa9e4066Sahrens 
3244fa9e4066Sahrens 		if (dn->dn_type == DMU_OT_DNODE) {
3245fa9e4066Sahrens 			dnode_phys_t *dnp = db->db.db_data;
3246fa9e4066Sahrens 			for (i = db->db.db_size >> DNODE_SHIFT; i > 0;
3247fa9e4066Sahrens 			    i--, dnp++) {
3248fa9e4066Sahrens 				if (dnp->dn_type != DMU_OT_NONE)
3249fa9e4066Sahrens 					fill++;
3250fa9e4066Sahrens 			}
3251fa9e4066Sahrens 		} else {
325243466aaeSMax Grossman 			if (BP_IS_HOLE(bp)) {
325343466aaeSMax Grossman 				fill = 0;
325443466aaeSMax Grossman 			} else {
325543466aaeSMax Grossman 				fill = 1;
325643466aaeSMax Grossman 			}
3257fa9e4066Sahrens 		}
3258fa9e4066Sahrens 	} else {
3259e14bb325SJeff Bonwick 		blkptr_t *ibp = db->db.db_data;
3260fa9e4066Sahrens 		ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3261e14bb325SJeff Bonwick 		for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
3262e14bb325SJeff Bonwick 			if (BP_IS_HOLE(ibp))
3263fa9e4066Sahrens 				continue;
32645d7b4d43SMatthew Ahrens 			fill += BP_GET_FILL(ibp);
3265fa9e4066Sahrens 		}
3266fa9e4066Sahrens 	}
3267744947dcSTom Erickson 	DB_DNODE_EXIT(db);
3268fa9e4066Sahrens 
32695d7b4d43SMatthew Ahrens 	if (!BP_IS_EMBEDDED(bp))
32705d7b4d43SMatthew Ahrens 		bp->blk_fill = fill;
3271c717a561Smaybee 
3272c717a561Smaybee 	mutex_exit(&db->db_mtx);
327311ceac77SAlex Reece 
327411ceac77SAlex Reece 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
327511ceac77SAlex Reece 	*db->db_blkptr = *bp;
327611ceac77SAlex Reece 	rw_exit(&dn->dn_struct_rwlock);
3277c717a561Smaybee }
3278fa9e4066Sahrens 
32798df0bcf0SPaul Dagnelie /* ARGSUSED */
32808df0bcf0SPaul Dagnelie /*
32818df0bcf0SPaul Dagnelie  * This function gets called just prior to running through the compression
32828df0bcf0SPaul Dagnelie  * stage of the zio pipeline. If we're an indirect block comprised of only
32838df0bcf0SPaul Dagnelie  * holes, then we want this indirect to be compressed away to a hole. In
32848df0bcf0SPaul Dagnelie  * order to do that we must zero out any information about the holes that
32858df0bcf0SPaul Dagnelie  * this indirect points to prior to before we try to compress it.
32868df0bcf0SPaul Dagnelie  */
32878df0bcf0SPaul Dagnelie static void
32888df0bcf0SPaul Dagnelie dbuf_write_children_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
32898df0bcf0SPaul Dagnelie {
32908df0bcf0SPaul Dagnelie 	dmu_buf_impl_t *db = vdb;
32918df0bcf0SPaul Dagnelie 	dnode_t *dn;
32928df0bcf0SPaul Dagnelie 	blkptr_t *bp;
32931a01181fSGeorge Wilson 	unsigned int epbs, i;
32948df0bcf0SPaul Dagnelie 
32958df0bcf0SPaul Dagnelie 	ASSERT3U(db->db_level, >, 0);
32968df0bcf0SPaul Dagnelie 	DB_DNODE_ENTER(db);
32978df0bcf0SPaul Dagnelie 	dn = DB_DNODE(db);
32988df0bcf0SPaul Dagnelie 	epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
32991a01181fSGeorge Wilson 	ASSERT3U(epbs, <, 31);
33008df0bcf0SPaul Dagnelie 
33018df0bcf0SPaul Dagnelie 	/* Determine if all our children are holes */
33028df0bcf0SPaul Dagnelie 	for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++) {
33038df0bcf0SPaul Dagnelie 		if (!BP_IS_HOLE(bp))
33048df0bcf0SPaul Dagnelie 			break;
33058df0bcf0SPaul Dagnelie 	}
33068df0bcf0SPaul Dagnelie 
33078df0bcf0SPaul Dagnelie 	/*
33088df0bcf0SPaul Dagnelie 	 * If all the children are holes, then zero them all out so that
33098df0bcf0SPaul Dagnelie 	 * we may get compressed away.
33108df0bcf0SPaul Dagnelie 	 */
33118df0bcf0SPaul Dagnelie 	if (i == 1 << epbs) {
33121a01181fSGeorge Wilson 		/*
33131a01181fSGeorge Wilson 		 * We only found holes. Grab the rwlock to prevent
33141a01181fSGeorge Wilson 		 * anybody from reading the blocks we're about to
33151a01181fSGeorge Wilson 		 * zero out.
33161a01181fSGeorge Wilson 		 */
33171a01181fSGeorge Wilson 		rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
33188df0bcf0SPaul Dagnelie 		bzero(db->db.db_data, db->db.db_size);
33191a01181fSGeorge Wilson 		rw_exit(&dn->dn_struct_rwlock);
33208df0bcf0SPaul Dagnelie 	}
33218df0bcf0SPaul Dagnelie 	DB_DNODE_EXIT(db);
33228df0bcf0SPaul Dagnelie }
33238df0bcf0SPaul Dagnelie 
332469962b56SMatthew Ahrens /*
332569962b56SMatthew Ahrens  * The SPA will call this callback several times for each zio - once
332669962b56SMatthew Ahrens  * for every physical child i/o (zio->io_phys_children times).  This
332769962b56SMatthew Ahrens  * allows the DMU to monitor the progress of each logical i/o.  For example,
332869962b56SMatthew Ahrens  * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
332969962b56SMatthew Ahrens  * block.  There may be a long delay before all copies/fragments are completed,
333069962b56SMatthew Ahrens  * so this callback allows us to retire dirty space gradually, as the physical
333169962b56SMatthew Ahrens  * i/os complete.
333269962b56SMatthew Ahrens  */
333369962b56SMatthew Ahrens /* ARGSUSED */
333469962b56SMatthew Ahrens static void
333569962b56SMatthew Ahrens dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
333669962b56SMatthew Ahrens {
333769962b56SMatthew Ahrens 	dmu_buf_impl_t *db = arg;
333869962b56SMatthew Ahrens 	objset_t *os = db->db_objset;
333969962b56SMatthew Ahrens 	dsl_pool_t *dp = dmu_objset_pool(os);
334069962b56SMatthew Ahrens 	dbuf_dirty_record_t *dr;
334169962b56SMatthew Ahrens 	int delta = 0;
334269962b56SMatthew Ahrens 
334369962b56SMatthew Ahrens 	dr = db->db_data_pending;
334469962b56SMatthew Ahrens 	ASSERT3U(dr->dr_txg, ==, zio->io_txg);
334569962b56SMatthew Ahrens 
334669962b56SMatthew Ahrens 	/*
334769962b56SMatthew Ahrens 	 * The callback will be called io_phys_children times.  Retire one
334869962b56SMatthew Ahrens 	 * portion of our dirty space each time we are called.  Any rounding
334969962b56SMatthew Ahrens 	 * error will be cleaned up by dsl_pool_sync()'s call to
335069962b56SMatthew Ahrens 	 * dsl_pool_undirty_space().
335169962b56SMatthew Ahrens 	 */
335269962b56SMatthew Ahrens 	delta = dr->dr_accounted / zio->io_phys_children;
335369962b56SMatthew Ahrens 	dsl_pool_undirty_space(dp, delta, zio->io_txg);
335469962b56SMatthew Ahrens }
335569962b56SMatthew Ahrens 
3356c717a561Smaybee /* ARGSUSED */
3357c717a561Smaybee static void
3358c717a561Smaybee dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
3359c717a561Smaybee {
3360c717a561Smaybee 	dmu_buf_impl_t *db = vdb;
3361b24ab676SJeff Bonwick 	blkptr_t *bp_orig = &zio->io_bp_orig;
336243466aaeSMax Grossman 	blkptr_t *bp = db->db_blkptr;
336343466aaeSMax Grossman 	objset_t *os = db->db_objset;
336443466aaeSMax Grossman 	dmu_tx_t *tx = os->os_synctx;
3365c717a561Smaybee 	dbuf_dirty_record_t **drp, *dr;
3366c717a561Smaybee 
3367fb09f5aaSMadhav Suresh 	ASSERT0(zio->io_error);
3368b24ab676SJeff Bonwick 	ASSERT(db->db_blkptr == bp);
3369b24ab676SJeff Bonwick 
337080901aeaSGeorge Wilson 	/*
337180901aeaSGeorge Wilson 	 * For nopwrites and rewrites we ensure that the bp matches our
337280901aeaSGeorge Wilson 	 * original and bypass all the accounting.
337380901aeaSGeorge Wilson 	 */
337480901aeaSGeorge Wilson 	if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
3375b24ab676SJeff Bonwick 		ASSERT(BP_EQUAL(bp, bp_orig));
3376b24ab676SJeff Bonwick 	} else {
337743466aaeSMax Grossman 		dsl_dataset_t *ds = os->os_dsl_dataset;
3378b24ab676SJeff Bonwick 		(void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
3379b24ab676SJeff Bonwick 		dsl_dataset_block_born(ds, bp, tx);
3380b24ab676SJeff Bonwick 	}
3381c717a561Smaybee 
3382c717a561Smaybee 	mutex_enter(&db->db_mtx);
3383c717a561Smaybee 
3384b24ab676SJeff Bonwick 	DBUF_VERIFY(db);
3385b24ab676SJeff Bonwick 
3386c717a561Smaybee 	drp = &db->db_last_dirty;
338717f17c2dSbonwick 	while ((dr = *drp) != db->db_data_pending)
338817f17c2dSbonwick 		drp = &dr->dr_next;
338917f17c2dSbonwick 	ASSERT(!list_link_active(&dr->dr_dirty_node));
3390b24ab676SJeff Bonwick 	ASSERT(dr->dr_dbuf == db);
339117f17c2dSbonwick 	ASSERT(dr->dr_next == NULL);
339217f17c2dSbonwick 	*drp = dr->dr_next;
3393c717a561Smaybee 
33940a586ceaSMark Shellenbaum #ifdef ZFS_DEBUG
33950a586ceaSMark Shellenbaum 	if (db->db_blkid == DMU_SPILL_BLKID) {
3396744947dcSTom Erickson 		dnode_t *dn;
3397744947dcSTom Erickson 
3398744947dcSTom Erickson 		DB_DNODE_ENTER(db);
3399744947dcSTom Erickson 		dn = DB_DNODE(db);
34000a586ceaSMark Shellenbaum 		ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
34010a586ceaSMark Shellenbaum 		ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
34020a586ceaSMark Shellenbaum 		    db->db_blkptr == &dn->dn_phys->dn_spill);
3403744947dcSTom Erickson 		DB_DNODE_EXIT(db);
34040a586ceaSMark Shellenbaum 	}
34050a586ceaSMark Shellenbaum #endif
34060a586ceaSMark Shellenbaum 
3407c717a561Smaybee 	if (db->db_level == 0) {
34080a586ceaSMark Shellenbaum 		ASSERT(db->db_blkid != DMU_BONUS_BLKID);
3409c717a561Smaybee 		ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
341082c9918fSTim Haley 		if (db->db_state != DB_NOFILL) {
341182c9918fSTim Haley 			if (dr->dt.dl.dr_data != db->db_buf)
3412dcbf3bd6SGeorge Wilson 				arc_buf_destroy(dr->dt.dl.dr_data, db);
341382c9918fSTim Haley 		}
3414c717a561Smaybee 	} else {
3415744947dcSTom Erickson 		dnode_t *dn;
3416744947dcSTom Erickson 
3417744947dcSTom Erickson 		DB_DNODE_ENTER(db);
3418744947dcSTom Erickson 		dn = DB_DNODE(db);
3419c717a561Smaybee 		ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
342043466aaeSMax Grossman 		ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
3421c717a561Smaybee 		if (!BP_IS_HOLE(db->db_blkptr)) {
3422c717a561Smaybee 			int epbs =
3423c717a561Smaybee 			    dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
342443466aaeSMax Grossman 			ASSERT3U(db->db_blkid, <=,
342543466aaeSMax Grossman 			    dn->dn_phys->dn_maxblkid >> (db->db_level * epbs));
3426c717a561Smaybee 			ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
3427c717a561Smaybee 			    db->db.db_size);
3428c717a561Smaybee 		}
3429744947dcSTom Erickson 		DB_DNODE_EXIT(db);
3430c25056deSgw 		mutex_destroy(&dr->dt.di.dr_mtx);
3431c25056deSgw 		list_destroy(&dr->dt.di.dr_children);
3432c717a561Smaybee 	}
3433c717a561Smaybee 	kmem_free(dr, sizeof (dbuf_dirty_record_t));
3434fa9e4066Sahrens 
3435fa9e4066Sahrens 	cv_broadcast(&db->db_changed);
3436fa9e4066Sahrens 	ASSERT(db->db_dirtycnt > 0);
3437fa9e4066Sahrens 	db->db_dirtycnt -= 1;
3438c717a561Smaybee 	db->db_data_pending = NULL;
343943466aaeSMax Grossman 	dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg);
3440b24ab676SJeff Bonwick }
3441b24ab676SJeff Bonwick 
3442b24ab676SJeff Bonwick static void
3443b24ab676SJeff Bonwick dbuf_write_nofill_ready(zio_t *zio)
3444b24ab676SJeff Bonwick {
3445b24ab676SJeff Bonwick 	dbuf_write_ready(zio, NULL, zio->io_private);
3446b24ab676SJeff Bonwick }
3447b24ab676SJeff Bonwick 
3448b24ab676SJeff Bonwick static void
3449b24ab676SJeff Bonwick dbuf_write_nofill_done(zio_t *zio)
3450b24ab676SJeff Bonwick {
3451b24ab676SJeff Bonwick 	dbuf_write_done(zio, NULL, zio->io_private);
3452b24ab676SJeff Bonwick }
3453b24ab676SJeff Bonwick 
3454b24ab676SJeff Bonwick static void
3455b24ab676SJeff Bonwick dbuf_write_override_ready(zio_t *zio)
3456b24ab676SJeff Bonwick {
3457b24ab676SJeff Bonwick 	dbuf_dirty_record_t *dr = zio->io_private;
3458b24ab676SJeff Bonwick 	dmu_buf_impl_t *db = dr->dr_dbuf;
3459b24ab676SJeff Bonwick 
3460b24ab676SJeff Bonwick 	dbuf_write_ready(zio, NULL, db);
3461b24ab676SJeff Bonwick }
3462b24ab676SJeff Bonwick 
3463b24ab676SJeff Bonwick static void
3464b24ab676SJeff Bonwick dbuf_write_override_done(zio_t *zio)
3465b24ab676SJeff Bonwick {
3466b24ab676SJeff Bonwick 	dbuf_dirty_record_t *dr = zio->io_private;
3467b24ab676SJeff Bonwick 	dmu_buf_impl_t *db = dr->dr_dbuf;
3468b24ab676SJeff Bonwick 	blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
3469b24ab676SJeff Bonwick 
3470b24ab676SJeff Bonwick 	mutex_enter(&db->db_mtx);
3471b24ab676SJeff Bonwick 	if (!BP_EQUAL(zio->io_bp, obp)) {
3472b24ab676SJeff Bonwick 		if (!BP_IS_HOLE(obp))
3473b24ab676SJeff Bonwick 			dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
3474b24ab676SJeff Bonwick 		arc_release(dr->dt.dl.dr_data, db);
3475b24ab676SJeff Bonwick 	}
3476fa9e4066Sahrens 	mutex_exit(&db->db_mtx);
34774ee0199eSRobert Mustacchi 	dbuf_write_done(zio, NULL, db);
3478770499e1SDan Kimmel 
3479770499e1SDan Kimmel 	if (zio->io_abd != NULL)
3480770499e1SDan Kimmel 		abd_put(zio->io_abd);
3481b24ab676SJeff Bonwick }
3482b24ab676SJeff Bonwick 
3483*5cabbc6bSPrashanth Sreenivasa typedef struct dbuf_remap_impl_callback_arg {
3484*5cabbc6bSPrashanth Sreenivasa 	objset_t	*drica_os;
3485*5cabbc6bSPrashanth Sreenivasa 	uint64_t	drica_blk_birth;
3486*5cabbc6bSPrashanth Sreenivasa 	dmu_tx_t	*drica_tx;
3487*5cabbc6bSPrashanth Sreenivasa } dbuf_remap_impl_callback_arg_t;
3488*5cabbc6bSPrashanth Sreenivasa 
3489*5cabbc6bSPrashanth Sreenivasa static void
3490*5cabbc6bSPrashanth Sreenivasa dbuf_remap_impl_callback(uint64_t vdev, uint64_t offset, uint64_t size,
3491*5cabbc6bSPrashanth Sreenivasa     void *arg)
3492*5cabbc6bSPrashanth Sreenivasa {
3493*5cabbc6bSPrashanth Sreenivasa 	dbuf_remap_impl_callback_arg_t *drica = arg;
3494*5cabbc6bSPrashanth Sreenivasa 	objset_t *os = drica->drica_os;
3495*5cabbc6bSPrashanth Sreenivasa 	spa_t *spa = dmu_objset_spa(os);
3496*5cabbc6bSPrashanth Sreenivasa 	dmu_tx_t *tx = drica->drica_tx;
3497*5cabbc6bSPrashanth Sreenivasa 
3498*5cabbc6bSPrashanth Sreenivasa 	ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)));
3499*5cabbc6bSPrashanth Sreenivasa 
3500*5cabbc6bSPrashanth Sreenivasa 	if (os == spa_meta_objset(spa)) {
3501*5cabbc6bSPrashanth Sreenivasa 		spa_vdev_indirect_mark_obsolete(spa, vdev, offset, size, tx);
3502*5cabbc6bSPrashanth Sreenivasa 	} else {
3503*5cabbc6bSPrashanth Sreenivasa 		dsl_dataset_block_remapped(dmu_objset_ds(os), vdev, offset,
3504*5cabbc6bSPrashanth Sreenivasa 		    size, drica->drica_blk_birth, tx);
3505*5cabbc6bSPrashanth Sreenivasa 	}
3506*5cabbc6bSPrashanth Sreenivasa }
3507*5cabbc6bSPrashanth Sreenivasa 
3508*5cabbc6bSPrashanth Sreenivasa static void
3509*5cabbc6bSPrashanth Sreenivasa dbuf_remap_impl(dnode_t *dn, blkptr_t *bp, dmu_tx_t *tx)
3510*5cabbc6bSPrashanth Sreenivasa {
3511*5cabbc6bSPrashanth Sreenivasa 	blkptr_t bp_copy = *bp;
3512*5cabbc6bSPrashanth Sreenivasa 	spa_t *spa = dmu_objset_spa(dn->dn_objset);
3513*5cabbc6bSPrashanth Sreenivasa 	dbuf_remap_impl_callback_arg_t drica;
3514*5cabbc6bSPrashanth Sreenivasa 
3515*5cabbc6bSPrashanth Sreenivasa 	ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)));
3516*5cabbc6bSPrashanth Sreenivasa 
3517*5cabbc6bSPrashanth Sreenivasa 	drica.drica_os = dn->dn_objset;
3518*5cabbc6bSPrashanth Sreenivasa 	drica.drica_blk_birth = bp->blk_birth;
3519*5cabbc6bSPrashanth Sreenivasa 	drica.drica_tx = tx;
3520*5cabbc6bSPrashanth Sreenivasa 	if (spa_remap_blkptr(spa, &bp_copy, dbuf_remap_impl_callback,
3521*5cabbc6bSPrashanth Sreenivasa 	    &drica)) {
3522*5cabbc6bSPrashanth Sreenivasa 		/*
3523*5cabbc6bSPrashanth Sreenivasa 		 * The struct_rwlock prevents dbuf_read_impl() from
3524*5cabbc6bSPrashanth Sreenivasa 		 * dereferencing the BP while we are changing it.  To
3525*5cabbc6bSPrashanth Sreenivasa 		 * avoid lock contention, only grab it when we are actually
3526*5cabbc6bSPrashanth Sreenivasa 		 * changing the BP.
3527*5cabbc6bSPrashanth Sreenivasa 		 */
3528*5cabbc6bSPrashanth Sreenivasa 		rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3529*5cabbc6bSPrashanth Sreenivasa 		*bp = bp_copy;
3530*5cabbc6bSPrashanth Sreenivasa 		rw_exit(&dn->dn_struct_rwlock);
3531*5cabbc6bSPrashanth Sreenivasa 	}
3532*5cabbc6bSPrashanth Sreenivasa }
3533*5cabbc6bSPrashanth Sreenivasa 
3534*5cabbc6bSPrashanth Sreenivasa /*
3535*5cabbc6bSPrashanth Sreenivasa  * Returns true if a dbuf_remap would modify the dbuf. We do this by attempting
3536*5cabbc6bSPrashanth Sreenivasa  * to remap a copy of every bp in the dbuf.
3537*5cabbc6bSPrashanth Sreenivasa  */
3538*5cabbc6bSPrashanth Sreenivasa boolean_t
3539*5cabbc6bSPrashanth Sreenivasa dbuf_can_remap(const dmu_buf_impl_t *db)
3540*5cabbc6bSPrashanth Sreenivasa {
3541*5cabbc6bSPrashanth Sreenivasa 	spa_t *spa = dmu_objset_spa(db->db_objset);
3542*5cabbc6bSPrashanth Sreenivasa 	blkptr_t *bp = db->db.db_data;
3543*5cabbc6bSPrashanth Sreenivasa 	boolean_t ret = B_FALSE;
3544*5cabbc6bSPrashanth Sreenivasa 
3545*5cabbc6bSPrashanth Sreenivasa 	ASSERT3U(db->db_level, >, 0);
3546*5cabbc6bSPrashanth Sreenivasa 	ASSERT3S(db->db_state, ==, DB_CACHED);
3547*5cabbc6bSPrashanth Sreenivasa 
3548*5cabbc6bSPrashanth Sreenivasa 	ASSERT(spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL));
3549*5cabbc6bSPrashanth Sreenivasa 
3550*5cabbc6bSPrashanth Sreenivasa 	spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
3551*5cabbc6bSPrashanth Sreenivasa 	for (int i = 0; i < db->db.db_size >> SPA_BLKPTRSHIFT; i++) {
3552*5cabbc6bSPrashanth Sreenivasa 		blkptr_t bp_copy = bp[i];
3553*5cabbc6bSPrashanth Sreenivasa 		if (spa_remap_blkptr(spa, &bp_copy, NULL, NULL)) {
3554*5cabbc6bSPrashanth Sreenivasa 			ret = B_TRUE;
3555*5cabbc6bSPrashanth Sreenivasa 			break;
3556*5cabbc6bSPrashanth Sreenivasa 		}
3557*5cabbc6bSPrashanth Sreenivasa 	}
3558*5cabbc6bSPrashanth Sreenivasa 	spa_config_exit(spa, SCL_VDEV, FTAG);
3559*5cabbc6bSPrashanth Sreenivasa 
3560*5cabbc6bSPrashanth Sreenivasa 	return (ret);
3561*5cabbc6bSPrashanth Sreenivasa }
3562*5cabbc6bSPrashanth Sreenivasa 
3563*5cabbc6bSPrashanth Sreenivasa boolean_t
3564*5cabbc6bSPrashanth Sreenivasa dnode_needs_remap(const dnode_t *dn)
3565*5cabbc6bSPrashanth Sreenivasa {
3566*5cabbc6bSPrashanth Sreenivasa 	spa_t *spa = dmu_objset_spa(dn->dn_objset);
3567*5cabbc6bSPrashanth Sreenivasa 	boolean_t ret = B_FALSE;
3568*5cabbc6bSPrashanth Sreenivasa 
3569*5cabbc6bSPrashanth Sreenivasa 	if (dn->dn_phys->dn_nlevels == 0) {
3570*5cabbc6bSPrashanth Sreenivasa 		return (B_FALSE);
3571*5cabbc6bSPrashanth Sreenivasa 	}
3572*5cabbc6bSPrashanth Sreenivasa 
3573*5cabbc6bSPrashanth Sreenivasa 	ASSERT(spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL));
3574*5cabbc6bSPrashanth Sreenivasa 
3575*5cabbc6bSPrashanth Sreenivasa 	spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER);
3576*5cabbc6bSPrashanth Sreenivasa 	for (int j = 0; j < dn->dn_phys->dn_nblkptr; j++) {
3577*5cabbc6bSPrashanth Sreenivasa 		blkptr_t bp_copy = dn->dn_phys->dn_blkptr[j];
3578*5cabbc6bSPrashanth Sreenivasa 		if (spa_remap_blkptr(spa, &bp_copy, NULL, NULL)) {
3579*5cabbc6bSPrashanth Sreenivasa 			ret = B_TRUE;
3580*5cabbc6bSPrashanth Sreenivasa 			break;
3581*5cabbc6bSPrashanth Sreenivasa 		}
3582*5cabbc6bSPrashanth Sreenivasa 	}
3583*5cabbc6bSPrashanth Sreenivasa 	spa_config_exit(spa, SCL_VDEV, FTAG);
3584*5cabbc6bSPrashanth Sreenivasa 
3585*5cabbc6bSPrashanth Sreenivasa 	return (ret);
3586*5cabbc6bSPrashanth Sreenivasa }
3587*5cabbc6bSPrashanth Sreenivasa 
3588*5cabbc6bSPrashanth Sreenivasa /*
3589*5cabbc6bSPrashanth Sreenivasa  * Remap any existing BP's to concrete vdevs, if possible.
3590*5cabbc6bSPrashanth Sreenivasa  */
3591*5cabbc6bSPrashanth Sreenivasa static void
3592*5cabbc6bSPrashanth Sreenivasa dbuf_remap(dnode_t *dn, dmu_buf_impl_t *db, dmu_tx_t *tx)
3593*5cabbc6bSPrashanth Sreenivasa {
3594*5cabbc6bSPrashanth Sreenivasa 	spa_t *spa = dmu_objset_spa(db->db_objset);
3595*5cabbc6bSPrashanth Sreenivasa 	ASSERT(dsl_pool_sync_context(spa_get_dsl(spa)));
3596*5cabbc6bSPrashanth Sreenivasa 
3597*5cabbc6bSPrashanth Sreenivasa 	if (!spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL))
3598*5cabbc6bSPrashanth Sreenivasa 		return;
3599*5cabbc6bSPrashanth Sreenivasa 
3600*5cabbc6bSPrashanth Sreenivasa 	if (db->db_level > 0) {
3601*5cabbc6bSPrashanth Sreenivasa 		blkptr_t *bp = db->db.db_data;
3602*5cabbc6bSPrashanth Sreenivasa 		for (int i = 0; i < db->db.db_size >> SPA_BLKPTRSHIFT; i++) {
3603*5cabbc6bSPrashanth Sreenivasa 			dbuf_remap_impl(dn, &bp[i], tx);
3604*5cabbc6bSPrashanth Sreenivasa 		}
3605*5cabbc6bSPrashanth Sreenivasa 	} else if (db->db.db_object == DMU_META_DNODE_OBJECT) {
3606*5cabbc6bSPrashanth Sreenivasa 		dnode_phys_t *dnp = db->db.db_data;
3607*5cabbc6bSPrashanth Sreenivasa 		ASSERT3U(db->db_dnode_handle->dnh_dnode->dn_type, ==,
3608*5cabbc6bSPrashanth Sreenivasa 		    DMU_OT_DNODE);
3609*5cabbc6bSPrashanth Sreenivasa 		for (int i = 0; i < db->db.db_size >> DNODE_SHIFT; i++) {
3610*5cabbc6bSPrashanth Sreenivasa 			for (int j = 0; j < dnp[i].dn_nblkptr; j++) {
3611*5cabbc6bSPrashanth Sreenivasa 				dbuf_remap_impl(dn, &dnp[i].dn_blkptr[j], tx);
3612*5cabbc6bSPrashanth Sreenivasa 			}
3613*5cabbc6bSPrashanth Sreenivasa 		}
3614*5cabbc6bSPrashanth Sreenivasa 	}
3615*5cabbc6bSPrashanth Sreenivasa }
3616*5cabbc6bSPrashanth Sreenivasa 
3617*5cabbc6bSPrashanth Sreenivasa 
36183e30c24aSWill Andrews /* Issue I/O to commit a dirty buffer to disk. */
3619b24ab676SJeff Bonwick static void
3620b24ab676SJeff Bonwick dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
3621b24ab676SJeff Bonwick {
3622b24ab676SJeff Bonwick 	dmu_buf_impl_t *db = dr->dr_dbuf;
3623744947dcSTom Erickson 	dnode_t *dn;
3624744947dcSTom Erickson 	objset_t *os;
3625b24ab676SJeff Bonwick 	dmu_buf_impl_t *parent = db->db_parent;
3626b24ab676SJeff Bonwick 	uint64_t txg = tx->tx_txg;
36277802d7bfSMatthew Ahrens 	zbookmark_phys_t zb;
3628b24ab676SJeff Bonwick 	zio_prop_t zp;
3629b24ab676SJeff Bonwick 	zio_t *zio;
36300a586ceaSMark Shellenbaum 	int wp_flag = 0;
3631b24ab676SJeff Bonwick 
363211ceac77SAlex Reece 	ASSERT(dmu_tx_is_syncing(tx));
363311ceac77SAlex Reece 
3634744947dcSTom Erickson 	DB_DNODE_ENTER(db);
3635744947dcSTom Erickson 	dn = DB_DNODE(db);
3636744947dcSTom Erickson 	os = dn->dn_objset;
3637744947dcSTom Erickson 
3638b24ab676SJeff Bonwick 	if (db->db_state != DB_NOFILL) {
3639b24ab676SJeff Bonwick 		if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
3640b24ab676SJeff Bonwick 			/*
3641b24ab676SJeff Bonwick 			 * Private object buffers are released here rather
3642b24ab676SJeff Bonwick 			 * than in dbuf_dirty() since they are only modified
3643b24ab676SJeff Bonwick 			 * in the syncing context and we don't want the
3644b24ab676SJeff Bonwick 			 * overhead of making multiple copies of the data.
3645b24ab676SJeff Bonwick 			 */
3646b24ab676SJeff Bonwick 			if (BP_IS_HOLE(db->db_blkptr)) {
3647b24ab676SJeff Bonwick 				arc_buf_thaw(data);
3648b24ab676SJeff Bonwick 			} else {
36493f9d6ad7SLin Ling 				dbuf_release_bp(db);
3650b24ab676SJeff Bonwick 			}
3651*5cabbc6bSPrashanth Sreenivasa 			dbuf_remap(dn, db, tx);
3652b24ab676SJeff Bonwick 		}
3653b24ab676SJeff Bonwick 	}
3654b24ab676SJeff Bonwick 
3655b24ab676SJeff Bonwick 	if (parent != dn->dn_dbuf) {
36563e30c24aSWill Andrews 		/* Our parent is an indirect block. */
36573e30c24aSWill Andrews 		/* We have a dirty parent that has been scheduled for write. */
3658b24ab676SJeff Bonwick 		ASSERT(parent && parent->db_data_pending);
36593e30c24aSWill Andrews 		/* Our parent's buffer is one level closer to the dnode. */
3660b24ab676SJeff Bonwick 		ASSERT(db->db_level == parent->db_level-1);
36613e30c24aSWill Andrews 		/*
36623e30c24aSWill Andrews 		 * We're about to modify our parent's db_data by modifying
36633e30c24aSWill Andrews 		 * our block pointer, so the parent must be released.
36643e30c24aSWill Andrews 		 */
3665b24ab676SJeff Bonwick 		ASSERT(arc_released(parent->db_buf));
3666b24ab676SJeff Bonwick 		zio = parent->db_data_pending->dr_zio;
3667b24ab676SJeff Bonwick 	} else {
36683e30c24aSWill Andrews 		/* Our parent is the dnode itself. */
36690a586ceaSMark Shellenbaum 		ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
36700a586ceaSMark Shellenbaum 		    db->db_blkid != DMU_SPILL_BLKID) ||
36710a586ceaSMark Shellenbaum 		    (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
36720a586ceaSMark Shellenbaum 		if (db->db_blkid != DMU_SPILL_BLKID)
36730a586ceaSMark Shellenbaum 			ASSERT3P(db->db_blkptr, ==,
36740a586ceaSMark Shellenbaum 			    &dn->dn_phys->dn_blkptr[db->db_blkid]);
3675b24ab676SJeff Bonwick 		zio = dn->dn_zio;
3676b24ab676SJeff Bonwick 	}
3677b24ab676SJeff Bonwick 
3678b24ab676SJeff Bonwick 	ASSERT(db->db_level == 0 || data == db->db_buf);
3679b24ab676SJeff Bonwick 	ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
3680b24ab676SJeff Bonwick 	ASSERT(zio);
3681fa9e4066Sahrens 
3682b24ab676SJeff Bonwick 	SET_BOOKMARK(&zb, os->os_dsl_dataset ?
3683b24ab676SJeff Bonwick 	    os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
3684b24ab676SJeff Bonwick 	    db->db.db_object, db->db_level, db->db_blkid);
3685b24ab676SJeff Bonwick 
36860a586ceaSMark Shellenbaum 	if (db->db_blkid == DMU_SPILL_BLKID)
36870a586ceaSMark Shellenbaum 		wp_flag = WP_SPILL;
36880a586ceaSMark Shellenbaum 	wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
36890a586ceaSMark Shellenbaum 
3690adaec86aSMatthew Ahrens 	dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
3691744947dcSTom Erickson 	DB_DNODE_EXIT(db);
3692b24ab676SJeff Bonwick 
369311ceac77SAlex Reece 	/*
369411ceac77SAlex Reece 	 * We copy the blkptr now (rather than when we instantiate the dirty
369511ceac77SAlex Reece 	 * record), because its value can change between open context and
369611ceac77SAlex Reece 	 * syncing context. We do not need to hold dn_struct_rwlock to read
369711ceac77SAlex Reece 	 * db_blkptr because we are in syncing context.
369811ceac77SAlex Reece 	 */
369911ceac77SAlex Reece 	dr->dr_bp_copy = *db->db_blkptr;
370011ceac77SAlex Reece 
37015d7b4d43SMatthew Ahrens 	if (db->db_level == 0 &&
37025d7b4d43SMatthew Ahrens 	    dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
37035d7b4d43SMatthew Ahrens 		/*
37045d7b4d43SMatthew Ahrens 		 * The BP for this block has been provided by open context
37055d7b4d43SMatthew Ahrens 		 * (by dmu_sync() or dmu_buf_write_embedded()).
37065d7b4d43SMatthew Ahrens 		 */
3707770499e1SDan Kimmel 		abd_t *contents = (data != NULL) ?
3708770499e1SDan Kimmel 		    abd_get_from_buf(data->b_data, arc_buf_size(data)) : NULL;
37095d7b4d43SMatthew Ahrens 
37105602294fSDan Kimmel 		dr->dr_zio = zio_write(zio, os->os_spa, txg, &dr->dr_bp_copy,
37115602294fSDan Kimmel 		    contents, db->db.db_size, db->db.db_size, &zp,
37128df0bcf0SPaul Dagnelie 		    dbuf_write_override_ready, NULL, NULL,
37138df0bcf0SPaul Dagnelie 		    dbuf_write_override_done,
371469962b56SMatthew Ahrens 		    dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
3715b24ab676SJeff Bonwick 		mutex_enter(&db->db_mtx);
3716b24ab676SJeff Bonwick 		dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
3717b24ab676SJeff Bonwick 		zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
371880901aeaSGeorge Wilson 		    dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
3719b24ab676SJeff Bonwick 		mutex_exit(&db->db_mtx);
3720b24ab676SJeff Bonwick 	} else if (db->db_state == DB_NOFILL) {
3721810e43b2SBill Pijewski 		ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF ||
3722810e43b2SBill Pijewski 		    zp.zp_checksum == ZIO_CHECKSUM_NOPARITY);
3723b24ab676SJeff Bonwick 		dr->dr_zio = zio_write(zio, os->os_spa, txg,
37245602294fSDan Kimmel 		    &dr->dr_bp_copy, NULL, db->db.db_size, db->db.db_size, &zp,
37258df0bcf0SPaul Dagnelie 		    dbuf_write_nofill_ready, NULL, NULL,
37268df0bcf0SPaul Dagnelie 		    dbuf_write_nofill_done, db,
3727b24ab676SJeff Bonwick 		    ZIO_PRIORITY_ASYNC_WRITE,
3728b24ab676SJeff Bonwick 		    ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
3729b24ab676SJeff Bonwick 	} else {
3730b24ab676SJeff Bonwick 		ASSERT(arc_released(data));
37318df0bcf0SPaul Dagnelie 
37328df0bcf0SPaul Dagnelie 		/*
37338df0bcf0SPaul Dagnelie 		 * For indirect blocks, we want to setup the children
37348df0bcf0SPaul Dagnelie 		 * ready callback so that we can properly handle an indirect
37358df0bcf0SPaul Dagnelie 		 * block that only contains holes.
37368df0bcf0SPaul Dagnelie 		 */
37378df0bcf0SPaul Dagnelie 		arc_done_func_t *children_ready_cb = NULL;
37388df0bcf0SPaul Dagnelie 		if (db->db_level != 0)
37398df0bcf0SPaul Dagnelie 			children_ready_cb = dbuf_write_children_ready;
37408df0bcf0SPaul Dagnelie 
3741b24ab676SJeff Bonwick 		dr->dr_zio = arc_write(zio, os->os_spa, txg,
374211ceac77SAlex Reece 		    &dr->dr_bp_copy, data, DBUF_IS_L2CACHEABLE(db),
3743dcbf3bd6SGeorge Wilson 		    &zp, dbuf_write_ready, children_ready_cb,
374469962b56SMatthew Ahrens 		    dbuf_write_physdone, dbuf_write_done, db,
374569962b56SMatthew Ahrens 		    ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
3746b24ab676SJeff Bonwick 	}
3747fa9e4066Sahrens }
3748