]>
Commit | Line | Data |
---|---|---|
34dc7c2f BB |
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 | /* | |
428870ff | 22 | * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. |
ef3c1dea | 23 | * Copyright 2011 Nexenta Systems, Inc. All rights reserved. |
64fc7762 | 24 | * Copyright (c) 2012, 2017 by Delphix. All rights reserved. |
3a17a7a9 | 25 | * Copyright (c) 2013 by Saso Kiselkov. All rights reserved. |
0c66c32d | 26 | * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved. |
34dc7c2f BB |
27 | */ |
28 | ||
34dc7c2f | 29 | #include <sys/zfs_context.h> |
c28b2279 | 30 | #include <sys/arc.h> |
34dc7c2f | 31 | #include <sys/dmu.h> |
ea97f8ce | 32 | #include <sys/dmu_send.h> |
34dc7c2f BB |
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> | |
428870ff BB |
42 | #include <sys/sa.h> |
43 | #include <sys/sa_impl.h> | |
9b67f605 MA |
44 | #include <sys/zfeature.h> |
45 | #include <sys/blkptr.h> | |
9bd274dd | 46 | #include <sys/range_tree.h> |
49ee64e5 | 47 | #include <sys/trace_dbuf.h> |
d3c2ae1c | 48 | #include <sys/callb.h> |
a6255b7f | 49 | #include <sys/abd.h> |
34dc7c2f | 50 | |
fc5bb51f BB |
51 | struct dbuf_hold_impl_data { |
52 | /* Function arguments */ | |
53 | dnode_t *dh_dn; | |
54 | uint8_t dh_level; | |
55 | uint64_t dh_blkid; | |
fcff0f35 PD |
56 | boolean_t dh_fail_sparse; |
57 | boolean_t dh_fail_uncached; | |
fc5bb51f BB |
58 | void *dh_tag; |
59 | dmu_buf_impl_t **dh_dbp; | |
60 | /* Local variables */ | |
61 | dmu_buf_impl_t *dh_db; | |
62 | dmu_buf_impl_t *dh_parent; | |
63 | blkptr_t *dh_bp; | |
64 | int dh_err; | |
65 | dbuf_dirty_record_t *dh_dr; | |
fc5bb51f BB |
66 | int dh_depth; |
67 | }; | |
68 | ||
69 | static void __dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh, | |
fcff0f35 PD |
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); | |
fc5bb51f BB |
73 | static int __dbuf_hold_impl(struct dbuf_hold_impl_data *dh); |
74 | ||
d3c2ae1c | 75 | uint_t zfs_dbuf_evict_key; |
b663a23d | 76 | |
13fe0198 | 77 | static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx); |
b128c09f | 78 | static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx); |
34dc7c2f | 79 | |
0c66c32d | 80 | extern inline void dmu_buf_init_user(dmu_buf_user_t *dbu, |
39efbde7 GM |
81 | dmu_buf_evict_func_t *evict_func_sync, |
82 | dmu_buf_evict_func_t *evict_func_async, | |
83 | dmu_buf_t **clear_on_evict_dbufp); | |
0c66c32d | 84 | |
34dc7c2f BB |
85 | /* |
86 | * Global data structures and functions for the dbuf cache. | |
87 | */ | |
d3c2ae1c | 88 | static kmem_cache_t *dbuf_kmem_cache; |
0c66c32d | 89 | static taskq_t *dbu_evict_taskq; |
34dc7c2f | 90 | |
d3c2ae1c GW |
91 | static kthread_t *dbuf_cache_evict_thread; |
92 | static kmutex_t dbuf_evict_lock; | |
93 | static kcondvar_t dbuf_evict_cv; | |
94 | static boolean_t dbuf_evict_thread_exit; | |
95 | ||
96 | /* | |
97 | * LRU cache of dbufs. The dbuf cache maintains a list of dbufs that | |
98 | * are not currently held but have been recently released. These dbufs | |
99 | * are not eligible for arc eviction until they are aged out of the cache. | |
100 | * Dbufs are added to the dbuf cache once the last hold is released. If a | |
101 | * dbuf is later accessed and still exists in the dbuf cache, then it will | |
102 | * be removed from the cache and later re-added to the head of the cache. | |
103 | * Dbufs that are aged out of the cache will be immediately destroyed and | |
104 | * become eligible for arc eviction. | |
105 | */ | |
64fc7762 | 106 | static multilist_t *dbuf_cache; |
d3c2ae1c GW |
107 | static refcount_t dbuf_cache_size; |
108 | unsigned long dbuf_cache_max_bytes = 100 * 1024 * 1024; | |
109 | ||
110 | /* Cap the size of the dbuf cache to log2 fraction of arc size. */ | |
111 | int dbuf_cache_max_shift = 5; | |
112 | ||
113 | /* | |
114 | * The dbuf cache uses a three-stage eviction policy: | |
115 | * - A low water marker designates when the dbuf eviction thread | |
116 | * should stop evicting from the dbuf cache. | |
117 | * - When we reach the maximum size (aka mid water mark), we | |
118 | * signal the eviction thread to run. | |
119 | * - The high water mark indicates when the eviction thread | |
120 | * is unable to keep up with the incoming load and eviction must | |
121 | * happen in the context of the calling thread. | |
122 | * | |
123 | * The dbuf cache: | |
124 | * (max size) | |
125 | * low water mid water hi water | |
126 | * +----------------------------------------+----------+----------+ | |
127 | * | | | | | |
128 | * | | | | | |
129 | * | | | | | |
130 | * | | | | | |
131 | * +----------------------------------------+----------+----------+ | |
132 | * stop signal evict | |
133 | * evicting eviction directly | |
134 | * thread | |
135 | * | |
136 | * The high and low water marks indicate the operating range for the eviction | |
137 | * thread. The low water mark is, by default, 90% of the total size of the | |
138 | * cache and the high water mark is at 110% (both of these percentages can be | |
139 | * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct, | |
140 | * respectively). The eviction thread will try to ensure that the cache remains | |
141 | * within this range by waking up every second and checking if the cache is | |
142 | * above the low water mark. The thread can also be woken up by callers adding | |
143 | * elements into the cache if the cache is larger than the mid water (i.e max | |
144 | * cache size). Once the eviction thread is woken up and eviction is required, | |
145 | * it will continue evicting buffers until it's able to reduce the cache size | |
146 | * to the low water mark. If the cache size continues to grow and hits the high | |
4e33ba4c | 147 | * water mark, then callers adding elements to the cache will begin to evict |
d3c2ae1c GW |
148 | * directly from the cache until the cache is no longer above the high water |
149 | * mark. | |
150 | */ | |
151 | ||
152 | /* | |
153 | * The percentage above and below the maximum cache size. | |
154 | */ | |
155 | uint_t dbuf_cache_hiwater_pct = 10; | |
156 | uint_t dbuf_cache_lowater_pct = 10; | |
157 | ||
34dc7c2f BB |
158 | /* ARGSUSED */ |
159 | static int | |
160 | dbuf_cons(void *vdb, void *unused, int kmflag) | |
161 | { | |
162 | dmu_buf_impl_t *db = vdb; | |
163 | bzero(db, sizeof (dmu_buf_impl_t)); | |
164 | ||
165 | mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL); | |
166 | cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL); | |
d3c2ae1c | 167 | multilist_link_init(&db->db_cache_link); |
34dc7c2f | 168 | refcount_create(&db->db_holds); |
8951cb8d | 169 | |
34dc7c2f BB |
170 | return (0); |
171 | } | |
172 | ||
173 | /* ARGSUSED */ | |
174 | static void | |
175 | dbuf_dest(void *vdb, void *unused) | |
176 | { | |
177 | dmu_buf_impl_t *db = vdb; | |
178 | mutex_destroy(&db->db_mtx); | |
179 | cv_destroy(&db->db_changed); | |
d3c2ae1c | 180 | ASSERT(!multilist_link_active(&db->db_cache_link)); |
34dc7c2f BB |
181 | refcount_destroy(&db->db_holds); |
182 | } | |
183 | ||
184 | /* | |
185 | * dbuf hash table routines | |
186 | */ | |
187 | static dbuf_hash_table_t dbuf_hash_table; | |
188 | ||
189 | static uint64_t dbuf_hash_count; | |
190 | ||
191 | static uint64_t | |
192 | dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid) | |
193 | { | |
194 | uintptr_t osv = (uintptr_t)os; | |
195 | uint64_t crc = -1ULL; | |
196 | ||
197 | ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); | |
198 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF]; | |
199 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF]; | |
200 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF]; | |
201 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF]; | |
202 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF]; | |
203 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF]; | |
204 | ||
205 | crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16); | |
206 | ||
207 | return (crc); | |
208 | } | |
209 | ||
34dc7c2f BB |
210 | #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \ |
211 | ((dbuf)->db.db_object == (obj) && \ | |
212 | (dbuf)->db_objset == (os) && \ | |
213 | (dbuf)->db_level == (level) && \ | |
214 | (dbuf)->db_blkid == (blkid)) | |
215 | ||
216 | dmu_buf_impl_t * | |
6ebebace | 217 | dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid) |
34dc7c2f BB |
218 | { |
219 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
d6320ddb BB |
220 | uint64_t hv; |
221 | uint64_t idx; | |
34dc7c2f BB |
222 | dmu_buf_impl_t *db; |
223 | ||
d3c2ae1c | 224 | hv = dbuf_hash(os, obj, level, blkid); |
d6320ddb BB |
225 | idx = hv & h->hash_table_mask; |
226 | ||
34dc7c2f BB |
227 | mutex_enter(DBUF_HASH_MUTEX(h, idx)); |
228 | for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) { | |
229 | if (DBUF_EQUAL(db, os, obj, level, blkid)) { | |
230 | mutex_enter(&db->db_mtx); | |
231 | if (db->db_state != DB_EVICTING) { | |
232 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); | |
233 | return (db); | |
234 | } | |
235 | mutex_exit(&db->db_mtx); | |
236 | } | |
237 | } | |
238 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); | |
239 | return (NULL); | |
240 | } | |
241 | ||
6ebebace JG |
242 | static dmu_buf_impl_t * |
243 | dbuf_find_bonus(objset_t *os, uint64_t object) | |
244 | { | |
245 | dnode_t *dn; | |
246 | dmu_buf_impl_t *db = NULL; | |
247 | ||
248 | if (dnode_hold(os, object, FTAG, &dn) == 0) { | |
249 | rw_enter(&dn->dn_struct_rwlock, RW_READER); | |
250 | if (dn->dn_bonus != NULL) { | |
251 | db = dn->dn_bonus; | |
252 | mutex_enter(&db->db_mtx); | |
253 | } | |
254 | rw_exit(&dn->dn_struct_rwlock); | |
255 | dnode_rele(dn, FTAG); | |
256 | } | |
257 | return (db); | |
258 | } | |
259 | ||
34dc7c2f BB |
260 | /* |
261 | * Insert an entry into the hash table. If there is already an element | |
262 | * equal to elem in the hash table, then the already existing element | |
263 | * will be returned and the new element will not be inserted. | |
264 | * Otherwise returns NULL. | |
265 | */ | |
266 | static dmu_buf_impl_t * | |
267 | dbuf_hash_insert(dmu_buf_impl_t *db) | |
268 | { | |
269 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
428870ff | 270 | objset_t *os = db->db_objset; |
34dc7c2f BB |
271 | uint64_t obj = db->db.db_object; |
272 | int level = db->db_level; | |
d6320ddb | 273 | uint64_t blkid, hv, idx; |
34dc7c2f BB |
274 | dmu_buf_impl_t *dbf; |
275 | ||
d6320ddb | 276 | blkid = db->db_blkid; |
d3c2ae1c | 277 | hv = dbuf_hash(os, obj, level, blkid); |
d6320ddb BB |
278 | idx = hv & h->hash_table_mask; |
279 | ||
34dc7c2f BB |
280 | mutex_enter(DBUF_HASH_MUTEX(h, idx)); |
281 | for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) { | |
282 | if (DBUF_EQUAL(dbf, os, obj, level, blkid)) { | |
283 | mutex_enter(&dbf->db_mtx); | |
284 | if (dbf->db_state != DB_EVICTING) { | |
285 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); | |
286 | return (dbf); | |
287 | } | |
288 | mutex_exit(&dbf->db_mtx); | |
289 | } | |
290 | } | |
291 | ||
292 | mutex_enter(&db->db_mtx); | |
293 | db->db_hash_next = h->hash_table[idx]; | |
294 | h->hash_table[idx] = db; | |
295 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); | |
bc89ac84 | 296 | atomic_inc_64(&dbuf_hash_count); |
34dc7c2f BB |
297 | |
298 | return (NULL); | |
299 | } | |
300 | ||
301 | /* | |
bd089c54 | 302 | * Remove an entry from the hash table. It must be in the EVICTING state. |
34dc7c2f BB |
303 | */ |
304 | static void | |
305 | dbuf_hash_remove(dmu_buf_impl_t *db) | |
306 | { | |
307 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
d6320ddb | 308 | uint64_t hv, idx; |
34dc7c2f BB |
309 | dmu_buf_impl_t *dbf, **dbp; |
310 | ||
d3c2ae1c | 311 | hv = dbuf_hash(db->db_objset, db->db.db_object, |
d6320ddb BB |
312 | db->db_level, db->db_blkid); |
313 | idx = hv & h->hash_table_mask; | |
314 | ||
34dc7c2f | 315 | /* |
4e33ba4c | 316 | * We mustn't hold db_mtx to maintain lock ordering: |
34dc7c2f BB |
317 | * DBUF_HASH_MUTEX > db_mtx. |
318 | */ | |
319 | ASSERT(refcount_is_zero(&db->db_holds)); | |
320 | ASSERT(db->db_state == DB_EVICTING); | |
321 | ASSERT(!MUTEX_HELD(&db->db_mtx)); | |
322 | ||
323 | mutex_enter(DBUF_HASH_MUTEX(h, idx)); | |
324 | dbp = &h->hash_table[idx]; | |
325 | while ((dbf = *dbp) != db) { | |
326 | dbp = &dbf->db_hash_next; | |
327 | ASSERT(dbf != NULL); | |
328 | } | |
329 | *dbp = db->db_hash_next; | |
330 | db->db_hash_next = NULL; | |
331 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); | |
bc89ac84 | 332 | atomic_dec_64(&dbuf_hash_count); |
34dc7c2f BB |
333 | } |
334 | ||
0c66c32d JG |
335 | typedef enum { |
336 | DBVU_EVICTING, | |
337 | DBVU_NOT_EVICTING | |
338 | } dbvu_verify_type_t; | |
339 | ||
340 | static void | |
341 | dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type) | |
342 | { | |
343 | #ifdef ZFS_DEBUG | |
344 | int64_t holds; | |
345 | ||
346 | if (db->db_user == NULL) | |
347 | return; | |
348 | ||
349 | /* Only data blocks support the attachment of user data. */ | |
350 | ASSERT(db->db_level == 0); | |
351 | ||
352 | /* Clients must resolve a dbuf before attaching user data. */ | |
353 | ASSERT(db->db.db_data != NULL); | |
354 | ASSERT3U(db->db_state, ==, DB_CACHED); | |
355 | ||
356 | holds = refcount_count(&db->db_holds); | |
357 | if (verify_type == DBVU_EVICTING) { | |
358 | /* | |
359 | * Immediate eviction occurs when holds == dirtycnt. | |
360 | * For normal eviction buffers, holds is zero on | |
361 | * eviction, except when dbuf_fix_old_data() calls | |
362 | * dbuf_clear_data(). However, the hold count can grow | |
363 | * during eviction even though db_mtx is held (see | |
364 | * dmu_bonus_hold() for an example), so we can only | |
365 | * test the generic invariant that holds >= dirtycnt. | |
366 | */ | |
367 | ASSERT3U(holds, >=, db->db_dirtycnt); | |
368 | } else { | |
bc4501f7 | 369 | if (db->db_user_immediate_evict == TRUE) |
0c66c32d JG |
370 | ASSERT3U(holds, >=, db->db_dirtycnt); |
371 | else | |
372 | ASSERT3U(holds, >, 0); | |
373 | } | |
374 | #endif | |
375 | } | |
376 | ||
34dc7c2f BB |
377 | static void |
378 | dbuf_evict_user(dmu_buf_impl_t *db) | |
379 | { | |
0c66c32d JG |
380 | dmu_buf_user_t *dbu = db->db_user; |
381 | ||
34dc7c2f BB |
382 | ASSERT(MUTEX_HELD(&db->db_mtx)); |
383 | ||
0c66c32d | 384 | if (dbu == NULL) |
34dc7c2f BB |
385 | return; |
386 | ||
0c66c32d JG |
387 | dbuf_verify_user(db, DBVU_EVICTING); |
388 | db->db_user = NULL; | |
389 | ||
390 | #ifdef ZFS_DEBUG | |
391 | if (dbu->dbu_clear_on_evict_dbufp != NULL) | |
392 | *dbu->dbu_clear_on_evict_dbufp = NULL; | |
393 | #endif | |
394 | ||
395 | /* | |
39efbde7 GM |
396 | * There are two eviction callbacks - one that we call synchronously |
397 | * and one that we invoke via a taskq. The async one is useful for | |
398 | * avoiding lock order reversals and limiting stack depth. | |
399 | * | |
400 | * Note that if we have a sync callback but no async callback, | |
401 | * it's likely that the sync callback will free the structure | |
402 | * containing the dbu. In that case we need to take care to not | |
403 | * dereference dbu after calling the sync evict func. | |
0c66c32d | 404 | */ |
a7004725 | 405 | boolean_t has_async = (dbu->dbu_evict_func_async != NULL); |
39efbde7 GM |
406 | |
407 | if (dbu->dbu_evict_func_sync != NULL) | |
408 | dbu->dbu_evict_func_sync(dbu); | |
409 | ||
410 | if (has_async) { | |
411 | taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func_async, | |
412 | dbu, 0, &dbu->dbu_tqent); | |
413 | } | |
34dc7c2f BB |
414 | } |
415 | ||
572e2857 BB |
416 | boolean_t |
417 | dbuf_is_metadata(dmu_buf_impl_t *db) | |
418 | { | |
cc79a5c2 BB |
419 | /* |
420 | * Consider indirect blocks and spill blocks to be meta data. | |
421 | */ | |
422 | if (db->db_level > 0 || db->db_blkid == DMU_SPILL_BLKID) { | |
572e2857 BB |
423 | return (B_TRUE); |
424 | } else { | |
425 | boolean_t is_metadata; | |
426 | ||
427 | DB_DNODE_ENTER(db); | |
9ae529ec | 428 | is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type); |
572e2857 BB |
429 | DB_DNODE_EXIT(db); |
430 | ||
431 | return (is_metadata); | |
432 | } | |
433 | } | |
434 | ||
d3c2ae1c GW |
435 | |
436 | /* | |
437 | * This function *must* return indices evenly distributed between all | |
438 | * sublists of the multilist. This is needed due to how the dbuf eviction | |
439 | * code is laid out; dbuf_evict_thread() assumes dbufs are evenly | |
440 | * distributed between all sublists and uses this assumption when | |
441 | * deciding which sublist to evict from and how much to evict from it. | |
442 | */ | |
443 | unsigned int | |
444 | dbuf_cache_multilist_index_func(multilist_t *ml, void *obj) | |
34dc7c2f | 445 | { |
d3c2ae1c GW |
446 | dmu_buf_impl_t *db = obj; |
447 | ||
448 | /* | |
449 | * The assumption here, is the hash value for a given | |
450 | * dmu_buf_impl_t will remain constant throughout it's lifetime | |
451 | * (i.e. it's objset, object, level and blkid fields don't change). | |
452 | * Thus, we don't need to store the dbuf's sublist index | |
453 | * on insertion, as this index can be recalculated on removal. | |
454 | * | |
455 | * Also, the low order bits of the hash value are thought to be | |
456 | * distributed evenly. Otherwise, in the case that the multilist | |
457 | * has a power of two number of sublists, each sublists' usage | |
458 | * would not be evenly distributed. | |
459 | */ | |
460 | return (dbuf_hash(db->db_objset, db->db.db_object, | |
461 | db->db_level, db->db_blkid) % | |
462 | multilist_get_num_sublists(ml)); | |
463 | } | |
464 | ||
e71cade6 | 465 | static inline unsigned long |
466 | dbuf_cache_target_bytes(void) | |
467 | { | |
468 | return MIN(dbuf_cache_max_bytes, | |
469 | arc_target_bytes() >> dbuf_cache_max_shift); | |
470 | } | |
471 | ||
d3c2ae1c GW |
472 | static inline boolean_t |
473 | dbuf_cache_above_hiwater(void) | |
474 | { | |
e71cade6 | 475 | uint64_t dbuf_cache_target = dbuf_cache_target_bytes(); |
476 | ||
d3c2ae1c | 477 | uint64_t dbuf_cache_hiwater_bytes = |
e71cade6 | 478 | (dbuf_cache_target * dbuf_cache_hiwater_pct) / 100; |
d3c2ae1c GW |
479 | |
480 | return (refcount_count(&dbuf_cache_size) > | |
e71cade6 | 481 | dbuf_cache_target + dbuf_cache_hiwater_bytes); |
d3c2ae1c GW |
482 | } |
483 | ||
484 | static inline boolean_t | |
485 | dbuf_cache_above_lowater(void) | |
486 | { | |
e71cade6 | 487 | uint64_t dbuf_cache_target = dbuf_cache_target_bytes(); |
488 | ||
d3c2ae1c | 489 | uint64_t dbuf_cache_lowater_bytes = |
e71cade6 | 490 | (dbuf_cache_target * dbuf_cache_lowater_pct) / 100; |
d3c2ae1c GW |
491 | |
492 | return (refcount_count(&dbuf_cache_size) > | |
e71cade6 | 493 | dbuf_cache_target - dbuf_cache_lowater_bytes); |
d3c2ae1c GW |
494 | } |
495 | ||
496 | /* | |
497 | * Evict the oldest eligible dbuf from the dbuf cache. | |
498 | */ | |
499 | static void | |
500 | dbuf_evict_one(void) | |
501 | { | |
64fc7762 MA |
502 | int idx = multilist_get_random_index(dbuf_cache); |
503 | multilist_sublist_t *mls = multilist_sublist_lock(dbuf_cache, idx); | |
1c27024e | 504 | |
d3c2ae1c GW |
505 | ASSERT(!MUTEX_HELD(&dbuf_evict_lock)); |
506 | ||
507 | /* | |
508 | * Set the thread's tsd to indicate that it's processing evictions. | |
509 | * Once a thread stops evicting from the dbuf cache it will | |
510 | * reset its tsd to NULL. | |
511 | */ | |
512 | ASSERT3P(tsd_get(zfs_dbuf_evict_key), ==, NULL); | |
513 | (void) tsd_set(zfs_dbuf_evict_key, (void *)B_TRUE); | |
514 | ||
1c27024e | 515 | dmu_buf_impl_t *db = multilist_sublist_tail(mls); |
d3c2ae1c GW |
516 | while (db != NULL && mutex_tryenter(&db->db_mtx) == 0) { |
517 | db = multilist_sublist_prev(mls, db); | |
518 | } | |
519 | ||
520 | DTRACE_PROBE2(dbuf__evict__one, dmu_buf_impl_t *, db, | |
521 | multilist_sublist_t *, mls); | |
522 | ||
523 | if (db != NULL) { | |
524 | multilist_sublist_remove(mls, db); | |
525 | multilist_sublist_unlock(mls); | |
526 | (void) refcount_remove_many(&dbuf_cache_size, | |
527 | db->db.db_size, db); | |
528 | dbuf_destroy(db); | |
529 | } else { | |
530 | multilist_sublist_unlock(mls); | |
531 | } | |
532 | (void) tsd_set(zfs_dbuf_evict_key, NULL); | |
533 | } | |
534 | ||
535 | /* | |
536 | * The dbuf evict thread is responsible for aging out dbufs from the | |
537 | * cache. Once the cache has reached it's maximum size, dbufs are removed | |
538 | * and destroyed. The eviction thread will continue running until the size | |
539 | * of the dbuf cache is at or below the maximum size. Once the dbuf is aged | |
540 | * out of the cache it is destroyed and becomes eligible for arc eviction. | |
541 | */ | |
867959b5 | 542 | /* ARGSUSED */ |
d3c2ae1c | 543 | static void |
c25b8f99 | 544 | dbuf_evict_thread(void *unused) |
d3c2ae1c GW |
545 | { |
546 | callb_cpr_t cpr; | |
547 | ||
548 | CALLB_CPR_INIT(&cpr, &dbuf_evict_lock, callb_generic_cpr, FTAG); | |
549 | ||
550 | mutex_enter(&dbuf_evict_lock); | |
551 | while (!dbuf_evict_thread_exit) { | |
552 | while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) { | |
553 | CALLB_CPR_SAFE_BEGIN(&cpr); | |
554 | (void) cv_timedwait_sig_hires(&dbuf_evict_cv, | |
555 | &dbuf_evict_lock, SEC2NSEC(1), MSEC2NSEC(1), 0); | |
556 | CALLB_CPR_SAFE_END(&cpr, &dbuf_evict_lock); | |
557 | } | |
558 | mutex_exit(&dbuf_evict_lock); | |
559 | ||
560 | /* | |
561 | * Keep evicting as long as we're above the low water mark | |
562 | * for the cache. We do this without holding the locks to | |
563 | * minimize lock contention. | |
564 | */ | |
565 | while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) { | |
566 | dbuf_evict_one(); | |
567 | } | |
568 | ||
569 | mutex_enter(&dbuf_evict_lock); | |
570 | } | |
571 | ||
572 | dbuf_evict_thread_exit = B_FALSE; | |
573 | cv_broadcast(&dbuf_evict_cv); | |
574 | CALLB_CPR_EXIT(&cpr); /* drops dbuf_evict_lock */ | |
575 | thread_exit(); | |
576 | } | |
577 | ||
578 | /* | |
579 | * Wake up the dbuf eviction thread if the dbuf cache is at its max size. | |
580 | * If the dbuf cache is at its high water mark, then evict a dbuf from the | |
581 | * dbuf cache using the callers context. | |
582 | */ | |
583 | static void | |
584 | dbuf_evict_notify(void) | |
585 | { | |
586 | ||
587 | /* | |
588 | * We use thread specific data to track when a thread has | |
589 | * started processing evictions. This allows us to avoid deeply | |
590 | * nested stacks that would have a call flow similar to this: | |
591 | * | |
592 | * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify() | |
593 | * ^ | | |
594 | * | | | |
595 | * +-----dbuf_destroy()<--dbuf_evict_one()<--------+ | |
596 | * | |
597 | * The dbuf_eviction_thread will always have its tsd set until | |
598 | * that thread exits. All other threads will only set their tsd | |
599 | * if they are participating in the eviction process. This only | |
600 | * happens if the eviction thread is unable to process evictions | |
601 | * fast enough. To keep the dbuf cache size in check, other threads | |
602 | * can evict from the dbuf cache directly. Those threads will set | |
603 | * their tsd values so that we ensure that they only evict one dbuf | |
604 | * from the dbuf cache. | |
605 | */ | |
606 | if (tsd_get(zfs_dbuf_evict_key) != NULL) | |
607 | return; | |
608 | ||
38240ebd MA |
609 | /* |
610 | * We check if we should evict without holding the dbuf_evict_lock, | |
611 | * because it's OK to occasionally make the wrong decision here, | |
612 | * and grabbing the lock results in massive lock contention. | |
613 | */ | |
e71cade6 | 614 | if (refcount_count(&dbuf_cache_size) > dbuf_cache_target_bytes()) { |
38240ebd | 615 | if (dbuf_cache_above_hiwater()) |
d3c2ae1c | 616 | dbuf_evict_one(); |
38240ebd | 617 | cv_signal(&dbuf_evict_cv); |
d3c2ae1c | 618 | } |
34dc7c2f BB |
619 | } |
620 | ||
d3c2ae1c GW |
621 | |
622 | ||
34dc7c2f BB |
623 | void |
624 | dbuf_init(void) | |
625 | { | |
626 | uint64_t hsize = 1ULL << 16; | |
627 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
628 | int i; | |
629 | ||
630 | /* | |
631 | * The hash table is big enough to fill all of physical memory | |
69de3421 TC |
632 | * with an average block size of zfs_arc_average_blocksize (default 8K). |
633 | * By default, the table will take up | |
634 | * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). | |
34dc7c2f | 635 | */ |
69de3421 | 636 | while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE) |
34dc7c2f BB |
637 | hsize <<= 1; |
638 | ||
639 | retry: | |
640 | h->hash_table_mask = hsize - 1; | |
00b46022 | 641 | #if defined(_KERNEL) && defined(HAVE_SPL) |
d1d7e268 MK |
642 | /* |
643 | * Large allocations which do not require contiguous pages | |
644 | * should be using vmem_alloc() in the linux kernel | |
645 | */ | |
79c76d5b | 646 | h->hash_table = vmem_zalloc(hsize * sizeof (void *), KM_SLEEP); |
00b46022 | 647 | #else |
34dc7c2f | 648 | h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP); |
00b46022 | 649 | #endif |
34dc7c2f BB |
650 | if (h->hash_table == NULL) { |
651 | /* XXX - we should really return an error instead of assert */ | |
652 | ASSERT(hsize > (1ULL << 10)); | |
653 | hsize >>= 1; | |
654 | goto retry; | |
655 | } | |
656 | ||
d3c2ae1c | 657 | dbuf_kmem_cache = kmem_cache_create("dmu_buf_impl_t", |
34dc7c2f BB |
658 | sizeof (dmu_buf_impl_t), |
659 | 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0); | |
660 | ||
661 | for (i = 0; i < DBUF_MUTEXES; i++) | |
40d06e3c | 662 | mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL); |
e0b0ca98 BB |
663 | |
664 | dbuf_stats_init(h); | |
0c66c32d | 665 | |
d3c2ae1c GW |
666 | /* |
667 | * Setup the parameters for the dbuf cache. We cap the size of the | |
668 | * dbuf cache to 1/32nd (default) of the size of the ARC. | |
669 | */ | |
670 | dbuf_cache_max_bytes = MIN(dbuf_cache_max_bytes, | |
e71cade6 | 671 | arc_target_bytes() >> dbuf_cache_max_shift); |
d3c2ae1c | 672 | |
0c66c32d JG |
673 | /* |
674 | * All entries are queued via taskq_dispatch_ent(), so min/maxalloc | |
675 | * configuration is not required. | |
676 | */ | |
1229323d | 677 | dbu_evict_taskq = taskq_create("dbu_evict", 1, defclsyspri, 0, 0, 0); |
d3c2ae1c | 678 | |
64fc7762 | 679 | dbuf_cache = multilist_create(sizeof (dmu_buf_impl_t), |
d3c2ae1c | 680 | offsetof(dmu_buf_impl_t, db_cache_link), |
d3c2ae1c GW |
681 | dbuf_cache_multilist_index_func); |
682 | refcount_create(&dbuf_cache_size); | |
683 | ||
684 | tsd_create(&zfs_dbuf_evict_key, NULL); | |
685 | dbuf_evict_thread_exit = B_FALSE; | |
686 | mutex_init(&dbuf_evict_lock, NULL, MUTEX_DEFAULT, NULL); | |
687 | cv_init(&dbuf_evict_cv, NULL, CV_DEFAULT, NULL); | |
688 | dbuf_cache_evict_thread = thread_create(NULL, 0, dbuf_evict_thread, | |
689 | NULL, 0, &p0, TS_RUN, minclsyspri); | |
34dc7c2f BB |
690 | } |
691 | ||
692 | void | |
693 | dbuf_fini(void) | |
694 | { | |
695 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
696 | int i; | |
697 | ||
e0b0ca98 BB |
698 | dbuf_stats_destroy(); |
699 | ||
34dc7c2f BB |
700 | for (i = 0; i < DBUF_MUTEXES; i++) |
701 | mutex_destroy(&h->hash_mutexes[i]); | |
00b46022 | 702 | #if defined(_KERNEL) && defined(HAVE_SPL) |
d1d7e268 MK |
703 | /* |
704 | * Large allocations which do not require contiguous pages | |
705 | * should be using vmem_free() in the linux kernel | |
706 | */ | |
00b46022 BB |
707 | vmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *)); |
708 | #else | |
34dc7c2f | 709 | kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *)); |
00b46022 | 710 | #endif |
d3c2ae1c | 711 | kmem_cache_destroy(dbuf_kmem_cache); |
0c66c32d | 712 | taskq_destroy(dbu_evict_taskq); |
d3c2ae1c GW |
713 | |
714 | mutex_enter(&dbuf_evict_lock); | |
715 | dbuf_evict_thread_exit = B_TRUE; | |
716 | while (dbuf_evict_thread_exit) { | |
717 | cv_signal(&dbuf_evict_cv); | |
718 | cv_wait(&dbuf_evict_cv, &dbuf_evict_lock); | |
719 | } | |
720 | mutex_exit(&dbuf_evict_lock); | |
721 | tsd_destroy(&zfs_dbuf_evict_key); | |
722 | ||
723 | mutex_destroy(&dbuf_evict_lock); | |
724 | cv_destroy(&dbuf_evict_cv); | |
725 | ||
726 | refcount_destroy(&dbuf_cache_size); | |
64fc7762 | 727 | multilist_destroy(dbuf_cache); |
34dc7c2f BB |
728 | } |
729 | ||
730 | /* | |
731 | * Other stuff. | |
732 | */ | |
733 | ||
734 | #ifdef ZFS_DEBUG | |
735 | static void | |
736 | dbuf_verify(dmu_buf_impl_t *db) | |
737 | { | |
572e2857 | 738 | dnode_t *dn; |
428870ff | 739 | dbuf_dirty_record_t *dr; |
34dc7c2f BB |
740 | |
741 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
742 | ||
743 | if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY)) | |
744 | return; | |
745 | ||
746 | ASSERT(db->db_objset != NULL); | |
572e2857 BB |
747 | DB_DNODE_ENTER(db); |
748 | dn = DB_DNODE(db); | |
34dc7c2f BB |
749 | if (dn == NULL) { |
750 | ASSERT(db->db_parent == NULL); | |
751 | ASSERT(db->db_blkptr == NULL); | |
752 | } else { | |
753 | ASSERT3U(db->db.db_object, ==, dn->dn_object); | |
754 | ASSERT3P(db->db_objset, ==, dn->dn_objset); | |
755 | ASSERT3U(db->db_level, <, dn->dn_nlevels); | |
572e2857 BB |
756 | ASSERT(db->db_blkid == DMU_BONUS_BLKID || |
757 | db->db_blkid == DMU_SPILL_BLKID || | |
8951cb8d | 758 | !avl_is_empty(&dn->dn_dbufs)); |
34dc7c2f | 759 | } |
428870ff BB |
760 | if (db->db_blkid == DMU_BONUS_BLKID) { |
761 | ASSERT(dn != NULL); | |
762 | ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); | |
763 | ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID); | |
764 | } else if (db->db_blkid == DMU_SPILL_BLKID) { | |
34dc7c2f | 765 | ASSERT(dn != NULL); |
c99c9001 | 766 | ASSERT0(db->db.db_offset); |
34dc7c2f BB |
767 | } else { |
768 | ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size); | |
769 | } | |
770 | ||
428870ff BB |
771 | for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next) |
772 | ASSERT(dr->dr_dbuf == db); | |
773 | ||
774 | for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next) | |
775 | ASSERT(dr->dr_dbuf == db); | |
776 | ||
b128c09f BB |
777 | /* |
778 | * We can't assert that db_size matches dn_datablksz because it | |
779 | * can be momentarily different when another thread is doing | |
780 | * dnode_set_blksz(). | |
781 | */ | |
782 | if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) { | |
428870ff | 783 | dr = db->db_data_pending; |
b128c09f BB |
784 | /* |
785 | * It should only be modified in syncing context, so | |
786 | * make sure we only have one copy of the data. | |
787 | */ | |
788 | ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf); | |
34dc7c2f BB |
789 | } |
790 | ||
791 | /* verify db->db_blkptr */ | |
792 | if (db->db_blkptr) { | |
793 | if (db->db_parent == dn->dn_dbuf) { | |
794 | /* db is pointed to by the dnode */ | |
795 | /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */ | |
9babb374 | 796 | if (DMU_OBJECT_IS_SPECIAL(db->db.db_object)) |
34dc7c2f BB |
797 | ASSERT(db->db_parent == NULL); |
798 | else | |
799 | ASSERT(db->db_parent != NULL); | |
428870ff BB |
800 | if (db->db_blkid != DMU_SPILL_BLKID) |
801 | ASSERT3P(db->db_blkptr, ==, | |
802 | &dn->dn_phys->dn_blkptr[db->db_blkid]); | |
34dc7c2f BB |
803 | } else { |
804 | /* db is pointed to by an indirect block */ | |
1fde1e37 | 805 | ASSERTV(int epb = db->db_parent->db.db_size >> |
02730c33 | 806 | SPA_BLKPTRSHIFT); |
34dc7c2f BB |
807 | ASSERT3U(db->db_parent->db_level, ==, db->db_level+1); |
808 | ASSERT3U(db->db_parent->db.db_object, ==, | |
809 | db->db.db_object); | |
810 | /* | |
811 | * dnode_grow_indblksz() can make this fail if we don't | |
812 | * have the struct_rwlock. XXX indblksz no longer | |
813 | * grows. safe to do this now? | |
814 | */ | |
572e2857 | 815 | if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) { |
34dc7c2f BB |
816 | ASSERT3P(db->db_blkptr, ==, |
817 | ((blkptr_t *)db->db_parent->db.db_data + | |
818 | db->db_blkid % epb)); | |
819 | } | |
820 | } | |
821 | } | |
822 | if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) && | |
428870ff BB |
823 | (db->db_buf == NULL || db->db_buf->b_data) && |
824 | db->db.db_data && db->db_blkid != DMU_BONUS_BLKID && | |
34dc7c2f BB |
825 | db->db_state != DB_FILL && !dn->dn_free_txg) { |
826 | /* | |
827 | * If the blkptr isn't set but they have nonzero data, | |
828 | * it had better be dirty, otherwise we'll lose that | |
829 | * data when we evict this buffer. | |
bc77ba73 PD |
830 | * |
831 | * There is an exception to this rule for indirect blocks; in | |
832 | * this case, if the indirect block is a hole, we fill in a few | |
833 | * fields on each of the child blocks (importantly, birth time) | |
834 | * to prevent hole birth times from being lost when you | |
835 | * partially fill in a hole. | |
34dc7c2f BB |
836 | */ |
837 | if (db->db_dirtycnt == 0) { | |
bc77ba73 PD |
838 | if (db->db_level == 0) { |
839 | uint64_t *buf = db->db.db_data; | |
840 | int i; | |
34dc7c2f | 841 | |
bc77ba73 PD |
842 | for (i = 0; i < db->db.db_size >> 3; i++) { |
843 | ASSERT(buf[i] == 0); | |
844 | } | |
845 | } else { | |
bc77ba73 PD |
846 | blkptr_t *bps = db->db.db_data; |
847 | ASSERT3U(1 << DB_DNODE(db)->dn_indblkshift, ==, | |
848 | db->db.db_size); | |
849 | /* | |
850 | * We want to verify that all the blkptrs in the | |
851 | * indirect block are holes, but we may have | |
852 | * automatically set up a few fields for them. | |
853 | * We iterate through each blkptr and verify | |
854 | * they only have those fields set. | |
855 | */ | |
1c27024e | 856 | for (int i = 0; |
bc77ba73 PD |
857 | i < db->db.db_size / sizeof (blkptr_t); |
858 | i++) { | |
859 | blkptr_t *bp = &bps[i]; | |
860 | ASSERT(ZIO_CHECKSUM_IS_ZERO( | |
861 | &bp->blk_cksum)); | |
862 | ASSERT( | |
863 | DVA_IS_EMPTY(&bp->blk_dva[0]) && | |
864 | DVA_IS_EMPTY(&bp->blk_dva[1]) && | |
865 | DVA_IS_EMPTY(&bp->blk_dva[2])); | |
866 | ASSERT0(bp->blk_fill); | |
867 | ASSERT0(bp->blk_pad[0]); | |
868 | ASSERT0(bp->blk_pad[1]); | |
869 | ASSERT(!BP_IS_EMBEDDED(bp)); | |
870 | ASSERT(BP_IS_HOLE(bp)); | |
871 | ASSERT0(bp->blk_phys_birth); | |
872 | } | |
34dc7c2f BB |
873 | } |
874 | } | |
875 | } | |
572e2857 | 876 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
877 | } |
878 | #endif | |
879 | ||
0c66c32d JG |
880 | static void |
881 | dbuf_clear_data(dmu_buf_impl_t *db) | |
882 | { | |
883 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
884 | dbuf_evict_user(db); | |
d3c2ae1c | 885 | ASSERT3P(db->db_buf, ==, NULL); |
0c66c32d JG |
886 | db->db.db_data = NULL; |
887 | if (db->db_state != DB_NOFILL) | |
888 | db->db_state = DB_UNCACHED; | |
889 | } | |
890 | ||
34dc7c2f BB |
891 | static void |
892 | dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf) | |
893 | { | |
894 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
0c66c32d JG |
895 | ASSERT(buf != NULL); |
896 | ||
34dc7c2f | 897 | db->db_buf = buf; |
0c66c32d JG |
898 | ASSERT(buf->b_data != NULL); |
899 | db->db.db_data = buf->b_data; | |
34dc7c2f BB |
900 | } |
901 | ||
428870ff BB |
902 | /* |
903 | * Loan out an arc_buf for read. Return the loaned arc_buf. | |
904 | */ | |
905 | arc_buf_t * | |
906 | dbuf_loan_arcbuf(dmu_buf_impl_t *db) | |
907 | { | |
908 | arc_buf_t *abuf; | |
909 | ||
d3c2ae1c | 910 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
428870ff BB |
911 | mutex_enter(&db->db_mtx); |
912 | if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) { | |
913 | int blksz = db->db.db_size; | |
b0bc7a84 | 914 | spa_t *spa = db->db_objset->os_spa; |
572e2857 | 915 | |
428870ff | 916 | mutex_exit(&db->db_mtx); |
2aa34383 | 917 | abuf = arc_loan_buf(spa, B_FALSE, blksz); |
428870ff BB |
918 | bcopy(db->db.db_data, abuf->b_data, blksz); |
919 | } else { | |
920 | abuf = db->db_buf; | |
921 | arc_loan_inuse_buf(abuf, db); | |
d3c2ae1c | 922 | db->db_buf = NULL; |
0c66c32d | 923 | dbuf_clear_data(db); |
428870ff BB |
924 | mutex_exit(&db->db_mtx); |
925 | } | |
926 | return (abuf); | |
927 | } | |
928 | ||
fcff0f35 PD |
929 | /* |
930 | * Calculate which level n block references the data at the level 0 offset | |
931 | * provided. | |
932 | */ | |
34dc7c2f | 933 | uint64_t |
031d7c2f | 934 | dbuf_whichblock(const dnode_t *dn, const int64_t level, const uint64_t offset) |
34dc7c2f | 935 | { |
fcff0f35 PD |
936 | if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) { |
937 | /* | |
938 | * The level n blkid is equal to the level 0 blkid divided by | |
939 | * the number of level 0s in a level n block. | |
940 | * | |
941 | * The level 0 blkid is offset >> datablkshift = | |
942 | * offset / 2^datablkshift. | |
943 | * | |
944 | * The number of level 0s in a level n is the number of block | |
945 | * pointers in an indirect block, raised to the power of level. | |
946 | * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level = | |
947 | * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)). | |
948 | * | |
949 | * Thus, the level n blkid is: offset / | |
950 | * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT))) | |
951 | * = offset / 2^(datablkshift + level * | |
952 | * (indblkshift - SPA_BLKPTRSHIFT)) | |
953 | * = offset >> (datablkshift + level * | |
954 | * (indblkshift - SPA_BLKPTRSHIFT)) | |
955 | */ | |
031d7c2f GN |
956 | |
957 | const unsigned exp = dn->dn_datablkshift + | |
958 | level * (dn->dn_indblkshift - SPA_BLKPTRSHIFT); | |
959 | ||
960 | if (exp >= 8 * sizeof (offset)) { | |
961 | /* This only happens on the highest indirection level */ | |
962 | ASSERT3U(level, ==, dn->dn_nlevels - 1); | |
963 | return (0); | |
964 | } | |
965 | ||
966 | ASSERT3U(exp, <, 8 * sizeof (offset)); | |
967 | ||
968 | return (offset >> exp); | |
34dc7c2f BB |
969 | } else { |
970 | ASSERT3U(offset, <, dn->dn_datablksz); | |
971 | return (0); | |
972 | } | |
973 | } | |
974 | ||
975 | static void | |
d4a72f23 TC |
976 | dbuf_read_done(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp, |
977 | arc_buf_t *buf, void *vdb) | |
34dc7c2f BB |
978 | { |
979 | dmu_buf_impl_t *db = vdb; | |
980 | ||
981 | mutex_enter(&db->db_mtx); | |
982 | ASSERT3U(db->db_state, ==, DB_READ); | |
983 | /* | |
984 | * All reads are synchronous, so we must have a hold on the dbuf | |
985 | */ | |
986 | ASSERT(refcount_count(&db->db_holds) > 0); | |
987 | ASSERT(db->db_buf == NULL); | |
988 | ASSERT(db->db.db_data == NULL); | |
989 | if (db->db_level == 0 && db->db_freed_in_flight) { | |
990 | /* we were freed in flight; disregard any error */ | |
d4a72f23 TC |
991 | if (buf == NULL) { |
992 | buf = arc_alloc_buf(db->db_objset->os_spa, | |
993 | db, DBUF_GET_BUFC_TYPE(db), db->db.db_size); | |
994 | } | |
34dc7c2f BB |
995 | arc_release(buf, db); |
996 | bzero(buf->b_data, db->db.db_size); | |
997 | arc_buf_freeze(buf); | |
998 | db->db_freed_in_flight = FALSE; | |
999 | dbuf_set_data(db, buf); | |
1000 | db->db_state = DB_CACHED; | |
d4a72f23 | 1001 | } else if (buf != NULL) { |
34dc7c2f BB |
1002 | dbuf_set_data(db, buf); |
1003 | db->db_state = DB_CACHED; | |
1004 | } else { | |
428870ff | 1005 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f | 1006 | ASSERT3P(db->db_buf, ==, NULL); |
34dc7c2f BB |
1007 | db->db_state = DB_UNCACHED; |
1008 | } | |
1009 | cv_broadcast(&db->db_changed); | |
428870ff | 1010 | dbuf_rele_and_unlock(db, NULL); |
34dc7c2f BB |
1011 | } |
1012 | ||
5f6d0b6f | 1013 | static int |
7f60329a | 1014 | dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) |
34dc7c2f | 1015 | { |
572e2857 | 1016 | dnode_t *dn; |
5dbd68a3 | 1017 | zbookmark_phys_t zb; |
2a432414 | 1018 | uint32_t aflags = ARC_FLAG_NOWAIT; |
b5256303 | 1019 | int err, zio_flags = 0; |
34dc7c2f | 1020 | |
572e2857 BB |
1021 | DB_DNODE_ENTER(db); |
1022 | dn = DB_DNODE(db); | |
34dc7c2f BB |
1023 | ASSERT(!refcount_is_zero(&db->db_holds)); |
1024 | /* We need the struct_rwlock to prevent db_blkptr from changing. */ | |
b128c09f | 1025 | ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); |
34dc7c2f BB |
1026 | ASSERT(MUTEX_HELD(&db->db_mtx)); |
1027 | ASSERT(db->db_state == DB_UNCACHED); | |
1028 | ASSERT(db->db_buf == NULL); | |
1029 | ||
428870ff | 1030 | if (db->db_blkid == DMU_BONUS_BLKID) { |
50c957f7 NB |
1031 | /* |
1032 | * The bonus length stored in the dnode may be less than | |
1033 | * the maximum available space in the bonus buffer. | |
1034 | */ | |
9babb374 | 1035 | int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen); |
50c957f7 | 1036 | int max_bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots); |
b5256303 TC |
1037 | arc_buf_t *dn_buf = (dn->dn_dbuf != NULL) ? |
1038 | dn->dn_dbuf->db_buf : NULL; | |
1039 | ||
1040 | /* if the underlying dnode block is encrypted, decrypt it */ | |
1041 | if (dn_buf != NULL && dn->dn_objset->os_encrypted && | |
1042 | DMU_OT_IS_ENCRYPTED(dn->dn_bonustype) && | |
1043 | (flags & DB_RF_NO_DECRYPT) == 0 && | |
1044 | arc_is_encrypted(dn_buf)) { | |
1045 | err = arc_untransform(dn_buf, dn->dn_objset->os_spa, | |
1046 | dmu_objset_id(dn->dn_objset), B_TRUE); | |
1047 | if (err != 0) { | |
1048 | DB_DNODE_EXIT(db); | |
1049 | mutex_exit(&db->db_mtx); | |
1050 | return (err); | |
1051 | } | |
1052 | } | |
34dc7c2f BB |
1053 | |
1054 | ASSERT3U(bonuslen, <=, db->db.db_size); | |
a3fd9d9e | 1055 | db->db.db_data = kmem_alloc(max_bonuslen, KM_SLEEP); |
25458cbe | 1056 | arc_space_consume(max_bonuslen, ARC_SPACE_BONUS); |
50c957f7 NB |
1057 | if (bonuslen < max_bonuslen) |
1058 | bzero(db->db.db_data, max_bonuslen); | |
9babb374 BB |
1059 | if (bonuslen) |
1060 | bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen); | |
572e2857 | 1061 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
1062 | db->db_state = DB_CACHED; |
1063 | mutex_exit(&db->db_mtx); | |
5f6d0b6f | 1064 | return (0); |
34dc7c2f BB |
1065 | } |
1066 | ||
b128c09f BB |
1067 | /* |
1068 | * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync() | |
1069 | * processes the delete record and clears the bp while we are waiting | |
1070 | * for the dn_mtx (resulting in a "no" from block_freed). | |
1071 | */ | |
1072 | if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) || | |
1073 | (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) || | |
1074 | BP_IS_HOLE(db->db_blkptr)))) { | |
34dc7c2f BB |
1075 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); |
1076 | ||
2aa34383 DK |
1077 | dbuf_set_data(db, arc_alloc_buf(db->db_objset->os_spa, db, type, |
1078 | db->db.db_size)); | |
34dc7c2f | 1079 | bzero(db->db.db_data, db->db.db_size); |
bc77ba73 PD |
1080 | |
1081 | if (db->db_blkptr != NULL && db->db_level > 0 && | |
1082 | BP_IS_HOLE(db->db_blkptr) && | |
1083 | db->db_blkptr->blk_birth != 0) { | |
1084 | blkptr_t *bps = db->db.db_data; | |
1c27024e | 1085 | for (int i = 0; i < ((1 << |
bc77ba73 PD |
1086 | DB_DNODE(db)->dn_indblkshift) / sizeof (blkptr_t)); |
1087 | i++) { | |
1088 | blkptr_t *bp = &bps[i]; | |
1089 | ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, | |
1090 | 1 << dn->dn_indblkshift); | |
1091 | BP_SET_LSIZE(bp, | |
1092 | BP_GET_LEVEL(db->db_blkptr) == 1 ? | |
1093 | dn->dn_datablksz : | |
1094 | BP_GET_LSIZE(db->db_blkptr)); | |
1095 | BP_SET_TYPE(bp, BP_GET_TYPE(db->db_blkptr)); | |
1096 | BP_SET_LEVEL(bp, | |
1097 | BP_GET_LEVEL(db->db_blkptr) - 1); | |
1098 | BP_SET_BIRTH(bp, db->db_blkptr->blk_birth, 0); | |
1099 | } | |
1100 | } | |
1101 | DB_DNODE_EXIT(db); | |
34dc7c2f | 1102 | db->db_state = DB_CACHED; |
34dc7c2f | 1103 | mutex_exit(&db->db_mtx); |
5f6d0b6f | 1104 | return (0); |
34dc7c2f BB |
1105 | } |
1106 | ||
572e2857 BB |
1107 | DB_DNODE_EXIT(db); |
1108 | ||
34dc7c2f BB |
1109 | db->db_state = DB_READ; |
1110 | mutex_exit(&db->db_mtx); | |
1111 | ||
b128c09f | 1112 | if (DBUF_IS_L2CACHEABLE(db)) |
2a432414 | 1113 | aflags |= ARC_FLAG_L2CACHE; |
b128c09f | 1114 | |
428870ff BB |
1115 | SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ? |
1116 | db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET, | |
1117 | db->db.db_object, db->db_level, db->db_blkid); | |
34dc7c2f | 1118 | |
b5256303 TC |
1119 | /* |
1120 | * All bps of an encrypted os should have the encryption bit set. | |
1121 | * If this is not true it indicates tampering and we report an error. | |
1122 | */ | |
1123 | if (db->db_objset->os_encrypted && !BP_USES_CRYPT(db->db_blkptr)) { | |
1124 | spa_log_error(db->db_objset->os_spa, &zb); | |
1125 | zfs_panic_recover("unencrypted block in encrypted " | |
1126 | "object set %llu", dmu_objset_id(db->db_objset)); | |
1127 | return (SET_ERROR(EIO)); | |
1128 | } | |
1129 | ||
34dc7c2f | 1130 | dbuf_add_ref(db, NULL); |
b128c09f | 1131 | |
b5256303 TC |
1132 | zio_flags = (flags & DB_RF_CANFAIL) ? |
1133 | ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED; | |
1134 | ||
1135 | if ((flags & DB_RF_NO_DECRYPT) && BP_IS_PROTECTED(db->db_blkptr)) | |
1136 | zio_flags |= ZIO_FLAG_RAW; | |
1137 | ||
5f6d0b6f | 1138 | err = arc_read(zio, db->db_objset->os_spa, db->db_blkptr, |
b5256303 | 1139 | dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ, zio_flags, |
34dc7c2f | 1140 | &aflags, &zb); |
5f6d0b6f | 1141 | |
da8d5748 | 1142 | return (err); |
34dc7c2f BB |
1143 | } |
1144 | ||
2aa34383 DK |
1145 | /* |
1146 | * This is our just-in-time copy function. It makes a copy of buffers that | |
1147 | * have been modified in a previous transaction group before we access them in | |
1148 | * the current active group. | |
1149 | * | |
1150 | * This function is used in three places: when we are dirtying a buffer for the | |
1151 | * first time in a txg, when we are freeing a range in a dnode that includes | |
1152 | * this buffer, and when we are accessing a buffer which was received compressed | |
1153 | * and later referenced in a WRITE_BYREF record. | |
1154 | * | |
1155 | * Note that when we are called from dbuf_free_range() we do not put a hold on | |
1156 | * the buffer, we just traverse the active dbuf list for the dnode. | |
1157 | */ | |
1158 | static void | |
1159 | dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg) | |
1160 | { | |
1161 | dbuf_dirty_record_t *dr = db->db_last_dirty; | |
1162 | ||
1163 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
1164 | ASSERT(db->db.db_data != NULL); | |
1165 | ASSERT(db->db_level == 0); | |
1166 | ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT); | |
1167 | ||
1168 | if (dr == NULL || | |
1169 | (dr->dt.dl.dr_data != | |
1170 | ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf))) | |
1171 | return; | |
1172 | ||
1173 | /* | |
1174 | * If the last dirty record for this dbuf has not yet synced | |
1175 | * and its referencing the dbuf data, either: | |
1176 | * reset the reference to point to a new copy, | |
1177 | * or (if there a no active holders) | |
1178 | * just null out the current db_data pointer. | |
1179 | */ | |
4807c0ba | 1180 | ASSERT3U(dr->dr_txg, >=, txg - 2); |
2aa34383 | 1181 | if (db->db_blkid == DMU_BONUS_BLKID) { |
2aa34383 DK |
1182 | dnode_t *dn = DB_DNODE(db); |
1183 | int bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots); | |
a3fd9d9e | 1184 | dr->dt.dl.dr_data = kmem_alloc(bonuslen, KM_SLEEP); |
2aa34383 DK |
1185 | arc_space_consume(bonuslen, ARC_SPACE_BONUS); |
1186 | bcopy(db->db.db_data, dr->dt.dl.dr_data, bonuslen); | |
1187 | } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) { | |
b5256303 | 1188 | dnode_t *dn = DB_DNODE(db); |
2aa34383 DK |
1189 | int size = arc_buf_size(db->db_buf); |
1190 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); | |
1191 | spa_t *spa = db->db_objset->os_spa; | |
1192 | enum zio_compress compress_type = | |
1193 | arc_get_compression(db->db_buf); | |
1194 | ||
b5256303 TC |
1195 | if (arc_is_encrypted(db->db_buf)) { |
1196 | boolean_t byteorder; | |
1197 | uint8_t salt[ZIO_DATA_SALT_LEN]; | |
1198 | uint8_t iv[ZIO_DATA_IV_LEN]; | |
1199 | uint8_t mac[ZIO_DATA_MAC_LEN]; | |
1200 | ||
1201 | arc_get_raw_params(db->db_buf, &byteorder, salt, | |
1202 | iv, mac); | |
1203 | dr->dt.dl.dr_data = arc_alloc_raw_buf(spa, db, | |
1204 | dmu_objset_id(dn->dn_objset), byteorder, salt, iv, | |
1205 | mac, dn->dn_type, size, arc_buf_lsize(db->db_buf), | |
1206 | compress_type); | |
1207 | } else if (compress_type != ZIO_COMPRESS_OFF) { | |
2aa34383 DK |
1208 | ASSERT3U(type, ==, ARC_BUFC_DATA); |
1209 | dr->dt.dl.dr_data = arc_alloc_compressed_buf(spa, db, | |
1210 | size, arc_buf_lsize(db->db_buf), compress_type); | |
b5256303 TC |
1211 | } else { |
1212 | dr->dt.dl.dr_data = arc_alloc_buf(spa, db, type, size); | |
2aa34383 DK |
1213 | } |
1214 | bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size); | |
1215 | } else { | |
1216 | db->db_buf = NULL; | |
1217 | dbuf_clear_data(db); | |
1218 | } | |
1219 | } | |
1220 | ||
34dc7c2f BB |
1221 | int |
1222 | dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) | |
1223 | { | |
1224 | int err = 0; | |
b0bc7a84 | 1225 | boolean_t prefetch; |
572e2857 | 1226 | dnode_t *dn; |
34dc7c2f BB |
1227 | |
1228 | /* | |
1229 | * We don't have to hold the mutex to check db_state because it | |
1230 | * can't be freed while we have a hold on the buffer. | |
1231 | */ | |
1232 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
1233 | ||
b128c09f | 1234 | if (db->db_state == DB_NOFILL) |
2e528b49 | 1235 | return (SET_ERROR(EIO)); |
b128c09f | 1236 | |
572e2857 BB |
1237 | DB_DNODE_ENTER(db); |
1238 | dn = DB_DNODE(db); | |
34dc7c2f | 1239 | if ((flags & DB_RF_HAVESTRUCT) == 0) |
572e2857 | 1240 | rw_enter(&dn->dn_struct_rwlock, RW_READER); |
34dc7c2f | 1241 | |
428870ff | 1242 | prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && |
572e2857 | 1243 | (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL && |
b128c09f | 1244 | DBUF_IS_CACHEABLE(db); |
34dc7c2f BB |
1245 | |
1246 | mutex_enter(&db->db_mtx); | |
1247 | if (db->db_state == DB_CACHED) { | |
b5256303 TC |
1248 | spa_t *spa = dn->dn_objset->os_spa; |
1249 | ||
2aa34383 | 1250 | /* |
b5256303 TC |
1251 | * If the arc buf is compressed or encrypted, we need to |
1252 | * untransform it to read the data. This could happen during | |
1253 | * the "zfs receive" of a stream which is deduplicated and | |
1254 | * either raw or compressed. We do not need to do this if the | |
1255 | * caller wants raw encrypted data. | |
2aa34383 | 1256 | */ |
b5256303 TC |
1257 | if (db->db_buf != NULL && (flags & DB_RF_NO_DECRYPT) == 0 && |
1258 | (arc_is_encrypted(db->db_buf) || | |
1259 | arc_get_compression(db->db_buf) != ZIO_COMPRESS_OFF)) { | |
1260 | dbuf_fix_old_data(db, spa_syncing_txg(spa)); | |
1261 | err = arc_untransform(db->db_buf, spa, | |
1262 | dmu_objset_id(db->db_objset), B_FALSE); | |
2aa34383 DK |
1263 | dbuf_set_data(db, db->db_buf); |
1264 | } | |
34dc7c2f BB |
1265 | mutex_exit(&db->db_mtx); |
1266 | if (prefetch) | |
755065f3 | 1267 | dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE); |
34dc7c2f | 1268 | if ((flags & DB_RF_HAVESTRUCT) == 0) |
572e2857 BB |
1269 | rw_exit(&dn->dn_struct_rwlock); |
1270 | DB_DNODE_EXIT(db); | |
34dc7c2f | 1271 | } else if (db->db_state == DB_UNCACHED) { |
572e2857 | 1272 | spa_t *spa = dn->dn_objset->os_spa; |
a0043383 | 1273 | boolean_t need_wait = B_FALSE; |
572e2857 | 1274 | |
b0319c1f | 1275 | if (zio == NULL && |
a0043383 | 1276 | db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) { |
572e2857 | 1277 | zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); |
a0043383 MA |
1278 | need_wait = B_TRUE; |
1279 | } | |
7f60329a | 1280 | err = dbuf_read_impl(db, zio, flags); |
34dc7c2f BB |
1281 | |
1282 | /* dbuf_read_impl has dropped db_mtx for us */ | |
1283 | ||
5f6d0b6f | 1284 | if (!err && prefetch) |
755065f3 | 1285 | dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE); |
34dc7c2f BB |
1286 | |
1287 | if ((flags & DB_RF_HAVESTRUCT) == 0) | |
572e2857 BB |
1288 | rw_exit(&dn->dn_struct_rwlock); |
1289 | DB_DNODE_EXIT(db); | |
34dc7c2f | 1290 | |
a0043383 | 1291 | if (!err && need_wait) |
34dc7c2f BB |
1292 | err = zio_wait(zio); |
1293 | } else { | |
e49f1e20 WA |
1294 | /* |
1295 | * Another reader came in while the dbuf was in flight | |
1296 | * between UNCACHED and CACHED. Either a writer will finish | |
1297 | * writing the buffer (sending the dbuf to CACHED) or the | |
1298 | * first reader's request will reach the read_done callback | |
1299 | * and send the dbuf to CACHED. Otherwise, a failure | |
1300 | * occurred and the dbuf went to UNCACHED. | |
1301 | */ | |
34dc7c2f BB |
1302 | mutex_exit(&db->db_mtx); |
1303 | if (prefetch) | |
755065f3 | 1304 | dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE); |
34dc7c2f | 1305 | if ((flags & DB_RF_HAVESTRUCT) == 0) |
572e2857 BB |
1306 | rw_exit(&dn->dn_struct_rwlock); |
1307 | DB_DNODE_EXIT(db); | |
34dc7c2f | 1308 | |
e49f1e20 | 1309 | /* Skip the wait per the caller's request. */ |
34dc7c2f BB |
1310 | mutex_enter(&db->db_mtx); |
1311 | if ((flags & DB_RF_NEVERWAIT) == 0) { | |
1312 | while (db->db_state == DB_READ || | |
1313 | db->db_state == DB_FILL) { | |
1314 | ASSERT(db->db_state == DB_READ || | |
1315 | (flags & DB_RF_HAVESTRUCT) == 0); | |
64dbba36 AL |
1316 | DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *, |
1317 | db, zio_t *, zio); | |
34dc7c2f BB |
1318 | cv_wait(&db->db_changed, &db->db_mtx); |
1319 | } | |
1320 | if (db->db_state == DB_UNCACHED) | |
2e528b49 | 1321 | err = SET_ERROR(EIO); |
34dc7c2f BB |
1322 | } |
1323 | mutex_exit(&db->db_mtx); | |
1324 | } | |
1325 | ||
34dc7c2f BB |
1326 | return (err); |
1327 | } | |
1328 | ||
1329 | static void | |
1330 | dbuf_noread(dmu_buf_impl_t *db) | |
1331 | { | |
1332 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
428870ff | 1333 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f BB |
1334 | mutex_enter(&db->db_mtx); |
1335 | while (db->db_state == DB_READ || db->db_state == DB_FILL) | |
1336 | cv_wait(&db->db_changed, &db->db_mtx); | |
1337 | if (db->db_state == DB_UNCACHED) { | |
1338 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); | |
b0bc7a84 | 1339 | spa_t *spa = db->db_objset->os_spa; |
34dc7c2f BB |
1340 | |
1341 | ASSERT(db->db_buf == NULL); | |
1342 | ASSERT(db->db.db_data == NULL); | |
2aa34383 | 1343 | dbuf_set_data(db, arc_alloc_buf(spa, db, type, db->db.db_size)); |
34dc7c2f | 1344 | db->db_state = DB_FILL; |
b128c09f | 1345 | } else if (db->db_state == DB_NOFILL) { |
0c66c32d | 1346 | dbuf_clear_data(db); |
34dc7c2f BB |
1347 | } else { |
1348 | ASSERT3U(db->db_state, ==, DB_CACHED); | |
1349 | } | |
1350 | mutex_exit(&db->db_mtx); | |
1351 | } | |
1352 | ||
34dc7c2f BB |
1353 | void |
1354 | dbuf_unoverride(dbuf_dirty_record_t *dr) | |
1355 | { | |
1356 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
428870ff | 1357 | blkptr_t *bp = &dr->dt.dl.dr_overridden_by; |
34dc7c2f BB |
1358 | uint64_t txg = dr->dr_txg; |
1359 | ||
1360 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
00710365 AS |
1361 | /* |
1362 | * This assert is valid because dmu_sync() expects to be called by | |
1363 | * a zilog's get_data while holding a range lock. This call only | |
1364 | * comes from dbuf_dirty() callers who must also hold a range lock. | |
1365 | */ | |
34dc7c2f BB |
1366 | ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC); |
1367 | ASSERT(db->db_level == 0); | |
1368 | ||
428870ff | 1369 | if (db->db_blkid == DMU_BONUS_BLKID || |
34dc7c2f BB |
1370 | dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN) |
1371 | return; | |
1372 | ||
428870ff BB |
1373 | ASSERT(db->db_data_pending != dr); |
1374 | ||
34dc7c2f | 1375 | /* free this block */ |
b0bc7a84 MG |
1376 | if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite) |
1377 | zio_free(db->db_objset->os_spa, txg, bp); | |
428870ff | 1378 | |
34dc7c2f | 1379 | dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; |
03c6040b | 1380 | dr->dt.dl.dr_nopwrite = B_FALSE; |
b5256303 | 1381 | dr->dt.dl.dr_raw = B_FALSE; |
03c6040b | 1382 | |
34dc7c2f BB |
1383 | /* |
1384 | * Release the already-written buffer, so we leave it in | |
1385 | * a consistent dirty state. Note that all callers are | |
1386 | * modifying the buffer, so they will immediately do | |
1387 | * another (redundant) arc_release(). Therefore, leave | |
1388 | * the buf thawed to save the effort of freezing & | |
1389 | * immediately re-thawing it. | |
1390 | */ | |
1391 | arc_release(dr->dt.dl.dr_data, db); | |
1392 | } | |
1393 | ||
b128c09f BB |
1394 | /* |
1395 | * Evict (if its unreferenced) or clear (if its referenced) any level-0 | |
1396 | * data blocks in the free range, so that any future readers will find | |
b0bc7a84 | 1397 | * empty blocks. |
b128c09f | 1398 | */ |
34dc7c2f | 1399 | void |
8951cb8d AR |
1400 | dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid, |
1401 | dmu_tx_t *tx) | |
34dc7c2f | 1402 | { |
0c66c32d JG |
1403 | dmu_buf_impl_t *db_search; |
1404 | dmu_buf_impl_t *db, *db_next; | |
34dc7c2f | 1405 | uint64_t txg = tx->tx_txg; |
8951cb8d | 1406 | avl_index_t where; |
8951cb8d | 1407 | |
9c9531cb GM |
1408 | if (end_blkid > dn->dn_maxblkid && |
1409 | !(start_blkid == DMU_SPILL_BLKID || end_blkid == DMU_SPILL_BLKID)) | |
8951cb8d AR |
1410 | end_blkid = dn->dn_maxblkid; |
1411 | dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid); | |
34dc7c2f | 1412 | |
0c66c32d | 1413 | db_search = kmem_alloc(sizeof (dmu_buf_impl_t), KM_SLEEP); |
8951cb8d AR |
1414 | db_search->db_level = 0; |
1415 | db_search->db_blkid = start_blkid; | |
9925c28c | 1416 | db_search->db_state = DB_SEARCH; |
ea97f8ce | 1417 | |
b663a23d | 1418 | mutex_enter(&dn->dn_dbufs_mtx); |
8951cb8d AR |
1419 | db = avl_find(&dn->dn_dbufs, db_search, &where); |
1420 | ASSERT3P(db, ==, NULL); | |
9c9531cb | 1421 | |
8951cb8d AR |
1422 | db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER); |
1423 | ||
1424 | for (; db != NULL; db = db_next) { | |
1425 | db_next = AVL_NEXT(&dn->dn_dbufs, db); | |
428870ff | 1426 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
b128c09f | 1427 | |
8951cb8d AR |
1428 | if (db->db_level != 0 || db->db_blkid > end_blkid) { |
1429 | break; | |
1430 | } | |
1431 | ASSERT3U(db->db_blkid, >=, start_blkid); | |
34dc7c2f BB |
1432 | |
1433 | /* found a level 0 buffer in the range */ | |
13fe0198 MA |
1434 | mutex_enter(&db->db_mtx); |
1435 | if (dbuf_undirty(db, tx)) { | |
1436 | /* mutex has been dropped and dbuf destroyed */ | |
34dc7c2f | 1437 | continue; |
13fe0198 | 1438 | } |
34dc7c2f | 1439 | |
34dc7c2f | 1440 | if (db->db_state == DB_UNCACHED || |
b128c09f | 1441 | db->db_state == DB_NOFILL || |
34dc7c2f BB |
1442 | db->db_state == DB_EVICTING) { |
1443 | ASSERT(db->db.db_data == NULL); | |
1444 | mutex_exit(&db->db_mtx); | |
1445 | continue; | |
1446 | } | |
1447 | if (db->db_state == DB_READ || db->db_state == DB_FILL) { | |
1448 | /* will be handled in dbuf_read_done or dbuf_rele */ | |
1449 | db->db_freed_in_flight = TRUE; | |
1450 | mutex_exit(&db->db_mtx); | |
1451 | continue; | |
1452 | } | |
1453 | if (refcount_count(&db->db_holds) == 0) { | |
1454 | ASSERT(db->db_buf); | |
d3c2ae1c | 1455 | dbuf_destroy(db); |
34dc7c2f BB |
1456 | continue; |
1457 | } | |
1458 | /* The dbuf is referenced */ | |
1459 | ||
1460 | if (db->db_last_dirty != NULL) { | |
1461 | dbuf_dirty_record_t *dr = db->db_last_dirty; | |
1462 | ||
1463 | if (dr->dr_txg == txg) { | |
1464 | /* | |
1465 | * This buffer is "in-use", re-adjust the file | |
1466 | * size to reflect that this buffer may | |
1467 | * contain new data when we sync. | |
1468 | */ | |
428870ff BB |
1469 | if (db->db_blkid != DMU_SPILL_BLKID && |
1470 | db->db_blkid > dn->dn_maxblkid) | |
34dc7c2f BB |
1471 | dn->dn_maxblkid = db->db_blkid; |
1472 | dbuf_unoverride(dr); | |
1473 | } else { | |
1474 | /* | |
1475 | * This dbuf is not dirty in the open context. | |
1476 | * Either uncache it (if its not referenced in | |
1477 | * the open context) or reset its contents to | |
1478 | * empty. | |
1479 | */ | |
1480 | dbuf_fix_old_data(db, txg); | |
1481 | } | |
1482 | } | |
1483 | /* clear the contents if its cached */ | |
1484 | if (db->db_state == DB_CACHED) { | |
1485 | ASSERT(db->db.db_data != NULL); | |
1486 | arc_release(db->db_buf, db); | |
1487 | bzero(db->db.db_data, db->db.db_size); | |
1488 | arc_buf_freeze(db->db_buf); | |
1489 | } | |
1490 | ||
1491 | mutex_exit(&db->db_mtx); | |
1492 | } | |
8951cb8d | 1493 | |
8951cb8d | 1494 | kmem_free(db_search, sizeof (dmu_buf_impl_t)); |
34dc7c2f BB |
1495 | mutex_exit(&dn->dn_dbufs_mtx); |
1496 | } | |
1497 | ||
34dc7c2f BB |
1498 | void |
1499 | dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx) | |
1500 | { | |
1501 | arc_buf_t *buf, *obuf; | |
1502 | int osize = db->db.db_size; | |
1503 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); | |
572e2857 | 1504 | dnode_t *dn; |
34dc7c2f | 1505 | |
428870ff | 1506 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f | 1507 | |
572e2857 BB |
1508 | DB_DNODE_ENTER(db); |
1509 | dn = DB_DNODE(db); | |
1510 | ||
34dc7c2f | 1511 | /* XXX does *this* func really need the lock? */ |
572e2857 | 1512 | ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); |
34dc7c2f BB |
1513 | |
1514 | /* | |
b0bc7a84 | 1515 | * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held |
34dc7c2f BB |
1516 | * is OK, because there can be no other references to the db |
1517 | * when we are changing its size, so no concurrent DB_FILL can | |
1518 | * be happening. | |
1519 | */ | |
1520 | /* | |
1521 | * XXX we should be doing a dbuf_read, checking the return | |
1522 | * value and returning that up to our callers | |
1523 | */ | |
b0bc7a84 | 1524 | dmu_buf_will_dirty(&db->db, tx); |
34dc7c2f BB |
1525 | |
1526 | /* create the data buffer for the new block */ | |
2aa34383 | 1527 | buf = arc_alloc_buf(dn->dn_objset->os_spa, db, type, size); |
34dc7c2f BB |
1528 | |
1529 | /* copy old block data to the new block */ | |
1530 | obuf = db->db_buf; | |
1531 | bcopy(obuf->b_data, buf->b_data, MIN(osize, size)); | |
1532 | /* zero the remainder */ | |
1533 | if (size > osize) | |
1534 | bzero((uint8_t *)buf->b_data + osize, size - osize); | |
1535 | ||
1536 | mutex_enter(&db->db_mtx); | |
1537 | dbuf_set_data(db, buf); | |
d3c2ae1c | 1538 | arc_buf_destroy(obuf, db); |
34dc7c2f BB |
1539 | db->db.db_size = size; |
1540 | ||
1541 | if (db->db_level == 0) { | |
1542 | ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); | |
1543 | db->db_last_dirty->dt.dl.dr_data = buf; | |
1544 | } | |
1545 | mutex_exit(&db->db_mtx); | |
1546 | ||
3ec3bc21 | 1547 | dmu_objset_willuse_space(dn->dn_objset, size - osize, tx); |
572e2857 | 1548 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
1549 | } |
1550 | ||
428870ff BB |
1551 | void |
1552 | dbuf_release_bp(dmu_buf_impl_t *db) | |
1553 | { | |
b0bc7a84 | 1554 | ASSERTV(objset_t *os = db->db_objset); |
428870ff BB |
1555 | |
1556 | ASSERT(dsl_pool_sync_context(dmu_objset_pool(os))); | |
1557 | ASSERT(arc_released(os->os_phys_buf) || | |
1558 | list_link_active(&os->os_dsl_dataset->ds_synced_link)); | |
1559 | ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf)); | |
1560 | ||
294f6806 | 1561 | (void) arc_release(db->db_buf, db); |
428870ff BB |
1562 | } |
1563 | ||
5a28a973 MA |
1564 | /* |
1565 | * We already have a dirty record for this TXG, and we are being | |
1566 | * dirtied again. | |
1567 | */ | |
1568 | static void | |
1569 | dbuf_redirty(dbuf_dirty_record_t *dr) | |
1570 | { | |
1571 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
1572 | ||
1573 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
1574 | ||
1575 | if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) { | |
1576 | /* | |
1577 | * If this buffer has already been written out, | |
1578 | * we now need to reset its state. | |
1579 | */ | |
1580 | dbuf_unoverride(dr); | |
1581 | if (db->db.db_object != DMU_META_DNODE_OBJECT && | |
1582 | db->db_state != DB_NOFILL) { | |
1583 | /* Already released on initial dirty, so just thaw. */ | |
1584 | ASSERT(arc_released(db->db_buf)); | |
1585 | arc_buf_thaw(db->db_buf); | |
1586 | } | |
1587 | } | |
1588 | } | |
1589 | ||
34dc7c2f BB |
1590 | dbuf_dirty_record_t * |
1591 | dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) | |
1592 | { | |
572e2857 BB |
1593 | dnode_t *dn; |
1594 | objset_t *os; | |
34dc7c2f BB |
1595 | dbuf_dirty_record_t **drp, *dr; |
1596 | int drop_struct_lock = FALSE; | |
1597 | int txgoff = tx->tx_txg & TXG_MASK; | |
1598 | ||
1599 | ASSERT(tx->tx_txg != 0); | |
1600 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
1601 | DMU_TX_DIRTY_BUF(tx, db); | |
1602 | ||
572e2857 BB |
1603 | DB_DNODE_ENTER(db); |
1604 | dn = DB_DNODE(db); | |
34dc7c2f BB |
1605 | /* |
1606 | * Shouldn't dirty a regular buffer in syncing context. Private | |
1607 | * objects may be dirtied in syncing context, but only if they | |
1608 | * were already pre-dirtied in open context. | |
34dc7c2f | 1609 | */ |
cc9bb3e5 GM |
1610 | #ifdef DEBUG |
1611 | if (dn->dn_objset->os_dsl_dataset != NULL) { | |
1612 | rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, | |
1613 | RW_READER, FTAG); | |
1614 | } | |
34dc7c2f BB |
1615 | ASSERT(!dmu_tx_is_syncing(tx) || |
1616 | BP_IS_HOLE(dn->dn_objset->os_rootbp) || | |
9babb374 BB |
1617 | DMU_OBJECT_IS_SPECIAL(dn->dn_object) || |
1618 | dn->dn_objset->os_dsl_dataset == NULL); | |
cc9bb3e5 GM |
1619 | if (dn->dn_objset->os_dsl_dataset != NULL) |
1620 | rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, FTAG); | |
1621 | #endif | |
34dc7c2f BB |
1622 | /* |
1623 | * We make this assert for private objects as well, but after we | |
1624 | * check if we're already dirty. They are allowed to re-dirty | |
1625 | * in syncing context. | |
1626 | */ | |
1627 | ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || | |
1628 | dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == | |
1629 | (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); | |
1630 | ||
1631 | mutex_enter(&db->db_mtx); | |
1632 | /* | |
1633 | * XXX make this true for indirects too? The problem is that | |
1634 | * transactions created with dmu_tx_create_assigned() from | |
1635 | * syncing context don't bother holding ahead. | |
1636 | */ | |
1637 | ASSERT(db->db_level != 0 || | |
b128c09f BB |
1638 | db->db_state == DB_CACHED || db->db_state == DB_FILL || |
1639 | db->db_state == DB_NOFILL); | |
34dc7c2f BB |
1640 | |
1641 | mutex_enter(&dn->dn_mtx); | |
1642 | /* | |
1643 | * Don't set dirtyctx to SYNC if we're just modifying this as we | |
1644 | * initialize the objset. | |
1645 | */ | |
cc9bb3e5 GM |
1646 | if (dn->dn_dirtyctx == DN_UNDIRTIED) { |
1647 | if (dn->dn_objset->os_dsl_dataset != NULL) { | |
1648 | rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, | |
1649 | RW_READER, FTAG); | |
1650 | } | |
1651 | if (!BP_IS_HOLE(dn->dn_objset->os_rootbp)) { | |
1652 | dn->dn_dirtyctx = (dmu_tx_is_syncing(tx) ? | |
1653 | DN_DIRTY_SYNC : DN_DIRTY_OPEN); | |
1654 | ASSERT(dn->dn_dirtyctx_firstset == NULL); | |
1655 | dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP); | |
1656 | } | |
1657 | if (dn->dn_objset->os_dsl_dataset != NULL) { | |
1658 | rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, | |
1659 | FTAG); | |
1660 | } | |
34dc7c2f BB |
1661 | } |
1662 | mutex_exit(&dn->dn_mtx); | |
1663 | ||
428870ff BB |
1664 | if (db->db_blkid == DMU_SPILL_BLKID) |
1665 | dn->dn_have_spill = B_TRUE; | |
1666 | ||
34dc7c2f BB |
1667 | /* |
1668 | * If this buffer is already dirty, we're done. | |
1669 | */ | |
1670 | drp = &db->db_last_dirty; | |
1671 | ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg || | |
1672 | db->db.db_object == DMU_META_DNODE_OBJECT); | |
1673 | while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg) | |
1674 | drp = &dr->dr_next; | |
1675 | if (dr && dr->dr_txg == tx->tx_txg) { | |
572e2857 BB |
1676 | DB_DNODE_EXIT(db); |
1677 | ||
5a28a973 | 1678 | dbuf_redirty(dr); |
34dc7c2f BB |
1679 | mutex_exit(&db->db_mtx); |
1680 | return (dr); | |
1681 | } | |
1682 | ||
1683 | /* | |
1684 | * Only valid if not already dirty. | |
1685 | */ | |
9babb374 BB |
1686 | ASSERT(dn->dn_object == 0 || |
1687 | dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == | |
34dc7c2f BB |
1688 | (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); |
1689 | ||
1690 | ASSERT3U(dn->dn_nlevels, >, db->db_level); | |
34dc7c2f BB |
1691 | |
1692 | /* | |
1693 | * We should only be dirtying in syncing context if it's the | |
9babb374 BB |
1694 | * mos or we're initializing the os or it's a special object. |
1695 | * However, we are allowed to dirty in syncing context provided | |
1696 | * we already dirtied it in open context. Hence we must make | |
1697 | * this assertion only if we're not already dirty. | |
34dc7c2f | 1698 | */ |
572e2857 | 1699 | os = dn->dn_objset; |
3b7f360c | 1700 | VERIFY3U(tx->tx_txg, <=, spa_final_dirty_txg(os->os_spa)); |
cc9bb3e5 GM |
1701 | #ifdef DEBUG |
1702 | if (dn->dn_objset->os_dsl_dataset != NULL) | |
1703 | rrw_enter(&os->os_dsl_dataset->ds_bp_rwlock, RW_READER, FTAG); | |
9babb374 BB |
1704 | ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) || |
1705 | os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp)); | |
cc9bb3e5 GM |
1706 | if (dn->dn_objset->os_dsl_dataset != NULL) |
1707 | rrw_exit(&os->os_dsl_dataset->ds_bp_rwlock, FTAG); | |
1708 | #endif | |
34dc7c2f BB |
1709 | ASSERT(db->db.db_size != 0); |
1710 | ||
1711 | dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); | |
1712 | ||
428870ff | 1713 | if (db->db_blkid != DMU_BONUS_BLKID) { |
3ec3bc21 | 1714 | dmu_objset_willuse_space(os, db->db.db_size, tx); |
34dc7c2f BB |
1715 | } |
1716 | ||
1717 | /* | |
1718 | * If this buffer is dirty in an old transaction group we need | |
1719 | * to make a copy of it so that the changes we make in this | |
1720 | * transaction group won't leak out when we sync the older txg. | |
1721 | */ | |
79c76d5b | 1722 | dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP); |
98f72a53 | 1723 | list_link_init(&dr->dr_dirty_node); |
34dc7c2f BB |
1724 | if (db->db_level == 0) { |
1725 | void *data_old = db->db_buf; | |
1726 | ||
b128c09f | 1727 | if (db->db_state != DB_NOFILL) { |
428870ff | 1728 | if (db->db_blkid == DMU_BONUS_BLKID) { |
b128c09f BB |
1729 | dbuf_fix_old_data(db, tx->tx_txg); |
1730 | data_old = db->db.db_data; | |
1731 | } else if (db->db.db_object != DMU_META_DNODE_OBJECT) { | |
1732 | /* | |
1733 | * Release the data buffer from the cache so | |
1734 | * that we can modify it without impacting | |
1735 | * possible other users of this cached data | |
1736 | * block. Note that indirect blocks and | |
1737 | * private objects are not released until the | |
1738 | * syncing state (since they are only modified | |
1739 | * then). | |
1740 | */ | |
1741 | arc_release(db->db_buf, db); | |
1742 | dbuf_fix_old_data(db, tx->tx_txg); | |
1743 | data_old = db->db_buf; | |
1744 | } | |
1745 | ASSERT(data_old != NULL); | |
34dc7c2f | 1746 | } |
34dc7c2f BB |
1747 | dr->dt.dl.dr_data = data_old; |
1748 | } else { | |
448d7aaa | 1749 | mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_NOLOCKDEP, NULL); |
34dc7c2f BB |
1750 | list_create(&dr->dt.di.dr_children, |
1751 | sizeof (dbuf_dirty_record_t), | |
1752 | offsetof(dbuf_dirty_record_t, dr_dirty_node)); | |
1753 | } | |
e8b96c60 MA |
1754 | if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL) |
1755 | dr->dr_accounted = db->db.db_size; | |
34dc7c2f BB |
1756 | dr->dr_dbuf = db; |
1757 | dr->dr_txg = tx->tx_txg; | |
1758 | dr->dr_next = *drp; | |
1759 | *drp = dr; | |
1760 | ||
1761 | /* | |
1762 | * We could have been freed_in_flight between the dbuf_noread | |
1763 | * and dbuf_dirty. We win, as though the dbuf_noread() had | |
1764 | * happened after the free. | |
1765 | */ | |
428870ff BB |
1766 | if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && |
1767 | db->db_blkid != DMU_SPILL_BLKID) { | |
34dc7c2f | 1768 | mutex_enter(&dn->dn_mtx); |
9bd274dd MA |
1769 | if (dn->dn_free_ranges[txgoff] != NULL) { |
1770 | range_tree_clear(dn->dn_free_ranges[txgoff], | |
1771 | db->db_blkid, 1); | |
1772 | } | |
34dc7c2f BB |
1773 | mutex_exit(&dn->dn_mtx); |
1774 | db->db_freed_in_flight = FALSE; | |
1775 | } | |
1776 | ||
1777 | /* | |
1778 | * This buffer is now part of this txg | |
1779 | */ | |
1780 | dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg); | |
1781 | db->db_dirtycnt += 1; | |
1782 | ASSERT3U(db->db_dirtycnt, <=, 3); | |
1783 | ||
1784 | mutex_exit(&db->db_mtx); | |
1785 | ||
428870ff BB |
1786 | if (db->db_blkid == DMU_BONUS_BLKID || |
1787 | db->db_blkid == DMU_SPILL_BLKID) { | |
34dc7c2f BB |
1788 | mutex_enter(&dn->dn_mtx); |
1789 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
1790 | list_insert_tail(&dn->dn_dirty_records[txgoff], dr); | |
1791 | mutex_exit(&dn->dn_mtx); | |
1792 | dnode_setdirty(dn, tx); | |
572e2857 | 1793 | DB_DNODE_EXIT(db); |
34dc7c2f | 1794 | return (dr); |
98ace739 MA |
1795 | } |
1796 | ||
1797 | /* | |
1798 | * The dn_struct_rwlock prevents db_blkptr from changing | |
1799 | * due to a write from syncing context completing | |
1800 | * while we are running, so we want to acquire it before | |
1801 | * looking at db_blkptr. | |
1802 | */ | |
1803 | if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) { | |
1804 | rw_enter(&dn->dn_struct_rwlock, RW_READER); | |
1805 | drop_struct_lock = TRUE; | |
1806 | } | |
1807 | ||
2ade4a99 MA |
1808 | /* |
1809 | * We need to hold the dn_struct_rwlock to make this assertion, | |
1810 | * because it protects dn_phys / dn_next_nlevels from changing. | |
1811 | */ | |
1812 | ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) || | |
1813 | dn->dn_phys->dn_nlevels > db->db_level || | |
1814 | dn->dn_next_nlevels[txgoff] > db->db_level || | |
1815 | dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level || | |
1816 | dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level); | |
1817 | ||
3ec3bc21 BB |
1818 | /* |
1819 | * If we are overwriting a dedup BP, then unless it is snapshotted, | |
1820 | * when we get to syncing context we will need to decrement its | |
1821 | * refcount in the DDT. Prefetch the relevant DDT block so that | |
1822 | * syncing context won't have to wait for the i/o. | |
1823 | */ | |
1824 | ddt_prefetch(os->os_spa, db->db_blkptr); | |
34dc7c2f | 1825 | |
b128c09f BB |
1826 | if (db->db_level == 0) { |
1827 | dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock); | |
1828 | ASSERT(dn->dn_maxblkid >= db->db_blkid); | |
1829 | } | |
1830 | ||
34dc7c2f BB |
1831 | if (db->db_level+1 < dn->dn_nlevels) { |
1832 | dmu_buf_impl_t *parent = db->db_parent; | |
1833 | dbuf_dirty_record_t *di; | |
1834 | int parent_held = FALSE; | |
1835 | ||
1836 | if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) { | |
1837 | int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; | |
1838 | ||
1839 | parent = dbuf_hold_level(dn, db->db_level+1, | |
1840 | db->db_blkid >> epbs, FTAG); | |
428870ff | 1841 | ASSERT(parent != NULL); |
34dc7c2f BB |
1842 | parent_held = TRUE; |
1843 | } | |
1844 | if (drop_struct_lock) | |
1845 | rw_exit(&dn->dn_struct_rwlock); | |
1846 | ASSERT3U(db->db_level+1, ==, parent->db_level); | |
1847 | di = dbuf_dirty(parent, tx); | |
1848 | if (parent_held) | |
1849 | dbuf_rele(parent, FTAG); | |
1850 | ||
1851 | mutex_enter(&db->db_mtx); | |
e8b96c60 MA |
1852 | /* |
1853 | * Since we've dropped the mutex, it's possible that | |
1854 | * dbuf_undirty() might have changed this out from under us. | |
1855 | */ | |
34dc7c2f BB |
1856 | if (db->db_last_dirty == dr || |
1857 | dn->dn_object == DMU_META_DNODE_OBJECT) { | |
1858 | mutex_enter(&di->dt.di.dr_mtx); | |
1859 | ASSERT3U(di->dr_txg, ==, tx->tx_txg); | |
1860 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
1861 | list_insert_tail(&di->dt.di.dr_children, dr); | |
1862 | mutex_exit(&di->dt.di.dr_mtx); | |
1863 | dr->dr_parent = di; | |
1864 | } | |
1865 | mutex_exit(&db->db_mtx); | |
1866 | } else { | |
1867 | ASSERT(db->db_level+1 == dn->dn_nlevels); | |
1868 | ASSERT(db->db_blkid < dn->dn_nblkptr); | |
572e2857 | 1869 | ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf); |
34dc7c2f BB |
1870 | mutex_enter(&dn->dn_mtx); |
1871 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
1872 | list_insert_tail(&dn->dn_dirty_records[txgoff], dr); | |
1873 | mutex_exit(&dn->dn_mtx); | |
1874 | if (drop_struct_lock) | |
1875 | rw_exit(&dn->dn_struct_rwlock); | |
1876 | } | |
1877 | ||
1878 | dnode_setdirty(dn, tx); | |
572e2857 | 1879 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
1880 | return (dr); |
1881 | } | |
1882 | ||
13fe0198 | 1883 | /* |
e49f1e20 WA |
1884 | * Undirty a buffer in the transaction group referenced by the given |
1885 | * transaction. Return whether this evicted the dbuf. | |
13fe0198 MA |
1886 | */ |
1887 | static boolean_t | |
34dc7c2f BB |
1888 | dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx) |
1889 | { | |
572e2857 | 1890 | dnode_t *dn; |
34dc7c2f BB |
1891 | uint64_t txg = tx->tx_txg; |
1892 | dbuf_dirty_record_t *dr, **drp; | |
1893 | ||
1894 | ASSERT(txg != 0); | |
4bda3bd0 MA |
1895 | |
1896 | /* | |
1897 | * Due to our use of dn_nlevels below, this can only be called | |
1898 | * in open context, unless we are operating on the MOS. | |
1899 | * From syncing context, dn_nlevels may be different from the | |
1900 | * dn_nlevels used when dbuf was dirtied. | |
1901 | */ | |
1902 | ASSERT(db->db_objset == | |
1903 | dmu_objset_pool(db->db_objset)->dp_meta_objset || | |
1904 | txg != spa_syncing_txg(dmu_objset_spa(db->db_objset))); | |
428870ff | 1905 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
13fe0198 MA |
1906 | ASSERT0(db->db_level); |
1907 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
34dc7c2f | 1908 | |
34dc7c2f BB |
1909 | /* |
1910 | * If this buffer is not dirty, we're done. | |
1911 | */ | |
1912 | for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next) | |
1913 | if (dr->dr_txg <= txg) | |
1914 | break; | |
13fe0198 MA |
1915 | if (dr == NULL || dr->dr_txg < txg) |
1916 | return (B_FALSE); | |
34dc7c2f | 1917 | ASSERT(dr->dr_txg == txg); |
428870ff | 1918 | ASSERT(dr->dr_dbuf == db); |
34dc7c2f | 1919 | |
572e2857 BB |
1920 | DB_DNODE_ENTER(db); |
1921 | dn = DB_DNODE(db); | |
1922 | ||
34dc7c2f BB |
1923 | dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); |
1924 | ||
1925 | ASSERT(db->db.db_size != 0); | |
1926 | ||
4bda3bd0 MA |
1927 | dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset), |
1928 | dr->dr_accounted, txg); | |
34dc7c2f BB |
1929 | |
1930 | *drp = dr->dr_next; | |
1931 | ||
ef3c1dea GR |
1932 | /* |
1933 | * Note that there are three places in dbuf_dirty() | |
1934 | * where this dirty record may be put on a list. | |
1935 | * Make sure to do a list_remove corresponding to | |
1936 | * every one of those list_insert calls. | |
1937 | */ | |
34dc7c2f BB |
1938 | if (dr->dr_parent) { |
1939 | mutex_enter(&dr->dr_parent->dt.di.dr_mtx); | |
1940 | list_remove(&dr->dr_parent->dt.di.dr_children, dr); | |
1941 | mutex_exit(&dr->dr_parent->dt.di.dr_mtx); | |
ef3c1dea | 1942 | } else if (db->db_blkid == DMU_SPILL_BLKID || |
4bda3bd0 | 1943 | db->db_level + 1 == dn->dn_nlevels) { |
b128c09f | 1944 | ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf); |
34dc7c2f BB |
1945 | mutex_enter(&dn->dn_mtx); |
1946 | list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr); | |
1947 | mutex_exit(&dn->dn_mtx); | |
1948 | } | |
572e2857 | 1949 | DB_DNODE_EXIT(db); |
34dc7c2f | 1950 | |
13fe0198 MA |
1951 | if (db->db_state != DB_NOFILL) { |
1952 | dbuf_unoverride(dr); | |
34dc7c2f | 1953 | |
34dc7c2f | 1954 | ASSERT(db->db_buf != NULL); |
13fe0198 MA |
1955 | ASSERT(dr->dt.dl.dr_data != NULL); |
1956 | if (dr->dt.dl.dr_data != db->db_buf) | |
d3c2ae1c | 1957 | arc_buf_destroy(dr->dt.dl.dr_data, db); |
34dc7c2f | 1958 | } |
58c4aa00 | 1959 | |
34dc7c2f BB |
1960 | kmem_free(dr, sizeof (dbuf_dirty_record_t)); |
1961 | ||
1962 | ASSERT(db->db_dirtycnt > 0); | |
1963 | db->db_dirtycnt -= 1; | |
1964 | ||
1965 | if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) { | |
d3c2ae1c GW |
1966 | ASSERT(db->db_state == DB_NOFILL || arc_released(db->db_buf)); |
1967 | dbuf_destroy(db); | |
13fe0198 | 1968 | return (B_TRUE); |
34dc7c2f BB |
1969 | } |
1970 | ||
13fe0198 | 1971 | return (B_FALSE); |
34dc7c2f BB |
1972 | } |
1973 | ||
b5256303 TC |
1974 | static void |
1975 | dmu_buf_will_dirty_impl(dmu_buf_t *db_fake, int flags, dmu_tx_t *tx) | |
34dc7c2f | 1976 | { |
b0bc7a84 | 1977 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; |
34dc7c2f BB |
1978 | |
1979 | ASSERT(tx->tx_txg != 0); | |
1980 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
1981 | ||
5a28a973 MA |
1982 | /* |
1983 | * Quick check for dirtyness. For already dirty blocks, this | |
1984 | * reduces runtime of this function by >90%, and overall performance | |
1985 | * by 50% for some workloads (e.g. file deletion with indirect blocks | |
1986 | * cached). | |
1987 | */ | |
1988 | mutex_enter(&db->db_mtx); | |
1989 | ||
1c27024e | 1990 | dbuf_dirty_record_t *dr; |
5a28a973 MA |
1991 | for (dr = db->db_last_dirty; |
1992 | dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) { | |
1993 | /* | |
1994 | * It's possible that it is already dirty but not cached, | |
1995 | * because there are some calls to dbuf_dirty() that don't | |
1996 | * go through dmu_buf_will_dirty(). | |
1997 | */ | |
1998 | if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) { | |
1999 | /* This dbuf is already dirty and cached. */ | |
2000 | dbuf_redirty(dr); | |
2001 | mutex_exit(&db->db_mtx); | |
2002 | return; | |
2003 | } | |
2004 | } | |
2005 | mutex_exit(&db->db_mtx); | |
2006 | ||
572e2857 BB |
2007 | DB_DNODE_ENTER(db); |
2008 | if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock)) | |
b5256303 | 2009 | flags |= DB_RF_HAVESTRUCT; |
572e2857 | 2010 | DB_DNODE_EXIT(db); |
b5256303 | 2011 | (void) dbuf_read(db, NULL, flags); |
34dc7c2f BB |
2012 | (void) dbuf_dirty(db, tx); |
2013 | } | |
2014 | ||
b5256303 TC |
2015 | void |
2016 | dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx) | |
2017 | { | |
2018 | dmu_buf_will_dirty_impl(db_fake, | |
2019 | DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH, tx); | |
2020 | } | |
2021 | ||
b128c09f BB |
2022 | void |
2023 | dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) | |
2024 | { | |
2025 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
2026 | ||
2027 | db->db_state = DB_NOFILL; | |
2028 | ||
2029 | dmu_buf_will_fill(db_fake, tx); | |
2030 | } | |
2031 | ||
34dc7c2f BB |
2032 | void |
2033 | dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) | |
2034 | { | |
2035 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
2036 | ||
428870ff | 2037 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f BB |
2038 | ASSERT(tx->tx_txg != 0); |
2039 | ASSERT(db->db_level == 0); | |
2040 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
2041 | ||
2042 | ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT || | |
2043 | dmu_tx_private_ok(tx)); | |
2044 | ||
2045 | dbuf_noread(db); | |
2046 | (void) dbuf_dirty(db, tx); | |
2047 | } | |
2048 | ||
b5256303 TC |
2049 | /* |
2050 | * This function is effectively the same as dmu_buf_will_dirty(), but | |
2051 | * indicates the caller expects raw encrypted data in the db. It will | |
2052 | * also set the raw flag on the created dirty record. | |
2053 | */ | |
2054 | void | |
2055 | dmu_buf_will_change_crypt_params(dmu_buf_t *db_fake, dmu_tx_t *tx) | |
2056 | { | |
2057 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
2058 | dbuf_dirty_record_t *dr; | |
2059 | ||
2060 | dmu_buf_will_dirty_impl(db_fake, | |
2061 | DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_NO_DECRYPT, tx); | |
2062 | ||
2063 | dr = db->db_last_dirty; | |
2064 | while (dr != NULL && dr->dr_txg > tx->tx_txg) | |
2065 | dr = dr->dr_next; | |
2066 | ||
2067 | ASSERT3P(dr, !=, NULL); | |
2068 | ASSERT3U(dr->dr_txg, ==, tx->tx_txg); | |
2069 | dr->dt.dl.dr_raw = B_TRUE; | |
2070 | } | |
2071 | ||
34dc7c2f BB |
2072 | #pragma weak dmu_buf_fill_done = dbuf_fill_done |
2073 | /* ARGSUSED */ | |
2074 | void | |
2075 | dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx) | |
2076 | { | |
2077 | mutex_enter(&db->db_mtx); | |
2078 | DBUF_VERIFY(db); | |
2079 | ||
2080 | if (db->db_state == DB_FILL) { | |
2081 | if (db->db_level == 0 && db->db_freed_in_flight) { | |
428870ff | 2082 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f BB |
2083 | /* we were freed while filling */ |
2084 | /* XXX dbuf_undirty? */ | |
2085 | bzero(db->db.db_data, db->db.db_size); | |
2086 | db->db_freed_in_flight = FALSE; | |
2087 | } | |
2088 | db->db_state = DB_CACHED; | |
2089 | cv_broadcast(&db->db_changed); | |
2090 | } | |
2091 | mutex_exit(&db->db_mtx); | |
2092 | } | |
2093 | ||
9b67f605 MA |
2094 | void |
2095 | dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data, | |
2096 | bp_embedded_type_t etype, enum zio_compress comp, | |
2097 | int uncompressed_size, int compressed_size, int byteorder, | |
2098 | dmu_tx_t *tx) | |
2099 | { | |
2100 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; | |
2101 | struct dirty_leaf *dl; | |
2102 | dmu_object_type_t type; | |
2103 | ||
241b5415 MA |
2104 | if (etype == BP_EMBEDDED_TYPE_DATA) { |
2105 | ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset), | |
2106 | SPA_FEATURE_EMBEDDED_DATA)); | |
2107 | } | |
2108 | ||
9b67f605 MA |
2109 | DB_DNODE_ENTER(db); |
2110 | type = DB_DNODE(db)->dn_type; | |
2111 | DB_DNODE_EXIT(db); | |
2112 | ||
2113 | ASSERT0(db->db_level); | |
2114 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); | |
2115 | ||
2116 | dmu_buf_will_not_fill(dbuf, tx); | |
2117 | ||
2118 | ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); | |
2119 | dl = &db->db_last_dirty->dt.dl; | |
2120 | encode_embedded_bp_compressed(&dl->dr_overridden_by, | |
2121 | data, comp, uncompressed_size, compressed_size); | |
2122 | BPE_SET_ETYPE(&dl->dr_overridden_by, etype); | |
2123 | BP_SET_TYPE(&dl->dr_overridden_by, type); | |
2124 | BP_SET_LEVEL(&dl->dr_overridden_by, 0); | |
2125 | BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder); | |
2126 | ||
2127 | dl->dr_override_state = DR_OVERRIDDEN; | |
2128 | dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg; | |
2129 | } | |
2130 | ||
9babb374 BB |
2131 | /* |
2132 | * Directly assign a provided arc buf to a given dbuf if it's not referenced | |
2133 | * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf. | |
2134 | */ | |
2135 | void | |
2136 | dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx) | |
2137 | { | |
2138 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
428870ff | 2139 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
9babb374 | 2140 | ASSERT(db->db_level == 0); |
2aa34383 | 2141 | ASSERT3U(dbuf_is_metadata(db), ==, arc_is_metadata(buf)); |
9babb374 | 2142 | ASSERT(buf != NULL); |
2aa34383 | 2143 | ASSERT(arc_buf_lsize(buf) == db->db.db_size); |
9babb374 BB |
2144 | ASSERT(tx->tx_txg != 0); |
2145 | ||
2146 | arc_return_buf(buf, db); | |
2147 | ASSERT(arc_released(buf)); | |
2148 | ||
2149 | mutex_enter(&db->db_mtx); | |
2150 | ||
2151 | while (db->db_state == DB_READ || db->db_state == DB_FILL) | |
2152 | cv_wait(&db->db_changed, &db->db_mtx); | |
2153 | ||
2154 | ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED); | |
2155 | ||
2156 | if (db->db_state == DB_CACHED && | |
2157 | refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) { | |
440a3eb9 TC |
2158 | /* |
2159 | * In practice, we will never have a case where we have an | |
2160 | * encrypted arc buffer while additional holds exist on the | |
2161 | * dbuf. We don't handle this here so we simply assert that | |
2162 | * fact instead. | |
2163 | */ | |
2164 | ASSERT(!arc_is_encrypted(buf)); | |
9babb374 BB |
2165 | mutex_exit(&db->db_mtx); |
2166 | (void) dbuf_dirty(db, tx); | |
2167 | bcopy(buf->b_data, db->db.db_data, db->db.db_size); | |
d3c2ae1c | 2168 | arc_buf_destroy(buf, db); |
428870ff | 2169 | xuio_stat_wbuf_copied(); |
9babb374 BB |
2170 | return; |
2171 | } | |
2172 | ||
428870ff | 2173 | xuio_stat_wbuf_nocopy(); |
9babb374 BB |
2174 | if (db->db_state == DB_CACHED) { |
2175 | dbuf_dirty_record_t *dr = db->db_last_dirty; | |
2176 | ||
2177 | ASSERT(db->db_buf != NULL); | |
2178 | if (dr != NULL && dr->dr_txg == tx->tx_txg) { | |
2179 | ASSERT(dr->dt.dl.dr_data == db->db_buf); | |
440a3eb9 TC |
2180 | IMPLY(arc_is_encrypted(buf), dr->dt.dl.dr_raw); |
2181 | ||
9babb374 BB |
2182 | if (!arc_released(db->db_buf)) { |
2183 | ASSERT(dr->dt.dl.dr_override_state == | |
2184 | DR_OVERRIDDEN); | |
2185 | arc_release(db->db_buf, db); | |
2186 | } | |
2187 | dr->dt.dl.dr_data = buf; | |
d3c2ae1c | 2188 | arc_buf_destroy(db->db_buf, db); |
9babb374 BB |
2189 | } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) { |
2190 | arc_release(db->db_buf, db); | |
d3c2ae1c | 2191 | arc_buf_destroy(db->db_buf, db); |
9babb374 BB |
2192 | } |
2193 | db->db_buf = NULL; | |
2194 | } | |
2195 | ASSERT(db->db_buf == NULL); | |
2196 | dbuf_set_data(db, buf); | |
2197 | db->db_state = DB_FILL; | |
2198 | mutex_exit(&db->db_mtx); | |
2199 | (void) dbuf_dirty(db, tx); | |
b0bc7a84 | 2200 | dmu_buf_fill_done(&db->db, tx); |
9babb374 BB |
2201 | } |
2202 | ||
34dc7c2f | 2203 | void |
d3c2ae1c | 2204 | dbuf_destroy(dmu_buf_impl_t *db) |
34dc7c2f | 2205 | { |
572e2857 | 2206 | dnode_t *dn; |
34dc7c2f | 2207 | dmu_buf_impl_t *parent = db->db_parent; |
572e2857 | 2208 | dmu_buf_impl_t *dndb; |
34dc7c2f BB |
2209 | |
2210 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
2211 | ASSERT(refcount_is_zero(&db->db_holds)); | |
2212 | ||
d3c2ae1c GW |
2213 | if (db->db_buf != NULL) { |
2214 | arc_buf_destroy(db->db_buf, db); | |
2215 | db->db_buf = NULL; | |
2216 | } | |
34dc7c2f | 2217 | |
d3c2ae1c GW |
2218 | if (db->db_blkid == DMU_BONUS_BLKID) { |
2219 | int slots = DB_DNODE(db)->dn_num_slots; | |
2220 | int bonuslen = DN_SLOTS_TO_BONUSLEN(slots); | |
b5256303 TC |
2221 | if (db->db.db_data != NULL) { |
2222 | kmem_free(db->db.db_data, bonuslen); | |
2223 | arc_space_return(bonuslen, ARC_SPACE_BONUS); | |
2224 | db->db_state = DB_UNCACHED; | |
2225 | } | |
34dc7c2f BB |
2226 | } |
2227 | ||
d3c2ae1c GW |
2228 | dbuf_clear_data(db); |
2229 | ||
2230 | if (multilist_link_active(&db->db_cache_link)) { | |
64fc7762 | 2231 | multilist_remove(dbuf_cache, db); |
d3c2ae1c GW |
2232 | (void) refcount_remove_many(&dbuf_cache_size, |
2233 | db->db.db_size, db); | |
2234 | } | |
2235 | ||
b128c09f | 2236 | ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL); |
34dc7c2f BB |
2237 | ASSERT(db->db_data_pending == NULL); |
2238 | ||
2239 | db->db_state = DB_EVICTING; | |
2240 | db->db_blkptr = NULL; | |
2241 | ||
d3c2ae1c GW |
2242 | /* |
2243 | * Now that db_state is DB_EVICTING, nobody else can find this via | |
2244 | * the hash table. We can now drop db_mtx, which allows us to | |
2245 | * acquire the dn_dbufs_mtx. | |
2246 | */ | |
2247 | mutex_exit(&db->db_mtx); | |
2248 | ||
572e2857 BB |
2249 | DB_DNODE_ENTER(db); |
2250 | dn = DB_DNODE(db); | |
2251 | dndb = dn->dn_dbuf; | |
d3c2ae1c GW |
2252 | if (db->db_blkid != DMU_BONUS_BLKID) { |
2253 | boolean_t needlock = !MUTEX_HELD(&dn->dn_dbufs_mtx); | |
2254 | if (needlock) | |
2255 | mutex_enter(&dn->dn_dbufs_mtx); | |
8951cb8d | 2256 | avl_remove(&dn->dn_dbufs, db); |
73ad4a9f | 2257 | atomic_dec_32(&dn->dn_dbufs_count); |
572e2857 BB |
2258 | membar_producer(); |
2259 | DB_DNODE_EXIT(db); | |
d3c2ae1c GW |
2260 | if (needlock) |
2261 | mutex_exit(&dn->dn_dbufs_mtx); | |
572e2857 BB |
2262 | /* |
2263 | * Decrementing the dbuf count means that the hold corresponding | |
2264 | * to the removed dbuf is no longer discounted in dnode_move(), | |
2265 | * so the dnode cannot be moved until after we release the hold. | |
2266 | * The membar_producer() ensures visibility of the decremented | |
2267 | * value in dnode_move(), since DB_DNODE_EXIT doesn't actually | |
2268 | * release any lock. | |
2269 | */ | |
34dc7c2f | 2270 | dnode_rele(dn, db); |
572e2857 | 2271 | db->db_dnode_handle = NULL; |
d3c2ae1c GW |
2272 | |
2273 | dbuf_hash_remove(db); | |
572e2857 BB |
2274 | } else { |
2275 | DB_DNODE_EXIT(db); | |
34dc7c2f BB |
2276 | } |
2277 | ||
d3c2ae1c | 2278 | ASSERT(refcount_is_zero(&db->db_holds)); |
34dc7c2f | 2279 | |
d3c2ae1c GW |
2280 | db->db_parent = NULL; |
2281 | ||
2282 | ASSERT(db->db_buf == NULL); | |
2283 | ASSERT(db->db.db_data == NULL); | |
2284 | ASSERT(db->db_hash_next == NULL); | |
2285 | ASSERT(db->db_blkptr == NULL); | |
2286 | ASSERT(db->db_data_pending == NULL); | |
2287 | ASSERT(!multilist_link_active(&db->db_cache_link)); | |
2288 | ||
2289 | kmem_cache_free(dbuf_kmem_cache, db); | |
2290 | arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF); | |
34dc7c2f BB |
2291 | |
2292 | /* | |
572e2857 | 2293 | * If this dbuf is referenced from an indirect dbuf, |
34dc7c2f BB |
2294 | * decrement the ref count on the indirect dbuf. |
2295 | */ | |
2296 | if (parent && parent != dndb) | |
2297 | dbuf_rele(parent, db); | |
2298 | } | |
2299 | ||
fcff0f35 PD |
2300 | /* |
2301 | * Note: While bpp will always be updated if the function returns success, | |
2302 | * parentp will not be updated if the dnode does not have dn_dbuf filled in; | |
2303 | * this happens when the dnode is the meta-dnode, or a userused or groupused | |
2304 | * object. | |
2305 | */ | |
bf701a83 BB |
2306 | __attribute__((always_inline)) |
2307 | static inline int | |
34dc7c2f | 2308 | dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse, |
fc5bb51f | 2309 | dmu_buf_impl_t **parentp, blkptr_t **bpp, struct dbuf_hold_impl_data *dh) |
34dc7c2f | 2310 | { |
34dc7c2f BB |
2311 | *parentp = NULL; |
2312 | *bpp = NULL; | |
2313 | ||
428870ff BB |
2314 | ASSERT(blkid != DMU_BONUS_BLKID); |
2315 | ||
2316 | if (blkid == DMU_SPILL_BLKID) { | |
2317 | mutex_enter(&dn->dn_mtx); | |
2318 | if (dn->dn_have_spill && | |
2319 | (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) | |
50c957f7 | 2320 | *bpp = DN_SPILL_BLKPTR(dn->dn_phys); |
428870ff BB |
2321 | else |
2322 | *bpp = NULL; | |
2323 | dbuf_add_ref(dn->dn_dbuf, NULL); | |
2324 | *parentp = dn->dn_dbuf; | |
2325 | mutex_exit(&dn->dn_mtx); | |
2326 | return (0); | |
2327 | } | |
34dc7c2f | 2328 | |
1c27024e | 2329 | int nlevels = |
32d41fb7 | 2330 | (dn->dn_phys->dn_nlevels == 0) ? 1 : dn->dn_phys->dn_nlevels; |
1c27024e | 2331 | int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; |
34dc7c2f BB |
2332 | |
2333 | ASSERT3U(level * epbs, <, 64); | |
2334 | ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); | |
32d41fb7 PD |
2335 | /* |
2336 | * This assertion shouldn't trip as long as the max indirect block size | |
2337 | * is less than 1M. The reason for this is that up to that point, | |
2338 | * the number of levels required to address an entire object with blocks | |
2339 | * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In | |
2340 | * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55 | |
2341 | * (i.e. we can address the entire object), objects will all use at most | |
2342 | * N-1 levels and the assertion won't overflow. However, once epbs is | |
2343 | * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be | |
2344 | * enough to address an entire object, so objects will have 5 levels, | |
2345 | * but then this assertion will overflow. | |
2346 | * | |
2347 | * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we | |
2348 | * need to redo this logic to handle overflows. | |
2349 | */ | |
2350 | ASSERT(level >= nlevels || | |
2351 | ((nlevels - level - 1) * epbs) + | |
2352 | highbit64(dn->dn_phys->dn_nblkptr) <= 64); | |
34dc7c2f | 2353 | if (level >= nlevels || |
32d41fb7 PD |
2354 | blkid >= ((uint64_t)dn->dn_phys->dn_nblkptr << |
2355 | ((nlevels - level - 1) * epbs)) || | |
2356 | (fail_sparse && | |
2357 | blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) { | |
34dc7c2f | 2358 | /* the buffer has no parent yet */ |
2e528b49 | 2359 | return (SET_ERROR(ENOENT)); |
34dc7c2f BB |
2360 | } else if (level < nlevels-1) { |
2361 | /* this block is referenced from an indirect block */ | |
fc5bb51f BB |
2362 | int err; |
2363 | if (dh == NULL) { | |
fcff0f35 PD |
2364 | err = dbuf_hold_impl(dn, level+1, |
2365 | blkid >> epbs, fail_sparse, FALSE, NULL, parentp); | |
d1d7e268 | 2366 | } else { |
fc5bb51f | 2367 | __dbuf_hold_impl_init(dh + 1, dn, dh->dh_level + 1, |
fcff0f35 PD |
2368 | blkid >> epbs, fail_sparse, FALSE, NULL, |
2369 | parentp, dh->dh_depth + 1); | |
fc5bb51f BB |
2370 | err = __dbuf_hold_impl(dh + 1); |
2371 | } | |
34dc7c2f BB |
2372 | if (err) |
2373 | return (err); | |
2374 | err = dbuf_read(*parentp, NULL, | |
2375 | (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL)); | |
2376 | if (err) { | |
2377 | dbuf_rele(*parentp, NULL); | |
2378 | *parentp = NULL; | |
2379 | return (err); | |
2380 | } | |
2381 | *bpp = ((blkptr_t *)(*parentp)->db.db_data) + | |
2382 | (blkid & ((1ULL << epbs) - 1)); | |
32d41fb7 PD |
2383 | if (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs))) |
2384 | ASSERT(BP_IS_HOLE(*bpp)); | |
34dc7c2f BB |
2385 | return (0); |
2386 | } else { | |
2387 | /* the block is referenced from the dnode */ | |
2388 | ASSERT3U(level, ==, nlevels-1); | |
2389 | ASSERT(dn->dn_phys->dn_nblkptr == 0 || | |
2390 | blkid < dn->dn_phys->dn_nblkptr); | |
2391 | if (dn->dn_dbuf) { | |
2392 | dbuf_add_ref(dn->dn_dbuf, NULL); | |
2393 | *parentp = dn->dn_dbuf; | |
2394 | } | |
2395 | *bpp = &dn->dn_phys->dn_blkptr[blkid]; | |
2396 | return (0); | |
2397 | } | |
2398 | } | |
2399 | ||
2400 | static dmu_buf_impl_t * | |
2401 | dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid, | |
2402 | dmu_buf_impl_t *parent, blkptr_t *blkptr) | |
2403 | { | |
428870ff | 2404 | objset_t *os = dn->dn_objset; |
34dc7c2f BB |
2405 | dmu_buf_impl_t *db, *odb; |
2406 | ||
2407 | ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); | |
2408 | ASSERT(dn->dn_type != DMU_OT_NONE); | |
2409 | ||
d3c2ae1c | 2410 | db = kmem_cache_alloc(dbuf_kmem_cache, KM_SLEEP); |
34dc7c2f BB |
2411 | |
2412 | db->db_objset = os; | |
2413 | db->db.db_object = dn->dn_object; | |
2414 | db->db_level = level; | |
2415 | db->db_blkid = blkid; | |
2416 | db->db_last_dirty = NULL; | |
2417 | db->db_dirtycnt = 0; | |
572e2857 | 2418 | db->db_dnode_handle = dn->dn_handle; |
34dc7c2f BB |
2419 | db->db_parent = parent; |
2420 | db->db_blkptr = blkptr; | |
2421 | ||
0c66c32d | 2422 | db->db_user = NULL; |
bc4501f7 JG |
2423 | db->db_user_immediate_evict = FALSE; |
2424 | db->db_freed_in_flight = FALSE; | |
2425 | db->db_pending_evict = FALSE; | |
34dc7c2f | 2426 | |
428870ff | 2427 | if (blkid == DMU_BONUS_BLKID) { |
34dc7c2f | 2428 | ASSERT3P(parent, ==, dn->dn_dbuf); |
50c957f7 | 2429 | db->db.db_size = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) - |
34dc7c2f BB |
2430 | (dn->dn_nblkptr-1) * sizeof (blkptr_t); |
2431 | ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); | |
428870ff | 2432 | db->db.db_offset = DMU_BONUS_BLKID; |
34dc7c2f BB |
2433 | db->db_state = DB_UNCACHED; |
2434 | /* the bonus dbuf is not placed in the hash table */ | |
25458cbe | 2435 | arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF); |
34dc7c2f | 2436 | return (db); |
428870ff BB |
2437 | } else if (blkid == DMU_SPILL_BLKID) { |
2438 | db->db.db_size = (blkptr != NULL) ? | |
2439 | BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE; | |
2440 | db->db.db_offset = 0; | |
34dc7c2f BB |
2441 | } else { |
2442 | int blocksize = | |
e8b96c60 | 2443 | db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz; |
34dc7c2f BB |
2444 | db->db.db_size = blocksize; |
2445 | db->db.db_offset = db->db_blkid * blocksize; | |
2446 | } | |
2447 | ||
2448 | /* | |
2449 | * Hold the dn_dbufs_mtx while we get the new dbuf | |
2450 | * in the hash table *and* added to the dbufs list. | |
2451 | * This prevents a possible deadlock with someone | |
2452 | * trying to look up this dbuf before its added to the | |
2453 | * dn_dbufs list. | |
2454 | */ | |
2455 | mutex_enter(&dn->dn_dbufs_mtx); | |
2456 | db->db_state = DB_EVICTING; | |
2457 | if ((odb = dbuf_hash_insert(db)) != NULL) { | |
2458 | /* someone else inserted it first */ | |
d3c2ae1c | 2459 | kmem_cache_free(dbuf_kmem_cache, db); |
34dc7c2f BB |
2460 | mutex_exit(&dn->dn_dbufs_mtx); |
2461 | return (odb); | |
2462 | } | |
8951cb8d | 2463 | avl_add(&dn->dn_dbufs, db); |
9c9531cb | 2464 | |
34dc7c2f BB |
2465 | db->db_state = DB_UNCACHED; |
2466 | mutex_exit(&dn->dn_dbufs_mtx); | |
25458cbe | 2467 | arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF); |
34dc7c2f BB |
2468 | |
2469 | if (parent && parent != dn->dn_dbuf) | |
2470 | dbuf_add_ref(parent, db); | |
2471 | ||
2472 | ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || | |
2473 | refcount_count(&dn->dn_holds) > 0); | |
2474 | (void) refcount_add(&dn->dn_holds, db); | |
73ad4a9f | 2475 | atomic_inc_32(&dn->dn_dbufs_count); |
34dc7c2f BB |
2476 | |
2477 | dprintf_dbuf(db, "db=%p\n", db); | |
2478 | ||
2479 | return (db); | |
2480 | } | |
2481 | ||
fcff0f35 PD |
2482 | typedef struct dbuf_prefetch_arg { |
2483 | spa_t *dpa_spa; /* The spa to issue the prefetch in. */ | |
2484 | zbookmark_phys_t dpa_zb; /* The target block to prefetch. */ | |
2485 | int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */ | |
2486 | int dpa_curlevel; /* The current level that we're reading */ | |
d3c2ae1c | 2487 | dnode_t *dpa_dnode; /* The dnode associated with the prefetch */ |
fcff0f35 PD |
2488 | zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */ |
2489 | zio_t *dpa_zio; /* The parent zio_t for all prefetches. */ | |
2490 | arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */ | |
2491 | } dbuf_prefetch_arg_t; | |
2492 | ||
2493 | /* | |
2494 | * Actually issue the prefetch read for the block given. | |
2495 | */ | |
2496 | static void | |
2497 | dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp) | |
2498 | { | |
fcff0f35 PD |
2499 | if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) |
2500 | return; | |
2501 | ||
1c27024e DB |
2502 | arc_flags_t aflags = |
2503 | dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH; | |
fcff0f35 PD |
2504 | |
2505 | ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp)); | |
2506 | ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level); | |
2507 | ASSERT(dpa->dpa_zio != NULL); | |
2508 | (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL, | |
2509 | dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, | |
2510 | &aflags, &dpa->dpa_zb); | |
2511 | } | |
2512 | ||
2513 | /* | |
2514 | * Called when an indirect block above our prefetch target is read in. This | |
2515 | * will either read in the next indirect block down the tree or issue the actual | |
2516 | * prefetch if the next block down is our target. | |
2517 | */ | |
2518 | static void | |
d4a72f23 TC |
2519 | dbuf_prefetch_indirect_done(zio_t *zio, const zbookmark_phys_t *zb, |
2520 | const blkptr_t *iobp, arc_buf_t *abuf, void *private) | |
fcff0f35 PD |
2521 | { |
2522 | dbuf_prefetch_arg_t *dpa = private; | |
fcff0f35 PD |
2523 | |
2524 | ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel); | |
2525 | ASSERT3S(dpa->dpa_curlevel, >, 0); | |
d3c2ae1c GW |
2526 | |
2527 | /* | |
2528 | * The dpa_dnode is only valid if we are called with a NULL | |
2529 | * zio. This indicates that the arc_read() returned without | |
2530 | * first calling zio_read() to issue a physical read. Once | |
2531 | * a physical read is made the dpa_dnode must be invalidated | |
2532 | * as the locks guarding it may have been dropped. If the | |
2533 | * dpa_dnode is still valid, then we want to add it to the dbuf | |
2534 | * cache. To do so, we must hold the dbuf associated with the block | |
2535 | * we just prefetched, read its contents so that we associate it | |
2536 | * with an arc_buf_t, and then release it. | |
2537 | */ | |
fcff0f35 PD |
2538 | if (zio != NULL) { |
2539 | ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel); | |
b5256303 | 2540 | if (zio->io_flags & ZIO_FLAG_RAW_COMPRESS) { |
d3c2ae1c GW |
2541 | ASSERT3U(BP_GET_PSIZE(zio->io_bp), ==, zio->io_size); |
2542 | } else { | |
2543 | ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size); | |
2544 | } | |
fcff0f35 | 2545 | ASSERT3P(zio->io_spa, ==, dpa->dpa_spa); |
d3c2ae1c GW |
2546 | |
2547 | dpa->dpa_dnode = NULL; | |
2548 | } else if (dpa->dpa_dnode != NULL) { | |
2549 | uint64_t curblkid = dpa->dpa_zb.zb_blkid >> | |
2550 | (dpa->dpa_epbs * (dpa->dpa_curlevel - | |
2551 | dpa->dpa_zb.zb_level)); | |
2552 | dmu_buf_impl_t *db = dbuf_hold_level(dpa->dpa_dnode, | |
2553 | dpa->dpa_curlevel, curblkid, FTAG); | |
2554 | (void) dbuf_read(db, NULL, | |
2555 | DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_HAVESTRUCT); | |
2556 | dbuf_rele(db, FTAG); | |
fcff0f35 PD |
2557 | } |
2558 | ||
d4a72f23 TC |
2559 | if (abuf == NULL) { |
2560 | kmem_free(dpa, sizeof (*dpa)); | |
2561 | return; | |
2562 | } | |
fcff0f35 | 2563 | |
d4a72f23 | 2564 | dpa->dpa_curlevel--; |
1c27024e | 2565 | uint64_t nextblkid = dpa->dpa_zb.zb_blkid >> |
fcff0f35 | 2566 | (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level)); |
1c27024e | 2567 | blkptr_t *bp = ((blkptr_t *)abuf->b_data) + |
fcff0f35 | 2568 | P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs); |
d4a72f23 TC |
2569 | |
2570 | if (BP_IS_HOLE(bp)) { | |
fcff0f35 PD |
2571 | kmem_free(dpa, sizeof (*dpa)); |
2572 | } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) { | |
2573 | ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid); | |
2574 | dbuf_issue_final_prefetch(dpa, bp); | |
2575 | kmem_free(dpa, sizeof (*dpa)); | |
2576 | } else { | |
2577 | arc_flags_t iter_aflags = ARC_FLAG_NOWAIT; | |
2578 | zbookmark_phys_t zb; | |
2579 | ||
7c351e31 | 2580 | /* flag if L2ARC eligible, l2arc_noprefetch then decides */ |
2581 | if (dpa->dpa_aflags & ARC_FLAG_L2CACHE) | |
2582 | iter_aflags |= ARC_FLAG_L2CACHE; | |
2583 | ||
fcff0f35 PD |
2584 | ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp)); |
2585 | ||
2586 | SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset, | |
2587 | dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid); | |
2588 | ||
2589 | (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, | |
2590 | bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio, | |
2591 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, | |
2592 | &iter_aflags, &zb); | |
2593 | } | |
d3c2ae1c GW |
2594 | |
2595 | arc_buf_destroy(abuf, private); | |
fcff0f35 PD |
2596 | } |
2597 | ||
2598 | /* | |
2599 | * Issue prefetch reads for the given block on the given level. If the indirect | |
2600 | * blocks above that block are not in memory, we will read them in | |
2601 | * asynchronously. As a result, this call never blocks waiting for a read to | |
b5256303 TC |
2602 | * complete. Note that the prefetch might fail if the dataset is encrypted and |
2603 | * the encryption key is unmapped before the IO completes. | |
fcff0f35 | 2604 | */ |
34dc7c2f | 2605 | void |
fcff0f35 PD |
2606 | dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio, |
2607 | arc_flags_t aflags) | |
34dc7c2f | 2608 | { |
fcff0f35 PD |
2609 | blkptr_t bp; |
2610 | int epbs, nlevels, curlevel; | |
2611 | uint64_t curblkid; | |
34dc7c2f | 2612 | |
428870ff | 2613 | ASSERT(blkid != DMU_BONUS_BLKID); |
34dc7c2f BB |
2614 | ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); |
2615 | ||
7f60329a MA |
2616 | if (blkid > dn->dn_maxblkid) |
2617 | return; | |
2618 | ||
34dc7c2f BB |
2619 | if (dnode_block_freed(dn, blkid)) |
2620 | return; | |
2621 | ||
fcff0f35 PD |
2622 | /* |
2623 | * This dnode hasn't been written to disk yet, so there's nothing to | |
2624 | * prefetch. | |
2625 | */ | |
2626 | nlevels = dn->dn_phys->dn_nlevels; | |
2627 | if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0) | |
2628 | return; | |
2629 | ||
2630 | epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; | |
2631 | if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level)) | |
2632 | return; | |
2633 | ||
1c27024e | 2634 | dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object, |
fcff0f35 PD |
2635 | level, blkid); |
2636 | if (db != NULL) { | |
2637 | mutex_exit(&db->db_mtx); | |
572e2857 | 2638 | /* |
fcff0f35 PD |
2639 | * This dbuf already exists. It is either CACHED, or |
2640 | * (we assume) about to be read or filled. | |
572e2857 | 2641 | */ |
572e2857 | 2642 | return; |
34dc7c2f BB |
2643 | } |
2644 | ||
fcff0f35 PD |
2645 | /* |
2646 | * Find the closest ancestor (indirect block) of the target block | |
2647 | * that is present in the cache. In this indirect block, we will | |
2648 | * find the bp that is at curlevel, curblkid. | |
2649 | */ | |
2650 | curlevel = level; | |
2651 | curblkid = blkid; | |
2652 | while (curlevel < nlevels - 1) { | |
2653 | int parent_level = curlevel + 1; | |
2654 | uint64_t parent_blkid = curblkid >> epbs; | |
2655 | dmu_buf_impl_t *db; | |
2656 | ||
2657 | if (dbuf_hold_impl(dn, parent_level, parent_blkid, | |
2658 | FALSE, TRUE, FTAG, &db) == 0) { | |
2659 | blkptr_t *bpp = db->db_buf->b_data; | |
2660 | bp = bpp[P2PHASE(curblkid, 1 << epbs)]; | |
2661 | dbuf_rele(db, FTAG); | |
2662 | break; | |
2663 | } | |
428870ff | 2664 | |
fcff0f35 PD |
2665 | curlevel = parent_level; |
2666 | curblkid = parent_blkid; | |
2667 | } | |
34dc7c2f | 2668 | |
fcff0f35 PD |
2669 | if (curlevel == nlevels - 1) { |
2670 | /* No cached indirect blocks found. */ | |
2671 | ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr); | |
2672 | bp = dn->dn_phys->dn_blkptr[curblkid]; | |
34dc7c2f | 2673 | } |
fcff0f35 PD |
2674 | if (BP_IS_HOLE(&bp)) |
2675 | return; | |
2676 | ||
2677 | ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp)); | |
2678 | ||
1c27024e | 2679 | zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL, |
fcff0f35 PD |
2680 | ZIO_FLAG_CANFAIL); |
2681 | ||
1c27024e DB |
2682 | dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP); |
2683 | dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset; | |
fcff0f35 PD |
2684 | SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET, |
2685 | dn->dn_object, level, blkid); | |
2686 | dpa->dpa_curlevel = curlevel; | |
2687 | dpa->dpa_prio = prio; | |
2688 | dpa->dpa_aflags = aflags; | |
2689 | dpa->dpa_spa = dn->dn_objset->os_spa; | |
d3c2ae1c | 2690 | dpa->dpa_dnode = dn; |
fcff0f35 PD |
2691 | dpa->dpa_epbs = epbs; |
2692 | dpa->dpa_zio = pio; | |
2693 | ||
7c351e31 | 2694 | /* flag if L2ARC eligible, l2arc_noprefetch then decides */ |
2695 | if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level)) | |
2696 | dpa->dpa_aflags |= ARC_FLAG_L2CACHE; | |
2697 | ||
fcff0f35 PD |
2698 | /* |
2699 | * If we have the indirect just above us, no need to do the asynchronous | |
2700 | * prefetch chain; we'll just run the last step ourselves. If we're at | |
2701 | * a higher level, though, we want to issue the prefetches for all the | |
2702 | * indirect blocks asynchronously, so we can go on with whatever we were | |
2703 | * doing. | |
2704 | */ | |
2705 | if (curlevel == level) { | |
2706 | ASSERT3U(curblkid, ==, blkid); | |
2707 | dbuf_issue_final_prefetch(dpa, &bp); | |
2708 | kmem_free(dpa, sizeof (*dpa)); | |
2709 | } else { | |
2710 | arc_flags_t iter_aflags = ARC_FLAG_NOWAIT; | |
2711 | zbookmark_phys_t zb; | |
2712 | ||
7c351e31 | 2713 | /* flag if L2ARC eligible, l2arc_noprefetch then decides */ |
2714 | if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level)) | |
2715 | iter_aflags |= ARC_FLAG_L2CACHE; | |
2716 | ||
fcff0f35 PD |
2717 | SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET, |
2718 | dn->dn_object, curlevel, curblkid); | |
2719 | (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, | |
2720 | &bp, dbuf_prefetch_indirect_done, dpa, prio, | |
2721 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, | |
2722 | &iter_aflags, &zb); | |
2723 | } | |
2724 | /* | |
2725 | * We use pio here instead of dpa_zio since it's possible that | |
2726 | * dpa may have already been freed. | |
2727 | */ | |
2728 | zio_nowait(pio); | |
34dc7c2f BB |
2729 | } |
2730 | ||
d1d7e268 | 2731 | #define DBUF_HOLD_IMPL_MAX_DEPTH 20 |
fc5bb51f | 2732 | |
71a24c3c TC |
2733 | /* |
2734 | * Helper function for __dbuf_hold_impl() to copy a buffer. Handles | |
2735 | * the case of encrypted, compressed and uncompressed buffers by | |
2736 | * allocating the new buffer, respectively, with arc_alloc_raw_buf(), | |
2737 | * arc_alloc_compressed_buf() or arc_alloc_buf().* | |
2738 | * | |
2739 | * NOTE: Declared noinline to avoid stack bloat in __dbuf_hold_impl(). | |
2740 | */ | |
2741 | noinline static void | |
2742 | dbuf_hold_copy(struct dbuf_hold_impl_data *dh) | |
2743 | { | |
2744 | dnode_t *dn = dh->dh_dn; | |
2745 | dmu_buf_impl_t *db = dh->dh_db; | |
2746 | dbuf_dirty_record_t *dr = dh->dh_dr; | |
2747 | arc_buf_t *data = dr->dt.dl.dr_data; | |
2748 | ||
2749 | enum zio_compress compress_type = arc_get_compression(data); | |
2750 | ||
2751 | if (arc_is_encrypted(data)) { | |
2752 | boolean_t byteorder; | |
2753 | uint8_t salt[ZIO_DATA_SALT_LEN]; | |
2754 | uint8_t iv[ZIO_DATA_IV_LEN]; | |
2755 | uint8_t mac[ZIO_DATA_MAC_LEN]; | |
2756 | ||
2757 | arc_get_raw_params(data, &byteorder, salt, iv, mac); | |
2758 | dbuf_set_data(db, arc_alloc_raw_buf(dn->dn_objset->os_spa, db, | |
2759 | dmu_objset_id(dn->dn_objset), byteorder, salt, iv, mac, | |
2760 | dn->dn_type, arc_buf_size(data), arc_buf_lsize(data), | |
2761 | compress_type)); | |
2762 | } else if (compress_type != ZIO_COMPRESS_OFF) { | |
2763 | dbuf_set_data(db, arc_alloc_compressed_buf( | |
2764 | dn->dn_objset->os_spa, db, arc_buf_size(data), | |
2765 | arc_buf_lsize(data), compress_type)); | |
2766 | } else { | |
2767 | dbuf_set_data(db, arc_alloc_buf(dn->dn_objset->os_spa, db, | |
2768 | DBUF_GET_BUFC_TYPE(db), db->db.db_size)); | |
2769 | } | |
2770 | ||
2771 | bcopy(data->b_data, db->db.db_data, arc_buf_size(data)); | |
2772 | } | |
2773 | ||
34dc7c2f BB |
2774 | /* |
2775 | * Returns with db_holds incremented, and db_mtx not held. | |
2776 | * Note: dn_struct_rwlock must be held. | |
2777 | */ | |
fc5bb51f BB |
2778 | static int |
2779 | __dbuf_hold_impl(struct dbuf_hold_impl_data *dh) | |
34dc7c2f | 2780 | { |
fc5bb51f BB |
2781 | ASSERT3S(dh->dh_depth, <, DBUF_HOLD_IMPL_MAX_DEPTH); |
2782 | dh->dh_parent = NULL; | |
34dc7c2f | 2783 | |
fc5bb51f BB |
2784 | ASSERT(dh->dh_blkid != DMU_BONUS_BLKID); |
2785 | ASSERT(RW_LOCK_HELD(&dh->dh_dn->dn_struct_rwlock)); | |
2786 | ASSERT3U(dh->dh_dn->dn_nlevels, >, dh->dh_level); | |
34dc7c2f | 2787 | |
fc5bb51f | 2788 | *(dh->dh_dbp) = NULL; |
d3c2ae1c | 2789 | |
34dc7c2f | 2790 | /* dbuf_find() returns with db_mtx held */ |
6ebebace JG |
2791 | dh->dh_db = dbuf_find(dh->dh_dn->dn_objset, dh->dh_dn->dn_object, |
2792 | dh->dh_level, dh->dh_blkid); | |
fc5bb51f BB |
2793 | |
2794 | if (dh->dh_db == NULL) { | |
2795 | dh->dh_bp = NULL; | |
2796 | ||
fcff0f35 PD |
2797 | if (dh->dh_fail_uncached) |
2798 | return (SET_ERROR(ENOENT)); | |
2799 | ||
fc5bb51f BB |
2800 | ASSERT3P(dh->dh_parent, ==, NULL); |
2801 | dh->dh_err = dbuf_findbp(dh->dh_dn, dh->dh_level, dh->dh_blkid, | |
02730c33 | 2802 | dh->dh_fail_sparse, &dh->dh_parent, &dh->dh_bp, dh); |
fc5bb51f | 2803 | if (dh->dh_fail_sparse) { |
d1d7e268 MK |
2804 | if (dh->dh_err == 0 && |
2805 | dh->dh_bp && BP_IS_HOLE(dh->dh_bp)) | |
2e528b49 | 2806 | dh->dh_err = SET_ERROR(ENOENT); |
fc5bb51f BB |
2807 | if (dh->dh_err) { |
2808 | if (dh->dh_parent) | |
2809 | dbuf_rele(dh->dh_parent, NULL); | |
2810 | return (dh->dh_err); | |
34dc7c2f BB |
2811 | } |
2812 | } | |
fc5bb51f BB |
2813 | if (dh->dh_err && dh->dh_err != ENOENT) |
2814 | return (dh->dh_err); | |
2815 | dh->dh_db = dbuf_create(dh->dh_dn, dh->dh_level, dh->dh_blkid, | |
02730c33 | 2816 | dh->dh_parent, dh->dh_bp); |
34dc7c2f BB |
2817 | } |
2818 | ||
fcff0f35 PD |
2819 | if (dh->dh_fail_uncached && dh->dh_db->db_state != DB_CACHED) { |
2820 | mutex_exit(&dh->dh_db->db_mtx); | |
2821 | return (SET_ERROR(ENOENT)); | |
2822 | } | |
2823 | ||
d3c2ae1c | 2824 | if (dh->dh_db->db_buf != NULL) |
fc5bb51f | 2825 | ASSERT3P(dh->dh_db->db.db_data, ==, dh->dh_db->db_buf->b_data); |
34dc7c2f | 2826 | |
fc5bb51f | 2827 | ASSERT(dh->dh_db->db_buf == NULL || arc_referenced(dh->dh_db->db_buf)); |
34dc7c2f BB |
2828 | |
2829 | /* | |
2830 | * If this buffer is currently syncing out, and we are are | |
2831 | * still referencing it from db_data, we need to make a copy | |
2832 | * of it in case we decide we want to dirty it again in this txg. | |
2833 | */ | |
fc5bb51f BB |
2834 | if (dh->dh_db->db_level == 0 && |
2835 | dh->dh_db->db_blkid != DMU_BONUS_BLKID && | |
2836 | dh->dh_dn->dn_object != DMU_META_DNODE_OBJECT && | |
2837 | dh->dh_db->db_state == DB_CACHED && dh->dh_db->db_data_pending) { | |
2838 | dh->dh_dr = dh->dh_db->db_data_pending; | |
71a24c3c TC |
2839 | if (dh->dh_dr->dt.dl.dr_data == dh->dh_db->db_buf) |
2840 | dbuf_hold_copy(dh); | |
34dc7c2f BB |
2841 | } |
2842 | ||
d3c2ae1c GW |
2843 | if (multilist_link_active(&dh->dh_db->db_cache_link)) { |
2844 | ASSERT(refcount_is_zero(&dh->dh_db->db_holds)); | |
64fc7762 | 2845 | multilist_remove(dbuf_cache, dh->dh_db); |
d3c2ae1c GW |
2846 | (void) refcount_remove_many(&dbuf_cache_size, |
2847 | dh->dh_db->db.db_size, dh->dh_db); | |
2848 | } | |
fc5bb51f | 2849 | (void) refcount_add(&dh->dh_db->db_holds, dh->dh_tag); |
fc5bb51f BB |
2850 | DBUF_VERIFY(dh->dh_db); |
2851 | mutex_exit(&dh->dh_db->db_mtx); | |
34dc7c2f BB |
2852 | |
2853 | /* NOTE: we can't rele the parent until after we drop the db_mtx */ | |
fc5bb51f BB |
2854 | if (dh->dh_parent) |
2855 | dbuf_rele(dh->dh_parent, NULL); | |
34dc7c2f | 2856 | |
fc5bb51f BB |
2857 | ASSERT3P(DB_DNODE(dh->dh_db), ==, dh->dh_dn); |
2858 | ASSERT3U(dh->dh_db->db_blkid, ==, dh->dh_blkid); | |
2859 | ASSERT3U(dh->dh_db->db_level, ==, dh->dh_level); | |
2860 | *(dh->dh_dbp) = dh->dh_db; | |
34dc7c2f BB |
2861 | |
2862 | return (0); | |
2863 | } | |
2864 | ||
fc5bb51f BB |
2865 | /* |
2866 | * The following code preserves the recursive function dbuf_hold_impl() | |
2867 | * but moves the local variables AND function arguments to the heap to | |
2868 | * minimize the stack frame size. Enough space is initially allocated | |
2869 | * on the stack for 20 levels of recursion. | |
2870 | */ | |
2871 | int | |
fcff0f35 PD |
2872 | dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, |
2873 | boolean_t fail_sparse, boolean_t fail_uncached, | |
fc5bb51f BB |
2874 | void *tag, dmu_buf_impl_t **dbp) |
2875 | { | |
2876 | struct dbuf_hold_impl_data *dh; | |
2877 | int error; | |
2878 | ||
d9eea113 | 2879 | dh = kmem_alloc(sizeof (struct dbuf_hold_impl_data) * |
79c76d5b | 2880 | DBUF_HOLD_IMPL_MAX_DEPTH, KM_SLEEP); |
fcff0f35 | 2881 | __dbuf_hold_impl_init(dh, dn, level, blkid, fail_sparse, |
02730c33 | 2882 | fail_uncached, tag, dbp, 0); |
fc5bb51f BB |
2883 | |
2884 | error = __dbuf_hold_impl(dh); | |
2885 | ||
d1d7e268 | 2886 | kmem_free(dh, sizeof (struct dbuf_hold_impl_data) * |
fc5bb51f BB |
2887 | DBUF_HOLD_IMPL_MAX_DEPTH); |
2888 | ||
2889 | return (error); | |
2890 | } | |
2891 | ||
2892 | static void | |
2893 | __dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh, | |
fcff0f35 | 2894 | dnode_t *dn, uint8_t level, uint64_t blkid, |
4ea3f864 GM |
2895 | boolean_t fail_sparse, boolean_t fail_uncached, |
2896 | void *tag, dmu_buf_impl_t **dbp, int depth) | |
fc5bb51f BB |
2897 | { |
2898 | dh->dh_dn = dn; | |
2899 | dh->dh_level = level; | |
2900 | dh->dh_blkid = blkid; | |
fcff0f35 | 2901 | |
fc5bb51f | 2902 | dh->dh_fail_sparse = fail_sparse; |
fcff0f35 PD |
2903 | dh->dh_fail_uncached = fail_uncached; |
2904 | ||
fc5bb51f BB |
2905 | dh->dh_tag = tag; |
2906 | dh->dh_dbp = dbp; | |
d9eea113 MA |
2907 | |
2908 | dh->dh_db = NULL; | |
2909 | dh->dh_parent = NULL; | |
2910 | dh->dh_bp = NULL; | |
2911 | dh->dh_err = 0; | |
2912 | dh->dh_dr = NULL; | |
d9eea113 | 2913 | |
fc5bb51f BB |
2914 | dh->dh_depth = depth; |
2915 | } | |
2916 | ||
34dc7c2f BB |
2917 | dmu_buf_impl_t * |
2918 | dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag) | |
2919 | { | |
fcff0f35 | 2920 | return (dbuf_hold_level(dn, 0, blkid, tag)); |
34dc7c2f BB |
2921 | } |
2922 | ||
2923 | dmu_buf_impl_t * | |
2924 | dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag) | |
2925 | { | |
2926 | dmu_buf_impl_t *db; | |
fcff0f35 | 2927 | int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db); |
34dc7c2f BB |
2928 | return (err ? NULL : db); |
2929 | } | |
2930 | ||
2931 | void | |
2932 | dbuf_create_bonus(dnode_t *dn) | |
2933 | { | |
2934 | ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); | |
2935 | ||
2936 | ASSERT(dn->dn_bonus == NULL); | |
428870ff BB |
2937 | dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL); |
2938 | } | |
2939 | ||
2940 | int | |
2941 | dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx) | |
2942 | { | |
2943 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
572e2857 BB |
2944 | dnode_t *dn; |
2945 | ||
428870ff | 2946 | if (db->db_blkid != DMU_SPILL_BLKID) |
2e528b49 | 2947 | return (SET_ERROR(ENOTSUP)); |
428870ff BB |
2948 | if (blksz == 0) |
2949 | blksz = SPA_MINBLOCKSIZE; | |
f1512ee6 MA |
2950 | ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset))); |
2951 | blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE); | |
428870ff | 2952 | |
572e2857 BB |
2953 | DB_DNODE_ENTER(db); |
2954 | dn = DB_DNODE(db); | |
2955 | rw_enter(&dn->dn_struct_rwlock, RW_WRITER); | |
428870ff | 2956 | dbuf_new_size(db, blksz, tx); |
572e2857 BB |
2957 | rw_exit(&dn->dn_struct_rwlock); |
2958 | DB_DNODE_EXIT(db); | |
428870ff BB |
2959 | |
2960 | return (0); | |
2961 | } | |
2962 | ||
2963 | void | |
2964 | dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx) | |
2965 | { | |
2966 | dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx); | |
34dc7c2f BB |
2967 | } |
2968 | ||
2969 | #pragma weak dmu_buf_add_ref = dbuf_add_ref | |
2970 | void | |
2971 | dbuf_add_ref(dmu_buf_impl_t *db, void *tag) | |
2972 | { | |
d3c2ae1c GW |
2973 | int64_t holds = refcount_add(&db->db_holds, tag); |
2974 | VERIFY3S(holds, >, 1); | |
34dc7c2f BB |
2975 | } |
2976 | ||
6ebebace JG |
2977 | #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref |
2978 | boolean_t | |
2979 | dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid, | |
2980 | void *tag) | |
2981 | { | |
2982 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
2983 | dmu_buf_impl_t *found_db; | |
2984 | boolean_t result = B_FALSE; | |
2985 | ||
d617648c | 2986 | if (blkid == DMU_BONUS_BLKID) |
6ebebace JG |
2987 | found_db = dbuf_find_bonus(os, obj); |
2988 | else | |
2989 | found_db = dbuf_find(os, obj, 0, blkid); | |
2990 | ||
2991 | if (found_db != NULL) { | |
2992 | if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) { | |
2993 | (void) refcount_add(&db->db_holds, tag); | |
2994 | result = B_TRUE; | |
2995 | } | |
d617648c | 2996 | mutex_exit(&found_db->db_mtx); |
6ebebace JG |
2997 | } |
2998 | return (result); | |
2999 | } | |
3000 | ||
572e2857 BB |
3001 | /* |
3002 | * If you call dbuf_rele() you had better not be referencing the dnode handle | |
3003 | * unless you have some other direct or indirect hold on the dnode. (An indirect | |
3004 | * hold is a hold on one of the dnode's dbufs, including the bonus buffer.) | |
3005 | * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the | |
3006 | * dnode's parent dbuf evicting its dnode handles. | |
3007 | */ | |
34dc7c2f BB |
3008 | void |
3009 | dbuf_rele(dmu_buf_impl_t *db, void *tag) | |
428870ff BB |
3010 | { |
3011 | mutex_enter(&db->db_mtx); | |
3012 | dbuf_rele_and_unlock(db, tag); | |
3013 | } | |
3014 | ||
b0bc7a84 MG |
3015 | void |
3016 | dmu_buf_rele(dmu_buf_t *db, void *tag) | |
3017 | { | |
3018 | dbuf_rele((dmu_buf_impl_t *)db, tag); | |
3019 | } | |
3020 | ||
428870ff BB |
3021 | /* |
3022 | * dbuf_rele() for an already-locked dbuf. This is necessary to allow | |
3023 | * db_dirtycnt and db_holds to be updated atomically. | |
3024 | */ | |
3025 | void | |
3026 | dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag) | |
34dc7c2f BB |
3027 | { |
3028 | int64_t holds; | |
3029 | ||
428870ff | 3030 | ASSERT(MUTEX_HELD(&db->db_mtx)); |
34dc7c2f BB |
3031 | DBUF_VERIFY(db); |
3032 | ||
572e2857 BB |
3033 | /* |
3034 | * Remove the reference to the dbuf before removing its hold on the | |
3035 | * dnode so we can guarantee in dnode_move() that a referenced bonus | |
3036 | * buffer has a corresponding dnode hold. | |
3037 | */ | |
34dc7c2f BB |
3038 | holds = refcount_remove(&db->db_holds, tag); |
3039 | ASSERT(holds >= 0); | |
3040 | ||
3041 | /* | |
3042 | * We can't freeze indirects if there is a possibility that they | |
3043 | * may be modified in the current syncing context. | |
3044 | */ | |
d3c2ae1c GW |
3045 | if (db->db_buf != NULL && |
3046 | holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) { | |
34dc7c2f | 3047 | arc_buf_freeze(db->db_buf); |
d3c2ae1c | 3048 | } |
34dc7c2f BB |
3049 | |
3050 | if (holds == db->db_dirtycnt && | |
bc4501f7 | 3051 | db->db_level == 0 && db->db_user_immediate_evict) |
34dc7c2f BB |
3052 | dbuf_evict_user(db); |
3053 | ||
3054 | if (holds == 0) { | |
428870ff | 3055 | if (db->db_blkid == DMU_BONUS_BLKID) { |
4c7b7eed | 3056 | dnode_t *dn; |
bc4501f7 | 3057 | boolean_t evict_dbuf = db->db_pending_evict; |
572e2857 BB |
3058 | |
3059 | /* | |
4c7b7eed JG |
3060 | * If the dnode moves here, we cannot cross this |
3061 | * barrier until the move completes. | |
572e2857 BB |
3062 | */ |
3063 | DB_DNODE_ENTER(db); | |
4c7b7eed JG |
3064 | |
3065 | dn = DB_DNODE(db); | |
3066 | atomic_dec_32(&dn->dn_dbufs_count); | |
3067 | ||
3068 | /* | |
3069 | * Decrementing the dbuf count means that the bonus | |
3070 | * buffer's dnode hold is no longer discounted in | |
3071 | * dnode_move(). The dnode cannot move until after | |
bc4501f7 | 3072 | * the dnode_rele() below. |
4c7b7eed | 3073 | */ |
572e2857 | 3074 | DB_DNODE_EXIT(db); |
4c7b7eed JG |
3075 | |
3076 | /* | |
3077 | * Do not reference db after its lock is dropped. | |
3078 | * Another thread may evict it. | |
3079 | */ | |
3080 | mutex_exit(&db->db_mtx); | |
3081 | ||
bc4501f7 | 3082 | if (evict_dbuf) |
4c7b7eed | 3083 | dnode_evict_bonus(dn); |
bc4501f7 JG |
3084 | |
3085 | dnode_rele(dn, db); | |
34dc7c2f BB |
3086 | } else if (db->db_buf == NULL) { |
3087 | /* | |
3088 | * This is a special case: we never associated this | |
3089 | * dbuf with any data allocated from the ARC. | |
3090 | */ | |
b128c09f BB |
3091 | ASSERT(db->db_state == DB_UNCACHED || |
3092 | db->db_state == DB_NOFILL); | |
d3c2ae1c | 3093 | dbuf_destroy(db); |
34dc7c2f | 3094 | } else if (arc_released(db->db_buf)) { |
34dc7c2f BB |
3095 | /* |
3096 | * This dbuf has anonymous data associated with it. | |
3097 | */ | |
d3c2ae1c | 3098 | dbuf_destroy(db); |
34dc7c2f | 3099 | } else { |
d3c2ae1c GW |
3100 | boolean_t do_arc_evict = B_FALSE; |
3101 | blkptr_t bp; | |
3102 | spa_t *spa = dmu_objset_spa(db->db_objset); | |
3103 | ||
3104 | if (!DBUF_IS_CACHEABLE(db) && | |
3105 | db->db_blkptr != NULL && | |
3106 | !BP_IS_HOLE(db->db_blkptr) && | |
3107 | !BP_IS_EMBEDDED(db->db_blkptr)) { | |
3108 | do_arc_evict = B_TRUE; | |
3109 | bp = *db->db_blkptr; | |
3110 | } | |
1eb5bfa3 | 3111 | |
d3c2ae1c GW |
3112 | if (!DBUF_IS_CACHEABLE(db) || |
3113 | db->db_pending_evict) { | |
3114 | dbuf_destroy(db); | |
3115 | } else if (!multilist_link_active(&db->db_cache_link)) { | |
64fc7762 | 3116 | multilist_insert(dbuf_cache, db); |
d3c2ae1c GW |
3117 | (void) refcount_add_many(&dbuf_cache_size, |
3118 | db->db.db_size, db); | |
b128c09f | 3119 | mutex_exit(&db->db_mtx); |
d3c2ae1c GW |
3120 | |
3121 | dbuf_evict_notify(); | |
bd089c54 | 3122 | } |
d3c2ae1c GW |
3123 | |
3124 | if (do_arc_evict) | |
3125 | arc_freed(spa, &bp); | |
34dc7c2f BB |
3126 | } |
3127 | } else { | |
3128 | mutex_exit(&db->db_mtx); | |
3129 | } | |
d3c2ae1c | 3130 | |
34dc7c2f BB |
3131 | } |
3132 | ||
3133 | #pragma weak dmu_buf_refcount = dbuf_refcount | |
3134 | uint64_t | |
3135 | dbuf_refcount(dmu_buf_impl_t *db) | |
3136 | { | |
3137 | return (refcount_count(&db->db_holds)); | |
3138 | } | |
3139 | ||
3140 | void * | |
0c66c32d JG |
3141 | dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user, |
3142 | dmu_buf_user_t *new_user) | |
34dc7c2f | 3143 | { |
0c66c32d JG |
3144 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; |
3145 | ||
3146 | mutex_enter(&db->db_mtx); | |
3147 | dbuf_verify_user(db, DBVU_NOT_EVICTING); | |
3148 | if (db->db_user == old_user) | |
3149 | db->db_user = new_user; | |
3150 | else | |
3151 | old_user = db->db_user; | |
3152 | dbuf_verify_user(db, DBVU_NOT_EVICTING); | |
3153 | mutex_exit(&db->db_mtx); | |
3154 | ||
3155 | return (old_user); | |
34dc7c2f BB |
3156 | } |
3157 | ||
3158 | void * | |
0c66c32d | 3159 | dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user) |
34dc7c2f | 3160 | { |
0c66c32d | 3161 | return (dmu_buf_replace_user(db_fake, NULL, user)); |
34dc7c2f BB |
3162 | } |
3163 | ||
3164 | void * | |
0c66c32d | 3165 | dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user) |
34dc7c2f BB |
3166 | { |
3167 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
34dc7c2f | 3168 | |
bc4501f7 | 3169 | db->db_user_immediate_evict = TRUE; |
0c66c32d JG |
3170 | return (dmu_buf_set_user(db_fake, user)); |
3171 | } | |
34dc7c2f | 3172 | |
0c66c32d JG |
3173 | void * |
3174 | dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user) | |
3175 | { | |
3176 | return (dmu_buf_replace_user(db_fake, user, NULL)); | |
34dc7c2f BB |
3177 | } |
3178 | ||
3179 | void * | |
3180 | dmu_buf_get_user(dmu_buf_t *db_fake) | |
3181 | { | |
3182 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
34dc7c2f | 3183 | |
0c66c32d JG |
3184 | dbuf_verify_user(db, DBVU_NOT_EVICTING); |
3185 | return (db->db_user); | |
3186 | } | |
3187 | ||
3188 | void | |
3189 | dmu_buf_user_evict_wait() | |
3190 | { | |
3191 | taskq_wait(dbu_evict_taskq); | |
34dc7c2f BB |
3192 | } |
3193 | ||
03c6040b GW |
3194 | blkptr_t * |
3195 | dmu_buf_get_blkptr(dmu_buf_t *db) | |
3196 | { | |
3197 | dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; | |
3198 | return (dbi->db_blkptr); | |
3199 | } | |
3200 | ||
8bea9815 MA |
3201 | objset_t * |
3202 | dmu_buf_get_objset(dmu_buf_t *db) | |
3203 | { | |
3204 | dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; | |
3205 | return (dbi->db_objset); | |
3206 | } | |
3207 | ||
2bce8049 MA |
3208 | dnode_t * |
3209 | dmu_buf_dnode_enter(dmu_buf_t *db) | |
3210 | { | |
3211 | dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; | |
3212 | DB_DNODE_ENTER(dbi); | |
3213 | return (DB_DNODE(dbi)); | |
3214 | } | |
3215 | ||
3216 | void | |
3217 | dmu_buf_dnode_exit(dmu_buf_t *db) | |
3218 | { | |
3219 | dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; | |
3220 | DB_DNODE_EXIT(dbi); | |
3221 | } | |
3222 | ||
34dc7c2f BB |
3223 | static void |
3224 | dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db) | |
3225 | { | |
3226 | /* ASSERT(dmu_tx_is_syncing(tx) */ | |
3227 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
3228 | ||
3229 | if (db->db_blkptr != NULL) | |
3230 | return; | |
3231 | ||
428870ff | 3232 | if (db->db_blkid == DMU_SPILL_BLKID) { |
50c957f7 | 3233 | db->db_blkptr = DN_SPILL_BLKPTR(dn->dn_phys); |
428870ff BB |
3234 | BP_ZERO(db->db_blkptr); |
3235 | return; | |
3236 | } | |
34dc7c2f BB |
3237 | if (db->db_level == dn->dn_phys->dn_nlevels-1) { |
3238 | /* | |
3239 | * This buffer was allocated at a time when there was | |
3240 | * no available blkptrs from the dnode, or it was | |
3241 | * inappropriate to hook it in (i.e., nlevels mis-match). | |
3242 | */ | |
3243 | ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr); | |
3244 | ASSERT(db->db_parent == NULL); | |
3245 | db->db_parent = dn->dn_dbuf; | |
3246 | db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid]; | |
3247 | DBUF_VERIFY(db); | |
3248 | } else { | |
3249 | dmu_buf_impl_t *parent = db->db_parent; | |
3250 | int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; | |
3251 | ||
3252 | ASSERT(dn->dn_phys->dn_nlevels > 1); | |
3253 | if (parent == NULL) { | |
3254 | mutex_exit(&db->db_mtx); | |
3255 | rw_enter(&dn->dn_struct_rwlock, RW_READER); | |
fcff0f35 PD |
3256 | parent = dbuf_hold_level(dn, db->db_level + 1, |
3257 | db->db_blkid >> epbs, db); | |
34dc7c2f BB |
3258 | rw_exit(&dn->dn_struct_rwlock); |
3259 | mutex_enter(&db->db_mtx); | |
3260 | db->db_parent = parent; | |
3261 | } | |
3262 | db->db_blkptr = (blkptr_t *)parent->db.db_data + | |
3263 | (db->db_blkid & ((1ULL << epbs) - 1)); | |
3264 | DBUF_VERIFY(db); | |
3265 | } | |
3266 | } | |
3267 | ||
b5256303 TC |
3268 | /* |
3269 | * Ensure the dbuf's data is untransformed if the associated dirty | |
3270 | * record requires it. This is used by dbuf_sync_leaf() to ensure | |
3271 | * that a dnode block is decrypted before we write new data to it. | |
3272 | * For raw writes we assert that the buffer is already encrypted. | |
3273 | */ | |
3274 | static void | |
3275 | dbuf_check_crypt(dbuf_dirty_record_t *dr) | |
3276 | { | |
3277 | int err; | |
3278 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
3279 | ||
3280 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
3281 | ||
3282 | if (!dr->dt.dl.dr_raw && arc_is_encrypted(db->db_buf)) { | |
3283 | /* | |
3284 | * Unfortunately, there is currently no mechanism for | |
3285 | * syncing context to handle decryption errors. An error | |
3286 | * here is only possible if an attacker maliciously | |
3287 | * changed a dnode block and updated the associated | |
3288 | * checksums going up the block tree. | |
3289 | */ | |
3290 | err = arc_untransform(db->db_buf, db->db_objset->os_spa, | |
3291 | dmu_objset_id(db->db_objset), B_TRUE); | |
3292 | if (err) | |
3293 | panic("Invalid dnode block MAC"); | |
3294 | } else if (dr->dt.dl.dr_raw) { | |
3295 | /* | |
3296 | * Writing raw encrypted data requires the db's arc buffer | |
3297 | * to be converted to raw by the caller. | |
3298 | */ | |
3299 | ASSERT(arc_is_encrypted(db->db_buf)); | |
3300 | } | |
3301 | } | |
3302 | ||
d1d7e268 MK |
3303 | /* |
3304 | * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it | |
60948de1 BB |
3305 | * is critical the we not allow the compiler to inline this function in to |
3306 | * dbuf_sync_list() thereby drastically bloating the stack usage. | |
3307 | */ | |
3308 | noinline static void | |
34dc7c2f BB |
3309 | dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx) |
3310 | { | |
3311 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
572e2857 | 3312 | dnode_t *dn; |
34dc7c2f BB |
3313 | zio_t *zio; |
3314 | ||
3315 | ASSERT(dmu_tx_is_syncing(tx)); | |
3316 | ||
3317 | dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); | |
3318 | ||
3319 | mutex_enter(&db->db_mtx); | |
3320 | ||
3321 | ASSERT(db->db_level > 0); | |
3322 | DBUF_VERIFY(db); | |
3323 | ||
e49f1e20 | 3324 | /* Read the block if it hasn't been read yet. */ |
34dc7c2f BB |
3325 | if (db->db_buf == NULL) { |
3326 | mutex_exit(&db->db_mtx); | |
3327 | (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED); | |
3328 | mutex_enter(&db->db_mtx); | |
3329 | } | |
3330 | ASSERT3U(db->db_state, ==, DB_CACHED); | |
34dc7c2f BB |
3331 | ASSERT(db->db_buf != NULL); |
3332 | ||
572e2857 BB |
3333 | DB_DNODE_ENTER(db); |
3334 | dn = DB_DNODE(db); | |
e49f1e20 | 3335 | /* Indirect block size must match what the dnode thinks it is. */ |
572e2857 | 3336 | ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); |
34dc7c2f | 3337 | dbuf_check_blkptr(dn, db); |
572e2857 | 3338 | DB_DNODE_EXIT(db); |
34dc7c2f | 3339 | |
e49f1e20 | 3340 | /* Provide the pending dirty record to child dbufs */ |
34dc7c2f BB |
3341 | db->db_data_pending = dr; |
3342 | ||
34dc7c2f | 3343 | mutex_exit(&db->db_mtx); |
b128c09f | 3344 | dbuf_write(dr, db->db_buf, tx); |
34dc7c2f BB |
3345 | |
3346 | zio = dr->dr_zio; | |
3347 | mutex_enter(&dr->dt.di.dr_mtx); | |
4bda3bd0 | 3348 | dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx); |
34dc7c2f BB |
3349 | ASSERT(list_head(&dr->dt.di.dr_children) == NULL); |
3350 | mutex_exit(&dr->dt.di.dr_mtx); | |
3351 | zio_nowait(zio); | |
3352 | } | |
3353 | ||
d1d7e268 MK |
3354 | /* |
3355 | * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is | |
60948de1 BB |
3356 | * critical the we not allow the compiler to inline this function in to |
3357 | * dbuf_sync_list() thereby drastically bloating the stack usage. | |
3358 | */ | |
3359 | noinline static void | |
34dc7c2f BB |
3360 | dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx) |
3361 | { | |
3362 | arc_buf_t **datap = &dr->dt.dl.dr_data; | |
3363 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
572e2857 BB |
3364 | dnode_t *dn; |
3365 | objset_t *os; | |
34dc7c2f | 3366 | uint64_t txg = tx->tx_txg; |
34dc7c2f BB |
3367 | |
3368 | ASSERT(dmu_tx_is_syncing(tx)); | |
3369 | ||
3370 | dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); | |
3371 | ||
3372 | mutex_enter(&db->db_mtx); | |
3373 | /* | |
3374 | * To be synced, we must be dirtied. But we | |
3375 | * might have been freed after the dirty. | |
3376 | */ | |
3377 | if (db->db_state == DB_UNCACHED) { | |
3378 | /* This buffer has been freed since it was dirtied */ | |
3379 | ASSERT(db->db.db_data == NULL); | |
3380 | } else if (db->db_state == DB_FILL) { | |
3381 | /* This buffer was freed and is now being re-filled */ | |
3382 | ASSERT(db->db.db_data != dr->dt.dl.dr_data); | |
3383 | } else { | |
b128c09f | 3384 | ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL); |
34dc7c2f BB |
3385 | } |
3386 | DBUF_VERIFY(db); | |
3387 | ||
572e2857 BB |
3388 | DB_DNODE_ENTER(db); |
3389 | dn = DB_DNODE(db); | |
3390 | ||
428870ff BB |
3391 | if (db->db_blkid == DMU_SPILL_BLKID) { |
3392 | mutex_enter(&dn->dn_mtx); | |
81edd3e8 P |
3393 | if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) { |
3394 | /* | |
3395 | * In the previous transaction group, the bonus buffer | |
3396 | * was entirely used to store the attributes for the | |
3397 | * dnode which overrode the dn_spill field. However, | |
3398 | * when adding more attributes to the file a spill | |
3399 | * block was required to hold the extra attributes. | |
3400 | * | |
3401 | * Make sure to clear the garbage left in the dn_spill | |
3402 | * field from the previous attributes in the bonus | |
3403 | * buffer. Otherwise, after writing out the spill | |
3404 | * block to the new allocated dva, it will free | |
3405 | * the old block pointed to by the invalid dn_spill. | |
3406 | */ | |
3407 | db->db_blkptr = NULL; | |
3408 | } | |
428870ff BB |
3409 | dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR; |
3410 | mutex_exit(&dn->dn_mtx); | |
3411 | } | |
3412 | ||
34dc7c2f BB |
3413 | /* |
3414 | * If this is a bonus buffer, simply copy the bonus data into the | |
3415 | * dnode. It will be written out when the dnode is synced (and it | |
3416 | * will be synced, since it must have been dirty for dbuf_sync to | |
3417 | * be called). | |
3418 | */ | |
428870ff | 3419 | if (db->db_blkid == DMU_BONUS_BLKID) { |
34dc7c2f BB |
3420 | dbuf_dirty_record_t **drp; |
3421 | ||
3422 | ASSERT(*datap != NULL); | |
c99c9001 | 3423 | ASSERT0(db->db_level); |
b5256303 | 3424 | ASSERT3U(DN_MAX_BONUS_LEN(dn->dn_phys), <=, |
50c957f7 | 3425 | DN_SLOTS_TO_BONUSLEN(dn->dn_phys->dn_extra_slots + 1)); |
b5256303 TC |
3426 | bcopy(*datap, DN_BONUS(dn->dn_phys), |
3427 | DN_MAX_BONUS_LEN(dn->dn_phys)); | |
572e2857 BB |
3428 | DB_DNODE_EXIT(db); |
3429 | ||
34dc7c2f | 3430 | if (*datap != db->db.db_data) { |
50c957f7 NB |
3431 | int slots = DB_DNODE(db)->dn_num_slots; |
3432 | int bonuslen = DN_SLOTS_TO_BONUSLEN(slots); | |
a3fd9d9e | 3433 | kmem_free(*datap, bonuslen); |
25458cbe | 3434 | arc_space_return(bonuslen, ARC_SPACE_BONUS); |
34dc7c2f BB |
3435 | } |
3436 | db->db_data_pending = NULL; | |
3437 | drp = &db->db_last_dirty; | |
3438 | while (*drp != dr) | |
3439 | drp = &(*drp)->dr_next; | |
3440 | ASSERT(dr->dr_next == NULL); | |
428870ff | 3441 | ASSERT(dr->dr_dbuf == db); |
34dc7c2f | 3442 | *drp = dr->dr_next; |
753972fc BB |
3443 | if (dr->dr_dbuf->db_level != 0) { |
3444 | mutex_destroy(&dr->dt.di.dr_mtx); | |
3445 | list_destroy(&dr->dt.di.dr_children); | |
3446 | } | |
34dc7c2f BB |
3447 | kmem_free(dr, sizeof (dbuf_dirty_record_t)); |
3448 | ASSERT(db->db_dirtycnt > 0); | |
3449 | db->db_dirtycnt -= 1; | |
428870ff | 3450 | dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg); |
34dc7c2f BB |
3451 | return; |
3452 | } | |
3453 | ||
572e2857 BB |
3454 | os = dn->dn_objset; |
3455 | ||
34dc7c2f BB |
3456 | /* |
3457 | * This function may have dropped the db_mtx lock allowing a dmu_sync | |
3458 | * operation to sneak in. As a result, we need to ensure that we | |
3459 | * don't check the dr_override_state until we have returned from | |
3460 | * dbuf_check_blkptr. | |
3461 | */ | |
3462 | dbuf_check_blkptr(dn, db); | |
3463 | ||
3464 | /* | |
572e2857 | 3465 | * If this buffer is in the middle of an immediate write, |
34dc7c2f BB |
3466 | * wait for the synchronous IO to complete. |
3467 | */ | |
3468 | while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) { | |
3469 | ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT); | |
3470 | cv_wait(&db->db_changed, &db->db_mtx); | |
3471 | ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN); | |
3472 | } | |
3473 | ||
b5256303 TC |
3474 | /* |
3475 | * If this is a dnode block, ensure it is appropriately encrypted | |
3476 | * or decrypted, depending on what we are writing to it this txg. | |
3477 | */ | |
3478 | if (os->os_encrypted && dn->dn_object == DMU_META_DNODE_OBJECT) | |
3479 | dbuf_check_crypt(dr); | |
3480 | ||
9babb374 BB |
3481 | if (db->db_state != DB_NOFILL && |
3482 | dn->dn_object != DMU_META_DNODE_OBJECT && | |
3483 | refcount_count(&db->db_holds) > 1 && | |
428870ff | 3484 | dr->dt.dl.dr_override_state != DR_OVERRIDDEN && |
9babb374 BB |
3485 | *datap == db->db_buf) { |
3486 | /* | |
3487 | * If this buffer is currently "in use" (i.e., there | |
3488 | * are active holds and db_data still references it), | |
3489 | * then make a copy before we start the write so that | |
3490 | * any modifications from the open txg will not leak | |
3491 | * into this write. | |
3492 | * | |
3493 | * NOTE: this copy does not need to be made for | |
3494 | * objects only modified in the syncing context (e.g. | |
3495 | * DNONE_DNODE blocks). | |
3496 | */ | |
2aa34383 | 3497 | int psize = arc_buf_size(*datap); |
b5256303 | 3498 | int lsize = arc_buf_lsize(*datap); |
9babb374 | 3499 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); |
2aa34383 DK |
3500 | enum zio_compress compress_type = arc_get_compression(*datap); |
3501 | ||
b5256303 TC |
3502 | if (arc_is_encrypted(*datap)) { |
3503 | boolean_t byteorder; | |
3504 | uint8_t salt[ZIO_DATA_SALT_LEN]; | |
3505 | uint8_t iv[ZIO_DATA_IV_LEN]; | |
3506 | uint8_t mac[ZIO_DATA_MAC_LEN]; | |
3507 | ||
3508 | arc_get_raw_params(*datap, &byteorder, salt, iv, mac); | |
3509 | *datap = arc_alloc_raw_buf(os->os_spa, db, | |
3510 | dmu_objset_id(os), byteorder, salt, iv, mac, | |
3511 | dn->dn_type, psize, lsize, compress_type); | |
3512 | } else if (compress_type != ZIO_COMPRESS_OFF) { | |
2aa34383 DK |
3513 | ASSERT3U(type, ==, ARC_BUFC_DATA); |
3514 | *datap = arc_alloc_compressed_buf(os->os_spa, db, | |
3515 | psize, lsize, compress_type); | |
b5256303 TC |
3516 | } else { |
3517 | *datap = arc_alloc_buf(os->os_spa, db, type, psize); | |
2aa34383 DK |
3518 | } |
3519 | bcopy(db->db.db_data, (*datap)->b_data, psize); | |
b128c09f | 3520 | } |
34dc7c2f BB |
3521 | db->db_data_pending = dr; |
3522 | ||
3523 | mutex_exit(&db->db_mtx); | |
3524 | ||
b128c09f | 3525 | dbuf_write(dr, *datap, tx); |
34dc7c2f BB |
3526 | |
3527 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
572e2857 | 3528 | if (dn->dn_object == DMU_META_DNODE_OBJECT) { |
34dc7c2f | 3529 | list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr); |
572e2857 BB |
3530 | DB_DNODE_EXIT(db); |
3531 | } else { | |
3532 | /* | |
3533 | * Although zio_nowait() does not "wait for an IO", it does | |
3534 | * initiate the IO. If this is an empty write it seems plausible | |
3535 | * that the IO could actually be completed before the nowait | |
3536 | * returns. We need to DB_DNODE_EXIT() first in case | |
3537 | * zio_nowait() invalidates the dbuf. | |
3538 | */ | |
3539 | DB_DNODE_EXIT(db); | |
34dc7c2f | 3540 | zio_nowait(dr->dr_zio); |
572e2857 | 3541 | } |
34dc7c2f BB |
3542 | } |
3543 | ||
3544 | void | |
4bda3bd0 | 3545 | dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx) |
34dc7c2f BB |
3546 | { |
3547 | dbuf_dirty_record_t *dr; | |
3548 | ||
c65aa5b2 | 3549 | while ((dr = list_head(list))) { |
34dc7c2f BB |
3550 | if (dr->dr_zio != NULL) { |
3551 | /* | |
3552 | * If we find an already initialized zio then we | |
3553 | * are processing the meta-dnode, and we have finished. | |
3554 | * The dbufs for all dnodes are put back on the list | |
3555 | * during processing, so that we can zio_wait() | |
3556 | * these IOs after initiating all child IOs. | |
3557 | */ | |
3558 | ASSERT3U(dr->dr_dbuf->db.db_object, ==, | |
3559 | DMU_META_DNODE_OBJECT); | |
3560 | break; | |
3561 | } | |
4bda3bd0 MA |
3562 | if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID && |
3563 | dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) { | |
3564 | VERIFY3U(dr->dr_dbuf->db_level, ==, level); | |
3565 | } | |
34dc7c2f BB |
3566 | list_remove(list, dr); |
3567 | if (dr->dr_dbuf->db_level > 0) | |
3568 | dbuf_sync_indirect(dr, tx); | |
3569 | else | |
3570 | dbuf_sync_leaf(dr, tx); | |
3571 | } | |
3572 | } | |
3573 | ||
34dc7c2f BB |
3574 | /* ARGSUSED */ |
3575 | static void | |
3576 | dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb) | |
3577 | { | |
3578 | dmu_buf_impl_t *db = vdb; | |
572e2857 | 3579 | dnode_t *dn; |
b128c09f | 3580 | blkptr_t *bp = zio->io_bp; |
34dc7c2f | 3581 | blkptr_t *bp_orig = &zio->io_bp_orig; |
428870ff BB |
3582 | spa_t *spa = zio->io_spa; |
3583 | int64_t delta; | |
34dc7c2f | 3584 | uint64_t fill = 0; |
428870ff | 3585 | int i; |
34dc7c2f | 3586 | |
463a8cfe AR |
3587 | ASSERT3P(db->db_blkptr, !=, NULL); |
3588 | ASSERT3P(&db->db_data_pending->dr_bp_copy, ==, bp); | |
b128c09f | 3589 | |
572e2857 BB |
3590 | DB_DNODE_ENTER(db); |
3591 | dn = DB_DNODE(db); | |
428870ff BB |
3592 | delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig); |
3593 | dnode_diduse_space(dn, delta - zio->io_prev_space_delta); | |
3594 | zio->io_prev_space_delta = delta; | |
34dc7c2f | 3595 | |
b0bc7a84 MG |
3596 | if (bp->blk_birth != 0) { |
3597 | ASSERT((db->db_blkid != DMU_SPILL_BLKID && | |
3598 | BP_GET_TYPE(bp) == dn->dn_type) || | |
3599 | (db->db_blkid == DMU_SPILL_BLKID && | |
9b67f605 MA |
3600 | BP_GET_TYPE(bp) == dn->dn_bonustype) || |
3601 | BP_IS_EMBEDDED(bp)); | |
b0bc7a84 | 3602 | ASSERT(BP_GET_LEVEL(bp) == db->db_level); |
34dc7c2f BB |
3603 | } |
3604 | ||
3605 | mutex_enter(&db->db_mtx); | |
3606 | ||
428870ff BB |
3607 | #ifdef ZFS_DEBUG |
3608 | if (db->db_blkid == DMU_SPILL_BLKID) { | |
428870ff | 3609 | ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); |
463a8cfe | 3610 | ASSERT(!(BP_IS_HOLE(bp)) && |
50c957f7 | 3611 | db->db_blkptr == DN_SPILL_BLKPTR(dn->dn_phys)); |
428870ff BB |
3612 | } |
3613 | #endif | |
3614 | ||
34dc7c2f BB |
3615 | if (db->db_level == 0) { |
3616 | mutex_enter(&dn->dn_mtx); | |
428870ff BB |
3617 | if (db->db_blkid > dn->dn_phys->dn_maxblkid && |
3618 | db->db_blkid != DMU_SPILL_BLKID) | |
34dc7c2f BB |
3619 | dn->dn_phys->dn_maxblkid = db->db_blkid; |
3620 | mutex_exit(&dn->dn_mtx); | |
3621 | ||
3622 | if (dn->dn_type == DMU_OT_DNODE) { | |
50c957f7 NB |
3623 | i = 0; |
3624 | while (i < db->db.db_size) { | |
817b1b6e MA |
3625 | dnode_phys_t *dnp = |
3626 | (void *)(((char *)db->db.db_data) + i); | |
50c957f7 NB |
3627 | |
3628 | i += DNODE_MIN_SIZE; | |
3629 | if (dnp->dn_type != DMU_OT_NONE) { | |
34dc7c2f | 3630 | fill++; |
50c957f7 NB |
3631 | i += dnp->dn_extra_slots * |
3632 | DNODE_MIN_SIZE; | |
3633 | } | |
34dc7c2f BB |
3634 | } |
3635 | } else { | |
b0bc7a84 MG |
3636 | if (BP_IS_HOLE(bp)) { |
3637 | fill = 0; | |
3638 | } else { | |
3639 | fill = 1; | |
3640 | } | |
34dc7c2f BB |
3641 | } |
3642 | } else { | |
b128c09f | 3643 | blkptr_t *ibp = db->db.db_data; |
34dc7c2f | 3644 | ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); |
b128c09f BB |
3645 | for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) { |
3646 | if (BP_IS_HOLE(ibp)) | |
34dc7c2f | 3647 | continue; |
9b67f605 | 3648 | fill += BP_GET_FILL(ibp); |
34dc7c2f BB |
3649 | } |
3650 | } | |
572e2857 | 3651 | DB_DNODE_EXIT(db); |
34dc7c2f | 3652 | |
9b67f605 | 3653 | if (!BP_IS_EMBEDDED(bp)) |
b5256303 | 3654 | BP_SET_FILL(bp, fill); |
34dc7c2f BB |
3655 | |
3656 | mutex_exit(&db->db_mtx); | |
463a8cfe AR |
3657 | |
3658 | rw_enter(&dn->dn_struct_rwlock, RW_WRITER); | |
3659 | *db->db_blkptr = *bp; | |
3660 | rw_exit(&dn->dn_struct_rwlock); | |
34dc7c2f BB |
3661 | } |
3662 | ||
bc77ba73 PD |
3663 | /* ARGSUSED */ |
3664 | /* | |
3665 | * This function gets called just prior to running through the compression | |
3666 | * stage of the zio pipeline. If we're an indirect block comprised of only | |
3667 | * holes, then we want this indirect to be compressed away to a hole. In | |
3668 | * order to do that we must zero out any information about the holes that | |
3669 | * this indirect points to prior to before we try to compress it. | |
3670 | */ | |
3671 | static void | |
3672 | dbuf_write_children_ready(zio_t *zio, arc_buf_t *buf, void *vdb) | |
3673 | { | |
3674 | dmu_buf_impl_t *db = vdb; | |
3675 | dnode_t *dn; | |
3676 | blkptr_t *bp; | |
721ed0ee | 3677 | unsigned int epbs, i; |
bc77ba73 PD |
3678 | |
3679 | ASSERT3U(db->db_level, >, 0); | |
3680 | DB_DNODE_ENTER(db); | |
3681 | dn = DB_DNODE(db); | |
3682 | epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; | |
721ed0ee | 3683 | ASSERT3U(epbs, <, 31); |
bc77ba73 PD |
3684 | |
3685 | /* Determine if all our children are holes */ | |
3f93077b | 3686 | for (i = 0, bp = db->db.db_data; i < 1ULL << epbs; i++, bp++) { |
bc77ba73 PD |
3687 | if (!BP_IS_HOLE(bp)) |
3688 | break; | |
3689 | } | |
3690 | ||
3691 | /* | |
3692 | * If all the children are holes, then zero them all out so that | |
3693 | * we may get compressed away. | |
3694 | */ | |
3f93077b | 3695 | if (i == 1ULL << epbs) { |
721ed0ee GM |
3696 | /* |
3697 | * We only found holes. Grab the rwlock to prevent | |
3698 | * anybody from reading the blocks we're about to | |
3699 | * zero out. | |
3700 | */ | |
3701 | rw_enter(&dn->dn_struct_rwlock, RW_WRITER); | |
bc77ba73 | 3702 | bzero(db->db.db_data, db->db.db_size); |
721ed0ee | 3703 | rw_exit(&dn->dn_struct_rwlock); |
bc77ba73 PD |
3704 | } |
3705 | DB_DNODE_EXIT(db); | |
3706 | } | |
3707 | ||
e8b96c60 MA |
3708 | /* |
3709 | * The SPA will call this callback several times for each zio - once | |
3710 | * for every physical child i/o (zio->io_phys_children times). This | |
3711 | * allows the DMU to monitor the progress of each logical i/o. For example, | |
3712 | * there may be 2 copies of an indirect block, or many fragments of a RAID-Z | |
3713 | * block. There may be a long delay before all copies/fragments are completed, | |
3714 | * so this callback allows us to retire dirty space gradually, as the physical | |
3715 | * i/os complete. | |
3716 | */ | |
3717 | /* ARGSUSED */ | |
3718 | static void | |
3719 | dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg) | |
3720 | { | |
3721 | dmu_buf_impl_t *db = arg; | |
3722 | objset_t *os = db->db_objset; | |
3723 | dsl_pool_t *dp = dmu_objset_pool(os); | |
3724 | dbuf_dirty_record_t *dr; | |
3725 | int delta = 0; | |
3726 | ||
3727 | dr = db->db_data_pending; | |
3728 | ASSERT3U(dr->dr_txg, ==, zio->io_txg); | |
3729 | ||
3730 | /* | |
3731 | * The callback will be called io_phys_children times. Retire one | |
3732 | * portion of our dirty space each time we are called. Any rounding | |
3733 | * error will be cleaned up by dsl_pool_sync()'s call to | |
3734 | * dsl_pool_undirty_space(). | |
3735 | */ | |
3736 | delta = dr->dr_accounted / zio->io_phys_children; | |
3737 | dsl_pool_undirty_space(dp, delta, zio->io_txg); | |
3738 | } | |
3739 | ||
34dc7c2f BB |
3740 | /* ARGSUSED */ |
3741 | static void | |
3742 | dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb) | |
3743 | { | |
3744 | dmu_buf_impl_t *db = vdb; | |
428870ff | 3745 | blkptr_t *bp_orig = &zio->io_bp_orig; |
b0bc7a84 MG |
3746 | blkptr_t *bp = db->db_blkptr; |
3747 | objset_t *os = db->db_objset; | |
3748 | dmu_tx_t *tx = os->os_synctx; | |
34dc7c2f BB |
3749 | dbuf_dirty_record_t **drp, *dr; |
3750 | ||
c99c9001 | 3751 | ASSERT0(zio->io_error); |
428870ff BB |
3752 | ASSERT(db->db_blkptr == bp); |
3753 | ||
03c6040b GW |
3754 | /* |
3755 | * For nopwrites and rewrites we ensure that the bp matches our | |
3756 | * original and bypass all the accounting. | |
3757 | */ | |
3758 | if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) { | |
428870ff BB |
3759 | ASSERT(BP_EQUAL(bp, bp_orig)); |
3760 | } else { | |
b0bc7a84 | 3761 | dsl_dataset_t *ds = os->os_dsl_dataset; |
428870ff BB |
3762 | (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE); |
3763 | dsl_dataset_block_born(ds, bp, tx); | |
3764 | } | |
34dc7c2f BB |
3765 | |
3766 | mutex_enter(&db->db_mtx); | |
3767 | ||
428870ff BB |
3768 | DBUF_VERIFY(db); |
3769 | ||
34dc7c2f BB |
3770 | drp = &db->db_last_dirty; |
3771 | while ((dr = *drp) != db->db_data_pending) | |
3772 | drp = &dr->dr_next; | |
3773 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
428870ff | 3774 | ASSERT(dr->dr_dbuf == db); |
34dc7c2f BB |
3775 | ASSERT(dr->dr_next == NULL); |
3776 | *drp = dr->dr_next; | |
3777 | ||
428870ff BB |
3778 | #ifdef ZFS_DEBUG |
3779 | if (db->db_blkid == DMU_SPILL_BLKID) { | |
572e2857 BB |
3780 | dnode_t *dn; |
3781 | ||
3782 | DB_DNODE_ENTER(db); | |
3783 | dn = DB_DNODE(db); | |
428870ff BB |
3784 | ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); |
3785 | ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && | |
50c957f7 | 3786 | db->db_blkptr == DN_SPILL_BLKPTR(dn->dn_phys)); |
572e2857 | 3787 | DB_DNODE_EXIT(db); |
428870ff BB |
3788 | } |
3789 | #endif | |
3790 | ||
34dc7c2f | 3791 | if (db->db_level == 0) { |
428870ff | 3792 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f | 3793 | ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); |
b128c09f BB |
3794 | if (db->db_state != DB_NOFILL) { |
3795 | if (dr->dt.dl.dr_data != db->db_buf) | |
d3c2ae1c | 3796 | arc_buf_destroy(dr->dt.dl.dr_data, db); |
b128c09f | 3797 | } |
34dc7c2f | 3798 | } else { |
572e2857 BB |
3799 | dnode_t *dn; |
3800 | ||
3801 | DB_DNODE_ENTER(db); | |
3802 | dn = DB_DNODE(db); | |
34dc7c2f | 3803 | ASSERT(list_head(&dr->dt.di.dr_children) == NULL); |
b0bc7a84 | 3804 | ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift); |
34dc7c2f | 3805 | if (!BP_IS_HOLE(db->db_blkptr)) { |
1fde1e37 BB |
3806 | ASSERTV(int epbs = dn->dn_phys->dn_indblkshift - |
3807 | SPA_BLKPTRSHIFT); | |
b0bc7a84 MG |
3808 | ASSERT3U(db->db_blkid, <=, |
3809 | dn->dn_phys->dn_maxblkid >> (db->db_level * epbs)); | |
34dc7c2f BB |
3810 | ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, |
3811 | db->db.db_size); | |
34dc7c2f | 3812 | } |
572e2857 | 3813 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
3814 | mutex_destroy(&dr->dt.di.dr_mtx); |
3815 | list_destroy(&dr->dt.di.dr_children); | |
3816 | } | |
3817 | kmem_free(dr, sizeof (dbuf_dirty_record_t)); | |
3818 | ||
3819 | cv_broadcast(&db->db_changed); | |
3820 | ASSERT(db->db_dirtycnt > 0); | |
3821 | db->db_dirtycnt -= 1; | |
3822 | db->db_data_pending = NULL; | |
b0bc7a84 | 3823 | dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg); |
428870ff BB |
3824 | } |
3825 | ||
3826 | static void | |
3827 | dbuf_write_nofill_ready(zio_t *zio) | |
3828 | { | |
3829 | dbuf_write_ready(zio, NULL, zio->io_private); | |
3830 | } | |
3831 | ||
3832 | static void | |
3833 | dbuf_write_nofill_done(zio_t *zio) | |
3834 | { | |
3835 | dbuf_write_done(zio, NULL, zio->io_private); | |
3836 | } | |
3837 | ||
3838 | static void | |
3839 | dbuf_write_override_ready(zio_t *zio) | |
3840 | { | |
3841 | dbuf_dirty_record_t *dr = zio->io_private; | |
3842 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
3843 | ||
3844 | dbuf_write_ready(zio, NULL, db); | |
3845 | } | |
3846 | ||
3847 | static void | |
3848 | dbuf_write_override_done(zio_t *zio) | |
3849 | { | |
3850 | dbuf_dirty_record_t *dr = zio->io_private; | |
3851 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
3852 | blkptr_t *obp = &dr->dt.dl.dr_overridden_by; | |
3853 | ||
3854 | mutex_enter(&db->db_mtx); | |
3855 | if (!BP_EQUAL(zio->io_bp, obp)) { | |
3856 | if (!BP_IS_HOLE(obp)) | |
3857 | dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp); | |
3858 | arc_release(dr->dt.dl.dr_data, db); | |
3859 | } | |
34dc7c2f BB |
3860 | mutex_exit(&db->db_mtx); |
3861 | ||
428870ff | 3862 | dbuf_write_done(zio, NULL, db); |
a6255b7f DQ |
3863 | |
3864 | if (zio->io_abd != NULL) | |
3865 | abd_put(zio->io_abd); | |
428870ff BB |
3866 | } |
3867 | ||
e49f1e20 | 3868 | /* Issue I/O to commit a dirty buffer to disk. */ |
428870ff BB |
3869 | static void |
3870 | dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx) | |
3871 | { | |
3872 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
572e2857 BB |
3873 | dnode_t *dn; |
3874 | objset_t *os; | |
428870ff BB |
3875 | dmu_buf_impl_t *parent = db->db_parent; |
3876 | uint64_t txg = tx->tx_txg; | |
5dbd68a3 | 3877 | zbookmark_phys_t zb; |
428870ff BB |
3878 | zio_prop_t zp; |
3879 | zio_t *zio; | |
3880 | int wp_flag = 0; | |
34dc7c2f | 3881 | |
463a8cfe AR |
3882 | ASSERT(dmu_tx_is_syncing(tx)); |
3883 | ||
572e2857 BB |
3884 | DB_DNODE_ENTER(db); |
3885 | dn = DB_DNODE(db); | |
3886 | os = dn->dn_objset; | |
3887 | ||
428870ff BB |
3888 | if (db->db_state != DB_NOFILL) { |
3889 | if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) { | |
3890 | /* | |
3891 | * Private object buffers are released here rather | |
3892 | * than in dbuf_dirty() since they are only modified | |
3893 | * in the syncing context and we don't want the | |
3894 | * overhead of making multiple copies of the data. | |
3895 | */ | |
3896 | if (BP_IS_HOLE(db->db_blkptr)) { | |
3897 | arc_buf_thaw(data); | |
3898 | } else { | |
3899 | dbuf_release_bp(db); | |
3900 | } | |
3901 | } | |
3902 | } | |
3903 | ||
3904 | if (parent != dn->dn_dbuf) { | |
e49f1e20 WA |
3905 | /* Our parent is an indirect block. */ |
3906 | /* We have a dirty parent that has been scheduled for write. */ | |
428870ff | 3907 | ASSERT(parent && parent->db_data_pending); |
e49f1e20 | 3908 | /* Our parent's buffer is one level closer to the dnode. */ |
428870ff | 3909 | ASSERT(db->db_level == parent->db_level-1); |
e49f1e20 WA |
3910 | /* |
3911 | * We're about to modify our parent's db_data by modifying | |
3912 | * our block pointer, so the parent must be released. | |
3913 | */ | |
428870ff BB |
3914 | ASSERT(arc_released(parent->db_buf)); |
3915 | zio = parent->db_data_pending->dr_zio; | |
3916 | } else { | |
e49f1e20 | 3917 | /* Our parent is the dnode itself. */ |
428870ff BB |
3918 | ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 && |
3919 | db->db_blkid != DMU_SPILL_BLKID) || | |
3920 | (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0)); | |
3921 | if (db->db_blkid != DMU_SPILL_BLKID) | |
3922 | ASSERT3P(db->db_blkptr, ==, | |
3923 | &dn->dn_phys->dn_blkptr[db->db_blkid]); | |
3924 | zio = dn->dn_zio; | |
3925 | } | |
3926 | ||
3927 | ASSERT(db->db_level == 0 || data == db->db_buf); | |
3928 | ASSERT3U(db->db_blkptr->blk_birth, <=, txg); | |
3929 | ASSERT(zio); | |
3930 | ||
3931 | SET_BOOKMARK(&zb, os->os_dsl_dataset ? | |
3932 | os->os_dsl_dataset->ds_object : DMU_META_OBJSET, | |
3933 | db->db.db_object, db->db_level, db->db_blkid); | |
3934 | ||
3935 | if (db->db_blkid == DMU_SPILL_BLKID) | |
3936 | wp_flag = WP_SPILL; | |
3937 | wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0; | |
3938 | ||
82644107 | 3939 | dmu_write_policy(os, dn, db->db_level, wp_flag, &zp); |
572e2857 | 3940 | DB_DNODE_EXIT(db); |
428870ff | 3941 | |
463a8cfe AR |
3942 | /* |
3943 | * We copy the blkptr now (rather than when we instantiate the dirty | |
3944 | * record), because its value can change between open context and | |
3945 | * syncing context. We do not need to hold dn_struct_rwlock to read | |
3946 | * db_blkptr because we are in syncing context. | |
3947 | */ | |
3948 | dr->dr_bp_copy = *db->db_blkptr; | |
3949 | ||
9b67f605 MA |
3950 | if (db->db_level == 0 && |
3951 | dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { | |
3952 | /* | |
3953 | * The BP for this block has been provided by open context | |
3954 | * (by dmu_sync() or dmu_buf_write_embedded()). | |
3955 | */ | |
a6255b7f DQ |
3956 | abd_t *contents = (data != NULL) ? |
3957 | abd_get_from_buf(data->b_data, arc_buf_size(data)) : NULL; | |
9b67f605 | 3958 | |
428870ff | 3959 | dr->dr_zio = zio_write(zio, os->os_spa, txg, |
2aa34383 DK |
3960 | &dr->dr_bp_copy, contents, db->db.db_size, db->db.db_size, |
3961 | &zp, dbuf_write_override_ready, NULL, NULL, | |
bc77ba73 | 3962 | dbuf_write_override_done, |
e8b96c60 | 3963 | dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); |
428870ff BB |
3964 | mutex_enter(&db->db_mtx); |
3965 | dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; | |
3966 | zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by, | |
03c6040b | 3967 | dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite); |
428870ff BB |
3968 | mutex_exit(&db->db_mtx); |
3969 | } else if (db->db_state == DB_NOFILL) { | |
3c67d83a TH |
3970 | ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF || |
3971 | zp.zp_checksum == ZIO_CHECKSUM_NOPARITY); | |
428870ff | 3972 | dr->dr_zio = zio_write(zio, os->os_spa, txg, |
2aa34383 | 3973 | &dr->dr_bp_copy, NULL, db->db.db_size, db->db.db_size, &zp, |
bc77ba73 PD |
3974 | dbuf_write_nofill_ready, NULL, NULL, |
3975 | dbuf_write_nofill_done, db, | |
428870ff BB |
3976 | ZIO_PRIORITY_ASYNC_WRITE, |
3977 | ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb); | |
3978 | } else { | |
3979 | ASSERT(arc_released(data)); | |
bc77ba73 PD |
3980 | |
3981 | /* | |
3982 | * For indirect blocks, we want to setup the children | |
3983 | * ready callback so that we can properly handle an indirect | |
3984 | * block that only contains holes. | |
3985 | */ | |
1c27024e | 3986 | arc_write_done_func_t *children_ready_cb = NULL; |
bc77ba73 PD |
3987 | if (db->db_level != 0) |
3988 | children_ready_cb = dbuf_write_children_ready; | |
3989 | ||
428870ff | 3990 | dr->dr_zio = arc_write(zio, os->os_spa, txg, |
463a8cfe | 3991 | &dr->dr_bp_copy, data, DBUF_IS_L2CACHEABLE(db), |
d3c2ae1c GW |
3992 | &zp, dbuf_write_ready, |
3993 | children_ready_cb, dbuf_write_physdone, | |
3994 | dbuf_write_done, db, ZIO_PRIORITY_ASYNC_WRITE, | |
3995 | ZIO_FLAG_MUSTSUCCEED, &zb); | |
428870ff | 3996 | } |
34dc7c2f | 3997 | } |
c28b2279 BB |
3998 | |
3999 | #if defined(_KERNEL) && defined(HAVE_SPL) | |
8f576c23 BB |
4000 | EXPORT_SYMBOL(dbuf_find); |
4001 | EXPORT_SYMBOL(dbuf_is_metadata); | |
d3c2ae1c | 4002 | EXPORT_SYMBOL(dbuf_destroy); |
8f576c23 BB |
4003 | EXPORT_SYMBOL(dbuf_loan_arcbuf); |
4004 | EXPORT_SYMBOL(dbuf_whichblock); | |
4005 | EXPORT_SYMBOL(dbuf_read); | |
4006 | EXPORT_SYMBOL(dbuf_unoverride); | |
4007 | EXPORT_SYMBOL(dbuf_free_range); | |
4008 | EXPORT_SYMBOL(dbuf_new_size); | |
4009 | EXPORT_SYMBOL(dbuf_release_bp); | |
4010 | EXPORT_SYMBOL(dbuf_dirty); | |
b5256303 | 4011 | EXPORT_SYMBOL(dmu_buf_will_change_crypt_params); |
c28b2279 | 4012 | EXPORT_SYMBOL(dmu_buf_will_dirty); |
8f576c23 BB |
4013 | EXPORT_SYMBOL(dmu_buf_will_not_fill); |
4014 | EXPORT_SYMBOL(dmu_buf_will_fill); | |
4015 | EXPORT_SYMBOL(dmu_buf_fill_done); | |
4047414a | 4016 | EXPORT_SYMBOL(dmu_buf_rele); |
8f576c23 | 4017 | EXPORT_SYMBOL(dbuf_assign_arcbuf); |
8f576c23 BB |
4018 | EXPORT_SYMBOL(dbuf_prefetch); |
4019 | EXPORT_SYMBOL(dbuf_hold_impl); | |
4020 | EXPORT_SYMBOL(dbuf_hold); | |
4021 | EXPORT_SYMBOL(dbuf_hold_level); | |
4022 | EXPORT_SYMBOL(dbuf_create_bonus); | |
4023 | EXPORT_SYMBOL(dbuf_spill_set_blksz); | |
4024 | EXPORT_SYMBOL(dbuf_rm_spill); | |
4025 | EXPORT_SYMBOL(dbuf_add_ref); | |
4026 | EXPORT_SYMBOL(dbuf_rele); | |
4027 | EXPORT_SYMBOL(dbuf_rele_and_unlock); | |
4028 | EXPORT_SYMBOL(dbuf_refcount); | |
4029 | EXPORT_SYMBOL(dbuf_sync_list); | |
4030 | EXPORT_SYMBOL(dmu_buf_set_user); | |
4031 | EXPORT_SYMBOL(dmu_buf_set_user_ie); | |
8f576c23 | 4032 | EXPORT_SYMBOL(dmu_buf_get_user); |
0f699108 | 4033 | EXPORT_SYMBOL(dmu_buf_get_blkptr); |
d3c2ae1c | 4034 | |
02730c33 | 4035 | /* BEGIN CSTYLED */ |
d3c2ae1c GW |
4036 | module_param(dbuf_cache_max_bytes, ulong, 0644); |
4037 | MODULE_PARM_DESC(dbuf_cache_max_bytes, | |
02730c33 | 4038 | "Maximum size in bytes of the dbuf cache."); |
d3c2ae1c GW |
4039 | |
4040 | module_param(dbuf_cache_hiwater_pct, uint, 0644); | |
4041 | MODULE_PARM_DESC(dbuf_cache_hiwater_pct, | |
f974e25d | 4042 | "Percentage over dbuf_cache_max_bytes when dbufs must be evicted " |
4043 | "directly."); | |
d3c2ae1c GW |
4044 | |
4045 | module_param(dbuf_cache_lowater_pct, uint, 0644); | |
4046 | MODULE_PARM_DESC(dbuf_cache_lowater_pct, | |
f974e25d | 4047 | "Percentage below dbuf_cache_max_bytes when the evict thread stops " |
4048 | "evicting dbufs."); | |
d3c2ae1c GW |
4049 | |
4050 | module_param(dbuf_cache_max_shift, int, 0644); | |
4051 | MODULE_PARM_DESC(dbuf_cache_max_shift, | |
f974e25d | 4052 | "Cap the size of the dbuf cache to a log2 fraction of arc size."); |
02730c33 | 4053 | /* END CSTYLED */ |
c28b2279 | 4054 | #endif |