xref: /illumos-gate/usr/src/uts/common/fs/zfs/dbuf.c (revision 1702cce751c5cb7ead878d0205a6c90b027e3de8)
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 	/*
565 	 * We check if we should evict without holding the dbuf_evict_lock,
566 	 * because it's OK to occasionally make the wrong decision here,
567 	 * and grabbing the lock results in massive lock contention.
568 	 */
569 	if (refcount_count(&dbuf_cache_size) > dbuf_cache_max_bytes) {
570 		if (dbuf_cache_above_hiwater())
571 			dbuf_evict_one();
572 		cv_signal(&dbuf_evict_cv);
573 	}
574 }
575 
576 void
577 dbuf_init(void)
578 {
579 	uint64_t hsize = 1ULL << 16;
580 	dbuf_hash_table_t *h = &dbuf_hash_table;
581 	int i;
582 
583 	/*
584 	 * The hash table is big enough to fill all of physical memory
585 	 * with an average 4K block size.  The table will take up
586 	 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
587 	 */
588 	while (hsize * 4096 < physmem * PAGESIZE)
589 		hsize <<= 1;
590 
591 retry:
592 	h->hash_table_mask = hsize - 1;
593 	h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
594 	if (h->hash_table == NULL) {
595 		/* XXX - we should really return an error instead of assert */
596 		ASSERT(hsize > (1ULL << 10));
597 		hsize >>= 1;
598 		goto retry;
599 	}
600 
601 	dbuf_kmem_cache = kmem_cache_create("dmu_buf_impl_t",
602 	    sizeof (dmu_buf_impl_t),
603 	    0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
604 
605 	for (i = 0; i < DBUF_MUTEXES; i++)
606 		mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
607 
608 	/*
609 	 * Setup the parameters for the dbuf cache. We cap the size of the
610 	 * dbuf cache to 1/32nd (default) of the size of the ARC.
611 	 */
612 	dbuf_cache_max_bytes = MIN(dbuf_cache_max_bytes,
613 	    arc_max_bytes() >> dbuf_cache_max_shift);
614 
615 	/*
616 	 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
617 	 * configuration is not required.
618 	 */
619 	dbu_evict_taskq = taskq_create("dbu_evict", 1, minclsyspri, 0, 0, 0);
620 
621 	dbuf_cache = multilist_create(sizeof (dmu_buf_impl_t),
622 	    offsetof(dmu_buf_impl_t, db_cache_link),
623 	    dbuf_cache_multilist_index_func);
624 	refcount_create(&dbuf_cache_size);
625 
626 	tsd_create(&zfs_dbuf_evict_key, NULL);
627 	dbuf_evict_thread_exit = B_FALSE;
628 	mutex_init(&dbuf_evict_lock, NULL, MUTEX_DEFAULT, NULL);
629 	cv_init(&dbuf_evict_cv, NULL, CV_DEFAULT, NULL);
630 	dbuf_cache_evict_thread = thread_create(NULL, 0, dbuf_evict_thread,
631 	    NULL, 0, &p0, TS_RUN, minclsyspri);
632 }
633 
634 void
635 dbuf_fini(void)
636 {
637 	dbuf_hash_table_t *h = &dbuf_hash_table;
638 	int i;
639 
640 	for (i = 0; i < DBUF_MUTEXES; i++)
641 		mutex_destroy(&h->hash_mutexes[i]);
642 	kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
643 	kmem_cache_destroy(dbuf_kmem_cache);
644 	taskq_destroy(dbu_evict_taskq);
645 
646 	mutex_enter(&dbuf_evict_lock);
647 	dbuf_evict_thread_exit = B_TRUE;
648 	while (dbuf_evict_thread_exit) {
649 		cv_signal(&dbuf_evict_cv);
650 		cv_wait(&dbuf_evict_cv, &dbuf_evict_lock);
651 	}
652 	mutex_exit(&dbuf_evict_lock);
653 	tsd_destroy(&zfs_dbuf_evict_key);
654 
655 	mutex_destroy(&dbuf_evict_lock);
656 	cv_destroy(&dbuf_evict_cv);
657 
658 	refcount_destroy(&dbuf_cache_size);
659 	multilist_destroy(dbuf_cache);
660 }
661 
662 /*
663  * Other stuff.
664  */
665 
666 #ifdef ZFS_DEBUG
667 static void
668 dbuf_verify(dmu_buf_impl_t *db)
669 {
670 	dnode_t *dn;
671 	dbuf_dirty_record_t *dr;
672 
673 	ASSERT(MUTEX_HELD(&db->db_mtx));
674 
675 	if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
676 		return;
677 
678 	ASSERT(db->db_objset != NULL);
679 	DB_DNODE_ENTER(db);
680 	dn = DB_DNODE(db);
681 	if (dn == NULL) {
682 		ASSERT(db->db_parent == NULL);
683 		ASSERT(db->db_blkptr == NULL);
684 	} else {
685 		ASSERT3U(db->db.db_object, ==, dn->dn_object);
686 		ASSERT3P(db->db_objset, ==, dn->dn_objset);
687 		ASSERT3U(db->db_level, <, dn->dn_nlevels);
688 		ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
689 		    db->db_blkid == DMU_SPILL_BLKID ||
690 		    !avl_is_empty(&dn->dn_dbufs));
691 	}
692 	if (db->db_blkid == DMU_BONUS_BLKID) {
693 		ASSERT(dn != NULL);
694 		ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
695 		ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
696 	} else if (db->db_blkid == DMU_SPILL_BLKID) {
697 		ASSERT(dn != NULL);
698 		ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
699 		ASSERT0(db->db.db_offset);
700 	} else {
701 		ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
702 	}
703 
704 	for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
705 		ASSERT(dr->dr_dbuf == db);
706 
707 	for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
708 		ASSERT(dr->dr_dbuf == db);
709 
710 	/*
711 	 * We can't assert that db_size matches dn_datablksz because it
712 	 * can be momentarily different when another thread is doing
713 	 * dnode_set_blksz().
714 	 */
715 	if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
716 		dr = db->db_data_pending;
717 		/*
718 		 * It should only be modified in syncing context, so
719 		 * make sure we only have one copy of the data.
720 		 */
721 		ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
722 	}
723 
724 	/* verify db->db_blkptr */
725 	if (db->db_blkptr) {
726 		if (db->db_parent == dn->dn_dbuf) {
727 			/* db is pointed to by the dnode */
728 			/* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
729 			if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
730 				ASSERT(db->db_parent == NULL);
731 			else
732 				ASSERT(db->db_parent != NULL);
733 			if (db->db_blkid != DMU_SPILL_BLKID)
734 				ASSERT3P(db->db_blkptr, ==,
735 				    &dn->dn_phys->dn_blkptr[db->db_blkid]);
736 		} else {
737 			/* db is pointed to by an indirect block */
738 			int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
739 			ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
740 			ASSERT3U(db->db_parent->db.db_object, ==,
741 			    db->db.db_object);
742 			/*
743 			 * dnode_grow_indblksz() can make this fail if we don't
744 			 * have the struct_rwlock.  XXX indblksz no longer
745 			 * grows.  safe to do this now?
746 			 */
747 			if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
748 				ASSERT3P(db->db_blkptr, ==,
749 				    ((blkptr_t *)db->db_parent->db.db_data +
750 				    db->db_blkid % epb));
751 			}
752 		}
753 	}
754 	if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
755 	    (db->db_buf == NULL || db->db_buf->b_data) &&
756 	    db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
757 	    db->db_state != DB_FILL && !dn->dn_free_txg) {
758 		/*
759 		 * If the blkptr isn't set but they have nonzero data,
760 		 * it had better be dirty, otherwise we'll lose that
761 		 * data when we evict this buffer.
762 		 *
763 		 * There is an exception to this rule for indirect blocks; in
764 		 * this case, if the indirect block is a hole, we fill in a few
765 		 * fields on each of the child blocks (importantly, birth time)
766 		 * to prevent hole birth times from being lost when you
767 		 * partially fill in a hole.
768 		 */
769 		if (db->db_dirtycnt == 0) {
770 			if (db->db_level == 0) {
771 				uint64_t *buf = db->db.db_data;
772 				int i;
773 
774 				for (i = 0; i < db->db.db_size >> 3; i++) {
775 					ASSERT(buf[i] == 0);
776 				}
777 			} else {
778 				blkptr_t *bps = db->db.db_data;
779 				ASSERT3U(1 << DB_DNODE(db)->dn_indblkshift, ==,
780 				    db->db.db_size);
781 				/*
782 				 * We want to verify that all the blkptrs in the
783 				 * indirect block are holes, but we may have
784 				 * automatically set up a few fields for them.
785 				 * We iterate through each blkptr and verify
786 				 * they only have those fields set.
787 				 */
788 				for (int i = 0;
789 				    i < db->db.db_size / sizeof (blkptr_t);
790 				    i++) {
791 					blkptr_t *bp = &bps[i];
792 					ASSERT(ZIO_CHECKSUM_IS_ZERO(
793 					    &bp->blk_cksum));
794 					ASSERT(
795 					    DVA_IS_EMPTY(&bp->blk_dva[0]) &&
796 					    DVA_IS_EMPTY(&bp->blk_dva[1]) &&
797 					    DVA_IS_EMPTY(&bp->blk_dva[2]));
798 					ASSERT0(bp->blk_fill);
799 					ASSERT0(bp->blk_pad[0]);
800 					ASSERT0(bp->blk_pad[1]);
801 					ASSERT(!BP_IS_EMBEDDED(bp));
802 					ASSERT(BP_IS_HOLE(bp));
803 					ASSERT0(bp->blk_phys_birth);
804 				}
805 			}
806 		}
807 	}
808 	DB_DNODE_EXIT(db);
809 }
810 #endif
811 
812 static void
813 dbuf_clear_data(dmu_buf_impl_t *db)
814 {
815 	ASSERT(MUTEX_HELD(&db->db_mtx));
816 	dbuf_evict_user(db);
817 	ASSERT3P(db->db_buf, ==, NULL);
818 	db->db.db_data = NULL;
819 	if (db->db_state != DB_NOFILL)
820 		db->db_state = DB_UNCACHED;
821 }
822 
823 static void
824 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
825 {
826 	ASSERT(MUTEX_HELD(&db->db_mtx));
827 	ASSERT(buf != NULL);
828 
829 	db->db_buf = buf;
830 	ASSERT(buf->b_data != NULL);
831 	db->db.db_data = buf->b_data;
832 }
833 
834 /*
835  * Loan out an arc_buf for read.  Return the loaned arc_buf.
836  */
837 arc_buf_t *
838 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
839 {
840 	arc_buf_t *abuf;
841 
842 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
843 	mutex_enter(&db->db_mtx);
844 	if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
845 		int blksz = db->db.db_size;
846 		spa_t *spa = db->db_objset->os_spa;
847 
848 		mutex_exit(&db->db_mtx);
849 		abuf = arc_loan_buf(spa, B_FALSE, blksz);
850 		bcopy(db->db.db_data, abuf->b_data, blksz);
851 	} else {
852 		abuf = db->db_buf;
853 		arc_loan_inuse_buf(abuf, db);
854 		db->db_buf = NULL;
855 		dbuf_clear_data(db);
856 		mutex_exit(&db->db_mtx);
857 	}
858 	return (abuf);
859 }
860 
861 /*
862  * Calculate which level n block references the data at the level 0 offset
863  * provided.
864  */
865 uint64_t
866 dbuf_whichblock(dnode_t *dn, int64_t level, uint64_t offset)
867 {
868 	if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) {
869 		/*
870 		 * The level n blkid is equal to the level 0 blkid divided by
871 		 * the number of level 0s in a level n block.
872 		 *
873 		 * The level 0 blkid is offset >> datablkshift =
874 		 * offset / 2^datablkshift.
875 		 *
876 		 * The number of level 0s in a level n is the number of block
877 		 * pointers in an indirect block, raised to the power of level.
878 		 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
879 		 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
880 		 *
881 		 * Thus, the level n blkid is: offset /
882 		 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
883 		 * = offset / 2^(datablkshift + level *
884 		 *   (indblkshift - SPA_BLKPTRSHIFT))
885 		 * = offset >> (datablkshift + level *
886 		 *   (indblkshift - SPA_BLKPTRSHIFT))
887 		 */
888 		return (offset >> (dn->dn_datablkshift + level *
889 		    (dn->dn_indblkshift - SPA_BLKPTRSHIFT)));
890 	} else {
891 		ASSERT3U(offset, <, dn->dn_datablksz);
892 		return (0);
893 	}
894 }
895 
896 static void
897 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
898 {
899 	dmu_buf_impl_t *db = vdb;
900 
901 	mutex_enter(&db->db_mtx);
902 	ASSERT3U(db->db_state, ==, DB_READ);
903 	/*
904 	 * All reads are synchronous, so we must have a hold on the dbuf
905 	 */
906 	ASSERT(refcount_count(&db->db_holds) > 0);
907 	ASSERT(db->db_buf == NULL);
908 	ASSERT(db->db.db_data == NULL);
909 	if (db->db_level == 0 && db->db_freed_in_flight) {
910 		/* we were freed in flight; disregard any error */
911 		arc_release(buf, db);
912 		bzero(buf->b_data, db->db.db_size);
913 		arc_buf_freeze(buf);
914 		db->db_freed_in_flight = FALSE;
915 		dbuf_set_data(db, buf);
916 		db->db_state = DB_CACHED;
917 	} else if (zio == NULL || zio->io_error == 0) {
918 		dbuf_set_data(db, buf);
919 		db->db_state = DB_CACHED;
920 	} else {
921 		ASSERT(db->db_blkid != DMU_BONUS_BLKID);
922 		ASSERT3P(db->db_buf, ==, NULL);
923 		arc_buf_destroy(buf, db);
924 		db->db_state = DB_UNCACHED;
925 	}
926 	cv_broadcast(&db->db_changed);
927 	dbuf_rele_and_unlock(db, NULL);
928 }
929 
930 static void
931 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
932 {
933 	dnode_t *dn;
934 	zbookmark_phys_t zb;
935 	arc_flags_t aflags = ARC_FLAG_NOWAIT;
936 
937 	DB_DNODE_ENTER(db);
938 	dn = DB_DNODE(db);
939 	ASSERT(!refcount_is_zero(&db->db_holds));
940 	/* We need the struct_rwlock to prevent db_blkptr from changing. */
941 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
942 	ASSERT(MUTEX_HELD(&db->db_mtx));
943 	ASSERT(db->db_state == DB_UNCACHED);
944 	ASSERT(db->db_buf == NULL);
945 
946 	if (db->db_blkid == DMU_BONUS_BLKID) {
947 		int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
948 
949 		ASSERT3U(bonuslen, <=, db->db.db_size);
950 		db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN);
951 		arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
952 		if (bonuslen < DN_MAX_BONUSLEN)
953 			bzero(db->db.db_data, DN_MAX_BONUSLEN);
954 		if (bonuslen)
955 			bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
956 		DB_DNODE_EXIT(db);
957 		db->db_state = DB_CACHED;
958 		mutex_exit(&db->db_mtx);
959 		return;
960 	}
961 
962 	/*
963 	 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
964 	 * processes the delete record and clears the bp while we are waiting
965 	 * for the dn_mtx (resulting in a "no" from block_freed).
966 	 */
967 	if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
968 	    (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
969 	    BP_IS_HOLE(db->db_blkptr)))) {
970 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
971 
972 		dbuf_set_data(db, arc_alloc_buf(db->db_objset->os_spa, db, type,
973 		    db->db.db_size));
974 		bzero(db->db.db_data, db->db.db_size);
975 
976 		if (db->db_blkptr != NULL && db->db_level > 0 &&
977 		    BP_IS_HOLE(db->db_blkptr) &&
978 		    db->db_blkptr->blk_birth != 0) {
979 			blkptr_t *bps = db->db.db_data;
980 			for (int i = 0; i < ((1 <<
981 			    DB_DNODE(db)->dn_indblkshift) / sizeof (blkptr_t));
982 			    i++) {
983 				blkptr_t *bp = &bps[i];
984 				ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
985 				    1 << dn->dn_indblkshift);
986 				BP_SET_LSIZE(bp,
987 				    BP_GET_LEVEL(db->db_blkptr) == 1 ?
988 				    dn->dn_datablksz :
989 				    BP_GET_LSIZE(db->db_blkptr));
990 				BP_SET_TYPE(bp, BP_GET_TYPE(db->db_blkptr));
991 				BP_SET_LEVEL(bp,
992 				    BP_GET_LEVEL(db->db_blkptr) - 1);
993 				BP_SET_BIRTH(bp, db->db_blkptr->blk_birth, 0);
994 			}
995 		}
996 		DB_DNODE_EXIT(db);
997 		db->db_state = DB_CACHED;
998 		mutex_exit(&db->db_mtx);
999 		return;
1000 	}
1001 
1002 	DB_DNODE_EXIT(db);
1003 
1004 	db->db_state = DB_READ;
1005 	mutex_exit(&db->db_mtx);
1006 
1007 	if (DBUF_IS_L2CACHEABLE(db))
1008 		aflags |= ARC_FLAG_L2CACHE;
1009 
1010 	SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
1011 	    db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
1012 	    db->db.db_object, db->db_level, db->db_blkid);
1013 
1014 	dbuf_add_ref(db, NULL);
1015 
1016 	(void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr,
1017 	    dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
1018 	    (flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
1019 	    &aflags, &zb);
1020 }
1021 
1022 /*
1023  * This is our just-in-time copy function.  It makes a copy of buffers that
1024  * have been modified in a previous transaction group before we access them in
1025  * the current active group.
1026  *
1027  * This function is used in three places: when we are dirtying a buffer for the
1028  * first time in a txg, when we are freeing a range in a dnode that includes
1029  * this buffer, and when we are accessing a buffer which was received compressed
1030  * and later referenced in a WRITE_BYREF record.
1031  *
1032  * Note that when we are called from dbuf_free_range() we do not put a hold on
1033  * the buffer, we just traverse the active dbuf list for the dnode.
1034  */
1035 static void
1036 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
1037 {
1038 	dbuf_dirty_record_t *dr = db->db_last_dirty;
1039 
1040 	ASSERT(MUTEX_HELD(&db->db_mtx));
1041 	ASSERT(db->db.db_data != NULL);
1042 	ASSERT(db->db_level == 0);
1043 	ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
1044 
1045 	if (dr == NULL ||
1046 	    (dr->dt.dl.dr_data !=
1047 	    ((db->db_blkid  == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
1048 		return;
1049 
1050 	/*
1051 	 * If the last dirty record for this dbuf has not yet synced
1052 	 * and its referencing the dbuf data, either:
1053 	 *	reset the reference to point to a new copy,
1054 	 * or (if there a no active holders)
1055 	 *	just null out the current db_data pointer.
1056 	 */
1057 	ASSERT(dr->dr_txg >= txg - 2);
1058 	if (db->db_blkid == DMU_BONUS_BLKID) {
1059 		/* Note that the data bufs here are zio_bufs */
1060 		dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
1061 		arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
1062 		bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
1063 	} else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
1064 		int size = arc_buf_size(db->db_buf);
1065 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1066 		spa_t *spa = db->db_objset->os_spa;
1067 		enum zio_compress compress_type =
1068 		    arc_get_compression(db->db_buf);
1069 
1070 		if (compress_type == ZIO_COMPRESS_OFF) {
1071 			dr->dt.dl.dr_data = arc_alloc_buf(spa, db, type, size);
1072 		} else {
1073 			ASSERT3U(type, ==, ARC_BUFC_DATA);
1074 			dr->dt.dl.dr_data = arc_alloc_compressed_buf(spa, db,
1075 			    size, arc_buf_lsize(db->db_buf), compress_type);
1076 		}
1077 		bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
1078 	} else {
1079 		db->db_buf = NULL;
1080 		dbuf_clear_data(db);
1081 	}
1082 }
1083 
1084 int
1085 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
1086 {
1087 	int err = 0;
1088 	boolean_t prefetch;
1089 	dnode_t *dn;
1090 
1091 	/*
1092 	 * We don't have to hold the mutex to check db_state because it
1093 	 * can't be freed while we have a hold on the buffer.
1094 	 */
1095 	ASSERT(!refcount_is_zero(&db->db_holds));
1096 
1097 	if (db->db_state == DB_NOFILL)
1098 		return (SET_ERROR(EIO));
1099 
1100 	DB_DNODE_ENTER(db);
1101 	dn = DB_DNODE(db);
1102 	if ((flags & DB_RF_HAVESTRUCT) == 0)
1103 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
1104 
1105 	prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1106 	    (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
1107 	    DBUF_IS_CACHEABLE(db);
1108 
1109 	mutex_enter(&db->db_mtx);
1110 	if (db->db_state == DB_CACHED) {
1111 		/*
1112 		 * If the arc buf is compressed, we need to decompress it to
1113 		 * read the data. This could happen during the "zfs receive" of
1114 		 * a stream which is compressed and deduplicated.
1115 		 */
1116 		if (db->db_buf != NULL &&
1117 		    arc_get_compression(db->db_buf) != ZIO_COMPRESS_OFF) {
1118 			dbuf_fix_old_data(db,
1119 			    spa_syncing_txg(dmu_objset_spa(db->db_objset)));
1120 			err = arc_decompress(db->db_buf);
1121 			dbuf_set_data(db, db->db_buf);
1122 		}
1123 		mutex_exit(&db->db_mtx);
1124 		if (prefetch)
1125 			dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1126 		if ((flags & DB_RF_HAVESTRUCT) == 0)
1127 			rw_exit(&dn->dn_struct_rwlock);
1128 		DB_DNODE_EXIT(db);
1129 	} else if (db->db_state == DB_UNCACHED) {
1130 		spa_t *spa = dn->dn_objset->os_spa;
1131 		boolean_t need_wait = B_FALSE;
1132 
1133 		if (zio == NULL &&
1134 		    db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) {
1135 			zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
1136 			need_wait = B_TRUE;
1137 		}
1138 		dbuf_read_impl(db, zio, flags);
1139 
1140 		/* dbuf_read_impl has dropped db_mtx for us */
1141 
1142 		if (prefetch)
1143 			dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1144 
1145 		if ((flags & DB_RF_HAVESTRUCT) == 0)
1146 			rw_exit(&dn->dn_struct_rwlock);
1147 		DB_DNODE_EXIT(db);
1148 
1149 		if (need_wait)
1150 			err = zio_wait(zio);
1151 	} else {
1152 		/*
1153 		 * Another reader came in while the dbuf was in flight
1154 		 * between UNCACHED and CACHED.  Either a writer will finish
1155 		 * writing the buffer (sending the dbuf to CACHED) or the
1156 		 * first reader's request will reach the read_done callback
1157 		 * and send the dbuf to CACHED.  Otherwise, a failure
1158 		 * occurred and the dbuf went to UNCACHED.
1159 		 */
1160 		mutex_exit(&db->db_mtx);
1161 		if (prefetch)
1162 			dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1163 		if ((flags & DB_RF_HAVESTRUCT) == 0)
1164 			rw_exit(&dn->dn_struct_rwlock);
1165 		DB_DNODE_EXIT(db);
1166 
1167 		/* Skip the wait per the caller's request. */
1168 		mutex_enter(&db->db_mtx);
1169 		if ((flags & DB_RF_NEVERWAIT) == 0) {
1170 			while (db->db_state == DB_READ ||
1171 			    db->db_state == DB_FILL) {
1172 				ASSERT(db->db_state == DB_READ ||
1173 				    (flags & DB_RF_HAVESTRUCT) == 0);
1174 				DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *,
1175 				    db, zio_t *, zio);
1176 				cv_wait(&db->db_changed, &db->db_mtx);
1177 			}
1178 			if (db->db_state == DB_UNCACHED)
1179 				err = SET_ERROR(EIO);
1180 		}
1181 		mutex_exit(&db->db_mtx);
1182 	}
1183 
1184 	return (err);
1185 }
1186 
1187 static void
1188 dbuf_noread(dmu_buf_impl_t *db)
1189 {
1190 	ASSERT(!refcount_is_zero(&db->db_holds));
1191 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1192 	mutex_enter(&db->db_mtx);
1193 	while (db->db_state == DB_READ || db->db_state == DB_FILL)
1194 		cv_wait(&db->db_changed, &db->db_mtx);
1195 	if (db->db_state == DB_UNCACHED) {
1196 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1197 		spa_t *spa = db->db_objset->os_spa;
1198 
1199 		ASSERT(db->db_buf == NULL);
1200 		ASSERT(db->db.db_data == NULL);
1201 		dbuf_set_data(db, arc_alloc_buf(spa, db, type, db->db.db_size));
1202 		db->db_state = DB_FILL;
1203 	} else if (db->db_state == DB_NOFILL) {
1204 		dbuf_clear_data(db);
1205 	} else {
1206 		ASSERT3U(db->db_state, ==, DB_CACHED);
1207 	}
1208 	mutex_exit(&db->db_mtx);
1209 }
1210 
1211 void
1212 dbuf_unoverride(dbuf_dirty_record_t *dr)
1213 {
1214 	dmu_buf_impl_t *db = dr->dr_dbuf;
1215 	blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
1216 	uint64_t txg = dr->dr_txg;
1217 
1218 	ASSERT(MUTEX_HELD(&db->db_mtx));
1219 	/*
1220 	 * This assert is valid because dmu_sync() expects to be called by
1221 	 * a zilog's get_data while holding a range lock.  This call only
1222 	 * comes from dbuf_dirty() callers who must also hold a range lock.
1223 	 */
1224 	ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
1225 	ASSERT(db->db_level == 0);
1226 
1227 	if (db->db_blkid == DMU_BONUS_BLKID ||
1228 	    dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
1229 		return;
1230 
1231 	ASSERT(db->db_data_pending != dr);
1232 
1233 	/* free this block */
1234 	if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite)
1235 		zio_free(db->db_objset->os_spa, txg, bp);
1236 
1237 	dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1238 	dr->dt.dl.dr_nopwrite = B_FALSE;
1239 
1240 	/*
1241 	 * Release the already-written buffer, so we leave it in
1242 	 * a consistent dirty state.  Note that all callers are
1243 	 * modifying the buffer, so they will immediately do
1244 	 * another (redundant) arc_release().  Therefore, leave
1245 	 * the buf thawed to save the effort of freezing &
1246 	 * immediately re-thawing it.
1247 	 */
1248 	arc_release(dr->dt.dl.dr_data, db);
1249 }
1250 
1251 /*
1252  * Evict (if its unreferenced) or clear (if its referenced) any level-0
1253  * data blocks in the free range, so that any future readers will find
1254  * empty blocks.
1255  */
1256 void
1257 dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
1258     dmu_tx_t *tx)
1259 {
1260 	dmu_buf_impl_t db_search;
1261 	dmu_buf_impl_t *db, *db_next;
1262 	uint64_t txg = tx->tx_txg;
1263 	avl_index_t where;
1264 
1265 	if (end_blkid > dn->dn_maxblkid &&
1266 	    !(start_blkid == DMU_SPILL_BLKID || end_blkid == DMU_SPILL_BLKID))
1267 		end_blkid = dn->dn_maxblkid;
1268 	dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid);
1269 
1270 	db_search.db_level = 0;
1271 	db_search.db_blkid = start_blkid;
1272 	db_search.db_state = DB_SEARCH;
1273 
1274 	mutex_enter(&dn->dn_dbufs_mtx);
1275 	db = avl_find(&dn->dn_dbufs, &db_search, &where);
1276 	ASSERT3P(db, ==, NULL);
1277 
1278 	db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1279 
1280 	for (; db != NULL; db = db_next) {
1281 		db_next = AVL_NEXT(&dn->dn_dbufs, db);
1282 		ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1283 
1284 		if (db->db_level != 0 || db->db_blkid > end_blkid) {
1285 			break;
1286 		}
1287 		ASSERT3U(db->db_blkid, >=, start_blkid);
1288 
1289 		/* found a level 0 buffer in the range */
1290 		mutex_enter(&db->db_mtx);
1291 		if (dbuf_undirty(db, tx)) {
1292 			/* mutex has been dropped and dbuf destroyed */
1293 			continue;
1294 		}
1295 
1296 		if (db->db_state == DB_UNCACHED ||
1297 		    db->db_state == DB_NOFILL ||
1298 		    db->db_state == DB_EVICTING) {
1299 			ASSERT(db->db.db_data == NULL);
1300 			mutex_exit(&db->db_mtx);
1301 			continue;
1302 		}
1303 		if (db->db_state == DB_READ || db->db_state == DB_FILL) {
1304 			/* will be handled in dbuf_read_done or dbuf_rele */
1305 			db->db_freed_in_flight = TRUE;
1306 			mutex_exit(&db->db_mtx);
1307 			continue;
1308 		}
1309 		if (refcount_count(&db->db_holds) == 0) {
1310 			ASSERT(db->db_buf);
1311 			dbuf_destroy(db);
1312 			continue;
1313 		}
1314 		/* The dbuf is referenced */
1315 
1316 		if (db->db_last_dirty != NULL) {
1317 			dbuf_dirty_record_t *dr = db->db_last_dirty;
1318 
1319 			if (dr->dr_txg == txg) {
1320 				/*
1321 				 * This buffer is "in-use", re-adjust the file
1322 				 * size to reflect that this buffer may
1323 				 * contain new data when we sync.
1324 				 */
1325 				if (db->db_blkid != DMU_SPILL_BLKID &&
1326 				    db->db_blkid > dn->dn_maxblkid)
1327 					dn->dn_maxblkid = db->db_blkid;
1328 				dbuf_unoverride(dr);
1329 			} else {
1330 				/*
1331 				 * This dbuf is not dirty in the open context.
1332 				 * Either uncache it (if its not referenced in
1333 				 * the open context) or reset its contents to
1334 				 * empty.
1335 				 */
1336 				dbuf_fix_old_data(db, txg);
1337 			}
1338 		}
1339 		/* clear the contents if its cached */
1340 		if (db->db_state == DB_CACHED) {
1341 			ASSERT(db->db.db_data != NULL);
1342 			arc_release(db->db_buf, db);
1343 			bzero(db->db.db_data, db->db.db_size);
1344 			arc_buf_freeze(db->db_buf);
1345 		}
1346 
1347 		mutex_exit(&db->db_mtx);
1348 	}
1349 	mutex_exit(&dn->dn_dbufs_mtx);
1350 }
1351 
1352 void
1353 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
1354 {
1355 	arc_buf_t *buf, *obuf;
1356 	int osize = db->db.db_size;
1357 	arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1358 	dnode_t *dn;
1359 
1360 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1361 
1362 	DB_DNODE_ENTER(db);
1363 	dn = DB_DNODE(db);
1364 
1365 	/* XXX does *this* func really need the lock? */
1366 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1367 
1368 	/*
1369 	 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1370 	 * is OK, because there can be no other references to the db
1371 	 * when we are changing its size, so no concurrent DB_FILL can
1372 	 * be happening.
1373 	 */
1374 	/*
1375 	 * XXX we should be doing a dbuf_read, checking the return
1376 	 * value and returning that up to our callers
1377 	 */
1378 	dmu_buf_will_dirty(&db->db, tx);
1379 
1380 	/* create the data buffer for the new block */
1381 	buf = arc_alloc_buf(dn->dn_objset->os_spa, db, type, size);
1382 
1383 	/* copy old block data to the new block */
1384 	obuf = db->db_buf;
1385 	bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
1386 	/* zero the remainder */
1387 	if (size > osize)
1388 		bzero((uint8_t *)buf->b_data + osize, size - osize);
1389 
1390 	mutex_enter(&db->db_mtx);
1391 	dbuf_set_data(db, buf);
1392 	arc_buf_destroy(obuf, db);
1393 	db->db.db_size = size;
1394 
1395 	if (db->db_level == 0) {
1396 		ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1397 		db->db_last_dirty->dt.dl.dr_data = buf;
1398 	}
1399 	mutex_exit(&db->db_mtx);
1400 
1401 	dmu_objset_willuse_space(dn->dn_objset, size - osize, tx);
1402 	DB_DNODE_EXIT(db);
1403 }
1404 
1405 void
1406 dbuf_release_bp(dmu_buf_impl_t *db)
1407 {
1408 	objset_t *os = db->db_objset;
1409 
1410 	ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1411 	ASSERT(arc_released(os->os_phys_buf) ||
1412 	    list_link_active(&os->os_dsl_dataset->ds_synced_link));
1413 	ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1414 
1415 	(void) arc_release(db->db_buf, db);
1416 }
1417 
1418 /*
1419  * We already have a dirty record for this TXG, and we are being
1420  * dirtied again.
1421  */
1422 static void
1423 dbuf_redirty(dbuf_dirty_record_t *dr)
1424 {
1425 	dmu_buf_impl_t *db = dr->dr_dbuf;
1426 
1427 	ASSERT(MUTEX_HELD(&db->db_mtx));
1428 
1429 	if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1430 		/*
1431 		 * If this buffer has already been written out,
1432 		 * we now need to reset its state.
1433 		 */
1434 		dbuf_unoverride(dr);
1435 		if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1436 		    db->db_state != DB_NOFILL) {
1437 			/* Already released on initial dirty, so just thaw. */
1438 			ASSERT(arc_released(db->db_buf));
1439 			arc_buf_thaw(db->db_buf);
1440 		}
1441 	}
1442 }
1443 
1444 dbuf_dirty_record_t *
1445 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1446 {
1447 	dnode_t *dn;
1448 	objset_t *os;
1449 	dbuf_dirty_record_t **drp, *dr;
1450 	int drop_struct_lock = FALSE;
1451 	int txgoff = tx->tx_txg & TXG_MASK;
1452 
1453 	ASSERT(tx->tx_txg != 0);
1454 	ASSERT(!refcount_is_zero(&db->db_holds));
1455 	DMU_TX_DIRTY_BUF(tx, db);
1456 
1457 	DB_DNODE_ENTER(db);
1458 	dn = DB_DNODE(db);
1459 	/*
1460 	 * Shouldn't dirty a regular buffer in syncing context.  Private
1461 	 * objects may be dirtied in syncing context, but only if they
1462 	 * were already pre-dirtied in open context.
1463 	 */
1464 #ifdef DEBUG
1465 	if (dn->dn_objset->os_dsl_dataset != NULL) {
1466 		rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1467 		    RW_READER, FTAG);
1468 	}
1469 	ASSERT(!dmu_tx_is_syncing(tx) ||
1470 	    BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1471 	    DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1472 	    dn->dn_objset->os_dsl_dataset == NULL);
1473 	if (dn->dn_objset->os_dsl_dataset != NULL)
1474 		rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, FTAG);
1475 #endif
1476 	/*
1477 	 * We make this assert for private objects as well, but after we
1478 	 * check if we're already dirty.  They are allowed to re-dirty
1479 	 * in syncing context.
1480 	 */
1481 	ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1482 	    dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1483 	    (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1484 
1485 	mutex_enter(&db->db_mtx);
1486 	/*
1487 	 * XXX make this true for indirects too?  The problem is that
1488 	 * transactions created with dmu_tx_create_assigned() from
1489 	 * syncing context don't bother holding ahead.
1490 	 */
1491 	ASSERT(db->db_level != 0 ||
1492 	    db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1493 	    db->db_state == DB_NOFILL);
1494 
1495 	mutex_enter(&dn->dn_mtx);
1496 	/*
1497 	 * Don't set dirtyctx to SYNC if we're just modifying this as we
1498 	 * initialize the objset.
1499 	 */
1500 	if (dn->dn_dirtyctx == DN_UNDIRTIED) {
1501 		if (dn->dn_objset->os_dsl_dataset != NULL) {
1502 			rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1503 			    RW_READER, FTAG);
1504 		}
1505 		if (!BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1506 			dn->dn_dirtyctx = (dmu_tx_is_syncing(tx) ?
1507 			    DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1508 			ASSERT(dn->dn_dirtyctx_firstset == NULL);
1509 			dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1510 		}
1511 		if (dn->dn_objset->os_dsl_dataset != NULL) {
1512 			rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1513 			    FTAG);
1514 		}
1515 	}
1516 	mutex_exit(&dn->dn_mtx);
1517 
1518 	if (db->db_blkid == DMU_SPILL_BLKID)
1519 		dn->dn_have_spill = B_TRUE;
1520 
1521 	/*
1522 	 * If this buffer is already dirty, we're done.
1523 	 */
1524 	drp = &db->db_last_dirty;
1525 	ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1526 	    db->db.db_object == DMU_META_DNODE_OBJECT);
1527 	while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1528 		drp = &dr->dr_next;
1529 	if (dr && dr->dr_txg == tx->tx_txg) {
1530 		DB_DNODE_EXIT(db);
1531 
1532 		dbuf_redirty(dr);
1533 		mutex_exit(&db->db_mtx);
1534 		return (dr);
1535 	}
1536 
1537 	/*
1538 	 * Only valid if not already dirty.
1539 	 */
1540 	ASSERT(dn->dn_object == 0 ||
1541 	    dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1542 	    (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1543 
1544 	ASSERT3U(dn->dn_nlevels, >, db->db_level);
1545 
1546 	/*
1547 	 * We should only be dirtying in syncing context if it's the
1548 	 * mos or we're initializing the os or it's a special object.
1549 	 * However, we are allowed to dirty in syncing context provided
1550 	 * we already dirtied it in open context.  Hence we must make
1551 	 * this assertion only if we're not already dirty.
1552 	 */
1553 	os = dn->dn_objset;
1554 	VERIFY3U(tx->tx_txg, <=, spa_final_dirty_txg(os->os_spa));
1555 #ifdef DEBUG
1556 	if (dn->dn_objset->os_dsl_dataset != NULL)
1557 		rrw_enter(&os->os_dsl_dataset->ds_bp_rwlock, RW_READER, FTAG);
1558 	ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1559 	    os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1560 	if (dn->dn_objset->os_dsl_dataset != NULL)
1561 		rrw_exit(&os->os_dsl_dataset->ds_bp_rwlock, FTAG);
1562 #endif
1563 	ASSERT(db->db.db_size != 0);
1564 
1565 	dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1566 
1567 	if (db->db_blkid != DMU_BONUS_BLKID) {
1568 		dmu_objset_willuse_space(os, db->db.db_size, tx);
1569 	}
1570 
1571 	/*
1572 	 * If this buffer is dirty in an old transaction group we need
1573 	 * to make a copy of it so that the changes we make in this
1574 	 * transaction group won't leak out when we sync the older txg.
1575 	 */
1576 	dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1577 	if (db->db_level == 0) {
1578 		void *data_old = db->db_buf;
1579 
1580 		if (db->db_state != DB_NOFILL) {
1581 			if (db->db_blkid == DMU_BONUS_BLKID) {
1582 				dbuf_fix_old_data(db, tx->tx_txg);
1583 				data_old = db->db.db_data;
1584 			} else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1585 				/*
1586 				 * Release the data buffer from the cache so
1587 				 * that we can modify it without impacting
1588 				 * possible other users of this cached data
1589 				 * block.  Note that indirect blocks and
1590 				 * private objects are not released until the
1591 				 * syncing state (since they are only modified
1592 				 * then).
1593 				 */
1594 				arc_release(db->db_buf, db);
1595 				dbuf_fix_old_data(db, tx->tx_txg);
1596 				data_old = db->db_buf;
1597 			}
1598 			ASSERT(data_old != NULL);
1599 		}
1600 		dr->dt.dl.dr_data = data_old;
1601 	} else {
1602 		mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1603 		list_create(&dr->dt.di.dr_children,
1604 		    sizeof (dbuf_dirty_record_t),
1605 		    offsetof(dbuf_dirty_record_t, dr_dirty_node));
1606 	}
1607 	if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL)
1608 		dr->dr_accounted = db->db.db_size;
1609 	dr->dr_dbuf = db;
1610 	dr->dr_txg = tx->tx_txg;
1611 	dr->dr_next = *drp;
1612 	*drp = dr;
1613 
1614 	/*
1615 	 * We could have been freed_in_flight between the dbuf_noread
1616 	 * and dbuf_dirty.  We win, as though the dbuf_noread() had
1617 	 * happened after the free.
1618 	 */
1619 	if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1620 	    db->db_blkid != DMU_SPILL_BLKID) {
1621 		mutex_enter(&dn->dn_mtx);
1622 		if (dn->dn_free_ranges[txgoff] != NULL) {
1623 			range_tree_clear(dn->dn_free_ranges[txgoff],
1624 			    db->db_blkid, 1);
1625 		}
1626 		mutex_exit(&dn->dn_mtx);
1627 		db->db_freed_in_flight = FALSE;
1628 	}
1629 
1630 	/*
1631 	 * This buffer is now part of this txg
1632 	 */
1633 	dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1634 	db->db_dirtycnt += 1;
1635 	ASSERT3U(db->db_dirtycnt, <=, 3);
1636 
1637 	mutex_exit(&db->db_mtx);
1638 
1639 	if (db->db_blkid == DMU_BONUS_BLKID ||
1640 	    db->db_blkid == DMU_SPILL_BLKID) {
1641 		mutex_enter(&dn->dn_mtx);
1642 		ASSERT(!list_link_active(&dr->dr_dirty_node));
1643 		list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1644 		mutex_exit(&dn->dn_mtx);
1645 		dnode_setdirty(dn, tx);
1646 		DB_DNODE_EXIT(db);
1647 		return (dr);
1648 	}
1649 
1650 	/*
1651 	 * The dn_struct_rwlock prevents db_blkptr from changing
1652 	 * due to a write from syncing context completing
1653 	 * while we are running, so we want to acquire it before
1654 	 * looking at db_blkptr.
1655 	 */
1656 	if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1657 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
1658 		drop_struct_lock = TRUE;
1659 	}
1660 
1661 	/*
1662 	 * We need to hold the dn_struct_rwlock to make this assertion,
1663 	 * because it protects dn_phys / dn_next_nlevels from changing.
1664 	 */
1665 	ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1666 	    dn->dn_phys->dn_nlevels > db->db_level ||
1667 	    dn->dn_next_nlevels[txgoff] > db->db_level ||
1668 	    dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1669 	    dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1670 
1671 	/*
1672 	 * If we are overwriting a dedup BP, then unless it is snapshotted,
1673 	 * when we get to syncing context we will need to decrement its
1674 	 * refcount in the DDT.  Prefetch the relevant DDT block so that
1675 	 * syncing context won't have to wait for the i/o.
1676 	 */
1677 	ddt_prefetch(os->os_spa, db->db_blkptr);
1678 
1679 	if (db->db_level == 0) {
1680 		dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1681 		ASSERT(dn->dn_maxblkid >= db->db_blkid);
1682 	}
1683 
1684 	if (db->db_level+1 < dn->dn_nlevels) {
1685 		dmu_buf_impl_t *parent = db->db_parent;
1686 		dbuf_dirty_record_t *di;
1687 		int parent_held = FALSE;
1688 
1689 		if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1690 			int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1691 
1692 			parent = dbuf_hold_level(dn, db->db_level+1,
1693 			    db->db_blkid >> epbs, FTAG);
1694 			ASSERT(parent != NULL);
1695 			parent_held = TRUE;
1696 		}
1697 		if (drop_struct_lock)
1698 			rw_exit(&dn->dn_struct_rwlock);
1699 		ASSERT3U(db->db_level+1, ==, parent->db_level);
1700 		di = dbuf_dirty(parent, tx);
1701 		if (parent_held)
1702 			dbuf_rele(parent, FTAG);
1703 
1704 		mutex_enter(&db->db_mtx);
1705 		/*
1706 		 * Since we've dropped the mutex, it's possible that
1707 		 * dbuf_undirty() might have changed this out from under us.
1708 		 */
1709 		if (db->db_last_dirty == dr ||
1710 		    dn->dn_object == DMU_META_DNODE_OBJECT) {
1711 			mutex_enter(&di->dt.di.dr_mtx);
1712 			ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1713 			ASSERT(!list_link_active(&dr->dr_dirty_node));
1714 			list_insert_tail(&di->dt.di.dr_children, dr);
1715 			mutex_exit(&di->dt.di.dr_mtx);
1716 			dr->dr_parent = di;
1717 		}
1718 		mutex_exit(&db->db_mtx);
1719 	} else {
1720 		ASSERT(db->db_level+1 == dn->dn_nlevels);
1721 		ASSERT(db->db_blkid < dn->dn_nblkptr);
1722 		ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1723 		mutex_enter(&dn->dn_mtx);
1724 		ASSERT(!list_link_active(&dr->dr_dirty_node));
1725 		list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1726 		mutex_exit(&dn->dn_mtx);
1727 		if (drop_struct_lock)
1728 			rw_exit(&dn->dn_struct_rwlock);
1729 	}
1730 
1731 	dnode_setdirty(dn, tx);
1732 	DB_DNODE_EXIT(db);
1733 	return (dr);
1734 }
1735 
1736 /*
1737  * Undirty a buffer in the transaction group referenced by the given
1738  * transaction.  Return whether this evicted the dbuf.
1739  */
1740 static boolean_t
1741 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1742 {
1743 	dnode_t *dn;
1744 	uint64_t txg = tx->tx_txg;
1745 	dbuf_dirty_record_t *dr, **drp;
1746 
1747 	ASSERT(txg != 0);
1748 
1749 	/*
1750 	 * Due to our use of dn_nlevels below, this can only be called
1751 	 * in open context, unless we are operating on the MOS.
1752 	 * From syncing context, dn_nlevels may be different from the
1753 	 * dn_nlevels used when dbuf was dirtied.
1754 	 */
1755 	ASSERT(db->db_objset ==
1756 	    dmu_objset_pool(db->db_objset)->dp_meta_objset ||
1757 	    txg != spa_syncing_txg(dmu_objset_spa(db->db_objset)));
1758 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1759 	ASSERT0(db->db_level);
1760 	ASSERT(MUTEX_HELD(&db->db_mtx));
1761 
1762 	/*
1763 	 * If this buffer is not dirty, we're done.
1764 	 */
1765 	for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1766 		if (dr->dr_txg <= txg)
1767 			break;
1768 	if (dr == NULL || dr->dr_txg < txg)
1769 		return (B_FALSE);
1770 	ASSERT(dr->dr_txg == txg);
1771 	ASSERT(dr->dr_dbuf == db);
1772 
1773 	DB_DNODE_ENTER(db);
1774 	dn = DB_DNODE(db);
1775 
1776 	dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1777 
1778 	ASSERT(db->db.db_size != 0);
1779 
1780 	dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset),
1781 	    dr->dr_accounted, txg);
1782 
1783 	*drp = dr->dr_next;
1784 
1785 	/*
1786 	 * Note that there are three places in dbuf_dirty()
1787 	 * where this dirty record may be put on a list.
1788 	 * Make sure to do a list_remove corresponding to
1789 	 * every one of those list_insert calls.
1790 	 */
1791 	if (dr->dr_parent) {
1792 		mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1793 		list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1794 		mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1795 	} else if (db->db_blkid == DMU_SPILL_BLKID ||
1796 	    db->db_level + 1 == dn->dn_nlevels) {
1797 		ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1798 		mutex_enter(&dn->dn_mtx);
1799 		list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1800 		mutex_exit(&dn->dn_mtx);
1801 	}
1802 	DB_DNODE_EXIT(db);
1803 
1804 	if (db->db_state != DB_NOFILL) {
1805 		dbuf_unoverride(dr);
1806 
1807 		ASSERT(db->db_buf != NULL);
1808 		ASSERT(dr->dt.dl.dr_data != NULL);
1809 		if (dr->dt.dl.dr_data != db->db_buf)
1810 			arc_buf_destroy(dr->dt.dl.dr_data, db);
1811 	}
1812 
1813 	kmem_free(dr, sizeof (dbuf_dirty_record_t));
1814 
1815 	ASSERT(db->db_dirtycnt > 0);
1816 	db->db_dirtycnt -= 1;
1817 
1818 	if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1819 		ASSERT(db->db_state == DB_NOFILL || arc_released(db->db_buf));
1820 		dbuf_destroy(db);
1821 		return (B_TRUE);
1822 	}
1823 
1824 	return (B_FALSE);
1825 }
1826 
1827 void
1828 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
1829 {
1830 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1831 	int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1832 
1833 	ASSERT(tx->tx_txg != 0);
1834 	ASSERT(!refcount_is_zero(&db->db_holds));
1835 
1836 	/*
1837 	 * Quick check for dirtyness.  For already dirty blocks, this
1838 	 * reduces runtime of this function by >90%, and overall performance
1839 	 * by 50% for some workloads (e.g. file deletion with indirect blocks
1840 	 * cached).
1841 	 */
1842 	mutex_enter(&db->db_mtx);
1843 	dbuf_dirty_record_t *dr;
1844 	for (dr = db->db_last_dirty;
1845 	    dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) {
1846 		/*
1847 		 * It's possible that it is already dirty but not cached,
1848 		 * because there are some calls to dbuf_dirty() that don't
1849 		 * go through dmu_buf_will_dirty().
1850 		 */
1851 		if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) {
1852 			/* This dbuf is already dirty and cached. */
1853 			dbuf_redirty(dr);
1854 			mutex_exit(&db->db_mtx);
1855 			return;
1856 		}
1857 	}
1858 	mutex_exit(&db->db_mtx);
1859 
1860 	DB_DNODE_ENTER(db);
1861 	if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
1862 		rf |= DB_RF_HAVESTRUCT;
1863 	DB_DNODE_EXIT(db);
1864 	(void) dbuf_read(db, NULL, rf);
1865 	(void) dbuf_dirty(db, tx);
1866 }
1867 
1868 void
1869 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1870 {
1871 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1872 
1873 	db->db_state = DB_NOFILL;
1874 
1875 	dmu_buf_will_fill(db_fake, tx);
1876 }
1877 
1878 void
1879 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1880 {
1881 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1882 
1883 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1884 	ASSERT(tx->tx_txg != 0);
1885 	ASSERT(db->db_level == 0);
1886 	ASSERT(!refcount_is_zero(&db->db_holds));
1887 
1888 	ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1889 	    dmu_tx_private_ok(tx));
1890 
1891 	dbuf_noread(db);
1892 	(void) dbuf_dirty(db, tx);
1893 }
1894 
1895 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1896 /* ARGSUSED */
1897 void
1898 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1899 {
1900 	mutex_enter(&db->db_mtx);
1901 	DBUF_VERIFY(db);
1902 
1903 	if (db->db_state == DB_FILL) {
1904 		if (db->db_level == 0 && db->db_freed_in_flight) {
1905 			ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1906 			/* we were freed while filling */
1907 			/* XXX dbuf_undirty? */
1908 			bzero(db->db.db_data, db->db.db_size);
1909 			db->db_freed_in_flight = FALSE;
1910 		}
1911 		db->db_state = DB_CACHED;
1912 		cv_broadcast(&db->db_changed);
1913 	}
1914 	mutex_exit(&db->db_mtx);
1915 }
1916 
1917 void
1918 dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data,
1919     bp_embedded_type_t etype, enum zio_compress comp,
1920     int uncompressed_size, int compressed_size, int byteorder,
1921     dmu_tx_t *tx)
1922 {
1923 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
1924 	struct dirty_leaf *dl;
1925 	dmu_object_type_t type;
1926 
1927 	if (etype == BP_EMBEDDED_TYPE_DATA) {
1928 		ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset),
1929 		    SPA_FEATURE_EMBEDDED_DATA));
1930 	}
1931 
1932 	DB_DNODE_ENTER(db);
1933 	type = DB_DNODE(db)->dn_type;
1934 	DB_DNODE_EXIT(db);
1935 
1936 	ASSERT0(db->db_level);
1937 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1938 
1939 	dmu_buf_will_not_fill(dbuf, tx);
1940 
1941 	ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1942 	dl = &db->db_last_dirty->dt.dl;
1943 	encode_embedded_bp_compressed(&dl->dr_overridden_by,
1944 	    data, comp, uncompressed_size, compressed_size);
1945 	BPE_SET_ETYPE(&dl->dr_overridden_by, etype);
1946 	BP_SET_TYPE(&dl->dr_overridden_by, type);
1947 	BP_SET_LEVEL(&dl->dr_overridden_by, 0);
1948 	BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder);
1949 
1950 	dl->dr_override_state = DR_OVERRIDDEN;
1951 	dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg;
1952 }
1953 
1954 /*
1955  * Directly assign a provided arc buf to a given dbuf if it's not referenced
1956  * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1957  */
1958 void
1959 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
1960 {
1961 	ASSERT(!refcount_is_zero(&db->db_holds));
1962 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1963 	ASSERT(db->db_level == 0);
1964 	ASSERT3U(dbuf_is_metadata(db), ==, arc_is_metadata(buf));
1965 	ASSERT(buf != NULL);
1966 	ASSERT(arc_buf_lsize(buf) == db->db.db_size);
1967 	ASSERT(tx->tx_txg != 0);
1968 
1969 	arc_return_buf(buf, db);
1970 	ASSERT(arc_released(buf));
1971 
1972 	mutex_enter(&db->db_mtx);
1973 
1974 	while (db->db_state == DB_READ || db->db_state == DB_FILL)
1975 		cv_wait(&db->db_changed, &db->db_mtx);
1976 
1977 	ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
1978 
1979 	if (db->db_state == DB_CACHED &&
1980 	    refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
1981 		mutex_exit(&db->db_mtx);
1982 		(void) dbuf_dirty(db, tx);
1983 		bcopy(buf->b_data, db->db.db_data, db->db.db_size);
1984 		arc_buf_destroy(buf, db);
1985 		xuio_stat_wbuf_copied();
1986 		return;
1987 	}
1988 
1989 	xuio_stat_wbuf_nocopy();
1990 	if (db->db_state == DB_CACHED) {
1991 		dbuf_dirty_record_t *dr = db->db_last_dirty;
1992 
1993 		ASSERT(db->db_buf != NULL);
1994 		if (dr != NULL && dr->dr_txg == tx->tx_txg) {
1995 			ASSERT(dr->dt.dl.dr_data == db->db_buf);
1996 			if (!arc_released(db->db_buf)) {
1997 				ASSERT(dr->dt.dl.dr_override_state ==
1998 				    DR_OVERRIDDEN);
1999 				arc_release(db->db_buf, db);
2000 			}
2001 			dr->dt.dl.dr_data = buf;
2002 			arc_buf_destroy(db->db_buf, db);
2003 		} else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
2004 			arc_release(db->db_buf, db);
2005 			arc_buf_destroy(db->db_buf, db);
2006 		}
2007 		db->db_buf = NULL;
2008 	}
2009 	ASSERT(db->db_buf == NULL);
2010 	dbuf_set_data(db, buf);
2011 	db->db_state = DB_FILL;
2012 	mutex_exit(&db->db_mtx);
2013 	(void) dbuf_dirty(db, tx);
2014 	dmu_buf_fill_done(&db->db, tx);
2015 }
2016 
2017 void
2018 dbuf_destroy(dmu_buf_impl_t *db)
2019 {
2020 	dnode_t *dn;
2021 	dmu_buf_impl_t *parent = db->db_parent;
2022 	dmu_buf_impl_t *dndb;
2023 
2024 	ASSERT(MUTEX_HELD(&db->db_mtx));
2025 	ASSERT(refcount_is_zero(&db->db_holds));
2026 
2027 	if (db->db_buf != NULL) {
2028 		arc_buf_destroy(db->db_buf, db);
2029 		db->db_buf = NULL;
2030 	}
2031 
2032 	if (db->db_blkid == DMU_BONUS_BLKID) {
2033 		ASSERT(db->db.db_data != NULL);
2034 		zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
2035 		arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
2036 		db->db_state = DB_UNCACHED;
2037 	}
2038 
2039 	dbuf_clear_data(db);
2040 
2041 	if (multilist_link_active(&db->db_cache_link)) {
2042 		multilist_remove(dbuf_cache, db);
2043 		(void) refcount_remove_many(&dbuf_cache_size,
2044 		    db->db.db_size, db);
2045 	}
2046 
2047 	ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
2048 	ASSERT(db->db_data_pending == NULL);
2049 
2050 	db->db_state = DB_EVICTING;
2051 	db->db_blkptr = NULL;
2052 
2053 	/*
2054 	 * Now that db_state is DB_EVICTING, nobody else can find this via
2055 	 * the hash table.  We can now drop db_mtx, which allows us to
2056 	 * acquire the dn_dbufs_mtx.
2057 	 */
2058 	mutex_exit(&db->db_mtx);
2059 
2060 	DB_DNODE_ENTER(db);
2061 	dn = DB_DNODE(db);
2062 	dndb = dn->dn_dbuf;
2063 	if (db->db_blkid != DMU_BONUS_BLKID) {
2064 		boolean_t needlock = !MUTEX_HELD(&dn->dn_dbufs_mtx);
2065 		if (needlock)
2066 			mutex_enter(&dn->dn_dbufs_mtx);
2067 		avl_remove(&dn->dn_dbufs, db);
2068 		atomic_dec_32(&dn->dn_dbufs_count);
2069 		membar_producer();
2070 		DB_DNODE_EXIT(db);
2071 		if (needlock)
2072 			mutex_exit(&dn->dn_dbufs_mtx);
2073 		/*
2074 		 * Decrementing the dbuf count means that the hold corresponding
2075 		 * to the removed dbuf is no longer discounted in dnode_move(),
2076 		 * so the dnode cannot be moved until after we release the hold.
2077 		 * The membar_producer() ensures visibility of the decremented
2078 		 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
2079 		 * release any lock.
2080 		 */
2081 		dnode_rele(dn, db);
2082 		db->db_dnode_handle = NULL;
2083 
2084 		dbuf_hash_remove(db);
2085 	} else {
2086 		DB_DNODE_EXIT(db);
2087 	}
2088 
2089 	ASSERT(refcount_is_zero(&db->db_holds));
2090 
2091 	db->db_parent = NULL;
2092 
2093 	ASSERT(db->db_buf == NULL);
2094 	ASSERT(db->db.db_data == NULL);
2095 	ASSERT(db->db_hash_next == NULL);
2096 	ASSERT(db->db_blkptr == NULL);
2097 	ASSERT(db->db_data_pending == NULL);
2098 	ASSERT(!multilist_link_active(&db->db_cache_link));
2099 
2100 	kmem_cache_free(dbuf_kmem_cache, db);
2101 	arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2102 
2103 	/*
2104 	 * If this dbuf is referenced from an indirect dbuf,
2105 	 * decrement the ref count on the indirect dbuf.
2106 	 */
2107 	if (parent && parent != dndb)
2108 		dbuf_rele(parent, db);
2109 }
2110 
2111 /*
2112  * Note: While bpp will always be updated if the function returns success,
2113  * parentp will not be updated if the dnode does not have dn_dbuf filled in;
2114  * this happens when the dnode is the meta-dnode, or a userused or groupused
2115  * object.
2116  */
2117 static int
2118 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
2119     dmu_buf_impl_t **parentp, blkptr_t **bpp)
2120 {
2121 	*parentp = NULL;
2122 	*bpp = NULL;
2123 
2124 	ASSERT(blkid != DMU_BONUS_BLKID);
2125 
2126 	if (blkid == DMU_SPILL_BLKID) {
2127 		mutex_enter(&dn->dn_mtx);
2128 		if (dn->dn_have_spill &&
2129 		    (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
2130 			*bpp = &dn->dn_phys->dn_spill;
2131 		else
2132 			*bpp = NULL;
2133 		dbuf_add_ref(dn->dn_dbuf, NULL);
2134 		*parentp = dn->dn_dbuf;
2135 		mutex_exit(&dn->dn_mtx);
2136 		return (0);
2137 	}
2138 
2139 	int nlevels =
2140 	    (dn->dn_phys->dn_nlevels == 0) ? 1 : dn->dn_phys->dn_nlevels;
2141 	int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2142 
2143 	ASSERT3U(level * epbs, <, 64);
2144 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2145 	/*
2146 	 * This assertion shouldn't trip as long as the max indirect block size
2147 	 * is less than 1M.  The reason for this is that up to that point,
2148 	 * the number of levels required to address an entire object with blocks
2149 	 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64.  In
2150 	 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55
2151 	 * (i.e. we can address the entire object), objects will all use at most
2152 	 * N-1 levels and the assertion won't overflow.  However, once epbs is
2153 	 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66.  Then, 4 levels will not be
2154 	 * enough to address an entire object, so objects will have 5 levels,
2155 	 * but then this assertion will overflow.
2156 	 *
2157 	 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we
2158 	 * need to redo this logic to handle overflows.
2159 	 */
2160 	ASSERT(level >= nlevels ||
2161 	    ((nlevels - level - 1) * epbs) +
2162 	    highbit64(dn->dn_phys->dn_nblkptr) <= 64);
2163 	if (level >= nlevels ||
2164 	    blkid >= ((uint64_t)dn->dn_phys->dn_nblkptr <<
2165 	    ((nlevels - level - 1) * epbs)) ||
2166 	    (fail_sparse &&
2167 	    blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
2168 		/* the buffer has no parent yet */
2169 		return (SET_ERROR(ENOENT));
2170 	} else if (level < nlevels-1) {
2171 		/* this block is referenced from an indirect block */
2172 		int err = dbuf_hold_impl(dn, level+1,
2173 		    blkid >> epbs, fail_sparse, FALSE, NULL, parentp);
2174 		if (err)
2175 			return (err);
2176 		err = dbuf_read(*parentp, NULL,
2177 		    (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
2178 		if (err) {
2179 			dbuf_rele(*parentp, NULL);
2180 			*parentp = NULL;
2181 			return (err);
2182 		}
2183 		*bpp = ((blkptr_t *)(*parentp)->db.db_data) +
2184 		    (blkid & ((1ULL << epbs) - 1));
2185 		if (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))
2186 			ASSERT(BP_IS_HOLE(*bpp));
2187 		return (0);
2188 	} else {
2189 		/* the block is referenced from the dnode */
2190 		ASSERT3U(level, ==, nlevels-1);
2191 		ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
2192 		    blkid < dn->dn_phys->dn_nblkptr);
2193 		if (dn->dn_dbuf) {
2194 			dbuf_add_ref(dn->dn_dbuf, NULL);
2195 			*parentp = dn->dn_dbuf;
2196 		}
2197 		*bpp = &dn->dn_phys->dn_blkptr[blkid];
2198 		return (0);
2199 	}
2200 }
2201 
2202 static dmu_buf_impl_t *
2203 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
2204     dmu_buf_impl_t *parent, blkptr_t *blkptr)
2205 {
2206 	objset_t *os = dn->dn_objset;
2207 	dmu_buf_impl_t *db, *odb;
2208 
2209 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2210 	ASSERT(dn->dn_type != DMU_OT_NONE);
2211 
2212 	db = kmem_cache_alloc(dbuf_kmem_cache, KM_SLEEP);
2213 
2214 	db->db_objset = os;
2215 	db->db.db_object = dn->dn_object;
2216 	db->db_level = level;
2217 	db->db_blkid = blkid;
2218 	db->db_last_dirty = NULL;
2219 	db->db_dirtycnt = 0;
2220 	db->db_dnode_handle = dn->dn_handle;
2221 	db->db_parent = parent;
2222 	db->db_blkptr = blkptr;
2223 
2224 	db->db_user = NULL;
2225 	db->db_user_immediate_evict = FALSE;
2226 	db->db_freed_in_flight = FALSE;
2227 	db->db_pending_evict = FALSE;
2228 
2229 	if (blkid == DMU_BONUS_BLKID) {
2230 		ASSERT3P(parent, ==, dn->dn_dbuf);
2231 		db->db.db_size = DN_MAX_BONUSLEN -
2232 		    (dn->dn_nblkptr-1) * sizeof (blkptr_t);
2233 		ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
2234 		db->db.db_offset = DMU_BONUS_BLKID;
2235 		db->db_state = DB_UNCACHED;
2236 		/* the bonus dbuf is not placed in the hash table */
2237 		arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2238 		return (db);
2239 	} else if (blkid == DMU_SPILL_BLKID) {
2240 		db->db.db_size = (blkptr != NULL) ?
2241 		    BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
2242 		db->db.db_offset = 0;
2243 	} else {
2244 		int blocksize =
2245 		    db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
2246 		db->db.db_size = blocksize;
2247 		db->db.db_offset = db->db_blkid * blocksize;
2248 	}
2249 
2250 	/*
2251 	 * Hold the dn_dbufs_mtx while we get the new dbuf
2252 	 * in the hash table *and* added to the dbufs list.
2253 	 * This prevents a possible deadlock with someone
2254 	 * trying to look up this dbuf before its added to the
2255 	 * dn_dbufs list.
2256 	 */
2257 	mutex_enter(&dn->dn_dbufs_mtx);
2258 	db->db_state = DB_EVICTING;
2259 	if ((odb = dbuf_hash_insert(db)) != NULL) {
2260 		/* someone else inserted it first */
2261 		kmem_cache_free(dbuf_kmem_cache, db);
2262 		mutex_exit(&dn->dn_dbufs_mtx);
2263 		return (odb);
2264 	}
2265 	avl_add(&dn->dn_dbufs, db);
2266 
2267 	db->db_state = DB_UNCACHED;
2268 	mutex_exit(&dn->dn_dbufs_mtx);
2269 	arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2270 
2271 	if (parent && parent != dn->dn_dbuf)
2272 		dbuf_add_ref(parent, db);
2273 
2274 	ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
2275 	    refcount_count(&dn->dn_holds) > 0);
2276 	(void) refcount_add(&dn->dn_holds, db);
2277 	atomic_inc_32(&dn->dn_dbufs_count);
2278 
2279 	dprintf_dbuf(db, "db=%p\n", db);
2280 
2281 	return (db);
2282 }
2283 
2284 typedef struct dbuf_prefetch_arg {
2285 	spa_t *dpa_spa;	/* The spa to issue the prefetch in. */
2286 	zbookmark_phys_t dpa_zb; /* The target block to prefetch. */
2287 	int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */
2288 	int dpa_curlevel; /* The current level that we're reading */
2289 	dnode_t *dpa_dnode; /* The dnode associated with the prefetch */
2290 	zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */
2291 	zio_t *dpa_zio; /* The parent zio_t for all prefetches. */
2292 	arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */
2293 } dbuf_prefetch_arg_t;
2294 
2295 /*
2296  * Actually issue the prefetch read for the block given.
2297  */
2298 static void
2299 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp)
2300 {
2301 	if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
2302 		return;
2303 
2304 	arc_flags_t aflags =
2305 	    dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
2306 
2307 	ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2308 	ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level);
2309 	ASSERT(dpa->dpa_zio != NULL);
2310 	(void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL,
2311 	    dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2312 	    &aflags, &dpa->dpa_zb);
2313 }
2314 
2315 /*
2316  * Called when an indirect block above our prefetch target is read in.  This
2317  * will either read in the next indirect block down the tree or issue the actual
2318  * prefetch if the next block down is our target.
2319  */
2320 static void
2321 dbuf_prefetch_indirect_done(zio_t *zio, arc_buf_t *abuf, void *private)
2322 {
2323 	dbuf_prefetch_arg_t *dpa = private;
2324 
2325 	ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel);
2326 	ASSERT3S(dpa->dpa_curlevel, >, 0);
2327 
2328 	/*
2329 	 * The dpa_dnode is only valid if we are called with a NULL
2330 	 * zio. This indicates that the arc_read() returned without
2331 	 * first calling zio_read() to issue a physical read. Once
2332 	 * a physical read is made the dpa_dnode must be invalidated
2333 	 * as the locks guarding it may have been dropped. If the
2334 	 * dpa_dnode is still valid, then we want to add it to the dbuf
2335 	 * cache. To do so, we must hold the dbuf associated with the block
2336 	 * we just prefetched, read its contents so that we associate it
2337 	 * with an arc_buf_t, and then release it.
2338 	 */
2339 	if (zio != NULL) {
2340 		ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel);
2341 		if (zio->io_flags & ZIO_FLAG_RAW) {
2342 			ASSERT3U(BP_GET_PSIZE(zio->io_bp), ==, zio->io_size);
2343 		} else {
2344 			ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size);
2345 		}
2346 		ASSERT3P(zio->io_spa, ==, dpa->dpa_spa);
2347 
2348 		dpa->dpa_dnode = NULL;
2349 	} else if (dpa->dpa_dnode != NULL) {
2350 		uint64_t curblkid = dpa->dpa_zb.zb_blkid >>
2351 		    (dpa->dpa_epbs * (dpa->dpa_curlevel -
2352 		    dpa->dpa_zb.zb_level));
2353 		dmu_buf_impl_t *db = dbuf_hold_level(dpa->dpa_dnode,
2354 		    dpa->dpa_curlevel, curblkid, FTAG);
2355 		(void) dbuf_read(db, NULL,
2356 		    DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_HAVESTRUCT);
2357 		dbuf_rele(db, FTAG);
2358 	}
2359 
2360 	dpa->dpa_curlevel--;
2361 
2362 	uint64_t nextblkid = dpa->dpa_zb.zb_blkid >>
2363 	    (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level));
2364 	blkptr_t *bp = ((blkptr_t *)abuf->b_data) +
2365 	    P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs);
2366 	if (BP_IS_HOLE(bp) || (zio != NULL && zio->io_error != 0)) {
2367 		kmem_free(dpa, sizeof (*dpa));
2368 	} else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) {
2369 		ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid);
2370 		dbuf_issue_final_prefetch(dpa, bp);
2371 		kmem_free(dpa, sizeof (*dpa));
2372 	} else {
2373 		arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2374 		zbookmark_phys_t zb;
2375 
2376 		ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2377 
2378 		SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset,
2379 		    dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid);
2380 
2381 		(void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2382 		    bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio,
2383 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2384 		    &iter_aflags, &zb);
2385 	}
2386 
2387 	arc_buf_destroy(abuf, private);
2388 }
2389 
2390 /*
2391  * Issue prefetch reads for the given block on the given level.  If the indirect
2392  * blocks above that block are not in memory, we will read them in
2393  * asynchronously.  As a result, this call never blocks waiting for a read to
2394  * complete.
2395  */
2396 void
2397 dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio,
2398     arc_flags_t aflags)
2399 {
2400 	blkptr_t bp;
2401 	int epbs, nlevels, curlevel;
2402 	uint64_t curblkid;
2403 
2404 	ASSERT(blkid != DMU_BONUS_BLKID);
2405 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2406 
2407 	if (blkid > dn->dn_maxblkid)
2408 		return;
2409 
2410 	if (dnode_block_freed(dn, blkid))
2411 		return;
2412 
2413 	/*
2414 	 * This dnode hasn't been written to disk yet, so there's nothing to
2415 	 * prefetch.
2416 	 */
2417 	nlevels = dn->dn_phys->dn_nlevels;
2418 	if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0)
2419 		return;
2420 
2421 	epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2422 	if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level))
2423 		return;
2424 
2425 	dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object,
2426 	    level, blkid);
2427 	if (db != NULL) {
2428 		mutex_exit(&db->db_mtx);
2429 		/*
2430 		 * This dbuf already exists.  It is either CACHED, or
2431 		 * (we assume) about to be read or filled.
2432 		 */
2433 		return;
2434 	}
2435 
2436 	/*
2437 	 * Find the closest ancestor (indirect block) of the target block
2438 	 * that is present in the cache.  In this indirect block, we will
2439 	 * find the bp that is at curlevel, curblkid.
2440 	 */
2441 	curlevel = level;
2442 	curblkid = blkid;
2443 	while (curlevel < nlevels - 1) {
2444 		int parent_level = curlevel + 1;
2445 		uint64_t parent_blkid = curblkid >> epbs;
2446 		dmu_buf_impl_t *db;
2447 
2448 		if (dbuf_hold_impl(dn, parent_level, parent_blkid,
2449 		    FALSE, TRUE, FTAG, &db) == 0) {
2450 			blkptr_t *bpp = db->db_buf->b_data;
2451 			bp = bpp[P2PHASE(curblkid, 1 << epbs)];
2452 			dbuf_rele(db, FTAG);
2453 			break;
2454 		}
2455 
2456 		curlevel = parent_level;
2457 		curblkid = parent_blkid;
2458 	}
2459 
2460 	if (curlevel == nlevels - 1) {
2461 		/* No cached indirect blocks found. */
2462 		ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr);
2463 		bp = dn->dn_phys->dn_blkptr[curblkid];
2464 	}
2465 	if (BP_IS_HOLE(&bp))
2466 		return;
2467 
2468 	ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp));
2469 
2470 	zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL,
2471 	    ZIO_FLAG_CANFAIL);
2472 
2473 	dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP);
2474 	dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
2475 	SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2476 	    dn->dn_object, level, blkid);
2477 	dpa->dpa_curlevel = curlevel;
2478 	dpa->dpa_prio = prio;
2479 	dpa->dpa_aflags = aflags;
2480 	dpa->dpa_spa = dn->dn_objset->os_spa;
2481 	dpa->dpa_dnode = dn;
2482 	dpa->dpa_epbs = epbs;
2483 	dpa->dpa_zio = pio;
2484 
2485 	/*
2486 	 * If we have the indirect just above us, no need to do the asynchronous
2487 	 * prefetch chain; we'll just run the last step ourselves.  If we're at
2488 	 * a higher level, though, we want to issue the prefetches for all the
2489 	 * indirect blocks asynchronously, so we can go on with whatever we were
2490 	 * doing.
2491 	 */
2492 	if (curlevel == level) {
2493 		ASSERT3U(curblkid, ==, blkid);
2494 		dbuf_issue_final_prefetch(dpa, &bp);
2495 		kmem_free(dpa, sizeof (*dpa));
2496 	} else {
2497 		arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2498 		zbookmark_phys_t zb;
2499 
2500 		SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2501 		    dn->dn_object, curlevel, curblkid);
2502 		(void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2503 		    &bp, dbuf_prefetch_indirect_done, dpa, prio,
2504 		    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2505 		    &iter_aflags, &zb);
2506 	}
2507 	/*
2508 	 * We use pio here instead of dpa_zio since it's possible that
2509 	 * dpa may have already been freed.
2510 	 */
2511 	zio_nowait(pio);
2512 }
2513 
2514 /*
2515  * Returns with db_holds incremented, and db_mtx not held.
2516  * Note: dn_struct_rwlock must be held.
2517  */
2518 int
2519 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid,
2520     boolean_t fail_sparse, boolean_t fail_uncached,
2521     void *tag, dmu_buf_impl_t **dbp)
2522 {
2523 	dmu_buf_impl_t *db, *parent = NULL;
2524 
2525 	ASSERT(blkid != DMU_BONUS_BLKID);
2526 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2527 	ASSERT3U(dn->dn_nlevels, >, level);
2528 
2529 	*dbp = NULL;
2530 top:
2531 	/* dbuf_find() returns with db_mtx held */
2532 	db = dbuf_find(dn->dn_objset, dn->dn_object, level, blkid);
2533 
2534 	if (db == NULL) {
2535 		blkptr_t *bp = NULL;
2536 		int err;
2537 
2538 		if (fail_uncached)
2539 			return (SET_ERROR(ENOENT));
2540 
2541 		ASSERT3P(parent, ==, NULL);
2542 		err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
2543 		if (fail_sparse) {
2544 			if (err == 0 && bp && BP_IS_HOLE(bp))
2545 				err = SET_ERROR(ENOENT);
2546 			if (err) {
2547 				if (parent)
2548 					dbuf_rele(parent, NULL);
2549 				return (err);
2550 			}
2551 		}
2552 		if (err && err != ENOENT)
2553 			return (err);
2554 		db = dbuf_create(dn, level, blkid, parent, bp);
2555 	}
2556 
2557 	if (fail_uncached && db->db_state != DB_CACHED) {
2558 		mutex_exit(&db->db_mtx);
2559 		return (SET_ERROR(ENOENT));
2560 	}
2561 
2562 	if (db->db_buf != NULL)
2563 		ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
2564 
2565 	ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
2566 
2567 	/*
2568 	 * If this buffer is currently syncing out, and we are are
2569 	 * still referencing it from db_data, we need to make a copy
2570 	 * of it in case we decide we want to dirty it again in this txg.
2571 	 */
2572 	if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
2573 	    dn->dn_object != DMU_META_DNODE_OBJECT &&
2574 	    db->db_state == DB_CACHED && db->db_data_pending) {
2575 		dbuf_dirty_record_t *dr = db->db_data_pending;
2576 
2577 		if (dr->dt.dl.dr_data == db->db_buf) {
2578 			arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2579 
2580 			dbuf_set_data(db,
2581 			    arc_alloc_buf(dn->dn_objset->os_spa, db, type,
2582 			    db->db.db_size));
2583 			bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
2584 			    db->db.db_size);
2585 		}
2586 	}
2587 
2588 	if (multilist_link_active(&db->db_cache_link)) {
2589 		ASSERT(refcount_is_zero(&db->db_holds));
2590 		multilist_remove(dbuf_cache, db);
2591 		(void) refcount_remove_many(&dbuf_cache_size,
2592 		    db->db.db_size, db);
2593 	}
2594 	(void) refcount_add(&db->db_holds, tag);
2595 	DBUF_VERIFY(db);
2596 	mutex_exit(&db->db_mtx);
2597 
2598 	/* NOTE: we can't rele the parent until after we drop the db_mtx */
2599 	if (parent)
2600 		dbuf_rele(parent, NULL);
2601 
2602 	ASSERT3P(DB_DNODE(db), ==, dn);
2603 	ASSERT3U(db->db_blkid, ==, blkid);
2604 	ASSERT3U(db->db_level, ==, level);
2605 	*dbp = db;
2606 
2607 	return (0);
2608 }
2609 
2610 dmu_buf_impl_t *
2611 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
2612 {
2613 	return (dbuf_hold_level(dn, 0, blkid, tag));
2614 }
2615 
2616 dmu_buf_impl_t *
2617 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
2618 {
2619 	dmu_buf_impl_t *db;
2620 	int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db);
2621 	return (err ? NULL : db);
2622 }
2623 
2624 void
2625 dbuf_create_bonus(dnode_t *dn)
2626 {
2627 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
2628 
2629 	ASSERT(dn->dn_bonus == NULL);
2630 	dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
2631 }
2632 
2633 int
2634 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
2635 {
2636 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2637 	dnode_t *dn;
2638 
2639 	if (db->db_blkid != DMU_SPILL_BLKID)
2640 		return (SET_ERROR(ENOTSUP));
2641 	if (blksz == 0)
2642 		blksz = SPA_MINBLOCKSIZE;
2643 	ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset)));
2644 	blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
2645 
2646 	DB_DNODE_ENTER(db);
2647 	dn = DB_DNODE(db);
2648 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2649 	dbuf_new_size(db, blksz, tx);
2650 	rw_exit(&dn->dn_struct_rwlock);
2651 	DB_DNODE_EXIT(db);
2652 
2653 	return (0);
2654 }
2655 
2656 void
2657 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
2658 {
2659 	dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
2660 }
2661 
2662 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2663 void
2664 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
2665 {
2666 	int64_t holds = refcount_add(&db->db_holds, tag);
2667 	ASSERT3S(holds, >, 1);
2668 }
2669 
2670 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
2671 boolean_t
2672 dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid,
2673     void *tag)
2674 {
2675 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2676 	dmu_buf_impl_t *found_db;
2677 	boolean_t result = B_FALSE;
2678 
2679 	if (db->db_blkid == DMU_BONUS_BLKID)
2680 		found_db = dbuf_find_bonus(os, obj);
2681 	else
2682 		found_db = dbuf_find(os, obj, 0, blkid);
2683 
2684 	if (found_db != NULL) {
2685 		if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) {
2686 			(void) refcount_add(&db->db_holds, tag);
2687 			result = B_TRUE;
2688 		}
2689 		mutex_exit(&db->db_mtx);
2690 	}
2691 	return (result);
2692 }
2693 
2694 /*
2695  * If you call dbuf_rele() you had better not be referencing the dnode handle
2696  * unless you have some other direct or indirect hold on the dnode. (An indirect
2697  * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2698  * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2699  * dnode's parent dbuf evicting its dnode handles.
2700  */
2701 void
2702 dbuf_rele(dmu_buf_impl_t *db, void *tag)
2703 {
2704 	mutex_enter(&db->db_mtx);
2705 	dbuf_rele_and_unlock(db, tag);
2706 }
2707 
2708 void
2709 dmu_buf_rele(dmu_buf_t *db, void *tag)
2710 {
2711 	dbuf_rele((dmu_buf_impl_t *)db, tag);
2712 }
2713 
2714 /*
2715  * dbuf_rele() for an already-locked dbuf.  This is necessary to allow
2716  * db_dirtycnt and db_holds to be updated atomically.
2717  */
2718 void
2719 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2720 {
2721 	int64_t holds;
2722 
2723 	ASSERT(MUTEX_HELD(&db->db_mtx));
2724 	DBUF_VERIFY(db);
2725 
2726 	/*
2727 	 * Remove the reference to the dbuf before removing its hold on the
2728 	 * dnode so we can guarantee in dnode_move() that a referenced bonus
2729 	 * buffer has a corresponding dnode hold.
2730 	 */
2731 	holds = refcount_remove(&db->db_holds, tag);
2732 	ASSERT(holds >= 0);
2733 
2734 	/*
2735 	 * We can't freeze indirects if there is a possibility that they
2736 	 * may be modified in the current syncing context.
2737 	 */
2738 	if (db->db_buf != NULL &&
2739 	    holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) {
2740 		arc_buf_freeze(db->db_buf);
2741 	}
2742 
2743 	if (holds == db->db_dirtycnt &&
2744 	    db->db_level == 0 && db->db_user_immediate_evict)
2745 		dbuf_evict_user(db);
2746 
2747 	if (holds == 0) {
2748 		if (db->db_blkid == DMU_BONUS_BLKID) {
2749 			dnode_t *dn;
2750 			boolean_t evict_dbuf = db->db_pending_evict;
2751 
2752 			/*
2753 			 * If the dnode moves here, we cannot cross this
2754 			 * barrier until the move completes.
2755 			 */
2756 			DB_DNODE_ENTER(db);
2757 
2758 			dn = DB_DNODE(db);
2759 			atomic_dec_32(&dn->dn_dbufs_count);
2760 
2761 			/*
2762 			 * Decrementing the dbuf count means that the bonus
2763 			 * buffer's dnode hold is no longer discounted in
2764 			 * dnode_move(). The dnode cannot move until after
2765 			 * the dnode_rele() below.
2766 			 */
2767 			DB_DNODE_EXIT(db);
2768 
2769 			/*
2770 			 * Do not reference db after its lock is dropped.
2771 			 * Another thread may evict it.
2772 			 */
2773 			mutex_exit(&db->db_mtx);
2774 
2775 			if (evict_dbuf)
2776 				dnode_evict_bonus(dn);
2777 
2778 			dnode_rele(dn, db);
2779 		} else if (db->db_buf == NULL) {
2780 			/*
2781 			 * This is a special case: we never associated this
2782 			 * dbuf with any data allocated from the ARC.
2783 			 */
2784 			ASSERT(db->db_state == DB_UNCACHED ||
2785 			    db->db_state == DB_NOFILL);
2786 			dbuf_destroy(db);
2787 		} else if (arc_released(db->db_buf)) {
2788 			/*
2789 			 * This dbuf has anonymous data associated with it.
2790 			 */
2791 			dbuf_destroy(db);
2792 		} else {
2793 			boolean_t do_arc_evict = B_FALSE;
2794 			blkptr_t bp;
2795 			spa_t *spa = dmu_objset_spa(db->db_objset);
2796 
2797 			if (!DBUF_IS_CACHEABLE(db) &&
2798 			    db->db_blkptr != NULL &&
2799 			    !BP_IS_HOLE(db->db_blkptr) &&
2800 			    !BP_IS_EMBEDDED(db->db_blkptr)) {
2801 				do_arc_evict = B_TRUE;
2802 				bp = *db->db_blkptr;
2803 			}
2804 
2805 			if (!DBUF_IS_CACHEABLE(db) ||
2806 			    db->db_pending_evict) {
2807 				dbuf_destroy(db);
2808 			} else if (!multilist_link_active(&db->db_cache_link)) {
2809 				multilist_insert(dbuf_cache, db);
2810 				(void) refcount_add_many(&dbuf_cache_size,
2811 				    db->db.db_size, db);
2812 				mutex_exit(&db->db_mtx);
2813 
2814 				dbuf_evict_notify();
2815 			}
2816 
2817 			if (do_arc_evict)
2818 				arc_freed(spa, &bp);
2819 		}
2820 	} else {
2821 		mutex_exit(&db->db_mtx);
2822 	}
2823 
2824 }
2825 
2826 #pragma weak dmu_buf_refcount = dbuf_refcount
2827 uint64_t
2828 dbuf_refcount(dmu_buf_impl_t *db)
2829 {
2830 	return (refcount_count(&db->db_holds));
2831 }
2832 
2833 void *
2834 dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user,
2835     dmu_buf_user_t *new_user)
2836 {
2837 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2838 
2839 	mutex_enter(&db->db_mtx);
2840 	dbuf_verify_user(db, DBVU_NOT_EVICTING);
2841 	if (db->db_user == old_user)
2842 		db->db_user = new_user;
2843 	else
2844 		old_user = db->db_user;
2845 	dbuf_verify_user(db, DBVU_NOT_EVICTING);
2846 	mutex_exit(&db->db_mtx);
2847 
2848 	return (old_user);
2849 }
2850 
2851 void *
2852 dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2853 {
2854 	return (dmu_buf_replace_user(db_fake, NULL, user));
2855 }
2856 
2857 void *
2858 dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2859 {
2860 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2861 
2862 	db->db_user_immediate_evict = TRUE;
2863 	return (dmu_buf_set_user(db_fake, user));
2864 }
2865 
2866 void *
2867 dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2868 {
2869 	return (dmu_buf_replace_user(db_fake, user, NULL));
2870 }
2871 
2872 void *
2873 dmu_buf_get_user(dmu_buf_t *db_fake)
2874 {
2875 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2876 
2877 	dbuf_verify_user(db, DBVU_NOT_EVICTING);
2878 	return (db->db_user);
2879 }
2880 
2881 void
2882 dmu_buf_user_evict_wait()
2883 {
2884 	taskq_wait(dbu_evict_taskq);
2885 }
2886 
2887 blkptr_t *
2888 dmu_buf_get_blkptr(dmu_buf_t *db)
2889 {
2890 	dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2891 	return (dbi->db_blkptr);
2892 }
2893 
2894 objset_t *
2895 dmu_buf_get_objset(dmu_buf_t *db)
2896 {
2897 	dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2898 	return (dbi->db_objset);
2899 }
2900 
2901 dnode_t *
2902 dmu_buf_dnode_enter(dmu_buf_t *db)
2903 {
2904 	dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2905 	DB_DNODE_ENTER(dbi);
2906 	return (DB_DNODE(dbi));
2907 }
2908 
2909 void
2910 dmu_buf_dnode_exit(dmu_buf_t *db)
2911 {
2912 	dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2913 	DB_DNODE_EXIT(dbi);
2914 }
2915 
2916 static void
2917 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
2918 {
2919 	/* ASSERT(dmu_tx_is_syncing(tx) */
2920 	ASSERT(MUTEX_HELD(&db->db_mtx));
2921 
2922 	if (db->db_blkptr != NULL)
2923 		return;
2924 
2925 	if (db->db_blkid == DMU_SPILL_BLKID) {
2926 		db->db_blkptr = &dn->dn_phys->dn_spill;
2927 		BP_ZERO(db->db_blkptr);
2928 		return;
2929 	}
2930 	if (db->db_level == dn->dn_phys->dn_nlevels-1) {
2931 		/*
2932 		 * This buffer was allocated at a time when there was
2933 		 * no available blkptrs from the dnode, or it was
2934 		 * inappropriate to hook it in (i.e., nlevels mis-match).
2935 		 */
2936 		ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
2937 		ASSERT(db->db_parent == NULL);
2938 		db->db_parent = dn->dn_dbuf;
2939 		db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
2940 		DBUF_VERIFY(db);
2941 	} else {
2942 		dmu_buf_impl_t *parent = db->db_parent;
2943 		int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2944 
2945 		ASSERT(dn->dn_phys->dn_nlevels > 1);
2946 		if (parent == NULL) {
2947 			mutex_exit(&db->db_mtx);
2948 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
2949 			parent = dbuf_hold_level(dn, db->db_level + 1,
2950 			    db->db_blkid >> epbs, db);
2951 			rw_exit(&dn->dn_struct_rwlock);
2952 			mutex_enter(&db->db_mtx);
2953 			db->db_parent = parent;
2954 		}
2955 		db->db_blkptr = (blkptr_t *)parent->db.db_data +
2956 		    (db->db_blkid & ((1ULL << epbs) - 1));
2957 		DBUF_VERIFY(db);
2958 	}
2959 }
2960 
2961 static void
2962 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2963 {
2964 	dmu_buf_impl_t *db = dr->dr_dbuf;
2965 	dnode_t *dn;
2966 	zio_t *zio;
2967 
2968 	ASSERT(dmu_tx_is_syncing(tx));
2969 
2970 	dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2971 
2972 	mutex_enter(&db->db_mtx);
2973 
2974 	ASSERT(db->db_level > 0);
2975 	DBUF_VERIFY(db);
2976 
2977 	/* Read the block if it hasn't been read yet. */
2978 	if (db->db_buf == NULL) {
2979 		mutex_exit(&db->db_mtx);
2980 		(void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
2981 		mutex_enter(&db->db_mtx);
2982 	}
2983 	ASSERT3U(db->db_state, ==, DB_CACHED);
2984 	ASSERT(db->db_buf != NULL);
2985 
2986 	DB_DNODE_ENTER(db);
2987 	dn = DB_DNODE(db);
2988 	/* Indirect block size must match what the dnode thinks it is. */
2989 	ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2990 	dbuf_check_blkptr(dn, db);
2991 	DB_DNODE_EXIT(db);
2992 
2993 	/* Provide the pending dirty record to child dbufs */
2994 	db->db_data_pending = dr;
2995 
2996 	mutex_exit(&db->db_mtx);
2997 	dbuf_write(dr, db->db_buf, tx);
2998 
2999 	zio = dr->dr_zio;
3000 	mutex_enter(&dr->dt.di.dr_mtx);
3001 	dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx);
3002 	ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3003 	mutex_exit(&dr->dt.di.dr_mtx);
3004 	zio_nowait(zio);
3005 }
3006 
3007 static void
3008 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
3009 {
3010 	arc_buf_t **datap = &dr->dt.dl.dr_data;
3011 	dmu_buf_impl_t *db = dr->dr_dbuf;
3012 	dnode_t *dn;
3013 	objset_t *os;
3014 	uint64_t txg = tx->tx_txg;
3015 
3016 	ASSERT(dmu_tx_is_syncing(tx));
3017 
3018 	dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
3019 
3020 	mutex_enter(&db->db_mtx);
3021 	/*
3022 	 * To be synced, we must be dirtied.  But we
3023 	 * might have been freed after the dirty.
3024 	 */
3025 	if (db->db_state == DB_UNCACHED) {
3026 		/* This buffer has been freed since it was dirtied */
3027 		ASSERT(db->db.db_data == NULL);
3028 	} else if (db->db_state == DB_FILL) {
3029 		/* This buffer was freed and is now being re-filled */
3030 		ASSERT(db->db.db_data != dr->dt.dl.dr_data);
3031 	} else {
3032 		ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
3033 	}
3034 	DBUF_VERIFY(db);
3035 
3036 	DB_DNODE_ENTER(db);
3037 	dn = DB_DNODE(db);
3038 
3039 	if (db->db_blkid == DMU_SPILL_BLKID) {
3040 		mutex_enter(&dn->dn_mtx);
3041 		dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
3042 		mutex_exit(&dn->dn_mtx);
3043 	}
3044 
3045 	/*
3046 	 * If this is a bonus buffer, simply copy the bonus data into the
3047 	 * dnode.  It will be written out when the dnode is synced (and it
3048 	 * will be synced, since it must have been dirty for dbuf_sync to
3049 	 * be called).
3050 	 */
3051 	if (db->db_blkid == DMU_BONUS_BLKID) {
3052 		dbuf_dirty_record_t **drp;
3053 
3054 		ASSERT(*datap != NULL);
3055 		ASSERT0(db->db_level);
3056 		ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN);
3057 		bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen);
3058 		DB_DNODE_EXIT(db);
3059 
3060 		if (*datap != db->db.db_data) {
3061 			zio_buf_free(*datap, DN_MAX_BONUSLEN);
3062 			arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
3063 		}
3064 		db->db_data_pending = NULL;
3065 		drp = &db->db_last_dirty;
3066 		while (*drp != dr)
3067 			drp = &(*drp)->dr_next;
3068 		ASSERT(dr->dr_next == NULL);
3069 		ASSERT(dr->dr_dbuf == db);
3070 		*drp = dr->dr_next;
3071 		kmem_free(dr, sizeof (dbuf_dirty_record_t));
3072 		ASSERT(db->db_dirtycnt > 0);
3073 		db->db_dirtycnt -= 1;
3074 		dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
3075 		return;
3076 	}
3077 
3078 	os = dn->dn_objset;
3079 
3080 	/*
3081 	 * This function may have dropped the db_mtx lock allowing a dmu_sync
3082 	 * operation to sneak in. As a result, we need to ensure that we
3083 	 * don't check the dr_override_state until we have returned from
3084 	 * dbuf_check_blkptr.
3085 	 */
3086 	dbuf_check_blkptr(dn, db);
3087 
3088 	/*
3089 	 * If this buffer is in the middle of an immediate write,
3090 	 * wait for the synchronous IO to complete.
3091 	 */
3092 	while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
3093 		ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
3094 		cv_wait(&db->db_changed, &db->db_mtx);
3095 		ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
3096 	}
3097 
3098 	if (db->db_state != DB_NOFILL &&
3099 	    dn->dn_object != DMU_META_DNODE_OBJECT &&
3100 	    refcount_count(&db->db_holds) > 1 &&
3101 	    dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
3102 	    *datap == db->db_buf) {
3103 		/*
3104 		 * If this buffer is currently "in use" (i.e., there
3105 		 * are active holds and db_data still references it),
3106 		 * then make a copy before we start the write so that
3107 		 * any modifications from the open txg will not leak
3108 		 * into this write.
3109 		 *
3110 		 * NOTE: this copy does not need to be made for
3111 		 * objects only modified in the syncing context (e.g.
3112 		 * DNONE_DNODE blocks).
3113 		 */
3114 		int psize = arc_buf_size(*datap);
3115 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
3116 		enum zio_compress compress_type = arc_get_compression(*datap);
3117 
3118 		if (compress_type == ZIO_COMPRESS_OFF) {
3119 			*datap = arc_alloc_buf(os->os_spa, db, type, psize);
3120 		} else {
3121 			ASSERT3U(type, ==, ARC_BUFC_DATA);
3122 			int lsize = arc_buf_lsize(*datap);
3123 			*datap = arc_alloc_compressed_buf(os->os_spa, db,
3124 			    psize, lsize, compress_type);
3125 		}
3126 		bcopy(db->db.db_data, (*datap)->b_data, psize);
3127 	}
3128 	db->db_data_pending = dr;
3129 
3130 	mutex_exit(&db->db_mtx);
3131 
3132 	dbuf_write(dr, *datap, tx);
3133 
3134 	ASSERT(!list_link_active(&dr->dr_dirty_node));
3135 	if (dn->dn_object == DMU_META_DNODE_OBJECT) {
3136 		list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
3137 		DB_DNODE_EXIT(db);
3138 	} else {
3139 		/*
3140 		 * Although zio_nowait() does not "wait for an IO", it does
3141 		 * initiate the IO. If this is an empty write it seems plausible
3142 		 * that the IO could actually be completed before the nowait
3143 		 * returns. We need to DB_DNODE_EXIT() first in case
3144 		 * zio_nowait() invalidates the dbuf.
3145 		 */
3146 		DB_DNODE_EXIT(db);
3147 		zio_nowait(dr->dr_zio);
3148 	}
3149 }
3150 
3151 void
3152 dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx)
3153 {
3154 	dbuf_dirty_record_t *dr;
3155 
3156 	while (dr = list_head(list)) {
3157 		if (dr->dr_zio != NULL) {
3158 			/*
3159 			 * If we find an already initialized zio then we
3160 			 * are processing the meta-dnode, and we have finished.
3161 			 * The dbufs for all dnodes are put back on the list
3162 			 * during processing, so that we can zio_wait()
3163 			 * these IOs after initiating all child IOs.
3164 			 */
3165 			ASSERT3U(dr->dr_dbuf->db.db_object, ==,
3166 			    DMU_META_DNODE_OBJECT);
3167 			break;
3168 		}
3169 		if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
3170 		    dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
3171 			VERIFY3U(dr->dr_dbuf->db_level, ==, level);
3172 		}
3173 		list_remove(list, dr);
3174 		if (dr->dr_dbuf->db_level > 0)
3175 			dbuf_sync_indirect(dr, tx);
3176 		else
3177 			dbuf_sync_leaf(dr, tx);
3178 	}
3179 }
3180 
3181 /* ARGSUSED */
3182 static void
3183 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3184 {
3185 	dmu_buf_impl_t *db = vdb;
3186 	dnode_t *dn;
3187 	blkptr_t *bp = zio->io_bp;
3188 	blkptr_t *bp_orig = &zio->io_bp_orig;
3189 	spa_t *spa = zio->io_spa;
3190 	int64_t delta;
3191 	uint64_t fill = 0;
3192 	int i;
3193 
3194 	ASSERT3P(db->db_blkptr, !=, NULL);
3195 	ASSERT3P(&db->db_data_pending->dr_bp_copy, ==, bp);
3196 
3197 	DB_DNODE_ENTER(db);
3198 	dn = DB_DNODE(db);
3199 	delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
3200 	dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
3201 	zio->io_prev_space_delta = delta;
3202 
3203 	if (bp->blk_birth != 0) {
3204 		ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
3205 		    BP_GET_TYPE(bp) == dn->dn_type) ||
3206 		    (db->db_blkid == DMU_SPILL_BLKID &&
3207 		    BP_GET_TYPE(bp) == dn->dn_bonustype) ||
3208 		    BP_IS_EMBEDDED(bp));
3209 		ASSERT(BP_GET_LEVEL(bp) == db->db_level);
3210 	}
3211 
3212 	mutex_enter(&db->db_mtx);
3213 
3214 #ifdef ZFS_DEBUG
3215 	if (db->db_blkid == DMU_SPILL_BLKID) {
3216 		ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3217 		ASSERT(!(BP_IS_HOLE(bp)) &&
3218 		    db->db_blkptr == &dn->dn_phys->dn_spill);
3219 	}
3220 #endif
3221 
3222 	if (db->db_level == 0) {
3223 		mutex_enter(&dn->dn_mtx);
3224 		if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
3225 		    db->db_blkid != DMU_SPILL_BLKID)
3226 			dn->dn_phys->dn_maxblkid = db->db_blkid;
3227 		mutex_exit(&dn->dn_mtx);
3228 
3229 		if (dn->dn_type == DMU_OT_DNODE) {
3230 			dnode_phys_t *dnp = db->db.db_data;
3231 			for (i = db->db.db_size >> DNODE_SHIFT; i > 0;
3232 			    i--, dnp++) {
3233 				if (dnp->dn_type != DMU_OT_NONE)
3234 					fill++;
3235 			}
3236 		} else {
3237 			if (BP_IS_HOLE(bp)) {
3238 				fill = 0;
3239 			} else {
3240 				fill = 1;
3241 			}
3242 		}
3243 	} else {
3244 		blkptr_t *ibp = db->db.db_data;
3245 		ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3246 		for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
3247 			if (BP_IS_HOLE(ibp))
3248 				continue;
3249 			fill += BP_GET_FILL(ibp);
3250 		}
3251 	}
3252 	DB_DNODE_EXIT(db);
3253 
3254 	if (!BP_IS_EMBEDDED(bp))
3255 		bp->blk_fill = fill;
3256 
3257 	mutex_exit(&db->db_mtx);
3258 
3259 	rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3260 	*db->db_blkptr = *bp;
3261 	rw_exit(&dn->dn_struct_rwlock);
3262 }
3263 
3264 /* ARGSUSED */
3265 /*
3266  * This function gets called just prior to running through the compression
3267  * stage of the zio pipeline. If we're an indirect block comprised of only
3268  * holes, then we want this indirect to be compressed away to a hole. In
3269  * order to do that we must zero out any information about the holes that
3270  * this indirect points to prior to before we try to compress it.
3271  */
3272 static void
3273 dbuf_write_children_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3274 {
3275 	dmu_buf_impl_t *db = vdb;
3276 	dnode_t *dn;
3277 	blkptr_t *bp;
3278 	unsigned int epbs, i;
3279 
3280 	ASSERT3U(db->db_level, >, 0);
3281 	DB_DNODE_ENTER(db);
3282 	dn = DB_DNODE(db);
3283 	epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3284 	ASSERT3U(epbs, <, 31);
3285 
3286 	/* Determine if all our children are holes */
3287 	for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++) {
3288 		if (!BP_IS_HOLE(bp))
3289 			break;
3290 	}
3291 
3292 	/*
3293 	 * If all the children are holes, then zero them all out so that
3294 	 * we may get compressed away.
3295 	 */
3296 	if (i == 1 << epbs) {
3297 		/*
3298 		 * We only found holes. Grab the rwlock to prevent
3299 		 * anybody from reading the blocks we're about to
3300 		 * zero out.
3301 		 */
3302 		rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3303 		bzero(db->db.db_data, db->db.db_size);
3304 		rw_exit(&dn->dn_struct_rwlock);
3305 	}
3306 	DB_DNODE_EXIT(db);
3307 }
3308 
3309 /*
3310  * The SPA will call this callback several times for each zio - once
3311  * for every physical child i/o (zio->io_phys_children times).  This
3312  * allows the DMU to monitor the progress of each logical i/o.  For example,
3313  * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
3314  * block.  There may be a long delay before all copies/fragments are completed,
3315  * so this callback allows us to retire dirty space gradually, as the physical
3316  * i/os complete.
3317  */
3318 /* ARGSUSED */
3319 static void
3320 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
3321 {
3322 	dmu_buf_impl_t *db = arg;
3323 	objset_t *os = db->db_objset;
3324 	dsl_pool_t *dp = dmu_objset_pool(os);
3325 	dbuf_dirty_record_t *dr;
3326 	int delta = 0;
3327 
3328 	dr = db->db_data_pending;
3329 	ASSERT3U(dr->dr_txg, ==, zio->io_txg);
3330 
3331 	/*
3332 	 * The callback will be called io_phys_children times.  Retire one
3333 	 * portion of our dirty space each time we are called.  Any rounding
3334 	 * error will be cleaned up by dsl_pool_sync()'s call to
3335 	 * dsl_pool_undirty_space().
3336 	 */
3337 	delta = dr->dr_accounted / zio->io_phys_children;
3338 	dsl_pool_undirty_space(dp, delta, zio->io_txg);
3339 }
3340 
3341 /* ARGSUSED */
3342 static void
3343 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
3344 {
3345 	dmu_buf_impl_t *db = vdb;
3346 	blkptr_t *bp_orig = &zio->io_bp_orig;
3347 	blkptr_t *bp = db->db_blkptr;
3348 	objset_t *os = db->db_objset;
3349 	dmu_tx_t *tx = os->os_synctx;
3350 	dbuf_dirty_record_t **drp, *dr;
3351 
3352 	ASSERT0(zio->io_error);
3353 	ASSERT(db->db_blkptr == bp);
3354 
3355 	/*
3356 	 * For nopwrites and rewrites we ensure that the bp matches our
3357 	 * original and bypass all the accounting.
3358 	 */
3359 	if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
3360 		ASSERT(BP_EQUAL(bp, bp_orig));
3361 	} else {
3362 		dsl_dataset_t *ds = os->os_dsl_dataset;
3363 		(void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
3364 		dsl_dataset_block_born(ds, bp, tx);
3365 	}
3366 
3367 	mutex_enter(&db->db_mtx);
3368 
3369 	DBUF_VERIFY(db);
3370 
3371 	drp = &db->db_last_dirty;
3372 	while ((dr = *drp) != db->db_data_pending)
3373 		drp = &dr->dr_next;
3374 	ASSERT(!list_link_active(&dr->dr_dirty_node));
3375 	ASSERT(dr->dr_dbuf == db);
3376 	ASSERT(dr->dr_next == NULL);
3377 	*drp = dr->dr_next;
3378 
3379 #ifdef ZFS_DEBUG
3380 	if (db->db_blkid == DMU_SPILL_BLKID) {
3381 		dnode_t *dn;
3382 
3383 		DB_DNODE_ENTER(db);
3384 		dn = DB_DNODE(db);
3385 		ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3386 		ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
3387 		    db->db_blkptr == &dn->dn_phys->dn_spill);
3388 		DB_DNODE_EXIT(db);
3389 	}
3390 #endif
3391 
3392 	if (db->db_level == 0) {
3393 		ASSERT(db->db_blkid != DMU_BONUS_BLKID);
3394 		ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
3395 		if (db->db_state != DB_NOFILL) {
3396 			if (dr->dt.dl.dr_data != db->db_buf)
3397 				arc_buf_destroy(dr->dt.dl.dr_data, db);
3398 		}
3399 	} else {
3400 		dnode_t *dn;
3401 
3402 		DB_DNODE_ENTER(db);
3403 		dn = DB_DNODE(db);
3404 		ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3405 		ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
3406 		if (!BP_IS_HOLE(db->db_blkptr)) {
3407 			int epbs =
3408 			    dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3409 			ASSERT3U(db->db_blkid, <=,
3410 			    dn->dn_phys->dn_maxblkid >> (db->db_level * epbs));
3411 			ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
3412 			    db->db.db_size);
3413 		}
3414 		DB_DNODE_EXIT(db);
3415 		mutex_destroy(&dr->dt.di.dr_mtx);
3416 		list_destroy(&dr->dt.di.dr_children);
3417 	}
3418 	kmem_free(dr, sizeof (dbuf_dirty_record_t));
3419 
3420 	cv_broadcast(&db->db_changed);
3421 	ASSERT(db->db_dirtycnt > 0);
3422 	db->db_dirtycnt -= 1;
3423 	db->db_data_pending = NULL;
3424 	dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg);
3425 }
3426 
3427 static void
3428 dbuf_write_nofill_ready(zio_t *zio)
3429 {
3430 	dbuf_write_ready(zio, NULL, zio->io_private);
3431 }
3432 
3433 static void
3434 dbuf_write_nofill_done(zio_t *zio)
3435 {
3436 	dbuf_write_done(zio, NULL, zio->io_private);
3437 }
3438 
3439 static void
3440 dbuf_write_override_ready(zio_t *zio)
3441 {
3442 	dbuf_dirty_record_t *dr = zio->io_private;
3443 	dmu_buf_impl_t *db = dr->dr_dbuf;
3444 
3445 	dbuf_write_ready(zio, NULL, db);
3446 }
3447 
3448 static void
3449 dbuf_write_override_done(zio_t *zio)
3450 {
3451 	dbuf_dirty_record_t *dr = zio->io_private;
3452 	dmu_buf_impl_t *db = dr->dr_dbuf;
3453 	blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
3454 
3455 	mutex_enter(&db->db_mtx);
3456 	if (!BP_EQUAL(zio->io_bp, obp)) {
3457 		if (!BP_IS_HOLE(obp))
3458 			dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
3459 		arc_release(dr->dt.dl.dr_data, db);
3460 	}
3461 	mutex_exit(&db->db_mtx);
3462 	dbuf_write_done(zio, NULL, db);
3463 
3464 	if (zio->io_abd != NULL)
3465 		abd_put(zio->io_abd);
3466 }
3467 
3468 /* Issue I/O to commit a dirty buffer to disk. */
3469 static void
3470 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
3471 {
3472 	dmu_buf_impl_t *db = dr->dr_dbuf;
3473 	dnode_t *dn;
3474 	objset_t *os;
3475 	dmu_buf_impl_t *parent = db->db_parent;
3476 	uint64_t txg = tx->tx_txg;
3477 	zbookmark_phys_t zb;
3478 	zio_prop_t zp;
3479 	zio_t *zio;
3480 	int wp_flag = 0;
3481 
3482 	ASSERT(dmu_tx_is_syncing(tx));
3483 
3484 	DB_DNODE_ENTER(db);
3485 	dn = DB_DNODE(db);
3486 	os = dn->dn_objset;
3487 
3488 	if (db->db_state != DB_NOFILL) {
3489 		if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
3490 			/*
3491 			 * Private object buffers are released here rather
3492 			 * than in dbuf_dirty() since they are only modified
3493 			 * in the syncing context and we don't want the
3494 			 * overhead of making multiple copies of the data.
3495 			 */
3496 			if (BP_IS_HOLE(db->db_blkptr)) {
3497 				arc_buf_thaw(data);
3498 			} else {
3499 				dbuf_release_bp(db);
3500 			}
3501 		}
3502 	}
3503 
3504 	if (parent != dn->dn_dbuf) {
3505 		/* Our parent is an indirect block. */
3506 		/* We have a dirty parent that has been scheduled for write. */
3507 		ASSERT(parent && parent->db_data_pending);
3508 		/* Our parent's buffer is one level closer to the dnode. */
3509 		ASSERT(db->db_level == parent->db_level-1);
3510 		/*
3511 		 * We're about to modify our parent's db_data by modifying
3512 		 * our block pointer, so the parent must be released.
3513 		 */
3514 		ASSERT(arc_released(parent->db_buf));
3515 		zio = parent->db_data_pending->dr_zio;
3516 	} else {
3517 		/* Our parent is the dnode itself. */
3518 		ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
3519 		    db->db_blkid != DMU_SPILL_BLKID) ||
3520 		    (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
3521 		if (db->db_blkid != DMU_SPILL_BLKID)
3522 			ASSERT3P(db->db_blkptr, ==,
3523 			    &dn->dn_phys->dn_blkptr[db->db_blkid]);
3524 		zio = dn->dn_zio;
3525 	}
3526 
3527 	ASSERT(db->db_level == 0 || data == db->db_buf);
3528 	ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
3529 	ASSERT(zio);
3530 
3531 	SET_BOOKMARK(&zb, os->os_dsl_dataset ?
3532 	    os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
3533 	    db->db.db_object, db->db_level, db->db_blkid);
3534 
3535 	if (db->db_blkid == DMU_SPILL_BLKID)
3536 		wp_flag = WP_SPILL;
3537 	wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
3538 
3539 	dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
3540 	DB_DNODE_EXIT(db);
3541 
3542 	/*
3543 	 * We copy the blkptr now (rather than when we instantiate the dirty
3544 	 * record), because its value can change between open context and
3545 	 * syncing context. We do not need to hold dn_struct_rwlock to read
3546 	 * db_blkptr because we are in syncing context.
3547 	 */
3548 	dr->dr_bp_copy = *db->db_blkptr;
3549 
3550 	if (db->db_level == 0 &&
3551 	    dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
3552 		/*
3553 		 * The BP for this block has been provided by open context
3554 		 * (by dmu_sync() or dmu_buf_write_embedded()).
3555 		 */
3556 		abd_t *contents = (data != NULL) ?
3557 		    abd_get_from_buf(data->b_data, arc_buf_size(data)) : NULL;
3558 
3559 		dr->dr_zio = zio_write(zio, os->os_spa, txg, &dr->dr_bp_copy,
3560 		    contents, db->db.db_size, db->db.db_size, &zp,
3561 		    dbuf_write_override_ready, NULL, NULL,
3562 		    dbuf_write_override_done,
3563 		    dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
3564 		mutex_enter(&db->db_mtx);
3565 		dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
3566 		zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
3567 		    dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
3568 		mutex_exit(&db->db_mtx);
3569 	} else if (db->db_state == DB_NOFILL) {
3570 		ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF ||
3571 		    zp.zp_checksum == ZIO_CHECKSUM_NOPARITY);
3572 		dr->dr_zio = zio_write(zio, os->os_spa, txg,
3573 		    &dr->dr_bp_copy, NULL, db->db.db_size, db->db.db_size, &zp,
3574 		    dbuf_write_nofill_ready, NULL, NULL,
3575 		    dbuf_write_nofill_done, db,
3576 		    ZIO_PRIORITY_ASYNC_WRITE,
3577 		    ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
3578 	} else {
3579 		ASSERT(arc_released(data));
3580 
3581 		/*
3582 		 * For indirect blocks, we want to setup the children
3583 		 * ready callback so that we can properly handle an indirect
3584 		 * block that only contains holes.
3585 		 */
3586 		arc_done_func_t *children_ready_cb = NULL;
3587 		if (db->db_level != 0)
3588 			children_ready_cb = dbuf_write_children_ready;
3589 
3590 		dr->dr_zio = arc_write(zio, os->os_spa, txg,
3591 		    &dr->dr_bp_copy, data, DBUF_IS_L2CACHEABLE(db),
3592 		    &zp, dbuf_write_ready, children_ready_cb,
3593 		    dbuf_write_physdone, dbuf_write_done, db,
3594 		    ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
3595 	}
3596 }
3597