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 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved.
26 * Copyright (c) 2013, Joyent, Inc. All rights reserved.
27 */
28
29 #include <sys/zfs_context.h>
30 #include <sys/dmu.h>
31 #include <sys/dmu_send.h>
32 #include <sys/dmu_impl.h>
33 #include <sys/dbuf.h>
34 #include <sys/dmu_objset.h>
35 #include <sys/dsl_dataset.h>
36 #include <sys/dsl_dir.h>
37 #include <sys/dmu_tx.h>
38 #include <sys/spa.h>
39 #include <sys/zio.h>
40 #include <sys/dmu_zfetch.h>
41 #include <sys/sa.h>
42 #include <sys/sa_impl.h>
43
44 /*
45 * Number of times that zfs_free_range() took the slow path while doing
46 * a zfs receive. A nonzero value indicates a potential performance problem.
47 */
48 uint64_t zfs_free_range_recv_miss;
49
50 static void dbuf_destroy(dmu_buf_impl_t *db);
51 static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
52 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
53
54 /*
55 * Global data structures and functions for the dbuf cache.
56 */
57 static kmem_cache_t *dbuf_cache;
58
59 /* ARGSUSED */
60 static int
61 dbuf_cons(void *vdb, void *unused, int kmflag)
62 {
63 dmu_buf_impl_t *db = vdb;
64 bzero(db, sizeof (dmu_buf_impl_t));
65
66 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
67 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
68 refcount_create(&db->db_holds);
69 return (0);
70 }
71
72 /* ARGSUSED */
73 static void
74 dbuf_dest(void *vdb, void *unused)
75 {
76 dmu_buf_impl_t *db = vdb;
77 mutex_destroy(&db->db_mtx);
78 cv_destroy(&db->db_changed);
79 refcount_destroy(&db->db_holds);
80 }
81
82 /*
83 * dbuf hash table routines
84 */
85 static dbuf_hash_table_t dbuf_hash_table;
86
87 static uint64_t dbuf_hash_count;
88
89 static uint64_t
90 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
91 {
92 uintptr_t osv = (uintptr_t)os;
93 uint64_t crc = -1ULL;
94
95 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
96 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF];
97 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF];
98 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF];
99 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF];
100 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF];
101 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF];
102
103 crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16);
104
105 return (crc);
106 }
107
108 #define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid);
109
110 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
111 ((dbuf)->db.db_object == (obj) && \
112 (dbuf)->db_objset == (os) && \
113 (dbuf)->db_level == (level) && \
114 (dbuf)->db_blkid == (blkid))
115
116 dmu_buf_impl_t *
117 dbuf_find(dnode_t *dn, uint8_t level, uint64_t blkid)
118 {
119 dbuf_hash_table_t *h = &dbuf_hash_table;
120 objset_t *os = dn->dn_objset;
121 uint64_t obj = dn->dn_object;
122 uint64_t hv = DBUF_HASH(os, obj, level, blkid);
123 uint64_t idx = hv & h->hash_table_mask;
124 dmu_buf_impl_t *db;
125
126 mutex_enter(DBUF_HASH_MUTEX(h, idx));
127 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
128 if (DBUF_EQUAL(db, os, obj, level, blkid)) {
129 mutex_enter(&db->db_mtx);
130 if (db->db_state != DB_EVICTING) {
131 mutex_exit(DBUF_HASH_MUTEX(h, idx));
132 return (db);
133 }
134 mutex_exit(&db->db_mtx);
135 }
136 }
137 mutex_exit(DBUF_HASH_MUTEX(h, idx));
138 return (NULL);
139 }
140
141 /*
142 * Insert an entry into the hash table. If there is already an element
143 * equal to elem in the hash table, then the already existing element
144 * will be returned and the new element will not be inserted.
145 * Otherwise returns NULL.
146 */
147 static dmu_buf_impl_t *
148 dbuf_hash_insert(dmu_buf_impl_t *db)
149 {
150 dbuf_hash_table_t *h = &dbuf_hash_table;
151 objset_t *os = db->db_objset;
152 uint64_t obj = db->db.db_object;
153 int level = db->db_level;
154 uint64_t blkid = db->db_blkid;
155 uint64_t hv = DBUF_HASH(os, obj, level, blkid);
156 uint64_t idx = hv & h->hash_table_mask;
157 dmu_buf_impl_t *dbf;
158
159 mutex_enter(DBUF_HASH_MUTEX(h, idx));
160 for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) {
161 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
162 mutex_enter(&dbf->db_mtx);
163 if (dbf->db_state != DB_EVICTING) {
164 mutex_exit(DBUF_HASH_MUTEX(h, idx));
165 return (dbf);
166 }
167 mutex_exit(&dbf->db_mtx);
168 }
169 }
170
171 mutex_enter(&db->db_mtx);
172 db->db_hash_next = h->hash_table[idx];
173 h->hash_table[idx] = db;
174 mutex_exit(DBUF_HASH_MUTEX(h, idx));
175 atomic_add_64(&dbuf_hash_count, 1);
176
177 return (NULL);
178 }
179
180 /*
181 * Remove an entry from the hash table. This operation will
182 * fail if there are any existing holds on the db.
183 */
184 static void
185 dbuf_hash_remove(dmu_buf_impl_t *db)
186 {
187 dbuf_hash_table_t *h = &dbuf_hash_table;
188 uint64_t hv = DBUF_HASH(db->db_objset, db->db.db_object,
189 db->db_level, db->db_blkid);
190 uint64_t idx = hv & h->hash_table_mask;
191 dmu_buf_impl_t *dbf, **dbp;
192
193 /*
194 * We musn't hold db_mtx to maintin lock ordering:
195 * DBUF_HASH_MUTEX > db_mtx.
196 */
197 ASSERT(refcount_is_zero(&db->db_holds));
198 ASSERT(db->db_state == DB_EVICTING);
199 ASSERT(!MUTEX_HELD(&db->db_mtx));
200
201 mutex_enter(DBUF_HASH_MUTEX(h, idx));
202 dbp = &h->hash_table[idx];
203 while ((dbf = *dbp) != db) {
204 dbp = &dbf->db_hash_next;
205 ASSERT(dbf != NULL);
206 }
207 *dbp = db->db_hash_next;
208 db->db_hash_next = NULL;
209 mutex_exit(DBUF_HASH_MUTEX(h, idx));
210 atomic_add_64(&dbuf_hash_count, -1);
211 }
212
213 static arc_evict_func_t dbuf_do_evict;
214
215 static void
216 dbuf_evict_user(dmu_buf_impl_t *db)
217 {
218 ASSERT(MUTEX_HELD(&db->db_mtx));
219
220 if (db->db_level != 0 || db->db_evict_func == NULL)
221 return;
222
223 if (db->db_user_data_ptr_ptr)
224 *db->db_user_data_ptr_ptr = db->db.db_data;
225 db->db_evict_func(&db->db, db->db_user_ptr);
226 db->db_user_ptr = NULL;
227 db->db_user_data_ptr_ptr = NULL;
228 db->db_evict_func = NULL;
229 }
230
231 boolean_t
232 dbuf_is_metadata(dmu_buf_impl_t *db)
233 {
234 if (db->db_level > 0) {
235 return (B_TRUE);
236 } else {
237 boolean_t is_metadata;
238
239 DB_DNODE_ENTER(db);
240 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type);
241 DB_DNODE_EXIT(db);
242
243 return (is_metadata);
244 }
245 }
246
247 void
248 dbuf_evict(dmu_buf_impl_t *db)
249 {
250 ASSERT(MUTEX_HELD(&db->db_mtx));
251 ASSERT(db->db_buf == NULL);
252 ASSERT(db->db_data_pending == NULL);
253
254 dbuf_clear(db);
255 dbuf_destroy(db);
256 }
257
258 void
259 dbuf_init(void)
260 {
261 uint64_t hsize = 1ULL << 16;
262 dbuf_hash_table_t *h = &dbuf_hash_table;
263 int i;
264
265 /*
266 * The hash table is big enough to fill all of physical memory
267 * with an average 4K block size. The table will take up
268 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
269 */
270 while (hsize * 4096 < physmem * PAGESIZE)
271 hsize <<= 1;
272
273 retry:
274 h->hash_table_mask = hsize - 1;
275 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
276 if (h->hash_table == NULL) {
277 /* XXX - we should really return an error instead of assert */
278 ASSERT(hsize > (1ULL << 10));
279 hsize >>= 1;
280 goto retry;
281 }
282
283 dbuf_cache = kmem_cache_create("dmu_buf_impl_t",
284 sizeof (dmu_buf_impl_t),
285 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
286
287 for (i = 0; i < DBUF_MUTEXES; i++)
288 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
289 }
290
291 void
292 dbuf_fini(void)
293 {
294 dbuf_hash_table_t *h = &dbuf_hash_table;
295 int i;
296
297 for (i = 0; i < DBUF_MUTEXES; i++)
298 mutex_destroy(&h->hash_mutexes[i]);
299 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
300 kmem_cache_destroy(dbuf_cache);
301 }
302
303 /*
304 * Other stuff.
305 */
306
307 #ifdef ZFS_DEBUG
308 static void
309 dbuf_verify(dmu_buf_impl_t *db)
310 {
311 dnode_t *dn;
312 dbuf_dirty_record_t *dr;
313
314 ASSERT(MUTEX_HELD(&db->db_mtx));
315
316 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
317 return;
318
319 ASSERT(db->db_objset != NULL);
320 DB_DNODE_ENTER(db);
321 dn = DB_DNODE(db);
322 if (dn == NULL) {
323 ASSERT(db->db_parent == NULL);
324 ASSERT(db->db_blkptr == NULL);
325 } else {
326 ASSERT3U(db->db.db_object, ==, dn->dn_object);
327 ASSERT3P(db->db_objset, ==, dn->dn_objset);
328 ASSERT3U(db->db_level, <, dn->dn_nlevels);
329 ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
330 db->db_blkid == DMU_SPILL_BLKID ||
331 !list_is_empty(&dn->dn_dbufs));
332 }
333 if (db->db_blkid == DMU_BONUS_BLKID) {
334 ASSERT(dn != NULL);
335 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
336 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
337 } else if (db->db_blkid == DMU_SPILL_BLKID) {
338 ASSERT(dn != NULL);
339 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
340 ASSERT0(db->db.db_offset);
341 } else {
342 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
343 }
344
345 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
346 ASSERT(dr->dr_dbuf == db);
347
348 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
349 ASSERT(dr->dr_dbuf == db);
350
351 /*
352 * We can't assert that db_size matches dn_datablksz because it
353 * can be momentarily different when another thread is doing
354 * dnode_set_blksz().
355 */
356 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
357 dr = db->db_data_pending;
358 /*
359 * It should only be modified in syncing context, so
360 * make sure we only have one copy of the data.
361 */
362 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
363 }
364
365 /* verify db->db_blkptr */
366 if (db->db_blkptr) {
367 if (db->db_parent == dn->dn_dbuf) {
368 /* db is pointed to by the dnode */
369 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
370 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
371 ASSERT(db->db_parent == NULL);
372 else
373 ASSERT(db->db_parent != NULL);
374 if (db->db_blkid != DMU_SPILL_BLKID)
375 ASSERT3P(db->db_blkptr, ==,
376 &dn->dn_phys->dn_blkptr[db->db_blkid]);
377 } else {
378 /* db is pointed to by an indirect block */
379 int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
380 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
381 ASSERT3U(db->db_parent->db.db_object, ==,
382 db->db.db_object);
383 /*
384 * dnode_grow_indblksz() can make this fail if we don't
385 * have the struct_rwlock. XXX indblksz no longer
386 * grows. safe to do this now?
387 */
388 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
389 ASSERT3P(db->db_blkptr, ==,
390 ((blkptr_t *)db->db_parent->db.db_data +
391 db->db_blkid % epb));
392 }
393 }
394 }
395 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
396 (db->db_buf == NULL || db->db_buf->b_data) &&
397 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
398 db->db_state != DB_FILL && !dn->dn_free_txg) {
399 /*
400 * If the blkptr isn't set but they have nonzero data,
401 * it had better be dirty, otherwise we'll lose that
402 * data when we evict this buffer.
403 */
404 if (db->db_dirtycnt == 0) {
405 uint64_t *buf = db->db.db_data;
406 int i;
407
408 for (i = 0; i < db->db.db_size >> 3; i++) {
409 ASSERT(buf[i] == 0);
410 }
411 }
412 }
413 DB_DNODE_EXIT(db);
414 }
415 #endif
416
417 static void
418 dbuf_update_data(dmu_buf_impl_t *db)
419 {
420 ASSERT(MUTEX_HELD(&db->db_mtx));
421 if (db->db_level == 0 && db->db_user_data_ptr_ptr) {
422 ASSERT(!refcount_is_zero(&db->db_holds));
423 *db->db_user_data_ptr_ptr = db->db.db_data;
424 }
425 }
426
427 static void
428 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
429 {
430 ASSERT(MUTEX_HELD(&db->db_mtx));
431 ASSERT(db->db_buf == NULL || !arc_has_callback(db->db_buf));
432 db->db_buf = buf;
433 if (buf != NULL) {
434 ASSERT(buf->b_data != NULL);
435 db->db.db_data = buf->b_data;
436 if (!arc_released(buf))
437 arc_set_callback(buf, dbuf_do_evict, db);
438 dbuf_update_data(db);
439 } else {
440 dbuf_evict_user(db);
441 db->db.db_data = NULL;
442 if (db->db_state != DB_NOFILL)
443 db->db_state = DB_UNCACHED;
444 }
445 }
446
447 /*
448 * Loan out an arc_buf for read. Return the loaned arc_buf.
449 */
450 arc_buf_t *
451 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
452 {
453 arc_buf_t *abuf;
454
455 mutex_enter(&db->db_mtx);
456 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
457 int blksz = db->db.db_size;
458 spa_t *spa;
459
460 mutex_exit(&db->db_mtx);
461 DB_GET_SPA(&spa, db);
462 abuf = arc_loan_buf(spa, blksz);
463 bcopy(db->db.db_data, abuf->b_data, blksz);
464 } else {
465 abuf = db->db_buf;
466 arc_loan_inuse_buf(abuf, db);
467 dbuf_set_data(db, NULL);
468 mutex_exit(&db->db_mtx);
469 }
470 return (abuf);
471 }
472
473 uint64_t
474 dbuf_whichblock(dnode_t *dn, uint64_t offset)
475 {
476 if (dn->dn_datablkshift) {
477 return (offset >> dn->dn_datablkshift);
478 } else {
479 ASSERT3U(offset, <, dn->dn_datablksz);
480 return (0);
481 }
482 }
483
484 static void
485 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
486 {
487 dmu_buf_impl_t *db = vdb;
488
489 mutex_enter(&db->db_mtx);
490 ASSERT3U(db->db_state, ==, DB_READ);
491 /*
492 * All reads are synchronous, so we must have a hold on the dbuf
493 */
494 ASSERT(refcount_count(&db->db_holds) > 0);
495 ASSERT(db->db_buf == NULL);
496 ASSERT(db->db.db_data == NULL);
497 if (db->db_level == 0 && db->db_freed_in_flight) {
498 /* we were freed in flight; disregard any error */
499 arc_release(buf, db);
500 bzero(buf->b_data, db->db.db_size);
501 arc_buf_freeze(buf);
502 db->db_freed_in_flight = FALSE;
503 dbuf_set_data(db, buf);
504 db->db_state = DB_CACHED;
505 } else if (zio == NULL || zio->io_error == 0) {
506 dbuf_set_data(db, buf);
507 db->db_state = DB_CACHED;
508 } else {
509 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
510 ASSERT3P(db->db_buf, ==, NULL);
511 VERIFY(arc_buf_remove_ref(buf, db));
512 db->db_state = DB_UNCACHED;
513 }
514 cv_broadcast(&db->db_changed);
515 dbuf_rele_and_unlock(db, NULL);
516 }
517
518 static void
519 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t *flags)
520 {
521 dnode_t *dn;
522 spa_t *spa;
523 zbookmark_t zb;
524 uint32_t aflags = ARC_NOWAIT;
525
526 DB_DNODE_ENTER(db);
527 dn = DB_DNODE(db);
528 ASSERT(!refcount_is_zero(&db->db_holds));
529 /* We need the struct_rwlock to prevent db_blkptr from changing. */
530 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
531 ASSERT(MUTEX_HELD(&db->db_mtx));
532 ASSERT(db->db_state == DB_UNCACHED);
533 ASSERT(db->db_buf == NULL);
534
535 if (db->db_blkid == DMU_BONUS_BLKID) {
536 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
537
538 ASSERT3U(bonuslen, <=, db->db.db_size);
539 db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN);
540 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
541 if (bonuslen < DN_MAX_BONUSLEN)
542 bzero(db->db.db_data, DN_MAX_BONUSLEN);
543 if (bonuslen)
544 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
545 DB_DNODE_EXIT(db);
546 dbuf_update_data(db);
547 db->db_state = DB_CACHED;
548 mutex_exit(&db->db_mtx);
549 return;
550 }
551
552 /*
553 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
554 * processes the delete record and clears the bp while we are waiting
555 * for the dn_mtx (resulting in a "no" from block_freed).
556 */
557 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
558 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
559 BP_IS_HOLE(db->db_blkptr)))) {
560 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
561
562 dbuf_set_data(db, arc_buf_alloc(dn->dn_objset->os_spa,
563 db->db.db_size, db, type));
564 DB_DNODE_EXIT(db);
565 bzero(db->db.db_data, db->db.db_size);
566 db->db_state = DB_CACHED;
567 *flags |= DB_RF_CACHED;
568 mutex_exit(&db->db_mtx);
569 return;
570 }
571
572 spa = dn->dn_objset->os_spa;
573 DB_DNODE_EXIT(db);
574
575 db->db_state = DB_READ;
576 mutex_exit(&db->db_mtx);
577
578 if (DBUF_IS_L2CACHEABLE(db))
579 aflags |= ARC_L2CACHE;
580 if (DBUF_IS_L2COMPRESSIBLE(db))
581 aflags |= ARC_L2COMPRESS;
582
583 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
584 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
585 db->db.db_object, db->db_level, db->db_blkid);
586
587 dbuf_add_ref(db, NULL);
588
589 (void) arc_read(zio, spa, db->db_blkptr,
590 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
591 (*flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
592 &aflags, &zb);
593 if (aflags & ARC_CACHED)
594 *flags |= DB_RF_CACHED;
595 }
596
597 int
598 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
599 {
600 int err = 0;
601 int havepzio = (zio != NULL);
602 int prefetch;
603 dnode_t *dn;
604
605 /*
606 * We don't have to hold the mutex to check db_state because it
607 * can't be freed while we have a hold on the buffer.
608 */
609 ASSERT(!refcount_is_zero(&db->db_holds));
610
611 if (db->db_state == DB_NOFILL)
612 return (SET_ERROR(EIO));
613
614 DB_DNODE_ENTER(db);
615 dn = DB_DNODE(db);
616 if ((flags & DB_RF_HAVESTRUCT) == 0)
617 rw_enter(&dn->dn_struct_rwlock, RW_READER);
618
619 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
620 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
621 DBUF_IS_CACHEABLE(db);
622
623 mutex_enter(&db->db_mtx);
624 if (db->db_state == DB_CACHED) {
625 mutex_exit(&db->db_mtx);
626 if (prefetch)
627 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
628 db->db.db_size, TRUE);
629 if ((flags & DB_RF_HAVESTRUCT) == 0)
630 rw_exit(&dn->dn_struct_rwlock);
631 DB_DNODE_EXIT(db);
632 } else if (db->db_state == DB_UNCACHED) {
633 spa_t *spa = dn->dn_objset->os_spa;
634
635 if (zio == NULL)
636 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
637 dbuf_read_impl(db, zio, &flags);
638
639 /* dbuf_read_impl has dropped db_mtx for us */
640
641 if (prefetch)
642 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
643 db->db.db_size, flags & DB_RF_CACHED);
644
645 if ((flags & DB_RF_HAVESTRUCT) == 0)
646 rw_exit(&dn->dn_struct_rwlock);
647 DB_DNODE_EXIT(db);
648
649 if (!havepzio)
650 err = zio_wait(zio);
651 } else {
652 /*
653 * Another reader came in while the dbuf was in flight
654 * between UNCACHED and CACHED. Either a writer will finish
655 * writing the buffer (sending the dbuf to CACHED) or the
656 * first reader's request will reach the read_done callback
657 * and send the dbuf to CACHED. Otherwise, a failure
658 * occurred and the dbuf went to UNCACHED.
659 */
660 mutex_exit(&db->db_mtx);
661 if (prefetch)
662 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
663 db->db.db_size, TRUE);
664 if ((flags & DB_RF_HAVESTRUCT) == 0)
665 rw_exit(&dn->dn_struct_rwlock);
666 DB_DNODE_EXIT(db);
667
668 /* Skip the wait per the caller's request. */
669 mutex_enter(&db->db_mtx);
670 if ((flags & DB_RF_NEVERWAIT) == 0) {
671 while (db->db_state == DB_READ ||
672 db->db_state == DB_FILL) {
673 ASSERT(db->db_state == DB_READ ||
674 (flags & DB_RF_HAVESTRUCT) == 0);
675 cv_wait(&db->db_changed, &db->db_mtx);
676 }
677 if (db->db_state == DB_UNCACHED)
678 err = SET_ERROR(EIO);
679 }
680 mutex_exit(&db->db_mtx);
681 }
682
683 ASSERT(err || havepzio || db->db_state == DB_CACHED);
684 return (err);
685 }
686
687 static void
688 dbuf_noread(dmu_buf_impl_t *db)
689 {
690 ASSERT(!refcount_is_zero(&db->db_holds));
691 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
692 mutex_enter(&db->db_mtx);
693 while (db->db_state == DB_READ || db->db_state == DB_FILL)
694 cv_wait(&db->db_changed, &db->db_mtx);
695 if (db->db_state == DB_UNCACHED) {
696 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
697 spa_t *spa;
698
699 ASSERT(db->db_buf == NULL);
700 ASSERT(db->db.db_data == NULL);
701 DB_GET_SPA(&spa, db);
702 dbuf_set_data(db, arc_buf_alloc(spa, db->db.db_size, db, type));
703 db->db_state = DB_FILL;
704 } else if (db->db_state == DB_NOFILL) {
705 dbuf_set_data(db, NULL);
706 } else {
707 ASSERT3U(db->db_state, ==, DB_CACHED);
708 }
709 mutex_exit(&db->db_mtx);
710 }
711
712 /*
713 * This is our just-in-time copy function. It makes a copy of
714 * buffers, that have been modified in a previous transaction
715 * group, before we modify them in the current active group.
716 *
717 * This function is used in two places: when we are dirtying a
718 * buffer for the first time in a txg, and when we are freeing
719 * a range in a dnode that includes this buffer.
720 *
721 * Note that when we are called from dbuf_free_range() we do
722 * not put a hold on the buffer, we just traverse the active
723 * dbuf list for the dnode.
724 */
725 static void
726 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
727 {
728 dbuf_dirty_record_t *dr = db->db_last_dirty;
729
730 ASSERT(MUTEX_HELD(&db->db_mtx));
731 ASSERT(db->db.db_data != NULL);
732 ASSERT(db->db_level == 0);
733 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
734
735 if (dr == NULL ||
736 (dr->dt.dl.dr_data !=
737 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
738 return;
739
740 /*
741 * If the last dirty record for this dbuf has not yet synced
742 * and its referencing the dbuf data, either:
743 * reset the reference to point to a new copy,
744 * or (if there a no active holders)
745 * just null out the current db_data pointer.
746 */
747 ASSERT(dr->dr_txg >= txg - 2);
748 if (db->db_blkid == DMU_BONUS_BLKID) {
749 /* Note that the data bufs here are zio_bufs */
750 dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
751 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
752 bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
753 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
754 int size = db->db.db_size;
755 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
756 spa_t *spa;
757
758 DB_GET_SPA(&spa, db);
759 dr->dt.dl.dr_data = arc_buf_alloc(spa, size, db, type);
760 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
761 } else {
762 dbuf_set_data(db, NULL);
763 }
764 }
765
766 void
767 dbuf_unoverride(dbuf_dirty_record_t *dr)
768 {
769 dmu_buf_impl_t *db = dr->dr_dbuf;
770 blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
771 uint64_t txg = dr->dr_txg;
772
773 ASSERT(MUTEX_HELD(&db->db_mtx));
774 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
775 ASSERT(db->db_level == 0);
776
777 if (db->db_blkid == DMU_BONUS_BLKID ||
778 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
779 return;
780
781 ASSERT(db->db_data_pending != dr);
782
783 /* free this block */
784 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite) {
785 spa_t *spa;
786
787 DB_GET_SPA(&spa, db);
788 zio_free(spa, txg, bp);
789 }
790 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
791 dr->dt.dl.dr_nopwrite = B_FALSE;
792
793 /*
794 * Release the already-written buffer, so we leave it in
795 * a consistent dirty state. Note that all callers are
796 * modifying the buffer, so they will immediately do
797 * another (redundant) arc_release(). Therefore, leave
798 * the buf thawed to save the effort of freezing &
799 * immediately re-thawing it.
800 */
801 arc_release(dr->dt.dl.dr_data, db);
802 }
803
804 /*
805 * Evict (if its unreferenced) or clear (if its referenced) any level-0
806 * data blocks in the free range, so that any future readers will find
807 * empty blocks. Also, if we happen across any level-1 dbufs in the
808 * range that have not already been marked dirty, mark them dirty so
809 * they stay in memory.
810 *
811 * This is a no-op if the dataset is in the middle of an incremental
812 * receive; see comment below for details.
813 */
814 void
815 dbuf_free_range(dnode_t *dn, uint64_t start, uint64_t end, dmu_tx_t *tx)
816 {
817 dmu_buf_impl_t *db, *db_next;
818 uint64_t txg = tx->tx_txg;
819 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
820 uint64_t first_l1 = start >> epbs;
821 uint64_t last_l1 = end >> epbs;
822
823 if (end > dn->dn_maxblkid && (end != DMU_SPILL_BLKID)) {
824 end = dn->dn_maxblkid;
825 last_l1 = end >> epbs;
826 }
827 dprintf_dnode(dn, "start=%llu end=%llu\n", start, end);
828
829 mutex_enter(&dn->dn_dbufs_mtx);
830 if (start >= dn->dn_unlisted_l0_blkid * dn->dn_datablksz) {
831 /* There can't be any dbufs in this range; no need to search. */
832 mutex_exit(&dn->dn_dbufs_mtx);
833 return;
834 } else if (dmu_objset_is_receiving(dn->dn_objset)) {
835 /*
836 * If we are receiving, we expect there to be no dbufs in
837 * the range to be freed, because receive modifies each
838 * block at most once, and in offset order. If this is
839 * not the case, it can lead to performance problems,
840 * so note that we unexpectedly took the slow path.
841 */
842 atomic_inc_64(&zfs_free_range_recv_miss);
843 }
844
845 for (db = list_head(&dn->dn_dbufs); db != NULL; db = db_next) {
846 db_next = list_next(&dn->dn_dbufs, db);
847 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
848
849 if (db->db_level == 1 &&
850 db->db_blkid >= first_l1 && db->db_blkid <= last_l1) {
851 mutex_enter(&db->db_mtx);
852 if (db->db_last_dirty &&
853 db->db_last_dirty->dr_txg < txg) {
854 dbuf_add_ref(db, FTAG);
855 mutex_exit(&db->db_mtx);
856 dbuf_will_dirty(db, tx);
857 dbuf_rele(db, FTAG);
858 } else {
859 mutex_exit(&db->db_mtx);
860 }
861 }
862
863 if (db->db_level != 0)
864 continue;
865 dprintf_dbuf(db, "found buf %s\n", "");
866 if (db->db_blkid < start || db->db_blkid > end)
867 continue;
868
869 /* found a level 0 buffer in the range */
870 mutex_enter(&db->db_mtx);
871 if (dbuf_undirty(db, tx)) {
872 /* mutex has been dropped and dbuf destroyed */
873 continue;
874 }
875
876 if (db->db_state == DB_UNCACHED ||
877 db->db_state == DB_NOFILL ||
878 db->db_state == DB_EVICTING) {
879 ASSERT(db->db.db_data == NULL);
880 mutex_exit(&db->db_mtx);
881 continue;
882 }
883 if (db->db_state == DB_READ || db->db_state == DB_FILL) {
884 /* will be handled in dbuf_read_done or dbuf_rele */
885 db->db_freed_in_flight = TRUE;
886 mutex_exit(&db->db_mtx);
887 continue;
888 }
889 if (refcount_count(&db->db_holds) == 0) {
890 ASSERT(db->db_buf);
891 dbuf_clear(db);
892 continue;
893 }
894 /* The dbuf is referenced */
895
896 if (db->db_last_dirty != NULL) {
897 dbuf_dirty_record_t *dr = db->db_last_dirty;
898
899 if (dr->dr_txg == txg) {
900 /*
901 * This buffer is "in-use", re-adjust the file
902 * size to reflect that this buffer may
903 * contain new data when we sync.
904 */
905 if (db->db_blkid != DMU_SPILL_BLKID &&
906 db->db_blkid > dn->dn_maxblkid)
907 dn->dn_maxblkid = db->db_blkid;
908 dbuf_unoverride(dr);
909 } else {
910 /*
911 * This dbuf is not dirty in the open context.
912 * Either uncache it (if its not referenced in
913 * the open context) or reset its contents to
914 * empty.
915 */
916 dbuf_fix_old_data(db, txg);
917 }
918 }
919 /* clear the contents if its cached */
920 if (db->db_state == DB_CACHED) {
921 ASSERT(db->db.db_data != NULL);
922 arc_release(db->db_buf, db);
923 bzero(db->db.db_data, db->db.db_size);
924 arc_buf_freeze(db->db_buf);
925 }
926
927 mutex_exit(&db->db_mtx);
928 }
929 mutex_exit(&dn->dn_dbufs_mtx);
930 }
931
932 static int
933 dbuf_block_freeable(dmu_buf_impl_t *db)
934 {
935 dsl_dataset_t *ds = db->db_objset->os_dsl_dataset;
936 uint64_t birth_txg = 0;
937
938 /*
939 * We don't need any locking to protect db_blkptr:
940 * If it's syncing, then db_last_dirty will be set
941 * so we'll ignore db_blkptr.
942 */
943 ASSERT(MUTEX_HELD(&db->db_mtx));
944 if (db->db_last_dirty)
945 birth_txg = db->db_last_dirty->dr_txg;
946 else if (db->db_blkptr)
947 birth_txg = db->db_blkptr->blk_birth;
948
949 /*
950 * If we don't exist or are in a snapshot, we can't be freed.
951 * Don't pass the bp to dsl_dataset_block_freeable() since we
952 * are holding the db_mtx lock and might deadlock if we are
953 * prefetching a dedup-ed block.
954 */
955 if (birth_txg)
956 return (ds == NULL ||
957 dsl_dataset_block_freeable(ds, NULL, birth_txg));
958 else
959 return (FALSE);
960 }
961
962 void
963 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
964 {
965 arc_buf_t *buf, *obuf;
966 int osize = db->db.db_size;
967 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
968 dnode_t *dn;
969
970 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
971
972 DB_DNODE_ENTER(db);
973 dn = DB_DNODE(db);
974
975 /* XXX does *this* func really need the lock? */
976 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
977
978 /*
979 * This call to dbuf_will_dirty() with the dn_struct_rwlock held
980 * is OK, because there can be no other references to the db
981 * when we are changing its size, so no concurrent DB_FILL can
982 * be happening.
983 */
984 /*
985 * XXX we should be doing a dbuf_read, checking the return
986 * value and returning that up to our callers
987 */
988 dbuf_will_dirty(db, tx);
989
990 /* create the data buffer for the new block */
991 buf = arc_buf_alloc(dn->dn_objset->os_spa, size, db, type);
992
993 /* copy old block data to the new block */
994 obuf = db->db_buf;
995 bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
996 /* zero the remainder */
997 if (size > osize)
998 bzero((uint8_t *)buf->b_data + osize, size - osize);
999
1000 mutex_enter(&db->db_mtx);
1001 dbuf_set_data(db, buf);
1002 VERIFY(arc_buf_remove_ref(obuf, db));
1003 db->db.db_size = size;
1004
1005 if (db->db_level == 0) {
1006 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1007 db->db_last_dirty->dt.dl.dr_data = buf;
1008 }
1009 mutex_exit(&db->db_mtx);
1010
1011 dnode_willuse_space(dn, size-osize, tx);
1012 DB_DNODE_EXIT(db);
1013 }
1014
1015 void
1016 dbuf_release_bp(dmu_buf_impl_t *db)
1017 {
1018 objset_t *os;
1019
1020 DB_GET_OBJSET(&os, db);
1021 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1022 ASSERT(arc_released(os->os_phys_buf) ||
1023 list_link_active(&os->os_dsl_dataset->ds_synced_link));
1024 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1025
1026 (void) arc_release(db->db_buf, db);
1027 }
1028
1029 dbuf_dirty_record_t *
1030 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1031 {
1032 dnode_t *dn;
1033 objset_t *os;
1034 dbuf_dirty_record_t **drp, *dr;
1035 int drop_struct_lock = FALSE;
1036 boolean_t do_free_accounting = B_FALSE;
1037 int txgoff = tx->tx_txg & TXG_MASK;
1038
1039 ASSERT(tx->tx_txg != 0);
1040 ASSERT(!refcount_is_zero(&db->db_holds));
1041 DMU_TX_DIRTY_BUF(tx, db);
1042
1043 DB_DNODE_ENTER(db);
1044 dn = DB_DNODE(db);
1045 /*
1046 * Shouldn't dirty a regular buffer in syncing context. Private
1047 * objects may be dirtied in syncing context, but only if they
1048 * were already pre-dirtied in open context.
1049 */
1050 ASSERT(!dmu_tx_is_syncing(tx) ||
1051 BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1052 DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1053 dn->dn_objset->os_dsl_dataset == NULL);
1054 /*
1055 * We make this assert for private objects as well, but after we
1056 * check if we're already dirty. They are allowed to re-dirty
1057 * in syncing context.
1058 */
1059 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1060 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1061 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1062
1063 mutex_enter(&db->db_mtx);
1064 /*
1065 * XXX make this true for indirects too? The problem is that
1066 * transactions created with dmu_tx_create_assigned() from
1067 * syncing context don't bother holding ahead.
1068 */
1069 ASSERT(db->db_level != 0 ||
1070 db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1071 db->db_state == DB_NOFILL);
1072
1073 mutex_enter(&dn->dn_mtx);
1074 /*
1075 * Don't set dirtyctx to SYNC if we're just modifying this as we
1076 * initialize the objset.
1077 */
1078 if (dn->dn_dirtyctx == DN_UNDIRTIED &&
1079 !BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1080 dn->dn_dirtyctx =
1081 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1082 ASSERT(dn->dn_dirtyctx_firstset == NULL);
1083 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1084 }
1085 mutex_exit(&dn->dn_mtx);
1086
1087 if (db->db_blkid == DMU_SPILL_BLKID)
1088 dn->dn_have_spill = B_TRUE;
1089
1090 /*
1091 * If this buffer is already dirty, we're done.
1092 */
1093 drp = &db->db_last_dirty;
1094 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1095 db->db.db_object == DMU_META_DNODE_OBJECT);
1096 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1097 drp = &dr->dr_next;
1098 if (dr && dr->dr_txg == tx->tx_txg) {
1099 DB_DNODE_EXIT(db);
1100
1101 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1102 /*
1103 * If this buffer has already been written out,
1104 * we now need to reset its state.
1105 */
1106 dbuf_unoverride(dr);
1107 if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1108 db->db_state != DB_NOFILL)
1109 arc_buf_thaw(db->db_buf);
1110 }
1111 mutex_exit(&db->db_mtx);
1112 return (dr);
1113 }
1114
1115 /*
1116 * Only valid if not already dirty.
1117 */
1118 ASSERT(dn->dn_object == 0 ||
1119 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1120 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1121
1122 ASSERT3U(dn->dn_nlevels, >, db->db_level);
1123 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1124 dn->dn_phys->dn_nlevels > db->db_level ||
1125 dn->dn_next_nlevels[txgoff] > db->db_level ||
1126 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1127 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1128
1129 /*
1130 * We should only be dirtying in syncing context if it's the
1131 * mos or we're initializing the os or it's a special object.
1132 * However, we are allowed to dirty in syncing context provided
1133 * we already dirtied it in open context. Hence we must make
1134 * this assertion only if we're not already dirty.
1135 */
1136 os = dn->dn_objset;
1137 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1138 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1139 ASSERT(db->db.db_size != 0);
1140
1141 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1142
1143 if (db->db_blkid != DMU_BONUS_BLKID) {
1144 /*
1145 * Update the accounting.
1146 * Note: we delay "free accounting" until after we drop
1147 * the db_mtx. This keeps us from grabbing other locks
1148 * (and possibly deadlocking) in bp_get_dsize() while
1149 * also holding the db_mtx.
1150 */
1151 dnode_willuse_space(dn, db->db.db_size, tx);
1152 do_free_accounting = dbuf_block_freeable(db);
1153 }
1154
1155 /*
1156 * If this buffer is dirty in an old transaction group we need
1157 * to make a copy of it so that the changes we make in this
1158 * transaction group won't leak out when we sync the older txg.
1159 */
1160 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1161 if (db->db_level == 0) {
1162 void *data_old = db->db_buf;
1163
1164 if (db->db_state != DB_NOFILL) {
1165 if (db->db_blkid == DMU_BONUS_BLKID) {
1166 dbuf_fix_old_data(db, tx->tx_txg);
1167 data_old = db->db.db_data;
1168 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1169 /*
1170 * Release the data buffer from the cache so
1171 * that we can modify it without impacting
1172 * possible other users of this cached data
1173 * block. Note that indirect blocks and
1174 * private objects are not released until the
1175 * syncing state (since they are only modified
1176 * then).
1177 */
1178 arc_release(db->db_buf, db);
1179 dbuf_fix_old_data(db, tx->tx_txg);
1180 data_old = db->db_buf;
1181 }
1182 ASSERT(data_old != NULL);
1183 }
1184 dr->dt.dl.dr_data = data_old;
1185 } else {
1186 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1187 list_create(&dr->dt.di.dr_children,
1188 sizeof (dbuf_dirty_record_t),
1189 offsetof(dbuf_dirty_record_t, dr_dirty_node));
1190 }
1191 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL)
1192 dr->dr_accounted = db->db.db_size;
1193 dr->dr_dbuf = db;
1194 dr->dr_txg = tx->tx_txg;
1195 dr->dr_next = *drp;
1196 *drp = dr;
1197
1198 /*
1199 * We could have been freed_in_flight between the dbuf_noread
1200 * and dbuf_dirty. We win, as though the dbuf_noread() had
1201 * happened after the free.
1202 */
1203 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1204 db->db_blkid != DMU_SPILL_BLKID) {
1205 mutex_enter(&dn->dn_mtx);
1206 dnode_clear_range(dn, db->db_blkid, 1, tx);
1207 mutex_exit(&dn->dn_mtx);
1208 db->db_freed_in_flight = FALSE;
1209 }
1210
1211 /*
1212 * This buffer is now part of this txg
1213 */
1214 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1215 db->db_dirtycnt += 1;
1216 ASSERT3U(db->db_dirtycnt, <=, 3);
1217
1218 mutex_exit(&db->db_mtx);
1219
1220 if (db->db_blkid == DMU_BONUS_BLKID ||
1221 db->db_blkid == DMU_SPILL_BLKID) {
1222 mutex_enter(&dn->dn_mtx);
1223 ASSERT(!list_link_active(&dr->dr_dirty_node));
1224 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1225 mutex_exit(&dn->dn_mtx);
1226 dnode_setdirty(dn, tx);
1227 DB_DNODE_EXIT(db);
1228 return (dr);
1229 } else if (do_free_accounting) {
1230 blkptr_t *bp = db->db_blkptr;
1231 int64_t willfree = (bp && !BP_IS_HOLE(bp)) ?
1232 bp_get_dsize(os->os_spa, bp) : db->db.db_size;
1233 /*
1234 * This is only a guess -- if the dbuf is dirty
1235 * in a previous txg, we don't know how much
1236 * space it will use on disk yet. We should
1237 * really have the struct_rwlock to access
1238 * db_blkptr, but since this is just a guess,
1239 * it's OK if we get an odd answer.
1240 */
1241 ddt_prefetch(os->os_spa, bp);
1242 dnode_willuse_space(dn, -willfree, tx);
1243 }
1244
1245 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1246 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1247 drop_struct_lock = TRUE;
1248 }
1249
1250 if (db->db_level == 0) {
1251 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1252 ASSERT(dn->dn_maxblkid >= db->db_blkid);
1253 }
1254
1255 if (db->db_level+1 < dn->dn_nlevels) {
1256 dmu_buf_impl_t *parent = db->db_parent;
1257 dbuf_dirty_record_t *di;
1258 int parent_held = FALSE;
1259
1260 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1261 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1262
1263 parent = dbuf_hold_level(dn, db->db_level+1,
1264 db->db_blkid >> epbs, FTAG);
1265 ASSERT(parent != NULL);
1266 parent_held = TRUE;
1267 }
1268 if (drop_struct_lock)
1269 rw_exit(&dn->dn_struct_rwlock);
1270 ASSERT3U(db->db_level+1, ==, parent->db_level);
1271 di = dbuf_dirty(parent, tx);
1272 if (parent_held)
1273 dbuf_rele(parent, FTAG);
1274
1275 mutex_enter(&db->db_mtx);
1276 /*
1277 * Since we've dropped the mutex, it's possible that
1278 * dbuf_undirty() might have changed this out from under us.
1279 */
1280 if (db->db_last_dirty == dr ||
1281 dn->dn_object == DMU_META_DNODE_OBJECT) {
1282 mutex_enter(&di->dt.di.dr_mtx);
1283 ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1284 ASSERT(!list_link_active(&dr->dr_dirty_node));
1285 list_insert_tail(&di->dt.di.dr_children, dr);
1286 mutex_exit(&di->dt.di.dr_mtx);
1287 dr->dr_parent = di;
1288 }
1289 mutex_exit(&db->db_mtx);
1290 } else {
1291 ASSERT(db->db_level+1 == dn->dn_nlevels);
1292 ASSERT(db->db_blkid < dn->dn_nblkptr);
1293 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1294 mutex_enter(&dn->dn_mtx);
1295 ASSERT(!list_link_active(&dr->dr_dirty_node));
1296 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1297 mutex_exit(&dn->dn_mtx);
1298 if (drop_struct_lock)
1299 rw_exit(&dn->dn_struct_rwlock);
1300 }
1301
1302 dnode_setdirty(dn, tx);
1303 DB_DNODE_EXIT(db);
1304 return (dr);
1305 }
1306
1307 /*
1308 * Undirty a buffer in the transaction group referenced by the given
1309 * transaction. Return whether this evicted the dbuf.
1310 */
1311 static boolean_t
1312 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1313 {
1314 dnode_t *dn;
1315 uint64_t txg = tx->tx_txg;
1316 dbuf_dirty_record_t *dr, **drp;
1317
1318 ASSERT(txg != 0);
1319 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1320 ASSERT0(db->db_level);
1321 ASSERT(MUTEX_HELD(&db->db_mtx));
1322
1323 /*
1324 * If this buffer is not dirty, we're done.
1325 */
1326 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1327 if (dr->dr_txg <= txg)
1328 break;
1329 if (dr == NULL || dr->dr_txg < txg)
1330 return (B_FALSE);
1331 ASSERT(dr->dr_txg == txg);
1332 ASSERT(dr->dr_dbuf == db);
1333
1334 DB_DNODE_ENTER(db);
1335 dn = DB_DNODE(db);
1336
1337 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1338
1339 ASSERT(db->db.db_size != 0);
1340
1341 /*
1342 * Any space we accounted for in dp_dirty_* will be cleaned up by
1343 * dsl_pool_sync(). This is relatively rare so the discrepancy
1344 * is not a big deal.
1345 */
1346
1347 *drp = dr->dr_next;
1348
1349 /*
1350 * Note that there are three places in dbuf_dirty()
1351 * where this dirty record may be put on a list.
1352 * Make sure to do a list_remove corresponding to
1353 * every one of those list_insert calls.
1354 */
1355 if (dr->dr_parent) {
1356 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1357 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1358 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1359 } else if (db->db_blkid == DMU_SPILL_BLKID ||
1360 db->db_level+1 == dn->dn_nlevels) {
1361 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1362 mutex_enter(&dn->dn_mtx);
1363 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1364 mutex_exit(&dn->dn_mtx);
1365 }
1366 DB_DNODE_EXIT(db);
1367
1368 if (db->db_state != DB_NOFILL) {
1369 dbuf_unoverride(dr);
1370
1371 ASSERT(db->db_buf != NULL);
1372 ASSERT(dr->dt.dl.dr_data != NULL);
1373 if (dr->dt.dl.dr_data != db->db_buf)
1374 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db));
1375 }
1376 kmem_free(dr, sizeof (dbuf_dirty_record_t));
1377
1378 ASSERT(db->db_dirtycnt > 0);
1379 db->db_dirtycnt -= 1;
1380
1381 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1382 arc_buf_t *buf = db->db_buf;
1383
1384 ASSERT(db->db_state == DB_NOFILL || arc_released(buf));
1385 dbuf_set_data(db, NULL);
1386 VERIFY(arc_buf_remove_ref(buf, db));
1387 dbuf_evict(db);
1388 return (B_TRUE);
1389 }
1390
1391 return (B_FALSE);
1392 }
1393
1394 #pragma weak dmu_buf_will_dirty = dbuf_will_dirty
1395 void
1396 dbuf_will_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1397 {
1398 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1399
1400 ASSERT(tx->tx_txg != 0);
1401 ASSERT(!refcount_is_zero(&db->db_holds));
1402
1403 DB_DNODE_ENTER(db);
1404 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
1405 rf |= DB_RF_HAVESTRUCT;
1406 DB_DNODE_EXIT(db);
1407 (void) dbuf_read(db, NULL, rf);
1408 (void) dbuf_dirty(db, tx);
1409 }
1410
1411 void
1412 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1413 {
1414 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1415
1416 db->db_state = DB_NOFILL;
1417
1418 dmu_buf_will_fill(db_fake, tx);
1419 }
1420
1421 void
1422 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1423 {
1424 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1425
1426 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1427 ASSERT(tx->tx_txg != 0);
1428 ASSERT(db->db_level == 0);
1429 ASSERT(!refcount_is_zero(&db->db_holds));
1430
1431 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1432 dmu_tx_private_ok(tx));
1433
1434 dbuf_noread(db);
1435 (void) dbuf_dirty(db, tx);
1436 }
1437
1438 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1439 /* ARGSUSED */
1440 void
1441 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1442 {
1443 mutex_enter(&db->db_mtx);
1444 DBUF_VERIFY(db);
1445
1446 if (db->db_state == DB_FILL) {
1447 if (db->db_level == 0 && db->db_freed_in_flight) {
1448 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1449 /* we were freed while filling */
1450 /* XXX dbuf_undirty? */
1451 bzero(db->db.db_data, db->db.db_size);
1452 db->db_freed_in_flight = FALSE;
1453 }
1454 db->db_state = DB_CACHED;
1455 cv_broadcast(&db->db_changed);
1456 }
1457 mutex_exit(&db->db_mtx);
1458 }
1459
1460 /*
1461 * Directly assign a provided arc buf to a given dbuf if it's not referenced
1462 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1463 */
1464 void
1465 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
1466 {
1467 ASSERT(!refcount_is_zero(&db->db_holds));
1468 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1469 ASSERT(db->db_level == 0);
1470 ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA);
1471 ASSERT(buf != NULL);
1472 ASSERT(arc_buf_size(buf) == db->db.db_size);
1473 ASSERT(tx->tx_txg != 0);
1474
1475 arc_return_buf(buf, db);
1476 ASSERT(arc_released(buf));
1477
1478 mutex_enter(&db->db_mtx);
1479
1480 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1481 cv_wait(&db->db_changed, &db->db_mtx);
1482
1483 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
1484
1485 if (db->db_state == DB_CACHED &&
1486 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
1487 mutex_exit(&db->db_mtx);
1488 (void) dbuf_dirty(db, tx);
1489 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
1490 VERIFY(arc_buf_remove_ref(buf, db));
1491 xuio_stat_wbuf_copied();
1492 return;
1493 }
1494
1495 xuio_stat_wbuf_nocopy();
1496 if (db->db_state == DB_CACHED) {
1497 dbuf_dirty_record_t *dr = db->db_last_dirty;
1498
1499 ASSERT(db->db_buf != NULL);
1500 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
1501 ASSERT(dr->dt.dl.dr_data == db->db_buf);
1502 if (!arc_released(db->db_buf)) {
1503 ASSERT(dr->dt.dl.dr_override_state ==
1504 DR_OVERRIDDEN);
1505 arc_release(db->db_buf, db);
1506 }
1507 dr->dt.dl.dr_data = buf;
1508 VERIFY(arc_buf_remove_ref(db->db_buf, db));
1509 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
1510 arc_release(db->db_buf, db);
1511 VERIFY(arc_buf_remove_ref(db->db_buf, db));
1512 }
1513 db->db_buf = NULL;
1514 }
1515 ASSERT(db->db_buf == NULL);
1516 dbuf_set_data(db, buf);
1517 db->db_state = DB_FILL;
1518 mutex_exit(&db->db_mtx);
1519 (void) dbuf_dirty(db, tx);
1520 dbuf_fill_done(db, tx);
1521 }
1522
1523 /*
1524 * "Clear" the contents of this dbuf. This will mark the dbuf
1525 * EVICTING and clear *most* of its references. Unfortunately,
1526 * when we are not holding the dn_dbufs_mtx, we can't clear the
1527 * entry in the dn_dbufs list. We have to wait until dbuf_destroy()
1528 * in this case. For callers from the DMU we will usually see:
1529 * dbuf_clear()->arc_buf_evict()->dbuf_do_evict()->dbuf_destroy()
1530 * For the arc callback, we will usually see:
1531 * dbuf_do_evict()->dbuf_clear();dbuf_destroy()
1532 * Sometimes, though, we will get a mix of these two:
1533 * DMU: dbuf_clear()->arc_buf_evict()
1534 * ARC: dbuf_do_evict()->dbuf_destroy()
1535 */
1536 void
1537 dbuf_clear(dmu_buf_impl_t *db)
1538 {
1539 dnode_t *dn;
1540 dmu_buf_impl_t *parent = db->db_parent;
1541 dmu_buf_impl_t *dndb;
1542 int dbuf_gone = FALSE;
1543
1544 ASSERT(MUTEX_HELD(&db->db_mtx));
1545 ASSERT(refcount_is_zero(&db->db_holds));
1546
1547 dbuf_evict_user(db);
1548
1549 if (db->db_state == DB_CACHED) {
1550 ASSERT(db->db.db_data != NULL);
1551 if (db->db_blkid == DMU_BONUS_BLKID) {
1552 zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
1553 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
1554 }
1555 db->db.db_data = NULL;
1556 db->db_state = DB_UNCACHED;
1557 }
1558
1559 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
1560 ASSERT(db->db_data_pending == NULL);
1561
1562 db->db_state = DB_EVICTING;
1563 db->db_blkptr = NULL;
1564
1565 DB_DNODE_ENTER(db);
1566 dn = DB_DNODE(db);
1567 dndb = dn->dn_dbuf;
1568 if (db->db_blkid != DMU_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) {
1569 list_remove(&dn->dn_dbufs, db);
1570 (void) atomic_dec_32_nv(&dn->dn_dbufs_count);
1571 membar_producer();
1572 DB_DNODE_EXIT(db);
1573 /*
1574 * Decrementing the dbuf count means that the hold corresponding
1575 * to the removed dbuf is no longer discounted in dnode_move(),
1576 * so the dnode cannot be moved until after we release the hold.
1577 * The membar_producer() ensures visibility of the decremented
1578 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
1579 * release any lock.
1580 */
1581 dnode_rele(dn, db);
1582 db->db_dnode_handle = NULL;
1583 } else {
1584 DB_DNODE_EXIT(db);
1585 }
1586
1587 if (db->db_buf)
1588 dbuf_gone = arc_buf_evict(db->db_buf);
1589
1590 if (!dbuf_gone)
1591 mutex_exit(&db->db_mtx);
1592
1593 /*
1594 * If this dbuf is referenced from an indirect dbuf,
1595 * decrement the ref count on the indirect dbuf.
1596 */
1597 if (parent && parent != dndb)
1598 dbuf_rele(parent, db);
1599 }
1600
1601 static int
1602 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
1603 dmu_buf_impl_t **parentp, blkptr_t **bpp)
1604 {
1605 int nlevels, epbs;
1606
1607 *parentp = NULL;
1608 *bpp = NULL;
1609
1610 ASSERT(blkid != DMU_BONUS_BLKID);
1611
1612 if (blkid == DMU_SPILL_BLKID) {
1613 mutex_enter(&dn->dn_mtx);
1614 if (dn->dn_have_spill &&
1615 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
1616 *bpp = &dn->dn_phys->dn_spill;
1617 else
1618 *bpp = NULL;
1619 dbuf_add_ref(dn->dn_dbuf, NULL);
1620 *parentp = dn->dn_dbuf;
1621 mutex_exit(&dn->dn_mtx);
1622 return (0);
1623 }
1624
1625 if (dn->dn_phys->dn_nlevels == 0)
1626 nlevels = 1;
1627 else
1628 nlevels = dn->dn_phys->dn_nlevels;
1629
1630 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1631
1632 ASSERT3U(level * epbs, <, 64);
1633 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1634 if (level >= nlevels ||
1635 (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
1636 /* the buffer has no parent yet */
1637 return (SET_ERROR(ENOENT));
1638 } else if (level < nlevels-1) {
1639 /* this block is referenced from an indirect block */
1640 int err = dbuf_hold_impl(dn, level+1,
1641 blkid >> epbs, fail_sparse, NULL, parentp);
1642 if (err)
1643 return (err);
1644 err = dbuf_read(*parentp, NULL,
1645 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
1646 if (err) {
1647 dbuf_rele(*parentp, NULL);
1648 *parentp = NULL;
1649 return (err);
1650 }
1651 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
1652 (blkid & ((1ULL << epbs) - 1));
1653 return (0);
1654 } else {
1655 /* the block is referenced from the dnode */
1656 ASSERT3U(level, ==, nlevels-1);
1657 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
1658 blkid < dn->dn_phys->dn_nblkptr);
1659 if (dn->dn_dbuf) {
1660 dbuf_add_ref(dn->dn_dbuf, NULL);
1661 *parentp = dn->dn_dbuf;
1662 }
1663 *bpp = &dn->dn_phys->dn_blkptr[blkid];
1664 return (0);
1665 }
1666 }
1667
1668 static dmu_buf_impl_t *
1669 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
1670 dmu_buf_impl_t *parent, blkptr_t *blkptr)
1671 {
1672 objset_t *os = dn->dn_objset;
1673 dmu_buf_impl_t *db, *odb;
1674
1675 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1676 ASSERT(dn->dn_type != DMU_OT_NONE);
1677
1678 db = kmem_cache_alloc(dbuf_cache, KM_SLEEP);
1679
1680 db->db_objset = os;
1681 db->db.db_object = dn->dn_object;
1682 db->db_level = level;
1683 db->db_blkid = blkid;
1684 db->db_last_dirty = NULL;
1685 db->db_dirtycnt = 0;
1686 db->db_dnode_handle = dn->dn_handle;
1687 db->db_parent = parent;
1688 db->db_blkptr = blkptr;
1689
1690 db->db_user_ptr = NULL;
1691 db->db_user_data_ptr_ptr = NULL;
1692 db->db_evict_func = NULL;
1693 db->db_immediate_evict = 0;
1694 db->db_freed_in_flight = 0;
1695
1696 if (blkid == DMU_BONUS_BLKID) {
1697 ASSERT3P(parent, ==, dn->dn_dbuf);
1698 db->db.db_size = DN_MAX_BONUSLEN -
1699 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
1700 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
1701 db->db.db_offset = DMU_BONUS_BLKID;
1702 db->db_state = DB_UNCACHED;
1703 /* the bonus dbuf is not placed in the hash table */
1704 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1705 return (db);
1706 } else if (blkid == DMU_SPILL_BLKID) {
1707 db->db.db_size = (blkptr != NULL) ?
1708 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
1709 db->db.db_offset = 0;
1710 } else {
1711 int blocksize =
1712 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
1713 db->db.db_size = blocksize;
1714 db->db.db_offset = db->db_blkid * blocksize;
1715 }
1716
1717 /*
1718 * Hold the dn_dbufs_mtx while we get the new dbuf
1719 * in the hash table *and* added to the dbufs list.
1720 * This prevents a possible deadlock with someone
1721 * trying to look up this dbuf before its added to the
1722 * dn_dbufs list.
1723 */
1724 mutex_enter(&dn->dn_dbufs_mtx);
1725 db->db_state = DB_EVICTING;
1726 if ((odb = dbuf_hash_insert(db)) != NULL) {
1727 /* someone else inserted it first */
1728 kmem_cache_free(dbuf_cache, db);
1729 mutex_exit(&dn->dn_dbufs_mtx);
1730 return (odb);
1731 }
1732 list_insert_head(&dn->dn_dbufs, db);
1733 if (db->db_level == 0 && db->db_blkid >=
1734 dn->dn_unlisted_l0_blkid)
1735 dn->dn_unlisted_l0_blkid = db->db_blkid + 1;
1736 db->db_state = DB_UNCACHED;
1737 mutex_exit(&dn->dn_dbufs_mtx);
1738 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1739
1740 if (parent && parent != dn->dn_dbuf)
1741 dbuf_add_ref(parent, db);
1742
1743 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1744 refcount_count(&dn->dn_holds) > 0);
1745 (void) refcount_add(&dn->dn_holds, db);
1746 (void) atomic_inc_32_nv(&dn->dn_dbufs_count);
1747
1748 dprintf_dbuf(db, "db=%p\n", db);
1749
1750 return (db);
1751 }
1752
1753 static int
1754 dbuf_do_evict(void *private)
1755 {
1756 arc_buf_t *buf = private;
1757 dmu_buf_impl_t *db = buf->b_private;
1758
1759 if (!MUTEX_HELD(&db->db_mtx))
1760 mutex_enter(&db->db_mtx);
1761
1762 ASSERT(refcount_is_zero(&db->db_holds));
1763
1764 if (db->db_state != DB_EVICTING) {
1765 ASSERT(db->db_state == DB_CACHED);
1766 DBUF_VERIFY(db);
1767 db->db_buf = NULL;
1768 dbuf_evict(db);
1769 } else {
1770 mutex_exit(&db->db_mtx);
1771 dbuf_destroy(db);
1772 }
1773 return (0);
1774 }
1775
1776 static void
1777 dbuf_destroy(dmu_buf_impl_t *db)
1778 {
1779 ASSERT(refcount_is_zero(&db->db_holds));
1780
1781 if (db->db_blkid != DMU_BONUS_BLKID) {
1782 /*
1783 * If this dbuf is still on the dn_dbufs list,
1784 * remove it from that list.
1785 */
1786 if (db->db_dnode_handle != NULL) {
1787 dnode_t *dn;
1788
1789 DB_DNODE_ENTER(db);
1790 dn = DB_DNODE(db);
1791 mutex_enter(&dn->dn_dbufs_mtx);
1792 list_remove(&dn->dn_dbufs, db);
1793 (void) atomic_dec_32_nv(&dn->dn_dbufs_count);
1794 mutex_exit(&dn->dn_dbufs_mtx);
1795 DB_DNODE_EXIT(db);
1796 /*
1797 * Decrementing the dbuf count means that the hold
1798 * corresponding to the removed dbuf is no longer
1799 * discounted in dnode_move(), so the dnode cannot be
1800 * moved until after we release the hold.
1801 */
1802 dnode_rele(dn, db);
1803 db->db_dnode_handle = NULL;
1804 }
1805 dbuf_hash_remove(db);
1806 }
1807 db->db_parent = NULL;
1808 db->db_buf = NULL;
1809
1810 ASSERT(!list_link_active(&db->db_link));
1811 ASSERT(db->db.db_data == NULL);
1812 ASSERT(db->db_hash_next == NULL);
1813 ASSERT(db->db_blkptr == NULL);
1814 ASSERT(db->db_data_pending == NULL);
1815
1816 kmem_cache_free(dbuf_cache, db);
1817 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1818 }
1819
1820 void
1821 dbuf_prefetch(dnode_t *dn, uint64_t blkid, zio_priority_t prio)
1822 {
1823 dmu_buf_impl_t *db = NULL;
1824 blkptr_t *bp = NULL;
1825
1826 ASSERT(blkid != DMU_BONUS_BLKID);
1827 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1828
1829 if (dnode_block_freed(dn, blkid))
1830 return;
1831
1832 /* dbuf_find() returns with db_mtx held */
1833 if (db = dbuf_find(dn, 0, blkid)) {
1834 /*
1835 * This dbuf is already in the cache. We assume that
1836 * it is already CACHED, or else about to be either
1837 * read or filled.
1838 */
1839 mutex_exit(&db->db_mtx);
1840 return;
1841 }
1842
1843 if (dbuf_findbp(dn, 0, blkid, TRUE, &db, &bp) == 0) {
1844 if (bp && !BP_IS_HOLE(bp)) {
1845 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
1846 uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH;
1847 zbookmark_t zb;
1848
1849 SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
1850 dn->dn_object, 0, blkid);
1851
1852 (void) arc_read(NULL, dn->dn_objset->os_spa,
1853 bp, NULL, NULL, prio,
1854 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1855 &aflags, &zb);
1856 }
1857 if (db)
1858 dbuf_rele(db, NULL);
1859 }
1860 }
1861
1862 /*
1863 * Returns with db_holds incremented, and db_mtx not held.
1864 * Note: dn_struct_rwlock must be held.
1865 */
1866 int
1867 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse,
1868 void *tag, dmu_buf_impl_t **dbp)
1869 {
1870 dmu_buf_impl_t *db, *parent = NULL;
1871
1872 ASSERT(blkid != DMU_BONUS_BLKID);
1873 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1874 ASSERT3U(dn->dn_nlevels, >, level);
1875
1876 *dbp = NULL;
1877 top:
1878 /* dbuf_find() returns with db_mtx held */
1879 db = dbuf_find(dn, level, blkid);
1880
1881 if (db == NULL) {
1882 blkptr_t *bp = NULL;
1883 int err;
1884
1885 ASSERT3P(parent, ==, NULL);
1886 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
1887 if (fail_sparse) {
1888 if (err == 0 && bp && BP_IS_HOLE(bp))
1889 err = SET_ERROR(ENOENT);
1890 if (err) {
1891 if (parent)
1892 dbuf_rele(parent, NULL);
1893 return (err);
1894 }
1895 }
1896 if (err && err != ENOENT)
1897 return (err);
1898 db = dbuf_create(dn, level, blkid, parent, bp);
1899 }
1900
1901 if (db->db_buf && refcount_is_zero(&db->db_holds)) {
1902 arc_buf_add_ref(db->db_buf, db);
1903 if (db->db_buf->b_data == NULL) {
1904 dbuf_clear(db);
1905 if (parent) {
1906 dbuf_rele(parent, NULL);
1907 parent = NULL;
1908 }
1909 goto top;
1910 }
1911 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
1912 }
1913
1914 ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
1915
1916 /*
1917 * If this buffer is currently syncing out, and we are are
1918 * still referencing it from db_data, we need to make a copy
1919 * of it in case we decide we want to dirty it again in this txg.
1920 */
1921 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1922 dn->dn_object != DMU_META_DNODE_OBJECT &&
1923 db->db_state == DB_CACHED && db->db_data_pending) {
1924 dbuf_dirty_record_t *dr = db->db_data_pending;
1925
1926 if (dr->dt.dl.dr_data == db->db_buf) {
1927 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1928
1929 dbuf_set_data(db,
1930 arc_buf_alloc(dn->dn_objset->os_spa,
1931 db->db.db_size, db, type));
1932 bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
1933 db->db.db_size);
1934 }
1935 }
1936
1937 (void) refcount_add(&db->db_holds, tag);
1938 dbuf_update_data(db);
1939 DBUF_VERIFY(db);
1940 mutex_exit(&db->db_mtx);
1941
1942 /* NOTE: we can't rele the parent until after we drop the db_mtx */
1943 if (parent)
1944 dbuf_rele(parent, NULL);
1945
1946 ASSERT3P(DB_DNODE(db), ==, dn);
1947 ASSERT3U(db->db_blkid, ==, blkid);
1948 ASSERT3U(db->db_level, ==, level);
1949 *dbp = db;
1950
1951 return (0);
1952 }
1953
1954 dmu_buf_impl_t *
1955 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
1956 {
1957 dmu_buf_impl_t *db;
1958 int err = dbuf_hold_impl(dn, 0, blkid, FALSE, tag, &db);
1959 return (err ? NULL : db);
1960 }
1961
1962 dmu_buf_impl_t *
1963 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
1964 {
1965 dmu_buf_impl_t *db;
1966 int err = dbuf_hold_impl(dn, level, blkid, FALSE, tag, &db);
1967 return (err ? NULL : db);
1968 }
1969
1970 void
1971 dbuf_create_bonus(dnode_t *dn)
1972 {
1973 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1974
1975 ASSERT(dn->dn_bonus == NULL);
1976 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
1977 }
1978
1979 int
1980 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
1981 {
1982 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1983 dnode_t *dn;
1984
1985 if (db->db_blkid != DMU_SPILL_BLKID)
1986 return (SET_ERROR(ENOTSUP));
1987 if (blksz == 0)
1988 blksz = SPA_MINBLOCKSIZE;
1989 if (blksz > SPA_MAXBLOCKSIZE)
1990 blksz = SPA_MAXBLOCKSIZE;
1991 else
1992 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
1993
1994 DB_DNODE_ENTER(db);
1995 dn = DB_DNODE(db);
1996 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1997 dbuf_new_size(db, blksz, tx);
1998 rw_exit(&dn->dn_struct_rwlock);
1999 DB_DNODE_EXIT(db);
2000
2001 return (0);
2002 }
2003
2004 void
2005 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
2006 {
2007 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
2008 }
2009
2010 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2011 void
2012 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
2013 {
2014 int64_t holds = refcount_add(&db->db_holds, tag);
2015 ASSERT(holds > 1);
2016 }
2017
2018 /*
2019 * If you call dbuf_rele() you had better not be referencing the dnode handle
2020 * unless you have some other direct or indirect hold on the dnode. (An indirect
2021 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2022 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2023 * dnode's parent dbuf evicting its dnode handles.
2024 */
2025 #pragma weak dmu_buf_rele = dbuf_rele
2026 void
2027 dbuf_rele(dmu_buf_impl_t *db, void *tag)
2028 {
2029 mutex_enter(&db->db_mtx);
2030 dbuf_rele_and_unlock(db, tag);
2031 }
2032
2033 /*
2034 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2035 * db_dirtycnt and db_holds to be updated atomically.
2036 */
2037 void
2038 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2039 {
2040 int64_t holds;
2041
2042 ASSERT(MUTEX_HELD(&db->db_mtx));
2043 DBUF_VERIFY(db);
2044
2045 /*
2046 * Remove the reference to the dbuf before removing its hold on the
2047 * dnode so we can guarantee in dnode_move() that a referenced bonus
2048 * buffer has a corresponding dnode hold.
2049 */
2050 holds = refcount_remove(&db->db_holds, tag);
2051 ASSERT(holds >= 0);
2052
2053 /*
2054 * We can't freeze indirects if there is a possibility that they
2055 * may be modified in the current syncing context.
2056 */
2057 if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0))
2058 arc_buf_freeze(db->db_buf);
2059
2060 if (holds == db->db_dirtycnt &&
2061 db->db_level == 0 && db->db_immediate_evict)
2062 dbuf_evict_user(db);
2063
2064 if (holds == 0) {
2065 if (db->db_blkid == DMU_BONUS_BLKID) {
2066 mutex_exit(&db->db_mtx);
2067
2068 /*
2069 * If the dnode moves here, we cannot cross this barrier
2070 * until the move completes.
2071 */
2072 DB_DNODE_ENTER(db);
2073 (void) atomic_dec_32_nv(&DB_DNODE(db)->dn_dbufs_count);
2074 DB_DNODE_EXIT(db);
2075 /*
2076 * The bonus buffer's dnode hold is no longer discounted
2077 * in dnode_move(). The dnode cannot move until after
2078 * the dnode_rele().
2079 */
2080 dnode_rele(DB_DNODE(db), db);
2081 } else if (db->db_buf == NULL) {
2082 /*
2083 * This is a special case: we never associated this
2084 * dbuf with any data allocated from the ARC.
2085 */
2086 ASSERT(db->db_state == DB_UNCACHED ||
2087 db->db_state == DB_NOFILL);
2088 dbuf_evict(db);
2089 } else if (arc_released(db->db_buf)) {
2090 arc_buf_t *buf = db->db_buf;
2091 /*
2092 * This dbuf has anonymous data associated with it.
2093 */
2094 dbuf_set_data(db, NULL);
2095 VERIFY(arc_buf_remove_ref(buf, db));
2096 dbuf_evict(db);
2097 } else {
2098 VERIFY(!arc_buf_remove_ref(db->db_buf, db));
2099
2100 /*
2101 * A dbuf will be eligible for eviction if either the
2102 * 'primarycache' property is set or a duplicate
2103 * copy of this buffer is already cached in the arc.
2104 *
2105 * In the case of the 'primarycache' a buffer
2106 * is considered for eviction if it matches the
2107 * criteria set in the property.
2108 *
2109 * To decide if our buffer is considered a
2110 * duplicate, we must call into the arc to determine
2111 * if multiple buffers are referencing the same
2112 * block on-disk. If so, then we simply evict
2113 * ourselves.
2114 */
2115 if (!DBUF_IS_CACHEABLE(db) ||
2116 arc_buf_eviction_needed(db->db_buf))
2117 dbuf_clear(db);
2118 else
2119 mutex_exit(&db->db_mtx);
2120 }
2121 } else {
2122 mutex_exit(&db->db_mtx);
2123 }
2124 }
2125
2126 #pragma weak dmu_buf_refcount = dbuf_refcount
2127 uint64_t
2128 dbuf_refcount(dmu_buf_impl_t *db)
2129 {
2130 return (refcount_count(&db->db_holds));
2131 }
2132
2133 void *
2134 dmu_buf_set_user(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr,
2135 dmu_buf_evict_func_t *evict_func)
2136 {
2137 return (dmu_buf_update_user(db_fake, NULL, user_ptr,
2138 user_data_ptr_ptr, evict_func));
2139 }
2140
2141 void *
2142 dmu_buf_set_user_ie(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr,
2143 dmu_buf_evict_func_t *evict_func)
2144 {
2145 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2146
2147 db->db_immediate_evict = TRUE;
2148 return (dmu_buf_update_user(db_fake, NULL, user_ptr,
2149 user_data_ptr_ptr, evict_func));
2150 }
2151
2152 void *
2153 dmu_buf_update_user(dmu_buf_t *db_fake, void *old_user_ptr, void *user_ptr,
2154 void *user_data_ptr_ptr, dmu_buf_evict_func_t *evict_func)
2155 {
2156 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2157 ASSERT(db->db_level == 0);
2158
2159 ASSERT((user_ptr == NULL) == (evict_func == NULL));
2160
2161 mutex_enter(&db->db_mtx);
2162
2163 if (db->db_user_ptr == old_user_ptr) {
2164 db->db_user_ptr = user_ptr;
2165 db->db_user_data_ptr_ptr = user_data_ptr_ptr;
2166 db->db_evict_func = evict_func;
2167
2168 dbuf_update_data(db);
2169 } else {
2170 old_user_ptr = db->db_user_ptr;
2171 }
2172
2173 mutex_exit(&db->db_mtx);
2174 return (old_user_ptr);
2175 }
2176
2177 void *
2178 dmu_buf_get_user(dmu_buf_t *db_fake)
2179 {
2180 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2181 ASSERT(!refcount_is_zero(&db->db_holds));
2182
2183 return (db->db_user_ptr);
2184 }
2185
2186 boolean_t
2187 dmu_buf_freeable(dmu_buf_t *dbuf)
2188 {
2189 boolean_t res = B_FALSE;
2190 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2191
2192 if (db->db_blkptr)
2193 res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset,
2194 db->db_blkptr, db->db_blkptr->blk_birth);
2195
2196 return (res);
2197 }
2198
2199 blkptr_t *
2200 dmu_buf_get_blkptr(dmu_buf_t *db)
2201 {
2202 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2203 return (dbi->db_blkptr);
2204 }
2205
2206 static void
2207 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
2208 {
2209 /* ASSERT(dmu_tx_is_syncing(tx) */
2210 ASSERT(MUTEX_HELD(&db->db_mtx));
2211
2212 if (db->db_blkptr != NULL)
2213 return;
2214
2215 if (db->db_blkid == DMU_SPILL_BLKID) {
2216 db->db_blkptr = &dn->dn_phys->dn_spill;
2217 BP_ZERO(db->db_blkptr);
2218 return;
2219 }
2220 if (db->db_level == dn->dn_phys->dn_nlevels-1) {
2221 /*
2222 * This buffer was allocated at a time when there was
2223 * no available blkptrs from the dnode, or it was
2224 * inappropriate to hook it in (i.e., nlevels mis-match).
2225 */
2226 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
2227 ASSERT(db->db_parent == NULL);
2228 db->db_parent = dn->dn_dbuf;
2229 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
2230 DBUF_VERIFY(db);
2231 } else {
2232 dmu_buf_impl_t *parent = db->db_parent;
2233 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2234
2235 ASSERT(dn->dn_phys->dn_nlevels > 1);
2236 if (parent == NULL) {
2237 mutex_exit(&db->db_mtx);
2238 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2239 (void) dbuf_hold_impl(dn, db->db_level+1,
2240 db->db_blkid >> epbs, FALSE, db, &parent);
2241 rw_exit(&dn->dn_struct_rwlock);
2242 mutex_enter(&db->db_mtx);
2243 db->db_parent = parent;
2244 }
2245 db->db_blkptr = (blkptr_t *)parent->db.db_data +
2246 (db->db_blkid & ((1ULL << epbs) - 1));
2247 DBUF_VERIFY(db);
2248 }
2249 }
2250
2251 static void
2252 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2253 {
2254 dmu_buf_impl_t *db = dr->dr_dbuf;
2255 dnode_t *dn;
2256 zio_t *zio;
2257
2258 ASSERT(dmu_tx_is_syncing(tx));
2259
2260 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2261
2262 mutex_enter(&db->db_mtx);
2263
2264 ASSERT(db->db_level > 0);
2265 DBUF_VERIFY(db);
2266
2267 /* Read the block if it hasn't been read yet. */
2268 if (db->db_buf == NULL) {
2269 mutex_exit(&db->db_mtx);
2270 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
2271 mutex_enter(&db->db_mtx);
2272 }
2273 ASSERT3U(db->db_state, ==, DB_CACHED);
2274 ASSERT(db->db_buf != NULL);
2275
2276 DB_DNODE_ENTER(db);
2277 dn = DB_DNODE(db);
2278 /* Indirect block size must match what the dnode thinks it is. */
2279 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2280 dbuf_check_blkptr(dn, db);
2281 DB_DNODE_EXIT(db);
2282
2283 /* Provide the pending dirty record to child dbufs */
2284 db->db_data_pending = dr;
2285
2286 mutex_exit(&db->db_mtx);
2287 dbuf_write(dr, db->db_buf, tx);
2288
2289 zio = dr->dr_zio;
2290 mutex_enter(&dr->dt.di.dr_mtx);
2291 dbuf_sync_list(&dr->dt.di.dr_children, tx);
2292 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2293 mutex_exit(&dr->dt.di.dr_mtx);
2294 zio_nowait(zio);
2295 }
2296
2297 static void
2298 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2299 {
2300 arc_buf_t **datap = &dr->dt.dl.dr_data;
2301 dmu_buf_impl_t *db = dr->dr_dbuf;
2302 dnode_t *dn;
2303 objset_t *os;
2304 uint64_t txg = tx->tx_txg;
2305
2306 ASSERT(dmu_tx_is_syncing(tx));
2307
2308 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2309
2310 mutex_enter(&db->db_mtx);
2311 /*
2312 * To be synced, we must be dirtied. But we
2313 * might have been freed after the dirty.
2314 */
2315 if (db->db_state == DB_UNCACHED) {
2316 /* This buffer has been freed since it was dirtied */
2317 ASSERT(db->db.db_data == NULL);
2318 } else if (db->db_state == DB_FILL) {
2319 /* This buffer was freed and is now being re-filled */
2320 ASSERT(db->db.db_data != dr->dt.dl.dr_data);
2321 } else {
2322 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
2323 }
2324 DBUF_VERIFY(db);
2325
2326 DB_DNODE_ENTER(db);
2327 dn = DB_DNODE(db);
2328
2329 if (db->db_blkid == DMU_SPILL_BLKID) {
2330 mutex_enter(&dn->dn_mtx);
2331 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
2332 mutex_exit(&dn->dn_mtx);
2333 }
2334
2335 /*
2336 * If this is a bonus buffer, simply copy the bonus data into the
2337 * dnode. It will be written out when the dnode is synced (and it
2338 * will be synced, since it must have been dirty for dbuf_sync to
2339 * be called).
2340 */
2341 if (db->db_blkid == DMU_BONUS_BLKID) {
2342 dbuf_dirty_record_t **drp;
2343
2344 ASSERT(*datap != NULL);
2345 ASSERT0(db->db_level);
2346 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN);
2347 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen);
2348 DB_DNODE_EXIT(db);
2349
2350 if (*datap != db->db.db_data) {
2351 zio_buf_free(*datap, DN_MAX_BONUSLEN);
2352 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
2353 }
2354 db->db_data_pending = NULL;
2355 drp = &db->db_last_dirty;
2356 while (*drp != dr)
2357 drp = &(*drp)->dr_next;
2358 ASSERT(dr->dr_next == NULL);
2359 ASSERT(dr->dr_dbuf == db);
2360 *drp = dr->dr_next;
2361 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2362 ASSERT(db->db_dirtycnt > 0);
2363 db->db_dirtycnt -= 1;
2364 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
2365 return;
2366 }
2367
2368 os = dn->dn_objset;
2369
2370 /*
2371 * This function may have dropped the db_mtx lock allowing a dmu_sync
2372 * operation to sneak in. As a result, we need to ensure that we
2373 * don't check the dr_override_state until we have returned from
2374 * dbuf_check_blkptr.
2375 */
2376 dbuf_check_blkptr(dn, db);
2377
2378 /*
2379 * If this buffer is in the middle of an immediate write,
2380 * wait for the synchronous IO to complete.
2381 */
2382 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
2383 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
2384 cv_wait(&db->db_changed, &db->db_mtx);
2385 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
2386 }
2387
2388 if (db->db_state != DB_NOFILL &&
2389 dn->dn_object != DMU_META_DNODE_OBJECT &&
2390 refcount_count(&db->db_holds) > 1 &&
2391 dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
2392 *datap == db->db_buf) {
2393 /*
2394 * If this buffer is currently "in use" (i.e., there
2395 * are active holds and db_data still references it),
2396 * then make a copy before we start the write so that
2397 * any modifications from the open txg will not leak
2398 * into this write.
2399 *
2400 * NOTE: this copy does not need to be made for
2401 * objects only modified in the syncing context (e.g.
2402 * DNONE_DNODE blocks).
2403 */
2404 int blksz = arc_buf_size(*datap);
2405 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2406 *datap = arc_buf_alloc(os->os_spa, blksz, db, type);
2407 bcopy(db->db.db_data, (*datap)->b_data, blksz);
2408 }
2409 db->db_data_pending = dr;
2410
2411 mutex_exit(&db->db_mtx);
2412
2413 dbuf_write(dr, *datap, tx);
2414
2415 ASSERT(!list_link_active(&dr->dr_dirty_node));
2416 if (dn->dn_object == DMU_META_DNODE_OBJECT) {
2417 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
2418 DB_DNODE_EXIT(db);
2419 } else {
2420 /*
2421 * Although zio_nowait() does not "wait for an IO", it does
2422 * initiate the IO. If this is an empty write it seems plausible
2423 * that the IO could actually be completed before the nowait
2424 * returns. We need to DB_DNODE_EXIT() first in case
2425 * zio_nowait() invalidates the dbuf.
2426 */
2427 DB_DNODE_EXIT(db);
2428 zio_nowait(dr->dr_zio);
2429 }
2430 }
2431
2432 void
2433 dbuf_sync_list(list_t *list, dmu_tx_t *tx)
2434 {
2435 dbuf_dirty_record_t *dr;
2436
2437 while (dr = list_head(list)) {
2438 if (dr->dr_zio != NULL) {
2439 /*
2440 * If we find an already initialized zio then we
2441 * are processing the meta-dnode, and we have finished.
2442 * The dbufs for all dnodes are put back on the list
2443 * during processing, so that we can zio_wait()
2444 * these IOs after initiating all child IOs.
2445 */
2446 ASSERT3U(dr->dr_dbuf->db.db_object, ==,
2447 DMU_META_DNODE_OBJECT);
2448 break;
2449 }
2450 list_remove(list, dr);
2451 if (dr->dr_dbuf->db_level > 0)
2452 dbuf_sync_indirect(dr, tx);
2453 else
2454 dbuf_sync_leaf(dr, tx);
2455 }
2456 }
2457
2458 /* ARGSUSED */
2459 static void
2460 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
2461 {
2462 dmu_buf_impl_t *db = vdb;
2463 dnode_t *dn;
2464 blkptr_t *bp = zio->io_bp;
2465 blkptr_t *bp_orig = &zio->io_bp_orig;
2466 spa_t *spa = zio->io_spa;
2467 int64_t delta;
2468 uint64_t fill = 0;
2469 int i;
2470
2471 ASSERT(db->db_blkptr == bp);
2472
2473 DB_DNODE_ENTER(db);
2474 dn = DB_DNODE(db);
2475 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
2476 dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
2477 zio->io_prev_space_delta = delta;
2478
2479 if (BP_IS_HOLE(bp)) {
2480 ASSERT(bp->blk_fill == 0);
2481 DB_DNODE_EXIT(db);
2482 return;
2483 }
2484
2485 ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
2486 BP_GET_TYPE(bp) == dn->dn_type) ||
2487 (db->db_blkid == DMU_SPILL_BLKID &&
2488 BP_GET_TYPE(bp) == dn->dn_bonustype));
2489 ASSERT(BP_GET_LEVEL(bp) == db->db_level);
2490
2491 mutex_enter(&db->db_mtx);
2492
2493 #ifdef ZFS_DEBUG
2494 if (db->db_blkid == DMU_SPILL_BLKID) {
2495 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2496 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2497 db->db_blkptr == &dn->dn_phys->dn_spill);
2498 }
2499 #endif
2500
2501 if (db->db_level == 0) {
2502 mutex_enter(&dn->dn_mtx);
2503 if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
2504 db->db_blkid != DMU_SPILL_BLKID)
2505 dn->dn_phys->dn_maxblkid = db->db_blkid;
2506 mutex_exit(&dn->dn_mtx);
2507
2508 if (dn->dn_type == DMU_OT_DNODE) {
2509 dnode_phys_t *dnp = db->db.db_data;
2510 for (i = db->db.db_size >> DNODE_SHIFT; i > 0;
2511 i--, dnp++) {
2512 if (dnp->dn_type != DMU_OT_NONE)
2513 fill++;
2514 }
2515 } else {
2516 fill = 1;
2517 }
2518 } else {
2519 blkptr_t *ibp = db->db.db_data;
2520 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2521 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
2522 if (BP_IS_HOLE(ibp))
2523 continue;
2524 fill += ibp->blk_fill;
2525 }
2526 }
2527 DB_DNODE_EXIT(db);
2528
2529 bp->blk_fill = fill;
2530
2531 mutex_exit(&db->db_mtx);
2532 }
2533
2534 /*
2535 * The SPA will call this callback several times for each zio - once
2536 * for every physical child i/o (zio->io_phys_children times). This
2537 * allows the DMU to monitor the progress of each logical i/o. For example,
2538 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
2539 * block. There may be a long delay before all copies/fragments are completed,
2540 * so this callback allows us to retire dirty space gradually, as the physical
2541 * i/os complete.
2542 */
2543 /* ARGSUSED */
2544 static void
2545 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
2546 {
2547 dmu_buf_impl_t *db = arg;
2548 objset_t *os = db->db_objset;
2549 dsl_pool_t *dp = dmu_objset_pool(os);
2550 dbuf_dirty_record_t *dr;
2551 int delta = 0;
2552
2553 dr = db->db_data_pending;
2554 ASSERT3U(dr->dr_txg, ==, zio->io_txg);
2555
2556 /*
2557 * The callback will be called io_phys_children times. Retire one
2558 * portion of our dirty space each time we are called. Any rounding
2559 * error will be cleaned up by dsl_pool_sync()'s call to
2560 * dsl_pool_undirty_space().
2561 */
2562 delta = dr->dr_accounted / zio->io_phys_children;
2563 dsl_pool_undirty_space(dp, delta, zio->io_txg);
2564 }
2565
2566 /* ARGSUSED */
2567 static void
2568 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
2569 {
2570 dmu_buf_impl_t *db = vdb;
2571 blkptr_t *bp = zio->io_bp;
2572 blkptr_t *bp_orig = &zio->io_bp_orig;
2573 uint64_t txg = zio->io_txg;
2574 dbuf_dirty_record_t **drp, *dr;
2575
2576 ASSERT0(zio->io_error);
2577 ASSERT(db->db_blkptr == bp);
2578
2579 /*
2580 * For nopwrites and rewrites we ensure that the bp matches our
2581 * original and bypass all the accounting.
2582 */
2583 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
2584 ASSERT(BP_EQUAL(bp, bp_orig));
2585 } else {
2586 objset_t *os;
2587 dsl_dataset_t *ds;
2588 dmu_tx_t *tx;
2589
2590 DB_GET_OBJSET(&os, db);
2591 ds = os->os_dsl_dataset;
2592 tx = os->os_synctx;
2593
2594 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
2595 dsl_dataset_block_born(ds, bp, tx);
2596 }
2597
2598 mutex_enter(&db->db_mtx);
2599
2600 DBUF_VERIFY(db);
2601
2602 drp = &db->db_last_dirty;
2603 while ((dr = *drp) != db->db_data_pending)
2604 drp = &dr->dr_next;
2605 ASSERT(!list_link_active(&dr->dr_dirty_node));
2606 ASSERT(dr->dr_txg == txg);
2607 ASSERT(dr->dr_dbuf == db);
2608 ASSERT(dr->dr_next == NULL);
2609 *drp = dr->dr_next;
2610
2611 #ifdef ZFS_DEBUG
2612 if (db->db_blkid == DMU_SPILL_BLKID) {
2613 dnode_t *dn;
2614
2615 DB_DNODE_ENTER(db);
2616 dn = DB_DNODE(db);
2617 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2618 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2619 db->db_blkptr == &dn->dn_phys->dn_spill);
2620 DB_DNODE_EXIT(db);
2621 }
2622 #endif
2623
2624 if (db->db_level == 0) {
2625 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2626 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
2627 if (db->db_state != DB_NOFILL) {
2628 if (dr->dt.dl.dr_data != db->db_buf)
2629 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data,
2630 db));
2631 else if (!arc_released(db->db_buf))
2632 arc_set_callback(db->db_buf, dbuf_do_evict, db);
2633 }
2634 } else {
2635 dnode_t *dn;
2636
2637 DB_DNODE_ENTER(db);
2638 dn = DB_DNODE(db);
2639 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2640 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2641 if (!BP_IS_HOLE(db->db_blkptr)) {
2642 int epbs =
2643 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2644 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
2645 db->db.db_size);
2646 ASSERT3U(dn->dn_phys->dn_maxblkid
2647 >> (db->db_level * epbs), >=, db->db_blkid);
2648 arc_set_callback(db->db_buf, dbuf_do_evict, db);
2649 }
2650 DB_DNODE_EXIT(db);
2651 mutex_destroy(&dr->dt.di.dr_mtx);
2652 list_destroy(&dr->dt.di.dr_children);
2653 }
2654 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2655
2656 cv_broadcast(&db->db_changed);
2657 ASSERT(db->db_dirtycnt > 0);
2658 db->db_dirtycnt -= 1;
2659 db->db_data_pending = NULL;
2660
2661 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
2662 }
2663
2664 static void
2665 dbuf_write_nofill_ready(zio_t *zio)
2666 {
2667 dbuf_write_ready(zio, NULL, zio->io_private);
2668 }
2669
2670 static void
2671 dbuf_write_nofill_done(zio_t *zio)
2672 {
2673 dbuf_write_done(zio, NULL, zio->io_private);
2674 }
2675
2676 static void
2677 dbuf_write_override_ready(zio_t *zio)
2678 {
2679 dbuf_dirty_record_t *dr = zio->io_private;
2680 dmu_buf_impl_t *db = dr->dr_dbuf;
2681
2682 dbuf_write_ready(zio, NULL, db);
2683 }
2684
2685 static void
2686 dbuf_write_override_done(zio_t *zio)
2687 {
2688 dbuf_dirty_record_t *dr = zio->io_private;
2689 dmu_buf_impl_t *db = dr->dr_dbuf;
2690 blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
2691
2692 mutex_enter(&db->db_mtx);
2693 if (!BP_EQUAL(zio->io_bp, obp)) {
2694 if (!BP_IS_HOLE(obp))
2695 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
2696 arc_release(dr->dt.dl.dr_data, db);
2697 }
2698 mutex_exit(&db->db_mtx);
2699
2700 dbuf_write_done(zio, NULL, db);
2701 }
2702
2703 /* Issue I/O to commit a dirty buffer to disk. */
2704 static void
2705 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
2706 {
2707 dmu_buf_impl_t *db = dr->dr_dbuf;
2708 dnode_t *dn;
2709 objset_t *os;
2710 dmu_buf_impl_t *parent = db->db_parent;
2711 uint64_t txg = tx->tx_txg;
2712 zbookmark_t zb;
2713 zio_prop_t zp;
2714 zio_t *zio;
2715 int wp_flag = 0;
2716
2717 DB_DNODE_ENTER(db);
2718 dn = DB_DNODE(db);
2719 os = dn->dn_objset;
2720
2721 if (db->db_state != DB_NOFILL) {
2722 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
2723 /*
2724 * Private object buffers are released here rather
2725 * than in dbuf_dirty() since they are only modified
2726 * in the syncing context and we don't want the
2727 * overhead of making multiple copies of the data.
2728 */
2729 if (BP_IS_HOLE(db->db_blkptr)) {
2730 arc_buf_thaw(data);
2731 } else {
2732 dbuf_release_bp(db);
2733 }
2734 }
2735 }
2736
2737 if (parent != dn->dn_dbuf) {
2738 /* Our parent is an indirect block. */
2739 /* We have a dirty parent that has been scheduled for write. */
2740 ASSERT(parent && parent->db_data_pending);
2741 /* Our parent's buffer is one level closer to the dnode. */
2742 ASSERT(db->db_level == parent->db_level-1);
2743 /*
2744 * We're about to modify our parent's db_data by modifying
2745 * our block pointer, so the parent must be released.
2746 */
2747 ASSERT(arc_released(parent->db_buf));
2748 zio = parent->db_data_pending->dr_zio;
2749 } else {
2750 /* Our parent is the dnode itself. */
2751 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
2752 db->db_blkid != DMU_SPILL_BLKID) ||
2753 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
2754 if (db->db_blkid != DMU_SPILL_BLKID)
2755 ASSERT3P(db->db_blkptr, ==,
2756 &dn->dn_phys->dn_blkptr[db->db_blkid]);
2757 zio = dn->dn_zio;
2758 }
2759
2760 ASSERT(db->db_level == 0 || data == db->db_buf);
2761 ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
2762 ASSERT(zio);
2763
2764 SET_BOOKMARK(&zb, os->os_dsl_dataset ?
2765 os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
2766 db->db.db_object, db->db_level, db->db_blkid);
2767
2768 if (db->db_blkid == DMU_SPILL_BLKID)
2769 wp_flag = WP_SPILL;
2770 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
2771
2772 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
2773 DB_DNODE_EXIT(db);
2774
2775 if (db->db_level == 0 && dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
2776 ASSERT(db->db_state != DB_NOFILL);
2777 dr->dr_zio = zio_write(zio, os->os_spa, txg,
2778 db->db_blkptr, data->b_data, arc_buf_size(data), &zp,
2779 dbuf_write_override_ready, NULL, dbuf_write_override_done,
2780 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
2781 mutex_enter(&db->db_mtx);
2782 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
2783 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
2784 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
2785 mutex_exit(&db->db_mtx);
2786 } else if (db->db_state == DB_NOFILL) {
2787 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF ||
2788 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY);
2789 dr->dr_zio = zio_write(zio, os->os_spa, txg,
2790 db->db_blkptr, NULL, db->db.db_size, &zp,
2791 dbuf_write_nofill_ready, NULL, dbuf_write_nofill_done, db,
2792 ZIO_PRIORITY_ASYNC_WRITE,
2793 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
2794 } else {
2795 ASSERT(arc_released(data));
2796 dr->dr_zio = arc_write(zio, os->os_spa, txg,
2797 db->db_blkptr, data, DBUF_IS_L2CACHEABLE(db),
2798 DBUF_IS_L2COMPRESSIBLE(db), &zp, dbuf_write_ready,
2799 dbuf_write_physdone, dbuf_write_done, db,
2800 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
2801 }
2802 }