1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright 2011 Nexenta Systems, Inc. All rights reserved.
24 * Copyright (c) 2012, 2014 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 #include <sys/range_tree.h>
44
45 /*
46 * Number of times that zfs_free_range() took the slow path while doing
47 * a zfs receive. A nonzero value indicates a potential performance problem.
48 */
49 uint64_t zfs_free_range_recv_miss;
50
51 static void dbuf_destroy(dmu_buf_impl_t *db);
52 static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
53 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
54
55 /*
56 * Global data structures and functions for the dbuf cache.
57 */
58 static kmem_cache_t *dbuf_cache;
59
60 /* ARGSUSED */
61 static int
62 dbuf_cons(void *vdb, void *unused, int kmflag)
63 {
64 dmu_buf_impl_t *db = vdb;
65 bzero(db, sizeof (dmu_buf_impl_t));
66
67 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
68 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
69 refcount_create(&db->db_holds);
70 return (0);
71 }
72
73 /* ARGSUSED */
74 static void
75 dbuf_dest(void *vdb, void *unused)
76 {
77 dmu_buf_impl_t *db = vdb;
78 mutex_destroy(&db->db_mtx);
79 cv_destroy(&db->db_changed);
80 refcount_destroy(&db->db_holds);
81 }
82
83 /*
84 * dbuf hash table routines
85 */
86 static dbuf_hash_table_t dbuf_hash_table;
87
88 static uint64_t dbuf_hash_count;
89
90 static uint64_t
91 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
92 {
93 uintptr_t osv = (uintptr_t)os;
94 uint64_t crc = -1ULL;
95
96 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
97 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF];
98 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF];
99 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF];
100 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF];
101 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF];
102 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF];
103
104 crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16);
105
106 return (crc);
107 }
108
109 #define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid);
110
111 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
112 ((dbuf)->db.db_object == (obj) && \
113 (dbuf)->db_objset == (os) && \
114 (dbuf)->db_level == (level) && \
115 (dbuf)->db_blkid == (blkid))
116
117 dmu_buf_impl_t *
118 dbuf_find(dnode_t *dn, uint8_t level, uint64_t blkid)
119 {
120 dbuf_hash_table_t *h = &dbuf_hash_table;
121 objset_t *os = dn->dn_objset;
122 uint64_t obj = dn->dn_object;
123 uint64_t hv = DBUF_HASH(os, obj, level, blkid);
124 uint64_t idx = hv & h->hash_table_mask;
125 dmu_buf_impl_t *db;
126
127 mutex_enter(DBUF_HASH_MUTEX(h, idx));
128 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
129 if (DBUF_EQUAL(db, os, obj, level, blkid)) {
130 mutex_enter(&db->db_mtx);
131 if (db->db_state != DB_EVICTING) {
132 mutex_exit(DBUF_HASH_MUTEX(h, idx));
133 return (db);
134 }
135 mutex_exit(&db->db_mtx);
136 }
137 }
138 mutex_exit(DBUF_HASH_MUTEX(h, idx));
139 return (NULL);
140 }
141
142 /*
143 * Insert an entry into the hash table. If there is already an element
144 * equal to elem in the hash table, then the already existing element
145 * will be returned and the new element will not be inserted.
146 * Otherwise returns NULL.
147 */
148 static dmu_buf_impl_t *
149 dbuf_hash_insert(dmu_buf_impl_t *db)
150 {
151 dbuf_hash_table_t *h = &dbuf_hash_table;
152 objset_t *os = db->db_objset;
153 uint64_t obj = db->db.db_object;
154 int level = db->db_level;
155 uint64_t blkid = db->db_blkid;
156 uint64_t hv = DBUF_HASH(os, obj, level, blkid);
157 uint64_t idx = hv & h->hash_table_mask;
158 dmu_buf_impl_t *dbf;
159
160 mutex_enter(DBUF_HASH_MUTEX(h, idx));
161 for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) {
162 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
163 mutex_enter(&dbf->db_mtx);
164 if (dbf->db_state != DB_EVICTING) {
165 mutex_exit(DBUF_HASH_MUTEX(h, idx));
166 return (dbf);
167 }
168 mutex_exit(&dbf->db_mtx);
169 }
170 }
171
172 mutex_enter(&db->db_mtx);
173 db->db_hash_next = h->hash_table[idx];
174 h->hash_table[idx] = db;
175 mutex_exit(DBUF_HASH_MUTEX(h, idx));
176 atomic_add_64(&dbuf_hash_count, 1);
177
178 return (NULL);
179 }
180
181 /*
182 * Remove an entry from the hash table. This operation will
183 * fail if there are any existing holds on the db.
184 */
185 static void
186 dbuf_hash_remove(dmu_buf_impl_t *db)
187 {
188 dbuf_hash_table_t *h = &dbuf_hash_table;
189 uint64_t hv = DBUF_HASH(db->db_objset, db->db.db_object,
190 db->db_level, db->db_blkid);
191 uint64_t idx = hv & h->hash_table_mask;
192 dmu_buf_impl_t *dbf, **dbp;
193
194 /*
195 * We musn't hold db_mtx to maintin lock ordering:
196 * DBUF_HASH_MUTEX > db_mtx.
197 */
198 ASSERT(refcount_is_zero(&db->db_holds));
199 ASSERT(db->db_state == DB_EVICTING);
200 ASSERT(!MUTEX_HELD(&db->db_mtx));
201
202 mutex_enter(DBUF_HASH_MUTEX(h, idx));
203 dbp = &h->hash_table[idx];
204 while ((dbf = *dbp) != db) {
205 dbp = &dbf->db_hash_next;
206 ASSERT(dbf != NULL);
207 }
208 *dbp = db->db_hash_next;
209 db->db_hash_next = NULL;
210 mutex_exit(DBUF_HASH_MUTEX(h, idx));
211 atomic_add_64(&dbuf_hash_count, -1);
212 }
213
214 static arc_evict_func_t dbuf_do_evict;
215
216 static void
217 dbuf_evict_user(dmu_buf_impl_t *db)
218 {
219 ASSERT(MUTEX_HELD(&db->db_mtx));
220
221 if (db->db_level != 0 || db->db_evict_func == NULL)
222 return;
223
224 if (db->db_user_data_ptr_ptr)
225 *db->db_user_data_ptr_ptr = db->db.db_data;
226 db->db_evict_func(&db->db, db->db_user_ptr);
227 db->db_user_ptr = NULL;
228 db->db_user_data_ptr_ptr = NULL;
229 db->db_evict_func = NULL;
230 }
231
232 boolean_t
233 dbuf_is_metadata(dmu_buf_impl_t *db)
234 {
235 if (db->db_level > 0) {
236 return (B_TRUE);
237 } else {
238 boolean_t is_metadata;
239
240 DB_DNODE_ENTER(db);
241 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type);
242 DB_DNODE_EXIT(db);
243
244 return (is_metadata);
245 }
246 }
247
248 void
249 dbuf_evict(dmu_buf_impl_t *db)
250 {
251 ASSERT(MUTEX_HELD(&db->db_mtx));
252 ASSERT(db->db_buf == NULL);
253 ASSERT(db->db_data_pending == NULL);
254
255 dbuf_clear(db);
256 dbuf_destroy(db);
257 }
258
259 void
260 dbuf_init(void)
261 {
262 uint64_t hsize = 1ULL << 16;
263 dbuf_hash_table_t *h = &dbuf_hash_table;
264 int i;
265
266 /*
267 * The hash table is big enough to fill all of physical memory
268 * with an average 4K block size. The table will take up
269 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
270 */
271 while (hsize * 4096 < physmem * PAGESIZE)
272 hsize <<= 1;
273
274 retry:
275 h->hash_table_mask = hsize - 1;
276 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
277 if (h->hash_table == NULL) {
278 /* XXX - we should really return an error instead of assert */
279 ASSERT(hsize > (1ULL << 10));
280 hsize >>= 1;
281 goto retry;
282 }
283
284 dbuf_cache = kmem_cache_create("dmu_buf_impl_t",
285 sizeof (dmu_buf_impl_t),
286 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
287
288 for (i = 0; i < DBUF_MUTEXES; i++)
289 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
290 }
291
292 void
293 dbuf_fini(void)
294 {
295 dbuf_hash_table_t *h = &dbuf_hash_table;
296 int i;
297
298 for (i = 0; i < DBUF_MUTEXES; i++)
299 mutex_destroy(&h->hash_mutexes[i]);
300 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
301 kmem_cache_destroy(dbuf_cache);
302 }
303
304 /*
305 * Other stuff.
306 */
307
308 #ifdef ZFS_DEBUG
309 static void
310 dbuf_verify(dmu_buf_impl_t *db)
311 {
312 dnode_t *dn;
313 dbuf_dirty_record_t *dr;
314
315 ASSERT(MUTEX_HELD(&db->db_mtx));
316
317 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
318 return;
319
320 ASSERT(db->db_objset != NULL);
321 DB_DNODE_ENTER(db);
322 dn = DB_DNODE(db);
323 if (dn == NULL) {
324 ASSERT(db->db_parent == NULL);
325 ASSERT(db->db_blkptr == NULL);
326 } else {
327 ASSERT3U(db->db.db_object, ==, dn->dn_object);
328 ASSERT3P(db->db_objset, ==, dn->dn_objset);
329 ASSERT3U(db->db_level, <, dn->dn_nlevels);
330 ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
331 db->db_blkid == DMU_SPILL_BLKID ||
332 !list_is_empty(&dn->dn_dbufs));
333 }
334 if (db->db_blkid == DMU_BONUS_BLKID) {
335 ASSERT(dn != NULL);
336 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
337 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
338 } else if (db->db_blkid == DMU_SPILL_BLKID) {
339 ASSERT(dn != NULL);
340 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
341 ASSERT0(db->db.db_offset);
342 } else {
343 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
344 }
345
346 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
347 ASSERT(dr->dr_dbuf == db);
348
349 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
350 ASSERT(dr->dr_dbuf == db);
351
352 /*
353 * We can't assert that db_size matches dn_datablksz because it
354 * can be momentarily different when another thread is doing
355 * dnode_set_blksz().
356 */
357 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
358 dr = db->db_data_pending;
359 /*
360 * It should only be modified in syncing context, so
361 * make sure we only have one copy of the data.
362 */
363 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
364 }
365
366 /* verify db->db_blkptr */
367 if (db->db_blkptr) {
368 if (db->db_parent == dn->dn_dbuf) {
369 /* db is pointed to by the dnode */
370 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
371 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
372 ASSERT(db->db_parent == NULL);
373 else
374 ASSERT(db->db_parent != NULL);
375 if (db->db_blkid != DMU_SPILL_BLKID)
376 ASSERT3P(db->db_blkptr, ==,
377 &dn->dn_phys->dn_blkptr[db->db_blkid]);
378 } else {
379 /* db is pointed to by an indirect block */
380 int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
381 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
382 ASSERT3U(db->db_parent->db.db_object, ==,
383 db->db.db_object);
384 /*
385 * dnode_grow_indblksz() can make this fail if we don't
386 * have the struct_rwlock. XXX indblksz no longer
387 * grows. safe to do this now?
388 */
389 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
390 ASSERT3P(db->db_blkptr, ==,
391 ((blkptr_t *)db->db_parent->db.db_data +
392 db->db_blkid % epb));
393 }
394 }
395 }
396 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
397 (db->db_buf == NULL || db->db_buf->b_data) &&
398 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
399 db->db_state != DB_FILL && !dn->dn_free_txg) {
400 /*
401 * If the blkptr isn't set but they have nonzero data,
402 * it had better be dirty, otherwise we'll lose that
403 * data when we evict this buffer.
404 */
405 if (db->db_dirtycnt == 0) {
406 uint64_t *buf = db->db.db_data;
407 int i;
408
409 for (i = 0; i < db->db.db_size >> 3; i++) {
410 ASSERT(buf[i] == 0);
411 }
412 }
413 }
414 DB_DNODE_EXIT(db);
415 }
416 #endif
417
418 static void
419 dbuf_update_data(dmu_buf_impl_t *db)
420 {
421 ASSERT(MUTEX_HELD(&db->db_mtx));
422 if (db->db_level == 0 && db->db_user_data_ptr_ptr) {
423 ASSERT(!refcount_is_zero(&db->db_holds));
424 *db->db_user_data_ptr_ptr = db->db.db_data;
425 }
426 }
427
428 static void
429 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
430 {
431 ASSERT(MUTEX_HELD(&db->db_mtx));
432 ASSERT(db->db_buf == NULL || !arc_has_callback(db->db_buf));
433 db->db_buf = buf;
434 if (buf != NULL) {
435 ASSERT(buf->b_data != NULL);
436 db->db.db_data = buf->b_data;
437 if (!arc_released(buf))
438 arc_set_callback(buf, dbuf_do_evict, db);
439 dbuf_update_data(db);
440 } else {
441 dbuf_evict_user(db);
442 db->db.db_data = NULL;
443 if (db->db_state != DB_NOFILL)
444 db->db_state = DB_UNCACHED;
445 }
446 }
447
448 /*
449 * Loan out an arc_buf for read. Return the loaned arc_buf.
450 */
451 arc_buf_t *
452 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
453 {
454 arc_buf_t *abuf;
455
456 mutex_enter(&db->db_mtx);
457 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
458 int blksz = db->db.db_size;
459 spa_t *spa = db->db_objset->os_spa;
460
461 mutex_exit(&db->db_mtx);
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 zbookmark_t zb;
523 uint32_t aflags = ARC_NOWAIT;
524
525 DB_DNODE_ENTER(db);
526 dn = DB_DNODE(db);
527 ASSERT(!refcount_is_zero(&db->db_holds));
528 /* We need the struct_rwlock to prevent db_blkptr from changing. */
529 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
530 ASSERT(MUTEX_HELD(&db->db_mtx));
531 ASSERT(db->db_state == DB_UNCACHED);
532 ASSERT(db->db_buf == NULL);
533
534 if (db->db_blkid == DMU_BONUS_BLKID) {
535 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
536
537 ASSERT3U(bonuslen, <=, db->db.db_size);
538 db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN);
539 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
540 if (bonuslen < DN_MAX_BONUSLEN)
541 bzero(db->db.db_data, DN_MAX_BONUSLEN);
542 if (bonuslen)
543 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
544 DB_DNODE_EXIT(db);
545 dbuf_update_data(db);
546 db->db_state = DB_CACHED;
547 mutex_exit(&db->db_mtx);
548 return;
549 }
550
551 /*
552 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
553 * processes the delete record and clears the bp while we are waiting
554 * for the dn_mtx (resulting in a "no" from block_freed).
555 */
556 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
557 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
558 BP_IS_HOLE(db->db_blkptr)))) {
559 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
560
561 DB_DNODE_EXIT(db);
562 dbuf_set_data(db, arc_buf_alloc(db->db_objset->os_spa,
563 db->db.db_size, db, type));
564 bzero(db->db.db_data, db->db.db_size);
565 db->db_state = DB_CACHED;
566 *flags |= DB_RF_CACHED;
567 mutex_exit(&db->db_mtx);
568 return;
569 }
570
571 DB_DNODE_EXIT(db);
572
573 db->db_state = DB_READ;
574 mutex_exit(&db->db_mtx);
575
576 if (DBUF_IS_L2CACHEABLE(db))
577 aflags |= ARC_L2CACHE;
578 if (DBUF_IS_L2COMPRESSIBLE(db))
579 aflags |= ARC_L2COMPRESS;
580
581 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
582 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
583 db->db.db_object, db->db_level, db->db_blkid);
584
585 dbuf_add_ref(db, NULL);
586
587 (void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr,
588 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
589 (*flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
590 &aflags, &zb);
591 if (aflags & ARC_CACHED)
592 *flags |= DB_RF_CACHED;
593 }
594
595 int
596 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
597 {
598 int err = 0;
599 boolean_t havepzio = (zio != NULL);
600 boolean_t prefetch;
601 dnode_t *dn;
602
603 /*
604 * We don't have to hold the mutex to check db_state because it
605 * can't be freed while we have a hold on the buffer.
606 */
607 ASSERT(!refcount_is_zero(&db->db_holds));
608
609 if (db->db_state == DB_NOFILL)
610 return (SET_ERROR(EIO));
611
612 DB_DNODE_ENTER(db);
613 dn = DB_DNODE(db);
614 if ((flags & DB_RF_HAVESTRUCT) == 0)
615 rw_enter(&dn->dn_struct_rwlock, RW_READER);
616
617 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
618 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
619 DBUF_IS_CACHEABLE(db);
620
621 mutex_enter(&db->db_mtx);
622 if (db->db_state == DB_CACHED) {
623 mutex_exit(&db->db_mtx);
624 if (prefetch)
625 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
626 db->db.db_size, TRUE);
627 if ((flags & DB_RF_HAVESTRUCT) == 0)
628 rw_exit(&dn->dn_struct_rwlock);
629 DB_DNODE_EXIT(db);
630 } else if (db->db_state == DB_UNCACHED) {
631 spa_t *spa = dn->dn_objset->os_spa;
632
633 if (zio == NULL)
634 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
635 dbuf_read_impl(db, zio, &flags);
636
637 /* dbuf_read_impl has dropped db_mtx for us */
638
639 if (prefetch)
640 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
641 db->db.db_size, flags & DB_RF_CACHED);
642
643 if ((flags & DB_RF_HAVESTRUCT) == 0)
644 rw_exit(&dn->dn_struct_rwlock);
645 DB_DNODE_EXIT(db);
646
647 if (!havepzio)
648 err = zio_wait(zio);
649 } else {
650 /*
651 * Another reader came in while the dbuf was in flight
652 * between UNCACHED and CACHED. Either a writer will finish
653 * writing the buffer (sending the dbuf to CACHED) or the
654 * first reader's request will reach the read_done callback
655 * and send the dbuf to CACHED. Otherwise, a failure
656 * occurred and the dbuf went to UNCACHED.
657 */
658 mutex_exit(&db->db_mtx);
659 if (prefetch)
660 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
661 db->db.db_size, TRUE);
662 if ((flags & DB_RF_HAVESTRUCT) == 0)
663 rw_exit(&dn->dn_struct_rwlock);
664 DB_DNODE_EXIT(db);
665
666 /* Skip the wait per the caller's request. */
667 mutex_enter(&db->db_mtx);
668 if ((flags & DB_RF_NEVERWAIT) == 0) {
669 while (db->db_state == DB_READ ||
670 db->db_state == DB_FILL) {
671 ASSERT(db->db_state == DB_READ ||
672 (flags & DB_RF_HAVESTRUCT) == 0);
673 cv_wait(&db->db_changed, &db->db_mtx);
674 }
675 if (db->db_state == DB_UNCACHED)
676 err = SET_ERROR(EIO);
677 }
678 mutex_exit(&db->db_mtx);
679 }
680
681 ASSERT(err || havepzio || db->db_state == DB_CACHED);
682 return (err);
683 }
684
685 static void
686 dbuf_noread(dmu_buf_impl_t *db)
687 {
688 ASSERT(!refcount_is_zero(&db->db_holds));
689 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
690 mutex_enter(&db->db_mtx);
691 while (db->db_state == DB_READ || db->db_state == DB_FILL)
692 cv_wait(&db->db_changed, &db->db_mtx);
693 if (db->db_state == DB_UNCACHED) {
694 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
695 spa_t *spa = db->db_objset->os_spa;
696
697 ASSERT(db->db_buf == NULL);
698 ASSERT(db->db.db_data == NULL);
699 dbuf_set_data(db, arc_buf_alloc(spa, db->db.db_size, db, type));
700 db->db_state = DB_FILL;
701 } else if (db->db_state == DB_NOFILL) {
702 dbuf_set_data(db, NULL);
703 } else {
704 ASSERT3U(db->db_state, ==, DB_CACHED);
705 }
706 mutex_exit(&db->db_mtx);
707 }
708
709 /*
710 * This is our just-in-time copy function. It makes a copy of
711 * buffers, that have been modified in a previous transaction
712 * group, before we modify them in the current active group.
713 *
714 * This function is used in two places: when we are dirtying a
715 * buffer for the first time in a txg, and when we are freeing
716 * a range in a dnode that includes this buffer.
717 *
718 * Note that when we are called from dbuf_free_range() we do
719 * not put a hold on the buffer, we just traverse the active
720 * dbuf list for the dnode.
721 */
722 static void
723 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
724 {
725 dbuf_dirty_record_t *dr = db->db_last_dirty;
726
727 ASSERT(MUTEX_HELD(&db->db_mtx));
728 ASSERT(db->db.db_data != NULL);
729 ASSERT(db->db_level == 0);
730 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
731
732 if (dr == NULL ||
733 (dr->dt.dl.dr_data !=
734 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
735 return;
736
737 /*
738 * If the last dirty record for this dbuf has not yet synced
739 * and its referencing the dbuf data, either:
740 * reset the reference to point to a new copy,
741 * or (if there a no active holders)
742 * just null out the current db_data pointer.
743 */
744 ASSERT(dr->dr_txg >= txg - 2);
745 if (db->db_blkid == DMU_BONUS_BLKID) {
746 /* Note that the data bufs here are zio_bufs */
747 dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
748 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
749 bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
750 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
751 int size = db->db.db_size;
752 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
753 spa_t *spa = db->db_objset->os_spa;
754
755 dr->dt.dl.dr_data = arc_buf_alloc(spa, size, db, type);
756 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
757 } else {
758 dbuf_set_data(db, NULL);
759 }
760 }
761
762 void
763 dbuf_unoverride(dbuf_dirty_record_t *dr)
764 {
765 dmu_buf_impl_t *db = dr->dr_dbuf;
766 blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
767 uint64_t txg = dr->dr_txg;
768
769 ASSERT(MUTEX_HELD(&db->db_mtx));
770 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
771 ASSERT(db->db_level == 0);
772
773 if (db->db_blkid == DMU_BONUS_BLKID ||
774 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
775 return;
776
777 ASSERT(db->db_data_pending != dr);
778
779 /* free this block */
780 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite)
781 zio_free(db->db_objset->os_spa, txg, bp);
782
783 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
784 dr->dt.dl.dr_nopwrite = B_FALSE;
785
786 /*
787 * Release the already-written buffer, so we leave it in
788 * a consistent dirty state. Note that all callers are
789 * modifying the buffer, so they will immediately do
790 * another (redundant) arc_release(). Therefore, leave
791 * the buf thawed to save the effort of freezing &
792 * immediately re-thawing it.
793 */
794 arc_release(dr->dt.dl.dr_data, db);
795 }
796
797 /*
798 * Evict (if its unreferenced) or clear (if its referenced) any level-0
799 * data blocks in the free range, so that any future readers will find
800 * empty blocks.
801 *
802 * This is a no-op if the dataset is in the middle of an incremental
803 * receive; see comment below for details.
804 */
805 void
806 dbuf_free_range(dnode_t *dn, uint64_t start, uint64_t end, dmu_tx_t *tx)
807 {
808 dmu_buf_impl_t *db, *db_next;
809 uint64_t txg = tx->tx_txg;
810
811 if (end > dn->dn_maxblkid && (end != DMU_SPILL_BLKID))
812 end = dn->dn_maxblkid;
813 dprintf_dnode(dn, "start=%llu end=%llu\n", start, end);
814
815 mutex_enter(&dn->dn_dbufs_mtx);
816 if (start >= dn->dn_unlisted_l0_blkid * dn->dn_datablksz) {
817 /* There can't be any dbufs in this range; no need to search. */
818 mutex_exit(&dn->dn_dbufs_mtx);
819 return;
820 } else if (dmu_objset_is_receiving(dn->dn_objset)) {
821 /*
822 * If we are receiving, we expect there to be no dbufs in
823 * the range to be freed, because receive modifies each
824 * block at most once, and in offset order. If this is
825 * not the case, it can lead to performance problems,
826 * so note that we unexpectedly took the slow path.
827 */
828 atomic_inc_64(&zfs_free_range_recv_miss);
829 }
830
831 for (db = list_head(&dn->dn_dbufs); db != NULL; db = db_next) {
832 db_next = list_next(&dn->dn_dbufs, db);
833 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
834
835 if (db->db_level != 0)
836 continue;
837 if (db->db_blkid < start || db->db_blkid > end)
838 continue;
839
840 /* found a level 0 buffer in the range */
841 mutex_enter(&db->db_mtx);
842 if (dbuf_undirty(db, tx)) {
843 /* mutex has been dropped and dbuf destroyed */
844 continue;
845 }
846
847 if (db->db_state == DB_UNCACHED ||
848 db->db_state == DB_NOFILL ||
849 db->db_state == DB_EVICTING) {
850 ASSERT(db->db.db_data == NULL);
851 mutex_exit(&db->db_mtx);
852 continue;
853 }
854 if (db->db_state == DB_READ || db->db_state == DB_FILL) {
855 /* will be handled in dbuf_read_done or dbuf_rele */
856 db->db_freed_in_flight = TRUE;
857 mutex_exit(&db->db_mtx);
858 continue;
859 }
860 if (refcount_count(&db->db_holds) == 0) {
861 ASSERT(db->db_buf);
862 dbuf_clear(db);
863 continue;
864 }
865 /* The dbuf is referenced */
866
867 if (db->db_last_dirty != NULL) {
868 dbuf_dirty_record_t *dr = db->db_last_dirty;
869
870 if (dr->dr_txg == txg) {
871 /*
872 * This buffer is "in-use", re-adjust the file
873 * size to reflect that this buffer may
874 * contain new data when we sync.
875 */
876 if (db->db_blkid != DMU_SPILL_BLKID &&
877 db->db_blkid > dn->dn_maxblkid)
878 dn->dn_maxblkid = db->db_blkid;
879 dbuf_unoverride(dr);
880 } else {
881 /*
882 * This dbuf is not dirty in the open context.
883 * Either uncache it (if its not referenced in
884 * the open context) or reset its contents to
885 * empty.
886 */
887 dbuf_fix_old_data(db, txg);
888 }
889 }
890 /* clear the contents if its cached */
891 if (db->db_state == DB_CACHED) {
892 ASSERT(db->db.db_data != NULL);
893 arc_release(db->db_buf, db);
894 bzero(db->db.db_data, db->db.db_size);
895 arc_buf_freeze(db->db_buf);
896 }
897
898 mutex_exit(&db->db_mtx);
899 }
900 mutex_exit(&dn->dn_dbufs_mtx);
901 }
902
903 static int
904 dbuf_block_freeable(dmu_buf_impl_t *db)
905 {
906 dsl_dataset_t *ds = db->db_objset->os_dsl_dataset;
907 uint64_t birth_txg = 0;
908
909 /*
910 * We don't need any locking to protect db_blkptr:
911 * If it's syncing, then db_last_dirty will be set
912 * so we'll ignore db_blkptr.
913 *
914 * This logic ensures that only block births for
915 * filled blocks are considered.
916 */
917 ASSERT(MUTEX_HELD(&db->db_mtx));
918 if (db->db_last_dirty && (db->db_blkptr == NULL ||
919 !BP_IS_HOLE(db->db_blkptr))) {
920 birth_txg = db->db_last_dirty->dr_txg;
921 } else if (db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) {
922 birth_txg = db->db_blkptr->blk_birth;
923 }
924
925 /*
926 * If this block don't exist or is in a snapshot, it can't be freed.
927 * Don't pass the bp to dsl_dataset_block_freeable() since we
928 * are holding the db_mtx lock and might deadlock if we are
929 * prefetching a dedup-ed block.
930 */
931 if (birth_txg != 0)
932 return (ds == NULL ||
933 dsl_dataset_block_freeable(ds, NULL, birth_txg));
934 else
935 return (B_FALSE);
936 }
937
938 void
939 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
940 {
941 arc_buf_t *buf, *obuf;
942 int osize = db->db.db_size;
943 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
944 dnode_t *dn;
945
946 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
947
948 DB_DNODE_ENTER(db);
949 dn = DB_DNODE(db);
950
951 /* XXX does *this* func really need the lock? */
952 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
953
954 /*
955 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
956 * is OK, because there can be no other references to the db
957 * when we are changing its size, so no concurrent DB_FILL can
958 * be happening.
959 */
960 /*
961 * XXX we should be doing a dbuf_read, checking the return
962 * value and returning that up to our callers
963 */
964 dmu_buf_will_dirty(&db->db, tx);
965
966 /* create the data buffer for the new block */
967 buf = arc_buf_alloc(dn->dn_objset->os_spa, size, db, type);
968
969 /* copy old block data to the new block */
970 obuf = db->db_buf;
971 bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
972 /* zero the remainder */
973 if (size > osize)
974 bzero((uint8_t *)buf->b_data + osize, size - osize);
975
976 mutex_enter(&db->db_mtx);
977 dbuf_set_data(db, buf);
978 VERIFY(arc_buf_remove_ref(obuf, db));
979 db->db.db_size = size;
980
981 if (db->db_level == 0) {
982 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
983 db->db_last_dirty->dt.dl.dr_data = buf;
984 }
985 mutex_exit(&db->db_mtx);
986
987 dnode_willuse_space(dn, size-osize, tx);
988 DB_DNODE_EXIT(db);
989 }
990
991 void
992 dbuf_release_bp(dmu_buf_impl_t *db)
993 {
994 objset_t *os = db->db_objset;
995
996 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
997 ASSERT(arc_released(os->os_phys_buf) ||
998 list_link_active(&os->os_dsl_dataset->ds_synced_link));
999 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1000
1001 (void) arc_release(db->db_buf, db);
1002 }
1003
1004 dbuf_dirty_record_t *
1005 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1006 {
1007 dnode_t *dn;
1008 objset_t *os;
1009 dbuf_dirty_record_t **drp, *dr;
1010 int drop_struct_lock = FALSE;
1011 boolean_t do_free_accounting = B_FALSE;
1012 int txgoff = tx->tx_txg & TXG_MASK;
1013
1014 ASSERT(tx->tx_txg != 0);
1015 ASSERT(!refcount_is_zero(&db->db_holds));
1016 DMU_TX_DIRTY_BUF(tx, db);
1017
1018 DB_DNODE_ENTER(db);
1019 dn = DB_DNODE(db);
1020 /*
1021 * Shouldn't dirty a regular buffer in syncing context. Private
1022 * objects may be dirtied in syncing context, but only if they
1023 * were already pre-dirtied in open context.
1024 */
1025 ASSERT(!dmu_tx_is_syncing(tx) ||
1026 BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1027 DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1028 dn->dn_objset->os_dsl_dataset == NULL);
1029 /*
1030 * We make this assert for private objects as well, but after we
1031 * check if we're already dirty. They are allowed to re-dirty
1032 * in syncing context.
1033 */
1034 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1035 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1036 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1037
1038 mutex_enter(&db->db_mtx);
1039 /*
1040 * XXX make this true for indirects too? The problem is that
1041 * transactions created with dmu_tx_create_assigned() from
1042 * syncing context don't bother holding ahead.
1043 */
1044 ASSERT(db->db_level != 0 ||
1045 db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1046 db->db_state == DB_NOFILL);
1047
1048 mutex_enter(&dn->dn_mtx);
1049 /*
1050 * Don't set dirtyctx to SYNC if we're just modifying this as we
1051 * initialize the objset.
1052 */
1053 if (dn->dn_dirtyctx == DN_UNDIRTIED &&
1054 !BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1055 dn->dn_dirtyctx =
1056 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1057 ASSERT(dn->dn_dirtyctx_firstset == NULL);
1058 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1059 }
1060 mutex_exit(&dn->dn_mtx);
1061
1062 if (db->db_blkid == DMU_SPILL_BLKID)
1063 dn->dn_have_spill = B_TRUE;
1064
1065 /*
1066 * If this buffer is already dirty, we're done.
1067 */
1068 drp = &db->db_last_dirty;
1069 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1070 db->db.db_object == DMU_META_DNODE_OBJECT);
1071 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1072 drp = &dr->dr_next;
1073 if (dr && dr->dr_txg == tx->tx_txg) {
1074 DB_DNODE_EXIT(db);
1075
1076 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1077 /*
1078 * If this buffer has already been written out,
1079 * we now need to reset its state.
1080 */
1081 dbuf_unoverride(dr);
1082 if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1083 db->db_state != DB_NOFILL)
1084 arc_buf_thaw(db->db_buf);
1085 }
1086 mutex_exit(&db->db_mtx);
1087 return (dr);
1088 }
1089
1090 /*
1091 * Only valid if not already dirty.
1092 */
1093 ASSERT(dn->dn_object == 0 ||
1094 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1095 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1096
1097 ASSERT3U(dn->dn_nlevels, >, db->db_level);
1098 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1099 dn->dn_phys->dn_nlevels > db->db_level ||
1100 dn->dn_next_nlevels[txgoff] > db->db_level ||
1101 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1102 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1103
1104 /*
1105 * We should only be dirtying in syncing context if it's the
1106 * mos or we're initializing the os or it's a special object.
1107 * However, we are allowed to dirty in syncing context provided
1108 * we already dirtied it in open context. Hence we must make
1109 * this assertion only if we're not already dirty.
1110 */
1111 os = dn->dn_objset;
1112 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1113 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1114 ASSERT(db->db.db_size != 0);
1115
1116 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1117
1118 if (db->db_blkid != DMU_BONUS_BLKID) {
1119 /*
1120 * Update the accounting.
1121 * Note: we delay "free accounting" until after we drop
1122 * the db_mtx. This keeps us from grabbing other locks
1123 * (and possibly deadlocking) in bp_get_dsize() while
1124 * also holding the db_mtx.
1125 */
1126 dnode_willuse_space(dn, db->db.db_size, tx);
1127 do_free_accounting = dbuf_block_freeable(db);
1128 }
1129
1130 /*
1131 * If this buffer is dirty in an old transaction group we need
1132 * to make a copy of it so that the changes we make in this
1133 * transaction group won't leak out when we sync the older txg.
1134 */
1135 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1136 if (db->db_level == 0) {
1137 void *data_old = db->db_buf;
1138
1139 if (db->db_state != DB_NOFILL) {
1140 if (db->db_blkid == DMU_BONUS_BLKID) {
1141 dbuf_fix_old_data(db, tx->tx_txg);
1142 data_old = db->db.db_data;
1143 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1144 /*
1145 * Release the data buffer from the cache so
1146 * that we can modify it without impacting
1147 * possible other users of this cached data
1148 * block. Note that indirect blocks and
1149 * private objects are not released until the
1150 * syncing state (since they are only modified
1151 * then).
1152 */
1153 arc_release(db->db_buf, db);
1154 dbuf_fix_old_data(db, tx->tx_txg);
1155 data_old = db->db_buf;
1156 }
1157 ASSERT(data_old != NULL);
1158 }
1159 dr->dt.dl.dr_data = data_old;
1160 } else {
1161 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1162 list_create(&dr->dt.di.dr_children,
1163 sizeof (dbuf_dirty_record_t),
1164 offsetof(dbuf_dirty_record_t, dr_dirty_node));
1165 }
1166 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL)
1167 dr->dr_accounted = db->db.db_size;
1168 dr->dr_dbuf = db;
1169 dr->dr_txg = tx->tx_txg;
1170 dr->dr_next = *drp;
1171 *drp = dr;
1172
1173 /*
1174 * We could have been freed_in_flight between the dbuf_noread
1175 * and dbuf_dirty. We win, as though the dbuf_noread() had
1176 * happened after the free.
1177 */
1178 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1179 db->db_blkid != DMU_SPILL_BLKID) {
1180 mutex_enter(&dn->dn_mtx);
1181 if (dn->dn_free_ranges[txgoff] != NULL) {
1182 range_tree_clear(dn->dn_free_ranges[txgoff],
1183 db->db_blkid, 1);
1184 }
1185 mutex_exit(&dn->dn_mtx);
1186 db->db_freed_in_flight = FALSE;
1187 }
1188
1189 /*
1190 * This buffer is now part of this txg
1191 */
1192 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1193 db->db_dirtycnt += 1;
1194 ASSERT3U(db->db_dirtycnt, <=, 3);
1195
1196 mutex_exit(&db->db_mtx);
1197
1198 if (db->db_blkid == DMU_BONUS_BLKID ||
1199 db->db_blkid == DMU_SPILL_BLKID) {
1200 mutex_enter(&dn->dn_mtx);
1201 ASSERT(!list_link_active(&dr->dr_dirty_node));
1202 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1203 mutex_exit(&dn->dn_mtx);
1204 dnode_setdirty(dn, tx);
1205 DB_DNODE_EXIT(db);
1206 return (dr);
1207 } else if (do_free_accounting) {
1208 blkptr_t *bp = db->db_blkptr;
1209 int64_t willfree = (bp && !BP_IS_HOLE(bp)) ?
1210 bp_get_dsize(os->os_spa, bp) : db->db.db_size;
1211 /*
1212 * This is only a guess -- if the dbuf is dirty
1213 * in a previous txg, we don't know how much
1214 * space it will use on disk yet. We should
1215 * really have the struct_rwlock to access
1216 * db_blkptr, but since this is just a guess,
1217 * it's OK if we get an odd answer.
1218 */
1219 ddt_prefetch(os->os_spa, bp);
1220 dnode_willuse_space(dn, -willfree, tx);
1221 }
1222
1223 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1224 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1225 drop_struct_lock = TRUE;
1226 }
1227
1228 if (db->db_level == 0) {
1229 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1230 ASSERT(dn->dn_maxblkid >= db->db_blkid);
1231 }
1232
1233 if (db->db_level+1 < dn->dn_nlevels) {
1234 dmu_buf_impl_t *parent = db->db_parent;
1235 dbuf_dirty_record_t *di;
1236 int parent_held = FALSE;
1237
1238 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1239 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1240
1241 parent = dbuf_hold_level(dn, db->db_level+1,
1242 db->db_blkid >> epbs, FTAG);
1243 ASSERT(parent != NULL);
1244 parent_held = TRUE;
1245 }
1246 if (drop_struct_lock)
1247 rw_exit(&dn->dn_struct_rwlock);
1248 ASSERT3U(db->db_level+1, ==, parent->db_level);
1249 di = dbuf_dirty(parent, tx);
1250 if (parent_held)
1251 dbuf_rele(parent, FTAG);
1252
1253 mutex_enter(&db->db_mtx);
1254 /*
1255 * Since we've dropped the mutex, it's possible that
1256 * dbuf_undirty() might have changed this out from under us.
1257 */
1258 if (db->db_last_dirty == dr ||
1259 dn->dn_object == DMU_META_DNODE_OBJECT) {
1260 mutex_enter(&di->dt.di.dr_mtx);
1261 ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1262 ASSERT(!list_link_active(&dr->dr_dirty_node));
1263 list_insert_tail(&di->dt.di.dr_children, dr);
1264 mutex_exit(&di->dt.di.dr_mtx);
1265 dr->dr_parent = di;
1266 }
1267 mutex_exit(&db->db_mtx);
1268 } else {
1269 ASSERT(db->db_level+1 == dn->dn_nlevels);
1270 ASSERT(db->db_blkid < dn->dn_nblkptr);
1271 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1272 mutex_enter(&dn->dn_mtx);
1273 ASSERT(!list_link_active(&dr->dr_dirty_node));
1274 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1275 mutex_exit(&dn->dn_mtx);
1276 if (drop_struct_lock)
1277 rw_exit(&dn->dn_struct_rwlock);
1278 }
1279
1280 dnode_setdirty(dn, tx);
1281 DB_DNODE_EXIT(db);
1282 return (dr);
1283 }
1284
1285 /*
1286 * Undirty a buffer in the transaction group referenced by the given
1287 * transaction. Return whether this evicted the dbuf.
1288 */
1289 static boolean_t
1290 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1291 {
1292 dnode_t *dn;
1293 uint64_t txg = tx->tx_txg;
1294 dbuf_dirty_record_t *dr, **drp;
1295
1296 ASSERT(txg != 0);
1297 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1298 ASSERT0(db->db_level);
1299 ASSERT(MUTEX_HELD(&db->db_mtx));
1300
1301 /*
1302 * If this buffer is not dirty, we're done.
1303 */
1304 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1305 if (dr->dr_txg <= txg)
1306 break;
1307 if (dr == NULL || dr->dr_txg < txg)
1308 return (B_FALSE);
1309 ASSERT(dr->dr_txg == txg);
1310 ASSERT(dr->dr_dbuf == db);
1311
1312 DB_DNODE_ENTER(db);
1313 dn = DB_DNODE(db);
1314
1315 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1316
1317 ASSERT(db->db.db_size != 0);
1318
1319 /*
1320 * Any space we accounted for in dp_dirty_* will be cleaned up by
1321 * dsl_pool_sync(). This is relatively rare so the discrepancy
1322 * is not a big deal.
1323 */
1324
1325 *drp = dr->dr_next;
1326
1327 /*
1328 * Note that there are three places in dbuf_dirty()
1329 * where this dirty record may be put on a list.
1330 * Make sure to do a list_remove corresponding to
1331 * every one of those list_insert calls.
1332 */
1333 if (dr->dr_parent) {
1334 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1335 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1336 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1337 } else if (db->db_blkid == DMU_SPILL_BLKID ||
1338 db->db_level+1 == dn->dn_nlevels) {
1339 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1340 mutex_enter(&dn->dn_mtx);
1341 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1342 mutex_exit(&dn->dn_mtx);
1343 }
1344 DB_DNODE_EXIT(db);
1345
1346 if (db->db_state != DB_NOFILL) {
1347 dbuf_unoverride(dr);
1348
1349 ASSERT(db->db_buf != NULL);
1350 ASSERT(dr->dt.dl.dr_data != NULL);
1351 if (dr->dt.dl.dr_data != db->db_buf)
1352 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db));
1353 }
1354 kmem_free(dr, sizeof (dbuf_dirty_record_t));
1355
1356 ASSERT(db->db_dirtycnt > 0);
1357 db->db_dirtycnt -= 1;
1358
1359 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1360 arc_buf_t *buf = db->db_buf;
1361
1362 ASSERT(db->db_state == DB_NOFILL || arc_released(buf));
1363 dbuf_set_data(db, NULL);
1364 VERIFY(arc_buf_remove_ref(buf, db));
1365 dbuf_evict(db);
1366 return (B_TRUE);
1367 }
1368
1369 return (B_FALSE);
1370 }
1371
1372 void
1373 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
1374 {
1375 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1376 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1377
1378 ASSERT(tx->tx_txg != 0);
1379 ASSERT(!refcount_is_zero(&db->db_holds));
1380
1381 DB_DNODE_ENTER(db);
1382 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
1383 rf |= DB_RF_HAVESTRUCT;
1384 DB_DNODE_EXIT(db);
1385 (void) dbuf_read(db, NULL, rf);
1386 (void) dbuf_dirty(db, tx);
1387 }
1388
1389 void
1390 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1391 {
1392 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1393
1394 db->db_state = DB_NOFILL;
1395
1396 dmu_buf_will_fill(db_fake, tx);
1397 }
1398
1399 void
1400 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1401 {
1402 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1403
1404 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1405 ASSERT(tx->tx_txg != 0);
1406 ASSERT(db->db_level == 0);
1407 ASSERT(!refcount_is_zero(&db->db_holds));
1408
1409 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1410 dmu_tx_private_ok(tx));
1411
1412 dbuf_noread(db);
1413 (void) dbuf_dirty(db, tx);
1414 }
1415
1416 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1417 /* ARGSUSED */
1418 void
1419 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1420 {
1421 mutex_enter(&db->db_mtx);
1422 DBUF_VERIFY(db);
1423
1424 if (db->db_state == DB_FILL) {
1425 if (db->db_level == 0 && db->db_freed_in_flight) {
1426 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1427 /* we were freed while filling */
1428 /* XXX dbuf_undirty? */
1429 bzero(db->db.db_data, db->db.db_size);
1430 db->db_freed_in_flight = FALSE;
1431 }
1432 db->db_state = DB_CACHED;
1433 cv_broadcast(&db->db_changed);
1434 }
1435 mutex_exit(&db->db_mtx);
1436 }
1437
1438 /*
1439 * Directly assign a provided arc buf to a given dbuf if it's not referenced
1440 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1441 */
1442 void
1443 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
1444 {
1445 ASSERT(!refcount_is_zero(&db->db_holds));
1446 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1447 ASSERT(db->db_level == 0);
1448 ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA);
1449 ASSERT(buf != NULL);
1450 ASSERT(arc_buf_size(buf) == db->db.db_size);
1451 ASSERT(tx->tx_txg != 0);
1452
1453 arc_return_buf(buf, db);
1454 ASSERT(arc_released(buf));
1455
1456 mutex_enter(&db->db_mtx);
1457
1458 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1459 cv_wait(&db->db_changed, &db->db_mtx);
1460
1461 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
1462
1463 if (db->db_state == DB_CACHED &&
1464 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
1465 mutex_exit(&db->db_mtx);
1466 (void) dbuf_dirty(db, tx);
1467 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
1468 VERIFY(arc_buf_remove_ref(buf, db));
1469 xuio_stat_wbuf_copied();
1470 return;
1471 }
1472
1473 xuio_stat_wbuf_nocopy();
1474 if (db->db_state == DB_CACHED) {
1475 dbuf_dirty_record_t *dr = db->db_last_dirty;
1476
1477 ASSERT(db->db_buf != NULL);
1478 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
1479 ASSERT(dr->dt.dl.dr_data == db->db_buf);
1480 if (!arc_released(db->db_buf)) {
1481 ASSERT(dr->dt.dl.dr_override_state ==
1482 DR_OVERRIDDEN);
1483 arc_release(db->db_buf, db);
1484 }
1485 dr->dt.dl.dr_data = buf;
1486 VERIFY(arc_buf_remove_ref(db->db_buf, db));
1487 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
1488 arc_release(db->db_buf, db);
1489 VERIFY(arc_buf_remove_ref(db->db_buf, db));
1490 }
1491 db->db_buf = NULL;
1492 }
1493 ASSERT(db->db_buf == NULL);
1494 dbuf_set_data(db, buf);
1495 db->db_state = DB_FILL;
1496 mutex_exit(&db->db_mtx);
1497 (void) dbuf_dirty(db, tx);
1498 dmu_buf_fill_done(&db->db, tx);
1499 }
1500
1501 /*
1502 * "Clear" the contents of this dbuf. This will mark the dbuf
1503 * EVICTING and clear *most* of its references. Unfortunately,
1504 * when we are not holding the dn_dbufs_mtx, we can't clear the
1505 * entry in the dn_dbufs list. We have to wait until dbuf_destroy()
1506 * in this case. For callers from the DMU we will usually see:
1507 * dbuf_clear()->arc_buf_evict()->dbuf_do_evict()->dbuf_destroy()
1508 * For the arc callback, we will usually see:
1509 * dbuf_do_evict()->dbuf_clear();dbuf_destroy()
1510 * Sometimes, though, we will get a mix of these two:
1511 * DMU: dbuf_clear()->arc_buf_evict()
1512 * ARC: dbuf_do_evict()->dbuf_destroy()
1513 */
1514 void
1515 dbuf_clear(dmu_buf_impl_t *db)
1516 {
1517 dnode_t *dn;
1518 dmu_buf_impl_t *parent = db->db_parent;
1519 dmu_buf_impl_t *dndb;
1520 int dbuf_gone = FALSE;
1521
1522 ASSERT(MUTEX_HELD(&db->db_mtx));
1523 ASSERT(refcount_is_zero(&db->db_holds));
1524
1525 dbuf_evict_user(db);
1526
1527 if (db->db_state == DB_CACHED) {
1528 ASSERT(db->db.db_data != NULL);
1529 if (db->db_blkid == DMU_BONUS_BLKID) {
1530 zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
1531 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
1532 }
1533 db->db.db_data = NULL;
1534 db->db_state = DB_UNCACHED;
1535 }
1536
1537 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
1538 ASSERT(db->db_data_pending == NULL);
1539
1540 db->db_state = DB_EVICTING;
1541 db->db_blkptr = NULL;
1542
1543 DB_DNODE_ENTER(db);
1544 dn = DB_DNODE(db);
1545 dndb = dn->dn_dbuf;
1546 if (db->db_blkid != DMU_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) {
1547 list_remove(&dn->dn_dbufs, db);
1548 (void) atomic_dec_32_nv(&dn->dn_dbufs_count);
1549 membar_producer();
1550 DB_DNODE_EXIT(db);
1551 /*
1552 * Decrementing the dbuf count means that the hold corresponding
1553 * to the removed dbuf is no longer discounted in dnode_move(),
1554 * so the dnode cannot be moved until after we release the hold.
1555 * The membar_producer() ensures visibility of the decremented
1556 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
1557 * release any lock.
1558 */
1559 dnode_rele(dn, db);
1560 db->db_dnode_handle = NULL;
1561 } else {
1562 DB_DNODE_EXIT(db);
1563 }
1564
1565 if (db->db_buf)
1566 dbuf_gone = arc_buf_evict(db->db_buf);
1567
1568 if (!dbuf_gone)
1569 mutex_exit(&db->db_mtx);
1570
1571 /*
1572 * If this dbuf is referenced from an indirect dbuf,
1573 * decrement the ref count on the indirect dbuf.
1574 */
1575 if (parent && parent != dndb)
1576 dbuf_rele(parent, db);
1577 }
1578
1579 static int
1580 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
1581 dmu_buf_impl_t **parentp, blkptr_t **bpp)
1582 {
1583 int nlevels, epbs;
1584
1585 *parentp = NULL;
1586 *bpp = NULL;
1587
1588 ASSERT(blkid != DMU_BONUS_BLKID);
1589
1590 if (blkid == DMU_SPILL_BLKID) {
1591 mutex_enter(&dn->dn_mtx);
1592 if (dn->dn_have_spill &&
1593 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
1594 *bpp = &dn->dn_phys->dn_spill;
1595 else
1596 *bpp = NULL;
1597 dbuf_add_ref(dn->dn_dbuf, NULL);
1598 *parentp = dn->dn_dbuf;
1599 mutex_exit(&dn->dn_mtx);
1600 return (0);
1601 }
1602
1603 if (dn->dn_phys->dn_nlevels == 0)
1604 nlevels = 1;
1605 else
1606 nlevels = dn->dn_phys->dn_nlevels;
1607
1608 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1609
1610 ASSERT3U(level * epbs, <, 64);
1611 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1612 if (level >= nlevels ||
1613 (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
1614 /* the buffer has no parent yet */
1615 return (SET_ERROR(ENOENT));
1616 } else if (level < nlevels-1) {
1617 /* this block is referenced from an indirect block */
1618 int err = dbuf_hold_impl(dn, level+1,
1619 blkid >> epbs, fail_sparse, NULL, parentp);
1620 if (err)
1621 return (err);
1622 err = dbuf_read(*parentp, NULL,
1623 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
1624 if (err) {
1625 dbuf_rele(*parentp, NULL);
1626 *parentp = NULL;
1627 return (err);
1628 }
1629 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
1630 (blkid & ((1ULL << epbs) - 1));
1631 return (0);
1632 } else {
1633 /* the block is referenced from the dnode */
1634 ASSERT3U(level, ==, nlevels-1);
1635 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
1636 blkid < dn->dn_phys->dn_nblkptr);
1637 if (dn->dn_dbuf) {
1638 dbuf_add_ref(dn->dn_dbuf, NULL);
1639 *parentp = dn->dn_dbuf;
1640 }
1641 *bpp = &dn->dn_phys->dn_blkptr[blkid];
1642 return (0);
1643 }
1644 }
1645
1646 static dmu_buf_impl_t *
1647 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
1648 dmu_buf_impl_t *parent, blkptr_t *blkptr)
1649 {
1650 objset_t *os = dn->dn_objset;
1651 dmu_buf_impl_t *db, *odb;
1652
1653 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1654 ASSERT(dn->dn_type != DMU_OT_NONE);
1655
1656 db = kmem_cache_alloc(dbuf_cache, KM_SLEEP);
1657
1658 db->db_objset = os;
1659 db->db.db_object = dn->dn_object;
1660 db->db_level = level;
1661 db->db_blkid = blkid;
1662 db->db_last_dirty = NULL;
1663 db->db_dirtycnt = 0;
1664 db->db_dnode_handle = dn->dn_handle;
1665 db->db_parent = parent;
1666 db->db_blkptr = blkptr;
1667
1668 db->db_user_ptr = NULL;
1669 db->db_user_data_ptr_ptr = NULL;
1670 db->db_evict_func = NULL;
1671 db->db_immediate_evict = 0;
1672 db->db_freed_in_flight = 0;
1673
1674 if (blkid == DMU_BONUS_BLKID) {
1675 ASSERT3P(parent, ==, dn->dn_dbuf);
1676 db->db.db_size = DN_MAX_BONUSLEN -
1677 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
1678 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
1679 db->db.db_offset = DMU_BONUS_BLKID;
1680 db->db_state = DB_UNCACHED;
1681 /* the bonus dbuf is not placed in the hash table */
1682 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1683 return (db);
1684 } else if (blkid == DMU_SPILL_BLKID) {
1685 db->db.db_size = (blkptr != NULL) ?
1686 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
1687 db->db.db_offset = 0;
1688 } else {
1689 int blocksize =
1690 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
1691 db->db.db_size = blocksize;
1692 db->db.db_offset = db->db_blkid * blocksize;
1693 }
1694
1695 /*
1696 * Hold the dn_dbufs_mtx while we get the new dbuf
1697 * in the hash table *and* added to the dbufs list.
1698 * This prevents a possible deadlock with someone
1699 * trying to look up this dbuf before its added to the
1700 * dn_dbufs list.
1701 */
1702 mutex_enter(&dn->dn_dbufs_mtx);
1703 db->db_state = DB_EVICTING;
1704 if ((odb = dbuf_hash_insert(db)) != NULL) {
1705 /* someone else inserted it first */
1706 kmem_cache_free(dbuf_cache, db);
1707 mutex_exit(&dn->dn_dbufs_mtx);
1708 return (odb);
1709 }
1710 list_insert_head(&dn->dn_dbufs, db);
1711 if (db->db_level == 0 && db->db_blkid >=
1712 dn->dn_unlisted_l0_blkid)
1713 dn->dn_unlisted_l0_blkid = db->db_blkid + 1;
1714 db->db_state = DB_UNCACHED;
1715 mutex_exit(&dn->dn_dbufs_mtx);
1716 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1717
1718 if (parent && parent != dn->dn_dbuf)
1719 dbuf_add_ref(parent, db);
1720
1721 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1722 refcount_count(&dn->dn_holds) > 0);
1723 (void) refcount_add(&dn->dn_holds, db);
1724 (void) atomic_inc_32_nv(&dn->dn_dbufs_count);
1725
1726 dprintf_dbuf(db, "db=%p\n", db);
1727
1728 return (db);
1729 }
1730
1731 static int
1732 dbuf_do_evict(void *private)
1733 {
1734 arc_buf_t *buf = private;
1735 dmu_buf_impl_t *db = buf->b_private;
1736
1737 if (!MUTEX_HELD(&db->db_mtx))
1738 mutex_enter(&db->db_mtx);
1739
1740 ASSERT(refcount_is_zero(&db->db_holds));
1741
1742 if (db->db_state != DB_EVICTING) {
1743 ASSERT(db->db_state == DB_CACHED);
1744 DBUF_VERIFY(db);
1745 db->db_buf = NULL;
1746 dbuf_evict(db);
1747 } else {
1748 mutex_exit(&db->db_mtx);
1749 dbuf_destroy(db);
1750 }
1751 return (0);
1752 }
1753
1754 static void
1755 dbuf_destroy(dmu_buf_impl_t *db)
1756 {
1757 ASSERT(refcount_is_zero(&db->db_holds));
1758
1759 if (db->db_blkid != DMU_BONUS_BLKID) {
1760 /*
1761 * If this dbuf is still on the dn_dbufs list,
1762 * remove it from that list.
1763 */
1764 if (db->db_dnode_handle != NULL) {
1765 dnode_t *dn;
1766
1767 DB_DNODE_ENTER(db);
1768 dn = DB_DNODE(db);
1769 mutex_enter(&dn->dn_dbufs_mtx);
1770 list_remove(&dn->dn_dbufs, db);
1771 (void) atomic_dec_32_nv(&dn->dn_dbufs_count);
1772 mutex_exit(&dn->dn_dbufs_mtx);
1773 DB_DNODE_EXIT(db);
1774 /*
1775 * Decrementing the dbuf count means that the hold
1776 * corresponding to the removed dbuf is no longer
1777 * discounted in dnode_move(), so the dnode cannot be
1778 * moved until after we release the hold.
1779 */
1780 dnode_rele(dn, db);
1781 db->db_dnode_handle = NULL;
1782 }
1783 dbuf_hash_remove(db);
1784 }
1785 db->db_parent = NULL;
1786 db->db_buf = NULL;
1787
1788 ASSERT(!list_link_active(&db->db_link));
1789 ASSERT(db->db.db_data == NULL);
1790 ASSERT(db->db_hash_next == NULL);
1791 ASSERT(db->db_blkptr == NULL);
1792 ASSERT(db->db_data_pending == NULL);
1793
1794 kmem_cache_free(dbuf_cache, db);
1795 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1796 }
1797
1798 void
1799 dbuf_prefetch(dnode_t *dn, uint64_t blkid, zio_priority_t prio)
1800 {
1801 dmu_buf_impl_t *db = NULL;
1802 blkptr_t *bp = NULL;
1803
1804 ASSERT(blkid != DMU_BONUS_BLKID);
1805 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1806
1807 if (dnode_block_freed(dn, blkid))
1808 return;
1809
1810 /* dbuf_find() returns with db_mtx held */
1811 if (db = dbuf_find(dn, 0, blkid)) {
1812 /*
1813 * This dbuf is already in the cache. We assume that
1814 * it is already CACHED, or else about to be either
1815 * read or filled.
1816 */
1817 mutex_exit(&db->db_mtx);
1818 return;
1819 }
1820
1821 if (dbuf_findbp(dn, 0, blkid, TRUE, &db, &bp) == 0) {
1822 if (bp && !BP_IS_HOLE(bp)) {
1823 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
1824 uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH;
1825 zbookmark_t zb;
1826
1827 SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
1828 dn->dn_object, 0, blkid);
1829
1830 (void) arc_read(NULL, dn->dn_objset->os_spa,
1831 bp, NULL, NULL, prio,
1832 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1833 &aflags, &zb);
1834 }
1835 if (db)
1836 dbuf_rele(db, NULL);
1837 }
1838 }
1839
1840 /*
1841 * Returns with db_holds incremented, and db_mtx not held.
1842 * Note: dn_struct_rwlock must be held.
1843 */
1844 int
1845 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse,
1846 void *tag, dmu_buf_impl_t **dbp)
1847 {
1848 dmu_buf_impl_t *db, *parent = NULL;
1849
1850 ASSERT(blkid != DMU_BONUS_BLKID);
1851 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1852 ASSERT3U(dn->dn_nlevels, >, level);
1853
1854 *dbp = NULL;
1855 top:
1856 /* dbuf_find() returns with db_mtx held */
1857 db = dbuf_find(dn, level, blkid);
1858
1859 if (db == NULL) {
1860 blkptr_t *bp = NULL;
1861 int err;
1862
1863 ASSERT3P(parent, ==, NULL);
1864 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
1865 if (fail_sparse) {
1866 if (err == 0 && bp && BP_IS_HOLE(bp))
1867 err = SET_ERROR(ENOENT);
1868 if (err) {
1869 if (parent)
1870 dbuf_rele(parent, NULL);
1871 return (err);
1872 }
1873 }
1874 if (err && err != ENOENT)
1875 return (err);
1876 db = dbuf_create(dn, level, blkid, parent, bp);
1877 }
1878
1879 if (db->db_buf && refcount_is_zero(&db->db_holds)) {
1880 arc_buf_add_ref(db->db_buf, db);
1881 if (db->db_buf->b_data == NULL) {
1882 dbuf_clear(db);
1883 if (parent) {
1884 dbuf_rele(parent, NULL);
1885 parent = NULL;
1886 }
1887 goto top;
1888 }
1889 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
1890 }
1891
1892 ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
1893
1894 /*
1895 * If this buffer is currently syncing out, and we are are
1896 * still referencing it from db_data, we need to make a copy
1897 * of it in case we decide we want to dirty it again in this txg.
1898 */
1899 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1900 dn->dn_object != DMU_META_DNODE_OBJECT &&
1901 db->db_state == DB_CACHED && db->db_data_pending) {
1902 dbuf_dirty_record_t *dr = db->db_data_pending;
1903
1904 if (dr->dt.dl.dr_data == db->db_buf) {
1905 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1906
1907 dbuf_set_data(db,
1908 arc_buf_alloc(dn->dn_objset->os_spa,
1909 db->db.db_size, db, type));
1910 bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
1911 db->db.db_size);
1912 }
1913 }
1914
1915 (void) refcount_add(&db->db_holds, tag);
1916 dbuf_update_data(db);
1917 DBUF_VERIFY(db);
1918 mutex_exit(&db->db_mtx);
1919
1920 /* NOTE: we can't rele the parent until after we drop the db_mtx */
1921 if (parent)
1922 dbuf_rele(parent, NULL);
1923
1924 ASSERT3P(DB_DNODE(db), ==, dn);
1925 ASSERT3U(db->db_blkid, ==, blkid);
1926 ASSERT3U(db->db_level, ==, level);
1927 *dbp = db;
1928
1929 return (0);
1930 }
1931
1932 dmu_buf_impl_t *
1933 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
1934 {
1935 dmu_buf_impl_t *db;
1936 int err = dbuf_hold_impl(dn, 0, blkid, FALSE, tag, &db);
1937 return (err ? NULL : db);
1938 }
1939
1940 dmu_buf_impl_t *
1941 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
1942 {
1943 dmu_buf_impl_t *db;
1944 int err = dbuf_hold_impl(dn, level, blkid, FALSE, tag, &db);
1945 return (err ? NULL : db);
1946 }
1947
1948 void
1949 dbuf_create_bonus(dnode_t *dn)
1950 {
1951 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1952
1953 ASSERT(dn->dn_bonus == NULL);
1954 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
1955 }
1956
1957 int
1958 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
1959 {
1960 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1961 dnode_t *dn;
1962
1963 if (db->db_blkid != DMU_SPILL_BLKID)
1964 return (SET_ERROR(ENOTSUP));
1965 if (blksz == 0)
1966 blksz = SPA_MINBLOCKSIZE;
1967 if (blksz > SPA_MAXBLOCKSIZE)
1968 blksz = SPA_MAXBLOCKSIZE;
1969 else
1970 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
1971
1972 DB_DNODE_ENTER(db);
1973 dn = DB_DNODE(db);
1974 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1975 dbuf_new_size(db, blksz, tx);
1976 rw_exit(&dn->dn_struct_rwlock);
1977 DB_DNODE_EXIT(db);
1978
1979 return (0);
1980 }
1981
1982 void
1983 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
1984 {
1985 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
1986 }
1987
1988 #pragma weak dmu_buf_add_ref = dbuf_add_ref
1989 void
1990 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
1991 {
1992 int64_t holds = refcount_add(&db->db_holds, tag);
1993 ASSERT(holds > 1);
1994 }
1995
1996 /*
1997 * If you call dbuf_rele() you had better not be referencing the dnode handle
1998 * unless you have some other direct or indirect hold on the dnode. (An indirect
1999 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2000 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2001 * dnode's parent dbuf evicting its dnode handles.
2002 */
2003 void
2004 dbuf_rele(dmu_buf_impl_t *db, void *tag)
2005 {
2006 mutex_enter(&db->db_mtx);
2007 dbuf_rele_and_unlock(db, tag);
2008 }
2009
2010 void
2011 dmu_buf_rele(dmu_buf_t *db, void *tag)
2012 {
2013 dbuf_rele((dmu_buf_impl_t *)db, tag);
2014 }
2015
2016 /*
2017 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2018 * db_dirtycnt and db_holds to be updated atomically.
2019 */
2020 void
2021 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2022 {
2023 int64_t holds;
2024
2025 ASSERT(MUTEX_HELD(&db->db_mtx));
2026 DBUF_VERIFY(db);
2027
2028 /*
2029 * Remove the reference to the dbuf before removing its hold on the
2030 * dnode so we can guarantee in dnode_move() that a referenced bonus
2031 * buffer has a corresponding dnode hold.
2032 */
2033 holds = refcount_remove(&db->db_holds, tag);
2034 ASSERT(holds >= 0);
2035
2036 /*
2037 * We can't freeze indirects if there is a possibility that they
2038 * may be modified in the current syncing context.
2039 */
2040 if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0))
2041 arc_buf_freeze(db->db_buf);
2042
2043 if (holds == db->db_dirtycnt &&
2044 db->db_level == 0 && db->db_immediate_evict)
2045 dbuf_evict_user(db);
2046
2047 if (holds == 0) {
2048 if (db->db_blkid == DMU_BONUS_BLKID) {
2049 mutex_exit(&db->db_mtx);
2050
2051 /*
2052 * If the dnode moves here, we cannot cross this barrier
2053 * until the move completes.
2054 */
2055 DB_DNODE_ENTER(db);
2056 (void) atomic_dec_32_nv(&DB_DNODE(db)->dn_dbufs_count);
2057 DB_DNODE_EXIT(db);
2058 /*
2059 * The bonus buffer's dnode hold is no longer discounted
2060 * in dnode_move(). The dnode cannot move until after
2061 * the dnode_rele().
2062 */
2063 dnode_rele(DB_DNODE(db), db);
2064 } else if (db->db_buf == NULL) {
2065 /*
2066 * This is a special case: we never associated this
2067 * dbuf with any data allocated from the ARC.
2068 */
2069 ASSERT(db->db_state == DB_UNCACHED ||
2070 db->db_state == DB_NOFILL);
2071 dbuf_evict(db);
2072 } else if (arc_released(db->db_buf)) {
2073 arc_buf_t *buf = db->db_buf;
2074 /*
2075 * This dbuf has anonymous data associated with it.
2076 */
2077 dbuf_set_data(db, NULL);
2078 VERIFY(arc_buf_remove_ref(buf, db));
2079 dbuf_evict(db);
2080 } else {
2081 VERIFY(!arc_buf_remove_ref(db->db_buf, db));
2082
2083 /*
2084 * A dbuf will be eligible for eviction if either the
2085 * 'primarycache' property is set or a duplicate
2086 * copy of this buffer is already cached in the arc.
2087 *
2088 * In the case of the 'primarycache' a buffer
2089 * is considered for eviction if it matches the
2090 * criteria set in the property.
2091 *
2092 * To decide if our buffer is considered a
2093 * duplicate, we must call into the arc to determine
2094 * if multiple buffers are referencing the same
2095 * block on-disk. If so, then we simply evict
2096 * ourselves.
2097 */
2098 if (!DBUF_IS_CACHEABLE(db) ||
2099 arc_buf_eviction_needed(db->db_buf))
2100 dbuf_clear(db);
2101 else
2102 mutex_exit(&db->db_mtx);
2103 }
2104 } else {
2105 mutex_exit(&db->db_mtx);
2106 }
2107 }
2108
2109 #pragma weak dmu_buf_refcount = dbuf_refcount
2110 uint64_t
2111 dbuf_refcount(dmu_buf_impl_t *db)
2112 {
2113 return (refcount_count(&db->db_holds));
2114 }
2115
2116 void *
2117 dmu_buf_set_user(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr,
2118 dmu_buf_evict_func_t *evict_func)
2119 {
2120 return (dmu_buf_update_user(db_fake, NULL, user_ptr,
2121 user_data_ptr_ptr, evict_func));
2122 }
2123
2124 void *
2125 dmu_buf_set_user_ie(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr,
2126 dmu_buf_evict_func_t *evict_func)
2127 {
2128 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2129
2130 db->db_immediate_evict = TRUE;
2131 return (dmu_buf_update_user(db_fake, NULL, user_ptr,
2132 user_data_ptr_ptr, evict_func));
2133 }
2134
2135 void *
2136 dmu_buf_update_user(dmu_buf_t *db_fake, void *old_user_ptr, void *user_ptr,
2137 void *user_data_ptr_ptr, dmu_buf_evict_func_t *evict_func)
2138 {
2139 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2140 ASSERT(db->db_level == 0);
2141
2142 ASSERT((user_ptr == NULL) == (evict_func == NULL));
2143
2144 mutex_enter(&db->db_mtx);
2145
2146 if (db->db_user_ptr == old_user_ptr) {
2147 db->db_user_ptr = user_ptr;
2148 db->db_user_data_ptr_ptr = user_data_ptr_ptr;
2149 db->db_evict_func = evict_func;
2150
2151 dbuf_update_data(db);
2152 } else {
2153 old_user_ptr = db->db_user_ptr;
2154 }
2155
2156 mutex_exit(&db->db_mtx);
2157 return (old_user_ptr);
2158 }
2159
2160 void *
2161 dmu_buf_get_user(dmu_buf_t *db_fake)
2162 {
2163 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2164 ASSERT(!refcount_is_zero(&db->db_holds));
2165
2166 return (db->db_user_ptr);
2167 }
2168
2169 boolean_t
2170 dmu_buf_freeable(dmu_buf_t *dbuf)
2171 {
2172 boolean_t res = B_FALSE;
2173 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2174
2175 if (db->db_blkptr)
2176 res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset,
2177 db->db_blkptr, db->db_blkptr->blk_birth);
2178
2179 return (res);
2180 }
2181
2182 blkptr_t *
2183 dmu_buf_get_blkptr(dmu_buf_t *db)
2184 {
2185 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2186 return (dbi->db_blkptr);
2187 }
2188
2189 static void
2190 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
2191 {
2192 /* ASSERT(dmu_tx_is_syncing(tx) */
2193 ASSERT(MUTEX_HELD(&db->db_mtx));
2194
2195 if (db->db_blkptr != NULL)
2196 return;
2197
2198 if (db->db_blkid == DMU_SPILL_BLKID) {
2199 db->db_blkptr = &dn->dn_phys->dn_spill;
2200 BP_ZERO(db->db_blkptr);
2201 return;
2202 }
2203 if (db->db_level == dn->dn_phys->dn_nlevels-1) {
2204 /*
2205 * This buffer was allocated at a time when there was
2206 * no available blkptrs from the dnode, or it was
2207 * inappropriate to hook it in (i.e., nlevels mis-match).
2208 */
2209 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
2210 ASSERT(db->db_parent == NULL);
2211 db->db_parent = dn->dn_dbuf;
2212 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
2213 DBUF_VERIFY(db);
2214 } else {
2215 dmu_buf_impl_t *parent = db->db_parent;
2216 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2217
2218 ASSERT(dn->dn_phys->dn_nlevels > 1);
2219 if (parent == NULL) {
2220 mutex_exit(&db->db_mtx);
2221 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2222 (void) dbuf_hold_impl(dn, db->db_level+1,
2223 db->db_blkid >> epbs, FALSE, db, &parent);
2224 rw_exit(&dn->dn_struct_rwlock);
2225 mutex_enter(&db->db_mtx);
2226 db->db_parent = parent;
2227 }
2228 db->db_blkptr = (blkptr_t *)parent->db.db_data +
2229 (db->db_blkid & ((1ULL << epbs) - 1));
2230 DBUF_VERIFY(db);
2231 }
2232 }
2233
2234 static void
2235 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2236 {
2237 dmu_buf_impl_t *db = dr->dr_dbuf;
2238 dnode_t *dn;
2239 zio_t *zio;
2240
2241 ASSERT(dmu_tx_is_syncing(tx));
2242
2243 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2244
2245 mutex_enter(&db->db_mtx);
2246
2247 ASSERT(db->db_level > 0);
2248 DBUF_VERIFY(db);
2249
2250 /* Read the block if it hasn't been read yet. */
2251 if (db->db_buf == NULL) {
2252 mutex_exit(&db->db_mtx);
2253 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
2254 mutex_enter(&db->db_mtx);
2255 }
2256 ASSERT3U(db->db_state, ==, DB_CACHED);
2257 ASSERT(db->db_buf != NULL);
2258
2259 DB_DNODE_ENTER(db);
2260 dn = DB_DNODE(db);
2261 /* Indirect block size must match what the dnode thinks it is. */
2262 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2263 dbuf_check_blkptr(dn, db);
2264 DB_DNODE_EXIT(db);
2265
2266 /* Provide the pending dirty record to child dbufs */
2267 db->db_data_pending = dr;
2268
2269 mutex_exit(&db->db_mtx);
2270 dbuf_write(dr, db->db_buf, tx);
2271
2272 zio = dr->dr_zio;
2273 mutex_enter(&dr->dt.di.dr_mtx);
2274 dbuf_sync_list(&dr->dt.di.dr_children, tx);
2275 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2276 mutex_exit(&dr->dt.di.dr_mtx);
2277 zio_nowait(zio);
2278 }
2279
2280 static void
2281 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2282 {
2283 arc_buf_t **datap = &dr->dt.dl.dr_data;
2284 dmu_buf_impl_t *db = dr->dr_dbuf;
2285 dnode_t *dn;
2286 objset_t *os;
2287 uint64_t txg = tx->tx_txg;
2288
2289 ASSERT(dmu_tx_is_syncing(tx));
2290
2291 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2292
2293 mutex_enter(&db->db_mtx);
2294 /*
2295 * To be synced, we must be dirtied. But we
2296 * might have been freed after the dirty.
2297 */
2298 if (db->db_state == DB_UNCACHED) {
2299 /* This buffer has been freed since it was dirtied */
2300 ASSERT(db->db.db_data == NULL);
2301 } else if (db->db_state == DB_FILL) {
2302 /* This buffer was freed and is now being re-filled */
2303 ASSERT(db->db.db_data != dr->dt.dl.dr_data);
2304 } else {
2305 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
2306 }
2307 DBUF_VERIFY(db);
2308
2309 DB_DNODE_ENTER(db);
2310 dn = DB_DNODE(db);
2311
2312 if (db->db_blkid == DMU_SPILL_BLKID) {
2313 mutex_enter(&dn->dn_mtx);
2314 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
2315 mutex_exit(&dn->dn_mtx);
2316 }
2317
2318 /*
2319 * If this is a bonus buffer, simply copy the bonus data into the
2320 * dnode. It will be written out when the dnode is synced (and it
2321 * will be synced, since it must have been dirty for dbuf_sync to
2322 * be called).
2323 */
2324 if (db->db_blkid == DMU_BONUS_BLKID) {
2325 dbuf_dirty_record_t **drp;
2326
2327 ASSERT(*datap != NULL);
2328 ASSERT0(db->db_level);
2329 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN);
2330 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen);
2331 DB_DNODE_EXIT(db);
2332
2333 if (*datap != db->db.db_data) {
2334 zio_buf_free(*datap, DN_MAX_BONUSLEN);
2335 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
2336 }
2337 db->db_data_pending = NULL;
2338 drp = &db->db_last_dirty;
2339 while (*drp != dr)
2340 drp = &(*drp)->dr_next;
2341 ASSERT(dr->dr_next == NULL);
2342 ASSERT(dr->dr_dbuf == db);
2343 *drp = dr->dr_next;
2344 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2345 ASSERT(db->db_dirtycnt > 0);
2346 db->db_dirtycnt -= 1;
2347 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
2348 return;
2349 }
2350
2351 os = dn->dn_objset;
2352
2353 /*
2354 * This function may have dropped the db_mtx lock allowing a dmu_sync
2355 * operation to sneak in. As a result, we need to ensure that we
2356 * don't check the dr_override_state until we have returned from
2357 * dbuf_check_blkptr.
2358 */
2359 dbuf_check_blkptr(dn, db);
2360
2361 /*
2362 * If this buffer is in the middle of an immediate write,
2363 * wait for the synchronous IO to complete.
2364 */
2365 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
2366 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
2367 cv_wait(&db->db_changed, &db->db_mtx);
2368 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
2369 }
2370
2371 if (db->db_state != DB_NOFILL &&
2372 dn->dn_object != DMU_META_DNODE_OBJECT &&
2373 refcount_count(&db->db_holds) > 1 &&
2374 dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
2375 *datap == db->db_buf) {
2376 /*
2377 * If this buffer is currently "in use" (i.e., there
2378 * are active holds and db_data still references it),
2379 * then make a copy before we start the write so that
2380 * any modifications from the open txg will not leak
2381 * into this write.
2382 *
2383 * NOTE: this copy does not need to be made for
2384 * objects only modified in the syncing context (e.g.
2385 * DNONE_DNODE blocks).
2386 */
2387 int blksz = arc_buf_size(*datap);
2388 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2389 *datap = arc_buf_alloc(os->os_spa, blksz, db, type);
2390 bcopy(db->db.db_data, (*datap)->b_data, blksz);
2391 }
2392 db->db_data_pending = dr;
2393
2394 mutex_exit(&db->db_mtx);
2395
2396 dbuf_write(dr, *datap, tx);
2397
2398 ASSERT(!list_link_active(&dr->dr_dirty_node));
2399 if (dn->dn_object == DMU_META_DNODE_OBJECT) {
2400 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
2401 DB_DNODE_EXIT(db);
2402 } else {
2403 /*
2404 * Although zio_nowait() does not "wait for an IO", it does
2405 * initiate the IO. If this is an empty write it seems plausible
2406 * that the IO could actually be completed before the nowait
2407 * returns. We need to DB_DNODE_EXIT() first in case
2408 * zio_nowait() invalidates the dbuf.
2409 */
2410 DB_DNODE_EXIT(db);
2411 zio_nowait(dr->dr_zio);
2412 }
2413 }
2414
2415 void
2416 dbuf_sync_list(list_t *list, dmu_tx_t *tx)
2417 {
2418 dbuf_dirty_record_t *dr;
2419
2420 while (dr = list_head(list)) {
2421 if (dr->dr_zio != NULL) {
2422 /*
2423 * If we find an already initialized zio then we
2424 * are processing the meta-dnode, and we have finished.
2425 * The dbufs for all dnodes are put back on the list
2426 * during processing, so that we can zio_wait()
2427 * these IOs after initiating all child IOs.
2428 */
2429 ASSERT3U(dr->dr_dbuf->db.db_object, ==,
2430 DMU_META_DNODE_OBJECT);
2431 break;
2432 }
2433 list_remove(list, dr);
2434 if (dr->dr_dbuf->db_level > 0)
2435 dbuf_sync_indirect(dr, tx);
2436 else
2437 dbuf_sync_leaf(dr, tx);
2438 }
2439 }
2440
2441 /* ARGSUSED */
2442 static void
2443 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
2444 {
2445 dmu_buf_impl_t *db = vdb;
2446 dnode_t *dn;
2447 blkptr_t *bp = zio->io_bp;
2448 blkptr_t *bp_orig = &zio->io_bp_orig;
2449 spa_t *spa = zio->io_spa;
2450 int64_t delta;
2451 uint64_t fill = 0;
2452 int i;
2453
2454 ASSERT(db->db_blkptr == bp);
2455
2456 DB_DNODE_ENTER(db);
2457 dn = DB_DNODE(db);
2458 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
2459 dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
2460 zio->io_prev_space_delta = delta;
2461
2462 if (bp->blk_birth != 0) {
2463 ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
2464 BP_GET_TYPE(bp) == dn->dn_type) ||
2465 (db->db_blkid == DMU_SPILL_BLKID &&
2466 BP_GET_TYPE(bp) == dn->dn_bonustype));
2467 ASSERT(BP_GET_LEVEL(bp) == db->db_level);
2468 }
2469
2470 mutex_enter(&db->db_mtx);
2471
2472 #ifdef ZFS_DEBUG
2473 if (db->db_blkid == DMU_SPILL_BLKID) {
2474 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2475 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2476 db->db_blkptr == &dn->dn_phys->dn_spill);
2477 }
2478 #endif
2479
2480 if (db->db_level == 0) {
2481 mutex_enter(&dn->dn_mtx);
2482 if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
2483 db->db_blkid != DMU_SPILL_BLKID)
2484 dn->dn_phys->dn_maxblkid = db->db_blkid;
2485 mutex_exit(&dn->dn_mtx);
2486
2487 if (dn->dn_type == DMU_OT_DNODE) {
2488 dnode_phys_t *dnp = db->db.db_data;
2489 for (i = db->db.db_size >> DNODE_SHIFT; i > 0;
2490 i--, dnp++) {
2491 if (dnp->dn_type != DMU_OT_NONE)
2492 fill++;
2493 }
2494 } else {
2495 if (BP_IS_HOLE(bp)) {
2496 fill = 0;
2497 } else {
2498 fill = 1;
2499 }
2500 }
2501 } else {
2502 blkptr_t *ibp = db->db.db_data;
2503 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2504 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
2505 if (BP_IS_HOLE(ibp))
2506 continue;
2507 fill += ibp->blk_fill;
2508 }
2509 }
2510 DB_DNODE_EXIT(db);
2511
2512 bp->blk_fill = fill;
2513
2514 mutex_exit(&db->db_mtx);
2515 }
2516
2517 /*
2518 * The SPA will call this callback several times for each zio - once
2519 * for every physical child i/o (zio->io_phys_children times). This
2520 * allows the DMU to monitor the progress of each logical i/o. For example,
2521 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
2522 * block. There may be a long delay before all copies/fragments are completed,
2523 * so this callback allows us to retire dirty space gradually, as the physical
2524 * i/os complete.
2525 */
2526 /* ARGSUSED */
2527 static void
2528 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
2529 {
2530 dmu_buf_impl_t *db = arg;
2531 objset_t *os = db->db_objset;
2532 dsl_pool_t *dp = dmu_objset_pool(os);
2533 dbuf_dirty_record_t *dr;
2534 int delta = 0;
2535
2536 dr = db->db_data_pending;
2537 ASSERT3U(dr->dr_txg, ==, zio->io_txg);
2538
2539 /*
2540 * The callback will be called io_phys_children times. Retire one
2541 * portion of our dirty space each time we are called. Any rounding
2542 * error will be cleaned up by dsl_pool_sync()'s call to
2543 * dsl_pool_undirty_space().
2544 */
2545 delta = dr->dr_accounted / zio->io_phys_children;
2546 dsl_pool_undirty_space(dp, delta, zio->io_txg);
2547 }
2548
2549 /* ARGSUSED */
2550 static void
2551 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
2552 {
2553 dmu_buf_impl_t *db = vdb;
2554 blkptr_t *bp_orig = &zio->io_bp_orig;
2555 blkptr_t *bp = db->db_blkptr;
2556 objset_t *os = db->db_objset;
2557 dmu_tx_t *tx = os->os_synctx;
2558 dbuf_dirty_record_t **drp, *dr;
2559
2560 ASSERT0(zio->io_error);
2561 ASSERT(db->db_blkptr == bp);
2562
2563 /*
2564 * For nopwrites and rewrites we ensure that the bp matches our
2565 * original and bypass all the accounting.
2566 */
2567 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
2568 ASSERT(BP_EQUAL(bp, bp_orig));
2569 } else {
2570 dsl_dataset_t *ds = os->os_dsl_dataset;
2571 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
2572 dsl_dataset_block_born(ds, bp, tx);
2573 }
2574
2575 mutex_enter(&db->db_mtx);
2576
2577 DBUF_VERIFY(db);
2578
2579 drp = &db->db_last_dirty;
2580 while ((dr = *drp) != db->db_data_pending)
2581 drp = &dr->dr_next;
2582 ASSERT(!list_link_active(&dr->dr_dirty_node));
2583 ASSERT(dr->dr_dbuf == db);
2584 ASSERT(dr->dr_next == NULL);
2585 *drp = dr->dr_next;
2586
2587 #ifdef ZFS_DEBUG
2588 if (db->db_blkid == DMU_SPILL_BLKID) {
2589 dnode_t *dn;
2590
2591 DB_DNODE_ENTER(db);
2592 dn = DB_DNODE(db);
2593 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2594 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2595 db->db_blkptr == &dn->dn_phys->dn_spill);
2596 DB_DNODE_EXIT(db);
2597 }
2598 #endif
2599
2600 if (db->db_level == 0) {
2601 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2602 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
2603 if (db->db_state != DB_NOFILL) {
2604 if (dr->dt.dl.dr_data != db->db_buf)
2605 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data,
2606 db));
2607 else if (!arc_released(db->db_buf))
2608 arc_set_callback(db->db_buf, dbuf_do_evict, db);
2609 }
2610 } else {
2611 dnode_t *dn;
2612
2613 DB_DNODE_ENTER(db);
2614 dn = DB_DNODE(db);
2615 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2616 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
2617 if (!BP_IS_HOLE(db->db_blkptr)) {
2618 int epbs =
2619 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2620 ASSERT3U(db->db_blkid, <=,
2621 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs));
2622 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
2623 db->db.db_size);
2624 arc_set_callback(db->db_buf, dbuf_do_evict, db);
2625 }
2626 DB_DNODE_EXIT(db);
2627 mutex_destroy(&dr->dt.di.dr_mtx);
2628 list_destroy(&dr->dt.di.dr_children);
2629 }
2630 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2631
2632 cv_broadcast(&db->db_changed);
2633 ASSERT(db->db_dirtycnt > 0);
2634 db->db_dirtycnt -= 1;
2635 db->db_data_pending = NULL;
2636 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg);
2637 }
2638
2639 static void
2640 dbuf_write_nofill_ready(zio_t *zio)
2641 {
2642 dbuf_write_ready(zio, NULL, zio->io_private);
2643 }
2644
2645 static void
2646 dbuf_write_nofill_done(zio_t *zio)
2647 {
2648 dbuf_write_done(zio, NULL, zio->io_private);
2649 }
2650
2651 static void
2652 dbuf_write_override_ready(zio_t *zio)
2653 {
2654 dbuf_dirty_record_t *dr = zio->io_private;
2655 dmu_buf_impl_t *db = dr->dr_dbuf;
2656
2657 dbuf_write_ready(zio, NULL, db);
2658 }
2659
2660 static void
2661 dbuf_write_override_done(zio_t *zio)
2662 {
2663 dbuf_dirty_record_t *dr = zio->io_private;
2664 dmu_buf_impl_t *db = dr->dr_dbuf;
2665 blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
2666
2667 mutex_enter(&db->db_mtx);
2668 if (!BP_EQUAL(zio->io_bp, obp)) {
2669 if (!BP_IS_HOLE(obp))
2670 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
2671 arc_release(dr->dt.dl.dr_data, db);
2672 }
2673 mutex_exit(&db->db_mtx);
2674
2675 dbuf_write_done(zio, NULL, db);
2676 }
2677
2678 /* Issue I/O to commit a dirty buffer to disk. */
2679 static void
2680 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
2681 {
2682 dmu_buf_impl_t *db = dr->dr_dbuf;
2683 dnode_t *dn;
2684 objset_t *os;
2685 dmu_buf_impl_t *parent = db->db_parent;
2686 uint64_t txg = tx->tx_txg;
2687 zbookmark_t zb;
2688 zio_prop_t zp;
2689 zio_t *zio;
2690 int wp_flag = 0;
2691
2692 DB_DNODE_ENTER(db);
2693 dn = DB_DNODE(db);
2694 os = dn->dn_objset;
2695
2696 if (db->db_state != DB_NOFILL) {
2697 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
2698 /*
2699 * Private object buffers are released here rather
2700 * than in dbuf_dirty() since they are only modified
2701 * in the syncing context and we don't want the
2702 * overhead of making multiple copies of the data.
2703 */
2704 if (BP_IS_HOLE(db->db_blkptr)) {
2705 arc_buf_thaw(data);
2706 } else {
2707 dbuf_release_bp(db);
2708 }
2709 }
2710 }
2711
2712 if (parent != dn->dn_dbuf) {
2713 /* Our parent is an indirect block. */
2714 /* We have a dirty parent that has been scheduled for write. */
2715 ASSERT(parent && parent->db_data_pending);
2716 /* Our parent's buffer is one level closer to the dnode. */
2717 ASSERT(db->db_level == parent->db_level-1);
2718 /*
2719 * We're about to modify our parent's db_data by modifying
2720 * our block pointer, so the parent must be released.
2721 */
2722 ASSERT(arc_released(parent->db_buf));
2723 zio = parent->db_data_pending->dr_zio;
2724 } else {
2725 /* Our parent is the dnode itself. */
2726 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
2727 db->db_blkid != DMU_SPILL_BLKID) ||
2728 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
2729 if (db->db_blkid != DMU_SPILL_BLKID)
2730 ASSERT3P(db->db_blkptr, ==,
2731 &dn->dn_phys->dn_blkptr[db->db_blkid]);
2732 zio = dn->dn_zio;
2733 }
2734
2735 ASSERT(db->db_level == 0 || data == db->db_buf);
2736 ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
2737 ASSERT(zio);
2738
2739 SET_BOOKMARK(&zb, os->os_dsl_dataset ?
2740 os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
2741 db->db.db_object, db->db_level, db->db_blkid);
2742
2743 if (db->db_blkid == DMU_SPILL_BLKID)
2744 wp_flag = WP_SPILL;
2745 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
2746
2747 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
2748 DB_DNODE_EXIT(db);
2749
2750 if (db->db_level == 0 && dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
2751 ASSERT(db->db_state != DB_NOFILL);
2752 dr->dr_zio = zio_write(zio, os->os_spa, txg,
2753 db->db_blkptr, data->b_data, arc_buf_size(data), &zp,
2754 dbuf_write_override_ready, NULL, dbuf_write_override_done,
2755 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
2756 mutex_enter(&db->db_mtx);
2757 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
2758 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
2759 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
2760 mutex_exit(&db->db_mtx);
2761 } else if (db->db_state == DB_NOFILL) {
2762 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF ||
2763 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY);
2764 dr->dr_zio = zio_write(zio, os->os_spa, txg,
2765 db->db_blkptr, NULL, db->db.db_size, &zp,
2766 dbuf_write_nofill_ready, NULL, dbuf_write_nofill_done, db,
2767 ZIO_PRIORITY_ASYNC_WRITE,
2768 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
2769 } else {
2770 ASSERT(arc_released(data));
2771 dr->dr_zio = arc_write(zio, os->os_spa, txg,
2772 db->db_blkptr, data, DBUF_IS_L2CACHEABLE(db),
2773 DBUF_IS_L2COMPRESSIBLE(db), &zp, dbuf_write_ready,
2774 dbuf_write_physdone, dbuf_write_done, db,
2775 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
2776 }
2777 }