Print this page
Possibility to physically reserve space without writing leaf blocks
| 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, 2014 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 (c) 2014, 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>
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 48 #include <sys/zfeature.h>
49 49 #ifdef _KERNEL
50 50 #include <sys/vmsystm.h>
51 51 #include <sys/zfs_znode.h>
52 52 #endif
53 53
54 54 /*
55 55 * Enable/disable nopwrite feature.
56 56 */
57 57 int zfs_nopwrite_enabled = 1;
58 58
59 59 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
60 60 { DMU_BSWAP_UINT8, TRUE, "unallocated" },
61 61 { DMU_BSWAP_ZAP, TRUE, "object directory" },
62 62 { DMU_BSWAP_UINT64, TRUE, "object array" },
63 63 { DMU_BSWAP_UINT8, TRUE, "packed nvlist" },
64 64 { DMU_BSWAP_UINT64, TRUE, "packed nvlist size" },
65 65 { DMU_BSWAP_UINT64, TRUE, "bpobj" },
66 66 { DMU_BSWAP_UINT64, TRUE, "bpobj header" },
67 67 { DMU_BSWAP_UINT64, TRUE, "SPA space map header" },
68 68 { DMU_BSWAP_UINT64, TRUE, "SPA space map" },
69 69 { DMU_BSWAP_UINT64, TRUE, "ZIL intent log" },
70 70 { DMU_BSWAP_DNODE, TRUE, "DMU dnode" },
71 71 { DMU_BSWAP_OBJSET, TRUE, "DMU objset" },
72 72 { DMU_BSWAP_UINT64, TRUE, "DSL directory" },
73 73 { DMU_BSWAP_ZAP, TRUE, "DSL directory child map"},
74 74 { DMU_BSWAP_ZAP, TRUE, "DSL dataset snap map" },
75 75 { DMU_BSWAP_ZAP, TRUE, "DSL props" },
76 76 { DMU_BSWAP_UINT64, TRUE, "DSL dataset" },
77 77 { DMU_BSWAP_ZNODE, TRUE, "ZFS znode" },
78 78 { DMU_BSWAP_OLDACL, TRUE, "ZFS V0 ACL" },
79 79 { DMU_BSWAP_UINT8, FALSE, "ZFS plain file" },
80 80 { DMU_BSWAP_ZAP, TRUE, "ZFS directory" },
81 81 { DMU_BSWAP_ZAP, TRUE, "ZFS master node" },
82 82 { DMU_BSWAP_ZAP, TRUE, "ZFS delete queue" },
83 83 { DMU_BSWAP_UINT8, FALSE, "zvol object" },
84 84 { DMU_BSWAP_ZAP, TRUE, "zvol prop" },
85 85 { DMU_BSWAP_UINT8, FALSE, "other uint8[]" },
86 86 { DMU_BSWAP_UINT64, FALSE, "other uint64[]" },
87 87 { DMU_BSWAP_ZAP, TRUE, "other ZAP" },
88 88 { DMU_BSWAP_ZAP, TRUE, "persistent error log" },
89 89 { DMU_BSWAP_UINT8, TRUE, "SPA history" },
90 90 { DMU_BSWAP_UINT64, TRUE, "SPA history offsets" },
91 91 { DMU_BSWAP_ZAP, TRUE, "Pool properties" },
92 92 { DMU_BSWAP_ZAP, TRUE, "DSL permissions" },
93 93 { DMU_BSWAP_ACL, TRUE, "ZFS ACL" },
94 94 { DMU_BSWAP_UINT8, TRUE, "ZFS SYSACL" },
95 95 { DMU_BSWAP_UINT8, TRUE, "FUID table" },
96 96 { DMU_BSWAP_UINT64, TRUE, "FUID table size" },
97 97 { DMU_BSWAP_ZAP, TRUE, "DSL dataset next clones"},
98 98 { DMU_BSWAP_ZAP, TRUE, "scan work queue" },
99 99 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group used" },
100 100 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group quota" },
101 101 { DMU_BSWAP_ZAP, TRUE, "snapshot refcount tags"},
102 102 { DMU_BSWAP_ZAP, TRUE, "DDT ZAP algorithm" },
103 103 { DMU_BSWAP_ZAP, TRUE, "DDT statistics" },
104 104 { DMU_BSWAP_UINT8, TRUE, "System attributes" },
105 105 { DMU_BSWAP_ZAP, TRUE, "SA master node" },
106 106 { DMU_BSWAP_ZAP, TRUE, "SA attr registration" },
107 107 { DMU_BSWAP_ZAP, TRUE, "SA attr layouts" },
108 108 { DMU_BSWAP_ZAP, TRUE, "scan translations" },
109 109 { DMU_BSWAP_UINT8, FALSE, "deduplicated block" },
110 110 { DMU_BSWAP_ZAP, TRUE, "DSL deadlist map" },
111 111 { DMU_BSWAP_UINT64, TRUE, "DSL deadlist map hdr" },
112 112 { DMU_BSWAP_ZAP, TRUE, "DSL dir clones" },
113 113 { DMU_BSWAP_UINT64, TRUE, "bpobj subobj" }
114 114 };
115 115
116 116 const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
117 117 { byteswap_uint8_array, "uint8" },
118 118 { byteswap_uint16_array, "uint16" },
119 119 { byteswap_uint32_array, "uint32" },
120 120 { byteswap_uint64_array, "uint64" },
121 121 { zap_byteswap, "zap" },
122 122 { dnode_buf_byteswap, "dnode" },
123 123 { dmu_objset_byteswap, "objset" },
124 124 { zfs_znode_byteswap, "znode" },
125 125 { zfs_oldacl_byteswap, "oldacl" },
126 126 { zfs_acl_byteswap, "acl" }
127 127 };
128 128
129 129 int
130 130 dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset,
131 131 void *tag, dmu_buf_t **dbp)
132 132 {
133 133 dnode_t *dn;
134 134 uint64_t blkid;
135 135 dmu_buf_impl_t *db;
136 136 int err;
137 137
138 138 err = dnode_hold(os, object, FTAG, &dn);
139 139 if (err)
140 140 return (err);
141 141 blkid = dbuf_whichblock(dn, offset);
142 142 rw_enter(&dn->dn_struct_rwlock, RW_READER);
143 143 db = dbuf_hold(dn, blkid, tag);
144 144 rw_exit(&dn->dn_struct_rwlock);
145 145 dnode_rele(dn, FTAG);
146 146
147 147 if (db == NULL) {
148 148 *dbp = NULL;
149 149 return (SET_ERROR(EIO));
150 150 }
151 151
152 152 *dbp = &db->db;
153 153 return (err);
154 154 }
155 155
156 156 int
157 157 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
158 158 void *tag, dmu_buf_t **dbp, int flags)
159 159 {
160 160 int err;
161 161 int db_flags = DB_RF_CANFAIL;
162 162
163 163 if (flags & DMU_READ_NO_PREFETCH)
164 164 db_flags |= DB_RF_NOPREFETCH;
165 165
166 166 err = dmu_buf_hold_noread(os, object, offset, tag, dbp);
167 167 if (err == 0) {
168 168 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp);
169 169 err = dbuf_read(db, NULL, db_flags);
170 170 if (err != 0) {
171 171 dbuf_rele(db, tag);
172 172 *dbp = NULL;
173 173 }
174 174 }
175 175
176 176 return (err);
177 177 }
178 178
179 179 int
180 180 dmu_bonus_max(void)
181 181 {
182 182 return (DN_MAX_BONUSLEN);
183 183 }
184 184
185 185 int
186 186 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
187 187 {
188 188 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
189 189 dnode_t *dn;
190 190 int error;
191 191
192 192 DB_DNODE_ENTER(db);
193 193 dn = DB_DNODE(db);
194 194
195 195 if (dn->dn_bonus != db) {
196 196 error = SET_ERROR(EINVAL);
197 197 } else if (newsize < 0 || newsize > db_fake->db_size) {
198 198 error = SET_ERROR(EINVAL);
199 199 } else {
200 200 dnode_setbonuslen(dn, newsize, tx);
201 201 error = 0;
202 202 }
203 203
204 204 DB_DNODE_EXIT(db);
205 205 return (error);
206 206 }
207 207
208 208 int
209 209 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
210 210 {
211 211 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
212 212 dnode_t *dn;
213 213 int error;
214 214
215 215 DB_DNODE_ENTER(db);
216 216 dn = DB_DNODE(db);
217 217
218 218 if (!DMU_OT_IS_VALID(type)) {
219 219 error = SET_ERROR(EINVAL);
220 220 } else if (dn->dn_bonus != db) {
221 221 error = SET_ERROR(EINVAL);
222 222 } else {
223 223 dnode_setbonus_type(dn, type, tx);
224 224 error = 0;
225 225 }
226 226
227 227 DB_DNODE_EXIT(db);
228 228 return (error);
229 229 }
230 230
231 231 dmu_object_type_t
232 232 dmu_get_bonustype(dmu_buf_t *db_fake)
233 233 {
234 234 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
235 235 dnode_t *dn;
236 236 dmu_object_type_t type;
237 237
238 238 DB_DNODE_ENTER(db);
239 239 dn = DB_DNODE(db);
240 240 type = dn->dn_bonustype;
241 241 DB_DNODE_EXIT(db);
242 242
243 243 return (type);
244 244 }
245 245
246 246 int
247 247 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
248 248 {
249 249 dnode_t *dn;
250 250 int error;
251 251
252 252 error = dnode_hold(os, object, FTAG, &dn);
253 253 dbuf_rm_spill(dn, tx);
254 254 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
255 255 dnode_rm_spill(dn, tx);
256 256 rw_exit(&dn->dn_struct_rwlock);
257 257 dnode_rele(dn, FTAG);
258 258 return (error);
259 259 }
260 260
261 261 /*
262 262 * returns ENOENT, EIO, or 0.
263 263 */
264 264 int
265 265 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
266 266 {
267 267 dnode_t *dn;
268 268 dmu_buf_impl_t *db;
269 269 int error;
270 270
271 271 error = dnode_hold(os, object, FTAG, &dn);
272 272 if (error)
273 273 return (error);
274 274
275 275 rw_enter(&dn->dn_struct_rwlock, RW_READER);
276 276 if (dn->dn_bonus == NULL) {
277 277 rw_exit(&dn->dn_struct_rwlock);
278 278 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
279 279 if (dn->dn_bonus == NULL)
280 280 dbuf_create_bonus(dn);
281 281 }
282 282 db = dn->dn_bonus;
283 283
284 284 /* as long as the bonus buf is held, the dnode will be held */
285 285 if (refcount_add(&db->db_holds, tag) == 1) {
286 286 VERIFY(dnode_add_ref(dn, db));
287 287 atomic_inc_32(&dn->dn_dbufs_count);
288 288 }
289 289
290 290 /*
291 291 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
292 292 * hold and incrementing the dbuf count to ensure that dnode_move() sees
293 293 * a dnode hold for every dbuf.
294 294 */
295 295 rw_exit(&dn->dn_struct_rwlock);
296 296
297 297 dnode_rele(dn, FTAG);
298 298
299 299 VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH));
300 300
301 301 *dbp = &db->db;
302 302 return (0);
303 303 }
304 304
305 305 /*
306 306 * returns ENOENT, EIO, or 0.
307 307 *
308 308 * This interface will allocate a blank spill dbuf when a spill blk
309 309 * doesn't already exist on the dnode.
310 310 *
311 311 * if you only want to find an already existing spill db, then
312 312 * dmu_spill_hold_existing() should be used.
313 313 */
314 314 int
315 315 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
316 316 {
317 317 dmu_buf_impl_t *db = NULL;
318 318 int err;
319 319
320 320 if ((flags & DB_RF_HAVESTRUCT) == 0)
321 321 rw_enter(&dn->dn_struct_rwlock, RW_READER);
322 322
323 323 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
324 324
325 325 if ((flags & DB_RF_HAVESTRUCT) == 0)
326 326 rw_exit(&dn->dn_struct_rwlock);
327 327
328 328 ASSERT(db != NULL);
329 329 err = dbuf_read(db, NULL, flags);
330 330 if (err == 0)
331 331 *dbp = &db->db;
332 332 else
333 333 dbuf_rele(db, tag);
334 334 return (err);
335 335 }
336 336
337 337 int
338 338 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
339 339 {
340 340 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
341 341 dnode_t *dn;
342 342 int err;
343 343
344 344 DB_DNODE_ENTER(db);
345 345 dn = DB_DNODE(db);
346 346
347 347 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
348 348 err = SET_ERROR(EINVAL);
349 349 } else {
350 350 rw_enter(&dn->dn_struct_rwlock, RW_READER);
351 351
352 352 if (!dn->dn_have_spill) {
353 353 err = SET_ERROR(ENOENT);
354 354 } else {
355 355 err = dmu_spill_hold_by_dnode(dn,
356 356 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
357 357 }
358 358
359 359 rw_exit(&dn->dn_struct_rwlock);
360 360 }
361 361
362 362 DB_DNODE_EXIT(db);
363 363 return (err);
364 364 }
365 365
366 366 int
367 367 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
368 368 {
369 369 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
370 370 dnode_t *dn;
371 371 int err;
372 372
373 373 DB_DNODE_ENTER(db);
374 374 dn = DB_DNODE(db);
375 375 err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp);
376 376 DB_DNODE_EXIT(db);
377 377
378 378 return (err);
379 379 }
380 380
381 381 /*
382 382 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
383 383 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
384 384 * and can induce severe lock contention when writing to several files
385 385 * whose dnodes are in the same block.
386 386 */
387 387 static int
388 388 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
389 389 int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
390 390 {
391 391 dmu_buf_t **dbp;
392 392 uint64_t blkid, nblks, i;
393 393 uint32_t dbuf_flags;
394 394 int err;
395 395 zio_t *zio;
396 396
397 397 ASSERT(length <= DMU_MAX_ACCESS);
398 398
399 399 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT;
400 400 if (flags & DMU_READ_NO_PREFETCH || length > zfetch_array_rd_sz)
401 401 dbuf_flags |= DB_RF_NOPREFETCH;
402 402
403 403 rw_enter(&dn->dn_struct_rwlock, RW_READER);
404 404 if (dn->dn_datablkshift) {
405 405 int blkshift = dn->dn_datablkshift;
406 406 nblks = (P2ROUNDUP(offset+length, 1ULL<<blkshift) -
407 407 P2ALIGN(offset, 1ULL<<blkshift)) >> blkshift;
408 408 } else {
409 409 if (offset + length > dn->dn_datablksz) {
410 410 zfs_panic_recover("zfs: accessing past end of object "
411 411 "%llx/%llx (size=%u access=%llu+%llu)",
412 412 (longlong_t)dn->dn_objset->
413 413 os_dsl_dataset->ds_object,
414 414 (longlong_t)dn->dn_object, dn->dn_datablksz,
415 415 (longlong_t)offset, (longlong_t)length);
416 416 rw_exit(&dn->dn_struct_rwlock);
417 417 return (SET_ERROR(EIO));
418 418 }
419 419 nblks = 1;
420 420 }
421 421 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
422 422
423 423 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
424 424 blkid = dbuf_whichblock(dn, offset);
425 425 for (i = 0; i < nblks; i++) {
426 426 dmu_buf_impl_t *db = dbuf_hold(dn, blkid+i, tag);
427 427 if (db == NULL) {
428 428 rw_exit(&dn->dn_struct_rwlock);
429 429 dmu_buf_rele_array(dbp, nblks, tag);
430 430 zio_nowait(zio);
431 431 return (SET_ERROR(EIO));
432 432 }
433 433 /* initiate async i/o */
434 434 if (read) {
435 435 (void) dbuf_read(db, zio, dbuf_flags);
436 436 }
437 437 dbp[i] = &db->db;
438 438 }
439 439 rw_exit(&dn->dn_struct_rwlock);
440 440
441 441 /* wait for async i/o */
442 442 err = zio_wait(zio);
443 443 if (err) {
444 444 dmu_buf_rele_array(dbp, nblks, tag);
445 445 return (err);
446 446 }
447 447
448 448 /* wait for other io to complete */
449 449 if (read) {
450 450 for (i = 0; i < nblks; i++) {
451 451 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
452 452 mutex_enter(&db->db_mtx);
453 453 while (db->db_state == DB_READ ||
454 454 db->db_state == DB_FILL)
455 455 cv_wait(&db->db_changed, &db->db_mtx);
456 456 if (db->db_state == DB_UNCACHED)
457 457 err = SET_ERROR(EIO);
458 458 mutex_exit(&db->db_mtx);
459 459 if (err) {
460 460 dmu_buf_rele_array(dbp, nblks, tag);
461 461 return (err);
462 462 }
463 463 }
464 464 }
465 465
466 466 *numbufsp = nblks;
467 467 *dbpp = dbp;
468 468 return (0);
469 469 }
470 470
471 471 static int
472 472 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
473 473 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
474 474 {
475 475 dnode_t *dn;
476 476 int err;
477 477
478 478 err = dnode_hold(os, object, FTAG, &dn);
479 479 if (err)
480 480 return (err);
481 481
482 482 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
483 483 numbufsp, dbpp, DMU_READ_PREFETCH);
484 484
485 485 dnode_rele(dn, FTAG);
486 486
487 487 return (err);
488 488 }
489 489
490 490 int
491 491 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
492 492 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
493 493 {
494 494 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
495 495 dnode_t *dn;
496 496 int err;
497 497
498 498 DB_DNODE_ENTER(db);
499 499 dn = DB_DNODE(db);
500 500 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
501 501 numbufsp, dbpp, DMU_READ_PREFETCH);
502 502 DB_DNODE_EXIT(db);
503 503
504 504 return (err);
505 505 }
506 506
507 507 void
508 508 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
509 509 {
510 510 int i;
511 511 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
512 512
513 513 if (numbufs == 0)
514 514 return;
515 515
516 516 for (i = 0; i < numbufs; i++) {
517 517 if (dbp[i])
518 518 dbuf_rele(dbp[i], tag);
519 519 }
520 520
521 521 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
522 522 }
523 523
524 524 /*
525 525 * Issue prefetch i/os for the given blocks.
526 526 *
527 527 * Note: The assumption is that we *know* these blocks will be needed
528 528 * almost immediately. Therefore, the prefetch i/os will be issued at
529 529 * ZIO_PRIORITY_SYNC_READ
530 530 *
531 531 * Note: indirect blocks and other metadata will be read synchronously,
532 532 * causing this function to block if they are not already cached.
533 533 */
534 534 void
535 535 dmu_prefetch(objset_t *os, uint64_t object, uint64_t offset, uint64_t len)
536 536 {
537 537 dnode_t *dn;
538 538 uint64_t blkid;
539 539 int nblks, err;
540 540
541 541 if (zfs_prefetch_disable)
542 542 return;
543 543
544 544 if (len == 0) { /* they're interested in the bonus buffer */
545 545 dn = DMU_META_DNODE(os);
546 546
547 547 if (object == 0 || object >= DN_MAX_OBJECT)
548 548 return;
549 549
550 550 rw_enter(&dn->dn_struct_rwlock, RW_READER);
551 551 blkid = dbuf_whichblock(dn, object * sizeof (dnode_phys_t));
552 552 dbuf_prefetch(dn, blkid, ZIO_PRIORITY_SYNC_READ);
553 553 rw_exit(&dn->dn_struct_rwlock);
554 554 return;
555 555 }
556 556
557 557 /*
558 558 * XXX - Note, if the dnode for the requested object is not
559 559 * already cached, we will do a *synchronous* read in the
560 560 * dnode_hold() call. The same is true for any indirects.
561 561 */
562 562 err = dnode_hold(os, object, FTAG, &dn);
563 563 if (err != 0)
564 564 return;
565 565
566 566 rw_enter(&dn->dn_struct_rwlock, RW_READER);
567 567 if (dn->dn_datablkshift) {
568 568 int blkshift = dn->dn_datablkshift;
569 569 nblks = (P2ROUNDUP(offset + len, 1 << blkshift) -
570 570 P2ALIGN(offset, 1 << blkshift)) >> blkshift;
571 571 } else {
572 572 nblks = (offset < dn->dn_datablksz);
573 573 }
574 574
575 575 if (nblks != 0) {
576 576 blkid = dbuf_whichblock(dn, offset);
577 577 for (int i = 0; i < nblks; i++)
578 578 dbuf_prefetch(dn, blkid + i, ZIO_PRIORITY_SYNC_READ);
579 579 }
580 580
581 581 rw_exit(&dn->dn_struct_rwlock);
582 582
583 583 dnode_rele(dn, FTAG);
584 584 }
585 585
586 586 /*
587 587 * Get the next "chunk" of file data to free. We traverse the file from
588 588 * the end so that the file gets shorter over time (if we crashes in the
589 589 * middle, this will leave us in a better state). We find allocated file
590 590 * data by simply searching the allocated level 1 indirects.
591 591 *
592 592 * On input, *start should be the first offset that does not need to be
593 593 * freed (e.g. "offset + length"). On return, *start will be the first
594 594 * offset that should be freed.
595 595 */
596 596 static int
597 597 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum)
598 598 {
599 599 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1);
600 600 /* bytes of data covered by a level-1 indirect block */
601 601 uint64_t iblkrange =
602 602 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
603 603
604 604 ASSERT3U(minimum, <=, *start);
605 605
606 606 if (*start - minimum <= iblkrange * maxblks) {
607 607 *start = minimum;
608 608 return (0);
609 609 }
610 610 ASSERT(ISP2(iblkrange));
611 611
612 612 for (uint64_t blks = 0; *start > minimum && blks < maxblks; blks++) {
613 613 int err;
614 614
615 615 /*
616 616 * dnode_next_offset(BACKWARDS) will find an allocated L1
617 617 * indirect block at or before the input offset. We must
618 618 * decrement *start so that it is at the end of the region
619 619 * to search.
620 620 */
621 621 (*start)--;
622 622 err = dnode_next_offset(dn,
623 623 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
624 624
625 625 /* if there are no indirect blocks before start, we are done */
626 626 if (err == ESRCH) {
627 627 *start = minimum;
628 628 break;
629 629 } else if (err != 0) {
630 630 return (err);
631 631 }
632 632
633 633 /* set start to the beginning of this L1 indirect */
634 634 *start = P2ALIGN(*start, iblkrange);
635 635 }
636 636 if (*start < minimum)
637 637 *start = minimum;
638 638 return (0);
639 639 }
640 640
641 641 static int
642 642 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
643 643 uint64_t length)
644 644 {
645 645 uint64_t object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
646 646 int err;
647 647
648 648 if (offset >= object_size)
649 649 return (0);
650 650
651 651 if (length == DMU_OBJECT_END || offset + length > object_size)
652 652 length = object_size - offset;
653 653
654 654 while (length != 0) {
655 655 uint64_t chunk_end, chunk_begin;
656 656
657 657 chunk_end = chunk_begin = offset + length;
658 658
659 659 /* move chunk_begin backwards to the beginning of this chunk */
660 660 err = get_next_chunk(dn, &chunk_begin, offset);
661 661 if (err)
662 662 return (err);
663 663 ASSERT3U(chunk_begin, >=, offset);
664 664 ASSERT3U(chunk_begin, <=, chunk_end);
665 665
666 666 dmu_tx_t *tx = dmu_tx_create(os);
667 667 dmu_tx_hold_free(tx, dn->dn_object,
668 668 chunk_begin, chunk_end - chunk_begin);
669 669
670 670 /*
671 671 * Mark this transaction as typically resulting in a net
672 672 * reduction in space used.
673 673 */
674 674 dmu_tx_mark_netfree(tx);
675 675 err = dmu_tx_assign(tx, TXG_WAIT);
676 676 if (err) {
677 677 dmu_tx_abort(tx);
678 678 return (err);
679 679 }
680 680 dnode_free_range(dn, chunk_begin, chunk_end - chunk_begin, tx);
681 681 dmu_tx_commit(tx);
682 682
683 683 length -= chunk_end - chunk_begin;
684 684 }
685 685 return (0);
686 686 }
687 687
688 688 int
689 689 dmu_free_long_range(objset_t *os, uint64_t object,
690 690 uint64_t offset, uint64_t length)
691 691 {
692 692 dnode_t *dn;
693 693 int err;
694 694
695 695 err = dnode_hold(os, object, FTAG, &dn);
696 696 if (err != 0)
697 697 return (err);
698 698 err = dmu_free_long_range_impl(os, dn, offset, length);
699 699
700 700 /*
701 701 * It is important to zero out the maxblkid when freeing the entire
702 702 * file, so that (a) subsequent calls to dmu_free_long_range_impl()
703 703 * will take the fast path, and (b) dnode_reallocate() can verify
704 704 * that the entire file has been freed.
705 705 */
706 706 if (err == 0 && offset == 0 && length == DMU_OBJECT_END)
707 707 dn->dn_maxblkid = 0;
708 708
709 709 dnode_rele(dn, FTAG);
710 710 return (err);
711 711 }
712 712
713 713 int
714 714 dmu_free_long_object(objset_t *os, uint64_t object)
715 715 {
716 716 dmu_tx_t *tx;
717 717 int err;
718 718
719 719 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END);
720 720 if (err != 0)
721 721 return (err);
722 722
723 723 tx = dmu_tx_create(os);
724 724 dmu_tx_hold_bonus(tx, object);
725 725 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END);
726 726 dmu_tx_mark_netfree(tx);
727 727 err = dmu_tx_assign(tx, TXG_WAIT);
728 728 if (err == 0) {
729 729 err = dmu_object_free(os, object, tx);
730 730 dmu_tx_commit(tx);
731 731 } else {
732 732 dmu_tx_abort(tx);
733 733 }
734 734
735 735 return (err);
736 736 }
737 737
738 738 int
739 739 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
740 740 uint64_t size, dmu_tx_t *tx)
741 741 {
742 742 dnode_t *dn;
743 743 int err = dnode_hold(os, object, FTAG, &dn);
744 744 if (err)
745 745 return (err);
746 746 ASSERT(offset < UINT64_MAX);
747 747 ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
748 748 dnode_free_range(dn, offset, size, tx);
749 749 dnode_rele(dn, FTAG);
750 750 return (0);
751 751 }
752 752
753 753 int
754 754 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
755 755 void *buf, uint32_t flags)
756 756 {
757 757 dnode_t *dn;
758 758 dmu_buf_t **dbp;
759 759 int numbufs, err;
760 760
761 761 err = dnode_hold(os, object, FTAG, &dn);
762 762 if (err)
763 763 return (err);
764 764
765 765 /*
766 766 * Deal with odd block sizes, where there can't be data past the first
767 767 * block. If we ever do the tail block optimization, we will need to
768 768 * handle that here as well.
769 769 */
770 770 if (dn->dn_maxblkid == 0) {
771 771 int newsz = offset > dn->dn_datablksz ? 0 :
772 772 MIN(size, dn->dn_datablksz - offset);
773 773 bzero((char *)buf + newsz, size - newsz);
774 774 size = newsz;
775 775 }
776 776
777 777 while (size > 0) {
778 778 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
779 779 int i;
780 780
781 781 /*
782 782 * NB: we could do this block-at-a-time, but it's nice
783 783 * to be reading in parallel.
784 784 */
785 785 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
786 786 TRUE, FTAG, &numbufs, &dbp, flags);
787 787 if (err)
788 788 break;
789 789
790 790 for (i = 0; i < numbufs; i++) {
791 791 int tocpy;
792 792 int bufoff;
793 793 dmu_buf_t *db = dbp[i];
794 794
795 795 ASSERT(size > 0);
796 796
797 797 bufoff = offset - db->db_offset;
798 798 tocpy = (int)MIN(db->db_size - bufoff, size);
799 799
800 800 bcopy((char *)db->db_data + bufoff, buf, tocpy);
801 801
802 802 offset += tocpy;
|
↓ open down ↓ |
802 lines elided |
↑ open up ↑ |
803 803 size -= tocpy;
804 804 buf = (char *)buf + tocpy;
805 805 }
806 806 dmu_buf_rele_array(dbp, numbufs, FTAG);
807 807 }
808 808 dnode_rele(dn, FTAG);
809 809 return (err);
810 810 }
811 811
812 812 void
813 +dmu_write_zero(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, dmu_tx_t *tx)
814 +{
815 + dmu_buf_t **dbp;
816 + int numbufs, i;
817 +
818 + VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
819 + FALSE, FTAG, &numbufs, &dbp));
820 +
821 + for (i = 0; i < numbufs; i++) {
822 + dmu_buf_t *db = dbp[i];
823 +
824 + dmu_buf_will_zero_fill(db, tx);
825 +
826 + memset(db->db_data, 0, db->db_size);
827 +
828 + dmu_buf_fill_done(db, tx);
829 + }
830 +
831 + dmu_buf_rele_array(dbp, numbufs, FTAG);
832 +}
833 +
834 +void
813 835 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
814 836 const void *buf, dmu_tx_t *tx)
815 837 {
816 838 dmu_buf_t **dbp;
817 839 int numbufs, i;
818 840
819 841 if (size == 0)
820 842 return;
821 843
822 844 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
823 845 FALSE, FTAG, &numbufs, &dbp));
824 846
825 847 for (i = 0; i < numbufs; i++) {
826 848 int tocpy;
827 849 int bufoff;
828 850 dmu_buf_t *db = dbp[i];
829 851
830 852 ASSERT(size > 0);
831 853
832 854 bufoff = offset - db->db_offset;
833 855 tocpy = (int)MIN(db->db_size - bufoff, size);
834 856
835 857 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
836 858
837 859 if (tocpy == db->db_size)
838 860 dmu_buf_will_fill(db, tx);
839 861 else
840 862 dmu_buf_will_dirty(db, tx);
841 863
842 864 bcopy(buf, (char *)db->db_data + bufoff, tocpy);
843 865
844 866 if (tocpy == db->db_size)
845 867 dmu_buf_fill_done(db, tx);
846 868
847 869 offset += tocpy;
848 870 size -= tocpy;
849 871 buf = (char *)buf + tocpy;
850 872 }
851 873 dmu_buf_rele_array(dbp, numbufs, FTAG);
852 874 }
853 875
854 876 void
855 877 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
856 878 dmu_tx_t *tx)
857 879 {
858 880 dmu_buf_t **dbp;
859 881 int numbufs, i;
860 882
861 883 if (size == 0)
862 884 return;
863 885
864 886 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
865 887 FALSE, FTAG, &numbufs, &dbp));
866 888
867 889 for (i = 0; i < numbufs; i++) {
868 890 dmu_buf_t *db = dbp[i];
869 891
870 892 dmu_buf_will_not_fill(db, tx);
871 893 }
872 894 dmu_buf_rele_array(dbp, numbufs, FTAG);
873 895 }
874 896
875 897 void
876 898 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset,
877 899 void *data, uint8_t etype, uint8_t comp, int uncompressed_size,
878 900 int compressed_size, int byteorder, dmu_tx_t *tx)
879 901 {
880 902 dmu_buf_t *db;
881 903
882 904 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES);
883 905 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS);
884 906 VERIFY0(dmu_buf_hold_noread(os, object, offset,
885 907 FTAG, &db));
886 908
887 909 dmu_buf_write_embedded(db,
888 910 data, (bp_embedded_type_t)etype, (enum zio_compress)comp,
889 911 uncompressed_size, compressed_size, byteorder, tx);
890 912
891 913 dmu_buf_rele(db, FTAG);
892 914 }
893 915
894 916 /*
895 917 * DMU support for xuio
896 918 */
897 919 kstat_t *xuio_ksp = NULL;
898 920
899 921 int
900 922 dmu_xuio_init(xuio_t *xuio, int nblk)
901 923 {
902 924 dmu_xuio_t *priv;
903 925 uio_t *uio = &xuio->xu_uio;
904 926
905 927 uio->uio_iovcnt = nblk;
906 928 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
907 929
908 930 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
909 931 priv->cnt = nblk;
910 932 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
911 933 priv->iovp = uio->uio_iov;
912 934 XUIO_XUZC_PRIV(xuio) = priv;
913 935
914 936 if (XUIO_XUZC_RW(xuio) == UIO_READ)
915 937 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
916 938 else
917 939 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
918 940
919 941 return (0);
920 942 }
921 943
922 944 void
923 945 dmu_xuio_fini(xuio_t *xuio)
924 946 {
925 947 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
926 948 int nblk = priv->cnt;
927 949
928 950 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
929 951 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
930 952 kmem_free(priv, sizeof (dmu_xuio_t));
931 953
932 954 if (XUIO_XUZC_RW(xuio) == UIO_READ)
933 955 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
934 956 else
935 957 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
936 958 }
937 959
938 960 /*
939 961 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
940 962 * and increase priv->next by 1.
941 963 */
942 964 int
943 965 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
944 966 {
945 967 struct iovec *iov;
946 968 uio_t *uio = &xuio->xu_uio;
947 969 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
948 970 int i = priv->next++;
949 971
950 972 ASSERT(i < priv->cnt);
951 973 ASSERT(off + n <= arc_buf_size(abuf));
952 974 iov = uio->uio_iov + i;
953 975 iov->iov_base = (char *)abuf->b_data + off;
954 976 iov->iov_len = n;
955 977 priv->bufs[i] = abuf;
956 978 return (0);
957 979 }
958 980
959 981 int
960 982 dmu_xuio_cnt(xuio_t *xuio)
961 983 {
962 984 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
963 985 return (priv->cnt);
964 986 }
965 987
966 988 arc_buf_t *
967 989 dmu_xuio_arcbuf(xuio_t *xuio, int i)
968 990 {
969 991 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
970 992
971 993 ASSERT(i < priv->cnt);
972 994 return (priv->bufs[i]);
973 995 }
974 996
975 997 void
976 998 dmu_xuio_clear(xuio_t *xuio, int i)
977 999 {
978 1000 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
979 1001
980 1002 ASSERT(i < priv->cnt);
981 1003 priv->bufs[i] = NULL;
982 1004 }
983 1005
984 1006 static void
985 1007 xuio_stat_init(void)
986 1008 {
987 1009 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
988 1010 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
989 1011 KSTAT_FLAG_VIRTUAL);
990 1012 if (xuio_ksp != NULL) {
991 1013 xuio_ksp->ks_data = &xuio_stats;
992 1014 kstat_install(xuio_ksp);
993 1015 }
994 1016 }
995 1017
996 1018 static void
997 1019 xuio_stat_fini(void)
998 1020 {
999 1021 if (xuio_ksp != NULL) {
1000 1022 kstat_delete(xuio_ksp);
1001 1023 xuio_ksp = NULL;
1002 1024 }
1003 1025 }
1004 1026
1005 1027 void
1006 1028 xuio_stat_wbuf_copied()
1007 1029 {
1008 1030 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1009 1031 }
1010 1032
1011 1033 void
1012 1034 xuio_stat_wbuf_nocopy()
1013 1035 {
1014 1036 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
1015 1037 }
1016 1038
1017 1039 #ifdef _KERNEL
1018 1040 static int
1019 1041 dmu_read_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size)
1020 1042 {
1021 1043 dmu_buf_t **dbp;
1022 1044 int numbufs, i, err;
1023 1045 xuio_t *xuio = NULL;
1024 1046
1025 1047 /*
1026 1048 * NB: we could do this block-at-a-time, but it's nice
1027 1049 * to be reading in parallel.
1028 1050 */
1029 1051 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1030 1052 TRUE, FTAG, &numbufs, &dbp, 0);
1031 1053 if (err)
1032 1054 return (err);
1033 1055
1034 1056 if (uio->uio_extflg == UIO_XUIO)
1035 1057 xuio = (xuio_t *)uio;
1036 1058
1037 1059 for (i = 0; i < numbufs; i++) {
1038 1060 int tocpy;
1039 1061 int bufoff;
1040 1062 dmu_buf_t *db = dbp[i];
1041 1063
1042 1064 ASSERT(size > 0);
1043 1065
1044 1066 bufoff = uio->uio_loffset - db->db_offset;
1045 1067 tocpy = (int)MIN(db->db_size - bufoff, size);
1046 1068
1047 1069 if (xuio) {
1048 1070 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
1049 1071 arc_buf_t *dbuf_abuf = dbi->db_buf;
1050 1072 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
1051 1073 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
1052 1074 if (!err) {
1053 1075 uio->uio_resid -= tocpy;
1054 1076 uio->uio_loffset += tocpy;
1055 1077 }
1056 1078
1057 1079 if (abuf == dbuf_abuf)
1058 1080 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
1059 1081 else
1060 1082 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
1061 1083 } else {
1062 1084 err = uiomove((char *)db->db_data + bufoff, tocpy,
1063 1085 UIO_READ, uio);
1064 1086 }
1065 1087 if (err)
1066 1088 break;
1067 1089
1068 1090 size -= tocpy;
1069 1091 }
1070 1092 dmu_buf_rele_array(dbp, numbufs, FTAG);
1071 1093
1072 1094 return (err);
1073 1095 }
1074 1096
1075 1097 /*
1076 1098 * Read 'size' bytes into the uio buffer.
1077 1099 * From object zdb->db_object.
1078 1100 * Starting at offset uio->uio_loffset.
1079 1101 *
1080 1102 * If the caller already has a dbuf in the target object
1081 1103 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(),
1082 1104 * because we don't have to find the dnode_t for the object.
1083 1105 */
1084 1106 int
1085 1107 dmu_read_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size)
1086 1108 {
1087 1109 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1088 1110 dnode_t *dn;
1089 1111 int err;
1090 1112
1091 1113 if (size == 0)
1092 1114 return (0);
1093 1115
1094 1116 DB_DNODE_ENTER(db);
1095 1117 dn = DB_DNODE(db);
1096 1118 err = dmu_read_uio_dnode(dn, uio, size);
1097 1119 DB_DNODE_EXIT(db);
1098 1120
1099 1121 return (err);
1100 1122 }
1101 1123
1102 1124 /*
1103 1125 * Read 'size' bytes into the uio buffer.
1104 1126 * From the specified object
1105 1127 * Starting at offset uio->uio_loffset.
1106 1128 */
1107 1129 int
1108 1130 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
1109 1131 {
1110 1132 dnode_t *dn;
1111 1133 int err;
1112 1134
1113 1135 if (size == 0)
1114 1136 return (0);
1115 1137
1116 1138 err = dnode_hold(os, object, FTAG, &dn);
1117 1139 if (err)
1118 1140 return (err);
1119 1141
1120 1142 err = dmu_read_uio_dnode(dn, uio, size);
1121 1143
1122 1144 dnode_rele(dn, FTAG);
1123 1145
1124 1146 return (err);
1125 1147 }
1126 1148
1127 1149 static int
1128 1150 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1129 1151 {
1130 1152 dmu_buf_t **dbp;
1131 1153 int numbufs;
1132 1154 int err = 0;
1133 1155 int i;
1134 1156
1135 1157 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1136 1158 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1137 1159 if (err)
1138 1160 return (err);
1139 1161
1140 1162 for (i = 0; i < numbufs; i++) {
1141 1163 int tocpy;
1142 1164 int bufoff;
1143 1165 dmu_buf_t *db = dbp[i];
1144 1166
1145 1167 ASSERT(size > 0);
1146 1168
1147 1169 bufoff = uio->uio_loffset - db->db_offset;
1148 1170 tocpy = (int)MIN(db->db_size - bufoff, size);
1149 1171
1150 1172 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1151 1173
1152 1174 if (tocpy == db->db_size)
1153 1175 dmu_buf_will_fill(db, tx);
1154 1176 else
1155 1177 dmu_buf_will_dirty(db, tx);
1156 1178
1157 1179 /*
1158 1180 * XXX uiomove could block forever (eg. nfs-backed
1159 1181 * pages). There needs to be a uiolockdown() function
1160 1182 * to lock the pages in memory, so that uiomove won't
1161 1183 * block.
1162 1184 */
1163 1185 err = uiomove((char *)db->db_data + bufoff, tocpy,
1164 1186 UIO_WRITE, uio);
1165 1187
1166 1188 if (tocpy == db->db_size)
1167 1189 dmu_buf_fill_done(db, tx);
1168 1190
1169 1191 if (err)
1170 1192 break;
1171 1193
1172 1194 size -= tocpy;
1173 1195 }
1174 1196
1175 1197 dmu_buf_rele_array(dbp, numbufs, FTAG);
1176 1198 return (err);
1177 1199 }
1178 1200
1179 1201 /*
1180 1202 * Write 'size' bytes from the uio buffer.
1181 1203 * To object zdb->db_object.
1182 1204 * Starting at offset uio->uio_loffset.
1183 1205 *
1184 1206 * If the caller already has a dbuf in the target object
1185 1207 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(),
1186 1208 * because we don't have to find the dnode_t for the object.
1187 1209 */
1188 1210 int
1189 1211 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1190 1212 dmu_tx_t *tx)
1191 1213 {
1192 1214 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1193 1215 dnode_t *dn;
1194 1216 int err;
1195 1217
1196 1218 if (size == 0)
1197 1219 return (0);
1198 1220
1199 1221 DB_DNODE_ENTER(db);
1200 1222 dn = DB_DNODE(db);
1201 1223 err = dmu_write_uio_dnode(dn, uio, size, tx);
1202 1224 DB_DNODE_EXIT(db);
1203 1225
1204 1226 return (err);
1205 1227 }
1206 1228
1207 1229 /*
1208 1230 * Write 'size' bytes from the uio buffer.
1209 1231 * To the specified object.
1210 1232 * Starting at offset uio->uio_loffset.
1211 1233 */
1212 1234 int
1213 1235 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1214 1236 dmu_tx_t *tx)
1215 1237 {
1216 1238 dnode_t *dn;
1217 1239 int err;
1218 1240
1219 1241 if (size == 0)
1220 1242 return (0);
1221 1243
1222 1244 err = dnode_hold(os, object, FTAG, &dn);
1223 1245 if (err)
1224 1246 return (err);
1225 1247
1226 1248 err = dmu_write_uio_dnode(dn, uio, size, tx);
1227 1249
1228 1250 dnode_rele(dn, FTAG);
1229 1251
1230 1252 return (err);
1231 1253 }
1232 1254
1233 1255 int
1234 1256 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1235 1257 page_t *pp, dmu_tx_t *tx)
1236 1258 {
1237 1259 dmu_buf_t **dbp;
1238 1260 int numbufs, i;
1239 1261 int err;
1240 1262
1241 1263 if (size == 0)
1242 1264 return (0);
1243 1265
1244 1266 err = dmu_buf_hold_array(os, object, offset, size,
1245 1267 FALSE, FTAG, &numbufs, &dbp);
1246 1268 if (err)
1247 1269 return (err);
1248 1270
1249 1271 for (i = 0; i < numbufs; i++) {
1250 1272 int tocpy, copied, thiscpy;
1251 1273 int bufoff;
1252 1274 dmu_buf_t *db = dbp[i];
1253 1275 caddr_t va;
1254 1276
1255 1277 ASSERT(size > 0);
1256 1278 ASSERT3U(db->db_size, >=, PAGESIZE);
1257 1279
1258 1280 bufoff = offset - db->db_offset;
1259 1281 tocpy = (int)MIN(db->db_size - bufoff, size);
1260 1282
1261 1283 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1262 1284
1263 1285 if (tocpy == db->db_size)
1264 1286 dmu_buf_will_fill(db, tx);
1265 1287 else
1266 1288 dmu_buf_will_dirty(db, tx);
1267 1289
1268 1290 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1269 1291 ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff);
1270 1292 thiscpy = MIN(PAGESIZE, tocpy - copied);
1271 1293 va = zfs_map_page(pp, S_READ);
1272 1294 bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1273 1295 zfs_unmap_page(pp, va);
1274 1296 pp = pp->p_next;
1275 1297 bufoff += PAGESIZE;
1276 1298 }
1277 1299
1278 1300 if (tocpy == db->db_size)
1279 1301 dmu_buf_fill_done(db, tx);
1280 1302
1281 1303 offset += tocpy;
1282 1304 size -= tocpy;
1283 1305 }
1284 1306 dmu_buf_rele_array(dbp, numbufs, FTAG);
1285 1307 return (err);
1286 1308 }
1287 1309 #endif
1288 1310
1289 1311 /*
1290 1312 * Allocate a loaned anonymous arc buffer.
1291 1313 */
1292 1314 arc_buf_t *
1293 1315 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1294 1316 {
1295 1317 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1296 1318
1297 1319 return (arc_loan_buf(db->db_objset->os_spa, size));
1298 1320 }
1299 1321
1300 1322 /*
1301 1323 * Free a loaned arc buffer.
1302 1324 */
1303 1325 void
1304 1326 dmu_return_arcbuf(arc_buf_t *buf)
1305 1327 {
1306 1328 arc_return_buf(buf, FTAG);
1307 1329 VERIFY(arc_buf_remove_ref(buf, FTAG));
1308 1330 }
1309 1331
1310 1332 /*
1311 1333 * When possible directly assign passed loaned arc buffer to a dbuf.
1312 1334 * If this is not possible copy the contents of passed arc buf via
1313 1335 * dmu_write().
1314 1336 */
1315 1337 void
1316 1338 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1317 1339 dmu_tx_t *tx)
1318 1340 {
1319 1341 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1320 1342 dnode_t *dn;
1321 1343 dmu_buf_impl_t *db;
1322 1344 uint32_t blksz = (uint32_t)arc_buf_size(buf);
1323 1345 uint64_t blkid;
1324 1346
1325 1347 DB_DNODE_ENTER(dbuf);
1326 1348 dn = DB_DNODE(dbuf);
1327 1349 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1328 1350 blkid = dbuf_whichblock(dn, offset);
1329 1351 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1330 1352 rw_exit(&dn->dn_struct_rwlock);
1331 1353 DB_DNODE_EXIT(dbuf);
1332 1354
1333 1355 /*
1334 1356 * We can only assign if the offset is aligned, the arc buf is the
1335 1357 * same size as the dbuf, and the dbuf is not metadata. It
1336 1358 * can't be metadata because the loaned arc buf comes from the
1337 1359 * user-data kmem arena.
1338 1360 */
1339 1361 if (offset == db->db.db_offset && blksz == db->db.db_size &&
1340 1362 DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA) {
1341 1363 dbuf_assign_arcbuf(db, buf, tx);
1342 1364 dbuf_rele(db, FTAG);
1343 1365 } else {
1344 1366 objset_t *os;
1345 1367 uint64_t object;
1346 1368
1347 1369 DB_DNODE_ENTER(dbuf);
1348 1370 dn = DB_DNODE(dbuf);
1349 1371 os = dn->dn_objset;
1350 1372 object = dn->dn_object;
1351 1373 DB_DNODE_EXIT(dbuf);
1352 1374
1353 1375 dbuf_rele(db, FTAG);
1354 1376 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1355 1377 dmu_return_arcbuf(buf);
1356 1378 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1357 1379 }
1358 1380 }
1359 1381
1360 1382 typedef struct {
1361 1383 dbuf_dirty_record_t *dsa_dr;
1362 1384 dmu_sync_cb_t *dsa_done;
1363 1385 zgd_t *dsa_zgd;
1364 1386 dmu_tx_t *dsa_tx;
1365 1387 } dmu_sync_arg_t;
1366 1388
1367 1389 /* ARGSUSED */
1368 1390 static void
1369 1391 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1370 1392 {
1371 1393 dmu_sync_arg_t *dsa = varg;
1372 1394 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1373 1395 blkptr_t *bp = zio->io_bp;
1374 1396
1375 1397 if (zio->io_error == 0) {
1376 1398 if (BP_IS_HOLE(bp)) {
1377 1399 /*
1378 1400 * A block of zeros may compress to a hole, but the
1379 1401 * block size still needs to be known for replay.
1380 1402 */
1381 1403 BP_SET_LSIZE(bp, db->db_size);
1382 1404 } else if (!BP_IS_EMBEDDED(bp)) {
1383 1405 ASSERT(BP_GET_LEVEL(bp) == 0);
1384 1406 bp->blk_fill = 1;
1385 1407 }
1386 1408 }
1387 1409 }
1388 1410
1389 1411 static void
1390 1412 dmu_sync_late_arrival_ready(zio_t *zio)
1391 1413 {
1392 1414 dmu_sync_ready(zio, NULL, zio->io_private);
1393 1415 }
1394 1416
1395 1417 /* ARGSUSED */
1396 1418 static void
1397 1419 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1398 1420 {
1399 1421 dmu_sync_arg_t *dsa = varg;
1400 1422 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1401 1423 dmu_buf_impl_t *db = dr->dr_dbuf;
1402 1424
1403 1425 mutex_enter(&db->db_mtx);
1404 1426 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1405 1427 if (zio->io_error == 0) {
1406 1428 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1407 1429 if (dr->dt.dl.dr_nopwrite) {
1408 1430 blkptr_t *bp = zio->io_bp;
1409 1431 blkptr_t *bp_orig = &zio->io_bp_orig;
1410 1432 uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1411 1433
1412 1434 ASSERT(BP_EQUAL(bp, bp_orig));
1413 1435 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1414 1436 ASSERT(zio_checksum_table[chksum].ci_dedup);
1415 1437 }
1416 1438 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1417 1439 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1418 1440 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1419 1441 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by))
1420 1442 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1421 1443 } else {
1422 1444 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1423 1445 }
1424 1446 cv_broadcast(&db->db_changed);
1425 1447 mutex_exit(&db->db_mtx);
1426 1448
1427 1449 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1428 1450
1429 1451 kmem_free(dsa, sizeof (*dsa));
1430 1452 }
1431 1453
1432 1454 static void
1433 1455 dmu_sync_late_arrival_done(zio_t *zio)
1434 1456 {
1435 1457 blkptr_t *bp = zio->io_bp;
1436 1458 dmu_sync_arg_t *dsa = zio->io_private;
1437 1459 blkptr_t *bp_orig = &zio->io_bp_orig;
1438 1460
1439 1461 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1440 1462 /*
1441 1463 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE)
1442 1464 * then there is nothing to do here. Otherwise, free the
1443 1465 * newly allocated block in this txg.
1444 1466 */
1445 1467 if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
1446 1468 ASSERT(BP_EQUAL(bp, bp_orig));
1447 1469 } else {
1448 1470 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1449 1471 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1450 1472 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1451 1473 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1452 1474 }
1453 1475 }
1454 1476
1455 1477 dmu_tx_commit(dsa->dsa_tx);
1456 1478
1457 1479 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1458 1480
1459 1481 kmem_free(dsa, sizeof (*dsa));
1460 1482 }
1461 1483
1462 1484 static int
1463 1485 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1464 1486 zio_prop_t *zp, zbookmark_phys_t *zb)
1465 1487 {
1466 1488 dmu_sync_arg_t *dsa;
1467 1489 dmu_tx_t *tx;
1468 1490
1469 1491 tx = dmu_tx_create(os);
1470 1492 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1471 1493 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1472 1494 dmu_tx_abort(tx);
1473 1495 /* Make zl_get_data do txg_waited_synced() */
1474 1496 return (SET_ERROR(EIO));
1475 1497 }
1476 1498
1477 1499 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1478 1500 dsa->dsa_dr = NULL;
1479 1501 dsa->dsa_done = done;
1480 1502 dsa->dsa_zgd = zgd;
1481 1503 dsa->dsa_tx = tx;
1482 1504
1483 1505 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1484 1506 zgd->zgd_db->db_data, zgd->zgd_db->db_size, zp,
1485 1507 dmu_sync_late_arrival_ready, NULL, dmu_sync_late_arrival_done, dsa,
1486 1508 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1487 1509
1488 1510 return (0);
1489 1511 }
1490 1512
1491 1513 /*
1492 1514 * Intent log support: sync the block associated with db to disk.
1493 1515 * N.B. and XXX: the caller is responsible for making sure that the
1494 1516 * data isn't changing while dmu_sync() is writing it.
1495 1517 *
1496 1518 * Return values:
1497 1519 *
1498 1520 * EEXIST: this txg has already been synced, so there's nothing to do.
1499 1521 * The caller should not log the write.
1500 1522 *
1501 1523 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1502 1524 * The caller should not log the write.
1503 1525 *
1504 1526 * EALREADY: this block is already in the process of being synced.
1505 1527 * The caller should track its progress (somehow).
1506 1528 *
1507 1529 * EIO: could not do the I/O.
1508 1530 * The caller should do a txg_wait_synced().
1509 1531 *
1510 1532 * 0: the I/O has been initiated.
1511 1533 * The caller should log this blkptr in the done callback.
1512 1534 * It is possible that the I/O will fail, in which case
1513 1535 * the error will be reported to the done callback and
1514 1536 * propagated to pio from zio_done().
1515 1537 */
1516 1538 int
1517 1539 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1518 1540 {
1519 1541 blkptr_t *bp = zgd->zgd_bp;
1520 1542 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1521 1543 objset_t *os = db->db_objset;
1522 1544 dsl_dataset_t *ds = os->os_dsl_dataset;
1523 1545 dbuf_dirty_record_t *dr;
1524 1546 dmu_sync_arg_t *dsa;
1525 1547 zbookmark_phys_t zb;
1526 1548 zio_prop_t zp;
1527 1549 dnode_t *dn;
1528 1550
1529 1551 ASSERT(pio != NULL);
1530 1552 ASSERT(txg != 0);
1531 1553
1532 1554 SET_BOOKMARK(&zb, ds->ds_object,
1533 1555 db->db.db_object, db->db_level, db->db_blkid);
1534 1556
1535 1557 DB_DNODE_ENTER(db);
1536 1558 dn = DB_DNODE(db);
1537 1559 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1538 1560 DB_DNODE_EXIT(db);
1539 1561
1540 1562 /*
1541 1563 * If we're frozen (running ziltest), we always need to generate a bp.
1542 1564 */
1543 1565 if (txg > spa_freeze_txg(os->os_spa))
1544 1566 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1545 1567
1546 1568 /*
1547 1569 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1548 1570 * and us. If we determine that this txg is not yet syncing,
1549 1571 * but it begins to sync a moment later, that's OK because the
1550 1572 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1551 1573 */
1552 1574 mutex_enter(&db->db_mtx);
1553 1575
1554 1576 if (txg <= spa_last_synced_txg(os->os_spa)) {
1555 1577 /*
1556 1578 * This txg has already synced. There's nothing to do.
1557 1579 */
1558 1580 mutex_exit(&db->db_mtx);
1559 1581 return (SET_ERROR(EEXIST));
1560 1582 }
1561 1583
1562 1584 if (txg <= spa_syncing_txg(os->os_spa)) {
1563 1585 /*
1564 1586 * This txg is currently syncing, so we can't mess with
1565 1587 * the dirty record anymore; just write a new log block.
1566 1588 */
1567 1589 mutex_exit(&db->db_mtx);
1568 1590 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1569 1591 }
1570 1592
1571 1593 dr = db->db_last_dirty;
1572 1594 while (dr && dr->dr_txg != txg)
1573 1595 dr = dr->dr_next;
1574 1596
1575 1597 if (dr == NULL) {
1576 1598 /*
1577 1599 * There's no dr for this dbuf, so it must have been freed.
1578 1600 * There's no need to log writes to freed blocks, so we're done.
1579 1601 */
1580 1602 mutex_exit(&db->db_mtx);
1581 1603 return (SET_ERROR(ENOENT));
1582 1604 }
1583 1605
1584 1606 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
1585 1607
1586 1608 /*
1587 1609 * Assume the on-disk data is X, the current syncing data is Y,
1588 1610 * and the current in-memory data is Z (currently in dmu_sync).
1589 1611 * X and Z are identical but Y is has been modified. Normally,
1590 1612 * when X and Z are the same we will perform a nopwrite but if Y
1591 1613 * is different we must disable nopwrite since the resulting write
1592 1614 * of Y to disk can free the block containing X. If we allowed a
1593 1615 * nopwrite to occur the block pointing to Z would reference a freed
1594 1616 * block. Since this is a rare case we simplify this by disabling
1595 1617 * nopwrite if the current dmu_sync-ing dbuf has been modified in
1596 1618 * a previous transaction.
1597 1619 */
1598 1620 if (dr->dr_next)
1599 1621 zp.zp_nopwrite = B_FALSE;
1600 1622
1601 1623 ASSERT(dr->dr_txg == txg);
1602 1624 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1603 1625 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1604 1626 /*
1605 1627 * We have already issued a sync write for this buffer,
1606 1628 * or this buffer has already been synced. It could not
1607 1629 * have been dirtied since, or we would have cleared the state.
1608 1630 */
1609 1631 mutex_exit(&db->db_mtx);
1610 1632 return (SET_ERROR(EALREADY));
1611 1633 }
1612 1634
1613 1635 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1614 1636 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1615 1637 mutex_exit(&db->db_mtx);
1616 1638
1617 1639 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1618 1640 dsa->dsa_dr = dr;
1619 1641 dsa->dsa_done = done;
1620 1642 dsa->dsa_zgd = zgd;
1621 1643 dsa->dsa_tx = NULL;
1622 1644
1623 1645 zio_nowait(arc_write(pio, os->os_spa, txg,
1624 1646 bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db),
1625 1647 DBUF_IS_L2COMPRESSIBLE(db), &zp, dmu_sync_ready,
1626 1648 NULL, dmu_sync_done, dsa, ZIO_PRIORITY_SYNC_WRITE,
1627 1649 ZIO_FLAG_CANFAIL, &zb));
1628 1650
1629 1651 return (0);
1630 1652 }
1631 1653
1632 1654 int
1633 1655 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1634 1656 dmu_tx_t *tx)
1635 1657 {
1636 1658 dnode_t *dn;
1637 1659 int err;
1638 1660
1639 1661 err = dnode_hold(os, object, FTAG, &dn);
1640 1662 if (err)
1641 1663 return (err);
1642 1664 err = dnode_set_blksz(dn, size, ibs, tx);
1643 1665 dnode_rele(dn, FTAG);
1644 1666 return (err);
1645 1667 }
1646 1668
1647 1669 void
1648 1670 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1649 1671 dmu_tx_t *tx)
1650 1672 {
1651 1673 dnode_t *dn;
1652 1674
1653 1675 /*
1654 1676 * Send streams include each object's checksum function. This
1655 1677 * check ensures that the receiving system can understand the
1656 1678 * checksum function transmitted.
1657 1679 */
1658 1680 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS);
1659 1681
1660 1682 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1661 1683 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS);
1662 1684 dn->dn_checksum = checksum;
1663 1685 dnode_setdirty(dn, tx);
1664 1686 dnode_rele(dn, FTAG);
1665 1687 }
1666 1688
1667 1689 void
1668 1690 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1669 1691 dmu_tx_t *tx)
1670 1692 {
1671 1693 dnode_t *dn;
1672 1694
1673 1695 /*
1674 1696 * Send streams include each object's compression function. This
1675 1697 * check ensures that the receiving system can understand the
1676 1698 * compression function transmitted.
1677 1699 */
1678 1700 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS);
1679 1701
1680 1702 VERIFY0(dnode_hold(os, object, FTAG, &dn));
1681 1703 dn->dn_compress = compress;
1682 1704 dnode_setdirty(dn, tx);
1683 1705 dnode_rele(dn, FTAG);
1684 1706 }
1685 1707
1686 1708 int zfs_mdcomp_disable = 0;
1687 1709
1688 1710 /*
1689 1711 * When the "redundant_metadata" property is set to "most", only indirect
1690 1712 * blocks of this level and higher will have an additional ditto block.
1691 1713 */
1692 1714 int zfs_redundant_metadata_most_ditto_level = 2;
1693 1715
1694 1716 void
1695 1717 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
1696 1718 {
1697 1719 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
1698 1720 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
1699 1721 (wp & WP_SPILL));
1700 1722 enum zio_checksum checksum = os->os_checksum;
1701 1723 enum zio_compress compress = os->os_compress;
1702 1724 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
1703 1725 boolean_t dedup = B_FALSE;
1704 1726 boolean_t nopwrite = B_FALSE;
1705 1727 boolean_t dedup_verify = os->os_dedup_verify;
1706 1728 int copies = os->os_copies;
1707 1729
1708 1730 /*
1709 1731 * We maintain different write policies for each of the following
1710 1732 * types of data:
1711 1733 * 1. metadata
1712 1734 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1713 1735 * 3. all other level 0 blocks
1714 1736 */
1715 1737 if (ismd) {
1716 1738 /*
1717 1739 * XXX -- we should design a compression algorithm
1718 1740 * that specializes in arrays of bps.
1719 1741 */
1720 1742 boolean_t lz4_ac = spa_feature_is_active(os->os_spa,
1721 1743 SPA_FEATURE_LZ4_COMPRESS);
1722 1744
1723 1745 if (zfs_mdcomp_disable) {
1724 1746 compress = ZIO_COMPRESS_EMPTY;
1725 1747 } else if (lz4_ac) {
1726 1748 compress = ZIO_COMPRESS_LZ4;
1727 1749 } else {
1728 1750 compress = ZIO_COMPRESS_LZJB;
1729 1751 }
1730 1752
1731 1753 /*
1732 1754 * Metadata always gets checksummed. If the data
1733 1755 * checksum is multi-bit correctable, and it's not a
1734 1756 * ZBT-style checksum, then it's suitable for metadata
1735 1757 * as well. Otherwise, the metadata checksum defaults
1736 1758 * to fletcher4.
1737 1759 */
1738 1760 if (zio_checksum_table[checksum].ci_correctable < 1 ||
1739 1761 zio_checksum_table[checksum].ci_eck)
1740 1762 checksum = ZIO_CHECKSUM_FLETCHER_4;
1741 1763
1742 1764 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL ||
1743 1765 (os->os_redundant_metadata ==
1744 1766 ZFS_REDUNDANT_METADATA_MOST &&
1745 1767 (level >= zfs_redundant_metadata_most_ditto_level ||
1746 1768 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL))))
1747 1769 copies++;
1748 1770 } else if (wp & WP_NOFILL) {
1749 1771 ASSERT(level == 0);
1750 1772
1751 1773 /*
1752 1774 * If we're writing preallocated blocks, we aren't actually
1753 1775 * writing them so don't set any policy properties. These
1754 1776 * blocks are currently only used by an external subsystem
1755 1777 * outside of zfs (i.e. dump) and not written by the zio
1756 1778 * pipeline.
1757 1779 */
1758 1780 compress = ZIO_COMPRESS_OFF;
1759 1781 checksum = ZIO_CHECKSUM_NOPARITY;
1760 1782 } else {
1761 1783 compress = zio_compress_select(dn->dn_compress, compress);
1762 1784
1763 1785 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
1764 1786 zio_checksum_select(dn->dn_checksum, checksum) :
1765 1787 dedup_checksum;
1766 1788
1767 1789 /*
1768 1790 * Determine dedup setting. If we are in dmu_sync(),
1769 1791 * we won't actually dedup now because that's all
1770 1792 * done in syncing context; but we do want to use the
1771 1793 * dedup checkum. If the checksum is not strong
1772 1794 * enough to ensure unique signatures, force
1773 1795 * dedup_verify.
1774 1796 */
1775 1797 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
1776 1798 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
1777 1799 if (!zio_checksum_table[checksum].ci_dedup)
1778 1800 dedup_verify = B_TRUE;
1779 1801 }
1780 1802
1781 1803 /*
1782 1804 * Enable nopwrite if we have a cryptographically secure
1783 1805 * checksum that has no known collisions (i.e. SHA-256)
1784 1806 * and compression is enabled. We don't enable nopwrite if
1785 1807 * dedup is enabled as the two features are mutually exclusive.
1786 1808 */
1787 1809 nopwrite = (!dedup && zio_checksum_table[checksum].ci_dedup &&
1788 1810 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
|
↓ open down ↓ |
966 lines elided |
↑ open up ↑ |
1789 1811 }
1790 1812
1791 1813 zp->zp_checksum = checksum;
1792 1814 zp->zp_compress = compress;
1793 1815 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
1794 1816 zp->zp_level = level;
1795 1817 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa));
1796 1818 zp->zp_dedup = dedup;
1797 1819 zp->zp_dedup_verify = dedup && dedup_verify;
1798 1820 zp->zp_nopwrite = nopwrite;
1821 + zp->zp_zero_write = B_FALSE;
1799 1822 }
1800 1823
1801 1824 int
1802 1825 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
1803 1826 {
1804 1827 dnode_t *dn;
1805 1828 int i, err;
1806 1829
1807 1830 err = dnode_hold(os, object, FTAG, &dn);
1808 1831 if (err)
1809 1832 return (err);
1810 1833 /*
1811 1834 * Sync any current changes before
1812 1835 * we go trundling through the block pointers.
1813 1836 */
1814 1837 for (i = 0; i < TXG_SIZE; i++) {
1815 1838 if (list_link_active(&dn->dn_dirty_link[i]))
1816 1839 break;
1817 1840 }
1818 1841 if (i != TXG_SIZE) {
1819 1842 dnode_rele(dn, FTAG);
1820 1843 txg_wait_synced(dmu_objset_pool(os), 0);
1821 1844 err = dnode_hold(os, object, FTAG, &dn);
1822 1845 if (err)
1823 1846 return (err);
1824 1847 }
1825 1848
1826 1849 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
1827 1850 dnode_rele(dn, FTAG);
1828 1851
1829 1852 return (err);
1830 1853 }
1831 1854
1832 1855 void
1833 1856 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
1834 1857 {
1835 1858 dnode_phys_t *dnp;
1836 1859
1837 1860 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1838 1861 mutex_enter(&dn->dn_mtx);
1839 1862
1840 1863 dnp = dn->dn_phys;
1841 1864
1842 1865 doi->doi_data_block_size = dn->dn_datablksz;
1843 1866 doi->doi_metadata_block_size = dn->dn_indblkshift ?
1844 1867 1ULL << dn->dn_indblkshift : 0;
1845 1868 doi->doi_type = dn->dn_type;
1846 1869 doi->doi_bonus_type = dn->dn_bonustype;
1847 1870 doi->doi_bonus_size = dn->dn_bonuslen;
1848 1871 doi->doi_indirection = dn->dn_nlevels;
1849 1872 doi->doi_checksum = dn->dn_checksum;
1850 1873 doi->doi_compress = dn->dn_compress;
1851 1874 doi->doi_nblkptr = dn->dn_nblkptr;
1852 1875 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
1853 1876 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
1854 1877 doi->doi_fill_count = 0;
1855 1878 for (int i = 0; i < dnp->dn_nblkptr; i++)
1856 1879 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]);
1857 1880
1858 1881 mutex_exit(&dn->dn_mtx);
1859 1882 rw_exit(&dn->dn_struct_rwlock);
1860 1883 }
1861 1884
1862 1885 /*
1863 1886 * Get information on a DMU object.
1864 1887 * If doi is NULL, just indicates whether the object exists.
1865 1888 */
1866 1889 int
1867 1890 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
1868 1891 {
1869 1892 dnode_t *dn;
1870 1893 int err = dnode_hold(os, object, FTAG, &dn);
1871 1894
1872 1895 if (err)
1873 1896 return (err);
1874 1897
1875 1898 if (doi != NULL)
1876 1899 dmu_object_info_from_dnode(dn, doi);
1877 1900
1878 1901 dnode_rele(dn, FTAG);
1879 1902 return (0);
1880 1903 }
1881 1904
1882 1905 /*
1883 1906 * As above, but faster; can be used when you have a held dbuf in hand.
1884 1907 */
1885 1908 void
1886 1909 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
1887 1910 {
1888 1911 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1889 1912
1890 1913 DB_DNODE_ENTER(db);
1891 1914 dmu_object_info_from_dnode(DB_DNODE(db), doi);
1892 1915 DB_DNODE_EXIT(db);
1893 1916 }
1894 1917
1895 1918 /*
1896 1919 * Faster still when you only care about the size.
1897 1920 * This is specifically optimized for zfs_getattr().
1898 1921 */
1899 1922 void
1900 1923 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
1901 1924 u_longlong_t *nblk512)
1902 1925 {
1903 1926 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1904 1927 dnode_t *dn;
1905 1928
1906 1929 DB_DNODE_ENTER(db);
1907 1930 dn = DB_DNODE(db);
1908 1931
1909 1932 *blksize = dn->dn_datablksz;
1910 1933 /* add 1 for dnode space */
1911 1934 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
1912 1935 SPA_MINBLOCKSHIFT) + 1;
1913 1936 DB_DNODE_EXIT(db);
1914 1937 }
1915 1938
1916 1939 void
1917 1940 byteswap_uint64_array(void *vbuf, size_t size)
1918 1941 {
1919 1942 uint64_t *buf = vbuf;
1920 1943 size_t count = size >> 3;
1921 1944 int i;
1922 1945
1923 1946 ASSERT((size & 7) == 0);
1924 1947
1925 1948 for (i = 0; i < count; i++)
1926 1949 buf[i] = BSWAP_64(buf[i]);
1927 1950 }
1928 1951
1929 1952 void
1930 1953 byteswap_uint32_array(void *vbuf, size_t size)
1931 1954 {
1932 1955 uint32_t *buf = vbuf;
1933 1956 size_t count = size >> 2;
1934 1957 int i;
1935 1958
1936 1959 ASSERT((size & 3) == 0);
1937 1960
1938 1961 for (i = 0; i < count; i++)
1939 1962 buf[i] = BSWAP_32(buf[i]);
1940 1963 }
1941 1964
1942 1965 void
1943 1966 byteswap_uint16_array(void *vbuf, size_t size)
1944 1967 {
1945 1968 uint16_t *buf = vbuf;
1946 1969 size_t count = size >> 1;
1947 1970 int i;
1948 1971
1949 1972 ASSERT((size & 1) == 0);
1950 1973
1951 1974 for (i = 0; i < count; i++)
1952 1975 buf[i] = BSWAP_16(buf[i]);
1953 1976 }
1954 1977
1955 1978 /* ARGSUSED */
1956 1979 void
1957 1980 byteswap_uint8_array(void *vbuf, size_t size)
1958 1981 {
1959 1982 }
1960 1983
1961 1984 void
1962 1985 dmu_init(void)
1963 1986 {
1964 1987 zfs_dbgmsg_init();
1965 1988 sa_cache_init();
1966 1989 xuio_stat_init();
1967 1990 dmu_objset_init();
1968 1991 dnode_init();
1969 1992 dbuf_init();
1970 1993 zfetch_init();
1971 1994 l2arc_init();
1972 1995 arc_init();
1973 1996 }
1974 1997
1975 1998 void
1976 1999 dmu_fini(void)
1977 2000 {
1978 2001 arc_fini(); /* arc depends on l2arc, so arc must go first */
1979 2002 l2arc_fini();
1980 2003 zfetch_fini();
1981 2004 dbuf_fini();
1982 2005 dnode_fini();
1983 2006 dmu_objset_fini();
1984 2007 xuio_stat_fini();
1985 2008 sa_cache_fini();
1986 2009 zfs_dbgmsg_fini();
1987 2010 }
|
↓ open down ↓ |
179 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX