]>
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 | |
1d3ba0bf | 9 | * or https://opensource.org/licenses/CDDL-1.0. |
34dc7c2f BB |
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. |
ba67d821 | 24 | * Copyright (c) 2012, 2020 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. |
10b3c7f5 MN |
27 | * Copyright (c) 2019, Klara Inc. |
28 | * Copyright (c) 2019, Allan Jude | |
34dc7c2f BB |
29 | */ |
30 | ||
34dc7c2f | 31 | #include <sys/zfs_context.h> |
c28b2279 | 32 | #include <sys/arc.h> |
34dc7c2f | 33 | #include <sys/dmu.h> |
ea97f8ce | 34 | #include <sys/dmu_send.h> |
34dc7c2f BB |
35 | #include <sys/dmu_impl.h> |
36 | #include <sys/dbuf.h> | |
37 | #include <sys/dmu_objset.h> | |
38 | #include <sys/dsl_dataset.h> | |
39 | #include <sys/dsl_dir.h> | |
40 | #include <sys/dmu_tx.h> | |
41 | #include <sys/spa.h> | |
42 | #include <sys/zio.h> | |
43 | #include <sys/dmu_zfetch.h> | |
428870ff BB |
44 | #include <sys/sa.h> |
45 | #include <sys/sa_impl.h> | |
9b67f605 MA |
46 | #include <sys/zfeature.h> |
47 | #include <sys/blkptr.h> | |
9bd274dd | 48 | #include <sys/range_tree.h> |
e5d1c27e | 49 | #include <sys/trace_zfs.h> |
d3c2ae1c | 50 | #include <sys/callb.h> |
a6255b7f | 51 | #include <sys/abd.h> |
a1d477c2 | 52 | #include <sys/vdev.h> |
3f387973 | 53 | #include <cityhash.h> |
2e5dc449 | 54 | #include <sys/spa_impl.h> |
c4c162c1 | 55 | #include <sys/wmsum.h> |
c9d62d13 | 56 | #include <sys/vdev_impl.h> |
34dc7c2f | 57 | |
18168da7 | 58 | static kstat_t *dbuf_ksp; |
5e021f56 GDN |
59 | |
60 | typedef struct dbuf_stats { | |
61 | /* | |
62 | * Various statistics about the size of the dbuf cache. | |
63 | */ | |
64 | kstat_named_t cache_count; | |
65 | kstat_named_t cache_size_bytes; | |
66 | kstat_named_t cache_size_bytes_max; | |
67 | /* | |
68 | * Statistics regarding the bounds on the dbuf cache size. | |
69 | */ | |
70 | kstat_named_t cache_target_bytes; | |
71 | kstat_named_t cache_lowater_bytes; | |
72 | kstat_named_t cache_hiwater_bytes; | |
73 | /* | |
74 | * Total number of dbuf cache evictions that have occurred. | |
75 | */ | |
76 | kstat_named_t cache_total_evicts; | |
77 | /* | |
78 | * The distribution of dbuf levels in the dbuf cache and | |
79 | * the total size of all dbufs at each level. | |
80 | */ | |
81 | kstat_named_t cache_levels[DN_MAX_LEVELS]; | |
82 | kstat_named_t cache_levels_bytes[DN_MAX_LEVELS]; | |
83 | /* | |
84 | * Statistics about the dbuf hash table. | |
85 | */ | |
86 | kstat_named_t hash_hits; | |
87 | kstat_named_t hash_misses; | |
88 | kstat_named_t hash_collisions; | |
89 | kstat_named_t hash_elements; | |
90 | kstat_named_t hash_elements_max; | |
91 | /* | |
92 | * Number of sublists containing more than one dbuf in the dbuf | |
93 | * hash table. Keep track of the longest hash chain. | |
94 | */ | |
95 | kstat_named_t hash_chains; | |
96 | kstat_named_t hash_chain_max; | |
97 | /* | |
98 | * Number of times a dbuf_create() discovers that a dbuf was | |
99 | * already created and in the dbuf hash table. | |
100 | */ | |
101 | kstat_named_t hash_insert_race; | |
2e5dc449 MA |
102 | /* |
103 | * Statistics about the size of the metadata dbuf cache. | |
104 | */ | |
105 | kstat_named_t metadata_cache_count; | |
106 | kstat_named_t metadata_cache_size_bytes; | |
107 | kstat_named_t metadata_cache_size_bytes_max; | |
108 | /* | |
109 | * For diagnostic purposes, this is incremented whenever we can't add | |
110 | * something to the metadata cache because it's full, and instead put | |
111 | * the data in the regular dbuf cache. | |
112 | */ | |
113 | kstat_named_t metadata_cache_overflow; | |
5e021f56 GDN |
114 | } dbuf_stats_t; |
115 | ||
116 | dbuf_stats_t dbuf_stats = { | |
117 | { "cache_count", KSTAT_DATA_UINT64 }, | |
118 | { "cache_size_bytes", KSTAT_DATA_UINT64 }, | |
119 | { "cache_size_bytes_max", KSTAT_DATA_UINT64 }, | |
120 | { "cache_target_bytes", KSTAT_DATA_UINT64 }, | |
121 | { "cache_lowater_bytes", KSTAT_DATA_UINT64 }, | |
122 | { "cache_hiwater_bytes", KSTAT_DATA_UINT64 }, | |
123 | { "cache_total_evicts", KSTAT_DATA_UINT64 }, | |
124 | { { "cache_levels_N", KSTAT_DATA_UINT64 } }, | |
125 | { { "cache_levels_bytes_N", KSTAT_DATA_UINT64 } }, | |
126 | { "hash_hits", KSTAT_DATA_UINT64 }, | |
127 | { "hash_misses", KSTAT_DATA_UINT64 }, | |
128 | { "hash_collisions", KSTAT_DATA_UINT64 }, | |
129 | { "hash_elements", KSTAT_DATA_UINT64 }, | |
130 | { "hash_elements_max", KSTAT_DATA_UINT64 }, | |
131 | { "hash_chains", KSTAT_DATA_UINT64 }, | |
132 | { "hash_chain_max", KSTAT_DATA_UINT64 }, | |
2e5dc449 MA |
133 | { "hash_insert_race", KSTAT_DATA_UINT64 }, |
134 | { "metadata_cache_count", KSTAT_DATA_UINT64 }, | |
135 | { "metadata_cache_size_bytes", KSTAT_DATA_UINT64 }, | |
136 | { "metadata_cache_size_bytes_max", KSTAT_DATA_UINT64 }, | |
137 | { "metadata_cache_overflow", KSTAT_DATA_UINT64 } | |
5e021f56 GDN |
138 | }; |
139 | ||
c4c162c1 AM |
140 | struct { |
141 | wmsum_t cache_count; | |
142 | wmsum_t cache_total_evicts; | |
143 | wmsum_t cache_levels[DN_MAX_LEVELS]; | |
144 | wmsum_t cache_levels_bytes[DN_MAX_LEVELS]; | |
145 | wmsum_t hash_hits; | |
146 | wmsum_t hash_misses; | |
147 | wmsum_t hash_collisions; | |
148 | wmsum_t hash_chains; | |
149 | wmsum_t hash_insert_race; | |
150 | wmsum_t metadata_cache_count; | |
151 | wmsum_t metadata_cache_overflow; | |
152 | } dbuf_sums; | |
153 | ||
5e021f56 | 154 | #define DBUF_STAT_INCR(stat, val) \ |
c4c162c1 | 155 | wmsum_add(&dbuf_sums.stat, val); |
5e021f56 GDN |
156 | #define DBUF_STAT_DECR(stat, val) \ |
157 | DBUF_STAT_INCR(stat, -(val)); | |
158 | #define DBUF_STAT_BUMP(stat) \ | |
159 | DBUF_STAT_INCR(stat, 1); | |
160 | #define DBUF_STAT_BUMPDOWN(stat) \ | |
161 | DBUF_STAT_INCR(stat, -1); | |
162 | #define DBUF_STAT_MAX(stat, v) { \ | |
163 | uint64_t _m; \ | |
164 | while ((v) > (_m = dbuf_stats.stat.value.ui64) && \ | |
165 | (_m != atomic_cas_64(&dbuf_stats.stat.value.ui64, _m, (v))))\ | |
166 | continue; \ | |
167 | } | |
168 | ||
13fe0198 | 169 | static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx); |
b128c09f | 170 | static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx); |
fa3922df | 171 | static void dbuf_sync_leaf_verify_bonus_dnode(dbuf_dirty_record_t *dr); |
8b3547a4 | 172 | static int dbuf_read_verify_dnode_crypt(dmu_buf_impl_t *db, uint32_t flags); |
34dc7c2f BB |
173 | |
174 | /* | |
175 | * Global data structures and functions for the dbuf cache. | |
176 | */ | |
d3c2ae1c | 177 | static kmem_cache_t *dbuf_kmem_cache; |
0c66c32d | 178 | static taskq_t *dbu_evict_taskq; |
34dc7c2f | 179 | |
d3c2ae1c GW |
180 | static kthread_t *dbuf_cache_evict_thread; |
181 | static kmutex_t dbuf_evict_lock; | |
182 | static kcondvar_t dbuf_evict_cv; | |
183 | static boolean_t dbuf_evict_thread_exit; | |
184 | ||
185 | /* | |
2e5dc449 MA |
186 | * There are two dbuf caches; each dbuf can only be in one of them at a time. |
187 | * | |
188 | * 1. Cache of metadata dbufs, to help make read-heavy administrative commands | |
189 | * from /sbin/zfs run faster. The "metadata cache" specifically stores dbufs | |
190 | * that represent the metadata that describes filesystems/snapshots/ | |
191 | * bookmarks/properties/etc. We only evict from this cache when we export a | |
192 | * pool, to short-circuit as much I/O as possible for all administrative | |
193 | * commands that need the metadata. There is no eviction policy for this | |
194 | * cache, because we try to only include types in it which would occupy a | |
195 | * very small amount of space per object but create a large impact on the | |
196 | * performance of these commands. Instead, after it reaches a maximum size | |
197 | * (which should only happen on very small memory systems with a very large | |
198 | * number of filesystem objects), we stop taking new dbufs into the | |
199 | * metadata cache, instead putting them in the normal dbuf cache. | |
200 | * | |
201 | * 2. LRU cache of dbufs. The dbuf cache maintains a list of dbufs that | |
202 | * are not currently held but have been recently released. These dbufs | |
203 | * are not eligible for arc eviction until they are aged out of the cache. | |
204 | * Dbufs that are aged out of the cache will be immediately destroyed and | |
205 | * become eligible for arc eviction. | |
206 | * | |
207 | * Dbufs are added to these caches once the last hold is released. If a dbuf is | |
208 | * later accessed and still exists in the dbuf cache, then it will be removed | |
209 | * from the cache and later re-added to the head of the cache. | |
210 | * | |
211 | * If a given dbuf meets the requirements for the metadata cache, it will go | |
212 | * there, otherwise it will be considered for the generic LRU dbuf cache. The | |
213 | * caches and the refcounts tracking their sizes are stored in an array indexed | |
214 | * by those caches' matching enum values (from dbuf_cached_state_t). | |
d3c2ae1c | 215 | */ |
2e5dc449 | 216 | typedef struct dbuf_cache { |
ffdf019c AM |
217 | multilist_t cache; |
218 | zfs_refcount_t size ____cacheline_aligned; | |
2e5dc449 MA |
219 | } dbuf_cache_t; |
220 | dbuf_cache_t dbuf_caches[DB_CACHE_MAX]; | |
d3c2ae1c | 221 | |
2e5dc449 | 222 | /* Size limits for the caches */ |
18168da7 AZ |
223 | static unsigned long dbuf_cache_max_bytes = ULONG_MAX; |
224 | static unsigned long dbuf_metadata_cache_max_bytes = ULONG_MAX; | |
8348fac3 | 225 | |
2e5dc449 | 226 | /* Set the default sizes of the caches to log2 fraction of arc size */ |
18168da7 AZ |
227 | static int dbuf_cache_shift = 5; |
228 | static int dbuf_metadata_cache_shift = 6; | |
d3c2ae1c | 229 | |
8348fac3 RM |
230 | static unsigned long dbuf_cache_target_bytes(void); |
231 | static unsigned long dbuf_metadata_cache_target_bytes(void); | |
232 | ||
d3c2ae1c | 233 | /* |
2e5dc449 | 234 | * The LRU dbuf cache uses a three-stage eviction policy: |
d3c2ae1c GW |
235 | * - A low water marker designates when the dbuf eviction thread |
236 | * should stop evicting from the dbuf cache. | |
237 | * - When we reach the maximum size (aka mid water mark), we | |
238 | * signal the eviction thread to run. | |
239 | * - The high water mark indicates when the eviction thread | |
240 | * is unable to keep up with the incoming load and eviction must | |
241 | * happen in the context of the calling thread. | |
242 | * | |
243 | * The dbuf cache: | |
244 | * (max size) | |
245 | * low water mid water hi water | |
246 | * +----------------------------------------+----------+----------+ | |
247 | * | | | | | |
248 | * | | | | | |
249 | * | | | | | |
250 | * | | | | | |
251 | * +----------------------------------------+----------+----------+ | |
252 | * stop signal evict | |
253 | * evicting eviction directly | |
254 | * thread | |
255 | * | |
256 | * The high and low water marks indicate the operating range for the eviction | |
257 | * thread. The low water mark is, by default, 90% of the total size of the | |
258 | * cache and the high water mark is at 110% (both of these percentages can be | |
259 | * changed by setting dbuf_cache_lowater_pct and dbuf_cache_hiwater_pct, | |
260 | * respectively). The eviction thread will try to ensure that the cache remains | |
261 | * within this range by waking up every second and checking if the cache is | |
262 | * above the low water mark. The thread can also be woken up by callers adding | |
263 | * elements into the cache if the cache is larger than the mid water (i.e max | |
264 | * cache size). Once the eviction thread is woken up and eviction is required, | |
265 | * it will continue evicting buffers until it's able to reduce the cache size | |
266 | * to the low water mark. If the cache size continues to grow and hits the high | |
4e33ba4c | 267 | * water mark, then callers adding elements to the cache will begin to evict |
d3c2ae1c GW |
268 | * directly from the cache until the cache is no longer above the high water |
269 | * mark. | |
270 | */ | |
271 | ||
272 | /* | |
273 | * The percentage above and below the maximum cache size. | |
274 | */ | |
18168da7 AZ |
275 | static uint_t dbuf_cache_hiwater_pct = 10; |
276 | static uint_t dbuf_cache_lowater_pct = 10; | |
d3c2ae1c | 277 | |
34dc7c2f BB |
278 | static int |
279 | dbuf_cons(void *vdb, void *unused, int kmflag) | |
280 | { | |
14e4e3cb | 281 | (void) unused, (void) kmflag; |
34dc7c2f | 282 | dmu_buf_impl_t *db = vdb; |
861166b0 | 283 | memset(db, 0, sizeof (dmu_buf_impl_t)); |
34dc7c2f BB |
284 | |
285 | mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL); | |
f664f1ee | 286 | rw_init(&db->db_rwlock, NULL, RW_DEFAULT, NULL); |
34dc7c2f | 287 | cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL); |
d3c2ae1c | 288 | multilist_link_init(&db->db_cache_link); |
424fd7c3 | 289 | zfs_refcount_create(&db->db_holds); |
8951cb8d | 290 | |
34dc7c2f BB |
291 | return (0); |
292 | } | |
293 | ||
34dc7c2f BB |
294 | static void |
295 | dbuf_dest(void *vdb, void *unused) | |
296 | { | |
14e4e3cb | 297 | (void) unused; |
34dc7c2f BB |
298 | dmu_buf_impl_t *db = vdb; |
299 | mutex_destroy(&db->db_mtx); | |
f664f1ee | 300 | rw_destroy(&db->db_rwlock); |
34dc7c2f | 301 | cv_destroy(&db->db_changed); |
d3c2ae1c | 302 | ASSERT(!multilist_link_active(&db->db_cache_link)); |
424fd7c3 | 303 | zfs_refcount_destroy(&db->db_holds); |
34dc7c2f BB |
304 | } |
305 | ||
306 | /* | |
307 | * dbuf hash table routines | |
308 | */ | |
309 | static dbuf_hash_table_t dbuf_hash_table; | |
310 | ||
37fb3e43 PD |
311 | /* |
312 | * We use Cityhash for this. It's fast, and has good hash properties without | |
313 | * requiring any large static buffers. | |
314 | */ | |
34dc7c2f BB |
315 | static uint64_t |
316 | dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid) | |
317 | { | |
37fb3e43 | 318 | return (cityhash4((uintptr_t)os, obj, (uint64_t)lvl, blkid)); |
34dc7c2f BB |
319 | } |
320 | ||
8b3547a4 MM |
321 | #define DTRACE_SET_STATE(db, why) \ |
322 | DTRACE_PROBE2(dbuf__state_change, dmu_buf_impl_t *, db, \ | |
323 | const char *, why) | |
324 | ||
34dc7c2f BB |
325 | #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \ |
326 | ((dbuf)->db.db_object == (obj) && \ | |
327 | (dbuf)->db_objset == (os) && \ | |
328 | (dbuf)->db_level == (level) && \ | |
329 | (dbuf)->db_blkid == (blkid)) | |
330 | ||
331 | dmu_buf_impl_t * | |
6ebebace | 332 | dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid) |
34dc7c2f BB |
333 | { |
334 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
d6320ddb BB |
335 | uint64_t hv; |
336 | uint64_t idx; | |
34dc7c2f BB |
337 | dmu_buf_impl_t *db; |
338 | ||
d3c2ae1c | 339 | hv = dbuf_hash(os, obj, level, blkid); |
d6320ddb BB |
340 | idx = hv & h->hash_table_mask; |
341 | ||
223b04d2 | 342 | mutex_enter(DBUF_HASH_MUTEX(h, idx)); |
34dc7c2f BB |
343 | for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) { |
344 | if (DBUF_EQUAL(db, os, obj, level, blkid)) { | |
345 | mutex_enter(&db->db_mtx); | |
346 | if (db->db_state != DB_EVICTING) { | |
223b04d2 | 347 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); |
34dc7c2f BB |
348 | return (db); |
349 | } | |
350 | mutex_exit(&db->db_mtx); | |
351 | } | |
352 | } | |
223b04d2 | 353 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); |
34dc7c2f BB |
354 | return (NULL); |
355 | } | |
356 | ||
6ebebace JG |
357 | static dmu_buf_impl_t * |
358 | dbuf_find_bonus(objset_t *os, uint64_t object) | |
359 | { | |
360 | dnode_t *dn; | |
361 | dmu_buf_impl_t *db = NULL; | |
362 | ||
363 | if (dnode_hold(os, object, FTAG, &dn) == 0) { | |
364 | rw_enter(&dn->dn_struct_rwlock, RW_READER); | |
365 | if (dn->dn_bonus != NULL) { | |
366 | db = dn->dn_bonus; | |
367 | mutex_enter(&db->db_mtx); | |
368 | } | |
369 | rw_exit(&dn->dn_struct_rwlock); | |
370 | dnode_rele(dn, FTAG); | |
371 | } | |
372 | return (db); | |
373 | } | |
374 | ||
34dc7c2f BB |
375 | /* |
376 | * Insert an entry into the hash table. If there is already an element | |
377 | * equal to elem in the hash table, then the already existing element | |
378 | * will be returned and the new element will not be inserted. | |
379 | * Otherwise returns NULL. | |
380 | */ | |
381 | static dmu_buf_impl_t * | |
382 | dbuf_hash_insert(dmu_buf_impl_t *db) | |
383 | { | |
384 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
428870ff | 385 | objset_t *os = db->db_objset; |
34dc7c2f BB |
386 | uint64_t obj = db->db.db_object; |
387 | int level = db->db_level; | |
d6320ddb | 388 | uint64_t blkid, hv, idx; |
34dc7c2f | 389 | dmu_buf_impl_t *dbf; |
5e021f56 | 390 | uint32_t i; |
34dc7c2f | 391 | |
d6320ddb | 392 | blkid = db->db_blkid; |
d3c2ae1c | 393 | hv = dbuf_hash(os, obj, level, blkid); |
d6320ddb BB |
394 | idx = hv & h->hash_table_mask; |
395 | ||
223b04d2 | 396 | mutex_enter(DBUF_HASH_MUTEX(h, idx)); |
5e021f56 GDN |
397 | for (dbf = h->hash_table[idx], i = 0; dbf != NULL; |
398 | dbf = dbf->db_hash_next, i++) { | |
34dc7c2f BB |
399 | if (DBUF_EQUAL(dbf, os, obj, level, blkid)) { |
400 | mutex_enter(&dbf->db_mtx); | |
401 | if (dbf->db_state != DB_EVICTING) { | |
223b04d2 | 402 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); |
34dc7c2f BB |
403 | return (dbf); |
404 | } | |
405 | mutex_exit(&dbf->db_mtx); | |
406 | } | |
407 | } | |
408 | ||
5e021f56 GDN |
409 | if (i > 0) { |
410 | DBUF_STAT_BUMP(hash_collisions); | |
411 | if (i == 1) | |
412 | DBUF_STAT_BUMP(hash_chains); | |
413 | ||
414 | DBUF_STAT_MAX(hash_chain_max, i); | |
415 | } | |
416 | ||
34dc7c2f BB |
417 | mutex_enter(&db->db_mtx); |
418 | db->db_hash_next = h->hash_table[idx]; | |
419 | h->hash_table[idx] = db; | |
223b04d2 | 420 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); |
c4c162c1 AM |
421 | uint64_t he = atomic_inc_64_nv(&dbuf_stats.hash_elements.value.ui64); |
422 | DBUF_STAT_MAX(hash_elements_max, he); | |
34dc7c2f BB |
423 | |
424 | return (NULL); | |
425 | } | |
426 | ||
2e5dc449 MA |
427 | /* |
428 | * This returns whether this dbuf should be stored in the metadata cache, which | |
429 | * is based on whether it's from one of the dnode types that store data related | |
430 | * to traversing dataset hierarchies. | |
431 | */ | |
432 | static boolean_t | |
433 | dbuf_include_in_metadata_cache(dmu_buf_impl_t *db) | |
434 | { | |
435 | DB_DNODE_ENTER(db); | |
436 | dmu_object_type_t type = DB_DNODE(db)->dn_type; | |
437 | DB_DNODE_EXIT(db); | |
438 | ||
439 | /* Check if this dbuf is one of the types we care about */ | |
440 | if (DMU_OT_IS_METADATA_CACHED(type)) { | |
441 | /* If we hit this, then we set something up wrong in dmu_ot */ | |
442 | ASSERT(DMU_OT_IS_METADATA(type)); | |
443 | ||
444 | /* | |
445 | * Sanity check for small-memory systems: don't allocate too | |
446 | * much memory for this purpose. | |
447 | */ | |
424fd7c3 TS |
448 | if (zfs_refcount_count( |
449 | &dbuf_caches[DB_DBUF_METADATA_CACHE].size) > | |
8348fac3 | 450 | dbuf_metadata_cache_target_bytes()) { |
2e5dc449 MA |
451 | DBUF_STAT_BUMP(metadata_cache_overflow); |
452 | return (B_FALSE); | |
453 | } | |
454 | ||
455 | return (B_TRUE); | |
456 | } | |
457 | ||
458 | return (B_FALSE); | |
459 | } | |
460 | ||
34dc7c2f | 461 | /* |
bd089c54 | 462 | * Remove an entry from the hash table. It must be in the EVICTING state. |
34dc7c2f BB |
463 | */ |
464 | static void | |
465 | dbuf_hash_remove(dmu_buf_impl_t *db) | |
466 | { | |
467 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
d6320ddb | 468 | uint64_t hv, idx; |
34dc7c2f BB |
469 | dmu_buf_impl_t *dbf, **dbp; |
470 | ||
d3c2ae1c | 471 | hv = dbuf_hash(db->db_objset, db->db.db_object, |
d6320ddb BB |
472 | db->db_level, db->db_blkid); |
473 | idx = hv & h->hash_table_mask; | |
474 | ||
34dc7c2f | 475 | /* |
4e33ba4c | 476 | * We mustn't hold db_mtx to maintain lock ordering: |
223b04d2 | 477 | * DBUF_HASH_MUTEX > db_mtx. |
34dc7c2f | 478 | */ |
424fd7c3 | 479 | ASSERT(zfs_refcount_is_zero(&db->db_holds)); |
34dc7c2f BB |
480 | ASSERT(db->db_state == DB_EVICTING); |
481 | ASSERT(!MUTEX_HELD(&db->db_mtx)); | |
482 | ||
223b04d2 | 483 | mutex_enter(DBUF_HASH_MUTEX(h, idx)); |
34dc7c2f BB |
484 | dbp = &h->hash_table[idx]; |
485 | while ((dbf = *dbp) != db) { | |
486 | dbp = &dbf->db_hash_next; | |
487 | ASSERT(dbf != NULL); | |
488 | } | |
489 | *dbp = db->db_hash_next; | |
490 | db->db_hash_next = NULL; | |
5e021f56 GDN |
491 | if (h->hash_table[idx] && |
492 | h->hash_table[idx]->db_hash_next == NULL) | |
493 | DBUF_STAT_BUMPDOWN(hash_chains); | |
223b04d2 | 494 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); |
c4c162c1 | 495 | atomic_dec_64(&dbuf_stats.hash_elements.value.ui64); |
34dc7c2f BB |
496 | } |
497 | ||
0c66c32d JG |
498 | typedef enum { |
499 | DBVU_EVICTING, | |
500 | DBVU_NOT_EVICTING | |
501 | } dbvu_verify_type_t; | |
502 | ||
503 | static void | |
504 | dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type) | |
505 | { | |
506 | #ifdef ZFS_DEBUG | |
507 | int64_t holds; | |
508 | ||
509 | if (db->db_user == NULL) | |
510 | return; | |
511 | ||
512 | /* Only data blocks support the attachment of user data. */ | |
513 | ASSERT(db->db_level == 0); | |
514 | ||
515 | /* Clients must resolve a dbuf before attaching user data. */ | |
516 | ASSERT(db->db.db_data != NULL); | |
517 | ASSERT3U(db->db_state, ==, DB_CACHED); | |
518 | ||
424fd7c3 | 519 | holds = zfs_refcount_count(&db->db_holds); |
0c66c32d JG |
520 | if (verify_type == DBVU_EVICTING) { |
521 | /* | |
522 | * Immediate eviction occurs when holds == dirtycnt. | |
523 | * For normal eviction buffers, holds is zero on | |
524 | * eviction, except when dbuf_fix_old_data() calls | |
525 | * dbuf_clear_data(). However, the hold count can grow | |
526 | * during eviction even though db_mtx is held (see | |
527 | * dmu_bonus_hold() for an example), so we can only | |
528 | * test the generic invariant that holds >= dirtycnt. | |
529 | */ | |
530 | ASSERT3U(holds, >=, db->db_dirtycnt); | |
531 | } else { | |
bc4501f7 | 532 | if (db->db_user_immediate_evict == TRUE) |
0c66c32d JG |
533 | ASSERT3U(holds, >=, db->db_dirtycnt); |
534 | else | |
535 | ASSERT3U(holds, >, 0); | |
536 | } | |
537 | #endif | |
538 | } | |
539 | ||
34dc7c2f BB |
540 | static void |
541 | dbuf_evict_user(dmu_buf_impl_t *db) | |
542 | { | |
0c66c32d JG |
543 | dmu_buf_user_t *dbu = db->db_user; |
544 | ||
34dc7c2f BB |
545 | ASSERT(MUTEX_HELD(&db->db_mtx)); |
546 | ||
0c66c32d | 547 | if (dbu == NULL) |
34dc7c2f BB |
548 | return; |
549 | ||
0c66c32d JG |
550 | dbuf_verify_user(db, DBVU_EVICTING); |
551 | db->db_user = NULL; | |
552 | ||
553 | #ifdef ZFS_DEBUG | |
554 | if (dbu->dbu_clear_on_evict_dbufp != NULL) | |
555 | *dbu->dbu_clear_on_evict_dbufp = NULL; | |
556 | #endif | |
557 | ||
558 | /* | |
39efbde7 GM |
559 | * There are two eviction callbacks - one that we call synchronously |
560 | * and one that we invoke via a taskq. The async one is useful for | |
561 | * avoiding lock order reversals and limiting stack depth. | |
562 | * | |
563 | * Note that if we have a sync callback but no async callback, | |
564 | * it's likely that the sync callback will free the structure | |
565 | * containing the dbu. In that case we need to take care to not | |
566 | * dereference dbu after calling the sync evict func. | |
0c66c32d | 567 | */ |
a7004725 | 568 | boolean_t has_async = (dbu->dbu_evict_func_async != NULL); |
39efbde7 GM |
569 | |
570 | if (dbu->dbu_evict_func_sync != NULL) | |
571 | dbu->dbu_evict_func_sync(dbu); | |
572 | ||
573 | if (has_async) { | |
574 | taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func_async, | |
575 | dbu, 0, &dbu->dbu_tqent); | |
576 | } | |
34dc7c2f BB |
577 | } |
578 | ||
572e2857 BB |
579 | boolean_t |
580 | dbuf_is_metadata(dmu_buf_impl_t *db) | |
581 | { | |
cc79a5c2 BB |
582 | /* |
583 | * Consider indirect blocks and spill blocks to be meta data. | |
584 | */ | |
585 | if (db->db_level > 0 || db->db_blkid == DMU_SPILL_BLKID) { | |
572e2857 BB |
586 | return (B_TRUE); |
587 | } else { | |
588 | boolean_t is_metadata; | |
589 | ||
590 | DB_DNODE_ENTER(db); | |
9ae529ec | 591 | is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type); |
572e2857 BB |
592 | DB_DNODE_EXIT(db); |
593 | ||
594 | return (is_metadata); | |
595 | } | |
596 | } | |
597 | ||
c9d62d13 GA |
598 | /* |
599 | * We want to exclude buffers that are on a special allocation class from | |
600 | * L2ARC. | |
601 | */ | |
602 | boolean_t | |
603 | dbuf_is_l2cacheable(dmu_buf_impl_t *db) | |
604 | { | |
605 | vdev_t *vd = NULL; | |
606 | zfs_cache_type_t cache = db->db_objset->os_secondary_cache; | |
607 | blkptr_t *bp = db->db_blkptr; | |
608 | ||
609 | if (bp != NULL && !BP_IS_HOLE(bp)) { | |
610 | uint64_t vdev = DVA_GET_VDEV(bp->blk_dva); | |
611 | vdev_t *rvd = db->db_objset->os_spa->spa_root_vdev; | |
612 | ||
613 | if (vdev < rvd->vdev_children) | |
614 | vd = rvd->vdev_child[vdev]; | |
615 | ||
616 | if (cache == ZFS_CACHE_ALL || | |
617 | (dbuf_is_metadata(db) && cache == ZFS_CACHE_METADATA)) { | |
618 | if (vd == NULL) | |
619 | return (B_TRUE); | |
620 | ||
621 | if ((vd->vdev_alloc_bias != VDEV_BIAS_SPECIAL && | |
622 | vd->vdev_alloc_bias != VDEV_BIAS_DEDUP) || | |
623 | l2arc_exclude_special == 0) | |
624 | return (B_TRUE); | |
625 | } | |
626 | } | |
627 | ||
628 | return (B_FALSE); | |
629 | } | |
630 | ||
631 | static inline boolean_t | |
632 | dnode_level_is_l2cacheable(blkptr_t *bp, dnode_t *dn, int64_t level) | |
633 | { | |
634 | vdev_t *vd = NULL; | |
635 | zfs_cache_type_t cache = dn->dn_objset->os_secondary_cache; | |
636 | ||
637 | if (bp != NULL && !BP_IS_HOLE(bp)) { | |
638 | uint64_t vdev = DVA_GET_VDEV(bp->blk_dva); | |
639 | vdev_t *rvd = dn->dn_objset->os_spa->spa_root_vdev; | |
640 | ||
641 | if (vdev < rvd->vdev_children) | |
642 | vd = rvd->vdev_child[vdev]; | |
643 | ||
644 | if (cache == ZFS_CACHE_ALL || ((level > 0 || | |
645 | DMU_OT_IS_METADATA(dn->dn_handle->dnh_dnode->dn_type)) && | |
646 | cache == ZFS_CACHE_METADATA)) { | |
647 | if (vd == NULL) | |
648 | return (B_TRUE); | |
649 | ||
650 | if ((vd->vdev_alloc_bias != VDEV_BIAS_SPECIAL && | |
651 | vd->vdev_alloc_bias != VDEV_BIAS_DEDUP) || | |
652 | l2arc_exclude_special == 0) | |
653 | return (B_TRUE); | |
654 | } | |
655 | } | |
656 | ||
657 | return (B_FALSE); | |
658 | } | |
659 | ||
d3c2ae1c GW |
660 | |
661 | /* | |
662 | * This function *must* return indices evenly distributed between all | |
663 | * sublists of the multilist. This is needed due to how the dbuf eviction | |
664 | * code is laid out; dbuf_evict_thread() assumes dbufs are evenly | |
665 | * distributed between all sublists and uses this assumption when | |
666 | * deciding which sublist to evict from and how much to evict from it. | |
667 | */ | |
65c7cc49 | 668 | static unsigned int |
d3c2ae1c | 669 | dbuf_cache_multilist_index_func(multilist_t *ml, void *obj) |
34dc7c2f | 670 | { |
d3c2ae1c GW |
671 | dmu_buf_impl_t *db = obj; |
672 | ||
673 | /* | |
674 | * The assumption here, is the hash value for a given | |
675 | * dmu_buf_impl_t will remain constant throughout it's lifetime | |
676 | * (i.e. it's objset, object, level and blkid fields don't change). | |
677 | * Thus, we don't need to store the dbuf's sublist index | |
678 | * on insertion, as this index can be recalculated on removal. | |
679 | * | |
680 | * Also, the low order bits of the hash value are thought to be | |
681 | * distributed evenly. Otherwise, in the case that the multilist | |
682 | * has a power of two number of sublists, each sublists' usage | |
5b7053a9 AM |
683 | * would not be evenly distributed. In this context full 64bit |
684 | * division would be a waste of time, so limit it to 32 bits. | |
d3c2ae1c | 685 | */ |
5b7053a9 | 686 | return ((unsigned int)dbuf_hash(db->db_objset, db->db.db_object, |
d3c2ae1c GW |
687 | db->db_level, db->db_blkid) % |
688 | multilist_get_num_sublists(ml)); | |
689 | } | |
690 | ||
8348fac3 RM |
691 | /* |
692 | * The target size of the dbuf cache can grow with the ARC target, | |
693 | * unless limited by the tunable dbuf_cache_max_bytes. | |
694 | */ | |
e71cade6 | 695 | static inline unsigned long |
696 | dbuf_cache_target_bytes(void) | |
697 | { | |
8348fac3 RM |
698 | return (MIN(dbuf_cache_max_bytes, |
699 | arc_target_bytes() >> dbuf_cache_shift)); | |
700 | } | |
701 | ||
702 | /* | |
703 | * The target size of the dbuf metadata cache can grow with the ARC target, | |
704 | * unless limited by the tunable dbuf_metadata_cache_max_bytes. | |
705 | */ | |
706 | static inline unsigned long | |
707 | dbuf_metadata_cache_target_bytes(void) | |
708 | { | |
709 | return (MIN(dbuf_metadata_cache_max_bytes, | |
710 | arc_target_bytes() >> dbuf_metadata_cache_shift)); | |
e71cade6 | 711 | } |
712 | ||
5e021f56 GDN |
713 | static inline uint64_t |
714 | dbuf_cache_hiwater_bytes(void) | |
d3c2ae1c | 715 | { |
e71cade6 | 716 | uint64_t dbuf_cache_target = dbuf_cache_target_bytes(); |
5e021f56 GDN |
717 | return (dbuf_cache_target + |
718 | (dbuf_cache_target * dbuf_cache_hiwater_pct) / 100); | |
719 | } | |
e71cade6 | 720 | |
5e021f56 GDN |
721 | static inline uint64_t |
722 | dbuf_cache_lowater_bytes(void) | |
723 | { | |
724 | uint64_t dbuf_cache_target = dbuf_cache_target_bytes(); | |
725 | return (dbuf_cache_target - | |
726 | (dbuf_cache_target * dbuf_cache_lowater_pct) / 100); | |
727 | } | |
d3c2ae1c | 728 | |
d3c2ae1c GW |
729 | static inline boolean_t |
730 | dbuf_cache_above_lowater(void) | |
731 | { | |
424fd7c3 | 732 | return (zfs_refcount_count(&dbuf_caches[DB_DBUF_CACHE].size) > |
2e5dc449 | 733 | dbuf_cache_lowater_bytes()); |
d3c2ae1c GW |
734 | } |
735 | ||
736 | /* | |
737 | * Evict the oldest eligible dbuf from the dbuf cache. | |
738 | */ | |
739 | static void | |
740 | dbuf_evict_one(void) | |
741 | { | |
ffdf019c | 742 | int idx = multilist_get_random_index(&dbuf_caches[DB_DBUF_CACHE].cache); |
2e5dc449 | 743 | multilist_sublist_t *mls = multilist_sublist_lock( |
ffdf019c | 744 | &dbuf_caches[DB_DBUF_CACHE].cache, idx); |
1c27024e | 745 | |
d3c2ae1c GW |
746 | ASSERT(!MUTEX_HELD(&dbuf_evict_lock)); |
747 | ||
1c27024e | 748 | dmu_buf_impl_t *db = multilist_sublist_tail(mls); |
d3c2ae1c GW |
749 | while (db != NULL && mutex_tryenter(&db->db_mtx) == 0) { |
750 | db = multilist_sublist_prev(mls, db); | |
751 | } | |
752 | ||
753 | DTRACE_PROBE2(dbuf__evict__one, dmu_buf_impl_t *, db, | |
754 | multilist_sublist_t *, mls); | |
755 | ||
756 | if (db != NULL) { | |
757 | multilist_sublist_remove(mls, db); | |
758 | multilist_sublist_unlock(mls); | |
424fd7c3 TS |
759 | (void) zfs_refcount_remove_many( |
760 | &dbuf_caches[DB_DBUF_CACHE].size, db->db.db_size, db); | |
5e021f56 GDN |
761 | DBUF_STAT_BUMPDOWN(cache_levels[db->db_level]); |
762 | DBUF_STAT_BUMPDOWN(cache_count); | |
763 | DBUF_STAT_DECR(cache_levels_bytes[db->db_level], | |
764 | db->db.db_size); | |
2e5dc449 MA |
765 | ASSERT3U(db->db_caching_status, ==, DB_DBUF_CACHE); |
766 | db->db_caching_status = DB_NO_CACHE; | |
d3c2ae1c | 767 | dbuf_destroy(db); |
5e021f56 | 768 | DBUF_STAT_BUMP(cache_total_evicts); |
d3c2ae1c GW |
769 | } else { |
770 | multilist_sublist_unlock(mls); | |
771 | } | |
d3c2ae1c GW |
772 | } |
773 | ||
774 | /* | |
775 | * The dbuf evict thread is responsible for aging out dbufs from the | |
776 | * cache. Once the cache has reached it's maximum size, dbufs are removed | |
777 | * and destroyed. The eviction thread will continue running until the size | |
778 | * of the dbuf cache is at or below the maximum size. Once the dbuf is aged | |
779 | * out of the cache it is destroyed and becomes eligible for arc eviction. | |
780 | */ | |
460748d4 | 781 | static __attribute__((noreturn)) void |
c25b8f99 | 782 | dbuf_evict_thread(void *unused) |
d3c2ae1c | 783 | { |
14e4e3cb | 784 | (void) unused; |
d3c2ae1c GW |
785 | callb_cpr_t cpr; |
786 | ||
787 | CALLB_CPR_INIT(&cpr, &dbuf_evict_lock, callb_generic_cpr, FTAG); | |
788 | ||
789 | mutex_enter(&dbuf_evict_lock); | |
790 | while (!dbuf_evict_thread_exit) { | |
791 | while (!dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) { | |
792 | CALLB_CPR_SAFE_BEGIN(&cpr); | |
ac6e5fb2 | 793 | (void) cv_timedwait_idle_hires(&dbuf_evict_cv, |
d3c2ae1c GW |
794 | &dbuf_evict_lock, SEC2NSEC(1), MSEC2NSEC(1), 0); |
795 | CALLB_CPR_SAFE_END(&cpr, &dbuf_evict_lock); | |
796 | } | |
797 | mutex_exit(&dbuf_evict_lock); | |
798 | ||
799 | /* | |
800 | * Keep evicting as long as we're above the low water mark | |
801 | * for the cache. We do this without holding the locks to | |
802 | * minimize lock contention. | |
803 | */ | |
804 | while (dbuf_cache_above_lowater() && !dbuf_evict_thread_exit) { | |
805 | dbuf_evict_one(); | |
806 | } | |
807 | ||
808 | mutex_enter(&dbuf_evict_lock); | |
809 | } | |
810 | ||
811 | dbuf_evict_thread_exit = B_FALSE; | |
812 | cv_broadcast(&dbuf_evict_cv); | |
813 | CALLB_CPR_EXIT(&cpr); /* drops dbuf_evict_lock */ | |
814 | thread_exit(); | |
815 | } | |
816 | ||
817 | /* | |
818 | * Wake up the dbuf eviction thread if the dbuf cache is at its max size. | |
819 | * If the dbuf cache is at its high water mark, then evict a dbuf from the | |
7ada752a | 820 | * dbuf cache using the caller's context. |
d3c2ae1c GW |
821 | */ |
822 | static void | |
cbd8f5b7 | 823 | dbuf_evict_notify(uint64_t size) |
d3c2ae1c | 824 | { |
38240ebd MA |
825 | /* |
826 | * We check if we should evict without holding the dbuf_evict_lock, | |
827 | * because it's OK to occasionally make the wrong decision here, | |
828 | * and grabbing the lock results in massive lock contention. | |
829 | */ | |
cbd8f5b7 AM |
830 | if (size > dbuf_cache_target_bytes()) { |
831 | if (size > dbuf_cache_hiwater_bytes()) | |
d3c2ae1c | 832 | dbuf_evict_one(); |
38240ebd | 833 | cv_signal(&dbuf_evict_cv); |
d3c2ae1c | 834 | } |
34dc7c2f BB |
835 | } |
836 | ||
5e021f56 GDN |
837 | static int |
838 | dbuf_kstat_update(kstat_t *ksp, int rw) | |
839 | { | |
840 | dbuf_stats_t *ds = ksp->ks_data; | |
d3c2ae1c | 841 | |
c4c162c1 | 842 | if (rw == KSTAT_WRITE) |
5e021f56 | 843 | return (SET_ERROR(EACCES)); |
5e021f56 | 844 | |
c4c162c1 AM |
845 | ds->cache_count.value.ui64 = |
846 | wmsum_value(&dbuf_sums.cache_count); | |
847 | ds->cache_size_bytes.value.ui64 = | |
848 | zfs_refcount_count(&dbuf_caches[DB_DBUF_CACHE].size); | |
849 | ds->cache_target_bytes.value.ui64 = dbuf_cache_target_bytes(); | |
850 | ds->cache_hiwater_bytes.value.ui64 = dbuf_cache_hiwater_bytes(); | |
851 | ds->cache_lowater_bytes.value.ui64 = dbuf_cache_lowater_bytes(); | |
852 | ds->cache_total_evicts.value.ui64 = | |
853 | wmsum_value(&dbuf_sums.cache_total_evicts); | |
854 | for (int i = 0; i < DN_MAX_LEVELS; i++) { | |
855 | ds->cache_levels[i].value.ui64 = | |
856 | wmsum_value(&dbuf_sums.cache_levels[i]); | |
857 | ds->cache_levels_bytes[i].value.ui64 = | |
858 | wmsum_value(&dbuf_sums.cache_levels_bytes[i]); | |
859 | } | |
860 | ds->hash_hits.value.ui64 = | |
861 | wmsum_value(&dbuf_sums.hash_hits); | |
862 | ds->hash_misses.value.ui64 = | |
863 | wmsum_value(&dbuf_sums.hash_misses); | |
864 | ds->hash_collisions.value.ui64 = | |
865 | wmsum_value(&dbuf_sums.hash_collisions); | |
866 | ds->hash_chains.value.ui64 = | |
867 | wmsum_value(&dbuf_sums.hash_chains); | |
868 | ds->hash_insert_race.value.ui64 = | |
869 | wmsum_value(&dbuf_sums.hash_insert_race); | |
870 | ds->metadata_cache_count.value.ui64 = | |
871 | wmsum_value(&dbuf_sums.metadata_cache_count); | |
872 | ds->metadata_cache_size_bytes.value.ui64 = zfs_refcount_count( | |
873 | &dbuf_caches[DB_DBUF_METADATA_CACHE].size); | |
874 | ds->metadata_cache_overflow.value.ui64 = | |
875 | wmsum_value(&dbuf_sums.metadata_cache_overflow); | |
5e021f56 GDN |
876 | return (0); |
877 | } | |
d3c2ae1c | 878 | |
34dc7c2f BB |
879 | void |
880 | dbuf_init(void) | |
881 | { | |
882 | uint64_t hsize = 1ULL << 16; | |
883 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
884 | int i; | |
885 | ||
886 | /* | |
490c845e | 887 | * The hash table is big enough to fill one eighth of physical memory |
69de3421 TC |
888 | * with an average block size of zfs_arc_average_blocksize (default 8K). |
889 | * By default, the table will take up | |
890 | * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). | |
34dc7c2f | 891 | */ |
490c845e | 892 | while (hsize * zfs_arc_average_blocksize < arc_all_memory() / 8) |
34dc7c2f BB |
893 | hsize <<= 1; |
894 | ||
895 | retry: | |
896 | h->hash_table_mask = hsize - 1; | |
93ce2b4c | 897 | #if defined(_KERNEL) |
d1d7e268 MK |
898 | /* |
899 | * Large allocations which do not require contiguous pages | |
900 | * should be using vmem_alloc() in the linux kernel | |
901 | */ | |
79c76d5b | 902 | h->hash_table = vmem_zalloc(hsize * sizeof (void *), KM_SLEEP); |
00b46022 | 903 | #else |
34dc7c2f | 904 | h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP); |
00b46022 | 905 | #endif |
34dc7c2f BB |
906 | if (h->hash_table == NULL) { |
907 | /* XXX - we should really return an error instead of assert */ | |
908 | ASSERT(hsize > (1ULL << 10)); | |
909 | hsize >>= 1; | |
910 | goto retry; | |
911 | } | |
912 | ||
d3c2ae1c | 913 | dbuf_kmem_cache = kmem_cache_create("dmu_buf_impl_t", |
34dc7c2f BB |
914 | sizeof (dmu_buf_impl_t), |
915 | 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0); | |
916 | ||
223b04d2 BB |
917 | for (i = 0; i < DBUF_MUTEXES; i++) |
918 | mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL); | |
e0b0ca98 BB |
919 | |
920 | dbuf_stats_init(h); | |
0c66c32d JG |
921 | |
922 | /* | |
923 | * All entries are queued via taskq_dispatch_ent(), so min/maxalloc | |
924 | * configuration is not required. | |
925 | */ | |
1229323d | 926 | dbu_evict_taskq = taskq_create("dbu_evict", 1, defclsyspri, 0, 0, 0); |
d3c2ae1c | 927 | |
2e5dc449 | 928 | for (dbuf_cached_state_t dcs = 0; dcs < DB_CACHE_MAX; dcs++) { |
ffdf019c AM |
929 | multilist_create(&dbuf_caches[dcs].cache, |
930 | sizeof (dmu_buf_impl_t), | |
2e5dc449 MA |
931 | offsetof(dmu_buf_impl_t, db_cache_link), |
932 | dbuf_cache_multilist_index_func); | |
424fd7c3 | 933 | zfs_refcount_create(&dbuf_caches[dcs].size); |
2e5dc449 | 934 | } |
d3c2ae1c | 935 | |
d3c2ae1c GW |
936 | dbuf_evict_thread_exit = B_FALSE; |
937 | mutex_init(&dbuf_evict_lock, NULL, MUTEX_DEFAULT, NULL); | |
938 | cv_init(&dbuf_evict_cv, NULL, CV_DEFAULT, NULL); | |
939 | dbuf_cache_evict_thread = thread_create(NULL, 0, dbuf_evict_thread, | |
940 | NULL, 0, &p0, TS_RUN, minclsyspri); | |
5e021f56 | 941 | |
c4c162c1 AM |
942 | wmsum_init(&dbuf_sums.cache_count, 0); |
943 | wmsum_init(&dbuf_sums.cache_total_evicts, 0); | |
944 | for (i = 0; i < DN_MAX_LEVELS; i++) { | |
945 | wmsum_init(&dbuf_sums.cache_levels[i], 0); | |
946 | wmsum_init(&dbuf_sums.cache_levels_bytes[i], 0); | |
947 | } | |
948 | wmsum_init(&dbuf_sums.hash_hits, 0); | |
949 | wmsum_init(&dbuf_sums.hash_misses, 0); | |
950 | wmsum_init(&dbuf_sums.hash_collisions, 0); | |
951 | wmsum_init(&dbuf_sums.hash_chains, 0); | |
952 | wmsum_init(&dbuf_sums.hash_insert_race, 0); | |
953 | wmsum_init(&dbuf_sums.metadata_cache_count, 0); | |
954 | wmsum_init(&dbuf_sums.metadata_cache_overflow, 0); | |
955 | ||
5e021f56 GDN |
956 | dbuf_ksp = kstat_create("zfs", 0, "dbufstats", "misc", |
957 | KSTAT_TYPE_NAMED, sizeof (dbuf_stats) / sizeof (kstat_named_t), | |
958 | KSTAT_FLAG_VIRTUAL); | |
959 | if (dbuf_ksp != NULL) { | |
5e021f56 GDN |
960 | for (i = 0; i < DN_MAX_LEVELS; i++) { |
961 | snprintf(dbuf_stats.cache_levels[i].name, | |
962 | KSTAT_STRLEN, "cache_level_%d", i); | |
963 | dbuf_stats.cache_levels[i].data_type = | |
964 | KSTAT_DATA_UINT64; | |
965 | snprintf(dbuf_stats.cache_levels_bytes[i].name, | |
966 | KSTAT_STRLEN, "cache_level_%d_bytes", i); | |
967 | dbuf_stats.cache_levels_bytes[i].data_type = | |
968 | KSTAT_DATA_UINT64; | |
969 | } | |
741db5a3 AM |
970 | dbuf_ksp->ks_data = &dbuf_stats; |
971 | dbuf_ksp->ks_update = dbuf_kstat_update; | |
972 | kstat_install(dbuf_ksp); | |
5e021f56 | 973 | } |
34dc7c2f BB |
974 | } |
975 | ||
976 | void | |
977 | dbuf_fini(void) | |
978 | { | |
979 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
980 | int i; | |
981 | ||
e0b0ca98 BB |
982 | dbuf_stats_destroy(); |
983 | ||
223b04d2 BB |
984 | for (i = 0; i < DBUF_MUTEXES; i++) |
985 | mutex_destroy(&h->hash_mutexes[i]); | |
93ce2b4c | 986 | #if defined(_KERNEL) |
d1d7e268 MK |
987 | /* |
988 | * Large allocations which do not require contiguous pages | |
989 | * should be using vmem_free() in the linux kernel | |
990 | */ | |
00b46022 BB |
991 | vmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *)); |
992 | #else | |
34dc7c2f | 993 | kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *)); |
00b46022 | 994 | #endif |
d3c2ae1c | 995 | kmem_cache_destroy(dbuf_kmem_cache); |
0c66c32d | 996 | taskq_destroy(dbu_evict_taskq); |
d3c2ae1c GW |
997 | |
998 | mutex_enter(&dbuf_evict_lock); | |
999 | dbuf_evict_thread_exit = B_TRUE; | |
1000 | while (dbuf_evict_thread_exit) { | |
1001 | cv_signal(&dbuf_evict_cv); | |
1002 | cv_wait(&dbuf_evict_cv, &dbuf_evict_lock); | |
1003 | } | |
1004 | mutex_exit(&dbuf_evict_lock); | |
d3c2ae1c GW |
1005 | |
1006 | mutex_destroy(&dbuf_evict_lock); | |
1007 | cv_destroy(&dbuf_evict_cv); | |
1008 | ||
2e5dc449 | 1009 | for (dbuf_cached_state_t dcs = 0; dcs < DB_CACHE_MAX; dcs++) { |
424fd7c3 | 1010 | zfs_refcount_destroy(&dbuf_caches[dcs].size); |
ffdf019c | 1011 | multilist_destroy(&dbuf_caches[dcs].cache); |
2e5dc449 | 1012 | } |
5e021f56 GDN |
1013 | |
1014 | if (dbuf_ksp != NULL) { | |
1015 | kstat_delete(dbuf_ksp); | |
1016 | dbuf_ksp = NULL; | |
1017 | } | |
c4c162c1 AM |
1018 | |
1019 | wmsum_fini(&dbuf_sums.cache_count); | |
1020 | wmsum_fini(&dbuf_sums.cache_total_evicts); | |
1021 | for (i = 0; i < DN_MAX_LEVELS; i++) { | |
1022 | wmsum_fini(&dbuf_sums.cache_levels[i]); | |
1023 | wmsum_fini(&dbuf_sums.cache_levels_bytes[i]); | |
1024 | } | |
1025 | wmsum_fini(&dbuf_sums.hash_hits); | |
1026 | wmsum_fini(&dbuf_sums.hash_misses); | |
1027 | wmsum_fini(&dbuf_sums.hash_collisions); | |
1028 | wmsum_fini(&dbuf_sums.hash_chains); | |
1029 | wmsum_fini(&dbuf_sums.hash_insert_race); | |
1030 | wmsum_fini(&dbuf_sums.metadata_cache_count); | |
1031 | wmsum_fini(&dbuf_sums.metadata_cache_overflow); | |
34dc7c2f BB |
1032 | } |
1033 | ||
1034 | /* | |
1035 | * Other stuff. | |
1036 | */ | |
1037 | ||
1038 | #ifdef ZFS_DEBUG | |
1039 | static void | |
1040 | dbuf_verify(dmu_buf_impl_t *db) | |
1041 | { | |
572e2857 | 1042 | dnode_t *dn; |
428870ff | 1043 | dbuf_dirty_record_t *dr; |
cccbed9f | 1044 | uint32_t txg_prev; |
34dc7c2f BB |
1045 | |
1046 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
1047 | ||
1048 | if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY)) | |
1049 | return; | |
1050 | ||
1051 | ASSERT(db->db_objset != NULL); | |
572e2857 BB |
1052 | DB_DNODE_ENTER(db); |
1053 | dn = DB_DNODE(db); | |
34dc7c2f BB |
1054 | if (dn == NULL) { |
1055 | ASSERT(db->db_parent == NULL); | |
1056 | ASSERT(db->db_blkptr == NULL); | |
1057 | } else { | |
1058 | ASSERT3U(db->db.db_object, ==, dn->dn_object); | |
1059 | ASSERT3P(db->db_objset, ==, dn->dn_objset); | |
1060 | ASSERT3U(db->db_level, <, dn->dn_nlevels); | |
572e2857 BB |
1061 | ASSERT(db->db_blkid == DMU_BONUS_BLKID || |
1062 | db->db_blkid == DMU_SPILL_BLKID || | |
8951cb8d | 1063 | !avl_is_empty(&dn->dn_dbufs)); |
34dc7c2f | 1064 | } |
428870ff BB |
1065 | if (db->db_blkid == DMU_BONUS_BLKID) { |
1066 | ASSERT(dn != NULL); | |
1067 | ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); | |
1068 | ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID); | |
1069 | } else if (db->db_blkid == DMU_SPILL_BLKID) { | |
34dc7c2f | 1070 | ASSERT(dn != NULL); |
c99c9001 | 1071 | ASSERT0(db->db.db_offset); |
34dc7c2f BB |
1072 | } else { |
1073 | ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size); | |
1074 | } | |
1075 | ||
cccbed9f | 1076 | if ((dr = list_head(&db->db_dirty_records)) != NULL) { |
428870ff | 1077 | ASSERT(dr->dr_dbuf == db); |
cccbed9f MM |
1078 | txg_prev = dr->dr_txg; |
1079 | for (dr = list_next(&db->db_dirty_records, dr); dr != NULL; | |
1080 | dr = list_next(&db->db_dirty_records, dr)) { | |
1081 | ASSERT(dr->dr_dbuf == db); | |
1082 | ASSERT(txg_prev > dr->dr_txg); | |
1083 | txg_prev = dr->dr_txg; | |
1084 | } | |
1085 | } | |
428870ff | 1086 | |
b128c09f BB |
1087 | /* |
1088 | * We can't assert that db_size matches dn_datablksz because it | |
1089 | * can be momentarily different when another thread is doing | |
1090 | * dnode_set_blksz(). | |
1091 | */ | |
1092 | if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) { | |
428870ff | 1093 | dr = db->db_data_pending; |
b128c09f BB |
1094 | /* |
1095 | * It should only be modified in syncing context, so | |
1096 | * make sure we only have one copy of the data. | |
1097 | */ | |
1098 | ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf); | |
34dc7c2f BB |
1099 | } |
1100 | ||
1101 | /* verify db->db_blkptr */ | |
1102 | if (db->db_blkptr) { | |
1103 | if (db->db_parent == dn->dn_dbuf) { | |
1104 | /* db is pointed to by the dnode */ | |
1105 | /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */ | |
9babb374 | 1106 | if (DMU_OBJECT_IS_SPECIAL(db->db.db_object)) |
34dc7c2f BB |
1107 | ASSERT(db->db_parent == NULL); |
1108 | else | |
1109 | ASSERT(db->db_parent != NULL); | |
428870ff BB |
1110 | if (db->db_blkid != DMU_SPILL_BLKID) |
1111 | ASSERT3P(db->db_blkptr, ==, | |
1112 | &dn->dn_phys->dn_blkptr[db->db_blkid]); | |
34dc7c2f BB |
1113 | } else { |
1114 | /* db is pointed to by an indirect block */ | |
2a8ba608 MM |
1115 | int epb __maybe_unused = db->db_parent->db.db_size >> |
1116 | SPA_BLKPTRSHIFT; | |
34dc7c2f BB |
1117 | ASSERT3U(db->db_parent->db_level, ==, db->db_level+1); |
1118 | ASSERT3U(db->db_parent->db.db_object, ==, | |
1119 | db->db.db_object); | |
1120 | /* | |
1121 | * dnode_grow_indblksz() can make this fail if we don't | |
f664f1ee | 1122 | * have the parent's rwlock. XXX indblksz no longer |
34dc7c2f BB |
1123 | * grows. safe to do this now? |
1124 | */ | |
f664f1ee | 1125 | if (RW_LOCK_HELD(&db->db_parent->db_rwlock)) { |
34dc7c2f BB |
1126 | ASSERT3P(db->db_blkptr, ==, |
1127 | ((blkptr_t *)db->db_parent->db.db_data + | |
1128 | db->db_blkid % epb)); | |
1129 | } | |
1130 | } | |
1131 | } | |
1132 | if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) && | |
428870ff BB |
1133 | (db->db_buf == NULL || db->db_buf->b_data) && |
1134 | db->db.db_data && db->db_blkid != DMU_BONUS_BLKID && | |
34dc7c2f BB |
1135 | db->db_state != DB_FILL && !dn->dn_free_txg) { |
1136 | /* | |
1137 | * If the blkptr isn't set but they have nonzero data, | |
1138 | * it had better be dirty, otherwise we'll lose that | |
1139 | * data when we evict this buffer. | |
bc77ba73 PD |
1140 | * |
1141 | * There is an exception to this rule for indirect blocks; in | |
1142 | * this case, if the indirect block is a hole, we fill in a few | |
1143 | * fields on each of the child blocks (importantly, birth time) | |
1144 | * to prevent hole birth times from being lost when you | |
1145 | * partially fill in a hole. | |
34dc7c2f BB |
1146 | */ |
1147 | if (db->db_dirtycnt == 0) { | |
bc77ba73 PD |
1148 | if (db->db_level == 0) { |
1149 | uint64_t *buf = db->db.db_data; | |
1150 | int i; | |
34dc7c2f | 1151 | |
bc77ba73 PD |
1152 | for (i = 0; i < db->db.db_size >> 3; i++) { |
1153 | ASSERT(buf[i] == 0); | |
1154 | } | |
1155 | } else { | |
bc77ba73 PD |
1156 | blkptr_t *bps = db->db.db_data; |
1157 | ASSERT3U(1 << DB_DNODE(db)->dn_indblkshift, ==, | |
1158 | db->db.db_size); | |
1159 | /* | |
1160 | * We want to verify that all the blkptrs in the | |
1161 | * indirect block are holes, but we may have | |
1162 | * automatically set up a few fields for them. | |
1163 | * We iterate through each blkptr and verify | |
1164 | * they only have those fields set. | |
1165 | */ | |
1c27024e | 1166 | for (int i = 0; |
bc77ba73 PD |
1167 | i < db->db.db_size / sizeof (blkptr_t); |
1168 | i++) { | |
1169 | blkptr_t *bp = &bps[i]; | |
1170 | ASSERT(ZIO_CHECKSUM_IS_ZERO( | |
1171 | &bp->blk_cksum)); | |
1172 | ASSERT( | |
1173 | DVA_IS_EMPTY(&bp->blk_dva[0]) && | |
1174 | DVA_IS_EMPTY(&bp->blk_dva[1]) && | |
1175 | DVA_IS_EMPTY(&bp->blk_dva[2])); | |
1176 | ASSERT0(bp->blk_fill); | |
1177 | ASSERT0(bp->blk_pad[0]); | |
1178 | ASSERT0(bp->blk_pad[1]); | |
1179 | ASSERT(!BP_IS_EMBEDDED(bp)); | |
1180 | ASSERT(BP_IS_HOLE(bp)); | |
1181 | ASSERT0(bp->blk_phys_birth); | |
1182 | } | |
34dc7c2f BB |
1183 | } |
1184 | } | |
1185 | } | |
572e2857 | 1186 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
1187 | } |
1188 | #endif | |
1189 | ||
0c66c32d JG |
1190 | static void |
1191 | dbuf_clear_data(dmu_buf_impl_t *db) | |
1192 | { | |
1193 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
1194 | dbuf_evict_user(db); | |
d3c2ae1c | 1195 | ASSERT3P(db->db_buf, ==, NULL); |
0c66c32d | 1196 | db->db.db_data = NULL; |
8b3547a4 | 1197 | if (db->db_state != DB_NOFILL) { |
0c66c32d | 1198 | db->db_state = DB_UNCACHED; |
8b3547a4 MM |
1199 | DTRACE_SET_STATE(db, "clear data"); |
1200 | } | |
0c66c32d JG |
1201 | } |
1202 | ||
34dc7c2f BB |
1203 | static void |
1204 | dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf) | |
1205 | { | |
1206 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
0c66c32d JG |
1207 | ASSERT(buf != NULL); |
1208 | ||
34dc7c2f | 1209 | db->db_buf = buf; |
0c66c32d JG |
1210 | ASSERT(buf->b_data != NULL); |
1211 | db->db.db_data = buf->b_data; | |
34dc7c2f BB |
1212 | } |
1213 | ||
8b3547a4 MM |
1214 | static arc_buf_t * |
1215 | dbuf_alloc_arcbuf(dmu_buf_impl_t *db) | |
1216 | { | |
1217 | spa_t *spa = db->db_objset->os_spa; | |
1218 | ||
1219 | return (arc_alloc_buf(spa, db, DBUF_GET_BUFC_TYPE(db), db->db.db_size)); | |
1220 | } | |
1221 | ||
428870ff BB |
1222 | /* |
1223 | * Loan out an arc_buf for read. Return the loaned arc_buf. | |
1224 | */ | |
1225 | arc_buf_t * | |
1226 | dbuf_loan_arcbuf(dmu_buf_impl_t *db) | |
1227 | { | |
1228 | arc_buf_t *abuf; | |
1229 | ||
d3c2ae1c | 1230 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
428870ff | 1231 | mutex_enter(&db->db_mtx); |
424fd7c3 | 1232 | if (arc_released(db->db_buf) || zfs_refcount_count(&db->db_holds) > 1) { |
428870ff | 1233 | int blksz = db->db.db_size; |
b0bc7a84 | 1234 | spa_t *spa = db->db_objset->os_spa; |
572e2857 | 1235 | |
428870ff | 1236 | mutex_exit(&db->db_mtx); |
2aa34383 | 1237 | abuf = arc_loan_buf(spa, B_FALSE, blksz); |
861166b0 | 1238 | memcpy(abuf->b_data, db->db.db_data, blksz); |
428870ff BB |
1239 | } else { |
1240 | abuf = db->db_buf; | |
1241 | arc_loan_inuse_buf(abuf, db); | |
d3c2ae1c | 1242 | db->db_buf = NULL; |
0c66c32d | 1243 | dbuf_clear_data(db); |
428870ff BB |
1244 | mutex_exit(&db->db_mtx); |
1245 | } | |
1246 | return (abuf); | |
1247 | } | |
1248 | ||
fcff0f35 PD |
1249 | /* |
1250 | * Calculate which level n block references the data at the level 0 offset | |
1251 | * provided. | |
1252 | */ | |
34dc7c2f | 1253 | uint64_t |
031d7c2f | 1254 | dbuf_whichblock(const dnode_t *dn, const int64_t level, const uint64_t offset) |
34dc7c2f | 1255 | { |
fcff0f35 PD |
1256 | if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) { |
1257 | /* | |
1258 | * The level n blkid is equal to the level 0 blkid divided by | |
1259 | * the number of level 0s in a level n block. | |
1260 | * | |
1261 | * The level 0 blkid is offset >> datablkshift = | |
1262 | * offset / 2^datablkshift. | |
1263 | * | |
1264 | * The number of level 0s in a level n is the number of block | |
1265 | * pointers in an indirect block, raised to the power of level. | |
1266 | * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level = | |
1267 | * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)). | |
1268 | * | |
1269 | * Thus, the level n blkid is: offset / | |
fe8a7982 | 1270 | * ((2^datablkshift)*(2^(level*(indblkshift-SPA_BLKPTRSHIFT)))) |
fcff0f35 PD |
1271 | * = offset / 2^(datablkshift + level * |
1272 | * (indblkshift - SPA_BLKPTRSHIFT)) | |
1273 | * = offset >> (datablkshift + level * | |
1274 | * (indblkshift - SPA_BLKPTRSHIFT)) | |
1275 | */ | |
031d7c2f GN |
1276 | |
1277 | const unsigned exp = dn->dn_datablkshift + | |
1278 | level * (dn->dn_indblkshift - SPA_BLKPTRSHIFT); | |
1279 | ||
1280 | if (exp >= 8 * sizeof (offset)) { | |
1281 | /* This only happens on the highest indirection level */ | |
1282 | ASSERT3U(level, ==, dn->dn_nlevels - 1); | |
1283 | return (0); | |
1284 | } | |
1285 | ||
1286 | ASSERT3U(exp, <, 8 * sizeof (offset)); | |
1287 | ||
1288 | return (offset >> exp); | |
34dc7c2f BB |
1289 | } else { |
1290 | ASSERT3U(offset, <, dn->dn_datablksz); | |
1291 | return (0); | |
1292 | } | |
1293 | } | |
1294 | ||
f664f1ee PD |
1295 | /* |
1296 | * This function is used to lock the parent of the provided dbuf. This should be | |
1297 | * used when modifying or reading db_blkptr. | |
1298 | */ | |
1299 | db_lock_type_t | |
a926aab9 | 1300 | dmu_buf_lock_parent(dmu_buf_impl_t *db, krw_t rw, const void *tag) |
f664f1ee PD |
1301 | { |
1302 | enum db_lock_type ret = DLT_NONE; | |
1303 | if (db->db_parent != NULL) { | |
1304 | rw_enter(&db->db_parent->db_rwlock, rw); | |
1305 | ret = DLT_PARENT; | |
1306 | } else if (dmu_objset_ds(db->db_objset) != NULL) { | |
1307 | rrw_enter(&dmu_objset_ds(db->db_objset)->ds_bp_rwlock, rw, | |
1308 | tag); | |
1309 | ret = DLT_OBJSET; | |
1310 | } | |
1311 | /* | |
1312 | * We only return a DLT_NONE lock when it's the top-most indirect block | |
1313 | * of the meta-dnode of the MOS. | |
1314 | */ | |
1315 | return (ret); | |
1316 | } | |
1317 | ||
1318 | /* | |
1319 | * We need to pass the lock type in because it's possible that the block will | |
1320 | * move from being the topmost indirect block in a dnode (and thus, have no | |
1321 | * parent) to not the top-most via an indirection increase. This would cause a | |
1322 | * panic if we didn't pass the lock type in. | |
1323 | */ | |
1324 | void | |
a926aab9 | 1325 | dmu_buf_unlock_parent(dmu_buf_impl_t *db, db_lock_type_t type, const void *tag) |
f664f1ee PD |
1326 | { |
1327 | if (type == DLT_PARENT) | |
1328 | rw_exit(&db->db_parent->db_rwlock); | |
1329 | else if (type == DLT_OBJSET) | |
1330 | rrw_exit(&dmu_objset_ds(db->db_objset)->ds_bp_rwlock, tag); | |
1331 | } | |
1332 | ||
34dc7c2f | 1333 | static void |
d4a72f23 TC |
1334 | dbuf_read_done(zio_t *zio, const zbookmark_phys_t *zb, const blkptr_t *bp, |
1335 | arc_buf_t *buf, void *vdb) | |
34dc7c2f | 1336 | { |
14e4e3cb | 1337 | (void) zb, (void) bp; |
34dc7c2f BB |
1338 | dmu_buf_impl_t *db = vdb; |
1339 | ||
1340 | mutex_enter(&db->db_mtx); | |
1341 | ASSERT3U(db->db_state, ==, DB_READ); | |
1342 | /* | |
1343 | * All reads are synchronous, so we must have a hold on the dbuf | |
1344 | */ | |
424fd7c3 | 1345 | ASSERT(zfs_refcount_count(&db->db_holds) > 0); |
34dc7c2f BB |
1346 | ASSERT(db->db_buf == NULL); |
1347 | ASSERT(db->db.db_data == NULL); | |
c3bd3fb4 TC |
1348 | if (buf == NULL) { |
1349 | /* i/o error */ | |
1350 | ASSERT(zio == NULL || zio->io_error != 0); | |
1351 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); | |
1352 | ASSERT3P(db->db_buf, ==, NULL); | |
1353 | db->db_state = DB_UNCACHED; | |
8b3547a4 | 1354 | DTRACE_SET_STATE(db, "i/o error"); |
c3bd3fb4 TC |
1355 | } else if (db->db_level == 0 && db->db_freed_in_flight) { |
1356 | /* freed in flight */ | |
1357 | ASSERT(zio == NULL || zio->io_error == 0); | |
34dc7c2f | 1358 | arc_release(buf, db); |
861166b0 | 1359 | memset(buf->b_data, 0, db->db.db_size); |
34dc7c2f BB |
1360 | arc_buf_freeze(buf); |
1361 | db->db_freed_in_flight = FALSE; | |
1362 | dbuf_set_data(db, buf); | |
1363 | db->db_state = DB_CACHED; | |
8b3547a4 | 1364 | DTRACE_SET_STATE(db, "freed in flight"); |
c3bd3fb4 TC |
1365 | } else { |
1366 | /* success */ | |
1367 | ASSERT(zio == NULL || zio->io_error == 0); | |
34dc7c2f BB |
1368 | dbuf_set_data(db, buf); |
1369 | db->db_state = DB_CACHED; | |
8b3547a4 | 1370 | DTRACE_SET_STATE(db, "successful read"); |
34dc7c2f BB |
1371 | } |
1372 | cv_broadcast(&db->db_changed); | |
3d503a76 | 1373 | dbuf_rele_and_unlock(db, NULL, B_FALSE); |
34dc7c2f BB |
1374 | } |
1375 | ||
8b3547a4 MM |
1376 | /* |
1377 | * Shortcut for performing reads on bonus dbufs. Returns | |
1378 | * an error if we fail to verify the dnode associated with | |
1379 | * a decrypted block. Otherwise success. | |
1380 | */ | |
1381 | static int | |
1382 | dbuf_read_bonus(dmu_buf_impl_t *db, dnode_t *dn, uint32_t flags) | |
1383 | { | |
1384 | int bonuslen, max_bonuslen, err; | |
1385 | ||
1386 | err = dbuf_read_verify_dnode_crypt(db, flags); | |
1387 | if (err) | |
1388 | return (err); | |
1389 | ||
1390 | bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen); | |
1391 | max_bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots); | |
1392 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
1393 | ASSERT(DB_DNODE_HELD(db)); | |
1394 | ASSERT3U(bonuslen, <=, db->db.db_size); | |
1395 | db->db.db_data = kmem_alloc(max_bonuslen, KM_SLEEP); | |
1396 | arc_space_consume(max_bonuslen, ARC_SPACE_BONUS); | |
1397 | if (bonuslen < max_bonuslen) | |
861166b0 | 1398 | memset(db->db.db_data, 0, max_bonuslen); |
8b3547a4 | 1399 | if (bonuslen) |
861166b0 | 1400 | memcpy(db->db.db_data, DN_BONUS(dn->dn_phys), bonuslen); |
8b3547a4 MM |
1401 | db->db_state = DB_CACHED; |
1402 | DTRACE_SET_STATE(db, "bonus buffer filled"); | |
1403 | return (0); | |
1404 | } | |
1405 | ||
1406 | static void | |
1407 | dbuf_handle_indirect_hole(dmu_buf_impl_t *db, dnode_t *dn) | |
1408 | { | |
1409 | blkptr_t *bps = db->db.db_data; | |
1410 | uint32_t indbs = 1ULL << dn->dn_indblkshift; | |
1411 | int n_bps = indbs >> SPA_BLKPTRSHIFT; | |
1412 | ||
1413 | for (int i = 0; i < n_bps; i++) { | |
1414 | blkptr_t *bp = &bps[i]; | |
1415 | ||
1416 | ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, indbs); | |
1417 | BP_SET_LSIZE(bp, BP_GET_LEVEL(db->db_blkptr) == 1 ? | |
1418 | dn->dn_datablksz : BP_GET_LSIZE(db->db_blkptr)); | |
1419 | BP_SET_TYPE(bp, BP_GET_TYPE(db->db_blkptr)); | |
1420 | BP_SET_LEVEL(bp, BP_GET_LEVEL(db->db_blkptr) - 1); | |
1421 | BP_SET_BIRTH(bp, db->db_blkptr->blk_birth, 0); | |
1422 | } | |
1423 | } | |
1424 | ||
1425 | /* | |
1426 | * Handle reads on dbufs that are holes, if necessary. This function | |
1427 | * requires that the dbuf's mutex is held. Returns success (0) if action | |
1428 | * was taken, ENOENT if no action was taken. | |
1429 | */ | |
1430 | static int | |
14e4e3cb | 1431 | dbuf_read_hole(dmu_buf_impl_t *db, dnode_t *dn) |
8b3547a4 MM |
1432 | { |
1433 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
1434 | ||
1435 | int is_hole = db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr); | |
1436 | /* | |
1437 | * For level 0 blocks only, if the above check fails: | |
1438 | * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync() | |
1439 | * processes the delete record and clears the bp while we are waiting | |
1440 | * for the dn_mtx (resulting in a "no" from block_freed). | |
1441 | */ | |
1442 | if (!is_hole && db->db_level == 0) { | |
1443 | is_hole = dnode_block_freed(dn, db->db_blkid) || | |
1444 | BP_IS_HOLE(db->db_blkptr); | |
1445 | } | |
1446 | ||
1447 | if (is_hole) { | |
1448 | dbuf_set_data(db, dbuf_alloc_arcbuf(db)); | |
861166b0 | 1449 | memset(db->db.db_data, 0, db->db.db_size); |
8b3547a4 MM |
1450 | |
1451 | if (db->db_blkptr != NULL && db->db_level > 0 && | |
1452 | BP_IS_HOLE(db->db_blkptr) && | |
1453 | db->db_blkptr->blk_birth != 0) { | |
1454 | dbuf_handle_indirect_hole(db, dn); | |
1455 | } | |
1456 | db->db_state = DB_CACHED; | |
1457 | DTRACE_SET_STATE(db, "hole read satisfied"); | |
1458 | return (0); | |
1459 | } | |
1460 | return (ENOENT); | |
1461 | } | |
69830602 TC |
1462 | |
1463 | /* | |
1464 | * This function ensures that, when doing a decrypting read of a block, | |
1465 | * we make sure we have decrypted the dnode associated with it. We must do | |
1466 | * this so that we ensure we are fully authenticating the checksum-of-MACs | |
1467 | * tree from the root of the objset down to this block. Indirect blocks are | |
1468 | * always verified against their secure checksum-of-MACs assuming that the | |
1469 | * dnode containing them is correct. Now that we are doing a decrypting read, | |
1470 | * we can be sure that the key is loaded and verify that assumption. This is | |
1471 | * especially important considering that we always read encrypted dnode | |
1472 | * blocks as raw data (without verifying their MACs) to start, and | |
1473 | * decrypt / authenticate them when we need to read an encrypted bonus buffer. | |
1474 | */ | |
1475 | static int | |
1476 | dbuf_read_verify_dnode_crypt(dmu_buf_impl_t *db, uint32_t flags) | |
1477 | { | |
1478 | int err = 0; | |
1479 | objset_t *os = db->db_objset; | |
1480 | arc_buf_t *dnode_abuf; | |
1481 | dnode_t *dn; | |
1482 | zbookmark_phys_t zb; | |
1483 | ||
1484 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
1485 | ||
1486 | if (!os->os_encrypted || os->os_raw_receive || | |
1487 | (flags & DB_RF_NO_DECRYPT) != 0) | |
1488 | return (0); | |
1489 | ||
1490 | DB_DNODE_ENTER(db); | |
1491 | dn = DB_DNODE(db); | |
1492 | dnode_abuf = (dn->dn_dbuf != NULL) ? dn->dn_dbuf->db_buf : NULL; | |
1493 | ||
1494 | if (dnode_abuf == NULL || !arc_is_encrypted(dnode_abuf)) { | |
1495 | DB_DNODE_EXIT(db); | |
1496 | return (0); | |
1497 | } | |
1498 | ||
1499 | SET_BOOKMARK(&zb, dmu_objset_id(os), | |
1500 | DMU_META_DNODE_OBJECT, 0, dn->dn_dbuf->db_blkid); | |
1501 | err = arc_untransform(dnode_abuf, os->os_spa, &zb, B_TRUE); | |
1502 | ||
1503 | /* | |
1504 | * An error code of EACCES tells us that the key is still not | |
1505 | * available. This is ok if we are only reading authenticated | |
1506 | * (and therefore non-encrypted) blocks. | |
1507 | */ | |
1508 | if (err == EACCES && ((db->db_blkid != DMU_BONUS_BLKID && | |
1509 | !DMU_OT_IS_ENCRYPTED(dn->dn_type)) || | |
1510 | (db->db_blkid == DMU_BONUS_BLKID && | |
1511 | !DMU_OT_IS_ENCRYPTED(dn->dn_bonustype)))) | |
1512 | err = 0; | |
1513 | ||
69830602 TC |
1514 | DB_DNODE_EXIT(db); |
1515 | ||
1516 | return (err); | |
1517 | } | |
1518 | ||
f664f1ee PD |
1519 | /* |
1520 | * Drops db_mtx and the parent lock specified by dblt and tag before | |
1521 | * returning. | |
1522 | */ | |
5f6d0b6f | 1523 | static int |
f664f1ee | 1524 | dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags, |
a926aab9 | 1525 | db_lock_type_t dblt, const void *tag) |
34dc7c2f | 1526 | { |
572e2857 | 1527 | dnode_t *dn; |
5dbd68a3 | 1528 | zbookmark_phys_t zb; |
2a432414 | 1529 | uint32_t aflags = ARC_FLAG_NOWAIT; |
8b3547a4 | 1530 | int err, zio_flags; |
34dc7c2f | 1531 | |
8b3547a4 | 1532 | err = zio_flags = 0; |
572e2857 BB |
1533 | DB_DNODE_ENTER(db); |
1534 | dn = DB_DNODE(db); | |
424fd7c3 | 1535 | ASSERT(!zfs_refcount_is_zero(&db->db_holds)); |
34dc7c2f BB |
1536 | ASSERT(MUTEX_HELD(&db->db_mtx)); |
1537 | ASSERT(db->db_state == DB_UNCACHED); | |
1538 | ASSERT(db->db_buf == NULL); | |
f664f1ee PD |
1539 | ASSERT(db->db_parent == NULL || |
1540 | RW_LOCK_HELD(&db->db_parent->db_rwlock)); | |
34dc7c2f | 1541 | |
428870ff | 1542 | if (db->db_blkid == DMU_BONUS_BLKID) { |
8b3547a4 MM |
1543 | err = dbuf_read_bonus(db, dn, flags); |
1544 | goto early_unlock; | |
34dc7c2f BB |
1545 | } |
1546 | ||
14e4e3cb | 1547 | err = dbuf_read_hole(db, dn); |
8b3547a4 MM |
1548 | if (err == 0) |
1549 | goto early_unlock; | |
34dc7c2f | 1550 | |
30af21b0 PD |
1551 | /* |
1552 | * Any attempt to read a redacted block should result in an error. This | |
1553 | * will never happen under normal conditions, but can be useful for | |
1554 | * debugging purposes. | |
1555 | */ | |
1556 | if (BP_IS_REDACTED(db->db_blkptr)) { | |
1557 | ASSERT(dsl_dataset_feature_is_active( | |
1558 | db->db_objset->os_dsl_dataset, | |
1559 | SPA_FEATURE_REDACTED_DATASETS)); | |
8b3547a4 MM |
1560 | err = SET_ERROR(EIO); |
1561 | goto early_unlock; | |
30af21b0 PD |
1562 | } |
1563 | ||
370bbf66 TC |
1564 | SET_BOOKMARK(&zb, dmu_objset_id(db->db_objset), |
1565 | db->db.db_object, db->db_level, db->db_blkid); | |
1566 | ||
b5256303 TC |
1567 | /* |
1568 | * All bps of an encrypted os should have the encryption bit set. | |
1569 | * If this is not true it indicates tampering and we report an error. | |
1570 | */ | |
1571 | if (db->db_objset->os_encrypted && !BP_USES_CRYPT(db->db_blkptr)) { | |
1572 | spa_log_error(db->db_objset->os_spa, &zb); | |
1573 | zfs_panic_recover("unencrypted block in encrypted " | |
1574 | "object set %llu", dmu_objset_id(db->db_objset)); | |
8b3547a4 MM |
1575 | err = SET_ERROR(EIO); |
1576 | goto early_unlock; | |
b5256303 TC |
1577 | } |
1578 | ||
69830602 | 1579 | err = dbuf_read_verify_dnode_crypt(db, flags); |
8b3547a4 MM |
1580 | if (err != 0) |
1581 | goto early_unlock; | |
69830602 TC |
1582 | |
1583 | DB_DNODE_EXIT(db); | |
1584 | ||
1585 | db->db_state = DB_READ; | |
8b3547a4 | 1586 | DTRACE_SET_STATE(db, "read issued"); |
69830602 TC |
1587 | mutex_exit(&db->db_mtx); |
1588 | ||
c9d62d13 | 1589 | if (dbuf_is_l2cacheable(db)) |
69830602 TC |
1590 | aflags |= ARC_FLAG_L2CACHE; |
1591 | ||
34dc7c2f | 1592 | dbuf_add_ref(db, NULL); |
b128c09f | 1593 | |
b5256303 TC |
1594 | zio_flags = (flags & DB_RF_CANFAIL) ? |
1595 | ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED; | |
1596 | ||
1597 | if ((flags & DB_RF_NO_DECRYPT) && BP_IS_PROTECTED(db->db_blkptr)) | |
1598 | zio_flags |= ZIO_FLAG_RAW; | |
f664f1ee PD |
1599 | /* |
1600 | * The zio layer will copy the provided blkptr later, but we need to | |
1601 | * do this now so that we can release the parent's rwlock. We have to | |
1602 | * do that now so that if dbuf_read_done is called synchronously (on | |
1603 | * an l1 cache hit) we don't acquire the db_mtx while holding the | |
1604 | * parent's rwlock, which would be a lock ordering violation. | |
1605 | */ | |
1606 | blkptr_t bp = *db->db_blkptr; | |
1607 | dmu_buf_unlock_parent(db, dblt, tag); | |
1608 | (void) arc_read(zio, db->db_objset->os_spa, &bp, | |
b5256303 | 1609 | dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ, zio_flags, |
34dc7c2f | 1610 | &aflags, &zb); |
da8d5748 | 1611 | return (err); |
8b3547a4 MM |
1612 | early_unlock: |
1613 | DB_DNODE_EXIT(db); | |
1614 | mutex_exit(&db->db_mtx); | |
1615 | dmu_buf_unlock_parent(db, dblt, tag); | |
1616 | return (err); | |
34dc7c2f BB |
1617 | } |
1618 | ||
2aa34383 DK |
1619 | /* |
1620 | * This is our just-in-time copy function. It makes a copy of buffers that | |
1621 | * have been modified in a previous transaction group before we access them in | |
1622 | * the current active group. | |
1623 | * | |
1624 | * This function is used in three places: when we are dirtying a buffer for the | |
1625 | * first time in a txg, when we are freeing a range in a dnode that includes | |
1626 | * this buffer, and when we are accessing a buffer which was received compressed | |
1627 | * and later referenced in a WRITE_BYREF record. | |
1628 | * | |
1629 | * Note that when we are called from dbuf_free_range() we do not put a hold on | |
1630 | * the buffer, we just traverse the active dbuf list for the dnode. | |
1631 | */ | |
1632 | static void | |
1633 | dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg) | |
1634 | { | |
cccbed9f | 1635 | dbuf_dirty_record_t *dr = list_head(&db->db_dirty_records); |
2aa34383 DK |
1636 | |
1637 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
1638 | ASSERT(db->db.db_data != NULL); | |
1639 | ASSERT(db->db_level == 0); | |
1640 | ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT); | |
1641 | ||
1642 | if (dr == NULL || | |
1643 | (dr->dt.dl.dr_data != | |
1644 | ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf))) | |
1645 | return; | |
1646 | ||
1647 | /* | |
1648 | * If the last dirty record for this dbuf has not yet synced | |
1649 | * and its referencing the dbuf data, either: | |
1650 | * reset the reference to point to a new copy, | |
1651 | * or (if there a no active holders) | |
1652 | * just null out the current db_data pointer. | |
1653 | */ | |
4807c0ba | 1654 | ASSERT3U(dr->dr_txg, >=, txg - 2); |
2aa34383 | 1655 | if (db->db_blkid == DMU_BONUS_BLKID) { |
2aa34383 DK |
1656 | dnode_t *dn = DB_DNODE(db); |
1657 | int bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots); | |
a3fd9d9e | 1658 | dr->dt.dl.dr_data = kmem_alloc(bonuslen, KM_SLEEP); |
2aa34383 | 1659 | arc_space_consume(bonuslen, ARC_SPACE_BONUS); |
861166b0 | 1660 | memcpy(dr->dt.dl.dr_data, db->db.db_data, bonuslen); |
424fd7c3 | 1661 | } else if (zfs_refcount_count(&db->db_holds) > db->db_dirtycnt) { |
a81b8124 AR |
1662 | dnode_t *dn = DB_DNODE(db); |
1663 | int size = arc_buf_size(db->db_buf); | |
1664 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); | |
1665 | spa_t *spa = db->db_objset->os_spa; | |
1666 | enum zio_compress compress_type = | |
1667 | arc_get_compression(db->db_buf); | |
1668 | uint8_t complevel = arc_get_complevel(db->db_buf); | |
1669 | ||
1670 | if (arc_is_encrypted(db->db_buf)) { | |
1671 | boolean_t byteorder; | |
1672 | uint8_t salt[ZIO_DATA_SALT_LEN]; | |
1673 | uint8_t iv[ZIO_DATA_IV_LEN]; | |
1674 | uint8_t mac[ZIO_DATA_MAC_LEN]; | |
1675 | ||
1676 | arc_get_raw_params(db->db_buf, &byteorder, salt, | |
1677 | iv, mac); | |
1678 | dr->dt.dl.dr_data = arc_alloc_raw_buf(spa, db, | |
1679 | dmu_objset_id(dn->dn_objset), byteorder, salt, iv, | |
1680 | mac, dn->dn_type, size, arc_buf_lsize(db->db_buf), | |
1681 | compress_type, complevel); | |
1682 | } else if (compress_type != ZIO_COMPRESS_OFF) { | |
1683 | ASSERT3U(type, ==, ARC_BUFC_DATA); | |
1684 | dr->dt.dl.dr_data = arc_alloc_compressed_buf(spa, db, | |
1685 | size, arc_buf_lsize(db->db_buf), compress_type, | |
1686 | complevel); | |
1687 | } else { | |
1688 | dr->dt.dl.dr_data = arc_alloc_buf(spa, db, type, size); | |
1689 | } | |
861166b0 | 1690 | memcpy(dr->dt.dl.dr_data->b_data, db->db.db_data, size); |
2aa34383 DK |
1691 | } else { |
1692 | db->db_buf = NULL; | |
1693 | dbuf_clear_data(db); | |
1694 | } | |
1695 | } | |
1696 | ||
34dc7c2f BB |
1697 | int |
1698 | dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) | |
1699 | { | |
1700 | int err = 0; | |
b0bc7a84 | 1701 | boolean_t prefetch; |
572e2857 | 1702 | dnode_t *dn; |
34dc7c2f BB |
1703 | |
1704 | /* | |
1705 | * We don't have to hold the mutex to check db_state because it | |
1706 | * can't be freed while we have a hold on the buffer. | |
1707 | */ | |
424fd7c3 | 1708 | ASSERT(!zfs_refcount_is_zero(&db->db_holds)); |
34dc7c2f | 1709 | |
b128c09f | 1710 | if (db->db_state == DB_NOFILL) |
2e528b49 | 1711 | return (SET_ERROR(EIO)); |
b128c09f | 1712 | |
572e2857 BB |
1713 | DB_DNODE_ENTER(db); |
1714 | dn = DB_DNODE(db); | |
34dc7c2f | 1715 | |
428870ff | 1716 | prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && |
572e2857 | 1717 | (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL && |
b128c09f | 1718 | DBUF_IS_CACHEABLE(db); |
34dc7c2f BB |
1719 | |
1720 | mutex_enter(&db->db_mtx); | |
1721 | if (db->db_state == DB_CACHED) { | |
b5256303 TC |
1722 | spa_t *spa = dn->dn_objset->os_spa; |
1723 | ||
2aa34383 | 1724 | /* |
69830602 TC |
1725 | * Ensure that this block's dnode has been decrypted if |
1726 | * the caller has requested decrypted data. | |
2aa34383 | 1727 | */ |
69830602 TC |
1728 | err = dbuf_read_verify_dnode_crypt(db, flags); |
1729 | ||
1730 | /* | |
1731 | * If the arc buf is compressed or encrypted and the caller | |
1732 | * requested uncompressed data, we need to untransform it | |
1733 | * before returning. We also call arc_untransform() on any | |
1734 | * unauthenticated blocks, which will verify their MAC if | |
1735 | * the key is now available. | |
1736 | */ | |
1737 | if (err == 0 && db->db_buf != NULL && | |
1738 | (flags & DB_RF_NO_DECRYPT) == 0 && | |
b5256303 | 1739 | (arc_is_encrypted(db->db_buf) || |
69830602 | 1740 | arc_is_unauthenticated(db->db_buf) || |
b5256303 | 1741 | arc_get_compression(db->db_buf) != ZIO_COMPRESS_OFF)) { |
a2c2ed1b TC |
1742 | zbookmark_phys_t zb; |
1743 | ||
1744 | SET_BOOKMARK(&zb, dmu_objset_id(db->db_objset), | |
1745 | db->db.db_object, db->db_level, db->db_blkid); | |
b5256303 | 1746 | dbuf_fix_old_data(db, spa_syncing_txg(spa)); |
a2c2ed1b | 1747 | err = arc_untransform(db->db_buf, spa, &zb, B_FALSE); |
2aa34383 DK |
1748 | dbuf_set_data(db, db->db_buf); |
1749 | } | |
34dc7c2f | 1750 | mutex_exit(&db->db_mtx); |
f664f1ee PD |
1751 | if (err == 0 && prefetch) { |
1752 | dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE, | |
891568c9 | 1753 | B_FALSE, flags & DB_RF_HAVESTRUCT); |
f664f1ee | 1754 | } |
572e2857 | 1755 | DB_DNODE_EXIT(db); |
5e021f56 | 1756 | DBUF_STAT_BUMP(hash_hits); |
34dc7c2f | 1757 | } else if (db->db_state == DB_UNCACHED) { |
572e2857 | 1758 | spa_t *spa = dn->dn_objset->os_spa; |
a0043383 | 1759 | boolean_t need_wait = B_FALSE; |
572e2857 | 1760 | |
f664f1ee PD |
1761 | db_lock_type_t dblt = dmu_buf_lock_parent(db, RW_READER, FTAG); |
1762 | ||
b0319c1f | 1763 | if (zio == NULL && |
a0043383 | 1764 | db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) { |
572e2857 | 1765 | zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); |
a0043383 MA |
1766 | need_wait = B_TRUE; |
1767 | } | |
f664f1ee PD |
1768 | err = dbuf_read_impl(db, zio, flags, dblt, FTAG); |
1769 | /* | |
1770 | * dbuf_read_impl has dropped db_mtx and our parent's rwlock | |
1771 | * for us | |
1772 | */ | |
1773 | if (!err && prefetch) { | |
1774 | dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE, | |
891568c9 | 1775 | db->db_state != DB_CACHED, |
f664f1ee PD |
1776 | flags & DB_RF_HAVESTRUCT); |
1777 | } | |
34dc7c2f | 1778 | |
572e2857 | 1779 | DB_DNODE_EXIT(db); |
5e021f56 | 1780 | DBUF_STAT_BUMP(hash_misses); |
34dc7c2f | 1781 | |
5e7f3ace TC |
1782 | /* |
1783 | * If we created a zio_root we must execute it to avoid | |
1784 | * leaking it, even if it isn't attached to any work due | |
1785 | * to an error in dbuf_read_impl(). | |
1786 | */ | |
1787 | if (need_wait) { | |
1788 | if (err == 0) | |
1789 | err = zio_wait(zio); | |
1790 | else | |
1791 | VERIFY0(zio_wait(zio)); | |
1792 | } | |
34dc7c2f | 1793 | } else { |
e49f1e20 WA |
1794 | /* |
1795 | * Another reader came in while the dbuf was in flight | |
1796 | * between UNCACHED and CACHED. Either a writer will finish | |
1797 | * writing the buffer (sending the dbuf to CACHED) or the | |
1798 | * first reader's request will reach the read_done callback | |
1799 | * and send the dbuf to CACHED. Otherwise, a failure | |
1800 | * occurred and the dbuf went to UNCACHED. | |
1801 | */ | |
34dc7c2f | 1802 | mutex_exit(&db->db_mtx); |
f664f1ee PD |
1803 | if (prefetch) { |
1804 | dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE, | |
891568c9 | 1805 | B_TRUE, flags & DB_RF_HAVESTRUCT); |
f664f1ee | 1806 | } |
572e2857 | 1807 | DB_DNODE_EXIT(db); |
5e021f56 | 1808 | DBUF_STAT_BUMP(hash_misses); |
34dc7c2f | 1809 | |
e49f1e20 | 1810 | /* Skip the wait per the caller's request. */ |
34dc7c2f | 1811 | if ((flags & DB_RF_NEVERWAIT) == 0) { |
28caa74b | 1812 | mutex_enter(&db->db_mtx); |
34dc7c2f BB |
1813 | while (db->db_state == DB_READ || |
1814 | db->db_state == DB_FILL) { | |
1815 | ASSERT(db->db_state == DB_READ || | |
1816 | (flags & DB_RF_HAVESTRUCT) == 0); | |
64dbba36 AL |
1817 | DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *, |
1818 | db, zio_t *, zio); | |
34dc7c2f BB |
1819 | cv_wait(&db->db_changed, &db->db_mtx); |
1820 | } | |
1821 | if (db->db_state == DB_UNCACHED) | |
2e528b49 | 1822 | err = SET_ERROR(EIO); |
28caa74b | 1823 | mutex_exit(&db->db_mtx); |
34dc7c2f | 1824 | } |
34dc7c2f BB |
1825 | } |
1826 | ||
34dc7c2f BB |
1827 | return (err); |
1828 | } | |
1829 | ||
1830 | static void | |
1831 | dbuf_noread(dmu_buf_impl_t *db) | |
1832 | { | |
424fd7c3 | 1833 | ASSERT(!zfs_refcount_is_zero(&db->db_holds)); |
428870ff | 1834 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f BB |
1835 | mutex_enter(&db->db_mtx); |
1836 | while (db->db_state == DB_READ || db->db_state == DB_FILL) | |
1837 | cv_wait(&db->db_changed, &db->db_mtx); | |
1838 | if (db->db_state == DB_UNCACHED) { | |
34dc7c2f BB |
1839 | ASSERT(db->db_buf == NULL); |
1840 | ASSERT(db->db.db_data == NULL); | |
8b3547a4 | 1841 | dbuf_set_data(db, dbuf_alloc_arcbuf(db)); |
34dc7c2f | 1842 | db->db_state = DB_FILL; |
8b3547a4 | 1843 | DTRACE_SET_STATE(db, "assigning filled buffer"); |
b128c09f | 1844 | } else if (db->db_state == DB_NOFILL) { |
0c66c32d | 1845 | dbuf_clear_data(db); |
34dc7c2f BB |
1846 | } else { |
1847 | ASSERT3U(db->db_state, ==, DB_CACHED); | |
1848 | } | |
1849 | mutex_exit(&db->db_mtx); | |
1850 | } | |
1851 | ||
34dc7c2f BB |
1852 | void |
1853 | dbuf_unoverride(dbuf_dirty_record_t *dr) | |
1854 | { | |
1855 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
428870ff | 1856 | blkptr_t *bp = &dr->dt.dl.dr_overridden_by; |
34dc7c2f BB |
1857 | uint64_t txg = dr->dr_txg; |
1858 | ||
1859 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
00710365 AS |
1860 | /* |
1861 | * This assert is valid because dmu_sync() expects to be called by | |
1862 | * a zilog's get_data while holding a range lock. This call only | |
1863 | * comes from dbuf_dirty() callers who must also hold a range lock. | |
1864 | */ | |
34dc7c2f BB |
1865 | ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC); |
1866 | ASSERT(db->db_level == 0); | |
1867 | ||
428870ff | 1868 | if (db->db_blkid == DMU_BONUS_BLKID || |
34dc7c2f BB |
1869 | dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN) |
1870 | return; | |
1871 | ||
428870ff BB |
1872 | ASSERT(db->db_data_pending != dr); |
1873 | ||
34dc7c2f | 1874 | /* free this block */ |
b0bc7a84 MG |
1875 | if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite) |
1876 | zio_free(db->db_objset->os_spa, txg, bp); | |
428870ff | 1877 | |
34dc7c2f | 1878 | dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; |
03c6040b | 1879 | dr->dt.dl.dr_nopwrite = B_FALSE; |
0c03d21a | 1880 | dr->dt.dl.dr_has_raw_params = B_FALSE; |
03c6040b | 1881 | |
34dc7c2f BB |
1882 | /* |
1883 | * Release the already-written buffer, so we leave it in | |
1884 | * a consistent dirty state. Note that all callers are | |
1885 | * modifying the buffer, so they will immediately do | |
1886 | * another (redundant) arc_release(). Therefore, leave | |
1887 | * the buf thawed to save the effort of freezing & | |
1888 | * immediately re-thawing it. | |
1889 | */ | |
1890 | arc_release(dr->dt.dl.dr_data, db); | |
1891 | } | |
1892 | ||
b128c09f BB |
1893 | /* |
1894 | * Evict (if its unreferenced) or clear (if its referenced) any level-0 | |
1895 | * data blocks in the free range, so that any future readers will find | |
b0bc7a84 | 1896 | * empty blocks. |
b128c09f | 1897 | */ |
34dc7c2f | 1898 | void |
8951cb8d AR |
1899 | dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid, |
1900 | dmu_tx_t *tx) | |
34dc7c2f | 1901 | { |
0c66c32d JG |
1902 | dmu_buf_impl_t *db_search; |
1903 | dmu_buf_impl_t *db, *db_next; | |
34dc7c2f | 1904 | uint64_t txg = tx->tx_txg; |
8951cb8d | 1905 | avl_index_t where; |
7b49bbc8 | 1906 | dbuf_dirty_record_t *dr; |
8951cb8d | 1907 | |
9c9531cb GM |
1908 | if (end_blkid > dn->dn_maxblkid && |
1909 | !(start_blkid == DMU_SPILL_BLKID || end_blkid == DMU_SPILL_BLKID)) | |
8951cb8d | 1910 | end_blkid = dn->dn_maxblkid; |
8e739b2c RE |
1911 | dprintf_dnode(dn, "start=%llu end=%llu\n", (u_longlong_t)start_blkid, |
1912 | (u_longlong_t)end_blkid); | |
34dc7c2f | 1913 | |
0c66c32d | 1914 | db_search = kmem_alloc(sizeof (dmu_buf_impl_t), KM_SLEEP); |
8951cb8d AR |
1915 | db_search->db_level = 0; |
1916 | db_search->db_blkid = start_blkid; | |
9925c28c | 1917 | db_search->db_state = DB_SEARCH; |
ea97f8ce | 1918 | |
b663a23d | 1919 | mutex_enter(&dn->dn_dbufs_mtx); |
8951cb8d AR |
1920 | db = avl_find(&dn->dn_dbufs, db_search, &where); |
1921 | ASSERT3P(db, ==, NULL); | |
9c9531cb | 1922 | |
8951cb8d AR |
1923 | db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER); |
1924 | ||
1925 | for (; db != NULL; db = db_next) { | |
1926 | db_next = AVL_NEXT(&dn->dn_dbufs, db); | |
428870ff | 1927 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
b128c09f | 1928 | |
8951cb8d AR |
1929 | if (db->db_level != 0 || db->db_blkid > end_blkid) { |
1930 | break; | |
1931 | } | |
1932 | ASSERT3U(db->db_blkid, >=, start_blkid); | |
34dc7c2f BB |
1933 | |
1934 | /* found a level 0 buffer in the range */ | |
13fe0198 MA |
1935 | mutex_enter(&db->db_mtx); |
1936 | if (dbuf_undirty(db, tx)) { | |
1937 | /* mutex has been dropped and dbuf destroyed */ | |
34dc7c2f | 1938 | continue; |
13fe0198 | 1939 | } |
34dc7c2f | 1940 | |
34dc7c2f | 1941 | if (db->db_state == DB_UNCACHED || |
b128c09f | 1942 | db->db_state == DB_NOFILL || |
34dc7c2f BB |
1943 | db->db_state == DB_EVICTING) { |
1944 | ASSERT(db->db.db_data == NULL); | |
1945 | mutex_exit(&db->db_mtx); | |
1946 | continue; | |
1947 | } | |
1948 | if (db->db_state == DB_READ || db->db_state == DB_FILL) { | |
1949 | /* will be handled in dbuf_read_done or dbuf_rele */ | |
1950 | db->db_freed_in_flight = TRUE; | |
1951 | mutex_exit(&db->db_mtx); | |
1952 | continue; | |
1953 | } | |
424fd7c3 | 1954 | if (zfs_refcount_count(&db->db_holds) == 0) { |
34dc7c2f | 1955 | ASSERT(db->db_buf); |
d3c2ae1c | 1956 | dbuf_destroy(db); |
34dc7c2f BB |
1957 | continue; |
1958 | } | |
1959 | /* The dbuf is referenced */ | |
1960 | ||
7b49bbc8 MM |
1961 | dr = list_head(&db->db_dirty_records); |
1962 | if (dr != NULL) { | |
34dc7c2f BB |
1963 | if (dr->dr_txg == txg) { |
1964 | /* | |
1965 | * This buffer is "in-use", re-adjust the file | |
1966 | * size to reflect that this buffer may | |
1967 | * contain new data when we sync. | |
1968 | */ | |
428870ff BB |
1969 | if (db->db_blkid != DMU_SPILL_BLKID && |
1970 | db->db_blkid > dn->dn_maxblkid) | |
34dc7c2f BB |
1971 | dn->dn_maxblkid = db->db_blkid; |
1972 | dbuf_unoverride(dr); | |
1973 | } else { | |
1974 | /* | |
1975 | * This dbuf is not dirty in the open context. | |
1976 | * Either uncache it (if its not referenced in | |
1977 | * the open context) or reset its contents to | |
1978 | * empty. | |
1979 | */ | |
1980 | dbuf_fix_old_data(db, txg); | |
1981 | } | |
1982 | } | |
1983 | /* clear the contents if its cached */ | |
1984 | if (db->db_state == DB_CACHED) { | |
1985 | ASSERT(db->db.db_data != NULL); | |
1986 | arc_release(db->db_buf, db); | |
f664f1ee | 1987 | rw_enter(&db->db_rwlock, RW_WRITER); |
861166b0 | 1988 | memset(db->db.db_data, 0, db->db.db_size); |
f664f1ee | 1989 | rw_exit(&db->db_rwlock); |
34dc7c2f BB |
1990 | arc_buf_freeze(db->db_buf); |
1991 | } | |
1992 | ||
1993 | mutex_exit(&db->db_mtx); | |
1994 | } | |
8951cb8d | 1995 | |
34dc7c2f | 1996 | mutex_exit(&dn->dn_dbufs_mtx); |
547df816 | 1997 | kmem_free(db_search, sizeof (dmu_buf_impl_t)); |
34dc7c2f BB |
1998 | } |
1999 | ||
34dc7c2f BB |
2000 | void |
2001 | dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx) | |
2002 | { | |
28caa74b | 2003 | arc_buf_t *buf, *old_buf; |
cccbed9f | 2004 | dbuf_dirty_record_t *dr; |
34dc7c2f BB |
2005 | int osize = db->db.db_size; |
2006 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); | |
572e2857 | 2007 | dnode_t *dn; |
34dc7c2f | 2008 | |
428870ff | 2009 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f | 2010 | |
572e2857 BB |
2011 | DB_DNODE_ENTER(db); |
2012 | dn = DB_DNODE(db); | |
2013 | ||
34dc7c2f BB |
2014 | /* |
2015 | * XXX we should be doing a dbuf_read, checking the return | |
2016 | * value and returning that up to our callers | |
2017 | */ | |
b0bc7a84 | 2018 | dmu_buf_will_dirty(&db->db, tx); |
34dc7c2f BB |
2019 | |
2020 | /* create the data buffer for the new block */ | |
2aa34383 | 2021 | buf = arc_alloc_buf(dn->dn_objset->os_spa, db, type, size); |
34dc7c2f BB |
2022 | |
2023 | /* copy old block data to the new block */ | |
28caa74b | 2024 | old_buf = db->db_buf; |
861166b0 | 2025 | memcpy(buf->b_data, old_buf->b_data, MIN(osize, size)); |
34dc7c2f BB |
2026 | /* zero the remainder */ |
2027 | if (size > osize) | |
861166b0 | 2028 | memset((uint8_t *)buf->b_data + osize, 0, size - osize); |
34dc7c2f BB |
2029 | |
2030 | mutex_enter(&db->db_mtx); | |
2031 | dbuf_set_data(db, buf); | |
28caa74b | 2032 | arc_buf_destroy(old_buf, db); |
34dc7c2f BB |
2033 | db->db.db_size = size; |
2034 | ||
cccbed9f | 2035 | dr = list_head(&db->db_dirty_records); |
7b49bbc8 MM |
2036 | /* dirty record added by dmu_buf_will_dirty() */ |
2037 | VERIFY(dr != NULL); | |
cccbed9f MM |
2038 | if (db->db_level == 0) |
2039 | dr->dt.dl.dr_data = buf; | |
2040 | ASSERT3U(dr->dr_txg, ==, tx->tx_txg); | |
2041 | ASSERT3U(dr->dr_accounted, ==, osize); | |
2042 | dr->dr_accounted = size; | |
34dc7c2f BB |
2043 | mutex_exit(&db->db_mtx); |
2044 | ||
3ec3bc21 | 2045 | dmu_objset_willuse_space(dn->dn_objset, size - osize, tx); |
572e2857 | 2046 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
2047 | } |
2048 | ||
428870ff BB |
2049 | void |
2050 | dbuf_release_bp(dmu_buf_impl_t *db) | |
2051 | { | |
2a8ba608 | 2052 | objset_t *os __maybe_unused = db->db_objset; |
428870ff BB |
2053 | |
2054 | ASSERT(dsl_pool_sync_context(dmu_objset_pool(os))); | |
2055 | ASSERT(arc_released(os->os_phys_buf) || | |
2056 | list_link_active(&os->os_dsl_dataset->ds_synced_link)); | |
2057 | ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf)); | |
2058 | ||
294f6806 | 2059 | (void) arc_release(db->db_buf, db); |
428870ff BB |
2060 | } |
2061 | ||
5a28a973 MA |
2062 | /* |
2063 | * We already have a dirty record for this TXG, and we are being | |
2064 | * dirtied again. | |
2065 | */ | |
2066 | static void | |
2067 | dbuf_redirty(dbuf_dirty_record_t *dr) | |
2068 | { | |
2069 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
2070 | ||
2071 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
2072 | ||
2073 | if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) { | |
2074 | /* | |
2075 | * If this buffer has already been written out, | |
2076 | * we now need to reset its state. | |
2077 | */ | |
2078 | dbuf_unoverride(dr); | |
2079 | if (db->db.db_object != DMU_META_DNODE_OBJECT && | |
2080 | db->db_state != DB_NOFILL) { | |
2081 | /* Already released on initial dirty, so just thaw. */ | |
2082 | ASSERT(arc_released(db->db_buf)); | |
2083 | arc_buf_thaw(db->db_buf); | |
2084 | } | |
2085 | } | |
2086 | } | |
2087 | ||
ba67d821 MA |
2088 | dbuf_dirty_record_t * |
2089 | dbuf_dirty_lightweight(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx) | |
2090 | { | |
2091 | rw_enter(&dn->dn_struct_rwlock, RW_READER); | |
2092 | IMPLY(dn->dn_objset->os_raw_receive, dn->dn_maxblkid >= blkid); | |
2093 | dnode_new_blkid(dn, blkid, tx, B_TRUE, B_FALSE); | |
2094 | ASSERT(dn->dn_maxblkid >= blkid); | |
2095 | ||
2096 | dbuf_dirty_record_t *dr = kmem_zalloc(sizeof (*dr), KM_SLEEP); | |
2097 | list_link_init(&dr->dr_dirty_node); | |
2098 | list_link_init(&dr->dr_dbuf_node); | |
2099 | dr->dr_dnode = dn; | |
2100 | dr->dr_txg = tx->tx_txg; | |
2101 | dr->dt.dll.dr_blkid = blkid; | |
2102 | dr->dr_accounted = dn->dn_datablksz; | |
2103 | ||
2104 | /* | |
2105 | * There should not be any dbuf for the block that we're dirtying. | |
2106 | * Otherwise the buffer contents could be inconsistent between the | |
2107 | * dbuf and the lightweight dirty record. | |
2108 | */ | |
2109 | ASSERT3P(NULL, ==, dbuf_find(dn->dn_objset, dn->dn_object, 0, blkid)); | |
2110 | ||
2111 | mutex_enter(&dn->dn_mtx); | |
2112 | int txgoff = tx->tx_txg & TXG_MASK; | |
2113 | if (dn->dn_free_ranges[txgoff] != NULL) { | |
2114 | range_tree_clear(dn->dn_free_ranges[txgoff], blkid, 1); | |
2115 | } | |
2116 | ||
2117 | if (dn->dn_nlevels == 1) { | |
2118 | ASSERT3U(blkid, <, dn->dn_nblkptr); | |
2119 | list_insert_tail(&dn->dn_dirty_records[txgoff], dr); | |
2120 | mutex_exit(&dn->dn_mtx); | |
2121 | rw_exit(&dn->dn_struct_rwlock); | |
2122 | dnode_setdirty(dn, tx); | |
2123 | } else { | |
2124 | mutex_exit(&dn->dn_mtx); | |
2125 | ||
2126 | int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; | |
2127 | dmu_buf_impl_t *parent_db = dbuf_hold_level(dn, | |
2128 | 1, blkid >> epbs, FTAG); | |
2129 | rw_exit(&dn->dn_struct_rwlock); | |
2130 | if (parent_db == NULL) { | |
2131 | kmem_free(dr, sizeof (*dr)); | |
2132 | return (NULL); | |
2133 | } | |
2134 | int err = dbuf_read(parent_db, NULL, | |
2135 | (DB_RF_NOPREFETCH | DB_RF_CANFAIL)); | |
2136 | if (err != 0) { | |
2137 | dbuf_rele(parent_db, FTAG); | |
2138 | kmem_free(dr, sizeof (*dr)); | |
2139 | return (NULL); | |
2140 | } | |
2141 | ||
2142 | dbuf_dirty_record_t *parent_dr = dbuf_dirty(parent_db, tx); | |
2143 | dbuf_rele(parent_db, FTAG); | |
2144 | mutex_enter(&parent_dr->dt.di.dr_mtx); | |
2145 | ASSERT3U(parent_dr->dr_txg, ==, tx->tx_txg); | |
2146 | list_insert_tail(&parent_dr->dt.di.dr_children, dr); | |
2147 | mutex_exit(&parent_dr->dt.di.dr_mtx); | |
2148 | dr->dr_parent = parent_dr; | |
2149 | } | |
2150 | ||
2151 | dmu_objset_willuse_space(dn->dn_objset, dr->dr_accounted, tx); | |
2152 | ||
2153 | return (dr); | |
2154 | } | |
2155 | ||
34dc7c2f BB |
2156 | dbuf_dirty_record_t * |
2157 | dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) | |
2158 | { | |
572e2857 BB |
2159 | dnode_t *dn; |
2160 | objset_t *os; | |
cccbed9f | 2161 | dbuf_dirty_record_t *dr, *dr_next, *dr_head; |
34dc7c2f | 2162 | int txgoff = tx->tx_txg & TXG_MASK; |
f664f1ee | 2163 | boolean_t drop_struct_rwlock = B_FALSE; |
34dc7c2f BB |
2164 | |
2165 | ASSERT(tx->tx_txg != 0); | |
424fd7c3 | 2166 | ASSERT(!zfs_refcount_is_zero(&db->db_holds)); |
34dc7c2f BB |
2167 | DMU_TX_DIRTY_BUF(tx, db); |
2168 | ||
572e2857 BB |
2169 | DB_DNODE_ENTER(db); |
2170 | dn = DB_DNODE(db); | |
34dc7c2f BB |
2171 | /* |
2172 | * Shouldn't dirty a regular buffer in syncing context. Private | |
2173 | * objects may be dirtied in syncing context, but only if they | |
2174 | * were already pre-dirtied in open context. | |
34dc7c2f | 2175 | */ |
6d8da841 | 2176 | #ifdef ZFS_DEBUG |
cc9bb3e5 GM |
2177 | if (dn->dn_objset->os_dsl_dataset != NULL) { |
2178 | rrw_enter(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, | |
2179 | RW_READER, FTAG); | |
2180 | } | |
34dc7c2f BB |
2181 | ASSERT(!dmu_tx_is_syncing(tx) || |
2182 | BP_IS_HOLE(dn->dn_objset->os_rootbp) || | |
9babb374 BB |
2183 | DMU_OBJECT_IS_SPECIAL(dn->dn_object) || |
2184 | dn->dn_objset->os_dsl_dataset == NULL); | |
cc9bb3e5 GM |
2185 | if (dn->dn_objset->os_dsl_dataset != NULL) |
2186 | rrw_exit(&dn->dn_objset->os_dsl_dataset->ds_bp_rwlock, FTAG); | |
2187 | #endif | |
34dc7c2f BB |
2188 | /* |
2189 | * We make this assert for private objects as well, but after we | |
2190 | * check if we're already dirty. They are allowed to re-dirty | |
2191 | * in syncing context. | |
2192 | */ | |
2193 | ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || | |
2194 | dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == | |
2195 | (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); | |
2196 | ||
2197 | mutex_enter(&db->db_mtx); | |
2198 | /* | |
2199 | * XXX make this true for indirects too? The problem is that | |
2200 | * transactions created with dmu_tx_create_assigned() from | |
2201 | * syncing context don't bother holding ahead. | |
2202 | */ | |
2203 | ASSERT(db->db_level != 0 || | |
b128c09f BB |
2204 | db->db_state == DB_CACHED || db->db_state == DB_FILL || |
2205 | db->db_state == DB_NOFILL); | |
34dc7c2f BB |
2206 | |
2207 | mutex_enter(&dn->dn_mtx); | |
28caa74b | 2208 | dnode_set_dirtyctx(dn, tx, db); |
edc1e713 TC |
2209 | if (tx->tx_txg > dn->dn_dirty_txg) |
2210 | dn->dn_dirty_txg = tx->tx_txg; | |
34dc7c2f BB |
2211 | mutex_exit(&dn->dn_mtx); |
2212 | ||
428870ff BB |
2213 | if (db->db_blkid == DMU_SPILL_BLKID) |
2214 | dn->dn_have_spill = B_TRUE; | |
2215 | ||
34dc7c2f BB |
2216 | /* |
2217 | * If this buffer is already dirty, we're done. | |
2218 | */ | |
cccbed9f MM |
2219 | dr_head = list_head(&db->db_dirty_records); |
2220 | ASSERT(dr_head == NULL || dr_head->dr_txg <= tx->tx_txg || | |
34dc7c2f | 2221 | db->db.db_object == DMU_META_DNODE_OBJECT); |
cccbed9f MM |
2222 | dr_next = dbuf_find_dirty_lte(db, tx->tx_txg); |
2223 | if (dr_next && dr_next->dr_txg == tx->tx_txg) { | |
572e2857 BB |
2224 | DB_DNODE_EXIT(db); |
2225 | ||
cccbed9f | 2226 | dbuf_redirty(dr_next); |
34dc7c2f | 2227 | mutex_exit(&db->db_mtx); |
cccbed9f | 2228 | return (dr_next); |
34dc7c2f BB |
2229 | } |
2230 | ||
2231 | /* | |
2232 | * Only valid if not already dirty. | |
2233 | */ | |
9babb374 BB |
2234 | ASSERT(dn->dn_object == 0 || |
2235 | dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == | |
34dc7c2f BB |
2236 | (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); |
2237 | ||
2238 | ASSERT3U(dn->dn_nlevels, >, db->db_level); | |
34dc7c2f BB |
2239 | |
2240 | /* | |
2241 | * We should only be dirtying in syncing context if it's the | |
9babb374 BB |
2242 | * mos or we're initializing the os or it's a special object. |
2243 | * However, we are allowed to dirty in syncing context provided | |
2244 | * we already dirtied it in open context. Hence we must make | |
2245 | * this assertion only if we're not already dirty. | |
34dc7c2f | 2246 | */ |
572e2857 | 2247 | os = dn->dn_objset; |
3b7f360c | 2248 | VERIFY3U(tx->tx_txg, <=, spa_final_dirty_txg(os->os_spa)); |
6d8da841 | 2249 | #ifdef ZFS_DEBUG |
cc9bb3e5 GM |
2250 | if (dn->dn_objset->os_dsl_dataset != NULL) |
2251 | rrw_enter(&os->os_dsl_dataset->ds_bp_rwlock, RW_READER, FTAG); | |
9babb374 BB |
2252 | ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) || |
2253 | os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp)); | |
cc9bb3e5 GM |
2254 | if (dn->dn_objset->os_dsl_dataset != NULL) |
2255 | rrw_exit(&os->os_dsl_dataset->ds_bp_rwlock, FTAG); | |
2256 | #endif | |
34dc7c2f BB |
2257 | ASSERT(db->db.db_size != 0); |
2258 | ||
2259 | dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); | |
2260 | ||
428870ff | 2261 | if (db->db_blkid != DMU_BONUS_BLKID) { |
3ec3bc21 | 2262 | dmu_objset_willuse_space(os, db->db.db_size, tx); |
34dc7c2f BB |
2263 | } |
2264 | ||
2265 | /* | |
2266 | * If this buffer is dirty in an old transaction group we need | |
2267 | * to make a copy of it so that the changes we make in this | |
2268 | * transaction group won't leak out when we sync the older txg. | |
2269 | */ | |
79c76d5b | 2270 | dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP); |
98f72a53 | 2271 | list_link_init(&dr->dr_dirty_node); |
cccbed9f | 2272 | list_link_init(&dr->dr_dbuf_node); |
ba67d821 | 2273 | dr->dr_dnode = dn; |
34dc7c2f BB |
2274 | if (db->db_level == 0) { |
2275 | void *data_old = db->db_buf; | |
2276 | ||
b128c09f | 2277 | if (db->db_state != DB_NOFILL) { |
428870ff | 2278 | if (db->db_blkid == DMU_BONUS_BLKID) { |
b128c09f BB |
2279 | dbuf_fix_old_data(db, tx->tx_txg); |
2280 | data_old = db->db.db_data; | |
2281 | } else if (db->db.db_object != DMU_META_DNODE_OBJECT) { | |
2282 | /* | |
2283 | * Release the data buffer from the cache so | |
2284 | * that we can modify it without impacting | |
2285 | * possible other users of this cached data | |
2286 | * block. Note that indirect blocks and | |
2287 | * private objects are not released until the | |
2288 | * syncing state (since they are only modified | |
2289 | * then). | |
2290 | */ | |
2291 | arc_release(db->db_buf, db); | |
2292 | dbuf_fix_old_data(db, tx->tx_txg); | |
2293 | data_old = db->db_buf; | |
2294 | } | |
2295 | ASSERT(data_old != NULL); | |
34dc7c2f | 2296 | } |
34dc7c2f BB |
2297 | dr->dt.dl.dr_data = data_old; |
2298 | } else { | |
448d7aaa | 2299 | mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_NOLOCKDEP, NULL); |
34dc7c2f BB |
2300 | list_create(&dr->dt.di.dr_children, |
2301 | sizeof (dbuf_dirty_record_t), | |
2302 | offsetof(dbuf_dirty_record_t, dr_dirty_node)); | |
2303 | } | |
0f8ff49e | 2304 | if (db->db_blkid != DMU_BONUS_BLKID) |
e8b96c60 | 2305 | dr->dr_accounted = db->db.db_size; |
34dc7c2f BB |
2306 | dr->dr_dbuf = db; |
2307 | dr->dr_txg = tx->tx_txg; | |
cccbed9f | 2308 | list_insert_before(&db->db_dirty_records, dr_next, dr); |
34dc7c2f BB |
2309 | |
2310 | /* | |
2311 | * We could have been freed_in_flight between the dbuf_noread | |
2312 | * and dbuf_dirty. We win, as though the dbuf_noread() had | |
2313 | * happened after the free. | |
2314 | */ | |
428870ff BB |
2315 | if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && |
2316 | db->db_blkid != DMU_SPILL_BLKID) { | |
34dc7c2f | 2317 | mutex_enter(&dn->dn_mtx); |
9bd274dd MA |
2318 | if (dn->dn_free_ranges[txgoff] != NULL) { |
2319 | range_tree_clear(dn->dn_free_ranges[txgoff], | |
2320 | db->db_blkid, 1); | |
2321 | } | |
34dc7c2f BB |
2322 | mutex_exit(&dn->dn_mtx); |
2323 | db->db_freed_in_flight = FALSE; | |
2324 | } | |
2325 | ||
2326 | /* | |
2327 | * This buffer is now part of this txg | |
2328 | */ | |
2329 | dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg); | |
2330 | db->db_dirtycnt += 1; | |
2331 | ASSERT3U(db->db_dirtycnt, <=, 3); | |
2332 | ||
2333 | mutex_exit(&db->db_mtx); | |
2334 | ||
428870ff BB |
2335 | if (db->db_blkid == DMU_BONUS_BLKID || |
2336 | db->db_blkid == DMU_SPILL_BLKID) { | |
34dc7c2f BB |
2337 | mutex_enter(&dn->dn_mtx); |
2338 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
2339 | list_insert_tail(&dn->dn_dirty_records[txgoff], dr); | |
2340 | mutex_exit(&dn->dn_mtx); | |
2341 | dnode_setdirty(dn, tx); | |
572e2857 | 2342 | DB_DNODE_EXIT(db); |
34dc7c2f | 2343 | return (dr); |
98ace739 MA |
2344 | } |
2345 | ||
98ace739 MA |
2346 | if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) { |
2347 | rw_enter(&dn->dn_struct_rwlock, RW_READER); | |
f664f1ee PD |
2348 | drop_struct_rwlock = B_TRUE; |
2349 | } | |
2350 | ||
2351 | /* | |
2352 | * If we are overwriting a dedup BP, then unless it is snapshotted, | |
2353 | * when we get to syncing context we will need to decrement its | |
2354 | * refcount in the DDT. Prefetch the relevant DDT block so that | |
2355 | * syncing context won't have to wait for the i/o. | |
2356 | */ | |
2357 | if (db->db_blkptr != NULL) { | |
2358 | db_lock_type_t dblt = dmu_buf_lock_parent(db, RW_READER, FTAG); | |
2359 | ddt_prefetch(os->os_spa, db->db_blkptr); | |
2360 | dmu_buf_unlock_parent(db, dblt, FTAG); | |
98ace739 MA |
2361 | } |
2362 | ||
2ade4a99 MA |
2363 | /* |
2364 | * We need to hold the dn_struct_rwlock to make this assertion, | |
2365 | * because it protects dn_phys / dn_next_nlevels from changing. | |
2366 | */ | |
2367 | ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) || | |
2368 | dn->dn_phys->dn_nlevels > db->db_level || | |
2369 | dn->dn_next_nlevels[txgoff] > db->db_level || | |
2370 | dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level || | |
2371 | dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level); | |
2372 | ||
34dc7c2f | 2373 | |
b128c09f | 2374 | if (db->db_level == 0) { |
69830602 TC |
2375 | ASSERT(!db->db_objset->os_raw_receive || |
2376 | dn->dn_maxblkid >= db->db_blkid); | |
369aa501 | 2377 | dnode_new_blkid(dn, db->db_blkid, tx, |
f664f1ee | 2378 | drop_struct_rwlock, B_FALSE); |
b128c09f BB |
2379 | ASSERT(dn->dn_maxblkid >= db->db_blkid); |
2380 | } | |
2381 | ||
34dc7c2f BB |
2382 | if (db->db_level+1 < dn->dn_nlevels) { |
2383 | dmu_buf_impl_t *parent = db->db_parent; | |
2384 | dbuf_dirty_record_t *di; | |
2385 | int parent_held = FALSE; | |
2386 | ||
2387 | if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) { | |
2388 | int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; | |
f664f1ee | 2389 | parent = dbuf_hold_level(dn, db->db_level + 1, |
34dc7c2f | 2390 | db->db_blkid >> epbs, FTAG); |
428870ff | 2391 | ASSERT(parent != NULL); |
34dc7c2f BB |
2392 | parent_held = TRUE; |
2393 | } | |
f664f1ee | 2394 | if (drop_struct_rwlock) |
34dc7c2f | 2395 | rw_exit(&dn->dn_struct_rwlock); |
f664f1ee | 2396 | ASSERT3U(db->db_level + 1, ==, parent->db_level); |
34dc7c2f BB |
2397 | di = dbuf_dirty(parent, tx); |
2398 | if (parent_held) | |
2399 | dbuf_rele(parent, FTAG); | |
2400 | ||
2401 | mutex_enter(&db->db_mtx); | |
e8b96c60 MA |
2402 | /* |
2403 | * Since we've dropped the mutex, it's possible that | |
2404 | * dbuf_undirty() might have changed this out from under us. | |
2405 | */ | |
cccbed9f | 2406 | if (list_head(&db->db_dirty_records) == dr || |
34dc7c2f BB |
2407 | dn->dn_object == DMU_META_DNODE_OBJECT) { |
2408 | mutex_enter(&di->dt.di.dr_mtx); | |
2409 | ASSERT3U(di->dr_txg, ==, tx->tx_txg); | |
2410 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
2411 | list_insert_tail(&di->dt.di.dr_children, dr); | |
2412 | mutex_exit(&di->dt.di.dr_mtx); | |
2413 | dr->dr_parent = di; | |
2414 | } | |
2415 | mutex_exit(&db->db_mtx); | |
2416 | } else { | |
f664f1ee | 2417 | ASSERT(db->db_level + 1 == dn->dn_nlevels); |
34dc7c2f | 2418 | ASSERT(db->db_blkid < dn->dn_nblkptr); |
572e2857 | 2419 | ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf); |
34dc7c2f BB |
2420 | mutex_enter(&dn->dn_mtx); |
2421 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
2422 | list_insert_tail(&dn->dn_dirty_records[txgoff], dr); | |
2423 | mutex_exit(&dn->dn_mtx); | |
f664f1ee | 2424 | if (drop_struct_rwlock) |
34dc7c2f BB |
2425 | rw_exit(&dn->dn_struct_rwlock); |
2426 | } | |
2427 | ||
2428 | dnode_setdirty(dn, tx); | |
572e2857 | 2429 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
2430 | return (dr); |
2431 | } | |
2432 | ||
fa3922df MM |
2433 | static void |
2434 | dbuf_undirty_bonus(dbuf_dirty_record_t *dr) | |
2435 | { | |
2436 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
2437 | ||
2438 | if (dr->dt.dl.dr_data != db->db.db_data) { | |
ba67d821 | 2439 | struct dnode *dn = dr->dr_dnode; |
fa3922df MM |
2440 | int max_bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots); |
2441 | ||
2442 | kmem_free(dr->dt.dl.dr_data, max_bonuslen); | |
2443 | arc_space_return(max_bonuslen, ARC_SPACE_BONUS); | |
2444 | } | |
2445 | db->db_data_pending = NULL; | |
2446 | ASSERT(list_next(&db->db_dirty_records, dr) == NULL); | |
2447 | list_remove(&db->db_dirty_records, dr); | |
2448 | if (dr->dr_dbuf->db_level != 0) { | |
2449 | mutex_destroy(&dr->dt.di.dr_mtx); | |
2450 | list_destroy(&dr->dt.di.dr_children); | |
2451 | } | |
2452 | kmem_free(dr, sizeof (dbuf_dirty_record_t)); | |
2453 | ASSERT3U(db->db_dirtycnt, >, 0); | |
2454 | db->db_dirtycnt -= 1; | |
2455 | } | |
2456 | ||
13fe0198 | 2457 | /* |
e49f1e20 WA |
2458 | * Undirty a buffer in the transaction group referenced by the given |
2459 | * transaction. Return whether this evicted the dbuf. | |
13fe0198 MA |
2460 | */ |
2461 | static boolean_t | |
34dc7c2f BB |
2462 | dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx) |
2463 | { | |
34dc7c2f | 2464 | uint64_t txg = tx->tx_txg; |
34dc7c2f BB |
2465 | |
2466 | ASSERT(txg != 0); | |
4bda3bd0 MA |
2467 | |
2468 | /* | |
2469 | * Due to our use of dn_nlevels below, this can only be called | |
2470 | * in open context, unless we are operating on the MOS. | |
2471 | * From syncing context, dn_nlevels may be different from the | |
2472 | * dn_nlevels used when dbuf was dirtied. | |
2473 | */ | |
2474 | ASSERT(db->db_objset == | |
2475 | dmu_objset_pool(db->db_objset)->dp_meta_objset || | |
2476 | txg != spa_syncing_txg(dmu_objset_spa(db->db_objset))); | |
428870ff | 2477 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
13fe0198 MA |
2478 | ASSERT0(db->db_level); |
2479 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
34dc7c2f | 2480 | |
34dc7c2f BB |
2481 | /* |
2482 | * If this buffer is not dirty, we're done. | |
2483 | */ | |
ba67d821 | 2484 | dbuf_dirty_record_t *dr = dbuf_find_dirty_eq(db, txg); |
cccbed9f | 2485 | if (dr == NULL) |
13fe0198 | 2486 | return (B_FALSE); |
428870ff | 2487 | ASSERT(dr->dr_dbuf == db); |
34dc7c2f | 2488 | |
ba67d821 | 2489 | dnode_t *dn = dr->dr_dnode; |
572e2857 | 2490 | |
34dc7c2f BB |
2491 | dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); |
2492 | ||
2493 | ASSERT(db->db.db_size != 0); | |
2494 | ||
4bda3bd0 MA |
2495 | dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset), |
2496 | dr->dr_accounted, txg); | |
34dc7c2f | 2497 | |
cccbed9f | 2498 | list_remove(&db->db_dirty_records, dr); |
34dc7c2f | 2499 | |
ef3c1dea GR |
2500 | /* |
2501 | * Note that there are three places in dbuf_dirty() | |
2502 | * where this dirty record may be put on a list. | |
2503 | * Make sure to do a list_remove corresponding to | |
2504 | * every one of those list_insert calls. | |
2505 | */ | |
34dc7c2f BB |
2506 | if (dr->dr_parent) { |
2507 | mutex_enter(&dr->dr_parent->dt.di.dr_mtx); | |
2508 | list_remove(&dr->dr_parent->dt.di.dr_children, dr); | |
2509 | mutex_exit(&dr->dr_parent->dt.di.dr_mtx); | |
ef3c1dea | 2510 | } else if (db->db_blkid == DMU_SPILL_BLKID || |
4bda3bd0 | 2511 | db->db_level + 1 == dn->dn_nlevels) { |
b128c09f | 2512 | ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf); |
34dc7c2f BB |
2513 | mutex_enter(&dn->dn_mtx); |
2514 | list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr); | |
2515 | mutex_exit(&dn->dn_mtx); | |
2516 | } | |
2517 | ||
13fe0198 MA |
2518 | if (db->db_state != DB_NOFILL) { |
2519 | dbuf_unoverride(dr); | |
34dc7c2f | 2520 | |
34dc7c2f | 2521 | ASSERT(db->db_buf != NULL); |
13fe0198 MA |
2522 | ASSERT(dr->dt.dl.dr_data != NULL); |
2523 | if (dr->dt.dl.dr_data != db->db_buf) | |
d3c2ae1c | 2524 | arc_buf_destroy(dr->dt.dl.dr_data, db); |
34dc7c2f | 2525 | } |
58c4aa00 | 2526 | |
34dc7c2f BB |
2527 | kmem_free(dr, sizeof (dbuf_dirty_record_t)); |
2528 | ||
2529 | ASSERT(db->db_dirtycnt > 0); | |
2530 | db->db_dirtycnt -= 1; | |
2531 | ||
424fd7c3 | 2532 | if (zfs_refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) { |
d3c2ae1c GW |
2533 | ASSERT(db->db_state == DB_NOFILL || arc_released(db->db_buf)); |
2534 | dbuf_destroy(db); | |
13fe0198 | 2535 | return (B_TRUE); |
34dc7c2f BB |
2536 | } |
2537 | ||
13fe0198 | 2538 | return (B_FALSE); |
34dc7c2f BB |
2539 | } |
2540 | ||
b5256303 TC |
2541 | static void |
2542 | dmu_buf_will_dirty_impl(dmu_buf_t *db_fake, int flags, dmu_tx_t *tx) | |
34dc7c2f | 2543 | { |
b0bc7a84 | 2544 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; |
34dc7c2f BB |
2545 | |
2546 | ASSERT(tx->tx_txg != 0); | |
424fd7c3 | 2547 | ASSERT(!zfs_refcount_is_zero(&db->db_holds)); |
34dc7c2f | 2548 | |
5a28a973 | 2549 | /* |
e1cfd73f | 2550 | * Quick check for dirtiness. For already dirty blocks, this |
5a28a973 MA |
2551 | * reduces runtime of this function by >90%, and overall performance |
2552 | * by 50% for some workloads (e.g. file deletion with indirect blocks | |
2553 | * cached). | |
2554 | */ | |
2555 | mutex_enter(&db->db_mtx); | |
2556 | ||
cccbed9f MM |
2557 | if (db->db_state == DB_CACHED) { |
2558 | dbuf_dirty_record_t *dr = dbuf_find_dirty_eq(db, tx->tx_txg); | |
5a28a973 MA |
2559 | /* |
2560 | * It's possible that it is already dirty but not cached, | |
2561 | * because there are some calls to dbuf_dirty() that don't | |
2562 | * go through dmu_buf_will_dirty(). | |
2563 | */ | |
cccbed9f | 2564 | if (dr != NULL) { |
5a28a973 MA |
2565 | /* This dbuf is already dirty and cached. */ |
2566 | dbuf_redirty(dr); | |
2567 | mutex_exit(&db->db_mtx); | |
2568 | return; | |
2569 | } | |
2570 | } | |
2571 | mutex_exit(&db->db_mtx); | |
2572 | ||
572e2857 BB |
2573 | DB_DNODE_ENTER(db); |
2574 | if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock)) | |
b5256303 | 2575 | flags |= DB_RF_HAVESTRUCT; |
572e2857 | 2576 | DB_DNODE_EXIT(db); |
b5256303 | 2577 | (void) dbuf_read(db, NULL, flags); |
34dc7c2f BB |
2578 | (void) dbuf_dirty(db, tx); |
2579 | } | |
2580 | ||
b5256303 TC |
2581 | void |
2582 | dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx) | |
2583 | { | |
2584 | dmu_buf_will_dirty_impl(db_fake, | |
2585 | DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH, tx); | |
2586 | } | |
2587 | ||
a73e8fdb PZ |
2588 | boolean_t |
2589 | dmu_buf_is_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx) | |
2590 | { | |
2591 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
cccbed9f | 2592 | dbuf_dirty_record_t *dr; |
a73e8fdb PZ |
2593 | |
2594 | mutex_enter(&db->db_mtx); | |
cccbed9f | 2595 | dr = dbuf_find_dirty_eq(db, tx->tx_txg); |
a73e8fdb | 2596 | mutex_exit(&db->db_mtx); |
cccbed9f | 2597 | return (dr != NULL); |
a73e8fdb PZ |
2598 | } |
2599 | ||
b128c09f BB |
2600 | void |
2601 | dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) | |
2602 | { | |
2603 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
2604 | ||
2605 | db->db_state = DB_NOFILL; | |
8b3547a4 | 2606 | DTRACE_SET_STATE(db, "allocating NOFILL buffer"); |
b128c09f BB |
2607 | dmu_buf_will_fill(db_fake, tx); |
2608 | } | |
2609 | ||
34dc7c2f BB |
2610 | void |
2611 | dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) | |
2612 | { | |
2613 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
2614 | ||
428870ff | 2615 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f BB |
2616 | ASSERT(tx->tx_txg != 0); |
2617 | ASSERT(db->db_level == 0); | |
424fd7c3 | 2618 | ASSERT(!zfs_refcount_is_zero(&db->db_holds)); |
34dc7c2f BB |
2619 | |
2620 | ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT || | |
2621 | dmu_tx_private_ok(tx)); | |
2622 | ||
2623 | dbuf_noread(db); | |
2624 | (void) dbuf_dirty(db, tx); | |
2625 | } | |
2626 | ||
b5256303 TC |
2627 | /* |
2628 | * This function is effectively the same as dmu_buf_will_dirty(), but | |
0c03d21a MA |
2629 | * indicates the caller expects raw encrypted data in the db, and provides |
2630 | * the crypt params (byteorder, salt, iv, mac) which should be stored in the | |
2631 | * blkptr_t when this dbuf is written. This is only used for blocks of | |
2632 | * dnodes, during raw receive. | |
b5256303 TC |
2633 | */ |
2634 | void | |
0c03d21a MA |
2635 | dmu_buf_set_crypt_params(dmu_buf_t *db_fake, boolean_t byteorder, |
2636 | const uint8_t *salt, const uint8_t *iv, const uint8_t *mac, dmu_tx_t *tx) | |
b5256303 TC |
2637 | { |
2638 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
2639 | dbuf_dirty_record_t *dr; | |
2640 | ||
0c03d21a MA |
2641 | /* |
2642 | * dr_has_raw_params is only processed for blocks of dnodes | |
2643 | * (see dbuf_sync_dnode_leaf_crypt()). | |
2644 | */ | |
2645 | ASSERT3U(db->db.db_object, ==, DMU_META_DNODE_OBJECT); | |
2646 | ASSERT3U(db->db_level, ==, 0); | |
2647 | ASSERT(db->db_objset->os_raw_receive); | |
2648 | ||
b5256303 TC |
2649 | dmu_buf_will_dirty_impl(db_fake, |
2650 | DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_NO_DECRYPT, tx); | |
2651 | ||
cccbed9f | 2652 | dr = dbuf_find_dirty_eq(db, tx->tx_txg); |
b5256303 TC |
2653 | |
2654 | ASSERT3P(dr, !=, NULL); | |
0c03d21a MA |
2655 | |
2656 | dr->dt.dl.dr_has_raw_params = B_TRUE; | |
2657 | dr->dt.dl.dr_byteorder = byteorder; | |
861166b0 AZ |
2658 | memcpy(dr->dt.dl.dr_salt, salt, ZIO_DATA_SALT_LEN); |
2659 | memcpy(dr->dt.dl.dr_iv, iv, ZIO_DATA_IV_LEN); | |
2660 | memcpy(dr->dt.dl.dr_mac, mac, ZIO_DATA_MAC_LEN); | |
b5256303 TC |
2661 | } |
2662 | ||
30af21b0 PD |
2663 | static void |
2664 | dbuf_override_impl(dmu_buf_impl_t *db, const blkptr_t *bp, dmu_tx_t *tx) | |
2665 | { | |
2666 | struct dirty_leaf *dl; | |
cccbed9f | 2667 | dbuf_dirty_record_t *dr; |
30af21b0 | 2668 | |
cccbed9f MM |
2669 | dr = list_head(&db->db_dirty_records); |
2670 | ASSERT3U(dr->dr_txg, ==, tx->tx_txg); | |
2671 | dl = &dr->dt.dl; | |
30af21b0 PD |
2672 | dl->dr_overridden_by = *bp; |
2673 | dl->dr_override_state = DR_OVERRIDDEN; | |
cccbed9f | 2674 | dl->dr_overridden_by.blk_birth = dr->dr_txg; |
30af21b0 PD |
2675 | } |
2676 | ||
34dc7c2f | 2677 | void |
30af21b0 | 2678 | dmu_buf_fill_done(dmu_buf_t *dbuf, dmu_tx_t *tx) |
34dc7c2f | 2679 | { |
14e4e3cb | 2680 | (void) tx; |
30af21b0 | 2681 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; |
8b3547a4 | 2682 | dbuf_states_t old_state; |
34dc7c2f BB |
2683 | mutex_enter(&db->db_mtx); |
2684 | DBUF_VERIFY(db); | |
2685 | ||
8b3547a4 MM |
2686 | old_state = db->db_state; |
2687 | db->db_state = DB_CACHED; | |
2688 | if (old_state == DB_FILL) { | |
34dc7c2f | 2689 | if (db->db_level == 0 && db->db_freed_in_flight) { |
428870ff | 2690 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f BB |
2691 | /* we were freed while filling */ |
2692 | /* XXX dbuf_undirty? */ | |
861166b0 | 2693 | memset(db->db.db_data, 0, db->db.db_size); |
34dc7c2f | 2694 | db->db_freed_in_flight = FALSE; |
8b3547a4 MM |
2695 | DTRACE_SET_STATE(db, |
2696 | "fill done handling freed in flight"); | |
2697 | } else { | |
2698 | DTRACE_SET_STATE(db, "fill done"); | |
34dc7c2f | 2699 | } |
34dc7c2f BB |
2700 | cv_broadcast(&db->db_changed); |
2701 | } | |
2702 | mutex_exit(&db->db_mtx); | |
2703 | } | |
2704 | ||
9b67f605 MA |
2705 | void |
2706 | dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data, | |
2707 | bp_embedded_type_t etype, enum zio_compress comp, | |
2708 | int uncompressed_size, int compressed_size, int byteorder, | |
2709 | dmu_tx_t *tx) | |
2710 | { | |
2711 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; | |
2712 | struct dirty_leaf *dl; | |
2713 | dmu_object_type_t type; | |
cccbed9f | 2714 | dbuf_dirty_record_t *dr; |
9b67f605 | 2715 | |
241b5415 MA |
2716 | if (etype == BP_EMBEDDED_TYPE_DATA) { |
2717 | ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset), | |
2718 | SPA_FEATURE_EMBEDDED_DATA)); | |
2719 | } | |
2720 | ||
9b67f605 MA |
2721 | DB_DNODE_ENTER(db); |
2722 | type = DB_DNODE(db)->dn_type; | |
2723 | DB_DNODE_EXIT(db); | |
2724 | ||
2725 | ASSERT0(db->db_level); | |
2726 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); | |
2727 | ||
2728 | dmu_buf_will_not_fill(dbuf, tx); | |
2729 | ||
cccbed9f MM |
2730 | dr = list_head(&db->db_dirty_records); |
2731 | ASSERT3U(dr->dr_txg, ==, tx->tx_txg); | |
2732 | dl = &dr->dt.dl; | |
9b67f605 MA |
2733 | encode_embedded_bp_compressed(&dl->dr_overridden_by, |
2734 | data, comp, uncompressed_size, compressed_size); | |
2735 | BPE_SET_ETYPE(&dl->dr_overridden_by, etype); | |
2736 | BP_SET_TYPE(&dl->dr_overridden_by, type); | |
2737 | BP_SET_LEVEL(&dl->dr_overridden_by, 0); | |
2738 | BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder); | |
2739 | ||
2740 | dl->dr_override_state = DR_OVERRIDDEN; | |
cccbed9f | 2741 | dl->dr_overridden_by.blk_birth = dr->dr_txg; |
9b67f605 MA |
2742 | } |
2743 | ||
30af21b0 PD |
2744 | void |
2745 | dmu_buf_redact(dmu_buf_t *dbuf, dmu_tx_t *tx) | |
2746 | { | |
2747 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; | |
2748 | dmu_object_type_t type; | |
2749 | ASSERT(dsl_dataset_feature_is_active(db->db_objset->os_dsl_dataset, | |
2750 | SPA_FEATURE_REDACTED_DATASETS)); | |
2751 | ||
2752 | DB_DNODE_ENTER(db); | |
2753 | type = DB_DNODE(db)->dn_type; | |
2754 | DB_DNODE_EXIT(db); | |
2755 | ||
2756 | ASSERT0(db->db_level); | |
2757 | dmu_buf_will_not_fill(dbuf, tx); | |
2758 | ||
2759 | blkptr_t bp = { { { {0} } } }; | |
2760 | BP_SET_TYPE(&bp, type); | |
2761 | BP_SET_LEVEL(&bp, 0); | |
2762 | BP_SET_BIRTH(&bp, tx->tx_txg, 0); | |
2763 | BP_SET_REDACTED(&bp); | |
2764 | BPE_SET_LSIZE(&bp, dbuf->db_size); | |
2765 | ||
2766 | dbuf_override_impl(db, &bp, tx); | |
2767 | } | |
2768 | ||
9babb374 BB |
2769 | /* |
2770 | * Directly assign a provided arc buf to a given dbuf if it's not referenced | |
2771 | * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf. | |
2772 | */ | |
2773 | void | |
2774 | dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx) | |
2775 | { | |
424fd7c3 | 2776 | ASSERT(!zfs_refcount_is_zero(&db->db_holds)); |
428870ff | 2777 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
9babb374 | 2778 | ASSERT(db->db_level == 0); |
2aa34383 | 2779 | ASSERT3U(dbuf_is_metadata(db), ==, arc_is_metadata(buf)); |
9babb374 | 2780 | ASSERT(buf != NULL); |
caf9dd20 | 2781 | ASSERT3U(arc_buf_lsize(buf), ==, db->db.db_size); |
9babb374 BB |
2782 | ASSERT(tx->tx_txg != 0); |
2783 | ||
2784 | arc_return_buf(buf, db); | |
2785 | ASSERT(arc_released(buf)); | |
2786 | ||
2787 | mutex_enter(&db->db_mtx); | |
2788 | ||
2789 | while (db->db_state == DB_READ || db->db_state == DB_FILL) | |
2790 | cv_wait(&db->db_changed, &db->db_mtx); | |
2791 | ||
2792 | ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED); | |
2793 | ||
2794 | if (db->db_state == DB_CACHED && | |
424fd7c3 | 2795 | zfs_refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) { |
440a3eb9 TC |
2796 | /* |
2797 | * In practice, we will never have a case where we have an | |
2798 | * encrypted arc buffer while additional holds exist on the | |
2799 | * dbuf. We don't handle this here so we simply assert that | |
2800 | * fact instead. | |
2801 | */ | |
2802 | ASSERT(!arc_is_encrypted(buf)); | |
9babb374 BB |
2803 | mutex_exit(&db->db_mtx); |
2804 | (void) dbuf_dirty(db, tx); | |
861166b0 | 2805 | memcpy(db->db.db_data, buf->b_data, db->db.db_size); |
d3c2ae1c | 2806 | arc_buf_destroy(buf, db); |
9babb374 BB |
2807 | return; |
2808 | } | |
2809 | ||
2810 | if (db->db_state == DB_CACHED) { | |
cccbed9f | 2811 | dbuf_dirty_record_t *dr = list_head(&db->db_dirty_records); |
9babb374 BB |
2812 | |
2813 | ASSERT(db->db_buf != NULL); | |
2814 | if (dr != NULL && dr->dr_txg == tx->tx_txg) { | |
2815 | ASSERT(dr->dt.dl.dr_data == db->db_buf); | |
440a3eb9 | 2816 | |
9babb374 BB |
2817 | if (!arc_released(db->db_buf)) { |
2818 | ASSERT(dr->dt.dl.dr_override_state == | |
2819 | DR_OVERRIDDEN); | |
2820 | arc_release(db->db_buf, db); | |
2821 | } | |
2822 | dr->dt.dl.dr_data = buf; | |
d3c2ae1c | 2823 | arc_buf_destroy(db->db_buf, db); |
9babb374 BB |
2824 | } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) { |
2825 | arc_release(db->db_buf, db); | |
d3c2ae1c | 2826 | arc_buf_destroy(db->db_buf, db); |
9babb374 BB |
2827 | } |
2828 | db->db_buf = NULL; | |
2829 | } | |
2830 | ASSERT(db->db_buf == NULL); | |
2831 | dbuf_set_data(db, buf); | |
2832 | db->db_state = DB_FILL; | |
8b3547a4 | 2833 | DTRACE_SET_STATE(db, "filling assigned arcbuf"); |
9babb374 BB |
2834 | mutex_exit(&db->db_mtx); |
2835 | (void) dbuf_dirty(db, tx); | |
b0bc7a84 | 2836 | dmu_buf_fill_done(&db->db, tx); |
9babb374 BB |
2837 | } |
2838 | ||
34dc7c2f | 2839 | void |
d3c2ae1c | 2840 | dbuf_destroy(dmu_buf_impl_t *db) |
34dc7c2f | 2841 | { |
572e2857 | 2842 | dnode_t *dn; |
34dc7c2f | 2843 | dmu_buf_impl_t *parent = db->db_parent; |
572e2857 | 2844 | dmu_buf_impl_t *dndb; |
34dc7c2f BB |
2845 | |
2846 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
424fd7c3 | 2847 | ASSERT(zfs_refcount_is_zero(&db->db_holds)); |
34dc7c2f | 2848 | |
d3c2ae1c GW |
2849 | if (db->db_buf != NULL) { |
2850 | arc_buf_destroy(db->db_buf, db); | |
2851 | db->db_buf = NULL; | |
2852 | } | |
34dc7c2f | 2853 | |
d3c2ae1c GW |
2854 | if (db->db_blkid == DMU_BONUS_BLKID) { |
2855 | int slots = DB_DNODE(db)->dn_num_slots; | |
2856 | int bonuslen = DN_SLOTS_TO_BONUSLEN(slots); | |
b5256303 TC |
2857 | if (db->db.db_data != NULL) { |
2858 | kmem_free(db->db.db_data, bonuslen); | |
2859 | arc_space_return(bonuslen, ARC_SPACE_BONUS); | |
2860 | db->db_state = DB_UNCACHED; | |
8b3547a4 | 2861 | DTRACE_SET_STATE(db, "buffer cleared"); |
b5256303 | 2862 | } |
34dc7c2f BB |
2863 | } |
2864 | ||
d3c2ae1c GW |
2865 | dbuf_clear_data(db); |
2866 | ||
2867 | if (multilist_link_active(&db->db_cache_link)) { | |
2e5dc449 MA |
2868 | ASSERT(db->db_caching_status == DB_DBUF_CACHE || |
2869 | db->db_caching_status == DB_DBUF_METADATA_CACHE); | |
2870 | ||
ffdf019c | 2871 | multilist_remove(&dbuf_caches[db->db_caching_status].cache, db); |
424fd7c3 | 2872 | (void) zfs_refcount_remove_many( |
2e5dc449 | 2873 | &dbuf_caches[db->db_caching_status].size, |
d3c2ae1c | 2874 | db->db.db_size, db); |
2e5dc449 MA |
2875 | |
2876 | if (db->db_caching_status == DB_DBUF_METADATA_CACHE) { | |
2877 | DBUF_STAT_BUMPDOWN(metadata_cache_count); | |
2878 | } else { | |
2879 | DBUF_STAT_BUMPDOWN(cache_levels[db->db_level]); | |
2880 | DBUF_STAT_BUMPDOWN(cache_count); | |
2881 | DBUF_STAT_DECR(cache_levels_bytes[db->db_level], | |
2882 | db->db.db_size); | |
2883 | } | |
2884 | db->db_caching_status = DB_NO_CACHE; | |
d3c2ae1c GW |
2885 | } |
2886 | ||
b128c09f | 2887 | ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL); |
34dc7c2f | 2888 | ASSERT(db->db_data_pending == NULL); |
28caa74b | 2889 | ASSERT(list_is_empty(&db->db_dirty_records)); |
34dc7c2f BB |
2890 | |
2891 | db->db_state = DB_EVICTING; | |
8b3547a4 | 2892 | DTRACE_SET_STATE(db, "buffer eviction started"); |
34dc7c2f BB |
2893 | db->db_blkptr = NULL; |
2894 | ||
d3c2ae1c GW |
2895 | /* |
2896 | * Now that db_state is DB_EVICTING, nobody else can find this via | |
2897 | * the hash table. We can now drop db_mtx, which allows us to | |
2898 | * acquire the dn_dbufs_mtx. | |
2899 | */ | |
2900 | mutex_exit(&db->db_mtx); | |
2901 | ||
572e2857 BB |
2902 | DB_DNODE_ENTER(db); |
2903 | dn = DB_DNODE(db); | |
2904 | dndb = dn->dn_dbuf; | |
d3c2ae1c GW |
2905 | if (db->db_blkid != DMU_BONUS_BLKID) { |
2906 | boolean_t needlock = !MUTEX_HELD(&dn->dn_dbufs_mtx); | |
2907 | if (needlock) | |
a649768a JD |
2908 | mutex_enter_nested(&dn->dn_dbufs_mtx, |
2909 | NESTED_SINGLE); | |
8951cb8d | 2910 | avl_remove(&dn->dn_dbufs, db); |
572e2857 BB |
2911 | membar_producer(); |
2912 | DB_DNODE_EXIT(db); | |
d3c2ae1c GW |
2913 | if (needlock) |
2914 | mutex_exit(&dn->dn_dbufs_mtx); | |
572e2857 BB |
2915 | /* |
2916 | * Decrementing the dbuf count means that the hold corresponding | |
2917 | * to the removed dbuf is no longer discounted in dnode_move(), | |
2918 | * so the dnode cannot be moved until after we release the hold. | |
2919 | * The membar_producer() ensures visibility of the decremented | |
2920 | * value in dnode_move(), since DB_DNODE_EXIT doesn't actually | |
2921 | * release any lock. | |
2922 | */ | |
1fac63e5 | 2923 | mutex_enter(&dn->dn_mtx); |
3d503a76 | 2924 | dnode_rele_and_unlock(dn, db, B_TRUE); |
572e2857 | 2925 | db->db_dnode_handle = NULL; |
d3c2ae1c GW |
2926 | |
2927 | dbuf_hash_remove(db); | |
572e2857 BB |
2928 | } else { |
2929 | DB_DNODE_EXIT(db); | |
34dc7c2f BB |
2930 | } |
2931 | ||
424fd7c3 | 2932 | ASSERT(zfs_refcount_is_zero(&db->db_holds)); |
34dc7c2f | 2933 | |
d3c2ae1c GW |
2934 | db->db_parent = NULL; |
2935 | ||
2936 | ASSERT(db->db_buf == NULL); | |
2937 | ASSERT(db->db.db_data == NULL); | |
2938 | ASSERT(db->db_hash_next == NULL); | |
2939 | ASSERT(db->db_blkptr == NULL); | |
2940 | ASSERT(db->db_data_pending == NULL); | |
2e5dc449 | 2941 | ASSERT3U(db->db_caching_status, ==, DB_NO_CACHE); |
d3c2ae1c GW |
2942 | ASSERT(!multilist_link_active(&db->db_cache_link)); |
2943 | ||
34dc7c2f | 2944 | /* |
572e2857 | 2945 | * If this dbuf is referenced from an indirect dbuf, |
34dc7c2f BB |
2946 | * decrement the ref count on the indirect dbuf. |
2947 | */ | |
1fac63e5 MA |
2948 | if (parent && parent != dndb) { |
2949 | mutex_enter(&parent->db_mtx); | |
3d503a76 | 2950 | dbuf_rele_and_unlock(parent, db, B_TRUE); |
1fac63e5 | 2951 | } |
c175f5eb BB |
2952 | |
2953 | kmem_cache_free(dbuf_kmem_cache, db); | |
2954 | arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF); | |
34dc7c2f BB |
2955 | } |
2956 | ||
fcff0f35 PD |
2957 | /* |
2958 | * Note: While bpp will always be updated if the function returns success, | |
2959 | * parentp will not be updated if the dnode does not have dn_dbuf filled in; | |
9c5167d1 | 2960 | * this happens when the dnode is the meta-dnode, or {user|group|project}used |
fcff0f35 PD |
2961 | * object. |
2962 | */ | |
bf701a83 BB |
2963 | __attribute__((always_inline)) |
2964 | static inline int | |
34dc7c2f | 2965 | dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse, |
adb726eb | 2966 | dmu_buf_impl_t **parentp, blkptr_t **bpp) |
34dc7c2f | 2967 | { |
34dc7c2f BB |
2968 | *parentp = NULL; |
2969 | *bpp = NULL; | |
2970 | ||
428870ff BB |
2971 | ASSERT(blkid != DMU_BONUS_BLKID); |
2972 | ||
2973 | if (blkid == DMU_SPILL_BLKID) { | |
2974 | mutex_enter(&dn->dn_mtx); | |
2975 | if (dn->dn_have_spill && | |
2976 | (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) | |
50c957f7 | 2977 | *bpp = DN_SPILL_BLKPTR(dn->dn_phys); |
428870ff BB |
2978 | else |
2979 | *bpp = NULL; | |
2980 | dbuf_add_ref(dn->dn_dbuf, NULL); | |
2981 | *parentp = dn->dn_dbuf; | |
2982 | mutex_exit(&dn->dn_mtx); | |
2983 | return (0); | |
2984 | } | |
34dc7c2f | 2985 | |
1c27024e | 2986 | int nlevels = |
32d41fb7 | 2987 | (dn->dn_phys->dn_nlevels == 0) ? 1 : dn->dn_phys->dn_nlevels; |
1c27024e | 2988 | int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; |
34dc7c2f BB |
2989 | |
2990 | ASSERT3U(level * epbs, <, 64); | |
2991 | ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); | |
32d41fb7 PD |
2992 | /* |
2993 | * This assertion shouldn't trip as long as the max indirect block size | |
2994 | * is less than 1M. The reason for this is that up to that point, | |
2995 | * the number of levels required to address an entire object with blocks | |
2996 | * of size SPA_MINBLOCKSIZE satisfies nlevels * epbs + 1 <= 64. In | |
2997 | * other words, if N * epbs + 1 > 64, then if (N-1) * epbs + 1 > 55 | |
2998 | * (i.e. we can address the entire object), objects will all use at most | |
2999 | * N-1 levels and the assertion won't overflow. However, once epbs is | |
3000 | * 13, 4 * 13 + 1 = 53, but 5 * 13 + 1 = 66. Then, 4 levels will not be | |
3001 | * enough to address an entire object, so objects will have 5 levels, | |
3002 | * but then this assertion will overflow. | |
3003 | * | |
3004 | * All this is to say that if we ever increase DN_MAX_INDBLKSHIFT, we | |
3005 | * need to redo this logic to handle overflows. | |
3006 | */ | |
3007 | ASSERT(level >= nlevels || | |
3008 | ((nlevels - level - 1) * epbs) + | |
3009 | highbit64(dn->dn_phys->dn_nblkptr) <= 64); | |
34dc7c2f | 3010 | if (level >= nlevels || |
32d41fb7 PD |
3011 | blkid >= ((uint64_t)dn->dn_phys->dn_nblkptr << |
3012 | ((nlevels - level - 1) * epbs)) || | |
3013 | (fail_sparse && | |
3014 | blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) { | |
34dc7c2f | 3015 | /* the buffer has no parent yet */ |
2e528b49 | 3016 | return (SET_ERROR(ENOENT)); |
34dc7c2f BB |
3017 | } else if (level < nlevels-1) { |
3018 | /* this block is referenced from an indirect block */ | |
fc5bb51f | 3019 | int err; |
64b6c47d TN |
3020 | |
3021 | err = dbuf_hold_impl(dn, level + 1, | |
adb726eb | 3022 | blkid >> epbs, fail_sparse, FALSE, NULL, parentp); |
64b6c47d | 3023 | |
34dc7c2f BB |
3024 | if (err) |
3025 | return (err); | |
3026 | err = dbuf_read(*parentp, NULL, | |
3027 | (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL)); | |
3028 | if (err) { | |
3029 | dbuf_rele(*parentp, NULL); | |
3030 | *parentp = NULL; | |
3031 | return (err); | |
3032 | } | |
f664f1ee | 3033 | rw_enter(&(*parentp)->db_rwlock, RW_READER); |
34dc7c2f BB |
3034 | *bpp = ((blkptr_t *)(*parentp)->db.db_data) + |
3035 | (blkid & ((1ULL << epbs) - 1)); | |
32d41fb7 PD |
3036 | if (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs))) |
3037 | ASSERT(BP_IS_HOLE(*bpp)); | |
f664f1ee | 3038 | rw_exit(&(*parentp)->db_rwlock); |
34dc7c2f BB |
3039 | return (0); |
3040 | } else { | |
3041 | /* the block is referenced from the dnode */ | |
3042 | ASSERT3U(level, ==, nlevels-1); | |
3043 | ASSERT(dn->dn_phys->dn_nblkptr == 0 || | |
3044 | blkid < dn->dn_phys->dn_nblkptr); | |
3045 | if (dn->dn_dbuf) { | |
3046 | dbuf_add_ref(dn->dn_dbuf, NULL); | |
3047 | *parentp = dn->dn_dbuf; | |
3048 | } | |
3049 | *bpp = &dn->dn_phys->dn_blkptr[blkid]; | |
3050 | return (0); | |
3051 | } | |
3052 | } | |
3053 | ||
3054 | static dmu_buf_impl_t * | |
3055 | dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid, | |
3056 | dmu_buf_impl_t *parent, blkptr_t *blkptr) | |
3057 | { | |
428870ff | 3058 | objset_t *os = dn->dn_objset; |
34dc7c2f BB |
3059 | dmu_buf_impl_t *db, *odb; |
3060 | ||
3061 | ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); | |
3062 | ASSERT(dn->dn_type != DMU_OT_NONE); | |
3063 | ||
d3c2ae1c | 3064 | db = kmem_cache_alloc(dbuf_kmem_cache, KM_SLEEP); |
34dc7c2f | 3065 | |
cccbed9f MM |
3066 | list_create(&db->db_dirty_records, sizeof (dbuf_dirty_record_t), |
3067 | offsetof(dbuf_dirty_record_t, dr_dbuf_node)); | |
3068 | ||
34dc7c2f BB |
3069 | db->db_objset = os; |
3070 | db->db.db_object = dn->dn_object; | |
3071 | db->db_level = level; | |
3072 | db->db_blkid = blkid; | |
34dc7c2f | 3073 | db->db_dirtycnt = 0; |
572e2857 | 3074 | db->db_dnode_handle = dn->dn_handle; |
34dc7c2f BB |
3075 | db->db_parent = parent; |
3076 | db->db_blkptr = blkptr; | |
3077 | ||
0c66c32d | 3078 | db->db_user = NULL; |
bc4501f7 JG |
3079 | db->db_user_immediate_evict = FALSE; |
3080 | db->db_freed_in_flight = FALSE; | |
3081 | db->db_pending_evict = FALSE; | |
34dc7c2f | 3082 | |
428870ff | 3083 | if (blkid == DMU_BONUS_BLKID) { |
34dc7c2f | 3084 | ASSERT3P(parent, ==, dn->dn_dbuf); |
50c957f7 | 3085 | db->db.db_size = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) - |
34dc7c2f BB |
3086 | (dn->dn_nblkptr-1) * sizeof (blkptr_t); |
3087 | ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); | |
428870ff | 3088 | db->db.db_offset = DMU_BONUS_BLKID; |
34dc7c2f | 3089 | db->db_state = DB_UNCACHED; |
8b3547a4 | 3090 | DTRACE_SET_STATE(db, "bonus buffer created"); |
2e5dc449 | 3091 | db->db_caching_status = DB_NO_CACHE; |
34dc7c2f | 3092 | /* the bonus dbuf is not placed in the hash table */ |
25458cbe | 3093 | arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF); |
34dc7c2f | 3094 | return (db); |
428870ff BB |
3095 | } else if (blkid == DMU_SPILL_BLKID) { |
3096 | db->db.db_size = (blkptr != NULL) ? | |
3097 | BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE; | |
3098 | db->db.db_offset = 0; | |
34dc7c2f BB |
3099 | } else { |
3100 | int blocksize = | |
e8b96c60 | 3101 | db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz; |
34dc7c2f BB |
3102 | db->db.db_size = blocksize; |
3103 | db->db.db_offset = db->db_blkid * blocksize; | |
3104 | } | |
3105 | ||
3106 | /* | |
3107 | * Hold the dn_dbufs_mtx while we get the new dbuf | |
3108 | * in the hash table *and* added to the dbufs list. | |
3109 | * This prevents a possible deadlock with someone | |
e1cfd73f | 3110 | * trying to look up this dbuf before it's added to the |
34dc7c2f BB |
3111 | * dn_dbufs list. |
3112 | */ | |
3113 | mutex_enter(&dn->dn_dbufs_mtx); | |
8b3547a4 | 3114 | db->db_state = DB_EVICTING; /* not worth logging this state change */ |
34dc7c2f BB |
3115 | if ((odb = dbuf_hash_insert(db)) != NULL) { |
3116 | /* someone else inserted it first */ | |
34dc7c2f | 3117 | mutex_exit(&dn->dn_dbufs_mtx); |
490c845e | 3118 | kmem_cache_free(dbuf_kmem_cache, db); |
5e021f56 | 3119 | DBUF_STAT_BUMP(hash_insert_race); |
34dc7c2f BB |
3120 | return (odb); |
3121 | } | |
8951cb8d | 3122 | avl_add(&dn->dn_dbufs, db); |
9c9531cb | 3123 | |
34dc7c2f | 3124 | db->db_state = DB_UNCACHED; |
8b3547a4 | 3125 | DTRACE_SET_STATE(db, "regular buffer created"); |
2e5dc449 | 3126 | db->db_caching_status = DB_NO_CACHE; |
34dc7c2f | 3127 | mutex_exit(&dn->dn_dbufs_mtx); |
25458cbe | 3128 | arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_DBUF); |
34dc7c2f BB |
3129 | |
3130 | if (parent && parent != dn->dn_dbuf) | |
3131 | dbuf_add_ref(parent, db); | |
3132 | ||
3133 | ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || | |
424fd7c3 | 3134 | zfs_refcount_count(&dn->dn_holds) > 0); |
c13060e4 | 3135 | (void) zfs_refcount_add(&dn->dn_holds, db); |
34dc7c2f BB |
3136 | |
3137 | dprintf_dbuf(db, "db=%p\n", db); | |
3138 | ||
3139 | return (db); | |
3140 | } | |
3141 | ||
30af21b0 PD |
3142 | /* |
3143 | * This function returns a block pointer and information about the object, | |
3144 | * given a dnode and a block. This is a publicly accessible version of | |
3145 | * dbuf_findbp that only returns some information, rather than the | |
3146 | * dbuf. Note that the dnode passed in must be held, and the dn_struct_rwlock | |
3147 | * should be locked as (at least) a reader. | |
3148 | */ | |
3149 | int | |
3150 | dbuf_dnode_findbp(dnode_t *dn, uint64_t level, uint64_t blkid, | |
3151 | blkptr_t *bp, uint16_t *datablkszsec, uint8_t *indblkshift) | |
3152 | { | |
3153 | dmu_buf_impl_t *dbp = NULL; | |
3154 | blkptr_t *bp2; | |
3155 | int err = 0; | |
3156 | ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); | |
3157 | ||
3158 | err = dbuf_findbp(dn, level, blkid, B_FALSE, &dbp, &bp2); | |
3159 | if (err == 0) { | |
3160 | *bp = *bp2; | |
3161 | if (dbp != NULL) | |
3162 | dbuf_rele(dbp, NULL); | |
3163 | if (datablkszsec != NULL) | |
3164 | *datablkszsec = dn->dn_phys->dn_datablkszsec; | |
3165 | if (indblkshift != NULL) | |
3166 | *indblkshift = dn->dn_phys->dn_indblkshift; | |
3167 | } | |
3168 | ||
3169 | return (err); | |
3170 | } | |
3171 | ||
fcff0f35 PD |
3172 | typedef struct dbuf_prefetch_arg { |
3173 | spa_t *dpa_spa; /* The spa to issue the prefetch in. */ | |
3174 | zbookmark_phys_t dpa_zb; /* The target block to prefetch. */ | |
3175 | int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */ | |
3176 | int dpa_curlevel; /* The current level that we're reading */ | |
d3c2ae1c | 3177 | dnode_t *dpa_dnode; /* The dnode associated with the prefetch */ |
fcff0f35 PD |
3178 | zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */ |
3179 | zio_t *dpa_zio; /* The parent zio_t for all prefetches. */ | |
3180 | arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */ | |
af20b970 MM |
3181 | dbuf_prefetch_fn dpa_cb; /* prefetch completion callback */ |
3182 | void *dpa_arg; /* prefetch completion arg */ | |
fcff0f35 PD |
3183 | } dbuf_prefetch_arg_t; |
3184 | ||
af20b970 MM |
3185 | static void |
3186 | dbuf_prefetch_fini(dbuf_prefetch_arg_t *dpa, boolean_t io_done) | |
3187 | { | |
6aa8c21a AM |
3188 | if (dpa->dpa_cb != NULL) { |
3189 | dpa->dpa_cb(dpa->dpa_arg, dpa->dpa_zb.zb_level, | |
3190 | dpa->dpa_zb.zb_blkid, io_done); | |
3191 | } | |
af20b970 MM |
3192 | kmem_free(dpa, sizeof (*dpa)); |
3193 | } | |
3194 | ||
3195 | static void | |
3196 | dbuf_issue_final_prefetch_done(zio_t *zio, const zbookmark_phys_t *zb, | |
3197 | const blkptr_t *iobp, arc_buf_t *abuf, void *private) | |
3198 | { | |
14e4e3cb | 3199 | (void) zio, (void) zb, (void) iobp; |
af20b970 MM |
3200 | dbuf_prefetch_arg_t *dpa = private; |
3201 | ||
af20b970 MM |
3202 | if (abuf != NULL) |
3203 | arc_buf_destroy(abuf, private); | |
9619bcde BB |
3204 | |
3205 | dbuf_prefetch_fini(dpa, B_TRUE); | |
af20b970 MM |
3206 | } |
3207 | ||
fcff0f35 PD |
3208 | /* |
3209 | * Actually issue the prefetch read for the block given. | |
3210 | */ | |
3211 | static void | |
3212 | dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp) | |
3213 | { | |
30af21b0 PD |
3214 | ASSERT(!BP_IS_REDACTED(bp) || |
3215 | dsl_dataset_feature_is_active( | |
3216 | dpa->dpa_dnode->dn_objset->os_dsl_dataset, | |
3217 | SPA_FEATURE_REDACTED_DATASETS)); | |
3218 | ||
3219 | if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp) || BP_IS_REDACTED(bp)) | |
af20b970 | 3220 | return (dbuf_prefetch_fini(dpa, B_FALSE)); |
fcff0f35 | 3221 | |
4515b1d0 | 3222 | int zio_flags = ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE; |
1c27024e | 3223 | arc_flags_t aflags = |
1e4732cb MM |
3224 | dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH | |
3225 | ARC_FLAG_NO_BUF; | |
fcff0f35 | 3226 | |
4515b1d0 TC |
3227 | /* dnodes are always read as raw and then converted later */ |
3228 | if (BP_GET_TYPE(bp) == DMU_OT_DNODE && BP_IS_PROTECTED(bp) && | |
3229 | dpa->dpa_curlevel == 0) | |
3230 | zio_flags |= ZIO_FLAG_RAW; | |
3231 | ||
fcff0f35 PD |
3232 | ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp)); |
3233 | ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level); | |
3234 | ASSERT(dpa->dpa_zio != NULL); | |
af20b970 MM |
3235 | (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, |
3236 | dbuf_issue_final_prefetch_done, dpa, | |
4515b1d0 | 3237 | dpa->dpa_prio, zio_flags, &aflags, &dpa->dpa_zb); |
fcff0f35 PD |
3238 | } |
3239 | ||
3240 | /* | |
3241 | * Called when an indirect block above our prefetch target is read in. This | |
3242 | * will either read in the next indirect block down the tree or issue the actual | |
3243 | * prefetch if the next block down is our target. | |
3244 | */ | |
3245 | static void | |
d4a72f23 TC |
3246 | dbuf_prefetch_indirect_done(zio_t *zio, const zbookmark_phys_t *zb, |
3247 | const blkptr_t *iobp, arc_buf_t *abuf, void *private) | |
fcff0f35 | 3248 | { |
14e4e3cb | 3249 | (void) zb, (void) iobp; |
fcff0f35 | 3250 | dbuf_prefetch_arg_t *dpa = private; |
fcff0f35 PD |
3251 | |
3252 | ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel); | |
3253 | ASSERT3S(dpa->dpa_curlevel, >, 0); | |
d3c2ae1c | 3254 | |
c3bd3fb4 TC |
3255 | if (abuf == NULL) { |
3256 | ASSERT(zio == NULL || zio->io_error != 0); | |
d954ca19 RY |
3257 | dbuf_prefetch_fini(dpa, B_TRUE); |
3258 | return; | |
c3bd3fb4 TC |
3259 | } |
3260 | ASSERT(zio == NULL || zio->io_error == 0); | |
3261 | ||
d3c2ae1c GW |
3262 | /* |
3263 | * The dpa_dnode is only valid if we are called with a NULL | |
3264 | * zio. This indicates that the arc_read() returned without | |
3265 | * first calling zio_read() to issue a physical read. Once | |
3266 | * a physical read is made the dpa_dnode must be invalidated | |
3267 | * as the locks guarding it may have been dropped. If the | |
3268 | * dpa_dnode is still valid, then we want to add it to the dbuf | |
3269 | * cache. To do so, we must hold the dbuf associated with the block | |
3270 | * we just prefetched, read its contents so that we associate it | |
3271 | * with an arc_buf_t, and then release it. | |
3272 | */ | |
fcff0f35 PD |
3273 | if (zio != NULL) { |
3274 | ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel); | |
b5256303 | 3275 | if (zio->io_flags & ZIO_FLAG_RAW_COMPRESS) { |
d3c2ae1c GW |
3276 | ASSERT3U(BP_GET_PSIZE(zio->io_bp), ==, zio->io_size); |
3277 | } else { | |
3278 | ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size); | |
3279 | } | |
fcff0f35 | 3280 | ASSERT3P(zio->io_spa, ==, dpa->dpa_spa); |
d3c2ae1c GW |
3281 | |
3282 | dpa->dpa_dnode = NULL; | |
3283 | } else if (dpa->dpa_dnode != NULL) { | |
3284 | uint64_t curblkid = dpa->dpa_zb.zb_blkid >> | |
3285 | (dpa->dpa_epbs * (dpa->dpa_curlevel - | |
3286 | dpa->dpa_zb.zb_level)); | |
3287 | dmu_buf_impl_t *db = dbuf_hold_level(dpa->dpa_dnode, | |
3288 | dpa->dpa_curlevel, curblkid, FTAG); | |
305781da | 3289 | if (db == NULL) { |
305781da | 3290 | arc_buf_destroy(abuf, private); |
d954ca19 RY |
3291 | dbuf_prefetch_fini(dpa, B_TRUE); |
3292 | return; | |
305781da | 3293 | } |
d3c2ae1c GW |
3294 | (void) dbuf_read(db, NULL, |
3295 | DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH | DB_RF_HAVESTRUCT); | |
3296 | dbuf_rele(db, FTAG); | |
fcff0f35 PD |
3297 | } |
3298 | ||
d4a72f23 | 3299 | dpa->dpa_curlevel--; |
1c27024e | 3300 | uint64_t nextblkid = dpa->dpa_zb.zb_blkid >> |
fcff0f35 | 3301 | (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level)); |
1c27024e | 3302 | blkptr_t *bp = ((blkptr_t *)abuf->b_data) + |
fcff0f35 | 3303 | P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs); |
d4a72f23 | 3304 | |
30af21b0 PD |
3305 | ASSERT(!BP_IS_REDACTED(bp) || |
3306 | dsl_dataset_feature_is_active( | |
3307 | dpa->dpa_dnode->dn_objset->os_dsl_dataset, | |
3308 | SPA_FEATURE_REDACTED_DATASETS)); | |
3309 | if (BP_IS_HOLE(bp) || BP_IS_REDACTED(bp)) { | |
d954ca19 | 3310 | arc_buf_destroy(abuf, private); |
af20b970 | 3311 | dbuf_prefetch_fini(dpa, B_TRUE); |
d954ca19 | 3312 | return; |
fcff0f35 PD |
3313 | } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) { |
3314 | ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid); | |
3315 | dbuf_issue_final_prefetch(dpa, bp); | |
fcff0f35 PD |
3316 | } else { |
3317 | arc_flags_t iter_aflags = ARC_FLAG_NOWAIT; | |
3318 | zbookmark_phys_t zb; | |
3319 | ||
7c351e31 | 3320 | /* flag if L2ARC eligible, l2arc_noprefetch then decides */ |
3321 | if (dpa->dpa_aflags & ARC_FLAG_L2CACHE) | |
3322 | iter_aflags |= ARC_FLAG_L2CACHE; | |
3323 | ||
fcff0f35 PD |
3324 | ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp)); |
3325 | ||
3326 | SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset, | |
3327 | dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid); | |
3328 | ||
3329 | (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, | |
6aa8c21a AM |
3330 | bp, dbuf_prefetch_indirect_done, dpa, |
3331 | ZIO_PRIORITY_SYNC_READ, | |
fcff0f35 PD |
3332 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, |
3333 | &iter_aflags, &zb); | |
3334 | } | |
d3c2ae1c GW |
3335 | |
3336 | arc_buf_destroy(abuf, private); | |
fcff0f35 PD |
3337 | } |
3338 | ||
3339 | /* | |
3340 | * Issue prefetch reads for the given block on the given level. If the indirect | |
3341 | * blocks above that block are not in memory, we will read them in | |
3342 | * asynchronously. As a result, this call never blocks waiting for a read to | |
b5256303 TC |
3343 | * complete. Note that the prefetch might fail if the dataset is encrypted and |
3344 | * the encryption key is unmapped before the IO completes. | |
fcff0f35 | 3345 | */ |
af20b970 MM |
3346 | int |
3347 | dbuf_prefetch_impl(dnode_t *dn, int64_t level, uint64_t blkid, | |
3348 | zio_priority_t prio, arc_flags_t aflags, dbuf_prefetch_fn cb, | |
3349 | void *arg) | |
34dc7c2f | 3350 | { |
fcff0f35 PD |
3351 | blkptr_t bp; |
3352 | int epbs, nlevels, curlevel; | |
3353 | uint64_t curblkid; | |
34dc7c2f | 3354 | |
428870ff | 3355 | ASSERT(blkid != DMU_BONUS_BLKID); |
34dc7c2f BB |
3356 | ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); |
3357 | ||
7f60329a | 3358 | if (blkid > dn->dn_maxblkid) |
af20b970 | 3359 | goto no_issue; |
7f60329a | 3360 | |
f664f1ee | 3361 | if (level == 0 && dnode_block_freed(dn, blkid)) |
af20b970 | 3362 | goto no_issue; |
34dc7c2f | 3363 | |
fcff0f35 PD |
3364 | /* |
3365 | * This dnode hasn't been written to disk yet, so there's nothing to | |
3366 | * prefetch. | |
3367 | */ | |
3368 | nlevels = dn->dn_phys->dn_nlevels; | |
3369 | if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0) | |
af20b970 | 3370 | goto no_issue; |
fcff0f35 PD |
3371 | |
3372 | epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; | |
3373 | if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level)) | |
af20b970 | 3374 | goto no_issue; |
fcff0f35 | 3375 | |
1c27024e | 3376 | dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object, |
fcff0f35 PD |
3377 | level, blkid); |
3378 | if (db != NULL) { | |
3379 | mutex_exit(&db->db_mtx); | |
572e2857 | 3380 | /* |
fcff0f35 PD |
3381 | * This dbuf already exists. It is either CACHED, or |
3382 | * (we assume) about to be read or filled. | |
572e2857 | 3383 | */ |
af20b970 | 3384 | goto no_issue; |
34dc7c2f BB |
3385 | } |
3386 | ||
fcff0f35 PD |
3387 | /* |
3388 | * Find the closest ancestor (indirect block) of the target block | |
3389 | * that is present in the cache. In this indirect block, we will | |
3390 | * find the bp that is at curlevel, curblkid. | |
3391 | */ | |
3392 | curlevel = level; | |
3393 | curblkid = blkid; | |
3394 | while (curlevel < nlevels - 1) { | |
3395 | int parent_level = curlevel + 1; | |
3396 | uint64_t parent_blkid = curblkid >> epbs; | |
3397 | dmu_buf_impl_t *db; | |
3398 | ||
3399 | if (dbuf_hold_impl(dn, parent_level, parent_blkid, | |
3400 | FALSE, TRUE, FTAG, &db) == 0) { | |
3401 | blkptr_t *bpp = db->db_buf->b_data; | |
3402 | bp = bpp[P2PHASE(curblkid, 1 << epbs)]; | |
3403 | dbuf_rele(db, FTAG); | |
3404 | break; | |
3405 | } | |
428870ff | 3406 | |
fcff0f35 PD |
3407 | curlevel = parent_level; |
3408 | curblkid = parent_blkid; | |
3409 | } | |
34dc7c2f | 3410 | |
fcff0f35 PD |
3411 | if (curlevel == nlevels - 1) { |
3412 | /* No cached indirect blocks found. */ | |
3413 | ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr); | |
3414 | bp = dn->dn_phys->dn_blkptr[curblkid]; | |
34dc7c2f | 3415 | } |
30af21b0 PD |
3416 | ASSERT(!BP_IS_REDACTED(&bp) || |
3417 | dsl_dataset_feature_is_active(dn->dn_objset->os_dsl_dataset, | |
3418 | SPA_FEATURE_REDACTED_DATASETS)); | |
3419 | if (BP_IS_HOLE(&bp) || BP_IS_REDACTED(&bp)) | |
af20b970 | 3420 | goto no_issue; |
fcff0f35 PD |
3421 | |
3422 | ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp)); | |
3423 | ||
1c27024e | 3424 | zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL, |
fcff0f35 PD |
3425 | ZIO_FLAG_CANFAIL); |
3426 | ||
1c27024e DB |
3427 | dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP); |
3428 | dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset; | |
fcff0f35 PD |
3429 | SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET, |
3430 | dn->dn_object, level, blkid); | |
3431 | dpa->dpa_curlevel = curlevel; | |
3432 | dpa->dpa_prio = prio; | |
3433 | dpa->dpa_aflags = aflags; | |
3434 | dpa->dpa_spa = dn->dn_objset->os_spa; | |
d3c2ae1c | 3435 | dpa->dpa_dnode = dn; |
fcff0f35 PD |
3436 | dpa->dpa_epbs = epbs; |
3437 | dpa->dpa_zio = pio; | |
af20b970 MM |
3438 | dpa->dpa_cb = cb; |
3439 | dpa->dpa_arg = arg; | |
fcff0f35 | 3440 | |
7c351e31 | 3441 | /* flag if L2ARC eligible, l2arc_noprefetch then decides */ |
c9d62d13 | 3442 | if (dnode_level_is_l2cacheable(&bp, dn, level)) |
7c351e31 | 3443 | dpa->dpa_aflags |= ARC_FLAG_L2CACHE; |
3444 | ||
fcff0f35 PD |
3445 | /* |
3446 | * If we have the indirect just above us, no need to do the asynchronous | |
3447 | * prefetch chain; we'll just run the last step ourselves. If we're at | |
3448 | * a higher level, though, we want to issue the prefetches for all the | |
3449 | * indirect blocks asynchronously, so we can go on with whatever we were | |
3450 | * doing. | |
3451 | */ | |
3452 | if (curlevel == level) { | |
3453 | ASSERT3U(curblkid, ==, blkid); | |
3454 | dbuf_issue_final_prefetch(dpa, &bp); | |
fcff0f35 PD |
3455 | } else { |
3456 | arc_flags_t iter_aflags = ARC_FLAG_NOWAIT; | |
3457 | zbookmark_phys_t zb; | |
3458 | ||
7c351e31 | 3459 | /* flag if L2ARC eligible, l2arc_noprefetch then decides */ |
c9d62d13 | 3460 | if (dnode_level_is_l2cacheable(&bp, dn, level)) |
7c351e31 | 3461 | iter_aflags |= ARC_FLAG_L2CACHE; |
3462 | ||
fcff0f35 PD |
3463 | SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET, |
3464 | dn->dn_object, curlevel, curblkid); | |
3465 | (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, | |
6aa8c21a AM |
3466 | &bp, dbuf_prefetch_indirect_done, dpa, |
3467 | ZIO_PRIORITY_SYNC_READ, | |
fcff0f35 PD |
3468 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, |
3469 | &iter_aflags, &zb); | |
3470 | } | |
3471 | /* | |
3472 | * We use pio here instead of dpa_zio since it's possible that | |
3473 | * dpa may have already been freed. | |
3474 | */ | |
3475 | zio_nowait(pio); | |
af20b970 MM |
3476 | return (1); |
3477 | no_issue: | |
3478 | if (cb != NULL) | |
6aa8c21a | 3479 | cb(arg, level, blkid, B_FALSE); |
af20b970 MM |
3480 | return (0); |
3481 | } | |
3482 | ||
3483 | int | |
3484 | dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio, | |
3485 | arc_flags_t aflags) | |
3486 | { | |
3487 | ||
3488 | return (dbuf_prefetch_impl(dn, level, blkid, prio, aflags, NULL, NULL)); | |
34dc7c2f BB |
3489 | } |
3490 | ||
71a24c3c | 3491 | /* |
64b6c47d | 3492 | * Helper function for dbuf_hold_impl() to copy a buffer. Handles |
71a24c3c TC |
3493 | * the case of encrypted, compressed and uncompressed buffers by |
3494 | * allocating the new buffer, respectively, with arc_alloc_raw_buf(), | |
3495 | * arc_alloc_compressed_buf() or arc_alloc_buf().* | |
3496 | * | |
64b6c47d | 3497 | * NOTE: Declared noinline to avoid stack bloat in dbuf_hold_impl(). |
71a24c3c TC |
3498 | */ |
3499 | noinline static void | |
64b6c47d | 3500 | dbuf_hold_copy(dnode_t *dn, dmu_buf_impl_t *db) |
71a24c3c | 3501 | { |
64b6c47d | 3502 | dbuf_dirty_record_t *dr = db->db_data_pending; |
a81b8124 AR |
3503 | arc_buf_t *data = dr->dt.dl.dr_data; |
3504 | enum zio_compress compress_type = arc_get_compression(data); | |
3505 | uint8_t complevel = arc_get_complevel(data); | |
3506 | ||
3507 | if (arc_is_encrypted(data)) { | |
3508 | boolean_t byteorder; | |
3509 | uint8_t salt[ZIO_DATA_SALT_LEN]; | |
3510 | uint8_t iv[ZIO_DATA_IV_LEN]; | |
3511 | uint8_t mac[ZIO_DATA_MAC_LEN]; | |
3512 | ||
3513 | arc_get_raw_params(data, &byteorder, salt, iv, mac); | |
3514 | dbuf_set_data(db, arc_alloc_raw_buf(dn->dn_objset->os_spa, db, | |
3515 | dmu_objset_id(dn->dn_objset), byteorder, salt, iv, mac, | |
3516 | dn->dn_type, arc_buf_size(data), arc_buf_lsize(data), | |
3517 | compress_type, complevel)); | |
3518 | } else if (compress_type != ZIO_COMPRESS_OFF) { | |
3519 | dbuf_set_data(db, arc_alloc_compressed_buf( | |
3520 | dn->dn_objset->os_spa, db, arc_buf_size(data), | |
3521 | arc_buf_lsize(data), compress_type, complevel)); | |
3522 | } else { | |
3523 | dbuf_set_data(db, arc_alloc_buf(dn->dn_objset->os_spa, db, | |
3524 | DBUF_GET_BUFC_TYPE(db), db->db.db_size)); | |
3525 | } | |
71a24c3c | 3526 | |
f664f1ee | 3527 | rw_enter(&db->db_rwlock, RW_WRITER); |
861166b0 | 3528 | memcpy(db->db.db_data, data->b_data, arc_buf_size(data)); |
f664f1ee | 3529 | rw_exit(&db->db_rwlock); |
71a24c3c TC |
3530 | } |
3531 | ||
34dc7c2f BB |
3532 | /* |
3533 | * Returns with db_holds incremented, and db_mtx not held. | |
3534 | * Note: dn_struct_rwlock must be held. | |
3535 | */ | |
64b6c47d TN |
3536 | int |
3537 | dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, | |
3538 | boolean_t fail_sparse, boolean_t fail_uncached, | |
a926aab9 | 3539 | const void *tag, dmu_buf_impl_t **dbp) |
34dc7c2f | 3540 | { |
64b6c47d | 3541 | dmu_buf_impl_t *db, *parent = NULL; |
d3c2ae1c | 3542 | |
37f03da8 | 3543 | /* If the pool has been created, verify the tx_sync_lock is not held */ |
64b6c47d | 3544 | spa_t *spa = dn->dn_objset->os_spa; |
37f03da8 SH |
3545 | dsl_pool_t *dp = spa->spa_dsl_pool; |
3546 | if (dp != NULL) { | |
3547 | ASSERT(!MUTEX_HELD(&dp->dp_tx.tx_sync_lock)); | |
3548 | } | |
3549 | ||
64b6c47d TN |
3550 | ASSERT(blkid != DMU_BONUS_BLKID); |
3551 | ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); | |
3552 | ASSERT3U(dn->dn_nlevels, >, level); | |
3553 | ||
3554 | *dbp = NULL; | |
3555 | ||
34dc7c2f | 3556 | /* dbuf_find() returns with db_mtx held */ |
64b6c47d | 3557 | db = dbuf_find(dn->dn_objset, dn->dn_object, level, blkid); |
fc5bb51f | 3558 | |
64b6c47d TN |
3559 | if (db == NULL) { |
3560 | blkptr_t *bp = NULL; | |
3561 | int err; | |
fc5bb51f | 3562 | |
64b6c47d | 3563 | if (fail_uncached) |
fcff0f35 PD |
3564 | return (SET_ERROR(ENOENT)); |
3565 | ||
64b6c47d TN |
3566 | ASSERT3P(parent, ==, NULL); |
3567 | err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp); | |
3568 | if (fail_sparse) { | |
3569 | if (err == 0 && bp && BP_IS_HOLE(bp)) | |
3570 | err = SET_ERROR(ENOENT); | |
3571 | if (err) { | |
3572 | if (parent) | |
3573 | dbuf_rele(parent, NULL); | |
3574 | return (err); | |
34dc7c2f BB |
3575 | } |
3576 | } | |
64b6c47d TN |
3577 | if (err && err != ENOENT) |
3578 | return (err); | |
3579 | db = dbuf_create(dn, level, blkid, parent, bp); | |
34dc7c2f BB |
3580 | } |
3581 | ||
64b6c47d TN |
3582 | if (fail_uncached && db->db_state != DB_CACHED) { |
3583 | mutex_exit(&db->db_mtx); | |
fcff0f35 PD |
3584 | return (SET_ERROR(ENOENT)); |
3585 | } | |
3586 | ||
64b6c47d TN |
3587 | if (db->db_buf != NULL) { |
3588 | arc_buf_access(db->db_buf); | |
3589 | ASSERT3P(db->db.db_data, ==, db->db_buf->b_data); | |
0873bb63 | 3590 | } |
34dc7c2f | 3591 | |
64b6c47d | 3592 | ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf)); |
34dc7c2f BB |
3593 | |
3594 | /* | |
e1cfd73f | 3595 | * If this buffer is currently syncing out, and we are |
34dc7c2f BB |
3596 | * still referencing it from db_data, we need to make a copy |
3597 | * of it in case we decide we want to dirty it again in this txg. | |
3598 | */ | |
64b6c47d TN |
3599 | if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && |
3600 | dn->dn_object != DMU_META_DNODE_OBJECT && | |
3601 | db->db_state == DB_CACHED && db->db_data_pending) { | |
3602 | dbuf_dirty_record_t *dr = db->db_data_pending; | |
3603 | if (dr->dt.dl.dr_data == db->db_buf) | |
3604 | dbuf_hold_copy(dn, db); | |
3605 | } | |
3606 | ||
3607 | if (multilist_link_active(&db->db_cache_link)) { | |
3608 | ASSERT(zfs_refcount_is_zero(&db->db_holds)); | |
3609 | ASSERT(db->db_caching_status == DB_DBUF_CACHE || | |
3610 | db->db_caching_status == DB_DBUF_METADATA_CACHE); | |
3611 | ||
ffdf019c | 3612 | multilist_remove(&dbuf_caches[db->db_caching_status].cache, db); |
424fd7c3 | 3613 | (void) zfs_refcount_remove_many( |
64b6c47d TN |
3614 | &dbuf_caches[db->db_caching_status].size, |
3615 | db->db.db_size, db); | |
2e5dc449 | 3616 | |
64b6c47d | 3617 | if (db->db_caching_status == DB_DBUF_METADATA_CACHE) { |
2e5dc449 MA |
3618 | DBUF_STAT_BUMPDOWN(metadata_cache_count); |
3619 | } else { | |
64b6c47d | 3620 | DBUF_STAT_BUMPDOWN(cache_levels[db->db_level]); |
2e5dc449 | 3621 | DBUF_STAT_BUMPDOWN(cache_count); |
64b6c47d TN |
3622 | DBUF_STAT_DECR(cache_levels_bytes[db->db_level], |
3623 | db->db.db_size); | |
2e5dc449 | 3624 | } |
64b6c47d | 3625 | db->db_caching_status = DB_NO_CACHE; |
d3c2ae1c | 3626 | } |
64b6c47d TN |
3627 | (void) zfs_refcount_add(&db->db_holds, tag); |
3628 | DBUF_VERIFY(db); | |
3629 | mutex_exit(&db->db_mtx); | |
34dc7c2f BB |
3630 | |
3631 | /* NOTE: we can't rele the parent until after we drop the db_mtx */ | |
64b6c47d TN |
3632 | if (parent) |
3633 | dbuf_rele(parent, NULL); | |
34dc7c2f | 3634 | |
64b6c47d TN |
3635 | ASSERT3P(DB_DNODE(db), ==, dn); |
3636 | ASSERT3U(db->db_blkid, ==, blkid); | |
3637 | ASSERT3U(db->db_level, ==, level); | |
3638 | *dbp = db; | |
34dc7c2f BB |
3639 | |
3640 | return (0); | |
3641 | } | |
3642 | ||
3643 | dmu_buf_impl_t * | |
a926aab9 | 3644 | dbuf_hold(dnode_t *dn, uint64_t blkid, const void *tag) |
34dc7c2f | 3645 | { |
fcff0f35 | 3646 | return (dbuf_hold_level(dn, 0, blkid, tag)); |
34dc7c2f BB |
3647 | } |
3648 | ||
3649 | dmu_buf_impl_t * | |
a926aab9 | 3650 | dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, const void *tag) |
34dc7c2f BB |
3651 | { |
3652 | dmu_buf_impl_t *db; | |
fcff0f35 | 3653 | int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db); |
34dc7c2f BB |
3654 | return (err ? NULL : db); |
3655 | } | |
3656 | ||
3657 | void | |
3658 | dbuf_create_bonus(dnode_t *dn) | |
3659 | { | |
3660 | ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); | |
3661 | ||
3662 | ASSERT(dn->dn_bonus == NULL); | |
428870ff BB |
3663 | dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL); |
3664 | } | |
3665 | ||
3666 | int | |
3667 | dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx) | |
3668 | { | |
3669 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
572e2857 | 3670 | |
428870ff | 3671 | if (db->db_blkid != DMU_SPILL_BLKID) |
2e528b49 | 3672 | return (SET_ERROR(ENOTSUP)); |
428870ff BB |
3673 | if (blksz == 0) |
3674 | blksz = SPA_MINBLOCKSIZE; | |
f1512ee6 MA |
3675 | ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset))); |
3676 | blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE); | |
428870ff | 3677 | |
428870ff | 3678 | dbuf_new_size(db, blksz, tx); |
428870ff BB |
3679 | |
3680 | return (0); | |
3681 | } | |
3682 | ||
3683 | void | |
3684 | dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx) | |
3685 | { | |
3686 | dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx); | |
34dc7c2f BB |
3687 | } |
3688 | ||
3689 | #pragma weak dmu_buf_add_ref = dbuf_add_ref | |
3690 | void | |
a926aab9 | 3691 | dbuf_add_ref(dmu_buf_impl_t *db, const void *tag) |
34dc7c2f | 3692 | { |
c13060e4 | 3693 | int64_t holds = zfs_refcount_add(&db->db_holds, tag); |
d3c2ae1c | 3694 | VERIFY3S(holds, >, 1); |
34dc7c2f BB |
3695 | } |
3696 | ||
6ebebace JG |
3697 | #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref |
3698 | boolean_t | |
3699 | dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid, | |
a926aab9 | 3700 | const void *tag) |
6ebebace JG |
3701 | { |
3702 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
3703 | dmu_buf_impl_t *found_db; | |
3704 | boolean_t result = B_FALSE; | |
3705 | ||
d617648c | 3706 | if (blkid == DMU_BONUS_BLKID) |
6ebebace JG |
3707 | found_db = dbuf_find_bonus(os, obj); |
3708 | else | |
3709 | found_db = dbuf_find(os, obj, 0, blkid); | |
3710 | ||
3711 | if (found_db != NULL) { | |
3712 | if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) { | |
c13060e4 | 3713 | (void) zfs_refcount_add(&db->db_holds, tag); |
6ebebace JG |
3714 | result = B_TRUE; |
3715 | } | |
d617648c | 3716 | mutex_exit(&found_db->db_mtx); |
6ebebace JG |
3717 | } |
3718 | return (result); | |
3719 | } | |
3720 | ||
572e2857 BB |
3721 | /* |
3722 | * If you call dbuf_rele() you had better not be referencing the dnode handle | |
3723 | * unless you have some other direct or indirect hold on the dnode. (An indirect | |
3724 | * hold is a hold on one of the dnode's dbufs, including the bonus buffer.) | |
3725 | * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the | |
3726 | * dnode's parent dbuf evicting its dnode handles. | |
3727 | */ | |
34dc7c2f | 3728 | void |
a926aab9 | 3729 | dbuf_rele(dmu_buf_impl_t *db, const void *tag) |
428870ff BB |
3730 | { |
3731 | mutex_enter(&db->db_mtx); | |
3d503a76 | 3732 | dbuf_rele_and_unlock(db, tag, B_FALSE); |
428870ff BB |
3733 | } |
3734 | ||
b0bc7a84 | 3735 | void |
a926aab9 | 3736 | dmu_buf_rele(dmu_buf_t *db, const void *tag) |
b0bc7a84 MG |
3737 | { |
3738 | dbuf_rele((dmu_buf_impl_t *)db, tag); | |
3739 | } | |
3740 | ||
428870ff BB |
3741 | /* |
3742 | * dbuf_rele() for an already-locked dbuf. This is necessary to allow | |
1fac63e5 MA |
3743 | * db_dirtycnt and db_holds to be updated atomically. The 'evicting' |
3744 | * argument should be set if we are already in the dbuf-evicting code | |
3745 | * path, in which case we don't want to recursively evict. This allows us to | |
3746 | * avoid deeply nested stacks that would have a call flow similar to this: | |
3747 | * | |
3748 | * dbuf_rele()-->dbuf_rele_and_unlock()-->dbuf_evict_notify() | |
3749 | * ^ | | |
3750 | * | | | |
3751 | * +-----dbuf_destroy()<--dbuf_evict_one()<--------+ | |
3752 | * | |
428870ff BB |
3753 | */ |
3754 | void | |
a926aab9 | 3755 | dbuf_rele_and_unlock(dmu_buf_impl_t *db, const void *tag, boolean_t evicting) |
34dc7c2f BB |
3756 | { |
3757 | int64_t holds; | |
cbd8f5b7 | 3758 | uint64_t size; |
34dc7c2f | 3759 | |
428870ff | 3760 | ASSERT(MUTEX_HELD(&db->db_mtx)); |
34dc7c2f BB |
3761 | DBUF_VERIFY(db); |
3762 | ||
572e2857 BB |
3763 | /* |
3764 | * Remove the reference to the dbuf before removing its hold on the | |
3765 | * dnode so we can guarantee in dnode_move() that a referenced bonus | |
3766 | * buffer has a corresponding dnode hold. | |
3767 | */ | |
424fd7c3 | 3768 | holds = zfs_refcount_remove(&db->db_holds, tag); |
34dc7c2f BB |
3769 | ASSERT(holds >= 0); |
3770 | ||
3771 | /* | |
3772 | * We can't freeze indirects if there is a possibility that they | |
3773 | * may be modified in the current syncing context. | |
3774 | */ | |
d3c2ae1c GW |
3775 | if (db->db_buf != NULL && |
3776 | holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) { | |
34dc7c2f | 3777 | arc_buf_freeze(db->db_buf); |
d3c2ae1c | 3778 | } |
34dc7c2f BB |
3779 | |
3780 | if (holds == db->db_dirtycnt && | |
bc4501f7 | 3781 | db->db_level == 0 && db->db_user_immediate_evict) |
34dc7c2f BB |
3782 | dbuf_evict_user(db); |
3783 | ||
3784 | if (holds == 0) { | |
428870ff | 3785 | if (db->db_blkid == DMU_BONUS_BLKID) { |
4c7b7eed | 3786 | dnode_t *dn; |
bc4501f7 | 3787 | boolean_t evict_dbuf = db->db_pending_evict; |
572e2857 BB |
3788 | |
3789 | /* | |
4c7b7eed JG |
3790 | * If the dnode moves here, we cannot cross this |
3791 | * barrier until the move completes. | |
572e2857 BB |
3792 | */ |
3793 | DB_DNODE_ENTER(db); | |
4c7b7eed JG |
3794 | |
3795 | dn = DB_DNODE(db); | |
3796 | atomic_dec_32(&dn->dn_dbufs_count); | |
3797 | ||
3798 | /* | |
3799 | * Decrementing the dbuf count means that the bonus | |
3800 | * buffer's dnode hold is no longer discounted in | |
3801 | * dnode_move(). The dnode cannot move until after | |
bc4501f7 | 3802 | * the dnode_rele() below. |
4c7b7eed | 3803 | */ |
572e2857 | 3804 | DB_DNODE_EXIT(db); |
4c7b7eed JG |
3805 | |
3806 | /* | |
3807 | * Do not reference db after its lock is dropped. | |
3808 | * Another thread may evict it. | |
3809 | */ | |
3810 | mutex_exit(&db->db_mtx); | |
3811 | ||
bc4501f7 | 3812 | if (evict_dbuf) |
4c7b7eed | 3813 | dnode_evict_bonus(dn); |
bc4501f7 JG |
3814 | |
3815 | dnode_rele(dn, db); | |
34dc7c2f BB |
3816 | } else if (db->db_buf == NULL) { |
3817 | /* | |
3818 | * This is a special case: we never associated this | |
3819 | * dbuf with any data allocated from the ARC. | |
3820 | */ | |
b128c09f BB |
3821 | ASSERT(db->db_state == DB_UNCACHED || |
3822 | db->db_state == DB_NOFILL); | |
d3c2ae1c | 3823 | dbuf_destroy(db); |
34dc7c2f | 3824 | } else if (arc_released(db->db_buf)) { |
34dc7c2f BB |
3825 | /* |
3826 | * This dbuf has anonymous data associated with it. | |
3827 | */ | |
d3c2ae1c | 3828 | dbuf_destroy(db); |
34dc7c2f | 3829 | } else { |
d3c2ae1c GW |
3830 | boolean_t do_arc_evict = B_FALSE; |
3831 | blkptr_t bp; | |
3832 | spa_t *spa = dmu_objset_spa(db->db_objset); | |
3833 | ||
3834 | if (!DBUF_IS_CACHEABLE(db) && | |
3835 | db->db_blkptr != NULL && | |
3836 | !BP_IS_HOLE(db->db_blkptr) && | |
3837 | !BP_IS_EMBEDDED(db->db_blkptr)) { | |
3838 | do_arc_evict = B_TRUE; | |
3839 | bp = *db->db_blkptr; | |
3840 | } | |
1eb5bfa3 | 3841 | |
d3c2ae1c GW |
3842 | if (!DBUF_IS_CACHEABLE(db) || |
3843 | db->db_pending_evict) { | |
3844 | dbuf_destroy(db); | |
3845 | } else if (!multilist_link_active(&db->db_cache_link)) { | |
2e5dc449 MA |
3846 | ASSERT3U(db->db_caching_status, ==, |
3847 | DB_NO_CACHE); | |
3848 | ||
3849 | dbuf_cached_state_t dcs = | |
3850 | dbuf_include_in_metadata_cache(db) ? | |
3851 | DB_DBUF_METADATA_CACHE : DB_DBUF_CACHE; | |
3852 | db->db_caching_status = dcs; | |
3853 | ||
ffdf019c | 3854 | multilist_insert(&dbuf_caches[dcs].cache, db); |
c4c162c1 | 3855 | uint64_t db_size = db->db.db_size; |
cbd8f5b7 | 3856 | size = zfs_refcount_add_many( |
c4c162c1 AM |
3857 | &dbuf_caches[dcs].size, db_size, db); |
3858 | uint8_t db_level = db->db_level; | |
3859 | mutex_exit(&db->db_mtx); | |
2e5dc449 MA |
3860 | |
3861 | if (dcs == DB_DBUF_METADATA_CACHE) { | |
3862 | DBUF_STAT_BUMP(metadata_cache_count); | |
3863 | DBUF_STAT_MAX( | |
3864 | metadata_cache_size_bytes_max, | |
cbd8f5b7 | 3865 | size); |
2e5dc449 | 3866 | } else { |
2e5dc449 | 3867 | DBUF_STAT_BUMP(cache_count); |
2e5dc449 | 3868 | DBUF_STAT_MAX(cache_size_bytes_max, |
cbd8f5b7 | 3869 | size); |
c4c162c1 AM |
3870 | DBUF_STAT_BUMP(cache_levels[db_level]); |
3871 | DBUF_STAT_INCR( | |
3872 | cache_levels_bytes[db_level], | |
3873 | db_size); | |
2e5dc449 | 3874 | } |
d3c2ae1c | 3875 | |
cbd8f5b7 AM |
3876 | if (dcs == DB_DBUF_CACHE && !evicting) |
3877 | dbuf_evict_notify(size); | |
bd089c54 | 3878 | } |
d3c2ae1c GW |
3879 | |
3880 | if (do_arc_evict) | |
3881 | arc_freed(spa, &bp); | |
34dc7c2f BB |
3882 | } |
3883 | } else { | |
3884 | mutex_exit(&db->db_mtx); | |
3885 | } | |
d3c2ae1c | 3886 | |
34dc7c2f BB |
3887 | } |
3888 | ||
3889 | #pragma weak dmu_buf_refcount = dbuf_refcount | |
3890 | uint64_t | |
3891 | dbuf_refcount(dmu_buf_impl_t *db) | |
3892 | { | |
424fd7c3 | 3893 | return (zfs_refcount_count(&db->db_holds)); |
34dc7c2f BB |
3894 | } |
3895 | ||
cd32e5db TC |
3896 | uint64_t |
3897 | dmu_buf_user_refcount(dmu_buf_t *db_fake) | |
3898 | { | |
3899 | uint64_t holds; | |
3900 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
3901 | ||
3902 | mutex_enter(&db->db_mtx); | |
424fd7c3 TS |
3903 | ASSERT3U(zfs_refcount_count(&db->db_holds), >=, db->db_dirtycnt); |
3904 | holds = zfs_refcount_count(&db->db_holds) - db->db_dirtycnt; | |
cd32e5db TC |
3905 | mutex_exit(&db->db_mtx); |
3906 | ||
3907 | return (holds); | |
3908 | } | |
3909 | ||
34dc7c2f | 3910 | void * |
0c66c32d JG |
3911 | dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user, |
3912 | dmu_buf_user_t *new_user) | |
34dc7c2f | 3913 | { |
0c66c32d JG |
3914 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; |
3915 | ||
3916 | mutex_enter(&db->db_mtx); | |
3917 | dbuf_verify_user(db, DBVU_NOT_EVICTING); | |
3918 | if (db->db_user == old_user) | |
3919 | db->db_user = new_user; | |
3920 | else | |
3921 | old_user = db->db_user; | |
3922 | dbuf_verify_user(db, DBVU_NOT_EVICTING); | |
3923 | mutex_exit(&db->db_mtx); | |
3924 | ||
3925 | return (old_user); | |
34dc7c2f BB |
3926 | } |
3927 | ||
3928 | void * | |
0c66c32d | 3929 | dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user) |
34dc7c2f | 3930 | { |
0c66c32d | 3931 | return (dmu_buf_replace_user(db_fake, NULL, user)); |
34dc7c2f BB |
3932 | } |
3933 | ||
3934 | void * | |
0c66c32d | 3935 | dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user) |
34dc7c2f BB |
3936 | { |
3937 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
34dc7c2f | 3938 | |
bc4501f7 | 3939 | db->db_user_immediate_evict = TRUE; |
0c66c32d JG |
3940 | return (dmu_buf_set_user(db_fake, user)); |
3941 | } | |
34dc7c2f | 3942 | |
0c66c32d JG |
3943 | void * |
3944 | dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user) | |
3945 | { | |
3946 | return (dmu_buf_replace_user(db_fake, user, NULL)); | |
34dc7c2f BB |
3947 | } |
3948 | ||
3949 | void * | |
3950 | dmu_buf_get_user(dmu_buf_t *db_fake) | |
3951 | { | |
3952 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
34dc7c2f | 3953 | |
0c66c32d JG |
3954 | dbuf_verify_user(db, DBVU_NOT_EVICTING); |
3955 | return (db->db_user); | |
3956 | } | |
3957 | ||
3958 | void | |
493b6e56 | 3959 | dmu_buf_user_evict_wait(void) |
0c66c32d JG |
3960 | { |
3961 | taskq_wait(dbu_evict_taskq); | |
34dc7c2f BB |
3962 | } |
3963 | ||
03c6040b GW |
3964 | blkptr_t * |
3965 | dmu_buf_get_blkptr(dmu_buf_t *db) | |
3966 | { | |
3967 | dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; | |
3968 | return (dbi->db_blkptr); | |
3969 | } | |
3970 | ||
8bea9815 MA |
3971 | objset_t * |
3972 | dmu_buf_get_objset(dmu_buf_t *db) | |
3973 | { | |
3974 | dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; | |
3975 | return (dbi->db_objset); | |
3976 | } | |
3977 | ||
2bce8049 MA |
3978 | dnode_t * |
3979 | dmu_buf_dnode_enter(dmu_buf_t *db) | |
3980 | { | |
3981 | dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; | |
3982 | DB_DNODE_ENTER(dbi); | |
3983 | return (DB_DNODE(dbi)); | |
3984 | } | |
3985 | ||
3986 | void | |
3987 | dmu_buf_dnode_exit(dmu_buf_t *db) | |
3988 | { | |
3989 | dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; | |
3990 | DB_DNODE_EXIT(dbi); | |
3991 | } | |
3992 | ||
34dc7c2f BB |
3993 | static void |
3994 | dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db) | |
3995 | { | |
3996 | /* ASSERT(dmu_tx_is_syncing(tx) */ | |
3997 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
3998 | ||
3999 | if (db->db_blkptr != NULL) | |
4000 | return; | |
4001 | ||
428870ff | 4002 | if (db->db_blkid == DMU_SPILL_BLKID) { |
50c957f7 | 4003 | db->db_blkptr = DN_SPILL_BLKPTR(dn->dn_phys); |
428870ff BB |
4004 | BP_ZERO(db->db_blkptr); |
4005 | return; | |
4006 | } | |
34dc7c2f BB |
4007 | if (db->db_level == dn->dn_phys->dn_nlevels-1) { |
4008 | /* | |
4009 | * This buffer was allocated at a time when there was | |
4010 | * no available blkptrs from the dnode, or it was | |
e1cfd73f | 4011 | * inappropriate to hook it in (i.e., nlevels mismatch). |
34dc7c2f BB |
4012 | */ |
4013 | ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr); | |
4014 | ASSERT(db->db_parent == NULL); | |
4015 | db->db_parent = dn->dn_dbuf; | |
4016 | db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid]; | |
4017 | DBUF_VERIFY(db); | |
4018 | } else { | |
4019 | dmu_buf_impl_t *parent = db->db_parent; | |
4020 | int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; | |
4021 | ||
4022 | ASSERT(dn->dn_phys->dn_nlevels > 1); | |
4023 | if (parent == NULL) { | |
4024 | mutex_exit(&db->db_mtx); | |
4025 | rw_enter(&dn->dn_struct_rwlock, RW_READER); | |
fcff0f35 PD |
4026 | parent = dbuf_hold_level(dn, db->db_level + 1, |
4027 | db->db_blkid >> epbs, db); | |
34dc7c2f BB |
4028 | rw_exit(&dn->dn_struct_rwlock); |
4029 | mutex_enter(&db->db_mtx); | |
4030 | db->db_parent = parent; | |
4031 | } | |
4032 | db->db_blkptr = (blkptr_t *)parent->db.db_data + | |
4033 | (db->db_blkid & ((1ULL << epbs) - 1)); | |
4034 | DBUF_VERIFY(db); | |
4035 | } | |
4036 | } | |
4037 | ||
fa3922df MM |
4038 | static void |
4039 | dbuf_sync_bonus(dbuf_dirty_record_t *dr, dmu_tx_t *tx) | |
4040 | { | |
4041 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
4042 | void *data = dr->dt.dl.dr_data; | |
4043 | ||
4044 | ASSERT0(db->db_level); | |
4045 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
fa3922df MM |
4046 | ASSERT(db->db_blkid == DMU_BONUS_BLKID); |
4047 | ASSERT(data != NULL); | |
4048 | ||
ba67d821 | 4049 | dnode_t *dn = dr->dr_dnode; |
fa3922df MM |
4050 | ASSERT3U(DN_MAX_BONUS_LEN(dn->dn_phys), <=, |
4051 | DN_SLOTS_TO_BONUSLEN(dn->dn_phys->dn_extra_slots + 1)); | |
861166b0 | 4052 | memcpy(DN_BONUS(dn->dn_phys), data, DN_MAX_BONUS_LEN(dn->dn_phys)); |
fa3922df MM |
4053 | |
4054 | dbuf_sync_leaf_verify_bonus_dnode(dr); | |
4055 | ||
4056 | dbuf_undirty_bonus(dr); | |
4057 | dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg, B_FALSE); | |
4058 | } | |
4059 | ||
b5256303 | 4060 | /* |
0c03d21a MA |
4061 | * When syncing out a blocks of dnodes, adjust the block to deal with |
4062 | * encryption. Normally, we make sure the block is decrypted before writing | |
4063 | * it. If we have crypt params, then we are writing a raw (encrypted) block, | |
4064 | * from a raw receive. In this case, set the ARC buf's crypt params so | |
4065 | * that the BP will be filled with the correct byteorder, salt, iv, and mac. | |
b5256303 TC |
4066 | */ |
4067 | static void | |
0c03d21a | 4068 | dbuf_prepare_encrypted_dnode_leaf(dbuf_dirty_record_t *dr) |
b5256303 TC |
4069 | { |
4070 | int err; | |
4071 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
4072 | ||
4073 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
0c03d21a MA |
4074 | ASSERT3U(db->db.db_object, ==, DMU_META_DNODE_OBJECT); |
4075 | ASSERT3U(db->db_level, ==, 0); | |
b5256303 | 4076 | |
0c03d21a | 4077 | if (!db->db_objset->os_raw_receive && arc_is_encrypted(db->db_buf)) { |
a2c2ed1b TC |
4078 | zbookmark_phys_t zb; |
4079 | ||
b5256303 TC |
4080 | /* |
4081 | * Unfortunately, there is currently no mechanism for | |
4082 | * syncing context to handle decryption errors. An error | |
4083 | * here is only possible if an attacker maliciously | |
4084 | * changed a dnode block and updated the associated | |
4085 | * checksums going up the block tree. | |
4086 | */ | |
a2c2ed1b TC |
4087 | SET_BOOKMARK(&zb, dmu_objset_id(db->db_objset), |
4088 | db->db.db_object, db->db_level, db->db_blkid); | |
b5256303 | 4089 | err = arc_untransform(db->db_buf, db->db_objset->os_spa, |
a2c2ed1b | 4090 | &zb, B_TRUE); |
b5256303 TC |
4091 | if (err) |
4092 | panic("Invalid dnode block MAC"); | |
0c03d21a MA |
4093 | } else if (dr->dt.dl.dr_has_raw_params) { |
4094 | (void) arc_release(dr->dt.dl.dr_data, db); | |
4095 | arc_convert_to_raw(dr->dt.dl.dr_data, | |
4096 | dmu_objset_id(db->db_objset), | |
4097 | dr->dt.dl.dr_byteorder, DMU_OT_DNODE, | |
4098 | dr->dt.dl.dr_salt, dr->dt.dl.dr_iv, dr->dt.dl.dr_mac); | |
b5256303 TC |
4099 | } |
4100 | } | |
4101 | ||
d1d7e268 MK |
4102 | /* |
4103 | * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it | |
60948de1 BB |
4104 | * is critical the we not allow the compiler to inline this function in to |
4105 | * dbuf_sync_list() thereby drastically bloating the stack usage. | |
4106 | */ | |
4107 | noinline static void | |
34dc7c2f BB |
4108 | dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx) |
4109 | { | |
4110 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
ba67d821 | 4111 | dnode_t *dn = dr->dr_dnode; |
34dc7c2f BB |
4112 | |
4113 | ASSERT(dmu_tx_is_syncing(tx)); | |
4114 | ||
4115 | dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); | |
4116 | ||
4117 | mutex_enter(&db->db_mtx); | |
4118 | ||
4119 | ASSERT(db->db_level > 0); | |
4120 | DBUF_VERIFY(db); | |
4121 | ||
e49f1e20 | 4122 | /* Read the block if it hasn't been read yet. */ |
34dc7c2f BB |
4123 | if (db->db_buf == NULL) { |
4124 | mutex_exit(&db->db_mtx); | |
4125 | (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED); | |
4126 | mutex_enter(&db->db_mtx); | |
4127 | } | |
4128 | ASSERT3U(db->db_state, ==, DB_CACHED); | |
34dc7c2f BB |
4129 | ASSERT(db->db_buf != NULL); |
4130 | ||
e49f1e20 | 4131 | /* Indirect block size must match what the dnode thinks it is. */ |
572e2857 | 4132 | ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); |
34dc7c2f BB |
4133 | dbuf_check_blkptr(dn, db); |
4134 | ||
e49f1e20 | 4135 | /* Provide the pending dirty record to child dbufs */ |
34dc7c2f BB |
4136 | db->db_data_pending = dr; |
4137 | ||
34dc7c2f | 4138 | mutex_exit(&db->db_mtx); |
a1d477c2 | 4139 | |
b128c09f | 4140 | dbuf_write(dr, db->db_buf, tx); |
34dc7c2f | 4141 | |
ba67d821 | 4142 | zio_t *zio = dr->dr_zio; |
34dc7c2f | 4143 | mutex_enter(&dr->dt.di.dr_mtx); |
4bda3bd0 | 4144 | dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx); |
34dc7c2f BB |
4145 | ASSERT(list_head(&dr->dt.di.dr_children) == NULL); |
4146 | mutex_exit(&dr->dt.di.dr_mtx); | |
4147 | zio_nowait(zio); | |
4148 | } | |
4149 | ||
0e37a0f4 SD |
4150 | /* |
4151 | * Verify that the size of the data in our bonus buffer does not exceed | |
4152 | * its recorded size. | |
4153 | * | |
4154 | * The purpose of this verification is to catch any cases in development | |
4155 | * where the size of a phys structure (i.e space_map_phys_t) grows and, | |
4156 | * due to incorrect feature management, older pools expect to read more | |
4157 | * data even though they didn't actually write it to begin with. | |
4158 | * | |
4159 | * For a example, this would catch an error in the feature logic where we | |
4160 | * open an older pool and we expect to write the space map histogram of | |
4161 | * a space map with size SPACE_MAP_SIZE_V0. | |
4162 | */ | |
4163 | static void | |
4164 | dbuf_sync_leaf_verify_bonus_dnode(dbuf_dirty_record_t *dr) | |
4165 | { | |
fa3922df | 4166 | #ifdef ZFS_DEBUG |
ba67d821 | 4167 | dnode_t *dn = dr->dr_dnode; |
0e37a0f4 SD |
4168 | |
4169 | /* | |
4170 | * Encrypted bonus buffers can have data past their bonuslen. | |
4171 | * Skip the verification of these blocks. | |
4172 | */ | |
4173 | if (DMU_OT_IS_ENCRYPTED(dn->dn_bonustype)) | |
4174 | return; | |
4175 | ||
4176 | uint16_t bonuslen = dn->dn_phys->dn_bonuslen; | |
4177 | uint16_t maxbonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots); | |
4178 | ASSERT3U(bonuslen, <=, maxbonuslen); | |
4179 | ||
4180 | arc_buf_t *datap = dr->dt.dl.dr_data; | |
4181 | char *datap_end = ((char *)datap) + bonuslen; | |
4182 | char *datap_max = ((char *)datap) + maxbonuslen; | |
4183 | ||
4184 | /* ensure that everything is zero after our data */ | |
4185 | for (; datap_end < datap_max; datap_end++) | |
4186 | ASSERT(*datap_end == 0); | |
0e37a0f4 | 4187 | #endif |
fa3922df | 4188 | } |
0e37a0f4 | 4189 | |
ba67d821 MA |
4190 | static blkptr_t * |
4191 | dbuf_lightweight_bp(dbuf_dirty_record_t *dr) | |
4192 | { | |
4193 | /* This must be a lightweight dirty record. */ | |
4194 | ASSERT3P(dr->dr_dbuf, ==, NULL); | |
4195 | dnode_t *dn = dr->dr_dnode; | |
4196 | ||
4197 | if (dn->dn_phys->dn_nlevels == 1) { | |
4198 | VERIFY3U(dr->dt.dll.dr_blkid, <, dn->dn_phys->dn_nblkptr); | |
4199 | return (&dn->dn_phys->dn_blkptr[dr->dt.dll.dr_blkid]); | |
4200 | } else { | |
4201 | dmu_buf_impl_t *parent_db = dr->dr_parent->dr_dbuf; | |
4202 | int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; | |
4203 | VERIFY3U(parent_db->db_level, ==, 1); | |
4204 | VERIFY3P(parent_db->db_dnode_handle->dnh_dnode, ==, dn); | |
4205 | VERIFY3U(dr->dt.dll.dr_blkid >> epbs, ==, parent_db->db_blkid); | |
4206 | blkptr_t *bp = parent_db->db.db_data; | |
4207 | return (&bp[dr->dt.dll.dr_blkid & ((1 << epbs) - 1)]); | |
4208 | } | |
4209 | } | |
4210 | ||
4211 | static void | |
4212 | dbuf_lightweight_ready(zio_t *zio) | |
4213 | { | |
4214 | dbuf_dirty_record_t *dr = zio->io_private; | |
4215 | blkptr_t *bp = zio->io_bp; | |
4216 | ||
4217 | if (zio->io_error != 0) | |
4218 | return; | |
4219 | ||
4220 | dnode_t *dn = dr->dr_dnode; | |
4221 | ||
4222 | blkptr_t *bp_orig = dbuf_lightweight_bp(dr); | |
4223 | spa_t *spa = dmu_objset_spa(dn->dn_objset); | |
4224 | int64_t delta = bp_get_dsize_sync(spa, bp) - | |
4225 | bp_get_dsize_sync(spa, bp_orig); | |
4226 | dnode_diduse_space(dn, delta); | |
4227 | ||
4228 | uint64_t blkid = dr->dt.dll.dr_blkid; | |
4229 | mutex_enter(&dn->dn_mtx); | |
4230 | if (blkid > dn->dn_phys->dn_maxblkid) { | |
4231 | ASSERT0(dn->dn_objset->os_raw_receive); | |
4232 | dn->dn_phys->dn_maxblkid = blkid; | |
4233 | } | |
4234 | mutex_exit(&dn->dn_mtx); | |
4235 | ||
4236 | if (!BP_IS_EMBEDDED(bp)) { | |
4237 | uint64_t fill = BP_IS_HOLE(bp) ? 0 : 1; | |
4238 | BP_SET_FILL(bp, fill); | |
4239 | } | |
4240 | ||
4241 | dmu_buf_impl_t *parent_db; | |
4242 | EQUIV(dr->dr_parent == NULL, dn->dn_phys->dn_nlevels == 1); | |
4243 | if (dr->dr_parent == NULL) { | |
4244 | parent_db = dn->dn_dbuf; | |
4245 | } else { | |
4246 | parent_db = dr->dr_parent->dr_dbuf; | |
4247 | } | |
4248 | rw_enter(&parent_db->db_rwlock, RW_WRITER); | |
4249 | *bp_orig = *bp; | |
4250 | rw_exit(&parent_db->db_rwlock); | |
4251 | } | |
4252 | ||
4253 | static void | |
4254 | dbuf_lightweight_physdone(zio_t *zio) | |
4255 | { | |
4256 | dbuf_dirty_record_t *dr = zio->io_private; | |
4257 | dsl_pool_t *dp = spa_get_dsl(zio->io_spa); | |
4258 | ASSERT3U(dr->dr_txg, ==, zio->io_txg); | |
4259 | ||
4260 | /* | |
4261 | * The callback will be called io_phys_children times. Retire one | |
4262 | * portion of our dirty space each time we are called. Any rounding | |
4263 | * error will be cleaned up by dbuf_lightweight_done(). | |
4264 | */ | |
4265 | int delta = dr->dr_accounted / zio->io_phys_children; | |
4266 | dsl_pool_undirty_space(dp, delta, zio->io_txg); | |
4267 | } | |
4268 | ||
4269 | static void | |
4270 | dbuf_lightweight_done(zio_t *zio) | |
4271 | { | |
4272 | dbuf_dirty_record_t *dr = zio->io_private; | |
4273 | ||
4274 | VERIFY0(zio->io_error); | |
4275 | ||
4276 | objset_t *os = dr->dr_dnode->dn_objset; | |
4277 | dmu_tx_t *tx = os->os_synctx; | |
4278 | ||
4279 | if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) { | |
4280 | ASSERT(BP_EQUAL(zio->io_bp, &zio->io_bp_orig)); | |
4281 | } else { | |
4282 | dsl_dataset_t *ds = os->os_dsl_dataset; | |
4283 | (void) dsl_dataset_block_kill(ds, &zio->io_bp_orig, tx, B_TRUE); | |
4284 | dsl_dataset_block_born(ds, zio->io_bp, tx); | |
4285 | } | |
4286 | ||
4287 | /* | |
4288 | * See comment in dbuf_write_done(). | |
4289 | */ | |
4290 | if (zio->io_phys_children == 0) { | |
4291 | dsl_pool_undirty_space(dmu_objset_pool(os), | |
4292 | dr->dr_accounted, zio->io_txg); | |
4293 | } else { | |
4294 | dsl_pool_undirty_space(dmu_objset_pool(os), | |
4295 | dr->dr_accounted % zio->io_phys_children, zio->io_txg); | |
4296 | } | |
4297 | ||
4298 | abd_free(dr->dt.dll.dr_abd); | |
4299 | kmem_free(dr, sizeof (*dr)); | |
4300 | } | |
4301 | ||
4302 | noinline static void | |
4303 | dbuf_sync_lightweight(dbuf_dirty_record_t *dr, dmu_tx_t *tx) | |
4304 | { | |
4305 | dnode_t *dn = dr->dr_dnode; | |
4306 | zio_t *pio; | |
4307 | if (dn->dn_phys->dn_nlevels == 1) { | |
4308 | pio = dn->dn_zio; | |
4309 | } else { | |
4310 | pio = dr->dr_parent->dr_zio; | |
4311 | } | |
4312 | ||
4313 | zbookmark_phys_t zb = { | |
4314 | .zb_objset = dmu_objset_id(dn->dn_objset), | |
4315 | .zb_object = dn->dn_object, | |
4316 | .zb_level = 0, | |
4317 | .zb_blkid = dr->dt.dll.dr_blkid, | |
4318 | }; | |
4319 | ||
4320 | /* | |
4321 | * See comment in dbuf_write(). This is so that zio->io_bp_orig | |
4322 | * will have the old BP in dbuf_lightweight_done(). | |
4323 | */ | |
4324 | dr->dr_bp_copy = *dbuf_lightweight_bp(dr); | |
4325 | ||
4326 | dr->dr_zio = zio_write(pio, dmu_objset_spa(dn->dn_objset), | |
4327 | dmu_tx_get_txg(tx), &dr->dr_bp_copy, dr->dt.dll.dr_abd, | |
4328 | dn->dn_datablksz, abd_get_size(dr->dt.dll.dr_abd), | |
4329 | &dr->dt.dll.dr_props, dbuf_lightweight_ready, NULL, | |
4330 | dbuf_lightweight_physdone, dbuf_lightweight_done, dr, | |
4331 | ZIO_PRIORITY_ASYNC_WRITE, | |
4332 | ZIO_FLAG_MUSTSUCCEED | dr->dt.dll.dr_flags, &zb); | |
4333 | ||
4334 | zio_nowait(dr->dr_zio); | |
4335 | } | |
4336 | ||
d1d7e268 MK |
4337 | /* |
4338 | * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is | |
60948de1 BB |
4339 | * critical the we not allow the compiler to inline this function in to |
4340 | * dbuf_sync_list() thereby drastically bloating the stack usage. | |
4341 | */ | |
4342 | noinline static void | |
34dc7c2f BB |
4343 | dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx) |
4344 | { | |
4345 | arc_buf_t **datap = &dr->dt.dl.dr_data; | |
4346 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
ba67d821 | 4347 | dnode_t *dn = dr->dr_dnode; |
572e2857 | 4348 | objset_t *os; |
34dc7c2f | 4349 | uint64_t txg = tx->tx_txg; |
34dc7c2f BB |
4350 | |
4351 | ASSERT(dmu_tx_is_syncing(tx)); | |
4352 | ||
4353 | dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); | |
4354 | ||
4355 | mutex_enter(&db->db_mtx); | |
4356 | /* | |
4357 | * To be synced, we must be dirtied. But we | |
4358 | * might have been freed after the dirty. | |
4359 | */ | |
4360 | if (db->db_state == DB_UNCACHED) { | |
4361 | /* This buffer has been freed since it was dirtied */ | |
4362 | ASSERT(db->db.db_data == NULL); | |
4363 | } else if (db->db_state == DB_FILL) { | |
4364 | /* This buffer was freed and is now being re-filled */ | |
4365 | ASSERT(db->db.db_data != dr->dt.dl.dr_data); | |
4366 | } else { | |
b128c09f | 4367 | ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL); |
34dc7c2f BB |
4368 | } |
4369 | DBUF_VERIFY(db); | |
4370 | ||
428870ff BB |
4371 | if (db->db_blkid == DMU_SPILL_BLKID) { |
4372 | mutex_enter(&dn->dn_mtx); | |
81edd3e8 P |
4373 | if (!(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) { |
4374 | /* | |
4375 | * In the previous transaction group, the bonus buffer | |
4376 | * was entirely used to store the attributes for the | |
4377 | * dnode which overrode the dn_spill field. However, | |
4378 | * when adding more attributes to the file a spill | |
4379 | * block was required to hold the extra attributes. | |
4380 | * | |
4381 | * Make sure to clear the garbage left in the dn_spill | |
4382 | * field from the previous attributes in the bonus | |
4383 | * buffer. Otherwise, after writing out the spill | |
4384 | * block to the new allocated dva, it will free | |
4385 | * the old block pointed to by the invalid dn_spill. | |
4386 | */ | |
4387 | db->db_blkptr = NULL; | |
4388 | } | |
428870ff BB |
4389 | dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR; |
4390 | mutex_exit(&dn->dn_mtx); | |
4391 | } | |
4392 | ||
34dc7c2f BB |
4393 | /* |
4394 | * If this is a bonus buffer, simply copy the bonus data into the | |
4395 | * dnode. It will be written out when the dnode is synced (and it | |
4396 | * will be synced, since it must have been dirty for dbuf_sync to | |
4397 | * be called). | |
4398 | */ | |
428870ff | 4399 | if (db->db_blkid == DMU_BONUS_BLKID) { |
428870ff | 4400 | ASSERT(dr->dr_dbuf == db); |
fa3922df | 4401 | dbuf_sync_bonus(dr, tx); |
34dc7c2f BB |
4402 | return; |
4403 | } | |
4404 | ||
572e2857 BB |
4405 | os = dn->dn_objset; |
4406 | ||
34dc7c2f BB |
4407 | /* |
4408 | * This function may have dropped the db_mtx lock allowing a dmu_sync | |
4409 | * operation to sneak in. As a result, we need to ensure that we | |
4410 | * don't check the dr_override_state until we have returned from | |
4411 | * dbuf_check_blkptr. | |
4412 | */ | |
4413 | dbuf_check_blkptr(dn, db); | |
4414 | ||
4415 | /* | |
572e2857 | 4416 | * If this buffer is in the middle of an immediate write, |
34dc7c2f BB |
4417 | * wait for the synchronous IO to complete. |
4418 | */ | |
4419 | while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) { | |
4420 | ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT); | |
4421 | cv_wait(&db->db_changed, &db->db_mtx); | |
4422 | ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN); | |
4423 | } | |
4424 | ||
b5256303 TC |
4425 | /* |
4426 | * If this is a dnode block, ensure it is appropriately encrypted | |
4427 | * or decrypted, depending on what we are writing to it this txg. | |
4428 | */ | |
4429 | if (os->os_encrypted && dn->dn_object == DMU_META_DNODE_OBJECT) | |
0c03d21a | 4430 | dbuf_prepare_encrypted_dnode_leaf(dr); |
b5256303 | 4431 | |
9babb374 BB |
4432 | if (db->db_state != DB_NOFILL && |
4433 | dn->dn_object != DMU_META_DNODE_OBJECT && | |
424fd7c3 | 4434 | zfs_refcount_count(&db->db_holds) > 1 && |
428870ff | 4435 | dr->dt.dl.dr_override_state != DR_OVERRIDDEN && |
9babb374 BB |
4436 | *datap == db->db_buf) { |
4437 | /* | |
4438 | * If this buffer is currently "in use" (i.e., there | |
4439 | * are active holds and db_data still references it), | |
4440 | * then make a copy before we start the write so that | |
4441 | * any modifications from the open txg will not leak | |
4442 | * into this write. | |
4443 | * | |
4444 | * NOTE: this copy does not need to be made for | |
4445 | * objects only modified in the syncing context (e.g. | |
4446 | * DNONE_DNODE blocks). | |
4447 | */ | |
a81b8124 AR |
4448 | int psize = arc_buf_size(*datap); |
4449 | int lsize = arc_buf_lsize(*datap); | |
4450 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); | |
4451 | enum zio_compress compress_type = arc_get_compression(*datap); | |
4452 | uint8_t complevel = arc_get_complevel(*datap); | |
4453 | ||
4454 | if (arc_is_encrypted(*datap)) { | |
4455 | boolean_t byteorder; | |
4456 | uint8_t salt[ZIO_DATA_SALT_LEN]; | |
4457 | uint8_t iv[ZIO_DATA_IV_LEN]; | |
4458 | uint8_t mac[ZIO_DATA_MAC_LEN]; | |
4459 | ||
4460 | arc_get_raw_params(*datap, &byteorder, salt, iv, mac); | |
4461 | *datap = arc_alloc_raw_buf(os->os_spa, db, | |
4462 | dmu_objset_id(os), byteorder, salt, iv, mac, | |
4463 | dn->dn_type, psize, lsize, compress_type, | |
4464 | complevel); | |
4465 | } else if (compress_type != ZIO_COMPRESS_OFF) { | |
4466 | ASSERT3U(type, ==, ARC_BUFC_DATA); | |
4467 | *datap = arc_alloc_compressed_buf(os->os_spa, db, | |
4468 | psize, lsize, compress_type, complevel); | |
4469 | } else { | |
4470 | *datap = arc_alloc_buf(os->os_spa, db, type, psize); | |
4471 | } | |
861166b0 | 4472 | memcpy((*datap)->b_data, db->db.db_data, psize); |
b128c09f | 4473 | } |
34dc7c2f BB |
4474 | db->db_data_pending = dr; |
4475 | ||
4476 | mutex_exit(&db->db_mtx); | |
4477 | ||
b128c09f | 4478 | dbuf_write(dr, *datap, tx); |
34dc7c2f BB |
4479 | |
4480 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
572e2857 | 4481 | if (dn->dn_object == DMU_META_DNODE_OBJECT) { |
3fa93bb8 | 4482 | list_insert_tail(&dn->dn_dirty_records[txg & TXG_MASK], dr); |
572e2857 | 4483 | } else { |
34dc7c2f | 4484 | zio_nowait(dr->dr_zio); |
572e2857 | 4485 | } |
34dc7c2f BB |
4486 | } |
4487 | ||
4488 | void | |
4bda3bd0 | 4489 | dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx) |
34dc7c2f BB |
4490 | { |
4491 | dbuf_dirty_record_t *dr; | |
4492 | ||
c65aa5b2 | 4493 | while ((dr = list_head(list))) { |
34dc7c2f BB |
4494 | if (dr->dr_zio != NULL) { |
4495 | /* | |
4496 | * If we find an already initialized zio then we | |
4497 | * are processing the meta-dnode, and we have finished. | |
4498 | * The dbufs for all dnodes are put back on the list | |
4499 | * during processing, so that we can zio_wait() | |
4500 | * these IOs after initiating all child IOs. | |
4501 | */ | |
4502 | ASSERT3U(dr->dr_dbuf->db.db_object, ==, | |
4503 | DMU_META_DNODE_OBJECT); | |
4504 | break; | |
4505 | } | |
4506 | list_remove(list, dr); | |
ba67d821 MA |
4507 | if (dr->dr_dbuf == NULL) { |
4508 | dbuf_sync_lightweight(dr, tx); | |
4509 | } else { | |
4510 | if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID && | |
4511 | dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) { | |
4512 | VERIFY3U(dr->dr_dbuf->db_level, ==, level); | |
4513 | } | |
4514 | if (dr->dr_dbuf->db_level > 0) | |
4515 | dbuf_sync_indirect(dr, tx); | |
4516 | else | |
4517 | dbuf_sync_leaf(dr, tx); | |
4518 | } | |
34dc7c2f BB |
4519 | } |
4520 | } | |
4521 | ||
34dc7c2f BB |
4522 | static void |
4523 | dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb) | |
4524 | { | |
14e4e3cb | 4525 | (void) buf; |
34dc7c2f | 4526 | dmu_buf_impl_t *db = vdb; |
572e2857 | 4527 | dnode_t *dn; |
b128c09f | 4528 | blkptr_t *bp = zio->io_bp; |
34dc7c2f | 4529 | blkptr_t *bp_orig = &zio->io_bp_orig; |
428870ff BB |
4530 | spa_t *spa = zio->io_spa; |
4531 | int64_t delta; | |
34dc7c2f | 4532 | uint64_t fill = 0; |
428870ff | 4533 | int i; |
34dc7c2f | 4534 | |
463a8cfe AR |
4535 | ASSERT3P(db->db_blkptr, !=, NULL); |
4536 | ASSERT3P(&db->db_data_pending->dr_bp_copy, ==, bp); | |
b128c09f | 4537 | |
572e2857 BB |
4538 | DB_DNODE_ENTER(db); |
4539 | dn = DB_DNODE(db); | |
428870ff BB |
4540 | delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig); |
4541 | dnode_diduse_space(dn, delta - zio->io_prev_space_delta); | |
4542 | zio->io_prev_space_delta = delta; | |
34dc7c2f | 4543 | |
b0bc7a84 MG |
4544 | if (bp->blk_birth != 0) { |
4545 | ASSERT((db->db_blkid != DMU_SPILL_BLKID && | |
4546 | BP_GET_TYPE(bp) == dn->dn_type) || | |
4547 | (db->db_blkid == DMU_SPILL_BLKID && | |
9b67f605 MA |
4548 | BP_GET_TYPE(bp) == dn->dn_bonustype) || |
4549 | BP_IS_EMBEDDED(bp)); | |
b0bc7a84 | 4550 | ASSERT(BP_GET_LEVEL(bp) == db->db_level); |
34dc7c2f BB |
4551 | } |
4552 | ||
4553 | mutex_enter(&db->db_mtx); | |
4554 | ||
428870ff BB |
4555 | #ifdef ZFS_DEBUG |
4556 | if (db->db_blkid == DMU_SPILL_BLKID) { | |
428870ff | 4557 | ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); |
463a8cfe | 4558 | ASSERT(!(BP_IS_HOLE(bp)) && |
50c957f7 | 4559 | db->db_blkptr == DN_SPILL_BLKPTR(dn->dn_phys)); |
428870ff BB |
4560 | } |
4561 | #endif | |
4562 | ||
34dc7c2f BB |
4563 | if (db->db_level == 0) { |
4564 | mutex_enter(&dn->dn_mtx); | |
428870ff | 4565 | if (db->db_blkid > dn->dn_phys->dn_maxblkid && |
69830602 TC |
4566 | db->db_blkid != DMU_SPILL_BLKID) { |
4567 | ASSERT0(db->db_objset->os_raw_receive); | |
34dc7c2f | 4568 | dn->dn_phys->dn_maxblkid = db->db_blkid; |
69830602 | 4569 | } |
34dc7c2f BB |
4570 | mutex_exit(&dn->dn_mtx); |
4571 | ||
4572 | if (dn->dn_type == DMU_OT_DNODE) { | |
50c957f7 NB |
4573 | i = 0; |
4574 | while (i < db->db.db_size) { | |
817b1b6e MA |
4575 | dnode_phys_t *dnp = |
4576 | (void *)(((char *)db->db.db_data) + i); | |
50c957f7 NB |
4577 | |
4578 | i += DNODE_MIN_SIZE; | |
4579 | if (dnp->dn_type != DMU_OT_NONE) { | |
34dc7c2f | 4580 | fill++; |
50c957f7 NB |
4581 | i += dnp->dn_extra_slots * |
4582 | DNODE_MIN_SIZE; | |
4583 | } | |
34dc7c2f BB |
4584 | } |
4585 | } else { | |
b0bc7a84 MG |
4586 | if (BP_IS_HOLE(bp)) { |
4587 | fill = 0; | |
4588 | } else { | |
4589 | fill = 1; | |
4590 | } | |
34dc7c2f BB |
4591 | } |
4592 | } else { | |
b128c09f | 4593 | blkptr_t *ibp = db->db.db_data; |
34dc7c2f | 4594 | ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); |
b128c09f BB |
4595 | for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) { |
4596 | if (BP_IS_HOLE(ibp)) | |
34dc7c2f | 4597 | continue; |
9b67f605 | 4598 | fill += BP_GET_FILL(ibp); |
34dc7c2f BB |
4599 | } |
4600 | } | |
572e2857 | 4601 | DB_DNODE_EXIT(db); |
34dc7c2f | 4602 | |
9b67f605 | 4603 | if (!BP_IS_EMBEDDED(bp)) |
b5256303 | 4604 | BP_SET_FILL(bp, fill); |
34dc7c2f BB |
4605 | |
4606 | mutex_exit(&db->db_mtx); | |
463a8cfe | 4607 | |
f664f1ee | 4608 | db_lock_type_t dblt = dmu_buf_lock_parent(db, RW_WRITER, FTAG); |
463a8cfe | 4609 | *db->db_blkptr = *bp; |
f664f1ee | 4610 | dmu_buf_unlock_parent(db, dblt, FTAG); |
34dc7c2f BB |
4611 | } |
4612 | ||
bc77ba73 PD |
4613 | /* |
4614 | * This function gets called just prior to running through the compression | |
4615 | * stage of the zio pipeline. If we're an indirect block comprised of only | |
4616 | * holes, then we want this indirect to be compressed away to a hole. In | |
4617 | * order to do that we must zero out any information about the holes that | |
4618 | * this indirect points to prior to before we try to compress it. | |
4619 | */ | |
4620 | static void | |
4621 | dbuf_write_children_ready(zio_t *zio, arc_buf_t *buf, void *vdb) | |
4622 | { | |
14e4e3cb | 4623 | (void) zio, (void) buf; |
bc77ba73 PD |
4624 | dmu_buf_impl_t *db = vdb; |
4625 | dnode_t *dn; | |
4626 | blkptr_t *bp; | |
721ed0ee | 4627 | unsigned int epbs, i; |
bc77ba73 PD |
4628 | |
4629 | ASSERT3U(db->db_level, >, 0); | |
4630 | DB_DNODE_ENTER(db); | |
4631 | dn = DB_DNODE(db); | |
4632 | epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; | |
721ed0ee | 4633 | ASSERT3U(epbs, <, 31); |
bc77ba73 PD |
4634 | |
4635 | /* Determine if all our children are holes */ | |
3f93077b | 4636 | for (i = 0, bp = db->db.db_data; i < 1ULL << epbs; i++, bp++) { |
bc77ba73 PD |
4637 | if (!BP_IS_HOLE(bp)) |
4638 | break; | |
4639 | } | |
4640 | ||
4641 | /* | |
4642 | * If all the children are holes, then zero them all out so that | |
4643 | * we may get compressed away. | |
4644 | */ | |
3f93077b | 4645 | if (i == 1ULL << epbs) { |
721ed0ee GM |
4646 | /* |
4647 | * We only found holes. Grab the rwlock to prevent | |
4648 | * anybody from reading the blocks we're about to | |
4649 | * zero out. | |
4650 | */ | |
f664f1ee | 4651 | rw_enter(&db->db_rwlock, RW_WRITER); |
861166b0 | 4652 | memset(db->db.db_data, 0, db->db.db_size); |
f664f1ee | 4653 | rw_exit(&db->db_rwlock); |
bc77ba73 PD |
4654 | } |
4655 | DB_DNODE_EXIT(db); | |
4656 | } | |
4657 | ||
e8b96c60 MA |
4658 | /* |
4659 | * The SPA will call this callback several times for each zio - once | |
4660 | * for every physical child i/o (zio->io_phys_children times). This | |
4661 | * allows the DMU to monitor the progress of each logical i/o. For example, | |
4662 | * there may be 2 copies of an indirect block, or many fragments of a RAID-Z | |
4663 | * block. There may be a long delay before all copies/fragments are completed, | |
4664 | * so this callback allows us to retire dirty space gradually, as the physical | |
4665 | * i/os complete. | |
4666 | */ | |
e8b96c60 MA |
4667 | static void |
4668 | dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg) | |
4669 | { | |
14e4e3cb | 4670 | (void) buf; |
e8b96c60 MA |
4671 | dmu_buf_impl_t *db = arg; |
4672 | objset_t *os = db->db_objset; | |
4673 | dsl_pool_t *dp = dmu_objset_pool(os); | |
4674 | dbuf_dirty_record_t *dr; | |
4675 | int delta = 0; | |
4676 | ||
4677 | dr = db->db_data_pending; | |
4678 | ASSERT3U(dr->dr_txg, ==, zio->io_txg); | |
4679 | ||
4680 | /* | |
4681 | * The callback will be called io_phys_children times. Retire one | |
4682 | * portion of our dirty space each time we are called. Any rounding | |
0f8ff49e | 4683 | * error will be cleaned up by dbuf_write_done(). |
e8b96c60 MA |
4684 | */ |
4685 | delta = dr->dr_accounted / zio->io_phys_children; | |
4686 | dsl_pool_undirty_space(dp, delta, zio->io_txg); | |
4687 | } | |
4688 | ||
34dc7c2f BB |
4689 | static void |
4690 | dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb) | |
4691 | { | |
14e4e3cb | 4692 | (void) buf; |
34dc7c2f | 4693 | dmu_buf_impl_t *db = vdb; |
428870ff | 4694 | blkptr_t *bp_orig = &zio->io_bp_orig; |
b0bc7a84 MG |
4695 | blkptr_t *bp = db->db_blkptr; |
4696 | objset_t *os = db->db_objset; | |
4697 | dmu_tx_t *tx = os->os_synctx; | |
34dc7c2f | 4698 | |
c99c9001 | 4699 | ASSERT0(zio->io_error); |
428870ff BB |
4700 | ASSERT(db->db_blkptr == bp); |
4701 | ||
03c6040b GW |
4702 | /* |
4703 | * For nopwrites and rewrites we ensure that the bp matches our | |
4704 | * original and bypass all the accounting. | |
4705 | */ | |
4706 | if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) { | |
428870ff BB |
4707 | ASSERT(BP_EQUAL(bp, bp_orig)); |
4708 | } else { | |
b0bc7a84 | 4709 | dsl_dataset_t *ds = os->os_dsl_dataset; |
428870ff BB |
4710 | (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE); |
4711 | dsl_dataset_block_born(ds, bp, tx); | |
4712 | } | |
34dc7c2f BB |
4713 | |
4714 | mutex_enter(&db->db_mtx); | |
4715 | ||
428870ff BB |
4716 | DBUF_VERIFY(db); |
4717 | ||
ba67d821 MA |
4718 | dbuf_dirty_record_t *dr = db->db_data_pending; |
4719 | dnode_t *dn = dr->dr_dnode; | |
34dc7c2f | 4720 | ASSERT(!list_link_active(&dr->dr_dirty_node)); |
428870ff | 4721 | ASSERT(dr->dr_dbuf == db); |
cccbed9f MM |
4722 | ASSERT(list_next(&db->db_dirty_records, dr) == NULL); |
4723 | list_remove(&db->db_dirty_records, dr); | |
34dc7c2f | 4724 | |
428870ff BB |
4725 | #ifdef ZFS_DEBUG |
4726 | if (db->db_blkid == DMU_SPILL_BLKID) { | |
428870ff BB |
4727 | ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); |
4728 | ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && | |
50c957f7 | 4729 | db->db_blkptr == DN_SPILL_BLKPTR(dn->dn_phys)); |
428870ff BB |
4730 | } |
4731 | #endif | |
4732 | ||
34dc7c2f | 4733 | if (db->db_level == 0) { |
428870ff | 4734 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f | 4735 | ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); |
b128c09f BB |
4736 | if (db->db_state != DB_NOFILL) { |
4737 | if (dr->dt.dl.dr_data != db->db_buf) | |
d3c2ae1c | 4738 | arc_buf_destroy(dr->dt.dl.dr_data, db); |
b128c09f | 4739 | } |
34dc7c2f | 4740 | } else { |
34dc7c2f | 4741 | ASSERT(list_head(&dr->dt.di.dr_children) == NULL); |
b0bc7a84 | 4742 | ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift); |
34dc7c2f | 4743 | if (!BP_IS_HOLE(db->db_blkptr)) { |
2a8ba608 MM |
4744 | int epbs __maybe_unused = dn->dn_phys->dn_indblkshift - |
4745 | SPA_BLKPTRSHIFT; | |
b0bc7a84 MG |
4746 | ASSERT3U(db->db_blkid, <=, |
4747 | dn->dn_phys->dn_maxblkid >> (db->db_level * epbs)); | |
34dc7c2f BB |
4748 | ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, |
4749 | db->db.db_size); | |
34dc7c2f BB |
4750 | } |
4751 | mutex_destroy(&dr->dt.di.dr_mtx); | |
4752 | list_destroy(&dr->dt.di.dr_children); | |
4753 | } | |
34dc7c2f BB |
4754 | |
4755 | cv_broadcast(&db->db_changed); | |
4756 | ASSERT(db->db_dirtycnt > 0); | |
4757 | db->db_dirtycnt -= 1; | |
4758 | db->db_data_pending = NULL; | |
3d503a76 | 4759 | dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg, B_FALSE); |
0f8ff49e SD |
4760 | |
4761 | /* | |
4762 | * If we didn't do a physical write in this ZIO and we | |
4763 | * still ended up here, it means that the space of the | |
4764 | * dbuf that we just released (and undirtied) above hasn't | |
4765 | * been marked as undirtied in the pool's accounting. | |
4766 | * | |
4767 | * Thus, we undirty that space in the pool's view of the | |
4768 | * world here. For physical writes this type of update | |
4769 | * happens in dbuf_write_physdone(). | |
4770 | * | |
4771 | * If we did a physical write, cleanup any rounding errors | |
4772 | * that came up due to writing multiple copies of a block | |
4773 | * on disk [see dbuf_write_physdone()]. | |
4774 | */ | |
4775 | if (zio->io_phys_children == 0) { | |
4776 | dsl_pool_undirty_space(dmu_objset_pool(os), | |
4777 | dr->dr_accounted, zio->io_txg); | |
4778 | } else { | |
4779 | dsl_pool_undirty_space(dmu_objset_pool(os), | |
4780 | dr->dr_accounted % zio->io_phys_children, zio->io_txg); | |
4781 | } | |
4782 | ||
4783 | kmem_free(dr, sizeof (dbuf_dirty_record_t)); | |
428870ff BB |
4784 | } |
4785 | ||
4786 | static void | |
4787 | dbuf_write_nofill_ready(zio_t *zio) | |
4788 | { | |
4789 | dbuf_write_ready(zio, NULL, zio->io_private); | |
4790 | } | |
4791 | ||
4792 | static void | |
4793 | dbuf_write_nofill_done(zio_t *zio) | |
4794 | { | |
4795 | dbuf_write_done(zio, NULL, zio->io_private); | |
4796 | } | |
4797 | ||
4798 | static void | |
4799 | dbuf_write_override_ready(zio_t *zio) | |
4800 | { | |
4801 | dbuf_dirty_record_t *dr = zio->io_private; | |
4802 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
4803 | ||
4804 | dbuf_write_ready(zio, NULL, db); | |
4805 | } | |
4806 | ||
4807 | static void | |
4808 | dbuf_write_override_done(zio_t *zio) | |
4809 | { | |
4810 | dbuf_dirty_record_t *dr = zio->io_private; | |
4811 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
4812 | blkptr_t *obp = &dr->dt.dl.dr_overridden_by; | |
4813 | ||
4814 | mutex_enter(&db->db_mtx); | |
4815 | if (!BP_EQUAL(zio->io_bp, obp)) { | |
4816 | if (!BP_IS_HOLE(obp)) | |
4817 | dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp); | |
4818 | arc_release(dr->dt.dl.dr_data, db); | |
4819 | } | |
34dc7c2f BB |
4820 | mutex_exit(&db->db_mtx); |
4821 | ||
428870ff | 4822 | dbuf_write_done(zio, NULL, db); |
a6255b7f DQ |
4823 | |
4824 | if (zio->io_abd != NULL) | |
e2af2acc | 4825 | abd_free(zio->io_abd); |
428870ff BB |
4826 | } |
4827 | ||
a1d477c2 MA |
4828 | typedef struct dbuf_remap_impl_callback_arg { |
4829 | objset_t *drica_os; | |
4830 | uint64_t drica_blk_birth; | |
4831 | dmu_tx_t *drica_tx; | |
4832 | } dbuf_remap_impl_callback_arg_t; | |
4833 | ||
4834 | static void | |
4835 | dbuf_remap_impl_callback(uint64_t vdev, uint64_t offset, uint64_t size, | |
4836 | void *arg) | |
4837 | { | |
4838 | dbuf_remap_impl_callback_arg_t *drica = arg; | |
4839 | objset_t *os = drica->drica_os; | |
4840 | spa_t *spa = dmu_objset_spa(os); | |
4841 | dmu_tx_t *tx = drica->drica_tx; | |
4842 | ||
4843 | ASSERT(dsl_pool_sync_context(spa_get_dsl(spa))); | |
4844 | ||
4845 | if (os == spa_meta_objset(spa)) { | |
4846 | spa_vdev_indirect_mark_obsolete(spa, vdev, offset, size, tx); | |
4847 | } else { | |
4848 | dsl_dataset_block_remapped(dmu_objset_ds(os), vdev, offset, | |
4849 | size, drica->drica_blk_birth, tx); | |
4850 | } | |
4851 | } | |
4852 | ||
4853 | static void | |
f664f1ee | 4854 | dbuf_remap_impl(dnode_t *dn, blkptr_t *bp, krwlock_t *rw, dmu_tx_t *tx) |
a1d477c2 MA |
4855 | { |
4856 | blkptr_t bp_copy = *bp; | |
4857 | spa_t *spa = dmu_objset_spa(dn->dn_objset); | |
4858 | dbuf_remap_impl_callback_arg_t drica; | |
4859 | ||
4860 | ASSERT(dsl_pool_sync_context(spa_get_dsl(spa))); | |
4861 | ||
4862 | drica.drica_os = dn->dn_objset; | |
4863 | drica.drica_blk_birth = bp->blk_birth; | |
4864 | drica.drica_tx = tx; | |
4865 | if (spa_remap_blkptr(spa, &bp_copy, dbuf_remap_impl_callback, | |
4866 | &drica)) { | |
37f03da8 SH |
4867 | /* |
4868 | * If the blkptr being remapped is tracked by a livelist, | |
4869 | * then we need to make sure the livelist reflects the update. | |
4870 | * First, cancel out the old blkptr by appending a 'FREE' | |
4871 | * entry. Next, add an 'ALLOC' to track the new version. This | |
4872 | * way we avoid trying to free an inaccurate blkptr at delete. | |
4873 | * Note that embedded blkptrs are not tracked in livelists. | |
4874 | */ | |
4875 | if (dn->dn_objset != spa_meta_objset(spa)) { | |
4876 | dsl_dataset_t *ds = dmu_objset_ds(dn->dn_objset); | |
4877 | if (dsl_deadlist_is_open(&ds->ds_dir->dd_livelist) && | |
4878 | bp->blk_birth > ds->ds_dir->dd_origin_txg) { | |
4879 | ASSERT(!BP_IS_EMBEDDED(bp)); | |
4880 | ASSERT(dsl_dir_is_clone(ds->ds_dir)); | |
4881 | ASSERT(spa_feature_is_enabled(spa, | |
4882 | SPA_FEATURE_LIVELIST)); | |
4883 | bplist_append(&ds->ds_dir->dd_pending_frees, | |
4884 | bp); | |
4885 | bplist_append(&ds->ds_dir->dd_pending_allocs, | |
4886 | &bp_copy); | |
4887 | } | |
4888 | } | |
4889 | ||
a1d477c2 | 4890 | /* |
f664f1ee | 4891 | * The db_rwlock prevents dbuf_read_impl() from |
a1d477c2 MA |
4892 | * dereferencing the BP while we are changing it. To |
4893 | * avoid lock contention, only grab it when we are actually | |
4894 | * changing the BP. | |
4895 | */ | |
f664f1ee PD |
4896 | if (rw != NULL) |
4897 | rw_enter(rw, RW_WRITER); | |
a1d477c2 | 4898 | *bp = bp_copy; |
f664f1ee PD |
4899 | if (rw != NULL) |
4900 | rw_exit(rw); | |
a1d477c2 MA |
4901 | } |
4902 | } | |
4903 | ||
a1d477c2 MA |
4904 | /* |
4905 | * Remap any existing BP's to concrete vdevs, if possible. | |
4906 | */ | |
4907 | static void | |
4908 | dbuf_remap(dnode_t *dn, dmu_buf_impl_t *db, dmu_tx_t *tx) | |
4909 | { | |
4910 | spa_t *spa = dmu_objset_spa(db->db_objset); | |
4911 | ASSERT(dsl_pool_sync_context(spa_get_dsl(spa))); | |
4912 | ||
4913 | if (!spa_feature_is_active(spa, SPA_FEATURE_DEVICE_REMOVAL)) | |
4914 | return; | |
4915 | ||
4916 | if (db->db_level > 0) { | |
4917 | blkptr_t *bp = db->db.db_data; | |
4918 | for (int i = 0; i < db->db.db_size >> SPA_BLKPTRSHIFT; i++) { | |
f664f1ee | 4919 | dbuf_remap_impl(dn, &bp[i], &db->db_rwlock, tx); |
a1d477c2 MA |
4920 | } |
4921 | } else if (db->db.db_object == DMU_META_DNODE_OBJECT) { | |
4922 | dnode_phys_t *dnp = db->db.db_data; | |
4923 | ASSERT3U(db->db_dnode_handle->dnh_dnode->dn_type, ==, | |
4924 | DMU_OT_DNODE); | |
4925 | for (int i = 0; i < db->db.db_size >> DNODE_SHIFT; | |
4926 | i += dnp[i].dn_extra_slots + 1) { | |
4927 | for (int j = 0; j < dnp[i].dn_nblkptr; j++) { | |
f664f1ee PD |
4928 | krwlock_t *lock = (dn->dn_dbuf == NULL ? NULL : |
4929 | &dn->dn_dbuf->db_rwlock); | |
4930 | dbuf_remap_impl(dn, &dnp[i].dn_blkptr[j], lock, | |
4931 | tx); | |
a1d477c2 MA |
4932 | } |
4933 | } | |
4934 | } | |
4935 | } | |
4936 | ||
4937 | ||
e49f1e20 | 4938 | /* Issue I/O to commit a dirty buffer to disk. */ |
428870ff BB |
4939 | static void |
4940 | dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx) | |
4941 | { | |
4942 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
ba67d821 | 4943 | dnode_t *dn = dr->dr_dnode; |
572e2857 | 4944 | objset_t *os; |
428870ff BB |
4945 | dmu_buf_impl_t *parent = db->db_parent; |
4946 | uint64_t txg = tx->tx_txg; | |
5dbd68a3 | 4947 | zbookmark_phys_t zb; |
428870ff | 4948 | zio_prop_t zp; |
28caa74b | 4949 | zio_t *pio; /* parent I/O */ |
428870ff | 4950 | int wp_flag = 0; |
34dc7c2f | 4951 | |
463a8cfe AR |
4952 | ASSERT(dmu_tx_is_syncing(tx)); |
4953 | ||
572e2857 BB |
4954 | os = dn->dn_objset; |
4955 | ||
428870ff BB |
4956 | if (db->db_state != DB_NOFILL) { |
4957 | if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) { | |
4958 | /* | |
4959 | * Private object buffers are released here rather | |
4960 | * than in dbuf_dirty() since they are only modified | |
4961 | * in the syncing context and we don't want the | |
4962 | * overhead of making multiple copies of the data. | |
4963 | */ | |
4964 | if (BP_IS_HOLE(db->db_blkptr)) { | |
4965 | arc_buf_thaw(data); | |
4966 | } else { | |
4967 | dbuf_release_bp(db); | |
4968 | } | |
a1d477c2 | 4969 | dbuf_remap(dn, db, tx); |
428870ff BB |
4970 | } |
4971 | } | |
4972 | ||
4973 | if (parent != dn->dn_dbuf) { | |
e49f1e20 WA |
4974 | /* Our parent is an indirect block. */ |
4975 | /* We have a dirty parent that has been scheduled for write. */ | |
428870ff | 4976 | ASSERT(parent && parent->db_data_pending); |
e49f1e20 | 4977 | /* Our parent's buffer is one level closer to the dnode. */ |
428870ff | 4978 | ASSERT(db->db_level == parent->db_level-1); |
e49f1e20 WA |
4979 | /* |
4980 | * We're about to modify our parent's db_data by modifying | |
4981 | * our block pointer, so the parent must be released. | |
4982 | */ | |
428870ff | 4983 | ASSERT(arc_released(parent->db_buf)); |
28caa74b | 4984 | pio = parent->db_data_pending->dr_zio; |
428870ff | 4985 | } else { |
e49f1e20 | 4986 | /* Our parent is the dnode itself. */ |
428870ff BB |
4987 | ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 && |
4988 | db->db_blkid != DMU_SPILL_BLKID) || | |
4989 | (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0)); | |
4990 | if (db->db_blkid != DMU_SPILL_BLKID) | |
4991 | ASSERT3P(db->db_blkptr, ==, | |
4992 | &dn->dn_phys->dn_blkptr[db->db_blkid]); | |
28caa74b | 4993 | pio = dn->dn_zio; |
428870ff BB |
4994 | } |
4995 | ||
4996 | ASSERT(db->db_level == 0 || data == db->db_buf); | |
4997 | ASSERT3U(db->db_blkptr->blk_birth, <=, txg); | |
28caa74b | 4998 | ASSERT(pio); |
428870ff BB |
4999 | |
5000 | SET_BOOKMARK(&zb, os->os_dsl_dataset ? | |
5001 | os->os_dsl_dataset->ds_object : DMU_META_OBJSET, | |
5002 | db->db.db_object, db->db_level, db->db_blkid); | |
5003 | ||
5004 | if (db->db_blkid == DMU_SPILL_BLKID) | |
5005 | wp_flag = WP_SPILL; | |
5006 | wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0; | |
5007 | ||
82644107 | 5008 | dmu_write_policy(os, dn, db->db_level, wp_flag, &zp); |
428870ff | 5009 | |
463a8cfe AR |
5010 | /* |
5011 | * We copy the blkptr now (rather than when we instantiate the dirty | |
5012 | * record), because its value can change between open context and | |
5013 | * syncing context. We do not need to hold dn_struct_rwlock to read | |
5014 | * db_blkptr because we are in syncing context. | |
5015 | */ | |
5016 | dr->dr_bp_copy = *db->db_blkptr; | |
5017 | ||
9b67f605 MA |
5018 | if (db->db_level == 0 && |
5019 | dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { | |
5020 | /* | |
5021 | * The BP for this block has been provided by open context | |
5022 | * (by dmu_sync() or dmu_buf_write_embedded()). | |
5023 | */ | |
a6255b7f DQ |
5024 | abd_t *contents = (data != NULL) ? |
5025 | abd_get_from_buf(data->b_data, arc_buf_size(data)) : NULL; | |
9b67f605 | 5026 | |
28caa74b MM |
5027 | dr->dr_zio = zio_write(pio, os->os_spa, txg, &dr->dr_bp_copy, |
5028 | contents, db->db.db_size, db->db.db_size, &zp, | |
5029 | dbuf_write_override_ready, NULL, NULL, | |
bc77ba73 | 5030 | dbuf_write_override_done, |
e8b96c60 | 5031 | dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); |
428870ff BB |
5032 | mutex_enter(&db->db_mtx); |
5033 | dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; | |
5034 | zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by, | |
03c6040b | 5035 | dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite); |
428870ff BB |
5036 | mutex_exit(&db->db_mtx); |
5037 | } else if (db->db_state == DB_NOFILL) { | |
3c67d83a TH |
5038 | ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF || |
5039 | zp.zp_checksum == ZIO_CHECKSUM_NOPARITY); | |
28caa74b | 5040 | dr->dr_zio = zio_write(pio, os->os_spa, txg, |
2aa34383 | 5041 | &dr->dr_bp_copy, NULL, db->db.db_size, db->db.db_size, &zp, |
bc77ba73 PD |
5042 | dbuf_write_nofill_ready, NULL, NULL, |
5043 | dbuf_write_nofill_done, db, | |
428870ff BB |
5044 | ZIO_PRIORITY_ASYNC_WRITE, |
5045 | ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb); | |
5046 | } else { | |
5047 | ASSERT(arc_released(data)); | |
bc77ba73 PD |
5048 | |
5049 | /* | |
5050 | * For indirect blocks, we want to setup the children | |
5051 | * ready callback so that we can properly handle an indirect | |
5052 | * block that only contains holes. | |
5053 | */ | |
1c27024e | 5054 | arc_write_done_func_t *children_ready_cb = NULL; |
bc77ba73 PD |
5055 | if (db->db_level != 0) |
5056 | children_ready_cb = dbuf_write_children_ready; | |
5057 | ||
28caa74b | 5058 | dr->dr_zio = arc_write(pio, os->os_spa, txg, |
c9d62d13 | 5059 | &dr->dr_bp_copy, data, dbuf_is_l2cacheable(db), |
d3c2ae1c GW |
5060 | &zp, dbuf_write_ready, |
5061 | children_ready_cb, dbuf_write_physdone, | |
5062 | dbuf_write_done, db, ZIO_PRIORITY_ASYNC_WRITE, | |
5063 | ZIO_FLAG_MUSTSUCCEED, &zb); | |
428870ff | 5064 | } |
34dc7c2f | 5065 | } |
c28b2279 | 5066 | |
8f576c23 BB |
5067 | EXPORT_SYMBOL(dbuf_find); |
5068 | EXPORT_SYMBOL(dbuf_is_metadata); | |
d3c2ae1c | 5069 | EXPORT_SYMBOL(dbuf_destroy); |
8f576c23 BB |
5070 | EXPORT_SYMBOL(dbuf_loan_arcbuf); |
5071 | EXPORT_SYMBOL(dbuf_whichblock); | |
5072 | EXPORT_SYMBOL(dbuf_read); | |
5073 | EXPORT_SYMBOL(dbuf_unoverride); | |
5074 | EXPORT_SYMBOL(dbuf_free_range); | |
5075 | EXPORT_SYMBOL(dbuf_new_size); | |
5076 | EXPORT_SYMBOL(dbuf_release_bp); | |
5077 | EXPORT_SYMBOL(dbuf_dirty); | |
0c03d21a | 5078 | EXPORT_SYMBOL(dmu_buf_set_crypt_params); |
c28b2279 | 5079 | EXPORT_SYMBOL(dmu_buf_will_dirty); |
a73e8fdb | 5080 | EXPORT_SYMBOL(dmu_buf_is_dirty); |
8f576c23 BB |
5081 | EXPORT_SYMBOL(dmu_buf_will_not_fill); |
5082 | EXPORT_SYMBOL(dmu_buf_will_fill); | |
5083 | EXPORT_SYMBOL(dmu_buf_fill_done); | |
4047414a | 5084 | EXPORT_SYMBOL(dmu_buf_rele); |
8f576c23 | 5085 | EXPORT_SYMBOL(dbuf_assign_arcbuf); |
8f576c23 BB |
5086 | EXPORT_SYMBOL(dbuf_prefetch); |
5087 | EXPORT_SYMBOL(dbuf_hold_impl); | |
5088 | EXPORT_SYMBOL(dbuf_hold); | |
5089 | EXPORT_SYMBOL(dbuf_hold_level); | |
5090 | EXPORT_SYMBOL(dbuf_create_bonus); | |
5091 | EXPORT_SYMBOL(dbuf_spill_set_blksz); | |
5092 | EXPORT_SYMBOL(dbuf_rm_spill); | |
5093 | EXPORT_SYMBOL(dbuf_add_ref); | |
5094 | EXPORT_SYMBOL(dbuf_rele); | |
5095 | EXPORT_SYMBOL(dbuf_rele_and_unlock); | |
5096 | EXPORT_SYMBOL(dbuf_refcount); | |
5097 | EXPORT_SYMBOL(dbuf_sync_list); | |
5098 | EXPORT_SYMBOL(dmu_buf_set_user); | |
5099 | EXPORT_SYMBOL(dmu_buf_set_user_ie); | |
8f576c23 | 5100 | EXPORT_SYMBOL(dmu_buf_get_user); |
0f699108 | 5101 | EXPORT_SYMBOL(dmu_buf_get_blkptr); |
d3c2ae1c | 5102 | |
03fdcb9a | 5103 | ZFS_MODULE_PARAM(zfs_dbuf_cache, dbuf_cache_, max_bytes, ULONG, ZMOD_RW, |
02730c33 | 5104 | "Maximum size in bytes of the dbuf cache."); |
d3c2ae1c | 5105 | |
03fdcb9a | 5106 | ZFS_MODULE_PARAM(zfs_dbuf_cache, dbuf_cache_, hiwater_pct, UINT, ZMOD_RW, |
7ada752a | 5107 | "Percentage over dbuf_cache_max_bytes for direct dbuf eviction."); |
d3c2ae1c | 5108 | |
03fdcb9a | 5109 | ZFS_MODULE_PARAM(zfs_dbuf_cache, dbuf_cache_, lowater_pct, UINT, ZMOD_RW, |
7ada752a | 5110 | "Percentage below dbuf_cache_max_bytes when dbuf eviction stops."); |
d3c2ae1c | 5111 | |
03fdcb9a | 5112 | ZFS_MODULE_PARAM(zfs_dbuf, dbuf_, metadata_cache_max_bytes, ULONG, ZMOD_RW, |
7ada752a | 5113 | "Maximum size in bytes of dbuf metadata cache."); |
2e5dc449 | 5114 | |
03fdcb9a | 5115 | ZFS_MODULE_PARAM(zfs_dbuf, dbuf_, cache_shift, INT, ZMOD_RW, |
7ada752a | 5116 | "Set size of dbuf cache to log2 fraction of arc size."); |
2e5dc449 | 5117 | |
03fdcb9a | 5118 | ZFS_MODULE_PARAM(zfs_dbuf, dbuf_, metadata_cache_shift, INT, ZMOD_RW, |
7ada752a | 5119 | "Set size of dbuf metadata cache to log2 fraction of arc size."); |