]>
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> |
a1d477c2 | 50 | #include <sys/vdev.h> |
34dc7c2f | 51 | |
5e021f56 GDN |
52 | kstat_t *dbuf_ksp; |
53 | ||
54 | typedef struct dbuf_stats { | |
55 | /* | |
56 | * Various statistics about the size of the dbuf cache. | |
57 | */ | |
58 | kstat_named_t cache_count; | |
59 | kstat_named_t cache_size_bytes; | |
60 | kstat_named_t cache_size_bytes_max; | |
61 | /* | |
62 | * Statistics regarding the bounds on the dbuf cache size. | |
63 | */ | |
64 | kstat_named_t cache_target_bytes; | |
65 | kstat_named_t cache_lowater_bytes; | |
66 | kstat_named_t cache_hiwater_bytes; | |
67 | /* | |
68 | * Total number of dbuf cache evictions that have occurred. | |
69 | */ | |
70 | kstat_named_t cache_total_evicts; | |
71 | /* | |
72 | * The distribution of dbuf levels in the dbuf cache and | |
73 | * the total size of all dbufs at each level. | |
74 | */ | |
75 | kstat_named_t cache_levels[DN_MAX_LEVELS]; | |
76 | kstat_named_t cache_levels_bytes[DN_MAX_LEVELS]; | |
77 | /* | |
78 | * Statistics about the dbuf hash table. | |
79 | */ | |
80 | kstat_named_t hash_hits; | |
81 | kstat_named_t hash_misses; | |
82 | kstat_named_t hash_collisions; | |
83 | kstat_named_t hash_elements; | |
84 | kstat_named_t hash_elements_max; | |
85 | /* | |
86 | * Number of sublists containing more than one dbuf in the dbuf | |
87 | * hash table. Keep track of the longest hash chain. | |
88 | */ | |
89 | kstat_named_t hash_chains; | |
90 | kstat_named_t hash_chain_max; | |
91 | /* | |
92 | * Number of times a dbuf_create() discovers that a dbuf was | |
93 | * already created and in the dbuf hash table. | |
94 | */ | |
95 | kstat_named_t hash_insert_race; | |
96 | } dbuf_stats_t; | |
97 | ||
98 | dbuf_stats_t dbuf_stats = { | |
99 | { "cache_count", KSTAT_DATA_UINT64 }, | |
100 | { "cache_size_bytes", KSTAT_DATA_UINT64 }, | |
101 | { "cache_size_bytes_max", KSTAT_DATA_UINT64 }, | |
102 | { "cache_target_bytes", KSTAT_DATA_UINT64 }, | |
103 | { "cache_lowater_bytes", KSTAT_DATA_UINT64 }, | |
104 | { "cache_hiwater_bytes", KSTAT_DATA_UINT64 }, | |
105 | { "cache_total_evicts", KSTAT_DATA_UINT64 }, | |
106 | { { "cache_levels_N", KSTAT_DATA_UINT64 } }, | |
107 | { { "cache_levels_bytes_N", KSTAT_DATA_UINT64 } }, | |
108 | { "hash_hits", KSTAT_DATA_UINT64 }, | |
109 | { "hash_misses", KSTAT_DATA_UINT64 }, | |
110 | { "hash_collisions", KSTAT_DATA_UINT64 }, | |
111 | { "hash_elements", KSTAT_DATA_UINT64 }, | |
112 | { "hash_elements_max", KSTAT_DATA_UINT64 }, | |
113 | { "hash_chains", KSTAT_DATA_UINT64 }, | |
114 | { "hash_chain_max", KSTAT_DATA_UINT64 }, | |
115 | { "hash_insert_race", KSTAT_DATA_UINT64 } | |
116 | }; | |
117 | ||
118 | #define DBUF_STAT_INCR(stat, val) \ | |
119 | atomic_add_64(&dbuf_stats.stat.value.ui64, (val)); | |
120 | #define DBUF_STAT_DECR(stat, val) \ | |
121 | DBUF_STAT_INCR(stat, -(val)); | |
122 | #define DBUF_STAT_BUMP(stat) \ | |
123 | DBUF_STAT_INCR(stat, 1); | |
124 | #define DBUF_STAT_BUMPDOWN(stat) \ | |
125 | DBUF_STAT_INCR(stat, -1); | |
126 | #define DBUF_STAT_MAX(stat, v) { \ | |
127 | uint64_t _m; \ | |
128 | while ((v) > (_m = dbuf_stats.stat.value.ui64) && \ | |
129 | (_m != atomic_cas_64(&dbuf_stats.stat.value.ui64, _m, (v))))\ | |
130 | continue; \ | |
131 | } | |
132 | ||
fc5bb51f BB |
133 | struct dbuf_hold_impl_data { |
134 | /* Function arguments */ | |
135 | dnode_t *dh_dn; | |
136 | uint8_t dh_level; | |
137 | uint64_t dh_blkid; | |
fcff0f35 PD |
138 | boolean_t dh_fail_sparse; |
139 | boolean_t dh_fail_uncached; | |
fc5bb51f BB |
140 | void *dh_tag; |
141 | dmu_buf_impl_t **dh_dbp; | |
142 | /* Local variables */ | |
143 | dmu_buf_impl_t *dh_db; | |
144 | dmu_buf_impl_t *dh_parent; | |
145 | blkptr_t *dh_bp; | |
146 | int dh_err; | |
147 | dbuf_dirty_record_t *dh_dr; | |
fc5bb51f BB |
148 | int dh_depth; |
149 | }; | |
150 | ||
151 | static void __dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh, | |
fcff0f35 PD |
152 | dnode_t *dn, uint8_t level, uint64_t blkid, boolean_t fail_sparse, |
153 | boolean_t fail_uncached, | |
154 | void *tag, dmu_buf_impl_t **dbp, int depth); | |
fc5bb51f BB |
155 | static int __dbuf_hold_impl(struct dbuf_hold_impl_data *dh); |
156 | ||
d3c2ae1c | 157 | uint_t zfs_dbuf_evict_key; |
b663a23d | 158 | |
13fe0198 | 159 | static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx); |
b128c09f | 160 | static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx); |
34dc7c2f | 161 | |
0c66c32d | 162 | extern inline void dmu_buf_init_user(dmu_buf_user_t *dbu, |
39efbde7 GM |
163 | dmu_buf_evict_func_t *evict_func_sync, |
164 | dmu_buf_evict_func_t *evict_func_async, | |
165 | dmu_buf_t **clear_on_evict_dbufp); | |
0c66c32d | 166 | |
34dc7c2f BB |
167 | /* |
168 | * Global data structures and functions for the dbuf cache. | |
169 | */ | |
d3c2ae1c | 170 | static kmem_cache_t *dbuf_kmem_cache; |
0c66c32d | 171 | static taskq_t *dbu_evict_taskq; |
34dc7c2f | 172 | |
d3c2ae1c GW |
173 | static kthread_t *dbuf_cache_evict_thread; |
174 | static kmutex_t dbuf_evict_lock; | |
175 | static kcondvar_t dbuf_evict_cv; | |
176 | static boolean_t dbuf_evict_thread_exit; | |
177 | ||
178 | /* | |
179 | * LRU cache of dbufs. The dbuf cache maintains a list of dbufs that | |
180 | * are not currently held but have been recently released. These dbufs | |
181 | * are not eligible for arc eviction until they are aged out of the cache. | |
182 | * Dbufs are added to the dbuf cache once the last hold is released. If a | |
183 | * dbuf is later accessed and still exists in the dbuf cache, then it will | |
184 | * be removed from the cache and later re-added to the head of the cache. | |
185 | * Dbufs that are aged out of the cache will be immediately destroyed and | |
186 | * become eligible for arc eviction. | |
187 | */ | |
64fc7762 | 188 | static multilist_t *dbuf_cache; |
d3c2ae1c | 189 | static refcount_t dbuf_cache_size; |
de4f8d5d | 190 | unsigned long dbuf_cache_max_bytes = 0; |
d3c2ae1c | 191 | |
de4f8d5d BB |
192 | /* Set the default size of the dbuf cache to log2 fraction of arc size. */ |
193 | int dbuf_cache_shift = 5; | |
d3c2ae1c GW |
194 | |
195 | /* | |
196 | * The dbuf cache uses a three-stage eviction policy: | |
197 | * - A low water marker designates when the dbuf eviction thread | |
198 | * should stop evicting from the dbuf cache. | |
199 | * - When we reach the maximum size (aka mid water mark), we | |
200 | * signal the eviction thread to run. | |
201 | * - The high water mark indicates when the eviction thread | |
202 | * is unable to keep up with the incoming load and eviction must | |
203 | * happen in the context of the calling thread. | |
204 | * | |
205 | * The dbuf cache: | |
206 | * (max size) | |
207 | * low water mid water hi water | |
208 | * +----------------------------------------+----------+----------+ | |
209 | * | | | | | |
210 | * | | | | | |
211 | * | | | | | |
212 | * | | | | | |
213 | * +----------------------------------------+----------+----------+ | |
214 | * stop signal evict | |
215 | * evicting eviction directly | |
216 | * thread | |
217 | * | |
218 | * The high and low water marks indicate the operating range for the eviction | |
219 | * thread. The low water mark is, by default, 90% of the total size of the | |
220 | * cache and the high water mark is at 110% (both of these percentages can be | |
221 | * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct, | |
222 | * respectively). The eviction thread will try to ensure that the cache remains | |
223 | * within this range by waking up every second and checking if the cache is | |
224 | * above the low water mark. The thread can also be woken up by callers adding | |
225 | * elements into the cache if the cache is larger than the mid water (i.e max | |
226 | * cache size). Once the eviction thread is woken up and eviction is required, | |
227 | * it will continue evicting buffers until it's able to reduce the cache size | |
228 | * to the low water mark. If the cache size continues to grow and hits the high | |
4e33ba4c | 229 | * water mark, then callers adding elements to the cache will begin to evict |
d3c2ae1c GW |
230 | * directly from the cache until the cache is no longer above the high water |
231 | * mark. | |
232 | */ | |
233 | ||
234 | /* | |
235 | * The percentage above and below the maximum cache size. | |
236 | */ | |
237 | uint_t dbuf_cache_hiwater_pct = 10; | |
238 | uint_t dbuf_cache_lowater_pct = 10; | |
239 | ||
34dc7c2f BB |
240 | /* ARGSUSED */ |
241 | static int | |
242 | dbuf_cons(void *vdb, void *unused, int kmflag) | |
243 | { | |
244 | dmu_buf_impl_t *db = vdb; | |
245 | bzero(db, sizeof (dmu_buf_impl_t)); | |
246 | ||
247 | mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL); | |
248 | cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL); | |
d3c2ae1c | 249 | multilist_link_init(&db->db_cache_link); |
34dc7c2f | 250 | refcount_create(&db->db_holds); |
8951cb8d | 251 | |
34dc7c2f BB |
252 | return (0); |
253 | } | |
254 | ||
255 | /* ARGSUSED */ | |
256 | static void | |
257 | dbuf_dest(void *vdb, void *unused) | |
258 | { | |
259 | dmu_buf_impl_t *db = vdb; | |
260 | mutex_destroy(&db->db_mtx); | |
261 | cv_destroy(&db->db_changed); | |
d3c2ae1c | 262 | ASSERT(!multilist_link_active(&db->db_cache_link)); |
34dc7c2f BB |
263 | refcount_destroy(&db->db_holds); |
264 | } | |
265 | ||
266 | /* | |
267 | * dbuf hash table routines | |
268 | */ | |
269 | static dbuf_hash_table_t dbuf_hash_table; | |
270 | ||
271 | static uint64_t dbuf_hash_count; | |
272 | ||
273 | static uint64_t | |
274 | dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid) | |
275 | { | |
276 | uintptr_t osv = (uintptr_t)os; | |
277 | uint64_t crc = -1ULL; | |
278 | ||
279 | ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); | |
280 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF]; | |
281 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF]; | |
282 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF]; | |
283 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF]; | |
284 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF]; | |
285 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF]; | |
286 | ||
287 | crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16); | |
288 | ||
289 | return (crc); | |
290 | } | |
291 | ||
34dc7c2f BB |
292 | #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \ |
293 | ((dbuf)->db.db_object == (obj) && \ | |
294 | (dbuf)->db_objset == (os) && \ | |
295 | (dbuf)->db_level == (level) && \ | |
296 | (dbuf)->db_blkid == (blkid)) | |
297 | ||
298 | dmu_buf_impl_t * | |
6ebebace | 299 | dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid) |
34dc7c2f BB |
300 | { |
301 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
d6320ddb BB |
302 | uint64_t hv; |
303 | uint64_t idx; | |
34dc7c2f BB |
304 | dmu_buf_impl_t *db; |
305 | ||
d3c2ae1c | 306 | hv = dbuf_hash(os, obj, level, blkid); |
d6320ddb BB |
307 | idx = hv & h->hash_table_mask; |
308 | ||
34dc7c2f BB |
309 | mutex_enter(DBUF_HASH_MUTEX(h, idx)); |
310 | for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) { | |
311 | if (DBUF_EQUAL(db, os, obj, level, blkid)) { | |
312 | mutex_enter(&db->db_mtx); | |
313 | if (db->db_state != DB_EVICTING) { | |
314 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); | |
315 | return (db); | |
316 | } | |
317 | mutex_exit(&db->db_mtx); | |
318 | } | |
319 | } | |
320 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); | |
321 | return (NULL); | |
322 | } | |
323 | ||
6ebebace JG |
324 | static dmu_buf_impl_t * |
325 | dbuf_find_bonus(objset_t *os, uint64_t object) | |
326 | { | |
327 | dnode_t *dn; | |
328 | dmu_buf_impl_t *db = NULL; | |
329 | ||
330 | if (dnode_hold(os, object, FTAG, &dn) == 0) { | |
331 | rw_enter(&dn->dn_struct_rwlock, RW_READER); | |
332 | if (dn->dn_bonus != NULL) { | |
333 | db = dn->dn_bonus; | |
334 | mutex_enter(&db->db_mtx); | |
335 | } | |
336 | rw_exit(&dn->dn_struct_rwlock); | |
337 | dnode_rele(dn, FTAG); | |
338 | } | |
339 | return (db); | |
340 | } | |
341 | ||
34dc7c2f BB |
342 | /* |
343 | * Insert an entry into the hash table. If there is already an element | |
344 | * equal to elem in the hash table, then the already existing element | |
345 | * will be returned and the new element will not be inserted. | |
346 | * Otherwise returns NULL. | |
347 | */ | |
348 | static dmu_buf_impl_t * | |
349 | dbuf_hash_insert(dmu_buf_impl_t *db) | |
350 | { | |
351 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
428870ff | 352 | objset_t *os = db->db_objset; |
34dc7c2f BB |
353 | uint64_t obj = db->db.db_object; |
354 | int level = db->db_level; | |
d6320ddb | 355 | uint64_t blkid, hv, idx; |
34dc7c2f | 356 | dmu_buf_impl_t *dbf; |
5e021f56 | 357 | uint32_t i; |
34dc7c2f | 358 | |
d6320ddb | 359 | blkid = db->db_blkid; |
d3c2ae1c | 360 | hv = dbuf_hash(os, obj, level, blkid); |
d6320ddb BB |
361 | idx = hv & h->hash_table_mask; |
362 | ||
34dc7c2f | 363 | mutex_enter(DBUF_HASH_MUTEX(h, idx)); |
5e021f56 GDN |
364 | for (dbf = h->hash_table[idx], i = 0; dbf != NULL; |
365 | dbf = dbf->db_hash_next, i++) { | |
34dc7c2f BB |
366 | if (DBUF_EQUAL(dbf, os, obj, level, blkid)) { |
367 | mutex_enter(&dbf->db_mtx); | |
368 | if (dbf->db_state != DB_EVICTING) { | |
369 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); | |
370 | return (dbf); | |
371 | } | |
372 | mutex_exit(&dbf->db_mtx); | |
373 | } | |
374 | } | |
375 | ||
5e021f56 GDN |
376 | if (i > 0) { |
377 | DBUF_STAT_BUMP(hash_collisions); | |
378 | if (i == 1) | |
379 | DBUF_STAT_BUMP(hash_chains); | |
380 | ||
381 | DBUF_STAT_MAX(hash_chain_max, i); | |
382 | } | |
383 | ||
34dc7c2f BB |
384 | mutex_enter(&db->db_mtx); |
385 | db->db_hash_next = h->hash_table[idx]; | |
386 | h->hash_table[idx] = db; | |
387 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); | |
bc89ac84 | 388 | atomic_inc_64(&dbuf_hash_count); |
5e021f56 | 389 | DBUF_STAT_MAX(hash_elements_max, dbuf_hash_count); |
34dc7c2f BB |
390 | |
391 | return (NULL); | |
392 | } | |
393 | ||
394 | /* | |
bd089c54 | 395 | * Remove an entry from the hash table. It must be in the EVICTING state. |
34dc7c2f BB |
396 | */ |
397 | static void | |
398 | dbuf_hash_remove(dmu_buf_impl_t *db) | |
399 | { | |
400 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
d6320ddb | 401 | uint64_t hv, idx; |
34dc7c2f BB |
402 | dmu_buf_impl_t *dbf, **dbp; |
403 | ||
d3c2ae1c | 404 | hv = dbuf_hash(db->db_objset, db->db.db_object, |
d6320ddb BB |
405 | db->db_level, db->db_blkid); |
406 | idx = hv & h->hash_table_mask; | |
407 | ||
34dc7c2f | 408 | /* |
4e33ba4c | 409 | * We mustn't hold db_mtx to maintain lock ordering: |
34dc7c2f BB |
410 | * DBUF_HASH_MUTEX > db_mtx. |
411 | */ | |
412 | ASSERT(refcount_is_zero(&db->db_holds)); | |
413 | ASSERT(db->db_state == DB_EVICTING); | |
414 | ASSERT(!MUTEX_HELD(&db->db_mtx)); | |
415 | ||
416 | mutex_enter(DBUF_HASH_MUTEX(h, idx)); | |
417 | dbp = &h->hash_table[idx]; | |
418 | while ((dbf = *dbp) != db) { | |
419 | dbp = &dbf->db_hash_next; | |
420 | ASSERT(dbf != NULL); | |
421 | } | |
422 | *dbp = db->db_hash_next; | |
423 | db->db_hash_next = NULL; | |
5e021f56 GDN |
424 | if (h->hash_table[idx] && |
425 | h->hash_table[idx]->db_hash_next == NULL) | |
426 | DBUF_STAT_BUMPDOWN(hash_chains); | |
34dc7c2f | 427 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); |
bc89ac84 | 428 | atomic_dec_64(&dbuf_hash_count); |
34dc7c2f BB |
429 | } |
430 | ||
0c66c32d JG |
431 | typedef enum { |
432 | DBVU_EVICTING, | |
433 | DBVU_NOT_EVICTING | |
434 | } dbvu_verify_type_t; | |
435 | ||
436 | static void | |
437 | dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type) | |
438 | { | |
439 | #ifdef ZFS_DEBUG | |
440 | int64_t holds; | |
441 | ||
442 | if (db->db_user == NULL) | |
443 | return; | |
444 | ||
445 | /* Only data blocks support the attachment of user data. */ | |
446 | ASSERT(db->db_level == 0); | |
447 | ||
448 | /* Clients must resolve a dbuf before attaching user data. */ | |
449 | ASSERT(db->db.db_data != NULL); | |
450 | ASSERT3U(db->db_state, ==, DB_CACHED); | |
451 | ||
452 | holds = refcount_count(&db->db_holds); | |
453 | if (verify_type == DBVU_EVICTING) { | |
454 | /* | |
455 | * Immediate eviction occurs when holds == dirtycnt. | |
456 | * For normal eviction buffers, holds is zero on | |
457 | * eviction, except when dbuf_fix_old_data() calls | |
458 | * dbuf_clear_data(). However, the hold count can grow | |
459 | * during eviction even though db_mtx is held (see | |
460 | * dmu_bonus_hold() for an example), so we can only | |
461 | * test the generic invariant that holds >= dirtycnt. | |
462 | */ | |
463 | ASSERT3U(holds, >=, db->db_dirtycnt); | |
464 | } else { | |
bc4501f7 | 465 | if (db->db_user_immediate_evict == TRUE) |
0c66c32d JG |
466 | ASSERT3U(holds, >=, db->db_dirtycnt); |
467 | else | |
468 | ASSERT3U(holds, >, 0); | |
469 | } | |
470 | #endif | |
471 | } | |
472 | ||
34dc7c2f BB |
473 | static void |
474 | dbuf_evict_user(dmu_buf_impl_t *db) | |
475 | { | |
0c66c32d JG |
476 | dmu_buf_user_t *dbu = db->db_user; |
477 | ||
34dc7c2f BB |
478 | ASSERT(MUTEX_HELD(&db->db_mtx)); |
479 | ||
0c66c32d | 480 | if (dbu == NULL) |
34dc7c2f BB |
481 | return; |
482 | ||
0c66c32d JG |
483 | dbuf_verify_user(db, DBVU_EVICTING); |
484 | db->db_user = NULL; | |
485 | ||
486 | #ifdef ZFS_DEBUG | |
487 | if (dbu->dbu_clear_on_evict_dbufp != NULL) | |
488 | *dbu->dbu_clear_on_evict_dbufp = NULL; | |
489 | #endif | |
490 | ||
491 | /* | |
39efbde7 GM |
492 | * There are two eviction callbacks - one that we call synchronously |
493 | * and one that we invoke via a taskq. The async one is useful for | |
494 | * avoiding lock order reversals and limiting stack depth. | |
495 | * | |
496 | * Note that if we have a sync callback but no async callback, | |
497 | * it's likely that the sync callback will free the structure | |
498 | * containing the dbu. In that case we need to take care to not | |
499 | * dereference dbu after calling the sync evict func. | |
0c66c32d | 500 | */ |
a7004725 | 501 | boolean_t has_async = (dbu->dbu_evict_func_async != NULL); |
39efbde7 GM |
502 | |
503 | if (dbu->dbu_evict_func_sync != NULL) | |
504 | dbu->dbu_evict_func_sync(dbu); | |
505 | ||
506 | if (has_async) { | |
507 | taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func_async, | |
508 | dbu, 0, &dbu->dbu_tqent); | |
509 | } | |
34dc7c2f BB |
510 | } |
511 | ||
572e2857 BB |
512 | boolean_t |
513 | dbuf_is_metadata(dmu_buf_impl_t *db) | |
514 | { | |
cc79a5c2 BB |
515 | /* |
516 | * Consider indirect blocks and spill blocks to be meta data. | |
517 | */ | |
518 | if (db->db_level > 0 || db->db_blkid == DMU_SPILL_BLKID) { | |
572e2857 BB |
519 | return (B_TRUE); |
520 | } else { | |
521 | boolean_t is_metadata; | |
522 | ||
523 | DB_DNODE_ENTER(db); | |
9ae529ec | 524 | is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type); |
572e2857 BB |
525 | DB_DNODE_EXIT(db); |
526 | ||
527 | return (is_metadata); | |
528 | } | |
529 | } | |
530 | ||
d3c2ae1c GW |
531 | |
532 | /* | |
533 | * This function *must* return indices evenly distributed between all | |
534 | * sublists of the multilist. This is needed due to how the dbuf eviction | |
535 | * code is laid out; dbuf_evict_thread() assumes dbufs are evenly | |
536 | * distributed between all sublists and uses this assumption when | |
537 | * deciding which sublist to evict from and how much to evict from it. | |
538 | */ | |
539 | unsigned int | |
540 | dbuf_cache_multilist_index_func(multilist_t *ml, void *obj) | |
34dc7c2f | 541 | { |
d3c2ae1c GW |
542 | dmu_buf_impl_t *db = obj; |
543 | ||
544 | /* | |
545 | * The assumption here, is the hash value for a given | |
546 | * dmu_buf_impl_t will remain constant throughout it's lifetime | |
547 | * (i.e. it's objset, object, level and blkid fields don't change). | |
548 | * Thus, we don't need to store the dbuf's sublist index | |
549 | * on insertion, as this index can be recalculated on removal. | |
550 | * | |
551 | * Also, the low order bits of the hash value are thought to be | |
552 | * distributed evenly. Otherwise, in the case that the multilist | |
553 | * has a power of two number of sublists, each sublists' usage | |
554 | * would not be evenly distributed. | |
555 | */ | |
556 | return (dbuf_hash(db->db_objset, db->db.db_object, | |
557 | db->db_level, db->db_blkid) % | |
558 | multilist_get_num_sublists(ml)); | |
559 | } | |
560 | ||
e71cade6 | 561 | static inline unsigned long |
562 | dbuf_cache_target_bytes(void) | |
563 | { | |
564 | return MIN(dbuf_cache_max_bytes, | |
de4f8d5d | 565 | arc_target_bytes() >> dbuf_cache_shift); |
e71cade6 | 566 | } |
567 | ||
5e021f56 GDN |
568 | static inline uint64_t |
569 | dbuf_cache_hiwater_bytes(void) | |
d3c2ae1c | 570 | { |
e71cade6 | 571 | uint64_t dbuf_cache_target = dbuf_cache_target_bytes(); |
5e021f56 GDN |
572 | return (dbuf_cache_target + |
573 | (dbuf_cache_target * dbuf_cache_hiwater_pct) / 100); | |
574 | } | |
e71cade6 | 575 | |
5e021f56 GDN |
576 | static inline uint64_t |
577 | dbuf_cache_lowater_bytes(void) | |
578 | { | |
579 | uint64_t dbuf_cache_target = dbuf_cache_target_bytes(); | |
580 | return (dbuf_cache_target - | |
581 | (dbuf_cache_target * dbuf_cache_lowater_pct) / 100); | |
582 | } | |
d3c2ae1c | 583 | |
5e021f56 GDN |
584 | static inline boolean_t |
585 | dbuf_cache_above_hiwater(void) | |
586 | { | |
587 | return (refcount_count(&dbuf_cache_size) > dbuf_cache_hiwater_bytes()); | |
d3c2ae1c GW |
588 | } |
589 | ||
590 | static inline boolean_t | |
591 | dbuf_cache_above_lowater(void) | |
592 | { | |
5e021f56 | 593 | return (refcount_count(&dbuf_cache_size) > dbuf_cache_lowater_bytes()); |
d3c2ae1c GW |
594 | } |
595 | ||
596 | /* | |
597 | * Evict the oldest eligible dbuf from the dbuf cache. | |
598 | */ | |
599 | static void | |
600 | dbuf_evict_one(void) | |
601 | { | |
64fc7762 MA |
602 | int idx = multilist_get_random_index(dbuf_cache); |
603 | multilist_sublist_t *mls = multilist_sublist_lock(dbuf_cache, idx); | |
1c27024e | 604 | |
d3c2ae1c GW |
605 | ASSERT(!MUTEX_HELD(&dbuf_evict_lock)); |
606 | ||
607 | /* | |
608 | * Set the thread's tsd to indicate that it's processing evictions. | |
609 | * Once a thread stops evicting from the dbuf cache it will | |
610 | * reset its tsd to NULL. | |
611 | */ | |
612 | ASSERT3P(tsd_get(zfs_dbuf_evict_key), ==, NULL); | |
613 | (void) tsd_set(zfs_dbuf_evict_key, (void *)B_TRUE); | |
614 | ||
1c27024e | 615 | dmu_buf_impl_t *db = multilist_sublist_tail(mls); |
d3c2ae1c GW |
616 | while (db != NULL && mutex_tryenter(&db->db_mtx) == 0) { |
617 | db = multilist_sublist_prev(mls, db); | |
618 | } | |
619 | ||
620 | DTRACE_PROBE2(dbuf__evict__one, dmu_buf_impl_t *, db, | |
621 | multilist_sublist_t *, mls); | |
622 | ||
623 | if (db != NULL) { | |
624 | multilist_sublist_remove(mls, db); | |
625 | multilist_sublist_unlock(mls); | |
626 | (void) refcount_remove_many(&dbuf_cache_size, | |
627 | db->db.db_size, db); | |
5e021f56 GDN |
628 | DBUF_STAT_BUMPDOWN(cache_levels[db->db_level]); |
629 | DBUF_STAT_BUMPDOWN(cache_count); | |
630 | DBUF_STAT_DECR(cache_levels_bytes[db->db_level], | |
631 | db->db.db_size); | |
d3c2ae1c | 632 | dbuf_destroy(db); |
5e021f56 GDN |
633 | DBUF_STAT_MAX(cache_size_bytes_max, |
634 | refcount_count(&dbuf_cache_size)); | |
635 | DBUF_STAT_BUMP(cache_total_evicts); | |
d3c2ae1c GW |
636 | } else { |
637 | multilist_sublist_unlock(mls); | |
638 | } | |
639 | (void) tsd_set(zfs_dbuf_evict_key, NULL); | |
640 | } | |
641 | ||
642 | /* | |
643 | * The dbuf evict thread is responsible for aging out dbufs from the | |
644 | * cache. Once the cache has reached it's maximum size, dbufs are removed | |
645 | * and destroyed. The eviction thread will continue running until the size | |
646 | * of the dbuf cache is at or below the maximum size. Once the dbuf is aged | |
647 | * out of the cache it is destroyed and becomes eligible for arc eviction. | |
648 | */ | |
867959b5 | 649 | /* ARGSUSED */ |
d3c2ae1c | 650 | static void |
c25b8f99 | 651 | dbuf_evict_thread(void *unused) |
d3c2ae1c GW |
652 | { |
653 | callb_cpr_t cpr; | |
654 | ||
655 | CALLB_CPR_INIT(&cpr, &dbuf_evict_lock, callb_generic_cpr, FTAG); | |
656 | ||
657 | mutex_enter(&dbuf_evict_lock); | |
658 | while (!dbuf_evict_thread_exit) { | |
659 | while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) { | |
660 | CALLB_CPR_SAFE_BEGIN(&cpr); | |
661 | (void) cv_timedwait_sig_hires(&dbuf_evict_cv, | |
662 | &dbuf_evict_lock, SEC2NSEC(1), MSEC2NSEC(1), 0); | |
663 | CALLB_CPR_SAFE_END(&cpr, &dbuf_evict_lock); | |
664 | } | |
665 | mutex_exit(&dbuf_evict_lock); | |
666 | ||
667 | /* | |
668 | * Keep evicting as long as we're above the low water mark | |
669 | * for the cache. We do this without holding the locks to | |
670 | * minimize lock contention. | |
671 | */ | |
672 | while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) { | |
673 | dbuf_evict_one(); | |
674 | } | |
675 | ||
676 | mutex_enter(&dbuf_evict_lock); | |
677 | } | |
678 | ||
679 | dbuf_evict_thread_exit = B_FALSE; | |
680 | cv_broadcast(&dbuf_evict_cv); | |
681 | CALLB_CPR_EXIT(&cpr); /* drops dbuf_evict_lock */ | |
682 | thread_exit(); | |
683 | } | |
684 | ||
685 | /* | |
686 | * Wake up the dbuf eviction thread if the dbuf cache is at its max size. | |
687 | * If the dbuf cache is at its high water mark, then evict a dbuf from the | |
688 | * dbuf cache using the callers context. | |
689 | */ | |
690 | static void | |
691 | dbuf_evict_notify(void) | |
692 | { | |
693 | ||
694 | /* | |
695 | * We use thread specific data to track when a thread has | |
696 | * started processing evictions. This allows us to avoid deeply | |
697 | * nested stacks that would have a call flow similar to this: | |
698 | * | |
699 | * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify() | |
700 | * ^ | | |
701 | * | | | |
702 | * +-----dbuf_destroy()<--dbuf_evict_one()<--------+ | |
703 | * | |
704 | * The dbuf_eviction_thread will always have its tsd set until | |
705 | * that thread exits. All other threads will only set their tsd | |
706 | * if they are participating in the eviction process. This only | |
707 | * happens if the eviction thread is unable to process evictions | |
708 | * fast enough. To keep the dbuf cache size in check, other threads | |
709 | * can evict from the dbuf cache directly. Those threads will set | |
710 | * their tsd values so that we ensure that they only evict one dbuf | |
711 | * from the dbuf cache. | |
712 | */ | |
713 | if (tsd_get(zfs_dbuf_evict_key) != NULL) | |
714 | return; | |
715 | ||
38240ebd MA |
716 | /* |
717 | * We check if we should evict without holding the dbuf_evict_lock, | |
718 | * because it's OK to occasionally make the wrong decision here, | |
719 | * and grabbing the lock results in massive lock contention. | |
720 | */ | |
e71cade6 | 721 | if (refcount_count(&dbuf_cache_size) > dbuf_cache_target_bytes()) { |
38240ebd | 722 | if (dbuf_cache_above_hiwater()) |
d3c2ae1c | 723 | dbuf_evict_one(); |
38240ebd | 724 | cv_signal(&dbuf_evict_cv); |
d3c2ae1c | 725 | } |
34dc7c2f BB |
726 | } |
727 | ||
5e021f56 GDN |
728 | static int |
729 | dbuf_kstat_update(kstat_t *ksp, int rw) | |
730 | { | |
731 | dbuf_stats_t *ds = ksp->ks_data; | |
d3c2ae1c | 732 | |
5e021f56 GDN |
733 | if (rw == KSTAT_WRITE) { |
734 | return (SET_ERROR(EACCES)); | |
735 | } else { | |
736 | ds->cache_size_bytes.value.ui64 = | |
737 | refcount_count(&dbuf_cache_size); | |
738 | ds->cache_target_bytes.value.ui64 = dbuf_cache_target_bytes(); | |
739 | ds->cache_hiwater_bytes.value.ui64 = dbuf_cache_hiwater_bytes(); | |
740 | ds->cache_lowater_bytes.value.ui64 = dbuf_cache_lowater_bytes(); | |
741 | ds->hash_elements.value.ui64 = dbuf_hash_count; | |
742 | } | |
743 | ||
744 | return (0); | |
745 | } | |
d3c2ae1c | 746 | |
34dc7c2f BB |
747 | void |
748 | dbuf_init(void) | |
749 | { | |
750 | uint64_t hsize = 1ULL << 16; | |
751 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
752 | int i; | |
753 | ||
754 | /* | |
755 | * The hash table is big enough to fill all of physical memory | |
69de3421 TC |
756 | * with an average block size of zfs_arc_average_blocksize (default 8K). |
757 | * By default, the table will take up | |
758 | * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). | |
34dc7c2f | 759 | */ |
69de3421 | 760 | while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE) |
34dc7c2f BB |
761 | hsize <<= 1; |
762 | ||
763 | retry: | |
764 | h->hash_table_mask = hsize - 1; | |
93ce2b4c | 765 | #if defined(_KERNEL) |
d1d7e268 MK |
766 | /* |
767 | * Large allocations which do not require contiguous pages | |
768 | * should be using vmem_alloc() in the linux kernel | |
769 | */ | |
79c76d5b | 770 | h->hash_table = vmem_zalloc(hsize * sizeof (void *), KM_SLEEP); |
00b46022 | 771 | #else |
34dc7c2f | 772 | h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP); |
00b46022 | 773 | #endif |
34dc7c2f BB |
774 | if (h->hash_table == NULL) { |
775 | /* XXX - we should really return an error instead of assert */ | |
776 | ASSERT(hsize > (1ULL << 10)); | |
777 | hsize >>= 1; | |
778 | goto retry; | |
779 | } | |
780 | ||
d3c2ae1c | 781 | dbuf_kmem_cache = kmem_cache_create("dmu_buf_impl_t", |
34dc7c2f BB |
782 | sizeof (dmu_buf_impl_t), |
783 | 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0); | |
784 | ||
785 | for (i = 0; i < DBUF_MUTEXES; i++) | |
40d06e3c | 786 | mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL); |
e0b0ca98 BB |
787 | |
788 | dbuf_stats_init(h); | |
0c66c32d | 789 | |
d3c2ae1c | 790 | /* |
de4f8d5d BB |
791 | * Setup the parameters for the dbuf cache. We set the size of the |
792 | * dbuf cache to 1/32nd (default) of the target size of the ARC. If | |
793 | * the value has been specified as a module option and it's not | |
794 | * greater than the target size of the ARC, then we honor that value. | |
d3c2ae1c | 795 | */ |
de4f8d5d BB |
796 | if (dbuf_cache_max_bytes == 0 || |
797 | dbuf_cache_max_bytes >= arc_target_bytes()) { | |
798 | dbuf_cache_max_bytes = arc_target_bytes() >> dbuf_cache_shift; | |
799 | } | |
d3c2ae1c | 800 | |
0c66c32d JG |
801 | /* |
802 | * All entries are queued via taskq_dispatch_ent(), so min/maxalloc | |
803 | * configuration is not required. | |
804 | */ | |
1229323d | 805 | dbu_evict_taskq = taskq_create("dbu_evict", 1, defclsyspri, 0, 0, 0); |
d3c2ae1c | 806 | |
64fc7762 | 807 | dbuf_cache = multilist_create(sizeof (dmu_buf_impl_t), |
d3c2ae1c | 808 | offsetof(dmu_buf_impl_t, db_cache_link), |
d3c2ae1c GW |
809 | dbuf_cache_multilist_index_func); |
810 | refcount_create(&dbuf_cache_size); | |
811 | ||
812 | tsd_create(&zfs_dbuf_evict_key, NULL); | |
813 | dbuf_evict_thread_exit = B_FALSE; | |
814 | mutex_init(&dbuf_evict_lock, NULL, MUTEX_DEFAULT, NULL); | |
815 | cv_init(&dbuf_evict_cv, NULL, CV_DEFAULT, NULL); | |
816 | dbuf_cache_evict_thread = thread_create(NULL, 0, dbuf_evict_thread, | |
817 | NULL, 0, &p0, TS_RUN, minclsyspri); | |
5e021f56 GDN |
818 | |
819 | dbuf_ksp = kstat_create("zfs", 0, "dbufstats", "misc", | |
820 | KSTAT_TYPE_NAMED, sizeof (dbuf_stats) / sizeof (kstat_named_t), | |
821 | KSTAT_FLAG_VIRTUAL); | |
822 | if (dbuf_ksp != NULL) { | |
823 | dbuf_ksp->ks_data = &dbuf_stats; | |
824 | dbuf_ksp->ks_update = dbuf_kstat_update; | |
825 | kstat_install(dbuf_ksp); | |
826 | ||
827 | for (i = 0; i < DN_MAX_LEVELS; i++) { | |
828 | snprintf(dbuf_stats.cache_levels[i].name, | |
829 | KSTAT_STRLEN, "cache_level_%d", i); | |
830 | dbuf_stats.cache_levels[i].data_type = | |
831 | KSTAT_DATA_UINT64; | |
832 | snprintf(dbuf_stats.cache_levels_bytes[i].name, | |
833 | KSTAT_STRLEN, "cache_level_%d_bytes", i); | |
834 | dbuf_stats.cache_levels_bytes[i].data_type = | |
835 | KSTAT_DATA_UINT64; | |
836 | } | |
837 | } | |
34dc7c2f BB |
838 | } |
839 | ||
840 | void | |
841 | dbuf_fini(void) | |
842 | { | |
843 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
844 | int i; | |
845 | ||
e0b0ca98 BB |
846 | dbuf_stats_destroy(); |
847 | ||
34dc7c2f BB |
848 | for (i = 0; i < DBUF_MUTEXES; i++) |
849 | mutex_destroy(&h->hash_mutexes[i]); | |
93ce2b4c | 850 | #if defined(_KERNEL) |
d1d7e268 MK |
851 | /* |
852 | * Large allocations which do not require contiguous pages | |
853 | * should be using vmem_free() in the linux kernel | |
854 | */ | |
00b46022 BB |
855 | vmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *)); |
856 | #else | |
34dc7c2f | 857 | kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *)); |
00b46022 | 858 | #endif |
d3c2ae1c | 859 | kmem_cache_destroy(dbuf_kmem_cache); |
0c66c32d | 860 | taskq_destroy(dbu_evict_taskq); |
d3c2ae1c GW |
861 | |
862 | mutex_enter(&dbuf_evict_lock); | |
863 | dbuf_evict_thread_exit = B_TRUE; | |
864 | while (dbuf_evict_thread_exit) { | |
865 | cv_signal(&dbuf_evict_cv); | |
866 | cv_wait(&dbuf_evict_cv, &dbuf_evict_lock); | |
867 | } | |
868 | mutex_exit(&dbuf_evict_lock); | |
869 | tsd_destroy(&zfs_dbuf_evict_key); | |
870 | ||
871 | mutex_destroy(&dbuf_evict_lock); | |
872 | cv_destroy(&dbuf_evict_cv); | |
873 | ||
874 | refcount_destroy(&dbuf_cache_size); | |
64fc7762 | 875 | multilist_destroy(dbuf_cache); |
5e021f56 GDN |
876 | |
877 | if (dbuf_ksp != NULL) { | |
878 | kstat_delete(dbuf_ksp); | |
879 | dbuf_ksp = NULL; | |
880 | } | |
34dc7c2f BB |
881 | } |
882 | ||
883 | /* | |
884 | * Other stuff. | |
885 | */ | |
886 | ||
887 | #ifdef ZFS_DEBUG | |
888 | static void | |
889 | dbuf_verify(dmu_buf_impl_t *db) | |
890 | { | |
572e2857 | 891 | dnode_t *dn; |
428870ff | 892 | dbuf_dirty_record_t *dr; |
34dc7c2f BB |
893 | |
894 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
895 | ||
896 | if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY)) | |
897 | return; | |
898 | ||
899 | ASSERT(db->db_objset != NULL); | |
572e2857 BB |
900 | DB_DNODE_ENTER(db); |
901 | dn = DB_DNODE(db); | |
34dc7c2f BB |
902 | if (dn == NULL) { |
903 | ASSERT(db->db_parent == NULL); | |
904 | ASSERT(db->db_blkptr == NULL); | |
905 | } else { | |
906 | ASSERT3U(db->db.db_object, ==, dn->dn_object); | |
907 | ASSERT3P(db->db_objset, ==, dn->dn_objset); | |
908 | ASSERT3U(db->db_level, <, dn->dn_nlevels); | |
572e2857 BB |
909 | ASSERT(db->db_blkid == DMU_BONUS_BLKID || |
910 | db->db_blkid == DMU_SPILL_BLKID || | |
8951cb8d | 911 | !avl_is_empty(&dn->dn_dbufs)); |
34dc7c2f | 912 | } |
428870ff BB |
913 | if (db->db_blkid == DMU_BONUS_BLKID) { |
914 | ASSERT(dn != NULL); | |
915 | ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); | |
916 | ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID); | |
917 | } else if (db->db_blkid == DMU_SPILL_BLKID) { | |
34dc7c2f | 918 | ASSERT(dn != NULL); |
c99c9001 | 919 | ASSERT0(db->db.db_offset); |
34dc7c2f BB |
920 | } else { |
921 | ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size); | |
922 | } | |
923 | ||
428870ff BB |
924 | for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next) |
925 | ASSERT(dr->dr_dbuf == db); | |
926 | ||
927 | for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next) | |
928 | ASSERT(dr->dr_dbuf == db); | |
929 | ||
b128c09f BB |
930 | /* |
931 | * We can't assert that db_size matches dn_datablksz because it | |
932 | * can be momentarily different when another thread is doing | |
933 | * dnode_set_blksz(). | |
934 | */ | |
935 | if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) { | |
428870ff | 936 | dr = db->db_data_pending; |
b128c09f BB |
937 | /* |
938 | * It should only be modified in syncing context, so | |
939 | * make sure we only have one copy of the data. | |
940 | */ | |
941 | ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf); | |
34dc7c2f BB |
942 | } |
943 | ||
944 | /* verify db->db_blkptr */ | |
945 | if (db->db_blkptr) { | |
946 | if (db->db_parent == dn->dn_dbuf) { | |
947 | /* db is pointed to by the dnode */ | |
948 | /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */ | |
9babb374 | 949 | if (DMU_OBJECT_IS_SPECIAL(db->db.db_object)) |
34dc7c2f BB |
950 | ASSERT(db->db_parent == NULL); |
951 | else | |
952 | ASSERT(db->db_parent != NULL); | |
428870ff BB |
953 | if (db->db_blkid != DMU_SPILL_BLKID) |
954 | ASSERT3P(db->db_blkptr, ==, | |
955 | &dn->dn_phys->dn_blkptr[db->db_blkid]); | |
34dc7c2f BB |
956 | } else { |
957 | /* db is pointed to by an indirect block */ | |
1fde1e37 | 958 | ASSERTV(int epb = db->db_parent->db.db_size >> |
02730c33 | 959 | SPA_BLKPTRSHIFT); |
34dc7c2f BB |
960 | ASSERT3U(db->db_parent->db_level, ==, db->db_level+1); |
961 | ASSERT3U(db->db_parent->db.db_object, ==, | |
962 | db->db.db_object); | |
963 | /* | |
964 | * dnode_grow_indblksz() can make this fail if we don't | |
965 | * have the struct_rwlock. XXX indblksz no longer | |
966 | * grows. safe to do this now? | |
967 | */ | |
572e2857 | 968 | if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) { |
34dc7c2f BB |
969 | ASSERT3P(db->db_blkptr, ==, |
970 | ((blkptr_t *)db->db_parent->db.db_data + | |
971 | db->db_blkid % epb)); | |
972 | } | |
973 | } | |
974 | } | |
975 | if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) && | |
428870ff BB |
976 | (db->db_buf == NULL || db->db_buf->b_data) && |
977 | db->db.db_data && db->db_blkid != DMU_BONUS_BLKID && | |
34dc7c2f BB |
978 | db->db_state != DB_FILL && !dn->dn_free_txg) { |
979 | /* | |
980 | * If the blkptr isn't set but they have nonzero data, | |
981 | * it had better be dirty, otherwise we'll lose that | |
982 | * data when we evict this buffer. | |
bc77ba73 PD |
983 | * |
984 | * There is an exception to this rule for indirect blocks; in | |
985 | * this case, if the indirect block is a hole, we fill in a few | |
986 | * fields on each of the child blocks (importantly, birth time) | |
987 | * to prevent hole birth times from being lost when you | |
988 | * partially fill in a hole. | |
34dc7c2f BB |
989 | */ |
990 | if (db->db_dirtycnt == 0) { | |
bc77ba73 PD |
991 | if (db->db_level == 0) { |
992 | uint64_t *buf = db->db.db_data; | |
993 | int i; | |
34dc7c2f | 994 | |
bc77ba73 PD |
995 | for (i = 0; i < db->db.db_size >> 3; i++) { |
996 | ASSERT(buf[i] == 0); | |
997 | } | |
998 | } else { | |
bc77ba73 PD |
999 | blkptr_t *bps = db->db.db_data; |
1000 | ASSERT3U(1 << DB_DNODE(db)->dn_indblkshift, ==, | |
1001 | db->db.db_size); | |
1002 | /* | |
1003 | * We want to verify that all the blkptrs in the | |
1004 | * indirect block are holes, but we may have | |
1005 | * automatically set up a few fields for them. | |
1006 | * We iterate through each blkptr and verify | |
1007 | * they only have those fields set. | |
1008 | */ | |
1c27024e | 1009 | for (int i = 0; |
bc77ba73 PD |
1010 | i < db->db.db_size / sizeof (blkptr_t); |
1011 | i++) { | |
1012 | blkptr_t *bp = &bps[i]; | |
1013 | ASSERT(ZIO_CHECKSUM_IS_ZERO( | |
1014 | &bp->blk_cksum)); | |
1015 | ASSERT( | |
1016 | DVA_IS_EMPTY(&bp->blk_dva[0]) && | |
1017 | DVA_IS_EMPTY(&bp->blk_dva[1]) && | |
1018 | DVA_IS_EMPTY(&bp->blk_dva[2])); | |
1019 | ASSERT0(bp->blk_fill); | |
1020 | ASSERT0(bp->blk_pad[0]); | |
1021 | ASSERT0(bp->blk_pad[1]); | |
1022 | ASSERT(!BP_IS_EMBEDDED(bp)); | |
1023 | ASSERT(BP_IS_HOLE(bp)); | |
1024 | ASSERT0(bp->blk_phys_birth); | |
1025 | } | |
34dc7c2f BB |
1026 | } |
1027 | } | |
1028 | } | |
572e2857 | 1029 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
1030 | } |
1031 | #endif | |
1032 | ||
0c66c32d JG |
1033 | static void |
1034 | dbuf_clear_data(dmu_buf_impl_t *db) | |
1035 | { | |
1036 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
1037 | dbuf_evict_user(db); | |
d3c2ae1c | 1038 | ASSERT3P(db->db_buf, ==, NULL); |
0c66c32d JG |
1039 | db->db.db_data = NULL; |
1040 | if (db->db_state != DB_NOFILL) | |
1041 | db->db_state = DB_UNCACHED; | |
1042 | } | |
1043 | ||
34dc7c2f BB |
1044 | static void |
1045 | dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf) | |
1046 | { | |
1047 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
0c66c32d JG |
1048 | ASSERT(buf != NULL); |
1049 | ||
34dc7c2f | 1050 | db->db_buf = buf; |
0c66c32d JG |
1051 | ASSERT(buf->b_data != NULL); |
1052 | db->db.db_data = buf->b_data; | |
34dc7c2f BB |
1053 | } |
1054 | ||
428870ff BB |
1055 | /* |
1056 | * Loan out an arc_buf for read. Return the loaned arc_buf. | |
1057 | */ | |
1058 | arc_buf_t * | |
1059 | dbuf_loan_arcbuf(dmu_buf_impl_t *db) | |
1060 | { | |
1061 | arc_buf_t *abuf; | |
1062 | ||
d3c2ae1c | 1063 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
428870ff BB |
1064 | mutex_enter(&db->db_mtx); |
1065 | if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) { | |
1066 | int blksz = db->db.db_size; | |
b0bc7a84 | 1067 | spa_t *spa = db->db_objset->os_spa; |
572e2857 | 1068 | |
428870ff | 1069 | mutex_exit(&db->db_mtx); |
2aa34383 | 1070 | abuf = arc_loan_buf(spa, B_FALSE, blksz); |
428870ff BB |
1071 | bcopy(db->db.db_data, abuf->b_data, blksz); |
1072 | } else { | |
1073 | abuf = db->db_buf; | |
1074 | arc_loan_inuse_buf(abuf, db); | |
d3c2ae1c | 1075 | db->db_buf = NULL; |
0c66c32d | 1076 | dbuf_clear_data(db); |
428870ff BB |
1077 | mutex_exit(&db->db_mtx); |
1078 | } | |
1079 | return (abuf); | |
1080 | } | |
1081 | ||
fcff0f35 PD |
1082 | /* |
1083 | * Calculate which level n block references the data at the level 0 offset | |
1084 | * provided. | |
1085 | */ | |
34dc7c2f | 1086 | uint64_t |
031d7c2f | 1087 | dbuf_whichblock(const dnode_t *dn, const int64_t level, const uint64_t offset) |
34dc7c2f | 1088 | { |
fcff0f35 PD |
1089 | if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) { |
1090 | /* | |
1091 | * The level n blkid is equal to the level 0 blkid divided by | |
1092 | * the number of level 0s in a level n block. | |
1093 | * | |
1094 | * The level 0 blkid is offset >> datablkshift = | |
1095 | * offset / 2^datablkshift. | |
1096 | * | |
1097 | * The number of level 0s in a level n is the number of block | |
1098 | * pointers in an indirect block, raised to the power of level. | |
1099 | * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level = | |
1100 | * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)). | |
1101 | * | |
1102 | * Thus, the level n blkid is: offset / | |
1103 | * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT))) | |
1104 | * = offset / 2^(datablkshift + level * | |
1105 | * (indblkshift - SPA_BLKPTRSHIFT)) | |
1106 | * = offset >> (datablkshift + level * | |
1107 | * (indblkshift - SPA_BLKPTRSHIFT)) | |
1108 | */ | |
031d7c2f GN |
1109 | |
1110 | const unsigned exp = dn->dn_datablkshift + | |
1111 | level * (dn->dn_indblkshift - SPA_BLKPTRSHIFT); | |
1112 | ||
1113 | if (exp >= 8 * sizeof (offset)) { | |
1114 | /* This only happens on the highest indirection level */ | |
1115 | ASSERT3U(level, ==, dn->dn_nlevels - 1); | |
1116 | return (0); | |
1117 | } | |
1118 | ||
1119 | ASSERT3U(exp, <, 8 * sizeof (offset)); | |
1120 | ||
1121 | return (offset >> exp); | |
34dc7c2f BB |
1122 | } else { |
1123 | ASSERT3U(offset, <, dn->dn_datablksz); | |
1124 | return (0); | |
1125 | } | |
1126 | } | |
1127 | ||
1128 | static void | |
d4a72f23 TC |
1129 | dbuf_read_done(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp, |
1130 | arc_buf_t *buf, void *vdb) | |
34dc7c2f BB |
1131 | { |
1132 | dmu_buf_impl_t *db = vdb; | |
1133 | ||
1134 | mutex_enter(&db->db_mtx); | |
1135 | ASSERT3U(db->db_state, ==, DB_READ); | |
1136 | /* | |
1137 | * All reads are synchronous, so we must have a hold on the dbuf | |
1138 | */ | |
1139 | ASSERT(refcount_count(&db->db_holds) > 0); | |
1140 | ASSERT(db->db_buf == NULL); | |
1141 | ASSERT(db->db.db_data == NULL); | |
1142 | if (db->db_level == 0 && db->db_freed_in_flight) { | |
1143 | /* we were freed in flight; disregard any error */ | |
d4a72f23 TC |
1144 | if (buf == NULL) { |
1145 | buf = arc_alloc_buf(db->db_objset->os_spa, | |
1146 | db, DBUF_GET_BUFC_TYPE(db), db->db.db_size); | |
1147 | } | |
34dc7c2f BB |
1148 | arc_release(buf, db); |
1149 | bzero(buf->b_data, db->db.db_size); | |
1150 | arc_buf_freeze(buf); | |
1151 | db->db_freed_in_flight = FALSE; | |
1152 | dbuf_set_data(db, buf); | |
1153 | db->db_state = DB_CACHED; | |
d4a72f23 | 1154 | } else if (buf != NULL) { |
34dc7c2f BB |
1155 | dbuf_set_data(db, buf); |
1156 | db->db_state = DB_CACHED; | |
1157 | } else { | |
428870ff | 1158 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f | 1159 | ASSERT3P(db->db_buf, ==, NULL); |
34dc7c2f BB |
1160 | db->db_state = DB_UNCACHED; |
1161 | } | |
1162 | cv_broadcast(&db->db_changed); | |
428870ff | 1163 | dbuf_rele_and_unlock(db, NULL); |
34dc7c2f BB |
1164 | } |
1165 | ||
5f6d0b6f | 1166 | static int |
7f60329a | 1167 | dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) |
34dc7c2f | 1168 | { |
572e2857 | 1169 | dnode_t *dn; |
5dbd68a3 | 1170 | zbookmark_phys_t zb; |
2a432414 | 1171 | uint32_t aflags = ARC_FLAG_NOWAIT; |
b5256303 | 1172 | int err, zio_flags = 0; |
34dc7c2f | 1173 | |
572e2857 BB |
1174 | DB_DNODE_ENTER(db); |
1175 | dn = DB_DNODE(db); | |
34dc7c2f BB |
1176 | ASSERT(!refcount_is_zero(&db->db_holds)); |
1177 | /* We need the struct_rwlock to prevent db_blkptr from changing. */ | |
b128c09f | 1178 | ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); |
34dc7c2f BB |
1179 | ASSERT(MUTEX_HELD(&db->db_mtx)); |
1180 | ASSERT(db->db_state == DB_UNCACHED); | |
1181 | ASSERT(db->db_buf == NULL); | |
1182 | ||
428870ff | 1183 | if (db->db_blkid == DMU_BONUS_BLKID) { |
50c957f7 NB |
1184 | /* |
1185 | * The bonus length stored in the dnode may be less than | |
1186 | * the maximum available space in the bonus buffer. | |
1187 | */ | |
9babb374 | 1188 | int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen); |
50c957f7 | 1189 | int max_bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots); |
b5256303 TC |
1190 | arc_buf_t *dn_buf = (dn->dn_dbuf != NULL) ? |
1191 | dn->dn_dbuf->db_buf : NULL; | |
1192 | ||
1193 | /* if the underlying dnode block is encrypted, decrypt it */ | |
1194 | if (dn_buf != NULL && dn->dn_objset->os_encrypted && | |
1195 | DMU_OT_IS_ENCRYPTED(dn->dn_bonustype) && | |
1196 | (flags & DB_RF_NO_DECRYPT) == 0 && | |
1197 | arc_is_encrypted(dn_buf)) { | |
a2c2ed1b TC |
1198 | SET_BOOKMARK(&zb, dmu_objset_id(db->db_objset), |
1199 | DMU_META_DNODE_OBJECT, 0, dn->dn_dbuf->db_blkid); | |
b5256303 | 1200 | err = arc_untransform(dn_buf, dn->dn_objset->os_spa, |
a2c2ed1b | 1201 | &zb, B_TRUE); |
b5256303 TC |
1202 | if (err != 0) { |
1203 | DB_DNODE_EXIT(db); | |
1204 | mutex_exit(&db->db_mtx); | |
1205 | return (err); | |
1206 | } | |
1207 | } | |
34dc7c2f BB |
1208 | |
1209 | ASSERT3U(bonuslen, <=, db->db.db_size); | |
a3fd9d9e | 1210 | db->db.db_data = kmem_alloc(max_bonuslen, KM_SLEEP); |
25458cbe | 1211 | arc_space_consume(max_bonuslen, ARC_SPACE_BONUS); |
50c957f7 NB |
1212 | if (bonuslen < max_bonuslen) |
1213 | bzero(db->db.db_data, max_bonuslen); | |
9babb374 BB |
1214 | if (bonuslen) |
1215 | bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen); | |
572e2857 | 1216 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
1217 | db->db_state = DB_CACHED; |
1218 | mutex_exit(&db->db_mtx); | |
5f6d0b6f | 1219 | return (0); |
34dc7c2f BB |
1220 | } |
1221 | ||
b128c09f BB |
1222 | /* |
1223 | * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync() | |
1224 | * processes the delete record and clears the bp while we are waiting | |
1225 | * for the dn_mtx (resulting in a "no" from block_freed). | |
1226 | */ | |
1227 | if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) || | |
1228 | (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) || | |
1229 | BP_IS_HOLE(db->db_blkptr)))) { | |
34dc7c2f BB |
1230 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); |
1231 | ||
2aa34383 DK |
1232 | dbuf_set_data(db, arc_alloc_buf(db->db_objset->os_spa, db, type, |
1233 | db->db.db_size)); | |
34dc7c2f | 1234 | bzero(db->db.db_data, db->db.db_size); |
bc77ba73 PD |
1235 | |
1236 | if (db->db_blkptr != NULL && db->db_level > 0 && | |
1237 | BP_IS_HOLE(db->db_blkptr) && | |
1238 | db->db_blkptr->blk_birth != 0) { | |
1239 | blkptr_t *bps = db->db.db_data; | |
1c27024e | 1240 | for (int i = 0; i < ((1 << |
bc77ba73 PD |
1241 | DB_DNODE(db)->dn_indblkshift) / sizeof (blkptr_t)); |
1242 | i++) { | |
1243 | blkptr_t *bp = &bps[i]; | |
1244 | ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, | |
1245 | 1 << dn->dn_indblkshift); | |
1246 | BP_SET_LSIZE(bp, | |
1247 | BP_GET_LEVEL(db->db_blkptr) == 1 ? | |
1248 | dn->dn_datablksz : | |
1249 | BP_GET_LSIZE(db->db_blkptr)); | |
1250 | BP_SET_TYPE(bp, BP_GET_TYPE(db->db_blkptr)); | |
1251 | BP_SET_LEVEL(bp, | |
1252 | BP_GET_LEVEL(db->db_blkptr) - 1); | |
1253 | BP_SET_BIRTH(bp, db->db_blkptr->blk_birth, 0); | |
1254 | } | |
1255 | } | |
1256 | DB_DNODE_EXIT(db); | |
34dc7c2f | 1257 | db->db_state = DB_CACHED; |
34dc7c2f | 1258 | mutex_exit(&db->db_mtx); |
5f6d0b6f | 1259 | return (0); |
34dc7c2f BB |
1260 | } |
1261 | ||
572e2857 BB |
1262 | DB_DNODE_EXIT(db); |
1263 | ||
34dc7c2f BB |
1264 | db->db_state = DB_READ; |
1265 | mutex_exit(&db->db_mtx); | |
1266 | ||
b128c09f | 1267 | if (DBUF_IS_L2CACHEABLE(db)) |
2a432414 | 1268 | aflags |= ARC_FLAG_L2CACHE; |
b128c09f | 1269 | |
a2c2ed1b | 1270 | SET_BOOKMARK(&zb, dmu_objset_id(db->db_objset), |
428870ff | 1271 | db->db.db_object, db->db_level, db->db_blkid); |
34dc7c2f | 1272 | |
b5256303 TC |
1273 | /* |
1274 | * All bps of an encrypted os should have the encryption bit set. | |
1275 | * If this is not true it indicates tampering and we report an error. | |
1276 | */ | |
1277 | if (db->db_objset->os_encrypted && !BP_USES_CRYPT(db->db_blkptr)) { | |
1278 | spa_log_error(db->db_objset->os_spa, &zb); | |
1279 | zfs_panic_recover("unencrypted block in encrypted " | |
1280 | "object set %llu", dmu_objset_id(db->db_objset)); | |
1281 | return (SET_ERROR(EIO)); | |
1282 | } | |
1283 | ||
34dc7c2f | 1284 | dbuf_add_ref(db, NULL); |
b128c09f | 1285 | |
b5256303 TC |
1286 | zio_flags = (flags & DB_RF_CANFAIL) ? |
1287 | ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED; | |
1288 | ||
1289 | if ((flags & DB_RF_NO_DECRYPT) && BP_IS_PROTECTED(db->db_blkptr)) | |
1290 | zio_flags |= ZIO_FLAG_RAW; | |
1291 | ||
5f6d0b6f | 1292 | err = arc_read(zio, db->db_objset->os_spa, db->db_blkptr, |
b5256303 | 1293 | dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ, zio_flags, |
34dc7c2f | 1294 | &aflags, &zb); |
5f6d0b6f | 1295 | |
da8d5748 | 1296 | return (err); |
34dc7c2f BB |
1297 | } |
1298 | ||
2aa34383 DK |
1299 | /* |
1300 | * This is our just-in-time copy function. It makes a copy of buffers that | |
1301 | * have been modified in a previous transaction group before we access them in | |
1302 | * the current active group. | |
1303 | * | |
1304 | * This function is used in three places: when we are dirtying a buffer for the | |
1305 | * first time in a txg, when we are freeing a range in a dnode that includes | |
1306 | * this buffer, and when we are accessing a buffer which was received compressed | |
1307 | * and later referenced in a WRITE_BYREF record. | |
1308 | * | |
1309 | * Note that when we are called from dbuf_free_range() we do not put a hold on | |
1310 | * the buffer, we just traverse the active dbuf list for the dnode. | |
1311 | */ | |
1312 | static void | |
1313 | dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg) | |
1314 | { | |
1315 | dbuf_dirty_record_t *dr = db->db_last_dirty; | |
1316 | ||
1317 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
1318 | ASSERT(db->db.db_data != NULL); | |
1319 | ASSERT(db->db_level == 0); | |
1320 | ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT); | |
1321 | ||
1322 | if (dr == NULL || | |
1323 | (dr->dt.dl.dr_data != | |
1324 | ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf))) | |
1325 | return; | |
1326 | ||
1327 | /* | |
1328 | * If the last dirty record for this dbuf has not yet synced | |
1329 | * and its referencing the dbuf data, either: | |
1330 | * reset the reference to point to a new copy, | |
1331 | * or (if there a no active holders) | |
1332 | * just null out the current db_data pointer. | |
1333 | */ | |
4807c0ba | 1334 | ASSERT3U(dr->dr_txg, >=, txg - 2); |
2aa34383 | 1335 | if (db->db_blkid == DMU_BONUS_BLKID) { |
2aa34383 DK |
1336 | dnode_t *dn = DB_DNODE(db); |
1337 | int bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots); | |
a3fd9d9e | 1338 | dr->dt.dl.dr_data = kmem_alloc(bonuslen, KM_SLEEP); |
2aa34383 DK |
1339 | arc_space_consume(bonuslen, ARC_SPACE_BONUS); |
1340 | bcopy(db->db.db_data, dr->dt.dl.dr_data, bonuslen); | |
1341 | } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) { | |
b5256303 | 1342 | dnode_t *dn = DB_DNODE(db); |
2aa34383 DK |
1343 | int size = arc_buf_size(db->db_buf); |
1344 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); | |
1345 | spa_t *spa = db->db_objset->os_spa; | |
1346 | enum zio_compress compress_type = | |
1347 | arc_get_compression(db->db_buf); | |
1348 | ||
b5256303 TC |
1349 | if (arc_is_encrypted(db->db_buf)) { |
1350 | boolean_t byteorder; | |
1351 | uint8_t salt[ZIO_DATA_SALT_LEN]; | |
1352 | uint8_t iv[ZIO_DATA_IV_LEN]; | |
1353 | uint8_t mac[ZIO_DATA_MAC_LEN]; | |
1354 | ||
1355 | arc_get_raw_params(db->db_buf, &byteorder, salt, | |
1356 | iv, mac); | |
1357 | dr->dt.dl.dr_data = arc_alloc_raw_buf(spa, db, | |
1358 | dmu_objset_id(dn->dn_objset), byteorder, salt, iv, | |
1359 | mac, dn->dn_type, size, arc_buf_lsize(db->db_buf), | |
1360 | compress_type); | |
1361 | } else if (compress_type != ZIO_COMPRESS_OFF) { | |
2aa34383 DK |
1362 | ASSERT3U(type, ==, ARC_BUFC_DATA); |
1363 | dr->dt.dl.dr_data = arc_alloc_compressed_buf(spa, db, | |
1364 | size, arc_buf_lsize(db->db_buf), compress_type); | |
b5256303 TC |
1365 | } else { |
1366 | dr->dt.dl.dr_data = arc_alloc_buf(spa, db, type, size); | |
2aa34383 DK |
1367 | } |
1368 | bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size); | |
1369 | } else { | |
1370 | db->db_buf = NULL; | |
1371 | dbuf_clear_data(db); | |
1372 | } | |
1373 | } | |
1374 | ||
34dc7c2f BB |
1375 | int |
1376 | dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) | |
1377 | { | |
1378 | int err = 0; | |
b0bc7a84 | 1379 | boolean_t prefetch; |
572e2857 | 1380 | dnode_t *dn; |
34dc7c2f BB |
1381 | |
1382 | /* | |
1383 | * We don't have to hold the mutex to check db_state because it | |
1384 | * can't be freed while we have a hold on the buffer. | |
1385 | */ | |
1386 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
1387 | ||
b128c09f | 1388 | if (db->db_state == DB_NOFILL) |
2e528b49 | 1389 | return (SET_ERROR(EIO)); |
b128c09f | 1390 | |
572e2857 BB |
1391 | DB_DNODE_ENTER(db); |
1392 | dn = DB_DNODE(db); | |
34dc7c2f | 1393 | if ((flags & DB_RF_HAVESTRUCT) == 0) |
572e2857 | 1394 | rw_enter(&dn->dn_struct_rwlock, RW_READER); |
34dc7c2f | 1395 | |
428870ff | 1396 | prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && |
572e2857 | 1397 | (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL && |
b128c09f | 1398 | DBUF_IS_CACHEABLE(db); |
34dc7c2f BB |
1399 | |
1400 | mutex_enter(&db->db_mtx); | |
1401 | if (db->db_state == DB_CACHED) { | |
b5256303 TC |
1402 | spa_t *spa = dn->dn_objset->os_spa; |
1403 | ||
2aa34383 | 1404 | /* |
b5256303 TC |
1405 | * If the arc buf is compressed or encrypted, we need to |
1406 | * untransform it to read the data. This could happen during | |
1407 | * the "zfs receive" of a stream which is deduplicated and | |
1408 | * either raw or compressed. We do not need to do this if the | |
1409 | * caller wants raw encrypted data. | |
2aa34383 | 1410 | */ |
b5256303 TC |
1411 | if (db->db_buf != NULL && (flags & DB_RF_NO_DECRYPT) == 0 && |
1412 | (arc_is_encrypted(db->db_buf) || | |
1413 | arc_get_compression(db->db_buf) != ZIO_COMPRESS_OFF)) { | |
a2c2ed1b TC |
1414 | zbookmark_phys_t zb; |
1415 | ||
1416 | SET_BOOKMARK(&zb, dmu_objset_id(db->db_objset), | |
1417 | db->db.db_object, db->db_level, db->db_blkid); | |
b5256303 | 1418 | dbuf_fix_old_data(db, spa_syncing_txg(spa)); |
a2c2ed1b | 1419 | err = arc_untransform(db->db_buf, spa, &zb, B_FALSE); |
2aa34383 DK |
1420 | dbuf_set_data(db, db->db_buf); |
1421 | } | |
34dc7c2f BB |
1422 | mutex_exit(&db->db_mtx); |
1423 | if (prefetch) | |
755065f3 | 1424 | dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE); |
34dc7c2f | 1425 | if ((flags & DB_RF_HAVESTRUCT) == 0) |
572e2857 BB |
1426 | rw_exit(&dn->dn_struct_rwlock); |
1427 | DB_DNODE_EXIT(db); | |
5e021f56 | 1428 | DBUF_STAT_BUMP(hash_hits); |
34dc7c2f | 1429 | } else if (db->db_state == DB_UNCACHED) { |
572e2857 | 1430 | spa_t *spa = dn->dn_objset->os_spa; |
a0043383 | 1431 | boolean_t need_wait = B_FALSE; |
572e2857 | 1432 | |
b0319c1f | 1433 | if (zio == NULL && |
a0043383 | 1434 | db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) { |
572e2857 | 1435 | zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); |
a0043383 MA |
1436 | need_wait = B_TRUE; |
1437 | } | |
7f60329a | 1438 | err = dbuf_read_impl(db, zio, flags); |
34dc7c2f BB |
1439 | |
1440 | /* dbuf_read_impl has dropped db_mtx for us */ | |
1441 | ||
5f6d0b6f | 1442 | if (!err && prefetch) |
755065f3 | 1443 | dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE); |
34dc7c2f BB |
1444 | |
1445 | if ((flags & DB_RF_HAVESTRUCT) == 0) | |
572e2857 BB |
1446 | rw_exit(&dn->dn_struct_rwlock); |
1447 | DB_DNODE_EXIT(db); | |
5e021f56 | 1448 | DBUF_STAT_BUMP(hash_misses); |
34dc7c2f | 1449 | |
a0043383 | 1450 | if (!err && need_wait) |
34dc7c2f BB |
1451 | err = zio_wait(zio); |
1452 | } else { | |
e49f1e20 WA |
1453 | /* |
1454 | * Another reader came in while the dbuf was in flight | |
1455 | * between UNCACHED and CACHED. Either a writer will finish | |
1456 | * writing the buffer (sending the dbuf to CACHED) or the | |
1457 | * first reader's request will reach the read_done callback | |
1458 | * and send the dbuf to CACHED. Otherwise, a failure | |
1459 | * occurred and the dbuf went to UNCACHED. | |
1460 | */ | |
34dc7c2f BB |
1461 | mutex_exit(&db->db_mtx); |
1462 | if (prefetch) | |
755065f3 | 1463 | dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE); |
34dc7c2f | 1464 | if ((flags & DB_RF_HAVESTRUCT) == 0) |
572e2857 BB |
1465 | rw_exit(&dn->dn_struct_rwlock); |
1466 | DB_DNODE_EXIT(db); | |
5e021f56 | 1467 | DBUF_STAT_BUMP(hash_misses); |
34dc7c2f | 1468 | |
e49f1e20 | 1469 | /* Skip the wait per the caller's request. */ |
34dc7c2f BB |
1470 | mutex_enter(&db->db_mtx); |
1471 | if ((flags & DB_RF_NEVERWAIT) == 0) { | |
1472 | while (db->db_state == DB_READ || | |
1473 | db->db_state == DB_FILL) { | |
1474 | ASSERT(db->db_state == DB_READ || | |
1475 | (flags & DB_RF_HAVESTRUCT) == 0); | |
64dbba36 AL |
1476 | DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *, |
1477 | db, zio_t *, zio); | |
34dc7c2f BB |
1478 | cv_wait(&db->db_changed, &db->db_mtx); |
1479 | } | |
1480 | if (db->db_state == DB_UNCACHED) | |
2e528b49 | 1481 | err = SET_ERROR(EIO); |
34dc7c2f BB |
1482 | } |
1483 | mutex_exit(&db->db_mtx); | |
1484 | } | |
1485 | ||
34dc7c2f BB |
1486 | return (err); |
1487 | } | |
1488 | ||
1489 | static void | |
1490 | dbuf_noread(dmu_buf_impl_t *db) | |
1491 | { | |
1492 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
428870ff | 1493 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f BB |
1494 | mutex_enter(&db->db_mtx); |
1495 | while (db->db_state == DB_READ || db->db_state == DB_FILL) | |
1496 | cv_wait(&db->db_changed, &db->db_mtx); | |
1497 | if (db->db_state == DB_UNCACHED) { | |
1498 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); | |
b0bc7a84 | 1499 | spa_t *spa = db->db_objset->os_spa; |
34dc7c2f BB |
1500 | |
1501 | ASSERT(db->db_buf == NULL); | |
1502 | ASSERT(db->db.db_data == NULL); | |
2aa34383 | 1503 | dbuf_set_data(db, arc_alloc_buf(spa, db, type, db->db.db_size)); |
34dc7c2f | 1504 | db->db_state = DB_FILL; |
b128c09f | 1505 | } else if (db->db_state == DB_NOFILL) { |
0c66c32d | 1506 | dbuf_clear_data(db); |
34dc7c2f BB |
1507 | } else { |
1508 | ASSERT3U(db->db_state, ==, DB_CACHED); | |
1509 | } | |
1510 | mutex_exit(&db->db_mtx); | |
1511 | } | |
1512 | ||
34dc7c2f BB |
1513 | void |
1514 | dbuf_unoverride(dbuf_dirty_record_t *dr) | |
1515 | { | |
1516 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
428870ff | 1517 | blkptr_t *bp = &dr->dt.dl.dr_overridden_by; |
34dc7c2f BB |
1518 | uint64_t txg = dr->dr_txg; |
1519 | ||
1520 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
00710365 AS |
1521 | /* |
1522 | * This assert is valid because dmu_sync() expects to be called by | |
1523 | * a zilog's get_data while holding a range lock. This call only | |
1524 | * comes from dbuf_dirty() callers who must also hold a range lock. | |
1525 | */ | |
34dc7c2f BB |
1526 | ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC); |
1527 | ASSERT(db->db_level == 0); | |
1528 | ||
428870ff | 1529 | if (db->db_blkid == DMU_BONUS_BLKID || |
34dc7c2f BB |
1530 | dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN) |
1531 | return; | |
1532 | ||
428870ff BB |
1533 | ASSERT(db->db_data_pending != dr); |
1534 | ||
34dc7c2f | 1535 | /* free this block */ |
b0bc7a84 MG |
1536 | if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite) |
1537 | zio_free(db->db_objset->os_spa, txg, bp); | |
428870ff | 1538 | |
34dc7c2f | 1539 | dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; |
03c6040b | 1540 | dr->dt.dl.dr_nopwrite = B_FALSE; |
0c03d21a | 1541 | dr->dt.dl.dr_has_raw_params = B_FALSE; |
03c6040b | 1542 | |
34dc7c2f BB |
1543 | /* |
1544 | * Release the already-written buffer, so we leave it in | |
1545 | * a consistent dirty state. Note that all callers are | |
1546 | * modifying the buffer, so they will immediately do | |
1547 | * another (redundant) arc_release(). Therefore, leave | |
1548 | * the buf thawed to save the effort of freezing & | |
1549 | * immediately re-thawing it. | |
1550 | */ | |
1551 | arc_release(dr->dt.dl.dr_data, db); | |
1552 | } | |
1553 | ||
b128c09f BB |
1554 | /* |
1555 | * Evict (if its unreferenced) or clear (if its referenced) any level-0 | |
1556 | * data blocks in the free range, so that any future readers will find | |
b0bc7a84 | 1557 | * empty blocks. |
b128c09f | 1558 | */ |
34dc7c2f | 1559 | void |
8951cb8d AR |
1560 | dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid, |
1561 | dmu_tx_t *tx) | |
34dc7c2f | 1562 | { |
0c66c32d JG |
1563 | dmu_buf_impl_t *db_search; |
1564 | dmu_buf_impl_t *db, *db_next; | |
34dc7c2f | 1565 | uint64_t txg = tx->tx_txg; |
8951cb8d | 1566 | avl_index_t where; |
8951cb8d | 1567 | |
9c9531cb GM |
1568 | if (end_blkid > dn->dn_maxblkid && |
1569 | !(start_blkid == DMU_SPILL_BLKID || end_blkid == DMU_SPILL_BLKID)) | |
8951cb8d AR |
1570 | end_blkid = dn->dn_maxblkid; |
1571 | dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid); | |
34dc7c2f | 1572 | |
0c66c32d | 1573 | db_search = kmem_alloc(sizeof (dmu_buf_impl_t), KM_SLEEP); |
8951cb8d AR |
1574 | db_search->db_level = 0; |
1575 | db_search->db_blkid = start_blkid; | |
9925c28c | 1576 | db_search->db_state = DB_SEARCH; |
ea97f8ce | 1577 | |
b663a23d | 1578 | mutex_enter(&dn->dn_dbufs_mtx); |
8951cb8d AR |
1579 | db = avl_find(&dn->dn_dbufs, db_search, &where); |
1580 | ASSERT3P(db, ==, NULL); | |
9c9531cb | 1581 | |
8951cb8d AR |
1582 | db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER); |
1583 | ||
1584 | for (; db != NULL; db = db_next) { | |
1585 | db_next = AVL_NEXT(&dn->dn_dbufs, db); | |
428870ff | 1586 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
b128c09f | 1587 | |
8951cb8d AR |
1588 | if (db->db_level != 0 || db->db_blkid > end_blkid) { |
1589 | break; | |
1590 | } | |
1591 | ASSERT3U(db->db_blkid, >=, start_blkid); | |
34dc7c2f BB |
1592 | |
1593 | /* found a level 0 buffer in the range */ | |
13fe0198 MA |
1594 | mutex_enter(&db->db_mtx); |
1595 | if (dbuf_undirty(db, tx)) { | |
1596 | /* mutex has been dropped and dbuf destroyed */ | |
34dc7c2f | 1597 | continue; |
13fe0198 | 1598 | } |
34dc7c2f | 1599 | |
34dc7c2f | 1600 | if (db->db_state == DB_UNCACHED || |
b128c09f | 1601 | db->db_state == DB_NOFILL || |
34dc7c2f BB |
1602 | db->db_state == DB_EVICTING) { |
1603 | ASSERT(db->db.db_data == NULL); | |
1604 | mutex_exit(&db->db_mtx); | |
1605 | continue; | |
1606 | } | |
1607 | if (db->db_state == DB_READ || db->db_state == DB_FILL) { | |
1608 | /* will be handled in dbuf_read_done or dbuf_rele */ | |
1609 | db->db_freed_in_flight = TRUE; | |
1610 | mutex_exit(&db->db_mtx); | |
1611 | continue; | |
1612 | } | |
1613 | if (refcount_count(&db->db_holds) == 0) { | |
1614 | ASSERT(db->db_buf); | |
d3c2ae1c | 1615 | dbuf_destroy(db); |
34dc7c2f BB |
1616 | continue; |
1617 | } | |
1618 | /* The dbuf is referenced */ | |
1619 | ||
1620 | if (db->db_last_dirty != NULL) { | |
1621 | dbuf_dirty_record_t *dr = db->db_last_dirty; | |
1622 | ||
1623 | if (dr->dr_txg == txg) { | |
1624 | /* | |
1625 | * This buffer is "in-use", re-adjust the file | |
1626 | * size to reflect that this buffer may | |
1627 | * contain new data when we sync. | |
1628 | */ | |
428870ff BB |
1629 | if (db->db_blkid != DMU_SPILL_BLKID && |
1630 | db->db_blkid > dn->dn_maxblkid) | |
34dc7c2f BB |
1631 | dn->dn_maxblkid = db->db_blkid; |
1632 | dbuf_unoverride(dr); | |
1633 | } else { | |
1634 | /* | |
1635 | * This dbuf is not dirty in the open context. | |
1636 | * Either uncache it (if its not referenced in | |
1637 | * the open context) or reset its contents to | |
1638 | * empty. | |
1639 | */ | |
1640 | dbuf_fix_old_data(db, txg); | |
1641 | } | |
1642 | } | |
1643 | /* clear the contents if its cached */ | |
1644 | if (db->db_state == DB_CACHED) { | |
1645 | ASSERT(db->db.db_data != NULL); | |
1646 | arc_release(db->db_buf, db); | |
1647 | bzero(db->db.db_data, db->db.db_size); | |
1648 | arc_buf_freeze(db->db_buf); | |
1649 | } | |
1650 | ||
1651 | mutex_exit(&db->db_mtx); | |
1652 | } | |
8951cb8d | 1653 | |
8951cb8d | 1654 | kmem_free(db_search, sizeof (dmu_buf_impl_t)); |
34dc7c2f BB |
1655 | mutex_exit(&dn->dn_dbufs_mtx); |
1656 | } | |
1657 | ||
34dc7c2f BB |
1658 | void |
1659 | dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx) | |
1660 | { | |
1661 | arc_buf_t *buf, *obuf; | |
1662 | int osize = db->db.db_size; | |
1663 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); | |
572e2857 | 1664 | dnode_t *dn; |
34dc7c2f | 1665 | |
428870ff | 1666 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f | 1667 | |
572e2857 BB |
1668 | DB_DNODE_ENTER(db); |
1669 | dn = DB_DNODE(db); | |
1670 | ||
34dc7c2f | 1671 | /* XXX does *this* func really need the lock? */ |
572e2857 | 1672 | ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); |
34dc7c2f BB |
1673 | |
1674 | /* | |
b0bc7a84 | 1675 | * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held |
34dc7c2f BB |
1676 | * is OK, because there can be no other references to the db |
1677 | * when we are changing its size, so no concurrent DB_FILL can | |
1678 | * be happening. | |
1679 | */ | |
1680 | /* | |
1681 | * XXX we should be doing a dbuf_read, checking the return | |
1682 | * value and returning that up to our callers | |
1683 | */ | |
b0bc7a84 | 1684 | dmu_buf_will_dirty(&db->db, tx); |
34dc7c2f BB |
1685 | |
1686 | /* create the data buffer for the new block */ | |
2aa34383 | 1687 | buf = arc_alloc_buf(dn->dn_objset->os_spa, db, type, size); |
34dc7c2f BB |
1688 | |
1689 | /* copy old block data to the new block */ | |
1690 | obuf = db->db_buf; | |
1691 | bcopy(obuf->b_data, buf->b_data, MIN(osize, size)); | |
1692 | /* zero the remainder */ | |
1693 | if (size > osize) | |
1694 | bzero((uint8_t *)buf->b_data + osize, size - osize); | |
1695 | ||
1696 | mutex_enter(&db->db_mtx); | |
1697 | dbuf_set_data(db, buf); | |
d3c2ae1c | 1698 | arc_buf_destroy(obuf, db); |
34dc7c2f BB |
1699 | db->db.db_size = size; |
1700 | ||
1701 | if (db->db_level == 0) { | |
1702 | ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); | |
1703 | db->db_last_dirty->dt.dl.dr_data = buf; | |
1704 | } | |
1705 | mutex_exit(&db->db_mtx); | |
1706 | ||
3ec3bc21 | 1707 | dmu_objset_willuse_space(dn->dn_objset, size - osize, tx); |
572e2857 | 1708 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
1709 | } |
1710 | ||
428870ff BB |
1711 | void |
1712 | dbuf_release_bp(dmu_buf_impl_t *db) | |
1713 | { | |
b0bc7a84 | 1714 | ASSERTV(objset_t *os = db->db_objset); |
428870ff BB |
1715 | |
1716 | ASSERT(dsl_pool_sync_context(dmu_objset_pool(os))); | |
1717 | ASSERT(arc_released(os->os_phys_buf) || | |
1718 | list_link_active(&os->os_dsl_dataset->ds_synced_link)); | |
1719 | ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf)); | |
1720 | ||
294f6806 | 1721 | (void) arc_release(db->db_buf, db); |
428870ff BB |
1722 | } |
1723 | ||
5a28a973 MA |
1724 | /* |
1725 | * We already have a dirty record for this TXG, and we are being | |
1726 | * dirtied again. | |
1727 | */ | |
1728 | static void | |
1729 | dbuf_redirty(dbuf_dirty_record_t *dr) | |
1730 | { | |
1731 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
1732 | ||
1733 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
1734 | ||
1735 | if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) { | |
1736 | /* | |
1737 | * If this buffer has already been written out, | |
1738 | * we now need to reset its state. | |
1739 | */ | |
1740 | dbuf_unoverride(dr); | |
1741 | if (db->db.db_object != DMU_META_DNODE_OBJECT && | |
1742 | db->db_state != DB_NOFILL) { | |
1743 | /* Already released on initial dirty, so just thaw. */ | |
1744 | ASSERT(arc_released(db->db_buf)); | |
1745 | arc_buf_thaw(db->db_buf); | |
1746 | } | |
1747 | } | |
1748 | } | |
1749 | ||
34dc7c2f BB |
1750 | dbuf_dirty_record_t * |
1751 | dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) | |
1752 | { | |
572e2857 BB |
1753 | dnode_t *dn; |
1754 | objset_t *os; | |
34dc7c2f BB |
1755 | dbuf_dirty_record_t **drp, *dr; |
1756 | int drop_struct_lock = FALSE; | |
1757 | int txgoff = tx->tx_txg & TXG_MASK; | |
1758 | ||
1759 | ASSERT(tx->tx_txg != 0); | |
1760 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
1761 | DMU_TX_DIRTY_BUF(tx, db); | |
1762 | ||
572e2857 BB |
1763 | DB_DNODE_ENTER(db); |
1764 | dn = DB_DNODE(db); | |
34dc7c2f BB |
1765 | /* |
1766 | * Shouldn't dirty a regular buffer in syncing context. Private | |
1767 | * objects may be dirtied in syncing context, but only if they | |
1768 | * were already pre-dirtied in open context. | |
34dc7c2f | 1769 | */ |
cc9bb3e5 GM |
1770 | #ifdef DEBUG |
1771 | if (dn->dn_objset->os_dsl_dataset != NULL) { | |
1772 | rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, | |
1773 | RW_READER, FTAG); | |
1774 | } | |
34dc7c2f BB |
1775 | ASSERT(!dmu_tx_is_syncing(tx) || |
1776 | BP_IS_HOLE(dn->dn_objset->os_rootbp) || | |
9babb374 BB |
1777 | DMU_OBJECT_IS_SPECIAL(dn->dn_object) || |
1778 | dn->dn_objset->os_dsl_dataset == NULL); | |
cc9bb3e5 GM |
1779 | if (dn->dn_objset->os_dsl_dataset != NULL) |
1780 | rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, FTAG); | |
1781 | #endif | |
34dc7c2f BB |
1782 | /* |
1783 | * We make this assert for private objects as well, but after we | |
1784 | * check if we're already dirty. They are allowed to re-dirty | |
1785 | * in syncing context. | |
1786 | */ | |
1787 | ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || | |
1788 | dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == | |
1789 | (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); | |
1790 | ||
1791 | mutex_enter(&db->db_mtx); | |
1792 | /* | |
1793 | * XXX make this true for indirects too? The problem is that | |
1794 | * transactions created with dmu_tx_create_assigned() from | |
1795 | * syncing context don't bother holding ahead. | |
1796 | */ | |
1797 | ASSERT(db->db_level != 0 || | |
b128c09f BB |
1798 | db->db_state == DB_CACHED || db->db_state == DB_FILL || |
1799 | db->db_state == DB_NOFILL); | |
34dc7c2f BB |
1800 | |
1801 | mutex_enter(&dn->dn_mtx); | |
1802 | /* | |
1803 | * Don't set dirtyctx to SYNC if we're just modifying this as we | |
1804 | * initialize the objset. | |
1805 | */ | |
cc9bb3e5 GM |
1806 | if (dn->dn_dirtyctx == DN_UNDIRTIED) { |
1807 | if (dn->dn_objset->os_dsl_dataset != NULL) { | |
1808 | rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, | |
1809 | RW_READER, FTAG); | |
1810 | } | |
1811 | if (!BP_IS_HOLE(dn->dn_objset->os_rootbp)) { | |
1812 | dn->dn_dirtyctx = (dmu_tx_is_syncing(tx) ? | |
1813 | DN_DIRTY_SYNC : DN_DIRTY_OPEN); | |
1814 | ASSERT(dn->dn_dirtyctx_firstset == NULL); | |
1815 | dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP); | |
1816 | } | |
1817 | if (dn->dn_objset->os_dsl_dataset != NULL) { | |
1818 | rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, | |
1819 | FTAG); | |
1820 | } | |
34dc7c2f | 1821 | } |
edc1e713 TC |
1822 | |
1823 | if (tx->tx_txg > dn->dn_dirty_txg) | |
1824 | dn->dn_dirty_txg = tx->tx_txg; | |
34dc7c2f BB |
1825 | mutex_exit(&dn->dn_mtx); |
1826 | ||
428870ff BB |
1827 | if (db->db_blkid == DMU_SPILL_BLKID) |
1828 | dn->dn_have_spill = B_TRUE; | |
1829 | ||
34dc7c2f BB |
1830 | /* |
1831 | * If this buffer is already dirty, we're done. | |
1832 | */ | |
1833 | drp = &db->db_last_dirty; | |
1834 | ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg || | |
1835 | db->db.db_object == DMU_META_DNODE_OBJECT); | |
1836 | while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg) | |
1837 | drp = &dr->dr_next; | |
1838 | if (dr && dr->dr_txg == tx->tx_txg) { | |
572e2857 BB |
1839 | DB_DNODE_EXIT(db); |
1840 | ||
5a28a973 | 1841 | dbuf_redirty(dr); |
34dc7c2f BB |
1842 | mutex_exit(&db->db_mtx); |
1843 | return (dr); | |
1844 | } | |
1845 | ||
1846 | /* | |
1847 | * Only valid if not already dirty. | |
1848 | */ | |
9babb374 BB |
1849 | ASSERT(dn->dn_object == 0 || |
1850 | dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == | |
34dc7c2f BB |
1851 | (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); |
1852 | ||
1853 | ASSERT3U(dn->dn_nlevels, >, db->db_level); | |
34dc7c2f BB |
1854 | |
1855 | /* | |
1856 | * We should only be dirtying in syncing context if it's the | |
9babb374 BB |
1857 | * mos or we're initializing the os or it's a special object. |
1858 | * However, we are allowed to dirty in syncing context provided | |
1859 | * we already dirtied it in open context. Hence we must make | |
1860 | * this assertion only if we're not already dirty. | |
34dc7c2f | 1861 | */ |
572e2857 | 1862 | os = dn->dn_objset; |
3b7f360c | 1863 | VERIFY3U(tx->tx_txg, <=, spa_final_dirty_txg(os->os_spa)); |
cc9bb3e5 GM |
1864 | #ifdef DEBUG |
1865 | if (dn->dn_objset->os_dsl_dataset != NULL) | |
1866 | rrw_enter(&os->os_dsl_dataset->ds_bp_rwlock, RW_READER, FTAG); | |
9babb374 BB |
1867 | ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) || |
1868 | os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp)); | |
cc9bb3e5 GM |
1869 | if (dn->dn_objset->os_dsl_dataset != NULL) |
1870 | rrw_exit(&os->os_dsl_dataset->ds_bp_rwlock, FTAG); | |
1871 | #endif | |
34dc7c2f BB |
1872 | ASSERT(db->db.db_size != 0); |
1873 | ||
1874 | dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); | |
1875 | ||
428870ff | 1876 | if (db->db_blkid != DMU_BONUS_BLKID) { |
3ec3bc21 | 1877 | dmu_objset_willuse_space(os, db->db.db_size, tx); |
34dc7c2f BB |
1878 | } |
1879 | ||
1880 | /* | |
1881 | * If this buffer is dirty in an old transaction group we need | |
1882 | * to make a copy of it so that the changes we make in this | |
1883 | * transaction group won't leak out when we sync the older txg. | |
1884 | */ | |
79c76d5b | 1885 | dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP); |
98f72a53 | 1886 | list_link_init(&dr->dr_dirty_node); |
34dc7c2f BB |
1887 | if (db->db_level == 0) { |
1888 | void *data_old = db->db_buf; | |
1889 | ||
b128c09f | 1890 | if (db->db_state != DB_NOFILL) { |
428870ff | 1891 | if (db->db_blkid == DMU_BONUS_BLKID) { |
b128c09f BB |
1892 | dbuf_fix_old_data(db, tx->tx_txg); |
1893 | data_old = db->db.db_data; | |
1894 | } else if (db->db.db_object != DMU_META_DNODE_OBJECT) { | |
1895 | /* | |
1896 | * Release the data buffer from the cache so | |
1897 | * that we can modify it without impacting | |
1898 | * possible other users of this cached data | |
1899 | * block. Note that indirect blocks and | |
1900 | * private objects are not released until the | |
1901 | * syncing state (since they are only modified | |
1902 | * then). | |
1903 | */ | |
1904 | arc_release(db->db_buf, db); | |
1905 | dbuf_fix_old_data(db, tx->tx_txg); | |
1906 | data_old = db->db_buf; | |
1907 | } | |
1908 | ASSERT(data_old != NULL); | |
34dc7c2f | 1909 | } |
34dc7c2f BB |
1910 | dr->dt.dl.dr_data = data_old; |
1911 | } else { | |
448d7aaa | 1912 | mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_NOLOCKDEP, NULL); |
34dc7c2f BB |
1913 | list_create(&dr->dt.di.dr_children, |
1914 | sizeof (dbuf_dirty_record_t), | |
1915 | offsetof(dbuf_dirty_record_t, dr_dirty_node)); | |
1916 | } | |
e8b96c60 MA |
1917 | if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL) |
1918 | dr->dr_accounted = db->db.db_size; | |
34dc7c2f BB |
1919 | dr->dr_dbuf = db; |
1920 | dr->dr_txg = tx->tx_txg; | |
1921 | dr->dr_next = *drp; | |
1922 | *drp = dr; | |
1923 | ||
1924 | /* | |
1925 | * We could have been freed_in_flight between the dbuf_noread | |
1926 | * and dbuf_dirty. We win, as though the dbuf_noread() had | |
1927 | * happened after the free. | |
1928 | */ | |
428870ff BB |
1929 | if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && |
1930 | db->db_blkid != DMU_SPILL_BLKID) { | |
34dc7c2f | 1931 | mutex_enter(&dn->dn_mtx); |
9bd274dd MA |
1932 | if (dn->dn_free_ranges[txgoff] != NULL) { |
1933 | range_tree_clear(dn->dn_free_ranges[txgoff], | |
1934 | db->db_blkid, 1); | |
1935 | } | |
34dc7c2f BB |
1936 | mutex_exit(&dn->dn_mtx); |
1937 | db->db_freed_in_flight = FALSE; | |
1938 | } | |
1939 | ||
1940 | /* | |
1941 | * This buffer is now part of this txg | |
1942 | */ | |
1943 | dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg); | |
1944 | db->db_dirtycnt += 1; | |
1945 | ASSERT3U(db->db_dirtycnt, <=, 3); | |
1946 | ||
1947 | mutex_exit(&db->db_mtx); | |
1948 | ||
428870ff BB |
1949 | if (db->db_blkid == DMU_BONUS_BLKID || |
1950 | db->db_blkid == DMU_SPILL_BLKID) { | |
34dc7c2f BB |
1951 | mutex_enter(&dn->dn_mtx); |
1952 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
1953 | list_insert_tail(&dn->dn_dirty_records[txgoff], dr); | |
1954 | mutex_exit(&dn->dn_mtx); | |
1955 | dnode_setdirty(dn, tx); | |
572e2857 | 1956 | DB_DNODE_EXIT(db); |
34dc7c2f | 1957 | return (dr); |
98ace739 MA |
1958 | } |
1959 | ||
1960 | /* | |
1961 | * The dn_struct_rwlock prevents db_blkptr from changing | |
1962 | * due to a write from syncing context completing | |
1963 | * while we are running, so we want to acquire it before | |
1964 | * looking at db_blkptr. | |
1965 | */ | |
1966 | if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) { | |
1967 | rw_enter(&dn->dn_struct_rwlock, RW_READER); | |
1968 | drop_struct_lock = TRUE; | |
1969 | } | |
1970 | ||
2ade4a99 MA |
1971 | /* |
1972 | * We need to hold the dn_struct_rwlock to make this assertion, | |
1973 | * because it protects dn_phys / dn_next_nlevels from changing. | |
1974 | */ | |
1975 | ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) || | |
1976 | dn->dn_phys->dn_nlevels > db->db_level || | |
1977 | dn->dn_next_nlevels[txgoff] > db->db_level || | |
1978 | dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level || | |
1979 | dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level); | |
1980 | ||
3ec3bc21 BB |
1981 | /* |
1982 | * If we are overwriting a dedup BP, then unless it is snapshotted, | |
1983 | * when we get to syncing context we will need to decrement its | |
1984 | * refcount in the DDT. Prefetch the relevant DDT block so that | |
1985 | * syncing context won't have to wait for the i/o. | |
1986 | */ | |
1987 | ddt_prefetch(os->os_spa, db->db_blkptr); | |
34dc7c2f | 1988 | |
b128c09f BB |
1989 | if (db->db_level == 0) { |
1990 | dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock); | |
1991 | ASSERT(dn->dn_maxblkid >= db->db_blkid); | |
1992 | } | |
1993 | ||
34dc7c2f BB |
1994 | if (db->db_level+1 < dn->dn_nlevels) { |
1995 | dmu_buf_impl_t *parent = db->db_parent; | |
1996 | dbuf_dirty_record_t *di; | |
1997 | int parent_held = FALSE; | |
1998 | ||
1999 | if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) { | |
2000 | int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; | |
2001 | ||
2002 | parent = dbuf_hold_level(dn, db->db_level+1, | |
2003 | db->db_blkid >> epbs, FTAG); | |
428870ff | 2004 | ASSERT(parent != NULL); |
34dc7c2f BB |
2005 | parent_held = TRUE; |
2006 | } | |
2007 | if (drop_struct_lock) | |
2008 | rw_exit(&dn->dn_struct_rwlock); | |
2009 | ASSERT3U(db->db_level+1, ==, parent->db_level); | |
2010 | di = dbuf_dirty(parent, tx); | |
2011 | if (parent_held) | |
2012 | dbuf_rele(parent, FTAG); | |
2013 | ||
2014 | mutex_enter(&db->db_mtx); | |
e8b96c60 MA |
2015 | /* |
2016 | * Since we've dropped the mutex, it's possible that | |
2017 | * dbuf_undirty() might have changed this out from under us. | |
2018 | */ | |
34dc7c2f BB |
2019 | if (db->db_last_dirty == dr || |
2020 | dn->dn_object == DMU_META_DNODE_OBJECT) { | |
2021 | mutex_enter(&di->dt.di.dr_mtx); | |
2022 | ASSERT3U(di->dr_txg, ==, tx->tx_txg); | |
2023 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
2024 | list_insert_tail(&di->dt.di.dr_children, dr); | |
2025 | mutex_exit(&di->dt.di.dr_mtx); | |
2026 | dr->dr_parent = di; | |
2027 | } | |
2028 | mutex_exit(&db->db_mtx); | |
2029 | } else { | |
2030 | ASSERT(db->db_level+1 == dn->dn_nlevels); | |
2031 | ASSERT(db->db_blkid < dn->dn_nblkptr); | |
572e2857 | 2032 | ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf); |
34dc7c2f BB |
2033 | mutex_enter(&dn->dn_mtx); |
2034 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
2035 | list_insert_tail(&dn->dn_dirty_records[txgoff], dr); | |
2036 | mutex_exit(&dn->dn_mtx); | |
2037 | if (drop_struct_lock) | |
2038 | rw_exit(&dn->dn_struct_rwlock); | |
2039 | } | |
2040 | ||
2041 | dnode_setdirty(dn, tx); | |
572e2857 | 2042 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
2043 | return (dr); |
2044 | } | |
2045 | ||
13fe0198 | 2046 | /* |
e49f1e20 WA |
2047 | * Undirty a buffer in the transaction group referenced by the given |
2048 | * transaction. Return whether this evicted the dbuf. | |
13fe0198 MA |
2049 | */ |
2050 | static boolean_t | |
34dc7c2f BB |
2051 | dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx) |
2052 | { | |
572e2857 | 2053 | dnode_t *dn; |
34dc7c2f BB |
2054 | uint64_t txg = tx->tx_txg; |
2055 | dbuf_dirty_record_t *dr, **drp; | |
2056 | ||
2057 | ASSERT(txg != 0); | |
4bda3bd0 MA |
2058 | |
2059 | /* | |
2060 | * Due to our use of dn_nlevels below, this can only be called | |
2061 | * in open context, unless we are operating on the MOS. | |
2062 | * From syncing context, dn_nlevels may be different from the | |
2063 | * dn_nlevels used when dbuf was dirtied. | |
2064 | */ | |
2065 | ASSERT(db->db_objset == | |
2066 | dmu_objset_pool(db->db_objset)->dp_meta_objset || | |
2067 | txg != spa_syncing_txg(dmu_objset_spa(db->db_objset))); | |
428870ff | 2068 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
13fe0198 MA |
2069 | ASSERT0(db->db_level); |
2070 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
34dc7c2f | 2071 | |
34dc7c2f BB |
2072 | /* |
2073 | * If this buffer is not dirty, we're done. | |
2074 | */ | |
2075 | for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next) | |
2076 | if (dr->dr_txg <= txg) | |
2077 | break; | |
13fe0198 MA |
2078 | if (dr == NULL || dr->dr_txg < txg) |
2079 | return (B_FALSE); | |
34dc7c2f | 2080 | ASSERT(dr->dr_txg == txg); |
428870ff | 2081 | ASSERT(dr->dr_dbuf == db); |
34dc7c2f | 2082 | |
572e2857 BB |
2083 | DB_DNODE_ENTER(db); |
2084 | dn = DB_DNODE(db); | |
2085 | ||
34dc7c2f BB |
2086 | dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); |
2087 | ||
2088 | ASSERT(db->db.db_size != 0); | |
2089 | ||
4bda3bd0 MA |
2090 | dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset), |
2091 | dr->dr_accounted, txg); | |
34dc7c2f BB |
2092 | |
2093 | *drp = dr->dr_next; | |
2094 | ||
ef3c1dea GR |
2095 | /* |
2096 | * Note that there are three places in dbuf_dirty() | |
2097 | * where this dirty record may be put on a list. | |
2098 | * Make sure to do a list_remove corresponding to | |
2099 | * every one of those list_insert calls. | |
2100 | */ | |
34dc7c2f BB |
2101 | if (dr->dr_parent) { |
2102 | mutex_enter(&dr->dr_parent->dt.di.dr_mtx); | |
2103 | list_remove(&dr->dr_parent->dt.di.dr_children, dr); | |
2104 | mutex_exit(&dr->dr_parent->dt.di.dr_mtx); | |
ef3c1dea | 2105 | } else if (db->db_blkid == DMU_SPILL_BLKID || |
4bda3bd0 | 2106 | db->db_level + 1 == dn->dn_nlevels) { |
b128c09f | 2107 | ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf); |
34dc7c2f BB |
2108 | mutex_enter(&dn->dn_mtx); |
2109 | list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr); | |
2110 | mutex_exit(&dn->dn_mtx); | |
2111 | } | |
572e2857 | 2112 | DB_DNODE_EXIT(db); |
34dc7c2f | 2113 | |
13fe0198 MA |
2114 | if (db->db_state != DB_NOFILL) { |
2115 | dbuf_unoverride(dr); | |
34dc7c2f | 2116 | |
34dc7c2f | 2117 | ASSERT(db->db_buf != NULL); |
13fe0198 MA |
2118 | ASSERT(dr->dt.dl.dr_data != NULL); |
2119 | if (dr->dt.dl.dr_data != db->db_buf) | |
d3c2ae1c | 2120 | arc_buf_destroy(dr->dt.dl.dr_data, db); |
34dc7c2f | 2121 | } |
58c4aa00 | 2122 | |
34dc7c2f BB |
2123 | kmem_free(dr, sizeof (dbuf_dirty_record_t)); |
2124 | ||
2125 | ASSERT(db->db_dirtycnt > 0); | |
2126 | db->db_dirtycnt -= 1; | |
2127 | ||
2128 | if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) { | |
d3c2ae1c GW |
2129 | ASSERT(db->db_state == DB_NOFILL || arc_released(db->db_buf)); |
2130 | dbuf_destroy(db); | |
13fe0198 | 2131 | return (B_TRUE); |
34dc7c2f BB |
2132 | } |
2133 | ||
13fe0198 | 2134 | return (B_FALSE); |
34dc7c2f BB |
2135 | } |
2136 | ||
b5256303 TC |
2137 | static void |
2138 | dmu_buf_will_dirty_impl(dmu_buf_t *db_fake, int flags, dmu_tx_t *tx) | |
34dc7c2f | 2139 | { |
b0bc7a84 | 2140 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; |
34dc7c2f BB |
2141 | |
2142 | ASSERT(tx->tx_txg != 0); | |
2143 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
2144 | ||
5a28a973 MA |
2145 | /* |
2146 | * Quick check for dirtyness. For already dirty blocks, this | |
2147 | * reduces runtime of this function by >90%, and overall performance | |
2148 | * by 50% for some workloads (e.g. file deletion with indirect blocks | |
2149 | * cached). | |
2150 | */ | |
2151 | mutex_enter(&db->db_mtx); | |
2152 | ||
1c27024e | 2153 | dbuf_dirty_record_t *dr; |
5a28a973 MA |
2154 | for (dr = db->db_last_dirty; |
2155 | dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) { | |
2156 | /* | |
2157 | * It's possible that it is already dirty but not cached, | |
2158 | * because there are some calls to dbuf_dirty() that don't | |
2159 | * go through dmu_buf_will_dirty(). | |
2160 | */ | |
2161 | if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) { | |
2162 | /* This dbuf is already dirty and cached. */ | |
2163 | dbuf_redirty(dr); | |
2164 | mutex_exit(&db->db_mtx); | |
2165 | return; | |
2166 | } | |
2167 | } | |
2168 | mutex_exit(&db->db_mtx); | |
2169 | ||
572e2857 BB |
2170 | DB_DNODE_ENTER(db); |
2171 | if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock)) | |
b5256303 | 2172 | flags |= DB_RF_HAVESTRUCT; |
572e2857 | 2173 | DB_DNODE_EXIT(db); |
b5256303 | 2174 | (void) dbuf_read(db, NULL, flags); |
34dc7c2f BB |
2175 | (void) dbuf_dirty(db, tx); |
2176 | } | |
2177 | ||
b5256303 TC |
2178 | void |
2179 | dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx) | |
2180 | { | |
2181 | dmu_buf_will_dirty_impl(db_fake, | |
2182 | DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH, tx); | |
2183 | } | |
2184 | ||
b128c09f BB |
2185 | void |
2186 | dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) | |
2187 | { | |
2188 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
2189 | ||
2190 | db->db_state = DB_NOFILL; | |
2191 | ||
2192 | dmu_buf_will_fill(db_fake, tx); | |
2193 | } | |
2194 | ||
34dc7c2f BB |
2195 | void |
2196 | dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) | |
2197 | { | |
2198 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
2199 | ||
428870ff | 2200 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f BB |
2201 | ASSERT(tx->tx_txg != 0); |
2202 | ASSERT(db->db_level == 0); | |
2203 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
2204 | ||
2205 | ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT || | |
2206 | dmu_tx_private_ok(tx)); | |
2207 | ||
2208 | dbuf_noread(db); | |
2209 | (void) dbuf_dirty(db, tx); | |
2210 | } | |
2211 | ||
b5256303 TC |
2212 | /* |
2213 | * This function is effectively the same as dmu_buf_will_dirty(), but | |
0c03d21a MA |
2214 | * indicates the caller expects raw encrypted data in the db, and provides |
2215 | * the crypt params (byteorder, salt, iv, mac) which should be stored in the | |
2216 | * blkptr_t when this dbuf is written. This is only used for blocks of | |
2217 | * dnodes, during raw receive. | |
b5256303 TC |
2218 | */ |
2219 | void | |
0c03d21a MA |
2220 | dmu_buf_set_crypt_params(dmu_buf_t *db_fake, boolean_t byteorder, |
2221 | const uint8_t *salt, const uint8_t *iv, const uint8_t *mac, dmu_tx_t *tx) | |
b5256303 TC |
2222 | { |
2223 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
2224 | dbuf_dirty_record_t *dr; | |
2225 | ||
0c03d21a MA |
2226 | /* |
2227 | * dr_has_raw_params is only processed for blocks of dnodes | |
2228 | * (see dbuf_sync_dnode_leaf_crypt()). | |
2229 | */ | |
2230 | ASSERT3U(db->db.db_object, ==, DMU_META_DNODE_OBJECT); | |
2231 | ASSERT3U(db->db_level, ==, 0); | |
2232 | ASSERT(db->db_objset->os_raw_receive); | |
2233 | ||
b5256303 TC |
2234 | dmu_buf_will_dirty_impl(db_fake, |
2235 | DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_NO_DECRYPT, tx); | |
2236 | ||
2237 | dr = db->db_last_dirty; | |
2238 | while (dr != NULL && dr->dr_txg > tx->tx_txg) | |
2239 | dr = dr->dr_next; | |
2240 | ||
2241 | ASSERT3P(dr, !=, NULL); | |
2242 | ASSERT3U(dr->dr_txg, ==, tx->tx_txg); | |
0c03d21a MA |
2243 | |
2244 | dr->dt.dl.dr_has_raw_params = B_TRUE; | |
2245 | dr->dt.dl.dr_byteorder = byteorder; | |
2246 | bcopy(salt, dr->dt.dl.dr_salt, ZIO_DATA_SALT_LEN); | |
2247 | bcopy(iv, dr->dt.dl.dr_iv, ZIO_DATA_IV_LEN); | |
2248 | bcopy(mac, dr->dt.dl.dr_mac, ZIO_DATA_MAC_LEN); | |
b5256303 TC |
2249 | } |
2250 | ||
34dc7c2f BB |
2251 | #pragma weak dmu_buf_fill_done = dbuf_fill_done |
2252 | /* ARGSUSED */ | |
2253 | void | |
2254 | dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx) | |
2255 | { | |
2256 | mutex_enter(&db->db_mtx); | |
2257 | DBUF_VERIFY(db); | |
2258 | ||
2259 | if (db->db_state == DB_FILL) { | |
2260 | if (db->db_level == 0 && db->db_freed_in_flight) { | |
428870ff | 2261 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f BB |
2262 | /* we were freed while filling */ |
2263 | /* XXX dbuf_undirty? */ | |
2264 | bzero(db->db.db_data, db->db.db_size); | |
2265 | db->db_freed_in_flight = FALSE; | |
2266 | } | |
2267 | db->db_state = DB_CACHED; | |
2268 | cv_broadcast(&db->db_changed); | |
2269 | } | |
2270 | mutex_exit(&db->db_mtx); | |
2271 | } | |
2272 | ||
9b67f605 MA |
2273 | void |
2274 | dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data, | |
2275 | bp_embedded_type_t etype, enum zio_compress comp, | |
2276 | int uncompressed_size, int compressed_size, int byteorder, | |
2277 | dmu_tx_t *tx) | |
2278 | { | |
2279 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; | |
2280 | struct dirty_leaf *dl; | |
2281 | dmu_object_type_t type; | |
2282 | ||
241b5415 MA |
2283 | if (etype == BP_EMBEDDED_TYPE_DATA) { |
2284 | ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset), | |
2285 | SPA_FEATURE_EMBEDDED_DATA)); | |
2286 | } | |
2287 | ||
9b67f605 MA |
2288 | DB_DNODE_ENTER(db); |
2289 | type = DB_DNODE(db)->dn_type; | |
2290 | DB_DNODE_EXIT(db); | |
2291 | ||
2292 | ASSERT0(db->db_level); | |
2293 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); | |
2294 | ||
2295 | dmu_buf_will_not_fill(dbuf, tx); | |
2296 | ||
2297 | ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); | |
2298 | dl = &db->db_last_dirty->dt.dl; | |
2299 | encode_embedded_bp_compressed(&dl->dr_overridden_by, | |
2300 | data, comp, uncompressed_size, compressed_size); | |
2301 | BPE_SET_ETYPE(&dl->dr_overridden_by, etype); | |
2302 | BP_SET_TYPE(&dl->dr_overridden_by, type); | |
2303 | BP_SET_LEVEL(&dl->dr_overridden_by, 0); | |
2304 | BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder); | |
2305 | ||
2306 | dl->dr_override_state = DR_OVERRIDDEN; | |
2307 | dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg; | |
2308 | } | |
2309 | ||
9babb374 BB |
2310 | /* |
2311 | * Directly assign a provided arc buf to a given dbuf if it's not referenced | |
2312 | * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf. | |
2313 | */ | |
2314 | void | |
2315 | dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx) | |
2316 | { | |
2317 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
428870ff | 2318 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
9babb374 | 2319 | ASSERT(db->db_level == 0); |
2aa34383 | 2320 | ASSERT3U(dbuf_is_metadata(db), ==, arc_is_metadata(buf)); |
9babb374 | 2321 | ASSERT(buf != NULL); |
2aa34383 | 2322 | ASSERT(arc_buf_lsize(buf) == db->db.db_size); |
9babb374 BB |
2323 | ASSERT(tx->tx_txg != 0); |
2324 | ||
2325 | arc_return_buf(buf, db); | |
2326 | ASSERT(arc_released(buf)); | |
2327 | ||
2328 | mutex_enter(&db->db_mtx); | |
2329 | ||
2330 | while (db->db_state == DB_READ || db->db_state == DB_FILL) | |
2331 | cv_wait(&db->db_changed, &db->db_mtx); | |
2332 | ||
2333 | ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED); | |
2334 | ||
2335 | if (db->db_state == DB_CACHED && | |
2336 | refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) { | |
440a3eb9 TC |
2337 | /* |
2338 | * In practice, we will never have a case where we have an | |
2339 | * encrypted arc buffer while additional holds exist on the | |
2340 | * dbuf. We don't handle this here so we simply assert that | |
2341 | * fact instead. | |
2342 | */ | |
2343 | ASSERT(!arc_is_encrypted(buf)); | |
9babb374 BB |
2344 | mutex_exit(&db->db_mtx); |
2345 | (void) dbuf_dirty(db, tx); | |
2346 | bcopy(buf->b_data, db->db.db_data, db->db.db_size); | |
d3c2ae1c | 2347 | arc_buf_destroy(buf, db); |
428870ff | 2348 | xuio_stat_wbuf_copied(); |
9babb374 BB |
2349 | return; |
2350 | } | |
2351 | ||
428870ff | 2352 | xuio_stat_wbuf_nocopy(); |
9babb374 BB |
2353 | if (db->db_state == DB_CACHED) { |
2354 | dbuf_dirty_record_t *dr = db->db_last_dirty; | |
2355 | ||
2356 | ASSERT(db->db_buf != NULL); | |
2357 | if (dr != NULL && dr->dr_txg == tx->tx_txg) { | |
2358 | ASSERT(dr->dt.dl.dr_data == db->db_buf); | |
440a3eb9 | 2359 | |
9babb374 BB |
2360 | if (!arc_released(db->db_buf)) { |
2361 | ASSERT(dr->dt.dl.dr_override_state == | |
2362 | DR_OVERRIDDEN); | |
2363 | arc_release(db->db_buf, db); | |
2364 | } | |
2365 | dr->dt.dl.dr_data = buf; | |
d3c2ae1c | 2366 | arc_buf_destroy(db->db_buf, db); |
9babb374 BB |
2367 | } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) { |
2368 | arc_release(db->db_buf, db); | |
d3c2ae1c | 2369 | arc_buf_destroy(db->db_buf, db); |
9babb374 BB |
2370 | } |
2371 | db->db_buf = NULL; | |
2372 | } | |
2373 | ASSERT(db->db_buf == NULL); | |
2374 | dbuf_set_data(db, buf); | |
2375 | db->db_state = DB_FILL; | |
2376 | mutex_exit(&db->db_mtx); | |
2377 | (void) dbuf_dirty(db, tx); | |
b0bc7a84 | 2378 | dmu_buf_fill_done(&db->db, tx); |
9babb374 BB |
2379 | } |
2380 | ||
34dc7c2f | 2381 | void |
d3c2ae1c | 2382 | dbuf_destroy(dmu_buf_impl_t *db) |
34dc7c2f | 2383 | { |
572e2857 | 2384 | dnode_t *dn; |
34dc7c2f | 2385 | dmu_buf_impl_t *parent = db->db_parent; |
572e2857 | 2386 | dmu_buf_impl_t *dndb; |
34dc7c2f BB |
2387 | |
2388 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
2389 | ASSERT(refcount_is_zero(&db->db_holds)); | |
2390 | ||
d3c2ae1c GW |
2391 | if (db->db_buf != NULL) { |
2392 | arc_buf_destroy(db->db_buf, db); | |
2393 | db->db_buf = NULL; | |
2394 | } | |
34dc7c2f | 2395 | |
d3c2ae1c GW |
2396 | if (db->db_blkid == DMU_BONUS_BLKID) { |
2397 | int slots = DB_DNODE(db)->dn_num_slots; | |
2398 | int bonuslen = DN_SLOTS_TO_BONUSLEN(slots); | |
b5256303 TC |
2399 | if (db->db.db_data != NULL) { |
2400 | kmem_free(db->db.db_data, bonuslen); | |
2401 | arc_space_return(bonuslen, ARC_SPACE_BONUS); | |
2402 | db->db_state = DB_UNCACHED; | |
2403 | } | |
34dc7c2f BB |
2404 | } |
2405 | ||
d3c2ae1c GW |
2406 | dbuf_clear_data(db); |
2407 | ||
2408 | if (multilist_link_active(&db->db_cache_link)) { | |
64fc7762 | 2409 | multilist_remove(dbuf_cache, db); |
d3c2ae1c GW |
2410 | (void) refcount_remove_many(&dbuf_cache_size, |
2411 | db->db.db_size, db); | |
5e021f56 GDN |
2412 | DBUF_STAT_BUMPDOWN(cache_levels[db->db_level]); |
2413 | DBUF_STAT_BUMPDOWN(cache_count); | |
2414 | DBUF_STAT_DECR(cache_levels_bytes[db->db_level], | |
2415 | db->db.db_size); | |
d3c2ae1c GW |
2416 | } |
2417 | ||
b128c09f | 2418 | ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL); |
34dc7c2f BB |
2419 | ASSERT(db->db_data_pending == NULL); |
2420 | ||
2421 | db->db_state = DB_EVICTING; | |
2422 | db->db_blkptr = NULL; | |
2423 | ||
d3c2ae1c GW |
2424 | /* |
2425 | * Now that db_state is DB_EVICTING, nobody else can find this via | |
2426 | * the hash table. We can now drop db_mtx, which allows us to | |
2427 | * acquire the dn_dbufs_mtx. | |
2428 | */ | |
2429 | mutex_exit(&db->db_mtx); | |
2430 | ||
572e2857 BB |
2431 | DB_DNODE_ENTER(db); |
2432 | dn = DB_DNODE(db); | |
2433 | dndb = dn->dn_dbuf; | |
d3c2ae1c GW |
2434 | if (db->db_blkid != DMU_BONUS_BLKID) { |
2435 | boolean_t needlock = !MUTEX_HELD(&dn->dn_dbufs_mtx); | |
2436 | if (needlock) | |
2437 | mutex_enter(&dn->dn_dbufs_mtx); | |
8951cb8d | 2438 | avl_remove(&dn->dn_dbufs, db); |
73ad4a9f | 2439 | atomic_dec_32(&dn->dn_dbufs_count); |
572e2857 BB |
2440 | membar_producer(); |
2441 | DB_DNODE_EXIT(db); | |
d3c2ae1c GW |
2442 | if (needlock) |
2443 | mutex_exit(&dn->dn_dbufs_mtx); | |
572e2857 BB |
2444 | /* |
2445 | * Decrementing the dbuf count means that the hold corresponding | |
2446 | * to the removed dbuf is no longer discounted in dnode_move(), | |
2447 | * so the dnode cannot be moved until after we release the hold. | |
2448 | * The membar_producer() ensures visibility of the decremented | |
2449 | * value in dnode_move(), since DB_DNODE_EXIT doesn't actually | |
2450 | * release any lock. | |
2451 | */ | |
34dc7c2f | 2452 | dnode_rele(dn, db); |
572e2857 | 2453 | db->db_dnode_handle = NULL; |
d3c2ae1c GW |
2454 | |
2455 | dbuf_hash_remove(db); | |
572e2857 BB |
2456 | } else { |
2457 | DB_DNODE_EXIT(db); | |
34dc7c2f BB |
2458 | } |
2459 | ||
d3c2ae1c | 2460 | ASSERT(refcount_is_zero(&db->db_holds)); |
34dc7c2f | 2461 | |
d3c2ae1c GW |
2462 | db->db_parent = NULL; |
2463 | ||
2464 | ASSERT(db->db_buf == NULL); | |
2465 | ASSERT(db->db.db_data == NULL); | |
2466 | ASSERT(db->db_hash_next == NULL); | |
2467 | ASSERT(db->db_blkptr == NULL); | |
2468 | ASSERT(db->db_data_pending == NULL); | |
2469 | ASSERT(!multilist_link_active(&db->db_cache_link)); | |
2470 | ||
2471 | kmem_cache_free(dbuf_kmem_cache, db); | |
2472 | arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF); | |
34dc7c2f BB |
2473 | |
2474 | /* | |
572e2857 | 2475 | * If this dbuf is referenced from an indirect dbuf, |
34dc7c2f BB |
2476 | * decrement the ref count on the indirect dbuf. |
2477 | */ | |
2478 | if (parent && parent != dndb) | |
2479 | dbuf_rele(parent, db); | |
2480 | } | |
2481 | ||
fcff0f35 PD |
2482 | /* |
2483 | * Note: While bpp will always be updated if the function returns success, | |
2484 | * parentp will not be updated if the dnode does not have dn_dbuf filled in; | |
9c5167d1 | 2485 | * this happens when the dnode is the meta-dnode, or {user|group|project}used |
fcff0f35 PD |
2486 | * object. |
2487 | */ | |
bf701a83 BB |
2488 | __attribute__((always_inline)) |
2489 | static inline int | |
34dc7c2f | 2490 | dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse, |
fc5bb51f | 2491 | dmu_buf_impl_t **parentp, blkptr_t **bpp, struct dbuf_hold_impl_data *dh) |
34dc7c2f | 2492 | { |
34dc7c2f BB |
2493 | *parentp = NULL; |
2494 | *bpp = NULL; | |
2495 | ||
428870ff BB |
2496 | ASSERT(blkid != DMU_BONUS_BLKID); |
2497 | ||
2498 | if (blkid == DMU_SPILL_BLKID) { | |
2499 | mutex_enter(&dn->dn_mtx); | |
2500 | if (dn->dn_have_spill && | |
2501 | (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) | |
50c957f7 | 2502 | *bpp = DN_SPILL_BLKPTR(dn->dn_phys); |
428870ff BB |
2503 | else |
2504 | *bpp = NULL; | |
2505 | dbuf_add_ref(dn->dn_dbuf, NULL); | |
2506 | *parentp = dn->dn_dbuf; | |
2507 | mutex_exit(&dn->dn_mtx); | |
2508 | return (0); | |
2509 | } | |
34dc7c2f | 2510 | |
1c27024e | 2511 | int nlevels = |
32d41fb7 | 2512 | (dn->dn_phys->dn_nlevels == 0) ? 1 : dn->dn_phys->dn_nlevels; |
1c27024e | 2513 | int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; |
34dc7c2f BB |
2514 | |
2515 | ASSERT3U(level * epbs, <, 64); | |
2516 | ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); | |
32d41fb7 PD |
2517 | /* |
2518 | * This assertion shouldn't trip as long as the max indirect block size | |
2519 | * is less than 1M. The reason for this is that up to that point, | |
2520 | * the number of levels required to address an entire object with blocks | |
2521 | * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In | |
2522 | * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55 | |
2523 | * (i.e. we can address the entire object), objects will all use at most | |
2524 | * N-1 levels and the assertion won't overflow. However, once epbs is | |
2525 | * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be | |
2526 | * enough to address an entire object, so objects will have 5 levels, | |
2527 | * but then this assertion will overflow. | |
2528 | * | |
2529 | * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we | |
2530 | * need to redo this logic to handle overflows. | |
2531 | */ | |
2532 | ASSERT(level >= nlevels || | |
2533 | ((nlevels - level - 1) * epbs) + | |
2534 | highbit64(dn->dn_phys->dn_nblkptr) <= 64); | |
34dc7c2f | 2535 | if (level >= nlevels || |
32d41fb7 PD |
2536 | blkid >= ((uint64_t)dn->dn_phys->dn_nblkptr << |
2537 | ((nlevels - level - 1) * epbs)) || | |
2538 | (fail_sparse && | |
2539 | blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) { | |
34dc7c2f | 2540 | /* the buffer has no parent yet */ |
2e528b49 | 2541 | return (SET_ERROR(ENOENT)); |
34dc7c2f BB |
2542 | } else if (level < nlevels-1) { |
2543 | /* this block is referenced from an indirect block */ | |
fc5bb51f BB |
2544 | int err; |
2545 | if (dh == NULL) { | |
fcff0f35 PD |
2546 | err = dbuf_hold_impl(dn, level+1, |
2547 | blkid >> epbs, fail_sparse, FALSE, NULL, parentp); | |
d1d7e268 | 2548 | } else { |
fc5bb51f | 2549 | __dbuf_hold_impl_init(dh + 1, dn, dh->dh_level + 1, |
fcff0f35 PD |
2550 | blkid >> epbs, fail_sparse, FALSE, NULL, |
2551 | parentp, dh->dh_depth + 1); | |
fc5bb51f BB |
2552 | err = __dbuf_hold_impl(dh + 1); |
2553 | } | |
34dc7c2f BB |
2554 | if (err) |
2555 | return (err); | |
2556 | err = dbuf_read(*parentp, NULL, | |
2557 | (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL)); | |
2558 | if (err) { | |
2559 | dbuf_rele(*parentp, NULL); | |
2560 | *parentp = NULL; | |
2561 | return (err); | |
2562 | } | |
2563 | *bpp = ((blkptr_t *)(*parentp)->db.db_data) + | |
2564 | (blkid & ((1ULL << epbs) - 1)); | |
32d41fb7 PD |
2565 | if (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs))) |
2566 | ASSERT(BP_IS_HOLE(*bpp)); | |
34dc7c2f BB |
2567 | return (0); |
2568 | } else { | |
2569 | /* the block is referenced from the dnode */ | |
2570 | ASSERT3U(level, ==, nlevels-1); | |
2571 | ASSERT(dn->dn_phys->dn_nblkptr == 0 || | |
2572 | blkid < dn->dn_phys->dn_nblkptr); | |
2573 | if (dn->dn_dbuf) { | |
2574 | dbuf_add_ref(dn->dn_dbuf, NULL); | |
2575 | *parentp = dn->dn_dbuf; | |
2576 | } | |
2577 | *bpp = &dn->dn_phys->dn_blkptr[blkid]; | |
2578 | return (0); | |
2579 | } | |
2580 | } | |
2581 | ||
2582 | static dmu_buf_impl_t * | |
2583 | dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid, | |
2584 | dmu_buf_impl_t *parent, blkptr_t *blkptr) | |
2585 | { | |
428870ff | 2586 | objset_t *os = dn->dn_objset; |
34dc7c2f BB |
2587 | dmu_buf_impl_t *db, *odb; |
2588 | ||
2589 | ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); | |
2590 | ASSERT(dn->dn_type != DMU_OT_NONE); | |
2591 | ||
d3c2ae1c | 2592 | db = kmem_cache_alloc(dbuf_kmem_cache, KM_SLEEP); |
34dc7c2f BB |
2593 | |
2594 | db->db_objset = os; | |
2595 | db->db.db_object = dn->dn_object; | |
2596 | db->db_level = level; | |
2597 | db->db_blkid = blkid; | |
2598 | db->db_last_dirty = NULL; | |
2599 | db->db_dirtycnt = 0; | |
572e2857 | 2600 | db->db_dnode_handle = dn->dn_handle; |
34dc7c2f BB |
2601 | db->db_parent = parent; |
2602 | db->db_blkptr = blkptr; | |
2603 | ||
0c66c32d | 2604 | db->db_user = NULL; |
bc4501f7 JG |
2605 | db->db_user_immediate_evict = FALSE; |
2606 | db->db_freed_in_flight = FALSE; | |
2607 | db->db_pending_evict = FALSE; | |
34dc7c2f | 2608 | |
428870ff | 2609 | if (blkid == DMU_BONUS_BLKID) { |
34dc7c2f | 2610 | ASSERT3P(parent, ==, dn->dn_dbuf); |
50c957f7 | 2611 | db->db.db_size = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) - |
34dc7c2f BB |
2612 | (dn->dn_nblkptr-1) * sizeof (blkptr_t); |
2613 | ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); | |
428870ff | 2614 | db->db.db_offset = DMU_BONUS_BLKID; |
34dc7c2f BB |
2615 | db->db_state = DB_UNCACHED; |
2616 | /* the bonus dbuf is not placed in the hash table */ | |
25458cbe | 2617 | arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF); |
34dc7c2f | 2618 | return (db); |
428870ff BB |
2619 | } else if (blkid == DMU_SPILL_BLKID) { |
2620 | db->db.db_size = (blkptr != NULL) ? | |
2621 | BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE; | |
2622 | db->db.db_offset = 0; | |
34dc7c2f BB |
2623 | } else { |
2624 | int blocksize = | |
e8b96c60 | 2625 | db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz; |
34dc7c2f BB |
2626 | db->db.db_size = blocksize; |
2627 | db->db.db_offset = db->db_blkid * blocksize; | |
2628 | } | |
2629 | ||
2630 | /* | |
2631 | * Hold the dn_dbufs_mtx while we get the new dbuf | |
2632 | * in the hash table *and* added to the dbufs list. | |
2633 | * This prevents a possible deadlock with someone | |
2634 | * trying to look up this dbuf before its added to the | |
2635 | * dn_dbufs list. | |
2636 | */ | |
2637 | mutex_enter(&dn->dn_dbufs_mtx); | |
2638 | db->db_state = DB_EVICTING; | |
2639 | if ((odb = dbuf_hash_insert(db)) != NULL) { | |
2640 | /* someone else inserted it first */ | |
d3c2ae1c | 2641 | kmem_cache_free(dbuf_kmem_cache, db); |
34dc7c2f | 2642 | mutex_exit(&dn->dn_dbufs_mtx); |
5e021f56 | 2643 | DBUF_STAT_BUMP(hash_insert_race); |
34dc7c2f BB |
2644 | return (odb); |
2645 | } | |
8951cb8d | 2646 | avl_add(&dn->dn_dbufs, db); |
9c9531cb | 2647 | |
34dc7c2f BB |
2648 | db->db_state = DB_UNCACHED; |
2649 | mutex_exit(&dn->dn_dbufs_mtx); | |
25458cbe | 2650 | arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF); |
34dc7c2f BB |
2651 | |
2652 | if (parent && parent != dn->dn_dbuf) | |
2653 | dbuf_add_ref(parent, db); | |
2654 | ||
2655 | ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || | |
2656 | refcount_count(&dn->dn_holds) > 0); | |
2657 | (void) refcount_add(&dn->dn_holds, db); | |
73ad4a9f | 2658 | atomic_inc_32(&dn->dn_dbufs_count); |
34dc7c2f BB |
2659 | |
2660 | dprintf_dbuf(db, "db=%p\n", db); | |
2661 | ||
2662 | return (db); | |
2663 | } | |
2664 | ||
fcff0f35 PD |
2665 | typedef struct dbuf_prefetch_arg { |
2666 | spa_t *dpa_spa; /* The spa to issue the prefetch in. */ | |
2667 | zbookmark_phys_t dpa_zb; /* The target block to prefetch. */ | |
2668 | int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */ | |
2669 | int dpa_curlevel; /* The current level that we're reading */ | |
d3c2ae1c | 2670 | dnode_t *dpa_dnode; /* The dnode associated with the prefetch */ |
fcff0f35 PD |
2671 | zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */ |
2672 | zio_t *dpa_zio; /* The parent zio_t for all prefetches. */ | |
2673 | arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */ | |
2674 | } dbuf_prefetch_arg_t; | |
2675 | ||
2676 | /* | |
2677 | * Actually issue the prefetch read for the block given. | |
2678 | */ | |
2679 | static void | |
2680 | dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp) | |
2681 | { | |
fcff0f35 PD |
2682 | if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) |
2683 | return; | |
2684 | ||
4515b1d0 | 2685 | int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE; |
1c27024e DB |
2686 | arc_flags_t aflags = |
2687 | dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH; | |
fcff0f35 | 2688 | |
4515b1d0 TC |
2689 | /* dnodes are always read as raw and then converted later */ |
2690 | if (BP_GET_TYPE(bp) == DMU_OT_DNODE && BP_IS_PROTECTED(bp) && | |
2691 | dpa->dpa_curlevel == 0) | |
2692 | zio_flags |= ZIO_FLAG_RAW; | |
2693 | ||
fcff0f35 PD |
2694 | ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp)); |
2695 | ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level); | |
2696 | ASSERT(dpa->dpa_zio != NULL); | |
2697 | (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL, | |
4515b1d0 | 2698 | dpa->dpa_prio, zio_flags, &aflags, &dpa->dpa_zb); |
fcff0f35 PD |
2699 | } |
2700 | ||
2701 | /* | |
2702 | * Called when an indirect block above our prefetch target is read in. This | |
2703 | * will either read in the next indirect block down the tree or issue the actual | |
2704 | * prefetch if the next block down is our target. | |
2705 | */ | |
2706 | static void | |
d4a72f23 TC |
2707 | dbuf_prefetch_indirect_done(zio_t *zio, const zbookmark_phys_t *zb, |
2708 | const blkptr_t *iobp, arc_buf_t *abuf, void *private) | |
fcff0f35 PD |
2709 | { |
2710 | dbuf_prefetch_arg_t *dpa = private; | |
fcff0f35 PD |
2711 | |
2712 | ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel); | |
2713 | ASSERT3S(dpa->dpa_curlevel, >, 0); | |
d3c2ae1c GW |
2714 | |
2715 | /* | |
2716 | * The dpa_dnode is only valid if we are called with a NULL | |
2717 | * zio. This indicates that the arc_read() returned without | |
2718 | * first calling zio_read() to issue a physical read. Once | |
2719 | * a physical read is made the dpa_dnode must be invalidated | |
2720 | * as the locks guarding it may have been dropped. If the | |
2721 | * dpa_dnode is still valid, then we want to add it to the dbuf | |
2722 | * cache. To do so, we must hold the dbuf associated with the block | |
2723 | * we just prefetched, read its contents so that we associate it | |
2724 | * with an arc_buf_t, and then release it. | |
2725 | */ | |
fcff0f35 PD |
2726 | if (zio != NULL) { |
2727 | ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel); | |
b5256303 | 2728 | if (zio->io_flags & ZIO_FLAG_RAW_COMPRESS) { |
d3c2ae1c GW |
2729 | ASSERT3U(BP_GET_PSIZE(zio->io_bp), ==, zio->io_size); |
2730 | } else { | |
2731 | ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size); | |
2732 | } | |
fcff0f35 | 2733 | ASSERT3P(zio->io_spa, ==, dpa->dpa_spa); |
d3c2ae1c GW |
2734 | |
2735 | dpa->dpa_dnode = NULL; | |
2736 | } else if (dpa->dpa_dnode != NULL) { | |
2737 | uint64_t curblkid = dpa->dpa_zb.zb_blkid >> | |
2738 | (dpa->dpa_epbs * (dpa->dpa_curlevel - | |
2739 | dpa->dpa_zb.zb_level)); | |
2740 | dmu_buf_impl_t *db = dbuf_hold_level(dpa->dpa_dnode, | |
2741 | dpa->dpa_curlevel, curblkid, FTAG); | |
2742 | (void) dbuf_read(db, NULL, | |
2743 | DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_HAVESTRUCT); | |
2744 | dbuf_rele(db, FTAG); | |
fcff0f35 PD |
2745 | } |
2746 | ||
d4a72f23 TC |
2747 | if (abuf == NULL) { |
2748 | kmem_free(dpa, sizeof (*dpa)); | |
2749 | return; | |
2750 | } | |
fcff0f35 | 2751 | |
d4a72f23 | 2752 | dpa->dpa_curlevel--; |
1c27024e | 2753 | uint64_t nextblkid = dpa->dpa_zb.zb_blkid >> |
fcff0f35 | 2754 | (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level)); |
1c27024e | 2755 | blkptr_t *bp = ((blkptr_t *)abuf->b_data) + |
fcff0f35 | 2756 | P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs); |
d4a72f23 TC |
2757 | |
2758 | if (BP_IS_HOLE(bp)) { | |
fcff0f35 PD |
2759 | kmem_free(dpa, sizeof (*dpa)); |
2760 | } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) { | |
2761 | ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid); | |
2762 | dbuf_issue_final_prefetch(dpa, bp); | |
2763 | kmem_free(dpa, sizeof (*dpa)); | |
2764 | } else { | |
2765 | arc_flags_t iter_aflags = ARC_FLAG_NOWAIT; | |
2766 | zbookmark_phys_t zb; | |
2767 | ||
7c351e31 | 2768 | /* flag if L2ARC eligible, l2arc_noprefetch then decides */ |
2769 | if (dpa->dpa_aflags & ARC_FLAG_L2CACHE) | |
2770 | iter_aflags |= ARC_FLAG_L2CACHE; | |
2771 | ||
fcff0f35 PD |
2772 | ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp)); |
2773 | ||
2774 | SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset, | |
2775 | dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid); | |
2776 | ||
2777 | (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, | |
2778 | bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio, | |
2779 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, | |
2780 | &iter_aflags, &zb); | |
2781 | } | |
d3c2ae1c GW |
2782 | |
2783 | arc_buf_destroy(abuf, private); | |
fcff0f35 PD |
2784 | } |
2785 | ||
2786 | /* | |
2787 | * Issue prefetch reads for the given block on the given level. If the indirect | |
2788 | * blocks above that block are not in memory, we will read them in | |
2789 | * asynchronously. As a result, this call never blocks waiting for a read to | |
b5256303 TC |
2790 | * complete. Note that the prefetch might fail if the dataset is encrypted and |
2791 | * the encryption key is unmapped before the IO completes. | |
fcff0f35 | 2792 | */ |
34dc7c2f | 2793 | void |
fcff0f35 PD |
2794 | dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio, |
2795 | arc_flags_t aflags) | |
34dc7c2f | 2796 | { |
fcff0f35 PD |
2797 | blkptr_t bp; |
2798 | int epbs, nlevels, curlevel; | |
2799 | uint64_t curblkid; | |
34dc7c2f | 2800 | |
428870ff | 2801 | ASSERT(blkid != DMU_BONUS_BLKID); |
34dc7c2f BB |
2802 | ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); |
2803 | ||
7f60329a MA |
2804 | if (blkid > dn->dn_maxblkid) |
2805 | return; | |
2806 | ||
34dc7c2f BB |
2807 | if (dnode_block_freed(dn, blkid)) |
2808 | return; | |
2809 | ||
fcff0f35 PD |
2810 | /* |
2811 | * This dnode hasn't been written to disk yet, so there's nothing to | |
2812 | * prefetch. | |
2813 | */ | |
2814 | nlevels = dn->dn_phys->dn_nlevels; | |
2815 | if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0) | |
2816 | return; | |
2817 | ||
2818 | epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; | |
2819 | if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level)) | |
2820 | return; | |
2821 | ||
1c27024e | 2822 | dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object, |
fcff0f35 PD |
2823 | level, blkid); |
2824 | if (db != NULL) { | |
2825 | mutex_exit(&db->db_mtx); | |
572e2857 | 2826 | /* |
fcff0f35 PD |
2827 | * This dbuf already exists. It is either CACHED, or |
2828 | * (we assume) about to be read or filled. | |
572e2857 | 2829 | */ |
572e2857 | 2830 | return; |
34dc7c2f BB |
2831 | } |
2832 | ||
fcff0f35 PD |
2833 | /* |
2834 | * Find the closest ancestor (indirect block) of the target block | |
2835 | * that is present in the cache. In this indirect block, we will | |
2836 | * find the bp that is at curlevel, curblkid. | |
2837 | */ | |
2838 | curlevel = level; | |
2839 | curblkid = blkid; | |
2840 | while (curlevel < nlevels - 1) { | |
2841 | int parent_level = curlevel + 1; | |
2842 | uint64_t parent_blkid = curblkid >> epbs; | |
2843 | dmu_buf_impl_t *db; | |
2844 | ||
2845 | if (dbuf_hold_impl(dn, parent_level, parent_blkid, | |
2846 | FALSE, TRUE, FTAG, &db) == 0) { | |
2847 | blkptr_t *bpp = db->db_buf->b_data; | |
2848 | bp = bpp[P2PHASE(curblkid, 1 << epbs)]; | |
2849 | dbuf_rele(db, FTAG); | |
2850 | break; | |
2851 | } | |
428870ff | 2852 | |
fcff0f35 PD |
2853 | curlevel = parent_level; |
2854 | curblkid = parent_blkid; | |
2855 | } | |
34dc7c2f | 2856 | |
fcff0f35 PD |
2857 | if (curlevel == nlevels - 1) { |
2858 | /* No cached indirect blocks found. */ | |
2859 | ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr); | |
2860 | bp = dn->dn_phys->dn_blkptr[curblkid]; | |
34dc7c2f | 2861 | } |
fcff0f35 PD |
2862 | if (BP_IS_HOLE(&bp)) |
2863 | return; | |
2864 | ||
2865 | ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp)); | |
2866 | ||
1c27024e | 2867 | zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL, |
fcff0f35 PD |
2868 | ZIO_FLAG_CANFAIL); |
2869 | ||
1c27024e DB |
2870 | dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP); |
2871 | dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset; | |
fcff0f35 PD |
2872 | SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET, |
2873 | dn->dn_object, level, blkid); | |
2874 | dpa->dpa_curlevel = curlevel; | |
2875 | dpa->dpa_prio = prio; | |
2876 | dpa->dpa_aflags = aflags; | |
2877 | dpa->dpa_spa = dn->dn_objset->os_spa; | |
d3c2ae1c | 2878 | dpa->dpa_dnode = dn; |
fcff0f35 PD |
2879 | dpa->dpa_epbs = epbs; |
2880 | dpa->dpa_zio = pio; | |
2881 | ||
7c351e31 | 2882 | /* flag if L2ARC eligible, l2arc_noprefetch then decides */ |
2883 | if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level)) | |
2884 | dpa->dpa_aflags |= ARC_FLAG_L2CACHE; | |
2885 | ||
fcff0f35 PD |
2886 | /* |
2887 | * If we have the indirect just above us, no need to do the asynchronous | |
2888 | * prefetch chain; we'll just run the last step ourselves. If we're at | |
2889 | * a higher level, though, we want to issue the prefetches for all the | |
2890 | * indirect blocks asynchronously, so we can go on with whatever we were | |
2891 | * doing. | |
2892 | */ | |
2893 | if (curlevel == level) { | |
2894 | ASSERT3U(curblkid, ==, blkid); | |
2895 | dbuf_issue_final_prefetch(dpa, &bp); | |
2896 | kmem_free(dpa, sizeof (*dpa)); | |
2897 | } else { | |
2898 | arc_flags_t iter_aflags = ARC_FLAG_NOWAIT; | |
2899 | zbookmark_phys_t zb; | |
2900 | ||
7c351e31 | 2901 | /* flag if L2ARC eligible, l2arc_noprefetch then decides */ |
2902 | if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level)) | |
2903 | iter_aflags |= ARC_FLAG_L2CACHE; | |
2904 | ||
fcff0f35 PD |
2905 | SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET, |
2906 | dn->dn_object, curlevel, curblkid); | |
2907 | (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, | |
2908 | &bp, dbuf_prefetch_indirect_done, dpa, prio, | |
2909 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, | |
2910 | &iter_aflags, &zb); | |
2911 | } | |
2912 | /* | |
2913 | * We use pio here instead of dpa_zio since it's possible that | |
2914 | * dpa may have already been freed. | |
2915 | */ | |
2916 | zio_nowait(pio); | |
34dc7c2f BB |
2917 | } |
2918 | ||
d1d7e268 | 2919 | #define DBUF_HOLD_IMPL_MAX_DEPTH 20 |
fc5bb51f | 2920 | |
71a24c3c TC |
2921 | /* |
2922 | * Helper function for __dbuf_hold_impl() to copy a buffer. Handles | |
2923 | * the case of encrypted, compressed and uncompressed buffers by | |
2924 | * allocating the new buffer, respectively, with arc_alloc_raw_buf(), | |
2925 | * arc_alloc_compressed_buf() or arc_alloc_buf().* | |
2926 | * | |
2927 | * NOTE: Declared noinline to avoid stack bloat in __dbuf_hold_impl(). | |
2928 | */ | |
2929 | noinline static void | |
2930 | dbuf_hold_copy(struct dbuf_hold_impl_data *dh) | |
2931 | { | |
2932 | dnode_t *dn = dh->dh_dn; | |
2933 | dmu_buf_impl_t *db = dh->dh_db; | |
2934 | dbuf_dirty_record_t *dr = dh->dh_dr; | |
2935 | arc_buf_t *data = dr->dt.dl.dr_data; | |
2936 | ||
2937 | enum zio_compress compress_type = arc_get_compression(data); | |
2938 | ||
2939 | if (arc_is_encrypted(data)) { | |
2940 | boolean_t byteorder; | |
2941 | uint8_t salt[ZIO_DATA_SALT_LEN]; | |
2942 | uint8_t iv[ZIO_DATA_IV_LEN]; | |
2943 | uint8_t mac[ZIO_DATA_MAC_LEN]; | |
2944 | ||
2945 | arc_get_raw_params(data, &byteorder, salt, iv, mac); | |
2946 | dbuf_set_data(db, arc_alloc_raw_buf(dn->dn_objset->os_spa, db, | |
2947 | dmu_objset_id(dn->dn_objset), byteorder, salt, iv, mac, | |
2948 | dn->dn_type, arc_buf_size(data), arc_buf_lsize(data), | |
2949 | compress_type)); | |
2950 | } else if (compress_type != ZIO_COMPRESS_OFF) { | |
2951 | dbuf_set_data(db, arc_alloc_compressed_buf( | |
2952 | dn->dn_objset->os_spa, db, arc_buf_size(data), | |
2953 | arc_buf_lsize(data), compress_type)); | |
2954 | } else { | |
2955 | dbuf_set_data(db, arc_alloc_buf(dn->dn_objset->os_spa, db, | |
2956 | DBUF_GET_BUFC_TYPE(db), db->db.db_size)); | |
2957 | } | |
2958 | ||
2959 | bcopy(data->b_data, db->db.db_data, arc_buf_size(data)); | |
2960 | } | |
2961 | ||
34dc7c2f BB |
2962 | /* |
2963 | * Returns with db_holds incremented, and db_mtx not held. | |
2964 | * Note: dn_struct_rwlock must be held. | |
2965 | */ | |
fc5bb51f BB |
2966 | static int |
2967 | __dbuf_hold_impl(struct dbuf_hold_impl_data *dh) | |
34dc7c2f | 2968 | { |
fc5bb51f BB |
2969 | ASSERT3S(dh->dh_depth, <, DBUF_HOLD_IMPL_MAX_DEPTH); |
2970 | dh->dh_parent = NULL; | |
34dc7c2f | 2971 | |
fc5bb51f BB |
2972 | ASSERT(dh->dh_blkid != DMU_BONUS_BLKID); |
2973 | ASSERT(RW_LOCK_HELD(&dh->dh_dn->dn_struct_rwlock)); | |
2974 | ASSERT3U(dh->dh_dn->dn_nlevels, >, dh->dh_level); | |
34dc7c2f | 2975 | |
fc5bb51f | 2976 | *(dh->dh_dbp) = NULL; |
d3c2ae1c | 2977 | |
34dc7c2f | 2978 | /* dbuf_find() returns with db_mtx held */ |
6ebebace JG |
2979 | dh->dh_db = dbuf_find(dh->dh_dn->dn_objset, dh->dh_dn->dn_object, |
2980 | dh->dh_level, dh->dh_blkid); | |
fc5bb51f BB |
2981 | |
2982 | if (dh->dh_db == NULL) { | |
2983 | dh->dh_bp = NULL; | |
2984 | ||
fcff0f35 PD |
2985 | if (dh->dh_fail_uncached) |
2986 | return (SET_ERROR(ENOENT)); | |
2987 | ||
fc5bb51f BB |
2988 | ASSERT3P(dh->dh_parent, ==, NULL); |
2989 | dh->dh_err = dbuf_findbp(dh->dh_dn, dh->dh_level, dh->dh_blkid, | |
02730c33 | 2990 | dh->dh_fail_sparse, &dh->dh_parent, &dh->dh_bp, dh); |
fc5bb51f | 2991 | if (dh->dh_fail_sparse) { |
d1d7e268 MK |
2992 | if (dh->dh_err == 0 && |
2993 | dh->dh_bp && BP_IS_HOLE(dh->dh_bp)) | |
2e528b49 | 2994 | dh->dh_err = SET_ERROR(ENOENT); |
fc5bb51f BB |
2995 | if (dh->dh_err) { |
2996 | if (dh->dh_parent) | |
2997 | dbuf_rele(dh->dh_parent, NULL); | |
2998 | return (dh->dh_err); | |
34dc7c2f BB |
2999 | } |
3000 | } | |
fc5bb51f BB |
3001 | if (dh->dh_err && dh->dh_err != ENOENT) |
3002 | return (dh->dh_err); | |
3003 | dh->dh_db = dbuf_create(dh->dh_dn, dh->dh_level, dh->dh_blkid, | |
02730c33 | 3004 | dh->dh_parent, dh->dh_bp); |
34dc7c2f BB |
3005 | } |
3006 | ||
fcff0f35 PD |
3007 | if (dh->dh_fail_uncached && dh->dh_db->db_state != DB_CACHED) { |
3008 | mutex_exit(&dh->dh_db->db_mtx); | |
3009 | return (SET_ERROR(ENOENT)); | |
3010 | } | |
3011 | ||
0873bb63 BB |
3012 | if (dh->dh_db->db_buf != NULL) { |
3013 | arc_buf_access(dh->dh_db->db_buf); | |
fc5bb51f | 3014 | ASSERT3P(dh->dh_db->db.db_data, ==, dh->dh_db->db_buf->b_data); |
0873bb63 | 3015 | } |
34dc7c2f | 3016 | |
fc5bb51f | 3017 | ASSERT(dh->dh_db->db_buf == NULL || arc_referenced(dh->dh_db->db_buf)); |
34dc7c2f BB |
3018 | |
3019 | /* | |
3020 | * If this buffer is currently syncing out, and we are are | |
3021 | * still referencing it from db_data, we need to make a copy | |
3022 | * of it in case we decide we want to dirty it again in this txg. | |
3023 | */ | |
fc5bb51f BB |
3024 | if (dh->dh_db->db_level == 0 && |
3025 | dh->dh_db->db_blkid != DMU_BONUS_BLKID && | |
3026 | dh->dh_dn->dn_object != DMU_META_DNODE_OBJECT && | |
3027 | dh->dh_db->db_state == DB_CACHED && dh->dh_db->db_data_pending) { | |
3028 | dh->dh_dr = dh->dh_db->db_data_pending; | |
71a24c3c TC |
3029 | if (dh->dh_dr->dt.dl.dr_data == dh->dh_db->db_buf) |
3030 | dbuf_hold_copy(dh); | |
34dc7c2f BB |
3031 | } |
3032 | ||
d3c2ae1c GW |
3033 | if (multilist_link_active(&dh->dh_db->db_cache_link)) { |
3034 | ASSERT(refcount_is_zero(&dh->dh_db->db_holds)); | |
64fc7762 | 3035 | multilist_remove(dbuf_cache, dh->dh_db); |
d3c2ae1c GW |
3036 | (void) refcount_remove_many(&dbuf_cache_size, |
3037 | dh->dh_db->db.db_size, dh->dh_db); | |
5e021f56 GDN |
3038 | DBUF_STAT_BUMPDOWN(cache_levels[dh->dh_db->db_level]); |
3039 | DBUF_STAT_BUMPDOWN(cache_count); | |
3040 | DBUF_STAT_DECR(cache_levels_bytes[dh->dh_db->db_level], | |
3041 | dh->dh_db->db.db_size); | |
d3c2ae1c | 3042 | } |
fc5bb51f | 3043 | (void) refcount_add(&dh->dh_db->db_holds, dh->dh_tag); |
fc5bb51f BB |
3044 | DBUF_VERIFY(dh->dh_db); |
3045 | mutex_exit(&dh->dh_db->db_mtx); | |
34dc7c2f BB |
3046 | |
3047 | /* NOTE: we can't rele the parent until after we drop the db_mtx */ | |
fc5bb51f BB |
3048 | if (dh->dh_parent) |
3049 | dbuf_rele(dh->dh_parent, NULL); | |
34dc7c2f | 3050 | |
fc5bb51f BB |
3051 | ASSERT3P(DB_DNODE(dh->dh_db), ==, dh->dh_dn); |
3052 | ASSERT3U(dh->dh_db->db_blkid, ==, dh->dh_blkid); | |
3053 | ASSERT3U(dh->dh_db->db_level, ==, dh->dh_level); | |
3054 | *(dh->dh_dbp) = dh->dh_db; | |
34dc7c2f BB |
3055 | |
3056 | return (0); | |
3057 | } | |
3058 | ||
fc5bb51f BB |
3059 | /* |
3060 | * The following code preserves the recursive function dbuf_hold_impl() | |
3061 | * but moves the local variables AND function arguments to the heap to | |
3062 | * minimize the stack frame size. Enough space is initially allocated | |
3063 | * on the stack for 20 levels of recursion. | |
3064 | */ | |
3065 | int | |
fcff0f35 PD |
3066 | dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, |
3067 | boolean_t fail_sparse, boolean_t fail_uncached, | |
fc5bb51f BB |
3068 | void *tag, dmu_buf_impl_t **dbp) |
3069 | { | |
3070 | struct dbuf_hold_impl_data *dh; | |
3071 | int error; | |
3072 | ||
d9eea113 | 3073 | dh = kmem_alloc(sizeof (struct dbuf_hold_impl_data) * |
79c76d5b | 3074 | DBUF_HOLD_IMPL_MAX_DEPTH, KM_SLEEP); |
fcff0f35 | 3075 | __dbuf_hold_impl_init(dh, dn, level, blkid, fail_sparse, |
02730c33 | 3076 | fail_uncached, tag, dbp, 0); |
fc5bb51f BB |
3077 | |
3078 | error = __dbuf_hold_impl(dh); | |
3079 | ||
d1d7e268 | 3080 | kmem_free(dh, sizeof (struct dbuf_hold_impl_data) * |
fc5bb51f BB |
3081 | DBUF_HOLD_IMPL_MAX_DEPTH); |
3082 | ||
3083 | return (error); | |
3084 | } | |
3085 | ||
3086 | static void | |
3087 | __dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh, | |
fcff0f35 | 3088 | dnode_t *dn, uint8_t level, uint64_t blkid, |
4ea3f864 GM |
3089 | boolean_t fail_sparse, boolean_t fail_uncached, |
3090 | void *tag, dmu_buf_impl_t **dbp, int depth) | |
fc5bb51f BB |
3091 | { |
3092 | dh->dh_dn = dn; | |
3093 | dh->dh_level = level; | |
3094 | dh->dh_blkid = blkid; | |
fcff0f35 | 3095 | |
fc5bb51f | 3096 | dh->dh_fail_sparse = fail_sparse; |
fcff0f35 PD |
3097 | dh->dh_fail_uncached = fail_uncached; |
3098 | ||
fc5bb51f BB |
3099 | dh->dh_tag = tag; |
3100 | dh->dh_dbp = dbp; | |
d9eea113 MA |
3101 | |
3102 | dh->dh_db = NULL; | |
3103 | dh->dh_parent = NULL; | |
3104 | dh->dh_bp = NULL; | |
3105 | dh->dh_err = 0; | |
3106 | dh->dh_dr = NULL; | |
d9eea113 | 3107 | |
fc5bb51f BB |
3108 | dh->dh_depth = depth; |
3109 | } | |
3110 | ||
34dc7c2f BB |
3111 | dmu_buf_impl_t * |
3112 | dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag) | |
3113 | { | |
fcff0f35 | 3114 | return (dbuf_hold_level(dn, 0, blkid, tag)); |
34dc7c2f BB |
3115 | } |
3116 | ||
3117 | dmu_buf_impl_t * | |
3118 | dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag) | |
3119 | { | |
3120 | dmu_buf_impl_t *db; | |
fcff0f35 | 3121 | int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db); |
34dc7c2f BB |
3122 | return (err ? NULL : db); |
3123 | } | |
3124 | ||
3125 | void | |
3126 | dbuf_create_bonus(dnode_t *dn) | |
3127 | { | |
3128 | ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); | |
3129 | ||
3130 | ASSERT(dn->dn_bonus == NULL); | |
428870ff BB |
3131 | dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL); |
3132 | } | |
3133 | ||
3134 | int | |
3135 | dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx) | |
3136 | { | |
3137 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
572e2857 BB |
3138 | dnode_t *dn; |
3139 | ||
428870ff | 3140 | if (db->db_blkid != DMU_SPILL_BLKID) |
2e528b49 | 3141 | return (SET_ERROR(ENOTSUP)); |
428870ff BB |
3142 | if (blksz == 0) |
3143 | blksz = SPA_MINBLOCKSIZE; | |
f1512ee6 MA |
3144 | ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset))); |
3145 | blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE); | |
428870ff | 3146 | |
572e2857 BB |
3147 | DB_DNODE_ENTER(db); |
3148 | dn = DB_DNODE(db); | |
3149 | rw_enter(&dn->dn_struct_rwlock, RW_WRITER); | |
428870ff | 3150 | dbuf_new_size(db, blksz, tx); |
572e2857 BB |
3151 | rw_exit(&dn->dn_struct_rwlock); |
3152 | DB_DNODE_EXIT(db); | |
428870ff BB |
3153 | |
3154 | return (0); | |
3155 | } | |
3156 | ||
3157 | void | |
3158 | dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx) | |
3159 | { | |
3160 | dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx); | |
34dc7c2f BB |
3161 | } |
3162 | ||
3163 | #pragma weak dmu_buf_add_ref = dbuf_add_ref | |
3164 | void | |
3165 | dbuf_add_ref(dmu_buf_impl_t *db, void *tag) | |
3166 | { | |
d3c2ae1c GW |
3167 | int64_t holds = refcount_add(&db->db_holds, tag); |
3168 | VERIFY3S(holds, >, 1); | |
34dc7c2f BB |
3169 | } |
3170 | ||
6ebebace JG |
3171 | #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref |
3172 | boolean_t | |
3173 | dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid, | |
3174 | void *tag) | |
3175 | { | |
3176 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
3177 | dmu_buf_impl_t *found_db; | |
3178 | boolean_t result = B_FALSE; | |
3179 | ||
d617648c | 3180 | if (blkid == DMU_BONUS_BLKID) |
6ebebace JG |
3181 | found_db = dbuf_find_bonus(os, obj); |
3182 | else | |
3183 | found_db = dbuf_find(os, obj, 0, blkid); | |
3184 | ||
3185 | if (found_db != NULL) { | |
3186 | if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) { | |
3187 | (void) refcount_add(&db->db_holds, tag); | |
3188 | result = B_TRUE; | |
3189 | } | |
d617648c | 3190 | mutex_exit(&found_db->db_mtx); |
6ebebace JG |
3191 | } |
3192 | return (result); | |
3193 | } | |
3194 | ||
572e2857 BB |
3195 | /* |
3196 | * If you call dbuf_rele() you had better not be referencing the dnode handle | |
3197 | * unless you have some other direct or indirect hold on the dnode. (An indirect | |
3198 | * hold is a hold on one of the dnode's dbufs, including the bonus buffer.) | |
3199 | * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the | |
3200 | * dnode's parent dbuf evicting its dnode handles. | |
3201 | */ | |
34dc7c2f BB |
3202 | void |
3203 | dbuf_rele(dmu_buf_impl_t *db, void *tag) | |
428870ff BB |
3204 | { |
3205 | mutex_enter(&db->db_mtx); | |
3206 | dbuf_rele_and_unlock(db, tag); | |
3207 | } | |
3208 | ||
b0bc7a84 MG |
3209 | void |
3210 | dmu_buf_rele(dmu_buf_t *db, void *tag) | |
3211 | { | |
3212 | dbuf_rele((dmu_buf_impl_t *)db, tag); | |
3213 | } | |
3214 | ||
428870ff BB |
3215 | /* |
3216 | * dbuf_rele() for an already-locked dbuf. This is necessary to allow | |
3217 | * db_dirtycnt and db_holds to be updated atomically. | |
3218 | */ | |
3219 | void | |
3220 | dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag) | |
34dc7c2f BB |
3221 | { |
3222 | int64_t holds; | |
3223 | ||
428870ff | 3224 | ASSERT(MUTEX_HELD(&db->db_mtx)); |
34dc7c2f BB |
3225 | DBUF_VERIFY(db); |
3226 | ||
572e2857 BB |
3227 | /* |
3228 | * Remove the reference to the dbuf before removing its hold on the | |
3229 | * dnode so we can guarantee in dnode_move() that a referenced bonus | |
3230 | * buffer has a corresponding dnode hold. | |
3231 | */ | |
34dc7c2f BB |
3232 | holds = refcount_remove(&db->db_holds, tag); |
3233 | ASSERT(holds >= 0); | |
3234 | ||
3235 | /* | |
3236 | * We can't freeze indirects if there is a possibility that they | |
3237 | * may be modified in the current syncing context. | |
3238 | */ | |
d3c2ae1c GW |
3239 | if (db->db_buf != NULL && |
3240 | holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) { | |
34dc7c2f | 3241 | arc_buf_freeze(db->db_buf); |
d3c2ae1c | 3242 | } |
34dc7c2f BB |
3243 | |
3244 | if (holds == db->db_dirtycnt && | |
bc4501f7 | 3245 | db->db_level == 0 && db->db_user_immediate_evict) |
34dc7c2f BB |
3246 | dbuf_evict_user(db); |
3247 | ||
3248 | if (holds == 0) { | |
428870ff | 3249 | if (db->db_blkid == DMU_BONUS_BLKID) { |
4c7b7eed | 3250 | dnode_t *dn; |
bc4501f7 | 3251 | boolean_t evict_dbuf = db->db_pending_evict; |
572e2857 BB |
3252 | |
3253 | /* | |
4c7b7eed JG |
3254 | * If the dnode moves here, we cannot cross this |
3255 | * barrier until the move completes. | |
572e2857 BB |
3256 | */ |
3257 | DB_DNODE_ENTER(db); | |
4c7b7eed JG |
3258 | |
3259 | dn = DB_DNODE(db); | |
3260 | atomic_dec_32(&dn->dn_dbufs_count); | |
3261 | ||
3262 | /* | |
3263 | * Decrementing the dbuf count means that the bonus | |
3264 | * buffer's dnode hold is no longer discounted in | |
3265 | * dnode_move(). The dnode cannot move until after | |
bc4501f7 | 3266 | * the dnode_rele() below. |
4c7b7eed | 3267 | */ |
572e2857 | 3268 | DB_DNODE_EXIT(db); |
4c7b7eed JG |
3269 | |
3270 | /* | |
3271 | * Do not reference db after its lock is dropped. | |
3272 | * Another thread may evict it. | |
3273 | */ | |
3274 | mutex_exit(&db->db_mtx); | |
3275 | ||
bc4501f7 | 3276 | if (evict_dbuf) |
4c7b7eed | 3277 | dnode_evict_bonus(dn); |
bc4501f7 JG |
3278 | |
3279 | dnode_rele(dn, db); | |
34dc7c2f BB |
3280 | } else if (db->db_buf == NULL) { |
3281 | /* | |
3282 | * This is a special case: we never associated this | |
3283 | * dbuf with any data allocated from the ARC. | |
3284 | */ | |
b128c09f BB |
3285 | ASSERT(db->db_state == DB_UNCACHED || |
3286 | db->db_state == DB_NOFILL); | |
d3c2ae1c | 3287 | dbuf_destroy(db); |
34dc7c2f | 3288 | } else if (arc_released(db->db_buf)) { |
34dc7c2f BB |
3289 | /* |
3290 | * This dbuf has anonymous data associated with it. | |
3291 | */ | |
d3c2ae1c | 3292 | dbuf_destroy(db); |
34dc7c2f | 3293 | } else { |
d3c2ae1c GW |
3294 | boolean_t do_arc_evict = B_FALSE; |
3295 | blkptr_t bp; | |
3296 | spa_t *spa = dmu_objset_spa(db->db_objset); | |
3297 | ||
3298 | if (!DBUF_IS_CACHEABLE(db) && | |
3299 | db->db_blkptr != NULL && | |
3300 | !BP_IS_HOLE(db->db_blkptr) && | |
3301 | !BP_IS_EMBEDDED(db->db_blkptr)) { | |
3302 | do_arc_evict = B_TRUE; | |
3303 | bp = *db->db_blkptr; | |
3304 | } | |
1eb5bfa3 | 3305 | |
d3c2ae1c GW |
3306 | if (!DBUF_IS_CACHEABLE(db) || |
3307 | db->db_pending_evict) { | |
3308 | dbuf_destroy(db); | |
3309 | } else if (!multilist_link_active(&db->db_cache_link)) { | |
64fc7762 | 3310 | multilist_insert(dbuf_cache, db); |
d3c2ae1c GW |
3311 | (void) refcount_add_many(&dbuf_cache_size, |
3312 | db->db.db_size, db); | |
5e021f56 GDN |
3313 | DBUF_STAT_BUMP(cache_levels[db->db_level]); |
3314 | DBUF_STAT_BUMP(cache_count); | |
3315 | DBUF_STAT_INCR(cache_levels_bytes[db->db_level], | |
3316 | db->db.db_size); | |
3317 | DBUF_STAT_MAX(cache_size_bytes_max, | |
3318 | refcount_count(&dbuf_cache_size)); | |
b128c09f | 3319 | mutex_exit(&db->db_mtx); |
d3c2ae1c GW |
3320 | |
3321 | dbuf_evict_notify(); | |
bd089c54 | 3322 | } |
d3c2ae1c GW |
3323 | |
3324 | if (do_arc_evict) | |
3325 | arc_freed(spa, &bp); | |
34dc7c2f BB |
3326 | } |
3327 | } else { | |
3328 | mutex_exit(&db->db_mtx); | |
3329 | } | |
d3c2ae1c | 3330 | |
34dc7c2f BB |
3331 | } |
3332 | ||
3333 | #pragma weak dmu_buf_refcount = dbuf_refcount | |
3334 | uint64_t | |
3335 | dbuf_refcount(dmu_buf_impl_t *db) | |
3336 | { | |
3337 | return (refcount_count(&db->db_holds)); | |
3338 | } | |
3339 | ||
3340 | void * | |
0c66c32d JG |
3341 | dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user, |
3342 | dmu_buf_user_t *new_user) | |
34dc7c2f | 3343 | { |
0c66c32d JG |
3344 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; |
3345 | ||
3346 | mutex_enter(&db->db_mtx); | |
3347 | dbuf_verify_user(db, DBVU_NOT_EVICTING); | |
3348 | if (db->db_user == old_user) | |
3349 | db->db_user = new_user; | |
3350 | else | |
3351 | old_user = db->db_user; | |
3352 | dbuf_verify_user(db, DBVU_NOT_EVICTING); | |
3353 | mutex_exit(&db->db_mtx); | |
3354 | ||
3355 | return (old_user); | |
34dc7c2f BB |
3356 | } |
3357 | ||
3358 | void * | |
0c66c32d | 3359 | dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user) |
34dc7c2f | 3360 | { |
0c66c32d | 3361 | return (dmu_buf_replace_user(db_fake, NULL, user)); |
34dc7c2f BB |
3362 | } |
3363 | ||
3364 | void * | |
0c66c32d | 3365 | dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user) |
34dc7c2f BB |
3366 | { |
3367 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
34dc7c2f | 3368 | |
bc4501f7 | 3369 | db->db_user_immediate_evict = TRUE; |
0c66c32d JG |
3370 | return (dmu_buf_set_user(db_fake, user)); |
3371 | } | |
34dc7c2f | 3372 | |
0c66c32d JG |
3373 | void * |
3374 | dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user) | |
3375 | { | |
3376 | return (dmu_buf_replace_user(db_fake, user, NULL)); | |
34dc7c2f BB |
3377 | } |
3378 | ||
3379 | void * | |
3380 | dmu_buf_get_user(dmu_buf_t *db_fake) | |
3381 | { | |
3382 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
34dc7c2f | 3383 | |
0c66c32d JG |
3384 | dbuf_verify_user(db, DBVU_NOT_EVICTING); |
3385 | return (db->db_user); | |
3386 | } | |
3387 | ||
3388 | void | |
3389 | dmu_buf_user_evict_wait() | |
3390 | { | |
3391 | taskq_wait(dbu_evict_taskq); | |
34dc7c2f BB |
3392 | } |
3393 | ||
03c6040b GW |
3394 | blkptr_t * |
3395 | dmu_buf_get_blkptr(dmu_buf_t *db) | |
3396 | { | |
3397 | dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; | |
3398 | return (dbi->db_blkptr); | |
3399 | } | |
3400 | ||
8bea9815 MA |
3401 | objset_t * |
3402 | dmu_buf_get_objset(dmu_buf_t *db) | |
3403 | { | |
3404 | dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; | |
3405 | return (dbi->db_objset); | |
3406 | } | |
3407 | ||
2bce8049 MA |
3408 | dnode_t * |
3409 | dmu_buf_dnode_enter(dmu_buf_t *db) | |
3410 | { | |
3411 | dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; | |
3412 | DB_DNODE_ENTER(dbi); | |
3413 | return (DB_DNODE(dbi)); | |
3414 | } | |
3415 | ||
3416 | void | |
3417 | dmu_buf_dnode_exit(dmu_buf_t *db) | |
3418 | { | |
3419 | dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; | |
3420 | DB_DNODE_EXIT(dbi); | |
3421 | } | |
3422 | ||
34dc7c2f BB |
3423 | static void |
3424 | dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db) | |
3425 | { | |
3426 | /* ASSERT(dmu_tx_is_syncing(tx) */ | |
3427 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
3428 | ||
3429 | if (db->db_blkptr != NULL) | |
3430 | return; | |
3431 | ||
428870ff | 3432 | if (db->db_blkid == DMU_SPILL_BLKID) { |
50c957f7 | 3433 | db->db_blkptr = DN_SPILL_BLKPTR(dn->dn_phys); |
428870ff BB |
3434 | BP_ZERO(db->db_blkptr); |
3435 | return; | |
3436 | } | |
34dc7c2f BB |
3437 | if (db->db_level == dn->dn_phys->dn_nlevels-1) { |
3438 | /* | |
3439 | * This buffer was allocated at a time when there was | |
3440 | * no available blkptrs from the dnode, or it was | |
3441 | * inappropriate to hook it in (i.e., nlevels mis-match). | |
3442 | */ | |
3443 | ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr); | |
3444 | ASSERT(db->db_parent == NULL); | |
3445 | db->db_parent = dn->dn_dbuf; | |
3446 | db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid]; | |
3447 | DBUF_VERIFY(db); | |
3448 | } else { | |
3449 | dmu_buf_impl_t *parent = db->db_parent; | |
3450 | int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; | |
3451 | ||
3452 | ASSERT(dn->dn_phys->dn_nlevels > 1); | |
3453 | if (parent == NULL) { | |
3454 | mutex_exit(&db->db_mtx); | |
3455 | rw_enter(&dn->dn_struct_rwlock, RW_READER); | |
fcff0f35 PD |
3456 | parent = dbuf_hold_level(dn, db->db_level + 1, |
3457 | db->db_blkid >> epbs, db); | |
34dc7c2f BB |
3458 | rw_exit(&dn->dn_struct_rwlock); |
3459 | mutex_enter(&db->db_mtx); | |
3460 | db->db_parent = parent; | |
3461 | } | |
3462 | db->db_blkptr = (blkptr_t *)parent->db.db_data + | |
3463 | (db->db_blkid & ((1ULL << epbs) - 1)); | |
3464 | DBUF_VERIFY(db); | |
3465 | } | |
3466 | } | |
3467 | ||
b5256303 | 3468 | /* |
0c03d21a MA |
3469 | * When syncing out a blocks of dnodes, adjust the block to deal with |
3470 | * encryption. Normally, we make sure the block is decrypted before writing | |
3471 | * it. If we have crypt params, then we are writing a raw (encrypted) block, | |
3472 | * from a raw receive. In this case, set the ARC buf's crypt params so | |
3473 | * that the BP will be filled with the correct byteorder, salt, iv, and mac. | |
b5256303 TC |
3474 | */ |
3475 | static void | |
0c03d21a | 3476 | dbuf_prepare_encrypted_dnode_leaf(dbuf_dirty_record_t *dr) |
b5256303 TC |
3477 | { |
3478 | int err; | |
3479 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
3480 | ||
3481 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
0c03d21a MA |
3482 | ASSERT3U(db->db.db_object, ==, DMU_META_DNODE_OBJECT); |
3483 | ASSERT3U(db->db_level, ==, 0); | |
b5256303 | 3484 | |
0c03d21a | 3485 | if (!db->db_objset->os_raw_receive && arc_is_encrypted(db->db_buf)) { |
a2c2ed1b TC |
3486 | zbookmark_phys_t zb; |
3487 | ||
b5256303 TC |
3488 | /* |
3489 | * Unfortunately, there is currently no mechanism for | |
3490 | * syncing context to handle decryption errors. An error | |
3491 | * here is only possible if an attacker maliciously | |
3492 | * changed a dnode block and updated the associated | |
3493 | * checksums going up the block tree. | |
3494 | */ | |
a2c2ed1b TC |
3495 | SET_BOOKMARK(&zb, dmu_objset_id(db->db_objset), |
3496 | db->db.db_object, db->db_level, db->db_blkid); | |
b5256303 | 3497 | err = arc_untransform(db->db_buf, db->db_objset->os_spa, |
a2c2ed1b | 3498 | &zb, B_TRUE); |
b5256303 TC |
3499 | if (err) |
3500 | panic("Invalid dnode block MAC"); | |
0c03d21a MA |
3501 | } else if (dr->dt.dl.dr_has_raw_params) { |
3502 | (void) arc_release(dr->dt.dl.dr_data, db); | |
3503 | arc_convert_to_raw(dr->dt.dl.dr_data, | |
3504 | dmu_objset_id(db->db_objset), | |
3505 | dr->dt.dl.dr_byteorder, DMU_OT_DNODE, | |
3506 | dr->dt.dl.dr_salt, dr->dt.dl.dr_iv, dr->dt.dl.dr_mac); | |
b5256303 TC |
3507 | } |
3508 | } | |
3509 | ||
d1d7e268 MK |
3510 | /* |
3511 | * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it | |
60948de1 BB |
3512 | * is critical the we not allow the compiler to inline this function in to |
3513 | * dbuf_sync_list() thereby drastically bloating the stack usage. | |
3514 | */ | |
3515 | noinline static void | |
34dc7c2f BB |
3516 | dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx) |
3517 | { | |
3518 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
572e2857 | 3519 | dnode_t *dn; |
34dc7c2f BB |
3520 | zio_t *zio; |
3521 | ||
3522 | ASSERT(dmu_tx_is_syncing(tx)); | |
3523 | ||
3524 | dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); | |
3525 | ||
3526 | mutex_enter(&db->db_mtx); | |
3527 | ||
3528 | ASSERT(db->db_level > 0); | |
3529 | DBUF_VERIFY(db); | |
3530 | ||
e49f1e20 | 3531 | /* Read the block if it hasn't been read yet. */ |
34dc7c2f BB |
3532 | if (db->db_buf == NULL) { |
3533 | mutex_exit(&db->db_mtx); | |
3534 | (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED); | |
3535 | mutex_enter(&db->db_mtx); | |
3536 | } | |
3537 | ASSERT3U(db->db_state, ==, DB_CACHED); | |
34dc7c2f BB |
3538 | ASSERT(db->db_buf != NULL); |
3539 | ||
572e2857 BB |
3540 | DB_DNODE_ENTER(db); |
3541 | dn = DB_DNODE(db); | |
e49f1e20 | 3542 | /* Indirect block size must match what the dnode thinks it is. */ |
572e2857 | 3543 | ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); |
34dc7c2f | 3544 | dbuf_check_blkptr(dn, db); |
572e2857 | 3545 | DB_DNODE_EXIT(db); |
34dc7c2f | 3546 | |
e49f1e20 | 3547 | /* Provide the pending dirty record to child dbufs */ |
34dc7c2f BB |
3548 | db->db_data_pending = dr; |
3549 | ||
34dc7c2f | 3550 | mutex_exit(&db->db_mtx); |
a1d477c2 | 3551 | |
b128c09f | 3552 | dbuf_write(dr, db->db_buf, tx); |
34dc7c2f BB |
3553 | |
3554 | zio = dr->dr_zio; | |
3555 | mutex_enter(&dr->dt.di.dr_mtx); | |
4bda3bd0 | 3556 | dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx); |
34dc7c2f BB |
3557 | ASSERT(list_head(&dr->dt.di.dr_children) == NULL); |
3558 | mutex_exit(&dr->dt.di.dr_mtx); | |
3559 | zio_nowait(zio); | |
3560 | } | |
3561 | ||
d1d7e268 MK |
3562 | /* |
3563 | * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is | |
60948de1 BB |
3564 | * critical the we not allow the compiler to inline this function in to |
3565 | * dbuf_sync_list() thereby drastically bloating the stack usage. | |
3566 | */ | |
3567 | noinline static void | |
34dc7c2f BB |
3568 | dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx) |
3569 | { | |
3570 | arc_buf_t **datap = &dr->dt.dl.dr_data; | |
3571 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
572e2857 BB |
3572 | dnode_t *dn; |
3573 | objset_t *os; | |
34dc7c2f | 3574 | uint64_t txg = tx->tx_txg; |
34dc7c2f BB |
3575 | |
3576 | ASSERT(dmu_tx_is_syncing(tx)); | |
3577 | ||
3578 | dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); | |
3579 | ||
3580 | mutex_enter(&db->db_mtx); | |
3581 | /* | |
3582 | * To be synced, we must be dirtied. But we | |
3583 | * might have been freed after the dirty. | |
3584 | */ | |
3585 | if (db->db_state == DB_UNCACHED) { | |
3586 | /* This buffer has been freed since it was dirtied */ | |
3587 | ASSERT(db->db.db_data == NULL); | |
3588 | } else if (db->db_state == DB_FILL) { | |
3589 | /* This buffer was freed and is now being re-filled */ | |
3590 | ASSERT(db->db.db_data != dr->dt.dl.dr_data); | |
3591 | } else { | |
b128c09f | 3592 | ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL); |
34dc7c2f BB |
3593 | } |
3594 | DBUF_VERIFY(db); | |
3595 | ||
572e2857 BB |
3596 | DB_DNODE_ENTER(db); |
3597 | dn = DB_DNODE(db); | |
3598 | ||
428870ff BB |
3599 | if (db->db_blkid == DMU_SPILL_BLKID) { |
3600 | mutex_enter(&dn->dn_mtx); | |
81edd3e8 P |
3601 | if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) { |
3602 | /* | |
3603 | * In the previous transaction group, the bonus buffer | |
3604 | * was entirely used to store the attributes for the | |
3605 | * dnode which overrode the dn_spill field. However, | |
3606 | * when adding more attributes to the file a spill | |
3607 | * block was required to hold the extra attributes. | |
3608 | * | |
3609 | * Make sure to clear the garbage left in the dn_spill | |
3610 | * field from the previous attributes in the bonus | |
3611 | * buffer. Otherwise, after writing out the spill | |
3612 | * block to the new allocated dva, it will free | |
3613 | * the old block pointed to by the invalid dn_spill. | |
3614 | */ | |
3615 | db->db_blkptr = NULL; | |
3616 | } | |
428870ff BB |
3617 | dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR; |
3618 | mutex_exit(&dn->dn_mtx); | |
3619 | } | |
3620 | ||
34dc7c2f BB |
3621 | /* |
3622 | * If this is a bonus buffer, simply copy the bonus data into the | |
3623 | * dnode. It will be written out when the dnode is synced (and it | |
3624 | * will be synced, since it must have been dirty for dbuf_sync to | |
3625 | * be called). | |
3626 | */ | |
428870ff | 3627 | if (db->db_blkid == DMU_BONUS_BLKID) { |
34dc7c2f BB |
3628 | dbuf_dirty_record_t **drp; |
3629 | ||
3630 | ASSERT(*datap != NULL); | |
c99c9001 | 3631 | ASSERT0(db->db_level); |
b5256303 | 3632 | ASSERT3U(DN_MAX_BONUS_LEN(dn->dn_phys), <=, |
50c957f7 | 3633 | DN_SLOTS_TO_BONUSLEN(dn->dn_phys->dn_extra_slots + 1)); |
b5256303 TC |
3634 | bcopy(*datap, DN_BONUS(dn->dn_phys), |
3635 | DN_MAX_BONUS_LEN(dn->dn_phys)); | |
572e2857 BB |
3636 | DB_DNODE_EXIT(db); |
3637 | ||
34dc7c2f | 3638 | if (*datap != db->db.db_data) { |
50c957f7 NB |
3639 | int slots = DB_DNODE(db)->dn_num_slots; |
3640 | int bonuslen = DN_SLOTS_TO_BONUSLEN(slots); | |
a3fd9d9e | 3641 | kmem_free(*datap, bonuslen); |
25458cbe | 3642 | arc_space_return(bonuslen, ARC_SPACE_BONUS); |
34dc7c2f BB |
3643 | } |
3644 | db->db_data_pending = NULL; | |
3645 | drp = &db->db_last_dirty; | |
3646 | while (*drp != dr) | |
3647 | drp = &(*drp)->dr_next; | |
3648 | ASSERT(dr->dr_next == NULL); | |
428870ff | 3649 | ASSERT(dr->dr_dbuf == db); |
34dc7c2f | 3650 | *drp = dr->dr_next; |
753972fc BB |
3651 | if (dr->dr_dbuf->db_level != 0) { |
3652 | mutex_destroy(&dr->dt.di.dr_mtx); | |
3653 | list_destroy(&dr->dt.di.dr_children); | |
3654 | } | |
34dc7c2f BB |
3655 | kmem_free(dr, sizeof (dbuf_dirty_record_t)); |
3656 | ASSERT(db->db_dirtycnt > 0); | |
3657 | db->db_dirtycnt -= 1; | |
428870ff | 3658 | dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg); |
34dc7c2f BB |
3659 | return; |
3660 | } | |
3661 | ||
572e2857 BB |
3662 | os = dn->dn_objset; |
3663 | ||
34dc7c2f BB |
3664 | /* |
3665 | * This function may have dropped the db_mtx lock allowing a dmu_sync | |
3666 | * operation to sneak in. As a result, we need to ensure that we | |
3667 | * don't check the dr_override_state until we have returned from | |
3668 | * dbuf_check_blkptr. | |
3669 | */ | |
3670 | dbuf_check_blkptr(dn, db); | |
3671 | ||
3672 | /* | |
572e2857 | 3673 | * If this buffer is in the middle of an immediate write, |
34dc7c2f BB |
3674 | * wait for the synchronous IO to complete. |
3675 | */ | |
3676 | while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) { | |
3677 | ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT); | |
3678 | cv_wait(&db->db_changed, &db->db_mtx); | |
3679 | ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN); | |
3680 | } | |
3681 | ||
b5256303 TC |
3682 | /* |
3683 | * If this is a dnode block, ensure it is appropriately encrypted | |
3684 | * or decrypted, depending on what we are writing to it this txg. | |
3685 | */ | |
3686 | if (os->os_encrypted && dn->dn_object == DMU_META_DNODE_OBJECT) | |
0c03d21a | 3687 | dbuf_prepare_encrypted_dnode_leaf(dr); |
b5256303 | 3688 | |
9babb374 BB |
3689 | if (db->db_state != DB_NOFILL && |
3690 | dn->dn_object != DMU_META_DNODE_OBJECT && | |
3691 | refcount_count(&db->db_holds) > 1 && | |
428870ff | 3692 | dr->dt.dl.dr_override_state != DR_OVERRIDDEN && |
9babb374 BB |
3693 | *datap == db->db_buf) { |
3694 | /* | |
3695 | * If this buffer is currently "in use" (i.e., there | |
3696 | * are active holds and db_data still references it), | |
3697 | * then make a copy before we start the write so that | |
3698 | * any modifications from the open txg will not leak | |
3699 | * into this write. | |
3700 | * | |
3701 | * NOTE: this copy does not need to be made for | |
3702 | * objects only modified in the syncing context (e.g. | |
3703 | * DNONE_DNODE blocks). | |
3704 | */ | |
2aa34383 | 3705 | int psize = arc_buf_size(*datap); |
b5256303 | 3706 | int lsize = arc_buf_lsize(*datap); |
9babb374 | 3707 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); |
2aa34383 DK |
3708 | enum zio_compress compress_type = arc_get_compression(*datap); |
3709 | ||
b5256303 TC |
3710 | if (arc_is_encrypted(*datap)) { |
3711 | boolean_t byteorder; | |
3712 | uint8_t salt[ZIO_DATA_SALT_LEN]; | |
3713 | uint8_t iv[ZIO_DATA_IV_LEN]; | |
3714 | uint8_t mac[ZIO_DATA_MAC_LEN]; | |
3715 | ||
3716 | arc_get_raw_params(*datap, &byteorder, salt, iv, mac); | |
3717 | *datap = arc_alloc_raw_buf(os->os_spa, db, | |
3718 | dmu_objset_id(os), byteorder, salt, iv, mac, | |
3719 | dn->dn_type, psize, lsize, compress_type); | |
3720 | } else if (compress_type != ZIO_COMPRESS_OFF) { | |
2aa34383 DK |
3721 | ASSERT3U(type, ==, ARC_BUFC_DATA); |
3722 | *datap = arc_alloc_compressed_buf(os->os_spa, db, | |
3723 | psize, lsize, compress_type); | |
b5256303 TC |
3724 | } else { |
3725 | *datap = arc_alloc_buf(os->os_spa, db, type, psize); | |
2aa34383 DK |
3726 | } |
3727 | bcopy(db->db.db_data, (*datap)->b_data, psize); | |
b128c09f | 3728 | } |
34dc7c2f BB |
3729 | db->db_data_pending = dr; |
3730 | ||
3731 | mutex_exit(&db->db_mtx); | |
3732 | ||
b128c09f | 3733 | dbuf_write(dr, *datap, tx); |
34dc7c2f BB |
3734 | |
3735 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
572e2857 | 3736 | if (dn->dn_object == DMU_META_DNODE_OBJECT) { |
34dc7c2f | 3737 | list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr); |
572e2857 BB |
3738 | DB_DNODE_EXIT(db); |
3739 | } else { | |
3740 | /* | |
3741 | * Although zio_nowait() does not "wait for an IO", it does | |
3742 | * initiate the IO. If this is an empty write it seems plausible | |
3743 | * that the IO could actually be completed before the nowait | |
3744 | * returns. We need to DB_DNODE_EXIT() first in case | |
3745 | * zio_nowait() invalidates the dbuf. | |
3746 | */ | |
3747 | DB_DNODE_EXIT(db); | |
34dc7c2f | 3748 | zio_nowait(dr->dr_zio); |
572e2857 | 3749 | } |
34dc7c2f BB |
3750 | } |
3751 | ||
3752 | void | |
4bda3bd0 | 3753 | dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx) |
34dc7c2f BB |
3754 | { |
3755 | dbuf_dirty_record_t *dr; | |
3756 | ||
c65aa5b2 | 3757 | while ((dr = list_head(list))) { |
34dc7c2f BB |
3758 | if (dr->dr_zio != NULL) { |
3759 | /* | |
3760 | * If we find an already initialized zio then we | |
3761 | * are processing the meta-dnode, and we have finished. | |
3762 | * The dbufs for all dnodes are put back on the list | |
3763 | * during processing, so that we can zio_wait() | |
3764 | * these IOs after initiating all child IOs. | |
3765 | */ | |
3766 | ASSERT3U(dr->dr_dbuf->db.db_object, ==, | |
3767 | DMU_META_DNODE_OBJECT); | |
3768 | break; | |
3769 | } | |
4bda3bd0 MA |
3770 | if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID && |
3771 | dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) { | |
3772 | VERIFY3U(dr->dr_dbuf->db_level, ==, level); | |
3773 | } | |
34dc7c2f BB |
3774 | list_remove(list, dr); |
3775 | if (dr->dr_dbuf->db_level > 0) | |
3776 | dbuf_sync_indirect(dr, tx); | |
3777 | else | |
3778 | dbuf_sync_leaf(dr, tx); | |
3779 | } | |
3780 | } | |
3781 | ||
34dc7c2f BB |
3782 | /* ARGSUSED */ |
3783 | static void | |
3784 | dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb) | |
3785 | { | |
3786 | dmu_buf_impl_t *db = vdb; | |
572e2857 | 3787 | dnode_t *dn; |
b128c09f | 3788 | blkptr_t *bp = zio->io_bp; |
34dc7c2f | 3789 | blkptr_t *bp_orig = &zio->io_bp_orig; |
428870ff BB |
3790 | spa_t *spa = zio->io_spa; |
3791 | int64_t delta; | |
34dc7c2f | 3792 | uint64_t fill = 0; |
428870ff | 3793 | int i; |
34dc7c2f | 3794 | |
463a8cfe AR |
3795 | ASSERT3P(db->db_blkptr, !=, NULL); |
3796 | ASSERT3P(&db->db_data_pending->dr_bp_copy, ==, bp); | |
b128c09f | 3797 | |
572e2857 BB |
3798 | DB_DNODE_ENTER(db); |
3799 | dn = DB_DNODE(db); | |
428870ff BB |
3800 | delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig); |
3801 | dnode_diduse_space(dn, delta - zio->io_prev_space_delta); | |
3802 | zio->io_prev_space_delta = delta; | |
34dc7c2f | 3803 | |
b0bc7a84 MG |
3804 | if (bp->blk_birth != 0) { |
3805 | ASSERT((db->db_blkid != DMU_SPILL_BLKID && | |
3806 | BP_GET_TYPE(bp) == dn->dn_type) || | |
3807 | (db->db_blkid == DMU_SPILL_BLKID && | |
9b67f605 MA |
3808 | BP_GET_TYPE(bp) == dn->dn_bonustype) || |
3809 | BP_IS_EMBEDDED(bp)); | |
b0bc7a84 | 3810 | ASSERT(BP_GET_LEVEL(bp) == db->db_level); |
34dc7c2f BB |
3811 | } |
3812 | ||
3813 | mutex_enter(&db->db_mtx); | |
3814 | ||
428870ff BB |
3815 | #ifdef ZFS_DEBUG |
3816 | if (db->db_blkid == DMU_SPILL_BLKID) { | |
428870ff | 3817 | ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); |
463a8cfe | 3818 | ASSERT(!(BP_IS_HOLE(bp)) && |
50c957f7 | 3819 | db->db_blkptr == DN_SPILL_BLKPTR(dn->dn_phys)); |
428870ff BB |
3820 | } |
3821 | #endif | |
3822 | ||
34dc7c2f BB |
3823 | if (db->db_level == 0) { |
3824 | mutex_enter(&dn->dn_mtx); | |
428870ff BB |
3825 | if (db->db_blkid > dn->dn_phys->dn_maxblkid && |
3826 | db->db_blkid != DMU_SPILL_BLKID) | |
34dc7c2f BB |
3827 | dn->dn_phys->dn_maxblkid = db->db_blkid; |
3828 | mutex_exit(&dn->dn_mtx); | |
3829 | ||
3830 | if (dn->dn_type == DMU_OT_DNODE) { | |
50c957f7 NB |
3831 | i = 0; |
3832 | while (i < db->db.db_size) { | |
817b1b6e MA |
3833 | dnode_phys_t *dnp = |
3834 | (void *)(((char *)db->db.db_data) + i); | |
50c957f7 NB |
3835 | |
3836 | i += DNODE_MIN_SIZE; | |
3837 | if (dnp->dn_type != DMU_OT_NONE) { | |
34dc7c2f | 3838 | fill++; |
50c957f7 NB |
3839 | i += dnp->dn_extra_slots * |
3840 | DNODE_MIN_SIZE; | |
3841 | } | |
34dc7c2f BB |
3842 | } |
3843 | } else { | |
b0bc7a84 MG |
3844 | if (BP_IS_HOLE(bp)) { |
3845 | fill = 0; | |
3846 | } else { | |
3847 | fill = 1; | |
3848 | } | |
34dc7c2f BB |
3849 | } |
3850 | } else { | |
b128c09f | 3851 | blkptr_t *ibp = db->db.db_data; |
34dc7c2f | 3852 | ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); |
b128c09f BB |
3853 | for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) { |
3854 | if (BP_IS_HOLE(ibp)) | |
34dc7c2f | 3855 | continue; |
9b67f605 | 3856 | fill += BP_GET_FILL(ibp); |
34dc7c2f BB |
3857 | } |
3858 | } | |
572e2857 | 3859 | DB_DNODE_EXIT(db); |
34dc7c2f | 3860 | |
9b67f605 | 3861 | if (!BP_IS_EMBEDDED(bp)) |
b5256303 | 3862 | BP_SET_FILL(bp, fill); |
34dc7c2f BB |
3863 | |
3864 | mutex_exit(&db->db_mtx); | |
463a8cfe AR |
3865 | |
3866 | rw_enter(&dn->dn_struct_rwlock, RW_WRITER); | |
3867 | *db->db_blkptr = *bp; | |
3868 | rw_exit(&dn->dn_struct_rwlock); | |
34dc7c2f BB |
3869 | } |
3870 | ||
bc77ba73 PD |
3871 | /* ARGSUSED */ |
3872 | /* | |
3873 | * This function gets called just prior to running through the compression | |
3874 | * stage of the zio pipeline. If we're an indirect block comprised of only | |
3875 | * holes, then we want this indirect to be compressed away to a hole. In | |
3876 | * order to do that we must zero out any information about the holes that | |
3877 | * this indirect points to prior to before we try to compress it. | |
3878 | */ | |
3879 | static void | |
3880 | dbuf_write_children_ready(zio_t *zio, arc_buf_t *buf, void *vdb) | |
3881 | { | |
3882 | dmu_buf_impl_t *db = vdb; | |
3883 | dnode_t *dn; | |
3884 | blkptr_t *bp; | |
721ed0ee | 3885 | unsigned int epbs, i; |
bc77ba73 PD |
3886 | |
3887 | ASSERT3U(db->db_level, >, 0); | |
3888 | DB_DNODE_ENTER(db); | |
3889 | dn = DB_DNODE(db); | |
3890 | epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; | |
721ed0ee | 3891 | ASSERT3U(epbs, <, 31); |
bc77ba73 PD |
3892 | |
3893 | /* Determine if all our children are holes */ | |
3f93077b | 3894 | for (i = 0, bp = db->db.db_data; i < 1ULL << epbs; i++, bp++) { |
bc77ba73 PD |
3895 | if (!BP_IS_HOLE(bp)) |
3896 | break; | |
3897 | } | |
3898 | ||
3899 | /* | |
3900 | * If all the children are holes, then zero them all out so that | |
3901 | * we may get compressed away. | |
3902 | */ | |
3f93077b | 3903 | if (i == 1ULL << epbs) { |
721ed0ee GM |
3904 | /* |
3905 | * We only found holes. Grab the rwlock to prevent | |
3906 | * anybody from reading the blocks we're about to | |
3907 | * zero out. | |
3908 | */ | |
3909 | rw_enter(&dn->dn_struct_rwlock, RW_WRITER); | |
bc77ba73 | 3910 | bzero(db->db.db_data, db->db.db_size); |
721ed0ee | 3911 | rw_exit(&dn->dn_struct_rwlock); |
bc77ba73 PD |
3912 | } |
3913 | DB_DNODE_EXIT(db); | |
3914 | } | |
3915 | ||
e8b96c60 MA |
3916 | /* |
3917 | * The SPA will call this callback several times for each zio - once | |
3918 | * for every physical child i/o (zio->io_phys_children times). This | |
3919 | * allows the DMU to monitor the progress of each logical i/o. For example, | |
3920 | * there may be 2 copies of an indirect block, or many fragments of a RAID-Z | |
3921 | * block. There may be a long delay before all copies/fragments are completed, | |
3922 | * so this callback allows us to retire dirty space gradually, as the physical | |
3923 | * i/os complete. | |
3924 | */ | |
3925 | /* ARGSUSED */ | |
3926 | static void | |
3927 | dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg) | |
3928 | { | |
3929 | dmu_buf_impl_t *db = arg; | |
3930 | objset_t *os = db->db_objset; | |
3931 | dsl_pool_t *dp = dmu_objset_pool(os); | |
3932 | dbuf_dirty_record_t *dr; | |
3933 | int delta = 0; | |
3934 | ||
3935 | dr = db->db_data_pending; | |
3936 | ASSERT3U(dr->dr_txg, ==, zio->io_txg); | |
3937 | ||
3938 | /* | |
3939 | * The callback will be called io_phys_children times. Retire one | |
3940 | * portion of our dirty space each time we are called. Any rounding | |
3941 | * error will be cleaned up by dsl_pool_sync()'s call to | |
3942 | * dsl_pool_undirty_space(). | |
3943 | */ | |
3944 | delta = dr->dr_accounted / zio->io_phys_children; | |
3945 | dsl_pool_undirty_space(dp, delta, zio->io_txg); | |
3946 | } | |
3947 | ||
34dc7c2f BB |
3948 | /* ARGSUSED */ |
3949 | static void | |
3950 | dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb) | |
3951 | { | |
3952 | dmu_buf_impl_t *db = vdb; | |
428870ff | 3953 | blkptr_t *bp_orig = &zio->io_bp_orig; |
b0bc7a84 MG |
3954 | blkptr_t *bp = db->db_blkptr; |
3955 | objset_t *os = db->db_objset; | |
3956 | dmu_tx_t *tx = os->os_synctx; | |
34dc7c2f BB |
3957 | dbuf_dirty_record_t **drp, *dr; |
3958 | ||
c99c9001 | 3959 | ASSERT0(zio->io_error); |
428870ff BB |
3960 | ASSERT(db->db_blkptr == bp); |
3961 | ||
03c6040b GW |
3962 | /* |
3963 | * For nopwrites and rewrites we ensure that the bp matches our | |
3964 | * original and bypass all the accounting. | |
3965 | */ | |
3966 | if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) { | |
428870ff BB |
3967 | ASSERT(BP_EQUAL(bp, bp_orig)); |
3968 | } else { | |
b0bc7a84 | 3969 | dsl_dataset_t *ds = os->os_dsl_dataset; |
428870ff BB |
3970 | (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE); |
3971 | dsl_dataset_block_born(ds, bp, tx); | |
3972 | } | |
34dc7c2f BB |
3973 | |
3974 | mutex_enter(&db->db_mtx); | |
3975 | ||
428870ff BB |
3976 | DBUF_VERIFY(db); |
3977 | ||
34dc7c2f BB |
3978 | drp = &db->db_last_dirty; |
3979 | while ((dr = *drp) != db->db_data_pending) | |
3980 | drp = &dr->dr_next; | |
3981 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
428870ff | 3982 | ASSERT(dr->dr_dbuf == db); |
34dc7c2f BB |
3983 | ASSERT(dr->dr_next == NULL); |
3984 | *drp = dr->dr_next; | |
3985 | ||
428870ff BB |
3986 | #ifdef ZFS_DEBUG |
3987 | if (db->db_blkid == DMU_SPILL_BLKID) { | |
572e2857 BB |
3988 | dnode_t *dn; |
3989 | ||
3990 | DB_DNODE_ENTER(db); | |
3991 | dn = DB_DNODE(db); | |
428870ff BB |
3992 | ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); |
3993 | ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && | |
50c957f7 | 3994 | db->db_blkptr == DN_SPILL_BLKPTR(dn->dn_phys)); |
572e2857 | 3995 | DB_DNODE_EXIT(db); |
428870ff BB |
3996 | } |
3997 | #endif | |
3998 | ||
34dc7c2f | 3999 | if (db->db_level == 0) { |
428870ff | 4000 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f | 4001 | ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); |
b128c09f BB |
4002 | if (db->db_state != DB_NOFILL) { |
4003 | if (dr->dt.dl.dr_data != db->db_buf) | |
d3c2ae1c | 4004 | arc_buf_destroy(dr->dt.dl.dr_data, db); |
b128c09f | 4005 | } |
34dc7c2f | 4006 | } else { |
572e2857 BB |
4007 | dnode_t *dn; |
4008 | ||
4009 | DB_DNODE_ENTER(db); | |
4010 | dn = DB_DNODE(db); | |
34dc7c2f | 4011 | ASSERT(list_head(&dr->dt.di.dr_children) == NULL); |
b0bc7a84 | 4012 | ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift); |
34dc7c2f | 4013 | if (!BP_IS_HOLE(db->db_blkptr)) { |
1fde1e37 BB |
4014 | ASSERTV(int epbs = dn->dn_phys->dn_indblkshift - |
4015 | SPA_BLKPTRSHIFT); | |
b0bc7a84 MG |
4016 | ASSERT3U(db->db_blkid, <=, |
4017 | dn->dn_phys->dn_maxblkid >> (db->db_level * epbs)); | |
34dc7c2f BB |
4018 | ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, |
4019 | db->db.db_size); | |
34dc7c2f | 4020 | } |
572e2857 | 4021 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
4022 | mutex_destroy(&dr->dt.di.dr_mtx); |
4023 | list_destroy(&dr->dt.di.dr_children); | |
4024 | } | |
4025 | kmem_free(dr, sizeof (dbuf_dirty_record_t)); | |
4026 | ||
4027 | cv_broadcast(&db->db_changed); | |
4028 | ASSERT(db->db_dirtycnt > 0); | |
4029 | db->db_dirtycnt -= 1; | |
4030 | db->db_data_pending = NULL; | |
b0bc7a84 | 4031 | dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg); |
428870ff BB |
4032 | } |
4033 | ||
4034 | static void | |
4035 | dbuf_write_nofill_ready(zio_t *zio) | |
4036 | { | |
4037 | dbuf_write_ready(zio, NULL, zio->io_private); | |
4038 | } | |
4039 | ||
4040 | static void | |
4041 | dbuf_write_nofill_done(zio_t *zio) | |
4042 | { | |
4043 | dbuf_write_done(zio, NULL, zio->io_private); | |
4044 | } | |
4045 | ||
4046 | static void | |
4047 | dbuf_write_override_ready(zio_t *zio) | |
4048 | { | |
4049 | dbuf_dirty_record_t *dr = zio->io_private; | |
4050 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
4051 | ||
4052 | dbuf_write_ready(zio, NULL, db); | |
4053 | } | |
4054 | ||
4055 | static void | |
4056 | dbuf_write_override_done(zio_t *zio) | |
4057 | { | |
4058 | dbuf_dirty_record_t *dr = zio->io_private; | |
4059 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
4060 | blkptr_t *obp = &dr->dt.dl.dr_overridden_by; | |
4061 | ||
4062 | mutex_enter(&db->db_mtx); | |
4063 | if (!BP_EQUAL(zio->io_bp, obp)) { | |
4064 | if (!BP_IS_HOLE(obp)) | |
4065 | dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp); | |
4066 | arc_release(dr->dt.dl.dr_data, db); | |
4067 | } | |
34dc7c2f BB |
4068 | mutex_exit(&db->db_mtx); |
4069 | ||
428870ff | 4070 | dbuf_write_done(zio, NULL, db); |
a6255b7f DQ |
4071 | |
4072 | if (zio->io_abd != NULL) | |
4073 | abd_put(zio->io_abd); | |
428870ff BB |
4074 | } |
4075 | ||
a1d477c2 MA |
4076 | typedef struct dbuf_remap_impl_callback_arg { |
4077 | objset_t *drica_os; | |
4078 | uint64_t drica_blk_birth; | |
4079 | dmu_tx_t *drica_tx; | |
4080 | } dbuf_remap_impl_callback_arg_t; | |
4081 | ||
4082 | static void | |
4083 | dbuf_remap_impl_callback(uint64_t vdev, uint64_t offset, uint64_t size, | |
4084 | void *arg) | |
4085 | { | |
4086 | dbuf_remap_impl_callback_arg_t *drica = arg; | |
4087 | objset_t *os = drica->drica_os; | |
4088 | spa_t *spa = dmu_objset_spa(os); | |
4089 | dmu_tx_t *tx = drica->drica_tx; | |
4090 | ||
4091 | ASSERT(dsl_pool_sync_context(spa_get_dsl(spa))); | |
4092 | ||
4093 | if (os == spa_meta_objset(spa)) { | |
4094 | spa_vdev_indirect_mark_obsolete(spa, vdev, offset, size, tx); | |
4095 | } else { | |
4096 | dsl_dataset_block_remapped(dmu_objset_ds(os), vdev, offset, | |
4097 | size, drica->drica_blk_birth, tx); | |
4098 | } | |
4099 | } | |
4100 | ||
4101 | static void | |
4102 | dbuf_remap_impl(dnode_t *dn, blkptr_t *bp, dmu_tx_t *tx) | |
4103 | { | |
4104 | blkptr_t bp_copy = *bp; | |
4105 | spa_t *spa = dmu_objset_spa(dn->dn_objset); | |
4106 | dbuf_remap_impl_callback_arg_t drica; | |
4107 | ||
4108 | ASSERT(dsl_pool_sync_context(spa_get_dsl(spa))); | |
4109 | ||
4110 | drica.drica_os = dn->dn_objset; | |
4111 | drica.drica_blk_birth = bp->blk_birth; | |
4112 | drica.drica_tx = tx; | |
4113 | if (spa_remap_blkptr(spa, &bp_copy, dbuf_remap_impl_callback, | |
4114 | &drica)) { | |
4115 | /* | |
4116 | * The struct_rwlock prevents dbuf_read_impl() from | |
4117 | * dereferencing the BP while we are changing it. To | |
4118 | * avoid lock contention, only grab it when we are actually | |
4119 | * changing the BP. | |
4120 | */ | |
4121 | rw_enter(&dn->dn_struct_rwlock, RW_WRITER); | |
4122 | *bp = bp_copy; | |
4123 | rw_exit(&dn->dn_struct_rwlock); | |
4124 | } | |
4125 | } | |
4126 | ||
4127 | /* | |
4128 | * Returns true if a dbuf_remap would modify the dbuf. We do this by attempting | |
4129 | * to remap a copy of every bp in the dbuf. | |
4130 | */ | |
4131 | boolean_t | |
4132 | dbuf_can_remap(const dmu_buf_impl_t *db) | |
4133 | { | |
4134 | spa_t *spa = dmu_objset_spa(db->db_objset); | |
4135 | blkptr_t *bp = db->db.db_data; | |
4136 | boolean_t ret = B_FALSE; | |
4137 | ||
4138 | ASSERT3U(db->db_level, >, 0); | |
4139 | ASSERT3S(db->db_state, ==, DB_CACHED); | |
4140 | ||
4141 | ASSERT(spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL)); | |
4142 | ||
4143 | spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER); | |
4144 | for (int i = 0; i < db->db.db_size >> SPA_BLKPTRSHIFT; i++) { | |
4145 | blkptr_t bp_copy = bp[i]; | |
4146 | if (spa_remap_blkptr(spa, &bp_copy, NULL, NULL)) { | |
4147 | ret = B_TRUE; | |
4148 | break; | |
4149 | } | |
4150 | } | |
4151 | spa_config_exit(spa, SCL_VDEV, FTAG); | |
4152 | ||
4153 | return (ret); | |
4154 | } | |
4155 | ||
4156 | boolean_t | |
4157 | dnode_needs_remap(const dnode_t *dn) | |
4158 | { | |
4159 | spa_t *spa = dmu_objset_spa(dn->dn_objset); | |
4160 | boolean_t ret = B_FALSE; | |
4161 | ||
4162 | if (dn->dn_phys->dn_nlevels == 0) { | |
4163 | return (B_FALSE); | |
4164 | } | |
4165 | ||
4166 | ASSERT(spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL)); | |
4167 | ||
4168 | spa_config_enter(spa, SCL_VDEV, FTAG, RW_READER); | |
4169 | for (int j = 0; j < dn->dn_phys->dn_nblkptr; j++) { | |
4170 | blkptr_t bp_copy = dn->dn_phys->dn_blkptr[j]; | |
4171 | if (spa_remap_blkptr(spa, &bp_copy, NULL, NULL)) { | |
4172 | ret = B_TRUE; | |
4173 | break; | |
4174 | } | |
4175 | } | |
4176 | spa_config_exit(spa, SCL_VDEV, FTAG); | |
4177 | ||
4178 | return (ret); | |
4179 | } | |
4180 | ||
4181 | /* | |
4182 | * Remap any existing BP's to concrete vdevs, if possible. | |
4183 | */ | |
4184 | static void | |
4185 | dbuf_remap(dnode_t *dn, dmu_buf_impl_t *db, dmu_tx_t *tx) | |
4186 | { | |
4187 | spa_t *spa = dmu_objset_spa(db->db_objset); | |
4188 | ASSERT(dsl_pool_sync_context(spa_get_dsl(spa))); | |
4189 | ||
4190 | if (!spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL)) | |
4191 | return; | |
4192 | ||
4193 | if (db->db_level > 0) { | |
4194 | blkptr_t *bp = db->db.db_data; | |
4195 | for (int i = 0; i < db->db.db_size >> SPA_BLKPTRSHIFT; i++) { | |
4196 | dbuf_remap_impl(dn, &bp[i], tx); | |
4197 | } | |
4198 | } else if (db->db.db_object == DMU_META_DNODE_OBJECT) { | |
4199 | dnode_phys_t *dnp = db->db.db_data; | |
4200 | ASSERT3U(db->db_dnode_handle->dnh_dnode->dn_type, ==, | |
4201 | DMU_OT_DNODE); | |
4202 | for (int i = 0; i < db->db.db_size >> DNODE_SHIFT; | |
4203 | i += dnp[i].dn_extra_slots + 1) { | |
4204 | for (int j = 0; j < dnp[i].dn_nblkptr; j++) { | |
4205 | dbuf_remap_impl(dn, &dnp[i].dn_blkptr[j], tx); | |
4206 | } | |
4207 | } | |
4208 | } | |
4209 | } | |
4210 | ||
4211 | ||
e49f1e20 | 4212 | /* Issue I/O to commit a dirty buffer to disk. */ |
428870ff BB |
4213 | static void |
4214 | dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx) | |
4215 | { | |
4216 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
572e2857 BB |
4217 | dnode_t *dn; |
4218 | objset_t *os; | |
428870ff BB |
4219 | dmu_buf_impl_t *parent = db->db_parent; |
4220 | uint64_t txg = tx->tx_txg; | |
5dbd68a3 | 4221 | zbookmark_phys_t zb; |
428870ff BB |
4222 | zio_prop_t zp; |
4223 | zio_t *zio; | |
4224 | int wp_flag = 0; | |
34dc7c2f | 4225 | |
463a8cfe AR |
4226 | ASSERT(dmu_tx_is_syncing(tx)); |
4227 | ||
572e2857 BB |
4228 | DB_DNODE_ENTER(db); |
4229 | dn = DB_DNODE(db); | |
4230 | os = dn->dn_objset; | |
4231 | ||
428870ff BB |
4232 | if (db->db_state != DB_NOFILL) { |
4233 | if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) { | |
4234 | /* | |
4235 | * Private object buffers are released here rather | |
4236 | * than in dbuf_dirty() since they are only modified | |
4237 | * in the syncing context and we don't want the | |
4238 | * overhead of making multiple copies of the data. | |
4239 | */ | |
4240 | if (BP_IS_HOLE(db->db_blkptr)) { | |
4241 | arc_buf_thaw(data); | |
4242 | } else { | |
4243 | dbuf_release_bp(db); | |
4244 | } | |
a1d477c2 | 4245 | dbuf_remap(dn, db, tx); |
428870ff BB |
4246 | } |
4247 | } | |
4248 | ||
4249 | if (parent != dn->dn_dbuf) { | |
e49f1e20 WA |
4250 | /* Our parent is an indirect block. */ |
4251 | /* We have a dirty parent that has been scheduled for write. */ | |
428870ff | 4252 | ASSERT(parent && parent->db_data_pending); |
e49f1e20 | 4253 | /* Our parent's buffer is one level closer to the dnode. */ |
428870ff | 4254 | ASSERT(db->db_level == parent->db_level-1); |
e49f1e20 WA |
4255 | /* |
4256 | * We're about to modify our parent's db_data by modifying | |
4257 | * our block pointer, so the parent must be released. | |
4258 | */ | |
428870ff BB |
4259 | ASSERT(arc_released(parent->db_buf)); |
4260 | zio = parent->db_data_pending->dr_zio; | |
4261 | } else { | |
e49f1e20 | 4262 | /* Our parent is the dnode itself. */ |
428870ff BB |
4263 | ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 && |
4264 | db->db_blkid != DMU_SPILL_BLKID) || | |
4265 | (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0)); | |
4266 | if (db->db_blkid != DMU_SPILL_BLKID) | |
4267 | ASSERT3P(db->db_blkptr, ==, | |
4268 | &dn->dn_phys->dn_blkptr[db->db_blkid]); | |
4269 | zio = dn->dn_zio; | |
4270 | } | |
4271 | ||
4272 | ASSERT(db->db_level == 0 || data == db->db_buf); | |
4273 | ASSERT3U(db->db_blkptr->blk_birth, <=, txg); | |
4274 | ASSERT(zio); | |
4275 | ||
4276 | SET_BOOKMARK(&zb, os->os_dsl_dataset ? | |
4277 | os->os_dsl_dataset->ds_object : DMU_META_OBJSET, | |
4278 | db->db.db_object, db->db_level, db->db_blkid); | |
4279 | ||
4280 | if (db->db_blkid == DMU_SPILL_BLKID) | |
4281 | wp_flag = WP_SPILL; | |
4282 | wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0; | |
4283 | ||
82644107 | 4284 | dmu_write_policy(os, dn, db->db_level, wp_flag, &zp); |
572e2857 | 4285 | DB_DNODE_EXIT(db); |
428870ff | 4286 | |
463a8cfe AR |
4287 | /* |
4288 | * We copy the blkptr now (rather than when we instantiate the dirty | |
4289 | * record), because its value can change between open context and | |
4290 | * syncing context. We do not need to hold dn_struct_rwlock to read | |
4291 | * db_blkptr because we are in syncing context. | |
4292 | */ | |
4293 | dr->dr_bp_copy = *db->db_blkptr; | |
4294 | ||
9b67f605 MA |
4295 | if (db->db_level == 0 && |
4296 | dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { | |
4297 | /* | |
4298 | * The BP for this block has been provided by open context | |
4299 | * (by dmu_sync() or dmu_buf_write_embedded()). | |
4300 | */ | |
a6255b7f DQ |
4301 | abd_t *contents = (data != NULL) ? |
4302 | abd_get_from_buf(data->b_data, arc_buf_size(data)) : NULL; | |
9b67f605 | 4303 | |
428870ff | 4304 | dr->dr_zio = zio_write(zio, os->os_spa, txg, |
2aa34383 DK |
4305 | &dr->dr_bp_copy, contents, db->db.db_size, db->db.db_size, |
4306 | &zp, dbuf_write_override_ready, NULL, NULL, | |
bc77ba73 | 4307 | dbuf_write_override_done, |
e8b96c60 | 4308 | dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); |
428870ff BB |
4309 | mutex_enter(&db->db_mtx); |
4310 | dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; | |
4311 | zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by, | |
03c6040b | 4312 | dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite); |
428870ff BB |
4313 | mutex_exit(&db->db_mtx); |
4314 | } else if (db->db_state == DB_NOFILL) { | |
3c67d83a TH |
4315 | ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF || |
4316 | zp.zp_checksum == ZIO_CHECKSUM_NOPARITY); | |
428870ff | 4317 | dr->dr_zio = zio_write(zio, os->os_spa, txg, |
2aa34383 | 4318 | &dr->dr_bp_copy, NULL, db->db.db_size, db->db.db_size, &zp, |
bc77ba73 PD |
4319 | dbuf_write_nofill_ready, NULL, NULL, |
4320 | dbuf_write_nofill_done, db, | |
428870ff BB |
4321 | ZIO_PRIORITY_ASYNC_WRITE, |
4322 | ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb); | |
4323 | } else { | |
4324 | ASSERT(arc_released(data)); | |
bc77ba73 PD |
4325 | |
4326 | /* | |
4327 | * For indirect blocks, we want to setup the children | |
4328 | * ready callback so that we can properly handle an indirect | |
4329 | * block that only contains holes. | |
4330 | */ | |
1c27024e | 4331 | arc_write_done_func_t *children_ready_cb = NULL; |
bc77ba73 PD |
4332 | if (db->db_level != 0) |
4333 | children_ready_cb = dbuf_write_children_ready; | |
4334 | ||
428870ff | 4335 | dr->dr_zio = arc_write(zio, os->os_spa, txg, |
463a8cfe | 4336 | &dr->dr_bp_copy, data, DBUF_IS_L2CACHEABLE(db), |
d3c2ae1c GW |
4337 | &zp, dbuf_write_ready, |
4338 | children_ready_cb, dbuf_write_physdone, | |
4339 | dbuf_write_done, db, ZIO_PRIORITY_ASYNC_WRITE, | |
4340 | ZIO_FLAG_MUSTSUCCEED, &zb); | |
428870ff | 4341 | } |
34dc7c2f | 4342 | } |
c28b2279 | 4343 | |
93ce2b4c | 4344 | #if defined(_KERNEL) |
8f576c23 BB |
4345 | EXPORT_SYMBOL(dbuf_find); |
4346 | EXPORT_SYMBOL(dbuf_is_metadata); | |
d3c2ae1c | 4347 | EXPORT_SYMBOL(dbuf_destroy); |
8f576c23 BB |
4348 | EXPORT_SYMBOL(dbuf_loan_arcbuf); |
4349 | EXPORT_SYMBOL(dbuf_whichblock); | |
4350 | EXPORT_SYMBOL(dbuf_read); | |
4351 | EXPORT_SYMBOL(dbuf_unoverride); | |
4352 | EXPORT_SYMBOL(dbuf_free_range); | |
4353 | EXPORT_SYMBOL(dbuf_new_size); | |
4354 | EXPORT_SYMBOL(dbuf_release_bp); | |
4355 | EXPORT_SYMBOL(dbuf_dirty); | |
0c03d21a | 4356 | EXPORT_SYMBOL(dmu_buf_set_crypt_params); |
c28b2279 | 4357 | EXPORT_SYMBOL(dmu_buf_will_dirty); |
8f576c23 BB |
4358 | EXPORT_SYMBOL(dmu_buf_will_not_fill); |
4359 | EXPORT_SYMBOL(dmu_buf_will_fill); | |
4360 | EXPORT_SYMBOL(dmu_buf_fill_done); | |
4047414a | 4361 | EXPORT_SYMBOL(dmu_buf_rele); |
8f576c23 | 4362 | EXPORT_SYMBOL(dbuf_assign_arcbuf); |
8f576c23 BB |
4363 | EXPORT_SYMBOL(dbuf_prefetch); |
4364 | EXPORT_SYMBOL(dbuf_hold_impl); | |
4365 | EXPORT_SYMBOL(dbuf_hold); | |
4366 | EXPORT_SYMBOL(dbuf_hold_level); | |
4367 | EXPORT_SYMBOL(dbuf_create_bonus); | |
4368 | EXPORT_SYMBOL(dbuf_spill_set_blksz); | |
4369 | EXPORT_SYMBOL(dbuf_rm_spill); | |
4370 | EXPORT_SYMBOL(dbuf_add_ref); | |
4371 | EXPORT_SYMBOL(dbuf_rele); | |
4372 | EXPORT_SYMBOL(dbuf_rele_and_unlock); | |
4373 | EXPORT_SYMBOL(dbuf_refcount); | |
4374 | EXPORT_SYMBOL(dbuf_sync_list); | |
4375 | EXPORT_SYMBOL(dmu_buf_set_user); | |
4376 | EXPORT_SYMBOL(dmu_buf_set_user_ie); | |
8f576c23 | 4377 | EXPORT_SYMBOL(dmu_buf_get_user); |
0f699108 | 4378 | EXPORT_SYMBOL(dmu_buf_get_blkptr); |
d3c2ae1c | 4379 | |
02730c33 | 4380 | /* BEGIN CSTYLED */ |
d3c2ae1c GW |
4381 | module_param(dbuf_cache_max_bytes, ulong, 0644); |
4382 | MODULE_PARM_DESC(dbuf_cache_max_bytes, | |
02730c33 | 4383 | "Maximum size in bytes of the dbuf cache."); |
d3c2ae1c GW |
4384 | |
4385 | module_param(dbuf_cache_hiwater_pct, uint, 0644); | |
4386 | MODULE_PARM_DESC(dbuf_cache_hiwater_pct, | |
f974e25d | 4387 | "Percentage over dbuf_cache_max_bytes when dbufs must be evicted " |
4388 | "directly."); | |
d3c2ae1c GW |
4389 | |
4390 | module_param(dbuf_cache_lowater_pct, uint, 0644); | |
4391 | MODULE_PARM_DESC(dbuf_cache_lowater_pct, | |
f974e25d | 4392 | "Percentage below dbuf_cache_max_bytes when the evict thread stops " |
4393 | "evicting dbufs."); | |
d3c2ae1c | 4394 | |
de4f8d5d BB |
4395 | module_param(dbuf_cache_shift, int, 0644); |
4396 | MODULE_PARM_DESC(dbuf_cache_shift, | |
4397 | "Set the size of the dbuf cache to a log2 fraction of arc size."); | |
02730c33 | 4398 | /* END CSTYLED */ |
c28b2279 | 4399 | #endif |