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