]>
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. |
34dc7c2f BB |
23 | */ |
24 | ||
34dc7c2f | 25 | #include <sys/zfs_context.h> |
c28b2279 | 26 | #include <sys/arc.h> |
34dc7c2f BB |
27 | #include <sys/dmu.h> |
28 | #include <sys/dmu_impl.h> | |
29 | #include <sys/dbuf.h> | |
30 | #include <sys/dmu_objset.h> | |
31 | #include <sys/dsl_dataset.h> | |
32 | #include <sys/dsl_dir.h> | |
33 | #include <sys/dmu_tx.h> | |
34 | #include <sys/spa.h> | |
35 | #include <sys/zio.h> | |
36 | #include <sys/dmu_zfetch.h> | |
428870ff BB |
37 | #include <sys/sa.h> |
38 | #include <sys/sa_impl.h> | |
34dc7c2f | 39 | |
fc5bb51f BB |
40 | struct dbuf_hold_impl_data { |
41 | /* Function arguments */ | |
42 | dnode_t *dh_dn; | |
43 | uint8_t dh_level; | |
44 | uint64_t dh_blkid; | |
45 | int dh_fail_sparse; | |
46 | void *dh_tag; | |
47 | dmu_buf_impl_t **dh_dbp; | |
48 | /* Local variables */ | |
49 | dmu_buf_impl_t *dh_db; | |
50 | dmu_buf_impl_t *dh_parent; | |
51 | blkptr_t *dh_bp; | |
52 | int dh_err; | |
53 | dbuf_dirty_record_t *dh_dr; | |
54 | arc_buf_contents_t dh_type; | |
55 | int dh_depth; | |
56 | }; | |
57 | ||
58 | static void __dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh, | |
59 | dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse, | |
60 | void *tag, dmu_buf_impl_t **dbp, int depth); | |
61 | static int __dbuf_hold_impl(struct dbuf_hold_impl_data *dh); | |
62 | ||
34dc7c2f BB |
63 | static void dbuf_destroy(dmu_buf_impl_t *db); |
64 | static int dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx); | |
b128c09f | 65 | static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx); |
34dc7c2f BB |
66 | |
67 | /* | |
68 | * Global data structures and functions for the dbuf cache. | |
69 | */ | |
70 | static kmem_cache_t *dbuf_cache; | |
71 | ||
72 | /* ARGSUSED */ | |
73 | static int | |
74 | dbuf_cons(void *vdb, void *unused, int kmflag) | |
75 | { | |
76 | dmu_buf_impl_t *db = vdb; | |
77 | bzero(db, sizeof (dmu_buf_impl_t)); | |
78 | ||
79 | mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL); | |
80 | cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL); | |
81 | refcount_create(&db->db_holds); | |
98f72a53 | 82 | list_link_init(&db->db_link); |
34dc7c2f BB |
83 | return (0); |
84 | } | |
85 | ||
86 | /* ARGSUSED */ | |
87 | static void | |
88 | dbuf_dest(void *vdb, void *unused) | |
89 | { | |
90 | dmu_buf_impl_t *db = vdb; | |
91 | mutex_destroy(&db->db_mtx); | |
92 | cv_destroy(&db->db_changed); | |
93 | refcount_destroy(&db->db_holds); | |
94 | } | |
95 | ||
96 | /* | |
97 | * dbuf hash table routines | |
98 | */ | |
99 | static dbuf_hash_table_t dbuf_hash_table; | |
100 | ||
101 | static uint64_t dbuf_hash_count; | |
102 | ||
103 | static uint64_t | |
104 | dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid) | |
105 | { | |
106 | uintptr_t osv = (uintptr_t)os; | |
107 | uint64_t crc = -1ULL; | |
108 | ||
109 | ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); | |
110 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF]; | |
111 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF]; | |
112 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF]; | |
113 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF]; | |
114 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF]; | |
115 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF]; | |
116 | ||
117 | crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16); | |
118 | ||
119 | return (crc); | |
120 | } | |
121 | ||
122 | #define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid); | |
123 | ||
124 | #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \ | |
125 | ((dbuf)->db.db_object == (obj) && \ | |
126 | (dbuf)->db_objset == (os) && \ | |
127 | (dbuf)->db_level == (level) && \ | |
128 | (dbuf)->db_blkid == (blkid)) | |
129 | ||
130 | dmu_buf_impl_t * | |
131 | dbuf_find(dnode_t *dn, uint8_t level, uint64_t blkid) | |
132 | { | |
133 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
428870ff | 134 | objset_t *os = dn->dn_objset; |
d6320ddb BB |
135 | uint64_t obj; |
136 | uint64_t hv; | |
137 | uint64_t idx; | |
34dc7c2f BB |
138 | dmu_buf_impl_t *db; |
139 | ||
d6320ddb BB |
140 | obj = dn->dn_object; |
141 | hv = DBUF_HASH(os, obj, level, blkid); | |
142 | idx = hv & h->hash_table_mask; | |
143 | ||
34dc7c2f BB |
144 | mutex_enter(DBUF_HASH_MUTEX(h, idx)); |
145 | for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) { | |
146 | if (DBUF_EQUAL(db, os, obj, level, blkid)) { | |
147 | mutex_enter(&db->db_mtx); | |
148 | if (db->db_state != DB_EVICTING) { | |
149 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); | |
150 | return (db); | |
151 | } | |
152 | mutex_exit(&db->db_mtx); | |
153 | } | |
154 | } | |
155 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); | |
156 | return (NULL); | |
157 | } | |
158 | ||
159 | /* | |
160 | * Insert an entry into the hash table. If there is already an element | |
161 | * equal to elem in the hash table, then the already existing element | |
162 | * will be returned and the new element will not be inserted. | |
163 | * Otherwise returns NULL. | |
164 | */ | |
165 | static dmu_buf_impl_t * | |
166 | dbuf_hash_insert(dmu_buf_impl_t *db) | |
167 | { | |
168 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
428870ff | 169 | objset_t *os = db->db_objset; |
34dc7c2f BB |
170 | uint64_t obj = db->db.db_object; |
171 | int level = db->db_level; | |
d6320ddb | 172 | uint64_t blkid, hv, idx; |
34dc7c2f BB |
173 | dmu_buf_impl_t *dbf; |
174 | ||
d6320ddb BB |
175 | blkid = db->db_blkid; |
176 | hv = DBUF_HASH(os, obj, level, blkid); | |
177 | idx = hv & h->hash_table_mask; | |
178 | ||
34dc7c2f BB |
179 | mutex_enter(DBUF_HASH_MUTEX(h, idx)); |
180 | for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) { | |
181 | if (DBUF_EQUAL(dbf, os, obj, level, blkid)) { | |
182 | mutex_enter(&dbf->db_mtx); | |
183 | if (dbf->db_state != DB_EVICTING) { | |
184 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); | |
185 | return (dbf); | |
186 | } | |
187 | mutex_exit(&dbf->db_mtx); | |
188 | } | |
189 | } | |
190 | ||
191 | mutex_enter(&db->db_mtx); | |
192 | db->db_hash_next = h->hash_table[idx]; | |
193 | h->hash_table[idx] = db; | |
194 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); | |
195 | atomic_add_64(&dbuf_hash_count, 1); | |
196 | ||
197 | return (NULL); | |
198 | } | |
199 | ||
200 | /* | |
201 | * Remove an entry from the hash table. This operation will | |
202 | * fail if there are any existing holds on the db. | |
203 | */ | |
204 | static void | |
205 | dbuf_hash_remove(dmu_buf_impl_t *db) | |
206 | { | |
207 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
d6320ddb | 208 | uint64_t hv, idx; |
34dc7c2f BB |
209 | dmu_buf_impl_t *dbf, **dbp; |
210 | ||
d6320ddb BB |
211 | hv = DBUF_HASH(db->db_objset, db->db.db_object, |
212 | db->db_level, db->db_blkid); | |
213 | idx = hv & h->hash_table_mask; | |
214 | ||
34dc7c2f BB |
215 | /* |
216 | * We musn't hold db_mtx to maintin lock ordering: | |
217 | * DBUF_HASH_MUTEX > db_mtx. | |
218 | */ | |
219 | ASSERT(refcount_is_zero(&db->db_holds)); | |
220 | ASSERT(db->db_state == DB_EVICTING); | |
221 | ASSERT(!MUTEX_HELD(&db->db_mtx)); | |
222 | ||
223 | mutex_enter(DBUF_HASH_MUTEX(h, idx)); | |
224 | dbp = &h->hash_table[idx]; | |
225 | while ((dbf = *dbp) != db) { | |
226 | dbp = &dbf->db_hash_next; | |
227 | ASSERT(dbf != NULL); | |
228 | } | |
229 | *dbp = db->db_hash_next; | |
230 | db->db_hash_next = NULL; | |
231 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); | |
232 | atomic_add_64(&dbuf_hash_count, -1); | |
233 | } | |
234 | ||
235 | static arc_evict_func_t dbuf_do_evict; | |
236 | ||
237 | static void | |
238 | dbuf_evict_user(dmu_buf_impl_t *db) | |
239 | { | |
240 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
241 | ||
242 | if (db->db_level != 0 || db->db_evict_func == NULL) | |
243 | return; | |
244 | ||
245 | if (db->db_user_data_ptr_ptr) | |
246 | *db->db_user_data_ptr_ptr = db->db.db_data; | |
247 | db->db_evict_func(&db->db, db->db_user_ptr); | |
248 | db->db_user_ptr = NULL; | |
249 | db->db_user_data_ptr_ptr = NULL; | |
250 | db->db_evict_func = NULL; | |
251 | } | |
252 | ||
572e2857 BB |
253 | boolean_t |
254 | dbuf_is_metadata(dmu_buf_impl_t *db) | |
255 | { | |
256 | if (db->db_level > 0) { | |
257 | return (B_TRUE); | |
258 | } else { | |
259 | boolean_t is_metadata; | |
260 | ||
261 | DB_DNODE_ENTER(db); | |
262 | is_metadata = dmu_ot[DB_DNODE(db)->dn_type].ot_metadata; | |
263 | DB_DNODE_EXIT(db); | |
264 | ||
265 | return (is_metadata); | |
266 | } | |
267 | } | |
268 | ||
34dc7c2f BB |
269 | void |
270 | dbuf_evict(dmu_buf_impl_t *db) | |
271 | { | |
272 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
273 | ASSERT(db->db_buf == NULL); | |
274 | ASSERT(db->db_data_pending == NULL); | |
275 | ||
276 | dbuf_clear(db); | |
277 | dbuf_destroy(db); | |
278 | } | |
279 | ||
280 | void | |
281 | dbuf_init(void) | |
282 | { | |
283 | uint64_t hsize = 1ULL << 16; | |
284 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
285 | int i; | |
286 | ||
287 | /* | |
288 | * The hash table is big enough to fill all of physical memory | |
289 | * with an average 4K block size. The table will take up | |
290 | * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers). | |
291 | */ | |
292 | while (hsize * 4096 < physmem * PAGESIZE) | |
293 | hsize <<= 1; | |
294 | ||
295 | retry: | |
296 | h->hash_table_mask = hsize - 1; | |
00b46022 BB |
297 | #if defined(_KERNEL) && defined(HAVE_SPL) |
298 | /* Large allocations which do not require contiguous pages | |
299 | * should be using vmem_alloc() in the linux kernel */ | |
300 | h->hash_table = vmem_zalloc(hsize * sizeof (void *), KM_SLEEP); | |
301 | #else | |
34dc7c2f | 302 | h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP); |
00b46022 | 303 | #endif |
34dc7c2f BB |
304 | if (h->hash_table == NULL) { |
305 | /* XXX - we should really return an error instead of assert */ | |
306 | ASSERT(hsize > (1ULL << 10)); | |
307 | hsize >>= 1; | |
308 | goto retry; | |
309 | } | |
310 | ||
311 | dbuf_cache = kmem_cache_create("dmu_buf_impl_t", | |
312 | sizeof (dmu_buf_impl_t), | |
313 | 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0); | |
314 | ||
315 | for (i = 0; i < DBUF_MUTEXES; i++) | |
316 | mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL); | |
317 | } | |
318 | ||
319 | void | |
320 | dbuf_fini(void) | |
321 | { | |
322 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
323 | int i; | |
324 | ||
325 | for (i = 0; i < DBUF_MUTEXES; i++) | |
326 | mutex_destroy(&h->hash_mutexes[i]); | |
00b46022 BB |
327 | #if defined(_KERNEL) && defined(HAVE_SPL) |
328 | /* Large allocations which do not require contiguous pages | |
329 | * should be using vmem_free() in the linux kernel */ | |
330 | vmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *)); | |
331 | #else | |
34dc7c2f | 332 | kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *)); |
00b46022 | 333 | #endif |
34dc7c2f BB |
334 | kmem_cache_destroy(dbuf_cache); |
335 | } | |
336 | ||
337 | /* | |
338 | * Other stuff. | |
339 | */ | |
340 | ||
341 | #ifdef ZFS_DEBUG | |
342 | static void | |
343 | dbuf_verify(dmu_buf_impl_t *db) | |
344 | { | |
572e2857 | 345 | dnode_t *dn; |
428870ff | 346 | dbuf_dirty_record_t *dr; |
34dc7c2f BB |
347 | |
348 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
349 | ||
350 | if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY)) | |
351 | return; | |
352 | ||
353 | ASSERT(db->db_objset != NULL); | |
572e2857 BB |
354 | DB_DNODE_ENTER(db); |
355 | dn = DB_DNODE(db); | |
34dc7c2f BB |
356 | if (dn == NULL) { |
357 | ASSERT(db->db_parent == NULL); | |
358 | ASSERT(db->db_blkptr == NULL); | |
359 | } else { | |
360 | ASSERT3U(db->db.db_object, ==, dn->dn_object); | |
361 | ASSERT3P(db->db_objset, ==, dn->dn_objset); | |
362 | ASSERT3U(db->db_level, <, dn->dn_nlevels); | |
572e2857 BB |
363 | ASSERT(db->db_blkid == DMU_BONUS_BLKID || |
364 | db->db_blkid == DMU_SPILL_BLKID || | |
365 | !list_is_empty(&dn->dn_dbufs)); | |
34dc7c2f | 366 | } |
428870ff BB |
367 | if (db->db_blkid == DMU_BONUS_BLKID) { |
368 | ASSERT(dn != NULL); | |
369 | ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); | |
370 | ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID); | |
371 | } else if (db->db_blkid == DMU_SPILL_BLKID) { | |
34dc7c2f BB |
372 | ASSERT(dn != NULL); |
373 | ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); | |
428870ff | 374 | ASSERT3U(db->db.db_offset, ==, 0); |
34dc7c2f BB |
375 | } else { |
376 | ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size); | |
377 | } | |
378 | ||
428870ff BB |
379 | for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next) |
380 | ASSERT(dr->dr_dbuf == db); | |
381 | ||
382 | for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next) | |
383 | ASSERT(dr->dr_dbuf == db); | |
384 | ||
b128c09f BB |
385 | /* |
386 | * We can't assert that db_size matches dn_datablksz because it | |
387 | * can be momentarily different when another thread is doing | |
388 | * dnode_set_blksz(). | |
389 | */ | |
390 | if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) { | |
428870ff | 391 | dr = db->db_data_pending; |
b128c09f BB |
392 | /* |
393 | * It should only be modified in syncing context, so | |
394 | * make sure we only have one copy of the data. | |
395 | */ | |
396 | ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf); | |
34dc7c2f BB |
397 | } |
398 | ||
399 | /* verify db->db_blkptr */ | |
400 | if (db->db_blkptr) { | |
401 | if (db->db_parent == dn->dn_dbuf) { | |
402 | /* db is pointed to by the dnode */ | |
403 | /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */ | |
9babb374 | 404 | if (DMU_OBJECT_IS_SPECIAL(db->db.db_object)) |
34dc7c2f BB |
405 | ASSERT(db->db_parent == NULL); |
406 | else | |
407 | ASSERT(db->db_parent != NULL); | |
428870ff BB |
408 | if (db->db_blkid != DMU_SPILL_BLKID) |
409 | ASSERT3P(db->db_blkptr, ==, | |
410 | &dn->dn_phys->dn_blkptr[db->db_blkid]); | |
34dc7c2f BB |
411 | } else { |
412 | /* db is pointed to by an indirect block */ | |
1fde1e37 BB |
413 | ASSERTV(int epb = db->db_parent->db.db_size >> |
414 | SPA_BLKPTRSHIFT); | |
34dc7c2f BB |
415 | ASSERT3U(db->db_parent->db_level, ==, db->db_level+1); |
416 | ASSERT3U(db->db_parent->db.db_object, ==, | |
417 | db->db.db_object); | |
418 | /* | |
419 | * dnode_grow_indblksz() can make this fail if we don't | |
420 | * have the struct_rwlock. XXX indblksz no longer | |
421 | * grows. safe to do this now? | |
422 | */ | |
572e2857 | 423 | if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) { |
34dc7c2f BB |
424 | ASSERT3P(db->db_blkptr, ==, |
425 | ((blkptr_t *)db->db_parent->db.db_data + | |
426 | db->db_blkid % epb)); | |
427 | } | |
428 | } | |
429 | } | |
430 | if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) && | |
428870ff BB |
431 | (db->db_buf == NULL || db->db_buf->b_data) && |
432 | db->db.db_data && db->db_blkid != DMU_BONUS_BLKID && | |
34dc7c2f BB |
433 | db->db_state != DB_FILL && !dn->dn_free_txg) { |
434 | /* | |
435 | * If the blkptr isn't set but they have nonzero data, | |
436 | * it had better be dirty, otherwise we'll lose that | |
437 | * data when we evict this buffer. | |
438 | */ | |
439 | if (db->db_dirtycnt == 0) { | |
1fde1e37 | 440 | ASSERTV(uint64_t *buf = db->db.db_data); |
34dc7c2f BB |
441 | int i; |
442 | ||
443 | for (i = 0; i < db->db.db_size >> 3; i++) { | |
444 | ASSERT(buf[i] == 0); | |
445 | } | |
446 | } | |
447 | } | |
572e2857 | 448 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
449 | } |
450 | #endif | |
451 | ||
452 | static void | |
453 | dbuf_update_data(dmu_buf_impl_t *db) | |
454 | { | |
455 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
456 | if (db->db_level == 0 && db->db_user_data_ptr_ptr) { | |
457 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
458 | *db->db_user_data_ptr_ptr = db->db.db_data; | |
459 | } | |
460 | } | |
461 | ||
462 | static void | |
463 | dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf) | |
464 | { | |
465 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
466 | ASSERT(db->db_buf == NULL || !arc_has_callback(db->db_buf)); | |
467 | db->db_buf = buf; | |
468 | if (buf != NULL) { | |
469 | ASSERT(buf->b_data != NULL); | |
470 | db->db.db_data = buf->b_data; | |
471 | if (!arc_released(buf)) | |
472 | arc_set_callback(buf, dbuf_do_evict, db); | |
473 | dbuf_update_data(db); | |
474 | } else { | |
475 | dbuf_evict_user(db); | |
476 | db->db.db_data = NULL; | |
b128c09f BB |
477 | if (db->db_state != DB_NOFILL) |
478 | db->db_state = DB_UNCACHED; | |
34dc7c2f BB |
479 | } |
480 | } | |
481 | ||
428870ff BB |
482 | /* |
483 | * Loan out an arc_buf for read. Return the loaned arc_buf. | |
484 | */ | |
485 | arc_buf_t * | |
486 | dbuf_loan_arcbuf(dmu_buf_impl_t *db) | |
487 | { | |
488 | arc_buf_t *abuf; | |
489 | ||
490 | mutex_enter(&db->db_mtx); | |
491 | if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) { | |
492 | int blksz = db->db.db_size; | |
572e2857 BB |
493 | spa_t *spa; |
494 | ||
428870ff | 495 | mutex_exit(&db->db_mtx); |
572e2857 BB |
496 | DB_GET_SPA(&spa, db); |
497 | abuf = arc_loan_buf(spa, blksz); | |
428870ff BB |
498 | bcopy(db->db.db_data, abuf->b_data, blksz); |
499 | } else { | |
500 | abuf = db->db_buf; | |
501 | arc_loan_inuse_buf(abuf, db); | |
502 | dbuf_set_data(db, NULL); | |
503 | mutex_exit(&db->db_mtx); | |
504 | } | |
505 | return (abuf); | |
506 | } | |
507 | ||
34dc7c2f BB |
508 | uint64_t |
509 | dbuf_whichblock(dnode_t *dn, uint64_t offset) | |
510 | { | |
511 | if (dn->dn_datablkshift) { | |
512 | return (offset >> dn->dn_datablkshift); | |
513 | } else { | |
514 | ASSERT3U(offset, <, dn->dn_datablksz); | |
515 | return (0); | |
516 | } | |
517 | } | |
518 | ||
519 | static void | |
520 | dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb) | |
521 | { | |
522 | dmu_buf_impl_t *db = vdb; | |
523 | ||
524 | mutex_enter(&db->db_mtx); | |
525 | ASSERT3U(db->db_state, ==, DB_READ); | |
526 | /* | |
527 | * All reads are synchronous, so we must have a hold on the dbuf | |
528 | */ | |
529 | ASSERT(refcount_count(&db->db_holds) > 0); | |
530 | ASSERT(db->db_buf == NULL); | |
531 | ASSERT(db->db.db_data == NULL); | |
532 | if (db->db_level == 0 && db->db_freed_in_flight) { | |
533 | /* we were freed in flight; disregard any error */ | |
534 | arc_release(buf, db); | |
535 | bzero(buf->b_data, db->db.db_size); | |
536 | arc_buf_freeze(buf); | |
537 | db->db_freed_in_flight = FALSE; | |
538 | dbuf_set_data(db, buf); | |
539 | db->db_state = DB_CACHED; | |
540 | } else if (zio == NULL || zio->io_error == 0) { | |
541 | dbuf_set_data(db, buf); | |
542 | db->db_state = DB_CACHED; | |
543 | } else { | |
428870ff | 544 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f BB |
545 | ASSERT3P(db->db_buf, ==, NULL); |
546 | VERIFY(arc_buf_remove_ref(buf, db) == 1); | |
547 | db->db_state = DB_UNCACHED; | |
548 | } | |
549 | cv_broadcast(&db->db_changed); | |
428870ff | 550 | dbuf_rele_and_unlock(db, NULL); |
34dc7c2f BB |
551 | } |
552 | ||
553 | static void | |
554 | dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t *flags) | |
555 | { | |
572e2857 BB |
556 | dnode_t *dn; |
557 | spa_t *spa; | |
34dc7c2f BB |
558 | zbookmark_t zb; |
559 | uint32_t aflags = ARC_NOWAIT; | |
b128c09f | 560 | arc_buf_t *pbuf; |
34dc7c2f | 561 | |
572e2857 BB |
562 | DB_DNODE_ENTER(db); |
563 | dn = DB_DNODE(db); | |
34dc7c2f BB |
564 | ASSERT(!refcount_is_zero(&db->db_holds)); |
565 | /* We need the struct_rwlock to prevent db_blkptr from changing. */ | |
b128c09f | 566 | ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); |
34dc7c2f BB |
567 | ASSERT(MUTEX_HELD(&db->db_mtx)); |
568 | ASSERT(db->db_state == DB_UNCACHED); | |
569 | ASSERT(db->db_buf == NULL); | |
570 | ||
428870ff | 571 | if (db->db_blkid == DMU_BONUS_BLKID) { |
9babb374 | 572 | int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen); |
34dc7c2f BB |
573 | |
574 | ASSERT3U(bonuslen, <=, db->db.db_size); | |
575 | db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN); | |
d164b209 | 576 | arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); |
34dc7c2f BB |
577 | if (bonuslen < DN_MAX_BONUSLEN) |
578 | bzero(db->db.db_data, DN_MAX_BONUSLEN); | |
9babb374 BB |
579 | if (bonuslen) |
580 | bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen); | |
572e2857 | 581 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
582 | dbuf_update_data(db); |
583 | db->db_state = DB_CACHED; | |
584 | mutex_exit(&db->db_mtx); | |
585 | return; | |
586 | } | |
587 | ||
b128c09f BB |
588 | /* |
589 | * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync() | |
590 | * processes the delete record and clears the bp while we are waiting | |
591 | * for the dn_mtx (resulting in a "no" from block_freed). | |
592 | */ | |
593 | if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) || | |
594 | (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) || | |
595 | BP_IS_HOLE(db->db_blkptr)))) { | |
34dc7c2f BB |
596 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); |
597 | ||
b128c09f | 598 | dbuf_set_data(db, arc_buf_alloc(dn->dn_objset->os_spa, |
34dc7c2f | 599 | db->db.db_size, db, type)); |
572e2857 | 600 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
601 | bzero(db->db.db_data, db->db.db_size); |
602 | db->db_state = DB_CACHED; | |
603 | *flags |= DB_RF_CACHED; | |
604 | mutex_exit(&db->db_mtx); | |
605 | return; | |
606 | } | |
607 | ||
572e2857 BB |
608 | spa = dn->dn_objset->os_spa; |
609 | DB_DNODE_EXIT(db); | |
610 | ||
34dc7c2f BB |
611 | db->db_state = DB_READ; |
612 | mutex_exit(&db->db_mtx); | |
613 | ||
b128c09f BB |
614 | if (DBUF_IS_L2CACHEABLE(db)) |
615 | aflags |= ARC_L2CACHE; | |
616 | ||
428870ff BB |
617 | SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ? |
618 | db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET, | |
619 | db->db.db_object, db->db_level, db->db_blkid); | |
34dc7c2f BB |
620 | |
621 | dbuf_add_ref(db, NULL); | |
622 | /* ZIO_FLAG_CANFAIL callers have to check the parent zio's error */ | |
b128c09f BB |
623 | |
624 | if (db->db_parent) | |
625 | pbuf = db->db_parent->db_buf; | |
626 | else | |
627 | pbuf = db->db_objset->os_phys_buf; | |
628 | ||
572e2857 | 629 | (void) dsl_read(zio, spa, db->db_blkptr, pbuf, |
34dc7c2f BB |
630 | dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ, |
631 | (*flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED, | |
632 | &aflags, &zb); | |
633 | if (aflags & ARC_CACHED) | |
634 | *flags |= DB_RF_CACHED; | |
635 | } | |
636 | ||
637 | int | |
638 | dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) | |
639 | { | |
640 | int err = 0; | |
641 | int havepzio = (zio != NULL); | |
642 | int prefetch; | |
572e2857 | 643 | dnode_t *dn; |
34dc7c2f BB |
644 | |
645 | /* | |
646 | * We don't have to hold the mutex to check db_state because it | |
647 | * can't be freed while we have a hold on the buffer. | |
648 | */ | |
649 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
650 | ||
b128c09f BB |
651 | if (db->db_state == DB_NOFILL) |
652 | return (EIO); | |
653 | ||
572e2857 BB |
654 | DB_DNODE_ENTER(db); |
655 | dn = DB_DNODE(db); | |
34dc7c2f | 656 | if ((flags & DB_RF_HAVESTRUCT) == 0) |
572e2857 | 657 | rw_enter(&dn->dn_struct_rwlock, RW_READER); |
34dc7c2f | 658 | |
428870ff | 659 | prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && |
572e2857 | 660 | (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL && |
b128c09f | 661 | DBUF_IS_CACHEABLE(db); |
34dc7c2f BB |
662 | |
663 | mutex_enter(&db->db_mtx); | |
664 | if (db->db_state == DB_CACHED) { | |
665 | mutex_exit(&db->db_mtx); | |
666 | if (prefetch) | |
572e2857 | 667 | dmu_zfetch(&dn->dn_zfetch, db->db.db_offset, |
34dc7c2f BB |
668 | db->db.db_size, TRUE); |
669 | if ((flags & DB_RF_HAVESTRUCT) == 0) | |
572e2857 BB |
670 | rw_exit(&dn->dn_struct_rwlock); |
671 | DB_DNODE_EXIT(db); | |
34dc7c2f | 672 | } else if (db->db_state == DB_UNCACHED) { |
572e2857 BB |
673 | spa_t *spa = dn->dn_objset->os_spa; |
674 | ||
675 | if (zio == NULL) | |
676 | zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); | |
34dc7c2f BB |
677 | dbuf_read_impl(db, zio, &flags); |
678 | ||
679 | /* dbuf_read_impl has dropped db_mtx for us */ | |
680 | ||
681 | if (prefetch) | |
572e2857 | 682 | dmu_zfetch(&dn->dn_zfetch, db->db.db_offset, |
34dc7c2f BB |
683 | db->db.db_size, flags & DB_RF_CACHED); |
684 | ||
685 | if ((flags & DB_RF_HAVESTRUCT) == 0) | |
572e2857 BB |
686 | rw_exit(&dn->dn_struct_rwlock); |
687 | DB_DNODE_EXIT(db); | |
34dc7c2f BB |
688 | |
689 | if (!havepzio) | |
690 | err = zio_wait(zio); | |
691 | } else { | |
692 | mutex_exit(&db->db_mtx); | |
693 | if (prefetch) | |
572e2857 | 694 | dmu_zfetch(&dn->dn_zfetch, db->db.db_offset, |
34dc7c2f BB |
695 | db->db.db_size, TRUE); |
696 | if ((flags & DB_RF_HAVESTRUCT) == 0) | |
572e2857 BB |
697 | rw_exit(&dn->dn_struct_rwlock); |
698 | DB_DNODE_EXIT(db); | |
34dc7c2f BB |
699 | |
700 | mutex_enter(&db->db_mtx); | |
701 | if ((flags & DB_RF_NEVERWAIT) == 0) { | |
702 | while (db->db_state == DB_READ || | |
703 | db->db_state == DB_FILL) { | |
704 | ASSERT(db->db_state == DB_READ || | |
705 | (flags & DB_RF_HAVESTRUCT) == 0); | |
706 | cv_wait(&db->db_changed, &db->db_mtx); | |
707 | } | |
708 | if (db->db_state == DB_UNCACHED) | |
709 | err = EIO; | |
710 | } | |
711 | mutex_exit(&db->db_mtx); | |
712 | } | |
713 | ||
714 | ASSERT(err || havepzio || db->db_state == DB_CACHED); | |
715 | return (err); | |
716 | } | |
717 | ||
718 | static void | |
719 | dbuf_noread(dmu_buf_impl_t *db) | |
720 | { | |
721 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
428870ff | 722 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f BB |
723 | mutex_enter(&db->db_mtx); |
724 | while (db->db_state == DB_READ || db->db_state == DB_FILL) | |
725 | cv_wait(&db->db_changed, &db->db_mtx); | |
726 | if (db->db_state == DB_UNCACHED) { | |
727 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); | |
572e2857 | 728 | spa_t *spa; |
34dc7c2f BB |
729 | |
730 | ASSERT(db->db_buf == NULL); | |
731 | ASSERT(db->db.db_data == NULL); | |
572e2857 BB |
732 | DB_GET_SPA(&spa, db); |
733 | dbuf_set_data(db, arc_buf_alloc(spa, db->db.db_size, db, type)); | |
34dc7c2f | 734 | db->db_state = DB_FILL; |
b128c09f BB |
735 | } else if (db->db_state == DB_NOFILL) { |
736 | dbuf_set_data(db, NULL); | |
34dc7c2f BB |
737 | } else { |
738 | ASSERT3U(db->db_state, ==, DB_CACHED); | |
739 | } | |
740 | mutex_exit(&db->db_mtx); | |
741 | } | |
742 | ||
743 | /* | |
744 | * This is our just-in-time copy function. It makes a copy of | |
745 | * buffers, that have been modified in a previous transaction | |
746 | * group, before we modify them in the current active group. | |
747 | * | |
748 | * This function is used in two places: when we are dirtying a | |
749 | * buffer for the first time in a txg, and when we are freeing | |
750 | * a range in a dnode that includes this buffer. | |
751 | * | |
752 | * Note that when we are called from dbuf_free_range() we do | |
753 | * not put a hold on the buffer, we just traverse the active | |
754 | * dbuf list for the dnode. | |
755 | */ | |
756 | static void | |
757 | dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg) | |
758 | { | |
759 | dbuf_dirty_record_t *dr = db->db_last_dirty; | |
760 | ||
761 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
762 | ASSERT(db->db.db_data != NULL); | |
763 | ASSERT(db->db_level == 0); | |
764 | ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT); | |
765 | ||
766 | if (dr == NULL || | |
767 | (dr->dt.dl.dr_data != | |
428870ff | 768 | ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf))) |
34dc7c2f BB |
769 | return; |
770 | ||
771 | /* | |
772 | * If the last dirty record for this dbuf has not yet synced | |
773 | * and its referencing the dbuf data, either: | |
572e2857 | 774 | * reset the reference to point to a new copy, |
34dc7c2f BB |
775 | * or (if there a no active holders) |
776 | * just null out the current db_data pointer. | |
777 | */ | |
778 | ASSERT(dr->dr_txg >= txg - 2); | |
428870ff | 779 | if (db->db_blkid == DMU_BONUS_BLKID) { |
34dc7c2f BB |
780 | /* Note that the data bufs here are zio_bufs */ |
781 | dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN); | |
d164b209 | 782 | arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); |
34dc7c2f BB |
783 | bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN); |
784 | } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) { | |
785 | int size = db->db.db_size; | |
786 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); | |
572e2857 BB |
787 | spa_t *spa; |
788 | ||
789 | DB_GET_SPA(&spa, db); | |
790 | dr->dt.dl.dr_data = arc_buf_alloc(spa, size, db, type); | |
34dc7c2f BB |
791 | bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size); |
792 | } else { | |
793 | dbuf_set_data(db, NULL); | |
794 | } | |
795 | } | |
796 | ||
797 | void | |
798 | dbuf_unoverride(dbuf_dirty_record_t *dr) | |
799 | { | |
800 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
428870ff | 801 | blkptr_t *bp = &dr->dt.dl.dr_overridden_by; |
34dc7c2f BB |
802 | uint64_t txg = dr->dr_txg; |
803 | ||
804 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
805 | ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC); | |
806 | ASSERT(db->db_level == 0); | |
807 | ||
428870ff | 808 | if (db->db_blkid == DMU_BONUS_BLKID || |
34dc7c2f BB |
809 | dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN) |
810 | return; | |
811 | ||
428870ff BB |
812 | ASSERT(db->db_data_pending != dr); |
813 | ||
34dc7c2f | 814 | /* free this block */ |
572e2857 BB |
815 | if (!BP_IS_HOLE(bp)) { |
816 | spa_t *spa; | |
428870ff | 817 | |
572e2857 BB |
818 | DB_GET_SPA(&spa, db); |
819 | zio_free(spa, txg, bp); | |
820 | } | |
34dc7c2f BB |
821 | dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; |
822 | /* | |
823 | * Release the already-written buffer, so we leave it in | |
824 | * a consistent dirty state. Note that all callers are | |
825 | * modifying the buffer, so they will immediately do | |
826 | * another (redundant) arc_release(). Therefore, leave | |
827 | * the buf thawed to save the effort of freezing & | |
828 | * immediately re-thawing it. | |
829 | */ | |
830 | arc_release(dr->dt.dl.dr_data, db); | |
831 | } | |
832 | ||
b128c09f BB |
833 | /* |
834 | * Evict (if its unreferenced) or clear (if its referenced) any level-0 | |
835 | * data blocks in the free range, so that any future readers will find | |
836 | * empty blocks. Also, if we happen accross any level-1 dbufs in the | |
837 | * range that have not already been marked dirty, mark them dirty so | |
838 | * they stay in memory. | |
839 | */ | |
34dc7c2f | 840 | void |
b128c09f | 841 | dbuf_free_range(dnode_t *dn, uint64_t start, uint64_t end, dmu_tx_t *tx) |
34dc7c2f BB |
842 | { |
843 | dmu_buf_impl_t *db, *db_next; | |
844 | uint64_t txg = tx->tx_txg; | |
b128c09f BB |
845 | int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; |
846 | uint64_t first_l1 = start >> epbs; | |
847 | uint64_t last_l1 = end >> epbs; | |
34dc7c2f | 848 | |
428870ff | 849 | if (end > dn->dn_maxblkid && (end != DMU_SPILL_BLKID)) { |
b128c09f BB |
850 | end = dn->dn_maxblkid; |
851 | last_l1 = end >> epbs; | |
852 | } | |
853 | dprintf_dnode(dn, "start=%llu end=%llu\n", start, end); | |
34dc7c2f BB |
854 | mutex_enter(&dn->dn_dbufs_mtx); |
855 | for (db = list_head(&dn->dn_dbufs); db; db = db_next) { | |
856 | db_next = list_next(&dn->dn_dbufs, db); | |
428870ff | 857 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
b128c09f BB |
858 | |
859 | if (db->db_level == 1 && | |
860 | db->db_blkid >= first_l1 && db->db_blkid <= last_l1) { | |
861 | mutex_enter(&db->db_mtx); | |
862 | if (db->db_last_dirty && | |
863 | db->db_last_dirty->dr_txg < txg) { | |
864 | dbuf_add_ref(db, FTAG); | |
865 | mutex_exit(&db->db_mtx); | |
866 | dbuf_will_dirty(db, tx); | |
867 | dbuf_rele(db, FTAG); | |
868 | } else { | |
869 | mutex_exit(&db->db_mtx); | |
870 | } | |
871 | } | |
872 | ||
34dc7c2f BB |
873 | if (db->db_level != 0) |
874 | continue; | |
875 | dprintf_dbuf(db, "found buf %s\n", ""); | |
b128c09f | 876 | if (db->db_blkid < start || db->db_blkid > end) |
34dc7c2f BB |
877 | continue; |
878 | ||
879 | /* found a level 0 buffer in the range */ | |
880 | if (dbuf_undirty(db, tx)) | |
881 | continue; | |
882 | ||
883 | mutex_enter(&db->db_mtx); | |
884 | if (db->db_state == DB_UNCACHED || | |
b128c09f | 885 | db->db_state == DB_NOFILL || |
34dc7c2f BB |
886 | db->db_state == DB_EVICTING) { |
887 | ASSERT(db->db.db_data == NULL); | |
888 | mutex_exit(&db->db_mtx); | |
889 | continue; | |
890 | } | |
891 | if (db->db_state == DB_READ || db->db_state == DB_FILL) { | |
892 | /* will be handled in dbuf_read_done or dbuf_rele */ | |
893 | db->db_freed_in_flight = TRUE; | |
894 | mutex_exit(&db->db_mtx); | |
895 | continue; | |
896 | } | |
897 | if (refcount_count(&db->db_holds) == 0) { | |
898 | ASSERT(db->db_buf); | |
899 | dbuf_clear(db); | |
900 | continue; | |
901 | } | |
902 | /* The dbuf is referenced */ | |
903 | ||
904 | if (db->db_last_dirty != NULL) { | |
905 | dbuf_dirty_record_t *dr = db->db_last_dirty; | |
906 | ||
907 | if (dr->dr_txg == txg) { | |
908 | /* | |
909 | * This buffer is "in-use", re-adjust the file | |
910 | * size to reflect that this buffer may | |
911 | * contain new data when we sync. | |
912 | */ | |
428870ff BB |
913 | if (db->db_blkid != DMU_SPILL_BLKID && |
914 | db->db_blkid > dn->dn_maxblkid) | |
34dc7c2f BB |
915 | dn->dn_maxblkid = db->db_blkid; |
916 | dbuf_unoverride(dr); | |
917 | } else { | |
918 | /* | |
919 | * This dbuf is not dirty in the open context. | |
920 | * Either uncache it (if its not referenced in | |
921 | * the open context) or reset its contents to | |
922 | * empty. | |
923 | */ | |
924 | dbuf_fix_old_data(db, txg); | |
925 | } | |
926 | } | |
927 | /* clear the contents if its cached */ | |
928 | if (db->db_state == DB_CACHED) { | |
929 | ASSERT(db->db.db_data != NULL); | |
930 | arc_release(db->db_buf, db); | |
931 | bzero(db->db.db_data, db->db.db_size); | |
932 | arc_buf_freeze(db->db_buf); | |
933 | } | |
934 | ||
935 | mutex_exit(&db->db_mtx); | |
936 | } | |
937 | mutex_exit(&dn->dn_dbufs_mtx); | |
938 | } | |
939 | ||
940 | static int | |
941 | dbuf_block_freeable(dmu_buf_impl_t *db) | |
942 | { | |
943 | dsl_dataset_t *ds = db->db_objset->os_dsl_dataset; | |
944 | uint64_t birth_txg = 0; | |
945 | ||
946 | /* | |
947 | * We don't need any locking to protect db_blkptr: | |
948 | * If it's syncing, then db_last_dirty will be set | |
949 | * so we'll ignore db_blkptr. | |
950 | */ | |
951 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
952 | if (db->db_last_dirty) | |
953 | birth_txg = db->db_last_dirty->dr_txg; | |
954 | else if (db->db_blkptr) | |
955 | birth_txg = db->db_blkptr->blk_birth; | |
956 | ||
572e2857 BB |
957 | /* |
958 | * If we don't exist or are in a snapshot, we can't be freed. | |
959 | * Don't pass the bp to dsl_dataset_block_freeable() since we | |
960 | * are holding the db_mtx lock and might deadlock if we are | |
961 | * prefetching a dedup-ed block. | |
962 | */ | |
34dc7c2f BB |
963 | if (birth_txg) |
964 | return (ds == NULL || | |
572e2857 | 965 | dsl_dataset_block_freeable(ds, NULL, birth_txg)); |
34dc7c2f BB |
966 | else |
967 | return (FALSE); | |
968 | } | |
969 | ||
970 | void | |
971 | dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx) | |
972 | { | |
973 | arc_buf_t *buf, *obuf; | |
974 | int osize = db->db.db_size; | |
975 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); | |
572e2857 | 976 | dnode_t *dn; |
34dc7c2f | 977 | |
428870ff | 978 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f | 979 | |
572e2857 BB |
980 | DB_DNODE_ENTER(db); |
981 | dn = DB_DNODE(db); | |
982 | ||
34dc7c2f | 983 | /* XXX does *this* func really need the lock? */ |
572e2857 | 984 | ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); |
34dc7c2f BB |
985 | |
986 | /* | |
987 | * This call to dbuf_will_dirty() with the dn_struct_rwlock held | |
988 | * is OK, because there can be no other references to the db | |
989 | * when we are changing its size, so no concurrent DB_FILL can | |
990 | * be happening. | |
991 | */ | |
992 | /* | |
993 | * XXX we should be doing a dbuf_read, checking the return | |
994 | * value and returning that up to our callers | |
995 | */ | |
996 | dbuf_will_dirty(db, tx); | |
997 | ||
998 | /* create the data buffer for the new block */ | |
572e2857 | 999 | buf = arc_buf_alloc(dn->dn_objset->os_spa, size, db, type); |
34dc7c2f BB |
1000 | |
1001 | /* copy old block data to the new block */ | |
1002 | obuf = db->db_buf; | |
1003 | bcopy(obuf->b_data, buf->b_data, MIN(osize, size)); | |
1004 | /* zero the remainder */ | |
1005 | if (size > osize) | |
1006 | bzero((uint8_t *)buf->b_data + osize, size - osize); | |
1007 | ||
1008 | mutex_enter(&db->db_mtx); | |
1009 | dbuf_set_data(db, buf); | |
1010 | VERIFY(arc_buf_remove_ref(obuf, db) == 1); | |
1011 | db->db.db_size = size; | |
1012 | ||
1013 | if (db->db_level == 0) { | |
1014 | ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); | |
1015 | db->db_last_dirty->dt.dl.dr_data = buf; | |
1016 | } | |
1017 | mutex_exit(&db->db_mtx); | |
1018 | ||
572e2857 BB |
1019 | dnode_willuse_space(dn, size-osize, tx); |
1020 | DB_DNODE_EXIT(db); | |
34dc7c2f BB |
1021 | } |
1022 | ||
428870ff BB |
1023 | void |
1024 | dbuf_release_bp(dmu_buf_impl_t *db) | |
1025 | { | |
572e2857 | 1026 | objset_t *os; |
428870ff BB |
1027 | zbookmark_t zb; |
1028 | ||
572e2857 | 1029 | DB_GET_OBJSET(&os, db); |
428870ff BB |
1030 | ASSERT(dsl_pool_sync_context(dmu_objset_pool(os))); |
1031 | ASSERT(arc_released(os->os_phys_buf) || | |
1032 | list_link_active(&os->os_dsl_dataset->ds_synced_link)); | |
1033 | ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf)); | |
1034 | ||
1035 | zb.zb_objset = os->os_dsl_dataset ? | |
1036 | os->os_dsl_dataset->ds_object : 0; | |
1037 | zb.zb_object = db->db.db_object; | |
1038 | zb.zb_level = db->db_level; | |
1039 | zb.zb_blkid = db->db_blkid; | |
1040 | (void) arc_release_bp(db->db_buf, db, | |
1041 | db->db_blkptr, os->os_spa, &zb); | |
1042 | } | |
1043 | ||
34dc7c2f BB |
1044 | dbuf_dirty_record_t * |
1045 | dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) | |
1046 | { | |
572e2857 BB |
1047 | dnode_t *dn; |
1048 | objset_t *os; | |
34dc7c2f BB |
1049 | dbuf_dirty_record_t **drp, *dr; |
1050 | int drop_struct_lock = FALSE; | |
b128c09f | 1051 | boolean_t do_free_accounting = B_FALSE; |
34dc7c2f BB |
1052 | int txgoff = tx->tx_txg & TXG_MASK; |
1053 | ||
1054 | ASSERT(tx->tx_txg != 0); | |
1055 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
1056 | DMU_TX_DIRTY_BUF(tx, db); | |
1057 | ||
572e2857 BB |
1058 | DB_DNODE_ENTER(db); |
1059 | dn = DB_DNODE(db); | |
34dc7c2f BB |
1060 | /* |
1061 | * Shouldn't dirty a regular buffer in syncing context. Private | |
1062 | * objects may be dirtied in syncing context, but only if they | |
1063 | * were already pre-dirtied in open context. | |
34dc7c2f BB |
1064 | */ |
1065 | ASSERT(!dmu_tx_is_syncing(tx) || | |
1066 | BP_IS_HOLE(dn->dn_objset->os_rootbp) || | |
9babb374 BB |
1067 | DMU_OBJECT_IS_SPECIAL(dn->dn_object) || |
1068 | dn->dn_objset->os_dsl_dataset == NULL); | |
34dc7c2f BB |
1069 | /* |
1070 | * We make this assert for private objects as well, but after we | |
1071 | * check if we're already dirty. They are allowed to re-dirty | |
1072 | * in syncing context. | |
1073 | */ | |
1074 | ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || | |
1075 | dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == | |
1076 | (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); | |
1077 | ||
1078 | mutex_enter(&db->db_mtx); | |
1079 | /* | |
1080 | * XXX make this true for indirects too? The problem is that | |
1081 | * transactions created with dmu_tx_create_assigned() from | |
1082 | * syncing context don't bother holding ahead. | |
1083 | */ | |
1084 | ASSERT(db->db_level != 0 || | |
b128c09f BB |
1085 | db->db_state == DB_CACHED || db->db_state == DB_FILL || |
1086 | db->db_state == DB_NOFILL); | |
34dc7c2f BB |
1087 | |
1088 | mutex_enter(&dn->dn_mtx); | |
1089 | /* | |
1090 | * Don't set dirtyctx to SYNC if we're just modifying this as we | |
1091 | * initialize the objset. | |
1092 | */ | |
1093 | if (dn->dn_dirtyctx == DN_UNDIRTIED && | |
1094 | !BP_IS_HOLE(dn->dn_objset->os_rootbp)) { | |
1095 | dn->dn_dirtyctx = | |
1096 | (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN); | |
1097 | ASSERT(dn->dn_dirtyctx_firstset == NULL); | |
1098 | dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP); | |
1099 | } | |
1100 | mutex_exit(&dn->dn_mtx); | |
1101 | ||
428870ff BB |
1102 | if (db->db_blkid == DMU_SPILL_BLKID) |
1103 | dn->dn_have_spill = B_TRUE; | |
1104 | ||
34dc7c2f BB |
1105 | /* |
1106 | * If this buffer is already dirty, we're done. | |
1107 | */ | |
1108 | drp = &db->db_last_dirty; | |
1109 | ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg || | |
1110 | db->db.db_object == DMU_META_DNODE_OBJECT); | |
1111 | while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg) | |
1112 | drp = &dr->dr_next; | |
1113 | if (dr && dr->dr_txg == tx->tx_txg) { | |
572e2857 BB |
1114 | DB_DNODE_EXIT(db); |
1115 | ||
428870ff | 1116 | if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) { |
34dc7c2f BB |
1117 | /* |
1118 | * If this buffer has already been written out, | |
1119 | * we now need to reset its state. | |
1120 | */ | |
1121 | dbuf_unoverride(dr); | |
428870ff BB |
1122 | if (db->db.db_object != DMU_META_DNODE_OBJECT && |
1123 | db->db_state != DB_NOFILL) | |
34dc7c2f BB |
1124 | arc_buf_thaw(db->db_buf); |
1125 | } | |
1126 | mutex_exit(&db->db_mtx); | |
1127 | return (dr); | |
1128 | } | |
1129 | ||
1130 | /* | |
1131 | * Only valid if not already dirty. | |
1132 | */ | |
9babb374 BB |
1133 | ASSERT(dn->dn_object == 0 || |
1134 | dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == | |
34dc7c2f BB |
1135 | (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); |
1136 | ||
1137 | ASSERT3U(dn->dn_nlevels, >, db->db_level); | |
1138 | ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) || | |
1139 | dn->dn_phys->dn_nlevels > db->db_level || | |
1140 | dn->dn_next_nlevels[txgoff] > db->db_level || | |
1141 | dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level || | |
1142 | dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level); | |
1143 | ||
1144 | /* | |
1145 | * We should only be dirtying in syncing context if it's the | |
9babb374 BB |
1146 | * mos or we're initializing the os or it's a special object. |
1147 | * However, we are allowed to dirty in syncing context provided | |
1148 | * we already dirtied it in open context. Hence we must make | |
1149 | * this assertion only if we're not already dirty. | |
34dc7c2f | 1150 | */ |
572e2857 | 1151 | os = dn->dn_objset; |
9babb374 BB |
1152 | ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) || |
1153 | os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp)); | |
34dc7c2f BB |
1154 | ASSERT(db->db.db_size != 0); |
1155 | ||
1156 | dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); | |
1157 | ||
428870ff | 1158 | if (db->db_blkid != DMU_BONUS_BLKID) { |
34dc7c2f BB |
1159 | /* |
1160 | * Update the accounting. | |
b128c09f BB |
1161 | * Note: we delay "free accounting" until after we drop |
1162 | * the db_mtx. This keeps us from grabbing other locks | |
428870ff | 1163 | * (and possibly deadlocking) in bp_get_dsize() while |
b128c09f | 1164 | * also holding the db_mtx. |
34dc7c2f | 1165 | */ |
34dc7c2f | 1166 | dnode_willuse_space(dn, db->db.db_size, tx); |
b128c09f | 1167 | do_free_accounting = dbuf_block_freeable(db); |
34dc7c2f BB |
1168 | } |
1169 | ||
1170 | /* | |
1171 | * If this buffer is dirty in an old transaction group we need | |
1172 | * to make a copy of it so that the changes we make in this | |
1173 | * transaction group won't leak out when we sync the older txg. | |
1174 | */ | |
1175 | dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP); | |
98f72a53 | 1176 | list_link_init(&dr->dr_dirty_node); |
34dc7c2f BB |
1177 | if (db->db_level == 0) { |
1178 | void *data_old = db->db_buf; | |
1179 | ||
b128c09f | 1180 | if (db->db_state != DB_NOFILL) { |
428870ff | 1181 | if (db->db_blkid == DMU_BONUS_BLKID) { |
b128c09f BB |
1182 | dbuf_fix_old_data(db, tx->tx_txg); |
1183 | data_old = db->db.db_data; | |
1184 | } else if (db->db.db_object != DMU_META_DNODE_OBJECT) { | |
1185 | /* | |
1186 | * Release the data buffer from the cache so | |
1187 | * that we can modify it without impacting | |
1188 | * possible other users of this cached data | |
1189 | * block. Note that indirect blocks and | |
1190 | * private objects are not released until the | |
1191 | * syncing state (since they are only modified | |
1192 | * then). | |
1193 | */ | |
1194 | arc_release(db->db_buf, db); | |
1195 | dbuf_fix_old_data(db, tx->tx_txg); | |
1196 | data_old = db->db_buf; | |
1197 | } | |
1198 | ASSERT(data_old != NULL); | |
34dc7c2f | 1199 | } |
34dc7c2f BB |
1200 | dr->dt.dl.dr_data = data_old; |
1201 | } else { | |
1202 | mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL); | |
1203 | list_create(&dr->dt.di.dr_children, | |
1204 | sizeof (dbuf_dirty_record_t), | |
1205 | offsetof(dbuf_dirty_record_t, dr_dirty_node)); | |
1206 | } | |
1207 | dr->dr_dbuf = db; | |
1208 | dr->dr_txg = tx->tx_txg; | |
1209 | dr->dr_next = *drp; | |
1210 | *drp = dr; | |
1211 | ||
1212 | /* | |
1213 | * We could have been freed_in_flight between the dbuf_noread | |
1214 | * and dbuf_dirty. We win, as though the dbuf_noread() had | |
1215 | * happened after the free. | |
1216 | */ | |
428870ff BB |
1217 | if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && |
1218 | db->db_blkid != DMU_SPILL_BLKID) { | |
34dc7c2f BB |
1219 | mutex_enter(&dn->dn_mtx); |
1220 | dnode_clear_range(dn, db->db_blkid, 1, tx); | |
1221 | mutex_exit(&dn->dn_mtx); | |
1222 | db->db_freed_in_flight = FALSE; | |
1223 | } | |
1224 | ||
1225 | /* | |
1226 | * This buffer is now part of this txg | |
1227 | */ | |
1228 | dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg); | |
1229 | db->db_dirtycnt += 1; | |
1230 | ASSERT3U(db->db_dirtycnt, <=, 3); | |
1231 | ||
1232 | mutex_exit(&db->db_mtx); | |
1233 | ||
428870ff BB |
1234 | if (db->db_blkid == DMU_BONUS_BLKID || |
1235 | db->db_blkid == DMU_SPILL_BLKID) { | |
34dc7c2f BB |
1236 | mutex_enter(&dn->dn_mtx); |
1237 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
1238 | list_insert_tail(&dn->dn_dirty_records[txgoff], dr); | |
1239 | mutex_exit(&dn->dn_mtx); | |
1240 | dnode_setdirty(dn, tx); | |
572e2857 | 1241 | DB_DNODE_EXIT(db); |
34dc7c2f | 1242 | return (dr); |
b128c09f BB |
1243 | } else if (do_free_accounting) { |
1244 | blkptr_t *bp = db->db_blkptr; | |
1245 | int64_t willfree = (bp && !BP_IS_HOLE(bp)) ? | |
428870ff | 1246 | bp_get_dsize(os->os_spa, bp) : db->db.db_size; |
b128c09f BB |
1247 | /* |
1248 | * This is only a guess -- if the dbuf is dirty | |
1249 | * in a previous txg, we don't know how much | |
1250 | * space it will use on disk yet. We should | |
1251 | * really have the struct_rwlock to access | |
1252 | * db_blkptr, but since this is just a guess, | |
1253 | * it's OK if we get an odd answer. | |
1254 | */ | |
572e2857 | 1255 | ddt_prefetch(os->os_spa, bp); |
b128c09f | 1256 | dnode_willuse_space(dn, -willfree, tx); |
34dc7c2f BB |
1257 | } |
1258 | ||
1259 | if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) { | |
1260 | rw_enter(&dn->dn_struct_rwlock, RW_READER); | |
1261 | drop_struct_lock = TRUE; | |
1262 | } | |
1263 | ||
b128c09f BB |
1264 | if (db->db_level == 0) { |
1265 | dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock); | |
1266 | ASSERT(dn->dn_maxblkid >= db->db_blkid); | |
1267 | } | |
1268 | ||
34dc7c2f BB |
1269 | if (db->db_level+1 < dn->dn_nlevels) { |
1270 | dmu_buf_impl_t *parent = db->db_parent; | |
1271 | dbuf_dirty_record_t *di; | |
1272 | int parent_held = FALSE; | |
1273 | ||
1274 | if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) { | |
1275 | int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; | |
1276 | ||
1277 | parent = dbuf_hold_level(dn, db->db_level+1, | |
1278 | db->db_blkid >> epbs, FTAG); | |
428870ff | 1279 | ASSERT(parent != NULL); |
34dc7c2f BB |
1280 | parent_held = TRUE; |
1281 | } | |
1282 | if (drop_struct_lock) | |
1283 | rw_exit(&dn->dn_struct_rwlock); | |
1284 | ASSERT3U(db->db_level+1, ==, parent->db_level); | |
1285 | di = dbuf_dirty(parent, tx); | |
1286 | if (parent_held) | |
1287 | dbuf_rele(parent, FTAG); | |
1288 | ||
1289 | mutex_enter(&db->db_mtx); | |
1290 | /* possible race with dbuf_undirty() */ | |
1291 | if (db->db_last_dirty == dr || | |
1292 | dn->dn_object == DMU_META_DNODE_OBJECT) { | |
1293 | mutex_enter(&di->dt.di.dr_mtx); | |
1294 | ASSERT3U(di->dr_txg, ==, tx->tx_txg); | |
1295 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
1296 | list_insert_tail(&di->dt.di.dr_children, dr); | |
1297 | mutex_exit(&di->dt.di.dr_mtx); | |
1298 | dr->dr_parent = di; | |
1299 | } | |
1300 | mutex_exit(&db->db_mtx); | |
1301 | } else { | |
1302 | ASSERT(db->db_level+1 == dn->dn_nlevels); | |
1303 | ASSERT(db->db_blkid < dn->dn_nblkptr); | |
572e2857 | 1304 | ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf); |
34dc7c2f BB |
1305 | mutex_enter(&dn->dn_mtx); |
1306 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
1307 | list_insert_tail(&dn->dn_dirty_records[txgoff], dr); | |
1308 | mutex_exit(&dn->dn_mtx); | |
1309 | if (drop_struct_lock) | |
1310 | rw_exit(&dn->dn_struct_rwlock); | |
1311 | } | |
1312 | ||
1313 | dnode_setdirty(dn, tx); | |
572e2857 | 1314 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
1315 | return (dr); |
1316 | } | |
1317 | ||
1318 | static int | |
1319 | dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx) | |
1320 | { | |
572e2857 | 1321 | dnode_t *dn; |
34dc7c2f BB |
1322 | uint64_t txg = tx->tx_txg; |
1323 | dbuf_dirty_record_t *dr, **drp; | |
1324 | ||
1325 | ASSERT(txg != 0); | |
428870ff | 1326 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f BB |
1327 | |
1328 | mutex_enter(&db->db_mtx); | |
34dc7c2f BB |
1329 | /* |
1330 | * If this buffer is not dirty, we're done. | |
1331 | */ | |
1332 | for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next) | |
1333 | if (dr->dr_txg <= txg) | |
1334 | break; | |
1335 | if (dr == NULL || dr->dr_txg < txg) { | |
1336 | mutex_exit(&db->db_mtx); | |
1337 | return (0); | |
1338 | } | |
1339 | ASSERT(dr->dr_txg == txg); | |
428870ff | 1340 | ASSERT(dr->dr_dbuf == db); |
34dc7c2f | 1341 | |
572e2857 BB |
1342 | DB_DNODE_ENTER(db); |
1343 | dn = DB_DNODE(db); | |
1344 | ||
34dc7c2f BB |
1345 | /* |
1346 | * If this buffer is currently held, we cannot undirty | |
1347 | * it, since one of the current holders may be in the | |
1348 | * middle of an update. Note that users of dbuf_undirty() | |
1349 | * should not place a hold on the dbuf before the call. | |
1350 | */ | |
1351 | if (refcount_count(&db->db_holds) > db->db_dirtycnt) { | |
1352 | mutex_exit(&db->db_mtx); | |
1353 | /* Make sure we don't toss this buffer at sync phase */ | |
1354 | mutex_enter(&dn->dn_mtx); | |
1355 | dnode_clear_range(dn, db->db_blkid, 1, tx); | |
1356 | mutex_exit(&dn->dn_mtx); | |
572e2857 | 1357 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
1358 | return (0); |
1359 | } | |
1360 | ||
1361 | dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); | |
1362 | ||
1363 | ASSERT(db->db.db_size != 0); | |
1364 | ||
1365 | /* XXX would be nice to fix up dn_towrite_space[] */ | |
1366 | ||
1367 | *drp = dr->dr_next; | |
1368 | ||
1369 | if (dr->dr_parent) { | |
1370 | mutex_enter(&dr->dr_parent->dt.di.dr_mtx); | |
1371 | list_remove(&dr->dr_parent->dt.di.dr_children, dr); | |
1372 | mutex_exit(&dr->dr_parent->dt.di.dr_mtx); | |
1373 | } else if (db->db_level+1 == dn->dn_nlevels) { | |
b128c09f | 1374 | ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf); |
34dc7c2f BB |
1375 | mutex_enter(&dn->dn_mtx); |
1376 | list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr); | |
1377 | mutex_exit(&dn->dn_mtx); | |
1378 | } | |
572e2857 | 1379 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
1380 | |
1381 | if (db->db_level == 0) { | |
b128c09f BB |
1382 | if (db->db_state != DB_NOFILL) { |
1383 | dbuf_unoverride(dr); | |
34dc7c2f | 1384 | |
b128c09f BB |
1385 | ASSERT(db->db_buf != NULL); |
1386 | ASSERT(dr->dt.dl.dr_data != NULL); | |
1387 | if (dr->dt.dl.dr_data != db->db_buf) | |
1388 | VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, | |
1389 | db) == 1); | |
1390 | } | |
34dc7c2f BB |
1391 | } else { |
1392 | ASSERT(db->db_buf != NULL); | |
1393 | ASSERT(list_head(&dr->dt.di.dr_children) == NULL); | |
1394 | mutex_destroy(&dr->dt.di.dr_mtx); | |
1395 | list_destroy(&dr->dt.di.dr_children); | |
1396 | } | |
1397 | kmem_free(dr, sizeof (dbuf_dirty_record_t)); | |
1398 | ||
1399 | ASSERT(db->db_dirtycnt > 0); | |
1400 | db->db_dirtycnt -= 1; | |
1401 | ||
1402 | if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) { | |
1403 | arc_buf_t *buf = db->db_buf; | |
1404 | ||
428870ff | 1405 | ASSERT(db->db_state == DB_NOFILL || arc_released(buf)); |
34dc7c2f BB |
1406 | dbuf_set_data(db, NULL); |
1407 | VERIFY(arc_buf_remove_ref(buf, db) == 1); | |
1408 | dbuf_evict(db); | |
1409 | return (1); | |
1410 | } | |
1411 | ||
1412 | mutex_exit(&db->db_mtx); | |
1413 | return (0); | |
1414 | } | |
1415 | ||
1416 | #pragma weak dmu_buf_will_dirty = dbuf_will_dirty | |
1417 | void | |
1418 | dbuf_will_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) | |
1419 | { | |
1420 | int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH; | |
1421 | ||
1422 | ASSERT(tx->tx_txg != 0); | |
1423 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
1424 | ||
572e2857 BB |
1425 | DB_DNODE_ENTER(db); |
1426 | if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock)) | |
34dc7c2f | 1427 | rf |= DB_RF_HAVESTRUCT; |
572e2857 | 1428 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
1429 | (void) dbuf_read(db, NULL, rf); |
1430 | (void) dbuf_dirty(db, tx); | |
1431 | } | |
1432 | ||
b128c09f BB |
1433 | void |
1434 | dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) | |
1435 | { | |
1436 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
1437 | ||
1438 | db->db_state = DB_NOFILL; | |
1439 | ||
1440 | dmu_buf_will_fill(db_fake, tx); | |
1441 | } | |
1442 | ||
34dc7c2f BB |
1443 | void |
1444 | dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) | |
1445 | { | |
1446 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
1447 | ||
428870ff | 1448 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f BB |
1449 | ASSERT(tx->tx_txg != 0); |
1450 | ASSERT(db->db_level == 0); | |
1451 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
1452 | ||
1453 | ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT || | |
1454 | dmu_tx_private_ok(tx)); | |
1455 | ||
1456 | dbuf_noread(db); | |
1457 | (void) dbuf_dirty(db, tx); | |
1458 | } | |
1459 | ||
1460 | #pragma weak dmu_buf_fill_done = dbuf_fill_done | |
1461 | /* ARGSUSED */ | |
1462 | void | |
1463 | dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx) | |
1464 | { | |
1465 | mutex_enter(&db->db_mtx); | |
1466 | DBUF_VERIFY(db); | |
1467 | ||
1468 | if (db->db_state == DB_FILL) { | |
1469 | if (db->db_level == 0 && db->db_freed_in_flight) { | |
428870ff | 1470 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f BB |
1471 | /* we were freed while filling */ |
1472 | /* XXX dbuf_undirty? */ | |
1473 | bzero(db->db.db_data, db->db.db_size); | |
1474 | db->db_freed_in_flight = FALSE; | |
1475 | } | |
1476 | db->db_state = DB_CACHED; | |
1477 | cv_broadcast(&db->db_changed); | |
1478 | } | |
1479 | mutex_exit(&db->db_mtx); | |
1480 | } | |
1481 | ||
9babb374 BB |
1482 | /* |
1483 | * Directly assign a provided arc buf to a given dbuf if it's not referenced | |
1484 | * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf. | |
1485 | */ | |
1486 | void | |
1487 | dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx) | |
1488 | { | |
1489 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
428870ff | 1490 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
9babb374 BB |
1491 | ASSERT(db->db_level == 0); |
1492 | ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA); | |
1493 | ASSERT(buf != NULL); | |
1494 | ASSERT(arc_buf_size(buf) == db->db.db_size); | |
1495 | ASSERT(tx->tx_txg != 0); | |
1496 | ||
1497 | arc_return_buf(buf, db); | |
1498 | ASSERT(arc_released(buf)); | |
1499 | ||
1500 | mutex_enter(&db->db_mtx); | |
1501 | ||
1502 | while (db->db_state == DB_READ || db->db_state == DB_FILL) | |
1503 | cv_wait(&db->db_changed, &db->db_mtx); | |
1504 | ||
1505 | ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED); | |
1506 | ||
1507 | if (db->db_state == DB_CACHED && | |
1508 | refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) { | |
1509 | mutex_exit(&db->db_mtx); | |
1510 | (void) dbuf_dirty(db, tx); | |
1511 | bcopy(buf->b_data, db->db.db_data, db->db.db_size); | |
1512 | VERIFY(arc_buf_remove_ref(buf, db) == 1); | |
428870ff | 1513 | xuio_stat_wbuf_copied(); |
9babb374 BB |
1514 | return; |
1515 | } | |
1516 | ||
428870ff | 1517 | xuio_stat_wbuf_nocopy(); |
9babb374 BB |
1518 | if (db->db_state == DB_CACHED) { |
1519 | dbuf_dirty_record_t *dr = db->db_last_dirty; | |
1520 | ||
1521 | ASSERT(db->db_buf != NULL); | |
1522 | if (dr != NULL && dr->dr_txg == tx->tx_txg) { | |
1523 | ASSERT(dr->dt.dl.dr_data == db->db_buf); | |
1524 | if (!arc_released(db->db_buf)) { | |
1525 | ASSERT(dr->dt.dl.dr_override_state == | |
1526 | DR_OVERRIDDEN); | |
1527 | arc_release(db->db_buf, db); | |
1528 | } | |
1529 | dr->dt.dl.dr_data = buf; | |
1530 | VERIFY(arc_buf_remove_ref(db->db_buf, db) == 1); | |
1531 | } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) { | |
1532 | arc_release(db->db_buf, db); | |
1533 | VERIFY(arc_buf_remove_ref(db->db_buf, db) == 1); | |
1534 | } | |
1535 | db->db_buf = NULL; | |
1536 | } | |
1537 | ASSERT(db->db_buf == NULL); | |
1538 | dbuf_set_data(db, buf); | |
1539 | db->db_state = DB_FILL; | |
1540 | mutex_exit(&db->db_mtx); | |
1541 | (void) dbuf_dirty(db, tx); | |
1542 | dbuf_fill_done(db, tx); | |
1543 | } | |
1544 | ||
34dc7c2f BB |
1545 | /* |
1546 | * "Clear" the contents of this dbuf. This will mark the dbuf | |
1547 | * EVICTING and clear *most* of its references. Unfortunetely, | |
1548 | * when we are not holding the dn_dbufs_mtx, we can't clear the | |
1549 | * entry in the dn_dbufs list. We have to wait until dbuf_destroy() | |
1550 | * in this case. For callers from the DMU we will usually see: | |
1551 | * dbuf_clear()->arc_buf_evict()->dbuf_do_evict()->dbuf_destroy() | |
1552 | * For the arc callback, we will usually see: | |
572e2857 | 1553 | * dbuf_do_evict()->dbuf_clear();dbuf_destroy() |
34dc7c2f BB |
1554 | * Sometimes, though, we will get a mix of these two: |
1555 | * DMU: dbuf_clear()->arc_buf_evict() | |
1556 | * ARC: dbuf_do_evict()->dbuf_destroy() | |
1557 | */ | |
1558 | void | |
1559 | dbuf_clear(dmu_buf_impl_t *db) | |
1560 | { | |
572e2857 | 1561 | dnode_t *dn; |
34dc7c2f | 1562 | dmu_buf_impl_t *parent = db->db_parent; |
572e2857 | 1563 | dmu_buf_impl_t *dndb; |
34dc7c2f BB |
1564 | int dbuf_gone = FALSE; |
1565 | ||
1566 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
1567 | ASSERT(refcount_is_zero(&db->db_holds)); | |
1568 | ||
1569 | dbuf_evict_user(db); | |
1570 | ||
1571 | if (db->db_state == DB_CACHED) { | |
1572 | ASSERT(db->db.db_data != NULL); | |
428870ff | 1573 | if (db->db_blkid == DMU_BONUS_BLKID) { |
34dc7c2f | 1574 | zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN); |
d164b209 | 1575 | arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); |
34dc7c2f BB |
1576 | } |
1577 | db->db.db_data = NULL; | |
1578 | db->db_state = DB_UNCACHED; | |
1579 | } | |
1580 | ||
b128c09f | 1581 | ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL); |
34dc7c2f BB |
1582 | ASSERT(db->db_data_pending == NULL); |
1583 | ||
1584 | db->db_state = DB_EVICTING; | |
1585 | db->db_blkptr = NULL; | |
1586 | ||
572e2857 BB |
1587 | DB_DNODE_ENTER(db); |
1588 | dn = DB_DNODE(db); | |
1589 | dndb = dn->dn_dbuf; | |
428870ff | 1590 | if (db->db_blkid != DMU_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) { |
34dc7c2f | 1591 | list_remove(&dn->dn_dbufs, db); |
572e2857 BB |
1592 | (void) atomic_dec_32_nv(&dn->dn_dbufs_count); |
1593 | membar_producer(); | |
1594 | DB_DNODE_EXIT(db); | |
1595 | /* | |
1596 | * Decrementing the dbuf count means that the hold corresponding | |
1597 | * to the removed dbuf is no longer discounted in dnode_move(), | |
1598 | * so the dnode cannot be moved until after we release the hold. | |
1599 | * The membar_producer() ensures visibility of the decremented | |
1600 | * value in dnode_move(), since DB_DNODE_EXIT doesn't actually | |
1601 | * release any lock. | |
1602 | */ | |
34dc7c2f | 1603 | dnode_rele(dn, db); |
572e2857 BB |
1604 | db->db_dnode_handle = NULL; |
1605 | } else { | |
1606 | DB_DNODE_EXIT(db); | |
34dc7c2f BB |
1607 | } |
1608 | ||
1609 | if (db->db_buf) | |
1610 | dbuf_gone = arc_buf_evict(db->db_buf); | |
1611 | ||
1612 | if (!dbuf_gone) | |
1613 | mutex_exit(&db->db_mtx); | |
1614 | ||
1615 | /* | |
572e2857 | 1616 | * If this dbuf is referenced from an indirect dbuf, |
34dc7c2f BB |
1617 | * decrement the ref count on the indirect dbuf. |
1618 | */ | |
1619 | if (parent && parent != dndb) | |
1620 | dbuf_rele(parent, db); | |
1621 | } | |
1622 | ||
bf701a83 BB |
1623 | __attribute__((always_inline)) |
1624 | static inline int | |
34dc7c2f | 1625 | dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse, |
fc5bb51f | 1626 | dmu_buf_impl_t **parentp, blkptr_t **bpp, struct dbuf_hold_impl_data *dh) |
34dc7c2f BB |
1627 | { |
1628 | int nlevels, epbs; | |
1629 | ||
1630 | *parentp = NULL; | |
1631 | *bpp = NULL; | |
1632 | ||
428870ff BB |
1633 | ASSERT(blkid != DMU_BONUS_BLKID); |
1634 | ||
1635 | if (blkid == DMU_SPILL_BLKID) { | |
1636 | mutex_enter(&dn->dn_mtx); | |
1637 | if (dn->dn_have_spill && | |
1638 | (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) | |
1639 | *bpp = &dn->dn_phys->dn_spill; | |
1640 | else | |
1641 | *bpp = NULL; | |
1642 | dbuf_add_ref(dn->dn_dbuf, NULL); | |
1643 | *parentp = dn->dn_dbuf; | |
1644 | mutex_exit(&dn->dn_mtx); | |
1645 | return (0); | |
1646 | } | |
34dc7c2f BB |
1647 | |
1648 | if (dn->dn_phys->dn_nlevels == 0) | |
1649 | nlevels = 1; | |
1650 | else | |
1651 | nlevels = dn->dn_phys->dn_nlevels; | |
1652 | ||
1653 | epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; | |
1654 | ||
1655 | ASSERT3U(level * epbs, <, 64); | |
1656 | ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); | |
1657 | if (level >= nlevels || | |
1658 | (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) { | |
1659 | /* the buffer has no parent yet */ | |
1660 | return (ENOENT); | |
1661 | } else if (level < nlevels-1) { | |
1662 | /* this block is referenced from an indirect block */ | |
fc5bb51f BB |
1663 | int err; |
1664 | if (dh == NULL) { | |
1665 | err = dbuf_hold_impl(dn, level+1, blkid >> epbs, | |
1666 | fail_sparse, NULL, parentp); | |
1667 | } | |
1668 | else { | |
1669 | __dbuf_hold_impl_init(dh + 1, dn, dh->dh_level + 1, | |
1670 | blkid >> epbs, fail_sparse, NULL, | |
1671 | parentp, dh->dh_depth + 1); | |
1672 | err = __dbuf_hold_impl(dh + 1); | |
1673 | } | |
34dc7c2f BB |
1674 | if (err) |
1675 | return (err); | |
1676 | err = dbuf_read(*parentp, NULL, | |
1677 | (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL)); | |
1678 | if (err) { | |
1679 | dbuf_rele(*parentp, NULL); | |
1680 | *parentp = NULL; | |
1681 | return (err); | |
1682 | } | |
1683 | *bpp = ((blkptr_t *)(*parentp)->db.db_data) + | |
1684 | (blkid & ((1ULL << epbs) - 1)); | |
1685 | return (0); | |
1686 | } else { | |
1687 | /* the block is referenced from the dnode */ | |
1688 | ASSERT3U(level, ==, nlevels-1); | |
1689 | ASSERT(dn->dn_phys->dn_nblkptr == 0 || | |
1690 | blkid < dn->dn_phys->dn_nblkptr); | |
1691 | if (dn->dn_dbuf) { | |
1692 | dbuf_add_ref(dn->dn_dbuf, NULL); | |
1693 | *parentp = dn->dn_dbuf; | |
1694 | } | |
1695 | *bpp = &dn->dn_phys->dn_blkptr[blkid]; | |
1696 | return (0); | |
1697 | } | |
1698 | } | |
1699 | ||
1700 | static dmu_buf_impl_t * | |
1701 | dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid, | |
1702 | dmu_buf_impl_t *parent, blkptr_t *blkptr) | |
1703 | { | |
428870ff | 1704 | objset_t *os = dn->dn_objset; |
34dc7c2f BB |
1705 | dmu_buf_impl_t *db, *odb; |
1706 | ||
1707 | ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); | |
1708 | ASSERT(dn->dn_type != DMU_OT_NONE); | |
1709 | ||
1710 | db = kmem_cache_alloc(dbuf_cache, KM_SLEEP); | |
1711 | ||
1712 | db->db_objset = os; | |
1713 | db->db.db_object = dn->dn_object; | |
1714 | db->db_level = level; | |
1715 | db->db_blkid = blkid; | |
1716 | db->db_last_dirty = NULL; | |
1717 | db->db_dirtycnt = 0; | |
572e2857 | 1718 | db->db_dnode_handle = dn->dn_handle; |
34dc7c2f BB |
1719 | db->db_parent = parent; |
1720 | db->db_blkptr = blkptr; | |
1721 | ||
1722 | db->db_user_ptr = NULL; | |
1723 | db->db_user_data_ptr_ptr = NULL; | |
1724 | db->db_evict_func = NULL; | |
1725 | db->db_immediate_evict = 0; | |
1726 | db->db_freed_in_flight = 0; | |
1727 | ||
428870ff | 1728 | if (blkid == DMU_BONUS_BLKID) { |
34dc7c2f BB |
1729 | ASSERT3P(parent, ==, dn->dn_dbuf); |
1730 | db->db.db_size = DN_MAX_BONUSLEN - | |
1731 | (dn->dn_nblkptr-1) * sizeof (blkptr_t); | |
1732 | ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); | |
428870ff | 1733 | db->db.db_offset = DMU_BONUS_BLKID; |
34dc7c2f BB |
1734 | db->db_state = DB_UNCACHED; |
1735 | /* the bonus dbuf is not placed in the hash table */ | |
d164b209 | 1736 | arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); |
34dc7c2f | 1737 | return (db); |
428870ff BB |
1738 | } else if (blkid == DMU_SPILL_BLKID) { |
1739 | db->db.db_size = (blkptr != NULL) ? | |
1740 | BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE; | |
1741 | db->db.db_offset = 0; | |
34dc7c2f BB |
1742 | } else { |
1743 | int blocksize = | |
1744 | db->db_level ? 1<<dn->dn_indblkshift : dn->dn_datablksz; | |
1745 | db->db.db_size = blocksize; | |
1746 | db->db.db_offset = db->db_blkid * blocksize; | |
1747 | } | |
1748 | ||
1749 | /* | |
1750 | * Hold the dn_dbufs_mtx while we get the new dbuf | |
1751 | * in the hash table *and* added to the dbufs list. | |
1752 | * This prevents a possible deadlock with someone | |
1753 | * trying to look up this dbuf before its added to the | |
1754 | * dn_dbufs list. | |
1755 | */ | |
1756 | mutex_enter(&dn->dn_dbufs_mtx); | |
1757 | db->db_state = DB_EVICTING; | |
1758 | if ((odb = dbuf_hash_insert(db)) != NULL) { | |
1759 | /* someone else inserted it first */ | |
1760 | kmem_cache_free(dbuf_cache, db); | |
1761 | mutex_exit(&dn->dn_dbufs_mtx); | |
1762 | return (odb); | |
1763 | } | |
1764 | list_insert_head(&dn->dn_dbufs, db); | |
1765 | db->db_state = DB_UNCACHED; | |
1766 | mutex_exit(&dn->dn_dbufs_mtx); | |
d164b209 | 1767 | arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); |
34dc7c2f BB |
1768 | |
1769 | if (parent && parent != dn->dn_dbuf) | |
1770 | dbuf_add_ref(parent, db); | |
1771 | ||
1772 | ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || | |
1773 | refcount_count(&dn->dn_holds) > 0); | |
1774 | (void) refcount_add(&dn->dn_holds, db); | |
572e2857 | 1775 | (void) atomic_inc_32_nv(&dn->dn_dbufs_count); |
34dc7c2f BB |
1776 | |
1777 | dprintf_dbuf(db, "db=%p\n", db); | |
1778 | ||
1779 | return (db); | |
1780 | } | |
1781 | ||
1782 | static int | |
1783 | dbuf_do_evict(void *private) | |
1784 | { | |
1785 | arc_buf_t *buf = private; | |
1786 | dmu_buf_impl_t *db = buf->b_private; | |
1787 | ||
1788 | if (!MUTEX_HELD(&db->db_mtx)) | |
1789 | mutex_enter(&db->db_mtx); | |
1790 | ||
1791 | ASSERT(refcount_is_zero(&db->db_holds)); | |
1792 | ||
1793 | if (db->db_state != DB_EVICTING) { | |
1794 | ASSERT(db->db_state == DB_CACHED); | |
1795 | DBUF_VERIFY(db); | |
1796 | db->db_buf = NULL; | |
1797 | dbuf_evict(db); | |
1798 | } else { | |
1799 | mutex_exit(&db->db_mtx); | |
1800 | dbuf_destroy(db); | |
1801 | } | |
1802 | return (0); | |
1803 | } | |
1804 | ||
1805 | static void | |
1806 | dbuf_destroy(dmu_buf_impl_t *db) | |
1807 | { | |
1808 | ASSERT(refcount_is_zero(&db->db_holds)); | |
1809 | ||
428870ff | 1810 | if (db->db_blkid != DMU_BONUS_BLKID) { |
34dc7c2f BB |
1811 | /* |
1812 | * If this dbuf is still on the dn_dbufs list, | |
1813 | * remove it from that list. | |
1814 | */ | |
572e2857 BB |
1815 | if (db->db_dnode_handle != NULL) { |
1816 | dnode_t *dn; | |
34dc7c2f | 1817 | |
572e2857 BB |
1818 | DB_DNODE_ENTER(db); |
1819 | dn = DB_DNODE(db); | |
34dc7c2f BB |
1820 | mutex_enter(&dn->dn_dbufs_mtx); |
1821 | list_remove(&dn->dn_dbufs, db); | |
572e2857 | 1822 | (void) atomic_dec_32_nv(&dn->dn_dbufs_count); |
34dc7c2f | 1823 | mutex_exit(&dn->dn_dbufs_mtx); |
572e2857 BB |
1824 | DB_DNODE_EXIT(db); |
1825 | /* | |
1826 | * Decrementing the dbuf count means that the hold | |
1827 | * corresponding to the removed dbuf is no longer | |
1828 | * discounted in dnode_move(), so the dnode cannot be | |
1829 | * moved until after we release the hold. | |
1830 | */ | |
34dc7c2f | 1831 | dnode_rele(dn, db); |
572e2857 | 1832 | db->db_dnode_handle = NULL; |
34dc7c2f BB |
1833 | } |
1834 | dbuf_hash_remove(db); | |
1835 | } | |
1836 | db->db_parent = NULL; | |
1837 | db->db_buf = NULL; | |
1838 | ||
1839 | ASSERT(!list_link_active(&db->db_link)); | |
1840 | ASSERT(db->db.db_data == NULL); | |
1841 | ASSERT(db->db_hash_next == NULL); | |
1842 | ASSERT(db->db_blkptr == NULL); | |
1843 | ASSERT(db->db_data_pending == NULL); | |
1844 | ||
1845 | kmem_cache_free(dbuf_cache, db); | |
d164b209 | 1846 | arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); |
34dc7c2f BB |
1847 | } |
1848 | ||
1849 | void | |
1850 | dbuf_prefetch(dnode_t *dn, uint64_t blkid) | |
1851 | { | |
1852 | dmu_buf_impl_t *db = NULL; | |
1853 | blkptr_t *bp = NULL; | |
1854 | ||
428870ff | 1855 | ASSERT(blkid != DMU_BONUS_BLKID); |
34dc7c2f BB |
1856 | ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); |
1857 | ||
1858 | if (dnode_block_freed(dn, blkid)) | |
1859 | return; | |
1860 | ||
1861 | /* dbuf_find() returns with db_mtx held */ | |
c65aa5b2 | 1862 | if ((db = dbuf_find(dn, 0, blkid))) { |
572e2857 BB |
1863 | /* |
1864 | * This dbuf is already in the cache. We assume that | |
1865 | * it is already CACHED, or else about to be either | |
1866 | * read or filled. | |
1867 | */ | |
34dc7c2f | 1868 | mutex_exit(&db->db_mtx); |
572e2857 | 1869 | return; |
34dc7c2f BB |
1870 | } |
1871 | ||
fc5bb51f | 1872 | if (dbuf_findbp(dn, 0, blkid, TRUE, &db, &bp, NULL) == 0) { |
34dc7c2f | 1873 | if (bp && !BP_IS_HOLE(bp)) { |
428870ff BB |
1874 | int priority = dn->dn_type == DMU_OT_DDT_ZAP ? |
1875 | ZIO_PRIORITY_DDT_PREFETCH : ZIO_PRIORITY_ASYNC_READ; | |
b128c09f | 1876 | arc_buf_t *pbuf; |
428870ff | 1877 | dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset; |
34dc7c2f BB |
1878 | uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH; |
1879 | zbookmark_t zb; | |
428870ff BB |
1880 | |
1881 | SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET, | |
1882 | dn->dn_object, 0, blkid); | |
34dc7c2f | 1883 | |
b128c09f BB |
1884 | if (db) |
1885 | pbuf = db->db_buf; | |
1886 | else | |
1887 | pbuf = dn->dn_objset->os_phys_buf; | |
1888 | ||
428870ff BB |
1889 | (void) dsl_read(NULL, dn->dn_objset->os_spa, |
1890 | bp, pbuf, NULL, NULL, priority, | |
34dc7c2f BB |
1891 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, |
1892 | &aflags, &zb); | |
1893 | } | |
1894 | if (db) | |
1895 | dbuf_rele(db, NULL); | |
1896 | } | |
1897 | } | |
1898 | ||
fc5bb51f BB |
1899 | #define DBUF_HOLD_IMPL_MAX_DEPTH 20 |
1900 | ||
34dc7c2f BB |
1901 | /* |
1902 | * Returns with db_holds incremented, and db_mtx not held. | |
1903 | * Note: dn_struct_rwlock must be held. | |
1904 | */ | |
fc5bb51f BB |
1905 | static int |
1906 | __dbuf_hold_impl(struct dbuf_hold_impl_data *dh) | |
34dc7c2f | 1907 | { |
fc5bb51f BB |
1908 | ASSERT3S(dh->dh_depth, <, DBUF_HOLD_IMPL_MAX_DEPTH); |
1909 | dh->dh_parent = NULL; | |
34dc7c2f | 1910 | |
fc5bb51f BB |
1911 | ASSERT(dh->dh_blkid != DMU_BONUS_BLKID); |
1912 | ASSERT(RW_LOCK_HELD(&dh->dh_dn->dn_struct_rwlock)); | |
1913 | ASSERT3U(dh->dh_dn->dn_nlevels, >, dh->dh_level); | |
34dc7c2f | 1914 | |
fc5bb51f | 1915 | *(dh->dh_dbp) = NULL; |
34dc7c2f BB |
1916 | top: |
1917 | /* dbuf_find() returns with db_mtx held */ | |
fc5bb51f BB |
1918 | dh->dh_db = dbuf_find(dh->dh_dn, dh->dh_level, dh->dh_blkid); |
1919 | ||
1920 | if (dh->dh_db == NULL) { | |
1921 | dh->dh_bp = NULL; | |
1922 | ||
1923 | ASSERT3P(dh->dh_parent, ==, NULL); | |
1924 | dh->dh_err = dbuf_findbp(dh->dh_dn, dh->dh_level, dh->dh_blkid, | |
1925 | dh->dh_fail_sparse, &dh->dh_parent, | |
1926 | &dh->dh_bp, dh); | |
1927 | if (dh->dh_fail_sparse) { | |
1928 | if (dh->dh_err == 0 && dh->dh_bp && BP_IS_HOLE(dh->dh_bp)) | |
1929 | dh->dh_err = ENOENT; | |
1930 | if (dh->dh_err) { | |
1931 | if (dh->dh_parent) | |
1932 | dbuf_rele(dh->dh_parent, NULL); | |
1933 | return (dh->dh_err); | |
34dc7c2f BB |
1934 | } |
1935 | } | |
fc5bb51f BB |
1936 | if (dh->dh_err && dh->dh_err != ENOENT) |
1937 | return (dh->dh_err); | |
1938 | dh->dh_db = dbuf_create(dh->dh_dn, dh->dh_level, dh->dh_blkid, | |
1939 | dh->dh_parent, dh->dh_bp); | |
34dc7c2f BB |
1940 | } |
1941 | ||
fc5bb51f BB |
1942 | if (dh->dh_db->db_buf && refcount_is_zero(&dh->dh_db->db_holds)) { |
1943 | arc_buf_add_ref(dh->dh_db->db_buf, dh->dh_db); | |
1944 | if (dh->dh_db->db_buf->b_data == NULL) { | |
1945 | dbuf_clear(dh->dh_db); | |
1946 | if (dh->dh_parent) { | |
1947 | dbuf_rele(dh->dh_parent, NULL); | |
1948 | dh->dh_parent = NULL; | |
34dc7c2f BB |
1949 | } |
1950 | goto top; | |
1951 | } | |
fc5bb51f | 1952 | ASSERT3P(dh->dh_db->db.db_data, ==, dh->dh_db->db_buf->b_data); |
34dc7c2f BB |
1953 | } |
1954 | ||
fc5bb51f | 1955 | ASSERT(dh->dh_db->db_buf == NULL || arc_referenced(dh->dh_db->db_buf)); |
34dc7c2f BB |
1956 | |
1957 | /* | |
1958 | * If this buffer is currently syncing out, and we are are | |
1959 | * still referencing it from db_data, we need to make a copy | |
1960 | * of it in case we decide we want to dirty it again in this txg. | |
1961 | */ | |
fc5bb51f BB |
1962 | if (dh->dh_db->db_level == 0 && |
1963 | dh->dh_db->db_blkid != DMU_BONUS_BLKID && | |
1964 | dh->dh_dn->dn_object != DMU_META_DNODE_OBJECT && | |
1965 | dh->dh_db->db_state == DB_CACHED && dh->dh_db->db_data_pending) { | |
1966 | dh->dh_dr = dh->dh_db->db_data_pending; | |
1967 | ||
1968 | if (dh->dh_dr->dt.dl.dr_data == dh->dh_db->db_buf) { | |
1969 | dh->dh_type = DBUF_GET_BUFC_TYPE(dh->dh_db); | |
1970 | ||
1971 | dbuf_set_data(dh->dh_db, | |
1972 | arc_buf_alloc(dh->dh_dn->dn_objset->os_spa, | |
1973 | dh->dh_db->db.db_size, dh->dh_db, dh->dh_type)); | |
1974 | bcopy(dh->dh_dr->dt.dl.dr_data->b_data, | |
1975 | dh->dh_db->db.db_data, dh->dh_db->db.db_size); | |
34dc7c2f BB |
1976 | } |
1977 | } | |
1978 | ||
fc5bb51f BB |
1979 | (void) refcount_add(&dh->dh_db->db_holds, dh->dh_tag); |
1980 | dbuf_update_data(dh->dh_db); | |
1981 | DBUF_VERIFY(dh->dh_db); | |
1982 | mutex_exit(&dh->dh_db->db_mtx); | |
34dc7c2f BB |
1983 | |
1984 | /* NOTE: we can't rele the parent until after we drop the db_mtx */ | |
fc5bb51f BB |
1985 | if (dh->dh_parent) |
1986 | dbuf_rele(dh->dh_parent, NULL); | |
34dc7c2f | 1987 | |
fc5bb51f BB |
1988 | ASSERT3P(DB_DNODE(dh->dh_db), ==, dh->dh_dn); |
1989 | ASSERT3U(dh->dh_db->db_blkid, ==, dh->dh_blkid); | |
1990 | ASSERT3U(dh->dh_db->db_level, ==, dh->dh_level); | |
1991 | *(dh->dh_dbp) = dh->dh_db; | |
34dc7c2f BB |
1992 | |
1993 | return (0); | |
1994 | } | |
1995 | ||
fc5bb51f BB |
1996 | /* |
1997 | * The following code preserves the recursive function dbuf_hold_impl() | |
1998 | * but moves the local variables AND function arguments to the heap to | |
1999 | * minimize the stack frame size. Enough space is initially allocated | |
2000 | * on the stack for 20 levels of recursion. | |
2001 | */ | |
2002 | int | |
2003 | dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse, | |
2004 | void *tag, dmu_buf_impl_t **dbp) | |
2005 | { | |
2006 | struct dbuf_hold_impl_data *dh; | |
2007 | int error; | |
2008 | ||
2009 | dh = kmem_zalloc(sizeof(struct dbuf_hold_impl_data) * | |
2010 | DBUF_HOLD_IMPL_MAX_DEPTH, KM_SLEEP); | |
2011 | __dbuf_hold_impl_init(dh, dn, level, blkid, fail_sparse, tag, dbp, 0); | |
2012 | ||
2013 | error = __dbuf_hold_impl(dh); | |
2014 | ||
2015 | kmem_free(dh, sizeof(struct dbuf_hold_impl_data) * | |
2016 | DBUF_HOLD_IMPL_MAX_DEPTH); | |
2017 | ||
2018 | return (error); | |
2019 | } | |
2020 | ||
2021 | static void | |
2022 | __dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh, | |
2023 | dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse, | |
2024 | void *tag, dmu_buf_impl_t **dbp, int depth) | |
2025 | { | |
2026 | dh->dh_dn = dn; | |
2027 | dh->dh_level = level; | |
2028 | dh->dh_blkid = blkid; | |
2029 | dh->dh_fail_sparse = fail_sparse; | |
2030 | dh->dh_tag = tag; | |
2031 | dh->dh_dbp = dbp; | |
2032 | dh->dh_depth = depth; | |
2033 | } | |
2034 | ||
34dc7c2f BB |
2035 | dmu_buf_impl_t * |
2036 | dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag) | |
2037 | { | |
2038 | dmu_buf_impl_t *db; | |
2039 | int err = dbuf_hold_impl(dn, 0, blkid, FALSE, tag, &db); | |
2040 | return (err ? NULL : db); | |
2041 | } | |
2042 | ||
2043 | dmu_buf_impl_t * | |
2044 | dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag) | |
2045 | { | |
2046 | dmu_buf_impl_t *db; | |
2047 | int err = dbuf_hold_impl(dn, level, blkid, FALSE, tag, &db); | |
2048 | return (err ? NULL : db); | |
2049 | } | |
2050 | ||
2051 | void | |
2052 | dbuf_create_bonus(dnode_t *dn) | |
2053 | { | |
2054 | ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); | |
2055 | ||
2056 | ASSERT(dn->dn_bonus == NULL); | |
428870ff BB |
2057 | dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL); |
2058 | } | |
2059 | ||
2060 | int | |
2061 | dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx) | |
2062 | { | |
2063 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
572e2857 BB |
2064 | dnode_t *dn; |
2065 | ||
428870ff BB |
2066 | if (db->db_blkid != DMU_SPILL_BLKID) |
2067 | return (ENOTSUP); | |
2068 | if (blksz == 0) | |
2069 | blksz = SPA_MINBLOCKSIZE; | |
2070 | if (blksz > SPA_MAXBLOCKSIZE) | |
2071 | blksz = SPA_MAXBLOCKSIZE; | |
2072 | else | |
2073 | blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE); | |
2074 | ||
572e2857 BB |
2075 | DB_DNODE_ENTER(db); |
2076 | dn = DB_DNODE(db); | |
2077 | rw_enter(&dn->dn_struct_rwlock, RW_WRITER); | |
428870ff | 2078 | dbuf_new_size(db, blksz, tx); |
572e2857 BB |
2079 | rw_exit(&dn->dn_struct_rwlock); |
2080 | DB_DNODE_EXIT(db); | |
428870ff BB |
2081 | |
2082 | return (0); | |
2083 | } | |
2084 | ||
2085 | void | |
2086 | dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx) | |
2087 | { | |
2088 | dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx); | |
34dc7c2f BB |
2089 | } |
2090 | ||
2091 | #pragma weak dmu_buf_add_ref = dbuf_add_ref | |
2092 | void | |
2093 | dbuf_add_ref(dmu_buf_impl_t *db, void *tag) | |
2094 | { | |
1fde1e37 | 2095 | VERIFY(refcount_add(&db->db_holds, tag) > 1); |
34dc7c2f BB |
2096 | } |
2097 | ||
572e2857 BB |
2098 | /* |
2099 | * If you call dbuf_rele() you had better not be referencing the dnode handle | |
2100 | * unless you have some other direct or indirect hold on the dnode. (An indirect | |
2101 | * hold is a hold on one of the dnode's dbufs, including the bonus buffer.) | |
2102 | * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the | |
2103 | * dnode's parent dbuf evicting its dnode handles. | |
2104 | */ | |
34dc7c2f BB |
2105 | #pragma weak dmu_buf_rele = dbuf_rele |
2106 | void | |
2107 | dbuf_rele(dmu_buf_impl_t *db, void *tag) | |
428870ff BB |
2108 | { |
2109 | mutex_enter(&db->db_mtx); | |
2110 | dbuf_rele_and_unlock(db, tag); | |
2111 | } | |
2112 | ||
2113 | /* | |
2114 | * dbuf_rele() for an already-locked dbuf. This is necessary to allow | |
2115 | * db_dirtycnt and db_holds to be updated atomically. | |
2116 | */ | |
2117 | void | |
2118 | dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag) | |
34dc7c2f BB |
2119 | { |
2120 | int64_t holds; | |
2121 | ||
428870ff | 2122 | ASSERT(MUTEX_HELD(&db->db_mtx)); |
34dc7c2f BB |
2123 | DBUF_VERIFY(db); |
2124 | ||
572e2857 BB |
2125 | /* |
2126 | * Remove the reference to the dbuf before removing its hold on the | |
2127 | * dnode so we can guarantee in dnode_move() that a referenced bonus | |
2128 | * buffer has a corresponding dnode hold. | |
2129 | */ | |
34dc7c2f BB |
2130 | holds = refcount_remove(&db->db_holds, tag); |
2131 | ASSERT(holds >= 0); | |
2132 | ||
2133 | /* | |
2134 | * We can't freeze indirects if there is a possibility that they | |
2135 | * may be modified in the current syncing context. | |
2136 | */ | |
2137 | if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) | |
2138 | arc_buf_freeze(db->db_buf); | |
2139 | ||
2140 | if (holds == db->db_dirtycnt && | |
2141 | db->db_level == 0 && db->db_immediate_evict) | |
2142 | dbuf_evict_user(db); | |
2143 | ||
2144 | if (holds == 0) { | |
428870ff | 2145 | if (db->db_blkid == DMU_BONUS_BLKID) { |
34dc7c2f | 2146 | mutex_exit(&db->db_mtx); |
572e2857 BB |
2147 | |
2148 | /* | |
2149 | * If the dnode moves here, we cannot cross this barrier | |
2150 | * until the move completes. | |
2151 | */ | |
2152 | DB_DNODE_ENTER(db); | |
2153 | (void) atomic_dec_32_nv(&DB_DNODE(db)->dn_dbufs_count); | |
2154 | DB_DNODE_EXIT(db); | |
2155 | /* | |
2156 | * The bonus buffer's dnode hold is no longer discounted | |
2157 | * in dnode_move(). The dnode cannot move until after | |
2158 | * the dnode_rele(). | |
2159 | */ | |
2160 | dnode_rele(DB_DNODE(db), db); | |
34dc7c2f BB |
2161 | } else if (db->db_buf == NULL) { |
2162 | /* | |
2163 | * This is a special case: we never associated this | |
2164 | * dbuf with any data allocated from the ARC. | |
2165 | */ | |
b128c09f BB |
2166 | ASSERT(db->db_state == DB_UNCACHED || |
2167 | db->db_state == DB_NOFILL); | |
34dc7c2f BB |
2168 | dbuf_evict(db); |
2169 | } else if (arc_released(db->db_buf)) { | |
2170 | arc_buf_t *buf = db->db_buf; | |
2171 | /* | |
2172 | * This dbuf has anonymous data associated with it. | |
2173 | */ | |
2174 | dbuf_set_data(db, NULL); | |
2175 | VERIFY(arc_buf_remove_ref(buf, db) == 1); | |
2176 | dbuf_evict(db); | |
2177 | } else { | |
2178 | VERIFY(arc_buf_remove_ref(db->db_buf, db) == 0); | |
b128c09f BB |
2179 | if (!DBUF_IS_CACHEABLE(db)) |
2180 | dbuf_clear(db); | |
2181 | else | |
2182 | mutex_exit(&db->db_mtx); | |
34dc7c2f BB |
2183 | } |
2184 | } else { | |
2185 | mutex_exit(&db->db_mtx); | |
2186 | } | |
2187 | } | |
2188 | ||
2189 | #pragma weak dmu_buf_refcount = dbuf_refcount | |
2190 | uint64_t | |
2191 | dbuf_refcount(dmu_buf_impl_t *db) | |
2192 | { | |
2193 | return (refcount_count(&db->db_holds)); | |
2194 | } | |
2195 | ||
2196 | void * | |
2197 | dmu_buf_set_user(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr, | |
2198 | dmu_buf_evict_func_t *evict_func) | |
2199 | { | |
2200 | return (dmu_buf_update_user(db_fake, NULL, user_ptr, | |
2201 | user_data_ptr_ptr, evict_func)); | |
2202 | } | |
2203 | ||
2204 | void * | |
2205 | dmu_buf_set_user_ie(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr, | |
2206 | dmu_buf_evict_func_t *evict_func) | |
2207 | { | |
2208 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
2209 | ||
2210 | db->db_immediate_evict = TRUE; | |
2211 | return (dmu_buf_update_user(db_fake, NULL, user_ptr, | |
2212 | user_data_ptr_ptr, evict_func)); | |
2213 | } | |
2214 | ||
2215 | void * | |
2216 | dmu_buf_update_user(dmu_buf_t *db_fake, void *old_user_ptr, void *user_ptr, | |
2217 | void *user_data_ptr_ptr, dmu_buf_evict_func_t *evict_func) | |
2218 | { | |
2219 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
2220 | ASSERT(db->db_level == 0); | |
2221 | ||
2222 | ASSERT((user_ptr == NULL) == (evict_func == NULL)); | |
2223 | ||
2224 | mutex_enter(&db->db_mtx); | |
2225 | ||
2226 | if (db->db_user_ptr == old_user_ptr) { | |
2227 | db->db_user_ptr = user_ptr; | |
2228 | db->db_user_data_ptr_ptr = user_data_ptr_ptr; | |
2229 | db->db_evict_func = evict_func; | |
2230 | ||
2231 | dbuf_update_data(db); | |
2232 | } else { | |
2233 | old_user_ptr = db->db_user_ptr; | |
2234 | } | |
2235 | ||
2236 | mutex_exit(&db->db_mtx); | |
2237 | return (old_user_ptr); | |
2238 | } | |
2239 | ||
2240 | void * | |
2241 | dmu_buf_get_user(dmu_buf_t *db_fake) | |
2242 | { | |
2243 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
2244 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
2245 | ||
2246 | return (db->db_user_ptr); | |
2247 | } | |
2248 | ||
9babb374 BB |
2249 | boolean_t |
2250 | dmu_buf_freeable(dmu_buf_t *dbuf) | |
2251 | { | |
2252 | boolean_t res = B_FALSE; | |
2253 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; | |
2254 | ||
2255 | if (db->db_blkptr) | |
2256 | res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset, | |
428870ff | 2257 | db->db_blkptr, db->db_blkptr->blk_birth); |
9babb374 BB |
2258 | |
2259 | return (res); | |
2260 | } | |
2261 | ||
34dc7c2f BB |
2262 | static void |
2263 | dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db) | |
2264 | { | |
2265 | /* ASSERT(dmu_tx_is_syncing(tx) */ | |
2266 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
2267 | ||
2268 | if (db->db_blkptr != NULL) | |
2269 | return; | |
2270 | ||
428870ff BB |
2271 | if (db->db_blkid == DMU_SPILL_BLKID) { |
2272 | db->db_blkptr = &dn->dn_phys->dn_spill; | |
2273 | BP_ZERO(db->db_blkptr); | |
2274 | return; | |
2275 | } | |
34dc7c2f BB |
2276 | if (db->db_level == dn->dn_phys->dn_nlevels-1) { |
2277 | /* | |
2278 | * This buffer was allocated at a time when there was | |
2279 | * no available blkptrs from the dnode, or it was | |
2280 | * inappropriate to hook it in (i.e., nlevels mis-match). | |
2281 | */ | |
2282 | ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr); | |
2283 | ASSERT(db->db_parent == NULL); | |
2284 | db->db_parent = dn->dn_dbuf; | |
2285 | db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid]; | |
2286 | DBUF_VERIFY(db); | |
2287 | } else { | |
2288 | dmu_buf_impl_t *parent = db->db_parent; | |
2289 | int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; | |
2290 | ||
2291 | ASSERT(dn->dn_phys->dn_nlevels > 1); | |
2292 | if (parent == NULL) { | |
2293 | mutex_exit(&db->db_mtx); | |
2294 | rw_enter(&dn->dn_struct_rwlock, RW_READER); | |
2295 | (void) dbuf_hold_impl(dn, db->db_level+1, | |
2296 | db->db_blkid >> epbs, FALSE, db, &parent); | |
2297 | rw_exit(&dn->dn_struct_rwlock); | |
2298 | mutex_enter(&db->db_mtx); | |
2299 | db->db_parent = parent; | |
2300 | } | |
2301 | db->db_blkptr = (blkptr_t *)parent->db.db_data + | |
2302 | (db->db_blkid & ((1ULL << epbs) - 1)); | |
2303 | DBUF_VERIFY(db); | |
2304 | } | |
2305 | } | |
2306 | ||
60948de1 BB |
2307 | /* dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it |
2308 | * is critical the we not allow the compiler to inline this function in to | |
2309 | * dbuf_sync_list() thereby drastically bloating the stack usage. | |
2310 | */ | |
2311 | noinline static void | |
34dc7c2f BB |
2312 | dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx) |
2313 | { | |
2314 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
572e2857 | 2315 | dnode_t *dn; |
34dc7c2f BB |
2316 | zio_t *zio; |
2317 | ||
2318 | ASSERT(dmu_tx_is_syncing(tx)); | |
2319 | ||
2320 | dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); | |
2321 | ||
2322 | mutex_enter(&db->db_mtx); | |
2323 | ||
2324 | ASSERT(db->db_level > 0); | |
2325 | DBUF_VERIFY(db); | |
2326 | ||
2327 | if (db->db_buf == NULL) { | |
2328 | mutex_exit(&db->db_mtx); | |
2329 | (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED); | |
2330 | mutex_enter(&db->db_mtx); | |
2331 | } | |
2332 | ASSERT3U(db->db_state, ==, DB_CACHED); | |
34dc7c2f BB |
2333 | ASSERT(db->db_buf != NULL); |
2334 | ||
572e2857 BB |
2335 | DB_DNODE_ENTER(db); |
2336 | dn = DB_DNODE(db); | |
2337 | ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); | |
34dc7c2f | 2338 | dbuf_check_blkptr(dn, db); |
572e2857 | 2339 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
2340 | |
2341 | db->db_data_pending = dr; | |
2342 | ||
34dc7c2f | 2343 | mutex_exit(&db->db_mtx); |
b128c09f | 2344 | dbuf_write(dr, db->db_buf, tx); |
34dc7c2f BB |
2345 | |
2346 | zio = dr->dr_zio; | |
2347 | mutex_enter(&dr->dt.di.dr_mtx); | |
2348 | dbuf_sync_list(&dr->dt.di.dr_children, tx); | |
2349 | ASSERT(list_head(&dr->dt.di.dr_children) == NULL); | |
2350 | mutex_exit(&dr->dt.di.dr_mtx); | |
2351 | zio_nowait(zio); | |
2352 | } | |
2353 | ||
60948de1 BB |
2354 | /* dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is |
2355 | * critical the we not allow the compiler to inline this function in to | |
2356 | * dbuf_sync_list() thereby drastically bloating the stack usage. | |
2357 | */ | |
2358 | noinline static void | |
34dc7c2f BB |
2359 | dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx) |
2360 | { | |
2361 | arc_buf_t **datap = &dr->dt.dl.dr_data; | |
2362 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
572e2857 BB |
2363 | dnode_t *dn; |
2364 | objset_t *os; | |
34dc7c2f | 2365 | uint64_t txg = tx->tx_txg; |
34dc7c2f BB |
2366 | |
2367 | ASSERT(dmu_tx_is_syncing(tx)); | |
2368 | ||
2369 | dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); | |
2370 | ||
2371 | mutex_enter(&db->db_mtx); | |
2372 | /* | |
2373 | * To be synced, we must be dirtied. But we | |
2374 | * might have been freed after the dirty. | |
2375 | */ | |
2376 | if (db->db_state == DB_UNCACHED) { | |
2377 | /* This buffer has been freed since it was dirtied */ | |
2378 | ASSERT(db->db.db_data == NULL); | |
2379 | } else if (db->db_state == DB_FILL) { | |
2380 | /* This buffer was freed and is now being re-filled */ | |
2381 | ASSERT(db->db.db_data != dr->dt.dl.dr_data); | |
2382 | } else { | |
b128c09f | 2383 | ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL); |
34dc7c2f BB |
2384 | } |
2385 | DBUF_VERIFY(db); | |
2386 | ||
572e2857 BB |
2387 | DB_DNODE_ENTER(db); |
2388 | dn = DB_DNODE(db); | |
2389 | ||
428870ff BB |
2390 | if (db->db_blkid == DMU_SPILL_BLKID) { |
2391 | mutex_enter(&dn->dn_mtx); | |
2392 | dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR; | |
2393 | mutex_exit(&dn->dn_mtx); | |
2394 | } | |
2395 | ||
34dc7c2f BB |
2396 | /* |
2397 | * If this is a bonus buffer, simply copy the bonus data into the | |
2398 | * dnode. It will be written out when the dnode is synced (and it | |
2399 | * will be synced, since it must have been dirty for dbuf_sync to | |
2400 | * be called). | |
2401 | */ | |
428870ff | 2402 | if (db->db_blkid == DMU_BONUS_BLKID) { |
34dc7c2f BB |
2403 | dbuf_dirty_record_t **drp; |
2404 | ||
2405 | ASSERT(*datap != NULL); | |
2406 | ASSERT3U(db->db_level, ==, 0); | |
2407 | ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN); | |
2408 | bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen); | |
572e2857 BB |
2409 | DB_DNODE_EXIT(db); |
2410 | ||
34dc7c2f BB |
2411 | if (*datap != db->db.db_data) { |
2412 | zio_buf_free(*datap, DN_MAX_BONUSLEN); | |
d164b209 | 2413 | arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); |
34dc7c2f BB |
2414 | } |
2415 | db->db_data_pending = NULL; | |
2416 | drp = &db->db_last_dirty; | |
2417 | while (*drp != dr) | |
2418 | drp = &(*drp)->dr_next; | |
2419 | ASSERT(dr->dr_next == NULL); | |
428870ff | 2420 | ASSERT(dr->dr_dbuf == db); |
34dc7c2f | 2421 | *drp = dr->dr_next; |
753972fc BB |
2422 | if (dr->dr_dbuf->db_level != 0) { |
2423 | mutex_destroy(&dr->dt.di.dr_mtx); | |
2424 | list_destroy(&dr->dt.di.dr_children); | |
2425 | } | |
34dc7c2f BB |
2426 | kmem_free(dr, sizeof (dbuf_dirty_record_t)); |
2427 | ASSERT(db->db_dirtycnt > 0); | |
2428 | db->db_dirtycnt -= 1; | |
428870ff | 2429 | dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg); |
34dc7c2f BB |
2430 | return; |
2431 | } | |
2432 | ||
572e2857 BB |
2433 | os = dn->dn_objset; |
2434 | ||
34dc7c2f BB |
2435 | /* |
2436 | * This function may have dropped the db_mtx lock allowing a dmu_sync | |
2437 | * operation to sneak in. As a result, we need to ensure that we | |
2438 | * don't check the dr_override_state until we have returned from | |
2439 | * dbuf_check_blkptr. | |
2440 | */ | |
2441 | dbuf_check_blkptr(dn, db); | |
2442 | ||
2443 | /* | |
572e2857 | 2444 | * If this buffer is in the middle of an immediate write, |
34dc7c2f BB |
2445 | * wait for the synchronous IO to complete. |
2446 | */ | |
2447 | while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) { | |
2448 | ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT); | |
2449 | cv_wait(&db->db_changed, &db->db_mtx); | |
2450 | ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN); | |
2451 | } | |
2452 | ||
9babb374 BB |
2453 | if (db->db_state != DB_NOFILL && |
2454 | dn->dn_object != DMU_META_DNODE_OBJECT && | |
2455 | refcount_count(&db->db_holds) > 1 && | |
428870ff | 2456 | dr->dt.dl.dr_override_state != DR_OVERRIDDEN && |
9babb374 BB |
2457 | *datap == db->db_buf) { |
2458 | /* | |
2459 | * If this buffer is currently "in use" (i.e., there | |
2460 | * are active holds and db_data still references it), | |
2461 | * then make a copy before we start the write so that | |
2462 | * any modifications from the open txg will not leak | |
2463 | * into this write. | |
2464 | * | |
2465 | * NOTE: this copy does not need to be made for | |
2466 | * objects only modified in the syncing context (e.g. | |
2467 | * DNONE_DNODE blocks). | |
2468 | */ | |
2469 | int blksz = arc_buf_size(*datap); | |
2470 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); | |
2471 | *datap = arc_buf_alloc(os->os_spa, blksz, db, type); | |
2472 | bcopy(db->db.db_data, (*datap)->b_data, blksz); | |
b128c09f | 2473 | } |
34dc7c2f BB |
2474 | db->db_data_pending = dr; |
2475 | ||
2476 | mutex_exit(&db->db_mtx); | |
2477 | ||
b128c09f | 2478 | dbuf_write(dr, *datap, tx); |
34dc7c2f BB |
2479 | |
2480 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
572e2857 | 2481 | if (dn->dn_object == DMU_META_DNODE_OBJECT) { |
34dc7c2f | 2482 | list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr); |
572e2857 BB |
2483 | DB_DNODE_EXIT(db); |
2484 | } else { | |
2485 | /* | |
2486 | * Although zio_nowait() does not "wait for an IO", it does | |
2487 | * initiate the IO. If this is an empty write it seems plausible | |
2488 | * that the IO could actually be completed before the nowait | |
2489 | * returns. We need to DB_DNODE_EXIT() first in case | |
2490 | * zio_nowait() invalidates the dbuf. | |
2491 | */ | |
2492 | DB_DNODE_EXIT(db); | |
34dc7c2f | 2493 | zio_nowait(dr->dr_zio); |
572e2857 | 2494 | } |
34dc7c2f BB |
2495 | } |
2496 | ||
2497 | void | |
2498 | dbuf_sync_list(list_t *list, dmu_tx_t *tx) | |
2499 | { | |
2500 | dbuf_dirty_record_t *dr; | |
2501 | ||
c65aa5b2 | 2502 | while ((dr = list_head(list))) { |
34dc7c2f BB |
2503 | if (dr->dr_zio != NULL) { |
2504 | /* | |
2505 | * If we find an already initialized zio then we | |
2506 | * are processing the meta-dnode, and we have finished. | |
2507 | * The dbufs for all dnodes are put back on the list | |
2508 | * during processing, so that we can zio_wait() | |
2509 | * these IOs after initiating all child IOs. | |
2510 | */ | |
2511 | ASSERT3U(dr->dr_dbuf->db.db_object, ==, | |
2512 | DMU_META_DNODE_OBJECT); | |
2513 | break; | |
2514 | } | |
2515 | list_remove(list, dr); | |
2516 | if (dr->dr_dbuf->db_level > 0) | |
2517 | dbuf_sync_indirect(dr, tx); | |
2518 | else | |
2519 | dbuf_sync_leaf(dr, tx); | |
2520 | } | |
2521 | } | |
2522 | ||
34dc7c2f BB |
2523 | /* ARGSUSED */ |
2524 | static void | |
2525 | dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb) | |
2526 | { | |
2527 | dmu_buf_impl_t *db = vdb; | |
572e2857 | 2528 | dnode_t *dn; |
b128c09f | 2529 | blkptr_t *bp = zio->io_bp; |
34dc7c2f | 2530 | blkptr_t *bp_orig = &zio->io_bp_orig; |
428870ff BB |
2531 | spa_t *spa = zio->io_spa; |
2532 | int64_t delta; | |
34dc7c2f | 2533 | uint64_t fill = 0; |
428870ff | 2534 | int i; |
34dc7c2f | 2535 | |
b128c09f BB |
2536 | ASSERT(db->db_blkptr == bp); |
2537 | ||
572e2857 BB |
2538 | DB_DNODE_ENTER(db); |
2539 | dn = DB_DNODE(db); | |
428870ff BB |
2540 | delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig); |
2541 | dnode_diduse_space(dn, delta - zio->io_prev_space_delta); | |
2542 | zio->io_prev_space_delta = delta; | |
34dc7c2f | 2543 | |
b128c09f | 2544 | if (BP_IS_HOLE(bp)) { |
428870ff | 2545 | ASSERT(bp->blk_fill == 0); |
572e2857 | 2546 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
2547 | return; |
2548 | } | |
2549 | ||
428870ff BB |
2550 | ASSERT((db->db_blkid != DMU_SPILL_BLKID && |
2551 | BP_GET_TYPE(bp) == dn->dn_type) || | |
2552 | (db->db_blkid == DMU_SPILL_BLKID && | |
2553 | BP_GET_TYPE(bp) == dn->dn_bonustype)); | |
b128c09f BB |
2554 | ASSERT(BP_GET_LEVEL(bp) == db->db_level); |
2555 | ||
34dc7c2f BB |
2556 | mutex_enter(&db->db_mtx); |
2557 | ||
428870ff BB |
2558 | #ifdef ZFS_DEBUG |
2559 | if (db->db_blkid == DMU_SPILL_BLKID) { | |
428870ff BB |
2560 | ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); |
2561 | ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && | |
2562 | db->db_blkptr == &dn->dn_phys->dn_spill); | |
2563 | } | |
2564 | #endif | |
2565 | ||
34dc7c2f BB |
2566 | if (db->db_level == 0) { |
2567 | mutex_enter(&dn->dn_mtx); | |
428870ff BB |
2568 | if (db->db_blkid > dn->dn_phys->dn_maxblkid && |
2569 | db->db_blkid != DMU_SPILL_BLKID) | |
34dc7c2f BB |
2570 | dn->dn_phys->dn_maxblkid = db->db_blkid; |
2571 | mutex_exit(&dn->dn_mtx); | |
2572 | ||
2573 | if (dn->dn_type == DMU_OT_DNODE) { | |
2574 | dnode_phys_t *dnp = db->db.db_data; | |
2575 | for (i = db->db.db_size >> DNODE_SHIFT; i > 0; | |
2576 | i--, dnp++) { | |
2577 | if (dnp->dn_type != DMU_OT_NONE) | |
2578 | fill++; | |
2579 | } | |
2580 | } else { | |
2581 | fill = 1; | |
2582 | } | |
2583 | } else { | |
b128c09f | 2584 | blkptr_t *ibp = db->db.db_data; |
34dc7c2f | 2585 | ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); |
b128c09f BB |
2586 | for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) { |
2587 | if (BP_IS_HOLE(ibp)) | |
34dc7c2f | 2588 | continue; |
b128c09f | 2589 | fill += ibp->blk_fill; |
34dc7c2f BB |
2590 | } |
2591 | } | |
572e2857 | 2592 | DB_DNODE_EXIT(db); |
34dc7c2f | 2593 | |
b128c09f | 2594 | bp->blk_fill = fill; |
34dc7c2f BB |
2595 | |
2596 | mutex_exit(&db->db_mtx); | |
34dc7c2f BB |
2597 | } |
2598 | ||
2599 | /* ARGSUSED */ | |
2600 | static void | |
2601 | dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb) | |
2602 | { | |
2603 | dmu_buf_impl_t *db = vdb; | |
428870ff BB |
2604 | blkptr_t *bp = zio->io_bp; |
2605 | blkptr_t *bp_orig = &zio->io_bp_orig; | |
34dc7c2f BB |
2606 | uint64_t txg = zio->io_txg; |
2607 | dbuf_dirty_record_t **drp, *dr; | |
2608 | ||
2609 | ASSERT3U(zio->io_error, ==, 0); | |
428870ff BB |
2610 | ASSERT(db->db_blkptr == bp); |
2611 | ||
2612 | if (zio->io_flags & ZIO_FLAG_IO_REWRITE) { | |
2613 | ASSERT(BP_EQUAL(bp, bp_orig)); | |
2614 | } else { | |
572e2857 BB |
2615 | objset_t *os; |
2616 | dsl_dataset_t *ds; | |
2617 | dmu_tx_t *tx; | |
2618 | ||
2619 | DB_GET_OBJSET(&os, db); | |
2620 | ds = os->os_dsl_dataset; | |
2621 | tx = os->os_synctx; | |
428870ff BB |
2622 | |
2623 | (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE); | |
2624 | dsl_dataset_block_born(ds, bp, tx); | |
2625 | } | |
34dc7c2f BB |
2626 | |
2627 | mutex_enter(&db->db_mtx); | |
2628 | ||
428870ff BB |
2629 | DBUF_VERIFY(db); |
2630 | ||
34dc7c2f BB |
2631 | drp = &db->db_last_dirty; |
2632 | while ((dr = *drp) != db->db_data_pending) | |
2633 | drp = &dr->dr_next; | |
2634 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
2635 | ASSERT(dr->dr_txg == txg); | |
428870ff | 2636 | ASSERT(dr->dr_dbuf == db); |
34dc7c2f BB |
2637 | ASSERT(dr->dr_next == NULL); |
2638 | *drp = dr->dr_next; | |
2639 | ||
428870ff BB |
2640 | #ifdef ZFS_DEBUG |
2641 | if (db->db_blkid == DMU_SPILL_BLKID) { | |
572e2857 BB |
2642 | dnode_t *dn; |
2643 | ||
2644 | DB_DNODE_ENTER(db); | |
2645 | dn = DB_DNODE(db); | |
428870ff BB |
2646 | ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); |
2647 | ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && | |
2648 | db->db_blkptr == &dn->dn_phys->dn_spill); | |
572e2857 | 2649 | DB_DNODE_EXIT(db); |
428870ff BB |
2650 | } |
2651 | #endif | |
2652 | ||
34dc7c2f | 2653 | if (db->db_level == 0) { |
428870ff | 2654 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f | 2655 | ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); |
b128c09f BB |
2656 | if (db->db_state != DB_NOFILL) { |
2657 | if (dr->dt.dl.dr_data != db->db_buf) | |
2658 | VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, | |
2659 | db) == 1); | |
428870ff | 2660 | else if (!arc_released(db->db_buf)) |
b128c09f | 2661 | arc_set_callback(db->db_buf, dbuf_do_evict, db); |
b128c09f | 2662 | } |
34dc7c2f | 2663 | } else { |
572e2857 BB |
2664 | dnode_t *dn; |
2665 | ||
2666 | DB_DNODE_ENTER(db); | |
2667 | dn = DB_DNODE(db); | |
34dc7c2f BB |
2668 | ASSERT(list_head(&dr->dt.di.dr_children) == NULL); |
2669 | ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); | |
2670 | if (!BP_IS_HOLE(db->db_blkptr)) { | |
1fde1e37 BB |
2671 | ASSERTV(int epbs = dn->dn_phys->dn_indblkshift - |
2672 | SPA_BLKPTRSHIFT); | |
34dc7c2f BB |
2673 | ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, |
2674 | db->db.db_size); | |
2675 | ASSERT3U(dn->dn_phys->dn_maxblkid | |
2676 | >> (db->db_level * epbs), >=, db->db_blkid); | |
2677 | arc_set_callback(db->db_buf, dbuf_do_evict, db); | |
2678 | } | |
572e2857 | 2679 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
2680 | mutex_destroy(&dr->dt.di.dr_mtx); |
2681 | list_destroy(&dr->dt.di.dr_children); | |
2682 | } | |
2683 | kmem_free(dr, sizeof (dbuf_dirty_record_t)); | |
2684 | ||
2685 | cv_broadcast(&db->db_changed); | |
2686 | ASSERT(db->db_dirtycnt > 0); | |
2687 | db->db_dirtycnt -= 1; | |
2688 | db->db_data_pending = NULL; | |
428870ff BB |
2689 | dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg); |
2690 | } | |
2691 | ||
2692 | static void | |
2693 | dbuf_write_nofill_ready(zio_t *zio) | |
2694 | { | |
2695 | dbuf_write_ready(zio, NULL, zio->io_private); | |
2696 | } | |
2697 | ||
2698 | static void | |
2699 | dbuf_write_nofill_done(zio_t *zio) | |
2700 | { | |
2701 | dbuf_write_done(zio, NULL, zio->io_private); | |
2702 | } | |
2703 | ||
2704 | static void | |
2705 | dbuf_write_override_ready(zio_t *zio) | |
2706 | { | |
2707 | dbuf_dirty_record_t *dr = zio->io_private; | |
2708 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
2709 | ||
2710 | dbuf_write_ready(zio, NULL, db); | |
2711 | } | |
2712 | ||
2713 | static void | |
2714 | dbuf_write_override_done(zio_t *zio) | |
2715 | { | |
2716 | dbuf_dirty_record_t *dr = zio->io_private; | |
2717 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
2718 | blkptr_t *obp = &dr->dt.dl.dr_overridden_by; | |
2719 | ||
2720 | mutex_enter(&db->db_mtx); | |
2721 | if (!BP_EQUAL(zio->io_bp, obp)) { | |
2722 | if (!BP_IS_HOLE(obp)) | |
2723 | dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp); | |
2724 | arc_release(dr->dt.dl.dr_data, db); | |
2725 | } | |
34dc7c2f BB |
2726 | mutex_exit(&db->db_mtx); |
2727 | ||
428870ff BB |
2728 | dbuf_write_done(zio, NULL, db); |
2729 | } | |
2730 | ||
2731 | static void | |
2732 | dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx) | |
2733 | { | |
2734 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
572e2857 BB |
2735 | dnode_t *dn; |
2736 | objset_t *os; | |
428870ff BB |
2737 | dmu_buf_impl_t *parent = db->db_parent; |
2738 | uint64_t txg = tx->tx_txg; | |
2739 | zbookmark_t zb; | |
2740 | zio_prop_t zp; | |
2741 | zio_t *zio; | |
2742 | int wp_flag = 0; | |
34dc7c2f | 2743 | |
572e2857 BB |
2744 | DB_DNODE_ENTER(db); |
2745 | dn = DB_DNODE(db); | |
2746 | os = dn->dn_objset; | |
2747 | ||
428870ff BB |
2748 | if (db->db_state != DB_NOFILL) { |
2749 | if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) { | |
2750 | /* | |
2751 | * Private object buffers are released here rather | |
2752 | * than in dbuf_dirty() since they are only modified | |
2753 | * in the syncing context and we don't want the | |
2754 | * overhead of making multiple copies of the data. | |
2755 | */ | |
2756 | if (BP_IS_HOLE(db->db_blkptr)) { | |
2757 | arc_buf_thaw(data); | |
2758 | } else { | |
2759 | dbuf_release_bp(db); | |
2760 | } | |
2761 | } | |
2762 | } | |
2763 | ||
2764 | if (parent != dn->dn_dbuf) { | |
2765 | ASSERT(parent && parent->db_data_pending); | |
2766 | ASSERT(db->db_level == parent->db_level-1); | |
2767 | ASSERT(arc_released(parent->db_buf)); | |
2768 | zio = parent->db_data_pending->dr_zio; | |
2769 | } else { | |
2770 | ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 && | |
2771 | db->db_blkid != DMU_SPILL_BLKID) || | |
2772 | (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0)); | |
2773 | if (db->db_blkid != DMU_SPILL_BLKID) | |
2774 | ASSERT3P(db->db_blkptr, ==, | |
2775 | &dn->dn_phys->dn_blkptr[db->db_blkid]); | |
2776 | zio = dn->dn_zio; | |
2777 | } | |
2778 | ||
2779 | ASSERT(db->db_level == 0 || data == db->db_buf); | |
2780 | ASSERT3U(db->db_blkptr->blk_birth, <=, txg); | |
2781 | ASSERT(zio); | |
2782 | ||
2783 | SET_BOOKMARK(&zb, os->os_dsl_dataset ? | |
2784 | os->os_dsl_dataset->ds_object : DMU_META_OBJSET, | |
2785 | db->db.db_object, db->db_level, db->db_blkid); | |
2786 | ||
2787 | if (db->db_blkid == DMU_SPILL_BLKID) | |
2788 | wp_flag = WP_SPILL; | |
2789 | wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0; | |
2790 | ||
2791 | dmu_write_policy(os, dn, db->db_level, wp_flag, &zp); | |
572e2857 | 2792 | DB_DNODE_EXIT(db); |
428870ff BB |
2793 | |
2794 | if (db->db_level == 0 && dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { | |
2795 | ASSERT(db->db_state != DB_NOFILL); | |
2796 | dr->dr_zio = zio_write(zio, os->os_spa, txg, | |
2797 | db->db_blkptr, data->b_data, arc_buf_size(data), &zp, | |
2798 | dbuf_write_override_ready, dbuf_write_override_done, dr, | |
2799 | ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); | |
2800 | mutex_enter(&db->db_mtx); | |
2801 | dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; | |
2802 | zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by, | |
2803 | dr->dt.dl.dr_copies); | |
2804 | mutex_exit(&db->db_mtx); | |
2805 | } else if (db->db_state == DB_NOFILL) { | |
2806 | ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF); | |
2807 | dr->dr_zio = zio_write(zio, os->os_spa, txg, | |
2808 | db->db_blkptr, NULL, db->db.db_size, &zp, | |
2809 | dbuf_write_nofill_ready, dbuf_write_nofill_done, db, | |
2810 | ZIO_PRIORITY_ASYNC_WRITE, | |
2811 | ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb); | |
2812 | } else { | |
2813 | ASSERT(arc_released(data)); | |
2814 | dr->dr_zio = arc_write(zio, os->os_spa, txg, | |
2815 | db->db_blkptr, data, DBUF_IS_L2CACHEABLE(db), &zp, | |
2816 | dbuf_write_ready, dbuf_write_done, db, | |
2817 | ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); | |
2818 | } | |
34dc7c2f | 2819 | } |
c28b2279 BB |
2820 | |
2821 | #if defined(_KERNEL) && defined(HAVE_SPL) | |
2822 | EXPORT_SYMBOL(dmu_buf_rele); | |
2823 | EXPORT_SYMBOL(dmu_buf_will_dirty); | |
2824 | #endif |