]> git.proxmox.com Git - mirror_zfs.git/blob - module/zfs/dbuf.c
Reduce noise in tracing logs
[mirror_zfs.git] / module / zfs / dbuf.c
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) 2014 Spectra Logic Corporation, All rights reserved.
27 */
28
29 #include <sys/zfs_context.h>
30 #include <sys/arc.h>
31 #include <sys/dmu.h>
32 #include <sys/dmu_send.h>
33 #include <sys/dmu_impl.h>
34 #include <sys/dbuf.h>
35 #include <sys/dmu_objset.h>
36 #include <sys/dsl_dataset.h>
37 #include <sys/dsl_dir.h>
38 #include <sys/dmu_tx.h>
39 #include <sys/spa.h>
40 #include <sys/zio.h>
41 #include <sys/dmu_zfetch.h>
42 #include <sys/sa.h>
43 #include <sys/sa_impl.h>
44 #include <sys/zfeature.h>
45 #include <sys/blkptr.h>
46 #include <sys/range_tree.h>
47 #include <sys/trace_dbuf.h>
48 #include <sys/callb.h>
49
50 struct dbuf_hold_impl_data {
51 /* Function arguments */
52 dnode_t *dh_dn;
53 uint8_t dh_level;
54 uint64_t dh_blkid;
55 boolean_t dh_fail_sparse;
56 boolean_t dh_fail_uncached;
57 void *dh_tag;
58 dmu_buf_impl_t **dh_dbp;
59 /* Local variables */
60 dmu_buf_impl_t *dh_db;
61 dmu_buf_impl_t *dh_parent;
62 blkptr_t *dh_bp;
63 int dh_err;
64 dbuf_dirty_record_t *dh_dr;
65 arc_buf_contents_t dh_type;
66 int dh_depth;
67 };
68
69 static void __dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh,
70 dnode_t *dn, uint8_t level, uint64_t blkid, boolean_t fail_sparse,
71 boolean_t fail_uncached,
72 void *tag, dmu_buf_impl_t **dbp, int depth);
73 static int __dbuf_hold_impl(struct dbuf_hold_impl_data *dh);
74
75 uint_t zfs_dbuf_evict_key;
76 /*
77 * Number of times that zfs_free_range() took the slow path while doing
78 * a zfs receive. A nonzero value indicates a potential performance problem.
79 */
80 uint64_t zfs_free_range_recv_miss;
81
82 static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx);
83 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx);
84
85 #ifndef __lint
86 extern inline void dmu_buf_init_user(dmu_buf_user_t *dbu,
87 dmu_buf_evict_func_t *evict_func, dmu_buf_t **clear_on_evict_dbufp);
88 #endif /* ! __lint */
89
90 /*
91 * Global data structures and functions for the dbuf cache.
92 */
93 static kmem_cache_t *dbuf_kmem_cache;
94 static taskq_t *dbu_evict_taskq;
95
96 static kthread_t *dbuf_cache_evict_thread;
97 static kmutex_t dbuf_evict_lock;
98 static kcondvar_t dbuf_evict_cv;
99 static boolean_t dbuf_evict_thread_exit;
100
101 /*
102 * LRU cache of dbufs. The dbuf cache maintains a list of dbufs that
103 * are not currently held but have been recently released. These dbufs
104 * are not eligible for arc eviction until they are aged out of the cache.
105 * Dbufs are added to the dbuf cache once the last hold is released. If a
106 * dbuf is later accessed and still exists in the dbuf cache, then it will
107 * be removed from the cache and later re-added to the head of the cache.
108 * Dbufs that are aged out of the cache will be immediately destroyed and
109 * become eligible for arc eviction.
110 */
111 static multilist_t dbuf_cache;
112 static refcount_t dbuf_cache_size;
113 unsigned long dbuf_cache_max_bytes = 100 * 1024 * 1024;
114
115 /* Cap the size of the dbuf cache to log2 fraction of arc size. */
116 int dbuf_cache_max_shift = 5;
117
118 /*
119 * The dbuf cache uses a three-stage eviction policy:
120 * - A low water marker designates when the dbuf eviction thread
121 * should stop evicting from the dbuf cache.
122 * - When we reach the maximum size (aka mid water mark), we
123 * signal the eviction thread to run.
124 * - The high water mark indicates when the eviction thread
125 * is unable to keep up with the incoming load and eviction must
126 * happen in the context of the calling thread.
127 *
128 * The dbuf cache:
129 * (max size)
130 * low water mid water hi water
131 * +----------------------------------------+----------+----------+
132 * | | | |
133 * | | | |
134 * | | | |
135 * | | | |
136 * +----------------------------------------+----------+----------+
137 * stop signal evict
138 * evicting eviction directly
139 * thread
140 *
141 * The high and low water marks indicate the operating range for the eviction
142 * thread. The low water mark is, by default, 90% of the total size of the
143 * cache and the high water mark is at 110% (both of these percentages can be
144 * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct,
145 * respectively). The eviction thread will try to ensure that the cache remains
146 * within this range by waking up every second and checking if the cache is
147 * above the low water mark. The thread can also be woken up by callers adding
148 * elements into the cache if the cache is larger than the mid water (i.e max
149 * cache size). Once the eviction thread is woken up and eviction is required,
150 * it will continue evicting buffers until it's able to reduce the cache size
151 * to the low water mark. If the cache size continues to grow and hits the high
152 * water mark, then callers adding elments to the cache will begin to evict
153 * directly from the cache until the cache is no longer above the high water
154 * mark.
155 */
156
157 /*
158 * The percentage above and below the maximum cache size.
159 */
160 uint_t dbuf_cache_hiwater_pct = 10;
161 uint_t dbuf_cache_lowater_pct = 10;
162
163 /* ARGSUSED */
164 static int
165 dbuf_cons(void *vdb, void *unused, int kmflag)
166 {
167 dmu_buf_impl_t *db = vdb;
168 bzero(db, sizeof (dmu_buf_impl_t));
169
170 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL);
171 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL);
172 multilist_link_init(&db->db_cache_link);
173 refcount_create(&db->db_holds);
174 multilist_link_init(&db->db_cache_link);
175
176 return (0);
177 }
178
179 /* ARGSUSED */
180 static void
181 dbuf_dest(void *vdb, void *unused)
182 {
183 dmu_buf_impl_t *db = vdb;
184 mutex_destroy(&db->db_mtx);
185 cv_destroy(&db->db_changed);
186 ASSERT(!multilist_link_active(&db->db_cache_link));
187 refcount_destroy(&db->db_holds);
188 }
189
190 /*
191 * dbuf hash table routines
192 */
193 static dbuf_hash_table_t dbuf_hash_table;
194
195 static uint64_t dbuf_hash_count;
196
197 static uint64_t
198 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid)
199 {
200 uintptr_t osv = (uintptr_t)os;
201 uint64_t crc = -1ULL;
202
203 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY);
204 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF];
205 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF];
206 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF];
207 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF];
208 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF];
209 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF];
210
211 crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16);
212
213 return (crc);
214 }
215
216 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \
217 ((dbuf)->db.db_object == (obj) && \
218 (dbuf)->db_objset == (os) && \
219 (dbuf)->db_level == (level) && \
220 (dbuf)->db_blkid == (blkid))
221
222 dmu_buf_impl_t *
223 dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid)
224 {
225 dbuf_hash_table_t *h = &dbuf_hash_table;
226 uint64_t hv;
227 uint64_t idx;
228 dmu_buf_impl_t *db;
229
230 hv = dbuf_hash(os, obj, level, blkid);
231 idx = hv & h->hash_table_mask;
232
233 mutex_enter(DBUF_HASH_MUTEX(h, idx));
234 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) {
235 if (DBUF_EQUAL(db, os, obj, level, blkid)) {
236 mutex_enter(&db->db_mtx);
237 if (db->db_state != DB_EVICTING) {
238 mutex_exit(DBUF_HASH_MUTEX(h, idx));
239 return (db);
240 }
241 mutex_exit(&db->db_mtx);
242 }
243 }
244 mutex_exit(DBUF_HASH_MUTEX(h, idx));
245 return (NULL);
246 }
247
248 static dmu_buf_impl_t *
249 dbuf_find_bonus(objset_t *os, uint64_t object)
250 {
251 dnode_t *dn;
252 dmu_buf_impl_t *db = NULL;
253
254 if (dnode_hold(os, object, FTAG, &dn) == 0) {
255 rw_enter(&dn->dn_struct_rwlock, RW_READER);
256 if (dn->dn_bonus != NULL) {
257 db = dn->dn_bonus;
258 mutex_enter(&db->db_mtx);
259 }
260 rw_exit(&dn->dn_struct_rwlock);
261 dnode_rele(dn, FTAG);
262 }
263 return (db);
264 }
265
266 /*
267 * Insert an entry into the hash table. If there is already an element
268 * equal to elem in the hash table, then the already existing element
269 * will be returned and the new element will not be inserted.
270 * Otherwise returns NULL.
271 */
272 static dmu_buf_impl_t *
273 dbuf_hash_insert(dmu_buf_impl_t *db)
274 {
275 dbuf_hash_table_t *h = &dbuf_hash_table;
276 objset_t *os = db->db_objset;
277 uint64_t obj = db->db.db_object;
278 int level = db->db_level;
279 uint64_t blkid, hv, idx;
280 dmu_buf_impl_t *dbf;
281
282 blkid = db->db_blkid;
283 hv = dbuf_hash(os, obj, level, blkid);
284 idx = hv & h->hash_table_mask;
285
286 mutex_enter(DBUF_HASH_MUTEX(h, idx));
287 for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) {
288 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) {
289 mutex_enter(&dbf->db_mtx);
290 if (dbf->db_state != DB_EVICTING) {
291 mutex_exit(DBUF_HASH_MUTEX(h, idx));
292 return (dbf);
293 }
294 mutex_exit(&dbf->db_mtx);
295 }
296 }
297
298 mutex_enter(&db->db_mtx);
299 db->db_hash_next = h->hash_table[idx];
300 h->hash_table[idx] = db;
301 mutex_exit(DBUF_HASH_MUTEX(h, idx));
302 atomic_inc_64(&dbuf_hash_count);
303
304 return (NULL);
305 }
306
307 /*
308 * Remove an entry from the hash table. It must be in the EVICTING state.
309 */
310 static void
311 dbuf_hash_remove(dmu_buf_impl_t *db)
312 {
313 dbuf_hash_table_t *h = &dbuf_hash_table;
314 uint64_t hv, idx;
315 dmu_buf_impl_t *dbf, **dbp;
316
317 hv = dbuf_hash(db->db_objset, db->db.db_object,
318 db->db_level, db->db_blkid);
319 idx = hv & h->hash_table_mask;
320
321 /*
322 * We musn't hold db_mtx to maintain lock ordering:
323 * DBUF_HASH_MUTEX > db_mtx.
324 */
325 ASSERT(refcount_is_zero(&db->db_holds));
326 ASSERT(db->db_state == DB_EVICTING);
327 ASSERT(!MUTEX_HELD(&db->db_mtx));
328
329 mutex_enter(DBUF_HASH_MUTEX(h, idx));
330 dbp = &h->hash_table[idx];
331 while ((dbf = *dbp) != db) {
332 dbp = &dbf->db_hash_next;
333 ASSERT(dbf != NULL);
334 }
335 *dbp = db->db_hash_next;
336 db->db_hash_next = NULL;
337 mutex_exit(DBUF_HASH_MUTEX(h, idx));
338 atomic_dec_64(&dbuf_hash_count);
339 }
340
341 typedef enum {
342 DBVU_EVICTING,
343 DBVU_NOT_EVICTING
344 } dbvu_verify_type_t;
345
346 static void
347 dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type)
348 {
349 #ifdef ZFS_DEBUG
350 int64_t holds;
351
352 if (db->db_user == NULL)
353 return;
354
355 /* Only data blocks support the attachment of user data. */
356 ASSERT(db->db_level == 0);
357
358 /* Clients must resolve a dbuf before attaching user data. */
359 ASSERT(db->db.db_data != NULL);
360 ASSERT3U(db->db_state, ==, DB_CACHED);
361
362 holds = refcount_count(&db->db_holds);
363 if (verify_type == DBVU_EVICTING) {
364 /*
365 * Immediate eviction occurs when holds == dirtycnt.
366 * For normal eviction buffers, holds is zero on
367 * eviction, except when dbuf_fix_old_data() calls
368 * dbuf_clear_data(). However, the hold count can grow
369 * during eviction even though db_mtx is held (see
370 * dmu_bonus_hold() for an example), so we can only
371 * test the generic invariant that holds >= dirtycnt.
372 */
373 ASSERT3U(holds, >=, db->db_dirtycnt);
374 } else {
375 if (db->db_user_immediate_evict == TRUE)
376 ASSERT3U(holds, >=, db->db_dirtycnt);
377 else
378 ASSERT3U(holds, >, 0);
379 }
380 #endif
381 }
382
383 static void
384 dbuf_evict_user(dmu_buf_impl_t *db)
385 {
386 dmu_buf_user_t *dbu = db->db_user;
387
388 ASSERT(MUTEX_HELD(&db->db_mtx));
389
390 if (dbu == NULL)
391 return;
392
393 dbuf_verify_user(db, DBVU_EVICTING);
394 db->db_user = NULL;
395
396 #ifdef ZFS_DEBUG
397 if (dbu->dbu_clear_on_evict_dbufp != NULL)
398 *dbu->dbu_clear_on_evict_dbufp = NULL;
399 #endif
400
401 /*
402 * Invoke the callback from a taskq to avoid lock order reversals
403 * and limit stack depth.
404 */
405 taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func, dbu, 0,
406 &dbu->dbu_tqent);
407 }
408
409 boolean_t
410 dbuf_is_metadata(dmu_buf_impl_t *db)
411 {
412 /*
413 * Consider indirect blocks and spill blocks to be meta data.
414 */
415 if (db->db_level > 0 || db->db_blkid == DMU_SPILL_BLKID) {
416 return (B_TRUE);
417 } else {
418 boolean_t is_metadata;
419
420 DB_DNODE_ENTER(db);
421 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type);
422 DB_DNODE_EXIT(db);
423
424 return (is_metadata);
425 }
426 }
427
428
429 /*
430 * This function *must* return indices evenly distributed between all
431 * sublists of the multilist. This is needed due to how the dbuf eviction
432 * code is laid out; dbuf_evict_thread() assumes dbufs are evenly
433 * distributed between all sublists and uses this assumption when
434 * deciding which sublist to evict from and how much to evict from it.
435 */
436 unsigned int
437 dbuf_cache_multilist_index_func(multilist_t *ml, void *obj)
438 {
439 dmu_buf_impl_t *db = obj;
440
441 /*
442 * The assumption here, is the hash value for a given
443 * dmu_buf_impl_t will remain constant throughout it's lifetime
444 * (i.e. it's objset, object, level and blkid fields don't change).
445 * Thus, we don't need to store the dbuf's sublist index
446 * on insertion, as this index can be recalculated on removal.
447 *
448 * Also, the low order bits of the hash value are thought to be
449 * distributed evenly. Otherwise, in the case that the multilist
450 * has a power of two number of sublists, each sublists' usage
451 * would not be evenly distributed.
452 */
453 return (dbuf_hash(db->db_objset, db->db.db_object,
454 db->db_level, db->db_blkid) %
455 multilist_get_num_sublists(ml));
456 }
457
458 static inline boolean_t
459 dbuf_cache_above_hiwater(void)
460 {
461 uint64_t dbuf_cache_hiwater_bytes =
462 (dbuf_cache_max_bytes * dbuf_cache_hiwater_pct) / 100;
463
464 return (refcount_count(&dbuf_cache_size) >
465 dbuf_cache_max_bytes + dbuf_cache_hiwater_bytes);
466 }
467
468 static inline boolean_t
469 dbuf_cache_above_lowater(void)
470 {
471 uint64_t dbuf_cache_lowater_bytes =
472 (dbuf_cache_max_bytes * dbuf_cache_lowater_pct) / 100;
473
474 return (refcount_count(&dbuf_cache_size) >
475 dbuf_cache_max_bytes - dbuf_cache_lowater_bytes);
476 }
477
478 /*
479 * Evict the oldest eligible dbuf from the dbuf cache.
480 */
481 static void
482 dbuf_evict_one(void)
483 {
484 int idx = multilist_get_random_index(&dbuf_cache);
485 multilist_sublist_t *mls = multilist_sublist_lock(&dbuf_cache, idx);
486 dmu_buf_impl_t *db;
487 ASSERT(!MUTEX_HELD(&dbuf_evict_lock));
488
489 /*
490 * Set the thread's tsd to indicate that it's processing evictions.
491 * Once a thread stops evicting from the dbuf cache it will
492 * reset its tsd to NULL.
493 */
494 ASSERT3P(tsd_get(zfs_dbuf_evict_key), ==, NULL);
495 (void) tsd_set(zfs_dbuf_evict_key, (void *)B_TRUE);
496
497 db = multilist_sublist_tail(mls);
498 while (db != NULL && mutex_tryenter(&db->db_mtx) == 0) {
499 db = multilist_sublist_prev(mls, db);
500 }
501
502 DTRACE_PROBE2(dbuf__evict__one, dmu_buf_impl_t *, db,
503 multilist_sublist_t *, mls);
504
505 if (db != NULL) {
506 multilist_sublist_remove(mls, db);
507 multilist_sublist_unlock(mls);
508 (void) refcount_remove_many(&dbuf_cache_size,
509 db->db.db_size, db);
510 dbuf_destroy(db);
511 } else {
512 multilist_sublist_unlock(mls);
513 }
514 (void) tsd_set(zfs_dbuf_evict_key, NULL);
515 }
516
517 /*
518 * The dbuf evict thread is responsible for aging out dbufs from the
519 * cache. Once the cache has reached it's maximum size, dbufs are removed
520 * and destroyed. The eviction thread will continue running until the size
521 * of the dbuf cache is at or below the maximum size. Once the dbuf is aged
522 * out of the cache it is destroyed and becomes eligible for arc eviction.
523 */
524 static void
525 dbuf_evict_thread(void)
526 {
527 callb_cpr_t cpr;
528
529 CALLB_CPR_INIT(&cpr, &dbuf_evict_lock, callb_generic_cpr, FTAG);
530
531 mutex_enter(&dbuf_evict_lock);
532 while (!dbuf_evict_thread_exit) {
533 while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
534 CALLB_CPR_SAFE_BEGIN(&cpr);
535 (void) cv_timedwait_sig_hires(&dbuf_evict_cv,
536 &dbuf_evict_lock, SEC2NSEC(1), MSEC2NSEC(1), 0);
537 CALLB_CPR_SAFE_END(&cpr, &dbuf_evict_lock);
538 }
539 mutex_exit(&dbuf_evict_lock);
540
541 /*
542 * Keep evicting as long as we're above the low water mark
543 * for the cache. We do this without holding the locks to
544 * minimize lock contention.
545 */
546 while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) {
547 dbuf_evict_one();
548 }
549
550 mutex_enter(&dbuf_evict_lock);
551 }
552
553 dbuf_evict_thread_exit = B_FALSE;
554 cv_broadcast(&dbuf_evict_cv);
555 CALLB_CPR_EXIT(&cpr); /* drops dbuf_evict_lock */
556 thread_exit();
557 }
558
559 /*
560 * Wake up the dbuf eviction thread if the dbuf cache is at its max size.
561 * If the dbuf cache is at its high water mark, then evict a dbuf from the
562 * dbuf cache using the callers context.
563 */
564 static void
565 dbuf_evict_notify(void)
566 {
567
568 /*
569 * We use thread specific data to track when a thread has
570 * started processing evictions. This allows us to avoid deeply
571 * nested stacks that would have a call flow similar to this:
572 *
573 * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify()
574 * ^ |
575 * | |
576 * +-----dbuf_destroy()<--dbuf_evict_one()<--------+
577 *
578 * The dbuf_eviction_thread will always have its tsd set until
579 * that thread exits. All other threads will only set their tsd
580 * if they are participating in the eviction process. This only
581 * happens if the eviction thread is unable to process evictions
582 * fast enough. To keep the dbuf cache size in check, other threads
583 * can evict from the dbuf cache directly. Those threads will set
584 * their tsd values so that we ensure that they only evict one dbuf
585 * from the dbuf cache.
586 */
587 if (tsd_get(zfs_dbuf_evict_key) != NULL)
588 return;
589
590 if (refcount_count(&dbuf_cache_size) > dbuf_cache_max_bytes) {
591 boolean_t evict_now = B_FALSE;
592
593 mutex_enter(&dbuf_evict_lock);
594 if (refcount_count(&dbuf_cache_size) > dbuf_cache_max_bytes) {
595 evict_now = dbuf_cache_above_hiwater();
596 cv_signal(&dbuf_evict_cv);
597 }
598 mutex_exit(&dbuf_evict_lock);
599
600 if (evict_now) {
601 dbuf_evict_one();
602 }
603 }
604 }
605
606
607
608 void
609 dbuf_init(void)
610 {
611 uint64_t hsize = 1ULL << 16;
612 dbuf_hash_table_t *h = &dbuf_hash_table;
613 int i;
614
615 /*
616 * The hash table is big enough to fill all of physical memory
617 * with an average block size of zfs_arc_average_blocksize (default 8K).
618 * By default, the table will take up
619 * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers).
620 */
621 while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE)
622 hsize <<= 1;
623
624 retry:
625 h->hash_table_mask = hsize - 1;
626 #if defined(_KERNEL) && defined(HAVE_SPL)
627 /*
628 * Large allocations which do not require contiguous pages
629 * should be using vmem_alloc() in the linux kernel
630 */
631 h->hash_table = vmem_zalloc(hsize * sizeof (void *), KM_SLEEP);
632 #else
633 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP);
634 #endif
635 if (h->hash_table == NULL) {
636 /* XXX - we should really return an error instead of assert */
637 ASSERT(hsize > (1ULL << 10));
638 hsize >>= 1;
639 goto retry;
640 }
641
642 dbuf_kmem_cache = kmem_cache_create("dmu_buf_impl_t",
643 sizeof (dmu_buf_impl_t),
644 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0);
645
646 for (i = 0; i < DBUF_MUTEXES; i++)
647 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL);
648
649 dbuf_stats_init(h);
650
651 /*
652 * Setup the parameters for the dbuf cache. We cap the size of the
653 * dbuf cache to 1/32nd (default) of the size of the ARC.
654 */
655 dbuf_cache_max_bytes = MIN(dbuf_cache_max_bytes,
656 arc_max_bytes() >> dbuf_cache_max_shift);
657
658 /*
659 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc
660 * configuration is not required.
661 */
662 dbu_evict_taskq = taskq_create("dbu_evict", 1, defclsyspri, 0, 0, 0);
663
664 multilist_create(&dbuf_cache, sizeof (dmu_buf_impl_t),
665 offsetof(dmu_buf_impl_t, db_cache_link),
666 zfs_arc_num_sublists_per_state,
667 dbuf_cache_multilist_index_func);
668 refcount_create(&dbuf_cache_size);
669
670 tsd_create(&zfs_dbuf_evict_key, NULL);
671 dbuf_evict_thread_exit = B_FALSE;
672 mutex_init(&dbuf_evict_lock, NULL, MUTEX_DEFAULT, NULL);
673 cv_init(&dbuf_evict_cv, NULL, CV_DEFAULT, NULL);
674 dbuf_cache_evict_thread = thread_create(NULL, 0, dbuf_evict_thread,
675 NULL, 0, &p0, TS_RUN, minclsyspri);
676 }
677
678 void
679 dbuf_fini(void)
680 {
681 dbuf_hash_table_t *h = &dbuf_hash_table;
682 int i;
683
684 dbuf_stats_destroy();
685
686 for (i = 0; i < DBUF_MUTEXES; i++)
687 mutex_destroy(&h->hash_mutexes[i]);
688 #if defined(_KERNEL) && defined(HAVE_SPL)
689 /*
690 * Large allocations which do not require contiguous pages
691 * should be using vmem_free() in the linux kernel
692 */
693 vmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
694 #else
695 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *));
696 #endif
697 kmem_cache_destroy(dbuf_kmem_cache);
698 taskq_destroy(dbu_evict_taskq);
699
700 mutex_enter(&dbuf_evict_lock);
701 dbuf_evict_thread_exit = B_TRUE;
702 while (dbuf_evict_thread_exit) {
703 cv_signal(&dbuf_evict_cv);
704 cv_wait(&dbuf_evict_cv, &dbuf_evict_lock);
705 }
706 mutex_exit(&dbuf_evict_lock);
707 tsd_destroy(&zfs_dbuf_evict_key);
708
709 mutex_destroy(&dbuf_evict_lock);
710 cv_destroy(&dbuf_evict_cv);
711
712 refcount_destroy(&dbuf_cache_size);
713 multilist_destroy(&dbuf_cache);
714 }
715
716 /*
717 * Other stuff.
718 */
719
720 #ifdef ZFS_DEBUG
721 static void
722 dbuf_verify(dmu_buf_impl_t *db)
723 {
724 dnode_t *dn;
725 dbuf_dirty_record_t *dr;
726
727 ASSERT(MUTEX_HELD(&db->db_mtx));
728
729 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY))
730 return;
731
732 ASSERT(db->db_objset != NULL);
733 DB_DNODE_ENTER(db);
734 dn = DB_DNODE(db);
735 if (dn == NULL) {
736 ASSERT(db->db_parent == NULL);
737 ASSERT(db->db_blkptr == NULL);
738 } else {
739 ASSERT3U(db->db.db_object, ==, dn->dn_object);
740 ASSERT3P(db->db_objset, ==, dn->dn_objset);
741 ASSERT3U(db->db_level, <, dn->dn_nlevels);
742 ASSERT(db->db_blkid == DMU_BONUS_BLKID ||
743 db->db_blkid == DMU_SPILL_BLKID ||
744 !avl_is_empty(&dn->dn_dbufs));
745 }
746 if (db->db_blkid == DMU_BONUS_BLKID) {
747 ASSERT(dn != NULL);
748 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
749 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID);
750 } else if (db->db_blkid == DMU_SPILL_BLKID) {
751 ASSERT(dn != NULL);
752 ASSERT0(db->db.db_offset);
753 } else {
754 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size);
755 }
756
757 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next)
758 ASSERT(dr->dr_dbuf == db);
759
760 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next)
761 ASSERT(dr->dr_dbuf == db);
762
763 /*
764 * We can't assert that db_size matches dn_datablksz because it
765 * can be momentarily different when another thread is doing
766 * dnode_set_blksz().
767 */
768 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) {
769 dr = db->db_data_pending;
770 /*
771 * It should only be modified in syncing context, so
772 * make sure we only have one copy of the data.
773 */
774 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf);
775 }
776
777 /* verify db->db_blkptr */
778 if (db->db_blkptr) {
779 if (db->db_parent == dn->dn_dbuf) {
780 /* db is pointed to by the dnode */
781 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */
782 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object))
783 ASSERT(db->db_parent == NULL);
784 else
785 ASSERT(db->db_parent != NULL);
786 if (db->db_blkid != DMU_SPILL_BLKID)
787 ASSERT3P(db->db_blkptr, ==,
788 &dn->dn_phys->dn_blkptr[db->db_blkid]);
789 } else {
790 /* db is pointed to by an indirect block */
791 ASSERTV(int epb = db->db_parent->db.db_size >>
792 SPA_BLKPTRSHIFT);
793 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1);
794 ASSERT3U(db->db_parent->db.db_object, ==,
795 db->db.db_object);
796 /*
797 * dnode_grow_indblksz() can make this fail if we don't
798 * have the struct_rwlock. XXX indblksz no longer
799 * grows. safe to do this now?
800 */
801 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
802 ASSERT3P(db->db_blkptr, ==,
803 ((blkptr_t *)db->db_parent->db.db_data +
804 db->db_blkid % epb));
805 }
806 }
807 }
808 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) &&
809 (db->db_buf == NULL || db->db_buf->b_data) &&
810 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID &&
811 db->db_state != DB_FILL && !dn->dn_free_txg) {
812 /*
813 * If the blkptr isn't set but they have nonzero data,
814 * it had better be dirty, otherwise we'll lose that
815 * data when we evict this buffer.
816 *
817 * There is an exception to this rule for indirect blocks; in
818 * this case, if the indirect block is a hole, we fill in a few
819 * fields on each of the child blocks (importantly, birth time)
820 * to prevent hole birth times from being lost when you
821 * partially fill in a hole.
822 */
823 if (db->db_dirtycnt == 0) {
824 if (db->db_level == 0) {
825 uint64_t *buf = db->db.db_data;
826 int i;
827
828 for (i = 0; i < db->db.db_size >> 3; i++) {
829 ASSERT(buf[i] == 0);
830 }
831 } else {
832 int i;
833 blkptr_t *bps = db->db.db_data;
834 ASSERT3U(1 << DB_DNODE(db)->dn_indblkshift, ==,
835 db->db.db_size);
836 /*
837 * We want to verify that all the blkptrs in the
838 * indirect block are holes, but we may have
839 * automatically set up a few fields for them.
840 * We iterate through each blkptr and verify
841 * they only have those fields set.
842 */
843 for (i = 0;
844 i < db->db.db_size / sizeof (blkptr_t);
845 i++) {
846 blkptr_t *bp = &bps[i];
847 ASSERT(ZIO_CHECKSUM_IS_ZERO(
848 &bp->blk_cksum));
849 ASSERT(
850 DVA_IS_EMPTY(&bp->blk_dva[0]) &&
851 DVA_IS_EMPTY(&bp->blk_dva[1]) &&
852 DVA_IS_EMPTY(&bp->blk_dva[2]));
853 ASSERT0(bp->blk_fill);
854 ASSERT0(bp->blk_pad[0]);
855 ASSERT0(bp->blk_pad[1]);
856 ASSERT(!BP_IS_EMBEDDED(bp));
857 ASSERT(BP_IS_HOLE(bp));
858 ASSERT0(bp->blk_phys_birth);
859 }
860 }
861 }
862 }
863 DB_DNODE_EXIT(db);
864 }
865 #endif
866
867 static void
868 dbuf_clear_data(dmu_buf_impl_t *db)
869 {
870 ASSERT(MUTEX_HELD(&db->db_mtx));
871 dbuf_evict_user(db);
872 ASSERT3P(db->db_buf, ==, NULL);
873 db->db.db_data = NULL;
874 if (db->db_state != DB_NOFILL)
875 db->db_state = DB_UNCACHED;
876 }
877
878 static void
879 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf)
880 {
881 ASSERT(MUTEX_HELD(&db->db_mtx));
882 ASSERT(buf != NULL);
883
884 db->db_buf = buf;
885 ASSERT(buf->b_data != NULL);
886 db->db.db_data = buf->b_data;
887 }
888
889 /*
890 * Loan out an arc_buf for read. Return the loaned arc_buf.
891 */
892 arc_buf_t *
893 dbuf_loan_arcbuf(dmu_buf_impl_t *db)
894 {
895 arc_buf_t *abuf;
896
897 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
898 mutex_enter(&db->db_mtx);
899 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) {
900 int blksz = db->db.db_size;
901 spa_t *spa = db->db_objset->os_spa;
902
903 mutex_exit(&db->db_mtx);
904 abuf = arc_loan_buf(spa, B_FALSE, blksz);
905 bcopy(db->db.db_data, abuf->b_data, blksz);
906 } else {
907 abuf = db->db_buf;
908 arc_loan_inuse_buf(abuf, db);
909 db->db_buf = NULL;
910 dbuf_clear_data(db);
911 mutex_exit(&db->db_mtx);
912 }
913 return (abuf);
914 }
915
916 /*
917 * Calculate which level n block references the data at the level 0 offset
918 * provided.
919 */
920 uint64_t
921 dbuf_whichblock(dnode_t *dn, int64_t level, uint64_t offset)
922 {
923 if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) {
924 /*
925 * The level n blkid is equal to the level 0 blkid divided by
926 * the number of level 0s in a level n block.
927 *
928 * The level 0 blkid is offset >> datablkshift =
929 * offset / 2^datablkshift.
930 *
931 * The number of level 0s in a level n is the number of block
932 * pointers in an indirect block, raised to the power of level.
933 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level =
934 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)).
935 *
936 * Thus, the level n blkid is: offset /
937 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT)))
938 * = offset / 2^(datablkshift + level *
939 * (indblkshift - SPA_BLKPTRSHIFT))
940 * = offset >> (datablkshift + level *
941 * (indblkshift - SPA_BLKPTRSHIFT))
942 */
943 return (offset >> (dn->dn_datablkshift + level *
944 (dn->dn_indblkshift - SPA_BLKPTRSHIFT)));
945 } else {
946 ASSERT3U(offset, <, dn->dn_datablksz);
947 return (0);
948 }
949 }
950
951 static void
952 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb)
953 {
954 dmu_buf_impl_t *db = vdb;
955
956 mutex_enter(&db->db_mtx);
957 ASSERT3U(db->db_state, ==, DB_READ);
958 /*
959 * All reads are synchronous, so we must have a hold on the dbuf
960 */
961 ASSERT(refcount_count(&db->db_holds) > 0);
962 ASSERT(db->db_buf == NULL);
963 ASSERT(db->db.db_data == NULL);
964 if (db->db_level == 0 && db->db_freed_in_flight) {
965 /* we were freed in flight; disregard any error */
966 arc_release(buf, db);
967 bzero(buf->b_data, db->db.db_size);
968 arc_buf_freeze(buf);
969 db->db_freed_in_flight = FALSE;
970 dbuf_set_data(db, buf);
971 db->db_state = DB_CACHED;
972 } else if (zio == NULL || zio->io_error == 0) {
973 dbuf_set_data(db, buf);
974 db->db_state = DB_CACHED;
975 } else {
976 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
977 ASSERT3P(db->db_buf, ==, NULL);
978 arc_buf_destroy(buf, db);
979 db->db_state = DB_UNCACHED;
980 }
981 cv_broadcast(&db->db_changed);
982 dbuf_rele_and_unlock(db, NULL);
983 }
984
985 static int
986 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
987 {
988 dnode_t *dn;
989 zbookmark_phys_t zb;
990 uint32_t aflags = ARC_FLAG_NOWAIT;
991 int err;
992
993 DB_DNODE_ENTER(db);
994 dn = DB_DNODE(db);
995 ASSERT(!refcount_is_zero(&db->db_holds));
996 /* We need the struct_rwlock to prevent db_blkptr from changing. */
997 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
998 ASSERT(MUTEX_HELD(&db->db_mtx));
999 ASSERT(db->db_state == DB_UNCACHED);
1000 ASSERT(db->db_buf == NULL);
1001
1002 if (db->db_blkid == DMU_BONUS_BLKID) {
1003 /*
1004 * The bonus length stored in the dnode may be less than
1005 * the maximum available space in the bonus buffer.
1006 */
1007 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen);
1008 int max_bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
1009
1010 ASSERT3U(bonuslen, <=, db->db.db_size);
1011 db->db.db_data = zio_buf_alloc(max_bonuslen);
1012 arc_space_consume(max_bonuslen, ARC_SPACE_BONUS);
1013 if (bonuslen < max_bonuslen)
1014 bzero(db->db.db_data, max_bonuslen);
1015 if (bonuslen)
1016 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen);
1017 DB_DNODE_EXIT(db);
1018 db->db_state = DB_CACHED;
1019 mutex_exit(&db->db_mtx);
1020 return (0);
1021 }
1022
1023 /*
1024 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync()
1025 * processes the delete record and clears the bp while we are waiting
1026 * for the dn_mtx (resulting in a "no" from block_freed).
1027 */
1028 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) ||
1029 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) ||
1030 BP_IS_HOLE(db->db_blkptr)))) {
1031 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1032
1033 dbuf_set_data(db, arc_alloc_buf(db->db_objset->os_spa, db, type,
1034 db->db.db_size));
1035 bzero(db->db.db_data, db->db.db_size);
1036
1037 if (db->db_blkptr != NULL && db->db_level > 0 &&
1038 BP_IS_HOLE(db->db_blkptr) &&
1039 db->db_blkptr->blk_birth != 0) {
1040 blkptr_t *bps = db->db.db_data;
1041 int i;
1042 for (i = 0; i < ((1 <<
1043 DB_DNODE(db)->dn_indblkshift) / sizeof (blkptr_t));
1044 i++) {
1045 blkptr_t *bp = &bps[i];
1046 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
1047 1 << dn->dn_indblkshift);
1048 BP_SET_LSIZE(bp,
1049 BP_GET_LEVEL(db->db_blkptr) == 1 ?
1050 dn->dn_datablksz :
1051 BP_GET_LSIZE(db->db_blkptr));
1052 BP_SET_TYPE(bp, BP_GET_TYPE(db->db_blkptr));
1053 BP_SET_LEVEL(bp,
1054 BP_GET_LEVEL(db->db_blkptr) - 1);
1055 BP_SET_BIRTH(bp, db->db_blkptr->blk_birth, 0);
1056 }
1057 }
1058 DB_DNODE_EXIT(db);
1059 db->db_state = DB_CACHED;
1060 mutex_exit(&db->db_mtx);
1061 return (0);
1062 }
1063
1064 DB_DNODE_EXIT(db);
1065
1066 db->db_state = DB_READ;
1067 mutex_exit(&db->db_mtx);
1068
1069 if (DBUF_IS_L2CACHEABLE(db))
1070 aflags |= ARC_FLAG_L2CACHE;
1071
1072 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ?
1073 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET,
1074 db->db.db_object, db->db_level, db->db_blkid);
1075
1076 dbuf_add_ref(db, NULL);
1077
1078 err = arc_read(zio, db->db_objset->os_spa, db->db_blkptr,
1079 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ,
1080 (flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED,
1081 &aflags, &zb);
1082
1083 return (err);
1084 }
1085
1086 /*
1087 * This is our just-in-time copy function. It makes a copy of buffers that
1088 * have been modified in a previous transaction group before we access them in
1089 * the current active group.
1090 *
1091 * This function is used in three places: when we are dirtying a buffer for the
1092 * first time in a txg, when we are freeing a range in a dnode that includes
1093 * this buffer, and when we are accessing a buffer which was received compressed
1094 * and later referenced in a WRITE_BYREF record.
1095 *
1096 * Note that when we are called from dbuf_free_range() we do not put a hold on
1097 * the buffer, we just traverse the active dbuf list for the dnode.
1098 */
1099 static void
1100 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg)
1101 {
1102 dbuf_dirty_record_t *dr = db->db_last_dirty;
1103
1104 ASSERT(MUTEX_HELD(&db->db_mtx));
1105 ASSERT(db->db.db_data != NULL);
1106 ASSERT(db->db_level == 0);
1107 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT);
1108
1109 if (dr == NULL ||
1110 (dr->dt.dl.dr_data !=
1111 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf)))
1112 return;
1113
1114 /*
1115 * If the last dirty record for this dbuf has not yet synced
1116 * and its referencing the dbuf data, either:
1117 * reset the reference to point to a new copy,
1118 * or (if there a no active holders)
1119 * just null out the current db_data pointer.
1120 */
1121 ASSERT(dr->dr_txg >= txg - 2);
1122 if (db->db_blkid == DMU_BONUS_BLKID) {
1123 /* Note that the data bufs here are zio_bufs */
1124 dnode_t *dn = DB_DNODE(db);
1125 int bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
1126 dr->dt.dl.dr_data = zio_buf_alloc(bonuslen);
1127 arc_space_consume(bonuslen, ARC_SPACE_BONUS);
1128 bcopy(db->db.db_data, dr->dt.dl.dr_data, bonuslen);
1129 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) {
1130 int size = arc_buf_size(db->db_buf);
1131 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1132 spa_t *spa = db->db_objset->os_spa;
1133 enum zio_compress compress_type =
1134 arc_get_compression(db->db_buf);
1135
1136 if (compress_type == ZIO_COMPRESS_OFF) {
1137 dr->dt.dl.dr_data = arc_alloc_buf(spa, db, type, size);
1138 } else {
1139 ASSERT3U(type, ==, ARC_BUFC_DATA);
1140 dr->dt.dl.dr_data = arc_alloc_compressed_buf(spa, db,
1141 size, arc_buf_lsize(db->db_buf), compress_type);
1142 }
1143 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size);
1144 } else {
1145 db->db_buf = NULL;
1146 dbuf_clear_data(db);
1147 }
1148 }
1149
1150 int
1151 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
1152 {
1153 int err = 0;
1154 boolean_t havepzio = (zio != NULL);
1155 boolean_t prefetch;
1156 dnode_t *dn;
1157
1158 /*
1159 * We don't have to hold the mutex to check db_state because it
1160 * can't be freed while we have a hold on the buffer.
1161 */
1162 ASSERT(!refcount_is_zero(&db->db_holds));
1163
1164 if (db->db_state == DB_NOFILL)
1165 return (SET_ERROR(EIO));
1166
1167 DB_DNODE_ENTER(db);
1168 dn = DB_DNODE(db);
1169 if ((flags & DB_RF_HAVESTRUCT) == 0)
1170 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1171
1172 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1173 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
1174 DBUF_IS_CACHEABLE(db);
1175
1176 mutex_enter(&db->db_mtx);
1177 if (db->db_state == DB_CACHED) {
1178 /*
1179 * If the arc buf is compressed, we need to decompress it to
1180 * read the data. This could happen during the "zfs receive" of
1181 * a stream which is compressed and deduplicated.
1182 */
1183 if (db->db_buf != NULL &&
1184 arc_get_compression(db->db_buf) != ZIO_COMPRESS_OFF) {
1185 dbuf_fix_old_data(db,
1186 spa_syncing_txg(dmu_objset_spa(db->db_objset)));
1187 err = arc_decompress(db->db_buf);
1188 dbuf_set_data(db, db->db_buf);
1189 }
1190 mutex_exit(&db->db_mtx);
1191 if (prefetch)
1192 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1193 if ((flags & DB_RF_HAVESTRUCT) == 0)
1194 rw_exit(&dn->dn_struct_rwlock);
1195 DB_DNODE_EXIT(db);
1196 } else if (db->db_state == DB_UNCACHED) {
1197 spa_t *spa = dn->dn_objset->os_spa;
1198
1199 if (zio == NULL)
1200 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL);
1201
1202 err = dbuf_read_impl(db, zio, flags);
1203
1204 /* dbuf_read_impl has dropped db_mtx for us */
1205
1206 if (!err && prefetch)
1207 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1208
1209 if ((flags & DB_RF_HAVESTRUCT) == 0)
1210 rw_exit(&dn->dn_struct_rwlock);
1211 DB_DNODE_EXIT(db);
1212
1213 if (!err && !havepzio)
1214 err = zio_wait(zio);
1215 } else {
1216 /*
1217 * Another reader came in while the dbuf was in flight
1218 * between UNCACHED and CACHED. Either a writer will finish
1219 * writing the buffer (sending the dbuf to CACHED) or the
1220 * first reader's request will reach the read_done callback
1221 * and send the dbuf to CACHED. Otherwise, a failure
1222 * occurred and the dbuf went to UNCACHED.
1223 */
1224 mutex_exit(&db->db_mtx);
1225 if (prefetch)
1226 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE);
1227 if ((flags & DB_RF_HAVESTRUCT) == 0)
1228 rw_exit(&dn->dn_struct_rwlock);
1229 DB_DNODE_EXIT(db);
1230
1231 /* Skip the wait per the caller's request. */
1232 mutex_enter(&db->db_mtx);
1233 if ((flags & DB_RF_NEVERWAIT) == 0) {
1234 while (db->db_state == DB_READ ||
1235 db->db_state == DB_FILL) {
1236 ASSERT(db->db_state == DB_READ ||
1237 (flags & DB_RF_HAVESTRUCT) == 0);
1238 DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *,
1239 db, zio_t *, zio);
1240 cv_wait(&db->db_changed, &db->db_mtx);
1241 }
1242 if (db->db_state == DB_UNCACHED)
1243 err = SET_ERROR(EIO);
1244 }
1245 mutex_exit(&db->db_mtx);
1246 }
1247
1248 ASSERT(err || havepzio || db->db_state == DB_CACHED);
1249 return (err);
1250 }
1251
1252 static void
1253 dbuf_noread(dmu_buf_impl_t *db)
1254 {
1255 ASSERT(!refcount_is_zero(&db->db_holds));
1256 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1257 mutex_enter(&db->db_mtx);
1258 while (db->db_state == DB_READ || db->db_state == DB_FILL)
1259 cv_wait(&db->db_changed, &db->db_mtx);
1260 if (db->db_state == DB_UNCACHED) {
1261 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1262 spa_t *spa = db->db_objset->os_spa;
1263
1264 ASSERT(db->db_buf == NULL);
1265 ASSERT(db->db.db_data == NULL);
1266 dbuf_set_data(db, arc_alloc_buf(spa, db, type, db->db.db_size));
1267 db->db_state = DB_FILL;
1268 } else if (db->db_state == DB_NOFILL) {
1269 dbuf_clear_data(db);
1270 } else {
1271 ASSERT3U(db->db_state, ==, DB_CACHED);
1272 }
1273 mutex_exit(&db->db_mtx);
1274 }
1275
1276 void
1277 dbuf_unoverride(dbuf_dirty_record_t *dr)
1278 {
1279 dmu_buf_impl_t *db = dr->dr_dbuf;
1280 blkptr_t *bp = &dr->dt.dl.dr_overridden_by;
1281 uint64_t txg = dr->dr_txg;
1282
1283 ASSERT(MUTEX_HELD(&db->db_mtx));
1284 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC);
1285 ASSERT(db->db_level == 0);
1286
1287 if (db->db_blkid == DMU_BONUS_BLKID ||
1288 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN)
1289 return;
1290
1291 ASSERT(db->db_data_pending != dr);
1292
1293 /* free this block */
1294 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite)
1295 zio_free(db->db_objset->os_spa, txg, bp);
1296
1297 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1298 dr->dt.dl.dr_nopwrite = B_FALSE;
1299
1300 /*
1301 * Release the already-written buffer, so we leave it in
1302 * a consistent dirty state. Note that all callers are
1303 * modifying the buffer, so they will immediately do
1304 * another (redundant) arc_release(). Therefore, leave
1305 * the buf thawed to save the effort of freezing &
1306 * immediately re-thawing it.
1307 */
1308 arc_release(dr->dt.dl.dr_data, db);
1309 }
1310
1311 /*
1312 * Evict (if its unreferenced) or clear (if its referenced) any level-0
1313 * data blocks in the free range, so that any future readers will find
1314 * empty blocks.
1315 *
1316 * This is a no-op if the dataset is in the middle of an incremental
1317 * receive; see comment below for details.
1318 */
1319 void
1320 dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
1321 dmu_tx_t *tx)
1322 {
1323 dmu_buf_impl_t *db_search;
1324 dmu_buf_impl_t *db, *db_next;
1325 uint64_t txg = tx->tx_txg;
1326 avl_index_t where;
1327 boolean_t freespill =
1328 (start_blkid == DMU_SPILL_BLKID || end_blkid == DMU_SPILL_BLKID);
1329
1330 if (end_blkid > dn->dn_maxblkid && !freespill)
1331 end_blkid = dn->dn_maxblkid;
1332 dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid);
1333
1334 db_search = kmem_alloc(sizeof (dmu_buf_impl_t), KM_SLEEP);
1335 db_search->db_level = 0;
1336 db_search->db_blkid = start_blkid;
1337 db_search->db_state = DB_SEARCH;
1338
1339 mutex_enter(&dn->dn_dbufs_mtx);
1340 if (start_blkid >= dn->dn_unlisted_l0_blkid && !freespill) {
1341 /* There can't be any dbufs in this range; no need to search. */
1342 #ifdef DEBUG
1343 db = avl_find(&dn->dn_dbufs, db_search, &where);
1344 ASSERT3P(db, ==, NULL);
1345 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1346 ASSERT(db == NULL || db->db_level > 0);
1347 #endif
1348 goto out;
1349 } else if (dmu_objset_is_receiving(dn->dn_objset)) {
1350 /*
1351 * If we are receiving, we expect there to be no dbufs in
1352 * the range to be freed, because receive modifies each
1353 * block at most once, and in offset order. If this is
1354 * not the case, it can lead to performance problems,
1355 * so note that we unexpectedly took the slow path.
1356 */
1357 atomic_inc_64(&zfs_free_range_recv_miss);
1358 }
1359
1360 db = avl_find(&dn->dn_dbufs, db_search, &where);
1361 ASSERT3P(db, ==, NULL);
1362 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1363
1364 for (; db != NULL; db = db_next) {
1365 db_next = AVL_NEXT(&dn->dn_dbufs, db);
1366 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1367
1368 if (db->db_level != 0 || db->db_blkid > end_blkid) {
1369 break;
1370 }
1371 ASSERT3U(db->db_blkid, >=, start_blkid);
1372
1373 /* found a level 0 buffer in the range */
1374 mutex_enter(&db->db_mtx);
1375 if (dbuf_undirty(db, tx)) {
1376 /* mutex has been dropped and dbuf destroyed */
1377 continue;
1378 }
1379
1380 if (db->db_state == DB_UNCACHED ||
1381 db->db_state == DB_NOFILL ||
1382 db->db_state == DB_EVICTING) {
1383 ASSERT(db->db.db_data == NULL);
1384 mutex_exit(&db->db_mtx);
1385 continue;
1386 }
1387 if (db->db_state == DB_READ || db->db_state == DB_FILL) {
1388 /* will be handled in dbuf_read_done or dbuf_rele */
1389 db->db_freed_in_flight = TRUE;
1390 mutex_exit(&db->db_mtx);
1391 continue;
1392 }
1393 if (refcount_count(&db->db_holds) == 0) {
1394 ASSERT(db->db_buf);
1395 dbuf_destroy(db);
1396 continue;
1397 }
1398 /* The dbuf is referenced */
1399
1400 if (db->db_last_dirty != NULL) {
1401 dbuf_dirty_record_t *dr = db->db_last_dirty;
1402
1403 if (dr->dr_txg == txg) {
1404 /*
1405 * This buffer is "in-use", re-adjust the file
1406 * size to reflect that this buffer may
1407 * contain new data when we sync.
1408 */
1409 if (db->db_blkid != DMU_SPILL_BLKID &&
1410 db->db_blkid > dn->dn_maxblkid)
1411 dn->dn_maxblkid = db->db_blkid;
1412 dbuf_unoverride(dr);
1413 } else {
1414 /*
1415 * This dbuf is not dirty in the open context.
1416 * Either uncache it (if its not referenced in
1417 * the open context) or reset its contents to
1418 * empty.
1419 */
1420 dbuf_fix_old_data(db, txg);
1421 }
1422 }
1423 /* clear the contents if its cached */
1424 if (db->db_state == DB_CACHED) {
1425 ASSERT(db->db.db_data != NULL);
1426 arc_release(db->db_buf, db);
1427 bzero(db->db.db_data, db->db.db_size);
1428 arc_buf_freeze(db->db_buf);
1429 }
1430
1431 mutex_exit(&db->db_mtx);
1432 }
1433
1434 out:
1435 kmem_free(db_search, sizeof (dmu_buf_impl_t));
1436 mutex_exit(&dn->dn_dbufs_mtx);
1437 }
1438
1439 static int
1440 dbuf_block_freeable(dmu_buf_impl_t *db)
1441 {
1442 dsl_dataset_t *ds = db->db_objset->os_dsl_dataset;
1443 uint64_t birth_txg = 0;
1444
1445 /*
1446 * We don't need any locking to protect db_blkptr:
1447 * If it's syncing, then db_last_dirty will be set
1448 * so we'll ignore db_blkptr.
1449 *
1450 * This logic ensures that only block births for
1451 * filled blocks are considered.
1452 */
1453 ASSERT(MUTEX_HELD(&db->db_mtx));
1454 if (db->db_last_dirty && (db->db_blkptr == NULL ||
1455 !BP_IS_HOLE(db->db_blkptr))) {
1456 birth_txg = db->db_last_dirty->dr_txg;
1457 } else if (db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) {
1458 birth_txg = db->db_blkptr->blk_birth;
1459 }
1460
1461 /*
1462 * If this block don't exist or is in a snapshot, it can't be freed.
1463 * Don't pass the bp to dsl_dataset_block_freeable() since we
1464 * are holding the db_mtx lock and might deadlock if we are
1465 * prefetching a dedup-ed block.
1466 */
1467 if (birth_txg != 0)
1468 return (ds == NULL ||
1469 dsl_dataset_block_freeable(ds, NULL, birth_txg));
1470 else
1471 return (B_FALSE);
1472 }
1473
1474 void
1475 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx)
1476 {
1477 arc_buf_t *buf, *obuf;
1478 int osize = db->db.db_size;
1479 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
1480 dnode_t *dn;
1481
1482 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1483
1484 DB_DNODE_ENTER(db);
1485 dn = DB_DNODE(db);
1486
1487 /* XXX does *this* func really need the lock? */
1488 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
1489
1490 /*
1491 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held
1492 * is OK, because there can be no other references to the db
1493 * when we are changing its size, so no concurrent DB_FILL can
1494 * be happening.
1495 */
1496 /*
1497 * XXX we should be doing a dbuf_read, checking the return
1498 * value and returning that up to our callers
1499 */
1500 dmu_buf_will_dirty(&db->db, tx);
1501
1502 /* create the data buffer for the new block */
1503 buf = arc_alloc_buf(dn->dn_objset->os_spa, db, type, size);
1504
1505 /* copy old block data to the new block */
1506 obuf = db->db_buf;
1507 bcopy(obuf->b_data, buf->b_data, MIN(osize, size));
1508 /* zero the remainder */
1509 if (size > osize)
1510 bzero((uint8_t *)buf->b_data + osize, size - osize);
1511
1512 mutex_enter(&db->db_mtx);
1513 dbuf_set_data(db, buf);
1514 arc_buf_destroy(obuf, db);
1515 db->db.db_size = size;
1516
1517 if (db->db_level == 0) {
1518 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
1519 db->db_last_dirty->dt.dl.dr_data = buf;
1520 }
1521 mutex_exit(&db->db_mtx);
1522
1523 dnode_willuse_space(dn, size-osize, tx);
1524 DB_DNODE_EXIT(db);
1525 }
1526
1527 void
1528 dbuf_release_bp(dmu_buf_impl_t *db)
1529 {
1530 ASSERTV(objset_t *os = db->db_objset);
1531
1532 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os)));
1533 ASSERT(arc_released(os->os_phys_buf) ||
1534 list_link_active(&os->os_dsl_dataset->ds_synced_link));
1535 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf));
1536
1537 (void) arc_release(db->db_buf, db);
1538 }
1539
1540 /*
1541 * We already have a dirty record for this TXG, and we are being
1542 * dirtied again.
1543 */
1544 static void
1545 dbuf_redirty(dbuf_dirty_record_t *dr)
1546 {
1547 dmu_buf_impl_t *db = dr->dr_dbuf;
1548
1549 ASSERT(MUTEX_HELD(&db->db_mtx));
1550
1551 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) {
1552 /*
1553 * If this buffer has already been written out,
1554 * we now need to reset its state.
1555 */
1556 dbuf_unoverride(dr);
1557 if (db->db.db_object != DMU_META_DNODE_OBJECT &&
1558 db->db_state != DB_NOFILL) {
1559 /* Already released on initial dirty, so just thaw. */
1560 ASSERT(arc_released(db->db_buf));
1561 arc_buf_thaw(db->db_buf);
1562 }
1563 }
1564 }
1565
1566 dbuf_dirty_record_t *
1567 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1568 {
1569 dnode_t *dn;
1570 objset_t *os;
1571 dbuf_dirty_record_t **drp, *dr;
1572 int drop_struct_lock = FALSE;
1573 boolean_t do_free_accounting = B_FALSE;
1574 int txgoff = tx->tx_txg & TXG_MASK;
1575
1576 ASSERT(tx->tx_txg != 0);
1577 ASSERT(!refcount_is_zero(&db->db_holds));
1578 DMU_TX_DIRTY_BUF(tx, db);
1579
1580 DB_DNODE_ENTER(db);
1581 dn = DB_DNODE(db);
1582 /*
1583 * Shouldn't dirty a regular buffer in syncing context. Private
1584 * objects may be dirtied in syncing context, but only if they
1585 * were already pre-dirtied in open context.
1586 */
1587 ASSERT(!dmu_tx_is_syncing(tx) ||
1588 BP_IS_HOLE(dn->dn_objset->os_rootbp) ||
1589 DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1590 dn->dn_objset->os_dsl_dataset == NULL);
1591 /*
1592 * We make this assert for private objects as well, but after we
1593 * check if we're already dirty. They are allowed to re-dirty
1594 * in syncing context.
1595 */
1596 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
1597 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1598 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1599
1600 mutex_enter(&db->db_mtx);
1601 /*
1602 * XXX make this true for indirects too? The problem is that
1603 * transactions created with dmu_tx_create_assigned() from
1604 * syncing context don't bother holding ahead.
1605 */
1606 ASSERT(db->db_level != 0 ||
1607 db->db_state == DB_CACHED || db->db_state == DB_FILL ||
1608 db->db_state == DB_NOFILL);
1609
1610 mutex_enter(&dn->dn_mtx);
1611 /*
1612 * Don't set dirtyctx to SYNC if we're just modifying this as we
1613 * initialize the objset.
1614 */
1615 if (dn->dn_dirtyctx == DN_UNDIRTIED &&
1616 !BP_IS_HOLE(dn->dn_objset->os_rootbp)) {
1617 dn->dn_dirtyctx =
1618 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN);
1619 ASSERT(dn->dn_dirtyctx_firstset == NULL);
1620 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP);
1621 }
1622 mutex_exit(&dn->dn_mtx);
1623
1624 if (db->db_blkid == DMU_SPILL_BLKID)
1625 dn->dn_have_spill = B_TRUE;
1626
1627 /*
1628 * If this buffer is already dirty, we're done.
1629 */
1630 drp = &db->db_last_dirty;
1631 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg ||
1632 db->db.db_object == DMU_META_DNODE_OBJECT);
1633 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg)
1634 drp = &dr->dr_next;
1635 if (dr && dr->dr_txg == tx->tx_txg) {
1636 DB_DNODE_EXIT(db);
1637
1638 dbuf_redirty(dr);
1639 mutex_exit(&db->db_mtx);
1640 return (dr);
1641 }
1642
1643 /*
1644 * Only valid if not already dirty.
1645 */
1646 ASSERT(dn->dn_object == 0 ||
1647 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx ==
1648 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN));
1649
1650 ASSERT3U(dn->dn_nlevels, >, db->db_level);
1651 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) ||
1652 dn->dn_phys->dn_nlevels > db->db_level ||
1653 dn->dn_next_nlevels[txgoff] > db->db_level ||
1654 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level ||
1655 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level);
1656
1657 /*
1658 * We should only be dirtying in syncing context if it's the
1659 * mos or we're initializing the os or it's a special object.
1660 * However, we are allowed to dirty in syncing context provided
1661 * we already dirtied it in open context. Hence we must make
1662 * this assertion only if we're not already dirty.
1663 */
1664 os = dn->dn_objset;
1665 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) ||
1666 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp));
1667 ASSERT(db->db.db_size != 0);
1668
1669 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1670
1671 if (db->db_blkid != DMU_BONUS_BLKID) {
1672 /*
1673 * Update the accounting.
1674 * Note: we delay "free accounting" until after we drop
1675 * the db_mtx. This keeps us from grabbing other locks
1676 * (and possibly deadlocking) in bp_get_dsize() while
1677 * also holding the db_mtx.
1678 */
1679 dnode_willuse_space(dn, db->db.db_size, tx);
1680 do_free_accounting = dbuf_block_freeable(db);
1681 }
1682
1683 /*
1684 * If this buffer is dirty in an old transaction group we need
1685 * to make a copy of it so that the changes we make in this
1686 * transaction group won't leak out when we sync the older txg.
1687 */
1688 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP);
1689 list_link_init(&dr->dr_dirty_node);
1690 if (db->db_level == 0) {
1691 void *data_old = db->db_buf;
1692
1693 if (db->db_state != DB_NOFILL) {
1694 if (db->db_blkid == DMU_BONUS_BLKID) {
1695 dbuf_fix_old_data(db, tx->tx_txg);
1696 data_old = db->db.db_data;
1697 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) {
1698 /*
1699 * Release the data buffer from the cache so
1700 * that we can modify it without impacting
1701 * possible other users of this cached data
1702 * block. Note that indirect blocks and
1703 * private objects are not released until the
1704 * syncing state (since they are only modified
1705 * then).
1706 */
1707 arc_release(db->db_buf, db);
1708 dbuf_fix_old_data(db, tx->tx_txg);
1709 data_old = db->db_buf;
1710 }
1711 ASSERT(data_old != NULL);
1712 }
1713 dr->dt.dl.dr_data = data_old;
1714 } else {
1715 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_NOLOCKDEP, NULL);
1716 list_create(&dr->dt.di.dr_children,
1717 sizeof (dbuf_dirty_record_t),
1718 offsetof(dbuf_dirty_record_t, dr_dirty_node));
1719 }
1720 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL)
1721 dr->dr_accounted = db->db.db_size;
1722 dr->dr_dbuf = db;
1723 dr->dr_txg = tx->tx_txg;
1724 dr->dr_next = *drp;
1725 *drp = dr;
1726
1727 /*
1728 * We could have been freed_in_flight between the dbuf_noread
1729 * and dbuf_dirty. We win, as though the dbuf_noread() had
1730 * happened after the free.
1731 */
1732 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
1733 db->db_blkid != DMU_SPILL_BLKID) {
1734 mutex_enter(&dn->dn_mtx);
1735 if (dn->dn_free_ranges[txgoff] != NULL) {
1736 range_tree_clear(dn->dn_free_ranges[txgoff],
1737 db->db_blkid, 1);
1738 }
1739 mutex_exit(&dn->dn_mtx);
1740 db->db_freed_in_flight = FALSE;
1741 }
1742
1743 /*
1744 * This buffer is now part of this txg
1745 */
1746 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg);
1747 db->db_dirtycnt += 1;
1748 ASSERT3U(db->db_dirtycnt, <=, 3);
1749
1750 mutex_exit(&db->db_mtx);
1751
1752 if (db->db_blkid == DMU_BONUS_BLKID ||
1753 db->db_blkid == DMU_SPILL_BLKID) {
1754 mutex_enter(&dn->dn_mtx);
1755 ASSERT(!list_link_active(&dr->dr_dirty_node));
1756 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1757 mutex_exit(&dn->dn_mtx);
1758 dnode_setdirty(dn, tx);
1759 DB_DNODE_EXIT(db);
1760 return (dr);
1761 }
1762
1763 /*
1764 * The dn_struct_rwlock prevents db_blkptr from changing
1765 * due to a write from syncing context completing
1766 * while we are running, so we want to acquire it before
1767 * looking at db_blkptr.
1768 */
1769 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
1770 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1771 drop_struct_lock = TRUE;
1772 }
1773
1774 if (do_free_accounting) {
1775 blkptr_t *bp = db->db_blkptr;
1776 int64_t willfree = (bp && !BP_IS_HOLE(bp)) ?
1777 bp_get_dsize(os->os_spa, bp) : db->db.db_size;
1778 /*
1779 * This is only a guess -- if the dbuf is dirty
1780 * in a previous txg, we don't know how much
1781 * space it will use on disk yet. We should
1782 * really have the struct_rwlock to access
1783 * db_blkptr, but since this is just a guess,
1784 * it's OK if we get an odd answer.
1785 */
1786 ddt_prefetch(os->os_spa, bp);
1787 dnode_willuse_space(dn, -willfree, tx);
1788 }
1789
1790 if (db->db_level == 0) {
1791 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock);
1792 ASSERT(dn->dn_maxblkid >= db->db_blkid);
1793 }
1794
1795 if (db->db_level+1 < dn->dn_nlevels) {
1796 dmu_buf_impl_t *parent = db->db_parent;
1797 dbuf_dirty_record_t *di;
1798 int parent_held = FALSE;
1799
1800 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) {
1801 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1802
1803 parent = dbuf_hold_level(dn, db->db_level+1,
1804 db->db_blkid >> epbs, FTAG);
1805 ASSERT(parent != NULL);
1806 parent_held = TRUE;
1807 }
1808 if (drop_struct_lock)
1809 rw_exit(&dn->dn_struct_rwlock);
1810 ASSERT3U(db->db_level+1, ==, parent->db_level);
1811 di = dbuf_dirty(parent, tx);
1812 if (parent_held)
1813 dbuf_rele(parent, FTAG);
1814
1815 mutex_enter(&db->db_mtx);
1816 /*
1817 * Since we've dropped the mutex, it's possible that
1818 * dbuf_undirty() might have changed this out from under us.
1819 */
1820 if (db->db_last_dirty == dr ||
1821 dn->dn_object == DMU_META_DNODE_OBJECT) {
1822 mutex_enter(&di->dt.di.dr_mtx);
1823 ASSERT3U(di->dr_txg, ==, tx->tx_txg);
1824 ASSERT(!list_link_active(&dr->dr_dirty_node));
1825 list_insert_tail(&di->dt.di.dr_children, dr);
1826 mutex_exit(&di->dt.di.dr_mtx);
1827 dr->dr_parent = di;
1828 }
1829 mutex_exit(&db->db_mtx);
1830 } else {
1831 ASSERT(db->db_level+1 == dn->dn_nlevels);
1832 ASSERT(db->db_blkid < dn->dn_nblkptr);
1833 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf);
1834 mutex_enter(&dn->dn_mtx);
1835 ASSERT(!list_link_active(&dr->dr_dirty_node));
1836 list_insert_tail(&dn->dn_dirty_records[txgoff], dr);
1837 mutex_exit(&dn->dn_mtx);
1838 if (drop_struct_lock)
1839 rw_exit(&dn->dn_struct_rwlock);
1840 }
1841
1842 dnode_setdirty(dn, tx);
1843 DB_DNODE_EXIT(db);
1844 return (dr);
1845 }
1846
1847 /*
1848 * Undirty a buffer in the transaction group referenced by the given
1849 * transaction. Return whether this evicted the dbuf.
1850 */
1851 static boolean_t
1852 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx)
1853 {
1854 dnode_t *dn;
1855 uint64_t txg = tx->tx_txg;
1856 dbuf_dirty_record_t *dr, **drp;
1857
1858 ASSERT(txg != 0);
1859
1860 /*
1861 * Due to our use of dn_nlevels below, this can only be called
1862 * in open context, unless we are operating on the MOS.
1863 * From syncing context, dn_nlevels may be different from the
1864 * dn_nlevels used when dbuf was dirtied.
1865 */
1866 ASSERT(db->db_objset ==
1867 dmu_objset_pool(db->db_objset)->dp_meta_objset ||
1868 txg != spa_syncing_txg(dmu_objset_spa(db->db_objset)));
1869 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1870 ASSERT0(db->db_level);
1871 ASSERT(MUTEX_HELD(&db->db_mtx));
1872
1873 /*
1874 * If this buffer is not dirty, we're done.
1875 */
1876 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next)
1877 if (dr->dr_txg <= txg)
1878 break;
1879 if (dr == NULL || dr->dr_txg < txg)
1880 return (B_FALSE);
1881 ASSERT(dr->dr_txg == txg);
1882 ASSERT(dr->dr_dbuf == db);
1883
1884 DB_DNODE_ENTER(db);
1885 dn = DB_DNODE(db);
1886
1887 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size);
1888
1889 ASSERT(db->db.db_size != 0);
1890
1891 dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset),
1892 dr->dr_accounted, txg);
1893
1894 *drp = dr->dr_next;
1895
1896 /*
1897 * Note that there are three places in dbuf_dirty()
1898 * where this dirty record may be put on a list.
1899 * Make sure to do a list_remove corresponding to
1900 * every one of those list_insert calls.
1901 */
1902 if (dr->dr_parent) {
1903 mutex_enter(&dr->dr_parent->dt.di.dr_mtx);
1904 list_remove(&dr->dr_parent->dt.di.dr_children, dr);
1905 mutex_exit(&dr->dr_parent->dt.di.dr_mtx);
1906 } else if (db->db_blkid == DMU_SPILL_BLKID ||
1907 db->db_level + 1 == dn->dn_nlevels) {
1908 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf);
1909 mutex_enter(&dn->dn_mtx);
1910 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr);
1911 mutex_exit(&dn->dn_mtx);
1912 }
1913 DB_DNODE_EXIT(db);
1914
1915 if (db->db_state != DB_NOFILL) {
1916 dbuf_unoverride(dr);
1917
1918 ASSERT(db->db_buf != NULL);
1919 ASSERT(dr->dt.dl.dr_data != NULL);
1920 if (dr->dt.dl.dr_data != db->db_buf)
1921 arc_buf_destroy(dr->dt.dl.dr_data, db);
1922 }
1923
1924 kmem_free(dr, sizeof (dbuf_dirty_record_t));
1925
1926 ASSERT(db->db_dirtycnt > 0);
1927 db->db_dirtycnt -= 1;
1928
1929 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) {
1930 ASSERT(db->db_state == DB_NOFILL || arc_released(db->db_buf));
1931 dbuf_destroy(db);
1932 return (B_TRUE);
1933 }
1934
1935 return (B_FALSE);
1936 }
1937
1938 void
1939 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx)
1940 {
1941 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1942 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH;
1943 dbuf_dirty_record_t *dr;
1944
1945 ASSERT(tx->tx_txg != 0);
1946 ASSERT(!refcount_is_zero(&db->db_holds));
1947
1948 /*
1949 * Quick check for dirtyness. For already dirty blocks, this
1950 * reduces runtime of this function by >90%, and overall performance
1951 * by 50% for some workloads (e.g. file deletion with indirect blocks
1952 * cached).
1953 */
1954 mutex_enter(&db->db_mtx);
1955
1956 for (dr = db->db_last_dirty;
1957 dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) {
1958 /*
1959 * It's possible that it is already dirty but not cached,
1960 * because there are some calls to dbuf_dirty() that don't
1961 * go through dmu_buf_will_dirty().
1962 */
1963 if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) {
1964 /* This dbuf is already dirty and cached. */
1965 dbuf_redirty(dr);
1966 mutex_exit(&db->db_mtx);
1967 return;
1968 }
1969 }
1970 mutex_exit(&db->db_mtx);
1971
1972 DB_DNODE_ENTER(db);
1973 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock))
1974 rf |= DB_RF_HAVESTRUCT;
1975 DB_DNODE_EXIT(db);
1976 (void) dbuf_read(db, NULL, rf);
1977 (void) dbuf_dirty(db, tx);
1978 }
1979
1980 void
1981 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1982 {
1983 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1984
1985 db->db_state = DB_NOFILL;
1986
1987 dmu_buf_will_fill(db_fake, tx);
1988 }
1989
1990 void
1991 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx)
1992 {
1993 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1994
1995 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
1996 ASSERT(tx->tx_txg != 0);
1997 ASSERT(db->db_level == 0);
1998 ASSERT(!refcount_is_zero(&db->db_holds));
1999
2000 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT ||
2001 dmu_tx_private_ok(tx));
2002
2003 dbuf_noread(db);
2004 (void) dbuf_dirty(db, tx);
2005 }
2006
2007 #pragma weak dmu_buf_fill_done = dbuf_fill_done
2008 /* ARGSUSED */
2009 void
2010 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx)
2011 {
2012 mutex_enter(&db->db_mtx);
2013 DBUF_VERIFY(db);
2014
2015 if (db->db_state == DB_FILL) {
2016 if (db->db_level == 0 && db->db_freed_in_flight) {
2017 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2018 /* we were freed while filling */
2019 /* XXX dbuf_undirty? */
2020 bzero(db->db.db_data, db->db.db_size);
2021 db->db_freed_in_flight = FALSE;
2022 }
2023 db->db_state = DB_CACHED;
2024 cv_broadcast(&db->db_changed);
2025 }
2026 mutex_exit(&db->db_mtx);
2027 }
2028
2029 void
2030 dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data,
2031 bp_embedded_type_t etype, enum zio_compress comp,
2032 int uncompressed_size, int compressed_size, int byteorder,
2033 dmu_tx_t *tx)
2034 {
2035 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
2036 struct dirty_leaf *dl;
2037 dmu_object_type_t type;
2038
2039 if (etype == BP_EMBEDDED_TYPE_DATA) {
2040 ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset),
2041 SPA_FEATURE_EMBEDDED_DATA));
2042 }
2043
2044 DB_DNODE_ENTER(db);
2045 type = DB_DNODE(db)->dn_type;
2046 DB_DNODE_EXIT(db);
2047
2048 ASSERT0(db->db_level);
2049 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2050
2051 dmu_buf_will_not_fill(dbuf, tx);
2052
2053 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg);
2054 dl = &db->db_last_dirty->dt.dl;
2055 encode_embedded_bp_compressed(&dl->dr_overridden_by,
2056 data, comp, uncompressed_size, compressed_size);
2057 BPE_SET_ETYPE(&dl->dr_overridden_by, etype);
2058 BP_SET_TYPE(&dl->dr_overridden_by, type);
2059 BP_SET_LEVEL(&dl->dr_overridden_by, 0);
2060 BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder);
2061
2062 dl->dr_override_state = DR_OVERRIDDEN;
2063 dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg;
2064 }
2065
2066 /*
2067 * Directly assign a provided arc buf to a given dbuf if it's not referenced
2068 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf.
2069 */
2070 void
2071 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx)
2072 {
2073 ASSERT(!refcount_is_zero(&db->db_holds));
2074 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
2075 ASSERT(db->db_level == 0);
2076 ASSERT3U(dbuf_is_metadata(db), ==, arc_is_metadata(buf));
2077 ASSERT(buf != NULL);
2078 ASSERT(arc_buf_lsize(buf) == db->db.db_size);
2079 ASSERT(tx->tx_txg != 0);
2080
2081 arc_return_buf(buf, db);
2082 ASSERT(arc_released(buf));
2083
2084 mutex_enter(&db->db_mtx);
2085
2086 while (db->db_state == DB_READ || db->db_state == DB_FILL)
2087 cv_wait(&db->db_changed, &db->db_mtx);
2088
2089 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED);
2090
2091 if (db->db_state == DB_CACHED &&
2092 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) {
2093 mutex_exit(&db->db_mtx);
2094 (void) dbuf_dirty(db, tx);
2095 bcopy(buf->b_data, db->db.db_data, db->db.db_size);
2096 arc_buf_destroy(buf, db);
2097 xuio_stat_wbuf_copied();
2098 return;
2099 }
2100
2101 xuio_stat_wbuf_nocopy();
2102 if (db->db_state == DB_CACHED) {
2103 dbuf_dirty_record_t *dr = db->db_last_dirty;
2104
2105 ASSERT(db->db_buf != NULL);
2106 if (dr != NULL && dr->dr_txg == tx->tx_txg) {
2107 ASSERT(dr->dt.dl.dr_data == db->db_buf);
2108 if (!arc_released(db->db_buf)) {
2109 ASSERT(dr->dt.dl.dr_override_state ==
2110 DR_OVERRIDDEN);
2111 arc_release(db->db_buf, db);
2112 }
2113 dr->dt.dl.dr_data = buf;
2114 arc_buf_destroy(db->db_buf, db);
2115 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) {
2116 arc_release(db->db_buf, db);
2117 arc_buf_destroy(db->db_buf, db);
2118 }
2119 db->db_buf = NULL;
2120 }
2121 ASSERT(db->db_buf == NULL);
2122 dbuf_set_data(db, buf);
2123 db->db_state = DB_FILL;
2124 mutex_exit(&db->db_mtx);
2125 (void) dbuf_dirty(db, tx);
2126 dmu_buf_fill_done(&db->db, tx);
2127 }
2128
2129 void
2130 dbuf_destroy(dmu_buf_impl_t *db)
2131 {
2132 dnode_t *dn;
2133 dmu_buf_impl_t *parent = db->db_parent;
2134 dmu_buf_impl_t *dndb;
2135
2136 ASSERT(MUTEX_HELD(&db->db_mtx));
2137 ASSERT(refcount_is_zero(&db->db_holds));
2138
2139 if (db->db_buf != NULL) {
2140 arc_buf_destroy(db->db_buf, db);
2141 db->db_buf = NULL;
2142 }
2143
2144 if (db->db_blkid == DMU_BONUS_BLKID) {
2145 int slots = DB_DNODE(db)->dn_num_slots;
2146 int bonuslen = DN_SLOTS_TO_BONUSLEN(slots);
2147 ASSERT(db->db.db_data != NULL);
2148 zio_buf_free(db->db.db_data, bonuslen);
2149 arc_space_return(bonuslen, ARC_SPACE_BONUS);
2150 db->db_state = DB_UNCACHED;
2151 }
2152
2153 dbuf_clear_data(db);
2154
2155 if (multilist_link_active(&db->db_cache_link)) {
2156 multilist_remove(&dbuf_cache, db);
2157 (void) refcount_remove_many(&dbuf_cache_size,
2158 db->db.db_size, db);
2159 }
2160
2161 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL);
2162 ASSERT(db->db_data_pending == NULL);
2163
2164 db->db_state = DB_EVICTING;
2165 db->db_blkptr = NULL;
2166
2167 /*
2168 * Now that db_state is DB_EVICTING, nobody else can find this via
2169 * the hash table. We can now drop db_mtx, which allows us to
2170 * acquire the dn_dbufs_mtx.
2171 */
2172 mutex_exit(&db->db_mtx);
2173
2174 DB_DNODE_ENTER(db);
2175 dn = DB_DNODE(db);
2176 dndb = dn->dn_dbuf;
2177 if (db->db_blkid != DMU_BONUS_BLKID) {
2178 boolean_t needlock = !MUTEX_HELD(&dn->dn_dbufs_mtx);
2179 if (needlock)
2180 mutex_enter(&dn->dn_dbufs_mtx);
2181 avl_remove(&dn->dn_dbufs, db);
2182 atomic_dec_32(&dn->dn_dbufs_count);
2183 membar_producer();
2184 DB_DNODE_EXIT(db);
2185 if (needlock)
2186 mutex_exit(&dn->dn_dbufs_mtx);
2187 /*
2188 * Decrementing the dbuf count means that the hold corresponding
2189 * to the removed dbuf is no longer discounted in dnode_move(),
2190 * so the dnode cannot be moved until after we release the hold.
2191 * The membar_producer() ensures visibility of the decremented
2192 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually
2193 * release any lock.
2194 */
2195 dnode_rele(dn, db);
2196 db->db_dnode_handle = NULL;
2197
2198 dbuf_hash_remove(db);
2199 } else {
2200 DB_DNODE_EXIT(db);
2201 }
2202
2203 ASSERT(refcount_is_zero(&db->db_holds));
2204
2205 db->db_parent = NULL;
2206
2207 ASSERT(db->db_buf == NULL);
2208 ASSERT(db->db.db_data == NULL);
2209 ASSERT(db->db_hash_next == NULL);
2210 ASSERT(db->db_blkptr == NULL);
2211 ASSERT(db->db_data_pending == NULL);
2212 ASSERT(!multilist_link_active(&db->db_cache_link));
2213
2214 kmem_cache_free(dbuf_kmem_cache, db);
2215 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
2216
2217 /*
2218 * If this dbuf is referenced from an indirect dbuf,
2219 * decrement the ref count on the indirect dbuf.
2220 */
2221 if (parent && parent != dndb)
2222 dbuf_rele(parent, db);
2223 }
2224
2225 /*
2226 * Note: While bpp will always be updated if the function returns success,
2227 * parentp will not be updated if the dnode does not have dn_dbuf filled in;
2228 * this happens when the dnode is the meta-dnode, or a userused or groupused
2229 * object.
2230 */
2231 __attribute__((always_inline))
2232 static inline int
2233 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse,
2234 dmu_buf_impl_t **parentp, blkptr_t **bpp, struct dbuf_hold_impl_data *dh)
2235 {
2236 int nlevels, epbs;
2237
2238 *parentp = NULL;
2239 *bpp = NULL;
2240
2241 ASSERT(blkid != DMU_BONUS_BLKID);
2242
2243 if (blkid == DMU_SPILL_BLKID) {
2244 mutex_enter(&dn->dn_mtx);
2245 if (dn->dn_have_spill &&
2246 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR))
2247 *bpp = DN_SPILL_BLKPTR(dn->dn_phys);
2248 else
2249 *bpp = NULL;
2250 dbuf_add_ref(dn->dn_dbuf, NULL);
2251 *parentp = dn->dn_dbuf;
2252 mutex_exit(&dn->dn_mtx);
2253 return (0);
2254 }
2255
2256 nlevels =
2257 (dn->dn_phys->dn_nlevels == 0) ? 1 : dn->dn_phys->dn_nlevels;
2258 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
2259
2260 ASSERT3U(level * epbs, <, 64);
2261 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2262 /*
2263 * This assertion shouldn't trip as long as the max indirect block size
2264 * is less than 1M. The reason for this is that up to that point,
2265 * the number of levels required to address an entire object with blocks
2266 * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In
2267 * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55
2268 * (i.e. we can address the entire object), objects will all use at most
2269 * N-1 levels and the assertion won't overflow. However, once epbs is
2270 * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be
2271 * enough to address an entire object, so objects will have 5 levels,
2272 * but then this assertion will overflow.
2273 *
2274 * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we
2275 * need to redo this logic to handle overflows.
2276 */
2277 ASSERT(level >= nlevels ||
2278 ((nlevels - level - 1) * epbs) +
2279 highbit64(dn->dn_phys->dn_nblkptr) <= 64);
2280 if (level >= nlevels ||
2281 blkid >= ((uint64_t)dn->dn_phys->dn_nblkptr <<
2282 ((nlevels - level - 1) * epbs)) ||
2283 (fail_sparse &&
2284 blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) {
2285 /* the buffer has no parent yet */
2286 return (SET_ERROR(ENOENT));
2287 } else if (level < nlevels-1) {
2288 /* this block is referenced from an indirect block */
2289 int err;
2290 if (dh == NULL) {
2291 err = dbuf_hold_impl(dn, level+1,
2292 blkid >> epbs, fail_sparse, FALSE, NULL, parentp);
2293 } else {
2294 __dbuf_hold_impl_init(dh + 1, dn, dh->dh_level + 1,
2295 blkid >> epbs, fail_sparse, FALSE, NULL,
2296 parentp, dh->dh_depth + 1);
2297 err = __dbuf_hold_impl(dh + 1);
2298 }
2299 if (err)
2300 return (err);
2301 err = dbuf_read(*parentp, NULL,
2302 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL));
2303 if (err) {
2304 dbuf_rele(*parentp, NULL);
2305 *parentp = NULL;
2306 return (err);
2307 }
2308 *bpp = ((blkptr_t *)(*parentp)->db.db_data) +
2309 (blkid & ((1ULL << epbs) - 1));
2310 if (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))
2311 ASSERT(BP_IS_HOLE(*bpp));
2312 return (0);
2313 } else {
2314 /* the block is referenced from the dnode */
2315 ASSERT3U(level, ==, nlevels-1);
2316 ASSERT(dn->dn_phys->dn_nblkptr == 0 ||
2317 blkid < dn->dn_phys->dn_nblkptr);
2318 if (dn->dn_dbuf) {
2319 dbuf_add_ref(dn->dn_dbuf, NULL);
2320 *parentp = dn->dn_dbuf;
2321 }
2322 *bpp = &dn->dn_phys->dn_blkptr[blkid];
2323 return (0);
2324 }
2325 }
2326
2327 static dmu_buf_impl_t *
2328 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid,
2329 dmu_buf_impl_t *parent, blkptr_t *blkptr)
2330 {
2331 objset_t *os = dn->dn_objset;
2332 dmu_buf_impl_t *db, *odb;
2333
2334 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2335 ASSERT(dn->dn_type != DMU_OT_NONE);
2336
2337 db = kmem_cache_alloc(dbuf_kmem_cache, KM_SLEEP);
2338
2339 db->db_objset = os;
2340 db->db.db_object = dn->dn_object;
2341 db->db_level = level;
2342 db->db_blkid = blkid;
2343 db->db_last_dirty = NULL;
2344 db->db_dirtycnt = 0;
2345 db->db_dnode_handle = dn->dn_handle;
2346 db->db_parent = parent;
2347 db->db_blkptr = blkptr;
2348
2349 db->db_user = NULL;
2350 db->db_user_immediate_evict = FALSE;
2351 db->db_freed_in_flight = FALSE;
2352 db->db_pending_evict = FALSE;
2353
2354 if (blkid == DMU_BONUS_BLKID) {
2355 ASSERT3P(parent, ==, dn->dn_dbuf);
2356 db->db.db_size = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
2357 (dn->dn_nblkptr-1) * sizeof (blkptr_t);
2358 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen);
2359 db->db.db_offset = DMU_BONUS_BLKID;
2360 db->db_state = DB_UNCACHED;
2361 /* the bonus dbuf is not placed in the hash table */
2362 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
2363 return (db);
2364 } else if (blkid == DMU_SPILL_BLKID) {
2365 db->db.db_size = (blkptr != NULL) ?
2366 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE;
2367 db->db.db_offset = 0;
2368 } else {
2369 int blocksize =
2370 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz;
2371 db->db.db_size = blocksize;
2372 db->db.db_offset = db->db_blkid * blocksize;
2373 }
2374
2375 /*
2376 * Hold the dn_dbufs_mtx while we get the new dbuf
2377 * in the hash table *and* added to the dbufs list.
2378 * This prevents a possible deadlock with someone
2379 * trying to look up this dbuf before its added to the
2380 * dn_dbufs list.
2381 */
2382 mutex_enter(&dn->dn_dbufs_mtx);
2383 db->db_state = DB_EVICTING;
2384 if ((odb = dbuf_hash_insert(db)) != NULL) {
2385 /* someone else inserted it first */
2386 kmem_cache_free(dbuf_kmem_cache, db);
2387 mutex_exit(&dn->dn_dbufs_mtx);
2388 return (odb);
2389 }
2390 avl_add(&dn->dn_dbufs, db);
2391 if (db->db_level == 0 && db->db_blkid >=
2392 dn->dn_unlisted_l0_blkid)
2393 dn->dn_unlisted_l0_blkid = db->db_blkid + 1;
2394 db->db_state = DB_UNCACHED;
2395 mutex_exit(&dn->dn_dbufs_mtx);
2396 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF);
2397
2398 if (parent && parent != dn->dn_dbuf)
2399 dbuf_add_ref(parent, db);
2400
2401 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT ||
2402 refcount_count(&dn->dn_holds) > 0);
2403 (void) refcount_add(&dn->dn_holds, db);
2404 atomic_inc_32(&dn->dn_dbufs_count);
2405
2406 dprintf_dbuf(db, "db=%p\n", db);
2407
2408 return (db);
2409 }
2410
2411 typedef struct dbuf_prefetch_arg {
2412 spa_t *dpa_spa; /* The spa to issue the prefetch in. */
2413 zbookmark_phys_t dpa_zb; /* The target block to prefetch. */
2414 int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */
2415 int dpa_curlevel; /* The current level that we're reading */
2416 dnode_t *dpa_dnode; /* The dnode associated with the prefetch */
2417 zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */
2418 zio_t *dpa_zio; /* The parent zio_t for all prefetches. */
2419 arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */
2420 } dbuf_prefetch_arg_t;
2421
2422 /*
2423 * Actually issue the prefetch read for the block given.
2424 */
2425 static void
2426 dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp)
2427 {
2428 arc_flags_t aflags;
2429 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp))
2430 return;
2431
2432 aflags = dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH;
2433
2434 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2435 ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level);
2436 ASSERT(dpa->dpa_zio != NULL);
2437 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL,
2438 dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2439 &aflags, &dpa->dpa_zb);
2440 }
2441
2442 /*
2443 * Called when an indirect block above our prefetch target is read in. This
2444 * will either read in the next indirect block down the tree or issue the actual
2445 * prefetch if the next block down is our target.
2446 */
2447 static void
2448 dbuf_prefetch_indirect_done(zio_t *zio, arc_buf_t *abuf, void *private)
2449 {
2450 dbuf_prefetch_arg_t *dpa = private;
2451 uint64_t nextblkid;
2452 blkptr_t *bp;
2453
2454 ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel);
2455 ASSERT3S(dpa->dpa_curlevel, >, 0);
2456
2457 /*
2458 * The dpa_dnode is only valid if we are called with a NULL
2459 * zio. This indicates that the arc_read() returned without
2460 * first calling zio_read() to issue a physical read. Once
2461 * a physical read is made the dpa_dnode must be invalidated
2462 * as the locks guarding it may have been dropped. If the
2463 * dpa_dnode is still valid, then we want to add it to the dbuf
2464 * cache. To do so, we must hold the dbuf associated with the block
2465 * we just prefetched, read its contents so that we associate it
2466 * with an arc_buf_t, and then release it.
2467 */
2468 if (zio != NULL) {
2469 ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel);
2470 if (zio->io_flags & ZIO_FLAG_RAW) {
2471 ASSERT3U(BP_GET_PSIZE(zio->io_bp), ==, zio->io_size);
2472 } else {
2473 ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size);
2474 }
2475 ASSERT3P(zio->io_spa, ==, dpa->dpa_spa);
2476
2477 dpa->dpa_dnode = NULL;
2478 } else if (dpa->dpa_dnode != NULL) {
2479 uint64_t curblkid = dpa->dpa_zb.zb_blkid >>
2480 (dpa->dpa_epbs * (dpa->dpa_curlevel -
2481 dpa->dpa_zb.zb_level));
2482 dmu_buf_impl_t *db = dbuf_hold_level(dpa->dpa_dnode,
2483 dpa->dpa_curlevel, curblkid, FTAG);
2484 (void) dbuf_read(db, NULL,
2485 DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_HAVESTRUCT);
2486 dbuf_rele(db, FTAG);
2487 }
2488
2489 dpa->dpa_curlevel--;
2490
2491 nextblkid = dpa->dpa_zb.zb_blkid >>
2492 (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level));
2493 bp = ((blkptr_t *)abuf->b_data) +
2494 P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs);
2495 if (BP_IS_HOLE(bp) || (zio != NULL && zio->io_error != 0)) {
2496 kmem_free(dpa, sizeof (*dpa));
2497 } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) {
2498 ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid);
2499 dbuf_issue_final_prefetch(dpa, bp);
2500 kmem_free(dpa, sizeof (*dpa));
2501 } else {
2502 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2503 zbookmark_phys_t zb;
2504
2505 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp));
2506
2507 SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset,
2508 dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid);
2509
2510 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2511 bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio,
2512 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2513 &iter_aflags, &zb);
2514 }
2515
2516 arc_buf_destroy(abuf, private);
2517 }
2518
2519 /*
2520 * Issue prefetch reads for the given block on the given level. If the indirect
2521 * blocks above that block are not in memory, we will read them in
2522 * asynchronously. As a result, this call never blocks waiting for a read to
2523 * complete.
2524 */
2525 void
2526 dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio,
2527 arc_flags_t aflags)
2528 {
2529 blkptr_t bp;
2530 int epbs, nlevels, curlevel;
2531 uint64_t curblkid;
2532 dmu_buf_impl_t *db;
2533 zio_t *pio;
2534 dbuf_prefetch_arg_t *dpa;
2535 dsl_dataset_t *ds;
2536
2537 ASSERT(blkid != DMU_BONUS_BLKID);
2538 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock));
2539
2540 if (blkid > dn->dn_maxblkid)
2541 return;
2542
2543 if (dnode_block_freed(dn, blkid))
2544 return;
2545
2546 /*
2547 * This dnode hasn't been written to disk yet, so there's nothing to
2548 * prefetch.
2549 */
2550 nlevels = dn->dn_phys->dn_nlevels;
2551 if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0)
2552 return;
2553
2554 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
2555 if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level))
2556 return;
2557
2558 db = dbuf_find(dn->dn_objset, dn->dn_object,
2559 level, blkid);
2560 if (db != NULL) {
2561 mutex_exit(&db->db_mtx);
2562 /*
2563 * This dbuf already exists. It is either CACHED, or
2564 * (we assume) about to be read or filled.
2565 */
2566 return;
2567 }
2568
2569 /*
2570 * Find the closest ancestor (indirect block) of the target block
2571 * that is present in the cache. In this indirect block, we will
2572 * find the bp that is at curlevel, curblkid.
2573 */
2574 curlevel = level;
2575 curblkid = blkid;
2576 while (curlevel < nlevels - 1) {
2577 int parent_level = curlevel + 1;
2578 uint64_t parent_blkid = curblkid >> epbs;
2579 dmu_buf_impl_t *db;
2580
2581 if (dbuf_hold_impl(dn, parent_level, parent_blkid,
2582 FALSE, TRUE, FTAG, &db) == 0) {
2583 blkptr_t *bpp = db->db_buf->b_data;
2584 bp = bpp[P2PHASE(curblkid, 1 << epbs)];
2585 dbuf_rele(db, FTAG);
2586 break;
2587 }
2588
2589 curlevel = parent_level;
2590 curblkid = parent_blkid;
2591 }
2592
2593 if (curlevel == nlevels - 1) {
2594 /* No cached indirect blocks found. */
2595 ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr);
2596 bp = dn->dn_phys->dn_blkptr[curblkid];
2597 }
2598 if (BP_IS_HOLE(&bp))
2599 return;
2600
2601 ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp));
2602
2603 pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL,
2604 ZIO_FLAG_CANFAIL);
2605
2606 dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP);
2607 ds = dn->dn_objset->os_dsl_dataset;
2608 SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2609 dn->dn_object, level, blkid);
2610 dpa->dpa_curlevel = curlevel;
2611 dpa->dpa_prio = prio;
2612 dpa->dpa_aflags = aflags;
2613 dpa->dpa_spa = dn->dn_objset->os_spa;
2614 dpa->dpa_dnode = dn;
2615 dpa->dpa_epbs = epbs;
2616 dpa->dpa_zio = pio;
2617
2618 /*
2619 * If we have the indirect just above us, no need to do the asynchronous
2620 * prefetch chain; we'll just run the last step ourselves. If we're at
2621 * a higher level, though, we want to issue the prefetches for all the
2622 * indirect blocks asynchronously, so we can go on with whatever we were
2623 * doing.
2624 */
2625 if (curlevel == level) {
2626 ASSERT3U(curblkid, ==, blkid);
2627 dbuf_issue_final_prefetch(dpa, &bp);
2628 kmem_free(dpa, sizeof (*dpa));
2629 } else {
2630 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT;
2631 zbookmark_phys_t zb;
2632
2633 SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET,
2634 dn->dn_object, curlevel, curblkid);
2635 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa,
2636 &bp, dbuf_prefetch_indirect_done, dpa, prio,
2637 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE,
2638 &iter_aflags, &zb);
2639 }
2640 /*
2641 * We use pio here instead of dpa_zio since it's possible that
2642 * dpa may have already been freed.
2643 */
2644 zio_nowait(pio);
2645 }
2646
2647 #define DBUF_HOLD_IMPL_MAX_DEPTH 20
2648
2649 /*
2650 * Returns with db_holds incremented, and db_mtx not held.
2651 * Note: dn_struct_rwlock must be held.
2652 */
2653 static int
2654 __dbuf_hold_impl(struct dbuf_hold_impl_data *dh)
2655 {
2656 ASSERT3S(dh->dh_depth, <, DBUF_HOLD_IMPL_MAX_DEPTH);
2657 dh->dh_parent = NULL;
2658
2659 ASSERT(dh->dh_blkid != DMU_BONUS_BLKID);
2660 ASSERT(RW_LOCK_HELD(&dh->dh_dn->dn_struct_rwlock));
2661 ASSERT3U(dh->dh_dn->dn_nlevels, >, dh->dh_level);
2662
2663 *(dh->dh_dbp) = NULL;
2664
2665 /* dbuf_find() returns with db_mtx held */
2666 dh->dh_db = dbuf_find(dh->dh_dn->dn_objset, dh->dh_dn->dn_object,
2667 dh->dh_level, dh->dh_blkid);
2668
2669 if (dh->dh_db == NULL) {
2670 dh->dh_bp = NULL;
2671
2672 if (dh->dh_fail_uncached)
2673 return (SET_ERROR(ENOENT));
2674
2675 ASSERT3P(dh->dh_parent, ==, NULL);
2676 dh->dh_err = dbuf_findbp(dh->dh_dn, dh->dh_level, dh->dh_blkid,
2677 dh->dh_fail_sparse, &dh->dh_parent,
2678 &dh->dh_bp, dh);
2679 if (dh->dh_fail_sparse) {
2680 if (dh->dh_err == 0 &&
2681 dh->dh_bp && BP_IS_HOLE(dh->dh_bp))
2682 dh->dh_err = SET_ERROR(ENOENT);
2683 if (dh->dh_err) {
2684 if (dh->dh_parent)
2685 dbuf_rele(dh->dh_parent, NULL);
2686 return (dh->dh_err);
2687 }
2688 }
2689 if (dh->dh_err && dh->dh_err != ENOENT)
2690 return (dh->dh_err);
2691 dh->dh_db = dbuf_create(dh->dh_dn, dh->dh_level, dh->dh_blkid,
2692 dh->dh_parent, dh->dh_bp);
2693 }
2694
2695 if (dh->dh_fail_uncached && dh->dh_db->db_state != DB_CACHED) {
2696 mutex_exit(&dh->dh_db->db_mtx);
2697 return (SET_ERROR(ENOENT));
2698 }
2699
2700 if (dh->dh_db->db_buf != NULL)
2701 ASSERT3P(dh->dh_db->db.db_data, ==, dh->dh_db->db_buf->b_data);
2702
2703 ASSERT(dh->dh_db->db_buf == NULL || arc_referenced(dh->dh_db->db_buf));
2704
2705 /*
2706 * If this buffer is currently syncing out, and we are are
2707 * still referencing it from db_data, we need to make a copy
2708 * of it in case we decide we want to dirty it again in this txg.
2709 */
2710 if (dh->dh_db->db_level == 0 &&
2711 dh->dh_db->db_blkid != DMU_BONUS_BLKID &&
2712 dh->dh_dn->dn_object != DMU_META_DNODE_OBJECT &&
2713 dh->dh_db->db_state == DB_CACHED && dh->dh_db->db_data_pending) {
2714 dh->dh_dr = dh->dh_db->db_data_pending;
2715
2716 if (dh->dh_dr->dt.dl.dr_data == dh->dh_db->db_buf) {
2717 dh->dh_type = DBUF_GET_BUFC_TYPE(dh->dh_db);
2718
2719 dbuf_set_data(dh->dh_db,
2720 arc_alloc_buf(dh->dh_dn->dn_objset->os_spa,
2721 dh->dh_db, dh->dh_type, dh->dh_db->db.db_size));
2722 bcopy(dh->dh_dr->dt.dl.dr_data->b_data,
2723 dh->dh_db->db.db_data, dh->dh_db->db.db_size);
2724 }
2725 }
2726
2727 if (multilist_link_active(&dh->dh_db->db_cache_link)) {
2728 ASSERT(refcount_is_zero(&dh->dh_db->db_holds));
2729 multilist_remove(&dbuf_cache, dh->dh_db);
2730 (void) refcount_remove_many(&dbuf_cache_size,
2731 dh->dh_db->db.db_size, dh->dh_db);
2732 }
2733 (void) refcount_add(&dh->dh_db->db_holds, dh->dh_tag);
2734 DBUF_VERIFY(dh->dh_db);
2735 mutex_exit(&dh->dh_db->db_mtx);
2736
2737 /* NOTE: we can't rele the parent until after we drop the db_mtx */
2738 if (dh->dh_parent)
2739 dbuf_rele(dh->dh_parent, NULL);
2740
2741 ASSERT3P(DB_DNODE(dh->dh_db), ==, dh->dh_dn);
2742 ASSERT3U(dh->dh_db->db_blkid, ==, dh->dh_blkid);
2743 ASSERT3U(dh->dh_db->db_level, ==, dh->dh_level);
2744 *(dh->dh_dbp) = dh->dh_db;
2745
2746 return (0);
2747 }
2748
2749 /*
2750 * The following code preserves the recursive function dbuf_hold_impl()
2751 * but moves the local variables AND function arguments to the heap to
2752 * minimize the stack frame size. Enough space is initially allocated
2753 * on the stack for 20 levels of recursion.
2754 */
2755 int
2756 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid,
2757 boolean_t fail_sparse, boolean_t fail_uncached,
2758 void *tag, dmu_buf_impl_t **dbp)
2759 {
2760 struct dbuf_hold_impl_data *dh;
2761 int error;
2762
2763 dh = kmem_alloc(sizeof (struct dbuf_hold_impl_data) *
2764 DBUF_HOLD_IMPL_MAX_DEPTH, KM_SLEEP);
2765 __dbuf_hold_impl_init(dh, dn, level, blkid, fail_sparse,
2766 fail_uncached, tag, dbp, 0);
2767
2768 error = __dbuf_hold_impl(dh);
2769
2770 kmem_free(dh, sizeof (struct dbuf_hold_impl_data) *
2771 DBUF_HOLD_IMPL_MAX_DEPTH);
2772
2773 return (error);
2774 }
2775
2776 static void
2777 __dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh,
2778 dnode_t *dn, uint8_t level, uint64_t blkid,
2779 boolean_t fail_sparse, boolean_t fail_uncached,
2780 void *tag, dmu_buf_impl_t **dbp, int depth)
2781 {
2782 dh->dh_dn = dn;
2783 dh->dh_level = level;
2784 dh->dh_blkid = blkid;
2785
2786 dh->dh_fail_sparse = fail_sparse;
2787 dh->dh_fail_uncached = fail_uncached;
2788
2789 dh->dh_tag = tag;
2790 dh->dh_dbp = dbp;
2791
2792 dh->dh_db = NULL;
2793 dh->dh_parent = NULL;
2794 dh->dh_bp = NULL;
2795 dh->dh_err = 0;
2796 dh->dh_dr = NULL;
2797 dh->dh_type = 0;
2798
2799 dh->dh_depth = depth;
2800 }
2801
2802 dmu_buf_impl_t *
2803 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag)
2804 {
2805 return (dbuf_hold_level(dn, 0, blkid, tag));
2806 }
2807
2808 dmu_buf_impl_t *
2809 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag)
2810 {
2811 dmu_buf_impl_t *db;
2812 int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db);
2813 return (err ? NULL : db);
2814 }
2815
2816 void
2817 dbuf_create_bonus(dnode_t *dn)
2818 {
2819 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
2820
2821 ASSERT(dn->dn_bonus == NULL);
2822 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL);
2823 }
2824
2825 int
2826 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx)
2827 {
2828 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2829 dnode_t *dn;
2830
2831 if (db->db_blkid != DMU_SPILL_BLKID)
2832 return (SET_ERROR(ENOTSUP));
2833 if (blksz == 0)
2834 blksz = SPA_MINBLOCKSIZE;
2835 ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset)));
2836 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE);
2837
2838 DB_DNODE_ENTER(db);
2839 dn = DB_DNODE(db);
2840 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
2841 dbuf_new_size(db, blksz, tx);
2842 rw_exit(&dn->dn_struct_rwlock);
2843 DB_DNODE_EXIT(db);
2844
2845 return (0);
2846 }
2847
2848 void
2849 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx)
2850 {
2851 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx);
2852 }
2853
2854 #pragma weak dmu_buf_add_ref = dbuf_add_ref
2855 void
2856 dbuf_add_ref(dmu_buf_impl_t *db, void *tag)
2857 {
2858 int64_t holds = refcount_add(&db->db_holds, tag);
2859 VERIFY3S(holds, >, 1);
2860 }
2861
2862 #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref
2863 boolean_t
2864 dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid,
2865 void *tag)
2866 {
2867 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2868 dmu_buf_impl_t *found_db;
2869 boolean_t result = B_FALSE;
2870
2871 if (blkid == DMU_BONUS_BLKID)
2872 found_db = dbuf_find_bonus(os, obj);
2873 else
2874 found_db = dbuf_find(os, obj, 0, blkid);
2875
2876 if (found_db != NULL) {
2877 if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) {
2878 (void) refcount_add(&db->db_holds, tag);
2879 result = B_TRUE;
2880 }
2881 mutex_exit(&found_db->db_mtx);
2882 }
2883 return (result);
2884 }
2885
2886 /*
2887 * If you call dbuf_rele() you had better not be referencing the dnode handle
2888 * unless you have some other direct or indirect hold on the dnode. (An indirect
2889 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.)
2890 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the
2891 * dnode's parent dbuf evicting its dnode handles.
2892 */
2893 void
2894 dbuf_rele(dmu_buf_impl_t *db, void *tag)
2895 {
2896 mutex_enter(&db->db_mtx);
2897 dbuf_rele_and_unlock(db, tag);
2898 }
2899
2900 void
2901 dmu_buf_rele(dmu_buf_t *db, void *tag)
2902 {
2903 dbuf_rele((dmu_buf_impl_t *)db, tag);
2904 }
2905
2906 /*
2907 * dbuf_rele() for an already-locked dbuf. This is necessary to allow
2908 * db_dirtycnt and db_holds to be updated atomically.
2909 */
2910 void
2911 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag)
2912 {
2913 int64_t holds;
2914
2915 ASSERT(MUTEX_HELD(&db->db_mtx));
2916 DBUF_VERIFY(db);
2917
2918 /*
2919 * Remove the reference to the dbuf before removing its hold on the
2920 * dnode so we can guarantee in dnode_move() that a referenced bonus
2921 * buffer has a corresponding dnode hold.
2922 */
2923 holds = refcount_remove(&db->db_holds, tag);
2924 ASSERT(holds >= 0);
2925
2926 /*
2927 * We can't freeze indirects if there is a possibility that they
2928 * may be modified in the current syncing context.
2929 */
2930 if (db->db_buf != NULL &&
2931 holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) {
2932 arc_buf_freeze(db->db_buf);
2933 }
2934
2935 if (holds == db->db_dirtycnt &&
2936 db->db_level == 0 && db->db_user_immediate_evict)
2937 dbuf_evict_user(db);
2938
2939 if (holds == 0) {
2940 if (db->db_blkid == DMU_BONUS_BLKID) {
2941 dnode_t *dn;
2942 boolean_t evict_dbuf = db->db_pending_evict;
2943
2944 /*
2945 * If the dnode moves here, we cannot cross this
2946 * barrier until the move completes.
2947 */
2948 DB_DNODE_ENTER(db);
2949
2950 dn = DB_DNODE(db);
2951 atomic_dec_32(&dn->dn_dbufs_count);
2952
2953 /*
2954 * Decrementing the dbuf count means that the bonus
2955 * buffer's dnode hold is no longer discounted in
2956 * dnode_move(). The dnode cannot move until after
2957 * the dnode_rele() below.
2958 */
2959 DB_DNODE_EXIT(db);
2960
2961 /*
2962 * Do not reference db after its lock is dropped.
2963 * Another thread may evict it.
2964 */
2965 mutex_exit(&db->db_mtx);
2966
2967 if (evict_dbuf)
2968 dnode_evict_bonus(dn);
2969
2970 dnode_rele(dn, db);
2971 } else if (db->db_buf == NULL) {
2972 /*
2973 * This is a special case: we never associated this
2974 * dbuf with any data allocated from the ARC.
2975 */
2976 ASSERT(db->db_state == DB_UNCACHED ||
2977 db->db_state == DB_NOFILL);
2978 dbuf_destroy(db);
2979 } else if (arc_released(db->db_buf)) {
2980 /*
2981 * This dbuf has anonymous data associated with it.
2982 */
2983 dbuf_destroy(db);
2984 } else {
2985 boolean_t do_arc_evict = B_FALSE;
2986 blkptr_t bp;
2987 spa_t *spa = dmu_objset_spa(db->db_objset);
2988
2989 if (!DBUF_IS_CACHEABLE(db) &&
2990 db->db_blkptr != NULL &&
2991 !BP_IS_HOLE(db->db_blkptr) &&
2992 !BP_IS_EMBEDDED(db->db_blkptr)) {
2993 do_arc_evict = B_TRUE;
2994 bp = *db->db_blkptr;
2995 }
2996
2997 if (!DBUF_IS_CACHEABLE(db) ||
2998 db->db_pending_evict) {
2999 dbuf_destroy(db);
3000 } else if (!multilist_link_active(&db->db_cache_link)) {
3001 multilist_insert(&dbuf_cache, db);
3002 (void) refcount_add_many(&dbuf_cache_size,
3003 db->db.db_size, db);
3004 mutex_exit(&db->db_mtx);
3005
3006 dbuf_evict_notify();
3007 }
3008
3009 if (do_arc_evict)
3010 arc_freed(spa, &bp);
3011 }
3012 } else {
3013 mutex_exit(&db->db_mtx);
3014 }
3015
3016 }
3017
3018 #pragma weak dmu_buf_refcount = dbuf_refcount
3019 uint64_t
3020 dbuf_refcount(dmu_buf_impl_t *db)
3021 {
3022 return (refcount_count(&db->db_holds));
3023 }
3024
3025 void *
3026 dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user,
3027 dmu_buf_user_t *new_user)
3028 {
3029 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3030
3031 mutex_enter(&db->db_mtx);
3032 dbuf_verify_user(db, DBVU_NOT_EVICTING);
3033 if (db->db_user == old_user)
3034 db->db_user = new_user;
3035 else
3036 old_user = db->db_user;
3037 dbuf_verify_user(db, DBVU_NOT_EVICTING);
3038 mutex_exit(&db->db_mtx);
3039
3040 return (old_user);
3041 }
3042
3043 void *
3044 dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
3045 {
3046 return (dmu_buf_replace_user(db_fake, NULL, user));
3047 }
3048
3049 void *
3050 dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user)
3051 {
3052 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3053
3054 db->db_user_immediate_evict = TRUE;
3055 return (dmu_buf_set_user(db_fake, user));
3056 }
3057
3058 void *
3059 dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user)
3060 {
3061 return (dmu_buf_replace_user(db_fake, user, NULL));
3062 }
3063
3064 void *
3065 dmu_buf_get_user(dmu_buf_t *db_fake)
3066 {
3067 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
3068
3069 dbuf_verify_user(db, DBVU_NOT_EVICTING);
3070 return (db->db_user);
3071 }
3072
3073 void
3074 dmu_buf_user_evict_wait()
3075 {
3076 taskq_wait(dbu_evict_taskq);
3077 }
3078
3079 boolean_t
3080 dmu_buf_freeable(dmu_buf_t *dbuf)
3081 {
3082 boolean_t res = B_FALSE;
3083 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf;
3084
3085 if (db->db_blkptr)
3086 res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset,
3087 db->db_blkptr, db->db_blkptr->blk_birth);
3088
3089 return (res);
3090 }
3091
3092 blkptr_t *
3093 dmu_buf_get_blkptr(dmu_buf_t *db)
3094 {
3095 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3096 return (dbi->db_blkptr);
3097 }
3098
3099 objset_t *
3100 dmu_buf_get_objset(dmu_buf_t *db)
3101 {
3102 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3103 return (dbi->db_objset);
3104 }
3105
3106 dnode_t *
3107 dmu_buf_dnode_enter(dmu_buf_t *db)
3108 {
3109 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3110 DB_DNODE_ENTER(dbi);
3111 return (DB_DNODE(dbi));
3112 }
3113
3114 void
3115 dmu_buf_dnode_exit(dmu_buf_t *db)
3116 {
3117 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
3118 DB_DNODE_EXIT(dbi);
3119 }
3120
3121 static void
3122 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db)
3123 {
3124 /* ASSERT(dmu_tx_is_syncing(tx) */
3125 ASSERT(MUTEX_HELD(&db->db_mtx));
3126
3127 if (db->db_blkptr != NULL)
3128 return;
3129
3130 if (db->db_blkid == DMU_SPILL_BLKID) {
3131 db->db_blkptr = DN_SPILL_BLKPTR(dn->dn_phys);
3132 BP_ZERO(db->db_blkptr);
3133 return;
3134 }
3135 if (db->db_level == dn->dn_phys->dn_nlevels-1) {
3136 /*
3137 * This buffer was allocated at a time when there was
3138 * no available blkptrs from the dnode, or it was
3139 * inappropriate to hook it in (i.e., nlevels mis-match).
3140 */
3141 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr);
3142 ASSERT(db->db_parent == NULL);
3143 db->db_parent = dn->dn_dbuf;
3144 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid];
3145 DBUF_VERIFY(db);
3146 } else {
3147 dmu_buf_impl_t *parent = db->db_parent;
3148 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3149
3150 ASSERT(dn->dn_phys->dn_nlevels > 1);
3151 if (parent == NULL) {
3152 mutex_exit(&db->db_mtx);
3153 rw_enter(&dn->dn_struct_rwlock, RW_READER);
3154 parent = dbuf_hold_level(dn, db->db_level + 1,
3155 db->db_blkid >> epbs, db);
3156 rw_exit(&dn->dn_struct_rwlock);
3157 mutex_enter(&db->db_mtx);
3158 db->db_parent = parent;
3159 }
3160 db->db_blkptr = (blkptr_t *)parent->db.db_data +
3161 (db->db_blkid & ((1ULL << epbs) - 1));
3162 DBUF_VERIFY(db);
3163 }
3164 }
3165
3166 /*
3167 * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it
3168 * is critical the we not allow the compiler to inline this function in to
3169 * dbuf_sync_list() thereby drastically bloating the stack usage.
3170 */
3171 noinline static void
3172 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
3173 {
3174 dmu_buf_impl_t *db = dr->dr_dbuf;
3175 dnode_t *dn;
3176 zio_t *zio;
3177
3178 ASSERT(dmu_tx_is_syncing(tx));
3179
3180 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
3181
3182 mutex_enter(&db->db_mtx);
3183
3184 ASSERT(db->db_level > 0);
3185 DBUF_VERIFY(db);
3186
3187 /* Read the block if it hasn't been read yet. */
3188 if (db->db_buf == NULL) {
3189 mutex_exit(&db->db_mtx);
3190 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED);
3191 mutex_enter(&db->db_mtx);
3192 }
3193 ASSERT3U(db->db_state, ==, DB_CACHED);
3194 ASSERT(db->db_buf != NULL);
3195
3196 DB_DNODE_ENTER(db);
3197 dn = DB_DNODE(db);
3198 /* Indirect block size must match what the dnode thinks it is. */
3199 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3200 dbuf_check_blkptr(dn, db);
3201 DB_DNODE_EXIT(db);
3202
3203 /* Provide the pending dirty record to child dbufs */
3204 db->db_data_pending = dr;
3205
3206 mutex_exit(&db->db_mtx);
3207 dbuf_write(dr, db->db_buf, tx);
3208
3209 zio = dr->dr_zio;
3210 mutex_enter(&dr->dt.di.dr_mtx);
3211 dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx);
3212 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3213 mutex_exit(&dr->dt.di.dr_mtx);
3214 zio_nowait(zio);
3215 }
3216
3217 /*
3218 * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is
3219 * critical the we not allow the compiler to inline this function in to
3220 * dbuf_sync_list() thereby drastically bloating the stack usage.
3221 */
3222 noinline static void
3223 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx)
3224 {
3225 arc_buf_t **datap = &dr->dt.dl.dr_data;
3226 dmu_buf_impl_t *db = dr->dr_dbuf;
3227 dnode_t *dn;
3228 objset_t *os;
3229 uint64_t txg = tx->tx_txg;
3230
3231 ASSERT(dmu_tx_is_syncing(tx));
3232
3233 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr);
3234
3235 mutex_enter(&db->db_mtx);
3236 /*
3237 * To be synced, we must be dirtied. But we
3238 * might have been freed after the dirty.
3239 */
3240 if (db->db_state == DB_UNCACHED) {
3241 /* This buffer has been freed since it was dirtied */
3242 ASSERT(db->db.db_data == NULL);
3243 } else if (db->db_state == DB_FILL) {
3244 /* This buffer was freed and is now being re-filled */
3245 ASSERT(db->db.db_data != dr->dt.dl.dr_data);
3246 } else {
3247 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL);
3248 }
3249 DBUF_VERIFY(db);
3250
3251 DB_DNODE_ENTER(db);
3252 dn = DB_DNODE(db);
3253
3254 if (db->db_blkid == DMU_SPILL_BLKID) {
3255 mutex_enter(&dn->dn_mtx);
3256 if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) {
3257 /*
3258 * In the previous transaction group, the bonus buffer
3259 * was entirely used to store the attributes for the
3260 * dnode which overrode the dn_spill field. However,
3261 * when adding more attributes to the file a spill
3262 * block was required to hold the extra attributes.
3263 *
3264 * Make sure to clear the garbage left in the dn_spill
3265 * field from the previous attributes in the bonus
3266 * buffer. Otherwise, after writing out the spill
3267 * block to the new allocated dva, it will free
3268 * the old block pointed to by the invalid dn_spill.
3269 */
3270 db->db_blkptr = NULL;
3271 }
3272 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR;
3273 mutex_exit(&dn->dn_mtx);
3274 }
3275
3276 /*
3277 * If this is a bonus buffer, simply copy the bonus data into the
3278 * dnode. It will be written out when the dnode is synced (and it
3279 * will be synced, since it must have been dirty for dbuf_sync to
3280 * be called).
3281 */
3282 if (db->db_blkid == DMU_BONUS_BLKID) {
3283 dbuf_dirty_record_t **drp;
3284
3285 ASSERT(*datap != NULL);
3286 ASSERT0(db->db_level);
3287 ASSERT3U(dn->dn_phys->dn_bonuslen, <=,
3288 DN_SLOTS_TO_BONUSLEN(dn->dn_phys->dn_extra_slots + 1));
3289 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen);
3290 DB_DNODE_EXIT(db);
3291
3292 if (*datap != db->db.db_data) {
3293 int slots = DB_DNODE(db)->dn_num_slots;
3294 int bonuslen = DN_SLOTS_TO_BONUSLEN(slots);
3295 zio_buf_free(*datap, bonuslen);
3296 arc_space_return(bonuslen, ARC_SPACE_BONUS);
3297 }
3298 db->db_data_pending = NULL;
3299 drp = &db->db_last_dirty;
3300 while (*drp != dr)
3301 drp = &(*drp)->dr_next;
3302 ASSERT(dr->dr_next == NULL);
3303 ASSERT(dr->dr_dbuf == db);
3304 *drp = dr->dr_next;
3305 if (dr->dr_dbuf->db_level != 0) {
3306 mutex_destroy(&dr->dt.di.dr_mtx);
3307 list_destroy(&dr->dt.di.dr_children);
3308 }
3309 kmem_free(dr, sizeof (dbuf_dirty_record_t));
3310 ASSERT(db->db_dirtycnt > 0);
3311 db->db_dirtycnt -= 1;
3312 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg);
3313 return;
3314 }
3315
3316 os = dn->dn_objset;
3317
3318 /*
3319 * This function may have dropped the db_mtx lock allowing a dmu_sync
3320 * operation to sneak in. As a result, we need to ensure that we
3321 * don't check the dr_override_state until we have returned from
3322 * dbuf_check_blkptr.
3323 */
3324 dbuf_check_blkptr(dn, db);
3325
3326 /*
3327 * If this buffer is in the middle of an immediate write,
3328 * wait for the synchronous IO to complete.
3329 */
3330 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) {
3331 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT);
3332 cv_wait(&db->db_changed, &db->db_mtx);
3333 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN);
3334 }
3335
3336 if (db->db_state != DB_NOFILL &&
3337 dn->dn_object != DMU_META_DNODE_OBJECT &&
3338 refcount_count(&db->db_holds) > 1 &&
3339 dr->dt.dl.dr_override_state != DR_OVERRIDDEN &&
3340 *datap == db->db_buf) {
3341 /*
3342 * If this buffer is currently "in use" (i.e., there
3343 * are active holds and db_data still references it),
3344 * then make a copy before we start the write so that
3345 * any modifications from the open txg will not leak
3346 * into this write.
3347 *
3348 * NOTE: this copy does not need to be made for
3349 * objects only modified in the syncing context (e.g.
3350 * DNONE_DNODE blocks).
3351 */
3352 int psize = arc_buf_size(*datap);
3353 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db);
3354 enum zio_compress compress_type = arc_get_compression(*datap);
3355
3356 if (compress_type == ZIO_COMPRESS_OFF) {
3357 *datap = arc_alloc_buf(os->os_spa, db, type, psize);
3358 } else {
3359 int lsize = arc_buf_lsize(*datap);
3360 ASSERT3U(type, ==, ARC_BUFC_DATA);
3361 *datap = arc_alloc_compressed_buf(os->os_spa, db,
3362 psize, lsize, compress_type);
3363 }
3364 bcopy(db->db.db_data, (*datap)->b_data, psize);
3365 }
3366 db->db_data_pending = dr;
3367
3368 mutex_exit(&db->db_mtx);
3369
3370 dbuf_write(dr, *datap, tx);
3371
3372 ASSERT(!list_link_active(&dr->dr_dirty_node));
3373 if (dn->dn_object == DMU_META_DNODE_OBJECT) {
3374 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr);
3375 DB_DNODE_EXIT(db);
3376 } else {
3377 /*
3378 * Although zio_nowait() does not "wait for an IO", it does
3379 * initiate the IO. If this is an empty write it seems plausible
3380 * that the IO could actually be completed before the nowait
3381 * returns. We need to DB_DNODE_EXIT() first in case
3382 * zio_nowait() invalidates the dbuf.
3383 */
3384 DB_DNODE_EXIT(db);
3385 zio_nowait(dr->dr_zio);
3386 }
3387 }
3388
3389 void
3390 dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx)
3391 {
3392 dbuf_dirty_record_t *dr;
3393
3394 while ((dr = list_head(list))) {
3395 if (dr->dr_zio != NULL) {
3396 /*
3397 * If we find an already initialized zio then we
3398 * are processing the meta-dnode, and we have finished.
3399 * The dbufs for all dnodes are put back on the list
3400 * during processing, so that we can zio_wait()
3401 * these IOs after initiating all child IOs.
3402 */
3403 ASSERT3U(dr->dr_dbuf->db.db_object, ==,
3404 DMU_META_DNODE_OBJECT);
3405 break;
3406 }
3407 if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
3408 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
3409 VERIFY3U(dr->dr_dbuf->db_level, ==, level);
3410 }
3411 list_remove(list, dr);
3412 if (dr->dr_dbuf->db_level > 0)
3413 dbuf_sync_indirect(dr, tx);
3414 else
3415 dbuf_sync_leaf(dr, tx);
3416 }
3417 }
3418
3419 /* ARGSUSED */
3420 static void
3421 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3422 {
3423 dmu_buf_impl_t *db = vdb;
3424 dnode_t *dn;
3425 blkptr_t *bp = zio->io_bp;
3426 blkptr_t *bp_orig = &zio->io_bp_orig;
3427 spa_t *spa = zio->io_spa;
3428 int64_t delta;
3429 uint64_t fill = 0;
3430 int i;
3431
3432 ASSERT3P(db->db_blkptr, !=, NULL);
3433 ASSERT3P(&db->db_data_pending->dr_bp_copy, ==, bp);
3434
3435 DB_DNODE_ENTER(db);
3436 dn = DB_DNODE(db);
3437 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig);
3438 dnode_diduse_space(dn, delta - zio->io_prev_space_delta);
3439 zio->io_prev_space_delta = delta;
3440
3441 if (bp->blk_birth != 0) {
3442 ASSERT((db->db_blkid != DMU_SPILL_BLKID &&
3443 BP_GET_TYPE(bp) == dn->dn_type) ||
3444 (db->db_blkid == DMU_SPILL_BLKID &&
3445 BP_GET_TYPE(bp) == dn->dn_bonustype) ||
3446 BP_IS_EMBEDDED(bp));
3447 ASSERT(BP_GET_LEVEL(bp) == db->db_level);
3448 }
3449
3450 mutex_enter(&db->db_mtx);
3451
3452 #ifdef ZFS_DEBUG
3453 if (db->db_blkid == DMU_SPILL_BLKID) {
3454 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3455 ASSERT(!(BP_IS_HOLE(bp)) &&
3456 db->db_blkptr == DN_SPILL_BLKPTR(dn->dn_phys));
3457 }
3458 #endif
3459
3460 if (db->db_level == 0) {
3461 mutex_enter(&dn->dn_mtx);
3462 if (db->db_blkid > dn->dn_phys->dn_maxblkid &&
3463 db->db_blkid != DMU_SPILL_BLKID)
3464 dn->dn_phys->dn_maxblkid = db->db_blkid;
3465 mutex_exit(&dn->dn_mtx);
3466
3467 if (dn->dn_type == DMU_OT_DNODE) {
3468 i = 0;
3469 while (i < db->db.db_size) {
3470 dnode_phys_t *dnp = db->db.db_data + i;
3471
3472 i += DNODE_MIN_SIZE;
3473 if (dnp->dn_type != DMU_OT_NONE) {
3474 fill++;
3475 i += dnp->dn_extra_slots *
3476 DNODE_MIN_SIZE;
3477 }
3478 }
3479 } else {
3480 if (BP_IS_HOLE(bp)) {
3481 fill = 0;
3482 } else {
3483 fill = 1;
3484 }
3485 }
3486 } else {
3487 blkptr_t *ibp = db->db.db_data;
3488 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift);
3489 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) {
3490 if (BP_IS_HOLE(ibp))
3491 continue;
3492 fill += BP_GET_FILL(ibp);
3493 }
3494 }
3495 DB_DNODE_EXIT(db);
3496
3497 if (!BP_IS_EMBEDDED(bp))
3498 bp->blk_fill = fill;
3499
3500 mutex_exit(&db->db_mtx);
3501
3502 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
3503 *db->db_blkptr = *bp;
3504 rw_exit(&dn->dn_struct_rwlock);
3505 }
3506
3507 /* ARGSUSED */
3508 /*
3509 * This function gets called just prior to running through the compression
3510 * stage of the zio pipeline. If we're an indirect block comprised of only
3511 * holes, then we want this indirect to be compressed away to a hole. In
3512 * order to do that we must zero out any information about the holes that
3513 * this indirect points to prior to before we try to compress it.
3514 */
3515 static void
3516 dbuf_write_children_ready(zio_t *zio, arc_buf_t *buf, void *vdb)
3517 {
3518 dmu_buf_impl_t *db = vdb;
3519 dnode_t *dn;
3520 blkptr_t *bp;
3521 uint64_t i;
3522 int epbs;
3523
3524 ASSERT3U(db->db_level, >, 0);
3525 DB_DNODE_ENTER(db);
3526 dn = DB_DNODE(db);
3527 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
3528
3529 /* Determine if all our children are holes */
3530 for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++) {
3531 if (!BP_IS_HOLE(bp))
3532 break;
3533 }
3534
3535 /*
3536 * If all the children are holes, then zero them all out so that
3537 * we may get compressed away.
3538 */
3539 if (i == 1 << epbs) {
3540 /* didn't find any non-holes */
3541 bzero(db->db.db_data, db->db.db_size);
3542 }
3543 DB_DNODE_EXIT(db);
3544 }
3545
3546 /*
3547 * The SPA will call this callback several times for each zio - once
3548 * for every physical child i/o (zio->io_phys_children times). This
3549 * allows the DMU to monitor the progress of each logical i/o. For example,
3550 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z
3551 * block. There may be a long delay before all copies/fragments are completed,
3552 * so this callback allows us to retire dirty space gradually, as the physical
3553 * i/os complete.
3554 */
3555 /* ARGSUSED */
3556 static void
3557 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg)
3558 {
3559 dmu_buf_impl_t *db = arg;
3560 objset_t *os = db->db_objset;
3561 dsl_pool_t *dp = dmu_objset_pool(os);
3562 dbuf_dirty_record_t *dr;
3563 int delta = 0;
3564
3565 dr = db->db_data_pending;
3566 ASSERT3U(dr->dr_txg, ==, zio->io_txg);
3567
3568 /*
3569 * The callback will be called io_phys_children times. Retire one
3570 * portion of our dirty space each time we are called. Any rounding
3571 * error will be cleaned up by dsl_pool_sync()'s call to
3572 * dsl_pool_undirty_space().
3573 */
3574 delta = dr->dr_accounted / zio->io_phys_children;
3575 dsl_pool_undirty_space(dp, delta, zio->io_txg);
3576 }
3577
3578 /* ARGSUSED */
3579 static void
3580 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb)
3581 {
3582 dmu_buf_impl_t *db = vdb;
3583 blkptr_t *bp_orig = &zio->io_bp_orig;
3584 blkptr_t *bp = db->db_blkptr;
3585 objset_t *os = db->db_objset;
3586 dmu_tx_t *tx = os->os_synctx;
3587 dbuf_dirty_record_t **drp, *dr;
3588
3589 ASSERT0(zio->io_error);
3590 ASSERT(db->db_blkptr == bp);
3591
3592 /*
3593 * For nopwrites and rewrites we ensure that the bp matches our
3594 * original and bypass all the accounting.
3595 */
3596 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) {
3597 ASSERT(BP_EQUAL(bp, bp_orig));
3598 } else {
3599 dsl_dataset_t *ds = os->os_dsl_dataset;
3600 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE);
3601 dsl_dataset_block_born(ds, bp, tx);
3602 }
3603
3604 mutex_enter(&db->db_mtx);
3605
3606 DBUF_VERIFY(db);
3607
3608 drp = &db->db_last_dirty;
3609 while ((dr = *drp) != db->db_data_pending)
3610 drp = &dr->dr_next;
3611 ASSERT(!list_link_active(&dr->dr_dirty_node));
3612 ASSERT(dr->dr_dbuf == db);
3613 ASSERT(dr->dr_next == NULL);
3614 *drp = dr->dr_next;
3615
3616 #ifdef ZFS_DEBUG
3617 if (db->db_blkid == DMU_SPILL_BLKID) {
3618 dnode_t *dn;
3619
3620 DB_DNODE_ENTER(db);
3621 dn = DB_DNODE(db);
3622 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR);
3623 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) &&
3624 db->db_blkptr == DN_SPILL_BLKPTR(dn->dn_phys));
3625 DB_DNODE_EXIT(db);
3626 }
3627 #endif
3628
3629 if (db->db_level == 0) {
3630 ASSERT(db->db_blkid != DMU_BONUS_BLKID);
3631 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
3632 if (db->db_state != DB_NOFILL) {
3633 if (dr->dt.dl.dr_data != db->db_buf)
3634 arc_buf_destroy(dr->dt.dl.dr_data, db);
3635 }
3636 } else {
3637 dnode_t *dn;
3638
3639 DB_DNODE_ENTER(db);
3640 dn = DB_DNODE(db);
3641 ASSERT(list_head(&dr->dt.di.dr_children) == NULL);
3642 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift);
3643 if (!BP_IS_HOLE(db->db_blkptr)) {
3644 ASSERTV(int epbs = dn->dn_phys->dn_indblkshift -
3645 SPA_BLKPTRSHIFT);
3646 ASSERT3U(db->db_blkid, <=,
3647 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs));
3648 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==,
3649 db->db.db_size);
3650 }
3651 DB_DNODE_EXIT(db);
3652 mutex_destroy(&dr->dt.di.dr_mtx);
3653 list_destroy(&dr->dt.di.dr_children);
3654 }
3655 kmem_free(dr, sizeof (dbuf_dirty_record_t));
3656
3657 cv_broadcast(&db->db_changed);
3658 ASSERT(db->db_dirtycnt > 0);
3659 db->db_dirtycnt -= 1;
3660 db->db_data_pending = NULL;
3661 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg);
3662 }
3663
3664 static void
3665 dbuf_write_nofill_ready(zio_t *zio)
3666 {
3667 dbuf_write_ready(zio, NULL, zio->io_private);
3668 }
3669
3670 static void
3671 dbuf_write_nofill_done(zio_t *zio)
3672 {
3673 dbuf_write_done(zio, NULL, zio->io_private);
3674 }
3675
3676 static void
3677 dbuf_write_override_ready(zio_t *zio)
3678 {
3679 dbuf_dirty_record_t *dr = zio->io_private;
3680 dmu_buf_impl_t *db = dr->dr_dbuf;
3681
3682 dbuf_write_ready(zio, NULL, db);
3683 }
3684
3685 static void
3686 dbuf_write_override_done(zio_t *zio)
3687 {
3688 dbuf_dirty_record_t *dr = zio->io_private;
3689 dmu_buf_impl_t *db = dr->dr_dbuf;
3690 blkptr_t *obp = &dr->dt.dl.dr_overridden_by;
3691
3692 mutex_enter(&db->db_mtx);
3693 if (!BP_EQUAL(zio->io_bp, obp)) {
3694 if (!BP_IS_HOLE(obp))
3695 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp);
3696 arc_release(dr->dt.dl.dr_data, db);
3697 }
3698 mutex_exit(&db->db_mtx);
3699
3700 dbuf_write_done(zio, NULL, db);
3701 }
3702
3703 /* Issue I/O to commit a dirty buffer to disk. */
3704 static void
3705 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
3706 {
3707 dmu_buf_impl_t *db = dr->dr_dbuf;
3708 dnode_t *dn;
3709 objset_t *os;
3710 dmu_buf_impl_t *parent = db->db_parent;
3711 uint64_t txg = tx->tx_txg;
3712 zbookmark_phys_t zb;
3713 zio_prop_t zp;
3714 zio_t *zio;
3715 int wp_flag = 0;
3716
3717 ASSERT(dmu_tx_is_syncing(tx));
3718
3719 DB_DNODE_ENTER(db);
3720 dn = DB_DNODE(db);
3721 os = dn->dn_objset;
3722
3723 if (db->db_state != DB_NOFILL) {
3724 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) {
3725 /*
3726 * Private object buffers are released here rather
3727 * than in dbuf_dirty() since they are only modified
3728 * in the syncing context and we don't want the
3729 * overhead of making multiple copies of the data.
3730 */
3731 if (BP_IS_HOLE(db->db_blkptr)) {
3732 arc_buf_thaw(data);
3733 } else {
3734 dbuf_release_bp(db);
3735 }
3736 }
3737 }
3738
3739 if (parent != dn->dn_dbuf) {
3740 /* Our parent is an indirect block. */
3741 /* We have a dirty parent that has been scheduled for write. */
3742 ASSERT(parent && parent->db_data_pending);
3743 /* Our parent's buffer is one level closer to the dnode. */
3744 ASSERT(db->db_level == parent->db_level-1);
3745 /*
3746 * We're about to modify our parent's db_data by modifying
3747 * our block pointer, so the parent must be released.
3748 */
3749 ASSERT(arc_released(parent->db_buf));
3750 zio = parent->db_data_pending->dr_zio;
3751 } else {
3752 /* Our parent is the dnode itself. */
3753 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 &&
3754 db->db_blkid != DMU_SPILL_BLKID) ||
3755 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0));
3756 if (db->db_blkid != DMU_SPILL_BLKID)
3757 ASSERT3P(db->db_blkptr, ==,
3758 &dn->dn_phys->dn_blkptr[db->db_blkid]);
3759 zio = dn->dn_zio;
3760 }
3761
3762 ASSERT(db->db_level == 0 || data == db->db_buf);
3763 ASSERT3U(db->db_blkptr->blk_birth, <=, txg);
3764 ASSERT(zio);
3765
3766 SET_BOOKMARK(&zb, os->os_dsl_dataset ?
3767 os->os_dsl_dataset->ds_object : DMU_META_OBJSET,
3768 db->db.db_object, db->db_level, db->db_blkid);
3769
3770 if (db->db_blkid == DMU_SPILL_BLKID)
3771 wp_flag = WP_SPILL;
3772 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0;
3773
3774 dmu_write_policy(os, dn, db->db_level, wp_flag,
3775 (data != NULL && arc_get_compression(data) != ZIO_COMPRESS_OFF) ?
3776 arc_get_compression(data) : ZIO_COMPRESS_INHERIT, &zp);
3777 DB_DNODE_EXIT(db);
3778
3779 /*
3780 * We copy the blkptr now (rather than when we instantiate the dirty
3781 * record), because its value can change between open context and
3782 * syncing context. We do not need to hold dn_struct_rwlock to read
3783 * db_blkptr because we are in syncing context.
3784 */
3785 dr->dr_bp_copy = *db->db_blkptr;
3786
3787 if (db->db_level == 0 &&
3788 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
3789 /*
3790 * The BP for this block has been provided by open context
3791 * (by dmu_sync() or dmu_buf_write_embedded()).
3792 */
3793 void *contents = (data != NULL) ? data->b_data : NULL;
3794
3795 dr->dr_zio = zio_write(zio, os->os_spa, txg,
3796 &dr->dr_bp_copy, contents, db->db.db_size, db->db.db_size,
3797 &zp, dbuf_write_override_ready, NULL, NULL,
3798 dbuf_write_override_done,
3799 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
3800 mutex_enter(&db->db_mtx);
3801 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
3802 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by,
3803 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite);
3804 mutex_exit(&db->db_mtx);
3805 } else if (db->db_state == DB_NOFILL) {
3806 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF);
3807 dr->dr_zio = zio_write(zio, os->os_spa, txg,
3808 &dr->dr_bp_copy, NULL, db->db.db_size, db->db.db_size, &zp,
3809 dbuf_write_nofill_ready, NULL, NULL,
3810 dbuf_write_nofill_done, db,
3811 ZIO_PRIORITY_ASYNC_WRITE,
3812 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb);
3813 } else {
3814 arc_done_func_t *children_ready_cb = NULL;
3815 ASSERT(arc_released(data));
3816
3817 /*
3818 * For indirect blocks, we want to setup the children
3819 * ready callback so that we can properly handle an indirect
3820 * block that only contains holes.
3821 */
3822 if (db->db_level != 0)
3823 children_ready_cb = dbuf_write_children_ready;
3824
3825 dr->dr_zio = arc_write(zio, os->os_spa, txg,
3826 &dr->dr_bp_copy, data, DBUF_IS_L2CACHEABLE(db),
3827 &zp, dbuf_write_ready,
3828 children_ready_cb, dbuf_write_physdone,
3829 dbuf_write_done, db, ZIO_PRIORITY_ASYNC_WRITE,
3830 ZIO_FLAG_MUSTSUCCEED, &zb);
3831 }
3832 }
3833
3834 #if defined(_KERNEL) && defined(HAVE_SPL)
3835 EXPORT_SYMBOL(dbuf_find);
3836 EXPORT_SYMBOL(dbuf_is_metadata);
3837 EXPORT_SYMBOL(dbuf_destroy);
3838 EXPORT_SYMBOL(dbuf_loan_arcbuf);
3839 EXPORT_SYMBOL(dbuf_whichblock);
3840 EXPORT_SYMBOL(dbuf_read);
3841 EXPORT_SYMBOL(dbuf_unoverride);
3842 EXPORT_SYMBOL(dbuf_free_range);
3843 EXPORT_SYMBOL(dbuf_new_size);
3844 EXPORT_SYMBOL(dbuf_release_bp);
3845 EXPORT_SYMBOL(dbuf_dirty);
3846 EXPORT_SYMBOL(dmu_buf_will_dirty);
3847 EXPORT_SYMBOL(dmu_buf_will_not_fill);
3848 EXPORT_SYMBOL(dmu_buf_will_fill);
3849 EXPORT_SYMBOL(dmu_buf_fill_done);
3850 EXPORT_SYMBOL(dmu_buf_rele);
3851 EXPORT_SYMBOL(dbuf_assign_arcbuf);
3852 EXPORT_SYMBOL(dbuf_prefetch);
3853 EXPORT_SYMBOL(dbuf_hold_impl);
3854 EXPORT_SYMBOL(dbuf_hold);
3855 EXPORT_SYMBOL(dbuf_hold_level);
3856 EXPORT_SYMBOL(dbuf_create_bonus);
3857 EXPORT_SYMBOL(dbuf_spill_set_blksz);
3858 EXPORT_SYMBOL(dbuf_rm_spill);
3859 EXPORT_SYMBOL(dbuf_add_ref);
3860 EXPORT_SYMBOL(dbuf_rele);
3861 EXPORT_SYMBOL(dbuf_rele_and_unlock);
3862 EXPORT_SYMBOL(dbuf_refcount);
3863 EXPORT_SYMBOL(dbuf_sync_list);
3864 EXPORT_SYMBOL(dmu_buf_set_user);
3865 EXPORT_SYMBOL(dmu_buf_set_user_ie);
3866 EXPORT_SYMBOL(dmu_buf_get_user);
3867 EXPORT_SYMBOL(dmu_buf_freeable);
3868 EXPORT_SYMBOL(dmu_buf_get_blkptr);
3869
3870
3871 module_param(dbuf_cache_max_bytes, ulong, 0644);
3872 MODULE_PARM_DESC(dbuf_cache_max_bytes,
3873 "Maximum size in bytes of the dbuf cache.");
3874
3875 module_param(dbuf_cache_hiwater_pct, uint, 0644);
3876 MODULE_PARM_DESC(dbuf_cache_hiwater_pct,
3877 "Percentage over dbuf_cache_max_bytes when dbufs \
3878 much be evicted directly.");
3879
3880 module_param(dbuf_cache_lowater_pct, uint, 0644);
3881 MODULE_PARM_DESC(dbuf_cache_lowater_pct,
3882 "Percentage below dbuf_cache_max_bytes \
3883 when the evict thread stop evicting dbufs.");
3884
3885 module_param(dbuf_cache_max_shift, int, 0644);
3886 MODULE_PARM_DESC(dbuf_cache_max_shift,
3887 "Cap the size of the dbuf cache to log2 fraction of arc size.");
3888
3889 #endif