]>
Commit | Line | Data |
---|---|---|
34dc7c2f BB |
1 | /* |
2 | * CDDL HEADER START | |
3 | * | |
4 | * The contents of this file are subject to the terms of the | |
5 | * Common Development and Distribution License (the "License"). | |
6 | * You may not use this file except in compliance with the License. | |
7 | * | |
8 | * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE | |
9 | * or http://www.opensolaris.org/os/licensing. | |
10 | * See the License for the specific language governing permissions | |
11 | * and limitations under the License. | |
12 | * | |
13 | * When distributing Covered Code, include this CDDL HEADER in each | |
14 | * file and include the License file at usr/src/OPENSOLARIS.LICENSE. | |
15 | * If applicable, add the following below this CDDL HEADER, with the | |
16 | * fields enclosed by brackets "[]" replaced with your own identifying | |
17 | * information: Portions Copyright [yyyy] [name of copyright owner] | |
18 | * | |
19 | * CDDL HEADER END | |
20 | */ | |
21 | /* | |
428870ff | 22 | * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. |
ef3c1dea | 23 | * Copyright 2011 Nexenta Systems, Inc. All rights reserved. |
9bd274dd | 24 | * Copyright (c) 2012, 2014 by Delphix. All rights reserved. |
3a17a7a9 | 25 | * Copyright (c) 2013 by Saso Kiselkov. All rights reserved. |
34dc7c2f BB |
26 | */ |
27 | ||
34dc7c2f | 28 | #include <sys/zfs_context.h> |
c28b2279 | 29 | #include <sys/arc.h> |
34dc7c2f | 30 | #include <sys/dmu.h> |
ea97f8ce | 31 | #include <sys/dmu_send.h> |
34dc7c2f BB |
32 | #include <sys/dmu_impl.h> |
33 | #include <sys/dbuf.h> | |
34 | #include <sys/dmu_objset.h> | |
35 | #include <sys/dsl_dataset.h> | |
36 | #include <sys/dsl_dir.h> | |
37 | #include <sys/dmu_tx.h> | |
38 | #include <sys/spa.h> | |
39 | #include <sys/zio.h> | |
40 | #include <sys/dmu_zfetch.h> | |
428870ff BB |
41 | #include <sys/sa.h> |
42 | #include <sys/sa_impl.h> | |
9b67f605 MA |
43 | #include <sys/zfeature.h> |
44 | #include <sys/blkptr.h> | |
9bd274dd | 45 | #include <sys/range_tree.h> |
34dc7c2f | 46 | |
fc5bb51f BB |
47 | struct dbuf_hold_impl_data { |
48 | /* Function arguments */ | |
49 | dnode_t *dh_dn; | |
50 | uint8_t dh_level; | |
51 | uint64_t dh_blkid; | |
52 | int dh_fail_sparse; | |
53 | void *dh_tag; | |
54 | dmu_buf_impl_t **dh_dbp; | |
55 | /* Local variables */ | |
56 | dmu_buf_impl_t *dh_db; | |
57 | dmu_buf_impl_t *dh_parent; | |
58 | blkptr_t *dh_bp; | |
59 | int dh_err; | |
60 | dbuf_dirty_record_t *dh_dr; | |
61 | arc_buf_contents_t dh_type; | |
62 | int dh_depth; | |
63 | }; | |
64 | ||
65 | static void __dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh, | |
66 | dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse, | |
67 | void *tag, dmu_buf_impl_t **dbp, int depth); | |
68 | static int __dbuf_hold_impl(struct dbuf_hold_impl_data *dh); | |
69 | ||
b663a23d MA |
70 | /* |
71 | * Number of times that zfs_free_range() took the slow path while doing | |
72 | * a zfs receive. A nonzero value indicates a potential performance problem. | |
73 | */ | |
74 | uint64_t zfs_free_range_recv_miss; | |
75 | ||
34dc7c2f | 76 | static void dbuf_destroy(dmu_buf_impl_t *db); |
13fe0198 | 77 | static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx); |
b128c09f | 78 | static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx); |
34dc7c2f BB |
79 | |
80 | /* | |
81 | * Global data structures and functions for the dbuf cache. | |
82 | */ | |
83 | static kmem_cache_t *dbuf_cache; | |
84 | ||
85 | /* ARGSUSED */ | |
86 | static int | |
87 | dbuf_cons(void *vdb, void *unused, int kmflag) | |
88 | { | |
89 | dmu_buf_impl_t *db = vdb; | |
90 | bzero(db, sizeof (dmu_buf_impl_t)); | |
91 | ||
92 | mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL); | |
93 | cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL); | |
94 | refcount_create(&db->db_holds); | |
98f72a53 | 95 | list_link_init(&db->db_link); |
34dc7c2f BB |
96 | return (0); |
97 | } | |
98 | ||
99 | /* ARGSUSED */ | |
100 | static void | |
101 | dbuf_dest(void *vdb, void *unused) | |
102 | { | |
103 | dmu_buf_impl_t *db = vdb; | |
104 | mutex_destroy(&db->db_mtx); | |
105 | cv_destroy(&db->db_changed); | |
106 | refcount_destroy(&db->db_holds); | |
107 | } | |
108 | ||
109 | /* | |
110 | * dbuf hash table routines | |
111 | */ | |
112 | static dbuf_hash_table_t dbuf_hash_table; | |
113 | ||
114 | static uint64_t dbuf_hash_count; | |
115 | ||
116 | static uint64_t | |
117 | dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid) | |
118 | { | |
119 | uintptr_t osv = (uintptr_t)os; | |
120 | uint64_t crc = -1ULL; | |
121 | ||
122 | ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); | |
123 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF]; | |
124 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF]; | |
125 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF]; | |
126 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF]; | |
127 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF]; | |
128 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF]; | |
129 | ||
130 | crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16); | |
131 | ||
132 | return (crc); | |
133 | } | |
134 | ||
135 | #define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid); | |
136 | ||
137 | #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \ | |
138 | ((dbuf)->db.db_object == (obj) && \ | |
139 | (dbuf)->db_objset == (os) && \ | |
140 | (dbuf)->db_level == (level) && \ | |
141 | (dbuf)->db_blkid == (blkid)) | |
142 | ||
143 | dmu_buf_impl_t * | |
144 | dbuf_find(dnode_t *dn, uint8_t level, uint64_t blkid) | |
145 | { | |
146 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
428870ff | 147 | objset_t *os = dn->dn_objset; |
d6320ddb BB |
148 | uint64_t obj; |
149 | uint64_t hv; | |
150 | uint64_t idx; | |
34dc7c2f BB |
151 | dmu_buf_impl_t *db; |
152 | ||
d6320ddb BB |
153 | obj = dn->dn_object; |
154 | hv = DBUF_HASH(os, obj, level, blkid); | |
155 | idx = hv & h->hash_table_mask; | |
156 | ||
34dc7c2f BB |
157 | mutex_enter(DBUF_HASH_MUTEX(h, idx)); |
158 | for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) { | |
159 | if (DBUF_EQUAL(db, os, obj, level, blkid)) { | |
160 | mutex_enter(&db->db_mtx); | |
161 | if (db->db_state != DB_EVICTING) { | |
162 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); | |
163 | return (db); | |
164 | } | |
165 | mutex_exit(&db->db_mtx); | |
166 | } | |
167 | } | |
168 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); | |
169 | return (NULL); | |
170 | } | |
171 | ||
172 | /* | |
173 | * Insert an entry into the hash table. If there is already an element | |
174 | * equal to elem in the hash table, then the already existing element | |
175 | * will be returned and the new element will not be inserted. | |
176 | * Otherwise returns NULL. | |
177 | */ | |
178 | static dmu_buf_impl_t * | |
179 | dbuf_hash_insert(dmu_buf_impl_t *db) | |
180 | { | |
181 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
428870ff | 182 | objset_t *os = db->db_objset; |
34dc7c2f BB |
183 | uint64_t obj = db->db.db_object; |
184 | int level = db->db_level; | |
d6320ddb | 185 | uint64_t blkid, hv, idx; |
34dc7c2f BB |
186 | dmu_buf_impl_t *dbf; |
187 | ||
d6320ddb BB |
188 | blkid = db->db_blkid; |
189 | hv = DBUF_HASH(os, obj, level, blkid); | |
190 | idx = hv & h->hash_table_mask; | |
191 | ||
34dc7c2f BB |
192 | mutex_enter(DBUF_HASH_MUTEX(h, idx)); |
193 | for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) { | |
194 | if (DBUF_EQUAL(dbf, os, obj, level, blkid)) { | |
195 | mutex_enter(&dbf->db_mtx); | |
196 | if (dbf->db_state != DB_EVICTING) { | |
197 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); | |
198 | return (dbf); | |
199 | } | |
200 | mutex_exit(&dbf->db_mtx); | |
201 | } | |
202 | } | |
203 | ||
204 | mutex_enter(&db->db_mtx); | |
205 | db->db_hash_next = h->hash_table[idx]; | |
206 | h->hash_table[idx] = db; | |
207 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); | |
208 | atomic_add_64(&dbuf_hash_count, 1); | |
209 | ||
210 | return (NULL); | |
211 | } | |
212 | ||
213 | /* | |
214 | * Remove an entry from the hash table. This operation will | |
215 | * fail if there are any existing holds on the db. | |
216 | */ | |
217 | static void | |
218 | dbuf_hash_remove(dmu_buf_impl_t *db) | |
219 | { | |
220 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
d6320ddb | 221 | uint64_t hv, idx; |
34dc7c2f BB |
222 | dmu_buf_impl_t *dbf, **dbp; |
223 | ||
d6320ddb BB |
224 | hv = DBUF_HASH(db->db_objset, db->db.db_object, |
225 | db->db_level, db->db_blkid); | |
226 | idx = hv & h->hash_table_mask; | |
227 | ||
34dc7c2f BB |
228 | /* |
229 | * We musn't hold db_mtx to maintin lock ordering: | |
230 | * DBUF_HASH_MUTEX > db_mtx. | |
231 | */ | |
232 | ASSERT(refcount_is_zero(&db->db_holds)); | |
233 | ASSERT(db->db_state == DB_EVICTING); | |
234 | ASSERT(!MUTEX_HELD(&db->db_mtx)); | |
235 | ||
236 | mutex_enter(DBUF_HASH_MUTEX(h, idx)); | |
237 | dbp = &h->hash_table[idx]; | |
238 | while ((dbf = *dbp) != db) { | |
239 | dbp = &dbf->db_hash_next; | |
240 | ASSERT(dbf != NULL); | |
241 | } | |
242 | *dbp = db->db_hash_next; | |
243 | db->db_hash_next = NULL; | |
244 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); | |
245 | atomic_add_64(&dbuf_hash_count, -1); | |
246 | } | |
247 | ||
248 | static arc_evict_func_t dbuf_do_evict; | |
249 | ||
250 | static void | |
251 | dbuf_evict_user(dmu_buf_impl_t *db) | |
252 | { | |
253 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
254 | ||
255 | if (db->db_level != 0 || db->db_evict_func == NULL) | |
256 | return; | |
257 | ||
258 | if (db->db_user_data_ptr_ptr) | |
259 | *db->db_user_data_ptr_ptr = db->db.db_data; | |
260 | db->db_evict_func(&db->db, db->db_user_ptr); | |
261 | db->db_user_ptr = NULL; | |
262 | db->db_user_data_ptr_ptr = NULL; | |
263 | db->db_evict_func = NULL; | |
264 | } | |
265 | ||
572e2857 BB |
266 | boolean_t |
267 | dbuf_is_metadata(dmu_buf_impl_t *db) | |
268 | { | |
cc79a5c2 BB |
269 | /* |
270 | * Consider indirect blocks and spill blocks to be meta data. | |
271 | */ | |
272 | if (db->db_level > 0 || db->db_blkid == DMU_SPILL_BLKID) { | |
572e2857 BB |
273 | return (B_TRUE); |
274 | } else { | |
275 | boolean_t is_metadata; | |
276 | ||
277 | DB_DNODE_ENTER(db); | |
9ae529ec | 278 | is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type); |
572e2857 BB |
279 | DB_DNODE_EXIT(db); |
280 | ||
281 | return (is_metadata); | |
282 | } | |
283 | } | |
284 | ||
34dc7c2f BB |
285 | void |
286 | dbuf_evict(dmu_buf_impl_t *db) | |
287 | { | |
288 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
289 | ASSERT(db->db_buf == NULL); | |
290 | ASSERT(db->db_data_pending == NULL); | |
291 | ||
292 | dbuf_clear(db); | |
293 | dbuf_destroy(db); | |
294 | } | |
295 | ||
296 | void | |
297 | dbuf_init(void) | |
298 | { | |
299 | uint64_t hsize = 1ULL << 16; | |
300 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
301 | int i; | |
302 | ||
303 | /* | |
304 | * The hash table is big enough to fill all of physical memory | |
305 | * with an average 4K block size. The table will take up | |
306 | * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers). | |
307 | */ | |
308 | while (hsize * 4096 < physmem * PAGESIZE) | |
309 | hsize <<= 1; | |
310 | ||
311 | retry: | |
312 | h->hash_table_mask = hsize - 1; | |
00b46022 | 313 | #if defined(_KERNEL) && defined(HAVE_SPL) |
d1d7e268 MK |
314 | /* |
315 | * Large allocations which do not require contiguous pages | |
316 | * should be using vmem_alloc() in the linux kernel | |
317 | */ | |
b8d06fca | 318 | h->hash_table = vmem_zalloc(hsize * sizeof (void *), KM_PUSHPAGE); |
00b46022 | 319 | #else |
34dc7c2f | 320 | h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP); |
00b46022 | 321 | #endif |
34dc7c2f BB |
322 | if (h->hash_table == NULL) { |
323 | /* XXX - we should really return an error instead of assert */ | |
324 | ASSERT(hsize > (1ULL << 10)); | |
325 | hsize >>= 1; | |
326 | goto retry; | |
327 | } | |
328 | ||
329 | dbuf_cache = kmem_cache_create("dmu_buf_impl_t", | |
330 | sizeof (dmu_buf_impl_t), | |
331 | 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0); | |
332 | ||
333 | for (i = 0; i < DBUF_MUTEXES; i++) | |
334 | mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL); | |
e0b0ca98 BB |
335 | |
336 | dbuf_stats_init(h); | |
34dc7c2f BB |
337 | } |
338 | ||
339 | void | |
340 | dbuf_fini(void) | |
341 | { | |
342 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
343 | int i; | |
344 | ||
e0b0ca98 BB |
345 | dbuf_stats_destroy(); |
346 | ||
34dc7c2f BB |
347 | for (i = 0; i < DBUF_MUTEXES; i++) |
348 | mutex_destroy(&h->hash_mutexes[i]); | |
00b46022 | 349 | #if defined(_KERNEL) && defined(HAVE_SPL) |
d1d7e268 MK |
350 | /* |
351 | * Large allocations which do not require contiguous pages | |
352 | * should be using vmem_free() in the linux kernel | |
353 | */ | |
00b46022 BB |
354 | vmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *)); |
355 | #else | |
34dc7c2f | 356 | kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *)); |
00b46022 | 357 | #endif |
34dc7c2f BB |
358 | kmem_cache_destroy(dbuf_cache); |
359 | } | |
360 | ||
361 | /* | |
362 | * Other stuff. | |
363 | */ | |
364 | ||
365 | #ifdef ZFS_DEBUG | |
366 | static void | |
367 | dbuf_verify(dmu_buf_impl_t *db) | |
368 | { | |
572e2857 | 369 | dnode_t *dn; |
428870ff | 370 | dbuf_dirty_record_t *dr; |
34dc7c2f BB |
371 | |
372 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
373 | ||
374 | if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY)) | |
375 | return; | |
376 | ||
377 | ASSERT(db->db_objset != NULL); | |
572e2857 BB |
378 | DB_DNODE_ENTER(db); |
379 | dn = DB_DNODE(db); | |
34dc7c2f BB |
380 | if (dn == NULL) { |
381 | ASSERT(db->db_parent == NULL); | |
382 | ASSERT(db->db_blkptr == NULL); | |
383 | } else { | |
384 | ASSERT3U(db->db.db_object, ==, dn->dn_object); | |
385 | ASSERT3P(db->db_objset, ==, dn->dn_objset); | |
386 | ASSERT3U(db->db_level, <, dn->dn_nlevels); | |
572e2857 BB |
387 | ASSERT(db->db_blkid == DMU_BONUS_BLKID || |
388 | db->db_blkid == DMU_SPILL_BLKID || | |
389 | !list_is_empty(&dn->dn_dbufs)); | |
34dc7c2f | 390 | } |
428870ff BB |
391 | if (db->db_blkid == DMU_BONUS_BLKID) { |
392 | ASSERT(dn != NULL); | |
393 | ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); | |
394 | ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID); | |
395 | } else if (db->db_blkid == DMU_SPILL_BLKID) { | |
34dc7c2f BB |
396 | ASSERT(dn != NULL); |
397 | ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); | |
c99c9001 | 398 | ASSERT0(db->db.db_offset); |
34dc7c2f BB |
399 | } else { |
400 | ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size); | |
401 | } | |
402 | ||
428870ff BB |
403 | for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next) |
404 | ASSERT(dr->dr_dbuf == db); | |
405 | ||
406 | for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next) | |
407 | ASSERT(dr->dr_dbuf == db); | |
408 | ||
b128c09f BB |
409 | /* |
410 | * We can't assert that db_size matches dn_datablksz because it | |
411 | * can be momentarily different when another thread is doing | |
412 | * dnode_set_blksz(). | |
413 | */ | |
414 | if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) { | |
428870ff | 415 | dr = db->db_data_pending; |
b128c09f BB |
416 | /* |
417 | * It should only be modified in syncing context, so | |
418 | * make sure we only have one copy of the data. | |
419 | */ | |
420 | ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf); | |
34dc7c2f BB |
421 | } |
422 | ||
423 | /* verify db->db_blkptr */ | |
424 | if (db->db_blkptr) { | |
425 | if (db->db_parent == dn->dn_dbuf) { | |
426 | /* db is pointed to by the dnode */ | |
427 | /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */ | |
9babb374 | 428 | if (DMU_OBJECT_IS_SPECIAL(db->db.db_object)) |
34dc7c2f BB |
429 | ASSERT(db->db_parent == NULL); |
430 | else | |
431 | ASSERT(db->db_parent != NULL); | |
428870ff BB |
432 | if (db->db_blkid != DMU_SPILL_BLKID) |
433 | ASSERT3P(db->db_blkptr, ==, | |
434 | &dn->dn_phys->dn_blkptr[db->db_blkid]); | |
34dc7c2f BB |
435 | } else { |
436 | /* db is pointed to by an indirect block */ | |
1fde1e37 BB |
437 | ASSERTV(int epb = db->db_parent->db.db_size >> |
438 | SPA_BLKPTRSHIFT); | |
34dc7c2f BB |
439 | ASSERT3U(db->db_parent->db_level, ==, db->db_level+1); |
440 | ASSERT3U(db->db_parent->db.db_object, ==, | |
441 | db->db.db_object); | |
442 | /* | |
443 | * dnode_grow_indblksz() can make this fail if we don't | |
444 | * have the struct_rwlock. XXX indblksz no longer | |
445 | * grows. safe to do this now? | |
446 | */ | |
572e2857 | 447 | if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) { |
34dc7c2f BB |
448 | ASSERT3P(db->db_blkptr, ==, |
449 | ((blkptr_t *)db->db_parent->db.db_data + | |
450 | db->db_blkid % epb)); | |
451 | } | |
452 | } | |
453 | } | |
454 | if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) && | |
428870ff BB |
455 | (db->db_buf == NULL || db->db_buf->b_data) && |
456 | db->db.db_data && db->db_blkid != DMU_BONUS_BLKID && | |
34dc7c2f BB |
457 | db->db_state != DB_FILL && !dn->dn_free_txg) { |
458 | /* | |
459 | * If the blkptr isn't set but they have nonzero data, | |
460 | * it had better be dirty, otherwise we'll lose that | |
461 | * data when we evict this buffer. | |
462 | */ | |
463 | if (db->db_dirtycnt == 0) { | |
1fde1e37 | 464 | ASSERTV(uint64_t *buf = db->db.db_data); |
34dc7c2f BB |
465 | int i; |
466 | ||
467 | for (i = 0; i < db->db.db_size >> 3; i++) { | |
468 | ASSERT(buf[i] == 0); | |
469 | } | |
470 | } | |
471 | } | |
572e2857 | 472 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
473 | } |
474 | #endif | |
475 | ||
476 | static void | |
477 | dbuf_update_data(dmu_buf_impl_t *db) | |
478 | { | |
479 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
480 | if (db->db_level == 0 && db->db_user_data_ptr_ptr) { | |
481 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
482 | *db->db_user_data_ptr_ptr = db->db.db_data; | |
483 | } | |
484 | } | |
485 | ||
486 | static void | |
487 | dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf) | |
488 | { | |
489 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
490 | ASSERT(db->db_buf == NULL || !arc_has_callback(db->db_buf)); | |
491 | db->db_buf = buf; | |
492 | if (buf != NULL) { | |
493 | ASSERT(buf->b_data != NULL); | |
494 | db->db.db_data = buf->b_data; | |
495 | if (!arc_released(buf)) | |
496 | arc_set_callback(buf, dbuf_do_evict, db); | |
497 | dbuf_update_data(db); | |
498 | } else { | |
499 | dbuf_evict_user(db); | |
500 | db->db.db_data = NULL; | |
b128c09f BB |
501 | if (db->db_state != DB_NOFILL) |
502 | db->db_state = DB_UNCACHED; | |
34dc7c2f BB |
503 | } |
504 | } | |
505 | ||
428870ff BB |
506 | /* |
507 | * Loan out an arc_buf for read. Return the loaned arc_buf. | |
508 | */ | |
509 | arc_buf_t * | |
510 | dbuf_loan_arcbuf(dmu_buf_impl_t *db) | |
511 | { | |
512 | arc_buf_t *abuf; | |
513 | ||
514 | mutex_enter(&db->db_mtx); | |
515 | if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) { | |
516 | int blksz = db->db.db_size; | |
b0bc7a84 | 517 | spa_t *spa = db->db_objset->os_spa; |
572e2857 | 518 | |
428870ff | 519 | mutex_exit(&db->db_mtx); |
572e2857 | 520 | abuf = arc_loan_buf(spa, blksz); |
428870ff BB |
521 | bcopy(db->db.db_data, abuf->b_data, blksz); |
522 | } else { | |
523 | abuf = db->db_buf; | |
524 | arc_loan_inuse_buf(abuf, db); | |
525 | dbuf_set_data(db, NULL); | |
526 | mutex_exit(&db->db_mtx); | |
527 | } | |
528 | return (abuf); | |
529 | } | |
530 | ||
34dc7c2f BB |
531 | uint64_t |
532 | dbuf_whichblock(dnode_t *dn, uint64_t offset) | |
533 | { | |
534 | if (dn->dn_datablkshift) { | |
535 | return (offset >> dn->dn_datablkshift); | |
536 | } else { | |
537 | ASSERT3U(offset, <, dn->dn_datablksz); | |
538 | return (0); | |
539 | } | |
540 | } | |
541 | ||
542 | static void | |
543 | dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb) | |
544 | { | |
545 | dmu_buf_impl_t *db = vdb; | |
546 | ||
547 | mutex_enter(&db->db_mtx); | |
548 | ASSERT3U(db->db_state, ==, DB_READ); | |
549 | /* | |
550 | * All reads are synchronous, so we must have a hold on the dbuf | |
551 | */ | |
552 | ASSERT(refcount_count(&db->db_holds) > 0); | |
553 | ASSERT(db->db_buf == NULL); | |
554 | ASSERT(db->db.db_data == NULL); | |
555 | if (db->db_level == 0 && db->db_freed_in_flight) { | |
556 | /* we were freed in flight; disregard any error */ | |
557 | arc_release(buf, db); | |
558 | bzero(buf->b_data, db->db.db_size); | |
559 | arc_buf_freeze(buf); | |
560 | db->db_freed_in_flight = FALSE; | |
561 | dbuf_set_data(db, buf); | |
562 | db->db_state = DB_CACHED; | |
563 | } else if (zio == NULL || zio->io_error == 0) { | |
564 | dbuf_set_data(db, buf); | |
565 | db->db_state = DB_CACHED; | |
566 | } else { | |
428870ff | 567 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f | 568 | ASSERT3P(db->db_buf, ==, NULL); |
13fe0198 | 569 | VERIFY(arc_buf_remove_ref(buf, db)); |
34dc7c2f BB |
570 | db->db_state = DB_UNCACHED; |
571 | } | |
572 | cv_broadcast(&db->db_changed); | |
428870ff | 573 | dbuf_rele_and_unlock(db, NULL); |
34dc7c2f BB |
574 | } |
575 | ||
576 | static void | |
577 | dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t *flags) | |
578 | { | |
572e2857 | 579 | dnode_t *dn; |
34dc7c2f BB |
580 | zbookmark_t zb; |
581 | uint32_t aflags = ARC_NOWAIT; | |
582 | ||
572e2857 BB |
583 | DB_DNODE_ENTER(db); |
584 | dn = DB_DNODE(db); | |
34dc7c2f BB |
585 | ASSERT(!refcount_is_zero(&db->db_holds)); |
586 | /* We need the struct_rwlock to prevent db_blkptr from changing. */ | |
b128c09f | 587 | ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); |
34dc7c2f BB |
588 | ASSERT(MUTEX_HELD(&db->db_mtx)); |
589 | ASSERT(db->db_state == DB_UNCACHED); | |
590 | ASSERT(db->db_buf == NULL); | |
591 | ||
428870ff | 592 | if (db->db_blkid == DMU_BONUS_BLKID) { |
9babb374 | 593 | int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen); |
34dc7c2f BB |
594 | |
595 | ASSERT3U(bonuslen, <=, db->db.db_size); | |
596 | db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN); | |
d164b209 | 597 | arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); |
34dc7c2f BB |
598 | if (bonuslen < DN_MAX_BONUSLEN) |
599 | bzero(db->db.db_data, DN_MAX_BONUSLEN); | |
9babb374 BB |
600 | if (bonuslen) |
601 | bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen); | |
572e2857 | 602 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
603 | dbuf_update_data(db); |
604 | db->db_state = DB_CACHED; | |
605 | mutex_exit(&db->db_mtx); | |
606 | return; | |
607 | } | |
608 | ||
b128c09f BB |
609 | /* |
610 | * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync() | |
611 | * processes the delete record and clears the bp while we are waiting | |
612 | * for the dn_mtx (resulting in a "no" from block_freed). | |
613 | */ | |
614 | if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) || | |
615 | (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) || | |
616 | BP_IS_HOLE(db->db_blkptr)))) { | |
34dc7c2f BB |
617 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); |
618 | ||
572e2857 | 619 | DB_DNODE_EXIT(db); |
b0bc7a84 MG |
620 | dbuf_set_data(db, arc_buf_alloc(db->db_objset->os_spa, |
621 | db->db.db_size, db, type)); | |
34dc7c2f BB |
622 | bzero(db->db.db_data, db->db.db_size); |
623 | db->db_state = DB_CACHED; | |
624 | *flags |= DB_RF_CACHED; | |
625 | mutex_exit(&db->db_mtx); | |
626 | return; | |
627 | } | |
628 | ||
572e2857 BB |
629 | DB_DNODE_EXIT(db); |
630 | ||
34dc7c2f BB |
631 | db->db_state = DB_READ; |
632 | mutex_exit(&db->db_mtx); | |
633 | ||
b128c09f BB |
634 | if (DBUF_IS_L2CACHEABLE(db)) |
635 | aflags |= ARC_L2CACHE; | |
3a17a7a9 SK |
636 | if (DBUF_IS_L2COMPRESSIBLE(db)) |
637 | aflags |= ARC_L2COMPRESS; | |
b128c09f | 638 | |
428870ff BB |
639 | SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ? |
640 | db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET, | |
641 | db->db.db_object, db->db_level, db->db_blkid); | |
34dc7c2f BB |
642 | |
643 | dbuf_add_ref(db, NULL); | |
b128c09f | 644 | |
b0bc7a84 | 645 | (void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr, |
34dc7c2f BB |
646 | dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ, |
647 | (*flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED, | |
648 | &aflags, &zb); | |
649 | if (aflags & ARC_CACHED) | |
650 | *flags |= DB_RF_CACHED; | |
651 | } | |
652 | ||
653 | int | |
654 | dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) | |
655 | { | |
656 | int err = 0; | |
b0bc7a84 MG |
657 | boolean_t havepzio = (zio != NULL); |
658 | boolean_t prefetch; | |
572e2857 | 659 | dnode_t *dn; |
34dc7c2f BB |
660 | |
661 | /* | |
662 | * We don't have to hold the mutex to check db_state because it | |
663 | * can't be freed while we have a hold on the buffer. | |
664 | */ | |
665 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
666 | ||
b128c09f | 667 | if (db->db_state == DB_NOFILL) |
2e528b49 | 668 | return (SET_ERROR(EIO)); |
b128c09f | 669 | |
572e2857 BB |
670 | DB_DNODE_ENTER(db); |
671 | dn = DB_DNODE(db); | |
34dc7c2f | 672 | if ((flags & DB_RF_HAVESTRUCT) == 0) |
572e2857 | 673 | rw_enter(&dn->dn_struct_rwlock, RW_READER); |
34dc7c2f | 674 | |
428870ff | 675 | prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && |
572e2857 | 676 | (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL && |
b128c09f | 677 | DBUF_IS_CACHEABLE(db); |
34dc7c2f BB |
678 | |
679 | mutex_enter(&db->db_mtx); | |
680 | if (db->db_state == DB_CACHED) { | |
681 | mutex_exit(&db->db_mtx); | |
682 | if (prefetch) | |
572e2857 | 683 | dmu_zfetch(&dn->dn_zfetch, db->db.db_offset, |
34dc7c2f BB |
684 | db->db.db_size, TRUE); |
685 | if ((flags & DB_RF_HAVESTRUCT) == 0) | |
572e2857 BB |
686 | rw_exit(&dn->dn_struct_rwlock); |
687 | DB_DNODE_EXIT(db); | |
34dc7c2f | 688 | } else if (db->db_state == DB_UNCACHED) { |
572e2857 BB |
689 | spa_t *spa = dn->dn_objset->os_spa; |
690 | ||
691 | if (zio == NULL) | |
692 | zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); | |
34dc7c2f BB |
693 | dbuf_read_impl(db, zio, &flags); |
694 | ||
695 | /* dbuf_read_impl has dropped db_mtx for us */ | |
696 | ||
697 | if (prefetch) | |
572e2857 | 698 | dmu_zfetch(&dn->dn_zfetch, db->db.db_offset, |
34dc7c2f BB |
699 | db->db.db_size, flags & DB_RF_CACHED); |
700 | ||
701 | if ((flags & DB_RF_HAVESTRUCT) == 0) | |
572e2857 BB |
702 | rw_exit(&dn->dn_struct_rwlock); |
703 | DB_DNODE_EXIT(db); | |
34dc7c2f BB |
704 | |
705 | if (!havepzio) | |
706 | err = zio_wait(zio); | |
707 | } else { | |
e49f1e20 WA |
708 | /* |
709 | * Another reader came in while the dbuf was in flight | |
710 | * between UNCACHED and CACHED. Either a writer will finish | |
711 | * writing the buffer (sending the dbuf to CACHED) or the | |
712 | * first reader's request will reach the read_done callback | |
713 | * and send the dbuf to CACHED. Otherwise, a failure | |
714 | * occurred and the dbuf went to UNCACHED. | |
715 | */ | |
34dc7c2f BB |
716 | mutex_exit(&db->db_mtx); |
717 | if (prefetch) | |
572e2857 | 718 | dmu_zfetch(&dn->dn_zfetch, db->db.db_offset, |
34dc7c2f BB |
719 | db->db.db_size, TRUE); |
720 | if ((flags & DB_RF_HAVESTRUCT) == 0) | |
572e2857 BB |
721 | rw_exit(&dn->dn_struct_rwlock); |
722 | DB_DNODE_EXIT(db); | |
34dc7c2f | 723 | |
e49f1e20 | 724 | /* Skip the wait per the caller's request. */ |
34dc7c2f BB |
725 | mutex_enter(&db->db_mtx); |
726 | if ((flags & DB_RF_NEVERWAIT) == 0) { | |
727 | while (db->db_state == DB_READ || | |
728 | db->db_state == DB_FILL) { | |
729 | ASSERT(db->db_state == DB_READ || | |
730 | (flags & DB_RF_HAVESTRUCT) == 0); | |
731 | cv_wait(&db->db_changed, &db->db_mtx); | |
732 | } | |
733 | if (db->db_state == DB_UNCACHED) | |
2e528b49 | 734 | err = SET_ERROR(EIO); |
34dc7c2f BB |
735 | } |
736 | mutex_exit(&db->db_mtx); | |
737 | } | |
738 | ||
739 | ASSERT(err || havepzio || db->db_state == DB_CACHED); | |
740 | return (err); | |
741 | } | |
742 | ||
743 | static void | |
744 | dbuf_noread(dmu_buf_impl_t *db) | |
745 | { | |
746 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
428870ff | 747 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f BB |
748 | mutex_enter(&db->db_mtx); |
749 | while (db->db_state == DB_READ || db->db_state == DB_FILL) | |
750 | cv_wait(&db->db_changed, &db->db_mtx); | |
751 | if (db->db_state == DB_UNCACHED) { | |
752 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); | |
b0bc7a84 | 753 | spa_t *spa = db->db_objset->os_spa; |
34dc7c2f BB |
754 | |
755 | ASSERT(db->db_buf == NULL); | |
756 | ASSERT(db->db.db_data == NULL); | |
572e2857 | 757 | dbuf_set_data(db, arc_buf_alloc(spa, db->db.db_size, db, type)); |
34dc7c2f | 758 | db->db_state = DB_FILL; |
b128c09f BB |
759 | } else if (db->db_state == DB_NOFILL) { |
760 | dbuf_set_data(db, NULL); | |
34dc7c2f BB |
761 | } else { |
762 | ASSERT3U(db->db_state, ==, DB_CACHED); | |
763 | } | |
764 | mutex_exit(&db->db_mtx); | |
765 | } | |
766 | ||
767 | /* | |
768 | * This is our just-in-time copy function. It makes a copy of | |
769 | * buffers, that have been modified in a previous transaction | |
770 | * group, before we modify them in the current active group. | |
771 | * | |
772 | * This function is used in two places: when we are dirtying a | |
773 | * buffer for the first time in a txg, and when we are freeing | |
774 | * a range in a dnode that includes this buffer. | |
775 | * | |
776 | * Note that when we are called from dbuf_free_range() we do | |
777 | * not put a hold on the buffer, we just traverse the active | |
778 | * dbuf list for the dnode. | |
779 | */ | |
780 | static void | |
781 | dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg) | |
782 | { | |
783 | dbuf_dirty_record_t *dr = db->db_last_dirty; | |
784 | ||
785 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
786 | ASSERT(db->db.db_data != NULL); | |
787 | ASSERT(db->db_level == 0); | |
788 | ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT); | |
789 | ||
790 | if (dr == NULL || | |
791 | (dr->dt.dl.dr_data != | |
428870ff | 792 | ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf))) |
34dc7c2f BB |
793 | return; |
794 | ||
795 | /* | |
796 | * If the last dirty record for this dbuf has not yet synced | |
797 | * and its referencing the dbuf data, either: | |
572e2857 | 798 | * reset the reference to point to a new copy, |
34dc7c2f BB |
799 | * or (if there a no active holders) |
800 | * just null out the current db_data pointer. | |
801 | */ | |
802 | ASSERT(dr->dr_txg >= txg - 2); | |
428870ff | 803 | if (db->db_blkid == DMU_BONUS_BLKID) { |
34dc7c2f BB |
804 | /* Note that the data bufs here are zio_bufs */ |
805 | dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN); | |
d164b209 | 806 | arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); |
34dc7c2f BB |
807 | bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN); |
808 | } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) { | |
809 | int size = db->db.db_size; | |
810 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); | |
b0bc7a84 | 811 | spa_t *spa = db->db_objset->os_spa; |
572e2857 | 812 | |
572e2857 | 813 | dr->dt.dl.dr_data = arc_buf_alloc(spa, size, db, type); |
34dc7c2f BB |
814 | bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size); |
815 | } else { | |
816 | dbuf_set_data(db, NULL); | |
817 | } | |
818 | } | |
819 | ||
820 | void | |
821 | dbuf_unoverride(dbuf_dirty_record_t *dr) | |
822 | { | |
823 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
428870ff | 824 | blkptr_t *bp = &dr->dt.dl.dr_overridden_by; |
34dc7c2f BB |
825 | uint64_t txg = dr->dr_txg; |
826 | ||
827 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
828 | ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC); | |
829 | ASSERT(db->db_level == 0); | |
830 | ||
428870ff | 831 | if (db->db_blkid == DMU_BONUS_BLKID || |
34dc7c2f BB |
832 | dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN) |
833 | return; | |
834 | ||
428870ff BB |
835 | ASSERT(db->db_data_pending != dr); |
836 | ||
34dc7c2f | 837 | /* free this block */ |
b0bc7a84 MG |
838 | if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite) |
839 | zio_free(db->db_objset->os_spa, txg, bp); | |
428870ff | 840 | |
34dc7c2f | 841 | dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; |
03c6040b GW |
842 | dr->dt.dl.dr_nopwrite = B_FALSE; |
843 | ||
34dc7c2f BB |
844 | /* |
845 | * Release the already-written buffer, so we leave it in | |
846 | * a consistent dirty state. Note that all callers are | |
847 | * modifying the buffer, so they will immediately do | |
848 | * another (redundant) arc_release(). Therefore, leave | |
849 | * the buf thawed to save the effort of freezing & | |
850 | * immediately re-thawing it. | |
851 | */ | |
852 | arc_release(dr->dt.dl.dr_data, db); | |
853 | } | |
854 | ||
b128c09f BB |
855 | /* |
856 | * Evict (if its unreferenced) or clear (if its referenced) any level-0 | |
857 | * data blocks in the free range, so that any future readers will find | |
b0bc7a84 | 858 | * empty blocks. |
ea97f8ce MA |
859 | * |
860 | * This is a no-op if the dataset is in the middle of an incremental | |
861 | * receive; see comment below for details. | |
b128c09f | 862 | */ |
34dc7c2f | 863 | void |
b128c09f | 864 | dbuf_free_range(dnode_t *dn, uint64_t start, uint64_t end, dmu_tx_t *tx) |
34dc7c2f BB |
865 | { |
866 | dmu_buf_impl_t *db, *db_next; | |
867 | uint64_t txg = tx->tx_txg; | |
868 | ||
b0bc7a84 | 869 | if (end > dn->dn_maxblkid && (end != DMU_SPILL_BLKID)) |
b128c09f | 870 | end = dn->dn_maxblkid; |
b128c09f | 871 | dprintf_dnode(dn, "start=%llu end=%llu\n", start, end); |
ea97f8ce | 872 | |
b663a23d MA |
873 | mutex_enter(&dn->dn_dbufs_mtx); |
874 | if (start >= dn->dn_unlisted_l0_blkid * dn->dn_datablksz) { | |
875 | /* There can't be any dbufs in this range; no need to search. */ | |
876 | mutex_exit(&dn->dn_dbufs_mtx); | |
877 | return; | |
878 | } else if (dmu_objset_is_receiving(dn->dn_objset)) { | |
ea97f8ce | 879 | /* |
b663a23d MA |
880 | * If we are receiving, we expect there to be no dbufs in |
881 | * the range to be freed, because receive modifies each | |
882 | * block at most once, and in offset order. If this is | |
883 | * not the case, it can lead to performance problems, | |
884 | * so note that we unexpectedly took the slow path. | |
ea97f8ce | 885 | */ |
b663a23d | 886 | atomic_inc_64(&zfs_free_range_recv_miss); |
ea97f8ce MA |
887 | } |
888 | ||
e8b96c60 | 889 | for (db = list_head(&dn->dn_dbufs); db != NULL; db = db_next) { |
34dc7c2f | 890 | db_next = list_next(&dn->dn_dbufs, db); |
428870ff | 891 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
b128c09f | 892 | |
34dc7c2f BB |
893 | if (db->db_level != 0) |
894 | continue; | |
b128c09f | 895 | if (db->db_blkid < start || db->db_blkid > end) |
34dc7c2f BB |
896 | continue; |
897 | ||
898 | /* found a level 0 buffer in the range */ | |
13fe0198 MA |
899 | mutex_enter(&db->db_mtx); |
900 | if (dbuf_undirty(db, tx)) { | |
901 | /* mutex has been dropped and dbuf destroyed */ | |
34dc7c2f | 902 | continue; |
13fe0198 | 903 | } |
34dc7c2f | 904 | |
34dc7c2f | 905 | if (db->db_state == DB_UNCACHED || |
b128c09f | 906 | db->db_state == DB_NOFILL || |
34dc7c2f BB |
907 | db->db_state == DB_EVICTING) { |
908 | ASSERT(db->db.db_data == NULL); | |
909 | mutex_exit(&db->db_mtx); | |
910 | continue; | |
911 | } | |
912 | if (db->db_state == DB_READ || db->db_state == DB_FILL) { | |
913 | /* will be handled in dbuf_read_done or dbuf_rele */ | |
914 | db->db_freed_in_flight = TRUE; | |
915 | mutex_exit(&db->db_mtx); | |
916 | continue; | |
917 | } | |
918 | if (refcount_count(&db->db_holds) == 0) { | |
919 | ASSERT(db->db_buf); | |
920 | dbuf_clear(db); | |
921 | continue; | |
922 | } | |
923 | /* The dbuf is referenced */ | |
924 | ||
925 | if (db->db_last_dirty != NULL) { | |
926 | dbuf_dirty_record_t *dr = db->db_last_dirty; | |
927 | ||
928 | if (dr->dr_txg == txg) { | |
929 | /* | |
930 | * This buffer is "in-use", re-adjust the file | |
931 | * size to reflect that this buffer may | |
932 | * contain new data when we sync. | |
933 | */ | |
428870ff BB |
934 | if (db->db_blkid != DMU_SPILL_BLKID && |
935 | db->db_blkid > dn->dn_maxblkid) | |
34dc7c2f BB |
936 | dn->dn_maxblkid = db->db_blkid; |
937 | dbuf_unoverride(dr); | |
938 | } else { | |
939 | /* | |
940 | * This dbuf is not dirty in the open context. | |
941 | * Either uncache it (if its not referenced in | |
942 | * the open context) or reset its contents to | |
943 | * empty. | |
944 | */ | |
945 | dbuf_fix_old_data(db, txg); | |
946 | } | |
947 | } | |
948 | /* clear the contents if its cached */ | |
949 | if (db->db_state == DB_CACHED) { | |
950 | ASSERT(db->db.db_data != NULL); | |
951 | arc_release(db->db_buf, db); | |
952 | bzero(db->db.db_data, db->db.db_size); | |
953 | arc_buf_freeze(db->db_buf); | |
954 | } | |
955 | ||
956 | mutex_exit(&db->db_mtx); | |
957 | } | |
958 | mutex_exit(&dn->dn_dbufs_mtx); | |
959 | } | |
960 | ||
961 | static int | |
962 | dbuf_block_freeable(dmu_buf_impl_t *db) | |
963 | { | |
964 | dsl_dataset_t *ds = db->db_objset->os_dsl_dataset; | |
965 | uint64_t birth_txg = 0; | |
966 | ||
967 | /* | |
968 | * We don't need any locking to protect db_blkptr: | |
969 | * If it's syncing, then db_last_dirty will be set | |
970 | * so we'll ignore db_blkptr. | |
b0bc7a84 MG |
971 | * |
972 | * This logic ensures that only block births for | |
973 | * filled blocks are considered. | |
34dc7c2f BB |
974 | */ |
975 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
b0bc7a84 MG |
976 | if (db->db_last_dirty && (db->db_blkptr == NULL || |
977 | !BP_IS_HOLE(db->db_blkptr))) { | |
34dc7c2f | 978 | birth_txg = db->db_last_dirty->dr_txg; |
b0bc7a84 | 979 | } else if (db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) { |
34dc7c2f | 980 | birth_txg = db->db_blkptr->blk_birth; |
b0bc7a84 | 981 | } |
34dc7c2f | 982 | |
572e2857 | 983 | /* |
b0bc7a84 | 984 | * If this block don't exist or is in a snapshot, it can't be freed. |
572e2857 BB |
985 | * Don't pass the bp to dsl_dataset_block_freeable() since we |
986 | * are holding the db_mtx lock and might deadlock if we are | |
987 | * prefetching a dedup-ed block. | |
988 | */ | |
b0bc7a84 | 989 | if (birth_txg != 0) |
34dc7c2f | 990 | return (ds == NULL || |
572e2857 | 991 | dsl_dataset_block_freeable(ds, NULL, birth_txg)); |
34dc7c2f | 992 | else |
b0bc7a84 | 993 | return (B_FALSE); |
34dc7c2f BB |
994 | } |
995 | ||
996 | void | |
997 | dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx) | |
998 | { | |
999 | arc_buf_t *buf, *obuf; | |
1000 | int osize = db->db.db_size; | |
1001 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); | |
572e2857 | 1002 | dnode_t *dn; |
34dc7c2f | 1003 | |
428870ff | 1004 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f | 1005 | |
572e2857 BB |
1006 | DB_DNODE_ENTER(db); |
1007 | dn = DB_DNODE(db); | |
1008 | ||
34dc7c2f | 1009 | /* XXX does *this* func really need the lock? */ |
572e2857 | 1010 | ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); |
34dc7c2f BB |
1011 | |
1012 | /* | |
b0bc7a84 | 1013 | * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held |
34dc7c2f BB |
1014 | * is OK, because there can be no other references to the db |
1015 | * when we are changing its size, so no concurrent DB_FILL can | |
1016 | * be happening. | |
1017 | */ | |
1018 | /* | |
1019 | * XXX we should be doing a dbuf_read, checking the return | |
1020 | * value and returning that up to our callers | |
1021 | */ | |
b0bc7a84 | 1022 | dmu_buf_will_dirty(&db->db, tx); |
34dc7c2f BB |
1023 | |
1024 | /* create the data buffer for the new block */ | |
572e2857 | 1025 | buf = arc_buf_alloc(dn->dn_objset->os_spa, size, db, type); |
34dc7c2f BB |
1026 | |
1027 | /* copy old block data to the new block */ | |
1028 | obuf = db->db_buf; | |
1029 | bcopy(obuf->b_data, buf->b_data, MIN(osize, size)); | |
1030 | /* zero the remainder */ | |
1031 | if (size > osize) | |
1032 | bzero((uint8_t *)buf->b_data + osize, size - osize); | |
1033 | ||
1034 | mutex_enter(&db->db_mtx); | |
1035 | dbuf_set_data(db, buf); | |
13fe0198 | 1036 | VERIFY(arc_buf_remove_ref(obuf, db)); |
34dc7c2f BB |
1037 | db->db.db_size = size; |
1038 | ||
1039 | if (db->db_level == 0) { | |
1040 | ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); | |
1041 | db->db_last_dirty->dt.dl.dr_data = buf; | |
1042 | } | |
1043 | mutex_exit(&db->db_mtx); | |
1044 | ||
572e2857 BB |
1045 | dnode_willuse_space(dn, size-osize, tx); |
1046 | DB_DNODE_EXIT(db); | |
34dc7c2f BB |
1047 | } |
1048 | ||
428870ff BB |
1049 | void |
1050 | dbuf_release_bp(dmu_buf_impl_t *db) | |
1051 | { | |
b0bc7a84 | 1052 | ASSERTV(objset_t *os = db->db_objset); |
428870ff BB |
1053 | |
1054 | ASSERT(dsl_pool_sync_context(dmu_objset_pool(os))); | |
1055 | ASSERT(arc_released(os->os_phys_buf) || | |
1056 | list_link_active(&os->os_dsl_dataset->ds_synced_link)); | |
1057 | ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf)); | |
1058 | ||
294f6806 | 1059 | (void) arc_release(db->db_buf, db); |
428870ff BB |
1060 | } |
1061 | ||
34dc7c2f BB |
1062 | dbuf_dirty_record_t * |
1063 | dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) | |
1064 | { | |
572e2857 BB |
1065 | dnode_t *dn; |
1066 | objset_t *os; | |
34dc7c2f BB |
1067 | dbuf_dirty_record_t **drp, *dr; |
1068 | int drop_struct_lock = FALSE; | |
b128c09f | 1069 | boolean_t do_free_accounting = B_FALSE; |
34dc7c2f BB |
1070 | int txgoff = tx->tx_txg & TXG_MASK; |
1071 | ||
1072 | ASSERT(tx->tx_txg != 0); | |
1073 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
1074 | DMU_TX_DIRTY_BUF(tx, db); | |
1075 | ||
572e2857 BB |
1076 | DB_DNODE_ENTER(db); |
1077 | dn = DB_DNODE(db); | |
34dc7c2f BB |
1078 | /* |
1079 | * Shouldn't dirty a regular buffer in syncing context. Private | |
1080 | * objects may be dirtied in syncing context, but only if they | |
1081 | * were already pre-dirtied in open context. | |
34dc7c2f BB |
1082 | */ |
1083 | ASSERT(!dmu_tx_is_syncing(tx) || | |
1084 | BP_IS_HOLE(dn->dn_objset->os_rootbp) || | |
9babb374 BB |
1085 | DMU_OBJECT_IS_SPECIAL(dn->dn_object) || |
1086 | dn->dn_objset->os_dsl_dataset == NULL); | |
34dc7c2f BB |
1087 | /* |
1088 | * We make this assert for private objects as well, but after we | |
1089 | * check if we're already dirty. They are allowed to re-dirty | |
1090 | * in syncing context. | |
1091 | */ | |
1092 | ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || | |
1093 | dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == | |
1094 | (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); | |
1095 | ||
1096 | mutex_enter(&db->db_mtx); | |
1097 | /* | |
1098 | * XXX make this true for indirects too? The problem is that | |
1099 | * transactions created with dmu_tx_create_assigned() from | |
1100 | * syncing context don't bother holding ahead. | |
1101 | */ | |
1102 | ASSERT(db->db_level != 0 || | |
b128c09f BB |
1103 | db->db_state == DB_CACHED || db->db_state == DB_FILL || |
1104 | db->db_state == DB_NOFILL); | |
34dc7c2f BB |
1105 | |
1106 | mutex_enter(&dn->dn_mtx); | |
1107 | /* | |
1108 | * Don't set dirtyctx to SYNC if we're just modifying this as we | |
1109 | * initialize the objset. | |
1110 | */ | |
1111 | if (dn->dn_dirtyctx == DN_UNDIRTIED && | |
1112 | !BP_IS_HOLE(dn->dn_objset->os_rootbp)) { | |
1113 | dn->dn_dirtyctx = | |
1114 | (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN); | |
1115 | ASSERT(dn->dn_dirtyctx_firstset == NULL); | |
beb98269 | 1116 | dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_PUSHPAGE); |
34dc7c2f BB |
1117 | } |
1118 | mutex_exit(&dn->dn_mtx); | |
1119 | ||
428870ff BB |
1120 | if (db->db_blkid == DMU_SPILL_BLKID) |
1121 | dn->dn_have_spill = B_TRUE; | |
1122 | ||
34dc7c2f BB |
1123 | /* |
1124 | * If this buffer is already dirty, we're done. | |
1125 | */ | |
1126 | drp = &db->db_last_dirty; | |
1127 | ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg || | |
1128 | db->db.db_object == DMU_META_DNODE_OBJECT); | |
1129 | while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg) | |
1130 | drp = &dr->dr_next; | |
1131 | if (dr && dr->dr_txg == tx->tx_txg) { | |
572e2857 BB |
1132 | DB_DNODE_EXIT(db); |
1133 | ||
428870ff | 1134 | if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) { |
34dc7c2f BB |
1135 | /* |
1136 | * If this buffer has already been written out, | |
1137 | * we now need to reset its state. | |
1138 | */ | |
1139 | dbuf_unoverride(dr); | |
428870ff BB |
1140 | if (db->db.db_object != DMU_META_DNODE_OBJECT && |
1141 | db->db_state != DB_NOFILL) | |
34dc7c2f BB |
1142 | arc_buf_thaw(db->db_buf); |
1143 | } | |
1144 | mutex_exit(&db->db_mtx); | |
1145 | return (dr); | |
1146 | } | |
1147 | ||
1148 | /* | |
1149 | * Only valid if not already dirty. | |
1150 | */ | |
9babb374 BB |
1151 | ASSERT(dn->dn_object == 0 || |
1152 | dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == | |
34dc7c2f BB |
1153 | (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); |
1154 | ||
1155 | ASSERT3U(dn->dn_nlevels, >, db->db_level); | |
1156 | ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) || | |
1157 | dn->dn_phys->dn_nlevels > db->db_level || | |
1158 | dn->dn_next_nlevels[txgoff] > db->db_level || | |
1159 | dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level || | |
1160 | dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level); | |
1161 | ||
1162 | /* | |
1163 | * We should only be dirtying in syncing context if it's the | |
9babb374 BB |
1164 | * mos or we're initializing the os or it's a special object. |
1165 | * However, we are allowed to dirty in syncing context provided | |
1166 | * we already dirtied it in open context. Hence we must make | |
1167 | * this assertion only if we're not already dirty. | |
34dc7c2f | 1168 | */ |
572e2857 | 1169 | os = dn->dn_objset; |
9babb374 BB |
1170 | ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) || |
1171 | os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp)); | |
34dc7c2f BB |
1172 | ASSERT(db->db.db_size != 0); |
1173 | ||
1174 | dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); | |
1175 | ||
428870ff | 1176 | if (db->db_blkid != DMU_BONUS_BLKID) { |
34dc7c2f BB |
1177 | /* |
1178 | * Update the accounting. | |
b128c09f BB |
1179 | * Note: we delay "free accounting" until after we drop |
1180 | * the db_mtx. This keeps us from grabbing other locks | |
428870ff | 1181 | * (and possibly deadlocking) in bp_get_dsize() while |
b128c09f | 1182 | * also holding the db_mtx. |
34dc7c2f | 1183 | */ |
34dc7c2f | 1184 | dnode_willuse_space(dn, db->db.db_size, tx); |
b128c09f | 1185 | do_free_accounting = dbuf_block_freeable(db); |
34dc7c2f BB |
1186 | } |
1187 | ||
1188 | /* | |
1189 | * If this buffer is dirty in an old transaction group we need | |
1190 | * to make a copy of it so that the changes we make in this | |
1191 | * transaction group won't leak out when we sync the older txg. | |
1192 | */ | |
beb98269 | 1193 | dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_PUSHPAGE); |
98f72a53 | 1194 | list_link_init(&dr->dr_dirty_node); |
34dc7c2f BB |
1195 | if (db->db_level == 0) { |
1196 | void *data_old = db->db_buf; | |
1197 | ||
b128c09f | 1198 | if (db->db_state != DB_NOFILL) { |
428870ff | 1199 | if (db->db_blkid == DMU_BONUS_BLKID) { |
b128c09f BB |
1200 | dbuf_fix_old_data(db, tx->tx_txg); |
1201 | data_old = db->db.db_data; | |
1202 | } else if (db->db.db_object != DMU_META_DNODE_OBJECT) { | |
1203 | /* | |
1204 | * Release the data buffer from the cache so | |
1205 | * that we can modify it without impacting | |
1206 | * possible other users of this cached data | |
1207 | * block. Note that indirect blocks and | |
1208 | * private objects are not released until the | |
1209 | * syncing state (since they are only modified | |
1210 | * then). | |
1211 | */ | |
1212 | arc_release(db->db_buf, db); | |
1213 | dbuf_fix_old_data(db, tx->tx_txg); | |
1214 | data_old = db->db_buf; | |
1215 | } | |
1216 | ASSERT(data_old != NULL); | |
34dc7c2f | 1217 | } |
34dc7c2f BB |
1218 | dr->dt.dl.dr_data = data_old; |
1219 | } else { | |
1220 | mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL); | |
1221 | list_create(&dr->dt.di.dr_children, | |
1222 | sizeof (dbuf_dirty_record_t), | |
1223 | offsetof(dbuf_dirty_record_t, dr_dirty_node)); | |
1224 | } | |
e8b96c60 MA |
1225 | if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL) |
1226 | dr->dr_accounted = db->db.db_size; | |
34dc7c2f BB |
1227 | dr->dr_dbuf = db; |
1228 | dr->dr_txg = tx->tx_txg; | |
1229 | dr->dr_next = *drp; | |
1230 | *drp = dr; | |
1231 | ||
1232 | /* | |
1233 | * We could have been freed_in_flight between the dbuf_noread | |
1234 | * and dbuf_dirty. We win, as though the dbuf_noread() had | |
1235 | * happened after the free. | |
1236 | */ | |
428870ff BB |
1237 | if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && |
1238 | db->db_blkid != DMU_SPILL_BLKID) { | |
34dc7c2f | 1239 | mutex_enter(&dn->dn_mtx); |
9bd274dd MA |
1240 | if (dn->dn_free_ranges[txgoff] != NULL) { |
1241 | range_tree_clear(dn->dn_free_ranges[txgoff], | |
1242 | db->db_blkid, 1); | |
1243 | } | |
34dc7c2f BB |
1244 | mutex_exit(&dn->dn_mtx); |
1245 | db->db_freed_in_flight = FALSE; | |
1246 | } | |
1247 | ||
1248 | /* | |
1249 | * This buffer is now part of this txg | |
1250 | */ | |
1251 | dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg); | |
1252 | db->db_dirtycnt += 1; | |
1253 | ASSERT3U(db->db_dirtycnt, <=, 3); | |
1254 | ||
1255 | mutex_exit(&db->db_mtx); | |
1256 | ||
428870ff BB |
1257 | if (db->db_blkid == DMU_BONUS_BLKID || |
1258 | db->db_blkid == DMU_SPILL_BLKID) { | |
34dc7c2f BB |
1259 | mutex_enter(&dn->dn_mtx); |
1260 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
1261 | list_insert_tail(&dn->dn_dirty_records[txgoff], dr); | |
1262 | mutex_exit(&dn->dn_mtx); | |
1263 | dnode_setdirty(dn, tx); | |
572e2857 | 1264 | DB_DNODE_EXIT(db); |
34dc7c2f | 1265 | return (dr); |
b128c09f BB |
1266 | } else if (do_free_accounting) { |
1267 | blkptr_t *bp = db->db_blkptr; | |
1268 | int64_t willfree = (bp && !BP_IS_HOLE(bp)) ? | |
428870ff | 1269 | bp_get_dsize(os->os_spa, bp) : db->db.db_size; |
b128c09f BB |
1270 | /* |
1271 | * This is only a guess -- if the dbuf is dirty | |
1272 | * in a previous txg, we don't know how much | |
1273 | * space it will use on disk yet. We should | |
1274 | * really have the struct_rwlock to access | |
1275 | * db_blkptr, but since this is just a guess, | |
1276 | * it's OK if we get an odd answer. | |
1277 | */ | |
572e2857 | 1278 | ddt_prefetch(os->os_spa, bp); |
b128c09f | 1279 | dnode_willuse_space(dn, -willfree, tx); |
34dc7c2f BB |
1280 | } |
1281 | ||
1282 | if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) { | |
1283 | rw_enter(&dn->dn_struct_rwlock, RW_READER); | |
1284 | drop_struct_lock = TRUE; | |
1285 | } | |
1286 | ||
b128c09f BB |
1287 | if (db->db_level == 0) { |
1288 | dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock); | |
1289 | ASSERT(dn->dn_maxblkid >= db->db_blkid); | |
1290 | } | |
1291 | ||
34dc7c2f BB |
1292 | if (db->db_level+1 < dn->dn_nlevels) { |
1293 | dmu_buf_impl_t *parent = db->db_parent; | |
1294 | dbuf_dirty_record_t *di; | |
1295 | int parent_held = FALSE; | |
1296 | ||
1297 | if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) { | |
1298 | int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; | |
1299 | ||
1300 | parent = dbuf_hold_level(dn, db->db_level+1, | |
1301 | db->db_blkid >> epbs, FTAG); | |
428870ff | 1302 | ASSERT(parent != NULL); |
34dc7c2f BB |
1303 | parent_held = TRUE; |
1304 | } | |
1305 | if (drop_struct_lock) | |
1306 | rw_exit(&dn->dn_struct_rwlock); | |
1307 | ASSERT3U(db->db_level+1, ==, parent->db_level); | |
1308 | di = dbuf_dirty(parent, tx); | |
1309 | if (parent_held) | |
1310 | dbuf_rele(parent, FTAG); | |
1311 | ||
1312 | mutex_enter(&db->db_mtx); | |
e8b96c60 MA |
1313 | /* |
1314 | * Since we've dropped the mutex, it's possible that | |
1315 | * dbuf_undirty() might have changed this out from under us. | |
1316 | */ | |
34dc7c2f BB |
1317 | if (db->db_last_dirty == dr || |
1318 | dn->dn_object == DMU_META_DNODE_OBJECT) { | |
1319 | mutex_enter(&di->dt.di.dr_mtx); | |
1320 | ASSERT3U(di->dr_txg, ==, tx->tx_txg); | |
1321 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
1322 | list_insert_tail(&di->dt.di.dr_children, dr); | |
1323 | mutex_exit(&di->dt.di.dr_mtx); | |
1324 | dr->dr_parent = di; | |
1325 | } | |
1326 | mutex_exit(&db->db_mtx); | |
1327 | } else { | |
1328 | ASSERT(db->db_level+1 == dn->dn_nlevels); | |
1329 | ASSERT(db->db_blkid < dn->dn_nblkptr); | |
572e2857 | 1330 | ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf); |
34dc7c2f BB |
1331 | mutex_enter(&dn->dn_mtx); |
1332 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
1333 | list_insert_tail(&dn->dn_dirty_records[txgoff], dr); | |
1334 | mutex_exit(&dn->dn_mtx); | |
1335 | if (drop_struct_lock) | |
1336 | rw_exit(&dn->dn_struct_rwlock); | |
1337 | } | |
1338 | ||
1339 | dnode_setdirty(dn, tx); | |
572e2857 | 1340 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
1341 | return (dr); |
1342 | } | |
1343 | ||
13fe0198 | 1344 | /* |
e49f1e20 WA |
1345 | * Undirty a buffer in the transaction group referenced by the given |
1346 | * transaction. Return whether this evicted the dbuf. | |
13fe0198 MA |
1347 | */ |
1348 | static boolean_t | |
34dc7c2f BB |
1349 | dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx) |
1350 | { | |
572e2857 | 1351 | dnode_t *dn; |
34dc7c2f BB |
1352 | uint64_t txg = tx->tx_txg; |
1353 | dbuf_dirty_record_t *dr, **drp; | |
1354 | ||
1355 | ASSERT(txg != 0); | |
428870ff | 1356 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
13fe0198 MA |
1357 | ASSERT0(db->db_level); |
1358 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
34dc7c2f | 1359 | |
34dc7c2f BB |
1360 | /* |
1361 | * If this buffer is not dirty, we're done. | |
1362 | */ | |
1363 | for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next) | |
1364 | if (dr->dr_txg <= txg) | |
1365 | break; | |
13fe0198 MA |
1366 | if (dr == NULL || dr->dr_txg < txg) |
1367 | return (B_FALSE); | |
34dc7c2f | 1368 | ASSERT(dr->dr_txg == txg); |
428870ff | 1369 | ASSERT(dr->dr_dbuf == db); |
34dc7c2f | 1370 | |
572e2857 BB |
1371 | DB_DNODE_ENTER(db); |
1372 | dn = DB_DNODE(db); | |
1373 | ||
34dc7c2f BB |
1374 | dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); |
1375 | ||
1376 | ASSERT(db->db.db_size != 0); | |
1377 | ||
e8b96c60 MA |
1378 | /* |
1379 | * Any space we accounted for in dp_dirty_* will be cleaned up by | |
1380 | * dsl_pool_sync(). This is relatively rare so the discrepancy | |
1381 | * is not a big deal. | |
1382 | */ | |
34dc7c2f BB |
1383 | |
1384 | *drp = dr->dr_next; | |
1385 | ||
ef3c1dea GR |
1386 | /* |
1387 | * Note that there are three places in dbuf_dirty() | |
1388 | * where this dirty record may be put on a list. | |
1389 | * Make sure to do a list_remove corresponding to | |
1390 | * every one of those list_insert calls. | |
1391 | */ | |
34dc7c2f BB |
1392 | if (dr->dr_parent) { |
1393 | mutex_enter(&dr->dr_parent->dt.di.dr_mtx); | |
1394 | list_remove(&dr->dr_parent->dt.di.dr_children, dr); | |
1395 | mutex_exit(&dr->dr_parent->dt.di.dr_mtx); | |
ef3c1dea GR |
1396 | } else if (db->db_blkid == DMU_SPILL_BLKID || |
1397 | db->db_level+1 == dn->dn_nlevels) { | |
b128c09f | 1398 | ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf); |
34dc7c2f BB |
1399 | mutex_enter(&dn->dn_mtx); |
1400 | list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr); | |
1401 | mutex_exit(&dn->dn_mtx); | |
1402 | } | |
572e2857 | 1403 | DB_DNODE_EXIT(db); |
34dc7c2f | 1404 | |
13fe0198 MA |
1405 | if (db->db_state != DB_NOFILL) { |
1406 | dbuf_unoverride(dr); | |
34dc7c2f | 1407 | |
34dc7c2f | 1408 | ASSERT(db->db_buf != NULL); |
13fe0198 MA |
1409 | ASSERT(dr->dt.dl.dr_data != NULL); |
1410 | if (dr->dt.dl.dr_data != db->db_buf) | |
1411 | VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db)); | |
34dc7c2f BB |
1412 | } |
1413 | kmem_free(dr, sizeof (dbuf_dirty_record_t)); | |
1414 | ||
1415 | ASSERT(db->db_dirtycnt > 0); | |
1416 | db->db_dirtycnt -= 1; | |
1417 | ||
1418 | if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) { | |
1419 | arc_buf_t *buf = db->db_buf; | |
1420 | ||
428870ff | 1421 | ASSERT(db->db_state == DB_NOFILL || arc_released(buf)); |
34dc7c2f | 1422 | dbuf_set_data(db, NULL); |
13fe0198 | 1423 | VERIFY(arc_buf_remove_ref(buf, db)); |
34dc7c2f | 1424 | dbuf_evict(db); |
13fe0198 | 1425 | return (B_TRUE); |
34dc7c2f BB |
1426 | } |
1427 | ||
13fe0198 | 1428 | return (B_FALSE); |
34dc7c2f BB |
1429 | } |
1430 | ||
34dc7c2f | 1431 | void |
b0bc7a84 | 1432 | dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx) |
34dc7c2f | 1433 | { |
b0bc7a84 | 1434 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; |
34dc7c2f BB |
1435 | int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH; |
1436 | ||
1437 | ASSERT(tx->tx_txg != 0); | |
1438 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
1439 | ||
572e2857 BB |
1440 | DB_DNODE_ENTER(db); |
1441 | if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock)) | |
34dc7c2f | 1442 | rf |= DB_RF_HAVESTRUCT; |
572e2857 | 1443 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
1444 | (void) dbuf_read(db, NULL, rf); |
1445 | (void) dbuf_dirty(db, tx); | |
1446 | } | |
1447 | ||
b128c09f BB |
1448 | void |
1449 | dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) | |
1450 | { | |
1451 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
1452 | ||
1453 | db->db_state = DB_NOFILL; | |
1454 | ||
1455 | dmu_buf_will_fill(db_fake, tx); | |
1456 | } | |
1457 | ||
34dc7c2f BB |
1458 | void |
1459 | dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) | |
1460 | { | |
1461 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
1462 | ||
428870ff | 1463 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f BB |
1464 | ASSERT(tx->tx_txg != 0); |
1465 | ASSERT(db->db_level == 0); | |
1466 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
1467 | ||
1468 | ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT || | |
1469 | dmu_tx_private_ok(tx)); | |
1470 | ||
1471 | dbuf_noread(db); | |
1472 | (void) dbuf_dirty(db, tx); | |
1473 | } | |
1474 | ||
1475 | #pragma weak dmu_buf_fill_done = dbuf_fill_done | |
1476 | /* ARGSUSED */ | |
1477 | void | |
1478 | dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx) | |
1479 | { | |
1480 | mutex_enter(&db->db_mtx); | |
1481 | DBUF_VERIFY(db); | |
1482 | ||
1483 | if (db->db_state == DB_FILL) { | |
1484 | if (db->db_level == 0 && db->db_freed_in_flight) { | |
428870ff | 1485 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f BB |
1486 | /* we were freed while filling */ |
1487 | /* XXX dbuf_undirty? */ | |
1488 | bzero(db->db.db_data, db->db.db_size); | |
1489 | db->db_freed_in_flight = FALSE; | |
1490 | } | |
1491 | db->db_state = DB_CACHED; | |
1492 | cv_broadcast(&db->db_changed); | |
1493 | } | |
1494 | mutex_exit(&db->db_mtx); | |
1495 | } | |
1496 | ||
9b67f605 MA |
1497 | void |
1498 | dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data, | |
1499 | bp_embedded_type_t etype, enum zio_compress comp, | |
1500 | int uncompressed_size, int compressed_size, int byteorder, | |
1501 | dmu_tx_t *tx) | |
1502 | { | |
1503 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; | |
1504 | struct dirty_leaf *dl; | |
1505 | dmu_object_type_t type; | |
1506 | ||
1507 | DB_DNODE_ENTER(db); | |
1508 | type = DB_DNODE(db)->dn_type; | |
1509 | DB_DNODE_EXIT(db); | |
1510 | ||
1511 | ASSERT0(db->db_level); | |
1512 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); | |
1513 | ||
1514 | dmu_buf_will_not_fill(dbuf, tx); | |
1515 | ||
1516 | ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); | |
1517 | dl = &db->db_last_dirty->dt.dl; | |
1518 | encode_embedded_bp_compressed(&dl->dr_overridden_by, | |
1519 | data, comp, uncompressed_size, compressed_size); | |
1520 | BPE_SET_ETYPE(&dl->dr_overridden_by, etype); | |
1521 | BP_SET_TYPE(&dl->dr_overridden_by, type); | |
1522 | BP_SET_LEVEL(&dl->dr_overridden_by, 0); | |
1523 | BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder); | |
1524 | ||
1525 | dl->dr_override_state = DR_OVERRIDDEN; | |
1526 | dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg; | |
1527 | } | |
1528 | ||
9babb374 BB |
1529 | /* |
1530 | * Directly assign a provided arc buf to a given dbuf if it's not referenced | |
1531 | * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf. | |
1532 | */ | |
1533 | void | |
1534 | dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx) | |
1535 | { | |
1536 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
428870ff | 1537 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
9babb374 BB |
1538 | ASSERT(db->db_level == 0); |
1539 | ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA); | |
1540 | ASSERT(buf != NULL); | |
1541 | ASSERT(arc_buf_size(buf) == db->db.db_size); | |
1542 | ASSERT(tx->tx_txg != 0); | |
1543 | ||
1544 | arc_return_buf(buf, db); | |
1545 | ASSERT(arc_released(buf)); | |
1546 | ||
1547 | mutex_enter(&db->db_mtx); | |
1548 | ||
1549 | while (db->db_state == DB_READ || db->db_state == DB_FILL) | |
1550 | cv_wait(&db->db_changed, &db->db_mtx); | |
1551 | ||
1552 | ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED); | |
1553 | ||
1554 | if (db->db_state == DB_CACHED && | |
1555 | refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) { | |
1556 | mutex_exit(&db->db_mtx); | |
1557 | (void) dbuf_dirty(db, tx); | |
1558 | bcopy(buf->b_data, db->db.db_data, db->db.db_size); | |
13fe0198 | 1559 | VERIFY(arc_buf_remove_ref(buf, db)); |
428870ff | 1560 | xuio_stat_wbuf_copied(); |
9babb374 BB |
1561 | return; |
1562 | } | |
1563 | ||
428870ff | 1564 | xuio_stat_wbuf_nocopy(); |
9babb374 BB |
1565 | if (db->db_state == DB_CACHED) { |
1566 | dbuf_dirty_record_t *dr = db->db_last_dirty; | |
1567 | ||
1568 | ASSERT(db->db_buf != NULL); | |
1569 | if (dr != NULL && dr->dr_txg == tx->tx_txg) { | |
1570 | ASSERT(dr->dt.dl.dr_data == db->db_buf); | |
1571 | if (!arc_released(db->db_buf)) { | |
1572 | ASSERT(dr->dt.dl.dr_override_state == | |
1573 | DR_OVERRIDDEN); | |
1574 | arc_release(db->db_buf, db); | |
1575 | } | |
1576 | dr->dt.dl.dr_data = buf; | |
13fe0198 | 1577 | VERIFY(arc_buf_remove_ref(db->db_buf, db)); |
9babb374 BB |
1578 | } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) { |
1579 | arc_release(db->db_buf, db); | |
13fe0198 | 1580 | VERIFY(arc_buf_remove_ref(db->db_buf, db)); |
9babb374 BB |
1581 | } |
1582 | db->db_buf = NULL; | |
1583 | } | |
1584 | ASSERT(db->db_buf == NULL); | |
1585 | dbuf_set_data(db, buf); | |
1586 | db->db_state = DB_FILL; | |
1587 | mutex_exit(&db->db_mtx); | |
1588 | (void) dbuf_dirty(db, tx); | |
b0bc7a84 | 1589 | dmu_buf_fill_done(&db->db, tx); |
9babb374 BB |
1590 | } |
1591 | ||
34dc7c2f BB |
1592 | /* |
1593 | * "Clear" the contents of this dbuf. This will mark the dbuf | |
e8b96c60 | 1594 | * EVICTING and clear *most* of its references. Unfortunately, |
34dc7c2f BB |
1595 | * when we are not holding the dn_dbufs_mtx, we can't clear the |
1596 | * entry in the dn_dbufs list. We have to wait until dbuf_destroy() | |
1597 | * in this case. For callers from the DMU we will usually see: | |
1598 | * dbuf_clear()->arc_buf_evict()->dbuf_do_evict()->dbuf_destroy() | |
1599 | * For the arc callback, we will usually see: | |
572e2857 | 1600 | * dbuf_do_evict()->dbuf_clear();dbuf_destroy() |
34dc7c2f BB |
1601 | * Sometimes, though, we will get a mix of these two: |
1602 | * DMU: dbuf_clear()->arc_buf_evict() | |
1603 | * ARC: dbuf_do_evict()->dbuf_destroy() | |
1604 | */ | |
1605 | void | |
1606 | dbuf_clear(dmu_buf_impl_t *db) | |
1607 | { | |
572e2857 | 1608 | dnode_t *dn; |
34dc7c2f | 1609 | dmu_buf_impl_t *parent = db->db_parent; |
572e2857 | 1610 | dmu_buf_impl_t *dndb; |
34dc7c2f BB |
1611 | int dbuf_gone = FALSE; |
1612 | ||
1613 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
1614 | ASSERT(refcount_is_zero(&db->db_holds)); | |
1615 | ||
1616 | dbuf_evict_user(db); | |
1617 | ||
1618 | if (db->db_state == DB_CACHED) { | |
1619 | ASSERT(db->db.db_data != NULL); | |
428870ff | 1620 | if (db->db_blkid == DMU_BONUS_BLKID) { |
34dc7c2f | 1621 | zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN); |
d164b209 | 1622 | arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); |
34dc7c2f BB |
1623 | } |
1624 | db->db.db_data = NULL; | |
1625 | db->db_state = DB_UNCACHED; | |
1626 | } | |
1627 | ||
b128c09f | 1628 | ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL); |
34dc7c2f BB |
1629 | ASSERT(db->db_data_pending == NULL); |
1630 | ||
1631 | db->db_state = DB_EVICTING; | |
1632 | db->db_blkptr = NULL; | |
1633 | ||
572e2857 BB |
1634 | DB_DNODE_ENTER(db); |
1635 | dn = DB_DNODE(db); | |
1636 | dndb = dn->dn_dbuf; | |
428870ff | 1637 | if (db->db_blkid != DMU_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) { |
34dc7c2f | 1638 | list_remove(&dn->dn_dbufs, db); |
572e2857 BB |
1639 | (void) atomic_dec_32_nv(&dn->dn_dbufs_count); |
1640 | membar_producer(); | |
1641 | DB_DNODE_EXIT(db); | |
1642 | /* | |
1643 | * Decrementing the dbuf count means that the hold corresponding | |
1644 | * to the removed dbuf is no longer discounted in dnode_move(), | |
1645 | * so the dnode cannot be moved until after we release the hold. | |
1646 | * The membar_producer() ensures visibility of the decremented | |
1647 | * value in dnode_move(), since DB_DNODE_EXIT doesn't actually | |
1648 | * release any lock. | |
1649 | */ | |
34dc7c2f | 1650 | dnode_rele(dn, db); |
572e2857 BB |
1651 | db->db_dnode_handle = NULL; |
1652 | } else { | |
1653 | DB_DNODE_EXIT(db); | |
34dc7c2f BB |
1654 | } |
1655 | ||
1656 | if (db->db_buf) | |
1657 | dbuf_gone = arc_buf_evict(db->db_buf); | |
1658 | ||
1659 | if (!dbuf_gone) | |
1660 | mutex_exit(&db->db_mtx); | |
1661 | ||
1662 | /* | |
572e2857 | 1663 | * If this dbuf is referenced from an indirect dbuf, |
34dc7c2f BB |
1664 | * decrement the ref count on the indirect dbuf. |
1665 | */ | |
1666 | if (parent && parent != dndb) | |
1667 | dbuf_rele(parent, db); | |
1668 | } | |
1669 | ||
bf701a83 BB |
1670 | __attribute__((always_inline)) |
1671 | static inline int | |
34dc7c2f | 1672 | dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse, |
fc5bb51f | 1673 | dmu_buf_impl_t **parentp, blkptr_t **bpp, struct dbuf_hold_impl_data *dh) |
34dc7c2f BB |
1674 | { |
1675 | int nlevels, epbs; | |
1676 | ||
1677 | *parentp = NULL; | |
1678 | *bpp = NULL; | |
1679 | ||
428870ff BB |
1680 | ASSERT(blkid != DMU_BONUS_BLKID); |
1681 | ||
1682 | if (blkid == DMU_SPILL_BLKID) { | |
1683 | mutex_enter(&dn->dn_mtx); | |
1684 | if (dn->dn_have_spill && | |
1685 | (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) | |
1686 | *bpp = &dn->dn_phys->dn_spill; | |
1687 | else | |
1688 | *bpp = NULL; | |
1689 | dbuf_add_ref(dn->dn_dbuf, NULL); | |
1690 | *parentp = dn->dn_dbuf; | |
1691 | mutex_exit(&dn->dn_mtx); | |
1692 | return (0); | |
1693 | } | |
34dc7c2f BB |
1694 | |
1695 | if (dn->dn_phys->dn_nlevels == 0) | |
1696 | nlevels = 1; | |
1697 | else | |
1698 | nlevels = dn->dn_phys->dn_nlevels; | |
1699 | ||
1700 | epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; | |
1701 | ||
1702 | ASSERT3U(level * epbs, <, 64); | |
1703 | ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); | |
1704 | if (level >= nlevels || | |
1705 | (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) { | |
1706 | /* the buffer has no parent yet */ | |
2e528b49 | 1707 | return (SET_ERROR(ENOENT)); |
34dc7c2f BB |
1708 | } else if (level < nlevels-1) { |
1709 | /* this block is referenced from an indirect block */ | |
fc5bb51f BB |
1710 | int err; |
1711 | if (dh == NULL) { | |
1712 | err = dbuf_hold_impl(dn, level+1, blkid >> epbs, | |
1713 | fail_sparse, NULL, parentp); | |
d1d7e268 | 1714 | } else { |
fc5bb51f BB |
1715 | __dbuf_hold_impl_init(dh + 1, dn, dh->dh_level + 1, |
1716 | blkid >> epbs, fail_sparse, NULL, | |
1717 | parentp, dh->dh_depth + 1); | |
1718 | err = __dbuf_hold_impl(dh + 1); | |
1719 | } | |
34dc7c2f BB |
1720 | if (err) |
1721 | return (err); | |
1722 | err = dbuf_read(*parentp, NULL, | |
1723 | (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL)); | |
1724 | if (err) { | |
1725 | dbuf_rele(*parentp, NULL); | |
1726 | *parentp = NULL; | |
1727 | return (err); | |
1728 | } | |
1729 | *bpp = ((blkptr_t *)(*parentp)->db.db_data) + | |
1730 | (blkid & ((1ULL << epbs) - 1)); | |
1731 | return (0); | |
1732 | } else { | |
1733 | /* the block is referenced from the dnode */ | |
1734 | ASSERT3U(level, ==, nlevels-1); | |
1735 | ASSERT(dn->dn_phys->dn_nblkptr == 0 || | |
1736 | blkid < dn->dn_phys->dn_nblkptr); | |
1737 | if (dn->dn_dbuf) { | |
1738 | dbuf_add_ref(dn->dn_dbuf, NULL); | |
1739 | *parentp = dn->dn_dbuf; | |
1740 | } | |
1741 | *bpp = &dn->dn_phys->dn_blkptr[blkid]; | |
1742 | return (0); | |
1743 | } | |
1744 | } | |
1745 | ||
1746 | static dmu_buf_impl_t * | |
1747 | dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid, | |
1748 | dmu_buf_impl_t *parent, blkptr_t *blkptr) | |
1749 | { | |
428870ff | 1750 | objset_t *os = dn->dn_objset; |
34dc7c2f BB |
1751 | dmu_buf_impl_t *db, *odb; |
1752 | ||
1753 | ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); | |
1754 | ASSERT(dn->dn_type != DMU_OT_NONE); | |
1755 | ||
b8d06fca | 1756 | db = kmem_cache_alloc(dbuf_cache, KM_PUSHPAGE); |
34dc7c2f BB |
1757 | |
1758 | db->db_objset = os; | |
1759 | db->db.db_object = dn->dn_object; | |
1760 | db->db_level = level; | |
1761 | db->db_blkid = blkid; | |
1762 | db->db_last_dirty = NULL; | |
1763 | db->db_dirtycnt = 0; | |
572e2857 | 1764 | db->db_dnode_handle = dn->dn_handle; |
34dc7c2f BB |
1765 | db->db_parent = parent; |
1766 | db->db_blkptr = blkptr; | |
1767 | ||
1768 | db->db_user_ptr = NULL; | |
1769 | db->db_user_data_ptr_ptr = NULL; | |
1770 | db->db_evict_func = NULL; | |
1771 | db->db_immediate_evict = 0; | |
1772 | db->db_freed_in_flight = 0; | |
1773 | ||
428870ff | 1774 | if (blkid == DMU_BONUS_BLKID) { |
34dc7c2f BB |
1775 | ASSERT3P(parent, ==, dn->dn_dbuf); |
1776 | db->db.db_size = DN_MAX_BONUSLEN - | |
1777 | (dn->dn_nblkptr-1) * sizeof (blkptr_t); | |
1778 | ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); | |
428870ff | 1779 | db->db.db_offset = DMU_BONUS_BLKID; |
34dc7c2f BB |
1780 | db->db_state = DB_UNCACHED; |
1781 | /* the bonus dbuf is not placed in the hash table */ | |
d164b209 | 1782 | arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); |
34dc7c2f | 1783 | return (db); |
428870ff BB |
1784 | } else if (blkid == DMU_SPILL_BLKID) { |
1785 | db->db.db_size = (blkptr != NULL) ? | |
1786 | BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE; | |
1787 | db->db.db_offset = 0; | |
34dc7c2f BB |
1788 | } else { |
1789 | int blocksize = | |
e8b96c60 | 1790 | db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz; |
34dc7c2f BB |
1791 | db->db.db_size = blocksize; |
1792 | db->db.db_offset = db->db_blkid * blocksize; | |
1793 | } | |
1794 | ||
1795 | /* | |
1796 | * Hold the dn_dbufs_mtx while we get the new dbuf | |
1797 | * in the hash table *and* added to the dbufs list. | |
1798 | * This prevents a possible deadlock with someone | |
1799 | * trying to look up this dbuf before its added to the | |
1800 | * dn_dbufs list. | |
1801 | */ | |
1802 | mutex_enter(&dn->dn_dbufs_mtx); | |
1803 | db->db_state = DB_EVICTING; | |
1804 | if ((odb = dbuf_hash_insert(db)) != NULL) { | |
1805 | /* someone else inserted it first */ | |
1806 | kmem_cache_free(dbuf_cache, db); | |
1807 | mutex_exit(&dn->dn_dbufs_mtx); | |
1808 | return (odb); | |
1809 | } | |
1810 | list_insert_head(&dn->dn_dbufs, db); | |
b663a23d MA |
1811 | if (db->db_level == 0 && db->db_blkid >= |
1812 | dn->dn_unlisted_l0_blkid) | |
1813 | dn->dn_unlisted_l0_blkid = db->db_blkid + 1; | |
34dc7c2f BB |
1814 | db->db_state = DB_UNCACHED; |
1815 | mutex_exit(&dn->dn_dbufs_mtx); | |
d164b209 | 1816 | arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); |
34dc7c2f BB |
1817 | |
1818 | if (parent && parent != dn->dn_dbuf) | |
1819 | dbuf_add_ref(parent, db); | |
1820 | ||
1821 | ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || | |
1822 | refcount_count(&dn->dn_holds) > 0); | |
1823 | (void) refcount_add(&dn->dn_holds, db); | |
572e2857 | 1824 | (void) atomic_inc_32_nv(&dn->dn_dbufs_count); |
34dc7c2f BB |
1825 | |
1826 | dprintf_dbuf(db, "db=%p\n", db); | |
1827 | ||
1828 | return (db); | |
1829 | } | |
1830 | ||
1831 | static int | |
1832 | dbuf_do_evict(void *private) | |
1833 | { | |
1834 | arc_buf_t *buf = private; | |
1835 | dmu_buf_impl_t *db = buf->b_private; | |
1836 | ||
1837 | if (!MUTEX_HELD(&db->db_mtx)) | |
1838 | mutex_enter(&db->db_mtx); | |
1839 | ||
1840 | ASSERT(refcount_is_zero(&db->db_holds)); | |
1841 | ||
1842 | if (db->db_state != DB_EVICTING) { | |
1843 | ASSERT(db->db_state == DB_CACHED); | |
1844 | DBUF_VERIFY(db); | |
1845 | db->db_buf = NULL; | |
1846 | dbuf_evict(db); | |
1847 | } else { | |
1848 | mutex_exit(&db->db_mtx); | |
1849 | dbuf_destroy(db); | |
1850 | } | |
1851 | return (0); | |
1852 | } | |
1853 | ||
1854 | static void | |
1855 | dbuf_destroy(dmu_buf_impl_t *db) | |
1856 | { | |
1857 | ASSERT(refcount_is_zero(&db->db_holds)); | |
1858 | ||
428870ff | 1859 | if (db->db_blkid != DMU_BONUS_BLKID) { |
34dc7c2f BB |
1860 | /* |
1861 | * If this dbuf is still on the dn_dbufs list, | |
1862 | * remove it from that list. | |
1863 | */ | |
572e2857 BB |
1864 | if (db->db_dnode_handle != NULL) { |
1865 | dnode_t *dn; | |
34dc7c2f | 1866 | |
572e2857 BB |
1867 | DB_DNODE_ENTER(db); |
1868 | dn = DB_DNODE(db); | |
34dc7c2f BB |
1869 | mutex_enter(&dn->dn_dbufs_mtx); |
1870 | list_remove(&dn->dn_dbufs, db); | |
572e2857 | 1871 | (void) atomic_dec_32_nv(&dn->dn_dbufs_count); |
34dc7c2f | 1872 | mutex_exit(&dn->dn_dbufs_mtx); |
572e2857 BB |
1873 | DB_DNODE_EXIT(db); |
1874 | /* | |
1875 | * Decrementing the dbuf count means that the hold | |
1876 | * corresponding to the removed dbuf is no longer | |
1877 | * discounted in dnode_move(), so the dnode cannot be | |
1878 | * moved until after we release the hold. | |
1879 | */ | |
34dc7c2f | 1880 | dnode_rele(dn, db); |
572e2857 | 1881 | db->db_dnode_handle = NULL; |
34dc7c2f BB |
1882 | } |
1883 | dbuf_hash_remove(db); | |
1884 | } | |
1885 | db->db_parent = NULL; | |
1886 | db->db_buf = NULL; | |
1887 | ||
1888 | ASSERT(!list_link_active(&db->db_link)); | |
1889 | ASSERT(db->db.db_data == NULL); | |
1890 | ASSERT(db->db_hash_next == NULL); | |
1891 | ASSERT(db->db_blkptr == NULL); | |
1892 | ASSERT(db->db_data_pending == NULL); | |
1893 | ||
1894 | kmem_cache_free(dbuf_cache, db); | |
d164b209 | 1895 | arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); |
34dc7c2f BB |
1896 | } |
1897 | ||
1898 | void | |
e8b96c60 | 1899 | dbuf_prefetch(dnode_t *dn, uint64_t blkid, zio_priority_t prio) |
34dc7c2f BB |
1900 | { |
1901 | dmu_buf_impl_t *db = NULL; | |
1902 | blkptr_t *bp = NULL; | |
1903 | ||
428870ff | 1904 | ASSERT(blkid != DMU_BONUS_BLKID); |
34dc7c2f BB |
1905 | ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); |
1906 | ||
1907 | if (dnode_block_freed(dn, blkid)) | |
1908 | return; | |
1909 | ||
1910 | /* dbuf_find() returns with db_mtx held */ | |
c65aa5b2 | 1911 | if ((db = dbuf_find(dn, 0, blkid))) { |
572e2857 BB |
1912 | /* |
1913 | * This dbuf is already in the cache. We assume that | |
1914 | * it is already CACHED, or else about to be either | |
1915 | * read or filled. | |
1916 | */ | |
34dc7c2f | 1917 | mutex_exit(&db->db_mtx); |
572e2857 | 1918 | return; |
34dc7c2f BB |
1919 | } |
1920 | ||
fc5bb51f | 1921 | if (dbuf_findbp(dn, 0, blkid, TRUE, &db, &bp, NULL) == 0) { |
9b67f605 | 1922 | if (bp && !BP_IS_HOLE(bp) && !BP_IS_EMBEDDED(bp)) { |
428870ff | 1923 | dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset; |
34dc7c2f BB |
1924 | uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH; |
1925 | zbookmark_t zb; | |
428870ff BB |
1926 | |
1927 | SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET, | |
1928 | dn->dn_object, 0, blkid); | |
34dc7c2f | 1929 | |
294f6806 | 1930 | (void) arc_read(NULL, dn->dn_objset->os_spa, |
e8b96c60 | 1931 | bp, NULL, NULL, prio, |
34dc7c2f BB |
1932 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, |
1933 | &aflags, &zb); | |
1934 | } | |
1935 | if (db) | |
1936 | dbuf_rele(db, NULL); | |
1937 | } | |
1938 | } | |
1939 | ||
d1d7e268 | 1940 | #define DBUF_HOLD_IMPL_MAX_DEPTH 20 |
fc5bb51f | 1941 | |
34dc7c2f BB |
1942 | /* |
1943 | * Returns with db_holds incremented, and db_mtx not held. | |
1944 | * Note: dn_struct_rwlock must be held. | |
1945 | */ | |
fc5bb51f BB |
1946 | static int |
1947 | __dbuf_hold_impl(struct dbuf_hold_impl_data *dh) | |
34dc7c2f | 1948 | { |
fc5bb51f BB |
1949 | ASSERT3S(dh->dh_depth, <, DBUF_HOLD_IMPL_MAX_DEPTH); |
1950 | dh->dh_parent = NULL; | |
34dc7c2f | 1951 | |
fc5bb51f BB |
1952 | ASSERT(dh->dh_blkid != DMU_BONUS_BLKID); |
1953 | ASSERT(RW_LOCK_HELD(&dh->dh_dn->dn_struct_rwlock)); | |
1954 | ASSERT3U(dh->dh_dn->dn_nlevels, >, dh->dh_level); | |
34dc7c2f | 1955 | |
fc5bb51f | 1956 | *(dh->dh_dbp) = NULL; |
34dc7c2f BB |
1957 | top: |
1958 | /* dbuf_find() returns with db_mtx held */ | |
fc5bb51f BB |
1959 | dh->dh_db = dbuf_find(dh->dh_dn, dh->dh_level, dh->dh_blkid); |
1960 | ||
1961 | if (dh->dh_db == NULL) { | |
1962 | dh->dh_bp = NULL; | |
1963 | ||
1964 | ASSERT3P(dh->dh_parent, ==, NULL); | |
1965 | dh->dh_err = dbuf_findbp(dh->dh_dn, dh->dh_level, dh->dh_blkid, | |
1966 | dh->dh_fail_sparse, &dh->dh_parent, | |
1967 | &dh->dh_bp, dh); | |
1968 | if (dh->dh_fail_sparse) { | |
d1d7e268 MK |
1969 | if (dh->dh_err == 0 && |
1970 | dh->dh_bp && BP_IS_HOLE(dh->dh_bp)) | |
2e528b49 | 1971 | dh->dh_err = SET_ERROR(ENOENT); |
fc5bb51f BB |
1972 | if (dh->dh_err) { |
1973 | if (dh->dh_parent) | |
1974 | dbuf_rele(dh->dh_parent, NULL); | |
1975 | return (dh->dh_err); | |
34dc7c2f BB |
1976 | } |
1977 | } | |
fc5bb51f BB |
1978 | if (dh->dh_err && dh->dh_err != ENOENT) |
1979 | return (dh->dh_err); | |
1980 | dh->dh_db = dbuf_create(dh->dh_dn, dh->dh_level, dh->dh_blkid, | |
1981 | dh->dh_parent, dh->dh_bp); | |
34dc7c2f BB |
1982 | } |
1983 | ||
fc5bb51f BB |
1984 | if (dh->dh_db->db_buf && refcount_is_zero(&dh->dh_db->db_holds)) { |
1985 | arc_buf_add_ref(dh->dh_db->db_buf, dh->dh_db); | |
1986 | if (dh->dh_db->db_buf->b_data == NULL) { | |
1987 | dbuf_clear(dh->dh_db); | |
1988 | if (dh->dh_parent) { | |
1989 | dbuf_rele(dh->dh_parent, NULL); | |
1990 | dh->dh_parent = NULL; | |
34dc7c2f BB |
1991 | } |
1992 | goto top; | |
1993 | } | |
fc5bb51f | 1994 | ASSERT3P(dh->dh_db->db.db_data, ==, dh->dh_db->db_buf->b_data); |
34dc7c2f BB |
1995 | } |
1996 | ||
fc5bb51f | 1997 | ASSERT(dh->dh_db->db_buf == NULL || arc_referenced(dh->dh_db->db_buf)); |
34dc7c2f BB |
1998 | |
1999 | /* | |
2000 | * If this buffer is currently syncing out, and we are are | |
2001 | * still referencing it from db_data, we need to make a copy | |
2002 | * of it in case we decide we want to dirty it again in this txg. | |
2003 | */ | |
fc5bb51f BB |
2004 | if (dh->dh_db->db_level == 0 && |
2005 | dh->dh_db->db_blkid != DMU_BONUS_BLKID && | |
2006 | dh->dh_dn->dn_object != DMU_META_DNODE_OBJECT && | |
2007 | dh->dh_db->db_state == DB_CACHED && dh->dh_db->db_data_pending) { | |
2008 | dh->dh_dr = dh->dh_db->db_data_pending; | |
2009 | ||
2010 | if (dh->dh_dr->dt.dl.dr_data == dh->dh_db->db_buf) { | |
2011 | dh->dh_type = DBUF_GET_BUFC_TYPE(dh->dh_db); | |
2012 | ||
2013 | dbuf_set_data(dh->dh_db, | |
2014 | arc_buf_alloc(dh->dh_dn->dn_objset->os_spa, | |
2015 | dh->dh_db->db.db_size, dh->dh_db, dh->dh_type)); | |
2016 | bcopy(dh->dh_dr->dt.dl.dr_data->b_data, | |
2017 | dh->dh_db->db.db_data, dh->dh_db->db.db_size); | |
34dc7c2f BB |
2018 | } |
2019 | } | |
2020 | ||
fc5bb51f BB |
2021 | (void) refcount_add(&dh->dh_db->db_holds, dh->dh_tag); |
2022 | dbuf_update_data(dh->dh_db); | |
2023 | DBUF_VERIFY(dh->dh_db); | |
2024 | mutex_exit(&dh->dh_db->db_mtx); | |
34dc7c2f BB |
2025 | |
2026 | /* NOTE: we can't rele the parent until after we drop the db_mtx */ | |
fc5bb51f BB |
2027 | if (dh->dh_parent) |
2028 | dbuf_rele(dh->dh_parent, NULL); | |
34dc7c2f | 2029 | |
fc5bb51f BB |
2030 | ASSERT3P(DB_DNODE(dh->dh_db), ==, dh->dh_dn); |
2031 | ASSERT3U(dh->dh_db->db_blkid, ==, dh->dh_blkid); | |
2032 | ASSERT3U(dh->dh_db->db_level, ==, dh->dh_level); | |
2033 | *(dh->dh_dbp) = dh->dh_db; | |
34dc7c2f BB |
2034 | |
2035 | return (0); | |
2036 | } | |
2037 | ||
fc5bb51f BB |
2038 | /* |
2039 | * The following code preserves the recursive function dbuf_hold_impl() | |
2040 | * but moves the local variables AND function arguments to the heap to | |
2041 | * minimize the stack frame size. Enough space is initially allocated | |
2042 | * on the stack for 20 levels of recursion. | |
2043 | */ | |
2044 | int | |
2045 | dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse, | |
2046 | void *tag, dmu_buf_impl_t **dbp) | |
2047 | { | |
2048 | struct dbuf_hold_impl_data *dh; | |
2049 | int error; | |
2050 | ||
d1d7e268 | 2051 | dh = kmem_zalloc(sizeof (struct dbuf_hold_impl_data) * |
b8d06fca | 2052 | DBUF_HOLD_IMPL_MAX_DEPTH, KM_PUSHPAGE); |
fc5bb51f BB |
2053 | __dbuf_hold_impl_init(dh, dn, level, blkid, fail_sparse, tag, dbp, 0); |
2054 | ||
2055 | error = __dbuf_hold_impl(dh); | |
2056 | ||
d1d7e268 | 2057 | kmem_free(dh, sizeof (struct dbuf_hold_impl_data) * |
fc5bb51f BB |
2058 | DBUF_HOLD_IMPL_MAX_DEPTH); |
2059 | ||
2060 | return (error); | |
2061 | } | |
2062 | ||
2063 | static void | |
2064 | __dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh, | |
2065 | dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse, | |
2066 | void *tag, dmu_buf_impl_t **dbp, int depth) | |
2067 | { | |
2068 | dh->dh_dn = dn; | |
2069 | dh->dh_level = level; | |
2070 | dh->dh_blkid = blkid; | |
2071 | dh->dh_fail_sparse = fail_sparse; | |
2072 | dh->dh_tag = tag; | |
2073 | dh->dh_dbp = dbp; | |
2074 | dh->dh_depth = depth; | |
2075 | } | |
2076 | ||
34dc7c2f BB |
2077 | dmu_buf_impl_t * |
2078 | dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag) | |
2079 | { | |
2080 | dmu_buf_impl_t *db; | |
2081 | int err = dbuf_hold_impl(dn, 0, blkid, FALSE, tag, &db); | |
2082 | return (err ? NULL : db); | |
2083 | } | |
2084 | ||
2085 | dmu_buf_impl_t * | |
2086 | dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag) | |
2087 | { | |
2088 | dmu_buf_impl_t *db; | |
2089 | int err = dbuf_hold_impl(dn, level, blkid, FALSE, tag, &db); | |
2090 | return (err ? NULL : db); | |
2091 | } | |
2092 | ||
2093 | void | |
2094 | dbuf_create_bonus(dnode_t *dn) | |
2095 | { | |
2096 | ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); | |
2097 | ||
2098 | ASSERT(dn->dn_bonus == NULL); | |
428870ff BB |
2099 | dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL); |
2100 | } | |
2101 | ||
2102 | int | |
2103 | dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx) | |
2104 | { | |
2105 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
572e2857 BB |
2106 | dnode_t *dn; |
2107 | ||
428870ff | 2108 | if (db->db_blkid != DMU_SPILL_BLKID) |
2e528b49 | 2109 | return (SET_ERROR(ENOTSUP)); |
428870ff BB |
2110 | if (blksz == 0) |
2111 | blksz = SPA_MINBLOCKSIZE; | |
2112 | if (blksz > SPA_MAXBLOCKSIZE) | |
2113 | blksz = SPA_MAXBLOCKSIZE; | |
2114 | else | |
2115 | blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE); | |
2116 | ||
572e2857 BB |
2117 | DB_DNODE_ENTER(db); |
2118 | dn = DB_DNODE(db); | |
2119 | rw_enter(&dn->dn_struct_rwlock, RW_WRITER); | |
428870ff | 2120 | dbuf_new_size(db, blksz, tx); |
572e2857 BB |
2121 | rw_exit(&dn->dn_struct_rwlock); |
2122 | DB_DNODE_EXIT(db); | |
428870ff BB |
2123 | |
2124 | return (0); | |
2125 | } | |
2126 | ||
2127 | void | |
2128 | dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx) | |
2129 | { | |
2130 | dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx); | |
34dc7c2f BB |
2131 | } |
2132 | ||
2133 | #pragma weak dmu_buf_add_ref = dbuf_add_ref | |
2134 | void | |
2135 | dbuf_add_ref(dmu_buf_impl_t *db, void *tag) | |
2136 | { | |
1fde1e37 | 2137 | VERIFY(refcount_add(&db->db_holds, tag) > 1); |
34dc7c2f BB |
2138 | } |
2139 | ||
572e2857 BB |
2140 | /* |
2141 | * If you call dbuf_rele() you had better not be referencing the dnode handle | |
2142 | * unless you have some other direct or indirect hold on the dnode. (An indirect | |
2143 | * hold is a hold on one of the dnode's dbufs, including the bonus buffer.) | |
2144 | * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the | |
2145 | * dnode's parent dbuf evicting its dnode handles. | |
2146 | */ | |
34dc7c2f BB |
2147 | void |
2148 | dbuf_rele(dmu_buf_impl_t *db, void *tag) | |
428870ff BB |
2149 | { |
2150 | mutex_enter(&db->db_mtx); | |
2151 | dbuf_rele_and_unlock(db, tag); | |
2152 | } | |
2153 | ||
b0bc7a84 MG |
2154 | void |
2155 | dmu_buf_rele(dmu_buf_t *db, void *tag) | |
2156 | { | |
2157 | dbuf_rele((dmu_buf_impl_t *)db, tag); | |
2158 | } | |
2159 | ||
428870ff BB |
2160 | /* |
2161 | * dbuf_rele() for an already-locked dbuf. This is necessary to allow | |
2162 | * db_dirtycnt and db_holds to be updated atomically. | |
2163 | */ | |
2164 | void | |
2165 | dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag) | |
34dc7c2f BB |
2166 | { |
2167 | int64_t holds; | |
2168 | ||
428870ff | 2169 | ASSERT(MUTEX_HELD(&db->db_mtx)); |
34dc7c2f BB |
2170 | DBUF_VERIFY(db); |
2171 | ||
572e2857 BB |
2172 | /* |
2173 | * Remove the reference to the dbuf before removing its hold on the | |
2174 | * dnode so we can guarantee in dnode_move() that a referenced bonus | |
2175 | * buffer has a corresponding dnode hold. | |
2176 | */ | |
34dc7c2f BB |
2177 | holds = refcount_remove(&db->db_holds, tag); |
2178 | ASSERT(holds >= 0); | |
2179 | ||
2180 | /* | |
2181 | * We can't freeze indirects if there is a possibility that they | |
2182 | * may be modified in the current syncing context. | |
2183 | */ | |
2184 | if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) | |
2185 | arc_buf_freeze(db->db_buf); | |
2186 | ||
2187 | if (holds == db->db_dirtycnt && | |
2188 | db->db_level == 0 && db->db_immediate_evict) | |
2189 | dbuf_evict_user(db); | |
2190 | ||
2191 | if (holds == 0) { | |
428870ff | 2192 | if (db->db_blkid == DMU_BONUS_BLKID) { |
34dc7c2f | 2193 | mutex_exit(&db->db_mtx); |
572e2857 BB |
2194 | |
2195 | /* | |
2196 | * If the dnode moves here, we cannot cross this barrier | |
2197 | * until the move completes. | |
2198 | */ | |
2199 | DB_DNODE_ENTER(db); | |
2200 | (void) atomic_dec_32_nv(&DB_DNODE(db)->dn_dbufs_count); | |
2201 | DB_DNODE_EXIT(db); | |
2202 | /* | |
2203 | * The bonus buffer's dnode hold is no longer discounted | |
2204 | * in dnode_move(). The dnode cannot move until after | |
2205 | * the dnode_rele(). | |
2206 | */ | |
2207 | dnode_rele(DB_DNODE(db), db); | |
34dc7c2f BB |
2208 | } else if (db->db_buf == NULL) { |
2209 | /* | |
2210 | * This is a special case: we never associated this | |
2211 | * dbuf with any data allocated from the ARC. | |
2212 | */ | |
b128c09f BB |
2213 | ASSERT(db->db_state == DB_UNCACHED || |
2214 | db->db_state == DB_NOFILL); | |
34dc7c2f BB |
2215 | dbuf_evict(db); |
2216 | } else if (arc_released(db->db_buf)) { | |
2217 | arc_buf_t *buf = db->db_buf; | |
2218 | /* | |
2219 | * This dbuf has anonymous data associated with it. | |
2220 | */ | |
2221 | dbuf_set_data(db, NULL); | |
13fe0198 | 2222 | VERIFY(arc_buf_remove_ref(buf, db)); |
34dc7c2f BB |
2223 | dbuf_evict(db); |
2224 | } else { | |
13fe0198 | 2225 | VERIFY(!arc_buf_remove_ref(db->db_buf, db)); |
1eb5bfa3 GW |
2226 | |
2227 | /* | |
2228 | * A dbuf will be eligible for eviction if either the | |
2229 | * 'primarycache' property is set or a duplicate | |
2230 | * copy of this buffer is already cached in the arc. | |
2231 | * | |
2232 | * In the case of the 'primarycache' a buffer | |
2233 | * is considered for eviction if it matches the | |
2234 | * criteria set in the property. | |
2235 | * | |
2236 | * To decide if our buffer is considered a | |
2237 | * duplicate, we must call into the arc to determine | |
2238 | * if multiple buffers are referencing the same | |
2239 | * block on-disk. If so, then we simply evict | |
2240 | * ourselves. | |
2241 | */ | |
2242 | if (!DBUF_IS_CACHEABLE(db) || | |
2243 | arc_buf_eviction_needed(db->db_buf)) | |
b128c09f BB |
2244 | dbuf_clear(db); |
2245 | else | |
2246 | mutex_exit(&db->db_mtx); | |
34dc7c2f BB |
2247 | } |
2248 | } else { | |
2249 | mutex_exit(&db->db_mtx); | |
2250 | } | |
2251 | } | |
2252 | ||
2253 | #pragma weak dmu_buf_refcount = dbuf_refcount | |
2254 | uint64_t | |
2255 | dbuf_refcount(dmu_buf_impl_t *db) | |
2256 | { | |
2257 | return (refcount_count(&db->db_holds)); | |
2258 | } | |
2259 | ||
2260 | void * | |
2261 | dmu_buf_set_user(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr, | |
2262 | dmu_buf_evict_func_t *evict_func) | |
2263 | { | |
2264 | return (dmu_buf_update_user(db_fake, NULL, user_ptr, | |
2265 | user_data_ptr_ptr, evict_func)); | |
2266 | } | |
2267 | ||
2268 | void * | |
2269 | dmu_buf_set_user_ie(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr, | |
2270 | dmu_buf_evict_func_t *evict_func) | |
2271 | { | |
2272 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
2273 | ||
2274 | db->db_immediate_evict = TRUE; | |
2275 | return (dmu_buf_update_user(db_fake, NULL, user_ptr, | |
2276 | user_data_ptr_ptr, evict_func)); | |
2277 | } | |
2278 | ||
2279 | void * | |
2280 | dmu_buf_update_user(dmu_buf_t *db_fake, void *old_user_ptr, void *user_ptr, | |
2281 | void *user_data_ptr_ptr, dmu_buf_evict_func_t *evict_func) | |
2282 | { | |
2283 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
2284 | ASSERT(db->db_level == 0); | |
2285 | ||
2286 | ASSERT((user_ptr == NULL) == (evict_func == NULL)); | |
2287 | ||
2288 | mutex_enter(&db->db_mtx); | |
2289 | ||
2290 | if (db->db_user_ptr == old_user_ptr) { | |
2291 | db->db_user_ptr = user_ptr; | |
2292 | db->db_user_data_ptr_ptr = user_data_ptr_ptr; | |
2293 | db->db_evict_func = evict_func; | |
2294 | ||
2295 | dbuf_update_data(db); | |
2296 | } else { | |
2297 | old_user_ptr = db->db_user_ptr; | |
2298 | } | |
2299 | ||
2300 | mutex_exit(&db->db_mtx); | |
2301 | return (old_user_ptr); | |
2302 | } | |
2303 | ||
2304 | void * | |
2305 | dmu_buf_get_user(dmu_buf_t *db_fake) | |
2306 | { | |
2307 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
2308 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
2309 | ||
2310 | return (db->db_user_ptr); | |
2311 | } | |
2312 | ||
9babb374 BB |
2313 | boolean_t |
2314 | dmu_buf_freeable(dmu_buf_t *dbuf) | |
2315 | { | |
2316 | boolean_t res = B_FALSE; | |
2317 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; | |
2318 | ||
2319 | if (db->db_blkptr) | |
2320 | res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset, | |
428870ff | 2321 | db->db_blkptr, db->db_blkptr->blk_birth); |
9babb374 BB |
2322 | |
2323 | return (res); | |
2324 | } | |
2325 | ||
03c6040b GW |
2326 | blkptr_t * |
2327 | dmu_buf_get_blkptr(dmu_buf_t *db) | |
2328 | { | |
2329 | dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; | |
2330 | return (dbi->db_blkptr); | |
2331 | } | |
2332 | ||
34dc7c2f BB |
2333 | static void |
2334 | dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db) | |
2335 | { | |
2336 | /* ASSERT(dmu_tx_is_syncing(tx) */ | |
2337 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
2338 | ||
2339 | if (db->db_blkptr != NULL) | |
2340 | return; | |
2341 | ||
428870ff BB |
2342 | if (db->db_blkid == DMU_SPILL_BLKID) { |
2343 | db->db_blkptr = &dn->dn_phys->dn_spill; | |
2344 | BP_ZERO(db->db_blkptr); | |
2345 | return; | |
2346 | } | |
34dc7c2f BB |
2347 | if (db->db_level == dn->dn_phys->dn_nlevels-1) { |
2348 | /* | |
2349 | * This buffer was allocated at a time when there was | |
2350 | * no available blkptrs from the dnode, or it was | |
2351 | * inappropriate to hook it in (i.e., nlevels mis-match). | |
2352 | */ | |
2353 | ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr); | |
2354 | ASSERT(db->db_parent == NULL); | |
2355 | db->db_parent = dn->dn_dbuf; | |
2356 | db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid]; | |
2357 | DBUF_VERIFY(db); | |
2358 | } else { | |
2359 | dmu_buf_impl_t *parent = db->db_parent; | |
2360 | int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; | |
2361 | ||
2362 | ASSERT(dn->dn_phys->dn_nlevels > 1); | |
2363 | if (parent == NULL) { | |
2364 | mutex_exit(&db->db_mtx); | |
2365 | rw_enter(&dn->dn_struct_rwlock, RW_READER); | |
2366 | (void) dbuf_hold_impl(dn, db->db_level+1, | |
2367 | db->db_blkid >> epbs, FALSE, db, &parent); | |
2368 | rw_exit(&dn->dn_struct_rwlock); | |
2369 | mutex_enter(&db->db_mtx); | |
2370 | db->db_parent = parent; | |
2371 | } | |
2372 | db->db_blkptr = (blkptr_t *)parent->db.db_data + | |
2373 | (db->db_blkid & ((1ULL << epbs) - 1)); | |
2374 | DBUF_VERIFY(db); | |
2375 | } | |
2376 | } | |
2377 | ||
d1d7e268 MK |
2378 | /* |
2379 | * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it | |
60948de1 BB |
2380 | * is critical the we not allow the compiler to inline this function in to |
2381 | * dbuf_sync_list() thereby drastically bloating the stack usage. | |
2382 | */ | |
2383 | noinline static void | |
34dc7c2f BB |
2384 | dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx) |
2385 | { | |
2386 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
572e2857 | 2387 | dnode_t *dn; |
34dc7c2f BB |
2388 | zio_t *zio; |
2389 | ||
2390 | ASSERT(dmu_tx_is_syncing(tx)); | |
2391 | ||
2392 | dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); | |
2393 | ||
2394 | mutex_enter(&db->db_mtx); | |
2395 | ||
2396 | ASSERT(db->db_level > 0); | |
2397 | DBUF_VERIFY(db); | |
2398 | ||
e49f1e20 | 2399 | /* Read the block if it hasn't been read yet. */ |
34dc7c2f BB |
2400 | if (db->db_buf == NULL) { |
2401 | mutex_exit(&db->db_mtx); | |
2402 | (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED); | |
2403 | mutex_enter(&db->db_mtx); | |
2404 | } | |
2405 | ASSERT3U(db->db_state, ==, DB_CACHED); | |
34dc7c2f BB |
2406 | ASSERT(db->db_buf != NULL); |
2407 | ||
572e2857 BB |
2408 | DB_DNODE_ENTER(db); |
2409 | dn = DB_DNODE(db); | |
e49f1e20 | 2410 | /* Indirect block size must match what the dnode thinks it is. */ |
572e2857 | 2411 | ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); |
34dc7c2f | 2412 | dbuf_check_blkptr(dn, db); |
572e2857 | 2413 | DB_DNODE_EXIT(db); |
34dc7c2f | 2414 | |
e49f1e20 | 2415 | /* Provide the pending dirty record to child dbufs */ |
34dc7c2f BB |
2416 | db->db_data_pending = dr; |
2417 | ||
34dc7c2f | 2418 | mutex_exit(&db->db_mtx); |
b128c09f | 2419 | dbuf_write(dr, db->db_buf, tx); |
34dc7c2f BB |
2420 | |
2421 | zio = dr->dr_zio; | |
2422 | mutex_enter(&dr->dt.di.dr_mtx); | |
2423 | dbuf_sync_list(&dr->dt.di.dr_children, tx); | |
2424 | ASSERT(list_head(&dr->dt.di.dr_children) == NULL); | |
2425 | mutex_exit(&dr->dt.di.dr_mtx); | |
2426 | zio_nowait(zio); | |
2427 | } | |
2428 | ||
d1d7e268 MK |
2429 | /* |
2430 | * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is | |
60948de1 BB |
2431 | * critical the we not allow the compiler to inline this function in to |
2432 | * dbuf_sync_list() thereby drastically bloating the stack usage. | |
2433 | */ | |
2434 | noinline static void | |
34dc7c2f BB |
2435 | dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx) |
2436 | { | |
2437 | arc_buf_t **datap = &dr->dt.dl.dr_data; | |
2438 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
572e2857 BB |
2439 | dnode_t *dn; |
2440 | objset_t *os; | |
34dc7c2f | 2441 | uint64_t txg = tx->tx_txg; |
34dc7c2f BB |
2442 | |
2443 | ASSERT(dmu_tx_is_syncing(tx)); | |
2444 | ||
2445 | dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); | |
2446 | ||
2447 | mutex_enter(&db->db_mtx); | |
2448 | /* | |
2449 | * To be synced, we must be dirtied. But we | |
2450 | * might have been freed after the dirty. | |
2451 | */ | |
2452 | if (db->db_state == DB_UNCACHED) { | |
2453 | /* This buffer has been freed since it was dirtied */ | |
2454 | ASSERT(db->db.db_data == NULL); | |
2455 | } else if (db->db_state == DB_FILL) { | |
2456 | /* This buffer was freed and is now being re-filled */ | |
2457 | ASSERT(db->db.db_data != dr->dt.dl.dr_data); | |
2458 | } else { | |
b128c09f | 2459 | ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL); |
34dc7c2f BB |
2460 | } |
2461 | DBUF_VERIFY(db); | |
2462 | ||
572e2857 BB |
2463 | DB_DNODE_ENTER(db); |
2464 | dn = DB_DNODE(db); | |
2465 | ||
428870ff BB |
2466 | if (db->db_blkid == DMU_SPILL_BLKID) { |
2467 | mutex_enter(&dn->dn_mtx); | |
2468 | dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR; | |
2469 | mutex_exit(&dn->dn_mtx); | |
2470 | } | |
2471 | ||
34dc7c2f BB |
2472 | /* |
2473 | * If this is a bonus buffer, simply copy the bonus data into the | |
2474 | * dnode. It will be written out when the dnode is synced (and it | |
2475 | * will be synced, since it must have been dirty for dbuf_sync to | |
2476 | * be called). | |
2477 | */ | |
428870ff | 2478 | if (db->db_blkid == DMU_BONUS_BLKID) { |
34dc7c2f BB |
2479 | dbuf_dirty_record_t **drp; |
2480 | ||
2481 | ASSERT(*datap != NULL); | |
c99c9001 | 2482 | ASSERT0(db->db_level); |
34dc7c2f BB |
2483 | ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN); |
2484 | bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen); | |
572e2857 BB |
2485 | DB_DNODE_EXIT(db); |
2486 | ||
34dc7c2f BB |
2487 | if (*datap != db->db.db_data) { |
2488 | zio_buf_free(*datap, DN_MAX_BONUSLEN); | |
d164b209 | 2489 | arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); |
34dc7c2f BB |
2490 | } |
2491 | db->db_data_pending = NULL; | |
2492 | drp = &db->db_last_dirty; | |
2493 | while (*drp != dr) | |
2494 | drp = &(*drp)->dr_next; | |
2495 | ASSERT(dr->dr_next == NULL); | |
428870ff | 2496 | ASSERT(dr->dr_dbuf == db); |
34dc7c2f | 2497 | *drp = dr->dr_next; |
753972fc BB |
2498 | if (dr->dr_dbuf->db_level != 0) { |
2499 | mutex_destroy(&dr->dt.di.dr_mtx); | |
2500 | list_destroy(&dr->dt.di.dr_children); | |
2501 | } | |
34dc7c2f BB |
2502 | kmem_free(dr, sizeof (dbuf_dirty_record_t)); |
2503 | ASSERT(db->db_dirtycnt > 0); | |
2504 | db->db_dirtycnt -= 1; | |
428870ff | 2505 | dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg); |
34dc7c2f BB |
2506 | return; |
2507 | } | |
2508 | ||
572e2857 BB |
2509 | os = dn->dn_objset; |
2510 | ||
34dc7c2f BB |
2511 | /* |
2512 | * This function may have dropped the db_mtx lock allowing a dmu_sync | |
2513 | * operation to sneak in. As a result, we need to ensure that we | |
2514 | * don't check the dr_override_state until we have returned from | |
2515 | * dbuf_check_blkptr. | |
2516 | */ | |
2517 | dbuf_check_blkptr(dn, db); | |
2518 | ||
2519 | /* | |
572e2857 | 2520 | * If this buffer is in the middle of an immediate write, |
34dc7c2f BB |
2521 | * wait for the synchronous IO to complete. |
2522 | */ | |
2523 | while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) { | |
2524 | ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT); | |
2525 | cv_wait(&db->db_changed, &db->db_mtx); | |
2526 | ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN); | |
2527 | } | |
2528 | ||
9babb374 BB |
2529 | if (db->db_state != DB_NOFILL && |
2530 | dn->dn_object != DMU_META_DNODE_OBJECT && | |
2531 | refcount_count(&db->db_holds) > 1 && | |
428870ff | 2532 | dr->dt.dl.dr_override_state != DR_OVERRIDDEN && |
9babb374 BB |
2533 | *datap == db->db_buf) { |
2534 | /* | |
2535 | * If this buffer is currently "in use" (i.e., there | |
2536 | * are active holds and db_data still references it), | |
2537 | * then make a copy before we start the write so that | |
2538 | * any modifications from the open txg will not leak | |
2539 | * into this write. | |
2540 | * | |
2541 | * NOTE: this copy does not need to be made for | |
2542 | * objects only modified in the syncing context (e.g. | |
2543 | * DNONE_DNODE blocks). | |
2544 | */ | |
2545 | int blksz = arc_buf_size(*datap); | |
2546 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); | |
2547 | *datap = arc_buf_alloc(os->os_spa, blksz, db, type); | |
2548 | bcopy(db->db.db_data, (*datap)->b_data, blksz); | |
b128c09f | 2549 | } |
34dc7c2f BB |
2550 | db->db_data_pending = dr; |
2551 | ||
2552 | mutex_exit(&db->db_mtx); | |
2553 | ||
b128c09f | 2554 | dbuf_write(dr, *datap, tx); |
34dc7c2f BB |
2555 | |
2556 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
572e2857 | 2557 | if (dn->dn_object == DMU_META_DNODE_OBJECT) { |
34dc7c2f | 2558 | list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr); |
572e2857 BB |
2559 | DB_DNODE_EXIT(db); |
2560 | } else { | |
2561 | /* | |
2562 | * Although zio_nowait() does not "wait for an IO", it does | |
2563 | * initiate the IO. If this is an empty write it seems plausible | |
2564 | * that the IO could actually be completed before the nowait | |
2565 | * returns. We need to DB_DNODE_EXIT() first in case | |
2566 | * zio_nowait() invalidates the dbuf. | |
2567 | */ | |
2568 | DB_DNODE_EXIT(db); | |
34dc7c2f | 2569 | zio_nowait(dr->dr_zio); |
572e2857 | 2570 | } |
34dc7c2f BB |
2571 | } |
2572 | ||
2573 | void | |
2574 | dbuf_sync_list(list_t *list, dmu_tx_t *tx) | |
2575 | { | |
2576 | dbuf_dirty_record_t *dr; | |
2577 | ||
c65aa5b2 | 2578 | while ((dr = list_head(list))) { |
34dc7c2f BB |
2579 | if (dr->dr_zio != NULL) { |
2580 | /* | |
2581 | * If we find an already initialized zio then we | |
2582 | * are processing the meta-dnode, and we have finished. | |
2583 | * The dbufs for all dnodes are put back on the list | |
2584 | * during processing, so that we can zio_wait() | |
2585 | * these IOs after initiating all child IOs. | |
2586 | */ | |
2587 | ASSERT3U(dr->dr_dbuf->db.db_object, ==, | |
2588 | DMU_META_DNODE_OBJECT); | |
2589 | break; | |
2590 | } | |
2591 | list_remove(list, dr); | |
2592 | if (dr->dr_dbuf->db_level > 0) | |
2593 | dbuf_sync_indirect(dr, tx); | |
2594 | else | |
2595 | dbuf_sync_leaf(dr, tx); | |
2596 | } | |
2597 | } | |
2598 | ||
34dc7c2f BB |
2599 | /* ARGSUSED */ |
2600 | static void | |
2601 | dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb) | |
2602 | { | |
2603 | dmu_buf_impl_t *db = vdb; | |
572e2857 | 2604 | dnode_t *dn; |
b128c09f | 2605 | blkptr_t *bp = zio->io_bp; |
34dc7c2f | 2606 | blkptr_t *bp_orig = &zio->io_bp_orig; |
428870ff BB |
2607 | spa_t *spa = zio->io_spa; |
2608 | int64_t delta; | |
34dc7c2f | 2609 | uint64_t fill = 0; |
428870ff | 2610 | int i; |
34dc7c2f | 2611 | |
9b67f605 | 2612 | ASSERT3P(db->db_blkptr, ==, bp); |
b128c09f | 2613 | |
572e2857 BB |
2614 | DB_DNODE_ENTER(db); |
2615 | dn = DB_DNODE(db); | |
428870ff BB |
2616 | delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig); |
2617 | dnode_diduse_space(dn, delta - zio->io_prev_space_delta); | |
2618 | zio->io_prev_space_delta = delta; | |
34dc7c2f | 2619 | |
b0bc7a84 MG |
2620 | if (bp->blk_birth != 0) { |
2621 | ASSERT((db->db_blkid != DMU_SPILL_BLKID && | |
2622 | BP_GET_TYPE(bp) == dn->dn_type) || | |
2623 | (db->db_blkid == DMU_SPILL_BLKID && | |
9b67f605 MA |
2624 | BP_GET_TYPE(bp) == dn->dn_bonustype) || |
2625 | BP_IS_EMBEDDED(bp)); | |
b0bc7a84 | 2626 | ASSERT(BP_GET_LEVEL(bp) == db->db_level); |
34dc7c2f BB |
2627 | } |
2628 | ||
2629 | mutex_enter(&db->db_mtx); | |
2630 | ||
428870ff BB |
2631 | #ifdef ZFS_DEBUG |
2632 | if (db->db_blkid == DMU_SPILL_BLKID) { | |
428870ff BB |
2633 | ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); |
2634 | ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && | |
2635 | db->db_blkptr == &dn->dn_phys->dn_spill); | |
2636 | } | |
2637 | #endif | |
2638 | ||
34dc7c2f BB |
2639 | if (db->db_level == 0) { |
2640 | mutex_enter(&dn->dn_mtx); | |
428870ff BB |
2641 | if (db->db_blkid > dn->dn_phys->dn_maxblkid && |
2642 | db->db_blkid != DMU_SPILL_BLKID) | |
34dc7c2f BB |
2643 | dn->dn_phys->dn_maxblkid = db->db_blkid; |
2644 | mutex_exit(&dn->dn_mtx); | |
2645 | ||
2646 | if (dn->dn_type == DMU_OT_DNODE) { | |
2647 | dnode_phys_t *dnp = db->db.db_data; | |
2648 | for (i = db->db.db_size >> DNODE_SHIFT; i > 0; | |
2649 | i--, dnp++) { | |
2650 | if (dnp->dn_type != DMU_OT_NONE) | |
2651 | fill++; | |
2652 | } | |
2653 | } else { | |
b0bc7a84 MG |
2654 | if (BP_IS_HOLE(bp)) { |
2655 | fill = 0; | |
2656 | } else { | |
2657 | fill = 1; | |
2658 | } | |
34dc7c2f BB |
2659 | } |
2660 | } else { | |
b128c09f | 2661 | blkptr_t *ibp = db->db.db_data; |
34dc7c2f | 2662 | ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); |
b128c09f BB |
2663 | for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) { |
2664 | if (BP_IS_HOLE(ibp)) | |
34dc7c2f | 2665 | continue; |
9b67f605 | 2666 | fill += BP_GET_FILL(ibp); |
34dc7c2f BB |
2667 | } |
2668 | } | |
572e2857 | 2669 | DB_DNODE_EXIT(db); |
34dc7c2f | 2670 | |
9b67f605 MA |
2671 | if (!BP_IS_EMBEDDED(bp)) |
2672 | bp->blk_fill = fill; | |
34dc7c2f BB |
2673 | |
2674 | mutex_exit(&db->db_mtx); | |
34dc7c2f BB |
2675 | } |
2676 | ||
e8b96c60 MA |
2677 | /* |
2678 | * The SPA will call this callback several times for each zio - once | |
2679 | * for every physical child i/o (zio->io_phys_children times). This | |
2680 | * allows the DMU to monitor the progress of each logical i/o. For example, | |
2681 | * there may be 2 copies of an indirect block, or many fragments of a RAID-Z | |
2682 | * block. There may be a long delay before all copies/fragments are completed, | |
2683 | * so this callback allows us to retire dirty space gradually, as the physical | |
2684 | * i/os complete. | |
2685 | */ | |
2686 | /* ARGSUSED */ | |
2687 | static void | |
2688 | dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg) | |
2689 | { | |
2690 | dmu_buf_impl_t *db = arg; | |
2691 | objset_t *os = db->db_objset; | |
2692 | dsl_pool_t *dp = dmu_objset_pool(os); | |
2693 | dbuf_dirty_record_t *dr; | |
2694 | int delta = 0; | |
2695 | ||
2696 | dr = db->db_data_pending; | |
2697 | ASSERT3U(dr->dr_txg, ==, zio->io_txg); | |
2698 | ||
2699 | /* | |
2700 | * The callback will be called io_phys_children times. Retire one | |
2701 | * portion of our dirty space each time we are called. Any rounding | |
2702 | * error will be cleaned up by dsl_pool_sync()'s call to | |
2703 | * dsl_pool_undirty_space(). | |
2704 | */ | |
2705 | delta = dr->dr_accounted / zio->io_phys_children; | |
2706 | dsl_pool_undirty_space(dp, delta, zio->io_txg); | |
2707 | } | |
2708 | ||
34dc7c2f BB |
2709 | /* ARGSUSED */ |
2710 | static void | |
2711 | dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb) | |
2712 | { | |
2713 | dmu_buf_impl_t *db = vdb; | |
428870ff | 2714 | blkptr_t *bp_orig = &zio->io_bp_orig; |
b0bc7a84 MG |
2715 | blkptr_t *bp = db->db_blkptr; |
2716 | objset_t *os = db->db_objset; | |
2717 | dmu_tx_t *tx = os->os_synctx; | |
34dc7c2f BB |
2718 | dbuf_dirty_record_t **drp, *dr; |
2719 | ||
c99c9001 | 2720 | ASSERT0(zio->io_error); |
428870ff BB |
2721 | ASSERT(db->db_blkptr == bp); |
2722 | ||
03c6040b GW |
2723 | /* |
2724 | * For nopwrites and rewrites we ensure that the bp matches our | |
2725 | * original and bypass all the accounting. | |
2726 | */ | |
2727 | if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) { | |
428870ff BB |
2728 | ASSERT(BP_EQUAL(bp, bp_orig)); |
2729 | } else { | |
b0bc7a84 | 2730 | dsl_dataset_t *ds = os->os_dsl_dataset; |
428870ff BB |
2731 | (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE); |
2732 | dsl_dataset_block_born(ds, bp, tx); | |
2733 | } | |
34dc7c2f BB |
2734 | |
2735 | mutex_enter(&db->db_mtx); | |
2736 | ||
428870ff BB |
2737 | DBUF_VERIFY(db); |
2738 | ||
34dc7c2f BB |
2739 | drp = &db->db_last_dirty; |
2740 | while ((dr = *drp) != db->db_data_pending) | |
2741 | drp = &dr->dr_next; | |
2742 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
428870ff | 2743 | ASSERT(dr->dr_dbuf == db); |
34dc7c2f BB |
2744 | ASSERT(dr->dr_next == NULL); |
2745 | *drp = dr->dr_next; | |
2746 | ||
428870ff BB |
2747 | #ifdef ZFS_DEBUG |
2748 | if (db->db_blkid == DMU_SPILL_BLKID) { | |
572e2857 BB |
2749 | dnode_t *dn; |
2750 | ||
2751 | DB_DNODE_ENTER(db); | |
2752 | dn = DB_DNODE(db); | |
428870ff BB |
2753 | ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); |
2754 | ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && | |
2755 | db->db_blkptr == &dn->dn_phys->dn_spill); | |
572e2857 | 2756 | DB_DNODE_EXIT(db); |
428870ff BB |
2757 | } |
2758 | #endif | |
2759 | ||
34dc7c2f | 2760 | if (db->db_level == 0) { |
428870ff | 2761 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f | 2762 | ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); |
b128c09f BB |
2763 | if (db->db_state != DB_NOFILL) { |
2764 | if (dr->dt.dl.dr_data != db->db_buf) | |
2765 | VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, | |
13fe0198 | 2766 | db)); |
428870ff | 2767 | else if (!arc_released(db->db_buf)) |
b128c09f | 2768 | arc_set_callback(db->db_buf, dbuf_do_evict, db); |
b128c09f | 2769 | } |
34dc7c2f | 2770 | } else { |
572e2857 BB |
2771 | dnode_t *dn; |
2772 | ||
2773 | DB_DNODE_ENTER(db); | |
2774 | dn = DB_DNODE(db); | |
34dc7c2f | 2775 | ASSERT(list_head(&dr->dt.di.dr_children) == NULL); |
b0bc7a84 | 2776 | ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift); |
34dc7c2f | 2777 | if (!BP_IS_HOLE(db->db_blkptr)) { |
1fde1e37 BB |
2778 | ASSERTV(int epbs = dn->dn_phys->dn_indblkshift - |
2779 | SPA_BLKPTRSHIFT); | |
b0bc7a84 MG |
2780 | ASSERT3U(db->db_blkid, <=, |
2781 | dn->dn_phys->dn_maxblkid >> (db->db_level * epbs)); | |
34dc7c2f BB |
2782 | ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, |
2783 | db->db.db_size); | |
9b67f605 MA |
2784 | if (!arc_released(db->db_buf)) |
2785 | arc_set_callback(db->db_buf, dbuf_do_evict, db); | |
34dc7c2f | 2786 | } |
572e2857 | 2787 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
2788 | mutex_destroy(&dr->dt.di.dr_mtx); |
2789 | list_destroy(&dr->dt.di.dr_children); | |
2790 | } | |
2791 | kmem_free(dr, sizeof (dbuf_dirty_record_t)); | |
2792 | ||
2793 | cv_broadcast(&db->db_changed); | |
2794 | ASSERT(db->db_dirtycnt > 0); | |
2795 | db->db_dirtycnt -= 1; | |
2796 | db->db_data_pending = NULL; | |
b0bc7a84 | 2797 | dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg); |
428870ff BB |
2798 | } |
2799 | ||
2800 | static void | |
2801 | dbuf_write_nofill_ready(zio_t *zio) | |
2802 | { | |
2803 | dbuf_write_ready(zio, NULL, zio->io_private); | |
2804 | } | |
2805 | ||
2806 | static void | |
2807 | dbuf_write_nofill_done(zio_t *zio) | |
2808 | { | |
2809 | dbuf_write_done(zio, NULL, zio->io_private); | |
2810 | } | |
2811 | ||
2812 | static void | |
2813 | dbuf_write_override_ready(zio_t *zio) | |
2814 | { | |
2815 | dbuf_dirty_record_t *dr = zio->io_private; | |
2816 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
2817 | ||
2818 | dbuf_write_ready(zio, NULL, db); | |
2819 | } | |
2820 | ||
2821 | static void | |
2822 | dbuf_write_override_done(zio_t *zio) | |
2823 | { | |
2824 | dbuf_dirty_record_t *dr = zio->io_private; | |
2825 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
2826 | blkptr_t *obp = &dr->dt.dl.dr_overridden_by; | |
2827 | ||
2828 | mutex_enter(&db->db_mtx); | |
2829 | if (!BP_EQUAL(zio->io_bp, obp)) { | |
2830 | if (!BP_IS_HOLE(obp)) | |
2831 | dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp); | |
2832 | arc_release(dr->dt.dl.dr_data, db); | |
2833 | } | |
34dc7c2f BB |
2834 | mutex_exit(&db->db_mtx); |
2835 | ||
428870ff BB |
2836 | dbuf_write_done(zio, NULL, db); |
2837 | } | |
2838 | ||
e49f1e20 | 2839 | /* Issue I/O to commit a dirty buffer to disk. */ |
428870ff BB |
2840 | static void |
2841 | dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx) | |
2842 | { | |
2843 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
572e2857 BB |
2844 | dnode_t *dn; |
2845 | objset_t *os; | |
428870ff BB |
2846 | dmu_buf_impl_t *parent = db->db_parent; |
2847 | uint64_t txg = tx->tx_txg; | |
2848 | zbookmark_t zb; | |
2849 | zio_prop_t zp; | |
2850 | zio_t *zio; | |
2851 | int wp_flag = 0; | |
34dc7c2f | 2852 | |
572e2857 BB |
2853 | DB_DNODE_ENTER(db); |
2854 | dn = DB_DNODE(db); | |
2855 | os = dn->dn_objset; | |
2856 | ||
428870ff BB |
2857 | if (db->db_state != DB_NOFILL) { |
2858 | if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) { | |
2859 | /* | |
2860 | * Private object buffers are released here rather | |
2861 | * than in dbuf_dirty() since they are only modified | |
2862 | * in the syncing context and we don't want the | |
2863 | * overhead of making multiple copies of the data. | |
2864 | */ | |
2865 | if (BP_IS_HOLE(db->db_blkptr)) { | |
2866 | arc_buf_thaw(data); | |
2867 | } else { | |
2868 | dbuf_release_bp(db); | |
2869 | } | |
2870 | } | |
2871 | } | |
2872 | ||
2873 | if (parent != dn->dn_dbuf) { | |
e49f1e20 WA |
2874 | /* Our parent is an indirect block. */ |
2875 | /* We have a dirty parent that has been scheduled for write. */ | |
428870ff | 2876 | ASSERT(parent && parent->db_data_pending); |
e49f1e20 | 2877 | /* Our parent's buffer is one level closer to the dnode. */ |
428870ff | 2878 | ASSERT(db->db_level == parent->db_level-1); |
e49f1e20 WA |
2879 | /* |
2880 | * We're about to modify our parent's db_data by modifying | |
2881 | * our block pointer, so the parent must be released. | |
2882 | */ | |
428870ff BB |
2883 | ASSERT(arc_released(parent->db_buf)); |
2884 | zio = parent->db_data_pending->dr_zio; | |
2885 | } else { | |
e49f1e20 | 2886 | /* Our parent is the dnode itself. */ |
428870ff BB |
2887 | ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 && |
2888 | db->db_blkid != DMU_SPILL_BLKID) || | |
2889 | (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0)); | |
2890 | if (db->db_blkid != DMU_SPILL_BLKID) | |
2891 | ASSERT3P(db->db_blkptr, ==, | |
2892 | &dn->dn_phys->dn_blkptr[db->db_blkid]); | |
2893 | zio = dn->dn_zio; | |
2894 | } | |
2895 | ||
2896 | ASSERT(db->db_level == 0 || data == db->db_buf); | |
2897 | ASSERT3U(db->db_blkptr->blk_birth, <=, txg); | |
2898 | ASSERT(zio); | |
2899 | ||
2900 | SET_BOOKMARK(&zb, os->os_dsl_dataset ? | |
2901 | os->os_dsl_dataset->ds_object : DMU_META_OBJSET, | |
2902 | db->db.db_object, db->db_level, db->db_blkid); | |
2903 | ||
2904 | if (db->db_blkid == DMU_SPILL_BLKID) | |
2905 | wp_flag = WP_SPILL; | |
2906 | wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0; | |
2907 | ||
2908 | dmu_write_policy(os, dn, db->db_level, wp_flag, &zp); | |
572e2857 | 2909 | DB_DNODE_EXIT(db); |
428870ff | 2910 | |
9b67f605 MA |
2911 | if (db->db_level == 0 && |
2912 | dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { | |
2913 | /* | |
2914 | * The BP for this block has been provided by open context | |
2915 | * (by dmu_sync() or dmu_buf_write_embedded()). | |
2916 | */ | |
2917 | void *contents = (data != NULL) ? data->b_data : NULL; | |
2918 | ||
428870ff | 2919 | dr->dr_zio = zio_write(zio, os->os_spa, txg, |
9b67f605 | 2920 | db->db_blkptr, contents, db->db.db_size, &zp, |
e8b96c60 MA |
2921 | dbuf_write_override_ready, NULL, dbuf_write_override_done, |
2922 | dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); | |
428870ff BB |
2923 | mutex_enter(&db->db_mtx); |
2924 | dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; | |
2925 | zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by, | |
03c6040b | 2926 | dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite); |
428870ff BB |
2927 | mutex_exit(&db->db_mtx); |
2928 | } else if (db->db_state == DB_NOFILL) { | |
2929 | ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF); | |
2930 | dr->dr_zio = zio_write(zio, os->os_spa, txg, | |
2931 | db->db_blkptr, NULL, db->db.db_size, &zp, | |
e8b96c60 | 2932 | dbuf_write_nofill_ready, NULL, dbuf_write_nofill_done, db, |
428870ff BB |
2933 | ZIO_PRIORITY_ASYNC_WRITE, |
2934 | ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb); | |
2935 | } else { | |
2936 | ASSERT(arc_released(data)); | |
2937 | dr->dr_zio = arc_write(zio, os->os_spa, txg, | |
3a17a7a9 SK |
2938 | db->db_blkptr, data, DBUF_IS_L2CACHEABLE(db), |
2939 | DBUF_IS_L2COMPRESSIBLE(db), &zp, dbuf_write_ready, | |
e8b96c60 MA |
2940 | dbuf_write_physdone, dbuf_write_done, db, |
2941 | ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); | |
428870ff | 2942 | } |
34dc7c2f | 2943 | } |
c28b2279 BB |
2944 | |
2945 | #if defined(_KERNEL) && defined(HAVE_SPL) | |
8f576c23 BB |
2946 | EXPORT_SYMBOL(dbuf_find); |
2947 | EXPORT_SYMBOL(dbuf_is_metadata); | |
2948 | EXPORT_SYMBOL(dbuf_evict); | |
2949 | EXPORT_SYMBOL(dbuf_loan_arcbuf); | |
2950 | EXPORT_SYMBOL(dbuf_whichblock); | |
2951 | EXPORT_SYMBOL(dbuf_read); | |
2952 | EXPORT_SYMBOL(dbuf_unoverride); | |
2953 | EXPORT_SYMBOL(dbuf_free_range); | |
2954 | EXPORT_SYMBOL(dbuf_new_size); | |
2955 | EXPORT_SYMBOL(dbuf_release_bp); | |
2956 | EXPORT_SYMBOL(dbuf_dirty); | |
c28b2279 | 2957 | EXPORT_SYMBOL(dmu_buf_will_dirty); |
8f576c23 BB |
2958 | EXPORT_SYMBOL(dmu_buf_will_not_fill); |
2959 | EXPORT_SYMBOL(dmu_buf_will_fill); | |
2960 | EXPORT_SYMBOL(dmu_buf_fill_done); | |
4047414a | 2961 | EXPORT_SYMBOL(dmu_buf_rele); |
8f576c23 BB |
2962 | EXPORT_SYMBOL(dbuf_assign_arcbuf); |
2963 | EXPORT_SYMBOL(dbuf_clear); | |
2964 | EXPORT_SYMBOL(dbuf_prefetch); | |
2965 | EXPORT_SYMBOL(dbuf_hold_impl); | |
2966 | EXPORT_SYMBOL(dbuf_hold); | |
2967 | EXPORT_SYMBOL(dbuf_hold_level); | |
2968 | EXPORT_SYMBOL(dbuf_create_bonus); | |
2969 | EXPORT_SYMBOL(dbuf_spill_set_blksz); | |
2970 | EXPORT_SYMBOL(dbuf_rm_spill); | |
2971 | EXPORT_SYMBOL(dbuf_add_ref); | |
2972 | EXPORT_SYMBOL(dbuf_rele); | |
2973 | EXPORT_SYMBOL(dbuf_rele_and_unlock); | |
2974 | EXPORT_SYMBOL(dbuf_refcount); | |
2975 | EXPORT_SYMBOL(dbuf_sync_list); | |
2976 | EXPORT_SYMBOL(dmu_buf_set_user); | |
2977 | EXPORT_SYMBOL(dmu_buf_set_user_ie); | |
2978 | EXPORT_SYMBOL(dmu_buf_update_user); | |
2979 | EXPORT_SYMBOL(dmu_buf_get_user); | |
2980 | EXPORT_SYMBOL(dmu_buf_freeable); | |
c28b2279 | 2981 | #endif |