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 /*
1338 * Note: This code will probably work even if there are concurrent
1339 * holders, but it is untested in that scenerio, as the ZPL and
1340 * ztest have additional locking (the range locks) that prevents
1341 * that type of concurrent access.
1342 */
1343 ASSERT3U(refcount_count(&db->db_holds), ==, db->db_dirtycnt);
1344
1345 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1346
1347 ASSERT(db->db.db_size != 0);
1348
1349 /*
1350 * Any space we accounted for in dp_dirty_* will be cleaned up by
1351 * dsl_pool_sync(). This is relatively rare so the discrepancy
1352 * is not a big deal.
1353 */
1354
1355 *drp = dr->dr_next;
1356
1357 /*
1358 * Note that there are three places in dbuf_dirty()
1359 * where this dirty record may be put on a list.
1360 * Make sure to do a list_remove corresponding to
1361 * every one of those list_insert calls.
1362 */
1363 if (dr->dr_parent) {
1364 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1365 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1366 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1367 } else if (db->db_blkid == DMU_SPILL_BLKID ||
1368 db->db_level+1 == dn->dn_nlevels) {
1369 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1370 mutex_enter(&dn->dn_mtx);
1371 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1372 mutex_exit(&dn->dn_mtx);
1373 }
1374 DB_DNODE_EXIT(db);
1375
1376 if (db->db_state != DB_NOFILL) {
1377 dbuf_unoverride(dr);
1378
1379 ASSERT(db->db_buf != NULL);
1380 ASSERT(dr->dt.dl.dr_data != NULL);
1381 if (dr->dt.dl.dr_data != db->db_buf)
1382 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db));
1383 }
1384 kmem_free(dr, sizeof (dbuf_dirty_record_t));
1385
1386 ASSERT(db->db_dirtycnt > 0);
1387 db->db_dirtycnt -= 1;
1388
1389 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1390 arc_buf_t *buf = db->db_buf;
1391
1392 ASSERT(db->db_state == DB_NOFILL || arc_released(buf));
1393 dbuf_set_data(db, NULL);
1394 VERIFY(arc_buf_remove_ref(buf, db));
1395 dbuf_evict(db);
1396 return (B_TRUE);
1397 }
1398
1399 return (B_FALSE);
1400 }
1401
1402 #pragma weak dmu_buf_will_dirty = dbuf_will_dirty
1403 void
1404 dbuf_will_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1405 {
1406 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1407
1408 ASSERT(tx->tx_txg != 0);
1409 ASSERT(!refcount_is_zero(&db->db_holds));
1410
1411 DB_DNODE_ENTER(db);
1412 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
1413 rf |= DB_RF_HAVESTRUCT;
1414 DB_DNODE_EXIT(db);
1415 (void) dbuf_read(db, NULL, rf);
1416 (void) dbuf_dirty(db, tx);
1417 }
1418
1419 void
1420 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1421 {
1422 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1423
1424 db->db_state = DB_NOFILL;
1425
1426 dmu_buf_will_fill(db_fake, tx);
1427 }
1428
1429 void
1430 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1431 {
1432 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1433
1434 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1435 ASSERT(tx->tx_txg != 0);
1436 ASSERT(db->db_level == 0);
1437 ASSERT(!refcount_is_zero(&db->db_holds));
1438
1439 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1440 dmu_tx_private_ok(tx));
1441
1442 dbuf_noread(db);
1443 (void) dbuf_dirty(db, tx);
1444 }
1445
1446 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1447 /* ARGSUSED */
1448 void
1449 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1450 {
1451 mutex_enter(&db->db_mtx);
1452 DBUF_VERIFY(db);
1453
1454 if (db->db_state == DB_FILL) {
1455 if (db->db_level == 0 && db->db_freed_in_flight) {
1456 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1457 /* we were freed while filling */
1458 /* XXX dbuf_undirty? */
1459 bzero(db->db.db_data, db->db.db_size);
1460 db->db_freed_in_flight = FALSE;
1461 }
1462 db->db_state = DB_CACHED;
1463 cv_broadcast(&db->db_changed);
1464 }
1465 mutex_exit(&db->db_mtx);
1466 }
1467
1468 /*
1469 * Directly assign a provided arc buf to a given dbuf if it's not referenced
1470 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1471 */
1472 void
1473 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
1474 {
1475 ASSERT(!refcount_is_zero(&db->db_holds));
1476 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1477 ASSERT(db->db_level == 0);
1478 ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA);
1479 ASSERT(buf != NULL);
1480 ASSERT(arc_buf_size(buf) == db->db.db_size);
1481 ASSERT(tx->tx_txg != 0);
1482
1483 arc_return_buf(buf, db);
1484 ASSERT(arc_released(buf));
1485
1486 mutex_enter(&db->db_mtx);
1487
1488 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1489 cv_wait(&db->db_changed, &db->db_mtx);
1490
1491 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
1492
1493 if (db->db_state == DB_CACHED &&
1494 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
1495 mutex_exit(&db->db_mtx);
1496 (void) dbuf_dirty(db, tx);
1497 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
1498 VERIFY(arc_buf_remove_ref(buf, db));
1499 xuio_stat_wbuf_copied();
1500 return;
1501 }
1502
1503 xuio_stat_wbuf_nocopy();
1504 if (db->db_state == DB_CACHED) {
1505 dbuf_dirty_record_t *dr = db->db_last_dirty;
1506
1507 ASSERT(db->db_buf != NULL);
1508 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
1509 ASSERT(dr->dt.dl.dr_data == db->db_buf);
1510 if (!arc_released(db->db_buf)) {
1511 ASSERT(dr->dt.dl.dr_override_state ==
1512 DR_OVERRIDDEN);
1513 arc_release(db->db_buf, db);
1514 }
1515 dr->dt.dl.dr_data = buf;
1516 VERIFY(arc_buf_remove_ref(db->db_buf, db));
1517 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
1518 arc_release(db->db_buf, db);
1519 VERIFY(arc_buf_remove_ref(db->db_buf, db));
1520 }
1521 db->db_buf = NULL;
1522 }
1523 ASSERT(db->db_buf == NULL);
1524 dbuf_set_data(db, buf);
1525 db->db_state = DB_FILL;
1526 mutex_exit(&db->db_mtx);
1527 (void) dbuf_dirty(db, tx);
1528 dbuf_fill_done(db, tx);
1529 }
1530
1531 /*
1532 * "Clear" the contents of this dbuf. This will mark the dbuf
1533 * EVICTING and clear *most* of its references. Unfortunately,
1534 * when we are not holding the dn_dbufs_mtx, we can't clear the
1535 * entry in the dn_dbufs list. We have to wait until dbuf_destroy()
1536 * in this case. For callers from the DMU we will usually see:
1537 * dbuf_clear()->arc_buf_evict()->dbuf_do_evict()->dbuf_destroy()
1538 * For the arc callback, we will usually see:
1539 * dbuf_do_evict()->dbuf_clear();dbuf_destroy()
1540 * Sometimes, though, we will get a mix of these two:
1541 * DMU: dbuf_clear()->arc_buf_evict()
1542 * ARC: dbuf_do_evict()->dbuf_destroy()
1543 */
1544 void
1545 dbuf_clear(dmu_buf_impl_t *db)
1546 {
1547 dnode_t *dn;
1548 dmu_buf_impl_t *parent = db->db_parent;
1549 dmu_buf_impl_t *dndb;
1550 int dbuf_gone = FALSE;
1551
1552 ASSERT(MUTEX_HELD(&db->db_mtx));
1553 ASSERT(refcount_is_zero(&db->db_holds));
1554
1555 dbuf_evict_user(db);
1556
1557 if (db->db_state == DB_CACHED) {
1558 ASSERT(db->db.db_data != NULL);
1559 if (db->db_blkid == DMU_BONUS_BLKID) {
1560 zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
1561 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
1562 }
1563 db->db.db_data = NULL;
1564 db->db_state = DB_UNCACHED;
1565 }
1566
1567 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
1568 ASSERT(db->db_data_pending == NULL);
1569
1570 db->db_state = DB_EVICTING;
1571 db->db_blkptr = NULL;
1572
1573 DB_DNODE_ENTER(db);
1574 dn = DB_DNODE(db);
1575 dndb = dn->dn_dbuf;
1576 if (db->db_blkid != DMU_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) {
1577 list_remove(&dn->dn_dbufs, db);
1578 (void) atomic_dec_32_nv(&dn->dn_dbufs_count);
1579 membar_producer();
1580 DB_DNODE_EXIT(db);
1581 /*
1582 * Decrementing the dbuf count means that the hold corresponding
1583 * to the removed dbuf is no longer discounted in dnode_move(),
1584 * so the dnode cannot be moved until after we release the hold.
1585 * The membar_producer() ensures visibility of the decremented
1586 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
1587 * release any lock.
1588 */
1589 dnode_rele(dn, db);
1590 db->db_dnode_handle = NULL;
1591 } else {
1592 DB_DNODE_EXIT(db);
1593 }
1594
1595 if (db->db_buf)
1596 dbuf_gone = arc_buf_evict(db->db_buf);
1597
1598 if (!dbuf_gone)
1599 mutex_exit(&db->db_mtx);
1600
1601 /*
1602 * If this dbuf is referenced from an indirect dbuf,
1603 * decrement the ref count on the indirect dbuf.
1604 */
1605 if (parent && parent != dndb)
1606 dbuf_rele(parent, db);
1607 }
1608
1609 static int
1610 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
1611 dmu_buf_impl_t **parentp, blkptr_t **bpp)
1612 {
1613 int nlevels, epbs;
1614
1615 *parentp = NULL;
1616 *bpp = NULL;
1617
1618 ASSERT(blkid != DMU_BONUS_BLKID);
1619
1620 if (blkid == DMU_SPILL_BLKID) {
1621 mutex_enter(&dn->dn_mtx);
1622 if (dn->dn_have_spill &&
1623 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
1624 *bpp = &dn->dn_phys->dn_spill;
1625 else
1626 *bpp = NULL;
1627 dbuf_add_ref(dn->dn_dbuf, NULL);
1628 *parentp = dn->dn_dbuf;
1629 mutex_exit(&dn->dn_mtx);
1630 return (0);
1631 }
1632
1633 if (dn->dn_phys->dn_nlevels == 0)
1634 nlevels = 1;
1635 else
1636 nlevels = dn->dn_phys->dn_nlevels;
1637
1638 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1639
1640 ASSERT3U(level * epbs, <, 64);
1641 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1642 if (level >= nlevels ||
1643 (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
1644 /* the buffer has no parent yet */
1645 return (SET_ERROR(ENOENT));
1646 } else if (level < nlevels-1) {
1647 /* this block is referenced from an indirect block */
1648 int err = dbuf_hold_impl(dn, level+1,
1649 blkid >> epbs, fail_sparse, NULL, parentp);
1650 if (err)
1651 return (err);
1652 err = dbuf_read(*parentp, NULL,
1653 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
1654 if (err) {
1655 dbuf_rele(*parentp, NULL);
1656 *parentp = NULL;
1657 return (err);
1658 }
1659 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
1660 (blkid & ((1ULL << epbs) - 1));
1661 return (0);
1662 } else {
1663 /* the block is referenced from the dnode */
1664 ASSERT3U(level, ==, nlevels-1);
1665 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
1666 blkid < dn->dn_phys->dn_nblkptr);
1667 if (dn->dn_dbuf) {
1668 dbuf_add_ref(dn->dn_dbuf, NULL);
1669 *parentp = dn->dn_dbuf;
1670 }
1671 *bpp = &dn->dn_phys->dn_blkptr[blkid];
1672 return (0);
1673 }
1674 }
1675
1676 static dmu_buf_impl_t *
1677 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
1678 dmu_buf_impl_t *parent, blkptr_t *blkptr)
1679 {
1680 objset_t *os = dn->dn_objset;
1681 dmu_buf_impl_t *db, *odb;
1682
1683 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1684 ASSERT(dn->dn_type != DMU_OT_NONE);
1685
1686 db = kmem_cache_alloc(dbuf_cache, KM_SLEEP);
1687
1688 db->db_objset = os;
1689 db->db.db_object = dn->dn_object;
1690 db->db_level = level;
1691 db->db_blkid = blkid;
1692 db->db_last_dirty = NULL;
1693 db->db_dirtycnt = 0;
1694 db->db_dnode_handle = dn->dn_handle;
1695 db->db_parent = parent;
1696 db->db_blkptr = blkptr;
1697
1698 db->db_user_ptr = NULL;
1699 db->db_user_data_ptr_ptr = NULL;
1700 db->db_evict_func = NULL;
1701 db->db_immediate_evict = 0;
1702 db->db_freed_in_flight = 0;
1703
1704 if (blkid == DMU_BONUS_BLKID) {
1705 ASSERT3P(parent, ==, dn->dn_dbuf);
1706 db->db.db_size = DN_MAX_BONUSLEN -
1707 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
1708 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
1709 db->db.db_offset = DMU_BONUS_BLKID;
1710 db->db_state = DB_UNCACHED;
1711 /* the bonus dbuf is not placed in the hash table */
1712 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1713 return (db);
1714 } else if (blkid == DMU_SPILL_BLKID) {
1715 db->db.db_size = (blkptr != NULL) ?
1716 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
1717 db->db.db_offset = 0;
1718 } else {
1719 int blocksize =
1720 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
1721 db->db.db_size = blocksize;
1722 db->db.db_offset = db->db_blkid * blocksize;
1723 }
1724
1725 /*
1726 * Hold the dn_dbufs_mtx while we get the new dbuf
1727 * in the hash table *and* added to the dbufs list.
1728 * This prevents a possible deadlock with someone
1729 * trying to look up this dbuf before its added to the
1730 * dn_dbufs list.
1731 */
1732 mutex_enter(&dn->dn_dbufs_mtx);
1733 db->db_state = DB_EVICTING;
1734 if ((odb = dbuf_hash_insert(db)) != NULL) {
1735 /* someone else inserted it first */
1736 kmem_cache_free(dbuf_cache, db);
1737 mutex_exit(&dn->dn_dbufs_mtx);
1738 return (odb);
1739 }
1740 list_insert_head(&dn->dn_dbufs, db);
1741 if (db->db_level == 0 && db->db_blkid >=
1742 dn->dn_unlisted_l0_blkid)
1743 dn->dn_unlisted_l0_blkid = db->db_blkid + 1;
1744 db->db_state = DB_UNCACHED;
1745 mutex_exit(&dn->dn_dbufs_mtx);
1746 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1747
1748 if (parent && parent != dn->dn_dbuf)
1749 dbuf_add_ref(parent, db);
1750
1751 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1752 refcount_count(&dn->dn_holds) > 0);
1753 (void) refcount_add(&dn->dn_holds, db);
1754 (void) atomic_inc_32_nv(&dn->dn_dbufs_count);
1755
1756 dprintf_dbuf(db, "db=%p\n", db);
1757
1758 return (db);
1759 }
1760
1761 static int
1762 dbuf_do_evict(void *private)
1763 {
1764 arc_buf_t *buf = private;
1765 dmu_buf_impl_t *db = buf->b_private;
1766
1767 if (!MUTEX_HELD(&db->db_mtx))
1768 mutex_enter(&db->db_mtx);
1769
1770 ASSERT(refcount_is_zero(&db->db_holds));
1771
1772 if (db->db_state != DB_EVICTING) {
1773 ASSERT(db->db_state == DB_CACHED);
1774 DBUF_VERIFY(db);
1775 db->db_buf = NULL;
1776 dbuf_evict(db);
1777 } else {
1778 mutex_exit(&db->db_mtx);
1779 dbuf_destroy(db);
1780 }
1781 return (0);
1782 }
1783
1784 static void
1785 dbuf_destroy(dmu_buf_impl_t *db)
1786 {
1787 ASSERT(refcount_is_zero(&db->db_holds));
1788
1789 if (db->db_blkid != DMU_BONUS_BLKID) {
1790 /*
1791 * If this dbuf is still on the dn_dbufs list,
1792 * remove it from that list.
1793 */
1794 if (db->db_dnode_handle != NULL) {
1795 dnode_t *dn;
1796
1797 DB_DNODE_ENTER(db);
1798 dn = DB_DNODE(db);
1799 mutex_enter(&dn->dn_dbufs_mtx);
1800 list_remove(&dn->dn_dbufs, db);
1801 (void) atomic_dec_32_nv(&dn->dn_dbufs_count);
1802 mutex_exit(&dn->dn_dbufs_mtx);
1803 DB_DNODE_EXIT(db);
1804 /*
1805 * Decrementing the dbuf count means that the hold
1806 * corresponding to the removed dbuf is no longer
1807 * discounted in dnode_move(), so the dnode cannot be
1808 * moved until after we release the hold.
1809 */
1810 dnode_rele(dn, db);
1811 db->db_dnode_handle = NULL;
1812 }
1813 dbuf_hash_remove(db);
1814 }
1815 db->db_parent = NULL;
1816 db->db_buf = NULL;
1817
1818 ASSERT(!list_link_active(&db->db_link));
1819 ASSERT(db->db.db_data == NULL);
1820 ASSERT(db->db_hash_next == NULL);
1821 ASSERT(db->db_blkptr == NULL);
1822 ASSERT(db->db_data_pending == NULL);
1823
1824 kmem_cache_free(dbuf_cache, db);
1825 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1826 }
1827
1828 void
1829 dbuf_prefetch(dnode_t *dn, uint64_t blkid, zio_priority_t prio)
1830 {
1831 dmu_buf_impl_t *db = NULL;
1832 blkptr_t *bp = NULL;
1833
1834 ASSERT(blkid != DMU_BONUS_BLKID);
1835 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1836
1837 if (dnode_block_freed(dn, blkid))
1838 return;
1839
1840 /* dbuf_find() returns with db_mtx held */
1841 if (db = dbuf_find(dn, 0, blkid)) {
1842 /*
1843 * This dbuf is already in the cache. We assume that
1844 * it is already CACHED, or else about to be either
1845 * read or filled.
1846 */
1847 mutex_exit(&db->db_mtx);
1848 return;
1849 }
1850
1851 if (dbuf_findbp(dn, 0, blkid, TRUE, &db, &bp) == 0) {
1852 if (bp && !BP_IS_HOLE(bp)) {
1853 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
1854 uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH;
1855 zbookmark_t zb;
1856
1857 SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
1858 dn->dn_object, 0, blkid);
1859
1860 (void) arc_read(NULL, dn->dn_objset->os_spa,
1861 bp, NULL, NULL, prio,
1862 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1863 &aflags, &zb);
1864 }
1865 if (db)
1866 dbuf_rele(db, NULL);
1867 }
1868 }
1869
1870 /*
1871 * Returns with db_holds incremented, and db_mtx not held.
1872 * Note: dn_struct_rwlock must be held.
1873 */
1874 int
1875 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse,
1876 void *tag, dmu_buf_impl_t **dbp)
1877 {
1878 dmu_buf_impl_t *db, *parent = NULL;
1879
1880 ASSERT(blkid != DMU_BONUS_BLKID);
1881 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1882 ASSERT3U(dn->dn_nlevels, >, level);
1883
1884 *dbp = NULL;
1885 top:
1886 /* dbuf_find() returns with db_mtx held */
1887 db = dbuf_find(dn, level, blkid);
1888
1889 if (db == NULL) {
1890 blkptr_t *bp = NULL;
1891 int err;
1892
1893 ASSERT3P(parent, ==, NULL);
1894 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
1895 if (fail_sparse) {
1896 if (err == 0 && bp && BP_IS_HOLE(bp))
1897 err = SET_ERROR(ENOENT);
1898 if (err) {
1899 if (parent)
1900 dbuf_rele(parent, NULL);
1901 return (err);
1902 }
1903 }
1904 if (err && err != ENOENT)
1905 return (err);
1906 db = dbuf_create(dn, level, blkid, parent, bp);
1907 }
1908
1909 if (db->db_buf && refcount_is_zero(&db->db_holds)) {
1910 arc_buf_add_ref(db->db_buf, db);
1911 if (db->db_buf->b_data == NULL) {
1912 dbuf_clear(db);
1913 if (parent) {
1914 dbuf_rele(parent, NULL);
1915 parent = NULL;
1916 }
1917 goto top;
1918 }
1919 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
1920 }
1921
1922 ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
1923
1924 /*
1925 * If this buffer is currently syncing out, and we are are
1926 * still referencing it from db_data, we need to make a copy
1927 * of it in case we decide we want to dirty it again in this txg.
1928 */
1929 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1930 dn->dn_object != DMU_META_DNODE_OBJECT &&
1931 db->db_state == DB_CACHED && db->db_data_pending) {
1932 dbuf_dirty_record_t *dr = db->db_data_pending;
1933
1934 if (dr->dt.dl.dr_data == db->db_buf) {
1935 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1936
1937 dbuf_set_data(db,
1938 arc_buf_alloc(dn->dn_objset->os_spa,
1939 db->db.db_size, db, type));
1940 bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
1941 db->db.db_size);
1942 }
1943 }
1944
1945 (void) refcount_add(&db->db_holds, tag);
1946 dbuf_update_data(db);
1947 DBUF_VERIFY(db);
1948 mutex_exit(&db->db_mtx);
1949
1950 /* NOTE: we can't rele the parent until after we drop the db_mtx */
1951 if (parent)
1952 dbuf_rele(parent, NULL);
1953
1954 ASSERT3P(DB_DNODE(db), ==, dn);
1955 ASSERT3U(db->db_blkid, ==, blkid);
1956 ASSERT3U(db->db_level, ==, level);
1957 *dbp = db;
1958
1959 return (0);
1960 }
1961
1962 dmu_buf_impl_t *
1963 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
1964 {
1965 dmu_buf_impl_t *db;
1966 int err = dbuf_hold_impl(dn, 0, blkid, FALSE, tag, &db);
1967 return (err ? NULL : db);
1968 }
1969
1970 dmu_buf_impl_t *
1971 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
1972 {
1973 dmu_buf_impl_t *db;
1974 int err = dbuf_hold_impl(dn, level, blkid, FALSE, tag, &db);
1975 return (err ? NULL : db);
1976 }
1977
1978 void
1979 dbuf_create_bonus(dnode_t *dn)
1980 {
1981 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1982
1983 ASSERT(dn->dn_bonus == NULL);
1984 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
1985 }
1986
1987 int
1988 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
1989 {
1990 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1991 dnode_t *dn;
1992
1993 if (db->db_blkid != DMU_SPILL_BLKID)
1994 return (SET_ERROR(ENOTSUP));
1995 if (blksz == 0)
1996 blksz = SPA_MINBLOCKSIZE;
1997 if (blksz > SPA_MAXBLOCKSIZE)
1998 blksz = SPA_MAXBLOCKSIZE;
1999 else
2000 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
2001
2002 DB_DNODE_ENTER(db);
2003 dn = DB_DNODE(db);
2004 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2005 dbuf_new_size(db, blksz, tx);
2006 rw_exit(&dn->dn_struct_rwlock);
2007 DB_DNODE_EXIT(db);
2008
2009 return (0);
2010 }
2011
2012 void
2013 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
2014 {
2015 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
2016 }
2017
2018 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2019 void
2020 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
2021 {
2022 int64_t holds = refcount_add(&db->db_holds, tag);
2023 ASSERT(holds > 1);
2024 }
2025
2026 /*
2027 * If you call dbuf_rele() you had better not be referencing the dnode handle
2028 * unless you have some other direct or indirect hold on the dnode. (An indirect
2029 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2030 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2031 * dnode's parent dbuf evicting its dnode handles.
2032 */
2033 #pragma weak dmu_buf_rele = dbuf_rele
2034 void
2035 dbuf_rele(dmu_buf_impl_t *db, void *tag)
2036 {
2037 mutex_enter(&db->db_mtx);
2038 dbuf_rele_and_unlock(db, tag);
2039 }
2040
2041 /*
2042 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2043 * db_dirtycnt and db_holds to be updated atomically.
2044 */
2045 void
2046 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2047 {
2048 int64_t holds;
2049
2050 ASSERT(MUTEX_HELD(&db->db_mtx));
2051 DBUF_VERIFY(db);
2052
2053 /*
2054 * Remove the reference to the dbuf before removing its hold on the
2055 * dnode so we can guarantee in dnode_move() that a referenced bonus
2056 * buffer has a corresponding dnode hold.
2057 */
2058 holds = refcount_remove(&db->db_holds, tag);
2059 ASSERT(holds >= 0);
2060
2061 /*
2062 * We can't freeze indirects if there is a possibility that they
2063 * may be modified in the current syncing context.
2064 */
2065 if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0))
2066 arc_buf_freeze(db->db_buf);
2067
2068 if (holds == db->db_dirtycnt &&
2069 db->db_level == 0 && db->db_immediate_evict)
2070 dbuf_evict_user(db);
2071
2072 if (holds == 0) {
2073 if (db->db_blkid == DMU_BONUS_BLKID) {
2074 mutex_exit(&db->db_mtx);
2075
2076 /*
2077 * If the dnode moves here, we cannot cross this barrier
2078 * until the move completes.
2079 */
2080 DB_DNODE_ENTER(db);
2081 (void) atomic_dec_32_nv(&DB_DNODE(db)->dn_dbufs_count);
2082 DB_DNODE_EXIT(db);
2083 /*
2084 * The bonus buffer's dnode hold is no longer discounted
2085 * in dnode_move(). The dnode cannot move until after
2086 * the dnode_rele().
2087 */
2088 dnode_rele(DB_DNODE(db), db);
2089 } else if (db->db_buf == NULL) {
2090 /*
2091 * This is a special case: we never associated this
2092 * dbuf with any data allocated from the ARC.
2093 */
2094 ASSERT(db->db_state == DB_UNCACHED ||
2095 db->db_state == DB_NOFILL);
2096 dbuf_evict(db);
2097 } else if (arc_released(db->db_buf)) {
2098 arc_buf_t *buf = db->db_buf;
2099 /*
2100 * This dbuf has anonymous data associated with it.
2101 */
2102 dbuf_set_data(db, NULL);
2103 VERIFY(arc_buf_remove_ref(buf, db));
2104 dbuf_evict(db);
2105 } else {
2106 VERIFY(!arc_buf_remove_ref(db->db_buf, db));
2107
2108 /*
2109 * A dbuf will be eligible for eviction if either the
2110 * 'primarycache' property is set or a duplicate
2111 * copy of this buffer is already cached in the arc.
2112 *
2113 * In the case of the 'primarycache' a buffer
2114 * is considered for eviction if it matches the
2115 * criteria set in the property.
2116 *
2117 * To decide if our buffer is considered a
2118 * duplicate, we must call into the arc to determine
2119 * if multiple buffers are referencing the same
2120 * block on-disk. If so, then we simply evict
2121 * ourselves.
2122 */
2123 if (!DBUF_IS_CACHEABLE(db) ||
2124 arc_buf_eviction_needed(db->db_buf))
2125 dbuf_clear(db);
2126 else
2127 mutex_exit(&db->db_mtx);
2128 }
2129 } else {
2130 mutex_exit(&db->db_mtx);
2131 }
2132 }
2133
2134 #pragma weak dmu_buf_refcount = dbuf_refcount
2135 uint64_t
2136 dbuf_refcount(dmu_buf_impl_t *db)
2137 {
2138 return (refcount_count(&db->db_holds));
2139 }
2140
2141 void *
2142 dmu_buf_set_user(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr,
2143 dmu_buf_evict_func_t *evict_func)
2144 {
2145 return (dmu_buf_update_user(db_fake, NULL, user_ptr,
2146 user_data_ptr_ptr, evict_func));
2147 }
2148
2149 void *
2150 dmu_buf_set_user_ie(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr,
2151 dmu_buf_evict_func_t *evict_func)
2152 {
2153 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2154
2155 db->db_immediate_evict = TRUE;
2156 return (dmu_buf_update_user(db_fake, NULL, user_ptr,
2157 user_data_ptr_ptr, evict_func));
2158 }
2159
2160 void *
2161 dmu_buf_update_user(dmu_buf_t *db_fake, void *old_user_ptr, void *user_ptr,
2162 void *user_data_ptr_ptr, dmu_buf_evict_func_t *evict_func)
2163 {
2164 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2165 ASSERT(db->db_level == 0);
2166
2167 ASSERT((user_ptr == NULL) == (evict_func == NULL));
2168
2169 mutex_enter(&db->db_mtx);
2170
2171 if (db->db_user_ptr == old_user_ptr) {
2172 db->db_user_ptr = user_ptr;
2173 db->db_user_data_ptr_ptr = user_data_ptr_ptr;
2174 db->db_evict_func = evict_func;
2175
2176 dbuf_update_data(db);
2177 } else {
2178 old_user_ptr = db->db_user_ptr;
2179 }
2180
2181 mutex_exit(&db->db_mtx);
2182 return (old_user_ptr);
2183 }
2184
2185 void *
2186 dmu_buf_get_user(dmu_buf_t *db_fake)
2187 {
2188 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2189 ASSERT(!refcount_is_zero(&db->db_holds));
2190
2191 return (db->db_user_ptr);
2192 }
2193
2194 boolean_t
2195 dmu_buf_freeable(dmu_buf_t *dbuf)
2196 {
2197 boolean_t res = B_FALSE;
2198 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2199
2200 if (db->db_blkptr)
2201 res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset,
2202 db->db_blkptr, db->db_blkptr->blk_birth);
2203
2204 return (res);
2205 }
2206
2207 blkptr_t *
2208 dmu_buf_get_blkptr(dmu_buf_t *db)
2209 {
2210 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2211 return (dbi->db_blkptr);
2212 }
2213
2214 static void
2215 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
2216 {
2217 /* ASSERT(dmu_tx_is_syncing(tx) */
2218 ASSERT(MUTEX_HELD(&db->db_mtx));
2219
2220 if (db->db_blkptr != NULL)
2221 return;
2222
2223 if (db->db_blkid == DMU_SPILL_BLKID) {
2224 db->db_blkptr = &dn->dn_phys->dn_spill;
2225 BP_ZERO(db->db_blkptr);
2226 return;
2227 }
2228 if (db->db_level == dn->dn_phys->dn_nlevels-1) {
2229 /*
2230 * This buffer was allocated at a time when there was
2231 * no available blkptrs from the dnode, or it was
2232 * inappropriate to hook it in (i.e., nlevels mis-match).
2233 */
2234 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
2235 ASSERT(db->db_parent == NULL);
2236 db->db_parent = dn->dn_dbuf;
2237 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
2238 DBUF_VERIFY(db);
2239 } else {
2240 dmu_buf_impl_t *parent = db->db_parent;
2241 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2242
2243 ASSERT(dn->dn_phys->dn_nlevels > 1);
2244 if (parent == NULL) {
2245 mutex_exit(&db->db_mtx);
2246 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2247 (void) dbuf_hold_impl(dn, db->db_level+1,
2248 db->db_blkid >> epbs, FALSE, db, &parent);
2249 rw_exit(&dn->dn_struct_rwlock);
2250 mutex_enter(&db->db_mtx);
2251 db->db_parent = parent;
2252 }
2253 db->db_blkptr = (blkptr_t *)parent->db.db_data +
2254 (db->db_blkid & ((1ULL << epbs) - 1));
2255 DBUF_VERIFY(db);
2256 }
2257 }
2258
2259 static void
2260 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2261 {
2262 dmu_buf_impl_t *db = dr->dr_dbuf;
2263 dnode_t *dn;
2264 zio_t *zio;
2265
2266 ASSERT(dmu_tx_is_syncing(tx));
2267
2268 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2269
2270 mutex_enter(&db->db_mtx);
2271
2272 ASSERT(db->db_level > 0);
2273 DBUF_VERIFY(db);
2274
2275 /* Read the block if it hasn't been read yet. */
2276 if (db->db_buf == NULL) {
2277 mutex_exit(&db->db_mtx);
2278 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
2279 mutex_enter(&db->db_mtx);
2280 }
2281 ASSERT3U(db->db_state, ==, DB_CACHED);
2282 ASSERT(db->db_buf != NULL);
2283
2284 DB_DNODE_ENTER(db);
2285 dn = DB_DNODE(db);
2286 /* Indirect block size must match what the dnode thinks it is. */
2287 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2288 dbuf_check_blkptr(dn, db);
2289 DB_DNODE_EXIT(db);
2290
2291 /* Provide the pending dirty record to child dbufs */
2292 db->db_data_pending = dr;
2293
2294 mutex_exit(&db->db_mtx);
2295 dbuf_write(dr, db->db_buf, tx);
2296
2297 zio = dr->dr_zio;
2298 mutex_enter(&dr->dt.di.dr_mtx);
2299 dbuf_sync_list(&dr->dt.di.dr_children, tx);
2300 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2301 mutex_exit(&dr->dt.di.dr_mtx);
2302 zio_nowait(zio);
2303 }
2304
2305 static void
2306 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2307 {
2308 arc_buf_t **datap = &dr->dt.dl.dr_data;
2309 dmu_buf_impl_t *db = dr->dr_dbuf;
2310 dnode_t *dn;
2311 objset_t *os;
2312 uint64_t txg = tx->tx_txg;
2313
2314 ASSERT(dmu_tx_is_syncing(tx));
2315
2316 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2317
2318 mutex_enter(&db->db_mtx);
2319 /*
2320 * To be synced, we must be dirtied. But we
2321 * might have been freed after the dirty.
2322 */
2323 if (db->db_state == DB_UNCACHED) {
2324 /* This buffer has been freed since it was dirtied */
2325 ASSERT(db->db.db_data == NULL);
2326 } else if (db->db_state == DB_FILL) {
2327 /* This buffer was freed and is now being re-filled */
2328 ASSERT(db->db.db_data != dr->dt.dl.dr_data);
2329 } else {
2330 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
2331 }
2332 DBUF_VERIFY(db);
2333
2334 DB_DNODE_ENTER(db);
2335 dn = DB_DNODE(db);
2336
2337 if (db->db_blkid == DMU_SPILL_BLKID) {
2338 mutex_enter(&dn->dn_mtx);
2339 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
2340 mutex_exit(&dn->dn_mtx);
2341 }
2342
2343 /*
2344 * If this is a bonus buffer, simply copy the bonus data into the
2345 * dnode. It will be written out when the dnode is synced (and it
2346 * will be synced, since it must have been dirty for dbuf_sync to
2347 * be called).
2348 */
2349 if (db->db_blkid == DMU_BONUS_BLKID) {
2350 dbuf_dirty_record_t **drp;
2351
2352 ASSERT(*datap != NULL);
2353 ASSERT0(db->db_level);
2354 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN);
2355 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen);
2356 DB_DNODE_EXIT(db);
2357
2358 if (*datap != db->db.db_data) {
2359 zio_buf_free(*datap, DN_MAX_BONUSLEN);
2360 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
2361 }
2362 db->db_data_pending = NULL;
2363 drp = &db->db_last_dirty;
2364 while (*drp != dr)
2365 drp = &(*drp)->dr_next;
2366 ASSERT(dr->dr_next == NULL);
2367 ASSERT(dr->dr_dbuf == db);
2368 *drp = dr->dr_next;
2369 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2370 ASSERT(db->db_dirtycnt > 0);
2371 db->db_dirtycnt -= 1;
2372 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
2373 return;
2374 }
2375
2376 os = dn->dn_objset;
2377
2378 /*
2379 * This function may have dropped the db_mtx lock allowing a dmu_sync
2380 * operation to sneak in. As a result, we need to ensure that we
2381 * don't check the dr_override_state until we have returned from
2382 * dbuf_check_blkptr.
2383 */
2384 dbuf_check_blkptr(dn, db);
2385
2386 /*
2387 * If this buffer is in the middle of an immediate write,
2388 * wait for the synchronous IO to complete.
2389 */
2390 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
2391 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
2392 cv_wait(&db->db_changed, &db->db_mtx);
2393 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
2394 }
2395
2396 if (db->db_state != DB_NOFILL &&
2397 dn->dn_object != DMU_META_DNODE_OBJECT &&
2398 refcount_count(&db->db_holds) > 1 &&
2399 dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
2400 *datap == db->db_buf) {
2401 /*
2402 * If this buffer is currently "in use" (i.e., there
2403 * are active holds and db_data still references it),
2404 * then make a copy before we start the write so that
2405 * any modifications from the open txg will not leak
2406 * into this write.
2407 *
2408 * NOTE: this copy does not need to be made for
2409 * objects only modified in the syncing context (e.g.
2410 * DNONE_DNODE blocks).
2411 */
2412 int blksz = arc_buf_size(*datap);
2413 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2414 *datap = arc_buf_alloc(os->os_spa, blksz, db, type);
2415 bcopy(db->db.db_data, (*datap)->b_data, blksz);
2416 }
2417 db->db_data_pending = dr;
2418
2419 mutex_exit(&db->db_mtx);
2420
2421 dbuf_write(dr, *datap, tx);
2422
2423 ASSERT(!list_link_active(&dr->dr_dirty_node));
2424 if (dn->dn_object == DMU_META_DNODE_OBJECT) {
2425 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
2426 DB_DNODE_EXIT(db);
2427 } else {
2428 /*
2429 * Although zio_nowait() does not "wait for an IO", it does
2430 * initiate the IO. If this is an empty write it seems plausible
2431 * that the IO could actually be completed before the nowait
2432 * returns. We need to DB_DNODE_EXIT() first in case
2433 * zio_nowait() invalidates the dbuf.
2434 */
2435 DB_DNODE_EXIT(db);
2436 zio_nowait(dr->dr_zio);
2437 }
2438 }
2439
2440 void
2441 dbuf_sync_list(list_t *list, dmu_tx_t *tx)
2442 {
2443 dbuf_dirty_record_t *dr;
2444
2445 while (dr = list_head(list)) {
2446 if (dr->dr_zio != NULL) {
2447 /*
2448 * If we find an already initialized zio then we
2449 * are processing the meta-dnode, and we have finished.
2450 * The dbufs for all dnodes are put back on the list
2451 * during processing, so that we can zio_wait()
2452 * these IOs after initiating all child IOs.
2453 */
2454 ASSERT3U(dr->dr_dbuf->db.db_object, ==,
2455 DMU_META_DNODE_OBJECT);
2456 break;
2457 }
2458 list_remove(list, dr);
2459 if (dr->dr_dbuf->db_level > 0)
2460 dbuf_sync_indirect(dr, tx);
2461 else
2462 dbuf_sync_leaf(dr, tx);
2463 }
2464 }
2465
2466 /* ARGSUSED */
2467 static void
2468 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
2469 {
2470 dmu_buf_impl_t *db = vdb;
2471 dnode_t *dn;
2472 blkptr_t *bp = zio->io_bp;
2473 blkptr_t *bp_orig = &zio->io_bp_orig;
2474 spa_t *spa = zio->io_spa;
2475 int64_t delta;
2476 uint64_t fill = 0;
2477 int i;
2478
2479 ASSERT(db->db_blkptr == bp);
2480
2481 DB_DNODE_ENTER(db);
2482 dn = DB_DNODE(db);
2483 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
2484 dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
2485 zio->io_prev_space_delta = delta;
2486
2487 if (BP_IS_HOLE(bp)) {
2488 ASSERT(bp->blk_fill == 0);
2489 DB_DNODE_EXIT(db);
2490 return;
2491 }
2492
2493 ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
2494 BP_GET_TYPE(bp) == dn->dn_type) ||
2495 (db->db_blkid == DMU_SPILL_BLKID &&
2496 BP_GET_TYPE(bp) == dn->dn_bonustype));
2497 ASSERT(BP_GET_LEVEL(bp) == db->db_level);
2498
2499 mutex_enter(&db->db_mtx);
2500
2501 #ifdef ZFS_DEBUG
2502 if (db->db_blkid == DMU_SPILL_BLKID) {
2503 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2504 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2505 db->db_blkptr == &dn->dn_phys->dn_spill);
2506 }
2507 #endif
2508
2509 if (db->db_level == 0) {
2510 mutex_enter(&dn->dn_mtx);
2511 if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
2512 db->db_blkid != DMU_SPILL_BLKID)
2513 dn->dn_phys->dn_maxblkid = db->db_blkid;
2514 mutex_exit(&dn->dn_mtx);
2515
2516 if (dn->dn_type == DMU_OT_DNODE) {
2517 dnode_phys_t *dnp = db->db.db_data;
2518 for (i = db->db.db_size >> DNODE_SHIFT; i > 0;
2519 i--, dnp++) {
2520 if (dnp->dn_type != DMU_OT_NONE)
2521 fill++;
2522 }
2523 } else {
2524 fill = 1;
2525 }
2526 } else {
2527 blkptr_t *ibp = db->db.db_data;
2528 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2529 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
2530 if (BP_IS_HOLE(ibp))
2531 continue;
2532 fill += ibp->blk_fill;
2533 }
2534 }
2535 DB_DNODE_EXIT(db);
2536
2537 bp->blk_fill = fill;
2538
2539 mutex_exit(&db->db_mtx);
2540 }
2541
2542 /*
2543 * The SPA will call this callback several times for each zio - once
2544 * for every physical child i/o (zio->io_phys_children times). This
2545 * allows the DMU to monitor the progress of each logical i/o. For example,
2546 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
2547 * block. There may be a long delay before all copies/fragments are completed,
2548 * so this callback allows us to retire dirty space gradually, as the physical
2549 * i/os complete.
2550 */
2551 /* ARGSUSED */
2552 static void
2553 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
2554 {
2555 dmu_buf_impl_t *db = arg;
2556 objset_t *os = db->db_objset;
2557 dsl_pool_t *dp = dmu_objset_pool(os);
2558 dbuf_dirty_record_t *dr;
2559 int delta = 0;
2560
2561 dr = db->db_data_pending;
2562 ASSERT3U(dr->dr_txg, ==, zio->io_txg);
2563
2564 /*
2565 * The callback will be called io_phys_children times. Retire one
2566 * portion of our dirty space each time we are called. Any rounding
2567 * error will be cleaned up by dsl_pool_sync()'s call to
2568 * dsl_pool_undirty_space().
2569 */
2570 delta = dr->dr_accounted / zio->io_phys_children;
2571 dsl_pool_undirty_space(dp, delta, zio->io_txg);
2572 }
2573
2574 /* ARGSUSED */
2575 static void
2576 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
2577 {
2578 dmu_buf_impl_t *db = vdb;
2579 blkptr_t *bp = zio->io_bp;
2580 blkptr_t *bp_orig = &zio->io_bp_orig;
2581 uint64_t txg = zio->io_txg;
2582 dbuf_dirty_record_t **drp, *dr;
2583
2584 ASSERT0(zio->io_error);
2585 ASSERT(db->db_blkptr == bp);
2586
2587 /*
2588 * For nopwrites and rewrites we ensure that the bp matches our
2589 * original and bypass all the accounting.
2590 */
2591 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
2592 ASSERT(BP_EQUAL(bp, bp_orig));
2593 } else {
2594 objset_t *os;
2595 dsl_dataset_t *ds;
2596 dmu_tx_t *tx;
2597
2598 DB_GET_OBJSET(&os, db);
2599 ds = os->os_dsl_dataset;
2600 tx = os->os_synctx;
2601
2602 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
2603 dsl_dataset_block_born(ds, bp, tx);
2604 }
2605
2606 mutex_enter(&db->db_mtx);
2607
2608 DBUF_VERIFY(db);
2609
2610 drp = &db->db_last_dirty;
2611 while ((dr = *drp) != db->db_data_pending)
2612 drp = &dr->dr_next;
2613 ASSERT(!list_link_active(&dr->dr_dirty_node));
2614 ASSERT(dr->dr_txg == txg);
2615 ASSERT(dr->dr_dbuf == db);
2616 ASSERT(dr->dr_next == NULL);
2617 *drp = dr->dr_next;
2618
2619 #ifdef ZFS_DEBUG
2620 if (db->db_blkid == DMU_SPILL_BLKID) {
2621 dnode_t *dn;
2622
2623 DB_DNODE_ENTER(db);
2624 dn = DB_DNODE(db);
2625 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2626 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2627 db->db_blkptr == &dn->dn_phys->dn_spill);
2628 DB_DNODE_EXIT(db);
2629 }
2630 #endif
2631
2632 if (db->db_level == 0) {
2633 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2634 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
2635 if (db->db_state != DB_NOFILL) {
2636 if (dr->dt.dl.dr_data != db->db_buf)
2637 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data,
2638 db));
2639 else if (!arc_released(db->db_buf))
2640 arc_set_callback(db->db_buf, dbuf_do_evict, db);
2641 }
2642 } else {
2643 dnode_t *dn;
2644
2645 DB_DNODE_ENTER(db);
2646 dn = DB_DNODE(db);
2647 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2648 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2649 if (!BP_IS_HOLE(db->db_blkptr)) {
2650 int epbs =
2651 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2652 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
2653 db->db.db_size);
2654 ASSERT3U(dn->dn_phys->dn_maxblkid
2655 >> (db->db_level * epbs), >=, db->db_blkid);
2656 arc_set_callback(db->db_buf, dbuf_do_evict, db);
2657 }
2658 DB_DNODE_EXIT(db);
2659 mutex_destroy(&dr->dt.di.dr_mtx);
2660 list_destroy(&dr->dt.di.dr_children);
2661 }
2662 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2663
2664 cv_broadcast(&db->db_changed);
2665 ASSERT(db->db_dirtycnt > 0);
2666 db->db_dirtycnt -= 1;
2667 db->db_data_pending = NULL;
2668
2669 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
2670 }
2671
2672 static void
2673 dbuf_write_nofill_ready(zio_t *zio)
2674 {
2675 dbuf_write_ready(zio, NULL, zio->io_private);
2676 }
2677
2678 static void
2679 dbuf_write_nofill_done(zio_t *zio)
2680 {
2681 dbuf_write_done(zio, NULL, zio->io_private);
2682 }
2683
2684 static void
2685 dbuf_write_override_ready(zio_t *zio)
2686 {
2687 dbuf_dirty_record_t *dr = zio->io_private;
2688 dmu_buf_impl_t *db = dr->dr_dbuf;
2689
2690 dbuf_write_ready(zio, NULL, db);
2691 }
2692
2693 static void
2694 dbuf_write_override_done(zio_t *zio)
2695 {
2696 dbuf_dirty_record_t *dr = zio->io_private;
2697 dmu_buf_impl_t *db = dr->dr_dbuf;
2698 blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
2699
2700 mutex_enter(&db->db_mtx);
2701 if (!BP_EQUAL(zio->io_bp, obp)) {
2702 if (!BP_IS_HOLE(obp))
2703 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
2704 arc_release(dr->dt.dl.dr_data, db);
2705 }
2706 mutex_exit(&db->db_mtx);
2707
2708 dbuf_write_done(zio, NULL, db);
2709 }
2710
2711 /* Issue I/O to commit a dirty buffer to disk. */
2712 static void
2713 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
2714 {
2715 dmu_buf_impl_t *db = dr->dr_dbuf;
2716 dnode_t *dn;
2717 objset_t *os;
2718 dmu_buf_impl_t *parent = db->db_parent;
2719 uint64_t txg = tx->tx_txg;
2720 zbookmark_t zb;
2721 zio_prop_t zp;
2722 zio_t *zio;
2723 int wp_flag = 0;
2724
2725 DB_DNODE_ENTER(db);
2726 dn = DB_DNODE(db);
2727 os = dn->dn_objset;
2728
2729 if (db->db_state != DB_NOFILL) {
2730 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
2731 /*
2732 * Private object buffers are released here rather
2733 * than in dbuf_dirty() since they are only modified
2734 * in the syncing context and we don't want the
2735 * overhead of making multiple copies of the data.
2736 */
2737 if (BP_IS_HOLE(db->db_blkptr)) {
2738 arc_buf_thaw(data);
2739 } else {
2740 dbuf_release_bp(db);
2741 }
2742 }
2743 }
2744
2745 if (parent != dn->dn_dbuf) {
2746 /* Our parent is an indirect block. */
2747 /* We have a dirty parent that has been scheduled for write. */
2748 ASSERT(parent && parent->db_data_pending);
2749 /* Our parent's buffer is one level closer to the dnode. */
2750 ASSERT(db->db_level == parent->db_level-1);
2751 /*
2752 * We're about to modify our parent's db_data by modifying
2753 * our block pointer, so the parent must be released.
2754 */
2755 ASSERT(arc_released(parent->db_buf));
2756 zio = parent->db_data_pending->dr_zio;
2757 } else {
2758 /* Our parent is the dnode itself. */
2759 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
2760 db->db_blkid != DMU_SPILL_BLKID) ||
2761 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
2762 if (db->db_blkid != DMU_SPILL_BLKID)
2763 ASSERT3P(db->db_blkptr, ==,
2764 &dn->dn_phys->dn_blkptr[db->db_blkid]);
2765 zio = dn->dn_zio;
2766 }
2767
2768 ASSERT(db->db_level == 0 || data == db->db_buf);
2769 ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
2770 ASSERT(zio);
2771
2772 SET_BOOKMARK(&zb, os->os_dsl_dataset ?
2773 os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
2774 db->db.db_object, db->db_level, db->db_blkid);
2775
2776 if (db->db_blkid == DMU_SPILL_BLKID)
2777 wp_flag = WP_SPILL;
2778 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
2779
2780 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
2781 DB_DNODE_EXIT(db);
2782
2783 if (db->db_level == 0 && dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
2784 ASSERT(db->db_state != DB_NOFILL);
2785 dr->dr_zio = zio_write(zio, os->os_spa, txg,
2786 db->db_blkptr, data->b_data, arc_buf_size(data), &zp,
2787 dbuf_write_override_ready, NULL, dbuf_write_override_done,
2788 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
2789 mutex_enter(&db->db_mtx);
2790 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
2791 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
2792 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
2793 mutex_exit(&db->db_mtx);
2794 } else if (db->db_state == DB_NOFILL) {
2795 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF ||
2796 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY);
2797 dr->dr_zio = zio_write(zio, os->os_spa, txg,
2798 db->db_blkptr, NULL, db->db.db_size, &zp,
2799 dbuf_write_nofill_ready, NULL, dbuf_write_nofill_done, db,
2800 ZIO_PRIORITY_ASYNC_WRITE,
2801 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
2802 } else {
2803 ASSERT(arc_released(data));
2804 dr->dr_zio = arc_write(zio, os->os_spa, txg,
2805 db->db_blkptr, data, DBUF_IS_L2CACHEABLE(db),
2806 DBUF_IS_L2COMPRESSIBLE(db), &zp, dbuf_write_ready,
2807 dbuf_write_physdone, dbuf_write_done, db,
2808 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
2809 }
2810 }