dmu.c revision 307266
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, 2016 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#ifdef illumos 1087 err = uiomove((char *)db->db_data + bufoff, tocpy, 1088 UIO_READ, uio); 1089#else 1090 err = vn_io_fault_uiomove((char *)db->db_data + bufoff, 1091 tocpy, uio); 1092#endif 1093 } 1094 if (err) 1095 break; 1096 1097 size -= tocpy; 1098 } 1099 dmu_buf_rele_array(dbp, numbufs, FTAG); 1100 1101 return (err); 1102} 1103 1104/* 1105 * Read 'size' bytes into the uio buffer. 1106 * From object zdb->db_object. 1107 * Starting at offset uio->uio_loffset. 1108 * 1109 * If the caller already has a dbuf in the target object 1110 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(), 1111 * because we don't have to find the dnode_t for the object. 1112 */ 1113int 1114dmu_read_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size) 1115{ 1116 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb; 1117 dnode_t *dn; 1118 int err; 1119 1120 if (size == 0) 1121 return (0); 1122 1123 DB_DNODE_ENTER(db); 1124 dn = DB_DNODE(db); 1125 err = dmu_read_uio_dnode(dn, uio, size); 1126 DB_DNODE_EXIT(db); 1127 1128 return (err); 1129} 1130 1131/* 1132 * Read 'size' bytes into the uio buffer. 1133 * From the specified object 1134 * Starting at offset uio->uio_loffset. 1135 */ 1136int 1137dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size) 1138{ 1139 dnode_t *dn; 1140 int err; 1141 1142 if (size == 0) 1143 return (0); 1144 1145 err = dnode_hold(os, object, FTAG, &dn); 1146 if (err) 1147 return (err); 1148 1149 err = dmu_read_uio_dnode(dn, uio, size); 1150 1151 dnode_rele(dn, FTAG); 1152 1153 return (err); 1154} 1155 1156static int 1157dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx) 1158{ 1159 dmu_buf_t **dbp; 1160 int numbufs; 1161 int err = 0; 1162 int i; 1163 1164 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size, 1165 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH); 1166 if (err) 1167 return (err); 1168 1169 for (i = 0; i < numbufs; i++) { 1170 int tocpy; 1171 int bufoff; 1172 dmu_buf_t *db = dbp[i]; 1173 1174 ASSERT(size > 0); 1175 1176 bufoff = uio->uio_loffset - db->db_offset; 1177 tocpy = (int)MIN(db->db_size - bufoff, size); 1178 1179 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 1180 1181 if (tocpy == db->db_size) 1182 dmu_buf_will_fill(db, tx); 1183 else 1184 dmu_buf_will_dirty(db, tx); 1185 1186#ifdef illumos 1187 /* 1188 * XXX uiomove could block forever (eg. nfs-backed 1189 * pages). There needs to be a uiolockdown() function 1190 * to lock the pages in memory, so that uiomove won't 1191 * block. 1192 */ 1193 err = uiomove((char *)db->db_data + bufoff, tocpy, 1194 UIO_WRITE, uio); 1195#else 1196 err = vn_io_fault_uiomove((char *)db->db_data + bufoff, tocpy, 1197 uio); 1198#endif 1199 1200 if (tocpy == db->db_size) 1201 dmu_buf_fill_done(db, tx); 1202 1203 if (err) 1204 break; 1205 1206 size -= tocpy; 1207 } 1208 1209 dmu_buf_rele_array(dbp, numbufs, FTAG); 1210 return (err); 1211} 1212 1213/* 1214 * Write 'size' bytes from the uio buffer. 1215 * To object zdb->db_object. 1216 * Starting at offset uio->uio_loffset. 1217 * 1218 * If the caller already has a dbuf in the target object 1219 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(), 1220 * because we don't have to find the dnode_t for the object. 1221 */ 1222int 1223dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size, 1224 dmu_tx_t *tx) 1225{ 1226 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb; 1227 dnode_t *dn; 1228 int err; 1229 1230 if (size == 0) 1231 return (0); 1232 1233 DB_DNODE_ENTER(db); 1234 dn = DB_DNODE(db); 1235 err = dmu_write_uio_dnode(dn, uio, size, tx); 1236 DB_DNODE_EXIT(db); 1237 1238 return (err); 1239} 1240 1241/* 1242 * Write 'size' bytes from the uio buffer. 1243 * To the specified object. 1244 * Starting at offset uio->uio_loffset. 1245 */ 1246int 1247dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size, 1248 dmu_tx_t *tx) 1249{ 1250 dnode_t *dn; 1251 int err; 1252 1253 if (size == 0) 1254 return (0); 1255 1256 err = dnode_hold(os, object, FTAG, &dn); 1257 if (err) 1258 return (err); 1259 1260 err = dmu_write_uio_dnode(dn, uio, size, tx); 1261 1262 dnode_rele(dn, FTAG); 1263 1264 return (err); 1265} 1266 1267#ifdef illumos 1268int 1269dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1270 page_t *pp, dmu_tx_t *tx) 1271{ 1272 dmu_buf_t **dbp; 1273 int numbufs, i; 1274 int err; 1275 1276 if (size == 0) 1277 return (0); 1278 1279 err = dmu_buf_hold_array(os, object, offset, size, 1280 FALSE, FTAG, &numbufs, &dbp); 1281 if (err) 1282 return (err); 1283 1284 for (i = 0; i < numbufs; i++) { 1285 int tocpy, copied, thiscpy; 1286 int bufoff; 1287 dmu_buf_t *db = dbp[i]; 1288 caddr_t va; 1289 1290 ASSERT(size > 0); 1291 ASSERT3U(db->db_size, >=, PAGESIZE); 1292 1293 bufoff = offset - db->db_offset; 1294 tocpy = (int)MIN(db->db_size - bufoff, size); 1295 1296 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 1297 1298 if (tocpy == db->db_size) 1299 dmu_buf_will_fill(db, tx); 1300 else 1301 dmu_buf_will_dirty(db, tx); 1302 1303 for (copied = 0; copied < tocpy; copied += PAGESIZE) { 1304 ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff); 1305 thiscpy = MIN(PAGESIZE, tocpy - copied); 1306 va = zfs_map_page(pp, S_READ); 1307 bcopy(va, (char *)db->db_data + bufoff, thiscpy); 1308 zfs_unmap_page(pp, va); 1309 pp = pp->p_next; 1310 bufoff += PAGESIZE; 1311 } 1312 1313 if (tocpy == db->db_size) 1314 dmu_buf_fill_done(db, tx); 1315 1316 offset += tocpy; 1317 size -= tocpy; 1318 } 1319 dmu_buf_rele_array(dbp, numbufs, FTAG); 1320 return (err); 1321} 1322 1323#else /* !illumos */ 1324 1325int 1326dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1327 vm_page_t *ma, dmu_tx_t *tx) 1328{ 1329 dmu_buf_t **dbp; 1330 struct sf_buf *sf; 1331 int numbufs, i; 1332 int err; 1333 1334 if (size == 0) 1335 return (0); 1336 1337 err = dmu_buf_hold_array(os, object, offset, size, 1338 FALSE, FTAG, &numbufs, &dbp); 1339 if (err) 1340 return (err); 1341 1342 for (i = 0; i < numbufs; i++) { 1343 int tocpy, copied, thiscpy; 1344 int bufoff; 1345 dmu_buf_t *db = dbp[i]; 1346 caddr_t va; 1347 1348 ASSERT(size > 0); 1349 ASSERT3U(db->db_size, >=, PAGESIZE); 1350 1351 bufoff = offset - db->db_offset; 1352 tocpy = (int)MIN(db->db_size - bufoff, size); 1353 1354 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 1355 1356 if (tocpy == db->db_size) 1357 dmu_buf_will_fill(db, tx); 1358 else 1359 dmu_buf_will_dirty(db, tx); 1360 1361 for (copied = 0; copied < tocpy; copied += PAGESIZE) { 1362 ASSERT3U(ptoa((*ma)->pindex), ==, db->db_offset + bufoff); 1363 thiscpy = MIN(PAGESIZE, tocpy - copied); 1364 va = zfs_map_page(*ma, &sf); 1365 bcopy(va, (char *)db->db_data + bufoff, thiscpy); 1366 zfs_unmap_page(sf); 1367 ma += 1; 1368 bufoff += PAGESIZE; 1369 } 1370 1371 if (tocpy == db->db_size) 1372 dmu_buf_fill_done(db, tx); 1373 1374 offset += tocpy; 1375 size -= tocpy; 1376 } 1377 dmu_buf_rele_array(dbp, numbufs, FTAG); 1378 return (err); 1379} 1380#endif /* illumos */ 1381#endif /* _KERNEL */ 1382 1383/* 1384 * Allocate a loaned anonymous arc buffer. 1385 */ 1386arc_buf_t * 1387dmu_request_arcbuf(dmu_buf_t *handle, int size) 1388{ 1389 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle; 1390 1391 return (arc_loan_buf(db->db_objset->os_spa, size)); 1392} 1393 1394/* 1395 * Free a loaned arc buffer. 1396 */ 1397void 1398dmu_return_arcbuf(arc_buf_t *buf) 1399{ 1400 arc_return_buf(buf, FTAG); 1401 arc_buf_destroy(buf, FTAG); 1402} 1403 1404/* 1405 * When possible directly assign passed loaned arc buffer to a dbuf. 1406 * If this is not possible copy the contents of passed arc buf via 1407 * dmu_write(). 1408 */ 1409void 1410dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf, 1411 dmu_tx_t *tx) 1412{ 1413 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle; 1414 dnode_t *dn; 1415 dmu_buf_impl_t *db; 1416 uint32_t blksz = (uint32_t)arc_buf_size(buf); 1417 uint64_t blkid; 1418 1419 DB_DNODE_ENTER(dbuf); 1420 dn = DB_DNODE(dbuf); 1421 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1422 blkid = dbuf_whichblock(dn, 0, offset); 1423 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL); 1424 rw_exit(&dn->dn_struct_rwlock); 1425 DB_DNODE_EXIT(dbuf); 1426 1427 /* 1428 * We can only assign if the offset is aligned, the arc buf is the 1429 * same size as the dbuf, and the dbuf is not metadata. It 1430 * can't be metadata because the loaned arc buf comes from the 1431 * user-data kmem arena. 1432 */ 1433 if (offset == db->db.db_offset && blksz == db->db.db_size && 1434 DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA) { 1435#ifdef _KERNEL 1436 curthread->td_ru.ru_oublock++; 1437#endif 1438 dbuf_assign_arcbuf(db, buf, tx); 1439 dbuf_rele(db, FTAG); 1440 } else { 1441 objset_t *os; 1442 uint64_t object; 1443 1444 DB_DNODE_ENTER(dbuf); 1445 dn = DB_DNODE(dbuf); 1446 os = dn->dn_objset; 1447 object = dn->dn_object; 1448 DB_DNODE_EXIT(dbuf); 1449 1450 dbuf_rele(db, FTAG); 1451 dmu_write(os, object, offset, blksz, buf->b_data, tx); 1452 dmu_return_arcbuf(buf); 1453 XUIOSTAT_BUMP(xuiostat_wbuf_copied); 1454 } 1455} 1456 1457typedef struct { 1458 dbuf_dirty_record_t *dsa_dr; 1459 dmu_sync_cb_t *dsa_done; 1460 zgd_t *dsa_zgd; 1461 dmu_tx_t *dsa_tx; 1462} dmu_sync_arg_t; 1463 1464/* ARGSUSED */ 1465static void 1466dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg) 1467{ 1468 dmu_sync_arg_t *dsa = varg; 1469 dmu_buf_t *db = dsa->dsa_zgd->zgd_db; 1470 blkptr_t *bp = zio->io_bp; 1471 1472 if (zio->io_error == 0) { 1473 if (BP_IS_HOLE(bp)) { 1474 /* 1475 * A block of zeros may compress to a hole, but the 1476 * block size still needs to be known for replay. 1477 */ 1478 BP_SET_LSIZE(bp, db->db_size); 1479 } else if (!BP_IS_EMBEDDED(bp)) { 1480 ASSERT(BP_GET_LEVEL(bp) == 0); 1481 bp->blk_fill = 1; 1482 } 1483 } 1484} 1485 1486static void 1487dmu_sync_late_arrival_ready(zio_t *zio) 1488{ 1489 dmu_sync_ready(zio, NULL, zio->io_private); 1490} 1491 1492/* ARGSUSED */ 1493static void 1494dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg) 1495{ 1496 dmu_sync_arg_t *dsa = varg; 1497 dbuf_dirty_record_t *dr = dsa->dsa_dr; 1498 dmu_buf_impl_t *db = dr->dr_dbuf; 1499 1500 mutex_enter(&db->db_mtx); 1501 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC); 1502 if (zio->io_error == 0) { 1503 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE); 1504 if (dr->dt.dl.dr_nopwrite) { 1505 blkptr_t *bp = zio->io_bp; 1506 blkptr_t *bp_orig = &zio->io_bp_orig; 1507 uint8_t chksum = BP_GET_CHECKSUM(bp_orig); 1508 1509 ASSERT(BP_EQUAL(bp, bp_orig)); 1510 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF); 1511 ASSERT(zio_checksum_table[chksum].ci_flags & 1512 ZCHECKSUM_FLAG_NOPWRITE); 1513 } 1514 dr->dt.dl.dr_overridden_by = *zio->io_bp; 1515 dr->dt.dl.dr_override_state = DR_OVERRIDDEN; 1516 dr->dt.dl.dr_copies = zio->io_prop.zp_copies; 1517 1518 /* 1519 * Old style holes are filled with all zeros, whereas 1520 * new-style holes maintain their lsize, type, level, 1521 * and birth time (see zio_write_compress). While we 1522 * need to reset the BP_SET_LSIZE() call that happened 1523 * in dmu_sync_ready for old style holes, we do *not* 1524 * want to wipe out the information contained in new 1525 * style holes. Thus, only zero out the block pointer if 1526 * it's an old style hole. 1527 */ 1528 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) && 1529 dr->dt.dl.dr_overridden_by.blk_birth == 0) 1530 BP_ZERO(&dr->dt.dl.dr_overridden_by); 1531 } else { 1532 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 1533 } 1534 cv_broadcast(&db->db_changed); 1535 mutex_exit(&db->db_mtx); 1536 1537 dsa->dsa_done(dsa->dsa_zgd, zio->io_error); 1538 1539 kmem_free(dsa, sizeof (*dsa)); 1540} 1541 1542static void 1543dmu_sync_late_arrival_done(zio_t *zio) 1544{ 1545 blkptr_t *bp = zio->io_bp; 1546 dmu_sync_arg_t *dsa = zio->io_private; 1547 blkptr_t *bp_orig = &zio->io_bp_orig; 1548 1549 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) { 1550 /* 1551 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE) 1552 * then there is nothing to do here. Otherwise, free the 1553 * newly allocated block in this txg. 1554 */ 1555 if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 1556 ASSERT(BP_EQUAL(bp, bp_orig)); 1557 } else { 1558 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig)); 1559 ASSERT(zio->io_bp->blk_birth == zio->io_txg); 1560 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa)); 1561 zio_free(zio->io_spa, zio->io_txg, zio->io_bp); 1562 } 1563 } 1564 1565 dmu_tx_commit(dsa->dsa_tx); 1566 1567 dsa->dsa_done(dsa->dsa_zgd, zio->io_error); 1568 1569 kmem_free(dsa, sizeof (*dsa)); 1570} 1571 1572static int 1573dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd, 1574 zio_prop_t *zp, zbookmark_phys_t *zb) 1575{ 1576 dmu_sync_arg_t *dsa; 1577 dmu_tx_t *tx; 1578 1579 tx = dmu_tx_create(os); 1580 dmu_tx_hold_space(tx, zgd->zgd_db->db_size); 1581 if (dmu_tx_assign(tx, TXG_WAIT) != 0) { 1582 dmu_tx_abort(tx); 1583 /* Make zl_get_data do txg_waited_synced() */ 1584 return (SET_ERROR(EIO)); 1585 } 1586 1587 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP); 1588 dsa->dsa_dr = NULL; 1589 dsa->dsa_done = done; 1590 dsa->dsa_zgd = zgd; 1591 dsa->dsa_tx = tx; 1592 1593 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), 1594 zgd->zgd_bp, zgd->zgd_db->db_data, zgd->zgd_db->db_size, 1595 zp, dmu_sync_late_arrival_ready, NULL, 1596 NULL, dmu_sync_late_arrival_done, dsa, ZIO_PRIORITY_SYNC_WRITE, 1597 ZIO_FLAG_CANFAIL, zb)); 1598 1599 return (0); 1600} 1601 1602/* 1603 * Intent log support: sync the block associated with db to disk. 1604 * N.B. and XXX: the caller is responsible for making sure that the 1605 * data isn't changing while dmu_sync() is writing it. 1606 * 1607 * Return values: 1608 * 1609 * EEXIST: this txg has already been synced, so there's nothing to do. 1610 * The caller should not log the write. 1611 * 1612 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do. 1613 * The caller should not log the write. 1614 * 1615 * EALREADY: this block is already in the process of being synced. 1616 * The caller should track its progress (somehow). 1617 * 1618 * EIO: could not do the I/O. 1619 * The caller should do a txg_wait_synced(). 1620 * 1621 * 0: the I/O has been initiated. 1622 * The caller should log this blkptr in the done callback. 1623 * It is possible that the I/O will fail, in which case 1624 * the error will be reported to the done callback and 1625 * propagated to pio from zio_done(). 1626 */ 1627int 1628dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd) 1629{ 1630 blkptr_t *bp = zgd->zgd_bp; 1631 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db; 1632 objset_t *os = db->db_objset; 1633 dsl_dataset_t *ds = os->os_dsl_dataset; 1634 dbuf_dirty_record_t *dr; 1635 dmu_sync_arg_t *dsa; 1636 zbookmark_phys_t zb; 1637 zio_prop_t zp; 1638 dnode_t *dn; 1639 1640 ASSERT(pio != NULL); 1641 ASSERT(txg != 0); 1642 1643 SET_BOOKMARK(&zb, ds->ds_object, 1644 db->db.db_object, db->db_level, db->db_blkid); 1645 1646 DB_DNODE_ENTER(db); 1647 dn = DB_DNODE(db); 1648 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp); 1649 DB_DNODE_EXIT(db); 1650 1651 /* 1652 * If we're frozen (running ziltest), we always need to generate a bp. 1653 */ 1654 if (txg > spa_freeze_txg(os->os_spa)) 1655 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); 1656 1657 /* 1658 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf() 1659 * and us. If we determine that this txg is not yet syncing, 1660 * but it begins to sync a moment later, that's OK because the 1661 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx. 1662 */ 1663 mutex_enter(&db->db_mtx); 1664 1665 if (txg <= spa_last_synced_txg(os->os_spa)) { 1666 /* 1667 * This txg has already synced. There's nothing to do. 1668 */ 1669 mutex_exit(&db->db_mtx); 1670 return (SET_ERROR(EEXIST)); 1671 } 1672 1673 if (txg <= spa_syncing_txg(os->os_spa)) { 1674 /* 1675 * This txg is currently syncing, so we can't mess with 1676 * the dirty record anymore; just write a new log block. 1677 */ 1678 mutex_exit(&db->db_mtx); 1679 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); 1680 } 1681 1682 dr = db->db_last_dirty; 1683 while (dr && dr->dr_txg != txg) 1684 dr = dr->dr_next; 1685 1686 if (dr == NULL) { 1687 /* 1688 * There's no dr for this dbuf, so it must have been freed. 1689 * There's no need to log writes to freed blocks, so we're done. 1690 */ 1691 mutex_exit(&db->db_mtx); 1692 return (SET_ERROR(ENOENT)); 1693 } 1694 1695 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg); 1696 1697 /* 1698 * Assume the on-disk data is X, the current syncing data (in 1699 * txg - 1) is Y, and the current in-memory data is Z (currently 1700 * in dmu_sync). 1701 * 1702 * We usually want to perform a nopwrite if X and Z are the 1703 * same. However, if Y is different (i.e. the BP is going to 1704 * change before this write takes effect), then a nopwrite will 1705 * be incorrect - we would override with X, which could have 1706 * been freed when Y was written. 1707 * 1708 * (Note that this is not a concern when we are nop-writing from 1709 * syncing context, because X and Y must be identical, because 1710 * all previous txgs have been synced.) 1711 * 1712 * Therefore, we disable nopwrite if the current BP could change 1713 * before this TXG. There are two ways it could change: by 1714 * being dirty (dr_next is non-NULL), or by being freed 1715 * (dnode_block_freed()). This behavior is verified by 1716 * zio_done(), which VERIFYs that the override BP is identical 1717 * to the on-disk BP. 1718 */ 1719 DB_DNODE_ENTER(db); 1720 dn = DB_DNODE(db); 1721 if (dr->dr_next != NULL || dnode_block_freed(dn, db->db_blkid)) 1722 zp.zp_nopwrite = B_FALSE; 1723 DB_DNODE_EXIT(db); 1724 1725 ASSERT(dr->dr_txg == txg); 1726 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC || 1727 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { 1728 /* 1729 * We have already issued a sync write for this buffer, 1730 * or this buffer has already been synced. It could not 1731 * have been dirtied since, or we would have cleared the state. 1732 */ 1733 mutex_exit(&db->db_mtx); 1734 return (SET_ERROR(EALREADY)); 1735 } 1736 1737 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); 1738 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC; 1739 mutex_exit(&db->db_mtx); 1740 1741 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP); 1742 dsa->dsa_dr = dr; 1743 dsa->dsa_done = done; 1744 dsa->dsa_zgd = zgd; 1745 dsa->dsa_tx = NULL; 1746 1747 zio_nowait(arc_write(pio, os->os_spa, txg, 1748 bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db), 1749 &zp, dmu_sync_ready, NULL, NULL, dmu_sync_done, dsa, 1750 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb)); 1751 1752 return (0); 1753} 1754 1755int 1756dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs, 1757 dmu_tx_t *tx) 1758{ 1759 dnode_t *dn; 1760 int err; 1761 1762 err = dnode_hold(os, object, FTAG, &dn); 1763 if (err) 1764 return (err); 1765 err = dnode_set_blksz(dn, size, ibs, tx); 1766 dnode_rele(dn, FTAG); 1767 return (err); 1768} 1769 1770void 1771dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum, 1772 dmu_tx_t *tx) 1773{ 1774 dnode_t *dn; 1775 1776 /* 1777 * Send streams include each object's checksum function. This 1778 * check ensures that the receiving system can understand the 1779 * checksum function transmitted. 1780 */ 1781 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS); 1782 1783 VERIFY0(dnode_hold(os, object, FTAG, &dn)); 1784 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS); 1785 dn->dn_checksum = checksum; 1786 dnode_setdirty(dn, tx); 1787 dnode_rele(dn, FTAG); 1788} 1789 1790void 1791dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress, 1792 dmu_tx_t *tx) 1793{ 1794 dnode_t *dn; 1795 1796 /* 1797 * Send streams include each object's compression function. This 1798 * check ensures that the receiving system can understand the 1799 * compression function transmitted. 1800 */ 1801 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS); 1802 1803 VERIFY0(dnode_hold(os, object, FTAG, &dn)); 1804 dn->dn_compress = compress; 1805 dnode_setdirty(dn, tx); 1806 dnode_rele(dn, FTAG); 1807} 1808 1809int zfs_mdcomp_disable = 0; 1810TUNABLE_INT("vfs.zfs.mdcomp_disable", &zfs_mdcomp_disable); 1811SYSCTL_INT(_vfs_zfs, OID_AUTO, mdcomp_disable, CTLFLAG_RW, 1812 &zfs_mdcomp_disable, 0, "Disable metadata compression"); 1813 1814/* 1815 * When the "redundant_metadata" property is set to "most", only indirect 1816 * blocks of this level and higher will have an additional ditto block. 1817 */ 1818int zfs_redundant_metadata_most_ditto_level = 2; 1819 1820void 1821dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp) 1822{ 1823 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET; 1824 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) || 1825 (wp & WP_SPILL)); 1826 enum zio_checksum checksum = os->os_checksum; 1827 enum zio_compress compress = os->os_compress; 1828 enum zio_checksum dedup_checksum = os->os_dedup_checksum; 1829 boolean_t dedup = B_FALSE; 1830 boolean_t nopwrite = B_FALSE; 1831 boolean_t dedup_verify = os->os_dedup_verify; 1832 int copies = os->os_copies; 1833 1834 /* 1835 * We maintain different write policies for each of the following 1836 * types of data: 1837 * 1. metadata 1838 * 2. preallocated blocks (i.e. level-0 blocks of a dump device) 1839 * 3. all other level 0 blocks 1840 */ 1841 if (ismd) { 1842 if (zfs_mdcomp_disable) { 1843 compress = ZIO_COMPRESS_EMPTY; 1844 } else { 1845 /* 1846 * XXX -- we should design a compression algorithm 1847 * that specializes in arrays of bps. 1848 */ 1849 compress = zio_compress_select(os->os_spa, 1850 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON); 1851 } 1852 1853 /* 1854 * Metadata always gets checksummed. If the data 1855 * checksum is multi-bit correctable, and it's not a 1856 * ZBT-style checksum, then it's suitable for metadata 1857 * as well. Otherwise, the metadata checksum defaults 1858 * to fletcher4. 1859 */ 1860 if (!(zio_checksum_table[checksum].ci_flags & 1861 ZCHECKSUM_FLAG_METADATA) || 1862 (zio_checksum_table[checksum].ci_flags & 1863 ZCHECKSUM_FLAG_EMBEDDED)) 1864 checksum = ZIO_CHECKSUM_FLETCHER_4; 1865 1866 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL || 1867 (os->os_redundant_metadata == 1868 ZFS_REDUNDANT_METADATA_MOST && 1869 (level >= zfs_redundant_metadata_most_ditto_level || 1870 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL)))) 1871 copies++; 1872 } else if (wp & WP_NOFILL) { 1873 ASSERT(level == 0); 1874 1875 /* 1876 * If we're writing preallocated blocks, we aren't actually 1877 * writing them so don't set any policy properties. These 1878 * blocks are currently only used by an external subsystem 1879 * outside of zfs (i.e. dump) and not written by the zio 1880 * pipeline. 1881 */ 1882 compress = ZIO_COMPRESS_OFF; 1883 checksum = ZIO_CHECKSUM_NOPARITY; 1884 } else { 1885 compress = zio_compress_select(os->os_spa, dn->dn_compress, 1886 compress); 1887 1888 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ? 1889 zio_checksum_select(dn->dn_checksum, checksum) : 1890 dedup_checksum; 1891 1892 /* 1893 * Determine dedup setting. If we are in dmu_sync(), 1894 * we won't actually dedup now because that's all 1895 * done in syncing context; but we do want to use the 1896 * dedup checkum. If the checksum is not strong 1897 * enough to ensure unique signatures, force 1898 * dedup_verify. 1899 */ 1900 if (dedup_checksum != ZIO_CHECKSUM_OFF) { 1901 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE; 1902 if (!(zio_checksum_table[checksum].ci_flags & 1903 ZCHECKSUM_FLAG_DEDUP)) 1904 dedup_verify = B_TRUE; 1905 } 1906 1907 /* 1908 * Enable nopwrite if we have secure enough checksum 1909 * algorithm (see comment in zio_nop_write) and 1910 * compression is enabled. We don't enable nopwrite if 1911 * dedup is enabled as the two features are mutually 1912 * exclusive. 1913 */ 1914 nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags & 1915 ZCHECKSUM_FLAG_NOPWRITE) && 1916 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled); 1917 } 1918 1919 zp->zp_checksum = checksum; 1920 zp->zp_compress = compress; 1921 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type; 1922 zp->zp_level = level; 1923 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa)); 1924 zp->zp_dedup = dedup; 1925 zp->zp_dedup_verify = dedup && dedup_verify; 1926 zp->zp_nopwrite = nopwrite; 1927} 1928 1929int 1930dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off) 1931{ 1932 dnode_t *dn; 1933 int err; 1934 1935 /* 1936 * Sync any current changes before 1937 * we go trundling through the block pointers. 1938 */ 1939 err = dmu_object_wait_synced(os, object); 1940 if (err) { 1941 return (err); 1942 } 1943 1944 err = dnode_hold(os, object, FTAG, &dn); 1945 if (err) { 1946 return (err); 1947 } 1948 1949 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0); 1950 dnode_rele(dn, FTAG); 1951 1952 return (err); 1953} 1954 1955/* 1956 * Given the ZFS object, if it contains any dirty nodes 1957 * this function flushes all dirty blocks to disk. This 1958 * ensures the DMU object info is updated. A more efficient 1959 * future version might just find the TXG with the maximum 1960 * ID and wait for that to be synced. 1961 */ 1962int 1963dmu_object_wait_synced(objset_t *os, uint64_t object) 1964{ 1965 dnode_t *dn; 1966 int error, i; 1967 1968 error = dnode_hold(os, object, FTAG, &dn); 1969 if (error) { 1970 return (error); 1971 } 1972 1973 for (i = 0; i < TXG_SIZE; i++) { 1974 if (list_link_active(&dn->dn_dirty_link[i])) { 1975 break; 1976 } 1977 } 1978 dnode_rele(dn, FTAG); 1979 if (i != TXG_SIZE) { 1980 txg_wait_synced(dmu_objset_pool(os), 0); 1981 } 1982 1983 return (0); 1984} 1985 1986void 1987dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi) 1988{ 1989 dnode_phys_t *dnp; 1990 1991 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1992 mutex_enter(&dn->dn_mtx); 1993 1994 dnp = dn->dn_phys; 1995 1996 doi->doi_data_block_size = dn->dn_datablksz; 1997 doi->doi_metadata_block_size = dn->dn_indblkshift ? 1998 1ULL << dn->dn_indblkshift : 0; 1999 doi->doi_type = dn->dn_type; 2000 doi->doi_bonus_type = dn->dn_bonustype; 2001 doi->doi_bonus_size = dn->dn_bonuslen; 2002 doi->doi_indirection = dn->dn_nlevels; 2003 doi->doi_checksum = dn->dn_checksum; 2004 doi->doi_compress = dn->dn_compress; 2005 doi->doi_nblkptr = dn->dn_nblkptr; 2006 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9; 2007 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz; 2008 doi->doi_fill_count = 0; 2009 for (int i = 0; i < dnp->dn_nblkptr; i++) 2010 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]); 2011 2012 mutex_exit(&dn->dn_mtx); 2013 rw_exit(&dn->dn_struct_rwlock); 2014} 2015 2016/* 2017 * Get information on a DMU object. 2018 * If doi is NULL, just indicates whether the object exists. 2019 */ 2020int 2021dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi) 2022{ 2023 dnode_t *dn; 2024 int err = dnode_hold(os, object, FTAG, &dn); 2025 2026 if (err) 2027 return (err); 2028 2029 if (doi != NULL) 2030 dmu_object_info_from_dnode(dn, doi); 2031 2032 dnode_rele(dn, FTAG); 2033 return (0); 2034} 2035 2036/* 2037 * As above, but faster; can be used when you have a held dbuf in hand. 2038 */ 2039void 2040dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi) 2041{ 2042 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2043 2044 DB_DNODE_ENTER(db); 2045 dmu_object_info_from_dnode(DB_DNODE(db), doi); 2046 DB_DNODE_EXIT(db); 2047} 2048 2049/* 2050 * Faster still when you only care about the size. 2051 * This is specifically optimized for zfs_getattr(). 2052 */ 2053void 2054dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize, 2055 u_longlong_t *nblk512) 2056{ 2057 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2058 dnode_t *dn; 2059 2060 DB_DNODE_ENTER(db); 2061 dn = DB_DNODE(db); 2062 2063 *blksize = dn->dn_datablksz; 2064 /* add 1 for dnode space */ 2065 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >> 2066 SPA_MINBLOCKSHIFT) + 1; 2067 DB_DNODE_EXIT(db); 2068} 2069 2070void 2071byteswap_uint64_array(void *vbuf, size_t size) 2072{ 2073 uint64_t *buf = vbuf; 2074 size_t count = size >> 3; 2075 int i; 2076 2077 ASSERT((size & 7) == 0); 2078 2079 for (i = 0; i < count; i++) 2080 buf[i] = BSWAP_64(buf[i]); 2081} 2082 2083void 2084byteswap_uint32_array(void *vbuf, size_t size) 2085{ 2086 uint32_t *buf = vbuf; 2087 size_t count = size >> 2; 2088 int i; 2089 2090 ASSERT((size & 3) == 0); 2091 2092 for (i = 0; i < count; i++) 2093 buf[i] = BSWAP_32(buf[i]); 2094} 2095 2096void 2097byteswap_uint16_array(void *vbuf, size_t size) 2098{ 2099 uint16_t *buf = vbuf; 2100 size_t count = size >> 1; 2101 int i; 2102 2103 ASSERT((size & 1) == 0); 2104 2105 for (i = 0; i < count; i++) 2106 buf[i] = BSWAP_16(buf[i]); 2107} 2108 2109/* ARGSUSED */ 2110void 2111byteswap_uint8_array(void *vbuf, size_t size) 2112{ 2113} 2114 2115void 2116dmu_init(void) 2117{ 2118 zfs_dbgmsg_init(); 2119 sa_cache_init(); 2120 xuio_stat_init(); 2121 dmu_objset_init(); 2122 dnode_init(); 2123 zfetch_init(); 2124 zio_compress_init(); 2125 l2arc_init(); 2126 arc_init(); 2127 dbuf_init(); 2128} 2129 2130void 2131dmu_fini(void) 2132{ 2133 arc_fini(); /* arc depends on l2arc, so arc must go first */ 2134 l2arc_fini(); 2135 zfetch_fini(); 2136 zio_compress_fini(); 2137 dbuf_fini(); 2138 dnode_fini(); 2139 dmu_objset_fini(); 2140 xuio_stat_fini(); 2141 sa_cache_fini(); 2142 zfs_dbgmsg_fini(); 2143} 2144