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, 2016 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 * Copyright (c) 2014 Integros [integros.com]
29 */
30
31 #include <sys/zfs_context.h>
32 #include <sys/dmu.h>
33 #include <sys/dmu_send.h>
34 #include <sys/dmu_impl.h>
35 #include <sys/dbuf.h>
36 #include <sys/dmu_objset.h>
37 #include <sys/dsl_dataset.h>
38 #include <sys/dsl_dir.h>
39 #include <sys/dmu_tx.h>
40 #include <sys/spa.h>
41 #include <sys/zio.h>
42 #include <sys/dmu_zfetch.h>
43 #include <sys/sa.h>
44 #include <sys/sa_impl.h>
45 #include <sys/zfeature.h>
46 #include <sys/blkptr.h>
47 #include <sys/range_tree.h>
48 #include <sys/callb.h>
49
50 uint_t zfs_dbuf_evict_key;
51
52 static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
53 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
54
55 #ifndef __lint
56 extern inline void dmu_buf_init_user(dmu_buf_user_t *dbu,
57 dmu_buf_evict_func_t *evict_func_sync,
58 dmu_buf_evict_func_t *evict_func_async,
59 dmu_buf_t **clear_on_evict_dbufp);
60 #endif /* ! __lint */
61
62 /*
63 * Global data structures and functions for the dbuf cache.
64 */
65 static kmem_cache_t *dbuf_kmem_cache;
66 static taskq_t *dbu_evict_taskq;
67
68 static kthread_t *dbuf_cache_evict_thread;
69 static kmutex_t dbuf_evict_lock;
70 static kcondvar_t dbuf_evict_cv;
71 static boolean_t dbuf_evict_thread_exit;
72
73 /*
74 * LRU cache of dbufs. The dbuf cache maintains a list of dbufs that
75 * are not currently held but have been recently released. These dbufs
76 * are not eligible for arc eviction until they are aged out of the cache.
77 * Dbufs are added to the dbuf cache once the last hold is released. If a
78 * dbuf is later accessed and still exists in the dbuf cache, then it will
79 * be removed from the cache and later re-added to the head of the cache.
80 * Dbufs that are aged out of the cache will be immediately destroyed and
81 * become eligible for arc eviction.
82 */
83 static multilist_t dbuf_cache;
84 static refcount_t dbuf_cache_size;
85 uint64_t dbuf_cache_max_bytes = 100 * 1024 * 1024;
86
87 /* Cap the size of the dbuf cache to log2 fraction of arc size. */
88 int dbuf_cache_max_shift = 5;
89
90 /*
91 * The dbuf cache uses a three-stage eviction policy:
92 * - A low water marker designates when the dbuf eviction thread
93 * should stop evicting from the dbuf cache.
94 * - When we reach the maximum size (aka mid water mark), we
95 * signal the eviction thread to run.
96 * - The high water mark indicates when the eviction thread
97 * is unable to keep up with the incoming load and eviction must
98 * happen in the context of the calling thread.
99 *
100 * The dbuf cache:
101 * (max size)
102 * low water mid water hi water
103 * +----------------------------------------+----------+----------+
104 * | | | |
105 * | | | |
106 * | | | |
107 * | | | |
108 * +----------------------------------------+----------+----------+
109 * stop signal evict
110 * evicting eviction directly
111 * thread
112 *
113 * The high and low water marks indicate the operating range for the eviction
114 * thread. The low water mark is, by default, 90% of the total size of the
115 * cache and the high water mark is at 110% (both of these percentages can be
116 * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct,
117 * respectively). The eviction thread will try to ensure that the cache remains
118 * within this range by waking up every second and checking if the cache is
119 * above the low water mark. The thread can also be woken up by callers adding
120 * elements into the cache if the cache is larger than the mid water (i.e max
121 * cache size). Once the eviction thread is woken up and eviction is required,
122 * it will continue evicting buffers until it's able to reduce the cache size
123 * to the low water mark. If the cache size continues to grow and hits the high
124 * water mark, then callers adding elments to the cache will begin to evict
125 * directly from the cache until the cache is no longer above the high water
126 * mark.
127 */
128
129 /*
130 * The percentage above and below the maximum cache size.
131 */
132 uint_t dbuf_cache_hiwater_pct = 10;
133 uint_t dbuf_cache_lowater_pct = 10;
134
135 /* ARGSUSED */
136 static int
dbuf_cons(void * vdb,void * unused,int kmflag)137 dbuf_cons(void *vdb, void *unused, int kmflag)
138 {
139 dmu_buf_impl_t *db = vdb;
140
141 bzero(db, sizeof (dmu_buf_impl_t));
142 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
143 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
144 multilist_link_init(&db->db_cache_link);
145 refcount_create(&db->db_holds);
146
147 return (0);
148 }
149
150 /* ARGSUSED */
151 static void
dbuf_dest(void * vdb,void * unused)152 dbuf_dest(void *vdb, void *unused)
153 {
154 dmu_buf_impl_t *db = vdb;
155
156 mutex_destroy(&db->db_mtx);
157 cv_destroy(&db->db_changed);
158 ASSERT(!multilist_link_active(&db->db_cache_link));
159 refcount_destroy(&db->db_holds);
160 }
161
162 /*
163 * dbuf hash table routines
164 */
165 static dbuf_hash_table_t dbuf_hash_table;
166
167 static uint64_t dbuf_hash_count;
168
169 static uint64_t
dbuf_hash(void * os,uint64_t obj,uint8_t lvl,uint64_t blkid)170 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
171 {
172 uintptr_t osv = (uintptr_t)os;
173 uint64_t crc = -1ULL;
174
175 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
176 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF];
177 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF];
178 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF];
179 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF];
180 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF];
181 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF];
182
183 crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16);
184
185 return (crc);
186 }
187
188 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
189 ((dbuf)->db.db_object == (obj) && \
190 (dbuf)->db_objset == (os) && \
191 (dbuf)->db_level == (level) && \
192 (dbuf)->db_blkid == (blkid))
193
194 dmu_buf_impl_t *
dbuf_find(objset_t * os,uint64_t obj,uint8_t level,uint64_t blkid)195 dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid)
196 {
197 dbuf_hash_table_t *h = &dbuf_hash_table;
198 uint64_t hv = dbuf_hash(os, obj, level, blkid);
199 uint64_t idx = hv & h->hash_table_mask;
200 dmu_buf_impl_t *db;
201
202 mutex_enter(DBUF_HASH_MUTEX(h, idx));
203 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
204 if (DBUF_EQUAL(db, os, obj, level, blkid)) {
205 mutex_enter(&db->db_mtx);
206 if (db->db_state != DB_EVICTING) {
207 mutex_exit(DBUF_HASH_MUTEX(h, idx));
208 return (db);
209 }
210 mutex_exit(&db->db_mtx);
211 }
212 }
213 mutex_exit(DBUF_HASH_MUTEX(h, idx));
214 return (NULL);
215 }
216
217 static dmu_buf_impl_t *
dbuf_find_bonus(objset_t * os,uint64_t object)218 dbuf_find_bonus(objset_t *os, uint64_t object)
219 {
220 dnode_t *dn;
221 dmu_buf_impl_t *db = NULL;
222
223 if (dnode_hold(os, object, FTAG, &dn) == 0) {
224 rw_enter(&dn->dn_struct_rwlock, RW_READER);
225 if (dn->dn_bonus != NULL) {
226 db = dn->dn_bonus;
227 mutex_enter(&db->db_mtx);
228 }
229 rw_exit(&dn->dn_struct_rwlock);
230 dnode_rele(dn, FTAG);
231 }
232 return (db);
233 }
234
235 /*
236 * Insert an entry into the hash table. If there is already an element
237 * equal to elem in the hash table, then the already existing element
238 * will be returned and the new element will not be inserted.
239 * Otherwise returns NULL.
240 */
241 static dmu_buf_impl_t *
dbuf_hash_insert(dmu_buf_impl_t * db)242 dbuf_hash_insert(dmu_buf_impl_t *db)
243 {
244 dbuf_hash_table_t *h = &dbuf_hash_table;
245 objset_t *os = db->db_objset;
246 uint64_t obj = db->db.db_object;
247 int level = db->db_level;
248 uint64_t blkid = db->db_blkid;
249 uint64_t hv = dbuf_hash(os, obj, level, blkid);
250 uint64_t idx = hv & h->hash_table_mask;
251 dmu_buf_impl_t *dbf;
252
253 mutex_enter(DBUF_HASH_MUTEX(h, idx));
254 for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) {
255 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
256 mutex_enter(&dbf->db_mtx);
257 if (dbf->db_state != DB_EVICTING) {
258 mutex_exit(DBUF_HASH_MUTEX(h, idx));
259 return (dbf);
260 }
261 mutex_exit(&dbf->db_mtx);
262 }
263 }
264
265 mutex_enter(&db->db_mtx);
266 db->db_hash_next = h->hash_table[idx];
267 h->hash_table[idx] = db;
268 mutex_exit(DBUF_HASH_MUTEX(h, idx));
269 atomic_inc_64(&dbuf_hash_count);
270
271 return (NULL);
272 }
273
274 /*
275 * Remove an entry from the hash table. It must be in the EVICTING state.
276 */
277 static void
dbuf_hash_remove(dmu_buf_impl_t * db)278 dbuf_hash_remove(dmu_buf_impl_t *db)
279 {
280 dbuf_hash_table_t *h = &dbuf_hash_table;
281 uint64_t hv = dbuf_hash(db->db_objset, db->db.db_object,
282 db->db_level, db->db_blkid);
283 uint64_t idx = hv & h->hash_table_mask;
284 dmu_buf_impl_t *dbf, **dbp;
285
286 /*
287 * We musn't hold db_mtx to maintain lock ordering:
288 * DBUF_HASH_MUTEX > db_mtx.
289 */
290 ASSERT(refcount_is_zero(&db->db_holds));
291 ASSERT(db->db_state == DB_EVICTING);
292 ASSERT(!MUTEX_HELD(&db->db_mtx));
293
294 mutex_enter(DBUF_HASH_MUTEX(h, idx));
295 dbp = &h->hash_table[idx];
296 while ((dbf = *dbp) != db) {
297 dbp = &dbf->db_hash_next;
298 ASSERT(dbf != NULL);
299 }
300 *dbp = db->db_hash_next;
301 db->db_hash_next = NULL;
302 mutex_exit(DBUF_HASH_MUTEX(h, idx));
303 atomic_dec_64(&dbuf_hash_count);
304 }
305
306 typedef enum {
307 DBVU_EVICTING,
308 DBVU_NOT_EVICTING
309 } dbvu_verify_type_t;
310
311 static void
dbuf_verify_user(dmu_buf_impl_t * db,dbvu_verify_type_t verify_type)312 dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type)
313 {
314 #ifdef ZFS_DEBUG
315 int64_t holds;
316
317 if (db->db_user == NULL)
318 return;
319
320 /* Only data blocks support the attachment of user data. */
321 ASSERT(db->db_level == 0);
322
323 /* Clients must resolve a dbuf before attaching user data. */
324 ASSERT(db->db.db_data != NULL);
325 ASSERT3U(db->db_state, ==, DB_CACHED);
326
327 holds = refcount_count(&db->db_holds);
328 if (verify_type == DBVU_EVICTING) {
329 /*
330 * Immediate eviction occurs when holds == dirtycnt.
331 * For normal eviction buffers, holds is zero on
332 * eviction, except when dbuf_fix_old_data() calls
333 * dbuf_clear_data(). However, the hold count can grow
334 * during eviction even though db_mtx is held (see
335 * dmu_bonus_hold() for an example), so we can only
336 * test the generic invariant that holds >= dirtycnt.
337 */
338 ASSERT3U(holds, >=, db->db_dirtycnt);
339 } else {
340 if (db->db_user_immediate_evict == TRUE)
341 ASSERT3U(holds, >=, db->db_dirtycnt);
342 else
343 ASSERT3U(holds, >, 0);
344 }
345 #endif
346 }
347
348 static void
dbuf_evict_user(dmu_buf_impl_t * db)349 dbuf_evict_user(dmu_buf_impl_t *db)
350 {
351 dmu_buf_user_t *dbu = db->db_user;
352
353 ASSERT(MUTEX_HELD(&db->db_mtx));
354
355 if (dbu == NULL)
356 return;
357
358 dbuf_verify_user(db, DBVU_EVICTING);
359 db->db_user = NULL;
360
361 #ifdef ZFS_DEBUG
362 if (dbu->dbu_clear_on_evict_dbufp != NULL)
363 *dbu->dbu_clear_on_evict_dbufp = NULL;
364 #endif
365
366 /*
367 * There are two eviction callbacks - one that we call synchronously
368 * and one that we invoke via a taskq. The async one is useful for
369 * avoiding lock order reversals and limiting stack depth.
370 *
371 * Note that if we have a sync callback but no async callback,
372 * it's likely that the sync callback will free the structure
373 * containing the dbu. In that case we need to take care to not
374 * dereference dbu after calling the sync evict func.
375 */
376 boolean_t has_async = (dbu->dbu_evict_func_async != NULL);
377
378 if (dbu->dbu_evict_func_sync != NULL)
379 dbu->dbu_evict_func_sync(dbu);
380
381 if (has_async) {
382 taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func_async,
383 dbu, 0, &dbu->dbu_tqent);
384 }
385 }
386
387 boolean_t
dbuf_is_metadata(dmu_buf_impl_t * db)388 dbuf_is_metadata(dmu_buf_impl_t *db)
389 {
390 if (db->db_level > 0) {
391 return (B_TRUE);
392 } else {
393 boolean_t is_metadata;
394
395 DB_DNODE_ENTER(db);
396 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type);
397 DB_DNODE_EXIT(db);
398
399 return (is_metadata);
400 }
401 }
402
403 /*
404 * This function *must* return indices evenly distributed between all
405 * sublists of the multilist. This is needed due to how the dbuf eviction
406 * code is laid out; dbuf_evict_thread() assumes dbufs are evenly
407 * distributed between all sublists and uses this assumption when
408 * deciding which sublist to evict from and how much to evict from it.
409 */
410 unsigned int
dbuf_cache_multilist_index_func(multilist_t * ml,void * obj)411 dbuf_cache_multilist_index_func(multilist_t *ml, void *obj)
412 {
413 dmu_buf_impl_t *db = obj;
414
415 /*
416 * The assumption here, is the hash value for a given
417 * dmu_buf_impl_t will remain constant throughout it's lifetime
418 * (i.e. it's objset, object, level and blkid fields don't change).
419 * Thus, we don't need to store the dbuf's sublist index
420 * on insertion, as this index can be recalculated on removal.
421 *
422 * Also, the low order bits of the hash value are thought to be
423 * distributed evenly. Otherwise, in the case that the multilist
424 * has a power of two number of sublists, each sublists' usage
425 * would not be evenly distributed.
426 */
427 return (dbuf_hash(db->db_objset, db->db.db_object,
428 db->db_level, db->db_blkid) %
429 multilist_get_num_sublists(ml));
430 }
431
432 static inline boolean_t
dbuf_cache_above_hiwater(void)433 dbuf_cache_above_hiwater(void)
434 {
435 uint64_t dbuf_cache_hiwater_bytes =
436 (dbuf_cache_max_bytes * dbuf_cache_hiwater_pct) / 100;
437
438 return (refcount_count(&dbuf_cache_size) >
439 dbuf_cache_max_bytes + dbuf_cache_hiwater_bytes);
440 }
441
442 static inline boolean_t
dbuf_cache_above_lowater(void)443 dbuf_cache_above_lowater(void)
444 {
445 uint64_t dbuf_cache_lowater_bytes =
446 (dbuf_cache_max_bytes * dbuf_cache_lowater_pct) / 100;
447
448 return (refcount_count(&dbuf_cache_size) >
449 dbuf_cache_max_bytes - dbuf_cache_lowater_bytes);
450 }
451
452 /*
453 * Evict the oldest eligible dbuf from the dbuf cache.
454 */
455 static void
dbuf_evict_one(void)456 dbuf_evict_one(void)
457 {
458 int idx = multilist_get_random_index(&dbuf_cache);
459 multilist_sublist_t *mls = multilist_sublist_lock(&dbuf_cache, idx);
460
461 ASSERT(!MUTEX_HELD(&dbuf_evict_lock));
462
463 /*
464 * Set the thread's tsd to indicate that it's processing evictions.
465 * Once a thread stops evicting from the dbuf cache it will
466 * reset its tsd to NULL.
467 */
468 ASSERT3P(tsd_get(zfs_dbuf_evict_key), ==, NULL);
469 (void) tsd_set(zfs_dbuf_evict_key, (void *)B_TRUE);
470
471 dmu_buf_impl_t *db = multilist_sublist_tail(mls);
472 while (db != NULL && mutex_tryenter(&db->db_mtx) == 0) {
473 db = multilist_sublist_prev(mls, db);
474 }
475
476 DTRACE_PROBE2(dbuf__evict__one, dmu_buf_impl_t *, db,
477 multilist_sublist_t *, mls);
478
479 if (db != NULL) {
480 multilist_sublist_remove(mls, db);
481 multilist_sublist_unlock(mls);
482 (void) refcount_remove_many(&dbuf_cache_size,
483 db->db.db_size, db);
484 dbuf_destroy(db);
485 } else {
486 multilist_sublist_unlock(mls);
487 }
488 (void) tsd_set(zfs_dbuf_evict_key, NULL);
489 }
490
491 /*
492 * The dbuf evict thread is responsible for aging out dbufs from the
493 * cache. Once the cache has reached it's maximum size, dbufs are removed
494 * and destroyed. The eviction thread will continue running until the size
495 * of the dbuf cache is at or below the maximum size. Once the dbuf is aged
496 * out of the cache it is destroyed and becomes eligible for arc eviction.
497 */
498 static void
dbuf_evict_thread(void * dummy __unused)499 dbuf_evict_thread(void *dummy __unused)
500 {
501 callb_cpr_t cpr;
502
503 CALLB_CPR_INIT(&cpr, &dbuf_evict_lock, callb_generic_cpr, FTAG);
504
505 mutex_enter(&dbuf_evict_lock);
506 while (!dbuf_evict_thread_exit) {
507 while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
508 CALLB_CPR_SAFE_BEGIN(&cpr);
509 (void) cv_timedwait_hires(&dbuf_evict_cv,
510 &dbuf_evict_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
511 CALLB_CPR_SAFE_END(&cpr, &dbuf_evict_lock);
512 }
513 mutex_exit(&dbuf_evict_lock);
514
515 /*
516 * Keep evicting as long as we're above the low water mark
517 * for the cache. We do this without holding the locks to
518 * minimize lock contention.
519 */
520 while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
521 dbuf_evict_one();
522 }
523
524 mutex_enter(&dbuf_evict_lock);
525 }
526
527 dbuf_evict_thread_exit = B_FALSE;
528 cv_broadcast(&dbuf_evict_cv);
529 CALLB_CPR_EXIT(&cpr); /* drops dbuf_evict_lock */
530 thread_exit();
531 }
532
533 /*
534 * Wake up the dbuf eviction thread if the dbuf cache is at its max size.
535 * If the dbuf cache is at its high water mark, then evict a dbuf from the
536 * dbuf cache using the callers context.
537 */
538 static void
dbuf_evict_notify(void)539 dbuf_evict_notify(void)
540 {
541
542 /*
543 * We use thread specific data to track when a thread has
544 * started processing evictions. This allows us to avoid deeply
545 * nested stacks that would have a call flow similar to this:
546 *
547 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify()
548 * ^ |
549 * | |
550 * +-----dbuf_destroy()<--dbuf_evict_one()<--------+
551 *
552 * The dbuf_eviction_thread will always have its tsd set until
553 * that thread exits. All other threads will only set their tsd
554 * if they are participating in the eviction process. This only
555 * happens if the eviction thread is unable to process evictions
556 * fast enough. To keep the dbuf cache size in check, other threads
557 * can evict from the dbuf cache directly. Those threads will set
558 * their tsd values so that we ensure that they only evict one dbuf
559 * from the dbuf cache.
560 */
561 if (tsd_get(zfs_dbuf_evict_key) != NULL)
562 return;
563
564 if (refcount_count(&dbuf_cache_size) > dbuf_cache_max_bytes) {
565 boolean_t evict_now = B_FALSE;
566
567 mutex_enter(&dbuf_evict_lock);
568 if (refcount_count(&dbuf_cache_size) > dbuf_cache_max_bytes) {
569 evict_now = dbuf_cache_above_hiwater();
570 cv_signal(&dbuf_evict_cv);
571 }
572 mutex_exit(&dbuf_evict_lock);
573
574 if (evict_now) {
575 dbuf_evict_one();
576 }
577 }
578 }
579
580 void
dbuf_init(void)581 dbuf_init(void)
582 {
583 uint64_t hsize = 1ULL << 16;
584 dbuf_hash_table_t *h = &dbuf_hash_table;
585 int i;
586
587 /*
588 * The hash table is big enough to fill all of physical memory
589 * with an average 4K block size. The table will take up
590 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers).
591 */
592 while (hsize * 4096 < (uint64_t)physmem * PAGESIZE)
593 hsize <<= 1;
594
595 retry:
596 h->hash_table_mask = hsize - 1;
597 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
598 if (h->hash_table == NULL) {
599 /* XXX - we should really return an error instead of assert */
600 ASSERT(hsize > (1ULL << 10));
601 hsize >>= 1;
602 goto retry;
603 }
604
605 dbuf_kmem_cache = kmem_cache_create("dmu_buf_impl_t",
606 sizeof (dmu_buf_impl_t),
607 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
608
609 for (i = 0; i < DBUF_MUTEXES; i++)
610 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
611
612 /*
613 * Setup the parameters for the dbuf cache. We cap the size of the
614 * dbuf cache to 1/32nd (default) of the size of the ARC.
615 */
616 dbuf_cache_max_bytes = MIN(dbuf_cache_max_bytes,
617 arc_max_bytes() >> dbuf_cache_max_shift);
618
619 /*
620 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
621 * configuration is not required.
622 */
623 dbu_evict_taskq = taskq_create("dbu_evict", 1, minclsyspri, 0, 0, 0);
624
625 multilist_create(&dbuf_cache, sizeof (dmu_buf_impl_t),
626 offsetof(dmu_buf_impl_t, db_cache_link),
627 zfs_arc_num_sublists_per_state,
628 dbuf_cache_multilist_index_func);
629 refcount_create(&dbuf_cache_size);
630
631 tsd_create(&zfs_dbuf_evict_key, NULL);
632 dbuf_evict_thread_exit = B_FALSE;
633 mutex_init(&dbuf_evict_lock, NULL, MUTEX_DEFAULT, NULL);
634 cv_init(&dbuf_evict_cv, NULL, CV_DEFAULT, NULL);
635 dbuf_cache_evict_thread = thread_create(NULL, 0, dbuf_evict_thread,
636 NULL, 0, &p0, TS_RUN, minclsyspri);
637 }
638
639 void
dbuf_fini(void)640 dbuf_fini(void)
641 {
642 dbuf_hash_table_t *h = &dbuf_hash_table;
643 int i;
644
645 for (i = 0; i < DBUF_MUTEXES; i++)
646 mutex_destroy(&h->hash_mutexes[i]);
647 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
648 kmem_cache_destroy(dbuf_kmem_cache);
649 taskq_destroy(dbu_evict_taskq);
650
651 mutex_enter(&dbuf_evict_lock);
652 dbuf_evict_thread_exit = B_TRUE;
653 while (dbuf_evict_thread_exit) {
654 cv_signal(&dbuf_evict_cv);
655 cv_wait(&dbuf_evict_cv, &dbuf_evict_lock);
656 }
657 mutex_exit(&dbuf_evict_lock);
658 tsd_destroy(&zfs_dbuf_evict_key);
659
660 mutex_destroy(&dbuf_evict_lock);
661 cv_destroy(&dbuf_evict_cv);
662
663 refcount_destroy(&dbuf_cache_size);
664 multilist_destroy(&dbuf_cache);
665 }
666
667 /*
668 * Other stuff.
669 */
670
671 #ifdef ZFS_DEBUG
672 static void
dbuf_verify(dmu_buf_impl_t * db)673 dbuf_verify(dmu_buf_impl_t *db)
674 {
675 dnode_t *dn;
676 dbuf_dirty_record_t *dr;
677
678 ASSERT(MUTEX_HELD(&db->db_mtx));
679
680 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
681 return;
682
683 ASSERT(db->db_objset != NULL);
684 DB_DNODE_ENTER(db);
685 dn = DB_DNODE(db);
686 if (dn == NULL) {
687 ASSERT(db->db_parent == NULL);
688 ASSERT(db->db_blkptr == NULL);
689 } else {
690 ASSERT3U(db->db.db_object, ==, dn->dn_object);
691 ASSERT3P(db->db_objset, ==, dn->dn_objset);
692 ASSERT3U(db->db_level, <, dn->dn_nlevels);
693 ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
694 db->db_blkid == DMU_SPILL_BLKID ||
695 !avl_is_empty(&dn->dn_dbufs));
696 }
697 if (db->db_blkid == DMU_BONUS_BLKID) {
698 ASSERT(dn != NULL);
699 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
700 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
701 } else if (db->db_blkid == DMU_SPILL_BLKID) {
702 ASSERT(dn != NULL);
703 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
704 ASSERT0(db->db.db_offset);
705 } else {
706 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
707 }
708
709 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
710 ASSERT(dr->dr_dbuf == db);
711
712 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
713 ASSERT(dr->dr_dbuf == db);
714
715 /*
716 * We can't assert that db_size matches dn_datablksz because it
717 * can be momentarily different when another thread is doing
718 * dnode_set_blksz().
719 */
720 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
721 dr = db->db_data_pending;
722 /*
723 * It should only be modified in syncing context, so
724 * make sure we only have one copy of the data.
725 */
726 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
727 }
728
729 /* verify db->db_blkptr */
730 if (db->db_blkptr) {
731 if (db->db_parent == dn->dn_dbuf) {
732 /* db is pointed to by the dnode */
733 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
734 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
735 ASSERT(db->db_parent == NULL);
736 else
737 ASSERT(db->db_parent != NULL);
738 if (db->db_blkid != DMU_SPILL_BLKID)
739 ASSERT3P(db->db_blkptr, ==,
740 &dn->dn_phys->dn_blkptr[db->db_blkid]);
741 } else {
742 /* db is pointed to by an indirect block */
743 int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT;
744 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
745 ASSERT3U(db->db_parent->db.db_object, ==,
746 db->db.db_object);
747 /*
748 * dnode_grow_indblksz() can make this fail if we don't
749 * have the struct_rwlock. XXX indblksz no longer
750 * grows. safe to do this now?
751 */
752 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
753 ASSERT3P(db->db_blkptr, ==,
754 ((blkptr_t *)db->db_parent->db.db_data +
755 db->db_blkid % epb));
756 }
757 }
758 }
759 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
760 (db->db_buf == NULL || db->db_buf->b_data) &&
761 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
762 db->db_state != DB_FILL && !dn->dn_free_txg) {
763 /*
764 * If the blkptr isn't set but they have nonzero data,
765 * it had better be dirty, otherwise we'll lose that
766 * data when we evict this buffer.
767 *
768 * There is an exception to this rule for indirect blocks; in
769 * this case, if the indirect block is a hole, we fill in a few
770 * fields on each of the child blocks (importantly, birth time)
771 * to prevent hole birth times from being lost when you
772 * partially fill in a hole.
773 */
774 if (db->db_dirtycnt == 0) {
775 if (db->db_level == 0) {
776 uint64_t *buf = db->db.db_data;
777 int i;
778
779 for (i = 0; i < db->db.db_size >> 3; i++) {
780 ASSERT(buf[i] == 0);
781 }
782 } else {
783 blkptr_t *bps = db->db.db_data;
784 ASSERT3U(1 << DB_DNODE(db)->dn_indblkshift, ==,
785 db->db.db_size);
786 /*
787 * We want to verify that all the blkptrs in the
788 * indirect block are holes, but we may have
789 * automatically set up a few fields for them.
790 * We iterate through each blkptr and verify
791 * they only have those fields set.
792 */
793 for (int i = 0;
794 i < db->db.db_size / sizeof (blkptr_t);
795 i++) {
796 blkptr_t *bp = &bps[i];
797 ASSERT(ZIO_CHECKSUM_IS_ZERO(
798 &bp->blk_cksum));
799 ASSERT(
800 DVA_IS_EMPTY(&bp->blk_dva[0]) &&
801 DVA_IS_EMPTY(&bp->blk_dva[1]) &&
802 DVA_IS_EMPTY(&bp->blk_dva[2]));
803 ASSERT0(bp->blk_fill);
804 ASSERT0(bp->blk_pad[0]);
805 ASSERT0(bp->blk_pad[1]);
806 ASSERT(!BP_IS_EMBEDDED(bp));
807 ASSERT(BP_IS_HOLE(bp));
808 ASSERT0(bp->blk_phys_birth);
809 }
810 }
811 }
812 }
813 DB_DNODE_EXIT(db);
814 }
815 #endif
816
817 static void
dbuf_clear_data(dmu_buf_impl_t * db)818 dbuf_clear_data(dmu_buf_impl_t *db)
819 {
820 ASSERT(MUTEX_HELD(&db->db_mtx));
821 dbuf_evict_user(db);
822 ASSERT3P(db->db_buf, ==, NULL);
823 db->db.db_data = NULL;
824 if (db->db_state != DB_NOFILL)
825 db->db_state = DB_UNCACHED;
826 }
827
828 static void
dbuf_set_data(dmu_buf_impl_t * db,arc_buf_t * buf)829 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
830 {
831 ASSERT(MUTEX_HELD(&db->db_mtx));
832 ASSERT(buf != NULL);
833
834 db->db_buf = buf;
835 ASSERT(buf->b_data != NULL);
836 db->db.db_data = buf->b_data;
837 }
838
839 /*
840 * Loan out an arc_buf for read. Return the loaned arc_buf.
841 */
842 arc_buf_t *
dbuf_loan_arcbuf(dmu_buf_impl_t * db)843 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
844 {
845 arc_buf_t *abuf;
846
847 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
848 mutex_enter(&db->db_mtx);
849 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
850 int blksz = db->db.db_size;
851 spa_t *spa = db->db_objset->os_spa;
852
853 mutex_exit(&db->db_mtx);
854 abuf = arc_loan_buf(spa, blksz);
855 bcopy(db->db.db_data, abuf->b_data, blksz);
856 } else {
857 abuf = db->db_buf;
858 arc_loan_inuse_buf(abuf, db);
859 db->db_buf = NULL;
860 dbuf_clear_data(db);
861 mutex_exit(&db->db_mtx);
862 }
863 return (abuf);
864 }
865
866 /*
867 * Calculate which level n block references the data at the level 0 offset
868 * provided.
869 */
870 uint64_t
dbuf_whichblock(dnode_t * dn,int64_t level,uint64_t offset)871 dbuf_whichblock(dnode_t *dn, int64_t level, uint64_t offset)
872 {
873 if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) {
874 /*
875 * The level n blkid is equal to the level 0 blkid divided by
876 * the number of level 0s in a level n block.
877 *
878 * The level 0 blkid is offset >> datablkshift =
879 * offset / 2^datablkshift.
880 *
881 * The number of level 0s in a level n is the number of block
882 * pointers in an indirect block, raised to the power of level.
883 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
884 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
885 *
886 * Thus, the level n blkid is: offset /
887 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
888 * = offset / 2^(datablkshift + level *
889 * (indblkshift - SPA_BLKPTRSHIFT))
890 * = offset >> (datablkshift + level *
891 * (indblkshift - SPA_BLKPTRSHIFT))
892 */
893 return (offset >> (dn->dn_datablkshift + level *
894 (dn->dn_indblkshift - SPA_BLKPTRSHIFT)));
895 } else {
896 ASSERT3U(offset, <, dn->dn_datablksz);
897 return (0);
898 }
899 }
900
901 static void
dbuf_read_done(zio_t * zio,arc_buf_t * buf,void * vdb)902 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
903 {
904 dmu_buf_impl_t *db = vdb;
905
906 mutex_enter(&db->db_mtx);
907 ASSERT3U(db->db_state, ==, DB_READ);
908 /*
909 * All reads are synchronous, so we must have a hold on the dbuf
910 */
911 ASSERT(refcount_count(&db->db_holds) > 0);
912 ASSERT(db->db_buf == NULL);
913 ASSERT(db->db.db_data == NULL);
914 if (db->db_level == 0 && db->db_freed_in_flight) {
915 /* we were freed in flight; disregard any error */
916 arc_release(buf, db);
917 bzero(buf->b_data, db->db.db_size);
918 arc_buf_freeze(buf);
919 db->db_freed_in_flight = FALSE;
920 dbuf_set_data(db, buf);
921 db->db_state = DB_CACHED;
922 } else if (zio == NULL || zio->io_error == 0) {
923 dbuf_set_data(db, buf);
924 db->db_state = DB_CACHED;
925 } else {
926 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
927 ASSERT3P(db->db_buf, ==, NULL);
928 arc_buf_destroy(buf, db);
929 db->db_state = DB_UNCACHED;
930 }
931 cv_broadcast(&db->db_changed);
932 dbuf_rele_and_unlock(db, NULL);
933 }
934
935 static void
dbuf_read_impl(dmu_buf_impl_t * db,zio_t * zio,uint32_t flags)936 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
937 {
938 dnode_t *dn;
939 zbookmark_phys_t zb;
940 arc_flags_t aflags = ARC_FLAG_NOWAIT;
941
942 DB_DNODE_ENTER(db);
943 dn = DB_DNODE(db);
944 ASSERT(!refcount_is_zero(&db->db_holds));
945 /* We need the struct_rwlock to prevent db_blkptr from changing. */
946 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
947 ASSERT(MUTEX_HELD(&db->db_mtx));
948 ASSERT(db->db_state == DB_UNCACHED);
949 ASSERT(db->db_buf == NULL);
950
951 if (db->db_blkid == DMU_BONUS_BLKID) {
952 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
953
954 ASSERT3U(bonuslen, <=, db->db.db_size);
955 db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN);
956 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
957 if (bonuslen < DN_MAX_BONUSLEN)
958 bzero(db->db.db_data, DN_MAX_BONUSLEN);
959 if (bonuslen)
960 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
961 DB_DNODE_EXIT(db);
962 db->db_state = DB_CACHED;
963 mutex_exit(&db->db_mtx);
964 return;
965 }
966
967 /*
968 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
969 * processes the delete record and clears the bp while we are waiting
970 * for the dn_mtx (resulting in a "no" from block_freed).
971 */
972 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
973 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
974 BP_IS_HOLE(db->db_blkptr)))) {
975 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
976
977 dbuf_set_data(db, arc_alloc_buf(db->db_objset->os_spa,
978 db->db.db_size, db, type));
979 bzero(db->db.db_data, db->db.db_size);
980
981 if (db->db_blkptr != NULL && db->db_level > 0 &&
982 BP_IS_HOLE(db->db_blkptr) &&
983 db->db_blkptr->blk_birth != 0) {
984 blkptr_t *bps = db->db.db_data;
985 for (int i = 0; i < ((1 <<
986 DB_DNODE(db)->dn_indblkshift) / sizeof (blkptr_t));
987 i++) {
988 blkptr_t *bp = &bps[i];
989 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
990 1 << dn->dn_indblkshift);
991 BP_SET_LSIZE(bp,
992 BP_GET_LEVEL(db->db_blkptr) == 1 ?
993 dn->dn_datablksz :
994 BP_GET_LSIZE(db->db_blkptr));
995 BP_SET_TYPE(bp, BP_GET_TYPE(db->db_blkptr));
996 BP_SET_LEVEL(bp,
997 BP_GET_LEVEL(db->db_blkptr) - 1);
998 BP_SET_BIRTH(bp, db->db_blkptr->blk_birth, 0);
999 }
1000 }
1001 DB_DNODE_EXIT(db);
1002 db->db_state = DB_CACHED;
1003 mutex_exit(&db->db_mtx);
1004 return;
1005 }
1006
1007 DB_DNODE_EXIT(db);
1008
1009 db->db_state = DB_READ;
1010 mutex_exit(&db->db_mtx);
1011
1012 if (DBUF_IS_L2CACHEABLE(db))
1013 aflags |= ARC_FLAG_L2CACHE;
1014
1015 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
1016 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
1017 db->db.db_object, db->db_level, db->db_blkid);
1018
1019 dbuf_add_ref(db, NULL);
1020
1021 (void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr,
1022 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
1023 (flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
1024 &aflags, &zb);
1025 }
1026
1027 int
dbuf_read(dmu_buf_impl_t * db,zio_t * zio,uint32_t flags)1028 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
1029 {
1030 int err = 0;
1031 boolean_t havepzio = (zio != NULL);
1032 boolean_t prefetch;
1033 dnode_t *dn;
1034
1035 /*
1036 * We don't have to hold the mutex to check db_state because it
1037 * can't be freed while we have a hold on the buffer.
1038 */
1039 ASSERT(!refcount_is_zero(&db->db_holds));
1040
1041 if (db->db_state == DB_NOFILL)
1042 return (SET_ERROR(EIO));
1043
1044 DB_DNODE_ENTER(db);
1045 dn = DB_DNODE(db);
1046 if ((flags & DB_RF_HAVESTRUCT) == 0)
1047 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1048
1049 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1050 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
1051 DBUF_IS_CACHEABLE(db);
1052
1053 mutex_enter(&db->db_mtx);
1054 if (db->db_state == DB_CACHED) {
1055 mutex_exit(&db->db_mtx);
1056 if (prefetch)
1057 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1058 if ((flags & DB_RF_HAVESTRUCT) == 0)
1059 rw_exit(&dn->dn_struct_rwlock);
1060 DB_DNODE_EXIT(db);
1061 } else if (db->db_state == DB_UNCACHED) {
1062 spa_t *spa = dn->dn_objset->os_spa;
1063
1064 if (zio == NULL)
1065 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
1066 dbuf_read_impl(db, zio, flags);
1067
1068 /* dbuf_read_impl has dropped db_mtx for us */
1069
1070 if (prefetch)
1071 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1072
1073 if ((flags & DB_RF_HAVESTRUCT) == 0)
1074 rw_exit(&dn->dn_struct_rwlock);
1075 DB_DNODE_EXIT(db);
1076
1077 if (!havepzio)
1078 err = zio_wait(zio);
1079 } else {
1080 /*
1081 * Another reader came in while the dbuf was in flight
1082 * between UNCACHED and CACHED. Either a writer will finish
1083 * writing the buffer (sending the dbuf to CACHED) or the
1084 * first reader's request will reach the read_done callback
1085 * and send the dbuf to CACHED. Otherwise, a failure
1086 * occurred and the dbuf went to UNCACHED.
1087 */
1088 mutex_exit(&db->db_mtx);
1089 if (prefetch)
1090 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1091 if ((flags & DB_RF_HAVESTRUCT) == 0)
1092 rw_exit(&dn->dn_struct_rwlock);
1093 DB_DNODE_EXIT(db);
1094
1095 /* Skip the wait per the caller's request. */
1096 mutex_enter(&db->db_mtx);
1097 if ((flags & DB_RF_NEVERWAIT) == 0) {
1098 while (db->db_state == DB_READ ||
1099 db->db_state == DB_FILL) {
1100 ASSERT(db->db_state == DB_READ ||
1101 (flags & DB_RF_HAVESTRUCT) == 0);
1102 DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *,
1103 db, zio_t *, zio);
1104 cv_wait(&db->db_changed, &db->db_mtx);
1105 }
1106 if (db->db_state == DB_UNCACHED)
1107 err = SET_ERROR(EIO);
1108 }
1109 mutex_exit(&db->db_mtx);
1110 }
1111
1112 ASSERT(err || havepzio || db->db_state == DB_CACHED);
1113 return (err);
1114 }
1115
1116 static void
dbuf_noread(dmu_buf_impl_t * db)1117 dbuf_noread(dmu_buf_impl_t *db)
1118 {
1119 ASSERT(!refcount_is_zero(&db->db_holds));
1120 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1121 mutex_enter(&db->db_mtx);
1122 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1123 cv_wait(&db->db_changed, &db->db_mtx);
1124 if (db->db_state == DB_UNCACHED) {
1125 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1126 spa_t *spa = db->db_objset->os_spa;
1127
1128 ASSERT(db->db_buf == NULL);
1129 ASSERT(db->db.db_data == NULL);
1130 dbuf_set_data(db, arc_alloc_buf(spa, db->db.db_size, db, type));
1131 db->db_state = DB_FILL;
1132 } else if (db->db_state == DB_NOFILL) {
1133 dbuf_clear_data(db);
1134 } else {
1135 ASSERT3U(db->db_state, ==, DB_CACHED);
1136 }
1137 mutex_exit(&db->db_mtx);
1138 }
1139
1140 /*
1141 * This is our just-in-time copy function. It makes a copy of
1142 * buffers, that have been modified in a previous transaction
1143 * group, before we modify them in the current active group.
1144 *
1145 * This function is used in two places: when we are dirtying a
1146 * buffer for the first time in a txg, and when we are freeing
1147 * a range in a dnode that includes this buffer.
1148 *
1149 * Note that when we are called from dbuf_free_range() we do
1150 * not put a hold on the buffer, we just traverse the active
1151 * dbuf list for the dnode.
1152 */
1153 static void
dbuf_fix_old_data(dmu_buf_impl_t * db,uint64_t txg)1154 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
1155 {
1156 dbuf_dirty_record_t *dr = db->db_last_dirty;
1157
1158 ASSERT(MUTEX_HELD(&db->db_mtx));
1159 ASSERT(db->db.db_data != NULL);
1160 ASSERT(db->db_level == 0);
1161 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
1162
1163 if (dr == NULL ||
1164 (dr->dt.dl.dr_data !=
1165 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
1166 return;
1167
1168 /*
1169 * If the last dirty record for this dbuf has not yet synced
1170 * and its referencing the dbuf data, either:
1171 * reset the reference to point to a new copy,
1172 * or (if there a no active holders)
1173 * just null out the current db_data pointer.
1174 */
1175 ASSERT(dr->dr_txg >= txg - 2);
1176 if (db->db_blkid == DMU_BONUS_BLKID) {
1177 /* Note that the data bufs here are zio_bufs */
1178 dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN);
1179 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
1180 bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN);
1181 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
1182 int size = db->db.db_size;
1183 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1184 spa_t *spa = db->db_objset->os_spa;
1185
1186 dr->dt.dl.dr_data = arc_alloc_buf(spa, size, db, type);
1187 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
1188 } else {
1189 db->db_buf = NULL;
1190 dbuf_clear_data(db);
1191 }
1192 }
1193
1194 void
dbuf_unoverride(dbuf_dirty_record_t * dr)1195 dbuf_unoverride(dbuf_dirty_record_t *dr)
1196 {
1197 dmu_buf_impl_t *db = dr->dr_dbuf;
1198 blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
1199 uint64_t txg = dr->dr_txg;
1200
1201 ASSERT(MUTEX_HELD(&db->db_mtx));
1202 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
1203 ASSERT(db->db_level == 0);
1204
1205 if (db->db_blkid == DMU_BONUS_BLKID ||
1206 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
1207 return;
1208
1209 ASSERT(db->db_data_pending != dr);
1210
1211 /* free this block */
1212 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite)
1213 zio_free(db->db_objset->os_spa, txg, bp);
1214
1215 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1216 dr->dt.dl.dr_nopwrite = B_FALSE;
1217
1218 /*
1219 * Release the already-written buffer, so we leave it in
1220 * a consistent dirty state. Note that all callers are
1221 * modifying the buffer, so they will immediately do
1222 * another (redundant) arc_release(). Therefore, leave
1223 * the buf thawed to save the effort of freezing &
1224 * immediately re-thawing it.
1225 */
1226 arc_release(dr->dt.dl.dr_data, db);
1227 }
1228
1229 /*
1230 * Evict (if its unreferenced) or clear (if its referenced) any level-0
1231 * data blocks in the free range, so that any future readers will find
1232 * empty blocks.
1233 */
1234 void
dbuf_free_range(dnode_t * dn,uint64_t start_blkid,uint64_t end_blkid,dmu_tx_t * tx)1235 dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
1236 dmu_tx_t *tx)
1237 {
1238 dmu_buf_impl_t db_search;
1239 dmu_buf_impl_t *db, *db_next;
1240 uint64_t txg = tx->tx_txg;
1241 avl_index_t where;
1242
1243 if (end_blkid > dn->dn_maxblkid &&
1244 !(start_blkid == DMU_SPILL_BLKID || end_blkid == DMU_SPILL_BLKID))
1245 end_blkid = dn->dn_maxblkid;
1246 dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid);
1247
1248 db_search.db_level = 0;
1249 db_search.db_blkid = start_blkid;
1250 db_search.db_state = DB_SEARCH;
1251
1252 mutex_enter(&dn->dn_dbufs_mtx);
1253 db = avl_find(&dn->dn_dbufs, &db_search, &where);
1254 ASSERT3P(db, ==, NULL);
1255
1256 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1257
1258 for (; db != NULL; db = db_next) {
1259 db_next = AVL_NEXT(&dn->dn_dbufs, db);
1260 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1261
1262 if (db->db_level != 0 || db->db_blkid > end_blkid) {
1263 break;
1264 }
1265 ASSERT3U(db->db_blkid, >=, start_blkid);
1266
1267 /* found a level 0 buffer in the range */
1268 mutex_enter(&db->db_mtx);
1269 if (dbuf_undirty(db, tx)) {
1270 /* mutex has been dropped and dbuf destroyed */
1271 continue;
1272 }
1273
1274 if (db->db_state == DB_UNCACHED ||
1275 db->db_state == DB_NOFILL ||
1276 db->db_state == DB_EVICTING) {
1277 ASSERT(db->db.db_data == NULL);
1278 mutex_exit(&db->db_mtx);
1279 continue;
1280 }
1281 if (db->db_state == DB_READ || db->db_state == DB_FILL) {
1282 /* will be handled in dbuf_read_done or dbuf_rele */
1283 db->db_freed_in_flight = TRUE;
1284 mutex_exit(&db->db_mtx);
1285 continue;
1286 }
1287 if (refcount_count(&db->db_holds) == 0) {
1288 ASSERT(db->db_buf);
1289 dbuf_destroy(db);
1290 continue;
1291 }
1292 /* The dbuf is referenced */
1293
1294 if (db->db_last_dirty != NULL) {
1295 dbuf_dirty_record_t *dr = db->db_last_dirty;
1296
1297 if (dr->dr_txg == txg) {
1298 /*
1299 * This buffer is "in-use", re-adjust the file
1300 * size to reflect that this buffer may
1301 * contain new data when we sync.
1302 */
1303 if (db->db_blkid != DMU_SPILL_BLKID &&
1304 db->db_blkid > dn->dn_maxblkid)
1305 dn->dn_maxblkid = db->db_blkid;
1306 dbuf_unoverride(dr);
1307 } else {
1308 /*
1309 * This dbuf is not dirty in the open context.
1310 * Either uncache it (if its not referenced in
1311 * the open context) or reset its contents to
1312 * empty.
1313 */
1314 dbuf_fix_old_data(db, txg);
1315 }
1316 }
1317 /* clear the contents if its cached */
1318 if (db->db_state == DB_CACHED) {
1319 ASSERT(db->db.db_data != NULL);
1320 arc_release(db->db_buf, db);
1321 bzero(db->db.db_data, db->db.db_size);
1322 arc_buf_freeze(db->db_buf);
1323 }
1324
1325 mutex_exit(&db->db_mtx);
1326 }
1327 mutex_exit(&dn->dn_dbufs_mtx);
1328 }
1329
1330 static int
dbuf_block_freeable(dmu_buf_impl_t * db)1331 dbuf_block_freeable(dmu_buf_impl_t *db)
1332 {
1333 dsl_dataset_t *ds = db->db_objset->os_dsl_dataset;
1334 uint64_t birth_txg = 0;
1335
1336 /*
1337 * We don't need any locking to protect db_blkptr:
1338 * If it's syncing, then db_last_dirty will be set
1339 * so we'll ignore db_blkptr.
1340 *
1341 * This logic ensures that only block births for
1342 * filled blocks are considered.
1343 */
1344 ASSERT(MUTEX_HELD(&db->db_mtx));
1345 if (db->db_last_dirty && (db->db_blkptr == NULL ||
1346 !BP_IS_HOLE(db->db_blkptr))) {
1347 birth_txg = db->db_last_dirty->dr_txg;
1348 } else if (db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) {
1349 birth_txg = db->db_blkptr->blk_birth;
1350 }
1351
1352 /*
1353 * If this block don't exist or is in a snapshot, it can't be freed.
1354 * Don't pass the bp to dsl_dataset_block_freeable() since we
1355 * are holding the db_mtx lock and might deadlock if we are
1356 * prefetching a dedup-ed block.
1357 */
1358 if (birth_txg != 0)
1359 return (ds == NULL ||
1360 dsl_dataset_block_freeable(ds, NULL, birth_txg));
1361 else
1362 return (B_FALSE);
1363 }
1364
1365 void
dbuf_new_size(dmu_buf_impl_t * db,int size,dmu_tx_t * tx)1366 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
1367 {
1368 arc_buf_t *buf, *obuf;
1369 int osize = db->db.db_size;
1370 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1371 dnode_t *dn;
1372
1373 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1374
1375 DB_DNODE_ENTER(db);
1376 dn = DB_DNODE(db);
1377
1378 /* XXX does *this* func really need the lock? */
1379 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1380
1381 /*
1382 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1383 * is OK, because there can be no other references to the db
1384 * when we are changing its size, so no concurrent DB_FILL can
1385 * be happening.
1386 */
1387 /*
1388 * XXX we should be doing a dbuf_read, checking the return
1389 * value and returning that up to our callers
1390 */
1391 dmu_buf_will_dirty(&db->db, tx);
1392
1393 /* create the data buffer for the new block */
1394 buf = arc_alloc_buf(dn->dn_objset->os_spa, size, db, type);
1395
1396 /* copy old block data to the new block */
1397 obuf = db->db_buf;
1398 bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
1399 /* zero the remainder */
1400 if (size > osize)
1401 bzero((uint8_t *)buf->b_data + osize, size - osize);
1402
1403 mutex_enter(&db->db_mtx);
1404 dbuf_set_data(db, buf);
1405 arc_buf_destroy(obuf, db);
1406 db->db.db_size = size;
1407
1408 if (db->db_level == 0) {
1409 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1410 db->db_last_dirty->dt.dl.dr_data = buf;
1411 }
1412 mutex_exit(&db->db_mtx);
1413
1414 dnode_willuse_space(dn, size-osize, tx);
1415 DB_DNODE_EXIT(db);
1416 }
1417
1418 void
dbuf_release_bp(dmu_buf_impl_t * db)1419 dbuf_release_bp(dmu_buf_impl_t *db)
1420 {
1421 objset_t *os = db->db_objset;
1422
1423 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1424 ASSERT(arc_released(os->os_phys_buf) ||
1425 list_link_active(&os->os_dsl_dataset->ds_synced_link));
1426 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1427
1428 (void) arc_release(db->db_buf, db);
1429 }
1430
1431 /*
1432 * We already have a dirty record for this TXG, and we are being
1433 * dirtied again.
1434 */
1435 static void
dbuf_redirty(dbuf_dirty_record_t * dr)1436 dbuf_redirty(dbuf_dirty_record_t *dr)
1437 {
1438 dmu_buf_impl_t *db = dr->dr_dbuf;
1439
1440 ASSERT(MUTEX_HELD(&db->db_mtx));
1441
1442 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1443 /*
1444 * If this buffer has already been written out,
1445 * we now need to reset its state.
1446 */
1447 dbuf_unoverride(dr);
1448 if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1449 db->db_state != DB_NOFILL) {
1450 /* Already released on initial dirty, so just thaw. */
1451 ASSERT(arc_released(db->db_buf));
1452 arc_buf_thaw(db->db_buf);
1453 }
1454 }
1455 }
1456
1457 dbuf_dirty_record_t *
dbuf_dirty(dmu_buf_impl_t * db,dmu_tx_t * tx)1458 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1459 {
1460 dnode_t *dn;
1461 objset_t *os;
1462 dbuf_dirty_record_t **drp, *dr;
1463 int drop_struct_lock = FALSE;
1464 boolean_t do_free_accounting = B_FALSE;
1465 int txgoff = tx->tx_txg & TXG_MASK;
1466
1467 ASSERT(tx->tx_txg != 0);
1468 ASSERT(!refcount_is_zero(&db->db_holds));
1469 DMU_TX_DIRTY_BUF(tx, db);
1470
1471 DB_DNODE_ENTER(db);
1472 dn = DB_DNODE(db);
1473 /*
1474 * Shouldn't dirty a regular buffer in syncing context. Private
1475 * objects may be dirtied in syncing context, but only if they
1476 * were already pre-dirtied in open context.
1477 */
1478 #ifdef DEBUG
1479 if (dn->dn_objset->os_dsl_dataset != NULL) {
1480 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1481 RW_READER, FTAG);
1482 }
1483 ASSERT(!dmu_tx_is_syncing(tx) ||
1484 BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1485 DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1486 dn->dn_objset->os_dsl_dataset == NULL);
1487 if (dn->dn_objset->os_dsl_dataset != NULL)
1488 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, FTAG);
1489 #endif
1490 /*
1491 * We make this assert for private objects as well, but after we
1492 * check if we're already dirty. They are allowed to re-dirty
1493 * in syncing context.
1494 */
1495 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1496 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1497 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1498
1499 mutex_enter(&db->db_mtx);
1500 /*
1501 * XXX make this true for indirects too? The problem is that
1502 * transactions created with dmu_tx_create_assigned() from
1503 * syncing context don't bother holding ahead.
1504 */
1505 ASSERT(db->db_level != 0 ||
1506 db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1507 db->db_state == DB_NOFILL);
1508
1509 mutex_enter(&dn->dn_mtx);
1510 /*
1511 * Don't set dirtyctx to SYNC if we're just modifying this as we
1512 * initialize the objset.
1513 */
1514 if (dn->dn_dirtyctx == DN_UNDIRTIED) {
1515 if (dn->dn_objset->os_dsl_dataset != NULL) {
1516 rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1517 RW_READER, FTAG);
1518 }
1519 if (!BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1520 dn->dn_dirtyctx = (dmu_tx_is_syncing(tx) ?
1521 DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1522 ASSERT(dn->dn_dirtyctx_firstset == NULL);
1523 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1524 }
1525 if (dn->dn_objset->os_dsl_dataset != NULL) {
1526 rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock,
1527 FTAG);
1528 }
1529 }
1530 mutex_exit(&dn->dn_mtx);
1531
1532 if (db->db_blkid == DMU_SPILL_BLKID)
1533 dn->dn_have_spill = B_TRUE;
1534
1535 /*
1536 * If this buffer is already dirty, we're done.
1537 */
1538 drp = &db->db_last_dirty;
1539 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1540 db->db.db_object == DMU_META_DNODE_OBJECT);
1541 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1542 drp = &dr->dr_next;
1543 if (dr && dr->dr_txg == tx->tx_txg) {
1544 DB_DNODE_EXIT(db);
1545
1546 dbuf_redirty(dr);
1547 mutex_exit(&db->db_mtx);
1548 return (dr);
1549 }
1550
1551 /*
1552 * Only valid if not already dirty.
1553 */
1554 ASSERT(dn->dn_object == 0 ||
1555 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1556 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1557
1558 ASSERT3U(dn->dn_nlevels, >, db->db_level);
1559 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1560 dn->dn_phys->dn_nlevels > db->db_level ||
1561 dn->dn_next_nlevels[txgoff] > db->db_level ||
1562 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1563 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1564
1565 /*
1566 * We should only be dirtying in syncing context if it's the
1567 * mos or we're initializing the os or it's a special object.
1568 * However, we are allowed to dirty in syncing context provided
1569 * we already dirtied it in open context. Hence we must make
1570 * this assertion only if we're not already dirty.
1571 */
1572 os = dn->dn_objset;
1573 #ifdef DEBUG
1574 if (dn->dn_objset->os_dsl_dataset != NULL)
1575 rrw_enter(&os->os_dsl_dataset->ds_bp_rwlock, RW_READER, FTAG);
1576 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1577 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1578 if (dn->dn_objset->os_dsl_dataset != NULL)
1579 rrw_exit(&os->os_dsl_dataset->ds_bp_rwlock, FTAG);
1580 #endif
1581 ASSERT(db->db.db_size != 0);
1582
1583 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1584
1585 if (db->db_blkid != DMU_BONUS_BLKID) {
1586 /*
1587 * Update the accounting.
1588 * Note: we delay "free accounting" until after we drop
1589 * the db_mtx. This keeps us from grabbing other locks
1590 * (and possibly deadlocking) in bp_get_dsize() while
1591 * also holding the db_mtx.
1592 */
1593 dnode_willuse_space(dn, db->db.db_size, tx);
1594 do_free_accounting = dbuf_block_freeable(db);
1595 }
1596
1597 /*
1598 * If this buffer is dirty in an old transaction group we need
1599 * to make a copy of it so that the changes we make in this
1600 * transaction group won't leak out when we sync the older txg.
1601 */
1602 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1603 if (db->db_level == 0) {
1604 void *data_old = db->db_buf;
1605
1606 if (db->db_state != DB_NOFILL) {
1607 if (db->db_blkid == DMU_BONUS_BLKID) {
1608 dbuf_fix_old_data(db, tx->tx_txg);
1609 data_old = db->db.db_data;
1610 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1611 /*
1612 * Release the data buffer from the cache so
1613 * that we can modify it without impacting
1614 * possible other users of this cached data
1615 * block. Note that indirect blocks and
1616 * private objects are not released until the
1617 * syncing state (since they are only modified
1618 * then).
1619 */
1620 arc_release(db->db_buf, db);
1621 dbuf_fix_old_data(db, tx->tx_txg);
1622 data_old = db->db_buf;
1623 }
1624 ASSERT(data_old != NULL);
1625 }
1626 dr->dt.dl.dr_data = data_old;
1627 } else {
1628 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL);
1629 list_create(&dr->dt.di.dr_children,
1630 sizeof (dbuf_dirty_record_t),
1631 offsetof(dbuf_dirty_record_t, dr_dirty_node));
1632 }
1633 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL)
1634 dr->dr_accounted = db->db.db_size;
1635 dr->dr_dbuf = db;
1636 dr->dr_txg = tx->tx_txg;
1637 dr->dr_next = *drp;
1638 *drp = dr;
1639
1640 /*
1641 * We could have been freed_in_flight between the dbuf_noread
1642 * and dbuf_dirty. We win, as though the dbuf_noread() had
1643 * happened after the free.
1644 */
1645 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1646 db->db_blkid != DMU_SPILL_BLKID) {
1647 mutex_enter(&dn->dn_mtx);
1648 if (dn->dn_free_ranges[txgoff] != NULL) {
1649 range_tree_clear(dn->dn_free_ranges[txgoff],
1650 db->db_blkid, 1);
1651 }
1652 mutex_exit(&dn->dn_mtx);
1653 db->db_freed_in_flight = FALSE;
1654 }
1655
1656 /*
1657 * This buffer is now part of this txg
1658 */
1659 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1660 db->db_dirtycnt += 1;
1661 ASSERT3U(db->db_dirtycnt, <=, 3);
1662
1663 mutex_exit(&db->db_mtx);
1664
1665 if (db->db_blkid == DMU_BONUS_BLKID ||
1666 db->db_blkid == DMU_SPILL_BLKID) {
1667 mutex_enter(&dn->dn_mtx);
1668 ASSERT(!list_link_active(&dr->dr_dirty_node));
1669 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1670 mutex_exit(&dn->dn_mtx);
1671 dnode_setdirty(dn, tx);
1672 DB_DNODE_EXIT(db);
1673 return (dr);
1674 }
1675
1676 /*
1677 * The dn_struct_rwlock prevents db_blkptr from changing
1678 * due to a write from syncing context completing
1679 * while we are running, so we want to acquire it before
1680 * looking at db_blkptr.
1681 */
1682 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1683 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1684 drop_struct_lock = TRUE;
1685 }
1686
1687 if (do_free_accounting) {
1688 blkptr_t *bp = db->db_blkptr;
1689 int64_t willfree = (bp && !BP_IS_HOLE(bp)) ?
1690 bp_get_dsize(os->os_spa, bp) : db->db.db_size;
1691 /*
1692 * This is only a guess -- if the dbuf is dirty
1693 * in a previous txg, we don't know how much
1694 * space it will use on disk yet. We should
1695 * really have the struct_rwlock to access
1696 * db_blkptr, but since this is just a guess,
1697 * it's OK if we get an odd answer.
1698 */
1699 ddt_prefetch(os->os_spa, bp);
1700 dnode_willuse_space(dn, -willfree, tx);
1701 }
1702
1703 if (db->db_level == 0) {
1704 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1705 ASSERT(dn->dn_maxblkid >= db->db_blkid);
1706 }
1707
1708 if (db->db_level+1 < dn->dn_nlevels) {
1709 dmu_buf_impl_t *parent = db->db_parent;
1710 dbuf_dirty_record_t *di;
1711 int parent_held = FALSE;
1712
1713 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1714 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1715
1716 parent = dbuf_hold_level(dn, db->db_level+1,
1717 db->db_blkid >> epbs, FTAG);
1718 ASSERT(parent != NULL);
1719 parent_held = TRUE;
1720 }
1721 if (drop_struct_lock)
1722 rw_exit(&dn->dn_struct_rwlock);
1723 ASSERT3U(db->db_level+1, ==, parent->db_level);
1724 di = dbuf_dirty(parent, tx);
1725 if (parent_held)
1726 dbuf_rele(parent, FTAG);
1727
1728 mutex_enter(&db->db_mtx);
1729 /*
1730 * Since we've dropped the mutex, it's possible that
1731 * dbuf_undirty() might have changed this out from under us.
1732 */
1733 if (db->db_last_dirty == dr ||
1734 dn->dn_object == DMU_META_DNODE_OBJECT) {
1735 mutex_enter(&di->dt.di.dr_mtx);
1736 ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1737 ASSERT(!list_link_active(&dr->dr_dirty_node));
1738 list_insert_tail(&di->dt.di.dr_children, dr);
1739 mutex_exit(&di->dt.di.dr_mtx);
1740 dr->dr_parent = di;
1741 }
1742 mutex_exit(&db->db_mtx);
1743 } else {
1744 ASSERT(db->db_level+1 == dn->dn_nlevels);
1745 ASSERT(db->db_blkid < dn->dn_nblkptr);
1746 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1747 mutex_enter(&dn->dn_mtx);
1748 ASSERT(!list_link_active(&dr->dr_dirty_node));
1749 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1750 mutex_exit(&dn->dn_mtx);
1751 if (drop_struct_lock)
1752 rw_exit(&dn->dn_struct_rwlock);
1753 }
1754
1755 dnode_setdirty(dn, tx);
1756 DB_DNODE_EXIT(db);
1757 return (dr);
1758 }
1759
1760 /*
1761 * Undirty a buffer in the transaction group referenced by the given
1762 * transaction. Return whether this evicted the dbuf.
1763 */
1764 static boolean_t
dbuf_undirty(dmu_buf_impl_t * db,dmu_tx_t * tx)1765 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1766 {
1767 dnode_t *dn;
1768 uint64_t txg = tx->tx_txg;
1769 dbuf_dirty_record_t *dr, **drp;
1770
1771 ASSERT(txg != 0);
1772
1773 /*
1774 * Due to our use of dn_nlevels below, this can only be called
1775 * in open context, unless we are operating on the MOS.
1776 * From syncing context, dn_nlevels may be different from the
1777 * dn_nlevels used when dbuf was dirtied.
1778 */
1779 ASSERT(db->db_objset ==
1780 dmu_objset_pool(db->db_objset)->dp_meta_objset ||
1781 txg != spa_syncing_txg(dmu_objset_spa(db->db_objset)));
1782 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1783 ASSERT0(db->db_level);
1784 ASSERT(MUTEX_HELD(&db->db_mtx));
1785
1786 /*
1787 * If this buffer is not dirty, we're done.
1788 */
1789 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1790 if (dr->dr_txg <= txg)
1791 break;
1792 if (dr == NULL || dr->dr_txg < txg)
1793 return (B_FALSE);
1794 ASSERT(dr->dr_txg == txg);
1795 ASSERT(dr->dr_dbuf == db);
1796
1797 DB_DNODE_ENTER(db);
1798 dn = DB_DNODE(db);
1799
1800 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1801
1802 ASSERT(db->db.db_size != 0);
1803
1804 dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset),
1805 dr->dr_accounted, txg);
1806
1807 *drp = dr->dr_next;
1808
1809 /*
1810 * Note that there are three places in dbuf_dirty()
1811 * where this dirty record may be put on a list.
1812 * Make sure to do a list_remove corresponding to
1813 * every one of those list_insert calls.
1814 */
1815 if (dr->dr_parent) {
1816 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1817 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1818 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1819 } else if (db->db_blkid == DMU_SPILL_BLKID ||
1820 db->db_level + 1 == dn->dn_nlevels) {
1821 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1822 mutex_enter(&dn->dn_mtx);
1823 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1824 mutex_exit(&dn->dn_mtx);
1825 }
1826 DB_DNODE_EXIT(db);
1827
1828 if (db->db_state != DB_NOFILL) {
1829 dbuf_unoverride(dr);
1830
1831 ASSERT(db->db_buf != NULL);
1832 ASSERT(dr->dt.dl.dr_data != NULL);
1833 if (dr->dt.dl.dr_data != db->db_buf)
1834 arc_buf_destroy(dr->dt.dl.dr_data, db);
1835 }
1836
1837 kmem_free(dr, sizeof (dbuf_dirty_record_t));
1838
1839 ASSERT(db->db_dirtycnt > 0);
1840 db->db_dirtycnt -= 1;
1841
1842 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1843 ASSERT(db->db_state == DB_NOFILL || arc_released(db->db_buf));
1844 dbuf_destroy(db);
1845 return (B_TRUE);
1846 }
1847
1848 return (B_FALSE);
1849 }
1850
1851 void
dmu_buf_will_dirty(dmu_buf_t * db_fake,dmu_tx_t * tx)1852 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
1853 {
1854 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1855 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1856
1857 ASSERT(tx->tx_txg != 0);
1858 ASSERT(!refcount_is_zero(&db->db_holds));
1859
1860 /*
1861 * Quick check for dirtyness. For already dirty blocks, this
1862 * reduces runtime of this function by >90%, and overall performance
1863 * by 50% for some workloads (e.g. file deletion with indirect blocks
1864 * cached).
1865 */
1866 mutex_enter(&db->db_mtx);
1867 dbuf_dirty_record_t *dr;
1868 for (dr = db->db_last_dirty;
1869 dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) {
1870 /*
1871 * It's possible that it is already dirty but not cached,
1872 * because there are some calls to dbuf_dirty() that don't
1873 * go through dmu_buf_will_dirty().
1874 */
1875 if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) {
1876 /* This dbuf is already dirty and cached. */
1877 dbuf_redirty(dr);
1878 mutex_exit(&db->db_mtx);
1879 return;
1880 }
1881 }
1882 mutex_exit(&db->db_mtx);
1883
1884 DB_DNODE_ENTER(db);
1885 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
1886 rf |= DB_RF_HAVESTRUCT;
1887 DB_DNODE_EXIT(db);
1888 (void) dbuf_read(db, NULL, rf);
1889 (void) dbuf_dirty(db, tx);
1890 }
1891
1892 void
dmu_buf_will_not_fill(dmu_buf_t * db_fake,dmu_tx_t * tx)1893 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1894 {
1895 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1896
1897 db->db_state = DB_NOFILL;
1898
1899 dmu_buf_will_fill(db_fake, tx);
1900 }
1901
1902 void
dmu_buf_will_fill(dmu_buf_t * db_fake,dmu_tx_t * tx)1903 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1904 {
1905 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1906
1907 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1908 ASSERT(tx->tx_txg != 0);
1909 ASSERT(db->db_level == 0);
1910 ASSERT(!refcount_is_zero(&db->db_holds));
1911
1912 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
1913 dmu_tx_private_ok(tx));
1914
1915 dbuf_noread(db);
1916 (void) dbuf_dirty(db, tx);
1917 }
1918
1919 #pragma weak dmu_buf_fill_done = dbuf_fill_done
1920 /* ARGSUSED */
1921 void
dbuf_fill_done(dmu_buf_impl_t * db,dmu_tx_t * tx)1922 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
1923 {
1924 mutex_enter(&db->db_mtx);
1925 DBUF_VERIFY(db);
1926
1927 if (db->db_state == DB_FILL) {
1928 if (db->db_level == 0 && db->db_freed_in_flight) {
1929 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1930 /* we were freed while filling */
1931 /* XXX dbuf_undirty? */
1932 bzero(db->db.db_data, db->db.db_size);
1933 db->db_freed_in_flight = FALSE;
1934 }
1935 db->db_state = DB_CACHED;
1936 cv_broadcast(&db->db_changed);
1937 }
1938 mutex_exit(&db->db_mtx);
1939 }
1940
1941 void
dmu_buf_write_embedded(dmu_buf_t * dbuf,void * data,bp_embedded_type_t etype,enum zio_compress comp,int uncompressed_size,int compressed_size,int byteorder,dmu_tx_t * tx)1942 dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data,
1943 bp_embedded_type_t etype, enum zio_compress comp,
1944 int uncompressed_size, int compressed_size, int byteorder,
1945 dmu_tx_t *tx)
1946 {
1947 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
1948 struct dirty_leaf *dl;
1949 dmu_object_type_t type;
1950
1951 if (etype == BP_EMBEDDED_TYPE_DATA) {
1952 ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset),
1953 SPA_FEATURE_EMBEDDED_DATA));
1954 }
1955
1956 DB_DNODE_ENTER(db);
1957 type = DB_DNODE(db)->dn_type;
1958 DB_DNODE_EXIT(db);
1959
1960 ASSERT0(db->db_level);
1961 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1962
1963 dmu_buf_will_not_fill(dbuf, tx);
1964
1965 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1966 dl = &db->db_last_dirty->dt.dl;
1967 encode_embedded_bp_compressed(&dl->dr_overridden_by,
1968 data, comp, uncompressed_size, compressed_size);
1969 BPE_SET_ETYPE(&dl->dr_overridden_by, etype);
1970 BP_SET_TYPE(&dl->dr_overridden_by, type);
1971 BP_SET_LEVEL(&dl->dr_overridden_by, 0);
1972 BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder);
1973
1974 dl->dr_override_state = DR_OVERRIDDEN;
1975 dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg;
1976 }
1977
1978 /*
1979 * Directly assign a provided arc buf to a given dbuf if it's not referenced
1980 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
1981 */
1982 void
dbuf_assign_arcbuf(dmu_buf_impl_t * db,arc_buf_t * buf,dmu_tx_t * tx)1983 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
1984 {
1985 ASSERT(!refcount_is_zero(&db->db_holds));
1986 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1987 ASSERT(db->db_level == 0);
1988 ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA);
1989 ASSERT(buf != NULL);
1990 ASSERT(arc_buf_size(buf) == db->db.db_size);
1991 ASSERT(tx->tx_txg != 0);
1992
1993 arc_return_buf(buf, db);
1994 ASSERT(arc_released(buf));
1995
1996 mutex_enter(&db->db_mtx);
1997
1998 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1999 cv_wait(&db->db_changed, &db->db_mtx);
2000
2001 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
2002
2003 if (db->db_state == DB_CACHED &&
2004 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
2005 mutex_exit(&db->db_mtx);
2006 (void) dbuf_dirty(db, tx);
2007 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
2008 arc_buf_destroy(buf, db);
2009 xuio_stat_wbuf_copied();
2010 return;
2011 }
2012
2013 xuio_stat_wbuf_nocopy();
2014 if (db->db_state == DB_CACHED) {
2015 dbuf_dirty_record_t *dr = db->db_last_dirty;
2016
2017 ASSERT(db->db_buf != NULL);
2018 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
2019 ASSERT(dr->dt.dl.dr_data == db->db_buf);
2020 if (!arc_released(db->db_buf)) {
2021 ASSERT(dr->dt.dl.dr_override_state ==
2022 DR_OVERRIDDEN);
2023 arc_release(db->db_buf, db);
2024 }
2025 dr->dt.dl.dr_data = buf;
2026 arc_buf_destroy(db->db_buf, db);
2027 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
2028 arc_release(db->db_buf, db);
2029 arc_buf_destroy(db->db_buf, db);
2030 }
2031 db->db_buf = NULL;
2032 }
2033 ASSERT(db->db_buf == NULL);
2034 dbuf_set_data(db, buf);
2035 db->db_state = DB_FILL;
2036 mutex_exit(&db->db_mtx);
2037 (void) dbuf_dirty(db, tx);
2038 dmu_buf_fill_done(&db->db, tx);
2039 }
2040
2041 void
dbuf_destroy(dmu_buf_impl_t * db)2042 dbuf_destroy(dmu_buf_impl_t *db)
2043 {
2044 dnode_t *dn;
2045 dmu_buf_impl_t *parent = db->db_parent;
2046 dmu_buf_impl_t *dndb;
2047
2048 ASSERT(MUTEX_HELD(&db->db_mtx));
2049 ASSERT(refcount_is_zero(&db->db_holds));
2050
2051 if (db->db_buf != NULL) {
2052 arc_buf_destroy(db->db_buf, db);
2053 db->db_buf = NULL;
2054 }
2055
2056 if (db->db_blkid == DMU_BONUS_BLKID) {
2057 ASSERT(db->db.db_data != NULL);
2058 zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN);
2059 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
2060 db->db_state = DB_UNCACHED;
2061 }
2062
2063 dbuf_clear_data(db);
2064
2065 if (multilist_link_active(&db->db_cache_link)) {
2066 multilist_remove(&dbuf_cache, db);
2067 (void) refcount_remove_many(&dbuf_cache_size,
2068 db->db.db_size, db);
2069 }
2070
2071 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
2072 ASSERT(db->db_data_pending == NULL);
2073
2074 db->db_state = DB_EVICTING;
2075 db->db_blkptr = NULL;
2076
2077 /*
2078 * Now that db_state is DB_EVICTING, nobody else can find this via
2079 * the hash table. We can now drop db_mtx, which allows us to
2080 * acquire the dn_dbufs_mtx.
2081 */
2082 mutex_exit(&db->db_mtx);
2083
2084 DB_DNODE_ENTER(db);
2085 dn = DB_DNODE(db);
2086 dndb = dn->dn_dbuf;
2087 if (db->db_blkid != DMU_BONUS_BLKID) {
2088 boolean_t needlock = !MUTEX_HELD(&dn->dn_dbufs_mtx);
2089 if (needlock)
2090 mutex_enter(&dn->dn_dbufs_mtx);
2091 avl_remove(&dn->dn_dbufs, db);
2092 atomic_dec_32(&dn->dn_dbufs_count);
2093 membar_producer();
2094 DB_DNODE_EXIT(db);
2095 if (needlock)
2096 mutex_exit(&dn->dn_dbufs_mtx);
2097 /*
2098 * Decrementing the dbuf count means that the hold corresponding
2099 * to the removed dbuf is no longer discounted in dnode_move(),
2100 * so the dnode cannot be moved until after we release the hold.
2101 * The membar_producer() ensures visibility of the decremented
2102 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
2103 * release any lock.
2104 */
2105 dnode_rele(dn, db);
2106 db->db_dnode_handle = NULL;
2107
2108 dbuf_hash_remove(db);
2109 } else {
2110 DB_DNODE_EXIT(db);
2111 }
2112
2113 ASSERT(refcount_is_zero(&db->db_holds));
2114
2115 db->db_parent = NULL;
2116
2117 ASSERT(db->db_buf == NULL);
2118 ASSERT(db->db.db_data == NULL);
2119 ASSERT(db->db_hash_next == NULL);
2120 ASSERT(db->db_blkptr == NULL);
2121 ASSERT(db->db_data_pending == NULL);
2122 ASSERT(!multilist_link_active(&db->db_cache_link));
2123
2124 kmem_cache_free(dbuf_kmem_cache, db);
2125 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2126
2127 /*
2128 * If this dbuf is referenced from an indirect dbuf,
2129 * decrement the ref count on the indirect dbuf.
2130 */
2131 if (parent && parent != dndb)
2132 dbuf_rele(parent, db);
2133 }
2134
2135 /*
2136 * Note: While bpp will always be updated if the function returns success,
2137 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
2138 * this happens when the dnode is the meta-dnode, or a userused or groupused
2139 * object.
2140 */
2141 static int
dbuf_findbp(dnode_t * dn,int level,uint64_t blkid,int fail_sparse,dmu_buf_impl_t ** parentp,blkptr_t ** bpp)2142 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
2143 dmu_buf_impl_t **parentp, blkptr_t **bpp)
2144 {
2145 int nlevels, epbs;
2146
2147 *parentp = NULL;
2148 *bpp = NULL;
2149
2150 ASSERT(blkid != DMU_BONUS_BLKID);
2151
2152 if (blkid == DMU_SPILL_BLKID) {
2153 mutex_enter(&dn->dn_mtx);
2154 if (dn->dn_have_spill &&
2155 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
2156 *bpp = &dn->dn_phys->dn_spill;
2157 else
2158 *bpp = NULL;
2159 dbuf_add_ref(dn->dn_dbuf, NULL);
2160 *parentp = dn->dn_dbuf;
2161 mutex_exit(&dn->dn_mtx);
2162 return (0);
2163 }
2164
2165 if (dn->dn_phys->dn_nlevels == 0)
2166 nlevels = 1;
2167 else
2168 nlevels = dn->dn_phys->dn_nlevels;
2169
2170 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2171
2172 ASSERT3U(level * epbs, <, 64);
2173 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2174 if (level >= nlevels ||
2175 (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
2176 /* the buffer has no parent yet */
2177 return (SET_ERROR(ENOENT));
2178 } else if (level < nlevels-1) {
2179 /* this block is referenced from an indirect block */
2180 int err = dbuf_hold_impl(dn, level+1,
2181 blkid >> epbs, fail_sparse, FALSE, NULL, parentp);
2182 if (err)
2183 return (err);
2184 err = dbuf_read(*parentp, NULL,
2185 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
2186 if (err) {
2187 dbuf_rele(*parentp, NULL);
2188 *parentp = NULL;
2189 return (err);
2190 }
2191 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
2192 (blkid & ((1ULL << epbs) - 1));
2193 return (0);
2194 } else {
2195 /* the block is referenced from the dnode */
2196 ASSERT3U(level, ==, nlevels-1);
2197 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
2198 blkid < dn->dn_phys->dn_nblkptr);
2199 if (dn->dn_dbuf) {
2200 dbuf_add_ref(dn->dn_dbuf, NULL);
2201 *parentp = dn->dn_dbuf;
2202 }
2203 *bpp = &dn->dn_phys->dn_blkptr[blkid];
2204 return (0);
2205 }
2206 }
2207
2208 static dmu_buf_impl_t *
dbuf_create(dnode_t * dn,uint8_t level,uint64_t blkid,dmu_buf_impl_t * parent,blkptr_t * blkptr)2209 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
2210 dmu_buf_impl_t *parent, blkptr_t *blkptr)
2211 {
2212 objset_t *os = dn->dn_objset;
2213 dmu_buf_impl_t *db, *odb;
2214
2215 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2216 ASSERT(dn->dn_type != DMU_OT_NONE);
2217
2218 db = kmem_cache_alloc(dbuf_kmem_cache, KM_SLEEP);
2219
2220 db->db_objset = os;
2221 db->db.db_object = dn->dn_object;
2222 db->db_level = level;
2223 db->db_blkid = blkid;
2224 db->db_last_dirty = NULL;
2225 db->db_dirtycnt = 0;
2226 db->db_dnode_handle = dn->dn_handle;
2227 db->db_parent = parent;
2228 db->db_blkptr = blkptr;
2229
2230 db->db_user = NULL;
2231 db->db_user_immediate_evict = FALSE;
2232 db->db_freed_in_flight = FALSE;
2233 db->db_pending_evict = FALSE;
2234
2235 if (blkid == DMU_BONUS_BLKID) {
2236 ASSERT3P(parent, ==, dn->dn_dbuf);
2237 db->db.db_size = DN_MAX_BONUSLEN -
2238 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
2239 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
2240 db->db.db_offset = DMU_BONUS_BLKID;
2241 db->db_state = DB_UNCACHED;
2242 /* the bonus dbuf is not placed in the hash table */
2243 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2244 return (db);
2245 } else if (blkid == DMU_SPILL_BLKID) {
2246 db->db.db_size = (blkptr != NULL) ?
2247 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
2248 db->db.db_offset = 0;
2249 } else {
2250 int blocksize =
2251 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
2252 db->db.db_size = blocksize;
2253 db->db.db_offset = db->db_blkid * blocksize;
2254 }
2255
2256 /*
2257 * Hold the dn_dbufs_mtx while we get the new dbuf
2258 * in the hash table *and* added to the dbufs list.
2259 * This prevents a possible deadlock with someone
2260 * trying to look up this dbuf before its added to the
2261 * dn_dbufs list.
2262 */
2263 mutex_enter(&dn->dn_dbufs_mtx);
2264 db->db_state = DB_EVICTING;
2265 if ((odb = dbuf_hash_insert(db)) != NULL) {
2266 /* someone else inserted it first */
2267 kmem_cache_free(dbuf_kmem_cache, db);
2268 mutex_exit(&dn->dn_dbufs_mtx);
2269 return (odb);
2270 }
2271 avl_add(&dn->dn_dbufs, db);
2272
2273 db->db_state = DB_UNCACHED;
2274 mutex_exit(&dn->dn_dbufs_mtx);
2275 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER);
2276
2277 if (parent && parent != dn->dn_dbuf)
2278 dbuf_add_ref(parent, db);
2279
2280 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
2281 refcount_count(&dn->dn_holds) > 0);
2282 (void) refcount_add(&dn->dn_holds, db);
2283 atomic_inc_32(&dn->dn_dbufs_count);
2284
2285 dprintf_dbuf(db, "db=%p\n", db);
2286
2287 return (db);
2288 }
2289
2290 typedef struct dbuf_prefetch_arg {
2291 spa_t *dpa_spa; /* The spa to issue the prefetch in. */
2292 zbookmark_phys_t dpa_zb; /* The target block to prefetch. */
2293 int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */
2294 int dpa_curlevel; /* The current level that we're reading */
2295 dnode_t *dpa_dnode; /* The dnode associated with the prefetch */
2296 zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */
2297 zio_t *dpa_zio; /* The parent zio_t for all prefetches. */
2298 arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */
2299 } dbuf_prefetch_arg_t;
2300
2301 /*
2302 * Actually issue the prefetch read for the block given.
2303 */
2304 static void
dbuf_issue_final_prefetch(dbuf_prefetch_arg_t * dpa,blkptr_t * bp)2305 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp)
2306 {
2307 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
2308 return;
2309
2310 arc_flags_t aflags =
2311 dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
2312
2313 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2314 ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level);
2315 ASSERT(dpa->dpa_zio != NULL);
2316 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL,
2317 dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2318 &aflags, &dpa->dpa_zb);
2319 }
2320
2321 /*
2322 * Called when an indirect block above our prefetch target is read in. This
2323 * will either read in the next indirect block down the tree or issue the actual
2324 * prefetch if the next block down is our target.
2325 */
2326 static void
dbuf_prefetch_indirect_done(zio_t * zio,arc_buf_t * abuf,void * private)2327 dbuf_prefetch_indirect_done(zio_t *zio, arc_buf_t *abuf, void *private)
2328 {
2329 dbuf_prefetch_arg_t *dpa = private;
2330
2331 ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel);
2332 ASSERT3S(dpa->dpa_curlevel, >, 0);
2333
2334 /*
2335 * The dpa_dnode is only valid if we are called with a NULL
2336 * zio. This indicates that the arc_read() returned without
2337 * first calling zio_read() to issue a physical read. Once
2338 * a physical read is made the dpa_dnode must be invalidated
2339 * as the locks guarding it may have been dropped. If the
2340 * dpa_dnode is still valid, then we want to add it to the dbuf
2341 * cache. To do so, we must hold the dbuf associated with the block
2342 * we just prefetched, read its contents so that we associate it
2343 * with an arc_buf_t, and then release it.
2344 */
2345 if (zio != NULL) {
2346 ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel);
2347 if (zio->io_flags & ZIO_FLAG_RAW) {
2348 ASSERT3U(BP_GET_PSIZE(zio->io_bp), ==, zio->io_size);
2349 } else {
2350 ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size);
2351 }
2352 ASSERT3P(zio->io_spa, ==, dpa->dpa_spa);
2353
2354 dpa->dpa_dnode = NULL;
2355 } else if (dpa->dpa_dnode != NULL) {
2356 uint64_t curblkid = dpa->dpa_zb.zb_blkid >>
2357 (dpa->dpa_epbs * (dpa->dpa_curlevel -
2358 dpa->dpa_zb.zb_level));
2359 dmu_buf_impl_t *db = dbuf_hold_level(dpa->dpa_dnode,
2360 dpa->dpa_curlevel, curblkid, FTAG);
2361 (void) dbuf_read(db, NULL,
2362 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_HAVESTRUCT);
2363 dbuf_rele(db, FTAG);
2364 }
2365
2366 dpa->dpa_curlevel--;
2367
2368 uint64_t nextblkid = dpa->dpa_zb.zb_blkid >>
2369 (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level));
2370 blkptr_t *bp = ((blkptr_t *)abuf->b_data) +
2371 P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs);
2372 if (BP_IS_HOLE(bp) || (zio != NULL && zio->io_error != 0)) {
2373 kmem_free(dpa, sizeof (*dpa));
2374 } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) {
2375 ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid);
2376 dbuf_issue_final_prefetch(dpa, bp);
2377 kmem_free(dpa, sizeof (*dpa));
2378 } else {
2379 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2380 zbookmark_phys_t zb;
2381
2382 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2383
2384 SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset,
2385 dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid);
2386
2387 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2388 bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio,
2389 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2390 &iter_aflags, &zb);
2391 }
2392
2393 arc_buf_destroy(abuf, private);
2394 }
2395
2396 /*
2397 * Issue prefetch reads for the given block on the given level. If the indirect
2398 * blocks above that block are not in memory, we will read them in
2399 * asynchronously. As a result, this call never blocks waiting for a read to
2400 * complete.
2401 */
2402 void
dbuf_prefetch(dnode_t * dn,int64_t level,uint64_t blkid,zio_priority_t prio,arc_flags_t aflags)2403 dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio,
2404 arc_flags_t aflags)
2405 {
2406 blkptr_t bp;
2407 int epbs, nlevels, curlevel;
2408 uint64_t curblkid;
2409
2410 ASSERT(blkid != DMU_BONUS_BLKID);
2411 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2412
2413 if (blkid > dn->dn_maxblkid)
2414 return;
2415
2416 if (dnode_block_freed(dn, blkid))
2417 return;
2418
2419 /*
2420 * This dnode hasn't been written to disk yet, so there's nothing to
2421 * prefetch.
2422 */
2423 nlevels = dn->dn_phys->dn_nlevels;
2424 if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0)
2425 return;
2426
2427 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2428 if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level))
2429 return;
2430
2431 dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object,
2432 level, blkid);
2433 if (db != NULL) {
2434 mutex_exit(&db->db_mtx);
2435 /*
2436 * This dbuf already exists. It is either CACHED, or
2437 * (we assume) about to be read or filled.
2438 */
2439 return;
2440 }
2441
2442 /*
2443 * Find the closest ancestor (indirect block) of the target block
2444 * that is present in the cache. In this indirect block, we will
2445 * find the bp that is at curlevel, curblkid.
2446 */
2447 curlevel = level;
2448 curblkid = blkid;
2449 while (curlevel < nlevels - 1) {
2450 int parent_level = curlevel + 1;
2451 uint64_t parent_blkid = curblkid >> epbs;
2452 dmu_buf_impl_t *db;
2453
2454 if (dbuf_hold_impl(dn, parent_level, parent_blkid,
2455 FALSE, TRUE, FTAG, &db) == 0) {
2456 blkptr_t *bpp = db->db_buf->b_data;
2457 bp = bpp[P2PHASE(curblkid, 1 << epbs)];
2458 dbuf_rele(db, FTAG);
2459 break;
2460 }
2461
2462 curlevel = parent_level;
2463 curblkid = parent_blkid;
2464 }
2465
2466 if (curlevel == nlevels - 1) {
2467 /* No cached indirect blocks found. */
2468 ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr);
2469 bp = dn->dn_phys->dn_blkptr[curblkid];
2470 }
2471 if (BP_IS_HOLE(&bp))
2472 return;
2473
2474 ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp));
2475
2476 zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL,
2477 ZIO_FLAG_CANFAIL);
2478
2479 dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP);
2480 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset;
2481 SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2482 dn->dn_object, level, blkid);
2483 dpa->dpa_curlevel = curlevel;
2484 dpa->dpa_prio = prio;
2485 dpa->dpa_aflags = aflags;
2486 dpa->dpa_spa = dn->dn_objset->os_spa;
2487 dpa->dpa_dnode = dn;
2488 dpa->dpa_epbs = epbs;
2489 dpa->dpa_zio = pio;
2490
2491 /*
2492 * If we have the indirect just above us, no need to do the asynchronous
2493 * prefetch chain; we'll just run the last step ourselves. If we're at
2494 * a higher level, though, we want to issue the prefetches for all the
2495 * indirect blocks asynchronously, so we can go on with whatever we were
2496 * doing.
2497 */
2498 if (curlevel == level) {
2499 ASSERT3U(curblkid, ==, blkid);
2500 dbuf_issue_final_prefetch(dpa, &bp);
2501 kmem_free(dpa, sizeof (*dpa));
2502 } else {
2503 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2504 zbookmark_phys_t zb;
2505
2506 SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2507 dn->dn_object, curlevel, curblkid);
2508 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2509 &bp, dbuf_prefetch_indirect_done, dpa, prio,
2510 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2511 &iter_aflags, &zb);
2512 }
2513 /*
2514 * We use pio here instead of dpa_zio since it's possible that
2515 * dpa may have already been freed.
2516 */
2517 zio_nowait(pio);
2518 }
2519
2520 /*
2521 * Returns with db_holds incremented, and db_mtx not held.
2522 * Note: dn_struct_rwlock must be held.
2523 */
2524 int
dbuf_hold_impl(dnode_t * dn,uint8_t level,uint64_t blkid,boolean_t fail_sparse,boolean_t fail_uncached,void * tag,dmu_buf_impl_t ** dbp)2525 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid,
2526 boolean_t fail_sparse, boolean_t fail_uncached,
2527 void *tag, dmu_buf_impl_t **dbp)
2528 {
2529 dmu_buf_impl_t *db, *parent = NULL;
2530
2531 ASSERT(blkid != DMU_BONUS_BLKID);
2532 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2533 ASSERT3U(dn->dn_nlevels, >, level);
2534
2535 *dbp = NULL;
2536 top:
2537 /* dbuf_find() returns with db_mtx held */
2538 db = dbuf_find(dn->dn_objset, dn->dn_object, level, blkid);
2539
2540 if (db == NULL) {
2541 blkptr_t *bp = NULL;
2542 int err;
2543
2544 if (fail_uncached)
2545 return (SET_ERROR(ENOENT));
2546
2547 ASSERT3P(parent, ==, NULL);
2548 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp);
2549 if (fail_sparse) {
2550 if (err == 0 && bp && BP_IS_HOLE(bp))
2551 err = SET_ERROR(ENOENT);
2552 if (err) {
2553 if (parent)
2554 dbuf_rele(parent, NULL);
2555 return (err);
2556 }
2557 }
2558 if (err && err != ENOENT)
2559 return (err);
2560 db = dbuf_create(dn, level, blkid, parent, bp);
2561 }
2562
2563 if (fail_uncached && db->db_state != DB_CACHED) {
2564 mutex_exit(&db->db_mtx);
2565 return (SET_ERROR(ENOENT));
2566 }
2567
2568 if (db->db_buf != NULL)
2569 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data);
2570
2571 ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf));
2572
2573 /*
2574 * If this buffer is currently syncing out, and we are are
2575 * still referencing it from db_data, we need to make a copy
2576 * of it in case we decide we want to dirty it again in this txg.
2577 */
2578 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
2579 dn->dn_object != DMU_META_DNODE_OBJECT &&
2580 db->db_state == DB_CACHED && db->db_data_pending) {
2581 dbuf_dirty_record_t *dr = db->db_data_pending;
2582
2583 if (dr->dt.dl.dr_data == db->db_buf) {
2584 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
2585
2586 dbuf_set_data(db,
2587 arc_alloc_buf(dn->dn_objset->os_spa,
2588 db->db.db_size, db, type));
2589 bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data,
2590 db->db.db_size);
2591 }
2592 }
2593
2594 if (multilist_link_active(&db->db_cache_link)) {
2595 ASSERT(refcount_is_zero(&db->db_holds));
2596 multilist_remove(&dbuf_cache, db);
2597 (void) refcount_remove_many(&dbuf_cache_size,
2598 db->db.db_size, db);
2599 }
2600 (void) refcount_add(&db->db_holds, tag);
2601 DBUF_VERIFY(db);
2602 mutex_exit(&db->db_mtx);
2603
2604 /* NOTE: we can't rele the parent until after we drop the db_mtx */
2605 if (parent)
2606 dbuf_rele(parent, NULL);
2607
2608 ASSERT3P(DB_DNODE(db), ==, dn);
2609 ASSERT3U(db->db_blkid, ==, blkid);
2610 ASSERT3U(db->db_level, ==, level);
2611 *dbp = db;
2612
2613 return (0);
2614 }
2615
2616 dmu_buf_impl_t *
dbuf_hold(dnode_t * dn,uint64_t blkid,void * tag)2617 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
2618 {
2619 return (dbuf_hold_level(dn, 0, blkid, tag));
2620 }
2621
2622 dmu_buf_impl_t *
dbuf_hold_level(dnode_t * dn,int level,uint64_t blkid,void * tag)2623 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
2624 {
2625 dmu_buf_impl_t *db;
2626 int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db);
2627 return (err ? NULL : db);
2628 }
2629
2630 void
dbuf_create_bonus(dnode_t * dn)2631 dbuf_create_bonus(dnode_t *dn)
2632 {
2633 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
2634
2635 ASSERT(dn->dn_bonus == NULL);
2636 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
2637 }
2638
2639 int
dbuf_spill_set_blksz(dmu_buf_t * db_fake,uint64_t blksz,dmu_tx_t * tx)2640 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
2641 {
2642 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2643 dnode_t *dn;
2644
2645 if (db->db_blkid != DMU_SPILL_BLKID)
2646 return (SET_ERROR(ENOTSUP));
2647 if (blksz == 0)
2648 blksz = SPA_MINBLOCKSIZE;
2649 ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset)));
2650 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
2651
2652 DB_DNODE_ENTER(db);
2653 dn = DB_DNODE(db);
2654 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2655 dbuf_new_size(db, blksz, tx);
2656 rw_exit(&dn->dn_struct_rwlock);
2657 DB_DNODE_EXIT(db);
2658
2659 return (0);
2660 }
2661
2662 void
dbuf_rm_spill(dnode_t * dn,dmu_tx_t * tx)2663 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
2664 {
2665 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
2666 }
2667
2668 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2669 void
dbuf_add_ref(dmu_buf_impl_t * db,void * tag)2670 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
2671 {
2672 int64_t holds = refcount_add(&db->db_holds, tag);
2673 ASSERT3S(holds, >, 1);
2674 }
2675
2676 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
2677 boolean_t
dbuf_try_add_ref(dmu_buf_t * db_fake,objset_t * os,uint64_t obj,uint64_t blkid,void * tag)2678 dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid,
2679 void *tag)
2680 {
2681 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2682 dmu_buf_impl_t *found_db;
2683 boolean_t result = B_FALSE;
2684
2685 if (db->db_blkid == DMU_BONUS_BLKID)
2686 found_db = dbuf_find_bonus(os, obj);
2687 else
2688 found_db = dbuf_find(os, obj, 0, blkid);
2689
2690 if (found_db != NULL) {
2691 if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) {
2692 (void) refcount_add(&db->db_holds, tag);
2693 result = B_TRUE;
2694 }
2695 mutex_exit(&db->db_mtx);
2696 }
2697 return (result);
2698 }
2699
2700 /*
2701 * If you call dbuf_rele() you had better not be referencing the dnode handle
2702 * unless you have some other direct or indirect hold on the dnode. (An indirect
2703 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2704 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2705 * dnode's parent dbuf evicting its dnode handles.
2706 */
2707 void
dbuf_rele(dmu_buf_impl_t * db,void * tag)2708 dbuf_rele(dmu_buf_impl_t *db, void *tag)
2709 {
2710 mutex_enter(&db->db_mtx);
2711 dbuf_rele_and_unlock(db, tag);
2712 }
2713
2714 void
dmu_buf_rele(dmu_buf_t * db,void * tag)2715 dmu_buf_rele(dmu_buf_t *db, void *tag)
2716 {
2717 dbuf_rele((dmu_buf_impl_t *)db, tag);
2718 }
2719
2720 /*
2721 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2722 * db_dirtycnt and db_holds to be updated atomically.
2723 */
2724 void
dbuf_rele_and_unlock(dmu_buf_impl_t * db,void * tag)2725 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2726 {
2727 int64_t holds;
2728
2729 ASSERT(MUTEX_HELD(&db->db_mtx));
2730 DBUF_VERIFY(db);
2731
2732 /*
2733 * Remove the reference to the dbuf before removing its hold on the
2734 * dnode so we can guarantee in dnode_move() that a referenced bonus
2735 * buffer has a corresponding dnode hold.
2736 */
2737 holds = refcount_remove(&db->db_holds, tag);
2738 ASSERT(holds >= 0);
2739
2740 /*
2741 * We can't freeze indirects if there is a possibility that they
2742 * may be modified in the current syncing context.
2743 */
2744 if (db->db_buf != NULL &&
2745 holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) {
2746 arc_buf_freeze(db->db_buf);
2747 }
2748
2749 if (holds == db->db_dirtycnt &&
2750 db->db_level == 0 && db->db_user_immediate_evict)
2751 dbuf_evict_user(db);
2752
2753 if (holds == 0) {
2754 if (db->db_blkid == DMU_BONUS_BLKID) {
2755 dnode_t *dn;
2756 boolean_t evict_dbuf = db->db_pending_evict;
2757
2758 /*
2759 * If the dnode moves here, we cannot cross this
2760 * barrier until the move completes.
2761 */
2762 DB_DNODE_ENTER(db);
2763
2764 dn = DB_DNODE(db);
2765 atomic_dec_32(&dn->dn_dbufs_count);
2766
2767 /*
2768 * Decrementing the dbuf count means that the bonus
2769 * buffer's dnode hold is no longer discounted in
2770 * dnode_move(). The dnode cannot move until after
2771 * the dnode_rele() below.
2772 */
2773 DB_DNODE_EXIT(db);
2774
2775 /*
2776 * Do not reference db after its lock is dropped.
2777 * Another thread may evict it.
2778 */
2779 mutex_exit(&db->db_mtx);
2780
2781 if (evict_dbuf)
2782 dnode_evict_bonus(dn);
2783
2784 dnode_rele(dn, db);
2785 } else if (db->db_buf == NULL) {
2786 /*
2787 * This is a special case: we never associated this
2788 * dbuf with any data allocated from the ARC.
2789 */
2790 ASSERT(db->db_state == DB_UNCACHED ||
2791 db->db_state == DB_NOFILL);
2792 dbuf_destroy(db);
2793 } else if (arc_released(db->db_buf)) {
2794 /*
2795 * This dbuf has anonymous data associated with it.
2796 */
2797 dbuf_destroy(db);
2798 } else {
2799 boolean_t do_arc_evict = B_FALSE;
2800 blkptr_t bp;
2801 spa_t *spa = dmu_objset_spa(db->db_objset);
2802
2803 if (!DBUF_IS_CACHEABLE(db) &&
2804 db->db_blkptr != NULL &&
2805 !BP_IS_HOLE(db->db_blkptr) &&
2806 !BP_IS_EMBEDDED(db->db_blkptr)) {
2807 do_arc_evict = B_TRUE;
2808 bp = *db->db_blkptr;
2809 }
2810
2811 if (!DBUF_IS_CACHEABLE(db) ||
2812 db->db_pending_evict) {
2813 dbuf_destroy(db);
2814 } else if (!multilist_link_active(&db->db_cache_link)) {
2815 multilist_insert(&dbuf_cache, db);
2816 (void) refcount_add_many(&dbuf_cache_size,
2817 db->db.db_size, db);
2818 mutex_exit(&db->db_mtx);
2819
2820 dbuf_evict_notify();
2821 }
2822
2823 if (do_arc_evict)
2824 arc_freed(spa, &bp);
2825 }
2826 } else {
2827 mutex_exit(&db->db_mtx);
2828 }
2829
2830 }
2831
2832 #pragma weak dmu_buf_refcount = dbuf_refcount
2833 uint64_t
dbuf_refcount(dmu_buf_impl_t * db)2834 dbuf_refcount(dmu_buf_impl_t *db)
2835 {
2836 return (refcount_count(&db->db_holds));
2837 }
2838
2839 void *
dmu_buf_replace_user(dmu_buf_t * db_fake,dmu_buf_user_t * old_user,dmu_buf_user_t * new_user)2840 dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user,
2841 dmu_buf_user_t *new_user)
2842 {
2843 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2844
2845 mutex_enter(&db->db_mtx);
2846 dbuf_verify_user(db, DBVU_NOT_EVICTING);
2847 if (db->db_user == old_user)
2848 db->db_user = new_user;
2849 else
2850 old_user = db->db_user;
2851 dbuf_verify_user(db, DBVU_NOT_EVICTING);
2852 mutex_exit(&db->db_mtx);
2853
2854 return (old_user);
2855 }
2856
2857 void *
dmu_buf_set_user(dmu_buf_t * db_fake,dmu_buf_user_t * user)2858 dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2859 {
2860 return (dmu_buf_replace_user(db_fake, NULL, user));
2861 }
2862
2863 void *
dmu_buf_set_user_ie(dmu_buf_t * db_fake,dmu_buf_user_t * user)2864 dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2865 {
2866 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2867
2868 db->db_user_immediate_evict = TRUE;
2869 return (dmu_buf_set_user(db_fake, user));
2870 }
2871
2872 void *
dmu_buf_remove_user(dmu_buf_t * db_fake,dmu_buf_user_t * user)2873 dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
2874 {
2875 return (dmu_buf_replace_user(db_fake, user, NULL));
2876 }
2877
2878 void *
dmu_buf_get_user(dmu_buf_t * db_fake)2879 dmu_buf_get_user(dmu_buf_t *db_fake)
2880 {
2881 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2882
2883 dbuf_verify_user(db, DBVU_NOT_EVICTING);
2884 return (db->db_user);
2885 }
2886
2887 void
dmu_buf_user_evict_wait()2888 dmu_buf_user_evict_wait()
2889 {
2890 taskq_wait(dbu_evict_taskq);
2891 }
2892
2893 boolean_t
dmu_buf_freeable(dmu_buf_t * dbuf)2894 dmu_buf_freeable(dmu_buf_t *dbuf)
2895 {
2896 boolean_t res = B_FALSE;
2897 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2898
2899 if (db->db_blkptr)
2900 res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset,
2901 db->db_blkptr, db->db_blkptr->blk_birth);
2902
2903 return (res);
2904 }
2905
2906 blkptr_t *
dmu_buf_get_blkptr(dmu_buf_t * db)2907 dmu_buf_get_blkptr(dmu_buf_t *db)
2908 {
2909 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2910 return (dbi->db_blkptr);
2911 }
2912
2913 objset_t *
dmu_buf_get_objset(dmu_buf_t * db)2914 dmu_buf_get_objset(dmu_buf_t *db)
2915 {
2916 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2917 return (dbi->db_objset);
2918 }
2919
2920 dnode_t *
dmu_buf_dnode_enter(dmu_buf_t * db)2921 dmu_buf_dnode_enter(dmu_buf_t *db)
2922 {
2923 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2924 DB_DNODE_ENTER(dbi);
2925 return (DB_DNODE(dbi));
2926 }
2927
2928 void
dmu_buf_dnode_exit(dmu_buf_t * db)2929 dmu_buf_dnode_exit(dmu_buf_t *db)
2930 {
2931 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
2932 DB_DNODE_EXIT(dbi);
2933 }
2934
2935 static void
dbuf_check_blkptr(dnode_t * dn,dmu_buf_impl_t * db)2936 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
2937 {
2938 /* ASSERT(dmu_tx_is_syncing(tx) */
2939 ASSERT(MUTEX_HELD(&db->db_mtx));
2940
2941 if (db->db_blkptr != NULL)
2942 return;
2943
2944 if (db->db_blkid == DMU_SPILL_BLKID) {
2945 db->db_blkptr = &dn->dn_phys->dn_spill;
2946 BP_ZERO(db->db_blkptr);
2947 return;
2948 }
2949 if (db->db_level == dn->dn_phys->dn_nlevels-1) {
2950 /*
2951 * This buffer was allocated at a time when there was
2952 * no available blkptrs from the dnode, or it was
2953 * inappropriate to hook it in (i.e., nlevels mis-match).
2954 */
2955 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
2956 ASSERT(db->db_parent == NULL);
2957 db->db_parent = dn->dn_dbuf;
2958 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
2959 DBUF_VERIFY(db);
2960 } else {
2961 dmu_buf_impl_t *parent = db->db_parent;
2962 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2963
2964 ASSERT(dn->dn_phys->dn_nlevels > 1);
2965 if (parent == NULL) {
2966 mutex_exit(&db->db_mtx);
2967 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2968 parent = dbuf_hold_level(dn, db->db_level + 1,
2969 db->db_blkid >> epbs, db);
2970 rw_exit(&dn->dn_struct_rwlock);
2971 mutex_enter(&db->db_mtx);
2972 db->db_parent = parent;
2973 }
2974 db->db_blkptr = (blkptr_t *)parent->db.db_data +
2975 (db->db_blkid & ((1ULL << epbs) - 1));
2976 DBUF_VERIFY(db);
2977 }
2978 }
2979
2980 static void
dbuf_sync_indirect(dbuf_dirty_record_t * dr,dmu_tx_t * tx)2981 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
2982 {
2983 dmu_buf_impl_t *db = dr->dr_dbuf;
2984 dnode_t *dn;
2985 zio_t *zio;
2986
2987 ASSERT(dmu_tx_is_syncing(tx));
2988
2989 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
2990
2991 mutex_enter(&db->db_mtx);
2992
2993 ASSERT(db->db_level > 0);
2994 DBUF_VERIFY(db);
2995
2996 /* Read the block if it hasn't been read yet. */
2997 if (db->db_buf == NULL) {
2998 mutex_exit(&db->db_mtx);
2999 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
3000 mutex_enter(&db->db_mtx);
3001 }
3002 ASSERT3U(db->db_state, ==, DB_CACHED);
3003 ASSERT(db->db_buf != NULL);
3004
3005 DB_DNODE_ENTER(db);
3006 dn = DB_DNODE(db);
3007 /* Indirect block size must match what the dnode thinks it is. */
3008 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3009 dbuf_check_blkptr(dn, db);
3010 DB_DNODE_EXIT(db);
3011
3012 /* Provide the pending dirty record to child dbufs */
3013 db->db_data_pending = dr;
3014
3015 mutex_exit(&db->db_mtx);
3016 dbuf_write(dr, db->db_buf, tx);
3017
3018 zio = dr->dr_zio;
3019 mutex_enter(&dr->dt.di.dr_mtx);
3020 dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx);
3021 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3022 mutex_exit(&dr->dt.di.dr_mtx);
3023 zio_nowait(zio);
3024 }
3025
3026 static void
dbuf_sync_leaf(dbuf_dirty_record_t * dr,dmu_tx_t * tx)3027 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
3028 {
3029 arc_buf_t **datap = &dr->dt.dl.dr_data;
3030 dmu_buf_impl_t *db = dr->dr_dbuf;
3031 dnode_t *dn;
3032 objset_t *os;
3033 uint64_t txg = tx->tx_txg;
3034
3035 ASSERT(dmu_tx_is_syncing(tx));
3036
3037 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
3038
3039 mutex_enter(&db->db_mtx);
3040 /*
3041 * To be synced, we must be dirtied. But we
3042 * might have been freed after the dirty.
3043 */
3044 if (db->db_state == DB_UNCACHED) {
3045 /* This buffer has been freed since it was dirtied */
3046 ASSERT(db->db.db_data == NULL);
3047 } else if (db->db_state == DB_FILL) {
3048 /* This buffer was freed and is now being re-filled */
3049 ASSERT(db->db.db_data != dr->dt.dl.dr_data);
3050 } else {
3051 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
3052 }
3053 DBUF_VERIFY(db);
3054
3055 DB_DNODE_ENTER(db);
3056 dn = DB_DNODE(db);
3057
3058 if (db->db_blkid == DMU_SPILL_BLKID) {
3059 mutex_enter(&dn->dn_mtx);
3060 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
3061 mutex_exit(&dn->dn_mtx);
3062 }
3063
3064 /*
3065 * If this is a bonus buffer, simply copy the bonus data into the
3066 * dnode. It will be written out when the dnode is synced (and it
3067 * will be synced, since it must have been dirty for dbuf_sync to
3068 * be called).
3069 */
3070 if (db->db_blkid == DMU_BONUS_BLKID) {
3071 dbuf_dirty_record_t **drp;
3072
3073 ASSERT(*datap != NULL);
3074 ASSERT0(db->db_level);
3075 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN);
3076 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen);
3077 DB_DNODE_EXIT(db);
3078
3079 if (*datap != db->db.db_data) {
3080 zio_buf_free(*datap, DN_MAX_BONUSLEN);
3081 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER);
3082 }
3083 db->db_data_pending = NULL;
3084 drp = &db->db_last_dirty;
3085 while (*drp != dr)
3086 drp = &(*drp)->dr_next;
3087 ASSERT(dr->dr_next == NULL);
3088 ASSERT(dr->dr_dbuf == db);
3089 *drp = dr->dr_next;
3090 if (dr->dr_dbuf->db_level != 0) {
3091 list_destroy(&dr->dt.di.dr_children);
3092 mutex_destroy(&dr->dt.di.dr_mtx);
3093 }
3094 kmem_free(dr, sizeof (dbuf_dirty_record_t));
3095 ASSERT(db->db_dirtycnt > 0);
3096 db->db_dirtycnt -= 1;
3097 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
3098 return;
3099 }
3100
3101 os = dn->dn_objset;
3102
3103 /*
3104 * This function may have dropped the db_mtx lock allowing a dmu_sync
3105 * operation to sneak in. As a result, we need to ensure that we
3106 * don't check the dr_override_state until we have returned from
3107 * dbuf_check_blkptr.
3108 */
3109 dbuf_check_blkptr(dn, db);
3110
3111 /*
3112 * If this buffer is in the middle of an immediate write,
3113 * wait for the synchronous IO to complete.
3114 */
3115 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
3116 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
3117 cv_wait(&db->db_changed, &db->db_mtx);
3118 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
3119 }
3120
3121 if (db->db_state != DB_NOFILL &&
3122 dn->dn_object != DMU_META_DNODE_OBJECT &&
3123 refcount_count(&db->db_holds) > 1 &&
3124 dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
3125 *datap == db->db_buf) {
3126 /*
3127 * If this buffer is currently "in use" (i.e., there
3128 * are active holds and db_data still references it),
3129 * then make a copy before we start the write so that
3130 * any modifications from the open txg will not leak
3131 * into this write.
3132 *
3133 * NOTE: this copy does not need to be made for
3134 * objects only modified in the syncing context (e.g.
3135 * DNONE_DNODE blocks).
3136 */
3137 int blksz = arc_buf_size(*datap);
3138 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
3139 *datap = arc_alloc_buf(os->os_spa, blksz, db, type);
3140 bcopy(db->db.db_data, (*datap)->b_data, blksz);
3141 }
3142 db->db_data_pending = dr;
3143
3144 mutex_exit(&db->db_mtx);
3145
3146 dbuf_write(dr, *datap, tx);
3147
3148 ASSERT(!list_link_active(&dr->dr_dirty_node));
3149 if (dn->dn_object == DMU_META_DNODE_OBJECT) {
3150 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
3151 DB_DNODE_EXIT(db);
3152 } else {
3153 /*
3154 * Although zio_nowait() does not "wait for an IO", it does
3155 * initiate the IO. If this is an empty write it seems plausible
3156 * that the IO could actually be completed before the nowait
3157 * returns. We need to DB_DNODE_EXIT() first in case
3158 * zio_nowait() invalidates the dbuf.
3159 */
3160 DB_DNODE_EXIT(db);
3161 zio_nowait(dr->dr_zio);
3162 }
3163 }
3164
3165 void
dbuf_sync_list(list_t * list,int level,dmu_tx_t * tx)3166 dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx)
3167 {
3168 dbuf_dirty_record_t *dr;
3169
3170 while (dr = list_head(list)) {
3171 if (dr->dr_zio != NULL) {
3172 /*
3173 * If we find an already initialized zio then we
3174 * are processing the meta-dnode, and we have finished.
3175 * The dbufs for all dnodes are put back on the list
3176 * during processing, so that we can zio_wait()
3177 * these IOs after initiating all child IOs.
3178 */
3179 ASSERT3U(dr->dr_dbuf->db.db_object, ==,
3180 DMU_META_DNODE_OBJECT);
3181 break;
3182 }
3183 if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
3184 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
3185 VERIFY3U(dr->dr_dbuf->db_level, ==, level);
3186 }
3187 list_remove(list, dr);
3188 if (dr->dr_dbuf->db_level > 0)
3189 dbuf_sync_indirect(dr, tx);
3190 else
3191 dbuf_sync_leaf(dr, tx);
3192 }
3193 }
3194
3195 /* ARGSUSED */
3196 static void
dbuf_write_ready(zio_t * zio,arc_buf_t * buf,void * vdb)3197 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3198 {
3199 dmu_buf_impl_t *db = vdb;
3200 dnode_t *dn;
3201 blkptr_t *bp = zio->io_bp;
3202 blkptr_t *bp_orig = &zio->io_bp_orig;
3203 spa_t *spa = zio->io_spa;
3204 int64_t delta;
3205 uint64_t fill = 0;
3206 int i;
3207
3208 ASSERT3P(db->db_blkptr, !=, NULL);
3209 ASSERT3P(&db->db_data_pending->dr_bp_copy, ==, bp);
3210
3211 DB_DNODE_ENTER(db);
3212 dn = DB_DNODE(db);
3213 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
3214 dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
3215 zio->io_prev_space_delta = delta;
3216
3217 if (bp->blk_birth != 0) {
3218 ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
3219 BP_GET_TYPE(bp) == dn->dn_type) ||
3220 (db->db_blkid == DMU_SPILL_BLKID &&
3221 BP_GET_TYPE(bp) == dn->dn_bonustype) ||
3222 BP_IS_EMBEDDED(bp));
3223 ASSERT(BP_GET_LEVEL(bp) == db->db_level);
3224 }
3225
3226 mutex_enter(&db->db_mtx);
3227
3228 #ifdef ZFS_DEBUG
3229 if (db->db_blkid == DMU_SPILL_BLKID) {
3230 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3231 ASSERT(!(BP_IS_HOLE(bp)) &&
3232 db->db_blkptr == &dn->dn_phys->dn_spill);
3233 }
3234 #endif
3235
3236 if (db->db_level == 0) {
3237 mutex_enter(&dn->dn_mtx);
3238 if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
3239 db->db_blkid != DMU_SPILL_BLKID)
3240 dn->dn_phys->dn_maxblkid = db->db_blkid;
3241 mutex_exit(&dn->dn_mtx);
3242
3243 if (dn->dn_type == DMU_OT_DNODE) {
3244 dnode_phys_t *dnp = db->db.db_data;
3245 for (i = db->db.db_size >> DNODE_SHIFT; i > 0;
3246 i--, dnp++) {
3247 if (dnp->dn_type != DMU_OT_NONE)
3248 fill++;
3249 }
3250 } else {
3251 if (BP_IS_HOLE(bp)) {
3252 fill = 0;
3253 } else {
3254 fill = 1;
3255 }
3256 }
3257 } else {
3258 blkptr_t *ibp = db->db.db_data;
3259 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3260 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
3261 if (BP_IS_HOLE(ibp))
3262 continue;
3263 fill += BP_GET_FILL(ibp);
3264 }
3265 }
3266 DB_DNODE_EXIT(db);
3267
3268 if (!BP_IS_EMBEDDED(bp))
3269 bp->blk_fill = fill;
3270
3271 mutex_exit(&db->db_mtx);
3272
3273 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3274 *db->db_blkptr = *bp;
3275 rw_exit(&dn->dn_struct_rwlock);
3276 }
3277
3278 /* ARGSUSED */
3279 /*
3280 * This function gets called just prior to running through the compression
3281 * stage of the zio pipeline. If we're an indirect block comprised of only
3282 * holes, then we want this indirect to be compressed away to a hole. In
3283 * order to do that we must zero out any information about the holes that
3284 * this indirect points to prior to before we try to compress it.
3285 */
3286 static void
dbuf_write_children_ready(zio_t * zio,arc_buf_t * buf,void * vdb)3287 dbuf_write_children_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3288 {
3289 dmu_buf_impl_t *db = vdb;
3290 dnode_t *dn;
3291 blkptr_t *bp;
3292 uint64_t i;
3293 int epbs;
3294
3295 ASSERT3U(db->db_level, >, 0);
3296 DB_DNODE_ENTER(db);
3297 dn = DB_DNODE(db);
3298 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3299
3300 /* Determine if all our children are holes */
3301 for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++) {
3302 if (!BP_IS_HOLE(bp))
3303 break;
3304 }
3305
3306 /*
3307 * If all the children are holes, then zero them all out so that
3308 * we may get compressed away.
3309 */
3310 if (i == 1 << epbs) {
3311 /* didn't find any non-holes */
3312 bzero(db->db.db_data, db->db.db_size);
3313 }
3314 DB_DNODE_EXIT(db);
3315 }
3316
3317 /*
3318 * The SPA will call this callback several times for each zio - once
3319 * for every physical child i/o (zio->io_phys_children times). This
3320 * allows the DMU to monitor the progress of each logical i/o. For example,
3321 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
3322 * block. There may be a long delay before all copies/fragments are completed,
3323 * so this callback allows us to retire dirty space gradually, as the physical
3324 * i/os complete.
3325 */
3326 /* ARGSUSED */
3327 static void
dbuf_write_physdone(zio_t * zio,arc_buf_t * buf,void * arg)3328 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
3329 {
3330 dmu_buf_impl_t *db = arg;
3331 objset_t *os = db->db_objset;
3332 dsl_pool_t *dp = dmu_objset_pool(os);
3333 dbuf_dirty_record_t *dr;
3334 int delta = 0;
3335
3336 dr = db->db_data_pending;
3337 ASSERT3U(dr->dr_txg, ==, zio->io_txg);
3338
3339 /*
3340 * The callback will be called io_phys_children times. Retire one
3341 * portion of our dirty space each time we are called. Any rounding
3342 * error will be cleaned up by dsl_pool_sync()'s call to
3343 * dsl_pool_undirty_space().
3344 */
3345 delta = dr->dr_accounted / zio->io_phys_children;
3346 dsl_pool_undirty_space(dp, delta, zio->io_txg);
3347 }
3348
3349 /* ARGSUSED */
3350 static void
dbuf_write_done(zio_t * zio,arc_buf_t * buf,void * vdb)3351 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
3352 {
3353 dmu_buf_impl_t *db = vdb;
3354 blkptr_t *bp_orig = &zio->io_bp_orig;
3355 blkptr_t *bp = db->db_blkptr;
3356 objset_t *os = db->db_objset;
3357 dmu_tx_t *tx = os->os_synctx;
3358 dbuf_dirty_record_t **drp, *dr;
3359
3360 ASSERT0(zio->io_error);
3361 ASSERT(db->db_blkptr == bp);
3362
3363 /*
3364 * For nopwrites and rewrites we ensure that the bp matches our
3365 * original and bypass all the accounting.
3366 */
3367 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
3368 ASSERT(BP_EQUAL(bp, bp_orig));
3369 } else {
3370 dsl_dataset_t *ds = os->os_dsl_dataset;
3371 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
3372 dsl_dataset_block_born(ds, bp, tx);
3373 }
3374
3375 mutex_enter(&db->db_mtx);
3376
3377 DBUF_VERIFY(db);
3378
3379 drp = &db->db_last_dirty;
3380 while ((dr = *drp) != db->db_data_pending)
3381 drp = &dr->dr_next;
3382 ASSERT(!list_link_active(&dr->dr_dirty_node));
3383 ASSERT(dr->dr_dbuf == db);
3384 ASSERT(dr->dr_next == NULL);
3385 *drp = dr->dr_next;
3386
3387 #ifdef ZFS_DEBUG
3388 if (db->db_blkid == DMU_SPILL_BLKID) {
3389 dnode_t *dn;
3390
3391 DB_DNODE_ENTER(db);
3392 dn = DB_DNODE(db);
3393 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3394 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
3395 db->db_blkptr == &dn->dn_phys->dn_spill);
3396 DB_DNODE_EXIT(db);
3397 }
3398 #endif
3399
3400 if (db->db_level == 0) {
3401 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
3402 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
3403 if (db->db_state != DB_NOFILL) {
3404 if (dr->dt.dl.dr_data != db->db_buf)
3405 arc_buf_destroy(dr->dt.dl.dr_data, db);
3406 }
3407 } else {
3408 dnode_t *dn;
3409
3410 DB_DNODE_ENTER(db);
3411 dn = DB_DNODE(db);
3412 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3413 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
3414 if (!BP_IS_HOLE(db->db_blkptr)) {
3415 int epbs =
3416 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3417 ASSERT3U(db->db_blkid, <=,
3418 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs));
3419 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
3420 db->db.db_size);
3421 }
3422 DB_DNODE_EXIT(db);
3423 mutex_destroy(&dr->dt.di.dr_mtx);
3424 list_destroy(&dr->dt.di.dr_children);
3425 }
3426 kmem_free(dr, sizeof (dbuf_dirty_record_t));
3427
3428 cv_broadcast(&db->db_changed);
3429 ASSERT(db->db_dirtycnt > 0);
3430 db->db_dirtycnt -= 1;
3431 db->db_data_pending = NULL;
3432 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg);
3433 }
3434
3435 static void
dbuf_write_nofill_ready(zio_t * zio)3436 dbuf_write_nofill_ready(zio_t *zio)
3437 {
3438 dbuf_write_ready(zio, NULL, zio->io_private);
3439 }
3440
3441 static void
dbuf_write_nofill_done(zio_t * zio)3442 dbuf_write_nofill_done(zio_t *zio)
3443 {
3444 dbuf_write_done(zio, NULL, zio->io_private);
3445 }
3446
3447 static void
dbuf_write_override_ready(zio_t * zio)3448 dbuf_write_override_ready(zio_t *zio)
3449 {
3450 dbuf_dirty_record_t *dr = zio->io_private;
3451 dmu_buf_impl_t *db = dr->dr_dbuf;
3452
3453 dbuf_write_ready(zio, NULL, db);
3454 }
3455
3456 static void
dbuf_write_override_done(zio_t * zio)3457 dbuf_write_override_done(zio_t *zio)
3458 {
3459 dbuf_dirty_record_t *dr = zio->io_private;
3460 dmu_buf_impl_t *db = dr->dr_dbuf;
3461 blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
3462
3463 mutex_enter(&db->db_mtx);
3464 if (!BP_EQUAL(zio->io_bp, obp)) {
3465 if (!BP_IS_HOLE(obp))
3466 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
3467 arc_release(dr->dt.dl.dr_data, db);
3468 }
3469 mutex_exit(&db->db_mtx);
3470
3471 dbuf_write_done(zio, NULL, db);
3472 }
3473
3474 /* Issue I/O to commit a dirty buffer to disk. */
3475 static void
dbuf_write(dbuf_dirty_record_t * dr,arc_buf_t * data,dmu_tx_t * tx)3476 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
3477 {
3478 dmu_buf_impl_t *db = dr->dr_dbuf;
3479 dnode_t *dn;
3480 objset_t *os;
3481 dmu_buf_impl_t *parent = db->db_parent;
3482 uint64_t txg = tx->tx_txg;
3483 zbookmark_phys_t zb;
3484 zio_prop_t zp;
3485 zio_t *zio;
3486 int wp_flag = 0;
3487
3488 ASSERT(dmu_tx_is_syncing(tx));
3489
3490 DB_DNODE_ENTER(db);
3491 dn = DB_DNODE(db);
3492 os = dn->dn_objset;
3493
3494 if (db->db_state != DB_NOFILL) {
3495 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
3496 /*
3497 * Private object buffers are released here rather
3498 * than in dbuf_dirty() since they are only modified
3499 * in the syncing context and we don't want the
3500 * overhead of making multiple copies of the data.
3501 */
3502 if (BP_IS_HOLE(db->db_blkptr)) {
3503 arc_buf_thaw(data);
3504 } else {
3505 dbuf_release_bp(db);
3506 }
3507 }
3508 }
3509
3510 if (parent != dn->dn_dbuf) {
3511 /* Our parent is an indirect block. */
3512 /* We have a dirty parent that has been scheduled for write. */
3513 ASSERT(parent && parent->db_data_pending);
3514 /* Our parent's buffer is one level closer to the dnode. */
3515 ASSERT(db->db_level == parent->db_level-1);
3516 /*
3517 * We're about to modify our parent's db_data by modifying
3518 * our block pointer, so the parent must be released.
3519 */
3520 ASSERT(arc_released(parent->db_buf));
3521 zio = parent->db_data_pending->dr_zio;
3522 } else {
3523 /* Our parent is the dnode itself. */
3524 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
3525 db->db_blkid != DMU_SPILL_BLKID) ||
3526 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
3527 if (db->db_blkid != DMU_SPILL_BLKID)
3528 ASSERT3P(db->db_blkptr, ==,
3529 &dn->dn_phys->dn_blkptr[db->db_blkid]);
3530 zio = dn->dn_zio;
3531 }
3532
3533 ASSERT(db->db_level == 0 || data == db->db_buf);
3534 ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
3535 ASSERT(zio);
3536
3537 SET_BOOKMARK(&zb, os->os_dsl_dataset ?
3538 os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
3539 db->db.db_object, db->db_level, db->db_blkid);
3540
3541 if (db->db_blkid == DMU_SPILL_BLKID)
3542 wp_flag = WP_SPILL;
3543 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
3544
3545 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp);
3546 DB_DNODE_EXIT(db);
3547
3548 /*
3549 * We copy the blkptr now (rather than when we instantiate the dirty
3550 * record), because its value can change between open context and
3551 * syncing context. We do not need to hold dn_struct_rwlock to read
3552 * db_blkptr because we are in syncing context.
3553 */
3554 dr->dr_bp_copy = *db->db_blkptr;
3555
3556 if (db->db_level == 0 &&
3557 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
3558 /*
3559 * The BP for this block has been provided by open context
3560 * (by dmu_sync() or dmu_buf_write_embedded()).
3561 */
3562 void *contents = (data != NULL) ? data->b_data : NULL;
3563
3564 dr->dr_zio = zio_write(zio, os->os_spa, txg,
3565 &dr->dr_bp_copy, contents, db->db.db_size, &zp,
3566 dbuf_write_override_ready, NULL, NULL,
3567 dbuf_write_override_done,
3568 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
3569 mutex_enter(&db->db_mtx);
3570 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
3571 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
3572 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
3573 mutex_exit(&db->db_mtx);
3574 } else if (db->db_state == DB_NOFILL) {
3575 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF ||
3576 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY);
3577 dr->dr_zio = zio_write(zio, os->os_spa, txg,
3578 &dr->dr_bp_copy, NULL, db->db.db_size, &zp,
3579 dbuf_write_nofill_ready, NULL, NULL,
3580 dbuf_write_nofill_done, db,
3581 ZIO_PRIORITY_ASYNC_WRITE,
3582 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
3583 } else {
3584 ASSERT(arc_released(data));
3585
3586 /*
3587 * For indirect blocks, we want to setup the children
3588 * ready callback so that we can properly handle an indirect
3589 * block that only contains holes.
3590 */
3591 arc_done_func_t *children_ready_cb = NULL;
3592 if (db->db_level != 0)
3593 children_ready_cb = dbuf_write_children_ready;
3594
3595 dr->dr_zio = arc_write(zio, os->os_spa, txg,
3596 &dr->dr_bp_copy, data, DBUF_IS_L2CACHEABLE(db),
3597 &zp, dbuf_write_ready, children_ready_cb,
3598 dbuf_write_physdone, dbuf_write_done, db,
3599 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
3600 }
3601 }
3602