xref: /illumos-gate/usr/src/uts/common/fs/zfs/dbuf.c (revision a1bf6e2eba3dc9acc46299eac9c6ef1bc1e15c2b)
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  */
24 
25 #include <sys/zfs_context.h>
26 #include <sys/dmu.h>
27 #include <sys/dmu_impl.h>
28 #include <sys/dbuf.h>
29 #include <sys/dmu_objset.h>
30 #include <sys/dsl_dataset.h>
31 #include <sys/dsl_dir.h>
32 #include <sys/dmu_tx.h>
33 #include <sys/spa.h>
34 #include <sys/zio.h>
35 #include <sys/dmu_zfetch.h>
36 #include <sys/sa.h>
37 #include <sys/sa_impl.h>
38 
39 static void dbuf_destroy(dmu_buf_impl_t *db);
40 static int dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
41 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
42 
43 /*
44  * Global data structures and functions for the dbuf cache.
45  */
46 static kmem_cache_t *dbuf_cache;
47 
48 /* ARGSUSED */
49 static int
50 dbuf_cons(void *vdb, void *unused, int kmflag)
51 {
52 	dmu_buf_impl_t *db = vdb;
53 	bzero(db, sizeof (dmu_buf_impl_t));
54 
55 	mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
56 	cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
57 	refcount_create(&db->db_holds);
58 	return (0);
59 }
60 
61 /* ARGSUSED */
62 static void
63 dbuf_dest(void *vdb, void *unused)
64 {
65 	dmu_buf_impl_t *db = vdb;
66 	mutex_destroy(&db->db_mtx);
67 	cv_destroy(&db->db_changed);
68 	refcount_destroy(&db->db_holds);
69 }
70 
71 /*
72  * dbuf hash table routines
73  */
74 static dbuf_hash_table_t dbuf_hash_table;
75 
76 static uint64_t dbuf_hash_count;
77 
78 static uint64_t
79 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
80 {
81 	uintptr_t osv = (uintptr_t)os;
82 	uint64_t crc = -1ULL;
83 
84 	ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
85 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF];
86 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF];
87 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF];
88 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF];
89 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF];
90 	crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF];
91 
92 	crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16);
93 
94 	return (crc);
95 }
96 
97 #define	DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid);
98 
99 #define	DBUF_EQUAL(dbuf, os, obj, level, blkid)		\
100 	((dbuf)->db.db_object == (obj) &&		\
101 	(dbuf)->db_objset == (os) &&			\
102 	(dbuf)->db_level == (level) &&			\
103 	(dbuf)->db_blkid == (blkid))
104 
105 dmu_buf_impl_t *
106 dbuf_find(dnode_t *dn, uint8_t level, uint64_t blkid)
107 {
108 	dbuf_hash_table_t *h = &dbuf_hash_table;
109 	objset_t *os = dn->dn_objset;
110 	uint64_t obj = dn->dn_object;
111 	uint64_t hv = DBUF_HASH(os, obj, level, blkid);
112 	uint64_t idx = hv & h->hash_table_mask;
113 	dmu_buf_impl_t *db;
114 
115 	mutex_enter(DBUF_HASH_MUTEX(h, idx));
116 	for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
117 		if (DBUF_EQUAL(db, os, obj, level, blkid)) {
118 			mutex_enter(&db->db_mtx);
119 			if (db->db_state != DB_EVICTING) {
120 				mutex_exit(DBUF_HASH_MUTEX(h, idx));
121 				return (db);
122 			}
123 			mutex_exit(&db->db_mtx);
124 		}
125 	}
126 	mutex_exit(DBUF_HASH_MUTEX(h, idx));
127 	return (NULL);
128 }
129 
130 /*
131  * Insert an entry into the hash table.  If there is already an element
132  * equal to elem in the hash table, then the already existing element
133  * will be returned and the new element will not be inserted.
134  * Otherwise returns NULL.
135  */
136 static dmu_buf_impl_t *
137 dbuf_hash_insert(dmu_buf_impl_t *db)
138 {
139 	dbuf_hash_table_t *h = &dbuf_hash_table;
140 	objset_t *os = db->db_objset;
141 	uint64_t obj = db->db.db_object;
142 	int level = db->db_level;
143 	uint64_t blkid = db->db_blkid;
144 	uint64_t hv = DBUF_HASH(os, obj, level, blkid);
145 	uint64_t idx = hv & h->hash_table_mask;
146 	dmu_buf_impl_t *dbf;
147 
148 	mutex_enter(DBUF_HASH_MUTEX(h, idx));
149 	for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) {
150 		if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
151 			mutex_enter(&dbf->db_mtx);
152 			if (dbf->db_state != DB_EVICTING) {
153 				mutex_exit(DBUF_HASH_MUTEX(h, idx));
154 				return (dbf);
155 			}
156 			mutex_exit(&dbf->db_mtx);
157 		}
158 	}
159 
160 	mutex_enter(&db->db_mtx);
161 	db->db_hash_next = h->hash_table[idx];
162 	h->hash_table[idx] = db;
163 	mutex_exit(DBUF_HASH_MUTEX(h, idx));
164 	atomic_add_64(&dbuf_hash_count, 1);
165 
166 	return (NULL);
167 }
168 
169 /*
170  * Remove an entry from the hash table.  This operation will
171  * fail if there are any existing holds on the db.
172  */
173 static void
174 dbuf_hash_remove(dmu_buf_impl_t *db)
175 {
176 	dbuf_hash_table_t *h = &dbuf_hash_table;
177 	uint64_t hv = DBUF_HASH(db->db_objset, db->db.db_object,
178 	    db->db_level, db->db_blkid);
179 	uint64_t idx = hv & h->hash_table_mask;
180 	dmu_buf_impl_t *dbf, **dbp;
181 
182 	/*
183 	 * We musn't hold db_mtx to maintin lock ordering:
184 	 * DBUF_HASH_MUTEX > db_mtx.
185 	 */
186 	ASSERT(refcount_is_zero(&db->db_holds));
187 	ASSERT(db->db_state == DB_EVICTING);
188 	ASSERT(!MUTEX_HELD(&db->db_mtx));
189 
190 	mutex_enter(DBUF_HASH_MUTEX(h, idx));
191 	dbp = &h->hash_table[idx];
192 	while ((dbf = *dbp) != db) {
193 		dbp = &dbf->db_hash_next;
194 		ASSERT(dbf != NULL);
195 	}
196 	*dbp = db->db_hash_next;
197 	db->db_hash_next = NULL;
198 	mutex_exit(DBUF_HASH_MUTEX(h, idx));
199 	atomic_add_64(&dbuf_hash_count, -1);
200 }
201 
202 static arc_evict_func_t dbuf_do_evict;
203 
204 static void
205 dbuf_evict_user(dmu_buf_impl_t *db)
206 {
207 	ASSERT(MUTEX_HELD(&db->db_mtx));
208 
209 	if (db->db_level != 0 || db->db_evict_func == NULL)
210 		return;
211 
212 	if (db->db_user_data_ptr_ptr)
213 		*db->db_user_data_ptr_ptr = db->db.db_data;
214 	db->db_evict_func(&db->db, db->db_user_ptr);
215 	db->db_user_ptr = NULL;
216 	db->db_user_data_ptr_ptr = NULL;
217 	db->db_evict_func = NULL;
218 }
219 
220 void
221 dbuf_evict(dmu_buf_impl_t *db)
222 {
223 	ASSERT(MUTEX_HELD(&db->db_mtx));
224 	ASSERT(db->db_buf == NULL);
225 	ASSERT(db->db_data_pending == NULL);
226 
227 	dbuf_clear(db);
228 	dbuf_destroy(db);
229 }
230 
231 void
232 dbuf_init(void)
233 {
234 	uint64_t hsize = 1ULL << 16;
235 	dbuf_hash_table_t *h = &dbuf_hash_table;
236 	int i;
237 
238 	/*
239 	 * The hash table is big enough to fill all of physical memory
240 	 * with an average 4K block size.  The table will take up
241 	 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
242 	 */
243 	while (hsize * 4096 < physmem * PAGESIZE)
244 		hsize <<= 1;
245 
246 retry:
247 	h->hash_table_mask = hsize - 1;
248 	h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
249 	if (h->hash_table == NULL) {
250 		/* XXX - we should really return an error instead of assert */
251 		ASSERT(hsize > (1ULL << 10));
252 		hsize >>= 1;
253 		goto retry;
254 	}
255 
256 	dbuf_cache = kmem_cache_create("dmu_buf_impl_t",
257 	    sizeof (dmu_buf_impl_t),
258 	    0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
259 
260 	for (i = 0; i < DBUF_MUTEXES; i++)
261 		mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
262 }
263 
264 void
265 dbuf_fini(void)
266 {
267 	dbuf_hash_table_t *h = &dbuf_hash_table;
268 	int i;
269 
270 	for (i = 0; i < DBUF_MUTEXES; i++)
271 		mutex_destroy(&h->hash_mutexes[i]);
272 	kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
273 	kmem_cache_destroy(dbuf_cache);
274 }
275 
276 /*
277  * Other stuff.
278  */
279 
280 #ifdef ZFS_DEBUG
281 static void
282 dbuf_verify(dmu_buf_impl_t *db)
283 {
284 	dnode_t *dn = db->db_dnode;
285 	dbuf_dirty_record_t *dr;
286 
287 	ASSERT(MUTEX_HELD(&db->db_mtx));
288 
289 	if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
290 		return;
291 
292 	ASSERT(db->db_objset != NULL);
293 	if (dn == NULL) {
294 		ASSERT(db->db_parent == NULL);
295 		ASSERT(db->db_blkptr == NULL);
296 	} else {
297 		ASSERT3U(db->db.db_object, ==, dn->dn_object);
298 		ASSERT3P(db->db_objset, ==, dn->dn_objset);
299 		ASSERT3U(db->db_level, <, dn->dn_nlevels);
300 		ASSERT(db->db_blkid == DMU_BONUS_BLKID || db->db_blkid ==
301 		    DMU_SPILL_BLKID || list_head(&dn->dn_dbufs));
302 	}
303 	if (db->db_blkid == DMU_BONUS_BLKID) {
304 		ASSERT(dn != NULL);
305 		ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
306 		ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
307 	} else if (db->db_blkid == DMU_SPILL_BLKID) {
308 		ASSERT(dn != NULL);
309 		ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
310 		ASSERT3U(db->db.db_offset, ==, 0);
311 	} else {
312 		ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
313 	}
314 
315 	for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
316 		ASSERT(dr->dr_dbuf == db);
317 
318 	for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
319 		ASSERT(dr->dr_dbuf == db);
320 
321 	/*
322 	 * We can't assert that db_size matches dn_datablksz because it
323 	 * can be momentarily different when another thread is doing
324 	 * dnode_set_blksz().
325 	 */
326 	if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
327 		dr = db->db_data_pending;
328 		/*
329 		 * It should only be modified in syncing context, so
330 		 * make sure we only have one copy of the data.
331 		 */
332 		ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
333 	}
334 
335 	/* verify db->db_blkptr */
336 	if (db->db_blkptr) {
337 		if (db->db_parent == dn->dn_dbuf) {
338 			/* db is pointed to by the dnode */
339 			/* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
340 			if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
341 				ASSERT(db->db_parent == NULL);
342 			else
343 				ASSERT(db->db_parent != NULL);
344 			if (db->db_blkid != DMU_SPILL_BLKID)
345 				ASSERT3P(db->db_blkptr, ==,
346 				    &dn->dn_phys->dn_blkptr[db->db_blkid]);
347 		} else {
348 			/* db is pointed to by an indirect block */
349 			int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
350 			ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
351 			ASSERT3U(db->db_parent->db.db_object, ==,
352 			    db->db.db_object);
353 			/*
354 			 * dnode_grow_indblksz() can make this fail if we don't
355 			 * have the struct_rwlock.  XXX indblksz no longer
356 			 * grows.  safe to do this now?
357 			 */
358 			if (RW_WRITE_HELD(&db->db_dnode->dn_struct_rwlock)) {
359 				ASSERT3P(db->db_blkptr, ==,
360 				    ((blkptr_t *)db->db_parent->db.db_data +
361 				    db->db_blkid % epb));
362 			}
363 		}
364 	}
365 	if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
366 	    (db->db_buf == NULL || db->db_buf->b_data) &&
367 	    db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
368 	    db->db_state != DB_FILL && !dn->dn_free_txg) {
369 		/*
370 		 * If the blkptr isn't set but they have nonzero data,
371 		 * it had better be dirty, otherwise we'll lose that
372 		 * data when we evict this buffer.
373 		 */
374 		if (db->db_dirtycnt == 0) {
375 			uint64_t *buf = db->db.db_data;
376 			int i;
377 
378 			for (i = 0; i < db->db.db_size >> 3; i++) {
379 				ASSERT(buf[i] == 0);
380 			}
381 		}
382 	}
383 }
384 #endif
385 
386 static void
387 dbuf_update_data(dmu_buf_impl_t *db)
388 {
389 	ASSERT(MUTEX_HELD(&db->db_mtx));
390 	if (db->db_level == 0 && db->db_user_data_ptr_ptr) {
391 		ASSERT(!refcount_is_zero(&db->db_holds));
392 		*db->db_user_data_ptr_ptr = db->db.db_data;
393 	}
394 }
395 
396 static void
397 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
398 {
399 	ASSERT(MUTEX_HELD(&db->db_mtx));
400 	ASSERT(db->db_buf == NULL || !arc_has_callback(db->db_buf));
401 	db->db_buf = buf;
402 	if (buf != NULL) {
403 		ASSERT(buf->b_data != NULL);
404 		db->db.db_data = buf->b_data;
405 		if (!arc_released(buf))
406 			arc_set_callback(buf, dbuf_do_evict, db);
407 		dbuf_update_data(db);
408 	} else {
409 		dbuf_evict_user(db);
410 		db->db.db_data = NULL;
411 		if (db->db_state != DB_NOFILL)
412 			db->db_state = DB_UNCACHED;
413 	}
414 }
415 
416 /*
417  * Loan out an arc_buf for read.  Return the loaned arc_buf.
418  */
419 arc_buf_t *
420 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
421 {
422 	arc_buf_t *abuf;
423 
424 	mutex_enter(&db->db_mtx);
425 	if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
426 		int blksz = db->db.db_size;
427 		mutex_exit(&db->db_mtx);
428 		abuf = arc_loan_buf(db->db_dnode->dn_objset->os_spa, blksz);
429 		bcopy(db->db.db_data, abuf->b_data, blksz);
430 	} else {
431 		abuf = db->db_buf;
432 		arc_loan_inuse_buf(abuf, db);
433 		dbuf_set_data(db, NULL);
434 		mutex_exit(&db->db_mtx);
435 	}
436 	return (abuf);
437 }
438 
439 uint64_t
440 dbuf_whichblock(dnode_t *dn, uint64_t offset)
441 {
442 	if (dn->dn_datablkshift) {
443 		return (offset >> dn->dn_datablkshift);
444 	} else {
445 		ASSERT3U(offset, <, dn->dn_datablksz);
446 		return (0);
447 	}
448 }
449 
450 static void
451 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
452 {
453 	dmu_buf_impl_t *db = vdb;
454 
455 	mutex_enter(&db->db_mtx);
456 	ASSERT3U(db->db_state, ==, DB_READ);
457 	/*
458 	 * All reads are synchronous, so we must have a hold on the dbuf
459 	 */
460 	ASSERT(refcount_count(&db->db_holds) > 0);
461 	ASSERT(db->db_buf == NULL);
462 	ASSERT(db->db.db_data == NULL);
463 	if (db->db_level == 0 && db->db_freed_in_flight) {
464 		/* we were freed in flight; disregard any error */
465 		arc_release(buf, db);
466 		bzero(buf->b_data, db->db.db_size);
467 		arc_buf_freeze(buf);
468 		db->db_freed_in_flight = FALSE;
469 		dbuf_set_data(db, buf);
470 		db->db_state = DB_CACHED;
471 	} else if (zio == NULL || zio->io_error == 0) {
472 		dbuf_set_data(db, buf);
473 		db->db_state = DB_CACHED;
474 	} else {
475 		ASSERT(db->db_blkid != DMU_BONUS_BLKID);
476 		ASSERT3P(db->db_buf, ==, NULL);
477 		VERIFY(arc_buf_remove_ref(buf, db) == 1);
478 		db->db_state = DB_UNCACHED;
479 	}
480 	cv_broadcast(&db->db_changed);
481 	dbuf_rele_and_unlock(db, NULL);
482 }
483 
484 static void
485 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t *flags)
486 {
487 	dnode_t *dn = db->db_dnode;
488 	zbookmark_t zb;
489 	uint32_t aflags = ARC_NOWAIT;
490 	arc_buf_t *pbuf;
491 
492 	ASSERT(!refcount_is_zero(&db->db_holds));
493 	/* We need the struct_rwlock to prevent db_blkptr from changing. */
494 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
495 	ASSERT(MUTEX_HELD(&db->db_mtx));
496 	ASSERT(db->db_state == DB_UNCACHED);
497 	ASSERT(db->db_buf == NULL);
498 
499 	if (db->db_blkid == DMU_BONUS_BLKID) {
500 		int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
501 
502 		ASSERT3U(bonuslen, <=, db->db.db_size);
503 		db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN);
504 		arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
505 		if (bonuslen < DN_MAX_BONUSLEN)
506 			bzero(db->db.db_data, DN_MAX_BONUSLEN);
507 		if (bonuslen)
508 			bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
509 		dbuf_update_data(db);
510 		db->db_state = DB_CACHED;
511 		mutex_exit(&db->db_mtx);
512 		return;
513 	}
514 
515 	/*
516 	 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
517 	 * processes the delete record and clears the bp while we are waiting
518 	 * for the dn_mtx (resulting in a "no" from block_freed).
519 	 */
520 	if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
521 	    (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
522 	    BP_IS_HOLE(db->db_blkptr)))) {
523 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
524 
525 		dbuf_set_data(db, arc_buf_alloc(dn->dn_objset->os_spa,
526 		    db->db.db_size, db, type));
527 		bzero(db->db.db_data, db->db.db_size);
528 		db->db_state = DB_CACHED;
529 		*flags |= DB_RF_CACHED;
530 		mutex_exit(&db->db_mtx);
531 		return;
532 	}
533 
534 	db->db_state = DB_READ;
535 	mutex_exit(&db->db_mtx);
536 
537 	if (DBUF_IS_L2CACHEABLE(db))
538 		aflags |= ARC_L2CACHE;
539 
540 	SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
541 	    db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
542 	    db->db.db_object, db->db_level, db->db_blkid);
543 
544 	dbuf_add_ref(db, NULL);
545 	/* ZIO_FLAG_CANFAIL callers have to check the parent zio's error */
546 
547 	if (db->db_parent)
548 		pbuf = db->db_parent->db_buf;
549 	else
550 		pbuf = db->db_objset->os_phys_buf;
551 
552 	(void) dsl_read(zio, dn->dn_objset->os_spa, db->db_blkptr, pbuf,
553 	    dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
554 	    (*flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
555 	    &aflags, &zb);
556 	if (aflags & ARC_CACHED)
557 		*flags |= DB_RF_CACHED;
558 }
559 
560 int
561 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
562 {
563 	int err = 0;
564 	int havepzio = (zio != NULL);
565 	int prefetch;
566 
567 	/*
568 	 * We don't have to hold the mutex to check db_state because it
569 	 * can't be freed while we have a hold on the buffer.
570 	 */
571 	ASSERT(!refcount_is_zero(&db->db_holds));
572 
573 	if (db->db_state == DB_NOFILL)
574 		return (EIO);
575 
576 	if ((flags & DB_RF_HAVESTRUCT) == 0)
577 		rw_enter(&db->db_dnode->dn_struct_rwlock, RW_READER);
578 
579 	prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
580 	    (flags & DB_RF_NOPREFETCH) == 0 && db->db_dnode != NULL &&
581 	    DBUF_IS_CACHEABLE(db);
582 
583 	mutex_enter(&db->db_mtx);
584 	if (db->db_state == DB_CACHED) {
585 		mutex_exit(&db->db_mtx);
586 		if (prefetch)
587 			dmu_zfetch(&db->db_dnode->dn_zfetch, db->db.db_offset,
588 			    db->db.db_size, TRUE);
589 		if ((flags & DB_RF_HAVESTRUCT) == 0)
590 			rw_exit(&db->db_dnode->dn_struct_rwlock);
591 	} else if (db->db_state == DB_UNCACHED) {
592 		if (zio == NULL) {
593 			zio = zio_root(db->db_dnode->dn_objset->os_spa,
594 			    NULL, NULL, ZIO_FLAG_CANFAIL);
595 		}
596 		dbuf_read_impl(db, zio, &flags);
597 
598 		/* dbuf_read_impl has dropped db_mtx for us */
599 
600 		if (prefetch)
601 			dmu_zfetch(&db->db_dnode->dn_zfetch, db->db.db_offset,
602 			    db->db.db_size, flags & DB_RF_CACHED);
603 
604 		if ((flags & DB_RF_HAVESTRUCT) == 0)
605 			rw_exit(&db->db_dnode->dn_struct_rwlock);
606 
607 		if (!havepzio)
608 			err = zio_wait(zio);
609 	} else {
610 		mutex_exit(&db->db_mtx);
611 		if (prefetch)
612 			dmu_zfetch(&db->db_dnode->dn_zfetch, db->db.db_offset,
613 			    db->db.db_size, TRUE);
614 		if ((flags & DB_RF_HAVESTRUCT) == 0)
615 			rw_exit(&db->db_dnode->dn_struct_rwlock);
616 
617 		mutex_enter(&db->db_mtx);
618 		if ((flags & DB_RF_NEVERWAIT) == 0) {
619 			while (db->db_state == DB_READ ||
620 			    db->db_state == DB_FILL) {
621 				ASSERT(db->db_state == DB_READ ||
622 				    (flags & DB_RF_HAVESTRUCT) == 0);
623 				cv_wait(&db->db_changed, &db->db_mtx);
624 			}
625 			if (db->db_state == DB_UNCACHED)
626 				err = EIO;
627 		}
628 		mutex_exit(&db->db_mtx);
629 	}
630 
631 	ASSERT(err || havepzio || db->db_state == DB_CACHED);
632 	return (err);
633 }
634 
635 static void
636 dbuf_noread(dmu_buf_impl_t *db)
637 {
638 	ASSERT(!refcount_is_zero(&db->db_holds));
639 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
640 	mutex_enter(&db->db_mtx);
641 	while (db->db_state == DB_READ || db->db_state == DB_FILL)
642 		cv_wait(&db->db_changed, &db->db_mtx);
643 	if (db->db_state == DB_UNCACHED) {
644 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
645 
646 		ASSERT(db->db_buf == NULL);
647 		ASSERT(db->db.db_data == NULL);
648 		dbuf_set_data(db, arc_buf_alloc(db->db_dnode->dn_objset->os_spa,
649 		    db->db.db_size, db, type));
650 		db->db_state = DB_FILL;
651 	} else if (db->db_state == DB_NOFILL) {
652 		dbuf_set_data(db, NULL);
653 	} else {
654 		ASSERT3U(db->db_state, ==, DB_CACHED);
655 	}
656 	mutex_exit(&db->db_mtx);
657 }
658 
659 /*
660  * This is our just-in-time copy function.  It makes a copy of
661  * buffers, that have been modified in a previous transaction
662  * group, before we modify them in the current active group.
663  *
664  * This function is used in two places: when we are dirtying a
665  * buffer for the first time in a txg, and when we are freeing
666  * a range in a dnode that includes this buffer.
667  *
668  * Note that when we are called from dbuf_free_range() we do
669  * not put a hold on the buffer, we just traverse the active
670  * dbuf list for the dnode.
671  */
672 static void
673 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
674 {
675 	dbuf_dirty_record_t *dr = db->db_last_dirty;
676 
677 	ASSERT(MUTEX_HELD(&db->db_mtx));
678 	ASSERT(db->db.db_data != NULL);
679 	ASSERT(db->db_level == 0);
680 	ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
681 
682 	if (dr == NULL ||
683 	    (dr->dt.dl.dr_data !=
684 	    ((db->db_blkid  == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
685 		return;
686 
687 	/*
688 	 * If the last dirty record for this dbuf has not yet synced
689 	 * and its referencing the dbuf data, either:
690 	 * 	reset the reference to point to a new copy,
691 	 * or (if there a no active holders)
692 	 *	just null out the current db_data pointer.
693 	 */
694 	ASSERT(dr->dr_txg >= txg - 2);
695 	if (db->db_blkid == DMU_BONUS_BLKID) {
696 		/* Note that the data bufs here are zio_bufs */
697 		dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
698 		arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
699 		bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
700 	} else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
701 		int size = db->db.db_size;
702 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
703 		dr->dt.dl.dr_data = arc_buf_alloc(
704 		    db->db_dnode->dn_objset->os_spa, size, db, type);
705 		bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
706 	} else {
707 		dbuf_set_data(db, NULL);
708 	}
709 }
710 
711 void
712 dbuf_unoverride(dbuf_dirty_record_t *dr)
713 {
714 	dmu_buf_impl_t *db = dr->dr_dbuf;
715 	blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
716 	uint64_t txg = dr->dr_txg;
717 
718 	ASSERT(MUTEX_HELD(&db->db_mtx));
719 	ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
720 	ASSERT(db->db_level == 0);
721 
722 	if (db->db_blkid == DMU_BONUS_BLKID ||
723 	    dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
724 		return;
725 
726 	ASSERT(db->db_data_pending != dr);
727 
728 	/* free this block */
729 	if (!BP_IS_HOLE(bp))
730 		zio_free(db->db_dnode->dn_objset->os_spa, txg, bp);
731 
732 	dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
733 	/*
734 	 * Release the already-written buffer, so we leave it in
735 	 * a consistent dirty state.  Note that all callers are
736 	 * modifying the buffer, so they will immediately do
737 	 * another (redundant) arc_release().  Therefore, leave
738 	 * the buf thawed to save the effort of freezing &
739 	 * immediately re-thawing it.
740 	 */
741 	arc_release(dr->dt.dl.dr_data, db);
742 }
743 
744 /*
745  * Evict (if its unreferenced) or clear (if its referenced) any level-0
746  * data blocks in the free range, so that any future readers will find
747  * empty blocks.  Also, if we happen accross any level-1 dbufs in the
748  * range that have not already been marked dirty, mark them dirty so
749  * they stay in memory.
750  */
751 void
752 dbuf_free_range(dnode_t *dn, uint64_t start, uint64_t end, dmu_tx_t *tx)
753 {
754 	dmu_buf_impl_t *db, *db_next;
755 	uint64_t txg = tx->tx_txg;
756 	int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
757 	uint64_t first_l1 = start >> epbs;
758 	uint64_t last_l1 = end >> epbs;
759 
760 	if (end > dn->dn_maxblkid && (end != DMU_SPILL_BLKID)) {
761 		end = dn->dn_maxblkid;
762 		last_l1 = end >> epbs;
763 	}
764 	dprintf_dnode(dn, "start=%llu end=%llu\n", start, end);
765 	mutex_enter(&dn->dn_dbufs_mtx);
766 	for (db = list_head(&dn->dn_dbufs); db; db = db_next) {
767 		db_next = list_next(&dn->dn_dbufs, db);
768 		ASSERT(db->db_blkid != DMU_BONUS_BLKID);
769 
770 		if (db->db_level == 1 &&
771 		    db->db_blkid >= first_l1 && db->db_blkid <= last_l1) {
772 			mutex_enter(&db->db_mtx);
773 			if (db->db_last_dirty &&
774 			    db->db_last_dirty->dr_txg < txg) {
775 				dbuf_add_ref(db, FTAG);
776 				mutex_exit(&db->db_mtx);
777 				dbuf_will_dirty(db, tx);
778 				dbuf_rele(db, FTAG);
779 			} else {
780 				mutex_exit(&db->db_mtx);
781 			}
782 		}
783 
784 		if (db->db_level != 0)
785 			continue;
786 		dprintf_dbuf(db, "found buf %s\n", "");
787 		if (db->db_blkid < start || db->db_blkid > end)
788 			continue;
789 
790 		/* found a level 0 buffer in the range */
791 		if (dbuf_undirty(db, tx))
792 			continue;
793 
794 		mutex_enter(&db->db_mtx);
795 		if (db->db_state == DB_UNCACHED ||
796 		    db->db_state == DB_NOFILL ||
797 		    db->db_state == DB_EVICTING) {
798 			ASSERT(db->db.db_data == NULL);
799 			mutex_exit(&db->db_mtx);
800 			continue;
801 		}
802 		if (db->db_state == DB_READ || db->db_state == DB_FILL) {
803 			/* will be handled in dbuf_read_done or dbuf_rele */
804 			db->db_freed_in_flight = TRUE;
805 			mutex_exit(&db->db_mtx);
806 			continue;
807 		}
808 		if (refcount_count(&db->db_holds) == 0) {
809 			ASSERT(db->db_buf);
810 			dbuf_clear(db);
811 			continue;
812 		}
813 		/* The dbuf is referenced */
814 
815 		if (db->db_last_dirty != NULL) {
816 			dbuf_dirty_record_t *dr = db->db_last_dirty;
817 
818 			if (dr->dr_txg == txg) {
819 				/*
820 				 * This buffer is "in-use", re-adjust the file
821 				 * size to reflect that this buffer may
822 				 * contain new data when we sync.
823 				 */
824 				if (db->db_blkid != DMU_SPILL_BLKID &&
825 				    db->db_blkid > dn->dn_maxblkid)
826 					dn->dn_maxblkid = db->db_blkid;
827 				dbuf_unoverride(dr);
828 			} else {
829 				/*
830 				 * This dbuf is not dirty in the open context.
831 				 * Either uncache it (if its not referenced in
832 				 * the open context) or reset its contents to
833 				 * empty.
834 				 */
835 				dbuf_fix_old_data(db, txg);
836 			}
837 		}
838 		/* clear the contents if its cached */
839 		if (db->db_state == DB_CACHED) {
840 			ASSERT(db->db.db_data != NULL);
841 			arc_release(db->db_buf, db);
842 			bzero(db->db.db_data, db->db.db_size);
843 			arc_buf_freeze(db->db_buf);
844 		}
845 
846 		mutex_exit(&db->db_mtx);
847 	}
848 	mutex_exit(&dn->dn_dbufs_mtx);
849 }
850 
851 static int
852 dbuf_block_freeable(dmu_buf_impl_t *db)
853 {
854 	dsl_dataset_t *ds = db->db_objset->os_dsl_dataset;
855 	uint64_t birth_txg = 0;
856 
857 	/*
858 	 * We don't need any locking to protect db_blkptr:
859 	 * If it's syncing, then db_last_dirty will be set
860 	 * so we'll ignore db_blkptr.
861 	 */
862 	ASSERT(MUTEX_HELD(&db->db_mtx));
863 	if (db->db_last_dirty)
864 		birth_txg = db->db_last_dirty->dr_txg;
865 	else if (db->db_blkptr)
866 		birth_txg = db->db_blkptr->blk_birth;
867 
868 	/* If we don't exist or are in a snapshot, we can't be freed */
869 	if (birth_txg)
870 		return (ds == NULL ||
871 		    dsl_dataset_block_freeable(ds, db->db_blkptr, birth_txg));
872 	else
873 		return (FALSE);
874 }
875 
876 void
877 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
878 {
879 	arc_buf_t *buf, *obuf;
880 	int osize = db->db.db_size;
881 	arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
882 
883 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
884 
885 	/* XXX does *this* func really need the lock? */
886 	ASSERT(RW_WRITE_HELD(&db->db_dnode->dn_struct_rwlock));
887 
888 	/*
889 	 * This call to dbuf_will_dirty() with the dn_struct_rwlock held
890 	 * is OK, because there can be no other references to the db
891 	 * when we are changing its size, so no concurrent DB_FILL can
892 	 * be happening.
893 	 */
894 	/*
895 	 * XXX we should be doing a dbuf_read, checking the return
896 	 * value and returning that up to our callers
897 	 */
898 	dbuf_will_dirty(db, tx);
899 
900 	/* create the data buffer for the new block */
901 	buf = arc_buf_alloc(db->db_dnode->dn_objset->os_spa, size, db, type);
902 
903 	/* copy old block data to the new block */
904 	obuf = db->db_buf;
905 	bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
906 	/* zero the remainder */
907 	if (size > osize)
908 		bzero((uint8_t *)buf->b_data + osize, size - osize);
909 
910 	mutex_enter(&db->db_mtx);
911 	dbuf_set_data(db, buf);
912 	VERIFY(arc_buf_remove_ref(obuf, db) == 1);
913 	db->db.db_size = size;
914 
915 	if (db->db_level == 0) {
916 		ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
917 		db->db_last_dirty->dt.dl.dr_data = buf;
918 	}
919 	mutex_exit(&db->db_mtx);
920 
921 	dnode_willuse_space(db->db_dnode, size-osize, tx);
922 }
923 
924 void
925 dbuf_release_bp(dmu_buf_impl_t *db)
926 {
927 	objset_t *os = db->db_dnode->dn_objset;
928 	zbookmark_t zb;
929 
930 	ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
931 	ASSERT(arc_released(os->os_phys_buf) ||
932 	    list_link_active(&os->os_dsl_dataset->ds_synced_link));
933 	ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
934 
935 	zb.zb_objset = os->os_dsl_dataset ?
936 	    os->os_dsl_dataset->ds_object : 0;
937 	zb.zb_object = db->db.db_object;
938 	zb.zb_level = db->db_level;
939 	zb.zb_blkid = db->db_blkid;
940 	(void) arc_release_bp(db->db_buf, db,
941 	    db->db_blkptr, os->os_spa, &zb);
942 }
943 
944 dbuf_dirty_record_t *
945 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
946 {
947 	dnode_t *dn = db->db_dnode;
948 	objset_t *os = dn->dn_objset;
949 	dbuf_dirty_record_t **drp, *dr;
950 	int drop_struct_lock = FALSE;
951 	boolean_t do_free_accounting = B_FALSE;
952 	int txgoff = tx->tx_txg & TXG_MASK;
953 
954 	ASSERT(tx->tx_txg != 0);
955 	ASSERT(!refcount_is_zero(&db->db_holds));
956 	DMU_TX_DIRTY_BUF(tx, db);
957 
958 	/*
959 	 * Shouldn't dirty a regular buffer in syncing context.  Private
960 	 * objects may be dirtied in syncing context, but only if they
961 	 * were already pre-dirtied in open context.
962 	 */
963 	ASSERT(!dmu_tx_is_syncing(tx) ||
964 	    BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
965 	    DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
966 	    dn->dn_objset->os_dsl_dataset == NULL);
967 	/*
968 	 * We make this assert for private objects as well, but after we
969 	 * check if we're already dirty.  They are allowed to re-dirty
970 	 * in syncing context.
971 	 */
972 	ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
973 	    dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
974 	    (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
975 
976 	mutex_enter(&db->db_mtx);
977 	/*
978 	 * XXX make this true for indirects too?  The problem is that
979 	 * transactions created with dmu_tx_create_assigned() from
980 	 * syncing context don't bother holding ahead.
981 	 */
982 	ASSERT(db->db_level != 0 ||
983 	    db->db_state == DB_CACHED || db->db_state == DB_FILL ||
984 	    db->db_state == DB_NOFILL);
985 
986 	mutex_enter(&dn->dn_mtx);
987 	/*
988 	 * Don't set dirtyctx to SYNC if we're just modifying this as we
989 	 * initialize the objset.
990 	 */
991 	if (dn->dn_dirtyctx == DN_UNDIRTIED &&
992 	    !BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
993 		dn->dn_dirtyctx =
994 		    (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN);
995 		ASSERT(dn->dn_dirtyctx_firstset == NULL);
996 		dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
997 	}
998 	mutex_exit(&dn->dn_mtx);
999 
1000 	if (db->db_blkid == DMU_SPILL_BLKID)
1001 		dn->dn_have_spill = B_TRUE;
1002 
1003 	/*
1004 	 * If this buffer is already dirty, we're done.
1005 	 */
1006 	drp = &db->db_last_dirty;
1007 	ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1008 	    db->db.db_object == DMU_META_DNODE_OBJECT);
1009 	while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1010 		drp = &dr->dr_next;
1011 	if (dr && dr->dr_txg == tx->tx_txg) {
1012 		if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1013 			/*
1014 			 * If this buffer has already been written out,
1015 			 * we now need to reset its state.
1016 			 */
1017 			dbuf_unoverride(dr);
1018 			if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1019 			    db->db_state != DB_NOFILL)
1020 				arc_buf_thaw(db->db_buf);
1021 		}
1022 		mutex_exit(&db->db_mtx);
1023 		return (dr);
1024 	}
1025 
1026 	/*
1027 	 * Only valid if not already dirty.
1028 	 */
1029 	ASSERT(dn->dn_object == 0 ||
1030 	    dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1031 	    (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1032 
1033 	ASSERT3U(dn->dn_nlevels, >, db->db_level);
1034 	ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1035 	    dn->dn_phys->dn_nlevels > db->db_level ||
1036 	    dn->dn_next_nlevels[txgoff] > db->db_level ||
1037 	    dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1038 	    dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1039 
1040 	/*
1041 	 * We should only be dirtying in syncing context if it's the
1042 	 * mos or we're initializing the os or it's a special object.
1043 	 * However, we are allowed to dirty in syncing context provided
1044 	 * we already dirtied it in open context.  Hence we must make
1045 	 * this assertion only if we're not already dirty.
1046 	 */
1047 	ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1048 	    os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1049 	ASSERT(db->db.db_size != 0);
1050 
1051 	dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1052 
1053 	if (db->db_blkid != DMU_BONUS_BLKID) {
1054 		/*
1055 		 * Update the accounting.
1056 		 * Note: we delay "free accounting" until after we drop
1057 		 * the db_mtx.  This keeps us from grabbing other locks
1058 		 * (and possibly deadlocking) in bp_get_dsize() while
1059 		 * also holding the db_mtx.
1060 		 */
1061 		dnode_willuse_space(dn, db->db.db_size, tx);
1062 		do_free_accounting = dbuf_block_freeable(db);
1063 	}
1064 
1065 	/*
1066 	 * If this buffer is dirty in an old transaction group we need
1067 	 * to make a copy of it so that the changes we make in this
1068 	 * transaction group won't leak out when we sync the older txg.
1069 	 */
1070 	dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1071 	if (db->db_level == 0) {
1072 		void *data_old = db->db_buf;
1073 
1074 		if (db->db_state != DB_NOFILL) {
1075 			if (db->db_blkid == DMU_BONUS_BLKID) {
1076 				dbuf_fix_old_data(db, tx->tx_txg);
1077 				data_old = db->db.db_data;
1078 			} else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1079 				/*
1080 				 * Release the data buffer from the cache so
1081 				 * that we can modify it without impacting
1082 				 * possible other users of this cached data
1083 				 * block.  Note that indirect blocks and
1084 				 * private objects are not released until the
1085 				 * syncing state (since they are only modified
1086 				 * then).
1087 				 */
1088 				arc_release(db->db_buf, db);
1089 				dbuf_fix_old_data(db, tx->tx_txg);
1090 				data_old = db->db_buf;
1091 			}
1092 			ASSERT(data_old != NULL);
1093 		}
1094 		dr->dt.dl.dr_data = data_old;
1095 	} else {
1096 		mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1097 		list_create(&dr->dt.di.dr_children,
1098 		    sizeof (dbuf_dirty_record_t),
1099 		    offsetof(dbuf_dirty_record_t, dr_dirty_node));
1100 	}
1101 	dr->dr_dbuf = db;
1102 	dr->dr_txg = tx->tx_txg;
1103 	dr->dr_next = *drp;
1104 	*drp = dr;
1105 
1106 	/*
1107 	 * We could have been freed_in_flight between the dbuf_noread
1108 	 * and dbuf_dirty.  We win, as though the dbuf_noread() had
1109 	 * happened after the free.
1110 	 */
1111 	if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1112 	    db->db_blkid != DMU_SPILL_BLKID) {
1113 		mutex_enter(&dn->dn_mtx);
1114 		dnode_clear_range(dn, db->db_blkid, 1, tx);
1115 		mutex_exit(&dn->dn_mtx);
1116 		db->db_freed_in_flight = FALSE;
1117 	}
1118 
1119 	/*
1120 	 * This buffer is now part of this txg
1121 	 */
1122 	dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1123 	db->db_dirtycnt += 1;
1124 	ASSERT3U(db->db_dirtycnt, <=, 3);
1125 
1126 	mutex_exit(&db->db_mtx);
1127 
1128 	if (db->db_blkid == DMU_BONUS_BLKID ||
1129 	    db->db_blkid == DMU_SPILL_BLKID) {
1130 		mutex_enter(&dn->dn_mtx);
1131 		ASSERT(!list_link_active(&dr->dr_dirty_node));
1132 		list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1133 		mutex_exit(&dn->dn_mtx);
1134 		dnode_setdirty(dn, tx);
1135 		return (dr);
1136 	} else if (do_free_accounting) {
1137 		blkptr_t *bp = db->db_blkptr;
1138 		int64_t willfree = (bp && !BP_IS_HOLE(bp)) ?
1139 		    bp_get_dsize(os->os_spa, bp) : db->db.db_size;
1140 		/*
1141 		 * This is only a guess -- if the dbuf is dirty
1142 		 * in a previous txg, we don't know how much
1143 		 * space it will use on disk yet.  We should
1144 		 * really have the struct_rwlock to access
1145 		 * db_blkptr, but since this is just a guess,
1146 		 * it's OK if we get an odd answer.
1147 		 */
1148 		dnode_willuse_space(dn, -willfree, tx);
1149 	}
1150 
1151 	if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1152 		rw_enter(&dn->dn_struct_rwlock, RW_READER);
1153 		drop_struct_lock = TRUE;
1154 	}
1155 
1156 	if (db->db_level == 0) {
1157 		dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1158 		ASSERT(dn->dn_maxblkid >= db->db_blkid);
1159 	}
1160 
1161 	if (db->db_level+1 < dn->dn_nlevels) {
1162 		dmu_buf_impl_t *parent = db->db_parent;
1163 		dbuf_dirty_record_t *di;
1164 		int parent_held = FALSE;
1165 
1166 		if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1167 			int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1168 
1169 			parent = dbuf_hold_level(dn, db->db_level+1,
1170 			    db->db_blkid >> epbs, FTAG);
1171 			ASSERT(parent != NULL);
1172 			parent_held = TRUE;
1173 		}
1174 		if (drop_struct_lock)
1175 			rw_exit(&dn->dn_struct_rwlock);
1176 		ASSERT3U(db->db_level+1, ==, parent->db_level);
1177 		di = dbuf_dirty(parent, tx);
1178 		if (parent_held)
1179 			dbuf_rele(parent, FTAG);
1180 
1181 		mutex_enter(&db->db_mtx);
1182 		/*  possible race with dbuf_undirty() */
1183 		if (db->db_last_dirty == dr ||
1184 		    dn->dn_object == DMU_META_DNODE_OBJECT) {
1185 			mutex_enter(&di->dt.di.dr_mtx);
1186 			ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1187 			ASSERT(!list_link_active(&dr->dr_dirty_node));
1188 			list_insert_tail(&di->dt.di.dr_children, dr);
1189 			mutex_exit(&di->dt.di.dr_mtx);
1190 			dr->dr_parent = di;
1191 		}
1192 		mutex_exit(&db->db_mtx);
1193 	} else {
1194 		ASSERT(db->db_level+1 == dn->dn_nlevels);
1195 		ASSERT(db->db_blkid < dn->dn_nblkptr);
1196 		ASSERT(db->db_parent == NULL ||
1197 		    db->db_parent == db->db_dnode->dn_dbuf);
1198 		mutex_enter(&dn->dn_mtx);
1199 		ASSERT(!list_link_active(&dr->dr_dirty_node));
1200 		list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1201 		mutex_exit(&dn->dn_mtx);
1202 		if (drop_struct_lock)
1203 			rw_exit(&dn->dn_struct_rwlock);
1204 	}
1205 
1206 	dnode_setdirty(dn, tx);
1207 	return (dr);
1208 }
1209 
1210 static int
1211 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1212 {
1213 	dnode_t *dn = db->db_dnode;
1214 	uint64_t txg = tx->tx_txg;
1215 	dbuf_dirty_record_t *dr, **drp;
1216 
1217 	ASSERT(txg != 0);
1218 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1219 
1220 	mutex_enter(&db->db_mtx);
1221 	/*
1222 	 * If this buffer is not dirty, we're done.
1223 	 */
1224 	for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1225 		if (dr->dr_txg <= txg)
1226 			break;
1227 	if (dr == NULL || dr->dr_txg < txg) {
1228 		mutex_exit(&db->db_mtx);
1229 		return (0);
1230 	}
1231 	ASSERT(dr->dr_txg == txg);
1232 	ASSERT(dr->dr_dbuf == db);
1233 
1234 	/*
1235 	 * If this buffer is currently held, we cannot undirty
1236 	 * it, since one of the current holders may be in the
1237 	 * middle of an update.  Note that users of dbuf_undirty()
1238 	 * should not place a hold on the dbuf before the call.
1239 	 */
1240 	if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
1241 		mutex_exit(&db->db_mtx);
1242 		/* Make sure we don't toss this buffer at sync phase */
1243 		mutex_enter(&dn->dn_mtx);
1244 		dnode_clear_range(dn, db->db_blkid, 1, tx);
1245 		mutex_exit(&dn->dn_mtx);
1246 		return (0);
1247 	}
1248 
1249 	dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1250 
1251 	ASSERT(db->db.db_size != 0);
1252 
1253 	/* XXX would be nice to fix up dn_towrite_space[] */
1254 
1255 	*drp = dr->dr_next;
1256 
1257 	if (dr->dr_parent) {
1258 		mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1259 		list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1260 		mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1261 	} else if (db->db_level+1 == dn->dn_nlevels) {
1262 		ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1263 		mutex_enter(&dn->dn_mtx);
1264 		list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1265 		mutex_exit(&dn->dn_mtx);
1266 	}
1267 
1268 	if (db->db_level == 0) {
1269 		if (db->db_state != DB_NOFILL) {
1270 			dbuf_unoverride(dr);
1271 
1272 			ASSERT(db->db_buf != NULL);
1273 			ASSERT(dr->dt.dl.dr_data != NULL);
1274 			if (dr->dt.dl.dr_data != db->db_buf)
1275 				VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data,
1276 				    db) == 1);
1277 		}
1278 	} else {
1279 		ASSERT(db->db_buf != NULL);
1280 		ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
1281 		mutex_destroy(&dr->dt.di.dr_mtx);
1282 		list_destroy(&dr->dt.di.dr_children);
1283 	}
1284 	kmem_free(dr, sizeof (dbuf_dirty_record_t));
1285 
1286 	ASSERT(db->db_dirtycnt > 0);
1287 	db->db_dirtycnt -= 1;
1288 
1289 	if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1290 		arc_buf_t *buf = db->db_buf;
1291 
1292 		ASSERT(db->db_state == DB_NOFILL || arc_released(buf));
1293 		dbuf_set_data(db, NULL);
1294 		VERIFY(arc_buf_remove_ref(buf, db) == 1);
1295 		dbuf_evict(db);
1296 		return (1);
1297 	}
1298 
1299 	mutex_exit(&db->db_mtx);
1300 	return (0);
1301 }
1302 
1303 #pragma weak dmu_buf_will_dirty = dbuf_will_dirty
1304 void
1305 dbuf_will_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1306 {
1307 	int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1308 
1309 	ASSERT(tx->tx_txg != 0);
1310 	ASSERT(!refcount_is_zero(&db->db_holds));
1311 
1312 	if (RW_WRITE_HELD(&db->db_dnode->dn_struct_rwlock))
1313 		rf |= DB_RF_HAVESTRUCT;
1314 	(void) dbuf_read(db, NULL, rf);
1315 	(void) dbuf_dirty(db, tx);
1316 }
1317 
1318 void
1319 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1320 {
1321 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1322 
1323 	db->db_state = DB_NOFILL;
1324 
1325 	dmu_buf_will_fill(db_fake, tx);
1326 }
1327 
1328 void
1329 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1330 {
1331 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1332 
1333 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1334 	ASSERT(tx->tx_txg != 0);
1335 	ASSERT(db->db_level == 0);
1336 	ASSERT(!refcount_is_zero(&db->db_holds));
1337 
1338 	ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1339 	    dmu_tx_private_ok(tx));
1340 
1341 	dbuf_noread(db);
1342 	(void) dbuf_dirty(db, tx);
1343 }
1344 
1345 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1346 /* ARGSUSED */
1347 void
1348 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1349 {
1350 	mutex_enter(&db->db_mtx);
1351 	DBUF_VERIFY(db);
1352 
1353 	if (db->db_state == DB_FILL) {
1354 		if (db->db_level == 0 && db->db_freed_in_flight) {
1355 			ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1356 			/* we were freed while filling */
1357 			/* XXX dbuf_undirty? */
1358 			bzero(db->db.db_data, db->db.db_size);
1359 			db->db_freed_in_flight = FALSE;
1360 		}
1361 		db->db_state = DB_CACHED;
1362 		cv_broadcast(&db->db_changed);
1363 	}
1364 	mutex_exit(&db->db_mtx);
1365 }
1366 
1367 /*
1368  * Directly assign a provided arc buf to a given dbuf if it's not referenced
1369  * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1370  */
1371 void
1372 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
1373 {
1374 	ASSERT(!refcount_is_zero(&db->db_holds));
1375 	ASSERT(db->db_dnode->dn_object != DMU_META_DNODE_OBJECT);
1376 	ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1377 	ASSERT(db->db_level == 0);
1378 	ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA);
1379 	ASSERT(buf != NULL);
1380 	ASSERT(arc_buf_size(buf) == db->db.db_size);
1381 	ASSERT(tx->tx_txg != 0);
1382 
1383 	arc_return_buf(buf, db);
1384 	ASSERT(arc_released(buf));
1385 
1386 	mutex_enter(&db->db_mtx);
1387 
1388 	while (db->db_state == DB_READ || db->db_state == DB_FILL)
1389 		cv_wait(&db->db_changed, &db->db_mtx);
1390 
1391 	ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
1392 
1393 	if (db->db_state == DB_CACHED &&
1394 	    refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
1395 		mutex_exit(&db->db_mtx);
1396 		(void) dbuf_dirty(db, tx);
1397 		bcopy(buf->b_data, db->db.db_data, db->db.db_size);
1398 		VERIFY(arc_buf_remove_ref(buf, db) == 1);
1399 		xuio_stat_wbuf_copied();
1400 		return;
1401 	}
1402 
1403 	xuio_stat_wbuf_nocopy();
1404 	if (db->db_state == DB_CACHED) {
1405 		dbuf_dirty_record_t *dr = db->db_last_dirty;
1406 
1407 		ASSERT(db->db_buf != NULL);
1408 		if (dr != NULL && dr->dr_txg == tx->tx_txg) {
1409 			ASSERT(dr->dt.dl.dr_data == db->db_buf);
1410 			if (!arc_released(db->db_buf)) {
1411 				ASSERT(dr->dt.dl.dr_override_state ==
1412 				    DR_OVERRIDDEN);
1413 				arc_release(db->db_buf, db);
1414 			}
1415 			dr->dt.dl.dr_data = buf;
1416 			VERIFY(arc_buf_remove_ref(db->db_buf, db) == 1);
1417 		} else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
1418 			arc_release(db->db_buf, db);
1419 			VERIFY(arc_buf_remove_ref(db->db_buf, db) == 1);
1420 		}
1421 		db->db_buf = NULL;
1422 	}
1423 	ASSERT(db->db_buf == NULL);
1424 	dbuf_set_data(db, buf);
1425 	db->db_state = DB_FILL;
1426 	mutex_exit(&db->db_mtx);
1427 	(void) dbuf_dirty(db, tx);
1428 	dbuf_fill_done(db, tx);
1429 }
1430 
1431 /*
1432  * "Clear" the contents of this dbuf.  This will mark the dbuf
1433  * EVICTING and clear *most* of its references.  Unfortunetely,
1434  * when we are not holding the dn_dbufs_mtx, we can't clear the
1435  * entry in the dn_dbufs list.  We have to wait until dbuf_destroy()
1436  * in this case.  For callers from the DMU we will usually see:
1437  *	dbuf_clear()->arc_buf_evict()->dbuf_do_evict()->dbuf_destroy()
1438  * For the arc callback, we will usually see:
1439  * 	dbuf_do_evict()->dbuf_clear();dbuf_destroy()
1440  * Sometimes, though, we will get a mix of these two:
1441  *	DMU: dbuf_clear()->arc_buf_evict()
1442  *	ARC: dbuf_do_evict()->dbuf_destroy()
1443  */
1444 void
1445 dbuf_clear(dmu_buf_impl_t *db)
1446 {
1447 	dnode_t *dn = db->db_dnode;
1448 	dmu_buf_impl_t *parent = db->db_parent;
1449 	dmu_buf_impl_t *dndb = dn->dn_dbuf;
1450 	int dbuf_gone = FALSE;
1451 
1452 	ASSERT(MUTEX_HELD(&db->db_mtx));
1453 	ASSERT(refcount_is_zero(&db->db_holds));
1454 
1455 	dbuf_evict_user(db);
1456 
1457 	if (db->db_state == DB_CACHED) {
1458 		ASSERT(db->db.db_data != NULL);
1459 		if (db->db_blkid == DMU_BONUS_BLKID) {
1460 			zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
1461 			arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
1462 		}
1463 		db->db.db_data = NULL;
1464 		db->db_state = DB_UNCACHED;
1465 	}
1466 
1467 	ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
1468 	ASSERT(db->db_data_pending == NULL);
1469 
1470 	db->db_state = DB_EVICTING;
1471 	db->db_blkptr = NULL;
1472 
1473 	if (db->db_blkid != DMU_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) {
1474 		list_remove(&dn->dn_dbufs, db);
1475 		dnode_rele(dn, db);
1476 		db->db_dnode = NULL;
1477 	}
1478 
1479 	if (db->db_buf)
1480 		dbuf_gone = arc_buf_evict(db->db_buf);
1481 
1482 	if (!dbuf_gone)
1483 		mutex_exit(&db->db_mtx);
1484 
1485 	/*
1486 	 * If this dbuf is referened from an indirect dbuf,
1487 	 * decrement the ref count on the indirect dbuf.
1488 	 */
1489 	if (parent && parent != dndb)
1490 		dbuf_rele(parent, db);
1491 }
1492 
1493 static int
1494 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
1495     dmu_buf_impl_t **parentp, blkptr_t **bpp)
1496 {
1497 	int nlevels, epbs;
1498 
1499 	*parentp = NULL;
1500 	*bpp = NULL;
1501 
1502 	ASSERT(blkid != DMU_BONUS_BLKID);
1503 
1504 	if (blkid == DMU_SPILL_BLKID) {
1505 		mutex_enter(&dn->dn_mtx);
1506 		if (dn->dn_have_spill &&
1507 		    (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
1508 			*bpp = &dn->dn_phys->dn_spill;
1509 		else
1510 			*bpp = NULL;
1511 		dbuf_add_ref(dn->dn_dbuf, NULL);
1512 		*parentp = dn->dn_dbuf;
1513 		mutex_exit(&dn->dn_mtx);
1514 		return (0);
1515 	}
1516 
1517 	if (dn->dn_phys->dn_nlevels == 0)
1518 		nlevels = 1;
1519 	else
1520 		nlevels = dn->dn_phys->dn_nlevels;
1521 
1522 	epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1523 
1524 	ASSERT3U(level * epbs, <, 64);
1525 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1526 	if (level >= nlevels ||
1527 	    (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
1528 		/* the buffer has no parent yet */
1529 		return (ENOENT);
1530 	} else if (level < nlevels-1) {
1531 		/* this block is referenced from an indirect block */
1532 		int err = dbuf_hold_impl(dn, level+1,
1533 		    blkid >> epbs, fail_sparse, NULL, parentp);
1534 		if (err)
1535 			return (err);
1536 		err = dbuf_read(*parentp, NULL,
1537 		    (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
1538 		if (err) {
1539 			dbuf_rele(*parentp, NULL);
1540 			*parentp = NULL;
1541 			return (err);
1542 		}
1543 		*bpp = ((blkptr_t *)(*parentp)->db.db_data) +
1544 		    (blkid & ((1ULL << epbs) - 1));
1545 		return (0);
1546 	} else {
1547 		/* the block is referenced from the dnode */
1548 		ASSERT3U(level, ==, nlevels-1);
1549 		ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
1550 		    blkid < dn->dn_phys->dn_nblkptr);
1551 		if (dn->dn_dbuf) {
1552 			dbuf_add_ref(dn->dn_dbuf, NULL);
1553 			*parentp = dn->dn_dbuf;
1554 		}
1555 		*bpp = &dn->dn_phys->dn_blkptr[blkid];
1556 		return (0);
1557 	}
1558 }
1559 
1560 static dmu_buf_impl_t *
1561 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
1562     dmu_buf_impl_t *parent, blkptr_t *blkptr)
1563 {
1564 	objset_t *os = dn->dn_objset;
1565 	dmu_buf_impl_t *db, *odb;
1566 
1567 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1568 	ASSERT(dn->dn_type != DMU_OT_NONE);
1569 
1570 	db = kmem_cache_alloc(dbuf_cache, KM_SLEEP);
1571 
1572 	db->db_objset = os;
1573 	db->db.db_object = dn->dn_object;
1574 	db->db_level = level;
1575 	db->db_blkid = blkid;
1576 	db->db_last_dirty = NULL;
1577 	db->db_dirtycnt = 0;
1578 	db->db_dnode = dn;
1579 	db->db_parent = parent;
1580 	db->db_blkptr = blkptr;
1581 
1582 	db->db_user_ptr = NULL;
1583 	db->db_user_data_ptr_ptr = NULL;
1584 	db->db_evict_func = NULL;
1585 	db->db_immediate_evict = 0;
1586 	db->db_freed_in_flight = 0;
1587 
1588 	if (blkid == DMU_BONUS_BLKID) {
1589 		ASSERT3P(parent, ==, dn->dn_dbuf);
1590 		db->db.db_size = DN_MAX_BONUSLEN -
1591 		    (dn->dn_nblkptr-1) * sizeof (blkptr_t);
1592 		ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
1593 		db->db.db_offset = DMU_BONUS_BLKID;
1594 		db->db_state = DB_UNCACHED;
1595 		/* the bonus dbuf is not placed in the hash table */
1596 		arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1597 		return (db);
1598 	} else if (blkid == DMU_SPILL_BLKID) {
1599 		db->db.db_size = (blkptr != NULL) ?
1600 		    BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
1601 		db->db.db_offset = 0;
1602 	} else {
1603 		int blocksize =
1604 		    db->db_level ? 1<<dn->dn_indblkshift :  dn->dn_datablksz;
1605 		db->db.db_size = blocksize;
1606 		db->db.db_offset = db->db_blkid * blocksize;
1607 	}
1608 
1609 	/*
1610 	 * Hold the dn_dbufs_mtx while we get the new dbuf
1611 	 * in the hash table *and* added to the dbufs list.
1612 	 * This prevents a possible deadlock with someone
1613 	 * trying to look up this dbuf before its added to the
1614 	 * dn_dbufs list.
1615 	 */
1616 	mutex_enter(&dn->dn_dbufs_mtx);
1617 	db->db_state = DB_EVICTING;
1618 	if ((odb = dbuf_hash_insert(db)) != NULL) {
1619 		/* someone else inserted it first */
1620 		kmem_cache_free(dbuf_cache, db);
1621 		mutex_exit(&dn->dn_dbufs_mtx);
1622 		return (odb);
1623 	}
1624 	list_insert_head(&dn->dn_dbufs, db);
1625 	db->db_state = DB_UNCACHED;
1626 	mutex_exit(&dn->dn_dbufs_mtx);
1627 	arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1628 
1629 	if (parent && parent != dn->dn_dbuf)
1630 		dbuf_add_ref(parent, db);
1631 
1632 	ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1633 	    refcount_count(&dn->dn_holds) > 0);
1634 	(void) refcount_add(&dn->dn_holds, db);
1635 
1636 	dprintf_dbuf(db, "db=%p\n", db);
1637 
1638 	return (db);
1639 }
1640 
1641 static int
1642 dbuf_do_evict(void *private)
1643 {
1644 	arc_buf_t *buf = private;
1645 	dmu_buf_impl_t *db = buf->b_private;
1646 
1647 	if (!MUTEX_HELD(&db->db_mtx))
1648 		mutex_enter(&db->db_mtx);
1649 
1650 	ASSERT(refcount_is_zero(&db->db_holds));
1651 
1652 	if (db->db_state != DB_EVICTING) {
1653 		ASSERT(db->db_state == DB_CACHED);
1654 		DBUF_VERIFY(db);
1655 		db->db_buf = NULL;
1656 		dbuf_evict(db);
1657 	} else {
1658 		mutex_exit(&db->db_mtx);
1659 		dbuf_destroy(db);
1660 	}
1661 	return (0);
1662 }
1663 
1664 static void
1665 dbuf_destroy(dmu_buf_impl_t *db)
1666 {
1667 	ASSERT(refcount_is_zero(&db->db_holds));
1668 
1669 	if (db->db_blkid != DMU_BONUS_BLKID) {
1670 		/*
1671 		 * If this dbuf is still on the dn_dbufs list,
1672 		 * remove it from that list.
1673 		 */
1674 		if (db->db_dnode) {
1675 			dnode_t *dn = db->db_dnode;
1676 
1677 			mutex_enter(&dn->dn_dbufs_mtx);
1678 			list_remove(&dn->dn_dbufs, db);
1679 			mutex_exit(&dn->dn_dbufs_mtx);
1680 
1681 			dnode_rele(dn, db);
1682 			db->db_dnode = NULL;
1683 		}
1684 		dbuf_hash_remove(db);
1685 	}
1686 	db->db_parent = NULL;
1687 	db->db_buf = NULL;
1688 
1689 	ASSERT(!list_link_active(&db->db_link));
1690 	ASSERT(db->db.db_data == NULL);
1691 	ASSERT(db->db_hash_next == NULL);
1692 	ASSERT(db->db_blkptr == NULL);
1693 	ASSERT(db->db_data_pending == NULL);
1694 
1695 	kmem_cache_free(dbuf_cache, db);
1696 	arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1697 }
1698 
1699 void
1700 dbuf_prefetch(dnode_t *dn, uint64_t blkid)
1701 {
1702 	dmu_buf_impl_t *db = NULL;
1703 	blkptr_t *bp = NULL;
1704 
1705 	ASSERT(blkid != DMU_BONUS_BLKID);
1706 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1707 
1708 	if (dnode_block_freed(dn, blkid))
1709 		return;
1710 
1711 	/* dbuf_find() returns with db_mtx held */
1712 	if (db = dbuf_find(dn, 0, blkid)) {
1713 		if (refcount_count(&db->db_holds) > 0) {
1714 			/*
1715 			 * This dbuf is active.  We assume that it is
1716 			 * already CACHED, or else about to be either
1717 			 * read or filled.
1718 			 */
1719 			mutex_exit(&db->db_mtx);
1720 			return;
1721 		}
1722 		mutex_exit(&db->db_mtx);
1723 		db = NULL;
1724 	}
1725 
1726 	if (dbuf_findbp(dn, 0, blkid, TRUE, &db, &bp) == 0) {
1727 		if (bp && !BP_IS_HOLE(bp)) {
1728 			int priority = dn->dn_type == DMU_OT_DDT_ZAP ?
1729 			    ZIO_PRIORITY_DDT_PREFETCH : ZIO_PRIORITY_ASYNC_READ;
1730 			arc_buf_t *pbuf;
1731 			dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
1732 			uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH;
1733 			zbookmark_t zb;
1734 
1735 			SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
1736 			    dn->dn_object, 0, blkid);
1737 
1738 			if (db)
1739 				pbuf = db->db_buf;
1740 			else
1741 				pbuf = dn->dn_objset->os_phys_buf;
1742 
1743 			(void) dsl_read(NULL, dn->dn_objset->os_spa,
1744 			    bp, pbuf, NULL, NULL, priority,
1745 			    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1746 			    &aflags, &zb);
1747 		}
1748 		if (db)
1749 			dbuf_rele(db, NULL);
1750 	}
1751 }
1752 
1753 /*
1754  * Returns with db_holds incremented, and db_mtx not held.
1755  * Note: dn_struct_rwlock must be held.
1756  */
1757 int
1758 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse,
1759     void *tag, dmu_buf_impl_t **dbp)
1760 {
1761 	dmu_buf_impl_t *db, *parent = NULL;
1762 
1763 	ASSERT(blkid != DMU_BONUS_BLKID);
1764 	ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1765 	ASSERT3U(dn->dn_nlevels, >, level);
1766 
1767 	*dbp = NULL;
1768 top:
1769 	/* dbuf_find() returns with db_mtx held */
1770 	db = dbuf_find(dn, level, blkid);
1771 
1772 	if (db == NULL) {
1773 		blkptr_t *bp = NULL;
1774 		int err;
1775 
1776 		ASSERT3P(parent, ==, NULL);
1777 		err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
1778 		if (fail_sparse) {
1779 			if (err == 0 && bp && BP_IS_HOLE(bp))
1780 				err = ENOENT;
1781 			if (err) {
1782 				if (parent)
1783 					dbuf_rele(parent, NULL);
1784 				return (err);
1785 			}
1786 		}
1787 		if (err && err != ENOENT)
1788 			return (err);
1789 		db = dbuf_create(dn, level, blkid, parent, bp);
1790 	}
1791 
1792 	if (db->db_buf && refcount_is_zero(&db->db_holds)) {
1793 		arc_buf_add_ref(db->db_buf, db);
1794 		if (db->db_buf->b_data == NULL) {
1795 			dbuf_clear(db);
1796 			if (parent) {
1797 				dbuf_rele(parent, NULL);
1798 				parent = NULL;
1799 			}
1800 			goto top;
1801 		}
1802 		ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
1803 	}
1804 
1805 	ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
1806 
1807 	/*
1808 	 * If this buffer is currently syncing out, and we are are
1809 	 * still referencing it from db_data, we need to make a copy
1810 	 * of it in case we decide we want to dirty it again in this txg.
1811 	 */
1812 	if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1813 	    dn->dn_object != DMU_META_DNODE_OBJECT &&
1814 	    db->db_state == DB_CACHED && db->db_data_pending) {
1815 		dbuf_dirty_record_t *dr = db->db_data_pending;
1816 
1817 		if (dr->dt.dl.dr_data == db->db_buf) {
1818 			arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1819 
1820 			dbuf_set_data(db,
1821 			    arc_buf_alloc(db->db_dnode->dn_objset->os_spa,
1822 			    db->db.db_size, db, type));
1823 			bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
1824 			    db->db.db_size);
1825 		}
1826 	}
1827 
1828 	(void) refcount_add(&db->db_holds, tag);
1829 	dbuf_update_data(db);
1830 	DBUF_VERIFY(db);
1831 	mutex_exit(&db->db_mtx);
1832 
1833 	/* NOTE: we can't rele the parent until after we drop the db_mtx */
1834 	if (parent)
1835 		dbuf_rele(parent, NULL);
1836 
1837 	ASSERT3P(db->db_dnode, ==, dn);
1838 	ASSERT3U(db->db_blkid, ==, blkid);
1839 	ASSERT3U(db->db_level, ==, level);
1840 	*dbp = db;
1841 
1842 	return (0);
1843 }
1844 
1845 dmu_buf_impl_t *
1846 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
1847 {
1848 	dmu_buf_impl_t *db;
1849 	int err = dbuf_hold_impl(dn, 0, blkid, FALSE, tag, &db);
1850 	return (err ? NULL : db);
1851 }
1852 
1853 dmu_buf_impl_t *
1854 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
1855 {
1856 	dmu_buf_impl_t *db;
1857 	int err = dbuf_hold_impl(dn, level, blkid, FALSE, tag, &db);
1858 	return (err ? NULL : db);
1859 }
1860 
1861 void
1862 dbuf_create_bonus(dnode_t *dn)
1863 {
1864 	ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1865 
1866 	ASSERT(dn->dn_bonus == NULL);
1867 	dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
1868 }
1869 
1870 int
1871 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
1872 {
1873 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1874 	if (db->db_blkid != DMU_SPILL_BLKID)
1875 		return (ENOTSUP);
1876 	if (blksz == 0)
1877 		blksz = SPA_MINBLOCKSIZE;
1878 	if (blksz > SPA_MAXBLOCKSIZE)
1879 		blksz = SPA_MAXBLOCKSIZE;
1880 	else
1881 		blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
1882 
1883 	rw_enter(&db->db_dnode->dn_struct_rwlock, RW_WRITER);
1884 	dbuf_new_size(db, blksz, tx);
1885 	rw_exit(&db->db_dnode->dn_struct_rwlock);
1886 
1887 	return (0);
1888 }
1889 
1890 void
1891 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
1892 {
1893 	dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
1894 }
1895 
1896 #pragma weak dmu_buf_add_ref = dbuf_add_ref
1897 void
1898 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
1899 {
1900 	int64_t holds = refcount_add(&db->db_holds, tag);
1901 	ASSERT(holds > 1);
1902 }
1903 
1904 #pragma weak dmu_buf_rele = dbuf_rele
1905 void
1906 dbuf_rele(dmu_buf_impl_t *db, void *tag)
1907 {
1908 	mutex_enter(&db->db_mtx);
1909 	dbuf_rele_and_unlock(db, tag);
1910 }
1911 
1912 /*
1913  * dbuf_rele() for an already-locked dbuf.  This is necessary to allow
1914  * db_dirtycnt and db_holds to be updated atomically.
1915  */
1916 void
1917 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
1918 {
1919 	int64_t holds;
1920 
1921 	ASSERT(MUTEX_HELD(&db->db_mtx));
1922 	DBUF_VERIFY(db);
1923 
1924 	holds = refcount_remove(&db->db_holds, tag);
1925 	ASSERT(holds >= 0);
1926 
1927 	/*
1928 	 * We can't freeze indirects if there is a possibility that they
1929 	 * may be modified in the current syncing context.
1930 	 */
1931 	if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0))
1932 		arc_buf_freeze(db->db_buf);
1933 
1934 	if (holds == db->db_dirtycnt &&
1935 	    db->db_level == 0 && db->db_immediate_evict)
1936 		dbuf_evict_user(db);
1937 
1938 	if (holds == 0) {
1939 		if (db->db_blkid == DMU_BONUS_BLKID) {
1940 			mutex_exit(&db->db_mtx);
1941 			dnode_rele(db->db_dnode, db);
1942 		} else if (db->db_buf == NULL) {
1943 			/*
1944 			 * This is a special case: we never associated this
1945 			 * dbuf with any data allocated from the ARC.
1946 			 */
1947 			ASSERT(db->db_state == DB_UNCACHED ||
1948 			    db->db_state == DB_NOFILL);
1949 			dbuf_evict(db);
1950 		} else if (arc_released(db->db_buf)) {
1951 			arc_buf_t *buf = db->db_buf;
1952 			/*
1953 			 * This dbuf has anonymous data associated with it.
1954 			 */
1955 			dbuf_set_data(db, NULL);
1956 			VERIFY(arc_buf_remove_ref(buf, db) == 1);
1957 			dbuf_evict(db);
1958 		} else {
1959 			VERIFY(arc_buf_remove_ref(db->db_buf, db) == 0);
1960 			if (!DBUF_IS_CACHEABLE(db))
1961 				dbuf_clear(db);
1962 			else
1963 				mutex_exit(&db->db_mtx);
1964 		}
1965 	} else {
1966 		mutex_exit(&db->db_mtx);
1967 	}
1968 }
1969 
1970 #pragma weak dmu_buf_refcount = dbuf_refcount
1971 uint64_t
1972 dbuf_refcount(dmu_buf_impl_t *db)
1973 {
1974 	return (refcount_count(&db->db_holds));
1975 }
1976 
1977 void *
1978 dmu_buf_set_user(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr,
1979     dmu_buf_evict_func_t *evict_func)
1980 {
1981 	return (dmu_buf_update_user(db_fake, NULL, user_ptr,
1982 	    user_data_ptr_ptr, evict_func));
1983 }
1984 
1985 void *
1986 dmu_buf_set_user_ie(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr,
1987     dmu_buf_evict_func_t *evict_func)
1988 {
1989 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1990 
1991 	db->db_immediate_evict = TRUE;
1992 	return (dmu_buf_update_user(db_fake, NULL, user_ptr,
1993 	    user_data_ptr_ptr, evict_func));
1994 }
1995 
1996 void *
1997 dmu_buf_update_user(dmu_buf_t *db_fake, void *old_user_ptr, void *user_ptr,
1998     void *user_data_ptr_ptr, dmu_buf_evict_func_t *evict_func)
1999 {
2000 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2001 	ASSERT(db->db_level == 0);
2002 
2003 	ASSERT((user_ptr == NULL) == (evict_func == NULL));
2004 
2005 	mutex_enter(&db->db_mtx);
2006 
2007 	if (db->db_user_ptr == old_user_ptr) {
2008 		db->db_user_ptr = user_ptr;
2009 		db->db_user_data_ptr_ptr = user_data_ptr_ptr;
2010 		db->db_evict_func = evict_func;
2011 
2012 		dbuf_update_data(db);
2013 	} else {
2014 		old_user_ptr = db->db_user_ptr;
2015 	}
2016 
2017 	mutex_exit(&db->db_mtx);
2018 	return (old_user_ptr);
2019 }
2020 
2021 void *
2022 dmu_buf_get_user(dmu_buf_t *db_fake)
2023 {
2024 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2025 	ASSERT(!refcount_is_zero(&db->db_holds));
2026 
2027 	return (db->db_user_ptr);
2028 }
2029 
2030 boolean_t
2031 dmu_buf_freeable(dmu_buf_t *dbuf)
2032 {
2033 	boolean_t res = B_FALSE;
2034 	dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2035 
2036 	if (db->db_blkptr)
2037 		res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset,
2038 		    db->db_blkptr, db->db_blkptr->blk_birth);
2039 
2040 	return (res);
2041 }
2042 
2043 static void
2044 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
2045 {
2046 	/* ASSERT(dmu_tx_is_syncing(tx) */
2047 	ASSERT(MUTEX_HELD(&db->db_mtx));
2048 
2049 	if (db->db_blkptr != NULL)
2050 		return;
2051 
2052 	if (db->db_blkid == DMU_SPILL_BLKID) {
2053 		db->db_blkptr = &dn->dn_phys->dn_spill;
2054 		BP_ZERO(db->db_blkptr);
2055 		return;
2056 	}
2057 	if (db->db_level == dn->dn_phys->dn_nlevels-1) {
2058 		/*
2059 		 * This buffer was allocated at a time when there was
2060 		 * no available blkptrs from the dnode, or it was
2061 		 * inappropriate to hook it in (i.e., nlevels mis-match).
2062 		 */
2063 		ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
2064 		ASSERT(db->db_parent == NULL);
2065 		db->db_parent = dn->dn_dbuf;
2066 		db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
2067 		DBUF_VERIFY(db);
2068 	} else {
2069 		dmu_buf_impl_t *parent = db->db_parent;
2070 		int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2071 
2072 		ASSERT(dn->dn_phys->dn_nlevels > 1);
2073 		if (parent == NULL) {
2074 			mutex_exit(&db->db_mtx);
2075 			rw_enter(&dn->dn_struct_rwlock, RW_READER);
2076 			(void) dbuf_hold_impl(dn, db->db_level+1,
2077 			    db->db_blkid >> epbs, FALSE, db, &parent);
2078 			rw_exit(&dn->dn_struct_rwlock);
2079 			mutex_enter(&db->db_mtx);
2080 			db->db_parent = parent;
2081 		}
2082 		db->db_blkptr = (blkptr_t *)parent->db.db_data +
2083 		    (db->db_blkid & ((1ULL << epbs) - 1));
2084 		DBUF_VERIFY(db);
2085 	}
2086 }
2087 
2088 static void
2089 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2090 {
2091 	dmu_buf_impl_t *db = dr->dr_dbuf;
2092 	dnode_t *dn = db->db_dnode;
2093 	zio_t *zio;
2094 
2095 	ASSERT(dmu_tx_is_syncing(tx));
2096 
2097 	dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2098 
2099 	mutex_enter(&db->db_mtx);
2100 
2101 	ASSERT(db->db_level > 0);
2102 	DBUF_VERIFY(db);
2103 
2104 	if (db->db_buf == NULL) {
2105 		mutex_exit(&db->db_mtx);
2106 		(void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
2107 		mutex_enter(&db->db_mtx);
2108 	}
2109 	ASSERT3U(db->db_state, ==, DB_CACHED);
2110 	ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2111 	ASSERT(db->db_buf != NULL);
2112 
2113 	dbuf_check_blkptr(dn, db);
2114 
2115 	db->db_data_pending = dr;
2116 
2117 	mutex_exit(&db->db_mtx);
2118 	dbuf_write(dr, db->db_buf, tx);
2119 
2120 	zio = dr->dr_zio;
2121 	mutex_enter(&dr->dt.di.dr_mtx);
2122 	dbuf_sync_list(&dr->dt.di.dr_children, tx);
2123 	ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2124 	mutex_exit(&dr->dt.di.dr_mtx);
2125 	zio_nowait(zio);
2126 }
2127 
2128 static void
2129 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2130 {
2131 	arc_buf_t **datap = &dr->dt.dl.dr_data;
2132 	dmu_buf_impl_t *db = dr->dr_dbuf;
2133 	dnode_t *dn = db->db_dnode;
2134 	objset_t *os = dn->dn_objset;
2135 	uint64_t txg = tx->tx_txg;
2136 
2137 	ASSERT(dmu_tx_is_syncing(tx));
2138 
2139 	dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2140 
2141 	mutex_enter(&db->db_mtx);
2142 	/*
2143 	 * To be synced, we must be dirtied.  But we
2144 	 * might have been freed after the dirty.
2145 	 */
2146 	if (db->db_state == DB_UNCACHED) {
2147 		/* This buffer has been freed since it was dirtied */
2148 		ASSERT(db->db.db_data == NULL);
2149 	} else if (db->db_state == DB_FILL) {
2150 		/* This buffer was freed and is now being re-filled */
2151 		ASSERT(db->db.db_data != dr->dt.dl.dr_data);
2152 	} else {
2153 		ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
2154 	}
2155 	DBUF_VERIFY(db);
2156 
2157 	if (db->db_blkid == DMU_SPILL_BLKID) {
2158 		mutex_enter(&dn->dn_mtx);
2159 		dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
2160 		mutex_exit(&dn->dn_mtx);
2161 	}
2162 
2163 	/*
2164 	 * If this is a bonus buffer, simply copy the bonus data into the
2165 	 * dnode.  It will be written out when the dnode is synced (and it
2166 	 * will be synced, since it must have been dirty for dbuf_sync to
2167 	 * be called).
2168 	 */
2169 	if (db->db_blkid == DMU_BONUS_BLKID) {
2170 		dbuf_dirty_record_t **drp;
2171 
2172 		ASSERT(*datap != NULL);
2173 		ASSERT3U(db->db_level, ==, 0);
2174 		ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN);
2175 		bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen);
2176 		if (*datap != db->db.db_data) {
2177 			zio_buf_free(*datap, DN_MAX_BONUSLEN);
2178 			arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
2179 		}
2180 		db->db_data_pending = NULL;
2181 		drp = &db->db_last_dirty;
2182 		while (*drp != dr)
2183 			drp = &(*drp)->dr_next;
2184 		ASSERT(dr->dr_next == NULL);
2185 		ASSERT(dr->dr_dbuf == db);
2186 		*drp = dr->dr_next;
2187 		kmem_free(dr, sizeof (dbuf_dirty_record_t));
2188 		ASSERT(db->db_dirtycnt > 0);
2189 		db->db_dirtycnt -= 1;
2190 		dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
2191 		return;
2192 	}
2193 
2194 	/*
2195 	 * This function may have dropped the db_mtx lock allowing a dmu_sync
2196 	 * operation to sneak in. As a result, we need to ensure that we
2197 	 * don't check the dr_override_state until we have returned from
2198 	 * dbuf_check_blkptr.
2199 	 */
2200 	dbuf_check_blkptr(dn, db);
2201 
2202 	/*
2203 	 * If this buffer is in the middle of an immdiate write,
2204 	 * wait for the synchronous IO to complete.
2205 	 */
2206 	while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
2207 		ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
2208 		cv_wait(&db->db_changed, &db->db_mtx);
2209 		ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
2210 	}
2211 
2212 	if (db->db_state != DB_NOFILL &&
2213 	    dn->dn_object != DMU_META_DNODE_OBJECT &&
2214 	    refcount_count(&db->db_holds) > 1 &&
2215 	    dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
2216 	    *datap == db->db_buf) {
2217 		/*
2218 		 * If this buffer is currently "in use" (i.e., there
2219 		 * are active holds and db_data still references it),
2220 		 * then make a copy before we start the write so that
2221 		 * any modifications from the open txg will not leak
2222 		 * into this write.
2223 		 *
2224 		 * NOTE: this copy does not need to be made for
2225 		 * objects only modified in the syncing context (e.g.
2226 		 * DNONE_DNODE blocks).
2227 		 */
2228 		int blksz = arc_buf_size(*datap);
2229 		arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2230 		*datap = arc_buf_alloc(os->os_spa, blksz, db, type);
2231 		bcopy(db->db.db_data, (*datap)->b_data, blksz);
2232 	}
2233 	db->db_data_pending = dr;
2234 
2235 	mutex_exit(&db->db_mtx);
2236 
2237 	dbuf_write(dr, *datap, tx);
2238 
2239 	ASSERT(!list_link_active(&dr->dr_dirty_node));
2240 	if (dn->dn_object == DMU_META_DNODE_OBJECT)
2241 		list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
2242 	else
2243 		zio_nowait(dr->dr_zio);
2244 }
2245 
2246 void
2247 dbuf_sync_list(list_t *list, dmu_tx_t *tx)
2248 {
2249 	dbuf_dirty_record_t *dr;
2250 
2251 	while (dr = list_head(list)) {
2252 		if (dr->dr_zio != NULL) {
2253 			/*
2254 			 * If we find an already initialized zio then we
2255 			 * are processing the meta-dnode, and we have finished.
2256 			 * The dbufs for all dnodes are put back on the list
2257 			 * during processing, so that we can zio_wait()
2258 			 * these IOs after initiating all child IOs.
2259 			 */
2260 			ASSERT3U(dr->dr_dbuf->db.db_object, ==,
2261 			    DMU_META_DNODE_OBJECT);
2262 			break;
2263 		}
2264 		list_remove(list, dr);
2265 		if (dr->dr_dbuf->db_level > 0)
2266 			dbuf_sync_indirect(dr, tx);
2267 		else
2268 			dbuf_sync_leaf(dr, tx);
2269 	}
2270 }
2271 
2272 /* ARGSUSED */
2273 static void
2274 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
2275 {
2276 	dmu_buf_impl_t *db = vdb;
2277 	blkptr_t *bp = zio->io_bp;
2278 	blkptr_t *bp_orig = &zio->io_bp_orig;
2279 	dnode_t *dn = db->db_dnode;
2280 	spa_t *spa = zio->io_spa;
2281 	int64_t delta;
2282 	uint64_t fill = 0;
2283 	int i;
2284 
2285 	ASSERT(db->db_blkptr == bp);
2286 
2287 	delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
2288 	dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
2289 	zio->io_prev_space_delta = delta;
2290 
2291 	if (BP_IS_HOLE(bp)) {
2292 		ASSERT(bp->blk_fill == 0);
2293 		return;
2294 	}
2295 
2296 	ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
2297 	    BP_GET_TYPE(bp) == dn->dn_type) ||
2298 	    (db->db_blkid == DMU_SPILL_BLKID &&
2299 	    BP_GET_TYPE(bp) == dn->dn_bonustype));
2300 	ASSERT(BP_GET_LEVEL(bp) == db->db_level);
2301 
2302 	mutex_enter(&db->db_mtx);
2303 
2304 #ifdef ZFS_DEBUG
2305 	if (db->db_blkid == DMU_SPILL_BLKID) {
2306 		dnode_t *dn = db->db_dnode;
2307 		ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2308 		ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2309 		    db->db_blkptr == &dn->dn_phys->dn_spill);
2310 	}
2311 #endif
2312 
2313 	if (db->db_level == 0) {
2314 		mutex_enter(&dn->dn_mtx);
2315 		if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
2316 		    db->db_blkid != DMU_SPILL_BLKID)
2317 			dn->dn_phys->dn_maxblkid = db->db_blkid;
2318 		mutex_exit(&dn->dn_mtx);
2319 
2320 		if (dn->dn_type == DMU_OT_DNODE) {
2321 			dnode_phys_t *dnp = db->db.db_data;
2322 			for (i = db->db.db_size >> DNODE_SHIFT; i > 0;
2323 			    i--, dnp++) {
2324 				if (dnp->dn_type != DMU_OT_NONE)
2325 					fill++;
2326 			}
2327 		} else {
2328 			fill = 1;
2329 		}
2330 	} else {
2331 		blkptr_t *ibp = db->db.db_data;
2332 		ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2333 		for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
2334 			if (BP_IS_HOLE(ibp))
2335 				continue;
2336 			fill += ibp->blk_fill;
2337 		}
2338 	}
2339 
2340 	bp->blk_fill = fill;
2341 
2342 	mutex_exit(&db->db_mtx);
2343 }
2344 
2345 /* ARGSUSED */
2346 static void
2347 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
2348 {
2349 	dmu_buf_impl_t *db = vdb;
2350 	blkptr_t *bp = zio->io_bp;
2351 	blkptr_t *bp_orig = &zio->io_bp_orig;
2352 	dnode_t *dn = db->db_dnode;
2353 	objset_t *os = dn->dn_objset;
2354 	uint64_t txg = zio->io_txg;
2355 	dbuf_dirty_record_t **drp, *dr;
2356 
2357 	ASSERT3U(zio->io_error, ==, 0);
2358 	ASSERT(db->db_blkptr == bp);
2359 
2360 	if (zio->io_flags & ZIO_FLAG_IO_REWRITE) {
2361 		ASSERT(BP_EQUAL(bp, bp_orig));
2362 	} else {
2363 		dsl_dataset_t *ds = os->os_dsl_dataset;
2364 		dmu_tx_t *tx = os->os_synctx;
2365 
2366 		(void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
2367 		dsl_dataset_block_born(ds, bp, tx);
2368 	}
2369 
2370 	mutex_enter(&db->db_mtx);
2371 
2372 	DBUF_VERIFY(db);
2373 
2374 	drp = &db->db_last_dirty;
2375 	while ((dr = *drp) != db->db_data_pending)
2376 		drp = &dr->dr_next;
2377 	ASSERT(!list_link_active(&dr->dr_dirty_node));
2378 	ASSERT(dr->dr_txg == txg);
2379 	ASSERT(dr->dr_dbuf == db);
2380 	ASSERT(dr->dr_next == NULL);
2381 	*drp = dr->dr_next;
2382 
2383 #ifdef ZFS_DEBUG
2384 	if (db->db_blkid == DMU_SPILL_BLKID) {
2385 		dnode_t *dn = db->db_dnode;
2386 		ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2387 		ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2388 		    db->db_blkptr == &dn->dn_phys->dn_spill);
2389 	}
2390 #endif
2391 
2392 	if (db->db_level == 0) {
2393 		ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2394 		ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
2395 		if (db->db_state != DB_NOFILL) {
2396 			if (dr->dt.dl.dr_data != db->db_buf)
2397 				VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data,
2398 				    db) == 1);
2399 			else if (!arc_released(db->db_buf))
2400 				arc_set_callback(db->db_buf, dbuf_do_evict, db);
2401 		}
2402 	} else {
2403 		ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2404 		ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2405 		if (!BP_IS_HOLE(db->db_blkptr)) {
2406 			int epbs =
2407 			    dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2408 			ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
2409 			    db->db.db_size);
2410 			ASSERT3U(dn->dn_phys->dn_maxblkid
2411 			    >> (db->db_level * epbs), >=, db->db_blkid);
2412 			arc_set_callback(db->db_buf, dbuf_do_evict, db);
2413 		}
2414 		mutex_destroy(&dr->dt.di.dr_mtx);
2415 		list_destroy(&dr->dt.di.dr_children);
2416 	}
2417 	kmem_free(dr, sizeof (dbuf_dirty_record_t));
2418 
2419 	cv_broadcast(&db->db_changed);
2420 	ASSERT(db->db_dirtycnt > 0);
2421 	db->db_dirtycnt -= 1;
2422 	db->db_data_pending = NULL;
2423 	dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
2424 }
2425 
2426 static void
2427 dbuf_write_nofill_ready(zio_t *zio)
2428 {
2429 	dbuf_write_ready(zio, NULL, zio->io_private);
2430 }
2431 
2432 static void
2433 dbuf_write_nofill_done(zio_t *zio)
2434 {
2435 	dbuf_write_done(zio, NULL, zio->io_private);
2436 }
2437 
2438 static void
2439 dbuf_write_override_ready(zio_t *zio)
2440 {
2441 	dbuf_dirty_record_t *dr = zio->io_private;
2442 	dmu_buf_impl_t *db = dr->dr_dbuf;
2443 
2444 	dbuf_write_ready(zio, NULL, db);
2445 }
2446 
2447 static void
2448 dbuf_write_override_done(zio_t *zio)
2449 {
2450 	dbuf_dirty_record_t *dr = zio->io_private;
2451 	dmu_buf_impl_t *db = dr->dr_dbuf;
2452 	blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
2453 
2454 	mutex_enter(&db->db_mtx);
2455 	if (!BP_EQUAL(zio->io_bp, obp)) {
2456 		if (!BP_IS_HOLE(obp))
2457 			dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
2458 		arc_release(dr->dt.dl.dr_data, db);
2459 	}
2460 	mutex_exit(&db->db_mtx);
2461 
2462 	dbuf_write_done(zio, NULL, db);
2463 }
2464 
2465 static void
2466 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
2467 {
2468 	dmu_buf_impl_t *db = dr->dr_dbuf;
2469 	dnode_t *dn = db->db_dnode;
2470 	objset_t *os = dn->dn_objset;
2471 	dmu_buf_impl_t *parent = db->db_parent;
2472 	uint64_t txg = tx->tx_txg;
2473 	zbookmark_t zb;
2474 	zio_prop_t zp;
2475 	zio_t *zio;
2476 	int wp_flag = 0;
2477 
2478 	if (db->db_state != DB_NOFILL) {
2479 		if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
2480 			/*
2481 			 * Private object buffers are released here rather
2482 			 * than in dbuf_dirty() since they are only modified
2483 			 * in the syncing context and we don't want the
2484 			 * overhead of making multiple copies of the data.
2485 			 */
2486 			if (BP_IS_HOLE(db->db_blkptr)) {
2487 				arc_buf_thaw(data);
2488 			} else {
2489 				dbuf_release_bp(db);
2490 			}
2491 		}
2492 	}
2493 
2494 	if (parent != dn->dn_dbuf) {
2495 		ASSERT(parent && parent->db_data_pending);
2496 		ASSERT(db->db_level == parent->db_level-1);
2497 		ASSERT(arc_released(parent->db_buf));
2498 		zio = parent->db_data_pending->dr_zio;
2499 	} else {
2500 		ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
2501 		    db->db_blkid != DMU_SPILL_BLKID) ||
2502 		    (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
2503 		if (db->db_blkid != DMU_SPILL_BLKID)
2504 			ASSERT3P(db->db_blkptr, ==,
2505 			    &dn->dn_phys->dn_blkptr[db->db_blkid]);
2506 		zio = dn->dn_zio;
2507 	}
2508 
2509 	ASSERT(db->db_level == 0 || data == db->db_buf);
2510 	ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
2511 	ASSERT(zio);
2512 
2513 	SET_BOOKMARK(&zb, os->os_dsl_dataset ?
2514 	    os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
2515 	    db->db.db_object, db->db_level, db->db_blkid);
2516 
2517 	if (db->db_blkid == DMU_SPILL_BLKID)
2518 		wp_flag = WP_SPILL;
2519 	wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
2520 
2521 	dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
2522 
2523 	if (db->db_level == 0 && dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
2524 		ASSERT(db->db_state != DB_NOFILL);
2525 		dr->dr_zio = zio_write(zio, os->os_spa, txg,
2526 		    db->db_blkptr, data->b_data, arc_buf_size(data), &zp,
2527 		    dbuf_write_override_ready, dbuf_write_override_done, dr,
2528 		    ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
2529 		mutex_enter(&db->db_mtx);
2530 		dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
2531 		zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
2532 		    dr->dt.dl.dr_copies);
2533 		mutex_exit(&db->db_mtx);
2534 	} else if (db->db_state == DB_NOFILL) {
2535 		ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF);
2536 		dr->dr_zio = zio_write(zio, os->os_spa, txg,
2537 		    db->db_blkptr, NULL, db->db.db_size, &zp,
2538 		    dbuf_write_nofill_ready, dbuf_write_nofill_done, db,
2539 		    ZIO_PRIORITY_ASYNC_WRITE,
2540 		    ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
2541 	} else {
2542 		ASSERT(arc_released(data));
2543 		dr->dr_zio = arc_write(zio, os->os_spa, txg,
2544 		    db->db_blkptr, data, DBUF_IS_L2CACHEABLE(db), &zp,
2545 		    dbuf_write_ready, dbuf_write_done, db,
2546 		    ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
2547 	}
2548 }
2549