Print this page
3756 want lz4 support for metadata compression
| 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
|
↓ open down ↓ |
13 lines elided |
↑ open up ↑ |
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) 2013 by Delphix. All rights reserved.
24 + * Copyright (c) 2013 Martin Matuska. All rights reserved.
24 25 */
25 26
26 27 #include <sys/dmu.h>
27 28 #include <sys/dmu_impl.h>
28 29 #include <sys/dmu_tx.h>
29 30 #include <sys/dbuf.h>
30 31 #include <sys/dnode.h>
31 32 #include <sys/zfs_context.h>
32 33 #include <sys/dmu_objset.h>
33 34 #include <sys/dmu_traverse.h>
34 35 #include <sys/dsl_dataset.h>
|
↓ open down ↓ |
1 lines elided |
↑ open up ↑ |
35 36 #include <sys/dsl_dir.h>
36 37 #include <sys/dsl_pool.h>
37 38 #include <sys/dsl_synctask.h>
38 39 #include <sys/dsl_prop.h>
39 40 #include <sys/dmu_zfetch.h>
40 41 #include <sys/zfs_ioctl.h>
41 42 #include <sys/zap.h>
42 43 #include <sys/zio_checksum.h>
43 44 #include <sys/zio_compress.h>
44 45 #include <sys/sa.h>
46 +#include <sys/zfeature.h>
45 47 #ifdef _KERNEL
46 48 #include <sys/vmsystm.h>
47 49 #include <sys/zfs_znode.h>
48 50 #endif
49 51
50 52 /*
51 53 * Enable/disable nopwrite feature.
52 54 */
53 55 int zfs_nopwrite_enabled = 1;
54 56
55 57 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = {
56 58 { DMU_BSWAP_UINT8, TRUE, "unallocated" },
57 59 { DMU_BSWAP_ZAP, TRUE, "object directory" },
58 60 { DMU_BSWAP_UINT64, TRUE, "object array" },
59 61 { DMU_BSWAP_UINT8, TRUE, "packed nvlist" },
60 62 { DMU_BSWAP_UINT64, TRUE, "packed nvlist size" },
61 63 { DMU_BSWAP_UINT64, TRUE, "bpobj" },
62 64 { DMU_BSWAP_UINT64, TRUE, "bpobj header" },
63 65 { DMU_BSWAP_UINT64, TRUE, "SPA space map header" },
64 66 { DMU_BSWAP_UINT64, TRUE, "SPA space map" },
65 67 { DMU_BSWAP_UINT64, TRUE, "ZIL intent log" },
66 68 { DMU_BSWAP_DNODE, TRUE, "DMU dnode" },
67 69 { DMU_BSWAP_OBJSET, TRUE, "DMU objset" },
68 70 { DMU_BSWAP_UINT64, TRUE, "DSL directory" },
69 71 { DMU_BSWAP_ZAP, TRUE, "DSL directory child map"},
70 72 { DMU_BSWAP_ZAP, TRUE, "DSL dataset snap map" },
71 73 { DMU_BSWAP_ZAP, TRUE, "DSL props" },
72 74 { DMU_BSWAP_UINT64, TRUE, "DSL dataset" },
73 75 { DMU_BSWAP_ZNODE, TRUE, "ZFS znode" },
74 76 { DMU_BSWAP_OLDACL, TRUE, "ZFS V0 ACL" },
75 77 { DMU_BSWAP_UINT8, FALSE, "ZFS plain file" },
76 78 { DMU_BSWAP_ZAP, TRUE, "ZFS directory" },
77 79 { DMU_BSWAP_ZAP, TRUE, "ZFS master node" },
78 80 { DMU_BSWAP_ZAP, TRUE, "ZFS delete queue" },
79 81 { DMU_BSWAP_UINT8, FALSE, "zvol object" },
80 82 { DMU_BSWAP_ZAP, TRUE, "zvol prop" },
81 83 { DMU_BSWAP_UINT8, FALSE, "other uint8[]" },
82 84 { DMU_BSWAP_UINT64, FALSE, "other uint64[]" },
83 85 { DMU_BSWAP_ZAP, TRUE, "other ZAP" },
84 86 { DMU_BSWAP_ZAP, TRUE, "persistent error log" },
85 87 { DMU_BSWAP_UINT8, TRUE, "SPA history" },
86 88 { DMU_BSWAP_UINT64, TRUE, "SPA history offsets" },
87 89 { DMU_BSWAP_ZAP, TRUE, "Pool properties" },
88 90 { DMU_BSWAP_ZAP, TRUE, "DSL permissions" },
89 91 { DMU_BSWAP_ACL, TRUE, "ZFS ACL" },
90 92 { DMU_BSWAP_UINT8, TRUE, "ZFS SYSACL" },
91 93 { DMU_BSWAP_UINT8, TRUE, "FUID table" },
92 94 { DMU_BSWAP_UINT64, TRUE, "FUID table size" },
93 95 { DMU_BSWAP_ZAP, TRUE, "DSL dataset next clones"},
94 96 { DMU_BSWAP_ZAP, TRUE, "scan work queue" },
95 97 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group used" },
96 98 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group quota" },
97 99 { DMU_BSWAP_ZAP, TRUE, "snapshot refcount tags"},
98 100 { DMU_BSWAP_ZAP, TRUE, "DDT ZAP algorithm" },
99 101 { DMU_BSWAP_ZAP, TRUE, "DDT statistics" },
100 102 { DMU_BSWAP_UINT8, TRUE, "System attributes" },
101 103 { DMU_BSWAP_ZAP, TRUE, "SA master node" },
102 104 { DMU_BSWAP_ZAP, TRUE, "SA attr registration" },
103 105 { DMU_BSWAP_ZAP, TRUE, "SA attr layouts" },
104 106 { DMU_BSWAP_ZAP, TRUE, "scan translations" },
105 107 { DMU_BSWAP_UINT8, FALSE, "deduplicated block" },
106 108 { DMU_BSWAP_ZAP, TRUE, "DSL deadlist map" },
107 109 { DMU_BSWAP_UINT64, TRUE, "DSL deadlist map hdr" },
108 110 { DMU_BSWAP_ZAP, TRUE, "DSL dir clones" },
109 111 { DMU_BSWAP_UINT64, TRUE, "bpobj subobj" }
110 112 };
111 113
112 114 const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = {
113 115 { byteswap_uint8_array, "uint8" },
114 116 { byteswap_uint16_array, "uint16" },
115 117 { byteswap_uint32_array, "uint32" },
116 118 { byteswap_uint64_array, "uint64" },
117 119 { zap_byteswap, "zap" },
118 120 { dnode_buf_byteswap, "dnode" },
119 121 { dmu_objset_byteswap, "objset" },
120 122 { zfs_znode_byteswap, "znode" },
121 123 { zfs_oldacl_byteswap, "oldacl" },
122 124 { zfs_acl_byteswap, "acl" }
123 125 };
124 126
125 127 int
126 128 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset,
127 129 void *tag, dmu_buf_t **dbp, int flags)
128 130 {
129 131 dnode_t *dn;
130 132 uint64_t blkid;
131 133 dmu_buf_impl_t *db;
132 134 int err;
133 135 int db_flags = DB_RF_CANFAIL;
134 136
135 137 if (flags & DMU_READ_NO_PREFETCH)
136 138 db_flags |= DB_RF_NOPREFETCH;
137 139
138 140 err = dnode_hold(os, object, FTAG, &dn);
139 141 if (err)
140 142 return (err);
141 143 blkid = dbuf_whichblock(dn, offset);
142 144 rw_enter(&dn->dn_struct_rwlock, RW_READER);
143 145 db = dbuf_hold(dn, blkid, tag);
144 146 rw_exit(&dn->dn_struct_rwlock);
145 147 if (db == NULL) {
146 148 err = SET_ERROR(EIO);
147 149 } else {
148 150 err = dbuf_read(db, NULL, db_flags);
149 151 if (err) {
150 152 dbuf_rele(db, tag);
151 153 db = NULL;
152 154 }
153 155 }
154 156
155 157 dnode_rele(dn, FTAG);
156 158 *dbp = &db->db; /* NULL db plus first field offset is NULL */
157 159 return (err);
158 160 }
159 161
160 162 int
161 163 dmu_bonus_max(void)
162 164 {
163 165 return (DN_MAX_BONUSLEN);
164 166 }
165 167
166 168 int
167 169 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx)
168 170 {
169 171 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
170 172 dnode_t *dn;
171 173 int error;
172 174
173 175 DB_DNODE_ENTER(db);
174 176 dn = DB_DNODE(db);
175 177
176 178 if (dn->dn_bonus != db) {
177 179 error = SET_ERROR(EINVAL);
178 180 } else if (newsize < 0 || newsize > db_fake->db_size) {
179 181 error = SET_ERROR(EINVAL);
180 182 } else {
181 183 dnode_setbonuslen(dn, newsize, tx);
182 184 error = 0;
183 185 }
184 186
185 187 DB_DNODE_EXIT(db);
186 188 return (error);
187 189 }
188 190
189 191 int
190 192 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx)
191 193 {
192 194 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
193 195 dnode_t *dn;
194 196 int error;
195 197
196 198 DB_DNODE_ENTER(db);
197 199 dn = DB_DNODE(db);
198 200
199 201 if (!DMU_OT_IS_VALID(type)) {
200 202 error = SET_ERROR(EINVAL);
201 203 } else if (dn->dn_bonus != db) {
202 204 error = SET_ERROR(EINVAL);
203 205 } else {
204 206 dnode_setbonus_type(dn, type, tx);
205 207 error = 0;
206 208 }
207 209
208 210 DB_DNODE_EXIT(db);
209 211 return (error);
210 212 }
211 213
212 214 dmu_object_type_t
213 215 dmu_get_bonustype(dmu_buf_t *db_fake)
214 216 {
215 217 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
216 218 dnode_t *dn;
217 219 dmu_object_type_t type;
218 220
219 221 DB_DNODE_ENTER(db);
220 222 dn = DB_DNODE(db);
221 223 type = dn->dn_bonustype;
222 224 DB_DNODE_EXIT(db);
223 225
224 226 return (type);
225 227 }
226 228
227 229 int
228 230 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx)
229 231 {
230 232 dnode_t *dn;
231 233 int error;
232 234
233 235 error = dnode_hold(os, object, FTAG, &dn);
234 236 dbuf_rm_spill(dn, tx);
235 237 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
236 238 dnode_rm_spill(dn, tx);
237 239 rw_exit(&dn->dn_struct_rwlock);
238 240 dnode_rele(dn, FTAG);
239 241 return (error);
240 242 }
241 243
242 244 /*
243 245 * returns ENOENT, EIO, or 0.
244 246 */
245 247 int
246 248 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp)
247 249 {
248 250 dnode_t *dn;
249 251 dmu_buf_impl_t *db;
250 252 int error;
251 253
252 254 error = dnode_hold(os, object, FTAG, &dn);
253 255 if (error)
254 256 return (error);
255 257
256 258 rw_enter(&dn->dn_struct_rwlock, RW_READER);
257 259 if (dn->dn_bonus == NULL) {
258 260 rw_exit(&dn->dn_struct_rwlock);
259 261 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
260 262 if (dn->dn_bonus == NULL)
261 263 dbuf_create_bonus(dn);
262 264 }
263 265 db = dn->dn_bonus;
264 266
265 267 /* as long as the bonus buf is held, the dnode will be held */
266 268 if (refcount_add(&db->db_holds, tag) == 1) {
267 269 VERIFY(dnode_add_ref(dn, db));
268 270 (void) atomic_inc_32_nv(&dn->dn_dbufs_count);
269 271 }
270 272
271 273 /*
272 274 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's
273 275 * hold and incrementing the dbuf count to ensure that dnode_move() sees
274 276 * a dnode hold for every dbuf.
275 277 */
276 278 rw_exit(&dn->dn_struct_rwlock);
277 279
278 280 dnode_rele(dn, FTAG);
279 281
280 282 VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH));
281 283
282 284 *dbp = &db->db;
283 285 return (0);
284 286 }
285 287
286 288 /*
287 289 * returns ENOENT, EIO, or 0.
288 290 *
289 291 * This interface will allocate a blank spill dbuf when a spill blk
290 292 * doesn't already exist on the dnode.
291 293 *
292 294 * if you only want to find an already existing spill db, then
293 295 * dmu_spill_hold_existing() should be used.
294 296 */
295 297 int
296 298 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp)
297 299 {
298 300 dmu_buf_impl_t *db = NULL;
299 301 int err;
300 302
301 303 if ((flags & DB_RF_HAVESTRUCT) == 0)
302 304 rw_enter(&dn->dn_struct_rwlock, RW_READER);
303 305
304 306 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag);
305 307
306 308 if ((flags & DB_RF_HAVESTRUCT) == 0)
307 309 rw_exit(&dn->dn_struct_rwlock);
308 310
309 311 ASSERT(db != NULL);
310 312 err = dbuf_read(db, NULL, flags);
311 313 if (err == 0)
312 314 *dbp = &db->db;
313 315 else
314 316 dbuf_rele(db, tag);
315 317 return (err);
316 318 }
317 319
318 320 int
319 321 dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
320 322 {
321 323 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
322 324 dnode_t *dn;
323 325 int err;
324 326
325 327 DB_DNODE_ENTER(db);
326 328 dn = DB_DNODE(db);
327 329
328 330 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) {
329 331 err = SET_ERROR(EINVAL);
330 332 } else {
331 333 rw_enter(&dn->dn_struct_rwlock, RW_READER);
332 334
333 335 if (!dn->dn_have_spill) {
334 336 err = SET_ERROR(ENOENT);
335 337 } else {
336 338 err = dmu_spill_hold_by_dnode(dn,
337 339 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp);
338 340 }
339 341
340 342 rw_exit(&dn->dn_struct_rwlock);
341 343 }
342 344
343 345 DB_DNODE_EXIT(db);
344 346 return (err);
345 347 }
346 348
347 349 int
348 350 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp)
349 351 {
350 352 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus;
351 353 dnode_t *dn;
352 354 int err;
353 355
354 356 DB_DNODE_ENTER(db);
355 357 dn = DB_DNODE(db);
356 358 err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp);
357 359 DB_DNODE_EXIT(db);
358 360
359 361 return (err);
360 362 }
361 363
362 364 /*
363 365 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces
364 366 * to take a held dnode rather than <os, object> -- the lookup is wasteful,
365 367 * and can induce severe lock contention when writing to several files
366 368 * whose dnodes are in the same block.
367 369 */
368 370 static int
369 371 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
370 372 int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags)
371 373 {
372 374 dsl_pool_t *dp = NULL;
373 375 dmu_buf_t **dbp;
374 376 uint64_t blkid, nblks, i;
375 377 uint32_t dbuf_flags;
376 378 int err;
377 379 zio_t *zio;
378 380 hrtime_t start;
379 381
380 382 ASSERT(length <= DMU_MAX_ACCESS);
381 383
382 384 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT;
383 385 if (flags & DMU_READ_NO_PREFETCH || length > zfetch_array_rd_sz)
384 386 dbuf_flags |= DB_RF_NOPREFETCH;
385 387
386 388 rw_enter(&dn->dn_struct_rwlock, RW_READER);
387 389 if (dn->dn_datablkshift) {
388 390 int blkshift = dn->dn_datablkshift;
389 391 nblks = (P2ROUNDUP(offset+length, 1ULL<<blkshift) -
390 392 P2ALIGN(offset, 1ULL<<blkshift)) >> blkshift;
391 393 } else {
392 394 if (offset + length > dn->dn_datablksz) {
393 395 zfs_panic_recover("zfs: accessing past end of object "
394 396 "%llx/%llx (size=%u access=%llu+%llu)",
395 397 (longlong_t)dn->dn_objset->
396 398 os_dsl_dataset->ds_object,
397 399 (longlong_t)dn->dn_object, dn->dn_datablksz,
398 400 (longlong_t)offset, (longlong_t)length);
399 401 rw_exit(&dn->dn_struct_rwlock);
400 402 return (SET_ERROR(EIO));
401 403 }
402 404 nblks = 1;
403 405 }
404 406 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP);
405 407
406 408 if (dn->dn_objset->os_dsl_dataset)
407 409 dp = dn->dn_objset->os_dsl_dataset->ds_dir->dd_pool;
408 410 start = gethrtime();
409 411 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL);
410 412 blkid = dbuf_whichblock(dn, offset);
411 413 for (i = 0; i < nblks; i++) {
412 414 dmu_buf_impl_t *db = dbuf_hold(dn, blkid+i, tag);
413 415 if (db == NULL) {
414 416 rw_exit(&dn->dn_struct_rwlock);
415 417 dmu_buf_rele_array(dbp, nblks, tag);
416 418 zio_nowait(zio);
417 419 return (SET_ERROR(EIO));
418 420 }
419 421 /* initiate async i/o */
420 422 if (read) {
421 423 (void) dbuf_read(db, zio, dbuf_flags);
422 424 }
423 425 dbp[i] = &db->db;
424 426 }
425 427 rw_exit(&dn->dn_struct_rwlock);
426 428
427 429 /* wait for async i/o */
428 430 err = zio_wait(zio);
429 431 /* track read overhead when we are in sync context */
430 432 if (dp && dsl_pool_sync_context(dp))
431 433 dp->dp_read_overhead += gethrtime() - start;
432 434 if (err) {
433 435 dmu_buf_rele_array(dbp, nblks, tag);
434 436 return (err);
435 437 }
436 438
437 439 /* wait for other io to complete */
438 440 if (read) {
439 441 for (i = 0; i < nblks; i++) {
440 442 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i];
441 443 mutex_enter(&db->db_mtx);
442 444 while (db->db_state == DB_READ ||
443 445 db->db_state == DB_FILL)
444 446 cv_wait(&db->db_changed, &db->db_mtx);
445 447 if (db->db_state == DB_UNCACHED)
446 448 err = SET_ERROR(EIO);
447 449 mutex_exit(&db->db_mtx);
448 450 if (err) {
449 451 dmu_buf_rele_array(dbp, nblks, tag);
450 452 return (err);
451 453 }
452 454 }
453 455 }
454 456
455 457 *numbufsp = nblks;
456 458 *dbpp = dbp;
457 459 return (0);
458 460 }
459 461
460 462 static int
461 463 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset,
462 464 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
463 465 {
464 466 dnode_t *dn;
465 467 int err;
466 468
467 469 err = dnode_hold(os, object, FTAG, &dn);
468 470 if (err)
469 471 return (err);
470 472
471 473 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
472 474 numbufsp, dbpp, DMU_READ_PREFETCH);
473 475
474 476 dnode_rele(dn, FTAG);
475 477
476 478 return (err);
477 479 }
478 480
479 481 int
480 482 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset,
481 483 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp)
482 484 {
483 485 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
484 486 dnode_t *dn;
485 487 int err;
486 488
487 489 DB_DNODE_ENTER(db);
488 490 dn = DB_DNODE(db);
489 491 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag,
490 492 numbufsp, dbpp, DMU_READ_PREFETCH);
491 493 DB_DNODE_EXIT(db);
492 494
493 495 return (err);
494 496 }
495 497
496 498 void
497 499 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag)
498 500 {
499 501 int i;
500 502 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake;
501 503
502 504 if (numbufs == 0)
503 505 return;
504 506
505 507 for (i = 0; i < numbufs; i++) {
506 508 if (dbp[i])
507 509 dbuf_rele(dbp[i], tag);
508 510 }
509 511
510 512 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs);
511 513 }
512 514
513 515 void
514 516 dmu_prefetch(objset_t *os, uint64_t object, uint64_t offset, uint64_t len)
515 517 {
516 518 dnode_t *dn;
517 519 uint64_t blkid;
518 520 int nblks, i, err;
519 521
520 522 if (zfs_prefetch_disable)
521 523 return;
522 524
523 525 if (len == 0) { /* they're interested in the bonus buffer */
524 526 dn = DMU_META_DNODE(os);
525 527
526 528 if (object == 0 || object >= DN_MAX_OBJECT)
527 529 return;
528 530
529 531 rw_enter(&dn->dn_struct_rwlock, RW_READER);
530 532 blkid = dbuf_whichblock(dn, object * sizeof (dnode_phys_t));
531 533 dbuf_prefetch(dn, blkid);
532 534 rw_exit(&dn->dn_struct_rwlock);
533 535 return;
534 536 }
535 537
536 538 /*
537 539 * XXX - Note, if the dnode for the requested object is not
538 540 * already cached, we will do a *synchronous* read in the
539 541 * dnode_hold() call. The same is true for any indirects.
540 542 */
541 543 err = dnode_hold(os, object, FTAG, &dn);
542 544 if (err != 0)
543 545 return;
544 546
545 547 rw_enter(&dn->dn_struct_rwlock, RW_READER);
546 548 if (dn->dn_datablkshift) {
547 549 int blkshift = dn->dn_datablkshift;
548 550 nblks = (P2ROUNDUP(offset+len, 1<<blkshift) -
549 551 P2ALIGN(offset, 1<<blkshift)) >> blkshift;
550 552 } else {
551 553 nblks = (offset < dn->dn_datablksz);
552 554 }
553 555
554 556 if (nblks != 0) {
555 557 blkid = dbuf_whichblock(dn, offset);
556 558 for (i = 0; i < nblks; i++)
557 559 dbuf_prefetch(dn, blkid+i);
558 560 }
559 561
560 562 rw_exit(&dn->dn_struct_rwlock);
561 563
562 564 dnode_rele(dn, FTAG);
563 565 }
564 566
565 567 /*
566 568 * Get the next "chunk" of file data to free. We traverse the file from
567 569 * the end so that the file gets shorter over time (if we crashes in the
568 570 * middle, this will leave us in a better state). We find allocated file
569 571 * data by simply searching the allocated level 1 indirects.
570 572 */
571 573 static int
572 574 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t limit)
573 575 {
574 576 uint64_t len = *start - limit;
575 577 uint64_t blkcnt = 0;
576 578 uint64_t maxblks = DMU_MAX_ACCESS / (1ULL << (dn->dn_indblkshift + 1));
577 579 uint64_t iblkrange =
578 580 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT);
579 581
580 582 ASSERT(limit <= *start);
581 583
582 584 if (len <= iblkrange * maxblks) {
583 585 *start = limit;
584 586 return (0);
585 587 }
586 588 ASSERT(ISP2(iblkrange));
587 589
588 590 while (*start > limit && blkcnt < maxblks) {
589 591 int err;
590 592
591 593 /* find next allocated L1 indirect */
592 594 err = dnode_next_offset(dn,
593 595 DNODE_FIND_BACKWARDS, start, 2, 1, 0);
594 596
595 597 /* if there are no more, then we are done */
596 598 if (err == ESRCH) {
597 599 *start = limit;
598 600 return (0);
599 601 } else if (err) {
600 602 return (err);
601 603 }
602 604 blkcnt += 1;
603 605
604 606 /* reset offset to end of "next" block back */
605 607 *start = P2ALIGN(*start, iblkrange);
606 608 if (*start <= limit)
607 609 *start = limit;
608 610 else
609 611 *start -= 1;
610 612 }
611 613 return (0);
612 614 }
613 615
614 616 static int
615 617 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset,
616 618 uint64_t length, boolean_t free_dnode)
617 619 {
618 620 dmu_tx_t *tx;
619 621 uint64_t object_size, start, end, len;
620 622 boolean_t trunc = (length == DMU_OBJECT_END);
621 623 int align, err;
622 624
623 625 align = 1 << dn->dn_datablkshift;
624 626 ASSERT(align > 0);
625 627 object_size = align == 1 ? dn->dn_datablksz :
626 628 (dn->dn_maxblkid + 1) << dn->dn_datablkshift;
627 629
628 630 end = offset + length;
629 631 if (trunc || end > object_size)
630 632 end = object_size;
631 633 if (end <= offset)
632 634 return (0);
633 635 length = end - offset;
634 636
635 637 while (length) {
636 638 start = end;
637 639 /* assert(offset <= start) */
638 640 err = get_next_chunk(dn, &start, offset);
639 641 if (err)
640 642 return (err);
641 643 len = trunc ? DMU_OBJECT_END : end - start;
642 644
643 645 tx = dmu_tx_create(os);
644 646 dmu_tx_hold_free(tx, dn->dn_object, start, len);
645 647 err = dmu_tx_assign(tx, TXG_WAIT);
646 648 if (err) {
647 649 dmu_tx_abort(tx);
648 650 return (err);
649 651 }
650 652
651 653 dnode_free_range(dn, start, trunc ? -1 : len, tx);
652 654
653 655 if (start == 0 && free_dnode) {
654 656 ASSERT(trunc);
655 657 dnode_free(dn, tx);
656 658 }
657 659
658 660 length -= end - start;
659 661
660 662 dmu_tx_commit(tx);
661 663 end = start;
662 664 }
663 665 return (0);
664 666 }
665 667
666 668 int
667 669 dmu_free_long_range(objset_t *os, uint64_t object,
668 670 uint64_t offset, uint64_t length)
669 671 {
670 672 dnode_t *dn;
671 673 int err;
672 674
673 675 err = dnode_hold(os, object, FTAG, &dn);
674 676 if (err != 0)
675 677 return (err);
676 678 err = dmu_free_long_range_impl(os, dn, offset, length, FALSE);
677 679 dnode_rele(dn, FTAG);
678 680 return (err);
679 681 }
680 682
681 683 int
682 684 dmu_free_object(objset_t *os, uint64_t object)
683 685 {
684 686 dnode_t *dn;
685 687 dmu_tx_t *tx;
686 688 int err;
687 689
688 690 err = dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED,
689 691 FTAG, &dn);
690 692 if (err != 0)
691 693 return (err);
692 694 if (dn->dn_nlevels == 1) {
693 695 tx = dmu_tx_create(os);
694 696 dmu_tx_hold_bonus(tx, object);
695 697 dmu_tx_hold_free(tx, dn->dn_object, 0, DMU_OBJECT_END);
696 698 err = dmu_tx_assign(tx, TXG_WAIT);
697 699 if (err == 0) {
698 700 dnode_free_range(dn, 0, DMU_OBJECT_END, tx);
699 701 dnode_free(dn, tx);
700 702 dmu_tx_commit(tx);
701 703 } else {
702 704 dmu_tx_abort(tx);
703 705 }
704 706 } else {
705 707 err = dmu_free_long_range_impl(os, dn, 0, DMU_OBJECT_END, TRUE);
706 708 }
707 709 dnode_rele(dn, FTAG);
708 710 return (err);
709 711 }
710 712
711 713 int
712 714 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset,
713 715 uint64_t size, dmu_tx_t *tx)
714 716 {
715 717 dnode_t *dn;
716 718 int err = dnode_hold(os, object, FTAG, &dn);
717 719 if (err)
718 720 return (err);
719 721 ASSERT(offset < UINT64_MAX);
720 722 ASSERT(size == -1ULL || size <= UINT64_MAX - offset);
721 723 dnode_free_range(dn, offset, size, tx);
722 724 dnode_rele(dn, FTAG);
723 725 return (0);
724 726 }
725 727
726 728 int
727 729 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
728 730 void *buf, uint32_t flags)
729 731 {
730 732 dnode_t *dn;
731 733 dmu_buf_t **dbp;
732 734 int numbufs, err;
733 735
734 736 err = dnode_hold(os, object, FTAG, &dn);
735 737 if (err)
736 738 return (err);
737 739
738 740 /*
739 741 * Deal with odd block sizes, where there can't be data past the first
740 742 * block. If we ever do the tail block optimization, we will need to
741 743 * handle that here as well.
742 744 */
743 745 if (dn->dn_maxblkid == 0) {
744 746 int newsz = offset > dn->dn_datablksz ? 0 :
745 747 MIN(size, dn->dn_datablksz - offset);
746 748 bzero((char *)buf + newsz, size - newsz);
747 749 size = newsz;
748 750 }
749 751
750 752 while (size > 0) {
751 753 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2);
752 754 int i;
753 755
754 756 /*
755 757 * NB: we could do this block-at-a-time, but it's nice
756 758 * to be reading in parallel.
757 759 */
758 760 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen,
759 761 TRUE, FTAG, &numbufs, &dbp, flags);
760 762 if (err)
761 763 break;
762 764
763 765 for (i = 0; i < numbufs; i++) {
764 766 int tocpy;
765 767 int bufoff;
766 768 dmu_buf_t *db = dbp[i];
767 769
768 770 ASSERT(size > 0);
769 771
770 772 bufoff = offset - db->db_offset;
771 773 tocpy = (int)MIN(db->db_size - bufoff, size);
772 774
773 775 bcopy((char *)db->db_data + bufoff, buf, tocpy);
774 776
775 777 offset += tocpy;
776 778 size -= tocpy;
777 779 buf = (char *)buf + tocpy;
778 780 }
779 781 dmu_buf_rele_array(dbp, numbufs, FTAG);
780 782 }
781 783 dnode_rele(dn, FTAG);
782 784 return (err);
783 785 }
784 786
785 787 void
786 788 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
787 789 const void *buf, dmu_tx_t *tx)
788 790 {
789 791 dmu_buf_t **dbp;
790 792 int numbufs, i;
791 793
792 794 if (size == 0)
793 795 return;
794 796
795 797 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
796 798 FALSE, FTAG, &numbufs, &dbp));
797 799
798 800 for (i = 0; i < numbufs; i++) {
799 801 int tocpy;
800 802 int bufoff;
801 803 dmu_buf_t *db = dbp[i];
802 804
803 805 ASSERT(size > 0);
804 806
805 807 bufoff = offset - db->db_offset;
806 808 tocpy = (int)MIN(db->db_size - bufoff, size);
807 809
808 810 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
809 811
810 812 if (tocpy == db->db_size)
811 813 dmu_buf_will_fill(db, tx);
812 814 else
813 815 dmu_buf_will_dirty(db, tx);
814 816
815 817 bcopy(buf, (char *)db->db_data + bufoff, tocpy);
816 818
817 819 if (tocpy == db->db_size)
818 820 dmu_buf_fill_done(db, tx);
819 821
820 822 offset += tocpy;
821 823 size -= tocpy;
822 824 buf = (char *)buf + tocpy;
823 825 }
824 826 dmu_buf_rele_array(dbp, numbufs, FTAG);
825 827 }
826 828
827 829 void
828 830 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
829 831 dmu_tx_t *tx)
830 832 {
831 833 dmu_buf_t **dbp;
832 834 int numbufs, i;
833 835
834 836 if (size == 0)
835 837 return;
836 838
837 839 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size,
838 840 FALSE, FTAG, &numbufs, &dbp));
839 841
840 842 for (i = 0; i < numbufs; i++) {
841 843 dmu_buf_t *db = dbp[i];
842 844
843 845 dmu_buf_will_not_fill(db, tx);
844 846 }
845 847 dmu_buf_rele_array(dbp, numbufs, FTAG);
846 848 }
847 849
848 850 /*
849 851 * DMU support for xuio
850 852 */
851 853 kstat_t *xuio_ksp = NULL;
852 854
853 855 int
854 856 dmu_xuio_init(xuio_t *xuio, int nblk)
855 857 {
856 858 dmu_xuio_t *priv;
857 859 uio_t *uio = &xuio->xu_uio;
858 860
859 861 uio->uio_iovcnt = nblk;
860 862 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP);
861 863
862 864 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP);
863 865 priv->cnt = nblk;
864 866 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP);
865 867 priv->iovp = uio->uio_iov;
866 868 XUIO_XUZC_PRIV(xuio) = priv;
867 869
868 870 if (XUIO_XUZC_RW(xuio) == UIO_READ)
869 871 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk);
870 872 else
871 873 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk);
872 874
873 875 return (0);
874 876 }
875 877
876 878 void
877 879 dmu_xuio_fini(xuio_t *xuio)
878 880 {
879 881 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
880 882 int nblk = priv->cnt;
881 883
882 884 kmem_free(priv->iovp, nblk * sizeof (iovec_t));
883 885 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *));
884 886 kmem_free(priv, sizeof (dmu_xuio_t));
885 887
886 888 if (XUIO_XUZC_RW(xuio) == UIO_READ)
887 889 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk);
888 890 else
889 891 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk);
890 892 }
891 893
892 894 /*
893 895 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf }
894 896 * and increase priv->next by 1.
895 897 */
896 898 int
897 899 dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n)
898 900 {
899 901 struct iovec *iov;
900 902 uio_t *uio = &xuio->xu_uio;
901 903 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
902 904 int i = priv->next++;
903 905
904 906 ASSERT(i < priv->cnt);
905 907 ASSERT(off + n <= arc_buf_size(abuf));
906 908 iov = uio->uio_iov + i;
907 909 iov->iov_base = (char *)abuf->b_data + off;
908 910 iov->iov_len = n;
909 911 priv->bufs[i] = abuf;
910 912 return (0);
911 913 }
912 914
913 915 int
914 916 dmu_xuio_cnt(xuio_t *xuio)
915 917 {
916 918 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
917 919 return (priv->cnt);
918 920 }
919 921
920 922 arc_buf_t *
921 923 dmu_xuio_arcbuf(xuio_t *xuio, int i)
922 924 {
923 925 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
924 926
925 927 ASSERT(i < priv->cnt);
926 928 return (priv->bufs[i]);
927 929 }
928 930
929 931 void
930 932 dmu_xuio_clear(xuio_t *xuio, int i)
931 933 {
932 934 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio);
933 935
934 936 ASSERT(i < priv->cnt);
935 937 priv->bufs[i] = NULL;
936 938 }
937 939
938 940 static void
939 941 xuio_stat_init(void)
940 942 {
941 943 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc",
942 944 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t),
943 945 KSTAT_FLAG_VIRTUAL);
944 946 if (xuio_ksp != NULL) {
945 947 xuio_ksp->ks_data = &xuio_stats;
946 948 kstat_install(xuio_ksp);
947 949 }
948 950 }
949 951
950 952 static void
951 953 xuio_stat_fini(void)
952 954 {
953 955 if (xuio_ksp != NULL) {
954 956 kstat_delete(xuio_ksp);
955 957 xuio_ksp = NULL;
956 958 }
957 959 }
958 960
959 961 void
960 962 xuio_stat_wbuf_copied()
961 963 {
962 964 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
963 965 }
964 966
965 967 void
966 968 xuio_stat_wbuf_nocopy()
967 969 {
968 970 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy);
969 971 }
970 972
971 973 #ifdef _KERNEL
972 974 int
973 975 dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size)
974 976 {
975 977 dmu_buf_t **dbp;
976 978 int numbufs, i, err;
977 979 xuio_t *xuio = NULL;
978 980
979 981 /*
980 982 * NB: we could do this block-at-a-time, but it's nice
981 983 * to be reading in parallel.
982 984 */
983 985 err = dmu_buf_hold_array(os, object, uio->uio_loffset, size, TRUE, FTAG,
984 986 &numbufs, &dbp);
985 987 if (err)
986 988 return (err);
987 989
988 990 if (uio->uio_extflg == UIO_XUIO)
989 991 xuio = (xuio_t *)uio;
990 992
991 993 for (i = 0; i < numbufs; i++) {
992 994 int tocpy;
993 995 int bufoff;
994 996 dmu_buf_t *db = dbp[i];
995 997
996 998 ASSERT(size > 0);
997 999
998 1000 bufoff = uio->uio_loffset - db->db_offset;
999 1001 tocpy = (int)MIN(db->db_size - bufoff, size);
1000 1002
1001 1003 if (xuio) {
1002 1004 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db;
1003 1005 arc_buf_t *dbuf_abuf = dbi->db_buf;
1004 1006 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi);
1005 1007 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy);
1006 1008 if (!err) {
1007 1009 uio->uio_resid -= tocpy;
1008 1010 uio->uio_loffset += tocpy;
1009 1011 }
1010 1012
1011 1013 if (abuf == dbuf_abuf)
1012 1014 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy);
1013 1015 else
1014 1016 XUIOSTAT_BUMP(xuiostat_rbuf_copied);
1015 1017 } else {
1016 1018 err = uiomove((char *)db->db_data + bufoff, tocpy,
1017 1019 UIO_READ, uio);
1018 1020 }
1019 1021 if (err)
1020 1022 break;
1021 1023
1022 1024 size -= tocpy;
1023 1025 }
1024 1026 dmu_buf_rele_array(dbp, numbufs, FTAG);
1025 1027
1026 1028 return (err);
1027 1029 }
1028 1030
1029 1031 static int
1030 1032 dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx)
1031 1033 {
1032 1034 dmu_buf_t **dbp;
1033 1035 int numbufs;
1034 1036 int err = 0;
1035 1037 int i;
1036 1038
1037 1039 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size,
1038 1040 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH);
1039 1041 if (err)
1040 1042 return (err);
1041 1043
1042 1044 for (i = 0; i < numbufs; i++) {
1043 1045 int tocpy;
1044 1046 int bufoff;
1045 1047 dmu_buf_t *db = dbp[i];
1046 1048
1047 1049 ASSERT(size > 0);
1048 1050
1049 1051 bufoff = uio->uio_loffset - db->db_offset;
1050 1052 tocpy = (int)MIN(db->db_size - bufoff, size);
1051 1053
1052 1054 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1053 1055
1054 1056 if (tocpy == db->db_size)
1055 1057 dmu_buf_will_fill(db, tx);
1056 1058 else
1057 1059 dmu_buf_will_dirty(db, tx);
1058 1060
1059 1061 /*
1060 1062 * XXX uiomove could block forever (eg. nfs-backed
1061 1063 * pages). There needs to be a uiolockdown() function
1062 1064 * to lock the pages in memory, so that uiomove won't
1063 1065 * block.
1064 1066 */
1065 1067 err = uiomove((char *)db->db_data + bufoff, tocpy,
1066 1068 UIO_WRITE, uio);
1067 1069
1068 1070 if (tocpy == db->db_size)
1069 1071 dmu_buf_fill_done(db, tx);
1070 1072
1071 1073 if (err)
1072 1074 break;
1073 1075
1074 1076 size -= tocpy;
1075 1077 }
1076 1078
1077 1079 dmu_buf_rele_array(dbp, numbufs, FTAG);
1078 1080 return (err);
1079 1081 }
1080 1082
1081 1083 int
1082 1084 dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size,
1083 1085 dmu_tx_t *tx)
1084 1086 {
1085 1087 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb;
1086 1088 dnode_t *dn;
1087 1089 int err;
1088 1090
1089 1091 if (size == 0)
1090 1092 return (0);
1091 1093
1092 1094 DB_DNODE_ENTER(db);
1093 1095 dn = DB_DNODE(db);
1094 1096 err = dmu_write_uio_dnode(dn, uio, size, tx);
1095 1097 DB_DNODE_EXIT(db);
1096 1098
1097 1099 return (err);
1098 1100 }
1099 1101
1100 1102 int
1101 1103 dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size,
1102 1104 dmu_tx_t *tx)
1103 1105 {
1104 1106 dnode_t *dn;
1105 1107 int err;
1106 1108
1107 1109 if (size == 0)
1108 1110 return (0);
1109 1111
1110 1112 err = dnode_hold(os, object, FTAG, &dn);
1111 1113 if (err)
1112 1114 return (err);
1113 1115
1114 1116 err = dmu_write_uio_dnode(dn, uio, size, tx);
1115 1117
1116 1118 dnode_rele(dn, FTAG);
1117 1119
1118 1120 return (err);
1119 1121 }
1120 1122
1121 1123 int
1122 1124 dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size,
1123 1125 page_t *pp, dmu_tx_t *tx)
1124 1126 {
1125 1127 dmu_buf_t **dbp;
1126 1128 int numbufs, i;
1127 1129 int err;
1128 1130
1129 1131 if (size == 0)
1130 1132 return (0);
1131 1133
1132 1134 err = dmu_buf_hold_array(os, object, offset, size,
1133 1135 FALSE, FTAG, &numbufs, &dbp);
1134 1136 if (err)
1135 1137 return (err);
1136 1138
1137 1139 for (i = 0; i < numbufs; i++) {
1138 1140 int tocpy, copied, thiscpy;
1139 1141 int bufoff;
1140 1142 dmu_buf_t *db = dbp[i];
1141 1143 caddr_t va;
1142 1144
1143 1145 ASSERT(size > 0);
1144 1146 ASSERT3U(db->db_size, >=, PAGESIZE);
1145 1147
1146 1148 bufoff = offset - db->db_offset;
1147 1149 tocpy = (int)MIN(db->db_size - bufoff, size);
1148 1150
1149 1151 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size);
1150 1152
1151 1153 if (tocpy == db->db_size)
1152 1154 dmu_buf_will_fill(db, tx);
1153 1155 else
1154 1156 dmu_buf_will_dirty(db, tx);
1155 1157
1156 1158 for (copied = 0; copied < tocpy; copied += PAGESIZE) {
1157 1159 ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff);
1158 1160 thiscpy = MIN(PAGESIZE, tocpy - copied);
1159 1161 va = zfs_map_page(pp, S_READ);
1160 1162 bcopy(va, (char *)db->db_data + bufoff, thiscpy);
1161 1163 zfs_unmap_page(pp, va);
1162 1164 pp = pp->p_next;
1163 1165 bufoff += PAGESIZE;
1164 1166 }
1165 1167
1166 1168 if (tocpy == db->db_size)
1167 1169 dmu_buf_fill_done(db, tx);
1168 1170
1169 1171 offset += tocpy;
1170 1172 size -= tocpy;
1171 1173 }
1172 1174 dmu_buf_rele_array(dbp, numbufs, FTAG);
1173 1175 return (err);
1174 1176 }
1175 1177 #endif
1176 1178
1177 1179 /*
1178 1180 * Allocate a loaned anonymous arc buffer.
1179 1181 */
1180 1182 arc_buf_t *
1181 1183 dmu_request_arcbuf(dmu_buf_t *handle, int size)
1182 1184 {
1183 1185 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle;
1184 1186 spa_t *spa;
1185 1187
1186 1188 DB_GET_SPA(&spa, db);
1187 1189 return (arc_loan_buf(spa, size));
1188 1190 }
1189 1191
1190 1192 /*
1191 1193 * Free a loaned arc buffer.
1192 1194 */
1193 1195 void
1194 1196 dmu_return_arcbuf(arc_buf_t *buf)
1195 1197 {
1196 1198 arc_return_buf(buf, FTAG);
1197 1199 VERIFY(arc_buf_remove_ref(buf, FTAG));
1198 1200 }
1199 1201
1200 1202 /*
1201 1203 * When possible directly assign passed loaned arc buffer to a dbuf.
1202 1204 * If this is not possible copy the contents of passed arc buf via
1203 1205 * dmu_write().
1204 1206 */
1205 1207 void
1206 1208 dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf,
1207 1209 dmu_tx_t *tx)
1208 1210 {
1209 1211 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle;
1210 1212 dnode_t *dn;
1211 1213 dmu_buf_impl_t *db;
1212 1214 uint32_t blksz = (uint32_t)arc_buf_size(buf);
1213 1215 uint64_t blkid;
1214 1216
1215 1217 DB_DNODE_ENTER(dbuf);
1216 1218 dn = DB_DNODE(dbuf);
1217 1219 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1218 1220 blkid = dbuf_whichblock(dn, offset);
1219 1221 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL);
1220 1222 rw_exit(&dn->dn_struct_rwlock);
1221 1223 DB_DNODE_EXIT(dbuf);
1222 1224
1223 1225 if (offset == db->db.db_offset && blksz == db->db.db_size) {
1224 1226 dbuf_assign_arcbuf(db, buf, tx);
1225 1227 dbuf_rele(db, FTAG);
1226 1228 } else {
1227 1229 objset_t *os;
1228 1230 uint64_t object;
1229 1231
1230 1232 DB_DNODE_ENTER(dbuf);
1231 1233 dn = DB_DNODE(dbuf);
1232 1234 os = dn->dn_objset;
1233 1235 object = dn->dn_object;
1234 1236 DB_DNODE_EXIT(dbuf);
1235 1237
1236 1238 dbuf_rele(db, FTAG);
1237 1239 dmu_write(os, object, offset, blksz, buf->b_data, tx);
1238 1240 dmu_return_arcbuf(buf);
1239 1241 XUIOSTAT_BUMP(xuiostat_wbuf_copied);
1240 1242 }
1241 1243 }
1242 1244
1243 1245 typedef struct {
1244 1246 dbuf_dirty_record_t *dsa_dr;
1245 1247 dmu_sync_cb_t *dsa_done;
1246 1248 zgd_t *dsa_zgd;
1247 1249 dmu_tx_t *dsa_tx;
1248 1250 } dmu_sync_arg_t;
1249 1251
1250 1252 /* ARGSUSED */
1251 1253 static void
1252 1254 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg)
1253 1255 {
1254 1256 dmu_sync_arg_t *dsa = varg;
1255 1257 dmu_buf_t *db = dsa->dsa_zgd->zgd_db;
1256 1258 blkptr_t *bp = zio->io_bp;
1257 1259
1258 1260 if (zio->io_error == 0) {
1259 1261 if (BP_IS_HOLE(bp)) {
1260 1262 /*
1261 1263 * A block of zeros may compress to a hole, but the
1262 1264 * block size still needs to be known for replay.
1263 1265 */
1264 1266 BP_SET_LSIZE(bp, db->db_size);
1265 1267 } else {
1266 1268 ASSERT(BP_GET_LEVEL(bp) == 0);
1267 1269 bp->blk_fill = 1;
1268 1270 }
1269 1271 }
1270 1272 }
1271 1273
1272 1274 static void
1273 1275 dmu_sync_late_arrival_ready(zio_t *zio)
1274 1276 {
1275 1277 dmu_sync_ready(zio, NULL, zio->io_private);
1276 1278 }
1277 1279
1278 1280 /* ARGSUSED */
1279 1281 static void
1280 1282 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg)
1281 1283 {
1282 1284 dmu_sync_arg_t *dsa = varg;
1283 1285 dbuf_dirty_record_t *dr = dsa->dsa_dr;
1284 1286 dmu_buf_impl_t *db = dr->dr_dbuf;
1285 1287
1286 1288 mutex_enter(&db->db_mtx);
1287 1289 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC);
1288 1290 if (zio->io_error == 0) {
1289 1291 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE);
1290 1292 if (dr->dt.dl.dr_nopwrite) {
1291 1293 blkptr_t *bp = zio->io_bp;
1292 1294 blkptr_t *bp_orig = &zio->io_bp_orig;
1293 1295 uint8_t chksum = BP_GET_CHECKSUM(bp_orig);
1294 1296
1295 1297 ASSERT(BP_EQUAL(bp, bp_orig));
1296 1298 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF);
1297 1299 ASSERT(zio_checksum_table[chksum].ci_dedup);
1298 1300 }
1299 1301 dr->dt.dl.dr_overridden_by = *zio->io_bp;
1300 1302 dr->dt.dl.dr_override_state = DR_OVERRIDDEN;
1301 1303 dr->dt.dl.dr_copies = zio->io_prop.zp_copies;
1302 1304 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by))
1303 1305 BP_ZERO(&dr->dt.dl.dr_overridden_by);
1304 1306 } else {
1305 1307 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN;
1306 1308 }
1307 1309 cv_broadcast(&db->db_changed);
1308 1310 mutex_exit(&db->db_mtx);
1309 1311
1310 1312 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1311 1313
1312 1314 kmem_free(dsa, sizeof (*dsa));
1313 1315 }
1314 1316
1315 1317 static void
1316 1318 dmu_sync_late_arrival_done(zio_t *zio)
1317 1319 {
1318 1320 blkptr_t *bp = zio->io_bp;
1319 1321 dmu_sync_arg_t *dsa = zio->io_private;
1320 1322 blkptr_t *bp_orig = &zio->io_bp_orig;
1321 1323
1322 1324 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) {
1323 1325 /*
1324 1326 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE)
1325 1327 * then there is nothing to do here. Otherwise, free the
1326 1328 * newly allocated block in this txg.
1327 1329 */
1328 1330 if (zio->io_flags & ZIO_FLAG_NOPWRITE) {
1329 1331 ASSERT(BP_EQUAL(bp, bp_orig));
1330 1332 } else {
1331 1333 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig));
1332 1334 ASSERT(zio->io_bp->blk_birth == zio->io_txg);
1333 1335 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa));
1334 1336 zio_free(zio->io_spa, zio->io_txg, zio->io_bp);
1335 1337 }
1336 1338 }
1337 1339
1338 1340 dmu_tx_commit(dsa->dsa_tx);
1339 1341
1340 1342 dsa->dsa_done(dsa->dsa_zgd, zio->io_error);
1341 1343
1342 1344 kmem_free(dsa, sizeof (*dsa));
1343 1345 }
1344 1346
1345 1347 static int
1346 1348 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd,
1347 1349 zio_prop_t *zp, zbookmark_t *zb)
1348 1350 {
1349 1351 dmu_sync_arg_t *dsa;
1350 1352 dmu_tx_t *tx;
1351 1353
1352 1354 tx = dmu_tx_create(os);
1353 1355 dmu_tx_hold_space(tx, zgd->zgd_db->db_size);
1354 1356 if (dmu_tx_assign(tx, TXG_WAIT) != 0) {
1355 1357 dmu_tx_abort(tx);
1356 1358 /* Make zl_get_data do txg_waited_synced() */
1357 1359 return (SET_ERROR(EIO));
1358 1360 }
1359 1361
1360 1362 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1361 1363 dsa->dsa_dr = NULL;
1362 1364 dsa->dsa_done = done;
1363 1365 dsa->dsa_zgd = zgd;
1364 1366 dsa->dsa_tx = tx;
1365 1367
1366 1368 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp,
1367 1369 zgd->zgd_db->db_data, zgd->zgd_db->db_size, zp,
1368 1370 dmu_sync_late_arrival_ready, dmu_sync_late_arrival_done, dsa,
1369 1371 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb));
1370 1372
1371 1373 return (0);
1372 1374 }
1373 1375
1374 1376 /*
1375 1377 * Intent log support: sync the block associated with db to disk.
1376 1378 * N.B. and XXX: the caller is responsible for making sure that the
1377 1379 * data isn't changing while dmu_sync() is writing it.
1378 1380 *
1379 1381 * Return values:
1380 1382 *
1381 1383 * EEXIST: this txg has already been synced, so there's nothing to do.
1382 1384 * The caller should not log the write.
1383 1385 *
1384 1386 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do.
1385 1387 * The caller should not log the write.
1386 1388 *
1387 1389 * EALREADY: this block is already in the process of being synced.
1388 1390 * The caller should track its progress (somehow).
1389 1391 *
1390 1392 * EIO: could not do the I/O.
1391 1393 * The caller should do a txg_wait_synced().
1392 1394 *
1393 1395 * 0: the I/O has been initiated.
1394 1396 * The caller should log this blkptr in the done callback.
1395 1397 * It is possible that the I/O will fail, in which case
1396 1398 * the error will be reported to the done callback and
1397 1399 * propagated to pio from zio_done().
1398 1400 */
1399 1401 int
1400 1402 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
1401 1403 {
1402 1404 blkptr_t *bp = zgd->zgd_bp;
1403 1405 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db;
1404 1406 objset_t *os = db->db_objset;
1405 1407 dsl_dataset_t *ds = os->os_dsl_dataset;
1406 1408 dbuf_dirty_record_t *dr;
1407 1409 dmu_sync_arg_t *dsa;
1408 1410 zbookmark_t zb;
1409 1411 zio_prop_t zp;
|
↓ open down ↓ |
1355 lines elided |
↑ open up ↑ |
1410 1412 dnode_t *dn;
1411 1413
1412 1414 ASSERT(pio != NULL);
1413 1415 ASSERT(txg != 0);
1414 1416
1415 1417 SET_BOOKMARK(&zb, ds->ds_object,
1416 1418 db->db.db_object, db->db_level, db->db_blkid);
1417 1419
1418 1420 DB_DNODE_ENTER(db);
1419 1421 dn = DB_DNODE(db);
1420 - dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp);
1422 + dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp, txg);
1421 1423 DB_DNODE_EXIT(db);
1422 1424
1423 1425 /*
1424 1426 * If we're frozen (running ziltest), we always need to generate a bp.
1425 1427 */
1426 1428 if (txg > spa_freeze_txg(os->os_spa))
1427 1429 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1428 1430
1429 1431 /*
1430 1432 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf()
1431 1433 * and us. If we determine that this txg is not yet syncing,
1432 1434 * but it begins to sync a moment later, that's OK because the
1433 1435 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx.
1434 1436 */
1435 1437 mutex_enter(&db->db_mtx);
1436 1438
1437 1439 if (txg <= spa_last_synced_txg(os->os_spa)) {
1438 1440 /*
1439 1441 * This txg has already synced. There's nothing to do.
1440 1442 */
1441 1443 mutex_exit(&db->db_mtx);
1442 1444 return (SET_ERROR(EEXIST));
1443 1445 }
1444 1446
1445 1447 if (txg <= spa_syncing_txg(os->os_spa)) {
1446 1448 /*
1447 1449 * This txg is currently syncing, so we can't mess with
1448 1450 * the dirty record anymore; just write a new log block.
1449 1451 */
1450 1452 mutex_exit(&db->db_mtx);
1451 1453 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb));
1452 1454 }
1453 1455
1454 1456 dr = db->db_last_dirty;
1455 1457 while (dr && dr->dr_txg != txg)
1456 1458 dr = dr->dr_next;
1457 1459
1458 1460 if (dr == NULL) {
1459 1461 /*
1460 1462 * There's no dr for this dbuf, so it must have been freed.
1461 1463 * There's no need to log writes to freed blocks, so we're done.
1462 1464 */
1463 1465 mutex_exit(&db->db_mtx);
1464 1466 return (SET_ERROR(ENOENT));
1465 1467 }
1466 1468
1467 1469 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg);
1468 1470
1469 1471 /*
1470 1472 * Assume the on-disk data is X, the current syncing data is Y,
1471 1473 * and the current in-memory data is Z (currently in dmu_sync).
1472 1474 * X and Z are identical but Y is has been modified. Normally,
1473 1475 * when X and Z are the same we will perform a nopwrite but if Y
1474 1476 * is different we must disable nopwrite since the resulting write
1475 1477 * of Y to disk can free the block containing X. If we allowed a
1476 1478 * nopwrite to occur the block pointing to Z would reference a freed
1477 1479 * block. Since this is a rare case we simplify this by disabling
1478 1480 * nopwrite if the current dmu_sync-ing dbuf has been modified in
1479 1481 * a previous transaction.
1480 1482 */
1481 1483 if (dr->dr_next)
1482 1484 zp.zp_nopwrite = B_FALSE;
1483 1485
1484 1486 ASSERT(dr->dr_txg == txg);
1485 1487 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC ||
1486 1488 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) {
1487 1489 /*
1488 1490 * We have already issued a sync write for this buffer,
1489 1491 * or this buffer has already been synced. It could not
1490 1492 * have been dirtied since, or we would have cleared the state.
1491 1493 */
1492 1494 mutex_exit(&db->db_mtx);
1493 1495 return (SET_ERROR(EALREADY));
1494 1496 }
1495 1497
1496 1498 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN);
1497 1499 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC;
1498 1500 mutex_exit(&db->db_mtx);
1499 1501
1500 1502 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP);
1501 1503 dsa->dsa_dr = dr;
1502 1504 dsa->dsa_done = done;
1503 1505 dsa->dsa_zgd = zgd;
1504 1506 dsa->dsa_tx = NULL;
1505 1507
1506 1508 zio_nowait(arc_write(pio, os->os_spa, txg,
1507 1509 bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db), &zp,
1508 1510 dmu_sync_ready, dmu_sync_done, dsa,
1509 1511 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb));
1510 1512
1511 1513 return (0);
1512 1514 }
1513 1515
1514 1516 int
1515 1517 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs,
1516 1518 dmu_tx_t *tx)
1517 1519 {
1518 1520 dnode_t *dn;
1519 1521 int err;
1520 1522
1521 1523 err = dnode_hold(os, object, FTAG, &dn);
1522 1524 if (err)
1523 1525 return (err);
1524 1526 err = dnode_set_blksz(dn, size, ibs, tx);
1525 1527 dnode_rele(dn, FTAG);
1526 1528 return (err);
1527 1529 }
1528 1530
1529 1531 void
1530 1532 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum,
1531 1533 dmu_tx_t *tx)
1532 1534 {
1533 1535 dnode_t *dn;
1534 1536
1535 1537 /* XXX assumes dnode_hold will not get an i/o error */
1536 1538 (void) dnode_hold(os, object, FTAG, &dn);
1537 1539 ASSERT(checksum < ZIO_CHECKSUM_FUNCTIONS);
1538 1540 dn->dn_checksum = checksum;
1539 1541 dnode_setdirty(dn, tx);
1540 1542 dnode_rele(dn, FTAG);
1541 1543 }
1542 1544
1543 1545 void
1544 1546 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress,
1545 1547 dmu_tx_t *tx)
1546 1548 {
1547 1549 dnode_t *dn;
|
↓ open down ↓ |
117 lines elided |
↑ open up ↑ |
1548 1550
1549 1551 /* XXX assumes dnode_hold will not get an i/o error */
1550 1552 (void) dnode_hold(os, object, FTAG, &dn);
1551 1553 ASSERT(compress < ZIO_COMPRESS_FUNCTIONS);
1552 1554 dn->dn_compress = compress;
1553 1555 dnode_setdirty(dn, tx);
1554 1556 dnode_rele(dn, FTAG);
1555 1557 }
1556 1558
1557 1559 int zfs_mdcomp_disable = 0;
1560 +int zfs_mdcomp_lz4 = 0;
1558 1561
1559 1562 void
1560 -dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp)
1563 +dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp,
1564 + uint64_t txg)
1561 1565 {
1562 1566 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET;
1563 1567 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) ||
1564 1568 (wp & WP_SPILL));
1565 1569 enum zio_checksum checksum = os->os_checksum;
1566 1570 enum zio_compress compress = os->os_compress;
1567 1571 enum zio_checksum dedup_checksum = os->os_dedup_checksum;
1568 1572 boolean_t dedup = B_FALSE;
1569 1573 boolean_t nopwrite = B_FALSE;
1570 1574 boolean_t dedup_verify = os->os_dedup_verify;
1571 1575 int copies = os->os_copies;
1572 1576
1573 1577 /*
1574 1578 * We maintain different write policies for each of the following
|
↓ open down ↓ |
4 lines elided |
↑ open up ↑ |
1575 1579 * types of data:
1576 1580 * 1. metadata
1577 1581 * 2. preallocated blocks (i.e. level-0 blocks of a dump device)
1578 1582 * 3. all other level 0 blocks
1579 1583 */
1580 1584 if (ismd) {
1581 1585 /*
1582 1586 * XXX -- we should design a compression algorithm
1583 1587 * that specializes in arrays of bps.
1584 1588 */
1585 - compress = zfs_mdcomp_disable ? ZIO_COMPRESS_EMPTY :
1586 - ZIO_COMPRESS_LZJB;
1589 + if (zfs_mdcomp_disable)
1590 + compress = ZIO_COMPRESS_EMPTY;
1591 + else if (zfs_mdcomp_lz4 && os->os_spa != NULL) {
1592 + zfeature_info_t *feat = &spa_feature_table
1593 + [SPA_FEATURE_LZ4_COMPRESS];
1594 +
1595 + if (spa_feature_is_active(os->os_spa, feat))
1596 + compress = ZIO_COMPRESS_LZ4;
1597 + else if (spa_feature_is_enabled(os->os_spa, feat)) {
1598 + dmu_tx_t *tx;
1599 +
1600 + tx = dmu_tx_create_assigned(
1601 + spa_get_dsl(os->os_spa), txg);
1602 + spa_feature_incr(os->os_spa, feat, tx);
1603 + dmu_tx_commit(tx);
1604 + compress = ZIO_COMPRESS_LZ4;
1605 + } else
1606 + compress = ZIO_COMPRESS_LZJB;
1607 + } else
1608 + compress = ZIO_COMPRESS_LZJB;
1587 1609
1588 1610 /*
1589 1611 * Metadata always gets checksummed. If the data
1590 1612 * checksum is multi-bit correctable, and it's not a
1591 1613 * ZBT-style checksum, then it's suitable for metadata
1592 1614 * as well. Otherwise, the metadata checksum defaults
1593 1615 * to fletcher4.
1594 1616 */
1595 1617 if (zio_checksum_table[checksum].ci_correctable < 1 ||
1596 1618 zio_checksum_table[checksum].ci_eck)
1597 1619 checksum = ZIO_CHECKSUM_FLETCHER_4;
1598 1620 } else if (wp & WP_NOFILL) {
1599 1621 ASSERT(level == 0);
1600 1622
1601 1623 /*
1602 1624 * If we're writing preallocated blocks, we aren't actually
1603 1625 * writing them so don't set any policy properties. These
1604 1626 * blocks are currently only used by an external subsystem
1605 1627 * outside of zfs (i.e. dump) and not written by the zio
1606 1628 * pipeline.
1607 1629 */
1608 1630 compress = ZIO_COMPRESS_OFF;
1609 1631 checksum = ZIO_CHECKSUM_OFF;
1610 1632 } else {
1611 1633 compress = zio_compress_select(dn->dn_compress, compress);
1612 1634
1613 1635 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ?
1614 1636 zio_checksum_select(dn->dn_checksum, checksum) :
1615 1637 dedup_checksum;
1616 1638
1617 1639 /*
1618 1640 * Determine dedup setting. If we are in dmu_sync(),
1619 1641 * we won't actually dedup now because that's all
1620 1642 * done in syncing context; but we do want to use the
1621 1643 * dedup checkum. If the checksum is not strong
1622 1644 * enough to ensure unique signatures, force
1623 1645 * dedup_verify.
1624 1646 */
1625 1647 if (dedup_checksum != ZIO_CHECKSUM_OFF) {
1626 1648 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE;
1627 1649 if (!zio_checksum_table[checksum].ci_dedup)
1628 1650 dedup_verify = B_TRUE;
1629 1651 }
1630 1652
1631 1653 /*
1632 1654 * Enable nopwrite if we have a cryptographically secure
1633 1655 * checksum that has no known collisions (i.e. SHA-256)
1634 1656 * and compression is enabled. We don't enable nopwrite if
1635 1657 * dedup is enabled as the two features are mutually exclusive.
1636 1658 */
1637 1659 nopwrite = (!dedup && zio_checksum_table[checksum].ci_dedup &&
1638 1660 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled);
1639 1661 }
1640 1662
1641 1663 zp->zp_checksum = checksum;
1642 1664 zp->zp_compress = compress;
1643 1665 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type;
1644 1666 zp->zp_level = level;
1645 1667 zp->zp_copies = MIN(copies + ismd, spa_max_replication(os->os_spa));
1646 1668 zp->zp_dedup = dedup;
1647 1669 zp->zp_dedup_verify = dedup && dedup_verify;
1648 1670 zp->zp_nopwrite = nopwrite;
1649 1671 }
1650 1672
1651 1673 int
1652 1674 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off)
1653 1675 {
1654 1676 dnode_t *dn;
1655 1677 int i, err;
1656 1678
1657 1679 err = dnode_hold(os, object, FTAG, &dn);
1658 1680 if (err)
1659 1681 return (err);
1660 1682 /*
1661 1683 * Sync any current changes before
1662 1684 * we go trundling through the block pointers.
1663 1685 */
1664 1686 for (i = 0; i < TXG_SIZE; i++) {
1665 1687 if (list_link_active(&dn->dn_dirty_link[i]))
1666 1688 break;
1667 1689 }
1668 1690 if (i != TXG_SIZE) {
1669 1691 dnode_rele(dn, FTAG);
1670 1692 txg_wait_synced(dmu_objset_pool(os), 0);
1671 1693 err = dnode_hold(os, object, FTAG, &dn);
1672 1694 if (err)
1673 1695 return (err);
1674 1696 }
1675 1697
1676 1698 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0);
1677 1699 dnode_rele(dn, FTAG);
1678 1700
1679 1701 return (err);
1680 1702 }
1681 1703
1682 1704 void
1683 1705 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi)
1684 1706 {
1685 1707 dnode_phys_t *dnp;
1686 1708
1687 1709 rw_enter(&dn->dn_struct_rwlock, RW_READER);
1688 1710 mutex_enter(&dn->dn_mtx);
1689 1711
1690 1712 dnp = dn->dn_phys;
1691 1713
1692 1714 doi->doi_data_block_size = dn->dn_datablksz;
1693 1715 doi->doi_metadata_block_size = dn->dn_indblkshift ?
1694 1716 1ULL << dn->dn_indblkshift : 0;
1695 1717 doi->doi_type = dn->dn_type;
1696 1718 doi->doi_bonus_type = dn->dn_bonustype;
1697 1719 doi->doi_bonus_size = dn->dn_bonuslen;
1698 1720 doi->doi_indirection = dn->dn_nlevels;
1699 1721 doi->doi_checksum = dn->dn_checksum;
1700 1722 doi->doi_compress = dn->dn_compress;
1701 1723 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9;
1702 1724 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz;
1703 1725 doi->doi_fill_count = 0;
1704 1726 for (int i = 0; i < dnp->dn_nblkptr; i++)
1705 1727 doi->doi_fill_count += dnp->dn_blkptr[i].blk_fill;
1706 1728
1707 1729 mutex_exit(&dn->dn_mtx);
1708 1730 rw_exit(&dn->dn_struct_rwlock);
1709 1731 }
1710 1732
1711 1733 /*
1712 1734 * Get information on a DMU object.
1713 1735 * If doi is NULL, just indicates whether the object exists.
1714 1736 */
1715 1737 int
1716 1738 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi)
1717 1739 {
1718 1740 dnode_t *dn;
1719 1741 int err = dnode_hold(os, object, FTAG, &dn);
1720 1742
1721 1743 if (err)
1722 1744 return (err);
1723 1745
1724 1746 if (doi != NULL)
1725 1747 dmu_object_info_from_dnode(dn, doi);
1726 1748
1727 1749 dnode_rele(dn, FTAG);
1728 1750 return (0);
1729 1751 }
1730 1752
1731 1753 /*
1732 1754 * As above, but faster; can be used when you have a held dbuf in hand.
1733 1755 */
1734 1756 void
1735 1757 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi)
1736 1758 {
1737 1759 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1738 1760
1739 1761 DB_DNODE_ENTER(db);
1740 1762 dmu_object_info_from_dnode(DB_DNODE(db), doi);
1741 1763 DB_DNODE_EXIT(db);
1742 1764 }
1743 1765
1744 1766 /*
1745 1767 * Faster still when you only care about the size.
1746 1768 * This is specifically optimized for zfs_getattr().
1747 1769 */
1748 1770 void
1749 1771 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize,
1750 1772 u_longlong_t *nblk512)
1751 1773 {
1752 1774 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake;
1753 1775 dnode_t *dn;
1754 1776
1755 1777 DB_DNODE_ENTER(db);
1756 1778 dn = DB_DNODE(db);
1757 1779
1758 1780 *blksize = dn->dn_datablksz;
1759 1781 /* add 1 for dnode space */
1760 1782 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >>
1761 1783 SPA_MINBLOCKSHIFT) + 1;
1762 1784 DB_DNODE_EXIT(db);
1763 1785 }
1764 1786
1765 1787 void
1766 1788 byteswap_uint64_array(void *vbuf, size_t size)
1767 1789 {
1768 1790 uint64_t *buf = vbuf;
1769 1791 size_t count = size >> 3;
1770 1792 int i;
1771 1793
1772 1794 ASSERT((size & 7) == 0);
1773 1795
1774 1796 for (i = 0; i < count; i++)
1775 1797 buf[i] = BSWAP_64(buf[i]);
1776 1798 }
1777 1799
1778 1800 void
1779 1801 byteswap_uint32_array(void *vbuf, size_t size)
1780 1802 {
1781 1803 uint32_t *buf = vbuf;
1782 1804 size_t count = size >> 2;
1783 1805 int i;
1784 1806
1785 1807 ASSERT((size & 3) == 0);
1786 1808
1787 1809 for (i = 0; i < count; i++)
1788 1810 buf[i] = BSWAP_32(buf[i]);
1789 1811 }
1790 1812
1791 1813 void
1792 1814 byteswap_uint16_array(void *vbuf, size_t size)
1793 1815 {
1794 1816 uint16_t *buf = vbuf;
1795 1817 size_t count = size >> 1;
1796 1818 int i;
1797 1819
1798 1820 ASSERT((size & 1) == 0);
1799 1821
1800 1822 for (i = 0; i < count; i++)
1801 1823 buf[i] = BSWAP_16(buf[i]);
1802 1824 }
1803 1825
1804 1826 /* ARGSUSED */
1805 1827 void
1806 1828 byteswap_uint8_array(void *vbuf, size_t size)
1807 1829 {
1808 1830 }
1809 1831
1810 1832 void
1811 1833 dmu_init(void)
1812 1834 {
1813 1835 zfs_dbgmsg_init();
1814 1836 sa_cache_init();
1815 1837 xuio_stat_init();
1816 1838 dmu_objset_init();
1817 1839 dnode_init();
1818 1840 dbuf_init();
1819 1841 zfetch_init();
1820 1842 l2arc_init();
1821 1843 arc_init();
1822 1844 }
1823 1845
1824 1846 void
1825 1847 dmu_fini(void)
1826 1848 {
1827 1849 arc_fini();
1828 1850 l2arc_fini();
1829 1851 zfetch_fini();
1830 1852 dbuf_fini();
1831 1853 dnode_fini();
1832 1854 dmu_objset_fini();
1833 1855 xuio_stat_fini();
1834 1856 sa_cache_fini();
1835 1857 zfs_dbgmsg_fini();
1836 1858 }
|
↓ open down ↓ |
240 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX