]>
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. |
f664f1ee | 24 | * Copyright (c) 2012, 2019 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> |
e5d1c27e | 47 | #include <sys/trace_zfs.h> |
d3c2ae1c | 48 | #include <sys/callb.h> |
a6255b7f | 49 | #include <sys/abd.h> |
a1d477c2 | 50 | #include <sys/vdev.h> |
37fb3e43 | 51 | #include <sys/cityhash.h> |
2e5dc449 | 52 | #include <sys/spa_impl.h> |
34dc7c2f | 53 | |
5e021f56 GDN |
54 | kstat_t *dbuf_ksp; |
55 | ||
56 | typedef struct dbuf_stats { | |
57 | /* | |
58 | * Various statistics about the size of the dbuf cache. | |
59 | */ | |
60 | kstat_named_t cache_count; | |
61 | kstat_named_t cache_size_bytes; | |
62 | kstat_named_t cache_size_bytes_max; | |
63 | /* | |
64 | * Statistics regarding the bounds on the dbuf cache size. | |
65 | */ | |
66 | kstat_named_t cache_target_bytes; | |
67 | kstat_named_t cache_lowater_bytes; | |
68 | kstat_named_t cache_hiwater_bytes; | |
69 | /* | |
70 | * Total number of dbuf cache evictions that have occurred. | |
71 | */ | |
72 | kstat_named_t cache_total_evicts; | |
73 | /* | |
74 | * The distribution of dbuf levels in the dbuf cache and | |
75 | * the total size of all dbufs at each level. | |
76 | */ | |
77 | kstat_named_t cache_levels[DN_MAX_LEVELS]; | |
78 | kstat_named_t cache_levels_bytes[DN_MAX_LEVELS]; | |
79 | /* | |
80 | * Statistics about the dbuf hash table. | |
81 | */ | |
82 | kstat_named_t hash_hits; | |
83 | kstat_named_t hash_misses; | |
84 | kstat_named_t hash_collisions; | |
85 | kstat_named_t hash_elements; | |
86 | kstat_named_t hash_elements_max; | |
87 | /* | |
88 | * Number of sublists containing more than one dbuf in the dbuf | |
89 | * hash table. Keep track of the longest hash chain. | |
90 | */ | |
91 | kstat_named_t hash_chains; | |
92 | kstat_named_t hash_chain_max; | |
93 | /* | |
94 | * Number of times a dbuf_create() discovers that a dbuf was | |
95 | * already created and in the dbuf hash table. | |
96 | */ | |
97 | kstat_named_t hash_insert_race; | |
2e5dc449 MA |
98 | /* |
99 | * Statistics about the size of the metadata dbuf cache. | |
100 | */ | |
101 | kstat_named_t metadata_cache_count; | |
102 | kstat_named_t metadata_cache_size_bytes; | |
103 | kstat_named_t metadata_cache_size_bytes_max; | |
104 | /* | |
105 | * For diagnostic purposes, this is incremented whenever we can't add | |
106 | * something to the metadata cache because it's full, and instead put | |
107 | * the data in the regular dbuf cache. | |
108 | */ | |
109 | kstat_named_t metadata_cache_overflow; | |
5e021f56 GDN |
110 | } dbuf_stats_t; |
111 | ||
112 | dbuf_stats_t dbuf_stats = { | |
113 | { "cache_count", KSTAT_DATA_UINT64 }, | |
114 | { "cache_size_bytes", KSTAT_DATA_UINT64 }, | |
115 | { "cache_size_bytes_max", KSTAT_DATA_UINT64 }, | |
116 | { "cache_target_bytes", KSTAT_DATA_UINT64 }, | |
117 | { "cache_lowater_bytes", KSTAT_DATA_UINT64 }, | |
118 | { "cache_hiwater_bytes", KSTAT_DATA_UINT64 }, | |
119 | { "cache_total_evicts", KSTAT_DATA_UINT64 }, | |
120 | { { "cache_levels_N", KSTAT_DATA_UINT64 } }, | |
121 | { { "cache_levels_bytes_N", KSTAT_DATA_UINT64 } }, | |
122 | { "hash_hits", KSTAT_DATA_UINT64 }, | |
123 | { "hash_misses", KSTAT_DATA_UINT64 }, | |
124 | { "hash_collisions", KSTAT_DATA_UINT64 }, | |
125 | { "hash_elements", KSTAT_DATA_UINT64 }, | |
126 | { "hash_elements_max", KSTAT_DATA_UINT64 }, | |
127 | { "hash_chains", KSTAT_DATA_UINT64 }, | |
128 | { "hash_chain_max", KSTAT_DATA_UINT64 }, | |
2e5dc449 MA |
129 | { "hash_insert_race", KSTAT_DATA_UINT64 }, |
130 | { "metadata_cache_count", KSTAT_DATA_UINT64 }, | |
131 | { "metadata_cache_size_bytes", KSTAT_DATA_UINT64 }, | |
132 | { "metadata_cache_size_bytes_max", KSTAT_DATA_UINT64 }, | |
133 | { "metadata_cache_overflow", KSTAT_DATA_UINT64 } | |
5e021f56 GDN |
134 | }; |
135 | ||
136 | #define DBUF_STAT_INCR(stat, val) \ | |
137 | atomic_add_64(&dbuf_stats.stat.value.ui64, (val)); | |
138 | #define DBUF_STAT_DECR(stat, val) \ | |
139 | DBUF_STAT_INCR(stat, -(val)); | |
140 | #define DBUF_STAT_BUMP(stat) \ | |
141 | DBUF_STAT_INCR(stat, 1); | |
142 | #define DBUF_STAT_BUMPDOWN(stat) \ | |
143 | DBUF_STAT_INCR(stat, -1); | |
144 | #define DBUF_STAT_MAX(stat, v) { \ | |
145 | uint64_t _m; \ | |
146 | while ((v) > (_m = dbuf_stats.stat.value.ui64) && \ | |
147 | (_m != atomic_cas_64(&dbuf_stats.stat.value.ui64, _m, (v))))\ | |
148 | continue; \ | |
149 | } | |
150 | ||
13fe0198 | 151 | static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx); |
b128c09f | 152 | static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx); |
34dc7c2f | 153 | |
0c66c32d | 154 | extern inline void dmu_buf_init_user(dmu_buf_user_t *dbu, |
39efbde7 GM |
155 | dmu_buf_evict_func_t *evict_func_sync, |
156 | dmu_buf_evict_func_t *evict_func_async, | |
157 | dmu_buf_t **clear_on_evict_dbufp); | |
0c66c32d | 158 | |
34dc7c2f BB |
159 | /* |
160 | * Global data structures and functions for the dbuf cache. | |
161 | */ | |
d3c2ae1c | 162 | static kmem_cache_t *dbuf_kmem_cache; |
0c66c32d | 163 | static taskq_t *dbu_evict_taskq; |
34dc7c2f | 164 | |
d3c2ae1c GW |
165 | static kthread_t *dbuf_cache_evict_thread; |
166 | static kmutex_t dbuf_evict_lock; | |
167 | static kcondvar_t dbuf_evict_cv; | |
168 | static boolean_t dbuf_evict_thread_exit; | |
169 | ||
170 | /* | |
2e5dc449 MA |
171 | * There are two dbuf caches; each dbuf can only be in one of them at a time. |
172 | * | |
173 | * 1. Cache of metadata dbufs, to help make read-heavy administrative commands | |
174 | * from /sbin/zfs run faster. The "metadata cache" specifically stores dbufs | |
175 | * that represent the metadata that describes filesystems/snapshots/ | |
176 | * bookmarks/properties/etc. We only evict from this cache when we export a | |
177 | * pool, to short-circuit as much I/O as possible for all administrative | |
178 | * commands that need the metadata. There is no eviction policy for this | |
179 | * cache, because we try to only include types in it which would occupy a | |
180 | * very small amount of space per object but create a large impact on the | |
181 | * performance of these commands. Instead, after it reaches a maximum size | |
182 | * (which should only happen on very small memory systems with a very large | |
183 | * number of filesystem objects), we stop taking new dbufs into the | |
184 | * metadata cache, instead putting them in the normal dbuf cache. | |
185 | * | |
186 | * 2. LRU cache of dbufs. The dbuf cache maintains a list of dbufs that | |
187 | * are not currently held but have been recently released. These dbufs | |
188 | * are not eligible for arc eviction until they are aged out of the cache. | |
189 | * Dbufs that are aged out of the cache will be immediately destroyed and | |
190 | * become eligible for arc eviction. | |
191 | * | |
192 | * Dbufs are added to these caches once the last hold is released. If a dbuf is | |
193 | * later accessed and still exists in the dbuf cache, then it will be removed | |
194 | * from the cache and later re-added to the head of the cache. | |
195 | * | |
196 | * If a given dbuf meets the requirements for the metadata cache, it will go | |
197 | * there, otherwise it will be considered for the generic LRU dbuf cache. The | |
198 | * caches and the refcounts tracking their sizes are stored in an array indexed | |
199 | * by those caches' matching enum values (from dbuf_cached_state_t). | |
d3c2ae1c | 200 | */ |
2e5dc449 MA |
201 | typedef struct dbuf_cache { |
202 | multilist_t *cache; | |
c13060e4 | 203 | zfs_refcount_t size; |
2e5dc449 MA |
204 | } dbuf_cache_t; |
205 | dbuf_cache_t dbuf_caches[DB_CACHE_MAX]; | |
d3c2ae1c | 206 | |
2e5dc449 MA |
207 | /* Size limits for the caches */ |
208 | unsigned long dbuf_cache_max_bytes = 0; | |
209 | unsigned long dbuf_metadata_cache_max_bytes = 0; | |
210 | /* Set the default sizes of the caches to log2 fraction of arc size */ | |
de4f8d5d | 211 | int dbuf_cache_shift = 5; |
2e5dc449 | 212 | int dbuf_metadata_cache_shift = 6; |
d3c2ae1c GW |
213 | |
214 | /* | |
2e5dc449 | 215 | * The LRU dbuf cache uses a three-stage eviction policy: |
d3c2ae1c GW |
216 | * - A low water marker designates when the dbuf eviction thread |
217 | * should stop evicting from the dbuf cache. | |
218 | * - When we reach the maximum size (aka mid water mark), we | |
219 | * signal the eviction thread to run. | |
220 | * - The high water mark indicates when the eviction thread | |
221 | * is unable to keep up with the incoming load and eviction must | |
222 | * happen in the context of the calling thread. | |
223 | * | |
224 | * The dbuf cache: | |
225 | * (max size) | |
226 | * low water mid water hi water | |
227 | * +----------------------------------------+----------+----------+ | |
228 | * | | | | | |
229 | * | | | | | |
230 | * | | | | | |
231 | * | | | | | |
232 | * +----------------------------------------+----------+----------+ | |
233 | * stop signal evict | |
234 | * evicting eviction directly | |
235 | * thread | |
236 | * | |
237 | * The high and low water marks indicate the operating range for the eviction | |
238 | * thread. The low water mark is, by default, 90% of the total size of the | |
239 | * cache and the high water mark is at 110% (both of these percentages can be | |
240 | * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct, | |
241 | * respectively). The eviction thread will try to ensure that the cache remains | |
242 | * within this range by waking up every second and checking if the cache is | |
243 | * above the low water mark. The thread can also be woken up by callers adding | |
244 | * elements into the cache if the cache is larger than the mid water (i.e max | |
245 | * cache size). Once the eviction thread is woken up and eviction is required, | |
246 | * it will continue evicting buffers until it's able to reduce the cache size | |
247 | * to the low water mark. If the cache size continues to grow and hits the high | |
4e33ba4c | 248 | * water mark, then callers adding elements to the cache will begin to evict |
d3c2ae1c GW |
249 | * directly from the cache until the cache is no longer above the high water |
250 | * mark. | |
251 | */ | |
252 | ||
253 | /* | |
254 | * The percentage above and below the maximum cache size. | |
255 | */ | |
256 | uint_t dbuf_cache_hiwater_pct = 10; | |
257 | uint_t dbuf_cache_lowater_pct = 10; | |
258 | ||
34dc7c2f BB |
259 | /* ARGSUSED */ |
260 | static int | |
261 | dbuf_cons(void *vdb, void *unused, int kmflag) | |
262 | { | |
263 | dmu_buf_impl_t *db = vdb; | |
264 | bzero(db, sizeof (dmu_buf_impl_t)); | |
265 | ||
266 | mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL); | |
f664f1ee | 267 | rw_init(&db->db_rwlock, NULL, RW_DEFAULT, NULL); |
34dc7c2f | 268 | cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL); |
d3c2ae1c | 269 | multilist_link_init(&db->db_cache_link); |
424fd7c3 | 270 | zfs_refcount_create(&db->db_holds); |
8951cb8d | 271 | |
34dc7c2f BB |
272 | return (0); |
273 | } | |
274 | ||
275 | /* ARGSUSED */ | |
276 | static void | |
277 | dbuf_dest(void *vdb, void *unused) | |
278 | { | |
279 | dmu_buf_impl_t *db = vdb; | |
280 | mutex_destroy(&db->db_mtx); | |
f664f1ee | 281 | rw_destroy(&db->db_rwlock); |
34dc7c2f | 282 | cv_destroy(&db->db_changed); |
d3c2ae1c | 283 | ASSERT(!multilist_link_active(&db->db_cache_link)); |
424fd7c3 | 284 | zfs_refcount_destroy(&db->db_holds); |
34dc7c2f BB |
285 | } |
286 | ||
287 | /* | |
288 | * dbuf hash table routines | |
289 | */ | |
290 | static dbuf_hash_table_t dbuf_hash_table; | |
291 | ||
292 | static uint64_t dbuf_hash_count; | |
293 | ||
37fb3e43 PD |
294 | /* |
295 | * We use Cityhash for this. It's fast, and has good hash properties without | |
296 | * requiring any large static buffers. | |
297 | */ | |
34dc7c2f BB |
298 | static uint64_t |
299 | dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid) | |
300 | { | |
37fb3e43 | 301 | return (cityhash4((uintptr_t)os, obj, (uint64_t)lvl, blkid)); |
34dc7c2f BB |
302 | } |
303 | ||
34dc7c2f BB |
304 | #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \ |
305 | ((dbuf)->db.db_object == (obj) && \ | |
306 | (dbuf)->db_objset == (os) && \ | |
307 | (dbuf)->db_level == (level) && \ | |
308 | (dbuf)->db_blkid == (blkid)) | |
309 | ||
310 | dmu_buf_impl_t * | |
6ebebace | 311 | dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid) |
34dc7c2f BB |
312 | { |
313 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
d6320ddb BB |
314 | uint64_t hv; |
315 | uint64_t idx; | |
34dc7c2f BB |
316 | dmu_buf_impl_t *db; |
317 | ||
d3c2ae1c | 318 | hv = dbuf_hash(os, obj, level, blkid); |
d6320ddb BB |
319 | idx = hv & h->hash_table_mask; |
320 | ||
34dc7c2f BB |
321 | mutex_enter(DBUF_HASH_MUTEX(h, idx)); |
322 | for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) { | |
323 | if (DBUF_EQUAL(db, os, obj, level, blkid)) { | |
324 | mutex_enter(&db->db_mtx); | |
325 | if (db->db_state != DB_EVICTING) { | |
326 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); | |
327 | return (db); | |
328 | } | |
329 | mutex_exit(&db->db_mtx); | |
330 | } | |
331 | } | |
332 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); | |
333 | return (NULL); | |
334 | } | |
335 | ||
6ebebace JG |
336 | static dmu_buf_impl_t * |
337 | dbuf_find_bonus(objset_t *os, uint64_t object) | |
338 | { | |
339 | dnode_t *dn; | |
340 | dmu_buf_impl_t *db = NULL; | |
341 | ||
342 | if (dnode_hold(os, object, FTAG, &dn) == 0) { | |
343 | rw_enter(&dn->dn_struct_rwlock, RW_READER); | |
344 | if (dn->dn_bonus != NULL) { | |
345 | db = dn->dn_bonus; | |
346 | mutex_enter(&db->db_mtx); | |
347 | } | |
348 | rw_exit(&dn->dn_struct_rwlock); | |
349 | dnode_rele(dn, FTAG); | |
350 | } | |
351 | return (db); | |
352 | } | |
353 | ||
34dc7c2f BB |
354 | /* |
355 | * Insert an entry into the hash table. If there is already an element | |
356 | * equal to elem in the hash table, then the already existing element | |
357 | * will be returned and the new element will not be inserted. | |
358 | * Otherwise returns NULL. | |
359 | */ | |
360 | static dmu_buf_impl_t * | |
361 | dbuf_hash_insert(dmu_buf_impl_t *db) | |
362 | { | |
363 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
428870ff | 364 | objset_t *os = db->db_objset; |
34dc7c2f BB |
365 | uint64_t obj = db->db.db_object; |
366 | int level = db->db_level; | |
d6320ddb | 367 | uint64_t blkid, hv, idx; |
34dc7c2f | 368 | dmu_buf_impl_t *dbf; |
5e021f56 | 369 | uint32_t i; |
34dc7c2f | 370 | |
d6320ddb | 371 | blkid = db->db_blkid; |
d3c2ae1c | 372 | hv = dbuf_hash(os, obj, level, blkid); |
d6320ddb BB |
373 | idx = hv & h->hash_table_mask; |
374 | ||
34dc7c2f | 375 | mutex_enter(DBUF_HASH_MUTEX(h, idx)); |
5e021f56 GDN |
376 | for (dbf = h->hash_table[idx], i = 0; dbf != NULL; |
377 | dbf = dbf->db_hash_next, i++) { | |
34dc7c2f BB |
378 | if (DBUF_EQUAL(dbf, os, obj, level, blkid)) { |
379 | mutex_enter(&dbf->db_mtx); | |
380 | if (dbf->db_state != DB_EVICTING) { | |
381 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); | |
382 | return (dbf); | |
383 | } | |
384 | mutex_exit(&dbf->db_mtx); | |
385 | } | |
386 | } | |
387 | ||
5e021f56 GDN |
388 | if (i > 0) { |
389 | DBUF_STAT_BUMP(hash_collisions); | |
390 | if (i == 1) | |
391 | DBUF_STAT_BUMP(hash_chains); | |
392 | ||
393 | DBUF_STAT_MAX(hash_chain_max, i); | |
394 | } | |
395 | ||
34dc7c2f BB |
396 | mutex_enter(&db->db_mtx); |
397 | db->db_hash_next = h->hash_table[idx]; | |
398 | h->hash_table[idx] = db; | |
399 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); | |
bc89ac84 | 400 | atomic_inc_64(&dbuf_hash_count); |
5e021f56 | 401 | DBUF_STAT_MAX(hash_elements_max, dbuf_hash_count); |
34dc7c2f BB |
402 | |
403 | return (NULL); | |
404 | } | |
405 | ||
2e5dc449 MA |
406 | /* |
407 | * This returns whether this dbuf should be stored in the metadata cache, which | |
408 | * is based on whether it's from one of the dnode types that store data related | |
409 | * to traversing dataset hierarchies. | |
410 | */ | |
411 | static boolean_t | |
412 | dbuf_include_in_metadata_cache(dmu_buf_impl_t *db) | |
413 | { | |
414 | DB_DNODE_ENTER(db); | |
415 | dmu_object_type_t type = DB_DNODE(db)->dn_type; | |
416 | DB_DNODE_EXIT(db); | |
417 | ||
418 | /* Check if this dbuf is one of the types we care about */ | |
419 | if (DMU_OT_IS_METADATA_CACHED(type)) { | |
420 | /* If we hit this, then we set something up wrong in dmu_ot */ | |
421 | ASSERT(DMU_OT_IS_METADATA(type)); | |
422 | ||
423 | /* | |
424 | * Sanity check for small-memory systems: don't allocate too | |
425 | * much memory for this purpose. | |
426 | */ | |
424fd7c3 TS |
427 | if (zfs_refcount_count( |
428 | &dbuf_caches[DB_DBUF_METADATA_CACHE].size) > | |
2e5dc449 MA |
429 | dbuf_metadata_cache_max_bytes) { |
430 | DBUF_STAT_BUMP(metadata_cache_overflow); | |
431 | return (B_FALSE); | |
432 | } | |
433 | ||
434 | return (B_TRUE); | |
435 | } | |
436 | ||
437 | return (B_FALSE); | |
438 | } | |
439 | ||
34dc7c2f | 440 | /* |
bd089c54 | 441 | * Remove an entry from the hash table. It must be in the EVICTING state. |
34dc7c2f BB |
442 | */ |
443 | static void | |
444 | dbuf_hash_remove(dmu_buf_impl_t *db) | |
445 | { | |
446 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
d6320ddb | 447 | uint64_t hv, idx; |
34dc7c2f BB |
448 | dmu_buf_impl_t *dbf, **dbp; |
449 | ||
d3c2ae1c | 450 | hv = dbuf_hash(db->db_objset, db->db.db_object, |
d6320ddb BB |
451 | db->db_level, db->db_blkid); |
452 | idx = hv & h->hash_table_mask; | |
453 | ||
34dc7c2f | 454 | /* |
4e33ba4c | 455 | * We mustn't hold db_mtx to maintain lock ordering: |
34dc7c2f BB |
456 | * DBUF_HASH_MUTEX > db_mtx. |
457 | */ | |
424fd7c3 | 458 | ASSERT(zfs_refcount_is_zero(&db->db_holds)); |
34dc7c2f BB |
459 | ASSERT(db->db_state == DB_EVICTING); |
460 | ASSERT(!MUTEX_HELD(&db->db_mtx)); | |
461 | ||
462 | mutex_enter(DBUF_HASH_MUTEX(h, idx)); | |
463 | dbp = &h->hash_table[idx]; | |
464 | while ((dbf = *dbp) != db) { | |
465 | dbp = &dbf->db_hash_next; | |
466 | ASSERT(dbf != NULL); | |
467 | } | |
468 | *dbp = db->db_hash_next; | |
469 | db->db_hash_next = NULL; | |
5e021f56 GDN |
470 | if (h->hash_table[idx] && |
471 | h->hash_table[idx]->db_hash_next == NULL) | |
472 | DBUF_STAT_BUMPDOWN(hash_chains); | |
34dc7c2f | 473 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); |
bc89ac84 | 474 | atomic_dec_64(&dbuf_hash_count); |
34dc7c2f BB |
475 | } |
476 | ||
0c66c32d JG |
477 | typedef enum { |
478 | DBVU_EVICTING, | |
479 | DBVU_NOT_EVICTING | |
480 | } dbvu_verify_type_t; | |
481 | ||
482 | static void | |
483 | dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type) | |
484 | { | |
485 | #ifdef ZFS_DEBUG | |
486 | int64_t holds; | |
487 | ||
488 | if (db->db_user == NULL) | |
489 | return; | |
490 | ||
491 | /* Only data blocks support the attachment of user data. */ | |
492 | ASSERT(db->db_level == 0); | |
493 | ||
494 | /* Clients must resolve a dbuf before attaching user data. */ | |
495 | ASSERT(db->db.db_data != NULL); | |
496 | ASSERT3U(db->db_state, ==, DB_CACHED); | |
497 | ||
424fd7c3 | 498 | holds = zfs_refcount_count(&db->db_holds); |
0c66c32d JG |
499 | if (verify_type == DBVU_EVICTING) { |
500 | /* | |
501 | * Immediate eviction occurs when holds == dirtycnt. | |
502 | * For normal eviction buffers, holds is zero on | |
503 | * eviction, except when dbuf_fix_old_data() calls | |
504 | * dbuf_clear_data(). However, the hold count can grow | |
505 | * during eviction even though db_mtx is held (see | |
506 | * dmu_bonus_hold() for an example), so we can only | |
507 | * test the generic invariant that holds >= dirtycnt. | |
508 | */ | |
509 | ASSERT3U(holds, >=, db->db_dirtycnt); | |
510 | } else { | |
bc4501f7 | 511 | if (db->db_user_immediate_evict == TRUE) |
0c66c32d JG |
512 | ASSERT3U(holds, >=, db->db_dirtycnt); |
513 | else | |
514 | ASSERT3U(holds, >, 0); | |
515 | } | |
516 | #endif | |
517 | } | |
518 | ||
34dc7c2f BB |
519 | static void |
520 | dbuf_evict_user(dmu_buf_impl_t *db) | |
521 | { | |
0c66c32d JG |
522 | dmu_buf_user_t *dbu = db->db_user; |
523 | ||
34dc7c2f BB |
524 | ASSERT(MUTEX_HELD(&db->db_mtx)); |
525 | ||
0c66c32d | 526 | if (dbu == NULL) |
34dc7c2f BB |
527 | return; |
528 | ||
0c66c32d JG |
529 | dbuf_verify_user(db, DBVU_EVICTING); |
530 | db->db_user = NULL; | |
531 | ||
532 | #ifdef ZFS_DEBUG | |
533 | if (dbu->dbu_clear_on_evict_dbufp != NULL) | |
534 | *dbu->dbu_clear_on_evict_dbufp = NULL; | |
535 | #endif | |
536 | ||
537 | /* | |
39efbde7 GM |
538 | * There are two eviction callbacks - one that we call synchronously |
539 | * and one that we invoke via a taskq. The async one is useful for | |
540 | * avoiding lock order reversals and limiting stack depth. | |
541 | * | |
542 | * Note that if we have a sync callback but no async callback, | |
543 | * it's likely that the sync callback will free the structure | |
544 | * containing the dbu. In that case we need to take care to not | |
545 | * dereference dbu after calling the sync evict func. | |
0c66c32d | 546 | */ |
a7004725 | 547 | boolean_t has_async = (dbu->dbu_evict_func_async != NULL); |
39efbde7 GM |
548 | |
549 | if (dbu->dbu_evict_func_sync != NULL) | |
550 | dbu->dbu_evict_func_sync(dbu); | |
551 | ||
552 | if (has_async) { | |
553 | taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func_async, | |
554 | dbu, 0, &dbu->dbu_tqent); | |
555 | } | |
34dc7c2f BB |
556 | } |
557 | ||
572e2857 BB |
558 | boolean_t |
559 | dbuf_is_metadata(dmu_buf_impl_t *db) | |
560 | { | |
cc79a5c2 BB |
561 | /* |
562 | * Consider indirect blocks and spill blocks to be meta data. | |
563 | */ | |
564 | if (db->db_level > 0 || db->db_blkid == DMU_SPILL_BLKID) { | |
572e2857 BB |
565 | return (B_TRUE); |
566 | } else { | |
567 | boolean_t is_metadata; | |
568 | ||
569 | DB_DNODE_ENTER(db); | |
9ae529ec | 570 | is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type); |
572e2857 BB |
571 | DB_DNODE_EXIT(db); |
572 | ||
573 | return (is_metadata); | |
574 | } | |
575 | } | |
576 | ||
d3c2ae1c GW |
577 | |
578 | /* | |
579 | * This function *must* return indices evenly distributed between all | |
580 | * sublists of the multilist. This is needed due to how the dbuf eviction | |
581 | * code is laid out; dbuf_evict_thread() assumes dbufs are evenly | |
582 | * distributed between all sublists and uses this assumption when | |
583 | * deciding which sublist to evict from and how much to evict from it. | |
584 | */ | |
585 | unsigned int | |
586 | dbuf_cache_multilist_index_func(multilist_t *ml, void *obj) | |
34dc7c2f | 587 | { |
d3c2ae1c GW |
588 | dmu_buf_impl_t *db = obj; |
589 | ||
590 | /* | |
591 | * The assumption here, is the hash value for a given | |
592 | * dmu_buf_impl_t will remain constant throughout it's lifetime | |
593 | * (i.e. it's objset, object, level and blkid fields don't change). | |
594 | * Thus, we don't need to store the dbuf's sublist index | |
595 | * on insertion, as this index can be recalculated on removal. | |
596 | * | |
597 | * Also, the low order bits of the hash value are thought to be | |
598 | * distributed evenly. Otherwise, in the case that the multilist | |
599 | * has a power of two number of sublists, each sublists' usage | |
600 | * would not be evenly distributed. | |
601 | */ | |
602 | return (dbuf_hash(db->db_objset, db->db.db_object, | |
603 | db->db_level, db->db_blkid) % | |
604 | multilist_get_num_sublists(ml)); | |
605 | } | |
606 | ||
e71cade6 | 607 | static inline unsigned long |
608 | dbuf_cache_target_bytes(void) | |
609 | { | |
610 | return MIN(dbuf_cache_max_bytes, | |
de4f8d5d | 611 | arc_target_bytes() >> dbuf_cache_shift); |
e71cade6 | 612 | } |
613 | ||
5e021f56 GDN |
614 | static inline uint64_t |
615 | dbuf_cache_hiwater_bytes(void) | |
d3c2ae1c | 616 | { |
e71cade6 | 617 | uint64_t dbuf_cache_target = dbuf_cache_target_bytes(); |
5e021f56 GDN |
618 | return (dbuf_cache_target + |
619 | (dbuf_cache_target * dbuf_cache_hiwater_pct) / 100); | |
620 | } | |
e71cade6 | 621 | |
5e021f56 GDN |
622 | static inline uint64_t |
623 | dbuf_cache_lowater_bytes(void) | |
624 | { | |
625 | uint64_t dbuf_cache_target = dbuf_cache_target_bytes(); | |
626 | return (dbuf_cache_target - | |
627 | (dbuf_cache_target * dbuf_cache_lowater_pct) / 100); | |
628 | } | |
d3c2ae1c | 629 | |
d3c2ae1c GW |
630 | static inline boolean_t |
631 | dbuf_cache_above_lowater(void) | |
632 | { | |
424fd7c3 | 633 | return (zfs_refcount_count(&dbuf_caches[DB_DBUF_CACHE].size) > |
2e5dc449 | 634 | dbuf_cache_lowater_bytes()); |
d3c2ae1c GW |
635 | } |
636 | ||
637 | /* | |
638 | * Evict the oldest eligible dbuf from the dbuf cache. | |
639 | */ | |
640 | static void | |
641 | dbuf_evict_one(void) | |
642 | { | |
2e5dc449 MA |
643 | int idx = multilist_get_random_index(dbuf_caches[DB_DBUF_CACHE].cache); |
644 | multilist_sublist_t *mls = multilist_sublist_lock( | |
645 | dbuf_caches[DB_DBUF_CACHE].cache, idx); | |
1c27024e | 646 | |
d3c2ae1c GW |
647 | ASSERT(!MUTEX_HELD(&dbuf_evict_lock)); |
648 | ||
1c27024e | 649 | dmu_buf_impl_t *db = multilist_sublist_tail(mls); |
d3c2ae1c GW |
650 | while (db != NULL && mutex_tryenter(&db->db_mtx) == 0) { |
651 | db = multilist_sublist_prev(mls, db); | |
652 | } | |
653 | ||
654 | DTRACE_PROBE2(dbuf__evict__one, dmu_buf_impl_t *, db, | |
655 | multilist_sublist_t *, mls); | |
656 | ||
657 | if (db != NULL) { | |
658 | multilist_sublist_remove(mls, db); | |
659 | multilist_sublist_unlock(mls); | |
424fd7c3 TS |
660 | (void) zfs_refcount_remove_many( |
661 | &dbuf_caches[DB_DBUF_CACHE].size, db->db.db_size, db); | |
5e021f56 GDN |
662 | DBUF_STAT_BUMPDOWN(cache_levels[db->db_level]); |
663 | DBUF_STAT_BUMPDOWN(cache_count); | |
664 | DBUF_STAT_DECR(cache_levels_bytes[db->db_level], | |
665 | db->db.db_size); | |
2e5dc449 MA |
666 | ASSERT3U(db->db_caching_status, ==, DB_DBUF_CACHE); |
667 | db->db_caching_status = DB_NO_CACHE; | |
d3c2ae1c | 668 | dbuf_destroy(db); |
5e021f56 | 669 | DBUF_STAT_BUMP(cache_total_evicts); |
d3c2ae1c GW |
670 | } else { |
671 | multilist_sublist_unlock(mls); | |
672 | } | |
d3c2ae1c GW |
673 | } |
674 | ||
675 | /* | |
676 | * The dbuf evict thread is responsible for aging out dbufs from the | |
677 | * cache. Once the cache has reached it's maximum size, dbufs are removed | |
678 | * and destroyed. The eviction thread will continue running until the size | |
679 | * of the dbuf cache is at or below the maximum size. Once the dbuf is aged | |
680 | * out of the cache it is destroyed and becomes eligible for arc eviction. | |
681 | */ | |
867959b5 | 682 | /* ARGSUSED */ |
d3c2ae1c | 683 | static void |
c25b8f99 | 684 | dbuf_evict_thread(void *unused) |
d3c2ae1c GW |
685 | { |
686 | callb_cpr_t cpr; | |
687 | ||
688 | CALLB_CPR_INIT(&cpr, &dbuf_evict_lock, callb_generic_cpr, FTAG); | |
689 | ||
690 | mutex_enter(&dbuf_evict_lock); | |
691 | while (!dbuf_evict_thread_exit) { | |
692 | while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) { | |
693 | CALLB_CPR_SAFE_BEGIN(&cpr); | |
694 | (void) cv_timedwait_sig_hires(&dbuf_evict_cv, | |
695 | &dbuf_evict_lock, SEC2NSEC(1), MSEC2NSEC(1), 0); | |
696 | CALLB_CPR_SAFE_END(&cpr, &dbuf_evict_lock); | |
697 | } | |
698 | mutex_exit(&dbuf_evict_lock); | |
699 | ||
700 | /* | |
701 | * Keep evicting as long as we're above the low water mark | |
702 | * for the cache. We do this without holding the locks to | |
703 | * minimize lock contention. | |
704 | */ | |
705 | while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) { | |
706 | dbuf_evict_one(); | |
707 | } | |
708 | ||
709 | mutex_enter(&dbuf_evict_lock); | |
710 | } | |
711 | ||
712 | dbuf_evict_thread_exit = B_FALSE; | |
713 | cv_broadcast(&dbuf_evict_cv); | |
714 | CALLB_CPR_EXIT(&cpr); /* drops dbuf_evict_lock */ | |
715 | thread_exit(); | |
716 | } | |
717 | ||
718 | /* | |
719 | * Wake up the dbuf eviction thread if the dbuf cache is at its max size. | |
720 | * If the dbuf cache is at its high water mark, then evict a dbuf from the | |
721 | * dbuf cache using the callers context. | |
722 | */ | |
723 | static void | |
cbd8f5b7 | 724 | dbuf_evict_notify(uint64_t size) |
d3c2ae1c | 725 | { |
38240ebd MA |
726 | /* |
727 | * We check if we should evict without holding the dbuf_evict_lock, | |
728 | * because it's OK to occasionally make the wrong decision here, | |
729 | * and grabbing the lock results in massive lock contention. | |
730 | */ | |
cbd8f5b7 AM |
731 | if (size > dbuf_cache_target_bytes()) { |
732 | if (size > dbuf_cache_hiwater_bytes()) | |
d3c2ae1c | 733 | dbuf_evict_one(); |
38240ebd | 734 | cv_signal(&dbuf_evict_cv); |
d3c2ae1c | 735 | } |
34dc7c2f BB |
736 | } |
737 | ||
5e021f56 GDN |
738 | static int |
739 | dbuf_kstat_update(kstat_t *ksp, int rw) | |
740 | { | |
741 | dbuf_stats_t *ds = ksp->ks_data; | |
d3c2ae1c | 742 | |
5e021f56 GDN |
743 | if (rw == KSTAT_WRITE) { |
744 | return (SET_ERROR(EACCES)); | |
745 | } else { | |
424fd7c3 TS |
746 | ds->metadata_cache_size_bytes.value.ui64 = zfs_refcount_count( |
747 | &dbuf_caches[DB_DBUF_METADATA_CACHE].size); | |
5e021f56 | 748 | ds->cache_size_bytes.value.ui64 = |
424fd7c3 | 749 | zfs_refcount_count(&dbuf_caches[DB_DBUF_CACHE].size); |
5e021f56 GDN |
750 | ds->cache_target_bytes.value.ui64 = dbuf_cache_target_bytes(); |
751 | ds->cache_hiwater_bytes.value.ui64 = dbuf_cache_hiwater_bytes(); | |
752 | ds->cache_lowater_bytes.value.ui64 = dbuf_cache_lowater_bytes(); | |
753 | ds->hash_elements.value.ui64 = dbuf_hash_count; | |
754 | } | |
755 | ||
756 | return (0); | |
757 | } | |
d3c2ae1c | 758 | |
34dc7c2f BB |
759 | void |
760 | dbuf_init(void) | |
761 | { | |
762 | uint64_t hsize = 1ULL << 16; | |
763 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
764 | int i; | |
765 | ||
766 | /* | |
767 | * The hash table is big enough to fill all of physical memory | |
69de3421 TC |
768 | * with an average block size of zfs_arc_average_blocksize (default 8K). |
769 | * By default, the table will take up | |
770 | * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). | |
34dc7c2f | 771 | */ |
69de3421 | 772 | while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE) |
34dc7c2f BB |
773 | hsize <<= 1; |
774 | ||
775 | retry: | |
776 | h->hash_table_mask = hsize - 1; | |
93ce2b4c | 777 | #if defined(_KERNEL) |
d1d7e268 MK |
778 | /* |
779 | * Large allocations which do not require contiguous pages | |
780 | * should be using vmem_alloc() in the linux kernel | |
781 | */ | |
79c76d5b | 782 | h->hash_table = vmem_zalloc(hsize * sizeof (void *), KM_SLEEP); |
00b46022 | 783 | #else |
34dc7c2f | 784 | h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP); |
00b46022 | 785 | #endif |
34dc7c2f BB |
786 | if (h->hash_table == NULL) { |
787 | /* XXX - we should really return an error instead of assert */ | |
788 | ASSERT(hsize > (1ULL << 10)); | |
789 | hsize >>= 1; | |
790 | goto retry; | |
791 | } | |
792 | ||
d3c2ae1c | 793 | dbuf_kmem_cache = kmem_cache_create("dmu_buf_impl_t", |
34dc7c2f BB |
794 | sizeof (dmu_buf_impl_t), |
795 | 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0); | |
796 | ||
797 | for (i = 0; i < DBUF_MUTEXES; i++) | |
40d06e3c | 798 | mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL); |
e0b0ca98 BB |
799 | |
800 | dbuf_stats_init(h); | |
0c66c32d | 801 | |
d3c2ae1c | 802 | /* |
2e5dc449 MA |
803 | * Setup the parameters for the dbuf caches. We set the sizes of the |
804 | * dbuf cache and the metadata cache to 1/32nd and 1/16th (default) | |
805 | * of the target size of the ARC. If the values has been specified as | |
806 | * a module option and they're not greater than the target size of the | |
807 | * ARC, then we honor that value. | |
d3c2ae1c | 808 | */ |
de4f8d5d BB |
809 | if (dbuf_cache_max_bytes == 0 || |
810 | dbuf_cache_max_bytes >= arc_target_bytes()) { | |
811 | dbuf_cache_max_bytes = arc_target_bytes() >> dbuf_cache_shift; | |
812 | } | |
2e5dc449 MA |
813 | if (dbuf_metadata_cache_max_bytes == 0 || |
814 | dbuf_metadata_cache_max_bytes >= arc_target_bytes()) { | |
815 | dbuf_metadata_cache_max_bytes = | |
816 | arc_target_bytes() >> dbuf_metadata_cache_shift; | |
817 | } | |
d3c2ae1c | 818 | |
0c66c32d JG |
819 | /* |
820 | * All entries are queued via taskq_dispatch_ent(), so min/maxalloc | |
821 | * configuration is not required. | |
822 | */ | |
1229323d | 823 | dbu_evict_taskq = taskq_create("dbu_evict", 1, defclsyspri, 0, 0, 0); |
d3c2ae1c | 824 | |
2e5dc449 MA |
825 | for (dbuf_cached_state_t dcs = 0; dcs < DB_CACHE_MAX; dcs++) { |
826 | dbuf_caches[dcs].cache = | |
827 | multilist_create(sizeof (dmu_buf_impl_t), | |
828 | offsetof(dmu_buf_impl_t, db_cache_link), | |
829 | dbuf_cache_multilist_index_func); | |
424fd7c3 | 830 | zfs_refcount_create(&dbuf_caches[dcs].size); |
2e5dc449 | 831 | } |
d3c2ae1c | 832 | |
d3c2ae1c GW |
833 | dbuf_evict_thread_exit = B_FALSE; |
834 | mutex_init(&dbuf_evict_lock, NULL, MUTEX_DEFAULT, NULL); | |
835 | cv_init(&dbuf_evict_cv, NULL, CV_DEFAULT, NULL); | |
836 | dbuf_cache_evict_thread = thread_create(NULL, 0, dbuf_evict_thread, | |
837 | NULL, 0, &p0, TS_RUN, minclsyspri); | |
5e021f56 GDN |
838 | |
839 | dbuf_ksp = kstat_create("zfs", 0, "dbufstats", "misc", | |
840 | KSTAT_TYPE_NAMED, sizeof (dbuf_stats) / sizeof (kstat_named_t), | |
841 | KSTAT_FLAG_VIRTUAL); | |
842 | if (dbuf_ksp != NULL) { | |
5e021f56 GDN |
843 | for (i = 0; i < DN_MAX_LEVELS; i++) { |
844 | snprintf(dbuf_stats.cache_levels[i].name, | |
845 | KSTAT_STRLEN, "cache_level_%d", i); | |
846 | dbuf_stats.cache_levels[i].data_type = | |
847 | KSTAT_DATA_UINT64; | |
848 | snprintf(dbuf_stats.cache_levels_bytes[i].name, | |
849 | KSTAT_STRLEN, "cache_level_%d_bytes", i); | |
850 | dbuf_stats.cache_levels_bytes[i].data_type = | |
851 | KSTAT_DATA_UINT64; | |
852 | } | |
741db5a3 AM |
853 | dbuf_ksp->ks_data = &dbuf_stats; |
854 | dbuf_ksp->ks_update = dbuf_kstat_update; | |
855 | kstat_install(dbuf_ksp); | |
5e021f56 | 856 | } |
34dc7c2f BB |
857 | } |
858 | ||
859 | void | |
860 | dbuf_fini(void) | |
861 | { | |
862 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
863 | int i; | |
864 | ||
e0b0ca98 BB |
865 | dbuf_stats_destroy(); |
866 | ||
34dc7c2f BB |
867 | for (i = 0; i < DBUF_MUTEXES; i++) |
868 | mutex_destroy(&h->hash_mutexes[i]); | |
93ce2b4c | 869 | #if defined(_KERNEL) |
d1d7e268 MK |
870 | /* |
871 | * Large allocations which do not require contiguous pages | |
872 | * should be using vmem_free() in the linux kernel | |
873 | */ | |
00b46022 BB |
874 | vmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *)); |
875 | #else | |
34dc7c2f | 876 | kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *)); |
00b46022 | 877 | #endif |
d3c2ae1c | 878 | kmem_cache_destroy(dbuf_kmem_cache); |
0c66c32d | 879 | taskq_destroy(dbu_evict_taskq); |
d3c2ae1c GW |
880 | |
881 | mutex_enter(&dbuf_evict_lock); | |
882 | dbuf_evict_thread_exit = B_TRUE; | |
883 | while (dbuf_evict_thread_exit) { | |
884 | cv_signal(&dbuf_evict_cv); | |
885 | cv_wait(&dbuf_evict_cv, &dbuf_evict_lock); | |
886 | } | |
887 | mutex_exit(&dbuf_evict_lock); | |
d3c2ae1c GW |
888 | |
889 | mutex_destroy(&dbuf_evict_lock); | |
890 | cv_destroy(&dbuf_evict_cv); | |
891 | ||
2e5dc449 | 892 | for (dbuf_cached_state_t dcs = 0; dcs < DB_CACHE_MAX; dcs++) { |
424fd7c3 | 893 | zfs_refcount_destroy(&dbuf_caches[dcs].size); |
2e5dc449 MA |
894 | multilist_destroy(dbuf_caches[dcs].cache); |
895 | } | |
5e021f56 GDN |
896 | |
897 | if (dbuf_ksp != NULL) { | |
898 | kstat_delete(dbuf_ksp); | |
899 | dbuf_ksp = NULL; | |
900 | } | |
34dc7c2f BB |
901 | } |
902 | ||
903 | /* | |
904 | * Other stuff. | |
905 | */ | |
906 | ||
907 | #ifdef ZFS_DEBUG | |
908 | static void | |
909 | dbuf_verify(dmu_buf_impl_t *db) | |
910 | { | |
572e2857 | 911 | dnode_t *dn; |
428870ff | 912 | dbuf_dirty_record_t *dr; |
cccbed9f | 913 | uint32_t txg_prev; |
34dc7c2f BB |
914 | |
915 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
916 | ||
917 | if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY)) | |
918 | return; | |
919 | ||
920 | ASSERT(db->db_objset != NULL); | |
572e2857 BB |
921 | DB_DNODE_ENTER(db); |
922 | dn = DB_DNODE(db); | |
34dc7c2f BB |
923 | if (dn == NULL) { |
924 | ASSERT(db->db_parent == NULL); | |
925 | ASSERT(db->db_blkptr == NULL); | |
926 | } else { | |
927 | ASSERT3U(db->db.db_object, ==, dn->dn_object); | |
928 | ASSERT3P(db->db_objset, ==, dn->dn_objset); | |
929 | ASSERT3U(db->db_level, <, dn->dn_nlevels); | |
572e2857 BB |
930 | ASSERT(db->db_blkid == DMU_BONUS_BLKID || |
931 | db->db_blkid == DMU_SPILL_BLKID || | |
8951cb8d | 932 | !avl_is_empty(&dn->dn_dbufs)); |
34dc7c2f | 933 | } |
428870ff BB |
934 | if (db->db_blkid == DMU_BONUS_BLKID) { |
935 | ASSERT(dn != NULL); | |
936 | ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); | |
937 | ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID); | |
938 | } else if (db->db_blkid == DMU_SPILL_BLKID) { | |
34dc7c2f | 939 | ASSERT(dn != NULL); |
c99c9001 | 940 | ASSERT0(db->db.db_offset); |
34dc7c2f BB |
941 | } else { |
942 | ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size); | |
943 | } | |
944 | ||
cccbed9f | 945 | if ((dr = list_head(&db->db_dirty_records)) != NULL) { |
428870ff | 946 | ASSERT(dr->dr_dbuf == db); |
cccbed9f MM |
947 | txg_prev = dr->dr_txg; |
948 | for (dr = list_next(&db->db_dirty_records, dr); dr != NULL; | |
949 | dr = list_next(&db->db_dirty_records, dr)) { | |
950 | ASSERT(dr->dr_dbuf == db); | |
951 | ASSERT(txg_prev > dr->dr_txg); | |
952 | txg_prev = dr->dr_txg; | |
953 | } | |
954 | } | |
428870ff | 955 | |
b128c09f BB |
956 | /* |
957 | * We can't assert that db_size matches dn_datablksz because it | |
958 | * can be momentarily different when another thread is doing | |
959 | * dnode_set_blksz(). | |
960 | */ | |
961 | if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) { | |
428870ff | 962 | dr = db->db_data_pending; |
b128c09f BB |
963 | /* |
964 | * It should only be modified in syncing context, so | |
965 | * make sure we only have one copy of the data. | |
966 | */ | |
967 | ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf); | |
34dc7c2f BB |
968 | } |
969 | ||
970 | /* verify db->db_blkptr */ | |
971 | if (db->db_blkptr) { | |
972 | if (db->db_parent == dn->dn_dbuf) { | |
973 | /* db is pointed to by the dnode */ | |
974 | /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */ | |
9babb374 | 975 | if (DMU_OBJECT_IS_SPECIAL(db->db.db_object)) |
34dc7c2f BB |
976 | ASSERT(db->db_parent == NULL); |
977 | else | |
978 | ASSERT(db->db_parent != NULL); | |
428870ff BB |
979 | if (db->db_blkid != DMU_SPILL_BLKID) |
980 | ASSERT3P(db->db_blkptr, ==, | |
981 | &dn->dn_phys->dn_blkptr[db->db_blkid]); | |
34dc7c2f BB |
982 | } else { |
983 | /* db is pointed to by an indirect block */ | |
2a8ba608 MM |
984 | int epb __maybe_unused = db->db_parent->db.db_size >> |
985 | SPA_BLKPTRSHIFT; | |
34dc7c2f BB |
986 | ASSERT3U(db->db_parent->db_level, ==, db->db_level+1); |
987 | ASSERT3U(db->db_parent->db.db_object, ==, | |
988 | db->db.db_object); | |
989 | /* | |
990 | * dnode_grow_indblksz() can make this fail if we don't | |
f664f1ee | 991 | * have the parent's rwlock. XXX indblksz no longer |
34dc7c2f BB |
992 | * grows. safe to do this now? |
993 | */ | |
f664f1ee | 994 | if (RW_LOCK_HELD(&db->db_parent->db_rwlock)) { |
34dc7c2f BB |
995 | ASSERT3P(db->db_blkptr, ==, |
996 | ((blkptr_t *)db->db_parent->db.db_data + | |
997 | db->db_blkid % epb)); | |
998 | } | |
999 | } | |
1000 | } | |
1001 | if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) && | |
428870ff BB |
1002 | (db->db_buf == NULL || db->db_buf->b_data) && |
1003 | db->db.db_data && db->db_blkid != DMU_BONUS_BLKID && | |
34dc7c2f BB |
1004 | db->db_state != DB_FILL && !dn->dn_free_txg) { |
1005 | /* | |
1006 | * If the blkptr isn't set but they have nonzero data, | |
1007 | * it had better be dirty, otherwise we'll lose that | |
1008 | * data when we evict this buffer. | |
bc77ba73 PD |
1009 | * |
1010 | * There is an exception to this rule for indirect blocks; in | |
1011 | * this case, if the indirect block is a hole, we fill in a few | |
1012 | * fields on each of the child blocks (importantly, birth time) | |
1013 | * to prevent hole birth times from being lost when you | |
1014 | * partially fill in a hole. | |
34dc7c2f BB |
1015 | */ |
1016 | if (db->db_dirtycnt == 0) { | |
bc77ba73 PD |
1017 | if (db->db_level == 0) { |
1018 | uint64_t *buf = db->db.db_data; | |
1019 | int i; | |
34dc7c2f | 1020 | |
bc77ba73 PD |
1021 | for (i = 0; i < db->db.db_size >> 3; i++) { |
1022 | ASSERT(buf[i] == 0); | |
1023 | } | |
1024 | } else { | |
bc77ba73 PD |
1025 | blkptr_t *bps = db->db.db_data; |
1026 | ASSERT3U(1 << DB_DNODE(db)->dn_indblkshift, ==, | |
1027 | db->db.db_size); | |
1028 | /* | |
1029 | * We want to verify that all the blkptrs in the | |
1030 | * indirect block are holes, but we may have | |
1031 | * automatically set up a few fields for them. | |
1032 | * We iterate through each blkptr and verify | |
1033 | * they only have those fields set. | |
1034 | */ | |
1c27024e | 1035 | for (int i = 0; |
bc77ba73 PD |
1036 | i < db->db.db_size / sizeof (blkptr_t); |
1037 | i++) { | |
1038 | blkptr_t *bp = &bps[i]; | |
1039 | ASSERT(ZIO_CHECKSUM_IS_ZERO( | |
1040 | &bp->blk_cksum)); | |
1041 | ASSERT( | |
1042 | DVA_IS_EMPTY(&bp->blk_dva[0]) && | |
1043 | DVA_IS_EMPTY(&bp->blk_dva[1]) && | |
1044 | DVA_IS_EMPTY(&bp->blk_dva[2])); | |
1045 | ASSERT0(bp->blk_fill); | |
1046 | ASSERT0(bp->blk_pad[0]); | |
1047 | ASSERT0(bp->blk_pad[1]); | |
1048 | ASSERT(!BP_IS_EMBEDDED(bp)); | |
1049 | ASSERT(BP_IS_HOLE(bp)); | |
1050 | ASSERT0(bp->blk_phys_birth); | |
1051 | } | |
34dc7c2f BB |
1052 | } |
1053 | } | |
1054 | } | |
572e2857 | 1055 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
1056 | } |
1057 | #endif | |
1058 | ||
0c66c32d JG |
1059 | static void |
1060 | dbuf_clear_data(dmu_buf_impl_t *db) | |
1061 | { | |
1062 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
1063 | dbuf_evict_user(db); | |
d3c2ae1c | 1064 | ASSERT3P(db->db_buf, ==, NULL); |
0c66c32d JG |
1065 | db->db.db_data = NULL; |
1066 | if (db->db_state != DB_NOFILL) | |
1067 | db->db_state = DB_UNCACHED; | |
1068 | } | |
1069 | ||
34dc7c2f BB |
1070 | static void |
1071 | dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf) | |
1072 | { | |
1073 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
0c66c32d JG |
1074 | ASSERT(buf != NULL); |
1075 | ||
34dc7c2f | 1076 | db->db_buf = buf; |
0c66c32d JG |
1077 | ASSERT(buf->b_data != NULL); |
1078 | db->db.db_data = buf->b_data; | |
34dc7c2f BB |
1079 | } |
1080 | ||
428870ff BB |
1081 | /* |
1082 | * Loan out an arc_buf for read. Return the loaned arc_buf. | |
1083 | */ | |
1084 | arc_buf_t * | |
1085 | dbuf_loan_arcbuf(dmu_buf_impl_t *db) | |
1086 | { | |
1087 | arc_buf_t *abuf; | |
1088 | ||
d3c2ae1c | 1089 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
428870ff | 1090 | mutex_enter(&db->db_mtx); |
424fd7c3 | 1091 | if (arc_released(db->db_buf) || zfs_refcount_count(&db->db_holds) > 1) { |
428870ff | 1092 | int blksz = db->db.db_size; |
b0bc7a84 | 1093 | spa_t *spa = db->db_objset->os_spa; |
572e2857 | 1094 | |
428870ff | 1095 | mutex_exit(&db->db_mtx); |
2aa34383 | 1096 | abuf = arc_loan_buf(spa, B_FALSE, blksz); |
428870ff BB |
1097 | bcopy(db->db.db_data, abuf->b_data, blksz); |
1098 | } else { | |
1099 | abuf = db->db_buf; | |
1100 | arc_loan_inuse_buf(abuf, db); | |
d3c2ae1c | 1101 | db->db_buf = NULL; |
0c66c32d | 1102 | dbuf_clear_data(db); |
428870ff BB |
1103 | mutex_exit(&db->db_mtx); |
1104 | } | |
1105 | return (abuf); | |
1106 | } | |
1107 | ||
fcff0f35 PD |
1108 | /* |
1109 | * Calculate which level n block references the data at the level 0 offset | |
1110 | * provided. | |
1111 | */ | |
34dc7c2f | 1112 | uint64_t |
031d7c2f | 1113 | dbuf_whichblock(const dnode_t *dn, const int64_t level, const uint64_t offset) |
34dc7c2f | 1114 | { |
fcff0f35 PD |
1115 | if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) { |
1116 | /* | |
1117 | * The level n blkid is equal to the level 0 blkid divided by | |
1118 | * the number of level 0s in a level n block. | |
1119 | * | |
1120 | * The level 0 blkid is offset >> datablkshift = | |
1121 | * offset / 2^datablkshift. | |
1122 | * | |
1123 | * The number of level 0s in a level n is the number of block | |
1124 | * pointers in an indirect block, raised to the power of level. | |
1125 | * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level = | |
1126 | * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)). | |
1127 | * | |
1128 | * Thus, the level n blkid is: offset / | |
fe8a7982 | 1129 | * ((2^datablkshift)*(2^(level*(indblkshift-SPA_BLKPTRSHIFT)))) |
fcff0f35 PD |
1130 | * = offset / 2^(datablkshift + level * |
1131 | * (indblkshift - SPA_BLKPTRSHIFT)) | |
1132 | * = offset >> (datablkshift + level * | |
1133 | * (indblkshift - SPA_BLKPTRSHIFT)) | |
1134 | */ | |
031d7c2f GN |
1135 | |
1136 | const unsigned exp = dn->dn_datablkshift + | |
1137 | level * (dn->dn_indblkshift - SPA_BLKPTRSHIFT); | |
1138 | ||
1139 | if (exp >= 8 * sizeof (offset)) { | |
1140 | /* This only happens on the highest indirection level */ | |
1141 | ASSERT3U(level, ==, dn->dn_nlevels - 1); | |
1142 | return (0); | |
1143 | } | |
1144 | ||
1145 | ASSERT3U(exp, <, 8 * sizeof (offset)); | |
1146 | ||
1147 | return (offset >> exp); | |
34dc7c2f BB |
1148 | } else { |
1149 | ASSERT3U(offset, <, dn->dn_datablksz); | |
1150 | return (0); | |
1151 | } | |
1152 | } | |
1153 | ||
f664f1ee PD |
1154 | /* |
1155 | * This function is used to lock the parent of the provided dbuf. This should be | |
1156 | * used when modifying or reading db_blkptr. | |
1157 | */ | |
1158 | db_lock_type_t | |
1159 | dmu_buf_lock_parent(dmu_buf_impl_t *db, krw_t rw, void *tag) | |
1160 | { | |
1161 | enum db_lock_type ret = DLT_NONE; | |
1162 | if (db->db_parent != NULL) { | |
1163 | rw_enter(&db->db_parent->db_rwlock, rw); | |
1164 | ret = DLT_PARENT; | |
1165 | } else if (dmu_objset_ds(db->db_objset) != NULL) { | |
1166 | rrw_enter(&dmu_objset_ds(db->db_objset)->ds_bp_rwlock, rw, | |
1167 | tag); | |
1168 | ret = DLT_OBJSET; | |
1169 | } | |
1170 | /* | |
1171 | * We only return a DLT_NONE lock when it's the top-most indirect block | |
1172 | * of the meta-dnode of the MOS. | |
1173 | */ | |
1174 | return (ret); | |
1175 | } | |
1176 | ||
1177 | /* | |
1178 | * We need to pass the lock type in because it's possible that the block will | |
1179 | * move from being the topmost indirect block in a dnode (and thus, have no | |
1180 | * parent) to not the top-most via an indirection increase. This would cause a | |
1181 | * panic if we didn't pass the lock type in. | |
1182 | */ | |
1183 | void | |
1184 | dmu_buf_unlock_parent(dmu_buf_impl_t *db, db_lock_type_t type, void *tag) | |
1185 | { | |
1186 | if (type == DLT_PARENT) | |
1187 | rw_exit(&db->db_parent->db_rwlock); | |
1188 | else if (type == DLT_OBJSET) | |
1189 | rrw_exit(&dmu_objset_ds(db->db_objset)->ds_bp_rwlock, tag); | |
1190 | } | |
1191 | ||
34dc7c2f | 1192 | static void |
d4a72f23 TC |
1193 | dbuf_read_done(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp, |
1194 | arc_buf_t *buf, void *vdb) | |
34dc7c2f BB |
1195 | { |
1196 | dmu_buf_impl_t *db = vdb; | |
1197 | ||
1198 | mutex_enter(&db->db_mtx); | |
1199 | ASSERT3U(db->db_state, ==, DB_READ); | |
1200 | /* | |
1201 | * All reads are synchronous, so we must have a hold on the dbuf | |
1202 | */ | |
424fd7c3 | 1203 | ASSERT(zfs_refcount_count(&db->db_holds) > 0); |
34dc7c2f BB |
1204 | ASSERT(db->db_buf == NULL); |
1205 | ASSERT(db->db.db_data == NULL); | |
c3bd3fb4 TC |
1206 | if (buf == NULL) { |
1207 | /* i/o error */ | |
1208 | ASSERT(zio == NULL || zio->io_error != 0); | |
1209 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); | |
1210 | ASSERT3P(db->db_buf, ==, NULL); | |
1211 | db->db_state = DB_UNCACHED; | |
1212 | } else if (db->db_level == 0 && db->db_freed_in_flight) { | |
1213 | /* freed in flight */ | |
1214 | ASSERT(zio == NULL || zio->io_error == 0); | |
34dc7c2f BB |
1215 | arc_release(buf, db); |
1216 | bzero(buf->b_data, db->db.db_size); | |
1217 | arc_buf_freeze(buf); | |
1218 | db->db_freed_in_flight = FALSE; | |
1219 | dbuf_set_data(db, buf); | |
1220 | db->db_state = DB_CACHED; | |
c3bd3fb4 TC |
1221 | } else { |
1222 | /* success */ | |
1223 | ASSERT(zio == NULL || zio->io_error == 0); | |
34dc7c2f BB |
1224 | dbuf_set_data(db, buf); |
1225 | db->db_state = DB_CACHED; | |
34dc7c2f BB |
1226 | } |
1227 | cv_broadcast(&db->db_changed); | |
3d503a76 | 1228 | dbuf_rele_and_unlock(db, NULL, B_FALSE); |
34dc7c2f BB |
1229 | } |
1230 | ||
69830602 TC |
1231 | |
1232 | /* | |
1233 | * This function ensures that, when doing a decrypting read of a block, | |
1234 | * we make sure we have decrypted the dnode associated with it. We must do | |
1235 | * this so that we ensure we are fully authenticating the checksum-of-MACs | |
1236 | * tree from the root of the objset down to this block. Indirect blocks are | |
1237 | * always verified against their secure checksum-of-MACs assuming that the | |
1238 | * dnode containing them is correct. Now that we are doing a decrypting read, | |
1239 | * we can be sure that the key is loaded and verify that assumption. This is | |
1240 | * especially important considering that we always read encrypted dnode | |
1241 | * blocks as raw data (without verifying their MACs) to start, and | |
1242 | * decrypt / authenticate them when we need to read an encrypted bonus buffer. | |
1243 | */ | |
1244 | static int | |
1245 | dbuf_read_verify_dnode_crypt(dmu_buf_impl_t *db, uint32_t flags) | |
1246 | { | |
1247 | int err = 0; | |
1248 | objset_t *os = db->db_objset; | |
1249 | arc_buf_t *dnode_abuf; | |
1250 | dnode_t *dn; | |
1251 | zbookmark_phys_t zb; | |
1252 | ||
1253 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
1254 | ||
1255 | if (!os->os_encrypted || os->os_raw_receive || | |
1256 | (flags & DB_RF_NO_DECRYPT) != 0) | |
1257 | return (0); | |
1258 | ||
1259 | DB_DNODE_ENTER(db); | |
1260 | dn = DB_DNODE(db); | |
1261 | dnode_abuf = (dn->dn_dbuf != NULL) ? dn->dn_dbuf->db_buf : NULL; | |
1262 | ||
1263 | if (dnode_abuf == NULL || !arc_is_encrypted(dnode_abuf)) { | |
1264 | DB_DNODE_EXIT(db); | |
1265 | return (0); | |
1266 | } | |
1267 | ||
1268 | SET_BOOKMARK(&zb, dmu_objset_id(os), | |
1269 | DMU_META_DNODE_OBJECT, 0, dn->dn_dbuf->db_blkid); | |
1270 | err = arc_untransform(dnode_abuf, os->os_spa, &zb, B_TRUE); | |
1271 | ||
1272 | /* | |
1273 | * An error code of EACCES tells us that the key is still not | |
1274 | * available. This is ok if we are only reading authenticated | |
1275 | * (and therefore non-encrypted) blocks. | |
1276 | */ | |
1277 | if (err == EACCES && ((db->db_blkid != DMU_BONUS_BLKID && | |
1278 | !DMU_OT_IS_ENCRYPTED(dn->dn_type)) || | |
1279 | (db->db_blkid == DMU_BONUS_BLKID && | |
1280 | !DMU_OT_IS_ENCRYPTED(dn->dn_bonustype)))) | |
1281 | err = 0; | |
1282 | ||
69830602 TC |
1283 | DB_DNODE_EXIT(db); |
1284 | ||
1285 | return (err); | |
1286 | } | |
1287 | ||
f664f1ee PD |
1288 | /* |
1289 | * Drops db_mtx and the parent lock specified by dblt and tag before | |
1290 | * returning. | |
1291 | */ | |
5f6d0b6f | 1292 | static int |
f664f1ee PD |
1293 | dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags, |
1294 | db_lock_type_t dblt, void *tag) | |
34dc7c2f | 1295 | { |
572e2857 | 1296 | dnode_t *dn; |
5dbd68a3 | 1297 | zbookmark_phys_t zb; |
2a432414 | 1298 | uint32_t aflags = ARC_FLAG_NOWAIT; |
b5256303 | 1299 | int err, zio_flags = 0; |
34dc7c2f | 1300 | |
572e2857 BB |
1301 | DB_DNODE_ENTER(db); |
1302 | dn = DB_DNODE(db); | |
424fd7c3 | 1303 | ASSERT(!zfs_refcount_is_zero(&db->db_holds)); |
34dc7c2f BB |
1304 | ASSERT(MUTEX_HELD(&db->db_mtx)); |
1305 | ASSERT(db->db_state == DB_UNCACHED); | |
1306 | ASSERT(db->db_buf == NULL); | |
f664f1ee PD |
1307 | ASSERT(db->db_parent == NULL || |
1308 | RW_LOCK_HELD(&db->db_parent->db_rwlock)); | |
34dc7c2f | 1309 | |
428870ff | 1310 | if (db->db_blkid == DMU_BONUS_BLKID) { |
50c957f7 NB |
1311 | /* |
1312 | * The bonus length stored in the dnode may be less than | |
1313 | * the maximum available space in the bonus buffer. | |
1314 | */ | |
9babb374 | 1315 | int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen); |
50c957f7 | 1316 | int max_bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots); |
b5256303 TC |
1317 | |
1318 | /* if the underlying dnode block is encrypted, decrypt it */ | |
69830602 TC |
1319 | err = dbuf_read_verify_dnode_crypt(db, flags); |
1320 | if (err != 0) { | |
1321 | DB_DNODE_EXIT(db); | |
1322 | mutex_exit(&db->db_mtx); | |
1323 | return (err); | |
b5256303 | 1324 | } |
34dc7c2f BB |
1325 | |
1326 | ASSERT3U(bonuslen, <=, db->db.db_size); | |
a3fd9d9e | 1327 | db->db.db_data = kmem_alloc(max_bonuslen, KM_SLEEP); |
25458cbe | 1328 | arc_space_consume(max_bonuslen, ARC_SPACE_BONUS); |
50c957f7 NB |
1329 | if (bonuslen < max_bonuslen) |
1330 | bzero(db->db.db_data, max_bonuslen); | |
9babb374 BB |
1331 | if (bonuslen) |
1332 | bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen); | |
572e2857 | 1333 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
1334 | db->db_state = DB_CACHED; |
1335 | mutex_exit(&db->db_mtx); | |
f664f1ee | 1336 | dmu_buf_unlock_parent(db, dblt, tag); |
5f6d0b6f | 1337 | return (0); |
34dc7c2f BB |
1338 | } |
1339 | ||
b128c09f BB |
1340 | /* |
1341 | * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync() | |
1342 | * processes the delete record and clears the bp while we are waiting | |
1343 | * for the dn_mtx (resulting in a "no" from block_freed). | |
1344 | */ | |
1345 | if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) || | |
1346 | (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) || | |
1347 | BP_IS_HOLE(db->db_blkptr)))) { | |
34dc7c2f BB |
1348 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); |
1349 | ||
2aa34383 DK |
1350 | dbuf_set_data(db, arc_alloc_buf(db->db_objset->os_spa, db, type, |
1351 | db->db.db_size)); | |
34dc7c2f | 1352 | bzero(db->db.db_data, db->db.db_size); |
bc77ba73 PD |
1353 | |
1354 | if (db->db_blkptr != NULL && db->db_level > 0 && | |
1355 | BP_IS_HOLE(db->db_blkptr) && | |
1356 | db->db_blkptr->blk_birth != 0) { | |
1357 | blkptr_t *bps = db->db.db_data; | |
1c27024e | 1358 | for (int i = 0; i < ((1 << |
bc77ba73 PD |
1359 | DB_DNODE(db)->dn_indblkshift) / sizeof (blkptr_t)); |
1360 | i++) { | |
1361 | blkptr_t *bp = &bps[i]; | |
1362 | ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, | |
1363 | 1 << dn->dn_indblkshift); | |
1364 | BP_SET_LSIZE(bp, | |
1365 | BP_GET_LEVEL(db->db_blkptr) == 1 ? | |
1366 | dn->dn_datablksz : | |
1367 | BP_GET_LSIZE(db->db_blkptr)); | |
1368 | BP_SET_TYPE(bp, BP_GET_TYPE(db->db_blkptr)); | |
1369 | BP_SET_LEVEL(bp, | |
1370 | BP_GET_LEVEL(db->db_blkptr) - 1); | |
1371 | BP_SET_BIRTH(bp, db->db_blkptr->blk_birth, 0); | |
1372 | } | |
1373 | } | |
1374 | DB_DNODE_EXIT(db); | |
34dc7c2f | 1375 | db->db_state = DB_CACHED; |
34dc7c2f | 1376 | mutex_exit(&db->db_mtx); |
f664f1ee | 1377 | dmu_buf_unlock_parent(db, dblt, tag); |
5f6d0b6f | 1378 | return (0); |
34dc7c2f BB |
1379 | } |
1380 | ||
30af21b0 PD |
1381 | /* |
1382 | * Any attempt to read a redacted block should result in an error. This | |
1383 | * will never happen under normal conditions, but can be useful for | |
1384 | * debugging purposes. | |
1385 | */ | |
1386 | if (BP_IS_REDACTED(db->db_blkptr)) { | |
1387 | ASSERT(dsl_dataset_feature_is_active( | |
1388 | db->db_objset->os_dsl_dataset, | |
1389 | SPA_FEATURE_REDACTED_DATASETS)); | |
1390 | DB_DNODE_EXIT(db); | |
1391 | mutex_exit(&db->db_mtx); | |
1392 | return (SET_ERROR(EIO)); | |
1393 | } | |
1394 | ||
370bbf66 TC |
1395 | |
1396 | SET_BOOKMARK(&zb, dmu_objset_id(db->db_objset), | |
1397 | db->db.db_object, db->db_level, db->db_blkid); | |
1398 | ||
b5256303 TC |
1399 | /* |
1400 | * All bps of an encrypted os should have the encryption bit set. | |
1401 | * If this is not true it indicates tampering and we report an error. | |
1402 | */ | |
1403 | if (db->db_objset->os_encrypted && !BP_USES_CRYPT(db->db_blkptr)) { | |
1404 | spa_log_error(db->db_objset->os_spa, &zb); | |
1405 | zfs_panic_recover("unencrypted block in encrypted " | |
1406 | "object set %llu", dmu_objset_id(db->db_objset)); | |
69830602 TC |
1407 | DB_DNODE_EXIT(db); |
1408 | mutex_exit(&db->db_mtx); | |
f664f1ee | 1409 | dmu_buf_unlock_parent(db, dblt, tag); |
b5256303 TC |
1410 | return (SET_ERROR(EIO)); |
1411 | } | |
1412 | ||
69830602 TC |
1413 | err = dbuf_read_verify_dnode_crypt(db, flags); |
1414 | if (err != 0) { | |
1415 | DB_DNODE_EXIT(db); | |
f664f1ee | 1416 | dmu_buf_unlock_parent(db, dblt, tag); |
69830602 TC |
1417 | mutex_exit(&db->db_mtx); |
1418 | return (err); | |
1419 | } | |
1420 | ||
1421 | DB_DNODE_EXIT(db); | |
1422 | ||
1423 | db->db_state = DB_READ; | |
1424 | mutex_exit(&db->db_mtx); | |
1425 | ||
1426 | if (DBUF_IS_L2CACHEABLE(db)) | |
1427 | aflags |= ARC_FLAG_L2CACHE; | |
1428 | ||
34dc7c2f | 1429 | dbuf_add_ref(db, NULL); |
b128c09f | 1430 | |
b5256303 TC |
1431 | zio_flags = (flags & DB_RF_CANFAIL) ? |
1432 | ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED; | |
1433 | ||
1434 | if ((flags & DB_RF_NO_DECRYPT) && BP_IS_PROTECTED(db->db_blkptr)) | |
1435 | zio_flags |= ZIO_FLAG_RAW; | |
f664f1ee PD |
1436 | /* |
1437 | * The zio layer will copy the provided blkptr later, but we need to | |
1438 | * do this now so that we can release the parent's rwlock. We have to | |
1439 | * do that now so that if dbuf_read_done is called synchronously (on | |
1440 | * an l1 cache hit) we don't acquire the db_mtx while holding the | |
1441 | * parent's rwlock, which would be a lock ordering violation. | |
1442 | */ | |
1443 | blkptr_t bp = *db->db_blkptr; | |
1444 | dmu_buf_unlock_parent(db, dblt, tag); | |
1445 | (void) arc_read(zio, db->db_objset->os_spa, &bp, | |
b5256303 | 1446 | dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ, zio_flags, |
34dc7c2f | 1447 | &aflags, &zb); |
da8d5748 | 1448 | return (err); |
34dc7c2f BB |
1449 | } |
1450 | ||
2aa34383 DK |
1451 | /* |
1452 | * This is our just-in-time copy function. It makes a copy of buffers that | |
1453 | * have been modified in a previous transaction group before we access them in | |
1454 | * the current active group. | |
1455 | * | |
1456 | * This function is used in three places: when we are dirtying a buffer for the | |
1457 | * first time in a txg, when we are freeing a range in a dnode that includes | |
1458 | * this buffer, and when we are accessing a buffer which was received compressed | |
1459 | * and later referenced in a WRITE_BYREF record. | |
1460 | * | |
1461 | * Note that when we are called from dbuf_free_range() we do not put a hold on | |
1462 | * the buffer, we just traverse the active dbuf list for the dnode. | |
1463 | */ | |
1464 | static void | |
1465 | dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg) | |
1466 | { | |
cccbed9f | 1467 | dbuf_dirty_record_t *dr = list_head(&db->db_dirty_records); |
2aa34383 DK |
1468 | |
1469 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
1470 | ASSERT(db->db.db_data != NULL); | |
1471 | ASSERT(db->db_level == 0); | |
1472 | ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT); | |
1473 | ||
1474 | if (dr == NULL || | |
1475 | (dr->dt.dl.dr_data != | |
1476 | ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf))) | |
1477 | return; | |
1478 | ||
1479 | /* | |
1480 | * If the last dirty record for this dbuf has not yet synced | |
1481 | * and its referencing the dbuf data, either: | |
1482 | * reset the reference to point to a new copy, | |
1483 | * or (if there a no active holders) | |
1484 | * just null out the current db_data pointer. | |
1485 | */ | |
4807c0ba | 1486 | ASSERT3U(dr->dr_txg, >=, txg - 2); |
2aa34383 | 1487 | if (db->db_blkid == DMU_BONUS_BLKID) { |
2aa34383 DK |
1488 | dnode_t *dn = DB_DNODE(db); |
1489 | int bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots); | |
a3fd9d9e | 1490 | dr->dt.dl.dr_data = kmem_alloc(bonuslen, KM_SLEEP); |
2aa34383 DK |
1491 | arc_space_consume(bonuslen, ARC_SPACE_BONUS); |
1492 | bcopy(db->db.db_data, dr->dt.dl.dr_data, bonuslen); | |
424fd7c3 | 1493 | } else if (zfs_refcount_count(&db->db_holds) > db->db_dirtycnt) { |
b5256303 | 1494 | dnode_t *dn = DB_DNODE(db); |
2aa34383 DK |
1495 | int size = arc_buf_size(db->db_buf); |
1496 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); | |
1497 | spa_t *spa = db->db_objset->os_spa; | |
1498 | enum zio_compress compress_type = | |
1499 | arc_get_compression(db->db_buf); | |
1500 | ||
b5256303 TC |
1501 | if (arc_is_encrypted(db->db_buf)) { |
1502 | boolean_t byteorder; | |
1503 | uint8_t salt[ZIO_DATA_SALT_LEN]; | |
1504 | uint8_t iv[ZIO_DATA_IV_LEN]; | |
1505 | uint8_t mac[ZIO_DATA_MAC_LEN]; | |
1506 | ||
1507 | arc_get_raw_params(db->db_buf, &byteorder, salt, | |
1508 | iv, mac); | |
1509 | dr->dt.dl.dr_data = arc_alloc_raw_buf(spa, db, | |
1510 | dmu_objset_id(dn->dn_objset), byteorder, salt, iv, | |
1511 | mac, dn->dn_type, size, arc_buf_lsize(db->db_buf), | |
1512 | compress_type); | |
1513 | } else if (compress_type != ZIO_COMPRESS_OFF) { | |
2aa34383 DK |
1514 | ASSERT3U(type, ==, ARC_BUFC_DATA); |
1515 | dr->dt.dl.dr_data = arc_alloc_compressed_buf(spa, db, | |
1516 | size, arc_buf_lsize(db->db_buf), compress_type); | |
b5256303 TC |
1517 | } else { |
1518 | dr->dt.dl.dr_data = arc_alloc_buf(spa, db, type, size); | |
2aa34383 DK |
1519 | } |
1520 | bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size); | |
1521 | } else { | |
1522 | db->db_buf = NULL; | |
1523 | dbuf_clear_data(db); | |
1524 | } | |
1525 | } | |
1526 | ||
34dc7c2f BB |
1527 | int |
1528 | dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) | |
1529 | { | |
1530 | int err = 0; | |
b0bc7a84 | 1531 | boolean_t prefetch; |
572e2857 | 1532 | dnode_t *dn; |
34dc7c2f BB |
1533 | |
1534 | /* | |
1535 | * We don't have to hold the mutex to check db_state because it | |
1536 | * can't be freed while we have a hold on the buffer. | |
1537 | */ | |
424fd7c3 | 1538 | ASSERT(!zfs_refcount_is_zero(&db->db_holds)); |
34dc7c2f | 1539 | |
b128c09f | 1540 | if (db->db_state == DB_NOFILL) |
2e528b49 | 1541 | return (SET_ERROR(EIO)); |
b128c09f | 1542 | |
572e2857 BB |
1543 | DB_DNODE_ENTER(db); |
1544 | dn = DB_DNODE(db); | |
34dc7c2f | 1545 | |
428870ff | 1546 | prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && |
572e2857 | 1547 | (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL && |
b128c09f | 1548 | DBUF_IS_CACHEABLE(db); |
34dc7c2f BB |
1549 | |
1550 | mutex_enter(&db->db_mtx); | |
1551 | if (db->db_state == DB_CACHED) { | |
b5256303 TC |
1552 | spa_t *spa = dn->dn_objset->os_spa; |
1553 | ||
2aa34383 | 1554 | /* |
69830602 TC |
1555 | * Ensure that this block's dnode has been decrypted if |
1556 | * the caller has requested decrypted data. | |
2aa34383 | 1557 | */ |
69830602 TC |
1558 | err = dbuf_read_verify_dnode_crypt(db, flags); |
1559 | ||
1560 | /* | |
1561 | * If the arc buf is compressed or encrypted and the caller | |
1562 | * requested uncompressed data, we need to untransform it | |
1563 | * before returning. We also call arc_untransform() on any | |
1564 | * unauthenticated blocks, which will verify their MAC if | |
1565 | * the key is now available. | |
1566 | */ | |
1567 | if (err == 0 && db->db_buf != NULL && | |
1568 | (flags & DB_RF_NO_DECRYPT) == 0 && | |
b5256303 | 1569 | (arc_is_encrypted(db->db_buf) || |
69830602 | 1570 | arc_is_unauthenticated(db->db_buf) || |
b5256303 | 1571 | arc_get_compression(db->db_buf) != ZIO_COMPRESS_OFF)) { |
a2c2ed1b TC |
1572 | zbookmark_phys_t zb; |
1573 | ||
1574 | SET_BOOKMARK(&zb, dmu_objset_id(db->db_objset), | |
1575 | db->db.db_object, db->db_level, db->db_blkid); | |
b5256303 | 1576 | dbuf_fix_old_data(db, spa_syncing_txg(spa)); |
a2c2ed1b | 1577 | err = arc_untransform(db->db_buf, spa, &zb, B_FALSE); |
2aa34383 DK |
1578 | dbuf_set_data(db, db->db_buf); |
1579 | } | |
34dc7c2f | 1580 | mutex_exit(&db->db_mtx); |
f664f1ee PD |
1581 | if (err == 0 && prefetch) { |
1582 | dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE, | |
1583 | flags & DB_RF_HAVESTRUCT); | |
1584 | } | |
572e2857 | 1585 | DB_DNODE_EXIT(db); |
5e021f56 | 1586 | DBUF_STAT_BUMP(hash_hits); |
34dc7c2f | 1587 | } else if (db->db_state == DB_UNCACHED) { |
572e2857 | 1588 | spa_t *spa = dn->dn_objset->os_spa; |
a0043383 | 1589 | boolean_t need_wait = B_FALSE; |
572e2857 | 1590 | |
f664f1ee PD |
1591 | db_lock_type_t dblt = dmu_buf_lock_parent(db, RW_READER, FTAG); |
1592 | ||
b0319c1f | 1593 | if (zio == NULL && |
a0043383 | 1594 | db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) { |
572e2857 | 1595 | zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); |
a0043383 MA |
1596 | need_wait = B_TRUE; |
1597 | } | |
f664f1ee PD |
1598 | err = dbuf_read_impl(db, zio, flags, dblt, FTAG); |
1599 | /* | |
1600 | * dbuf_read_impl has dropped db_mtx and our parent's rwlock | |
1601 | * for us | |
1602 | */ | |
1603 | if (!err && prefetch) { | |
1604 | dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE, | |
1605 | flags & DB_RF_HAVESTRUCT); | |
1606 | } | |
34dc7c2f | 1607 | |
572e2857 | 1608 | DB_DNODE_EXIT(db); |
5e021f56 | 1609 | DBUF_STAT_BUMP(hash_misses); |
34dc7c2f | 1610 | |
5e7f3ace TC |
1611 | /* |
1612 | * If we created a zio_root we must execute it to avoid | |
1613 | * leaking it, even if it isn't attached to any work due | |
1614 | * to an error in dbuf_read_impl(). | |
1615 | */ | |
1616 | if (need_wait) { | |
1617 | if (err == 0) | |
1618 | err = zio_wait(zio); | |
1619 | else | |
1620 | VERIFY0(zio_wait(zio)); | |
1621 | } | |
34dc7c2f | 1622 | } else { |
e49f1e20 WA |
1623 | /* |
1624 | * Another reader came in while the dbuf was in flight | |
1625 | * between UNCACHED and CACHED. Either a writer will finish | |
1626 | * writing the buffer (sending the dbuf to CACHED) or the | |
1627 | * first reader's request will reach the read_done callback | |
1628 | * and send the dbuf to CACHED. Otherwise, a failure | |
1629 | * occurred and the dbuf went to UNCACHED. | |
1630 | */ | |
34dc7c2f | 1631 | mutex_exit(&db->db_mtx); |
f664f1ee PD |
1632 | if (prefetch) { |
1633 | dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE, | |
1634 | flags & DB_RF_HAVESTRUCT); | |
1635 | } | |
572e2857 | 1636 | DB_DNODE_EXIT(db); |
5e021f56 | 1637 | DBUF_STAT_BUMP(hash_misses); |
34dc7c2f | 1638 | |
e49f1e20 | 1639 | /* Skip the wait per the caller's request. */ |
34dc7c2f BB |
1640 | mutex_enter(&db->db_mtx); |
1641 | if ((flags & DB_RF_NEVERWAIT) == 0) { | |
1642 | while (db->db_state == DB_READ || | |
1643 | db->db_state == DB_FILL) { | |
1644 | ASSERT(db->db_state == DB_READ || | |
1645 | (flags & DB_RF_HAVESTRUCT) == 0); | |
64dbba36 AL |
1646 | DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *, |
1647 | db, zio_t *, zio); | |
34dc7c2f BB |
1648 | cv_wait(&db->db_changed, &db->db_mtx); |
1649 | } | |
1650 | if (db->db_state == DB_UNCACHED) | |
2e528b49 | 1651 | err = SET_ERROR(EIO); |
34dc7c2f BB |
1652 | } |
1653 | mutex_exit(&db->db_mtx); | |
1654 | } | |
1655 | ||
34dc7c2f BB |
1656 | return (err); |
1657 | } | |
1658 | ||
1659 | static void | |
1660 | dbuf_noread(dmu_buf_impl_t *db) | |
1661 | { | |
424fd7c3 | 1662 | ASSERT(!zfs_refcount_is_zero(&db->db_holds)); |
428870ff | 1663 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f BB |
1664 | mutex_enter(&db->db_mtx); |
1665 | while (db->db_state == DB_READ || db->db_state == DB_FILL) | |
1666 | cv_wait(&db->db_changed, &db->db_mtx); | |
1667 | if (db->db_state == DB_UNCACHED) { | |
1668 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); | |
b0bc7a84 | 1669 | spa_t *spa = db->db_objset->os_spa; |
34dc7c2f BB |
1670 | |
1671 | ASSERT(db->db_buf == NULL); | |
1672 | ASSERT(db->db.db_data == NULL); | |
2aa34383 | 1673 | dbuf_set_data(db, arc_alloc_buf(spa, db, type, db->db.db_size)); |
34dc7c2f | 1674 | db->db_state = DB_FILL; |
b128c09f | 1675 | } else if (db->db_state == DB_NOFILL) { |
0c66c32d | 1676 | dbuf_clear_data(db); |
34dc7c2f BB |
1677 | } else { |
1678 | ASSERT3U(db->db_state, ==, DB_CACHED); | |
1679 | } | |
1680 | mutex_exit(&db->db_mtx); | |
1681 | } | |
1682 | ||
34dc7c2f BB |
1683 | void |
1684 | dbuf_unoverride(dbuf_dirty_record_t *dr) | |
1685 | { | |
1686 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
428870ff | 1687 | blkptr_t *bp = &dr->dt.dl.dr_overridden_by; |
34dc7c2f BB |
1688 | uint64_t txg = dr->dr_txg; |
1689 | ||
1690 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
00710365 AS |
1691 | /* |
1692 | * This assert is valid because dmu_sync() expects to be called by | |
1693 | * a zilog's get_data while holding a range lock. This call only | |
1694 | * comes from dbuf_dirty() callers who must also hold a range lock. | |
1695 | */ | |
34dc7c2f BB |
1696 | ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC); |
1697 | ASSERT(db->db_level == 0); | |
1698 | ||
428870ff | 1699 | if (db->db_blkid == DMU_BONUS_BLKID || |
34dc7c2f BB |
1700 | dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN) |
1701 | return; | |
1702 | ||
428870ff BB |
1703 | ASSERT(db->db_data_pending != dr); |
1704 | ||
34dc7c2f | 1705 | /* free this block */ |
b0bc7a84 MG |
1706 | if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite) |
1707 | zio_free(db->db_objset->os_spa, txg, bp); | |
428870ff | 1708 | |
34dc7c2f | 1709 | dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; |
03c6040b | 1710 | dr->dt.dl.dr_nopwrite = B_FALSE; |
0c03d21a | 1711 | dr->dt.dl.dr_has_raw_params = B_FALSE; |
03c6040b | 1712 | |
34dc7c2f BB |
1713 | /* |
1714 | * Release the already-written buffer, so we leave it in | |
1715 | * a consistent dirty state. Note that all callers are | |
1716 | * modifying the buffer, so they will immediately do | |
1717 | * another (redundant) arc_release(). Therefore, leave | |
1718 | * the buf thawed to save the effort of freezing & | |
1719 | * immediately re-thawing it. | |
1720 | */ | |
1721 | arc_release(dr->dt.dl.dr_data, db); | |
1722 | } | |
1723 | ||
b128c09f BB |
1724 | /* |
1725 | * Evict (if its unreferenced) or clear (if its referenced) any level-0 | |
1726 | * data blocks in the free range, so that any future readers will find | |
b0bc7a84 | 1727 | * empty blocks. |
b128c09f | 1728 | */ |
34dc7c2f | 1729 | void |
8951cb8d AR |
1730 | dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid, |
1731 | dmu_tx_t *tx) | |
34dc7c2f | 1732 | { |
0c66c32d JG |
1733 | dmu_buf_impl_t *db_search; |
1734 | dmu_buf_impl_t *db, *db_next; | |
34dc7c2f | 1735 | uint64_t txg = tx->tx_txg; |
8951cb8d | 1736 | avl_index_t where; |
8951cb8d | 1737 | |
9c9531cb GM |
1738 | if (end_blkid > dn->dn_maxblkid && |
1739 | !(start_blkid == DMU_SPILL_BLKID || end_blkid == DMU_SPILL_BLKID)) | |
8951cb8d AR |
1740 | end_blkid = dn->dn_maxblkid; |
1741 | dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid); | |
34dc7c2f | 1742 | |
0c66c32d | 1743 | db_search = kmem_alloc(sizeof (dmu_buf_impl_t), KM_SLEEP); |
8951cb8d AR |
1744 | db_search->db_level = 0; |
1745 | db_search->db_blkid = start_blkid; | |
9925c28c | 1746 | db_search->db_state = DB_SEARCH; |
ea97f8ce | 1747 | |
b663a23d | 1748 | mutex_enter(&dn->dn_dbufs_mtx); |
8951cb8d AR |
1749 | db = avl_find(&dn->dn_dbufs, db_search, &where); |
1750 | ASSERT3P(db, ==, NULL); | |
9c9531cb | 1751 | |
8951cb8d AR |
1752 | db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER); |
1753 | ||
1754 | for (; db != NULL; db = db_next) { | |
1755 | db_next = AVL_NEXT(&dn->dn_dbufs, db); | |
428870ff | 1756 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
b128c09f | 1757 | |
8951cb8d AR |
1758 | if (db->db_level != 0 || db->db_blkid > end_blkid) { |
1759 | break; | |
1760 | } | |
1761 | ASSERT3U(db->db_blkid, >=, start_blkid); | |
34dc7c2f BB |
1762 | |
1763 | /* found a level 0 buffer in the range */ | |
13fe0198 MA |
1764 | mutex_enter(&db->db_mtx); |
1765 | if (dbuf_undirty(db, tx)) { | |
1766 | /* mutex has been dropped and dbuf destroyed */ | |
34dc7c2f | 1767 | continue; |
13fe0198 | 1768 | } |
34dc7c2f | 1769 | |
34dc7c2f | 1770 | if (db->db_state == DB_UNCACHED || |
b128c09f | 1771 | db->db_state == DB_NOFILL || |
34dc7c2f BB |
1772 | db->db_state == DB_EVICTING) { |
1773 | ASSERT(db->db.db_data == NULL); | |
1774 | mutex_exit(&db->db_mtx); | |
1775 | continue; | |
1776 | } | |
1777 | if (db->db_state == DB_READ || db->db_state == DB_FILL) { | |
1778 | /* will be handled in dbuf_read_done or dbuf_rele */ | |
1779 | db->db_freed_in_flight = TRUE; | |
1780 | mutex_exit(&db->db_mtx); | |
1781 | continue; | |
1782 | } | |
424fd7c3 | 1783 | if (zfs_refcount_count(&db->db_holds) == 0) { |
34dc7c2f | 1784 | ASSERT(db->db_buf); |
d3c2ae1c | 1785 | dbuf_destroy(db); |
34dc7c2f BB |
1786 | continue; |
1787 | } | |
1788 | /* The dbuf is referenced */ | |
1789 | ||
cccbed9f MM |
1790 | if (!list_is_empty(&db->db_dirty_records)) { |
1791 | dbuf_dirty_record_t *dr; | |
34dc7c2f | 1792 | |
cccbed9f | 1793 | dr = list_head(&db->db_dirty_records); |
34dc7c2f BB |
1794 | if (dr->dr_txg == txg) { |
1795 | /* | |
1796 | * This buffer is "in-use", re-adjust the file | |
1797 | * size to reflect that this buffer may | |
1798 | * contain new data when we sync. | |
1799 | */ | |
428870ff BB |
1800 | if (db->db_blkid != DMU_SPILL_BLKID && |
1801 | db->db_blkid > dn->dn_maxblkid) | |
34dc7c2f BB |
1802 | dn->dn_maxblkid = db->db_blkid; |
1803 | dbuf_unoverride(dr); | |
1804 | } else { | |
1805 | /* | |
1806 | * This dbuf is not dirty in the open context. | |
1807 | * Either uncache it (if its not referenced in | |
1808 | * the open context) or reset its contents to | |
1809 | * empty. | |
1810 | */ | |
1811 | dbuf_fix_old_data(db, txg); | |
1812 | } | |
1813 | } | |
1814 | /* clear the contents if its cached */ | |
1815 | if (db->db_state == DB_CACHED) { | |
1816 | ASSERT(db->db.db_data != NULL); | |
1817 | arc_release(db->db_buf, db); | |
f664f1ee | 1818 | rw_enter(&db->db_rwlock, RW_WRITER); |
34dc7c2f | 1819 | bzero(db->db.db_data, db->db.db_size); |
f664f1ee | 1820 | rw_exit(&db->db_rwlock); |
34dc7c2f BB |
1821 | arc_buf_freeze(db->db_buf); |
1822 | } | |
1823 | ||
1824 | mutex_exit(&db->db_mtx); | |
1825 | } | |
8951cb8d | 1826 | |
8951cb8d | 1827 | kmem_free(db_search, sizeof (dmu_buf_impl_t)); |
34dc7c2f BB |
1828 | mutex_exit(&dn->dn_dbufs_mtx); |
1829 | } | |
1830 | ||
34dc7c2f BB |
1831 | void |
1832 | dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx) | |
1833 | { | |
1834 | arc_buf_t *buf, *obuf; | |
cccbed9f | 1835 | dbuf_dirty_record_t *dr; |
34dc7c2f BB |
1836 | int osize = db->db.db_size; |
1837 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); | |
572e2857 | 1838 | dnode_t *dn; |
34dc7c2f | 1839 | |
428870ff | 1840 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f | 1841 | |
572e2857 BB |
1842 | DB_DNODE_ENTER(db); |
1843 | dn = DB_DNODE(db); | |
1844 | ||
34dc7c2f BB |
1845 | /* |
1846 | * XXX we should be doing a dbuf_read, checking the return | |
1847 | * value and returning that up to our callers | |
1848 | */ | |
b0bc7a84 | 1849 | dmu_buf_will_dirty(&db->db, tx); |
34dc7c2f BB |
1850 | |
1851 | /* create the data buffer for the new block */ | |
2aa34383 | 1852 | buf = arc_alloc_buf(dn->dn_objset->os_spa, db, type, size); |
34dc7c2f BB |
1853 | |
1854 | /* copy old block data to the new block */ | |
1855 | obuf = db->db_buf; | |
1856 | bcopy(obuf->b_data, buf->b_data, MIN(osize, size)); | |
1857 | /* zero the remainder */ | |
1858 | if (size > osize) | |
1859 | bzero((uint8_t *)buf->b_data + osize, size - osize); | |
1860 | ||
1861 | mutex_enter(&db->db_mtx); | |
1862 | dbuf_set_data(db, buf); | |
d3c2ae1c | 1863 | arc_buf_destroy(obuf, db); |
34dc7c2f BB |
1864 | db->db.db_size = size; |
1865 | ||
cccbed9f MM |
1866 | dr = list_head(&db->db_dirty_records); |
1867 | if (db->db_level == 0) | |
1868 | dr->dt.dl.dr_data = buf; | |
1869 | ASSERT3U(dr->dr_txg, ==, tx->tx_txg); | |
1870 | ASSERT3U(dr->dr_accounted, ==, osize); | |
1871 | dr->dr_accounted = size; | |
34dc7c2f BB |
1872 | mutex_exit(&db->db_mtx); |
1873 | ||
3ec3bc21 | 1874 | dmu_objset_willuse_space(dn->dn_objset, size - osize, tx); |
572e2857 | 1875 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
1876 | } |
1877 | ||
428870ff BB |
1878 | void |
1879 | dbuf_release_bp(dmu_buf_impl_t *db) | |
1880 | { | |
2a8ba608 | 1881 | objset_t *os __maybe_unused = db->db_objset; |
428870ff BB |
1882 | |
1883 | ASSERT(dsl_pool_sync_context(dmu_objset_pool(os))); | |
1884 | ASSERT(arc_released(os->os_phys_buf) || | |
1885 | list_link_active(&os->os_dsl_dataset->ds_synced_link)); | |
1886 | ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf)); | |
1887 | ||
294f6806 | 1888 | (void) arc_release(db->db_buf, db); |
428870ff BB |
1889 | } |
1890 | ||
5a28a973 MA |
1891 | /* |
1892 | * We already have a dirty record for this TXG, and we are being | |
1893 | * dirtied again. | |
1894 | */ | |
1895 | static void | |
1896 | dbuf_redirty(dbuf_dirty_record_t *dr) | |
1897 | { | |
1898 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
1899 | ||
1900 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
1901 | ||
1902 | if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) { | |
1903 | /* | |
1904 | * If this buffer has already been written out, | |
1905 | * we now need to reset its state. | |
1906 | */ | |
1907 | dbuf_unoverride(dr); | |
1908 | if (db->db.db_object != DMU_META_DNODE_OBJECT && | |
1909 | db->db_state != DB_NOFILL) { | |
1910 | /* Already released on initial dirty, so just thaw. */ | |
1911 | ASSERT(arc_released(db->db_buf)); | |
1912 | arc_buf_thaw(db->db_buf); | |
1913 | } | |
1914 | } | |
1915 | } | |
1916 | ||
34dc7c2f BB |
1917 | dbuf_dirty_record_t * |
1918 | dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) | |
1919 | { | |
572e2857 BB |
1920 | dnode_t *dn; |
1921 | objset_t *os; | |
cccbed9f | 1922 | dbuf_dirty_record_t *dr, *dr_next, *dr_head; |
34dc7c2f | 1923 | int txgoff = tx->tx_txg & TXG_MASK; |
f664f1ee | 1924 | boolean_t drop_struct_rwlock = B_FALSE; |
34dc7c2f BB |
1925 | |
1926 | ASSERT(tx->tx_txg != 0); | |
424fd7c3 | 1927 | ASSERT(!zfs_refcount_is_zero(&db->db_holds)); |
34dc7c2f BB |
1928 | DMU_TX_DIRTY_BUF(tx, db); |
1929 | ||
572e2857 BB |
1930 | DB_DNODE_ENTER(db); |
1931 | dn = DB_DNODE(db); | |
34dc7c2f BB |
1932 | /* |
1933 | * Shouldn't dirty a regular buffer in syncing context. Private | |
1934 | * objects may be dirtied in syncing context, but only if they | |
1935 | * were already pre-dirtied in open context. | |
34dc7c2f | 1936 | */ |
cc9bb3e5 GM |
1937 | #ifdef DEBUG |
1938 | if (dn->dn_objset->os_dsl_dataset != NULL) { | |
1939 | rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, | |
1940 | RW_READER, FTAG); | |
1941 | } | |
34dc7c2f BB |
1942 | ASSERT(!dmu_tx_is_syncing(tx) || |
1943 | BP_IS_HOLE(dn->dn_objset->os_rootbp) || | |
9babb374 BB |
1944 | DMU_OBJECT_IS_SPECIAL(dn->dn_object) || |
1945 | dn->dn_objset->os_dsl_dataset == NULL); | |
cc9bb3e5 GM |
1946 | if (dn->dn_objset->os_dsl_dataset != NULL) |
1947 | rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, FTAG); | |
1948 | #endif | |
34dc7c2f BB |
1949 | /* |
1950 | * We make this assert for private objects as well, but after we | |
1951 | * check if we're already dirty. They are allowed to re-dirty | |
1952 | * in syncing context. | |
1953 | */ | |
1954 | ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || | |
1955 | dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == | |
1956 | (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); | |
1957 | ||
1958 | mutex_enter(&db->db_mtx); | |
1959 | /* | |
1960 | * XXX make this true for indirects too? The problem is that | |
1961 | * transactions created with dmu_tx_create_assigned() from | |
1962 | * syncing context don't bother holding ahead. | |
1963 | */ | |
1964 | ASSERT(db->db_level != 0 || | |
b128c09f BB |
1965 | db->db_state == DB_CACHED || db->db_state == DB_FILL || |
1966 | db->db_state == DB_NOFILL); | |
34dc7c2f BB |
1967 | |
1968 | mutex_enter(&dn->dn_mtx); | |
1969 | /* | |
1970 | * Don't set dirtyctx to SYNC if we're just modifying this as we | |
1971 | * initialize the objset. | |
1972 | */ | |
cc9bb3e5 GM |
1973 | if (dn->dn_dirtyctx == DN_UNDIRTIED) { |
1974 | if (dn->dn_objset->os_dsl_dataset != NULL) { | |
1975 | rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, | |
1976 | RW_READER, FTAG); | |
1977 | } | |
1978 | if (!BP_IS_HOLE(dn->dn_objset->os_rootbp)) { | |
1979 | dn->dn_dirtyctx = (dmu_tx_is_syncing(tx) ? | |
1980 | DN_DIRTY_SYNC : DN_DIRTY_OPEN); | |
1981 | ASSERT(dn->dn_dirtyctx_firstset == NULL); | |
1982 | dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP); | |
1983 | } | |
1984 | if (dn->dn_objset->os_dsl_dataset != NULL) { | |
1985 | rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, | |
1986 | FTAG); | |
1987 | } | |
34dc7c2f | 1988 | } |
edc1e713 TC |
1989 | |
1990 | if (tx->tx_txg > dn->dn_dirty_txg) | |
1991 | dn->dn_dirty_txg = tx->tx_txg; | |
34dc7c2f BB |
1992 | mutex_exit(&dn->dn_mtx); |
1993 | ||
428870ff BB |
1994 | if (db->db_blkid == DMU_SPILL_BLKID) |
1995 | dn->dn_have_spill = B_TRUE; | |
1996 | ||
34dc7c2f BB |
1997 | /* |
1998 | * If this buffer is already dirty, we're done. | |
1999 | */ | |
cccbed9f MM |
2000 | dr_head = list_head(&db->db_dirty_records); |
2001 | ASSERT(dr_head == NULL || dr_head->dr_txg <= tx->tx_txg || | |
34dc7c2f | 2002 | db->db.db_object == DMU_META_DNODE_OBJECT); |
cccbed9f MM |
2003 | dr_next = dbuf_find_dirty_lte(db, tx->tx_txg); |
2004 | if (dr_next && dr_next->dr_txg == tx->tx_txg) { | |
572e2857 BB |
2005 | DB_DNODE_EXIT(db); |
2006 | ||
cccbed9f | 2007 | dbuf_redirty(dr_next); |
34dc7c2f | 2008 | mutex_exit(&db->db_mtx); |
cccbed9f | 2009 | return (dr_next); |
34dc7c2f BB |
2010 | } |
2011 | ||
2012 | /* | |
2013 | * Only valid if not already dirty. | |
2014 | */ | |
9babb374 BB |
2015 | ASSERT(dn->dn_object == 0 || |
2016 | dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == | |
34dc7c2f BB |
2017 | (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); |
2018 | ||
2019 | ASSERT3U(dn->dn_nlevels, >, db->db_level); | |
34dc7c2f BB |
2020 | |
2021 | /* | |
2022 | * We should only be dirtying in syncing context if it's the | |
9babb374 BB |
2023 | * mos or we're initializing the os or it's a special object. |
2024 | * However, we are allowed to dirty in syncing context provided | |
2025 | * we already dirtied it in open context. Hence we must make | |
2026 | * this assertion only if we're not already dirty. | |
34dc7c2f | 2027 | */ |
572e2857 | 2028 | os = dn->dn_objset; |
3b7f360c | 2029 | VERIFY3U(tx->tx_txg, <=, spa_final_dirty_txg(os->os_spa)); |
cc9bb3e5 GM |
2030 | #ifdef DEBUG |
2031 | if (dn->dn_objset->os_dsl_dataset != NULL) | |
2032 | rrw_enter(&os->os_dsl_dataset->ds_bp_rwlock, RW_READER, FTAG); | |
9babb374 BB |
2033 | ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) || |
2034 | os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp)); | |
cc9bb3e5 GM |
2035 | if (dn->dn_objset->os_dsl_dataset != NULL) |
2036 | rrw_exit(&os->os_dsl_dataset->ds_bp_rwlock, FTAG); | |
2037 | #endif | |
34dc7c2f BB |
2038 | ASSERT(db->db.db_size != 0); |
2039 | ||
2040 | dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); | |
2041 | ||
428870ff | 2042 | if (db->db_blkid != DMU_BONUS_BLKID) { |
3ec3bc21 | 2043 | dmu_objset_willuse_space(os, db->db.db_size, tx); |
34dc7c2f BB |
2044 | } |
2045 | ||
2046 | /* | |
2047 | * If this buffer is dirty in an old transaction group we need | |
2048 | * to make a copy of it so that the changes we make in this | |
2049 | * transaction group won't leak out when we sync the older txg. | |
2050 | */ | |
79c76d5b | 2051 | dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP); |
98f72a53 | 2052 | list_link_init(&dr->dr_dirty_node); |
cccbed9f | 2053 | list_link_init(&dr->dr_dbuf_node); |
34dc7c2f BB |
2054 | if (db->db_level == 0) { |
2055 | void *data_old = db->db_buf; | |
2056 | ||
b128c09f | 2057 | if (db->db_state != DB_NOFILL) { |
428870ff | 2058 | if (db->db_blkid == DMU_BONUS_BLKID) { |
b128c09f BB |
2059 | dbuf_fix_old_data(db, tx->tx_txg); |
2060 | data_old = db->db.db_data; | |
2061 | } else if (db->db.db_object != DMU_META_DNODE_OBJECT) { | |
2062 | /* | |
2063 | * Release the data buffer from the cache so | |
2064 | * that we can modify it without impacting | |
2065 | * possible other users of this cached data | |
2066 | * block. Note that indirect blocks and | |
2067 | * private objects are not released until the | |
2068 | * syncing state (since they are only modified | |
2069 | * then). | |
2070 | */ | |
2071 | arc_release(db->db_buf, db); | |
2072 | dbuf_fix_old_data(db, tx->tx_txg); | |
2073 | data_old = db->db_buf; | |
2074 | } | |
2075 | ASSERT(data_old != NULL); | |
34dc7c2f | 2076 | } |
34dc7c2f BB |
2077 | dr->dt.dl.dr_data = data_old; |
2078 | } else { | |
448d7aaa | 2079 | mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_NOLOCKDEP, NULL); |
34dc7c2f BB |
2080 | list_create(&dr->dt.di.dr_children, |
2081 | sizeof (dbuf_dirty_record_t), | |
2082 | offsetof(dbuf_dirty_record_t, dr_dirty_node)); | |
2083 | } | |
0f8ff49e | 2084 | if (db->db_blkid != DMU_BONUS_BLKID) |
e8b96c60 | 2085 | dr->dr_accounted = db->db.db_size; |
34dc7c2f BB |
2086 | dr->dr_dbuf = db; |
2087 | dr->dr_txg = tx->tx_txg; | |
cccbed9f | 2088 | list_insert_before(&db->db_dirty_records, dr_next, dr); |
34dc7c2f BB |
2089 | |
2090 | /* | |
2091 | * We could have been freed_in_flight between the dbuf_noread | |
2092 | * and dbuf_dirty. We win, as though the dbuf_noread() had | |
2093 | * happened after the free. | |
2094 | */ | |
428870ff BB |
2095 | if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && |
2096 | db->db_blkid != DMU_SPILL_BLKID) { | |
34dc7c2f | 2097 | mutex_enter(&dn->dn_mtx); |
9bd274dd MA |
2098 | if (dn->dn_free_ranges[txgoff] != NULL) { |
2099 | range_tree_clear(dn->dn_free_ranges[txgoff], | |
2100 | db->db_blkid, 1); | |
2101 | } | |
34dc7c2f BB |
2102 | mutex_exit(&dn->dn_mtx); |
2103 | db->db_freed_in_flight = FALSE; | |
2104 | } | |
2105 | ||
2106 | /* | |
2107 | * This buffer is now part of this txg | |
2108 | */ | |
2109 | dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg); | |
2110 | db->db_dirtycnt += 1; | |
2111 | ASSERT3U(db->db_dirtycnt, <=, 3); | |
2112 | ||
2113 | mutex_exit(&db->db_mtx); | |
2114 | ||
428870ff BB |
2115 | if (db->db_blkid == DMU_BONUS_BLKID || |
2116 | db->db_blkid == DMU_SPILL_BLKID) { | |
34dc7c2f BB |
2117 | mutex_enter(&dn->dn_mtx); |
2118 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
2119 | list_insert_tail(&dn->dn_dirty_records[txgoff], dr); | |
2120 | mutex_exit(&dn->dn_mtx); | |
2121 | dnode_setdirty(dn, tx); | |
572e2857 | 2122 | DB_DNODE_EXIT(db); |
34dc7c2f | 2123 | return (dr); |
98ace739 MA |
2124 | } |
2125 | ||
98ace739 MA |
2126 | if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) { |
2127 | rw_enter(&dn->dn_struct_rwlock, RW_READER); | |
f664f1ee PD |
2128 | drop_struct_rwlock = B_TRUE; |
2129 | } | |
2130 | ||
2131 | /* | |
2132 | * If we are overwriting a dedup BP, then unless it is snapshotted, | |
2133 | * when we get to syncing context we will need to decrement its | |
2134 | * refcount in the DDT. Prefetch the relevant DDT block so that | |
2135 | * syncing context won't have to wait for the i/o. | |
2136 | */ | |
2137 | if (db->db_blkptr != NULL) { | |
2138 | db_lock_type_t dblt = dmu_buf_lock_parent(db, RW_READER, FTAG); | |
2139 | ddt_prefetch(os->os_spa, db->db_blkptr); | |
2140 | dmu_buf_unlock_parent(db, dblt, FTAG); | |
98ace739 MA |
2141 | } |
2142 | ||
2ade4a99 MA |
2143 | /* |
2144 | * We need to hold the dn_struct_rwlock to make this assertion, | |
2145 | * because it protects dn_phys / dn_next_nlevels from changing. | |
2146 | */ | |
2147 | ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) || | |
2148 | dn->dn_phys->dn_nlevels > db->db_level || | |
2149 | dn->dn_next_nlevels[txgoff] > db->db_level || | |
2150 | dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level || | |
2151 | dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level); | |
2152 | ||
34dc7c2f | 2153 | |
b128c09f | 2154 | if (db->db_level == 0) { |
69830602 TC |
2155 | ASSERT(!db->db_objset->os_raw_receive || |
2156 | dn->dn_maxblkid >= db->db_blkid); | |
369aa501 | 2157 | dnode_new_blkid(dn, db->db_blkid, tx, |
f664f1ee | 2158 | drop_struct_rwlock, B_FALSE); |
b128c09f BB |
2159 | ASSERT(dn->dn_maxblkid >= db->db_blkid); |
2160 | } | |
2161 | ||
34dc7c2f BB |
2162 | if (db->db_level+1 < dn->dn_nlevels) { |
2163 | dmu_buf_impl_t *parent = db->db_parent; | |
2164 | dbuf_dirty_record_t *di; | |
2165 | int parent_held = FALSE; | |
2166 | ||
2167 | if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) { | |
2168 | int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; | |
f664f1ee | 2169 | parent = dbuf_hold_level(dn, db->db_level + 1, |
34dc7c2f | 2170 | db->db_blkid >> epbs, FTAG); |
428870ff | 2171 | ASSERT(parent != NULL); |
34dc7c2f BB |
2172 | parent_held = TRUE; |
2173 | } | |
f664f1ee | 2174 | if (drop_struct_rwlock) |
34dc7c2f | 2175 | rw_exit(&dn->dn_struct_rwlock); |
f664f1ee | 2176 | ASSERT3U(db->db_level + 1, ==, parent->db_level); |
34dc7c2f BB |
2177 | di = dbuf_dirty(parent, tx); |
2178 | if (parent_held) | |
2179 | dbuf_rele(parent, FTAG); | |
2180 | ||
2181 | mutex_enter(&db->db_mtx); | |
e8b96c60 MA |
2182 | /* |
2183 | * Since we've dropped the mutex, it's possible that | |
2184 | * dbuf_undirty() might have changed this out from under us. | |
2185 | */ | |
cccbed9f | 2186 | if (list_head(&db->db_dirty_records) == dr || |
34dc7c2f BB |
2187 | dn->dn_object == DMU_META_DNODE_OBJECT) { |
2188 | mutex_enter(&di->dt.di.dr_mtx); | |
2189 | ASSERT3U(di->dr_txg, ==, tx->tx_txg); | |
2190 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
2191 | list_insert_tail(&di->dt.di.dr_children, dr); | |
2192 | mutex_exit(&di->dt.di.dr_mtx); | |
2193 | dr->dr_parent = di; | |
2194 | } | |
2195 | mutex_exit(&db->db_mtx); | |
2196 | } else { | |
f664f1ee | 2197 | ASSERT(db->db_level + 1 == dn->dn_nlevels); |
34dc7c2f | 2198 | ASSERT(db->db_blkid < dn->dn_nblkptr); |
572e2857 | 2199 | ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf); |
34dc7c2f BB |
2200 | mutex_enter(&dn->dn_mtx); |
2201 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
2202 | list_insert_tail(&dn->dn_dirty_records[txgoff], dr); | |
2203 | mutex_exit(&dn->dn_mtx); | |
f664f1ee | 2204 | if (drop_struct_rwlock) |
34dc7c2f BB |
2205 | rw_exit(&dn->dn_struct_rwlock); |
2206 | } | |
2207 | ||
2208 | dnode_setdirty(dn, tx); | |
572e2857 | 2209 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
2210 | return (dr); |
2211 | } | |
2212 | ||
13fe0198 | 2213 | /* |
e49f1e20 WA |
2214 | * Undirty a buffer in the transaction group referenced by the given |
2215 | * transaction. Return whether this evicted the dbuf. | |
13fe0198 MA |
2216 | */ |
2217 | static boolean_t | |
34dc7c2f BB |
2218 | dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx) |
2219 | { | |
572e2857 | 2220 | dnode_t *dn; |
34dc7c2f | 2221 | uint64_t txg = tx->tx_txg; |
cccbed9f | 2222 | dbuf_dirty_record_t *dr; |
34dc7c2f BB |
2223 | |
2224 | ASSERT(txg != 0); | |
4bda3bd0 MA |
2225 | |
2226 | /* | |
2227 | * Due to our use of dn_nlevels below, this can only be called | |
2228 | * in open context, unless we are operating on the MOS. | |
2229 | * From syncing context, dn_nlevels may be different from the | |
2230 | * dn_nlevels used when dbuf was dirtied. | |
2231 | */ | |
2232 | ASSERT(db->db_objset == | |
2233 | dmu_objset_pool(db->db_objset)->dp_meta_objset || | |
2234 | txg != spa_syncing_txg(dmu_objset_spa(db->db_objset))); | |
428870ff | 2235 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
13fe0198 MA |
2236 | ASSERT0(db->db_level); |
2237 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
34dc7c2f | 2238 | |
34dc7c2f BB |
2239 | /* |
2240 | * If this buffer is not dirty, we're done. | |
2241 | */ | |
cccbed9f MM |
2242 | dr = dbuf_find_dirty_eq(db, txg); |
2243 | if (dr == NULL) | |
13fe0198 | 2244 | return (B_FALSE); |
428870ff | 2245 | ASSERT(dr->dr_dbuf == db); |
34dc7c2f | 2246 | |
572e2857 BB |
2247 | DB_DNODE_ENTER(db); |
2248 | dn = DB_DNODE(db); | |
2249 | ||
34dc7c2f BB |
2250 | dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); |
2251 | ||
2252 | ASSERT(db->db.db_size != 0); | |
2253 | ||
4bda3bd0 MA |
2254 | dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset), |
2255 | dr->dr_accounted, txg); | |
34dc7c2f | 2256 | |
cccbed9f | 2257 | list_remove(&db->db_dirty_records, dr); |
34dc7c2f | 2258 | |
ef3c1dea GR |
2259 | /* |
2260 | * Note that there are three places in dbuf_dirty() | |
2261 | * where this dirty record may be put on a list. | |
2262 | * Make sure to do a list_remove corresponding to | |
2263 | * every one of those list_insert calls. | |
2264 | */ | |
34dc7c2f BB |
2265 | if (dr->dr_parent) { |
2266 | mutex_enter(&dr->dr_parent->dt.di.dr_mtx); | |
2267 | list_remove(&dr->dr_parent->dt.di.dr_children, dr); | |
2268 | mutex_exit(&dr->dr_parent->dt.di.dr_mtx); | |
ef3c1dea | 2269 | } else if (db->db_blkid == DMU_SPILL_BLKID || |
4bda3bd0 | 2270 | db->db_level + 1 == dn->dn_nlevels) { |
b128c09f | 2271 | ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf); |
34dc7c2f BB |
2272 | mutex_enter(&dn->dn_mtx); |
2273 | list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr); | |
2274 | mutex_exit(&dn->dn_mtx); | |
2275 | } | |
572e2857 | 2276 | DB_DNODE_EXIT(db); |
34dc7c2f | 2277 | |
13fe0198 MA |
2278 | if (db->db_state != DB_NOFILL) { |
2279 | dbuf_unoverride(dr); | |
34dc7c2f | 2280 | |
34dc7c2f | 2281 | ASSERT(db->db_buf != NULL); |
13fe0198 MA |
2282 | ASSERT(dr->dt.dl.dr_data != NULL); |
2283 | if (dr->dt.dl.dr_data != db->db_buf) | |
d3c2ae1c | 2284 | arc_buf_destroy(dr->dt.dl.dr_data, db); |
34dc7c2f | 2285 | } |
58c4aa00 | 2286 | |
34dc7c2f BB |
2287 | kmem_free(dr, sizeof (dbuf_dirty_record_t)); |
2288 | ||
2289 | ASSERT(db->db_dirtycnt > 0); | |
2290 | db->db_dirtycnt -= 1; | |
2291 | ||
424fd7c3 | 2292 | if (zfs_refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) { |
d3c2ae1c GW |
2293 | ASSERT(db->db_state == DB_NOFILL || arc_released(db->db_buf)); |
2294 | dbuf_destroy(db); | |
13fe0198 | 2295 | return (B_TRUE); |
34dc7c2f BB |
2296 | } |
2297 | ||
13fe0198 | 2298 | return (B_FALSE); |
34dc7c2f BB |
2299 | } |
2300 | ||
b5256303 TC |
2301 | static void |
2302 | dmu_buf_will_dirty_impl(dmu_buf_t *db_fake, int flags, dmu_tx_t *tx) | |
34dc7c2f | 2303 | { |
b0bc7a84 | 2304 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; |
34dc7c2f BB |
2305 | |
2306 | ASSERT(tx->tx_txg != 0); | |
424fd7c3 | 2307 | ASSERT(!zfs_refcount_is_zero(&db->db_holds)); |
34dc7c2f | 2308 | |
5a28a973 | 2309 | /* |
e1cfd73f | 2310 | * Quick check for dirtiness. For already dirty blocks, this |
5a28a973 MA |
2311 | * reduces runtime of this function by >90%, and overall performance |
2312 | * by 50% for some workloads (e.g. file deletion with indirect blocks | |
2313 | * cached). | |
2314 | */ | |
2315 | mutex_enter(&db->db_mtx); | |
2316 | ||
cccbed9f MM |
2317 | if (db->db_state == DB_CACHED) { |
2318 | dbuf_dirty_record_t *dr = dbuf_find_dirty_eq(db, tx->tx_txg); | |
5a28a973 MA |
2319 | /* |
2320 | * It's possible that it is already dirty but not cached, | |
2321 | * because there are some calls to dbuf_dirty() that don't | |
2322 | * go through dmu_buf_will_dirty(). | |
2323 | */ | |
cccbed9f | 2324 | if (dr != NULL) { |
5a28a973 MA |
2325 | /* This dbuf is already dirty and cached. */ |
2326 | dbuf_redirty(dr); | |
2327 | mutex_exit(&db->db_mtx); | |
2328 | return; | |
2329 | } | |
2330 | } | |
2331 | mutex_exit(&db->db_mtx); | |
2332 | ||
572e2857 BB |
2333 | DB_DNODE_ENTER(db); |
2334 | if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock)) | |
b5256303 | 2335 | flags |= DB_RF_HAVESTRUCT; |
572e2857 | 2336 | DB_DNODE_EXIT(db); |
b5256303 | 2337 | (void) dbuf_read(db, NULL, flags); |
34dc7c2f BB |
2338 | (void) dbuf_dirty(db, tx); |
2339 | } | |
2340 | ||
b5256303 TC |
2341 | void |
2342 | dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx) | |
2343 | { | |
2344 | dmu_buf_will_dirty_impl(db_fake, | |
2345 | DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH, tx); | |
2346 | } | |
2347 | ||
a73e8fdb PZ |
2348 | boolean_t |
2349 | dmu_buf_is_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx) | |
2350 | { | |
2351 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
cccbed9f | 2352 | dbuf_dirty_record_t *dr; |
a73e8fdb PZ |
2353 | |
2354 | mutex_enter(&db->db_mtx); | |
cccbed9f | 2355 | dr = dbuf_find_dirty_eq(db, tx->tx_txg); |
a73e8fdb | 2356 | mutex_exit(&db->db_mtx); |
cccbed9f | 2357 | return (dr != NULL); |
a73e8fdb PZ |
2358 | } |
2359 | ||
b128c09f BB |
2360 | void |
2361 | dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) | |
2362 | { | |
2363 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
2364 | ||
2365 | db->db_state = DB_NOFILL; | |
2366 | ||
2367 | dmu_buf_will_fill(db_fake, tx); | |
2368 | } | |
2369 | ||
34dc7c2f BB |
2370 | void |
2371 | dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) | |
2372 | { | |
2373 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
2374 | ||
428870ff | 2375 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f BB |
2376 | ASSERT(tx->tx_txg != 0); |
2377 | ASSERT(db->db_level == 0); | |
424fd7c3 | 2378 | ASSERT(!zfs_refcount_is_zero(&db->db_holds)); |
34dc7c2f BB |
2379 | |
2380 | ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT || | |
2381 | dmu_tx_private_ok(tx)); | |
2382 | ||
2383 | dbuf_noread(db); | |
2384 | (void) dbuf_dirty(db, tx); | |
2385 | } | |
2386 | ||
b5256303 TC |
2387 | /* |
2388 | * This function is effectively the same as dmu_buf_will_dirty(), but | |
0c03d21a MA |
2389 | * indicates the caller expects raw encrypted data in the db, and provides |
2390 | * the crypt params (byteorder, salt, iv, mac) which should be stored in the | |
2391 | * blkptr_t when this dbuf is written. This is only used for blocks of | |
2392 | * dnodes, during raw receive. | |
b5256303 TC |
2393 | */ |
2394 | void | |
0c03d21a MA |
2395 | dmu_buf_set_crypt_params(dmu_buf_t *db_fake, boolean_t byteorder, |
2396 | const uint8_t *salt, const uint8_t *iv, const uint8_t *mac, dmu_tx_t *tx) | |
b5256303 TC |
2397 | { |
2398 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
2399 | dbuf_dirty_record_t *dr; | |
2400 | ||
0c03d21a MA |
2401 | /* |
2402 | * dr_has_raw_params is only processed for blocks of dnodes | |
2403 | * (see dbuf_sync_dnode_leaf_crypt()). | |
2404 | */ | |
2405 | ASSERT3U(db->db.db_object, ==, DMU_META_DNODE_OBJECT); | |
2406 | ASSERT3U(db->db_level, ==, 0); | |
2407 | ASSERT(db->db_objset->os_raw_receive); | |
2408 | ||
b5256303 TC |
2409 | dmu_buf_will_dirty_impl(db_fake, |
2410 | DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_NO_DECRYPT, tx); | |
2411 | ||
cccbed9f | 2412 | dr = dbuf_find_dirty_eq(db, tx->tx_txg); |
b5256303 TC |
2413 | |
2414 | ASSERT3P(dr, !=, NULL); | |
0c03d21a MA |
2415 | |
2416 | dr->dt.dl.dr_has_raw_params = B_TRUE; | |
2417 | dr->dt.dl.dr_byteorder = byteorder; | |
2418 | bcopy(salt, dr->dt.dl.dr_salt, ZIO_DATA_SALT_LEN); | |
2419 | bcopy(iv, dr->dt.dl.dr_iv, ZIO_DATA_IV_LEN); | |
2420 | bcopy(mac, dr->dt.dl.dr_mac, ZIO_DATA_MAC_LEN); | |
b5256303 TC |
2421 | } |
2422 | ||
30af21b0 PD |
2423 | static void |
2424 | dbuf_override_impl(dmu_buf_impl_t *db, const blkptr_t *bp, dmu_tx_t *tx) | |
2425 | { | |
2426 | struct dirty_leaf *dl; | |
cccbed9f | 2427 | dbuf_dirty_record_t *dr; |
30af21b0 | 2428 | |
cccbed9f MM |
2429 | dr = list_head(&db->db_dirty_records); |
2430 | ASSERT3U(dr->dr_txg, ==, tx->tx_txg); | |
2431 | dl = &dr->dt.dl; | |
30af21b0 PD |
2432 | dl->dr_overridden_by = *bp; |
2433 | dl->dr_override_state = DR_OVERRIDDEN; | |
cccbed9f | 2434 | dl->dr_overridden_by.blk_birth = dr->dr_txg; |
30af21b0 PD |
2435 | } |
2436 | ||
34dc7c2f BB |
2437 | /* ARGSUSED */ |
2438 | void | |
30af21b0 | 2439 | dmu_buf_fill_done(dmu_buf_t *dbuf, dmu_tx_t *tx) |
34dc7c2f | 2440 | { |
30af21b0 | 2441 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; |
34dc7c2f BB |
2442 | mutex_enter(&db->db_mtx); |
2443 | DBUF_VERIFY(db); | |
2444 | ||
2445 | if (db->db_state == DB_FILL) { | |
2446 | if (db->db_level == 0 && db->db_freed_in_flight) { | |
428870ff | 2447 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f BB |
2448 | /* we were freed while filling */ |
2449 | /* XXX dbuf_undirty? */ | |
2450 | bzero(db->db.db_data, db->db.db_size); | |
2451 | db->db_freed_in_flight = FALSE; | |
2452 | } | |
2453 | db->db_state = DB_CACHED; | |
2454 | cv_broadcast(&db->db_changed); | |
2455 | } | |
2456 | mutex_exit(&db->db_mtx); | |
2457 | } | |
2458 | ||
9b67f605 MA |
2459 | void |
2460 | dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data, | |
2461 | bp_embedded_type_t etype, enum zio_compress comp, | |
2462 | int uncompressed_size, int compressed_size, int byteorder, | |
2463 | dmu_tx_t *tx) | |
2464 | { | |
2465 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; | |
2466 | struct dirty_leaf *dl; | |
2467 | dmu_object_type_t type; | |
cccbed9f | 2468 | dbuf_dirty_record_t *dr; |
9b67f605 | 2469 | |
241b5415 MA |
2470 | if (etype == BP_EMBEDDED_TYPE_DATA) { |
2471 | ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset), | |
2472 | SPA_FEATURE_EMBEDDED_DATA)); | |
2473 | } | |
2474 | ||
9b67f605 MA |
2475 | DB_DNODE_ENTER(db); |
2476 | type = DB_DNODE(db)->dn_type; | |
2477 | DB_DNODE_EXIT(db); | |
2478 | ||
2479 | ASSERT0(db->db_level); | |
2480 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); | |
2481 | ||
2482 | dmu_buf_will_not_fill(dbuf, tx); | |
2483 | ||
cccbed9f MM |
2484 | dr = list_head(&db->db_dirty_records); |
2485 | ASSERT3U(dr->dr_txg, ==, tx->tx_txg); | |
2486 | dl = &dr->dt.dl; | |
9b67f605 MA |
2487 | encode_embedded_bp_compressed(&dl->dr_overridden_by, |
2488 | data, comp, uncompressed_size, compressed_size); | |
2489 | BPE_SET_ETYPE(&dl->dr_overridden_by, etype); | |
2490 | BP_SET_TYPE(&dl->dr_overridden_by, type); | |
2491 | BP_SET_LEVEL(&dl->dr_overridden_by, 0); | |
2492 | BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder); | |
2493 | ||
2494 | dl->dr_override_state = DR_OVERRIDDEN; | |
cccbed9f | 2495 | dl->dr_overridden_by.blk_birth = dr->dr_txg; |
9b67f605 MA |
2496 | } |
2497 | ||
30af21b0 PD |
2498 | void |
2499 | dmu_buf_redact(dmu_buf_t *dbuf, dmu_tx_t *tx) | |
2500 | { | |
2501 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; | |
2502 | dmu_object_type_t type; | |
2503 | ASSERT(dsl_dataset_feature_is_active(db->db_objset->os_dsl_dataset, | |
2504 | SPA_FEATURE_REDACTED_DATASETS)); | |
2505 | ||
2506 | DB_DNODE_ENTER(db); | |
2507 | type = DB_DNODE(db)->dn_type; | |
2508 | DB_DNODE_EXIT(db); | |
2509 | ||
2510 | ASSERT0(db->db_level); | |
2511 | dmu_buf_will_not_fill(dbuf, tx); | |
2512 | ||
2513 | blkptr_t bp = { { { {0} } } }; | |
2514 | BP_SET_TYPE(&bp, type); | |
2515 | BP_SET_LEVEL(&bp, 0); | |
2516 | BP_SET_BIRTH(&bp, tx->tx_txg, 0); | |
2517 | BP_SET_REDACTED(&bp); | |
2518 | BPE_SET_LSIZE(&bp, dbuf->db_size); | |
2519 | ||
2520 | dbuf_override_impl(db, &bp, tx); | |
2521 | } | |
2522 | ||
9babb374 BB |
2523 | /* |
2524 | * Directly assign a provided arc buf to a given dbuf if it's not referenced | |
2525 | * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf. | |
2526 | */ | |
2527 | void | |
2528 | dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx) | |
2529 | { | |
424fd7c3 | 2530 | ASSERT(!zfs_refcount_is_zero(&db->db_holds)); |
428870ff | 2531 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
9babb374 | 2532 | ASSERT(db->db_level == 0); |
2aa34383 | 2533 | ASSERT3U(dbuf_is_metadata(db), ==, arc_is_metadata(buf)); |
9babb374 | 2534 | ASSERT(buf != NULL); |
caf9dd20 | 2535 | ASSERT3U(arc_buf_lsize(buf), ==, db->db.db_size); |
9babb374 BB |
2536 | ASSERT(tx->tx_txg != 0); |
2537 | ||
2538 | arc_return_buf(buf, db); | |
2539 | ASSERT(arc_released(buf)); | |
2540 | ||
2541 | mutex_enter(&db->db_mtx); | |
2542 | ||
2543 | while (db->db_state == DB_READ || db->db_state == DB_FILL) | |
2544 | cv_wait(&db->db_changed, &db->db_mtx); | |
2545 | ||
2546 | ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED); | |
2547 | ||
2548 | if (db->db_state == DB_CACHED && | |
424fd7c3 | 2549 | zfs_refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) { |
440a3eb9 TC |
2550 | /* |
2551 | * In practice, we will never have a case where we have an | |
2552 | * encrypted arc buffer while additional holds exist on the | |
2553 | * dbuf. We don't handle this here so we simply assert that | |
2554 | * fact instead. | |
2555 | */ | |
2556 | ASSERT(!arc_is_encrypted(buf)); | |
9babb374 BB |
2557 | mutex_exit(&db->db_mtx); |
2558 | (void) dbuf_dirty(db, tx); | |
2559 | bcopy(buf->b_data, db->db.db_data, db->db.db_size); | |
d3c2ae1c | 2560 | arc_buf_destroy(buf, db); |
428870ff | 2561 | xuio_stat_wbuf_copied(); |
9babb374 BB |
2562 | return; |
2563 | } | |
2564 | ||
428870ff | 2565 | xuio_stat_wbuf_nocopy(); |
9babb374 | 2566 | if (db->db_state == DB_CACHED) { |
cccbed9f | 2567 | dbuf_dirty_record_t *dr = list_head(&db->db_dirty_records); |
9babb374 BB |
2568 | |
2569 | ASSERT(db->db_buf != NULL); | |
2570 | if (dr != NULL && dr->dr_txg == tx->tx_txg) { | |
2571 | ASSERT(dr->dt.dl.dr_data == db->db_buf); | |
440a3eb9 | 2572 | |
9babb374 BB |
2573 | if (!arc_released(db->db_buf)) { |
2574 | ASSERT(dr->dt.dl.dr_override_state == | |
2575 | DR_OVERRIDDEN); | |
2576 | arc_release(db->db_buf, db); | |
2577 | } | |
2578 | dr->dt.dl.dr_data = buf; | |
d3c2ae1c | 2579 | arc_buf_destroy(db->db_buf, db); |
9babb374 BB |
2580 | } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) { |
2581 | arc_release(db->db_buf, db); | |
d3c2ae1c | 2582 | arc_buf_destroy(db->db_buf, db); |
9babb374 BB |
2583 | } |
2584 | db->db_buf = NULL; | |
2585 | } | |
2586 | ASSERT(db->db_buf == NULL); | |
2587 | dbuf_set_data(db, buf); | |
2588 | db->db_state = DB_FILL; | |
2589 | mutex_exit(&db->db_mtx); | |
2590 | (void) dbuf_dirty(db, tx); | |
b0bc7a84 | 2591 | dmu_buf_fill_done(&db->db, tx); |
9babb374 BB |
2592 | } |
2593 | ||
34dc7c2f | 2594 | void |
d3c2ae1c | 2595 | dbuf_destroy(dmu_buf_impl_t *db) |
34dc7c2f | 2596 | { |
572e2857 | 2597 | dnode_t *dn; |
34dc7c2f | 2598 | dmu_buf_impl_t *parent = db->db_parent; |
572e2857 | 2599 | dmu_buf_impl_t *dndb; |
34dc7c2f BB |
2600 | |
2601 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
424fd7c3 | 2602 | ASSERT(zfs_refcount_is_zero(&db->db_holds)); |
34dc7c2f | 2603 | |
d3c2ae1c GW |
2604 | if (db->db_buf != NULL) { |
2605 | arc_buf_destroy(db->db_buf, db); | |
2606 | db->db_buf = NULL; | |
2607 | } | |
34dc7c2f | 2608 | |
d3c2ae1c GW |
2609 | if (db->db_blkid == DMU_BONUS_BLKID) { |
2610 | int slots = DB_DNODE(db)->dn_num_slots; | |
2611 | int bonuslen = DN_SLOTS_TO_BONUSLEN(slots); | |
b5256303 TC |
2612 | if (db->db.db_data != NULL) { |
2613 | kmem_free(db->db.db_data, bonuslen); | |
2614 | arc_space_return(bonuslen, ARC_SPACE_BONUS); | |
2615 | db->db_state = DB_UNCACHED; | |
2616 | } | |
34dc7c2f BB |
2617 | } |
2618 | ||
d3c2ae1c GW |
2619 | dbuf_clear_data(db); |
2620 | ||
2621 | if (multilist_link_active(&db->db_cache_link)) { | |
2e5dc449 MA |
2622 | ASSERT(db->db_caching_status == DB_DBUF_CACHE || |
2623 | db->db_caching_status == DB_DBUF_METADATA_CACHE); | |
2624 | ||
2625 | multilist_remove(dbuf_caches[db->db_caching_status].cache, db); | |
424fd7c3 | 2626 | (void) zfs_refcount_remove_many( |
2e5dc449 | 2627 | &dbuf_caches[db->db_caching_status].size, |
d3c2ae1c | 2628 | db->db.db_size, db); |
2e5dc449 MA |
2629 | |
2630 | if (db->db_caching_status == DB_DBUF_METADATA_CACHE) { | |
2631 | DBUF_STAT_BUMPDOWN(metadata_cache_count); | |
2632 | } else { | |
2633 | DBUF_STAT_BUMPDOWN(cache_levels[db->db_level]); | |
2634 | DBUF_STAT_BUMPDOWN(cache_count); | |
2635 | DBUF_STAT_DECR(cache_levels_bytes[db->db_level], | |
2636 | db->db.db_size); | |
2637 | } | |
2638 | db->db_caching_status = DB_NO_CACHE; | |
d3c2ae1c GW |
2639 | } |
2640 | ||
b128c09f | 2641 | ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL); |
34dc7c2f BB |
2642 | ASSERT(db->db_data_pending == NULL); |
2643 | ||
2644 | db->db_state = DB_EVICTING; | |
2645 | db->db_blkptr = NULL; | |
2646 | ||
d3c2ae1c GW |
2647 | /* |
2648 | * Now that db_state is DB_EVICTING, nobody else can find this via | |
2649 | * the hash table. We can now drop db_mtx, which allows us to | |
2650 | * acquire the dn_dbufs_mtx. | |
2651 | */ | |
2652 | mutex_exit(&db->db_mtx); | |
2653 | ||
572e2857 BB |
2654 | DB_DNODE_ENTER(db); |
2655 | dn = DB_DNODE(db); | |
2656 | dndb = dn->dn_dbuf; | |
d3c2ae1c GW |
2657 | if (db->db_blkid != DMU_BONUS_BLKID) { |
2658 | boolean_t needlock = !MUTEX_HELD(&dn->dn_dbufs_mtx); | |
2659 | if (needlock) | |
a649768a JD |
2660 | mutex_enter_nested(&dn->dn_dbufs_mtx, |
2661 | NESTED_SINGLE); | |
8951cb8d | 2662 | avl_remove(&dn->dn_dbufs, db); |
73ad4a9f | 2663 | atomic_dec_32(&dn->dn_dbufs_count); |
572e2857 BB |
2664 | membar_producer(); |
2665 | DB_DNODE_EXIT(db); | |
d3c2ae1c GW |
2666 | if (needlock) |
2667 | mutex_exit(&dn->dn_dbufs_mtx); | |
572e2857 BB |
2668 | /* |
2669 | * Decrementing the dbuf count means that the hold corresponding | |
2670 | * to the removed dbuf is no longer discounted in dnode_move(), | |
2671 | * so the dnode cannot be moved until after we release the hold. | |
2672 | * The membar_producer() ensures visibility of the decremented | |
2673 | * value in dnode_move(), since DB_DNODE_EXIT doesn't actually | |
2674 | * release any lock. | |
2675 | */ | |
1fac63e5 | 2676 | mutex_enter(&dn->dn_mtx); |
3d503a76 | 2677 | dnode_rele_and_unlock(dn, db, B_TRUE); |
572e2857 | 2678 | db->db_dnode_handle = NULL; |
d3c2ae1c GW |
2679 | |
2680 | dbuf_hash_remove(db); | |
572e2857 BB |
2681 | } else { |
2682 | DB_DNODE_EXIT(db); | |
34dc7c2f BB |
2683 | } |
2684 | ||
424fd7c3 | 2685 | ASSERT(zfs_refcount_is_zero(&db->db_holds)); |
34dc7c2f | 2686 | |
d3c2ae1c GW |
2687 | db->db_parent = NULL; |
2688 | ||
2689 | ASSERT(db->db_buf == NULL); | |
2690 | ASSERT(db->db.db_data == NULL); | |
2691 | ASSERT(db->db_hash_next == NULL); | |
2692 | ASSERT(db->db_blkptr == NULL); | |
2693 | ASSERT(db->db_data_pending == NULL); | |
2e5dc449 | 2694 | ASSERT3U(db->db_caching_status, ==, DB_NO_CACHE); |
d3c2ae1c GW |
2695 | ASSERT(!multilist_link_active(&db->db_cache_link)); |
2696 | ||
2697 | kmem_cache_free(dbuf_kmem_cache, db); | |
2698 | arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF); | |
34dc7c2f BB |
2699 | |
2700 | /* | |
572e2857 | 2701 | * If this dbuf is referenced from an indirect dbuf, |
34dc7c2f BB |
2702 | * decrement the ref count on the indirect dbuf. |
2703 | */ | |
1fac63e5 MA |
2704 | if (parent && parent != dndb) { |
2705 | mutex_enter(&parent->db_mtx); | |
3d503a76 | 2706 | dbuf_rele_and_unlock(parent, db, B_TRUE); |
1fac63e5 | 2707 | } |
34dc7c2f BB |
2708 | } |
2709 | ||
fcff0f35 PD |
2710 | /* |
2711 | * Note: While bpp will always be updated if the function returns success, | |
2712 | * parentp will not be updated if the dnode does not have dn_dbuf filled in; | |
9c5167d1 | 2713 | * this happens when the dnode is the meta-dnode, or {user|group|project}used |
fcff0f35 PD |
2714 | * object. |
2715 | */ | |
bf701a83 BB |
2716 | __attribute__((always_inline)) |
2717 | static inline int | |
34dc7c2f | 2718 | dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse, |
adb726eb | 2719 | dmu_buf_impl_t **parentp, blkptr_t **bpp) |
34dc7c2f | 2720 | { |
34dc7c2f BB |
2721 | *parentp = NULL; |
2722 | *bpp = NULL; | |
2723 | ||
428870ff BB |
2724 | ASSERT(blkid != DMU_BONUS_BLKID); |
2725 | ||
2726 | if (blkid == DMU_SPILL_BLKID) { | |
2727 | mutex_enter(&dn->dn_mtx); | |
2728 | if (dn->dn_have_spill && | |
2729 | (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) | |
50c957f7 | 2730 | *bpp = DN_SPILL_BLKPTR(dn->dn_phys); |
428870ff BB |
2731 | else |
2732 | *bpp = NULL; | |
2733 | dbuf_add_ref(dn->dn_dbuf, NULL); | |
2734 | *parentp = dn->dn_dbuf; | |
2735 | mutex_exit(&dn->dn_mtx); | |
2736 | return (0); | |
2737 | } | |
34dc7c2f | 2738 | |
1c27024e | 2739 | int nlevels = |
32d41fb7 | 2740 | (dn->dn_phys->dn_nlevels == 0) ? 1 : dn->dn_phys->dn_nlevels; |
1c27024e | 2741 | int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; |
34dc7c2f BB |
2742 | |
2743 | ASSERT3U(level * epbs, <, 64); | |
2744 | ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); | |
32d41fb7 PD |
2745 | /* |
2746 | * This assertion shouldn't trip as long as the max indirect block size | |
2747 | * is less than 1M. The reason for this is that up to that point, | |
2748 | * the number of levels required to address an entire object with blocks | |
2749 | * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In | |
2750 | * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55 | |
2751 | * (i.e. we can address the entire object), objects will all use at most | |
2752 | * N-1 levels and the assertion won't overflow. However, once epbs is | |
2753 | * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be | |
2754 | * enough to address an entire object, so objects will have 5 levels, | |
2755 | * but then this assertion will overflow. | |
2756 | * | |
2757 | * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we | |
2758 | * need to redo this logic to handle overflows. | |
2759 | */ | |
2760 | ASSERT(level >= nlevels || | |
2761 | ((nlevels - level - 1) * epbs) + | |
2762 | highbit64(dn->dn_phys->dn_nblkptr) <= 64); | |
34dc7c2f | 2763 | if (level >= nlevels || |
32d41fb7 PD |
2764 | blkid >= ((uint64_t)dn->dn_phys->dn_nblkptr << |
2765 | ((nlevels - level - 1) * epbs)) || | |
2766 | (fail_sparse && | |
2767 | blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) { | |
34dc7c2f | 2768 | /* the buffer has no parent yet */ |
2e528b49 | 2769 | return (SET_ERROR(ENOENT)); |
34dc7c2f BB |
2770 | } else if (level < nlevels-1) { |
2771 | /* this block is referenced from an indirect block */ | |
fc5bb51f | 2772 | int err; |
64b6c47d TN |
2773 | |
2774 | err = dbuf_hold_impl(dn, level + 1, | |
adb726eb | 2775 | blkid >> epbs, fail_sparse, FALSE, NULL, parentp); |
64b6c47d | 2776 | |
34dc7c2f BB |
2777 | if (err) |
2778 | return (err); | |
2779 | err = dbuf_read(*parentp, NULL, | |
2780 | (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL)); | |
2781 | if (err) { | |
2782 | dbuf_rele(*parentp, NULL); | |
2783 | *parentp = NULL; | |
2784 | return (err); | |
2785 | } | |
f664f1ee | 2786 | rw_enter(&(*parentp)->db_rwlock, RW_READER); |
34dc7c2f BB |
2787 | *bpp = ((blkptr_t *)(*parentp)->db.db_data) + |
2788 | (blkid & ((1ULL << epbs) - 1)); | |
32d41fb7 PD |
2789 | if (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs))) |
2790 | ASSERT(BP_IS_HOLE(*bpp)); | |
f664f1ee | 2791 | rw_exit(&(*parentp)->db_rwlock); |
34dc7c2f BB |
2792 | return (0); |
2793 | } else { | |
2794 | /* the block is referenced from the dnode */ | |
2795 | ASSERT3U(level, ==, nlevels-1); | |
2796 | ASSERT(dn->dn_phys->dn_nblkptr == 0 || | |
2797 | blkid < dn->dn_phys->dn_nblkptr); | |
2798 | if (dn->dn_dbuf) { | |
2799 | dbuf_add_ref(dn->dn_dbuf, NULL); | |
2800 | *parentp = dn->dn_dbuf; | |
2801 | } | |
2802 | *bpp = &dn->dn_phys->dn_blkptr[blkid]; | |
2803 | return (0); | |
2804 | } | |
2805 | } | |
2806 | ||
2807 | static dmu_buf_impl_t * | |
2808 | dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid, | |
2809 | dmu_buf_impl_t *parent, blkptr_t *blkptr) | |
2810 | { | |
428870ff | 2811 | objset_t *os = dn->dn_objset; |
34dc7c2f BB |
2812 | dmu_buf_impl_t *db, *odb; |
2813 | ||
2814 | ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); | |
2815 | ASSERT(dn->dn_type != DMU_OT_NONE); | |
2816 | ||
d3c2ae1c | 2817 | db = kmem_cache_alloc(dbuf_kmem_cache, KM_SLEEP); |
34dc7c2f | 2818 | |
cccbed9f MM |
2819 | list_create(&db->db_dirty_records, sizeof (dbuf_dirty_record_t), |
2820 | offsetof(dbuf_dirty_record_t, dr_dbuf_node)); | |
2821 | ||
34dc7c2f BB |
2822 | db->db_objset = os; |
2823 | db->db.db_object = dn->dn_object; | |
2824 | db->db_level = level; | |
2825 | db->db_blkid = blkid; | |
34dc7c2f | 2826 | db->db_dirtycnt = 0; |
572e2857 | 2827 | db->db_dnode_handle = dn->dn_handle; |
34dc7c2f BB |
2828 | db->db_parent = parent; |
2829 | db->db_blkptr = blkptr; | |
2830 | ||
0c66c32d | 2831 | db->db_user = NULL; |
bc4501f7 JG |
2832 | db->db_user_immediate_evict = FALSE; |
2833 | db->db_freed_in_flight = FALSE; | |
2834 | db->db_pending_evict = FALSE; | |
34dc7c2f | 2835 | |
428870ff | 2836 | if (blkid == DMU_BONUS_BLKID) { |
34dc7c2f | 2837 | ASSERT3P(parent, ==, dn->dn_dbuf); |
50c957f7 | 2838 | db->db.db_size = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) - |
34dc7c2f BB |
2839 | (dn->dn_nblkptr-1) * sizeof (blkptr_t); |
2840 | ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); | |
428870ff | 2841 | db->db.db_offset = DMU_BONUS_BLKID; |
34dc7c2f | 2842 | db->db_state = DB_UNCACHED; |
2e5dc449 | 2843 | db->db_caching_status = DB_NO_CACHE; |
34dc7c2f | 2844 | /* the bonus dbuf is not placed in the hash table */ |
25458cbe | 2845 | arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF); |
34dc7c2f | 2846 | return (db); |
428870ff BB |
2847 | } else if (blkid == DMU_SPILL_BLKID) { |
2848 | db->db.db_size = (blkptr != NULL) ? | |
2849 | BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE; | |
2850 | db->db.db_offset = 0; | |
34dc7c2f BB |
2851 | } else { |
2852 | int blocksize = | |
e8b96c60 | 2853 | db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz; |
34dc7c2f BB |
2854 | db->db.db_size = blocksize; |
2855 | db->db.db_offset = db->db_blkid * blocksize; | |
2856 | } | |
2857 | ||
2858 | /* | |
2859 | * Hold the dn_dbufs_mtx while we get the new dbuf | |
2860 | * in the hash table *and* added to the dbufs list. | |
2861 | * This prevents a possible deadlock with someone | |
e1cfd73f | 2862 | * trying to look up this dbuf before it's added to the |
34dc7c2f BB |
2863 | * dn_dbufs list. |
2864 | */ | |
2865 | mutex_enter(&dn->dn_dbufs_mtx); | |
2866 | db->db_state = DB_EVICTING; | |
2867 | if ((odb = dbuf_hash_insert(db)) != NULL) { | |
2868 | /* someone else inserted it first */ | |
d3c2ae1c | 2869 | kmem_cache_free(dbuf_kmem_cache, db); |
34dc7c2f | 2870 | mutex_exit(&dn->dn_dbufs_mtx); |
5e021f56 | 2871 | DBUF_STAT_BUMP(hash_insert_race); |
34dc7c2f BB |
2872 | return (odb); |
2873 | } | |
8951cb8d | 2874 | avl_add(&dn->dn_dbufs, db); |
9c9531cb | 2875 | |
34dc7c2f | 2876 | db->db_state = DB_UNCACHED; |
2e5dc449 | 2877 | db->db_caching_status = DB_NO_CACHE; |
34dc7c2f | 2878 | mutex_exit(&dn->dn_dbufs_mtx); |
25458cbe | 2879 | arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF); |
34dc7c2f BB |
2880 | |
2881 | if (parent && parent != dn->dn_dbuf) | |
2882 | dbuf_add_ref(parent, db); | |
2883 | ||
2884 | ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || | |
424fd7c3 | 2885 | zfs_refcount_count(&dn->dn_holds) > 0); |
c13060e4 | 2886 | (void) zfs_refcount_add(&dn->dn_holds, db); |
73ad4a9f | 2887 | atomic_inc_32(&dn->dn_dbufs_count); |
34dc7c2f BB |
2888 | |
2889 | dprintf_dbuf(db, "db=%p\n", db); | |
2890 | ||
2891 | return (db); | |
2892 | } | |
2893 | ||
30af21b0 PD |
2894 | /* |
2895 | * This function returns a block pointer and information about the object, | |
2896 | * given a dnode and a block. This is a publicly accessible version of | |
2897 | * dbuf_findbp that only returns some information, rather than the | |
2898 | * dbuf. Note that the dnode passed in must be held, and the dn_struct_rwlock | |
2899 | * should be locked as (at least) a reader. | |
2900 | */ | |
2901 | int | |
2902 | dbuf_dnode_findbp(dnode_t *dn, uint64_t level, uint64_t blkid, | |
2903 | blkptr_t *bp, uint16_t *datablkszsec, uint8_t *indblkshift) | |
2904 | { | |
2905 | dmu_buf_impl_t *dbp = NULL; | |
2906 | blkptr_t *bp2; | |
2907 | int err = 0; | |
2908 | ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); | |
2909 | ||
2910 | err = dbuf_findbp(dn, level, blkid, B_FALSE, &dbp, &bp2); | |
2911 | if (err == 0) { | |
2912 | *bp = *bp2; | |
2913 | if (dbp != NULL) | |
2914 | dbuf_rele(dbp, NULL); | |
2915 | if (datablkszsec != NULL) | |
2916 | *datablkszsec = dn->dn_phys->dn_datablkszsec; | |
2917 | if (indblkshift != NULL) | |
2918 | *indblkshift = dn->dn_phys->dn_indblkshift; | |
2919 | } | |
2920 | ||
2921 | return (err); | |
2922 | } | |
2923 | ||
fcff0f35 PD |
2924 | typedef struct dbuf_prefetch_arg { |
2925 | spa_t *dpa_spa; /* The spa to issue the prefetch in. */ | |
2926 | zbookmark_phys_t dpa_zb; /* The target block to prefetch. */ | |
2927 | int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */ | |
2928 | int dpa_curlevel; /* The current level that we're reading */ | |
d3c2ae1c | 2929 | dnode_t *dpa_dnode; /* The dnode associated with the prefetch */ |
fcff0f35 PD |
2930 | zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */ |
2931 | zio_t *dpa_zio; /* The parent zio_t for all prefetches. */ | |
2932 | arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */ | |
2933 | } dbuf_prefetch_arg_t; | |
2934 | ||
2935 | /* | |
2936 | * Actually issue the prefetch read for the block given. | |
2937 | */ | |
2938 | static void | |
2939 | dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp) | |
2940 | { | |
30af21b0 PD |
2941 | ASSERT(!BP_IS_REDACTED(bp) || |
2942 | dsl_dataset_feature_is_active( | |
2943 | dpa->dpa_dnode->dn_objset->os_dsl_dataset, | |
2944 | SPA_FEATURE_REDACTED_DATASETS)); | |
2945 | ||
2946 | if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp) || BP_IS_REDACTED(bp)) | |
fcff0f35 PD |
2947 | return; |
2948 | ||
4515b1d0 | 2949 | int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE; |
1c27024e DB |
2950 | arc_flags_t aflags = |
2951 | dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH; | |
fcff0f35 | 2952 | |
4515b1d0 TC |
2953 | /* dnodes are always read as raw and then converted later */ |
2954 | if (BP_GET_TYPE(bp) == DMU_OT_DNODE && BP_IS_PROTECTED(bp) && | |
2955 | dpa->dpa_curlevel == 0) | |
2956 | zio_flags |= ZIO_FLAG_RAW; | |
2957 | ||
fcff0f35 PD |
2958 | ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp)); |
2959 | ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level); | |
2960 | ASSERT(dpa->dpa_zio != NULL); | |
2961 | (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL, | |
4515b1d0 | 2962 | dpa->dpa_prio, zio_flags, &aflags, &dpa->dpa_zb); |
fcff0f35 PD |
2963 | } |
2964 | ||
2965 | /* | |
2966 | * Called when an indirect block above our prefetch target is read in. This | |
2967 | * will either read in the next indirect block down the tree or issue the actual | |
2968 | * prefetch if the next block down is our target. | |
2969 | */ | |
2970 | static void | |
d4a72f23 TC |
2971 | dbuf_prefetch_indirect_done(zio_t *zio, const zbookmark_phys_t *zb, |
2972 | const blkptr_t *iobp, arc_buf_t *abuf, void *private) | |
fcff0f35 PD |
2973 | { |
2974 | dbuf_prefetch_arg_t *dpa = private; | |
fcff0f35 PD |
2975 | |
2976 | ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel); | |
2977 | ASSERT3S(dpa->dpa_curlevel, >, 0); | |
d3c2ae1c | 2978 | |
c3bd3fb4 TC |
2979 | if (abuf == NULL) { |
2980 | ASSERT(zio == NULL || zio->io_error != 0); | |
2981 | kmem_free(dpa, sizeof (*dpa)); | |
2982 | return; | |
2983 | } | |
2984 | ASSERT(zio == NULL || zio->io_error == 0); | |
2985 | ||
d3c2ae1c GW |
2986 | /* |
2987 | * The dpa_dnode is only valid if we are called with a NULL | |
2988 | * zio. This indicates that the arc_read() returned without | |
2989 | * first calling zio_read() to issue a physical read. Once | |
2990 | * a physical read is made the dpa_dnode must be invalidated | |
2991 | * as the locks guarding it may have been dropped. If the | |
2992 | * dpa_dnode is still valid, then we want to add it to the dbuf | |
2993 | * cache. To do so, we must hold the dbuf associated with the block | |
2994 | * we just prefetched, read its contents so that we associate it | |
2995 | * with an arc_buf_t, and then release it. | |
2996 | */ | |
fcff0f35 PD |
2997 | if (zio != NULL) { |
2998 | ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel); | |
b5256303 | 2999 | if (zio->io_flags & ZIO_FLAG_RAW_COMPRESS) { |
d3c2ae1c GW |
3000 | ASSERT3U(BP_GET_PSIZE(zio->io_bp), ==, zio->io_size); |
3001 | } else { | |
3002 | ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size); | |
3003 | } | |
fcff0f35 | 3004 | ASSERT3P(zio->io_spa, ==, dpa->dpa_spa); |
d3c2ae1c GW |
3005 | |
3006 | dpa->dpa_dnode = NULL; | |
3007 | } else if (dpa->dpa_dnode != NULL) { | |
3008 | uint64_t curblkid = dpa->dpa_zb.zb_blkid >> | |
3009 | (dpa->dpa_epbs * (dpa->dpa_curlevel - | |
3010 | dpa->dpa_zb.zb_level)); | |
3011 | dmu_buf_impl_t *db = dbuf_hold_level(dpa->dpa_dnode, | |
3012 | dpa->dpa_curlevel, curblkid, FTAG); | |
305781da TC |
3013 | if (db == NULL) { |
3014 | kmem_free(dpa, sizeof (*dpa)); | |
3015 | arc_buf_destroy(abuf, private); | |
3016 | return; | |
3017 | } | |
3018 | ||
d3c2ae1c GW |
3019 | (void) dbuf_read(db, NULL, |
3020 | DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_HAVESTRUCT); | |
3021 | dbuf_rele(db, FTAG); | |
fcff0f35 PD |
3022 | } |
3023 | ||
d4a72f23 | 3024 | dpa->dpa_curlevel--; |
1c27024e | 3025 | uint64_t nextblkid = dpa->dpa_zb.zb_blkid >> |
fcff0f35 | 3026 | (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level)); |
1c27024e | 3027 | blkptr_t *bp = ((blkptr_t *)abuf->b_data) + |
fcff0f35 | 3028 | P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs); |
d4a72f23 | 3029 | |
30af21b0 PD |
3030 | ASSERT(!BP_IS_REDACTED(bp) || |
3031 | dsl_dataset_feature_is_active( | |
3032 | dpa->dpa_dnode->dn_objset->os_dsl_dataset, | |
3033 | SPA_FEATURE_REDACTED_DATASETS)); | |
3034 | if (BP_IS_HOLE(bp) || BP_IS_REDACTED(bp)) { | |
fcff0f35 PD |
3035 | kmem_free(dpa, sizeof (*dpa)); |
3036 | } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) { | |
3037 | ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid); | |
3038 | dbuf_issue_final_prefetch(dpa, bp); | |
3039 | kmem_free(dpa, sizeof (*dpa)); | |
3040 | } else { | |
3041 | arc_flags_t iter_aflags = ARC_FLAG_NOWAIT; | |
3042 | zbookmark_phys_t zb; | |
3043 | ||
7c351e31 | 3044 | /* flag if L2ARC eligible, l2arc_noprefetch then decides */ |
3045 | if (dpa->dpa_aflags & ARC_FLAG_L2CACHE) | |
3046 | iter_aflags |= ARC_FLAG_L2CACHE; | |
3047 | ||
fcff0f35 PD |
3048 | ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp)); |
3049 | ||
3050 | SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset, | |
3051 | dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid); | |
3052 | ||
3053 | (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, | |
3054 | bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio, | |
3055 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, | |
3056 | &iter_aflags, &zb); | |
3057 | } | |
d3c2ae1c GW |
3058 | |
3059 | arc_buf_destroy(abuf, private); | |
fcff0f35 PD |
3060 | } |
3061 | ||
3062 | /* | |
3063 | * Issue prefetch reads for the given block on the given level. If the indirect | |
3064 | * blocks above that block are not in memory, we will read them in | |
3065 | * asynchronously. As a result, this call never blocks waiting for a read to | |
b5256303 TC |
3066 | * complete. Note that the prefetch might fail if the dataset is encrypted and |
3067 | * the encryption key is unmapped before the IO completes. | |
fcff0f35 | 3068 | */ |
34dc7c2f | 3069 | void |
fcff0f35 PD |
3070 | dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio, |
3071 | arc_flags_t aflags) | |
34dc7c2f | 3072 | { |
fcff0f35 PD |
3073 | blkptr_t bp; |
3074 | int epbs, nlevels, curlevel; | |
3075 | uint64_t curblkid; | |
34dc7c2f | 3076 | |
428870ff | 3077 | ASSERT(blkid != DMU_BONUS_BLKID); |
34dc7c2f BB |
3078 | ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); |
3079 | ||
7f60329a MA |
3080 | if (blkid > dn->dn_maxblkid) |
3081 | return; | |
3082 | ||
f664f1ee | 3083 | if (level == 0 && dnode_block_freed(dn, blkid)) |
34dc7c2f BB |
3084 | return; |
3085 | ||
fcff0f35 PD |
3086 | /* |
3087 | * This dnode hasn't been written to disk yet, so there's nothing to | |
3088 | * prefetch. | |
3089 | */ | |
3090 | nlevels = dn->dn_phys->dn_nlevels; | |
3091 | if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0) | |
3092 | return; | |
3093 | ||
3094 | epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; | |
3095 | if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level)) | |
3096 | return; | |
3097 | ||
1c27024e | 3098 | dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object, |
fcff0f35 PD |
3099 | level, blkid); |
3100 | if (db != NULL) { | |
3101 | mutex_exit(&db->db_mtx); | |
572e2857 | 3102 | /* |
fcff0f35 PD |
3103 | * This dbuf already exists. It is either CACHED, or |
3104 | * (we assume) about to be read or filled. | |
572e2857 | 3105 | */ |
572e2857 | 3106 | return; |
34dc7c2f BB |
3107 | } |
3108 | ||
fcff0f35 PD |
3109 | /* |
3110 | * Find the closest ancestor (indirect block) of the target block | |
3111 | * that is present in the cache. In this indirect block, we will | |
3112 | * find the bp that is at curlevel, curblkid. | |
3113 | */ | |
3114 | curlevel = level; | |
3115 | curblkid = blkid; | |
3116 | while (curlevel < nlevels - 1) { | |
3117 | int parent_level = curlevel + 1; | |
3118 | uint64_t parent_blkid = curblkid >> epbs; | |
3119 | dmu_buf_impl_t *db; | |
3120 | ||
3121 | if (dbuf_hold_impl(dn, parent_level, parent_blkid, | |
3122 | FALSE, TRUE, FTAG, &db) == 0) { | |
3123 | blkptr_t *bpp = db->db_buf->b_data; | |
3124 | bp = bpp[P2PHASE(curblkid, 1 << epbs)]; | |
3125 | dbuf_rele(db, FTAG); | |
3126 | break; | |
3127 | } | |
428870ff | 3128 | |
fcff0f35 PD |
3129 | curlevel = parent_level; |
3130 | curblkid = parent_blkid; | |
3131 | } | |
34dc7c2f | 3132 | |
fcff0f35 PD |
3133 | if (curlevel == nlevels - 1) { |
3134 | /* No cached indirect blocks found. */ | |
3135 | ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr); | |
3136 | bp = dn->dn_phys->dn_blkptr[curblkid]; | |
34dc7c2f | 3137 | } |
30af21b0 PD |
3138 | ASSERT(!BP_IS_REDACTED(&bp) || |
3139 | dsl_dataset_feature_is_active(dn->dn_objset->os_dsl_dataset, | |
3140 | SPA_FEATURE_REDACTED_DATASETS)); | |
3141 | if (BP_IS_HOLE(&bp) || BP_IS_REDACTED(&bp)) | |
fcff0f35 PD |
3142 | return; |
3143 | ||
3144 | ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp)); | |
3145 | ||
1c27024e | 3146 | zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL, |
fcff0f35 PD |
3147 | ZIO_FLAG_CANFAIL); |
3148 | ||
1c27024e DB |
3149 | dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP); |
3150 | dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset; | |
fcff0f35 PD |
3151 | SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET, |
3152 | dn->dn_object, level, blkid); | |
3153 | dpa->dpa_curlevel = curlevel; | |
3154 | dpa->dpa_prio = prio; | |
3155 | dpa->dpa_aflags = aflags; | |
3156 | dpa->dpa_spa = dn->dn_objset->os_spa; | |
d3c2ae1c | 3157 | dpa->dpa_dnode = dn; |
fcff0f35 PD |
3158 | dpa->dpa_epbs = epbs; |
3159 | dpa->dpa_zio = pio; | |
3160 | ||
7c351e31 | 3161 | /* flag if L2ARC eligible, l2arc_noprefetch then decides */ |
3162 | if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level)) | |
3163 | dpa->dpa_aflags |= ARC_FLAG_L2CACHE; | |
3164 | ||
fcff0f35 PD |
3165 | /* |
3166 | * If we have the indirect just above us, no need to do the asynchronous | |
3167 | * prefetch chain; we'll just run the last step ourselves. If we're at | |
3168 | * a higher level, though, we want to issue the prefetches for all the | |
3169 | * indirect blocks asynchronously, so we can go on with whatever we were | |
3170 | * doing. | |
3171 | */ | |
3172 | if (curlevel == level) { | |
3173 | ASSERT3U(curblkid, ==, blkid); | |
3174 | dbuf_issue_final_prefetch(dpa, &bp); | |
3175 | kmem_free(dpa, sizeof (*dpa)); | |
3176 | } else { | |
3177 | arc_flags_t iter_aflags = ARC_FLAG_NOWAIT; | |
3178 | zbookmark_phys_t zb; | |
3179 | ||
7c351e31 | 3180 | /* flag if L2ARC eligible, l2arc_noprefetch then decides */ |
3181 | if (DNODE_LEVEL_IS_L2CACHEABLE(dn, level)) | |
3182 | iter_aflags |= ARC_FLAG_L2CACHE; | |
3183 | ||
fcff0f35 PD |
3184 | SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET, |
3185 | dn->dn_object, curlevel, curblkid); | |
3186 | (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, | |
3187 | &bp, dbuf_prefetch_indirect_done, dpa, prio, | |
3188 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, | |
3189 | &iter_aflags, &zb); | |
3190 | } | |
3191 | /* | |
3192 | * We use pio here instead of dpa_zio since it's possible that | |
3193 | * dpa may have already been freed. | |
3194 | */ | |
3195 | zio_nowait(pio); | |
34dc7c2f BB |
3196 | } |
3197 | ||
71a24c3c | 3198 | /* |
64b6c47d | 3199 | * Helper function for dbuf_hold_impl() to copy a buffer. Handles |
71a24c3c TC |
3200 | * the case of encrypted, compressed and uncompressed buffers by |
3201 | * allocating the new buffer, respectively, with arc_alloc_raw_buf(), | |
3202 | * arc_alloc_compressed_buf() or arc_alloc_buf().* | |
3203 | * | |
64b6c47d | 3204 | * NOTE: Declared noinline to avoid stack bloat in dbuf_hold_impl(). |
71a24c3c TC |
3205 | */ |
3206 | noinline static void | |
64b6c47d | 3207 | dbuf_hold_copy(dnode_t *dn, dmu_buf_impl_t *db) |
71a24c3c | 3208 | { |
64b6c47d | 3209 | dbuf_dirty_record_t *dr = db->db_data_pending; |
71a24c3c | 3210 | arc_buf_t *data = dr->dt.dl.dr_data; |
71a24c3c TC |
3211 | enum zio_compress compress_type = arc_get_compression(data); |
3212 | ||
3213 | if (arc_is_encrypted(data)) { | |
3214 | boolean_t byteorder; | |
3215 | uint8_t salt[ZIO_DATA_SALT_LEN]; | |
3216 | uint8_t iv[ZIO_DATA_IV_LEN]; | |
3217 | uint8_t mac[ZIO_DATA_MAC_LEN]; | |
3218 | ||
3219 | arc_get_raw_params(data, &byteorder, salt, iv, mac); | |
3220 | dbuf_set_data(db, arc_alloc_raw_buf(dn->dn_objset->os_spa, db, | |
3221 | dmu_objset_id(dn->dn_objset), byteorder, salt, iv, mac, | |
3222 | dn->dn_type, arc_buf_size(data), arc_buf_lsize(data), | |
3223 | compress_type)); | |
3224 | } else if (compress_type != ZIO_COMPRESS_OFF) { | |
3225 | dbuf_set_data(db, arc_alloc_compressed_buf( | |
3226 | dn->dn_objset->os_spa, db, arc_buf_size(data), | |
3227 | arc_buf_lsize(data), compress_type)); | |
3228 | } else { | |
3229 | dbuf_set_data(db, arc_alloc_buf(dn->dn_objset->os_spa, db, | |
3230 | DBUF_GET_BUFC_TYPE(db), db->db.db_size)); | |
3231 | } | |
3232 | ||
f664f1ee | 3233 | rw_enter(&db->db_rwlock, RW_WRITER); |
71a24c3c | 3234 | bcopy(data->b_data, db->db.db_data, arc_buf_size(data)); |
f664f1ee | 3235 | rw_exit(&db->db_rwlock); |
71a24c3c TC |
3236 | } |
3237 | ||
34dc7c2f BB |
3238 | /* |
3239 | * Returns with db_holds incremented, and db_mtx not held. | |
3240 | * Note: dn_struct_rwlock must be held. | |
3241 | */ | |
64b6c47d TN |
3242 | int |
3243 | dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, | |
3244 | boolean_t fail_sparse, boolean_t fail_uncached, | |
3245 | void *tag, dmu_buf_impl_t **dbp) | |
34dc7c2f | 3246 | { |
64b6c47d | 3247 | dmu_buf_impl_t *db, *parent = NULL; |
d3c2ae1c | 3248 | |
37f03da8 | 3249 | /* If the pool has been created, verify the tx_sync_lock is not held */ |
64b6c47d | 3250 | spa_t *spa = dn->dn_objset->os_spa; |
37f03da8 SH |
3251 | dsl_pool_t *dp = spa->spa_dsl_pool; |
3252 | if (dp != NULL) { | |
3253 | ASSERT(!MUTEX_HELD(&dp->dp_tx.tx_sync_lock)); | |
3254 | } | |
3255 | ||
64b6c47d TN |
3256 | ASSERT(blkid != DMU_BONUS_BLKID); |
3257 | ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); | |
3258 | ASSERT3U(dn->dn_nlevels, >, level); | |
3259 | ||
3260 | *dbp = NULL; | |
3261 | ||
34dc7c2f | 3262 | /* dbuf_find() returns with db_mtx held */ |
64b6c47d | 3263 | db = dbuf_find(dn->dn_objset, dn->dn_object, level, blkid); |
fc5bb51f | 3264 | |
64b6c47d TN |
3265 | if (db == NULL) { |
3266 | blkptr_t *bp = NULL; | |
3267 | int err; | |
fc5bb51f | 3268 | |
64b6c47d | 3269 | if (fail_uncached) |
fcff0f35 PD |
3270 | return (SET_ERROR(ENOENT)); |
3271 | ||
64b6c47d TN |
3272 | ASSERT3P(parent, ==, NULL); |
3273 | err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp); | |
3274 | if (fail_sparse) { | |
3275 | if (err == 0 && bp && BP_IS_HOLE(bp)) | |
3276 | err = SET_ERROR(ENOENT); | |
3277 | if (err) { | |
3278 | if (parent) | |
3279 | dbuf_rele(parent, NULL); | |
3280 | return (err); | |
34dc7c2f BB |
3281 | } |
3282 | } | |
64b6c47d TN |
3283 | if (err && err != ENOENT) |
3284 | return (err); | |
3285 | db = dbuf_create(dn, level, blkid, parent, bp); | |
34dc7c2f BB |
3286 | } |
3287 | ||
64b6c47d TN |
3288 | if (fail_uncached && db->db_state != DB_CACHED) { |
3289 | mutex_exit(&db->db_mtx); | |
fcff0f35 PD |
3290 | return (SET_ERROR(ENOENT)); |
3291 | } | |
3292 | ||
64b6c47d TN |
3293 | if (db->db_buf != NULL) { |
3294 | arc_buf_access(db->db_buf); | |
3295 | ASSERT3P(db->db.db_data, ==, db->db_buf->b_data); | |
0873bb63 | 3296 | } |
34dc7c2f | 3297 | |
64b6c47d | 3298 | ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf)); |
34dc7c2f BB |
3299 | |
3300 | /* | |
e1cfd73f | 3301 | * If this buffer is currently syncing out, and we are |
34dc7c2f BB |
3302 | * still referencing it from db_data, we need to make a copy |
3303 | * of it in case we decide we want to dirty it again in this txg. | |
3304 | */ | |
64b6c47d TN |
3305 | if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && |
3306 | dn->dn_object != DMU_META_DNODE_OBJECT && | |
3307 | db->db_state == DB_CACHED && db->db_data_pending) { | |
3308 | dbuf_dirty_record_t *dr = db->db_data_pending; | |
3309 | if (dr->dt.dl.dr_data == db->db_buf) | |
3310 | dbuf_hold_copy(dn, db); | |
3311 | } | |
3312 | ||
3313 | if (multilist_link_active(&db->db_cache_link)) { | |
3314 | ASSERT(zfs_refcount_is_zero(&db->db_holds)); | |
3315 | ASSERT(db->db_caching_status == DB_DBUF_CACHE || | |
3316 | db->db_caching_status == DB_DBUF_METADATA_CACHE); | |
3317 | ||
3318 | multilist_remove(dbuf_caches[db->db_caching_status].cache, db); | |
424fd7c3 | 3319 | (void) zfs_refcount_remove_many( |
64b6c47d TN |
3320 | &dbuf_caches[db->db_caching_status].size, |
3321 | db->db.db_size, db); | |
2e5dc449 | 3322 | |
64b6c47d | 3323 | if (db->db_caching_status == DB_DBUF_METADATA_CACHE) { |
2e5dc449 MA |
3324 | DBUF_STAT_BUMPDOWN(metadata_cache_count); |
3325 | } else { | |
64b6c47d | 3326 | DBUF_STAT_BUMPDOWN(cache_levels[db->db_level]); |
2e5dc449 | 3327 | DBUF_STAT_BUMPDOWN(cache_count); |
64b6c47d TN |
3328 | DBUF_STAT_DECR(cache_levels_bytes[db->db_level], |
3329 | db->db.db_size); | |
2e5dc449 | 3330 | } |
64b6c47d | 3331 | db->db_caching_status = DB_NO_CACHE; |
d3c2ae1c | 3332 | } |
64b6c47d TN |
3333 | (void) zfs_refcount_add(&db->db_holds, tag); |
3334 | DBUF_VERIFY(db); | |
3335 | mutex_exit(&db->db_mtx); | |
34dc7c2f BB |
3336 | |
3337 | /* NOTE: we can't rele the parent until after we drop the db_mtx */ | |
64b6c47d TN |
3338 | if (parent) |
3339 | dbuf_rele(parent, NULL); | |
34dc7c2f | 3340 | |
64b6c47d TN |
3341 | ASSERT3P(DB_DNODE(db), ==, dn); |
3342 | ASSERT3U(db->db_blkid, ==, blkid); | |
3343 | ASSERT3U(db->db_level, ==, level); | |
3344 | *dbp = db; | |
34dc7c2f BB |
3345 | |
3346 | return (0); | |
3347 | } | |
3348 | ||
3349 | dmu_buf_impl_t * | |
3350 | dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag) | |
3351 | { | |
fcff0f35 | 3352 | return (dbuf_hold_level(dn, 0, blkid, tag)); |
34dc7c2f BB |
3353 | } |
3354 | ||
3355 | dmu_buf_impl_t * | |
3356 | dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag) | |
3357 | { | |
3358 | dmu_buf_impl_t *db; | |
fcff0f35 | 3359 | int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db); |
34dc7c2f BB |
3360 | return (err ? NULL : db); |
3361 | } | |
3362 | ||
3363 | void | |
3364 | dbuf_create_bonus(dnode_t *dn) | |
3365 | { | |
3366 | ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); | |
3367 | ||
3368 | ASSERT(dn->dn_bonus == NULL); | |
428870ff BB |
3369 | dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL); |
3370 | } | |
3371 | ||
3372 | int | |
3373 | dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx) | |
3374 | { | |
3375 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
572e2857 | 3376 | |
428870ff | 3377 | if (db->db_blkid != DMU_SPILL_BLKID) |
2e528b49 | 3378 | return (SET_ERROR(ENOTSUP)); |
428870ff BB |
3379 | if (blksz == 0) |
3380 | blksz = SPA_MINBLOCKSIZE; | |
f1512ee6 MA |
3381 | ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset))); |
3382 | blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE); | |
428870ff | 3383 | |
428870ff | 3384 | dbuf_new_size(db, blksz, tx); |
428870ff BB |
3385 | |
3386 | return (0); | |
3387 | } | |
3388 | ||
3389 | void | |
3390 | dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx) | |
3391 | { | |
3392 | dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx); | |
34dc7c2f BB |
3393 | } |
3394 | ||
3395 | #pragma weak dmu_buf_add_ref = dbuf_add_ref | |
3396 | void | |
3397 | dbuf_add_ref(dmu_buf_impl_t *db, void *tag) | |
3398 | { | |
c13060e4 | 3399 | int64_t holds = zfs_refcount_add(&db->db_holds, tag); |
d3c2ae1c | 3400 | VERIFY3S(holds, >, 1); |
34dc7c2f BB |
3401 | } |
3402 | ||
6ebebace JG |
3403 | #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref |
3404 | boolean_t | |
3405 | dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid, | |
3406 | void *tag) | |
3407 | { | |
3408 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
3409 | dmu_buf_impl_t *found_db; | |
3410 | boolean_t result = B_FALSE; | |
3411 | ||
d617648c | 3412 | if (blkid == DMU_BONUS_BLKID) |
6ebebace JG |
3413 | found_db = dbuf_find_bonus(os, obj); |
3414 | else | |
3415 | found_db = dbuf_find(os, obj, 0, blkid); | |
3416 | ||
3417 | if (found_db != NULL) { | |
3418 | if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) { | |
c13060e4 | 3419 | (void) zfs_refcount_add(&db->db_holds, tag); |
6ebebace JG |
3420 | result = B_TRUE; |
3421 | } | |
d617648c | 3422 | mutex_exit(&found_db->db_mtx); |
6ebebace JG |
3423 | } |
3424 | return (result); | |
3425 | } | |
3426 | ||
572e2857 BB |
3427 | /* |
3428 | * If you call dbuf_rele() you had better not be referencing the dnode handle | |
3429 | * unless you have some other direct or indirect hold on the dnode. (An indirect | |
3430 | * hold is a hold on one of the dnode's dbufs, including the bonus buffer.) | |
3431 | * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the | |
3432 | * dnode's parent dbuf evicting its dnode handles. | |
3433 | */ | |
34dc7c2f BB |
3434 | void |
3435 | dbuf_rele(dmu_buf_impl_t *db, void *tag) | |
428870ff BB |
3436 | { |
3437 | mutex_enter(&db->db_mtx); | |
3d503a76 | 3438 | dbuf_rele_and_unlock(db, tag, B_FALSE); |
428870ff BB |
3439 | } |
3440 | ||
b0bc7a84 MG |
3441 | void |
3442 | dmu_buf_rele(dmu_buf_t *db, void *tag) | |
3443 | { | |
3444 | dbuf_rele((dmu_buf_impl_t *)db, tag); | |
3445 | } | |
3446 | ||
428870ff BB |
3447 | /* |
3448 | * dbuf_rele() for an already-locked dbuf. This is necessary to allow | |
1fac63e5 MA |
3449 | * db_dirtycnt and db_holds to be updated atomically. The 'evicting' |
3450 | * argument should be set if we are already in the dbuf-evicting code | |
3451 | * path, in which case we don't want to recursively evict. This allows us to | |
3452 | * avoid deeply nested stacks that would have a call flow similar to this: | |
3453 | * | |
3454 | * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify() | |
3455 | * ^ | | |
3456 | * | | | |
3457 | * +-----dbuf_destroy()<--dbuf_evict_one()<--------+ | |
3458 | * | |
428870ff BB |
3459 | */ |
3460 | void | |
3d503a76 | 3461 | dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag, boolean_t evicting) |
34dc7c2f BB |
3462 | { |
3463 | int64_t holds; | |
cbd8f5b7 | 3464 | uint64_t size; |
34dc7c2f | 3465 | |
428870ff | 3466 | ASSERT(MUTEX_HELD(&db->db_mtx)); |
34dc7c2f BB |
3467 | DBUF_VERIFY(db); |
3468 | ||
572e2857 BB |
3469 | /* |
3470 | * Remove the reference to the dbuf before removing its hold on the | |
3471 | * dnode so we can guarantee in dnode_move() that a referenced bonus | |
3472 | * buffer has a corresponding dnode hold. | |
3473 | */ | |
424fd7c3 | 3474 | holds = zfs_refcount_remove(&db->db_holds, tag); |
34dc7c2f BB |
3475 | ASSERT(holds >= 0); |
3476 | ||
3477 | /* | |
3478 | * We can't freeze indirects if there is a possibility that they | |
3479 | * may be modified in the current syncing context. | |
3480 | */ | |
d3c2ae1c GW |
3481 | if (db->db_buf != NULL && |
3482 | holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) { | |
34dc7c2f | 3483 | arc_buf_freeze(db->db_buf); |
d3c2ae1c | 3484 | } |
34dc7c2f BB |
3485 | |
3486 | if (holds == db->db_dirtycnt && | |
bc4501f7 | 3487 | db->db_level == 0 && db->db_user_immediate_evict) |
34dc7c2f BB |
3488 | dbuf_evict_user(db); |
3489 | ||
3490 | if (holds == 0) { | |
428870ff | 3491 | if (db->db_blkid == DMU_BONUS_BLKID) { |
4c7b7eed | 3492 | dnode_t *dn; |
bc4501f7 | 3493 | boolean_t evict_dbuf = db->db_pending_evict; |
572e2857 BB |
3494 | |
3495 | /* | |
4c7b7eed JG |
3496 | * If the dnode moves here, we cannot cross this |
3497 | * barrier until the move completes. | |
572e2857 BB |
3498 | */ |
3499 | DB_DNODE_ENTER(db); | |
4c7b7eed JG |
3500 | |
3501 | dn = DB_DNODE(db); | |
3502 | atomic_dec_32(&dn->dn_dbufs_count); | |
3503 | ||
3504 | /* | |
3505 | * Decrementing the dbuf count means that the bonus | |
3506 | * buffer's dnode hold is no longer discounted in | |
3507 | * dnode_move(). The dnode cannot move until after | |
bc4501f7 | 3508 | * the dnode_rele() below. |
4c7b7eed | 3509 | */ |
572e2857 | 3510 | DB_DNODE_EXIT(db); |
4c7b7eed JG |
3511 | |
3512 | /* | |
3513 | * Do not reference db after its lock is dropped. | |
3514 | * Another thread may evict it. | |
3515 | */ | |
3516 | mutex_exit(&db->db_mtx); | |
3517 | ||
bc4501f7 | 3518 | if (evict_dbuf) |
4c7b7eed | 3519 | dnode_evict_bonus(dn); |
bc4501f7 JG |
3520 | |
3521 | dnode_rele(dn, db); | |
34dc7c2f BB |
3522 | } else if (db->db_buf == NULL) { |
3523 | /* | |
3524 | * This is a special case: we never associated this | |
3525 | * dbuf with any data allocated from the ARC. | |
3526 | */ | |
b128c09f BB |
3527 | ASSERT(db->db_state == DB_UNCACHED || |
3528 | db->db_state == DB_NOFILL); | |
d3c2ae1c | 3529 | dbuf_destroy(db); |
34dc7c2f | 3530 | } else if (arc_released(db->db_buf)) { |
34dc7c2f BB |
3531 | /* |
3532 | * This dbuf has anonymous data associated with it. | |
3533 | */ | |
d3c2ae1c | 3534 | dbuf_destroy(db); |
34dc7c2f | 3535 | } else { |
d3c2ae1c GW |
3536 | boolean_t do_arc_evict = B_FALSE; |
3537 | blkptr_t bp; | |
3538 | spa_t *spa = dmu_objset_spa(db->db_objset); | |
3539 | ||
3540 | if (!DBUF_IS_CACHEABLE(db) && | |
3541 | db->db_blkptr != NULL && | |
3542 | !BP_IS_HOLE(db->db_blkptr) && | |
3543 | !BP_IS_EMBEDDED(db->db_blkptr)) { | |
3544 | do_arc_evict = B_TRUE; | |
3545 | bp = *db->db_blkptr; | |
3546 | } | |
1eb5bfa3 | 3547 | |
d3c2ae1c GW |
3548 | if (!DBUF_IS_CACHEABLE(db) || |
3549 | db->db_pending_evict) { | |
3550 | dbuf_destroy(db); | |
3551 | } else if (!multilist_link_active(&db->db_cache_link)) { | |
2e5dc449 MA |
3552 | ASSERT3U(db->db_caching_status, ==, |
3553 | DB_NO_CACHE); | |
3554 | ||
3555 | dbuf_cached_state_t dcs = | |
3556 | dbuf_include_in_metadata_cache(db) ? | |
3557 | DB_DBUF_METADATA_CACHE : DB_DBUF_CACHE; | |
3558 | db->db_caching_status = dcs; | |
3559 | ||
3560 | multilist_insert(dbuf_caches[dcs].cache, db); | |
cbd8f5b7 | 3561 | size = zfs_refcount_add_many( |
424fd7c3 | 3562 | &dbuf_caches[dcs].size, |
d3c2ae1c | 3563 | db->db.db_size, db); |
2e5dc449 MA |
3564 | |
3565 | if (dcs == DB_DBUF_METADATA_CACHE) { | |
3566 | DBUF_STAT_BUMP(metadata_cache_count); | |
3567 | DBUF_STAT_MAX( | |
3568 | metadata_cache_size_bytes_max, | |
cbd8f5b7 | 3569 | size); |
2e5dc449 MA |
3570 | } else { |
3571 | DBUF_STAT_BUMP( | |
3572 | cache_levels[db->db_level]); | |
3573 | DBUF_STAT_BUMP(cache_count); | |
3574 | DBUF_STAT_INCR( | |
3575 | cache_levels_bytes[db->db_level], | |
3576 | db->db.db_size); | |
3577 | DBUF_STAT_MAX(cache_size_bytes_max, | |
cbd8f5b7 | 3578 | size); |
2e5dc449 | 3579 | } |
b128c09f | 3580 | mutex_exit(&db->db_mtx); |
d3c2ae1c | 3581 | |
cbd8f5b7 AM |
3582 | if (dcs == DB_DBUF_CACHE && !evicting) |
3583 | dbuf_evict_notify(size); | |
bd089c54 | 3584 | } |
d3c2ae1c GW |
3585 | |
3586 | if (do_arc_evict) | |
3587 | arc_freed(spa, &bp); | |
34dc7c2f BB |
3588 | } |
3589 | } else { | |
3590 | mutex_exit(&db->db_mtx); | |
3591 | } | |
d3c2ae1c | 3592 | |
34dc7c2f BB |
3593 | } |
3594 | ||
3595 | #pragma weak dmu_buf_refcount = dbuf_refcount | |
3596 | uint64_t | |
3597 | dbuf_refcount(dmu_buf_impl_t *db) | |
3598 | { | |
424fd7c3 | 3599 | return (zfs_refcount_count(&db->db_holds)); |
34dc7c2f BB |
3600 | } |
3601 | ||
cd32e5db TC |
3602 | uint64_t |
3603 | dmu_buf_user_refcount(dmu_buf_t *db_fake) | |
3604 | { | |
3605 | uint64_t holds; | |
3606 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
3607 | ||
3608 | mutex_enter(&db->db_mtx); | |
424fd7c3 TS |
3609 | ASSERT3U(zfs_refcount_count(&db->db_holds), >=, db->db_dirtycnt); |
3610 | holds = zfs_refcount_count(&db->db_holds) - db->db_dirtycnt; | |
cd32e5db TC |
3611 | mutex_exit(&db->db_mtx); |
3612 | ||
3613 | return (holds); | |
3614 | } | |
3615 | ||
34dc7c2f | 3616 | void * |
0c66c32d JG |
3617 | dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user, |
3618 | dmu_buf_user_t *new_user) | |
34dc7c2f | 3619 | { |
0c66c32d JG |
3620 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; |
3621 | ||
3622 | mutex_enter(&db->db_mtx); | |
3623 | dbuf_verify_user(db, DBVU_NOT_EVICTING); | |
3624 | if (db->db_user == old_user) | |
3625 | db->db_user = new_user; | |
3626 | else | |
3627 | old_user = db->db_user; | |
3628 | dbuf_verify_user(db, DBVU_NOT_EVICTING); | |
3629 | mutex_exit(&db->db_mtx); | |
3630 | ||
3631 | return (old_user); | |
34dc7c2f BB |
3632 | } |
3633 | ||
3634 | void * | |
0c66c32d | 3635 | dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user) |
34dc7c2f | 3636 | { |
0c66c32d | 3637 | return (dmu_buf_replace_user(db_fake, NULL, user)); |
34dc7c2f BB |
3638 | } |
3639 | ||
3640 | void * | |
0c66c32d | 3641 | dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user) |
34dc7c2f BB |
3642 | { |
3643 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
34dc7c2f | 3644 | |
bc4501f7 | 3645 | db->db_user_immediate_evict = TRUE; |
0c66c32d JG |
3646 | return (dmu_buf_set_user(db_fake, user)); |
3647 | } | |
34dc7c2f | 3648 | |
0c66c32d JG |
3649 | void * |
3650 | dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user) | |
3651 | { | |
3652 | return (dmu_buf_replace_user(db_fake, user, NULL)); | |
34dc7c2f BB |
3653 | } |
3654 | ||
3655 | void * | |
3656 | dmu_buf_get_user(dmu_buf_t *db_fake) | |
3657 | { | |
3658 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
34dc7c2f | 3659 | |
0c66c32d JG |
3660 | dbuf_verify_user(db, DBVU_NOT_EVICTING); |
3661 | return (db->db_user); | |
3662 | } | |
3663 | ||
3664 | void | |
3665 | dmu_buf_user_evict_wait() | |
3666 | { | |
3667 | taskq_wait(dbu_evict_taskq); | |
34dc7c2f BB |
3668 | } |
3669 | ||
03c6040b GW |
3670 | blkptr_t * |
3671 | dmu_buf_get_blkptr(dmu_buf_t *db) | |
3672 | { | |
3673 | dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; | |
3674 | return (dbi->db_blkptr); | |
3675 | } | |
3676 | ||
8bea9815 MA |
3677 | objset_t * |
3678 | dmu_buf_get_objset(dmu_buf_t *db) | |
3679 | { | |
3680 | dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; | |
3681 | return (dbi->db_objset); | |
3682 | } | |
3683 | ||
2bce8049 MA |
3684 | dnode_t * |
3685 | dmu_buf_dnode_enter(dmu_buf_t *db) | |
3686 | { | |
3687 | dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; | |
3688 | DB_DNODE_ENTER(dbi); | |
3689 | return (DB_DNODE(dbi)); | |
3690 | } | |
3691 | ||
3692 | void | |
3693 | dmu_buf_dnode_exit(dmu_buf_t *db) | |
3694 | { | |
3695 | dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; | |
3696 | DB_DNODE_EXIT(dbi); | |
3697 | } | |
3698 | ||
34dc7c2f BB |
3699 | static void |
3700 | dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db) | |
3701 | { | |
3702 | /* ASSERT(dmu_tx_is_syncing(tx) */ | |
3703 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
3704 | ||
3705 | if (db->db_blkptr != NULL) | |
3706 | return; | |
3707 | ||
428870ff | 3708 | if (db->db_blkid == DMU_SPILL_BLKID) { |
50c957f7 | 3709 | db->db_blkptr = DN_SPILL_BLKPTR(dn->dn_phys); |
428870ff BB |
3710 | BP_ZERO(db->db_blkptr); |
3711 | return; | |
3712 | } | |
34dc7c2f BB |
3713 | if (db->db_level == dn->dn_phys->dn_nlevels-1) { |
3714 | /* | |
3715 | * This buffer was allocated at a time when there was | |
3716 | * no available blkptrs from the dnode, or it was | |
e1cfd73f | 3717 | * inappropriate to hook it in (i.e., nlevels mismatch). |
34dc7c2f BB |
3718 | */ |
3719 | ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr); | |
3720 | ASSERT(db->db_parent == NULL); | |
3721 | db->db_parent = dn->dn_dbuf; | |
3722 | db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid]; | |
3723 | DBUF_VERIFY(db); | |
3724 | } else { | |
3725 | dmu_buf_impl_t *parent = db->db_parent; | |
3726 | int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; | |
3727 | ||
3728 | ASSERT(dn->dn_phys->dn_nlevels > 1); | |
3729 | if (parent == NULL) { | |
3730 | mutex_exit(&db->db_mtx); | |
3731 | rw_enter(&dn->dn_struct_rwlock, RW_READER); | |
fcff0f35 PD |
3732 | parent = dbuf_hold_level(dn, db->db_level + 1, |
3733 | db->db_blkid >> epbs, db); | |
34dc7c2f BB |
3734 | rw_exit(&dn->dn_struct_rwlock); |
3735 | mutex_enter(&db->db_mtx); | |
3736 | db->db_parent = parent; | |
3737 | } | |
3738 | db->db_blkptr = (blkptr_t *)parent->db.db_data + | |
3739 | (db->db_blkid & ((1ULL << epbs) - 1)); | |
3740 | DBUF_VERIFY(db); | |
3741 | } | |
3742 | } | |
3743 | ||
b5256303 | 3744 | /* |
0c03d21a MA |
3745 | * When syncing out a blocks of dnodes, adjust the block to deal with |
3746 | * encryption. Normally, we make sure the block is decrypted before writing | |
3747 | * it. If we have crypt params, then we are writing a raw (encrypted) block, | |
3748 | * from a raw receive. In this case, set the ARC buf's crypt params so | |
3749 | * that the BP will be filled with the correct byteorder, salt, iv, and mac. | |
b5256303 TC |
3750 | */ |
3751 | static void | |
0c03d21a | 3752 | dbuf_prepare_encrypted_dnode_leaf(dbuf_dirty_record_t *dr) |
b5256303 TC |
3753 | { |
3754 | int err; | |
3755 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
3756 | ||
3757 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
0c03d21a MA |
3758 | ASSERT3U(db->db.db_object, ==, DMU_META_DNODE_OBJECT); |
3759 | ASSERT3U(db->db_level, ==, 0); | |
b5256303 | 3760 | |
0c03d21a | 3761 | if (!db->db_objset->os_raw_receive && arc_is_encrypted(db->db_buf)) { |
a2c2ed1b TC |
3762 | zbookmark_phys_t zb; |
3763 | ||
b5256303 TC |
3764 | /* |
3765 | * Unfortunately, there is currently no mechanism for | |
3766 | * syncing context to handle decryption errors. An error | |
3767 | * here is only possible if an attacker maliciously | |
3768 | * changed a dnode block and updated the associated | |
3769 | * checksums going up the block tree. | |
3770 | */ | |
a2c2ed1b TC |
3771 | SET_BOOKMARK(&zb, dmu_objset_id(db->db_objset), |
3772 | db->db.db_object, db->db_level, db->db_blkid); | |
b5256303 | 3773 | err = arc_untransform(db->db_buf, db->db_objset->os_spa, |
a2c2ed1b | 3774 | &zb, B_TRUE); |
b5256303 TC |
3775 | if (err) |
3776 | panic("Invalid dnode block MAC"); | |
0c03d21a MA |
3777 | } else if (dr->dt.dl.dr_has_raw_params) { |
3778 | (void) arc_release(dr->dt.dl.dr_data, db); | |
3779 | arc_convert_to_raw(dr->dt.dl.dr_data, | |
3780 | dmu_objset_id(db->db_objset), | |
3781 | dr->dt.dl.dr_byteorder, DMU_OT_DNODE, | |
3782 | dr->dt.dl.dr_salt, dr->dt.dl.dr_iv, dr->dt.dl.dr_mac); | |
b5256303 TC |
3783 | } |
3784 | } | |
3785 | ||
d1d7e268 MK |
3786 | /* |
3787 | * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it | |
60948de1 BB |
3788 | * is critical the we not allow the compiler to inline this function in to |
3789 | * dbuf_sync_list() thereby drastically bloating the stack usage. | |
3790 | */ | |
3791 | noinline static void | |
34dc7c2f BB |
3792 | dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx) |
3793 | { | |
3794 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
572e2857 | 3795 | dnode_t *dn; |
34dc7c2f BB |
3796 | zio_t *zio; |
3797 | ||
3798 | ASSERT(dmu_tx_is_syncing(tx)); | |
3799 | ||
3800 | dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); | |
3801 | ||
3802 | mutex_enter(&db->db_mtx); | |
3803 | ||
3804 | ASSERT(db->db_level > 0); | |
3805 | DBUF_VERIFY(db); | |
3806 | ||
e49f1e20 | 3807 | /* Read the block if it hasn't been read yet. */ |
34dc7c2f BB |
3808 | if (db->db_buf == NULL) { |
3809 | mutex_exit(&db->db_mtx); | |
3810 | (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED); | |
3811 | mutex_enter(&db->db_mtx); | |
3812 | } | |
3813 | ASSERT3U(db->db_state, ==, DB_CACHED); | |
34dc7c2f BB |
3814 | ASSERT(db->db_buf != NULL); |
3815 | ||
572e2857 BB |
3816 | DB_DNODE_ENTER(db); |
3817 | dn = DB_DNODE(db); | |
e49f1e20 | 3818 | /* Indirect block size must match what the dnode thinks it is. */ |
572e2857 | 3819 | ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); |
34dc7c2f | 3820 | dbuf_check_blkptr(dn, db); |
572e2857 | 3821 | DB_DNODE_EXIT(db); |
34dc7c2f | 3822 | |
e49f1e20 | 3823 | /* Provide the pending dirty record to child dbufs */ |
34dc7c2f BB |
3824 | db->db_data_pending = dr; |
3825 | ||
34dc7c2f | 3826 | mutex_exit(&db->db_mtx); |
a1d477c2 | 3827 | |
b128c09f | 3828 | dbuf_write(dr, db->db_buf, tx); |
34dc7c2f BB |
3829 | |
3830 | zio = dr->dr_zio; | |
3831 | mutex_enter(&dr->dt.di.dr_mtx); | |
4bda3bd0 | 3832 | dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx); |
34dc7c2f BB |
3833 | ASSERT(list_head(&dr->dt.di.dr_children) == NULL); |
3834 | mutex_exit(&dr->dt.di.dr_mtx); | |
3835 | zio_nowait(zio); | |
3836 | } | |
3837 | ||
0e37a0f4 SD |
3838 | #ifdef ZFS_DEBUG |
3839 | /* | |
3840 | * Verify that the size of the data in our bonus buffer does not exceed | |
3841 | * its recorded size. | |
3842 | * | |
3843 | * The purpose of this verification is to catch any cases in development | |
3844 | * where the size of a phys structure (i.e space_map_phys_t) grows and, | |
3845 | * due to incorrect feature management, older pools expect to read more | |
3846 | * data even though they didn't actually write it to begin with. | |
3847 | * | |
3848 | * For a example, this would catch an error in the feature logic where we | |
3849 | * open an older pool and we expect to write the space map histogram of | |
3850 | * a space map with size SPACE_MAP_SIZE_V0. | |
3851 | */ | |
3852 | static void | |
3853 | dbuf_sync_leaf_verify_bonus_dnode(dbuf_dirty_record_t *dr) | |
3854 | { | |
3855 | dnode_t *dn = DB_DNODE(dr->dr_dbuf); | |
3856 | ||
3857 | /* | |
3858 | * Encrypted bonus buffers can have data past their bonuslen. | |
3859 | * Skip the verification of these blocks. | |
3860 | */ | |
3861 | if (DMU_OT_IS_ENCRYPTED(dn->dn_bonustype)) | |
3862 | return; | |
3863 | ||
3864 | uint16_t bonuslen = dn->dn_phys->dn_bonuslen; | |
3865 | uint16_t maxbonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots); | |
3866 | ASSERT3U(bonuslen, <=, maxbonuslen); | |
3867 | ||
3868 | arc_buf_t *datap = dr->dt.dl.dr_data; | |
3869 | char *datap_end = ((char *)datap) + bonuslen; | |
3870 | char *datap_max = ((char *)datap) + maxbonuslen; | |
3871 | ||
3872 | /* ensure that everything is zero after our data */ | |
3873 | for (; datap_end < datap_max; datap_end++) | |
3874 | ASSERT(*datap_end == 0); | |
3875 | } | |
3876 | #endif | |
3877 | ||
d1d7e268 MK |
3878 | /* |
3879 | * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is | |
60948de1 BB |
3880 | * critical the we not allow the compiler to inline this function in to |
3881 | * dbuf_sync_list() thereby drastically bloating the stack usage. | |
3882 | */ | |
3883 | noinline static void | |
34dc7c2f BB |
3884 | dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx) |
3885 | { | |
3886 | arc_buf_t **datap = &dr->dt.dl.dr_data; | |
3887 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
572e2857 BB |
3888 | dnode_t *dn; |
3889 | objset_t *os; | |
34dc7c2f | 3890 | uint64_t txg = tx->tx_txg; |
34dc7c2f BB |
3891 | |
3892 | ASSERT(dmu_tx_is_syncing(tx)); | |
3893 | ||
3894 | dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); | |
3895 | ||
3896 | mutex_enter(&db->db_mtx); | |
3897 | /* | |
3898 | * To be synced, we must be dirtied. But we | |
3899 | * might have been freed after the dirty. | |
3900 | */ | |
3901 | if (db->db_state == DB_UNCACHED) { | |
3902 | /* This buffer has been freed since it was dirtied */ | |
3903 | ASSERT(db->db.db_data == NULL); | |
3904 | } else if (db->db_state == DB_FILL) { | |
3905 | /* This buffer was freed and is now being re-filled */ | |
3906 | ASSERT(db->db.db_data != dr->dt.dl.dr_data); | |
3907 | } else { | |
b128c09f | 3908 | ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL); |
34dc7c2f BB |
3909 | } |
3910 | DBUF_VERIFY(db); | |
3911 | ||
572e2857 BB |
3912 | DB_DNODE_ENTER(db); |
3913 | dn = DB_DNODE(db); | |
3914 | ||
428870ff BB |
3915 | if (db->db_blkid == DMU_SPILL_BLKID) { |
3916 | mutex_enter(&dn->dn_mtx); | |
81edd3e8 P |
3917 | if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) { |
3918 | /* | |
3919 | * In the previous transaction group, the bonus buffer | |
3920 | * was entirely used to store the attributes for the | |
3921 | * dnode which overrode the dn_spill field. However, | |
3922 | * when adding more attributes to the file a spill | |
3923 | * block was required to hold the extra attributes. | |
3924 | * | |
3925 | * Make sure to clear the garbage left in the dn_spill | |
3926 | * field from the previous attributes in the bonus | |
3927 | * buffer. Otherwise, after writing out the spill | |
3928 | * block to the new allocated dva, it will free | |
3929 | * the old block pointed to by the invalid dn_spill. | |
3930 | */ | |
3931 | db->db_blkptr = NULL; | |
3932 | } | |
428870ff BB |
3933 | dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR; |
3934 | mutex_exit(&dn->dn_mtx); | |
3935 | } | |
3936 | ||
34dc7c2f BB |
3937 | /* |
3938 | * If this is a bonus buffer, simply copy the bonus data into the | |
3939 | * dnode. It will be written out when the dnode is synced (and it | |
3940 | * will be synced, since it must have been dirty for dbuf_sync to | |
3941 | * be called). | |
3942 | */ | |
428870ff | 3943 | if (db->db_blkid == DMU_BONUS_BLKID) { |
34dc7c2f | 3944 | ASSERT(*datap != NULL); |
c99c9001 | 3945 | ASSERT0(db->db_level); |
b5256303 | 3946 | ASSERT3U(DN_MAX_BONUS_LEN(dn->dn_phys), <=, |
50c957f7 | 3947 | DN_SLOTS_TO_BONUSLEN(dn->dn_phys->dn_extra_slots + 1)); |
b5256303 TC |
3948 | bcopy(*datap, DN_BONUS(dn->dn_phys), |
3949 | DN_MAX_BONUS_LEN(dn->dn_phys)); | |
572e2857 BB |
3950 | DB_DNODE_EXIT(db); |
3951 | ||
0e37a0f4 SD |
3952 | #ifdef ZFS_DEBUG |
3953 | dbuf_sync_leaf_verify_bonus_dnode(dr); | |
3954 | #endif | |
3955 | ||
34dc7c2f | 3956 | if (*datap != db->db.db_data) { |
50c957f7 NB |
3957 | int slots = DB_DNODE(db)->dn_num_slots; |
3958 | int bonuslen = DN_SLOTS_TO_BONUSLEN(slots); | |
a3fd9d9e | 3959 | kmem_free(*datap, bonuslen); |
25458cbe | 3960 | arc_space_return(bonuslen, ARC_SPACE_BONUS); |
34dc7c2f BB |
3961 | } |
3962 | db->db_data_pending = NULL; | |
cccbed9f | 3963 | ASSERT(list_next(&db->db_dirty_records, dr) == NULL); |
428870ff | 3964 | ASSERT(dr->dr_dbuf == db); |
cccbed9f | 3965 | list_remove(&db->db_dirty_records, dr); |
753972fc BB |
3966 | if (dr->dr_dbuf->db_level != 0) { |
3967 | mutex_destroy(&dr->dt.di.dr_mtx); | |
3968 | list_destroy(&dr->dt.di.dr_children); | |
3969 | } | |
34dc7c2f BB |
3970 | kmem_free(dr, sizeof (dbuf_dirty_record_t)); |
3971 | ASSERT(db->db_dirtycnt > 0); | |
3972 | db->db_dirtycnt -= 1; | |
3d503a76 | 3973 | dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg, B_FALSE); |
34dc7c2f BB |
3974 | return; |
3975 | } | |
3976 | ||
572e2857 BB |
3977 | os = dn->dn_objset; |
3978 | ||
34dc7c2f BB |
3979 | /* |
3980 | * This function may have dropped the db_mtx lock allowing a dmu_sync | |
3981 | * operation to sneak in. As a result, we need to ensure that we | |
3982 | * don't check the dr_override_state until we have returned from | |
3983 | * dbuf_check_blkptr. | |
3984 | */ | |
3985 | dbuf_check_blkptr(dn, db); | |
3986 | ||
3987 | /* | |
572e2857 | 3988 | * If this buffer is in the middle of an immediate write, |
34dc7c2f BB |
3989 | * wait for the synchronous IO to complete. |
3990 | */ | |
3991 | while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) { | |
3992 | ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT); | |
3993 | cv_wait(&db->db_changed, &db->db_mtx); | |
3994 | ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN); | |
3995 | } | |
3996 | ||
b5256303 TC |
3997 | /* |
3998 | * If this is a dnode block, ensure it is appropriately encrypted | |
3999 | * or decrypted, depending on what we are writing to it this txg. | |
4000 | */ | |
4001 | if (os->os_encrypted && dn->dn_object == DMU_META_DNODE_OBJECT) | |
0c03d21a | 4002 | dbuf_prepare_encrypted_dnode_leaf(dr); |
b5256303 | 4003 | |
9babb374 BB |
4004 | if (db->db_state != DB_NOFILL && |
4005 | dn->dn_object != DMU_META_DNODE_OBJECT && | |
424fd7c3 | 4006 | zfs_refcount_count(&db->db_holds) > 1 && |
428870ff | 4007 | dr->dt.dl.dr_override_state != DR_OVERRIDDEN && |
9babb374 BB |
4008 | *datap == db->db_buf) { |
4009 | /* | |
4010 | * If this buffer is currently "in use" (i.e., there | |
4011 | * are active holds and db_data still references it), | |
4012 | * then make a copy before we start the write so that | |
4013 | * any modifications from the open txg will not leak | |
4014 | * into this write. | |
4015 | * | |
4016 | * NOTE: this copy does not need to be made for | |
4017 | * objects only modified in the syncing context (e.g. | |
4018 | * DNONE_DNODE blocks). | |
4019 | */ | |
2aa34383 | 4020 | int psize = arc_buf_size(*datap); |
b5256303 | 4021 | int lsize = arc_buf_lsize(*datap); |
9babb374 | 4022 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); |
2aa34383 DK |
4023 | enum zio_compress compress_type = arc_get_compression(*datap); |
4024 | ||
b5256303 TC |
4025 | if (arc_is_encrypted(*datap)) { |
4026 | boolean_t byteorder; | |
4027 | uint8_t salt[ZIO_DATA_SALT_LEN]; | |
4028 | uint8_t iv[ZIO_DATA_IV_LEN]; | |
4029 | uint8_t mac[ZIO_DATA_MAC_LEN]; | |
4030 | ||
4031 | arc_get_raw_params(*datap, &byteorder, salt, iv, mac); | |
4032 | *datap = arc_alloc_raw_buf(os->os_spa, db, | |
4033 | dmu_objset_id(os), byteorder, salt, iv, mac, | |
4034 | dn->dn_type, psize, lsize, compress_type); | |
4035 | } else if (compress_type != ZIO_COMPRESS_OFF) { | |
2aa34383 DK |
4036 | ASSERT3U(type, ==, ARC_BUFC_DATA); |
4037 | *datap = arc_alloc_compressed_buf(os->os_spa, db, | |
4038 | psize, lsize, compress_type); | |
b5256303 TC |
4039 | } else { |
4040 | *datap = arc_alloc_buf(os->os_spa, db, type, psize); | |
2aa34383 DK |
4041 | } |
4042 | bcopy(db->db.db_data, (*datap)->b_data, psize); | |
b128c09f | 4043 | } |
34dc7c2f BB |
4044 | db->db_data_pending = dr; |
4045 | ||
4046 | mutex_exit(&db->db_mtx); | |
4047 | ||
b128c09f | 4048 | dbuf_write(dr, *datap, tx); |
34dc7c2f BB |
4049 | |
4050 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
572e2857 | 4051 | if (dn->dn_object == DMU_META_DNODE_OBJECT) { |
3fa93bb8 | 4052 | list_insert_tail(&dn->dn_dirty_records[txg & TXG_MASK], dr); |
572e2857 BB |
4053 | DB_DNODE_EXIT(db); |
4054 | } else { | |
4055 | /* | |
4056 | * Although zio_nowait() does not "wait for an IO", it does | |
4057 | * initiate the IO. If this is an empty write it seems plausible | |
4058 | * that the IO could actually be completed before the nowait | |
4059 | * returns. We need to DB_DNODE_EXIT() first in case | |
4060 | * zio_nowait() invalidates the dbuf. | |
4061 | */ | |
4062 | DB_DNODE_EXIT(db); | |
34dc7c2f | 4063 | zio_nowait(dr->dr_zio); |
572e2857 | 4064 | } |
34dc7c2f BB |
4065 | } |
4066 | ||
4067 | void | |
4bda3bd0 | 4068 | dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx) |
34dc7c2f BB |
4069 | { |
4070 | dbuf_dirty_record_t *dr; | |
4071 | ||
c65aa5b2 | 4072 | while ((dr = list_head(list))) { |
34dc7c2f BB |
4073 | if (dr->dr_zio != NULL) { |
4074 | /* | |
4075 | * If we find an already initialized zio then we | |
4076 | * are processing the meta-dnode, and we have finished. | |
4077 | * The dbufs for all dnodes are put back on the list | |
4078 | * during processing, so that we can zio_wait() | |
4079 | * these IOs after initiating all child IOs. | |
4080 | */ | |
4081 | ASSERT3U(dr->dr_dbuf->db.db_object, ==, | |
4082 | DMU_META_DNODE_OBJECT); | |
4083 | break; | |
4084 | } | |
4bda3bd0 MA |
4085 | if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID && |
4086 | dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) { | |
4087 | VERIFY3U(dr->dr_dbuf->db_level, ==, level); | |
4088 | } | |
34dc7c2f BB |
4089 | list_remove(list, dr); |
4090 | if (dr->dr_dbuf->db_level > 0) | |
4091 | dbuf_sync_indirect(dr, tx); | |
4092 | else | |
4093 | dbuf_sync_leaf(dr, tx); | |
4094 | } | |
4095 | } | |
4096 | ||
34dc7c2f BB |
4097 | /* ARGSUSED */ |
4098 | static void | |
4099 | dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb) | |
4100 | { | |
4101 | dmu_buf_impl_t *db = vdb; | |
572e2857 | 4102 | dnode_t *dn; |
b128c09f | 4103 | blkptr_t *bp = zio->io_bp; |
34dc7c2f | 4104 | blkptr_t *bp_orig = &zio->io_bp_orig; |
428870ff BB |
4105 | spa_t *spa = zio->io_spa; |
4106 | int64_t delta; | |
34dc7c2f | 4107 | uint64_t fill = 0; |
428870ff | 4108 | int i; |
34dc7c2f | 4109 | |
463a8cfe AR |
4110 | ASSERT3P(db->db_blkptr, !=, NULL); |
4111 | ASSERT3P(&db->db_data_pending->dr_bp_copy, ==, bp); | |
b128c09f | 4112 | |
572e2857 BB |
4113 | DB_DNODE_ENTER(db); |
4114 | dn = DB_DNODE(db); | |
428870ff BB |
4115 | delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig); |
4116 | dnode_diduse_space(dn, delta - zio->io_prev_space_delta); | |
4117 | zio->io_prev_space_delta = delta; | |
34dc7c2f | 4118 | |
b0bc7a84 MG |
4119 | if (bp->blk_birth != 0) { |
4120 | ASSERT((db->db_blkid != DMU_SPILL_BLKID && | |
4121 | BP_GET_TYPE(bp) == dn->dn_type) || | |
4122 | (db->db_blkid == DMU_SPILL_BLKID && | |
9b67f605 MA |
4123 | BP_GET_TYPE(bp) == dn->dn_bonustype) || |
4124 | BP_IS_EMBEDDED(bp)); | |
b0bc7a84 | 4125 | ASSERT(BP_GET_LEVEL(bp) == db->db_level); |
34dc7c2f BB |
4126 | } |
4127 | ||
4128 | mutex_enter(&db->db_mtx); | |
4129 | ||
428870ff BB |
4130 | #ifdef ZFS_DEBUG |
4131 | if (db->db_blkid == DMU_SPILL_BLKID) { | |
428870ff | 4132 | ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); |
463a8cfe | 4133 | ASSERT(!(BP_IS_HOLE(bp)) && |
50c957f7 | 4134 | db->db_blkptr == DN_SPILL_BLKPTR(dn->dn_phys)); |
428870ff BB |
4135 | } |
4136 | #endif | |
4137 | ||
34dc7c2f BB |
4138 | if (db->db_level == 0) { |
4139 | mutex_enter(&dn->dn_mtx); | |
428870ff | 4140 | if (db->db_blkid > dn->dn_phys->dn_maxblkid && |
69830602 TC |
4141 | db->db_blkid != DMU_SPILL_BLKID) { |
4142 | ASSERT0(db->db_objset->os_raw_receive); | |
34dc7c2f | 4143 | dn->dn_phys->dn_maxblkid = db->db_blkid; |
69830602 | 4144 | } |
34dc7c2f BB |
4145 | mutex_exit(&dn->dn_mtx); |
4146 | ||
4147 | if (dn->dn_type == DMU_OT_DNODE) { | |
50c957f7 NB |
4148 | i = 0; |
4149 | while (i < db->db.db_size) { | |
817b1b6e MA |
4150 | dnode_phys_t *dnp = |
4151 | (void *)(((char *)db->db.db_data) + i); | |
50c957f7 NB |
4152 | |
4153 | i += DNODE_MIN_SIZE; | |
4154 | if (dnp->dn_type != DMU_OT_NONE) { | |
34dc7c2f | 4155 | fill++; |
50c957f7 NB |
4156 | i += dnp->dn_extra_slots * |
4157 | DNODE_MIN_SIZE; | |
4158 | } | |
34dc7c2f BB |
4159 | } |
4160 | } else { | |
b0bc7a84 MG |
4161 | if (BP_IS_HOLE(bp)) { |
4162 | fill = 0; | |
4163 | } else { | |
4164 | fill = 1; | |
4165 | } | |
34dc7c2f BB |
4166 | } |
4167 | } else { | |
b128c09f | 4168 | blkptr_t *ibp = db->db.db_data; |
34dc7c2f | 4169 | ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); |
b128c09f BB |
4170 | for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) { |
4171 | if (BP_IS_HOLE(ibp)) | |
34dc7c2f | 4172 | continue; |
9b67f605 | 4173 | fill += BP_GET_FILL(ibp); |
34dc7c2f BB |
4174 | } |
4175 | } | |
572e2857 | 4176 | DB_DNODE_EXIT(db); |
34dc7c2f | 4177 | |
9b67f605 | 4178 | if (!BP_IS_EMBEDDED(bp)) |
b5256303 | 4179 | BP_SET_FILL(bp, fill); |
34dc7c2f BB |
4180 | |
4181 | mutex_exit(&db->db_mtx); | |
463a8cfe | 4182 | |
f664f1ee | 4183 | db_lock_type_t dblt = dmu_buf_lock_parent(db, RW_WRITER, FTAG); |
463a8cfe | 4184 | *db->db_blkptr = *bp; |
f664f1ee | 4185 | dmu_buf_unlock_parent(db, dblt, FTAG); |
34dc7c2f BB |
4186 | } |
4187 | ||
bc77ba73 PD |
4188 | /* ARGSUSED */ |
4189 | /* | |
4190 | * This function gets called just prior to running through the compression | |
4191 | * stage of the zio pipeline. If we're an indirect block comprised of only | |
4192 | * holes, then we want this indirect to be compressed away to a hole. In | |
4193 | * order to do that we must zero out any information about the holes that | |
4194 | * this indirect points to prior to before we try to compress it. | |
4195 | */ | |
4196 | static void | |
4197 | dbuf_write_children_ready(zio_t *zio, arc_buf_t *buf, void *vdb) | |
4198 | { | |
4199 | dmu_buf_impl_t *db = vdb; | |
4200 | dnode_t *dn; | |
4201 | blkptr_t *bp; | |
721ed0ee | 4202 | unsigned int epbs, i; |
bc77ba73 PD |
4203 | |
4204 | ASSERT3U(db->db_level, >, 0); | |
4205 | DB_DNODE_ENTER(db); | |
4206 | dn = DB_DNODE(db); | |
4207 | epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; | |
721ed0ee | 4208 | ASSERT3U(epbs, <, 31); |
bc77ba73 PD |
4209 | |
4210 | /* Determine if all our children are holes */ | |
3f93077b | 4211 | for (i = 0, bp = db->db.db_data; i < 1ULL << epbs; i++, bp++) { |
bc77ba73 PD |
4212 | if (!BP_IS_HOLE(bp)) |
4213 | break; | |
4214 | } | |
4215 | ||
4216 | /* | |
4217 | * If all the children are holes, then zero them all out so that | |
4218 | * we may get compressed away. | |
4219 | */ | |
3f93077b | 4220 | if (i == 1ULL << epbs) { |
721ed0ee GM |
4221 | /* |
4222 | * We only found holes. Grab the rwlock to prevent | |
4223 | * anybody from reading the blocks we're about to | |
4224 | * zero out. | |
4225 | */ | |
f664f1ee | 4226 | rw_enter(&db->db_rwlock, RW_WRITER); |
bc77ba73 | 4227 | bzero(db->db.db_data, db->db.db_size); |
f664f1ee | 4228 | rw_exit(&db->db_rwlock); |
bc77ba73 PD |
4229 | } |
4230 | DB_DNODE_EXIT(db); | |
4231 | } | |
4232 | ||
e8b96c60 MA |
4233 | /* |
4234 | * The SPA will call this callback several times for each zio - once | |
4235 | * for every physical child i/o (zio->io_phys_children times). This | |
4236 | * allows the DMU to monitor the progress of each logical i/o. For example, | |
4237 | * there may be 2 copies of an indirect block, or many fragments of a RAID-Z | |
4238 | * block. There may be a long delay before all copies/fragments are completed, | |
4239 | * so this callback allows us to retire dirty space gradually, as the physical | |
4240 | * i/os complete. | |
4241 | */ | |
4242 | /* ARGSUSED */ | |
4243 | static void | |
4244 | dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg) | |
4245 | { | |
4246 | dmu_buf_impl_t *db = arg; | |
4247 | objset_t *os = db->db_objset; | |
4248 | dsl_pool_t *dp = dmu_objset_pool(os); | |
4249 | dbuf_dirty_record_t *dr; | |
4250 | int delta = 0; | |
4251 | ||
4252 | dr = db->db_data_pending; | |
4253 | ASSERT3U(dr->dr_txg, ==, zio->io_txg); | |
4254 | ||
4255 | /* | |
4256 | * The callback will be called io_phys_children times. Retire one | |
4257 | * portion of our dirty space each time we are called. Any rounding | |
0f8ff49e | 4258 | * error will be cleaned up by dbuf_write_done(). |
e8b96c60 MA |
4259 | */ |
4260 | delta = dr->dr_accounted / zio->io_phys_children; | |
4261 | dsl_pool_undirty_space(dp, delta, zio->io_txg); | |
4262 | } | |
4263 | ||
34dc7c2f BB |
4264 | /* ARGSUSED */ |
4265 | static void | |
4266 | dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb) | |
4267 | { | |
4268 | dmu_buf_impl_t *db = vdb; | |
428870ff | 4269 | blkptr_t *bp_orig = &zio->io_bp_orig; |
b0bc7a84 MG |
4270 | blkptr_t *bp = db->db_blkptr; |
4271 | objset_t *os = db->db_objset; | |
4272 | dmu_tx_t *tx = os->os_synctx; | |
cccbed9f | 4273 | dbuf_dirty_record_t *dr; |
34dc7c2f | 4274 | |
c99c9001 | 4275 | ASSERT0(zio->io_error); |
428870ff BB |
4276 | ASSERT(db->db_blkptr == bp); |
4277 | ||
03c6040b GW |
4278 | /* |
4279 | * For nopwrites and rewrites we ensure that the bp matches our | |
4280 | * original and bypass all the accounting. | |
4281 | */ | |
4282 | if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) { | |
428870ff BB |
4283 | ASSERT(BP_EQUAL(bp, bp_orig)); |
4284 | } else { | |
b0bc7a84 | 4285 | dsl_dataset_t *ds = os->os_dsl_dataset; |
428870ff BB |
4286 | (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE); |
4287 | dsl_dataset_block_born(ds, bp, tx); | |
4288 | } | |
34dc7c2f BB |
4289 | |
4290 | mutex_enter(&db->db_mtx); | |
4291 | ||
428870ff BB |
4292 | DBUF_VERIFY(db); |
4293 | ||
cccbed9f | 4294 | dr = db->db_data_pending; |
34dc7c2f | 4295 | ASSERT(!list_link_active(&dr->dr_dirty_node)); |
428870ff | 4296 | ASSERT(dr->dr_dbuf == db); |
cccbed9f MM |
4297 | ASSERT(list_next(&db->db_dirty_records, dr) == NULL); |
4298 | list_remove(&db->db_dirty_records, dr); | |
34dc7c2f | 4299 | |
428870ff BB |
4300 | #ifdef ZFS_DEBUG |
4301 | if (db->db_blkid == DMU_SPILL_BLKID) { | |
572e2857 BB |
4302 | dnode_t *dn; |
4303 | ||
4304 | DB_DNODE_ENTER(db); | |
4305 | dn = DB_DNODE(db); | |
428870ff BB |
4306 | ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); |
4307 | ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && | |
50c957f7 | 4308 | db->db_blkptr == DN_SPILL_BLKPTR(dn->dn_phys)); |
572e2857 | 4309 | DB_DNODE_EXIT(db); |
428870ff BB |
4310 | } |
4311 | #endif | |
4312 | ||
34dc7c2f | 4313 | if (db->db_level == 0) { |
428870ff | 4314 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f | 4315 | ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); |
b128c09f BB |
4316 | if (db->db_state != DB_NOFILL) { |
4317 | if (dr->dt.dl.dr_data != db->db_buf) | |
d3c2ae1c | 4318 | arc_buf_destroy(dr->dt.dl.dr_data, db); |
b128c09f | 4319 | } |
34dc7c2f | 4320 | } else { |
572e2857 BB |
4321 | dnode_t *dn; |
4322 | ||
4323 | DB_DNODE_ENTER(db); | |
4324 | dn = DB_DNODE(db); | |
34dc7c2f | 4325 | ASSERT(list_head(&dr->dt.di.dr_children) == NULL); |
b0bc7a84 | 4326 | ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift); |
34dc7c2f | 4327 | if (!BP_IS_HOLE(db->db_blkptr)) { |
2a8ba608 MM |
4328 | int epbs __maybe_unused = dn->dn_phys->dn_indblkshift - |
4329 | SPA_BLKPTRSHIFT; | |
b0bc7a84 MG |
4330 | ASSERT3U(db->db_blkid, <=, |
4331 | dn->dn_phys->dn_maxblkid >> (db->db_level * epbs)); | |
34dc7c2f BB |
4332 | ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, |
4333 | db->db.db_size); | |
34dc7c2f | 4334 | } |
572e2857 | 4335 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
4336 | mutex_destroy(&dr->dt.di.dr_mtx); |
4337 | list_destroy(&dr->dt.di.dr_children); | |
4338 | } | |
34dc7c2f BB |
4339 | |
4340 | cv_broadcast(&db->db_changed); | |
4341 | ASSERT(db->db_dirtycnt > 0); | |
4342 | db->db_dirtycnt -= 1; | |
4343 | db->db_data_pending = NULL; | |
3d503a76 | 4344 | dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg, B_FALSE); |
0f8ff49e SD |
4345 | |
4346 | /* | |
4347 | * If we didn't do a physical write in this ZIO and we | |
4348 | * still ended up here, it means that the space of the | |
4349 | * dbuf that we just released (and undirtied) above hasn't | |
4350 | * been marked as undirtied in the pool's accounting. | |
4351 | * | |
4352 | * Thus, we undirty that space in the pool's view of the | |
4353 | * world here. For physical writes this type of update | |
4354 | * happens in dbuf_write_physdone(). | |
4355 | * | |
4356 | * If we did a physical write, cleanup any rounding errors | |
4357 | * that came up due to writing multiple copies of a block | |
4358 | * on disk [see dbuf_write_physdone()]. | |
4359 | */ | |
4360 | if (zio->io_phys_children == 0) { | |
4361 | dsl_pool_undirty_space(dmu_objset_pool(os), | |
4362 | dr->dr_accounted, zio->io_txg); | |
4363 | } else { | |
4364 | dsl_pool_undirty_space(dmu_objset_pool(os), | |
4365 | dr->dr_accounted % zio->io_phys_children, zio->io_txg); | |
4366 | } | |
4367 | ||
4368 | kmem_free(dr, sizeof (dbuf_dirty_record_t)); | |
428870ff BB |
4369 | } |
4370 | ||
4371 | static void | |
4372 | dbuf_write_nofill_ready(zio_t *zio) | |
4373 | { | |
4374 | dbuf_write_ready(zio, NULL, zio->io_private); | |
4375 | } | |
4376 | ||
4377 | static void | |
4378 | dbuf_write_nofill_done(zio_t *zio) | |
4379 | { | |
4380 | dbuf_write_done(zio, NULL, zio->io_private); | |
4381 | } | |
4382 | ||
4383 | static void | |
4384 | dbuf_write_override_ready(zio_t *zio) | |
4385 | { | |
4386 | dbuf_dirty_record_t *dr = zio->io_private; | |
4387 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
4388 | ||
4389 | dbuf_write_ready(zio, NULL, db); | |
4390 | } | |
4391 | ||
4392 | static void | |
4393 | dbuf_write_override_done(zio_t *zio) | |
4394 | { | |
4395 | dbuf_dirty_record_t *dr = zio->io_private; | |
4396 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
4397 | blkptr_t *obp = &dr->dt.dl.dr_overridden_by; | |
4398 | ||
4399 | mutex_enter(&db->db_mtx); | |
4400 | if (!BP_EQUAL(zio->io_bp, obp)) { | |
4401 | if (!BP_IS_HOLE(obp)) | |
4402 | dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp); | |
4403 | arc_release(dr->dt.dl.dr_data, db); | |
4404 | } | |
34dc7c2f BB |
4405 | mutex_exit(&db->db_mtx); |
4406 | ||
428870ff | 4407 | dbuf_write_done(zio, NULL, db); |
a6255b7f DQ |
4408 | |
4409 | if (zio->io_abd != NULL) | |
4410 | abd_put(zio->io_abd); | |
428870ff BB |
4411 | } |
4412 | ||
a1d477c2 MA |
4413 | typedef struct dbuf_remap_impl_callback_arg { |
4414 | objset_t *drica_os; | |
4415 | uint64_t drica_blk_birth; | |
4416 | dmu_tx_t *drica_tx; | |
4417 | } dbuf_remap_impl_callback_arg_t; | |
4418 | ||
4419 | static void | |
4420 | dbuf_remap_impl_callback(uint64_t vdev, uint64_t offset, uint64_t size, | |
4421 | void *arg) | |
4422 | { | |
4423 | dbuf_remap_impl_callback_arg_t *drica = arg; | |
4424 | objset_t *os = drica->drica_os; | |
4425 | spa_t *spa = dmu_objset_spa(os); | |
4426 | dmu_tx_t *tx = drica->drica_tx; | |
4427 | ||
4428 | ASSERT(dsl_pool_sync_context(spa_get_dsl(spa))); | |
4429 | ||
4430 | if (os == spa_meta_objset(spa)) { | |
4431 | spa_vdev_indirect_mark_obsolete(spa, vdev, offset, size, tx); | |
4432 | } else { | |
4433 | dsl_dataset_block_remapped(dmu_objset_ds(os), vdev, offset, | |
4434 | size, drica->drica_blk_birth, tx); | |
4435 | } | |
4436 | } | |
4437 | ||
4438 | static void | |
f664f1ee | 4439 | dbuf_remap_impl(dnode_t *dn, blkptr_t *bp, krwlock_t *rw, dmu_tx_t *tx) |
a1d477c2 MA |
4440 | { |
4441 | blkptr_t bp_copy = *bp; | |
4442 | spa_t *spa = dmu_objset_spa(dn->dn_objset); | |
4443 | dbuf_remap_impl_callback_arg_t drica; | |
4444 | ||
4445 | ASSERT(dsl_pool_sync_context(spa_get_dsl(spa))); | |
4446 | ||
4447 | drica.drica_os = dn->dn_objset; | |
4448 | drica.drica_blk_birth = bp->blk_birth; | |
4449 | drica.drica_tx = tx; | |
4450 | if (spa_remap_blkptr(spa, &bp_copy, dbuf_remap_impl_callback, | |
4451 | &drica)) { | |
37f03da8 SH |
4452 | /* |
4453 | * If the blkptr being remapped is tracked by a livelist, | |
4454 | * then we need to make sure the livelist reflects the update. | |
4455 | * First, cancel out the old blkptr by appending a 'FREE' | |
4456 | * entry. Next, add an 'ALLOC' to track the new version. This | |
4457 | * way we avoid trying to free an inaccurate blkptr at delete. | |
4458 | * Note that embedded blkptrs are not tracked in livelists. | |
4459 | */ | |
4460 | if (dn->dn_objset != spa_meta_objset(spa)) { | |
4461 | dsl_dataset_t *ds = dmu_objset_ds(dn->dn_objset); | |
4462 | if (dsl_deadlist_is_open(&ds->ds_dir->dd_livelist) && | |
4463 | bp->blk_birth > ds->ds_dir->dd_origin_txg) { | |
4464 | ASSERT(!BP_IS_EMBEDDED(bp)); | |
4465 | ASSERT(dsl_dir_is_clone(ds->ds_dir)); | |
4466 | ASSERT(spa_feature_is_enabled(spa, | |
4467 | SPA_FEATURE_LIVELIST)); | |
4468 | bplist_append(&ds->ds_dir->dd_pending_frees, | |
4469 | bp); | |
4470 | bplist_append(&ds->ds_dir->dd_pending_allocs, | |
4471 | &bp_copy); | |
4472 | } | |
4473 | } | |
4474 | ||
a1d477c2 | 4475 | /* |
f664f1ee | 4476 | * The db_rwlock prevents dbuf_read_impl() from |
a1d477c2 MA |
4477 | * dereferencing the BP while we are changing it. To |
4478 | * avoid lock contention, only grab it when we are actually | |
4479 | * changing the BP. | |
4480 | */ | |
f664f1ee PD |
4481 | if (rw != NULL) |
4482 | rw_enter(rw, RW_WRITER); | |
a1d477c2 | 4483 | *bp = bp_copy; |
f664f1ee PD |
4484 | if (rw != NULL) |
4485 | rw_exit(rw); | |
a1d477c2 MA |
4486 | } |
4487 | } | |
4488 | ||
a1d477c2 MA |
4489 | /* |
4490 | * Remap any existing BP's to concrete vdevs, if possible. | |
4491 | */ | |
4492 | static void | |
4493 | dbuf_remap(dnode_t *dn, dmu_buf_impl_t *db, dmu_tx_t *tx) | |
4494 | { | |
4495 | spa_t *spa = dmu_objset_spa(db->db_objset); | |
4496 | ASSERT(dsl_pool_sync_context(spa_get_dsl(spa))); | |
4497 | ||
4498 | if (!spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL)) | |
4499 | return; | |
4500 | ||
4501 | if (db->db_level > 0) { | |
4502 | blkptr_t *bp = db->db.db_data; | |
4503 | for (int i = 0; i < db->db.db_size >> SPA_BLKPTRSHIFT; i++) { | |
f664f1ee | 4504 | dbuf_remap_impl(dn, &bp[i], &db->db_rwlock, tx); |
a1d477c2 MA |
4505 | } |
4506 | } else if (db->db.db_object == DMU_META_DNODE_OBJECT) { | |
4507 | dnode_phys_t *dnp = db->db.db_data; | |
4508 | ASSERT3U(db->db_dnode_handle->dnh_dnode->dn_type, ==, | |
4509 | DMU_OT_DNODE); | |
4510 | for (int i = 0; i < db->db.db_size >> DNODE_SHIFT; | |
4511 | i += dnp[i].dn_extra_slots + 1) { | |
4512 | for (int j = 0; j < dnp[i].dn_nblkptr; j++) { | |
f664f1ee PD |
4513 | krwlock_t *lock = (dn->dn_dbuf == NULL ? NULL : |
4514 | &dn->dn_dbuf->db_rwlock); | |
4515 | dbuf_remap_impl(dn, &dnp[i].dn_blkptr[j], lock, | |
4516 | tx); | |
a1d477c2 MA |
4517 | } |
4518 | } | |
4519 | } | |
4520 | } | |
4521 | ||
4522 | ||
e49f1e20 | 4523 | /* Issue I/O to commit a dirty buffer to disk. */ |
428870ff BB |
4524 | static void |
4525 | dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx) | |
4526 | { | |
4527 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
572e2857 BB |
4528 | dnode_t *dn; |
4529 | objset_t *os; | |
428870ff BB |
4530 | dmu_buf_impl_t *parent = db->db_parent; |
4531 | uint64_t txg = tx->tx_txg; | |
5dbd68a3 | 4532 | zbookmark_phys_t zb; |
428870ff BB |
4533 | zio_prop_t zp; |
4534 | zio_t *zio; | |
4535 | int wp_flag = 0; | |
34dc7c2f | 4536 | |
463a8cfe AR |
4537 | ASSERT(dmu_tx_is_syncing(tx)); |
4538 | ||
572e2857 BB |
4539 | DB_DNODE_ENTER(db); |
4540 | dn = DB_DNODE(db); | |
4541 | os = dn->dn_objset; | |
4542 | ||
428870ff BB |
4543 | if (db->db_state != DB_NOFILL) { |
4544 | if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) { | |
4545 | /* | |
4546 | * Private object buffers are released here rather | |
4547 | * than in dbuf_dirty() since they are only modified | |
4548 | * in the syncing context and we don't want the | |
4549 | * overhead of making multiple copies of the data. | |
4550 | */ | |
4551 | if (BP_IS_HOLE(db->db_blkptr)) { | |
4552 | arc_buf_thaw(data); | |
4553 | } else { | |
4554 | dbuf_release_bp(db); | |
4555 | } | |
a1d477c2 | 4556 | dbuf_remap(dn, db, tx); |
428870ff BB |
4557 | } |
4558 | } | |
4559 | ||
4560 | if (parent != dn->dn_dbuf) { | |
e49f1e20 WA |
4561 | /* Our parent is an indirect block. */ |
4562 | /* We have a dirty parent that has been scheduled for write. */ | |
428870ff | 4563 | ASSERT(parent && parent->db_data_pending); |
e49f1e20 | 4564 | /* Our parent's buffer is one level closer to the dnode. */ |
428870ff | 4565 | ASSERT(db->db_level == parent->db_level-1); |
e49f1e20 WA |
4566 | /* |
4567 | * We're about to modify our parent's db_data by modifying | |
4568 | * our block pointer, so the parent must be released. | |
4569 | */ | |
428870ff BB |
4570 | ASSERT(arc_released(parent->db_buf)); |
4571 | zio = parent->db_data_pending->dr_zio; | |
4572 | } else { | |
e49f1e20 | 4573 | /* Our parent is the dnode itself. */ |
428870ff BB |
4574 | ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 && |
4575 | db->db_blkid != DMU_SPILL_BLKID) || | |
4576 | (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0)); | |
4577 | if (db->db_blkid != DMU_SPILL_BLKID) | |
4578 | ASSERT3P(db->db_blkptr, ==, | |
4579 | &dn->dn_phys->dn_blkptr[db->db_blkid]); | |
4580 | zio = dn->dn_zio; | |
4581 | } | |
4582 | ||
4583 | ASSERT(db->db_level == 0 || data == db->db_buf); | |
4584 | ASSERT3U(db->db_blkptr->blk_birth, <=, txg); | |
4585 | ASSERT(zio); | |
4586 | ||
4587 | SET_BOOKMARK(&zb, os->os_dsl_dataset ? | |
4588 | os->os_dsl_dataset->ds_object : DMU_META_OBJSET, | |
4589 | db->db.db_object, db->db_level, db->db_blkid); | |
4590 | ||
4591 | if (db->db_blkid == DMU_SPILL_BLKID) | |
4592 | wp_flag = WP_SPILL; | |
4593 | wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0; | |
4594 | ||
82644107 | 4595 | dmu_write_policy(os, dn, db->db_level, wp_flag, &zp); |
572e2857 | 4596 | DB_DNODE_EXIT(db); |
428870ff | 4597 | |
463a8cfe AR |
4598 | /* |
4599 | * We copy the blkptr now (rather than when we instantiate the dirty | |
4600 | * record), because its value can change between open context and | |
4601 | * syncing context. We do not need to hold dn_struct_rwlock to read | |
4602 | * db_blkptr because we are in syncing context. | |
4603 | */ | |
4604 | dr->dr_bp_copy = *db->db_blkptr; | |
4605 | ||
9b67f605 MA |
4606 | if (db->db_level == 0 && |
4607 | dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { | |
4608 | /* | |
4609 | * The BP for this block has been provided by open context | |
4610 | * (by dmu_sync() or dmu_buf_write_embedded()). | |
4611 | */ | |
a6255b7f DQ |
4612 | abd_t *contents = (data != NULL) ? |
4613 | abd_get_from_buf(data->b_data, arc_buf_size(data)) : NULL; | |
9b67f605 | 4614 | |
428870ff | 4615 | dr->dr_zio = zio_write(zio, os->os_spa, txg, |
2aa34383 DK |
4616 | &dr->dr_bp_copy, contents, db->db.db_size, db->db.db_size, |
4617 | &zp, dbuf_write_override_ready, NULL, NULL, | |
bc77ba73 | 4618 | dbuf_write_override_done, |
e8b96c60 | 4619 | dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); |
428870ff BB |
4620 | mutex_enter(&db->db_mtx); |
4621 | dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; | |
4622 | zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by, | |
03c6040b | 4623 | dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite); |
428870ff BB |
4624 | mutex_exit(&db->db_mtx); |
4625 | } else if (db->db_state == DB_NOFILL) { | |
3c67d83a TH |
4626 | ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF || |
4627 | zp.zp_checksum == ZIO_CHECKSUM_NOPARITY); | |
428870ff | 4628 | dr->dr_zio = zio_write(zio, os->os_spa, txg, |
2aa34383 | 4629 | &dr->dr_bp_copy, NULL, db->db.db_size, db->db.db_size, &zp, |
bc77ba73 PD |
4630 | dbuf_write_nofill_ready, NULL, NULL, |
4631 | dbuf_write_nofill_done, db, | |
428870ff BB |
4632 | ZIO_PRIORITY_ASYNC_WRITE, |
4633 | ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb); | |
4634 | } else { | |
4635 | ASSERT(arc_released(data)); | |
bc77ba73 PD |
4636 | |
4637 | /* | |
4638 | * For indirect blocks, we want to setup the children | |
4639 | * ready callback so that we can properly handle an indirect | |
4640 | * block that only contains holes. | |
4641 | */ | |
1c27024e | 4642 | arc_write_done_func_t *children_ready_cb = NULL; |
bc77ba73 PD |
4643 | if (db->db_level != 0) |
4644 | children_ready_cb = dbuf_write_children_ready; | |
4645 | ||
428870ff | 4646 | dr->dr_zio = arc_write(zio, os->os_spa, txg, |
463a8cfe | 4647 | &dr->dr_bp_copy, data, DBUF_IS_L2CACHEABLE(db), |
d3c2ae1c GW |
4648 | &zp, dbuf_write_ready, |
4649 | children_ready_cb, dbuf_write_physdone, | |
4650 | dbuf_write_done, db, ZIO_PRIORITY_ASYNC_WRITE, | |
4651 | ZIO_FLAG_MUSTSUCCEED, &zb); | |
428870ff | 4652 | } |
34dc7c2f | 4653 | } |
c28b2279 | 4654 | |
8f576c23 BB |
4655 | EXPORT_SYMBOL(dbuf_find); |
4656 | EXPORT_SYMBOL(dbuf_is_metadata); | |
d3c2ae1c | 4657 | EXPORT_SYMBOL(dbuf_destroy); |
8f576c23 BB |
4658 | EXPORT_SYMBOL(dbuf_loan_arcbuf); |
4659 | EXPORT_SYMBOL(dbuf_whichblock); | |
4660 | EXPORT_SYMBOL(dbuf_read); | |
4661 | EXPORT_SYMBOL(dbuf_unoverride); | |
4662 | EXPORT_SYMBOL(dbuf_free_range); | |
4663 | EXPORT_SYMBOL(dbuf_new_size); | |
4664 | EXPORT_SYMBOL(dbuf_release_bp); | |
4665 | EXPORT_SYMBOL(dbuf_dirty); | |
0c03d21a | 4666 | EXPORT_SYMBOL(dmu_buf_set_crypt_params); |
c28b2279 | 4667 | EXPORT_SYMBOL(dmu_buf_will_dirty); |
a73e8fdb | 4668 | EXPORT_SYMBOL(dmu_buf_is_dirty); |
8f576c23 BB |
4669 | EXPORT_SYMBOL(dmu_buf_will_not_fill); |
4670 | EXPORT_SYMBOL(dmu_buf_will_fill); | |
4671 | EXPORT_SYMBOL(dmu_buf_fill_done); | |
4047414a | 4672 | EXPORT_SYMBOL(dmu_buf_rele); |
8f576c23 | 4673 | EXPORT_SYMBOL(dbuf_assign_arcbuf); |
8f576c23 BB |
4674 | EXPORT_SYMBOL(dbuf_prefetch); |
4675 | EXPORT_SYMBOL(dbuf_hold_impl); | |
4676 | EXPORT_SYMBOL(dbuf_hold); | |
4677 | EXPORT_SYMBOL(dbuf_hold_level); | |
4678 | EXPORT_SYMBOL(dbuf_create_bonus); | |
4679 | EXPORT_SYMBOL(dbuf_spill_set_blksz); | |
4680 | EXPORT_SYMBOL(dbuf_rm_spill); | |
4681 | EXPORT_SYMBOL(dbuf_add_ref); | |
4682 | EXPORT_SYMBOL(dbuf_rele); | |
4683 | EXPORT_SYMBOL(dbuf_rele_and_unlock); | |
4684 | EXPORT_SYMBOL(dbuf_refcount); | |
4685 | EXPORT_SYMBOL(dbuf_sync_list); | |
4686 | EXPORT_SYMBOL(dmu_buf_set_user); | |
4687 | EXPORT_SYMBOL(dmu_buf_set_user_ie); | |
8f576c23 | 4688 | EXPORT_SYMBOL(dmu_buf_get_user); |
0f699108 | 4689 | EXPORT_SYMBOL(dmu_buf_get_blkptr); |
d3c2ae1c | 4690 | |
02730c33 | 4691 | /* BEGIN CSTYLED */ |
03fdcb9a | 4692 | ZFS_MODULE_PARAM(zfs_dbuf_cache, dbuf_cache_, max_bytes, ULONG, ZMOD_RW, |
02730c33 | 4693 | "Maximum size in bytes of the dbuf cache."); |
d3c2ae1c | 4694 | |
03fdcb9a | 4695 | ZFS_MODULE_PARAM(zfs_dbuf_cache, dbuf_cache_, hiwater_pct, UINT, ZMOD_RW, |
f974e25d | 4696 | "Percentage over dbuf_cache_max_bytes when dbufs must be evicted " |
4697 | "directly."); | |
d3c2ae1c | 4698 | |
03fdcb9a | 4699 | ZFS_MODULE_PARAM(zfs_dbuf_cache, dbuf_cache_, lowater_pct, UINT, ZMOD_RW, |
f974e25d | 4700 | "Percentage below dbuf_cache_max_bytes when the evict thread stops " |
4701 | "evicting dbufs."); | |
d3c2ae1c | 4702 | |
03fdcb9a | 4703 | ZFS_MODULE_PARAM(zfs_dbuf, dbuf_, metadata_cache_max_bytes, ULONG, ZMOD_RW, |
2e5dc449 MA |
4704 | "Maximum size in bytes of the dbuf metadata cache."); |
4705 | ||
03fdcb9a | 4706 | ZFS_MODULE_PARAM(zfs_dbuf, dbuf_, cache_shift, INT, ZMOD_RW, |
de4f8d5d | 4707 | "Set the size of the dbuf cache to a log2 fraction of arc size."); |
2e5dc449 | 4708 | |
03fdcb9a MM |
4709 | ZFS_MODULE_PARAM(zfs_dbuf, dbuf_, metadata_cache_shift, INT, ZMOD_RW, |
4710 | "Set the size of the dbuf metadata cache to a log2 fraction of arc " | |
4711 | "size."); | |
02730c33 | 4712 | /* END CSTYLED */ |