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, 2015 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(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid)
129 {
130 dbuf_hash_table_t *h = &dbuf_hash_table;
131 uint64_t hv = DBUF_HASH(os, obj, level, blkid);
132 uint64_t idx = hv & h->hash_table_mask;
133 dmu_buf_impl_t *db;
134
135 mutex_enter(DBUF_HASH_MUTEX(h, idx));
136 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
137 if (DBUF_EQUAL(db, os, obj, level, blkid)) {
138 mutex_enter(&db->db_mtx);
139 if (db->db_state != DB_EVICTING) {
140 mutex_exit(DBUF_HASH_MUTEX(h, idx));
141 return (db);
142 }
143 mutex_exit(&db->db_mtx);
144 }
145 }
146 mutex_exit(DBUF_HASH_MUTEX(h, idx));
147 return (NULL);
148 }
149
150 static dmu_buf_impl_t *
151 dbuf_find_bonus(objset_t *os, uint64_t object)
152 {
153 dnode_t *dn;
154 dmu_buf_impl_t *db = NULL;
155
156 if (dnode_hold(os, object, FTAG, &dn) == 0) {
157 rw_enter(&dn->dn_struct_rwlock, RW_READER);
158 if (dn->dn_bonus != NULL) {
159 db = dn->dn_bonus;
160 mutex_enter(&db->db_mtx);
161 }
162 rw_exit(&dn->dn_struct_rwlock);
163 dnode_rele(dn, FTAG);
164 }
165 return (db);
166 }
167
168 /*
169 * Insert an entry into the hash table. If there is already an element
170 * equal to elem in the hash table, then the already existing element
171 * will be returned and the new element will not be inserted.
172 * Otherwise returns NULL.
173 */
174 static dmu_buf_impl_t *
175 dbuf_hash_insert(dmu_buf_impl_t *db)
176 {
177 dbuf_hash_table_t *h = &dbuf_hash_table;
178 objset_t *os = db->db_objset;
179 uint64_t obj = db->db.db_object;
180 int level = db->db_level;
181 uint64_t blkid = db->db_blkid;
182 uint64_t hv = DBUF_HASH(os, obj, level, blkid);
183 uint64_t idx = hv & h->hash_table_mask;
184 dmu_buf_impl_t *dbf;
185
186 mutex_enter(DBUF_HASH_MUTEX(h, idx));
187 for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) {
188 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
189 mutex_enter(&dbf->db_mtx);
190 if (dbf->db_state != DB_EVICTING) {
191 mutex_exit(DBUF_HASH_MUTEX(h, idx));
192 return (dbf);
193 }
194 mutex_exit(&dbf->db_mtx);
195 }
196 }
197
198 mutex_enter(&db->db_mtx);
199 db->db_hash_next = h->hash_table[idx];
200 h->hash_table[idx] = db;
201 mutex_exit(DBUF_HASH_MUTEX(h, idx));
202 atomic_inc_64(&dbuf_hash_count);
203
204 return (NULL);
205 }
206
207 /*
208 * Remove an entry from the hash table. It must be in the EVICTING state.
209 */
210 static void
211 dbuf_hash_remove(dmu_buf_impl_t *db)
212 {
213 dbuf_hash_table_t *h = &dbuf_hash_table;
214 uint64_t hv = DBUF_HASH(db->db_objset, db->db.db_object,
215 db->db_level, db->db_blkid);
216 uint64_t idx = hv & h->hash_table_mask;
217 dmu_buf_impl_t *dbf, **dbp;
218
219 /*
220 * We musn't hold db_mtx to maintain lock ordering:
221 * DBUF_HASH_MUTEX > db_mtx.
222 */
223 ASSERT(refcount_is_zero(&db->db_holds));
224 ASSERT(db->db_state == DB_EVICTING);
225 ASSERT(!MUTEX_HELD(&db->db_mtx));
226
227 mutex_enter(DBUF_HASH_MUTEX(h, idx));
228 dbp = &h->hash_table[idx];
229 while ((dbf = *dbp) != db) {
230 dbp = &dbf->db_hash_next;
231 ASSERT(dbf != NULL);
232 }
233 *dbp = db->db_hash_next;
234 db->db_hash_next = NULL;
235 mutex_exit(DBUF_HASH_MUTEX(h, idx));
236 atomic_dec_64(&dbuf_hash_count);
237 }
238
239 static arc_evict_func_t dbuf_do_evict;
240
241 typedef enum {
242 DBVU_EVICTING,
243 DBVU_NOT_EVICTING
244 } dbvu_verify_type_t;
245
246 static void
247 dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type)
248 {
249 #ifdef ZFS_DEBUG
250 int64_t holds;
251
252 if (db->db_user == NULL)
253 return;
254
255 /* Only data blocks support the attachment of user data. */
256 ASSERT(db->db_level == 0);
257
258 /* Clients must resolve a dbuf before attaching user data. */
259 ASSERT(db->db.db_data != NULL);
260 ASSERT3U(db->db_state, ==, DB_CACHED);
261
262 holds = refcount_count(&db->db_holds);
263 if (verify_type == DBVU_EVICTING) {
264 /*
265 * Immediate eviction occurs when holds == dirtycnt.
266 * For normal eviction buffers, holds is zero on
267 * eviction, except when dbuf_fix_old_data() calls
268 * dbuf_clear_data(). However, the hold count can grow
269 * during eviction even though db_mtx is held (see
270 * dmu_bonus_hold() for an example), so we can only
271 * test the generic invariant that holds >= dirtycnt.
272 */
273 ASSERT3U(holds, >=, db->db_dirtycnt);
274 } else {
275 if (db->db_user_immediate_evict == TRUE)
276 ASSERT3U(holds, >=, db->db_dirtycnt);
277 else
278 ASSERT3U(holds, >, 0);
279 }
280 #endif
281 }
282
283 static void
284 dbuf_evict_user(dmu_buf_impl_t *db)
285 {
286 dmu_buf_user_t *dbu = db->db_user;
287
288 ASSERT(MUTEX_HELD(&db->db_mtx));
289
290 if (dbu == NULL)
291 return;
292
293 dbuf_verify_user(db, DBVU_EVICTING);
294 db->db_user = NULL;
295
296 #ifdef ZFS_DEBUG
297 if (dbu->dbu_clear_on_evict_dbufp != NULL)
298 *dbu->dbu_clear_on_evict_dbufp = NULL;
299 #endif
300
301 /*
302 * Invoke the callback from a taskq to avoid lock order reversals
303 * and limit stack depth.
304 */
305 taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func, dbu, 0,
306 &dbu->dbu_tqent);
307 }
308
309 boolean_t
310 dbuf_is_metadata(dmu_buf_impl_t *db)
311 {
312 if (db->db_level > 0) {
313 return (B_TRUE);
314 } else {
315 boolean_t is_metadata;
316
317 DB_DNODE_ENTER(db);
318 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type);
319 DB_DNODE_EXIT(db);
320
321 return (is_metadata);
322 }
323 }
324
325 void
326 dbuf_evict(dmu_buf_impl_t *db)
327 {
328 ASSERT(MUTEX_HELD(&db->db_mtx));
329 ASSERT(db->db_buf == NULL);
330 ASSERT(db->db_data_pending == NULL);
331
332 dbuf_clear(db);
333 dbuf_destroy(db);
334 }
335
336 void
337 dbuf_init(void)
338 {
339 uint64_t hsize = 1ULL << 16;
340 dbuf_hash_table_t *h = &dbuf_hash_table;
341 int i;
342
343 /*
344 * The hash table is big enough to fill all of physical memory
345 * with an average 4K block size. The table will take up
346 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
347 */
348 while (hsize * 4096 < physmem * PAGESIZE)
349 hsize <<= 1;
350
351 retry:
352 h->hash_table_mask = hsize - 1;
353 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
354 if (h->hash_table == NULL) {
355 /* XXX - we should really return an error instead of assert */
356 ASSERT(hsize > (1ULL << 10));
357 hsize >>= 1;
358 goto retry;
359 }
360
361 dbuf_cache = kmem_cache_create("dmu_buf_impl_t",
362 sizeof (dmu_buf_impl_t),
363 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
364
365 for (i = 0; i < DBUF_MUTEXES; i++)
366 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
367
368 /*
369 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
370 * configuration is not required.
371 */
372 dbu_evict_taskq = taskq_create("dbu_evict", 1, minclsyspri, 0, 0, 0);
373 }
374
375 void
376 dbuf_fini(void)
377 {
378 dbuf_hash_table_t *h = &dbuf_hash_table;
379 int i;
380
381 for (i = 0; i < DBUF_MUTEXES; i++)
382 mutex_destroy(&h->hash_mutexes[i]);
383 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
384 kmem_cache_destroy(dbuf_cache);
385 taskq_destroy(dbu_evict_taskq);
386 }
387
388 /*
389 * Other stuff.
390 */
391
392 #ifdef ZFS_DEBUG
393 static void
394 dbuf_verify(dmu_buf_impl_t *db)
395 {
396 dnode_t *dn;
397 dbuf_dirty_record_t *dr;
398
399 ASSERT(MUTEX_HELD(&db->db_mtx));
400
401 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
402 return;
403
404 ASSERT(db->db_objset != NULL);
405 DB_DNODE_ENTER(db);
406 dn = DB_DNODE(db);
407 if (dn == NULL) {
408 ASSERT(db->db_parent == NULL);
409 ASSERT(db->db_blkptr == NULL);
410 } else {
411 ASSERT3U(db->db.db_object, ==, dn->dn_object);
412 ASSERT3P(db->db_objset, ==, dn->dn_objset);
413 ASSERT3U(db->db_level, <, dn->dn_nlevels);
414 ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
415 db->db_blkid == DMU_SPILL_BLKID ||
416 !avl_is_empty(&dn->dn_dbufs));
417 }
418 if (db->db_blkid == DMU_BONUS_BLKID) {
419 ASSERT(dn != NULL);
420 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
421 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
422 } else if (db->db_blkid == DMU_SPILL_BLKID) {
423 ASSERT(dn != NULL);
424 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
425 ASSERT0(db->db.db_offset);
426 } else {
427 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
428 }
429
430 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
431 ASSERT(dr->dr_dbuf == db);
432
433 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
434 ASSERT(dr->dr_dbuf == db);
435
436 /*
437 * We can't assert that db_size matches dn_datablksz because it
438 * can be momentarily different when another thread is doing
439 * dnode_set_blksz().
440 */
441 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
442 dr = db->db_data_pending;
443 /*
444 * It should only be modified in syncing context, so
445 * make sure we only have one copy of the data.
446 */
447 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
448 }
449
450 /* verify db->db_blkptr */
451 if (db->db_blkptr) {
452 if (db->db_parent == dn->dn_dbuf) {
453 /* db is pointed to by the dnode */
454 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
455 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
456 ASSERT(db->db_parent == NULL);
457 else
458 ASSERT(db->db_parent != NULL);
459 if (db->db_blkid != DMU_SPILL_BLKID)
460 ASSERT3P(db->db_blkptr, ==,
461 &dn->dn_phys->dn_blkptr[db->db_blkid]);
462 } else {
463 /* db is pointed to by an indirect block */
464 int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
465 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
466 ASSERT3U(db->db_parent->db.db_object, ==,
467 db->db.db_object);
468 /*
469 * dnode_grow_indblksz() can make this fail if we don't
470 * have the struct_rwlock. XXX indblksz no longer
471 * grows. safe to do this now?
472 */
473 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
474 ASSERT3P(db->db_blkptr, ==,
475 ((blkptr_t *)db->db_parent->db.db_data +
476 db->db_blkid % epb));
477 }
478 }
479 }
480 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
481 (db->db_buf == NULL || db->db_buf->b_data) &&
482 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
483 db->db_state != DB_FILL && !dn->dn_free_txg) {
484 /*
485 * If the blkptr isn't set but they have nonzero data,
486 * it had better be dirty, otherwise we'll lose that
487 * data when we evict this buffer.
488 */
489 if (db->db_dirtycnt == 0) {
490 uint64_t *buf = db->db.db_data;
491 int i;
492
493 for (i = 0; i < db->db.db_size >> 3; i++) {
494 ASSERT(buf[i] == 0);
495 }
496 }
497 }
498 DB_DNODE_EXIT(db);
499 }
500 #endif
501
502 static void
503 dbuf_clear_data(dmu_buf_impl_t *db)
504 {
505 ASSERT(MUTEX_HELD(&db->db_mtx));
506 dbuf_evict_user(db);
507 db->db_buf = NULL;
508 db->db.db_data = NULL;
509 if (db->db_state != DB_NOFILL)
510 db->db_state = DB_UNCACHED;
511 }
512
513 static void
514 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
515 {
516 ASSERT(MUTEX_HELD(&db->db_mtx));
517 ASSERT(buf != NULL);
518
519 db->db_buf = buf;
520 ASSERT(buf->b_data != NULL);
521 db->db.db_data = buf->b_data;
522 if (!arc_released(buf))
523 arc_set_callback(buf, dbuf_do_evict, db);
524 }
525
526 /*
527 * Loan out an arc_buf for read. Return the loaned arc_buf.
528 */
529 arc_buf_t *
530 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
531 {
532 arc_buf_t *abuf;
533
534 mutex_enter(&db->db_mtx);
535 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
536 int blksz = db->db.db_size;
537 spa_t *spa = db->db_objset->os_spa;
538
539 mutex_exit(&db->db_mtx);
540 abuf = arc_loan_buf(spa, blksz);
541 bcopy(db->db.db_data, abuf->b_data, blksz);
542 } else {
543 abuf = db->db_buf;
544 arc_loan_inuse_buf(abuf, db);
545 dbuf_clear_data(db);
546 mutex_exit(&db->db_mtx);
547 }
548 return (abuf);
549 }
550
551 /*
552 * Calculate which level n block references the data at the level 0 offset
553 * provided.
554 */
555 uint64_t
556 dbuf_whichblock(dnode_t *dn, int64_t level, uint64_t offset)
557 {
558 if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) {
559 /*
560 * The level n blkid is equal to the level 0 blkid divided by
561 * the number of level 0s in a level n block.
562 *
563 * The level 0 blkid is offset >> datablkshift =
564 * offset / 2^datablkshift.
565 *
566 * The number of level 0s in a level n is the number of block
567 * pointers in an indirect block, raised to the power of level.
568 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
569 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
570 *
571 * Thus, the level n blkid is: offset /
572 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
573 * = offset / 2^(datablkshift + level *
574 * (indblkshift - SPA_BLKPTRSHIFT))
575 * = offset >> (datablkshift + level *
576 * (indblkshift - SPA_BLKPTRSHIFT))
577 */
578 return (offset >> (dn->dn_datablkshift + level *
579 (dn->dn_indblkshift - SPA_BLKPTRSHIFT)));
580 } else {
581 ASSERT3U(offset, <, dn->dn_datablksz);
582 return (0);
583 }
584 }
585
586 static void
587 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
588 {
589 dmu_buf_impl_t *db = vdb;
590
591 mutex_enter(&db->db_mtx);
592 ASSERT3U(db->db_state, ==, DB_READ);
593 /*
594 * All reads are synchronous, so we must have a hold on the dbuf
595 */
596 ASSERT(refcount_count(&db->db_holds) > 0);
597 ASSERT(db->db_buf == NULL);
598 ASSERT(db->db.db_data == NULL);
599 if (db->db_level == 0 && db->db_freed_in_flight) {
600 /* we were freed in flight; disregard any error */
601 arc_release(buf, db);
602 bzero(buf->b_data, db->db.db_size);
603 arc_buf_freeze(buf);
604 db->db_freed_in_flight = FALSE;
605 dbuf_set_data(db, buf);
606 db->db_state = DB_CACHED;
607 } else if (zio == NULL || zio->io_error == 0) {
608 dbuf_set_data(db, buf);
609 db->db_state = DB_CACHED;
610 } else {
611 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
612 ASSERT3P(db->db_buf, ==, NULL);
613 VERIFY(arc_buf_remove_ref(buf, db));
614 db->db_state = DB_UNCACHED;
615 }
616 cv_broadcast(&db->db_changed);
617 dbuf_rele_and_unlock(db, NULL);
618 }
619
620 static void
621 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
622 {
623 dnode_t *dn;
624 zbookmark_phys_t zb;
625 arc_flags_t aflags = ARC_FLAG_NOWAIT;
626
627 DB_DNODE_ENTER(db);
628 dn = DB_DNODE(db);
629 ASSERT(!refcount_is_zero(&db->db_holds));
630 /* We need the struct_rwlock to prevent db_blkptr from changing. */
631 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
632 ASSERT(MUTEX_HELD(&db->db_mtx));
633 ASSERT(db->db_state == DB_UNCACHED);
634 ASSERT(db->db_buf == NULL);
635
636 if (db->db_blkid == DMU_BONUS_BLKID) {
637 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
638
639 ASSERT3U(bonuslen, <=, db->db.db_size);
640 db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN);
641 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
642 if (bonuslen < DN_MAX_BONUSLEN)
643 bzero(db->db.db_data, DN_MAX_BONUSLEN);
644 if (bonuslen) {
645 /*
646 * Absent byzantine on-disk corruption, we fully expect
647 * our bonuslen to be no more than DN_MAX_BONUSLEN --
648 * but we nonetheless explicitly clamp it on the
649 * bcopy() to prevent any on-disk corruption from
650 * becoming rampant in-kernel corruption.
651 */
652 if (bonuslen > DN_MAX_BONUSLEN) {
653 DTRACE_PROBE3(dbuf__read__impl__toolong, int,
654 bonuslen, dnode_t *, dn, dmu_buf_impl_t *,
655 db);
656 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data,
657 DN_MAX_BONUSLEN);
658 } else {
659 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data,
660 bonuslen);
661 }
662 }
663 DB_DNODE_EXIT(db);
664 db->db_state = DB_CACHED;
665 mutex_exit(&db->db_mtx);
666 return;
667 }
668
669 /*
670 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
671 * processes the delete record and clears the bp while we are waiting
672 * for the dn_mtx (resulting in a "no" from block_freed).
673 */
674 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
675 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
676 BP_IS_HOLE(db->db_blkptr)))) {
677 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
678
679 DB_DNODE_EXIT(db);
680 dbuf_set_data(db, arc_buf_alloc(db->db_objset->os_spa,
681 db->db.db_size, db, type));
682 bzero(db->db.db_data, db->db.db_size);
683 db->db_state = DB_CACHED;
684 mutex_exit(&db->db_mtx);
685 return;
686 }
687
688 DB_DNODE_EXIT(db);
689
690 db->db_state = DB_READ;
691 mutex_exit(&db->db_mtx);
692
693 if (DBUF_IS_L2CACHEABLE(db))
694 aflags |= ARC_FLAG_L2CACHE;
695 if (DBUF_IS_L2COMPRESSIBLE(db))
696 aflags |= ARC_FLAG_L2COMPRESS;
697
698 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
699 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
700 db->db.db_object, db->db_level, db->db_blkid);
701
702 dbuf_add_ref(db, NULL);
703
704 (void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr,
705 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
706 (flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
707 &aflags, &zb);
708 }
709
710 int
711 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
712 {
713 int err = 0;
714 boolean_t havepzio = (zio != NULL);
715 boolean_t prefetch;
716 dnode_t *dn;
717
718 /*
719 * We don't have to hold the mutex to check db_state because it
720 * can't be freed while we have a hold on the buffer.
721 */
722 ASSERT(!refcount_is_zero(&db->db_holds));
723
724 if (db->db_state == DB_NOFILL)
725 return (SET_ERROR(EIO));
726
727 DB_DNODE_ENTER(db);
728 dn = DB_DNODE(db);
729 if ((flags & DB_RF_HAVESTRUCT) == 0)
730 rw_enter(&dn->dn_struct_rwlock, RW_READER);
731
732 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
733 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
734 DBUF_IS_CACHEABLE(db);
735
736 mutex_enter(&db->db_mtx);
737 if (db->db_state == DB_CACHED) {
738 mutex_exit(&db->db_mtx);
739 if (prefetch)
740 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1);
741 if ((flags & DB_RF_HAVESTRUCT) == 0)
742 rw_exit(&dn->dn_struct_rwlock);
743 DB_DNODE_EXIT(db);
744 } else if (db->db_state == DB_UNCACHED) {
745 spa_t *spa = dn->dn_objset->os_spa;
746
747 if (zio == NULL)
748 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
749 dbuf_read_impl(db, zio, flags);
750
751 /* dbuf_read_impl has dropped db_mtx for us */
752
753 if (prefetch)
754 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1);
755
756 if ((flags & DB_RF_HAVESTRUCT) == 0)
757 rw_exit(&dn->dn_struct_rwlock);
758 DB_DNODE_EXIT(db);
759
760 if (!havepzio)
761 err = zio_wait(zio);
762 } else {
763 /*
764 * Another reader came in while the dbuf was in flight
765 * between UNCACHED and CACHED. Either a writer will finish
766 * writing the buffer (sending the dbuf to CACHED) or the
767 * first reader's request will reach the read_done callback
768 * and send the dbuf to CACHED. Otherwise, a failure
769 * occurred and the dbuf went to UNCACHED.
770 */
771 mutex_exit(&db->db_mtx);
772 if (prefetch)
773 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1);
774 if ((flags & DB_RF_HAVESTRUCT) == 0)
775 rw_exit(&dn->dn_struct_rwlock);
776 DB_DNODE_EXIT(db);
777
778 /* Skip the wait per the caller's request. */
779 mutex_enter(&db->db_mtx);
780 if ((flags & DB_RF_NEVERWAIT) == 0) {
781 while (db->db_state == DB_READ ||
782 db->db_state == DB_FILL) {
783 ASSERT(db->db_state == DB_READ ||
784 (flags & DB_RF_HAVESTRUCT) == 0);
785 DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *,
786 db, zio_t *, zio);
787 cv_wait(&db->db_changed, &db->db_mtx);
788 }
789 if (db->db_state == DB_UNCACHED)
790 err = SET_ERROR(EIO);
791 }
792 mutex_exit(&db->db_mtx);
793 }
794
795 ASSERT(err || havepzio || db->db_state == DB_CACHED);
796 return (err);
797 }
798
799 static void
800 dbuf_noread(dmu_buf_impl_t *db)
801 {
802 ASSERT(!refcount_is_zero(&db->db_holds));
803 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
804 mutex_enter(&db->db_mtx);
805 while (db->db_state == DB_READ || db->db_state == DB_FILL)
806 cv_wait(&db->db_changed, &db->db_mtx);
807 if (db->db_state == DB_UNCACHED) {
808 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
809 spa_t *spa = db->db_objset->os_spa;
810
811 ASSERT(db->db_buf == NULL);
812 ASSERT(db->db.db_data == NULL);
813 dbuf_set_data(db, arc_buf_alloc(spa, db->db.db_size, db, type));
814 db->db_state = DB_FILL;
815 } else if (db->db_state == DB_NOFILL) {
816 dbuf_clear_data(db);
817 } else {
818 ASSERT3U(db->db_state, ==, DB_CACHED);
819 }
820 mutex_exit(&db->db_mtx);
821 }
822
823 /*
824 * This is our just-in-time copy function. It makes a copy of
825 * buffers, that have been modified in a previous transaction
826 * group, before we modify them in the current active group.
827 *
828 * This function is used in two places: when we are dirtying a
829 * buffer for the first time in a txg, and when we are freeing
830 * a range in a dnode that includes this buffer.
831 *
832 * Note that when we are called from dbuf_free_range() we do
833 * not put a hold on the buffer, we just traverse the active
834 * dbuf list for the dnode.
835 */
836 static void
837 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
838 {
839 dbuf_dirty_record_t *dr = db->db_last_dirty;
840
841 ASSERT(MUTEX_HELD(&db->db_mtx));
842 ASSERT(db->db.db_data != NULL);
843 ASSERT(db->db_level == 0);
844 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
845
846 if (dr == NULL ||
847 (dr->dt.dl.dr_data !=
848 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
849 return;
850
851 /*
852 * If the last dirty record for this dbuf has not yet synced
853 * and its referencing the dbuf data, either:
854 * reset the reference to point to a new copy,
855 * or (if there a no active holders)
856 * just null out the current db_data pointer.
857 */
858 ASSERT(dr->dr_txg >= txg - 2);
859 if (db->db_blkid == DMU_BONUS_BLKID) {
860 /* Note that the data bufs here are zio_bufs */
861 dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
862 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
863 bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
864 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
865 int size = db->db.db_size;
866 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
867 spa_t *spa = db->db_objset->os_spa;
868
869 dr->dt.dl.dr_data = arc_buf_alloc(spa, size, db, type);
870 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
871 } else {
872 dbuf_clear_data(db);
873 }
874 }
875
876 void
877 dbuf_unoverride(dbuf_dirty_record_t *dr)
878 {
879 dmu_buf_impl_t *db = dr->dr_dbuf;
880 blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
881 uint64_t txg = dr->dr_txg;
882
883 ASSERT(MUTEX_HELD(&db->db_mtx));
884 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
885 ASSERT(db->db_level == 0);
886
887 if (db->db_blkid == DMU_BONUS_BLKID ||
888 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
889 return;
890
891 ASSERT(db->db_data_pending != dr);
892
893 /* free this block */
894 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite)
895 zio_free(db->db_objset->os_spa, txg, bp);
896
897 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
898 dr->dt.dl.dr_nopwrite = B_FALSE;
899
900 /*
901 * Release the already-written buffer, so we leave it in
902 * a consistent dirty state. Note that all callers are
903 * modifying the buffer, so they will immediately do
904 * another (redundant) arc_release(). Therefore, leave
905 * the buf thawed to save the effort of freezing &
906 * immediately re-thawing it.
907 */
908 arc_release(dr->dt.dl.dr_data, db);
909 }
910
911 /*
912 * Evict (if its unreferenced) or clear (if its referenced) any level-0
913 * data blocks in the free range, so that any future readers will find
914 * empty blocks.
915 *
916 * This is a no-op if the dataset is in the middle of an incremental
917 * receive; see comment below for details.
918 */
919 void
920 dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
921 dmu_tx_t *tx)
922 {
923 dmu_buf_impl_t db_search;
924 dmu_buf_impl_t *db, *db_next;
925 uint64_t txg = tx->tx_txg;
926 avl_index_t where;
927
928 if (end_blkid > dn->dn_maxblkid && (end_blkid != DMU_SPILL_BLKID))
929 end_blkid = dn->dn_maxblkid;
930 dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid);
931
932 db_search.db_level = 0;
933 db_search.db_blkid = start_blkid;
934 db_search.db_state = DB_SEARCH;
935
936 mutex_enter(&dn->dn_dbufs_mtx);
937 if (start_blkid >= dn->dn_unlisted_l0_blkid) {
938 /* There can't be any dbufs in this range; no need to search. */
939 #ifdef DEBUG
940 db = avl_find(&dn->dn_dbufs, &db_search, &where);
941 ASSERT3P(db, ==, NULL);
942 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
943 ASSERT(db == NULL || db->db_level > 0);
944 #endif
945 mutex_exit(&dn->dn_dbufs_mtx);
946 return;
947 } else if (dmu_objset_is_receiving(dn->dn_objset)) {
948 /*
949 * If we are receiving, we expect there to be no dbufs in
950 * the range to be freed, because receive modifies each
951 * block at most once, and in offset order. If this is
952 * not the case, it can lead to performance problems,
953 * so note that we unexpectedly took the slow path.
954 */
955 atomic_inc_64(&zfs_free_range_recv_miss);
956 }
957
958 db = avl_find(&dn->dn_dbufs, &db_search, &where);
959 ASSERT3P(db, ==, NULL);
960 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
961
962 for (; db != NULL; db = db_next) {
963 db_next = AVL_NEXT(&dn->dn_dbufs, db);
964 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
965
966 if (db->db_level != 0 || db->db_blkid > end_blkid) {
967 break;
968 }
969 ASSERT3U(db->db_blkid, >=, start_blkid);
970
971 /* found a level 0 buffer in the range */
972 mutex_enter(&db->db_mtx);
973 if (dbuf_undirty(db, tx)) {
974 /* mutex has been dropped and dbuf destroyed */
975 continue;
976 }
977
978 if (db->db_state == DB_UNCACHED ||
979 db->db_state == DB_NOFILL ||
980 db->db_state == DB_EVICTING) {
981 ASSERT(db->db.db_data == NULL);
982 mutex_exit(&db->db_mtx);
983 continue;
984 }
985 if (db->db_state == DB_READ || db->db_state == DB_FILL) {
986 /* will be handled in dbuf_read_done or dbuf_rele */
987 db->db_freed_in_flight = TRUE;
988 mutex_exit(&db->db_mtx);
989 continue;
990 }
991 if (refcount_count(&db->db_holds) == 0) {
992 ASSERT(db->db_buf);
993 dbuf_clear(db);
994 continue;
995 }
996 /* The dbuf is referenced */
997
998 if (db->db_last_dirty != NULL) {
999 dbuf_dirty_record_t *dr = db->db_last_dirty;
1000
1001 if (dr->dr_txg == txg) {
1002 /*
1003 * This buffer is "in-use", re-adjust the file
1004 * size to reflect that this buffer may
1005 * contain new data when we sync.
1006 */
1007 if (db->db_blkid != DMU_SPILL_BLKID &&
1008 db->db_blkid > dn->dn_maxblkid)
1009 dn->dn_maxblkid = db->db_blkid;
1010 dbuf_unoverride(dr);
1011 } else {
1012 /*
1013 * This dbuf is not dirty in the open context.
1014 * Either uncache it (if its not referenced in
1015 * the open context) or reset its contents to
1016 * empty.
1017 */
1018 dbuf_fix_old_data(db, txg);
1019 }
1020 }
1021 /* clear the contents if its cached */
1022 if (db->db_state == DB_CACHED) {
1023 ASSERT(db->db.db_data != NULL);
1024 arc_release(db->db_buf, db);
1025 bzero(db->db.db_data, db->db.db_size);
1026 arc_buf_freeze(db->db_buf);
1027 }
1028
1029 mutex_exit(&db->db_mtx);
1030 }
1031 mutex_exit(&dn->dn_dbufs_mtx);
1032 }
1033
1034 static int
1035 dbuf_block_freeable(dmu_buf_impl_t *db)
1036 {
1037 dsl_dataset_t *ds = db->db_objset->os_dsl_dataset;
1038 uint64_t birth_txg = 0;
1039
1040 /*
1041 * We don't need any locking to protect db_blkptr:
1042 * If it's syncing, then db_last_dirty will be set
1043 * so we'll ignore db_blkptr.
1044 *
1045 * This logic ensures that only block births for
1046 * filled blocks are considered.
1047 */
1048 ASSERT(MUTEX_HELD(&db->db_mtx));
1049 if (db->db_last_dirty && (db->db_blkptr == NULL ||
1050 !BP_IS_HOLE(db->db_blkptr))) {
1051 birth_txg = db->db_last_dirty->dr_txg;
1052 } else if (db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) {
1053 birth_txg = db->db_blkptr->blk_birth;
1054 }
1055
1056 /*
1057 * If this block don't exist or is in a snapshot, it can't be freed.
1058 * Don't pass the bp to dsl_dataset_block_freeable() since we
1059 * are holding the db_mtx lock and might deadlock if we are
1060 * prefetching a dedup-ed block.
1061 */
1062 if (birth_txg != 0)
1063 return (ds == NULL ||
1064 dsl_dataset_block_freeable(ds, NULL, birth_txg));
1065 else
1066 return (B_FALSE);
1067 }
1068
1069 void
1070 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
1071 {
1072 arc_buf_t *buf, *obuf;
1073 int osize = db->db.db_size;
1074 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1075 dnode_t *dn;
1076
1077 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1078
1079 DB_DNODE_ENTER(db);
1080 dn = DB_DNODE(db);
1081
1082 /* XXX does *this* func really need the lock? */
1083 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1084
1085 /*
1086 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1087 * is OK, because there can be no other references to the db
1088 * when we are changing its size, so no concurrent DB_FILL can
1089 * be happening.
1090 */
1091 /*
1092 * XXX we should be doing a dbuf_read, checking the return
1093 * value and returning that up to our callers
1094 */
1095 dmu_buf_will_dirty(&db->db, tx);
1096
1097 /* create the data buffer for the new block */
1098 buf = arc_buf_alloc(dn->dn_objset->os_spa, size, db, type);
1099
1100 /* copy old block data to the new block */
1101 obuf = db->db_buf;
1102 bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
1103 /* zero the remainder */
1104 if (size > osize)
1105 bzero((uint8_t *)buf->b_data + osize, size - osize);
1106
1107 mutex_enter(&db->db_mtx);
1108 dbuf_set_data(db, buf);
1109 VERIFY(arc_buf_remove_ref(obuf, db));
1110 db->db.db_size = size;
1111
1112 if (db->db_level == 0) {
1113 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1114 db->db_last_dirty->dt.dl.dr_data = buf;
1115 }
1116 mutex_exit(&db->db_mtx);
1117
1118 dnode_willuse_space(dn, size-osize, tx);
1119 DB_DNODE_EXIT(db);
1120 }
1121
1122 void
1123 dbuf_release_bp(dmu_buf_impl_t *db)
1124 {
1125 objset_t *os = db->db_objset;
1126
1127 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1128 ASSERT(arc_released(os->os_phys_buf) ||
1129 list_link_active(&os->os_dsl_dataset->ds_synced_link));
1130 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1131
1132 (void) arc_release(db->db_buf, db);
1133 }
1134
1135 /*
1136 * We already have a dirty record for this TXG, and we are being
1137 * dirtied again.
1138 */
1139 static void
1140 dbuf_redirty(dbuf_dirty_record_t *dr)
1141 {
1142 dmu_buf_impl_t *db = dr->dr_dbuf;
1143
1144 ASSERT(MUTEX_HELD(&db->db_mtx));
1145
1146 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1147 /*
1148 * If this buffer has already been written out,
1149 * we now need to reset its state.
1150 */
1151 dbuf_unoverride(dr);
1152 if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1153 db->db_state != DB_NOFILL) {
1154 /* Already released on initial dirty, so just thaw. */
1155 ASSERT(arc_released(db->db_buf));
1156 arc_buf_thaw(db->db_buf);
1157 }
1158 }
1159 }
1160
1161 dbuf_dirty_record_t *
1162 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1163 {
1164 dnode_t *dn;
1165 objset_t *os;
1166 dbuf_dirty_record_t **drp, *dr;
1167 int drop_struct_lock = FALSE;
1168 boolean_t do_free_accounting = B_FALSE;
1169 int txgoff = tx->tx_txg & TXG_MASK;
1170
1171 ASSERT(tx->tx_txg != 0);
1172 ASSERT(!refcount_is_zero(&db->db_holds));
1173 DMU_TX_DIRTY_BUF(tx, db);
1174
1175 DB_DNODE_ENTER(db);
1176 dn = DB_DNODE(db);
1177 /*
1178 * Shouldn't dirty a regular buffer in syncing context. Private
1179 * objects may be dirtied in syncing context, but only if they
1180 * were already pre-dirtied in open context.
1181 */
1182 ASSERT(!dmu_tx_is_syncing(tx) ||
1183 BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1184 DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1185 dn->dn_objset->os_dsl_dataset == NULL);
1186 /*
1187 * We make this assert for private objects as well, but after we
1188 * check if we're already dirty. They are allowed to re-dirty
1189 * in syncing context.
1190 */
1191 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1192 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1193 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1194
1195 mutex_enter(&db->db_mtx);
1196 /*
1197 * XXX make this true for indirects too? The problem is that
1198 * transactions created with dmu_tx_create_assigned() from
1199 * syncing context don't bother holding ahead.
1200 */
1201 ASSERT(db->db_level != 0 ||
1202 db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1203 db->db_state == DB_NOFILL);
1204
1205 mutex_enter(&dn->dn_mtx);
1206 /*
1207 * Don't set dirtyctx to SYNC if we're just modifying this as we
1208 * initialize the objset.
1209 */
1210 if (dn->dn_dirtyctx == DN_UNDIRTIED &&
1211 !BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1212 dn->dn_dirtyctx =
1213 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1214 ASSERT(dn->dn_dirtyctx_firstset == NULL);
1215 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1216 }
1217 mutex_exit(&dn->dn_mtx);
1218
1219 if (db->db_blkid == DMU_SPILL_BLKID)
1220 dn->dn_have_spill = B_TRUE;
1221
1222 /*
1223 * If this buffer is already dirty, we're done.
1224 */
1225 drp = &db->db_last_dirty;
1226 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1227 db->db.db_object == DMU_META_DNODE_OBJECT);
1228 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1229 drp = &dr->dr_next;
1230 if (dr && dr->dr_txg == tx->tx_txg) {
1231 DB_DNODE_EXIT(db);
1232
1233 dbuf_redirty(dr);
1234 mutex_exit(&db->db_mtx);
1235 return (dr);
1236 }
1237
1238 /*
1239 * Only valid if not already dirty.
1240 */
1241 ASSERT(dn->dn_object == 0 ||
1242 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1243 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1244
1245 ASSERT3U(dn->dn_nlevels, >, db->db_level);
1246 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1247 dn->dn_phys->dn_nlevels > db->db_level ||
1248 dn->dn_next_nlevels[txgoff] > db->db_level ||
1249 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1250 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1251
1252 /*
1253 * We should only be dirtying in syncing context if it's the
1254 * mos or we're initializing the os or it's a special object.
1255 * However, we are allowed to dirty in syncing context provided
1256 * we already dirtied it in open context. Hence we must make
1257 * this assertion only if we're not already dirty.
1258 */
1259 os = dn->dn_objset;
1260 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1261 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1262 ASSERT(db->db.db_size != 0);
1263
1264 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1265
1266 if (db->db_blkid != DMU_BONUS_BLKID) {
1267 /*
1268 * Update the accounting.
1269 * Note: we delay "free accounting" until after we drop
1270 * the db_mtx. This keeps us from grabbing other locks
1271 * (and possibly deadlocking) in bp_get_dsize() while
1272 * also holding the db_mtx.
1273 */
1274 dnode_willuse_space(dn, db->db.db_size, tx);
1275 do_free_accounting = dbuf_block_freeable(db);
1276 }
1277
1278 /*
1279 * If this buffer is dirty in an old transaction group we need
1280 * to make a copy of it so that the changes we make in this
1281 * transaction group won't leak out when we sync the older txg.
1282 */
1283 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1284 if (db->db_level == 0) {
1285 void *data_old = db->db_buf;
1286
1287 if (db->db_state != DB_NOFILL) {
1288 if (db->db_blkid == DMU_BONUS_BLKID) {
1289 dbuf_fix_old_data(db, tx->tx_txg);
1290 data_old = db->db.db_data;
1291 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1292 /*
1293 * Release the data buffer from the cache so
1294 * that we can modify it without impacting
1295 * possible other users of this cached data
1296 * block. Note that indirect blocks and
1297 * private objects are not released until the
1298 * syncing state (since they are only modified
1299 * then).
1300 */
1301 arc_release(db->db_buf, db);
1302 dbuf_fix_old_data(db, tx->tx_txg);
1303 data_old = db->db_buf;
1304 }
1305 ASSERT(data_old != NULL);
1306 }
1307 dr->dt.dl.dr_data = data_old;
1308 } else {
1309 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1310 list_create(&dr->dt.di.dr_children,
1311 sizeof (dbuf_dirty_record_t),
1312 offsetof(dbuf_dirty_record_t, dr_dirty_node));
1313 }
1314 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL)
1315 dr->dr_accounted = db->db.db_size;
1316 dr->dr_dbuf = db;
1317 dr->dr_txg = tx->tx_txg;
1318 dr->dr_next = *drp;
1319 *drp = dr;
1320
1321 /*
1322 * We could have been freed_in_flight between the dbuf_noread
1323 * and dbuf_dirty. We win, as though the dbuf_noread() had
1324 * happened after the free.
1325 */
1326 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1327 db->db_blkid != DMU_SPILL_BLKID) {
1328 mutex_enter(&dn->dn_mtx);
1329 if (dn->dn_free_ranges[txgoff] != NULL) {
1330 range_tree_clear(dn->dn_free_ranges[txgoff],
1331 db->db_blkid, 1);
1332 }
1333 mutex_exit(&dn->dn_mtx);
1334 db->db_freed_in_flight = FALSE;
1335 }
1336
1337 /*
1338 * This buffer is now part of this txg
1339 */
1340 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1341 db->db_dirtycnt += 1;
1342 ASSERT3U(db->db_dirtycnt, <=, 3);
1343
1344 mutex_exit(&db->db_mtx);
1345
1346 if (db->db_blkid == DMU_BONUS_BLKID ||
1347 db->db_blkid == DMU_SPILL_BLKID) {
1348 mutex_enter(&dn->dn_mtx);
1349 ASSERT(!list_link_active(&dr->dr_dirty_node));
1350 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1351 mutex_exit(&dn->dn_mtx);
1352 dnode_setdirty(dn, tx);
1353 DB_DNODE_EXIT(db);
1354 return (dr);
1355 } else if (do_free_accounting) {
1356 blkptr_t *bp = db->db_blkptr;
1357 int64_t willfree = (bp && !BP_IS_HOLE(bp)) ?
1358 bp_get_dsize(os->os_spa, bp) : db->db.db_size;
1359 /*
1360 * This is only a guess -- if the dbuf is dirty
1361 * in a previous txg, we don't know how much
1362 * space it will use on disk yet. We should
1363 * really have the struct_rwlock to access
1364 * db_blkptr, but since this is just a guess,
1365 * it's OK if we get an odd answer.
1366 */
1367 ddt_prefetch(os->os_spa, bp);
1368 dnode_willuse_space(dn, -willfree, tx);
1369 }
1370
1371 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1372 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1373 drop_struct_lock = TRUE;
1374 }
1375
1376 if (db->db_level == 0) {
1377 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1378 ASSERT(dn->dn_maxblkid >= db->db_blkid);
1379 }
1380
1381 if (db->db_level+1 < dn->dn_nlevels) {
1382 dmu_buf_impl_t *parent = db->db_parent;
1383 dbuf_dirty_record_t *di;
1384 int parent_held = FALSE;
1385
1386 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1387 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1388
1389 parent = dbuf_hold_level(dn, db->db_level+1,
1390 db->db_blkid >> epbs, FTAG);
1391 ASSERT(parent != NULL);
1392 parent_held = TRUE;
1393 }
1394 if (drop_struct_lock)
1395 rw_exit(&dn->dn_struct_rwlock);
1396 ASSERT3U(db->db_level+1, ==, parent->db_level);
1397 di = dbuf_dirty(parent, tx);
1398 if (parent_held)
1399 dbuf_rele(parent, FTAG);
1400
1401 mutex_enter(&db->db_mtx);
1402 /*
1403 * Since we've dropped the mutex, it's possible that
1404 * dbuf_undirty() might have changed this out from under us.
1405 */
1406 if (db->db_last_dirty == dr ||
1407 dn->dn_object == DMU_META_DNODE_OBJECT) {
1408 mutex_enter(&di->dt.di.dr_mtx);
1409 ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1410 ASSERT(!list_link_active(&dr->dr_dirty_node));
1411 list_insert_tail(&di->dt.di.dr_children, dr);
1412 mutex_exit(&di->dt.di.dr_mtx);
1413 dr->dr_parent = di;
1414 }
1415 mutex_exit(&db->db_mtx);
1416 } else {
1417 ASSERT(db->db_level+1 == dn->dn_nlevels);
1418 ASSERT(db->db_blkid < dn->dn_nblkptr);
1419 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1420 mutex_enter(&dn->dn_mtx);
1421 ASSERT(!list_link_active(&dr->dr_dirty_node));
1422 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1423 mutex_exit(&dn->dn_mtx);
1424 if (drop_struct_lock)
1425 rw_exit(&dn->dn_struct_rwlock);
1426 }
1427
1428 dnode_setdirty(dn, tx);
1429 DB_DNODE_EXIT(db);
1430 return (dr);
1431 }
1432
1433 /*
1434 * Undirty a buffer in the transaction group referenced by the given
1435 * transaction. Return whether this evicted the dbuf.
1436 */
1437 static boolean_t
1438 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1439 {
1440 dnode_t *dn;
1441 uint64_t txg = tx->tx_txg;
1442 dbuf_dirty_record_t *dr, **drp;
1443
1444 ASSERT(txg != 0);
1445
1446 /*
1447 * Due to our use of dn_nlevels below, this can only be called
1448 * in open context, unless we are operating on the MOS.
1449 * From syncing context, dn_nlevels may be different from the
1450 * dn_nlevels used when dbuf was dirtied.
1451 */
1452 ASSERT(db->db_objset ==
1453 dmu_objset_pool(db->db_objset)->dp_meta_objset ||
1454 txg != spa_syncing_txg(dmu_objset_spa(db->db_objset)));
1455 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1456 ASSERT0(db->db_level);
1457 ASSERT(MUTEX_HELD(&db->db_mtx));
1458
1459 /*
1460 * If this buffer is not dirty, we're done.
1461 */
1462 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1463 if (dr->dr_txg <= txg)
1464 break;
1465 if (dr == NULL || dr->dr_txg < txg)
1466 return (B_FALSE);
1467 ASSERT(dr->dr_txg == txg);
1468 ASSERT(dr->dr_dbuf == db);
1469
1470 DB_DNODE_ENTER(db);
1471 dn = DB_DNODE(db);
1472
1473 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1474
1475 ASSERT(db->db.db_size != 0);
1476
1477 dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset),
1478 dr->dr_accounted, txg);
1479
1480 *drp = dr->dr_next;
1481
1482 /*
1483 * Note that there are three places in dbuf_dirty()
1484 * where this dirty record may be put on a list.
1485 * Make sure to do a list_remove corresponding to
1486 * every one of those list_insert calls.
1487 */
1488 if (dr->dr_parent) {
1489 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1490 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1491 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1492 } else if (db->db_blkid == DMU_SPILL_BLKID ||
1493 db->db_level + 1 == dn->dn_nlevels) {
1494 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1495 mutex_enter(&dn->dn_mtx);
1496 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1497 mutex_exit(&dn->dn_mtx);
1498 }
1499 DB_DNODE_EXIT(db);
1500
1501 if (db->db_state != DB_NOFILL) {
1502 dbuf_unoverride(dr);
1503
1504 ASSERT(db->db_buf != NULL);
1505 ASSERT(dr->dt.dl.dr_data != NULL);
1506 if (dr->dt.dl.dr_data != db->db_buf)
1507 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db));
1508 }
1509
1510 kmem_free(dr, sizeof (dbuf_dirty_record_t));
1511
1512 ASSERT(db->db_dirtycnt > 0);
1513 db->db_dirtycnt -= 1;
1514
1515 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1516 arc_buf_t *buf = db->db_buf;
1517
1518 ASSERT(db->db_state == DB_NOFILL || arc_released(buf));
1519 dbuf_clear_data(db);
1520 VERIFY(arc_buf_remove_ref(buf, db));
1521 dbuf_evict(db);
1522 return (B_TRUE);
1523 }
1524
1525 return (B_FALSE);
1526 }
1527
1528 void
1529 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
1530 {
1531 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1532 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1533
1534 ASSERT(tx->tx_txg != 0);
1535 ASSERT(!refcount_is_zero(&db->db_holds));
1536
1537 /*
1538 * Quick check for dirtyness. For already dirty blocks, this
1539 * reduces runtime of this function by >90%, and overall performance
1540 * by 50% for some workloads (e.g. file deletion with indirect blocks
1541 * cached).
1542 */
1543 mutex_enter(&db->db_mtx);
1544 dbuf_dirty_record_t *dr;
1545 for (dr = db->db_last_dirty;
1546 dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) {
1547 /*
1548 * It's possible that it is already dirty but not cached,
1549 * because there are some calls to dbuf_dirty() that don't
1550 * go through dmu_buf_will_dirty().
1551 */
1552 if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) {
1553 /* This dbuf is already dirty and cached. */
1554 dbuf_redirty(dr);
1555 mutex_exit(&db->db_mtx);
1556 return;
1557 }
1558 }
1559 mutex_exit(&db->db_mtx);
1560
1561 DB_DNODE_ENTER(db);
1562 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
1563 rf |= DB_RF_HAVESTRUCT;
1564 DB_DNODE_EXIT(db);
1565 (void) dbuf_read(db, NULL, rf);
1566 (void) dbuf_dirty(db, tx);
1567 }
1568
1569 void
1570 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1571 {
1572 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1573
1574 db->db_state = DB_NOFILL;
1575
1576 dmu_buf_will_fill(db_fake, tx);
1577 }
1578
1579 void
1580 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1581 {
1582 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1583
1584 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1585 ASSERT(tx->tx_txg != 0);
1586 ASSERT(db->db_level == 0);
1587 ASSERT(!refcount_is_zero(&db->db_holds));
1588
1589 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1590 dmu_tx_private_ok(tx));
1591
1592 dbuf_noread(db);
1593 (void) dbuf_dirty(db, tx);
1594 }
1595
1596 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1597 /* ARGSUSED */
1598 void
1599 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1600 {
1601 mutex_enter(&db->db_mtx);
1602 DBUF_VERIFY(db);
1603
1604 if (db->db_state == DB_FILL) {
1605 if (db->db_level == 0 && db->db_freed_in_flight) {
1606 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1607 /* we were freed while filling */
1608 /* XXX dbuf_undirty? */
1609 bzero(db->db.db_data, db->db.db_size);
1610 db->db_freed_in_flight = FALSE;
1611 }
1612 db->db_state = DB_CACHED;
1613 cv_broadcast(&db->db_changed);
1614 }
1615 mutex_exit(&db->db_mtx);
1616 }
1617
1618 void
1619 dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data,
1620 bp_embedded_type_t etype, enum zio_compress comp,
1621 int uncompressed_size, int compressed_size, int byteorder,
1622 dmu_tx_t *tx)
1623 {
1624 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
1625 struct dirty_leaf *dl;
1626 dmu_object_type_t type;
1627
1628 if (etype == BP_EMBEDDED_TYPE_DATA) {
1629 ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset),
1630 SPA_FEATURE_EMBEDDED_DATA));
1631 }
1632
1633 DB_DNODE_ENTER(db);
1634 type = DB_DNODE(db)->dn_type;
1635 DB_DNODE_EXIT(db);
1636
1637 ASSERT0(db->db_level);
1638 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1639
1640 dmu_buf_will_not_fill(dbuf, tx);
1641
1642 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1643 dl = &db->db_last_dirty->dt.dl;
1644 encode_embedded_bp_compressed(&dl->dr_overridden_by,
1645 data, comp, uncompressed_size, compressed_size);
1646 BPE_SET_ETYPE(&dl->dr_overridden_by, etype);
1647 BP_SET_TYPE(&dl->dr_overridden_by, type);
1648 BP_SET_LEVEL(&dl->dr_overridden_by, 0);
1649 BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder);
1650
1651 dl->dr_override_state = DR_OVERRIDDEN;
1652 dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg;
1653 }
1654
1655 /*
1656 * Directly assign a provided arc buf to a given dbuf if it's not referenced
1657 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1658 */
1659 void
1660 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
1661 {
1662 ASSERT(!refcount_is_zero(&db->db_holds));
1663 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1664 ASSERT(db->db_level == 0);
1665 ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA);
1666 ASSERT(buf != NULL);
1667 ASSERT(arc_buf_size(buf) == db->db.db_size);
1668 ASSERT(tx->tx_txg != 0);
1669
1670 arc_return_buf(buf, db);
1671 ASSERT(arc_released(buf));
1672
1673 mutex_enter(&db->db_mtx);
1674
1675 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1676 cv_wait(&db->db_changed, &db->db_mtx);
1677
1678 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
1679
1680 if (db->db_state == DB_CACHED &&
1681 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
1682 mutex_exit(&db->db_mtx);
1683 (void) dbuf_dirty(db, tx);
1684 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
1685 VERIFY(arc_buf_remove_ref(buf, db));
1686 xuio_stat_wbuf_copied();
1687 return;
1688 }
1689
1690 xuio_stat_wbuf_nocopy();
1691 if (db->db_state == DB_CACHED) {
1692 dbuf_dirty_record_t *dr = db->db_last_dirty;
1693
1694 ASSERT(db->db_buf != NULL);
1695 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
1696 ASSERT(dr->dt.dl.dr_data == db->db_buf);
1697 if (!arc_released(db->db_buf)) {
1698 ASSERT(dr->dt.dl.dr_override_state ==
1699 DR_OVERRIDDEN);
1700 arc_release(db->db_buf, db);
1701 }
1702 dr->dt.dl.dr_data = buf;
1703 VERIFY(arc_buf_remove_ref(db->db_buf, db));
1704 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
1705 arc_release(db->db_buf, db);
1706 VERIFY(arc_buf_remove_ref(db->db_buf, db));
1707 }
1708 db->db_buf = NULL;
1709 }
1710 ASSERT(db->db_buf == NULL);
1711 dbuf_set_data(db, buf);
1712 db->db_state = DB_FILL;
1713 mutex_exit(&db->db_mtx);
1714 (void) dbuf_dirty(db, tx);
1715 dmu_buf_fill_done(&db->db, tx);
1716 }
1717
1718 /*
1719 * "Clear" the contents of this dbuf. This will mark the dbuf
1720 * EVICTING and clear *most* of its references. Unfortunately,
1721 * when we are not holding the dn_dbufs_mtx, we can't clear the
1722 * entry in the dn_dbufs list. We have to wait until dbuf_destroy()
1723 * in this case. For callers from the DMU we will usually see:
1724 * dbuf_clear()->arc_clear_callback()->dbuf_do_evict()->dbuf_destroy()
1725 * For the arc callback, we will usually see:
1726 * dbuf_do_evict()->dbuf_clear();dbuf_destroy()
1727 * Sometimes, though, we will get a mix of these two:
1728 * DMU: dbuf_clear()->arc_clear_callback()
1729 * ARC: dbuf_do_evict()->dbuf_destroy()
1730 *
1731 * This routine will dissociate the dbuf from the arc, by calling
1732 * arc_clear_callback(), but will not evict the data from the ARC.
1733 */
1734 void
1735 dbuf_clear(dmu_buf_impl_t *db)
1736 {
1737 dnode_t *dn;
1738 dmu_buf_impl_t *parent = db->db_parent;
1739 dmu_buf_impl_t *dndb;
1740 boolean_t dbuf_gone = B_FALSE;
1741
1742 ASSERT(MUTEX_HELD(&db->db_mtx));
1743 ASSERT(refcount_is_zero(&db->db_holds));
1744
1745 dbuf_evict_user(db);
1746
1747 if (db->db_state == DB_CACHED) {
1748 ASSERT(db->db.db_data != NULL);
1749 if (db->db_blkid == DMU_BONUS_BLKID) {
1750 zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
1751 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
1752 }
1753 db->db.db_data = NULL;
1754 db->db_state = DB_UNCACHED;
1755 }
1756
1757 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
1758 ASSERT(db->db_data_pending == NULL);
1759
1760 db->db_state = DB_EVICTING;
1761 db->db_blkptr = NULL;
1762
1763 DB_DNODE_ENTER(db);
1764 dn = DB_DNODE(db);
1765 dndb = dn->dn_dbuf;
1766 if (db->db_blkid != DMU_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) {
1767 avl_remove(&dn->dn_dbufs, db);
1768 atomic_dec_32(&dn->dn_dbufs_count);
1769 membar_producer();
1770 DB_DNODE_EXIT(db);
1771 /*
1772 * Decrementing the dbuf count means that the hold corresponding
1773 * to the removed dbuf is no longer discounted in dnode_move(),
1774 * so the dnode cannot be moved until after we release the hold.
1775 * The membar_producer() ensures visibility of the decremented
1776 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
1777 * release any lock.
1778 */
1779 dnode_rele(dn, db);
1780 db->db_dnode_handle = NULL;
1781 } else {
1782 DB_DNODE_EXIT(db);
1783 }
1784
1785 if (db->db_buf)
1786 dbuf_gone = arc_clear_callback(db->db_buf);
1787
1788 if (!dbuf_gone)
1789 mutex_exit(&db->db_mtx);
1790
1791 /*
1792 * If this dbuf is referenced from an indirect dbuf,
1793 * decrement the ref count on the indirect dbuf.
1794 */
1795 if (parent && parent != dndb)
1796 dbuf_rele(parent, db);
1797 }
1798
1799 /*
1800 * Note: While bpp will always be updated if the function returns success,
1801 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
1802 * this happens when the dnode is the meta-dnode, or a userused or groupused
1803 * object.
1804 */
1805 static int
1806 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
1807 dmu_buf_impl_t **parentp, blkptr_t **bpp)
1808 {
1809 int nlevels, epbs;
1810
1811 *parentp = NULL;
1812 *bpp = NULL;
1813
1814 ASSERT(blkid != DMU_BONUS_BLKID);
1815
1816 if (blkid == DMU_SPILL_BLKID) {
1817 mutex_enter(&dn->dn_mtx);
1818 if (dn->dn_have_spill &&
1819 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
1820 *bpp = &dn->dn_phys->dn_spill;
1821 else
1822 *bpp = NULL;
1823 dbuf_add_ref(dn->dn_dbuf, NULL);
1824 *parentp = dn->dn_dbuf;
1825 mutex_exit(&dn->dn_mtx);
1826 return (0);
1827 }
1828
1829 if (dn->dn_phys->dn_nlevels == 0)
1830 nlevels = 1;
1831 else
1832 nlevels = dn->dn_phys->dn_nlevels;
1833
1834 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1835
1836 ASSERT3U(level * epbs, <, 64);
1837 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1838 if (level >= nlevels ||
1839 (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
1840 /* the buffer has no parent yet */
1841 return (SET_ERROR(ENOENT));
1842 } else if (level < nlevels-1) {
1843 /* this block is referenced from an indirect block */
1844 int err = dbuf_hold_impl(dn, level+1,
1845 blkid >> epbs, fail_sparse, FALSE, NULL, parentp);
1846 if (err)
1847 return (err);
1848 err = dbuf_read(*parentp, NULL,
1849 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
1850 if (err) {
1851 dbuf_rele(*parentp, NULL);
1852 *parentp = NULL;
1853 return (err);
1854 }
1855 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
1856 (blkid & ((1ULL << epbs) - 1));
1857 return (0);
1858 } else {
1859 /* the block is referenced from the dnode */
1860 ASSERT3U(level, ==, nlevels-1);
1861 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
1862 blkid < dn->dn_phys->dn_nblkptr);
1863 if (dn->dn_dbuf) {
1864 dbuf_add_ref(dn->dn_dbuf, NULL);
1865 *parentp = dn->dn_dbuf;
1866 }
1867 *bpp = &dn->dn_phys->dn_blkptr[blkid];
1868 return (0);
1869 }
1870 }
1871
1872 static dmu_buf_impl_t *
1873 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
1874 dmu_buf_impl_t *parent, blkptr_t *blkptr)
1875 {
1876 objset_t *os = dn->dn_objset;
1877 dmu_buf_impl_t *db, *odb;
1878
1879 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
1880 ASSERT(dn->dn_type != DMU_OT_NONE);
1881
1882 db = kmem_cache_alloc(dbuf_cache, KM_SLEEP);
1883
1884 db->db_objset = os;
1885 db->db.db_object = dn->dn_object;
1886 db->db_level = level;
1887 db->db_blkid = blkid;
1888 db->db_last_dirty = NULL;
1889 db->db_dirtycnt = 0;
1890 db->db_dnode_handle = dn->dn_handle;
1891 db->db_parent = parent;
1892 db->db_blkptr = blkptr;
1893
1894 db->db_user = NULL;
1895 db->db_user_immediate_evict = FALSE;
1896 db->db_freed_in_flight = FALSE;
1897 db->db_pending_evict = FALSE;
1898
1899 if (blkid == DMU_BONUS_BLKID) {
1900 ASSERT3P(parent, ==, dn->dn_dbuf);
1901 db->db.db_size = DN_MAX_BONUSLEN -
1902 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
1903 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
1904 db->db.db_offset = DMU_BONUS_BLKID;
1905 db->db_state = DB_UNCACHED;
1906 /* the bonus dbuf is not placed in the hash table */
1907 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1908 return (db);
1909 } else if (blkid == DMU_SPILL_BLKID) {
1910 db->db.db_size = (blkptr != NULL) ?
1911 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
1912 db->db.db_offset = 0;
1913 } else {
1914 int blocksize =
1915 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
1916 db->db.db_size = blocksize;
1917 db->db.db_offset = db->db_blkid * blocksize;
1918 }
1919
1920 /*
1921 * Hold the dn_dbufs_mtx while we get the new dbuf
1922 * in the hash table *and* added to the dbufs list.
1923 * This prevents a possible deadlock with someone
1924 * trying to look up this dbuf before its added to the
1925 * dn_dbufs list.
1926 */
1927 mutex_enter(&dn->dn_dbufs_mtx);
1928 db->db_state = DB_EVICTING;
1929 if ((odb = dbuf_hash_insert(db)) != NULL) {
1930 /* someone else inserted it first */
1931 kmem_cache_free(dbuf_cache, db);
1932 mutex_exit(&dn->dn_dbufs_mtx);
1933 return (odb);
1934 }
1935 avl_add(&dn->dn_dbufs, db);
1936 if (db->db_level == 0 && db->db_blkid >=
1937 dn->dn_unlisted_l0_blkid)
1938 dn->dn_unlisted_l0_blkid = db->db_blkid + 1;
1939 db->db_state = DB_UNCACHED;
1940 mutex_exit(&dn->dn_dbufs_mtx);
1941 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
1942
1943 if (parent && parent != dn->dn_dbuf)
1944 dbuf_add_ref(parent, db);
1945
1946 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1947 refcount_count(&dn->dn_holds) > 0);
1948 (void) refcount_add(&dn->dn_holds, db);
1949 atomic_inc_32(&dn->dn_dbufs_count);
1950
1951 dprintf_dbuf(db, "db=%p\n", db);
1952
1953 return (db);
1954 }
1955
1956 static int
1957 dbuf_do_evict(void *private)
1958 {
1959 dmu_buf_impl_t *db = private;
1960
1961 if (!MUTEX_HELD(&db->db_mtx))
1962 mutex_enter(&db->db_mtx);
1963
1964 ASSERT(refcount_is_zero(&db->db_holds));
1965
1966 if (db->db_state != DB_EVICTING) {
1967 ASSERT(db->db_state == DB_CACHED);
1968 DBUF_VERIFY(db);
1969 db->db_buf = NULL;
1970 dbuf_evict(db);
1971 } else {
1972 mutex_exit(&db->db_mtx);
1973 dbuf_destroy(db);
1974 }
1975 return (0);
1976 }
1977
1978 static void
1979 dbuf_destroy(dmu_buf_impl_t *db)
1980 {
1981 ASSERT(refcount_is_zero(&db->db_holds));
1982
1983 if (db->db_blkid != DMU_BONUS_BLKID) {
1984 /*
1985 * If this dbuf is still on the dn_dbufs list,
1986 * remove it from that list.
1987 */
1988 if (db->db_dnode_handle != NULL) {
1989 dnode_t *dn;
1990
1991 DB_DNODE_ENTER(db);
1992 dn = DB_DNODE(db);
1993 mutex_enter(&dn->dn_dbufs_mtx);
1994 avl_remove(&dn->dn_dbufs, db);
1995 atomic_dec_32(&dn->dn_dbufs_count);
1996 mutex_exit(&dn->dn_dbufs_mtx);
1997 DB_DNODE_EXIT(db);
1998 /*
1999 * Decrementing the dbuf count means that the hold
2000 * corresponding to the removed dbuf is no longer
2001 * discounted in dnode_move(), so the dnode cannot be
2002 * moved until after we release the hold.
2003 */
2004 dnode_rele(dn, db);
2005 db->db_dnode_handle = NULL;
2006 }
2007 dbuf_hash_remove(db);
2008 }
2009 db->db_parent = NULL;
2010 db->db_buf = NULL;
2011
2012 ASSERT(db->db.db_data == NULL);
2013 ASSERT(db->db_hash_next == NULL);
2014 ASSERT(db->db_blkptr == NULL);
2015 ASSERT(db->db_data_pending == NULL);
2016
2017 kmem_cache_free(dbuf_cache, db);
2018 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2019 }
2020
2021 typedef struct dbuf_prefetch_arg {
2022 spa_t *dpa_spa; /* The spa to issue the prefetch in. */
2023 zbookmark_phys_t dpa_zb; /* The target block to prefetch. */
2024 int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */
2025 int dpa_curlevel; /* The current level that we're reading */
2026 zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */
2027 zio_t *dpa_zio; /* The parent zio_t for all prefetches. */
2028 arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */
2029 } dbuf_prefetch_arg_t;
2030
2031 /*
2032 * Actually issue the prefetch read for the block given.
2033 */
2034 static void
2035 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp)
2036 {
2037 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
2038 return;
2039
2040 arc_flags_t aflags =
2041 dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
2042
2043 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2044 ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level);
2045 ASSERT(dpa->dpa_zio != NULL);
2046 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL,
2047 dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2048 &aflags, &dpa->dpa_zb);
2049 }
2050
2051 /*
2052 * Called when an indirect block above our prefetch target is read in. This
2053 * will either read in the next indirect block down the tree or issue the actual
2054 * prefetch if the next block down is our target.
2055 */
2056 static void
2057 dbuf_prefetch_indirect_done(zio_t *zio, arc_buf_t *abuf, void *private)
2058 {
2059 dbuf_prefetch_arg_t *dpa = private;
2060
2061 ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel);
2062 ASSERT3S(dpa->dpa_curlevel, >, 0);
2063 if (zio != NULL) {
2064 ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel);
2065 ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size);
2066 ASSERT3P(zio->io_spa, ==, dpa->dpa_spa);
2067 }
2068
2069 dpa->dpa_curlevel--;
2070
2071 uint64_t nextblkid = dpa->dpa_zb.zb_blkid >>
2072 (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level));
2073 blkptr_t *bp = ((blkptr_t *)abuf->b_data) +
2074 P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs);
2075 if (BP_IS_HOLE(bp) || (zio != NULL && zio->io_error != 0)) {
2076 kmem_free(dpa, sizeof (*dpa));
2077 } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) {
2078 ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid);
2079 dbuf_issue_final_prefetch(dpa, bp);
2080 kmem_free(dpa, sizeof (*dpa));
2081 } else {
2082 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2083 zbookmark_phys_t zb;
2084
2085 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2086
2087 SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset,
2088 dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid);
2089
2090 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2091 bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio,
2092 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2093 &iter_aflags, &zb);
2094 }
2095 (void) arc_buf_remove_ref(abuf, private);
2096 }
2097
2098 /*
2099 * Issue prefetch reads for the given block on the given level. If the indirect
2100 * blocks above that block are not in memory, we will read them in
2101 * asynchronously. As a result, this call never blocks waiting for a read to
2102 * complete.
2103 */
2104 void
2105 dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio,
2106 arc_flags_t aflags)
2107 {
2108 blkptr_t bp;
2109 int epbs, nlevels, curlevel;
2110 uint64_t curblkid;
2111
2112 ASSERT(blkid != DMU_BONUS_BLKID);
2113 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2114
2115 if (blkid > dn->dn_maxblkid)
2116 return;
2117
2118 if (dnode_block_freed(dn, blkid))
2119 return;
2120
2121 /*
2122 * This dnode hasn't been written to disk yet, so there's nothing to
2123 * prefetch.
2124 */
2125 nlevels = dn->dn_phys->dn_nlevels;
2126 if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0)
2127 return;
2128
2129 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2130 if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level))
2131 return;
2132
2133 dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object,
2134 level, blkid);
2135 if (db != NULL) {
2136 mutex_exit(&db->db_mtx);
2137 /*
2138 * This dbuf already exists. It is either CACHED, or
2139 * (we assume) about to be read or filled.
2140 */
2141 return;
2142 }
2143
2144 /*
2145 * Find the closest ancestor (indirect block) of the target block
2146 * that is present in the cache. In this indirect block, we will
2147 * find the bp that is at curlevel, curblkid.
2148 */
2149 curlevel = level;
2150 curblkid = blkid;
2151 while (curlevel < nlevels - 1) {
2152 int parent_level = curlevel + 1;
2153 uint64_t parent_blkid = curblkid >> epbs;
2154 dmu_buf_impl_t *db;
2155
2156 if (dbuf_hold_impl(dn, parent_level, parent_blkid,
2157 FALSE, TRUE, FTAG, &db) == 0) {
2158 blkptr_t *bpp = db->db_buf->b_data;
2159 bp = bpp[P2PHASE(curblkid, 1 << epbs)];
2160 dbuf_rele(db, FTAG);
2161 break;
2162 }
2163
2164 curlevel = parent_level;
2165 curblkid = parent_blkid;
2166 }
2167
2168 if (curlevel == nlevels - 1) {
2169 /* No cached indirect blocks found. */
2170 ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr);
2171 bp = dn->dn_phys->dn_blkptr[curblkid];
2172 }
2173 if (BP_IS_HOLE(&bp))
2174 return;
2175
2176 ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp));
2177
2178 zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL,
2179 ZIO_FLAG_CANFAIL);
2180
2181 dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP);
2182 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
2183 SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2184 dn->dn_object, level, blkid);
2185 dpa->dpa_curlevel = curlevel;
2186 dpa->dpa_prio = prio;
2187 dpa->dpa_aflags = aflags;
2188 dpa->dpa_spa = dn->dn_objset->os_spa;
2189 dpa->dpa_epbs = epbs;
2190 dpa->dpa_zio = pio;
2191
2192 /*
2193 * If we have the indirect just above us, no need to do the asynchronous
2194 * prefetch chain; we'll just run the last step ourselves. If we're at
2195 * a higher level, though, we want to issue the prefetches for all the
2196 * indirect blocks asynchronously, so we can go on with whatever we were
2197 * doing.
2198 */
2199 if (curlevel == level) {
2200 ASSERT3U(curblkid, ==, blkid);
2201 dbuf_issue_final_prefetch(dpa, &bp);
2202 kmem_free(dpa, sizeof (*dpa));
2203 } else {
2204 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2205 zbookmark_phys_t zb;
2206
2207 SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2208 dn->dn_object, curlevel, curblkid);
2209 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2210 &bp, dbuf_prefetch_indirect_done, dpa, prio,
2211 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2212 &iter_aflags, &zb);
2213 }
2214 /*
2215 * We use pio here instead of dpa_zio since it's possible that
2216 * dpa may have already been freed.
2217 */
2218 zio_nowait(pio);
2219 }
2220
2221 /*
2222 * Returns with db_holds incremented, and db_mtx not held.
2223 * Note: dn_struct_rwlock must be held.
2224 */
2225 int
2226 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid,
2227 boolean_t fail_sparse, boolean_t fail_uncached,
2228 void *tag, dmu_buf_impl_t **dbp)
2229 {
2230 dmu_buf_impl_t *db, *parent = NULL;
2231
2232 ASSERT(blkid != DMU_BONUS_BLKID);
2233 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2234 ASSERT3U(dn->dn_nlevels, >, level);
2235
2236 *dbp = NULL;
2237 top:
2238 /* dbuf_find() returns with db_mtx held */
2239 db = dbuf_find(dn->dn_objset, dn->dn_object, level, blkid);
2240
2241 if (db == NULL) {
2242 blkptr_t *bp = NULL;
2243 int err;
2244
2245 if (fail_uncached)
2246 return (SET_ERROR(ENOENT));
2247
2248 ASSERT3P(parent, ==, NULL);
2249 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
2250 if (fail_sparse) {
2251 if (err == 0 && bp && BP_IS_HOLE(bp))
2252 err = SET_ERROR(ENOENT);
2253 if (err) {
2254 if (parent)
2255 dbuf_rele(parent, NULL);
2256 return (err);
2257 }
2258 }
2259 if (err && err != ENOENT)
2260 return (err);
2261 db = dbuf_create(dn, level, blkid, parent, bp);
2262 }
2263
2264 if (fail_uncached && db->db_state != DB_CACHED) {
2265 mutex_exit(&db->db_mtx);
2266 return (SET_ERROR(ENOENT));
2267 }
2268
2269 if (db->db_buf && refcount_is_zero(&db->db_holds)) {
2270 arc_buf_add_ref(db->db_buf, db);
2271 if (db->db_buf->b_data == NULL) {
2272 dbuf_clear(db);
2273 if (parent) {
2274 dbuf_rele(parent, NULL);
2275 parent = NULL;
2276 }
2277 goto top;
2278 }
2279 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
2280 }
2281
2282 ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
2283
2284 /*
2285 * If this buffer is currently syncing out, and we are are
2286 * still referencing it from db_data, we need to make a copy
2287 * of it in case we decide we want to dirty it again in this txg.
2288 */
2289 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
2290 dn->dn_object != DMU_META_DNODE_OBJECT &&
2291 db->db_state == DB_CACHED && db->db_data_pending) {
2292 dbuf_dirty_record_t *dr = db->db_data_pending;
2293
2294 if (dr->dt.dl.dr_data == db->db_buf) {
2295 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2296
2297 dbuf_set_data(db,
2298 arc_buf_alloc(dn->dn_objset->os_spa,
2299 db->db.db_size, db, type));
2300 bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
2301 db->db.db_size);
2302 }
2303 }
2304
2305 (void) refcount_add(&db->db_holds, tag);
2306 DBUF_VERIFY(db);
2307 mutex_exit(&db->db_mtx);
2308
2309 /* NOTE: we can't rele the parent until after we drop the db_mtx */
2310 if (parent)
2311 dbuf_rele(parent, NULL);
2312
2313 ASSERT3P(DB_DNODE(db), ==, dn);
2314 ASSERT3U(db->db_blkid, ==, blkid);
2315 ASSERT3U(db->db_level, ==, level);
2316 *dbp = db;
2317
2318 return (0);
2319 }
2320
2321 dmu_buf_impl_t *
2322 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
2323 {
2324 return (dbuf_hold_level(dn, 0, blkid, tag));
2325 }
2326
2327 dmu_buf_impl_t *
2328 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
2329 {
2330 dmu_buf_impl_t *db;
2331 int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db);
2332 return (err ? NULL : db);
2333 }
2334
2335 void
2336 dbuf_create_bonus(dnode_t *dn)
2337 {
2338 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
2339
2340 ASSERT(dn->dn_bonus == NULL);
2341 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
2342 }
2343
2344 int
2345 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
2346 {
2347 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2348 dnode_t *dn;
2349
2350 if (db->db_blkid != DMU_SPILL_BLKID)
2351 return (SET_ERROR(ENOTSUP));
2352 if (blksz == 0)
2353 blksz = SPA_MINBLOCKSIZE;
2354 ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset)));
2355 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
2356
2357 DB_DNODE_ENTER(db);
2358 dn = DB_DNODE(db);
2359 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2360 dbuf_new_size(db, blksz, tx);
2361 rw_exit(&dn->dn_struct_rwlock);
2362 DB_DNODE_EXIT(db);
2363
2364 return (0);
2365 }
2366
2367 void
2368 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
2369 {
2370 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
2371 }
2372
2373 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2374 void
2375 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
2376 {
2377 int64_t holds = refcount_add(&db->db_holds, tag);
2378 ASSERT(holds > 1);
2379 }
2380
2381 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
2382 boolean_t
2383 dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid,
2384 void *tag)
2385 {
2386 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2387 dmu_buf_impl_t *found_db;
2388 boolean_t result = B_FALSE;
2389
2390 if (db->db_blkid == DMU_BONUS_BLKID)
2391 found_db = dbuf_find_bonus(os, obj);
2392 else
2393 found_db = dbuf_find(os, obj, 0, blkid);
2394
2395 if (found_db != NULL) {
2396 if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) {
2397 (void) refcount_add(&db->db_holds, tag);
2398 result = B_TRUE;
2399 }
2400 mutex_exit(&db->db_mtx);
2401 }
2402 return (result);
2403 }
2404
2405 /*
2406 * If you call dbuf_rele() you had better not be referencing the dnode handle
2407 * unless you have some other direct or indirect hold on the dnode. (An indirect
2408 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2409 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2410 * dnode's parent dbuf evicting its dnode handles.
2411 */
2412 void
2413 dbuf_rele(dmu_buf_impl_t *db, void *tag)
2414 {
2415 mutex_enter(&db->db_mtx);
2416 dbuf_rele_and_unlock(db, tag);
2417 }
2418
2419 void
2420 dmu_buf_rele(dmu_buf_t *db, void *tag)
2421 {
2422 dbuf_rele((dmu_buf_impl_t *)db, tag);
2423 }
2424
2425 /*
2426 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2427 * db_dirtycnt and db_holds to be updated atomically.
2428 */
2429 void
2430 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2431 {
2432 int64_t holds;
2433
2434 ASSERT(MUTEX_HELD(&db->db_mtx));
2435 DBUF_VERIFY(db);
2436
2437 /*
2438 * Remove the reference to the dbuf before removing its hold on the
2439 * dnode so we can guarantee in dnode_move() that a referenced bonus
2440 * buffer has a corresponding dnode hold.
2441 */
2442 holds = refcount_remove(&db->db_holds, tag);
2443 ASSERT(holds >= 0);
2444
2445 /*
2446 * We can't freeze indirects if there is a possibility that they
2447 * may be modified in the current syncing context.
2448 */
2449 if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0))
2450 arc_buf_freeze(db->db_buf);
2451
2452 if (holds == db->db_dirtycnt &&
2453 db->db_level == 0 && db->db_user_immediate_evict)
2454 dbuf_evict_user(db);
2455
2456 if (holds == 0) {
2457 if (db->db_blkid == DMU_BONUS_BLKID) {
2458 dnode_t *dn;
2459 boolean_t evict_dbuf = db->db_pending_evict;
2460
2461 /*
2462 * If the dnode moves here, we cannot cross this
2463 * barrier until the move completes.
2464 */
2465 DB_DNODE_ENTER(db);
2466
2467 dn = DB_DNODE(db);
2468 atomic_dec_32(&dn->dn_dbufs_count);
2469
2470 /*
2471 * Decrementing the dbuf count means that the bonus
2472 * buffer's dnode hold is no longer discounted in
2473 * dnode_move(). The dnode cannot move until after
2474 * the dnode_rele() below.
2475 */
2476 DB_DNODE_EXIT(db);
2477
2478 /*
2479 * Do not reference db after its lock is dropped.
2480 * Another thread may evict it.
2481 */
2482 mutex_exit(&db->db_mtx);
2483
2484 if (evict_dbuf)
2485 dnode_evict_bonus(dn);
2486
2487 dnode_rele(dn, db);
2488 } else if (db->db_buf == NULL) {
2489 /*
2490 * This is a special case: we never associated this
2491 * dbuf with any data allocated from the ARC.
2492 */
2493 ASSERT(db->db_state == DB_UNCACHED ||
2494 db->db_state == DB_NOFILL);
2495 dbuf_evict(db);
2496 } else if (arc_released(db->db_buf)) {
2497 arc_buf_t *buf = db->db_buf;
2498 /*
2499 * This dbuf has anonymous data associated with it.
2500 */
2501 dbuf_clear_data(db);
2502 VERIFY(arc_buf_remove_ref(buf, db));
2503 dbuf_evict(db);
2504 } else {
2505 VERIFY(!arc_buf_remove_ref(db->db_buf, db));
2506
2507 /*
2508 * A dbuf will be eligible for eviction if either the
2509 * 'primarycache' property is set or a duplicate
2510 * copy of this buffer is already cached in the arc.
2511 *
2512 * In the case of the 'primarycache' a buffer
2513 * is considered for eviction if it matches the
2514 * criteria set in the property.
2515 *
2516 * To decide if our buffer is considered a
2517 * duplicate, we must call into the arc to determine
2518 * if multiple buffers are referencing the same
2519 * block on-disk. If so, then we simply evict
2520 * ourselves.
2521 */
2522 if (!DBUF_IS_CACHEABLE(db)) {
2523 if (db->db_blkptr != NULL &&
2524 !BP_IS_HOLE(db->db_blkptr) &&
2525 !BP_IS_EMBEDDED(db->db_blkptr)) {
2526 spa_t *spa =
2527 dmu_objset_spa(db->db_objset);
2528 blkptr_t bp = *db->db_blkptr;
2529 dbuf_clear(db);
2530 arc_freed(spa, &bp);
2531 } else {
2532 dbuf_clear(db);
2533 }
2534 } else if (db->db_pending_evict ||
2535 arc_buf_eviction_needed(db->db_buf)) {
2536 dbuf_clear(db);
2537 } else {
2538 mutex_exit(&db->db_mtx);
2539 }
2540 }
2541 } else {
2542 mutex_exit(&db->db_mtx);
2543 }
2544 }
2545
2546 #pragma weak dmu_buf_refcount = dbuf_refcount
2547 uint64_t
2548 dbuf_refcount(dmu_buf_impl_t *db)
2549 {
2550 return (refcount_count(&db->db_holds));
2551 }
2552
2553 void *
2554 dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user,
2555 dmu_buf_user_t *new_user)
2556 {
2557 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2558
2559 mutex_enter(&db->db_mtx);
2560 dbuf_verify_user(db, DBVU_NOT_EVICTING);
2561 if (db->db_user == old_user)
2562 db->db_user = new_user;
2563 else
2564 old_user = db->db_user;
2565 dbuf_verify_user(db, DBVU_NOT_EVICTING);
2566 mutex_exit(&db->db_mtx);
2567
2568 return (old_user);
2569 }
2570
2571 void *
2572 dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2573 {
2574 return (dmu_buf_replace_user(db_fake, NULL, user));
2575 }
2576
2577 void *
2578 dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2579 {
2580 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2581
2582 db->db_user_immediate_evict = TRUE;
2583 return (dmu_buf_set_user(db_fake, user));
2584 }
2585
2586 void *
2587 dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2588 {
2589 return (dmu_buf_replace_user(db_fake, user, NULL));
2590 }
2591
2592 void *
2593 dmu_buf_get_user(dmu_buf_t *db_fake)
2594 {
2595 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2596
2597 dbuf_verify_user(db, DBVU_NOT_EVICTING);
2598 return (db->db_user);
2599 }
2600
2601 void
2602 dmu_buf_user_evict_wait()
2603 {
2604 taskq_wait(dbu_evict_taskq);
2605 }
2606
2607 boolean_t
2608 dmu_buf_freeable(dmu_buf_t *dbuf)
2609 {
2610 boolean_t res = B_FALSE;
2611 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2612
2613 if (db->db_blkptr)
2614 res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset,
2615 db->db_blkptr, db->db_blkptr->blk_birth);
2616
2617 return (res);
2618 }
2619
2620 blkptr_t *
2621 dmu_buf_get_blkptr(dmu_buf_t *db)
2622 {
2623 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2624 return (dbi->db_blkptr);
2625 }
2626
2627 static void
2628 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
2629 {
2630 /* ASSERT(dmu_tx_is_syncing(tx) */
2631 ASSERT(MUTEX_HELD(&db->db_mtx));
2632
2633 if (db->db_blkptr != NULL)
2634 return;
2635
2636 if (db->db_blkid == DMU_SPILL_BLKID) {
2637 db->db_blkptr = &dn->dn_phys->dn_spill;
2638 BP_ZERO(db->db_blkptr);
2639 return;
2640 }
2641 if (db->db_level == dn->dn_phys->dn_nlevels-1) {
2642 /*
2643 * This buffer was allocated at a time when there was
2644 * no available blkptrs from the dnode, or it was
2645 * inappropriate to hook it in (i.e., nlevels mis-match).
2646 */
2647 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
2648 ASSERT(db->db_parent == NULL);
2649 db->db_parent = dn->dn_dbuf;
2650 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
2651 DBUF_VERIFY(db);
2652 } else {
2653 dmu_buf_impl_t *parent = db->db_parent;
2654 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2655
2656 ASSERT(dn->dn_phys->dn_nlevels > 1);
2657 if (parent == NULL) {
2658 mutex_exit(&db->db_mtx);
2659 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2660 parent = dbuf_hold_level(dn, db->db_level + 1,
2661 db->db_blkid >> epbs, db);
2662 rw_exit(&dn->dn_struct_rwlock);
2663 mutex_enter(&db->db_mtx);
2664 db->db_parent = parent;
2665 }
2666 db->db_blkptr = (blkptr_t *)parent->db.db_data +
2667 (db->db_blkid & ((1ULL << epbs) - 1));
2668 DBUF_VERIFY(db);
2669 }
2670 }
2671
2672 static void
2673 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2674 {
2675 dmu_buf_impl_t *db = dr->dr_dbuf;
2676 dnode_t *dn;
2677 zio_t *zio;
2678
2679 ASSERT(dmu_tx_is_syncing(tx));
2680
2681 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2682
2683 mutex_enter(&db->db_mtx);
2684
2685 ASSERT(db->db_level > 0);
2686 DBUF_VERIFY(db);
2687
2688 /* Read the block if it hasn't been read yet. */
2689 if (db->db_buf == NULL) {
2690 mutex_exit(&db->db_mtx);
2691 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
2692 mutex_enter(&db->db_mtx);
2693 }
2694 ASSERT3U(db->db_state, ==, DB_CACHED);
2695 ASSERT(db->db_buf != NULL);
2696
2697 DB_DNODE_ENTER(db);
2698 dn = DB_DNODE(db);
2699 /* Indirect block size must match what the dnode thinks it is. */
2700 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2701 dbuf_check_blkptr(dn, db);
2702 DB_DNODE_EXIT(db);
2703
2704 /* Provide the pending dirty record to child dbufs */
2705 db->db_data_pending = dr;
2706
2707 mutex_exit(&db->db_mtx);
2708 dbuf_write(dr, db->db_buf, tx);
2709
2710 zio = dr->dr_zio;
2711 mutex_enter(&dr->dt.di.dr_mtx);
2712 dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx);
2713 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
2714 mutex_exit(&dr->dt.di.dr_mtx);
2715 zio_nowait(zio);
2716 }
2717
2718 static void
2719 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2720 {
2721 arc_buf_t **datap = &dr->dt.dl.dr_data;
2722 dmu_buf_impl_t *db = dr->dr_dbuf;
2723 dnode_t *dn;
2724 objset_t *os;
2725 uint64_t txg = tx->tx_txg;
2726
2727 ASSERT(dmu_tx_is_syncing(tx));
2728
2729 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2730
2731 mutex_enter(&db->db_mtx);
2732 /*
2733 * To be synced, we must be dirtied. But we
2734 * might have been freed after the dirty.
2735 */
2736 if (db->db_state == DB_UNCACHED) {
2737 /* This buffer has been freed since it was dirtied */
2738 ASSERT(db->db.db_data == NULL);
2739 } else if (db->db_state == DB_FILL) {
2740 /* This buffer was freed and is now being re-filled */
2741 ASSERT(db->db.db_data != dr->dt.dl.dr_data);
2742 } else {
2743 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
2744 }
2745 DBUF_VERIFY(db);
2746
2747 DB_DNODE_ENTER(db);
2748 dn = DB_DNODE(db);
2749
2750 if (db->db_blkid == DMU_SPILL_BLKID) {
2751 mutex_enter(&dn->dn_mtx);
2752 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
2753 mutex_exit(&dn->dn_mtx);
2754 }
2755
2756 /*
2757 * If this is a bonus buffer, simply copy the bonus data into the
2758 * dnode. It will be written out when the dnode is synced (and it
2759 * will be synced, since it must have been dirty for dbuf_sync to
2760 * be called).
2761 */
2762 if (db->db_blkid == DMU_BONUS_BLKID) {
2763 dbuf_dirty_record_t **drp;
2764
2765 ASSERT(*datap != NULL);
2766 ASSERT0(db->db_level);
2767 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN);
2768 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen);
2769 DB_DNODE_EXIT(db);
2770
2771 if (*datap != db->db.db_data) {
2772 zio_buf_free(*datap, DN_MAX_BONUSLEN);
2773 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
2774 }
2775 db->db_data_pending = NULL;
2776 drp = &db->db_last_dirty;
2777 while (*drp != dr)
2778 drp = &(*drp)->dr_next;
2779 ASSERT(dr->dr_next == NULL);
2780 ASSERT(dr->dr_dbuf == db);
2781 *drp = dr->dr_next;
2782 kmem_free(dr, sizeof (dbuf_dirty_record_t));
2783 ASSERT(db->db_dirtycnt > 0);
2784 db->db_dirtycnt -= 1;
2785 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
2786 return;
2787 }
2788
2789 os = dn->dn_objset;
2790
2791 /*
2792 * This function may have dropped the db_mtx lock allowing a dmu_sync
2793 * operation to sneak in. As a result, we need to ensure that we
2794 * don't check the dr_override_state until we have returned from
2795 * dbuf_check_blkptr.
2796 */
2797 dbuf_check_blkptr(dn, db);
2798
2799 /*
2800 * If this buffer is in the middle of an immediate write,
2801 * wait for the synchronous IO to complete.
2802 */
2803 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
2804 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
2805 cv_wait(&db->db_changed, &db->db_mtx);
2806 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
2807 }
2808
2809 if (db->db_state != DB_NOFILL &&
2810 dn->dn_object != DMU_META_DNODE_OBJECT &&
2811 refcount_count(&db->db_holds) > 1 &&
2812 dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
2813 *datap == db->db_buf) {
2814 /*
2815 * If this buffer is currently "in use" (i.e., there
2816 * are active holds and db_data still references it),
2817 * then make a copy before we start the write so that
2818 * any modifications from the open txg will not leak
2819 * into this write.
2820 *
2821 * NOTE: this copy does not need to be made for
2822 * objects only modified in the syncing context (e.g.
2823 * DNONE_DNODE blocks).
2824 */
2825 int blksz = arc_buf_size(*datap);
2826 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2827 *datap = arc_buf_alloc(os->os_spa, blksz, db, type);
2828 bcopy(db->db.db_data, (*datap)->b_data, blksz);
2829 }
2830 db->db_data_pending = dr;
2831
2832 mutex_exit(&db->db_mtx);
2833
2834 dbuf_write(dr, *datap, tx);
2835
2836 ASSERT(!list_link_active(&dr->dr_dirty_node));
2837 if (dn->dn_object == DMU_META_DNODE_OBJECT) {
2838 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
2839 DB_DNODE_EXIT(db);
2840 } else {
2841 /*
2842 * Although zio_nowait() does not "wait for an IO", it does
2843 * initiate the IO. If this is an empty write it seems plausible
2844 * that the IO could actually be completed before the nowait
2845 * returns. We need to DB_DNODE_EXIT() first in case
2846 * zio_nowait() invalidates the dbuf.
2847 */
2848 DB_DNODE_EXIT(db);
2849 zio_nowait(dr->dr_zio);
2850 }
2851 }
2852
2853 void
2854 dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx)
2855 {
2856 dbuf_dirty_record_t *dr;
2857
2858 while (dr = list_head(list)) {
2859 if (dr->dr_zio != NULL) {
2860 /*
2861 * If we find an already initialized zio then we
2862 * are processing the meta-dnode, and we have finished.
2863 * The dbufs for all dnodes are put back on the list
2864 * during processing, so that we can zio_wait()
2865 * these IOs after initiating all child IOs.
2866 */
2867 ASSERT3U(dr->dr_dbuf->db.db_object, ==,
2868 DMU_META_DNODE_OBJECT);
2869 break;
2870 }
2871 if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
2872 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
2873 VERIFY3U(dr->dr_dbuf->db_level, ==, level);
2874 }
2875 list_remove(list, dr);
2876 if (dr->dr_dbuf->db_level > 0)
2877 dbuf_sync_indirect(dr, tx);
2878 else
2879 dbuf_sync_leaf(dr, tx);
2880 }
2881 }
2882
2883 /* ARGSUSED */
2884 static void
2885 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
2886 {
2887 dmu_buf_impl_t *db = vdb;
2888 dnode_t *dn;
2889 blkptr_t *bp = zio->io_bp;
2890 blkptr_t *bp_orig = &zio->io_bp_orig;
2891 spa_t *spa = zio->io_spa;
2892 int64_t delta;
2893 uint64_t fill = 0;
2894 int i;
2895
2896 ASSERT3P(db->db_blkptr, ==, bp);
2897
2898 DB_DNODE_ENTER(db);
2899 dn = DB_DNODE(db);
2900 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
2901 dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
2902 zio->io_prev_space_delta = delta;
2903
2904 if (bp->blk_birth != 0) {
2905 ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
2906 BP_GET_TYPE(bp) == dn->dn_type) ||
2907 (db->db_blkid == DMU_SPILL_BLKID &&
2908 BP_GET_TYPE(bp) == dn->dn_bonustype) ||
2909 BP_IS_EMBEDDED(bp));
2910 ASSERT(BP_GET_LEVEL(bp) == db->db_level);
2911 }
2912
2913 mutex_enter(&db->db_mtx);
2914
2915 #ifdef ZFS_DEBUG
2916 if (db->db_blkid == DMU_SPILL_BLKID) {
2917 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
2918 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
2919 db->db_blkptr == &dn->dn_phys->dn_spill);
2920 }
2921 #endif
2922
2923 if (db->db_level == 0) {
2924 mutex_enter(&dn->dn_mtx);
2925 if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
2926 db->db_blkid != DMU_SPILL_BLKID)
2927 dn->dn_phys->dn_maxblkid = db->db_blkid;
2928 mutex_exit(&dn->dn_mtx);
2929
2930 if (dn->dn_type == DMU_OT_DNODE) {
2931 dnode_phys_t *dnp = db->db.db_data;
2932 for (i = db->db.db_size >> DNODE_SHIFT; i > 0;
2933 i--, dnp++) {
2934 if (dnp->dn_type != DMU_OT_NONE)
2935 fill++;
2936 }
2937 } else {
2938 if (BP_IS_HOLE(bp)) {
2939 fill = 0;
2940 } else {
2941 fill = 1;
2942 }
2943 }
2944 } else {
2945 blkptr_t *ibp = db->db.db_data;
2946 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
2947 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
2948 if (BP_IS_HOLE(ibp))
2949 continue;
2950 fill += BP_GET_FILL(ibp);
2951 }
2952 }
2953 DB_DNODE_EXIT(db);
2954
2955 if (!BP_IS_EMBEDDED(bp))
2956 bp->blk_fill = fill;
2957
2958 mutex_exit(&db->db_mtx);
2959 }
2960
2961 /*
2962 * The SPA will call this callback several times for each zio - once
2963 * for every physical child i/o (zio->io_phys_children times). This
2964 * allows the DMU to monitor the progress of each logical i/o. For example,
2965 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
2966 * block. There may be a long delay before all copies/fragments are completed,
2967 * so this callback allows us to retire dirty space gradually, as the physical
2968 * i/os complete.
2969 */
2970 /* ARGSUSED */
2971 static void
2972 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
2973 {
2974 dmu_buf_impl_t *db = arg;
2975 objset_t *os = db->db_objset;
2976 dsl_pool_t *dp = dmu_objset_pool(os);
2977 dbuf_dirty_record_t *dr;
2978 int delta = 0;
2979
2980 dr = db->db_data_pending;
2981 ASSERT3U(dr->dr_txg, ==, zio->io_txg);
2982
2983 /*
2984 * The callback will be called io_phys_children times. Retire one
2985 * portion of our dirty space each time we are called. Any rounding
2986 * error will be cleaned up by dsl_pool_sync()'s call to
2987 * dsl_pool_undirty_space().
2988 */
2989 delta = dr->dr_accounted / zio->io_phys_children;
2990 dsl_pool_undirty_space(dp, delta, zio->io_txg);
2991 }
2992
2993 /* ARGSUSED */
2994 static void
2995 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
2996 {
2997 dmu_buf_impl_t *db = vdb;
2998 blkptr_t *bp_orig = &zio->io_bp_orig;
2999 blkptr_t *bp = db->db_blkptr;
3000 objset_t *os = db->db_objset;
3001 dmu_tx_t *tx = os->os_synctx;
3002 dbuf_dirty_record_t **drp, *dr;
3003
3004 ASSERT0(zio->io_error);
3005 ASSERT(db->db_blkptr == bp);
3006
3007 /*
3008 * For nopwrites and rewrites we ensure that the bp matches our
3009 * original and bypass all the accounting.
3010 */
3011 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
3012 ASSERT(BP_EQUAL(bp, bp_orig));
3013 } else {
3014 dsl_dataset_t *ds = os->os_dsl_dataset;
3015 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
3016 dsl_dataset_block_born(ds, bp, tx);
3017 }
3018
3019 mutex_enter(&db->db_mtx);
3020
3021 DBUF_VERIFY(db);
3022
3023 drp = &db->db_last_dirty;
3024 while ((dr = *drp) != db->db_data_pending)
3025 drp = &dr->dr_next;
3026 ASSERT(!list_link_active(&dr->dr_dirty_node));
3027 ASSERT(dr->dr_dbuf == db);
3028 ASSERT(dr->dr_next == NULL);
3029 *drp = dr->dr_next;
3030
3031 #ifdef ZFS_DEBUG
3032 if (db->db_blkid == DMU_SPILL_BLKID) {
3033 dnode_t *dn;
3034
3035 DB_DNODE_ENTER(db);
3036 dn = DB_DNODE(db);
3037 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3038 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
3039 db->db_blkptr == &dn->dn_phys->dn_spill);
3040 DB_DNODE_EXIT(db);
3041 }
3042 #endif
3043
3044 if (db->db_level == 0) {
3045 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
3046 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
3047 if (db->db_state != DB_NOFILL) {
3048 if (dr->dt.dl.dr_data != db->db_buf)
3049 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data,
3050 db));
3051 else if (!arc_released(db->db_buf))
3052 arc_set_callback(db->db_buf, dbuf_do_evict, db);
3053 }
3054 } else {
3055 dnode_t *dn;
3056
3057 DB_DNODE_ENTER(db);
3058 dn = DB_DNODE(db);
3059 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3060 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
3061 if (!BP_IS_HOLE(db->db_blkptr)) {
3062 int epbs =
3063 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3064 ASSERT3U(db->db_blkid, <=,
3065 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs));
3066 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
3067 db->db.db_size);
3068 if (!arc_released(db->db_buf))
3069 arc_set_callback(db->db_buf, dbuf_do_evict, db);
3070 }
3071 DB_DNODE_EXIT(db);
3072 mutex_destroy(&dr->dt.di.dr_mtx);
3073 list_destroy(&dr->dt.di.dr_children);
3074 }
3075 kmem_free(dr, sizeof (dbuf_dirty_record_t));
3076
3077 cv_broadcast(&db->db_changed);
3078 ASSERT(db->db_dirtycnt > 0);
3079 db->db_dirtycnt -= 1;
3080 db->db_data_pending = NULL;
3081 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg);
3082 }
3083
3084 static void
3085 dbuf_write_nofill_ready(zio_t *zio)
3086 {
3087 dbuf_write_ready(zio, NULL, zio->io_private);
3088 }
3089
3090 static void
3091 dbuf_write_nofill_done(zio_t *zio)
3092 {
3093 dbuf_write_done(zio, NULL, zio->io_private);
3094 }
3095
3096 static void
3097 dbuf_write_override_ready(zio_t *zio)
3098 {
3099 dbuf_dirty_record_t *dr = zio->io_private;
3100 dmu_buf_impl_t *db = dr->dr_dbuf;
3101
3102 dbuf_write_ready(zio, NULL, db);
3103 }
3104
3105 static void
3106 dbuf_write_override_done(zio_t *zio)
3107 {
3108 dbuf_dirty_record_t *dr = zio->io_private;
3109 dmu_buf_impl_t *db = dr->dr_dbuf;
3110 blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
3111
3112 mutex_enter(&db->db_mtx);
3113 if (!BP_EQUAL(zio->io_bp, obp)) {
3114 if (!BP_IS_HOLE(obp))
3115 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
3116 arc_release(dr->dt.dl.dr_data, db);
3117 }
3118 mutex_exit(&db->db_mtx);
3119
3120 dbuf_write_done(zio, NULL, db);
3121 }
3122
3123 /* Issue I/O to commit a dirty buffer to disk. */
3124 static void
3125 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
3126 {
3127 dmu_buf_impl_t *db = dr->dr_dbuf;
3128 dnode_t *dn;
3129 objset_t *os;
3130 dmu_buf_impl_t *parent = db->db_parent;
3131 uint64_t txg = tx->tx_txg;
3132 zbookmark_phys_t zb;
3133 zio_prop_t zp;
3134 zio_t *zio;
3135 int wp_flag = 0;
3136
3137 DB_DNODE_ENTER(db);
3138 dn = DB_DNODE(db);
3139 os = dn->dn_objset;
3140
3141 if (db->db_state != DB_NOFILL) {
3142 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
3143 /*
3144 * Private object buffers are released here rather
3145 * than in dbuf_dirty() since they are only modified
3146 * in the syncing context and we don't want the
3147 * overhead of making multiple copies of the data.
3148 */
3149 if (BP_IS_HOLE(db->db_blkptr)) {
3150 arc_buf_thaw(data);
3151 } else {
3152 dbuf_release_bp(db);
3153 }
3154 }
3155 }
3156
3157 if (parent != dn->dn_dbuf) {
3158 /* Our parent is an indirect block. */
3159 /* We have a dirty parent that has been scheduled for write. */
3160 ASSERT(parent && parent->db_data_pending);
3161 /* Our parent's buffer is one level closer to the dnode. */
3162 ASSERT(db->db_level == parent->db_level-1);
3163 /*
3164 * We're about to modify our parent's db_data by modifying
3165 * our block pointer, so the parent must be released.
3166 */
3167 ASSERT(arc_released(parent->db_buf));
3168 zio = parent->db_data_pending->dr_zio;
3169 } else {
3170 /* Our parent is the dnode itself. */
3171 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
3172 db->db_blkid != DMU_SPILL_BLKID) ||
3173 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
3174 if (db->db_blkid != DMU_SPILL_BLKID)
3175 ASSERT3P(db->db_blkptr, ==,
3176 &dn->dn_phys->dn_blkptr[db->db_blkid]);
3177 zio = dn->dn_zio;
3178 }
3179
3180 ASSERT(db->db_level == 0 || data == db->db_buf);
3181 ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
3182 ASSERT(zio);
3183
3184 SET_BOOKMARK(&zb, os->os_dsl_dataset ?
3185 os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
3186 db->db.db_object, db->db_level, db->db_blkid);
3187
3188 if (db->db_blkid == DMU_SPILL_BLKID)
3189 wp_flag = WP_SPILL;
3190 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
3191
3192 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
3193 DB_DNODE_EXIT(db);
3194
3195 if (db->db_level == 0 &&
3196 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
3197 /*
3198 * The BP for this block has been provided by open context
3199 * (by dmu_sync() or dmu_buf_write_embedded()).
3200 */
3201 void *contents = (data != NULL) ? data->b_data : NULL;
3202
3203 dr->dr_zio = zio_write(zio, os->os_spa, txg,
3204 db->db_blkptr, contents, db->db.db_size, &zp,
3205 dbuf_write_override_ready, NULL, dbuf_write_override_done,
3206 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
3207 mutex_enter(&db->db_mtx);
3208 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
3209 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
3210 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
3211 mutex_exit(&db->db_mtx);
3212 } else if (db->db_state == DB_NOFILL) {
3213 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF ||
3214 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY);
3215 dr->dr_zio = zio_write(zio, os->os_spa, txg,
3216 db->db_blkptr, NULL, db->db.db_size, &zp,
3217 dbuf_write_nofill_ready, NULL, dbuf_write_nofill_done, db,
3218 ZIO_PRIORITY_ASYNC_WRITE,
3219 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
3220 } else {
3221 ASSERT(arc_released(data));
3222 dr->dr_zio = arc_write(zio, os->os_spa, txg,
3223 db->db_blkptr, data, DBUF_IS_L2CACHEABLE(db),
3224 DBUF_IS_L2COMPRESSIBLE(db), &zp, dbuf_write_ready,
3225 dbuf_write_physdone, dbuf_write_done, db,
3226 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
3227 }
3228 }