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) 2013 by Delphix. All rights reserved.
25 */
26
27 #include <sys/zfs_context.h>
28 #include <sys/dmu.h>
29 #include <sys/dmu_impl.h>
30 #include <sys/dbuf.h>
31 #include <sys/dmu_objset.h>
32 #include <sys/dsl_dataset.h>
33 #include <sys/dsl_dir.h>
34 #include <sys/dmu_tx.h>
35 #include <sys/spa.h>
36 #include <sys/zio.h>
37 #include <sys/dmu_zfetch.h>
38 #include <sys/sa.h>
39 #include <sys/sa_impl.h>
40
41 static void dbuf_destroy(dmu_buf_impl_t *db);
42 static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
43 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
44
45 /*
46 * Global data structures and functions for the dbuf cache.
47 */
48 static kmem_cache_t *dbuf_cache;
49
50 /* ARGSUSED */
51 static int
52 dbuf_cons(void *vdb, void *unused, int kmflag)
53 {
54 dmu_buf_impl_t *db = vdb;
55 bzero(db, sizeof (dmu_buf_impl_t));
56
57 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
58 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
59 refcount_create(&db->db_holds);
60 return (0);
61 }
62
63 /* ARGSUSED */
64 static void
65 dbuf_dest(void *vdb, void *unused)
66 {
67 dmu_buf_impl_t *db = vdb;
68 mutex_destroy(&db->db_mtx);
69 cv_destroy(&db->db_changed);
70 refcount_destroy(&db->db_holds);
71 }
72
73 /*
74 * dbuf hash table routines
75 */
76 static dbuf_hash_table_t dbuf_hash_table;
77
78 static uint64_t dbuf_hash_count;
79
80 static uint64_t
81 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
82 {
83 uintptr_t osv = (uintptr_t)os;
84 uint64_t crc = -1ULL;
85
86 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
87 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF];
88 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF];
89 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF];
90 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF];
91 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF];
92 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF];
93
94 crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16);
95
96 return (crc);
97 }
98
99 #define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid);
100
101 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
102 ((dbuf)->db.db_object == (obj) && \
103 (dbuf)->db_objset == (os) && \
104 (dbuf)->db_level == (level) && \
105 (dbuf)->db_blkid == (blkid))
106
107 dmu_buf_impl_t *
108 dbuf_find(dnode_t *dn, uint8_t level, uint64_t blkid)
109 {
110 dbuf_hash_table_t *h = &dbuf_hash_table;
111 objset_t *os = dn->dn_objset;
112 uint64_t obj = dn->dn_object;
113 uint64_t hv = DBUF_HASH(os, obj, level, blkid);
114 uint64_t idx = hv & h->hash_table_mask;
115 dmu_buf_impl_t *db;
116
117 mutex_enter(DBUF_HASH_MUTEX(h, idx));
118 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
119 if (DBUF_EQUAL(db, os, obj, level, blkid)) {
120 mutex_enter(&db->db_mtx);
121 if (db->db_state != DB_EVICTING) {
122 mutex_exit(DBUF_HASH_MUTEX(h, idx));
123 return (db);
124 }
125 mutex_exit(&db->db_mtx);
126 }
127 }
128 mutex_exit(DBUF_HASH_MUTEX(h, idx));
129 return (NULL);
130 }
131
132 /*
133 * Insert an entry into the hash table. If there is already an element
134 * equal to elem in the hash table, then the already existing element
135 * will be returned and the new element will not be inserted.
136 * Otherwise returns NULL.
137 */
138 static dmu_buf_impl_t *
139 dbuf_hash_insert(dmu_buf_impl_t *db)
140 {
141 dbuf_hash_table_t *h = &dbuf_hash_table;
142 objset_t *os = db->db_objset;
143 uint64_t obj = db->db.db_object;
144 int level = db->db_level;
145 uint64_t blkid = db->db_blkid;
146 uint64_t hv = DBUF_HASH(os, obj, level, blkid);
147 uint64_t idx = hv & h->hash_table_mask;
148 dmu_buf_impl_t *dbf;
149
150 mutex_enter(DBUF_HASH_MUTEX(h, idx));
151 for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) {
152 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
153 mutex_enter(&dbf->db_mtx);
154 if (dbf->db_state != DB_EVICTING) {
155 mutex_exit(DBUF_HASH_MUTEX(h, idx));
156 return (dbf);
157 }
158 mutex_exit(&dbf->db_mtx);
159 }
160 }
161
162 mutex_enter(&db->db_mtx);
163 db->db_hash_next = h->hash_table[idx];
164 h->hash_table[idx] = db;
165 mutex_exit(DBUF_HASH_MUTEX(h, idx));
166 atomic_add_64(&dbuf_hash_count, 1);
167
168 return (NULL);
169 }
170
171 /*
172 * Remove an entry from the hash table. This operation will
173 * fail if there are any existing holds on the db.
174 */
175 static void
176 dbuf_hash_remove(dmu_buf_impl_t *db)
177 {
178 dbuf_hash_table_t *h = &dbuf_hash_table;
179 uint64_t hv = DBUF_HASH(db->db_objset, db->db.db_object,
180 db->db_level, db->db_blkid);
181 uint64_t idx = hv & h->hash_table_mask;
182 dmu_buf_impl_t *dbf, **dbp;
183
184 /*
185 * We musn't hold db_mtx to maintin lock ordering:
186 * DBUF_HASH_MUTEX > db_mtx.
187 */
188 ASSERT(refcount_is_zero(&db->db_holds));
189 ASSERT(db->db_state == DB_EVICTING);
190 ASSERT(!MUTEX_HELD(&db->db_mtx));
191
192 mutex_enter(DBUF_HASH_MUTEX(h, idx));
193 dbp = &h->hash_table[idx];
194 while ((dbf = *dbp) != db) {
195 dbp = &dbf->db_hash_next;
196 ASSERT(dbf != NULL);
197 }
198 *dbp = db->db_hash_next;
199 db->db_hash_next = NULL;
200 mutex_exit(DBUF_HASH_MUTEX(h, idx));
201 atomic_add_64(&dbuf_hash_count, -1);
202 }
203
204 static arc_evict_func_t dbuf_do_evict;
205
206 static void
207 dbuf_evict_user(dmu_buf_impl_t *db)
208 {
209 ASSERT(MUTEX_HELD(&db->db_mtx));
210
211 if (db->db_level != 0 || db->db_evict_func == NULL)
212 return;
213
214 if (db->db_user_data_ptr_ptr)
215 *db->db_user_data_ptr_ptr = db->db.db_data;
216 db->db_evict_func(&db->db, db->db_user_ptr);
217 db->db_user_ptr = NULL;
218 db->db_user_data_ptr_ptr = NULL;
219 db->db_evict_func = NULL;
220 }
221
222 boolean_t
223 dbuf_is_metadata(dmu_buf_impl_t *db)
224 {
225 if (db->db_level > 0) {
226 return (B_TRUE);
227 } else {
228 boolean_t is_metadata;
229
230 DB_DNODE_ENTER(db);
231 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type);
232 DB_DNODE_EXIT(db);
233
234 return (is_metadata);
235 }
236 }
237
238 void
239 dbuf_evict(dmu_buf_impl_t *db)
240 {
241 ASSERT(MUTEX_HELD(&db->db_mtx));
242 ASSERT(db->db_buf == NULL);
243 ASSERT(db->db_data_pending == NULL);
244
245 dbuf_clear(db);
246 dbuf_destroy(db);
247 }
248
249 void
250 dbuf_init(void)
251 {
252 uint64_t hsize = 1ULL << 16;
253 dbuf_hash_table_t *h = &dbuf_hash_table;
254 int i;
255
256 /*
257 * The hash table is big enough to fill all of physical memory
258 * with an average 4K block size. The table will take up
259 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
260 */
261 while (hsize * 4096 < physmem * PAGESIZE)
262 hsize <<= 1;
263
264 retry:
265 h->hash_table_mask = hsize - 1;
266 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
267 if (h->hash_table == NULL) {
268 /* XXX - we should really return an error instead of assert */
269 ASSERT(hsize > (1ULL << 10));
270 hsize >>= 1;
271 goto retry;
272 }
273
274 dbuf_cache = kmem_cache_create("dmu_buf_impl_t",
275 sizeof (dmu_buf_impl_t),
276 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
277
278 for (i = 0; i < DBUF_MUTEXES; i++)
279 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
280 }
281
282 void
283 dbuf_fini(void)
284 {
285 dbuf_hash_table_t *h = &dbuf_hash_table;
286 int i;
287
288 for (i = 0; i < DBUF_MUTEXES; i++)
289 mutex_destroy(&h->hash_mutexes[i]);
290 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
291 kmem_cache_destroy(dbuf_cache);
292 }
293
294 /*
295 * Other stuff.
296 */
297
298 #ifdef ZFS_DEBUG
299 static void
300 dbuf_verify(dmu_buf_impl_t *db)
301 {
302 dnode_t *dn;
303 dbuf_dirty_record_t *dr;
304
305 ASSERT(MUTEX_HELD(&db->db_mtx));
306
307 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
308 return;
309
310 ASSERT(db->db_objset != NULL);
311 DB_DNODE_ENTER(db);
312 dn = DB_DNODE(db);
313 if (dn == NULL) {
314 ASSERT(db->db_parent == NULL);
315 ASSERT(db->db_blkptr == NULL);
316 } else {
317 ASSERT3U(db->db.db_object, ==, dn->dn_object);
318 ASSERT3P(db->db_objset, ==, dn->dn_objset);
319 ASSERT3U(db->db_level, <, dn->dn_nlevels);
320 ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
321 db->db_blkid == DMU_SPILL_BLKID ||
322 !list_is_empty(&dn->dn_dbufs));
323 }
324 if (db->db_blkid == DMU_BONUS_BLKID) {
325 ASSERT(dn != NULL);
326 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
327 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
328 } else if (db->db_blkid == DMU_SPILL_BLKID) {
329 ASSERT(dn != NULL);
330 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
331 ASSERT0(db->db.db_offset);
332 } else {
333 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
334 }
335
336 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
337 ASSERT(dr->dr_dbuf == db);
338
339 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
340 ASSERT(dr->dr_dbuf == db);
341
342 /*
343 * We can't assert that db_size matches dn_datablksz because it
344 * can be momentarily different when another thread is doing
345 * dnode_set_blksz().
346 */
347 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
348 dr = db->db_data_pending;
349 /*
350 * It should only be modified in syncing context, so
351 * make sure we only have one copy of the data.
352 */
353 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
354 }
355
356 /* verify db->db_blkptr */
357 if (db->db_blkptr) {
358 if (db->db_parent == dn->dn_dbuf) {
359 /* db is pointed to by the dnode */
360 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
361 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
362 ASSERT(db->db_parent == NULL);
363 else
364 ASSERT(db->db_parent != NULL);
365 if (db->db_blkid != DMU_SPILL_BLKID)
366 ASSERT3P(db->db_blkptr, ==,
367 &dn->dn_phys->dn_blkptr[db->db_blkid]);
368 } else {
369 /* db is pointed to by an indirect block */
370 int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
371 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
372 ASSERT3U(db->db_parent->db.db_object, ==,
373 db->db.db_object);
374 /*
375 * dnode_grow_indblksz() can make this fail if we don't
376 * have the struct_rwlock. XXX indblksz no longer
377 * grows. safe to do this now?
378 */
379 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
380 ASSERT3P(db->db_blkptr, ==,
381 ((blkptr_t *)db->db_parent->db.db_data +
382 db->db_blkid % epb));
383 }
384 }
385 }
386 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
387 (db->db_buf == NULL || db->db_buf->b_data) &&
388 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
389 db->db_state != DB_FILL && !dn->dn_free_txg) {
390 /*
391 * If the blkptr isn't set but they have nonzero data,
392 * it had better be dirty, otherwise we'll lose that
393 * data when we evict this buffer.
394 */
395 if (db->db_dirtycnt == 0) {
396 uint64_t *buf = db->db.db_data;
397 int i;
398
399 for (i = 0; i < db->db.db_size >> 3; i++) {
400 ASSERT(buf[i] == 0);
401 }
402 }
403 }
404 DB_DNODE_EXIT(db);
405 }
406 #endif
407
408 static void
409 dbuf_update_data(dmu_buf_impl_t *db)
410 {
411 ASSERT(MUTEX_HELD(&db->db_mtx));
412 if (db->db_level == 0 && db->db_user_data_ptr_ptr) {
413 ASSERT(!refcount_is_zero(&db->db_holds));
414 *db->db_user_data_ptr_ptr = db->db.db_data;
415 }
416 }
417
418 static void
419 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
420 {
421 ASSERT(MUTEX_HELD(&db->db_mtx));
422 ASSERT(db->db_buf == NULL || !arc_has_callback(db->db_buf));
423 db->db_buf = buf;
424 if (buf != NULL) {
425 ASSERT(buf->b_data != NULL);
426 db->db.db_data = buf->b_data;
427 if (!arc_released(buf))
428 arc_set_callback(buf, dbuf_do_evict, db);
429 dbuf_update_data(db);
430 } else {
431 dbuf_evict_user(db);
432 db->db.db_data = NULL;
433 if (db->db_state != DB_NOFILL)
434 db->db_state = DB_UNCACHED;
435 }
436 }
437
438 /*
439 * Loan out an arc_buf for read. Return the loaned arc_buf.
440 */
441 arc_buf_t *
442 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
443 {
444 arc_buf_t *abuf;
445
446 mutex_enter(&db->db_mtx);
447 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
448 int blksz = db->db.db_size;
449 spa_t *spa;
450
451 mutex_exit(&db->db_mtx);
452 DB_GET_SPA(&spa, db);
453 abuf = arc_loan_buf(spa, blksz);
454 bcopy(db->db.db_data, abuf->b_data, blksz);
455 } else {
456 abuf = db->db_buf;
457 arc_loan_inuse_buf(abuf, db);
458 dbuf_set_data(db, NULL);
459 mutex_exit(&db->db_mtx);
460 }
461 return (abuf);
462 }
463
464 uint64_t
465 dbuf_whichblock(dnode_t *dn, uint64_t offset)
466 {
467 if (dn->dn_datablkshift) {
468 return (offset >> dn->dn_datablkshift);
469 } else {
470 ASSERT3U(offset, <, dn->dn_datablksz);
471 return (0);
472 }
473 }
474
475 static void
476 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
477 {
478 dmu_buf_impl_t *db = vdb;
479
480 mutex_enter(&db->db_mtx);
481 ASSERT3U(db->db_state, ==, DB_READ);
482 /*
483 * All reads are synchronous, so we must have a hold on the dbuf
484 */
485 ASSERT(refcount_count(&db->db_holds) > 0);
486 ASSERT(db->db_buf == NULL);
487 ASSERT(db->db.db_data == NULL);
488 if (db->db_level == 0 && db->db_freed_in_flight) {
489 /* we were freed in flight; disregard any error */
490 arc_release(buf, db);
491 bzero(buf->b_data, db->db.db_size);
492 arc_buf_freeze(buf);
493 db->db_freed_in_flight = FALSE;
494 dbuf_set_data(db, buf);
495 db->db_state = DB_CACHED;
496 } else if (zio == NULL || zio->io_error == 0) {
497 dbuf_set_data(db, buf);
498 db->db_state = DB_CACHED;
499 } else {
500 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
501 ASSERT3P(db->db_buf, ==, NULL);
502 VERIFY(arc_buf_remove_ref(buf, db));
503 db->db_state = DB_UNCACHED;
504 }
505 cv_broadcast(&db->db_changed);
506 dbuf_rele_and_unlock(db, NULL);
507 }
508
509 static void
510 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t *flags)
511 {
512 dnode_t *dn;
513 spa_t *spa;
514 zbookmark_t zb;
515 uint32_t aflags = ARC_NOWAIT;
516
517 DB_DNODE_ENTER(db);
518 dn = DB_DNODE(db);
519 ASSERT(!refcount_is_zero(&db->db_holds));
520 /* We need the struct_rwlock to prevent db_blkptr from changing. */
521 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
522 ASSERT(MUTEX_HELD(&db->db_mtx));
523 ASSERT(db->db_state == DB_UNCACHED);
524 ASSERT(db->db_buf == NULL);
525
526 if (db->db_blkid == DMU_BONUS_BLKID) {
527 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
528
529 ASSERT3U(bonuslen, <=, db->db.db_size);
530 db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN);
531 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
532 if (bonuslen < DN_MAX_BONUSLEN)
533 bzero(db->db.db_data, DN_MAX_BONUSLEN);
534 if (bonuslen)
535 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
536 DB_DNODE_EXIT(db);
537 dbuf_update_data(db);
538 db->db_state = DB_CACHED;
539 mutex_exit(&db->db_mtx);
540 return;
541 }
542
543 /*
544 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
545 * processes the delete record and clears the bp while we are waiting
546 * for the dn_mtx (resulting in a "no" from block_freed).
547 */
548 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
549 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
550 BP_IS_HOLE(db->db_blkptr)))) {
551 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
552
553 dbuf_set_data(db, arc_buf_alloc(dn->dn_objset->os_spa,
554 db->db.db_size, db, type));
555 DB_DNODE_EXIT(db);
556 bzero(db->db.db_data, db->db.db_size);
557 db->db_state = DB_CACHED;
558 *flags |= DB_RF_CACHED;
559 mutex_exit(&db->db_mtx);
560 return;
561 }
562
563 spa = dn->dn_objset->os_spa;
564 DB_DNODE_EXIT(db);
565
566 db->db_state = DB_READ;
567 mutex_exit(&db->db_mtx);
568
569 if (DBUF_IS_L2CACHEABLE(db))
570 aflags |= ARC_L2CACHE;
571
572 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
573 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
574 db->db.db_object, db->db_level, db->db_blkid);
575
576 dbuf_add_ref(db, NULL);
577
578 (void) arc_read(zio, spa, db->db_blkptr,
579 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
580 (*flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
581 &aflags, &zb);
582 if (aflags & ARC_CACHED)
583 *flags |= DB_RF_CACHED;
584 }
585
586 int
587 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
588 {
589 int err = 0;
590 int havepzio = (zio != NULL);
591 int prefetch;
592 dnode_t *dn;
593
594 /*
595 * We don't have to hold the mutex to check db_state because it
596 * can't be freed while we have a hold on the buffer.
597 */
598 ASSERT(!refcount_is_zero(&db->db_holds));
599
600 if (db->db_state == DB_NOFILL)
601 return (SET_ERROR(EIO));
602
603 DB_DNODE_ENTER(db);
604 dn = DB_DNODE(db);
605 if ((flags & DB_RF_HAVESTRUCT) == 0)
606 rw_enter(&dn->dn_struct_rwlock, RW_READER);
607
608 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
609 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
610 DBUF_IS_CACHEABLE(db);
611
612 mutex_enter(&db->db_mtx);
613 if (db->db_state == DB_CACHED) {
614 mutex_exit(&db->db_mtx);
615 if (prefetch)
616 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
617 db->db.db_size, TRUE);
618 if ((flags & DB_RF_HAVESTRUCT) == 0)
619 rw_exit(&dn->dn_struct_rwlock);
620 DB_DNODE_EXIT(db);
621 } else if (db->db_state == DB_UNCACHED) {
622 spa_t *spa = dn->dn_objset->os_spa;
623
624 if (zio == NULL)
625 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
626 dbuf_read_impl(db, zio, &flags);
627
628 /* dbuf_read_impl has dropped db_mtx for us */
629
630 if (prefetch)
631 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
632 db->db.db_size, flags & DB_RF_CACHED);
633
634 if ((flags & DB_RF_HAVESTRUCT) == 0)
635 rw_exit(&dn->dn_struct_rwlock);
636 DB_DNODE_EXIT(db);
637
638 if (!havepzio)
639 err = zio_wait(zio);
640 } else {
641 /*
642 * Another reader came in while the dbuf was in flight
643 * between UNCACHED and CACHED. Either a writer will finish
644 * writing the buffer (sending the dbuf to CACHED) or the
645 * first reader's request will reach the read_done callback
646 * and send the dbuf to CACHED. Otherwise, a failure
647 * occurred and the dbuf went to UNCACHED.
648 */
649 mutex_exit(&db->db_mtx);
650 if (prefetch)
651 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
652 db->db.db_size, TRUE);
653 if ((flags & DB_RF_HAVESTRUCT) == 0)
654 rw_exit(&dn->dn_struct_rwlock);
655 DB_DNODE_EXIT(db);
656
657 /* Skip the wait per the caller's request. */
658 mutex_enter(&db->db_mtx);
659 if ((flags & DB_RF_NEVERWAIT) == 0) {
660 while (db->db_state == DB_READ ||
661 db->db_state == DB_FILL) {
662 ASSERT(db->db_state == DB_READ ||
663 (flags & DB_RF_HAVESTRUCT) == 0);
664 cv_wait(&db->db_changed, &db->db_mtx);
665 }
666 if (db->db_state == DB_UNCACHED)
667 err = SET_ERROR(EIO);
668 }
669 mutex_exit(&db->db_mtx);
670 }
671
672 ASSERT(err || havepzio || db->db_state == DB_CACHED);
673 return (err);
674 }
675
676 static void
677 dbuf_noread(dmu_buf_impl_t *db)
678 {
679 ASSERT(!refcount_is_zero(&db->db_holds));
680 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
681 mutex_enter(&db->db_mtx);
682 while (db->db_state == DB_READ || db->db_state == DB_FILL)
683 cv_wait(&db->db_changed, &db->db_mtx);
684 if (db->db_state == DB_UNCACHED) {
685 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
686 spa_t *spa;
687
688 ASSERT(db->db_buf == NULL);
689 ASSERT(db->db.db_data == NULL);
690 DB_GET_SPA(&spa, db);
691 dbuf_set_data(db, arc_buf_alloc(spa, db->db.db_size, db, type));
692 db->db_state = DB_FILL;
693 } else if (db->db_state == DB_NOFILL) {
694 dbuf_set_data(db, NULL);
695 } else {
696 ASSERT3U(db->db_state, ==, DB_CACHED);
697 }
698 mutex_exit(&db->db_mtx);
699 }
700
701 /*
702 * This is our just-in-time copy function. It makes a copy of
703 * buffers, that have been modified in a previous transaction
704 * group, before we modify them in the current active group.
705 *
706 * This function is used in two places: when we are dirtying a
707 * buffer for the first time in a txg, and when we are freeing
708 * a range in a dnode that includes this buffer.
709 *
710 * Note that when we are called from dbuf_free_range() we do
711 * not put a hold on the buffer, we just traverse the active
712 * dbuf list for the dnode.
713 */
714 static void
715 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
716 {
717 dbuf_dirty_record_t *dr = db->db_last_dirty;
718
719 ASSERT(MUTEX_HELD(&db->db_mtx));
720 ASSERT(db->db.db_data != NULL);
721 ASSERT(db->db_level == 0);
722 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
723
724 if (dr == NULL ||
725 (dr->dt.dl.dr_data !=
726 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
727 return;
728
729 /*
730 * If the last dirty record for this dbuf has not yet synced
731 * and its referencing the dbuf data, either:
732 * reset the reference to point to a new copy,
733 * or (if there a no active holders)
734 * just null out the current db_data pointer.
735 */
736 ASSERT(dr->dr_txg >= txg - 2);
737 if (db->db_blkid == DMU_BONUS_BLKID) {
738 /* Note that the data bufs here are zio_bufs */
739 dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
740 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
741 bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
742 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
743 int size = db->db.db_size;
744 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
745 spa_t *spa;
746
747 DB_GET_SPA(&spa, db);
748 dr->dt.dl.dr_data = arc_buf_alloc(spa, size, db, type);
749 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
750 } else {
751 dbuf_set_data(db, NULL);
752 }
753 }
754
755 void
756 dbuf_unoverride(dbuf_dirty_record_t *dr)
757 {
758 dmu_buf_impl_t *db = dr->dr_dbuf;
759 blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
760 uint64_t txg = dr->dr_txg;
761
762 ASSERT(MUTEX_HELD(&db->db_mtx));
763 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
764 ASSERT(db->db_level == 0);
765
766 if (db->db_blkid == DMU_BONUS_BLKID ||
767 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
768 return;
769
770 ASSERT(db->db_data_pending != dr);
771
772 /* free this block */
773 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite) {
774 spa_t *spa;
775
776 DB_GET_SPA(&spa, db);
777 zio_free(spa, txg, bp);
778 }
779 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
780 dr->dt.dl.dr_nopwrite = B_FALSE;
781
782 /*
783 * Release the already-written buffer, so we leave it in
784 * a consistent dirty state. Note that all callers are
785 * modifying the buffer, so they will immediately do
786 * another (redundant) arc_release(). Therefore, leave
787 * the buf thawed to save the effort of freezing &
788 * immediately re-thawing it.
789 */
790 arc_release(dr->dt.dl.dr_data, db);
791 }
792
793 /*
794 * Evict (if its unreferenced) or clear (if its referenced) any level-0
795 * data blocks in the free range, so that any future readers will find
796 * empty blocks. Also, if we happen accross any level-1 dbufs in the
797 * range that have not already been marked dirty, mark them dirty so
798 * they stay in memory.
799 */
800 void
801 dbuf_free_range(dnode_t *dn, uint64_t start, uint64_t end, dmu_tx_t *tx)
802 {
803 dmu_buf_impl_t *db, *db_next;
804 uint64_t txg = tx->tx_txg;
805 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
806 uint64_t first_l1 = start >> epbs;
807 uint64_t last_l1 = end >> epbs;
808
809 if (end > dn->dn_maxblkid && (end != DMU_SPILL_BLKID)) {
810 end = dn->dn_maxblkid;
811 last_l1 = end >> epbs;
812 }
813 dprintf_dnode(dn, "start=%llu end=%llu\n", start, end);
814 mutex_enter(&dn->dn_dbufs_mtx);
815 for (db = list_head(&dn->dn_dbufs); db; db = db_next) {
816 db_next = list_next(&dn->dn_dbufs, db);
817 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
818
819 if (db->db_level == 1 &&
820 db->db_blkid >= first_l1 && db->db_blkid <= last_l1) {
821 mutex_enter(&db->db_mtx);
822 if (db->db_last_dirty &&
823 db->db_last_dirty->dr_txg < txg) {
824 dbuf_add_ref(db, FTAG);
825 mutex_exit(&db->db_mtx);
826 dbuf_will_dirty(db, tx);
827 dbuf_rele(db, FTAG);
828 } else {
829 mutex_exit(&db->db_mtx);
830 }
831 }
832
833 if (db->db_level != 0)
834 continue;
835 dprintf_dbuf(db, "found buf %s\n", "");
836 if (db->db_blkid < start || db->db_blkid > end)
837 continue;
838
839 /* found a level 0 buffer in the range */
840 mutex_enter(&db->db_mtx);
841 if (dbuf_undirty(db, tx)) {
842 /* mutex has been dropped and dbuf destroyed */
843 continue;
844 }
845
846 if (db->db_state == DB_UNCACHED ||
847 db->db_state == DB_NOFILL ||
848 db->db_state == DB_EVICTING) {
849 ASSERT(db->db.db_data == NULL);
850 mutex_exit(&db->db_mtx);
851 continue;
852 }
853 if (db->db_state == DB_READ || db->db_state == DB_FILL) {
854 /* will be handled in dbuf_read_done or dbuf_rele */
855 db->db_freed_in_flight = TRUE;
856 mutex_exit(&db->db_mtx);
857 continue;
858 }
859 if (refcount_count(&db->db_holds) == 0) {
860 ASSERT(db->db_buf);
861 dbuf_clear(db);
862 continue;
863 }
864 /* The dbuf is referenced */
865
866 if (db->db_last_dirty != NULL) {
867 dbuf_dirty_record_t *dr = db->db_last_dirty;
868
869 if (dr->dr_txg == txg) {
870 /*
871 * This buffer is "in-use", re-adjust the file
872 * size to reflect that this buffer may
873 * contain new data when we sync.
874 */
875 if (db->db_blkid != DMU_SPILL_BLKID &&
876 db->db_blkid > dn->dn_maxblkid)
877 dn->dn_maxblkid = db->db_blkid;
878 dbuf_unoverride(dr);
879 } else {
880 /*
881 * This dbuf is not dirty in the open context.
882 * Either uncache it (if its not referenced in
883 * the open context) or reset its contents to
884 * empty.
885 */
886 dbuf_fix_old_data(db, txg);
887 }
888 }
889 /* clear the contents if its cached */
890 if (db->db_state == DB_CACHED) {
891 ASSERT(db->db.db_data != NULL);
892 arc_release(db->db_buf, db);
893 bzero(db->db.db_data, db->db.db_size);
894 arc_buf_freeze(db->db_buf);
895 }
896
897 mutex_exit(&db->db_mtx);
898 }
899 mutex_exit(&dn->dn_dbufs_mtx);
900 }
901
902 static int
903 dbuf_block_freeable(dmu_buf_impl_t *db)
904 {
905 dsl_dataset_t *ds = db->db_objset->os_dsl_dataset;
906 uint64_t birth_txg = 0;
907
908 /*
909 * We don't need any locking to protect db_blkptr:
910 * If it's syncing, then db_last_dirty will be set
911 * so we'll ignore db_blkptr.
912 */
913 ASSERT(MUTEX_HELD(&db->db_mtx));
914 if (db->db_last_dirty)
915 birth_txg = db->db_last_dirty->dr_txg;
916 else if (db->db_blkptr)
917 birth_txg = db->db_blkptr->blk_birth;
918
919 /*
920 * If we don't exist or are in a snapshot, we can't be freed.
921 * Don't pass the bp to dsl_dataset_block_freeable() since we
922 * are holding the db_mtx lock and might deadlock if we are
923 * prefetching a dedup-ed block.
924 */
925 if (birth_txg)
926 return (ds == NULL ||
927 dsl_dataset_block_freeable(ds, NULL, birth_txg));
928 else
929 return (FALSE);
930 }
931
932 void
933 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
934 {
935 arc_buf_t *buf, *obuf;
936 int osize = db->db.db_size;
937 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
938 dnode_t *dn;
939
940 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
941
942 DB_DNODE_ENTER(db);
943 dn = DB_DNODE(db);
944
945 /* XXX does *this* func really need the lock? */
946 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
947
948 /*
949 * This call to dbuf_will_dirty() with the dn_struct_rwlock held
950 * is OK, because there can be no other references to the db
951 * when we are changing its size, so no concurrent DB_FILL can
952 * be happening.
953 */
954 /*
955 * XXX we should be doing a dbuf_read, checking the return
956 * value and returning that up to our callers
957 */
958 dbuf_will_dirty(db, tx);
959
960 /* create the data buffer for the new block */
961 buf = arc_buf_alloc(dn->dn_objset->os_spa, size, db, type);
962
963 /* copy old block data to the new block */
964 obuf = db->db_buf;
965 bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
966 /* zero the remainder */
967 if (size > osize)
968 bzero((uint8_t *)buf->b_data + osize, size - osize);
969
970 mutex_enter(&db->db_mtx);
971 dbuf_set_data(db, buf);
972 VERIFY(arc_buf_remove_ref(obuf, db));
973 db->db.db_size = size;
974
975 if (db->db_level == 0) {
976 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
977 db->db_last_dirty->dt.dl.dr_data = buf;
978 }
979 mutex_exit(&db->db_mtx);
980
981 dnode_willuse_space(dn, size-osize, tx);
982 DB_DNODE_EXIT(db);
983 }
984
985 void
986 dbuf_release_bp(dmu_buf_impl_t *db)
987 {
988 objset_t *os;
989
990 DB_GET_OBJSET(&os, db);
991 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
992 ASSERT(arc_released(os->os_phys_buf) ||
993 list_link_active(&os->os_dsl_dataset->ds_synced_link));
994 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
995
996 (void) arc_release(db->db_buf, db);
997 }
998
999 dbuf_dirty_record_t *
1000 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1001 {
1002 dnode_t *dn;
1003 objset_t *os;
1004 dbuf_dirty_record_t **drp, *dr;
1005 int drop_struct_lock = FALSE;
1006 boolean_t do_free_accounting = B_FALSE;
1007 int txgoff = tx->tx_txg & TXG_MASK;
1008
1009 ASSERT(tx->tx_txg != 0);
1010 ASSERT(!refcount_is_zero(&db->db_holds));
1011 DMU_TX_DIRTY_BUF(tx, db);
1012
1013 DB_DNODE_ENTER(db);
1014 dn = DB_DNODE(db);
1015 /*
1016 * Shouldn't dirty a regular buffer in syncing context. Private
1017 * objects may be dirtied in syncing context, but only if they
1018 * were already pre-dirtied in open context.
1019 */
1020 ASSERT(!dmu_tx_is_syncing(tx) ||
1021 BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1022 DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1023 dn->dn_objset->os_dsl_dataset == NULL);
1024 /*
1025 * We make this assert for private objects as well, but after we
1026 * check if we're already dirty. They are allowed to re-dirty
1027 * in syncing context.
1028 */
1029 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1030 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1031 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1032
1033 mutex_enter(&db->db_mtx);
1034 /*
1035 * XXX make this true for indirects too? The problem is that
1036 * transactions created with dmu_tx_create_assigned() from
1037 * syncing context don't bother holding ahead.
1038 */
1039 ASSERT(db->db_level != 0 ||
1040 db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1041 db->db_state == DB_NOFILL);
1042
1043 mutex_enter(&dn->dn_mtx);
1044 /*
1045 * Don't set dirtyctx to SYNC if we're just modifying this as we
1046 * initialize the objset.
1047 */
1048 if (dn->dn_dirtyctx == DN_UNDIRTIED &&
1049 !BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1050 dn->dn_dirtyctx =
1051 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1052 ASSERT(dn->dn_dirtyctx_firstset == NULL);
1053 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1054 }
1055 mutex_exit(&dn->dn_mtx);
1056
1057 if (db->db_blkid == DMU_SPILL_BLKID)
1058 dn->dn_have_spill = B_TRUE;
1059
1060 /*
1061 * If this buffer is already dirty, we're done.
1062 */
1063 drp = &db->db_last_dirty;
1064 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1065 db->db.db_object == DMU_META_DNODE_OBJECT);
1066 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1067 drp = &dr->dr_next;
1068 if (dr && dr->dr_txg == tx->tx_txg) {
1069 DB_DNODE_EXIT(db);
1070
1071 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1072 /*
1073 * If this buffer has already been written out,
1074 * we now need to reset its state.
1075 */
1076 dbuf_unoverride(dr);
1077 if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1078 db->db_state != DB_NOFILL)
1079 arc_buf_thaw(db->db_buf);
1080 }
1081 mutex_exit(&db->db_mtx);
1082 return (dr);
1083 }
1084
1085 /*
1086 * Only valid if not already dirty.
1087 */
1088 ASSERT(dn->dn_object == 0 ||
1089 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1090 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1091
1092 ASSERT3U(dn->dn_nlevels, >, db->db_level);
1093 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1094 dn->dn_phys->dn_nlevels > db->db_level ||
1095 dn->dn_next_nlevels[txgoff] > db->db_level ||
1096 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1097 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1098
1099 /*
1100 * We should only be dirtying in syncing context if it's the
1101 * mos or we're initializing the os or it's a special object.
1102 * However, we are allowed to dirty in syncing context provided
1103 * we already dirtied it in open context. Hence we must make
1104 * this assertion only if we're not already dirty.
1105 */
1106 os = dn->dn_objset;
1107 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1108 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1109 ASSERT(db->db.db_size != 0);
1110
1111 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1112
1113 if (db->db_blkid != DMU_BONUS_BLKID) {
1114 /*
1115 * Update the accounting.
1116 * Note: we delay "free accounting" until after we drop
1117 * the db_mtx. This keeps us from grabbing other locks
1118 * (and possibly deadlocking) in bp_get_dsize() while
1119 * also holding the db_mtx.
1120 */
1121 dnode_willuse_space(dn, db->db.db_size, tx);
1122 do_free_accounting = dbuf_block_freeable(db);
1123 }
1124
1125 /*
1126 * If this buffer is dirty in an old transaction group we need
1127 * to make a copy of it so that the changes we make in this
1128 * transaction group won't leak out when we sync the older txg.
1129 */
1130 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1131 if (db->db_level == 0) {
1132 void *data_old = db->db_buf;
1133
1134 if (db->db_state != DB_NOFILL) {
1135 if (db->db_blkid == DMU_BONUS_BLKID) {
1136 dbuf_fix_old_data(db, tx->tx_txg);
1137 data_old = db->db.db_data;
1138 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1139 /*
1140 * Release the data buffer from the cache so
1141 * that we can modify it without impacting
1142 * possible other users of this cached data
1143 * block. Note that indirect blocks and
1144 * private objects are not released until the
1145 * syncing state (since they are only modified
1146 * then).
1147 */
1148 arc_release(db->db_buf, db);
1149 dbuf_fix_old_data(db, tx->tx_txg);
1150 data_old = db->db_buf;
1151 }
1152 ASSERT(data_old != NULL);
1153 }
1154 dr->dt.dl.dr_data = data_old;
1155 } else {
1156 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1157 list_create(&dr->dt.di.dr_children,
1158 sizeof (dbuf_dirty_record_t),
1159 offsetof(dbuf_dirty_record_t, dr_dirty_node));
1160 }
1161 dr->dr_dbuf = db;
1162 dr->dr_txg = tx->tx_txg;
1163 dr->dr_next = *drp;
1164 *drp = dr;
1165
1166 /*
1167 * We could have been freed_in_flight between the dbuf_noread
1168 * and dbuf_dirty. We win, as though the dbuf_noread() had
1169 * happened after the free.
1170 */
1171 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1172 db->db_blkid != DMU_SPILL_BLKID) {
1173 mutex_enter(&dn->dn_mtx);
1174 dnode_clear_range(dn, db->db_blkid, 1, tx);
1175 mutex_exit(&dn->dn_mtx);
1176 db->db_freed_in_flight = FALSE;
1177 }
1178
1179 /*
1180 * This buffer is now part of this txg
1181 */
1182 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1183 db->db_dirtycnt += 1;
1184 ASSERT3U(db->db_dirtycnt, <=, 3);
1185
1186 mutex_exit(&db->db_mtx);
1187
1188 if (db->db_blkid == DMU_BONUS_BLKID ||
1189 db->db_blkid == DMU_SPILL_BLKID) {
1190 mutex_enter(&dn->dn_mtx);
1191 ASSERT(!list_link_active(&dr->dr_dirty_node));
1192 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1193 mutex_exit(&dn->dn_mtx);
1194 dnode_setdirty(dn, tx);
1195 DB_DNODE_EXIT(db);
1196 return (dr);
1197 } else if (do_free_accounting) {
1198 blkptr_t *bp = db->db_blkptr;
1199 int64_t willfree = (bp && !BP_IS_HOLE(bp)) ?
1200 bp_get_dsize(os->os_spa, bp) : db->db.db_size;
1201 /*
1202 * This is only a guess -- if the dbuf is dirty
1203 * in a previous txg, we don't know how much
1204 * space it will use on disk yet. We should
1205 * really have the struct_rwlock to access
1206 * db_blkptr, but since this is just a guess,
1207 * it's OK if we get an odd answer.
1208 */
1209 ddt_prefetch(os->os_spa, bp);
1210 dnode_willuse_space(dn, -willfree, tx);
1211 }
1212
1213 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1214 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1215 drop_struct_lock = TRUE;
1216 }
1217
1218 if (db->db_level == 0) {
1219 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1220 ASSERT(dn->dn_maxblkid >= db->db_blkid);
1221 }
1222
1223 if (db->db_level+1 < dn->dn_nlevels) {
1224 dmu_buf_impl_t *parent = db->db_parent;
1225 dbuf_dirty_record_t *di;
1226 int parent_held = FALSE;
1227
1228 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1229 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1230
1231 parent = dbuf_hold_level(dn, db->db_level+1,
1232 db->db_blkid >> epbs, FTAG);
1233 ASSERT(parent != NULL);
1234 parent_held = TRUE;
1235 }
1236 if (drop_struct_lock)
1237 rw_exit(&dn->dn_struct_rwlock);
1238 ASSERT3U(db->db_level+1, ==, parent->db_level);
1239 di = dbuf_dirty(parent, tx);
1240 if (parent_held)
1241 dbuf_rele(parent, FTAG);
1242
1243 mutex_enter(&db->db_mtx);
1244 /* possible race with dbuf_undirty() */
1245 if (db->db_last_dirty == dr ||
1246 dn->dn_object == DMU_META_DNODE_OBJECT) {
1247 mutex_enter(&di->dt.di.dr_mtx);
1248 ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1249 ASSERT(!list_link_active(&dr->dr_dirty_node));
1250 list_insert_tail(&di->dt.di.dr_children, dr);
1251 mutex_exit(&di->dt.di.dr_mtx);
1252 dr->dr_parent = di;
1253 }
1254 mutex_exit(&db->db_mtx);
1255 } else {
1256 ASSERT(db->db_level+1 == dn->dn_nlevels);
1257 ASSERT(db->db_blkid < dn->dn_nblkptr);
1258 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1259 mutex_enter(&dn->dn_mtx);
1260 ASSERT(!list_link_active(&dr->dr_dirty_node));
1261 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1262 mutex_exit(&dn->dn_mtx);
1263 if (drop_struct_lock)
1264 rw_exit(&dn->dn_struct_rwlock);
1265 }
1266
1267 dnode_setdirty(dn, tx);
1268 DB_DNODE_EXIT(db);
1269 return (dr);
1270 }
1271
1272 /*
1273 * Undirty a buffer in the transaction group referenced by the given
1274 * transaction. Return whether this evicted the dbuf.
1275 */
1276 static boolean_t
1277 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1278 {
1279 dnode_t *dn;
1280 uint64_t txg = tx->tx_txg;
1281 dbuf_dirty_record_t *dr, **drp;
1282
1283 ASSERT(txg != 0);
1284 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1285 ASSERT0(db->db_level);
1286 ASSERT(MUTEX_HELD(&db->db_mtx));
1287
1288 /*
1289 * If this buffer is not dirty, we're done.
1290 */
1291 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1292 if (dr->dr_txg <= txg)
1293 break;
1294 if (dr == NULL || dr->dr_txg < txg)
1295 return (B_FALSE);
1296 ASSERT(dr->dr_txg == txg);
1297 ASSERT(dr->dr_dbuf == db);
1298
1299 DB_DNODE_ENTER(db);
1300 dn = DB_DNODE(db);
1301
1302 /*
1303 * Note: This code will probably work even if there are concurrent
1304 * holders, but it is untested in that scenerio, as the ZPL and
1305 * ztest have additional locking (the range locks) that prevents
1306 * that type of concurrent access.
1307 */
1308 ASSERT3U(refcount_count(&db->db_holds), ==, db->db_dirtycnt);
1309
1310 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1311
1312 ASSERT(db->db.db_size != 0);
1313
1314 /* XXX would be nice to fix up dn_towrite_space[] */
1315
1316 *drp = dr->dr_next;
1317
1318 /*
1319 * Note that there are three places in dbuf_dirty()
1320 * where this dirty record may be put on a list.
1321 * Make sure to do a list_remove corresponding to
1322 * every one of those list_insert calls.
1323 */
1324 if (dr->dr_parent) {
1325 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1326 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1327 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1328 } else if (db->db_blkid == DMU_SPILL_BLKID ||
1329 db->db_level+1 == dn->dn_nlevels) {
1330 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1331 mutex_enter(&dn->dn_mtx);
1332 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1333 mutex_exit(&dn->dn_mtx);
1334 }
1335 DB_DNODE_EXIT(db);
1336
1337 if (db->db_state != DB_NOFILL) {
1338 dbuf_unoverride(dr);
1339
1340 ASSERT(db->db_buf != NULL);
1341 ASSERT(dr->dt.dl.dr_data != NULL);
1342 if (dr->dt.dl.dr_data != db->db_buf)
1343 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db));
1344 }
1345 kmem_free(dr, sizeof (dbuf_dirty_record_t));
1346
1347 ASSERT(db->db_dirtycnt > 0);
1348 db->db_dirtycnt -= 1;
1349
1350 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1351 arc_buf_t *buf = db->db_buf;
1352
1353 ASSERT(db->db_state == DB_NOFILL || arc_released(buf));
1354 dbuf_set_data(db, NULL);
1355 VERIFY(arc_buf_remove_ref(buf, db));
1356 dbuf_evict(db);
1357 return (B_TRUE);
1358 }
1359
1360 return (B_FALSE);
1361 }
1362
1363 #pragma weak dmu_buf_will_dirty = dbuf_will_dirty
1364 void
1365 dbuf_will_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1366 {
1367 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1368
1369 ASSERT(tx->tx_txg != 0);
1370 ASSERT(!refcount_is_zero(&db->db_holds));
1371
1372 DB_DNODE_ENTER(db);
1373 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
1374 rf |= DB_RF_HAVESTRUCT;
1375 DB_DNODE_EXIT(db);
1376 (void) dbuf_read(db, NULL, rf);
1377 (void) dbuf_dirty(db, tx);
1378 }
1379
1380 void
1381 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1382 {
1383 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1384
1385 db->db_state = DB_NOFILL;
1386
1387 dmu_buf_will_fill(db_fake, tx);
1388 }
1389
1390 void
1391 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1392 {
1393 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1394
1395 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1396 ASSERT(tx->tx_txg != 0);
1397 ASSERT(db->db_level == 0);
1398 ASSERT(!refcount_is_zero(&db->db_holds));
1399
1400 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1401 dmu_tx_private_ok(tx));
1402
1403 dbuf_noread(db);
1404 (void) dbuf_dirty(db, tx);
1405 }
1406
1407 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1408 /* ARGSUSED */
1409 void
1410 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1411 {
1412 mutex_enter(&db->db_mtx);
1413 DBUF_VERIFY(db);
1414
1415 if (db->db_state == DB_FILL) {
1416 if (db->db_level == 0 && db->db_freed_in_flight) {
1417 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1418 /* we were freed while filling */
1419 /* XXX dbuf_undirty? */
1420 bzero(db->db.db_data, db->db.db_size);
1421 db->db_freed_in_flight = FALSE;
1422 }
1423 db->db_state = DB_CACHED;
1424 cv_broadcast(&db->db_changed);
1425 }
1426 mutex_exit(&db->db_mtx);
1427 }
1428
1429 /*
1430 * Directly assign a provided arc buf to a given dbuf if it's not referenced
1431 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1432 */
1433 void
1434 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
1435 {
1436 ASSERT(!refcount_is_zero(&db->db_holds));
1437 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1438 ASSERT(db->db_level == 0);
1439 ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA);
1440 ASSERT(buf != NULL);
1441 ASSERT(arc_buf_size(buf) == db->db.db_size);
1442 ASSERT(tx->tx_txg != 0);
1443
1444 arc_return_buf(buf, db);
1445 ASSERT(arc_released(buf));
1446
1447 mutex_enter(&db->db_mtx);
1448
1449 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1450 cv_wait(&db->db_changed, &db->db_mtx);
1451
1452 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
1453
1454 if (db->db_state == DB_CACHED &&
1455 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
1456 mutex_exit(&db->db_mtx);
1457 (void) dbuf_dirty(db, tx);
1458 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
1459 VERIFY(arc_buf_remove_ref(buf, db));
1460 xuio_stat_wbuf_copied();
1461 return;
1462 }
1463
1464 xuio_stat_wbuf_nocopy();
1465 if (db->db_state == DB_CACHED) {
1466 dbuf_dirty_record_t *dr = db->db_last_dirty;
1467
1468 ASSERT(db->db_buf != NULL);
1469 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
1470 ASSERT(dr->dt.dl.dr_data == db->db_buf);
1471 if (!arc_released(db->db_buf)) {
1472 ASSERT(dr->dt.dl.dr_override_state ==
1473 DR_OVERRIDDEN);
1474 arc_release(db->db_buf, db);
1475 }
1476 dr->dt.dl.dr_data = buf;
1477 VERIFY(arc_buf_remove_ref(db->db_buf, db));
1478 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
1479 arc_release(db->db_buf, db);
1480 VERIFY(arc_buf_remove_ref(db->db_buf, db));
1481 }
1482 db->db_buf = NULL;
1483 }
1484 ASSERT(db->db_buf == NULL);
1485 dbuf_set_data(db, buf);
1486 db->db_state = DB_FILL;
1487 mutex_exit(&db->db_mtx);
1488 (void) dbuf_dirty(db, tx);
1489 dbuf_fill_done(db, tx);
1490 }
1491
1492 /*
1493 * "Clear" the contents of this dbuf. This will mark the dbuf
1494 * EVICTING and clear *most* of its references. Unfortunetely,
1495 * when we are not holding the dn_dbufs_mtx, we can't clear the
1496 * entry in the dn_dbufs list. We have to wait until dbuf_destroy()
1497 * in this case. For callers from the DMU we will usually see:
1498 * dbuf_clear()->arc_buf_evict()->dbuf_do_evict()->dbuf_destroy()
1499 * For the arc callback, we will usually see:
1500 * dbuf_do_evict()->dbuf_clear();dbuf_destroy()
1501 * Sometimes, though, we will get a mix of these two:
1502 * DMU: dbuf_clear()->arc_buf_evict()
1503 * ARC: dbuf_do_evict()->dbuf_destroy()
1504 */
1505 void
1506 dbuf_clear(dmu_buf_impl_t *db)
1507 {
1508 dnode_t *dn;
1509 dmu_buf_impl_t *parent = db->db_parent;
1510 dmu_buf_impl_t *dndb;
1511 int dbuf_gone = FALSE;
1512
1513 ASSERT(MUTEX_HELD(&db->db_mtx));
1514 ASSERT(refcount_is_zero(&db->db_holds));
1515
1516 dbuf_evict_user(db);
1517
1518 if (db->db_state == DB_CACHED) {
1519 ASSERT(db->db.db_data != NULL);
1520 if (db->db_blkid == DMU_BONUS_BLKID) {
1521 zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
1522 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
1523 }
1524 db->db.db_data = NULL;
1525 db->db_state = DB_UNCACHED;
1526 }
1527
1528 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
1529 ASSERT(db->db_data_pending == NULL);
1530
1531 db->db_state = DB_EVICTING;
1532 db->db_blkptr = NULL;
1533
1534 DB_DNODE_ENTER(db);
1535 dn = DB_DNODE(db);
1536 dndb = dn->dn_dbuf;
1537 if (db->db_blkid != DMU_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) {
1538 list_remove(&dn->dn_dbufs, db);
1539 (void) atomic_dec_32_nv(&dn->dn_dbufs_count);
1540 membar_producer();
1541 DB_DNODE_EXIT(db);
1542 /*
1543 * Decrementing the dbuf count means that the hold corresponding
1544 * to the removed dbuf is no longer discounted in dnode_move(),
1545 * so the dnode cannot be moved until after we release the hold.
1546 * The membar_producer() ensures visibility of the decremented
1547 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
1548 * release any lock.
1549 */
1550 dnode_rele(dn, db);
1551 db->db_dnode_handle = NULL;
1552 } else {
1553 DB_DNODE_EXIT(db);
1554 }
1555
1556 if (db->db_buf)
1557 dbuf_gone = arc_buf_evict(db->db_buf);
1558
1559 if (!dbuf_gone)
1560 mutex_exit(&db->db_mtx);
1561
1562 /*
1563 * If this dbuf is referenced from an indirect dbuf,
1564 * decrement the ref count on the indirect dbuf.
1565 */
1566 if (parent && parent != dndb)
1567 dbuf_rele(parent, db);
1568 }
1569
1570 static int
1571 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
1572 dmu_buf_impl_t **parentp, blkptr_t **bpp)
1573 {
1574 int nlevels, epbs;
1575
1576 *parentp = NULL;
1577 *bpp = NULL;
1578
1579 ASSERT(blkid != DMU_BONUS_BLKID);
1580
1581 if (blkid == DMU_SPILL_BLKID) {
1582 mutex_enter(&dn->dn_mtx);
1583 if (dn->dn_have_spill &&
1584 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
1585 *bpp = &dn->dn_phys->dn_spill;
1586 else
1587 *bpp = NULL;
1588 dbuf_add_ref(dn->dn_dbuf, NULL);
1589 *parentp = dn->dn_dbuf;
1590 mutex_exit(&dn->dn_mtx);
1591 return (0);
1592 }
1593
1594 if (dn->dn_phys->dn_nlevels == 0)
1595 nlevels = 1;
1596 else
1597 nlevels = dn->dn_phys->dn_nlevels;
1598
1599 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1600
1601 ASSERT3U(level * epbs, <, 64);
1602 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1603 if (level >= nlevels ||
1604 (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
1605 /* the buffer has no parent yet */
1606 return (SET_ERROR(ENOENT));
1607 } else if (level < nlevels-1) {
1608 /* this block is referenced from an indirect block */
1609 int err = dbuf_hold_impl(dn, level+1,
1610 blkid >> epbs, fail_sparse, NULL, parentp);
1611 if (err)
1612 return (err);
1613 err = dbuf_read(*parentp, NULL,
1614 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
1615 if (err) {
1616 dbuf_rele(*parentp, NULL);
1617 *parentp = NULL;
1618 return (err);
1619 }
1620 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
1621 (blkid & ((1ULL << epbs) - 1));
1622 return (0);
1623 } else {
1624 /* the block is referenced from the dnode */
1625 ASSERT3U(level, ==, nlevels-1);
1626 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
1627 blkid < dn->dn_phys->dn_nblkptr);
1628 if (dn->dn_dbuf) {
1629 dbuf_add_ref(dn->dn_dbuf, NULL);
1630 *parentp = dn->dn_dbuf;
1631 }
1632 *bpp = &dn->dn_phys->dn_blkptr[blkid];
1633 return (0);
1634 }
1635 }
1636
1637 static dmu_buf_impl_t *
1638 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
1639 dmu_buf_impl_t *parent, blkptr_t *blkptr)
1640 {
1641 objset_t *os = dn->dn_objset;
1642 dmu_buf_impl_t *db, *odb;
1643
1644 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1645 ASSERT(dn->dn_type != DMU_OT_NONE);
1646
1647 db = kmem_cache_alloc(dbuf_cache, KM_SLEEP);
1648
1649 db->db_objset = os;
1650 db->db.db_object = dn->dn_object;
1651 db->db_level = level;
1652 db->db_blkid = blkid;
1653 db->db_last_dirty = NULL;
1654 db->db_dirtycnt = 0;
1655 db->db_dnode_handle = dn->dn_handle;
1656 db->db_parent = parent;
1657 db->db_blkptr = blkptr;
1658
1659 db->db_user_ptr = NULL;
1660 db->db_user_data_ptr_ptr = NULL;
1661 db->db_evict_func = NULL;
1662 db->db_immediate_evict = 0;
1663 db->db_freed_in_flight = 0;
1664
1665 if (blkid == DMU_BONUS_BLKID) {
1666 ASSERT3P(parent, ==, dn->dn_dbuf);
1667 db->db.db_size = DN_MAX_BONUSLEN -
1668 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
1669 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
1670 db->db.db_offset = DMU_BONUS_BLKID;
1671 db->db_state = DB_UNCACHED;
1672 /* the bonus dbuf is not placed in the hash table */
1673 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1674 return (db);
1675 } else if (blkid == DMU_SPILL_BLKID) {
1676 db->db.db_size = (blkptr != NULL) ?
1677 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
1678 db->db.db_offset = 0;
1679 } else {
1680 int blocksize =
1681 db->db_level ? 1<<dn->dn_indblkshift : dn->dn_datablksz;
1682 db->db.db_size = blocksize;
1683 db->db.db_offset = db->db_blkid * blocksize;
1684 }
1685
1686 /*
1687 * Hold the dn_dbufs_mtx while we get the new dbuf
1688 * in the hash table *and* added to the dbufs list.
1689 * This prevents a possible deadlock with someone
1690 * trying to look up this dbuf before its added to the
1691 * dn_dbufs list.
1692 */
1693 mutex_enter(&dn->dn_dbufs_mtx);
1694 db->db_state = DB_EVICTING;
1695 if ((odb = dbuf_hash_insert(db)) != NULL) {
1696 /* someone else inserted it first */
1697 kmem_cache_free(dbuf_cache, db);
1698 mutex_exit(&dn->dn_dbufs_mtx);
1699 return (odb);
1700 }
1701 list_insert_head(&dn->dn_dbufs, db);
1702 db->db_state = DB_UNCACHED;
1703 mutex_exit(&dn->dn_dbufs_mtx);
1704 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1705
1706 if (parent && parent != dn->dn_dbuf)
1707 dbuf_add_ref(parent, db);
1708
1709 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1710 refcount_count(&dn->dn_holds) > 0);
1711 (void) refcount_add(&dn->dn_holds, db);
1712 (void) atomic_inc_32_nv(&dn->dn_dbufs_count);
1713
1714 dprintf_dbuf(db, "db=%p\n", db);
1715
1716 return (db);
1717 }
1718
1719 static int
1720 dbuf_do_evict(void *private)
1721 {
1722 arc_buf_t *buf = private;
1723 dmu_buf_impl_t *db = buf->b_private;
1724
1725 if (!MUTEX_HELD(&db->db_mtx))
1726 mutex_enter(&db->db_mtx);
1727
1728 ASSERT(refcount_is_zero(&db->db_holds));
1729
1730 if (db->db_state != DB_EVICTING) {
1731 ASSERT(db->db_state == DB_CACHED);
1732 DBUF_VERIFY(db);
1733 db->db_buf = NULL;
1734 dbuf_evict(db);
1735 } else {
1736 mutex_exit(&db->db_mtx);
1737 dbuf_destroy(db);
1738 }
1739 return (0);
1740 }
1741
1742 static void
1743 dbuf_destroy(dmu_buf_impl_t *db)
1744 {
1745 ASSERT(refcount_is_zero(&db->db_holds));
1746
1747 if (db->db_blkid != DMU_BONUS_BLKID) {
1748 /*
1749 * If this dbuf is still on the dn_dbufs list,
1750 * remove it from that list.
1751 */
1752 if (db->db_dnode_handle != NULL) {
1753 dnode_t *dn;
1754
1755 DB_DNODE_ENTER(db);
1756 dn = DB_DNODE(db);
1757 mutex_enter(&dn->dn_dbufs_mtx);
1758 list_remove(&dn->dn_dbufs, db);
1759 (void) atomic_dec_32_nv(&dn->dn_dbufs_count);
1760 mutex_exit(&dn->dn_dbufs_mtx);
1761 DB_DNODE_EXIT(db);
1762 /*
1763 * Decrementing the dbuf count means that the hold
1764 * corresponding to the removed dbuf is no longer
1765 * discounted in dnode_move(), so the dnode cannot be
1766 * moved until after we release the hold.
1767 */
1768 dnode_rele(dn, db);
1769 db->db_dnode_handle = NULL;
1770 }
1771 dbuf_hash_remove(db);
1772 }
1773 db->db_parent = NULL;
1774 db->db_buf = NULL;
1775
1776 ASSERT(!list_link_active(&db->db_link));
1777 ASSERT(db->db.db_data == NULL);
1778 ASSERT(db->db_hash_next == NULL);
1779 ASSERT(db->db_blkptr == NULL);
1780 ASSERT(db->db_data_pending == NULL);
1781
1782 kmem_cache_free(dbuf_cache, db);
1783 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1784 }
1785
1786 void
1787 dbuf_prefetch(dnode_t *dn, uint64_t blkid)
1788 {
1789 dmu_buf_impl_t *db = NULL;
1790 blkptr_t *bp = NULL;
1791
1792 ASSERT(blkid != DMU_BONUS_BLKID);
1793 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1794
1795 if (dnode_block_freed(dn, blkid))
1796 return;
1797
1798 /* dbuf_find() returns with db_mtx held */
1799 if (db = dbuf_find(dn, 0, blkid)) {
1800 /*
1801 * This dbuf is already in the cache. We assume that
1802 * it is already CACHED, or else about to be either
1803 * read or filled.
1804 */
1805 mutex_exit(&db->db_mtx);
1806 return;
1807 }
1808
1809 if (dbuf_findbp(dn, 0, blkid, TRUE, &db, &bp) == 0) {
1810 if (bp && !BP_IS_HOLE(bp)) {
1811 int priority = dn->dn_type == DMU_OT_DDT_ZAP ?
1812 ZIO_PRIORITY_DDT_PREFETCH : ZIO_PRIORITY_ASYNC_READ;
1813 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
1814 uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH;
1815 zbookmark_t zb;
1816
1817 SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
1818 dn->dn_object, 0, blkid);
1819
1820 (void) arc_read(NULL, dn->dn_objset->os_spa,
1821 bp, NULL, NULL, priority,
1822 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1823 &aflags, &zb);
1824 }
1825 if (db)
1826 dbuf_rele(db, NULL);
1827 }
1828 }
1829
1830 /*
1831 * Returns with db_holds incremented, and db_mtx not held.
1832 * Note: dn_struct_rwlock must be held.
1833 */
1834 int
1835 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse,
1836 void *tag, dmu_buf_impl_t **dbp)
1837 {
1838 dmu_buf_impl_t *db, *parent = NULL;
1839
1840 ASSERT(blkid != DMU_BONUS_BLKID);
1841 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1842 ASSERT3U(dn->dn_nlevels, >, level);
1843
1844 *dbp = NULL;
1845 top:
1846 /* dbuf_find() returns with db_mtx held */
1847 db = dbuf_find(dn, level, blkid);
1848
1849 if (db == NULL) {
1850 blkptr_t *bp = NULL;
1851 int err;
1852
1853 ASSERT3P(parent, ==, NULL);
1854 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
1855 if (fail_sparse) {
1856 if (err == 0 && bp && BP_IS_HOLE(bp))
1857 err = SET_ERROR(ENOENT);
1858 if (err) {
1859 if (parent)
1860 dbuf_rele(parent, NULL);
1861 return (err);
1862 }
1863 }
1864 if (err && err != ENOENT)
1865 return (err);
1866 db = dbuf_create(dn, level, blkid, parent, bp);
1867 }
1868
1869 if (db->db_buf && refcount_is_zero(&db->db_holds)) {
1870 arc_buf_add_ref(db->db_buf, db);
1871 if (db->db_buf->b_data == NULL) {
1872 dbuf_clear(db);
1873 if (parent) {
1874 dbuf_rele(parent, NULL);
1875 parent = NULL;
1876 }
1877 goto top;
1878 }
1879 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
1880 }
1881
1882 ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
1883
1884 /*
1885 * If this buffer is currently syncing out, and we are are
1886 * still referencing it from db_data, we need to make a copy
1887 * of it in case we decide we want to dirty it again in this txg.
1888 */
1889 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1890 dn->dn_object != DMU_META_DNODE_OBJECT &&
1891 db->db_state == DB_CACHED && db->db_data_pending) {
1892 dbuf_dirty_record_t *dr = db->db_data_pending;
1893
1894 if (dr->dt.dl.dr_data == db->db_buf) {
1895 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1896
1897 dbuf_set_data(db,
1898 arc_buf_alloc(dn->dn_objset->os_spa,
1899 db->db.db_size, db, type));
1900 bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
1901 db->db.db_size);
1902 }
1903 }
1904
1905 (void) refcount_add(&db->db_holds, tag);
1906 dbuf_update_data(db);
1907 DBUF_VERIFY(db);
1908 mutex_exit(&db->db_mtx);
1909
1910 /* NOTE: we can't rele the parent until after we drop the db_mtx */
1911 if (parent)
1912 dbuf_rele(parent, NULL);
1913
1914 ASSERT3P(DB_DNODE(db), ==, dn);
1915 ASSERT3U(db->db_blkid, ==, blkid);
1916 ASSERT3U(db->db_level, ==, level);
1917 *dbp = db;
1918
1919 return (0);
1920 }
1921
1922 dmu_buf_impl_t *
1923 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
1924 {
1925 dmu_buf_impl_t *db;
1926 int err = dbuf_hold_impl(dn, 0, blkid, FALSE, tag, &db);
1927 return (err ? NULL : db);
1928 }
1929
1930 dmu_buf_impl_t *
1931 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
1932 {
1933 dmu_buf_impl_t *db;
1934 int err = dbuf_hold_impl(dn, level, blkid, FALSE, tag, &db);
1935 return (err ? NULL : db);
1936 }
1937
1938 void
1939 dbuf_create_bonus(dnode_t *dn)
1940 {
1941 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1942
1943 ASSERT(dn->dn_bonus == NULL);
1944 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
1945 }
1946
1947 int
1948 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
1949 {
1950 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1951 dnode_t *dn;
1952
1953 if (db->db_blkid != DMU_SPILL_BLKID)
1954 return (SET_ERROR(ENOTSUP));
1955 if (blksz == 0)
1956 blksz = SPA_MINBLOCKSIZE;
1957 if (blksz > SPA_MAXBLOCKSIZE)
1958 blksz = SPA_MAXBLOCKSIZE;
1959 else
1960 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
1961
1962 DB_DNODE_ENTER(db);
1963 dn = DB_DNODE(db);
1964 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1965 dbuf_new_size(db, blksz, tx);
1966 rw_exit(&dn->dn_struct_rwlock);
1967 DB_DNODE_EXIT(db);
1968
1969 return (0);
1970 }
1971
1972 void
1973 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
1974 {
1975 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
1976 }
1977
1978 #pragma weak dmu_buf_add_ref = dbuf_add_ref
1979 void
1980 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
1981 {
1982 int64_t holds = refcount_add(&db->db_holds, tag);
1983 ASSERT(holds > 1);
1984 }
1985
1986 /*
1987 * If you call dbuf_rele() you had better not be referencing the dnode handle
1988 * unless you have some other direct or indirect hold on the dnode. (An indirect
1989 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
1990 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
1991 * dnode's parent dbuf evicting its dnode handles.
1992 */
1993 #pragma weak dmu_buf_rele = dbuf_rele
1994 void
1995 dbuf_rele(dmu_buf_impl_t *db, void *tag)
1996 {
1997 mutex_enter(&db->db_mtx);
1998 dbuf_rele_and_unlock(db, tag);
1999 }
2000
2001 /*
2002 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2003 * db_dirtycnt and db_holds to be updated atomically.
2004 */
2005 void
2006 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2007 {
2008 int64_t holds;
2009
2010 ASSERT(MUTEX_HELD(&db->db_mtx));
2011 DBUF_VERIFY(db);
2012
2013 /*
2014 * Remove the reference to the dbuf before removing its hold on the
2015 * dnode so we can guarantee in dnode_move() that a referenced bonus
2016 * buffer has a corresponding dnode hold.
2017 */
2018 holds = refcount_remove(&db->db_holds, tag);
2019 ASSERT(holds >= 0);
2020
2021 /*
2022 * We can't freeze indirects if there is a possibility that they
2023 * may be modified in the current syncing context.
2024 */
2025 if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0))
2026 arc_buf_freeze(db->db_buf);
2027
2028 if (holds == db->db_dirtycnt &&
2029 db->db_level == 0 && db->db_immediate_evict)
2030 dbuf_evict_user(db);
2031
2032 if (holds == 0) {
2033 if (db->db_blkid == DMU_BONUS_BLKID) {
2034 mutex_exit(&db->db_mtx);
2035
2036 /*
2037 * If the dnode moves here, we cannot cross this barrier
2038 * until the move completes.
2039 */
2040 DB_DNODE_ENTER(db);
2041 (void) atomic_dec_32_nv(&DB_DNODE(db)->dn_dbufs_count);
2042 DB_DNODE_EXIT(db);
2043 /*
2044 * The bonus buffer's dnode hold is no longer discounted
2045 * in dnode_move(). The dnode cannot move until after
2046 * the dnode_rele().
2047 */
2048 dnode_rele(DB_DNODE(db), db);
2049 } else if (db->db_buf == NULL) {
2050 /*
2051 * This is a special case: we never associated this
2052 * dbuf with any data allocated from the ARC.
2053 */
2054 ASSERT(db->db_state == DB_UNCACHED ||
2055 db->db_state == DB_NOFILL);
2056 dbuf_evict(db);
2057 } else if (arc_released(db->db_buf)) {
2058 arc_buf_t *buf = db->db_buf;
2059 /*
2060 * This dbuf has anonymous data associated with it.
2061 */
2062 dbuf_set_data(db, NULL);
2063 VERIFY(arc_buf_remove_ref(buf, db));
2064 dbuf_evict(db);
2065 } else {
2066 VERIFY(!arc_buf_remove_ref(db->db_buf, db));
2067
2068 /*
2069 * A dbuf will be eligible for eviction if either the
2070 * 'primarycache' property is set or a duplicate
2071 * copy of this buffer is already cached in the arc.
2072 *
2073 * In the case of the 'primarycache' a buffer
2074 * is considered for eviction if it matches the
2075 * criteria set in the property.
2076 *
2077 * To decide if our buffer is considered a
2078 * duplicate, we must call into the arc to determine
2079 * if multiple buffers are referencing the same
2080 * block on-disk. If so, then we simply evict
2081 * ourselves.
2082 */
2083 if (!DBUF_IS_CACHEABLE(db) ||
2084 arc_buf_eviction_needed(db->db_buf))
2085 dbuf_clear(db);
2086 else
2087 mutex_exit(&db->db_mtx);
2088 }
2089 } else {
2090 mutex_exit(&db->db_mtx);
2091 }
2092 }
2093
2094 #pragma weak dmu_buf_refcount = dbuf_refcount
2095 uint64_t
2096 dbuf_refcount(dmu_buf_impl_t *db)
2097 {
2098 return (refcount_count(&db->db_holds));
2099 }
2100
2101 void *
2102 dmu_buf_set_user(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr,
2103 dmu_buf_evict_func_t *evict_func)
2104 {
2105 return (dmu_buf_update_user(db_fake, NULL, user_ptr,
2106 user_data_ptr_ptr, evict_func));
2107 }
2108
2109 void *
2110 dmu_buf_set_user_ie(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr,
2111 dmu_buf_evict_func_t *evict_func)
2112 {
2113 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2114
2115 db->db_immediate_evict = TRUE;
2116 return (dmu_buf_update_user(db_fake, NULL, user_ptr,
2117 user_data_ptr_ptr, evict_func));
2118 }
2119
2120 void *
2121 dmu_buf_update_user(dmu_buf_t *db_fake, void *old_user_ptr, void *user_ptr,
2122 void *user_data_ptr_ptr, dmu_buf_evict_func_t *evict_func)
2123 {
2124 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2125 ASSERT(db->db_level == 0);
2126
2127 ASSERT((user_ptr == NULL) == (evict_func == NULL));
2128
2129 mutex_enter(&db->db_mtx);
2130
2131 if (db->db_user_ptr == old_user_ptr) {
2132 db->db_user_ptr = user_ptr;
2133 db->db_user_data_ptr_ptr = user_data_ptr_ptr;
2134 db->db_evict_func = evict_func;
2135
2136 dbuf_update_data(db);
2137 } else {
2138 old_user_ptr = db->db_user_ptr;
2139 }
2140
2141 mutex_exit(&db->db_mtx);
2142 return (old_user_ptr);
2143 }
2144
2145 void *
2146 dmu_buf_get_user(dmu_buf_t *db_fake)
2147 {
2148 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2149 ASSERT(!refcount_is_zero(&db->db_holds));
2150
2151 return (db->db_user_ptr);
2152 }
2153
2154 boolean_t
2155 dmu_buf_freeable(dmu_buf_t *dbuf)
2156 {
2157 boolean_t res = B_FALSE;
2158 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2159
2160 if (db->db_blkptr)
2161 res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset,
2162 db->db_blkptr, db->db_blkptr->blk_birth);
2163
2164 return (res);
2165 }
2166
2167 blkptr_t *
2168 dmu_buf_get_blkptr(dmu_buf_t *db)
2169 {
2170 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2171 return (dbi->db_blkptr);
2172 }
2173
2174 static void
2175 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
2176 {
2177 /* ASSERT(dmu_tx_is_syncing(tx) */
2178 ASSERT(MUTEX_HELD(&db->db_mtx));
2179
2180 if (db->db_blkptr != NULL)
2181 return;
2182
2183 if (db->db_blkid == DMU_SPILL_BLKID) {
2184 db->db_blkptr = &dn->dn_phys->dn_spill;
2185 BP_ZERO(db->db_blkptr);
2186 return;
2187 }
2188 if (db->db_level == dn->dn_phys->dn_nlevels-1) {
2189 /*
2190 * This buffer was allocated at a time when there was
2191 * no available blkptrs from the dnode, or it was
2192 * inappropriate to hook it in (i.e., nlevels mis-match).
2193 */
2194 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
2195 ASSERT(db->db_parent == NULL);
2196 db->db_parent = dn->dn_dbuf;
2197 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
2198 DBUF_VERIFY(db);
2199 } else {
2200 dmu_buf_impl_t *parent = db->db_parent;
2201 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2202
2203 ASSERT(dn->dn_phys->dn_nlevels > 1);
2204 if (parent == NULL) {
2205 mutex_exit(&db->db_mtx);
2206 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2207 (void) dbuf_hold_impl(dn, db->db_level+1,
2208 db->db_blkid >> epbs, FALSE, db, &parent);
2209 rw_exit(&dn->dn_struct_rwlock);
2210 mutex_enter(&db->db_mtx);
2211 db->db_parent = parent;
2212 }
2213 db->db_blkptr = (blkptr_t *)parent->db.db_data +
2214 (db->db_blkid & ((1ULL << epbs) - 1));
2215 DBUF_VERIFY(db);
2216 }
2217 }
2218
2219 static void
2220 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2221 {
2222 dmu_buf_impl_t *db = dr->dr_dbuf;
2223 dnode_t *dn;
2224 zio_t *zio;
2225
2226 ASSERT(dmu_tx_is_syncing(tx));
2227
2228 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2229
2230 mutex_enter(&db->db_mtx);
2231
2232 ASSERT(db->db_level > 0);
2233 DBUF_VERIFY(db);
2234
2235 /* Read the block if it hasn't been read yet. */
2236 if (db->db_buf == NULL) {
2237 mutex_exit(&db->db_mtx);
2238 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
2239 mutex_enter(&db->db_mtx);
2240 }
2241 ASSERT3U(db->db_state, ==, DB_CACHED);
2242 ASSERT(db->db_buf != NULL);
2243
2244 DB_DNODE_ENTER(db);
2245 dn = DB_DNODE(db);
2246 /* Indirect block size must match what the dnode thinks it is. */
2247 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2248 dbuf_check_blkptr(dn, db);
2249 DB_DNODE_EXIT(db);
2250
2251 /* Provide the pending dirty record to child dbufs */
2252 db->db_data_pending = dr;
2253
2254 mutex_exit(&db->db_mtx);
2255 dbuf_write(dr, db->db_buf, tx);
2256
2257 zio = dr->dr_zio;
2258 mutex_enter(&dr->dt.di.dr_mtx);
2259 dbuf_sync_list(&dr->dt.di.dr_children, tx);
2260 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2261 mutex_exit(&dr->dt.di.dr_mtx);
2262 zio_nowait(zio);
2263 }
2264
2265 static void
2266 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2267 {
2268 arc_buf_t **datap = &dr->dt.dl.dr_data;
2269 dmu_buf_impl_t *db = dr->dr_dbuf;
2270 dnode_t *dn;
2271 objset_t *os;
2272 uint64_t txg = tx->tx_txg;
2273
2274 ASSERT(dmu_tx_is_syncing(tx));
2275
2276 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2277
2278 mutex_enter(&db->db_mtx);
2279 /*
2280 * To be synced, we must be dirtied. But we
2281 * might have been freed after the dirty.
2282 */
2283 if (db->db_state == DB_UNCACHED) {
2284 /* This buffer has been freed since it was dirtied */
2285 ASSERT(db->db.db_data == NULL);
2286 } else if (db->db_state == DB_FILL) {
2287 /* This buffer was freed and is now being re-filled */
2288 ASSERT(db->db.db_data != dr->dt.dl.dr_data);
2289 } else {
2290 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
2291 }
2292 DBUF_VERIFY(db);
2293
2294 DB_DNODE_ENTER(db);
2295 dn = DB_DNODE(db);
2296
2297 if (db->db_blkid == DMU_SPILL_BLKID) {
2298 mutex_enter(&dn->dn_mtx);
2299 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
2300 mutex_exit(&dn->dn_mtx);
2301 }
2302
2303 /*
2304 * If this is a bonus buffer, simply copy the bonus data into the
2305 * dnode. It will be written out when the dnode is synced (and it
2306 * will be synced, since it must have been dirty for dbuf_sync to
2307 * be called).
2308 */
2309 if (db->db_blkid == DMU_BONUS_BLKID) {
2310 dbuf_dirty_record_t **drp;
2311
2312 ASSERT(*datap != NULL);
2313 ASSERT0(db->db_level);
2314 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN);
2315 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen);
2316 DB_DNODE_EXIT(db);
2317
2318 if (*datap != db->db.db_data) {
2319 zio_buf_free(*datap, DN_MAX_BONUSLEN);
2320 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
2321 }
2322 db->db_data_pending = NULL;
2323 drp = &db->db_last_dirty;
2324 while (*drp != dr)
2325 drp = &(*drp)->dr_next;
2326 ASSERT(dr->dr_next == NULL);
2327 ASSERT(dr->dr_dbuf == db);
2328 *drp = dr->dr_next;
2329 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2330 ASSERT(db->db_dirtycnt > 0);
2331 db->db_dirtycnt -= 1;
2332 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
2333 return;
2334 }
2335
2336 os = dn->dn_objset;
2337
2338 /*
2339 * This function may have dropped the db_mtx lock allowing a dmu_sync
2340 * operation to sneak in. As a result, we need to ensure that we
2341 * don't check the dr_override_state until we have returned from
2342 * dbuf_check_blkptr.
2343 */
2344 dbuf_check_blkptr(dn, db);
2345
2346 /*
2347 * If this buffer is in the middle of an immediate write,
2348 * wait for the synchronous IO to complete.
2349 */
2350 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
2351 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
2352 cv_wait(&db->db_changed, &db->db_mtx);
2353 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
2354 }
2355
2356 if (db->db_state != DB_NOFILL &&
2357 dn->dn_object != DMU_META_DNODE_OBJECT &&
2358 refcount_count(&db->db_holds) > 1 &&
2359 dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
2360 *datap == db->db_buf) {
2361 /*
2362 * If this buffer is currently "in use" (i.e., there
2363 * are active holds and db_data still references it),
2364 * then make a copy before we start the write so that
2365 * any modifications from the open txg will not leak
2366 * into this write.
2367 *
2368 * NOTE: this copy does not need to be made for
2369 * objects only modified in the syncing context (e.g.
2370 * DNONE_DNODE blocks).
2371 */
2372 int blksz = arc_buf_size(*datap);
2373 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2374 *datap = arc_buf_alloc(os->os_spa, blksz, db, type);
2375 bcopy(db->db.db_data, (*datap)->b_data, blksz);
2376 }
2377 db->db_data_pending = dr;
2378
2379 mutex_exit(&db->db_mtx);
2380
2381 dbuf_write(dr, *datap, tx);
2382
2383 ASSERT(!list_link_active(&dr->dr_dirty_node));
2384 if (dn->dn_object == DMU_META_DNODE_OBJECT) {
2385 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
2386 DB_DNODE_EXIT(db);
2387 } else {
2388 /*
2389 * Although zio_nowait() does not "wait for an IO", it does
2390 * initiate the IO. If this is an empty write it seems plausible
2391 * that the IO could actually be completed before the nowait
2392 * returns. We need to DB_DNODE_EXIT() first in case
2393 * zio_nowait() invalidates the dbuf.
2394 */
2395 DB_DNODE_EXIT(db);
2396 zio_nowait(dr->dr_zio);
2397 }
2398 }
2399
2400 void
2401 dbuf_sync_list(list_t *list, dmu_tx_t *tx)
2402 {
2403 dbuf_dirty_record_t *dr;
2404
2405 while (dr = list_head(list)) {
2406 if (dr->dr_zio != NULL) {
2407 /*
2408 * If we find an already initialized zio then we
2409 * are processing the meta-dnode, and we have finished.
2410 * The dbufs for all dnodes are put back on the list
2411 * during processing, so that we can zio_wait()
2412 * these IOs after initiating all child IOs.
2413 */
2414 ASSERT3U(dr->dr_dbuf->db.db_object, ==,
2415 DMU_META_DNODE_OBJECT);
2416 break;
2417 }
2418 list_remove(list, dr);
2419 if (dr->dr_dbuf->db_level > 0)
2420 dbuf_sync_indirect(dr, tx);
2421 else
2422 dbuf_sync_leaf(dr, tx);
2423 }
2424 }
2425
2426 /* ARGSUSED */
2427 static void
2428 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
2429 {
2430 dmu_buf_impl_t *db = vdb;
2431 dnode_t *dn;
2432 blkptr_t *bp = zio->io_bp;
2433 blkptr_t *bp_orig = &zio->io_bp_orig;
2434 spa_t *spa = zio->io_spa;
2435 int64_t delta;
2436 uint64_t fill = 0;
2437 int i;
2438
2439 ASSERT(db->db_blkptr == bp);
2440
2441 DB_DNODE_ENTER(db);
2442 dn = DB_DNODE(db);
2443 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
2444 dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
2445 zio->io_prev_space_delta = delta;
2446
2447 if (BP_IS_HOLE(bp)) {
2448 ASSERT(bp->blk_fill == 0);
2449 DB_DNODE_EXIT(db);
2450 return;
2451 }
2452
2453 ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
2454 BP_GET_TYPE(bp) == dn->dn_type) ||
2455 (db->db_blkid == DMU_SPILL_BLKID &&
2456 BP_GET_TYPE(bp) == dn->dn_bonustype));
2457 ASSERT(BP_GET_LEVEL(bp) == db->db_level);
2458
2459 mutex_enter(&db->db_mtx);
2460
2461 #ifdef ZFS_DEBUG
2462 if (db->db_blkid == DMU_SPILL_BLKID) {
2463 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2464 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2465 db->db_blkptr == &dn->dn_phys->dn_spill);
2466 }
2467 #endif
2468
2469 if (db->db_level == 0) {
2470 mutex_enter(&dn->dn_mtx);
2471 if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
2472 db->db_blkid != DMU_SPILL_BLKID)
2473 dn->dn_phys->dn_maxblkid = db->db_blkid;
2474 mutex_exit(&dn->dn_mtx);
2475
2476 if (dn->dn_type == DMU_OT_DNODE) {
2477 dnode_phys_t *dnp = db->db.db_data;
2478 for (i = db->db.db_size >> DNODE_SHIFT; i > 0;
2479 i--, dnp++) {
2480 if (dnp->dn_type != DMU_OT_NONE)
2481 fill++;
2482 }
2483 } else {
2484 fill = 1;
2485 }
2486 } else {
2487 blkptr_t *ibp = db->db.db_data;
2488 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2489 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
2490 if (BP_IS_HOLE(ibp))
2491 continue;
2492 fill += ibp->blk_fill;
2493 }
2494 }
2495 DB_DNODE_EXIT(db);
2496
2497 bp->blk_fill = fill;
2498
2499 mutex_exit(&db->db_mtx);
2500 }
2501
2502 /* ARGSUSED */
2503 static void
2504 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
2505 {
2506 dmu_buf_impl_t *db = vdb;
2507 blkptr_t *bp = zio->io_bp;
2508 blkptr_t *bp_orig = &zio->io_bp_orig;
2509 uint64_t txg = zio->io_txg;
2510 dbuf_dirty_record_t **drp, *dr;
2511
2512 ASSERT0(zio->io_error);
2513 ASSERT(db->db_blkptr == bp);
2514
2515 /*
2516 * For nopwrites and rewrites we ensure that the bp matches our
2517 * original and bypass all the accounting.
2518 */
2519 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
2520 ASSERT(BP_EQUAL(bp, bp_orig));
2521 } else {
2522 objset_t *os;
2523 dsl_dataset_t *ds;
2524 dmu_tx_t *tx;
2525
2526 DB_GET_OBJSET(&os, db);
2527 ds = os->os_dsl_dataset;
2528 tx = os->os_synctx;
2529
2530 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
2531 dsl_dataset_block_born(ds, bp, tx);
2532 }
2533
2534 mutex_enter(&db->db_mtx);
2535
2536 DBUF_VERIFY(db);
2537
2538 drp = &db->db_last_dirty;
2539 while ((dr = *drp) != db->db_data_pending)
2540 drp = &dr->dr_next;
2541 ASSERT(!list_link_active(&dr->dr_dirty_node));
2542 ASSERT(dr->dr_txg == txg);
2543 ASSERT(dr->dr_dbuf == db);
2544 ASSERT(dr->dr_next == NULL);
2545 *drp = dr->dr_next;
2546
2547 #ifdef ZFS_DEBUG
2548 if (db->db_blkid == DMU_SPILL_BLKID) {
2549 dnode_t *dn;
2550
2551 DB_DNODE_ENTER(db);
2552 dn = DB_DNODE(db);
2553 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2554 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2555 db->db_blkptr == &dn->dn_phys->dn_spill);
2556 DB_DNODE_EXIT(db);
2557 }
2558 #endif
2559
2560 if (db->db_level == 0) {
2561 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2562 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
2563 if (db->db_state != DB_NOFILL) {
2564 if (dr->dt.dl.dr_data != db->db_buf)
2565 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data,
2566 db));
2567 else if (!arc_released(db->db_buf))
2568 arc_set_callback(db->db_buf, dbuf_do_evict, db);
2569 }
2570 } else {
2571 dnode_t *dn;
2572
2573 DB_DNODE_ENTER(db);
2574 dn = DB_DNODE(db);
2575 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2576 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2577 if (!BP_IS_HOLE(db->db_blkptr)) {
2578 int epbs =
2579 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2580 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
2581 db->db.db_size);
2582 ASSERT3U(dn->dn_phys->dn_maxblkid
2583 >> (db->db_level * epbs), >=, db->db_blkid);
2584 arc_set_callback(db->db_buf, dbuf_do_evict, db);
2585 }
2586 DB_DNODE_EXIT(db);
2587 mutex_destroy(&dr->dt.di.dr_mtx);
2588 list_destroy(&dr->dt.di.dr_children);
2589 }
2590 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2591
2592 cv_broadcast(&db->db_changed);
2593 ASSERT(db->db_dirtycnt > 0);
2594 db->db_dirtycnt -= 1;
2595 db->db_data_pending = NULL;
2596 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
2597 }
2598
2599 static void
2600 dbuf_write_nofill_ready(zio_t *zio)
2601 {
2602 dbuf_write_ready(zio, NULL, zio->io_private);
2603 }
2604
2605 static void
2606 dbuf_write_nofill_done(zio_t *zio)
2607 {
2608 dbuf_write_done(zio, NULL, zio->io_private);
2609 }
2610
2611 static void
2612 dbuf_write_override_ready(zio_t *zio)
2613 {
2614 dbuf_dirty_record_t *dr = zio->io_private;
2615 dmu_buf_impl_t *db = dr->dr_dbuf;
2616
2617 dbuf_write_ready(zio, NULL, db);
2618 }
2619
2620 static void
2621 dbuf_write_override_done(zio_t *zio)
2622 {
2623 dbuf_dirty_record_t *dr = zio->io_private;
2624 dmu_buf_impl_t *db = dr->dr_dbuf;
2625 blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
2626
2627 mutex_enter(&db->db_mtx);
2628 if (!BP_EQUAL(zio->io_bp, obp)) {
2629 if (!BP_IS_HOLE(obp))
2630 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
2631 arc_release(dr->dt.dl.dr_data, db);
2632 }
2633 mutex_exit(&db->db_mtx);
2634
2635 dbuf_write_done(zio, NULL, db);
2636 }
2637
2638 /* Issue I/O to commit a dirty buffer to disk. */
2639 static void
2640 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
2641 {
2642 dmu_buf_impl_t *db = dr->dr_dbuf;
2643 dnode_t *dn;
2644 objset_t *os;
2645 dmu_buf_impl_t *parent = db->db_parent;
2646 uint64_t txg = tx->tx_txg;
2647 zbookmark_t zb;
2648 zio_prop_t zp;
2649 zio_t *zio;
2650 int wp_flag = 0;
2651
2652 DB_DNODE_ENTER(db);
2653 dn = DB_DNODE(db);
2654 os = dn->dn_objset;
2655
2656 if (db->db_state != DB_NOFILL) {
2657 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
2658 /*
2659 * Private object buffers are released here rather
2660 * than in dbuf_dirty() since they are only modified
2661 * in the syncing context and we don't want the
2662 * overhead of making multiple copies of the data.
2663 */
2664 if (BP_IS_HOLE(db->db_blkptr)) {
2665 arc_buf_thaw(data);
2666 } else {
2667 dbuf_release_bp(db);
2668 }
2669 }
2670 }
2671
2672 if (parent != dn->dn_dbuf) {
2673 /* Our parent is an indirect block. */
2674 /* We have a dirty parent that has been scheduled for write. */
2675 ASSERT(parent && parent->db_data_pending);
2676 /* Our parent's buffer is one level closer to the dnode. */
2677 ASSERT(db->db_level == parent->db_level-1);
2678 /*
2679 * We're about to modify our parent's db_data by modifying
2680 * our block pointer, so the parent must be released.
2681 */
2682 ASSERT(arc_released(parent->db_buf));
2683 zio = parent->db_data_pending->dr_zio;
2684 } else {
2685 /* Our parent is the dnode itself. */
2686 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
2687 db->db_blkid != DMU_SPILL_BLKID) ||
2688 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
2689 if (db->db_blkid != DMU_SPILL_BLKID)
2690 ASSERT3P(db->db_blkptr, ==,
2691 &dn->dn_phys->dn_blkptr[db->db_blkid]);
2692 zio = dn->dn_zio;
2693 }
2694
2695 ASSERT(db->db_level == 0 || data == db->db_buf);
2696 ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
2697 ASSERT(zio);
2698
2699 SET_BOOKMARK(&zb, os->os_dsl_dataset ?
2700 os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
2701 db->db.db_object, db->db_level, db->db_blkid);
2702
2703 if (db->db_blkid == DMU_SPILL_BLKID)
2704 wp_flag = WP_SPILL;
2705 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
2706
2707 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
2708 DB_DNODE_EXIT(db);
2709
2710 if (db->db_level == 0 && dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
2711 ASSERT(db->db_state != DB_NOFILL);
2712 dr->dr_zio = zio_write(zio, os->os_spa, txg,
2713 db->db_blkptr, data->b_data, arc_buf_size(data), &zp,
2714 dbuf_write_override_ready, dbuf_write_override_done, dr,
2715 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
2716 mutex_enter(&db->db_mtx);
2717 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
2718 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
2719 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
2720 mutex_exit(&db->db_mtx);
2721 } else if (db->db_state == DB_NOFILL) {
2722 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF);
2723 dr->dr_zio = zio_write(zio, os->os_spa, txg,
2724 db->db_blkptr, NULL, db->db.db_size, &zp,
2725 dbuf_write_nofill_ready, dbuf_write_nofill_done, db,
2726 ZIO_PRIORITY_ASYNC_WRITE,
2727 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
2728 } else {
2729 ASSERT(arc_released(data));
2730 dr->dr_zio = arc_write(zio, os->os_spa, txg,
2731 db->db_blkptr, data, DBUF_IS_L2CACHEABLE(db), &zp,
2732 dbuf_write_ready, dbuf_write_done, db,
2733 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
2734 }
2735 }