]>
Commit | Line | Data |
---|---|---|
34dc7c2f BB |
1 | /* |
2 | * CDDL HEADER START | |
3 | * | |
4 | * The contents of this file are subject to the terms of the | |
5 | * Common Development and Distribution License (the "License"). | |
6 | * You may not use this file except in compliance with the License. | |
7 | * | |
8 | * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE | |
9 | * or http://www.opensolaris.org/os/licensing. | |
10 | * See the License for the specific language governing permissions | |
11 | * and limitations under the License. | |
12 | * | |
13 | * When distributing Covered Code, include this CDDL HEADER in each | |
14 | * file and include the License file at usr/src/OPENSOLARIS.LICENSE. | |
15 | * If applicable, add the following below this CDDL HEADER, with the | |
16 | * fields enclosed by brackets "[]" replaced with your own identifying | |
17 | * information: Portions Copyright [yyyy] [name of copyright owner] | |
18 | * | |
19 | * CDDL HEADER END | |
20 | */ | |
21 | /* | |
428870ff | 22 | * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. |
ef3c1dea | 23 | * Copyright 2011 Nexenta Systems, Inc. All rights reserved. |
4bda3bd0 | 24 | * Copyright (c) 2012, 2015 by Delphix. All rights reserved. |
3a17a7a9 | 25 | * Copyright (c) 2013 by Saso Kiselkov. All rights reserved. |
0c66c32d | 26 | * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved. |
34dc7c2f BB |
27 | */ |
28 | ||
34dc7c2f | 29 | #include <sys/zfs_context.h> |
c28b2279 | 30 | #include <sys/arc.h> |
34dc7c2f | 31 | #include <sys/dmu.h> |
ea97f8ce | 32 | #include <sys/dmu_send.h> |
34dc7c2f BB |
33 | #include <sys/dmu_impl.h> |
34 | #include <sys/dbuf.h> | |
35 | #include <sys/dmu_objset.h> | |
36 | #include <sys/dsl_dataset.h> | |
37 | #include <sys/dsl_dir.h> | |
38 | #include <sys/dmu_tx.h> | |
39 | #include <sys/spa.h> | |
40 | #include <sys/zio.h> | |
41 | #include <sys/dmu_zfetch.h> | |
428870ff BB |
42 | #include <sys/sa.h> |
43 | #include <sys/sa_impl.h> | |
9b67f605 MA |
44 | #include <sys/zfeature.h> |
45 | #include <sys/blkptr.h> | |
9bd274dd | 46 | #include <sys/range_tree.h> |
49ee64e5 | 47 | #include <sys/trace_dbuf.h> |
34dc7c2f | 48 | |
fc5bb51f BB |
49 | struct dbuf_hold_impl_data { |
50 | /* Function arguments */ | |
51 | dnode_t *dh_dn; | |
52 | uint8_t dh_level; | |
53 | uint64_t dh_blkid; | |
54 | int dh_fail_sparse; | |
55 | void *dh_tag; | |
56 | dmu_buf_impl_t **dh_dbp; | |
57 | /* Local variables */ | |
58 | dmu_buf_impl_t *dh_db; | |
59 | dmu_buf_impl_t *dh_parent; | |
60 | blkptr_t *dh_bp; | |
61 | int dh_err; | |
62 | dbuf_dirty_record_t *dh_dr; | |
63 | arc_buf_contents_t dh_type; | |
64 | int dh_depth; | |
65 | }; | |
66 | ||
67 | static void __dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh, | |
68 | dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse, | |
69 | void *tag, dmu_buf_impl_t **dbp, int depth); | |
70 | static int __dbuf_hold_impl(struct dbuf_hold_impl_data *dh); | |
71 | ||
b663a23d MA |
72 | /* |
73 | * Number of times that zfs_free_range() took the slow path while doing | |
74 | * a zfs receive. A nonzero value indicates a potential performance problem. | |
75 | */ | |
76 | uint64_t zfs_free_range_recv_miss; | |
77 | ||
34dc7c2f | 78 | static void dbuf_destroy(dmu_buf_impl_t *db); |
13fe0198 | 79 | static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx); |
b128c09f | 80 | static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx); |
34dc7c2f | 81 | |
0c66c32d JG |
82 | #ifndef __lint |
83 | extern inline void dmu_buf_init_user(dmu_buf_user_t *dbu, | |
84 | dmu_buf_evict_func_t *evict_func, dmu_buf_t **clear_on_evict_dbufp); | |
85 | #endif /* ! __lint */ | |
86 | ||
34dc7c2f BB |
87 | /* |
88 | * Global data structures and functions for the dbuf cache. | |
89 | */ | |
90 | static kmem_cache_t *dbuf_cache; | |
0c66c32d | 91 | static taskq_t *dbu_evict_taskq; |
34dc7c2f BB |
92 | |
93 | /* ARGSUSED */ | |
94 | static int | |
95 | dbuf_cons(void *vdb, void *unused, int kmflag) | |
96 | { | |
97 | dmu_buf_impl_t *db = vdb; | |
98 | bzero(db, sizeof (dmu_buf_impl_t)); | |
99 | ||
100 | mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL); | |
101 | cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL); | |
102 | refcount_create(&db->db_holds); | |
8951cb8d | 103 | |
34dc7c2f BB |
104 | return (0); |
105 | } | |
106 | ||
107 | /* ARGSUSED */ | |
108 | static void | |
109 | dbuf_dest(void *vdb, void *unused) | |
110 | { | |
111 | dmu_buf_impl_t *db = vdb; | |
112 | mutex_destroy(&db->db_mtx); | |
113 | cv_destroy(&db->db_changed); | |
114 | refcount_destroy(&db->db_holds); | |
115 | } | |
116 | ||
117 | /* | |
118 | * dbuf hash table routines | |
119 | */ | |
120 | static dbuf_hash_table_t dbuf_hash_table; | |
121 | ||
122 | static uint64_t dbuf_hash_count; | |
123 | ||
124 | static uint64_t | |
125 | dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid) | |
126 | { | |
127 | uintptr_t osv = (uintptr_t)os; | |
128 | uint64_t crc = -1ULL; | |
129 | ||
130 | ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); | |
131 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF]; | |
132 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF]; | |
133 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF]; | |
134 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF]; | |
135 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF]; | |
136 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF]; | |
137 | ||
138 | crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16); | |
139 | ||
140 | return (crc); | |
141 | } | |
142 | ||
143 | #define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid); | |
144 | ||
145 | #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \ | |
146 | ((dbuf)->db.db_object == (obj) && \ | |
147 | (dbuf)->db_objset == (os) && \ | |
148 | (dbuf)->db_level == (level) && \ | |
149 | (dbuf)->db_blkid == (blkid)) | |
150 | ||
151 | dmu_buf_impl_t * | |
6ebebace | 152 | dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid) |
34dc7c2f BB |
153 | { |
154 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
d6320ddb BB |
155 | uint64_t hv; |
156 | uint64_t idx; | |
34dc7c2f BB |
157 | dmu_buf_impl_t *db; |
158 | ||
d6320ddb BB |
159 | hv = DBUF_HASH(os, obj, level, blkid); |
160 | idx = hv & h->hash_table_mask; | |
161 | ||
34dc7c2f BB |
162 | mutex_enter(DBUF_HASH_MUTEX(h, idx)); |
163 | for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) { | |
164 | if (DBUF_EQUAL(db, os, obj, level, blkid)) { | |
165 | mutex_enter(&db->db_mtx); | |
166 | if (db->db_state != DB_EVICTING) { | |
167 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); | |
168 | return (db); | |
169 | } | |
170 | mutex_exit(&db->db_mtx); | |
171 | } | |
172 | } | |
173 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); | |
174 | return (NULL); | |
175 | } | |
176 | ||
6ebebace JG |
177 | static dmu_buf_impl_t * |
178 | dbuf_find_bonus(objset_t *os, uint64_t object) | |
179 | { | |
180 | dnode_t *dn; | |
181 | dmu_buf_impl_t *db = NULL; | |
182 | ||
183 | if (dnode_hold(os, object, FTAG, &dn) == 0) { | |
184 | rw_enter(&dn->dn_struct_rwlock, RW_READER); | |
185 | if (dn->dn_bonus != NULL) { | |
186 | db = dn->dn_bonus; | |
187 | mutex_enter(&db->db_mtx); | |
188 | } | |
189 | rw_exit(&dn->dn_struct_rwlock); | |
190 | dnode_rele(dn, FTAG); | |
191 | } | |
192 | return (db); | |
193 | } | |
194 | ||
34dc7c2f BB |
195 | /* |
196 | * Insert an entry into the hash table. If there is already an element | |
197 | * equal to elem in the hash table, then the already existing element | |
198 | * will be returned and the new element will not be inserted. | |
199 | * Otherwise returns NULL. | |
200 | */ | |
201 | static dmu_buf_impl_t * | |
202 | dbuf_hash_insert(dmu_buf_impl_t *db) | |
203 | { | |
204 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
428870ff | 205 | objset_t *os = db->db_objset; |
34dc7c2f BB |
206 | uint64_t obj = db->db.db_object; |
207 | int level = db->db_level; | |
d6320ddb | 208 | uint64_t blkid, hv, idx; |
34dc7c2f BB |
209 | dmu_buf_impl_t *dbf; |
210 | ||
d6320ddb BB |
211 | blkid = db->db_blkid; |
212 | hv = DBUF_HASH(os, obj, level, blkid); | |
213 | idx = hv & h->hash_table_mask; | |
214 | ||
34dc7c2f BB |
215 | mutex_enter(DBUF_HASH_MUTEX(h, idx)); |
216 | for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) { | |
217 | if (DBUF_EQUAL(dbf, os, obj, level, blkid)) { | |
218 | mutex_enter(&dbf->db_mtx); | |
219 | if (dbf->db_state != DB_EVICTING) { | |
220 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); | |
221 | return (dbf); | |
222 | } | |
223 | mutex_exit(&dbf->db_mtx); | |
224 | } | |
225 | } | |
226 | ||
227 | mutex_enter(&db->db_mtx); | |
228 | db->db_hash_next = h->hash_table[idx]; | |
229 | h->hash_table[idx] = db; | |
230 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); | |
231 | atomic_add_64(&dbuf_hash_count, 1); | |
232 | ||
233 | return (NULL); | |
234 | } | |
235 | ||
236 | /* | |
bd089c54 | 237 | * Remove an entry from the hash table. It must be in the EVICTING state. |
34dc7c2f BB |
238 | */ |
239 | static void | |
240 | dbuf_hash_remove(dmu_buf_impl_t *db) | |
241 | { | |
242 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
d6320ddb | 243 | uint64_t hv, idx; |
34dc7c2f BB |
244 | dmu_buf_impl_t *dbf, **dbp; |
245 | ||
d6320ddb BB |
246 | hv = DBUF_HASH(db->db_objset, db->db.db_object, |
247 | db->db_level, db->db_blkid); | |
248 | idx = hv & h->hash_table_mask; | |
249 | ||
34dc7c2f | 250 | /* |
bd089c54 | 251 | * We musn't hold db_mtx to maintain lock ordering: |
34dc7c2f BB |
252 | * DBUF_HASH_MUTEX > db_mtx. |
253 | */ | |
254 | ASSERT(refcount_is_zero(&db->db_holds)); | |
255 | ASSERT(db->db_state == DB_EVICTING); | |
256 | ASSERT(!MUTEX_HELD(&db->db_mtx)); | |
257 | ||
258 | mutex_enter(DBUF_HASH_MUTEX(h, idx)); | |
259 | dbp = &h->hash_table[idx]; | |
260 | while ((dbf = *dbp) != db) { | |
261 | dbp = &dbf->db_hash_next; | |
262 | ASSERT(dbf != NULL); | |
263 | } | |
264 | *dbp = db->db_hash_next; | |
265 | db->db_hash_next = NULL; | |
266 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); | |
267 | atomic_add_64(&dbuf_hash_count, -1); | |
268 | } | |
269 | ||
270 | static arc_evict_func_t dbuf_do_evict; | |
271 | ||
0c66c32d JG |
272 | typedef enum { |
273 | DBVU_EVICTING, | |
274 | DBVU_NOT_EVICTING | |
275 | } dbvu_verify_type_t; | |
276 | ||
277 | static void | |
278 | dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type) | |
279 | { | |
280 | #ifdef ZFS_DEBUG | |
281 | int64_t holds; | |
282 | ||
283 | if (db->db_user == NULL) | |
284 | return; | |
285 | ||
286 | /* Only data blocks support the attachment of user data. */ | |
287 | ASSERT(db->db_level == 0); | |
288 | ||
289 | /* Clients must resolve a dbuf before attaching user data. */ | |
290 | ASSERT(db->db.db_data != NULL); | |
291 | ASSERT3U(db->db_state, ==, DB_CACHED); | |
292 | ||
293 | holds = refcount_count(&db->db_holds); | |
294 | if (verify_type == DBVU_EVICTING) { | |
295 | /* | |
296 | * Immediate eviction occurs when holds == dirtycnt. | |
297 | * For normal eviction buffers, holds is zero on | |
298 | * eviction, except when dbuf_fix_old_data() calls | |
299 | * dbuf_clear_data(). However, the hold count can grow | |
300 | * during eviction even though db_mtx is held (see | |
301 | * dmu_bonus_hold() for an example), so we can only | |
302 | * test the generic invariant that holds >= dirtycnt. | |
303 | */ | |
304 | ASSERT3U(holds, >=, db->db_dirtycnt); | |
305 | } else { | |
bc4501f7 | 306 | if (db->db_user_immediate_evict == TRUE) |
0c66c32d JG |
307 | ASSERT3U(holds, >=, db->db_dirtycnt); |
308 | else | |
309 | ASSERT3U(holds, >, 0); | |
310 | } | |
311 | #endif | |
312 | } | |
313 | ||
34dc7c2f BB |
314 | static void |
315 | dbuf_evict_user(dmu_buf_impl_t *db) | |
316 | { | |
0c66c32d JG |
317 | dmu_buf_user_t *dbu = db->db_user; |
318 | ||
34dc7c2f BB |
319 | ASSERT(MUTEX_HELD(&db->db_mtx)); |
320 | ||
0c66c32d | 321 | if (dbu == NULL) |
34dc7c2f BB |
322 | return; |
323 | ||
0c66c32d JG |
324 | dbuf_verify_user(db, DBVU_EVICTING); |
325 | db->db_user = NULL; | |
326 | ||
327 | #ifdef ZFS_DEBUG | |
328 | if (dbu->dbu_clear_on_evict_dbufp != NULL) | |
329 | *dbu->dbu_clear_on_evict_dbufp = NULL; | |
330 | #endif | |
331 | ||
332 | /* | |
333 | * Invoke the callback from a taskq to avoid lock order reversals | |
334 | * and limit stack depth. | |
335 | */ | |
336 | taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func, dbu, 0, | |
337 | &dbu->dbu_tqent); | |
34dc7c2f BB |
338 | } |
339 | ||
572e2857 BB |
340 | boolean_t |
341 | dbuf_is_metadata(dmu_buf_impl_t *db) | |
342 | { | |
cc79a5c2 BB |
343 | /* |
344 | * Consider indirect blocks and spill blocks to be meta data. | |
345 | */ | |
346 | if (db->db_level > 0 || db->db_blkid == DMU_SPILL_BLKID) { | |
572e2857 BB |
347 | return (B_TRUE); |
348 | } else { | |
349 | boolean_t is_metadata; | |
350 | ||
351 | DB_DNODE_ENTER(db); | |
9ae529ec | 352 | is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type); |
572e2857 BB |
353 | DB_DNODE_EXIT(db); |
354 | ||
355 | return (is_metadata); | |
356 | } | |
357 | } | |
358 | ||
34dc7c2f BB |
359 | void |
360 | dbuf_evict(dmu_buf_impl_t *db) | |
361 | { | |
362 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
363 | ASSERT(db->db_buf == NULL); | |
364 | ASSERT(db->db_data_pending == NULL); | |
365 | ||
366 | dbuf_clear(db); | |
367 | dbuf_destroy(db); | |
368 | } | |
369 | ||
370 | void | |
371 | dbuf_init(void) | |
372 | { | |
373 | uint64_t hsize = 1ULL << 16; | |
374 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
375 | int i; | |
376 | ||
377 | /* | |
378 | * The hash table is big enough to fill all of physical memory | |
69de3421 TC |
379 | * with an average block size of zfs_arc_average_blocksize (default 8K). |
380 | * By default, the table will take up | |
381 | * totalmem * sizeof(void*) / 8K (1MB per GB with 8-byte pointers). | |
34dc7c2f | 382 | */ |
69de3421 | 383 | while (hsize * zfs_arc_average_blocksize < physmem * PAGESIZE) |
34dc7c2f BB |
384 | hsize <<= 1; |
385 | ||
386 | retry: | |
387 | h->hash_table_mask = hsize - 1; | |
00b46022 | 388 | #if defined(_KERNEL) && defined(HAVE_SPL) |
d1d7e268 MK |
389 | /* |
390 | * Large allocations which do not require contiguous pages | |
391 | * should be using vmem_alloc() in the linux kernel | |
392 | */ | |
79c76d5b | 393 | h->hash_table = vmem_zalloc(hsize * sizeof (void *), KM_SLEEP); |
00b46022 | 394 | #else |
34dc7c2f | 395 | h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP); |
00b46022 | 396 | #endif |
34dc7c2f BB |
397 | if (h->hash_table == NULL) { |
398 | /* XXX - we should really return an error instead of assert */ | |
399 | ASSERT(hsize > (1ULL << 10)); | |
400 | hsize >>= 1; | |
401 | goto retry; | |
402 | } | |
403 | ||
404 | dbuf_cache = kmem_cache_create("dmu_buf_impl_t", | |
405 | sizeof (dmu_buf_impl_t), | |
406 | 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0); | |
407 | ||
408 | for (i = 0; i < DBUF_MUTEXES; i++) | |
40d06e3c | 409 | mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL); |
e0b0ca98 BB |
410 | |
411 | dbuf_stats_init(h); | |
0c66c32d JG |
412 | |
413 | /* | |
414 | * All entries are queued via taskq_dispatch_ent(), so min/maxalloc | |
415 | * configuration is not required. | |
416 | */ | |
1229323d | 417 | dbu_evict_taskq = taskq_create("dbu_evict", 1, defclsyspri, 0, 0, 0); |
34dc7c2f BB |
418 | } |
419 | ||
420 | void | |
421 | dbuf_fini(void) | |
422 | { | |
423 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
424 | int i; | |
425 | ||
e0b0ca98 BB |
426 | dbuf_stats_destroy(); |
427 | ||
34dc7c2f BB |
428 | for (i = 0; i < DBUF_MUTEXES; i++) |
429 | mutex_destroy(&h->hash_mutexes[i]); | |
00b46022 | 430 | #if defined(_KERNEL) && defined(HAVE_SPL) |
d1d7e268 MK |
431 | /* |
432 | * Large allocations which do not require contiguous pages | |
433 | * should be using vmem_free() in the linux kernel | |
434 | */ | |
00b46022 BB |
435 | vmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *)); |
436 | #else | |
34dc7c2f | 437 | kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *)); |
00b46022 | 438 | #endif |
34dc7c2f | 439 | kmem_cache_destroy(dbuf_cache); |
0c66c32d | 440 | taskq_destroy(dbu_evict_taskq); |
34dc7c2f BB |
441 | } |
442 | ||
443 | /* | |
444 | * Other stuff. | |
445 | */ | |
446 | ||
447 | #ifdef ZFS_DEBUG | |
448 | static void | |
449 | dbuf_verify(dmu_buf_impl_t *db) | |
450 | { | |
572e2857 | 451 | dnode_t *dn; |
428870ff | 452 | dbuf_dirty_record_t *dr; |
34dc7c2f BB |
453 | |
454 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
455 | ||
456 | if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY)) | |
457 | return; | |
458 | ||
459 | ASSERT(db->db_objset != NULL); | |
572e2857 BB |
460 | DB_DNODE_ENTER(db); |
461 | dn = DB_DNODE(db); | |
34dc7c2f BB |
462 | if (dn == NULL) { |
463 | ASSERT(db->db_parent == NULL); | |
464 | ASSERT(db->db_blkptr == NULL); | |
465 | } else { | |
466 | ASSERT3U(db->db.db_object, ==, dn->dn_object); | |
467 | ASSERT3P(db->db_objset, ==, dn->dn_objset); | |
468 | ASSERT3U(db->db_level, <, dn->dn_nlevels); | |
572e2857 BB |
469 | ASSERT(db->db_blkid == DMU_BONUS_BLKID || |
470 | db->db_blkid == DMU_SPILL_BLKID || | |
8951cb8d | 471 | !avl_is_empty(&dn->dn_dbufs)); |
34dc7c2f | 472 | } |
428870ff BB |
473 | if (db->db_blkid == DMU_BONUS_BLKID) { |
474 | ASSERT(dn != NULL); | |
475 | ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); | |
476 | ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID); | |
477 | } else if (db->db_blkid == DMU_SPILL_BLKID) { | |
34dc7c2f BB |
478 | ASSERT(dn != NULL); |
479 | ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); | |
c99c9001 | 480 | ASSERT0(db->db.db_offset); |
34dc7c2f BB |
481 | } else { |
482 | ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size); | |
483 | } | |
484 | ||
428870ff BB |
485 | for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next) |
486 | ASSERT(dr->dr_dbuf == db); | |
487 | ||
488 | for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next) | |
489 | ASSERT(dr->dr_dbuf == db); | |
490 | ||
b128c09f BB |
491 | /* |
492 | * We can't assert that db_size matches dn_datablksz because it | |
493 | * can be momentarily different when another thread is doing | |
494 | * dnode_set_blksz(). | |
495 | */ | |
496 | if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) { | |
428870ff | 497 | dr = db->db_data_pending; |
b128c09f BB |
498 | /* |
499 | * It should only be modified in syncing context, so | |
500 | * make sure we only have one copy of the data. | |
501 | */ | |
502 | ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf); | |
34dc7c2f BB |
503 | } |
504 | ||
505 | /* verify db->db_blkptr */ | |
506 | if (db->db_blkptr) { | |
507 | if (db->db_parent == dn->dn_dbuf) { | |
508 | /* db is pointed to by the dnode */ | |
509 | /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */ | |
9babb374 | 510 | if (DMU_OBJECT_IS_SPECIAL(db->db.db_object)) |
34dc7c2f BB |
511 | ASSERT(db->db_parent == NULL); |
512 | else | |
513 | ASSERT(db->db_parent != NULL); | |
428870ff BB |
514 | if (db->db_blkid != DMU_SPILL_BLKID) |
515 | ASSERT3P(db->db_blkptr, ==, | |
516 | &dn->dn_phys->dn_blkptr[db->db_blkid]); | |
34dc7c2f BB |
517 | } else { |
518 | /* db is pointed to by an indirect block */ | |
1fde1e37 BB |
519 | ASSERTV(int epb = db->db_parent->db.db_size >> |
520 | SPA_BLKPTRSHIFT); | |
34dc7c2f BB |
521 | ASSERT3U(db->db_parent->db_level, ==, db->db_level+1); |
522 | ASSERT3U(db->db_parent->db.db_object, ==, | |
523 | db->db.db_object); | |
524 | /* | |
525 | * dnode_grow_indblksz() can make this fail if we don't | |
526 | * have the struct_rwlock. XXX indblksz no longer | |
527 | * grows. safe to do this now? | |
528 | */ | |
572e2857 | 529 | if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) { |
34dc7c2f BB |
530 | ASSERT3P(db->db_blkptr, ==, |
531 | ((blkptr_t *)db->db_parent->db.db_data + | |
532 | db->db_blkid % epb)); | |
533 | } | |
534 | } | |
535 | } | |
536 | if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) && | |
428870ff BB |
537 | (db->db_buf == NULL || db->db_buf->b_data) && |
538 | db->db.db_data && db->db_blkid != DMU_BONUS_BLKID && | |
34dc7c2f BB |
539 | db->db_state != DB_FILL && !dn->dn_free_txg) { |
540 | /* | |
541 | * If the blkptr isn't set but they have nonzero data, | |
542 | * it had better be dirty, otherwise we'll lose that | |
543 | * data when we evict this buffer. | |
544 | */ | |
545 | if (db->db_dirtycnt == 0) { | |
1fde1e37 | 546 | ASSERTV(uint64_t *buf = db->db.db_data); |
34dc7c2f BB |
547 | int i; |
548 | ||
549 | for (i = 0; i < db->db.db_size >> 3; i++) { | |
550 | ASSERT(buf[i] == 0); | |
551 | } | |
552 | } | |
553 | } | |
572e2857 | 554 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
555 | } |
556 | #endif | |
557 | ||
0c66c32d JG |
558 | static void |
559 | dbuf_clear_data(dmu_buf_impl_t *db) | |
560 | { | |
561 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
562 | dbuf_evict_user(db); | |
563 | db->db_buf = NULL; | |
564 | db->db.db_data = NULL; | |
565 | if (db->db_state != DB_NOFILL) | |
566 | db->db_state = DB_UNCACHED; | |
567 | } | |
568 | ||
34dc7c2f BB |
569 | static void |
570 | dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf) | |
571 | { | |
572 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
0c66c32d JG |
573 | ASSERT(buf != NULL); |
574 | ||
34dc7c2f | 575 | db->db_buf = buf; |
0c66c32d JG |
576 | ASSERT(buf->b_data != NULL); |
577 | db->db.db_data = buf->b_data; | |
578 | if (!arc_released(buf)) | |
579 | arc_set_callback(buf, dbuf_do_evict, db); | |
34dc7c2f BB |
580 | } |
581 | ||
428870ff BB |
582 | /* |
583 | * Loan out an arc_buf for read. Return the loaned arc_buf. | |
584 | */ | |
585 | arc_buf_t * | |
586 | dbuf_loan_arcbuf(dmu_buf_impl_t *db) | |
587 | { | |
588 | arc_buf_t *abuf; | |
589 | ||
590 | mutex_enter(&db->db_mtx); | |
591 | if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) { | |
592 | int blksz = db->db.db_size; | |
b0bc7a84 | 593 | spa_t *spa = db->db_objset->os_spa; |
572e2857 | 594 | |
428870ff | 595 | mutex_exit(&db->db_mtx); |
572e2857 | 596 | abuf = arc_loan_buf(spa, blksz); |
428870ff BB |
597 | bcopy(db->db.db_data, abuf->b_data, blksz); |
598 | } else { | |
599 | abuf = db->db_buf; | |
600 | arc_loan_inuse_buf(abuf, db); | |
0c66c32d | 601 | dbuf_clear_data(db); |
428870ff BB |
602 | mutex_exit(&db->db_mtx); |
603 | } | |
604 | return (abuf); | |
605 | } | |
606 | ||
34dc7c2f BB |
607 | uint64_t |
608 | dbuf_whichblock(dnode_t *dn, uint64_t offset) | |
609 | { | |
610 | if (dn->dn_datablkshift) { | |
611 | return (offset >> dn->dn_datablkshift); | |
612 | } else { | |
613 | ASSERT3U(offset, <, dn->dn_datablksz); | |
614 | return (0); | |
615 | } | |
616 | } | |
617 | ||
618 | static void | |
619 | dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb) | |
620 | { | |
621 | dmu_buf_impl_t *db = vdb; | |
622 | ||
623 | mutex_enter(&db->db_mtx); | |
624 | ASSERT3U(db->db_state, ==, DB_READ); | |
625 | /* | |
626 | * All reads are synchronous, so we must have a hold on the dbuf | |
627 | */ | |
628 | ASSERT(refcount_count(&db->db_holds) > 0); | |
629 | ASSERT(db->db_buf == NULL); | |
630 | ASSERT(db->db.db_data == NULL); | |
631 | if (db->db_level == 0 && db->db_freed_in_flight) { | |
632 | /* we were freed in flight; disregard any error */ | |
633 | arc_release(buf, db); | |
634 | bzero(buf->b_data, db->db.db_size); | |
635 | arc_buf_freeze(buf); | |
636 | db->db_freed_in_flight = FALSE; | |
637 | dbuf_set_data(db, buf); | |
638 | db->db_state = DB_CACHED; | |
639 | } else if (zio == NULL || zio->io_error == 0) { | |
640 | dbuf_set_data(db, buf); | |
641 | db->db_state = DB_CACHED; | |
642 | } else { | |
428870ff | 643 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f | 644 | ASSERT3P(db->db_buf, ==, NULL); |
13fe0198 | 645 | VERIFY(arc_buf_remove_ref(buf, db)); |
34dc7c2f BB |
646 | db->db_state = DB_UNCACHED; |
647 | } | |
648 | cv_broadcast(&db->db_changed); | |
428870ff | 649 | dbuf_rele_and_unlock(db, NULL); |
34dc7c2f BB |
650 | } |
651 | ||
5f6d0b6f | 652 | static int |
34dc7c2f BB |
653 | dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t *flags) |
654 | { | |
572e2857 | 655 | dnode_t *dn; |
5dbd68a3 | 656 | zbookmark_phys_t zb; |
2a432414 | 657 | uint32_t aflags = ARC_FLAG_NOWAIT; |
5f6d0b6f | 658 | int err; |
34dc7c2f | 659 | |
572e2857 BB |
660 | DB_DNODE_ENTER(db); |
661 | dn = DB_DNODE(db); | |
34dc7c2f BB |
662 | ASSERT(!refcount_is_zero(&db->db_holds)); |
663 | /* We need the struct_rwlock to prevent db_blkptr from changing. */ | |
b128c09f | 664 | ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); |
34dc7c2f BB |
665 | ASSERT(MUTEX_HELD(&db->db_mtx)); |
666 | ASSERT(db->db_state == DB_UNCACHED); | |
667 | ASSERT(db->db_buf == NULL); | |
668 | ||
428870ff | 669 | if (db->db_blkid == DMU_BONUS_BLKID) { |
9babb374 | 670 | int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen); |
34dc7c2f BB |
671 | |
672 | ASSERT3U(bonuslen, <=, db->db.db_size); | |
673 | db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN); | |
d164b209 | 674 | arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); |
34dc7c2f BB |
675 | if (bonuslen < DN_MAX_BONUSLEN) |
676 | bzero(db->db.db_data, DN_MAX_BONUSLEN); | |
9babb374 BB |
677 | if (bonuslen) |
678 | bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen); | |
572e2857 | 679 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
680 | db->db_state = DB_CACHED; |
681 | mutex_exit(&db->db_mtx); | |
5f6d0b6f | 682 | return (0); |
34dc7c2f BB |
683 | } |
684 | ||
b128c09f BB |
685 | /* |
686 | * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync() | |
687 | * processes the delete record and clears the bp while we are waiting | |
688 | * for the dn_mtx (resulting in a "no" from block_freed). | |
689 | */ | |
690 | if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) || | |
691 | (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) || | |
692 | BP_IS_HOLE(db->db_blkptr)))) { | |
34dc7c2f BB |
693 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); |
694 | ||
572e2857 | 695 | DB_DNODE_EXIT(db); |
b0bc7a84 MG |
696 | dbuf_set_data(db, arc_buf_alloc(db->db_objset->os_spa, |
697 | db->db.db_size, db, type)); | |
34dc7c2f BB |
698 | bzero(db->db.db_data, db->db.db_size); |
699 | db->db_state = DB_CACHED; | |
700 | *flags |= DB_RF_CACHED; | |
701 | mutex_exit(&db->db_mtx); | |
5f6d0b6f | 702 | return (0); |
34dc7c2f BB |
703 | } |
704 | ||
572e2857 BB |
705 | DB_DNODE_EXIT(db); |
706 | ||
34dc7c2f BB |
707 | db->db_state = DB_READ; |
708 | mutex_exit(&db->db_mtx); | |
709 | ||
b128c09f | 710 | if (DBUF_IS_L2CACHEABLE(db)) |
2a432414 | 711 | aflags |= ARC_FLAG_L2CACHE; |
3a17a7a9 | 712 | if (DBUF_IS_L2COMPRESSIBLE(db)) |
2a432414 | 713 | aflags |= ARC_FLAG_L2COMPRESS; |
b128c09f | 714 | |
428870ff BB |
715 | SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ? |
716 | db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET, | |
717 | db->db.db_object, db->db_level, db->db_blkid); | |
34dc7c2f BB |
718 | |
719 | dbuf_add_ref(db, NULL); | |
b128c09f | 720 | |
5f6d0b6f | 721 | err = arc_read(zio, db->db_objset->os_spa, db->db_blkptr, |
34dc7c2f BB |
722 | dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ, |
723 | (*flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED, | |
724 | &aflags, &zb); | |
2a432414 | 725 | if (aflags & ARC_FLAG_CACHED) |
34dc7c2f | 726 | *flags |= DB_RF_CACHED; |
5f6d0b6f BB |
727 | |
728 | return (SET_ERROR(err)); | |
34dc7c2f BB |
729 | } |
730 | ||
731 | int | |
732 | dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) | |
733 | { | |
734 | int err = 0; | |
b0bc7a84 MG |
735 | boolean_t havepzio = (zio != NULL); |
736 | boolean_t prefetch; | |
572e2857 | 737 | dnode_t *dn; |
34dc7c2f BB |
738 | |
739 | /* | |
740 | * We don't have to hold the mutex to check db_state because it | |
741 | * can't be freed while we have a hold on the buffer. | |
742 | */ | |
743 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
744 | ||
b128c09f | 745 | if (db->db_state == DB_NOFILL) |
2e528b49 | 746 | return (SET_ERROR(EIO)); |
b128c09f | 747 | |
572e2857 BB |
748 | DB_DNODE_ENTER(db); |
749 | dn = DB_DNODE(db); | |
34dc7c2f | 750 | if ((flags & DB_RF_HAVESTRUCT) == 0) |
572e2857 | 751 | rw_enter(&dn->dn_struct_rwlock, RW_READER); |
34dc7c2f | 752 | |
428870ff | 753 | prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && |
572e2857 | 754 | (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL && |
b128c09f | 755 | DBUF_IS_CACHEABLE(db); |
34dc7c2f BB |
756 | |
757 | mutex_enter(&db->db_mtx); | |
758 | if (db->db_state == DB_CACHED) { | |
759 | mutex_exit(&db->db_mtx); | |
760 | if (prefetch) | |
572e2857 | 761 | dmu_zfetch(&dn->dn_zfetch, db->db.db_offset, |
34dc7c2f BB |
762 | db->db.db_size, TRUE); |
763 | if ((flags & DB_RF_HAVESTRUCT) == 0) | |
572e2857 BB |
764 | rw_exit(&dn->dn_struct_rwlock); |
765 | DB_DNODE_EXIT(db); | |
34dc7c2f | 766 | } else if (db->db_state == DB_UNCACHED) { |
572e2857 BB |
767 | spa_t *spa = dn->dn_objset->os_spa; |
768 | ||
769 | if (zio == NULL) | |
770 | zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); | |
5f6d0b6f BB |
771 | |
772 | err = dbuf_read_impl(db, zio, &flags); | |
34dc7c2f BB |
773 | |
774 | /* dbuf_read_impl has dropped db_mtx for us */ | |
775 | ||
5f6d0b6f | 776 | if (!err && prefetch) |
572e2857 | 777 | dmu_zfetch(&dn->dn_zfetch, db->db.db_offset, |
34dc7c2f BB |
778 | db->db.db_size, flags & DB_RF_CACHED); |
779 | ||
780 | if ((flags & DB_RF_HAVESTRUCT) == 0) | |
572e2857 BB |
781 | rw_exit(&dn->dn_struct_rwlock); |
782 | DB_DNODE_EXIT(db); | |
34dc7c2f | 783 | |
5f6d0b6f | 784 | if (!err && !havepzio) |
34dc7c2f BB |
785 | err = zio_wait(zio); |
786 | } else { | |
e49f1e20 WA |
787 | /* |
788 | * Another reader came in while the dbuf was in flight | |
789 | * between UNCACHED and CACHED. Either a writer will finish | |
790 | * writing the buffer (sending the dbuf to CACHED) or the | |
791 | * first reader's request will reach the read_done callback | |
792 | * and send the dbuf to CACHED. Otherwise, a failure | |
793 | * occurred and the dbuf went to UNCACHED. | |
794 | */ | |
34dc7c2f BB |
795 | mutex_exit(&db->db_mtx); |
796 | if (prefetch) | |
572e2857 | 797 | dmu_zfetch(&dn->dn_zfetch, db->db.db_offset, |
34dc7c2f BB |
798 | db->db.db_size, TRUE); |
799 | if ((flags & DB_RF_HAVESTRUCT) == 0) | |
572e2857 BB |
800 | rw_exit(&dn->dn_struct_rwlock); |
801 | DB_DNODE_EXIT(db); | |
34dc7c2f | 802 | |
e49f1e20 | 803 | /* Skip the wait per the caller's request. */ |
34dc7c2f BB |
804 | mutex_enter(&db->db_mtx); |
805 | if ((flags & DB_RF_NEVERWAIT) == 0) { | |
806 | while (db->db_state == DB_READ || | |
807 | db->db_state == DB_FILL) { | |
808 | ASSERT(db->db_state == DB_READ || | |
809 | (flags & DB_RF_HAVESTRUCT) == 0); | |
64dbba36 AL |
810 | DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *, |
811 | db, zio_t *, zio); | |
34dc7c2f BB |
812 | cv_wait(&db->db_changed, &db->db_mtx); |
813 | } | |
814 | if (db->db_state == DB_UNCACHED) | |
2e528b49 | 815 | err = SET_ERROR(EIO); |
34dc7c2f BB |
816 | } |
817 | mutex_exit(&db->db_mtx); | |
818 | } | |
819 | ||
820 | ASSERT(err || havepzio || db->db_state == DB_CACHED); | |
821 | return (err); | |
822 | } | |
823 | ||
824 | static void | |
825 | dbuf_noread(dmu_buf_impl_t *db) | |
826 | { | |
827 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
428870ff | 828 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f BB |
829 | mutex_enter(&db->db_mtx); |
830 | while (db->db_state == DB_READ || db->db_state == DB_FILL) | |
831 | cv_wait(&db->db_changed, &db->db_mtx); | |
832 | if (db->db_state == DB_UNCACHED) { | |
833 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); | |
b0bc7a84 | 834 | spa_t *spa = db->db_objset->os_spa; |
34dc7c2f BB |
835 | |
836 | ASSERT(db->db_buf == NULL); | |
837 | ASSERT(db->db.db_data == NULL); | |
572e2857 | 838 | dbuf_set_data(db, arc_buf_alloc(spa, db->db.db_size, db, type)); |
34dc7c2f | 839 | db->db_state = DB_FILL; |
b128c09f | 840 | } else if (db->db_state == DB_NOFILL) { |
0c66c32d | 841 | dbuf_clear_data(db); |
34dc7c2f BB |
842 | } else { |
843 | ASSERT3U(db->db_state, ==, DB_CACHED); | |
844 | } | |
845 | mutex_exit(&db->db_mtx); | |
846 | } | |
847 | ||
848 | /* | |
849 | * This is our just-in-time copy function. It makes a copy of | |
850 | * buffers, that have been modified in a previous transaction | |
851 | * group, before we modify them in the current active group. | |
852 | * | |
853 | * This function is used in two places: when we are dirtying a | |
854 | * buffer for the first time in a txg, and when we are freeing | |
855 | * a range in a dnode that includes this buffer. | |
856 | * | |
857 | * Note that when we are called from dbuf_free_range() we do | |
858 | * not put a hold on the buffer, we just traverse the active | |
859 | * dbuf list for the dnode. | |
860 | */ | |
861 | static void | |
862 | dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg) | |
863 | { | |
864 | dbuf_dirty_record_t *dr = db->db_last_dirty; | |
865 | ||
866 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
867 | ASSERT(db->db.db_data != NULL); | |
868 | ASSERT(db->db_level == 0); | |
869 | ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT); | |
870 | ||
871 | if (dr == NULL || | |
872 | (dr->dt.dl.dr_data != | |
428870ff | 873 | ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf))) |
34dc7c2f BB |
874 | return; |
875 | ||
876 | /* | |
877 | * If the last dirty record for this dbuf has not yet synced | |
878 | * and its referencing the dbuf data, either: | |
572e2857 | 879 | * reset the reference to point to a new copy, |
34dc7c2f BB |
880 | * or (if there a no active holders) |
881 | * just null out the current db_data pointer. | |
882 | */ | |
883 | ASSERT(dr->dr_txg >= txg - 2); | |
428870ff | 884 | if (db->db_blkid == DMU_BONUS_BLKID) { |
34dc7c2f BB |
885 | /* Note that the data bufs here are zio_bufs */ |
886 | dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN); | |
d164b209 | 887 | arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); |
34dc7c2f BB |
888 | bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN); |
889 | } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) { | |
890 | int size = db->db.db_size; | |
891 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); | |
b0bc7a84 | 892 | spa_t *spa = db->db_objset->os_spa; |
572e2857 | 893 | |
572e2857 | 894 | dr->dt.dl.dr_data = arc_buf_alloc(spa, size, db, type); |
34dc7c2f BB |
895 | bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size); |
896 | } else { | |
0c66c32d | 897 | dbuf_clear_data(db); |
34dc7c2f BB |
898 | } |
899 | } | |
900 | ||
901 | void | |
902 | dbuf_unoverride(dbuf_dirty_record_t *dr) | |
903 | { | |
904 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
428870ff | 905 | blkptr_t *bp = &dr->dt.dl.dr_overridden_by; |
34dc7c2f BB |
906 | uint64_t txg = dr->dr_txg; |
907 | ||
908 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
909 | ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC); | |
910 | ASSERT(db->db_level == 0); | |
911 | ||
428870ff | 912 | if (db->db_blkid == DMU_BONUS_BLKID || |
34dc7c2f BB |
913 | dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN) |
914 | return; | |
915 | ||
428870ff BB |
916 | ASSERT(db->db_data_pending != dr); |
917 | ||
34dc7c2f | 918 | /* free this block */ |
b0bc7a84 MG |
919 | if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite) |
920 | zio_free(db->db_objset->os_spa, txg, bp); | |
428870ff | 921 | |
34dc7c2f | 922 | dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; |
03c6040b GW |
923 | dr->dt.dl.dr_nopwrite = B_FALSE; |
924 | ||
34dc7c2f BB |
925 | /* |
926 | * Release the already-written buffer, so we leave it in | |
927 | * a consistent dirty state. Note that all callers are | |
928 | * modifying the buffer, so they will immediately do | |
929 | * another (redundant) arc_release(). Therefore, leave | |
930 | * the buf thawed to save the effort of freezing & | |
931 | * immediately re-thawing it. | |
932 | */ | |
933 | arc_release(dr->dt.dl.dr_data, db); | |
934 | } | |
935 | ||
b128c09f BB |
936 | /* |
937 | * Evict (if its unreferenced) or clear (if its referenced) any level-0 | |
938 | * data blocks in the free range, so that any future readers will find | |
b0bc7a84 | 939 | * empty blocks. |
ea97f8ce MA |
940 | * |
941 | * This is a no-op if the dataset is in the middle of an incremental | |
942 | * receive; see comment below for details. | |
b128c09f | 943 | */ |
34dc7c2f | 944 | void |
8951cb8d AR |
945 | dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid, |
946 | dmu_tx_t *tx) | |
34dc7c2f | 947 | { |
0c66c32d JG |
948 | dmu_buf_impl_t *db_search; |
949 | dmu_buf_impl_t *db, *db_next; | |
34dc7c2f | 950 | uint64_t txg = tx->tx_txg; |
8951cb8d | 951 | avl_index_t where; |
4254acb0 | 952 | boolean_t freespill = |
8951cb8d AR |
953 | (start_blkid == DMU_SPILL_BLKID || end_blkid == DMU_SPILL_BLKID); |
954 | ||
955 | if (end_blkid > dn->dn_maxblkid && !freespill) | |
956 | end_blkid = dn->dn_maxblkid; | |
957 | dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid); | |
34dc7c2f | 958 | |
0c66c32d | 959 | db_search = kmem_alloc(sizeof (dmu_buf_impl_t), KM_SLEEP); |
8951cb8d AR |
960 | db_search->db_level = 0; |
961 | db_search->db_blkid = start_blkid; | |
9925c28c | 962 | db_search->db_state = DB_SEARCH; |
ea97f8ce | 963 | |
b663a23d | 964 | mutex_enter(&dn->dn_dbufs_mtx); |
8951cb8d | 965 | if (start_blkid >= dn->dn_unlisted_l0_blkid && !freespill) { |
b663a23d | 966 | /* There can't be any dbufs in this range; no need to search. */ |
8951cb8d AR |
967 | #ifdef DEBUG |
968 | db = avl_find(&dn->dn_dbufs, db_search, &where); | |
969 | ASSERT3P(db, ==, NULL); | |
970 | db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER); | |
971 | ASSERT(db == NULL || db->db_level > 0); | |
972 | #endif | |
973 | goto out; | |
b663a23d | 974 | } else if (dmu_objset_is_receiving(dn->dn_objset)) { |
ea97f8ce | 975 | /* |
b663a23d MA |
976 | * If we are receiving, we expect there to be no dbufs in |
977 | * the range to be freed, because receive modifies each | |
978 | * block at most once, and in offset order. If this is | |
979 | * not the case, it can lead to performance problems, | |
980 | * so note that we unexpectedly took the slow path. | |
ea97f8ce | 981 | */ |
b663a23d | 982 | atomic_inc_64(&zfs_free_range_recv_miss); |
ea97f8ce MA |
983 | } |
984 | ||
8951cb8d AR |
985 | db = avl_find(&dn->dn_dbufs, db_search, &where); |
986 | ASSERT3P(db, ==, NULL); | |
987 | db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER); | |
988 | ||
989 | for (; db != NULL; db = db_next) { | |
990 | db_next = AVL_NEXT(&dn->dn_dbufs, db); | |
428870ff | 991 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
b128c09f | 992 | |
8951cb8d AR |
993 | if (db->db_level != 0 || db->db_blkid > end_blkid) { |
994 | break; | |
995 | } | |
996 | ASSERT3U(db->db_blkid, >=, start_blkid); | |
34dc7c2f BB |
997 | |
998 | /* found a level 0 buffer in the range */ | |
13fe0198 MA |
999 | mutex_enter(&db->db_mtx); |
1000 | if (dbuf_undirty(db, tx)) { | |
1001 | /* mutex has been dropped and dbuf destroyed */ | |
34dc7c2f | 1002 | continue; |
13fe0198 | 1003 | } |
34dc7c2f | 1004 | |
34dc7c2f | 1005 | if (db->db_state == DB_UNCACHED || |
b128c09f | 1006 | db->db_state == DB_NOFILL || |
34dc7c2f BB |
1007 | db->db_state == DB_EVICTING) { |
1008 | ASSERT(db->db.db_data == NULL); | |
1009 | mutex_exit(&db->db_mtx); | |
1010 | continue; | |
1011 | } | |
1012 | if (db->db_state == DB_READ || db->db_state == DB_FILL) { | |
1013 | /* will be handled in dbuf_read_done or dbuf_rele */ | |
1014 | db->db_freed_in_flight = TRUE; | |
1015 | mutex_exit(&db->db_mtx); | |
1016 | continue; | |
1017 | } | |
1018 | if (refcount_count(&db->db_holds) == 0) { | |
1019 | ASSERT(db->db_buf); | |
1020 | dbuf_clear(db); | |
1021 | continue; | |
1022 | } | |
1023 | /* The dbuf is referenced */ | |
1024 | ||
1025 | if (db->db_last_dirty != NULL) { | |
1026 | dbuf_dirty_record_t *dr = db->db_last_dirty; | |
1027 | ||
1028 | if (dr->dr_txg == txg) { | |
1029 | /* | |
1030 | * This buffer is "in-use", re-adjust the file | |
1031 | * size to reflect that this buffer may | |
1032 | * contain new data when we sync. | |
1033 | */ | |
428870ff BB |
1034 | if (db->db_blkid != DMU_SPILL_BLKID && |
1035 | db->db_blkid > dn->dn_maxblkid) | |
34dc7c2f BB |
1036 | dn->dn_maxblkid = db->db_blkid; |
1037 | dbuf_unoverride(dr); | |
1038 | } else { | |
1039 | /* | |
1040 | * This dbuf is not dirty in the open context. | |
1041 | * Either uncache it (if its not referenced in | |
1042 | * the open context) or reset its contents to | |
1043 | * empty. | |
1044 | */ | |
1045 | dbuf_fix_old_data(db, txg); | |
1046 | } | |
1047 | } | |
1048 | /* clear the contents if its cached */ | |
1049 | if (db->db_state == DB_CACHED) { | |
1050 | ASSERT(db->db.db_data != NULL); | |
1051 | arc_release(db->db_buf, db); | |
1052 | bzero(db->db.db_data, db->db.db_size); | |
1053 | arc_buf_freeze(db->db_buf); | |
1054 | } | |
1055 | ||
1056 | mutex_exit(&db->db_mtx); | |
1057 | } | |
8951cb8d AR |
1058 | |
1059 | out: | |
1060 | kmem_free(db_search, sizeof (dmu_buf_impl_t)); | |
34dc7c2f BB |
1061 | mutex_exit(&dn->dn_dbufs_mtx); |
1062 | } | |
1063 | ||
1064 | static int | |
1065 | dbuf_block_freeable(dmu_buf_impl_t *db) | |
1066 | { | |
1067 | dsl_dataset_t *ds = db->db_objset->os_dsl_dataset; | |
1068 | uint64_t birth_txg = 0; | |
1069 | ||
1070 | /* | |
1071 | * We don't need any locking to protect db_blkptr: | |
1072 | * If it's syncing, then db_last_dirty will be set | |
1073 | * so we'll ignore db_blkptr. | |
b0bc7a84 MG |
1074 | * |
1075 | * This logic ensures that only block births for | |
1076 | * filled blocks are considered. | |
34dc7c2f BB |
1077 | */ |
1078 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
b0bc7a84 MG |
1079 | if (db->db_last_dirty && (db->db_blkptr == NULL || |
1080 | !BP_IS_HOLE(db->db_blkptr))) { | |
34dc7c2f | 1081 | birth_txg = db->db_last_dirty->dr_txg; |
b0bc7a84 | 1082 | } else if (db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) { |
34dc7c2f | 1083 | birth_txg = db->db_blkptr->blk_birth; |
b0bc7a84 | 1084 | } |
34dc7c2f | 1085 | |
572e2857 | 1086 | /* |
b0bc7a84 | 1087 | * If this block don't exist or is in a snapshot, it can't be freed. |
572e2857 BB |
1088 | * Don't pass the bp to dsl_dataset_block_freeable() since we |
1089 | * are holding the db_mtx lock and might deadlock if we are | |
1090 | * prefetching a dedup-ed block. | |
1091 | */ | |
b0bc7a84 | 1092 | if (birth_txg != 0) |
34dc7c2f | 1093 | return (ds == NULL || |
572e2857 | 1094 | dsl_dataset_block_freeable(ds, NULL, birth_txg)); |
34dc7c2f | 1095 | else |
b0bc7a84 | 1096 | return (B_FALSE); |
34dc7c2f BB |
1097 | } |
1098 | ||
1099 | void | |
1100 | dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx) | |
1101 | { | |
1102 | arc_buf_t *buf, *obuf; | |
1103 | int osize = db->db.db_size; | |
1104 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); | |
572e2857 | 1105 | dnode_t *dn; |
34dc7c2f | 1106 | |
428870ff | 1107 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f | 1108 | |
572e2857 BB |
1109 | DB_DNODE_ENTER(db); |
1110 | dn = DB_DNODE(db); | |
1111 | ||
34dc7c2f | 1112 | /* XXX does *this* func really need the lock? */ |
572e2857 | 1113 | ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); |
34dc7c2f BB |
1114 | |
1115 | /* | |
b0bc7a84 | 1116 | * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held |
34dc7c2f BB |
1117 | * is OK, because there can be no other references to the db |
1118 | * when we are changing its size, so no concurrent DB_FILL can | |
1119 | * be happening. | |
1120 | */ | |
1121 | /* | |
1122 | * XXX we should be doing a dbuf_read, checking the return | |
1123 | * value and returning that up to our callers | |
1124 | */ | |
b0bc7a84 | 1125 | dmu_buf_will_dirty(&db->db, tx); |
34dc7c2f BB |
1126 | |
1127 | /* create the data buffer for the new block */ | |
572e2857 | 1128 | buf = arc_buf_alloc(dn->dn_objset->os_spa, size, db, type); |
34dc7c2f BB |
1129 | |
1130 | /* copy old block data to the new block */ | |
1131 | obuf = db->db_buf; | |
1132 | bcopy(obuf->b_data, buf->b_data, MIN(osize, size)); | |
1133 | /* zero the remainder */ | |
1134 | if (size > osize) | |
1135 | bzero((uint8_t *)buf->b_data + osize, size - osize); | |
1136 | ||
1137 | mutex_enter(&db->db_mtx); | |
1138 | dbuf_set_data(db, buf); | |
13fe0198 | 1139 | VERIFY(arc_buf_remove_ref(obuf, db)); |
34dc7c2f BB |
1140 | db->db.db_size = size; |
1141 | ||
1142 | if (db->db_level == 0) { | |
1143 | ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); | |
1144 | db->db_last_dirty->dt.dl.dr_data = buf; | |
1145 | } | |
1146 | mutex_exit(&db->db_mtx); | |
1147 | ||
572e2857 BB |
1148 | dnode_willuse_space(dn, size-osize, tx); |
1149 | DB_DNODE_EXIT(db); | |
34dc7c2f BB |
1150 | } |
1151 | ||
428870ff BB |
1152 | void |
1153 | dbuf_release_bp(dmu_buf_impl_t *db) | |
1154 | { | |
b0bc7a84 | 1155 | ASSERTV(objset_t *os = db->db_objset); |
428870ff BB |
1156 | |
1157 | ASSERT(dsl_pool_sync_context(dmu_objset_pool(os))); | |
1158 | ASSERT(arc_released(os->os_phys_buf) || | |
1159 | list_link_active(&os->os_dsl_dataset->ds_synced_link)); | |
1160 | ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf)); | |
1161 | ||
294f6806 | 1162 | (void) arc_release(db->db_buf, db); |
428870ff BB |
1163 | } |
1164 | ||
34dc7c2f BB |
1165 | dbuf_dirty_record_t * |
1166 | dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) | |
1167 | { | |
572e2857 BB |
1168 | dnode_t *dn; |
1169 | objset_t *os; | |
34dc7c2f BB |
1170 | dbuf_dirty_record_t **drp, *dr; |
1171 | int drop_struct_lock = FALSE; | |
b128c09f | 1172 | boolean_t do_free_accounting = B_FALSE; |
34dc7c2f BB |
1173 | int txgoff = tx->tx_txg & TXG_MASK; |
1174 | ||
1175 | ASSERT(tx->tx_txg != 0); | |
1176 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
1177 | DMU_TX_DIRTY_BUF(tx, db); | |
1178 | ||
572e2857 BB |
1179 | DB_DNODE_ENTER(db); |
1180 | dn = DB_DNODE(db); | |
34dc7c2f BB |
1181 | /* |
1182 | * Shouldn't dirty a regular buffer in syncing context. Private | |
1183 | * objects may be dirtied in syncing context, but only if they | |
1184 | * were already pre-dirtied in open context. | |
34dc7c2f BB |
1185 | */ |
1186 | ASSERT(!dmu_tx_is_syncing(tx) || | |
1187 | BP_IS_HOLE(dn->dn_objset->os_rootbp) || | |
9babb374 BB |
1188 | DMU_OBJECT_IS_SPECIAL(dn->dn_object) || |
1189 | dn->dn_objset->os_dsl_dataset == NULL); | |
34dc7c2f BB |
1190 | /* |
1191 | * We make this assert for private objects as well, but after we | |
1192 | * check if we're already dirty. They are allowed to re-dirty | |
1193 | * in syncing context. | |
1194 | */ | |
1195 | ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || | |
1196 | dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == | |
1197 | (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); | |
1198 | ||
1199 | mutex_enter(&db->db_mtx); | |
1200 | /* | |
1201 | * XXX make this true for indirects too? The problem is that | |
1202 | * transactions created with dmu_tx_create_assigned() from | |
1203 | * syncing context don't bother holding ahead. | |
1204 | */ | |
1205 | ASSERT(db->db_level != 0 || | |
b128c09f BB |
1206 | db->db_state == DB_CACHED || db->db_state == DB_FILL || |
1207 | db->db_state == DB_NOFILL); | |
34dc7c2f BB |
1208 | |
1209 | mutex_enter(&dn->dn_mtx); | |
1210 | /* | |
1211 | * Don't set dirtyctx to SYNC if we're just modifying this as we | |
1212 | * initialize the objset. | |
1213 | */ | |
1214 | if (dn->dn_dirtyctx == DN_UNDIRTIED && | |
1215 | !BP_IS_HOLE(dn->dn_objset->os_rootbp)) { | |
1216 | dn->dn_dirtyctx = | |
1217 | (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN); | |
1218 | ASSERT(dn->dn_dirtyctx_firstset == NULL); | |
79c76d5b | 1219 | dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP); |
34dc7c2f BB |
1220 | } |
1221 | mutex_exit(&dn->dn_mtx); | |
1222 | ||
428870ff BB |
1223 | if (db->db_blkid == DMU_SPILL_BLKID) |
1224 | dn->dn_have_spill = B_TRUE; | |
1225 | ||
34dc7c2f BB |
1226 | /* |
1227 | * If this buffer is already dirty, we're done. | |
1228 | */ | |
1229 | drp = &db->db_last_dirty; | |
1230 | ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg || | |
1231 | db->db.db_object == DMU_META_DNODE_OBJECT); | |
1232 | while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg) | |
1233 | drp = &dr->dr_next; | |
1234 | if (dr && dr->dr_txg == tx->tx_txg) { | |
572e2857 BB |
1235 | DB_DNODE_EXIT(db); |
1236 | ||
428870ff | 1237 | if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) { |
34dc7c2f BB |
1238 | /* |
1239 | * If this buffer has already been written out, | |
1240 | * we now need to reset its state. | |
1241 | */ | |
1242 | dbuf_unoverride(dr); | |
428870ff BB |
1243 | if (db->db.db_object != DMU_META_DNODE_OBJECT && |
1244 | db->db_state != DB_NOFILL) | |
34dc7c2f BB |
1245 | arc_buf_thaw(db->db_buf); |
1246 | } | |
1247 | mutex_exit(&db->db_mtx); | |
1248 | return (dr); | |
1249 | } | |
1250 | ||
1251 | /* | |
1252 | * Only valid if not already dirty. | |
1253 | */ | |
9babb374 BB |
1254 | ASSERT(dn->dn_object == 0 || |
1255 | dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == | |
34dc7c2f BB |
1256 | (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); |
1257 | ||
1258 | ASSERT3U(dn->dn_nlevels, >, db->db_level); | |
1259 | ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) || | |
1260 | dn->dn_phys->dn_nlevels > db->db_level || | |
1261 | dn->dn_next_nlevels[txgoff] > db->db_level || | |
1262 | dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level || | |
1263 | dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level); | |
1264 | ||
1265 | /* | |
1266 | * We should only be dirtying in syncing context if it's the | |
9babb374 BB |
1267 | * mos or we're initializing the os or it's a special object. |
1268 | * However, we are allowed to dirty in syncing context provided | |
1269 | * we already dirtied it in open context. Hence we must make | |
1270 | * this assertion only if we're not already dirty. | |
34dc7c2f | 1271 | */ |
572e2857 | 1272 | os = dn->dn_objset; |
9babb374 BB |
1273 | ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) || |
1274 | os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp)); | |
34dc7c2f BB |
1275 | ASSERT(db->db.db_size != 0); |
1276 | ||
1277 | dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); | |
1278 | ||
428870ff | 1279 | if (db->db_blkid != DMU_BONUS_BLKID) { |
34dc7c2f BB |
1280 | /* |
1281 | * Update the accounting. | |
b128c09f BB |
1282 | * Note: we delay "free accounting" until after we drop |
1283 | * the db_mtx. This keeps us from grabbing other locks | |
428870ff | 1284 | * (and possibly deadlocking) in bp_get_dsize() while |
b128c09f | 1285 | * also holding the db_mtx. |
34dc7c2f | 1286 | */ |
34dc7c2f | 1287 | dnode_willuse_space(dn, db->db.db_size, tx); |
b128c09f | 1288 | do_free_accounting = dbuf_block_freeable(db); |
34dc7c2f BB |
1289 | } |
1290 | ||
1291 | /* | |
1292 | * If this buffer is dirty in an old transaction group we need | |
1293 | * to make a copy of it so that the changes we make in this | |
1294 | * transaction group won't leak out when we sync the older txg. | |
1295 | */ | |
79c76d5b | 1296 | dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP); |
98f72a53 | 1297 | list_link_init(&dr->dr_dirty_node); |
34dc7c2f BB |
1298 | if (db->db_level == 0) { |
1299 | void *data_old = db->db_buf; | |
1300 | ||
b128c09f | 1301 | if (db->db_state != DB_NOFILL) { |
428870ff | 1302 | if (db->db_blkid == DMU_BONUS_BLKID) { |
b128c09f BB |
1303 | dbuf_fix_old_data(db, tx->tx_txg); |
1304 | data_old = db->db.db_data; | |
1305 | } else if (db->db.db_object != DMU_META_DNODE_OBJECT) { | |
1306 | /* | |
1307 | * Release the data buffer from the cache so | |
1308 | * that we can modify it without impacting | |
1309 | * possible other users of this cached data | |
1310 | * block. Note that indirect blocks and | |
1311 | * private objects are not released until the | |
1312 | * syncing state (since they are only modified | |
1313 | * then). | |
1314 | */ | |
1315 | arc_release(db->db_buf, db); | |
1316 | dbuf_fix_old_data(db, tx->tx_txg); | |
1317 | data_old = db->db_buf; | |
1318 | } | |
1319 | ASSERT(data_old != NULL); | |
34dc7c2f | 1320 | } |
34dc7c2f BB |
1321 | dr->dt.dl.dr_data = data_old; |
1322 | } else { | |
448d7aaa | 1323 | mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_NOLOCKDEP, NULL); |
34dc7c2f BB |
1324 | list_create(&dr->dt.di.dr_children, |
1325 | sizeof (dbuf_dirty_record_t), | |
1326 | offsetof(dbuf_dirty_record_t, dr_dirty_node)); | |
1327 | } | |
e8b96c60 MA |
1328 | if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL) |
1329 | dr->dr_accounted = db->db.db_size; | |
34dc7c2f BB |
1330 | dr->dr_dbuf = db; |
1331 | dr->dr_txg = tx->tx_txg; | |
1332 | dr->dr_next = *drp; | |
1333 | *drp = dr; | |
1334 | ||
1335 | /* | |
1336 | * We could have been freed_in_flight between the dbuf_noread | |
1337 | * and dbuf_dirty. We win, as though the dbuf_noread() had | |
1338 | * happened after the free. | |
1339 | */ | |
428870ff BB |
1340 | if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && |
1341 | db->db_blkid != DMU_SPILL_BLKID) { | |
34dc7c2f | 1342 | mutex_enter(&dn->dn_mtx); |
9bd274dd MA |
1343 | if (dn->dn_free_ranges[txgoff] != NULL) { |
1344 | range_tree_clear(dn->dn_free_ranges[txgoff], | |
1345 | db->db_blkid, 1); | |
1346 | } | |
34dc7c2f BB |
1347 | mutex_exit(&dn->dn_mtx); |
1348 | db->db_freed_in_flight = FALSE; | |
1349 | } | |
1350 | ||
1351 | /* | |
1352 | * This buffer is now part of this txg | |
1353 | */ | |
1354 | dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg); | |
1355 | db->db_dirtycnt += 1; | |
1356 | ASSERT3U(db->db_dirtycnt, <=, 3); | |
1357 | ||
1358 | mutex_exit(&db->db_mtx); | |
1359 | ||
428870ff BB |
1360 | if (db->db_blkid == DMU_BONUS_BLKID || |
1361 | db->db_blkid == DMU_SPILL_BLKID) { | |
34dc7c2f BB |
1362 | mutex_enter(&dn->dn_mtx); |
1363 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
1364 | list_insert_tail(&dn->dn_dirty_records[txgoff], dr); | |
1365 | mutex_exit(&dn->dn_mtx); | |
1366 | dnode_setdirty(dn, tx); | |
572e2857 | 1367 | DB_DNODE_EXIT(db); |
34dc7c2f | 1368 | return (dr); |
b128c09f BB |
1369 | } else if (do_free_accounting) { |
1370 | blkptr_t *bp = db->db_blkptr; | |
1371 | int64_t willfree = (bp && !BP_IS_HOLE(bp)) ? | |
428870ff | 1372 | bp_get_dsize(os->os_spa, bp) : db->db.db_size; |
b128c09f BB |
1373 | /* |
1374 | * This is only a guess -- if the dbuf is dirty | |
1375 | * in a previous txg, we don't know how much | |
1376 | * space it will use on disk yet. We should | |
1377 | * really have the struct_rwlock to access | |
1378 | * db_blkptr, but since this is just a guess, | |
1379 | * it's OK if we get an odd answer. | |
1380 | */ | |
572e2857 | 1381 | ddt_prefetch(os->os_spa, bp); |
b128c09f | 1382 | dnode_willuse_space(dn, -willfree, tx); |
34dc7c2f BB |
1383 | } |
1384 | ||
1385 | if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) { | |
1386 | rw_enter(&dn->dn_struct_rwlock, RW_READER); | |
1387 | drop_struct_lock = TRUE; | |
1388 | } | |
1389 | ||
b128c09f BB |
1390 | if (db->db_level == 0) { |
1391 | dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock); | |
1392 | ASSERT(dn->dn_maxblkid >= db->db_blkid); | |
1393 | } | |
1394 | ||
34dc7c2f BB |
1395 | if (db->db_level+1 < dn->dn_nlevels) { |
1396 | dmu_buf_impl_t *parent = db->db_parent; | |
1397 | dbuf_dirty_record_t *di; | |
1398 | int parent_held = FALSE; | |
1399 | ||
1400 | if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) { | |
1401 | int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; | |
1402 | ||
1403 | parent = dbuf_hold_level(dn, db->db_level+1, | |
1404 | db->db_blkid >> epbs, FTAG); | |
428870ff | 1405 | ASSERT(parent != NULL); |
34dc7c2f BB |
1406 | parent_held = TRUE; |
1407 | } | |
1408 | if (drop_struct_lock) | |
1409 | rw_exit(&dn->dn_struct_rwlock); | |
1410 | ASSERT3U(db->db_level+1, ==, parent->db_level); | |
1411 | di = dbuf_dirty(parent, tx); | |
1412 | if (parent_held) | |
1413 | dbuf_rele(parent, FTAG); | |
1414 | ||
1415 | mutex_enter(&db->db_mtx); | |
e8b96c60 MA |
1416 | /* |
1417 | * Since we've dropped the mutex, it's possible that | |
1418 | * dbuf_undirty() might have changed this out from under us. | |
1419 | */ | |
34dc7c2f BB |
1420 | if (db->db_last_dirty == dr || |
1421 | dn->dn_object == DMU_META_DNODE_OBJECT) { | |
1422 | mutex_enter(&di->dt.di.dr_mtx); | |
1423 | ASSERT3U(di->dr_txg, ==, tx->tx_txg); | |
1424 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
1425 | list_insert_tail(&di->dt.di.dr_children, dr); | |
1426 | mutex_exit(&di->dt.di.dr_mtx); | |
1427 | dr->dr_parent = di; | |
1428 | } | |
1429 | mutex_exit(&db->db_mtx); | |
1430 | } else { | |
1431 | ASSERT(db->db_level+1 == dn->dn_nlevels); | |
1432 | ASSERT(db->db_blkid < dn->dn_nblkptr); | |
572e2857 | 1433 | ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf); |
34dc7c2f BB |
1434 | mutex_enter(&dn->dn_mtx); |
1435 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
1436 | list_insert_tail(&dn->dn_dirty_records[txgoff], dr); | |
1437 | mutex_exit(&dn->dn_mtx); | |
1438 | if (drop_struct_lock) | |
1439 | rw_exit(&dn->dn_struct_rwlock); | |
1440 | } | |
1441 | ||
1442 | dnode_setdirty(dn, tx); | |
572e2857 | 1443 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
1444 | return (dr); |
1445 | } | |
1446 | ||
13fe0198 | 1447 | /* |
e49f1e20 WA |
1448 | * Undirty a buffer in the transaction group referenced by the given |
1449 | * transaction. Return whether this evicted the dbuf. | |
13fe0198 MA |
1450 | */ |
1451 | static boolean_t | |
34dc7c2f BB |
1452 | dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx) |
1453 | { | |
572e2857 | 1454 | dnode_t *dn; |
34dc7c2f BB |
1455 | uint64_t txg = tx->tx_txg; |
1456 | dbuf_dirty_record_t *dr, **drp; | |
1457 | ||
1458 | ASSERT(txg != 0); | |
4bda3bd0 MA |
1459 | |
1460 | /* | |
1461 | * Due to our use of dn_nlevels below, this can only be called | |
1462 | * in open context, unless we are operating on the MOS. | |
1463 | * From syncing context, dn_nlevels may be different from the | |
1464 | * dn_nlevels used when dbuf was dirtied. | |
1465 | */ | |
1466 | ASSERT(db->db_objset == | |
1467 | dmu_objset_pool(db->db_objset)->dp_meta_objset || | |
1468 | txg != spa_syncing_txg(dmu_objset_spa(db->db_objset))); | |
428870ff | 1469 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
13fe0198 MA |
1470 | ASSERT0(db->db_level); |
1471 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
34dc7c2f | 1472 | |
34dc7c2f BB |
1473 | /* |
1474 | * If this buffer is not dirty, we're done. | |
1475 | */ | |
1476 | for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next) | |
1477 | if (dr->dr_txg <= txg) | |
1478 | break; | |
13fe0198 MA |
1479 | if (dr == NULL || dr->dr_txg < txg) |
1480 | return (B_FALSE); | |
34dc7c2f | 1481 | ASSERT(dr->dr_txg == txg); |
428870ff | 1482 | ASSERT(dr->dr_dbuf == db); |
34dc7c2f | 1483 | |
572e2857 BB |
1484 | DB_DNODE_ENTER(db); |
1485 | dn = DB_DNODE(db); | |
1486 | ||
34dc7c2f BB |
1487 | dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); |
1488 | ||
1489 | ASSERT(db->db.db_size != 0); | |
1490 | ||
4bda3bd0 MA |
1491 | dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset), |
1492 | dr->dr_accounted, txg); | |
34dc7c2f BB |
1493 | |
1494 | *drp = dr->dr_next; | |
1495 | ||
ef3c1dea GR |
1496 | /* |
1497 | * Note that there are three places in dbuf_dirty() | |
1498 | * where this dirty record may be put on a list. | |
1499 | * Make sure to do a list_remove corresponding to | |
1500 | * every one of those list_insert calls. | |
1501 | */ | |
34dc7c2f BB |
1502 | if (dr->dr_parent) { |
1503 | mutex_enter(&dr->dr_parent->dt.di.dr_mtx); | |
1504 | list_remove(&dr->dr_parent->dt.di.dr_children, dr); | |
1505 | mutex_exit(&dr->dr_parent->dt.di.dr_mtx); | |
ef3c1dea | 1506 | } else if (db->db_blkid == DMU_SPILL_BLKID || |
4bda3bd0 | 1507 | db->db_level + 1 == dn->dn_nlevels) { |
b128c09f | 1508 | ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf); |
34dc7c2f BB |
1509 | mutex_enter(&dn->dn_mtx); |
1510 | list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr); | |
1511 | mutex_exit(&dn->dn_mtx); | |
1512 | } | |
572e2857 | 1513 | DB_DNODE_EXIT(db); |
34dc7c2f | 1514 | |
13fe0198 MA |
1515 | if (db->db_state != DB_NOFILL) { |
1516 | dbuf_unoverride(dr); | |
34dc7c2f | 1517 | |
34dc7c2f | 1518 | ASSERT(db->db_buf != NULL); |
13fe0198 MA |
1519 | ASSERT(dr->dt.dl.dr_data != NULL); |
1520 | if (dr->dt.dl.dr_data != db->db_buf) | |
1521 | VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db)); | |
34dc7c2f | 1522 | } |
58c4aa00 | 1523 | |
34dc7c2f BB |
1524 | kmem_free(dr, sizeof (dbuf_dirty_record_t)); |
1525 | ||
1526 | ASSERT(db->db_dirtycnt > 0); | |
1527 | db->db_dirtycnt -= 1; | |
1528 | ||
1529 | if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) { | |
1530 | arc_buf_t *buf = db->db_buf; | |
1531 | ||
428870ff | 1532 | ASSERT(db->db_state == DB_NOFILL || arc_released(buf)); |
0c66c32d | 1533 | dbuf_clear_data(db); |
13fe0198 | 1534 | VERIFY(arc_buf_remove_ref(buf, db)); |
34dc7c2f | 1535 | dbuf_evict(db); |
13fe0198 | 1536 | return (B_TRUE); |
34dc7c2f BB |
1537 | } |
1538 | ||
13fe0198 | 1539 | return (B_FALSE); |
34dc7c2f BB |
1540 | } |
1541 | ||
34dc7c2f | 1542 | void |
b0bc7a84 | 1543 | dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx) |
34dc7c2f | 1544 | { |
b0bc7a84 | 1545 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; |
34dc7c2f BB |
1546 | int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH; |
1547 | ||
1548 | ASSERT(tx->tx_txg != 0); | |
1549 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
1550 | ||
572e2857 BB |
1551 | DB_DNODE_ENTER(db); |
1552 | if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock)) | |
34dc7c2f | 1553 | rf |= DB_RF_HAVESTRUCT; |
572e2857 | 1554 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
1555 | (void) dbuf_read(db, NULL, rf); |
1556 | (void) dbuf_dirty(db, tx); | |
1557 | } | |
1558 | ||
b128c09f BB |
1559 | void |
1560 | dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) | |
1561 | { | |
1562 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
1563 | ||
1564 | db->db_state = DB_NOFILL; | |
1565 | ||
1566 | dmu_buf_will_fill(db_fake, tx); | |
1567 | } | |
1568 | ||
34dc7c2f BB |
1569 | void |
1570 | dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) | |
1571 | { | |
1572 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
1573 | ||
428870ff | 1574 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f BB |
1575 | ASSERT(tx->tx_txg != 0); |
1576 | ASSERT(db->db_level == 0); | |
1577 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
1578 | ||
1579 | ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT || | |
1580 | dmu_tx_private_ok(tx)); | |
1581 | ||
1582 | dbuf_noread(db); | |
1583 | (void) dbuf_dirty(db, tx); | |
1584 | } | |
1585 | ||
1586 | #pragma weak dmu_buf_fill_done = dbuf_fill_done | |
1587 | /* ARGSUSED */ | |
1588 | void | |
1589 | dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx) | |
1590 | { | |
1591 | mutex_enter(&db->db_mtx); | |
1592 | DBUF_VERIFY(db); | |
1593 | ||
1594 | if (db->db_state == DB_FILL) { | |
1595 | if (db->db_level == 0 && db->db_freed_in_flight) { | |
428870ff | 1596 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f BB |
1597 | /* we were freed while filling */ |
1598 | /* XXX dbuf_undirty? */ | |
1599 | bzero(db->db.db_data, db->db.db_size); | |
1600 | db->db_freed_in_flight = FALSE; | |
1601 | } | |
1602 | db->db_state = DB_CACHED; | |
1603 | cv_broadcast(&db->db_changed); | |
1604 | } | |
1605 | mutex_exit(&db->db_mtx); | |
1606 | } | |
1607 | ||
9b67f605 MA |
1608 | void |
1609 | dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data, | |
1610 | bp_embedded_type_t etype, enum zio_compress comp, | |
1611 | int uncompressed_size, int compressed_size, int byteorder, | |
1612 | dmu_tx_t *tx) | |
1613 | { | |
1614 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; | |
1615 | struct dirty_leaf *dl; | |
1616 | dmu_object_type_t type; | |
1617 | ||
241b5415 MA |
1618 | if (etype == BP_EMBEDDED_TYPE_DATA) { |
1619 | ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset), | |
1620 | SPA_FEATURE_EMBEDDED_DATA)); | |
1621 | } | |
1622 | ||
9b67f605 MA |
1623 | DB_DNODE_ENTER(db); |
1624 | type = DB_DNODE(db)->dn_type; | |
1625 | DB_DNODE_EXIT(db); | |
1626 | ||
1627 | ASSERT0(db->db_level); | |
1628 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); | |
1629 | ||
1630 | dmu_buf_will_not_fill(dbuf, tx); | |
1631 | ||
1632 | ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); | |
1633 | dl = &db->db_last_dirty->dt.dl; | |
1634 | encode_embedded_bp_compressed(&dl->dr_overridden_by, | |
1635 | data, comp, uncompressed_size, compressed_size); | |
1636 | BPE_SET_ETYPE(&dl->dr_overridden_by, etype); | |
1637 | BP_SET_TYPE(&dl->dr_overridden_by, type); | |
1638 | BP_SET_LEVEL(&dl->dr_overridden_by, 0); | |
1639 | BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder); | |
1640 | ||
1641 | dl->dr_override_state = DR_OVERRIDDEN; | |
1642 | dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg; | |
1643 | } | |
1644 | ||
9babb374 BB |
1645 | /* |
1646 | * Directly assign a provided arc buf to a given dbuf if it's not referenced | |
1647 | * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf. | |
1648 | */ | |
1649 | void | |
1650 | dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx) | |
1651 | { | |
1652 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
428870ff | 1653 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
9babb374 BB |
1654 | ASSERT(db->db_level == 0); |
1655 | ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA); | |
1656 | ASSERT(buf != NULL); | |
1657 | ASSERT(arc_buf_size(buf) == db->db.db_size); | |
1658 | ASSERT(tx->tx_txg != 0); | |
1659 | ||
1660 | arc_return_buf(buf, db); | |
1661 | ASSERT(arc_released(buf)); | |
1662 | ||
1663 | mutex_enter(&db->db_mtx); | |
1664 | ||
1665 | while (db->db_state == DB_READ || db->db_state == DB_FILL) | |
1666 | cv_wait(&db->db_changed, &db->db_mtx); | |
1667 | ||
1668 | ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED); | |
1669 | ||
1670 | if (db->db_state == DB_CACHED && | |
1671 | refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) { | |
1672 | mutex_exit(&db->db_mtx); | |
1673 | (void) dbuf_dirty(db, tx); | |
1674 | bcopy(buf->b_data, db->db.db_data, db->db.db_size); | |
13fe0198 | 1675 | VERIFY(arc_buf_remove_ref(buf, db)); |
428870ff | 1676 | xuio_stat_wbuf_copied(); |
9babb374 BB |
1677 | return; |
1678 | } | |
1679 | ||
428870ff | 1680 | xuio_stat_wbuf_nocopy(); |
9babb374 BB |
1681 | if (db->db_state == DB_CACHED) { |
1682 | dbuf_dirty_record_t *dr = db->db_last_dirty; | |
1683 | ||
1684 | ASSERT(db->db_buf != NULL); | |
1685 | if (dr != NULL && dr->dr_txg == tx->tx_txg) { | |
1686 | ASSERT(dr->dt.dl.dr_data == db->db_buf); | |
1687 | if (!arc_released(db->db_buf)) { | |
1688 | ASSERT(dr->dt.dl.dr_override_state == | |
1689 | DR_OVERRIDDEN); | |
1690 | arc_release(db->db_buf, db); | |
1691 | } | |
1692 | dr->dt.dl.dr_data = buf; | |
13fe0198 | 1693 | VERIFY(arc_buf_remove_ref(db->db_buf, db)); |
9babb374 BB |
1694 | } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) { |
1695 | arc_release(db->db_buf, db); | |
13fe0198 | 1696 | VERIFY(arc_buf_remove_ref(db->db_buf, db)); |
9babb374 BB |
1697 | } |
1698 | db->db_buf = NULL; | |
1699 | } | |
1700 | ASSERT(db->db_buf == NULL); | |
1701 | dbuf_set_data(db, buf); | |
1702 | db->db_state = DB_FILL; | |
1703 | mutex_exit(&db->db_mtx); | |
1704 | (void) dbuf_dirty(db, tx); | |
b0bc7a84 | 1705 | dmu_buf_fill_done(&db->db, tx); |
9babb374 BB |
1706 | } |
1707 | ||
34dc7c2f BB |
1708 | /* |
1709 | * "Clear" the contents of this dbuf. This will mark the dbuf | |
e8b96c60 | 1710 | * EVICTING and clear *most* of its references. Unfortunately, |
34dc7c2f BB |
1711 | * when we are not holding the dn_dbufs_mtx, we can't clear the |
1712 | * entry in the dn_dbufs list. We have to wait until dbuf_destroy() | |
1713 | * in this case. For callers from the DMU we will usually see: | |
bd089c54 | 1714 | * dbuf_clear()->arc_clear_callback()->dbuf_do_evict()->dbuf_destroy() |
34dc7c2f | 1715 | * For the arc callback, we will usually see: |
572e2857 | 1716 | * dbuf_do_evict()->dbuf_clear();dbuf_destroy() |
34dc7c2f | 1717 | * Sometimes, though, we will get a mix of these two: |
bd089c54 | 1718 | * DMU: dbuf_clear()->arc_clear_callback() |
34dc7c2f | 1719 | * ARC: dbuf_do_evict()->dbuf_destroy() |
bd089c54 MA |
1720 | * |
1721 | * This routine will dissociate the dbuf from the arc, by calling | |
1722 | * arc_clear_callback(), but will not evict the data from the ARC. | |
34dc7c2f BB |
1723 | */ |
1724 | void | |
1725 | dbuf_clear(dmu_buf_impl_t *db) | |
1726 | { | |
572e2857 | 1727 | dnode_t *dn; |
34dc7c2f | 1728 | dmu_buf_impl_t *parent = db->db_parent; |
572e2857 | 1729 | dmu_buf_impl_t *dndb; |
bd089c54 | 1730 | boolean_t dbuf_gone = B_FALSE; |
34dc7c2f BB |
1731 | |
1732 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
1733 | ASSERT(refcount_is_zero(&db->db_holds)); | |
1734 | ||
1735 | dbuf_evict_user(db); | |
1736 | ||
1737 | if (db->db_state == DB_CACHED) { | |
1738 | ASSERT(db->db.db_data != NULL); | |
428870ff | 1739 | if (db->db_blkid == DMU_BONUS_BLKID) { |
34dc7c2f | 1740 | zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN); |
d164b209 | 1741 | arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); |
34dc7c2f BB |
1742 | } |
1743 | db->db.db_data = NULL; | |
1744 | db->db_state = DB_UNCACHED; | |
1745 | } | |
1746 | ||
b128c09f | 1747 | ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL); |
34dc7c2f BB |
1748 | ASSERT(db->db_data_pending == NULL); |
1749 | ||
1750 | db->db_state = DB_EVICTING; | |
1751 | db->db_blkptr = NULL; | |
1752 | ||
572e2857 BB |
1753 | DB_DNODE_ENTER(db); |
1754 | dn = DB_DNODE(db); | |
1755 | dndb = dn->dn_dbuf; | |
428870ff | 1756 | if (db->db_blkid != DMU_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) { |
8951cb8d | 1757 | avl_remove(&dn->dn_dbufs, db); |
73ad4a9f | 1758 | atomic_dec_32(&dn->dn_dbufs_count); |
572e2857 BB |
1759 | membar_producer(); |
1760 | DB_DNODE_EXIT(db); | |
1761 | /* | |
1762 | * Decrementing the dbuf count means that the hold corresponding | |
1763 | * to the removed dbuf is no longer discounted in dnode_move(), | |
1764 | * so the dnode cannot be moved until after we release the hold. | |
1765 | * The membar_producer() ensures visibility of the decremented | |
1766 | * value in dnode_move(), since DB_DNODE_EXIT doesn't actually | |
1767 | * release any lock. | |
1768 | */ | |
34dc7c2f | 1769 | dnode_rele(dn, db); |
572e2857 BB |
1770 | db->db_dnode_handle = NULL; |
1771 | } else { | |
1772 | DB_DNODE_EXIT(db); | |
34dc7c2f BB |
1773 | } |
1774 | ||
1775 | if (db->db_buf) | |
bd089c54 | 1776 | dbuf_gone = arc_clear_callback(db->db_buf); |
34dc7c2f BB |
1777 | |
1778 | if (!dbuf_gone) | |
1779 | mutex_exit(&db->db_mtx); | |
1780 | ||
1781 | /* | |
572e2857 | 1782 | * If this dbuf is referenced from an indirect dbuf, |
34dc7c2f BB |
1783 | * decrement the ref count on the indirect dbuf. |
1784 | */ | |
1785 | if (parent && parent != dndb) | |
1786 | dbuf_rele(parent, db); | |
1787 | } | |
1788 | ||
bf701a83 BB |
1789 | __attribute__((always_inline)) |
1790 | static inline int | |
34dc7c2f | 1791 | dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse, |
fc5bb51f | 1792 | dmu_buf_impl_t **parentp, blkptr_t **bpp, struct dbuf_hold_impl_data *dh) |
34dc7c2f BB |
1793 | { |
1794 | int nlevels, epbs; | |
1795 | ||
1796 | *parentp = NULL; | |
1797 | *bpp = NULL; | |
1798 | ||
428870ff BB |
1799 | ASSERT(blkid != DMU_BONUS_BLKID); |
1800 | ||
1801 | if (blkid == DMU_SPILL_BLKID) { | |
1802 | mutex_enter(&dn->dn_mtx); | |
1803 | if (dn->dn_have_spill && | |
1804 | (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) | |
1805 | *bpp = &dn->dn_phys->dn_spill; | |
1806 | else | |
1807 | *bpp = NULL; | |
1808 | dbuf_add_ref(dn->dn_dbuf, NULL); | |
1809 | *parentp = dn->dn_dbuf; | |
1810 | mutex_exit(&dn->dn_mtx); | |
1811 | return (0); | |
1812 | } | |
34dc7c2f BB |
1813 | |
1814 | if (dn->dn_phys->dn_nlevels == 0) | |
1815 | nlevels = 1; | |
1816 | else | |
1817 | nlevels = dn->dn_phys->dn_nlevels; | |
1818 | ||
1819 | epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; | |
1820 | ||
1821 | ASSERT3U(level * epbs, <, 64); | |
1822 | ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); | |
1823 | if (level >= nlevels || | |
1824 | (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) { | |
1825 | /* the buffer has no parent yet */ | |
2e528b49 | 1826 | return (SET_ERROR(ENOENT)); |
34dc7c2f BB |
1827 | } else if (level < nlevels-1) { |
1828 | /* this block is referenced from an indirect block */ | |
fc5bb51f BB |
1829 | int err; |
1830 | if (dh == NULL) { | |
1831 | err = dbuf_hold_impl(dn, level+1, blkid >> epbs, | |
1832 | fail_sparse, NULL, parentp); | |
d1d7e268 | 1833 | } else { |
fc5bb51f BB |
1834 | __dbuf_hold_impl_init(dh + 1, dn, dh->dh_level + 1, |
1835 | blkid >> epbs, fail_sparse, NULL, | |
1836 | parentp, dh->dh_depth + 1); | |
1837 | err = __dbuf_hold_impl(dh + 1); | |
1838 | } | |
34dc7c2f BB |
1839 | if (err) |
1840 | return (err); | |
1841 | err = dbuf_read(*parentp, NULL, | |
1842 | (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL)); | |
1843 | if (err) { | |
1844 | dbuf_rele(*parentp, NULL); | |
1845 | *parentp = NULL; | |
1846 | return (err); | |
1847 | } | |
1848 | *bpp = ((blkptr_t *)(*parentp)->db.db_data) + | |
1849 | (blkid & ((1ULL << epbs) - 1)); | |
1850 | return (0); | |
1851 | } else { | |
1852 | /* the block is referenced from the dnode */ | |
1853 | ASSERT3U(level, ==, nlevels-1); | |
1854 | ASSERT(dn->dn_phys->dn_nblkptr == 0 || | |
1855 | blkid < dn->dn_phys->dn_nblkptr); | |
1856 | if (dn->dn_dbuf) { | |
1857 | dbuf_add_ref(dn->dn_dbuf, NULL); | |
1858 | *parentp = dn->dn_dbuf; | |
1859 | } | |
1860 | *bpp = &dn->dn_phys->dn_blkptr[blkid]; | |
1861 | return (0); | |
1862 | } | |
1863 | } | |
1864 | ||
1865 | static dmu_buf_impl_t * | |
1866 | dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid, | |
1867 | dmu_buf_impl_t *parent, blkptr_t *blkptr) | |
1868 | { | |
428870ff | 1869 | objset_t *os = dn->dn_objset; |
34dc7c2f BB |
1870 | dmu_buf_impl_t *db, *odb; |
1871 | ||
1872 | ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); | |
1873 | ASSERT(dn->dn_type != DMU_OT_NONE); | |
1874 | ||
79c76d5b | 1875 | db = kmem_cache_alloc(dbuf_cache, KM_SLEEP); |
34dc7c2f BB |
1876 | |
1877 | db->db_objset = os; | |
1878 | db->db.db_object = dn->dn_object; | |
1879 | db->db_level = level; | |
1880 | db->db_blkid = blkid; | |
1881 | db->db_last_dirty = NULL; | |
1882 | db->db_dirtycnt = 0; | |
572e2857 | 1883 | db->db_dnode_handle = dn->dn_handle; |
34dc7c2f BB |
1884 | db->db_parent = parent; |
1885 | db->db_blkptr = blkptr; | |
1886 | ||
0c66c32d | 1887 | db->db_user = NULL; |
bc4501f7 JG |
1888 | db->db_user_immediate_evict = FALSE; |
1889 | db->db_freed_in_flight = FALSE; | |
1890 | db->db_pending_evict = FALSE; | |
34dc7c2f | 1891 | |
428870ff | 1892 | if (blkid == DMU_BONUS_BLKID) { |
34dc7c2f BB |
1893 | ASSERT3P(parent, ==, dn->dn_dbuf); |
1894 | db->db.db_size = DN_MAX_BONUSLEN - | |
1895 | (dn->dn_nblkptr-1) * sizeof (blkptr_t); | |
1896 | ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); | |
428870ff | 1897 | db->db.db_offset = DMU_BONUS_BLKID; |
34dc7c2f BB |
1898 | db->db_state = DB_UNCACHED; |
1899 | /* the bonus dbuf is not placed in the hash table */ | |
d164b209 | 1900 | arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); |
34dc7c2f | 1901 | return (db); |
428870ff BB |
1902 | } else if (blkid == DMU_SPILL_BLKID) { |
1903 | db->db.db_size = (blkptr != NULL) ? | |
1904 | BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE; | |
1905 | db->db.db_offset = 0; | |
34dc7c2f BB |
1906 | } else { |
1907 | int blocksize = | |
e8b96c60 | 1908 | db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz; |
34dc7c2f BB |
1909 | db->db.db_size = blocksize; |
1910 | db->db.db_offset = db->db_blkid * blocksize; | |
1911 | } | |
1912 | ||
1913 | /* | |
1914 | * Hold the dn_dbufs_mtx while we get the new dbuf | |
1915 | * in the hash table *and* added to the dbufs list. | |
1916 | * This prevents a possible deadlock with someone | |
1917 | * trying to look up this dbuf before its added to the | |
1918 | * dn_dbufs list. | |
1919 | */ | |
1920 | mutex_enter(&dn->dn_dbufs_mtx); | |
1921 | db->db_state = DB_EVICTING; | |
1922 | if ((odb = dbuf_hash_insert(db)) != NULL) { | |
1923 | /* someone else inserted it first */ | |
1924 | kmem_cache_free(dbuf_cache, db); | |
1925 | mutex_exit(&dn->dn_dbufs_mtx); | |
1926 | return (odb); | |
1927 | } | |
8951cb8d | 1928 | avl_add(&dn->dn_dbufs, db); |
b663a23d MA |
1929 | if (db->db_level == 0 && db->db_blkid >= |
1930 | dn->dn_unlisted_l0_blkid) | |
1931 | dn->dn_unlisted_l0_blkid = db->db_blkid + 1; | |
34dc7c2f BB |
1932 | db->db_state = DB_UNCACHED; |
1933 | mutex_exit(&dn->dn_dbufs_mtx); | |
d164b209 | 1934 | arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); |
34dc7c2f BB |
1935 | |
1936 | if (parent && parent != dn->dn_dbuf) | |
1937 | dbuf_add_ref(parent, db); | |
1938 | ||
1939 | ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || | |
1940 | refcount_count(&dn->dn_holds) > 0); | |
1941 | (void) refcount_add(&dn->dn_holds, db); | |
73ad4a9f | 1942 | atomic_inc_32(&dn->dn_dbufs_count); |
34dc7c2f BB |
1943 | |
1944 | dprintf_dbuf(db, "db=%p\n", db); | |
1945 | ||
1946 | return (db); | |
1947 | } | |
1948 | ||
1949 | static int | |
1950 | dbuf_do_evict(void *private) | |
1951 | { | |
bd089c54 | 1952 | dmu_buf_impl_t *db = private; |
34dc7c2f BB |
1953 | |
1954 | if (!MUTEX_HELD(&db->db_mtx)) | |
1955 | mutex_enter(&db->db_mtx); | |
1956 | ||
1957 | ASSERT(refcount_is_zero(&db->db_holds)); | |
1958 | ||
1959 | if (db->db_state != DB_EVICTING) { | |
1960 | ASSERT(db->db_state == DB_CACHED); | |
1961 | DBUF_VERIFY(db); | |
1962 | db->db_buf = NULL; | |
1963 | dbuf_evict(db); | |
1964 | } else { | |
1965 | mutex_exit(&db->db_mtx); | |
1966 | dbuf_destroy(db); | |
1967 | } | |
1968 | return (0); | |
1969 | } | |
1970 | ||
1971 | static void | |
1972 | dbuf_destroy(dmu_buf_impl_t *db) | |
1973 | { | |
1974 | ASSERT(refcount_is_zero(&db->db_holds)); | |
1975 | ||
428870ff | 1976 | if (db->db_blkid != DMU_BONUS_BLKID) { |
34dc7c2f BB |
1977 | /* |
1978 | * If this dbuf is still on the dn_dbufs list, | |
1979 | * remove it from that list. | |
1980 | */ | |
572e2857 BB |
1981 | if (db->db_dnode_handle != NULL) { |
1982 | dnode_t *dn; | |
34dc7c2f | 1983 | |
572e2857 BB |
1984 | DB_DNODE_ENTER(db); |
1985 | dn = DB_DNODE(db); | |
34dc7c2f | 1986 | mutex_enter(&dn->dn_dbufs_mtx); |
8951cb8d | 1987 | avl_remove(&dn->dn_dbufs, db); |
73ad4a9f | 1988 | atomic_dec_32(&dn->dn_dbufs_count); |
34dc7c2f | 1989 | mutex_exit(&dn->dn_dbufs_mtx); |
572e2857 BB |
1990 | DB_DNODE_EXIT(db); |
1991 | /* | |
1992 | * Decrementing the dbuf count means that the hold | |
1993 | * corresponding to the removed dbuf is no longer | |
1994 | * discounted in dnode_move(), so the dnode cannot be | |
1995 | * moved until after we release the hold. | |
1996 | */ | |
34dc7c2f | 1997 | dnode_rele(dn, db); |
572e2857 | 1998 | db->db_dnode_handle = NULL; |
34dc7c2f BB |
1999 | } |
2000 | dbuf_hash_remove(db); | |
2001 | } | |
2002 | db->db_parent = NULL; | |
2003 | db->db_buf = NULL; | |
2004 | ||
34dc7c2f BB |
2005 | ASSERT(db->db.db_data == NULL); |
2006 | ASSERT(db->db_hash_next == NULL); | |
2007 | ASSERT(db->db_blkptr == NULL); | |
2008 | ASSERT(db->db_data_pending == NULL); | |
2009 | ||
2010 | kmem_cache_free(dbuf_cache, db); | |
d164b209 | 2011 | arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); |
34dc7c2f BB |
2012 | } |
2013 | ||
2014 | void | |
e8b96c60 | 2015 | dbuf_prefetch(dnode_t *dn, uint64_t blkid, zio_priority_t prio) |
34dc7c2f BB |
2016 | { |
2017 | dmu_buf_impl_t *db = NULL; | |
2018 | blkptr_t *bp = NULL; | |
2019 | ||
428870ff | 2020 | ASSERT(blkid != DMU_BONUS_BLKID); |
34dc7c2f BB |
2021 | ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); |
2022 | ||
2023 | if (dnode_block_freed(dn, blkid)) | |
2024 | return; | |
2025 | ||
2026 | /* dbuf_find() returns with db_mtx held */ | |
6ebebace | 2027 | if ((db = dbuf_find(dn->dn_objset, dn->dn_object, 0, blkid))) { |
572e2857 BB |
2028 | /* |
2029 | * This dbuf is already in the cache. We assume that | |
2030 | * it is already CACHED, or else about to be either | |
2031 | * read or filled. | |
2032 | */ | |
34dc7c2f | 2033 | mutex_exit(&db->db_mtx); |
572e2857 | 2034 | return; |
34dc7c2f BB |
2035 | } |
2036 | ||
fc5bb51f | 2037 | if (dbuf_findbp(dn, 0, blkid, TRUE, &db, &bp, NULL) == 0) { |
9b67f605 | 2038 | if (bp && !BP_IS_HOLE(bp) && !BP_IS_EMBEDDED(bp)) { |
428870ff | 2039 | dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset; |
2a432414 GW |
2040 | arc_flags_t aflags = |
2041 | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH; | |
5dbd68a3 | 2042 | zbookmark_phys_t zb; |
428870ff BB |
2043 | |
2044 | SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET, | |
2045 | dn->dn_object, 0, blkid); | |
34dc7c2f | 2046 | |
294f6806 | 2047 | (void) arc_read(NULL, dn->dn_objset->os_spa, |
e8b96c60 | 2048 | bp, NULL, NULL, prio, |
34dc7c2f BB |
2049 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, |
2050 | &aflags, &zb); | |
2051 | } | |
2052 | if (db) | |
2053 | dbuf_rele(db, NULL); | |
2054 | } | |
2055 | } | |
2056 | ||
d1d7e268 | 2057 | #define DBUF_HOLD_IMPL_MAX_DEPTH 20 |
fc5bb51f | 2058 | |
34dc7c2f BB |
2059 | /* |
2060 | * Returns with db_holds incremented, and db_mtx not held. | |
2061 | * Note: dn_struct_rwlock must be held. | |
2062 | */ | |
fc5bb51f BB |
2063 | static int |
2064 | __dbuf_hold_impl(struct dbuf_hold_impl_data *dh) | |
34dc7c2f | 2065 | { |
fc5bb51f BB |
2066 | ASSERT3S(dh->dh_depth, <, DBUF_HOLD_IMPL_MAX_DEPTH); |
2067 | dh->dh_parent = NULL; | |
34dc7c2f | 2068 | |
fc5bb51f BB |
2069 | ASSERT(dh->dh_blkid != DMU_BONUS_BLKID); |
2070 | ASSERT(RW_LOCK_HELD(&dh->dh_dn->dn_struct_rwlock)); | |
2071 | ASSERT3U(dh->dh_dn->dn_nlevels, >, dh->dh_level); | |
34dc7c2f | 2072 | |
fc5bb51f | 2073 | *(dh->dh_dbp) = NULL; |
34dc7c2f BB |
2074 | top: |
2075 | /* dbuf_find() returns with db_mtx held */ | |
6ebebace JG |
2076 | dh->dh_db = dbuf_find(dh->dh_dn->dn_objset, dh->dh_dn->dn_object, |
2077 | dh->dh_level, dh->dh_blkid); | |
fc5bb51f BB |
2078 | |
2079 | if (dh->dh_db == NULL) { | |
2080 | dh->dh_bp = NULL; | |
2081 | ||
2082 | ASSERT3P(dh->dh_parent, ==, NULL); | |
2083 | dh->dh_err = dbuf_findbp(dh->dh_dn, dh->dh_level, dh->dh_blkid, | |
2084 | dh->dh_fail_sparse, &dh->dh_parent, | |
2085 | &dh->dh_bp, dh); | |
2086 | if (dh->dh_fail_sparse) { | |
d1d7e268 MK |
2087 | if (dh->dh_err == 0 && |
2088 | dh->dh_bp && BP_IS_HOLE(dh->dh_bp)) | |
2e528b49 | 2089 | dh->dh_err = SET_ERROR(ENOENT); |
fc5bb51f BB |
2090 | if (dh->dh_err) { |
2091 | if (dh->dh_parent) | |
2092 | dbuf_rele(dh->dh_parent, NULL); | |
2093 | return (dh->dh_err); | |
34dc7c2f BB |
2094 | } |
2095 | } | |
fc5bb51f BB |
2096 | if (dh->dh_err && dh->dh_err != ENOENT) |
2097 | return (dh->dh_err); | |
2098 | dh->dh_db = dbuf_create(dh->dh_dn, dh->dh_level, dh->dh_blkid, | |
2099 | dh->dh_parent, dh->dh_bp); | |
34dc7c2f BB |
2100 | } |
2101 | ||
fc5bb51f BB |
2102 | if (dh->dh_db->db_buf && refcount_is_zero(&dh->dh_db->db_holds)) { |
2103 | arc_buf_add_ref(dh->dh_db->db_buf, dh->dh_db); | |
2104 | if (dh->dh_db->db_buf->b_data == NULL) { | |
2105 | dbuf_clear(dh->dh_db); | |
2106 | if (dh->dh_parent) { | |
2107 | dbuf_rele(dh->dh_parent, NULL); | |
2108 | dh->dh_parent = NULL; | |
34dc7c2f BB |
2109 | } |
2110 | goto top; | |
2111 | } | |
fc5bb51f | 2112 | ASSERT3P(dh->dh_db->db.db_data, ==, dh->dh_db->db_buf->b_data); |
34dc7c2f BB |
2113 | } |
2114 | ||
fc5bb51f | 2115 | ASSERT(dh->dh_db->db_buf == NULL || arc_referenced(dh->dh_db->db_buf)); |
34dc7c2f BB |
2116 | |
2117 | /* | |
2118 | * If this buffer is currently syncing out, and we are are | |
2119 | * still referencing it from db_data, we need to make a copy | |
2120 | * of it in case we decide we want to dirty it again in this txg. | |
2121 | */ | |
fc5bb51f BB |
2122 | if (dh->dh_db->db_level == 0 && |
2123 | dh->dh_db->db_blkid != DMU_BONUS_BLKID && | |
2124 | dh->dh_dn->dn_object != DMU_META_DNODE_OBJECT && | |
2125 | dh->dh_db->db_state == DB_CACHED && dh->dh_db->db_data_pending) { | |
2126 | dh->dh_dr = dh->dh_db->db_data_pending; | |
2127 | ||
2128 | if (dh->dh_dr->dt.dl.dr_data == dh->dh_db->db_buf) { | |
2129 | dh->dh_type = DBUF_GET_BUFC_TYPE(dh->dh_db); | |
2130 | ||
2131 | dbuf_set_data(dh->dh_db, | |
2132 | arc_buf_alloc(dh->dh_dn->dn_objset->os_spa, | |
2133 | dh->dh_db->db.db_size, dh->dh_db, dh->dh_type)); | |
2134 | bcopy(dh->dh_dr->dt.dl.dr_data->b_data, | |
2135 | dh->dh_db->db.db_data, dh->dh_db->db.db_size); | |
34dc7c2f BB |
2136 | } |
2137 | } | |
2138 | ||
fc5bb51f | 2139 | (void) refcount_add(&dh->dh_db->db_holds, dh->dh_tag); |
fc5bb51f BB |
2140 | DBUF_VERIFY(dh->dh_db); |
2141 | mutex_exit(&dh->dh_db->db_mtx); | |
34dc7c2f BB |
2142 | |
2143 | /* NOTE: we can't rele the parent until after we drop the db_mtx */ | |
fc5bb51f BB |
2144 | if (dh->dh_parent) |
2145 | dbuf_rele(dh->dh_parent, NULL); | |
34dc7c2f | 2146 | |
fc5bb51f BB |
2147 | ASSERT3P(DB_DNODE(dh->dh_db), ==, dh->dh_dn); |
2148 | ASSERT3U(dh->dh_db->db_blkid, ==, dh->dh_blkid); | |
2149 | ASSERT3U(dh->dh_db->db_level, ==, dh->dh_level); | |
2150 | *(dh->dh_dbp) = dh->dh_db; | |
34dc7c2f BB |
2151 | |
2152 | return (0); | |
2153 | } | |
2154 | ||
fc5bb51f BB |
2155 | /* |
2156 | * The following code preserves the recursive function dbuf_hold_impl() | |
2157 | * but moves the local variables AND function arguments to the heap to | |
2158 | * minimize the stack frame size. Enough space is initially allocated | |
2159 | * on the stack for 20 levels of recursion. | |
2160 | */ | |
2161 | int | |
2162 | dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse, | |
2163 | void *tag, dmu_buf_impl_t **dbp) | |
2164 | { | |
2165 | struct dbuf_hold_impl_data *dh; | |
2166 | int error; | |
2167 | ||
d1d7e268 | 2168 | dh = kmem_zalloc(sizeof (struct dbuf_hold_impl_data) * |
79c76d5b | 2169 | DBUF_HOLD_IMPL_MAX_DEPTH, KM_SLEEP); |
fc5bb51f BB |
2170 | __dbuf_hold_impl_init(dh, dn, level, blkid, fail_sparse, tag, dbp, 0); |
2171 | ||
2172 | error = __dbuf_hold_impl(dh); | |
2173 | ||
d1d7e268 | 2174 | kmem_free(dh, sizeof (struct dbuf_hold_impl_data) * |
fc5bb51f BB |
2175 | DBUF_HOLD_IMPL_MAX_DEPTH); |
2176 | ||
2177 | return (error); | |
2178 | } | |
2179 | ||
2180 | static void | |
2181 | __dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh, | |
2182 | dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse, | |
2183 | void *tag, dmu_buf_impl_t **dbp, int depth) | |
2184 | { | |
2185 | dh->dh_dn = dn; | |
2186 | dh->dh_level = level; | |
2187 | dh->dh_blkid = blkid; | |
2188 | dh->dh_fail_sparse = fail_sparse; | |
2189 | dh->dh_tag = tag; | |
2190 | dh->dh_dbp = dbp; | |
2191 | dh->dh_depth = depth; | |
2192 | } | |
2193 | ||
34dc7c2f BB |
2194 | dmu_buf_impl_t * |
2195 | dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag) | |
2196 | { | |
2197 | dmu_buf_impl_t *db; | |
2198 | int err = dbuf_hold_impl(dn, 0, blkid, FALSE, tag, &db); | |
2199 | return (err ? NULL : db); | |
2200 | } | |
2201 | ||
2202 | dmu_buf_impl_t * | |
2203 | dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag) | |
2204 | { | |
2205 | dmu_buf_impl_t *db; | |
2206 | int err = dbuf_hold_impl(dn, level, blkid, FALSE, tag, &db); | |
2207 | return (err ? NULL : db); | |
2208 | } | |
2209 | ||
2210 | void | |
2211 | dbuf_create_bonus(dnode_t *dn) | |
2212 | { | |
2213 | ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); | |
2214 | ||
2215 | ASSERT(dn->dn_bonus == NULL); | |
428870ff BB |
2216 | dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL); |
2217 | } | |
2218 | ||
2219 | int | |
2220 | dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx) | |
2221 | { | |
2222 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
572e2857 BB |
2223 | dnode_t *dn; |
2224 | ||
428870ff | 2225 | if (db->db_blkid != DMU_SPILL_BLKID) |
2e528b49 | 2226 | return (SET_ERROR(ENOTSUP)); |
428870ff BB |
2227 | if (blksz == 0) |
2228 | blksz = SPA_MINBLOCKSIZE; | |
f1512ee6 MA |
2229 | ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset))); |
2230 | blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE); | |
428870ff | 2231 | |
572e2857 BB |
2232 | DB_DNODE_ENTER(db); |
2233 | dn = DB_DNODE(db); | |
2234 | rw_enter(&dn->dn_struct_rwlock, RW_WRITER); | |
428870ff | 2235 | dbuf_new_size(db, blksz, tx); |
572e2857 BB |
2236 | rw_exit(&dn->dn_struct_rwlock); |
2237 | DB_DNODE_EXIT(db); | |
428870ff BB |
2238 | |
2239 | return (0); | |
2240 | } | |
2241 | ||
2242 | void | |
2243 | dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx) | |
2244 | { | |
2245 | dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx); | |
34dc7c2f BB |
2246 | } |
2247 | ||
2248 | #pragma weak dmu_buf_add_ref = dbuf_add_ref | |
2249 | void | |
2250 | dbuf_add_ref(dmu_buf_impl_t *db, void *tag) | |
2251 | { | |
1fde1e37 | 2252 | VERIFY(refcount_add(&db->db_holds, tag) > 1); |
34dc7c2f BB |
2253 | } |
2254 | ||
6ebebace JG |
2255 | #pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref |
2256 | boolean_t | |
2257 | dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid, | |
2258 | void *tag) | |
2259 | { | |
2260 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
2261 | dmu_buf_impl_t *found_db; | |
2262 | boolean_t result = B_FALSE; | |
2263 | ||
d617648c | 2264 | if (blkid == DMU_BONUS_BLKID) |
6ebebace JG |
2265 | found_db = dbuf_find_bonus(os, obj); |
2266 | else | |
2267 | found_db = dbuf_find(os, obj, 0, blkid); | |
2268 | ||
2269 | if (found_db != NULL) { | |
2270 | if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) { | |
2271 | (void) refcount_add(&db->db_holds, tag); | |
2272 | result = B_TRUE; | |
2273 | } | |
d617648c | 2274 | mutex_exit(&found_db->db_mtx); |
6ebebace JG |
2275 | } |
2276 | return (result); | |
2277 | } | |
2278 | ||
572e2857 BB |
2279 | /* |
2280 | * If you call dbuf_rele() you had better not be referencing the dnode handle | |
2281 | * unless you have some other direct or indirect hold on the dnode. (An indirect | |
2282 | * hold is a hold on one of the dnode's dbufs, including the bonus buffer.) | |
2283 | * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the | |
2284 | * dnode's parent dbuf evicting its dnode handles. | |
2285 | */ | |
34dc7c2f BB |
2286 | void |
2287 | dbuf_rele(dmu_buf_impl_t *db, void *tag) | |
428870ff BB |
2288 | { |
2289 | mutex_enter(&db->db_mtx); | |
2290 | dbuf_rele_and_unlock(db, tag); | |
2291 | } | |
2292 | ||
b0bc7a84 MG |
2293 | void |
2294 | dmu_buf_rele(dmu_buf_t *db, void *tag) | |
2295 | { | |
2296 | dbuf_rele((dmu_buf_impl_t *)db, tag); | |
2297 | } | |
2298 | ||
428870ff BB |
2299 | /* |
2300 | * dbuf_rele() for an already-locked dbuf. This is necessary to allow | |
2301 | * db_dirtycnt and db_holds to be updated atomically. | |
2302 | */ | |
2303 | void | |
2304 | dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag) | |
34dc7c2f BB |
2305 | { |
2306 | int64_t holds; | |
2307 | ||
428870ff | 2308 | ASSERT(MUTEX_HELD(&db->db_mtx)); |
34dc7c2f BB |
2309 | DBUF_VERIFY(db); |
2310 | ||
572e2857 BB |
2311 | /* |
2312 | * Remove the reference to the dbuf before removing its hold on the | |
2313 | * dnode so we can guarantee in dnode_move() that a referenced bonus | |
2314 | * buffer has a corresponding dnode hold. | |
2315 | */ | |
34dc7c2f BB |
2316 | holds = refcount_remove(&db->db_holds, tag); |
2317 | ASSERT(holds >= 0); | |
2318 | ||
2319 | /* | |
2320 | * We can't freeze indirects if there is a possibility that they | |
2321 | * may be modified in the current syncing context. | |
2322 | */ | |
2323 | if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) | |
2324 | arc_buf_freeze(db->db_buf); | |
2325 | ||
2326 | if (holds == db->db_dirtycnt && | |
bc4501f7 | 2327 | db->db_level == 0 && db->db_user_immediate_evict) |
34dc7c2f BB |
2328 | dbuf_evict_user(db); |
2329 | ||
2330 | if (holds == 0) { | |
428870ff | 2331 | if (db->db_blkid == DMU_BONUS_BLKID) { |
4c7b7eed | 2332 | dnode_t *dn; |
bc4501f7 | 2333 | boolean_t evict_dbuf = db->db_pending_evict; |
572e2857 BB |
2334 | |
2335 | /* | |
4c7b7eed JG |
2336 | * If the dnode moves here, we cannot cross this |
2337 | * barrier until the move completes. | |
572e2857 BB |
2338 | */ |
2339 | DB_DNODE_ENTER(db); | |
4c7b7eed JG |
2340 | |
2341 | dn = DB_DNODE(db); | |
2342 | atomic_dec_32(&dn->dn_dbufs_count); | |
2343 | ||
2344 | /* | |
2345 | * Decrementing the dbuf count means that the bonus | |
2346 | * buffer's dnode hold is no longer discounted in | |
2347 | * dnode_move(). The dnode cannot move until after | |
bc4501f7 | 2348 | * the dnode_rele() below. |
4c7b7eed | 2349 | */ |
572e2857 | 2350 | DB_DNODE_EXIT(db); |
4c7b7eed JG |
2351 | |
2352 | /* | |
2353 | * Do not reference db after its lock is dropped. | |
2354 | * Another thread may evict it. | |
2355 | */ | |
2356 | mutex_exit(&db->db_mtx); | |
2357 | ||
bc4501f7 | 2358 | if (evict_dbuf) |
4c7b7eed | 2359 | dnode_evict_bonus(dn); |
bc4501f7 JG |
2360 | |
2361 | dnode_rele(dn, db); | |
34dc7c2f BB |
2362 | } else if (db->db_buf == NULL) { |
2363 | /* | |
2364 | * This is a special case: we never associated this | |
2365 | * dbuf with any data allocated from the ARC. | |
2366 | */ | |
b128c09f BB |
2367 | ASSERT(db->db_state == DB_UNCACHED || |
2368 | db->db_state == DB_NOFILL); | |
34dc7c2f BB |
2369 | dbuf_evict(db); |
2370 | } else if (arc_released(db->db_buf)) { | |
2371 | arc_buf_t *buf = db->db_buf; | |
2372 | /* | |
2373 | * This dbuf has anonymous data associated with it. | |
2374 | */ | |
0c66c32d | 2375 | dbuf_clear_data(db); |
13fe0198 | 2376 | VERIFY(arc_buf_remove_ref(buf, db)); |
34dc7c2f BB |
2377 | dbuf_evict(db); |
2378 | } else { | |
13fe0198 | 2379 | VERIFY(!arc_buf_remove_ref(db->db_buf, db)); |
1eb5bfa3 GW |
2380 | |
2381 | /* | |
2382 | * A dbuf will be eligible for eviction if either the | |
2383 | * 'primarycache' property is set or a duplicate | |
2384 | * copy of this buffer is already cached in the arc. | |
2385 | * | |
2386 | * In the case of the 'primarycache' a buffer | |
2387 | * is considered for eviction if it matches the | |
2388 | * criteria set in the property. | |
2389 | * | |
2390 | * To decide if our buffer is considered a | |
2391 | * duplicate, we must call into the arc to determine | |
2392 | * if multiple buffers are referencing the same | |
2393 | * block on-disk. If so, then we simply evict | |
2394 | * ourselves. | |
2395 | */ | |
bd089c54 MA |
2396 | if (!DBUF_IS_CACHEABLE(db)) { |
2397 | if (db->db_blkptr != NULL && | |
2398 | !BP_IS_HOLE(db->db_blkptr) && | |
2399 | !BP_IS_EMBEDDED(db->db_blkptr)) { | |
2400 | spa_t *spa = | |
2401 | dmu_objset_spa(db->db_objset); | |
2402 | blkptr_t bp = *db->db_blkptr; | |
2403 | dbuf_clear(db); | |
2404 | arc_freed(spa, &bp); | |
2405 | } else { | |
2406 | dbuf_clear(db); | |
2407 | } | |
bc4501f7 | 2408 | } else if (db->db_pending_evict || |
0c66c32d | 2409 | arc_buf_eviction_needed(db->db_buf)) { |
b128c09f | 2410 | dbuf_clear(db); |
bd089c54 | 2411 | } else { |
b128c09f | 2412 | mutex_exit(&db->db_mtx); |
bd089c54 | 2413 | } |
34dc7c2f BB |
2414 | } |
2415 | } else { | |
2416 | mutex_exit(&db->db_mtx); | |
2417 | } | |
2418 | } | |
2419 | ||
2420 | #pragma weak dmu_buf_refcount = dbuf_refcount | |
2421 | uint64_t | |
2422 | dbuf_refcount(dmu_buf_impl_t *db) | |
2423 | { | |
2424 | return (refcount_count(&db->db_holds)); | |
2425 | } | |
2426 | ||
2427 | void * | |
0c66c32d JG |
2428 | dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user, |
2429 | dmu_buf_user_t *new_user) | |
34dc7c2f | 2430 | { |
0c66c32d JG |
2431 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; |
2432 | ||
2433 | mutex_enter(&db->db_mtx); | |
2434 | dbuf_verify_user(db, DBVU_NOT_EVICTING); | |
2435 | if (db->db_user == old_user) | |
2436 | db->db_user = new_user; | |
2437 | else | |
2438 | old_user = db->db_user; | |
2439 | dbuf_verify_user(db, DBVU_NOT_EVICTING); | |
2440 | mutex_exit(&db->db_mtx); | |
2441 | ||
2442 | return (old_user); | |
34dc7c2f BB |
2443 | } |
2444 | ||
2445 | void * | |
0c66c32d | 2446 | dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user) |
34dc7c2f | 2447 | { |
0c66c32d | 2448 | return (dmu_buf_replace_user(db_fake, NULL, user)); |
34dc7c2f BB |
2449 | } |
2450 | ||
2451 | void * | |
0c66c32d | 2452 | dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user) |
34dc7c2f BB |
2453 | { |
2454 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
34dc7c2f | 2455 | |
bc4501f7 | 2456 | db->db_user_immediate_evict = TRUE; |
0c66c32d JG |
2457 | return (dmu_buf_set_user(db_fake, user)); |
2458 | } | |
34dc7c2f | 2459 | |
0c66c32d JG |
2460 | void * |
2461 | dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user) | |
2462 | { | |
2463 | return (dmu_buf_replace_user(db_fake, user, NULL)); | |
34dc7c2f BB |
2464 | } |
2465 | ||
2466 | void * | |
2467 | dmu_buf_get_user(dmu_buf_t *db_fake) | |
2468 | { | |
2469 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
34dc7c2f | 2470 | |
0c66c32d JG |
2471 | dbuf_verify_user(db, DBVU_NOT_EVICTING); |
2472 | return (db->db_user); | |
2473 | } | |
2474 | ||
2475 | void | |
2476 | dmu_buf_user_evict_wait() | |
2477 | { | |
2478 | taskq_wait(dbu_evict_taskq); | |
34dc7c2f BB |
2479 | } |
2480 | ||
9babb374 BB |
2481 | boolean_t |
2482 | dmu_buf_freeable(dmu_buf_t *dbuf) | |
2483 | { | |
2484 | boolean_t res = B_FALSE; | |
2485 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; | |
2486 | ||
2487 | if (db->db_blkptr) | |
2488 | res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset, | |
428870ff | 2489 | db->db_blkptr, db->db_blkptr->blk_birth); |
9babb374 BB |
2490 | |
2491 | return (res); | |
2492 | } | |
2493 | ||
03c6040b GW |
2494 | blkptr_t * |
2495 | dmu_buf_get_blkptr(dmu_buf_t *db) | |
2496 | { | |
2497 | dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; | |
2498 | return (dbi->db_blkptr); | |
2499 | } | |
2500 | ||
34dc7c2f BB |
2501 | static void |
2502 | dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db) | |
2503 | { | |
2504 | /* ASSERT(dmu_tx_is_syncing(tx) */ | |
2505 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
2506 | ||
2507 | if (db->db_blkptr != NULL) | |
2508 | return; | |
2509 | ||
428870ff BB |
2510 | if (db->db_blkid == DMU_SPILL_BLKID) { |
2511 | db->db_blkptr = &dn->dn_phys->dn_spill; | |
2512 | BP_ZERO(db->db_blkptr); | |
2513 | return; | |
2514 | } | |
34dc7c2f BB |
2515 | if (db->db_level == dn->dn_phys->dn_nlevels-1) { |
2516 | /* | |
2517 | * This buffer was allocated at a time when there was | |
2518 | * no available blkptrs from the dnode, or it was | |
2519 | * inappropriate to hook it in (i.e., nlevels mis-match). | |
2520 | */ | |
2521 | ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr); | |
2522 | ASSERT(db->db_parent == NULL); | |
2523 | db->db_parent = dn->dn_dbuf; | |
2524 | db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid]; | |
2525 | DBUF_VERIFY(db); | |
2526 | } else { | |
2527 | dmu_buf_impl_t *parent = db->db_parent; | |
2528 | int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; | |
2529 | ||
2530 | ASSERT(dn->dn_phys->dn_nlevels > 1); | |
2531 | if (parent == NULL) { | |
2532 | mutex_exit(&db->db_mtx); | |
2533 | rw_enter(&dn->dn_struct_rwlock, RW_READER); | |
2534 | (void) dbuf_hold_impl(dn, db->db_level+1, | |
2535 | db->db_blkid >> epbs, FALSE, db, &parent); | |
2536 | rw_exit(&dn->dn_struct_rwlock); | |
2537 | mutex_enter(&db->db_mtx); | |
2538 | db->db_parent = parent; | |
2539 | } | |
2540 | db->db_blkptr = (blkptr_t *)parent->db.db_data + | |
2541 | (db->db_blkid & ((1ULL << epbs) - 1)); | |
2542 | DBUF_VERIFY(db); | |
2543 | } | |
2544 | } | |
2545 | ||
d1d7e268 MK |
2546 | /* |
2547 | * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it | |
60948de1 BB |
2548 | * is critical the we not allow the compiler to inline this function in to |
2549 | * dbuf_sync_list() thereby drastically bloating the stack usage. | |
2550 | */ | |
2551 | noinline static void | |
34dc7c2f BB |
2552 | dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx) |
2553 | { | |
2554 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
572e2857 | 2555 | dnode_t *dn; |
34dc7c2f BB |
2556 | zio_t *zio; |
2557 | ||
2558 | ASSERT(dmu_tx_is_syncing(tx)); | |
2559 | ||
2560 | dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); | |
2561 | ||
2562 | mutex_enter(&db->db_mtx); | |
2563 | ||
2564 | ASSERT(db->db_level > 0); | |
2565 | DBUF_VERIFY(db); | |
2566 | ||
e49f1e20 | 2567 | /* Read the block if it hasn't been read yet. */ |
34dc7c2f BB |
2568 | if (db->db_buf == NULL) { |
2569 | mutex_exit(&db->db_mtx); | |
2570 | (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED); | |
2571 | mutex_enter(&db->db_mtx); | |
2572 | } | |
2573 | ASSERT3U(db->db_state, ==, DB_CACHED); | |
34dc7c2f BB |
2574 | ASSERT(db->db_buf != NULL); |
2575 | ||
572e2857 BB |
2576 | DB_DNODE_ENTER(db); |
2577 | dn = DB_DNODE(db); | |
e49f1e20 | 2578 | /* Indirect block size must match what the dnode thinks it is. */ |
572e2857 | 2579 | ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); |
34dc7c2f | 2580 | dbuf_check_blkptr(dn, db); |
572e2857 | 2581 | DB_DNODE_EXIT(db); |
34dc7c2f | 2582 | |
e49f1e20 | 2583 | /* Provide the pending dirty record to child dbufs */ |
34dc7c2f BB |
2584 | db->db_data_pending = dr; |
2585 | ||
34dc7c2f | 2586 | mutex_exit(&db->db_mtx); |
b128c09f | 2587 | dbuf_write(dr, db->db_buf, tx); |
34dc7c2f BB |
2588 | |
2589 | zio = dr->dr_zio; | |
2590 | mutex_enter(&dr->dt.di.dr_mtx); | |
4bda3bd0 | 2591 | dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx); |
34dc7c2f BB |
2592 | ASSERT(list_head(&dr->dt.di.dr_children) == NULL); |
2593 | mutex_exit(&dr->dt.di.dr_mtx); | |
2594 | zio_nowait(zio); | |
2595 | } | |
2596 | ||
d1d7e268 MK |
2597 | /* |
2598 | * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is | |
60948de1 BB |
2599 | * critical the we not allow the compiler to inline this function in to |
2600 | * dbuf_sync_list() thereby drastically bloating the stack usage. | |
2601 | */ | |
2602 | noinline static void | |
34dc7c2f BB |
2603 | dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx) |
2604 | { | |
2605 | arc_buf_t **datap = &dr->dt.dl.dr_data; | |
2606 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
572e2857 BB |
2607 | dnode_t *dn; |
2608 | objset_t *os; | |
34dc7c2f | 2609 | uint64_t txg = tx->tx_txg; |
34dc7c2f BB |
2610 | |
2611 | ASSERT(dmu_tx_is_syncing(tx)); | |
2612 | ||
2613 | dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); | |
2614 | ||
2615 | mutex_enter(&db->db_mtx); | |
2616 | /* | |
2617 | * To be synced, we must be dirtied. But we | |
2618 | * might have been freed after the dirty. | |
2619 | */ | |
2620 | if (db->db_state == DB_UNCACHED) { | |
2621 | /* This buffer has been freed since it was dirtied */ | |
2622 | ASSERT(db->db.db_data == NULL); | |
2623 | } else if (db->db_state == DB_FILL) { | |
2624 | /* This buffer was freed and is now being re-filled */ | |
2625 | ASSERT(db->db.db_data != dr->dt.dl.dr_data); | |
2626 | } else { | |
b128c09f | 2627 | ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL); |
34dc7c2f BB |
2628 | } |
2629 | DBUF_VERIFY(db); | |
2630 | ||
572e2857 BB |
2631 | DB_DNODE_ENTER(db); |
2632 | dn = DB_DNODE(db); | |
2633 | ||
428870ff BB |
2634 | if (db->db_blkid == DMU_SPILL_BLKID) { |
2635 | mutex_enter(&dn->dn_mtx); | |
2636 | dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR; | |
2637 | mutex_exit(&dn->dn_mtx); | |
2638 | } | |
2639 | ||
34dc7c2f BB |
2640 | /* |
2641 | * If this is a bonus buffer, simply copy the bonus data into the | |
2642 | * dnode. It will be written out when the dnode is synced (and it | |
2643 | * will be synced, since it must have been dirty for dbuf_sync to | |
2644 | * be called). | |
2645 | */ | |
428870ff | 2646 | if (db->db_blkid == DMU_BONUS_BLKID) { |
34dc7c2f BB |
2647 | dbuf_dirty_record_t **drp; |
2648 | ||
2649 | ASSERT(*datap != NULL); | |
c99c9001 | 2650 | ASSERT0(db->db_level); |
34dc7c2f BB |
2651 | ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN); |
2652 | bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen); | |
572e2857 BB |
2653 | DB_DNODE_EXIT(db); |
2654 | ||
34dc7c2f BB |
2655 | if (*datap != db->db.db_data) { |
2656 | zio_buf_free(*datap, DN_MAX_BONUSLEN); | |
d164b209 | 2657 | arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); |
34dc7c2f BB |
2658 | } |
2659 | db->db_data_pending = NULL; | |
2660 | drp = &db->db_last_dirty; | |
2661 | while (*drp != dr) | |
2662 | drp = &(*drp)->dr_next; | |
2663 | ASSERT(dr->dr_next == NULL); | |
428870ff | 2664 | ASSERT(dr->dr_dbuf == db); |
34dc7c2f | 2665 | *drp = dr->dr_next; |
753972fc BB |
2666 | if (dr->dr_dbuf->db_level != 0) { |
2667 | mutex_destroy(&dr->dt.di.dr_mtx); | |
2668 | list_destroy(&dr->dt.di.dr_children); | |
2669 | } | |
34dc7c2f BB |
2670 | kmem_free(dr, sizeof (dbuf_dirty_record_t)); |
2671 | ASSERT(db->db_dirtycnt > 0); | |
2672 | db->db_dirtycnt -= 1; | |
428870ff | 2673 | dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg); |
34dc7c2f BB |
2674 | return; |
2675 | } | |
2676 | ||
572e2857 BB |
2677 | os = dn->dn_objset; |
2678 | ||
34dc7c2f BB |
2679 | /* |
2680 | * This function may have dropped the db_mtx lock allowing a dmu_sync | |
2681 | * operation to sneak in. As a result, we need to ensure that we | |
2682 | * don't check the dr_override_state until we have returned from | |
2683 | * dbuf_check_blkptr. | |
2684 | */ | |
2685 | dbuf_check_blkptr(dn, db); | |
2686 | ||
2687 | /* | |
572e2857 | 2688 | * If this buffer is in the middle of an immediate write, |
34dc7c2f BB |
2689 | * wait for the synchronous IO to complete. |
2690 | */ | |
2691 | while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) { | |
2692 | ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT); | |
2693 | cv_wait(&db->db_changed, &db->db_mtx); | |
2694 | ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN); | |
2695 | } | |
2696 | ||
9babb374 BB |
2697 | if (db->db_state != DB_NOFILL && |
2698 | dn->dn_object != DMU_META_DNODE_OBJECT && | |
2699 | refcount_count(&db->db_holds) > 1 && | |
428870ff | 2700 | dr->dt.dl.dr_override_state != DR_OVERRIDDEN && |
9babb374 BB |
2701 | *datap == db->db_buf) { |
2702 | /* | |
2703 | * If this buffer is currently "in use" (i.e., there | |
2704 | * are active holds and db_data still references it), | |
2705 | * then make a copy before we start the write so that | |
2706 | * any modifications from the open txg will not leak | |
2707 | * into this write. | |
2708 | * | |
2709 | * NOTE: this copy does not need to be made for | |
2710 | * objects only modified in the syncing context (e.g. | |
2711 | * DNONE_DNODE blocks). | |
2712 | */ | |
2713 | int blksz = arc_buf_size(*datap); | |
2714 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); | |
2715 | *datap = arc_buf_alloc(os->os_spa, blksz, db, type); | |
2716 | bcopy(db->db.db_data, (*datap)->b_data, blksz); | |
b128c09f | 2717 | } |
34dc7c2f BB |
2718 | db->db_data_pending = dr; |
2719 | ||
2720 | mutex_exit(&db->db_mtx); | |
2721 | ||
b128c09f | 2722 | dbuf_write(dr, *datap, tx); |
34dc7c2f BB |
2723 | |
2724 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
572e2857 | 2725 | if (dn->dn_object == DMU_META_DNODE_OBJECT) { |
34dc7c2f | 2726 | list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr); |
572e2857 BB |
2727 | DB_DNODE_EXIT(db); |
2728 | } else { | |
2729 | /* | |
2730 | * Although zio_nowait() does not "wait for an IO", it does | |
2731 | * initiate the IO. If this is an empty write it seems plausible | |
2732 | * that the IO could actually be completed before the nowait | |
2733 | * returns. We need to DB_DNODE_EXIT() first in case | |
2734 | * zio_nowait() invalidates the dbuf. | |
2735 | */ | |
2736 | DB_DNODE_EXIT(db); | |
34dc7c2f | 2737 | zio_nowait(dr->dr_zio); |
572e2857 | 2738 | } |
34dc7c2f BB |
2739 | } |
2740 | ||
2741 | void | |
4bda3bd0 | 2742 | dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx) |
34dc7c2f BB |
2743 | { |
2744 | dbuf_dirty_record_t *dr; | |
2745 | ||
c65aa5b2 | 2746 | while ((dr = list_head(list))) { |
34dc7c2f BB |
2747 | if (dr->dr_zio != NULL) { |
2748 | /* | |
2749 | * If we find an already initialized zio then we | |
2750 | * are processing the meta-dnode, and we have finished. | |
2751 | * The dbufs for all dnodes are put back on the list | |
2752 | * during processing, so that we can zio_wait() | |
2753 | * these IOs after initiating all child IOs. | |
2754 | */ | |
2755 | ASSERT3U(dr->dr_dbuf->db.db_object, ==, | |
2756 | DMU_META_DNODE_OBJECT); | |
2757 | break; | |
2758 | } | |
4bda3bd0 MA |
2759 | if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID && |
2760 | dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) { | |
2761 | VERIFY3U(dr->dr_dbuf->db_level, ==, level); | |
2762 | } | |
34dc7c2f BB |
2763 | list_remove(list, dr); |
2764 | if (dr->dr_dbuf->db_level > 0) | |
2765 | dbuf_sync_indirect(dr, tx); | |
2766 | else | |
2767 | dbuf_sync_leaf(dr, tx); | |
2768 | } | |
2769 | } | |
2770 | ||
34dc7c2f BB |
2771 | /* ARGSUSED */ |
2772 | static void | |
2773 | dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb) | |
2774 | { | |
2775 | dmu_buf_impl_t *db = vdb; | |
572e2857 | 2776 | dnode_t *dn; |
b128c09f | 2777 | blkptr_t *bp = zio->io_bp; |
34dc7c2f | 2778 | blkptr_t *bp_orig = &zio->io_bp_orig; |
428870ff BB |
2779 | spa_t *spa = zio->io_spa; |
2780 | int64_t delta; | |
34dc7c2f | 2781 | uint64_t fill = 0; |
428870ff | 2782 | int i; |
34dc7c2f | 2783 | |
9b67f605 | 2784 | ASSERT3P(db->db_blkptr, ==, bp); |
b128c09f | 2785 | |
572e2857 BB |
2786 | DB_DNODE_ENTER(db); |
2787 | dn = DB_DNODE(db); | |
428870ff BB |
2788 | delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig); |
2789 | dnode_diduse_space(dn, delta - zio->io_prev_space_delta); | |
2790 | zio->io_prev_space_delta = delta; | |
34dc7c2f | 2791 | |
b0bc7a84 MG |
2792 | if (bp->blk_birth != 0) { |
2793 | ASSERT((db->db_blkid != DMU_SPILL_BLKID && | |
2794 | BP_GET_TYPE(bp) == dn->dn_type) || | |
2795 | (db->db_blkid == DMU_SPILL_BLKID && | |
9b67f605 MA |
2796 | BP_GET_TYPE(bp) == dn->dn_bonustype) || |
2797 | BP_IS_EMBEDDED(bp)); | |
b0bc7a84 | 2798 | ASSERT(BP_GET_LEVEL(bp) == db->db_level); |
34dc7c2f BB |
2799 | } |
2800 | ||
2801 | mutex_enter(&db->db_mtx); | |
2802 | ||
428870ff BB |
2803 | #ifdef ZFS_DEBUG |
2804 | if (db->db_blkid == DMU_SPILL_BLKID) { | |
428870ff BB |
2805 | ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); |
2806 | ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && | |
2807 | db->db_blkptr == &dn->dn_phys->dn_spill); | |
2808 | } | |
2809 | #endif | |
2810 | ||
34dc7c2f BB |
2811 | if (db->db_level == 0) { |
2812 | mutex_enter(&dn->dn_mtx); | |
428870ff BB |
2813 | if (db->db_blkid > dn->dn_phys->dn_maxblkid && |
2814 | db->db_blkid != DMU_SPILL_BLKID) | |
34dc7c2f BB |
2815 | dn->dn_phys->dn_maxblkid = db->db_blkid; |
2816 | mutex_exit(&dn->dn_mtx); | |
2817 | ||
2818 | if (dn->dn_type == DMU_OT_DNODE) { | |
2819 | dnode_phys_t *dnp = db->db.db_data; | |
2820 | for (i = db->db.db_size >> DNODE_SHIFT; i > 0; | |
2821 | i--, dnp++) { | |
2822 | if (dnp->dn_type != DMU_OT_NONE) | |
2823 | fill++; | |
2824 | } | |
2825 | } else { | |
b0bc7a84 MG |
2826 | if (BP_IS_HOLE(bp)) { |
2827 | fill = 0; | |
2828 | } else { | |
2829 | fill = 1; | |
2830 | } | |
34dc7c2f BB |
2831 | } |
2832 | } else { | |
b128c09f | 2833 | blkptr_t *ibp = db->db.db_data; |
34dc7c2f | 2834 | ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); |
b128c09f BB |
2835 | for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) { |
2836 | if (BP_IS_HOLE(ibp)) | |
34dc7c2f | 2837 | continue; |
9b67f605 | 2838 | fill += BP_GET_FILL(ibp); |
34dc7c2f BB |
2839 | } |
2840 | } | |
572e2857 | 2841 | DB_DNODE_EXIT(db); |
34dc7c2f | 2842 | |
9b67f605 MA |
2843 | if (!BP_IS_EMBEDDED(bp)) |
2844 | bp->blk_fill = fill; | |
34dc7c2f BB |
2845 | |
2846 | mutex_exit(&db->db_mtx); | |
34dc7c2f BB |
2847 | } |
2848 | ||
e8b96c60 MA |
2849 | /* |
2850 | * The SPA will call this callback several times for each zio - once | |
2851 | * for every physical child i/o (zio->io_phys_children times). This | |
2852 | * allows the DMU to monitor the progress of each logical i/o. For example, | |
2853 | * there may be 2 copies of an indirect block, or many fragments of a RAID-Z | |
2854 | * block. There may be a long delay before all copies/fragments are completed, | |
2855 | * so this callback allows us to retire dirty space gradually, as the physical | |
2856 | * i/os complete. | |
2857 | */ | |
2858 | /* ARGSUSED */ | |
2859 | static void | |
2860 | dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg) | |
2861 | { | |
2862 | dmu_buf_impl_t *db = arg; | |
2863 | objset_t *os = db->db_objset; | |
2864 | dsl_pool_t *dp = dmu_objset_pool(os); | |
2865 | dbuf_dirty_record_t *dr; | |
2866 | int delta = 0; | |
2867 | ||
2868 | dr = db->db_data_pending; | |
2869 | ASSERT3U(dr->dr_txg, ==, zio->io_txg); | |
2870 | ||
2871 | /* | |
2872 | * The callback will be called io_phys_children times. Retire one | |
2873 | * portion of our dirty space each time we are called. Any rounding | |
2874 | * error will be cleaned up by dsl_pool_sync()'s call to | |
2875 | * dsl_pool_undirty_space(). | |
2876 | */ | |
2877 | delta = dr->dr_accounted / zio->io_phys_children; | |
2878 | dsl_pool_undirty_space(dp, delta, zio->io_txg); | |
2879 | } | |
2880 | ||
34dc7c2f BB |
2881 | /* ARGSUSED */ |
2882 | static void | |
2883 | dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb) | |
2884 | { | |
2885 | dmu_buf_impl_t *db = vdb; | |
428870ff | 2886 | blkptr_t *bp_orig = &zio->io_bp_orig; |
b0bc7a84 MG |
2887 | blkptr_t *bp = db->db_blkptr; |
2888 | objset_t *os = db->db_objset; | |
2889 | dmu_tx_t *tx = os->os_synctx; | |
34dc7c2f BB |
2890 | dbuf_dirty_record_t **drp, *dr; |
2891 | ||
c99c9001 | 2892 | ASSERT0(zio->io_error); |
428870ff BB |
2893 | ASSERT(db->db_blkptr == bp); |
2894 | ||
03c6040b GW |
2895 | /* |
2896 | * For nopwrites and rewrites we ensure that the bp matches our | |
2897 | * original and bypass all the accounting. | |
2898 | */ | |
2899 | if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) { | |
428870ff BB |
2900 | ASSERT(BP_EQUAL(bp, bp_orig)); |
2901 | } else { | |
b0bc7a84 | 2902 | dsl_dataset_t *ds = os->os_dsl_dataset; |
428870ff BB |
2903 | (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE); |
2904 | dsl_dataset_block_born(ds, bp, tx); | |
2905 | } | |
34dc7c2f BB |
2906 | |
2907 | mutex_enter(&db->db_mtx); | |
2908 | ||
428870ff BB |
2909 | DBUF_VERIFY(db); |
2910 | ||
34dc7c2f BB |
2911 | drp = &db->db_last_dirty; |
2912 | while ((dr = *drp) != db->db_data_pending) | |
2913 | drp = &dr->dr_next; | |
2914 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
428870ff | 2915 | ASSERT(dr->dr_dbuf == db); |
34dc7c2f BB |
2916 | ASSERT(dr->dr_next == NULL); |
2917 | *drp = dr->dr_next; | |
2918 | ||
428870ff BB |
2919 | #ifdef ZFS_DEBUG |
2920 | if (db->db_blkid == DMU_SPILL_BLKID) { | |
572e2857 BB |
2921 | dnode_t *dn; |
2922 | ||
2923 | DB_DNODE_ENTER(db); | |
2924 | dn = DB_DNODE(db); | |
428870ff BB |
2925 | ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); |
2926 | ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && | |
2927 | db->db_blkptr == &dn->dn_phys->dn_spill); | |
572e2857 | 2928 | DB_DNODE_EXIT(db); |
428870ff BB |
2929 | } |
2930 | #endif | |
2931 | ||
34dc7c2f | 2932 | if (db->db_level == 0) { |
428870ff | 2933 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f | 2934 | ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); |
b128c09f BB |
2935 | if (db->db_state != DB_NOFILL) { |
2936 | if (dr->dt.dl.dr_data != db->db_buf) | |
2937 | VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, | |
13fe0198 | 2938 | db)); |
428870ff | 2939 | else if (!arc_released(db->db_buf)) |
b128c09f | 2940 | arc_set_callback(db->db_buf, dbuf_do_evict, db); |
b128c09f | 2941 | } |
34dc7c2f | 2942 | } else { |
572e2857 BB |
2943 | dnode_t *dn; |
2944 | ||
2945 | DB_DNODE_ENTER(db); | |
2946 | dn = DB_DNODE(db); | |
34dc7c2f | 2947 | ASSERT(list_head(&dr->dt.di.dr_children) == NULL); |
b0bc7a84 | 2948 | ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift); |
34dc7c2f | 2949 | if (!BP_IS_HOLE(db->db_blkptr)) { |
1fde1e37 BB |
2950 | ASSERTV(int epbs = dn->dn_phys->dn_indblkshift - |
2951 | SPA_BLKPTRSHIFT); | |
b0bc7a84 MG |
2952 | ASSERT3U(db->db_blkid, <=, |
2953 | dn->dn_phys->dn_maxblkid >> (db->db_level * epbs)); | |
34dc7c2f BB |
2954 | ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, |
2955 | db->db.db_size); | |
9b67f605 MA |
2956 | if (!arc_released(db->db_buf)) |
2957 | arc_set_callback(db->db_buf, dbuf_do_evict, db); | |
34dc7c2f | 2958 | } |
572e2857 | 2959 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
2960 | mutex_destroy(&dr->dt.di.dr_mtx); |
2961 | list_destroy(&dr->dt.di.dr_children); | |
2962 | } | |
2963 | kmem_free(dr, sizeof (dbuf_dirty_record_t)); | |
2964 | ||
2965 | cv_broadcast(&db->db_changed); | |
2966 | ASSERT(db->db_dirtycnt > 0); | |
2967 | db->db_dirtycnt -= 1; | |
2968 | db->db_data_pending = NULL; | |
b0bc7a84 | 2969 | dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg); |
428870ff BB |
2970 | } |
2971 | ||
2972 | static void | |
2973 | dbuf_write_nofill_ready(zio_t *zio) | |
2974 | { | |
2975 | dbuf_write_ready(zio, NULL, zio->io_private); | |
2976 | } | |
2977 | ||
2978 | static void | |
2979 | dbuf_write_nofill_done(zio_t *zio) | |
2980 | { | |
2981 | dbuf_write_done(zio, NULL, zio->io_private); | |
2982 | } | |
2983 | ||
2984 | static void | |
2985 | dbuf_write_override_ready(zio_t *zio) | |
2986 | { | |
2987 | dbuf_dirty_record_t *dr = zio->io_private; | |
2988 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
2989 | ||
2990 | dbuf_write_ready(zio, NULL, db); | |
2991 | } | |
2992 | ||
2993 | static void | |
2994 | dbuf_write_override_done(zio_t *zio) | |
2995 | { | |
2996 | dbuf_dirty_record_t *dr = zio->io_private; | |
2997 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
2998 | blkptr_t *obp = &dr->dt.dl.dr_overridden_by; | |
2999 | ||
3000 | mutex_enter(&db->db_mtx); | |
3001 | if (!BP_EQUAL(zio->io_bp, obp)) { | |
3002 | if (!BP_IS_HOLE(obp)) | |
3003 | dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp); | |
3004 | arc_release(dr->dt.dl.dr_data, db); | |
3005 | } | |
34dc7c2f BB |
3006 | mutex_exit(&db->db_mtx); |
3007 | ||
428870ff BB |
3008 | dbuf_write_done(zio, NULL, db); |
3009 | } | |
3010 | ||
e49f1e20 | 3011 | /* Issue I/O to commit a dirty buffer to disk. */ |
428870ff BB |
3012 | static void |
3013 | dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx) | |
3014 | { | |
3015 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
572e2857 BB |
3016 | dnode_t *dn; |
3017 | objset_t *os; | |
428870ff BB |
3018 | dmu_buf_impl_t *parent = db->db_parent; |
3019 | uint64_t txg = tx->tx_txg; | |
5dbd68a3 | 3020 | zbookmark_phys_t zb; |
428870ff BB |
3021 | zio_prop_t zp; |
3022 | zio_t *zio; | |
3023 | int wp_flag = 0; | |
34dc7c2f | 3024 | |
572e2857 BB |
3025 | DB_DNODE_ENTER(db); |
3026 | dn = DB_DNODE(db); | |
3027 | os = dn->dn_objset; | |
3028 | ||
428870ff BB |
3029 | if (db->db_state != DB_NOFILL) { |
3030 | if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) { | |
3031 | /* | |
3032 | * Private object buffers are released here rather | |
3033 | * than in dbuf_dirty() since they are only modified | |
3034 | * in the syncing context and we don't want the | |
3035 | * overhead of making multiple copies of the data. | |
3036 | */ | |
3037 | if (BP_IS_HOLE(db->db_blkptr)) { | |
3038 | arc_buf_thaw(data); | |
3039 | } else { | |
3040 | dbuf_release_bp(db); | |
3041 | } | |
3042 | } | |
3043 | } | |
3044 | ||
3045 | if (parent != dn->dn_dbuf) { | |
e49f1e20 WA |
3046 | /* Our parent is an indirect block. */ |
3047 | /* We have a dirty parent that has been scheduled for write. */ | |
428870ff | 3048 | ASSERT(parent && parent->db_data_pending); |
e49f1e20 | 3049 | /* Our parent's buffer is one level closer to the dnode. */ |
428870ff | 3050 | ASSERT(db->db_level == parent->db_level-1); |
e49f1e20 WA |
3051 | /* |
3052 | * We're about to modify our parent's db_data by modifying | |
3053 | * our block pointer, so the parent must be released. | |
3054 | */ | |
428870ff BB |
3055 | ASSERT(arc_released(parent->db_buf)); |
3056 | zio = parent->db_data_pending->dr_zio; | |
3057 | } else { | |
e49f1e20 | 3058 | /* Our parent is the dnode itself. */ |
428870ff BB |
3059 | ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 && |
3060 | db->db_blkid != DMU_SPILL_BLKID) || | |
3061 | (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0)); | |
3062 | if (db->db_blkid != DMU_SPILL_BLKID) | |
3063 | ASSERT3P(db->db_blkptr, ==, | |
3064 | &dn->dn_phys->dn_blkptr[db->db_blkid]); | |
3065 | zio = dn->dn_zio; | |
3066 | } | |
3067 | ||
3068 | ASSERT(db->db_level == 0 || data == db->db_buf); | |
3069 | ASSERT3U(db->db_blkptr->blk_birth, <=, txg); | |
3070 | ASSERT(zio); | |
3071 | ||
3072 | SET_BOOKMARK(&zb, os->os_dsl_dataset ? | |
3073 | os->os_dsl_dataset->ds_object : DMU_META_OBJSET, | |
3074 | db->db.db_object, db->db_level, db->db_blkid); | |
3075 | ||
3076 | if (db->db_blkid == DMU_SPILL_BLKID) | |
3077 | wp_flag = WP_SPILL; | |
3078 | wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0; | |
3079 | ||
3080 | dmu_write_policy(os, dn, db->db_level, wp_flag, &zp); | |
572e2857 | 3081 | DB_DNODE_EXIT(db); |
428870ff | 3082 | |
9b67f605 MA |
3083 | if (db->db_level == 0 && |
3084 | dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { | |
3085 | /* | |
3086 | * The BP for this block has been provided by open context | |
3087 | * (by dmu_sync() or dmu_buf_write_embedded()). | |
3088 | */ | |
3089 | void *contents = (data != NULL) ? data->b_data : NULL; | |
3090 | ||
428870ff | 3091 | dr->dr_zio = zio_write(zio, os->os_spa, txg, |
9b67f605 | 3092 | db->db_blkptr, contents, db->db.db_size, &zp, |
e8b96c60 MA |
3093 | dbuf_write_override_ready, NULL, dbuf_write_override_done, |
3094 | dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); | |
428870ff BB |
3095 | mutex_enter(&db->db_mtx); |
3096 | dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; | |
3097 | zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by, | |
03c6040b | 3098 | dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite); |
428870ff BB |
3099 | mutex_exit(&db->db_mtx); |
3100 | } else if (db->db_state == DB_NOFILL) { | |
3101 | ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF); | |
3102 | dr->dr_zio = zio_write(zio, os->os_spa, txg, | |
3103 | db->db_blkptr, NULL, db->db.db_size, &zp, | |
e8b96c60 | 3104 | dbuf_write_nofill_ready, NULL, dbuf_write_nofill_done, db, |
428870ff BB |
3105 | ZIO_PRIORITY_ASYNC_WRITE, |
3106 | ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb); | |
3107 | } else { | |
3108 | ASSERT(arc_released(data)); | |
3109 | dr->dr_zio = arc_write(zio, os->os_spa, txg, | |
3a17a7a9 SK |
3110 | db->db_blkptr, data, DBUF_IS_L2CACHEABLE(db), |
3111 | DBUF_IS_L2COMPRESSIBLE(db), &zp, dbuf_write_ready, | |
e8b96c60 MA |
3112 | dbuf_write_physdone, dbuf_write_done, db, |
3113 | ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); | |
428870ff | 3114 | } |
34dc7c2f | 3115 | } |
c28b2279 BB |
3116 | |
3117 | #if defined(_KERNEL) && defined(HAVE_SPL) | |
8f576c23 BB |
3118 | EXPORT_SYMBOL(dbuf_find); |
3119 | EXPORT_SYMBOL(dbuf_is_metadata); | |
3120 | EXPORT_SYMBOL(dbuf_evict); | |
3121 | EXPORT_SYMBOL(dbuf_loan_arcbuf); | |
3122 | EXPORT_SYMBOL(dbuf_whichblock); | |
3123 | EXPORT_SYMBOL(dbuf_read); | |
3124 | EXPORT_SYMBOL(dbuf_unoverride); | |
3125 | EXPORT_SYMBOL(dbuf_free_range); | |
3126 | EXPORT_SYMBOL(dbuf_new_size); | |
3127 | EXPORT_SYMBOL(dbuf_release_bp); | |
3128 | EXPORT_SYMBOL(dbuf_dirty); | |
c28b2279 | 3129 | EXPORT_SYMBOL(dmu_buf_will_dirty); |
8f576c23 BB |
3130 | EXPORT_SYMBOL(dmu_buf_will_not_fill); |
3131 | EXPORT_SYMBOL(dmu_buf_will_fill); | |
3132 | EXPORT_SYMBOL(dmu_buf_fill_done); | |
4047414a | 3133 | EXPORT_SYMBOL(dmu_buf_rele); |
8f576c23 BB |
3134 | EXPORT_SYMBOL(dbuf_assign_arcbuf); |
3135 | EXPORT_SYMBOL(dbuf_clear); | |
3136 | EXPORT_SYMBOL(dbuf_prefetch); | |
3137 | EXPORT_SYMBOL(dbuf_hold_impl); | |
3138 | EXPORT_SYMBOL(dbuf_hold); | |
3139 | EXPORT_SYMBOL(dbuf_hold_level); | |
3140 | EXPORT_SYMBOL(dbuf_create_bonus); | |
3141 | EXPORT_SYMBOL(dbuf_spill_set_blksz); | |
3142 | EXPORT_SYMBOL(dbuf_rm_spill); | |
3143 | EXPORT_SYMBOL(dbuf_add_ref); | |
3144 | EXPORT_SYMBOL(dbuf_rele); | |
3145 | EXPORT_SYMBOL(dbuf_rele_and_unlock); | |
3146 | EXPORT_SYMBOL(dbuf_refcount); | |
3147 | EXPORT_SYMBOL(dbuf_sync_list); | |
3148 | EXPORT_SYMBOL(dmu_buf_set_user); | |
3149 | EXPORT_SYMBOL(dmu_buf_set_user_ie); | |
8f576c23 BB |
3150 | EXPORT_SYMBOL(dmu_buf_get_user); |
3151 | EXPORT_SYMBOL(dmu_buf_freeable); | |
0f699108 | 3152 | EXPORT_SYMBOL(dmu_buf_get_blkptr); |
c28b2279 | 3153 | #endif |