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