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 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
28 */
29
30 #include <sys/zfs_context.h>
31 #include <sys/dmu.h>
32 #include <sys/dmu_send.h>
33 #include <sys/dmu_impl.h>
34 #include <sys/dbuf.h>
35 #include <sys/dmu_objset.h>
36 #include <sys/dsl_dataset.h>
37 #include <sys/dsl_dir.h>
38 #include <sys/dmu_tx.h>
39 #include <sys/spa.h>
40 #include <sys/zio.h>
41 #include <sys/dmu_zfetch.h>
42 #include <sys/sa.h>
43 #include <sys/sa_impl.h>
44 #include <sys/zfeature.h>
45 #include <sys/blkptr.h>
46 #include <sys/range_tree.h>
47
48 /*
49 * Number of times that zfs_free_range() took the slow path while doing
50 * a zfs receive. A nonzero value indicates a potential performance problem.
51 */
52 uint64_t zfs_free_range_recv_miss;
53
54 static void dbuf_destroy(dmu_buf_impl_t *db);
55 static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
56 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
57
58 #ifndef __lint
59 extern inline void dmu_buf_init_user(dmu_buf_user_t *dbu,
60 dmu_buf_evict_func_t *evict_func, dmu_buf_t **clear_on_evict_dbufp);
61 #endif /* ! __lint */
62
63 /*
64 * Global data structures and functions for the dbuf cache.
65 */
66 static kmem_cache_t *dbuf_cache;
67 static taskq_t *dbu_evict_taskq;
68
69 /* ARGSUSED */
70 static int
71 dbuf_cons(void *vdb, void *unused, int kmflag)
72 {
73 dmu_buf_impl_t *db = vdb;
74 bzero(db, sizeof (dmu_buf_impl_t));
75
76 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
77 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
78 refcount_create(&db->db_holds);
79
80 return (0);
81 }
82
83 /* ARGSUSED */
84 static void
85 dbuf_dest(void *vdb, void *unused)
86 {
87 dmu_buf_impl_t *db = vdb;
88 mutex_destroy(&db->db_mtx);
89 cv_destroy(&db->db_changed);
90 refcount_destroy(&db->db_holds);
91 }
92
93 /*
94 * dbuf hash table routines
95 */
96 static dbuf_hash_table_t dbuf_hash_table;
97
98 static uint64_t dbuf_hash_count;
99
100 static uint64_t
101 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
102 {
103 uintptr_t osv = (uintptr_t)os;
104 uint64_t crc = -1ULL;
105
106 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
107 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF];
108 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF];
109 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF];
110 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF];
111 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF];
112 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF];
113
114 crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16);
115
116 return (crc);
117 }
118
119 #define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid);
120
121 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
122 ((dbuf)->db.db_object == (obj) && \
123 (dbuf)->db_objset == (os) && \
124 (dbuf)->db_level == (level) && \
125 (dbuf)->db_blkid == (blkid))
126
127 dmu_buf_impl_t *
128 dbuf_find(dnode_t *dn, uint8_t level, uint64_t blkid)
129 {
130 dbuf_hash_table_t *h = &dbuf_hash_table;
131 objset_t *os = dn->dn_objset;
132 uint64_t obj = dn->dn_object;
133 uint64_t hv = DBUF_HASH(os, obj, level, blkid);
134 uint64_t idx = hv & h->hash_table_mask;
135 dmu_buf_impl_t *db;
136
137 mutex_enter(DBUF_HASH_MUTEX(h, idx));
138 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
139 if (DBUF_EQUAL(db, os, obj, level, blkid)) {
140 mutex_enter(&db->db_mtx);
141 if (db->db_state != DB_EVICTING) {
142 mutex_exit(DBUF_HASH_MUTEX(h, idx));
143 return (db);
144 }
145 mutex_exit(&db->db_mtx);
146 }
147 }
148 mutex_exit(DBUF_HASH_MUTEX(h, idx));
149 return (NULL);
150 }
151
152 /*
153 * Insert an entry into the hash table. If there is already an element
154 * equal to elem in the hash table, then the already existing element
155 * will be returned and the new element will not be inserted.
156 * Otherwise returns NULL.
157 */
158 static dmu_buf_impl_t *
159 dbuf_hash_insert(dmu_buf_impl_t *db)
160 {
161 dbuf_hash_table_t *h = &dbuf_hash_table;
162 objset_t *os = db->db_objset;
163 uint64_t obj = db->db.db_object;
164 int level = db->db_level;
165 uint64_t blkid = db->db_blkid;
166 uint64_t hv = DBUF_HASH(os, obj, level, blkid);
167 uint64_t idx = hv & h->hash_table_mask;
168 dmu_buf_impl_t *dbf;
169
170 mutex_enter(DBUF_HASH_MUTEX(h, idx));
171 for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) {
172 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
173 mutex_enter(&dbf->db_mtx);
174 if (dbf->db_state != DB_EVICTING) {
175 mutex_exit(DBUF_HASH_MUTEX(h, idx));
176 return (dbf);
177 }
178 mutex_exit(&dbf->db_mtx);
179 }
180 }
181
182 mutex_enter(&db->db_mtx);
183 db->db_hash_next = h->hash_table[idx];
184 h->hash_table[idx] = db;
185 mutex_exit(DBUF_HASH_MUTEX(h, idx));
186 atomic_inc_64(&dbuf_hash_count);
187
188 return (NULL);
189 }
190
191 /*
192 * Remove an entry from the hash table. It must be in the EVICTING state.
193 */
194 static void
195 dbuf_hash_remove(dmu_buf_impl_t *db)
196 {
197 dbuf_hash_table_t *h = &dbuf_hash_table;
198 uint64_t hv = DBUF_HASH(db->db_objset, db->db.db_object,
199 db->db_level, db->db_blkid);
200 uint64_t idx = hv & h->hash_table_mask;
201 dmu_buf_impl_t *dbf, **dbp;
202
203 /*
204 * We musn't hold db_mtx to maintain lock ordering:
205 * DBUF_HASH_MUTEX > db_mtx.
206 */
207 ASSERT(refcount_is_zero(&db->db_holds));
208 ASSERT(db->db_state == DB_EVICTING);
209 ASSERT(!MUTEX_HELD(&db->db_mtx));
210
211 mutex_enter(DBUF_HASH_MUTEX(h, idx));
212 dbp = &h->hash_table[idx];
213 while ((dbf = *dbp) != db) {
214 dbp = &dbf->db_hash_next;
215 ASSERT(dbf != NULL);
216 }
217 *dbp = db->db_hash_next;
218 db->db_hash_next = NULL;
219 mutex_exit(DBUF_HASH_MUTEX(h, idx));
220 atomic_dec_64(&dbuf_hash_count);
221 }
222
223 static arc_evict_func_t dbuf_do_evict;
224
225 typedef enum {
226 DBVU_EVICTING,
227 DBVU_NOT_EVICTING
228 } dbvu_verify_type_t;
229
230 static void
231 dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type)
232 {
233 #ifdef ZFS_DEBUG
234 int64_t holds;
235
236 if (db->db_user == NULL)
237 return;
238
239 /* Only data blocks support the attachment of user data. */
240 ASSERT(db->db_level == 0);
241
242 /* Clients must resolve a dbuf before attaching user data. */
243 ASSERT(db->db.db_data != NULL);
244 ASSERT3U(db->db_state, ==, DB_CACHED);
245
246 holds = refcount_count(&db->db_holds);
247 if (verify_type == DBVU_EVICTING) {
248 /*
249 * Immediate eviction occurs when holds == dirtycnt.
250 * For normal eviction buffers, holds is zero on
251 * eviction, except when dbuf_fix_old_data() calls
252 * dbuf_clear_data(). However, the hold count can grow
253 * during eviction even though db_mtx is held (see
254 * dmu_bonus_hold() for an example), so we can only
255 * test the generic invariant that holds >= dirtycnt.
256 */
257 ASSERT3U(holds, >=, db->db_dirtycnt);
258 } else {
259 if (db->db_immediate_evict == TRUE)
260 ASSERT3U(holds, >=, db->db_dirtycnt);
261 else
262 ASSERT3U(holds, >, 0);
263 }
264 #endif
265 }
266
267 static void
268 dbuf_evict_user(dmu_buf_impl_t *db)
269 {
270 dmu_buf_user_t *dbu = db->db_user;
271
272 ASSERT(MUTEX_HELD(&db->db_mtx));
273
274 if (dbu == NULL)
275 return;
276
277 dbuf_verify_user(db, DBVU_EVICTING);
278 db->db_user = NULL;
279
280 #ifdef ZFS_DEBUG
281 if (dbu->dbu_clear_on_evict_dbufp != NULL)
282 *dbu->dbu_clear_on_evict_dbufp = NULL;
283 #endif
284
285 /*
286 * Invoke the callback from a taskq to avoid lock order reversals
287 * and limit stack depth.
288 */
289 taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func, dbu, 0,
290 &dbu->dbu_tqent);
291 }
292
293 boolean_t
294 dbuf_is_metadata(dmu_buf_impl_t *db)
295 {
296 if (db->db_level > 0) {
297 return (B_TRUE);
298 } else {
299 boolean_t is_metadata;
300
301 DB_DNODE_ENTER(db);
302 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type);
303 DB_DNODE_EXIT(db);
304
305 return (is_metadata);
306 }
307 }
308
309 void
310 dbuf_evict(dmu_buf_impl_t *db)
311 {
312 ASSERT(MUTEX_HELD(&db->db_mtx));
313 ASSERT(db->db_buf == NULL);
314 ASSERT(db->db_data_pending == NULL);
315
316 dbuf_clear(db);
317 dbuf_destroy(db);
318 }
319
320 void
321 dbuf_init(void)
322 {
323 uint64_t hsize = 1ULL << 16;
324 dbuf_hash_table_t *h = &dbuf_hash_table;
325 int i;
326
327 /*
328 * The hash table is big enough to fill all of physical memory
329 * with an average 4K block size. The table will take up
330 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
331 */
332 while (hsize * 4096 < physmem * PAGESIZE)
333 hsize <<= 1;
334
335 retry:
336 h->hash_table_mask = hsize - 1;
337 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
338 if (h->hash_table == NULL) {
339 /* XXX - we should really return an error instead of assert */
340 ASSERT(hsize > (1ULL << 10));
341 hsize >>= 1;
342 goto retry;
343 }
344
345 dbuf_cache = kmem_cache_create("dmu_buf_impl_t",
346 sizeof (dmu_buf_impl_t),
347 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
348
349 for (i = 0; i < DBUF_MUTEXES; i++)
350 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
351
352 /*
353 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
354 * configuration is not required.
355 */
356 dbu_evict_taskq = taskq_create("dbu_evict", 1, minclsyspri, 0, 0, 0);
357 }
358
359 void
360 dbuf_fini(void)
361 {
362 dbuf_hash_table_t *h = &dbuf_hash_table;
363 int i;
364
365 for (i = 0; i < DBUF_MUTEXES; i++)
366 mutex_destroy(&h->hash_mutexes[i]);
367 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
368 kmem_cache_destroy(dbuf_cache);
369 taskq_destroy(dbu_evict_taskq);
370 }
371
372 /*
373 * Other stuff.
374 */
375
376 #ifdef ZFS_DEBUG
377 static void
378 dbuf_verify(dmu_buf_impl_t *db)
379 {
380 dnode_t *dn;
381 dbuf_dirty_record_t *dr;
382
383 ASSERT(MUTEX_HELD(&db->db_mtx));
384
385 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
386 return;
387
388 ASSERT(db->db_objset != NULL);
389 DB_DNODE_ENTER(db);
390 dn = DB_DNODE(db);
391 if (dn == NULL) {
392 ASSERT(db->db_parent == NULL);
393 ASSERT(db->db_blkptr == NULL);
394 } else {
395 ASSERT3U(db->db.db_object, ==, dn->dn_object);
396 ASSERT3P(db->db_objset, ==, dn->dn_objset);
397 ASSERT3U(db->db_level, <, dn->dn_nlevels);
398 ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
399 db->db_blkid == DMU_SPILL_BLKID ||
400 !avl_is_empty(&dn->dn_dbufs));
401 }
402 if (db->db_blkid == DMU_BONUS_BLKID) {
403 ASSERT(dn != NULL);
404 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
405 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
406 } else if (db->db_blkid == DMU_SPILL_BLKID) {
407 ASSERT(dn != NULL);
408 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
409 ASSERT0(db->db.db_offset);
410 } else {
411 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
412 }
413
414 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
415 ASSERT(dr->dr_dbuf == db);
416
417 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
418 ASSERT(dr->dr_dbuf == db);
419
420 /*
421 * We can't assert that db_size matches dn_datablksz because it
422 * can be momentarily different when another thread is doing
423 * dnode_set_blksz().
424 */
425 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
426 dr = db->db_data_pending;
427 /*
428 * It should only be modified in syncing context, so
429 * make sure we only have one copy of the data.
430 */
431 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
432 }
433
434 /* verify db->db_blkptr */
435 if (db->db_blkptr) {
436 if (db->db_parent == dn->dn_dbuf) {
437 /* db is pointed to by the dnode */
438 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
439 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
440 ASSERT(db->db_parent == NULL);
441 else
442 ASSERT(db->db_parent != NULL);
443 if (db->db_blkid != DMU_SPILL_BLKID)
444 ASSERT3P(db->db_blkptr, ==,
445 &dn->dn_phys->dn_blkptr[db->db_blkid]);
446 } else {
447 /* db is pointed to by an indirect block */
448 int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
449 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
450 ASSERT3U(db->db_parent->db.db_object, ==,
451 db->db.db_object);
452 /*
453 * dnode_grow_indblksz() can make this fail if we don't
454 * have the struct_rwlock. XXX indblksz no longer
455 * grows. safe to do this now?
456 */
457 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
458 ASSERT3P(db->db_blkptr, ==,
459 ((blkptr_t *)db->db_parent->db.db_data +
460 db->db_blkid % epb));
461 }
462 }
463 }
464 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
465 (db->db_buf == NULL || db->db_buf->b_data) &&
466 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
467 db->db_state != DB_FILL && !dn->dn_free_txg) {
468 /*
469 * If the blkptr isn't set but they have nonzero data,
470 * it had better be dirty, otherwise we'll lose that
471 * data when we evict this buffer.
472 */
473 if (db->db_dirtycnt == 0) {
474 uint64_t *buf = db->db.db_data;
475 int i;
476
477 for (i = 0; i < db->db.db_size >> 3; i++) {
478 ASSERT(buf[i] == 0);
479 }
480 }
481 }
482 DB_DNODE_EXIT(db);
483 }
484 #endif
485
486 static void
487 dbuf_clear_data(dmu_buf_impl_t *db)
488 {
489 ASSERT(MUTEX_HELD(&db->db_mtx));
490 dbuf_evict_user(db);
491 db->db_buf = NULL;
492 db->db.db_data = NULL;
493 if (db->db_state != DB_NOFILL)
494 db->db_state = DB_UNCACHED;
495 }
496
497 static void
498 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
499 {
500 ASSERT(MUTEX_HELD(&db->db_mtx));
501 ASSERT(buf != NULL);
502
503 db->db_buf = buf;
504 ASSERT(buf->b_data != NULL);
505 db->db.db_data = buf->b_data;
506 if (!arc_released(buf))
507 arc_set_callback(buf, dbuf_do_evict, db);
508 }
509
510 /*
511 * Loan out an arc_buf for read. Return the loaned arc_buf.
512 */
513 arc_buf_t *
514 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
515 {
516 arc_buf_t *abuf;
517
518 mutex_enter(&db->db_mtx);
519 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
520 int blksz = db->db.db_size;
521 spa_t *spa = db->db_objset->os_spa;
522
523 mutex_exit(&db->db_mtx);
524 abuf = arc_loan_buf(spa, blksz);
525 bcopy(db->db.db_data, abuf->b_data, blksz);
526 } else {
527 abuf = db->db_buf;
528 arc_loan_inuse_buf(abuf, db);
529 dbuf_clear_data(db);
530 mutex_exit(&db->db_mtx);
531 }
532 return (abuf);
533 }
534
535 uint64_t
536 dbuf_whichblock(dnode_t *dn, uint64_t offset)
537 {
538 if (dn->dn_datablkshift) {
539 return (offset >> dn->dn_datablkshift);
540 } else {
541 ASSERT3U(offset, <, dn->dn_datablksz);
542 return (0);
543 }
544 }
545
546 static void
547 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
548 {
549 dmu_buf_impl_t *db = vdb;
550
551 mutex_enter(&db->db_mtx);
552 ASSERT3U(db->db_state, ==, DB_READ);
553 /*
554 * All reads are synchronous, so we must have a hold on the dbuf
555 */
556 ASSERT(refcount_count(&db->db_holds) > 0);
557 ASSERT(db->db_buf == NULL);
558 ASSERT(db->db.db_data == NULL);
559 if (db->db_level == 0 && db->db_freed_in_flight) {
560 /* we were freed in flight; disregard any error */
561 arc_release(buf, db);
562 bzero(buf->b_data, db->db.db_size);
563 arc_buf_freeze(buf);
564 db->db_freed_in_flight = FALSE;
565 dbuf_set_data(db, buf);
566 db->db_state = DB_CACHED;
567 } else if (zio == NULL || zio->io_error == 0) {
568 dbuf_set_data(db, buf);
569 db->db_state = DB_CACHED;
570 } else {
571 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
572 ASSERT3P(db->db_buf, ==, NULL);
573 VERIFY(arc_buf_remove_ref(buf, db));
574 db->db_state = DB_UNCACHED;
575 }
576 cv_broadcast(&db->db_changed);
577 dbuf_rele_and_unlock(db, NULL);
578 }
579
580 static void
581 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t *flags)
582 {
583 dnode_t *dn;
584 zbookmark_phys_t zb;
585 arc_flags_t aflags = ARC_FLAG_NOWAIT;
586
587 DB_DNODE_ENTER(db);
588 dn = DB_DNODE(db);
589 ASSERT(!refcount_is_zero(&db->db_holds));
590 /* We need the struct_rwlock to prevent db_blkptr from changing. */
591 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
592 ASSERT(MUTEX_HELD(&db->db_mtx));
593 ASSERT(db->db_state == DB_UNCACHED);
594 ASSERT(db->db_buf == NULL);
595
596 if (db->db_blkid == DMU_BONUS_BLKID) {
597 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
598
599 ASSERT3U(bonuslen, <=, db->db.db_size);
600 db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN);
601 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
602 if (bonuslen < DN_MAX_BONUSLEN)
603 bzero(db->db.db_data, DN_MAX_BONUSLEN);
604 if (bonuslen)
605 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
606 DB_DNODE_EXIT(db);
607 db->db_state = DB_CACHED;
608 mutex_exit(&db->db_mtx);
609 return;
610 }
611
612 /*
613 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
614 * processes the delete record and clears the bp while we are waiting
615 * for the dn_mtx (resulting in a "no" from block_freed).
616 */
617 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
618 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
619 BP_IS_HOLE(db->db_blkptr)))) {
620 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
621
622 DB_DNODE_EXIT(db);
623 dbuf_set_data(db, arc_buf_alloc(db->db_objset->os_spa,
624 db->db.db_size, db, type));
625 bzero(db->db.db_data, db->db.db_size);
626 db->db_state = DB_CACHED;
627 *flags |= DB_RF_CACHED;
628 mutex_exit(&db->db_mtx);
629 return;
630 }
631
632 DB_DNODE_EXIT(db);
633
634 db->db_state = DB_READ;
635 mutex_exit(&db->db_mtx);
636
637 if (DBUF_IS_L2CACHEABLE(db))
638 aflags |= ARC_FLAG_L2CACHE;
639 if (DBUF_IS_L2COMPRESSIBLE(db))
640 aflags |= ARC_FLAG_L2COMPRESS;
641
642 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
643 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
644 db->db.db_object, db->db_level, db->db_blkid);
645
646 dbuf_add_ref(db, NULL);
647
648 (void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr,
649 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
650 (*flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
651 &aflags, &zb);
652 if (aflags & ARC_FLAG_CACHED)
653 *flags |= DB_RF_CACHED;
654 }
655
656 int
657 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
658 {
659 int err = 0;
660 boolean_t havepzio = (zio != NULL);
661 boolean_t prefetch;
662 dnode_t *dn;
663
664 /*
665 * We don't have to hold the mutex to check db_state because it
666 * can't be freed while we have a hold on the buffer.
667 */
668 ASSERT(!refcount_is_zero(&db->db_holds));
669
670 if (db->db_state == DB_NOFILL)
671 return (SET_ERROR(EIO));
672
673 DB_DNODE_ENTER(db);
674 dn = DB_DNODE(db);
675 if ((flags & DB_RF_HAVESTRUCT) == 0)
676 rw_enter(&dn->dn_struct_rwlock, RW_READER);
677
678 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
679 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
680 DBUF_IS_CACHEABLE(db);
681
682 mutex_enter(&db->db_mtx);
683 if (db->db_state == DB_CACHED) {
684 mutex_exit(&db->db_mtx);
685 if (prefetch)
686 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
687 db->db.db_size, TRUE);
688 if ((flags & DB_RF_HAVESTRUCT) == 0)
689 rw_exit(&dn->dn_struct_rwlock);
690 DB_DNODE_EXIT(db);
691 } else if (db->db_state == DB_UNCACHED) {
692 spa_t *spa = dn->dn_objset->os_spa;
693
694 if (zio == NULL)
695 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
696 dbuf_read_impl(db, zio, &flags);
697
698 /* dbuf_read_impl has dropped db_mtx for us */
699
700 if (prefetch)
701 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
702 db->db.db_size, flags & DB_RF_CACHED);
703
704 if ((flags & DB_RF_HAVESTRUCT) == 0)
705 rw_exit(&dn->dn_struct_rwlock);
706 DB_DNODE_EXIT(db);
707
708 if (!havepzio)
709 err = zio_wait(zio);
710 } else {
711 /*
712 * Another reader came in while the dbuf was in flight
713 * between UNCACHED and CACHED. Either a writer will finish
714 * writing the buffer (sending the dbuf to CACHED) or the
715 * first reader's request will reach the read_done callback
716 * and send the dbuf to CACHED. Otherwise, a failure
717 * occurred and the dbuf went to UNCACHED.
718 */
719 mutex_exit(&db->db_mtx);
720 if (prefetch)
721 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset,
722 db->db.db_size, TRUE);
723 if ((flags & DB_RF_HAVESTRUCT) == 0)
724 rw_exit(&dn->dn_struct_rwlock);
725 DB_DNODE_EXIT(db);
726
727 /* Skip the wait per the caller's request. */
728 mutex_enter(&db->db_mtx);
729 if ((flags & DB_RF_NEVERWAIT) == 0) {
730 while (db->db_state == DB_READ ||
731 db->db_state == DB_FILL) {
732 ASSERT(db->db_state == DB_READ ||
733 (flags & DB_RF_HAVESTRUCT) == 0);
734 DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *,
735 db, zio_t *, zio);
736 cv_wait(&db->db_changed, &db->db_mtx);
737 }
738 if (db->db_state == DB_UNCACHED)
739 err = SET_ERROR(EIO);
740 }
741 mutex_exit(&db->db_mtx);
742 }
743
744 ASSERT(err || havepzio || db->db_state == DB_CACHED);
745 return (err);
746 }
747
748 static void
749 dbuf_noread(dmu_buf_impl_t *db)
750 {
751 ASSERT(!refcount_is_zero(&db->db_holds));
752 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
753 mutex_enter(&db->db_mtx);
754 while (db->db_state == DB_READ || db->db_state == DB_FILL)
755 cv_wait(&db->db_changed, &db->db_mtx);
756 if (db->db_state == DB_UNCACHED) {
757 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
758 spa_t *spa = db->db_objset->os_spa;
759
760 ASSERT(db->db_buf == NULL);
761 ASSERT(db->db.db_data == NULL);
762 dbuf_set_data(db, arc_buf_alloc(spa, db->db.db_size, db, type));
763 db->db_state = DB_FILL;
764 } else if (db->db_state == DB_NOFILL) {
765 dbuf_clear_data(db);
766 } else {
767 ASSERT3U(db->db_state, ==, DB_CACHED);
768 }
769 mutex_exit(&db->db_mtx);
770 }
771
772 /*
773 * This is our just-in-time copy function. It makes a copy of
774 * buffers, that have been modified in a previous transaction
775 * group, before we modify them in the current active group.
776 *
777 * This function is used in two places: when we are dirtying a
778 * buffer for the first time in a txg, and when we are freeing
779 * a range in a dnode that includes this buffer.
780 *
781 * Note that when we are called from dbuf_free_range() we do
782 * not put a hold on the buffer, we just traverse the active
783 * dbuf list for the dnode.
784 */
785 static void
786 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
787 {
788 dbuf_dirty_record_t *dr = db->db_last_dirty;
789
790 ASSERT(MUTEX_HELD(&db->db_mtx));
791 ASSERT(db->db.db_data != NULL);
792 ASSERT(db->db_level == 0);
793 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
794
795 if (dr == NULL ||
796 (dr->dt.dl.dr_data !=
797 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
798 return;
799
800 /*
801 * If the last dirty record for this dbuf has not yet synced
802 * and its referencing the dbuf data, either:
803 * reset the reference to point to a new copy,
804 * or (if there a no active holders)
805 * just null out the current db_data pointer.
806 */
807 ASSERT(dr->dr_txg >= txg - 2);
808 if (db->db_blkid == DMU_BONUS_BLKID) {
809 /* Note that the data bufs here are zio_bufs */
810 dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
811 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
812 bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
813 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
814 int size = db->db.db_size;
815 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
816 spa_t *spa = db->db_objset->os_spa;
817
818 dr->dt.dl.dr_data = arc_buf_alloc(spa, size, db, type);
819 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
820 } else {
821 dbuf_clear_data(db);
822 }
823 }
824
825 void
826 dbuf_unoverride(dbuf_dirty_record_t *dr)
827 {
828 dmu_buf_impl_t *db = dr->dr_dbuf;
829 blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
830 uint64_t txg = dr->dr_txg;
831
832 ASSERT(MUTEX_HELD(&db->db_mtx));
833 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
834 ASSERT(db->db_level == 0);
835
836 if (db->db_blkid == DMU_BONUS_BLKID ||
837 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
838 return;
839
840 ASSERT(db->db_data_pending != dr);
841
842 /* free this block */
843 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite)
844 zio_free(db->db_objset->os_spa, txg, bp);
845
846 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
847 dr->dt.dl.dr_nopwrite = B_FALSE;
848
849 /*
850 * Release the already-written buffer, so we leave it in
851 * a consistent dirty state. Note that all callers are
852 * modifying the buffer, so they will immediately do
853 * another (redundant) arc_release(). Therefore, leave
854 * the buf thawed to save the effort of freezing &
855 * immediately re-thawing it.
856 */
857 arc_release(dr->dt.dl.dr_data, db);
858 }
859
860 /*
861 * Evict (if its unreferenced) or clear (if its referenced) any level-0
862 * data blocks in the free range, so that any future readers will find
863 * empty blocks.
864 *
865 * This is a no-op if the dataset is in the middle of an incremental
866 * receive; see comment below for details.
867 */
868 void
869 dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
870 dmu_tx_t *tx)
871 {
872 dmu_buf_impl_t db_search;
873 dmu_buf_impl_t *db, *db_next;
874 uint64_t txg = tx->tx_txg;
875 avl_index_t where;
876
877 if (end_blkid > dn->dn_maxblkid && (end_blkid != DMU_SPILL_BLKID))
878 end_blkid = dn->dn_maxblkid;
879 dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid);
880
881 db_search.db_level = 0;
882 db_search.db_blkid = start_blkid;
883 db_search.db_state = DB_SEARCH;
884
885 mutex_enter(&dn->dn_dbufs_mtx);
886 if (start_blkid >= dn->dn_unlisted_l0_blkid) {
887 /* There can't be any dbufs in this range; no need to search. */
888 #ifdef DEBUG
889 db = avl_find(&dn->dn_dbufs, &db_search, &where);
890 ASSERT3P(db, ==, NULL);
891 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
892 ASSERT(db == NULL || db->db_level > 0);
893 #endif
894 mutex_exit(&dn->dn_dbufs_mtx);
895 return;
896 } else if (dmu_objset_is_receiving(dn->dn_objset)) {
897 /*
898 * If we are receiving, we expect there to be no dbufs in
899 * the range to be freed, because receive modifies each
900 * block at most once, and in offset order. If this is
901 * not the case, it can lead to performance problems,
902 * so note that we unexpectedly took the slow path.
903 */
904 atomic_inc_64(&zfs_free_range_recv_miss);
905 }
906
907 db = avl_find(&dn->dn_dbufs, &db_search, &where);
908 ASSERT3P(db, ==, NULL);
909 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
910
911 for (; db != NULL; db = db_next) {
912 db_next = AVL_NEXT(&dn->dn_dbufs, db);
913 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
914
915 if (db->db_level != 0 || db->db_blkid > end_blkid) {
916 break;
917 }
918 ASSERT3U(db->db_blkid, >=, start_blkid);
919
920 /* found a level 0 buffer in the range */
921 mutex_enter(&db->db_mtx);
922 if (dbuf_undirty(db, tx)) {
923 /* mutex has been dropped and dbuf destroyed */
924 continue;
925 }
926
927 if (db->db_state == DB_UNCACHED ||
928 db->db_state == DB_NOFILL ||
929 db->db_state == DB_EVICTING) {
930 ASSERT(db->db.db_data == NULL);
931 mutex_exit(&db->db_mtx);
932 continue;
933 }
934 if (db->db_state == DB_READ || db->db_state == DB_FILL) {
935 /* will be handled in dbuf_read_done or dbuf_rele */
936 db->db_freed_in_flight = TRUE;
937 mutex_exit(&db->db_mtx);
938 continue;
939 }
940 if (refcount_count(&db->db_holds) == 0) {
941 ASSERT(db->db_buf);
942 dbuf_clear(db);
943 continue;
944 }
945 /* The dbuf is referenced */
946
947 if (db->db_last_dirty != NULL) {
948 dbuf_dirty_record_t *dr = db->db_last_dirty;
949
950 if (dr->dr_txg == txg) {
951 /*
952 * This buffer is "in-use", re-adjust the file
953 * size to reflect that this buffer may
954 * contain new data when we sync.
955 */
956 if (db->db_blkid != DMU_SPILL_BLKID &&
957 db->db_blkid > dn->dn_maxblkid)
958 dn->dn_maxblkid = db->db_blkid;
959 dbuf_unoverride(dr);
960 } else {
961 /*
962 * This dbuf is not dirty in the open context.
963 * Either uncache it (if its not referenced in
964 * the open context) or reset its contents to
965 * empty.
966 */
967 dbuf_fix_old_data(db, txg);
968 }
969 }
970 /* clear the contents if its cached */
971 if (db->db_state == DB_CACHED) {
972 ASSERT(db->db.db_data != NULL);
973 arc_release(db->db_buf, db);
974 bzero(db->db.db_data, db->db.db_size);
975 arc_buf_freeze(db->db_buf);
976 }
977
978 mutex_exit(&db->db_mtx);
979 }
980 mutex_exit(&dn->dn_dbufs_mtx);
981 }
982
983 static int
984 dbuf_block_freeable(dmu_buf_impl_t *db)
985 {
986 dsl_dataset_t *ds = db->db_objset->os_dsl_dataset;
987 uint64_t birth_txg = 0;
988
989 /*
990 * We don't need any locking to protect db_blkptr:
991 * If it's syncing, then db_last_dirty will be set
992 * so we'll ignore db_blkptr.
993 *
994 * This logic ensures that only block births for
995 * filled blocks are considered.
996 */
997 ASSERT(MUTEX_HELD(&db->db_mtx));
998 if (db->db_last_dirty && (db->db_blkptr == NULL ||
999 !BP_IS_HOLE(db->db_blkptr))) {
1000 birth_txg = db->db_last_dirty->dr_txg;
1001 } else if (db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) {
1002 birth_txg = db->db_blkptr->blk_birth;
1003 }
1004
1005 /*
1006 * If this block don't exist or is in a snapshot, it can't be freed.
1007 * Don't pass the bp to dsl_dataset_block_freeable() since we
1008 * are holding the db_mtx lock and might deadlock if we are
1009 * prefetching a dedup-ed block.
1010 */
1011 if (birth_txg != 0)
1012 return (ds == NULL ||
1013 dsl_dataset_block_freeable(ds, NULL, birth_txg));
1014 else
1015 return (B_FALSE);
1016 }
1017
1018 void
1019 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
1020 {
1021 arc_buf_t *buf, *obuf;
1022 int osize = db->db.db_size;
1023 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1024 dnode_t *dn;
1025
1026 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1027
1028 DB_DNODE_ENTER(db);
1029 dn = DB_DNODE(db);
1030
1031 /* XXX does *this* func really need the lock? */
1032 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1033
1034 /*
1035 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1036 * is OK, because there can be no other references to the db
1037 * when we are changing its size, so no concurrent DB_FILL can
1038 * be happening.
1039 */
1040 /*
1041 * XXX we should be doing a dbuf_read, checking the return
1042 * value and returning that up to our callers
1043 */
1044 dmu_buf_will_dirty(&db->db, tx);
1045
1046 /* create the data buffer for the new block */
1047 buf = arc_buf_alloc(dn->dn_objset->os_spa, size, db, type);
1048
1049 /* copy old block data to the new block */
1050 obuf = db->db_buf;
1051 bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
1052 /* zero the remainder */
1053 if (size > osize)
1054 bzero((uint8_t *)buf->b_data + osize, size - osize);
1055
1056 mutex_enter(&db->db_mtx);
1057 dbuf_set_data(db, buf);
1058 VERIFY(arc_buf_remove_ref(obuf, db));
1059 db->db.db_size = size;
1060
1061 if (db->db_level == 0) {
1062 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1063 db->db_last_dirty->dt.dl.dr_data = buf;
1064 }
1065 mutex_exit(&db->db_mtx);
1066
1067 dnode_willuse_space(dn, size-osize, tx);
1068 DB_DNODE_EXIT(db);
1069 }
1070
1071 void
1072 dbuf_release_bp(dmu_buf_impl_t *db)
1073 {
1074 objset_t *os = db->db_objset;
1075
1076 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1077 ASSERT(arc_released(os->os_phys_buf) ||
1078 list_link_active(&os->os_dsl_dataset->ds_synced_link));
1079 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1080
1081 (void) arc_release(db->db_buf, db);
1082 }
1083
1084 dbuf_dirty_record_t *
1085 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1086 {
1087 dnode_t *dn;
1088 objset_t *os;
1089 dbuf_dirty_record_t **drp, *dr;
1090 int drop_struct_lock = FALSE;
1091 boolean_t do_free_accounting = B_FALSE;
1092 int txgoff = tx->tx_txg & TXG_MASK;
1093
1094 ASSERT(tx->tx_txg != 0);
1095 ASSERT(!refcount_is_zero(&db->db_holds));
1096 DMU_TX_DIRTY_BUF(tx, db);
1097
1098 DB_DNODE_ENTER(db);
1099 dn = DB_DNODE(db);
1100 /*
1101 * Shouldn't dirty a regular buffer in syncing context. Private
1102 * objects may be dirtied in syncing context, but only if they
1103 * were already pre-dirtied in open context.
1104 */
1105 ASSERT(!dmu_tx_is_syncing(tx) ||
1106 BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1107 DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1108 dn->dn_objset->os_dsl_dataset == NULL);
1109 /*
1110 * We make this assert for private objects as well, but after we
1111 * check if we're already dirty. They are allowed to re-dirty
1112 * in syncing context.
1113 */
1114 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1115 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1116 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1117
1118 mutex_enter(&db->db_mtx);
1119 /*
1120 * XXX make this true for indirects too? The problem is that
1121 * transactions created with dmu_tx_create_assigned() from
1122 * syncing context don't bother holding ahead.
1123 */
1124 ASSERT(db->db_level != 0 ||
1125 db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1126 db->db_state == DB_NOFILL);
1127
1128 mutex_enter(&dn->dn_mtx);
1129 /*
1130 * Don't set dirtyctx to SYNC if we're just modifying this as we
1131 * initialize the objset.
1132 */
1133 if (dn->dn_dirtyctx == DN_UNDIRTIED &&
1134 !BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1135 dn->dn_dirtyctx =
1136 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1137 ASSERT(dn->dn_dirtyctx_firstset == NULL);
1138 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1139 }
1140 mutex_exit(&dn->dn_mtx);
1141
1142 if (db->db_blkid == DMU_SPILL_BLKID)
1143 dn->dn_have_spill = B_TRUE;
1144
1145 /*
1146 * If this buffer is already dirty, we're done.
1147 */
1148 drp = &db->db_last_dirty;
1149 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1150 db->db.db_object == DMU_META_DNODE_OBJECT);
1151 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1152 drp = &dr->dr_next;
1153 if (dr && dr->dr_txg == tx->tx_txg) {
1154 DB_DNODE_EXIT(db);
1155
1156 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1157 /*
1158 * If this buffer has already been written out,
1159 * we now need to reset its state.
1160 */
1161 dbuf_unoverride(dr);
1162 if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1163 db->db_state != DB_NOFILL)
1164 arc_buf_thaw(db->db_buf);
1165 }
1166 mutex_exit(&db->db_mtx);
1167 return (dr);
1168 }
1169
1170 /*
1171 * Only valid if not already dirty.
1172 */
1173 ASSERT(dn->dn_object == 0 ||
1174 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1175 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1176
1177 ASSERT3U(dn->dn_nlevels, >, db->db_level);
1178 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1179 dn->dn_phys->dn_nlevels > db->db_level ||
1180 dn->dn_next_nlevels[txgoff] > db->db_level ||
1181 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1182 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1183
1184 /*
1185 * We should only be dirtying in syncing context if it's the
1186 * mos or we're initializing the os or it's a special object.
1187 * However, we are allowed to dirty in syncing context provided
1188 * we already dirtied it in open context. Hence we must make
1189 * this assertion only if we're not already dirty.
1190 */
1191 os = dn->dn_objset;
1192 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1193 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1194 ASSERT(db->db.db_size != 0);
1195
1196 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1197
1198 if (db->db_blkid != DMU_BONUS_BLKID) {
1199 /*
1200 * Update the accounting.
1201 * Note: we delay "free accounting" until after we drop
1202 * the db_mtx. This keeps us from grabbing other locks
1203 * (and possibly deadlocking) in bp_get_dsize() while
1204 * also holding the db_mtx.
1205 */
1206 dnode_willuse_space(dn, db->db.db_size, tx);
1207 do_free_accounting = dbuf_block_freeable(db);
1208 }
1209
1210 /*
1211 * If this buffer is dirty in an old transaction group we need
1212 * to make a copy of it so that the changes we make in this
1213 * transaction group won't leak out when we sync the older txg.
1214 */
1215 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1216 if (db->db_level == 0) {
1217 void *data_old = db->db_buf;
1218
1219 if (db->db_state != DB_NOFILL) {
1220 if (db->db_blkid == DMU_BONUS_BLKID) {
1221 dbuf_fix_old_data(db, tx->tx_txg);
1222 data_old = db->db.db_data;
1223 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1224 /*
1225 * Release the data buffer from the cache so
1226 * that we can modify it without impacting
1227 * possible other users of this cached data
1228 * block. Note that indirect blocks and
1229 * private objects are not released until the
1230 * syncing state (since they are only modified
1231 * then).
1232 */
1233 arc_release(db->db_buf, db);
1234 dbuf_fix_old_data(db, tx->tx_txg);
1235 data_old = db->db_buf;
1236 }
1237 ASSERT(data_old != NULL);
1238 }
1239 dr->dt.dl.dr_data = data_old;
1240 } else {
1241 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1242 list_create(&dr->dt.di.dr_children,
1243 sizeof (dbuf_dirty_record_t),
1244 offsetof(dbuf_dirty_record_t, dr_dirty_node));
1245 }
1246 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL)
1247 dr->dr_accounted = db->db.db_size;
1248 dr->dr_dbuf = db;
1249 dr->dr_txg = tx->tx_txg;
1250 dr->dr_next = *drp;
1251 *drp = dr;
1252
1253 /*
1254 * We could have been freed_in_flight between the dbuf_noread
1255 * and dbuf_dirty. We win, as though the dbuf_noread() had
1256 * happened after the free.
1257 */
1258 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1259 db->db_blkid != DMU_SPILL_BLKID) {
1260 mutex_enter(&dn->dn_mtx);
1261 if (dn->dn_free_ranges[txgoff] != NULL) {
1262 range_tree_clear(dn->dn_free_ranges[txgoff],
1263 db->db_blkid, 1);
1264 }
1265 mutex_exit(&dn->dn_mtx);
1266 db->db_freed_in_flight = FALSE;
1267 }
1268
1269 /*
1270 * This buffer is now part of this txg
1271 */
1272 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1273 db->db_dirtycnt += 1;
1274 ASSERT3U(db->db_dirtycnt, <=, 3);
1275
1276 mutex_exit(&db->db_mtx);
1277
1278 if (db->db_blkid == DMU_BONUS_BLKID ||
1279 db->db_blkid == DMU_SPILL_BLKID) {
1280 mutex_enter(&dn->dn_mtx);
1281 ASSERT(!list_link_active(&dr->dr_dirty_node));
1282 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1283 mutex_exit(&dn->dn_mtx);
1284 dnode_setdirty(dn, tx);
1285 DB_DNODE_EXIT(db);
1286 return (dr);
1287 } else if (do_free_accounting) {
1288 blkptr_t *bp = db->db_blkptr;
1289 int64_t willfree = (bp && !BP_IS_HOLE(bp)) ?
1290 bp_get_dsize(os->os_spa, bp) : db->db.db_size;
1291 /*
1292 * This is only a guess -- if the dbuf is dirty
1293 * in a previous txg, we don't know how much
1294 * space it will use on disk yet. We should
1295 * really have the struct_rwlock to access
1296 * db_blkptr, but since this is just a guess,
1297 * it's OK if we get an odd answer.
1298 */
1299 ddt_prefetch(os->os_spa, bp);
1300 dnode_willuse_space(dn, -willfree, tx);
1301 }
1302
1303 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1304 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1305 drop_struct_lock = TRUE;
1306 }
1307
1308 if (db->db_level == 0) {
1309 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1310 ASSERT(dn->dn_maxblkid >= db->db_blkid);
1311 }
1312
1313 if (db->db_level+1 < dn->dn_nlevels) {
1314 dmu_buf_impl_t *parent = db->db_parent;
1315 dbuf_dirty_record_t *di;
1316 int parent_held = FALSE;
1317
1318 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1319 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1320
1321 parent = dbuf_hold_level(dn, db->db_level+1,
1322 db->db_blkid >> epbs, FTAG);
1323 ASSERT(parent != NULL);
1324 parent_held = TRUE;
1325 }
1326 if (drop_struct_lock)
1327 rw_exit(&dn->dn_struct_rwlock);
1328 ASSERT3U(db->db_level+1, ==, parent->db_level);
1329 di = dbuf_dirty(parent, tx);
1330 if (parent_held)
1331 dbuf_rele(parent, FTAG);
1332
1333 mutex_enter(&db->db_mtx);
1334 /*
1335 * Since we've dropped the mutex, it's possible that
1336 * dbuf_undirty() might have changed this out from under us.
1337 */
1338 if (db->db_last_dirty == dr ||
1339 dn->dn_object == DMU_META_DNODE_OBJECT) {
1340 mutex_enter(&di->dt.di.dr_mtx);
1341 ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1342 ASSERT(!list_link_active(&dr->dr_dirty_node));
1343 list_insert_tail(&di->dt.di.dr_children, dr);
1344 mutex_exit(&di->dt.di.dr_mtx);
1345 dr->dr_parent = di;
1346 }
1347 mutex_exit(&db->db_mtx);
1348 } else {
1349 ASSERT(db->db_level+1 == dn->dn_nlevels);
1350 ASSERT(db->db_blkid < dn->dn_nblkptr);
1351 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1352 mutex_enter(&dn->dn_mtx);
1353 ASSERT(!list_link_active(&dr->dr_dirty_node));
1354 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1355 mutex_exit(&dn->dn_mtx);
1356 if (drop_struct_lock)
1357 rw_exit(&dn->dn_struct_rwlock);
1358 }
1359
1360 dnode_setdirty(dn, tx);
1361 DB_DNODE_EXIT(db);
1362 return (dr);
1363 }
1364
1365 /*
1366 * Undirty a buffer in the transaction group referenced by the given
1367 * transaction. Return whether this evicted the dbuf.
1368 */
1369 static boolean_t
1370 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1371 {
1372 dnode_t *dn;
1373 uint64_t txg = tx->tx_txg;
1374 dbuf_dirty_record_t *dr, **drp;
1375
1376 ASSERT(txg != 0);
1377 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1378 ASSERT0(db->db_level);
1379 ASSERT(MUTEX_HELD(&db->db_mtx));
1380
1381 /*
1382 * If this buffer is not dirty, we're done.
1383 */
1384 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1385 if (dr->dr_txg <= txg)
1386 break;
1387 if (dr == NULL || dr->dr_txg < txg)
1388 return (B_FALSE);
1389 ASSERT(dr->dr_txg == txg);
1390 ASSERT(dr->dr_dbuf == db);
1391
1392 DB_DNODE_ENTER(db);
1393 dn = DB_DNODE(db);
1394
1395 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1396
1397 ASSERT(db->db.db_size != 0);
1398
1399 /*
1400 * Any space we accounted for in dp_dirty_* will be cleaned up by
1401 * dsl_pool_sync(). This is relatively rare so the discrepancy
1402 * is not a big deal.
1403 */
1404
1405 *drp = dr->dr_next;
1406
1407 /*
1408 * Note that there are three places in dbuf_dirty()
1409 * where this dirty record may be put on a list.
1410 * Make sure to do a list_remove corresponding to
1411 * every one of those list_insert calls.
1412 */
1413 if (dr->dr_parent) {
1414 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1415 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1416 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1417 } else if (db->db_blkid == DMU_SPILL_BLKID ||
1418 db->db_level+1 == dn->dn_nlevels) {
1419 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1420 mutex_enter(&dn->dn_mtx);
1421 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1422 mutex_exit(&dn->dn_mtx);
1423 }
1424 DB_DNODE_EXIT(db);
1425
1426 if (db->db_state != DB_NOFILL) {
1427 dbuf_unoverride(dr);
1428
1429 ASSERT(db->db_buf != NULL);
1430 ASSERT(dr->dt.dl.dr_data != NULL);
1431 if (dr->dt.dl.dr_data != db->db_buf)
1432 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db));
1433 }
1434
1435 if (db->db_level != 0) {
1436 mutex_destroy(&dr->dt.di.dr_mtx);
1437 list_destroy(&dr->dt.di.dr_children);
1438 }
1439
1440 kmem_free(dr, sizeof (dbuf_dirty_record_t));
1441
1442 ASSERT(db->db_dirtycnt > 0);
1443 db->db_dirtycnt -= 1;
1444
1445 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1446 arc_buf_t *buf = db->db_buf;
1447
1448 ASSERT(db->db_state == DB_NOFILL || arc_released(buf));
1449 dbuf_clear_data(db);
1450 VERIFY(arc_buf_remove_ref(buf, db));
1451 dbuf_evict(db);
1452 return (B_TRUE);
1453 }
1454
1455 return (B_FALSE);
1456 }
1457
1458 void
1459 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
1460 {
1461 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1462 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1463
1464 ASSERT(tx->tx_txg != 0);
1465 ASSERT(!refcount_is_zero(&db->db_holds));
1466
1467 DB_DNODE_ENTER(db);
1468 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
1469 rf |= DB_RF_HAVESTRUCT;
1470 DB_DNODE_EXIT(db);
1471 (void) dbuf_read(db, NULL, rf);
1472 (void) dbuf_dirty(db, tx);
1473 }
1474
1475 void
1476 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1477 {
1478 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1479
1480 db->db_state = DB_NOFILL;
1481
1482 dmu_buf_will_fill(db_fake, tx);
1483 }
1484
1485 void
1486 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1487 {
1488 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1489
1490 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1491 ASSERT(tx->tx_txg != 0);
1492 ASSERT(db->db_level == 0);
1493 ASSERT(!refcount_is_zero(&db->db_holds));
1494
1495 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1496 dmu_tx_private_ok(tx));
1497
1498 dbuf_noread(db);
1499 (void) dbuf_dirty(db, tx);
1500 }
1501
1502 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1503 /* ARGSUSED */
1504 void
1505 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1506 {
1507 mutex_enter(&db->db_mtx);
1508 DBUF_VERIFY(db);
1509
1510 if (db->db_state == DB_FILL) {
1511 if (db->db_level == 0 && db->db_freed_in_flight) {
1512 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1513 /* we were freed while filling */
1514 /* XXX dbuf_undirty? */
1515 bzero(db->db.db_data, db->db.db_size);
1516 db->db_freed_in_flight = FALSE;
1517 }
1518 db->db_state = DB_CACHED;
1519 cv_broadcast(&db->db_changed);
1520 }
1521 mutex_exit(&db->db_mtx);
1522 }
1523
1524 void
1525 dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data,
1526 bp_embedded_type_t etype, enum zio_compress comp,
1527 int uncompressed_size, int compressed_size, int byteorder,
1528 dmu_tx_t *tx)
1529 {
1530 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
1531 struct dirty_leaf *dl;
1532 dmu_object_type_t type;
1533
1534 DB_DNODE_ENTER(db);
1535 type = DB_DNODE(db)->dn_type;
1536 DB_DNODE_EXIT(db);
1537
1538 ASSERT0(db->db_level);
1539 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1540
1541 dmu_buf_will_not_fill(dbuf, tx);
1542
1543 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1544 dl = &db->db_last_dirty->dt.dl;
1545 encode_embedded_bp_compressed(&dl->dr_overridden_by,
1546 data, comp, uncompressed_size, compressed_size);
1547 BPE_SET_ETYPE(&dl->dr_overridden_by, etype);
1548 BP_SET_TYPE(&dl->dr_overridden_by, type);
1549 BP_SET_LEVEL(&dl->dr_overridden_by, 0);
1550 BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder);
1551
1552 dl->dr_override_state = DR_OVERRIDDEN;
1553 dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg;
1554 }
1555
1556 /*
1557 * Directly assign a provided arc buf to a given dbuf if it's not referenced
1558 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1559 */
1560 void
1561 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
1562 {
1563 ASSERT(!refcount_is_zero(&db->db_holds));
1564 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1565 ASSERT(db->db_level == 0);
1566 ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA);
1567 ASSERT(buf != NULL);
1568 ASSERT(arc_buf_size(buf) == db->db.db_size);
1569 ASSERT(tx->tx_txg != 0);
1570
1571 arc_return_buf(buf, db);
1572 ASSERT(arc_released(buf));
1573
1574 mutex_enter(&db->db_mtx);
1575
1576 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1577 cv_wait(&db->db_changed, &db->db_mtx);
1578
1579 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
1580
1581 if (db->db_state == DB_CACHED &&
1582 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
1583 mutex_exit(&db->db_mtx);
1584 (void) dbuf_dirty(db, tx);
1585 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
1586 VERIFY(arc_buf_remove_ref(buf, db));
1587 xuio_stat_wbuf_copied();
1588 return;
1589 }
1590
1591 xuio_stat_wbuf_nocopy();
1592 if (db->db_state == DB_CACHED) {
1593 dbuf_dirty_record_t *dr = db->db_last_dirty;
1594
1595 ASSERT(db->db_buf != NULL);
1596 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
1597 ASSERT(dr->dt.dl.dr_data == db->db_buf);
1598 if (!arc_released(db->db_buf)) {
1599 ASSERT(dr->dt.dl.dr_override_state ==
1600 DR_OVERRIDDEN);
1601 arc_release(db->db_buf, db);
1602 }
1603 dr->dt.dl.dr_data = buf;
1604 VERIFY(arc_buf_remove_ref(db->db_buf, db));
1605 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
1606 arc_release(db->db_buf, db);
1607 VERIFY(arc_buf_remove_ref(db->db_buf, db));
1608 }
1609 db->db_buf = NULL;
1610 }
1611 ASSERT(db->db_buf == NULL);
1612 dbuf_set_data(db, buf);
1613 db->db_state = DB_FILL;
1614 mutex_exit(&db->db_mtx);
1615 (void) dbuf_dirty(db, tx);
1616 dmu_buf_fill_done(&db->db, tx);
1617 }
1618
1619 /*
1620 * "Clear" the contents of this dbuf. This will mark the dbuf
1621 * EVICTING and clear *most* of its references. Unfortunately,
1622 * when we are not holding the dn_dbufs_mtx, we can't clear the
1623 * entry in the dn_dbufs list. We have to wait until dbuf_destroy()
1624 * in this case. For callers from the DMU we will usually see:
1625 * dbuf_clear()->arc_clear_callback()->dbuf_do_evict()->dbuf_destroy()
1626 * For the arc callback, we will usually see:
1627 * dbuf_do_evict()->dbuf_clear();dbuf_destroy()
1628 * Sometimes, though, we will get a mix of these two:
1629 * DMU: dbuf_clear()->arc_clear_callback()
1630 * ARC: dbuf_do_evict()->dbuf_destroy()
1631 *
1632 * This routine will dissociate the dbuf from the arc, by calling
1633 * arc_clear_callback(), but will not evict the data from the ARC.
1634 */
1635 void
1636 dbuf_clear(dmu_buf_impl_t *db)
1637 {
1638 dnode_t *dn;
1639 dmu_buf_impl_t *parent = db->db_parent;
1640 dmu_buf_impl_t *dndb;
1641 boolean_t dbuf_gone = B_FALSE;
1642
1643 ASSERT(MUTEX_HELD(&db->db_mtx));
1644 ASSERT(refcount_is_zero(&db->db_holds));
1645
1646 dbuf_evict_user(db);
1647
1648 if (db->db_state == DB_CACHED) {
1649 ASSERT(db->db.db_data != NULL);
1650 if (db->db_blkid == DMU_BONUS_BLKID) {
1651 zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
1652 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
1653 }
1654 db->db.db_data = NULL;
1655 db->db_state = DB_UNCACHED;
1656 }
1657
1658 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
1659 ASSERT(db->db_data_pending == NULL);
1660
1661 db->db_state = DB_EVICTING;
1662 db->db_blkptr = NULL;
1663
1664 DB_DNODE_ENTER(db);
1665 dn = DB_DNODE(db);
1666 dndb = dn->dn_dbuf;
1667 if (db->db_blkid != DMU_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) {
1668 avl_remove(&dn->dn_dbufs, db);
1669 atomic_dec_32(&dn->dn_dbufs_count);
1670 membar_producer();
1671 DB_DNODE_EXIT(db);
1672 /*
1673 * Decrementing the dbuf count means that the hold corresponding
1674 * to the removed dbuf is no longer discounted in dnode_move(),
1675 * so the dnode cannot be moved until after we release the hold.
1676 * The membar_producer() ensures visibility of the decremented
1677 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
1678 * release any lock.
1679 */
1680 dnode_rele(dn, db);
1681 db->db_dnode_handle = NULL;
1682 } else {
1683 DB_DNODE_EXIT(db);
1684 }
1685
1686 if (db->db_buf)
1687 dbuf_gone = arc_clear_callback(db->db_buf);
1688
1689 if (!dbuf_gone)
1690 mutex_exit(&db->db_mtx);
1691
1692 /*
1693 * If this dbuf is referenced from an indirect dbuf,
1694 * decrement the ref count on the indirect dbuf.
1695 */
1696 if (parent && parent != dndb)
1697 dbuf_rele(parent, db);
1698 }
1699
1700 static int
1701 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
1702 dmu_buf_impl_t **parentp, blkptr_t **bpp)
1703 {
1704 int nlevels, epbs;
1705
1706 *parentp = NULL;
1707 *bpp = NULL;
1708
1709 ASSERT(blkid != DMU_BONUS_BLKID);
1710
1711 if (blkid == DMU_SPILL_BLKID) {
1712 mutex_enter(&dn->dn_mtx);
1713 if (dn->dn_have_spill &&
1714 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
1715 *bpp = &dn->dn_phys->dn_spill;
1716 else
1717 *bpp = NULL;
1718 dbuf_add_ref(dn->dn_dbuf, NULL);
1719 *parentp = dn->dn_dbuf;
1720 mutex_exit(&dn->dn_mtx);
1721 return (0);
1722 }
1723
1724 if (dn->dn_phys->dn_nlevels == 0)
1725 nlevels = 1;
1726 else
1727 nlevels = dn->dn_phys->dn_nlevels;
1728
1729 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1730
1731 ASSERT3U(level * epbs, <, 64);
1732 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1733 if (level >= nlevels ||
1734 (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
1735 /* the buffer has no parent yet */
1736 return (SET_ERROR(ENOENT));
1737 } else if (level < nlevels-1) {
1738 /* this block is referenced from an indirect block */
1739 int err = dbuf_hold_impl(dn, level+1,
1740 blkid >> epbs, fail_sparse, NULL, parentp);
1741 if (err)
1742 return (err);
1743 err = dbuf_read(*parentp, NULL,
1744 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
1745 if (err) {
1746 dbuf_rele(*parentp, NULL);
1747 *parentp = NULL;
1748 return (err);
1749 }
1750 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
1751 (blkid & ((1ULL << epbs) - 1));
1752 return (0);
1753 } else {
1754 /* the block is referenced from the dnode */
1755 ASSERT3U(level, ==, nlevels-1);
1756 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
1757 blkid < dn->dn_phys->dn_nblkptr);
1758 if (dn->dn_dbuf) {
1759 dbuf_add_ref(dn->dn_dbuf, NULL);
1760 *parentp = dn->dn_dbuf;
1761 }
1762 *bpp = &dn->dn_phys->dn_blkptr[blkid];
1763 return (0);
1764 }
1765 }
1766
1767 static dmu_buf_impl_t *
1768 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
1769 dmu_buf_impl_t *parent, blkptr_t *blkptr)
1770 {
1771 objset_t *os = dn->dn_objset;
1772 dmu_buf_impl_t *db, *odb;
1773
1774 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1775 ASSERT(dn->dn_type != DMU_OT_NONE);
1776
1777 db = kmem_cache_alloc(dbuf_cache, KM_SLEEP);
1778
1779 db->db_objset = os;
1780 db->db.db_object = dn->dn_object;
1781 db->db_level = level;
1782 db->db_blkid = blkid;
1783 db->db_last_dirty = NULL;
1784 db->db_dirtycnt = 0;
1785 db->db_dnode_handle = dn->dn_handle;
1786 db->db_parent = parent;
1787 db->db_blkptr = blkptr;
1788
1789 db->db_user = NULL;
1790 db->db_immediate_evict = 0;
1791 db->db_freed_in_flight = 0;
1792
1793 if (blkid == DMU_BONUS_BLKID) {
1794 ASSERT3P(parent, ==, dn->dn_dbuf);
1795 db->db.db_size = DN_MAX_BONUSLEN -
1796 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
1797 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
1798 db->db.db_offset = DMU_BONUS_BLKID;
1799 db->db_state = DB_UNCACHED;
1800 /* the bonus dbuf is not placed in the hash table */
1801 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1802 return (db);
1803 } else if (blkid == DMU_SPILL_BLKID) {
1804 db->db.db_size = (blkptr != NULL) ?
1805 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
1806 db->db.db_offset = 0;
1807 } else {
1808 int blocksize =
1809 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
1810 db->db.db_size = blocksize;
1811 db->db.db_offset = db->db_blkid * blocksize;
1812 }
1813
1814 /*
1815 * Hold the dn_dbufs_mtx while we get the new dbuf
1816 * in the hash table *and* added to the dbufs list.
1817 * This prevents a possible deadlock with someone
1818 * trying to look up this dbuf before its added to the
1819 * dn_dbufs list.
1820 */
1821 mutex_enter(&dn->dn_dbufs_mtx);
1822 db->db_state = DB_EVICTING;
1823 if ((odb = dbuf_hash_insert(db)) != NULL) {
1824 /* someone else inserted it first */
1825 kmem_cache_free(dbuf_cache, db);
1826 mutex_exit(&dn->dn_dbufs_mtx);
1827 return (odb);
1828 }
1829 avl_add(&dn->dn_dbufs, db);
1830 if (db->db_level == 0 && db->db_blkid >=
1831 dn->dn_unlisted_l0_blkid)
1832 dn->dn_unlisted_l0_blkid = db->db_blkid + 1;
1833 db->db_state = DB_UNCACHED;
1834 mutex_exit(&dn->dn_dbufs_mtx);
1835 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1836
1837 if (parent && parent != dn->dn_dbuf)
1838 dbuf_add_ref(parent, db);
1839
1840 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1841 refcount_count(&dn->dn_holds) > 0);
1842 (void) refcount_add(&dn->dn_holds, db);
1843 atomic_inc_32(&dn->dn_dbufs_count);
1844
1845 dprintf_dbuf(db, "db=%p\n", db);
1846
1847 return (db);
1848 }
1849
1850 static int
1851 dbuf_do_evict(void *private)
1852 {
1853 dmu_buf_impl_t *db = private;
1854
1855 if (!MUTEX_HELD(&db->db_mtx))
1856 mutex_enter(&db->db_mtx);
1857
1858 ASSERT(refcount_is_zero(&db->db_holds));
1859
1860 if (db->db_state != DB_EVICTING) {
1861 ASSERT(db->db_state == DB_CACHED);
1862 DBUF_VERIFY(db);
1863 db->db_buf = NULL;
1864 dbuf_evict(db);
1865 } else {
1866 mutex_exit(&db->db_mtx);
1867 dbuf_destroy(db);
1868 }
1869 return (0);
1870 }
1871
1872 static void
1873 dbuf_destroy(dmu_buf_impl_t *db)
1874 {
1875 ASSERT(refcount_is_zero(&db->db_holds));
1876
1877 if (db->db_blkid != DMU_BONUS_BLKID) {
1878 /*
1879 * If this dbuf is still on the dn_dbufs list,
1880 * remove it from that list.
1881 */
1882 if (db->db_dnode_handle != NULL) {
1883 dnode_t *dn;
1884
1885 DB_DNODE_ENTER(db);
1886 dn = DB_DNODE(db);
1887 mutex_enter(&dn->dn_dbufs_mtx);
1888 avl_remove(&dn->dn_dbufs, db);
1889 atomic_dec_32(&dn->dn_dbufs_count);
1890 mutex_exit(&dn->dn_dbufs_mtx);
1891 DB_DNODE_EXIT(db);
1892 /*
1893 * Decrementing the dbuf count means that the hold
1894 * corresponding to the removed dbuf is no longer
1895 * discounted in dnode_move(), so the dnode cannot be
1896 * moved until after we release the hold.
1897 */
1898 dnode_rele(dn, db);
1899 db->db_dnode_handle = NULL;
1900 }
1901 dbuf_hash_remove(db);
1902 }
1903 db->db_parent = NULL;
1904 db->db_buf = NULL;
1905
1906 ASSERT(db->db.db_data == NULL);
1907 ASSERT(db->db_hash_next == NULL);
1908 ASSERT(db->db_blkptr == NULL);
1909 ASSERT(db->db_data_pending == NULL);
1910
1911 kmem_cache_free(dbuf_cache, db);
1912 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1913 }
1914
1915 void
1916 dbuf_prefetch(dnode_t *dn, uint64_t blkid, zio_priority_t prio)
1917 {
1918 dmu_buf_impl_t *db = NULL;
1919 blkptr_t *bp = NULL;
1920
1921 ASSERT(blkid != DMU_BONUS_BLKID);
1922 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1923
1924 if (dnode_block_freed(dn, blkid))
1925 return;
1926
1927 /* dbuf_find() returns with db_mtx held */
1928 if (db = dbuf_find(dn, 0, blkid)) {
1929 /*
1930 * This dbuf is already in the cache. We assume that
1931 * it is already CACHED, or else about to be either
1932 * read or filled.
1933 */
1934 mutex_exit(&db->db_mtx);
1935 return;
1936 }
1937
1938 if (dbuf_findbp(dn, 0, blkid, TRUE, &db, &bp) == 0) {
1939 if (bp && !BP_IS_HOLE(bp) && !BP_IS_EMBEDDED(bp)) {
1940 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
1941 arc_flags_t aflags =
1942 ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
1943 zbookmark_phys_t zb;
1944
1945 SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET,
1946 dn->dn_object, 0, blkid);
1947
1948 (void) arc_read(NULL, dn->dn_objset->os_spa,
1949 bp, NULL, NULL, prio,
1950 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
1951 &aflags, &zb);
1952 }
1953 if (db)
1954 dbuf_rele(db, NULL);
1955 }
1956 }
1957
1958 /*
1959 * Returns with db_holds incremented, and db_mtx not held.
1960 * Note: dn_struct_rwlock must be held.
1961 */
1962 int
1963 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse,
1964 void *tag, dmu_buf_impl_t **dbp)
1965 {
1966 dmu_buf_impl_t *db, *parent = NULL;
1967
1968 ASSERT(blkid != DMU_BONUS_BLKID);
1969 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1970 ASSERT3U(dn->dn_nlevels, >, level);
1971
1972 *dbp = NULL;
1973 top:
1974 /* dbuf_find() returns with db_mtx held */
1975 db = dbuf_find(dn, level, blkid);
1976
1977 if (db == NULL) {
1978 blkptr_t *bp = NULL;
1979 int err;
1980
1981 ASSERT3P(parent, ==, NULL);
1982 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
1983 if (fail_sparse) {
1984 if (err == 0 && bp && BP_IS_HOLE(bp))
1985 err = SET_ERROR(ENOENT);
1986 if (err) {
1987 if (parent)
1988 dbuf_rele(parent, NULL);
1989 return (err);
1990 }
1991 }
1992 if (err && err != ENOENT)
1993 return (err);
1994 db = dbuf_create(dn, level, blkid, parent, bp);
1995 }
1996
1997 if (db->db_buf && refcount_is_zero(&db->db_holds)) {
1998 arc_buf_add_ref(db->db_buf, db);
1999 if (db->db_buf->b_data == NULL) {
2000 dbuf_clear(db);
2001 if (parent) {
2002 dbuf_rele(parent, NULL);
2003 parent = NULL;
2004 }
2005 goto top;
2006 }
2007 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
2008 }
2009
2010 ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
2011
2012 /*
2013 * If this buffer is currently syncing out, and we are are
2014 * still referencing it from db_data, we need to make a copy
2015 * of it in case we decide we want to dirty it again in this txg.
2016 */
2017 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
2018 dn->dn_object != DMU_META_DNODE_OBJECT &&
2019 db->db_state == DB_CACHED && db->db_data_pending) {
2020 dbuf_dirty_record_t *dr = db->db_data_pending;
2021
2022 if (dr->dt.dl.dr_data == db->db_buf) {
2023 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2024
2025 dbuf_set_data(db,
2026 arc_buf_alloc(dn->dn_objset->os_spa,
2027 db->db.db_size, db, type));
2028 bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
2029 db->db.db_size);
2030 }
2031 }
2032
2033 (void) refcount_add(&db->db_holds, tag);
2034 DBUF_VERIFY(db);
2035 mutex_exit(&db->db_mtx);
2036
2037 /* NOTE: we can't rele the parent until after we drop the db_mtx */
2038 if (parent)
2039 dbuf_rele(parent, NULL);
2040
2041 ASSERT3P(DB_DNODE(db), ==, dn);
2042 ASSERT3U(db->db_blkid, ==, blkid);
2043 ASSERT3U(db->db_level, ==, level);
2044 *dbp = db;
2045
2046 return (0);
2047 }
2048
2049 dmu_buf_impl_t *
2050 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
2051 {
2052 dmu_buf_impl_t *db;
2053 int err = dbuf_hold_impl(dn, 0, blkid, FALSE, tag, &db);
2054 return (err ? NULL : db);
2055 }
2056
2057 dmu_buf_impl_t *
2058 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
2059 {
2060 dmu_buf_impl_t *db;
2061 int err = dbuf_hold_impl(dn, level, blkid, FALSE, tag, &db);
2062 return (err ? NULL : db);
2063 }
2064
2065 void
2066 dbuf_create_bonus(dnode_t *dn)
2067 {
2068 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
2069
2070 ASSERT(dn->dn_bonus == NULL);
2071 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
2072 }
2073
2074 int
2075 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
2076 {
2077 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2078 dnode_t *dn;
2079
2080 if (db->db_blkid != DMU_SPILL_BLKID)
2081 return (SET_ERROR(ENOTSUP));
2082 if (blksz == 0)
2083 blksz = SPA_MINBLOCKSIZE;
2084 ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset)));
2085 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
2086
2087 DB_DNODE_ENTER(db);
2088 dn = DB_DNODE(db);
2089 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2090 dbuf_new_size(db, blksz, tx);
2091 rw_exit(&dn->dn_struct_rwlock);
2092 DB_DNODE_EXIT(db);
2093
2094 return (0);
2095 }
2096
2097 void
2098 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
2099 {
2100 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
2101 }
2102
2103 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2104 void
2105 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
2106 {
2107 int64_t holds = refcount_add(&db->db_holds, tag);
2108 ASSERT(holds > 1);
2109 }
2110
2111 /*
2112 * If you call dbuf_rele() you had better not be referencing the dnode handle
2113 * unless you have some other direct or indirect hold on the dnode. (An indirect
2114 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2115 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2116 * dnode's parent dbuf evicting its dnode handles.
2117 */
2118 void
2119 dbuf_rele(dmu_buf_impl_t *db, void *tag)
2120 {
2121 mutex_enter(&db->db_mtx);
2122 dbuf_rele_and_unlock(db, tag);
2123 }
2124
2125 void
2126 dmu_buf_rele(dmu_buf_t *db, void *tag)
2127 {
2128 dbuf_rele((dmu_buf_impl_t *)db, tag);
2129 }
2130
2131 /*
2132 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2133 * db_dirtycnt and db_holds to be updated atomically.
2134 */
2135 void
2136 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2137 {
2138 int64_t holds;
2139
2140 ASSERT(MUTEX_HELD(&db->db_mtx));
2141 DBUF_VERIFY(db);
2142
2143 /*
2144 * Remove the reference to the dbuf before removing its hold on the
2145 * dnode so we can guarantee in dnode_move() that a referenced bonus
2146 * buffer has a corresponding dnode hold.
2147 */
2148 holds = refcount_remove(&db->db_holds, tag);
2149 ASSERT(holds >= 0);
2150
2151 /*
2152 * We can't freeze indirects if there is a possibility that they
2153 * may be modified in the current syncing context.
2154 */
2155 if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0))
2156 arc_buf_freeze(db->db_buf);
2157
2158 if (holds == db->db_dirtycnt &&
2159 db->db_level == 0 && db->db_immediate_evict)
2160 dbuf_evict_user(db);
2161
2162 if (holds == 0) {
2163 if (db->db_blkid == DMU_BONUS_BLKID) {
2164 mutex_exit(&db->db_mtx);
2165
2166 /*
2167 * If the dnode moves here, we cannot cross this barrier
2168 * until the move completes.
2169 */
2170 DB_DNODE_ENTER(db);
2171 atomic_dec_32(&DB_DNODE(db)->dn_dbufs_count);
2172 DB_DNODE_EXIT(db);
2173 /*
2174 * The bonus buffer's dnode hold is no longer discounted
2175 * in dnode_move(). The dnode cannot move until after
2176 * the dnode_rele().
2177 */
2178 dnode_rele(DB_DNODE(db), db);
2179 } else if (db->db_buf == NULL) {
2180 /*
2181 * This is a special case: we never associated this
2182 * dbuf with any data allocated from the ARC.
2183 */
2184 ASSERT(db->db_state == DB_UNCACHED ||
2185 db->db_state == DB_NOFILL);
2186 dbuf_evict(db);
2187 } else if (arc_released(db->db_buf)) {
2188 arc_buf_t *buf = db->db_buf;
2189 /*
2190 * This dbuf has anonymous data associated with it.
2191 */
2192 dbuf_clear_data(db);
2193 VERIFY(arc_buf_remove_ref(buf, db));
2194 dbuf_evict(db);
2195 } else {
2196 VERIFY(!arc_buf_remove_ref(db->db_buf, db));
2197
2198 /*
2199 * A dbuf will be eligible for eviction if either the
2200 * 'primarycache' property is set or a duplicate
2201 * copy of this buffer is already cached in the arc.
2202 *
2203 * In the case of the 'primarycache' a buffer
2204 * is considered for eviction if it matches the
2205 * criteria set in the property.
2206 *
2207 * To decide if our buffer is considered a
2208 * duplicate, we must call into the arc to determine
2209 * if multiple buffers are referencing the same
2210 * block on-disk. If so, then we simply evict
2211 * ourselves.
2212 */
2213 if (!DBUF_IS_CACHEABLE(db)) {
2214 if (db->db_blkptr != NULL &&
2215 !BP_IS_HOLE(db->db_blkptr) &&
2216 !BP_IS_EMBEDDED(db->db_blkptr)) {
2217 spa_t *spa =
2218 dmu_objset_spa(db->db_objset);
2219 blkptr_t bp = *db->db_blkptr;
2220 dbuf_clear(db);
2221 arc_freed(spa, &bp);
2222 } else {
2223 dbuf_clear(db);
2224 }
2225 } else if (db->db_objset->os_evicting ||
2226 arc_buf_eviction_needed(db->db_buf)) {
2227 dbuf_clear(db);
2228 } else {
2229 mutex_exit(&db->db_mtx);
2230 }
2231 }
2232 } else {
2233 mutex_exit(&db->db_mtx);
2234 }
2235 }
2236
2237 #pragma weak dmu_buf_refcount = dbuf_refcount
2238 uint64_t
2239 dbuf_refcount(dmu_buf_impl_t *db)
2240 {
2241 return (refcount_count(&db->db_holds));
2242 }
2243
2244 void *
2245 dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user,
2246 dmu_buf_user_t *new_user)
2247 {
2248 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2249
2250 mutex_enter(&db->db_mtx);
2251 dbuf_verify_user(db, DBVU_NOT_EVICTING);
2252 if (db->db_user == old_user)
2253 db->db_user = new_user;
2254 else
2255 old_user = db->db_user;
2256 dbuf_verify_user(db, DBVU_NOT_EVICTING);
2257 mutex_exit(&db->db_mtx);
2258
2259 return (old_user);
2260 }
2261
2262 void *
2263 dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2264 {
2265 return (dmu_buf_replace_user(db_fake, NULL, user));
2266 }
2267
2268 void *
2269 dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2270 {
2271 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2272
2273 db->db_immediate_evict = TRUE;
2274 return (dmu_buf_set_user(db_fake, user));
2275 }
2276
2277 void *
2278 dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2279 {
2280 return (dmu_buf_replace_user(db_fake, user, NULL));
2281 }
2282
2283 void *
2284 dmu_buf_get_user(dmu_buf_t *db_fake)
2285 {
2286 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2287
2288 dbuf_verify_user(db, DBVU_NOT_EVICTING);
2289 return (db->db_user);
2290 }
2291
2292 void
2293 dmu_buf_user_evict_wait()
2294 {
2295 taskq_wait(dbu_evict_taskq);
2296 }
2297
2298 boolean_t
2299 dmu_buf_freeable(dmu_buf_t *dbuf)
2300 {
2301 boolean_t res = B_FALSE;
2302 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2303
2304 if (db->db_blkptr)
2305 res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset,
2306 db->db_blkptr, db->db_blkptr->blk_birth);
2307
2308 return (res);
2309 }
2310
2311 blkptr_t *
2312 dmu_buf_get_blkptr(dmu_buf_t *db)
2313 {
2314 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2315 return (dbi->db_blkptr);
2316 }
2317
2318 static void
2319 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
2320 {
2321 /* ASSERT(dmu_tx_is_syncing(tx) */
2322 ASSERT(MUTEX_HELD(&db->db_mtx));
2323
2324 if (db->db_blkptr != NULL)
2325 return;
2326
2327 if (db->db_blkid == DMU_SPILL_BLKID) {
2328 db->db_blkptr = &dn->dn_phys->dn_spill;
2329 BP_ZERO(db->db_blkptr);
2330 return;
2331 }
2332 if (db->db_level == dn->dn_phys->dn_nlevels-1) {
2333 /*
2334 * This buffer was allocated at a time when there was
2335 * no available blkptrs from the dnode, or it was
2336 * inappropriate to hook it in (i.e., nlevels mis-match).
2337 */
2338 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
2339 ASSERT(db->db_parent == NULL);
2340 db->db_parent = dn->dn_dbuf;
2341 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
2342 DBUF_VERIFY(db);
2343 } else {
2344 dmu_buf_impl_t *parent = db->db_parent;
2345 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2346
2347 ASSERT(dn->dn_phys->dn_nlevels > 1);
2348 if (parent == NULL) {
2349 mutex_exit(&db->db_mtx);
2350 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2351 (void) dbuf_hold_impl(dn, db->db_level+1,
2352 db->db_blkid >> epbs, FALSE, db, &parent);
2353 rw_exit(&dn->dn_struct_rwlock);
2354 mutex_enter(&db->db_mtx);
2355 db->db_parent = parent;
2356 }
2357 db->db_blkptr = (blkptr_t *)parent->db.db_data +
2358 (db->db_blkid & ((1ULL << epbs) - 1));
2359 DBUF_VERIFY(db);
2360 }
2361 }
2362
2363 static void
2364 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2365 {
2366 dmu_buf_impl_t *db = dr->dr_dbuf;
2367 dnode_t *dn;
2368 zio_t *zio;
2369
2370 ASSERT(dmu_tx_is_syncing(tx));
2371
2372 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2373
2374 mutex_enter(&db->db_mtx);
2375
2376 ASSERT(db->db_level > 0);
2377 DBUF_VERIFY(db);
2378
2379 /* Read the block if it hasn't been read yet. */
2380 if (db->db_buf == NULL) {
2381 mutex_exit(&db->db_mtx);
2382 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
2383 mutex_enter(&db->db_mtx);
2384 }
2385 ASSERT3U(db->db_state, ==, DB_CACHED);
2386 ASSERT(db->db_buf != NULL);
2387
2388 DB_DNODE_ENTER(db);
2389 dn = DB_DNODE(db);
2390 /* Indirect block size must match what the dnode thinks it is. */
2391 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2392 dbuf_check_blkptr(dn, db);
2393 DB_DNODE_EXIT(db);
2394
2395 /* Provide the pending dirty record to child dbufs */
2396 db->db_data_pending = dr;
2397
2398 mutex_exit(&db->db_mtx);
2399 dbuf_write(dr, db->db_buf, tx);
2400
2401 zio = dr->dr_zio;
2402 mutex_enter(&dr->dt.di.dr_mtx);
2403 dbuf_sync_list(&dr->dt.di.dr_children, tx);
2404 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2405 mutex_exit(&dr->dt.di.dr_mtx);
2406 zio_nowait(zio);
2407 }
2408
2409 static void
2410 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2411 {
2412 arc_buf_t **datap = &dr->dt.dl.dr_data;
2413 dmu_buf_impl_t *db = dr->dr_dbuf;
2414 dnode_t *dn;
2415 objset_t *os;
2416 uint64_t txg = tx->tx_txg;
2417
2418 ASSERT(dmu_tx_is_syncing(tx));
2419
2420 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2421
2422 mutex_enter(&db->db_mtx);
2423 /*
2424 * To be synced, we must be dirtied. But we
2425 * might have been freed after the dirty.
2426 */
2427 if (db->db_state == DB_UNCACHED) {
2428 /* This buffer has been freed since it was dirtied */
2429 ASSERT(db->db.db_data == NULL);
2430 } else if (db->db_state == DB_FILL) {
2431 /* This buffer was freed and is now being re-filled */
2432 ASSERT(db->db.db_data != dr->dt.dl.dr_data);
2433 } else {
2434 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
2435 }
2436 DBUF_VERIFY(db);
2437
2438 DB_DNODE_ENTER(db);
2439 dn = DB_DNODE(db);
2440
2441 if (db->db_blkid == DMU_SPILL_BLKID) {
2442 mutex_enter(&dn->dn_mtx);
2443 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
2444 mutex_exit(&dn->dn_mtx);
2445 }
2446
2447 /*
2448 * If this is a bonus buffer, simply copy the bonus data into the
2449 * dnode. It will be written out when the dnode is synced (and it
2450 * will be synced, since it must have been dirty for dbuf_sync to
2451 * be called).
2452 */
2453 if (db->db_blkid == DMU_BONUS_BLKID) {
2454 dbuf_dirty_record_t **drp;
2455
2456 ASSERT(*datap != NULL);
2457 ASSERT0(db->db_level);
2458 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN);
2459 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen);
2460 DB_DNODE_EXIT(db);
2461
2462 if (*datap != db->db.db_data) {
2463 zio_buf_free(*datap, DN_MAX_BONUSLEN);
2464 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
2465 }
2466 db->db_data_pending = NULL;
2467 drp = &db->db_last_dirty;
2468 while (*drp != dr)
2469 drp = &(*drp)->dr_next;
2470 ASSERT(dr->dr_next == NULL);
2471 ASSERT(dr->dr_dbuf == db);
2472 *drp = dr->dr_next;
2473 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2474 ASSERT(db->db_dirtycnt > 0);
2475 db->db_dirtycnt -= 1;
2476 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
2477 return;
2478 }
2479
2480 os = dn->dn_objset;
2481
2482 /*
2483 * This function may have dropped the db_mtx lock allowing a dmu_sync
2484 * operation to sneak in. As a result, we need to ensure that we
2485 * don't check the dr_override_state until we have returned from
2486 * dbuf_check_blkptr.
2487 */
2488 dbuf_check_blkptr(dn, db);
2489
2490 /*
2491 * If this buffer is in the middle of an immediate write,
2492 * wait for the synchronous IO to complete.
2493 */
2494 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
2495 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
2496 cv_wait(&db->db_changed, &db->db_mtx);
2497 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
2498 }
2499
2500 if (db->db_state != DB_NOFILL &&
2501 dn->dn_object != DMU_META_DNODE_OBJECT &&
2502 refcount_count(&db->db_holds) > 1 &&
2503 dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
2504 *datap == db->db_buf) {
2505 /*
2506 * If this buffer is currently "in use" (i.e., there
2507 * are active holds and db_data still references it),
2508 * then make a copy before we start the write so that
2509 * any modifications from the open txg will not leak
2510 * into this write.
2511 *
2512 * NOTE: this copy does not need to be made for
2513 * objects only modified in the syncing context (e.g.
2514 * DNONE_DNODE blocks).
2515 */
2516 int blksz = arc_buf_size(*datap);
2517 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2518 *datap = arc_buf_alloc(os->os_spa, blksz, db, type);
2519 bcopy(db->db.db_data, (*datap)->b_data, blksz);
2520 }
2521 db->db_data_pending = dr;
2522
2523 mutex_exit(&db->db_mtx);
2524
2525 dbuf_write(dr, *datap, tx);
2526
2527 ASSERT(!list_link_active(&dr->dr_dirty_node));
2528 if (dn->dn_object == DMU_META_DNODE_OBJECT) {
2529 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
2530 DB_DNODE_EXIT(db);
2531 } else {
2532 /*
2533 * Although zio_nowait() does not "wait for an IO", it does
2534 * initiate the IO. If this is an empty write it seems plausible
2535 * that the IO could actually be completed before the nowait
2536 * returns. We need to DB_DNODE_EXIT() first in case
2537 * zio_nowait() invalidates the dbuf.
2538 */
2539 DB_DNODE_EXIT(db);
2540 zio_nowait(dr->dr_zio);
2541 }
2542 }
2543
2544 void
2545 dbuf_sync_list(list_t *list, dmu_tx_t *tx)
2546 {
2547 dbuf_dirty_record_t *dr;
2548
2549 while (dr = list_head(list)) {
2550 if (dr->dr_zio != NULL) {
2551 /*
2552 * If we find an already initialized zio then we
2553 * are processing the meta-dnode, and we have finished.
2554 * The dbufs for all dnodes are put back on the list
2555 * during processing, so that we can zio_wait()
2556 * these IOs after initiating all child IOs.
2557 */
2558 ASSERT3U(dr->dr_dbuf->db.db_object, ==,
2559 DMU_META_DNODE_OBJECT);
2560 break;
2561 }
2562 list_remove(list, dr);
2563 if (dr->dr_dbuf->db_level > 0)
2564 dbuf_sync_indirect(dr, tx);
2565 else
2566 dbuf_sync_leaf(dr, tx);
2567 }
2568 }
2569
2570 /* ARGSUSED */
2571 static void
2572 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
2573 {
2574 dmu_buf_impl_t *db = vdb;
2575 dnode_t *dn;
2576 blkptr_t *bp = zio->io_bp;
2577 blkptr_t *bp_orig = &zio->io_bp_orig;
2578 spa_t *spa = zio->io_spa;
2579 int64_t delta;
2580 uint64_t fill = 0;
2581 int i;
2582
2583 ASSERT3P(db->db_blkptr, ==, bp);
2584
2585 DB_DNODE_ENTER(db);
2586 dn = DB_DNODE(db);
2587 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
2588 dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
2589 zio->io_prev_space_delta = delta;
2590
2591 if (bp->blk_birth != 0) {
2592 ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
2593 BP_GET_TYPE(bp) == dn->dn_type) ||
2594 (db->db_blkid == DMU_SPILL_BLKID &&
2595 BP_GET_TYPE(bp) == dn->dn_bonustype) ||
2596 BP_IS_EMBEDDED(bp));
2597 ASSERT(BP_GET_LEVEL(bp) == db->db_level);
2598 }
2599
2600 mutex_enter(&db->db_mtx);
2601
2602 #ifdef ZFS_DEBUG
2603 if (db->db_blkid == DMU_SPILL_BLKID) {
2604 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2605 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2606 db->db_blkptr == &dn->dn_phys->dn_spill);
2607 }
2608 #endif
2609
2610 if (db->db_level == 0) {
2611 mutex_enter(&dn->dn_mtx);
2612 if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
2613 db->db_blkid != DMU_SPILL_BLKID)
2614 dn->dn_phys->dn_maxblkid = db->db_blkid;
2615 mutex_exit(&dn->dn_mtx);
2616
2617 if (dn->dn_type == DMU_OT_DNODE) {
2618 dnode_phys_t *dnp = db->db.db_data;
2619 for (i = db->db.db_size >> DNODE_SHIFT; i > 0;
2620 i--, dnp++) {
2621 if (dnp->dn_type != DMU_OT_NONE)
2622 fill++;
2623 }
2624 } else {
2625 if (BP_IS_HOLE(bp)) {
2626 fill = 0;
2627 } else {
2628 fill = 1;
2629 }
2630 }
2631 } else {
2632 blkptr_t *ibp = db->db.db_data;
2633 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2634 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
2635 if (BP_IS_HOLE(ibp))
2636 continue;
2637 fill += BP_GET_FILL(ibp);
2638 }
2639 }
2640 DB_DNODE_EXIT(db);
2641
2642 if (!BP_IS_EMBEDDED(bp))
2643 bp->blk_fill = fill;
2644
2645 mutex_exit(&db->db_mtx);
2646 }
2647
2648 /*
2649 * The SPA will call this callback several times for each zio - once
2650 * for every physical child i/o (zio->io_phys_children times). This
2651 * allows the DMU to monitor the progress of each logical i/o. For example,
2652 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
2653 * block. There may be a long delay before all copies/fragments are completed,
2654 * so this callback allows us to retire dirty space gradually, as the physical
2655 * i/os complete.
2656 */
2657 /* ARGSUSED */
2658 static void
2659 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
2660 {
2661 dmu_buf_impl_t *db = arg;
2662 objset_t *os = db->db_objset;
2663 dsl_pool_t *dp = dmu_objset_pool(os);
2664 dbuf_dirty_record_t *dr;
2665 int delta = 0;
2666
2667 dr = db->db_data_pending;
2668 ASSERT3U(dr->dr_txg, ==, zio->io_txg);
2669
2670 /*
2671 * The callback will be called io_phys_children times. Retire one
2672 * portion of our dirty space each time we are called. Any rounding
2673 * error will be cleaned up by dsl_pool_sync()'s call to
2674 * dsl_pool_undirty_space().
2675 */
2676 delta = dr->dr_accounted / zio->io_phys_children;
2677 dsl_pool_undirty_space(dp, delta, zio->io_txg);
2678 }
2679
2680 /* ARGSUSED */
2681 static void
2682 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
2683 {
2684 dmu_buf_impl_t *db = vdb;
2685 blkptr_t *bp_orig = &zio->io_bp_orig;
2686 blkptr_t *bp = db->db_blkptr;
2687 objset_t *os = db->db_objset;
2688 dmu_tx_t *tx = os->os_synctx;
2689 dbuf_dirty_record_t **drp, *dr;
2690
2691 ASSERT0(zio->io_error);
2692 ASSERT(db->db_blkptr == bp);
2693
2694 /*
2695 * For nopwrites and rewrites we ensure that the bp matches our
2696 * original and bypass all the accounting.
2697 */
2698 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
2699 ASSERT(BP_EQUAL(bp, bp_orig));
2700 } else {
2701 dsl_dataset_t *ds = os->os_dsl_dataset;
2702 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
2703 dsl_dataset_block_born(ds, bp, tx);
2704 }
2705
2706 mutex_enter(&db->db_mtx);
2707
2708 DBUF_VERIFY(db);
2709
2710 drp = &db->db_last_dirty;
2711 while ((dr = *drp) != db->db_data_pending)
2712 drp = &dr->dr_next;
2713 ASSERT(!list_link_active(&dr->dr_dirty_node));
2714 ASSERT(dr->dr_dbuf == db);
2715 ASSERT(dr->dr_next == NULL);
2716 *drp = dr->dr_next;
2717
2718 #ifdef ZFS_DEBUG
2719 if (db->db_blkid == DMU_SPILL_BLKID) {
2720 dnode_t *dn;
2721
2722 DB_DNODE_ENTER(db);
2723 dn = DB_DNODE(db);
2724 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2725 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2726 db->db_blkptr == &dn->dn_phys->dn_spill);
2727 DB_DNODE_EXIT(db);
2728 }
2729 #endif
2730
2731 if (db->db_level == 0) {
2732 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2733 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
2734 if (db->db_state != DB_NOFILL) {
2735 if (dr->dt.dl.dr_data != db->db_buf)
2736 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data,
2737 db));
2738 else if (!arc_released(db->db_buf))
2739 arc_set_callback(db->db_buf, dbuf_do_evict, db);
2740 }
2741 } else {
2742 dnode_t *dn;
2743
2744 DB_DNODE_ENTER(db);
2745 dn = DB_DNODE(db);
2746 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2747 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
2748 if (!BP_IS_HOLE(db->db_blkptr)) {
2749 int epbs =
2750 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2751 ASSERT3U(db->db_blkid, <=,
2752 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs));
2753 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
2754 db->db.db_size);
2755 if (!arc_released(db->db_buf))
2756 arc_set_callback(db->db_buf, dbuf_do_evict, db);
2757 }
2758 DB_DNODE_EXIT(db);
2759 mutex_destroy(&dr->dt.di.dr_mtx);
2760 list_destroy(&dr->dt.di.dr_children);
2761 }
2762 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2763
2764 cv_broadcast(&db->db_changed);
2765 ASSERT(db->db_dirtycnt > 0);
2766 db->db_dirtycnt -= 1;
2767 db->db_data_pending = NULL;
2768 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg);
2769 }
2770
2771 static void
2772 dbuf_write_nofill_ready(zio_t *zio)
2773 {
2774 dbuf_write_ready(zio, NULL, zio->io_private);
2775 }
2776
2777 static void
2778 dbuf_write_nofill_done(zio_t *zio)
2779 {
2780 dbuf_write_done(zio, NULL, zio->io_private);
2781 }
2782
2783 static void
2784 dbuf_write_override_ready(zio_t *zio)
2785 {
2786 dbuf_dirty_record_t *dr = zio->io_private;
2787 dmu_buf_impl_t *db = dr->dr_dbuf;
2788
2789 dbuf_write_ready(zio, NULL, db);
2790 }
2791
2792 static void
2793 dbuf_write_override_done(zio_t *zio)
2794 {
2795 dbuf_dirty_record_t *dr = zio->io_private;
2796 dmu_buf_impl_t *db = dr->dr_dbuf;
2797 blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
2798
2799 mutex_enter(&db->db_mtx);
2800 if (!BP_EQUAL(zio->io_bp, obp)) {
2801 if (!BP_IS_HOLE(obp))
2802 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
2803 arc_release(dr->dt.dl.dr_data, db);
2804 }
2805 mutex_exit(&db->db_mtx);
2806
2807 dbuf_write_done(zio, NULL, db);
2808 }
2809
2810 /* Issue I/O to commit a dirty buffer to disk. */
2811 static void
2812 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
2813 {
2814 dmu_buf_impl_t *db = dr->dr_dbuf;
2815 dnode_t *dn;
2816 objset_t *os;
2817 dmu_buf_impl_t *parent = db->db_parent;
2818 uint64_t txg = tx->tx_txg;
2819 zbookmark_phys_t zb;
2820 zio_prop_t zp;
2821 zio_t *zio;
2822 int wp_flag = 0;
2823
2824 DB_DNODE_ENTER(db);
2825 dn = DB_DNODE(db);
2826 os = dn->dn_objset;
2827
2828 if (db->db_state != DB_NOFILL) {
2829 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
2830 /*
2831 * Private object buffers are released here rather
2832 * than in dbuf_dirty() since they are only modified
2833 * in the syncing context and we don't want the
2834 * overhead of making multiple copies of the data.
2835 */
2836 if (BP_IS_HOLE(db->db_blkptr)) {
2837 arc_buf_thaw(data);
2838 } else {
2839 dbuf_release_bp(db);
2840 }
2841 }
2842 }
2843
2844 if (parent != dn->dn_dbuf) {
2845 /* Our parent is an indirect block. */
2846 /* We have a dirty parent that has been scheduled for write. */
2847 ASSERT(parent && parent->db_data_pending);
2848 /* Our parent's buffer is one level closer to the dnode. */
2849 ASSERT(db->db_level == parent->db_level-1);
2850 /*
2851 * We're about to modify our parent's db_data by modifying
2852 * our block pointer, so the parent must be released.
2853 */
2854 ASSERT(arc_released(parent->db_buf));
2855 zio = parent->db_data_pending->dr_zio;
2856 } else {
2857 /* Our parent is the dnode itself. */
2858 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
2859 db->db_blkid != DMU_SPILL_BLKID) ||
2860 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
2861 if (db->db_blkid != DMU_SPILL_BLKID)
2862 ASSERT3P(db->db_blkptr, ==,
2863 &dn->dn_phys->dn_blkptr[db->db_blkid]);
2864 zio = dn->dn_zio;
2865 }
2866
2867 ASSERT(db->db_level == 0 || data == db->db_buf);
2868 ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
2869 ASSERT(zio);
2870
2871 SET_BOOKMARK(&zb, os->os_dsl_dataset ?
2872 os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
2873 db->db.db_object, db->db_level, db->db_blkid);
2874
2875 if (db->db_blkid == DMU_SPILL_BLKID)
2876 wp_flag = WP_SPILL;
2877 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
2878
2879 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
2880 DB_DNODE_EXIT(db);
2881
2882 if (db->db_level == 0 &&
2883 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
2884 /*
2885 * The BP for this block has been provided by open context
2886 * (by dmu_sync() or dmu_buf_write_embedded()).
2887 */
2888 void *contents = (data != NULL) ? data->b_data : NULL;
2889
2890 dr->dr_zio = zio_write(zio, os->os_spa, txg,
2891 db->db_blkptr, contents, db->db.db_size, &zp,
2892 dbuf_write_override_ready, NULL, dbuf_write_override_done,
2893 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
2894 mutex_enter(&db->db_mtx);
2895 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
2896 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
2897 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
2898 mutex_exit(&db->db_mtx);
2899 } else if (db->db_state == DB_NOFILL) {
2900 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF ||
2901 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY);
2902 dr->dr_zio = zio_write(zio, os->os_spa, txg,
2903 db->db_blkptr, NULL, db->db.db_size, &zp,
2904 dbuf_write_nofill_ready, NULL, dbuf_write_nofill_done, db,
2905 ZIO_PRIORITY_ASYNC_WRITE,
2906 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
2907 } else {
2908 ASSERT(arc_released(data));
2909 dr->dr_zio = arc_write(zio, os->os_spa, txg,
2910 db->db_blkptr, data, DBUF_IS_L2CACHEABLE(db),
2911 DBUF_IS_L2COMPRESSIBLE(db), &zp, dbuf_write_ready,
2912 dbuf_write_physdone, dbuf_write_done, db,
2913 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
2914 }
2915 }