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 */
27
28 #include <sys/zfs_context.h>
29 #include <sys/dmu.h>
30 #include <sys/dmu_send.h>
31 #include <sys/dmu_impl.h>
32 #include <sys/dbuf.h>
33 #include <sys/dmu_objset.h>
34 #include <sys/dsl_dataset.h>
35 #include <sys/dsl_dir.h>
36 #include <sys/dmu_tx.h>
37 #include <sys/spa.h>
38 #include <sys/zio.h>
39 #include <sys/dmu_zfetch.h>
40 #include <sys/sa.h>
41 #include <sys/sa_impl.h>
42
43 /*
44 * Number of times that zfs_free_range() took the slow path while doing
45 * a zfs receive. A nonzero value indicates a potential performance problem.
46 */
47 uint64_t zfs_free_range_recv_miss;
48
49 static void dbuf_destroy(dmu_buf_impl_t *db);
50 static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
51 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
52
53 /*
54 * Global data structures and functions for the dbuf cache.
55 */
56 static kmem_cache_t *dbuf_cache;
57
58 /* ARGSUSED */
59 static int
60 dbuf_cons(void *vdb, void *unused, int kmflag)
61 {
62 dmu_buf_impl_t *db = vdb;
63 bzero(db, sizeof (dmu_buf_impl_t));
64
65 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
66 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
67 refcount_create(&db->db_holds);
68 return (0);
69 }
70
71 /* ARGSUSED */
72 static void
73 dbuf_dest(void *vdb, void *unused)
74 {
75 dmu_buf_impl_t *db = vdb;
76 mutex_destroy(&db->db_mtx);
77 cv_destroy(&db->db_changed);
78 refcount_destroy(&db->db_holds);
79 }
80
81 /*
82 * dbuf hash table routines
83 */
84 static dbuf_hash_table_t dbuf_hash_table;
85
86 static uint64_t dbuf_hash_count;
87
88 static uint64_t
89 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
90 {
91 uintptr_t osv = (uintptr_t)os;
92 uint64_t crc = -1ULL;
93
94 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
95 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF];
96 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF];
97 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF];
98 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF];
99 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF];
100 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF];
101
102 crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16);
103
104 return (crc);
105 }
106
107 #define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid);
108
109 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
110 ((dbuf)->db.db_object == (obj) && \
111 (dbuf)->db_objset == (os) && \
112 (dbuf)->db_level == (level) && \
113 (dbuf)->db_blkid == (blkid))
114
115 dmu_buf_impl_t *
116 dbuf_find(dnode_t *dn, uint8_t level, uint64_t blkid)
117 {
118 dbuf_hash_table_t *h = &dbuf_hash_table;
119 objset_t *os = dn->dn_objset;
120 uint64_t obj = dn->dn_object;
121 uint64_t hv = DBUF_HASH(os, obj, level, blkid);
122 uint64_t idx = hv & h->hash_table_mask;
123 dmu_buf_impl_t *db;
124
125 mutex_enter(DBUF_HASH_MUTEX(h, idx));
126 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
127 if (DBUF_EQUAL(db, os, obj, level, blkid)) {
128 mutex_enter(&db->db_mtx);
129 if (db->db_state != DB_EVICTING) {
130 mutex_exit(DBUF_HASH_MUTEX(h, idx));
131 return (db);
132 }
133 mutex_exit(&db->db_mtx);
134 }
135 }
136 mutex_exit(DBUF_HASH_MUTEX(h, idx));
137 return (NULL);
138 }
139
140 /*
141 * Insert an entry into the hash table. If there is already an element
142 * equal to elem in the hash table, then the already existing element
143 * will be returned and the new element will not be inserted.
144 * Otherwise returns NULL.
145 */
146 static dmu_buf_impl_t *
147 dbuf_hash_insert(dmu_buf_impl_t *db)
148 {
149 dbuf_hash_table_t *h = &dbuf_hash_table;
150 objset_t *os = db->db_objset;
151 uint64_t obj = db->db.db_object;
152 int level = db->db_level;
153 uint64_t blkid = db->db_blkid;
154 uint64_t hv = DBUF_HASH(os, obj, level, blkid);
155 uint64_t idx = hv & h->hash_table_mask;
156 dmu_buf_impl_t *dbf;
157
158 mutex_enter(DBUF_HASH_MUTEX(h, idx));
159 for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) {
160 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
161 mutex_enter(&dbf->db_mtx);
162 if (dbf->db_state != DB_EVICTING) {
163 mutex_exit(DBUF_HASH_MUTEX(h, idx));
164 return (dbf);
165 }
166 mutex_exit(&dbf->db_mtx);
167 }
168 }
169
170 mutex_enter(&db->db_mtx);
171 db->db_hash_next = h->hash_table[idx];
172 h->hash_table[idx] = db;
173 mutex_exit(DBUF_HASH_MUTEX(h, idx));
174 atomic_add_64(&dbuf_hash_count, 1);
175
176 return (NULL);
177 }
178
179 /*
180 * Remove an entry from the hash table. This operation will
181 * fail if there are any existing holds on the db.
182 */
183 static void
184 dbuf_hash_remove(dmu_buf_impl_t *db)
185 {
186 dbuf_hash_table_t *h = &dbuf_hash_table;
187 uint64_t hv = DBUF_HASH(db->db_objset, db->db.db_object,
188 db->db_level, db->db_blkid);
189 uint64_t idx = hv & h->hash_table_mask;
190 dmu_buf_impl_t *dbf, **dbp;
191
192 /*
193 * We musn't hold db_mtx to maintin lock ordering:
194 * DBUF_HASH_MUTEX > db_mtx.
195 */
196 ASSERT(refcount_is_zero(&db->db_holds));
197 ASSERT(db->db_state == DB_EVICTING);
198 ASSERT(!MUTEX_HELD(&db->db_mtx));
199
200 mutex_enter(DBUF_HASH_MUTEX(h, idx));
201 dbp = &h->hash_table[idx];
202 while ((dbf = *dbp) != db) {
203 dbp = &dbf->db_hash_next;
204 ASSERT(dbf != NULL);
205 }
206 *dbp = db->db_hash_next;
207 db->db_hash_next = NULL;
208 mutex_exit(DBUF_HASH_MUTEX(h, idx));
209 atomic_add_64(&dbuf_hash_count, -1);
210 }
211
212 static arc_evict_func_t dbuf_do_evict;
213
214 static void
215 dbuf_evict_user(dmu_buf_impl_t *db)
216 {
217 ASSERT(MUTEX_HELD(&db->db_mtx));
218
219 if (db->db_level != 0 || db->db_evict_func == NULL)
220 return;
221
222 if (db->db_user_data_ptr_ptr)
223 *db->db_user_data_ptr_ptr = db->db.db_data;
224 db->db_evict_func(&db->db, db->db_user_ptr);
225 db->db_user_ptr = NULL;
226 db->db_user_data_ptr_ptr = NULL;
227 db->db_evict_func = NULL;
228 }
229
230 boolean_t
231 dbuf_is_metadata(dmu_buf_impl_t *db)
232 {
233 if (db->db_level > 0) {
234 return (B_TRUE);
235 } else {
236 boolean_t is_metadata;
237
238 DB_DNODE_ENTER(db);
239 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type);
240 DB_DNODE_EXIT(db);
241
242 return (is_metadata);
243 }
244 }
245
246 void
247 dbuf_evict(dmu_buf_impl_t *db)
248 {
249 ASSERT(MUTEX_HELD(&db->db_mtx));
250 ASSERT(db->db_buf == NULL);
251 ASSERT(db->db_data_pending == NULL);
252
253 dbuf_clear(db);
254 dbuf_destroy(db);
255 }
256
257 void
258 dbuf_init(void)
259 {
260 uint64_t hsize = 1ULL << 16;
261 dbuf_hash_table_t *h = &dbuf_hash_table;
262 int i;
263
264 /*
265 * The hash table is big enough to fill all of physical memory
266 * with an average 4K block size. The table will take up
267 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
268 */
269 while (hsize * 4096 < physmem * PAGESIZE)
270 hsize <<= 1;
271
272 retry:
273 h->hash_table_mask = hsize - 1;
274 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
275 if (h->hash_table == NULL) {
276 /* XXX - we should really return an error instead of assert */
277 ASSERT(hsize > (1ULL << 10));
278 hsize >>= 1;
279 goto retry;
280 }
281
282 dbuf_cache = kmem_cache_create("dmu_buf_impl_t",
283 sizeof (dmu_buf_impl_t),
284 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
285
286 for (i = 0; i < DBUF_MUTEXES; i++)
287 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
288 }
289
290 void
291 dbuf_fini(void)
292 {
293 dbuf_hash_table_t *h = &dbuf_hash_table;
294 int i;
295
296 for (i = 0; i < DBUF_MUTEXES; i++)
297 mutex_destroy(&h->hash_mutexes[i]);
298 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
299 kmem_cache_destroy(dbuf_cache);
300 }
301
302 /*
303 * Other stuff.
304 */
305
306 #ifdef ZFS_DEBUG
307 static void
308 dbuf_verify(dmu_buf_impl_t *db)
309 {
310 dnode_t *dn;
311 dbuf_dirty_record_t *dr;
312
313 ASSERT(MUTEX_HELD(&db->db_mtx));
314
315 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
316 return;
317
318 ASSERT(db->db_objset != NULL);
319 DB_DNODE_ENTER(db);
320 dn = DB_DNODE(db);
321 if (dn == NULL) {
322 ASSERT(db->db_parent == NULL);
323 ASSERT(db->db_blkptr == NULL);
324 } else {
325 ASSERT3U(db->db.db_object, ==, dn->dn_object);
326 ASSERT3P(db->db_objset, ==, dn->dn_objset);
327 ASSERT3U(db->db_level, <, dn->dn_nlevels);
328 ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
329 db->db_blkid == DMU_SPILL_BLKID ||
330 !list_is_empty(&dn->dn_dbufs));
331 }
332 if (db->db_blkid == DMU_BONUS_BLKID) {
333 ASSERT(dn != NULL);
334 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
335 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
336 } else if (db->db_blkid == DMU_SPILL_BLKID) {
337 ASSERT(dn != NULL);
338 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
339 ASSERT0(db->db.db_offset);
340 } else {
341 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
342 }
343
344 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
345 ASSERT(dr->dr_dbuf == db);
346
347 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
348 ASSERT(dr->dr_dbuf == db);
349
350 /*
351 * We can't assert that db_size matches dn_datablksz because it
352 * can be momentarily different when another thread is doing
353 * dnode_set_blksz().
354 */
355 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
356 dr = db->db_data_pending;
357 /*
358 * It should only be modified in syncing context, so
359 * make sure we only have one copy of the data.
360 */
361 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
362 }
363
364 /* verify db->db_blkptr */
365 if (db->db_blkptr) {
366 if (db->db_parent == dn->dn_dbuf) {
367 /* db is pointed to by the dnode */
368 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
369 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
370 ASSERT(db->db_parent == NULL);
371 else
372 ASSERT(db->db_parent != NULL);
373 if (db->db_blkid != DMU_SPILL_BLKID)
374 ASSERT3P(db->db_blkptr, ==,
375 &dn->dn_phys->dn_blkptr[db->db_blkid]);
376 } else {
377 /* db is pointed to by an indirect block */
378 int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
379 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
380 ASSERT3U(db->db_parent->db.db_object, ==,
381 db->db.db_object);
382 /*
383 * dnode_grow_indblksz() can make this fail if we don't
384 * have the struct_rwlock. XXX indblksz no longer
385 * grows. safe to do this now?
386 */
387 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
388 ASSERT3P(db->db_blkptr, ==,
389 ((blkptr_t *)db->db_parent->db.db_data +
390 db->db_blkid % epb));
391 }
392 }
393 }
394 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
395 (db->db_buf == NULL || db->db_buf->b_data) &&
396 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
397 db->db_state != DB_FILL && !dn->dn_free_txg) {
398 /*
399 * If the blkptr isn't set but they have nonzero data,
400 * it had better be dirty, otherwise we'll lose that
401 * data when we evict this buffer.
402 */
403 if (db->db_dirtycnt == 0) {
404 uint64_t *buf = db->db.db_data;
405 int i;
406
407 for (i = 0; i < db->db.db_size >> 3; i++) {
408 ASSERT(buf[i] == 0);
409 }
410 }
411 }
412 DB_DNODE_EXIT(db);
413 }
414 #endif
415
416 static void
417 dbuf_update_data(dmu_buf_impl_t *db)
418 {
419 ASSERT(MUTEX_HELD(&db->db_mtx));
420 if (db->db_level == 0 && db->db_user_data_ptr_ptr) {
421 ASSERT(!refcount_is_zero(&db->db_holds));
422 *db->db_user_data_ptr_ptr = db->db.db_data;
423 }
424 }
425
426 static void
427 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
428 {
429 ASSERT(MUTEX_HELD(&db->db_mtx));
430 ASSERT(db->db_buf == NULL || !arc_has_callback(db->db_buf));
431 db->db_buf = buf;
432 if (buf != NULL) {
433 ASSERT(buf->b_data != NULL);
434 db->db.db_data = buf->b_data;
435 if (!arc_released(buf))
436 arc_set_callback(buf, dbuf_do_evict, db);
437 dbuf_update_data(db);
438 } else {
439 dbuf_evict_user(db);
440 db->db.db_data = NULL;
441 if (db->db_state != DB_NOFILL)
442 db->db_state = DB_UNCACHED;
443 }
444 }
445
446 /*
447 * Loan out an arc_buf for read. Return the loaned arc_buf.
448 */
449 arc_buf_t *
450 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
451 {
452 arc_buf_t *abuf;
453
454 mutex_enter(&db->db_mtx);
455 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
456 int blksz = db->db.db_size;
457 spa_t *spa;
458
459 mutex_exit(&db->db_mtx);
460 DB_GET_SPA(&spa, db);
461 abuf = arc_loan_buf(spa, blksz);
462 bcopy(db->db.db_data, abuf->b_data, blksz);
463 } else {
464 abuf = db->db_buf;
465 arc_loan_inuse_buf(abuf, db);
466 dbuf_set_data(db, NULL);
467 mutex_exit(&db->db_mtx);
468 }
469 return (abuf);
470 }
471
472 uint64_t
473 dbuf_whichblock(dnode_t *dn, uint64_t offset)
474 {
475 if (dn->dn_datablkshift) {
476 return (offset >> dn->dn_datablkshift);
477 } else {
478 ASSERT3U(offset, <, dn->dn_datablksz);
479 return (0);
480 }
481 }
482
483 static void
484 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
485 {
486 dmu_buf_impl_t *db = vdb;
487
488 mutex_enter(&db->db_mtx);
489 ASSERT3U(db->db_state, ==, DB_READ);
490 /*
491 * All reads are synchronous, so we must have a hold on the dbuf
492 */
493 ASSERT(refcount_count(&db->db_holds) > 0);
494 ASSERT(db->db_buf == NULL);
495 ASSERT(db->db.db_data == NULL);
496 if (db->db_level == 0 && db->db_freed_in_flight) {
497 /* we were freed in flight; disregard any error */
498 arc_release(buf, db);
499 bzero(buf->b_data, db->db.db_size);
500 arc_buf_freeze(buf);
501 db->db_freed_in_flight = FALSE;
502 dbuf_set_data(db, buf);
503 db->db_state = DB_CACHED;
504 } else if (zio == NULL || zio->io_error == 0) {
505 dbuf_set_data(db, buf);
506 db->db_state = DB_CACHED;
507 } else {
508 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
509 ASSERT3P(db->db_buf, ==, NULL);
510 VERIFY(arc_buf_remove_ref(buf, db));
511 db->db_state = DB_UNCACHED;
512 }
513 cv_broadcast(&db->db_changed);
514 dbuf_rele_and_unlock(db, NULL);
515 }
516
517 static void
518 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t *flags)
519 {
520 dnode_t *dn;
521 spa_t *spa;
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 dbuf_set_data(db, arc_buf_alloc(dn->dn_objset->os_spa,
562 db->db.db_size, db, type));
563 DB_DNODE_EXIT(db);
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 spa = dn->dn_objset->os_spa;
572 DB_DNODE_EXIT(db);
573
574 db->db_state = DB_READ;
575 mutex_exit(&db->db_mtx);
576
577 if (DBUF_IS_L2CACHEABLE(db))
578 aflags |= ARC_L2CACHE;
579 if (DBUF_IS_L2COMPRESSIBLE(db))
580 aflags |= ARC_L2COMPRESS;
581
582 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
583 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
584 db->db.db_object, db->db_level, db->db_blkid);
585
586 dbuf_add_ref(db, NULL);
587
588 (void) arc_read(zio, spa, db->db_blkptr,
589 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
590 (*flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
591 &aflags, &zb);
592 if (aflags & ARC_CACHED)
593 *flags |= DB_RF_CACHED;
594 }
595
596 int
597 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
598 {
599 int err = 0;
600 int havepzio = (zio != NULL);
601 int prefetch;
602 dnode_t *dn;
603
604 /*
605 * We don't have to hold the mutex to check db_state because it
606 * can't be freed while we have a hold on the buffer.
607 */
608 ASSERT(!refcount_is_zero(&db->db_holds));
609
610 if (db->db_state == DB_NOFILL)
611 return (SET_ERROR(EIO));
612
613 DB_DNODE_ENTER(db);
614 dn = DB_DNODE(db);
615 if ((flags & DB_RF_HAVESTRUCT) == 0)
616 rw_enter(&dn->dn_struct_rwlock, RW_READER);
617
618 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
619 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
620 DBUF_IS_CACHEABLE(db);
621
622 mutex_enter(&db->db_mtx);
623 if (db->db_state == DB_CACHED) {
624 mutex_exit(&db->db_mtx);
625 if (prefetch)
626 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
627 db->db.db_size, TRUE);
628 if ((flags & DB_RF_HAVESTRUCT) == 0)
629 rw_exit(&dn->dn_struct_rwlock);
630 DB_DNODE_EXIT(db);
631 } else if (db->db_state == DB_UNCACHED) {
632 spa_t *spa = dn->dn_objset->os_spa;
633
634 if (zio == NULL)
635 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
636 dbuf_read_impl(db, zio, &flags);
637
638 /* dbuf_read_impl has dropped db_mtx for us */
639
640 if (prefetch)
641 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
642 db->db.db_size, flags & DB_RF_CACHED);
643
644 if ((flags & DB_RF_HAVESTRUCT) == 0)
645 rw_exit(&dn->dn_struct_rwlock);
646 DB_DNODE_EXIT(db);
647
648 if (!havepzio)
649 err = zio_wait(zio);
650 } else {
651 /*
652 * Another reader came in while the dbuf was in flight
653 * between UNCACHED and CACHED. Either a writer will finish
654 * writing the buffer (sending the dbuf to CACHED) or the
655 * first reader's request will reach the read_done callback
656 * and send the dbuf to CACHED. Otherwise, a failure
657 * occurred and the dbuf went to UNCACHED.
658 */
659 mutex_exit(&db->db_mtx);
660 if (prefetch)
661 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
662 db->db.db_size, TRUE);
663 if ((flags & DB_RF_HAVESTRUCT) == 0)
664 rw_exit(&dn->dn_struct_rwlock);
665 DB_DNODE_EXIT(db);
666
667 /* Skip the wait per the caller's request. */
668 mutex_enter(&db->db_mtx);
669 if ((flags & DB_RF_NEVERWAIT) == 0) {
670 while (db->db_state == DB_READ ||
671 db->db_state == DB_FILL) {
672 ASSERT(db->db_state == DB_READ ||
673 (flags & DB_RF_HAVESTRUCT) == 0);
674 cv_wait(&db->db_changed, &db->db_mtx);
675 }
676 if (db->db_state == DB_UNCACHED)
677 err = SET_ERROR(EIO);
678 }
679 mutex_exit(&db->db_mtx);
680 }
681
682 ASSERT(err || havepzio || db->db_state == DB_CACHED);
683 return (err);
684 }
685
686 static void
687 dbuf_noread(dmu_buf_impl_t *db)
688 {
689 ASSERT(!refcount_is_zero(&db->db_holds));
690 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
691 mutex_enter(&db->db_mtx);
692 while (db->db_state == DB_READ || db->db_state == DB_FILL)
693 cv_wait(&db->db_changed, &db->db_mtx);
694 if (db->db_state == DB_UNCACHED) {
695 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
696 spa_t *spa;
697
698 ASSERT(db->db_buf == NULL);
699 ASSERT(db->db.db_data == NULL);
700 DB_GET_SPA(&spa, db);
701 dbuf_set_data(db, arc_buf_alloc(spa, db->db.db_size, db, type));
702 db->db_state = DB_FILL;
703 } else if (db->db_state == DB_NOFILL) {
704 dbuf_set_data(db, NULL);
705 } else {
706 ASSERT3U(db->db_state, ==, DB_CACHED);
707 }
708 mutex_exit(&db->db_mtx);
709 }
710
711 /*
712 * This is our just-in-time copy function. It makes a copy of
713 * buffers, that have been modified in a previous transaction
714 * group, before we modify them in the current active group.
715 *
716 * This function is used in two places: when we are dirtying a
717 * buffer for the first time in a txg, and when we are freeing
718 * a range in a dnode that includes this buffer.
719 *
720 * Note that when we are called from dbuf_free_range() we do
721 * not put a hold on the buffer, we just traverse the active
722 * dbuf list for the dnode.
723 */
724 static void
725 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
726 {
727 dbuf_dirty_record_t *dr = db->db_last_dirty;
728
729 ASSERT(MUTEX_HELD(&db->db_mtx));
730 ASSERT(db->db.db_data != NULL);
731 ASSERT(db->db_level == 0);
732 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
733
734 if (dr == NULL ||
735 (dr->dt.dl.dr_data !=
736 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
737 return;
738
739 /*
740 * If the last dirty record for this dbuf has not yet synced
741 * and its referencing the dbuf data, either:
742 * reset the reference to point to a new copy,
743 * or (if there a no active holders)
744 * just null out the current db_data pointer.
745 */
746 ASSERT(dr->dr_txg >= txg - 2);
747 if (db->db_blkid == DMU_BONUS_BLKID) {
748 /* Note that the data bufs here are zio_bufs */
749 dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
750 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
751 bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
752 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
753 int size = db->db.db_size;
754 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
755 spa_t *spa;
756
757 DB_GET_SPA(&spa, db);
758 dr->dt.dl.dr_data = arc_buf_alloc(spa, size, db, type);
759 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
760 } else {
761 dbuf_set_data(db, NULL);
762 }
763 }
764
765 void
766 dbuf_unoverride(dbuf_dirty_record_t *dr)
767 {
768 dmu_buf_impl_t *db = dr->dr_dbuf;
769 blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
770 uint64_t txg = dr->dr_txg;
771
772 ASSERT(MUTEX_HELD(&db->db_mtx));
773 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
774 ASSERT(db->db_level == 0);
775
776 if (db->db_blkid == DMU_BONUS_BLKID ||
777 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
778 return;
779
780 ASSERT(db->db_data_pending != dr);
781
782 /* free this block */
783 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite) {
784 spa_t *spa;
785
786 DB_GET_SPA(&spa, db);
787 zio_free(spa, txg, bp);
788 }
789 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
790 dr->dt.dl.dr_nopwrite = B_FALSE;
791
792 /*
793 * Release the already-written buffer, so we leave it in
794 * a consistent dirty state. Note that all callers are
795 * modifying the buffer, so they will immediately do
796 * another (redundant) arc_release(). Therefore, leave
797 * the buf thawed to save the effort of freezing &
798 * immediately re-thawing it.
799 */
800 arc_release(dr->dt.dl.dr_data, db);
801 }
802
803 /*
804 * Evict (if its unreferenced) or clear (if its referenced) any level-0
805 * data blocks in the free range, so that any future readers will find
806 * empty blocks. Also, if we happen across any level-1 dbufs in the
807 * range that have not already been marked dirty, mark them dirty so
808 * they stay in memory.
809 *
810 * This is a no-op if the dataset is in the middle of an incremental
811 * receive; see comment below for details.
812 */
813 void
814 dbuf_free_range(dnode_t *dn, uint64_t start, uint64_t end, dmu_tx_t *tx)
815 {
816 dmu_buf_impl_t *db, *db_next;
817 uint64_t txg = tx->tx_txg;
818 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
819 uint64_t first_l1 = start >> epbs;
820 uint64_t last_l1 = end >> epbs;
821
822 if (end > dn->dn_maxblkid && (end != DMU_SPILL_BLKID)) {
823 end = dn->dn_maxblkid;
824 last_l1 = end >> epbs;
825 }
826 dprintf_dnode(dn, "start=%llu end=%llu\n", start, end);
827
828 mutex_enter(&dn->dn_dbufs_mtx);
829 if (start >= dn->dn_unlisted_l0_blkid * dn->dn_datablksz) {
830 /* There can't be any dbufs in this range; no need to search. */
831 mutex_exit(&dn->dn_dbufs_mtx);
832 return;
833 } else if (dmu_objset_is_receiving(dn->dn_objset)) {
834 /*
835 * If we are receiving, we expect there to be no dbufs in
836 * the range to be freed, because receive modifies each
837 * block at most once, and in offset order. If this is
838 * not the case, it can lead to performance problems,
839 * so note that we unexpectedly took the slow path.
840 */
841 atomic_inc_64(&zfs_free_range_recv_miss);
842 }
843
844 for (db = list_head(&dn->dn_dbufs); db != NULL; db = db_next) {
845 db_next = list_next(&dn->dn_dbufs, db);
846 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
847
848 if (db->db_level == 1 &&
849 db->db_blkid >= first_l1 && db->db_blkid <= last_l1) {
850 mutex_enter(&db->db_mtx);
851 if (db->db_last_dirty &&
852 db->db_last_dirty->dr_txg < txg) {
853 dbuf_add_ref(db, FTAG);
854 mutex_exit(&db->db_mtx);
855 dbuf_will_dirty(db, tx);
856 dbuf_rele(db, FTAG);
857 } else {
858 mutex_exit(&db->db_mtx);
859 }
860 }
861
862 if (db->db_level != 0)
863 continue;
864 dprintf_dbuf(db, "found buf %s\n", "");
865 if (db->db_blkid < start || db->db_blkid > end)
866 continue;
867
868 /* found a level 0 buffer in the range */
869 mutex_enter(&db->db_mtx);
870 if (dbuf_undirty(db, tx)) {
871 /* mutex has been dropped and dbuf destroyed */
872 continue;
873 }
874
875 if (db->db_state == DB_UNCACHED ||
876 db->db_state == DB_NOFILL ||
877 db->db_state == DB_EVICTING) {
878 ASSERT(db->db.db_data == NULL);
879 mutex_exit(&db->db_mtx);
880 continue;
881 }
882 if (db->db_state == DB_READ || db->db_state == DB_FILL) {
883 /* will be handled in dbuf_read_done or dbuf_rele */
884 db->db_freed_in_flight = TRUE;
885 mutex_exit(&db->db_mtx);
886 continue;
887 }
888 if (refcount_count(&db->db_holds) == 0) {
889 ASSERT(db->db_buf);
890 dbuf_clear(db);
891 continue;
892 }
893 /* The dbuf is referenced */
894
895 if (db->db_last_dirty != NULL) {
896 dbuf_dirty_record_t *dr = db->db_last_dirty;
897
898 if (dr->dr_txg == txg) {
899 /*
900 * This buffer is "in-use", re-adjust the file
901 * size to reflect that this buffer may
902 * contain new data when we sync.
903 */
904 if (db->db_blkid != DMU_SPILL_BLKID &&
905 db->db_blkid > dn->dn_maxblkid)
906 dn->dn_maxblkid = db->db_blkid;
907 dbuf_unoverride(dr);
908 } else {
909 /*
910 * This dbuf is not dirty in the open context.
911 * Either uncache it (if its not referenced in
912 * the open context) or reset its contents to
913 * empty.
914 */
915 dbuf_fix_old_data(db, txg);
916 }
917 }
918 /* clear the contents if its cached */
919 if (db->db_state == DB_CACHED) {
920 ASSERT(db->db.db_data != NULL);
921 arc_release(db->db_buf, db);
922 bzero(db->db.db_data, db->db.db_size);
923 arc_buf_freeze(db->db_buf);
924 }
925
926 mutex_exit(&db->db_mtx);
927 }
928 mutex_exit(&dn->dn_dbufs_mtx);
929 }
930
931 static int
932 dbuf_block_freeable(dmu_buf_impl_t *db)
933 {
934 dsl_dataset_t *ds = db->db_objset->os_dsl_dataset;
935 uint64_t birth_txg = 0;
936
937 /*
938 * We don't need any locking to protect db_blkptr:
939 * If it's syncing, then db_last_dirty will be set
940 * so we'll ignore db_blkptr.
941 */
942 ASSERT(MUTEX_HELD(&db->db_mtx));
943 if (db->db_last_dirty)
944 birth_txg = db->db_last_dirty->dr_txg;
945 else if (db->db_blkptr)
946 birth_txg = db->db_blkptr->blk_birth;
947
948 /*
949 * If we don't exist or are in a snapshot, we can't be freed.
950 * Don't pass the bp to dsl_dataset_block_freeable() since we
951 * are holding the db_mtx lock and might deadlock if we are
952 * prefetching a dedup-ed block.
953 */
954 if (birth_txg)
955 return (ds == NULL ||
956 dsl_dataset_block_freeable(ds, NULL, birth_txg));
957 else
958 return (FALSE);
959 }
960
961 void
962 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
963 {
964 arc_buf_t *buf, *obuf;
965 int osize = db->db.db_size;
966 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
967 dnode_t *dn;
968
969 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
970
971 DB_DNODE_ENTER(db);
972 dn = DB_DNODE(db);
973
974 /* XXX does *this* func really need the lock? */
975 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
976
977 /*
978 * This call to dbuf_will_dirty() with the dn_struct_rwlock held
979 * is OK, because there can be no other references to the db
980 * when we are changing its size, so no concurrent DB_FILL can
981 * be happening.
982 */
983 /*
984 * XXX we should be doing a dbuf_read, checking the return
985 * value and returning that up to our callers
986 */
987 dbuf_will_dirty(db, tx);
988
989 /* create the data buffer for the new block */
990 buf = arc_buf_alloc(dn->dn_objset->os_spa, size, db, type);
991
992 /* copy old block data to the new block */
993 obuf = db->db_buf;
994 bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
995 /* zero the remainder */
996 if (size > osize)
997 bzero((uint8_t *)buf->b_data + osize, size - osize);
998
999 mutex_enter(&db->db_mtx);
1000 dbuf_set_data(db, buf);
1001 VERIFY(arc_buf_remove_ref(obuf, db));
1002 db->db.db_size = size;
1003
1004 if (db->db_level == 0) {
1005 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1006 db->db_last_dirty->dt.dl.dr_data = buf;
1007 }
1008 mutex_exit(&db->db_mtx);
1009
1010 dnode_willuse_space(dn, size-osize, tx);
1011 DB_DNODE_EXIT(db);
1012 }
1013
1014 void
1015 dbuf_release_bp(dmu_buf_impl_t *db)
1016 {
1017 objset_t *os;
1018
1019 DB_GET_OBJSET(&os, db);
1020 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1021 ASSERT(arc_released(os->os_phys_buf) ||
1022 list_link_active(&os->os_dsl_dataset->ds_synced_link));
1023 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1024
1025 (void) arc_release(db->db_buf, db);
1026 }
1027
1028 dbuf_dirty_record_t *
1029 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1030 {
1031 dnode_t *dn;
1032 objset_t *os;
1033 dbuf_dirty_record_t **drp, *dr;
1034 int drop_struct_lock = FALSE;
1035 boolean_t do_free_accounting = B_FALSE;
1036 int txgoff = tx->tx_txg & TXG_MASK;
1037
1038 ASSERT(tx->tx_txg != 0);
1039 ASSERT(!refcount_is_zero(&db->db_holds));
1040 DMU_TX_DIRTY_BUF(tx, db);
1041
1042 DB_DNODE_ENTER(db);
1043 dn = DB_DNODE(db);
1044 /*
1045 * Shouldn't dirty a regular buffer in syncing context. Private
1046 * objects may be dirtied in syncing context, but only if they
1047 * were already pre-dirtied in open context.
1048 */
1049 ASSERT(!dmu_tx_is_syncing(tx) ||
1050 BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1051 DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1052 dn->dn_objset->os_dsl_dataset == NULL);
1053 /*
1054 * We make this assert for private objects as well, but after we
1055 * check if we're already dirty. They are allowed to re-dirty
1056 * in syncing context.
1057 */
1058 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1059 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1060 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1061
1062 mutex_enter(&db->db_mtx);
1063 /*
1064 * XXX make this true for indirects too? The problem is that
1065 * transactions created with dmu_tx_create_assigned() from
1066 * syncing context don't bother holding ahead.
1067 */
1068 ASSERT(db->db_level != 0 ||
1069 db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1070 db->db_state == DB_NOFILL);
1071
1072 mutex_enter(&dn->dn_mtx);
1073 /*
1074 * Don't set dirtyctx to SYNC if we're just modifying this as we
1075 * initialize the objset.
1076 */
1077 if (dn->dn_dirtyctx == DN_UNDIRTIED &&
1078 !BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1079 dn->dn_dirtyctx =
1080 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1081 ASSERT(dn->dn_dirtyctx_firstset == NULL);
1082 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1083 }
1084 mutex_exit(&dn->dn_mtx);
1085
1086 if (db->db_blkid == DMU_SPILL_BLKID)
1087 dn->dn_have_spill = B_TRUE;
1088
1089 /*
1090 * If this buffer is already dirty, we're done.
1091 */
1092 drp = &db->db_last_dirty;
1093 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1094 db->db.db_object == DMU_META_DNODE_OBJECT);
1095 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1096 drp = &dr->dr_next;
1097 if (dr && dr->dr_txg == tx->tx_txg) {
1098 DB_DNODE_EXIT(db);
1099
1100 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1101 /*
1102 * If this buffer has already been written out,
1103 * we now need to reset its state.
1104 */
1105 dbuf_unoverride(dr);
1106 if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1107 db->db_state != DB_NOFILL)
1108 arc_buf_thaw(db->db_buf);
1109 }
1110 mutex_exit(&db->db_mtx);
1111 return (dr);
1112 }
1113
1114 /*
1115 * Only valid if not already dirty.
1116 */
1117 ASSERT(dn->dn_object == 0 ||
1118 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1119 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1120
1121 ASSERT3U(dn->dn_nlevels, >, db->db_level);
1122 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1123 dn->dn_phys->dn_nlevels > db->db_level ||
1124 dn->dn_next_nlevels[txgoff] > db->db_level ||
1125 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1126 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1127
1128 /*
1129 * We should only be dirtying in syncing context if it's the
1130 * mos or we're initializing the os or it's a special object.
1131 * However, we are allowed to dirty in syncing context provided
1132 * we already dirtied it in open context. Hence we must make
1133 * this assertion only if we're not already dirty.
1134 */
1135 os = dn->dn_objset;
1136 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1137 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1138 ASSERT(db->db.db_size != 0);
1139
1140 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1141
1142 if (db->db_blkid != DMU_BONUS_BLKID) {
1143 /*
1144 * Update the accounting.
1145 * Note: we delay "free accounting" until after we drop
1146 * the db_mtx. This keeps us from grabbing other locks
1147 * (and possibly deadlocking) in bp_get_dsize() while
1148 * also holding the db_mtx.
1149 */
1150 dnode_willuse_space(dn, db->db.db_size, tx);
1151 do_free_accounting = dbuf_block_freeable(db);
1152 }
1153
1154 /*
1155 * If this buffer is dirty in an old transaction group we need
1156 * to make a copy of it so that the changes we make in this
1157 * transaction group won't leak out when we sync the older txg.
1158 */
1159 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1160 if (db->db_level == 0) {
1161 void *data_old = db->db_buf;
1162
1163 if (db->db_state != DB_NOFILL) {
1164 if (db->db_blkid == DMU_BONUS_BLKID) {
1165 dbuf_fix_old_data(db, tx->tx_txg);
1166 data_old = db->db.db_data;
1167 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1168 /*
1169 * Release the data buffer from the cache so
1170 * that we can modify it without impacting
1171 * possible other users of this cached data
1172 * block. Note that indirect blocks and
1173 * private objects are not released until the
1174 * syncing state (since they are only modified
1175 * then).
1176 */
1177 arc_release(db->db_buf, db);
1178 dbuf_fix_old_data(db, tx->tx_txg);
1179 data_old = db->db_buf;
1180 }
1181 ASSERT(data_old != NULL);
1182 }
1183 dr->dt.dl.dr_data = data_old;
1184 } else {
1185 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1186 list_create(&dr->dt.di.dr_children,
1187 sizeof (dbuf_dirty_record_t),
1188 offsetof(dbuf_dirty_record_t, dr_dirty_node));
1189 }
1190 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL)
1191 dr->dr_accounted = db->db.db_size;
1192 dr->dr_dbuf = db;
1193 dr->dr_txg = tx->tx_txg;
1194 dr->dr_next = *drp;
1195 *drp = dr;
1196
1197 /*
1198 * We could have been freed_in_flight between the dbuf_noread
1199 * and dbuf_dirty. We win, as though the dbuf_noread() had
1200 * happened after the free.
1201 */
1202 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1203 db->db_blkid != DMU_SPILL_BLKID) {
1204 mutex_enter(&dn->dn_mtx);
1205 dnode_clear_range(dn, db->db_blkid, 1, tx);
1206 mutex_exit(&dn->dn_mtx);
1207 db->db_freed_in_flight = FALSE;
1208 }
1209
1210 /*
1211 * This buffer is now part of this txg
1212 */
1213 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1214 db->db_dirtycnt += 1;
1215 ASSERT3U(db->db_dirtycnt, <=, 3);
1216
1217 mutex_exit(&db->db_mtx);
1218
1219 if (db->db_blkid == DMU_BONUS_BLKID ||
1220 db->db_blkid == DMU_SPILL_BLKID) {
1221 mutex_enter(&dn->dn_mtx);
1222 ASSERT(!list_link_active(&dr->dr_dirty_node));
1223 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1224 mutex_exit(&dn->dn_mtx);
1225 dnode_setdirty(dn, tx);
1226 DB_DNODE_EXIT(db);
1227 return (dr);
1228 } else if (do_free_accounting) {
1229 blkptr_t *bp = db->db_blkptr;
1230 int64_t willfree = (bp && !BP_IS_HOLE(bp)) ?
1231 bp_get_dsize(os->os_spa, bp) : db->db.db_size;
1232 /*
1233 * This is only a guess -- if the dbuf is dirty
1234 * in a previous txg, we don't know how much
1235 * space it will use on disk yet. We should
1236 * really have the struct_rwlock to access
1237 * db_blkptr, but since this is just a guess,
1238 * it's OK if we get an odd answer.
1239 */
1240 ddt_prefetch(os->os_spa, bp);
1241 dnode_willuse_space(dn, -willfree, tx);
1242 }
1243
1244 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1245 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1246 drop_struct_lock = TRUE;
1247 }
1248
1249 if (db->db_level == 0) {
1250 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1251 ASSERT(dn->dn_maxblkid >= db->db_blkid);
1252 }
1253
1254 if (db->db_level+1 < dn->dn_nlevels) {
1255 dmu_buf_impl_t *parent = db->db_parent;
1256 dbuf_dirty_record_t *di;
1257 int parent_held = FALSE;
1258
1259 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1260 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1261
1262 parent = dbuf_hold_level(dn, db->db_level+1,
1263 db->db_blkid >> epbs, FTAG);
1264 ASSERT(parent != NULL);
1265 parent_held = TRUE;
1266 }
1267 if (drop_struct_lock)
1268 rw_exit(&dn->dn_struct_rwlock);
1269 ASSERT3U(db->db_level+1, ==, parent->db_level);
1270 di = dbuf_dirty(parent, tx);
1271 if (parent_held)
1272 dbuf_rele(parent, FTAG);
1273
1274 mutex_enter(&db->db_mtx);
1275 /*
1276 * Since we've dropped the mutex, it's possible that
1277 * dbuf_undirty() might have changed this out from under us.
1278 */
1279 if (db->db_last_dirty == dr ||
1280 dn->dn_object == DMU_META_DNODE_OBJECT) {
1281 mutex_enter(&di->dt.di.dr_mtx);
1282 ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1283 ASSERT(!list_link_active(&dr->dr_dirty_node));
1284 list_insert_tail(&di->dt.di.dr_children, dr);
1285 mutex_exit(&di->dt.di.dr_mtx);
1286 dr->dr_parent = di;
1287 }
1288 mutex_exit(&db->db_mtx);
1289 } else {
1290 ASSERT(db->db_level+1 == dn->dn_nlevels);
1291 ASSERT(db->db_blkid < dn->dn_nblkptr);
1292 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1293 mutex_enter(&dn->dn_mtx);
1294 ASSERT(!list_link_active(&dr->dr_dirty_node));
1295 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1296 mutex_exit(&dn->dn_mtx);
1297 if (drop_struct_lock)
1298 rw_exit(&dn->dn_struct_rwlock);
1299 }
1300
1301 dnode_setdirty(dn, tx);
1302 DB_DNODE_EXIT(db);
1303 return (dr);
1304 }
1305
1306 /*
1307 * Undirty a buffer in the transaction group referenced by the given
1308 * transaction. Return whether this evicted the dbuf.
1309 */
1310 static boolean_t
1311 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1312 {
1313 dnode_t *dn;
1314 uint64_t txg = tx->tx_txg;
1315 dbuf_dirty_record_t *dr, **drp;
1316
1317 ASSERT(txg != 0);
1318 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1319 ASSERT0(db->db_level);
1320 ASSERT(MUTEX_HELD(&db->db_mtx));
1321
1322 /*
1323 * If this buffer is not dirty, we're done.
1324 */
1325 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1326 if (dr->dr_txg <= txg)
1327 break;
1328 if (dr == NULL || dr->dr_txg < txg)
1329 return (B_FALSE);
1330 ASSERT(dr->dr_txg == txg);
1331 ASSERT(dr->dr_dbuf == db);
1332
1333 DB_DNODE_ENTER(db);
1334 dn = DB_DNODE(db);
1335
1336 /*
1337 * Note: This code will probably work even if there are concurrent
1338 * holders, but it is untested in that scenerio, as the ZPL and
1339 * ztest have additional locking (the range locks) that prevents
1340 * that type of concurrent access.
1341 */
1342 ASSERT3U(refcount_count(&db->db_holds), ==, db->db_dirtycnt);
1343
1344 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1345
1346 ASSERT(db->db.db_size != 0);
1347
1348 /*
1349 * Any space we accounted for in dp_dirty_* will be cleaned up by
1350 * dsl_pool_sync(). This is relatively rare so the discrepancy
1351 * is not a big deal.
1352 */
1353
1354 *drp = dr->dr_next;
1355
1356 /*
1357 * Note that there are three places in dbuf_dirty()
1358 * where this dirty record may be put on a list.
1359 * Make sure to do a list_remove corresponding to
1360 * every one of those list_insert calls.
1361 */
1362 if (dr->dr_parent) {
1363 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1364 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1365 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1366 } else if (db->db_blkid == DMU_SPILL_BLKID ||
1367 db->db_level+1 == dn->dn_nlevels) {
1368 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1369 mutex_enter(&dn->dn_mtx);
1370 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1371 mutex_exit(&dn->dn_mtx);
1372 }
1373 DB_DNODE_EXIT(db);
1374
1375 if (db->db_state != DB_NOFILL) {
1376 dbuf_unoverride(dr);
1377
1378 ASSERT(db->db_buf != NULL);
1379 ASSERT(dr->dt.dl.dr_data != NULL);
1380 if (dr->dt.dl.dr_data != db->db_buf)
1381 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db));
1382 }
1383 kmem_free(dr, sizeof (dbuf_dirty_record_t));
1384
1385 ASSERT(db->db_dirtycnt > 0);
1386 db->db_dirtycnt -= 1;
1387
1388 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1389 arc_buf_t *buf = db->db_buf;
1390
1391 ASSERT(db->db_state == DB_NOFILL || arc_released(buf));
1392 dbuf_set_data(db, NULL);
1393 VERIFY(arc_buf_remove_ref(buf, db));
1394 dbuf_evict(db);
1395 return (B_TRUE);
1396 }
1397
1398 return (B_FALSE);
1399 }
1400
1401 #pragma weak dmu_buf_will_dirty = dbuf_will_dirty
1402 void
1403 dbuf_will_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1404 {
1405 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1406
1407 ASSERT(tx->tx_txg != 0);
1408 ASSERT(!refcount_is_zero(&db->db_holds));
1409
1410 DB_DNODE_ENTER(db);
1411 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
1412 rf |= DB_RF_HAVESTRUCT;
1413 DB_DNODE_EXIT(db);
1414 (void) dbuf_read(db, NULL, rf);
1415 (void) dbuf_dirty(db, tx);
1416 }
1417
1418 void
1419 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1420 {
1421 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1422
1423 db->db_state = DB_NOFILL;
1424
1425 dmu_buf_will_fill(db_fake, tx);
1426 }
1427
1428 void
1429 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1430 {
1431 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1432
1433 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1434 ASSERT(tx->tx_txg != 0);
1435 ASSERT(db->db_level == 0);
1436 ASSERT(!refcount_is_zero(&db->db_holds));
1437
1438 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1439 dmu_tx_private_ok(tx));
1440
1441 dbuf_noread(db);
1442 (void) dbuf_dirty(db, tx);
1443 }
1444
1445 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1446 /* ARGSUSED */
1447 void
1448 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1449 {
1450 mutex_enter(&db->db_mtx);
1451 DBUF_VERIFY(db);
1452
1453 if (db->db_state == DB_FILL) {
1454 if (db->db_level == 0 && db->db_freed_in_flight) {
1455 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1456 /* we were freed while filling */
1457 /* XXX dbuf_undirty? */
1458 bzero(db->db.db_data, db->db.db_size);
1459 db->db_freed_in_flight = FALSE;
1460 }
1461 db->db_state = DB_CACHED;
1462 cv_broadcast(&db->db_changed);
1463 }
1464 mutex_exit(&db->db_mtx);
1465 }
1466
1467 /*
1468 * Directly assign a provided arc buf to a given dbuf if it's not referenced
1469 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1470 */
1471 void
1472 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
1473 {
1474 ASSERT(!refcount_is_zero(&db->db_holds));
1475 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1476 ASSERT(db->db_level == 0);
1477 ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA);
1478 ASSERT(buf != NULL);
1479 ASSERT(arc_buf_size(buf) == db->db.db_size);
1480 ASSERT(tx->tx_txg != 0);
1481
1482 arc_return_buf(buf, db);
1483 ASSERT(arc_released(buf));
1484
1485 mutex_enter(&db->db_mtx);
1486
1487 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1488 cv_wait(&db->db_changed, &db->db_mtx);
1489
1490 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
1491
1492 if (db->db_state == DB_CACHED &&
1493 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
1494 mutex_exit(&db->db_mtx);
1495 (void) dbuf_dirty(db, tx);
1496 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
1497 VERIFY(arc_buf_remove_ref(buf, db));
1498 xuio_stat_wbuf_copied();
1499 return;
1500 }
1501
1502 xuio_stat_wbuf_nocopy();
1503 if (db->db_state == DB_CACHED) {
1504 dbuf_dirty_record_t *dr = db->db_last_dirty;
1505
1506 ASSERT(db->db_buf != NULL);
1507 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
1508 ASSERT(dr->dt.dl.dr_data == db->db_buf);
1509 if (!arc_released(db->db_buf)) {
1510 ASSERT(dr->dt.dl.dr_override_state ==
1511 DR_OVERRIDDEN);
1512 arc_release(db->db_buf, db);
1513 }
1514 dr->dt.dl.dr_data = buf;
1515 VERIFY(arc_buf_remove_ref(db->db_buf, db));
1516 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
1517 arc_release(db->db_buf, db);
1518 VERIFY(arc_buf_remove_ref(db->db_buf, db));
1519 }
1520 db->db_buf = NULL;
1521 }
1522 ASSERT(db->db_buf == NULL);
1523 dbuf_set_data(db, buf);
1524 db->db_state = DB_FILL;
1525 mutex_exit(&db->db_mtx);
1526 (void) dbuf_dirty(db, tx);
1527 dbuf_fill_done(db, tx);
1528 }
1529
1530 /*
1531 * "Clear" the contents of this dbuf. This will mark the dbuf
1532 * EVICTING and clear *most* of its references. Unfortunately,
1533 * when we are not holding the dn_dbufs_mtx, we can't clear the
1534 * entry in the dn_dbufs list. We have to wait until dbuf_destroy()
1535 * in this case. For callers from the DMU we will usually see:
1536 * dbuf_clear()->arc_buf_evict()->dbuf_do_evict()->dbuf_destroy()
1537 * For the arc callback, we will usually see:
1538 * dbuf_do_evict()->dbuf_clear();dbuf_destroy()
1539 * Sometimes, though, we will get a mix of these two:
1540 * DMU: dbuf_clear()->arc_buf_evict()
1541 * ARC: dbuf_do_evict()->dbuf_destroy()
1542 */
1543 void
1544 dbuf_clear(dmu_buf_impl_t *db)
1545 {
1546 dnode_t *dn;
1547 dmu_buf_impl_t *parent = db->db_parent;
1548 dmu_buf_impl_t *dndb;
1549 int dbuf_gone = FALSE;
1550
1551 ASSERT(MUTEX_HELD(&db->db_mtx));
1552 ASSERT(refcount_is_zero(&db->db_holds));
1553
1554 dbuf_evict_user(db);
1555
1556 if (db->db_state == DB_CACHED) {
1557 ASSERT(db->db.db_data != NULL);
1558 if (db->db_blkid == DMU_BONUS_BLKID) {
1559 zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
1560 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
1561 }
1562 db->db.db_data = NULL;
1563 db->db_state = DB_UNCACHED;
1564 }
1565
1566 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
1567 ASSERT(db->db_data_pending == NULL);
1568
1569 db->db_state = DB_EVICTING;
1570 db->db_blkptr = NULL;
1571
1572 DB_DNODE_ENTER(db);
1573 dn = DB_DNODE(db);
1574 dndb = dn->dn_dbuf;
1575 if (db->db_blkid != DMU_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) {
1576 list_remove(&dn->dn_dbufs, db);
1577 (void) atomic_dec_32_nv(&dn->dn_dbufs_count);
1578 membar_producer();
1579 DB_DNODE_EXIT(db);
1580 /*
1581 * Decrementing the dbuf count means that the hold corresponding
1582 * to the removed dbuf is no longer discounted in dnode_move(),
1583 * so the dnode cannot be moved until after we release the hold.
1584 * The membar_producer() ensures visibility of the decremented
1585 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
1586 * release any lock.
1587 */
1588 dnode_rele(dn, db);
1589 db->db_dnode_handle = NULL;
1590 } else {
1591 DB_DNODE_EXIT(db);
1592 }
1593
1594 if (db->db_buf)
1595 dbuf_gone = arc_buf_evict(db->db_buf);
1596
1597 if (!dbuf_gone)
1598 mutex_exit(&db->db_mtx);
1599
1600 /*
1601 * If this dbuf is referenced from an indirect dbuf,
1602 * decrement the ref count on the indirect dbuf.
1603 */
1604 if (parent && parent != dndb)
1605 dbuf_rele(parent, db);
1606 }
1607
1608 static int
1609 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
1610 dmu_buf_impl_t **parentp, blkptr_t **bpp)
1611 {
1612 int nlevels, epbs;
1613
1614 *parentp = NULL;
1615 *bpp = NULL;
1616
1617 ASSERT(blkid != DMU_BONUS_BLKID);
1618
1619 if (blkid == DMU_SPILL_BLKID) {
1620 mutex_enter(&dn->dn_mtx);
1621 if (dn->dn_have_spill &&
1622 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
1623 *bpp = &dn->dn_phys->dn_spill;
1624 else
1625 *bpp = NULL;
1626 dbuf_add_ref(dn->dn_dbuf, NULL);
1627 *parentp = dn->dn_dbuf;
1628 mutex_exit(&dn->dn_mtx);
1629 return (0);
1630 }
1631
1632 if (dn->dn_phys->dn_nlevels == 0)
1633 nlevels = 1;
1634 else
1635 nlevels = dn->dn_phys->dn_nlevels;
1636
1637 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1638
1639 ASSERT3U(level * epbs, <, 64);
1640 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1641 if (level >= nlevels ||
1642 (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
1643 /* the buffer has no parent yet */
1644 return (SET_ERROR(ENOENT));
1645 } else if (level < nlevels-1) {
1646 /* this block is referenced from an indirect block */
1647 int err = dbuf_hold_impl(dn, level+1,
1648 blkid >> epbs, fail_sparse, NULL, parentp);
1649 if (err)
1650 return (err);
1651 err = dbuf_read(*parentp, NULL,
1652 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
1653 if (err) {
1654 dbuf_rele(*parentp, NULL);
1655 *parentp = NULL;
1656 return (err);
1657 }
1658 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
1659 (blkid & ((1ULL << epbs) - 1));
1660 return (0);
1661 } else {
1662 /* the block is referenced from the dnode */
1663 ASSERT3U(level, ==, nlevels-1);
1664 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
1665 blkid < dn->dn_phys->dn_nblkptr);
1666 if (dn->dn_dbuf) {
1667 dbuf_add_ref(dn->dn_dbuf, NULL);
1668 *parentp = dn->dn_dbuf;
1669 }
1670 *bpp = &dn->dn_phys->dn_blkptr[blkid];
1671 return (0);
1672 }
1673 }
1674
1675 static dmu_buf_impl_t *
1676 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
1677 dmu_buf_impl_t *parent, blkptr_t *blkptr)
1678 {
1679 objset_t *os = dn->dn_objset;
1680 dmu_buf_impl_t *db, *odb;
1681
1682 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1683 ASSERT(dn->dn_type != DMU_OT_NONE);
1684
1685 db = kmem_cache_alloc(dbuf_cache, KM_SLEEP);
1686
1687 db->db_objset = os;
1688 db->db.db_object = dn->dn_object;
1689 db->db_level = level;
1690 db->db_blkid = blkid;
1691 db->db_last_dirty = NULL;
1692 db->db_dirtycnt = 0;
1693 db->db_dnode_handle = dn->dn_handle;
1694 db->db_parent = parent;
1695 db->db_blkptr = blkptr;
1696
1697 db->db_user_ptr = NULL;
1698 db->db_user_data_ptr_ptr = NULL;
1699 db->db_evict_func = NULL;
1700 db->db_immediate_evict = 0;
1701 db->db_freed_in_flight = 0;
1702
1703 if (blkid == DMU_BONUS_BLKID) {
1704 ASSERT3P(parent, ==, dn->dn_dbuf);
1705 db->db.db_size = DN_MAX_BONUSLEN -
1706 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
1707 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
1708 db->db.db_offset = DMU_BONUS_BLKID;
1709 db->db_state = DB_UNCACHED;
1710 /* the bonus dbuf is not placed in the hash table */
1711 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1712 return (db);
1713 } else if (blkid == DMU_SPILL_BLKID) {
1714 db->db.db_size = (blkptr != NULL) ?
1715 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
1716 db->db.db_offset = 0;
1717 } else {
1718 int blocksize =
1719 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
1720 db->db.db_size = blocksize;
1721 db->db.db_offset = db->db_blkid * blocksize;
1722 }
1723
1724 /*
1725 * Hold the dn_dbufs_mtx while we get the new dbuf
1726 * in the hash table *and* added to the dbufs list.
1727 * This prevents a possible deadlock with someone
1728 * trying to look up this dbuf before its added to the
1729 * dn_dbufs list.
1730 */
1731 mutex_enter(&dn->dn_dbufs_mtx);
1732 db->db_state = DB_EVICTING;
1733 if ((odb = dbuf_hash_insert(db)) != NULL) {
1734 /* someone else inserted it first */
1735 kmem_cache_free(dbuf_cache, db);
1736 mutex_exit(&dn->dn_dbufs_mtx);
1737 return (odb);
1738 }
1739 list_insert_head(&dn->dn_dbufs, db);
1740 if (db->db_level == 0 && db->db_blkid >=
1741 dn->dn_unlisted_l0_blkid)
1742 dn->dn_unlisted_l0_blkid = db->db_blkid + 1;
1743 db->db_state = DB_UNCACHED;
1744 mutex_exit(&dn->dn_dbufs_mtx);
1745 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1746
1747 if (parent && parent != dn->dn_dbuf)
1748 dbuf_add_ref(parent, db);
1749
1750 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1751 refcount_count(&dn->dn_holds) > 0);
1752 (void) refcount_add(&dn->dn_holds, db);
1753 (void) atomic_inc_32_nv(&dn->dn_dbufs_count);
1754
1755 dprintf_dbuf(db, "db=%p\n", db);
1756
1757 return (db);
1758 }
1759
1760 static int
1761 dbuf_do_evict(void *private)
1762 {
1763 arc_buf_t *buf = private;
1764 dmu_buf_impl_t *db = buf->b_private;
1765
1766 if (!MUTEX_HELD(&db->db_mtx))
1767 mutex_enter(&db->db_mtx);
1768
1769 ASSERT(refcount_is_zero(&db->db_holds));
1770
1771 if (db->db_state != DB_EVICTING) {
1772 ASSERT(db->db_state == DB_CACHED);
1773 DBUF_VERIFY(db);
1774 db->db_buf = NULL;
1775 dbuf_evict(db);
1776 } else {
1777 mutex_exit(&db->db_mtx);
1778 dbuf_destroy(db);
1779 }
1780 return (0);
1781 }
1782
1783 static void
1784 dbuf_destroy(dmu_buf_impl_t *db)
1785 {
1786 ASSERT(refcount_is_zero(&db->db_holds));
1787
1788 if (db->db_blkid != DMU_BONUS_BLKID) {
1789 /*
1790 * If this dbuf is still on the dn_dbufs list,
1791 * remove it from that list.
1792 */
1793 if (db->db_dnode_handle != NULL) {
1794 dnode_t *dn;
1795
1796 DB_DNODE_ENTER(db);
1797 dn = DB_DNODE(db);
1798 mutex_enter(&dn->dn_dbufs_mtx);
1799 list_remove(&dn->dn_dbufs, db);
1800 (void) atomic_dec_32_nv(&dn->dn_dbufs_count);
1801 mutex_exit(&dn->dn_dbufs_mtx);
1802 DB_DNODE_EXIT(db);
1803 /*
1804 * Decrementing the dbuf count means that the hold
1805 * corresponding to the removed dbuf is no longer
1806 * discounted in dnode_move(), so the dnode cannot be
1807 * moved until after we release the hold.
1808 */
1809 dnode_rele(dn, db);
1810 db->db_dnode_handle = NULL;
1811 }
1812 dbuf_hash_remove(db);
1813 }
1814 db->db_parent = NULL;
1815 db->db_buf = NULL;
1816
1817 ASSERT(!list_link_active(&db->db_link));
1818 ASSERT(db->db.db_data == NULL);
1819 ASSERT(db->db_hash_next == NULL);
1820 ASSERT(db->db_blkptr == NULL);
1821 ASSERT(db->db_data_pending == NULL);
1822
1823 kmem_cache_free(dbuf_cache, db);
1824 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1825 }
1826
1827 void
1828 dbuf_prefetch(dnode_t *dn, uint64_t blkid, zio_priority_t prio)
1829 {
1830 dmu_buf_impl_t *db = NULL;
1831 blkptr_t *bp = NULL;
1832
1833 ASSERT(blkid != DMU_BONUS_BLKID);
1834 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1835
1836 if (dnode_block_freed(dn, blkid))
1837 return;
1838
1839 /* dbuf_find() returns with db_mtx held */
1840 if (db = dbuf_find(dn, 0, blkid)) {
1841 /*
1842 * This dbuf is already in the cache. We assume that
1843 * it is already CACHED, or else about to be either
1844 * read or filled.
1845 */
1846 mutex_exit(&db->db_mtx);
1847 return;
1848 }
1849
1850 if (dbuf_findbp(dn, 0, blkid, TRUE, &db, &bp) == 0) {
1851 if (bp && !BP_IS_HOLE(bp)) {
1852 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
1853 uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH;
1854 zbookmark_t zb;
1855
1856 SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
1857 dn->dn_object, 0, blkid);
1858
1859 (void) arc_read(NULL, dn->dn_objset->os_spa,
1860 bp, NULL, NULL, prio,
1861 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1862 &aflags, &zb);
1863 }
1864 if (db)
1865 dbuf_rele(db, NULL);
1866 }
1867 }
1868
1869 /*
1870 * Returns with db_holds incremented, and db_mtx not held.
1871 * Note: dn_struct_rwlock must be held.
1872 */
1873 int
1874 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse,
1875 void *tag, dmu_buf_impl_t **dbp)
1876 {
1877 dmu_buf_impl_t *db, *parent = NULL;
1878
1879 ASSERT(blkid != DMU_BONUS_BLKID);
1880 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1881 ASSERT3U(dn->dn_nlevels, >, level);
1882
1883 *dbp = NULL;
1884 top:
1885 /* dbuf_find() returns with db_mtx held */
1886 db = dbuf_find(dn, level, blkid);
1887
1888 if (db == NULL) {
1889 blkptr_t *bp = NULL;
1890 int err;
1891
1892 ASSERT3P(parent, ==, NULL);
1893 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
1894 if (fail_sparse) {
1895 if (err == 0 && bp && BP_IS_HOLE(bp))
1896 err = SET_ERROR(ENOENT);
1897 if (err) {
1898 if (parent)
1899 dbuf_rele(parent, NULL);
1900 return (err);
1901 }
1902 }
1903 if (err && err != ENOENT)
1904 return (err);
1905 db = dbuf_create(dn, level, blkid, parent, bp);
1906 }
1907
1908 if (db->db_buf && refcount_is_zero(&db->db_holds)) {
1909 arc_buf_add_ref(db->db_buf, db);
1910 if (db->db_buf->b_data == NULL) {
1911 dbuf_clear(db);
1912 if (parent) {
1913 dbuf_rele(parent, NULL);
1914 parent = NULL;
1915 }
1916 goto top;
1917 }
1918 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
1919 }
1920
1921 ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
1922
1923 /*
1924 * If this buffer is currently syncing out, and we are are
1925 * still referencing it from db_data, we need to make a copy
1926 * of it in case we decide we want to dirty it again in this txg.
1927 */
1928 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1929 dn->dn_object != DMU_META_DNODE_OBJECT &&
1930 db->db_state == DB_CACHED && db->db_data_pending) {
1931 dbuf_dirty_record_t *dr = db->db_data_pending;
1932
1933 if (dr->dt.dl.dr_data == db->db_buf) {
1934 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1935
1936 dbuf_set_data(db,
1937 arc_buf_alloc(dn->dn_objset->os_spa,
1938 db->db.db_size, db, type));
1939 bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
1940 db->db.db_size);
1941 }
1942 }
1943
1944 (void) refcount_add(&db->db_holds, tag);
1945 dbuf_update_data(db);
1946 DBUF_VERIFY(db);
1947 mutex_exit(&db->db_mtx);
1948
1949 /* NOTE: we can't rele the parent until after we drop the db_mtx */
1950 if (parent)
1951 dbuf_rele(parent, NULL);
1952
1953 ASSERT3P(DB_DNODE(db), ==, dn);
1954 ASSERT3U(db->db_blkid, ==, blkid);
1955 ASSERT3U(db->db_level, ==, level);
1956 *dbp = db;
1957
1958 return (0);
1959 }
1960
1961 dmu_buf_impl_t *
1962 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
1963 {
1964 dmu_buf_impl_t *db;
1965 int err = dbuf_hold_impl(dn, 0, blkid, FALSE, tag, &db);
1966 return (err ? NULL : db);
1967 }
1968
1969 dmu_buf_impl_t *
1970 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
1971 {
1972 dmu_buf_impl_t *db;
1973 int err = dbuf_hold_impl(dn, level, blkid, FALSE, tag, &db);
1974 return (err ? NULL : db);
1975 }
1976
1977 void
1978 dbuf_create_bonus(dnode_t *dn)
1979 {
1980 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1981
1982 ASSERT(dn->dn_bonus == NULL);
1983 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
1984 }
1985
1986 int
1987 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
1988 {
1989 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1990 dnode_t *dn;
1991
1992 if (db->db_blkid != DMU_SPILL_BLKID)
1993 return (SET_ERROR(ENOTSUP));
1994 if (blksz == 0)
1995 blksz = SPA_MINBLOCKSIZE;
1996 if (blksz > SPA_MAXBLOCKSIZE)
1997 blksz = SPA_MAXBLOCKSIZE;
1998 else
1999 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
2000
2001 DB_DNODE_ENTER(db);
2002 dn = DB_DNODE(db);
2003 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2004 dbuf_new_size(db, blksz, tx);
2005 rw_exit(&dn->dn_struct_rwlock);
2006 DB_DNODE_EXIT(db);
2007
2008 return (0);
2009 }
2010
2011 void
2012 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
2013 {
2014 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
2015 }
2016
2017 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2018 void
2019 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
2020 {
2021 int64_t holds = refcount_add(&db->db_holds, tag);
2022 ASSERT(holds > 1);
2023 }
2024
2025 /*
2026 * If you call dbuf_rele() you had better not be referencing the dnode handle
2027 * unless you have some other direct or indirect hold on the dnode. (An indirect
2028 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2029 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2030 * dnode's parent dbuf evicting its dnode handles.
2031 */
2032 #pragma weak dmu_buf_rele = dbuf_rele
2033 void
2034 dbuf_rele(dmu_buf_impl_t *db, void *tag)
2035 {
2036 mutex_enter(&db->db_mtx);
2037 dbuf_rele_and_unlock(db, tag);
2038 }
2039
2040 /*
2041 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2042 * db_dirtycnt and db_holds to be updated atomically.
2043 */
2044 void
2045 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2046 {
2047 int64_t holds;
2048
2049 ASSERT(MUTEX_HELD(&db->db_mtx));
2050 DBUF_VERIFY(db);
2051
2052 /*
2053 * Remove the reference to the dbuf before removing its hold on the
2054 * dnode so we can guarantee in dnode_move() that a referenced bonus
2055 * buffer has a corresponding dnode hold.
2056 */
2057 holds = refcount_remove(&db->db_holds, tag);
2058 ASSERT(holds >= 0);
2059
2060 /*
2061 * We can't freeze indirects if there is a possibility that they
2062 * may be modified in the current syncing context.
2063 */
2064 if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0))
2065 arc_buf_freeze(db->db_buf);
2066
2067 if (holds == db->db_dirtycnt &&
2068 db->db_level == 0 && db->db_immediate_evict)
2069 dbuf_evict_user(db);
2070
2071 if (holds == 0) {
2072 if (db->db_blkid == DMU_BONUS_BLKID) {
2073 mutex_exit(&db->db_mtx);
2074
2075 /*
2076 * If the dnode moves here, we cannot cross this barrier
2077 * until the move completes.
2078 */
2079 DB_DNODE_ENTER(db);
2080 (void) atomic_dec_32_nv(&DB_DNODE(db)->dn_dbufs_count);
2081 DB_DNODE_EXIT(db);
2082 /*
2083 * The bonus buffer's dnode hold is no longer discounted
2084 * in dnode_move(). The dnode cannot move until after
2085 * the dnode_rele().
2086 */
2087 dnode_rele(DB_DNODE(db), db);
2088 } else if (db->db_buf == NULL) {
2089 /*
2090 * This is a special case: we never associated this
2091 * dbuf with any data allocated from the ARC.
2092 */
2093 ASSERT(db->db_state == DB_UNCACHED ||
2094 db->db_state == DB_NOFILL);
2095 dbuf_evict(db);
2096 } else if (arc_released(db->db_buf)) {
2097 arc_buf_t *buf = db->db_buf;
2098 /*
2099 * This dbuf has anonymous data associated with it.
2100 */
2101 dbuf_set_data(db, NULL);
2102 VERIFY(arc_buf_remove_ref(buf, db));
2103 dbuf_evict(db);
2104 } else {
2105 VERIFY(!arc_buf_remove_ref(db->db_buf, db));
2106
2107 /*
2108 * A dbuf will be eligible for eviction if either the
2109 * 'primarycache' property is set or a duplicate
2110 * copy of this buffer is already cached in the arc.
2111 *
2112 * In the case of the 'primarycache' a buffer
2113 * is considered for eviction if it matches the
2114 * criteria set in the property.
2115 *
2116 * To decide if our buffer is considered a
2117 * duplicate, we must call into the arc to determine
2118 * if multiple buffers are referencing the same
2119 * block on-disk. If so, then we simply evict
2120 * ourselves.
2121 */
2122 if (!DBUF_IS_CACHEABLE(db) ||
2123 arc_buf_eviction_needed(db->db_buf))
2124 dbuf_clear(db);
2125 else
2126 mutex_exit(&db->db_mtx);
2127 }
2128 } else {
2129 mutex_exit(&db->db_mtx);
2130 }
2131 }
2132
2133 #pragma weak dmu_buf_refcount = dbuf_refcount
2134 uint64_t
2135 dbuf_refcount(dmu_buf_impl_t *db)
2136 {
2137 return (refcount_count(&db->db_holds));
2138 }
2139
2140 void *
2141 dmu_buf_set_user(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr,
2142 dmu_buf_evict_func_t *evict_func)
2143 {
2144 return (dmu_buf_update_user(db_fake, NULL, user_ptr,
2145 user_data_ptr_ptr, evict_func));
2146 }
2147
2148 void *
2149 dmu_buf_set_user_ie(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr,
2150 dmu_buf_evict_func_t *evict_func)
2151 {
2152 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2153
2154 db->db_immediate_evict = TRUE;
2155 return (dmu_buf_update_user(db_fake, NULL, user_ptr,
2156 user_data_ptr_ptr, evict_func));
2157 }
2158
2159 void *
2160 dmu_buf_update_user(dmu_buf_t *db_fake, void *old_user_ptr, void *user_ptr,
2161 void *user_data_ptr_ptr, dmu_buf_evict_func_t *evict_func)
2162 {
2163 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2164 ASSERT(db->db_level == 0);
2165
2166 ASSERT((user_ptr == NULL) == (evict_func == NULL));
2167
2168 mutex_enter(&db->db_mtx);
2169
2170 if (db->db_user_ptr == old_user_ptr) {
2171 db->db_user_ptr = user_ptr;
2172 db->db_user_data_ptr_ptr = user_data_ptr_ptr;
2173 db->db_evict_func = evict_func;
2174
2175 dbuf_update_data(db);
2176 } else {
2177 old_user_ptr = db->db_user_ptr;
2178 }
2179
2180 mutex_exit(&db->db_mtx);
2181 return (old_user_ptr);
2182 }
2183
2184 void *
2185 dmu_buf_get_user(dmu_buf_t *db_fake)
2186 {
2187 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2188 ASSERT(!refcount_is_zero(&db->db_holds));
2189
2190 return (db->db_user_ptr);
2191 }
2192
2193 boolean_t
2194 dmu_buf_freeable(dmu_buf_t *dbuf)
2195 {
2196 boolean_t res = B_FALSE;
2197 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2198
2199 if (db->db_blkptr)
2200 res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset,
2201 db->db_blkptr, db->db_blkptr->blk_birth);
2202
2203 return (res);
2204 }
2205
2206 blkptr_t *
2207 dmu_buf_get_blkptr(dmu_buf_t *db)
2208 {
2209 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2210 return (dbi->db_blkptr);
2211 }
2212
2213 static void
2214 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
2215 {
2216 /* ASSERT(dmu_tx_is_syncing(tx) */
2217 ASSERT(MUTEX_HELD(&db->db_mtx));
2218
2219 if (db->db_blkptr != NULL)
2220 return;
2221
2222 if (db->db_blkid == DMU_SPILL_BLKID) {
2223 db->db_blkptr = &dn->dn_phys->dn_spill;
2224 BP_ZERO(db->db_blkptr);
2225 return;
2226 }
2227 if (db->db_level == dn->dn_phys->dn_nlevels-1) {
2228 /*
2229 * This buffer was allocated at a time when there was
2230 * no available blkptrs from the dnode, or it was
2231 * inappropriate to hook it in (i.e., nlevels mis-match).
2232 */
2233 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
2234 ASSERT(db->db_parent == NULL);
2235 db->db_parent = dn->dn_dbuf;
2236 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
2237 DBUF_VERIFY(db);
2238 } else {
2239 dmu_buf_impl_t *parent = db->db_parent;
2240 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2241
2242 ASSERT(dn->dn_phys->dn_nlevels > 1);
2243 if (parent == NULL) {
2244 mutex_exit(&db->db_mtx);
2245 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2246 (void) dbuf_hold_impl(dn, db->db_level+1,
2247 db->db_blkid >> epbs, FALSE, db, &parent);
2248 rw_exit(&dn->dn_struct_rwlock);
2249 mutex_enter(&db->db_mtx);
2250 db->db_parent = parent;
2251 }
2252 db->db_blkptr = (blkptr_t *)parent->db.db_data +
2253 (db->db_blkid & ((1ULL << epbs) - 1));
2254 DBUF_VERIFY(db);
2255 }
2256 }
2257
2258 static void
2259 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2260 {
2261 dmu_buf_impl_t *db = dr->dr_dbuf;
2262 dnode_t *dn;
2263 zio_t *zio;
2264
2265 ASSERT(dmu_tx_is_syncing(tx));
2266
2267 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2268
2269 mutex_enter(&db->db_mtx);
2270
2271 ASSERT(db->db_level > 0);
2272 DBUF_VERIFY(db);
2273
2274 /* Read the block if it hasn't been read yet. */
2275 if (db->db_buf == NULL) {
2276 mutex_exit(&db->db_mtx);
2277 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
2278 mutex_enter(&db->db_mtx);
2279 }
2280 ASSERT3U(db->db_state, ==, DB_CACHED);
2281 ASSERT(db->db_buf != NULL);
2282
2283 DB_DNODE_ENTER(db);
2284 dn = DB_DNODE(db);
2285 /* Indirect block size must match what the dnode thinks it is. */
2286 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2287 dbuf_check_blkptr(dn, db);
2288 DB_DNODE_EXIT(db);
2289
2290 /* Provide the pending dirty record to child dbufs */
2291 db->db_data_pending = dr;
2292
2293 mutex_exit(&db->db_mtx);
2294 dbuf_write(dr, db->db_buf, tx);
2295
2296 zio = dr->dr_zio;
2297 mutex_enter(&dr->dt.di.dr_mtx);
2298 dbuf_sync_list(&dr->dt.di.dr_children, tx);
2299 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2300 mutex_exit(&dr->dt.di.dr_mtx);
2301 zio_nowait(zio);
2302 }
2303
2304 static void
2305 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2306 {
2307 arc_buf_t **datap = &dr->dt.dl.dr_data;
2308 dmu_buf_impl_t *db = dr->dr_dbuf;
2309 dnode_t *dn;
2310 objset_t *os;
2311 uint64_t txg = tx->tx_txg;
2312
2313 ASSERT(dmu_tx_is_syncing(tx));
2314
2315 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2316
2317 mutex_enter(&db->db_mtx);
2318 /*
2319 * To be synced, we must be dirtied. But we
2320 * might have been freed after the dirty.
2321 */
2322 if (db->db_state == DB_UNCACHED) {
2323 /* This buffer has been freed since it was dirtied */
2324 ASSERT(db->db.db_data == NULL);
2325 } else if (db->db_state == DB_FILL) {
2326 /* This buffer was freed and is now being re-filled */
2327 ASSERT(db->db.db_data != dr->dt.dl.dr_data);
2328 } else {
2329 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
2330 }
2331 DBUF_VERIFY(db);
2332
2333 DB_DNODE_ENTER(db);
2334 dn = DB_DNODE(db);
2335
2336 if (db->db_blkid == DMU_SPILL_BLKID) {
2337 mutex_enter(&dn->dn_mtx);
2338 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
2339 mutex_exit(&dn->dn_mtx);
2340 }
2341
2342 /*
2343 * If this is a bonus buffer, simply copy the bonus data into the
2344 * dnode. It will be written out when the dnode is synced (and it
2345 * will be synced, since it must have been dirty for dbuf_sync to
2346 * be called).
2347 */
2348 if (db->db_blkid == DMU_BONUS_BLKID) {
2349 dbuf_dirty_record_t **drp;
2350
2351 ASSERT(*datap != NULL);
2352 ASSERT0(db->db_level);
2353 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN);
2354 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen);
2355 DB_DNODE_EXIT(db);
2356
2357 if (*datap != db->db.db_data) {
2358 zio_buf_free(*datap, DN_MAX_BONUSLEN);
2359 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
2360 }
2361 db->db_data_pending = NULL;
2362 drp = &db->db_last_dirty;
2363 while (*drp != dr)
2364 drp = &(*drp)->dr_next;
2365 ASSERT(dr->dr_next == NULL);
2366 ASSERT(dr->dr_dbuf == db);
2367 *drp = dr->dr_next;
2368 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2369 ASSERT(db->db_dirtycnt > 0);
2370 db->db_dirtycnt -= 1;
2371 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
2372 return;
2373 }
2374
2375 os = dn->dn_objset;
2376
2377 /*
2378 * This function may have dropped the db_mtx lock allowing a dmu_sync
2379 * operation to sneak in. As a result, we need to ensure that we
2380 * don't check the dr_override_state until we have returned from
2381 * dbuf_check_blkptr.
2382 */
2383 dbuf_check_blkptr(dn, db);
2384
2385 /*
2386 * If this buffer is in the middle of an immediate write,
2387 * wait for the synchronous IO to complete.
2388 */
2389 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
2390 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
2391 cv_wait(&db->db_changed, &db->db_mtx);
2392 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
2393 }
2394
2395 if (db->db_state != DB_NOFILL &&
2396 dn->dn_object != DMU_META_DNODE_OBJECT &&
2397 refcount_count(&db->db_holds) > 1 &&
2398 dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
2399 *datap == db->db_buf) {
2400 /*
2401 * If this buffer is currently "in use" (i.e., there
2402 * are active holds and db_data still references it),
2403 * then make a copy before we start the write so that
2404 * any modifications from the open txg will not leak
2405 * into this write.
2406 *
2407 * NOTE: this copy does not need to be made for
2408 * objects only modified in the syncing context (e.g.
2409 * DNONE_DNODE blocks).
2410 */
2411 int blksz = arc_buf_size(*datap);
2412 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2413 *datap = arc_buf_alloc(os->os_spa, blksz, db, type);
2414 bcopy(db->db.db_data, (*datap)->b_data, blksz);
2415 }
2416 db->db_data_pending = dr;
2417
2418 mutex_exit(&db->db_mtx);
2419
2420 dbuf_write(dr, *datap, tx);
2421
2422 ASSERT(!list_link_active(&dr->dr_dirty_node));
2423 if (dn->dn_object == DMU_META_DNODE_OBJECT) {
2424 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
2425 DB_DNODE_EXIT(db);
2426 } else {
2427 /*
2428 * Although zio_nowait() does not "wait for an IO", it does
2429 * initiate the IO. If this is an empty write it seems plausible
2430 * that the IO could actually be completed before the nowait
2431 * returns. We need to DB_DNODE_EXIT() first in case
2432 * zio_nowait() invalidates the dbuf.
2433 */
2434 DB_DNODE_EXIT(db);
2435 zio_nowait(dr->dr_zio);
2436 }
2437 }
2438
2439 void
2440 dbuf_sync_list(list_t *list, dmu_tx_t *tx)
2441 {
2442 dbuf_dirty_record_t *dr;
2443
2444 while (dr = list_head(list)) {
2445 if (dr->dr_zio != NULL) {
2446 /*
2447 * If we find an already initialized zio then we
2448 * are processing the meta-dnode, and we have finished.
2449 * The dbufs for all dnodes are put back on the list
2450 * during processing, so that we can zio_wait()
2451 * these IOs after initiating all child IOs.
2452 */
2453 ASSERT3U(dr->dr_dbuf->db.db_object, ==,
2454 DMU_META_DNODE_OBJECT);
2455 break;
2456 }
2457 list_remove(list, dr);
2458 if (dr->dr_dbuf->db_level > 0)
2459 dbuf_sync_indirect(dr, tx);
2460 else
2461 dbuf_sync_leaf(dr, tx);
2462 }
2463 }
2464
2465 /* ARGSUSED */
2466 static void
2467 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
2468 {
2469 dmu_buf_impl_t *db = vdb;
2470 dnode_t *dn;
2471 blkptr_t *bp = zio->io_bp;
2472 blkptr_t *bp_orig = &zio->io_bp_orig;
2473 spa_t *spa = zio->io_spa;
2474 int64_t delta;
2475 uint64_t fill = 0;
2476 int i;
2477
2478 ASSERT(db->db_blkptr == bp);
2479
2480 DB_DNODE_ENTER(db);
2481 dn = DB_DNODE(db);
2482 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
2483 dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
2484 zio->io_prev_space_delta = delta;
2485
2486 if (BP_IS_HOLE(bp)) {
2487 ASSERT(bp->blk_fill == 0);
2488 DB_DNODE_EXIT(db);
2489 return;
2490 }
2491
2492 ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
2493 BP_GET_TYPE(bp) == dn->dn_type) ||
2494 (db->db_blkid == DMU_SPILL_BLKID &&
2495 BP_GET_TYPE(bp) == dn->dn_bonustype));
2496 ASSERT(BP_GET_LEVEL(bp) == db->db_level);
2497
2498 mutex_enter(&db->db_mtx);
2499
2500 #ifdef ZFS_DEBUG
2501 if (db->db_blkid == DMU_SPILL_BLKID) {
2502 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2503 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2504 db->db_blkptr == &dn->dn_phys->dn_spill);
2505 }
2506 #endif
2507
2508 if (db->db_level == 0) {
2509 mutex_enter(&dn->dn_mtx);
2510 if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
2511 db->db_blkid != DMU_SPILL_BLKID)
2512 dn->dn_phys->dn_maxblkid = db->db_blkid;
2513 mutex_exit(&dn->dn_mtx);
2514
2515 if (dn->dn_type == DMU_OT_DNODE) {
2516 dnode_phys_t *dnp = db->db.db_data;
2517 for (i = db->db.db_size >> DNODE_SHIFT; i > 0;
2518 i--, dnp++) {
2519 if (dnp->dn_type != DMU_OT_NONE)
2520 fill++;
2521 }
2522 } else {
2523 fill = 1;
2524 }
2525 } else {
2526 blkptr_t *ibp = db->db.db_data;
2527 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2528 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
2529 if (BP_IS_HOLE(ibp))
2530 continue;
2531 fill += ibp->blk_fill;
2532 }
2533 }
2534 DB_DNODE_EXIT(db);
2535
2536 bp->blk_fill = fill;
2537
2538 mutex_exit(&db->db_mtx);
2539 }
2540
2541 /*
2542 * The SPA will call this callback several times for each zio - once
2543 * for every physical child i/o (zio->io_phys_children times). This
2544 * allows the DMU to monitor the progress of each logical i/o. For example,
2545 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
2546 * block. There may be a long delay before all copies/fragments are completed,
2547 * so this callback allows us to retire dirty space gradually, as the physical
2548 * i/os complete.
2549 */
2550 /* ARGSUSED */
2551 static void
2552 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
2553 {
2554 dmu_buf_impl_t *db = arg;
2555 objset_t *os = db->db_objset;
2556 dsl_pool_t *dp = dmu_objset_pool(os);
2557 dbuf_dirty_record_t *dr;
2558 int delta = 0;
2559
2560 dr = db->db_data_pending;
2561 ASSERT3U(dr->dr_txg, ==, zio->io_txg);
2562
2563 /*
2564 * The callback will be called io_phys_children times. Retire one
2565 * portion of our dirty space each time we are called. Any rounding
2566 * error will be cleaned up by dsl_pool_sync()'s call to
2567 * dsl_pool_undirty_space().
2568 */
2569 delta = dr->dr_accounted / zio->io_phys_children;
2570 dsl_pool_undirty_space(dp, delta, zio->io_txg);
2571 }
2572
2573 /* ARGSUSED */
2574 static void
2575 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
2576 {
2577 dmu_buf_impl_t *db = vdb;
2578 blkptr_t *bp = zio->io_bp;
2579 blkptr_t *bp_orig = &zio->io_bp_orig;
2580 uint64_t txg = zio->io_txg;
2581 dbuf_dirty_record_t **drp, *dr;
2582
2583 ASSERT0(zio->io_error);
2584 ASSERT(db->db_blkptr == bp);
2585
2586 /*
2587 * For nopwrites and rewrites we ensure that the bp matches our
2588 * original and bypass all the accounting.
2589 */
2590 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
2591 ASSERT(BP_EQUAL(bp, bp_orig));
2592 } else {
2593 objset_t *os;
2594 dsl_dataset_t *ds;
2595 dmu_tx_t *tx;
2596
2597 DB_GET_OBJSET(&os, db);
2598 ds = os->os_dsl_dataset;
2599 tx = os->os_synctx;
2600
2601 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
2602 dsl_dataset_block_born(ds, bp, tx);
2603 }
2604
2605 mutex_enter(&db->db_mtx);
2606
2607 DBUF_VERIFY(db);
2608
2609 drp = &db->db_last_dirty;
2610 while ((dr = *drp) != db->db_data_pending)
2611 drp = &dr->dr_next;
2612 ASSERT(!list_link_active(&dr->dr_dirty_node));
2613 ASSERT(dr->dr_txg == txg);
2614 ASSERT(dr->dr_dbuf == db);
2615 ASSERT(dr->dr_next == NULL);
2616 *drp = dr->dr_next;
2617
2618 #ifdef ZFS_DEBUG
2619 if (db->db_blkid == DMU_SPILL_BLKID) {
2620 dnode_t *dn;
2621
2622 DB_DNODE_ENTER(db);
2623 dn = DB_DNODE(db);
2624 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2625 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2626 db->db_blkptr == &dn->dn_phys->dn_spill);
2627 DB_DNODE_EXIT(db);
2628 }
2629 #endif
2630
2631 if (db->db_level == 0) {
2632 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2633 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
2634 if (db->db_state != DB_NOFILL) {
2635 if (dr->dt.dl.dr_data != db->db_buf)
2636 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data,
2637 db));
2638 else if (!arc_released(db->db_buf))
2639 arc_set_callback(db->db_buf, dbuf_do_evict, db);
2640 }
2641 } else {
2642 dnode_t *dn;
2643
2644 DB_DNODE_ENTER(db);
2645 dn = DB_DNODE(db);
2646 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2647 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2648 if (!BP_IS_HOLE(db->db_blkptr)) {
2649 int epbs =
2650 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2651 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
2652 db->db.db_size);
2653 ASSERT3U(dn->dn_phys->dn_maxblkid
2654 >> (db->db_level * epbs), >=, db->db_blkid);
2655 arc_set_callback(db->db_buf, dbuf_do_evict, db);
2656 }
2657 DB_DNODE_EXIT(db);
2658 mutex_destroy(&dr->dt.di.dr_mtx);
2659 list_destroy(&dr->dt.di.dr_children);
2660 }
2661 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2662
2663 cv_broadcast(&db->db_changed);
2664 ASSERT(db->db_dirtycnt > 0);
2665 db->db_dirtycnt -= 1;
2666 db->db_data_pending = NULL;
2667
2668 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
2669 }
2670
2671 static void
2672 dbuf_write_nofill_ready(zio_t *zio)
2673 {
2674 dbuf_write_ready(zio, NULL, zio->io_private);
2675 }
2676
2677 static void
2678 dbuf_write_nofill_done(zio_t *zio)
2679 {
2680 dbuf_write_done(zio, NULL, zio->io_private);
2681 }
2682
2683 static void
2684 dbuf_write_override_ready(zio_t *zio)
2685 {
2686 dbuf_dirty_record_t *dr = zio->io_private;
2687 dmu_buf_impl_t *db = dr->dr_dbuf;
2688
2689 dbuf_write_ready(zio, NULL, db);
2690 }
2691
2692 static void
2693 dbuf_write_override_done(zio_t *zio)
2694 {
2695 dbuf_dirty_record_t *dr = zio->io_private;
2696 dmu_buf_impl_t *db = dr->dr_dbuf;
2697 blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
2698
2699 mutex_enter(&db->db_mtx);
2700 if (!BP_EQUAL(zio->io_bp, obp)) {
2701 if (!BP_IS_HOLE(obp))
2702 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
2703 arc_release(dr->dt.dl.dr_data, db);
2704 }
2705 mutex_exit(&db->db_mtx);
2706
2707 dbuf_write_done(zio, NULL, db);
2708 }
2709
2710 /* Issue I/O to commit a dirty buffer to disk. */
2711 static void
2712 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
2713 {
2714 dmu_buf_impl_t *db = dr->dr_dbuf;
2715 dnode_t *dn;
2716 objset_t *os;
2717 dmu_buf_impl_t *parent = db->db_parent;
2718 uint64_t txg = tx->tx_txg;
2719 zbookmark_t zb;
2720 zio_prop_t zp;
2721 zio_t *zio;
2722 int wp_flag = 0;
2723
2724 DB_DNODE_ENTER(db);
2725 dn = DB_DNODE(db);
2726 os = dn->dn_objset;
2727
2728 if (db->db_state != DB_NOFILL) {
2729 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
2730 /*
2731 * Private object buffers are released here rather
2732 * than in dbuf_dirty() since they are only modified
2733 * in the syncing context and we don't want the
2734 * overhead of making multiple copies of the data.
2735 */
2736 if (BP_IS_HOLE(db->db_blkptr)) {
2737 arc_buf_thaw(data);
2738 } else {
2739 dbuf_release_bp(db);
2740 }
2741 }
2742 }
2743
2744 if (parent != dn->dn_dbuf) {
2745 /* Our parent is an indirect block. */
2746 /* We have a dirty parent that has been scheduled for write. */
2747 ASSERT(parent && parent->db_data_pending);
2748 /* Our parent's buffer is one level closer to the dnode. */
2749 ASSERT(db->db_level == parent->db_level-1);
2750 /*
2751 * We're about to modify our parent's db_data by modifying
2752 * our block pointer, so the parent must be released.
2753 */
2754 ASSERT(arc_released(parent->db_buf));
2755 zio = parent->db_data_pending->dr_zio;
2756 } else {
2757 /* Our parent is the dnode itself. */
2758 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
2759 db->db_blkid != DMU_SPILL_BLKID) ||
2760 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
2761 if (db->db_blkid != DMU_SPILL_BLKID)
2762 ASSERT3P(db->db_blkptr, ==,
2763 &dn->dn_phys->dn_blkptr[db->db_blkid]);
2764 zio = dn->dn_zio;
2765 }
2766
2767 ASSERT(db->db_level == 0 || data == db->db_buf);
2768 ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
2769 ASSERT(zio);
2770
2771 SET_BOOKMARK(&zb, os->os_dsl_dataset ?
2772 os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
2773 db->db.db_object, db->db_level, db->db_blkid);
2774
2775 if (db->db_blkid == DMU_SPILL_BLKID)
2776 wp_flag = WP_SPILL;
2777 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
2778
2779 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
2780 DB_DNODE_EXIT(db);
2781
2782 if (db->db_level == 0 && dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
2783 ASSERT(db->db_state != DB_NOFILL);
2784 dr->dr_zio = zio_write(zio, os->os_spa, txg,
2785 db->db_blkptr, data->b_data, arc_buf_size(data), &zp,
2786 dbuf_write_override_ready, NULL, dbuf_write_override_done,
2787 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
2788 mutex_enter(&db->db_mtx);
2789 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
2790 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
2791 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
2792 mutex_exit(&db->db_mtx);
2793 } else if (db->db_state == DB_NOFILL) {
2794 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF);
2795 dr->dr_zio = zio_write(zio, os->os_spa, txg,
2796 db->db_blkptr, NULL, db->db.db_size, &zp,
2797 dbuf_write_nofill_ready, NULL, dbuf_write_nofill_done, db,
2798 ZIO_PRIORITY_ASYNC_WRITE,
2799 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
2800 } else {
2801 ASSERT(arc_released(data));
2802 dr->dr_zio = arc_write(zio, os->os_spa, txg,
2803 db->db_blkptr, data, DBUF_IS_L2CACHEABLE(db),
2804 DBUF_IS_L2COMPRESSIBLE(db), &zp, dbuf_write_ready,
2805 dbuf_write_physdone, dbuf_write_done, db,
2806 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
2807 }
2808 }