xref: /illumos-gate/usr/src/uts/common/fs/zfs/dbuf.c (revision adaec86a)
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 
51 uint_t zfs_dbuf_evict_key;
52 
53 static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
54 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
55 
56 #ifndef __lint
57 extern inline void dmu_buf_init_user(dmu_buf_user_t *dbu,
58     dmu_buf_evict_func_t *evict_func_sync,
59     dmu_buf_evict_func_t *evict_func_async,
60     dmu_buf_t **clear_on_evict_dbufp);
61 #endif /* ! __lint */
62 
63 /*
64  * Global data structures and functions for the dbuf cache.
65  */
66 static kmem_cache_t *dbuf_kmem_cache;
67 static taskq_t *dbu_evict_taskq;
68 
69 static kthread_t *dbuf_cache_evict_thread;
70 static kmutex_t dbuf_evict_lock;
71 static kcondvar_t dbuf_evict_cv;
72 static boolean_t dbuf_evict_thread_exit;
73 
74 /*
75  * LRU cache of dbufs. The dbuf cache maintains a list of dbufs that
76  * are not currently held but have been recently released. These dbufs
77  * are not eligible for arc eviction until they are aged out of the cache.
78  * Dbufs are added to the dbuf cache once the last hold is released. If a
79  * dbuf is later accessed and still exists in the dbuf cache, then it will
80  * be removed from the cache and later re-added to the head of the cache.
81  * Dbufs that are aged out of the cache will be immediately destroyed and
82  * become eligible for arc eviction.
83  */
84 static multilist_t *dbuf_cache;
85 static refcount_t dbuf_cache_size;
86 uint64_t dbuf_cache_max_bytes = 100 * 1024 * 1024;
87 
88 /* Cap the size of the dbuf cache to log2 fraction of arc size. */
89 int dbuf_cache_max_shift = 5;
90 
91 /*
92  * The dbuf cache uses a three-stage eviction policy:
93  *	- A low water marker designates when the dbuf eviction thread
94  *	should stop evicting from the dbuf cache.
95  *	- When we reach the maximum size (aka mid water mark), we
96  *	signal the eviction thread to run.
97  *	- The high water mark indicates when the eviction thread
98  *	is unable to keep up with the incoming load and eviction must
99  *	happen in the context of the calling thread.
100  *
101  * The dbuf cache:
102  *                                                 (max size)
103  *                                      low water   mid water   hi water
104  * +----------------------------------------+----------+----------+
105  * |                                        |          |          |
106  * |                                        |          |          |
107  * |                                        |          |          |
108  * |                                        |          |          |
109  * +----------------------------------------+----------+----------+
110  *                                        stop        signal     evict
111  *                                      evicting     eviction   directly
112  *                                                    thread
113  *
114  * The high and low water marks indicate the operating range for the eviction
115  * thread. The low water mark is, by default, 90% of the total size of the
116  * cache and the high water mark is at 110% (both of these percentages can be
117  * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct,
118  * respectively). The eviction thread will try to ensure that the cache remains
119  * within this range by waking up every second and checking if the cache is
120  * above the low water mark. The thread can also be woken up by callers adding
121  * elements into the cache if the cache is larger than the mid water (i.e max
122  * cache size). Once the eviction thread is woken up and eviction is required,
123  * it will continue evicting buffers until it's able to reduce the cache size
124  * to the low water mark. If the cache size continues to grow and hits the high
125  * water mark, then callers adding elments to the cache will begin to evict
126  * directly from the cache until the cache is no longer above the high water
127  * mark.
128  */
129 
130 /*
131  * The percentage above and below the maximum cache size.
132  */
133 uint_t dbuf_cache_hiwater_pct = 10;
134 uint_t dbuf_cache_lowater_pct = 10;
135 
136 /* ARGSUSED */
137 static int
138 dbuf_cons(void *vdb, void *unused, int kmflag)
139 {
140 	dmu_buf_impl_t *db = vdb;
141 	bzero(db, sizeof (dmu_buf_impl_t));
142 
143 	mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
144 	cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
145 	multilist_link_init(&db->db_cache_link);
146 	refcount_create(&db->db_holds);
147 
148 	return (0);
149 }
150 
151 /* ARGSUSED */
152 static void
153 dbuf_dest(void *vdb, void *unused)
154 {
155 	dmu_buf_impl_t *db = vdb;
156 	mutex_destroy(&db->db_mtx);
157 	cv_destroy(&db->db_changed);
158 	ASSERT(!multilist_link_active(&db->db_cache_link));
159 	refcount_destroy(&db->db_holds);
160 }
161 
162 /*
163  * dbuf hash table routines
164  */
165 static dbuf_hash_table_t dbuf_hash_table;
166 
167 static uint64_t dbuf_hash_count;
168 
169 static uint64_t
170 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
171 {
172 	uintptr_t osv = (uintptr_t)os;
173 	uint64_t crc = -1ULL;
174 
175 	ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
176 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF];
177 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF];
178 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF];
179 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF];
180 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF];
181 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF];
182 
183 	crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16);
184 
185 	return (crc);
186 }
187 
188 #define	DBUF_EQUAL(dbuf, os, obj, level, blkid)		\
189 	((dbuf)->db.db_object == (obj) &&		\
190 	(dbuf)->db_objset == (os) &&			\
191 	(dbuf)->db_level == (level) &&			\
192 	(dbuf)->db_blkid == (blkid))
193 
194 dmu_buf_impl_t *
195 dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid)
196 {
197 	dbuf_hash_table_t *h = &dbuf_hash_table;
198 	uint64_t hv = dbuf_hash(os, obj, level, blkid);
199 	uint64_t idx = hv & h->hash_table_mask;
200 	dmu_buf_impl_t *db;
201 
202 	mutex_enter(DBUF_HASH_MUTEX(h, idx));
203 	for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
204 		if (DBUF_EQUAL(db, os, obj, level, blkid)) {
205 			mutex_enter(&db->db_mtx);
206 			if (db->db_state != DB_EVICTING) {
207 				mutex_exit(DBUF_HASH_MUTEX(h, idx));
208 				return (db);
209 			}
210 			mutex_exit(&db->db_mtx);
211 		}
212 	}
213 	mutex_exit(DBUF_HASH_MUTEX(h, idx));
214 	return (NULL);
215 }
216 
217 static dmu_buf_impl_t *
218 dbuf_find_bonus(objset_t *os, uint64_t object)
219 {
220 	dnode_t *dn;
221 	dmu_buf_impl_t *db = NULL;
222 
223 	if (dnode_hold(os, object, FTAG, &dn) == 0) {
224 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
225 		if (dn->dn_bonus != NULL) {
226 			db = dn->dn_bonus;
227 			mutex_enter(&db->db_mtx);
228 		}
229 		rw_exit(&dn->dn_struct_rwlock);
230 		dnode_rele(dn, FTAG);
231 	}
232 	return (db);
233 }
234 
235 /*
236  * Insert an entry into the hash table.  If there is already an element
237  * equal to elem in the hash table, then the already existing element
238  * will be returned and the new element will not be inserted.
239  * Otherwise returns NULL.
240  */
241 static dmu_buf_impl_t *
242 dbuf_hash_insert(dmu_buf_impl_t *db)
243 {
244 	dbuf_hash_table_t *h = &dbuf_hash_table;
245 	objset_t *os = db->db_objset;
246 	uint64_t obj = db->db.db_object;
247 	int level = db->db_level;
248 	uint64_t blkid = db->db_blkid;
249 	uint64_t hv = dbuf_hash(os, obj, level, blkid);
250 	uint64_t idx = hv & h->hash_table_mask;
251 	dmu_buf_impl_t *dbf;
252 
253 	mutex_enter(DBUF_HASH_MUTEX(h, idx));
254 	for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) {
255 		if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
256 			mutex_enter(&dbf->db_mtx);
257 			if (dbf->db_state != DB_EVICTING) {
258 				mutex_exit(DBUF_HASH_MUTEX(h, idx));
259 				return (dbf);
260 			}
261 			mutex_exit(&dbf->db_mtx);
262 		}
263 	}
264 
265 	mutex_enter(&db->db_mtx);
266 	db->db_hash_next = h->hash_table[idx];
267 	h->hash_table[idx] = db;
268 	mutex_exit(DBUF_HASH_MUTEX(h, idx));
269 	atomic_inc_64(&dbuf_hash_count);
270 
271 	return (NULL);
272 }
273 
274 /*
275  * Remove an entry from the hash table.  It must be in the EVICTING state.
276  */
277 static void
278 dbuf_hash_remove(dmu_buf_impl_t *db)
279 {
280 	dbuf_hash_table_t *h = &dbuf_hash_table;
281 	uint64_t hv = dbuf_hash(db->db_objset, db->db.db_object,
282 	    db->db_level, db->db_blkid);
283 	uint64_t idx = hv & h->hash_table_mask;
284 	dmu_buf_impl_t *dbf, **dbp;
285 
286 	/*
287 	 * We musn't hold db_mtx to maintain lock ordering:
288 	 * DBUF_HASH_MUTEX > db_mtx.
289 	 */
290 	ASSERT(refcount_is_zero(&db->db_holds));
291 	ASSERT(db->db_state == DB_EVICTING);
292 	ASSERT(!MUTEX_HELD(&db->db_mtx));
293 
294 	mutex_enter(DBUF_HASH_MUTEX(h, idx));
295 	dbp = &h->hash_table[idx];
296 	while ((dbf = *dbp) != db) {
297 		dbp = &dbf->db_hash_next;
298 		ASSERT(dbf != NULL);
299 	}
300 	*dbp = db->db_hash_next;
301 	db->db_hash_next = NULL;
302 	mutex_exit(DBUF_HASH_MUTEX(h, idx));
303 	atomic_dec_64(&dbuf_hash_count);
304 }
305 
306 typedef enum {
307 	DBVU_EVICTING,
308 	DBVU_NOT_EVICTING
309 } dbvu_verify_type_t;
310 
311 static void
312 dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type)
313 {
314 #ifdef ZFS_DEBUG
315 	int64_t holds;
316 
317 	if (db->db_user == NULL)
318 		return;
319 
320 	/* Only data blocks support the attachment of user data. */
321 	ASSERT(db->db_level == 0);
322 
323 	/* Clients must resolve a dbuf before attaching user data. */
324 	ASSERT(db->db.db_data != NULL);
325 	ASSERT3U(db->db_state, ==, DB_CACHED);
326 
327 	holds = refcount_count(&db->db_holds);
328 	if (verify_type == DBVU_EVICTING) {
329 		/*
330 		 * Immediate eviction occurs when holds == dirtycnt.
331 		 * For normal eviction buffers, holds is zero on
332 		 * eviction, except when dbuf_fix_old_data() calls
333 		 * dbuf_clear_data().  However, the hold count can grow
334 		 * during eviction even though db_mtx is held (see
335 		 * dmu_bonus_hold() for an example), so we can only
336 		 * test the generic invariant that holds >= dirtycnt.
337 		 */
338 		ASSERT3U(holds, >=, db->db_dirtycnt);
339 	} else {
340 		if (db->db_user_immediate_evict == TRUE)
341 			ASSERT3U(holds, >=, db->db_dirtycnt);
342 		else
343 			ASSERT3U(holds, >, 0);
344 	}
345 #endif
346 }
347 
348 static void
349 dbuf_evict_user(dmu_buf_impl_t *db)
350 {
351 	dmu_buf_user_t *dbu = db->db_user;
352 
353 	ASSERT(MUTEX_HELD(&db->db_mtx));
354 
355 	if (dbu == NULL)
356 		return;
357 
358 	dbuf_verify_user(db, DBVU_EVICTING);
359 	db->db_user = NULL;
360 
361 #ifdef ZFS_DEBUG
362 	if (dbu->dbu_clear_on_evict_dbufp != NULL)
363 		*dbu->dbu_clear_on_evict_dbufp = NULL;
364 #endif
365 
366 	/*
367 	 * There are two eviction callbacks - one that we call synchronously
368 	 * and one that we invoke via a taskq.  The async one is useful for
369 	 * avoiding lock order reversals and limiting stack depth.
370 	 *
371 	 * Note that if we have a sync callback but no async callback,
372 	 * it's likely that the sync callback will free the structure
373 	 * containing the dbu.  In that case we need to take care to not
374 	 * dereference dbu after calling the sync evict func.
375 	 */
376 	boolean_t has_async = (dbu->dbu_evict_func_async != NULL);
377 
378 	if (dbu->dbu_evict_func_sync != NULL)
379 		dbu->dbu_evict_func_sync(dbu);
380 
381 	if (has_async) {
382 		taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func_async,
383 		    dbu, 0, &dbu->dbu_tqent);
384 	}
385 }
386 
387 boolean_t
388 dbuf_is_metadata(dmu_buf_impl_t *db)
389 {
390 	if (db->db_level > 0) {
391 		return (B_TRUE);
392 	} else {
393 		boolean_t is_metadata;
394 
395 		DB_DNODE_ENTER(db);
396 		is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type);
397 		DB_DNODE_EXIT(db);
398 
399 		return (is_metadata);
400 	}
401 }
402 
403 /*
404  * This function *must* return indices evenly distributed between all
405  * sublists of the multilist. This is needed due to how the dbuf eviction
406  * code is laid out; dbuf_evict_thread() assumes dbufs are evenly
407  * distributed between all sublists and uses this assumption when
408  * deciding which sublist to evict from and how much to evict from it.
409  */
410 unsigned int
411 dbuf_cache_multilist_index_func(multilist_t *ml, void *obj)
412 {
413 	dmu_buf_impl_t *db = obj;
414 
415 	/*
416 	 * The assumption here, is the hash value for a given
417 	 * dmu_buf_impl_t will remain constant throughout it's lifetime
418 	 * (i.e. it's objset, object, level and blkid fields don't change).
419 	 * Thus, we don't need to store the dbuf's sublist index
420 	 * on insertion, as this index can be recalculated on removal.
421 	 *
422 	 * Also, the low order bits of the hash value are thought to be
423 	 * distributed evenly. Otherwise, in the case that the multilist
424 	 * has a power of two number of sublists, each sublists' usage
425 	 * would not be evenly distributed.
426 	 */
427 	return (dbuf_hash(db->db_objset, db->db.db_object,
428 	    db->db_level, db->db_blkid) %
429 	    multilist_get_num_sublists(ml));
430 }
431 
432 static inline boolean_t
433 dbuf_cache_above_hiwater(void)
434 {
435 	uint64_t dbuf_cache_hiwater_bytes =
436 	    (dbuf_cache_max_bytes * dbuf_cache_hiwater_pct) / 100;
437 
438 	return (refcount_count(&dbuf_cache_size) >
439 	    dbuf_cache_max_bytes + dbuf_cache_hiwater_bytes);
440 }
441 
442 static inline boolean_t
443 dbuf_cache_above_lowater(void)
444 {
445 	uint64_t dbuf_cache_lowater_bytes =
446 	    (dbuf_cache_max_bytes * dbuf_cache_lowater_pct) / 100;
447 
448 	return (refcount_count(&dbuf_cache_size) >
449 	    dbuf_cache_max_bytes - dbuf_cache_lowater_bytes);
450 }
451 
452 /*
453  * Evict the oldest eligible dbuf from the dbuf cache.
454  */
455 static void
456 dbuf_evict_one(void)
457 {
458 	int idx = multilist_get_random_index(dbuf_cache);
459 	multilist_sublist_t *mls = multilist_sublist_lock(dbuf_cache, idx);
460 
461 	ASSERT(!MUTEX_HELD(&dbuf_evict_lock));
462 
463 	/*
464 	 * Set the thread's tsd to indicate that it's processing evictions.
465 	 * Once a thread stops evicting from the dbuf cache it will
466 	 * reset its tsd to NULL.
467 	 */
468 	ASSERT3P(tsd_get(zfs_dbuf_evict_key), ==, NULL);
469 	(void) tsd_set(zfs_dbuf_evict_key, (void *)B_TRUE);
470 
471 	dmu_buf_impl_t *db = multilist_sublist_tail(mls);
472 	while (db != NULL && mutex_tryenter(&db->db_mtx) == 0) {
473 		db = multilist_sublist_prev(mls, db);
474 	}
475 
476 	DTRACE_PROBE2(dbuf__evict__one, dmu_buf_impl_t *, db,
477 	    multilist_sublist_t *, mls);
478 
479 	if (db != NULL) {
480 		multilist_sublist_remove(mls, db);
481 		multilist_sublist_unlock(mls);
482 		(void) refcount_remove_many(&dbuf_cache_size,
483 		    db->db.db_size, db);
484 		dbuf_destroy(db);
485 	} else {
486 		multilist_sublist_unlock(mls);
487 	}
488 	(void) tsd_set(zfs_dbuf_evict_key, NULL);
489 }
490 
491 /*
492  * The dbuf evict thread is responsible for aging out dbufs from the
493  * cache. Once the cache has reached it's maximum size, dbufs are removed
494  * and destroyed. The eviction thread will continue running until the size
495  * of the dbuf cache is at or below the maximum size. Once the dbuf is aged
496  * out of the cache it is destroyed and becomes eligible for arc eviction.
497  */
498 static void
499 dbuf_evict_thread(void)
500 {
501 	callb_cpr_t cpr;
502 
503 	CALLB_CPR_INIT(&cpr, &dbuf_evict_lock, callb_generic_cpr, FTAG);
504 
505 	mutex_enter(&dbuf_evict_lock);
506 	while (!dbuf_evict_thread_exit) {
507 		while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
508 			CALLB_CPR_SAFE_BEGIN(&cpr);
509 			(void) cv_timedwait_hires(&dbuf_evict_cv,
510 			    &dbuf_evict_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
511 			CALLB_CPR_SAFE_END(&cpr, &dbuf_evict_lock);
512 		}
513 		mutex_exit(&dbuf_evict_lock);
514 
515 		/*
516 		 * Keep evicting as long as we're above the low water mark
517 		 * for the cache. We do this without holding the locks to
518 		 * minimize lock contention.
519 		 */
520 		while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
521 			dbuf_evict_one();
522 		}
523 
524 		mutex_enter(&dbuf_evict_lock);
525 	}
526 
527 	dbuf_evict_thread_exit = B_FALSE;
528 	cv_broadcast(&dbuf_evict_cv);
529 	CALLB_CPR_EXIT(&cpr);	/* drops dbuf_evict_lock */
530 	thread_exit();
531 }
532 
533 /*
534  * Wake up the dbuf eviction thread if the dbuf cache is at its max size.
535  * If the dbuf cache is at its high water mark, then evict a dbuf from the
536  * dbuf cache using the callers context.
537  */
538 static void
539 dbuf_evict_notify(void)
540 {
541 
542 	/*
543 	 * We use thread specific data to track when a thread has
544 	 * started processing evictions. This allows us to avoid deeply
545 	 * nested stacks that would have a call flow similar to this:
546 	 *
547 	 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify()
548 	 *	^						|
549 	 *	|						|
550 	 *	+-----dbuf_destroy()<--dbuf_evict_one()<--------+
551 	 *
552 	 * The dbuf_eviction_thread will always have its tsd set until
553 	 * that thread exits. All other threads will only set their tsd
554 	 * if they are participating in the eviction process. This only
555 	 * happens if the eviction thread is unable to process evictions
556 	 * fast enough. To keep the dbuf cache size in check, other threads
557 	 * can evict from the dbuf cache directly. Those threads will set
558 	 * their tsd values so that we ensure that they only evict one dbuf
559 	 * from the dbuf cache.
560 	 */
561 	if (tsd_get(zfs_dbuf_evict_key) != NULL)
562 		return;
563 
564 	if (refcount_count(&dbuf_cache_size) > dbuf_cache_max_bytes) {
565 		boolean_t evict_now = B_FALSE;
566 
567 		mutex_enter(&dbuf_evict_lock);
568 		if (refcount_count(&dbuf_cache_size) > dbuf_cache_max_bytes) {
569 			evict_now = dbuf_cache_above_hiwater();
570 			cv_signal(&dbuf_evict_cv);
571 		}
572 		mutex_exit(&dbuf_evict_lock);
573 
574 		if (evict_now) {
575 			dbuf_evict_one();
576 		}
577 	}
578 }
579 
580 void
581 dbuf_init(void)
582 {
583 	uint64_t hsize = 1ULL << 16;
584 	dbuf_hash_table_t *h = &dbuf_hash_table;
585 	int i;
586 
587 	/*
588 	 * The hash table is big enough to fill all of physical memory
589 	 * with an average 4K block size.  The table will take up
590 	 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
591 	 */
592 	while (hsize * 4096 < physmem * PAGESIZE)
593 		hsize <<= 1;
594 
595 retry:
596 	h->hash_table_mask = hsize - 1;
597 	h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
598 	if (h->hash_table == NULL) {
599 		/* XXX - we should really return an error instead of assert */
600 		ASSERT(hsize > (1ULL << 10));
601 		hsize >>= 1;
602 		goto retry;
603 	}
604 
605 	dbuf_kmem_cache = kmem_cache_create("dmu_buf_impl_t",
606 	    sizeof (dmu_buf_impl_t),
607 	    0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
608 
609 	for (i = 0; i < DBUF_MUTEXES; i++)
610 		mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
611 
612 	/*
613 	 * Setup the parameters for the dbuf cache. We cap the size of the
614 	 * dbuf cache to 1/32nd (default) of the size of the ARC.
615 	 */
616 	dbuf_cache_max_bytes = MIN(dbuf_cache_max_bytes,
617 	    arc_max_bytes() >> dbuf_cache_max_shift);
618 
619 	/*
620 	 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
621 	 * configuration is not required.
622 	 */
623 	dbu_evict_taskq = taskq_create("dbu_evict", 1, minclsyspri, 0, 0, 0);
624 
625 	dbuf_cache = multilist_create(sizeof (dmu_buf_impl_t),
626 	    offsetof(dmu_buf_impl_t, db_cache_link),
627 	    dbuf_cache_multilist_index_func);
628 	refcount_create(&dbuf_cache_size);
629 
630 	tsd_create(&zfs_dbuf_evict_key, NULL);
631 	dbuf_evict_thread_exit = B_FALSE;
632 	mutex_init(&dbuf_evict_lock, NULL, MUTEX_DEFAULT, NULL);
633 	cv_init(&dbuf_evict_cv, NULL, CV_DEFAULT, NULL);
634 	dbuf_cache_evict_thread = thread_create(NULL, 0, dbuf_evict_thread,
635 	    NULL, 0, &p0, TS_RUN, minclsyspri);
636 }
637 
638 void
639 dbuf_fini(void)
640 {
641 	dbuf_hash_table_t *h = &dbuf_hash_table;
642 	int i;
643 
644 	for (i = 0; i < DBUF_MUTEXES; i++)
645 		mutex_destroy(&h->hash_mutexes[i]);
646 	kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
647 	kmem_cache_destroy(dbuf_kmem_cache);
648 	taskq_destroy(dbu_evict_taskq);
649 
650 	mutex_enter(&dbuf_evict_lock);
651 	dbuf_evict_thread_exit = B_TRUE;
652 	while (dbuf_evict_thread_exit) {
653 		cv_signal(&dbuf_evict_cv);
654 		cv_wait(&dbuf_evict_cv, &dbuf_evict_lock);
655 	}
656 	mutex_exit(&dbuf_evict_lock);
657 	tsd_destroy(&zfs_dbuf_evict_key);
658 
659 	mutex_destroy(&dbuf_evict_lock);
660 	cv_destroy(&dbuf_evict_cv);
661 
662 	refcount_destroy(&dbuf_cache_size);
663 	multilist_destroy(dbuf_cache);
664 }
665 
666 /*
667  * Other stuff.
668  */
669 
670 #ifdef ZFS_DEBUG
671 static void
672 dbuf_verify(dmu_buf_impl_t *db)
673 {
674 	dnode_t *dn;
675 	dbuf_dirty_record_t *dr;
676 
677 	ASSERT(MUTEX_HELD(&db->db_mtx));
678 
679 	if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
680 		return;
681 
682 	ASSERT(db->db_objset != NULL);
683 	DB_DNODE_ENTER(db);
684 	dn = DB_DNODE(db);
685 	if (dn == NULL) {
686 		ASSERT(db->db_parent == NULL);
687 		ASSERT(db->db_blkptr == NULL);
688 	} else {
689 		ASSERT3U(db->db.db_object, ==, dn->dn_object);
690 		ASSERT3P(db->db_objset, ==, dn->dn_objset);
691 		ASSERT3U(db->db_level, <, dn->dn_nlevels);
692 		ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
693 		    db->db_blkid == DMU_SPILL_BLKID ||
694 		    !avl_is_empty(&dn->dn_dbufs));
695 	}
696 	if (db->db_blkid == DMU_BONUS_BLKID) {
697 		ASSERT(dn != NULL);
698 		ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
699 		ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
700 	} else if (db->db_blkid == DMU_SPILL_BLKID) {
701 		ASSERT(dn != NULL);
702 		ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
703 		ASSERT0(db->db.db_offset);
704 	} else {
705 		ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
706 	}
707 
708 	for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
709 		ASSERT(dr->dr_dbuf == db);
710 
711 	for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
712 		ASSERT(dr->dr_dbuf == db);
713 
714 	/*
715 	 * We can't assert that db_size matches dn_datablksz because it
716 	 * can be momentarily different when another thread is doing
717 	 * dnode_set_blksz().
718 	 */
719 	if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
720 		dr = db->db_data_pending;
721 		/*
722 		 * It should only be modified in syncing context, so
723 		 * make sure we only have one copy of the data.
724 		 */
725 		ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
726 	}
727 
728 	/* verify db->db_blkptr */
729 	if (db->db_blkptr) {
730 		if (db->db_parent == dn->dn_dbuf) {
731 			/* db is pointed to by the dnode */
732 			/* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
733 			if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
734 				ASSERT(db->db_parent == NULL);
735 			else
736 				ASSERT(db->db_parent != NULL);
737 			if (db->db_blkid != DMU_SPILL_BLKID)
738 				ASSERT3P(db->db_blkptr, ==,
739 				    &dn->dn_phys->dn_blkptr[db->db_blkid]);
740 		} else {
741 			/* db is pointed to by an indirect block */
742 			int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
743 			ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
744 			ASSERT3U(db->db_parent->db.db_object, ==,
745 			    db->db.db_object);
746 			/*
747 			 * dnode_grow_indblksz() can make this fail if we don't
748 			 * have the struct_rwlock.  XXX indblksz no longer
749 			 * grows.  safe to do this now?
750 			 */
751 			if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
752 				ASSERT3P(db->db_blkptr, ==,
753 				    ((blkptr_t *)db->db_parent->db.db_data +
754 				    db->db_blkid % epb));
755 			}
756 		}
757 	}
758 	if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
759 	    (db->db_buf == NULL || db->db_buf->b_data) &&
760 	    db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
761 	    db->db_state != DB_FILL && !dn->dn_free_txg) {
762 		/*
763 		 * If the blkptr isn't set but they have nonzero data,
764 		 * it had better be dirty, otherwise we'll lose that
765 		 * data when we evict this buffer.
766 		 *
767 		 * There is an exception to this rule for indirect blocks; in
768 		 * this case, if the indirect block is a hole, we fill in a few
769 		 * fields on each of the child blocks (importantly, birth time)
770 		 * to prevent hole birth times from being lost when you
771 		 * partially fill in a hole.
772 		 */
773 		if (db->db_dirtycnt == 0) {
774 			if (db->db_level == 0) {
775 				uint64_t *buf = db->db.db_data;
776 				int i;
777 
778 				for (i = 0; i < db->db.db_size >> 3; i++) {
779 					ASSERT(buf[i] == 0);
780 				}
781 			} else {
782 				blkptr_t *bps = db->db.db_data;
783 				ASSERT3U(1 << DB_DNODE(db)->dn_indblkshift, ==,
784 				    db->db.db_size);
785 				/*
786 				 * We want to verify that all the blkptrs in the
787 				 * indirect block are holes, but we may have
788 				 * automatically set up a few fields for them.
789 				 * We iterate through each blkptr and verify
790 				 * they only have those fields set.
791 				 */
792 				for (int i = 0;
793 				    i < db->db.db_size / sizeof (blkptr_t);
794 				    i++) {
795 					blkptr_t *bp = &bps[i];
796 					ASSERT(ZIO_CHECKSUM_IS_ZERO(
797 					    &bp->blk_cksum));
798 					ASSERT(
799 					    DVA_IS_EMPTY(&bp->blk_dva[0]) &&
800 					    DVA_IS_EMPTY(&bp->blk_dva[1]) &&
801 					    DVA_IS_EMPTY(&bp->blk_dva[2]));
802 					ASSERT0(bp->blk_fill);
803 					ASSERT0(bp->blk_pad[0]);
804 					ASSERT0(bp->blk_pad[1]);
805 					ASSERT(!BP_IS_EMBEDDED(bp));
806 					ASSERT(BP_IS_HOLE(bp));
807 					ASSERT0(bp->blk_phys_birth);
808 				}
809 			}
810 		}
811 	}
812 	DB_DNODE_EXIT(db);
813 }
814 #endif
815 
816 static void
817 dbuf_clear_data(dmu_buf_impl_t *db)
818 {
819 	ASSERT(MUTEX_HELD(&db->db_mtx));
820 	dbuf_evict_user(db);
821 	ASSERT3P(db->db_buf, ==, NULL);
822 	db->db.db_data = NULL;
823 	if (db->db_state != DB_NOFILL)
824 		db->db_state = DB_UNCACHED;
825 }
826 
827 static void
828 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
829 {
830 	ASSERT(MUTEX_HELD(&db->db_mtx));
831 	ASSERT(buf != NULL);
832 
833 	db->db_buf = buf;
834 	ASSERT(buf->b_data != NULL);
835 	db->db.db_data = buf->b_data;
836 }
837 
838 /*
839  * Loan out an arc_buf for read.  Return the loaned arc_buf.
840  */
841 arc_buf_t *
842 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
843 {
844 	arc_buf_t *abuf;
845 
846 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
847 	mutex_enter(&db->db_mtx);
848 	if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
849 		int blksz = db->db.db_size;
850 		spa_t *spa = db->db_objset->os_spa;
851 
852 		mutex_exit(&db->db_mtx);
853 		abuf = arc_loan_buf(spa, B_FALSE, blksz);
854 		bcopy(db->db.db_data, abuf->b_data, blksz);
855 	} else {
856 		abuf = db->db_buf;
857 		arc_loan_inuse_buf(abuf, db);
858 		db->db_buf = NULL;
859 		dbuf_clear_data(db);
860 		mutex_exit(&db->db_mtx);
861 	}
862 	return (abuf);
863 }
864 
865 /*
866  * Calculate which level n block references the data at the level 0 offset
867  * provided.
868  */
869 uint64_t
870 dbuf_whichblock(dnode_t *dn, int64_t level, uint64_t offset)
871 {
872 	if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) {
873 		/*
874 		 * The level n blkid is equal to the level 0 blkid divided by
875 		 * the number of level 0s in a level n block.
876 		 *
877 		 * The level 0 blkid is offset >> datablkshift =
878 		 * offset / 2^datablkshift.
879 		 *
880 		 * The number of level 0s in a level n is the number of block
881 		 * pointers in an indirect block, raised to the power of level.
882 		 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
883 		 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
884 		 *
885 		 * Thus, the level n blkid is: offset /
886 		 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
887 		 * = offset / 2^(datablkshift + level *
888 		 *   (indblkshift - SPA_BLKPTRSHIFT))
889 		 * = offset >> (datablkshift + level *
890 		 *   (indblkshift - SPA_BLKPTRSHIFT))
891 		 */
892 		return (offset >> (dn->dn_datablkshift + level *
893 		    (dn->dn_indblkshift - SPA_BLKPTRSHIFT)));
894 	} else {
895 		ASSERT3U(offset, <, dn->dn_datablksz);
896 		return (0);
897 	}
898 }
899 
900 static void
901 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
902 {
903 	dmu_buf_impl_t *db = vdb;
904 
905 	mutex_enter(&db->db_mtx);
906 	ASSERT3U(db->db_state, ==, DB_READ);
907 	/*
908 	 * All reads are synchronous, so we must have a hold on the dbuf
909 	 */
910 	ASSERT(refcount_count(&db->db_holds) > 0);
911 	ASSERT(db->db_buf == NULL);
912 	ASSERT(db->db.db_data == NULL);
913 	if (db->db_level == 0 && db->db_freed_in_flight) {
914 		/* we were freed in flight; disregard any error */
915 		arc_release(buf, db);
916 		bzero(buf->b_data, db->db.db_size);
917 		arc_buf_freeze(buf);
918 		db->db_freed_in_flight = FALSE;
919 		dbuf_set_data(db, buf);
920 		db->db_state = DB_CACHED;
921 	} else if (zio == NULL || zio->io_error == 0) {
922 		dbuf_set_data(db, buf);
923 		db->db_state = DB_CACHED;
924 	} else {
925 		ASSERT(db->db_blkid != DMU_BONUS_BLKID);
926 		ASSERT3P(db->db_buf, ==, NULL);
927 		arc_buf_destroy(buf, db);
928 		db->db_state = DB_UNCACHED;
929 	}
930 	cv_broadcast(&db->db_changed);
931 	dbuf_rele_and_unlock(db, NULL);
932 }
933 
934 static void
935 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
936 {
937 	dnode_t *dn;
938 	zbookmark_phys_t zb;
939 	arc_flags_t aflags = ARC_FLAG_NOWAIT;
940 
941 	DB_DNODE_ENTER(db);
942 	dn = DB_DNODE(db);
943 	ASSERT(!refcount_is_zero(&db->db_holds));
944 	/* We need the struct_rwlock to prevent db_blkptr from changing. */
945 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
946 	ASSERT(MUTEX_HELD(&db->db_mtx));
947 	ASSERT(db->db_state == DB_UNCACHED);
948 	ASSERT(db->db_buf == NULL);
949 
950 	if (db->db_blkid == DMU_BONUS_BLKID) {
951 		int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
952 
953 		ASSERT3U(bonuslen, <=, db->db.db_size);
954 		db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN);
955 		arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
956 		if (bonuslen < DN_MAX_BONUSLEN)
957 			bzero(db->db.db_data, DN_MAX_BONUSLEN);
958 		if (bonuslen)
959 			bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
960 		DB_DNODE_EXIT(db);
961 		db->db_state = DB_CACHED;
962 		mutex_exit(&db->db_mtx);
963 		return;
964 	}
965 
966 	/*
967 	 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
968 	 * processes the delete record and clears the bp while we are waiting
969 	 * for the dn_mtx (resulting in a "no" from block_freed).
970 	 */
971 	if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
972 	    (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
973 	    BP_IS_HOLE(db->db_blkptr)))) {
974 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
975 
976 		dbuf_set_data(db, arc_alloc_buf(db->db_objset->os_spa, db, type,
977 		    db->db.db_size));
978 		bzero(db->db.db_data, db->db.db_size);
979 
980 		if (db->db_blkptr != NULL && db->db_level > 0 &&
981 		    BP_IS_HOLE(db->db_blkptr) &&
982 		    db->db_blkptr->blk_birth != 0) {
983 			blkptr_t *bps = db->db.db_data;
984 			for (int i = 0; i < ((1 <<
985 			    DB_DNODE(db)->dn_indblkshift) / sizeof (blkptr_t));
986 			    i++) {
987 				blkptr_t *bp = &bps[i];
988 				ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
989 				    1 << dn->dn_indblkshift);
990 				BP_SET_LSIZE(bp,
991 				    BP_GET_LEVEL(db->db_blkptr) == 1 ?
992 				    dn->dn_datablksz :
993 				    BP_GET_LSIZE(db->db_blkptr));
994 				BP_SET_TYPE(bp, BP_GET_TYPE(db->db_blkptr));
995 				BP_SET_LEVEL(bp,
996 				    BP_GET_LEVEL(db->db_blkptr) - 1);
997 				BP_SET_BIRTH(bp, db->db_blkptr->blk_birth, 0);
998 			}
999 		}
1000 		DB_DNODE_EXIT(db);
1001 		db->db_state = DB_CACHED;
1002 		mutex_exit(&db->db_mtx);
1003 		return;
1004 	}
1005 
1006 	DB_DNODE_EXIT(db);
1007 
1008 	db->db_state = DB_READ;
1009 	mutex_exit(&db->db_mtx);
1010 
1011 	if (DBUF_IS_L2CACHEABLE(db))
1012 		aflags |= ARC_FLAG_L2CACHE;
1013 
1014 	SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
1015 	    db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
1016 	    db->db.db_object, db->db_level, db->db_blkid);
1017 
1018 	dbuf_add_ref(db, NULL);
1019 
1020 	(void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr,
1021 	    dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
1022 	    (flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
1023 	    &aflags, &zb);
1024 }
1025 
1026 /*
1027  * This is our just-in-time copy function.  It makes a copy of buffers that
1028  * have been modified in a previous transaction group before we access them in
1029  * the current active group.
1030  *
1031  * This function is used in three places: when we are dirtying a buffer for the
1032  * first time in a txg, when we are freeing a range in a dnode that includes
1033  * this buffer, and when we are accessing a buffer which was received compressed
1034  * and later referenced in a WRITE_BYREF record.
1035  *
1036  * Note that when we are called from dbuf_free_range() we do not put a hold on
1037  * the buffer, we just traverse the active dbuf list for the dnode.
1038  */
1039 static void
1040 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
1041 {
1042 	dbuf_dirty_record_t *dr = db->db_last_dirty;
1043 
1044 	ASSERT(MUTEX_HELD(&db->db_mtx));
1045 	ASSERT(db->db.db_data != NULL);
1046 	ASSERT(db->db_level == 0);
1047 	ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
1048 
1049 	if (dr == NULL ||
1050 	    (dr->dt.dl.dr_data !=
1051 	    ((db->db_blkid  == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
1052 		return;
1053 
1054 	/*
1055 	 * If the last dirty record for this dbuf has not yet synced
1056 	 * and its referencing the dbuf data, either:
1057 	 *	reset the reference to point to a new copy,
1058 	 * or (if there a no active holders)
1059 	 *	just null out the current db_data pointer.
1060 	 */
1061 	ASSERT(dr->dr_txg >= txg - 2);
1062 	if (db->db_blkid == DMU_BONUS_BLKID) {
1063 		/* Note that the data bufs here are zio_bufs */
1064 		dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
1065 		arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
1066 		bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
1067 	} else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
1068 		int size = arc_buf_size(db->db_buf);
1069 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1070 		spa_t *spa = db->db_objset->os_spa;
1071 		enum zio_compress compress_type =
1072 		    arc_get_compression(db->db_buf);
1073 
1074 		if (compress_type == ZIO_COMPRESS_OFF) {
1075 			dr->dt.dl.dr_data = arc_alloc_buf(spa, db, type, size);
1076 		} else {
1077 			ASSERT3U(type, ==, ARC_BUFC_DATA);
1078 			dr->dt.dl.dr_data = arc_alloc_compressed_buf(spa, db,
1079 			    size, arc_buf_lsize(db->db_buf), compress_type);
1080 		}
1081 		bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
1082 	} else {
1083 		db->db_buf = NULL;
1084 		dbuf_clear_data(db);
1085 	}
1086 }
1087 
1088 int
1089 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
1090 {
1091 	int err = 0;
1092 	boolean_t prefetch;
1093 	dnode_t *dn;
1094 
1095 	/*
1096 	 * We don't have to hold the mutex to check db_state because it
1097 	 * can't be freed while we have a hold on the buffer.
1098 	 */
1099 	ASSERT(!refcount_is_zero(&db->db_holds));
1100 
1101 	if (db->db_state == DB_NOFILL)
1102 		return (SET_ERROR(EIO));
1103 
1104 	DB_DNODE_ENTER(db);
1105 	dn = DB_DNODE(db);
1106 	if ((flags & DB_RF_HAVESTRUCT) == 0)
1107 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
1108 
1109 	prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1110 	    (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
1111 	    DBUF_IS_CACHEABLE(db);
1112 
1113 	mutex_enter(&db->db_mtx);
1114 	if (db->db_state == DB_CACHED) {
1115 		/*
1116 		 * If the arc buf is compressed, we need to decompress it to
1117 		 * read the data. This could happen during the "zfs receive" of
1118 		 * a stream which is compressed and deduplicated.
1119 		 */
1120 		if (db->db_buf != NULL &&
1121 		    arc_get_compression(db->db_buf) != ZIO_COMPRESS_OFF) {
1122 			dbuf_fix_old_data(db,
1123 			    spa_syncing_txg(dmu_objset_spa(db->db_objset)));
1124 			err = arc_decompress(db->db_buf);
1125 			dbuf_set_data(db, db->db_buf);
1126 		}
1127 		mutex_exit(&db->db_mtx);
1128 		if (prefetch)
1129 			dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1130 		if ((flags & DB_RF_HAVESTRUCT) == 0)
1131 			rw_exit(&dn->dn_struct_rwlock);
1132 		DB_DNODE_EXIT(db);
1133 	} else if (db->db_state == DB_UNCACHED) {
1134 		spa_t *spa = dn->dn_objset->os_spa;
1135 		boolean_t need_wait = B_FALSE;
1136 
1137 		if (zio == NULL &&
1138 		    db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) {
1139 			zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
1140 			need_wait = B_TRUE;
1141 		}
1142 		dbuf_read_impl(db, zio, flags);
1143 
1144 		/* dbuf_read_impl has dropped db_mtx for us */
1145 
1146 		if (prefetch)
1147 			dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1148 
1149 		if ((flags & DB_RF_HAVESTRUCT) == 0)
1150 			rw_exit(&dn->dn_struct_rwlock);
1151 		DB_DNODE_EXIT(db);
1152 
1153 		if (need_wait)
1154 			err = zio_wait(zio);
1155 	} else {
1156 		/*
1157 		 * Another reader came in while the dbuf was in flight
1158 		 * between UNCACHED and CACHED.  Either a writer will finish
1159 		 * writing the buffer (sending the dbuf to CACHED) or the
1160 		 * first reader's request will reach the read_done callback
1161 		 * and send the dbuf to CACHED.  Otherwise, a failure
1162 		 * occurred and the dbuf went to UNCACHED.
1163 		 */
1164 		mutex_exit(&db->db_mtx);
1165 		if (prefetch)
1166 			dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1167 		if ((flags & DB_RF_HAVESTRUCT) == 0)
1168 			rw_exit(&dn->dn_struct_rwlock);
1169 		DB_DNODE_EXIT(db);
1170 
1171 		/* Skip the wait per the caller's request. */
1172 		mutex_enter(&db->db_mtx);
1173 		if ((flags & DB_RF_NEVERWAIT) == 0) {
1174 			while (db->db_state == DB_READ ||
1175 			    db->db_state == DB_FILL) {
1176 				ASSERT(db->db_state == DB_READ ||
1177 				    (flags & DB_RF_HAVESTRUCT) == 0);
1178 				DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *,
1179 				    db, zio_t *, zio);
1180 				cv_wait(&db->db_changed, &db->db_mtx);
1181 			}
1182 			if (db->db_state == DB_UNCACHED)
1183 				err = SET_ERROR(EIO);
1184 		}
1185 		mutex_exit(&db->db_mtx);
1186 	}
1187 
1188 	return (err);
1189 }
1190 
1191 static void
1192 dbuf_noread(dmu_buf_impl_t *db)
1193 {
1194 	ASSERT(!refcount_is_zero(&db->db_holds));
1195 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1196 	mutex_enter(&db->db_mtx);
1197 	while (db->db_state == DB_READ || db->db_state == DB_FILL)
1198 		cv_wait(&db->db_changed, &db->db_mtx);
1199 	if (db->db_state == DB_UNCACHED) {
1200 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1201 		spa_t *spa = db->db_objset->os_spa;
1202 
1203 		ASSERT(db->db_buf == NULL);
1204 		ASSERT(db->db.db_data == NULL);
1205 		dbuf_set_data(db, arc_alloc_buf(spa, db, type, db->db.db_size));
1206 		db->db_state = DB_FILL;
1207 	} else if (db->db_state == DB_NOFILL) {
1208 		dbuf_clear_data(db);
1209 	} else {
1210 		ASSERT3U(db->db_state, ==, DB_CACHED);
1211 	}
1212 	mutex_exit(&db->db_mtx);
1213 }
1214 
1215 void
1216 dbuf_unoverride(dbuf_dirty_record_t *dr)
1217 {
1218 	dmu_buf_impl_t *db = dr->dr_dbuf;
1219 	blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
1220 	uint64_t txg = dr->dr_txg;
1221 
1222 	ASSERT(MUTEX_HELD(&db->db_mtx));
1223 	/*
1224 	 * This assert is valid because dmu_sync() expects to be called by
1225 	 * a zilog's get_data while holding a range lock.  This call only
1226 	 * comes from dbuf_dirty() callers who must also hold a range lock.
1227 	 */
1228 	ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
1229 	ASSERT(db->db_level == 0);
1230 
1231 	if (db->db_blkid == DMU_BONUS_BLKID ||
1232 	    dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
1233 		return;
1234 
1235 	ASSERT(db->db_data_pending != dr);
1236 
1237 	/* free this block */
1238 	if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite)
1239 		zio_free(db->db_objset->os_spa, txg, bp);
1240 
1241 	dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1242 	dr->dt.dl.dr_nopwrite = B_FALSE;
1243 
1244 	/*
1245 	 * Release the already-written buffer, so we leave it in
1246 	 * a consistent dirty state.  Note that all callers are
1247 	 * modifying the buffer, so they will immediately do
1248 	 * another (redundant) arc_release().  Therefore, leave
1249 	 * the buf thawed to save the effort of freezing &
1250 	 * immediately re-thawing it.
1251 	 */
1252 	arc_release(dr->dt.dl.dr_data, db);
1253 }
1254 
1255 /*
1256  * Evict (if its unreferenced) or clear (if its referenced) any level-0
1257  * data blocks in the free range, so that any future readers will find
1258  * empty blocks.
1259  */
1260 void
1261 dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
1262     dmu_tx_t *tx)
1263 {
1264 	dmu_buf_impl_t db_search;
1265 	dmu_buf_impl_t *db, *db_next;
1266 	uint64_t txg = tx->tx_txg;
1267 	avl_index_t where;
1268 
1269 	if (end_blkid > dn->dn_maxblkid &&
1270 	    !(start_blkid == DMU_SPILL_BLKID || end_blkid == DMU_SPILL_BLKID))
1271 		end_blkid = dn->dn_maxblkid;
1272 	dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid);
1273 
1274 	db_search.db_level = 0;
1275 	db_search.db_blkid = start_blkid;
1276 	db_search.db_state = DB_SEARCH;
1277 
1278 	mutex_enter(&dn->dn_dbufs_mtx);
1279 	db = avl_find(&dn->dn_dbufs, &db_search, &where);
1280 	ASSERT3P(db, ==, NULL);
1281 
1282 	db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1283 
1284 	for (; db != NULL; db = db_next) {
1285 		db_next = AVL_NEXT(&dn->dn_dbufs, db);
1286 		ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1287 
1288 		if (db->db_level != 0 || db->db_blkid > end_blkid) {
1289 			break;
1290 		}
1291 		ASSERT3U(db->db_blkid, >=, start_blkid);
1292 
1293 		/* found a level 0 buffer in the range */
1294 		mutex_enter(&db->db_mtx);
1295 		if (dbuf_undirty(db, tx)) {
1296 			/* mutex has been dropped and dbuf destroyed */
1297 			continue;
1298 		}
1299 
1300 		if (db->db_state == DB_UNCACHED ||
1301 		    db->db_state == DB_NOFILL ||
1302 		    db->db_state == DB_EVICTING) {
1303 			ASSERT(db->db.db_data == NULL);
1304 			mutex_exit(&db->db_mtx);
1305 			continue;
1306 		}
1307 		if (db->db_state == DB_READ || db->db_state == DB_FILL) {
1308 			/* will be handled in dbuf_read_done or dbuf_rele */
1309 			db->db_freed_in_flight = TRUE;
1310 			mutex_exit(&db->db_mtx);
1311 			continue;
1312 		}
1313 		if (refcount_count(&db->db_holds) == 0) {
1314 			ASSERT(db->db_buf);
1315 			dbuf_destroy(db);
1316 			continue;
1317 		}
1318 		/* The dbuf is referenced */
1319 
1320 		if (db->db_last_dirty != NULL) {
1321 			dbuf_dirty_record_t *dr = db->db_last_dirty;
1322 
1323 			if (dr->dr_txg == txg) {
1324 				/*
1325 				 * This buffer is "in-use", re-adjust the file
1326 				 * size to reflect that this buffer may
1327 				 * contain new data when we sync.
1328 				 */
1329 				if (db->db_blkid != DMU_SPILL_BLKID &&
1330 				    db->db_blkid > dn->dn_maxblkid)
1331 					dn->dn_maxblkid = db->db_blkid;
1332 				dbuf_unoverride(dr);
1333 			} else {
1334 				/*
1335 				 * This dbuf is not dirty in the open context.
1336 				 * Either uncache it (if its not referenced in
1337 				 * the open context) or reset its contents to
1338 				 * empty.
1339 				 */
1340 				dbuf_fix_old_data(db, txg);
1341 			}
1342 		}
1343 		/* clear the contents if its cached */
1344 		if (db->db_state == DB_CACHED) {
1345 			ASSERT(db->db.db_data != NULL);
1346 			arc_release(db->db_buf, db);
1347 			bzero(db->db.db_data, db->db.db_size);
1348 			arc_buf_freeze(db->db_buf);
1349 		}
1350 
1351 		mutex_exit(&db->db_mtx);
1352 	}
1353 	mutex_exit(&dn->dn_dbufs_mtx);
1354 }
1355 
1356 void
1357 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
1358 {
1359 	arc_buf_t *buf, *obuf;
1360 	int osize = db->db.db_size;
1361 	arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1362 	dnode_t *dn;
1363 
1364 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1365 
1366 	DB_DNODE_ENTER(db);
1367 	dn = DB_DNODE(db);
1368 
1369 	/* XXX does *this* func really need the lock? */
1370 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1371 
1372 	/*
1373 	 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1374 	 * is OK, because there can be no other references to the db
1375 	 * when we are changing its size, so no concurrent DB_FILL can
1376 	 * be happening.
1377 	 */
1378 	/*
1379 	 * XXX we should be doing a dbuf_read, checking the return
1380 	 * value and returning that up to our callers
1381 	 */
1382 	dmu_buf_will_dirty(&db->db, tx);
1383 
1384 	/* create the data buffer for the new block */
1385 	buf = arc_alloc_buf(dn->dn_objset->os_spa, db, type, size);
1386 
1387 	/* copy old block data to the new block */
1388 	obuf = db->db_buf;
1389 	bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
1390 	/* zero the remainder */
1391 	if (size > osize)
1392 		bzero((uint8_t *)buf->b_data + osize, size - osize);
1393 
1394 	mutex_enter(&db->db_mtx);
1395 	dbuf_set_data(db, buf);
1396 	arc_buf_destroy(obuf, db);
1397 	db->db.db_size = size;
1398 
1399 	if (db->db_level == 0) {
1400 		ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1401 		db->db_last_dirty->dt.dl.dr_data = buf;
1402 	}
1403 	mutex_exit(&db->db_mtx);
1404 
1405 	dmu_objset_willuse_space(dn->dn_objset, size - osize, tx);
1406 	DB_DNODE_EXIT(db);
1407 }
1408 
1409 void
1410 dbuf_release_bp(dmu_buf_impl_t *db)
1411 {
1412 	objset_t *os = db->db_objset;
1413 
1414 	ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1415 	ASSERT(arc_released(os->os_phys_buf) ||
1416 	    list_link_active(&os->os_dsl_dataset->ds_synced_link));
1417 	ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1418 
1419 	(void) arc_release(db->db_buf, db);
1420 }
1421 
1422 /*
1423  * We already have a dirty record for this TXG, and we are being
1424  * dirtied again.
1425  */
1426 static void
1427 dbuf_redirty(dbuf_dirty_record_t *dr)
1428 {
1429 	dmu_buf_impl_t *db = dr->dr_dbuf;
1430 
1431 	ASSERT(MUTEX_HELD(&db->db_mtx));
1432 
1433 	if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1434 		/*
1435 		 * If this buffer has already been written out,
1436 		 * we now need to reset its state.
1437 		 */
1438 		dbuf_unoverride(dr);
1439 		if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1440 		    db->db_state != DB_NOFILL) {
1441 			/* Already released on initial dirty, so just thaw. */
1442 			ASSERT(arc_released(db->db_buf));
1443 			arc_buf_thaw(db->db_buf);
1444 		}
1445 	}
1446 }
1447 
1448 dbuf_dirty_record_t *
1449 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1450 {
1451 	dnode_t *dn;
1452 	objset_t *os;
1453 	dbuf_dirty_record_t **drp, *dr;
1454 	int drop_struct_lock = FALSE;
1455 	int txgoff = tx->tx_txg & TXG_MASK;
1456 
1457 	ASSERT(tx->tx_txg != 0);
1458 	ASSERT(!refcount_is_zero(&db->db_holds));
1459 	DMU_TX_DIRTY_BUF(tx, db);
1460 
1461 	DB_DNODE_ENTER(db);
1462 	dn = DB_DNODE(db);
1463 	/*
1464 	 * Shouldn't dirty a regular buffer in syncing context.  Private
1465 	 * objects may be dirtied in syncing context, but only if they
1466 	 * were already pre-dirtied in open context.
1467 	 */
1468 #ifdef DEBUG
1469 	if (dn->dn_objset->os_dsl_dataset != NULL) {
1470 		rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1471 		    RW_READER, FTAG);
1472 	}
1473 	ASSERT(!dmu_tx_is_syncing(tx) ||
1474 	    BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1475 	    DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1476 	    dn->dn_objset->os_dsl_dataset == NULL);
1477 	if (dn->dn_objset->os_dsl_dataset != NULL)
1478 		rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, FTAG);
1479 #endif
1480 	/*
1481 	 * We make this assert for private objects as well, but after we
1482 	 * check if we're already dirty.  They are allowed to re-dirty
1483 	 * in syncing context.
1484 	 */
1485 	ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1486 	    dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1487 	    (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1488 
1489 	mutex_enter(&db->db_mtx);
1490 	/*
1491 	 * XXX make this true for indirects too?  The problem is that
1492 	 * transactions created with dmu_tx_create_assigned() from
1493 	 * syncing context don't bother holding ahead.
1494 	 */
1495 	ASSERT(db->db_level != 0 ||
1496 	    db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1497 	    db->db_state == DB_NOFILL);
1498 
1499 	mutex_enter(&dn->dn_mtx);
1500 	/*
1501 	 * Don't set dirtyctx to SYNC if we're just modifying this as we
1502 	 * initialize the objset.
1503 	 */
1504 	if (dn->dn_dirtyctx == DN_UNDIRTIED) {
1505 		if (dn->dn_objset->os_dsl_dataset != NULL) {
1506 			rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1507 			    RW_READER, FTAG);
1508 		}
1509 		if (!BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1510 			dn->dn_dirtyctx = (dmu_tx_is_syncing(tx) ?
1511 			    DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1512 			ASSERT(dn->dn_dirtyctx_firstset == NULL);
1513 			dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1514 		}
1515 		if (dn->dn_objset->os_dsl_dataset != NULL) {
1516 			rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1517 			    FTAG);
1518 		}
1519 	}
1520 	mutex_exit(&dn->dn_mtx);
1521 
1522 	if (db->db_blkid == DMU_SPILL_BLKID)
1523 		dn->dn_have_spill = B_TRUE;
1524 
1525 	/*
1526 	 * If this buffer is already dirty, we're done.
1527 	 */
1528 	drp = &db->db_last_dirty;
1529 	ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1530 	    db->db.db_object == DMU_META_DNODE_OBJECT);
1531 	while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1532 		drp = &dr->dr_next;
1533 	if (dr && dr->dr_txg == tx->tx_txg) {
1534 		DB_DNODE_EXIT(db);
1535 
1536 		dbuf_redirty(dr);
1537 		mutex_exit(&db->db_mtx);
1538 		return (dr);
1539 	}
1540 
1541 	/*
1542 	 * Only valid if not already dirty.
1543 	 */
1544 	ASSERT(dn->dn_object == 0 ||
1545 	    dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1546 	    (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1547 
1548 	ASSERT3U(dn->dn_nlevels, >, db->db_level);
1549 	ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1550 	    dn->dn_phys->dn_nlevels > db->db_level ||
1551 	    dn->dn_next_nlevels[txgoff] > db->db_level ||
1552 	    dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1553 	    dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1554 
1555 	/*
1556 	 * We should only be dirtying in syncing context if it's the
1557 	 * mos or we're initializing the os or it's a special object.
1558 	 * However, we are allowed to dirty in syncing context provided
1559 	 * we already dirtied it in open context.  Hence we must make
1560 	 * this assertion only if we're not already dirty.
1561 	 */
1562 	os = dn->dn_objset;
1563 	VERIFY3U(tx->tx_txg, <=, spa_final_dirty_txg(os->os_spa));
1564 #ifdef DEBUG
1565 	if (dn->dn_objset->os_dsl_dataset != NULL)
1566 		rrw_enter(&os->os_dsl_dataset->ds_bp_rwlock, RW_READER, FTAG);
1567 	ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1568 	    os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1569 	if (dn->dn_objset->os_dsl_dataset != NULL)
1570 		rrw_exit(&os->os_dsl_dataset->ds_bp_rwlock, FTAG);
1571 #endif
1572 	ASSERT(db->db.db_size != 0);
1573 
1574 	dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1575 
1576 	if (db->db_blkid != DMU_BONUS_BLKID) {
1577 		dmu_objset_willuse_space(os, db->db.db_size, tx);
1578 	}
1579 
1580 	/*
1581 	 * If this buffer is dirty in an old transaction group we need
1582 	 * to make a copy of it so that the changes we make in this
1583 	 * transaction group won't leak out when we sync the older txg.
1584 	 */
1585 	dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1586 	if (db->db_level == 0) {
1587 		void *data_old = db->db_buf;
1588 
1589 		if (db->db_state != DB_NOFILL) {
1590 			if (db->db_blkid == DMU_BONUS_BLKID) {
1591 				dbuf_fix_old_data(db, tx->tx_txg);
1592 				data_old = db->db.db_data;
1593 			} else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1594 				/*
1595 				 * Release the data buffer from the cache so
1596 				 * that we can modify it without impacting
1597 				 * possible other users of this cached data
1598 				 * block.  Note that indirect blocks and
1599 				 * private objects are not released until the
1600 				 * syncing state (since they are only modified
1601 				 * then).
1602 				 */
1603 				arc_release(db->db_buf, db);
1604 				dbuf_fix_old_data(db, tx->tx_txg);
1605 				data_old = db->db_buf;
1606 			}
1607 			ASSERT(data_old != NULL);
1608 		}
1609 		dr->dt.dl.dr_data = data_old;
1610 	} else {
1611 		mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1612 		list_create(&dr->dt.di.dr_children,
1613 		    sizeof (dbuf_dirty_record_t),
1614 		    offsetof(dbuf_dirty_record_t, dr_dirty_node));
1615 	}
1616 	if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL)
1617 		dr->dr_accounted = db->db.db_size;
1618 	dr->dr_dbuf = db;
1619 	dr->dr_txg = tx->tx_txg;
1620 	dr->dr_next = *drp;
1621 	*drp = dr;
1622 
1623 	/*
1624 	 * We could have been freed_in_flight between the dbuf_noread
1625 	 * and dbuf_dirty.  We win, as though the dbuf_noread() had
1626 	 * happened after the free.
1627 	 */
1628 	if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1629 	    db->db_blkid != DMU_SPILL_BLKID) {
1630 		mutex_enter(&dn->dn_mtx);
1631 		if (dn->dn_free_ranges[txgoff] != NULL) {
1632 			range_tree_clear(dn->dn_free_ranges[txgoff],
1633 			    db->db_blkid, 1);
1634 		}
1635 		mutex_exit(&dn->dn_mtx);
1636 		db->db_freed_in_flight = FALSE;
1637 	}
1638 
1639 	/*
1640 	 * This buffer is now part of this txg
1641 	 */
1642 	dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1643 	db->db_dirtycnt += 1;
1644 	ASSERT3U(db->db_dirtycnt, <=, 3);
1645 
1646 	mutex_exit(&db->db_mtx);
1647 
1648 	if (db->db_blkid == DMU_BONUS_BLKID ||
1649 	    db->db_blkid == DMU_SPILL_BLKID) {
1650 		mutex_enter(&dn->dn_mtx);
1651 		ASSERT(!list_link_active(&dr->dr_dirty_node));
1652 		list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1653 		mutex_exit(&dn->dn_mtx);
1654 		dnode_setdirty(dn, tx);
1655 		DB_DNODE_EXIT(db);
1656 		return (dr);
1657 	}
1658 
1659 	/*
1660 	 * The dn_struct_rwlock prevents db_blkptr from changing
1661 	 * due to a write from syncing context completing
1662 	 * while we are running, so we want to acquire it before
1663 	 * looking at db_blkptr.
1664 	 */
1665 	if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1666 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
1667 		drop_struct_lock = TRUE;
1668 	}
1669 
1670 	/*
1671 	 * If we are overwriting a dedup BP, then unless it is snapshotted,
1672 	 * when we get to syncing context we will need to decrement its
1673 	 * refcount in the DDT.  Prefetch the relevant DDT block so that
1674 	 * syncing context won't have to wait for the i/o.
1675 	 */
1676 	ddt_prefetch(os->os_spa, db->db_blkptr);
1677 
1678 	if (db->db_level == 0) {
1679 		dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1680 		ASSERT(dn->dn_maxblkid >= db->db_blkid);
1681 	}
1682 
1683 	if (db->db_level+1 < dn->dn_nlevels) {
1684 		dmu_buf_impl_t *parent = db->db_parent;
1685 		dbuf_dirty_record_t *di;
1686 		int parent_held = FALSE;
1687 
1688 		if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1689 			int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1690 
1691 			parent = dbuf_hold_level(dn, db->db_level+1,
1692 			    db->db_blkid >> epbs, FTAG);
1693 			ASSERT(parent != NULL);
1694 			parent_held = TRUE;
1695 		}
1696 		if (drop_struct_lock)
1697 			rw_exit(&dn->dn_struct_rwlock);
1698 		ASSERT3U(db->db_level+1, ==, parent->db_level);
1699 		di = dbuf_dirty(parent, tx);
1700 		if (parent_held)
1701 			dbuf_rele(parent, FTAG);
1702 
1703 		mutex_enter(&db->db_mtx);
1704 		/*
1705 		 * Since we've dropped the mutex, it's possible that
1706 		 * dbuf_undirty() might have changed this out from under us.
1707 		 */
1708 		if (db->db_last_dirty == dr ||
1709 		    dn->dn_object == DMU_META_DNODE_OBJECT) {
1710 			mutex_enter(&di->dt.di.dr_mtx);
1711 			ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1712 			ASSERT(!list_link_active(&dr->dr_dirty_node));
1713 			list_insert_tail(&di->dt.di.dr_children, dr);
1714 			mutex_exit(&di->dt.di.dr_mtx);
1715 			dr->dr_parent = di;
1716 		}
1717 		mutex_exit(&db->db_mtx);
1718 	} else {
1719 		ASSERT(db->db_level+1 == dn->dn_nlevels);
1720 		ASSERT(db->db_blkid < dn->dn_nblkptr);
1721 		ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1722 		mutex_enter(&dn->dn_mtx);
1723 		ASSERT(!list_link_active(&dr->dr_dirty_node));
1724 		list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1725 		mutex_exit(&dn->dn_mtx);
1726 		if (drop_struct_lock)
1727 			rw_exit(&dn->dn_struct_rwlock);
1728 	}
1729 
1730 	dnode_setdirty(dn, tx);
1731 	DB_DNODE_EXIT(db);
1732 	return (dr);
1733 }
1734 
1735 /*
1736  * Undirty a buffer in the transaction group referenced by the given
1737  * transaction.  Return whether this evicted the dbuf.
1738  */
1739 static boolean_t
1740 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1741 {
1742 	dnode_t *dn;
1743 	uint64_t txg = tx->tx_txg;
1744 	dbuf_dirty_record_t *dr, **drp;
1745 
1746 	ASSERT(txg != 0);
1747 
1748 	/*
1749 	 * Due to our use of dn_nlevels below, this can only be called
1750 	 * in open context, unless we are operating on the MOS.
1751 	 * From syncing context, dn_nlevels may be different from the
1752 	 * dn_nlevels used when dbuf was dirtied.
1753 	 */
1754 	ASSERT(db->db_objset ==
1755 	    dmu_objset_pool(db->db_objset)->dp_meta_objset ||
1756 	    txg != spa_syncing_txg(dmu_objset_spa(db->db_objset)));
1757 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1758 	ASSERT0(db->db_level);
1759 	ASSERT(MUTEX_HELD(&db->db_mtx));
1760 
1761 	/*
1762 	 * If this buffer is not dirty, we're done.
1763 	 */
1764 	for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1765 		if (dr->dr_txg <= txg)
1766 			break;
1767 	if (dr == NULL || dr->dr_txg < txg)
1768 		return (B_FALSE);
1769 	ASSERT(dr->dr_txg == txg);
1770 	ASSERT(dr->dr_dbuf == db);
1771 
1772 	DB_DNODE_ENTER(db);
1773 	dn = DB_DNODE(db);
1774 
1775 	dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1776 
1777 	ASSERT(db->db.db_size != 0);
1778 
1779 	dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset),
1780 	    dr->dr_accounted, txg);
1781 
1782 	*drp = dr->dr_next;
1783 
1784 	/*
1785 	 * Note that there are three places in dbuf_dirty()
1786 	 * where this dirty record may be put on a list.
1787 	 * Make sure to do a list_remove corresponding to
1788 	 * every one of those list_insert calls.
1789 	 */
1790 	if (dr->dr_parent) {
1791 		mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1792 		list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1793 		mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1794 	} else if (db->db_blkid == DMU_SPILL_BLKID ||
1795 	    db->db_level + 1 == dn->dn_nlevels) {
1796 		ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1797 		mutex_enter(&dn->dn_mtx);
1798 		list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1799 		mutex_exit(&dn->dn_mtx);
1800 	}
1801 	DB_DNODE_EXIT(db);
1802 
1803 	if (db->db_state != DB_NOFILL) {
1804 		dbuf_unoverride(dr);
1805 
1806 		ASSERT(db->db_buf != NULL);
1807 		ASSERT(dr->dt.dl.dr_data != NULL);
1808 		if (dr->dt.dl.dr_data != db->db_buf)
1809 			arc_buf_destroy(dr->dt.dl.dr_data, db);
1810 	}
1811 
1812 	kmem_free(dr, sizeof (dbuf_dirty_record_t));
1813 
1814 	ASSERT(db->db_dirtycnt > 0);
1815 	db->db_dirtycnt -= 1;
1816 
1817 	if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1818 		ASSERT(db->db_state == DB_NOFILL || arc_released(db->db_buf));
1819 		dbuf_destroy(db);
1820 		return (B_TRUE);
1821 	}
1822 
1823 	return (B_FALSE);
1824 }
1825 
1826 void
1827 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
1828 {
1829 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1830 	int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1831 
1832 	ASSERT(tx->tx_txg != 0);
1833 	ASSERT(!refcount_is_zero(&db->db_holds));
1834 
1835 	/*
1836 	 * Quick check for dirtyness.  For already dirty blocks, this
1837 	 * reduces runtime of this function by >90%, and overall performance
1838 	 * by 50% for some workloads (e.g. file deletion with indirect blocks
1839 	 * cached).
1840 	 */
1841 	mutex_enter(&db->db_mtx);
1842 	dbuf_dirty_record_t *dr;
1843 	for (dr = db->db_last_dirty;
1844 	    dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) {
1845 		/*
1846 		 * It's possible that it is already dirty but not cached,
1847 		 * because there are some calls to dbuf_dirty() that don't
1848 		 * go through dmu_buf_will_dirty().
1849 		 */
1850 		if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) {
1851 			/* This dbuf is already dirty and cached. */
1852 			dbuf_redirty(dr);
1853 			mutex_exit(&db->db_mtx);
1854 			return;
1855 		}
1856 	}
1857 	mutex_exit(&db->db_mtx);
1858 
1859 	DB_DNODE_ENTER(db);
1860 	if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
1861 		rf |= DB_RF_HAVESTRUCT;
1862 	DB_DNODE_EXIT(db);
1863 	(void) dbuf_read(db, NULL, rf);
1864 	(void) dbuf_dirty(db, tx);
1865 }
1866 
1867 void
1868 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1869 {
1870 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1871 
1872 	db->db_state = DB_NOFILL;
1873 
1874 	dmu_buf_will_fill(db_fake, tx);
1875 }
1876 
1877 void
1878 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1879 {
1880 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1881 
1882 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1883 	ASSERT(tx->tx_txg != 0);
1884 	ASSERT(db->db_level == 0);
1885 	ASSERT(!refcount_is_zero(&db->db_holds));
1886 
1887 	ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1888 	    dmu_tx_private_ok(tx));
1889 
1890 	dbuf_noread(db);
1891 	(void) dbuf_dirty(db, tx);
1892 }
1893 
1894 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1895 /* ARGSUSED */
1896 void
1897 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1898 {
1899 	mutex_enter(&db->db_mtx);
1900 	DBUF_VERIFY(db);
1901 
1902 	if (db->db_state == DB_FILL) {
1903 		if (db->db_level == 0 && db->db_freed_in_flight) {
1904 			ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1905 			/* we were freed while filling */
1906 			/* XXX dbuf_undirty? */
1907 			bzero(db->db.db_data, db->db.db_size);
1908 			db->db_freed_in_flight = FALSE;
1909 		}
1910 		db->db_state = DB_CACHED;
1911 		cv_broadcast(&db->db_changed);
1912 	}
1913 	mutex_exit(&db->db_mtx);
1914 }
1915 
1916 void
1917 dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data,
1918     bp_embedded_type_t etype, enum zio_compress comp,
1919     int uncompressed_size, int compressed_size, int byteorder,
1920     dmu_tx_t *tx)
1921 {
1922 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
1923 	struct dirty_leaf *dl;
1924 	dmu_object_type_t type;
1925 
1926 	if (etype == BP_EMBEDDED_TYPE_DATA) {
1927 		ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset),
1928 		    SPA_FEATURE_EMBEDDED_DATA));
1929 	}
1930 
1931 	DB_DNODE_ENTER(db);
1932 	type = DB_DNODE(db)->dn_type;
1933 	DB_DNODE_EXIT(db);
1934 
1935 	ASSERT0(db->db_level);
1936 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1937 
1938 	dmu_buf_will_not_fill(dbuf, tx);
1939 
1940 	ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1941 	dl = &db->db_last_dirty->dt.dl;
1942 	encode_embedded_bp_compressed(&dl->dr_overridden_by,
1943 	    data, comp, uncompressed_size, compressed_size);
1944 	BPE_SET_ETYPE(&dl->dr_overridden_by, etype);
1945 	BP_SET_TYPE(&dl->dr_overridden_by, type);
1946 	BP_SET_LEVEL(&dl->dr_overridden_by, 0);
1947 	BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder);
1948 
1949 	dl->dr_override_state = DR_OVERRIDDEN;
1950 	dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg;
1951 }
1952 
1953 /*
1954  * Directly assign a provided arc buf to a given dbuf if it's not referenced
1955  * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1956  */
1957 void
1958 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
1959 {
1960 	ASSERT(!refcount_is_zero(&db->db_holds));
1961 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1962 	ASSERT(db->db_level == 0);
1963 	ASSERT3U(dbuf_is_metadata(db), ==, arc_is_metadata(buf));
1964 	ASSERT(buf != NULL);
1965 	ASSERT(arc_buf_lsize(buf) == db->db.db_size);
1966 	ASSERT(tx->tx_txg != 0);
1967 
1968 	arc_return_buf(buf, db);
1969 	ASSERT(arc_released(buf));
1970 
1971 	mutex_enter(&db->db_mtx);
1972 
1973 	while (db->db_state == DB_READ || db->db_state == DB_FILL)
1974 		cv_wait(&db->db_changed, &db->db_mtx);
1975 
1976 	ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
1977 
1978 	if (db->db_state == DB_CACHED &&
1979 	    refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
1980 		mutex_exit(&db->db_mtx);
1981 		(void) dbuf_dirty(db, tx);
1982 		bcopy(buf->b_data, db->db.db_data, db->db.db_size);
1983 		arc_buf_destroy(buf, db);
1984 		xuio_stat_wbuf_copied();
1985 		return;
1986 	}
1987 
1988 	xuio_stat_wbuf_nocopy();
1989 	if (db->db_state == DB_CACHED) {
1990 		dbuf_dirty_record_t *dr = db->db_last_dirty;
1991 
1992 		ASSERT(db->db_buf != NULL);
1993 		if (dr != NULL && dr->dr_txg == tx->tx_txg) {
1994 			ASSERT(dr->dt.dl.dr_data == db->db_buf);
1995 			if (!arc_released(db->db_buf)) {
1996 				ASSERT(dr->dt.dl.dr_override_state ==
1997 				    DR_OVERRIDDEN);
1998 				arc_release(db->db_buf, db);
1999 			}
2000 			dr->dt.dl.dr_data = buf;
2001 			arc_buf_destroy(db->db_buf, db);
2002 		} else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
2003 			arc_release(db->db_buf, db);
2004 			arc_buf_destroy(db->db_buf, db);
2005 		}
2006 		db->db_buf = NULL;
2007 	}
2008 	ASSERT(db->db_buf == NULL);
2009 	dbuf_set_data(db, buf);
2010 	db->db_state = DB_FILL;
2011 	mutex_exit(&db->db_mtx);
2012 	(void) dbuf_dirty(db, tx);
2013 	dmu_buf_fill_done(&db->db, tx);
2014 }
2015 
2016 void
2017 dbuf_destroy(dmu_buf_impl_t *db)
2018 {
2019 	dnode_t *dn;
2020 	dmu_buf_impl_t *parent = db->db_parent;
2021 	dmu_buf_impl_t *dndb;
2022 
2023 	ASSERT(MUTEX_HELD(&db->db_mtx));
2024 	ASSERT(refcount_is_zero(&db->db_holds));
2025 
2026 	if (db->db_buf != NULL) {
2027 		arc_buf_destroy(db->db_buf, db);
2028 		db->db_buf = NULL;
2029 	}
2030 
2031 	if (db->db_blkid == DMU_BONUS_BLKID) {
2032 		ASSERT(db->db.db_data != NULL);
2033 		zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
2034 		arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
2035 		db->db_state = DB_UNCACHED;
2036 	}
2037 
2038 	dbuf_clear_data(db);
2039 
2040 	if (multilist_link_active(&db->db_cache_link)) {
2041 		multilist_remove(dbuf_cache, db);
2042 		(void) refcount_remove_many(&dbuf_cache_size,
2043 		    db->db.db_size, db);
2044 	}
2045 
2046 	ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
2047 	ASSERT(db->db_data_pending == NULL);
2048 
2049 	db->db_state = DB_EVICTING;
2050 	db->db_blkptr = NULL;
2051 
2052 	/*
2053 	 * Now that db_state is DB_EVICTING, nobody else can find this via
2054 	 * the hash table.  We can now drop db_mtx, which allows us to
2055 	 * acquire the dn_dbufs_mtx.
2056 	 */
2057 	mutex_exit(&db->db_mtx);
2058 
2059 	DB_DNODE_ENTER(db);
2060 	dn = DB_DNODE(db);
2061 	dndb = dn->dn_dbuf;
2062 	if (db->db_blkid != DMU_BONUS_BLKID) {
2063 		boolean_t needlock = !MUTEX_HELD(&dn->dn_dbufs_mtx);
2064 		if (needlock)
2065 			mutex_enter(&dn->dn_dbufs_mtx);
2066 		avl_remove(&dn->dn_dbufs, db);
2067 		atomic_dec_32(&dn->dn_dbufs_count);
2068 		membar_producer();
2069 		DB_DNODE_EXIT(db);
2070 		if (needlock)
2071 			mutex_exit(&dn->dn_dbufs_mtx);
2072 		/*
2073 		 * Decrementing the dbuf count means that the hold corresponding
2074 		 * to the removed dbuf is no longer discounted in dnode_move(),
2075 		 * so the dnode cannot be moved until after we release the hold.
2076 		 * The membar_producer() ensures visibility of the decremented
2077 		 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
2078 		 * release any lock.
2079 		 */
2080 		dnode_rele(dn, db);
2081 		db->db_dnode_handle = NULL;
2082 
2083 		dbuf_hash_remove(db);
2084 	} else {
2085 		DB_DNODE_EXIT(db);
2086 	}
2087 
2088 	ASSERT(refcount_is_zero(&db->db_holds));
2089 
2090 	db->db_parent = NULL;
2091 
2092 	ASSERT(db->db_buf == NULL);
2093 	ASSERT(db->db.db_data == NULL);
2094 	ASSERT(db->db_hash_next == NULL);
2095 	ASSERT(db->db_blkptr == NULL);
2096 	ASSERT(db->db_data_pending == NULL);
2097 	ASSERT(!multilist_link_active(&db->db_cache_link));
2098 
2099 	kmem_cache_free(dbuf_kmem_cache, db);
2100 	arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2101 
2102 	/*
2103 	 * If this dbuf is referenced from an indirect dbuf,
2104 	 * decrement the ref count on the indirect dbuf.
2105 	 */
2106 	if (parent && parent != dndb)
2107 		dbuf_rele(parent, db);
2108 }
2109 
2110 /*
2111  * Note: While bpp will always be updated if the function returns success,
2112  * parentp will not be updated if the dnode does not have dn_dbuf filled in;
2113  * this happens when the dnode is the meta-dnode, or a userused or groupused
2114  * object.
2115  */
2116 static int
2117 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
2118     dmu_buf_impl_t **parentp, blkptr_t **bpp)
2119 {
2120 	*parentp = NULL;
2121 	*bpp = NULL;
2122 
2123 	ASSERT(blkid != DMU_BONUS_BLKID);
2124 
2125 	if (blkid == DMU_SPILL_BLKID) {
2126 		mutex_enter(&dn->dn_mtx);
2127 		if (dn->dn_have_spill &&
2128 		    (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
2129 			*bpp = &dn->dn_phys->dn_spill;
2130 		else
2131 			*bpp = NULL;
2132 		dbuf_add_ref(dn->dn_dbuf, NULL);
2133 		*parentp = dn->dn_dbuf;
2134 		mutex_exit(&dn->dn_mtx);
2135 		return (0);
2136 	}
2137 
2138 	int nlevels =
2139 	    (dn->dn_phys->dn_nlevels == 0) ? 1 : dn->dn_phys->dn_nlevels;
2140 	int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2141 
2142 	ASSERT3U(level * epbs, <, 64);
2143 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2144 	/*
2145 	 * This assertion shouldn't trip as long as the max indirect block size
2146 	 * is less than 1M.  The reason for this is that up to that point,
2147 	 * the number of levels required to address an entire object with blocks
2148 	 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64.  In
2149 	 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55
2150 	 * (i.e. we can address the entire object), objects will all use at most
2151 	 * N-1 levels and the assertion won't overflow.  However, once epbs is
2152 	 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66.  Then, 4 levels will not be
2153 	 * enough to address an entire object, so objects will have 5 levels,
2154 	 * but then this assertion will overflow.
2155 	 *
2156 	 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we
2157 	 * need to redo this logic to handle overflows.
2158 	 */
2159 	ASSERT(level >= nlevels ||
2160 	    ((nlevels - level - 1) * epbs) +
2161 	    highbit64(dn->dn_phys->dn_nblkptr) <= 64);
2162 	if (level >= nlevels ||
2163 	    blkid >= ((uint64_t)dn->dn_phys->dn_nblkptr <<
2164 	    ((nlevels - level - 1) * epbs)) ||
2165 	    (fail_sparse &&
2166 	    blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
2167 		/* the buffer has no parent yet */
2168 		return (SET_ERROR(ENOENT));
2169 	} else if (level < nlevels-1) {
2170 		/* this block is referenced from an indirect block */
2171 		int err = dbuf_hold_impl(dn, level+1,
2172 		    blkid >> epbs, fail_sparse, FALSE, NULL, parentp);
2173 		if (err)
2174 			return (err);
2175 		err = dbuf_read(*parentp, NULL,
2176 		    (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
2177 		if (err) {
2178 			dbuf_rele(*parentp, NULL);
2179 			*parentp = NULL;
2180 			return (err);
2181 		}
2182 		*bpp = ((blkptr_t *)(*parentp)->db.db_data) +
2183 		    (blkid & ((1ULL << epbs) - 1));
2184 		if (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))
2185 			ASSERT(BP_IS_HOLE(*bpp));
2186 		return (0);
2187 	} else {
2188 		/* the block is referenced from the dnode */
2189 		ASSERT3U(level, ==, nlevels-1);
2190 		ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
2191 		    blkid < dn->dn_phys->dn_nblkptr);
2192 		if (dn->dn_dbuf) {
2193 			dbuf_add_ref(dn->dn_dbuf, NULL);
2194 			*parentp = dn->dn_dbuf;
2195 		}
2196 		*bpp = &dn->dn_phys->dn_blkptr[blkid];
2197 		return (0);
2198 	}
2199 }
2200 
2201 static dmu_buf_impl_t *
2202 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
2203     dmu_buf_impl_t *parent, blkptr_t *blkptr)
2204 {
2205 	objset_t *os = dn->dn_objset;
2206 	dmu_buf_impl_t *db, *odb;
2207 
2208 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2209 	ASSERT(dn->dn_type != DMU_OT_NONE);
2210 
2211 	db = kmem_cache_alloc(dbuf_kmem_cache, KM_SLEEP);
2212 
2213 	db->db_objset = os;
2214 	db->db.db_object = dn->dn_object;
2215 	db->db_level = level;
2216 	db->db_blkid = blkid;
2217 	db->db_last_dirty = NULL;
2218 	db->db_dirtycnt = 0;
2219 	db->db_dnode_handle = dn->dn_handle;
2220 	db->db_parent = parent;
2221 	db->db_blkptr = blkptr;
2222 
2223 	db->db_user = NULL;
2224 	db->db_user_immediate_evict = FALSE;
2225 	db->db_freed_in_flight = FALSE;
2226 	db->db_pending_evict = FALSE;
2227 
2228 	if (blkid == DMU_BONUS_BLKID) {
2229 		ASSERT3P(parent, ==, dn->dn_dbuf);
2230 		db->db.db_size = DN_MAX_BONUSLEN -
2231 		    (dn->dn_nblkptr-1) * sizeof (blkptr_t);
2232 		ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
2233 		db->db.db_offset = DMU_BONUS_BLKID;
2234 		db->db_state = DB_UNCACHED;
2235 		/* the bonus dbuf is not placed in the hash table */
2236 		arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2237 		return (db);
2238 	} else if (blkid == DMU_SPILL_BLKID) {
2239 		db->db.db_size = (blkptr != NULL) ?
2240 		    BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
2241 		db->db.db_offset = 0;
2242 	} else {
2243 		int blocksize =
2244 		    db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
2245 		db->db.db_size = blocksize;
2246 		db->db.db_offset = db->db_blkid * blocksize;
2247 	}
2248 
2249 	/*
2250 	 * Hold the dn_dbufs_mtx while we get the new dbuf
2251 	 * in the hash table *and* added to the dbufs list.
2252 	 * This prevents a possible deadlock with someone
2253 	 * trying to look up this dbuf before its added to the
2254 	 * dn_dbufs list.
2255 	 */
2256 	mutex_enter(&dn->dn_dbufs_mtx);
2257 	db->db_state = DB_EVICTING;
2258 	if ((odb = dbuf_hash_insert(db)) != NULL) {
2259 		/* someone else inserted it first */
2260 		kmem_cache_free(dbuf_kmem_cache, db);
2261 		mutex_exit(&dn->dn_dbufs_mtx);
2262 		return (odb);
2263 	}
2264 	avl_add(&dn->dn_dbufs, db);
2265 
2266 	db->db_state = DB_UNCACHED;
2267 	mutex_exit(&dn->dn_dbufs_mtx);
2268 	arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2269 
2270 	if (parent && parent != dn->dn_dbuf)
2271 		dbuf_add_ref(parent, db);
2272 
2273 	ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
2274 	    refcount_count(&dn->dn_holds) > 0);
2275 	(void) refcount_add(&dn->dn_holds, db);
2276 	atomic_inc_32(&dn->dn_dbufs_count);
2277 
2278 	dprintf_dbuf(db, "db=%p\n", db);
2279 
2280 	return (db);
2281 }
2282 
2283 typedef struct dbuf_prefetch_arg {
2284 	spa_t *dpa_spa;	/* The spa to issue the prefetch in. */
2285 	zbookmark_phys_t dpa_zb; /* The target block to prefetch. */
2286 	int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */
2287 	int dpa_curlevel; /* The current level that we're reading */
2288 	dnode_t *dpa_dnode; /* The dnode associated with the prefetch */
2289 	zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */
2290 	zio_t *dpa_zio; /* The parent zio_t for all prefetches. */
2291 	arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */
2292 } dbuf_prefetch_arg_t;
2293 
2294 /*
2295  * Actually issue the prefetch read for the block given.
2296  */
2297 static void
2298 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp)
2299 {
2300 	if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
2301 		return;
2302 
2303 	arc_flags_t aflags =
2304 	    dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
2305 
2306 	ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2307 	ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level);
2308 	ASSERT(dpa->dpa_zio != NULL);
2309 	(void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL,
2310 	    dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2311 	    &aflags, &dpa->dpa_zb);
2312 }
2313 
2314 /*
2315  * Called when an indirect block above our prefetch target is read in.  This
2316  * will either read in the next indirect block down the tree or issue the actual
2317  * prefetch if the next block down is our target.
2318  */
2319 static void
2320 dbuf_prefetch_indirect_done(zio_t *zio, arc_buf_t *abuf, void *private)
2321 {
2322 	dbuf_prefetch_arg_t *dpa = private;
2323 
2324 	ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel);
2325 	ASSERT3S(dpa->dpa_curlevel, >, 0);
2326 
2327 	/*
2328 	 * The dpa_dnode is only valid if we are called with a NULL
2329 	 * zio. This indicates that the arc_read() returned without
2330 	 * first calling zio_read() to issue a physical read. Once
2331 	 * a physical read is made the dpa_dnode must be invalidated
2332 	 * as the locks guarding it may have been dropped. If the
2333 	 * dpa_dnode is still valid, then we want to add it to the dbuf
2334 	 * cache. To do so, we must hold the dbuf associated with the block
2335 	 * we just prefetched, read its contents so that we associate it
2336 	 * with an arc_buf_t, and then release it.
2337 	 */
2338 	if (zio != NULL) {
2339 		ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel);
2340 		if (zio->io_flags & ZIO_FLAG_RAW) {
2341 			ASSERT3U(BP_GET_PSIZE(zio->io_bp), ==, zio->io_size);
2342 		} else {
2343 			ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size);
2344 		}
2345 		ASSERT3P(zio->io_spa, ==, dpa->dpa_spa);
2346 
2347 		dpa->dpa_dnode = NULL;
2348 	} else if (dpa->dpa_dnode != NULL) {
2349 		uint64_t curblkid = dpa->dpa_zb.zb_blkid >>
2350 		    (dpa->dpa_epbs * (dpa->dpa_curlevel -
2351 		    dpa->dpa_zb.zb_level));
2352 		dmu_buf_impl_t *db = dbuf_hold_level(dpa->dpa_dnode,
2353 		    dpa->dpa_curlevel, curblkid, FTAG);
2354 		(void) dbuf_read(db, NULL,
2355 		    DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_HAVESTRUCT);
2356 		dbuf_rele(db, FTAG);
2357 	}
2358 
2359 	dpa->dpa_curlevel--;
2360 
2361 	uint64_t nextblkid = dpa->dpa_zb.zb_blkid >>
2362 	    (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level));
2363 	blkptr_t *bp = ((blkptr_t *)abuf->b_data) +
2364 	    P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs);
2365 	if (BP_IS_HOLE(bp) || (zio != NULL && zio->io_error != 0)) {
2366 		kmem_free(dpa, sizeof (*dpa));
2367 	} else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) {
2368 		ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid);
2369 		dbuf_issue_final_prefetch(dpa, bp);
2370 		kmem_free(dpa, sizeof (*dpa));
2371 	} else {
2372 		arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2373 		zbookmark_phys_t zb;
2374 
2375 		ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2376 
2377 		SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset,
2378 		    dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid);
2379 
2380 		(void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2381 		    bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio,
2382 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2383 		    &iter_aflags, &zb);
2384 	}
2385 
2386 	arc_buf_destroy(abuf, private);
2387 }
2388 
2389 /*
2390  * Issue prefetch reads for the given block on the given level.  If the indirect
2391  * blocks above that block are not in memory, we will read them in
2392  * asynchronously.  As a result, this call never blocks waiting for a read to
2393  * complete.
2394  */
2395 void
2396 dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio,
2397     arc_flags_t aflags)
2398 {
2399 	blkptr_t bp;
2400 	int epbs, nlevels, curlevel;
2401 	uint64_t curblkid;
2402 
2403 	ASSERT(blkid != DMU_BONUS_BLKID);
2404 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2405 
2406 	if (blkid > dn->dn_maxblkid)
2407 		return;
2408 
2409 	if (dnode_block_freed(dn, blkid))
2410 		return;
2411 
2412 	/*
2413 	 * This dnode hasn't been written to disk yet, so there's nothing to
2414 	 * prefetch.
2415 	 */
2416 	nlevels = dn->dn_phys->dn_nlevels;
2417 	if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0)
2418 		return;
2419 
2420 	epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2421 	if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level))
2422 		return;
2423 
2424 	dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object,
2425 	    level, blkid);
2426 	if (db != NULL) {
2427 		mutex_exit(&db->db_mtx);
2428 		/*
2429 		 * This dbuf already exists.  It is either CACHED, or
2430 		 * (we assume) about to be read or filled.
2431 		 */
2432 		return;
2433 	}
2434 
2435 	/*
2436 	 * Find the closest ancestor (indirect block) of the target block
2437 	 * that is present in the cache.  In this indirect block, we will
2438 	 * find the bp that is at curlevel, curblkid.
2439 	 */
2440 	curlevel = level;
2441 	curblkid = blkid;
2442 	while (curlevel < nlevels - 1) {
2443 		int parent_level = curlevel + 1;
2444 		uint64_t parent_blkid = curblkid >> epbs;
2445 		dmu_buf_impl_t *db;
2446 
2447 		if (dbuf_hold_impl(dn, parent_level, parent_blkid,
2448 		    FALSE, TRUE, FTAG, &db) == 0) {
2449 			blkptr_t *bpp = db->db_buf->b_data;
2450 			bp = bpp[P2PHASE(curblkid, 1 << epbs)];
2451 			dbuf_rele(db, FTAG);
2452 			break;
2453 		}
2454 
2455 		curlevel = parent_level;
2456 		curblkid = parent_blkid;
2457 	}
2458 
2459 	if (curlevel == nlevels - 1) {
2460 		/* No cached indirect blocks found. */
2461 		ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr);
2462 		bp = dn->dn_phys->dn_blkptr[curblkid];
2463 	}
2464 	if (BP_IS_HOLE(&bp))
2465 		return;
2466 
2467 	ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp));
2468 
2469 	zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL,
2470 	    ZIO_FLAG_CANFAIL);
2471 
2472 	dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP);
2473 	dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
2474 	SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2475 	    dn->dn_object, level, blkid);
2476 	dpa->dpa_curlevel = curlevel;
2477 	dpa->dpa_prio = prio;
2478 	dpa->dpa_aflags = aflags;
2479 	dpa->dpa_spa = dn->dn_objset->os_spa;
2480 	dpa->dpa_dnode = dn;
2481 	dpa->dpa_epbs = epbs;
2482 	dpa->dpa_zio = pio;
2483 
2484 	/*
2485 	 * If we have the indirect just above us, no need to do the asynchronous
2486 	 * prefetch chain; we'll just run the last step ourselves.  If we're at
2487 	 * a higher level, though, we want to issue the prefetches for all the
2488 	 * indirect blocks asynchronously, so we can go on with whatever we were
2489 	 * doing.
2490 	 */
2491 	if (curlevel == level) {
2492 		ASSERT3U(curblkid, ==, blkid);
2493 		dbuf_issue_final_prefetch(dpa, &bp);
2494 		kmem_free(dpa, sizeof (*dpa));
2495 	} else {
2496 		arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2497 		zbookmark_phys_t zb;
2498 
2499 		SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2500 		    dn->dn_object, curlevel, curblkid);
2501 		(void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2502 		    &bp, dbuf_prefetch_indirect_done, dpa, prio,
2503 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2504 		    &iter_aflags, &zb);
2505 	}
2506 	/*
2507 	 * We use pio here instead of dpa_zio since it's possible that
2508 	 * dpa may have already been freed.
2509 	 */
2510 	zio_nowait(pio);
2511 }
2512 
2513 /*
2514  * Returns with db_holds incremented, and db_mtx not held.
2515  * Note: dn_struct_rwlock must be held.
2516  */
2517 int
2518 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid,
2519     boolean_t fail_sparse, boolean_t fail_uncached,
2520     void *tag, dmu_buf_impl_t **dbp)
2521 {
2522 	dmu_buf_impl_t *db, *parent = NULL;
2523 
2524 	ASSERT(blkid != DMU_BONUS_BLKID);
2525 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2526 	ASSERT3U(dn->dn_nlevels, >, level);
2527 
2528 	*dbp = NULL;
2529 top:
2530 	/* dbuf_find() returns with db_mtx held */
2531 	db = dbuf_find(dn->dn_objset, dn->dn_object, level, blkid);
2532 
2533 	if (db == NULL) {
2534 		blkptr_t *bp = NULL;
2535 		int err;
2536 
2537 		if (fail_uncached)
2538 			return (SET_ERROR(ENOENT));
2539 
2540 		ASSERT3P(parent, ==, NULL);
2541 		err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
2542 		if (fail_sparse) {
2543 			if (err == 0 && bp && BP_IS_HOLE(bp))
2544 				err = SET_ERROR(ENOENT);
2545 			if (err) {
2546 				if (parent)
2547 					dbuf_rele(parent, NULL);
2548 				return (err);
2549 			}
2550 		}
2551 		if (err && err != ENOENT)
2552 			return (err);
2553 		db = dbuf_create(dn, level, blkid, parent, bp);
2554 	}
2555 
2556 	if (fail_uncached && db->db_state != DB_CACHED) {
2557 		mutex_exit(&db->db_mtx);
2558 		return (SET_ERROR(ENOENT));
2559 	}
2560 
2561 	if (db->db_buf != NULL)
2562 		ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
2563 
2564 	ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
2565 
2566 	/*
2567 	 * If this buffer is currently syncing out, and we are are
2568 	 * still referencing it from db_data, we need to make a copy
2569 	 * of it in case we decide we want to dirty it again in this txg.
2570 	 */
2571 	if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
2572 	    dn->dn_object != DMU_META_DNODE_OBJECT &&
2573 	    db->db_state == DB_CACHED && db->db_data_pending) {
2574 		dbuf_dirty_record_t *dr = db->db_data_pending;
2575 
2576 		if (dr->dt.dl.dr_data == db->db_buf) {
2577 			arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2578 
2579 			dbuf_set_data(db,
2580 			    arc_alloc_buf(dn->dn_objset->os_spa, db, type,
2581 			    db->db.db_size));
2582 			bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
2583 			    db->db.db_size);
2584 		}
2585 	}
2586 
2587 	if (multilist_link_active(&db->db_cache_link)) {
2588 		ASSERT(refcount_is_zero(&db->db_holds));
2589 		multilist_remove(dbuf_cache, db);
2590 		(void) refcount_remove_many(&dbuf_cache_size,
2591 		    db->db.db_size, db);
2592 	}
2593 	(void) refcount_add(&db->db_holds, tag);
2594 	DBUF_VERIFY(db);
2595 	mutex_exit(&db->db_mtx);
2596 
2597 	/* NOTE: we can't rele the parent until after we drop the db_mtx */
2598 	if (parent)
2599 		dbuf_rele(parent, NULL);
2600 
2601 	ASSERT3P(DB_DNODE(db), ==, dn);
2602 	ASSERT3U(db->db_blkid, ==, blkid);
2603 	ASSERT3U(db->db_level, ==, level);
2604 	*dbp = db;
2605 
2606 	return (0);
2607 }
2608 
2609 dmu_buf_impl_t *
2610 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
2611 {
2612 	return (dbuf_hold_level(dn, 0, blkid, tag));
2613 }
2614 
2615 dmu_buf_impl_t *
2616 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
2617 {
2618 	dmu_buf_impl_t *db;
2619 	int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db);
2620 	return (err ? NULL : db);
2621 }
2622 
2623 void
2624 dbuf_create_bonus(dnode_t *dn)
2625 {
2626 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
2627 
2628 	ASSERT(dn->dn_bonus == NULL);
2629 	dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
2630 }
2631 
2632 int
2633 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
2634 {
2635 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2636 	dnode_t *dn;
2637 
2638 	if (db->db_blkid != DMU_SPILL_BLKID)
2639 		return (SET_ERROR(ENOTSUP));
2640 	if (blksz == 0)
2641 		blksz = SPA_MINBLOCKSIZE;
2642 	ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset)));
2643 	blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
2644 
2645 	DB_DNODE_ENTER(db);
2646 	dn = DB_DNODE(db);
2647 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2648 	dbuf_new_size(db, blksz, tx);
2649 	rw_exit(&dn->dn_struct_rwlock);
2650 	DB_DNODE_EXIT(db);
2651 
2652 	return (0);
2653 }
2654 
2655 void
2656 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
2657 {
2658 	dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
2659 }
2660 
2661 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2662 void
2663 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
2664 {
2665 	int64_t holds = refcount_add(&db->db_holds, tag);
2666 	ASSERT3S(holds, >, 1);
2667 }
2668 
2669 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
2670 boolean_t
2671 dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid,
2672     void *tag)
2673 {
2674 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2675 	dmu_buf_impl_t *found_db;
2676 	boolean_t result = B_FALSE;
2677 
2678 	if (db->db_blkid == DMU_BONUS_BLKID)
2679 		found_db = dbuf_find_bonus(os, obj);
2680 	else
2681 		found_db = dbuf_find(os, obj, 0, blkid);
2682 
2683 	if (found_db != NULL) {
2684 		if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) {
2685 			(void) refcount_add(&db->db_holds, tag);
2686 			result = B_TRUE;
2687 		}
2688 		mutex_exit(&db->db_mtx);
2689 	}
2690 	return (result);
2691 }
2692 
2693 /*
2694  * If you call dbuf_rele() you had better not be referencing the dnode handle
2695  * unless you have some other direct or indirect hold on the dnode. (An indirect
2696  * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2697  * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2698  * dnode's parent dbuf evicting its dnode handles.
2699  */
2700 void
2701 dbuf_rele(dmu_buf_impl_t *db, void *tag)
2702 {
2703 	mutex_enter(&db->db_mtx);
2704 	dbuf_rele_and_unlock(db, tag);
2705 }
2706 
2707 void
2708 dmu_buf_rele(dmu_buf_t *db, void *tag)
2709 {
2710 	dbuf_rele((dmu_buf_impl_t *)db, tag);
2711 }
2712 
2713 /*
2714  * dbuf_rele() for an already-locked dbuf.  This is necessary to allow
2715  * db_dirtycnt and db_holds to be updated atomically.
2716  */
2717 void
2718 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2719 {
2720 	int64_t holds;
2721 
2722 	ASSERT(MUTEX_HELD(&db->db_mtx));
2723 	DBUF_VERIFY(db);
2724 
2725 	/*
2726 	 * Remove the reference to the dbuf before removing its hold on the
2727 	 * dnode so we can guarantee in dnode_move() that a referenced bonus
2728 	 * buffer has a corresponding dnode hold.
2729 	 */
2730 	holds = refcount_remove(&db->db_holds, tag);
2731 	ASSERT(holds >= 0);
2732 
2733 	/*
2734 	 * We can't freeze indirects if there is a possibility that they
2735 	 * may be modified in the current syncing context.
2736 	 */
2737 	if (db->db_buf != NULL &&
2738 	    holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) {
2739 		arc_buf_freeze(db->db_buf);
2740 	}
2741 
2742 	if (holds == db->db_dirtycnt &&
2743 	    db->db_level == 0 && db->db_user_immediate_evict)
2744 		dbuf_evict_user(db);
2745 
2746 	if (holds == 0) {
2747 		if (db->db_blkid == DMU_BONUS_BLKID) {
2748 			dnode_t *dn;
2749 			boolean_t evict_dbuf = db->db_pending_evict;
2750 
2751 			/*
2752 			 * If the dnode moves here, we cannot cross this
2753 			 * barrier until the move completes.
2754 			 */
2755 			DB_DNODE_ENTER(db);
2756 
2757 			dn = DB_DNODE(db);
2758 			atomic_dec_32(&dn->dn_dbufs_count);
2759 
2760 			/*
2761 			 * Decrementing the dbuf count means that the bonus
2762 			 * buffer's dnode hold is no longer discounted in
2763 			 * dnode_move(). The dnode cannot move until after
2764 			 * the dnode_rele() below.
2765 			 */
2766 			DB_DNODE_EXIT(db);
2767 
2768 			/*
2769 			 * Do not reference db after its lock is dropped.
2770 			 * Another thread may evict it.
2771 			 */
2772 			mutex_exit(&db->db_mtx);
2773 
2774 			if (evict_dbuf)
2775 				dnode_evict_bonus(dn);
2776 
2777 			dnode_rele(dn, db);
2778 		} else if (db->db_buf == NULL) {
2779 			/*
2780 			 * This is a special case: we never associated this
2781 			 * dbuf with any data allocated from the ARC.
2782 			 */
2783 			ASSERT(db->db_state == DB_UNCACHED ||
2784 			    db->db_state == DB_NOFILL);
2785 			dbuf_destroy(db);
2786 		} else if (arc_released(db->db_buf)) {
2787 			/*
2788 			 * This dbuf has anonymous data associated with it.
2789 			 */
2790 			dbuf_destroy(db);
2791 		} else {
2792 			boolean_t do_arc_evict = B_FALSE;
2793 			blkptr_t bp;
2794 			spa_t *spa = dmu_objset_spa(db->db_objset);
2795 
2796 			if (!DBUF_IS_CACHEABLE(db) &&
2797 			    db->db_blkptr != NULL &&
2798 			    !BP_IS_HOLE(db->db_blkptr) &&
2799 			    !BP_IS_EMBEDDED(db->db_blkptr)) {
2800 				do_arc_evict = B_TRUE;
2801 				bp = *db->db_blkptr;
2802 			}
2803 
2804 			if (!DBUF_IS_CACHEABLE(db) ||
2805 			    db->db_pending_evict) {
2806 				dbuf_destroy(db);
2807 			} else if (!multilist_link_active(&db->db_cache_link)) {
2808 				multilist_insert(dbuf_cache, db);
2809 				(void) refcount_add_many(&dbuf_cache_size,
2810 				    db->db.db_size, db);
2811 				mutex_exit(&db->db_mtx);
2812 
2813 				dbuf_evict_notify();
2814 			}
2815 
2816 			if (do_arc_evict)
2817 				arc_freed(spa, &bp);
2818 		}
2819 	} else {
2820 		mutex_exit(&db->db_mtx);
2821 	}
2822 
2823 }
2824 
2825 #pragma weak dmu_buf_refcount = dbuf_refcount
2826 uint64_t
2827 dbuf_refcount(dmu_buf_impl_t *db)
2828 {
2829 	return (refcount_count(&db->db_holds));
2830 }
2831 
2832 void *
2833 dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user,
2834     dmu_buf_user_t *new_user)
2835 {
2836 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2837 
2838 	mutex_enter(&db->db_mtx);
2839 	dbuf_verify_user(db, DBVU_NOT_EVICTING);
2840 	if (db->db_user == old_user)
2841 		db->db_user = new_user;
2842 	else
2843 		old_user = db->db_user;
2844 	dbuf_verify_user(db, DBVU_NOT_EVICTING);
2845 	mutex_exit(&db->db_mtx);
2846 
2847 	return (old_user);
2848 }
2849 
2850 void *
2851 dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2852 {
2853 	return (dmu_buf_replace_user(db_fake, NULL, user));
2854 }
2855 
2856 void *
2857 dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2858 {
2859 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2860 
2861 	db->db_user_immediate_evict = TRUE;
2862 	return (dmu_buf_set_user(db_fake, user));
2863 }
2864 
2865 void *
2866 dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2867 {
2868 	return (dmu_buf_replace_user(db_fake, user, NULL));
2869 }
2870 
2871 void *
2872 dmu_buf_get_user(dmu_buf_t *db_fake)
2873 {
2874 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2875 
2876 	dbuf_verify_user(db, DBVU_NOT_EVICTING);
2877 	return (db->db_user);
2878 }
2879 
2880 void
2881 dmu_buf_user_evict_wait()
2882 {
2883 	taskq_wait(dbu_evict_taskq);
2884 }
2885 
2886 blkptr_t *
2887 dmu_buf_get_blkptr(dmu_buf_t *db)
2888 {
2889 	dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2890 	return (dbi->db_blkptr);
2891 }
2892 
2893 objset_t *
2894 dmu_buf_get_objset(dmu_buf_t *db)
2895<