]>
Commit | Line | Data |
---|---|---|
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 | /* | |
22 | * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. | |
23 | * Copyright (c) 2013 by Delphix. All rights reserved. | |
24 | * Copyright (c) 2013 by Saso Kiselkov. All rights reserved. | |
25 | */ | |
26 | ||
27 | #include <sys/dmu.h> | |
28 | #include <sys/dmu_impl.h> | |
29 | #include <sys/dmu_tx.h> | |
30 | #include <sys/dbuf.h> | |
31 | #include <sys/dnode.h> | |
32 | #include <sys/zfs_context.h> | |
33 | #include <sys/dmu_objset.h> | |
34 | #include <sys/dmu_traverse.h> | |
35 | #include <sys/dsl_dataset.h> | |
36 | #include <sys/dsl_dir.h> | |
37 | #include <sys/dsl_pool.h> | |
38 | #include <sys/dsl_synctask.h> | |
39 | #include <sys/dsl_prop.h> | |
40 | #include <sys/dmu_zfetch.h> | |
41 | #include <sys/zfs_ioctl.h> | |
42 | #include <sys/zap.h> | |
43 | #include <sys/zio_checksum.h> | |
44 | #include <sys/zio_compress.h> | |
45 | #include <sys/sa.h> | |
46 | #ifdef _KERNEL | |
47 | #include <sys/vmsystm.h> | |
48 | #include <sys/zfs_znode.h> | |
49 | #endif | |
50 | ||
51 | /* | |
52 | * Enable/disable nopwrite feature. | |
53 | */ | |
54 | int zfs_nopwrite_enabled = 1; | |
55 | ||
56 | const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = { | |
57 | { DMU_BSWAP_UINT8, TRUE, "unallocated" }, | |
58 | { DMU_BSWAP_ZAP, TRUE, "object directory" }, | |
59 | { DMU_BSWAP_UINT64, TRUE, "object array" }, | |
60 | { DMU_BSWAP_UINT8, TRUE, "packed nvlist" }, | |
61 | { DMU_BSWAP_UINT64, TRUE, "packed nvlist size" }, | |
62 | { DMU_BSWAP_UINT64, TRUE, "bpobj" }, | |
63 | { DMU_BSWAP_UINT64, TRUE, "bpobj header" }, | |
64 | { DMU_BSWAP_UINT64, TRUE, "SPA space map header" }, | |
65 | { DMU_BSWAP_UINT64, TRUE, "SPA space map" }, | |
66 | { DMU_BSWAP_UINT64, TRUE, "ZIL intent log" }, | |
67 | { DMU_BSWAP_DNODE, TRUE, "DMU dnode" }, | |
68 | { DMU_BSWAP_OBJSET, TRUE, "DMU objset" }, | |
69 | { DMU_BSWAP_UINT64, TRUE, "DSL directory" }, | |
70 | { DMU_BSWAP_ZAP, TRUE, "DSL directory child map"}, | |
71 | { DMU_BSWAP_ZAP, TRUE, "DSL dataset snap map" }, | |
72 | { DMU_BSWAP_ZAP, TRUE, "DSL props" }, | |
73 | { DMU_BSWAP_UINT64, TRUE, "DSL dataset" }, | |
74 | { DMU_BSWAP_ZNODE, TRUE, "ZFS znode" }, | |
75 | { DMU_BSWAP_OLDACL, TRUE, "ZFS V0 ACL" }, | |
76 | { DMU_BSWAP_UINT8, FALSE, "ZFS plain file" }, | |
77 | { DMU_BSWAP_ZAP, TRUE, "ZFS directory" }, | |
78 | { DMU_BSWAP_ZAP, TRUE, "ZFS master node" }, | |
79 | { DMU_BSWAP_ZAP, TRUE, "ZFS delete queue" }, | |
80 | { DMU_BSWAP_UINT8, FALSE, "zvol object" }, | |
81 | { DMU_BSWAP_ZAP, TRUE, "zvol prop" }, | |
82 | { DMU_BSWAP_UINT8, FALSE, "other uint8[]" }, | |
83 | { DMU_BSWAP_UINT64, FALSE, "other uint64[]" }, | |
84 | { DMU_BSWAP_ZAP, TRUE, "other ZAP" }, | |
85 | { DMU_BSWAP_ZAP, TRUE, "persistent error log" }, | |
86 | { DMU_BSWAP_UINT8, TRUE, "SPA history" }, | |
87 | { DMU_BSWAP_UINT64, TRUE, "SPA history offsets" }, | |
88 | { DMU_BSWAP_ZAP, TRUE, "Pool properties" }, | |
89 | { DMU_BSWAP_ZAP, TRUE, "DSL permissions" }, | |
90 | { DMU_BSWAP_ACL, TRUE, "ZFS ACL" }, | |
91 | { DMU_BSWAP_UINT8, TRUE, "ZFS SYSACL" }, | |
92 | { DMU_BSWAP_UINT8, TRUE, "FUID table" }, | |
93 | { DMU_BSWAP_UINT64, TRUE, "FUID table size" }, | |
94 | { DMU_BSWAP_ZAP, TRUE, "DSL dataset next clones"}, | |
95 | { DMU_BSWAP_ZAP, TRUE, "scan work queue" }, | |
96 | { DMU_BSWAP_ZAP, TRUE, "ZFS user/group used" }, | |
97 | { DMU_BSWAP_ZAP, TRUE, "ZFS user/group quota" }, | |
98 | { DMU_BSWAP_ZAP, TRUE, "snapshot refcount tags"}, | |
99 | { DMU_BSWAP_ZAP, TRUE, "DDT ZAP algorithm" }, | |
100 | { DMU_BSWAP_ZAP, TRUE, "DDT statistics" }, | |
101 | { DMU_BSWAP_UINT8, TRUE, "System attributes" }, | |
102 | { DMU_BSWAP_ZAP, TRUE, "SA master node" }, | |
103 | { DMU_BSWAP_ZAP, TRUE, "SA attr registration" }, | |
104 | { DMU_BSWAP_ZAP, TRUE, "SA attr layouts" }, | |
105 | { DMU_BSWAP_ZAP, TRUE, "scan translations" }, | |
106 | { DMU_BSWAP_UINT8, FALSE, "deduplicated block" }, | |
107 | { DMU_BSWAP_ZAP, TRUE, "DSL deadlist map" }, | |
108 | { DMU_BSWAP_UINT64, TRUE, "DSL deadlist map hdr" }, | |
109 | { DMU_BSWAP_ZAP, TRUE, "DSL dir clones" }, | |
110 | { DMU_BSWAP_UINT64, TRUE, "bpobj subobj" } | |
111 | }; | |
112 | ||
113 | const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = { | |
114 | { byteswap_uint8_array, "uint8" }, | |
115 | { byteswap_uint16_array, "uint16" }, | |
116 | { byteswap_uint32_array, "uint32" }, | |
117 | { byteswap_uint64_array, "uint64" }, | |
118 | { zap_byteswap, "zap" }, | |
119 | { dnode_buf_byteswap, "dnode" }, | |
120 | { dmu_objset_byteswap, "objset" }, | |
121 | { zfs_znode_byteswap, "znode" }, | |
122 | { zfs_oldacl_byteswap, "oldacl" }, | |
123 | { zfs_acl_byteswap, "acl" } | |
124 | }; | |
125 | ||
126 | int | |
127 | dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset, | |
128 | void *tag, dmu_buf_t **dbp, int flags) | |
129 | { | |
130 | dnode_t *dn; | |
131 | uint64_t blkid; | |
132 | dmu_buf_impl_t *db; | |
133 | int err; | |
134 | int db_flags = DB_RF_CANFAIL; | |
135 | ||
136 | if (flags & DMU_READ_NO_PREFETCH) | |
137 | db_flags |= DB_RF_NOPREFETCH; | |
138 | ||
139 | err = dnode_hold(os, object, FTAG, &dn); | |
140 | if (err) | |
141 | return (err); | |
142 | blkid = dbuf_whichblock(dn, offset); | |
143 | rw_enter(&dn->dn_struct_rwlock, RW_READER); | |
144 | db = dbuf_hold(dn, blkid, tag); | |
145 | rw_exit(&dn->dn_struct_rwlock); | |
146 | if (db == NULL) { | |
147 | err = SET_ERROR(EIO); | |
148 | } else { | |
149 | err = dbuf_read(db, NULL, db_flags); | |
150 | if (err) { | |
151 | dbuf_rele(db, tag); | |
152 | db = NULL; | |
153 | } | |
154 | } | |
155 | ||
156 | dnode_rele(dn, FTAG); | |
157 | *dbp = &db->db; /* NULL db plus first field offset is NULL */ | |
158 | return (err); | |
159 | } | |
160 | ||
161 | int | |
162 | dmu_bonus_max(void) | |
163 | { | |
164 | return (DN_MAX_BONUSLEN); | |
165 | } | |
166 | ||
167 | int | |
168 | dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx) | |
169 | { | |
170 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
171 | dnode_t *dn; | |
172 | int error; | |
173 | ||
174 | DB_DNODE_ENTER(db); | |
175 | dn = DB_DNODE(db); | |
176 | ||
177 | if (dn->dn_bonus != db) { | |
178 | error = SET_ERROR(EINVAL); | |
179 | } else if (newsize < 0 || newsize > db_fake->db_size) { | |
180 | error = SET_ERROR(EINVAL); | |
181 | } else { | |
182 | dnode_setbonuslen(dn, newsize, tx); | |
183 | error = 0; | |
184 | } | |
185 | ||
186 | DB_DNODE_EXIT(db); | |
187 | return (error); | |
188 | } | |
189 | ||
190 | int | |
191 | dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx) | |
192 | { | |
193 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
194 | dnode_t *dn; | |
195 | int error; | |
196 | ||
197 | DB_DNODE_ENTER(db); | |
198 | dn = DB_DNODE(db); | |
199 | ||
200 | if (!DMU_OT_IS_VALID(type)) { | |
201 | error = SET_ERROR(EINVAL); | |
202 | } else if (dn->dn_bonus != db) { | |
203 | error = SET_ERROR(EINVAL); | |
204 | } else { | |
205 | dnode_setbonus_type(dn, type, tx); | |
206 | error = 0; | |
207 | } | |
208 | ||
209 | DB_DNODE_EXIT(db); | |
210 | return (error); | |
211 | } | |
212 | ||
213 | dmu_object_type_t | |
214 | dmu_get_bonustype(dmu_buf_t *db_fake) | |
215 | { | |
216 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
217 | dnode_t *dn; | |
218 | dmu_object_type_t type; | |
219 | ||
220 | DB_DNODE_ENTER(db); | |
221 | dn = DB_DNODE(db); | |
222 | type = dn->dn_bonustype; | |
223 | DB_DNODE_EXIT(db); | |
224 | ||
225 | return (type); | |
226 | } | |
227 | ||
228 | int | |
229 | dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx) | |
230 | { | |
231 | dnode_t *dn; | |
232 | int error; | |
233 | ||
234 | error = dnode_hold(os, object, FTAG, &dn); | |
235 | dbuf_rm_spill(dn, tx); | |
236 | rw_enter(&dn->dn_struct_rwlock, RW_WRITER); | |
237 | dnode_rm_spill(dn, tx); | |
238 | rw_exit(&dn->dn_struct_rwlock); | |
239 | dnode_rele(dn, FTAG); | |
240 | return (error); | |
241 | } | |
242 | ||
243 | /* | |
244 | * returns ENOENT, EIO, or 0. | |
245 | */ | |
246 | int | |
247 | dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp) | |
248 | { | |
249 | dnode_t *dn; | |
250 | dmu_buf_impl_t *db; | |
251 | int error; | |
252 | ||
253 | error = dnode_hold(os, object, FTAG, &dn); | |
254 | if (error) | |
255 | return (error); | |
256 | ||
257 | rw_enter(&dn->dn_struct_rwlock, RW_READER); | |
258 | if (dn->dn_bonus == NULL) { | |
259 | rw_exit(&dn->dn_struct_rwlock); | |
260 | rw_enter(&dn->dn_struct_rwlock, RW_WRITER); | |
261 | if (dn->dn_bonus == NULL) | |
262 | dbuf_create_bonus(dn); | |
263 | } | |
264 | db = dn->dn_bonus; | |
265 | ||
266 | /* as long as the bonus buf is held, the dnode will be held */ | |
267 | if (refcount_add(&db->db_holds, tag) == 1) { | |
268 | VERIFY(dnode_add_ref(dn, db)); | |
269 | (void) atomic_inc_32_nv(&dn->dn_dbufs_count); | |
270 | } | |
271 | ||
272 | /* | |
273 | * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's | |
274 | * hold and incrementing the dbuf count to ensure that dnode_move() sees | |
275 | * a dnode hold for every dbuf. | |
276 | */ | |
277 | rw_exit(&dn->dn_struct_rwlock); | |
278 | ||
279 | dnode_rele(dn, FTAG); | |
280 | ||
281 | VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH)); | |
282 | ||
283 | *dbp = &db->db; | |
284 | return (0); | |
285 | } | |
286 | ||
287 | /* | |
288 | * returns ENOENT, EIO, or 0. | |
289 | * | |
290 | * This interface will allocate a blank spill dbuf when a spill blk | |
291 | * doesn't already exist on the dnode. | |
292 | * | |
293 | * if you only want to find an already existing spill db, then | |
294 | * dmu_spill_hold_existing() should be used. | |
295 | */ | |
296 | int | |
297 | dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp) | |
298 | { | |
299 | dmu_buf_impl_t *db = NULL; | |
300 | int err; | |
301 | ||
302 | if ((flags & DB_RF_HAVESTRUCT) == 0) | |
303 | rw_enter(&dn->dn_struct_rwlock, RW_READER); | |
304 | ||
305 | db = dbuf_hold(dn, DMU_SPILL_BLKID, tag); | |
306 | ||
307 | if ((flags & DB_RF_HAVESTRUCT) == 0) | |
308 | rw_exit(&dn->dn_struct_rwlock); | |
309 | ||
310 | ASSERT(db != NULL); | |
311 | err = dbuf_read(db, NULL, flags); | |
312 | if (err == 0) | |
313 | *dbp = &db->db; | |
314 | else | |
315 | dbuf_rele(db, tag); | |
316 | return (err); | |
317 | } | |
318 | ||
319 | int | |
320 | dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp) | |
321 | { | |
322 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus; | |
323 | dnode_t *dn; | |
324 | int err; | |
325 | ||
326 | DB_DNODE_ENTER(db); | |
327 | dn = DB_DNODE(db); | |
328 | ||
329 | if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) { | |
330 | err = SET_ERROR(EINVAL); | |
331 | } else { | |
332 | rw_enter(&dn->dn_struct_rwlock, RW_READER); | |
333 | ||
334 | if (!dn->dn_have_spill) { | |
335 | err = SET_ERROR(ENOENT); | |
336 | } else { | |
337 | err = dmu_spill_hold_by_dnode(dn, | |
338 | DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp); | |
339 | } | |
340 | ||
341 | rw_exit(&dn->dn_struct_rwlock); | |
342 | } | |
343 | ||
344 | DB_DNODE_EXIT(db); | |
345 | return (err); | |
346 | } | |
347 | ||
348 | int | |
349 | dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp) | |
350 | { | |
351 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus; | |
352 | dnode_t *dn; | |
353 | int err; | |
354 | ||
355 | DB_DNODE_ENTER(db); | |
356 | dn = DB_DNODE(db); | |
357 | err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp); | |
358 | DB_DNODE_EXIT(db); | |
359 | ||
360 | return (err); | |
361 | } | |
362 | ||
363 | /* | |
364 | * Note: longer-term, we should modify all of the dmu_buf_*() interfaces | |
365 | * to take a held dnode rather than <os, object> -- the lookup is wasteful, | |
366 | * and can induce severe lock contention when writing to several files | |
367 | * whose dnodes are in the same block. | |
368 | */ | |
369 | static int | |
370 | dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length, | |
371 | int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags) | |
372 | { | |
373 | dsl_pool_t *dp = NULL; | |
374 | dmu_buf_t **dbp; | |
375 | uint64_t blkid, nblks, i; | |
376 | uint32_t dbuf_flags; | |
377 | int err; | |
378 | zio_t *zio; | |
379 | hrtime_t start = 0; | |
380 | ||
381 | ASSERT(length <= DMU_MAX_ACCESS); | |
382 | ||
383 | dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT; | |
384 | if (flags & DMU_READ_NO_PREFETCH || length > zfetch_array_rd_sz) | |
385 | dbuf_flags |= DB_RF_NOPREFETCH; | |
386 | ||
387 | rw_enter(&dn->dn_struct_rwlock, RW_READER); | |
388 | if (dn->dn_datablkshift) { | |
389 | int blkshift = dn->dn_datablkshift; | |
390 | nblks = (P2ROUNDUP(offset+length, 1ULL<<blkshift) - | |
391 | P2ALIGN(offset, 1ULL<<blkshift)) >> blkshift; | |
392 | } else { | |
393 | if (offset + length > dn->dn_datablksz) { | |
394 | zfs_panic_recover("zfs: accessing past end of object " | |
395 | "%llx/%llx (size=%u access=%llu+%llu)", | |
396 | (longlong_t)dn->dn_objset-> | |
397 | os_dsl_dataset->ds_object, | |
398 | (longlong_t)dn->dn_object, dn->dn_datablksz, | |
399 | (longlong_t)offset, (longlong_t)length); | |
400 | rw_exit(&dn->dn_struct_rwlock); | |
401 | return (SET_ERROR(EIO)); | |
402 | } | |
403 | nblks = 1; | |
404 | } | |
405 | dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_PUSHPAGE | KM_NODEBUG); | |
406 | ||
407 | if (dn->dn_objset->os_dsl_dataset) | |
408 | dp = dn->dn_objset->os_dsl_dataset->ds_dir->dd_pool; | |
409 | start = gethrtime(); | |
410 | zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL); | |
411 | blkid = dbuf_whichblock(dn, offset); | |
412 | for (i = 0; i < nblks; i++) { | |
413 | dmu_buf_impl_t *db = dbuf_hold(dn, blkid+i, tag); | |
414 | if (db == NULL) { | |
415 | rw_exit(&dn->dn_struct_rwlock); | |
416 | dmu_buf_rele_array(dbp, nblks, tag); | |
417 | zio_nowait(zio); | |
418 | return (SET_ERROR(EIO)); | |
419 | } | |
420 | /* initiate async i/o */ | |
421 | if (read) { | |
422 | (void) dbuf_read(db, zio, dbuf_flags); | |
423 | } | |
424 | dbp[i] = &db->db; | |
425 | } | |
426 | rw_exit(&dn->dn_struct_rwlock); | |
427 | ||
428 | /* wait for async i/o */ | |
429 | err = zio_wait(zio); | |
430 | /* track read overhead when we are in sync context */ | |
431 | if (dp && dsl_pool_sync_context(dp)) | |
432 | dp->dp_read_overhead += gethrtime() - start; | |
433 | if (err) { | |
434 | dmu_buf_rele_array(dbp, nblks, tag); | |
435 | return (err); | |
436 | } | |
437 | ||
438 | /* wait for other io to complete */ | |
439 | if (read) { | |
440 | for (i = 0; i < nblks; i++) { | |
441 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i]; | |
442 | mutex_enter(&db->db_mtx); | |
443 | while (db->db_state == DB_READ || | |
444 | db->db_state == DB_FILL) | |
445 | cv_wait(&db->db_changed, &db->db_mtx); | |
446 | if (db->db_state == DB_UNCACHED) | |
447 | err = SET_ERROR(EIO); | |
448 | mutex_exit(&db->db_mtx); | |
449 | if (err) { | |
450 | dmu_buf_rele_array(dbp, nblks, tag); | |
451 | return (err); | |
452 | } | |
453 | } | |
454 | } | |
455 | ||
456 | *numbufsp = nblks; | |
457 | *dbpp = dbp; | |
458 | return (0); | |
459 | } | |
460 | ||
461 | static int | |
462 | dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset, | |
463 | uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp) | |
464 | { | |
465 | dnode_t *dn; | |
466 | int err; | |
467 | ||
468 | err = dnode_hold(os, object, FTAG, &dn); | |
469 | if (err) | |
470 | return (err); | |
471 | ||
472 | err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag, | |
473 | numbufsp, dbpp, DMU_READ_PREFETCH); | |
474 | ||
475 | dnode_rele(dn, FTAG); | |
476 | ||
477 | return (err); | |
478 | } | |
479 | ||
480 | int | |
481 | dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset, | |
482 | uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp) | |
483 | { | |
484 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
485 | dnode_t *dn; | |
486 | int err; | |
487 | ||
488 | DB_DNODE_ENTER(db); | |
489 | dn = DB_DNODE(db); | |
490 | err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag, | |
491 | numbufsp, dbpp, DMU_READ_PREFETCH); | |
492 | DB_DNODE_EXIT(db); | |
493 | ||
494 | return (err); | |
495 | } | |
496 | ||
497 | void | |
498 | dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag) | |
499 | { | |
500 | int i; | |
501 | dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake; | |
502 | ||
503 | if (numbufs == 0) | |
504 | return; | |
505 | ||
506 | for (i = 0; i < numbufs; i++) { | |
507 | if (dbp[i]) | |
508 | dbuf_rele(dbp[i], tag); | |
509 | } | |
510 | ||
511 | kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs); | |
512 | } | |
513 | ||
514 | void | |
515 | dmu_prefetch(objset_t *os, uint64_t object, uint64_t offset, uint64_t len) | |
516 | { | |
517 | dnode_t *dn; | |
518 | uint64_t blkid; | |
519 | int nblks, i, err; | |
520 | ||
521 | if (zfs_prefetch_disable) | |
522 | return; | |
523 | ||
524 | if (len == 0) { /* they're interested in the bonus buffer */ | |
525 | dn = DMU_META_DNODE(os); | |
526 | ||
527 | if (object == 0 || object >= DN_MAX_OBJECT) | |
528 | return; | |
529 | ||
530 | rw_enter(&dn->dn_struct_rwlock, RW_READER); | |
531 | blkid = dbuf_whichblock(dn, object * sizeof (dnode_phys_t)); | |
532 | dbuf_prefetch(dn, blkid); | |
533 | rw_exit(&dn->dn_struct_rwlock); | |
534 | return; | |
535 | } | |
536 | ||
537 | /* | |
538 | * XXX - Note, if the dnode for the requested object is not | |
539 | * already cached, we will do a *synchronous* read in the | |
540 | * dnode_hold() call. The same is true for any indirects. | |
541 | */ | |
542 | err = dnode_hold(os, object, FTAG, &dn); | |
543 | if (err != 0) | |
544 | return; | |
545 | ||
546 | rw_enter(&dn->dn_struct_rwlock, RW_READER); | |
547 | if (dn->dn_datablkshift) { | |
548 | int blkshift = dn->dn_datablkshift; | |
549 | nblks = (P2ROUNDUP(offset+len, 1<<blkshift) - | |
550 | P2ALIGN(offset, 1<<blkshift)) >> blkshift; | |
551 | } else { | |
552 | nblks = (offset < dn->dn_datablksz); | |
553 | } | |
554 | ||
555 | if (nblks != 0) { | |
556 | blkid = dbuf_whichblock(dn, offset); | |
557 | for (i = 0; i < nblks; i++) | |
558 | dbuf_prefetch(dn, blkid+i); | |
559 | } | |
560 | ||
561 | rw_exit(&dn->dn_struct_rwlock); | |
562 | ||
563 | dnode_rele(dn, FTAG); | |
564 | } | |
565 | ||
566 | /* | |
567 | * Get the next "chunk" of file data to free. We traverse the file from | |
568 | * the end so that the file gets shorter over time (if we crashes in the | |
569 | * middle, this will leave us in a better state). We find allocated file | |
570 | * data by simply searching the allocated level 1 indirects. | |
571 | * | |
572 | * On input, *start should be the first offset that does not need to be | |
573 | * freed (e.g. "offset + length"). On return, *start will be the first | |
574 | * offset that should be freed. | |
575 | */ | |
576 | static int | |
577 | get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum) | |
578 | { | |
579 | uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1); | |
580 | /* bytes of data covered by a level-1 indirect block */ | |
581 | uint64_t iblkrange = | |
582 | dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT); | |
583 | uint64_t blks; | |
584 | ||
585 | ASSERT3U(minimum, <=, *start); | |
586 | ||
587 | if (*start - minimum <= iblkrange * maxblks) { | |
588 | *start = minimum; | |
589 | return (0); | |
590 | } | |
591 | ASSERT(ISP2(iblkrange)); | |
592 | ||
593 | for (blks = 0; *start > minimum && blks < maxblks; blks++) { | |
594 | int err; | |
595 | ||
596 | /* | |
597 | * dnode_next_offset(BACKWARDS) will find an allocated L1 | |
598 | * indirect block at or before the input offset. We must | |
599 | * decrement *start so that it is at the end of the region | |
600 | * to search. | |
601 | */ | |
602 | (*start)--; | |
603 | err = dnode_next_offset(dn, | |
604 | DNODE_FIND_BACKWARDS, start, 2, 1, 0); | |
605 | ||
606 | /* if there are no indirect blocks before start, we are done */ | |
607 | if (err == ESRCH) { | |
608 | *start = minimum; | |
609 | break; | |
610 | } else if (err != 0) { | |
611 | return (err); | |
612 | } | |
613 | ||
614 | /* set start to the beginning of this L1 indirect */ | |
615 | *start = P2ALIGN(*start, iblkrange); | |
616 | } | |
617 | if (*start < minimum) | |
618 | *start = minimum; | |
619 | return (0); | |
620 | } | |
621 | ||
622 | static int | |
623 | dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset, | |
624 | uint64_t length) | |
625 | { | |
626 | uint64_t object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz; | |
627 | int err; | |
628 | ||
629 | if (offset >= object_size) | |
630 | return (0); | |
631 | ||
632 | if (length == DMU_OBJECT_END || offset + length > object_size) | |
633 | length = object_size - offset; | |
634 | ||
635 | while (length != 0) { | |
636 | uint64_t chunk_end, chunk_begin; | |
637 | dmu_tx_t *tx; | |
638 | ||
639 | chunk_end = chunk_begin = offset + length; | |
640 | ||
641 | /* move chunk_begin backwards to the beginning of this chunk */ | |
642 | err = get_next_chunk(dn, &chunk_begin, offset); | |
643 | if (err) | |
644 | return (err); | |
645 | ASSERT3U(chunk_begin, >=, offset); | |
646 | ASSERT3U(chunk_begin, <=, chunk_end); | |
647 | ||
648 | tx = dmu_tx_create(os); | |
649 | dmu_tx_hold_free(tx, dn->dn_object, | |
650 | chunk_begin, chunk_end - chunk_begin); | |
651 | err = dmu_tx_assign(tx, TXG_WAIT); | |
652 | if (err) { | |
653 | dmu_tx_abort(tx); | |
654 | return (err); | |
655 | } | |
656 | dnode_free_range(dn, chunk_begin, chunk_end - chunk_begin, tx); | |
657 | dmu_tx_commit(tx); | |
658 | ||
659 | length -= chunk_end - chunk_begin; | |
660 | } | |
661 | return (0); | |
662 | } | |
663 | ||
664 | int | |
665 | dmu_free_long_range(objset_t *os, uint64_t object, | |
666 | uint64_t offset, uint64_t length) | |
667 | { | |
668 | dnode_t *dn; | |
669 | int err; | |
670 | ||
671 | err = dnode_hold(os, object, FTAG, &dn); | |
672 | if (err != 0) | |
673 | return (err); | |
674 | err = dmu_free_long_range_impl(os, dn, offset, length); | |
675 | dnode_rele(dn, FTAG); | |
676 | return (err); | |
677 | } | |
678 | ||
679 | int | |
680 | dmu_free_long_object(objset_t *os, uint64_t object) | |
681 | { | |
682 | dmu_tx_t *tx; | |
683 | int err; | |
684 | ||
685 | err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END); | |
686 | if (err != 0) | |
687 | return (err); | |
688 | ||
689 | tx = dmu_tx_create(os); | |
690 | dmu_tx_hold_bonus(tx, object); | |
691 | dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END); | |
692 | err = dmu_tx_assign(tx, TXG_WAIT); | |
693 | if (err == 0) { | |
694 | err = dmu_object_free(os, object, tx); | |
695 | dmu_tx_commit(tx); | |
696 | } else { | |
697 | dmu_tx_abort(tx); | |
698 | } | |
699 | ||
700 | return (err); | |
701 | } | |
702 | ||
703 | int | |
704 | dmu_free_range(objset_t *os, uint64_t object, uint64_t offset, | |
705 | uint64_t size, dmu_tx_t *tx) | |
706 | { | |
707 | dnode_t *dn; | |
708 | int err = dnode_hold(os, object, FTAG, &dn); | |
709 | if (err) | |
710 | return (err); | |
711 | ASSERT(offset < UINT64_MAX); | |
712 | ASSERT(size == -1ULL || size <= UINT64_MAX - offset); | |
713 | dnode_free_range(dn, offset, size, tx); | |
714 | dnode_rele(dn, FTAG); | |
715 | return (0); | |
716 | } | |
717 | ||
718 | int | |
719 | dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, | |
720 | void *buf, uint32_t flags) | |
721 | { | |
722 | dnode_t *dn; | |
723 | dmu_buf_t **dbp; | |
724 | int numbufs, err; | |
725 | ||
726 | err = dnode_hold(os, object, FTAG, &dn); | |
727 | if (err) | |
728 | return (err); | |
729 | ||
730 | /* | |
731 | * Deal with odd block sizes, where there can't be data past the first | |
732 | * block. If we ever do the tail block optimization, we will need to | |
733 | * handle that here as well. | |
734 | */ | |
735 | if (dn->dn_maxblkid == 0) { | |
736 | int newsz = offset > dn->dn_datablksz ? 0 : | |
737 | MIN(size, dn->dn_datablksz - offset); | |
738 | bzero((char *)buf + newsz, size - newsz); | |
739 | size = newsz; | |
740 | } | |
741 | ||
742 | while (size > 0) { | |
743 | uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2); | |
744 | int i; | |
745 | ||
746 | /* | |
747 | * NB: we could do this block-at-a-time, but it's nice | |
748 | * to be reading in parallel. | |
749 | */ | |
750 | err = dmu_buf_hold_array_by_dnode(dn, offset, mylen, | |
751 | TRUE, FTAG, &numbufs, &dbp, flags); | |
752 | if (err) | |
753 | break; | |
754 | ||
755 | for (i = 0; i < numbufs; i++) { | |
756 | int tocpy; | |
757 | int bufoff; | |
758 | dmu_buf_t *db = dbp[i]; | |
759 | ||
760 | ASSERT(size > 0); | |
761 | ||
762 | bufoff = offset - db->db_offset; | |
763 | tocpy = (int)MIN(db->db_size - bufoff, size); | |
764 | ||
765 | bcopy((char *)db->db_data + bufoff, buf, tocpy); | |
766 | ||
767 | offset += tocpy; | |
768 | size -= tocpy; | |
769 | buf = (char *)buf + tocpy; | |
770 | } | |
771 | dmu_buf_rele_array(dbp, numbufs, FTAG); | |
772 | } | |
773 | dnode_rele(dn, FTAG); | |
774 | return (err); | |
775 | } | |
776 | ||
777 | void | |
778 | dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, | |
779 | const void *buf, dmu_tx_t *tx) | |
780 | { | |
781 | dmu_buf_t **dbp; | |
782 | int numbufs, i; | |
783 | ||
784 | if (size == 0) | |
785 | return; | |
786 | ||
787 | VERIFY(0 == dmu_buf_hold_array(os, object, offset, size, | |
788 | FALSE, FTAG, &numbufs, &dbp)); | |
789 | ||
790 | for (i = 0; i < numbufs; i++) { | |
791 | int tocpy; | |
792 | int bufoff; | |
793 | dmu_buf_t *db = dbp[i]; | |
794 | ||
795 | ASSERT(size > 0); | |
796 | ||
797 | bufoff = offset - db->db_offset; | |
798 | tocpy = (int)MIN(db->db_size - bufoff, size); | |
799 | ||
800 | ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); | |
801 | ||
802 | if (tocpy == db->db_size) | |
803 | dmu_buf_will_fill(db, tx); | |
804 | else | |
805 | dmu_buf_will_dirty(db, tx); | |
806 | ||
807 | (void) memcpy((char *)db->db_data + bufoff, buf, tocpy); | |
808 | ||
809 | if (tocpy == db->db_size) | |
810 | dmu_buf_fill_done(db, tx); | |
811 | ||
812 | offset += tocpy; | |
813 | size -= tocpy; | |
814 | buf = (char *)buf + tocpy; | |
815 | } | |
816 | dmu_buf_rele_array(dbp, numbufs, FTAG); | |
817 | } | |
818 | ||
819 | void | |
820 | dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, | |
821 | dmu_tx_t *tx) | |
822 | { | |
823 | dmu_buf_t **dbp; | |
824 | int numbufs, i; | |
825 | ||
826 | if (size == 0) | |
827 | return; | |
828 | ||
829 | VERIFY(0 == dmu_buf_hold_array(os, object, offset, size, | |
830 | FALSE, FTAG, &numbufs, &dbp)); | |
831 | ||
832 | for (i = 0; i < numbufs; i++) { | |
833 | dmu_buf_t *db = dbp[i]; | |
834 | ||
835 | dmu_buf_will_not_fill(db, tx); | |
836 | } | |
837 | dmu_buf_rele_array(dbp, numbufs, FTAG); | |
838 | } | |
839 | ||
840 | /* | |
841 | * DMU support for xuio | |
842 | */ | |
843 | kstat_t *xuio_ksp = NULL; | |
844 | ||
845 | typedef struct xuio_stats { | |
846 | /* loaned yet not returned arc_buf */ | |
847 | kstat_named_t xuiostat_onloan_rbuf; | |
848 | kstat_named_t xuiostat_onloan_wbuf; | |
849 | /* whether a copy is made when loaning out a read buffer */ | |
850 | kstat_named_t xuiostat_rbuf_copied; | |
851 | kstat_named_t xuiostat_rbuf_nocopy; | |
852 | /* whether a copy is made when assigning a write buffer */ | |
853 | kstat_named_t xuiostat_wbuf_copied; | |
854 | kstat_named_t xuiostat_wbuf_nocopy; | |
855 | } xuio_stats_t; | |
856 | ||
857 | static xuio_stats_t xuio_stats = { | |
858 | { "onloan_read_buf", KSTAT_DATA_UINT64 }, | |
859 | { "onloan_write_buf", KSTAT_DATA_UINT64 }, | |
860 | { "read_buf_copied", KSTAT_DATA_UINT64 }, | |
861 | { "read_buf_nocopy", KSTAT_DATA_UINT64 }, | |
862 | { "write_buf_copied", KSTAT_DATA_UINT64 }, | |
863 | { "write_buf_nocopy", KSTAT_DATA_UINT64 } | |
864 | }; | |
865 | ||
866 | #define XUIOSTAT_INCR(stat, val) \ | |
867 | atomic_add_64(&xuio_stats.stat.value.ui64, (val)) | |
868 | #define XUIOSTAT_BUMP(stat) XUIOSTAT_INCR(stat, 1) | |
869 | ||
870 | int | |
871 | dmu_xuio_init(xuio_t *xuio, int nblk) | |
872 | { | |
873 | dmu_xuio_t *priv; | |
874 | uio_t *uio = &xuio->xu_uio; | |
875 | ||
876 | uio->uio_iovcnt = nblk; | |
877 | uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_PUSHPAGE); | |
878 | ||
879 | priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_PUSHPAGE); | |
880 | priv->cnt = nblk; | |
881 | priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_PUSHPAGE); | |
882 | priv->iovp = uio->uio_iov; | |
883 | XUIO_XUZC_PRIV(xuio) = priv; | |
884 | ||
885 | if (XUIO_XUZC_RW(xuio) == UIO_READ) | |
886 | XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk); | |
887 | else | |
888 | XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk); | |
889 | ||
890 | return (0); | |
891 | } | |
892 | ||
893 | void | |
894 | dmu_xuio_fini(xuio_t *xuio) | |
895 | { | |
896 | dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); | |
897 | int nblk = priv->cnt; | |
898 | ||
899 | kmem_free(priv->iovp, nblk * sizeof (iovec_t)); | |
900 | kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *)); | |
901 | kmem_free(priv, sizeof (dmu_xuio_t)); | |
902 | ||
903 | if (XUIO_XUZC_RW(xuio) == UIO_READ) | |
904 | XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk); | |
905 | else | |
906 | XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk); | |
907 | } | |
908 | ||
909 | /* | |
910 | * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf } | |
911 | * and increase priv->next by 1. | |
912 | */ | |
913 | int | |
914 | dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n) | |
915 | { | |
916 | struct iovec *iov; | |
917 | uio_t *uio = &xuio->xu_uio; | |
918 | dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); | |
919 | int i = priv->next++; | |
920 | ||
921 | ASSERT(i < priv->cnt); | |
922 | ASSERT(off + n <= arc_buf_size(abuf)); | |
923 | iov = uio->uio_iov + i; | |
924 | iov->iov_base = (char *)abuf->b_data + off; | |
925 | iov->iov_len = n; | |
926 | priv->bufs[i] = abuf; | |
927 | return (0); | |
928 | } | |
929 | ||
930 | int | |
931 | dmu_xuio_cnt(xuio_t *xuio) | |
932 | { | |
933 | dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); | |
934 | return (priv->cnt); | |
935 | } | |
936 | ||
937 | arc_buf_t * | |
938 | dmu_xuio_arcbuf(xuio_t *xuio, int i) | |
939 | { | |
940 | dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); | |
941 | ||
942 | ASSERT(i < priv->cnt); | |
943 | return (priv->bufs[i]); | |
944 | } | |
945 | ||
946 | void | |
947 | dmu_xuio_clear(xuio_t *xuio, int i) | |
948 | { | |
949 | dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); | |
950 | ||
951 | ASSERT(i < priv->cnt); | |
952 | priv->bufs[i] = NULL; | |
953 | } | |
954 | ||
955 | static void | |
956 | xuio_stat_init(void) | |
957 | { | |
958 | xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc", | |
959 | KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t), | |
960 | KSTAT_FLAG_VIRTUAL); | |
961 | if (xuio_ksp != NULL) { | |
962 | xuio_ksp->ks_data = &xuio_stats; | |
963 | kstat_install(xuio_ksp); | |
964 | } | |
965 | } | |
966 | ||
967 | static void | |
968 | xuio_stat_fini(void) | |
969 | { | |
970 | if (xuio_ksp != NULL) { | |
971 | kstat_delete(xuio_ksp); | |
972 | xuio_ksp = NULL; | |
973 | } | |
974 | } | |
975 | ||
976 | void | |
977 | xuio_stat_wbuf_copied() | |
978 | { | |
979 | XUIOSTAT_BUMP(xuiostat_wbuf_copied); | |
980 | } | |
981 | ||
982 | void | |
983 | xuio_stat_wbuf_nocopy() | |
984 | { | |
985 | XUIOSTAT_BUMP(xuiostat_wbuf_nocopy); | |
986 | } | |
987 | ||
988 | #ifdef _KERNEL | |
989 | ||
990 | /* | |
991 | * Copy up to size bytes between arg_buf and req based on the data direction | |
992 | * described by the req. If an entire req's data cannot be transfered the | |
993 | * req's is updated such that it's current index and bv offsets correctly | |
994 | * reference any residual data which could not be copied. The return value | |
995 | * is the number of bytes successfully copied to arg_buf. | |
996 | */ | |
997 | static int | |
998 | dmu_req_copy(void *arg_buf, int size, int *offset, struct request *req) | |
999 | { | |
1000 | struct bio_vec *bv; | |
1001 | struct req_iterator iter; | |
1002 | char *bv_buf; | |
1003 | int tocpy; | |
1004 | ||
1005 | *offset = 0; | |
1006 | rq_for_each_segment(bv, req, iter) { | |
1007 | ||
1008 | /* Fully consumed the passed arg_buf */ | |
1009 | ASSERT3S(*offset, <=, size); | |
1010 | if (size == *offset) | |
1011 | break; | |
1012 | ||
1013 | /* Skip fully consumed bv's */ | |
1014 | if (bv->bv_len == 0) | |
1015 | continue; | |
1016 | ||
1017 | tocpy = MIN(bv->bv_len, size - *offset); | |
1018 | ASSERT3S(tocpy, >=, 0); | |
1019 | ||
1020 | bv_buf = page_address(bv->bv_page) + bv->bv_offset; | |
1021 | ASSERT3P(bv_buf, !=, NULL); | |
1022 | ||
1023 | if (rq_data_dir(req) == WRITE) | |
1024 | memcpy(arg_buf + *offset, bv_buf, tocpy); | |
1025 | else | |
1026 | memcpy(bv_buf, arg_buf + *offset, tocpy); | |
1027 | ||
1028 | *offset += tocpy; | |
1029 | bv->bv_offset += tocpy; | |
1030 | bv->bv_len -= tocpy; | |
1031 | } | |
1032 | ||
1033 | return 0; | |
1034 | } | |
1035 | ||
1036 | static void | |
1037 | dmu_bio_put(struct bio *bio) | |
1038 | { | |
1039 | struct bio *bio_next; | |
1040 | ||
1041 | while (bio) { | |
1042 | bio_next = bio->bi_next; | |
1043 | bio_put(bio); | |
1044 | bio = bio_next; | |
1045 | } | |
1046 | } | |
1047 | ||
1048 | static int | |
1049 | dmu_bio_clone(struct bio *bio, struct bio **bio_copy) | |
1050 | { | |
1051 | struct bio *bio_root = NULL; | |
1052 | struct bio *bio_last = NULL; | |
1053 | struct bio *bio_new; | |
1054 | ||
1055 | if (bio == NULL) | |
1056 | return EINVAL; | |
1057 | ||
1058 | while (bio) { | |
1059 | bio_new = bio_clone(bio, GFP_NOIO); | |
1060 | if (bio_new == NULL) { | |
1061 | dmu_bio_put(bio_root); | |
1062 | return ENOMEM; | |
1063 | } | |
1064 | ||
1065 | if (bio_last) { | |
1066 | bio_last->bi_next = bio_new; | |
1067 | bio_last = bio_new; | |
1068 | } else { | |
1069 | bio_root = bio_new; | |
1070 | bio_last = bio_new; | |
1071 | } | |
1072 | ||
1073 | bio = bio->bi_next; | |
1074 | } | |
1075 | ||
1076 | *bio_copy = bio_root; | |
1077 | ||
1078 | return 0; | |
1079 | } | |
1080 | ||
1081 | int | |
1082 | dmu_read_req(objset_t *os, uint64_t object, struct request *req) | |
1083 | { | |
1084 | uint64_t size = blk_rq_bytes(req); | |
1085 | uint64_t offset = blk_rq_pos(req) << 9; | |
1086 | struct bio *bio_saved = req->bio; | |
1087 | dmu_buf_t **dbp; | |
1088 | int numbufs, i, err; | |
1089 | ||
1090 | /* | |
1091 | * NB: we could do this block-at-a-time, but it's nice | |
1092 | * to be reading in parallel. | |
1093 | */ | |
1094 | err = dmu_buf_hold_array(os, object, offset, size, TRUE, FTAG, | |
1095 | &numbufs, &dbp); | |
1096 | if (err) | |
1097 | return (err); | |
1098 | ||
1099 | /* | |
1100 | * Clone the bio list so the bv->bv_offset and bv->bv_len members | |
1101 | * can be safely modified. The original bio list is relinked in to | |
1102 | * the request when the function exits. This is required because | |
1103 | * some file systems blindly assume that these values will remain | |
1104 | * constant between bio_submit() and the IO completion callback. | |
1105 | */ | |
1106 | err = dmu_bio_clone(bio_saved, &req->bio); | |
1107 | if (err) | |
1108 | goto error; | |
1109 | ||
1110 | for (i = 0; i < numbufs; i++) { | |
1111 | int tocpy, didcpy, bufoff; | |
1112 | dmu_buf_t *db = dbp[i]; | |
1113 | ||
1114 | bufoff = offset - db->db_offset; | |
1115 | ASSERT3S(bufoff, >=, 0); | |
1116 | ||
1117 | tocpy = (int)MIN(db->db_size - bufoff, size); | |
1118 | if (tocpy == 0) | |
1119 | break; | |
1120 | ||
1121 | err = dmu_req_copy(db->db_data + bufoff, tocpy, &didcpy, req); | |
1122 | ||
1123 | if (didcpy < tocpy) | |
1124 | err = EIO; | |
1125 | ||
1126 | if (err) | |
1127 | break; | |
1128 | ||
1129 | size -= tocpy; | |
1130 | offset += didcpy; | |
1131 | err = 0; | |
1132 | } | |
1133 | ||
1134 | dmu_bio_put(req->bio); | |
1135 | req->bio = bio_saved; | |
1136 | error: | |
1137 | dmu_buf_rele_array(dbp, numbufs, FTAG); | |
1138 | ||
1139 | return (err); | |
1140 | } | |
1141 | ||
1142 | int | |
1143 | dmu_write_req(objset_t *os, uint64_t object, struct request *req, dmu_tx_t *tx) | |
1144 | { | |
1145 | uint64_t size = blk_rq_bytes(req); | |
1146 | uint64_t offset = blk_rq_pos(req) << 9; | |
1147 | struct bio *bio_saved = req->bio; | |
1148 | dmu_buf_t **dbp; | |
1149 | int numbufs; | |
1150 | int err = 0; | |
1151 | int i; | |
1152 | ||
1153 | if (size == 0) | |
1154 | return (0); | |
1155 | ||
1156 | err = dmu_buf_hold_array(os, object, offset, size, FALSE, FTAG, | |
1157 | &numbufs, &dbp); | |
1158 | if (err) | |
1159 | return (err); | |
1160 | ||
1161 | /* | |
1162 | * Clone the bio list so the bv->bv_offset and bv->bv_len members | |
1163 | * can be safely modified. The original bio list is relinked in to | |
1164 | * the request when the function exits. This is required because | |
1165 | * some file systems blindly assume that these values will remain | |
1166 | * constant between bio_submit() and the IO completion callback. | |
1167 | */ | |
1168 | err = dmu_bio_clone(bio_saved, &req->bio); | |
1169 | if (err) | |
1170 | goto error; | |
1171 | ||
1172 | for (i = 0; i < numbufs; i++) { | |
1173 | int tocpy, didcpy, bufoff; | |
1174 | dmu_buf_t *db = dbp[i]; | |
1175 | ||
1176 | bufoff = offset - db->db_offset; | |
1177 | ASSERT3S(bufoff, >=, 0); | |
1178 | ||
1179 | tocpy = (int)MIN(db->db_size - bufoff, size); | |
1180 | if (tocpy == 0) | |
1181 | break; | |
1182 | ||
1183 | ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); | |
1184 | ||
1185 | if (tocpy == db->db_size) | |
1186 | dmu_buf_will_fill(db, tx); | |
1187 | else | |
1188 | dmu_buf_will_dirty(db, tx); | |
1189 | ||
1190 | err = dmu_req_copy(db->db_data + bufoff, tocpy, &didcpy, req); | |
1191 | ||
1192 | if (tocpy == db->db_size) | |
1193 | dmu_buf_fill_done(db, tx); | |
1194 | ||
1195 | if (didcpy < tocpy) | |
1196 | err = EIO; | |
1197 | ||
1198 | if (err) | |
1199 | break; | |
1200 | ||
1201 | size -= tocpy; | |
1202 | offset += didcpy; | |
1203 | err = 0; | |
1204 | } | |
1205 | ||
1206 | dmu_bio_put(req->bio); | |
1207 | req->bio = bio_saved; | |
1208 | error: | |
1209 | dmu_buf_rele_array(dbp, numbufs, FTAG); | |
1210 | ||
1211 | return (err); | |
1212 | } | |
1213 | ||
1214 | int | |
1215 | dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size) | |
1216 | { | |
1217 | dmu_buf_t **dbp; | |
1218 | int numbufs, i, err; | |
1219 | xuio_t *xuio = NULL; | |
1220 | ||
1221 | /* | |
1222 | * NB: we could do this block-at-a-time, but it's nice | |
1223 | * to be reading in parallel. | |
1224 | */ | |
1225 | err = dmu_buf_hold_array(os, object, uio->uio_loffset, size, TRUE, FTAG, | |
1226 | &numbufs, &dbp); | |
1227 | if (err) | |
1228 | return (err); | |
1229 | ||
1230 | for (i = 0; i < numbufs; i++) { | |
1231 | int tocpy; | |
1232 | int bufoff; | |
1233 | dmu_buf_t *db = dbp[i]; | |
1234 | ||
1235 | ASSERT(size > 0); | |
1236 | ||
1237 | bufoff = uio->uio_loffset - db->db_offset; | |
1238 | tocpy = (int)MIN(db->db_size - bufoff, size); | |
1239 | ||
1240 | if (xuio) { | |
1241 | dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; | |
1242 | arc_buf_t *dbuf_abuf = dbi->db_buf; | |
1243 | arc_buf_t *abuf = dbuf_loan_arcbuf(dbi); | |
1244 | err = dmu_xuio_add(xuio, abuf, bufoff, tocpy); | |
1245 | if (!err) { | |
1246 | uio->uio_resid -= tocpy; | |
1247 | uio->uio_loffset += tocpy; | |
1248 | } | |
1249 | ||
1250 | if (abuf == dbuf_abuf) | |
1251 | XUIOSTAT_BUMP(xuiostat_rbuf_nocopy); | |
1252 | else | |
1253 | XUIOSTAT_BUMP(xuiostat_rbuf_copied); | |
1254 | } else { | |
1255 | err = uiomove((char *)db->db_data + bufoff, tocpy, | |
1256 | UIO_READ, uio); | |
1257 | } | |
1258 | if (err) | |
1259 | break; | |
1260 | ||
1261 | size -= tocpy; | |
1262 | } | |
1263 | dmu_buf_rele_array(dbp, numbufs, FTAG); | |
1264 | ||
1265 | return (err); | |
1266 | } | |
1267 | ||
1268 | static int | |
1269 | dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx) | |
1270 | { | |
1271 | dmu_buf_t **dbp; | |
1272 | int numbufs; | |
1273 | int err = 0; | |
1274 | int i; | |
1275 | ||
1276 | err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size, | |
1277 | FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH); | |
1278 | if (err) | |
1279 | return (err); | |
1280 | ||
1281 | for (i = 0; i < numbufs; i++) { | |
1282 | int tocpy; | |
1283 | int bufoff; | |
1284 | dmu_buf_t *db = dbp[i]; | |
1285 | ||
1286 | ASSERT(size > 0); | |
1287 | ||
1288 | bufoff = uio->uio_loffset - db->db_offset; | |
1289 | tocpy = (int)MIN(db->db_size - bufoff, size); | |
1290 | ||
1291 | ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); | |
1292 | ||
1293 | if (tocpy == db->db_size) | |
1294 | dmu_buf_will_fill(db, tx); | |
1295 | else | |
1296 | dmu_buf_will_dirty(db, tx); | |
1297 | ||
1298 | /* | |
1299 | * XXX uiomove could block forever (eg.nfs-backed | |
1300 | * pages). There needs to be a uiolockdown() function | |
1301 | * to lock the pages in memory, so that uiomove won't | |
1302 | * block. | |
1303 | */ | |
1304 | err = uiomove((char *)db->db_data + bufoff, tocpy, | |
1305 | UIO_WRITE, uio); | |
1306 | ||
1307 | if (tocpy == db->db_size) | |
1308 | dmu_buf_fill_done(db, tx); | |
1309 | ||
1310 | if (err) | |
1311 | break; | |
1312 | ||
1313 | size -= tocpy; | |
1314 | } | |
1315 | ||
1316 | dmu_buf_rele_array(dbp, numbufs, FTAG); | |
1317 | return (err); | |
1318 | } | |
1319 | ||
1320 | int | |
1321 | dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size, | |
1322 | dmu_tx_t *tx) | |
1323 | { | |
1324 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb; | |
1325 | dnode_t *dn; | |
1326 | int err; | |
1327 | ||
1328 | if (size == 0) | |
1329 | return (0); | |
1330 | ||
1331 | DB_DNODE_ENTER(db); | |
1332 | dn = DB_DNODE(db); | |
1333 | err = dmu_write_uio_dnode(dn, uio, size, tx); | |
1334 | DB_DNODE_EXIT(db); | |
1335 | ||
1336 | return (err); | |
1337 | } | |
1338 | ||
1339 | int | |
1340 | dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size, | |
1341 | dmu_tx_t *tx) | |
1342 | { | |
1343 | dnode_t *dn; | |
1344 | int err; | |
1345 | ||
1346 | if (size == 0) | |
1347 | return (0); | |
1348 | ||
1349 | err = dnode_hold(os, object, FTAG, &dn); | |
1350 | if (err) | |
1351 | return (err); | |
1352 | ||
1353 | err = dmu_write_uio_dnode(dn, uio, size, tx); | |
1354 | ||
1355 | dnode_rele(dn, FTAG); | |
1356 | ||
1357 | return (err); | |
1358 | } | |
1359 | #endif /* _KERNEL */ | |
1360 | ||
1361 | /* | |
1362 | * Allocate a loaned anonymous arc buffer. | |
1363 | */ | |
1364 | arc_buf_t * | |
1365 | dmu_request_arcbuf(dmu_buf_t *handle, int size) | |
1366 | { | |
1367 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle; | |
1368 | spa_t *spa; | |
1369 | ||
1370 | DB_GET_SPA(&spa, db); | |
1371 | return (arc_loan_buf(spa, size)); | |
1372 | } | |
1373 | ||
1374 | /* | |
1375 | * Free a loaned arc buffer. | |
1376 | */ | |
1377 | void | |
1378 | dmu_return_arcbuf(arc_buf_t *buf) | |
1379 | { | |
1380 | arc_return_buf(buf, FTAG); | |
1381 | VERIFY(arc_buf_remove_ref(buf, FTAG)); | |
1382 | } | |
1383 | ||
1384 | /* | |
1385 | * When possible directly assign passed loaned arc buffer to a dbuf. | |
1386 | * If this is not possible copy the contents of passed arc buf via | |
1387 | * dmu_write(). | |
1388 | */ | |
1389 | void | |
1390 | dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf, | |
1391 | dmu_tx_t *tx) | |
1392 | { | |
1393 | dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle; | |
1394 | dnode_t *dn; | |
1395 | dmu_buf_impl_t *db; | |
1396 | uint32_t blksz = (uint32_t)arc_buf_size(buf); | |
1397 | uint64_t blkid; | |
1398 | ||
1399 | DB_DNODE_ENTER(dbuf); | |
1400 | dn = DB_DNODE(dbuf); | |
1401 | rw_enter(&dn->dn_struct_rwlock, RW_READER); | |
1402 | blkid = dbuf_whichblock(dn, offset); | |
1403 | VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL); | |
1404 | rw_exit(&dn->dn_struct_rwlock); | |
1405 | DB_DNODE_EXIT(dbuf); | |
1406 | ||
1407 | if (offset == db->db.db_offset && blksz == db->db.db_size) { | |
1408 | dbuf_assign_arcbuf(db, buf, tx); | |
1409 | dbuf_rele(db, FTAG); | |
1410 | } else { | |
1411 | objset_t *os; | |
1412 | uint64_t object; | |
1413 | ||
1414 | DB_DNODE_ENTER(dbuf); | |
1415 | dn = DB_DNODE(dbuf); | |
1416 | os = dn->dn_objset; | |
1417 | object = dn->dn_object; | |
1418 | DB_DNODE_EXIT(dbuf); | |
1419 | ||
1420 | dbuf_rele(db, FTAG); | |
1421 | dmu_write(os, object, offset, blksz, buf->b_data, tx); | |
1422 | dmu_return_arcbuf(buf); | |
1423 | XUIOSTAT_BUMP(xuiostat_wbuf_copied); | |
1424 | } | |
1425 | } | |
1426 | ||
1427 | typedef struct { | |
1428 | dbuf_dirty_record_t *dsa_dr; | |
1429 | dmu_sync_cb_t *dsa_done; | |
1430 | zgd_t *dsa_zgd; | |
1431 | dmu_tx_t *dsa_tx; | |
1432 | } dmu_sync_arg_t; | |
1433 | ||
1434 | /* ARGSUSED */ | |
1435 | static void | |
1436 | dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg) | |
1437 | { | |
1438 | dmu_sync_arg_t *dsa = varg; | |
1439 | dmu_buf_t *db = dsa->dsa_zgd->zgd_db; | |
1440 | blkptr_t *bp = zio->io_bp; | |
1441 | ||
1442 | if (zio->io_error == 0) { | |
1443 | if (BP_IS_HOLE(bp)) { | |
1444 | /* | |
1445 | * A block of zeros may compress to a hole, but the | |
1446 | * block size still needs to be known for replay. | |
1447 | */ | |
1448 | BP_SET_LSIZE(bp, db->db_size); | |
1449 | } else { | |
1450 | ASSERT(BP_GET_LEVEL(bp) == 0); | |
1451 | bp->blk_fill = 1; | |
1452 | } | |
1453 | } | |
1454 | } | |
1455 | ||
1456 | static void | |
1457 | dmu_sync_late_arrival_ready(zio_t *zio) | |
1458 | { | |
1459 | dmu_sync_ready(zio, NULL, zio->io_private); | |
1460 | } | |
1461 | ||
1462 | /* ARGSUSED */ | |
1463 | static void | |
1464 | dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg) | |
1465 | { | |
1466 | dmu_sync_arg_t *dsa = varg; | |
1467 | dbuf_dirty_record_t *dr = dsa->dsa_dr; | |
1468 | dmu_buf_impl_t *db = dr->dr_dbuf; | |
1469 | ||
1470 | mutex_enter(&db->db_mtx); | |
1471 | ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC); | |
1472 | if (zio->io_error == 0) { | |
1473 | dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE); | |
1474 | if (dr->dt.dl.dr_nopwrite) { | |
1475 | ASSERTV(blkptr_t *bp = zio->io_bp); | |
1476 | ASSERTV(blkptr_t *bp_orig = &zio->io_bp_orig); | |
1477 | ASSERTV(uint8_t chksum = BP_GET_CHECKSUM(bp_orig)); | |
1478 | ||
1479 | ASSERT(BP_EQUAL(bp, bp_orig)); | |
1480 | ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF); | |
1481 | ASSERT(zio_checksum_table[chksum].ci_dedup); | |
1482 | } | |
1483 | dr->dt.dl.dr_overridden_by = *zio->io_bp; | |
1484 | dr->dt.dl.dr_override_state = DR_OVERRIDDEN; | |
1485 | dr->dt.dl.dr_copies = zio->io_prop.zp_copies; | |
1486 | if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by)) | |
1487 | BP_ZERO(&dr->dt.dl.dr_overridden_by); | |
1488 | } else { | |
1489 | dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; | |
1490 | } | |
1491 | cv_broadcast(&db->db_changed); | |
1492 | mutex_exit(&db->db_mtx); | |
1493 | ||
1494 | dsa->dsa_done(dsa->dsa_zgd, zio->io_error); | |
1495 | ||
1496 | kmem_free(dsa, sizeof (*dsa)); | |
1497 | } | |
1498 | ||
1499 | static void | |
1500 | dmu_sync_late_arrival_done(zio_t *zio) | |
1501 | { | |
1502 | blkptr_t *bp = zio->io_bp; | |
1503 | dmu_sync_arg_t *dsa = zio->io_private; | |
1504 | ASSERTV(blkptr_t *bp_orig = &zio->io_bp_orig); | |
1505 | ||
1506 | if (zio->io_error == 0 && !BP_IS_HOLE(bp)) { | |
1507 | /* | |
1508 | * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE) | |
1509 | * then there is nothing to do here. Otherwise, free the | |
1510 | * newly allocated block in this txg. | |
1511 | */ | |
1512 | if (zio->io_flags & ZIO_FLAG_NOPWRITE) { | |
1513 | ASSERT(BP_EQUAL(bp, bp_orig)); | |
1514 | } else { | |
1515 | ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig)); | |
1516 | ASSERT(zio->io_bp->blk_birth == zio->io_txg); | |
1517 | ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa)); | |
1518 | zio_free(zio->io_spa, zio->io_txg, zio->io_bp); | |
1519 | } | |
1520 | } | |
1521 | ||
1522 | dmu_tx_commit(dsa->dsa_tx); | |
1523 | ||
1524 | dsa->dsa_done(dsa->dsa_zgd, zio->io_error); | |
1525 | ||
1526 | kmem_free(dsa, sizeof (*dsa)); | |
1527 | } | |
1528 | ||
1529 | static int | |
1530 | dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd, | |
1531 | zio_prop_t *zp, zbookmark_t *zb) | |
1532 | { | |
1533 | dmu_sync_arg_t *dsa; | |
1534 | dmu_tx_t *tx; | |
1535 | ||
1536 | tx = dmu_tx_create(os); | |
1537 | dmu_tx_hold_space(tx, zgd->zgd_db->db_size); | |
1538 | if (dmu_tx_assign(tx, TXG_WAIT) != 0) { | |
1539 | dmu_tx_abort(tx); | |
1540 | /* Make zl_get_data do txg_waited_synced() */ | |
1541 | return (SET_ERROR(EIO)); | |
1542 | } | |
1543 | ||
1544 | dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_PUSHPAGE); | |
1545 | dsa->dsa_dr = NULL; | |
1546 | dsa->dsa_done = done; | |
1547 | dsa->dsa_zgd = zgd; | |
1548 | dsa->dsa_tx = tx; | |
1549 | ||
1550 | zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp, | |
1551 | zgd->zgd_db->db_data, zgd->zgd_db->db_size, zp, | |
1552 | dmu_sync_late_arrival_ready, dmu_sync_late_arrival_done, dsa, | |
1553 | ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL | ZIO_FLAG_FASTWRITE, zb)); | |
1554 | ||
1555 | return (0); | |
1556 | } | |
1557 | ||
1558 | /* | |
1559 | * Intent log support: sync the block associated with db to disk. | |
1560 | * N.B. and XXX: the caller is responsible for making sure that the | |
1561 | * data isn't changing while dmu_sync() is writing it. | |
1562 | * | |
1563 | * Return values: | |
1564 | * | |
1565 | * EEXIST: this txg has already been synced, so there's nothing to do. | |
1566 | * The caller should not log the write. | |
1567 | * | |
1568 | * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do. | |
1569 | * The caller should not log the write. | |
1570 | * | |
1571 | * EALREADY: this block is already in the process of being synced. | |
1572 | * The caller should track its progress (somehow). | |
1573 | * | |
1574 | * EIO: could not do the I/O. | |
1575 | * The caller should do a txg_wait_synced(). | |
1576 | * | |
1577 | * 0: the I/O has been initiated. | |
1578 | * The caller should log this blkptr in the done callback. | |
1579 | * It is possible that the I/O will fail, in which case | |
1580 | * the error will be reported to the done callback and | |
1581 | * propagated to pio from zio_done(). | |
1582 | */ | |
1583 | int | |
1584 | dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd) | |
1585 | { | |
1586 | blkptr_t *bp = zgd->zgd_bp; | |
1587 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db; | |
1588 | objset_t *os = db->db_objset; | |
1589 | dsl_dataset_t *ds = os->os_dsl_dataset; | |
1590 | dbuf_dirty_record_t *dr; | |
1591 | dmu_sync_arg_t *dsa; | |
1592 | zbookmark_t zb; | |
1593 | zio_prop_t zp; | |
1594 | dnode_t *dn; | |
1595 | ||
1596 | ASSERT(pio != NULL); | |
1597 | ASSERT(txg != 0); | |
1598 | ||
1599 | SET_BOOKMARK(&zb, ds->ds_object, | |
1600 | db->db.db_object, db->db_level, db->db_blkid); | |
1601 | ||
1602 | DB_DNODE_ENTER(db); | |
1603 | dn = DB_DNODE(db); | |
1604 | dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp); | |
1605 | DB_DNODE_EXIT(db); | |
1606 | ||
1607 | /* | |
1608 | * If we're frozen (running ziltest), we always need to generate a bp. | |
1609 | */ | |
1610 | if (txg > spa_freeze_txg(os->os_spa)) | |
1611 | return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); | |
1612 | ||
1613 | /* | |
1614 | * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf() | |
1615 | * and us. If we determine that this txg is not yet syncing, | |
1616 | * but it begins to sync a moment later, that's OK because the | |
1617 | * sync thread will block in dbuf_sync_leaf() until we drop db_mtx. | |
1618 | */ | |
1619 | mutex_enter(&db->db_mtx); | |
1620 | ||
1621 | if (txg <= spa_last_synced_txg(os->os_spa)) { | |
1622 | /* | |
1623 | * This txg has already synced. There's nothing to do. | |
1624 | */ | |
1625 | mutex_exit(&db->db_mtx); | |
1626 | return (SET_ERROR(EEXIST)); | |
1627 | } | |
1628 | ||
1629 | if (txg <= spa_syncing_txg(os->os_spa)) { | |
1630 | /* | |
1631 | * This txg is currently syncing, so we can't mess with | |
1632 | * the dirty record anymore; just write a new log block. | |
1633 | */ | |
1634 | mutex_exit(&db->db_mtx); | |
1635 | return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); | |
1636 | } | |
1637 | ||
1638 | dr = db->db_last_dirty; | |
1639 | while (dr && dr->dr_txg != txg) | |
1640 | dr = dr->dr_next; | |
1641 | ||
1642 | if (dr == NULL) { | |
1643 | /* | |
1644 | * There's no dr for this dbuf, so it must have been freed. | |
1645 | * There's no need to log writes to freed blocks, so we're done. | |
1646 | */ | |
1647 | mutex_exit(&db->db_mtx); | |
1648 | return (SET_ERROR(ENOENT)); | |
1649 | } | |
1650 | ||
1651 | ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg); | |
1652 | ||
1653 | /* | |
1654 | * Assume the on-disk data is X, the current syncing data is Y, | |
1655 | * and the current in-memory data is Z (currently in dmu_sync). | |
1656 | * X and Z are identical but Y is has been modified. Normally, | |
1657 | * when X and Z are the same we will perform a nopwrite but if Y | |
1658 | * is different we must disable nopwrite since the resulting write | |
1659 | * of Y to disk can free the block containing X. If we allowed a | |
1660 | * nopwrite to occur the block pointing to Z would reference a freed | |
1661 | * block. Since this is a rare case we simplify this by disabling | |
1662 | * nopwrite if the current dmu_sync-ing dbuf has been modified in | |
1663 | * a previous transaction. | |
1664 | */ | |
1665 | if (dr->dr_next) | |
1666 | zp.zp_nopwrite = B_FALSE; | |
1667 | ||
1668 | ASSERT(dr->dr_txg == txg); | |
1669 | if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC || | |
1670 | dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { | |
1671 | /* | |
1672 | * We have already issued a sync write for this buffer, | |
1673 | * or this buffer has already been synced. It could not | |
1674 | * have been dirtied since, or we would have cleared the state. | |
1675 | */ | |
1676 | mutex_exit(&db->db_mtx); | |
1677 | return (SET_ERROR(EALREADY)); | |
1678 | } | |
1679 | ||
1680 | ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); | |
1681 | dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC; | |
1682 | mutex_exit(&db->db_mtx); | |
1683 | ||
1684 | dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_PUSHPAGE); | |
1685 | dsa->dsa_dr = dr; | |
1686 | dsa->dsa_done = done; | |
1687 | dsa->dsa_zgd = zgd; | |
1688 | dsa->dsa_tx = NULL; | |
1689 | ||
1690 | zio_nowait(arc_write(pio, os->os_spa, txg, | |
1691 | bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db), | |
1692 | DBUF_IS_L2COMPRESSIBLE(db), &zp, dmu_sync_ready, dmu_sync_done, | |
1693 | dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL | ZIO_FLAG_FASTWRITE, &zb)); | |
1694 | ||
1695 | return (0); | |
1696 | } | |
1697 | ||
1698 | int | |
1699 | dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs, | |
1700 | dmu_tx_t *tx) | |
1701 | { | |
1702 | dnode_t *dn; | |
1703 | int err; | |
1704 | ||
1705 | err = dnode_hold(os, object, FTAG, &dn); | |
1706 | if (err) | |
1707 | return (err); | |
1708 | err = dnode_set_blksz(dn, size, ibs, tx); | |
1709 | dnode_rele(dn, FTAG); | |
1710 | return (err); | |
1711 | } | |
1712 | ||
1713 | void | |
1714 | dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum, | |
1715 | dmu_tx_t *tx) | |
1716 | { | |
1717 | dnode_t *dn; | |
1718 | ||
1719 | /* XXX assumes dnode_hold will not get an i/o error */ | |
1720 | (void) dnode_hold(os, object, FTAG, &dn); | |
1721 | ASSERT(checksum < ZIO_CHECKSUM_FUNCTIONS); | |
1722 | dn->dn_checksum = checksum; | |
1723 | dnode_setdirty(dn, tx); | |
1724 | dnode_rele(dn, FTAG); | |
1725 | } | |
1726 | ||
1727 | void | |
1728 | dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress, | |
1729 | dmu_tx_t *tx) | |
1730 | { | |
1731 | dnode_t *dn; | |
1732 | ||
1733 | /* XXX assumes dnode_hold will not get an i/o error */ | |
1734 | (void) dnode_hold(os, object, FTAG, &dn); | |
1735 | ASSERT(compress < ZIO_COMPRESS_FUNCTIONS); | |
1736 | dn->dn_compress = compress; | |
1737 | dnode_setdirty(dn, tx); | |
1738 | dnode_rele(dn, FTAG); | |
1739 | } | |
1740 | ||
1741 | int zfs_mdcomp_disable = 0; | |
1742 | ||
1743 | void | |
1744 | dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp) | |
1745 | { | |
1746 | dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET; | |
1747 | boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) || | |
1748 | (wp & WP_SPILL)); | |
1749 | enum zio_checksum checksum = os->os_checksum; | |
1750 | enum zio_compress compress = os->os_compress; | |
1751 | enum zio_checksum dedup_checksum = os->os_dedup_checksum; | |
1752 | boolean_t dedup = B_FALSE; | |
1753 | boolean_t nopwrite = B_FALSE; | |
1754 | boolean_t dedup_verify = os->os_dedup_verify; | |
1755 | int copies = os->os_copies; | |
1756 | ||
1757 | /* | |
1758 | * We maintain different write policies for each of the following | |
1759 | * types of data: | |
1760 | * 1. metadata | |
1761 | * 2. preallocated blocks (i.e. level-0 blocks of a dump device) | |
1762 | * 3. all other level 0 blocks | |
1763 | */ | |
1764 | if (ismd) { | |
1765 | /* | |
1766 | * XXX -- we should design a compression algorithm | |
1767 | * that specializes in arrays of bps. | |
1768 | */ | |
1769 | compress = zfs_mdcomp_disable ? ZIO_COMPRESS_EMPTY : | |
1770 | ZIO_COMPRESS_LZJB; | |
1771 | ||
1772 | /* | |
1773 | * Metadata always gets checksummed. If the data | |
1774 | * checksum is multi-bit correctable, and it's not a | |
1775 | * ZBT-style checksum, then it's suitable for metadata | |
1776 | * as well. Otherwise, the metadata checksum defaults | |
1777 | * to fletcher4. | |
1778 | */ | |
1779 | if (zio_checksum_table[checksum].ci_correctable < 1 || | |
1780 | zio_checksum_table[checksum].ci_eck) | |
1781 | checksum = ZIO_CHECKSUM_FLETCHER_4; | |
1782 | } else if (wp & WP_NOFILL) { | |
1783 | ASSERT(level == 0); | |
1784 | ||
1785 | /* | |
1786 | * If we're writing preallocated blocks, we aren't actually | |
1787 | * writing them so don't set any policy properties. These | |
1788 | * blocks are currently only used by an external subsystem | |
1789 | * outside of zfs (i.e. dump) and not written by the zio | |
1790 | * pipeline. | |
1791 | */ | |
1792 | compress = ZIO_COMPRESS_OFF; | |
1793 | checksum = ZIO_CHECKSUM_OFF; | |
1794 | } else { | |
1795 | compress = zio_compress_select(dn->dn_compress, compress); | |
1796 | ||
1797 | checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ? | |
1798 | zio_checksum_select(dn->dn_checksum, checksum) : | |
1799 | dedup_checksum; | |
1800 | ||
1801 | /* | |
1802 | * Determine dedup setting. If we are in dmu_sync(), | |
1803 | * we won't actually dedup now because that's all | |
1804 | * done in syncing context; but we do want to use the | |
1805 | * dedup checkum. If the checksum is not strong | |
1806 | * enough to ensure unique signatures, force | |
1807 | * dedup_verify. | |
1808 | */ | |
1809 | if (dedup_checksum != ZIO_CHECKSUM_OFF) { | |
1810 | dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE; | |
1811 | if (!zio_checksum_table[checksum].ci_dedup) | |
1812 | dedup_verify = B_TRUE; | |
1813 | } | |
1814 | ||
1815 | /* | |
1816 | * Enable nopwrite if we have a cryptographically secure | |
1817 | * checksum that has no known collisions (i.e. SHA-256) | |
1818 | * and compression is enabled. We don't enable nopwrite if | |
1819 | * dedup is enabled as the two features are mutually exclusive. | |
1820 | */ | |
1821 | nopwrite = (!dedup && zio_checksum_table[checksum].ci_dedup && | |
1822 | compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled); | |
1823 | } | |
1824 | ||
1825 | zp->zp_checksum = checksum; | |
1826 | zp->zp_compress = compress; | |
1827 | zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type; | |
1828 | zp->zp_level = level; | |
1829 | zp->zp_copies = MIN(copies + ismd, spa_max_replication(os->os_spa)); | |
1830 | zp->zp_dedup = dedup; | |
1831 | zp->zp_dedup_verify = dedup && dedup_verify; | |
1832 | zp->zp_nopwrite = nopwrite; | |
1833 | } | |
1834 | ||
1835 | int | |
1836 | dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off) | |
1837 | { | |
1838 | dnode_t *dn; | |
1839 | int i, err; | |
1840 | ||
1841 | err = dnode_hold(os, object, FTAG, &dn); | |
1842 | if (err) | |
1843 | return (err); | |
1844 | /* | |
1845 | * Sync any current changes before | |
1846 | * we go trundling through the block pointers. | |
1847 | */ | |
1848 | for (i = 0; i < TXG_SIZE; i++) { | |
1849 | if (list_link_active(&dn->dn_dirty_link[i])) | |
1850 | break; | |
1851 | } | |
1852 | if (i != TXG_SIZE) { | |
1853 | dnode_rele(dn, FTAG); | |
1854 | txg_wait_synced(dmu_objset_pool(os), 0); | |
1855 | err = dnode_hold(os, object, FTAG, &dn); | |
1856 | if (err) | |
1857 | return (err); | |
1858 | } | |
1859 | ||
1860 | err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0); | |
1861 | dnode_rele(dn, FTAG); | |
1862 | ||
1863 | return (err); | |
1864 | } | |
1865 | ||
1866 | void | |
1867 | __dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi) | |
1868 | { | |
1869 | dnode_phys_t *dnp = dn->dn_phys; | |
1870 | int i; | |
1871 | ||
1872 | doi->doi_data_block_size = dn->dn_datablksz; | |
1873 | doi->doi_metadata_block_size = dn->dn_indblkshift ? | |
1874 | 1ULL << dn->dn_indblkshift : 0; | |
1875 | doi->doi_type = dn->dn_type; | |
1876 | doi->doi_bonus_type = dn->dn_bonustype; | |
1877 | doi->doi_bonus_size = dn->dn_bonuslen; | |
1878 | doi->doi_indirection = dn->dn_nlevels; | |
1879 | doi->doi_checksum = dn->dn_checksum; | |
1880 | doi->doi_compress = dn->dn_compress; | |
1881 | doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9; | |
1882 | doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz; | |
1883 | doi->doi_fill_count = 0; | |
1884 | for (i = 0; i < dnp->dn_nblkptr; i++) | |
1885 | doi->doi_fill_count += dnp->dn_blkptr[i].blk_fill; | |
1886 | } | |
1887 | ||
1888 | void | |
1889 | dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi) | |
1890 | { | |
1891 | rw_enter(&dn->dn_struct_rwlock, RW_READER); | |
1892 | mutex_enter(&dn->dn_mtx); | |
1893 | ||
1894 | __dmu_object_info_from_dnode(dn, doi); | |
1895 | ||
1896 | mutex_exit(&dn->dn_mtx); | |
1897 | rw_exit(&dn->dn_struct_rwlock); | |
1898 | } | |
1899 | ||
1900 | /* | |
1901 | * Get information on a DMU object. | |
1902 | * If doi is NULL, just indicates whether the object exists. | |
1903 | */ | |
1904 | int | |
1905 | dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi) | |
1906 | { | |
1907 | dnode_t *dn; | |
1908 | int err = dnode_hold(os, object, FTAG, &dn); | |
1909 | ||
1910 | if (err) | |
1911 | return (err); | |
1912 | ||
1913 | if (doi != NULL) | |
1914 | dmu_object_info_from_dnode(dn, doi); | |
1915 | ||
1916 | dnode_rele(dn, FTAG); | |
1917 | return (0); | |
1918 | } | |
1919 | ||
1920 | /* | |
1921 | * As above, but faster; can be used when you have a held dbuf in hand. | |
1922 | */ | |
1923 | void | |
1924 | dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi) | |
1925 | { | |
1926 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
1927 | ||
1928 | DB_DNODE_ENTER(db); | |
1929 | dmu_object_info_from_dnode(DB_DNODE(db), doi); | |
1930 | DB_DNODE_EXIT(db); | |
1931 | } | |
1932 | ||
1933 | /* | |
1934 | * Faster still when you only care about the size. | |
1935 | * This is specifically optimized for zfs_getattr(). | |
1936 | */ | |
1937 | void | |
1938 | dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize, | |
1939 | u_longlong_t *nblk512) | |
1940 | { | |
1941 | dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; | |
1942 | dnode_t *dn; | |
1943 | ||
1944 | DB_DNODE_ENTER(db); | |
1945 | dn = DB_DNODE(db); | |
1946 | ||
1947 | *blksize = dn->dn_datablksz; | |
1948 | /* add 1 for dnode space */ | |
1949 | *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >> | |
1950 | SPA_MINBLOCKSHIFT) + 1; | |
1951 | DB_DNODE_EXIT(db); | |
1952 | } | |
1953 | ||
1954 | void | |
1955 | byteswap_uint64_array(void *vbuf, size_t size) | |
1956 | { | |
1957 | uint64_t *buf = vbuf; | |
1958 | size_t count = size >> 3; | |
1959 | int i; | |
1960 | ||
1961 | ASSERT((size & 7) == 0); | |
1962 | ||
1963 | for (i = 0; i < count; i++) | |
1964 | buf[i] = BSWAP_64(buf[i]); | |
1965 | } | |
1966 | ||
1967 | void | |
1968 | byteswap_uint32_array(void *vbuf, size_t size) | |
1969 | { | |
1970 | uint32_t *buf = vbuf; | |
1971 | size_t count = size >> 2; | |
1972 | int i; | |
1973 | ||
1974 | ASSERT((size & 3) == 0); | |
1975 | ||
1976 | for (i = 0; i < count; i++) | |
1977 | buf[i] = BSWAP_32(buf[i]); | |
1978 | } | |
1979 | ||
1980 | void | |
1981 | byteswap_uint16_array(void *vbuf, size_t size) | |
1982 | { | |
1983 | uint16_t *buf = vbuf; | |
1984 | size_t count = size >> 1; | |
1985 | int i; | |
1986 | ||
1987 | ASSERT((size & 1) == 0); | |
1988 | ||
1989 | for (i = 0; i < count; i++) | |
1990 | buf[i] = BSWAP_16(buf[i]); | |
1991 | } | |
1992 | ||
1993 | /* ARGSUSED */ | |
1994 | void | |
1995 | byteswap_uint8_array(void *vbuf, size_t size) | |
1996 | { | |
1997 | } | |
1998 | ||
1999 | void | |
2000 | dmu_init(void) | |
2001 | { | |
2002 | zfs_dbgmsg_init(); | |
2003 | sa_cache_init(); | |
2004 | xuio_stat_init(); | |
2005 | dmu_objset_init(); | |
2006 | dnode_init(); | |
2007 | dbuf_init(); | |
2008 | zfetch_init(); | |
2009 | dmu_tx_init(); | |
2010 | l2arc_init(); | |
2011 | arc_init(); | |
2012 | } | |
2013 | ||
2014 | void | |
2015 | dmu_fini(void) | |
2016 | { | |
2017 | arc_fini(); /* arc depends on l2arc, so arc must go first */ | |
2018 | l2arc_fini(); | |
2019 | dmu_tx_fini(); | |
2020 | zfetch_fini(); | |
2021 | dbuf_fini(); | |
2022 | dnode_fini(); | |
2023 | dmu_objset_fini(); | |
2024 | xuio_stat_fini(); | |
2025 | sa_cache_fini(); | |
2026 | zfs_dbgmsg_fini(); | |
2027 | } | |
2028 | ||
2029 | #if defined(_KERNEL) && defined(HAVE_SPL) | |
2030 | EXPORT_SYMBOL(dmu_bonus_hold); | |
2031 | EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus); | |
2032 | EXPORT_SYMBOL(dmu_buf_rele_array); | |
2033 | EXPORT_SYMBOL(dmu_prefetch); | |
2034 | EXPORT_SYMBOL(dmu_free_range); | |
2035 | EXPORT_SYMBOL(dmu_free_long_range); | |
2036 | EXPORT_SYMBOL(dmu_free_long_object); | |
2037 | EXPORT_SYMBOL(dmu_read); | |
2038 | EXPORT_SYMBOL(dmu_write); | |
2039 | EXPORT_SYMBOL(dmu_prealloc); | |
2040 | EXPORT_SYMBOL(dmu_object_info); | |
2041 | EXPORT_SYMBOL(dmu_object_info_from_dnode); | |
2042 | EXPORT_SYMBOL(dmu_object_info_from_db); | |
2043 | EXPORT_SYMBOL(dmu_object_size_from_db); | |
2044 | EXPORT_SYMBOL(dmu_object_set_blocksize); | |
2045 | EXPORT_SYMBOL(dmu_object_set_checksum); | |
2046 | EXPORT_SYMBOL(dmu_object_set_compress); | |
2047 | EXPORT_SYMBOL(dmu_write_policy); | |
2048 | EXPORT_SYMBOL(dmu_sync); | |
2049 | EXPORT_SYMBOL(dmu_request_arcbuf); | |
2050 | EXPORT_SYMBOL(dmu_return_arcbuf); | |
2051 | EXPORT_SYMBOL(dmu_assign_arcbuf); | |
2052 | EXPORT_SYMBOL(dmu_buf_hold); | |
2053 | EXPORT_SYMBOL(dmu_ot); | |
2054 | ||
2055 | module_param(zfs_mdcomp_disable, int, 0644); | |
2056 | MODULE_PARM_DESC(zfs_mdcomp_disable, "Disable meta data compression"); | |
2057 | ||
2058 | module_param(zfs_nopwrite_enabled, int, 0644); | |
2059 | MODULE_PARM_DESC(zfs_nopwrite_enabled, "Enable NOP writes"); | |
2060 | ||
2061 | #endif |