Print this page
usr/src/uts/common/fs/zfs/ddt.c
| Split |
Close |
| Expand all |
| Collapse all |
--- old/usr/src/uts/common/fs/zfs/dmu.c
+++ new/usr/src/uts/common/fs/zfs/dmu.c
1 1 /*
2 2 * CDDL HEADER START
3 3 *
4 4 * The contents of this file are subject to the terms of the
5 5 * Common Development and Distribution License (the "License").
6 6 * You may not use this file except in compliance with the License.
7 7 *
8 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 9 * or http://www.opensolaris.org/os/licensing.
10 10 * See the License for the specific language governing permissions
11 11 * and limitations under the License.
12 12 *
13 13 * When distributing Covered Code, include this CDDL HEADER in each
14 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 15 * If applicable, add the following below this CDDL HEADER, with the
16 16 * fields enclosed by brackets "[]" replaced with your own identifying
17 17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 18 *
19 19 * CDDL HEADER END
20 20 */
21 21 /*
22 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 23 * Copyright (c) 2011, 2017 by Delphix. All rights reserved.
24 24 */
25 25 /* Copyright (c) 2013 by Saso Kiselkov. All rights reserved. */
26 26 /* Copyright (c) 2013, Joyent, Inc. All rights reserved. */
27 27 /* Copyright 2016 Nexenta Systems, Inc. All rights reserved. */
28 28
29 29 #include <sys/dmu.h>
30 30 #include <sys/dmu_impl.h>
31 31 #include <sys/dmu_tx.h>
32 32 #include <sys/dbuf.h>
33 33 #include <sys/dnode.h>
34 34 #include <sys/zfs_context.h>
35 35 #include <sys/dmu_objset.h>
36 36 #include <sys/dmu_traverse.h>
37 37 #include <sys/dsl_dataset.h>
|
↓ open down ↓ |
37 lines elided |
↑ open up ↑ |
38 38 #include <sys/dsl_dir.h>
39 39 #include <sys/dsl_pool.h>
40 40 #include <sys/dsl_synctask.h>
41 41 #include <sys/dsl_prop.h>
42 42 #include <sys/dmu_zfetch.h>
43 43 #include <sys/zfs_ioctl.h>
44 44 #include <sys/zap.h>
45 45 #include <sys/zio_checksum.h>
46 46 #include <sys/zio_compress.h>
47 47 #include <sys/sa.h>
48 +#include <sys/spa_impl.h>
48 49 #include <sys/zfeature.h>
49 50 #include <sys/abd.h>
50 51 #ifdef _KERNEL
51 52 #include <sys/vmsystm.h>
52 53 #include <sys/zfs_znode.h>
54 +#include <sys/zfs_vfsops.h>
53 55 #endif
56 +#include <sys/special.h>
54 57
55 58 /*
56 59 * Enable/disable nopwrite feature.
57 60 */
58 61 int zfs_nopwrite_enabled = 1;
59 62
60 63 /*
61 64 * Tunable to control percentage of dirtied blocks from frees in one TXG.
62 65 * After this threshold is crossed, additional dirty blocks from frees
63 66 * wait until the next TXG.
64 67 * A value of zero will disable this throttle.
65 68 */
66 69 uint32_t zfs_per_txg_dirty_frees_percent = 30;
67 70
68 -/*
69 - * This can be used for testing, to ensure that certain actions happen
70 - * while in the middle of a remap (which might otherwise complete too
71 - * quickly).
72 - */
73 -int zfs_object_remap_one_indirect_delay_ticks = 0;
74 -
75 71 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
76 - { DMU_BSWAP_UINT8, TRUE, "unallocated" },
77 - { DMU_BSWAP_ZAP, TRUE, "object directory" },
78 - { DMU_BSWAP_UINT64, TRUE, "object array" },
79 - { DMU_BSWAP_UINT8, TRUE, "packed nvlist" },
80 - { DMU_BSWAP_UINT64, TRUE, "packed nvlist size" },
81 - { DMU_BSWAP_UINT64, TRUE, "bpobj" },
82 - { DMU_BSWAP_UINT64, TRUE, "bpobj header" },
83 - { DMU_BSWAP_UINT64, TRUE, "SPA space map header" },
84 - { DMU_BSWAP_UINT64, TRUE, "SPA space map" },
85 - { DMU_BSWAP_UINT64, TRUE, "ZIL intent log" },
86 - { DMU_BSWAP_DNODE, TRUE, "DMU dnode" },
87 - { DMU_BSWAP_OBJSET, TRUE, "DMU objset" },
88 - { DMU_BSWAP_UINT64, TRUE, "DSL directory" },
89 - { DMU_BSWAP_ZAP, TRUE, "DSL directory child map"},
90 - { DMU_BSWAP_ZAP, TRUE, "DSL dataset snap map" },
91 - { DMU_BSWAP_ZAP, TRUE, "DSL props" },
92 - { DMU_BSWAP_UINT64, TRUE, "DSL dataset" },
93 - { DMU_BSWAP_ZNODE, TRUE, "ZFS znode" },
94 - { DMU_BSWAP_OLDACL, TRUE, "ZFS V0 ACL" },
95 - { DMU_BSWAP_UINT8, FALSE, "ZFS plain file" },
96 - { DMU_BSWAP_ZAP, TRUE, "ZFS directory" },
97 - { DMU_BSWAP_ZAP, TRUE, "ZFS master node" },
98 - { DMU_BSWAP_ZAP, TRUE, "ZFS delete queue" },
99 - { DMU_BSWAP_UINT8, FALSE, "zvol object" },
100 - { DMU_BSWAP_ZAP, TRUE, "zvol prop" },
101 - { DMU_BSWAP_UINT8, FALSE, "other uint8[]" },
102 - { DMU_BSWAP_UINT64, FALSE, "other uint64[]" },
103 - { DMU_BSWAP_ZAP, TRUE, "other ZAP" },
104 - { DMU_BSWAP_ZAP, TRUE, "persistent error log" },
105 - { DMU_BSWAP_UINT8, TRUE, "SPA history" },
106 - { DMU_BSWAP_UINT64, TRUE, "SPA history offsets" },
107 - { DMU_BSWAP_ZAP, TRUE, "Pool properties" },
108 - { DMU_BSWAP_ZAP, TRUE, "DSL permissions" },
109 - { DMU_BSWAP_ACL, TRUE, "ZFS ACL" },
110 - { DMU_BSWAP_UINT8, TRUE, "ZFS SYSACL" },
111 - { DMU_BSWAP_UINT8, TRUE, "FUID table" },
112 - { DMU_BSWAP_UINT64, TRUE, "FUID table size" },
113 - { DMU_BSWAP_ZAP, TRUE, "DSL dataset next clones"},
114 - { DMU_BSWAP_ZAP, TRUE, "scan work queue" },
115 - { DMU_BSWAP_ZAP, TRUE, "ZFS user/group used" },
116 - { DMU_BSWAP_ZAP, TRUE, "ZFS user/group quota" },
117 - { DMU_BSWAP_ZAP, TRUE, "snapshot refcount tags"},
118 - { DMU_BSWAP_ZAP, TRUE, "DDT ZAP algorithm" },
119 - { DMU_BSWAP_ZAP, TRUE, "DDT statistics" },
120 - { DMU_BSWAP_UINT8, TRUE, "System attributes" },
121 - { DMU_BSWAP_ZAP, TRUE, "SA master node" },
122 - { DMU_BSWAP_ZAP, TRUE, "SA attr registration" },
123 - { DMU_BSWAP_ZAP, TRUE, "SA attr layouts" },
124 - { DMU_BSWAP_ZAP, TRUE, "scan translations" },
125 - { DMU_BSWAP_UINT8, FALSE, "deduplicated block" },
126 - { DMU_BSWAP_ZAP, TRUE, "DSL deadlist map" },
127 - { DMU_BSWAP_UINT64, TRUE, "DSL deadlist map hdr" },
128 - { DMU_BSWAP_ZAP, TRUE, "DSL dir clones" },
129 - { DMU_BSWAP_UINT64, TRUE, "bpobj subobj" }
72 + { DMU_BSWAP_UINT8, TRUE, FALSE, "unallocated" },
73 + { DMU_BSWAP_ZAP, TRUE, TRUE, "object directory" },
74 + { DMU_BSWAP_UINT64, TRUE, TRUE, "object array" },
75 + { DMU_BSWAP_UINT8, TRUE, FALSE, "packed nvlist" },
76 + { DMU_BSWAP_UINT64, TRUE, FALSE, "packed nvlist size" },
77 + { DMU_BSWAP_UINT64, TRUE, FALSE, "bpobj" },
78 + { DMU_BSWAP_UINT64, TRUE, FALSE, "bpobj header" },
79 + { DMU_BSWAP_UINT64, TRUE, FALSE, "SPA space map header" },
80 + { DMU_BSWAP_UINT64, TRUE, FALSE, "SPA space map" },
81 + { DMU_BSWAP_UINT64, TRUE, FALSE, "ZIL intent log" },
82 + { DMU_BSWAP_DNODE, TRUE, FALSE, "DMU dnode" },
83 + { DMU_BSWAP_OBJSET, TRUE, TRUE, "DMU objset" },
84 + { DMU_BSWAP_UINT64, TRUE, TRUE, "DSL directory" },
85 + { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL directory child map" },
86 + { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL dataset snap map" },
87 + { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL props" },
88 + { DMU_BSWAP_UINT64, TRUE, TRUE, "DSL dataset" },
89 + { DMU_BSWAP_ZNODE, TRUE, FALSE, "ZFS znode" },
90 + { DMU_BSWAP_OLDACL, TRUE, FALSE, "ZFS V0 ACL" },
91 + { DMU_BSWAP_UINT8, FALSE, FALSE, "ZFS plain file" },
92 + { DMU_BSWAP_ZAP, TRUE, FALSE, "ZFS directory" },
93 + { DMU_BSWAP_ZAP, TRUE, FALSE, "ZFS master node" },
94 + { DMU_BSWAP_ZAP, TRUE, FALSE, "ZFS delete queue" },
95 + { DMU_BSWAP_UINT8, FALSE, FALSE, "zvol object" },
96 + { DMU_BSWAP_ZAP, TRUE, TRUE, "zvol prop" },
97 + { DMU_BSWAP_UINT8, FALSE, FALSE, "other uint8[]" },
98 + { DMU_BSWAP_UINT64, FALSE, FALSE, "other uint64[]" },
99 + { DMU_BSWAP_ZAP, TRUE, FALSE, "other ZAP" },
100 + { DMU_BSWAP_ZAP, TRUE, FALSE, "persistent error log" },
101 + { DMU_BSWAP_UINT8, TRUE, FALSE, "SPA history" },
102 + { DMU_BSWAP_UINT64, TRUE, FALSE, "SPA history offsets" },
103 + { DMU_BSWAP_ZAP, TRUE, TRUE, "Pool properties" },
104 + { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL permissions" },
105 + { DMU_BSWAP_ACL, TRUE, FALSE, "ZFS ACL" },
106 + { DMU_BSWAP_UINT8, TRUE, FALSE, "ZFS SYSACL" },
107 + { DMU_BSWAP_UINT8, TRUE, FALSE, "FUID table" },
108 + { DMU_BSWAP_UINT64, TRUE, FALSE, "FUID table size" },
109 + { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL dataset next clones" },
110 + { DMU_BSWAP_ZAP, TRUE, FALSE, "scan work queue" },
111 + { DMU_BSWAP_ZAP, TRUE, FALSE, "ZFS user/group used" },
112 + { DMU_BSWAP_ZAP, TRUE, FALSE, "ZFS user/group quota" },
113 + { DMU_BSWAP_ZAP, TRUE, TRUE, "snapshot refcount tags" },
114 + { DMU_BSWAP_ZAP, TRUE, FALSE, "DDT ZAP algorithm" },
115 + { DMU_BSWAP_ZAP, TRUE, FALSE, "DDT statistics" },
116 + { DMU_BSWAP_UINT8, TRUE, FALSE, "System attributes" },
117 + { DMU_BSWAP_ZAP, TRUE, FALSE, "SA master node" },
118 + { DMU_BSWAP_ZAP, TRUE, FALSE, "SA attr registration" },
119 + { DMU_BSWAP_ZAP, TRUE, FALSE, "SA attr layouts" },
120 + { DMU_BSWAP_ZAP, TRUE, FALSE, "scan translations" },
121 + { DMU_BSWAP_UINT8, FALSE, FALSE, "deduplicated block" },
122 + { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL deadlist map" },
123 + { DMU_BSWAP_UINT64, TRUE, TRUE, "DSL deadlist map hdr" },
124 + { DMU_BSWAP_ZAP, TRUE, TRUE, "DSL dir clones" },
125 + { DMU_BSWAP_UINT64, TRUE, FALSE, "bpobj subobj" }
130 126 };
131 127
132 128 const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
133 129 { byteswap_uint8_array, "uint8" },
134 130 { byteswap_uint16_array, "uint16" },
135 131 { byteswap_uint32_array, "uint32" },
136 132 { byteswap_uint64_array, "uint64" },
137 133 { zap_byteswap, "zap" },
138 134 { dnode_buf_byteswap, "dnode" },
139 135 { dmu_objset_byteswap, "objset" },
140 136 { zfs_znode_byteswap, "znode" },
141 137 { zfs_oldacl_byteswap, "oldacl" },
142 138 { zfs_acl_byteswap, "acl" }
143 139 };
144 140
145 141 int
146 142 dmu_buf_hold_noread_by_dnode(dnode_t *dn, uint64_t offset,
147 143 void *tag, dmu_buf_t **dbp)
148 144 {
149 145 uint64_t blkid;
150 146 dmu_buf_impl_t *db;
151 147
152 148 blkid = dbuf_whichblock(dn, 0, offset);
153 149 rw_enter(&dn->dn_struct_rwlock, RW_READER);
154 150 db = dbuf_hold(dn, blkid, tag);
155 151 rw_exit(&dn->dn_struct_rwlock);
156 152
157 153 if (db == NULL) {
158 154 *dbp = NULL;
159 155 return (SET_ERROR(EIO));
160 156 }
161 157
162 158 *dbp = &db->db;
163 159 return (0);
164 160 }
165 161 int
166 162 dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset,
167 163 void *tag, dmu_buf_t **dbp)
168 164 {
169 165 dnode_t *dn;
170 166 uint64_t blkid;
171 167 dmu_buf_impl_t *db;
172 168 int err;
173 169
174 170 err = dnode_hold(os, object, FTAG, &dn);
175 171 if (err)
176 172 return (err);
177 173 blkid = dbuf_whichblock(dn, 0, offset);
178 174 rw_enter(&dn->dn_struct_rwlock, RW_READER);
179 175 db = dbuf_hold(dn, blkid, tag);
180 176 rw_exit(&dn->dn_struct_rwlock);
181 177 dnode_rele(dn, FTAG);
182 178
183 179 if (db == NULL) {
184 180 *dbp = NULL;
185 181 return (SET_ERROR(EIO));
186 182 }
187 183
188 184 *dbp = &db->db;
189 185 return (err);
190 186 }
191 187
192 188 int
193 189 dmu_buf_hold_by_dnode(dnode_t *dn, uint64_t offset,
194 190 void *tag, dmu_buf_t **dbp, int flags)
195 191 {
196 192 int err;
197 193 int db_flags = DB_RF_CANFAIL;
198 194
199 195 if (flags & DMU_READ_NO_PREFETCH)
200 196 db_flags |= DB_RF_NOPREFETCH;
201 197
202 198 err = dmu_buf_hold_noread_by_dnode(dn, offset, tag, dbp);
203 199 if (err == 0) {
204 200 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
205 201 err = dbuf_read(db, NULL, db_flags);
206 202 if (err != 0) {
207 203 dbuf_rele(db, tag);
208 204 *dbp = NULL;
209 205 }
210 206 }
211 207
212 208 return (err);
213 209 }
214 210
215 211 int
216 212 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
217 213 void *tag, dmu_buf_t **dbp, int flags)
218 214 {
219 215 int err;
220 216 int db_flags = DB_RF_CANFAIL;
221 217
222 218 if (flags & DMU_READ_NO_PREFETCH)
223 219 db_flags |= DB_RF_NOPREFETCH;
224 220
225 221 err = dmu_buf_hold_noread(os, object, offset, tag, dbp);
226 222 if (err == 0) {
227 223 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
228 224 err = dbuf_read(db, NULL, db_flags);
229 225 if (err != 0) {
230 226 dbuf_rele(db, tag);
231 227 *dbp = NULL;
232 228 }
233 229 }
234 230
235 231 return (err);
236 232 }
237 233
238 234 int
239 235 dmu_bonus_max(void)
240 236 {
241 237 return (DN_MAX_BONUSLEN);
242 238 }
243 239
244 240 int
245 241 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
246 242 {
247 243 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
248 244 dnode_t *dn;
249 245 int error;
250 246
251 247 DB_DNODE_ENTER(db);
252 248 dn = DB_DNODE(db);
253 249
254 250 if (dn->dn_bonus != db) {
255 251 error = SET_ERROR(EINVAL);
256 252 } else if (newsize < 0 || newsize > db_fake->db_size) {
257 253 error = SET_ERROR(EINVAL);
258 254 } else {
259 255 dnode_setbonuslen(dn, newsize, tx);
260 256 error = 0;
261 257 }
262 258
263 259 DB_DNODE_EXIT(db);
264 260 return (error);
265 261 }
266 262
267 263 int
268 264 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
269 265 {
270 266 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
271 267 dnode_t *dn;
272 268 int error;
273 269
274 270 DB_DNODE_ENTER(db);
275 271 dn = DB_DNODE(db);
276 272
277 273 if (!DMU_OT_IS_VALID(type)) {
278 274 error = SET_ERROR(EINVAL);
279 275 } else if (dn->dn_bonus != db) {
280 276 error = SET_ERROR(EINVAL);
281 277 } else {
282 278 dnode_setbonus_type(dn, type, tx);
283 279 error = 0;
284 280 }
285 281
286 282 DB_DNODE_EXIT(db);
287 283 return (error);
288 284 }
289 285
290 286 dmu_object_type_t
291 287 dmu_get_bonustype(dmu_buf_t *db_fake)
292 288 {
293 289 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
294 290 dnode_t *dn;
295 291 dmu_object_type_t type;
296 292
297 293 DB_DNODE_ENTER(db);
298 294 dn = DB_DNODE(db);
299 295 type = dn->dn_bonustype;
300 296 DB_DNODE_EXIT(db);
301 297
302 298 return (type);
303 299 }
304 300
305 301 int
306 302 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
307 303 {
308 304 dnode_t *dn;
309 305 int error;
310 306
311 307 error = dnode_hold(os, object, FTAG, &dn);
312 308 dbuf_rm_spill(dn, tx);
313 309 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
314 310 dnode_rm_spill(dn, tx);
315 311 rw_exit(&dn->dn_struct_rwlock);
316 312 dnode_rele(dn, FTAG);
317 313 return (error);
318 314 }
319 315
320 316 /*
321 317 * returns ENOENT, EIO, or 0.
322 318 */
323 319 int
324 320 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
325 321 {
326 322 dnode_t *dn;
327 323 dmu_buf_impl_t *db;
328 324 int error;
329 325
330 326 error = dnode_hold(os, object, FTAG, &dn);
331 327 if (error)
332 328 return (error);
333 329
334 330 rw_enter(&dn->dn_struct_rwlock, RW_READER);
335 331 if (dn->dn_bonus == NULL) {
336 332 rw_exit(&dn->dn_struct_rwlock);
337 333 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
338 334 if (dn->dn_bonus == NULL)
339 335 dbuf_create_bonus(dn);
340 336 }
341 337 db = dn->dn_bonus;
342 338
343 339 /* as long as the bonus buf is held, the dnode will be held */
344 340 if (refcount_add(&db->db_holds, tag) == 1) {
345 341 VERIFY(dnode_add_ref(dn, db));
346 342 atomic_inc_32(&dn->dn_dbufs_count);
347 343 }
348 344
349 345 /*
350 346 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
351 347 * hold and incrementing the dbuf count to ensure that dnode_move() sees
352 348 * a dnode hold for every dbuf.
353 349 */
354 350 rw_exit(&dn->dn_struct_rwlock);
355 351
356 352 dnode_rele(dn, FTAG);
357 353
358 354 VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH));
359 355
360 356 *dbp = &db->db;
361 357 return (0);
362 358 }
363 359
364 360 /*
365 361 * returns ENOENT, EIO, or 0.
366 362 *
367 363 * This interface will allocate a blank spill dbuf when a spill blk
368 364 * doesn't already exist on the dnode.
369 365 *
370 366 * if you only want to find an already existing spill db, then
371 367 * dmu_spill_hold_existing() should be used.
372 368 */
373 369 int
374 370 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
375 371 {
376 372 dmu_buf_impl_t *db = NULL;
377 373 int err;
378 374
379 375 if ((flags & DB_RF_HAVESTRUCT) == 0)
380 376 rw_enter(&dn->dn_struct_rwlock, RW_READER);
381 377
382 378 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
383 379
384 380 if ((flags & DB_RF_HAVESTRUCT) == 0)
385 381 rw_exit(&dn->dn_struct_rwlock);
386 382
387 383 ASSERT(db != NULL);
388 384 err = dbuf_read(db, NULL, flags);
389 385 if (err == 0)
390 386 *dbp = &db->db;
391 387 else
392 388 dbuf_rele(db, tag);
393 389 return (err);
394 390 }
395 391
396 392 int
397 393 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
398 394 {
399 395 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
400 396 dnode_t *dn;
401 397 int err;
402 398
403 399 DB_DNODE_ENTER(db);
404 400 dn = DB_DNODE(db);
405 401
406 402 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
407 403 err = SET_ERROR(EINVAL);
408 404 } else {
409 405 rw_enter(&dn->dn_struct_rwlock, RW_READER);
410 406
411 407 if (!dn->dn_have_spill) {
412 408 err = SET_ERROR(ENOENT);
413 409 } else {
414 410 err = dmu_spill_hold_by_dnode(dn,
415 411 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
416 412 }
417 413
418 414 rw_exit(&dn->dn_struct_rwlock);
419 415 }
420 416
421 417 DB_DNODE_EXIT(db);
422 418 return (err);
423 419 }
424 420
425 421 int
426 422 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
427 423 {
428 424 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
429 425 dnode_t *dn;
430 426 int err;
431 427
432 428 DB_DNODE_ENTER(db);
433 429 dn = DB_DNODE(db);
434 430 err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp);
435 431 DB_DNODE_EXIT(db);
436 432
437 433 return (err);
438 434 }
439 435
440 436 /*
441 437 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
442 438 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
443 439 * and can induce severe lock contention when writing to several files
444 440 * whose dnodes are in the same block.
445 441 */
446 442 static int
447 443 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
448 444 boolean_t read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
449 445 {
450 446 dmu_buf_t **dbp;
451 447 uint64_t blkid, nblks, i;
452 448 uint32_t dbuf_flags;
453 449 int err;
454 450 zio_t *zio;
455 451
456 452 ASSERT(length <= DMU_MAX_ACCESS);
457 453
458 454 /*
459 455 * Note: We directly notify the prefetch code of this read, so that
460 456 * we can tell it about the multi-block read. dbuf_read() only knows
461 457 * about the one block it is accessing.
462 458 */
463 459 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT |
464 460 DB_RF_NOPREFETCH;
465 461
466 462 rw_enter(&dn->dn_struct_rwlock, RW_READER);
467 463 if (dn->dn_datablkshift) {
468 464 int blkshift = dn->dn_datablkshift;
469 465 nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) -
470 466 P2ALIGN(offset, 1ULL << blkshift)) >> blkshift;
471 467 } else {
472 468 if (offset + length > dn->dn_datablksz) {
473 469 zfs_panic_recover("zfs: accessing past end of object "
474 470 "%llx/%llx (size=%u access=%llu+%llu)",
475 471 (longlong_t)dn->dn_objset->
476 472 os_dsl_dataset->ds_object,
477 473 (longlong_t)dn->dn_object, dn->dn_datablksz,
478 474 (longlong_t)offset, (longlong_t)length);
479 475 rw_exit(&dn->dn_struct_rwlock);
480 476 return (SET_ERROR(EIO));
481 477 }
482 478 nblks = 1;
483 479 }
484 480 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
485 481
486 482 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
487 483 blkid = dbuf_whichblock(dn, 0, offset);
488 484 for (i = 0; i < nblks; i++) {
489 485 dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag);
490 486 if (db == NULL) {
491 487 rw_exit(&dn->dn_struct_rwlock);
492 488 dmu_buf_rele_array(dbp, nblks, tag);
493 489 zio_nowait(zio);
494 490 return (SET_ERROR(EIO));
495 491 }
496 492
497 493 /* initiate async i/o */
498 494 if (read)
499 495 (void) dbuf_read(db, zio, dbuf_flags);
500 496 dbp[i] = &db->db;
501 497 }
502 498
503 499 if ((flags & DMU_READ_NO_PREFETCH) == 0 &&
504 500 DNODE_META_IS_CACHEABLE(dn) && length <= zfetch_array_rd_sz) {
505 501 dmu_zfetch(&dn->dn_zfetch, blkid, nblks,
506 502 read && DNODE_IS_CACHEABLE(dn));
507 503 }
508 504 rw_exit(&dn->dn_struct_rwlock);
509 505
510 506 /* wait for async i/o */
511 507 err = zio_wait(zio);
512 508 if (err) {
513 509 dmu_buf_rele_array(dbp, nblks, tag);
514 510 return (err);
515 511 }
516 512
517 513 /* wait for other io to complete */
518 514 if (read) {
519 515 for (i = 0; i < nblks; i++) {
520 516 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
521 517 mutex_enter(&db->db_mtx);
522 518 while (db->db_state == DB_READ ||
523 519 db->db_state == DB_FILL)
524 520 cv_wait(&db->db_changed, &db->db_mtx);
525 521 if (db->db_state == DB_UNCACHED)
526 522 err = SET_ERROR(EIO);
527 523 mutex_exit(&db->db_mtx);
528 524 if (err) {
529 525 dmu_buf_rele_array(dbp, nblks, tag);
530 526 return (err);
531 527 }
532 528 }
533 529 }
534 530
535 531 *numbufsp = nblks;
536 532 *dbpp = dbp;
537 533 return (0);
538 534 }
539 535
540 536 static int
541 537 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
542 538 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
543 539 {
544 540 dnode_t *dn;
545 541 int err;
546 542
547 543 err = dnode_hold(os, object, FTAG, &dn);
548 544 if (err)
549 545 return (err);
550 546
551 547 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
552 548 numbufsp, dbpp, DMU_READ_PREFETCH);
553 549
554 550 dnode_rele(dn, FTAG);
555 551
556 552 return (err);
557 553 }
558 554
559 555 int
560 556 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
561 557 uint64_t length, boolean_t read, void *tag, int *numbufsp,
562 558 dmu_buf_t ***dbpp)
563 559 {
564 560 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
565 561 dnode_t *dn;
566 562 int err;
567 563
568 564 DB_DNODE_ENTER(db);
569 565 dn = DB_DNODE(db);
570 566 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
571 567 numbufsp, dbpp, DMU_READ_PREFETCH);
572 568 DB_DNODE_EXIT(db);
573 569
574 570 return (err);
575 571 }
576 572
577 573 void
578 574 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
579 575 {
580 576 int i;
581 577 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
582 578
583 579 if (numbufs == 0)
584 580 return;
585 581
586 582 for (i = 0; i < numbufs; i++) {
587 583 if (dbp[i])
588 584 dbuf_rele(dbp[i], tag);
589 585 }
590 586
591 587 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
592 588 }
593 589
594 590 /*
595 591 * Issue prefetch i/os for the given blocks. If level is greater than 0, the
596 592 * indirect blocks prefeteched will be those that point to the blocks containing
597 593 * the data starting at offset, and continuing to offset + len.
598 594 *
599 595 * Note that if the indirect blocks above the blocks being prefetched are not in
600 596 * cache, they will be asychronously read in.
601 597 */
602 598 void
603 599 dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset,
604 600 uint64_t len, zio_priority_t pri)
605 601 {
606 602 dnode_t *dn;
607 603 uint64_t blkid;
608 604 int nblks, err;
609 605
610 606 if (len == 0) { /* they're interested in the bonus buffer */
611 607 dn = DMU_META_DNODE(os);
612 608
613 609 if (object == 0 || object >= DN_MAX_OBJECT)
614 610 return;
615 611
616 612 rw_enter(&dn->dn_struct_rwlock, RW_READER);
617 613 blkid = dbuf_whichblock(dn, level,
618 614 object * sizeof (dnode_phys_t));
619 615 dbuf_prefetch(dn, level, blkid, pri, 0);
620 616 rw_exit(&dn->dn_struct_rwlock);
621 617 return;
622 618 }
623 619
624 620 /*
625 621 * XXX - Note, if the dnode for the requested object is not
626 622 * already cached, we will do a *synchronous* read in the
627 623 * dnode_hold() call. The same is true for any indirects.
628 624 */
629 625 err = dnode_hold(os, object, FTAG, &dn);
630 626 if (err != 0)
631 627 return;
632 628
633 629 rw_enter(&dn->dn_struct_rwlock, RW_READER);
634 630 /*
635 631 * offset + len - 1 is the last byte we want to prefetch for, and offset
636 632 * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the
637 633 * last block we want to prefetch, and dbuf_whichblock(dn, level,
638 634 * offset) is the first. Then the number we need to prefetch is the
639 635 * last - first + 1.
640 636 */
641 637 if (level > 0 || dn->dn_datablkshift != 0) {
642 638 nblks = dbuf_whichblock(dn, level, offset + len - 1) -
643 639 dbuf_whichblock(dn, level, offset) + 1;
644 640 } else {
645 641 nblks = (offset < dn->dn_datablksz);
646 642 }
647 643
648 644 if (nblks != 0) {
649 645 blkid = dbuf_whichblock(dn, level, offset);
650 646 for (int i = 0; i < nblks; i++)
651 647 dbuf_prefetch(dn, level, blkid + i, pri, 0);
652 648 }
653 649
654 650 rw_exit(&dn->dn_struct_rwlock);
655 651
656 652 dnode_rele(dn, FTAG);
657 653 }
658 654
659 655 /*
660 656 * Get the next "chunk" of file data to free. We traverse the file from
661 657 * the end so that the file gets shorter over time (if we crashes in the
662 658 * middle, this will leave us in a better state). We find allocated file
663 659 * data by simply searching the allocated level 1 indirects.
664 660 *
665 661 * On input, *start should be the first offset that does not need to be
666 662 * freed (e.g. "offset + length"). On return, *start will be the first
667 663 * offset that should be freed.
668 664 */
669 665 static int
670 666 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum)
671 667 {
672 668 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
673 669 /* bytes of data covered by a level-1 indirect block */
674 670 uint64_t iblkrange =
675 671 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
676 672
677 673 ASSERT3U(minimum, <=, *start);
678 674
679 675 if (*start - minimum <= iblkrange * maxblks) {
680 676 *start = minimum;
681 677 return (0);
682 678 }
683 679 ASSERT(ISP2(iblkrange));
684 680
685 681 for (uint64_t blks = 0; *start > minimum && blks < maxblks; blks++) {
686 682 int err;
687 683
688 684 /*
689 685 * dnode_next_offset(BACKWARDS) will find an allocated L1
690 686 * indirect block at or before the input offset. We must
691 687 * decrement *start so that it is at the end of the region
692 688 * to search.
693 689 */
694 690 (*start)--;
695 691 err = dnode_next_offset(dn,
696 692 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
697 693
698 694 /* if there are no indirect blocks before start, we are done */
699 695 if (err == ESRCH) {
700 696 *start = minimum;
701 697 break;
702 698 } else if (err != 0) {
703 699 return (err);
704 700 }
|
↓ open down ↓ |
565 lines elided |
↑ open up ↑ |
705 701
706 702 /* set start to the beginning of this L1 indirect */
707 703 *start = P2ALIGN(*start, iblkrange);
708 704 }
709 705 if (*start < minimum)
710 706 *start = minimum;
711 707 return (0);
712 708 }
713 709
714 710 /*
715 - * If this objset is of type OST_ZFS return true if vfs's unmounted flag is set,
716 - * otherwise return false.
717 - * Used below in dmu_free_long_range_impl() to enable abort when unmounting
711 + * If this dnode is in the ZFS object set
712 + * return true if vfs's unmounted flag is set or the
713 + * zfsvfs is currently suspended, otherwise return false.
718 714 */
719 715 /*ARGSUSED*/
720 716 static boolean_t
721 -dmu_objset_zfs_unmounting(objset_t *os)
717 +dmu_dnode_fs_unmounting_or_suspended(dnode_t *freeing_dn)
722 718 {
723 719 #ifdef _KERNEL
724 - if (dmu_objset_type(os) == DMU_OST_ZFS)
725 - return (zfs_get_vfs_flag_unmounted(os));
726 -#endif
720 + boolean_t busy = B_FALSE;
721 + objset_t *os = freeing_dn->dn_objset;
722 + zfsvfs_t *zfsvfs;
723 +
724 + if (dmu_objset_type(os) == DMU_OST_ZFS) {
725 + mutex_enter(&os->os_user_ptr_lock);
726 + zfsvfs = dmu_objset_get_user(os);
727 + if (zfsvfs != NULL && zfsvfs->z_vfs != NULL &&
728 + ((zfsvfs->z_vfs->vfs_flag & VFS_UNMOUNTED) ||
729 + zfsvfs->z_busy))
730 + busy = B_TRUE;
731 + mutex_exit(&os->os_user_ptr_lock);
732 + }
733 +
734 + return (busy);
735 +#else
727 736 return (B_FALSE);
737 +#endif
728 738 }
729 739
730 740 static int
731 741 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
732 742 uint64_t length)
733 743 {
734 744 uint64_t object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
735 745 int err;
736 746 uint64_t dirty_frees_threshold;
737 747 dsl_pool_t *dp = dmu_objset_pool(os);
738 748
739 749 if (offset >= object_size)
740 750 return (0);
741 751
742 752 if (zfs_per_txg_dirty_frees_percent <= 100)
743 753 dirty_frees_threshold =
744 754 zfs_per_txg_dirty_frees_percent * zfs_dirty_data_max / 100;
745 755 else
746 756 dirty_frees_threshold = zfs_dirty_data_max / 4;
747 757
758 + if (length == DMU_OBJECT_END && offset == 0)
759 + dnode_evict_dbufs(dn, 0);
760 +
748 761 if (length == DMU_OBJECT_END || offset + length > object_size)
749 762 length = object_size - offset;
750 763
764 + mutex_enter(&dp->dp_lock);
765 + dp->dp_long_freeing_total += length;
766 + mutex_exit(&dp->dp_lock);
767 +
751 768 while (length != 0) {
752 769 uint64_t chunk_end, chunk_begin, chunk_len;
753 770 uint64_t long_free_dirty_all_txgs = 0;
754 771 dmu_tx_t *tx;
755 772
756 - if (dmu_objset_zfs_unmounting(dn->dn_objset))
773 + if (dmu_dnode_fs_unmounting_or_suspended(dn)) {
774 + mutex_enter(&dp->dp_lock);
775 + dp->dp_long_freeing_total -= length;
776 + mutex_exit(&dp->dp_lock);
777 +
757 778 return (SET_ERROR(EINTR));
779 + }
758 780
759 781 chunk_end = chunk_begin = offset + length;
760 782
761 783 /* move chunk_begin backwards to the beginning of this chunk */
762 784 err = get_next_chunk(dn, &chunk_begin, offset);
763 785 if (err)
764 786 return (err);
765 787 ASSERT3U(chunk_begin, >=, offset);
766 788 ASSERT3U(chunk_begin, <=, chunk_end);
767 789
768 790 chunk_len = chunk_end - chunk_begin;
769 791
770 792 mutex_enter(&dp->dp_lock);
771 793 for (int t = 0; t < TXG_SIZE; t++) {
772 794 long_free_dirty_all_txgs +=
773 795 dp->dp_long_free_dirty_pertxg[t];
774 796 }
775 797 mutex_exit(&dp->dp_lock);
776 798
777 799 /*
778 800 * To avoid filling up a TXG with just frees wait for
779 801 * the next TXG to open before freeing more chunks if
780 802 * we have reached the threshold of frees
781 803 */
782 804 if (dirty_frees_threshold != 0 &&
783 805 long_free_dirty_all_txgs >= dirty_frees_threshold) {
784 806 txg_wait_open(dp, 0);
785 807 continue;
786 808 }
787 809
788 810 tx = dmu_tx_create(os);
|
↓ open down ↓ |
21 lines elided |
↑ open up ↑ |
789 811 dmu_tx_hold_free(tx, dn->dn_object, chunk_begin, chunk_len);
790 812
791 813 /*
792 814 * Mark this transaction as typically resulting in a net
793 815 * reduction in space used.
794 816 */
795 817 dmu_tx_mark_netfree(tx);
796 818 err = dmu_tx_assign(tx, TXG_WAIT);
797 819 if (err) {
798 820 dmu_tx_abort(tx);
821 + mutex_enter(&dp->dp_lock);
822 + dp->dp_long_freeing_total -= length - chunk_len;
823 + mutex_exit(&dp->dp_lock);
799 824 return (err);
800 825 }
801 826
802 827 mutex_enter(&dp->dp_lock);
803 828 dp->dp_long_free_dirty_pertxg[dmu_tx_get_txg(tx) & TXG_MASK] +=
804 829 chunk_len;
805 830 mutex_exit(&dp->dp_lock);
806 831 DTRACE_PROBE3(free__long__range,
807 832 uint64_t, long_free_dirty_all_txgs, uint64_t, chunk_len,
808 833 uint64_t, dmu_tx_get_txg(tx));
809 834 dnode_free_range(dn, chunk_begin, chunk_len, tx);
810 835 dmu_tx_commit(tx);
811 836
812 837 length -= chunk_len;
813 838 }
814 839 return (0);
815 840 }
816 841
817 842 int
818 843 dmu_free_long_range(objset_t *os, uint64_t object,
819 844 uint64_t offset, uint64_t length)
820 845 {
821 846 dnode_t *dn;
822 847 int err;
823 848
824 849 err = dnode_hold(os, object, FTAG, &dn);
825 850 if (err != 0)
826 851 return (err);
827 852 err = dmu_free_long_range_impl(os, dn, offset, length);
828 853
829 854 /*
830 855 * It is important to zero out the maxblkid when freeing the entire
831 856 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
832 857 * will take the fast path, and (b) dnode_reallocate() can verify
833 858 * that the entire file has been freed.
834 859 */
835 860 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
836 861 dn->dn_maxblkid = 0;
837 862
838 863 dnode_rele(dn, FTAG);
839 864 return (err);
840 865 }
841 866
842 867 int
843 868 dmu_free_long_object(objset_t *os, uint64_t object)
844 869 {
845 870 dmu_tx_t *tx;
846 871 int err;
847 872
848 873 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
849 874 if (err != 0)
850 875 return (err);
851 876
852 877 tx = dmu_tx_create(os);
853 878 dmu_tx_hold_bonus(tx, object);
854 879 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
855 880 dmu_tx_mark_netfree(tx);
856 881 err = dmu_tx_assign(tx, TXG_WAIT);
857 882 if (err == 0) {
858 883 err = dmu_object_free(os, object, tx);
859 884 dmu_tx_commit(tx);
860 885 } else {
861 886 dmu_tx_abort(tx);
862 887 }
863 888
864 889 return (err);
865 890 }
866 891
867 892 int
868 893 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
869 894 uint64_t size, dmu_tx_t *tx)
870 895 {
871 896 dnode_t *dn;
872 897 int err = dnode_hold(os, object, FTAG, &dn);
873 898 if (err)
874 899 return (err);
875 900 ASSERT(offset < UINT64_MAX);
876 901 ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
877 902 dnode_free_range(dn, offset, size, tx);
878 903 dnode_rele(dn, FTAG);
879 904 return (0);
880 905 }
881 906
882 907 static int
883 908 dmu_read_impl(dnode_t *dn, uint64_t offset, uint64_t size,
884 909 void *buf, uint32_t flags)
885 910 {
886 911 dmu_buf_t **dbp;
887 912 int numbufs, err = 0;
888 913
889 914 /*
890 915 * Deal with odd block sizes, where there can't be data past the first
891 916 * block. If we ever do the tail block optimization, we will need to
892 917 * handle that here as well.
893 918 */
894 919 if (dn->dn_maxblkid == 0) {
895 920 int newsz = offset > dn->dn_datablksz ? 0 :
896 921 MIN(size, dn->dn_datablksz - offset);
897 922 bzero((char *)buf + newsz, size - newsz);
898 923 size = newsz;
899 924 }
900 925
901 926 while (size > 0) {
902 927 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
903 928 int i;
904 929
905 930 /*
906 931 * NB: we could do this block-at-a-time, but it's nice
907 932 * to be reading in parallel.
908 933 */
909 934 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
910 935 TRUE, FTAG, &numbufs, &dbp, flags);
911 936 if (err)
912 937 break;
913 938
914 939 for (i = 0; i < numbufs; i++) {
915 940 int tocpy;
916 941 int bufoff;
917 942 dmu_buf_t *db = dbp[i];
918 943
919 944 ASSERT(size > 0);
920 945
921 946 bufoff = offset - db->db_offset;
922 947 tocpy = (int)MIN(db->db_size - bufoff, size);
923 948
924 949 bcopy((char *)db->db_data + bufoff, buf, tocpy);
925 950
926 951 offset += tocpy;
927 952 size -= tocpy;
928 953 buf = (char *)buf + tocpy;
929 954 }
930 955 dmu_buf_rele_array(dbp, numbufs, FTAG);
931 956 }
932 957 return (err);
933 958 }
934 959
935 960 int
936 961 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
937 962 void *buf, uint32_t flags)
938 963 {
939 964 dnode_t *dn;
940 965 int err;
941 966
942 967 err = dnode_hold(os, object, FTAG, &dn);
943 968 if (err != 0)
944 969 return (err);
945 970
946 971 err = dmu_read_impl(dn, offset, size, buf, flags);
947 972 dnode_rele(dn, FTAG);
948 973 return (err);
949 974 }
950 975
951 976 int
952 977 dmu_read_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, void *buf,
953 978 uint32_t flags)
954 979 {
955 980 return (dmu_read_impl(dn, offset, size, buf, flags));
956 981 }
957 982
958 983 static void
959 984 dmu_write_impl(dmu_buf_t **dbp, int numbufs, uint64_t offset, uint64_t size,
960 985 const void *buf, dmu_tx_t *tx)
961 986 {
962 987 int i;
963 988
964 989 for (i = 0; i < numbufs; i++) {
965 990 int tocpy;
966 991 int bufoff;
967 992 dmu_buf_t *db = dbp[i];
968 993
969 994 ASSERT(size > 0);
970 995
971 996 bufoff = offset - db->db_offset;
972 997 tocpy = (int)MIN(db->db_size - bufoff, size);
973 998
974 999 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
975 1000
976 1001 if (tocpy == db->db_size)
977 1002 dmu_buf_will_fill(db, tx);
978 1003 else
979 1004 dmu_buf_will_dirty(db, tx);
980 1005
981 1006 bcopy(buf, (char *)db->db_data + bufoff, tocpy);
982 1007
983 1008 if (tocpy == db->db_size)
984 1009 dmu_buf_fill_done(db, tx);
985 1010
986 1011 offset += tocpy;
987 1012 size -= tocpy;
988 1013 buf = (char *)buf + tocpy;
989 1014 }
990 1015 }
991 1016
992 1017 void
993 1018 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
994 1019 const void *buf, dmu_tx_t *tx)
995 1020 {
996 1021 dmu_buf_t **dbp;
997 1022 int numbufs;
998 1023
999 1024 if (size == 0)
1000 1025 return;
1001 1026
1002 1027 VERIFY0(dmu_buf_hold_array(os, object, offset, size,
1003 1028 FALSE, FTAG, &numbufs, &dbp));
1004 1029 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1005 1030 dmu_buf_rele_array(dbp, numbufs, FTAG);
1006 1031 }
1007 1032
1008 1033 void
1009 1034 dmu_write_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size,
1010 1035 const void *buf, dmu_tx_t *tx)
1011 1036 {
1012 1037 dmu_buf_t **dbp;
1013 1038 int numbufs;
|
↓ open down ↓ |
205 lines elided |
↑ open up ↑ |
1014 1039
1015 1040 if (size == 0)
1016 1041 return;
1017 1042
1018 1043 VERIFY0(dmu_buf_hold_array_by_dnode(dn, offset, size,
1019 1044 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH));
1020 1045 dmu_write_impl(dbp, numbufs, offset, size, buf, tx);
1021 1046 dmu_buf_rele_array(dbp, numbufs, FTAG);
1022 1047 }
1023 1048
1024 -static int
1025 -dmu_object_remap_one_indirect(objset_t *os, dnode_t *dn,
1026 - uint64_t last_removal_txg, uint64_t offset)
1027 -{
1028 - uint64_t l1blkid = dbuf_whichblock(dn, 1, offset);
1029 - int err = 0;
1030 -
1031 - rw_enter(&dn->dn_struct_rwlock, RW_READER);
1032 - dmu_buf_impl_t *dbuf = dbuf_hold_level(dn, 1, l1blkid, FTAG);
1033 - ASSERT3P(dbuf, !=, NULL);
1034 -
1035 - /*
1036 - * If the block hasn't been written yet, this default will ensure
1037 - * we don't try to remap it.
1038 - */
1039 - uint64_t birth = UINT64_MAX;
1040 - ASSERT3U(last_removal_txg, !=, UINT64_MAX);
1041 - if (dbuf->db_blkptr != NULL)
1042 - birth = dbuf->db_blkptr->blk_birth;
1043 - rw_exit(&dn->dn_struct_rwlock);
1044 -
1045 - /*
1046 - * If this L1 was already written after the last removal, then we've
1047 - * already tried to remap it.
1048 - */
1049 - if (birth <= last_removal_txg &&
1050 - dbuf_read(dbuf, NULL, DB_RF_MUST_SUCCEED) == 0 &&
1051 - dbuf_can_remap(dbuf)) {
1052 - dmu_tx_t *tx = dmu_tx_create(os);
1053 - dmu_tx_hold_remap_l1indirect(tx, dn->dn_object);
1054 - err = dmu_tx_assign(tx, TXG_WAIT);
1055 - if (err == 0) {
1056 - (void) dbuf_dirty(dbuf, tx);
1057 - dmu_tx_commit(tx);
1058 - } else {
1059 - dmu_tx_abort(tx);
1060 - }
1061 - }
1062 -
1063 - dbuf_rele(dbuf, FTAG);
1064 -
1065 - delay(zfs_object_remap_one_indirect_delay_ticks);
1066 -
1067 - return (err);
1068 -}
1069 -
1070 -/*
1071 - * Remap all blockpointers in the object, if possible, so that they reference
1072 - * only concrete vdevs.
1073 - *
1074 - * To do this, iterate over the L0 blockpointers and remap any that reference
1075 - * an indirect vdev. Note that we only examine L0 blockpointers; since we
1076 - * cannot guarantee that we can remap all blockpointer anyways (due to split
1077 - * blocks), we do not want to make the code unnecessarily complicated to
1078 - * catch the unlikely case that there is an L1 block on an indirect vdev that
1079 - * contains no indirect blockpointers.
1080 - */
1081 -int
1082 -dmu_object_remap_indirects(objset_t *os, uint64_t object,
1083 - uint64_t last_removal_txg)
1084 -{
1085 - uint64_t offset, l1span;
1086 - int err;
1087 - dnode_t *dn;
1088 -
1089 - err = dnode_hold(os, object, FTAG, &dn);
1090 - if (err != 0) {
1091 - return (err);
1092 - }
1093 -
1094 - if (dn->dn_nlevels <= 1) {
1095 - if (issig(JUSTLOOKING) && issig(FORREAL)) {
1096 - err = SET_ERROR(EINTR);
1097 - }
1098 -
1099 - /*
1100 - * If the dnode has no indirect blocks, we cannot dirty them.
1101 - * We still want to remap the blkptr(s) in the dnode if
1102 - * appropriate, so mark it as dirty.
1103 - */
1104 - if (err == 0 && dnode_needs_remap(dn)) {
1105 - dmu_tx_t *tx = dmu_tx_create(os);
1106 - dmu_tx_hold_bonus(tx, dn->dn_object);
1107 - if ((err = dmu_tx_assign(tx, TXG_WAIT)) == 0) {
1108 - dnode_setdirty(dn, tx);
1109 - dmu_tx_commit(tx);
1110 - } else {
1111 - dmu_tx_abort(tx);
1112 - }
1113 - }
1114 -
1115 - dnode_rele(dn, FTAG);
1116 - return (err);
1117 - }
1118 -
1119 - offset = 0;
1120 - l1span = 1ULL << (dn->dn_indblkshift - SPA_BLKPTRSHIFT +
1121 - dn->dn_datablkshift);
1122 - /*
1123 - * Find the next L1 indirect that is not a hole.
1124 - */
1125 - while (dnode_next_offset(dn, 0, &offset, 2, 1, 0) == 0) {
1126 - if (issig(JUSTLOOKING) && issig(FORREAL)) {
1127 - err = SET_ERROR(EINTR);
1128 - break;
1129 - }
1130 - if ((err = dmu_object_remap_one_indirect(os, dn,
1131 - last_removal_txg, offset)) != 0) {
1132 - break;
1133 - }
1134 - offset += l1span;
1135 - }
1136 -
1137 - dnode_rele(dn, FTAG);
1138 - return (err);
1139 -}
1140 -
1141 1049 void
1142 1050 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1143 1051 dmu_tx_t *tx)
1144 1052 {
1145 1053 dmu_buf_t **dbp;
1146 1054 int numbufs, i;
1147 1055
1148 1056 if (size == 0)
1149 1057 return;
1150 1058
1151 1059 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
1152 1060 FALSE, FTAG, &numbufs, &dbp));
1153 1061
1154 1062 for (i = 0; i < numbufs; i++) {
1155 1063 dmu_buf_t *db = dbp[i];
1156 1064
1157 1065 dmu_buf_will_not_fill(db, tx);
1158 1066 }
1159 1067 dmu_buf_rele_array(dbp, numbufs, FTAG);
1160 1068 }
1161 1069
1162 1070 void
1163 1071 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
1164 1072 void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
1165 1073 int compressed_size, int byteorder, dmu_tx_t *tx)
1166 1074 {
1167 1075 dmu_buf_t *db;
1168 1076
1169 1077 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
1170 1078 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
1171 1079 VERIFY0(dmu_buf_hold_noread(os, object, offset,
1172 1080 FTAG, &db));
1173 1081
1174 1082 dmu_buf_write_embedded(db,
1175 1083 data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
1176 1084 uncompressed_size, compressed_size, byteorder, tx);
1177 1085
1178 1086 dmu_buf_rele(db, FTAG);
1179 1087 }
1180 1088
1181 1089 /*
1182 1090 * DMU support for xuio
1183 1091 */
1184 1092 kstat_t *xuio_ksp = NULL;
1185 1093
1186 1094 int
1187 1095 dmu_xuio_init(xuio_t *xuio, int nblk)
1188 1096 {
1189 1097 dmu_xuio_t *priv;
1190 1098 uio_t *uio = &xuio->xu_uio;
1191 1099
1192 1100 uio->uio_iovcnt = nblk;
1193 1101 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
1194 1102
1195 1103 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
1196 1104 priv->cnt = nblk;
1197 1105 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
1198 1106 priv->iovp = uio->uio_iov;
1199 1107 XUIO_XUZC_PRIV(xuio) = priv;
1200 1108
1201 1109 if (XUIO_XUZC_RW(xuio) == UIO_READ)
1202 1110 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
1203 1111 else
1204 1112 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
1205 1113
1206 1114 return (0);
1207 1115 }
1208 1116
1209 1117 void
1210 1118 dmu_xuio_fini(xuio_t *xuio)
1211 1119 {
1212 1120 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1213 1121 int nblk = priv->cnt;
1214 1122
1215 1123 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
1216 1124 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
1217 1125 kmem_free(priv, sizeof (dmu_xuio_t));
1218 1126
1219 1127 if (XUIO_XUZC_RW(xuio) == UIO_READ)
1220 1128 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
1221 1129 else
1222 1130 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
1223 1131 }
1224 1132
1225 1133 /*
1226 1134 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
1227 1135 * and increase priv->next by 1.
1228 1136 */
1229 1137 int
1230 1138 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
1231 1139 {
1232 1140 struct iovec *iov;
1233 1141 uio_t *uio = &xuio->xu_uio;
1234 1142 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1235 1143 int i = priv->next++;
1236 1144
1237 1145 ASSERT(i < priv->cnt);
1238 1146 ASSERT(off + n <= arc_buf_lsize(abuf));
1239 1147 iov = uio->uio_iov + i;
1240 1148 iov->iov_base = (char *)abuf->b_data + off;
1241 1149 iov->iov_len = n;
1242 1150 priv->bufs[i] = abuf;
1243 1151 return (0);
1244 1152 }
1245 1153
1246 1154 int
1247 1155 dmu_xuio_cnt(xuio_t *xuio)
1248 1156 {
1249 1157 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1250 1158 return (priv->cnt);
1251 1159 }
1252 1160
1253 1161 arc_buf_t *
1254 1162 dmu_xuio_arcbuf(xuio_t *xuio, int i)
1255 1163 {
1256 1164 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1257 1165
1258 1166 ASSERT(i < priv->cnt);
1259 1167 return (priv->bufs[i]);
1260 1168 }
1261 1169
1262 1170 void
1263 1171 dmu_xuio_clear(xuio_t *xuio, int i)
1264 1172 {
1265 1173 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
1266 1174
1267 1175 ASSERT(i < priv->cnt);
1268 1176 priv->bufs[i] = NULL;
1269 1177 }
1270 1178
1271 1179 static void
1272 1180 xuio_stat_init(void)
1273 1181 {
1274 1182 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
1275 1183 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
1276 1184 KSTAT_FLAG_VIRTUAL);
1277 1185 if (xuio_ksp != NULL) {
1278 1186 xuio_ksp->ks_data = &xuio_stats;
1279 1187 kstat_install(xuio_ksp);
1280 1188 }
1281 1189 }
1282 1190
1283 1191 static void
1284 1192 xuio_stat_fini(void)
1285 1193 {
1286 1194 if (xuio_ksp != NULL) {
1287 1195 kstat_delete(xuio_ksp);
1288 1196 xuio_ksp = NULL;
1289 1197 }
1290 1198 }
1291 1199
1292 1200 void
1293 1201 xuio_stat_wbuf_copied(void)
1294 1202 {
1295 1203 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1296 1204 }
1297 1205
1298 1206 void
1299 1207 xuio_stat_wbuf_nocopy(void)
1300 1208 {
1301 1209 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
1302 1210 }
1303 1211
1304 1212 #ifdef _KERNEL
1305 1213 static int
1306 1214 dmu_read_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size)
1307 1215 {
1308 1216 dmu_buf_t **dbp;
1309 1217 int numbufs, i, err;
1310 1218 xuio_t *xuio = NULL;
1311 1219
1312 1220 /*
1313 1221 * NB: we could do this block-at-a-time, but it's nice
1314 1222 * to be reading in parallel.
1315 1223 */
1316 1224 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1317 1225 TRUE, FTAG, &numbufs, &dbp, 0);
1318 1226 if (err)
1319 1227 return (err);
1320 1228
1321 1229 if (uio->uio_extflg == UIO_XUIO)
1322 1230 xuio = (xuio_t *)uio;
1323 1231
1324 1232 for (i = 0; i < numbufs; i++) {
1325 1233 int tocpy;
1326 1234 int bufoff;
1327 1235 dmu_buf_t *db = dbp[i];
1328 1236
1329 1237 ASSERT(size > 0);
1330 1238
1331 1239 bufoff = uio->uio_loffset - db->db_offset;
1332 1240 tocpy = (int)MIN(db->db_size - bufoff, size);
1333 1241
1334 1242 if (xuio) {
1335 1243 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
1336 1244 arc_buf_t *dbuf_abuf = dbi->db_buf;
1337 1245 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
1338 1246 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
1339 1247 if (!err) {
1340 1248 uio->uio_resid -= tocpy;
1341 1249 uio->uio_loffset += tocpy;
1342 1250 }
1343 1251
1344 1252 if (abuf == dbuf_abuf)
1345 1253 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
1346 1254 else
1347 1255 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
1348 1256 } else {
1349 1257 err = uiomove((char *)db->db_data + bufoff, tocpy,
1350 1258 UIO_READ, uio);
1351 1259 }
1352 1260 if (err)
1353 1261 break;
1354 1262
1355 1263 size -= tocpy;
1356 1264 }
1357 1265 dmu_buf_rele_array(dbp, numbufs, FTAG);
1358 1266
1359 1267 return (err);
1360 1268 }
1361 1269
1362 1270 /*
1363 1271 * Read 'size' bytes into the uio buffer.
1364 1272 * From object zdb->db_object.
1365 1273 * Starting at offset uio->uio_loffset.
1366 1274 *
1367 1275 * If the caller already has a dbuf in the target object
1368 1276 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1369 1277 * because we don't have to find the dnode_t for the object.
1370 1278 */
1371 1279 int
1372 1280 dmu_read_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size)
1373 1281 {
1374 1282 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1375 1283 dnode_t *dn;
1376 1284 int err;
1377 1285
1378 1286 if (size == 0)
1379 1287 return (0);
1380 1288
1381 1289 DB_DNODE_ENTER(db);
1382 1290 dn = DB_DNODE(db);
1383 1291 err = dmu_read_uio_dnode(dn, uio, size);
1384 1292 DB_DNODE_EXIT(db);
1385 1293
1386 1294 return (err);
1387 1295 }
1388 1296
1389 1297 /*
1390 1298 * Read 'size' bytes into the uio buffer.
1391 1299 * From the specified object
1392 1300 * Starting at offset uio->uio_loffset.
1393 1301 */
1394 1302 int
1395 1303 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
1396 1304 {
1397 1305 dnode_t *dn;
1398 1306 int err;
1399 1307
1400 1308 if (size == 0)
1401 1309 return (0);
1402 1310
1403 1311 err = dnode_hold(os, object, FTAG, &dn);
1404 1312 if (err)
1405 1313 return (err);
1406 1314
1407 1315 err = dmu_read_uio_dnode(dn, uio, size);
1408 1316
1409 1317 dnode_rele(dn, FTAG);
1410 1318
1411 1319 return (err);
1412 1320 }
1413 1321
1414 1322 static int
1415 1323 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1416 1324 {
1417 1325 dmu_buf_t **dbp;
1418 1326 int numbufs;
1419 1327 int err = 0;
1420 1328 int i;
1421 1329
1422 1330 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1423 1331 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1424 1332 if (err)
1425 1333 return (err);
1426 1334
1427 1335 for (i = 0; i < numbufs; i++) {
1428 1336 int tocpy;
1429 1337 int bufoff;
1430 1338 dmu_buf_t *db = dbp[i];
1431 1339
1432 1340 ASSERT(size > 0);
1433 1341
1434 1342 bufoff = uio->uio_loffset - db->db_offset;
1435 1343 tocpy = (int)MIN(db->db_size - bufoff, size);
1436 1344
1437 1345 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1438 1346
1439 1347 if (tocpy == db->db_size)
1440 1348 dmu_buf_will_fill(db, tx);
1441 1349 else
1442 1350 dmu_buf_will_dirty(db, tx);
1443 1351
1444 1352 /*
1445 1353 * XXX uiomove could block forever (eg. nfs-backed
1446 1354 * pages). There needs to be a uiolockdown() function
1447 1355 * to lock the pages in memory, so that uiomove won't
1448 1356 * block.
1449 1357 */
1450 1358 err = uiomove((char *)db->db_data + bufoff, tocpy,
1451 1359 UIO_WRITE, uio);
1452 1360
1453 1361 if (tocpy == db->db_size)
1454 1362 dmu_buf_fill_done(db, tx);
1455 1363
1456 1364 if (err)
1457 1365 break;
1458 1366
1459 1367 size -= tocpy;
1460 1368 }
1461 1369
1462 1370 dmu_buf_rele_array(dbp, numbufs, FTAG);
1463 1371 return (err);
1464 1372 }
1465 1373
1466 1374 /*
1467 1375 * Write 'size' bytes from the uio buffer.
1468 1376 * To object zdb->db_object.
1469 1377 * Starting at offset uio->uio_loffset.
1470 1378 *
1471 1379 * If the caller already has a dbuf in the target object
1472 1380 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1473 1381 * because we don't have to find the dnode_t for the object.
1474 1382 */
1475 1383 int
1476 1384 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1477 1385 dmu_tx_t *tx)
1478 1386 {
1479 1387 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1480 1388 dnode_t *dn;
1481 1389 int err;
1482 1390
1483 1391 if (size == 0)
1484 1392 return (0);
1485 1393
1486 1394 DB_DNODE_ENTER(db);
1487 1395 dn = DB_DNODE(db);
1488 1396 err = dmu_write_uio_dnode(dn, uio, size, tx);
1489 1397 DB_DNODE_EXIT(db);
1490 1398
1491 1399 return (err);
1492 1400 }
1493 1401
1494 1402 /*
1495 1403 * Write 'size' bytes from the uio buffer.
1496 1404 * To the specified object.
1497 1405 * Starting at offset uio->uio_loffset.
1498 1406 */
1499 1407 int
1500 1408 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1501 1409 dmu_tx_t *tx)
1502 1410 {
1503 1411 dnode_t *dn;
1504 1412 int err;
1505 1413
1506 1414 if (size == 0)
1507 1415 return (0);
1508 1416
1509 1417 err = dnode_hold(os, object, FTAG, &dn);
1510 1418 if (err)
1511 1419 return (err);
1512 1420
1513 1421 err = dmu_write_uio_dnode(dn, uio, size, tx);
1514 1422
1515 1423 dnode_rele(dn, FTAG);
1516 1424
1517 1425 return (err);
1518 1426 }
1519 1427
1520 1428 int
1521 1429 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1522 1430 page_t *pp, dmu_tx_t *tx)
1523 1431 {
1524 1432 dmu_buf_t **dbp;
1525 1433 int numbufs, i;
1526 1434 int err;
1527 1435
1528 1436 if (size == 0)
1529 1437 return (0);
1530 1438
1531 1439 err = dmu_buf_hold_array(os, object, offset, size,
1532 1440 FALSE, FTAG, &numbufs, &dbp);
1533 1441 if (err)
1534 1442 return (err);
1535 1443
1536 1444 for (i = 0; i < numbufs; i++) {
1537 1445 int tocpy, copied, thiscpy;
1538 1446 int bufoff;
1539 1447 dmu_buf_t *db = dbp[i];
1540 1448 caddr_t va;
1541 1449
1542 1450 ASSERT(size > 0);
1543 1451 ASSERT3U(db->db_size, >=, PAGESIZE);
1544 1452
1545 1453 bufoff = offset - db->db_offset;
1546 1454 tocpy = (int)MIN(db->db_size - bufoff, size);
1547 1455
1548 1456 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1549 1457
1550 1458 if (tocpy == db->db_size)
1551 1459 dmu_buf_will_fill(db, tx);
1552 1460 else
1553 1461 dmu_buf_will_dirty(db, tx);
1554 1462
1555 1463 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1556 1464 ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff);
1557 1465 thiscpy = MIN(PAGESIZE, tocpy - copied);
1558 1466 va = zfs_map_page(pp, S_READ);
1559 1467 bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1560 1468 zfs_unmap_page(pp, va);
1561 1469 pp = pp->p_next;
1562 1470 bufoff += PAGESIZE;
1563 1471 }
1564 1472
1565 1473 if (tocpy == db->db_size)
1566 1474 dmu_buf_fill_done(db, tx);
1567 1475
1568 1476 offset += tocpy;
1569 1477 size -= tocpy;
1570 1478 }
1571 1479 dmu_buf_rele_array(dbp, numbufs, FTAG);
1572 1480 return (err);
1573 1481 }
1574 1482 #endif
1575 1483
1576 1484 /*
1577 1485 * Allocate a loaned anonymous arc buffer.
1578 1486 */
1579 1487 arc_buf_t *
1580 1488 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1581 1489 {
1582 1490 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1583 1491
1584 1492 return (arc_loan_buf(db->db_objset->os_spa, B_FALSE, size));
1585 1493 }
1586 1494
1587 1495 /*
1588 1496 * Free a loaned arc buffer.
1589 1497 */
1590 1498 void
1591 1499 dmu_return_arcbuf(arc_buf_t *buf)
1592 1500 {
1593 1501 arc_return_buf(buf, FTAG);
1594 1502 arc_buf_destroy(buf, FTAG);
1595 1503 }
1596 1504
1597 1505 /*
1598 1506 * When possible directly assign passed loaned arc buffer to a dbuf.
1599 1507 * If this is not possible copy the contents of passed arc buf via
1600 1508 * dmu_write().
1601 1509 */
1602 1510 void
1603 1511 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1604 1512 dmu_tx_t *tx)
1605 1513 {
1606 1514 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1607 1515 dnode_t *dn;
1608 1516 dmu_buf_impl_t *db;
1609 1517 uint32_t blksz = (uint32_t)arc_buf_lsize(buf);
1610 1518 uint64_t blkid;
1611 1519
1612 1520 DB_DNODE_ENTER(dbuf);
1613 1521 dn = DB_DNODE(dbuf);
1614 1522 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1615 1523 blkid = dbuf_whichblock(dn, 0, offset);
1616 1524 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1617 1525 rw_exit(&dn->dn_struct_rwlock);
1618 1526 DB_DNODE_EXIT(dbuf);
1619 1527
1620 1528 /*
1621 1529 * We can only assign if the offset is aligned, the arc buf is the
1622 1530 * same size as the dbuf, and the dbuf is not metadata.
1623 1531 */
1624 1532 if (offset == db->db.db_offset && blksz == db->db.db_size) {
1625 1533 dbuf_assign_arcbuf(db, buf, tx);
1626 1534 dbuf_rele(db, FTAG);
1627 1535 } else {
1628 1536 objset_t *os;
1629 1537 uint64_t object;
1630 1538
1631 1539 /* compressed bufs must always be assignable to their dbuf */
1632 1540 ASSERT3U(arc_get_compression(buf), ==, ZIO_COMPRESS_OFF);
1633 1541 ASSERT(!(buf->b_flags & ARC_BUF_FLAG_COMPRESSED));
1634 1542
1635 1543 DB_DNODE_ENTER(dbuf);
1636 1544 dn = DB_DNODE(dbuf);
1637 1545 os = dn->dn_objset;
1638 1546 object = dn->dn_object;
1639 1547 DB_DNODE_EXIT(dbuf);
1640 1548
1641 1549 dbuf_rele(db, FTAG);
1642 1550 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1643 1551 dmu_return_arcbuf(buf);
1644 1552 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1645 1553 }
1646 1554 }
1647 1555
1648 1556 typedef struct {
1649 1557 dbuf_dirty_record_t *dsa_dr;
1650 1558 dmu_sync_cb_t *dsa_done;
1651 1559 zgd_t *dsa_zgd;
1652 1560 dmu_tx_t *dsa_tx;
1653 1561 } dmu_sync_arg_t;
1654 1562
1655 1563 /* ARGSUSED */
1656 1564 static void
1657 1565 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1658 1566 {
1659 1567 dmu_sync_arg_t *dsa = varg;
1660 1568 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1661 1569 blkptr_t *bp = zio->io_bp;
1662 1570
1663 1571 if (zio->io_error == 0) {
1664 1572 if (BP_IS_HOLE(bp)) {
1665 1573 /*
1666 1574 * A block of zeros may compress to a hole, but the
1667 1575 * block size still needs to be known for replay.
1668 1576 */
1669 1577 BP_SET_LSIZE(bp, db->db_size);
1670 1578 } else if (!BP_IS_EMBEDDED(bp)) {
1671 1579 ASSERT(BP_GET_LEVEL(bp) == 0);
1672 1580 bp->blk_fill = 1;
1673 1581 }
1674 1582 }
1675 1583 }
1676 1584
1677 1585 static void
1678 1586 dmu_sync_late_arrival_ready(zio_t *zio)
1679 1587 {
|
↓ open down ↓ |
529 lines elided |
↑ open up ↑ |
1680 1588 dmu_sync_ready(zio, NULL, zio->io_private);
1681 1589 }
1682 1590
1683 1591 /* ARGSUSED */
1684 1592 static void
1685 1593 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1686 1594 {
1687 1595 dmu_sync_arg_t *dsa = varg;
1688 1596 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1689 1597 dmu_buf_impl_t *db = dr->dr_dbuf;
1598 + zgd_t *zgd = dsa->dsa_zgd;
1690 1599
1600 + /*
1601 + * Record the vdev(s) backing this blkptr so they can be flushed after
1602 + * the writes for the lwb have completed.
1603 + */
1604 + if (zio->io_error == 0) {
1605 + zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp);
1606 + }
1607 +
1691 1608 mutex_enter(&db->db_mtx);
1692 1609 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1693 1610 if (zio->io_error == 0) {
1694 1611 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1695 1612 if (dr->dt.dl.dr_nopwrite) {
1696 1613 blkptr_t *bp = zio->io_bp;
1697 1614 blkptr_t *bp_orig = &zio->io_bp_orig;
1698 1615 uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1699 1616
1700 1617 ASSERT(BP_EQUAL(bp, bp_orig));
1701 1618 VERIFY(BP_EQUAL(bp, db->db_blkptr));
1702 1619 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1703 1620 ASSERT(zio_checksum_table[chksum].ci_flags &
1704 1621 ZCHECKSUM_FLAG_NOPWRITE);
1705 1622 }
1706 1623 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1707 1624 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1708 1625 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1709 1626
1710 1627 /*
1711 1628 * Old style holes are filled with all zeros, whereas
1712 1629 * new-style holes maintain their lsize, type, level,
1713 1630 * and birth time (see zio_write_compress). While we
1714 1631 * need to reset the BP_SET_LSIZE() call that happened
1715 1632 * in dmu_sync_ready for old style holes, we do *not*
1716 1633 * want to wipe out the information contained in new
1717 1634 * style holes. Thus, only zero out the block pointer if
1718 1635 * it's an old style hole.
1719 1636 */
1720 1637 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) &&
1721 1638 dr->dt.dl.dr_overridden_by.blk_birth == 0)
1722 1639 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1723 1640 } else {
1724 1641 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1725 1642 }
1726 1643 cv_broadcast(&db->db_changed);
1727 1644 mutex_exit(&db->db_mtx);
1728 1645
1729 1646 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
|
↓ open down ↓ |
29 lines elided |
↑ open up ↑ |
1730 1647
1731 1648 kmem_free(dsa, sizeof (*dsa));
1732 1649 }
1733 1650
1734 1651 static void
1735 1652 dmu_sync_late_arrival_done(zio_t *zio)
1736 1653 {
1737 1654 blkptr_t *bp = zio->io_bp;
1738 1655 dmu_sync_arg_t *dsa = zio->io_private;
1739 1656 blkptr_t *bp_orig = &zio->io_bp_orig;
1657 + zgd_t *zgd = dsa->dsa_zgd;
1740 1658
1741 - if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1742 - ASSERT(!(zio->io_flags & ZIO_FLAG_NOPWRITE));
1743 - ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1744 - ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1745 - ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1746 - zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1659 + if (zio->io_error == 0) {
1660 + /*
1661 + * Record the vdev(s) backing this blkptr so they can be
1662 + * flushed after the writes for the lwb have completed.
1663 + */
1664 + zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp);
1665 +
1666 + if (!BP_IS_HOLE(bp)) {
1667 + ASSERT(!(zio->io_flags & ZIO_FLAG_NOPWRITE));
1668 + ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1669 + ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1670 + ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1671 + zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1672 + }
1747 1673 }
1748 1674
1749 1675 dmu_tx_commit(dsa->dsa_tx);
1750 1676
1751 1677 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1752 1678
1753 1679 abd_put(zio->io_abd);
1754 1680 kmem_free(dsa, sizeof (*dsa));
1755 1681 }
1756 1682
1757 1683 static int
1758 1684 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1759 - zio_prop_t *zp, zbookmark_phys_t *zb)
1685 + zio_prop_t *zp, zbookmark_phys_t *zb, const zio_smartcomp_info_t *sc)
1760 1686 {
1761 1687 dmu_sync_arg_t *dsa;
1762 1688 dmu_tx_t *tx;
1763 1689
1764 1690 tx = dmu_tx_create(os);
1765 1691 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1766 1692 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1767 1693 dmu_tx_abort(tx);
1768 1694 /* Make zl_get_data do txg_waited_synced() */
1769 1695 return (SET_ERROR(EIO));
1770 1696 }
1771 1697
1772 1698 /*
1773 1699 * In order to prevent the zgd's lwb from being free'd prior to
1774 1700 * dmu_sync_late_arrival_done() being called, we have to ensure
1775 1701 * the lwb's "max txg" takes this tx's txg into account.
1776 1702 */
1777 1703 zil_lwb_add_txg(zgd->zgd_lwb, dmu_tx_get_txg(tx));
1778 1704
1779 1705 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1780 1706 dsa->dsa_dr = NULL;
1781 1707 dsa->dsa_done = done;
1782 1708 dsa->dsa_zgd = zgd;
1783 1709 dsa->dsa_tx = tx;
1784 1710
1785 1711 /*
1786 1712 * Since we are currently syncing this txg, it's nontrivial to
1787 1713 * determine what BP to nopwrite against, so we disable nopwrite.
1788 1714 *
1789 1715 * When syncing, the db_blkptr is initially the BP of the previous
1790 1716 * txg. We can not nopwrite against it because it will be changed
1791 1717 * (this is similar to the non-late-arrival case where the dbuf is
1792 1718 * dirty in a future txg).
1793 1719 *
1794 1720 * Then dbuf_write_ready() sets bp_blkptr to the location we will write.
1795 1721 * We can not nopwrite against it because although the BP will not
1796 1722 * (typically) be changed, the data has not yet been persisted to this
1797 1723 * location.
1798 1724 *
1799 1725 * Finally, when dbuf_write_done() is called, it is theoretically
1800 1726 * possible to always nopwrite, because the data that was written in
1801 1727 * this txg is the same data that we are trying to write. However we
|
↓ open down ↓ |
32 lines elided |
↑ open up ↑ |
1802 1728 * would need to check that this dbuf is not dirty in any future
1803 1729 * txg's (as we do in the normal dmu_sync() path). For simplicity, we
1804 1730 * don't nopwrite in this case.
1805 1731 */
1806 1732 zp->zp_nopwrite = B_FALSE;
1807 1733
1808 1734 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1809 1735 abd_get_from_buf(zgd->zgd_db->db_data, zgd->zgd_db->db_size),
1810 1736 zgd->zgd_db->db_size, zgd->zgd_db->db_size, zp,
1811 1737 dmu_sync_late_arrival_ready, NULL, NULL, dmu_sync_late_arrival_done,
1812 - dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1738 + dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb, sc));
1813 1739
1814 1740 return (0);
1815 1741 }
1816 1742
1817 1743 /*
1818 1744 * Intent log support: sync the block associated with db to disk.
1819 1745 * N.B. and XXX: the caller is responsible for making sure that the
1820 1746 * data isn't changing while dmu_sync() is writing it.
1821 1747 *
1822 1748 * Return values:
1823 1749 *
1824 1750 * EEXIST: this txg has already been synced, so there's nothing to do.
1825 1751 * The caller should not log the write.
1826 1752 *
1827 1753 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1828 1754 * The caller should not log the write.
1829 1755 *
1830 1756 * EALREADY: this block is already in the process of being synced.
1831 1757 * The caller should track its progress (somehow).
|
↓ open down ↓ |
9 lines elided |
↑ open up ↑ |
1832 1758 *
1833 1759 * EIO: could not do the I/O.
1834 1760 * The caller should do a txg_wait_synced().
1835 1761 *
1836 1762 * 0: the I/O has been initiated.
1837 1763 * The caller should log this blkptr in the done callback.
1838 1764 * It is possible that the I/O will fail, in which case
1839 1765 * the error will be reported to the done callback and
1840 1766 * propagated to pio from zio_done().
1841 1767 */
1768 +
1842 1769 int
1843 1770 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1844 1771 {
1845 1772 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1846 1773 objset_t *os = db->db_objset;
1847 1774 dsl_dataset_t *ds = os->os_dsl_dataset;
1848 1775 dbuf_dirty_record_t *dr;
1849 1776 dmu_sync_arg_t *dsa;
1850 1777 zbookmark_phys_t zb;
1851 1778 zio_prop_t zp;
1852 1779 dnode_t *dn;
1780 + int flags = 0;
1781 + zio_smartcomp_info_t sc;
1853 1782
1854 1783 ASSERT(pio != NULL);
1855 1784 ASSERT(txg != 0);
1856 1785
1857 1786 SET_BOOKMARK(&zb, ds->ds_object,
1858 1787 db->db.db_object, db->db_level, db->db_blkid);
1859 1788
1789 + /* write to special only if proper conditions hold */
1790 + if (spa_write_data_to_special(os->os_spa, os))
1791 + WP_SET_SPECIALCLASS(flags, B_TRUE);
1792 +
1860 1793 DB_DNODE_ENTER(db);
1861 1794 dn = DB_DNODE(db);
1862 - dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1795 + dmu_write_policy(os, dn, db->db_level, flags | WP_DMU_SYNC, &zp);
1796 + dnode_setup_zio_smartcomp(db, &sc);
1863 1797 DB_DNODE_EXIT(db);
1864 1798
1865 1799 /*
1866 1800 * If we're frozen (running ziltest), we always need to generate a bp.
1867 1801 */
1868 1802 if (txg > spa_freeze_txg(os->os_spa))
1869 - return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1803 + return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb,
1804 + &sc));
1870 1805
1871 1806 /*
1872 1807 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1873 1808 * and us. If we determine that this txg is not yet syncing,
1874 1809 * but it begins to sync a moment later, that's OK because the
1875 1810 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1876 1811 */
1877 1812 mutex_enter(&db->db_mtx);
1878 1813
1879 1814 if (txg <= spa_last_synced_txg(os->os_spa)) {
1880 1815 /*
1881 1816 * This txg has already synced. There's nothing to do.
1882 1817 */
|
↓ open down ↓ |
3 lines elided |
↑ open up ↑ |
1883 1818 mutex_exit(&db->db_mtx);
1884 1819 return (SET_ERROR(EEXIST));
1885 1820 }
1886 1821
1887 1822 if (txg <= spa_syncing_txg(os->os_spa)) {
1888 1823 /*
1889 1824 * This txg is currently syncing, so we can't mess with
1890 1825 * the dirty record anymore; just write a new log block.
1891 1826 */
1892 1827 mutex_exit(&db->db_mtx);
1893 - return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1828 + return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb,
1829 + &sc));
1894 1830 }
1895 1831
1896 1832 dr = db->db_last_dirty;
1897 1833 while (dr && dr->dr_txg != txg)
1898 1834 dr = dr->dr_next;
1899 1835
1900 1836 if (dr == NULL) {
1901 1837 /*
1902 1838 * There's no dr for this dbuf, so it must have been freed.
1903 1839 * There's no need to log writes to freed blocks, so we're done.
1904 1840 */
1905 1841 mutex_exit(&db->db_mtx);
1906 1842 return (SET_ERROR(ENOENT));
1907 1843 }
1908 1844
1909 1845 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
1910 1846
1911 1847 if (db->db_blkptr != NULL) {
1912 1848 /*
1913 1849 * We need to fill in zgd_bp with the current blkptr so that
1914 1850 * the nopwrite code can check if we're writing the same
1915 1851 * data that's already on disk. We can only nopwrite if we
1916 1852 * are sure that after making the copy, db_blkptr will not
1917 1853 * change until our i/o completes. We ensure this by
1918 1854 * holding the db_mtx, and only allowing nopwrite if the
1919 1855 * block is not already dirty (see below). This is verified
1920 1856 * by dmu_sync_done(), which VERIFYs that the db_blkptr has
1921 1857 * not changed.
1922 1858 */
1923 1859 *zgd->zgd_bp = *db->db_blkptr;
1924 1860 }
1925 1861
1926 1862 /*
1927 1863 * Assume the on-disk data is X, the current syncing data (in
1928 1864 * txg - 1) is Y, and the current in-memory data is Z (currently
1929 1865 * in dmu_sync).
1930 1866 *
1931 1867 * We usually want to perform a nopwrite if X and Z are the
1932 1868 * same. However, if Y is different (i.e. the BP is going to
1933 1869 * change before this write takes effect), then a nopwrite will
1934 1870 * be incorrect - we would override with X, which could have
1935 1871 * been freed when Y was written.
1936 1872 *
1937 1873 * (Note that this is not a concern when we are nop-writing from
1938 1874 * syncing context, because X and Y must be identical, because
1939 1875 * all previous txgs have been synced.)
1940 1876 *
1941 1877 * Therefore, we disable nopwrite if the current BP could change
1942 1878 * before this TXG. There are two ways it could change: by
1943 1879 * being dirty (dr_next is non-NULL), or by being freed
1944 1880 * (dnode_block_freed()). This behavior is verified by
1945 1881 * zio_done(), which VERIFYs that the override BP is identical
1946 1882 * to the on-disk BP.
1947 1883 */
1948 1884 DB_DNODE_ENTER(db);
1949 1885 dn = DB_DNODE(db);
1950 1886 if (dr->dr_next != NULL || dnode_block_freed(dn, db->db_blkid))
1951 1887 zp.zp_nopwrite = B_FALSE;
1952 1888 DB_DNODE_EXIT(db);
1953 1889
1954 1890 ASSERT(dr->dr_txg == txg);
1955 1891 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1956 1892 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1957 1893 /*
1958 1894 * We have already issued a sync write for this buffer,
1959 1895 * or this buffer has already been synced. It could not
1960 1896 * have been dirtied since, or we would have cleared the state.
1961 1897 */
1962 1898 mutex_exit(&db->db_mtx);
1963 1899 return (SET_ERROR(EALREADY));
1964 1900 }
1965 1901
1966 1902 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1967 1903 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1968 1904 mutex_exit(&db->db_mtx);
|
↓ open down ↓ |
65 lines elided |
↑ open up ↑ |
1969 1905
1970 1906 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1971 1907 dsa->dsa_dr = dr;
1972 1908 dsa->dsa_done = done;
1973 1909 dsa->dsa_zgd = zgd;
1974 1910 dsa->dsa_tx = NULL;
1975 1911
1976 1912 zio_nowait(arc_write(pio, os->os_spa, txg,
1977 1913 zgd->zgd_bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
1978 1914 &zp, dmu_sync_ready, NULL, NULL, dmu_sync_done, dsa,
1979 - ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb));
1915 + ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb, &sc));
1980 1916
1981 1917 return (0);
1982 1918 }
1983 1919
1984 1920 int
1985 1921 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1986 1922 dmu_tx_t *tx)
1987 1923 {
1988 1924 dnode_t *dn;
1989 1925 int err;
1990 1926
1991 1927 err = dnode_hold(os, object, FTAG, &dn);
1992 1928 if (err)
1993 1929 return (err);
1994 1930 err = dnode_set_blksz(dn, size, ibs, tx);
1995 1931 dnode_rele(dn, FTAG);
1996 1932 return (err);
1997 1933 }
1998 1934
1999 1935 void
2000 1936 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
2001 1937 dmu_tx_t *tx)
2002 1938 {
2003 1939 dnode_t *dn;
2004 1940
2005 1941 /*
2006 1942 * Send streams include each object's checksum function. This
2007 1943 * check ensures that the receiving system can understand the
2008 1944 * checksum function transmitted.
2009 1945 */
2010 1946 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS);
2011 1947
2012 1948 VERIFY0(dnode_hold(os, object, FTAG, &dn));
2013 1949 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS);
2014 1950 dn->dn_checksum = checksum;
2015 1951 dnode_setdirty(dn, tx);
2016 1952 dnode_rele(dn, FTAG);
2017 1953 }
2018 1954
2019 1955 void
2020 1956 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
2021 1957 dmu_tx_t *tx)
2022 1958 {
2023 1959 dnode_t *dn;
2024 1960
2025 1961 /*
2026 1962 * Send streams include each object's compression function. This
2027 1963 * check ensures that the receiving system can understand the
2028 1964 * compression function transmitted.
2029 1965 */
2030 1966 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS);
2031 1967
2032 1968 VERIFY0(dnode_hold(os, object, FTAG, &dn));
2033 1969 dn->dn_compress = compress;
2034 1970 dnode_setdirty(dn, tx);
2035 1971 dnode_rele(dn, FTAG);
2036 1972 }
2037 1973
2038 1974 int zfs_mdcomp_disable = 0;
2039 1975
2040 1976 /*
2041 1977 * When the "redundant_metadata" property is set to "most", only indirect
2042 1978 * blocks of this level and higher will have an additional ditto block.
2043 1979 */
2044 1980 int zfs_redundant_metadata_most_ditto_level = 2;
2045 1981
2046 1982 void
2047 1983 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
2048 1984 {
2049 1985 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
2050 1986 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
2051 1987 (wp & WP_SPILL));
2052 1988 enum zio_checksum checksum = os->os_checksum;
2053 1989 enum zio_compress compress = os->os_compress;
2054 1990 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
2055 1991 boolean_t dedup = B_FALSE;
2056 1992 boolean_t nopwrite = B_FALSE;
2057 1993 boolean_t dedup_verify = os->os_dedup_verify;
2058 1994 int copies = os->os_copies;
2059 1995
2060 1996 /*
2061 1997 * We maintain different write policies for each of the following
2062 1998 * types of data:
2063 1999 * 1. metadata
2064 2000 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
2065 2001 * 3. all other level 0 blocks
2066 2002 */
2067 2003 if (ismd) {
2068 2004 if (zfs_mdcomp_disable) {
2069 2005 compress = ZIO_COMPRESS_EMPTY;
2070 2006 } else {
2071 2007 /*
2072 2008 * XXX -- we should design a compression algorithm
2073 2009 * that specializes in arrays of bps.
2074 2010 */
2075 2011 compress = zio_compress_select(os->os_spa,
2076 2012 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON);
2077 2013 }
2078 2014
2079 2015 /*
2080 2016 * Metadata always gets checksummed. If the data
2081 2017 * checksum is multi-bit correctable, and it's not a
2082 2018 * ZBT-style checksum, then it's suitable for metadata
2083 2019 * as well. Otherwise, the metadata checksum defaults
2084 2020 * to fletcher4.
2085 2021 */
2086 2022 if (!(zio_checksum_table[checksum].ci_flags &
2087 2023 ZCHECKSUM_FLAG_METADATA) ||
2088 2024 (zio_checksum_table[checksum].ci_flags &
2089 2025 ZCHECKSUM_FLAG_EMBEDDED))
2090 2026 checksum = ZIO_CHECKSUM_FLETCHER_4;
2091 2027
2092 2028 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL ||
2093 2029 (os->os_redundant_metadata ==
2094 2030 ZFS_REDUNDANT_METADATA_MOST &&
2095 2031 (level >= zfs_redundant_metadata_most_ditto_level ||
2096 2032 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))))
2097 2033 copies++;
2098 2034 } else if (wp & WP_NOFILL) {
2099 2035 ASSERT(level == 0);
2100 2036
2101 2037 /*
2102 2038 * If we're writing preallocated blocks, we aren't actually
2103 2039 * writing them so don't set any policy properties. These
2104 2040 * blocks are currently only used by an external subsystem
2105 2041 * outside of zfs (i.e. dump) and not written by the zio
2106 2042 * pipeline.
2107 2043 */
2108 2044 compress = ZIO_COMPRESS_OFF;
2109 2045 checksum = ZIO_CHECKSUM_NOPARITY;
2110 2046 } else {
2111 2047 compress = zio_compress_select(os->os_spa, dn->dn_compress,
2112 2048 compress);
2113 2049
2114 2050 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
2115 2051 zio_checksum_select(dn->dn_checksum, checksum) :
2116 2052 dedup_checksum;
2117 2053
2118 2054 /*
2119 2055 * Determine dedup setting. If we are in dmu_sync(),
2120 2056 * we won't actually dedup now because that's all
2121 2057 * done in syncing context; but we do want to use the
2122 2058 * dedup checkum. If the checksum is not strong
2123 2059 * enough to ensure unique signatures, force
2124 2060 * dedup_verify.
2125 2061 */
2126 2062 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
2127 2063 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
2128 2064 if (!(zio_checksum_table[checksum].ci_flags &
2129 2065 ZCHECKSUM_FLAG_DEDUP))
2130 2066 dedup_verify = B_TRUE;
2131 2067 }
2132 2068
2133 2069 /*
2134 2070 * Enable nopwrite if we have secure enough checksum
|
↓ open down ↓ |
145 lines elided |
↑ open up ↑ |
2135 2071 * algorithm (see comment in zio_nop_write) and
2136 2072 * compression is enabled. We don't enable nopwrite if
2137 2073 * dedup is enabled as the two features are mutually
2138 2074 * exclusive.
2139 2075 */
2140 2076 nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags &
2141 2077 ZCHECKSUM_FLAG_NOPWRITE) &&
2142 2078 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
2143 2079 }
2144 2080
2081 + zp->zp_usesc = WP_GET_SPECIALCLASS(wp);
2145 2082 zp->zp_checksum = checksum;
2146 2083 zp->zp_compress = compress;
2147 2084 ASSERT3U(zp->zp_compress, !=, ZIO_COMPRESS_INHERIT);
2148 2085
2149 2086 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
2150 2087 zp->zp_level = level;
2151 2088 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
2152 2089 zp->zp_dedup = dedup;
2153 2090 zp->zp_dedup_verify = dedup && dedup_verify;
2091 + zp->zp_metadata = ismd;
2154 2092 zp->zp_nopwrite = nopwrite;
2093 + zp->zp_zpl_meta_to_special = os->os_zpl_meta_to_special;
2094 + zp->zp_usewbc = (zp->zp_usesc &&
2095 + os->os_wbc_mode == ZFS_WBC_MODE_ON && !ismd);
2096 +
2097 + /* explicitly control the number for copies for DDT */
2098 + if (DMU_OT_IS_DDT_META(type) &&
2099 + os->os_spa->spa_ddt_meta_copies > 0) {
2100 + zp->zp_copies =
2101 + MIN(os->os_spa->spa_ddt_meta_copies,
2102 + spa_max_replication(os->os_spa));
2103 + }
2104 +
2105 + DTRACE_PROBE2(dmu_wp, boolean_t, zp->zp_metadata,
2106 + boolean_t, zp->zp_usesc);
2155 2107 }
2156 2108
2157 2109 int
2158 2110 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
2159 2111 {
2160 2112 dnode_t *dn;
2161 2113 int err;
2162 2114
2163 2115 /*
2164 2116 * Sync any current changes before
2165 2117 * we go trundling through the block pointers.
2166 2118 */
2167 2119 err = dmu_object_wait_synced(os, object);
2168 2120 if (err) {
2169 2121 return (err);
2170 2122 }
2171 2123
2172 2124 err = dnode_hold(os, object, FTAG, &dn);
2173 2125 if (err) {
2174 2126 return (err);
2175 2127 }
2176 2128
2177 2129 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
2178 2130 dnode_rele(dn, FTAG);
2179 2131
2180 2132 return (err);
2181 2133 }
2182 2134
2183 2135 /*
2184 2136 * Given the ZFS object, if it contains any dirty nodes
2185 2137 * this function flushes all dirty blocks to disk. This
2186 2138 * ensures the DMU object info is updated. A more efficient
2187 2139 * future version might just find the TXG with the maximum
2188 2140 * ID and wait for that to be synced.
2189 2141 */
2190 2142 int
2191 2143 dmu_object_wait_synced(objset_t *os, uint64_t object)
2192 2144 {
2193 2145 dnode_t *dn;
2194 2146 int error, i;
2195 2147
2196 2148 error = dnode_hold(os, object, FTAG, &dn);
2197 2149 if (error) {
2198 2150 return (error);
2199 2151 }
2200 2152
2201 2153 for (i = 0; i < TXG_SIZE; i++) {
2202 2154 if (list_link_active(&dn->dn_dirty_link[i])) {
2203 2155 break;
2204 2156 }
2205 2157 }
2206 2158 dnode_rele(dn, FTAG);
2207 2159 if (i != TXG_SIZE) {
2208 2160 txg_wait_synced(dmu_objset_pool(os), 0);
2209 2161 }
2210 2162
2211 2163 return (0);
2212 2164 }
2213 2165
2214 2166 void
2215 2167 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
2216 2168 {
2217 2169 dnode_phys_t *dnp;
2218 2170
2219 2171 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2220 2172 mutex_enter(&dn->dn_mtx);
2221 2173
2222 2174 dnp = dn->dn_phys;
2223 2175
2224 2176 doi->doi_data_block_size = dn->dn_datablksz;
2225 2177 doi->doi_metadata_block_size = dn->dn_indblkshift ?
2226 2178 1ULL << dn->dn_indblkshift : 0;
2227 2179 doi->doi_type = dn->dn_type;
2228 2180 doi->doi_bonus_type = dn->dn_bonustype;
2229 2181 doi->doi_bonus_size = dn->dn_bonuslen;
2230 2182 doi->doi_indirection = dn->dn_nlevels;
2231 2183 doi->doi_checksum = dn->dn_checksum;
2232 2184 doi->doi_compress = dn->dn_compress;
2233 2185 doi->doi_nblkptr = dn->dn_nblkptr;
2234 2186 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
2235 2187 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
2236 2188 doi->doi_fill_count = 0;
2237 2189 for (int i = 0; i < dnp->dn_nblkptr; i++)
2238 2190 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
2239 2191
2240 2192 mutex_exit(&dn->dn_mtx);
2241 2193 rw_exit(&dn->dn_struct_rwlock);
2242 2194 }
2243 2195
2244 2196 /*
2245 2197 * Get information on a DMU object.
2246 2198 * If doi is NULL, just indicates whether the object exists.
2247 2199 */
2248 2200 int
2249 2201 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
2250 2202 {
2251 2203 dnode_t *dn;
2252 2204 int err = dnode_hold(os, object, FTAG, &dn);
2253 2205
2254 2206 if (err)
2255 2207 return (err);
2256 2208
2257 2209 if (doi != NULL)
2258 2210 dmu_object_info_from_dnode(dn, doi);
2259 2211
2260 2212 dnode_rele(dn, FTAG);
2261 2213 return (0);
2262 2214 }
2263 2215
2264 2216 /*
2265 2217 * As above, but faster; can be used when you have a held dbuf in hand.
2266 2218 */
2267 2219 void
2268 2220 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
2269 2221 {
2270 2222 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2271 2223
2272 2224 DB_DNODE_ENTER(db);
2273 2225 dmu_object_info_from_dnode(DB_DNODE(db), doi);
2274 2226 DB_DNODE_EXIT(db);
2275 2227 }
2276 2228
2277 2229 /*
2278 2230 * Faster still when you only care about the size.
2279 2231 * This is specifically optimized for zfs_getattr().
2280 2232 */
2281 2233 void
2282 2234 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
2283 2235 u_longlong_t *nblk512)
2284 2236 {
2285 2237 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
2286 2238 dnode_t *dn;
2287 2239
2288 2240 DB_DNODE_ENTER(db);
2289 2241 dn = DB_DNODE(db);
2290 2242
2291 2243 *blksize = dn->dn_datablksz;
2292 2244 /* add 1 for dnode space */
2293 2245 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
2294 2246 SPA_MINBLOCKSHIFT) + 1;
2295 2247 DB_DNODE_EXIT(db);
2296 2248 }
2297 2249
2298 2250 void
2299 2251 byteswap_uint64_array(void *vbuf, size_t size)
2300 2252 {
2301 2253 uint64_t *buf = vbuf;
2302 2254 size_t count = size >> 3;
2303 2255 int i;
2304 2256
2305 2257 ASSERT((size & 7) == 0);
2306 2258
2307 2259 for (i = 0; i < count; i++)
2308 2260 buf[i] = BSWAP_64(buf[i]);
2309 2261 }
2310 2262
2311 2263 void
2312 2264 byteswap_uint32_array(void *vbuf, size_t size)
2313 2265 {
2314 2266 uint32_t *buf = vbuf;
2315 2267 size_t count = size >> 2;
2316 2268 int i;
2317 2269
2318 2270 ASSERT((size & 3) == 0);
2319 2271
2320 2272 for (i = 0; i < count; i++)
2321 2273 buf[i] = BSWAP_32(buf[i]);
2322 2274 }
2323 2275
2324 2276 void
2325 2277 byteswap_uint16_array(void *vbuf, size_t size)
2326 2278 {
2327 2279 uint16_t *buf = vbuf;
2328 2280 size_t count = size >> 1;
2329 2281 int i;
2330 2282
2331 2283 ASSERT((size & 1) == 0);
2332 2284
2333 2285 for (i = 0; i < count; i++)
2334 2286 buf[i] = BSWAP_16(buf[i]);
2335 2287 }
2336 2288
2337 2289 /* ARGSUSED */
2338 2290 void
2339 2291 byteswap_uint8_array(void *vbuf, size_t size)
2340 2292 {
2341 2293 }
2342 2294
2343 2295 void
2344 2296 dmu_init(void)
2345 2297 {
2346 2298 abd_init();
2347 2299 zfs_dbgmsg_init();
2348 2300 sa_cache_init();
2349 2301 xuio_stat_init();
2350 2302 dmu_objset_init();
2351 2303 dnode_init();
2352 2304 zfetch_init();
2353 2305 l2arc_init();
2354 2306 arc_init();
2355 2307 dbuf_init();
2356 2308 }
2357 2309
2358 2310 void
2359 2311 dmu_fini(void)
2360 2312 {
2361 2313 arc_fini(); /* arc depends on l2arc, so arc must go first */
2362 2314 l2arc_fini();
2363 2315 zfetch_fini();
2364 2316 dbuf_fini();
2365 2317 dnode_fini();
2366 2318 dmu_objset_fini();
2367 2319 xuio_stat_fini();
2368 2320 sa_cache_fini();
2369 2321 zfs_dbgmsg_fini();
2370 2322 abd_fini();
2371 2323 }
|
↓ open down ↓ |
207 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX