]>
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> |
49ee64e5 | 46 | #include <sys/trace_dbuf.h> |
34dc7c2f | 47 | |
fc5bb51f BB |
48 | struct dbuf_hold_impl_data { |
49 | /* Function arguments */ | |
50 | dnode_t *dh_dn; | |
51 | uint8_t dh_level; | |
52 | uint64_t dh_blkid; | |
53 | int dh_fail_sparse; | |
54 | void *dh_tag; | |
55 | dmu_buf_impl_t **dh_dbp; | |
56 | /* Local variables */ | |
57 | dmu_buf_impl_t *dh_db; | |
58 | dmu_buf_impl_t *dh_parent; | |
59 | blkptr_t *dh_bp; | |
60 | int dh_err; | |
61 | dbuf_dirty_record_t *dh_dr; | |
62 | arc_buf_contents_t dh_type; | |
63 | int dh_depth; | |
64 | }; | |
65 | ||
66 | static void __dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh, | |
67 | dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse, | |
68 | void *tag, dmu_buf_impl_t **dbp, int depth); | |
69 | static int __dbuf_hold_impl(struct dbuf_hold_impl_data *dh); | |
70 | ||
b663a23d MA |
71 | /* |
72 | * Number of times that zfs_free_range() took the slow path while doing | |
73 | * a zfs receive. A nonzero value indicates a potential performance problem. | |
74 | */ | |
75 | uint64_t zfs_free_range_recv_miss; | |
76 | ||
34dc7c2f | 77 | static void dbuf_destroy(dmu_buf_impl_t *db); |
13fe0198 | 78 | static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx); |
b128c09f | 79 | static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx); |
34dc7c2f BB |
80 | |
81 | /* | |
82 | * Global data structures and functions for the dbuf cache. | |
83 | */ | |
84 | static kmem_cache_t *dbuf_cache; | |
85 | ||
86 | /* ARGSUSED */ | |
87 | static int | |
88 | dbuf_cons(void *vdb, void *unused, int kmflag) | |
89 | { | |
90 | dmu_buf_impl_t *db = vdb; | |
91 | bzero(db, sizeof (dmu_buf_impl_t)); | |
92 | ||
93 | mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL); | |
94 | cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL); | |
95 | refcount_create(&db->db_holds); | |
8951cb8d | 96 | |
34dc7c2f BB |
97 | return (0); |
98 | } | |
99 | ||
100 | /* ARGSUSED */ | |
101 | static void | |
102 | dbuf_dest(void *vdb, void *unused) | |
103 | { | |
104 | dmu_buf_impl_t *db = vdb; | |
105 | mutex_destroy(&db->db_mtx); | |
106 | cv_destroy(&db->db_changed); | |
107 | refcount_destroy(&db->db_holds); | |
108 | } | |
109 | ||
110 | /* | |
111 | * dbuf hash table routines | |
112 | */ | |
113 | static dbuf_hash_table_t dbuf_hash_table; | |
114 | ||
115 | static uint64_t dbuf_hash_count; | |
116 | ||
117 | static uint64_t | |
118 | dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid) | |
119 | { | |
120 | uintptr_t osv = (uintptr_t)os; | |
121 | uint64_t crc = -1ULL; | |
122 | ||
123 | ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); | |
124 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF]; | |
125 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF]; | |
126 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF]; | |
127 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF]; | |
128 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF]; | |
129 | crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF]; | |
130 | ||
131 | crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16); | |
132 | ||
133 | return (crc); | |
134 | } | |
135 | ||
136 | #define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid); | |
137 | ||
138 | #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \ | |
139 | ((dbuf)->db.db_object == (obj) && \ | |
140 | (dbuf)->db_objset == (os) && \ | |
141 | (dbuf)->db_level == (level) && \ | |
142 | (dbuf)->db_blkid == (blkid)) | |
143 | ||
144 | dmu_buf_impl_t * | |
145 | dbuf_find(dnode_t *dn, uint8_t level, uint64_t blkid) | |
146 | { | |
147 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
428870ff | 148 | objset_t *os = dn->dn_objset; |
d6320ddb BB |
149 | uint64_t obj; |
150 | uint64_t hv; | |
151 | uint64_t idx; | |
34dc7c2f BB |
152 | dmu_buf_impl_t *db; |
153 | ||
d6320ddb BB |
154 | obj = dn->dn_object; |
155 | hv = DBUF_HASH(os, obj, level, blkid); | |
156 | idx = hv & h->hash_table_mask; | |
157 | ||
34dc7c2f BB |
158 | mutex_enter(DBUF_HASH_MUTEX(h, idx)); |
159 | for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) { | |
160 | if (DBUF_EQUAL(db, os, obj, level, blkid)) { | |
161 | mutex_enter(&db->db_mtx); | |
162 | if (db->db_state != DB_EVICTING) { | |
163 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); | |
164 | return (db); | |
165 | } | |
166 | mutex_exit(&db->db_mtx); | |
167 | } | |
168 | } | |
169 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); | |
170 | return (NULL); | |
171 | } | |
172 | ||
173 | /* | |
174 | * Insert an entry into the hash table. If there is already an element | |
175 | * equal to elem in the hash table, then the already existing element | |
176 | * will be returned and the new element will not be inserted. | |
177 | * Otherwise returns NULL. | |
178 | */ | |
179 | static dmu_buf_impl_t * | |
180 | dbuf_hash_insert(dmu_buf_impl_t *db) | |
181 | { | |
182 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
428870ff | 183 | objset_t *os = db->db_objset; |
34dc7c2f BB |
184 | uint64_t obj = db->db.db_object; |
185 | int level = db->db_level; | |
d6320ddb | 186 | uint64_t blkid, hv, idx; |
34dc7c2f BB |
187 | dmu_buf_impl_t *dbf; |
188 | ||
d6320ddb BB |
189 | blkid = db->db_blkid; |
190 | hv = DBUF_HASH(os, obj, level, blkid); | |
191 | idx = hv & h->hash_table_mask; | |
192 | ||
34dc7c2f BB |
193 | mutex_enter(DBUF_HASH_MUTEX(h, idx)); |
194 | for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) { | |
195 | if (DBUF_EQUAL(dbf, os, obj, level, blkid)) { | |
196 | mutex_enter(&dbf->db_mtx); | |
197 | if (dbf->db_state != DB_EVICTING) { | |
198 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); | |
199 | return (dbf); | |
200 | } | |
201 | mutex_exit(&dbf->db_mtx); | |
202 | } | |
203 | } | |
204 | ||
205 | mutex_enter(&db->db_mtx); | |
206 | db->db_hash_next = h->hash_table[idx]; | |
207 | h->hash_table[idx] = db; | |
208 | mutex_exit(DBUF_HASH_MUTEX(h, idx)); | |
209 | atomic_add_64(&dbuf_hash_count, 1); | |
210 | ||
211 | return (NULL); | |
212 | } | |
213 | ||
214 | /* | |
bd089c54 | 215 | * Remove an entry from the hash table. It must be in the EVICTING state. |
34dc7c2f BB |
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 | 228 | /* |
bd089c54 | 229 | * We musn't hold db_mtx to maintain lock ordering: |
34dc7c2f BB |
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 | ||
34dc7c2f BB |
258 | db->db_evict_func(&db->db, db->db_user_ptr); |
259 | db->db_user_ptr = NULL; | |
34dc7c2f BB |
260 | db->db_evict_func = NULL; |
261 | } | |
262 | ||
572e2857 BB |
263 | boolean_t |
264 | dbuf_is_metadata(dmu_buf_impl_t *db) | |
265 | { | |
cc79a5c2 BB |
266 | /* |
267 | * Consider indirect blocks and spill blocks to be meta data. | |
268 | */ | |
269 | if (db->db_level > 0 || db->db_blkid == DMU_SPILL_BLKID) { | |
572e2857 BB |
270 | return (B_TRUE); |
271 | } else { | |
272 | boolean_t is_metadata; | |
273 | ||
274 | DB_DNODE_ENTER(db); | |
9ae529ec | 275 | is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type); |
572e2857 BB |
276 | DB_DNODE_EXIT(db); |
277 | ||
278 | return (is_metadata); | |
279 | } | |
280 | } | |
281 | ||
34dc7c2f BB |
282 | void |
283 | dbuf_evict(dmu_buf_impl_t *db) | |
284 | { | |
285 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
286 | ASSERT(db->db_buf == NULL); | |
287 | ASSERT(db->db_data_pending == NULL); | |
288 | ||
289 | dbuf_clear(db); | |
290 | dbuf_destroy(db); | |
291 | } | |
292 | ||
293 | void | |
294 | dbuf_init(void) | |
295 | { | |
296 | uint64_t hsize = 1ULL << 16; | |
297 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
298 | int i; | |
299 | ||
300 | /* | |
301 | * The hash table is big enough to fill all of physical memory | |
302 | * with an average 4K block size. The table will take up | |
303 | * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers). | |
304 | */ | |
305 | while (hsize * 4096 < physmem * PAGESIZE) | |
306 | hsize <<= 1; | |
307 | ||
308 | retry: | |
309 | h->hash_table_mask = hsize - 1; | |
00b46022 | 310 | #if defined(_KERNEL) && defined(HAVE_SPL) |
d1d7e268 MK |
311 | /* |
312 | * Large allocations which do not require contiguous pages | |
313 | * should be using vmem_alloc() in the linux kernel | |
314 | */ | |
79c76d5b | 315 | h->hash_table = vmem_zalloc(hsize * sizeof (void *), KM_SLEEP); |
00b46022 | 316 | #else |
34dc7c2f | 317 | h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP); |
00b46022 | 318 | #endif |
34dc7c2f BB |
319 | if (h->hash_table == NULL) { |
320 | /* XXX - we should really return an error instead of assert */ | |
321 | ASSERT(hsize > (1ULL << 10)); | |
322 | hsize >>= 1; | |
323 | goto retry; | |
324 | } | |
325 | ||
326 | dbuf_cache = kmem_cache_create("dmu_buf_impl_t", | |
327 | sizeof (dmu_buf_impl_t), | |
328 | 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0); | |
329 | ||
330 | for (i = 0; i < DBUF_MUTEXES; i++) | |
40d06e3c | 331 | mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL); |
e0b0ca98 BB |
332 | |
333 | dbuf_stats_init(h); | |
34dc7c2f BB |
334 | } |
335 | ||
336 | void | |
337 | dbuf_fini(void) | |
338 | { | |
339 | dbuf_hash_table_t *h = &dbuf_hash_table; | |
340 | int i; | |
341 | ||
e0b0ca98 BB |
342 | dbuf_stats_destroy(); |
343 | ||
34dc7c2f BB |
344 | for (i = 0; i < DBUF_MUTEXES; i++) |
345 | mutex_destroy(&h->hash_mutexes[i]); | |
00b46022 | 346 | #if defined(_KERNEL) && defined(HAVE_SPL) |
d1d7e268 MK |
347 | /* |
348 | * Large allocations which do not require contiguous pages | |
349 | * should be using vmem_free() in the linux kernel | |
350 | */ | |
00b46022 BB |
351 | vmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *)); |
352 | #else | |
34dc7c2f | 353 | kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *)); |
00b46022 | 354 | #endif |
34dc7c2f BB |
355 | kmem_cache_destroy(dbuf_cache); |
356 | } | |
357 | ||
358 | /* | |
359 | * Other stuff. | |
360 | */ | |
361 | ||
362 | #ifdef ZFS_DEBUG | |
363 | static void | |
364 | dbuf_verify(dmu_buf_impl_t *db) | |
365 | { | |
572e2857 | 366 | dnode_t *dn; |
428870ff | 367 | dbuf_dirty_record_t *dr; |
34dc7c2f BB |
368 | |
369 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
370 | ||
371 | if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY)) | |
372 | return; | |
373 | ||
374 | ASSERT(db->db_objset != NULL); | |
572e2857 BB |
375 | DB_DNODE_ENTER(db); |
376 | dn = DB_DNODE(db); | |
34dc7c2f BB |
377 | if (dn == NULL) { |
378 | ASSERT(db->db_parent == NULL); | |
379 | ASSERT(db->db_blkptr == NULL); | |
380 | } else { | |
381 | ASSERT3U(db->db.db_object, ==, dn->dn_object); | |
382 | ASSERT3P(db->db_objset, ==, dn->dn_objset); | |
383 | ASSERT3U(db->db_level, <, dn->dn_nlevels); | |
572e2857 BB |
384 | ASSERT(db->db_blkid == DMU_BONUS_BLKID || |
385 | db->db_blkid == DMU_SPILL_BLKID || | |
8951cb8d | 386 | !avl_is_empty(&dn->dn_dbufs)); |
34dc7c2f | 387 | } |
428870ff BB |
388 | if (db->db_blkid == DMU_BONUS_BLKID) { |
389 | ASSERT(dn != NULL); | |
390 | ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); | |
391 | ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID); | |
392 | } else if (db->db_blkid == DMU_SPILL_BLKID) { | |
34dc7c2f BB |
393 | ASSERT(dn != NULL); |
394 | ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); | |
c99c9001 | 395 | ASSERT0(db->db.db_offset); |
34dc7c2f BB |
396 | } else { |
397 | ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size); | |
398 | } | |
399 | ||
428870ff BB |
400 | for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next) |
401 | ASSERT(dr->dr_dbuf == db); | |
402 | ||
403 | for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next) | |
404 | ASSERT(dr->dr_dbuf == db); | |
405 | ||
b128c09f BB |
406 | /* |
407 | * We can't assert that db_size matches dn_datablksz because it | |
408 | * can be momentarily different when another thread is doing | |
409 | * dnode_set_blksz(). | |
410 | */ | |
411 | if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) { | |
428870ff | 412 | dr = db->db_data_pending; |
b128c09f BB |
413 | /* |
414 | * It should only be modified in syncing context, so | |
415 | * make sure we only have one copy of the data. | |
416 | */ | |
417 | ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf); | |
34dc7c2f BB |
418 | } |
419 | ||
420 | /* verify db->db_blkptr */ | |
421 | if (db->db_blkptr) { | |
422 | if (db->db_parent == dn->dn_dbuf) { | |
423 | /* db is pointed to by the dnode */ | |
424 | /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */ | |
9babb374 | 425 | if (DMU_OBJECT_IS_SPECIAL(db->db.db_object)) |
34dc7c2f BB |
426 | ASSERT(db->db_parent == NULL); |
427 | else | |
428 | ASSERT(db->db_parent != NULL); | |
428870ff BB |
429 | if (db->db_blkid != DMU_SPILL_BLKID) |
430 | ASSERT3P(db->db_blkptr, ==, | |
431 | &dn->dn_phys->dn_blkptr[db->db_blkid]); | |
34dc7c2f BB |
432 | } else { |
433 | /* db is pointed to by an indirect block */ | |
1fde1e37 BB |
434 | ASSERTV(int epb = db->db_parent->db.db_size >> |
435 | SPA_BLKPTRSHIFT); | |
34dc7c2f BB |
436 | ASSERT3U(db->db_parent->db_level, ==, db->db_level+1); |
437 | ASSERT3U(db->db_parent->db.db_object, ==, | |
438 | db->db.db_object); | |
439 | /* | |
440 | * dnode_grow_indblksz() can make this fail if we don't | |
441 | * have the struct_rwlock. XXX indblksz no longer | |
442 | * grows. safe to do this now? | |
443 | */ | |
572e2857 | 444 | if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) { |
34dc7c2f BB |
445 | ASSERT3P(db->db_blkptr, ==, |
446 | ((blkptr_t *)db->db_parent->db.db_data + | |
447 | db->db_blkid % epb)); | |
448 | } | |
449 | } | |
450 | } | |
451 | if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) && | |
428870ff BB |
452 | (db->db_buf == NULL || db->db_buf->b_data) && |
453 | db->db.db_data && db->db_blkid != DMU_BONUS_BLKID && | |
34dc7c2f BB |
454 | db->db_state != DB_FILL && !dn->dn_free_txg) { |
455 | /* | |
456 | * If the blkptr isn't set but they have nonzero data, | |
457 | * it had better be dirty, otherwise we'll lose that | |
458 | * data when we evict this buffer. | |
459 | */ | |
460 | if (db->db_dirtycnt == 0) { | |
1fde1e37 | 461 | ASSERTV(uint64_t *buf = db->db.db_data); |
34dc7c2f BB |
462 | int i; |
463 | ||
464 | for (i = 0; i < db->db.db_size >> 3; i++) { | |
465 | ASSERT(buf[i] == 0); | |
466 | } | |
467 | } | |
468 | } | |
572e2857 | 469 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
470 | } |
471 | #endif | |
472 | ||
34dc7c2f BB |
473 | static void |
474 | dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf) | |
475 | { | |
476 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
34dc7c2f BB |
477 | db->db_buf = buf; |
478 | if (buf != NULL) { | |
479 | ASSERT(buf->b_data != NULL); | |
480 | db->db.db_data = buf->b_data; | |
481 | if (!arc_released(buf)) | |
482 | arc_set_callback(buf, dbuf_do_evict, db); | |
34dc7c2f BB |
483 | } else { |
484 | dbuf_evict_user(db); | |
485 | db->db.db_data = NULL; | |
b128c09f BB |
486 | if (db->db_state != DB_NOFILL) |
487 | db->db_state = DB_UNCACHED; | |
34dc7c2f BB |
488 | } |
489 | } | |
490 | ||
428870ff BB |
491 | /* |
492 | * Loan out an arc_buf for read. Return the loaned arc_buf. | |
493 | */ | |
494 | arc_buf_t * | |
495 | dbuf_loan_arcbuf(dmu_buf_impl_t *db) | |
496 | { | |
497 | arc_buf_t *abuf; | |
498 | ||
499 | mutex_enter(&db->db_mtx); | |
500 | if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) { | |
501 | int blksz = db->db.db_size; | |
b0bc7a84 | 502 | spa_t *spa = db->db_objset->os_spa; |
572e2857 | 503 | |
428870ff | 504 | mutex_exit(&db->db_mtx); |
572e2857 | 505 | abuf = arc_loan_buf(spa, blksz); |
428870ff BB |
506 | bcopy(db->db.db_data, abuf->b_data, blksz); |
507 | } else { | |
508 | abuf = db->db_buf; | |
509 | arc_loan_inuse_buf(abuf, db); | |
510 | dbuf_set_data(db, NULL); | |
511 | mutex_exit(&db->db_mtx); | |
512 | } | |
513 | return (abuf); | |
514 | } | |
515 | ||
34dc7c2f BB |
516 | uint64_t |
517 | dbuf_whichblock(dnode_t *dn, uint64_t offset) | |
518 | { | |
519 | if (dn->dn_datablkshift) { | |
520 | return (offset >> dn->dn_datablkshift); | |
521 | } else { | |
522 | ASSERT3U(offset, <, dn->dn_datablksz); | |
523 | return (0); | |
524 | } | |
525 | } | |
526 | ||
527 | static void | |
528 | dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb) | |
529 | { | |
530 | dmu_buf_impl_t *db = vdb; | |
531 | ||
532 | mutex_enter(&db->db_mtx); | |
533 | ASSERT3U(db->db_state, ==, DB_READ); | |
534 | /* | |
535 | * All reads are synchronous, so we must have a hold on the dbuf | |
536 | */ | |
537 | ASSERT(refcount_count(&db->db_holds) > 0); | |
538 | ASSERT(db->db_buf == NULL); | |
539 | ASSERT(db->db.db_data == NULL); | |
540 | if (db->db_level == 0 && db->db_freed_in_flight) { | |
541 | /* we were freed in flight; disregard any error */ | |
542 | arc_release(buf, db); | |
543 | bzero(buf->b_data, db->db.db_size); | |
544 | arc_buf_freeze(buf); | |
545 | db->db_freed_in_flight = FALSE; | |
546 | dbuf_set_data(db, buf); | |
547 | db->db_state = DB_CACHED; | |
548 | } else if (zio == NULL || zio->io_error == 0) { | |
549 | dbuf_set_data(db, buf); | |
550 | db->db_state = DB_CACHED; | |
551 | } else { | |
428870ff | 552 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f | 553 | ASSERT3P(db->db_buf, ==, NULL); |
13fe0198 | 554 | VERIFY(arc_buf_remove_ref(buf, db)); |
34dc7c2f BB |
555 | db->db_state = DB_UNCACHED; |
556 | } | |
557 | cv_broadcast(&db->db_changed); | |
428870ff | 558 | dbuf_rele_and_unlock(db, NULL); |
34dc7c2f BB |
559 | } |
560 | ||
5f6d0b6f | 561 | static int |
34dc7c2f BB |
562 | dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t *flags) |
563 | { | |
572e2857 | 564 | dnode_t *dn; |
5dbd68a3 | 565 | zbookmark_phys_t zb; |
34dc7c2f | 566 | uint32_t aflags = ARC_NOWAIT; |
5f6d0b6f | 567 | int err; |
34dc7c2f | 568 | |
572e2857 BB |
569 | DB_DNODE_ENTER(db); |
570 | dn = DB_DNODE(db); | |
34dc7c2f BB |
571 | ASSERT(!refcount_is_zero(&db->db_holds)); |
572 | /* We need the struct_rwlock to prevent db_blkptr from changing. */ | |
b128c09f | 573 | ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); |
34dc7c2f BB |
574 | ASSERT(MUTEX_HELD(&db->db_mtx)); |
575 | ASSERT(db->db_state == DB_UNCACHED); | |
576 | ASSERT(db->db_buf == NULL); | |
577 | ||
428870ff | 578 | if (db->db_blkid == DMU_BONUS_BLKID) { |
9babb374 | 579 | int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen); |
34dc7c2f BB |
580 | |
581 | ASSERT3U(bonuslen, <=, db->db.db_size); | |
582 | db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN); | |
d164b209 | 583 | arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); |
34dc7c2f BB |
584 | if (bonuslen < DN_MAX_BONUSLEN) |
585 | bzero(db->db.db_data, DN_MAX_BONUSLEN); | |
9babb374 BB |
586 | if (bonuslen) |
587 | bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen); | |
572e2857 | 588 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
589 | db->db_state = DB_CACHED; |
590 | mutex_exit(&db->db_mtx); | |
5f6d0b6f | 591 | return (0); |
34dc7c2f BB |
592 | } |
593 | ||
b128c09f BB |
594 | /* |
595 | * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync() | |
596 | * processes the delete record and clears the bp while we are waiting | |
597 | * for the dn_mtx (resulting in a "no" from block_freed). | |
598 | */ | |
599 | if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) || | |
600 | (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) || | |
601 | BP_IS_HOLE(db->db_blkptr)))) { | |
34dc7c2f BB |
602 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); |
603 | ||
572e2857 | 604 | DB_DNODE_EXIT(db); |
b0bc7a84 MG |
605 | dbuf_set_data(db, arc_buf_alloc(db->db_objset->os_spa, |
606 | db->db.db_size, db, type)); | |
34dc7c2f BB |
607 | bzero(db->db.db_data, db->db.db_size); |
608 | db->db_state = DB_CACHED; | |
609 | *flags |= DB_RF_CACHED; | |
610 | mutex_exit(&db->db_mtx); | |
5f6d0b6f | 611 | return (0); |
34dc7c2f BB |
612 | } |
613 | ||
572e2857 BB |
614 | DB_DNODE_EXIT(db); |
615 | ||
34dc7c2f BB |
616 | db->db_state = DB_READ; |
617 | mutex_exit(&db->db_mtx); | |
618 | ||
b128c09f BB |
619 | if (DBUF_IS_L2CACHEABLE(db)) |
620 | aflags |= ARC_L2CACHE; | |
3a17a7a9 SK |
621 | if (DBUF_IS_L2COMPRESSIBLE(db)) |
622 | aflags |= ARC_L2COMPRESS; | |
b128c09f | 623 | |
428870ff BB |
624 | SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ? |
625 | db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET, | |
626 | db->db.db_object, db->db_level, db->db_blkid); | |
34dc7c2f BB |
627 | |
628 | dbuf_add_ref(db, NULL); | |
b128c09f | 629 | |
5f6d0b6f | 630 | err = arc_read(zio, db->db_objset->os_spa, db->db_blkptr, |
34dc7c2f BB |
631 | dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ, |
632 | (*flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED, | |
633 | &aflags, &zb); | |
634 | if (aflags & ARC_CACHED) | |
635 | *flags |= DB_RF_CACHED; | |
5f6d0b6f BB |
636 | |
637 | return (SET_ERROR(err)); | |
34dc7c2f BB |
638 | } |
639 | ||
640 | int | |
641 | dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) | |
642 | { | |
643 | int err = 0; | |
b0bc7a84 MG |
644 | boolean_t havepzio = (zio != NULL); |
645 | boolean_t prefetch; | |
572e2857 | 646 | dnode_t *dn; |
34dc7c2f BB |
647 | |
648 | /* | |
649 | * We don't have to hold the mutex to check db_state because it | |
650 | * can't be freed while we have a hold on the buffer. | |
651 | */ | |
652 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
653 | ||
b128c09f | 654 | if (db->db_state == DB_NOFILL) |
2e528b49 | 655 | return (SET_ERROR(EIO)); |
b128c09f | 656 | |
572e2857 BB |
657 | DB_DNODE_ENTER(db); |
658 | dn = DB_DNODE(db); | |
34dc7c2f | 659 | if ((flags & DB_RF_HAVESTRUCT) == 0) |
572e2857 | 660 | rw_enter(&dn->dn_struct_rwlock, RW_READER); |
34dc7c2f | 661 | |
428870ff | 662 | prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && |
572e2857 | 663 | (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL && |
b128c09f | 664 | DBUF_IS_CACHEABLE(db); |
34dc7c2f BB |
665 | |
666 | mutex_enter(&db->db_mtx); | |
667 | if (db->db_state == DB_CACHED) { | |
668 | mutex_exit(&db->db_mtx); | |
669 | if (prefetch) | |
572e2857 | 670 | dmu_zfetch(&dn->dn_zfetch, db->db.db_offset, |
34dc7c2f BB |
671 | db->db.db_size, TRUE); |
672 | if ((flags & DB_RF_HAVESTRUCT) == 0) | |
572e2857 BB |
673 | rw_exit(&dn->dn_struct_rwlock); |
674 | DB_DNODE_EXIT(db); | |
34dc7c2f | 675 | } else if (db->db_state == DB_UNCACHED) { |
572e2857 BB |
676 | spa_t *spa = dn->dn_objset->os_spa; |
677 | ||
678 | if (zio == NULL) | |
679 | zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); | |
5f6d0b6f BB |
680 | |
681 | err = dbuf_read_impl(db, zio, &flags); | |
34dc7c2f BB |
682 | |
683 | /* dbuf_read_impl has dropped db_mtx for us */ | |
684 | ||
5f6d0b6f | 685 | if (!err && prefetch) |
572e2857 | 686 | dmu_zfetch(&dn->dn_zfetch, db->db.db_offset, |
34dc7c2f BB |
687 | db->db.db_size, flags & DB_RF_CACHED); |
688 | ||
689 | if ((flags & DB_RF_HAVESTRUCT) == 0) | |
572e2857 BB |
690 | rw_exit(&dn->dn_struct_rwlock); |
691 | DB_DNODE_EXIT(db); | |
34dc7c2f | 692 | |
5f6d0b6f | 693 | if (!err && !havepzio) |
34dc7c2f BB |
694 | err = zio_wait(zio); |
695 | } else { | |
e49f1e20 WA |
696 | /* |
697 | * Another reader came in while the dbuf was in flight | |
698 | * between UNCACHED and CACHED. Either a writer will finish | |
699 | * writing the buffer (sending the dbuf to CACHED) or the | |
700 | * first reader's request will reach the read_done callback | |
701 | * and send the dbuf to CACHED. Otherwise, a failure | |
702 | * occurred and the dbuf went to UNCACHED. | |
703 | */ | |
34dc7c2f BB |
704 | mutex_exit(&db->db_mtx); |
705 | if (prefetch) | |
572e2857 | 706 | dmu_zfetch(&dn->dn_zfetch, db->db.db_offset, |
34dc7c2f BB |
707 | db->db.db_size, TRUE); |
708 | if ((flags & DB_RF_HAVESTRUCT) == 0) | |
572e2857 BB |
709 | rw_exit(&dn->dn_struct_rwlock); |
710 | DB_DNODE_EXIT(db); | |
34dc7c2f | 711 | |
e49f1e20 | 712 | /* Skip the wait per the caller's request. */ |
34dc7c2f BB |
713 | mutex_enter(&db->db_mtx); |
714 | if ((flags & DB_RF_NEVERWAIT) == 0) { | |
715 | while (db->db_state == DB_READ || | |
716 | db->db_state == DB_FILL) { | |
717 | ASSERT(db->db_state == DB_READ || | |
718 | (flags & DB_RF_HAVESTRUCT) == 0); | |
64dbba36 AL |
719 | DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *, |
720 | db, zio_t *, zio); | |
34dc7c2f BB |
721 | cv_wait(&db->db_changed, &db->db_mtx); |
722 | } | |
723 | if (db->db_state == DB_UNCACHED) | |
2e528b49 | 724 | err = SET_ERROR(EIO); |
34dc7c2f BB |
725 | } |
726 | mutex_exit(&db->db_mtx); | |
727 | } | |
728 | ||
729 | ASSERT(err || havepzio || db->db_state == DB_CACHED); | |
730 | return (err); | |
731 | } | |
732 | ||
733 | static void | |
734 | dbuf_noread(dmu_buf_impl_t *db) | |
735 | { | |
736 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
428870ff | 737 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f BB |
738 | mutex_enter(&db->db_mtx); |
739 | while (db->db_state == DB_READ || db->db_state == DB_FILL) | |
740 | cv_wait(&db->db_changed, &db->db_mtx); | |
741 | if (db->db_state == DB_UNCACHED) { | |
742 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); | |
b0bc7a84 | 743 | spa_t *spa = db->db_objset->os_spa; |
34dc7c2f BB |
744 | |
745 | ASSERT(db->db_buf == NULL); | |
746 | ASSERT(db->db.db_data == NULL); | |
572e2857 | 747 | dbuf_set_data(db, arc_buf_alloc(spa, db->db.db_size, db, type)); |
34dc7c2f | 748 | db->db_state = DB_FILL; |
b128c09f BB |
749 | } else if (db->db_state == DB_NOFILL) { |
750 | dbuf_set_data(db, NULL); | |
34dc7c2f BB |
751 | } else { |
752 | ASSERT3U(db->db_state, ==, DB_CACHED); | |
753 | } | |
754 | mutex_exit(&db->db_mtx); | |
755 | } | |
756 | ||
757 | /* | |
758 | * This is our just-in-time copy function. It makes a copy of | |
759 | * buffers, that have been modified in a previous transaction | |
760 | * group, before we modify them in the current active group. | |
761 | * | |
762 | * This function is used in two places: when we are dirtying a | |
763 | * buffer for the first time in a txg, and when we are freeing | |
764 | * a range in a dnode that includes this buffer. | |
765 | * | |
766 | * Note that when we are called from dbuf_free_range() we do | |
767 | * not put a hold on the buffer, we just traverse the active | |
768 | * dbuf list for the dnode. | |
769 | */ | |
770 | static void | |
771 | dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg) | |
772 | { | |
773 | dbuf_dirty_record_t *dr = db->db_last_dirty; | |
774 | ||
775 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
776 | ASSERT(db->db.db_data != NULL); | |
777 | ASSERT(db->db_level == 0); | |
778 | ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT); | |
779 | ||
780 | if (dr == NULL || | |
781 | (dr->dt.dl.dr_data != | |
428870ff | 782 | ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf))) |
34dc7c2f BB |
783 | return; |
784 | ||
785 | /* | |
786 | * If the last dirty record for this dbuf has not yet synced | |
787 | * and its referencing the dbuf data, either: | |
572e2857 | 788 | * reset the reference to point to a new copy, |
34dc7c2f BB |
789 | * or (if there a no active holders) |
790 | * just null out the current db_data pointer. | |
791 | */ | |
792 | ASSERT(dr->dr_txg >= txg - 2); | |
428870ff | 793 | if (db->db_blkid == DMU_BONUS_BLKID) { |
34dc7c2f BB |
794 | /* Note that the data bufs here are zio_bufs */ |
795 | dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN); | |
d164b209 | 796 | arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); |
34dc7c2f BB |
797 | bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN); |
798 | } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) { | |
799 | int size = db->db.db_size; | |
800 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); | |
b0bc7a84 | 801 | spa_t *spa = db->db_objset->os_spa; |
572e2857 | 802 | |
572e2857 | 803 | dr->dt.dl.dr_data = arc_buf_alloc(spa, size, db, type); |
34dc7c2f BB |
804 | bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size); |
805 | } else { | |
806 | dbuf_set_data(db, NULL); | |
807 | } | |
808 | } | |
809 | ||
810 | void | |
811 | dbuf_unoverride(dbuf_dirty_record_t *dr) | |
812 | { | |
813 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
428870ff | 814 | blkptr_t *bp = &dr->dt.dl.dr_overridden_by; |
34dc7c2f BB |
815 | uint64_t txg = dr->dr_txg; |
816 | ||
817 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
818 | ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC); | |
819 | ASSERT(db->db_level == 0); | |
820 | ||
428870ff | 821 | if (db->db_blkid == DMU_BONUS_BLKID || |
34dc7c2f BB |
822 | dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN) |
823 | return; | |
824 | ||
428870ff BB |
825 | ASSERT(db->db_data_pending != dr); |
826 | ||
34dc7c2f | 827 | /* free this block */ |
b0bc7a84 MG |
828 | if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite) |
829 | zio_free(db->db_objset->os_spa, txg, bp); | |
428870ff | 830 | |
34dc7c2f | 831 | dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; |
03c6040b GW |
832 | dr->dt.dl.dr_nopwrite = B_FALSE; |
833 | ||
34dc7c2f BB |
834 | /* |
835 | * Release the already-written buffer, so we leave it in | |
836 | * a consistent dirty state. Note that all callers are | |
837 | * modifying the buffer, so they will immediately do | |
838 | * another (redundant) arc_release(). Therefore, leave | |
839 | * the buf thawed to save the effort of freezing & | |
840 | * immediately re-thawing it. | |
841 | */ | |
842 | arc_release(dr->dt.dl.dr_data, db); | |
843 | } | |
844 | ||
b128c09f BB |
845 | /* |
846 | * Evict (if its unreferenced) or clear (if its referenced) any level-0 | |
847 | * data blocks in the free range, so that any future readers will find | |
b0bc7a84 | 848 | * empty blocks. |
ea97f8ce MA |
849 | * |
850 | * This is a no-op if the dataset is in the middle of an incremental | |
851 | * receive; see comment below for details. | |
b128c09f | 852 | */ |
34dc7c2f | 853 | void |
8951cb8d AR |
854 | dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid, |
855 | dmu_tx_t *tx) | |
34dc7c2f | 856 | { |
8951cb8d | 857 | dmu_buf_impl_t *db, *db_next, *db_search; |
34dc7c2f | 858 | uint64_t txg = tx->tx_txg; |
8951cb8d | 859 | avl_index_t where; |
4254acb0 | 860 | boolean_t freespill = |
8951cb8d AR |
861 | (start_blkid == DMU_SPILL_BLKID || end_blkid == DMU_SPILL_BLKID); |
862 | ||
863 | if (end_blkid > dn->dn_maxblkid && !freespill) | |
864 | end_blkid = dn->dn_maxblkid; | |
865 | dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid); | |
34dc7c2f | 866 | |
8951cb8d AR |
867 | db_seach = kmem_alloc(sizeof (dmu_buf_impl_t), KM_SLEEP); |
868 | db_search->db_level = 0; | |
869 | db_search->db_blkid = start_blkid; | |
9925c28c | 870 | db_search->db_state = DB_SEARCH; |
ea97f8ce | 871 | |
b663a23d | 872 | mutex_enter(&dn->dn_dbufs_mtx); |
8951cb8d | 873 | if (start_blkid >= dn->dn_unlisted_l0_blkid && !freespill) { |
b663a23d | 874 | /* There can't be any dbufs in this range; no need to search. */ |
8951cb8d AR |
875 | #ifdef DEBUG |
876 | db = avl_find(&dn->dn_dbufs, db_search, &where); | |
877 | ASSERT3P(db, ==, NULL); | |
878 | db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER); | |
879 | ASSERT(db == NULL || db->db_level > 0); | |
880 | #endif | |
881 | goto out; | |
b663a23d | 882 | } else if (dmu_objset_is_receiving(dn->dn_objset)) { |
ea97f8ce | 883 | /* |
b663a23d MA |
884 | * If we are receiving, we expect there to be no dbufs in |
885 | * the range to be freed, because receive modifies each | |
886 | * block at most once, and in offset order. If this is | |
887 | * not the case, it can lead to performance problems, | |
888 | * so note that we unexpectedly took the slow path. | |
ea97f8ce | 889 | */ |
b663a23d | 890 | atomic_inc_64(&zfs_free_range_recv_miss); |
ea97f8ce MA |
891 | } |
892 | ||
8951cb8d AR |
893 | db = avl_find(&dn->dn_dbufs, db_search, &where); |
894 | ASSERT3P(db, ==, NULL); | |
895 | db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER); | |
896 | ||
897 | for (; db != NULL; db = db_next) { | |
898 | db_next = AVL_NEXT(&dn->dn_dbufs, db); | |
428870ff | 899 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
b128c09f | 900 | |
8951cb8d AR |
901 | if (db->db_level != 0 || db->db_blkid > end_blkid) { |
902 | break; | |
903 | } | |
904 | ASSERT3U(db->db_blkid, >=, start_blkid); | |
34dc7c2f BB |
905 | |
906 | /* found a level 0 buffer in the range */ | |
13fe0198 MA |
907 | mutex_enter(&db->db_mtx); |
908 | if (dbuf_undirty(db, tx)) { | |
909 | /* mutex has been dropped and dbuf destroyed */ | |
34dc7c2f | 910 | continue; |
13fe0198 | 911 | } |
34dc7c2f | 912 | |
34dc7c2f | 913 | if (db->db_state == DB_UNCACHED || |
b128c09f | 914 | db->db_state == DB_NOFILL || |
34dc7c2f BB |
915 | db->db_state == DB_EVICTING) { |
916 | ASSERT(db->db.db_data == NULL); | |
917 | mutex_exit(&db->db_mtx); | |
918 | continue; | |
919 | } | |
920 | if (db->db_state == DB_READ || db->db_state == DB_FILL) { | |
921 | /* will be handled in dbuf_read_done or dbuf_rele */ | |
922 | db->db_freed_in_flight = TRUE; | |
923 | mutex_exit(&db->db_mtx); | |
924 | continue; | |
925 | } | |
926 | if (refcount_count(&db->db_holds) == 0) { | |
927 | ASSERT(db->db_buf); | |
928 | dbuf_clear(db); | |
929 | continue; | |
930 | } | |
931 | /* The dbuf is referenced */ | |
932 | ||
933 | if (db->db_last_dirty != NULL) { | |
934 | dbuf_dirty_record_t *dr = db->db_last_dirty; | |
935 | ||
936 | if (dr->dr_txg == txg) { | |
937 | /* | |
938 | * This buffer is "in-use", re-adjust the file | |
939 | * size to reflect that this buffer may | |
940 | * contain new data when we sync. | |
941 | */ | |
428870ff BB |
942 | if (db->db_blkid != DMU_SPILL_BLKID && |
943 | db->db_blkid > dn->dn_maxblkid) | |
34dc7c2f BB |
944 | dn->dn_maxblkid = db->db_blkid; |
945 | dbuf_unoverride(dr); | |
946 | } else { | |
947 | /* | |
948 | * This dbuf is not dirty in the open context. | |
949 | * Either uncache it (if its not referenced in | |
950 | * the open context) or reset its contents to | |
951 | * empty. | |
952 | */ | |
953 | dbuf_fix_old_data(db, txg); | |
954 | } | |
955 | } | |
956 | /* clear the contents if its cached */ | |
957 | if (db->db_state == DB_CACHED) { | |
958 | ASSERT(db->db.db_data != NULL); | |
959 | arc_release(db->db_buf, db); | |
960 | bzero(db->db.db_data, db->db.db_size); | |
961 | arc_buf_freeze(db->db_buf); | |
962 | } | |
963 | ||
964 | mutex_exit(&db->db_mtx); | |
965 | } | |
8951cb8d AR |
966 | |
967 | out: | |
968 | kmem_free(db_search, sizeof (dmu_buf_impl_t)); | |
34dc7c2f BB |
969 | mutex_exit(&dn->dn_dbufs_mtx); |
970 | } | |
971 | ||
972 | static int | |
973 | dbuf_block_freeable(dmu_buf_impl_t *db) | |
974 | { | |
975 | dsl_dataset_t *ds = db->db_objset->os_dsl_dataset; | |
976 | uint64_t birth_txg = 0; | |
977 | ||
978 | /* | |
979 | * We don't need any locking to protect db_blkptr: | |
980 | * If it's syncing, then db_last_dirty will be set | |
981 | * so we'll ignore db_blkptr. | |
b0bc7a84 MG |
982 | * |
983 | * This logic ensures that only block births for | |
984 | * filled blocks are considered. | |
34dc7c2f BB |
985 | */ |
986 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
b0bc7a84 MG |
987 | if (db->db_last_dirty && (db->db_blkptr == NULL || |
988 | !BP_IS_HOLE(db->db_blkptr))) { | |
34dc7c2f | 989 | birth_txg = db->db_last_dirty->dr_txg; |
b0bc7a84 | 990 | } else if (db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) { |
34dc7c2f | 991 | birth_txg = db->db_blkptr->blk_birth; |
b0bc7a84 | 992 | } |
34dc7c2f | 993 | |
572e2857 | 994 | /* |
b0bc7a84 | 995 | * If this block don't exist or is in a snapshot, it can't be freed. |
572e2857 BB |
996 | * Don't pass the bp to dsl_dataset_block_freeable() since we |
997 | * are holding the db_mtx lock and might deadlock if we are | |
998 | * prefetching a dedup-ed block. | |
999 | */ | |
b0bc7a84 | 1000 | if (birth_txg != 0) |
34dc7c2f | 1001 | return (ds == NULL || |
572e2857 | 1002 | dsl_dataset_block_freeable(ds, NULL, birth_txg)); |
34dc7c2f | 1003 | else |
b0bc7a84 | 1004 | return (B_FALSE); |
34dc7c2f BB |
1005 | } |
1006 | ||
1007 | void | |
1008 | dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx) | |
1009 | { | |
1010 | arc_buf_t *buf, *obuf; | |
1011 | int osize = db->db.db_size; | |
1012 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); | |
572e2857 | 1013 | dnode_t *dn; |
34dc7c2f | 1014 | |
428870ff | 1015 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f | 1016 | |
572e2857 BB |
1017 | DB_DNODE_ENTER(db); |
1018 | dn = DB_DNODE(db); | |
1019 | ||
34dc7c2f | 1020 | /* XXX does *this* func really need the lock? */ |
572e2857 | 1021 | ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); |
34dc7c2f BB |
1022 | |
1023 | /* | |
b0bc7a84 | 1024 | * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held |
34dc7c2f BB |
1025 | * is OK, because there can be no other references to the db |
1026 | * when we are changing its size, so no concurrent DB_FILL can | |
1027 | * be happening. | |
1028 | */ | |
1029 | /* | |
1030 | * XXX we should be doing a dbuf_read, checking the return | |
1031 | * value and returning that up to our callers | |
1032 | */ | |
b0bc7a84 | 1033 | dmu_buf_will_dirty(&db->db, tx); |
34dc7c2f BB |
1034 | |
1035 | /* create the data buffer for the new block */ | |
572e2857 | 1036 | buf = arc_buf_alloc(dn->dn_objset->os_spa, size, db, type); |
34dc7c2f BB |
1037 | |
1038 | /* copy old block data to the new block */ | |
1039 | obuf = db->db_buf; | |
1040 | bcopy(obuf->b_data, buf->b_data, MIN(osize, size)); | |
1041 | /* zero the remainder */ | |
1042 | if (size > osize) | |
1043 | bzero((uint8_t *)buf->b_data + osize, size - osize); | |
1044 | ||
1045 | mutex_enter(&db->db_mtx); | |
1046 | dbuf_set_data(db, buf); | |
13fe0198 | 1047 | VERIFY(arc_buf_remove_ref(obuf, db)); |
34dc7c2f BB |
1048 | db->db.db_size = size; |
1049 | ||
1050 | if (db->db_level == 0) { | |
1051 | ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); | |
1052 | db->db_last_dirty->dt.dl.dr_data = buf; | |
1053 | } | |
1054 | mutex_exit(&db->db_mtx); | |
1055 | ||
572e2857 BB |
1056 | dnode_willuse_space(dn, size-osize, tx); |
1057 | DB_DNODE_EXIT(db); | |
34dc7c2f BB |
1058 | } |
1059 | ||
428870ff BB |
1060 | void |
1061 | dbuf_release_bp(dmu_buf_impl_t *db) | |
1062 | { | |
b0bc7a84 | 1063 | ASSERTV(objset_t *os = db->db_objset); |
428870ff BB |
1064 | |
1065 | ASSERT(dsl_pool_sync_context(dmu_objset_pool(os))); | |
1066 | ASSERT(arc_released(os->os_phys_buf) || | |
1067 | list_link_active(&os->os_dsl_dataset->ds_synced_link)); | |
1068 | ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf)); | |
1069 | ||
294f6806 | 1070 | (void) arc_release(db->db_buf, db); |
428870ff BB |
1071 | } |
1072 | ||
34dc7c2f BB |
1073 | dbuf_dirty_record_t * |
1074 | dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) | |
1075 | { | |
572e2857 BB |
1076 | dnode_t *dn; |
1077 | objset_t *os; | |
34dc7c2f BB |
1078 | dbuf_dirty_record_t **drp, *dr; |
1079 | int drop_struct_lock = FALSE; | |
b128c09f | 1080 | boolean_t do_free_accounting = B_FALSE; |
34dc7c2f BB |
1081 | int txgoff = tx->tx_txg & TXG_MASK; |
1082 | ||
1083 | ASSERT(tx->tx_txg != 0); | |
1084 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
1085 | DMU_TX_DIRTY_BUF(tx, db); | |
1086 | ||
572e2857 BB |
1087 | DB_DNODE_ENTER(db); |
1088 | dn = DB_DNODE(db); | |
34dc7c2f BB |
1089 | /* |
1090 | * Shouldn't dirty a regular buffer in syncing context. Private | |
1091 | * objects may be dirtied in syncing context, but only if they | |
1092 | * were already pre-dirtied in open context. | |
34dc7c2f BB |
1093 | */ |
1094 | ASSERT(!dmu_tx_is_syncing(tx) || | |
1095 | BP_IS_HOLE(dn->dn_objset->os_rootbp) || | |
9babb374 BB |
1096 | DMU_OBJECT_IS_SPECIAL(dn->dn_object) || |
1097 | dn->dn_objset->os_dsl_dataset == NULL); | |
34dc7c2f BB |
1098 | /* |
1099 | * We make this assert for private objects as well, but after we | |
1100 | * check if we're already dirty. They are allowed to re-dirty | |
1101 | * in syncing context. | |
1102 | */ | |
1103 | ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || | |
1104 | dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == | |
1105 | (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); | |
1106 | ||
1107 | mutex_enter(&db->db_mtx); | |
1108 | /* | |
1109 | * XXX make this true for indirects too? The problem is that | |
1110 | * transactions created with dmu_tx_create_assigned() from | |
1111 | * syncing context don't bother holding ahead. | |
1112 | */ | |
1113 | ASSERT(db->db_level != 0 || | |
b128c09f BB |
1114 | db->db_state == DB_CACHED || db->db_state == DB_FILL || |
1115 | db->db_state == DB_NOFILL); | |
34dc7c2f BB |
1116 | |
1117 | mutex_enter(&dn->dn_mtx); | |
1118 | /* | |
1119 | * Don't set dirtyctx to SYNC if we're just modifying this as we | |
1120 | * initialize the objset. | |
1121 | */ | |
1122 | if (dn->dn_dirtyctx == DN_UNDIRTIED && | |
1123 | !BP_IS_HOLE(dn->dn_objset->os_rootbp)) { | |
1124 | dn->dn_dirtyctx = | |
1125 | (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN); | |
1126 | ASSERT(dn->dn_dirtyctx_firstset == NULL); | |
79c76d5b | 1127 | dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP); |
34dc7c2f BB |
1128 | } |
1129 | mutex_exit(&dn->dn_mtx); | |
1130 | ||
428870ff BB |
1131 | if (db->db_blkid == DMU_SPILL_BLKID) |
1132 | dn->dn_have_spill = B_TRUE; | |
1133 | ||
34dc7c2f BB |
1134 | /* |
1135 | * If this buffer is already dirty, we're done. | |
1136 | */ | |
1137 | drp = &db->db_last_dirty; | |
1138 | ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg || | |
1139 | db->db.db_object == DMU_META_DNODE_OBJECT); | |
1140 | while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg) | |
1141 | drp = &dr->dr_next; | |
1142 | if (dr && dr->dr_txg == tx->tx_txg) { | |
572e2857 BB |
1143 | DB_DNODE_EXIT(db); |
1144 | ||
428870ff | 1145 | if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) { |
34dc7c2f BB |
1146 | /* |
1147 | * If this buffer has already been written out, | |
1148 | * we now need to reset its state. | |
1149 | */ | |
1150 | dbuf_unoverride(dr); | |
428870ff BB |
1151 | if (db->db.db_object != DMU_META_DNODE_OBJECT && |
1152 | db->db_state != DB_NOFILL) | |
34dc7c2f BB |
1153 | arc_buf_thaw(db->db_buf); |
1154 | } | |
1155 | mutex_exit(&db->db_mtx); | |
1156 | return (dr); | |
1157 | } | |
1158 | ||
1159 | /* | |
1160 | * Only valid if not already dirty. | |
1161 | */ | |
9babb374 BB |
1162 | ASSERT(dn->dn_object == 0 || |
1163 | dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == | |
34dc7c2f BB |
1164 | (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); |
1165 | ||
1166 | ASSERT3U(dn->dn_nlevels, >, db->db_level); | |
1167 | ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) || | |
1168 | dn->dn_phys->dn_nlevels > db->db_level || | |
1169 | dn->dn_next_nlevels[txgoff] > db->db_level || | |
1170 | dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level || | |
1171 | dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level); | |
1172 | ||
1173 | /* | |
1174 | * We should only be dirtying in syncing context if it's the | |
9babb374 BB |
1175 | * mos or we're initializing the os or it's a special object. |
1176 | * However, we are allowed to dirty in syncing context provided | |
1177 | * we already dirtied it in open context. Hence we must make | |
1178 | * this assertion only if we're not already dirty. | |
34dc7c2f | 1179 | */ |
572e2857 | 1180 | os = dn->dn_objset; |
9babb374 BB |
1181 | ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) || |
1182 | os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp)); | |
34dc7c2f BB |
1183 | ASSERT(db->db.db_size != 0); |
1184 | ||
1185 | dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); | |
1186 | ||
428870ff | 1187 | if (db->db_blkid != DMU_BONUS_BLKID) { |
34dc7c2f BB |
1188 | /* |
1189 | * Update the accounting. | |
b128c09f BB |
1190 | * Note: we delay "free accounting" until after we drop |
1191 | * the db_mtx. This keeps us from grabbing other locks | |
428870ff | 1192 | * (and possibly deadlocking) in bp_get_dsize() while |
b128c09f | 1193 | * also holding the db_mtx. |
34dc7c2f | 1194 | */ |
34dc7c2f | 1195 | dnode_willuse_space(dn, db->db.db_size, tx); |
b128c09f | 1196 | do_free_accounting = dbuf_block_freeable(db); |
34dc7c2f BB |
1197 | } |
1198 | ||
1199 | /* | |
1200 | * If this buffer is dirty in an old transaction group we need | |
1201 | * to make a copy of it so that the changes we make in this | |
1202 | * transaction group won't leak out when we sync the older txg. | |
1203 | */ | |
79c76d5b | 1204 | dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP); |
98f72a53 | 1205 | list_link_init(&dr->dr_dirty_node); |
34dc7c2f BB |
1206 | if (db->db_level == 0) { |
1207 | void *data_old = db->db_buf; | |
1208 | ||
b128c09f | 1209 | if (db->db_state != DB_NOFILL) { |
428870ff | 1210 | if (db->db_blkid == DMU_BONUS_BLKID) { |
b128c09f BB |
1211 | dbuf_fix_old_data(db, tx->tx_txg); |
1212 | data_old = db->db.db_data; | |
1213 | } else if (db->db.db_object != DMU_META_DNODE_OBJECT) { | |
1214 | /* | |
1215 | * Release the data buffer from the cache so | |
1216 | * that we can modify it without impacting | |
1217 | * possible other users of this cached data | |
1218 | * block. Note that indirect blocks and | |
1219 | * private objects are not released until the | |
1220 | * syncing state (since they are only modified | |
1221 | * then). | |
1222 | */ | |
1223 | arc_release(db->db_buf, db); | |
1224 | dbuf_fix_old_data(db, tx->tx_txg); | |
1225 | data_old = db->db_buf; | |
1226 | } | |
1227 | ASSERT(data_old != NULL); | |
34dc7c2f | 1228 | } |
34dc7c2f BB |
1229 | dr->dt.dl.dr_data = data_old; |
1230 | } else { | |
1231 | mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL); | |
1232 | list_create(&dr->dt.di.dr_children, | |
1233 | sizeof (dbuf_dirty_record_t), | |
1234 | offsetof(dbuf_dirty_record_t, dr_dirty_node)); | |
1235 | } | |
e8b96c60 MA |
1236 | if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL) |
1237 | dr->dr_accounted = db->db.db_size; | |
34dc7c2f BB |
1238 | dr->dr_dbuf = db; |
1239 | dr->dr_txg = tx->tx_txg; | |
1240 | dr->dr_next = *drp; | |
1241 | *drp = dr; | |
1242 | ||
1243 | /* | |
1244 | * We could have been freed_in_flight between the dbuf_noread | |
1245 | * and dbuf_dirty. We win, as though the dbuf_noread() had | |
1246 | * happened after the free. | |
1247 | */ | |
428870ff BB |
1248 | if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && |
1249 | db->db_blkid != DMU_SPILL_BLKID) { | |
34dc7c2f | 1250 | mutex_enter(&dn->dn_mtx); |
9bd274dd MA |
1251 | if (dn->dn_free_ranges[txgoff] != NULL) { |
1252 | range_tree_clear(dn->dn_free_ranges[txgoff], | |
1253 | db->db_blkid, 1); | |
1254 | } | |
34dc7c2f BB |
1255 | mutex_exit(&dn->dn_mtx); |
1256 | db->db_freed_in_flight = FALSE; | |
1257 | } | |
1258 | ||
1259 | /* | |
1260 | * This buffer is now part of this txg | |
1261 | */ | |
1262 | dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg); | |
1263 | db->db_dirtycnt += 1; | |
1264 | ASSERT3U(db->db_dirtycnt, <=, 3); | |
1265 | ||
1266 | mutex_exit(&db->db_mtx); | |
1267 | ||
428870ff BB |
1268 | if (db->db_blkid == DMU_BONUS_BLKID || |
1269 | db->db_blkid == DMU_SPILL_BLKID) { | |
34dc7c2f BB |
1270 | mutex_enter(&dn->dn_mtx); |
1271 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
1272 | list_insert_tail(&dn->dn_dirty_records[txgoff], dr); | |
1273 | mutex_exit(&dn->dn_mtx); | |
1274 | dnode_setdirty(dn, tx); | |
572e2857 | 1275 | DB_DNODE_EXIT(db); |
34dc7c2f | 1276 | return (dr); |
b128c09f BB |
1277 | } else if (do_free_accounting) { |
1278 | blkptr_t *bp = db->db_blkptr; | |
1279 | int64_t willfree = (bp && !BP_IS_HOLE(bp)) ? | |
428870ff | 1280 | bp_get_dsize(os->os_spa, bp) : db->db.db_size; |
b128c09f BB |
1281 | /* |
1282 | * This is only a guess -- if the dbuf is dirty | |
1283 | * in a previous txg, we don't know how much | |
1284 | * space it will use on disk yet. We should | |
1285 | * really have the struct_rwlock to access | |
1286 | * db_blkptr, but since this is just a guess, | |
1287 | * it's OK if we get an odd answer. | |
1288 | */ | |
572e2857 | 1289 | ddt_prefetch(os->os_spa, bp); |
b128c09f | 1290 | dnode_willuse_space(dn, -willfree, tx); |
34dc7c2f BB |
1291 | } |
1292 | ||
1293 | if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) { | |
1294 | rw_enter(&dn->dn_struct_rwlock, RW_READER); | |
1295 | drop_struct_lock = TRUE; | |
1296 | } | |
1297 | ||
b128c09f BB |
1298 | if (db->db_level == 0) { |
1299 | dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock); | |
1300 | ASSERT(dn->dn_maxblkid >= db->db_blkid); | |
1301 | } | |
1302 | ||
34dc7c2f BB |
1303 | if (db->db_level+1 < dn->dn_nlevels) { |
1304 | dmu_buf_impl_t *parent = db->db_parent; | |
1305 | dbuf_dirty_record_t *di; | |
1306 | int parent_held = FALSE; | |
1307 | ||
1308 | if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) { | |
1309 | int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; | |
1310 | ||
1311 | parent = dbuf_hold_level(dn, db->db_level+1, | |
1312 | db->db_blkid >> epbs, FTAG); | |
428870ff | 1313 | ASSERT(parent != NULL); |
34dc7c2f BB |
1314 | parent_held = TRUE; |
1315 | } | |
1316 | if (drop_struct_lock) | |
1317 | rw_exit(&dn->dn_struct_rwlock); | |
1318 | ASSERT3U(db->db_level+1, ==, parent->db_level); | |
1319 | di = dbuf_dirty(parent, tx); | |
1320 | if (parent_held) | |
1321 | dbuf_rele(parent, FTAG); | |
1322 | ||
1323 | mutex_enter(&db->db_mtx); | |
e8b96c60 MA |
1324 | /* |
1325 | * Since we've dropped the mutex, it's possible that | |
1326 | * dbuf_undirty() might have changed this out from under us. | |
1327 | */ | |
34dc7c2f BB |
1328 | if (db->db_last_dirty == dr || |
1329 | dn->dn_object == DMU_META_DNODE_OBJECT) { | |
1330 | mutex_enter(&di->dt.di.dr_mtx); | |
1331 | ASSERT3U(di->dr_txg, ==, tx->tx_txg); | |
1332 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
1333 | list_insert_tail(&di->dt.di.dr_children, dr); | |
1334 | mutex_exit(&di->dt.di.dr_mtx); | |
1335 | dr->dr_parent = di; | |
1336 | } | |
1337 | mutex_exit(&db->db_mtx); | |
1338 | } else { | |
1339 | ASSERT(db->db_level+1 == dn->dn_nlevels); | |
1340 | ASSERT(db->db_blkid < dn->dn_nblkptr); | |
572e2857 | 1341 | ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf); |
34dc7c2f BB |
1342 | mutex_enter(&dn->dn_mtx); |
1343 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
1344 | list_insert_tail(&dn->dn_dirty_records[txgoff], dr); | |
1345 | mutex_exit(&dn->dn_mtx); | |
1346 | if (drop_struct_lock) | |
1347 | rw_exit(&dn->dn_struct_rwlock); | |
1348 | } | |
1349 | ||
1350 | dnode_setdirty(dn, tx); | |
572e2857 | 1351 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
1352 | return (dr); |
1353 | } | |
1354 | ||
13fe0198 | 1355 | /* |
e49f1e20 WA |
1356 | * Undirty a buffer in the transaction group referenced by the given |
1357 | * transaction. Return whether this evicted the dbuf. | |
13fe0198 MA |
1358 | */ |
1359 | static boolean_t | |
34dc7c2f BB |
1360 | dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx) |
1361 | { | |
572e2857 | 1362 | dnode_t *dn; |
34dc7c2f BB |
1363 | uint64_t txg = tx->tx_txg; |
1364 | dbuf_dirty_record_t *dr, **drp; | |
1365 | ||
1366 | ASSERT(txg != 0); | |
428870ff | 1367 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
13fe0198 MA |
1368 | ASSERT0(db->db_level); |
1369 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
34dc7c2f | 1370 | |
34dc7c2f BB |
1371 | /* |
1372 | * If this buffer is not dirty, we're done. | |
1373 | */ | |
1374 | for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next) | |
1375 | if (dr->dr_txg <= txg) | |
1376 | break; | |
13fe0198 MA |
1377 | if (dr == NULL || dr->dr_txg < txg) |
1378 | return (B_FALSE); | |
34dc7c2f | 1379 | ASSERT(dr->dr_txg == txg); |
428870ff | 1380 | ASSERT(dr->dr_dbuf == db); |
34dc7c2f | 1381 | |
572e2857 BB |
1382 | DB_DNODE_ENTER(db); |
1383 | dn = DB_DNODE(db); | |
1384 | ||
34dc7c2f BB |
1385 | dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); |
1386 | ||
1387 | ASSERT(db->db.db_size != 0); | |
1388 | ||
e8b96c60 MA |
1389 | /* |
1390 | * Any space we accounted for in dp_dirty_* will be cleaned up by | |
1391 | * dsl_pool_sync(). This is relatively rare so the discrepancy | |
1392 | * is not a big deal. | |
1393 | */ | |
34dc7c2f BB |
1394 | |
1395 | *drp = dr->dr_next; | |
1396 | ||
ef3c1dea GR |
1397 | /* |
1398 | * Note that there are three places in dbuf_dirty() | |
1399 | * where this dirty record may be put on a list. | |
1400 | * Make sure to do a list_remove corresponding to | |
1401 | * every one of those list_insert calls. | |
1402 | */ | |
34dc7c2f BB |
1403 | if (dr->dr_parent) { |
1404 | mutex_enter(&dr->dr_parent->dt.di.dr_mtx); | |
1405 | list_remove(&dr->dr_parent->dt.di.dr_children, dr); | |
1406 | mutex_exit(&dr->dr_parent->dt.di.dr_mtx); | |
ef3c1dea GR |
1407 | } else if (db->db_blkid == DMU_SPILL_BLKID || |
1408 | db->db_level+1 == dn->dn_nlevels) { | |
b128c09f | 1409 | ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf); |
34dc7c2f BB |
1410 | mutex_enter(&dn->dn_mtx); |
1411 | list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr); | |
1412 | mutex_exit(&dn->dn_mtx); | |
1413 | } | |
572e2857 | 1414 | DB_DNODE_EXIT(db); |
34dc7c2f | 1415 | |
13fe0198 MA |
1416 | if (db->db_state != DB_NOFILL) { |
1417 | dbuf_unoverride(dr); | |
34dc7c2f | 1418 | |
34dc7c2f | 1419 | ASSERT(db->db_buf != NULL); |
13fe0198 MA |
1420 | ASSERT(dr->dt.dl.dr_data != NULL); |
1421 | if (dr->dt.dl.dr_data != db->db_buf) | |
1422 | VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db)); | |
34dc7c2f | 1423 | } |
58c4aa00 JL |
1424 | |
1425 | if (db->db_level != 0) { | |
1426 | mutex_destroy(&dr->dt.di.dr_mtx); | |
1427 | list_destroy(&dr->dt.di.dr_children); | |
1428 | } | |
1429 | ||
34dc7c2f BB |
1430 | kmem_free(dr, sizeof (dbuf_dirty_record_t)); |
1431 | ||
1432 | ASSERT(db->db_dirtycnt > 0); | |
1433 | db->db_dirtycnt -= 1; | |
1434 | ||
1435 | if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) { | |
1436 | arc_buf_t *buf = db->db_buf; | |
1437 | ||
428870ff | 1438 | ASSERT(db->db_state == DB_NOFILL || arc_released(buf)); |
34dc7c2f | 1439 | dbuf_set_data(db, NULL); |
13fe0198 | 1440 | VERIFY(arc_buf_remove_ref(buf, db)); |
34dc7c2f | 1441 | dbuf_evict(db); |
13fe0198 | 1442 | return (B_TRUE); |
34dc7c2f BB |
1443 | } |
1444 | ||
13fe0198 | 1445 | return (B_FALSE); |
34dc7c2f BB |
1446 | } |
1447 | ||
34dc7c2f | 1448 | void |
b0bc7a84 | 1449 | dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx) |
34dc7c2f | 1450 | { |
b0bc7a84 | 1451 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; |
34dc7c2f BB |
1452 | int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH; |
1453 | ||
1454 | ASSERT(tx->tx_txg != 0); | |
1455 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
1456 | ||
572e2857 BB |
1457 | DB_DNODE_ENTER(db); |
1458 | if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock)) | |
34dc7c2f | 1459 | rf |= DB_RF_HAVESTRUCT; |
572e2857 | 1460 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
1461 | (void) dbuf_read(db, NULL, rf); |
1462 | (void) dbuf_dirty(db, tx); | |
1463 | } | |
1464 | ||
b128c09f BB |
1465 | void |
1466 | dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) | |
1467 | { | |
1468 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
1469 | ||
1470 | db->db_state = DB_NOFILL; | |
1471 | ||
1472 | dmu_buf_will_fill(db_fake, tx); | |
1473 | } | |
1474 | ||
34dc7c2f BB |
1475 | void |
1476 | dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) | |
1477 | { | |
1478 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
1479 | ||
428870ff | 1480 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f BB |
1481 | ASSERT(tx->tx_txg != 0); |
1482 | ASSERT(db->db_level == 0); | |
1483 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
1484 | ||
1485 | ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT || | |
1486 | dmu_tx_private_ok(tx)); | |
1487 | ||
1488 | dbuf_noread(db); | |
1489 | (void) dbuf_dirty(db, tx); | |
1490 | } | |
1491 | ||
1492 | #pragma weak dmu_buf_fill_done = dbuf_fill_done | |
1493 | /* ARGSUSED */ | |
1494 | void | |
1495 | dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx) | |
1496 | { | |
1497 | mutex_enter(&db->db_mtx); | |
1498 | DBUF_VERIFY(db); | |
1499 | ||
1500 | if (db->db_state == DB_FILL) { | |
1501 | if (db->db_level == 0 && db->db_freed_in_flight) { | |
428870ff | 1502 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f BB |
1503 | /* we were freed while filling */ |
1504 | /* XXX dbuf_undirty? */ | |
1505 | bzero(db->db.db_data, db->db.db_size); | |
1506 | db->db_freed_in_flight = FALSE; | |
1507 | } | |
1508 | db->db_state = DB_CACHED; | |
1509 | cv_broadcast(&db->db_changed); | |
1510 | } | |
1511 | mutex_exit(&db->db_mtx); | |
1512 | } | |
1513 | ||
9b67f605 MA |
1514 | void |
1515 | dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data, | |
1516 | bp_embedded_type_t etype, enum zio_compress comp, | |
1517 | int uncompressed_size, int compressed_size, int byteorder, | |
1518 | dmu_tx_t *tx) | |
1519 | { | |
1520 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; | |
1521 | struct dirty_leaf *dl; | |
1522 | dmu_object_type_t type; | |
1523 | ||
1524 | DB_DNODE_ENTER(db); | |
1525 | type = DB_DNODE(db)->dn_type; | |
1526 | DB_DNODE_EXIT(db); | |
1527 | ||
1528 | ASSERT0(db->db_level); | |
1529 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); | |
1530 | ||
1531 | dmu_buf_will_not_fill(dbuf, tx); | |
1532 | ||
1533 | ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); | |
1534 | dl = &db->db_last_dirty->dt.dl; | |
1535 | encode_embedded_bp_compressed(&dl->dr_overridden_by, | |
1536 | data, comp, uncompressed_size, compressed_size); | |
1537 | BPE_SET_ETYPE(&dl->dr_overridden_by, etype); | |
1538 | BP_SET_TYPE(&dl->dr_overridden_by, type); | |
1539 | BP_SET_LEVEL(&dl->dr_overridden_by, 0); | |
1540 | BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder); | |
1541 | ||
1542 | dl->dr_override_state = DR_OVERRIDDEN; | |
1543 | dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg; | |
1544 | } | |
1545 | ||
9babb374 BB |
1546 | /* |
1547 | * Directly assign a provided arc buf to a given dbuf if it's not referenced | |
1548 | * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf. | |
1549 | */ | |
1550 | void | |
1551 | dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx) | |
1552 | { | |
1553 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
428870ff | 1554 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
9babb374 BB |
1555 | ASSERT(db->db_level == 0); |
1556 | ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA); | |
1557 | ASSERT(buf != NULL); | |
1558 | ASSERT(arc_buf_size(buf) == db->db.db_size); | |
1559 | ASSERT(tx->tx_txg != 0); | |
1560 | ||
1561 | arc_return_buf(buf, db); | |
1562 | ASSERT(arc_released(buf)); | |
1563 | ||
1564 | mutex_enter(&db->db_mtx); | |
1565 | ||
1566 | while (db->db_state == DB_READ || db->db_state == DB_FILL) | |
1567 | cv_wait(&db->db_changed, &db->db_mtx); | |
1568 | ||
1569 | ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED); | |
1570 | ||
1571 | if (db->db_state == DB_CACHED && | |
1572 | refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) { | |
1573 | mutex_exit(&db->db_mtx); | |
1574 | (void) dbuf_dirty(db, tx); | |
1575 | bcopy(buf->b_data, db->db.db_data, db->db.db_size); | |
13fe0198 | 1576 | VERIFY(arc_buf_remove_ref(buf, db)); |
428870ff | 1577 | xuio_stat_wbuf_copied(); |
9babb374 BB |
1578 | return; |
1579 | } | |
1580 | ||
428870ff | 1581 | xuio_stat_wbuf_nocopy(); |
9babb374 BB |
1582 | if (db->db_state == DB_CACHED) { |
1583 | dbuf_dirty_record_t *dr = db->db_last_dirty; | |
1584 | ||
1585 | ASSERT(db->db_buf != NULL); | |
1586 | if (dr != NULL && dr->dr_txg == tx->tx_txg) { | |
1587 | ASSERT(dr->dt.dl.dr_data == db->db_buf); | |
1588 | if (!arc_released(db->db_buf)) { | |
1589 | ASSERT(dr->dt.dl.dr_override_state == | |
1590 | DR_OVERRIDDEN); | |
1591 | arc_release(db->db_buf, db); | |
1592 | } | |
1593 | dr->dt.dl.dr_data = buf; | |
13fe0198 | 1594 | VERIFY(arc_buf_remove_ref(db->db_buf, db)); |
9babb374 BB |
1595 | } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) { |
1596 | arc_release(db->db_buf, db); | |
13fe0198 | 1597 | VERIFY(arc_buf_remove_ref(db->db_buf, db)); |
9babb374 BB |
1598 | } |
1599 | db->db_buf = NULL; | |
1600 | } | |
1601 | ASSERT(db->db_buf == NULL); | |
1602 | dbuf_set_data(db, buf); | |
1603 | db->db_state = DB_FILL; | |
1604 | mutex_exit(&db->db_mtx); | |
1605 | (void) dbuf_dirty(db, tx); | |
b0bc7a84 | 1606 | dmu_buf_fill_done(&db->db, tx); |
9babb374 BB |
1607 | } |
1608 | ||
34dc7c2f BB |
1609 | /* |
1610 | * "Clear" the contents of this dbuf. This will mark the dbuf | |
e8b96c60 | 1611 | * EVICTING and clear *most* of its references. Unfortunately, |
34dc7c2f BB |
1612 | * when we are not holding the dn_dbufs_mtx, we can't clear the |
1613 | * entry in the dn_dbufs list. We have to wait until dbuf_destroy() | |
1614 | * in this case. For callers from the DMU we will usually see: | |
bd089c54 | 1615 | * dbuf_clear()->arc_clear_callback()->dbuf_do_evict()->dbuf_destroy() |
34dc7c2f | 1616 | * For the arc callback, we will usually see: |
572e2857 | 1617 | * dbuf_do_evict()->dbuf_clear();dbuf_destroy() |
34dc7c2f | 1618 | * Sometimes, though, we will get a mix of these two: |
bd089c54 | 1619 | * DMU: dbuf_clear()->arc_clear_callback() |
34dc7c2f | 1620 | * ARC: dbuf_do_evict()->dbuf_destroy() |
bd089c54 MA |
1621 | * |
1622 | * This routine will dissociate the dbuf from the arc, by calling | |
1623 | * arc_clear_callback(), but will not evict the data from the ARC. | |
34dc7c2f BB |
1624 | */ |
1625 | void | |
1626 | dbuf_clear(dmu_buf_impl_t *db) | |
1627 | { | |
572e2857 | 1628 | dnode_t *dn; |
34dc7c2f | 1629 | dmu_buf_impl_t *parent = db->db_parent; |
572e2857 | 1630 | dmu_buf_impl_t *dndb; |
bd089c54 | 1631 | boolean_t dbuf_gone = B_FALSE; |
34dc7c2f BB |
1632 | |
1633 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
1634 | ASSERT(refcount_is_zero(&db->db_holds)); | |
1635 | ||
1636 | dbuf_evict_user(db); | |
1637 | ||
1638 | if (db->db_state == DB_CACHED) { | |
1639 | ASSERT(db->db.db_data != NULL); | |
428870ff | 1640 | if (db->db_blkid == DMU_BONUS_BLKID) { |
34dc7c2f | 1641 | zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN); |
d164b209 | 1642 | arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); |
34dc7c2f BB |
1643 | } |
1644 | db->db.db_data = NULL; | |
1645 | db->db_state = DB_UNCACHED; | |
1646 | } | |
1647 | ||
b128c09f | 1648 | ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL); |
34dc7c2f BB |
1649 | ASSERT(db->db_data_pending == NULL); |
1650 | ||
1651 | db->db_state = DB_EVICTING; | |
1652 | db->db_blkptr = NULL; | |
1653 | ||
572e2857 BB |
1654 | DB_DNODE_ENTER(db); |
1655 | dn = DB_DNODE(db); | |
1656 | dndb = dn->dn_dbuf; | |
428870ff | 1657 | if (db->db_blkid != DMU_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) { |
8951cb8d | 1658 | avl_remove(&dn->dn_dbufs, db); |
73ad4a9f | 1659 | atomic_dec_32(&dn->dn_dbufs_count); |
572e2857 BB |
1660 | membar_producer(); |
1661 | DB_DNODE_EXIT(db); | |
1662 | /* | |
1663 | * Decrementing the dbuf count means that the hold corresponding | |
1664 | * to the removed dbuf is no longer discounted in dnode_move(), | |
1665 | * so the dnode cannot be moved until after we release the hold. | |
1666 | * The membar_producer() ensures visibility of the decremented | |
1667 | * value in dnode_move(), since DB_DNODE_EXIT doesn't actually | |
1668 | * release any lock. | |
1669 | */ | |
34dc7c2f | 1670 | dnode_rele(dn, db); |
572e2857 BB |
1671 | db->db_dnode_handle = NULL; |
1672 | } else { | |
1673 | DB_DNODE_EXIT(db); | |
34dc7c2f BB |
1674 | } |
1675 | ||
1676 | if (db->db_buf) | |
bd089c54 | 1677 | dbuf_gone = arc_clear_callback(db->db_buf); |
34dc7c2f BB |
1678 | |
1679 | if (!dbuf_gone) | |
1680 | mutex_exit(&db->db_mtx); | |
1681 | ||
1682 | /* | |
572e2857 | 1683 | * If this dbuf is referenced from an indirect dbuf, |
34dc7c2f BB |
1684 | * decrement the ref count on the indirect dbuf. |
1685 | */ | |
1686 | if (parent && parent != dndb) | |
1687 | dbuf_rele(parent, db); | |
1688 | } | |
1689 | ||
bf701a83 BB |
1690 | __attribute__((always_inline)) |
1691 | static inline int | |
34dc7c2f | 1692 | dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse, |
fc5bb51f | 1693 | dmu_buf_impl_t **parentp, blkptr_t **bpp, struct dbuf_hold_impl_data *dh) |
34dc7c2f BB |
1694 | { |
1695 | int nlevels, epbs; | |
1696 | ||
1697 | *parentp = NULL; | |
1698 | *bpp = NULL; | |
1699 | ||
428870ff BB |
1700 | ASSERT(blkid != DMU_BONUS_BLKID); |
1701 | ||
1702 | if (blkid == DMU_SPILL_BLKID) { | |
1703 | mutex_enter(&dn->dn_mtx); | |
1704 | if (dn->dn_have_spill && | |
1705 | (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) | |
1706 | *bpp = &dn->dn_phys->dn_spill; | |
1707 | else | |
1708 | *bpp = NULL; | |
1709 | dbuf_add_ref(dn->dn_dbuf, NULL); | |
1710 | *parentp = dn->dn_dbuf; | |
1711 | mutex_exit(&dn->dn_mtx); | |
1712 | return (0); | |
1713 | } | |
34dc7c2f BB |
1714 | |
1715 | if (dn->dn_phys->dn_nlevels == 0) | |
1716 | nlevels = 1; | |
1717 | else | |
1718 | nlevels = dn->dn_phys->dn_nlevels; | |
1719 | ||
1720 | epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; | |
1721 | ||
1722 | ASSERT3U(level * epbs, <, 64); | |
1723 | ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); | |
1724 | if (level >= nlevels || | |
1725 | (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) { | |
1726 | /* the buffer has no parent yet */ | |
2e528b49 | 1727 | return (SET_ERROR(ENOENT)); |
34dc7c2f BB |
1728 | } else if (level < nlevels-1) { |
1729 | /* this block is referenced from an indirect block */ | |
fc5bb51f BB |
1730 | int err; |
1731 | if (dh == NULL) { | |
1732 | err = dbuf_hold_impl(dn, level+1, blkid >> epbs, | |
1733 | fail_sparse, NULL, parentp); | |
d1d7e268 | 1734 | } else { |
fc5bb51f BB |
1735 | __dbuf_hold_impl_init(dh + 1, dn, dh->dh_level + 1, |
1736 | blkid >> epbs, fail_sparse, NULL, | |
1737 | parentp, dh->dh_depth + 1); | |
1738 | err = __dbuf_hold_impl(dh + 1); | |
1739 | } | |
34dc7c2f BB |
1740 | if (err) |
1741 | return (err); | |
1742 | err = dbuf_read(*parentp, NULL, | |
1743 | (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL)); | |
1744 | if (err) { | |
1745 | dbuf_rele(*parentp, NULL); | |
1746 | *parentp = NULL; | |
1747 | return (err); | |
1748 | } | |
1749 | *bpp = ((blkptr_t *)(*parentp)->db.db_data) + | |
1750 | (blkid & ((1ULL << epbs) - 1)); | |
1751 | return (0); | |
1752 | } else { | |
1753 | /* the block is referenced from the dnode */ | |
1754 | ASSERT3U(level, ==, nlevels-1); | |
1755 | ASSERT(dn->dn_phys->dn_nblkptr == 0 || | |
1756 | blkid < dn->dn_phys->dn_nblkptr); | |
1757 | if (dn->dn_dbuf) { | |
1758 | dbuf_add_ref(dn->dn_dbuf, NULL); | |
1759 | *parentp = dn->dn_dbuf; | |
1760 | } | |
1761 | *bpp = &dn->dn_phys->dn_blkptr[blkid]; | |
1762 | return (0); | |
1763 | } | |
1764 | } | |
1765 | ||
1766 | static dmu_buf_impl_t * | |
1767 | dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid, | |
1768 | dmu_buf_impl_t *parent, blkptr_t *blkptr) | |
1769 | { | |
428870ff | 1770 | objset_t *os = dn->dn_objset; |
34dc7c2f BB |
1771 | dmu_buf_impl_t *db, *odb; |
1772 | ||
1773 | ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); | |
1774 | ASSERT(dn->dn_type != DMU_OT_NONE); | |
1775 | ||
79c76d5b | 1776 | db = kmem_cache_alloc(dbuf_cache, KM_SLEEP); |
34dc7c2f BB |
1777 | |
1778 | db->db_objset = os; | |
1779 | db->db.db_object = dn->dn_object; | |
1780 | db->db_level = level; | |
1781 | db->db_blkid = blkid; | |
1782 | db->db_last_dirty = NULL; | |
1783 | db->db_dirtycnt = 0; | |
572e2857 | 1784 | db->db_dnode_handle = dn->dn_handle; |
34dc7c2f BB |
1785 | db->db_parent = parent; |
1786 | db->db_blkptr = blkptr; | |
1787 | ||
1788 | db->db_user_ptr = NULL; | |
34dc7c2f BB |
1789 | db->db_evict_func = NULL; |
1790 | db->db_immediate_evict = 0; | |
1791 | db->db_freed_in_flight = 0; | |
1792 | ||
428870ff | 1793 | if (blkid == DMU_BONUS_BLKID) { |
34dc7c2f BB |
1794 | ASSERT3P(parent, ==, dn->dn_dbuf); |
1795 | db->db.db_size = DN_MAX_BONUSLEN - | |
1796 | (dn->dn_nblkptr-1) * sizeof (blkptr_t); | |
1797 | ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); | |
428870ff | 1798 | db->db.db_offset = DMU_BONUS_BLKID; |
34dc7c2f BB |
1799 | db->db_state = DB_UNCACHED; |
1800 | /* the bonus dbuf is not placed in the hash table */ | |
d164b209 | 1801 | arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); |
34dc7c2f | 1802 | return (db); |
428870ff BB |
1803 | } else if (blkid == DMU_SPILL_BLKID) { |
1804 | db->db.db_size = (blkptr != NULL) ? | |
1805 | BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE; | |
1806 | db->db.db_offset = 0; | |
34dc7c2f BB |
1807 | } else { |
1808 | int blocksize = | |
e8b96c60 | 1809 | db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz; |
34dc7c2f BB |
1810 | db->db.db_size = blocksize; |
1811 | db->db.db_offset = db->db_blkid * blocksize; | |
1812 | } | |
1813 | ||
1814 | /* | |
1815 | * Hold the dn_dbufs_mtx while we get the new dbuf | |
1816 | * in the hash table *and* added to the dbufs list. | |
1817 | * This prevents a possible deadlock with someone | |
1818 | * trying to look up this dbuf before its added to the | |
1819 | * dn_dbufs list. | |
1820 | */ | |
1821 | mutex_enter(&dn->dn_dbufs_mtx); | |
1822 | db->db_state = DB_EVICTING; | |
1823 | if ((odb = dbuf_hash_insert(db)) != NULL) { | |
1824 | /* someone else inserted it first */ | |
1825 | kmem_cache_free(dbuf_cache, db); | |
1826 | mutex_exit(&dn->dn_dbufs_mtx); | |
1827 | return (odb); | |
1828 | } | |
8951cb8d | 1829 | avl_add(&dn->dn_dbufs, db); |
b663a23d MA |
1830 | if (db->db_level == 0 && db->db_blkid >= |
1831 | dn->dn_unlisted_l0_blkid) | |
1832 | dn->dn_unlisted_l0_blkid = db->db_blkid + 1; | |
34dc7c2f BB |
1833 | db->db_state = DB_UNCACHED; |
1834 | mutex_exit(&dn->dn_dbufs_mtx); | |
d164b209 | 1835 | arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); |
34dc7c2f BB |
1836 | |
1837 | if (parent && parent != dn->dn_dbuf) | |
1838 | dbuf_add_ref(parent, db); | |
1839 | ||
1840 | ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || | |
1841 | refcount_count(&dn->dn_holds) > 0); | |
1842 | (void) refcount_add(&dn->dn_holds, db); | |
73ad4a9f | 1843 | atomic_inc_32(&dn->dn_dbufs_count); |
34dc7c2f BB |
1844 | |
1845 | dprintf_dbuf(db, "db=%p\n", db); | |
1846 | ||
1847 | return (db); | |
1848 | } | |
1849 | ||
1850 | static int | |
1851 | dbuf_do_evict(void *private) | |
1852 | { | |
bd089c54 | 1853 | dmu_buf_impl_t *db = private; |
34dc7c2f BB |
1854 | |
1855 | if (!MUTEX_HELD(&db->db_mtx)) | |
1856 | mutex_enter(&db->db_mtx); | |
1857 | ||
1858 | ASSERT(refcount_is_zero(&db->db_holds)); | |
1859 | ||
1860 | if (db->db_state != DB_EVICTING) { | |
1861 | ASSERT(db->db_state == DB_CACHED); | |
1862 | DBUF_VERIFY(db); | |
1863 | db->db_buf = NULL; | |
1864 | dbuf_evict(db); | |
1865 | } else { | |
1866 | mutex_exit(&db->db_mtx); | |
1867 | dbuf_destroy(db); | |
1868 | } | |
1869 | return (0); | |
1870 | } | |
1871 | ||
1872 | static void | |
1873 | dbuf_destroy(dmu_buf_impl_t *db) | |
1874 | { | |
1875 | ASSERT(refcount_is_zero(&db->db_holds)); | |
1876 | ||
428870ff | 1877 | if (db->db_blkid != DMU_BONUS_BLKID) { |
34dc7c2f BB |
1878 | /* |
1879 | * If this dbuf is still on the dn_dbufs list, | |
1880 | * remove it from that list. | |
1881 | */ | |
572e2857 BB |
1882 | if (db->db_dnode_handle != NULL) { |
1883 | dnode_t *dn; | |
34dc7c2f | 1884 | |
572e2857 BB |
1885 | DB_DNODE_ENTER(db); |
1886 | dn = DB_DNODE(db); | |
34dc7c2f | 1887 | mutex_enter(&dn->dn_dbufs_mtx); |
8951cb8d | 1888 | avl_remove(&dn->dn_dbufs, db); |
73ad4a9f | 1889 | atomic_dec_32(&dn->dn_dbufs_count); |
34dc7c2f | 1890 | mutex_exit(&dn->dn_dbufs_mtx); |
572e2857 BB |
1891 | DB_DNODE_EXIT(db); |
1892 | /* | |
1893 | * Decrementing the dbuf count means that the hold | |
1894 | * corresponding to the removed dbuf is no longer | |
1895 | * discounted in dnode_move(), so the dnode cannot be | |
1896 | * moved until after we release the hold. | |
1897 | */ | |
34dc7c2f | 1898 | dnode_rele(dn, db); |
572e2857 | 1899 | db->db_dnode_handle = NULL; |
34dc7c2f BB |
1900 | } |
1901 | dbuf_hash_remove(db); | |
1902 | } | |
1903 | db->db_parent = NULL; | |
1904 | db->db_buf = NULL; | |
1905 | ||
34dc7c2f BB |
1906 | ASSERT(db->db.db_data == NULL); |
1907 | ASSERT(db->db_hash_next == NULL); | |
1908 | ASSERT(db->db_blkptr == NULL); | |
1909 | ASSERT(db->db_data_pending == NULL); | |
1910 | ||
1911 | kmem_cache_free(dbuf_cache, db); | |
d164b209 | 1912 | arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); |
34dc7c2f BB |
1913 | } |
1914 | ||
1915 | void | |
e8b96c60 | 1916 | dbuf_prefetch(dnode_t *dn, uint64_t blkid, zio_priority_t prio) |
34dc7c2f BB |
1917 | { |
1918 | dmu_buf_impl_t *db = NULL; | |
1919 | blkptr_t *bp = NULL; | |
1920 | ||
428870ff | 1921 | ASSERT(blkid != DMU_BONUS_BLKID); |
34dc7c2f BB |
1922 | ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); |
1923 | ||
1924 | if (dnode_block_freed(dn, blkid)) | |
1925 | return; | |
1926 | ||
1927 | /* dbuf_find() returns with db_mtx held */ | |
c65aa5b2 | 1928 | if ((db = dbuf_find(dn, 0, blkid))) { |
572e2857 BB |
1929 | /* |
1930 | * This dbuf is already in the cache. We assume that | |
1931 | * it is already CACHED, or else about to be either | |
1932 | * read or filled. | |
1933 | */ | |
34dc7c2f | 1934 | mutex_exit(&db->db_mtx); |
572e2857 | 1935 | return; |
34dc7c2f BB |
1936 | } |
1937 | ||
fc5bb51f | 1938 | if (dbuf_findbp(dn, 0, blkid, TRUE, &db, &bp, NULL) == 0) { |
9b67f605 | 1939 | if (bp && !BP_IS_HOLE(bp) && !BP_IS_EMBEDDED(bp)) { |
428870ff | 1940 | dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset; |
34dc7c2f | 1941 | uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH; |
5dbd68a3 | 1942 | zbookmark_phys_t zb; |
428870ff BB |
1943 | |
1944 | SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET, | |
1945 | dn->dn_object, 0, blkid); | |
34dc7c2f | 1946 | |
294f6806 | 1947 | (void) arc_read(NULL, dn->dn_objset->os_spa, |
e8b96c60 | 1948 | bp, NULL, NULL, prio, |
34dc7c2f BB |
1949 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, |
1950 | &aflags, &zb); | |
1951 | } | |
1952 | if (db) | |
1953 | dbuf_rele(db, NULL); | |
1954 | } | |
1955 | } | |
1956 | ||
d1d7e268 | 1957 | #define DBUF_HOLD_IMPL_MAX_DEPTH 20 |
fc5bb51f | 1958 | |
34dc7c2f BB |
1959 | /* |
1960 | * Returns with db_holds incremented, and db_mtx not held. | |
1961 | * Note: dn_struct_rwlock must be held. | |
1962 | */ | |
fc5bb51f BB |
1963 | static int |
1964 | __dbuf_hold_impl(struct dbuf_hold_impl_data *dh) | |
34dc7c2f | 1965 | { |
fc5bb51f BB |
1966 | ASSERT3S(dh->dh_depth, <, DBUF_HOLD_IMPL_MAX_DEPTH); |
1967 | dh->dh_parent = NULL; | |
34dc7c2f | 1968 | |
fc5bb51f BB |
1969 | ASSERT(dh->dh_blkid != DMU_BONUS_BLKID); |
1970 | ASSERT(RW_LOCK_HELD(&dh->dh_dn->dn_struct_rwlock)); | |
1971 | ASSERT3U(dh->dh_dn->dn_nlevels, >, dh->dh_level); | |
34dc7c2f | 1972 | |
fc5bb51f | 1973 | *(dh->dh_dbp) = NULL; |
34dc7c2f BB |
1974 | top: |
1975 | /* dbuf_find() returns with db_mtx held */ | |
fc5bb51f BB |
1976 | dh->dh_db = dbuf_find(dh->dh_dn, dh->dh_level, dh->dh_blkid); |
1977 | ||
1978 | if (dh->dh_db == NULL) { | |
1979 | dh->dh_bp = NULL; | |
1980 | ||
1981 | ASSERT3P(dh->dh_parent, ==, NULL); | |
1982 | dh->dh_err = dbuf_findbp(dh->dh_dn, dh->dh_level, dh->dh_blkid, | |
1983 | dh->dh_fail_sparse, &dh->dh_parent, | |
1984 | &dh->dh_bp, dh); | |
1985 | if (dh->dh_fail_sparse) { | |
d1d7e268 MK |
1986 | if (dh->dh_err == 0 && |
1987 | dh->dh_bp && BP_IS_HOLE(dh->dh_bp)) | |
2e528b49 | 1988 | dh->dh_err = SET_ERROR(ENOENT); |
fc5bb51f BB |
1989 | if (dh->dh_err) { |
1990 | if (dh->dh_parent) | |
1991 | dbuf_rele(dh->dh_parent, NULL); | |
1992 | return (dh->dh_err); | |
34dc7c2f BB |
1993 | } |
1994 | } | |
fc5bb51f BB |
1995 | if (dh->dh_err && dh->dh_err != ENOENT) |
1996 | return (dh->dh_err); | |
1997 | dh->dh_db = dbuf_create(dh->dh_dn, dh->dh_level, dh->dh_blkid, | |
1998 | dh->dh_parent, dh->dh_bp); | |
34dc7c2f BB |
1999 | } |
2000 | ||
fc5bb51f BB |
2001 | if (dh->dh_db->db_buf && refcount_is_zero(&dh->dh_db->db_holds)) { |
2002 | arc_buf_add_ref(dh->dh_db->db_buf, dh->dh_db); | |
2003 | if (dh->dh_db->db_buf->b_data == NULL) { | |
2004 | dbuf_clear(dh->dh_db); | |
2005 | if (dh->dh_parent) { | |
2006 | dbuf_rele(dh->dh_parent, NULL); | |
2007 | dh->dh_parent = NULL; | |
34dc7c2f BB |
2008 | } |
2009 | goto top; | |
2010 | } | |
fc5bb51f | 2011 | ASSERT3P(dh->dh_db->db.db_data, ==, dh->dh_db->db_buf->b_data); |
34dc7c2f BB |
2012 | } |
2013 | ||
fc5bb51f | 2014 | ASSERT(dh->dh_db->db_buf == NULL || arc_referenced(dh->dh_db->db_buf)); |
34dc7c2f BB |
2015 | |
2016 | /* | |
2017 | * If this buffer is currently syncing out, and we are are | |
2018 | * still referencing it from db_data, we need to make a copy | |
2019 | * of it in case we decide we want to dirty it again in this txg. | |
2020 | */ | |
fc5bb51f BB |
2021 | if (dh->dh_db->db_level == 0 && |
2022 | dh->dh_db->db_blkid != DMU_BONUS_BLKID && | |
2023 | dh->dh_dn->dn_object != DMU_META_DNODE_OBJECT && | |
2024 | dh->dh_db->db_state == DB_CACHED && dh->dh_db->db_data_pending) { | |
2025 | dh->dh_dr = dh->dh_db->db_data_pending; | |
2026 | ||
2027 | if (dh->dh_dr->dt.dl.dr_data == dh->dh_db->db_buf) { | |
2028 | dh->dh_type = DBUF_GET_BUFC_TYPE(dh->dh_db); | |
2029 | ||
2030 | dbuf_set_data(dh->dh_db, | |
2031 | arc_buf_alloc(dh->dh_dn->dn_objset->os_spa, | |
2032 | dh->dh_db->db.db_size, dh->dh_db, dh->dh_type)); | |
2033 | bcopy(dh->dh_dr->dt.dl.dr_data->b_data, | |
2034 | dh->dh_db->db.db_data, dh->dh_db->db.db_size); | |
34dc7c2f BB |
2035 | } |
2036 | } | |
2037 | ||
fc5bb51f | 2038 | (void) refcount_add(&dh->dh_db->db_holds, dh->dh_tag); |
fc5bb51f BB |
2039 | DBUF_VERIFY(dh->dh_db); |
2040 | mutex_exit(&dh->dh_db->db_mtx); | |
34dc7c2f BB |
2041 | |
2042 | /* NOTE: we can't rele the parent until after we drop the db_mtx */ | |
fc5bb51f BB |
2043 | if (dh->dh_parent) |
2044 | dbuf_rele(dh->dh_parent, NULL); | |
34dc7c2f | 2045 | |
fc5bb51f BB |
2046 | ASSERT3P(DB_DNODE(dh->dh_db), ==, dh->dh_dn); |
2047 | ASSERT3U(dh->dh_db->db_blkid, ==, dh->dh_blkid); | |
2048 | ASSERT3U(dh->dh_db->db_level, ==, dh->dh_level); | |
2049 | *(dh->dh_dbp) = dh->dh_db; | |
34dc7c2f BB |
2050 | |
2051 | return (0); | |
2052 | } | |
2053 | ||
fc5bb51f BB |
2054 | /* |
2055 | * The following code preserves the recursive function dbuf_hold_impl() | |
2056 | * but moves the local variables AND function arguments to the heap to | |
2057 | * minimize the stack frame size. Enough space is initially allocated | |
2058 | * on the stack for 20 levels of recursion. | |
2059 | */ | |
2060 | int | |
2061 | dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse, | |
2062 | void *tag, dmu_buf_impl_t **dbp) | |
2063 | { | |
2064 | struct dbuf_hold_impl_data *dh; | |
2065 | int error; | |
2066 | ||
d1d7e268 | 2067 | dh = kmem_zalloc(sizeof (struct dbuf_hold_impl_data) * |
79c76d5b | 2068 | DBUF_HOLD_IMPL_MAX_DEPTH, KM_SLEEP); |
fc5bb51f BB |
2069 | __dbuf_hold_impl_init(dh, dn, level, blkid, fail_sparse, tag, dbp, 0); |
2070 | ||
2071 | error = __dbuf_hold_impl(dh); | |
2072 | ||
d1d7e268 | 2073 | kmem_free(dh, sizeof (struct dbuf_hold_impl_data) * |
fc5bb51f BB |
2074 | DBUF_HOLD_IMPL_MAX_DEPTH); |
2075 | ||
2076 | return (error); | |
2077 | } | |
2078 | ||
2079 | static void | |
2080 | __dbuf_hold_impl_init(struct dbuf_hold_impl_data *dh, | |
2081 | dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse, | |
2082 | void *tag, dmu_buf_impl_t **dbp, int depth) | |
2083 | { | |
2084 | dh->dh_dn = dn; | |
2085 | dh->dh_level = level; | |
2086 | dh->dh_blkid = blkid; | |
2087 | dh->dh_fail_sparse = fail_sparse; | |
2088 | dh->dh_tag = tag; | |
2089 | dh->dh_dbp = dbp; | |
2090 | dh->dh_depth = depth; | |
2091 | } | |
2092 | ||
34dc7c2f BB |
2093 | dmu_buf_impl_t * |
2094 | dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag) | |
2095 | { | |
2096 | dmu_buf_impl_t *db; | |
2097 | int err = dbuf_hold_impl(dn, 0, blkid, FALSE, tag, &db); | |
2098 | return (err ? NULL : db); | |
2099 | } | |
2100 | ||
2101 | dmu_buf_impl_t * | |
2102 | dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag) | |
2103 | { | |
2104 | dmu_buf_impl_t *db; | |
2105 | int err = dbuf_hold_impl(dn, level, blkid, FALSE, tag, &db); | |
2106 | return (err ? NULL : db); | |
2107 | } | |
2108 | ||
2109 | void | |
2110 | dbuf_create_bonus(dnode_t *dn) | |
2111 | { | |
2112 | ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); | |
2113 | ||
2114 | ASSERT(dn->dn_bonus == NULL); | |
428870ff BB |
2115 | dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL); |
2116 | } | |
2117 | ||
2118 | int | |
2119 | dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx) | |
2120 | { | |
2121 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
572e2857 BB |
2122 | dnode_t *dn; |
2123 | ||
428870ff | 2124 | if (db->db_blkid != DMU_SPILL_BLKID) |
2e528b49 | 2125 | return (SET_ERROR(ENOTSUP)); |
428870ff BB |
2126 | if (blksz == 0) |
2127 | blksz = SPA_MINBLOCKSIZE; | |
2128 | if (blksz > SPA_MAXBLOCKSIZE) | |
2129 | blksz = SPA_MAXBLOCKSIZE; | |
2130 | else | |
2131 | blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE); | |
2132 | ||
572e2857 BB |
2133 | DB_DNODE_ENTER(db); |
2134 | dn = DB_DNODE(db); | |
2135 | rw_enter(&dn->dn_struct_rwlock, RW_WRITER); | |
428870ff | 2136 | dbuf_new_size(db, blksz, tx); |
572e2857 BB |
2137 | rw_exit(&dn->dn_struct_rwlock); |
2138 | DB_DNODE_EXIT(db); | |
428870ff BB |
2139 | |
2140 | return (0); | |
2141 | } | |
2142 | ||
2143 | void | |
2144 | dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx) | |
2145 | { | |
2146 | dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx); | |
34dc7c2f BB |
2147 | } |
2148 | ||
2149 | #pragma weak dmu_buf_add_ref = dbuf_add_ref | |
2150 | void | |
2151 | dbuf_add_ref(dmu_buf_impl_t *db, void *tag) | |
2152 | { | |
1fde1e37 | 2153 | VERIFY(refcount_add(&db->db_holds, tag) > 1); |
34dc7c2f BB |
2154 | } |
2155 | ||
572e2857 BB |
2156 | /* |
2157 | * If you call dbuf_rele() you had better not be referencing the dnode handle | |
2158 | * unless you have some other direct or indirect hold on the dnode. (An indirect | |
2159 | * hold is a hold on one of the dnode's dbufs, including the bonus buffer.) | |
2160 | * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the | |
2161 | * dnode's parent dbuf evicting its dnode handles. | |
2162 | */ | |
34dc7c2f BB |
2163 | void |
2164 | dbuf_rele(dmu_buf_impl_t *db, void *tag) | |
428870ff BB |
2165 | { |
2166 | mutex_enter(&db->db_mtx); | |
2167 | dbuf_rele_and_unlock(db, tag); | |
2168 | } | |
2169 | ||
b0bc7a84 MG |
2170 | void |
2171 | dmu_buf_rele(dmu_buf_t *db, void *tag) | |
2172 | { | |
2173 | dbuf_rele((dmu_buf_impl_t *)db, tag); | |
2174 | } | |
2175 | ||
428870ff BB |
2176 | /* |
2177 | * dbuf_rele() for an already-locked dbuf. This is necessary to allow | |
2178 | * db_dirtycnt and db_holds to be updated atomically. | |
2179 | */ | |
2180 | void | |
2181 | dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag) | |
34dc7c2f BB |
2182 | { |
2183 | int64_t holds; | |
2184 | ||
428870ff | 2185 | ASSERT(MUTEX_HELD(&db->db_mtx)); |
34dc7c2f BB |
2186 | DBUF_VERIFY(db); |
2187 | ||
572e2857 BB |
2188 | /* |
2189 | * Remove the reference to the dbuf before removing its hold on the | |
2190 | * dnode so we can guarantee in dnode_move() that a referenced bonus | |
2191 | * buffer has a corresponding dnode hold. | |
2192 | */ | |
34dc7c2f BB |
2193 | holds = refcount_remove(&db->db_holds, tag); |
2194 | ASSERT(holds >= 0); | |
2195 | ||
2196 | /* | |
2197 | * We can't freeze indirects if there is a possibility that they | |
2198 | * may be modified in the current syncing context. | |
2199 | */ | |
2200 | if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) | |
2201 | arc_buf_freeze(db->db_buf); | |
2202 | ||
2203 | if (holds == db->db_dirtycnt && | |
2204 | db->db_level == 0 && db->db_immediate_evict) | |
2205 | dbuf_evict_user(db); | |
2206 | ||
2207 | if (holds == 0) { | |
428870ff | 2208 | if (db->db_blkid == DMU_BONUS_BLKID) { |
4c7b7eed | 2209 | dnode_t *dn; |
572e2857 BB |
2210 | |
2211 | /* | |
4c7b7eed JG |
2212 | * If the dnode moves here, we cannot cross this |
2213 | * barrier until the move completes. | |
572e2857 BB |
2214 | */ |
2215 | DB_DNODE_ENTER(db); | |
4c7b7eed JG |
2216 | |
2217 | dn = DB_DNODE(db); | |
2218 | atomic_dec_32(&dn->dn_dbufs_count); | |
2219 | ||
2220 | /* | |
2221 | * Decrementing the dbuf count means that the bonus | |
2222 | * buffer's dnode hold is no longer discounted in | |
2223 | * dnode_move(). The dnode cannot move until after | |
2224 | * the dnode_rele_and_unlock() below. | |
2225 | */ | |
572e2857 | 2226 | DB_DNODE_EXIT(db); |
4c7b7eed JG |
2227 | |
2228 | /* | |
2229 | * Do not reference db after its lock is dropped. | |
2230 | * Another thread may evict it. | |
2231 | */ | |
2232 | mutex_exit(&db->db_mtx); | |
2233 | ||
572e2857 | 2234 | /* |
4c7b7eed JG |
2235 | * If the dnode has been freed, evict the bonus |
2236 | * buffer immediately. The data in the bonus | |
2237 | * buffer is no longer relevant and this prevents | |
2238 | * a stale bonus buffer from being associated | |
2239 | * with this dnode_t should the dnode_t be reused | |
2240 | * prior to being destroyed. | |
572e2857 | 2241 | */ |
4c7b7eed JG |
2242 | mutex_enter(&dn->dn_mtx); |
2243 | if (dn->dn_type == DMU_OT_NONE || | |
2244 | dn->dn_free_txg != 0) { | |
2245 | /* | |
2246 | * Drop dn_mtx. It is a leaf lock and | |
2247 | * cannot be held when dnode_evict_bonus() | |
2248 | * acquires other locks in order to | |
2249 | * perform the eviction. | |
2250 | * | |
2251 | * Freed dnodes cannot be reused until the | |
2252 | * last hold is released. Since this bonus | |
2253 | * buffer has a hold, the dnode will remain | |
2254 | * in the free state, even without dn_mtx | |
2255 | * held, until the dnode_rele_and_unlock() | |
2256 | * below. | |
2257 | */ | |
2258 | mutex_exit(&dn->dn_mtx); | |
2259 | dnode_evict_bonus(dn); | |
2260 | mutex_enter(&dn->dn_mtx); | |
2261 | } | |
2262 | dnode_rele_and_unlock(dn, db); | |
34dc7c2f BB |
2263 | } else if (db->db_buf == NULL) { |
2264 | /* | |
2265 | * This is a special case: we never associated this | |
2266 | * dbuf with any data allocated from the ARC. | |
2267 | */ | |
b128c09f BB |
2268 | ASSERT(db->db_state == DB_UNCACHED || |
2269 | db->db_state == DB_NOFILL); | |
34dc7c2f BB |
2270 | dbuf_evict(db); |
2271 | } else if (arc_released(db->db_buf)) { | |
2272 | arc_buf_t *buf = db->db_buf; | |
2273 | /* | |
2274 | * This dbuf has anonymous data associated with it. | |
2275 | */ | |
2276 | dbuf_set_data(db, NULL); | |
13fe0198 | 2277 | VERIFY(arc_buf_remove_ref(buf, db)); |
34dc7c2f BB |
2278 | dbuf_evict(db); |
2279 | } else { | |
13fe0198 | 2280 | VERIFY(!arc_buf_remove_ref(db->db_buf, db)); |
1eb5bfa3 GW |
2281 | |
2282 | /* | |
2283 | * A dbuf will be eligible for eviction if either the | |
2284 | * 'primarycache' property is set or a duplicate | |
2285 | * copy of this buffer is already cached in the arc. | |
2286 | * | |
2287 | * In the case of the 'primarycache' a buffer | |
2288 | * is considered for eviction if it matches the | |
2289 | * criteria set in the property. | |
2290 | * | |
2291 | * To decide if our buffer is considered a | |
2292 | * duplicate, we must call into the arc to determine | |
2293 | * if multiple buffers are referencing the same | |
2294 | * block on-disk. If so, then we simply evict | |
2295 | * ourselves. | |
2296 | */ | |
bd089c54 MA |
2297 | if (!DBUF_IS_CACHEABLE(db)) { |
2298 | if (db->db_blkptr != NULL && | |
2299 | !BP_IS_HOLE(db->db_blkptr) && | |
2300 | !BP_IS_EMBEDDED(db->db_blkptr)) { | |
2301 | spa_t *spa = | |
2302 | dmu_objset_spa(db->db_objset); | |
2303 | blkptr_t bp = *db->db_blkptr; | |
2304 | dbuf_clear(db); | |
2305 | arc_freed(spa, &bp); | |
2306 | } else { | |
2307 | dbuf_clear(db); | |
2308 | } | |
2309 | } else if (arc_buf_eviction_needed(db->db_buf)) { | |
b128c09f | 2310 | dbuf_clear(db); |
bd089c54 | 2311 | } else { |
b128c09f | 2312 | mutex_exit(&db->db_mtx); |
bd089c54 | 2313 | } |
34dc7c2f BB |
2314 | } |
2315 | } else { | |
2316 | mutex_exit(&db->db_mtx); | |
2317 | } | |
2318 | } | |
2319 | ||
2320 | #pragma weak dmu_buf_refcount = dbuf_refcount | |
2321 | uint64_t | |
2322 | dbuf_refcount(dmu_buf_impl_t *db) | |
2323 | { | |
2324 | return (refcount_count(&db->db_holds)); | |
2325 | } | |
2326 | ||
2327 | void * | |
d683ddbb | 2328 | dmu_buf_set_user(dmu_buf_t *db_fake, void *user_ptr, |
34dc7c2f BB |
2329 | dmu_buf_evict_func_t *evict_func) |
2330 | { | |
d683ddbb | 2331 | return (dmu_buf_update_user(db_fake, NULL, user_ptr, evict_func)); |
34dc7c2f BB |
2332 | } |
2333 | ||
2334 | void * | |
d683ddbb | 2335 | dmu_buf_set_user_ie(dmu_buf_t *db_fake, void *user_ptr, |
34dc7c2f BB |
2336 | dmu_buf_evict_func_t *evict_func) |
2337 | { | |
2338 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
2339 | ||
2340 | db->db_immediate_evict = TRUE; | |
d683ddbb | 2341 | return (dmu_buf_update_user(db_fake, NULL, user_ptr, evict_func)); |
34dc7c2f BB |
2342 | } |
2343 | ||
2344 | void * | |
2345 | dmu_buf_update_user(dmu_buf_t *db_fake, void *old_user_ptr, void *user_ptr, | |
d683ddbb | 2346 | dmu_buf_evict_func_t *evict_func) |
34dc7c2f BB |
2347 | { |
2348 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
2349 | ASSERT(db->db_level == 0); | |
2350 | ||
2351 | ASSERT((user_ptr == NULL) == (evict_func == NULL)); | |
2352 | ||
2353 | mutex_enter(&db->db_mtx); | |
2354 | ||
2355 | if (db->db_user_ptr == old_user_ptr) { | |
2356 | db->db_user_ptr = user_ptr; | |
34dc7c2f | 2357 | db->db_evict_func = evict_func; |
34dc7c2f BB |
2358 | } else { |
2359 | old_user_ptr = db->db_user_ptr; | |
2360 | } | |
2361 | ||
2362 | mutex_exit(&db->db_mtx); | |
2363 | return (old_user_ptr); | |
2364 | } | |
2365 | ||
2366 | void * | |
2367 | dmu_buf_get_user(dmu_buf_t *db_fake) | |
2368 | { | |
2369 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
2370 | ASSERT(!refcount_is_zero(&db->db_holds)); | |
2371 | ||
2372 | return (db->db_user_ptr); | |
2373 | } | |
2374 | ||
9babb374 BB |
2375 | boolean_t |
2376 | dmu_buf_freeable(dmu_buf_t *dbuf) | |
2377 | { | |
2378 | boolean_t res = B_FALSE; | |
2379 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; | |
2380 | ||
2381 | if (db->db_blkptr) | |
2382 | res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset, | |
428870ff | 2383 | db->db_blkptr, db->db_blkptr->blk_birth); |
9babb374 BB |
2384 | |
2385 | return (res); | |
2386 | } | |
2387 | ||
03c6040b GW |
2388 | blkptr_t * |
2389 | dmu_buf_get_blkptr(dmu_buf_t *db) | |
2390 | { | |
2391 | dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; | |
2392 | return (dbi->db_blkptr); | |
2393 | } | |
2394 | ||
34dc7c2f BB |
2395 | static void |
2396 | dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db) | |
2397 | { | |
2398 | /* ASSERT(dmu_tx_is_syncing(tx) */ | |
2399 | ASSERT(MUTEX_HELD(&db->db_mtx)); | |
2400 | ||
2401 | if (db->db_blkptr != NULL) | |
2402 | return; | |
2403 | ||
428870ff BB |
2404 | if (db->db_blkid == DMU_SPILL_BLKID) { |
2405 | db->db_blkptr = &dn->dn_phys->dn_spill; | |
2406 | BP_ZERO(db->db_blkptr); | |
2407 | return; | |
2408 | } | |
34dc7c2f BB |
2409 | if (db->db_level == dn->dn_phys->dn_nlevels-1) { |
2410 | /* | |
2411 | * This buffer was allocated at a time when there was | |
2412 | * no available blkptrs from the dnode, or it was | |
2413 | * inappropriate to hook it in (i.e., nlevels mis-match). | |
2414 | */ | |
2415 | ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr); | |
2416 | ASSERT(db->db_parent == NULL); | |
2417 | db->db_parent = dn->dn_dbuf; | |
2418 | db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid]; | |
2419 | DBUF_VERIFY(db); | |
2420 | } else { | |
2421 | dmu_buf_impl_t *parent = db->db_parent; | |
2422 | int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; | |
2423 | ||
2424 | ASSERT(dn->dn_phys->dn_nlevels > 1); | |
2425 | if (parent == NULL) { | |
2426 | mutex_exit(&db->db_mtx); | |
2427 | rw_enter(&dn->dn_struct_rwlock, RW_READER); | |
2428 | (void) dbuf_hold_impl(dn, db->db_level+1, | |
2429 | db->db_blkid >> epbs, FALSE, db, &parent); | |
2430 | rw_exit(&dn->dn_struct_rwlock); | |
2431 | mutex_enter(&db->db_mtx); | |
2432 | db->db_parent = parent; | |
2433 | } | |
2434 | db->db_blkptr = (blkptr_t *)parent->db.db_data + | |
2435 | (db->db_blkid & ((1ULL << epbs) - 1)); | |
2436 | DBUF_VERIFY(db); | |
2437 | } | |
2438 | } | |
2439 | ||
d1d7e268 MK |
2440 | /* |
2441 | * dbuf_sync_indirect() is called recursively from dbuf_sync_list() so it | |
60948de1 BB |
2442 | * is critical the we not allow the compiler to inline this function in to |
2443 | * dbuf_sync_list() thereby drastically bloating the stack usage. | |
2444 | */ | |
2445 | noinline static void | |
34dc7c2f BB |
2446 | dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx) |
2447 | { | |
2448 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
572e2857 | 2449 | dnode_t *dn; |
34dc7c2f BB |
2450 | zio_t *zio; |
2451 | ||
2452 | ASSERT(dmu_tx_is_syncing(tx)); | |
2453 | ||
2454 | dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); | |
2455 | ||
2456 | mutex_enter(&db->db_mtx); | |
2457 | ||
2458 | ASSERT(db->db_level > 0); | |
2459 | DBUF_VERIFY(db); | |
2460 | ||
e49f1e20 | 2461 | /* Read the block if it hasn't been read yet. */ |
34dc7c2f BB |
2462 | if (db->db_buf == NULL) { |
2463 | mutex_exit(&db->db_mtx); | |
2464 | (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED); | |
2465 | mutex_enter(&db->db_mtx); | |
2466 | } | |
2467 | ASSERT3U(db->db_state, ==, DB_CACHED); | |
34dc7c2f BB |
2468 | ASSERT(db->db_buf != NULL); |
2469 | ||
572e2857 BB |
2470 | DB_DNODE_ENTER(db); |
2471 | dn = DB_DNODE(db); | |
e49f1e20 | 2472 | /* Indirect block size must match what the dnode thinks it is. */ |
572e2857 | 2473 | ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); |
34dc7c2f | 2474 | dbuf_check_blkptr(dn, db); |
572e2857 | 2475 | DB_DNODE_EXIT(db); |
34dc7c2f | 2476 | |
e49f1e20 | 2477 | /* Provide the pending dirty record to child dbufs */ |
34dc7c2f BB |
2478 | db->db_data_pending = dr; |
2479 | ||
34dc7c2f | 2480 | mutex_exit(&db->db_mtx); |
b128c09f | 2481 | dbuf_write(dr, db->db_buf, tx); |
34dc7c2f BB |
2482 | |
2483 | zio = dr->dr_zio; | |
2484 | mutex_enter(&dr->dt.di.dr_mtx); | |
2485 | dbuf_sync_list(&dr->dt.di.dr_children, tx); | |
2486 | ASSERT(list_head(&dr->dt.di.dr_children) == NULL); | |
2487 | mutex_exit(&dr->dt.di.dr_mtx); | |
2488 | zio_nowait(zio); | |
2489 | } | |
2490 | ||
d1d7e268 MK |
2491 | /* |
2492 | * dbuf_sync_leaf() is called recursively from dbuf_sync_list() so it is | |
60948de1 BB |
2493 | * critical the we not allow the compiler to inline this function in to |
2494 | * dbuf_sync_list() thereby drastically bloating the stack usage. | |
2495 | */ | |
2496 | noinline static void | |
34dc7c2f BB |
2497 | dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx) |
2498 | { | |
2499 | arc_buf_t **datap = &dr->dt.dl.dr_data; | |
2500 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
572e2857 BB |
2501 | dnode_t *dn; |
2502 | objset_t *os; | |
34dc7c2f | 2503 | uint64_t txg = tx->tx_txg; |
34dc7c2f BB |
2504 | |
2505 | ASSERT(dmu_tx_is_syncing(tx)); | |
2506 | ||
2507 | dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); | |
2508 | ||
2509 | mutex_enter(&db->db_mtx); | |
2510 | /* | |
2511 | * To be synced, we must be dirtied. But we | |
2512 | * might have been freed after the dirty. | |
2513 | */ | |
2514 | if (db->db_state == DB_UNCACHED) { | |
2515 | /* This buffer has been freed since it was dirtied */ | |
2516 | ASSERT(db->db.db_data == NULL); | |
2517 | } else if (db->db_state == DB_FILL) { | |
2518 | /* This buffer was freed and is now being re-filled */ | |
2519 | ASSERT(db->db.db_data != dr->dt.dl.dr_data); | |
2520 | } else { | |
b128c09f | 2521 | ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL); |
34dc7c2f BB |
2522 | } |
2523 | DBUF_VERIFY(db); | |
2524 | ||
572e2857 BB |
2525 | DB_DNODE_ENTER(db); |
2526 | dn = DB_DNODE(db); | |
2527 | ||
428870ff BB |
2528 | if (db->db_blkid == DMU_SPILL_BLKID) { |
2529 | mutex_enter(&dn->dn_mtx); | |
2530 | dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR; | |
2531 | mutex_exit(&dn->dn_mtx); | |
2532 | } | |
2533 | ||
34dc7c2f BB |
2534 | /* |
2535 | * If this is a bonus buffer, simply copy the bonus data into the | |
2536 | * dnode. It will be written out when the dnode is synced (and it | |
2537 | * will be synced, since it must have been dirty for dbuf_sync to | |
2538 | * be called). | |
2539 | */ | |
428870ff | 2540 | if (db->db_blkid == DMU_BONUS_BLKID) { |
34dc7c2f BB |
2541 | dbuf_dirty_record_t **drp; |
2542 | ||
2543 | ASSERT(*datap != NULL); | |
c99c9001 | 2544 | ASSERT0(db->db_level); |
34dc7c2f BB |
2545 | ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN); |
2546 | bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen); | |
572e2857 BB |
2547 | DB_DNODE_EXIT(db); |
2548 | ||
34dc7c2f BB |
2549 | if (*datap != db->db.db_data) { |
2550 | zio_buf_free(*datap, DN_MAX_BONUSLEN); | |
d164b209 | 2551 | arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); |
34dc7c2f BB |
2552 | } |
2553 | db->db_data_pending = NULL; | |
2554 | drp = &db->db_last_dirty; | |
2555 | while (*drp != dr) | |
2556 | drp = &(*drp)->dr_next; | |
2557 | ASSERT(dr->dr_next == NULL); | |
428870ff | 2558 | ASSERT(dr->dr_dbuf == db); |
34dc7c2f | 2559 | *drp = dr->dr_next; |
753972fc BB |
2560 | if (dr->dr_dbuf->db_level != 0) { |
2561 | mutex_destroy(&dr->dt.di.dr_mtx); | |
2562 | list_destroy(&dr->dt.di.dr_children); | |
2563 | } | |
34dc7c2f BB |
2564 | kmem_free(dr, sizeof (dbuf_dirty_record_t)); |
2565 | ASSERT(db->db_dirtycnt > 0); | |
2566 | db->db_dirtycnt -= 1; | |
428870ff | 2567 | dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg); |
34dc7c2f BB |
2568 | return; |
2569 | } | |
2570 | ||
572e2857 BB |
2571 | os = dn->dn_objset; |
2572 | ||
34dc7c2f BB |
2573 | /* |
2574 | * This function may have dropped the db_mtx lock allowing a dmu_sync | |
2575 | * operation to sneak in. As a result, we need to ensure that we | |
2576 | * don't check the dr_override_state until we have returned from | |
2577 | * dbuf_check_blkptr. | |
2578 | */ | |
2579 | dbuf_check_blkptr(dn, db); | |
2580 | ||
2581 | /* | |
572e2857 | 2582 | * If this buffer is in the middle of an immediate write, |
34dc7c2f BB |
2583 | * wait for the synchronous IO to complete. |
2584 | */ | |
2585 | while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) { | |
2586 | ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT); | |
2587 | cv_wait(&db->db_changed, &db->db_mtx); | |
2588 | ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN); | |
2589 | } | |
2590 | ||
9babb374 BB |
2591 | if (db->db_state != DB_NOFILL && |
2592 | dn->dn_object != DMU_META_DNODE_OBJECT && | |
2593 | refcount_count(&db->db_holds) > 1 && | |
428870ff | 2594 | dr->dt.dl.dr_override_state != DR_OVERRIDDEN && |
9babb374 BB |
2595 | *datap == db->db_buf) { |
2596 | /* | |
2597 | * If this buffer is currently "in use" (i.e., there | |
2598 | * are active holds and db_data still references it), | |
2599 | * then make a copy before we start the write so that | |
2600 | * any modifications from the open txg will not leak | |
2601 | * into this write. | |
2602 | * | |
2603 | * NOTE: this copy does not need to be made for | |
2604 | * objects only modified in the syncing context (e.g. | |
2605 | * DNONE_DNODE blocks). | |
2606 | */ | |
2607 | int blksz = arc_buf_size(*datap); | |
2608 | arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); | |
2609 | *datap = arc_buf_alloc(os->os_spa, blksz, db, type); | |
2610 | bcopy(db->db.db_data, (*datap)->b_data, blksz); | |
b128c09f | 2611 | } |
34dc7c2f BB |
2612 | db->db_data_pending = dr; |
2613 | ||
2614 | mutex_exit(&db->db_mtx); | |
2615 | ||
b128c09f | 2616 | dbuf_write(dr, *datap, tx); |
34dc7c2f BB |
2617 | |
2618 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
572e2857 | 2619 | if (dn->dn_object == DMU_META_DNODE_OBJECT) { |
34dc7c2f | 2620 | list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr); |
572e2857 BB |
2621 | DB_DNODE_EXIT(db); |
2622 | } else { | |
2623 | /* | |
2624 | * Although zio_nowait() does not "wait for an IO", it does | |
2625 | * initiate the IO. If this is an empty write it seems plausible | |
2626 | * that the IO could actually be completed before the nowait | |
2627 | * returns. We need to DB_DNODE_EXIT() first in case | |
2628 | * zio_nowait() invalidates the dbuf. | |
2629 | */ | |
2630 | DB_DNODE_EXIT(db); | |
34dc7c2f | 2631 | zio_nowait(dr->dr_zio); |
572e2857 | 2632 | } |
34dc7c2f BB |
2633 | } |
2634 | ||
2635 | void | |
2636 | dbuf_sync_list(list_t *list, dmu_tx_t *tx) | |
2637 | { | |
2638 | dbuf_dirty_record_t *dr; | |
2639 | ||
c65aa5b2 | 2640 | while ((dr = list_head(list))) { |
34dc7c2f BB |
2641 | if (dr->dr_zio != NULL) { |
2642 | /* | |
2643 | * If we find an already initialized zio then we | |
2644 | * are processing the meta-dnode, and we have finished. | |
2645 | * The dbufs for all dnodes are put back on the list | |
2646 | * during processing, so that we can zio_wait() | |
2647 | * these IOs after initiating all child IOs. | |
2648 | */ | |
2649 | ASSERT3U(dr->dr_dbuf->db.db_object, ==, | |
2650 | DMU_META_DNODE_OBJECT); | |
2651 | break; | |
2652 | } | |
2653 | list_remove(list, dr); | |
2654 | if (dr->dr_dbuf->db_level > 0) | |
2655 | dbuf_sync_indirect(dr, tx); | |
2656 | else | |
2657 | dbuf_sync_leaf(dr, tx); | |
2658 | } | |
2659 | } | |
2660 | ||
34dc7c2f BB |
2661 | /* ARGSUSED */ |
2662 | static void | |
2663 | dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb) | |
2664 | { | |
2665 | dmu_buf_impl_t *db = vdb; | |
572e2857 | 2666 | dnode_t *dn; |
b128c09f | 2667 | blkptr_t *bp = zio->io_bp; |
34dc7c2f | 2668 | blkptr_t *bp_orig = &zio->io_bp_orig; |
428870ff BB |
2669 | spa_t *spa = zio->io_spa; |
2670 | int64_t delta; | |
34dc7c2f | 2671 | uint64_t fill = 0; |
428870ff | 2672 | int i; |
34dc7c2f | 2673 | |
9b67f605 | 2674 | ASSERT3P(db->db_blkptr, ==, bp); |
b128c09f | 2675 | |
572e2857 BB |
2676 | DB_DNODE_ENTER(db); |
2677 | dn = DB_DNODE(db); | |
428870ff BB |
2678 | delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig); |
2679 | dnode_diduse_space(dn, delta - zio->io_prev_space_delta); | |
2680 | zio->io_prev_space_delta = delta; | |
34dc7c2f | 2681 | |
b0bc7a84 MG |
2682 | if (bp->blk_birth != 0) { |
2683 | ASSERT((db->db_blkid != DMU_SPILL_BLKID && | |
2684 | BP_GET_TYPE(bp) == dn->dn_type) || | |
2685 | (db->db_blkid == DMU_SPILL_BLKID && | |
9b67f605 MA |
2686 | BP_GET_TYPE(bp) == dn->dn_bonustype) || |
2687 | BP_IS_EMBEDDED(bp)); | |
b0bc7a84 | 2688 | ASSERT(BP_GET_LEVEL(bp) == db->db_level); |
34dc7c2f BB |
2689 | } |
2690 | ||
2691 | mutex_enter(&db->db_mtx); | |
2692 | ||
428870ff BB |
2693 | #ifdef ZFS_DEBUG |
2694 | if (db->db_blkid == DMU_SPILL_BLKID) { | |
428870ff BB |
2695 | ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); |
2696 | ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && | |
2697 | db->db_blkptr == &dn->dn_phys->dn_spill); | |
2698 | } | |
2699 | #endif | |
2700 | ||
34dc7c2f BB |
2701 | if (db->db_level == 0) { |
2702 | mutex_enter(&dn->dn_mtx); | |
428870ff BB |
2703 | if (db->db_blkid > dn->dn_phys->dn_maxblkid && |
2704 | db->db_blkid != DMU_SPILL_BLKID) | |
34dc7c2f BB |
2705 | dn->dn_phys->dn_maxblkid = db->db_blkid; |
2706 | mutex_exit(&dn->dn_mtx); | |
2707 | ||
2708 | if (dn->dn_type == DMU_OT_DNODE) { | |
2709 | dnode_phys_t *dnp = db->db.db_data; | |
2710 | for (i = db->db.db_size >> DNODE_SHIFT; i > 0; | |
2711 | i--, dnp++) { | |
2712 | if (dnp->dn_type != DMU_OT_NONE) | |
2713 | fill++; | |
2714 | } | |
2715 | } else { | |
b0bc7a84 MG |
2716 | if (BP_IS_HOLE(bp)) { |
2717 | fill = 0; | |
2718 | } else { | |
2719 | fill = 1; | |
2720 | } | |
34dc7c2f BB |
2721 | } |
2722 | } else { | |
b128c09f | 2723 | blkptr_t *ibp = db->db.db_data; |
34dc7c2f | 2724 | ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); |
b128c09f BB |
2725 | for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) { |
2726 | if (BP_IS_HOLE(ibp)) | |
34dc7c2f | 2727 | continue; |
9b67f605 | 2728 | fill += BP_GET_FILL(ibp); |
34dc7c2f BB |
2729 | } |
2730 | } | |
572e2857 | 2731 | DB_DNODE_EXIT(db); |
34dc7c2f | 2732 | |
9b67f605 MA |
2733 | if (!BP_IS_EMBEDDED(bp)) |
2734 | bp->blk_fill = fill; | |
34dc7c2f BB |
2735 | |
2736 | mutex_exit(&db->db_mtx); | |
34dc7c2f BB |
2737 | } |
2738 | ||
e8b96c60 MA |
2739 | /* |
2740 | * The SPA will call this callback several times for each zio - once | |
2741 | * for every physical child i/o (zio->io_phys_children times). This | |
2742 | * allows the DMU to monitor the progress of each logical i/o. For example, | |
2743 | * there may be 2 copies of an indirect block, or many fragments of a RAID-Z | |
2744 | * block. There may be a long delay before all copies/fragments are completed, | |
2745 | * so this callback allows us to retire dirty space gradually, as the physical | |
2746 | * i/os complete. | |
2747 | */ | |
2748 | /* ARGSUSED */ | |
2749 | static void | |
2750 | dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg) | |
2751 | { | |
2752 | dmu_buf_impl_t *db = arg; | |
2753 | objset_t *os = db->db_objset; | |
2754 | dsl_pool_t *dp = dmu_objset_pool(os); | |
2755 | dbuf_dirty_record_t *dr; | |
2756 | int delta = 0; | |
2757 | ||
2758 | dr = db->db_data_pending; | |
2759 | ASSERT3U(dr->dr_txg, ==, zio->io_txg); | |
2760 | ||
2761 | /* | |
2762 | * The callback will be called io_phys_children times. Retire one | |
2763 | * portion of our dirty space each time we are called. Any rounding | |
2764 | * error will be cleaned up by dsl_pool_sync()'s call to | |
2765 | * dsl_pool_undirty_space(). | |
2766 | */ | |
2767 | delta = dr->dr_accounted / zio->io_phys_children; | |
2768 | dsl_pool_undirty_space(dp, delta, zio->io_txg); | |
2769 | } | |
2770 | ||
34dc7c2f BB |
2771 | /* ARGSUSED */ |
2772 | static void | |
2773 | dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb) | |
2774 | { | |
2775 | dmu_buf_impl_t *db = vdb; | |
428870ff | 2776 | blkptr_t *bp_orig = &zio->io_bp_orig; |
b0bc7a84 MG |
2777 | blkptr_t *bp = db->db_blkptr; |
2778 | objset_t *os = db->db_objset; | |
2779 | dmu_tx_t *tx = os->os_synctx; | |
34dc7c2f BB |
2780 | dbuf_dirty_record_t **drp, *dr; |
2781 | ||
c99c9001 | 2782 | ASSERT0(zio->io_error); |
428870ff BB |
2783 | ASSERT(db->db_blkptr == bp); |
2784 | ||
03c6040b GW |
2785 | /* |
2786 | * For nopwrites and rewrites we ensure that the bp matches our | |
2787 | * original and bypass all the accounting. | |
2788 | */ | |
2789 | if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) { | |
428870ff BB |
2790 | ASSERT(BP_EQUAL(bp, bp_orig)); |
2791 | } else { | |
b0bc7a84 | 2792 | dsl_dataset_t *ds = os->os_dsl_dataset; |
428870ff BB |
2793 | (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE); |
2794 | dsl_dataset_block_born(ds, bp, tx); | |
2795 | } | |
34dc7c2f BB |
2796 | |
2797 | mutex_enter(&db->db_mtx); | |
2798 | ||
428870ff BB |
2799 | DBUF_VERIFY(db); |
2800 | ||
34dc7c2f BB |
2801 | drp = &db->db_last_dirty; |
2802 | while ((dr = *drp) != db->db_data_pending) | |
2803 | drp = &dr->dr_next; | |
2804 | ASSERT(!list_link_active(&dr->dr_dirty_node)); | |
428870ff | 2805 | ASSERT(dr->dr_dbuf == db); |
34dc7c2f BB |
2806 | ASSERT(dr->dr_next == NULL); |
2807 | *drp = dr->dr_next; | |
2808 | ||
428870ff BB |
2809 | #ifdef ZFS_DEBUG |
2810 | if (db->db_blkid == DMU_SPILL_BLKID) { | |
572e2857 BB |
2811 | dnode_t *dn; |
2812 | ||
2813 | DB_DNODE_ENTER(db); | |
2814 | dn = DB_DNODE(db); | |
428870ff BB |
2815 | ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); |
2816 | ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && | |
2817 | db->db_blkptr == &dn->dn_phys->dn_spill); | |
572e2857 | 2818 | DB_DNODE_EXIT(db); |
428870ff BB |
2819 | } |
2820 | #endif | |
2821 | ||
34dc7c2f | 2822 | if (db->db_level == 0) { |
428870ff | 2823 | ASSERT(db->db_blkid != DMU_BONUS_BLKID); |
34dc7c2f | 2824 | ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); |
b128c09f BB |
2825 | if (db->db_state != DB_NOFILL) { |
2826 | if (dr->dt.dl.dr_data != db->db_buf) | |
2827 | VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, | |
13fe0198 | 2828 | db)); |
428870ff | 2829 | else if (!arc_released(db->db_buf)) |
b128c09f | 2830 | arc_set_callback(db->db_buf, dbuf_do_evict, db); |
b128c09f | 2831 | } |
34dc7c2f | 2832 | } else { |
572e2857 BB |
2833 | dnode_t *dn; |
2834 | ||
2835 | DB_DNODE_ENTER(db); | |
2836 | dn = DB_DNODE(db); | |
34dc7c2f | 2837 | ASSERT(list_head(&dr->dt.di.dr_children) == NULL); |
b0bc7a84 | 2838 | ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift); |
34dc7c2f | 2839 | if (!BP_IS_HOLE(db->db_blkptr)) { |
1fde1e37 BB |
2840 | ASSERTV(int epbs = dn->dn_phys->dn_indblkshift - |
2841 | SPA_BLKPTRSHIFT); | |
b0bc7a84 MG |
2842 | ASSERT3U(db->db_blkid, <=, |
2843 | dn->dn_phys->dn_maxblkid >> (db->db_level * epbs)); | |
34dc7c2f BB |
2844 | ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, |
2845 | db->db.db_size); | |
9b67f605 MA |
2846 | if (!arc_released(db->db_buf)) |
2847 | arc_set_callback(db->db_buf, dbuf_do_evict, db); | |
34dc7c2f | 2848 | } |
572e2857 | 2849 | DB_DNODE_EXIT(db); |
34dc7c2f BB |
2850 | mutex_destroy(&dr->dt.di.dr_mtx); |
2851 | list_destroy(&dr->dt.di.dr_children); | |
2852 | } | |
2853 | kmem_free(dr, sizeof (dbuf_dirty_record_t)); | |
2854 | ||
2855 | cv_broadcast(&db->db_changed); | |
2856 | ASSERT(db->db_dirtycnt > 0); | |
2857 | db->db_dirtycnt -= 1; | |
2858 | db->db_data_pending = NULL; | |
b0bc7a84 | 2859 | dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg); |
428870ff BB |
2860 | } |
2861 | ||
2862 | static void | |
2863 | dbuf_write_nofill_ready(zio_t *zio) | |
2864 | { | |
2865 | dbuf_write_ready(zio, NULL, zio->io_private); | |
2866 | } | |
2867 | ||
2868 | static void | |
2869 | dbuf_write_nofill_done(zio_t *zio) | |
2870 | { | |
2871 | dbuf_write_done(zio, NULL, zio->io_private); | |
2872 | } | |
2873 | ||
2874 | static void | |
2875 | dbuf_write_override_ready(zio_t *zio) | |
2876 | { | |
2877 | dbuf_dirty_record_t *dr = zio->io_private; | |
2878 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
2879 | ||
2880 | dbuf_write_ready(zio, NULL, db); | |
2881 | } | |
2882 | ||
2883 | static void | |
2884 | dbuf_write_override_done(zio_t *zio) | |
2885 | { | |
2886 | dbuf_dirty_record_t *dr = zio->io_private; | |
2887 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
2888 | blkptr_t *obp = &dr->dt.dl.dr_overridden_by; | |
2889 | ||
2890 | mutex_enter(&db->db_mtx); | |
2891 | if (!BP_EQUAL(zio->io_bp, obp)) { | |
2892 | if (!BP_IS_HOLE(obp)) | |
2893 | dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp); | |
2894 | arc_release(dr->dt.dl.dr_data, db); | |
2895 | } | |
34dc7c2f BB |
2896 | mutex_exit(&db->db_mtx); |
2897 | ||
428870ff BB |
2898 | dbuf_write_done(zio, NULL, db); |
2899 | } | |
2900 | ||
e49f1e20 | 2901 | /* Issue I/O to commit a dirty buffer to disk. */ |
428870ff BB |
2902 | static void |
2903 | dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx) | |
2904 | { | |
2905 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
572e2857 BB |
2906 | dnode_t *dn; |
2907 | objset_t *os; | |
428870ff BB |
2908 | dmu_buf_impl_t *parent = db->db_parent; |
2909 | uint64_t txg = tx->tx_txg; | |
5dbd68a3 | 2910 | zbookmark_phys_t zb; |
428870ff BB |
2911 | zio_prop_t zp; |
2912 | zio_t *zio; | |
2913 | int wp_flag = 0; | |
34dc7c2f | 2914 | |
572e2857 BB |
2915 | DB_DNODE_ENTER(db); |
2916 | dn = DB_DNODE(db); | |
2917 | os = dn->dn_objset; | |
2918 | ||
428870ff BB |
2919 | if (db->db_state != DB_NOFILL) { |
2920 | if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) { | |
2921 | /* | |
2922 | * Private object buffers are released here rather | |
2923 | * than in dbuf_dirty() since they are only modified | |
2924 | * in the syncing context and we don't want the | |
2925 | * overhead of making multiple copies of the data. | |
2926 | */ | |
2927 | if (BP_IS_HOLE(db->db_blkptr)) { | |
2928 | arc_buf_thaw(data); | |
2929 | } else { | |
2930 | dbuf_release_bp(db); | |
2931 | } | |
2932 | } | |
2933 | } | |
2934 | ||
2935 | if (parent != dn->dn_dbuf) { | |
e49f1e20 WA |
2936 | /* Our parent is an indirect block. */ |
2937 | /* We have a dirty parent that has been scheduled for write. */ | |
428870ff | 2938 | ASSERT(parent && parent->db_data_pending); |
e49f1e20 | 2939 | /* Our parent's buffer is one level closer to the dnode. */ |
428870ff | 2940 | ASSERT(db->db_level == parent->db_level-1); |
e49f1e20 WA |
2941 | /* |
2942 | * We're about to modify our parent's db_data by modifying | |
2943 | * our block pointer, so the parent must be released. | |
2944 | */ | |
428870ff BB |
2945 | ASSERT(arc_released(parent->db_buf)); |
2946 | zio = parent->db_data_pending->dr_zio; | |
2947 | } else { | |
e49f1e20 | 2948 | /* Our parent is the dnode itself. */ |
428870ff BB |
2949 | ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 && |
2950 | db->db_blkid != DMU_SPILL_BLKID) || | |
2951 | (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0)); | |
2952 | if (db->db_blkid != DMU_SPILL_BLKID) | |
2953 | ASSERT3P(db->db_blkptr, ==, | |
2954 | &dn->dn_phys->dn_blkptr[db->db_blkid]); | |
2955 | zio = dn->dn_zio; | |
2956 | } | |
2957 | ||
2958 | ASSERT(db->db_level == 0 || data == db->db_buf); | |
2959 | ASSERT3U(db->db_blkptr->blk_birth, <=, txg); | |
2960 | ASSERT(zio); | |
2961 | ||
2962 | SET_BOOKMARK(&zb, os->os_dsl_dataset ? | |
2963 | os->os_dsl_dataset->ds_object : DMU_META_OBJSET, | |
2964 | db->db.db_object, db->db_level, db->db_blkid); | |
2965 | ||
2966 | if (db->db_blkid == DMU_SPILL_BLKID) | |
2967 | wp_flag = WP_SPILL; | |
2968 | wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0; | |
2969 | ||
2970 | dmu_write_policy(os, dn, db->db_level, wp_flag, &zp); | |
572e2857 | 2971 | DB_DNODE_EXIT(db); |
428870ff | 2972 | |
9b67f605 MA |
2973 | if (db->db_level == 0 && |
2974 | dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { | |
2975 | /* | |
2976 | * The BP for this block has been provided by open context | |
2977 | * (by dmu_sync() or dmu_buf_write_embedded()). | |
2978 | */ | |
2979 | void *contents = (data != NULL) ? data->b_data : NULL; | |
2980 | ||
428870ff | 2981 | dr->dr_zio = zio_write(zio, os->os_spa, txg, |
9b67f605 | 2982 | db->db_blkptr, contents, db->db.db_size, &zp, |
e8b96c60 MA |
2983 | dbuf_write_override_ready, NULL, dbuf_write_override_done, |
2984 | dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); | |
428870ff BB |
2985 | mutex_enter(&db->db_mtx); |
2986 | dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; | |
2987 | zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by, | |
03c6040b | 2988 | dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite); |
428870ff BB |
2989 | mutex_exit(&db->db_mtx); |
2990 | } else if (db->db_state == DB_NOFILL) { | |
2991 | ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF); | |
2992 | dr->dr_zio = zio_write(zio, os->os_spa, txg, | |
2993 | db->db_blkptr, NULL, db->db.db_size, &zp, | |
e8b96c60 | 2994 | dbuf_write_nofill_ready, NULL, dbuf_write_nofill_done, db, |
428870ff BB |
2995 | ZIO_PRIORITY_ASYNC_WRITE, |
2996 | ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb); | |
2997 | } else { | |
2998 | ASSERT(arc_released(data)); | |
2999 | dr->dr_zio = arc_write(zio, os->os_spa, txg, | |
3a17a7a9 SK |
3000 | db->db_blkptr, data, DBUF_IS_L2CACHEABLE(db), |
3001 | DBUF_IS_L2COMPRESSIBLE(db), &zp, dbuf_write_ready, | |
e8b96c60 MA |
3002 | dbuf_write_physdone, dbuf_write_done, db, |
3003 | ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); | |
428870ff | 3004 | } |
34dc7c2f | 3005 | } |
c28b2279 BB |
3006 | |
3007 | #if defined(_KERNEL) && defined(HAVE_SPL) | |
8f576c23 BB |
3008 | EXPORT_SYMBOL(dbuf_find); |
3009 | EXPORT_SYMBOL(dbuf_is_metadata); | |
3010 | EXPORT_SYMBOL(dbuf_evict); | |
3011 | EXPORT_SYMBOL(dbuf_loan_arcbuf); | |
3012 | EXPORT_SYMBOL(dbuf_whichblock); | |
3013 | EXPORT_SYMBOL(dbuf_read); | |
3014 | EXPORT_SYMBOL(dbuf_unoverride); | |
3015 | EXPORT_SYMBOL(dbuf_free_range); | |
3016 | EXPORT_SYMBOL(dbuf_new_size); | |
3017 | EXPORT_SYMBOL(dbuf_release_bp); | |
3018 | EXPORT_SYMBOL(dbuf_dirty); | |
c28b2279 | 3019 | EXPORT_SYMBOL(dmu_buf_will_dirty); |
8f576c23 BB |
3020 | EXPORT_SYMBOL(dmu_buf_will_not_fill); |
3021 | EXPORT_SYMBOL(dmu_buf_will_fill); | |
3022 | EXPORT_SYMBOL(dmu_buf_fill_done); | |
4047414a | 3023 | EXPORT_SYMBOL(dmu_buf_rele); |
8f576c23 BB |
3024 | EXPORT_SYMBOL(dbuf_assign_arcbuf); |
3025 | EXPORT_SYMBOL(dbuf_clear); | |
3026 | EXPORT_SYMBOL(dbuf_prefetch); | |
3027 | EXPORT_SYMBOL(dbuf_hold_impl); | |
3028 | EXPORT_SYMBOL(dbuf_hold); | |
3029 | EXPORT_SYMBOL(dbuf_hold_level); | |
3030 | EXPORT_SYMBOL(dbuf_create_bonus); | |
3031 | EXPORT_SYMBOL(dbuf_spill_set_blksz); | |
3032 | EXPORT_SYMBOL(dbuf_rm_spill); | |
3033 | EXPORT_SYMBOL(dbuf_add_ref); | |
3034 | EXPORT_SYMBOL(dbuf_rele); | |
3035 | EXPORT_SYMBOL(dbuf_rele_and_unlock); | |
3036 | EXPORT_SYMBOL(dbuf_refcount); | |
3037 | EXPORT_SYMBOL(dbuf_sync_list); | |
3038 | EXPORT_SYMBOL(dmu_buf_set_user); | |
3039 | EXPORT_SYMBOL(dmu_buf_set_user_ie); | |
3040 | EXPORT_SYMBOL(dmu_buf_update_user); | |
3041 | EXPORT_SYMBOL(dmu_buf_get_user); | |
3042 | EXPORT_SYMBOL(dmu_buf_freeable); | |
0f699108 | 3043 | EXPORT_SYMBOL(dmu_buf_get_blkptr); |
c28b2279 | 3044 | #endif |