dmu.c revision 290757
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) 2011, 2015 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/* Copyright (c) 2014, Nexenta Systems, Inc. All rights reserved. */ 28 29#include <sys/dmu.h> 30#include <sys/dmu_impl.h> 31#include <sys/dmu_tx.h> 32#include <sys/dbuf.h> 33#include <sys/dnode.h> 34#include <sys/zfs_context.h> 35#include <sys/dmu_objset.h> 36#include <sys/dmu_traverse.h> 37#include <sys/dsl_dataset.h> 38#include <sys/dsl_dir.h> 39#include <sys/dsl_pool.h> 40#include <sys/dsl_synctask.h> 41#include <sys/dsl_prop.h> 42#include <sys/dmu_zfetch.h> 43#include <sys/zfs_ioctl.h> 44#include <sys/zap.h> 45#include <sys/zio_checksum.h> 46#include <sys/zio_compress.h> 47#include <sys/sa.h> 48#include <sys/zfeature.h> 49#ifdef _KERNEL 50#include <sys/vm.h> 51#include <sys/zfs_znode.h> 52#endif 53 54/* 55 * Enable/disable nopwrite feature. 56 */ 57int zfs_nopwrite_enabled = 1; 58SYSCTL_DECL(_vfs_zfs); 59TUNABLE_INT("vfs.zfs.nopwrite_enabled", &zfs_nopwrite_enabled); 60SYSCTL_INT(_vfs_zfs, OID_AUTO, nopwrite_enabled, CTLFLAG_RDTUN, 61 &zfs_nopwrite_enabled, 0, "Enable nopwrite feature"); 62 63const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = { 64 { DMU_BSWAP_UINT8, TRUE, "unallocated" }, 65 { DMU_BSWAP_ZAP, TRUE, "object directory" }, 66 { DMU_BSWAP_UINT64, TRUE, "object array" }, 67 { DMU_BSWAP_UINT8, TRUE, "packed nvlist" }, 68 { DMU_BSWAP_UINT64, TRUE, "packed nvlist size" }, 69 { DMU_BSWAP_UINT64, TRUE, "bpobj" }, 70 { DMU_BSWAP_UINT64, TRUE, "bpobj header" }, 71 { DMU_BSWAP_UINT64, TRUE, "SPA space map header" }, 72 { DMU_BSWAP_UINT64, TRUE, "SPA space map" }, 73 { DMU_BSWAP_UINT64, TRUE, "ZIL intent log" }, 74 { DMU_BSWAP_DNODE, TRUE, "DMU dnode" }, 75 { DMU_BSWAP_OBJSET, TRUE, "DMU objset" }, 76 { DMU_BSWAP_UINT64, TRUE, "DSL directory" }, 77 { DMU_BSWAP_ZAP, TRUE, "DSL directory child map"}, 78 { DMU_BSWAP_ZAP, TRUE, "DSL dataset snap map" }, 79 { DMU_BSWAP_ZAP, TRUE, "DSL props" }, 80 { DMU_BSWAP_UINT64, TRUE, "DSL dataset" }, 81 { DMU_BSWAP_ZNODE, TRUE, "ZFS znode" }, 82 { DMU_BSWAP_OLDACL, TRUE, "ZFS V0 ACL" }, 83 { DMU_BSWAP_UINT8, FALSE, "ZFS plain file" }, 84 { DMU_BSWAP_ZAP, TRUE, "ZFS directory" }, 85 { DMU_BSWAP_ZAP, TRUE, "ZFS master node" }, 86 { DMU_BSWAP_ZAP, TRUE, "ZFS delete queue" }, 87 { DMU_BSWAP_UINT8, FALSE, "zvol object" }, 88 { DMU_BSWAP_ZAP, TRUE, "zvol prop" }, 89 { DMU_BSWAP_UINT8, FALSE, "other uint8[]" }, 90 { DMU_BSWAP_UINT64, FALSE, "other uint64[]" }, 91 { DMU_BSWAP_ZAP, TRUE, "other ZAP" }, 92 { DMU_BSWAP_ZAP, TRUE, "persistent error log" }, 93 { DMU_BSWAP_UINT8, TRUE, "SPA history" }, 94 { DMU_BSWAP_UINT64, TRUE, "SPA history offsets" }, 95 { DMU_BSWAP_ZAP, TRUE, "Pool properties" }, 96 { DMU_BSWAP_ZAP, TRUE, "DSL permissions" }, 97 { DMU_BSWAP_ACL, TRUE, "ZFS ACL" }, 98 { DMU_BSWAP_UINT8, TRUE, "ZFS SYSACL" }, 99 { DMU_BSWAP_UINT8, TRUE, "FUID table" }, 100 { DMU_BSWAP_UINT64, TRUE, "FUID table size" }, 101 { DMU_BSWAP_ZAP, TRUE, "DSL dataset next clones"}, 102 { DMU_BSWAP_ZAP, TRUE, "scan work queue" }, 103 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group used" }, 104 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group quota" }, 105 { DMU_BSWAP_ZAP, TRUE, "snapshot refcount tags"}, 106 { DMU_BSWAP_ZAP, TRUE, "DDT ZAP algorithm" }, 107 { DMU_BSWAP_ZAP, TRUE, "DDT statistics" }, 108 { DMU_BSWAP_UINT8, TRUE, "System attributes" }, 109 { DMU_BSWAP_ZAP, TRUE, "SA master node" }, 110 { DMU_BSWAP_ZAP, TRUE, "SA attr registration" }, 111 { DMU_BSWAP_ZAP, TRUE, "SA attr layouts" }, 112 { DMU_BSWAP_ZAP, TRUE, "scan translations" }, 113 { DMU_BSWAP_UINT8, FALSE, "deduplicated block" }, 114 { DMU_BSWAP_ZAP, TRUE, "DSL deadlist map" }, 115 { DMU_BSWAP_UINT64, TRUE, "DSL deadlist map hdr" }, 116 { DMU_BSWAP_ZAP, TRUE, "DSL dir clones" }, 117 { DMU_BSWAP_UINT64, TRUE, "bpobj subobj" } 118}; 119 120const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = { 121 { byteswap_uint8_array, "uint8" }, 122 { byteswap_uint16_array, "uint16" }, 123 { byteswap_uint32_array, "uint32" }, 124 { byteswap_uint64_array, "uint64" }, 125 { zap_byteswap, "zap" }, 126 { dnode_buf_byteswap, "dnode" }, 127 { dmu_objset_byteswap, "objset" }, 128 { zfs_znode_byteswap, "znode" }, 129 { zfs_oldacl_byteswap, "oldacl" }, 130 { zfs_acl_byteswap, "acl" } 131}; 132 133int 134dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset, 135 void *tag, dmu_buf_t **dbp) 136{ 137 dnode_t *dn; 138 uint64_t blkid; 139 dmu_buf_impl_t *db; 140 int err; 141 142 err = dnode_hold(os, object, FTAG, &dn); 143 if (err) 144 return (err); 145 blkid = dbuf_whichblock(dn, 0, offset); 146 rw_enter(&dn->dn_struct_rwlock, RW_READER); 147 db = dbuf_hold(dn, blkid, tag); 148 rw_exit(&dn->dn_struct_rwlock); 149 dnode_rele(dn, FTAG); 150 151 if (db == NULL) { 152 *dbp = NULL; 153 return (SET_ERROR(EIO)); 154 } 155 156 *dbp = &db->db; 157 return (err); 158} 159 160int 161dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset, 162 void *tag, dmu_buf_t **dbp, int flags) 163{ 164 int err; 165 int db_flags = DB_RF_CANFAIL; 166 167 if (flags & DMU_READ_NO_PREFETCH) 168 db_flags |= DB_RF_NOPREFETCH; 169 170 err = dmu_buf_hold_noread(os, object, offset, tag, dbp); 171 if (err == 0) { 172 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp); 173 err = dbuf_read(db, NULL, db_flags); 174 if (err != 0) { 175 dbuf_rele(db, tag); 176 *dbp = NULL; 177 } 178 } 179 180 return (err); 181} 182 183int 184dmu_bonus_max(void) 185{ 186 return (DN_MAX_BONUSLEN); 187} 188 189int 190dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx) 191{ 192 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 193 dnode_t *dn; 194 int error; 195 196 DB_DNODE_ENTER(db); 197 dn = DB_DNODE(db); 198 199 if (dn->dn_bonus != db) { 200 error = SET_ERROR(EINVAL); 201 } else if (newsize < 0 || newsize > db_fake->db_size) { 202 error = SET_ERROR(EINVAL); 203 } else { 204 dnode_setbonuslen(dn, newsize, tx); 205 error = 0; 206 } 207 208 DB_DNODE_EXIT(db); 209 return (error); 210} 211 212int 213dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx) 214{ 215 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 216 dnode_t *dn; 217 int error; 218 219 DB_DNODE_ENTER(db); 220 dn = DB_DNODE(db); 221 222 if (!DMU_OT_IS_VALID(type)) { 223 error = SET_ERROR(EINVAL); 224 } else if (dn->dn_bonus != db) { 225 error = SET_ERROR(EINVAL); 226 } else { 227 dnode_setbonus_type(dn, type, tx); 228 error = 0; 229 } 230 231 DB_DNODE_EXIT(db); 232 return (error); 233} 234 235dmu_object_type_t 236dmu_get_bonustype(dmu_buf_t *db_fake) 237{ 238 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 239 dnode_t *dn; 240 dmu_object_type_t type; 241 242 DB_DNODE_ENTER(db); 243 dn = DB_DNODE(db); 244 type = dn->dn_bonustype; 245 DB_DNODE_EXIT(db); 246 247 return (type); 248} 249 250int 251dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx) 252{ 253 dnode_t *dn; 254 int error; 255 256 error = dnode_hold(os, object, FTAG, &dn); 257 dbuf_rm_spill(dn, tx); 258 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 259 dnode_rm_spill(dn, tx); 260 rw_exit(&dn->dn_struct_rwlock); 261 dnode_rele(dn, FTAG); 262 return (error); 263} 264 265/* 266 * returns ENOENT, EIO, or 0. 267 */ 268int 269dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp) 270{ 271 dnode_t *dn; 272 dmu_buf_impl_t *db; 273 int error; 274 275 error = dnode_hold(os, object, FTAG, &dn); 276 if (error) 277 return (error); 278 279 rw_enter(&dn->dn_struct_rwlock, RW_READER); 280 if (dn->dn_bonus == NULL) { 281 rw_exit(&dn->dn_struct_rwlock); 282 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 283 if (dn->dn_bonus == NULL) 284 dbuf_create_bonus(dn); 285 } 286 db = dn->dn_bonus; 287 288 /* as long as the bonus buf is held, the dnode will be held */ 289 if (refcount_add(&db->db_holds, tag) == 1) { 290 VERIFY(dnode_add_ref(dn, db)); 291 atomic_inc_32(&dn->dn_dbufs_count); 292 } 293 294 /* 295 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's 296 * hold and incrementing the dbuf count to ensure that dnode_move() sees 297 * a dnode hold for every dbuf. 298 */ 299 rw_exit(&dn->dn_struct_rwlock); 300 301 dnode_rele(dn, FTAG); 302 303 VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH)); 304 305 *dbp = &db->db; 306 return (0); 307} 308 309/* 310 * returns ENOENT, EIO, or 0. 311 * 312 * This interface will allocate a blank spill dbuf when a spill blk 313 * doesn't already exist on the dnode. 314 * 315 * if you only want to find an already existing spill db, then 316 * dmu_spill_hold_existing() should be used. 317 */ 318int 319dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp) 320{ 321 dmu_buf_impl_t *db = NULL; 322 int err; 323 324 if ((flags & DB_RF_HAVESTRUCT) == 0) 325 rw_enter(&dn->dn_struct_rwlock, RW_READER); 326 327 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag); 328 329 if ((flags & DB_RF_HAVESTRUCT) == 0) 330 rw_exit(&dn->dn_struct_rwlock); 331 332 ASSERT(db != NULL); 333 err = dbuf_read(db, NULL, flags); 334 if (err == 0) 335 *dbp = &db->db; 336 else 337 dbuf_rele(db, tag); 338 return (err); 339} 340 341int 342dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp) 343{ 344 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus; 345 dnode_t *dn; 346 int err; 347 348 DB_DNODE_ENTER(db); 349 dn = DB_DNODE(db); 350 351 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) { 352 err = SET_ERROR(EINVAL); 353 } else { 354 rw_enter(&dn->dn_struct_rwlock, RW_READER); 355 356 if (!dn->dn_have_spill) { 357 err = SET_ERROR(ENOENT); 358 } else { 359 err = dmu_spill_hold_by_dnode(dn, 360 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp); 361 } 362 363 rw_exit(&dn->dn_struct_rwlock); 364 } 365 366 DB_DNODE_EXIT(db); 367 return (err); 368} 369 370int 371dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp) 372{ 373 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus; 374 dnode_t *dn; 375 int err; 376 377 DB_DNODE_ENTER(db); 378 dn = DB_DNODE(db); 379 err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp); 380 DB_DNODE_EXIT(db); 381 382 return (err); 383} 384 385/* 386 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces 387 * to take a held dnode rather than <os, object> -- the lookup is wasteful, 388 * and can induce severe lock contention when writing to several files 389 * whose dnodes are in the same block. 390 */ 391static int 392dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length, 393 boolean_t read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags) 394{ 395 dmu_buf_t **dbp; 396 uint64_t blkid, nblks, i; 397 uint32_t dbuf_flags; 398 int err; 399 zio_t *zio; 400 401 ASSERT(length <= DMU_MAX_ACCESS); 402 403 /* 404 * Note: We directly notify the prefetch code of this read, so that 405 * we can tell it about the multi-block read. dbuf_read() only knows 406 * about the one block it is accessing. 407 */ 408 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT | 409 DB_RF_NOPREFETCH; 410 411 rw_enter(&dn->dn_struct_rwlock, RW_READER); 412 if (dn->dn_datablkshift) { 413 int blkshift = dn->dn_datablkshift; 414 nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) - 415 P2ALIGN(offset, 1ULL << blkshift)) >> blkshift; 416 } else { 417 if (offset + length > dn->dn_datablksz) { 418 zfs_panic_recover("zfs: accessing past end of object " 419 "%llx/%llx (size=%u access=%llu+%llu)", 420 (longlong_t)dn->dn_objset-> 421 os_dsl_dataset->ds_object, 422 (longlong_t)dn->dn_object, dn->dn_datablksz, 423 (longlong_t)offset, (longlong_t)length); 424 rw_exit(&dn->dn_struct_rwlock); 425 return (SET_ERROR(EIO)); 426 } 427 nblks = 1; 428 } 429 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP); 430 431 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL); 432 blkid = dbuf_whichblock(dn, 0, offset); 433 for (i = 0; i < nblks; i++) { 434 dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag); 435 if (db == NULL) { 436 rw_exit(&dn->dn_struct_rwlock); 437 dmu_buf_rele_array(dbp, nblks, tag); 438 zio_nowait(zio); 439 return (SET_ERROR(EIO)); 440 } 441 442 /* initiate async i/o */ 443 if (read) 444 (void) dbuf_read(db, zio, dbuf_flags); 445#ifdef _KERNEL 446 else 447 curthread->td_ru.ru_oublock++; 448#endif 449 dbp[i] = &db->db; 450 } 451 452 if ((flags & DMU_READ_NO_PREFETCH) == 0 && read && 453 length <= zfetch_array_rd_sz) { 454 dmu_zfetch(&dn->dn_zfetch, blkid, nblks); 455 } 456 rw_exit(&dn->dn_struct_rwlock); 457 458 /* wait for async i/o */ 459 err = zio_wait(zio); 460 if (err) { 461 dmu_buf_rele_array(dbp, nblks, tag); 462 return (err); 463 } 464 465 /* wait for other io to complete */ 466 if (read) { 467 for (i = 0; i < nblks; i++) { 468 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i]; 469 mutex_enter(&db->db_mtx); 470 while (db->db_state == DB_READ || 471 db->db_state == DB_FILL) 472 cv_wait(&db->db_changed, &db->db_mtx); 473 if (db->db_state == DB_UNCACHED) 474 err = SET_ERROR(EIO); 475 mutex_exit(&db->db_mtx); 476 if (err) { 477 dmu_buf_rele_array(dbp, nblks, tag); 478 return (err); 479 } 480 } 481 } 482 483 *numbufsp = nblks; 484 *dbpp = dbp; 485 return (0); 486} 487 488static int 489dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset, 490 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp) 491{ 492 dnode_t *dn; 493 int err; 494 495 err = dnode_hold(os, object, FTAG, &dn); 496 if (err) 497 return (err); 498 499 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag, 500 numbufsp, dbpp, DMU_READ_PREFETCH); 501 502 dnode_rele(dn, FTAG); 503 504 return (err); 505} 506 507int 508dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset, 509 uint64_t length, boolean_t read, void *tag, int *numbufsp, 510 dmu_buf_t ***dbpp) 511{ 512 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 513 dnode_t *dn; 514 int err; 515 516 DB_DNODE_ENTER(db); 517 dn = DB_DNODE(db); 518 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag, 519 numbufsp, dbpp, DMU_READ_PREFETCH); 520 DB_DNODE_EXIT(db); 521 522 return (err); 523} 524 525void 526dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag) 527{ 528 int i; 529 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake; 530 531 if (numbufs == 0) 532 return; 533 534 for (i = 0; i < numbufs; i++) { 535 if (dbp[i]) 536 dbuf_rele(dbp[i], tag); 537 } 538 539 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs); 540} 541 542/* 543 * Issue prefetch i/os for the given blocks. If level is greater than 0, the 544 * indirect blocks prefeteched will be those that point to the blocks containing 545 * the data starting at offset, and continuing to offset + len. 546 * 547 * Note that if the indirect blocks above the blocks being prefetched are not in 548 * cache, they will be asychronously read in. 549 */ 550void 551dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset, 552 uint64_t len, zio_priority_t pri) 553{ 554 dnode_t *dn; 555 uint64_t blkid; 556 int nblks, err; 557 558 if (len == 0) { /* they're interested in the bonus buffer */ 559 dn = DMU_META_DNODE(os); 560 561 if (object == 0 || object >= DN_MAX_OBJECT) 562 return; 563 564 rw_enter(&dn->dn_struct_rwlock, RW_READER); 565 blkid = dbuf_whichblock(dn, level, 566 object * sizeof (dnode_phys_t)); 567 dbuf_prefetch(dn, level, blkid, pri, 0); 568 rw_exit(&dn->dn_struct_rwlock); 569 return; 570 } 571 572 /* 573 * XXX - Note, if the dnode for the requested object is not 574 * already cached, we will do a *synchronous* read in the 575 * dnode_hold() call. The same is true for any indirects. 576 */ 577 err = dnode_hold(os, object, FTAG, &dn); 578 if (err != 0) 579 return; 580 581 rw_enter(&dn->dn_struct_rwlock, RW_READER); 582 /* 583 * offset + len - 1 is the last byte we want to prefetch for, and offset 584 * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the 585 * last block we want to prefetch, and dbuf_whichblock(dn, level, 586 * offset) is the first. Then the number we need to prefetch is the 587 * last - first + 1. 588 */ 589 if (level > 0 || dn->dn_datablkshift != 0) { 590 nblks = dbuf_whichblock(dn, level, offset + len - 1) - 591 dbuf_whichblock(dn, level, offset) + 1; 592 } else { 593 nblks = (offset < dn->dn_datablksz); 594 } 595 596 if (nblks != 0) { 597 blkid = dbuf_whichblock(dn, level, offset); 598 for (int i = 0; i < nblks; i++) 599 dbuf_prefetch(dn, level, blkid + i, pri, 0); 600 } 601 602 rw_exit(&dn->dn_struct_rwlock); 603 604 dnode_rele(dn, FTAG); 605} 606 607/* 608 * Get the next "chunk" of file data to free. We traverse the file from 609 * the end so that the file gets shorter over time (if we crashes in the 610 * middle, this will leave us in a better state). We find allocated file 611 * data by simply searching the allocated level 1 indirects. 612 * 613 * On input, *start should be the first offset that does not need to be 614 * freed (e.g. "offset + length"). On return, *start will be the first 615 * offset that should be freed. 616 */ 617static int 618get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum) 619{ 620 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1); 621 /* bytes of data covered by a level-1 indirect block */ 622 uint64_t iblkrange = 623 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT); 624 625 ASSERT3U(minimum, <=, *start); 626 627 if (*start - minimum <= iblkrange * maxblks) { 628 *start = minimum; 629 return (0); 630 } 631 ASSERT(ISP2(iblkrange)); 632 633 for (uint64_t blks = 0; *start > minimum && blks < maxblks; blks++) { 634 int err; 635 636 /* 637 * dnode_next_offset(BACKWARDS) will find an allocated L1 638 * indirect block at or before the input offset. We must 639 * decrement *start so that it is at the end of the region 640 * to search. 641 */ 642 (*start)--; 643 err = dnode_next_offset(dn, 644 DNODE_FIND_BACKWARDS, start, 2, 1, 0); 645 646 /* if there are no indirect blocks before start, we are done */ 647 if (err == ESRCH) { 648 *start = minimum; 649 break; 650 } else if (err != 0) { 651 return (err); 652 } 653 654 /* set start to the beginning of this L1 indirect */ 655 *start = P2ALIGN(*start, iblkrange); 656 } 657 if (*start < minimum) 658 *start = minimum; 659 return (0); 660} 661 662static int 663dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset, 664 uint64_t length) 665{ 666 uint64_t object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz; 667 int err; 668 669 if (offset >= object_size) 670 return (0); 671 672 if (length == DMU_OBJECT_END || offset + length > object_size) 673 length = object_size - offset; 674 675 while (length != 0) { 676 uint64_t chunk_end, chunk_begin; 677 678 chunk_end = chunk_begin = offset + length; 679 680 /* move chunk_begin backwards to the beginning of this chunk */ 681 err = get_next_chunk(dn, &chunk_begin, offset); 682 if (err) 683 return (err); 684 ASSERT3U(chunk_begin, >=, offset); 685 ASSERT3U(chunk_begin, <=, chunk_end); 686 687 dmu_tx_t *tx = dmu_tx_create(os); 688 dmu_tx_hold_free(tx, dn->dn_object, 689 chunk_begin, chunk_end - chunk_begin); 690 691 /* 692 * Mark this transaction as typically resulting in a net 693 * reduction in space used. 694 */ 695 dmu_tx_mark_netfree(tx); 696 err = dmu_tx_assign(tx, TXG_WAIT); 697 if (err) { 698 dmu_tx_abort(tx); 699 return (err); 700 } 701 dnode_free_range(dn, chunk_begin, chunk_end - chunk_begin, tx); 702 dmu_tx_commit(tx); 703 704 length -= chunk_end - chunk_begin; 705 } 706 return (0); 707} 708 709int 710dmu_free_long_range(objset_t *os, uint64_t object, 711 uint64_t offset, uint64_t length) 712{ 713 dnode_t *dn; 714 int err; 715 716 err = dnode_hold(os, object, FTAG, &dn); 717 if (err != 0) 718 return (err); 719 err = dmu_free_long_range_impl(os, dn, offset, length); 720 721 /* 722 * It is important to zero out the maxblkid when freeing the entire 723 * file, so that (a) subsequent calls to dmu_free_long_range_impl() 724 * will take the fast path, and (b) dnode_reallocate() can verify 725 * that the entire file has been freed. 726 */ 727 if (err == 0 && offset == 0 && length == DMU_OBJECT_END) 728 dn->dn_maxblkid = 0; 729 730 dnode_rele(dn, FTAG); 731 return (err); 732} 733 734int 735dmu_free_long_object(objset_t *os, uint64_t object) 736{ 737 dmu_tx_t *tx; 738 int err; 739 740 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END); 741 if (err != 0) 742 return (err); 743 744 tx = dmu_tx_create(os); 745 dmu_tx_hold_bonus(tx, object); 746 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END); 747 dmu_tx_mark_netfree(tx); 748 err = dmu_tx_assign(tx, TXG_WAIT); 749 if (err == 0) { 750 err = dmu_object_free(os, object, tx); 751 dmu_tx_commit(tx); 752 } else { 753 dmu_tx_abort(tx); 754 } 755 756 return (err); 757} 758 759int 760dmu_free_range(objset_t *os, uint64_t object, uint64_t offset, 761 uint64_t size, dmu_tx_t *tx) 762{ 763 dnode_t *dn; 764 int err = dnode_hold(os, object, FTAG, &dn); 765 if (err) 766 return (err); 767 ASSERT(offset < UINT64_MAX); 768 ASSERT(size == -1ULL || size <= UINT64_MAX - offset); 769 dnode_free_range(dn, offset, size, tx); 770 dnode_rele(dn, FTAG); 771 return (0); 772} 773 774int 775dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 776 void *buf, uint32_t flags) 777{ 778 dnode_t *dn; 779 dmu_buf_t **dbp; 780 int numbufs, err; 781 782 err = dnode_hold(os, object, FTAG, &dn); 783 if (err) 784 return (err); 785 786 /* 787 * Deal with odd block sizes, where there can't be data past the first 788 * block. If we ever do the tail block optimization, we will need to 789 * handle that here as well. 790 */ 791 if (dn->dn_maxblkid == 0) { 792 int newsz = offset > dn->dn_datablksz ? 0 : 793 MIN(size, dn->dn_datablksz - offset); 794 bzero((char *)buf + newsz, size - newsz); 795 size = newsz; 796 } 797 798 while (size > 0) { 799 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2); 800 int i; 801 802 /* 803 * NB: we could do this block-at-a-time, but it's nice 804 * to be reading in parallel. 805 */ 806 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen, 807 TRUE, FTAG, &numbufs, &dbp, flags); 808 if (err) 809 break; 810 811 for (i = 0; i < numbufs; i++) { 812 int tocpy; 813 int bufoff; 814 dmu_buf_t *db = dbp[i]; 815 816 ASSERT(size > 0); 817 818 bufoff = offset - db->db_offset; 819 tocpy = (int)MIN(db->db_size - bufoff, size); 820 821 bcopy((char *)db->db_data + bufoff, buf, tocpy); 822 823 offset += tocpy; 824 size -= tocpy; 825 buf = (char *)buf + tocpy; 826 } 827 dmu_buf_rele_array(dbp, numbufs, FTAG); 828 } 829 dnode_rele(dn, FTAG); 830 return (err); 831} 832 833void 834dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 835 const void *buf, dmu_tx_t *tx) 836{ 837 dmu_buf_t **dbp; 838 int numbufs, i; 839 840 if (size == 0) 841 return; 842 843 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size, 844 FALSE, FTAG, &numbufs, &dbp)); 845 846 for (i = 0; i < numbufs; i++) { 847 int tocpy; 848 int bufoff; 849 dmu_buf_t *db = dbp[i]; 850 851 ASSERT(size > 0); 852 853 bufoff = offset - db->db_offset; 854 tocpy = (int)MIN(db->db_size - bufoff, size); 855 856 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 857 858 if (tocpy == db->db_size) 859 dmu_buf_will_fill(db, tx); 860 else 861 dmu_buf_will_dirty(db, tx); 862 863 bcopy(buf, (char *)db->db_data + bufoff, tocpy); 864 865 if (tocpy == db->db_size) 866 dmu_buf_fill_done(db, tx); 867 868 offset += tocpy; 869 size -= tocpy; 870 buf = (char *)buf + tocpy; 871 } 872 dmu_buf_rele_array(dbp, numbufs, FTAG); 873} 874 875void 876dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 877 dmu_tx_t *tx) 878{ 879 dmu_buf_t **dbp; 880 int numbufs, i; 881 882 if (size == 0) 883 return; 884 885 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size, 886 FALSE, FTAG, &numbufs, &dbp)); 887 888 for (i = 0; i < numbufs; i++) { 889 dmu_buf_t *db = dbp[i]; 890 891 dmu_buf_will_not_fill(db, tx); 892 } 893 dmu_buf_rele_array(dbp, numbufs, FTAG); 894} 895 896void 897dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset, 898 void *data, uint8_t etype, uint8_t comp, int uncompressed_size, 899 int compressed_size, int byteorder, dmu_tx_t *tx) 900{ 901 dmu_buf_t *db; 902 903 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES); 904 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS); 905 VERIFY0(dmu_buf_hold_noread(os, object, offset, 906 FTAG, &db)); 907 908 dmu_buf_write_embedded(db, 909 data, (bp_embedded_type_t)etype, (enum zio_compress)comp, 910 uncompressed_size, compressed_size, byteorder, tx); 911 912 dmu_buf_rele(db, FTAG); 913} 914 915/* 916 * DMU support for xuio 917 */ 918kstat_t *xuio_ksp = NULL; 919 920int 921dmu_xuio_init(xuio_t *xuio, int nblk) 922{ 923 dmu_xuio_t *priv; 924 uio_t *uio = &xuio->xu_uio; 925 926 uio->uio_iovcnt = nblk; 927 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP); 928 929 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP); 930 priv->cnt = nblk; 931 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP); 932 priv->iovp = uio->uio_iov; 933 XUIO_XUZC_PRIV(xuio) = priv; 934 935 if (XUIO_XUZC_RW(xuio) == UIO_READ) 936 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk); 937 else 938 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk); 939 940 return (0); 941} 942 943void 944dmu_xuio_fini(xuio_t *xuio) 945{ 946 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 947 int nblk = priv->cnt; 948 949 kmem_free(priv->iovp, nblk * sizeof (iovec_t)); 950 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *)); 951 kmem_free(priv, sizeof (dmu_xuio_t)); 952 953 if (XUIO_XUZC_RW(xuio) == UIO_READ) 954 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk); 955 else 956 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk); 957} 958 959/* 960 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf } 961 * and increase priv->next by 1. 962 */ 963int 964dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n) 965{ 966 struct iovec *iov; 967 uio_t *uio = &xuio->xu_uio; 968 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 969 int i = priv->next++; 970 971 ASSERT(i < priv->cnt); 972 ASSERT(off + n <= arc_buf_size(abuf)); 973 iov = uio->uio_iov + i; 974 iov->iov_base = (char *)abuf->b_data + off; 975 iov->iov_len = n; 976 priv->bufs[i] = abuf; 977 return (0); 978} 979 980int 981dmu_xuio_cnt(xuio_t *xuio) 982{ 983 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 984 return (priv->cnt); 985} 986 987arc_buf_t * 988dmu_xuio_arcbuf(xuio_t *xuio, int i) 989{ 990 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 991 992 ASSERT(i < priv->cnt); 993 return (priv->bufs[i]); 994} 995 996void 997dmu_xuio_clear(xuio_t *xuio, int i) 998{ 999 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 1000 1001 ASSERT(i < priv->cnt); 1002 priv->bufs[i] = NULL; 1003} 1004 1005static void 1006xuio_stat_init(void) 1007{ 1008 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc", 1009 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t), 1010 KSTAT_FLAG_VIRTUAL); 1011 if (xuio_ksp != NULL) { 1012 xuio_ksp->ks_data = &xuio_stats; 1013 kstat_install(xuio_ksp); 1014 } 1015} 1016 1017static void 1018xuio_stat_fini(void) 1019{ 1020 if (xuio_ksp != NULL) { 1021 kstat_delete(xuio_ksp); 1022 xuio_ksp = NULL; 1023 } 1024} 1025 1026void 1027xuio_stat_wbuf_copied() 1028{ 1029 XUIOSTAT_BUMP(xuiostat_wbuf_copied); 1030} 1031 1032void 1033xuio_stat_wbuf_nocopy() 1034{ 1035 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy); 1036} 1037 1038#ifdef _KERNEL 1039static int 1040dmu_read_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size) 1041{ 1042 dmu_buf_t **dbp; 1043 int numbufs, i, err; 1044 xuio_t *xuio = NULL; 1045 1046 /* 1047 * NB: we could do this block-at-a-time, but it's nice 1048 * to be reading in parallel. 1049 */ 1050 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size, 1051 TRUE, FTAG, &numbufs, &dbp, 0); 1052 if (err) 1053 return (err); 1054 1055#ifdef UIO_XUIO 1056 if (uio->uio_extflg == UIO_XUIO) 1057 xuio = (xuio_t *)uio; 1058#endif 1059 1060 for (i = 0; i < numbufs; i++) { 1061 int tocpy; 1062 int bufoff; 1063 dmu_buf_t *db = dbp[i]; 1064 1065 ASSERT(size > 0); 1066 1067 bufoff = uio->uio_loffset - db->db_offset; 1068 tocpy = (int)MIN(db->db_size - bufoff, size); 1069 1070 if (xuio) { 1071 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 1072 arc_buf_t *dbuf_abuf = dbi->db_buf; 1073 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi); 1074 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy); 1075 if (!err) { 1076 uio->uio_resid -= tocpy; 1077 uio->uio_loffset += tocpy; 1078 } 1079 1080 if (abuf == dbuf_abuf) 1081 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy); 1082 else 1083 XUIOSTAT_BUMP(xuiostat_rbuf_copied); 1084 } else { 1085 err = uiomove((char *)db->db_data + bufoff, tocpy, 1086 UIO_READ, uio); 1087 } 1088 if (err) 1089 break; 1090 1091 size -= tocpy; 1092 } 1093 dmu_buf_rele_array(dbp, numbufs, FTAG); 1094 1095 return (err); 1096} 1097 1098/* 1099 * Read 'size' bytes into the uio buffer. 1100 * From object zdb->db_object. 1101 * Starting at offset uio->uio_loffset. 1102 * 1103 * If the caller already has a dbuf in the target object 1104 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(), 1105 * because we don't have to find the dnode_t for the object. 1106 */ 1107int 1108dmu_read_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size) 1109{ 1110 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb; 1111 dnode_t *dn; 1112 int err; 1113 1114 if (size == 0) 1115 return (0); 1116 1117 DB_DNODE_ENTER(db); 1118 dn = DB_DNODE(db); 1119 err = dmu_read_uio_dnode(dn, uio, size); 1120 DB_DNODE_EXIT(db); 1121 1122 return (err); 1123} 1124 1125/* 1126 * Read 'size' bytes into the uio buffer. 1127 * From the specified object 1128 * Starting at offset uio->uio_loffset. 1129 */ 1130int 1131dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size) 1132{ 1133 dnode_t *dn; 1134 int err; 1135 1136 if (size == 0) 1137 return (0); 1138 1139 err = dnode_hold(os, object, FTAG, &dn); 1140 if (err) 1141 return (err); 1142 1143 err = dmu_read_uio_dnode(dn, uio, size); 1144 1145 dnode_rele(dn, FTAG); 1146 1147 return (err); 1148} 1149 1150static int 1151dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx) 1152{ 1153 dmu_buf_t **dbp; 1154 int numbufs; 1155 int err = 0; 1156 int i; 1157 1158 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size, 1159 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH); 1160 if (err) 1161 return (err); 1162 1163 for (i = 0; i < numbufs; i++) { 1164 int tocpy; 1165 int bufoff; 1166 dmu_buf_t *db = dbp[i]; 1167 1168 ASSERT(size > 0); 1169 1170 bufoff = uio->uio_loffset - db->db_offset; 1171 tocpy = (int)MIN(db->db_size - bufoff, size); 1172 1173 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 1174 1175 if (tocpy == db->db_size) 1176 dmu_buf_will_fill(db, tx); 1177 else 1178 dmu_buf_will_dirty(db, tx); 1179 1180 /* 1181 * XXX uiomove could block forever (eg. nfs-backed 1182 * pages). There needs to be a uiolockdown() function 1183 * to lock the pages in memory, so that uiomove won't 1184 * block. 1185 */ 1186 err = uiomove((char *)db->db_data + bufoff, tocpy, 1187 UIO_WRITE, uio); 1188 1189 if (tocpy == db->db_size) 1190 dmu_buf_fill_done(db, tx); 1191 1192 if (err) 1193 break; 1194 1195 size -= tocpy; 1196 } 1197 1198 dmu_buf_rele_array(dbp, numbufs, FTAG); 1199 return (err); 1200} 1201 1202/* 1203 * Write 'size' bytes from the uio buffer. 1204 * To object zdb->db_object. 1205 * Starting at offset uio->uio_loffset. 1206 * 1207 * If the caller already has a dbuf in the target object 1208 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(), 1209 * because we don't have to find the dnode_t for the object. 1210 */ 1211int 1212dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size, 1213 dmu_tx_t *tx) 1214{ 1215 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb; 1216 dnode_t *dn; 1217 int err; 1218 1219 if (size == 0) 1220 return (0); 1221 1222 DB_DNODE_ENTER(db); 1223 dn = DB_DNODE(db); 1224 err = dmu_write_uio_dnode(dn, uio, size, tx); 1225 DB_DNODE_EXIT(db); 1226 1227 return (err); 1228} 1229 1230/* 1231 * Write 'size' bytes from the uio buffer. 1232 * To the specified object. 1233 * Starting at offset uio->uio_loffset. 1234 */ 1235int 1236dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size, 1237 dmu_tx_t *tx) 1238{ 1239 dnode_t *dn; 1240 int err; 1241 1242 if (size == 0) 1243 return (0); 1244 1245 err = dnode_hold(os, object, FTAG, &dn); 1246 if (err) 1247 return (err); 1248 1249 err = dmu_write_uio_dnode(dn, uio, size, tx); 1250 1251 dnode_rele(dn, FTAG); 1252 1253 return (err); 1254} 1255 1256#ifdef sun 1257int 1258dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1259 page_t *pp, dmu_tx_t *tx) 1260{ 1261 dmu_buf_t **dbp; 1262 int numbufs, i; 1263 int err; 1264 1265 if (size == 0) 1266 return (0); 1267 1268 err = dmu_buf_hold_array(os, object, offset, size, 1269 FALSE, FTAG, &numbufs, &dbp); 1270 if (err) 1271 return (err); 1272 1273 for (i = 0; i < numbufs; i++) { 1274 int tocpy, copied, thiscpy; 1275 int bufoff; 1276 dmu_buf_t *db = dbp[i]; 1277 caddr_t va; 1278 1279 ASSERT(size > 0); 1280 ASSERT3U(db->db_size, >=, PAGESIZE); 1281 1282 bufoff = offset - db->db_offset; 1283 tocpy = (int)MIN(db->db_size - bufoff, size); 1284 1285 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 1286 1287 if (tocpy == db->db_size) 1288 dmu_buf_will_fill(db, tx); 1289 else 1290 dmu_buf_will_dirty(db, tx); 1291 1292 for (copied = 0; copied < tocpy; copied += PAGESIZE) { 1293 ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff); 1294 thiscpy = MIN(PAGESIZE, tocpy - copied); 1295 va = zfs_map_page(pp, S_READ); 1296 bcopy(va, (char *)db->db_data + bufoff, thiscpy); 1297 zfs_unmap_page(pp, va); 1298 pp = pp->p_next; 1299 bufoff += PAGESIZE; 1300 } 1301 1302 if (tocpy == db->db_size) 1303 dmu_buf_fill_done(db, tx); 1304 1305 offset += tocpy; 1306 size -= tocpy; 1307 } 1308 dmu_buf_rele_array(dbp, numbufs, FTAG); 1309 return (err); 1310} 1311 1312#else 1313 1314int 1315dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1316 vm_page_t *ma, dmu_tx_t *tx) 1317{ 1318 dmu_buf_t **dbp; 1319 struct sf_buf *sf; 1320 int numbufs, i; 1321 int err; 1322 1323 if (size == 0) 1324 return (0); 1325 1326 err = dmu_buf_hold_array(os, object, offset, size, 1327 FALSE, FTAG, &numbufs, &dbp); 1328 if (err) 1329 return (err); 1330 1331 for (i = 0; i < numbufs; i++) { 1332 int tocpy, copied, thiscpy; 1333 int bufoff; 1334 dmu_buf_t *db = dbp[i]; 1335 caddr_t va; 1336 1337 ASSERT(size > 0); 1338 ASSERT3U(db->db_size, >=, PAGESIZE); 1339 1340 bufoff = offset - db->db_offset; 1341 tocpy = (int)MIN(db->db_size - bufoff, size); 1342 1343 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 1344 1345 if (tocpy == db->db_size) 1346 dmu_buf_will_fill(db, tx); 1347 else 1348 dmu_buf_will_dirty(db, tx); 1349 1350 for (copied = 0; copied < tocpy; copied += PAGESIZE) { 1351 ASSERT3U(ptoa((*ma)->pindex), ==, db->db_offset + bufoff); 1352 thiscpy = MIN(PAGESIZE, tocpy - copied); 1353 va = zfs_map_page(*ma, &sf); 1354 bcopy(va, (char *)db->db_data + bufoff, thiscpy); 1355 zfs_unmap_page(sf); 1356 ma += 1; 1357 bufoff += PAGESIZE; 1358 } 1359 1360 if (tocpy == db->db_size) 1361 dmu_buf_fill_done(db, tx); 1362 1363 offset += tocpy; 1364 size -= tocpy; 1365 } 1366 dmu_buf_rele_array(dbp, numbufs, FTAG); 1367 return (err); 1368} 1369#endif /* sun */ 1370#endif 1371 1372/* 1373 * Allocate a loaned anonymous arc buffer. 1374 */ 1375arc_buf_t * 1376dmu_request_arcbuf(dmu_buf_t *handle, int size) 1377{ 1378 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle; 1379 1380 return (arc_loan_buf(db->db_objset->os_spa, size)); 1381} 1382 1383/* 1384 * Free a loaned arc buffer. 1385 */ 1386void 1387dmu_return_arcbuf(arc_buf_t *buf) 1388{ 1389 arc_return_buf(buf, FTAG); 1390 VERIFY(arc_buf_remove_ref(buf, FTAG)); 1391} 1392 1393/* 1394 * When possible directly assign passed loaned arc buffer to a dbuf. 1395 * If this is not possible copy the contents of passed arc buf via 1396 * dmu_write(). 1397 */ 1398void 1399dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf, 1400 dmu_tx_t *tx) 1401{ 1402 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle; 1403 dnode_t *dn; 1404 dmu_buf_impl_t *db; 1405 uint32_t blksz = (uint32_t)arc_buf_size(buf); 1406 uint64_t blkid; 1407 1408 DB_DNODE_ENTER(dbuf); 1409 dn = DB_DNODE(dbuf); 1410 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1411 blkid = dbuf_whichblock(dn, 0, offset); 1412 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL); 1413 rw_exit(&dn->dn_struct_rwlock); 1414 DB_DNODE_EXIT(dbuf); 1415 1416 /* 1417 * We can only assign if the offset is aligned, the arc buf is the 1418 * same size as the dbuf, and the dbuf is not metadata. It 1419 * can't be metadata because the loaned arc buf comes from the 1420 * user-data kmem arena. 1421 */ 1422 if (offset == db->db.db_offset && blksz == db->db.db_size && 1423 DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA) { 1424 dbuf_assign_arcbuf(db, buf, tx); 1425 dbuf_rele(db, FTAG); 1426 } else { 1427 objset_t *os; 1428 uint64_t object; 1429 1430 DB_DNODE_ENTER(dbuf); 1431 dn = DB_DNODE(dbuf); 1432 os = dn->dn_objset; 1433 object = dn->dn_object; 1434 DB_DNODE_EXIT(dbuf); 1435 1436 dbuf_rele(db, FTAG); 1437 dmu_write(os, object, offset, blksz, buf->b_data, tx); 1438 dmu_return_arcbuf(buf); 1439 XUIOSTAT_BUMP(xuiostat_wbuf_copied); 1440 } 1441} 1442 1443typedef struct { 1444 dbuf_dirty_record_t *dsa_dr; 1445 dmu_sync_cb_t *dsa_done; 1446 zgd_t *dsa_zgd; 1447 dmu_tx_t *dsa_tx; 1448} dmu_sync_arg_t; 1449 1450/* ARGSUSED */ 1451static void 1452dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg) 1453{ 1454 dmu_sync_arg_t *dsa = varg; 1455 dmu_buf_t *db = dsa->dsa_zgd->zgd_db; 1456 blkptr_t *bp = zio->io_bp; 1457 1458 if (zio->io_error == 0) { 1459 if (BP_IS_HOLE(bp)) { 1460 /* 1461 * A block of zeros may compress to a hole, but the 1462 * block size still needs to be known for replay. 1463 */ 1464 BP_SET_LSIZE(bp, db->db_size); 1465 } else if (!BP_IS_EMBEDDED(bp)) { 1466 ASSERT(BP_GET_LEVEL(bp) == 0); 1467 bp->blk_fill = 1; 1468 } 1469 } 1470} 1471 1472static void 1473dmu_sync_late_arrival_ready(zio_t *zio) 1474{ 1475 dmu_sync_ready(zio, NULL, zio->io_private); 1476} 1477 1478/* ARGSUSED */ 1479static void 1480dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg) 1481{ 1482 dmu_sync_arg_t *dsa = varg; 1483 dbuf_dirty_record_t *dr = dsa->dsa_dr; 1484 dmu_buf_impl_t *db = dr->dr_dbuf; 1485 1486 mutex_enter(&db->db_mtx); 1487 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC); 1488 if (zio->io_error == 0) { 1489 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE); 1490 if (dr->dt.dl.dr_nopwrite) { 1491 blkptr_t *bp = zio->io_bp; 1492 blkptr_t *bp_orig = &zio->io_bp_orig; 1493 uint8_t chksum = BP_GET_CHECKSUM(bp_orig); 1494 1495 ASSERT(BP_EQUAL(bp, bp_orig)); 1496 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF); 1497 ASSERT(zio_checksum_table[chksum].ci_flags & 1498 ZCHECKSUM_FLAG_NOPWRITE); 1499 } 1500 dr->dt.dl.dr_overridden_by = *zio->io_bp; 1501 dr->dt.dl.dr_override_state = DR_OVERRIDDEN; 1502 dr->dt.dl.dr_copies = zio->io_prop.zp_copies; 1503 1504 /* 1505 * Old style holes are filled with all zeros, whereas 1506 * new-style holes maintain their lsize, type, level, 1507 * and birth time (see zio_write_compress). While we 1508 * need to reset the BP_SET_LSIZE() call that happened 1509 * in dmu_sync_ready for old style holes, we do *not* 1510 * want to wipe out the information contained in new 1511 * style holes. Thus, only zero out the block pointer if 1512 * it's an old style hole. 1513 */ 1514 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) && 1515 dr->dt.dl.dr_overridden_by.blk_birth == 0) 1516 BP_ZERO(&dr->dt.dl.dr_overridden_by); 1517 } else { 1518 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 1519 } 1520 cv_broadcast(&db->db_changed); 1521 mutex_exit(&db->db_mtx); 1522 1523 dsa->dsa_done(dsa->dsa_zgd, zio->io_error); 1524 1525 kmem_free(dsa, sizeof (*dsa)); 1526} 1527 1528static void 1529dmu_sync_late_arrival_done(zio_t *zio) 1530{ 1531 blkptr_t *bp = zio->io_bp; 1532 dmu_sync_arg_t *dsa = zio->io_private; 1533 blkptr_t *bp_orig = &zio->io_bp_orig; 1534 1535 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) { 1536 /* 1537 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE) 1538 * then there is nothing to do here. Otherwise, free the 1539 * newly allocated block in this txg. 1540 */ 1541 if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 1542 ASSERT(BP_EQUAL(bp, bp_orig)); 1543 } else { 1544 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig)); 1545 ASSERT(zio->io_bp->blk_birth == zio->io_txg); 1546 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa)); 1547 zio_free(zio->io_spa, zio->io_txg, zio->io_bp); 1548 } 1549 } 1550 1551 dmu_tx_commit(dsa->dsa_tx); 1552 1553 dsa->dsa_done(dsa->dsa_zgd, zio->io_error); 1554 1555 kmem_free(dsa, sizeof (*dsa)); 1556} 1557 1558static int 1559dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd, 1560 zio_prop_t *zp, zbookmark_phys_t *zb) 1561{ 1562 dmu_sync_arg_t *dsa; 1563 dmu_tx_t *tx; 1564 1565 tx = dmu_tx_create(os); 1566 dmu_tx_hold_space(tx, zgd->zgd_db->db_size); 1567 if (dmu_tx_assign(tx, TXG_WAIT) != 0) { 1568 dmu_tx_abort(tx); 1569 /* Make zl_get_data do txg_waited_synced() */ 1570 return (SET_ERROR(EIO)); 1571 } 1572 1573 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP); 1574 dsa->dsa_dr = NULL; 1575 dsa->dsa_done = done; 1576 dsa->dsa_zgd = zgd; 1577 dsa->dsa_tx = tx; 1578 1579 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp, 1580 zgd->zgd_db->db_data, zgd->zgd_db->db_size, zp, 1581 dmu_sync_late_arrival_ready, NULL, dmu_sync_late_arrival_done, dsa, 1582 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb)); 1583 1584 return (0); 1585} 1586 1587/* 1588 * Intent log support: sync the block associated with db to disk. 1589 * N.B. and XXX: the caller is responsible for making sure that the 1590 * data isn't changing while dmu_sync() is writing it. 1591 * 1592 * Return values: 1593 * 1594 * EEXIST: this txg has already been synced, so there's nothing to do. 1595 * The caller should not log the write. 1596 * 1597 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do. 1598 * The caller should not log the write. 1599 * 1600 * EALREADY: this block is already in the process of being synced. 1601 * The caller should track its progress (somehow). 1602 * 1603 * EIO: could not do the I/O. 1604 * The caller should do a txg_wait_synced(). 1605 * 1606 * 0: the I/O has been initiated. 1607 * The caller should log this blkptr in the done callback. 1608 * It is possible that the I/O will fail, in which case 1609 * the error will be reported to the done callback and 1610 * propagated to pio from zio_done(). 1611 */ 1612int 1613dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd) 1614{ 1615 blkptr_t *bp = zgd->zgd_bp; 1616 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db; 1617 objset_t *os = db->db_objset; 1618 dsl_dataset_t *ds = os->os_dsl_dataset; 1619 dbuf_dirty_record_t *dr; 1620 dmu_sync_arg_t *dsa; 1621 zbookmark_phys_t zb; 1622 zio_prop_t zp; 1623 dnode_t *dn; 1624 1625 ASSERT(pio != NULL); 1626 ASSERT(txg != 0); 1627 1628 SET_BOOKMARK(&zb, ds->ds_object, 1629 db->db.db_object, db->db_level, db->db_blkid); 1630 1631 DB_DNODE_ENTER(db); 1632 dn = DB_DNODE(db); 1633 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp); 1634 DB_DNODE_EXIT(db); 1635 1636 /* 1637 * If we're frozen (running ziltest), we always need to generate a bp. 1638 */ 1639 if (txg > spa_freeze_txg(os->os_spa)) 1640 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); 1641 1642 /* 1643 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf() 1644 * and us. If we determine that this txg is not yet syncing, 1645 * but it begins to sync a moment later, that's OK because the 1646 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx. 1647 */ 1648 mutex_enter(&db->db_mtx); 1649 1650 if (txg <= spa_last_synced_txg(os->os_spa)) { 1651 /* 1652 * This txg has already synced. There's nothing to do. 1653 */ 1654 mutex_exit(&db->db_mtx); 1655 return (SET_ERROR(EEXIST)); 1656 } 1657 1658 if (txg <= spa_syncing_txg(os->os_spa)) { 1659 /* 1660 * This txg is currently syncing, so we can't mess with 1661 * the dirty record anymore; just write a new log block. 1662 */ 1663 mutex_exit(&db->db_mtx); 1664 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); 1665 } 1666 1667 dr = db->db_last_dirty; 1668 while (dr && dr->dr_txg != txg) 1669 dr = dr->dr_next; 1670 1671 if (dr == NULL) { 1672 /* 1673 * There's no dr for this dbuf, so it must have been freed. 1674 * There's no need to log writes to freed blocks, so we're done. 1675 */ 1676 mutex_exit(&db->db_mtx); 1677 return (SET_ERROR(ENOENT)); 1678 } 1679 1680 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg); 1681 1682 /* 1683 * Assume the on-disk data is X, the current syncing data (in 1684 * txg - 1) is Y, and the current in-memory data is Z (currently 1685 * in dmu_sync). 1686 * 1687 * We usually want to perform a nopwrite if X and Z are the 1688 * same. However, if Y is different (i.e. the BP is going to 1689 * change before this write takes effect), then a nopwrite will 1690 * be incorrect - we would override with X, which could have 1691 * been freed when Y was written. 1692 * 1693 * (Note that this is not a concern when we are nop-writing from 1694 * syncing context, because X and Y must be identical, because 1695 * all previous txgs have been synced.) 1696 * 1697 * Therefore, we disable nopwrite if the current BP could change 1698 * before this TXG. There are two ways it could change: by 1699 * being dirty (dr_next is non-NULL), or by being freed 1700 * (dnode_block_freed()). This behavior is verified by 1701 * zio_done(), which VERIFYs that the override BP is identical 1702 * to the on-disk BP. 1703 */ 1704 DB_DNODE_ENTER(db); 1705 dn = DB_DNODE(db); 1706 if (dr->dr_next != NULL || dnode_block_freed(dn, db->db_blkid)) 1707 zp.zp_nopwrite = B_FALSE; 1708 DB_DNODE_EXIT(db); 1709 1710 ASSERT(dr->dr_txg == txg); 1711 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC || 1712 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { 1713 /* 1714 * We have already issued a sync write for this buffer, 1715 * or this buffer has already been synced. It could not 1716 * have been dirtied since, or we would have cleared the state. 1717 */ 1718 mutex_exit(&db->db_mtx); 1719 return (SET_ERROR(EALREADY)); 1720 } 1721 1722 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); 1723 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC; 1724 mutex_exit(&db->db_mtx); 1725 1726 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP); 1727 dsa->dsa_dr = dr; 1728 dsa->dsa_done = done; 1729 dsa->dsa_zgd = zgd; 1730 dsa->dsa_tx = NULL; 1731 1732 zio_nowait(arc_write(pio, os->os_spa, txg, 1733 bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db), 1734 DBUF_IS_L2COMPRESSIBLE(db), &zp, dmu_sync_ready, 1735 NULL, dmu_sync_done, dsa, ZIO_PRIORITY_SYNC_WRITE, 1736 ZIO_FLAG_CANFAIL, &zb)); 1737 1738 return (0); 1739} 1740 1741int 1742dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs, 1743 dmu_tx_t *tx) 1744{ 1745 dnode_t *dn; 1746 int err; 1747 1748 err = dnode_hold(os, object, FTAG, &dn); 1749 if (err) 1750 return (err); 1751 err = dnode_set_blksz(dn, size, ibs, tx); 1752 dnode_rele(dn, FTAG); 1753 return (err); 1754} 1755 1756void 1757dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum, 1758 dmu_tx_t *tx) 1759{ 1760 dnode_t *dn; 1761 1762 /* 1763 * Send streams include each object's checksum function. This 1764 * check ensures that the receiving system can understand the 1765 * checksum function transmitted. 1766 */ 1767 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS); 1768 1769 VERIFY0(dnode_hold(os, object, FTAG, &dn)); 1770 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS); 1771 dn->dn_checksum = checksum; 1772 dnode_setdirty(dn, tx); 1773 dnode_rele(dn, FTAG); 1774} 1775 1776void 1777dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress, 1778 dmu_tx_t *tx) 1779{ 1780 dnode_t *dn; 1781 1782 /* 1783 * Send streams include each object's compression function. This 1784 * check ensures that the receiving system can understand the 1785 * compression function transmitted. 1786 */ 1787 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS); 1788 1789 VERIFY0(dnode_hold(os, object, FTAG, &dn)); 1790 dn->dn_compress = compress; 1791 dnode_setdirty(dn, tx); 1792 dnode_rele(dn, FTAG); 1793} 1794 1795int zfs_mdcomp_disable = 0; 1796TUNABLE_INT("vfs.zfs.mdcomp_disable", &zfs_mdcomp_disable); 1797SYSCTL_INT(_vfs_zfs, OID_AUTO, mdcomp_disable, CTLFLAG_RW, 1798 &zfs_mdcomp_disable, 0, "Disable metadata compression"); 1799 1800/* 1801 * When the "redundant_metadata" property is set to "most", only indirect 1802 * blocks of this level and higher will have an additional ditto block. 1803 */ 1804int zfs_redundant_metadata_most_ditto_level = 2; 1805 1806void 1807dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp) 1808{ 1809 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET; 1810 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) || 1811 (wp & WP_SPILL)); 1812 enum zio_checksum checksum = os->os_checksum; 1813 enum zio_compress compress = os->os_compress; 1814 enum zio_checksum dedup_checksum = os->os_dedup_checksum; 1815 boolean_t dedup = B_FALSE; 1816 boolean_t nopwrite = B_FALSE; 1817 boolean_t dedup_verify = os->os_dedup_verify; 1818 int copies = os->os_copies; 1819 1820 /* 1821 * We maintain different write policies for each of the following 1822 * types of data: 1823 * 1. metadata 1824 * 2. preallocated blocks (i.e. level-0 blocks of a dump device) 1825 * 3. all other level 0 blocks 1826 */ 1827 if (ismd) { 1828 if (zfs_mdcomp_disable) { 1829 compress = ZIO_COMPRESS_EMPTY; 1830 } else { 1831 /* 1832 * XXX -- we should design a compression algorithm 1833 * that specializes in arrays of bps. 1834 */ 1835 compress = zio_compress_select(os->os_spa, 1836 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON); 1837 } 1838 1839 /* 1840 * Metadata always gets checksummed. If the data 1841 * checksum is multi-bit correctable, and it's not a 1842 * ZBT-style checksum, then it's suitable for metadata 1843 * as well. Otherwise, the metadata checksum defaults 1844 * to fletcher4. 1845 */ 1846 if (!(zio_checksum_table[checksum].ci_flags & 1847 ZCHECKSUM_FLAG_METADATA) || 1848 (zio_checksum_table[checksum].ci_flags & 1849 ZCHECKSUM_FLAG_EMBEDDED)) 1850 checksum = ZIO_CHECKSUM_FLETCHER_4; 1851 1852 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL || 1853 (os->os_redundant_metadata == 1854 ZFS_REDUNDANT_METADATA_MOST && 1855 (level >= zfs_redundant_metadata_most_ditto_level || 1856 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL)))) 1857 copies++; 1858 } else if (wp & WP_NOFILL) { 1859 ASSERT(level == 0); 1860 1861 /* 1862 * If we're writing preallocated blocks, we aren't actually 1863 * writing them so don't set any policy properties. These 1864 * blocks are currently only used by an external subsystem 1865 * outside of zfs (i.e. dump) and not written by the zio 1866 * pipeline. 1867 */ 1868 compress = ZIO_COMPRESS_OFF; 1869 checksum = ZIO_CHECKSUM_NOPARITY; 1870 } else { 1871 compress = zio_compress_select(os->os_spa, dn->dn_compress, 1872 compress); 1873 1874 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ? 1875 zio_checksum_select(dn->dn_checksum, checksum) : 1876 dedup_checksum; 1877 1878 /* 1879 * Determine dedup setting. If we are in dmu_sync(), 1880 * we won't actually dedup now because that's all 1881 * done in syncing context; but we do want to use the 1882 * dedup checkum. If the checksum is not strong 1883 * enough to ensure unique signatures, force 1884 * dedup_verify. 1885 */ 1886 if (dedup_checksum != ZIO_CHECKSUM_OFF) { 1887 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE; 1888 if (!(zio_checksum_table[checksum].ci_flags & 1889 ZCHECKSUM_FLAG_DEDUP)) 1890 dedup_verify = B_TRUE; 1891 } 1892 1893 /* 1894 * Enable nopwrite if we have secure enough checksum 1895 * algorithm (see comment in zio_nop_write) and 1896 * compression is enabled. We don't enable nopwrite if 1897 * dedup is enabled as the two features are mutually 1898 * exclusive. 1899 */ 1900 nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags & 1901 ZCHECKSUM_FLAG_NOPWRITE) && 1902 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled); 1903 } 1904 1905 zp->zp_checksum = checksum; 1906 zp->zp_compress = compress; 1907 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type; 1908 zp->zp_level = level; 1909 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa)); 1910 zp->zp_dedup = dedup; 1911 zp->zp_dedup_verify = dedup && dedup_verify; 1912 zp->zp_nopwrite = nopwrite; 1913} 1914 1915int 1916dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off) 1917{ 1918 dnode_t *dn; 1919 int err; 1920 1921 /* 1922 * Sync any current changes before 1923 * we go trundling through the block pointers. 1924 */ 1925 err = dmu_object_wait_synced(os, object); 1926 if (err) { 1927 return (err); 1928 } 1929 1930 err = dnode_hold(os, object, FTAG, &dn); 1931 if (err) { 1932 return (err); 1933 } 1934 1935 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0); 1936 dnode_rele(dn, FTAG); 1937 1938 return (err); 1939} 1940 1941/* 1942 * Given the ZFS object, if it contains any dirty nodes 1943 * this function flushes all dirty blocks to disk. This 1944 * ensures the DMU object info is updated. A more efficient 1945 * future version might just find the TXG with the maximum 1946 * ID and wait for that to be synced. 1947 */ 1948int 1949dmu_object_wait_synced(objset_t *os, uint64_t object) { 1950 dnode_t *dn; 1951 int error, i; 1952 1953 error = dnode_hold(os, object, FTAG, &dn); 1954 if (error) { 1955 return (error); 1956 } 1957 1958 for (i = 0; i < TXG_SIZE; i++) { 1959 if (list_link_active(&dn->dn_dirty_link[i])) { 1960 break; 1961 } 1962 } 1963 dnode_rele(dn, FTAG); 1964 if (i != TXG_SIZE) { 1965 txg_wait_synced(dmu_objset_pool(os), 0); 1966 } 1967 1968 return (0); 1969} 1970 1971void 1972dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi) 1973{ 1974 dnode_phys_t *dnp; 1975 1976 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1977 mutex_enter(&dn->dn_mtx); 1978 1979 dnp = dn->dn_phys; 1980 1981 doi->doi_data_block_size = dn->dn_datablksz; 1982 doi->doi_metadata_block_size = dn->dn_indblkshift ? 1983 1ULL << dn->dn_indblkshift : 0; 1984 doi->doi_type = dn->dn_type; 1985 doi->doi_bonus_type = dn->dn_bonustype; 1986 doi->doi_bonus_size = dn->dn_bonuslen; 1987 doi->doi_indirection = dn->dn_nlevels; 1988 doi->doi_checksum = dn->dn_checksum; 1989 doi->doi_compress = dn->dn_compress; 1990 doi->doi_nblkptr = dn->dn_nblkptr; 1991 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9; 1992 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz; 1993 doi->doi_fill_count = 0; 1994 for (int i = 0; i < dnp->dn_nblkptr; i++) 1995 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]); 1996 1997 mutex_exit(&dn->dn_mtx); 1998 rw_exit(&dn->dn_struct_rwlock); 1999} 2000 2001/* 2002 * Get information on a DMU object. 2003 * If doi is NULL, just indicates whether the object exists. 2004 */ 2005int 2006dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi) 2007{ 2008 dnode_t *dn; 2009 int err = dnode_hold(os, object, FTAG, &dn); 2010 2011 if (err) 2012 return (err); 2013 2014 if (doi != NULL) 2015 dmu_object_info_from_dnode(dn, doi); 2016 2017 dnode_rele(dn, FTAG); 2018 return (0); 2019} 2020 2021/* 2022 * As above, but faster; can be used when you have a held dbuf in hand. 2023 */ 2024void 2025dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi) 2026{ 2027 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2028 2029 DB_DNODE_ENTER(db); 2030 dmu_object_info_from_dnode(DB_DNODE(db), doi); 2031 DB_DNODE_EXIT(db); 2032} 2033 2034/* 2035 * Faster still when you only care about the size. 2036 * This is specifically optimized for zfs_getattr(). 2037 */ 2038void 2039dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize, 2040 u_longlong_t *nblk512) 2041{ 2042 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2043 dnode_t *dn; 2044 2045 DB_DNODE_ENTER(db); 2046 dn = DB_DNODE(db); 2047 2048 *blksize = dn->dn_datablksz; 2049 /* add 1 for dnode space */ 2050 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >> 2051 SPA_MINBLOCKSHIFT) + 1; 2052 DB_DNODE_EXIT(db); 2053} 2054 2055void 2056byteswap_uint64_array(void *vbuf, size_t size) 2057{ 2058 uint64_t *buf = vbuf; 2059 size_t count = size >> 3; 2060 int i; 2061 2062 ASSERT((size & 7) == 0); 2063 2064 for (i = 0; i < count; i++) 2065 buf[i] = BSWAP_64(buf[i]); 2066} 2067 2068void 2069byteswap_uint32_array(void *vbuf, size_t size) 2070{ 2071 uint32_t *buf = vbuf; 2072 size_t count = size >> 2; 2073 int i; 2074 2075 ASSERT((size & 3) == 0); 2076 2077 for (i = 0; i < count; i++) 2078 buf[i] = BSWAP_32(buf[i]); 2079} 2080 2081void 2082byteswap_uint16_array(void *vbuf, size_t size) 2083{ 2084 uint16_t *buf = vbuf; 2085 size_t count = size >> 1; 2086 int i; 2087 2088 ASSERT((size & 1) == 0); 2089 2090 for (i = 0; i < count; i++) 2091 buf[i] = BSWAP_16(buf[i]); 2092} 2093 2094/* ARGSUSED */ 2095void 2096byteswap_uint8_array(void *vbuf, size_t size) 2097{ 2098} 2099 2100void 2101dmu_init(void) 2102{ 2103 zfs_dbgmsg_init(); 2104 sa_cache_init(); 2105 xuio_stat_init(); 2106 dmu_objset_init(); 2107 dnode_init(); 2108 dbuf_init(); 2109 zfetch_init(); 2110 zio_compress_init(); 2111 l2arc_init(); 2112 arc_init(); 2113} 2114 2115void 2116dmu_fini(void) 2117{ 2118 arc_fini(); /* arc depends on l2arc, so arc must go first */ 2119 l2arc_fini(); 2120 zfetch_fini(); 2121 zio_compress_fini(); 2122 dbuf_fini(); 2123 dnode_fini(); 2124 dmu_objset_fini(); 2125 xuio_stat_fini(); 2126 sa_cache_fini(); 2127 zfs_dbgmsg_fini(); 2128} 2129