dmu.c revision 299433
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 && 453 DNODE_META_IS_CACHEABLE(dn) && length <= zfetch_array_rd_sz) { 454 dmu_zfetch(&dn->dn_zfetch, blkid, nblks, 455 read && DNODE_IS_CACHEABLE(dn)); 456 } 457 rw_exit(&dn->dn_struct_rwlock); 458 459 /* wait for async i/o */ 460 err = zio_wait(zio); 461 if (err) { 462 dmu_buf_rele_array(dbp, nblks, tag); 463 return (err); 464 } 465 466 /* wait for other io to complete */ 467 if (read) { 468 for (i = 0; i < nblks; i++) { 469 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i]; 470 mutex_enter(&db->db_mtx); 471 while (db->db_state == DB_READ || 472 db->db_state == DB_FILL) 473 cv_wait(&db->db_changed, &db->db_mtx); 474 if (db->db_state == DB_UNCACHED) 475 err = SET_ERROR(EIO); 476 mutex_exit(&db->db_mtx); 477 if (err) { 478 dmu_buf_rele_array(dbp, nblks, tag); 479 return (err); 480 } 481 } 482 } 483 484 *numbufsp = nblks; 485 *dbpp = dbp; 486 return (0); 487} 488 489static int 490dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset, 491 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp) 492{ 493 dnode_t *dn; 494 int err; 495 496 err = dnode_hold(os, object, FTAG, &dn); 497 if (err) 498 return (err); 499 500 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag, 501 numbufsp, dbpp, DMU_READ_PREFETCH); 502 503 dnode_rele(dn, FTAG); 504 505 return (err); 506} 507 508int 509dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset, 510 uint64_t length, boolean_t read, void *tag, int *numbufsp, 511 dmu_buf_t ***dbpp) 512{ 513 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 514 dnode_t *dn; 515 int err; 516 517 DB_DNODE_ENTER(db); 518 dn = DB_DNODE(db); 519 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag, 520 numbufsp, dbpp, DMU_READ_PREFETCH); 521 DB_DNODE_EXIT(db); 522 523 return (err); 524} 525 526void 527dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag) 528{ 529 int i; 530 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake; 531 532 if (numbufs == 0) 533 return; 534 535 for (i = 0; i < numbufs; i++) { 536 if (dbp[i]) 537 dbuf_rele(dbp[i], tag); 538 } 539 540 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs); 541} 542 543/* 544 * Issue prefetch i/os for the given blocks. If level is greater than 0, the 545 * indirect blocks prefeteched will be those that point to the blocks containing 546 * the data starting at offset, and continuing to offset + len. 547 * 548 * Note that if the indirect blocks above the blocks being prefetched are not in 549 * cache, they will be asychronously read in. 550 */ 551void 552dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset, 553 uint64_t len, zio_priority_t pri) 554{ 555 dnode_t *dn; 556 uint64_t blkid; 557 int nblks, err; 558 559 if (len == 0) { /* they're interested in the bonus buffer */ 560 dn = DMU_META_DNODE(os); 561 562 if (object == 0 || object >= DN_MAX_OBJECT) 563 return; 564 565 rw_enter(&dn->dn_struct_rwlock, RW_READER); 566 blkid = dbuf_whichblock(dn, level, 567 object * sizeof (dnode_phys_t)); 568 dbuf_prefetch(dn, level, blkid, pri, 0); 569 rw_exit(&dn->dn_struct_rwlock); 570 return; 571 } 572 573 /* 574 * XXX - Note, if the dnode for the requested object is not 575 * already cached, we will do a *synchronous* read in the 576 * dnode_hold() call. The same is true for any indirects. 577 */ 578 err = dnode_hold(os, object, FTAG, &dn); 579 if (err != 0) 580 return; 581 582 rw_enter(&dn->dn_struct_rwlock, RW_READER); 583 /* 584 * offset + len - 1 is the last byte we want to prefetch for, and offset 585 * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the 586 * last block we want to prefetch, and dbuf_whichblock(dn, level, 587 * offset) is the first. Then the number we need to prefetch is the 588 * last - first + 1. 589 */ 590 if (level > 0 || dn->dn_datablkshift != 0) { 591 nblks = dbuf_whichblock(dn, level, offset + len - 1) - 592 dbuf_whichblock(dn, level, offset) + 1; 593 } else { 594 nblks = (offset < dn->dn_datablksz); 595 } 596 597 if (nblks != 0) { 598 blkid = dbuf_whichblock(dn, level, offset); 599 for (int i = 0; i < nblks; i++) 600 dbuf_prefetch(dn, level, blkid + i, pri, 0); 601 } 602 603 rw_exit(&dn->dn_struct_rwlock); 604 605 dnode_rele(dn, FTAG); 606} 607 608/* 609 * Get the next "chunk" of file data to free. We traverse the file from 610 * the end so that the file gets shorter over time (if we crashes in the 611 * middle, this will leave us in a better state). We find allocated file 612 * data by simply searching the allocated level 1 indirects. 613 * 614 * On input, *start should be the first offset that does not need to be 615 * freed (e.g. "offset + length"). On return, *start will be the first 616 * offset that should be freed. 617 */ 618static int 619get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum) 620{ 621 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1); 622 /* bytes of data covered by a level-1 indirect block */ 623 uint64_t iblkrange = 624 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT); 625 626 ASSERT3U(minimum, <=, *start); 627 628 if (*start - minimum <= iblkrange * maxblks) { 629 *start = minimum; 630 return (0); 631 } 632 ASSERT(ISP2(iblkrange)); 633 634 for (uint64_t blks = 0; *start > minimum && blks < maxblks; blks++) { 635 int err; 636 637 /* 638 * dnode_next_offset(BACKWARDS) will find an allocated L1 639 * indirect block at or before the input offset. We must 640 * decrement *start so that it is at the end of the region 641 * to search. 642 */ 643 (*start)--; 644 err = dnode_next_offset(dn, 645 DNODE_FIND_BACKWARDS, start, 2, 1, 0); 646 647 /* if there are no indirect blocks before start, we are done */ 648 if (err == ESRCH) { 649 *start = minimum; 650 break; 651 } else if (err != 0) { 652 return (err); 653 } 654 655 /* set start to the beginning of this L1 indirect */ 656 *start = P2ALIGN(*start, iblkrange); 657 } 658 if (*start < minimum) 659 *start = minimum; 660 return (0); 661} 662 663static int 664dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset, 665 uint64_t length) 666{ 667 uint64_t object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz; 668 int err; 669 670 if (offset >= object_size) 671 return (0); 672 673 if (length == DMU_OBJECT_END || offset + length > object_size) 674 length = object_size - offset; 675 676 while (length != 0) { 677 uint64_t chunk_end, chunk_begin; 678 679 chunk_end = chunk_begin = offset + length; 680 681 /* move chunk_begin backwards to the beginning of this chunk */ 682 err = get_next_chunk(dn, &chunk_begin, offset); 683 if (err) 684 return (err); 685 ASSERT3U(chunk_begin, >=, offset); 686 ASSERT3U(chunk_begin, <=, chunk_end); 687 688 dmu_tx_t *tx = dmu_tx_create(os); 689 dmu_tx_hold_free(tx, dn->dn_object, 690 chunk_begin, chunk_end - chunk_begin); 691 692 /* 693 * Mark this transaction as typically resulting in a net 694 * reduction in space used. 695 */ 696 dmu_tx_mark_netfree(tx); 697 err = dmu_tx_assign(tx, TXG_WAIT); 698 if (err) { 699 dmu_tx_abort(tx); 700 return (err); 701 } 702 dnode_free_range(dn, chunk_begin, chunk_end - chunk_begin, tx); 703 dmu_tx_commit(tx); 704 705 length -= chunk_end - chunk_begin; 706 } 707 return (0); 708} 709 710int 711dmu_free_long_range(objset_t *os, uint64_t object, 712 uint64_t offset, uint64_t length) 713{ 714 dnode_t *dn; 715 int err; 716 717 err = dnode_hold(os, object, FTAG, &dn); 718 if (err != 0) 719 return (err); 720 err = dmu_free_long_range_impl(os, dn, offset, length); 721 722 /* 723 * It is important to zero out the maxblkid when freeing the entire 724 * file, so that (a) subsequent calls to dmu_free_long_range_impl() 725 * will take the fast path, and (b) dnode_reallocate() can verify 726 * that the entire file has been freed. 727 */ 728 if (err == 0 && offset == 0 && length == DMU_OBJECT_END) 729 dn->dn_maxblkid = 0; 730 731 dnode_rele(dn, FTAG); 732 return (err); 733} 734 735int 736dmu_free_long_object(objset_t *os, uint64_t object) 737{ 738 dmu_tx_t *tx; 739 int err; 740 741 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END); 742 if (err != 0) 743 return (err); 744 745 tx = dmu_tx_create(os); 746 dmu_tx_hold_bonus(tx, object); 747 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END); 748 dmu_tx_mark_netfree(tx); 749 err = dmu_tx_assign(tx, TXG_WAIT); 750 if (err == 0) { 751 err = dmu_object_free(os, object, tx); 752 dmu_tx_commit(tx); 753 } else { 754 dmu_tx_abort(tx); 755 } 756 757 return (err); 758} 759 760int 761dmu_free_range(objset_t *os, uint64_t object, uint64_t offset, 762 uint64_t size, dmu_tx_t *tx) 763{ 764 dnode_t *dn; 765 int err = dnode_hold(os, object, FTAG, &dn); 766 if (err) 767 return (err); 768 ASSERT(offset < UINT64_MAX); 769 ASSERT(size == -1ULL || size <= UINT64_MAX - offset); 770 dnode_free_range(dn, offset, size, tx); 771 dnode_rele(dn, FTAG); 772 return (0); 773} 774 775int 776dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 777 void *buf, uint32_t flags) 778{ 779 dnode_t *dn; 780 dmu_buf_t **dbp; 781 int numbufs, err; 782 783 err = dnode_hold(os, object, FTAG, &dn); 784 if (err) 785 return (err); 786 787 /* 788 * Deal with odd block sizes, where there can't be data past the first 789 * block. If we ever do the tail block optimization, we will need to 790 * handle that here as well. 791 */ 792 if (dn->dn_maxblkid == 0) { 793 int newsz = offset > dn->dn_datablksz ? 0 : 794 MIN(size, dn->dn_datablksz - offset); 795 bzero((char *)buf + newsz, size - newsz); 796 size = newsz; 797 } 798 799 while (size > 0) { 800 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2); 801 int i; 802 803 /* 804 * NB: we could do this block-at-a-time, but it's nice 805 * to be reading in parallel. 806 */ 807 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen, 808 TRUE, FTAG, &numbufs, &dbp, flags); 809 if (err) 810 break; 811 812 for (i = 0; i < numbufs; i++) { 813 int tocpy; 814 int bufoff; 815 dmu_buf_t *db = dbp[i]; 816 817 ASSERT(size > 0); 818 819 bufoff = offset - db->db_offset; 820 tocpy = (int)MIN(db->db_size - bufoff, size); 821 822 bcopy((char *)db->db_data + bufoff, buf, tocpy); 823 824 offset += tocpy; 825 size -= tocpy; 826 buf = (char *)buf + tocpy; 827 } 828 dmu_buf_rele_array(dbp, numbufs, FTAG); 829 } 830 dnode_rele(dn, FTAG); 831 return (err); 832} 833 834void 835dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 836 const void *buf, 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 int tocpy; 849 int bufoff; 850 dmu_buf_t *db = dbp[i]; 851 852 ASSERT(size > 0); 853 854 bufoff = offset - db->db_offset; 855 tocpy = (int)MIN(db->db_size - bufoff, size); 856 857 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 858 859 if (tocpy == db->db_size) 860 dmu_buf_will_fill(db, tx); 861 else 862 dmu_buf_will_dirty(db, tx); 863 864 bcopy(buf, (char *)db->db_data + bufoff, tocpy); 865 866 if (tocpy == db->db_size) 867 dmu_buf_fill_done(db, tx); 868 869 offset += tocpy; 870 size -= tocpy; 871 buf = (char *)buf + tocpy; 872 } 873 dmu_buf_rele_array(dbp, numbufs, FTAG); 874} 875 876void 877dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 878 dmu_tx_t *tx) 879{ 880 dmu_buf_t **dbp; 881 int numbufs, i; 882 883 if (size == 0) 884 return; 885 886 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size, 887 FALSE, FTAG, &numbufs, &dbp)); 888 889 for (i = 0; i < numbufs; i++) { 890 dmu_buf_t *db = dbp[i]; 891 892 dmu_buf_will_not_fill(db, tx); 893 } 894 dmu_buf_rele_array(dbp, numbufs, FTAG); 895} 896 897void 898dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset, 899 void *data, uint8_t etype, uint8_t comp, int uncompressed_size, 900 int compressed_size, int byteorder, dmu_tx_t *tx) 901{ 902 dmu_buf_t *db; 903 904 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES); 905 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS); 906 VERIFY0(dmu_buf_hold_noread(os, object, offset, 907 FTAG, &db)); 908 909 dmu_buf_write_embedded(db, 910 data, (bp_embedded_type_t)etype, (enum zio_compress)comp, 911 uncompressed_size, compressed_size, byteorder, tx); 912 913 dmu_buf_rele(db, FTAG); 914} 915 916/* 917 * DMU support for xuio 918 */ 919kstat_t *xuio_ksp = NULL; 920 921int 922dmu_xuio_init(xuio_t *xuio, int nblk) 923{ 924 dmu_xuio_t *priv; 925 uio_t *uio = &xuio->xu_uio; 926 927 uio->uio_iovcnt = nblk; 928 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP); 929 930 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP); 931 priv->cnt = nblk; 932 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP); 933 priv->iovp = uio->uio_iov; 934 XUIO_XUZC_PRIV(xuio) = priv; 935 936 if (XUIO_XUZC_RW(xuio) == UIO_READ) 937 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk); 938 else 939 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk); 940 941 return (0); 942} 943 944void 945dmu_xuio_fini(xuio_t *xuio) 946{ 947 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 948 int nblk = priv->cnt; 949 950 kmem_free(priv->iovp, nblk * sizeof (iovec_t)); 951 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *)); 952 kmem_free(priv, sizeof (dmu_xuio_t)); 953 954 if (XUIO_XUZC_RW(xuio) == UIO_READ) 955 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk); 956 else 957 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk); 958} 959 960/* 961 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf } 962 * and increase priv->next by 1. 963 */ 964int 965dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n) 966{ 967 struct iovec *iov; 968 uio_t *uio = &xuio->xu_uio; 969 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 970 int i = priv->next++; 971 972 ASSERT(i < priv->cnt); 973 ASSERT(off + n <= arc_buf_size(abuf)); 974 iov = uio->uio_iov + i; 975 iov->iov_base = (char *)abuf->b_data + off; 976 iov->iov_len = n; 977 priv->bufs[i] = abuf; 978 return (0); 979} 980 981int 982dmu_xuio_cnt(xuio_t *xuio) 983{ 984 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 985 return (priv->cnt); 986} 987 988arc_buf_t * 989dmu_xuio_arcbuf(xuio_t *xuio, int i) 990{ 991 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 992 993 ASSERT(i < priv->cnt); 994 return (priv->bufs[i]); 995} 996 997void 998dmu_xuio_clear(xuio_t *xuio, int i) 999{ 1000 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 1001 1002 ASSERT(i < priv->cnt); 1003 priv->bufs[i] = NULL; 1004} 1005 1006static void 1007xuio_stat_init(void) 1008{ 1009 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc", 1010 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t), 1011 KSTAT_FLAG_VIRTUAL); 1012 if (xuio_ksp != NULL) { 1013 xuio_ksp->ks_data = &xuio_stats; 1014 kstat_install(xuio_ksp); 1015 } 1016} 1017 1018static void 1019xuio_stat_fini(void) 1020{ 1021 if (xuio_ksp != NULL) { 1022 kstat_delete(xuio_ksp); 1023 xuio_ksp = NULL; 1024 } 1025} 1026 1027void 1028xuio_stat_wbuf_copied() 1029{ 1030 XUIOSTAT_BUMP(xuiostat_wbuf_copied); 1031} 1032 1033void 1034xuio_stat_wbuf_nocopy() 1035{ 1036 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy); 1037} 1038 1039#ifdef _KERNEL 1040static int 1041dmu_read_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size) 1042{ 1043 dmu_buf_t **dbp; 1044 int numbufs, i, err; 1045 xuio_t *xuio = NULL; 1046 1047 /* 1048 * NB: we could do this block-at-a-time, but it's nice 1049 * to be reading in parallel. 1050 */ 1051 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size, 1052 TRUE, FTAG, &numbufs, &dbp, 0); 1053 if (err) 1054 return (err); 1055 1056#ifdef UIO_XUIO 1057 if (uio->uio_extflg == UIO_XUIO) 1058 xuio = (xuio_t *)uio; 1059#endif 1060 1061 for (i = 0; i < numbufs; i++) { 1062 int tocpy; 1063 int bufoff; 1064 dmu_buf_t *db = dbp[i]; 1065 1066 ASSERT(size > 0); 1067 1068 bufoff = uio->uio_loffset - db->db_offset; 1069 tocpy = (int)MIN(db->db_size - bufoff, size); 1070 1071 if (xuio) { 1072 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 1073 arc_buf_t *dbuf_abuf = dbi->db_buf; 1074 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi); 1075 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy); 1076 if (!err) { 1077 uio->uio_resid -= tocpy; 1078 uio->uio_loffset += tocpy; 1079 } 1080 1081 if (abuf == dbuf_abuf) 1082 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy); 1083 else 1084 XUIOSTAT_BUMP(xuiostat_rbuf_copied); 1085 } else { 1086 err = uiomove((char *)db->db_data + bufoff, tocpy, 1087 UIO_READ, uio); 1088 } 1089 if (err) 1090 break; 1091 1092 size -= tocpy; 1093 } 1094 dmu_buf_rele_array(dbp, numbufs, FTAG); 1095 1096 return (err); 1097} 1098 1099/* 1100 * Read 'size' bytes into the uio buffer. 1101 * From object zdb->db_object. 1102 * Starting at offset uio->uio_loffset. 1103 * 1104 * If the caller already has a dbuf in the target object 1105 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(), 1106 * because we don't have to find the dnode_t for the object. 1107 */ 1108int 1109dmu_read_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size) 1110{ 1111 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb; 1112 dnode_t *dn; 1113 int err; 1114 1115 if (size == 0) 1116 return (0); 1117 1118 DB_DNODE_ENTER(db); 1119 dn = DB_DNODE(db); 1120 err = dmu_read_uio_dnode(dn, uio, size); 1121 DB_DNODE_EXIT(db); 1122 1123 return (err); 1124} 1125 1126/* 1127 * Read 'size' bytes into the uio buffer. 1128 * From the specified object 1129 * Starting at offset uio->uio_loffset. 1130 */ 1131int 1132dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size) 1133{ 1134 dnode_t *dn; 1135 int err; 1136 1137 if (size == 0) 1138 return (0); 1139 1140 err = dnode_hold(os, object, FTAG, &dn); 1141 if (err) 1142 return (err); 1143 1144 err = dmu_read_uio_dnode(dn, uio, size); 1145 1146 dnode_rele(dn, FTAG); 1147 1148 return (err); 1149} 1150 1151static int 1152dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx) 1153{ 1154 dmu_buf_t **dbp; 1155 int numbufs; 1156 int err = 0; 1157 int i; 1158 1159 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size, 1160 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH); 1161 if (err) 1162 return (err); 1163 1164 for (i = 0; i < numbufs; i++) { 1165 int tocpy; 1166 int bufoff; 1167 dmu_buf_t *db = dbp[i]; 1168 1169 ASSERT(size > 0); 1170 1171 bufoff = uio->uio_loffset - db->db_offset; 1172 tocpy = (int)MIN(db->db_size - bufoff, size); 1173 1174 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 1175 1176 if (tocpy == db->db_size) 1177 dmu_buf_will_fill(db, tx); 1178 else 1179 dmu_buf_will_dirty(db, tx); 1180 1181 /* 1182 * XXX uiomove could block forever (eg. nfs-backed 1183 * pages). There needs to be a uiolockdown() function 1184 * to lock the pages in memory, so that uiomove won't 1185 * block. 1186 */ 1187 err = uiomove((char *)db->db_data + bufoff, tocpy, 1188 UIO_WRITE, uio); 1189 1190 if (tocpy == db->db_size) 1191 dmu_buf_fill_done(db, tx); 1192 1193 if (err) 1194 break; 1195 1196 size -= tocpy; 1197 } 1198 1199 dmu_buf_rele_array(dbp, numbufs, FTAG); 1200 return (err); 1201} 1202 1203/* 1204 * Write 'size' bytes from the uio buffer. 1205 * To object zdb->db_object. 1206 * Starting at offset uio->uio_loffset. 1207 * 1208 * If the caller already has a dbuf in the target object 1209 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(), 1210 * because we don't have to find the dnode_t for the object. 1211 */ 1212int 1213dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size, 1214 dmu_tx_t *tx) 1215{ 1216 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb; 1217 dnode_t *dn; 1218 int err; 1219 1220 if (size == 0) 1221 return (0); 1222 1223 DB_DNODE_ENTER(db); 1224 dn = DB_DNODE(db); 1225 err = dmu_write_uio_dnode(dn, uio, size, tx); 1226 DB_DNODE_EXIT(db); 1227 1228 return (err); 1229} 1230 1231/* 1232 * Write 'size' bytes from the uio buffer. 1233 * To the specified object. 1234 * Starting at offset uio->uio_loffset. 1235 */ 1236int 1237dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size, 1238 dmu_tx_t *tx) 1239{ 1240 dnode_t *dn; 1241 int err; 1242 1243 if (size == 0) 1244 return (0); 1245 1246 err = dnode_hold(os, object, FTAG, &dn); 1247 if (err) 1248 return (err); 1249 1250 err = dmu_write_uio_dnode(dn, uio, size, tx); 1251 1252 dnode_rele(dn, FTAG); 1253 1254 return (err); 1255} 1256 1257#ifdef illumos 1258int 1259dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1260 page_t *pp, dmu_tx_t *tx) 1261{ 1262 dmu_buf_t **dbp; 1263 int numbufs, i; 1264 int err; 1265 1266 if (size == 0) 1267 return (0); 1268 1269 err = dmu_buf_hold_array(os, object, offset, size, 1270 FALSE, FTAG, &numbufs, &dbp); 1271 if (err) 1272 return (err); 1273 1274 for (i = 0; i < numbufs; i++) { 1275 int tocpy, copied, thiscpy; 1276 int bufoff; 1277 dmu_buf_t *db = dbp[i]; 1278 caddr_t va; 1279 1280 ASSERT(size > 0); 1281 ASSERT3U(db->db_size, >=, PAGESIZE); 1282 1283 bufoff = offset - db->db_offset; 1284 tocpy = (int)MIN(db->db_size - bufoff, size); 1285 1286 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 1287 1288 if (tocpy == db->db_size) 1289 dmu_buf_will_fill(db, tx); 1290 else 1291 dmu_buf_will_dirty(db, tx); 1292 1293 for (copied = 0; copied < tocpy; copied += PAGESIZE) { 1294 ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff); 1295 thiscpy = MIN(PAGESIZE, tocpy - copied); 1296 va = zfs_map_page(pp, S_READ); 1297 bcopy(va, (char *)db->db_data + bufoff, thiscpy); 1298 zfs_unmap_page(pp, va); 1299 pp = pp->p_next; 1300 bufoff += PAGESIZE; 1301 } 1302 1303 if (tocpy == db->db_size) 1304 dmu_buf_fill_done(db, tx); 1305 1306 offset += tocpy; 1307 size -= tocpy; 1308 } 1309 dmu_buf_rele_array(dbp, numbufs, FTAG); 1310 return (err); 1311} 1312 1313#else /* !illumos */ 1314 1315int 1316dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1317 vm_page_t *ma, dmu_tx_t *tx) 1318{ 1319 dmu_buf_t **dbp; 1320 struct sf_buf *sf; 1321 int numbufs, i; 1322 int err; 1323 1324 if (size == 0) 1325 return (0); 1326 1327 err = dmu_buf_hold_array(os, object, offset, size, 1328 FALSE, FTAG, &numbufs, &dbp); 1329 if (err) 1330 return (err); 1331 1332 for (i = 0; i < numbufs; i++) { 1333 int tocpy, copied, thiscpy; 1334 int bufoff; 1335 dmu_buf_t *db = dbp[i]; 1336 caddr_t va; 1337 1338 ASSERT(size > 0); 1339 ASSERT3U(db->db_size, >=, PAGESIZE); 1340 1341 bufoff = offset - db->db_offset; 1342 tocpy = (int)MIN(db->db_size - bufoff, size); 1343 1344 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 1345 1346 if (tocpy == db->db_size) 1347 dmu_buf_will_fill(db, tx); 1348 else 1349 dmu_buf_will_dirty(db, tx); 1350 1351 for (copied = 0; copied < tocpy; copied += PAGESIZE) { 1352 ASSERT3U(ptoa((*ma)->pindex), ==, db->db_offset + bufoff); 1353 thiscpy = MIN(PAGESIZE, tocpy - copied); 1354 va = zfs_map_page(*ma, &sf); 1355 bcopy(va, (char *)db->db_data + bufoff, thiscpy); 1356 zfs_unmap_page(sf); 1357 ma += 1; 1358 bufoff += PAGESIZE; 1359 } 1360 1361 if (tocpy == db->db_size) 1362 dmu_buf_fill_done(db, tx); 1363 1364 offset += tocpy; 1365 size -= tocpy; 1366 } 1367 dmu_buf_rele_array(dbp, numbufs, FTAG); 1368 return (err); 1369} 1370#endif /* illumos */ 1371#endif /* _KERNEL */ 1372 1373/* 1374 * Allocate a loaned anonymous arc buffer. 1375 */ 1376arc_buf_t * 1377dmu_request_arcbuf(dmu_buf_t *handle, int size) 1378{ 1379 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle; 1380 1381 return (arc_loan_buf(db->db_objset->os_spa, size)); 1382} 1383 1384/* 1385 * Free a loaned arc buffer. 1386 */ 1387void 1388dmu_return_arcbuf(arc_buf_t *buf) 1389{ 1390 arc_return_buf(buf, FTAG); 1391 VERIFY(arc_buf_remove_ref(buf, FTAG)); 1392} 1393 1394/* 1395 * When possible directly assign passed loaned arc buffer to a dbuf. 1396 * If this is not possible copy the contents of passed arc buf via 1397 * dmu_write(). 1398 */ 1399void 1400dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf, 1401 dmu_tx_t *tx) 1402{ 1403 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle; 1404 dnode_t *dn; 1405 dmu_buf_impl_t *db; 1406 uint32_t blksz = (uint32_t)arc_buf_size(buf); 1407 uint64_t blkid; 1408 1409 DB_DNODE_ENTER(dbuf); 1410 dn = DB_DNODE(dbuf); 1411 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1412 blkid = dbuf_whichblock(dn, 0, offset); 1413 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL); 1414 rw_exit(&dn->dn_struct_rwlock); 1415 DB_DNODE_EXIT(dbuf); 1416 1417 /* 1418 * We can only assign if the offset is aligned, the arc buf is the 1419 * same size as the dbuf, and the dbuf is not metadata. It 1420 * can't be metadata because the loaned arc buf comes from the 1421 * user-data kmem arena. 1422 */ 1423 if (offset == db->db.db_offset && blksz == db->db.db_size && 1424 DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA) { 1425#ifdef _KERNEL 1426 curthread->td_ru.ru_oublock++; 1427#endif 1428 dbuf_assign_arcbuf(db, buf, tx); 1429 dbuf_rele(db, FTAG); 1430 } else { 1431 objset_t *os; 1432 uint64_t object; 1433 1434 DB_DNODE_ENTER(dbuf); 1435 dn = DB_DNODE(dbuf); 1436 os = dn->dn_objset; 1437 object = dn->dn_object; 1438 DB_DNODE_EXIT(dbuf); 1439 1440 dbuf_rele(db, FTAG); 1441 dmu_write(os, object, offset, blksz, buf->b_data, tx); 1442 dmu_return_arcbuf(buf); 1443 XUIOSTAT_BUMP(xuiostat_wbuf_copied); 1444 } 1445} 1446 1447typedef struct { 1448 dbuf_dirty_record_t *dsa_dr; 1449 dmu_sync_cb_t *dsa_done; 1450 zgd_t *dsa_zgd; 1451 dmu_tx_t *dsa_tx; 1452} dmu_sync_arg_t; 1453 1454/* ARGSUSED */ 1455static void 1456dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg) 1457{ 1458 dmu_sync_arg_t *dsa = varg; 1459 dmu_buf_t *db = dsa->dsa_zgd->zgd_db; 1460 blkptr_t *bp = zio->io_bp; 1461 1462 if (zio->io_error == 0) { 1463 if (BP_IS_HOLE(bp)) { 1464 /* 1465 * A block of zeros may compress to a hole, but the 1466 * block size still needs to be known for replay. 1467 */ 1468 BP_SET_LSIZE(bp, db->db_size); 1469 } else if (!BP_IS_EMBEDDED(bp)) { 1470 ASSERT(BP_GET_LEVEL(bp) == 0); 1471 bp->blk_fill = 1; 1472 } 1473 } 1474} 1475 1476static void 1477dmu_sync_late_arrival_ready(zio_t *zio) 1478{ 1479 dmu_sync_ready(zio, NULL, zio->io_private); 1480} 1481 1482/* ARGSUSED */ 1483static void 1484dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg) 1485{ 1486 dmu_sync_arg_t *dsa = varg; 1487 dbuf_dirty_record_t *dr = dsa->dsa_dr; 1488 dmu_buf_impl_t *db = dr->dr_dbuf; 1489 1490 mutex_enter(&db->db_mtx); 1491 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC); 1492 if (zio->io_error == 0) { 1493 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE); 1494 if (dr->dt.dl.dr_nopwrite) { 1495 blkptr_t *bp = zio->io_bp; 1496 blkptr_t *bp_orig = &zio->io_bp_orig; 1497 uint8_t chksum = BP_GET_CHECKSUM(bp_orig); 1498 1499 ASSERT(BP_EQUAL(bp, bp_orig)); 1500 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF); 1501 ASSERT(zio_checksum_table[chksum].ci_flags & 1502 ZCHECKSUM_FLAG_NOPWRITE); 1503 } 1504 dr->dt.dl.dr_overridden_by = *zio->io_bp; 1505 dr->dt.dl.dr_override_state = DR_OVERRIDDEN; 1506 dr->dt.dl.dr_copies = zio->io_prop.zp_copies; 1507 1508 /* 1509 * Old style holes are filled with all zeros, whereas 1510 * new-style holes maintain their lsize, type, level, 1511 * and birth time (see zio_write_compress). While we 1512 * need to reset the BP_SET_LSIZE() call that happened 1513 * in dmu_sync_ready for old style holes, we do *not* 1514 * want to wipe out the information contained in new 1515 * style holes. Thus, only zero out the block pointer if 1516 * it's an old style hole. 1517 */ 1518 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) && 1519 dr->dt.dl.dr_overridden_by.blk_birth == 0) 1520 BP_ZERO(&dr->dt.dl.dr_overridden_by); 1521 } else { 1522 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 1523 } 1524 cv_broadcast(&db->db_changed); 1525 mutex_exit(&db->db_mtx); 1526 1527 dsa->dsa_done(dsa->dsa_zgd, zio->io_error); 1528 1529 kmem_free(dsa, sizeof (*dsa)); 1530} 1531 1532static void 1533dmu_sync_late_arrival_done(zio_t *zio) 1534{ 1535 blkptr_t *bp = zio->io_bp; 1536 dmu_sync_arg_t *dsa = zio->io_private; 1537 blkptr_t *bp_orig = &zio->io_bp_orig; 1538 1539 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) { 1540 /* 1541 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE) 1542 * then there is nothing to do here. Otherwise, free the 1543 * newly allocated block in this txg. 1544 */ 1545 if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 1546 ASSERT(BP_EQUAL(bp, bp_orig)); 1547 } else { 1548 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig)); 1549 ASSERT(zio->io_bp->blk_birth == zio->io_txg); 1550 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa)); 1551 zio_free(zio->io_spa, zio->io_txg, zio->io_bp); 1552 } 1553 } 1554 1555 dmu_tx_commit(dsa->dsa_tx); 1556 1557 dsa->dsa_done(dsa->dsa_zgd, zio->io_error); 1558 1559 kmem_free(dsa, sizeof (*dsa)); 1560} 1561 1562static int 1563dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd, 1564 zio_prop_t *zp, zbookmark_phys_t *zb) 1565{ 1566 dmu_sync_arg_t *dsa; 1567 dmu_tx_t *tx; 1568 1569 tx = dmu_tx_create(os); 1570 dmu_tx_hold_space(tx, zgd->zgd_db->db_size); 1571 if (dmu_tx_assign(tx, TXG_WAIT) != 0) { 1572 dmu_tx_abort(tx); 1573 /* Make zl_get_data do txg_waited_synced() */ 1574 return (SET_ERROR(EIO)); 1575 } 1576 1577 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP); 1578 dsa->dsa_dr = NULL; 1579 dsa->dsa_done = done; 1580 dsa->dsa_zgd = zgd; 1581 dsa->dsa_tx = tx; 1582 1583 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp, 1584 zgd->zgd_db->db_data, zgd->zgd_db->db_size, zp, 1585 dmu_sync_late_arrival_ready, NULL, dmu_sync_late_arrival_done, dsa, 1586 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb)); 1587 1588 return (0); 1589} 1590 1591/* 1592 * Intent log support: sync the block associated with db to disk. 1593 * N.B. and XXX: the caller is responsible for making sure that the 1594 * data isn't changing while dmu_sync() is writing it. 1595 * 1596 * Return values: 1597 * 1598 * EEXIST: this txg has already been synced, so there's nothing to do. 1599 * The caller should not log the write. 1600 * 1601 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do. 1602 * The caller should not log the write. 1603 * 1604 * EALREADY: this block is already in the process of being synced. 1605 * The caller should track its progress (somehow). 1606 * 1607 * EIO: could not do the I/O. 1608 * The caller should do a txg_wait_synced(). 1609 * 1610 * 0: the I/O has been initiated. 1611 * The caller should log this blkptr in the done callback. 1612 * It is possible that the I/O will fail, in which case 1613 * the error will be reported to the done callback and 1614 * propagated to pio from zio_done(). 1615 */ 1616int 1617dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd) 1618{ 1619 blkptr_t *bp = zgd->zgd_bp; 1620 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db; 1621 objset_t *os = db->db_objset; 1622 dsl_dataset_t *ds = os->os_dsl_dataset; 1623 dbuf_dirty_record_t *dr; 1624 dmu_sync_arg_t *dsa; 1625 zbookmark_phys_t zb; 1626 zio_prop_t zp; 1627 dnode_t *dn; 1628 1629 ASSERT(pio != NULL); 1630 ASSERT(txg != 0); 1631 1632 SET_BOOKMARK(&zb, ds->ds_object, 1633 db->db.db_object, db->db_level, db->db_blkid); 1634 1635 DB_DNODE_ENTER(db); 1636 dn = DB_DNODE(db); 1637 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp); 1638 DB_DNODE_EXIT(db); 1639 1640 /* 1641 * If we're frozen (running ziltest), we always need to generate a bp. 1642 */ 1643 if (txg > spa_freeze_txg(os->os_spa)) 1644 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); 1645 1646 /* 1647 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf() 1648 * and us. If we determine that this txg is not yet syncing, 1649 * but it begins to sync a moment later, that's OK because the 1650 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx. 1651 */ 1652 mutex_enter(&db->db_mtx); 1653 1654 if (txg <= spa_last_synced_txg(os->os_spa)) { 1655 /* 1656 * This txg has already synced. There's nothing to do. 1657 */ 1658 mutex_exit(&db->db_mtx); 1659 return (SET_ERROR(EEXIST)); 1660 } 1661 1662 if (txg <= spa_syncing_txg(os->os_spa)) { 1663 /* 1664 * This txg is currently syncing, so we can't mess with 1665 * the dirty record anymore; just write a new log block. 1666 */ 1667 mutex_exit(&db->db_mtx); 1668 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); 1669 } 1670 1671 dr = db->db_last_dirty; 1672 while (dr && dr->dr_txg != txg) 1673 dr = dr->dr_next; 1674 1675 if (dr == NULL) { 1676 /* 1677 * There's no dr for this dbuf, so it must have been freed. 1678 * There's no need to log writes to freed blocks, so we're done. 1679 */ 1680 mutex_exit(&db->db_mtx); 1681 return (SET_ERROR(ENOENT)); 1682 } 1683 1684 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg); 1685 1686 /* 1687 * Assume the on-disk data is X, the current syncing data (in 1688 * txg - 1) is Y, and the current in-memory data is Z (currently 1689 * in dmu_sync). 1690 * 1691 * We usually want to perform a nopwrite if X and Z are the 1692 * same. However, if Y is different (i.e. the BP is going to 1693 * change before this write takes effect), then a nopwrite will 1694 * be incorrect - we would override with X, which could have 1695 * been freed when Y was written. 1696 * 1697 * (Note that this is not a concern when we are nop-writing from 1698 * syncing context, because X and Y must be identical, because 1699 * all previous txgs have been synced.) 1700 * 1701 * Therefore, we disable nopwrite if the current BP could change 1702 * before this TXG. There are two ways it could change: by 1703 * being dirty (dr_next is non-NULL), or by being freed 1704 * (dnode_block_freed()). This behavior is verified by 1705 * zio_done(), which VERIFYs that the override BP is identical 1706 * to the on-disk BP. 1707 */ 1708 DB_DNODE_ENTER(db); 1709 dn = DB_DNODE(db); 1710 if (dr->dr_next != NULL || dnode_block_freed(dn, db->db_blkid)) 1711 zp.zp_nopwrite = B_FALSE; 1712 DB_DNODE_EXIT(db); 1713 1714 ASSERT(dr->dr_txg == txg); 1715 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC || 1716 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { 1717 /* 1718 * We have already issued a sync write for this buffer, 1719 * or this buffer has already been synced. It could not 1720 * have been dirtied since, or we would have cleared the state. 1721 */ 1722 mutex_exit(&db->db_mtx); 1723 return (SET_ERROR(EALREADY)); 1724 } 1725 1726 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); 1727 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC; 1728 mutex_exit(&db->db_mtx); 1729 1730 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP); 1731 dsa->dsa_dr = dr; 1732 dsa->dsa_done = done; 1733 dsa->dsa_zgd = zgd; 1734 dsa->dsa_tx = NULL; 1735 1736 zio_nowait(arc_write(pio, os->os_spa, txg, 1737 bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db), 1738 DBUF_IS_L2COMPRESSIBLE(db), &zp, dmu_sync_ready, 1739 NULL, dmu_sync_done, dsa, ZIO_PRIORITY_SYNC_WRITE, 1740 ZIO_FLAG_CANFAIL, &zb)); 1741 1742 return (0); 1743} 1744 1745int 1746dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs, 1747 dmu_tx_t *tx) 1748{ 1749 dnode_t *dn; 1750 int err; 1751 1752 err = dnode_hold(os, object, FTAG, &dn); 1753 if (err) 1754 return (err); 1755 err = dnode_set_blksz(dn, size, ibs, tx); 1756 dnode_rele(dn, FTAG); 1757 return (err); 1758} 1759 1760void 1761dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum, 1762 dmu_tx_t *tx) 1763{ 1764 dnode_t *dn; 1765 1766 /* 1767 * Send streams include each object's checksum function. This 1768 * check ensures that the receiving system can understand the 1769 * checksum function transmitted. 1770 */ 1771 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS); 1772 1773 VERIFY0(dnode_hold(os, object, FTAG, &dn)); 1774 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS); 1775 dn->dn_checksum = checksum; 1776 dnode_setdirty(dn, tx); 1777 dnode_rele(dn, FTAG); 1778} 1779 1780void 1781dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress, 1782 dmu_tx_t *tx) 1783{ 1784 dnode_t *dn; 1785 1786 /* 1787 * Send streams include each object's compression function. This 1788 * check ensures that the receiving system can understand the 1789 * compression function transmitted. 1790 */ 1791 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS); 1792 1793 VERIFY0(dnode_hold(os, object, FTAG, &dn)); 1794 dn->dn_compress = compress; 1795 dnode_setdirty(dn, tx); 1796 dnode_rele(dn, FTAG); 1797} 1798 1799int zfs_mdcomp_disable = 0; 1800TUNABLE_INT("vfs.zfs.mdcomp_disable", &zfs_mdcomp_disable); 1801SYSCTL_INT(_vfs_zfs, OID_AUTO, mdcomp_disable, CTLFLAG_RW, 1802 &zfs_mdcomp_disable, 0, "Disable metadata compression"); 1803 1804/* 1805 * When the "redundant_metadata" property is set to "most", only indirect 1806 * blocks of this level and higher will have an additional ditto block. 1807 */ 1808int zfs_redundant_metadata_most_ditto_level = 2; 1809 1810void 1811dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp) 1812{ 1813 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET; 1814 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) || 1815 (wp & WP_SPILL)); 1816 enum zio_checksum checksum = os->os_checksum; 1817 enum zio_compress compress = os->os_compress; 1818 enum zio_checksum dedup_checksum = os->os_dedup_checksum; 1819 boolean_t dedup = B_FALSE; 1820 boolean_t nopwrite = B_FALSE; 1821 boolean_t dedup_verify = os->os_dedup_verify; 1822 int copies = os->os_copies; 1823 1824 /* 1825 * We maintain different write policies for each of the following 1826 * types of data: 1827 * 1. metadata 1828 * 2. preallocated blocks (i.e. level-0 blocks of a dump device) 1829 * 3. all other level 0 blocks 1830 */ 1831 if (ismd) { 1832 if (zfs_mdcomp_disable) { 1833 compress = ZIO_COMPRESS_EMPTY; 1834 } else { 1835 /* 1836 * XXX -- we should design a compression algorithm 1837 * that specializes in arrays of bps. 1838 */ 1839 compress = zio_compress_select(os->os_spa, 1840 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON); 1841 } 1842 1843 /* 1844 * Metadata always gets checksummed. If the data 1845 * checksum is multi-bit correctable, and it's not a 1846 * ZBT-style checksum, then it's suitable for metadata 1847 * as well. Otherwise, the metadata checksum defaults 1848 * to fletcher4. 1849 */ 1850 if (!(zio_checksum_table[checksum].ci_flags & 1851 ZCHECKSUM_FLAG_METADATA) || 1852 (zio_checksum_table[checksum].ci_flags & 1853 ZCHECKSUM_FLAG_EMBEDDED)) 1854 checksum = ZIO_CHECKSUM_FLETCHER_4; 1855 1856 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL || 1857 (os->os_redundant_metadata == 1858 ZFS_REDUNDANT_METADATA_MOST && 1859 (level >= zfs_redundant_metadata_most_ditto_level || 1860 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL)))) 1861 copies++; 1862 } else if (wp & WP_NOFILL) { 1863 ASSERT(level == 0); 1864 1865 /* 1866 * If we're writing preallocated blocks, we aren't actually 1867 * writing them so don't set any policy properties. These 1868 * blocks are currently only used by an external subsystem 1869 * outside of zfs (i.e. dump) and not written by the zio 1870 * pipeline. 1871 */ 1872 compress = ZIO_COMPRESS_OFF; 1873 checksum = ZIO_CHECKSUM_NOPARITY; 1874 } else { 1875 compress = zio_compress_select(os->os_spa, dn->dn_compress, 1876 compress); 1877 1878 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ? 1879 zio_checksum_select(dn->dn_checksum, checksum) : 1880 dedup_checksum; 1881 1882 /* 1883 * Determine dedup setting. If we are in dmu_sync(), 1884 * we won't actually dedup now because that's all 1885 * done in syncing context; but we do want to use the 1886 * dedup checkum. If the checksum is not strong 1887 * enough to ensure unique signatures, force 1888 * dedup_verify. 1889 */ 1890 if (dedup_checksum != ZIO_CHECKSUM_OFF) { 1891 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE; 1892 if (!(zio_checksum_table[checksum].ci_flags & 1893 ZCHECKSUM_FLAG_DEDUP)) 1894 dedup_verify = B_TRUE; 1895 } 1896 1897 /* 1898 * Enable nopwrite if we have secure enough checksum 1899 * algorithm (see comment in zio_nop_write) and 1900 * compression is enabled. We don't enable nopwrite if 1901 * dedup is enabled as the two features are mutually 1902 * exclusive. 1903 */ 1904 nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags & 1905 ZCHECKSUM_FLAG_NOPWRITE) && 1906 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled); 1907 } 1908 1909 zp->zp_checksum = checksum; 1910 zp->zp_compress = compress; 1911 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type; 1912 zp->zp_level = level; 1913 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa)); 1914 zp->zp_dedup = dedup; 1915 zp->zp_dedup_verify = dedup && dedup_verify; 1916 zp->zp_nopwrite = nopwrite; 1917} 1918 1919int 1920dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off) 1921{ 1922 dnode_t *dn; 1923 int err; 1924 1925 /* 1926 * Sync any current changes before 1927 * we go trundling through the block pointers. 1928 */ 1929 err = dmu_object_wait_synced(os, object); 1930 if (err) { 1931 return (err); 1932 } 1933 1934 err = dnode_hold(os, object, FTAG, &dn); 1935 if (err) { 1936 return (err); 1937 } 1938 1939 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0); 1940 dnode_rele(dn, FTAG); 1941 1942 return (err); 1943} 1944 1945/* 1946 * Given the ZFS object, if it contains any dirty nodes 1947 * this function flushes all dirty blocks to disk. This 1948 * ensures the DMU object info is updated. A more efficient 1949 * future version might just find the TXG with the maximum 1950 * ID and wait for that to be synced. 1951 */ 1952int 1953dmu_object_wait_synced(objset_t *os, uint64_t object) 1954{ 1955 dnode_t *dn; 1956 int error, i; 1957 1958 error = dnode_hold(os, object, FTAG, &dn); 1959 if (error) { 1960 return (error); 1961 } 1962 1963 for (i = 0; i < TXG_SIZE; i++) { 1964 if (list_link_active(&dn->dn_dirty_link[i])) { 1965 break; 1966 } 1967 } 1968 dnode_rele(dn, FTAG); 1969 if (i != TXG_SIZE) { 1970 txg_wait_synced(dmu_objset_pool(os), 0); 1971 } 1972 1973 return (0); 1974} 1975 1976void 1977dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi) 1978{ 1979 dnode_phys_t *dnp; 1980 1981 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1982 mutex_enter(&dn->dn_mtx); 1983 1984 dnp = dn->dn_phys; 1985 1986 doi->doi_data_block_size = dn->dn_datablksz; 1987 doi->doi_metadata_block_size = dn->dn_indblkshift ? 1988 1ULL << dn->dn_indblkshift : 0; 1989 doi->doi_type = dn->dn_type; 1990 doi->doi_bonus_type = dn->dn_bonustype; 1991 doi->doi_bonus_size = dn->dn_bonuslen; 1992 doi->doi_indirection = dn->dn_nlevels; 1993 doi->doi_checksum = dn->dn_checksum; 1994 doi->doi_compress = dn->dn_compress; 1995 doi->doi_nblkptr = dn->dn_nblkptr; 1996 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9; 1997 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz; 1998 doi->doi_fill_count = 0; 1999 for (int i = 0; i < dnp->dn_nblkptr; i++) 2000 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]); 2001 2002 mutex_exit(&dn->dn_mtx); 2003 rw_exit(&dn->dn_struct_rwlock); 2004} 2005 2006/* 2007 * Get information on a DMU object. 2008 * If doi is NULL, just indicates whether the object exists. 2009 */ 2010int 2011dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi) 2012{ 2013 dnode_t *dn; 2014 int err = dnode_hold(os, object, FTAG, &dn); 2015 2016 if (err) 2017 return (err); 2018 2019 if (doi != NULL) 2020 dmu_object_info_from_dnode(dn, doi); 2021 2022 dnode_rele(dn, FTAG); 2023 return (0); 2024} 2025 2026/* 2027 * As above, but faster; can be used when you have a held dbuf in hand. 2028 */ 2029void 2030dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi) 2031{ 2032 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2033 2034 DB_DNODE_ENTER(db); 2035 dmu_object_info_from_dnode(DB_DNODE(db), doi); 2036 DB_DNODE_EXIT(db); 2037} 2038 2039/* 2040 * Faster still when you only care about the size. 2041 * This is specifically optimized for zfs_getattr(). 2042 */ 2043void 2044dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize, 2045 u_longlong_t *nblk512) 2046{ 2047 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2048 dnode_t *dn; 2049 2050 DB_DNODE_ENTER(db); 2051 dn = DB_DNODE(db); 2052 2053 *blksize = dn->dn_datablksz; 2054 /* add 1 for dnode space */ 2055 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >> 2056 SPA_MINBLOCKSHIFT) + 1; 2057 DB_DNODE_EXIT(db); 2058} 2059 2060void 2061byteswap_uint64_array(void *vbuf, size_t size) 2062{ 2063 uint64_t *buf = vbuf; 2064 size_t count = size >> 3; 2065 int i; 2066 2067 ASSERT((size & 7) == 0); 2068 2069 for (i = 0; i < count; i++) 2070 buf[i] = BSWAP_64(buf[i]); 2071} 2072 2073void 2074byteswap_uint32_array(void *vbuf, size_t size) 2075{ 2076 uint32_t *buf = vbuf; 2077 size_t count = size >> 2; 2078 int i; 2079 2080 ASSERT((size & 3) == 0); 2081 2082 for (i = 0; i < count; i++) 2083 buf[i] = BSWAP_32(buf[i]); 2084} 2085 2086void 2087byteswap_uint16_array(void *vbuf, size_t size) 2088{ 2089 uint16_t *buf = vbuf; 2090 size_t count = size >> 1; 2091 int i; 2092 2093 ASSERT((size & 1) == 0); 2094 2095 for (i = 0; i < count; i++) 2096 buf[i] = BSWAP_16(buf[i]); 2097} 2098 2099/* ARGSUSED */ 2100void 2101byteswap_uint8_array(void *vbuf, size_t size) 2102{ 2103} 2104 2105void 2106dmu_init(void) 2107{ 2108 zfs_dbgmsg_init(); 2109 sa_cache_init(); 2110 xuio_stat_init(); 2111 dmu_objset_init(); 2112 dnode_init(); 2113 dbuf_init(); 2114 zfetch_init(); 2115 zio_compress_init(); 2116 l2arc_init(); 2117 arc_init(); 2118} 2119 2120void 2121dmu_fini(void) 2122{ 2123 arc_fini(); /* arc depends on l2arc, so arc must go first */ 2124 l2arc_fini(); 2125 zfetch_fini(); 2126 zio_compress_fini(); 2127 dbuf_fini(); 2128 dnode_fini(); 2129 dmu_objset_fini(); 2130 xuio_stat_fini(); 2131 sa_cache_fini(); 2132 zfs_dbgmsg_fini(); 2133} 2134