1/* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21/* 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23 * Copyright (c) 2013 by Delphix. All rights reserved. 24 */ 25/* Copyright (c) 2013 by Saso Kiselkov. All rights reserved. */ 26/* Copyright (c) 2013, Joyent, Inc. All rights reserved. */ 27 28#include <sys/dmu.h> 29#include <sys/dmu_impl.h> 30#include <sys/dmu_tx.h> 31#include <sys/dbuf.h> 32#include <sys/dnode.h> 33#include <sys/zfs_context.h> 34#include <sys/dmu_objset.h> 35#include <sys/dmu_traverse.h> 36#include <sys/dsl_dataset.h> 37#include <sys/dsl_dir.h> 38#include <sys/dsl_pool.h> 39#include <sys/dsl_synctask.h> 40#include <sys/dsl_prop.h> 41#include <sys/dmu_zfetch.h> 42#include <sys/zfs_ioctl.h> 43#include <sys/zap.h> 44#include <sys/zio_checksum.h> 45#include <sys/zio_compress.h> 46#include <sys/sa.h> 47#ifdef _KERNEL 48#include <sys/zfs_znode.h> 49#endif 50 51/* 52 * Enable/disable nopwrite feature. 53 */ 54int zfs_nopwrite_enabled = 1; 55SYSCTL_DECL(_vfs_zfs); 56TUNABLE_INT("vfs.zfs.nopwrite_enabled", &zfs_nopwrite_enabled); 57SYSCTL_INT(_vfs_zfs, OID_AUTO, nopwrite_enabled, CTLFLAG_RDTUN, 58 &zfs_nopwrite_enabled, 0, "Enable nopwrite feature"); 59 60const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = { 61 { DMU_BSWAP_UINT8, TRUE, "unallocated" }, 62 { DMU_BSWAP_ZAP, TRUE, "object directory" }, 63 { DMU_BSWAP_UINT64, TRUE, "object array" }, 64 { DMU_BSWAP_UINT8, TRUE, "packed nvlist" }, 65 { DMU_BSWAP_UINT64, TRUE, "packed nvlist size" }, 66 { DMU_BSWAP_UINT64, TRUE, "bpobj" }, 67 { DMU_BSWAP_UINT64, TRUE, "bpobj header" }, 68 { DMU_BSWAP_UINT64, TRUE, "SPA space map header" }, 69 { DMU_BSWAP_UINT64, TRUE, "SPA space map" }, 70 { DMU_BSWAP_UINT64, TRUE, "ZIL intent log" }, 71 { DMU_BSWAP_DNODE, TRUE, "DMU dnode" }, 72 { DMU_BSWAP_OBJSET, TRUE, "DMU objset" }, 73 { DMU_BSWAP_UINT64, TRUE, "DSL directory" }, 74 { DMU_BSWAP_ZAP, TRUE, "DSL directory child map"}, 75 { DMU_BSWAP_ZAP, TRUE, "DSL dataset snap map" }, 76 { DMU_BSWAP_ZAP, TRUE, "DSL props" }, 77 { DMU_BSWAP_UINT64, TRUE, "DSL dataset" }, 78 { DMU_BSWAP_ZNODE, TRUE, "ZFS znode" }, 79 { DMU_BSWAP_OLDACL, TRUE, "ZFS V0 ACL" }, 80 { DMU_BSWAP_UINT8, FALSE, "ZFS plain file" }, 81 { DMU_BSWAP_ZAP, TRUE, "ZFS directory" }, 82 { DMU_BSWAP_ZAP, TRUE, "ZFS master node" }, 83 { DMU_BSWAP_ZAP, TRUE, "ZFS delete queue" }, 84 { DMU_BSWAP_UINT8, FALSE, "zvol object" }, 85 { DMU_BSWAP_ZAP, TRUE, "zvol prop" }, 86 { DMU_BSWAP_UINT8, FALSE, "other uint8[]" }, 87 { DMU_BSWAP_UINT64, FALSE, "other uint64[]" }, 88 { DMU_BSWAP_ZAP, TRUE, "other ZAP" }, 89 { DMU_BSWAP_ZAP, TRUE, "persistent error log" }, 90 { DMU_BSWAP_UINT8, TRUE, "SPA history" }, 91 { DMU_BSWAP_UINT64, TRUE, "SPA history offsets" }, 92 { DMU_BSWAP_ZAP, TRUE, "Pool properties" }, 93 { DMU_BSWAP_ZAP, TRUE, "DSL permissions" }, 94 { DMU_BSWAP_ACL, TRUE, "ZFS ACL" }, 95 { DMU_BSWAP_UINT8, TRUE, "ZFS SYSACL" }, 96 { DMU_BSWAP_UINT8, TRUE, "FUID table" }, 97 { DMU_BSWAP_UINT64, TRUE, "FUID table size" }, 98 { DMU_BSWAP_ZAP, TRUE, "DSL dataset next clones"}, 99 { DMU_BSWAP_ZAP, TRUE, "scan work queue" }, 100 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group used" }, 101 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group quota" }, 102 { DMU_BSWAP_ZAP, TRUE, "snapshot refcount tags"}, 103 { DMU_BSWAP_ZAP, TRUE, "DDT ZAP algorithm" }, 104 { DMU_BSWAP_ZAP, TRUE, "DDT statistics" }, 105 { DMU_BSWAP_UINT8, TRUE, "System attributes" }, 106 { DMU_BSWAP_ZAP, TRUE, "SA master node" }, 107 { DMU_BSWAP_ZAP, TRUE, "SA attr registration" }, 108 { DMU_BSWAP_ZAP, TRUE, "SA attr layouts" }, 109 { DMU_BSWAP_ZAP, TRUE, "scan translations" }, 110 { DMU_BSWAP_UINT8, FALSE, "deduplicated block" }, 111 { DMU_BSWAP_ZAP, TRUE, "DSL deadlist map" }, 112 { DMU_BSWAP_UINT64, TRUE, "DSL deadlist map hdr" }, 113 { DMU_BSWAP_ZAP, TRUE, "DSL dir clones" }, 114 { DMU_BSWAP_UINT64, TRUE, "bpobj subobj" } 115}; 116 117const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = { 118 { byteswap_uint8_array, "uint8" }, 119 { byteswap_uint16_array, "uint16" }, 120 { byteswap_uint32_array, "uint32" }, 121 { byteswap_uint64_array, "uint64" }, 122 { zap_byteswap, "zap" }, 123 { dnode_buf_byteswap, "dnode" }, 124 { dmu_objset_byteswap, "objset" }, 125 { zfs_znode_byteswap, "znode" }, 126 { zfs_oldacl_byteswap, "oldacl" }, 127 { zfs_acl_byteswap, "acl" } 128}; 129 130int 131dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset, 132 void *tag, dmu_buf_t **dbp, int flags) 133{ 134 dnode_t *dn; 135 uint64_t blkid; 136 dmu_buf_impl_t *db; 137 int err; 138 int db_flags = DB_RF_CANFAIL; 139 140 if (flags & DMU_READ_NO_PREFETCH) 141 db_flags |= DB_RF_NOPREFETCH; 142 143 err = dnode_hold(os, object, FTAG, &dn); 144 if (err) 145 return (err); 146 blkid = dbuf_whichblock(dn, offset); 147 rw_enter(&dn->dn_struct_rwlock, RW_READER); 148 db = dbuf_hold(dn, blkid, tag); 149 rw_exit(&dn->dn_struct_rwlock); 150 if (db == NULL) { 151 err = SET_ERROR(EIO); 152 } else { 153 err = dbuf_read(db, NULL, db_flags); 154 if (err) { 155 dbuf_rele(db, tag); 156 db = NULL; 157 } 158 } 159 160 dnode_rele(dn, FTAG); 161 *dbp = &db->db; /* NULL db plus first field offset is NULL */ 162 return (err); 163} 164 165int 166dmu_bonus_max(void) 167{ 168 return (DN_MAX_BONUSLEN); 169} 170 171int 172dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx) 173{ 174 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 175 dnode_t *dn; 176 int error; 177 178 DB_DNODE_ENTER(db); 179 dn = DB_DNODE(db); 180 181 if (dn->dn_bonus != db) { 182 error = SET_ERROR(EINVAL); 183 } else if (newsize < 0 || newsize > db_fake->db_size) { 184 error = SET_ERROR(EINVAL); 185 } else { 186 dnode_setbonuslen(dn, newsize, tx); 187 error = 0; 188 } 189 190 DB_DNODE_EXIT(db); 191 return (error); 192} 193 194int 195dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx) 196{ 197 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 198 dnode_t *dn; 199 int error; 200 201 DB_DNODE_ENTER(db); 202 dn = DB_DNODE(db); 203 204 if (!DMU_OT_IS_VALID(type)) { 205 error = SET_ERROR(EINVAL); 206 } else if (dn->dn_bonus != db) { 207 error = SET_ERROR(EINVAL); 208 } else { 209 dnode_setbonus_type(dn, type, tx); 210 error = 0; 211 } 212 213 DB_DNODE_EXIT(db); 214 return (error); 215} 216 217dmu_object_type_t 218dmu_get_bonustype(dmu_buf_t *db_fake) 219{ 220 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 221 dnode_t *dn; 222 dmu_object_type_t type; 223 224 DB_DNODE_ENTER(db); 225 dn = DB_DNODE(db); 226 type = dn->dn_bonustype; 227 DB_DNODE_EXIT(db); 228 229 return (type); 230} 231 232int 233dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx) 234{ 235 dnode_t *dn; 236 int error; 237 238 error = dnode_hold(os, object, FTAG, &dn); 239 dbuf_rm_spill(dn, tx); 240 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 241 dnode_rm_spill(dn, tx); 242 rw_exit(&dn->dn_struct_rwlock); 243 dnode_rele(dn, FTAG); 244 return (error); 245} 246 247/* 248 * returns ENOENT, EIO, or 0. 249 */ 250int 251dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp) 252{ 253 dnode_t *dn; 254 dmu_buf_impl_t *db; 255 int error; 256 257 error = dnode_hold(os, object, FTAG, &dn); 258 if (error) 259 return (error); 260 261 rw_enter(&dn->dn_struct_rwlock, RW_READER); 262 if (dn->dn_bonus == NULL) { 263 rw_exit(&dn->dn_struct_rwlock); 264 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 265 if (dn->dn_bonus == NULL) 266 dbuf_create_bonus(dn); 267 } 268 db = dn->dn_bonus; 269 270 /* as long as the bonus buf is held, the dnode will be held */ 271 if (refcount_add(&db->db_holds, tag) == 1) { 272 VERIFY(dnode_add_ref(dn, db)); 273 (void) atomic_inc_32_nv(&dn->dn_dbufs_count); 274 } 275 276 /* 277 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's 278 * hold and incrementing the dbuf count to ensure that dnode_move() sees 279 * a dnode hold for every dbuf. 280 */ 281 rw_exit(&dn->dn_struct_rwlock); 282 283 dnode_rele(dn, FTAG); 284 285 VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH)); 286 287 *dbp = &db->db; 288 return (0); 289} 290 291/* 292 * returns ENOENT, EIO, or 0. 293 * 294 * This interface will allocate a blank spill dbuf when a spill blk 295 * doesn't already exist on the dnode. 296 * 297 * if you only want to find an already existing spill db, then 298 * dmu_spill_hold_existing() should be used. 299 */ 300int 301dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp) 302{ 303 dmu_buf_impl_t *db = NULL; 304 int err; 305 306 if ((flags & DB_RF_HAVESTRUCT) == 0) 307 rw_enter(&dn->dn_struct_rwlock, RW_READER); 308 309 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag); 310 311 if ((flags & DB_RF_HAVESTRUCT) == 0) 312 rw_exit(&dn->dn_struct_rwlock); 313 314 ASSERT(db != NULL); 315 err = dbuf_read(db, NULL, flags); 316 if (err == 0) 317 *dbp = &db->db; 318 else 319 dbuf_rele(db, tag); 320 return (err); 321} 322 323int 324dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp) 325{ 326 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus; 327 dnode_t *dn; 328 int err; 329 330 DB_DNODE_ENTER(db); 331 dn = DB_DNODE(db); 332 333 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) { 334 err = SET_ERROR(EINVAL); 335 } else { 336 rw_enter(&dn->dn_struct_rwlock, RW_READER); 337 338 if (!dn->dn_have_spill) { 339 err = SET_ERROR(ENOENT); 340 } else { 341 err = dmu_spill_hold_by_dnode(dn, 342 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp); 343 } 344 345 rw_exit(&dn->dn_struct_rwlock); 346 } 347 348 DB_DNODE_EXIT(db); 349 return (err); 350} 351 352int 353dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp) 354{ 355 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus; 356 dnode_t *dn; 357 int err; 358 359 DB_DNODE_ENTER(db); 360 dn = DB_DNODE(db); 361 err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp); 362 DB_DNODE_EXIT(db); 363 364 return (err); 365} 366 367/* 368 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces 369 * to take a held dnode rather than <os, object> -- the lookup is wasteful, 370 * and can induce severe lock contention when writing to several files 371 * whose dnodes are in the same block. 372 */ 373static int 374dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length, 375 int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags) 376{ 377 dsl_pool_t *dp = NULL; 378 dmu_buf_t **dbp; 379 uint64_t blkid, nblks, i; 380 uint32_t dbuf_flags; 381 int err; 382 zio_t *zio; 383 hrtime_t start; 384 385 ASSERT(length <= DMU_MAX_ACCESS); 386 387 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT; 388 if (flags & DMU_READ_NO_PREFETCH || length > zfetch_array_rd_sz) 389 dbuf_flags |= DB_RF_NOPREFETCH; 390 391 rw_enter(&dn->dn_struct_rwlock, RW_READER); 392 if (dn->dn_datablkshift) { 393 int blkshift = dn->dn_datablkshift; 394 nblks = (P2ROUNDUP(offset+length, 1ULL<<blkshift) - 395 P2ALIGN(offset, 1ULL<<blkshift)) >> blkshift; 396 } else { 397 if (offset + length > dn->dn_datablksz) { 398 zfs_panic_recover("zfs: accessing past end of object " 399 "%llx/%llx (size=%u access=%llu+%llu)", 400 (longlong_t)dn->dn_objset-> 401 os_dsl_dataset->ds_object, 402 (longlong_t)dn->dn_object, dn->dn_datablksz, 403 (longlong_t)offset, (longlong_t)length); 404 rw_exit(&dn->dn_struct_rwlock); 405 return (SET_ERROR(EIO)); 406 } 407 nblks = 1; 408 } 409 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP); 410 411 if (dn->dn_objset->os_dsl_dataset) 412 dp = dn->dn_objset->os_dsl_dataset->ds_dir->dd_pool; 413 start = gethrtime(); 414 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL); 415 blkid = dbuf_whichblock(dn, offset); 416 for (i = 0; i < nblks; i++) { 417 dmu_buf_impl_t *db = dbuf_hold(dn, blkid+i, tag); 418 if (db == NULL) { 419 rw_exit(&dn->dn_struct_rwlock); 420 dmu_buf_rele_array(dbp, nblks, tag); 421 zio_nowait(zio); 422 return (SET_ERROR(EIO)); 423 } 424 /* initiate async i/o */ 425 if (read) 426 (void) dbuf_read(db, zio, dbuf_flags); 427#ifdef _KERNEL 428 else 429 curthread->td_ru.ru_oublock++; 430#endif 431 dbp[i] = &db->db; 432 } 433 rw_exit(&dn->dn_struct_rwlock); 434 435 /* wait for async i/o */ 436 err = zio_wait(zio); 437 /* track read overhead when we are in sync context */ 438 if (dp && dsl_pool_sync_context(dp)) 439 dp->dp_read_overhead += gethrtime() - start; 440 if (err) { 441 dmu_buf_rele_array(dbp, nblks, tag); 442 return (err); 443 } 444 445 /* wait for other io to complete */ 446 if (read) { 447 for (i = 0; i < nblks; i++) { 448 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i]; 449 mutex_enter(&db->db_mtx); 450 while (db->db_state == DB_READ || 451 db->db_state == DB_FILL) 452 cv_wait(&db->db_changed, &db->db_mtx); 453 if (db->db_state == DB_UNCACHED) 454 err = SET_ERROR(EIO); 455 mutex_exit(&db->db_mtx); 456 if (err) { 457 dmu_buf_rele_array(dbp, nblks, tag); 458 return (err); 459 } 460 } 461 } 462 463 *numbufsp = nblks; 464 *dbpp = dbp; 465 return (0); 466} 467 468static int 469dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset, 470 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp) 471{ 472 dnode_t *dn; 473 int err; 474 475 err = dnode_hold(os, object, FTAG, &dn); 476 if (err) 477 return (err); 478 479 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag, 480 numbufsp, dbpp, DMU_READ_PREFETCH); 481 482 dnode_rele(dn, FTAG); 483 484 return (err); 485} 486 487int 488dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset, 489 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp) 490{ 491 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 492 dnode_t *dn; 493 int err; 494 495 DB_DNODE_ENTER(db); 496 dn = DB_DNODE(db); 497 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag, 498 numbufsp, dbpp, DMU_READ_PREFETCH); 499 DB_DNODE_EXIT(db); 500 501 return (err); 502} 503 504void 505dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag) 506{ 507 int i; 508 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake; 509 510 if (numbufs == 0) 511 return; 512 513 for (i = 0; i < numbufs; i++) { 514 if (dbp[i]) 515 dbuf_rele(dbp[i], tag); 516 } 517 518 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs); 519} 520 521void 522dmu_prefetch(objset_t *os, uint64_t object, uint64_t offset, uint64_t len) 523{ 524 dnode_t *dn; 525 uint64_t blkid; 526 int nblks, i, err; 527 528 if (zfs_prefetch_disable) 529 return; 530 531 if (len == 0) { /* they're interested in the bonus buffer */ 532 dn = DMU_META_DNODE(os); 533 534 if (object == 0 || object >= DN_MAX_OBJECT) 535 return; 536 537 rw_enter(&dn->dn_struct_rwlock, RW_READER); 538 blkid = dbuf_whichblock(dn, object * sizeof (dnode_phys_t)); 539 dbuf_prefetch(dn, blkid); 540 rw_exit(&dn->dn_struct_rwlock); 541 return; 542 } 543 544 /* 545 * XXX - Note, if the dnode for the requested object is not 546 * already cached, we will do a *synchronous* read in the 547 * dnode_hold() call. The same is true for any indirects. 548 */ 549 err = dnode_hold(os, object, FTAG, &dn); 550 if (err != 0) 551 return; 552 553 rw_enter(&dn->dn_struct_rwlock, RW_READER); 554 if (dn->dn_datablkshift) { 555 int blkshift = dn->dn_datablkshift; 556 nblks = (P2ROUNDUP(offset+len, 1<<blkshift) - 557 P2ALIGN(offset, 1<<blkshift)) >> blkshift; 558 } else { 559 nblks = (offset < dn->dn_datablksz); 560 } 561 562 if (nblks != 0) { 563 blkid = dbuf_whichblock(dn, offset); 564 for (i = 0; i < nblks; i++) 565 dbuf_prefetch(dn, blkid+i); 566 } 567 568 rw_exit(&dn->dn_struct_rwlock); 569 570 dnode_rele(dn, FTAG); 571} 572 573/* 574 * Get the next "chunk" of file data to free. We traverse the file from 575 * the end so that the file gets shorter over time (if we crashes in the 576 * middle, this will leave us in a better state). We find allocated file 577 * data by simply searching the allocated level 1 indirects. 578 * 579 * On input, *start should be the first offset that does not need to be 580 * freed (e.g. "offset + length"). On return, *start will be the first 581 * offset that should be freed. 582 */ 583static int 584get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum) 585{ 586 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1); 587 /* bytes of data covered by a level-1 indirect block */ 588 uint64_t iblkrange = 589 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT); 590 591 ASSERT3U(minimum, <=, *start); 592 593 if (*start - minimum <= iblkrange * maxblks) { 594 *start = minimum; 595 return (0); 596 } 597 ASSERT(ISP2(iblkrange)); 598 599 for (uint64_t blks = 0; *start > minimum && blks < maxblks; blks++) { 600 int err; 601 602 /* 603 * dnode_next_offset(BACKWARDS) will find an allocated L1 604 * indirect block at or before the input offset. We must 605 * decrement *start so that it is at the end of the region 606 * to search. 607 */ 608 (*start)--; 609 err = dnode_next_offset(dn, 610 DNODE_FIND_BACKWARDS, start, 2, 1, 0); 611 612 /* if there are no indirect blocks before start, we are done */ 613 if (err == ESRCH) { 614 *start = minimum; 615 break; 616 } else if (err != 0) { 617 return (err); 618 } 619 620 /* set start to the beginning of this L1 indirect */ 621 *start = P2ALIGN(*start, iblkrange); 622 } 623 if (*start < minimum) 624 *start = minimum; 625 return (0); 626} 627 628static int 629dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset, 630 uint64_t length) 631{ 632 uint64_t object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz; 633 int err; 634 635 if (offset >= object_size) 636 return (0); 637 638 if (length == DMU_OBJECT_END || offset + length > object_size) 639 length = object_size - offset; 640 641 while (length != 0) { 642 uint64_t chunk_end, chunk_begin; 643 644 chunk_end = chunk_begin = offset + length; 645 646 /* move chunk_begin backwards to the beginning of this chunk */ 647 err = get_next_chunk(dn, &chunk_begin, offset); 648 if (err) 649 return (err); 650 ASSERT3U(chunk_begin, >=, offset); 651 ASSERT3U(chunk_begin, <=, chunk_end); 652 653 dmu_tx_t *tx = dmu_tx_create(os); 654 dmu_tx_hold_free(tx, dn->dn_object, 655 chunk_begin, chunk_end - chunk_begin); 656 err = dmu_tx_assign(tx, TXG_WAIT); 657 if (err) { 658 dmu_tx_abort(tx); 659 return (err); 660 } 661 dnode_free_range(dn, chunk_begin, chunk_end - chunk_begin, tx); 662 dmu_tx_commit(tx); 663 664 length -= chunk_end - chunk_begin; 665 } 666 return (0); 667} 668 669int 670dmu_free_long_range(objset_t *os, uint64_t object, 671 uint64_t offset, uint64_t length) 672{ 673 dnode_t *dn; 674 int err; 675 676 err = dnode_hold(os, object, FTAG, &dn); 677 if (err != 0) 678 return (err); 679 err = dmu_free_long_range_impl(os, dn, offset, length); 680 681 /* 682 * It is important to zero out the maxblkid when freeing the entire 683 * file, so that (a) subsequent calls to dmu_free_long_range_impl() 684 * will take the fast path, and (b) dnode_reallocate() can verify 685 * that the entire file has been freed. 686 */ 687 if (offset == 0 && length == DMU_OBJECT_END) 688 dn->dn_maxblkid = 0; 689 690 dnode_rele(dn, FTAG); 691 return (err); 692} 693 694int 695dmu_free_long_object(objset_t *os, uint64_t object) 696{ 697 dmu_tx_t *tx; 698 int err; 699 700 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END); 701 if (err != 0) 702 return (err); 703 704 tx = dmu_tx_create(os); 705 dmu_tx_hold_bonus(tx, object); 706 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END); 707 err = dmu_tx_assign(tx, TXG_WAIT); 708 if (err == 0) { 709 err = dmu_object_free(os, object, tx); 710 dmu_tx_commit(tx); 711 } else { 712 dmu_tx_abort(tx); 713 } 714 715 return (err); 716} 717 718int 719dmu_free_range(objset_t *os, uint64_t object, uint64_t offset, 720 uint64_t size, dmu_tx_t *tx) 721{ 722 dnode_t *dn; 723 int err = dnode_hold(os, object, FTAG, &dn); 724 if (err) 725 return (err); 726 ASSERT(offset < UINT64_MAX); 727 ASSERT(size == -1ULL || size <= UINT64_MAX - offset); 728 dnode_free_range(dn, offset, size, tx); 729 dnode_rele(dn, FTAG); 730 return (0); 731} 732 733int 734dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 735 void *buf, uint32_t flags) 736{ 737 dnode_t *dn; 738 dmu_buf_t **dbp; 739 int numbufs, err; 740 741 err = dnode_hold(os, object, FTAG, &dn); 742 if (err) 743 return (err); 744 745 /* 746 * Deal with odd block sizes, where there can't be data past the first 747 * block. If we ever do the tail block optimization, we will need to 748 * handle that here as well. 749 */ 750 if (dn->dn_maxblkid == 0) { 751 int newsz = offset > dn->dn_datablksz ? 0 : 752 MIN(size, dn->dn_datablksz - offset); 753 bzero((char *)buf + newsz, size - newsz); 754 size = newsz; 755 } 756 757 while (size > 0) { 758 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2); 759 int i; 760 761 /* 762 * NB: we could do this block-at-a-time, but it's nice 763 * to be reading in parallel. 764 */ 765 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen, 766 TRUE, FTAG, &numbufs, &dbp, flags); 767 if (err) 768 break; 769 770 for (i = 0; i < numbufs; i++) { 771 int tocpy; 772 int bufoff; 773 dmu_buf_t *db = dbp[i]; 774 775 ASSERT(size > 0); 776 777 bufoff = offset - db->db_offset; 778 tocpy = (int)MIN(db->db_size - bufoff, size); 779 780 bcopy((char *)db->db_data + bufoff, buf, tocpy); 781 782 offset += tocpy; 783 size -= tocpy; 784 buf = (char *)buf + tocpy; 785 } 786 dmu_buf_rele_array(dbp, numbufs, FTAG); 787 } 788 dnode_rele(dn, FTAG); 789 return (err); 790} 791 792void 793dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 794 const void *buf, dmu_tx_t *tx) 795{ 796 dmu_buf_t **dbp; 797 int numbufs, i; 798 799 if (size == 0) 800 return; 801 802 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size, 803 FALSE, FTAG, &numbufs, &dbp)); 804 805 for (i = 0; i < numbufs; i++) { 806 int tocpy; 807 int bufoff; 808 dmu_buf_t *db = dbp[i]; 809 810 ASSERT(size > 0); 811 812 bufoff = offset - db->db_offset; 813 tocpy = (int)MIN(db->db_size - bufoff, size); 814 815 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 816 817 if (tocpy == db->db_size) 818 dmu_buf_will_fill(db, tx); 819 else 820 dmu_buf_will_dirty(db, tx); 821 822 bcopy(buf, (char *)db->db_data + bufoff, tocpy); 823 824 if (tocpy == db->db_size) 825 dmu_buf_fill_done(db, tx); 826 827 offset += tocpy; 828 size -= tocpy; 829 buf = (char *)buf + tocpy; 830 } 831 dmu_buf_rele_array(dbp, numbufs, FTAG); 832} 833 834void 835dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 836 dmu_tx_t *tx) 837{ 838 dmu_buf_t **dbp; 839 int numbufs, i; 840 841 if (size == 0) 842 return; 843 844 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size, 845 FALSE, FTAG, &numbufs, &dbp)); 846 847 for (i = 0; i < numbufs; i++) { 848 dmu_buf_t *db = dbp[i]; 849 850 dmu_buf_will_not_fill(db, tx); 851 } 852 dmu_buf_rele_array(dbp, numbufs, FTAG); 853} 854 855/* 856 * DMU support for xuio 857 */ 858kstat_t *xuio_ksp = NULL; 859 860int 861dmu_xuio_init(xuio_t *xuio, int nblk) 862{ 863 dmu_xuio_t *priv; 864 uio_t *uio = &xuio->xu_uio; 865 866 uio->uio_iovcnt = nblk; 867 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP); 868 869 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP); 870 priv->cnt = nblk; 871 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP); 872 priv->iovp = uio->uio_iov; 873 XUIO_XUZC_PRIV(xuio) = priv; 874 875 if (XUIO_XUZC_RW(xuio) == UIO_READ) 876 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk); 877 else 878 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk); 879 880 return (0); 881} 882 883void 884dmu_xuio_fini(xuio_t *xuio) 885{ 886 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 887 int nblk = priv->cnt; 888 889 kmem_free(priv->iovp, nblk * sizeof (iovec_t)); 890 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *)); 891 kmem_free(priv, sizeof (dmu_xuio_t)); 892 893 if (XUIO_XUZC_RW(xuio) == UIO_READ) 894 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk); 895 else 896 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk); 897} 898 899/* 900 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf } 901 * and increase priv->next by 1. 902 */ 903int 904dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n) 905{ 906 struct iovec *iov; 907 uio_t *uio = &xuio->xu_uio; 908 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 909 int i = priv->next++; 910 911 ASSERT(i < priv->cnt); 912 ASSERT(off + n <= arc_buf_size(abuf)); 913 iov = uio->uio_iov + i; 914 iov->iov_base = (char *)abuf->b_data + off; 915 iov->iov_len = n; 916 priv->bufs[i] = abuf; 917 return (0); 918} 919 920int 921dmu_xuio_cnt(xuio_t *xuio) 922{ 923 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 924 return (priv->cnt); 925} 926 927arc_buf_t * 928dmu_xuio_arcbuf(xuio_t *xuio, int i) 929{ 930 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 931 932 ASSERT(i < priv->cnt); 933 return (priv->bufs[i]); 934} 935 936void 937dmu_xuio_clear(xuio_t *xuio, int i) 938{ 939 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 940 941 ASSERT(i < priv->cnt); 942 priv->bufs[i] = NULL; 943} 944 945static void 946xuio_stat_init(void) 947{ 948 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc", 949 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t), 950 KSTAT_FLAG_VIRTUAL); 951 if (xuio_ksp != NULL) { 952 xuio_ksp->ks_data = &xuio_stats; 953 kstat_install(xuio_ksp); 954 } 955} 956 957static void 958xuio_stat_fini(void) 959{ 960 if (xuio_ksp != NULL) { 961 kstat_delete(xuio_ksp); 962 xuio_ksp = NULL; 963 } 964} 965 966void 967xuio_stat_wbuf_copied() 968{ 969 XUIOSTAT_BUMP(xuiostat_wbuf_copied); 970} 971 972void 973xuio_stat_wbuf_nocopy() 974{ 975 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy); 976} 977 978#ifdef _KERNEL 979int 980dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size) 981{ 982 dmu_buf_t **dbp; 983 int numbufs, i, err; 984 xuio_t *xuio = NULL; 985 986 /* 987 * NB: we could do this block-at-a-time, but it's nice 988 * to be reading in parallel. 989 */ 990 err = dmu_buf_hold_array(os, object, uio->uio_loffset, size, TRUE, FTAG, 991 &numbufs, &dbp); 992 if (err) 993 return (err); 994 995#ifdef UIO_XUIO 996 if (uio->uio_extflg == UIO_XUIO) 997 xuio = (xuio_t *)uio; 998#endif 999 1000 for (i = 0; i < numbufs; i++) { 1001 int tocpy; 1002 int bufoff; 1003 dmu_buf_t *db = dbp[i]; 1004 1005 ASSERT(size > 0); 1006 1007 bufoff = uio->uio_loffset - db->db_offset; 1008 tocpy = (int)MIN(db->db_size - bufoff, size); 1009 1010 if (xuio) { 1011 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 1012 arc_buf_t *dbuf_abuf = dbi->db_buf; 1013 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi); 1014 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy); 1015 if (!err) { 1016 uio->uio_resid -= tocpy; 1017 uio->uio_loffset += tocpy; 1018 } 1019 1020 if (abuf == dbuf_abuf) 1021 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy); 1022 else 1023 XUIOSTAT_BUMP(xuiostat_rbuf_copied); 1024 } else { 1025 err = uiomove((char *)db->db_data + bufoff, tocpy, 1026 UIO_READ, uio); 1027 } 1028 if (err) 1029 break; 1030 1031 size -= tocpy; 1032 } 1033 dmu_buf_rele_array(dbp, numbufs, FTAG); 1034 1035 return (err); 1036} 1037 1038static int 1039dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx) 1040{ 1041 dmu_buf_t **dbp; 1042 int numbufs; 1043 int err = 0; 1044 int i; 1045 1046 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size, 1047 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH); 1048 if (err) 1049 return (err); 1050 1051 for (i = 0; i < numbufs; i++) { 1052 int tocpy; 1053 int bufoff; 1054 dmu_buf_t *db = dbp[i]; 1055 1056 ASSERT(size > 0); 1057 1058 bufoff = uio->uio_loffset - db->db_offset; 1059 tocpy = (int)MIN(db->db_size - bufoff, size); 1060 1061 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 1062 1063 if (tocpy == db->db_size) 1064 dmu_buf_will_fill(db, tx); 1065 else 1066 dmu_buf_will_dirty(db, tx); 1067 1068 /* 1069 * XXX uiomove could block forever (eg. nfs-backed 1070 * pages). There needs to be a uiolockdown() function 1071 * to lock the pages in memory, so that uiomove won't 1072 * block. 1073 */ 1074 err = uiomove((char *)db->db_data + bufoff, tocpy, 1075 UIO_WRITE, uio); 1076 1077 if (tocpy == db->db_size) 1078 dmu_buf_fill_done(db, tx); 1079 1080 if (err) 1081 break; 1082 1083 size -= tocpy; 1084 } 1085 1086 dmu_buf_rele_array(dbp, numbufs, FTAG); 1087 return (err); 1088} 1089 1090int 1091dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size, 1092 dmu_tx_t *tx) 1093{ 1094 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb; 1095 dnode_t *dn; 1096 int err; 1097 1098 if (size == 0) 1099 return (0); 1100 1101 DB_DNODE_ENTER(db); 1102 dn = DB_DNODE(db); 1103 err = dmu_write_uio_dnode(dn, uio, size, tx); 1104 DB_DNODE_EXIT(db); 1105 1106 return (err); 1107} 1108 1109int 1110dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size, 1111 dmu_tx_t *tx) 1112{ 1113 dnode_t *dn; 1114 int err; 1115 1116 if (size == 0) 1117 return (0); 1118 1119 err = dnode_hold(os, object, FTAG, &dn); 1120 if (err) 1121 return (err); 1122 1123 err = dmu_write_uio_dnode(dn, uio, size, tx); 1124 1125 dnode_rele(dn, FTAG); 1126 1127 return (err); 1128} 1129 1130#ifdef sun 1131int 1132dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1133 page_t *pp, dmu_tx_t *tx) 1134{ 1135 dmu_buf_t **dbp; 1136 int numbufs, i; 1137 int err; 1138 1139 if (size == 0) 1140 return (0); 1141 1142 err = dmu_buf_hold_array(os, object, offset, size, 1143 FALSE, FTAG, &numbufs, &dbp); 1144 if (err) 1145 return (err); 1146 1147 for (i = 0; i < numbufs; i++) { 1148 int tocpy, copied, thiscpy; 1149 int bufoff; 1150 dmu_buf_t *db = dbp[i]; 1151 caddr_t va; 1152 1153 ASSERT(size > 0); 1154 ASSERT3U(db->db_size, >=, PAGESIZE); 1155 1156 bufoff = offset - db->db_offset; 1157 tocpy = (int)MIN(db->db_size - bufoff, size); 1158 1159 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 1160 1161 if (tocpy == db->db_size) 1162 dmu_buf_will_fill(db, tx); 1163 else 1164 dmu_buf_will_dirty(db, tx); 1165 1166 for (copied = 0; copied < tocpy; copied += PAGESIZE) { 1167 ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff); 1168 thiscpy = MIN(PAGESIZE, tocpy - copied); 1169 va = zfs_map_page(pp, S_READ); 1170 bcopy(va, (char *)db->db_data + bufoff, thiscpy); 1171 zfs_unmap_page(pp, va); 1172 pp = pp->p_next; 1173 bufoff += PAGESIZE; 1174 } 1175 1176 if (tocpy == db->db_size) 1177 dmu_buf_fill_done(db, tx); 1178 1179 offset += tocpy; 1180 size -= tocpy; 1181 } 1182 dmu_buf_rele_array(dbp, numbufs, FTAG); 1183 return (err); 1184} 1185#endif /* sun */ 1186#endif 1187 1188/* 1189 * Allocate a loaned anonymous arc buffer. 1190 */ 1191arc_buf_t * 1192dmu_request_arcbuf(dmu_buf_t *handle, int size) 1193{ 1194 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle; 1195 spa_t *spa; 1196 1197 DB_GET_SPA(&spa, db); 1198 return (arc_loan_buf(spa, size)); 1199} 1200 1201/* 1202 * Free a loaned arc buffer. 1203 */ 1204void 1205dmu_return_arcbuf(arc_buf_t *buf) 1206{ 1207 arc_return_buf(buf, FTAG); 1208 VERIFY(arc_buf_remove_ref(buf, FTAG)); 1209} 1210 1211/* 1212 * When possible directly assign passed loaned arc buffer to a dbuf. 1213 * If this is not possible copy the contents of passed arc buf via 1214 * dmu_write(). 1215 */ 1216void 1217dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf, 1218 dmu_tx_t *tx) 1219{ 1220 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle; 1221 dnode_t *dn; 1222 dmu_buf_impl_t *db; 1223 uint32_t blksz = (uint32_t)arc_buf_size(buf); 1224 uint64_t blkid; 1225 1226 DB_DNODE_ENTER(dbuf); 1227 dn = DB_DNODE(dbuf); 1228 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1229 blkid = dbuf_whichblock(dn, offset); 1230 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL); 1231 rw_exit(&dn->dn_struct_rwlock); 1232 DB_DNODE_EXIT(dbuf); 1233 1234 if (offset == db->db.db_offset && blksz == db->db.db_size) { 1235 dbuf_assign_arcbuf(db, buf, tx); 1236 dbuf_rele(db, FTAG); 1237 } else { 1238 objset_t *os; 1239 uint64_t object; 1240 1241 DB_DNODE_ENTER(dbuf); 1242 dn = DB_DNODE(dbuf); 1243 os = dn->dn_objset; 1244 object = dn->dn_object; 1245 DB_DNODE_EXIT(dbuf); 1246 1247 dbuf_rele(db, FTAG); 1248 dmu_write(os, object, offset, blksz, buf->b_data, tx); 1249 dmu_return_arcbuf(buf); 1250 XUIOSTAT_BUMP(xuiostat_wbuf_copied); 1251 } 1252} 1253 1254typedef struct { 1255 dbuf_dirty_record_t *dsa_dr; 1256 dmu_sync_cb_t *dsa_done; 1257 zgd_t *dsa_zgd; 1258 dmu_tx_t *dsa_tx; 1259} dmu_sync_arg_t; 1260 1261/* ARGSUSED */ 1262static void 1263dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg) 1264{ 1265 dmu_sync_arg_t *dsa = varg; 1266 dmu_buf_t *db = dsa->dsa_zgd->zgd_db; 1267 blkptr_t *bp = zio->io_bp; 1268 1269 if (zio->io_error == 0) { 1270 if (BP_IS_HOLE(bp)) { 1271 /* 1272 * A block of zeros may compress to a hole, but the 1273 * block size still needs to be known for replay. 1274 */ 1275 BP_SET_LSIZE(bp, db->db_size); 1276 } else { 1277 ASSERT(BP_GET_LEVEL(bp) == 0); 1278 bp->blk_fill = 1; 1279 } 1280 } 1281} 1282 1283static void 1284dmu_sync_late_arrival_ready(zio_t *zio) 1285{ 1286 dmu_sync_ready(zio, NULL, zio->io_private); 1287} 1288 1289/* ARGSUSED */ 1290static void 1291dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg) 1292{ 1293 dmu_sync_arg_t *dsa = varg; 1294 dbuf_dirty_record_t *dr = dsa->dsa_dr; 1295 dmu_buf_impl_t *db = dr->dr_dbuf; 1296 1297 mutex_enter(&db->db_mtx); 1298 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC); 1299 if (zio->io_error == 0) { 1300 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE); 1301 if (dr->dt.dl.dr_nopwrite) { 1302 blkptr_t *bp = zio->io_bp; 1303 blkptr_t *bp_orig = &zio->io_bp_orig; 1304 uint8_t chksum = BP_GET_CHECKSUM(bp_orig); 1305 1306 ASSERT(BP_EQUAL(bp, bp_orig)); 1307 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF); 1308 ASSERT(zio_checksum_table[chksum].ci_dedup); 1309 } 1310 dr->dt.dl.dr_overridden_by = *zio->io_bp; 1311 dr->dt.dl.dr_override_state = DR_OVERRIDDEN; 1312 dr->dt.dl.dr_copies = zio->io_prop.zp_copies; 1313 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by)) 1314 BP_ZERO(&dr->dt.dl.dr_overridden_by); 1315 } else { 1316 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 1317 } 1318 cv_broadcast(&db->db_changed); 1319 mutex_exit(&db->db_mtx); 1320 1321 dsa->dsa_done(dsa->dsa_zgd, zio->io_error); 1322 1323 kmem_free(dsa, sizeof (*dsa)); 1324} 1325 1326static void 1327dmu_sync_late_arrival_done(zio_t *zio) 1328{ 1329 blkptr_t *bp = zio->io_bp; 1330 dmu_sync_arg_t *dsa = zio->io_private; 1331 blkptr_t *bp_orig = &zio->io_bp_orig; 1332 1333 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) { 1334 /* 1335 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE) 1336 * then there is nothing to do here. Otherwise, free the 1337 * newly allocated block in this txg. 1338 */ 1339 if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 1340 ASSERT(BP_EQUAL(bp, bp_orig)); 1341 } else { 1342 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig)); 1343 ASSERT(zio->io_bp->blk_birth == zio->io_txg); 1344 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa)); 1345 zio_free(zio->io_spa, zio->io_txg, zio->io_bp); 1346 } 1347 } 1348 1349 dmu_tx_commit(dsa->dsa_tx); 1350 1351 dsa->dsa_done(dsa->dsa_zgd, zio->io_error); 1352 1353 kmem_free(dsa, sizeof (*dsa)); 1354} 1355 1356static int 1357dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd, 1358 zio_prop_t *zp, zbookmark_t *zb) 1359{ 1360 dmu_sync_arg_t *dsa; 1361 dmu_tx_t *tx; 1362 1363 tx = dmu_tx_create(os); 1364 dmu_tx_hold_space(tx, zgd->zgd_db->db_size); 1365 if (dmu_tx_assign(tx, TXG_WAIT) != 0) { 1366 dmu_tx_abort(tx); 1367 /* Make zl_get_data do txg_waited_synced() */ 1368 return (SET_ERROR(EIO)); 1369 } 1370 1371 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP); 1372 dsa->dsa_dr = NULL; 1373 dsa->dsa_done = done; 1374 dsa->dsa_zgd = zgd; 1375 dsa->dsa_tx = tx; 1376 1377 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp, 1378 zgd->zgd_db->db_data, zgd->zgd_db->db_size, zp, 1379 dmu_sync_late_arrival_ready, dmu_sync_late_arrival_done, dsa, 1380 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb)); 1381 1382 return (0); 1383} 1384 1385/* 1386 * Intent log support: sync the block associated with db to disk. 1387 * N.B. and XXX: the caller is responsible for making sure that the 1388 * data isn't changing while dmu_sync() is writing it. 1389 * 1390 * Return values: 1391 * 1392 * EEXIST: this txg has already been synced, so there's nothing to do. 1393 * The caller should not log the write. 1394 * 1395 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do. 1396 * The caller should not log the write. 1397 * 1398 * EALREADY: this block is already in the process of being synced. 1399 * The caller should track its progress (somehow). 1400 * 1401 * EIO: could not do the I/O. 1402 * The caller should do a txg_wait_synced(). 1403 * 1404 * 0: the I/O has been initiated. 1405 * The caller should log this blkptr in the done callback. 1406 * It is possible that the I/O will fail, in which case 1407 * the error will be reported to the done callback and 1408 * propagated to pio from zio_done(). 1409 */ 1410int 1411dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd) 1412{ 1413 blkptr_t *bp = zgd->zgd_bp; 1414 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db; 1415 objset_t *os = db->db_objset; 1416 dsl_dataset_t *ds = os->os_dsl_dataset; 1417 dbuf_dirty_record_t *dr; 1418 dmu_sync_arg_t *dsa; 1419 zbookmark_t zb; 1420 zio_prop_t zp; 1421 dnode_t *dn; 1422 1423 ASSERT(pio != NULL); 1424 ASSERT(txg != 0); 1425 1426 SET_BOOKMARK(&zb, ds->ds_object, 1427 db->db.db_object, db->db_level, db->db_blkid); 1428 1429 DB_DNODE_ENTER(db); 1430 dn = DB_DNODE(db); 1431 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp); 1432 DB_DNODE_EXIT(db); 1433 1434 /* 1435 * If we're frozen (running ziltest), we always need to generate a bp. 1436 */ 1437 if (txg > spa_freeze_txg(os->os_spa)) 1438 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); 1439 1440 /* 1441 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf() 1442 * and us. If we determine that this txg is not yet syncing, 1443 * but it begins to sync a moment later, that's OK because the 1444 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx. 1445 */ 1446 mutex_enter(&db->db_mtx); 1447 1448 if (txg <= spa_last_synced_txg(os->os_spa)) { 1449 /* 1450 * This txg has already synced. There's nothing to do. 1451 */ 1452 mutex_exit(&db->db_mtx); 1453 return (SET_ERROR(EEXIST)); 1454 } 1455 1456 if (txg <= spa_syncing_txg(os->os_spa)) { 1457 /* 1458 * This txg is currently syncing, so we can't mess with 1459 * the dirty record anymore; just write a new log block. 1460 */ 1461 mutex_exit(&db->db_mtx); 1462 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); 1463 } 1464 1465 dr = db->db_last_dirty; 1466 while (dr && dr->dr_txg != txg) 1467 dr = dr->dr_next; 1468 1469 if (dr == NULL) { 1470 /* 1471 * There's no dr for this dbuf, so it must have been freed. 1472 * There's no need to log writes to freed blocks, so we're done. 1473 */ 1474 mutex_exit(&db->db_mtx); 1475 return (SET_ERROR(ENOENT)); 1476 } 1477 1478 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg); 1479 1480 /* 1481 * Assume the on-disk data is X, the current syncing data is Y, 1482 * and the current in-memory data is Z (currently in dmu_sync). 1483 * X and Z are identical but Y is has been modified. Normally, 1484 * when X and Z are the same we will perform a nopwrite but if Y 1485 * is different we must disable nopwrite since the resulting write 1486 * of Y to disk can free the block containing X. If we allowed a 1487 * nopwrite to occur the block pointing to Z would reference a freed 1488 * block. Since this is a rare case we simplify this by disabling 1489 * nopwrite if the current dmu_sync-ing dbuf has been modified in 1490 * a previous transaction. 1491 */ 1492 if (dr->dr_next) 1493 zp.zp_nopwrite = B_FALSE; 1494 1495 ASSERT(dr->dr_txg == txg); 1496 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC || 1497 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { 1498 /* 1499 * We have already issued a sync write for this buffer, 1500 * or this buffer has already been synced. It could not 1501 * have been dirtied since, or we would have cleared the state. 1502 */ 1503 mutex_exit(&db->db_mtx); 1504 return (SET_ERROR(EALREADY)); 1505 } 1506 1507 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); 1508 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC; 1509 mutex_exit(&db->db_mtx); 1510 1511 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP); 1512 dsa->dsa_dr = dr; 1513 dsa->dsa_done = done; 1514 dsa->dsa_zgd = zgd; 1515 dsa->dsa_tx = NULL; 1516 1517 zio_nowait(arc_write(pio, os->os_spa, txg, 1518 bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db), 1519 DBUF_IS_L2COMPRESSIBLE(db), &zp, dmu_sync_ready, dmu_sync_done, 1520 dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb)); 1521 1522 return (0); 1523} 1524 1525int 1526dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs, 1527 dmu_tx_t *tx) 1528{ 1529 dnode_t *dn; 1530 int err; 1531 1532 err = dnode_hold(os, object, FTAG, &dn); 1533 if (err) 1534 return (err); 1535 err = dnode_set_blksz(dn, size, ibs, tx); 1536 dnode_rele(dn, FTAG); 1537 return (err); 1538} 1539 1540void 1541dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum, 1542 dmu_tx_t *tx) 1543{ 1544 dnode_t *dn; 1545 1546 /* XXX assumes dnode_hold will not get an i/o error */ 1547 (void) dnode_hold(os, object, FTAG, &dn); 1548 ASSERT(checksum < ZIO_CHECKSUM_FUNCTIONS); 1549 dn->dn_checksum = checksum; 1550 dnode_setdirty(dn, tx); 1551 dnode_rele(dn, FTAG); 1552} 1553 1554void 1555dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress, 1556 dmu_tx_t *tx) 1557{ 1558 dnode_t *dn; 1559 1560 /* XXX assumes dnode_hold will not get an i/o error */ 1561 (void) dnode_hold(os, object, FTAG, &dn); 1562 ASSERT(compress < ZIO_COMPRESS_FUNCTIONS); 1563 dn->dn_compress = compress; 1564 dnode_setdirty(dn, tx); 1565 dnode_rele(dn, FTAG); 1566} 1567 1568int zfs_mdcomp_disable = 0; 1569TUNABLE_INT("vfs.zfs.mdcomp_disable", &zfs_mdcomp_disable); 1570SYSCTL_INT(_vfs_zfs, OID_AUTO, mdcomp_disable, CTLFLAG_RW, 1571 &zfs_mdcomp_disable, 0, "Disable metadata compression"); 1572 1573void 1574dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp) 1575{ 1576 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET; 1577 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) || 1578 (wp & WP_SPILL)); 1579 enum zio_checksum checksum = os->os_checksum; 1580 enum zio_compress compress = os->os_compress; 1581 enum zio_checksum dedup_checksum = os->os_dedup_checksum; 1582 boolean_t dedup = B_FALSE; 1583 boolean_t nopwrite = B_FALSE; 1584 boolean_t dedup_verify = os->os_dedup_verify; 1585 int copies = os->os_copies; 1586 1587 /* 1588 * We maintain different write policies for each of the following 1589 * types of data: 1590 * 1. metadata 1591 * 2. preallocated blocks (i.e. level-0 blocks of a dump device) 1592 * 3. all other level 0 blocks 1593 */ 1594 if (ismd) { 1595 /* 1596 * XXX -- we should design a compression algorithm 1597 * that specializes in arrays of bps. 1598 */ 1599 compress = zfs_mdcomp_disable ? ZIO_COMPRESS_EMPTY : 1600 ZIO_COMPRESS_LZJB; 1601 1602 /* 1603 * Metadata always gets checksummed. If the data 1604 * checksum is multi-bit correctable, and it's not a 1605 * ZBT-style checksum, then it's suitable for metadata 1606 * as well. Otherwise, the metadata checksum defaults 1607 * to fletcher4. 1608 */ 1609 if (zio_checksum_table[checksum].ci_correctable < 1 || 1610 zio_checksum_table[checksum].ci_eck) 1611 checksum = ZIO_CHECKSUM_FLETCHER_4; 1612 } else if (wp & WP_NOFILL) { 1613 ASSERT(level == 0); 1614 1615 /* 1616 * If we're writing preallocated blocks, we aren't actually 1617 * writing them so don't set any policy properties. These 1618 * blocks are currently only used by an external subsystem 1619 * outside of zfs (i.e. dump) and not written by the zio 1620 * pipeline. 1621 */ 1622 compress = ZIO_COMPRESS_OFF; 1623 checksum = ZIO_CHECKSUM_NOPARITY; 1624 } else { 1625 compress = zio_compress_select(dn->dn_compress, compress); 1626 1627 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ? 1628 zio_checksum_select(dn->dn_checksum, checksum) : 1629 dedup_checksum; 1630 1631 /* 1632 * Determine dedup setting. If we are in dmu_sync(), 1633 * we won't actually dedup now because that's all 1634 * done in syncing context; but we do want to use the 1635 * dedup checkum. If the checksum is not strong 1636 * enough to ensure unique signatures, force 1637 * dedup_verify. 1638 */ 1639 if (dedup_checksum != ZIO_CHECKSUM_OFF) { 1640 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE; 1641 if (!zio_checksum_table[checksum].ci_dedup) 1642 dedup_verify = B_TRUE; 1643 } 1644 1645 /* 1646 * Enable nopwrite if we have a cryptographically secure 1647 * checksum that has no known collisions (i.e. SHA-256) 1648 * and compression is enabled. We don't enable nopwrite if 1649 * dedup is enabled as the two features are mutually exclusive. 1650 */ 1651 nopwrite = (!dedup && zio_checksum_table[checksum].ci_dedup && 1652 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled); 1653 } 1654 1655 zp->zp_checksum = checksum; 1656 zp->zp_compress = compress; 1657 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type; 1658 zp->zp_level = level; 1659 zp->zp_copies = MIN(copies + ismd, spa_max_replication(os->os_spa)); 1660 zp->zp_dedup = dedup; 1661 zp->zp_dedup_verify = dedup && dedup_verify; 1662 zp->zp_nopwrite = nopwrite; 1663} 1664 1665int 1666dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off) 1667{ 1668 dnode_t *dn; 1669 int i, err; 1670 1671 err = dnode_hold(os, object, FTAG, &dn); 1672 if (err) 1673 return (err); 1674 /* 1675 * Sync any current changes before 1676 * we go trundling through the block pointers. 1677 */ 1678 for (i = 0; i < TXG_SIZE; i++) { 1679 if (list_link_active(&dn->dn_dirty_link[i])) 1680 break; 1681 } 1682 if (i != TXG_SIZE) { 1683 dnode_rele(dn, FTAG); 1684 txg_wait_synced(dmu_objset_pool(os), 0); 1685 err = dnode_hold(os, object, FTAG, &dn); 1686 if (err) 1687 return (err); 1688 } 1689 1690 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0); 1691 dnode_rele(dn, FTAG); 1692 1693 return (err); 1694} 1695 1696void 1697dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi) 1698{ 1699 dnode_phys_t *dnp; 1700 1701 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1702 mutex_enter(&dn->dn_mtx); 1703 1704 dnp = dn->dn_phys; 1705 1706 doi->doi_data_block_size = dn->dn_datablksz; 1707 doi->doi_metadata_block_size = dn->dn_indblkshift ? 1708 1ULL << dn->dn_indblkshift : 0; 1709 doi->doi_type = dn->dn_type; 1710 doi->doi_bonus_type = dn->dn_bonustype; 1711 doi->doi_bonus_size = dn->dn_bonuslen; 1712 doi->doi_indirection = dn->dn_nlevels; 1713 doi->doi_checksum = dn->dn_checksum; 1714 doi->doi_compress = dn->dn_compress; 1715 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9; 1716 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz; 1717 doi->doi_fill_count = 0; 1718 for (int i = 0; i < dnp->dn_nblkptr; i++) 1719 doi->doi_fill_count += dnp->dn_blkptr[i].blk_fill; 1720 1721 mutex_exit(&dn->dn_mtx); 1722 rw_exit(&dn->dn_struct_rwlock); 1723} 1724 1725/* 1726 * Get information on a DMU object. 1727 * If doi is NULL, just indicates whether the object exists. 1728 */ 1729int 1730dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi) 1731{ 1732 dnode_t *dn; 1733 int err = dnode_hold(os, object, FTAG, &dn); 1734 1735 if (err) 1736 return (err); 1737 1738 if (doi != NULL) 1739 dmu_object_info_from_dnode(dn, doi); 1740 1741 dnode_rele(dn, FTAG); 1742 return (0); 1743} 1744 1745/* 1746 * As above, but faster; can be used when you have a held dbuf in hand. 1747 */ 1748void 1749dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi) 1750{ 1751 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1752 1753 DB_DNODE_ENTER(db); 1754 dmu_object_info_from_dnode(DB_DNODE(db), doi); 1755 DB_DNODE_EXIT(db); 1756} 1757 1758/* 1759 * Faster still when you only care about the size. 1760 * This is specifically optimized for zfs_getattr(). 1761 */ 1762void 1763dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize, 1764 u_longlong_t *nblk512) 1765{ 1766 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1767 dnode_t *dn; 1768 1769 DB_DNODE_ENTER(db); 1770 dn = DB_DNODE(db); 1771 1772 *blksize = dn->dn_datablksz; 1773 /* add 1 for dnode space */ 1774 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >> 1775 SPA_MINBLOCKSHIFT) + 1; 1776 DB_DNODE_EXIT(db); 1777} 1778 1779void 1780byteswap_uint64_array(void *vbuf, size_t size) 1781{ 1782 uint64_t *buf = vbuf; 1783 size_t count = size >> 3; 1784 int i; 1785 1786 ASSERT((size & 7) == 0); 1787 1788 for (i = 0; i < count; i++) 1789 buf[i] = BSWAP_64(buf[i]); 1790} 1791 1792void 1793byteswap_uint32_array(void *vbuf, size_t size) 1794{ 1795 uint32_t *buf = vbuf; 1796 size_t count = size >> 2; 1797 int i; 1798 1799 ASSERT((size & 3) == 0); 1800 1801 for (i = 0; i < count; i++) 1802 buf[i] = BSWAP_32(buf[i]); 1803} 1804 1805void 1806byteswap_uint16_array(void *vbuf, size_t size) 1807{ 1808 uint16_t *buf = vbuf; 1809 size_t count = size >> 1; 1810 int i; 1811 1812 ASSERT((size & 1) == 0); 1813 1814 for (i = 0; i < count; i++) 1815 buf[i] = BSWAP_16(buf[i]); 1816} 1817 1818/* ARGSUSED */ 1819void 1820byteswap_uint8_array(void *vbuf, size_t size) 1821{ 1822} 1823 1824void 1825dmu_init(void) 1826{ 1827 zfs_dbgmsg_init(); 1828 sa_cache_init(); 1829 xuio_stat_init(); 1830 dmu_objset_init(); 1831 dnode_init(); 1832 dbuf_init(); 1833 zfetch_init(); 1834 zio_compress_init(); 1835 l2arc_init(); 1836 arc_init(); 1837} 1838 1839void 1840dmu_fini(void) 1841{ 1842 arc_fini(); /* arc depends on l2arc, so arc must go first */ 1843 l2arc_fini(); 1844 zfetch_fini(); 1845 zio_compress_fini(); 1846 dbuf_fini(); 1847 dnode_fini(); 1848 dmu_objset_fini(); 1849 xuio_stat_fini(); 1850 sa_cache_fini(); 1851 zfs_dbgmsg_fini(); 1852} 1853