xref: /illumos-gate/usr/src/uts/common/fs/zfs/dbuf.c (revision 5cabbc6b)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
24  * Copyright (c) 2012, 2017 by Delphix. All rights reserved.
25  * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
26  * Copyright (c) 2013, Joyent, Inc. All rights reserved.
27  * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
28  * Copyright (c) 2014 Integros [integros.com]
29  */
30 
31 #include <sys/zfs_context.h>
32 #include <sys/dmu.h>
33 #include <sys/dmu_send.h>
34 #include <sys/dmu_impl.h>
35 #include <sys/dbuf.h>
36 #include <sys/dmu_objset.h>
37 #include <sys/dsl_dataset.h>
38 #include <sys/dsl_dir.h>
39 #include <sys/dmu_tx.h>
40 #include <sys/spa.h>
41 #include <sys/zio.h>
42 #include <sys/dmu_zfetch.h>
43 #include <sys/sa.h>
44 #include <sys/sa_impl.h>
45 #include <sys/zfeature.h>
46 #include <sys/blkptr.h>
47 #include <sys/range_tree.h>
48 #include <sys/callb.h>
49 #include <sys/abd.h>
50 #include <sys/vdev.h>
51 
52 uint_t zfs_dbuf_evict_key;
53 
54 static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
55 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
56 
57 #ifndef __lint
58 extern inline void dmu_buf_init_user(dmu_buf_user_t *dbu,
59     dmu_buf_evict_func_t *evict_func_sync,
60     dmu_buf_evict_func_t *evict_func_async,
61     dmu_buf_t **clear_on_evict_dbufp);
62 #endif /* ! __lint */
63 
64 /*
65  * Global data structures and functions for the dbuf cache.
66  */
67 static kmem_cache_t *dbuf_kmem_cache;
68 static taskq_t *dbu_evict_taskq;
69 
70 static kthread_t *dbuf_cache_evict_thread;
71 static kmutex_t dbuf_evict_lock;
72 static kcondvar_t dbuf_evict_cv;
73 static boolean_t dbuf_evict_thread_exit;
74 
75 /*
76  * LRU cache of dbufs. The dbuf cache maintains a list of dbufs that
77  * are not currently held but have been recently released. These dbufs
78  * are not eligible for arc eviction until they are aged out of the cache.
79  * Dbufs are added to the dbuf cache once the last hold is released. If a
80  * dbuf is later accessed and still exists in the dbuf cache, then it will
81  * be removed from the cache and later re-added to the head of the cache.
82  * Dbufs that are aged out of the cache will be immediately destroyed and
83  * become eligible for arc eviction.
84  */
85 static multilist_t *dbuf_cache;
86 static refcount_t dbuf_cache_size;
87 uint64_t dbuf_cache_max_bytes = 100 * 1024 * 1024;
88 
89 /* Cap the size of the dbuf cache to log2 fraction of arc size. */
90 int dbuf_cache_max_shift = 5;
91 
92 /*
93  * The dbuf cache uses a three-stage eviction policy:
94  *	- A low water marker designates when the dbuf eviction thread
95  *	should stop evicting from the dbuf cache.
96  *	- When we reach the maximum size (aka mid water mark), we
97  *	signal the eviction thread to run.
98  *	- The high water mark indicates when the eviction thread
99  *	is unable to keep up with the incoming load and eviction must
100  *	happen in the context of the calling thread.
101  *
102  * The dbuf cache:
103  *                                                 (max size)
104  *                                      low water   mid water   hi water
105  * +----------------------------------------+----------+----------+
106  * |                                        |          |          |
107  * |                                        |          |          |
108  * |                                        |          |          |
109  * |                                        |          |          |
110  * +----------------------------------------+----------+----------+
111  *                                        stop        signal     evict
112  *                                      evicting     eviction   directly
113  *                                                    thread
114  *
115  * The high and low water marks indicate the operating range for the eviction
116  * thread. The low water mark is, by default, 90% of the total size of the
117  * cache and the high water mark is at 110% (both of these percentages can be
118  * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct,
119  * respectively). The eviction thread will try to ensure that the cache remains
120  * within this range by waking up every second and checking if the cache is
121  * above the low water mark. The thread can also be woken up by callers adding
122  * elements into the cache if the cache is larger than the mid water (i.e max
123  * cache size). Once the eviction thread is woken up and eviction is required,
124  * it will continue evicting buffers until it's able to reduce the cache size
125  * to the low water mark. If the cache size continues to grow and hits the high
126  * water mark, then callers adding elments to the cache will begin to evict
127  * directly from the cache until the cache is no longer above the high water
128  * mark.
129  */
130 
131 /*
132  * The percentage above and below the maximum cache size.
133  */
134 uint_t dbuf_cache_hiwater_pct = 10;
135 uint_t dbuf_cache_lowater_pct = 10;
136 
137 /* ARGSUSED */
138 static int
139 dbuf_cons(void *vdb, void *unused, int kmflag)
140 {
141 	dmu_buf_impl_t *db = vdb;
142 	bzero(db, sizeof (dmu_buf_impl_t));
143 
144 	mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
145 	cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
146 	multilist_link_init(&db->db_cache_link);
147 	refcount_create(&db->db_holds);
148 
149 	return (0);
150 }
151 
152 /* ARGSUSED */
153 static void
154 dbuf_dest(void *vdb, void *unused)
155 {
156 	dmu_buf_impl_t *db = vdb;
157 	mutex_destroy(&db->db_mtx);
158 	cv_destroy(&db->db_changed);
159 	ASSERT(!multilist_link_active(&db->db_cache_link));
160 	refcount_destroy(&db->db_holds);
161 }
162 
163 /*
164  * dbuf hash table routines
165  */
166 static dbuf_hash_table_t dbuf_hash_table;
167 
168 static uint64_t dbuf_hash_count;
169 
170 static uint64_t
171 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
172 {
173 	uintptr_t osv = (uintptr_t)os;
174 	uint64_t crc = -1ULL;
175 
176 	ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
177 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF];
178 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF];
179 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF];
180 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF];
181 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF];
182 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF];
183 
184 	crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16);
185 
186 	return (crc);
187 }
188 
189 #define	DBUF_EQUAL(dbuf, os, obj, level, blkid)		\
190 	((dbuf)->db.db_object == (obj) &&		\
191 	(dbuf)->db_objset == (os) &&			\
192 	(dbuf)->db_level == (level) &&			\
193 	(dbuf)->db_blkid == (blkid))
194 
195 dmu_buf_impl_t *
196 dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid)
197 {
198 	dbuf_hash_table_t *h = &dbuf_hash_table;
199 	uint64_t hv = dbuf_hash(os, obj, level, blkid);
200 	uint64_t idx = hv & h->hash_table_mask;
201 	dmu_buf_impl_t *db;
202 
203 	mutex_enter(DBUF_HASH_MUTEX(h, idx));
204 	for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
205 		if (DBUF_EQUAL(db, os, obj, level, blkid)) {
206 			mutex_enter(&db->db_mtx);
207 			if (db->db_state != DB_EVICTING) {
208 				mutex_exit(DBUF_HASH_MUTEX(h, idx));
209 				return (db);
210 			}
211 			mutex_exit(&db->db_mtx);
212 		}
213 	}
214 	mutex_exit(DBUF_HASH_MUTEX(h, idx));
215 	return (NULL);
216 }
217 
218 static dmu_buf_impl_t *
219 dbuf_find_bonus(objset_t *os, uint64_t object)
220 {
221 	dnode_t *dn;
222 	dmu_buf_impl_t *db = NULL;
223 
224 	if (dnode_hold(os, object, FTAG, &dn) == 0) {
225 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
226 		if (dn->dn_bonus != NULL) {
227 			db = dn->dn_bonus;
228 			mutex_enter(&db->db_mtx);
229 		}
230 		rw_exit(&dn->dn_struct_rwlock);
231 		dnode_rele(dn, FTAG);
232 	}
233 	return (db);
234 }
235 
236 /*
237  * Insert an entry into the hash table.  If there is already an element
238  * equal to elem in the hash table, then the already existing element
239  * will be returned and the new element will not be inserted.
240  * Otherwise returns NULL.
241  */
242 static dmu_buf_impl_t *
243 dbuf_hash_insert(dmu_buf_impl_t *db)
244 {
245 	dbuf_hash_table_t *h = &dbuf_hash_table;
246 	objset_t *os = db->db_objset;
247 	uint64_t obj = db->db.db_object;
248 	int level = db->db_level;
249 	uint64_t blkid = db->db_blkid;
250 	uint64_t hv = dbuf_hash(os, obj, level, blkid);
251 	uint64_t idx = hv & h->hash_table_mask;
252 	dmu_buf_impl_t *dbf;
253 
254 	mutex_enter(DBUF_HASH_MUTEX(h, idx));
255 	for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) {
256 		if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
257 			mutex_enter(&dbf->db_mtx);
258 			if (dbf->db_state != DB_EVICTING) {
259 				mutex_exit(DBUF_HASH_MUTEX(h, idx));
260 				return (dbf);
261 			}
262 			mutex_exit(&dbf->db_mtx);
263 		}
264 	}
265 
266 	mutex_enter(&db->db_mtx);
267 	db->db_hash_next = h->hash_table[idx];
268 	h->hash_table[idx] = db;
269 	mutex_exit(DBUF_HASH_MUTEX(h, idx));
270 	atomic_inc_64(&dbuf_hash_count);
271 
272 	return (NULL);
273 }
274 
275 /*
276  * Remove an entry from the hash table.  It must be in the EVICTING state.
277  */
278 static void
279 dbuf_hash_remove(dmu_buf_impl_t *db)
280 {
281 	dbuf_hash_table_t *h = &dbuf_hash_table;
282 	uint64_t hv = dbuf_hash(db->db_objset, db->db.db_object,
283 	    db->db_level, db->db_blkid);
284 	uint64_t idx = hv & h->hash_table_mask;
285 	dmu_buf_impl_t *dbf, **dbp;
286 
287 	/*
288 	 * We musn't hold db_mtx to maintain lock ordering:
289 	 * DBUF_HASH_MUTEX > db_mtx.
290 	 */
291 	ASSERT(refcount_is_zero(&db->db_holds));
292 	ASSERT(db->db_state == DB_EVICTING);
293 	ASSERT(!MUTEX_HELD(&db->db_mtx));
294 
295 	mutex_enter(DBUF_HASH_MUTEX(h, idx));
296 	dbp = &h->hash_table[idx];
297 	while ((dbf = *dbp) != db) {
298 		dbp = &dbf->db_hash_next;
299 		ASSERT(dbf != NULL);
300 	}
301 	*dbp = db->db_hash_next;
302 	db->db_hash_next = NULL;
303 	mutex_exit(DBUF_HASH_MUTEX(h, idx));
304 	atomic_dec_64(&dbuf_hash_count);
305 }
306 
307 typedef enum {
308 	DBVU_EVICTING,
309 	DBVU_NOT_EVICTING
310 } dbvu_verify_type_t;
311 
312 static void
313 dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type)
314 {
315 #ifdef ZFS_DEBUG
316 	int64_t holds;
317 
318 	if (db->db_user == NULL)
319 		return;
320 
321 	/* Only data blocks support the attachment of user data. */
322 	ASSERT(db->db_level == 0);
323 
324 	/* Clients must resolve a dbuf before attaching user data. */
325 	ASSERT(db->db.db_data != NULL);
326 	ASSERT3U(db->db_state, ==, DB_CACHED);
327 
328 	holds = refcount_count(&db->db_holds);
329 	if (verify_type == DBVU_EVICTING) {
330 		/*
331 		 * Immediate eviction occurs when holds == dirtycnt.
332 		 * For normal eviction buffers, holds is zero on
333 		 * eviction, except when dbuf_fix_old_data() calls
334 		 * dbuf_clear_data().  However, the hold count can grow
335 		 * during eviction even though db_mtx is held (see
336 		 * dmu_bonus_hold() for an example), so we can only
337 		 * test the generic invariant that holds >= dirtycnt.
338 		 */
339 		ASSERT3U(holds, >=, db->db_dirtycnt);
340 	} else {
341 		if (db->db_user_immediate_evict == TRUE)
342 			ASSERT3U(holds, >=, db->db_dirtycnt);
343 		else
344 			ASSERT3U(holds, >, 0);
345 	}
346 #endif
347 }
348 
349 static void
350 dbuf_evict_user(dmu_buf_impl_t *db)
351 {
352 	dmu_buf_user_t *dbu = db->db_user;
353 
354 	ASSERT(MUTEX_HELD(&db->db_mtx));
355 
356 	if (dbu == NULL)
357 		return;
358 
359 	dbuf_verify_user(db, DBVU_EVICTING);
360 	db->db_user = NULL;
361 
362 #ifdef ZFS_DEBUG
363 	if (dbu->dbu_clear_on_evict_dbufp != NULL)
364 		*dbu->dbu_clear_on_evict_dbufp = NULL;
365 #endif
366 
367 	/*
368 	 * There are two eviction callbacks - one that we call synchronously
369 	 * and one that we invoke via a taskq.  The async one is useful for
370 	 * avoiding lock order reversals and limiting stack depth.
371 	 *
372 	 * Note that if we have a sync callback but no async callback,
373 	 * it's likely that the sync callback will free the structure
374 	 * containing the dbu.  In that case we need to take care to not
375 	 * dereference dbu after calling the sync evict func.
376 	 */
377 	boolean_t has_async = (dbu->dbu_evict_func_async != NULL);
378 
379 	if (dbu->dbu_evict_func_sync != NULL)
380 		dbu->dbu_evict_func_sync(dbu);
381 
382 	if (has_async) {
383 		taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func_async,
384 		    dbu, 0, &dbu->dbu_tqent);
385 	}
386 }
387 
388 boolean_t
389 dbuf_is_metadata(dmu_buf_impl_t *db)
390 {
391 	if (db->db_level > 0) {
392 		return (B_TRUE);
393 	} else {
394 		boolean_t is_metadata;
395 
396 		DB_DNODE_ENTER(db);
397 		is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type);
398 		DB_DNODE_EXIT(db);
399 
400 		return (is_metadata);
401 	}
402 }
403 
404 /*
405  * This function *must* return indices evenly distributed between all
406  * sublists of the multilist. This is needed due to how the dbuf eviction
407  * code is laid out; dbuf_evict_thread() assumes dbufs are evenly
408  * distributed between all sublists and uses this assumption when
409  * deciding which sublist to evict from and how much to evict from it.
410  */
411 unsigned int
412 dbuf_cache_multilist_index_func(multilist_t *ml, void *obj)
413 {
414 	dmu_buf_impl_t *db = obj;
415 
416 	/*
417 	 * The assumption here, is the hash value for a given
418 	 * dmu_buf_impl_t will remain constant throughout it's lifetime
419 	 * (i.e. it's objset, object, level and blkid fields don't change).
420 	 * Thus, we don't need to store the dbuf's sublist index
421 	 * on insertion, as this index can be recalculated on removal.
422 	 *
423 	 * Also, the low order bits of the hash value are thought to be
424 	 * distributed evenly. Otherwise, in the case that the multilist
425 	 * has a power of two number of sublists, each sublists' usage
426 	 * would not be evenly distributed.
427 	 */
428 	return (dbuf_hash(db->db_objset, db->db.db_object,
429 	    db->db_level, db->db_blkid) %
430 	    multilist_get_num_sublists(ml));
431 }
432 
433 static inline boolean_t
434 dbuf_cache_above_hiwater(void)
435 {
436 	uint64_t dbuf_cache_hiwater_bytes =
437 	    (dbuf_cache_max_bytes * dbuf_cache_hiwater_pct) / 100;
438 
439 	return (refcount_count(&dbuf_cache_size) >
440 	    dbuf_cache_max_bytes + dbuf_cache_hiwater_bytes);
441 }
442 
443 static inline boolean_t
444 dbuf_cache_above_lowater(void)
445 {
446 	uint64_t dbuf_cache_lowater_bytes =
447 	    (dbuf_cache_max_bytes * dbuf_cache_lowater_pct) / 100;
448 
449 	return (refcount_count(&dbuf_cache_size) >
450 	    dbuf_cache_max_bytes - dbuf_cache_lowater_bytes);
451 }
452 
453 /*
454  * Evict the oldest eligible dbuf from the dbuf cache.
455  */
456 static void
457 dbuf_evict_one(void)
458 {
459 	int idx = multilist_get_random_index(dbuf_cache);
460 	multilist_sublist_t *mls = multilist_sublist_lock(dbuf_cache, idx);
461 
462 	ASSERT(!MUTEX_HELD(&dbuf_evict_lock));
463 
464 	/*
465 	 * Set the thread's tsd to indicate that it's processing evictions.
466 	 * Once a thread stops evicting from the dbuf cache it will
467 	 * reset its tsd to NULL.
468 	 */
469 	ASSERT3P(tsd_get(zfs_dbuf_evict_key), ==, NULL);
470 	(void) tsd_set(zfs_dbuf_evict_key, (void *)B_TRUE);
471 
472 	dmu_buf_impl_t *db = multilist_sublist_tail(mls);
473 	while (db != NULL && mutex_tryenter(&db->db_mtx) == 0) {
474 		db = multilist_sublist_prev(mls, db);
475 	}
476 
477 	DTRACE_PROBE2(dbuf__evict__one, dmu_buf_impl_t *, db,
478 	    multilist_sublist_t *, mls);
479 
480 	if (db != NULL) {
481 		multilist_sublist_remove(mls, db);
482 		multilist_sublist_unlock(mls);
483 		(void) refcount_remove_many(&dbuf_cache_size,
484 		    db->db.db_size, db);
485 		dbuf_destroy(db);
486 	} else {
487 		multilist_sublist_unlock(mls);
488 	}
489 	(void) tsd_set(zfs_dbuf_evict_key, NULL);
490 }
491 
492 /*
493  * The dbuf evict thread is responsible for aging out dbufs from the
494  * cache. Once the cache has reached it's maximum size, dbufs are removed
495  * and destroyed. The eviction thread will continue running until the size
496  * of the dbuf cache is at or below the maximum size. Once the dbuf is aged
497  * out of the cache it is destroyed and becomes eligible for arc eviction.
498  */
499 /* ARGSUSED */
500 static void
501 dbuf_evict_thread(void *unused)
502 {
503 	callb_cpr_t cpr;
504 
505 	CALLB_CPR_INIT(&cpr, &dbuf_evict_lock, callb_generic_cpr, FTAG);
506 
507 	mutex_enter(&dbuf_evict_lock);
508 	while (!dbuf_evict_thread_exit) {
509 		while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
510 			CALLB_CPR_SAFE_BEGIN(&cpr);
511 			(void) cv_timedwait_hires(&dbuf_evict_cv,
512 			    &dbuf_evict_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
513 			CALLB_CPR_SAFE_END(&cpr, &dbuf_evict_lock);
514 		}
515 		mutex_exit(&dbuf_evict_lock);
516 
517 		/*
518 		 * Keep evicting as long as we're above the low water mark
519 		 * for the cache. We do this without holding the locks to
520 		 * minimize lock contention.
521 		 */
522 		while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
523 			dbuf_evict_one();
524 		}
525 
526 		mutex_enter(&dbuf_evict_lock);
527 	}
528 
529 	dbuf_evict_thread_exit = B_FALSE;
530 	cv_broadcast(&dbuf_evict_cv);
531 	CALLB_CPR_EXIT(&cpr);	/* drops dbuf_evict_lock */
532 	thread_exit();
533 }
534 
535 /*
536  * Wake up the dbuf eviction thread if the dbuf cache is at its max size.
537  * If the dbuf cache is at its high water mark, then evict a dbuf from the
538  * dbuf cache using the callers context.
539  */
540 static void
541 dbuf_evict_notify(void)
542 {
543 
544 	/*
545 	 * We use thread specific data to track when a thread has
546 	 * started processing evictions. This allows us to avoid deeply
547 	 * nested stacks that would have a call flow similar to this:
548 	 *
549 	 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify()
550 	 *	^						|
551 	 *	|						|
552 	 *	+-----dbuf_destroy()<--dbuf_evict_one()<--------+
553 	 *
554 	 * The dbuf_eviction_thread will always have its tsd set until
555 	 * that thread exits. All other threads will only set their tsd
556 	 * if they are participating in the eviction process. This only
557 	 * happens if the eviction thread is unable to process evictions
558 	 * fast enough. To keep the dbuf cache size in check, other threads
559 	 * can evict from the dbuf cache directly. Those threads will set
560 	 * their tsd values so that we ensure that they only evict one dbuf
561 	 * from the dbuf cache.
562 	 */
563 	if (tsd_get(zfs_dbuf_evict_key) != NULL)
564 		return;
565 
566 	/*
567 	 * We check if we should evict without holding the dbuf_evict_lock,
568 	 * because it's OK to occasionally make the wrong decision here,
569 	 * and grabbing the lock results in massive lock contention.
570 	 */
571 	if (refcount_count(&dbuf_cache_size) > dbuf_cache_max_bytes) {
572 		if (dbuf_cache_above_hiwater())
573 			dbuf_evict_one();
574 		cv_signal(&dbuf_evict_cv);
575 	}
576 }
577 
578 void
579 dbuf_init(void)
580 {
581 	uint64_t hsize = 1ULL << 16;
582 	dbuf_hash_table_t *h = &dbuf_hash_table;
583 	int i;
584 
585 	/*
586 	 * The hash table is big enough to fill all of physical memory
587 	 * with an average 4K block size.  The table will take up
588 	 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
589 	 */
590 	while (hsize * 4096 < physmem * PAGESIZE)
591 		hsize <<= 1;
592 
593 retry:
594 	h->hash_table_mask = hsize - 1;
595 	h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
596 	if (h->hash_table == NULL) {
597 		/* XXX - we should really return an error instead of assert */
598 		ASSERT(hsize > (1ULL << 10));
599 		hsize >>= 1;
600 		goto retry;
601 	}
602 
603 	dbuf_kmem_cache = kmem_cache_create("dmu_buf_impl_t",
604 	    sizeof (dmu_buf_impl_t),
605 	    0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
606 
607 	for (i = 0; i < DBUF_MUTEXES; i++)
608 		mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
609 
610 	/*
611 	 * Setup the parameters for the dbuf cache. We cap the size of the
612 	 * dbuf cache to 1/32nd (default) of the size of the ARC.
613 	 */
614 	dbuf_cache_max_bytes = MIN(dbuf_cache_max_bytes,
615 	    arc_max_bytes() >> dbuf_cache_max_shift);
616 
617 	/*
618 	 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
619 	 * configuration is not required.
620 	 */
621 	dbu_evict_taskq = taskq_create("dbu_evict", 1, minclsyspri, 0, 0, 0);
622 
623 	dbuf_cache = multilist_create(sizeof (dmu_buf_impl_t),
624 	    offsetof(dmu_buf_impl_t, db_cache_link),
625 	    dbuf_cache_multilist_index_func);
626 	refcount_create(&dbuf_cache_size);
627 
628 	tsd_create(&zfs_dbuf_evict_key, NULL);
629 	dbuf_evict_thread_exit = B_FALSE;
630 	mutex_init(&dbuf_evict_lock, NULL, MUTEX_DEFAULT, NULL);
631 	cv_init(&dbuf_evict_cv, NULL, CV_DEFAULT, NULL);
632 	dbuf_cache_evict_thread = thread_create(NULL, 0, dbuf_evict_thread,
633 	    NULL, 0, &p0, TS_RUN, minclsyspri);
634 }
635 
636 void
637 dbuf_fini(void)
638 {
639 	dbuf_hash_table_t *h = &dbuf_hash_table;
640 	int i;
641 
642 	for (i = 0; i < DBUF_MUTEXES; i++)
643 		mutex_destroy(&h->hash_mutexes[i]);
644 	kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
645 	kmem_cache_destroy(dbuf_kmem_cache);
646 	taskq_destroy(dbu_evict_taskq);
647 
648 	mutex_enter(&dbuf_evict_lock);
649 	dbuf_evict_thread_exit = B_TRUE;
650 	while (dbuf_evict_thread_exit) {
651 		cv_signal(&dbuf_evict_cv);
652 		cv_wait(&dbuf_evict_cv, &dbuf_evict_lock);
653 	}
654 	mutex_exit(&dbuf_evict_lock);
655 	tsd_destroy(&zfs_dbuf_evict_key);
656 
657 	mutex_destroy(&dbuf_evict_lock);
658 	cv_destroy(&dbuf_evict_cv);
659 
660 	refcount_destroy(&dbuf_cache_size);
661 	multilist_destroy(dbuf_cache);
662 }
663 
664 /*
665  * Other stuff.
666  */
667 
668 #ifdef ZFS_DEBUG
669 static void
670 dbuf_verify(dmu_buf_impl_t *db)
671 {
672 	dnode_t *dn;
673 	dbuf_dirty_record_t *dr;
674 
675 	ASSERT(MUTEX_HELD(&db->db_mtx));
676 
677 	if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
678 		return;
679 
680 	ASSERT(db->db_objset != NULL);
681 	DB_DNODE_ENTER(db);
682 	dn = DB_DNODE(db);
683 	if (dn == NULL) {
684 		ASSERT(db->db_parent == NULL);
685 		ASSERT(db->db_blkptr == NULL);
686 	} else {
687 		ASSERT3U(db->db.db_object, ==, dn->dn_object);
688 		ASSERT3P(db->db_objset, ==, dn->dn_objset);
689 		ASSERT3U(db->db_level, <, dn->dn_nlevels);
690 		ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
691 		    db->db_blkid == DMU_SPILL_BLKID ||
692 		    !avl_is_empty(&dn->dn_dbufs));
693 	}
694 	if (db->db_blkid == DMU_BONUS_BLKID) {
695 		ASSERT(dn != NULL);
696 		ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
697 		ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
698 	} else if (db->db_blkid == DMU_SPILL_BLKID) {
699 		ASSERT(dn != NULL);
700 		ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
701 		ASSERT0(db->db.db_offset);
702 	} else {
703 		ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
704 	}
705 
706 	for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
707 		ASSERT(dr->dr_dbuf == db);
708 
709 	for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
710 		ASSERT(dr->dr_dbuf == db);
711 
712 	/*
713 	 * We can't assert that db_size matches dn_datablksz because it
714 	 * can be momentarily different when another thread is doing
715 	 * dnode_set_blksz().
716 	 */
717 	if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
718 		dr = db->db_data_pending;
719 		/*
720 		 * It should only be modified in syncing context, so
721 		 * make sure we only have one copy of the data.
722 		 */
723 		ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
724 	}
725 
726 	/* verify db->db_blkptr */
727 	if (db->db_blkptr) {
728 		if (db->db_parent == dn->dn_dbuf) {
729 			/* db is pointed to by the dnode */
730 			/* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
731 			if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
732 				ASSERT(db->db_parent == NULL);
733 			else
734 				ASSERT(db->db_parent != NULL);
735 			if (db->db_blkid != DMU_SPILL_BLKID)
736 				ASSERT3P(db->db_blkptr, ==,
737 				    &dn->dn_phys->dn_blkptr[db->db_blkid]);
738 		} else {
739 			/* db is pointed to by an indirect block */
740 			int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
741 			ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
742 			ASSERT3U(db->db_parent->db.db_object, ==,
743 			    db->db.db_object);
744 			/*
745 			 * dnode_grow_indblksz() can make this fail if we don't
746 			 * have the struct_rwlock.  XXX indblksz no longer
747 			 * grows.  safe to do this now?
748 			 */
749 			if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
750 				ASSERT3P(db->db_blkptr, ==,
751 				    ((blkptr_t *)db->db_parent->db.db_data +
752 				    db->db_blkid % epb));
753 			}
754 		}
755 	}
756 	if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
757 	    (db->db_buf == NULL || db->db_buf->b_data) &&
758 	    db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
759 	    db->db_state != DB_FILL && !dn->dn_free_txg) {
760 		/*
761 		 * If the blkptr isn't set but they have nonzero data,
762 		 * it had better be dirty, otherwise we'll lose that
763 		 * data when we evict this buffer.
764 		 *
765 		 * There is an exception to this rule for indirect blocks; in
766 		 * this case, if the indirect block is a hole, we fill in a few
767 		 * fields on each of the child blocks (importantly, birth time)
768 		 * to prevent hole birth times from being lost when you
769 		 * partially fill in a hole.
770 		 */
771 		if (db->db_dirtycnt == 0) {
772 			if (db->db_level == 0) {
773 				uint64_t *buf = db->db.db_data;
774 				int i;
775 
776 				for (i = 0; i < db->db.db_size >> 3; i++) {
777 					ASSERT(buf[i] == 0);
778 				}
779 			} else {
780 				blkptr_t *bps = db->db.db_data;
781 				ASSERT3U(1 << DB_DNODE(db)->dn_indblkshift, ==,
782 				    db->db.db_size);
783 				/*
784 				 * We want to verify that all the blkptrs in the
785 				 * indirect block are holes, but we may have
786 				 * automatically set up a few fields for them.
787 				 * We iterate through each blkptr and verify
788 				 * they only have those fields set.
789 				 */
790 				for (int i = 0;
791 				    i < db->db.db_size / sizeof (blkptr_t);
792 				    i++) {
793 					blkptr_t *bp = &bps[i];
794 					ASSERT(ZIO_CHECKSUM_IS_ZERO(
795 					    &bp->blk_cksum));
796 					ASSERT(
797 					    DVA_IS_EMPTY(&bp->blk_dva[0]) &&
798 					    DVA_IS_EMPTY(&bp->blk_dva[1]) &&
799 					    DVA_IS_EMPTY(&bp->blk_dva[2]));
800 					ASSERT0(bp->blk_fill);
801 					ASSERT0(bp->blk_pad[0]);
802 					ASSERT0(bp->blk_pad[1]);
803 					ASSERT(!BP_IS_EMBEDDED(bp));
804 					ASSERT(BP_IS_HOLE(bp));
805 					ASSERT0(bp->blk_phys_birth);
806 				}
807 			}
808 		}
809 	}
810 	DB_DNODE_EXIT(db);
811 }
812 #endif
813 
814 static void
815 dbuf_clear_data(dmu_buf_impl_t *db)
816 {
817 	ASSERT(MUTEX_HELD(&db->db_mtx));
818 	dbuf_evict_user(db);
819 	ASSERT3P(db->db_buf, ==, NULL);
820 	db->db.db_data = NULL;
821 	if (db->db_state != DB_NOFILL)
822 		db->db_state = DB_UNCACHED;
823 }
824 
825 static void
826 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
827 {
828 	ASSERT(MUTEX_HELD(&db->db_mtx));
829 	ASSERT(buf != NULL);
830 
831 	db->db_buf = buf;
832 	ASSERT(buf->b_data != NULL);
833 	db->db.db_data = buf->b_data;
834 }
835 
836 /*
837  * Loan out an arc_buf for read.  Return the loaned arc_buf.
838  */
839 arc_buf_t *
840 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
841 {
842 	arc_buf_t *abuf;
843 
844 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
845 	mutex_enter(&db->db_mtx);
846 	if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
847 		int blksz = db->db.db_size;
848 		spa_t *spa = db->db_objset->os_spa;
849 
850 		mutex_exit(&db->db_mtx);
851 		abuf = arc_loan_buf(spa, B_FALSE, blksz);
852 		bcopy(db->db.db_data, abuf->b_data, blksz);
853 	} else {
854 		abuf = db->db_buf;
855 		arc_loan_inuse_buf(abuf, db);
856 		db->db_buf = NULL;
857 		dbuf_clear_data(db);
858 		mutex_exit(&db->db_mtx);
859 	}
860 	return (abuf);
861 }
862 
863 /*
864  * Calculate which level n block references the data at the level 0 offset
865  * provided.
866  */
867 uint64_t
868 dbuf_whichblock(dnode_t *dn, int64_t level, uint64_t offset)
869 {
870 	if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) {
871 		/*
872 		 * The level n blkid is equal to the level 0 blkid divided by
873 		 * the number of level 0s in a level n block.
874 		 *
875 		 * The level 0 blkid is offset >> datablkshift =
876 		 * offset / 2^datablkshift.
877 		 *
878 		 * The number of level 0s in a level n is the number of block
879 		 * pointers in an indirect block, raised to the power of level.
880 		 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
881 		 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
882 		 *
883 		 * Thus, the level n blkid is: offset /
884 		 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
885 		 * = offset / 2^(datablkshift + level *
886 		 *   (indblkshift - SPA_BLKPTRSHIFT))
887 		 * = offset >> (datablkshift + level *
888 		 *   (indblkshift - SPA_BLKPTRSHIFT))
889 		 */
890 		return (offset >> (dn->dn_datablkshift + level *
891 		    (dn->dn_indblkshift - SPA_BLKPTRSHIFT)));
892 	} else {
893 		ASSERT3U(offset, <, dn->dn_datablksz);
894 		return (0);
895 	}
896 }
897 
898 static void
899 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
900 {
901 	dmu_buf_impl_t *db = vdb;
902 
903 	mutex_enter(&db->db_mtx);
904 	ASSERT3U(db->db_state, ==, DB_READ);
905 	/*
906 	 * All reads are synchronous, so we must have a hold on the dbuf
907 	 */
908 	ASSERT(refcount_count(&db->db_holds) > 0);
909 	ASSERT(db->db_buf == NULL);
910 	ASSERT(db->db.db_data == NULL);
911 	if (db->db_level == 0 && db->db_freed_in_flight) {
912 		/* we were freed in flight; disregard any error */
913 		arc_release(buf, db);
914 		bzero(buf->b_data, db->db.db_size);
915 		arc_buf_freeze(buf);
916 		db->db_freed_in_flight = FALSE;
917 		dbuf_set_data(db, buf);
918 		db->db_state = DB_CACHED;
919 	} else if (zio == NULL || zio->io_error == 0) {
920 		dbuf_set_data(db, buf);
921 		db->db_state = DB_CACHED;
922 	} else {
923 		ASSERT(db->db_blkid != DMU_BONUS_BLKID);
924 		ASSERT3P(db->db_buf, ==, NULL);
925 		arc_buf_destroy(buf, db);
926 		db->db_state = DB_UNCACHED;
927 	}
928 	cv_broadcast(&db->db_changed);
929 	dbuf_rele_and_unlock(db, NULL);
930 }
931 
932 static void
933 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
934 {
935 	dnode_t *dn;
936 	zbookmark_phys_t zb;
937 	arc_flags_t aflags = ARC_FLAG_NOWAIT;
938 
939 	DB_DNODE_ENTER(db);
940 	dn = DB_DNODE(db);
941 	ASSERT(!refcount_is_zero(&db->db_holds));
942 	/* We need the struct_rwlock to prevent db_blkptr from changing. */
943 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
944 	ASSERT(MUTEX_HELD(&db->db_mtx));
945 	ASSERT(db->db_state == DB_UNCACHED);
946 	ASSERT(db->db_buf == NULL);
947 
948 	if (db->db_blkid == DMU_BONUS_BLKID) {
949 		int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
950 
951 		ASSERT3U(bonuslen, <=, db->db.db_size);
952 		db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN);
953 		arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
954 		if (bonuslen < DN_MAX_BONUSLEN)
955 			bzero(db->db.db_data, DN_MAX_BONUSLEN);
956 		if (bonuslen)
957 			bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
958 		DB_DNODE_EXIT(db);
959 		db->db_state = DB_CACHED;
960 		mutex_exit(&db->db_mtx);
961 		return;
962 	}
963 
964 	/*
965 	 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
966 	 * processes the delete record and clears the bp while we are waiting
967 	 * for the dn_mtx (resulting in a "no" from block_freed).
968 	 */
969 	if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
970 	    (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
971 	    BP_IS_HOLE(db->db_blkptr)))) {
972 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
973 
974 		dbuf_set_data(db, arc_alloc_buf(db->db_objset->os_spa, db, type,
975 		    db->db.db_size));
976 		bzero(db->db.db_data, db->db.db_size);
977 
978 		if (db->db_blkptr != NULL && db->db_level > 0 &&
979 		    BP_IS_HOLE(db->db_blkptr) &&
980 		    db->db_blkptr->blk_birth != 0) {
981 			blkptr_t *bps = db->db.db_data;
982 			for (int i = 0; i < ((1 <<
983 			    DB_DNODE(db)->dn_indblkshift) / sizeof (blkptr_t));
984 			    i++) {
985 				blkptr_t *bp = &bps[i];
986 				ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
987 				    1 << dn->dn_indblkshift);
988 				BP_SET_LSIZE(bp,
989 				    BP_GET_LEVEL(db->db_blkptr) == 1 ?
990 				    dn->dn_datablksz :
991 				    BP_GET_LSIZE(db->db_blkptr));
992 				BP_SET_TYPE(bp, BP_GET_TYPE(db->db_blkptr));
993 				BP_SET_LEVEL(bp,
994 				    BP_GET_LEVEL(db->db_blkptr) - 1);
995 				BP_SET_BIRTH(bp, db->db_blkptr->blk_birth, 0);
996 			}
997 		}
998 		DB_DNODE_EXIT(db);
999 		db->db_state = DB_CACHED;
1000 		mutex_exit(&db->db_mtx);
1001 		return;
1002 	}
1003 
1004 	DB_DNODE_EXIT(db);
1005 
1006 	db->db_state = DB_READ;
1007 	mutex_exit(&db->db_mtx);
1008 
1009 	if (DBUF_IS_L2CACHEABLE(db))
1010 		aflags |= ARC_FLAG_L2CACHE;
1011 
1012 	SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
1013 	    db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
1014 	    db->db.db_object, db->db_level, db->db_blkid);
1015 
1016 	dbuf_add_ref(db, NULL);
1017 
1018 	(void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr,
1019 	    dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
1020 	    (flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
1021 	    &aflags, &zb);
1022 }
1023 
1024 /*
1025  * This is our just-in-time copy function.  It makes a copy of buffers that
1026  * have been modified in a previous transaction group before we access them in
1027  * the current active group.
1028  *
1029  * This function is used in three places: when we are dirtying a buffer for the
1030  * first time in a txg, when we are freeing a range in a dnode that includes
1031  * this buffer, and when we are accessing a buffer which was received compressed
1032  * and later referenced in a WRITE_BYREF record.
1033  *
1034  * Note that when we are called from dbuf_free_range() we do not put a hold on
1035  * the buffer, we just traverse the active dbuf list for the dnode.
1036  */
1037 static void
1038 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
1039 {
1040 	dbuf_dirty_record_t *dr = db->db_last_dirty;
1041 
1042 	ASSERT(MUTEX_HELD(&db->db_mtx));
1043 	ASSERT(db->db.db_data != NULL);
1044 	ASSERT(db->db_level == 0);
1045 	ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
1046 
1047 	if (dr == NULL ||
1048 	    (dr->dt.dl.dr_data !=
1049 	    ((db->db_blkid  == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
1050 		return;
1051 
1052 	/*
1053 	 * If the last dirty record for this dbuf has not yet synced
1054 	 * and its referencing the dbuf data, either:
1055 	 *	reset the reference to point to a new copy,
1056 	 * or (if there a no active holders)
1057 	 *	just null out the current db_data pointer.
1058 	 */
1059 	ASSERT(dr->dr_txg >= txg - 2);
1060 	if (db->db_blkid == DMU_BONUS_BLKID) {
1061 		/* Note that the data bufs here are zio_bufs */
1062 		dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
1063 		arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
1064 		bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
1065 	} else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
1066 		int size = arc_buf_size(db->db_buf);
1067 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1068 		spa_t *spa = db->db_objset->os_spa;
1069 		enum zio_compress compress_type =
1070 		    arc_get_compression(db->db_buf);
1071 
1072 		if (compress_type == ZIO_COMPRESS_OFF) {
1073 			dr->dt.dl.dr_data = arc_alloc_buf(spa, db, type, size);
1074 		} else {
1075 			ASSERT3U(type, ==, ARC_BUFC_DATA);
1076 			dr->dt.dl.dr_data = arc_alloc_compressed_buf(spa, db,
1077 			    size, arc_buf_lsize(db->db_buf), compress_type);
1078 		}
1079 		bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
1080 	} else {
1081 		db->db_buf = NULL;
1082 		dbuf_clear_data(db);
1083 	}
1084 }
1085 
1086 int
1087 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
1088 {
1089 	int err = 0;
1090 	boolean_t prefetch;
1091 	dnode_t *dn;
1092 
1093 	/*
1094 	 * We don't have to hold the mutex to check db_state because it
1095 	 * can't be freed while we have a hold on the buffer.
1096 	 */
1097 	ASSERT(!refcount_is_zero(&db->db_holds));
1098 
1099 	if (db->db_state == DB_NOFILL)
1100 		return (SET_ERROR(EIO));
1101 
1102 	DB_DNODE_ENTER(db);
1103 	dn = DB_DNODE(db);
1104 	if ((flags & DB_RF_HAVESTRUCT) == 0)
1105 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
1106 
1107 	prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1108 	    (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
1109 	    DBUF_IS_CACHEABLE(db);
1110 
1111 	mutex_enter(&db->db_mtx);
1112 	if (db->db_state == DB_CACHED) {
1113 		/*
1114 		 * If the arc buf is compressed, we need to decompress it to
1115 		 * read the data. This could happen during the "zfs receive" of
1116 		 * a stream which is compressed and deduplicated.
1117 		 */
1118 		if (db->db_buf != NULL &&
1119 		    arc_get_compression(db->db_buf) != ZIO_COMPRESS_OFF) {
1120 			dbuf_fix_old_data(db,
1121 			    spa_syncing_txg(dmu_objset_spa(db->db_objset)));
1122 			err = arc_decompress(db->db_buf);
1123 			dbuf_set_data(db, db->db_buf);
1124 		}
1125 		mutex_exit(&db->db_mtx);
1126 		if (prefetch)
1127 			dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1128 		if ((flags & DB_RF_HAVESTRUCT) == 0)
1129 			rw_exit(&dn->dn_struct_rwlock);
1130 		DB_DNODE_EXIT(db);
1131 	} else if (db->db_state == DB_UNCACHED) {
1132 		spa_t *spa = dn->dn_objset->os_spa;
1133 		boolean_t need_wait = B_FALSE;
1134 
1135 		if (zio == NULL &&
1136 		    db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) {
1137 			zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
1138 			need_wait = B_TRUE;
1139 		}
1140 		dbuf_read_impl(db, zio, flags);
1141 
1142 		/* dbuf_read_impl has dropped db_mtx for us */
1143 
1144 		if (prefetch)
1145 			dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1146 
1147 		if ((flags & DB_RF_HAVESTRUCT) == 0)
1148 			rw_exit(&dn->dn_struct_rwlock);
1149 		DB_DNODE_EXIT(db);
1150 
1151 		if (need_wait)
1152 			err = zio_wait(zio);
1153 	} else {
1154 		/*
1155 		 * Another reader came in while the dbuf was in flight
1156 		 * between UNCACHED and CACHED.  Either a writer will finish
1157 		 * writing the buffer (sending the dbuf to CACHED) or the
1158 		 * first reader's request will reach the read_done callback
1159 		 * and send the dbuf to CACHED.  Otherwise, a failure
1160 		 * occurred and the dbuf went to UNCACHED.
1161 		 */
1162 		mutex_exit(&db->db_mtx);
1163 		if (prefetch)
1164 			dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1165 		if ((flags & DB_RF_HAVESTRUCT) == 0)
1166 			rw_exit(&dn->dn_struct_rwlock);
1167 		DB_DNODE_EXIT(db);
1168 
1169 		/* Skip the wait per the caller's request. */
1170 		mutex_enter(&db->db_mtx);
1171 		if ((flags & DB_RF_NEVERWAIT) == 0) {
1172 			while (db->db_state == DB_READ ||
1173 			    db->db_state == DB_FILL) {
1174 				ASSERT(db->db_state == DB_READ ||
1175 				    (flags & DB_RF_HAVESTRUCT) == 0);
1176 				DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *,
1177 				    db, zio_t *, zio);
1178 				cv_wait(&db->db_changed, &db->db_mtx);
1179 			}
1180 			if (db->db_state == DB_UNCACHED)
1181 				err = SET_ERROR(EIO);
1182 		}
1183 		mutex_exit(&db->db_mtx);
1184 	}
1185 
1186 	return (err);
1187 }
1188 
1189 static void
1190 dbuf_noread(dmu_buf_impl_t *db)
1191 {
1192 	ASSERT(!refcount_is_zero(&db->db_holds));
1193 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1194 	mutex_enter(&db->db_mtx);
1195 	while (db->db_state == DB_READ || db->db_state == DB_FILL)
1196 		cv_wait(&db->db_changed, &db->db_mtx);
1197 	if (db->db_state == DB_UNCACHED) {
1198 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1199 		spa_t *spa = db->db_objset->os_spa;
1200 
1201 		ASSERT(db->db_buf == NULL);
1202 		ASSERT(db->db.db_data == NULL);
1203 		dbuf_set_data(db, arc_alloc_buf(spa, db, type, db->db.db_size));
1204 		db->db_state = DB_FILL;
1205 	} else if (db->db_state == DB_NOFILL) {
1206 		dbuf_clear_data(db);
1207 	} else {
1208 		ASSERT3U(db->db_state, ==, DB_CACHED);
1209 	}
1210 	mutex_exit(&db->db_mtx);
1211 }
1212 
1213 void
1214 dbuf_unoverride(dbuf_dirty_record_t *dr)
1215 {
1216 	dmu_buf_impl_t *db = dr->dr_dbuf;
1217 	blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
1218 	uint64_t txg = dr->dr_txg;
1219 
1220 	ASSERT(MUTEX_HELD(&db->db_mtx));
1221 	/*
1222 	 * This assert is valid because dmu_sync() expects to be called by
1223 	 * a zilog's get_data while holding a range lock.  This call only
1224 	 * comes from dbuf_dirty() callers who must also hold a range lock.
1225 	 */
1226 	ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
1227 	ASSERT(db->db_level == 0);
1228 
1229 	if (db->db_blkid == DMU_BONUS_BLKID ||
1230 	    dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
1231 		return;
1232 
1233 	ASSERT(db->db_data_pending != dr);
1234 
1235 	/* free this block */
1236 	if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite)
1237 		zio_free(db->db_objset->os_spa, txg, bp);
1238 
1239 	dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1240 	dr->dt.dl.dr_nopwrite = B_FALSE;
1241 
1242 	/*
1243 	 * Release the already-written buffer, so we leave it in
1244 	 * a consistent dirty state.  Note that all callers are
1245 	 * modifying the buffer, so they will immediately do
1246 	 * another (redundant) arc_release().  Therefore, leave
1247 	 * the buf thawed to save the effort of freezing &
1248 	 * immediately re-thawing it.
1249 	 */
1250 	arc_release(dr->dt.dl.dr_data, db);
1251 }
1252 
1253 /*
1254  * Evict (if its unreferenced) or clear (if its referenced) any level-0
1255  * data blocks in the free range, so that any future readers will find
1256  * empty blocks.
1257  */
1258 void
1259 dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
1260     dmu_tx_t *tx)
1261 {
1262 	dmu_buf_impl_t db_search;
1263 	dmu_buf_impl_t *db, *db_next;
1264 	uint64_t txg = tx->tx_txg;
1265 	avl_index_t where;
1266 
1267 	if (end_blkid > dn->dn_maxblkid &&
1268 	    !(start_blkid == DMU_SPILL_BLKID || end_blkid == DMU_SPILL_BLKID))
1269 		end_blkid = dn->dn_maxblkid;
1270 	dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid);
1271 
1272 	db_search.db_level = 0;
1273 	db_search.db_blkid = start_blkid;
1274 	db_search.db_state = DB_SEARCH;
1275 
1276 	mutex_enter(&dn->dn_dbufs_mtx);
1277 	db = avl_find(&dn->dn_dbufs, &db_search, &where);
1278 	ASSERT3P(db, ==, NULL);
1279 
1280 	db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1281 
1282 	for (; db != NULL; db = db_next) {
1283 		db_next = AVL_NEXT(&dn->dn_dbufs, db);
1284 		ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1285 
1286 		if (db->db_level != 0 || db->db_blkid > end_blkid) {
1287 			break;
1288 		}
1289 		ASSERT3U(db->db_blkid, >=, start_blkid);
1290 
1291 		/* found a level 0 buffer in the range */
1292 		mutex_enter(&db->db_mtx);
1293 		if (dbuf_undirty(db, tx)) {
1294 			/* mutex has been dropped and dbuf destroyed */
1295 			continue;
1296 		}
1297 
1298 		if (db->db_state == DB_UNCACHED ||
1299 		    db->db_state == DB_NOFILL ||
1300 		    db->db_state == DB_EVICTING) {
1301 			ASSERT(db->db.db_data == NULL);
1302 			mutex_exit(&db->db_mtx);
1303 			continue;
1304 		}
1305 		if (db->db_state == DB_READ || db->db_state == DB_FILL) {
1306 			/* will be handled in dbuf_read_done or dbuf_rele */
1307 			db->db_freed_in_flight = TRUE;
1308 			mutex_exit(&db->db_mtx);
1309 			continue;
1310 		}
1311 		if (refcount_count(&db->db_holds) == 0) {
1312 			ASSERT(db->db_buf);
1313 			dbuf_destroy(db);
1314 			continue;
1315 		}
1316 		/* The dbuf is referenced */
1317 
1318 		if (db->db_last_dirty != NULL) {
1319 			dbuf_dirty_record_t *dr = db->db_last_dirty;
1320 
1321 			if (dr->dr_txg == txg) {
1322 				/*
1323 				 * This buffer is "in-use", re-adjust the file
1324 				 * size to reflect that this buffer may
1325 				 * contain new data when we sync.
1326 				 */
1327 				if (db->db_blkid != DMU_SPILL_BLKID &&
1328 				    db->db_blkid > dn->dn_maxblkid)
1329 					dn->dn_maxblkid = db->db_blkid;
1330 				dbuf_unoverride(dr);
1331 			} else {
1332 				/*
1333 				 * This dbuf is not dirty in the open context.
1334 				 * Either uncache it (if its not referenced in
1335 				 * the open context) or reset its contents to
1336 				 * empty.
1337 				 */
1338 				dbuf_fix_old_data(db, txg);
1339 			}
1340 		}
1341 		/* clear the contents if its cached */
1342 		if (db->db_state == DB_CACHED) {
1343 			ASSERT(db->db.db_data != NULL);
1344 			arc_release(db->db_buf, db);
1345 			bzero(db->db.db_data, db->db.db_size);
1346 			arc_buf_freeze(db->db_buf);
1347 		}
1348 
1349 		mutex_exit(&db->db_mtx);
1350 	}
1351 	mutex_exit(&dn->dn_dbufs_mtx);
1352 }
1353 
1354 void
1355 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
1356 {
1357 	arc_buf_t *buf, *obuf;
1358 	int osize = db->db.db_size;
1359 	arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1360 	dnode_t *dn;
1361 
1362 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1363 
1364 	DB_DNODE_ENTER(db);
1365 	dn = DB_DNODE(db);
1366 
1367 	/* XXX does *this* func really need the lock? */
1368 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1369 
1370 	/*
1371 	 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1372 	 * is OK, because there can be no other references to the db
1373 	 * when we are changing its size, so no concurrent DB_FILL can
1374 	 * be happening.
1375 	 */
1376 	/*
1377 	 * XXX we should be doing a dbuf_read, checking the return
1378 	 * value and returning that up to our callers
1379 	 */
1380 	dmu_buf_will_dirty(&db->db, tx);
1381 
1382 	/* create the data buffer for the new block */
1383 	buf = arc_alloc_buf(dn->dn_objset->os_spa, db, type, size);
1384 
1385 	/* copy old block data to the new block */
1386 	obuf = db->db_buf;
1387 	bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
1388 	/* zero the remainder */
1389 	if (size > osize)
1390 		bzero((uint8_t *)buf->b_data + osize, size - osize);
1391 
1392 	mutex_enter(&db->db_mtx);
1393 	dbuf_set_data(db, buf);
1394 	arc_buf_destroy(obuf, db);
1395 	db->db.db_size = size;
1396 
1397 	if (db->db_level == 0) {
1398 		ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1399 		db->db_last_dirty->dt.dl.dr_data = buf;
1400 	}
1401 	mutex_exit(&db->db_mtx);
1402 
1403 	dmu_objset_willuse_space(dn->dn_objset, size - osize, tx);
1404 	DB_DNODE_EXIT(db);
1405 }
1406 
1407 void
1408 dbuf_release_bp(dmu_buf_impl_t *db)
1409 {
1410 	objset_t *os = db->db_objset;
1411 
1412 	ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1413 	ASSERT(arc_released(os->os_phys_buf) ||
1414 	    list_link_active(&os->os_dsl_dataset->ds_synced_link));
1415 	ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1416 
1417 	(void) arc_release(db->db_buf, db);
1418 }
1419 
1420 /*
1421  * We already have a dirty record for this TXG, and we are being
1422  * dirtied again.
1423  */
1424 static void
1425 dbuf_redirty(dbuf_dirty_record_t *dr)
1426 {
1427 	dmu_buf_impl_t *db = dr->dr_dbuf;
1428 
1429 	ASSERT(MUTEX_HELD(&db->db_mtx));
1430 
1431 	if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1432 		/*
1433 		 * If this buffer has already been written out,
1434 		 * we now need to reset its state.
1435 		 */
1436 		dbuf_unoverride(dr);
1437 		if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1438 		    db->db_state != DB_NOFILL) {
1439 			/* Already released on initial dirty, so just thaw. */
1440 			ASSERT(arc_released(db->db_buf));
1441 			arc_buf_thaw(db->db_buf);
1442 		}
1443 	}
1444 }
1445 
1446 dbuf_dirty_record_t *
1447 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1448 {
1449 	dnode_t *dn;
1450 	objset_t *os;
1451 	dbuf_dirty_record_t **drp, *dr;
1452 	int drop_struct_lock = FALSE;
1453 	int txgoff = tx->tx_txg & TXG_MASK;
1454 
1455 	ASSERT(tx->tx_txg != 0);
1456 	ASSERT(!refcount_is_zero(&db->db_holds));
1457 	DMU_TX_DIRTY_BUF(tx, db);
1458 
1459 	DB_DNODE_ENTER(db);
1460 	dn = DB_DNODE(db);
1461 	/*
1462 	 * Shouldn't dirty a regular buffer in syncing context.  Private
1463 	 * objects may be dirtied in syncing context, but only if they
1464 	 * were already pre-dirtied in open context.
1465 	 */
1466 #ifdef DEBUG
1467 	if (dn->dn_objset->os_dsl_dataset != NULL) {
1468 		rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1469 		    RW_READER, FTAG);
1470 	}
1471 	ASSERT(!dmu_tx_is_syncing(tx) ||
1472 	    BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1473 	    DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1474 	    dn->dn_objset->os_dsl_dataset == NULL);
1475 	if (dn->dn_objset->os_dsl_dataset != NULL)
1476 		rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, FTAG);
1477 #endif
1478 	/*
1479 	 * We make this assert for private objects as well, but after we
1480 	 * check if we're already dirty.  They are allowed to re-dirty
1481 	 * in syncing context.
1482 	 */
1483 	ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1484 	    dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1485 	    (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1486 
1487 	mutex_enter(&db->db_mtx);
1488 	/*
1489 	 * XXX make this true for indirects too?  The problem is that
1490 	 * transactions created with dmu_tx_create_assigned() from
1491 	 * syncing context don't bother holding ahead.
1492 	 */
1493 	ASSERT(db->db_level != 0 ||
1494 	    db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1495 	    db->db_state == DB_NOFILL);
1496 
1497 	mutex_enter(&dn->dn_mtx);
1498 	/*
1499 	 * Don't set dirtyctx to SYNC if we're just modifying this as we
1500 	 * initialize the objset.
1501 	 */
1502 	if (dn->dn_dirtyctx == DN_UNDIRTIED) {
1503 		if (dn->dn_objset->os_dsl_dataset != NULL) {
1504 			rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1505 			    RW_READER, FTAG);
1506 		}
1507 		if (!BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1508 			dn->dn_dirtyctx = (dmu_tx_is_syncing(tx) ?
1509 			    DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1510 			ASSERT(dn->dn_dirtyctx_firstset == NULL);
1511 			dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1512 		}
1513 		if (dn->dn_objset->os_dsl_dataset != NULL) {
1514 			rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1515 			    FTAG);
1516 		}
1517 	}
1518 	mutex_exit(&dn->dn_mtx);
1519 
1520 	if (db->db_blkid == DMU_SPILL_BLKID)
1521 		dn->dn_have_spill = B_TRUE;
1522 
1523 	/*
1524 	 * If this buffer is already dirty, we're done.
1525 	 */
1526 	drp = &db->db_last_dirty;
1527 	ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1528 	    db->db.db_object == DMU_META_DNODE_OBJECT);
1529 	while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1530 		drp = &dr->dr_next;
1531 	if (dr && dr->dr_txg == tx->tx_txg) {
1532 		DB_DNODE_EXIT(db);
1533 
1534 		dbuf_redirty(dr);
1535 		mutex_exit(&db->db_mtx);
1536 		return (dr);
1537 	}
1538 
1539 	/*
1540 	 * Only valid if not already dirty.
1541 	 */
1542 	ASSERT(dn->dn_object == 0 ||
1543 	    dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1544 	    (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1545 
1546 	ASSERT3U(dn->dn_nlevels, >, db->db_level);
1547 
1548 	/*
1549 	 * We should only be dirtying in syncing context if it's the
1550 	 * mos or we're initializing the os or it's a special object.
1551 	 * However, we are allowed to dirty in syncing context provided
1552 	 * we already dirtied it in open context.  Hence we must make
1553 	 * this assertion only if we're not already dirty.
1554 	 */
1555 	os = dn->dn_objset;
1556 	VERIFY3U(tx->tx_txg, <=, spa_final_dirty_txg(os->os_spa));
1557 #ifdef DEBUG
1558 	if (dn->dn_objset->os_dsl_dataset != NULL)
1559 		rrw_enter(&os->os_dsl_dataset->ds_bp_rwlock, RW_READER, FTAG);
1560 	ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1561 	    os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1562 	if (dn->dn_objset->os_dsl_dataset != NULL)
1563 		rrw_exit(&os->os_dsl_dataset->ds_bp_rwlock, FTAG);
1564 #endif
1565 	ASSERT(db->db.db_size != 0);
1566 
1567 	dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1568 
1569 	if (db->db_blkid != DMU_BONUS_BLKID) {
1570 		dmu_objset_willuse_space(os, db->db.db_size, tx);
1571 	}
1572 
1573 	/*
1574 	 * If this buffer is dirty in an old transaction group we need
1575 	 * to make a copy of it so that the changes we make in this
1576 	 * transaction group won't leak out when we sync the older txg.
1577 	 */
1578 	dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1579 	if (db->db_level == 0) {
1580 		void *data_old = db->db_buf;
1581 
1582 		if (db->db_state != DB_NOFILL) {
1583 			if (db->db_blkid == DMU_BONUS_BLKID) {
1584 				dbuf_fix_old_data(db, tx->tx_txg);
1585 				data_old = db->db.db_data;
1586 			} else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1587 				/*
1588 				 * Release the data buffer from the cache so
1589 				 * that we can modify it without impacting
1590 				 * possible other users of this cached data
1591 				 * block.  Note that indirect blocks and
1592 				 * private objects are not released until the
1593 				 * syncing state (since they are only modified
1594 				 * then).
1595 				 */
1596 				arc_release(db->db_buf, db);
1597 				dbuf_fix_old_data(db, tx->tx_txg);
1598 				data_old = db->db_buf;
1599 			}
1600 			ASSERT(data_old != NULL);
1601 		}
1602 		dr->dt.dl.dr_data = data_old;
1603 	} else {
1604 		mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1605 		list_create(&dr->dt.di.dr_children,
1606 		    sizeof (dbuf_dirty_record_t),
1607 		    offsetof(dbuf_dirty_record_t, dr_dirty_node));
1608 	}
1609 	if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL)
1610 		dr->dr_accounted = db->db.db_size;
1611 	dr->dr_dbuf = db;
1612 	dr->dr_txg = tx->tx_txg;
1613 	dr->dr_next = *drp;
1614 	*drp = dr;
1615 
1616 	/*
1617 	 * We could have been freed_in_flight between the dbuf_noread
1618 	 * and dbuf_dirty.  We win, as though the dbuf_noread() had
1619 	 * happened after the free.
1620 	 */
1621 	if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1622 	    db->db_blkid != DMU_SPILL_BLKID) {
1623 		mutex_enter(&dn->dn_mtx);
1624 		if (dn->dn_free_ranges[txgoff] != NULL) {
1625 			range_tree_clear(dn->dn_free_ranges[txgoff],
1626 			    db->db_blkid, 1);
1627 		}
1628 		mutex_exit(&dn->dn_mtx);
1629 		db->db_freed_in_flight = FALSE;
1630 	}
1631 
1632 	/*
1633 	 * This buffer is now part of this txg
1634 	 */
1635 	dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1636 	db->db_dirtycnt += 1;
1637 	ASSERT3U(db->db_dirtycnt, <=, 3);
1638 
1639 	mutex_exit(&db->db_mtx);
1640 
1641 	if (db->db_blkid == DMU_BONUS_BLKID ||
1642 	    db->db_blkid == DMU_SPILL_BLKID) {
1643 		mutex_enter(&dn->dn_mtx);
1644 		ASSERT(!list_link_active(&dr->dr_dirty_node));
1645 		list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1646 		mutex_exit(&dn->dn_mtx);
1647 		dnode_setdirty(dn, tx);
1648 		DB_DNODE_EXIT(db);
1649 		return (dr);
1650 	}
1651 
1652 	/*
1653 	 * The dn_struct_rwlock prevents db_blkptr from changing
1654 	 * due to a write from syncing context completing
1655 	 * while we are running, so we want to acquire it before
1656 	 * looking at db_blkptr.
1657 	 */
1658 	if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1659 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
1660 		drop_struct_lock = TRUE;
1661 	}
1662 
1663 	/*
1664 	 * We need to hold the dn_struct_rwlock to make this assertion,
1665 	 * because it protects dn_phys / dn_next_nlevels from changing.
1666 	 */
1667 	ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1668 	    dn->dn_phys->dn_nlevels > db->db_level ||
1669 	    dn->dn_next_nlevels[txgoff] > db->db_level ||
1670 	    dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1671 	    dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1672 
1673 	/*
1674 	 * If we are overwriting a dedup BP, then unless it is snapshotted,
1675 	 * when we get to syncing context we will need to decrement its
1676 	 * refcount in the DDT.  Prefetch the relevant DDT block so that
1677 	 * syncing context won't have to wait for the i/o.
1678 	 */
1679 	ddt_prefetch(os->os_spa, db->db_blkptr);
1680 
1681 	if (db->db_level == 0) {
1682 		dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1683 		ASSERT(dn->dn_maxblkid >= db->db_blkid);
1684 	}
1685 
1686 	if (db->db_level+1 < dn->dn_nlevels) {
1687 		dmu_buf_impl_t *parent = db->db_parent;
1688 		dbuf_dirty_record_t *di;
1689 		int parent_held = FALSE;
1690 
1691 		if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1692 			int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1693 
1694 			parent = dbuf_hold_level(dn, db->db_level+1,
1695 			    db->db_blkid >> epbs, FTAG);
1696 			ASSERT(parent != NULL);
1697 			parent_held = TRUE;
1698 		}
1699 		if (drop_struct_lock)
1700 			rw_exit(&dn->dn_struct_rwlock);
1701 		ASSERT3U(db->db_level+1, ==, parent->db_level);
1702 		di = dbuf_dirty(parent, tx);
1703 		if (parent_held)
1704 			dbuf_rele(parent, FTAG);
1705 
1706 		mutex_enter(&db->db_mtx);
1707 		/*
1708 		 * Since we've dropped the mutex, it's possible that
1709 		 * dbuf_undirty() might have changed this out from under us.
1710 		 */
1711 		if (db->db_last_dirty == dr ||
1712 		    dn->dn_object == DMU_META_DNODE_OBJECT) {
1713 			mutex_enter(&di->dt.di.dr_mtx);
1714 			ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1715 			ASSERT(!list_link_active(&dr->dr_dirty_node));
1716 			list_insert_tail(&di->dt.di.dr_children, dr);
1717 			mutex_exit(&di->dt.di.dr_mtx);
1718 			dr->dr_parent = di;
1719 		}
1720 		mutex_exit(&db->db_mtx);
1721 	} else {
1722 		ASSERT(db->db_level+1 == dn->dn_nlevels);
1723 		ASSERT(db->db_blkid < dn->dn_nblkptr);
1724 		ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1725 		mutex_enter(&dn->dn_mtx);
1726 		ASSERT(!list_link_active(&dr->dr_dirty_node));
1727 		list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1728 		mutex_exit(&dn->dn_mtx);
1729 		if (drop_struct_lock)
1730 			rw_exit(&dn->dn_struct_rwlock);
1731 	}
1732 
1733 	dnode_setdirty(dn, tx);
1734 	DB_DNODE_EXIT(db);
1735 	return (dr);
1736 }
1737 
1738 /*
1739  * Undirty a buffer in the transaction group referenced by the given
1740  * transaction.  Return whether this evicted the dbuf.
1741  */
1742 static boolean_t
1743 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1744 {
1745 	dnode_t *dn;
1746 	uint64_t txg = tx->tx_txg;
1747 	dbuf_dirty_record_t *dr, **drp;
1748 
1749 	ASSERT(txg != 0);
1750 
1751 	/*
1752 	 * Due to our use of dn_nlevels below, this can only be called
1753 	 * in open context, unless we are operating on the MOS.
1754 	 * From syncing context, dn_nlevels may be different from the
1755 	 * dn_nlevels used when dbuf was dirtied.
1756 	 */
1757 	ASSERT(db->db_objset ==
1758 	    dmu_objset_pool(db->db_objset)->dp_meta_objset ||
1759 	    txg != spa_syncing_txg(dmu_objset_spa(db->db_objset)));
1760 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1761 	ASSERT0(db->db_level);
1762 	ASSERT(MUTEX_HELD(&db->db_mtx));
1763 
1764 	/*
1765 	 * If this buffer is not dirty, we're done.
1766 	 */
1767 	for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1768 		if (dr->dr_txg <= txg)
1769 			break;
1770 	if (dr == NULL || dr->dr_txg < txg)
1771 		return (B_FALSE);
1772 	ASSERT(dr->dr_txg == txg);
1773 	ASSERT(dr->dr_dbuf == db);
1774 
1775 	DB_DNODE_ENTER(db);
1776 	dn = DB_DNODE(db);
1777 
1778 	dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1779 
1780 	ASSERT(db->db.db_size != 0);
1781 
1782 	dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset),
1783 	    dr->dr_accounted, txg);
1784 
1785 	*drp = dr->dr_next;
1786 
1787 	/*
1788 	 * Note that there are three places in dbuf_dirty()
1789 	 * where this dirty record may be put on a list.
1790 	 * Make sure to do a list_remove corresponding to
1791 	 * every one of those list_insert calls.
1792 	 */
1793 	if (dr->dr_parent) {
1794 		mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1795 		list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1796 		mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1797 	} else if (db->db_blkid == DMU_SPILL_BLKID ||
1798 	    db->db_level + 1 == dn->dn_nlevels) {
1799 		ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1800 		mutex_enter(&dn->dn_mtx);
1801 		list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1802 		mutex_exit(&dn->dn_mtx);
1803 	}
1804 	DB_DNODE_EXIT(db);
1805 
1806 	if (db->db_state != DB_NOFILL) {
1807 		dbuf_unoverride(dr);
1808 
1809 		ASSERT(db->db_buf != NULL);
1810 		ASSERT(dr->dt.dl.dr_data != NULL);
1811 		if (dr->dt.dl.dr_data != db->db_buf)
1812 			arc_buf_destroy(dr->dt.dl.dr_data, db);
1813 	}
1814 
1815 	kmem_free(dr, sizeof (dbuf_dirty_record_t));
1816 
1817 	ASSERT(db->db_dirtycnt > 0);
1818 	db->db_dirtycnt -= 1;
1819 
1820 	if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1821 		ASSERT(db->db_state == DB_NOFILL || arc_released(db->db_buf));
1822 		dbuf_destroy(db);
1823 		return (B_TRUE);
1824 	}
1825 
1826 	return (B_FALSE);
1827 }
1828 
1829 void
1830 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
1831 {
1832 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1833 	int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1834 
1835 	ASSERT(tx->tx_txg != 0);
1836 	ASSERT(!refcount_is_zero(&db->db_holds));
1837 
1838 	/*
1839 	 * Quick check for dirtyness.  For already dirty blocks, this
1840 	 * reduces runtime of this function by >90%, and overall performance
1841 	 * by 50% for some workloads (e.g. file deletion with indirect blocks
1842 	 * cached).
1843 	 */
1844 	mutex_enter(&db->db_mtx);
1845 	dbuf_dirty_record_t *dr;
1846 	for (dr = db->db_last_dirty;
1847 	    dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) {
1848 		/*
1849 		 * It's possible that it is already dirty but not cached,
1850 		 * because there are some calls to dbuf_dirty() that don't
1851 		 * go through dmu_buf_will_dirty().
1852 		 */
1853 		if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) {
1854 			/* This dbuf is already dirty and cached. */
1855 			dbuf_redirty(dr);
1856 			mutex_exit(&db->db_mtx);
1857 			return;
1858 		}
1859 	}
1860 	mutex_exit(&db->db_mtx);
1861 
1862 	DB_DNODE_ENTER(db);
1863 	if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
1864 		rf |= DB_RF_HAVESTRUCT;
1865 	DB_DNODE_EXIT(db);
1866 	(void) dbuf_read(db, NULL, rf);
1867 	(void) dbuf_dirty(db, tx);
1868 }
1869 
1870 void
1871 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1872 {
1873 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1874 
1875 	db->db_state = DB_NOFILL;
1876 
1877 	dmu_buf_will_fill(db_fake, tx);
1878 }
1879 
1880 void
1881 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1882 {
1883 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1884 
1885 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1886 	ASSERT(tx->tx_txg != 0);
1887 	ASSERT(db->db_level == 0);
1888 	ASSERT(!refcount_is_zero(&db->db_holds));
1889 
1890 	ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1891 	    dmu_tx_private_ok(tx));
1892 
1893 	dbuf_noread(db);
1894 	(void) dbuf_dirty(db, tx);
1895 }
1896 
1897 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1898 /* ARGSUSED */
1899 void
1900 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1901 {
1902 	mutex_enter(&db->db_mtx);
1903 	DBUF_VERIFY(db);
1904 
1905 	if (db->db_state == DB_FILL) {
1906 		if (db->db_level == 0 && db->db_freed_in_flight) {
1907 			ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1908 			/* we were freed while filling */
1909 			/* XXX dbuf_undirty? */
1910 			bzero(db->db.db_data, db->db.db_size);
1911 			db->db_freed_in_flight = FALSE;
1912 		}
1913 		db->db_state = DB_CACHED;
1914 		cv_broadcast(&db->db_changed);
1915 	}
1916 	mutex_exit(&db->db_mtx);
1917 }
1918 
1919 void
1920 dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data,
1921     bp_embedded_type_t etype, enum zio_compress comp,
1922     int uncompressed_size, int compressed_size, int byteorder,
1923     dmu_tx_t *tx)
1924 {
1925 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
1926 	struct dirty_leaf *dl;
1927 	dmu_object_type_t type;
1928 
1929 	if (etype == BP_EMBEDDED_TYPE_DATA) {
1930 		ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset),
1931 		    SPA_FEATURE_EMBEDDED_DATA));
1932 	}
1933 
1934 	DB_DNODE_ENTER(db);
1935 	type = DB_DNODE(db)->dn_type;
1936 	DB_DNODE_EXIT(db);
1937 
1938 	ASSERT0(db->db_level);
1939 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1940 
1941 	dmu_buf_will_not_fill(dbuf, tx);
1942 
1943 	ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1944 	dl = &db->db_last_dirty->dt.dl;
1945 	encode_embedded_bp_compressed(&dl->dr_overridden_by,
1946 	    data, comp, uncompressed_size, compressed_size);
1947 	BPE_SET_ETYPE(&dl->dr_overridden_by, etype);
1948 	BP_SET_TYPE(&dl->dr_overridden_by, type);
1949 	BP_SET_LEVEL(&dl->dr_overridden_by, 0);
1950 	BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder);
1951 
1952 	dl->dr_override_state = DR_OVERRIDDEN;
1953 	dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg;
1954 }
1955 
1956 /*
1957  * Directly assign a provided arc buf to a given dbuf if it's not referenced
1958  * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1959  */
1960 void
1961 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
1962 {
1963 	ASSERT(!refcount_is_zero(&db->db_holds));
1964 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1965 	ASSERT(db->db_level == 0);
1966 	ASSERT3U(dbuf_is_metadata(db), ==, arc_is_metadata(buf));
1967 	ASSERT(buf != NULL);
1968 	ASSERT(arc_buf_lsize(buf) == db->db.db_size);
1969 	ASSERT(tx->tx_txg != 0);
1970 
1971 	arc_return_buf(buf, db);
1972 	ASSERT(arc_released(buf));
1973 
1974 	mutex_enter(&db->db_mtx);
1975 
1976 	while (db->db_state == DB_READ || db->db_state == DB_FILL)
1977 		cv_wait(&db->db_changed, &db->db_mtx);
1978 
1979 	ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
1980 
1981 	if (db->db_state == DB_CACHED &&
1982 	    refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
1983 		mutex_exit(&db->db_mtx);
1984 		(void) dbuf_dirty(db, tx);
1985 		bcopy(buf->b_data, db->db.db_data, db->db.db_size);
1986 		arc_buf_destroy(buf, db);
1987 		xuio_stat_wbuf_copied();
1988 		return;
1989 	}
1990 
1991 	xuio_stat_wbuf_nocopy();
1992 	if (db->db_state == DB_CACHED) {
1993 		dbuf_dirty_record_t *dr = db->db_last_dirty;
1994 
1995 		ASSERT(db->db_buf != NULL);
1996 		if (dr != NULL && dr->dr_txg == tx->tx_txg) {
1997 			ASSERT(dr->dt.dl.dr_data == db->db_buf);
1998 			if (!arc_released(db->db_buf)) {
1999 				ASSERT(dr->dt.dl.dr_override_state ==
2000 				    DR_OVERRIDDEN);
2001 				arc_release(db->db_buf, db);
2002 			}
2003 			dr->dt.dl.dr_data = buf;
2004 			arc_buf_destroy(db->db_buf, db);
2005 		} else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
2006 			arc_release(db->db_buf, db);
2007 			arc_buf_destroy(db->db_buf, db);
2008 		}
2009 		db->db_buf = NULL;
2010 	}
2011 	ASSERT(db->db_buf == NULL);
2012 	dbuf_set_data(db, buf);
2013 	db->db_state = DB_FILL;
2014 	mutex_exit(&db->db_mtx);
2015 	(void) dbuf_dirty(db, tx);
2016 	dmu_buf_fill_done(&db->db, tx);
2017 }
2018 
2019 void
2020 dbuf_destroy(dmu_buf_impl_t *db)
2021 {
2022 	dnode_t *dn;
2023 	dmu_buf_impl_t *parent = db->db_parent;
2024 	dmu_buf_impl_t *dndb;
2025 
2026 	ASSERT(MUTEX_HELD(&db->db_mtx));
2027 	ASSERT(refcount_is_zero(&db->db_holds));
2028 
2029 	if (db->db_buf != NULL) {
2030 		arc_buf_destroy(db->db_buf, db);
2031 		db->db_buf = NULL;
2032 	}
2033 
2034 	if (db->db_blkid == DMU_BONUS_BLKID) {
2035 		ASSERT(db->db.db_data != NULL);
2036 		zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
2037 		arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
2038 		db->db_state = DB_UNCACHED;
2039 	}
2040 
2041 	dbuf_clear_data(db);
2042 
2043 	if (multilist_link_active(&db->db_cache_link)) {
2044 		multilist_remove(dbuf_cache, db);
2045 		(void) refcount_remove_many(&dbuf_cache_size,
2046 		    db->db.db_size, db);
2047 	}
2048 
2049 	ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
2050 	ASSERT(db->db_data_pending == NULL);
2051 
2052 	db->db_state = DB_EVICTING;
2053 	db->db_blkptr = NULL;
2054 
2055 	/*
2056 	 * Now that db_state is DB_EVICTING, nobody else can find this via
2057 	 * the hash table.  We can now drop db_mtx, which allows us to
2058 	 * acquire the dn_dbufs_mtx.
2059 	 */
2060 	mutex_exit(&db->db_mtx);
2061 
2062 	DB_DNODE_ENTER(db);
2063 	dn = DB_DNODE(db);
2064 	dndb = dn->dn_dbuf;
2065 	if (db->db_blkid != DMU_BONUS_BLKID) {
2066 		boolean_t needlock = !MUTEX_HELD(&dn->dn_dbufs_mtx);
2067 		if (needlock)
2068 			mutex_enter(&dn->dn_dbufs_mtx);
2069 		avl_remove(&dn->dn_dbufs, db);
2070 		atomic_dec_32(&dn->dn_dbufs_count);
2071 		membar_producer();
2072 		DB_DNODE_EXIT(db);
2073 		if (needlock)
2074 			mutex_exit(&dn->dn_dbufs_mtx);
2075 		/*
2076 		 * Decrementing the dbuf count means that the hold corresponding
2077 		 * to the removed dbuf is no longer discounted in dnode_move(),
2078 		 * so the dnode cannot be moved until after we release the hold.
2079 		 * The membar_producer() ensures visibility of the decremented
2080 		 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
2081 		 * release any lock.
2082 		 */
2083 		dnode_rele(dn, db);
2084 		db->db_dnode_handle = NULL;
2085 
2086 		dbuf_hash_remove(db);
2087 	} else {
2088 		DB_DNODE_EXIT(db);
2089 	}
2090 
2091 	ASSERT(refcount_is_zero(&db->db_holds));
2092 
2093 	db->db_parent = NULL;
2094 
2095 	ASSERT(db->db_buf == NULL);
2096 	ASSERT(db->db.db_data == NULL);
2097 	ASSERT(db->db_hash_next == NULL);
2098 	ASSERT(db->db_blkptr == NULL);
2099 	ASSERT(db->db_data_pending == NULL);
2100 	ASSERT(!multilist_link_active(&db->db_cache_link));
2101 
2102 	kmem_cache_free(dbuf_kmem_cache, db);
2103 	arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2104 
2105 	/*
2106 	 * If this dbuf is referenced from an indirect dbuf,
2107 	 * decrement the ref count on the indirect dbuf.
2108 	 */
2109 	if (parent && parent != dndb)
2110 		dbuf_rele(parent, db);
2111 }
2112 
2113 /*
2114  * Note: While bpp will always be updated if the function returns success,
2115  * parentp will not be updated if the dnode does not have dn_dbuf filled in;
2116  * this happens when the dnode is the meta-dnode, or a userused or groupused
2117  * object.
2118  */
2119 static int
2120 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
2121     dmu_buf_impl_t **parentp, blkptr_t **bpp)
2122 {
2123 	*parentp = NULL;
2124 	*bpp = NULL;
2125 
2126 	ASSERT(blkid != DMU_BONUS_BLKID);
2127 
2128 	if (blkid == DMU_SPILL_BLKID) {
2129 		mutex_enter(&dn->dn_mtx);
2130 		if (dn->dn_have_spill &&
2131 		    (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
2132 			*bpp = &dn->dn_phys->dn_spill;
2133 		else
2134 			*bpp = NULL;
2135 		dbuf_add_ref(dn->dn_dbuf, NULL);
2136 		*parentp = dn->dn_dbuf;
2137 		mutex_exit(&dn->dn_mtx);
2138 		return (0);
2139 	}
2140 
2141 	int nlevels =
2142 	    (dn->dn_phys->dn_nlevels == 0) ? 1 : dn->dn_phys->dn_nlevels;
2143 	int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2144 
2145 	ASSERT3U(level * epbs, <, 64);
2146 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2147 	/*
2148 	 * This assertion shouldn't trip as long as the max indirect block size
2149 	 * is less than 1M.  The reason for this is that up to that point,
2150 	 * the number of levels required to address an entire object with blocks
2151 	 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64.  In
2152 	 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55
2153 	 * (i.e. we can address the entire object), objects will all use at most
2154 	 * N-1 levels and the assertion won't overflow.  However, once epbs is
2155 	 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66.  Then, 4 levels will not be
2156 	 * enough to address an entire object, so objects will have 5 levels,
2157 	 * but then this assertion will overflow.
2158 	 *
2159 	 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we
2160 	 * need to redo this logic to handle overflows.
2161 	 */
2162 	ASSERT(level >= nlevels ||
2163 	    ((nlevels - level - 1) * epbs) +
2164 	    highbit64(dn->dn_phys->dn_nblkptr) <= 64);
2165 	if (level >= nlevels ||
2166 	    blkid >= ((uint64_t)dn->dn_phys->dn_nblkptr <<
2167 	    ((nlevels - level - 1) * epbs)) ||
2168 	    (fail_sparse &&
2169 	    blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
2170 		/* the buffer has no parent yet */
2171 		return (SET_ERROR(ENOENT));
2172 	} else if (level < nlevels-1) {
2173 		/* this block is referenced from an indirect block */
2174 		int err = dbuf_hold_impl(dn, level+1,
2175 		    blkid >> epbs, fail_sparse, FALSE, NULL, parentp);
2176 		if (err)
2177 			return (err);
2178 		err = dbuf_read(*parentp, NULL,
2179 		    (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
2180 		if (err) {
2181 			dbuf_rele(*parentp, NULL);
2182 			*parentp = NULL;
2183 			return (err);
2184 		}
2185 		*bpp = ((blkptr_t *)(*parentp)->db.db_data) +
2186 		    (blkid & ((1ULL << epbs) - 1));
2187 		if (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))
2188 			ASSERT(BP_IS_HOLE(*bpp));
2189 		return (0);
2190 	} else {
2191 		/* the block is referenced from the dnode */
2192 		ASSERT3U(level, ==, nlevels-1);
2193 		ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
2194 		    blkid < dn->dn_phys->dn_nblkptr);
2195 		if (dn->dn_dbuf) {
2196 			dbuf_add_ref(dn->dn_dbuf, NULL);
2197 			*parentp = dn->dn_dbuf;
2198 		}
2199 		*bpp = &dn->dn_phys->dn_blkptr[blkid];
2200 		return (0);
2201 	}
2202 }
2203 
2204 static dmu_buf_impl_t *
2205 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
2206     dmu_buf_impl_t *parent, blkptr_t *blkptr)
2207 {
2208 	objset_t *os = dn->dn_objset;
2209 	dmu_buf_impl_t *db, *odb;
2210 
2211 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2212 	ASSERT(dn->dn_type != DMU_OT_NONE);
2213 
2214 	db = kmem_cache_alloc(dbuf_kmem_cache, KM_SLEEP);
2215 
2216 	db->db_objset = os;
2217 	db->db.db_object = dn->dn_object;
2218 	db->db_level = level;
2219 	db->db_blkid = blkid;
2220 	db->db_last_dirty = NULL;
2221 	db->db_dirtycnt = 0;
2222 	db->db_dnode_handle = dn->dn_handle;
2223 	db->db_parent = parent;
2224 	db->db_blkptr = blkptr;
2225 
2226 	db->db_user = NULL;
2227 	db->db_user_immediate_evict = FALSE;
2228 	db->db_freed_in_flight = FALSE;
2229 	db->db_pending_evict = FALSE;
2230 
2231 	if (blkid == DMU_BONUS_BLKID) {
2232 		ASSERT3P(parent, ==, dn->dn_dbuf);
2233 		db->db.db_size = DN_MAX_BONUSLEN -
2234 		    (dn->dn_nblkptr-1) * sizeof (blkptr_t);
2235 		ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
2236 		db->db.db_offset = DMU_BONUS_BLKID;
2237 		db->db_state = DB_UNCACHED;
2238 		/* the bonus dbuf is not placed in the hash table */
2239 		arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2240 		return (db);
2241 	} else if (blkid == DMU_SPILL_BLKID) {
2242 		db->db.db_size = (blkptr != NULL) ?
2243 		    BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
2244 		db->db.db_offset = 0;
2245 	} else {
2246 		int blocksize =
2247 		    db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
2248 		db->db.db_size = blocksize;
2249 		db->db.db_offset = db->db_blkid * blocksize;
2250 	}
2251 
2252 	/*
2253 	 * Hold the dn_dbufs_mtx while we get the new dbuf
2254 	 * in the hash table *and* added to the dbufs list.
2255 	 * This prevents a possible deadlock with someone
2256 	 * trying to look up this dbuf before its added to the
2257 	 * dn_dbufs list.
2258 	 */
2259 	mutex_enter(&dn->dn_dbufs_mtx);
2260 	db->db_state = DB_EVICTING;
2261 	if ((odb = dbuf_hash_insert(db)) != NULL) {
2262 		/* someone else inserted it first */
2263 		kmem_cache_free(dbuf_kmem_cache, db);
2264 		mutex_exit(&dn->dn_dbufs_mtx);
2265 		return (odb);
2266 	}
2267 	avl_add(&dn->dn_dbufs, db);
2268 
2269 	db->db_state = DB_UNCACHED;
2270 	mutex_exit(&dn->dn_dbufs_mtx);
2271 	arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2272 
2273 	if (parent && parent != dn->dn_dbuf)
2274 		dbuf_add_ref(parent, db);
2275 
2276 	ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
2277 	    refcount_count(&dn->dn_holds) > 0);
2278 	(void) refcount_add(&dn->dn_holds, db);
2279 	atomic_inc_32(&dn->dn_dbufs_count);
2280 
2281 	dprintf_dbuf(db, "db=%p\n", db);
2282 
2283 	return (db);
2284 }
2285 
2286 typedef struct dbuf_prefetch_arg {
2287 	spa_t *dpa_spa;	/* The spa to issue the prefetch in. */
2288 	zbookmark_phys_t dpa_zb; /* The target block to prefetch. */
2289 	int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */
2290 	int dpa_curlevel; /* The current level that we're reading */
2291 	dnode_t *dpa_dnode; /* The dnode associated with the prefetch */
2292 	zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */
2293 	zio_t *dpa_zio; /* The parent zio_t for all prefetches. */
2294 	arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */
2295 } dbuf_prefetch_arg_t;
2296 
2297 /*
2298  * Actually issue the prefetch read for the block given.
2299  */
2300 static void
2301 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp)
2302 {
2303 	if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
2304 		return;
2305 
2306 	arc_flags_t aflags =
2307 	    dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
2308 
2309 	ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2310 	ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level);
2311 	ASSERT(dpa->dpa_zio != NULL);
2312 	(void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL,
2313 	    dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2314 	    &aflags, &dpa->dpa_zb);
2315 }
2316 
2317 /*
2318  * Called when an indirect block above our prefetch target is read in.  This
2319  * will either read in the next indirect block down the tree or issue the actual
2320  * prefetch if the next block down is our target.
2321  */
2322 static void
2323 dbuf_prefetch_indirect_done(zio_t *zio, arc_buf_t *abuf, void *private)
2324 {
2325 	dbuf_prefetch_arg_t *dpa = private;
2326 
2327 	ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel);
2328 	ASSERT3S(dpa->dpa_curlevel, >, 0);
2329 
2330 	/*
2331 	 * The dpa_dnode is only valid if we are called with a NULL
2332 	 * zio. This indicates that the arc_read() returned without
2333 	 * first calling zio_read() to issue a physical read. Once
2334 	 * a physical read is made the dpa_dnode must be invalidated
2335 	 * as the locks guarding it may have been dropped. If the
2336 	 * dpa_dnode is still valid, then we want to add it to the dbuf
2337 	 * cache. To do so, we must hold the dbuf associated with the block
2338 	 * we just prefetched, read its contents so that we associate it
2339 	 * with an arc_buf_t, and then release it.
2340 	 */
2341 	if (zio != NULL) {
2342 		ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel);
2343 		if (zio->io_flags & ZIO_FLAG_RAW) {
2344 			ASSERT3U(BP_GET_PSIZE(zio->io_bp), ==, zio->io_size);
2345 		} else {
2346 			ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size);
2347 		}
2348 		ASSERT3P(zio->io_spa, ==, dpa->dpa_spa);
2349 
2350 		dpa->dpa_dnode = NULL;
2351 	} else if (dpa->dpa_dnode != NULL) {
2352 		uint64_t curblkid = dpa->dpa_zb.zb_blkid >>
2353 		    (dpa->dpa_epbs * (dpa->dpa_curlevel -
2354 		    dpa->dpa_zb.zb_level));
2355 		dmu_buf_impl_t *db = dbuf_hold_level(dpa->dpa_dnode,
2356 		    dpa->dpa_curlevel, curblkid, FTAG);
2357 		(void) dbuf_read(db, NULL,
2358 		    DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_HAVESTRUCT);
2359 		dbuf_rele(db, FTAG);
2360 	}
2361 
2362 	dpa->dpa_curlevel--;
2363 
2364 	uint64_t nextblkid = dpa->dpa_zb.zb_blkid >>
2365 	    (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level));
2366 	blkptr_t *bp = ((blkptr_t *)abuf->b_data) +
2367 	    P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs);
2368 	if (BP_IS_HOLE(bp) || (zio != NULL && zio->io_error != 0)) {
2369 		kmem_free(dpa, sizeof (*dpa));
2370 	} else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) {
2371 		ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid);
2372 		dbuf_issue_final_prefetch(dpa, bp);
2373 		kmem_free(dpa, sizeof (*dpa));
2374 	} else {
2375 		arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2376 		zbookmark_phys_t zb;
2377 
2378 		/* flag if L2ARC eligible, l2arc_noprefetch then decides */
2379 		if (dpa->dpa_aflags & ARC_FLAG_L2CACHE)
2380 			iter_aflags |= ARC_FLAG_L2CACHE;
2381 
2382 		ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2383 
2384 		SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset,
2385 		    dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid);
2386 
2387 		(void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2388 		    bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio,
2389 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2390 		    &iter_aflags, &zb);
2391 	}
2392 
2393 	arc_buf_destroy(abuf, private);
2394 }
2395 
2396 /*
2397  * Issue prefetch reads for the given block on the given level.  If the indirect
2398  * blocks above that block are not in memory, we will read them in
2399  * asynchronously.  As a result, this call never blocks waiting for a read to
2400  * complete.
2401  */
2402 void
2403 dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio,
2404     arc_flags_t aflags)
2405 {
2406 	blkptr_t bp;
2407 	int epbs, nlevels, curlevel;
2408 	uint64_t curblkid;
2409 
2410 	ASSERT(blkid != DMU_BONUS_BLKID);
2411 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2412 
2413 	if (blkid > dn->dn_maxblkid)
2414 		return;
2415 
2416 	if (dnode_block_freed(dn, blkid))
2417 		return;
2418 
2419 	/*
2420 	 * This dnode hasn't been written to disk yet, so there's nothing to
2421 	 * prefetch.
2422 	 */
2423 	nlevels = dn->dn_phys->dn_nlevels;
2424 	if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0)
2425 		return;
2426 
2427 	epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2428 	if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level))
2429 		return;
2430 
2431 	dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object,
2432 	    level, blkid);
2433 	if (db != NULL) {
2434 		mutex_exit(&db->db_mtx);
2435 		/*
2436 		 * This dbuf already exists.  It is either CACHED, or
2437 		 * (we assume) about to be read or filled.
2438 		 */
2439 		return;
2440 	}
2441 
2442 	/*
2443 	 * Find the closest ancestor (indirect block) of the target block
2444 	 * that is present in the cache.  In this indirect block, we will
2445 	 * find the bp that is at curlevel, curblkid.
2446 	 */
2447 	curlevel = level;
2448 	curblkid = blkid;
2449 	while (curlevel < nlevels - 1) {
2450 		int parent_level = curlevel + 1;
2451 		uint64_t parent_blkid = curblkid >> epbs;
2452 		dmu_buf_impl_t *db;
2453 
2454 		if (dbuf_hold_impl(dn, parent_level, parent_blkid,
2455 		    FALSE, TRUE, FTAG, &db) == 0) {
2456 			blkptr_t *bpp = db->db_buf->b_data;
2457 			bp = bpp[P2PHASE(curblkid, 1 << epbs)];
2458 			dbuf_rele(db, FTAG);
2459 			break;
2460 		}
2461 
2462 		curlevel = parent_level;
2463 		curblkid = parent_blkid;
2464 	}
2465 
2466 	if (curlevel == nlevels - 1) {
2467 		/* No cached indirect blocks found. */
2468 		ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr);
2469 		bp = dn->dn_phys->dn_blkptr[curblkid];
2470 	}
2471 	if (BP_IS_HOLE(&bp))
2472 		return;
2473 
2474 	ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp));
2475 
2476 	zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL,
2477 	    ZIO_FLAG_CANFAIL);
2478 
2479 	dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP);
2480 	dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
2481 	SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2482 	    dn->dn_object, level, blkid);
2483 	dpa->dpa_curlevel = curlevel;
2484 	dpa->dpa_prio = prio;
2485 	dpa->dpa_aflags = aflags;
2486 	dpa->dpa_spa = dn->dn_objset->os_spa;
2487 	dpa->dpa_dnode = dn;
2488 	dpa->dpa_epbs = epbs;
2489 	dpa->dpa_zio = pio;
2490 
2491 	/* flag if L2ARC eligible, l2arc_noprefetch then decides */
2492 	if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level))
2493 		dpa->dpa_aflags |= ARC_FLAG_L2CACHE;
2494 
2495 	/*
2496 	 * If we have the indirect just above us, no need to do the asynchronous
2497 	 * prefetch chain; we'll just run the last step ourselves.  If we're at
2498 	 * a higher level, though, we want to issue the prefetches for all the
2499 	 * indirect blocks asynchronously, so we can go on with whatever we were
2500 	 * doing.
2501 	 */
2502 	if (curlevel == level) {
2503 		ASSERT3U(curblkid, ==, blkid);
2504 		dbuf_issue_final_prefetch(dpa, &bp);
2505 		kmem_free(dpa, sizeof (*dpa));
2506 	} else {
2507 		arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2508 		zbookmark_phys_t zb;
2509 
2510 		/* flag if L2ARC eligible, l2arc_noprefetch then decides */
2511 		if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level))
2512 			iter_aflags |= ARC_FLAG_L2CACHE;
2513 
2514 		SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2515 		    dn->dn_object, curlevel, curblkid);
2516 		(void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2517 		    &bp, dbuf_prefetch_indirect_done, dpa, prio,
2518 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2519 		    &iter_aflags, &zb);
2520 	}
2521 	/*
2522 	 * We use pio here instead of dpa_zio since it's possible that
2523 	 * dpa may have already been freed.
2524 	 */
2525 	zio_nowait(pio);
2526 }
2527 
2528 /*
2529  * Returns with db_holds incremented, and db_mtx not held.
2530  * Note: dn_struct_rwlock must be held.
2531  */
2532 int
2533 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid,
2534     boolean_t fail_sparse, boolean_t fail_uncached,
2535     void *tag, dmu_buf_impl_t **dbp)
2536 {
2537 	dmu_buf_impl_t *db, *parent = NULL;
2538 
2539 	ASSERT(blkid != DMU_BONUS_BLKID);
2540 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2541 	ASSERT3U(dn->dn_nlevels, >, level);
2542 
2543 	*dbp = NULL;
2544 top:
2545 	/* dbuf_find() returns with db_mtx held */
2546 	db = dbuf_find(dn->dn_objset, dn->dn_object, level, blkid);
2547 
2548 	if (db == NULL) {
2549 		blkptr_t *bp = NULL;
2550 		int err;
2551 
2552 		if (fail_uncached)
2553 			return (SET_ERROR(ENOENT));
2554 
2555 		ASSERT3P(parent, ==, NULL);
2556 		err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
2557 		if (fail_sparse) {
2558 			if (err == 0 && bp && BP_IS_HOLE(bp))
2559 				err = SET_ERROR(ENOENT);
2560 			if (err) {
2561 				if (parent)
2562 					dbuf_rele(parent, NULL);
2563 				return (err);
2564 			}
2565 		}
2566 		if (err && err != ENOENT)
2567 			return (err);
2568 		db = dbuf_create(dn, level, blkid, parent, bp);
2569 	}
2570 
2571 	if (fail_uncached && db->db_state != DB_CACHED) {
2572 		mutex_exit(&db->db_mtx);
2573 		return (SET_ERROR(ENOENT));
2574 	}
2575 
2576 	if (db->db_buf != NULL)
2577 		ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
2578 
2579 	ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
2580 
2581 	/*
2582 	 * If this buffer is currently syncing out, and we are are
2583 	 * still referencing it from db_data, we need to make a copy
2584 	 * of it in case we decide we want to dirty it again in this txg.
2585 	 */
2586 	if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
2587 	    dn->dn_object != DMU_META_DNODE_OBJECT &&
2588 	    db->db_state == DB_CACHED && db->db_data_pending) {
2589 		dbuf_dirty_record_t *dr = db->db_data_pending;
2590 
2591 		if (dr->dt.dl.dr_data == db->db_buf) {
2592 			arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2593 
2594 			dbuf_set_data(db,
2595 			    arc_alloc_buf(dn->dn_objset->os_spa, db, type,
2596 			    db->db.db_size));
2597 			bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
2598 			    db->db.db_size);
2599 		}
2600 	}
2601 
2602 	if (multilist_link_active(&db->db_cache_link)) {
2603 		ASSERT(refcount_is_zero(&db->db_holds));
2604 		multilist_remove(dbuf_cache, db);
2605 		(void) refcount_remove_many(&dbuf_cache_size,
2606 		    db->db.db_size, db);
2607 	}
2608 	(void) refcount_add(&db->db_holds, tag);
2609 	DBUF_VERIFY(db);
2610 	mutex_exit(&db->db_mtx);
2611 
2612 	/* NOTE: we can't rele the parent until after we drop the db_mtx */
2613 	if (parent)
2614 		dbuf_rele(parent, NULL);
2615 
2616 	ASSERT3P(DB_DNODE(db), ==, dn);
2617 	ASSERT3U(db->db_blkid, ==, blkid);
2618 	ASSERT3U(db->db_level, ==, level);
2619 	*dbp = db;
2620 
2621 	return (0);
2622 }
2623 
2624 dmu_buf_impl_t *
2625 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
2626 {
2627 	return (dbuf_hold_level(dn, 0, blkid, tag));
2628 }
2629 
2630 dmu_buf_impl_t *
2631 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
2632 {
2633 	dmu_buf_impl_t *db;
2634 	int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db);
2635 	return (err ? NULL : db);
2636 }
2637 
2638 void
2639 dbuf_create_bonus(dnode_t *dn)
2640 {
2641 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
2642 
2643 	ASSERT(dn->dn_bonus == NULL);
2644 	dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
2645 }
2646 
2647 int
2648 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
2649 {
2650 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2651 	dnode_t *dn;
2652 
2653 	if (db->db_blkid != DMU_SPILL_BLKID)
2654 		return (SET_ERROR(ENOTSUP));
2655 	if (blksz == 0)
2656 		blksz = SPA_MINBLOCKSIZE;
2657 	ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset)));
2658 	blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
2659 
2660 	DB_DNODE_ENTER(db);
2661 	dn = DB_DNODE(db);
2662 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2663 	dbuf_new_size(db, blksz, tx);
2664 	rw_exit(&dn->dn_struct_rwlock);
2665 	DB_DNODE_EXIT(db);
2666 
2667 	return (0);
2668 }
2669 
2670 void
2671 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
2672 {
2673 	dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
2674 }
2675 
2676 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2677 void
2678 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
2679 {
2680 	int64_t holds = refcount_add(&db->db_holds, tag);
2681 	ASSERT3S(holds, >, 1);
2682 }
2683 
2684 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
2685 boolean_t
2686 dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid,
2687     void *tag)
2688 {
2689 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2690 	dmu_buf_impl_t *found_db;
2691 	boolean_t result = B_FALSE;
2692 
2693 	if (db->db_blkid == DMU_BONUS_BLKID)
2694 		found_db = dbuf_find_bonus(os, obj);
2695 	else
2696 		found_db = dbuf_find(os, obj, 0, blkid);
2697 
2698 	if (found_db != NULL) {
2699 		if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) {
2700 			(void) refcount_add(&db->db_holds, tag);
2701 			result = B_TRUE;
2702 		}
2703 		mutex_exit(&db->db_mtx);
2704 	}
2705 	return (result);
2706 }
2707 
2708 /*
2709  * If you call dbuf_rele() you had better not be referencing the dnode handle
2710  * unless you have some other direct or indirect hold on the dnode. (An indirect
2711  * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2712  * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2713  * dnode's parent dbuf evicting its dnode handles.
2714  */
2715 void
2716 dbuf_rele(dmu_buf_impl_t *db, void *tag)
2717 {
2718 	mutex_enter(&db->db_mtx);
2719 	dbuf_rele_and_unlock(db, tag);
2720 }
2721 
2722 void
2723 dmu_buf_rele(dmu_buf_t *db, void *tag)
2724 {
2725 	dbuf_rele((dmu_buf_impl_t *)db, tag);
2726 }
2727 
2728 /*
2729  * dbuf_rele() for an already-locked dbuf.  This is necessary to allow
2730  * db_dirtycnt and db_holds to be updated atomically.
2731  */
2732 void
2733 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2734 {
2735 	int64_t holds;
2736 
2737 	ASSERT(MUTEX_HELD(&db->db_mtx));
2738 	DBUF_VERIFY(db);
2739 
2740 	/*
2741 	 * Remove the reference to the dbuf before removing its hold on the
2742 	 * dnode so we can guarantee in dnode_move() that a referenced bonus
2743 	 * buffer has a corresponding dnode hold.
2744 	 */
2745 	holds = refcount_remove(&db->db_holds, tag);
2746 	ASSERT(holds >= 0);
2747 
2748 	/*
2749 	 * We can't freeze indirects if there is a possibility that they
2750 	 * may be modified in the current syncing context.
2751 	 */
2752 	if (db->db_buf != NULL &&
2753 	    holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) {
2754 		arc_buf_freeze(db->db_buf);
2755 	}
2756 
2757 	if (holds == db->db_dirtycnt &&
2758 	    db->db_level == 0 && db->db_user_immediate_evict)
2759 		dbuf_evict_user(db);
2760 
2761 	if (holds == 0) {
2762 		if (db->db_blkid == DMU_BONUS_BLKID) {
2763 			dnode_t *dn;
2764 			boolean_t evict_dbuf = db->db_pending_evict;
2765 
2766 			/*
2767 			 * If the dnode moves here, we cannot cross this
2768 			 * barrier until the move completes.
2769 			 */
2770 			DB_DNODE_ENTER(db);
2771 
2772 			dn = DB_DNODE(db);
2773 			atomic_dec_32(&dn->dn_dbufs_count);
2774 
2775 			/*
2776 			 * Decrementing the dbuf count means that the bonus
2777 			 * buffer's dnode hold is no longer discounted in
2778 			 * dnode_move(). The dnode cannot move until after
2779 			 * the dnode_rele() below.
2780 			 */
2781 			DB_DNODE_EXIT(db);
2782 
2783 			/*
2784 			 * Do not reference db after its lock is dropped.
2785 			 * Another thread may evict it.
2786 			 */
2787 			mutex_exit(&db->db_mtx);
2788 
2789 			if (evict_dbuf)
2790 				dnode_evict_bonus(dn);
2791 
2792 			dnode_rele(dn, db);
2793 		} else if (db->db_buf == NULL) {
2794 			/*
2795 			 * This is a special case: we never associated this
2796 			 * dbuf with any data allocated from the ARC.
2797 			 */
2798 			ASSERT(db->db_state == DB_UNCACHED ||
2799 			    db->db_state == DB_NOFILL);
2800 			dbuf_destroy(db);
2801 		} else if (arc_released(db->db_buf)) {
2802 			/*
2803 			 * This dbuf has anonymous data associated with it.
2804 			 */
2805 			dbuf_destroy(db);
2806 		} else {
2807 			boolean_t do_arc_evict = B_FALSE;
2808 			blkptr_t bp;
2809 			spa_t *spa = dmu_objset_spa(db->db_objset);
2810 
2811 			if (!DBUF_IS_CACHEABLE(db) &&
2812 			    db->db_blkptr != NULL &&
2813 			    !BP_IS_HOLE(db->db_blkptr) &&
2814 			    !BP_IS_EMBEDDED(db->db_blkptr)) {
2815 				do_arc_evict = B_TRUE;
2816 				bp = *db->db_blkptr;
2817 			}
2818 
2819 			if (!DBUF_IS_CACHEABLE(db) ||
2820 			    db->db_pending_evict) {
2821 				dbuf_destroy(db);
2822 			} else if (!multilist_link_active(&db->db_cache_link)) {
2823 				multilist_insert(dbuf_cache, db);
2824 				(void) refcount_add_many(&dbuf_cache_size,
2825 				    db->db.db_size, db);
2826 				mutex_exit(&db->db_mtx);
2827 
2828 				dbuf_evict_notify();
2829 			}
2830 
2831 			if (do_arc_evict)
2832 				arc_freed(spa, &bp);
2833 		}
2834 	} else {
2835 		mutex_exit(&db->db_mtx);
2836 	}
2837 
2838 }
2839 
2840 #pragma weak dmu_buf_refcount = dbuf_refcount
2841 uint64_t
2842 dbuf_refcount(dmu_buf_impl_t *db)
2843 {
2844 	return (refcount_count(&db->db_holds));
2845 }
2846 
2847 void *
2848 dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user,
2849     dmu_buf_user_t *new_user)
2850 {
2851 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2852 
2853 	mutex_enter(&db->db_mtx);
2854 	dbuf_verify_user(db, DBVU_NOT_EVICTING);
2855 	if (db->db_user == old_user)
2856 		db->db_user = new_user;
2857 	else
2858 		old_user = db->db_user;
2859 	dbuf_verify_user(db, DBVU_NOT_EVICTING);
2860 	mutex_exit(&db->db_mtx);
2861 
2862 	return (old_user);
2863 }
2864 
2865 void *
2866 dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2867 {
2868 	return (dmu_buf_replace_user(db_fake, NULL, user));
2869 }
2870 
2871 void *
2872 dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2873 {
2874 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2875 
2876 	db->db_user_immediate_evict = TRUE;
2877 	return (dmu_buf_set_user(db_fake, user));
2878 }
2879 
2880 void *
2881 dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2882 {
2883 	return (dmu_buf_replace_user(db_fake, user, NULL));
2884 }
2885 
2886 void *
2887 dmu_buf_get_user(dmu_buf_t *db_fake)
2888 {
2889 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2890 
2891 	dbuf_verify_user(db, DBVU_NOT_EVICTING);
2892 	return (db->db_user);
2893 }
2894 
2895 void
2896 dmu_buf_user_evict_wait()
2897 {
2898 	taskq_wait(dbu_evict_taskq);
2899 }
2900 
2901