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/zfeature.h>
44 #include <sys/blkptr.h>
45 #include <sys/range_tree.h>
46
47 /*
48 * Number of times that zfs_free_range() took the slow path while doing
49 * a zfs receive. A nonzero value indicates a potential performance problem.
50 */
51 uint64_t zfs_free_range_recv_miss;
52
53 static void dbuf_destroy(dmu_buf_impl_t *db);
54 static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
55 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
56
57 /*
58 * Global data structures and functions for the dbuf cache.
59 */
60 static kmem_cache_t *dbuf_cache;
61
62 /* ARGSUSED */
63 static int
64 dbuf_cons(void *vdb, void *unused, int kmflag)
65 {
66 dmu_buf_impl_t *db = vdb;
67 bzero(db, sizeof (dmu_buf_impl_t));
68
69 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
70 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
71 refcount_create(&db->db_holds);
72
73 return (0);
74 }
75
76 /* ARGSUSED */
77 static void
78 dbuf_dest(void *vdb, void *unused)
79 {
80 dmu_buf_impl_t *db = vdb;
81 mutex_destroy(&db->db_mtx);
82 cv_destroy(&db->db_changed);
83 refcount_destroy(&db->db_holds);
84 }
85
86 /*
87 * dbuf hash table routines
88 */
89 static dbuf_hash_table_t dbuf_hash_table;
90
91 static uint64_t dbuf_hash_count;
92
93 static uint64_t
94 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
95 {
96 uintptr_t osv = (uintptr_t)os;
97 uint64_t crc = -1ULL;
98
99 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
100 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF];
101 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF];
102 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF];
103 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF];
104 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF];
105 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF];
106
107 crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16);
108
109 return (crc);
110 }
111
112 #define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid);
113
114 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
115 ((dbuf)->db.db_object == (obj) && \
116 (dbuf)->db_objset == (os) && \
117 (dbuf)->db_level == (level) && \
118 (dbuf)->db_blkid == (blkid))
119
120 dmu_buf_impl_t *
121 dbuf_find(dnode_t *dn, uint8_t level, uint64_t blkid)
122 {
123 dbuf_hash_table_t *h = &dbuf_hash_table;
124 objset_t *os = dn->dn_objset;
125 uint64_t obj = dn->dn_object;
126 uint64_t hv = DBUF_HASH(os, obj, level, blkid);
127 uint64_t idx = hv & h->hash_table_mask;
128 dmu_buf_impl_t *db;
129
130 mutex_enter(DBUF_HASH_MUTEX(h, idx));
131 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
132 if (DBUF_EQUAL(db, os, obj, level, blkid)) {
133 mutex_enter(&db->db_mtx);
134 if (db->db_state != DB_EVICTING) {
135 mutex_exit(DBUF_HASH_MUTEX(h, idx));
136 return (db);
137 }
138 mutex_exit(&db->db_mtx);
139 }
140 }
141 mutex_exit(DBUF_HASH_MUTEX(h, idx));
142 return (NULL);
143 }
144
145 /*
146 * Insert an entry into the hash table. If there is already an element
147 * equal to elem in the hash table, then the already existing element
148 * will be returned and the new element will not be inserted.
149 * Otherwise returns NULL.
150 */
151 static dmu_buf_impl_t *
152 dbuf_hash_insert(dmu_buf_impl_t *db)
153 {
154 dbuf_hash_table_t *h = &dbuf_hash_table;
155 objset_t *os = db->db_objset;
156 uint64_t obj = db->db.db_object;
157 int level = db->db_level;
158 uint64_t blkid = db->db_blkid;
159 uint64_t hv = DBUF_HASH(os, obj, level, blkid);
160 uint64_t idx = hv & h->hash_table_mask;
161 dmu_buf_impl_t *dbf;
162
163 mutex_enter(DBUF_HASH_MUTEX(h, idx));
164 for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) {
165 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
166 mutex_enter(&dbf->db_mtx);
167 if (dbf->db_state != DB_EVICTING) {
168 mutex_exit(DBUF_HASH_MUTEX(h, idx));
169 return (dbf);
170 }
171 mutex_exit(&dbf->db_mtx);
172 }
173 }
174
175 mutex_enter(&db->db_mtx);
176 db->db_hash_next = h->hash_table[idx];
177 h->hash_table[idx] = db;
178 mutex_exit(DBUF_HASH_MUTEX(h, idx));
179 atomic_inc_64(&dbuf_hash_count);
180
181 return (NULL);
182 }
183
184 /*
185 * Remove an entry from the hash table. It must be in the EVICTING state.
186 */
187 static void
188 dbuf_hash_remove(dmu_buf_impl_t *db)
189 {
190 dbuf_hash_table_t *h = &dbuf_hash_table;
191 uint64_t hv = DBUF_HASH(db->db_objset, db->db.db_object,
192 db->db_level, db->db_blkid);
193 uint64_t idx = hv & h->hash_table_mask;
194 dmu_buf_impl_t *dbf, **dbp;
195
196 /*
197 * We musn't hold db_mtx to maintain lock ordering:
198 * DBUF_HASH_MUTEX > db_mtx.
199 */
200 ASSERT(refcount_is_zero(&db->db_holds));
201 ASSERT(db->db_state == DB_EVICTING);
202 ASSERT(!MUTEX_HELD(&db->db_mtx));
203
204 mutex_enter(DBUF_HASH_MUTEX(h, idx));
205 dbp = &h->hash_table[idx];
206 while ((dbf = *dbp) != db) {
207 dbp = &dbf->db_hash_next;
208 ASSERT(dbf != NULL);
209 }
210 *dbp = db->db_hash_next;
211 db->db_hash_next = NULL;
212 mutex_exit(DBUF_HASH_MUTEX(h, idx));
213 atomic_dec_64(&dbuf_hash_count);
214 }
215
216 static arc_evict_func_t dbuf_do_evict;
217
218 static void
219 dbuf_evict_user(dmu_buf_impl_t *db)
220 {
221 ASSERT(MUTEX_HELD(&db->db_mtx));
222
223 if (db->db_level != 0 || db->db_evict_func == NULL)
224 return;
225
226 if (db->db_user_data_ptr_ptr)
227 *db->db_user_data_ptr_ptr = db->db.db_data;
228 db->db_evict_func(&db->db, db->db_user_ptr);
229 db->db_user_ptr = NULL;
230 db->db_user_data_ptr_ptr = NULL;
231 db->db_evict_func = NULL;
232 }
233
234 boolean_t
235 dbuf_is_metadata(dmu_buf_impl_t *db)
236 {
237 if (db->db_level > 0) {
238 return (B_TRUE);
239 } else {
240 boolean_t is_metadata;
241
242 DB_DNODE_ENTER(db);
243 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type);
244 DB_DNODE_EXIT(db);
245
246 return (is_metadata);
247 }
248 }
249
250 void
251 dbuf_evict(dmu_buf_impl_t *db)
252 {
253 ASSERT(MUTEX_HELD(&db->db_mtx));
254 ASSERT(db->db_buf == NULL);
255 ASSERT(db->db_data_pending == NULL);
256
257 dbuf_clear(db);
258 dbuf_destroy(db);
259 }
260
261 void
262 dbuf_init(void)
263 {
264 uint64_t hsize = 1ULL << 16;
265 dbuf_hash_table_t *h = &dbuf_hash_table;
266 int i;
267
268 /*
269 * The hash table is big enough to fill all of physical memory
270 * with an average 4K block size. The table will take up
271 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
272 */
273 while (hsize * 4096 < physmem * PAGESIZE)
274 hsize <<= 1;
275
276 retry:
277 h->hash_table_mask = hsize - 1;
278 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
279 if (h->hash_table == NULL) {
280 /* XXX - we should really return an error instead of assert */
281 ASSERT(hsize > (1ULL << 10));
282 hsize >>= 1;
283 goto retry;
284 }
285
286 dbuf_cache = kmem_cache_create("dmu_buf_impl_t",
287 sizeof (dmu_buf_impl_t),
288 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
289
290 for (i = 0; i < DBUF_MUTEXES; i++)
291 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
292 }
293
294 void
295 dbuf_fini(void)
296 {
297 dbuf_hash_table_t *h = &dbuf_hash_table;
298 int i;
299
300 for (i = 0; i < DBUF_MUTEXES; i++)
301 mutex_destroy(&h->hash_mutexes[i]);
302 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
303 kmem_cache_destroy(dbuf_cache);
304 }
305
306 /*
307 * Other stuff.
308 */
309
310 #ifdef ZFS_DEBUG
311 static void
312 dbuf_verify(dmu_buf_impl_t *db)
313 {
314 dnode_t *dn;
315 dbuf_dirty_record_t *dr;
316
317 ASSERT(MUTEX_HELD(&db->db_mtx));
318
319 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
320 return;
321
322 ASSERT(db->db_objset != NULL);
323 DB_DNODE_ENTER(db);
324 dn = DB_DNODE(db);
325 if (dn == NULL) {
326 ASSERT(db->db_parent == NULL);
327 ASSERT(db->db_blkptr == NULL);
328 } else {
329 ASSERT3U(db->db.db_object, ==, dn->dn_object);
330 ASSERT3P(db->db_objset, ==, dn->dn_objset);
331 ASSERT3U(db->db_level, <, dn->dn_nlevels);
332 ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
333 db->db_blkid == DMU_SPILL_BLKID ||
334 !avl_is_empty(&dn->dn_dbufs));
335 }
336 if (db->db_blkid == DMU_BONUS_BLKID) {
337 ASSERT(dn != NULL);
338 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
339 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
340 } else if (db->db_blkid == DMU_SPILL_BLKID) {
341 ASSERT(dn != NULL);
342 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
343 ASSERT0(db->db.db_offset);
344 } else {
345 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
346 }
347
348 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
349 ASSERT(dr->dr_dbuf == db);
350
351 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
352 ASSERT(dr->dr_dbuf == db);
353
354 /*
355 * We can't assert that db_size matches dn_datablksz because it
356 * can be momentarily different when another thread is doing
357 * dnode_set_blksz().
358 */
359 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
360 dr = db->db_data_pending;
361 /*
362 * It should only be modified in syncing context, so
363 * make sure we only have one copy of the data.
364 */
365 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
366 }
367
368 /* verify db->db_blkptr */
369 if (db->db_blkptr) {
370 if (db->db_parent == dn->dn_dbuf) {
371 /* db is pointed to by the dnode */
372 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
373 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
374 ASSERT(db->db_parent == NULL);
375 else
376 ASSERT(db->db_parent != NULL);
377 if (db->db_blkid != DMU_SPILL_BLKID)
378 ASSERT3P(db->db_blkptr, ==,
379 &dn->dn_phys->dn_blkptr[db->db_blkid]);
380 } else {
381 /* db is pointed to by an indirect block */
382 int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
383 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
384 ASSERT3U(db->db_parent->db.db_object, ==,
385 db->db.db_object);
386 /*
387 * dnode_grow_indblksz() can make this fail if we don't
388 * have the struct_rwlock. XXX indblksz no longer
389 * grows. safe to do this now?
390 */
391 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
392 ASSERT3P(db->db_blkptr, ==,
393 ((blkptr_t *)db->db_parent->db.db_data +
394 db->db_blkid % epb));
395 }
396 }
397 }
398 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
399 (db->db_buf == NULL || db->db_buf->b_data) &&
400 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
401 db->db_state != DB_FILL && !dn->dn_free_txg) {
402 /*
403 * If the blkptr isn't set but they have nonzero data,
404 * it had better be dirty, otherwise we'll lose that
405 * data when we evict this buffer.
406 */
407 if (db->db_dirtycnt == 0) {
408 uint64_t *buf = db->db.db_data;
409 int i;
410
411 for (i = 0; i < db->db.db_size >> 3; i++) {
412 ASSERT(buf[i] == 0);
413 }
414 }
415 }
416 DB_DNODE_EXIT(db);
417 }
418 #endif
419
420 static void
421 dbuf_update_data(dmu_buf_impl_t *db)
422 {
423 ASSERT(MUTEX_HELD(&db->db_mtx));
424 if (db->db_level == 0 && db->db_user_data_ptr_ptr) {
425 ASSERT(!refcount_is_zero(&db->db_holds));
426 *db->db_user_data_ptr_ptr = db->db.db_data;
427 }
428 }
429
430 static void
431 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
432 {
433 ASSERT(MUTEX_HELD(&db->db_mtx));
434 db->db_buf = buf;
435 if (buf != NULL) {
436 ASSERT(buf->b_data != NULL);
437 db->db.db_data = buf->b_data;
438 if (!arc_released(buf))
439 arc_set_callback(buf, dbuf_do_evict, db);
440 dbuf_update_data(db);
441 } else {
442 dbuf_evict_user(db);
443 db->db.db_data = NULL;
444 if (db->db_state != DB_NOFILL)
445 db->db_state = DB_UNCACHED;
446 }
447 }
448
449 /*
450 * Loan out an arc_buf for read. Return the loaned arc_buf.
451 */
452 arc_buf_t *
453 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
454 {
455 arc_buf_t *abuf;
456
457 mutex_enter(&db->db_mtx);
458 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
459 int blksz = db->db.db_size;
460 spa_t *spa = db->db_objset->os_spa;
461
462 mutex_exit(&db->db_mtx);
463 abuf = arc_loan_buf(spa, blksz);
464 bcopy(db->db.db_data, abuf->b_data, blksz);
465 } else {
466 abuf = db->db_buf;
467 arc_loan_inuse_buf(abuf, db);
468 dbuf_set_data(db, NULL);
469 mutex_exit(&db->db_mtx);
470 }
471 return (abuf);
472 }
473
474 uint64_t
475 dbuf_whichblock(dnode_t *dn, uint64_t offset)
476 {
477 if (dn->dn_datablkshift) {
478 return (offset >> dn->dn_datablkshift);
479 } else {
480 ASSERT3U(offset, <, dn->dn_datablksz);
481 return (0);
482 }
483 }
484
485 static void
486 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
487 {
488 dmu_buf_impl_t *db = vdb;
489
490 mutex_enter(&db->db_mtx);
491 ASSERT3U(db->db_state, ==, DB_READ);
492 /*
493 * All reads are synchronous, so we must have a hold on the dbuf
494 */
495 ASSERT(refcount_count(&db->db_holds) > 0);
496 ASSERT(db->db_buf == NULL);
497 ASSERT(db->db.db_data == NULL);
498 if (db->db_level == 0 && db->db_freed_in_flight) {
499 /* we were freed in flight; disregard any error */
500 arc_release(buf, db);
501 bzero(buf->b_data, db->db.db_size);
502 arc_buf_freeze(buf);
503 db->db_freed_in_flight = FALSE;
504 dbuf_set_data(db, buf);
505 db->db_state = DB_CACHED;
506 } else if (zio == NULL || zio->io_error == 0) {
507 dbuf_set_data(db, buf);
508 db->db_state = DB_CACHED;
509 } else {
510 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
511 ASSERT3P(db->db_buf, ==, NULL);
512 VERIFY(arc_buf_remove_ref(buf, db));
513 db->db_state = DB_UNCACHED;
514 }
515 cv_broadcast(&db->db_changed);
516 dbuf_rele_and_unlock(db, NULL);
517 }
518
519 static void
520 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t *flags)
521 {
522 dnode_t *dn;
523 zbookmark_phys_t zb;
524 uint32_t aflags = ARC_NOWAIT;
525
526 DB_DNODE_ENTER(db);
527 dn = DB_DNODE(db);
528 ASSERT(!refcount_is_zero(&db->db_holds));
529 /* We need the struct_rwlock to prevent db_blkptr from changing. */
530 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
531 ASSERT(MUTEX_HELD(&db->db_mtx));
532 ASSERT(db->db_state == DB_UNCACHED);
533 ASSERT(db->db_buf == NULL);
534
535 if (db->db_blkid == DMU_BONUS_BLKID) {
536 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
537
538 ASSERT3U(bonuslen, <=, db->db.db_size);
539 db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN);
540 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
541 if (bonuslen < DN_MAX_BONUSLEN)
542 bzero(db->db.db_data, DN_MAX_BONUSLEN);
543 if (bonuslen)
544 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
545 DB_DNODE_EXIT(db);
546 dbuf_update_data(db);
547 db->db_state = DB_CACHED;
548 mutex_exit(&db->db_mtx);
549 return;
550 }
551
552 /*
553 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
554 * processes the delete record and clears the bp while we are waiting
555 * for the dn_mtx (resulting in a "no" from block_freed).
556 */
557 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
558 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
559 BP_IS_HOLE(db->db_blkptr)))) {
560 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
561
562 DB_DNODE_EXIT(db);
563 dbuf_set_data(db, arc_buf_alloc(db->db_objset->os_spa,
564 db->db.db_size, db, type));
565 bzero(db->db.db_data, db->db.db_size);
566 db->db_state = DB_CACHED;
567 *flags |= DB_RF_CACHED;
568 mutex_exit(&db->db_mtx);
569 return;
570 }
571
572 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, db->db_objset->os_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 boolean_t havepzio = (zio != NULL);
601 boolean_t 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 DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *,
675 db, zio_t *, zio);
676 cv_wait(&db->db_changed, &db->db_mtx);
677 }
678 if (db->db_state == DB_UNCACHED)
679 err = SET_ERROR(EIO);
680 }
681 mutex_exit(&db->db_mtx);
682 }
683
684 ASSERT(err || havepzio || db->db_state == DB_CACHED);
685 return (err);
686 }
687
688 static void
689 dbuf_noread(dmu_buf_impl_t *db)
690 {
691 ASSERT(!refcount_is_zero(&db->db_holds));
692 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
693 mutex_enter(&db->db_mtx);
694 while (db->db_state == DB_READ || db->db_state == DB_FILL)
695 cv_wait(&db->db_changed, &db->db_mtx);
696 if (db->db_state == DB_UNCACHED) {
697 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
698 spa_t *spa = db->db_objset->os_spa;
699
700 ASSERT(db->db_buf == NULL);
701 ASSERT(db->db.db_data == NULL);
702 dbuf_set_data(db, arc_buf_alloc(spa, db->db.db_size, db, type));
703 db->db_state = DB_FILL;
704 } else if (db->db_state == DB_NOFILL) {
705 dbuf_set_data(db, NULL);
706 } else {
707 ASSERT3U(db->db_state, ==, DB_CACHED);
708 }
709 mutex_exit(&db->db_mtx);
710 }
711
712 /*
713 * This is our just-in-time copy function. It makes a copy of
714 * buffers, that have been modified in a previous transaction
715 * group, before we modify them in the current active group.
716 *
717 * This function is used in two places: when we are dirtying a
718 * buffer for the first time in a txg, and when we are freeing
719 * a range in a dnode that includes this buffer.
720 *
721 * Note that when we are called from dbuf_free_range() we do
722 * not put a hold on the buffer, we just traverse the active
723 * dbuf list for the dnode.
724 */
725 static void
726 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
727 {
728 dbuf_dirty_record_t *dr = db->db_last_dirty;
729
730 ASSERT(MUTEX_HELD(&db->db_mtx));
731 ASSERT(db->db.db_data != NULL);
732 ASSERT(db->db_level == 0);
733 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
734
735 if (dr == NULL ||
736 (dr->dt.dl.dr_data !=
737 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
738 return;
739
740 /*
741 * If the last dirty record for this dbuf has not yet synced
742 * and its referencing the dbuf data, either:
743 * reset the reference to point to a new copy,
744 * or (if there a no active holders)
745 * just null out the current db_data pointer.
746 */
747 ASSERT(dr->dr_txg >= txg - 2);
748 if (db->db_blkid == DMU_BONUS_BLKID) {
749 /* Note that the data bufs here are zio_bufs */
750 dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
751 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
752 bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
753 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
754 int size = db->db.db_size;
755 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
756 spa_t *spa = db->db_objset->os_spa;
757
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 zio_free(db->db_objset->os_spa, txg, bp);
785
786 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
787 dr->dt.dl.dr_nopwrite = B_FALSE;
788
789 /*
790 * Release the already-written buffer, so we leave it in
791 * a consistent dirty state. Note that all callers are
792 * modifying the buffer, so they will immediately do
793 * another (redundant) arc_release(). Therefore, leave
794 * the buf thawed to save the effort of freezing &
795 * immediately re-thawing it.
796 */
797 arc_release(dr->dt.dl.dr_data, db);
798 }
799
800 /*
801 * Evict (if its unreferenced) or clear (if its referenced) any level-0
802 * data blocks in the free range, so that any future readers will find
803 * empty blocks.
804 *
805 * This is a no-op if the dataset is in the middle of an incremental
806 * receive; see comment below for details.
807 */
808 void
809 dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
810 dmu_tx_t *tx)
811 {
812 dmu_buf_impl_t *db, *db_next, db_search;
813 uint64_t txg = tx->tx_txg;
814 avl_index_t where;
815
816 if (end_blkid > dn->dn_maxblkid && (end_blkid != DMU_SPILL_BLKID))
817 end_blkid = dn->dn_maxblkid;
818 dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid);
819
820 db_search.db_level = 0;
821 db_search.db_blkid = start_blkid;
822 db_search.db_state = DB_SEARCH;
823
824 mutex_enter(&dn->dn_dbufs_mtx);
825 if (start_blkid >= dn->dn_unlisted_l0_blkid) {
826 /* There can't be any dbufs in this range; no need to search. */
827 #ifdef DEBUG
828 db = avl_find(&dn->dn_dbufs, &db_search, &where);
829 ASSERT3P(db, ==, NULL);
830 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
831 ASSERT(db == NULL || db->db_level > 0);
832 #endif
833 mutex_exit(&dn->dn_dbufs_mtx);
834 return;
835 } else if (dmu_objset_is_receiving(dn->dn_objset)) {
836 /*
837 * If we are receiving, we expect there to be no dbufs in
838 * the range to be freed, because receive modifies each
839 * block at most once, and in offset order. If this is
840 * not the case, it can lead to performance problems,
841 * so note that we unexpectedly took the slow path.
842 */
843 atomic_inc_64(&zfs_free_range_recv_miss);
844 }
845
846 db = avl_find(&dn->dn_dbufs, &db_search, &where);
847 ASSERT3P(db, ==, NULL);
848 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
849
850 for (; db != NULL; db = db_next) {
851 db_next = AVL_NEXT(&dn->dn_dbufs, db);
852 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
853
854 if (db->db_level != 0 || db->db_blkid > end_blkid) {
855 break;
856 }
857 ASSERT3U(db->db_blkid, >=, start_blkid);
858
859 /* found a level 0 buffer in the range */
860 mutex_enter(&db->db_mtx);
861 if (dbuf_undirty(db, tx)) {
862 /* mutex has been dropped and dbuf destroyed */
863 continue;
864 }
865
866 if (db->db_state == DB_UNCACHED ||
867 db->db_state == DB_NOFILL ||
868 db->db_state == DB_EVICTING) {
869 ASSERT(db->db.db_data == NULL);
870 mutex_exit(&db->db_mtx);
871 continue;
872 }
873 if (db->db_state == DB_READ || db->db_state == DB_FILL) {
874 /* will be handled in dbuf_read_done or dbuf_rele */
875 db->db_freed_in_flight = TRUE;
876 mutex_exit(&db->db_mtx);
877 continue;
878 }
879 if (refcount_count(&db->db_holds) == 0) {
880 ASSERT(db->db_buf);
881 dbuf_clear(db);
882 continue;
883 }
884 /* The dbuf is referenced */
885
886 if (db->db_last_dirty != NULL) {
887 dbuf_dirty_record_t *dr = db->db_last_dirty;
888
889 if (dr->dr_txg == txg) {
890 /*
891 * This buffer is "in-use", re-adjust the file
892 * size to reflect that this buffer may
893 * contain new data when we sync.
894 */
895 if (db->db_blkid != DMU_SPILL_BLKID &&
896 db->db_blkid > dn->dn_maxblkid)
897 dn->dn_maxblkid = db->db_blkid;
898 dbuf_unoverride(dr);
899 } else {
900 /*
901 * This dbuf is not dirty in the open context.
902 * Either uncache it (if its not referenced in
903 * the open context) or reset its contents to
904 * empty.
905 */
906 dbuf_fix_old_data(db, txg);
907 }
908 }
909 /* clear the contents if its cached */
910 if (db->db_state == DB_CACHED) {
911 ASSERT(db->db.db_data != NULL);
912 arc_release(db->db_buf, db);
913 bzero(db->db.db_data, db->db.db_size);
914 arc_buf_freeze(db->db_buf);
915 }
916
917 mutex_exit(&db->db_mtx);
918 }
919 mutex_exit(&dn->dn_dbufs_mtx);
920 }
921
922 static int
923 dbuf_block_freeable(dmu_buf_impl_t *db)
924 {
925 dsl_dataset_t *ds = db->db_objset->os_dsl_dataset;
926 uint64_t birth_txg = 0;
927
928 /*
929 * We don't need any locking to protect db_blkptr:
930 * If it's syncing, then db_last_dirty will be set
931 * so we'll ignore db_blkptr.
932 *
933 * This logic ensures that only block births for
934 * filled blocks are considered.
935 */
936 ASSERT(MUTEX_HELD(&db->db_mtx));
937 if (db->db_last_dirty && (db->db_blkptr == NULL ||
938 !BP_IS_HOLE(db->db_blkptr))) {
939 birth_txg = db->db_last_dirty->dr_txg;
940 } else if (db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) {
941 birth_txg = db->db_blkptr->blk_birth;
942 }
943
944 /*
945 * If this block don't exist or is in a snapshot, it can't be freed.
946 * Don't pass the bp to dsl_dataset_block_freeable() since we
947 * are holding the db_mtx lock and might deadlock if we are
948 * prefetching a dedup-ed block.
949 */
950 if (birth_txg != 0)
951 return (ds == NULL ||
952 dsl_dataset_block_freeable(ds, NULL, birth_txg));
953 else
954 return (B_FALSE);
955 }
956
957 void
958 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
959 {
960 arc_buf_t *buf, *obuf;
961 int osize = db->db.db_size;
962 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
963 dnode_t *dn;
964
965 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
966
967 DB_DNODE_ENTER(db);
968 dn = DB_DNODE(db);
969
970 /* XXX does *this* func really need the lock? */
971 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
972
973 /*
974 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
975 * is OK, because there can be no other references to the db
976 * when we are changing its size, so no concurrent DB_FILL can
977 * be happening.
978 */
979 /*
980 * XXX we should be doing a dbuf_read, checking the return
981 * value and returning that up to our callers
982 */
983 dmu_buf_will_dirty(&db->db, tx);
984
985 /* create the data buffer for the new block */
986 buf = arc_buf_alloc(dn->dn_objset->os_spa, size, db, type);
987
988 /* copy old block data to the new block */
989 obuf = db->db_buf;
990 bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
991 /* zero the remainder */
992 if (size > osize)
993 bzero((uint8_t *)buf->b_data + osize, size - osize);
994
995 mutex_enter(&db->db_mtx);
996 dbuf_set_data(db, buf);
997 VERIFY(arc_buf_remove_ref(obuf, db));
998 db->db.db_size = size;
999
1000 if (db->db_level == 0) {
1001 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1002 db->db_last_dirty->dt.dl.dr_data = buf;
1003 }
1004 mutex_exit(&db->db_mtx);
1005
1006 dnode_willuse_space(dn, size-osize, tx);
1007 DB_DNODE_EXIT(db);
1008 }
1009
1010 void
1011 dbuf_release_bp(dmu_buf_impl_t *db)
1012 {
1013 objset_t *os = db->db_objset;
1014
1015 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1016 ASSERT(arc_released(os->os_phys_buf) ||
1017 list_link_active(&os->os_dsl_dataset->ds_synced_link));
1018 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1019
1020 (void) arc_release(db->db_buf, db);
1021 }
1022
1023 dbuf_dirty_record_t *
1024 dbuf_zero_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1025 {
1026 ASSERT(db->db_objset != NULL);
1027
1028 return (dbuf_dirty(db, tx, B_TRUE));
1029 }
1030
1031 dbuf_dirty_record_t *
1032 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx, boolean_t zero_write)
1033 {
1034 dnode_t *dn;
1035 objset_t *os;
1036 dbuf_dirty_record_t **drp, *dr;
1037 int drop_struct_lock = FALSE;
1038 boolean_t do_free_accounting = B_FALSE;
1039 int txgoff = tx->tx_txg & TXG_MASK;
1040
1041 ASSERT(tx->tx_txg != 0);
1042 ASSERT(!refcount_is_zero(&db->db_holds));
1043 DMU_TX_DIRTY_BUF(tx, db);
1044
1045 DB_DNODE_ENTER(db);
1046 dn = DB_DNODE(db);
1047 /*
1048 * Shouldn't dirty a regular buffer in syncing context. Private
1049 * objects may be dirtied in syncing context, but only if they
1050 * were already pre-dirtied in open context.
1051 */
1052 ASSERT(!dmu_tx_is_syncing(tx) ||
1053 BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1054 DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1055 dn->dn_objset->os_dsl_dataset == NULL);
1056 /*
1057 * We make this assert for private objects as well, but after we
1058 * check if we're already dirty. They are allowed to re-dirty
1059 * in syncing context.
1060 */
1061 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1062 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1063 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1064
1065 mutex_enter(&db->db_mtx);
1066 /*
1067 * XXX make this true for indirects too? The problem is that
1068 * transactions created with dmu_tx_create_assigned() from
1069 * syncing context don't bother holding ahead.
1070 */
1071 ASSERT(db->db_level != 0 ||
1072 db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1073 db->db_state == DB_NOFILL);
1074
1075 mutex_enter(&dn->dn_mtx);
1076 /*
1077 * Don't set dirtyctx to SYNC if we're just modifying this as we
1078 * initialize the objset.
1079 */
1080 if (dn->dn_dirtyctx == DN_UNDIRTIED &&
1081 !BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1082 dn->dn_dirtyctx =
1083 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1084 ASSERT(dn->dn_dirtyctx_firstset == NULL);
1085 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1086 }
1087 mutex_exit(&dn->dn_mtx);
1088
1089 if (db->db_blkid == DMU_SPILL_BLKID)
1090 dn->dn_have_spill = B_TRUE;
1091
1092 /*
1093 * If this buffer is already dirty, we're done.
1094 */
1095 drp = &db->db_last_dirty;
1096 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1097 db->db.db_object == DMU_META_DNODE_OBJECT);
1098 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1099 drp = &dr->dr_next;
1100 if (dr && dr->dr_txg == tx->tx_txg) {
1101 DB_DNODE_EXIT(db);
1102
1103 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1104 /*
1105 * If this buffer has already been written out,
1106 * we now need to reset its state.
1107 */
1108 dbuf_unoverride(dr);
1109 if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1110 db->db_state != DB_NOFILL)
1111 arc_buf_thaw(db->db_buf);
1112 }
1113 mutex_exit(&db->db_mtx);
1114 return (dr);
1115 }
1116
1117 /*
1118 * Only valid if not already dirty.
1119 */
1120 ASSERT(dn->dn_object == 0 ||
1121 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1122 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1123
1124 ASSERT3U(dn->dn_nlevels, >, db->db_level);
1125 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1126 dn->dn_phys->dn_nlevels > db->db_level ||
1127 dn->dn_next_nlevels[txgoff] > db->db_level ||
1128 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1129 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1130
1131 /*
1132 * We should only be dirtying in syncing context if it's the
1133 * mos or we're initializing the os or it's a special object.
1134 * However, we are allowed to dirty in syncing context provided
1135 * we already dirtied it in open context. Hence we must make
1136 * this assertion only if we're not already dirty.
1137 */
1138 os = dn->dn_objset;
1139 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1140 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1141 ASSERT(db->db.db_size != 0);
1142
1143 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1144
1145 if (db->db_blkid != DMU_BONUS_BLKID) {
1146 /*
1147 * Update the accounting.
1148 * Note: we delay "free accounting" until after we drop
1149 * the db_mtx. This keeps us from grabbing other locks
1150 * (and possibly deadlocking) in bp_get_dsize() while
1151 * also holding the db_mtx.
1152 */
1153 dnode_willuse_space(dn, db->db.db_size, tx);
1154 do_free_accounting = dbuf_block_freeable(db);
1155 }
1156
1157 /*
1158 * If this buffer is dirty in an old transaction group we need
1159 * to make a copy of it so that the changes we make in this
1160 * transaction group won't leak out when we sync the older txg.
1161 */
1162 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1163 dr->dr_zero_write = zero_write;
1164 if (db->db_level == 0) {
1165 void *data_old = db->db_buf;
1166
1167 if (db->db_state != DB_NOFILL) {
1168 if (db->db_blkid == DMU_BONUS_BLKID) {
1169 dbuf_fix_old_data(db, tx->tx_txg);
1170 data_old = db->db.db_data;
1171 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1172 /*
1173 * Release the data buffer from the cache so
1174 * that we can modify it without impacting
1175 * possible other users of this cached data
1176 * block. Note that indirect blocks and
1177 * private objects are not released until the
1178 * syncing state (since they are only modified
1179 * then).
1180 */
1181 arc_release(db->db_buf, db);
1182 dbuf_fix_old_data(db, tx->tx_txg);
1183 data_old = db->db_buf;
1184 }
1185 ASSERT(data_old != NULL);
1186 }
1187 dr->dt.dl.dr_data = data_old;
1188 } else {
1189 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1190 list_create(&dr->dt.di.dr_children,
1191 sizeof (dbuf_dirty_record_t),
1192 offsetof(dbuf_dirty_record_t, dr_dirty_node));
1193 }
1194 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL)
1195 dr->dr_accounted = db->db.db_size;
1196 dr->dr_dbuf = db;
1197 dr->dr_txg = tx->tx_txg;
1198 dr->dr_next = *drp;
1199 *drp = dr;
1200
1201 /*
1202 * We could have been freed_in_flight between the dbuf_noread
1203 * and dbuf_dirty. We win, as though the dbuf_noread() had
1204 * happened after the free.
1205 */
1206 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1207 db->db_blkid != DMU_SPILL_BLKID) {
1208 mutex_enter(&dn->dn_mtx);
1209 if (dn->dn_free_ranges[txgoff] != NULL) {
1210 range_tree_clear(dn->dn_free_ranges[txgoff],
1211 db->db_blkid, 1);
1212 }
1213 mutex_exit(&dn->dn_mtx);
1214 db->db_freed_in_flight = FALSE;
1215 }
1216
1217 /*
1218 * This buffer is now part of this txg
1219 */
1220 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1221 db->db_dirtycnt += 1;
1222 ASSERT3U(db->db_dirtycnt, <=, 3);
1223
1224 mutex_exit(&db->db_mtx);
1225
1226 if (db->db_blkid == DMU_BONUS_BLKID ||
1227 db->db_blkid == DMU_SPILL_BLKID) {
1228 mutex_enter(&dn->dn_mtx);
1229 ASSERT(!list_link_active(&dr->dr_dirty_node));
1230 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1231 mutex_exit(&dn->dn_mtx);
1232 dnode_setdirty(dn, tx);
1233 DB_DNODE_EXIT(db);
1234 return (dr);
1235 } else if (do_free_accounting) {
1236 blkptr_t *bp = db->db_blkptr;
1237 int64_t willfree = (bp && !BP_IS_HOLE(bp)) ?
1238 bp_get_dsize(os->os_spa, bp) : db->db.db_size;
1239 /*
1240 * This is only a guess -- if the dbuf is dirty
1241 * in a previous txg, we don't know how much
1242 * space it will use on disk yet. We should
1243 * really have the struct_rwlock to access
1244 * db_blkptr, but since this is just a guess,
1245 * it's OK if we get an odd answer.
1246 */
1247 ddt_prefetch(os->os_spa, bp);
1248 dnode_willuse_space(dn, -willfree, tx);
1249 }
1250
1251 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1252 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1253 drop_struct_lock = TRUE;
1254 }
1255
1256 if (db->db_level == 0) {
1257 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1258 ASSERT(dn->dn_maxblkid >= db->db_blkid);
1259 }
1260
1261 if (db->db_level+1 < dn->dn_nlevels) {
1262 dmu_buf_impl_t *parent = db->db_parent;
1263 dbuf_dirty_record_t *di;
1264 int parent_held = FALSE;
1265
1266 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1267 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1268
1269 parent = dbuf_hold_level(dn, db->db_level+1,
1270 db->db_blkid >> epbs, FTAG);
1271 ASSERT(parent != NULL);
1272 parent_held = TRUE;
1273 }
1274 if (drop_struct_lock)
1275 rw_exit(&dn->dn_struct_rwlock);
1276 ASSERT3U(db->db_level+1, ==, parent->db_level);
1277 di = dbuf_dirty(parent, tx, B_FALSE);
1278 if (parent_held)
1279 dbuf_rele(parent, FTAG);
1280
1281 mutex_enter(&db->db_mtx);
1282 /*
1283 * Since we've dropped the mutex, it's possible that
1284 * dbuf_undirty() might have changed this out from under us.
1285 */
1286 if (db->db_last_dirty == dr ||
1287 dn->dn_object == DMU_META_DNODE_OBJECT) {
1288 mutex_enter(&di->dt.di.dr_mtx);
1289 ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1290 ASSERT(!list_link_active(&dr->dr_dirty_node));
1291 list_insert_tail(&di->dt.di.dr_children, dr);
1292 mutex_exit(&di->dt.di.dr_mtx);
1293 dr->dr_parent = di;
1294 }
1295 mutex_exit(&db->db_mtx);
1296 } else {
1297 ASSERT(db->db_level+1 == dn->dn_nlevels);
1298 ASSERT(db->db_blkid < dn->dn_nblkptr);
1299 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1300 mutex_enter(&dn->dn_mtx);
1301 ASSERT(!list_link_active(&dr->dr_dirty_node));
1302 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1303 mutex_exit(&dn->dn_mtx);
1304 if (drop_struct_lock)
1305 rw_exit(&dn->dn_struct_rwlock);
1306 }
1307
1308 dnode_setdirty(dn, tx);
1309 DB_DNODE_EXIT(db);
1310 return (dr);
1311 }
1312
1313 /*
1314 * Undirty a buffer in the transaction group referenced by the given
1315 * transaction. Return whether this evicted the dbuf.
1316 */
1317 static boolean_t
1318 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1319 {
1320 dnode_t *dn;
1321 uint64_t txg = tx->tx_txg;
1322 dbuf_dirty_record_t *dr, **drp;
1323
1324 ASSERT(txg != 0);
1325 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1326 ASSERT0(db->db_level);
1327 ASSERT(MUTEX_HELD(&db->db_mtx));
1328
1329 /*
1330 * If this buffer is not dirty, we're done.
1331 */
1332 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1333 if (dr->dr_txg <= txg)
1334 break;
1335 if (dr == NULL || dr->dr_txg < txg)
1336 return (B_FALSE);
1337 ASSERT(dr->dr_txg == txg);
1338 ASSERT(dr->dr_dbuf == db);
1339
1340 DB_DNODE_ENTER(db);
1341 dn = DB_DNODE(db);
1342
1343 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1344
1345 ASSERT(db->db.db_size != 0);
1346
1347 /*
1348 * Any space we accounted for in dp_dirty_* will be cleaned up by
1349 * dsl_pool_sync(). This is relatively rare so the discrepancy
1350 * is not a big deal.
1351 */
1352
1353 *drp = dr->dr_next;
1354
1355 /*
1356 * Note that there are three places in dbuf_dirty()
1357 * where this dirty record may be put on a list.
1358 * Make sure to do a list_remove corresponding to
1359 * every one of those list_insert calls.
1360 */
1361 if (dr->dr_parent) {
1362 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1363 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1364 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1365 } else if (db->db_blkid == DMU_SPILL_BLKID ||
1366 db->db_level+1 == dn->dn_nlevels) {
1367 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1368 mutex_enter(&dn->dn_mtx);
1369 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1370 mutex_exit(&dn->dn_mtx);
1371 }
1372 DB_DNODE_EXIT(db);
1373
1374 if (db->db_state != DB_NOFILL) {
1375 dbuf_unoverride(dr);
1376
1377 ASSERT(db->db_buf != NULL);
1378 ASSERT(dr->dt.dl.dr_data != NULL);
1379 if (dr->dt.dl.dr_data != db->db_buf)
1380 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db));
1381 }
1382
1383 if (db->db_level != 0) {
1384 mutex_destroy(&dr->dt.di.dr_mtx);
1385 list_destroy(&dr->dt.di.dr_children);
1386 }
1387
1388 kmem_free(dr, sizeof (dbuf_dirty_record_t));
1389
1390 ASSERT(db->db_dirtycnt > 0);
1391 db->db_dirtycnt -= 1;
1392
1393 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1394 arc_buf_t *buf = db->db_buf;
1395
1396 ASSERT(db->db_state == DB_NOFILL || arc_released(buf));
1397 dbuf_set_data(db, NULL);
1398 VERIFY(arc_buf_remove_ref(buf, db));
1399 dbuf_evict(db);
1400 return (B_TRUE);
1401 }
1402
1403 return (B_FALSE);
1404 }
1405
1406 void
1407 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
1408 {
1409 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1410 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1411
1412 ASSERT(tx->tx_txg != 0);
1413 ASSERT(!refcount_is_zero(&db->db_holds));
1414
1415 DB_DNODE_ENTER(db);
1416 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
1417 rf |= DB_RF_HAVESTRUCT;
1418 DB_DNODE_EXIT(db);
1419 (void) dbuf_read(db, NULL, rf);
1420 (void) dbuf_dirty(db, tx, B_FALSE);
1421 }
1422
1423 void
1424 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1425 {
1426 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1427
1428 db->db_state = DB_NOFILL;
1429
1430 dmu_buf_will_fill(db_fake, tx);
1431 }
1432
1433 void
1434 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1435 {
1436 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1437
1438 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1439 ASSERT(tx->tx_txg != 0);
1440 ASSERT(db->db_level == 0);
1441 ASSERT(!refcount_is_zero(&db->db_holds));
1442
1443 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1444 dmu_tx_private_ok(tx));
1445
1446 dbuf_noread(db);
1447 (void) dbuf_dirty(db, tx, B_FALSE);
1448 }
1449
1450
1451 void
1452 dmu_buf_will_zero_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1453 {
1454 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1455
1456 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1457 ASSERT(tx->tx_txg != 0);
1458 ASSERT(db->db_level == 0);
1459 ASSERT(!refcount_is_zero(&db->db_holds));
1460
1461 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1462 dmu_tx_private_ok(tx));
1463
1464 dbuf_noread(db);
1465 (void) dbuf_zero_dirty(db, tx);
1466 }
1467
1468 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1469 /* ARGSUSED */
1470 void
1471 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1472 {
1473 mutex_enter(&db->db_mtx);
1474 DBUF_VERIFY(db);
1475
1476 if (db->db_state == DB_FILL) {
1477 if (db->db_level == 0 && db->db_freed_in_flight) {
1478 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1479 /* we were freed while filling */
1480 /* XXX dbuf_undirty? */
1481 bzero(db->db.db_data, db->db.db_size);
1482 db->db_freed_in_flight = FALSE;
1483 }
1484 db->db_state = DB_CACHED;
1485 cv_broadcast(&db->db_changed);
1486 }
1487 mutex_exit(&db->db_mtx);
1488 }
1489
1490 void
1491 dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data,
1492 bp_embedded_type_t etype, enum zio_compress comp,
1493 int uncompressed_size, int compressed_size, int byteorder,
1494 dmu_tx_t *tx)
1495 {
1496 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
1497 struct dirty_leaf *dl;
1498 dmu_object_type_t type;
1499
1500 DB_DNODE_ENTER(db);
1501 type = DB_DNODE(db)->dn_type;
1502 DB_DNODE_EXIT(db);
1503
1504 ASSERT0(db->db_level);
1505 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1506
1507 dmu_buf_will_not_fill(dbuf, tx);
1508
1509 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1510 dl = &db->db_last_dirty->dt.dl;
1511 encode_embedded_bp_compressed(&dl->dr_overridden_by,
1512 data, comp, uncompressed_size, compressed_size);
1513 BPE_SET_ETYPE(&dl->dr_overridden_by, etype);
1514 BP_SET_TYPE(&dl->dr_overridden_by, type);
1515 BP_SET_LEVEL(&dl->dr_overridden_by, 0);
1516 BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder);
1517
1518 dl->dr_override_state = DR_OVERRIDDEN;
1519 dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg;
1520 }
1521
1522 /*
1523 * Directly assign a provided arc buf to a given dbuf if it's not referenced
1524 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1525 */
1526 void
1527 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
1528 {
1529 ASSERT(!refcount_is_zero(&db->db_holds));
1530 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1531 ASSERT(db->db_level == 0);
1532 ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA);
1533 ASSERT(buf != NULL);
1534 ASSERT(arc_buf_size(buf) == db->db.db_size);
1535 ASSERT(tx->tx_txg != 0);
1536
1537 arc_return_buf(buf, db);
1538 ASSERT(arc_released(buf));
1539
1540 mutex_enter(&db->db_mtx);
1541
1542 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1543 cv_wait(&db->db_changed, &db->db_mtx);
1544
1545 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
1546
1547 if (db->db_state == DB_CACHED &&
1548 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
1549 mutex_exit(&db->db_mtx);
1550 (void) dbuf_dirty(db, tx, B_FALSE);
1551 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
1552 VERIFY(arc_buf_remove_ref(buf, db));
1553 xuio_stat_wbuf_copied();
1554 return;
1555 }
1556
1557 xuio_stat_wbuf_nocopy();
1558 if (db->db_state == DB_CACHED) {
1559 dbuf_dirty_record_t *dr = db->db_last_dirty;
1560
1561 ASSERT(db->db_buf != NULL);
1562 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
1563 ASSERT(dr->dt.dl.dr_data == db->db_buf);
1564 if (!arc_released(db->db_buf)) {
1565 ASSERT(dr->dt.dl.dr_override_state ==
1566 DR_OVERRIDDEN);
1567 arc_release(db->db_buf, db);
1568 }
1569 dr->dt.dl.dr_data = buf;
1570 VERIFY(arc_buf_remove_ref(db->db_buf, db));
1571 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
1572 arc_release(db->db_buf, db);
1573 VERIFY(arc_buf_remove_ref(db->db_buf, db));
1574 }
1575 db->db_buf = NULL;
1576 }
1577 ASSERT(db->db_buf == NULL);
1578 dbuf_set_data(db, buf);
1579 db->db_state = DB_FILL;
1580 mutex_exit(&db->db_mtx);
1581 (void) dbuf_dirty(db, tx, B_FALSE);
1582 dmu_buf_fill_done(&db->db, tx);
1583 }
1584
1585 /*
1586 * "Clear" the contents of this dbuf. This will mark the dbuf
1587 * EVICTING and clear *most* of its references. Unfortunately,
1588 * when we are not holding the dn_dbufs_mtx, we can't clear the
1589 * entry in the dn_dbufs list. We have to wait until dbuf_destroy()
1590 * in this case. For callers from the DMU we will usually see:
1591 * dbuf_clear()->arc_clear_callback()->dbuf_do_evict()->dbuf_destroy()
1592 * For the arc callback, we will usually see:
1593 * dbuf_do_evict()->dbuf_clear();dbuf_destroy()
1594 * Sometimes, though, we will get a mix of these two:
1595 * DMU: dbuf_clear()->arc_clear_callback()
1596 * ARC: dbuf_do_evict()->dbuf_destroy()
1597 *
1598 * This routine will dissociate the dbuf from the arc, by calling
1599 * arc_clear_callback(), but will not evict the data from the ARC.
1600 */
1601 void
1602 dbuf_clear(dmu_buf_impl_t *db)
1603 {
1604 dnode_t *dn;
1605 dmu_buf_impl_t *parent = db->db_parent;
1606 dmu_buf_impl_t *dndb;
1607 boolean_t dbuf_gone = B_FALSE;
1608
1609 ASSERT(MUTEX_HELD(&db->db_mtx));
1610 ASSERT(refcount_is_zero(&db->db_holds));
1611
1612 dbuf_evict_user(db);
1613
1614 if (db->db_state == DB_CACHED) {
1615 ASSERT(db->db.db_data != NULL);
1616 if (db->db_blkid == DMU_BONUS_BLKID) {
1617 zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
1618 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
1619 }
1620 db->db.db_data = NULL;
1621 db->db_state = DB_UNCACHED;
1622 }
1623
1624 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
1625 ASSERT(db->db_data_pending == NULL);
1626
1627 db->db_state = DB_EVICTING;
1628 db->db_blkptr = NULL;
1629
1630 DB_DNODE_ENTER(db);
1631 dn = DB_DNODE(db);
1632 dndb = dn->dn_dbuf;
1633 if (db->db_blkid != DMU_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) {
1634 avl_remove(&dn->dn_dbufs, db);
1635 atomic_dec_32(&dn->dn_dbufs_count);
1636 membar_producer();
1637 DB_DNODE_EXIT(db);
1638 /*
1639 * Decrementing the dbuf count means that the hold corresponding
1640 * to the removed dbuf is no longer discounted in dnode_move(),
1641 * so the dnode cannot be moved until after we release the hold.
1642 * The membar_producer() ensures visibility of the decremented
1643 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
1644 * release any lock.
1645 */
1646 dnode_rele(dn, db);
1647 db->db_dnode_handle = NULL;
1648 } else {
1649 DB_DNODE_EXIT(db);
1650 }
1651
1652 if (db->db_buf)
1653 dbuf_gone = arc_clear_callback(db->db_buf);
1654
1655 if (!dbuf_gone)
1656 mutex_exit(&db->db_mtx);
1657
1658 /*
1659 * If this dbuf is referenced from an indirect dbuf,
1660 * decrement the ref count on the indirect dbuf.
1661 */
1662 if (parent && parent != dndb)
1663 dbuf_rele(parent, db);
1664 }
1665
1666 static int
1667 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
1668 dmu_buf_impl_t **parentp, blkptr_t **bpp)
1669 {
1670 int nlevels, epbs;
1671
1672 *parentp = NULL;
1673 *bpp = NULL;
1674
1675 ASSERT(blkid != DMU_BONUS_BLKID);
1676
1677 if (blkid == DMU_SPILL_BLKID) {
1678 mutex_enter(&dn->dn_mtx);
1679 if (dn->dn_have_spill &&
1680 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
1681 *bpp = &dn->dn_phys->dn_spill;
1682 else
1683 *bpp = NULL;
1684 dbuf_add_ref(dn->dn_dbuf, NULL);
1685 *parentp = dn->dn_dbuf;
1686 mutex_exit(&dn->dn_mtx);
1687 return (0);
1688 }
1689
1690 if (dn->dn_phys->dn_nlevels == 0)
1691 nlevels = 1;
1692 else
1693 nlevels = dn->dn_phys->dn_nlevels;
1694
1695 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1696
1697 ASSERT3U(level * epbs, <, 64);
1698 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1699 if (level >= nlevels ||
1700 (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
1701 /* the buffer has no parent yet */
1702 return (SET_ERROR(ENOENT));
1703 } else if (level < nlevels-1) {
1704 /* this block is referenced from an indirect block */
1705 int err = dbuf_hold_impl(dn, level+1,
1706 blkid >> epbs, fail_sparse, NULL, parentp);
1707 if (err)
1708 return (err);
1709 err = dbuf_read(*parentp, NULL,
1710 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
1711 if (err) {
1712 dbuf_rele(*parentp, NULL);
1713 *parentp = NULL;
1714 return (err);
1715 }
1716 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
1717 (blkid & ((1ULL << epbs) - 1));
1718 return (0);
1719 } else {
1720 /* the block is referenced from the dnode */
1721 ASSERT3U(level, ==, nlevels-1);
1722 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
1723 blkid < dn->dn_phys->dn_nblkptr);
1724 if (dn->dn_dbuf) {
1725 dbuf_add_ref(dn->dn_dbuf, NULL);
1726 *parentp = dn->dn_dbuf;
1727 }
1728 *bpp = &dn->dn_phys->dn_blkptr[blkid];
1729 return (0);
1730 }
1731 }
1732
1733 static dmu_buf_impl_t *
1734 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
1735 dmu_buf_impl_t *parent, blkptr_t *blkptr)
1736 {
1737 objset_t *os = dn->dn_objset;
1738 dmu_buf_impl_t *db, *odb;
1739
1740 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1741 ASSERT(dn->dn_type != DMU_OT_NONE);
1742
1743 db = kmem_cache_alloc(dbuf_cache, KM_SLEEP);
1744
1745 db->db_objset = os;
1746 db->db.db_object = dn->dn_object;
1747 db->db_level = level;
1748 db->db_blkid = blkid;
1749 db->db_last_dirty = NULL;
1750 db->db_dirtycnt = 0;
1751 db->db_dnode_handle = dn->dn_handle;
1752 db->db_parent = parent;
1753 db->db_blkptr = blkptr;
1754
1755 db->db_user_ptr = NULL;
1756 db->db_user_data_ptr_ptr = NULL;
1757 db->db_evict_func = NULL;
1758 db->db_immediate_evict = 0;
1759 db->db_freed_in_flight = 0;
1760
1761 if (blkid == DMU_BONUS_BLKID) {
1762 ASSERT3P(parent, ==, dn->dn_dbuf);
1763 db->db.db_size = DN_MAX_BONUSLEN -
1764 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
1765 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
1766 db->db.db_offset = DMU_BONUS_BLKID;
1767 db->db_state = DB_UNCACHED;
1768 /* the bonus dbuf is not placed in the hash table */
1769 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1770 return (db);
1771 } else if (blkid == DMU_SPILL_BLKID) {
1772 db->db.db_size = (blkptr != NULL) ?
1773 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
1774 db->db.db_offset = 0;
1775 } else {
1776 int blocksize =
1777 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
1778 db->db.db_size = blocksize;
1779 db->db.db_offset = db->db_blkid * blocksize;
1780 }
1781
1782 /*
1783 * Hold the dn_dbufs_mtx while we get the new dbuf
1784 * in the hash table *and* added to the dbufs list.
1785 * This prevents a possible deadlock with someone
1786 * trying to look up this dbuf before its added to the
1787 * dn_dbufs list.
1788 */
1789 mutex_enter(&dn->dn_dbufs_mtx);
1790 db->db_state = DB_EVICTING;
1791 if ((odb = dbuf_hash_insert(db)) != NULL) {
1792 /* someone else inserted it first */
1793 kmem_cache_free(dbuf_cache, db);
1794 mutex_exit(&dn->dn_dbufs_mtx);
1795 return (odb);
1796 }
1797 avl_add(&dn->dn_dbufs, db);
1798 if (db->db_level == 0 && db->db_blkid >=
1799 dn->dn_unlisted_l0_blkid)
1800 dn->dn_unlisted_l0_blkid = db->db_blkid + 1;
1801 db->db_state = DB_UNCACHED;
1802 mutex_exit(&dn->dn_dbufs_mtx);
1803 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1804
1805 if (parent && parent != dn->dn_dbuf)
1806 dbuf_add_ref(parent, db);
1807
1808 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1809 refcount_count(&dn->dn_holds) > 0);
1810 (void) refcount_add(&dn->dn_holds, db);
1811 atomic_inc_32(&dn->dn_dbufs_count);
1812
1813 dprintf_dbuf(db, "db=%p\n", db);
1814
1815 return (db);
1816 }
1817
1818 static int
1819 dbuf_do_evict(void *private)
1820 {
1821 dmu_buf_impl_t *db = private;
1822
1823 if (!MUTEX_HELD(&db->db_mtx))
1824 mutex_enter(&db->db_mtx);
1825
1826 ASSERT(refcount_is_zero(&db->db_holds));
1827
1828 if (db->db_state != DB_EVICTING) {
1829 ASSERT(db->db_state == DB_CACHED);
1830 DBUF_VERIFY(db);
1831 db->db_buf = NULL;
1832 dbuf_evict(db);
1833 } else {
1834 mutex_exit(&db->db_mtx);
1835 dbuf_destroy(db);
1836 }
1837 return (0);
1838 }
1839
1840 static void
1841 dbuf_destroy(dmu_buf_impl_t *db)
1842 {
1843 ASSERT(refcount_is_zero(&db->db_holds));
1844
1845 if (db->db_blkid != DMU_BONUS_BLKID) {
1846 /*
1847 * If this dbuf is still on the dn_dbufs list,
1848 * remove it from that list.
1849 */
1850 if (db->db_dnode_handle != NULL) {
1851 dnode_t *dn;
1852
1853 DB_DNODE_ENTER(db);
1854 dn = DB_DNODE(db);
1855 mutex_enter(&dn->dn_dbufs_mtx);
1856 avl_remove(&dn->dn_dbufs, db);
1857 atomic_dec_32(&dn->dn_dbufs_count);
1858 mutex_exit(&dn->dn_dbufs_mtx);
1859 DB_DNODE_EXIT(db);
1860 /*
1861 * Decrementing the dbuf count means that the hold
1862 * corresponding to the removed dbuf is no longer
1863 * discounted in dnode_move(), so the dnode cannot be
1864 * moved until after we release the hold.
1865 */
1866 dnode_rele(dn, db);
1867 db->db_dnode_handle = NULL;
1868 }
1869 dbuf_hash_remove(db);
1870 }
1871 db->db_parent = NULL;
1872 db->db_buf = NULL;
1873
1874 ASSERT(db->db.db_data == NULL);
1875 ASSERT(db->db_hash_next == NULL);
1876 ASSERT(db->db_blkptr == NULL);
1877 ASSERT(db->db_data_pending == NULL);
1878
1879 kmem_cache_free(dbuf_cache, db);
1880 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1881 }
1882
1883 void
1884 dbuf_prefetch(dnode_t *dn, uint64_t blkid, zio_priority_t prio)
1885 {
1886 dmu_buf_impl_t *db = NULL;
1887 blkptr_t *bp = NULL;
1888
1889 ASSERT(blkid != DMU_BONUS_BLKID);
1890 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1891
1892 if (dnode_block_freed(dn, blkid))
1893 return;
1894
1895 /* dbuf_find() returns with db_mtx held */
1896 if (db = dbuf_find(dn, 0, blkid)) {
1897 /*
1898 * This dbuf is already in the cache. We assume that
1899 * it is already CACHED, or else about to be either
1900 * read or filled.
1901 */
1902 mutex_exit(&db->db_mtx);
1903 return;
1904 }
1905
1906 if (dbuf_findbp(dn, 0, blkid, TRUE, &db, &bp) == 0) {
1907 if (bp && !BP_IS_HOLE(bp) && !BP_IS_EMBEDDED(bp)) {
1908 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
1909 uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH;
1910 zbookmark_phys_t zb;
1911
1912 SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
1913 dn->dn_object, 0, blkid);
1914
1915 (void) arc_read(NULL, dn->dn_objset->os_spa,
1916 bp, NULL, NULL, prio,
1917 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1918 &aflags, &zb);
1919 }
1920 if (db)
1921 dbuf_rele(db, NULL);
1922 }
1923 }
1924
1925 /*
1926 * Returns with db_holds incremented, and db_mtx not held.
1927 * Note: dn_struct_rwlock must be held.
1928 */
1929 int
1930 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse,
1931 void *tag, dmu_buf_impl_t **dbp)
1932 {
1933 dmu_buf_impl_t *db, *parent = NULL;
1934
1935 ASSERT(blkid != DMU_BONUS_BLKID);
1936 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1937 ASSERT3U(dn->dn_nlevels, >, level);
1938
1939 *dbp = NULL;
1940 top:
1941 /* dbuf_find() returns with db_mtx held */
1942 db = dbuf_find(dn, level, blkid);
1943
1944 if (db == NULL) {
1945 blkptr_t *bp = NULL;
1946 int err;
1947
1948 ASSERT3P(parent, ==, NULL);
1949 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
1950 if (fail_sparse) {
1951 if (err == 0 && bp && BP_IS_HOLE(bp))
1952 err = SET_ERROR(ENOENT);
1953 if (err) {
1954 if (parent)
1955 dbuf_rele(parent, NULL);
1956 return (err);
1957 }
1958 }
1959 if (err && err != ENOENT)
1960 return (err);
1961 db = dbuf_create(dn, level, blkid, parent, bp);
1962 }
1963
1964 if (db->db_buf && refcount_is_zero(&db->db_holds)) {
1965 arc_buf_add_ref(db->db_buf, db);
1966 if (db->db_buf->b_data == NULL) {
1967 dbuf_clear(db);
1968 if (parent) {
1969 dbuf_rele(parent, NULL);
1970 parent = NULL;
1971 }
1972 goto top;
1973 }
1974 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
1975 }
1976
1977 ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
1978
1979 /*
1980 * If this buffer is currently syncing out, and we are are
1981 * still referencing it from db_data, we need to make a copy
1982 * of it in case we decide we want to dirty it again in this txg.
1983 */
1984 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1985 dn->dn_object != DMU_META_DNODE_OBJECT &&
1986 db->db_state == DB_CACHED && db->db_data_pending) {
1987 dbuf_dirty_record_t *dr = db->db_data_pending;
1988
1989 if (dr->dt.dl.dr_data == db->db_buf) {
1990 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1991
1992 dbuf_set_data(db,
1993 arc_buf_alloc(dn->dn_objset->os_spa,
1994 db->db.db_size, db, type));
1995 bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
1996 db->db.db_size);
1997 }
1998 }
1999
2000 (void) refcount_add(&db->db_holds, tag);
2001 dbuf_update_data(db);
2002 DBUF_VERIFY(db);
2003 mutex_exit(&db->db_mtx);
2004
2005 /* NOTE: we can't rele the parent until after we drop the db_mtx */
2006 if (parent)
2007 dbuf_rele(parent, NULL);
2008
2009 ASSERT3P(DB_DNODE(db), ==, dn);
2010 ASSERT3U(db->db_blkid, ==, blkid);
2011 ASSERT3U(db->db_level, ==, level);
2012 *dbp = db;
2013
2014 return (0);
2015 }
2016
2017 dmu_buf_impl_t *
2018 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
2019 {
2020 dmu_buf_impl_t *db;
2021 int err = dbuf_hold_impl(dn, 0, blkid, FALSE, tag, &db);
2022 return (err ? NULL : db);
2023 }
2024
2025 dmu_buf_impl_t *
2026 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
2027 {
2028 dmu_buf_impl_t *db;
2029 int err = dbuf_hold_impl(dn, level, blkid, FALSE, tag, &db);
2030 return (err ? NULL : db);
2031 }
2032
2033 void
2034 dbuf_create_bonus(dnode_t *dn)
2035 {
2036 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
2037
2038 ASSERT(dn->dn_bonus == NULL);
2039 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
2040 }
2041
2042 int
2043 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
2044 {
2045 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2046 dnode_t *dn;
2047
2048 if (db->db_blkid != DMU_SPILL_BLKID)
2049 return (SET_ERROR(ENOTSUP));
2050 if (blksz == 0)
2051 blksz = SPA_MINBLOCKSIZE;
2052 if (blksz > SPA_MAXBLOCKSIZE)
2053 blksz = SPA_MAXBLOCKSIZE;
2054 else
2055 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
2056
2057 DB_DNODE_ENTER(db);
2058 dn = DB_DNODE(db);
2059 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2060 dbuf_new_size(db, blksz, tx);
2061 rw_exit(&dn->dn_struct_rwlock);
2062 DB_DNODE_EXIT(db);
2063
2064 return (0);
2065 }
2066
2067 void
2068 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
2069 {
2070 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
2071 }
2072
2073 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2074 void
2075 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
2076 {
2077 int64_t holds = refcount_add(&db->db_holds, tag);
2078 ASSERT(holds > 1);
2079 }
2080
2081 /*
2082 * If you call dbuf_rele() you had better not be referencing the dnode handle
2083 * unless you have some other direct or indirect hold on the dnode. (An indirect
2084 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2085 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2086 * dnode's parent dbuf evicting its dnode handles.
2087 */
2088 void
2089 dbuf_rele(dmu_buf_impl_t *db, void *tag)
2090 {
2091 mutex_enter(&db->db_mtx);
2092 dbuf_rele_and_unlock(db, tag);
2093 }
2094
2095 void
2096 dmu_buf_rele(dmu_buf_t *db, void *tag)
2097 {
2098 dbuf_rele((dmu_buf_impl_t *)db, tag);
2099 }
2100
2101 /*
2102 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2103 * db_dirtycnt and db_holds to be updated atomically.
2104 */
2105 void
2106 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2107 {
2108 int64_t holds;
2109
2110 ASSERT(MUTEX_HELD(&db->db_mtx));
2111 DBUF_VERIFY(db);
2112
2113 /*
2114 * Remove the reference to the dbuf before removing its hold on the
2115 * dnode so we can guarantee in dnode_move() that a referenced bonus
2116 * buffer has a corresponding dnode hold.
2117 */
2118 holds = refcount_remove(&db->db_holds, tag);
2119 ASSERT(holds >= 0);
2120
2121 /*
2122 * We can't freeze indirects if there is a possibility that they
2123 * may be modified in the current syncing context.
2124 */
2125 if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0))
2126 arc_buf_freeze(db->db_buf);
2127
2128 if (holds == db->db_dirtycnt &&
2129 db->db_level == 0 && db->db_immediate_evict)
2130 dbuf_evict_user(db);
2131
2132 if (holds == 0) {
2133 if (db->db_blkid == DMU_BONUS_BLKID) {
2134 mutex_exit(&db->db_mtx);
2135
2136 /*
2137 * If the dnode moves here, we cannot cross this barrier
2138 * until the move completes.
2139 */
2140 DB_DNODE_ENTER(db);
2141 atomic_dec_32(&DB_DNODE(db)->dn_dbufs_count);
2142 DB_DNODE_EXIT(db);
2143 /*
2144 * The bonus buffer's dnode hold is no longer discounted
2145 * in dnode_move(). The dnode cannot move until after
2146 * the dnode_rele().
2147 */
2148 dnode_rele(DB_DNODE(db), db);
2149 } else if (db->db_buf == NULL) {
2150 /*
2151 * This is a special case: we never associated this
2152 * dbuf with any data allocated from the ARC.
2153 */
2154 ASSERT(db->db_state == DB_UNCACHED ||
2155 db->db_state == DB_NOFILL);
2156 dbuf_evict(db);
2157 } else if (arc_released(db->db_buf)) {
2158 arc_buf_t *buf = db->db_buf;
2159 /*
2160 * This dbuf has anonymous data associated with it.
2161 */
2162 dbuf_set_data(db, NULL);
2163 VERIFY(arc_buf_remove_ref(buf, db));
2164 dbuf_evict(db);
2165 } else {
2166 VERIFY(!arc_buf_remove_ref(db->db_buf, db));
2167
2168 /*
2169 * A dbuf will be eligible for eviction if either the
2170 * 'primarycache' property is set or a duplicate
2171 * copy of this buffer is already cached in the arc.
2172 *
2173 * In the case of the 'primarycache' a buffer
2174 * is considered for eviction if it matches the
2175 * criteria set in the property.
2176 *
2177 * To decide if our buffer is considered a
2178 * duplicate, we must call into the arc to determine
2179 * if multiple buffers are referencing the same
2180 * block on-disk. If so, then we simply evict
2181 * ourselves.
2182 */
2183 if (!DBUF_IS_CACHEABLE(db)) {
2184 if (db->db_blkptr != NULL &&
2185 !BP_IS_HOLE(db->db_blkptr) &&
2186 !BP_IS_EMBEDDED(db->db_blkptr)) {
2187 spa_t *spa =
2188 dmu_objset_spa(db->db_objset);
2189 blkptr_t bp = *db->db_blkptr;
2190 dbuf_clear(db);
2191 arc_freed(spa, &bp);
2192 } else {
2193 dbuf_clear(db);
2194 }
2195 } else if (arc_buf_eviction_needed(db->db_buf)) {
2196 dbuf_clear(db);
2197 } else {
2198 mutex_exit(&db->db_mtx);
2199 }
2200 }
2201 } else {
2202 mutex_exit(&db->db_mtx);
2203 }
2204 }
2205
2206 #pragma weak dmu_buf_refcount = dbuf_refcount
2207 uint64_t
2208 dbuf_refcount(dmu_buf_impl_t *db)
2209 {
2210 return (refcount_count(&db->db_holds));
2211 }
2212
2213 void *
2214 dmu_buf_set_user(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr,
2215 dmu_buf_evict_func_t *evict_func)
2216 {
2217 return (dmu_buf_update_user(db_fake, NULL, user_ptr,
2218 user_data_ptr_ptr, evict_func));
2219 }
2220
2221 void *
2222 dmu_buf_set_user_ie(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr,
2223 dmu_buf_evict_func_t *evict_func)
2224 {
2225 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2226
2227 db->db_immediate_evict = TRUE;
2228 return (dmu_buf_update_user(db_fake, NULL, user_ptr,
2229 user_data_ptr_ptr, evict_func));
2230 }
2231
2232 void *
2233 dmu_buf_update_user(dmu_buf_t *db_fake, void *old_user_ptr, void *user_ptr,
2234 void *user_data_ptr_ptr, dmu_buf_evict_func_t *evict_func)
2235 {
2236 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2237 ASSERT(db->db_level == 0);
2238
2239 ASSERT((user_ptr == NULL) == (evict_func == NULL));
2240
2241 mutex_enter(&db->db_mtx);
2242
2243 if (db->db_user_ptr == old_user_ptr) {
2244 db->db_user_ptr = user_ptr;
2245 db->db_user_data_ptr_ptr = user_data_ptr_ptr;
2246 db->db_evict_func = evict_func;
2247
2248 dbuf_update_data(db);
2249 } else {
2250 old_user_ptr = db->db_user_ptr;
2251 }
2252
2253 mutex_exit(&db->db_mtx);
2254 return (old_user_ptr);
2255 }
2256
2257 void *
2258 dmu_buf_get_user(dmu_buf_t *db_fake)
2259 {
2260 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2261 ASSERT(!refcount_is_zero(&db->db_holds));
2262
2263 return (db->db_user_ptr);
2264 }
2265
2266 boolean_t
2267 dmu_buf_freeable(dmu_buf_t *dbuf)
2268 {
2269 boolean_t res = B_FALSE;
2270 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2271
2272 if (db->db_blkptr)
2273 res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset,
2274 db->db_blkptr, db->db_blkptr->blk_birth);
2275
2276 return (res);
2277 }
2278
2279 blkptr_t *
2280 dmu_buf_get_blkptr(dmu_buf_t *db)
2281 {
2282 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2283 return (dbi->db_blkptr);
2284 }
2285
2286 static void
2287 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
2288 {
2289 /* ASSERT(dmu_tx_is_syncing(tx) */
2290 ASSERT(MUTEX_HELD(&db->db_mtx));
2291
2292 if (db->db_blkptr != NULL)
2293 return;
2294
2295 if (db->db_blkid == DMU_SPILL_BLKID) {
2296 db->db_blkptr = &dn->dn_phys->dn_spill;
2297 BP_ZERO(db->db_blkptr);
2298 return;
2299 }
2300 if (db->db_level == dn->dn_phys->dn_nlevels-1) {
2301 /*
2302 * This buffer was allocated at a time when there was
2303 * no available blkptrs from the dnode, or it was
2304 * inappropriate to hook it in (i.e., nlevels mis-match).
2305 */
2306 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
2307 ASSERT(db->db_parent == NULL);
2308 db->db_parent = dn->dn_dbuf;
2309 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
2310 DBUF_VERIFY(db);
2311 } else {
2312 dmu_buf_impl_t *parent = db->db_parent;
2313 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2314
2315 ASSERT(dn->dn_phys->dn_nlevels > 1);
2316 if (parent == NULL) {
2317 mutex_exit(&db->db_mtx);
2318 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2319 (void) dbuf_hold_impl(dn, db->db_level+1,
2320 db->db_blkid >> epbs, FALSE, db, &parent);
2321 rw_exit(&dn->dn_struct_rwlock);
2322 mutex_enter(&db->db_mtx);
2323 db->db_parent = parent;
2324 }
2325 db->db_blkptr = (blkptr_t *)parent->db.db_data +
2326 (db->db_blkid & ((1ULL << epbs) - 1));
2327 DBUF_VERIFY(db);
2328 }
2329 }
2330
2331 static void
2332 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2333 {
2334 dmu_buf_impl_t *db = dr->dr_dbuf;
2335 dnode_t *dn;
2336 zio_t *zio;
2337
2338 ASSERT(dmu_tx_is_syncing(tx));
2339
2340 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2341
2342 mutex_enter(&db->db_mtx);
2343
2344 ASSERT(db->db_level > 0);
2345 DBUF_VERIFY(db);
2346
2347 /* Read the block if it hasn't been read yet. */
2348 if (db->db_buf == NULL) {
2349 mutex_exit(&db->db_mtx);
2350 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
2351 mutex_enter(&db->db_mtx);
2352 }
2353 ASSERT3U(db->db_state, ==, DB_CACHED);
2354 ASSERT(db->db_buf != NULL);
2355
2356 DB_DNODE_ENTER(db);
2357 dn = DB_DNODE(db);
2358 /* Indirect block size must match what the dnode thinks it is. */
2359 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2360 dbuf_check_blkptr(dn, db);
2361 DB_DNODE_EXIT(db);
2362
2363 /* Provide the pending dirty record to child dbufs */
2364 db->db_data_pending = dr;
2365
2366 mutex_exit(&db->db_mtx);
2367 dbuf_write(dr, db->db_buf, tx);
2368
2369 zio = dr->dr_zio;
2370 mutex_enter(&dr->dt.di.dr_mtx);
2371 dbuf_sync_list(&dr->dt.di.dr_children, tx);
2372 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2373 mutex_exit(&dr->dt.di.dr_mtx);
2374 zio_nowait(zio);
2375 }
2376
2377 static void
2378 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2379 {
2380 arc_buf_t **datap = &dr->dt.dl.dr_data;
2381 dmu_buf_impl_t *db = dr->dr_dbuf;
2382 dnode_t *dn;
2383 objset_t *os;
2384 uint64_t txg = tx->tx_txg;
2385
2386 ASSERT(dmu_tx_is_syncing(tx));
2387
2388 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2389
2390 mutex_enter(&db->db_mtx);
2391 /*
2392 * To be synced, we must be dirtied. But we
2393 * might have been freed after the dirty.
2394 */
2395 if (db->db_state == DB_UNCACHED) {
2396 /* This buffer has been freed since it was dirtied */
2397 ASSERT(db->db.db_data == NULL);
2398 } else if (db->db_state == DB_FILL) {
2399 /* This buffer was freed and is now being re-filled */
2400 ASSERT(db->db.db_data != dr->dt.dl.dr_data);
2401 } else {
2402 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
2403 }
2404 DBUF_VERIFY(db);
2405
2406 DB_DNODE_ENTER(db);
2407 dn = DB_DNODE(db);
2408
2409 if (db->db_blkid == DMU_SPILL_BLKID) {
2410 mutex_enter(&dn->dn_mtx);
2411 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
2412 mutex_exit(&dn->dn_mtx);
2413 }
2414
2415 /*
2416 * If this is a bonus buffer, simply copy the bonus data into the
2417 * dnode. It will be written out when the dnode is synced (and it
2418 * will be synced, since it must have been dirty for dbuf_sync to
2419 * be called).
2420 */
2421 if (db->db_blkid == DMU_BONUS_BLKID) {
2422 dbuf_dirty_record_t **drp;
2423
2424 ASSERT(*datap != NULL);
2425 ASSERT0(db->db_level);
2426 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN);
2427 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen);
2428 DB_DNODE_EXIT(db);
2429
2430 if (*datap != db->db.db_data) {
2431 zio_buf_free(*datap, DN_MAX_BONUSLEN);
2432 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
2433 }
2434 db->db_data_pending = NULL;
2435 drp = &db->db_last_dirty;
2436 while (*drp != dr)
2437 drp = &(*drp)->dr_next;
2438 ASSERT(dr->dr_next == NULL);
2439 ASSERT(dr->dr_dbuf == db);
2440 *drp = dr->dr_next;
2441 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2442 ASSERT(db->db_dirtycnt > 0);
2443 db->db_dirtycnt -= 1;
2444 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
2445 return;
2446 }
2447
2448 os = dn->dn_objset;
2449
2450 /*
2451 * This function may have dropped the db_mtx lock allowing a dmu_sync
2452 * operation to sneak in. As a result, we need to ensure that we
2453 * don't check the dr_override_state until we have returned from
2454 * dbuf_check_blkptr.
2455 */
2456 dbuf_check_blkptr(dn, db);
2457
2458 /*
2459 * If this buffer is in the middle of an immediate write,
2460 * wait for the synchronous IO to complete.
2461 */
2462 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
2463 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
2464 cv_wait(&db->db_changed, &db->db_mtx);
2465 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
2466 }
2467
2468 if (db->db_state != DB_NOFILL &&
2469 dn->dn_object != DMU_META_DNODE_OBJECT &&
2470 refcount_count(&db->db_holds) > 1 &&
2471 dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
2472 *datap == db->db_buf) {
2473 /*
2474 * If this buffer is currently "in use" (i.e., there
2475 * are active holds and db_data still references it),
2476 * then make a copy before we start the write so that
2477 * any modifications from the open txg will not leak
2478 * into this write.
2479 *
2480 * NOTE: this copy does not need to be made for
2481 * objects only modified in the syncing context (e.g.
2482 * DNONE_DNODE blocks).
2483 */
2484 int blksz = arc_buf_size(*datap);
2485 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2486 *datap = arc_buf_alloc(os->os_spa, blksz, db, type);
2487 bcopy(db->db.db_data, (*datap)->b_data, blksz);
2488 }
2489 db->db_data_pending = dr;
2490
2491 mutex_exit(&db->db_mtx);
2492
2493 dbuf_write(dr, *datap, tx);
2494
2495 ASSERT(!list_link_active(&dr->dr_dirty_node));
2496 if (dn->dn_object == DMU_META_DNODE_OBJECT) {
2497 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
2498 DB_DNODE_EXIT(db);
2499 } else {
2500 /*
2501 * Although zio_nowait() does not "wait for an IO", it does
2502 * initiate the IO. If this is an empty write it seems plausible
2503 * that the IO could actually be completed before the nowait
2504 * returns. We need to DB_DNODE_EXIT() first in case
2505 * zio_nowait() invalidates the dbuf.
2506 */
2507 DB_DNODE_EXIT(db);
2508 zio_nowait(dr->dr_zio);
2509 }
2510 }
2511
2512 void
2513 dbuf_sync_list(list_t *list, dmu_tx_t *tx)
2514 {
2515 dbuf_dirty_record_t *dr;
2516
2517 while (dr = list_head(list)) {
2518 if (dr->dr_zio != NULL) {
2519 /*
2520 * If we find an already initialized zio then we
2521 * are processing the meta-dnode, and we have finished.
2522 * The dbufs for all dnodes are put back on the list
2523 * during processing, so that we can zio_wait()
2524 * these IOs after initiating all child IOs.
2525 */
2526 ASSERT3U(dr->dr_dbuf->db.db_object, ==,
2527 DMU_META_DNODE_OBJECT);
2528 break;
2529 }
2530 list_remove(list, dr);
2531 if (dr->dr_dbuf->db_level > 0)
2532 dbuf_sync_indirect(dr, tx);
2533 else
2534 dbuf_sync_leaf(dr, tx);
2535 }
2536 }
2537
2538 /* ARGSUSED */
2539 static void
2540 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
2541 {
2542 dmu_buf_impl_t *db = vdb;
2543 dnode_t *dn;
2544 blkptr_t *bp = zio->io_bp;
2545 blkptr_t *bp_orig = &zio->io_bp_orig;
2546 spa_t *spa = zio->io_spa;
2547 int64_t delta;
2548 uint64_t fill = 0;
2549 int i;
2550
2551 ASSERT3P(db->db_blkptr, ==, bp);
2552
2553 DB_DNODE_ENTER(db);
2554 dn = DB_DNODE(db);
2555 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
2556 dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
2557 zio->io_prev_space_delta = delta;
2558
2559 if (bp->blk_birth != 0) {
2560 ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
2561 BP_GET_TYPE(bp) == dn->dn_type) ||
2562 (db->db_blkid == DMU_SPILL_BLKID &&
2563 BP_GET_TYPE(bp) == dn->dn_bonustype) ||
2564 BP_IS_EMBEDDED(bp));
2565 ASSERT(BP_GET_LEVEL(bp) == db->db_level);
2566 }
2567
2568 mutex_enter(&db->db_mtx);
2569
2570 #ifdef ZFS_DEBUG
2571 if (db->db_blkid == DMU_SPILL_BLKID) {
2572 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2573 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2574 db->db_blkptr == &dn->dn_phys->dn_spill);
2575 }
2576 #endif
2577
2578 if (db->db_level == 0) {
2579 mutex_enter(&dn->dn_mtx);
2580 if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
2581 db->db_blkid != DMU_SPILL_BLKID)
2582 dn->dn_phys->dn_maxblkid = db->db_blkid;
2583 mutex_exit(&dn->dn_mtx);
2584
2585 if (dn->dn_type == DMU_OT_DNODE) {
2586 dnode_phys_t *dnp = db->db.db_data;
2587 for (i = db->db.db_size >> DNODE_SHIFT; i > 0;
2588 i--, dnp++) {
2589 if (dnp->dn_type != DMU_OT_NONE)
2590 fill++;
2591 }
2592 } else {
2593 if (BP_IS_HOLE(bp)) {
2594 fill = 0;
2595 } else {
2596 fill = 1;
2597 }
2598 }
2599 } else {
2600 blkptr_t *ibp = db->db.db_data;
2601 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2602 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
2603 if (BP_IS_HOLE(ibp))
2604 continue;
2605 fill += BP_GET_FILL(ibp);
2606 }
2607 }
2608 DB_DNODE_EXIT(db);
2609
2610 if (!BP_IS_EMBEDDED(bp))
2611 bp->blk_fill = fill;
2612
2613 mutex_exit(&db->db_mtx);
2614 }
2615
2616 /*
2617 * The SPA will call this callback several times for each zio - once
2618 * for every physical child i/o (zio->io_phys_children times). This
2619 * allows the DMU to monitor the progress of each logical i/o. For example,
2620 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
2621 * block. There may be a long delay before all copies/fragments are completed,
2622 * so this callback allows us to retire dirty space gradually, as the physical
2623 * i/os complete.
2624 */
2625 /* ARGSUSED */
2626 static void
2627 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
2628 {
2629 dmu_buf_impl_t *db = arg;
2630 objset_t *os = db->db_objset;
2631 dsl_pool_t *dp = dmu_objset_pool(os);
2632 dbuf_dirty_record_t *dr;
2633 int delta = 0;
2634
2635 dr = db->db_data_pending;
2636 ASSERT3U(dr->dr_txg, ==, zio->io_txg);
2637
2638 /*
2639 * The callback will be called io_phys_children times. Retire one
2640 * portion of our dirty space each time we are called. Any rounding
2641 * error will be cleaned up by dsl_pool_sync()'s call to
2642 * dsl_pool_undirty_space().
2643 */
2644 delta = dr->dr_accounted / zio->io_phys_children;
2645 dsl_pool_undirty_space(dp, delta, zio->io_txg);
2646 }
2647
2648 /* ARGSUSED */
2649 static void
2650 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
2651 {
2652 dmu_buf_impl_t *db = vdb;
2653 blkptr_t *bp_orig = &zio->io_bp_orig;
2654 blkptr_t *bp = db->db_blkptr;
2655 objset_t *os = db->db_objset;
2656 dmu_tx_t *tx = os->os_synctx;
2657 dbuf_dirty_record_t **drp, *dr;
2658
2659 ASSERT0(zio->io_error);
2660 ASSERT(db->db_blkptr == bp);
2661
2662 /*
2663 * For nopwrites and rewrites we ensure that the bp matches our
2664 * original and bypass all the accounting.
2665 */
2666 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
2667 ASSERT(BP_EQUAL(bp, bp_orig));
2668 } else {
2669 dsl_dataset_t *ds = os->os_dsl_dataset;
2670 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
2671 dsl_dataset_block_born(ds, bp, tx);
2672 }
2673
2674 mutex_enter(&db->db_mtx);
2675
2676 DBUF_VERIFY(db);
2677
2678 drp = &db->db_last_dirty;
2679 while ((dr = *drp) != db->db_data_pending)
2680 drp = &dr->dr_next;
2681 ASSERT(!list_link_active(&dr->dr_dirty_node));
2682 ASSERT(dr->dr_dbuf == db);
2683 ASSERT(dr->dr_next == NULL);
2684 *drp = dr->dr_next;
2685
2686 #ifdef ZFS_DEBUG
2687 if (db->db_blkid == DMU_SPILL_BLKID) {
2688 dnode_t *dn;
2689
2690 DB_DNODE_ENTER(db);
2691 dn = DB_DNODE(db);
2692 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2693 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2694 db->db_blkptr == &dn->dn_phys->dn_spill);
2695 DB_DNODE_EXIT(db);
2696 }
2697 #endif
2698
2699 if (db->db_level == 0) {
2700 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2701 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
2702 if (db->db_state != DB_NOFILL) {
2703 if (dr->dt.dl.dr_data != db->db_buf)
2704 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data,
2705 db));
2706 else if (!arc_released(db->db_buf))
2707 arc_set_callback(db->db_buf, dbuf_do_evict, db);
2708 }
2709 } else {
2710 dnode_t *dn;
2711
2712 DB_DNODE_ENTER(db);
2713 dn = DB_DNODE(db);
2714 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2715 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
2716 if (!BP_IS_HOLE(db->db_blkptr)) {
2717 int epbs =
2718 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2719 ASSERT3U(db->db_blkid, <=,
2720 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs));
2721 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
2722 db->db.db_size);
2723 if (!arc_released(db->db_buf))
2724 arc_set_callback(db->db_buf, dbuf_do_evict, db);
2725 }
2726 DB_DNODE_EXIT(db);
2727 mutex_destroy(&dr->dt.di.dr_mtx);
2728 list_destroy(&dr->dt.di.dr_children);
2729 }
2730 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2731
2732 cv_broadcast(&db->db_changed);
2733 ASSERT(db->db_dirtycnt > 0);
2734 db->db_dirtycnt -= 1;
2735 db->db_data_pending = NULL;
2736 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg);
2737 }
2738
2739 static void
2740 dbuf_write_nofill_ready(zio_t *zio)
2741 {
2742 dbuf_write_ready(zio, NULL, zio->io_private);
2743 }
2744
2745 static void
2746 dbuf_write_nofill_done(zio_t *zio)
2747 {
2748 dbuf_write_done(zio, NULL, zio->io_private);
2749 }
2750
2751 static void
2752 dbuf_write_override_ready(zio_t *zio)
2753 {
2754 dbuf_dirty_record_t *dr = zio->io_private;
2755 dmu_buf_impl_t *db = dr->dr_dbuf;
2756
2757 dbuf_write_ready(zio, NULL, db);
2758 }
2759
2760 static void
2761 dbuf_write_override_done(zio_t *zio)
2762 {
2763 dbuf_dirty_record_t *dr = zio->io_private;
2764 dmu_buf_impl_t *db = dr->dr_dbuf;
2765 blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
2766
2767 mutex_enter(&db->db_mtx);
2768 if (!BP_EQUAL(zio->io_bp, obp)) {
2769 if (!BP_IS_HOLE(obp))
2770 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
2771 arc_release(dr->dt.dl.dr_data, db);
2772 }
2773 mutex_exit(&db->db_mtx);
2774
2775 dbuf_write_done(zio, NULL, db);
2776 }
2777
2778 /* Issue I/O to commit a dirty buffer to disk. */
2779 static void
2780 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
2781 {
2782 dmu_buf_impl_t *db = dr->dr_dbuf;
2783 dnode_t *dn;
2784 objset_t *os;
2785 dmu_buf_impl_t *parent = db->db_parent;
2786 uint64_t txg = tx->tx_txg;
2787 zbookmark_phys_t zb;
2788 zio_prop_t zp;
2789 zio_t *zio;
2790 int wp_flag = 0;
2791
2792 DB_DNODE_ENTER(db);
2793 dn = DB_DNODE(db);
2794 os = dn->dn_objset;
2795
2796 if (db->db_state != DB_NOFILL) {
2797 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
2798 /*
2799 * Private object buffers are released here rather
2800 * than in dbuf_dirty() since they are only modified
2801 * in the syncing context and we don't want the
2802 * overhead of making multiple copies of the data.
2803 */
2804 if (BP_IS_HOLE(db->db_blkptr)) {
2805 arc_buf_thaw(data);
2806 } else {
2807 dbuf_release_bp(db);
2808 }
2809 }
2810 }
2811
2812 if (parent != dn->dn_dbuf) {
2813 /* Our parent is an indirect block. */
2814 /* We have a dirty parent that has been scheduled for write. */
2815 ASSERT(parent && parent->db_data_pending);
2816 /* Our parent's buffer is one level closer to the dnode. */
2817 ASSERT(db->db_level == parent->db_level-1);
2818 /*
2819 * We're about to modify our parent's db_data by modifying
2820 * our block pointer, so the parent must be released.
2821 */
2822 ASSERT(arc_released(parent->db_buf));
2823 zio = parent->db_data_pending->dr_zio;
2824 } else {
2825 /* Our parent is the dnode itself. */
2826 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
2827 db->db_blkid != DMU_SPILL_BLKID) ||
2828 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
2829 if (db->db_blkid != DMU_SPILL_BLKID)
2830 ASSERT3P(db->db_blkptr, ==,
2831 &dn->dn_phys->dn_blkptr[db->db_blkid]);
2832 zio = dn->dn_zio;
2833 }
2834
2835 ASSERT(db->db_level == 0 || data == db->db_buf);
2836 ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
2837 ASSERT(zio);
2838
2839 SET_BOOKMARK(&zb, os->os_dsl_dataset ?
2840 os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
2841 db->db.db_object, db->db_level, db->db_blkid);
2842
2843 if (db->db_blkid == DMU_SPILL_BLKID)
2844 wp_flag = WP_SPILL;
2845 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
2846
2847 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
2848 DB_DNODE_EXIT(db);
2849
2850 if (dr->dr_zero_write) {
2851 zp.zp_zero_write = B_TRUE;
2852
2853 if (!spa_feature_is_active(os->os_spa, SPA_FEATURE_SPACE_RESERVATION))
2854 {
2855 spa_feature_incr(os->os_spa,
2856 SPA_FEATURE_SPACE_RESERVATION, tx);
2857 }
2858 }
2859
2860 if (db->db_level == 0 &&
2861 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
2862 /*
2863 * The BP for this block has been provided by open context
2864 * (by dmu_sync() or dmu_buf_write_embedded()).
2865 */
2866 void *contents = (data != NULL) ? data->b_data : NULL;
2867
2868 dr->dr_zio = zio_write(zio, os->os_spa, txg,
2869 db->db_blkptr, contents, db->db.db_size, &zp,
2870 dbuf_write_override_ready, NULL, dbuf_write_override_done,
2871 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
2872 mutex_enter(&db->db_mtx);
2873 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
2874 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
2875 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
2876 mutex_exit(&db->db_mtx);
2877 } else if (db->db_state == DB_NOFILL) {
2878 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF ||
2879 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY);
2880 dr->dr_zio = zio_write(zio, os->os_spa, txg,
2881 db->db_blkptr, NULL, db->db.db_size, &zp,
2882 dbuf_write_nofill_ready, NULL, dbuf_write_nofill_done, db,
2883 ZIO_PRIORITY_ASYNC_WRITE,
2884 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
2885 } else {
2886 ASSERT(arc_released(data));
2887 dr->dr_zio = arc_write(zio, os->os_spa, txg,
2888 db->db_blkptr, data, DBUF_IS_L2CACHEABLE(db),
2889 DBUF_IS_L2COMPRESSIBLE(db), &zp, dbuf_write_ready,
2890 dbuf_write_physdone, dbuf_write_done, db,
2891 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
2892 }
2893 }