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_by_dnode(dnode_t *dn, uint64_t offset, 135 void *tag, dmu_buf_t **dbp) 136{ 137 uint64_t blkid; 138 dmu_buf_impl_t *db; 139 140 blkid = dbuf_whichblock(dn, 0, offset); 141 rw_enter(&dn->dn_struct_rwlock, RW_READER); 142 db = dbuf_hold(dn, blkid, tag); 143 rw_exit(&dn->dn_struct_rwlock); 144 145 if (db == NULL) { 146 *dbp = NULL; 147 return (SET_ERROR(EIO)); 148 } 149 150 *dbp = &db->db; 151 return (0); 152} 153int 154dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset, 155 void *tag, dmu_buf_t **dbp) 156{ 157 dnode_t *dn; 158 uint64_t blkid; 159 dmu_buf_impl_t *db; 160 int err; 161 162 err = dnode_hold(os, object, FTAG, &dn); 163 if (err) 164 return (err); 165 blkid = dbuf_whichblock(dn, 0, offset); 166 rw_enter(&dn->dn_struct_rwlock, RW_READER); 167 db = dbuf_hold(dn, blkid, tag); 168 rw_exit(&dn->dn_struct_rwlock); 169 dnode_rele(dn, FTAG); 170 171 if (db == NULL) { 172 *dbp = NULL; 173 return (SET_ERROR(EIO)); 174 } 175 176 *dbp = &db->db; 177 return (err); 178} 179 180int 181dmu_buf_hold_by_dnode(dnode_t *dn, uint64_t offset, 182 void *tag, dmu_buf_t **dbp, int flags) 183{ 184 int err; 185 int db_flags = DB_RF_CANFAIL; 186 187 if (flags & DMU_READ_NO_PREFETCH) 188 db_flags |= DB_RF_NOPREFETCH; 189 190 err = dmu_buf_hold_noread_by_dnode(dn, offset, tag, dbp); 191 if (err == 0) { 192 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp); 193 err = dbuf_read(db, NULL, db_flags); 194 if (err != 0) { 195 dbuf_rele(db, tag); 196 *dbp = NULL; 197 } 198 } 199 200 return (err); 201} 202 203int 204dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset, 205 void *tag, dmu_buf_t **dbp, int flags) 206{ 207 int err; 208 int db_flags = DB_RF_CANFAIL; 209 210 if (flags & DMU_READ_NO_PREFETCH) 211 db_flags |= DB_RF_NOPREFETCH; 212 213 err = dmu_buf_hold_noread(os, object, offset, tag, dbp); 214 if (err == 0) { 215 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp); 216 err = dbuf_read(db, NULL, db_flags); 217 if (err != 0) { 218 dbuf_rele(db, tag); 219 *dbp = NULL; 220 } 221 } 222 223 return (err); 224} 225 226int 227dmu_bonus_max(void) 228{ 229 return (DN_MAX_BONUSLEN); 230} 231 232int 233dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx) 234{ 235 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 236 dnode_t *dn; 237 int error; 238 239 DB_DNODE_ENTER(db); 240 dn = DB_DNODE(db); 241 242 if (dn->dn_bonus != db) { 243 error = SET_ERROR(EINVAL); 244 } else if (newsize < 0 || newsize > db_fake->db_size) { 245 error = SET_ERROR(EINVAL); 246 } else { 247 dnode_setbonuslen(dn, newsize, tx); 248 error = 0; 249 } 250 251 DB_DNODE_EXIT(db); 252 return (error); 253} 254 255int 256dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx) 257{ 258 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 259 dnode_t *dn; 260 int error; 261 262 DB_DNODE_ENTER(db); 263 dn = DB_DNODE(db); 264 265 if (!DMU_OT_IS_VALID(type)) { 266 error = SET_ERROR(EINVAL); 267 } else if (dn->dn_bonus != db) { 268 error = SET_ERROR(EINVAL); 269 } else { 270 dnode_setbonus_type(dn, type, tx); 271 error = 0; 272 } 273 274 DB_DNODE_EXIT(db); 275 return (error); 276} 277 278dmu_object_type_t 279dmu_get_bonustype(dmu_buf_t *db_fake) 280{ 281 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 282 dnode_t *dn; 283 dmu_object_type_t type; 284 285 DB_DNODE_ENTER(db); 286 dn = DB_DNODE(db); 287 type = dn->dn_bonustype; 288 DB_DNODE_EXIT(db); 289 290 return (type); 291} 292 293int 294dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx) 295{ 296 dnode_t *dn; 297 int error; 298 299 error = dnode_hold(os, object, FTAG, &dn); 300 dbuf_rm_spill(dn, tx); 301 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 302 dnode_rm_spill(dn, tx); 303 rw_exit(&dn->dn_struct_rwlock); 304 dnode_rele(dn, FTAG); 305 return (error); 306} 307 308/* 309 * returns ENOENT, EIO, or 0. 310 */ 311int 312dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp) 313{ 314 dnode_t *dn; 315 dmu_buf_impl_t *db; 316 int error; 317 318 error = dnode_hold(os, object, FTAG, &dn); 319 if (error) 320 return (error); 321 322 rw_enter(&dn->dn_struct_rwlock, RW_READER); 323 if (dn->dn_bonus == NULL) { 324 rw_exit(&dn->dn_struct_rwlock); 325 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 326 if (dn->dn_bonus == NULL) 327 dbuf_create_bonus(dn); 328 } 329 db = dn->dn_bonus; 330 331 /* as long as the bonus buf is held, the dnode will be held */ 332 if (refcount_add(&db->db_holds, tag) == 1) { 333 VERIFY(dnode_add_ref(dn, db)); 334 atomic_inc_32(&dn->dn_dbufs_count); 335 } 336 337 /* 338 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's 339 * hold and incrementing the dbuf count to ensure that dnode_move() sees 340 * a dnode hold for every dbuf. 341 */ 342 rw_exit(&dn->dn_struct_rwlock); 343 344 dnode_rele(dn, FTAG); 345 346 VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH)); 347 348 *dbp = &db->db; 349 return (0); 350} 351 352/* 353 * returns ENOENT, EIO, or 0. 354 * 355 * This interface will allocate a blank spill dbuf when a spill blk 356 * doesn't already exist on the dnode. 357 * 358 * if you only want to find an already existing spill db, then 359 * dmu_spill_hold_existing() should be used. 360 */ 361int 362dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp) 363{ 364 dmu_buf_impl_t *db = NULL; 365 int err; 366 367 if ((flags & DB_RF_HAVESTRUCT) == 0) 368 rw_enter(&dn->dn_struct_rwlock, RW_READER); 369 370 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag); 371 372 if ((flags & DB_RF_HAVESTRUCT) == 0) 373 rw_exit(&dn->dn_struct_rwlock); 374 375 ASSERT(db != NULL); 376 err = dbuf_read(db, NULL, flags); 377 if (err == 0) 378 *dbp = &db->db; 379 else 380 dbuf_rele(db, tag); 381 return (err); 382} 383 384int 385dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp) 386{ 387 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus; 388 dnode_t *dn; 389 int err; 390 391 DB_DNODE_ENTER(db); 392 dn = DB_DNODE(db); 393 394 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) { 395 err = SET_ERROR(EINVAL); 396 } else { 397 rw_enter(&dn->dn_struct_rwlock, RW_READER); 398 399 if (!dn->dn_have_spill) { 400 err = SET_ERROR(ENOENT); 401 } else { 402 err = dmu_spill_hold_by_dnode(dn, 403 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp); 404 } 405 406 rw_exit(&dn->dn_struct_rwlock); 407 } 408 409 DB_DNODE_EXIT(db); 410 return (err); 411} 412 413int 414dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp) 415{ 416 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus; 417 dnode_t *dn; 418 int err; 419 420 DB_DNODE_ENTER(db); 421 dn = DB_DNODE(db); 422 err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp); 423 DB_DNODE_EXIT(db); 424 425 return (err); 426} 427 428/* 429 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces 430 * to take a held dnode rather than <os, object> -- the lookup is wasteful, 431 * and can induce severe lock contention when writing to several files 432 * whose dnodes are in the same block. 433 */ 434static int 435dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length, 436 boolean_t read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags) 437{ 438 dmu_buf_t **dbp; 439 uint64_t blkid, nblks, i; 440 uint32_t dbuf_flags; 441 int err; 442 zio_t *zio; 443 444 ASSERT(length <= DMU_MAX_ACCESS); 445 446 /* 447 * Note: We directly notify the prefetch code of this read, so that 448 * we can tell it about the multi-block read. dbuf_read() only knows 449 * about the one block it is accessing. 450 */ 451 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT | 452 DB_RF_NOPREFETCH; 453 454 rw_enter(&dn->dn_struct_rwlock, RW_READER); 455 if (dn->dn_datablkshift) { 456 int blkshift = dn->dn_datablkshift; 457 nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) - 458 P2ALIGN(offset, 1ULL << blkshift)) >> blkshift; 459 } else { 460 if (offset + length > dn->dn_datablksz) { 461 zfs_panic_recover("zfs: accessing past end of object " 462 "%llx/%llx (size=%u access=%llu+%llu)", 463 (longlong_t)dn->dn_objset-> 464 os_dsl_dataset->ds_object, 465 (longlong_t)dn->dn_object, dn->dn_datablksz, 466 (longlong_t)offset, (longlong_t)length); 467 rw_exit(&dn->dn_struct_rwlock); 468 return (SET_ERROR(EIO)); 469 } 470 nblks = 1; 471 } 472 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP); 473 474 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL); 475 blkid = dbuf_whichblock(dn, 0, offset); 476 for (i = 0; i < nblks; i++) { 477 dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag); 478 if (db == NULL) { 479 rw_exit(&dn->dn_struct_rwlock); 480 dmu_buf_rele_array(dbp, nblks, tag); 481 zio_nowait(zio); 482 return (SET_ERROR(EIO)); 483 } 484 485 /* initiate async i/o */ 486 if (read) 487 (void) dbuf_read(db, zio, dbuf_flags); 488#ifdef _KERNEL 489 else 490 curthread->td_ru.ru_oublock++; 491#endif 492 dbp[i] = &db->db; 493 } 494 495 if ((flags & DMU_READ_NO_PREFETCH) == 0 && 496 DNODE_META_IS_CACHEABLE(dn) && length <= zfetch_array_rd_sz) { 497 dmu_zfetch(&dn->dn_zfetch, blkid, nblks, 498 read && DNODE_IS_CACHEABLE(dn)); 499 } 500 rw_exit(&dn->dn_struct_rwlock); 501 502 /* wait for async i/o */ 503 err = zio_wait(zio); 504 if (err) { 505 dmu_buf_rele_array(dbp, nblks, tag); 506 return (err); 507 } 508 509 /* wait for other io to complete */ 510 if (read) { 511 for (i = 0; i < nblks; i++) { 512 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i]; 513 mutex_enter(&db->db_mtx); 514 while (db->db_state == DB_READ || 515 db->db_state == DB_FILL) 516 cv_wait(&db->db_changed, &db->db_mtx); 517 if (db->db_state == DB_UNCACHED) 518 err = SET_ERROR(EIO); 519 mutex_exit(&db->db_mtx); 520 if (err) { 521 dmu_buf_rele_array(dbp, nblks, tag); 522 return (err); 523 } 524 } 525 } 526 527 *numbufsp = nblks; 528 *dbpp = dbp; 529 return (0); 530} 531 532static int 533dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset, 534 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp) 535{ 536 dnode_t *dn; 537 int err; 538 539 err = dnode_hold(os, object, FTAG, &dn); 540 if (err) 541 return (err); 542 543 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag, 544 numbufsp, dbpp, DMU_READ_PREFETCH); 545 546 dnode_rele(dn, FTAG); 547 548 return (err); 549} 550 551int 552dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset, 553 uint64_t length, boolean_t read, void *tag, int *numbufsp, 554 dmu_buf_t ***dbpp) 555{ 556 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 557 dnode_t *dn; 558 int err; 559 560 DB_DNODE_ENTER(db); 561 dn = DB_DNODE(db); 562 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag, 563 numbufsp, dbpp, DMU_READ_PREFETCH); 564 DB_DNODE_EXIT(db); 565 566 return (err); 567} 568 569void 570dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag) 571{ 572 int i; 573 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake; 574 575 if (numbufs == 0) 576 return; 577 578 for (i = 0; i < numbufs; i++) { 579 if (dbp[i]) 580 dbuf_rele(dbp[i], tag); 581 } 582 583 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs); 584} 585 586/* 587 * Issue prefetch i/os for the given blocks. If level is greater than 0, the 588 * indirect blocks prefeteched will be those that point to the blocks containing 589 * the data starting at offset, and continuing to offset + len. 590 * 591 * Note that if the indirect blocks above the blocks being prefetched are not in 592 * cache, they will be asychronously read in. 593 */ 594void 595dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset, 596 uint64_t len, zio_priority_t pri) 597{ 598 dnode_t *dn; 599 uint64_t blkid; 600 int nblks, err; 601 602 if (len == 0) { /* they're interested in the bonus buffer */ 603 dn = DMU_META_DNODE(os); 604 605 if (object == 0 || object >= DN_MAX_OBJECT) 606 return; 607 608 rw_enter(&dn->dn_struct_rwlock, RW_READER); 609 blkid = dbuf_whichblock(dn, level, 610 object * sizeof (dnode_phys_t)); 611 dbuf_prefetch(dn, level, blkid, pri, 0); 612 rw_exit(&dn->dn_struct_rwlock); 613 return; 614 } 615 616 /* 617 * XXX - Note, if the dnode for the requested object is not 618 * already cached, we will do a *synchronous* read in the 619 * dnode_hold() call. The same is true for any indirects. 620 */ 621 err = dnode_hold(os, object, FTAG, &dn); 622 if (err != 0) 623 return; 624 625 rw_enter(&dn->dn_struct_rwlock, RW_READER); 626 /* 627 * offset + len - 1 is the last byte we want to prefetch for, and offset 628 * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the 629 * last block we want to prefetch, and dbuf_whichblock(dn, level, 630 * offset) is the first. Then the number we need to prefetch is the 631 * last - first + 1. 632 */ 633 if (level > 0 || dn->dn_datablkshift != 0) { 634 nblks = dbuf_whichblock(dn, level, offset + len - 1) - 635 dbuf_whichblock(dn, level, offset) + 1; 636 } else { 637 nblks = (offset < dn->dn_datablksz); 638 } 639 640 if (nblks != 0) { 641 blkid = dbuf_whichblock(dn, level, offset); 642 for (int i = 0; i < nblks; i++) 643 dbuf_prefetch(dn, level, blkid + i, pri, 0); 644 } 645 646 rw_exit(&dn->dn_struct_rwlock); 647 648 dnode_rele(dn, FTAG); 649} 650 651/* 652 * Get the next "chunk" of file data to free. We traverse the file from 653 * the end so that the file gets shorter over time (if we crashes in the 654 * middle, this will leave us in a better state). We find allocated file 655 * data by simply searching the allocated level 1 indirects. 656 * 657 * On input, *start should be the first offset that does not need to be 658 * freed (e.g. "offset + length"). On return, *start will be the first 659 * offset that should be freed. 660 */ 661static int 662get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum) 663{ 664 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1); 665 /* bytes of data covered by a level-1 indirect block */ 666 uint64_t iblkrange = 667 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT); 668 669 ASSERT3U(minimum, <=, *start); 670 671 if (*start - minimum <= iblkrange * maxblks) { 672 *start = minimum; 673 return (0); 674 } 675 ASSERT(ISP2(iblkrange)); 676 677 for (uint64_t blks = 0; *start > minimum && blks < maxblks; blks++) { 678 int err; 679 680 /* 681 * dnode_next_offset(BACKWARDS) will find an allocated L1 682 * indirect block at or before the input offset. We must 683 * decrement *start so that it is at the end of the region 684 * to search. 685 */ 686 (*start)--; 687 err = dnode_next_offset(dn, 688 DNODE_FIND_BACKWARDS, start, 2, 1, 0); 689 690 /* if there are no indirect blocks before start, we are done */ 691 if (err == ESRCH) { 692 *start = minimum; 693 break; 694 } else if (err != 0) { 695 return (err); 696 } 697 698 /* set start to the beginning of this L1 indirect */ 699 *start = P2ALIGN(*start, iblkrange); 700 } 701 if (*start < minimum) 702 *start = minimum; 703 return (0); 704} 705 706static int 707dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset, 708 uint64_t length) 709{ 710 uint64_t object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz; 711 int err; 712 713 if (offset >= object_size) 714 return (0); 715 716 if (length == DMU_OBJECT_END || offset + length > object_size) 717 length = object_size - offset; 718 719 while (length != 0) { 720 uint64_t chunk_end, chunk_begin; 721 722 chunk_end = chunk_begin = offset + length; 723 724 /* move chunk_begin backwards to the beginning of this chunk */ 725 err = get_next_chunk(dn, &chunk_begin, offset); 726 if (err) 727 return (err); 728 ASSERT3U(chunk_begin, >=, offset); 729 ASSERT3U(chunk_begin, <=, chunk_end); 730 731 dmu_tx_t *tx = dmu_tx_create(os); 732 dmu_tx_hold_free(tx, dn->dn_object, 733 chunk_begin, chunk_end - chunk_begin); 734 735 /* 736 * Mark this transaction as typically resulting in a net 737 * reduction in space used. 738 */ 739 dmu_tx_mark_netfree(tx); 740 err = dmu_tx_assign(tx, TXG_WAIT); 741 if (err) { 742 dmu_tx_abort(tx); 743 return (err); 744 } 745 dnode_free_range(dn, chunk_begin, chunk_end - chunk_begin, tx); 746 dmu_tx_commit(tx); 747 748 length -= chunk_end - chunk_begin; 749 } 750 return (0); 751} 752 753int 754dmu_free_long_range(objset_t *os, uint64_t object, 755 uint64_t offset, uint64_t length) 756{ 757 dnode_t *dn; 758 int err; 759 760 err = dnode_hold(os, object, FTAG, &dn); 761 if (err != 0) 762 return (err); 763 err = dmu_free_long_range_impl(os, dn, offset, length); 764 765 /* 766 * It is important to zero out the maxblkid when freeing the entire 767 * file, so that (a) subsequent calls to dmu_free_long_range_impl() 768 * will take the fast path, and (b) dnode_reallocate() can verify 769 * that the entire file has been freed. 770 */ 771 if (err == 0 && offset == 0 && length == DMU_OBJECT_END) 772 dn->dn_maxblkid = 0; 773 774 dnode_rele(dn, FTAG); 775 return (err); 776} 777 778int 779dmu_free_long_object(objset_t *os, uint64_t object) 780{ 781 dmu_tx_t *tx; 782 int err; 783 784 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END); 785 if (err != 0) 786 return (err); 787 788 tx = dmu_tx_create(os); 789 dmu_tx_hold_bonus(tx, object); 790 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END); 791 dmu_tx_mark_netfree(tx); 792 err = dmu_tx_assign(tx, TXG_WAIT); 793 if (err == 0) { 794 err = dmu_object_free(os, object, tx); 795 dmu_tx_commit(tx); 796 } else { 797 dmu_tx_abort(tx); 798 } 799 800 return (err); 801} 802 803int 804dmu_free_range(objset_t *os, uint64_t object, uint64_t offset, 805 uint64_t size, dmu_tx_t *tx) 806{ 807 dnode_t *dn; 808 int err = dnode_hold(os, object, FTAG, &dn); 809 if (err) 810 return (err); 811 ASSERT(offset < UINT64_MAX); 812 ASSERT(size == -1ULL || size <= UINT64_MAX - offset); 813 dnode_free_range(dn, offset, size, tx); 814 dnode_rele(dn, FTAG); 815 return (0); 816} 817 818int 819dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 820 void *buf, uint32_t flags) 821{ 822 dnode_t *dn; 823 dmu_buf_t **dbp; 824 int numbufs, err; 825 826 err = dnode_hold(os, object, FTAG, &dn); 827 if (err) 828 return (err); 829 830 /* 831 * Deal with odd block sizes, where there can't be data past the first 832 * block. If we ever do the tail block optimization, we will need to 833 * handle that here as well. 834 */ 835 if (dn->dn_maxblkid == 0) { 836 int newsz = offset > dn->dn_datablksz ? 0 : 837 MIN(size, dn->dn_datablksz - offset); 838 bzero((char *)buf + newsz, size - newsz); 839 size = newsz; 840 } 841 842 while (size > 0) { 843 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2); 844 int i; 845 846 /* 847 * NB: we could do this block-at-a-time, but it's nice 848 * to be reading in parallel. 849 */ 850 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen, 851 TRUE, FTAG, &numbufs, &dbp, flags); 852 if (err) 853 break; 854 855 for (i = 0; i < numbufs; i++) { 856 int tocpy; 857 int bufoff; 858 dmu_buf_t *db = dbp[i]; 859 860 ASSERT(size > 0); 861 862 bufoff = offset - db->db_offset; 863 tocpy = (int)MIN(db->db_size - bufoff, size); 864 865 bcopy((char *)db->db_data + bufoff, buf, tocpy); 866 867 offset += tocpy; 868 size -= tocpy; 869 buf = (char *)buf + tocpy; 870 } 871 dmu_buf_rele_array(dbp, numbufs, FTAG); 872 } 873 dnode_rele(dn, FTAG); 874 return (err); 875} 876 877void 878dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 879 const void *buf, dmu_tx_t *tx) 880{ 881 dmu_buf_t **dbp; 882 int numbufs, i; 883 884 if (size == 0) 885 return; 886 887 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size, 888 FALSE, FTAG, &numbufs, &dbp)); 889 890 for (i = 0; i < numbufs; i++) { 891 int tocpy; 892 int bufoff; 893 dmu_buf_t *db = dbp[i]; 894 895 ASSERT(size > 0); 896 897 bufoff = offset - db->db_offset; 898 tocpy = (int)MIN(db->db_size - bufoff, size); 899 900 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 901 902 if (tocpy == db->db_size) 903 dmu_buf_will_fill(db, tx); 904 else 905 dmu_buf_will_dirty(db, tx); 906 907 bcopy(buf, (char *)db->db_data + bufoff, tocpy); 908 909 if (tocpy == db->db_size) 910 dmu_buf_fill_done(db, tx); 911 912 offset += tocpy; 913 size -= tocpy; 914 buf = (char *)buf + tocpy; 915 } 916 dmu_buf_rele_array(dbp, numbufs, FTAG); 917} 918 919void 920dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 921 dmu_tx_t *tx) 922{ 923 dmu_buf_t **dbp; 924 int numbufs, i; 925 926 if (size == 0) 927 return; 928 929 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size, 930 FALSE, FTAG, &numbufs, &dbp)); 931 932 for (i = 0; i < numbufs; i++) { 933 dmu_buf_t *db = dbp[i]; 934 935 dmu_buf_will_not_fill(db, tx); 936 } 937 dmu_buf_rele_array(dbp, numbufs, FTAG); 938} 939 940void 941dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset, 942 void *data, uint8_t etype, uint8_t comp, int uncompressed_size, 943 int compressed_size, int byteorder, dmu_tx_t *tx) 944{ 945 dmu_buf_t *db; 946 947 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES); 948 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS); 949 VERIFY0(dmu_buf_hold_noread(os, object, offset, 950 FTAG, &db)); 951 952 dmu_buf_write_embedded(db, 953 data, (bp_embedded_type_t)etype, (enum zio_compress)comp, 954 uncompressed_size, compressed_size, byteorder, tx); 955 956 dmu_buf_rele(db, FTAG); 957} 958 959/* 960 * DMU support for xuio 961 */ 962kstat_t *xuio_ksp = NULL; 963 964int 965dmu_xuio_init(xuio_t *xuio, int nblk) 966{ 967 dmu_xuio_t *priv; 968 uio_t *uio = &xuio->xu_uio; 969 970 uio->uio_iovcnt = nblk; 971 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP); 972 973 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP); 974 priv->cnt = nblk; 975 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP); 976 priv->iovp = uio->uio_iov; 977 XUIO_XUZC_PRIV(xuio) = priv; 978 979 if (XUIO_XUZC_RW(xuio) == UIO_READ) 980 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk); 981 else 982 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk); 983 984 return (0); 985} 986 987void 988dmu_xuio_fini(xuio_t *xuio) 989{ 990 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 991 int nblk = priv->cnt; 992 993 kmem_free(priv->iovp, nblk * sizeof (iovec_t)); 994 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *)); 995 kmem_free(priv, sizeof (dmu_xuio_t)); 996 997 if (XUIO_XUZC_RW(xuio) == UIO_READ) 998 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk); 999 else 1000 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk); 1001} 1002 1003/* 1004 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf } 1005 * and increase priv->next by 1. 1006 */ 1007int 1008dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n) 1009{ 1010 struct iovec *iov; 1011 uio_t *uio = &xuio->xu_uio; 1012 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 1013 int i = priv->next++; 1014 1015 ASSERT(i < priv->cnt); 1016 ASSERT(off + n <= arc_buf_size(abuf)); 1017 iov = uio->uio_iov + i; 1018 iov->iov_base = (char *)abuf->b_data + off; 1019 iov->iov_len = n; 1020 priv->bufs[i] = abuf; 1021 return (0); 1022} 1023 1024int 1025dmu_xuio_cnt(xuio_t *xuio) 1026{ 1027 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 1028 return (priv->cnt); 1029} 1030 1031arc_buf_t * 1032dmu_xuio_arcbuf(xuio_t *xuio, int i) 1033{ 1034 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 1035 1036 ASSERT(i < priv->cnt); 1037 return (priv->bufs[i]); 1038} 1039 1040void 1041dmu_xuio_clear(xuio_t *xuio, int i) 1042{ 1043 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 1044 1045 ASSERT(i < priv->cnt); 1046 priv->bufs[i] = NULL; 1047} 1048 1049static void 1050xuio_stat_init(void) 1051{ 1052 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc", 1053 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t), 1054 KSTAT_FLAG_VIRTUAL); 1055 if (xuio_ksp != NULL) { 1056 xuio_ksp->ks_data = &xuio_stats; 1057 kstat_install(xuio_ksp); 1058 } 1059} 1060 1061static void 1062xuio_stat_fini(void) 1063{ 1064 if (xuio_ksp != NULL) { 1065 kstat_delete(xuio_ksp); 1066 xuio_ksp = NULL; 1067 } 1068} 1069 1070void 1071xuio_stat_wbuf_copied() 1072{ 1073 XUIOSTAT_BUMP(xuiostat_wbuf_copied); 1074} 1075 1076void 1077xuio_stat_wbuf_nocopy() 1078{ 1079 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy); 1080} 1081 1082#ifdef _KERNEL 1083static int 1084dmu_read_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size) 1085{ 1086 dmu_buf_t **dbp; 1087 int numbufs, i, err; 1088 xuio_t *xuio = NULL; 1089 1090 /* 1091 * NB: we could do this block-at-a-time, but it's nice 1092 * to be reading in parallel. 1093 */ 1094 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size, 1095 TRUE, FTAG, &numbufs, &dbp, 0); 1096 if (err) 1097 return (err); 1098 1099#ifdef UIO_XUIO 1100 if (uio->uio_extflg == UIO_XUIO) 1101 xuio = (xuio_t *)uio; 1102#endif 1103 1104 for (i = 0; i < numbufs; i++) { 1105 int tocpy; 1106 int bufoff; 1107 dmu_buf_t *db = dbp[i]; 1108 1109 ASSERT(size > 0); 1110 1111 bufoff = uio->uio_loffset - db->db_offset; 1112 tocpy = (int)MIN(db->db_size - bufoff, size); 1113 1114 if (xuio) { 1115 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 1116 arc_buf_t *dbuf_abuf = dbi->db_buf; 1117 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi); 1118 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy); 1119 if (!err) { 1120 uio->uio_resid -= tocpy; 1121 uio->uio_loffset += tocpy; 1122 } 1123 1124 if (abuf == dbuf_abuf) 1125 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy); 1126 else 1127 XUIOSTAT_BUMP(xuiostat_rbuf_copied); 1128 } else { 1129#ifdef illumos 1130 err = uiomove((char *)db->db_data + bufoff, tocpy, 1131 UIO_READ, uio); 1132#else 1133 err = vn_io_fault_uiomove((char *)db->db_data + bufoff, 1134 tocpy, uio); 1135#endif 1136 } 1137 if (err) 1138 break; 1139 1140 size -= tocpy; 1141 } 1142 dmu_buf_rele_array(dbp, numbufs, FTAG); 1143 1144 return (err); 1145} 1146 1147/* 1148 * Read 'size' bytes into the uio buffer. 1149 * From object zdb->db_object. 1150 * Starting at offset uio->uio_loffset. 1151 * 1152 * If the caller already has a dbuf in the target object 1153 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(), 1154 * because we don't have to find the dnode_t for the object. 1155 */ 1156int 1157dmu_read_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size) 1158{ 1159 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb; 1160 dnode_t *dn; 1161 int err; 1162 1163 if (size == 0) 1164 return (0); 1165 1166 DB_DNODE_ENTER(db); 1167 dn = DB_DNODE(db); 1168 err = dmu_read_uio_dnode(dn, uio, size); 1169 DB_DNODE_EXIT(db); 1170 1171 return (err); 1172} 1173 1174/* 1175 * Read 'size' bytes into the uio buffer. 1176 * From the specified object 1177 * Starting at offset uio->uio_loffset. 1178 */ 1179int 1180dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size) 1181{ 1182 dnode_t *dn; 1183 int err; 1184 1185 if (size == 0) 1186 return (0); 1187 1188 err = dnode_hold(os, object, FTAG, &dn); 1189 if (err) 1190 return (err); 1191 1192 err = dmu_read_uio_dnode(dn, uio, size); 1193 1194 dnode_rele(dn, FTAG); 1195 1196 return (err); 1197} 1198 1199static int 1200dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx) 1201{ 1202 dmu_buf_t **dbp; 1203 int numbufs; 1204 int err = 0; 1205 int i; 1206 1207 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size, 1208 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH); 1209 if (err) 1210 return (err); 1211 1212 for (i = 0; i < numbufs; i++) { 1213 int tocpy; 1214 int bufoff; 1215 dmu_buf_t *db = dbp[i]; 1216 1217 ASSERT(size > 0); 1218 1219 bufoff = uio->uio_loffset - db->db_offset; 1220 tocpy = (int)MIN(db->db_size - bufoff, size); 1221 1222 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 1223 1224 if (tocpy == db->db_size) 1225 dmu_buf_will_fill(db, tx); 1226 else 1227 dmu_buf_will_dirty(db, tx); 1228 1229#ifdef illumos 1230 /* 1231 * XXX uiomove could block forever (eg. nfs-backed 1232 * pages). There needs to be a uiolockdown() function 1233 * to lock the pages in memory, so that uiomove won't 1234 * block. 1235 */ 1236 err = uiomove((char *)db->db_data + bufoff, tocpy, 1237 UIO_WRITE, uio); 1238#else 1239 err = vn_io_fault_uiomove((char *)db->db_data + bufoff, tocpy, 1240 uio); 1241#endif 1242 1243 if (tocpy == db->db_size) 1244 dmu_buf_fill_done(db, tx); 1245 1246 if (err) 1247 break; 1248 1249 size -= tocpy; 1250 } 1251 1252 dmu_buf_rele_array(dbp, numbufs, FTAG); 1253 return (err); 1254} 1255 1256/* 1257 * Write 'size' bytes from the uio buffer. 1258 * To object zdb->db_object. 1259 * Starting at offset uio->uio_loffset. 1260 * 1261 * If the caller already has a dbuf in the target object 1262 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(), 1263 * because we don't have to find the dnode_t for the object. 1264 */ 1265int 1266dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size, 1267 dmu_tx_t *tx) 1268{ 1269 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb; 1270 dnode_t *dn; 1271 int err; 1272 1273 if (size == 0) 1274 return (0); 1275 1276 DB_DNODE_ENTER(db); 1277 dn = DB_DNODE(db); 1278 err = dmu_write_uio_dnode(dn, uio, size, tx); 1279 DB_DNODE_EXIT(db); 1280 1281 return (err); 1282} 1283 1284/* 1285 * Write 'size' bytes from the uio buffer. 1286 * To the specified object. 1287 * Starting at offset uio->uio_loffset. 1288 */ 1289int 1290dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size, 1291 dmu_tx_t *tx) 1292{ 1293 dnode_t *dn; 1294 int err; 1295 1296 if (size == 0) 1297 return (0); 1298 1299 err = dnode_hold(os, object, FTAG, &dn); 1300 if (err) 1301 return (err); 1302 1303 err = dmu_write_uio_dnode(dn, uio, size, tx); 1304 1305 dnode_rele(dn, FTAG); 1306 1307 return (err); 1308} 1309 1310#ifdef illumos 1311int 1312dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1313 page_t *pp, dmu_tx_t *tx) 1314{ 1315 dmu_buf_t **dbp; 1316 int numbufs, i; 1317 int err; 1318 1319 if (size == 0) 1320 return (0); 1321 1322 err = dmu_buf_hold_array(os, object, offset, size, 1323 FALSE, FTAG, &numbufs, &dbp); 1324 if (err) 1325 return (err); 1326 1327 for (i = 0; i < numbufs; i++) { 1328 int tocpy, copied, thiscpy; 1329 int bufoff; 1330 dmu_buf_t *db = dbp[i]; 1331 caddr_t va; 1332 1333 ASSERT(size > 0); 1334 ASSERT3U(db->db_size, >=, PAGESIZE); 1335 1336 bufoff = offset - db->db_offset; 1337 tocpy = (int)MIN(db->db_size - bufoff, size); 1338 1339 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 1340 1341 if (tocpy == db->db_size) 1342 dmu_buf_will_fill(db, tx); 1343 else 1344 dmu_buf_will_dirty(db, tx); 1345 1346 for (copied = 0; copied < tocpy; copied += PAGESIZE) { 1347 ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff); 1348 thiscpy = MIN(PAGESIZE, tocpy - copied); 1349 va = zfs_map_page(pp, S_READ); 1350 bcopy(va, (char *)db->db_data + bufoff, thiscpy); 1351 zfs_unmap_page(pp, va); 1352 pp = pp->p_next; 1353 bufoff += PAGESIZE; 1354 } 1355 1356 if (tocpy == db->db_size) 1357 dmu_buf_fill_done(db, tx); 1358 1359 offset += tocpy; 1360 size -= tocpy; 1361 } 1362 dmu_buf_rele_array(dbp, numbufs, FTAG); 1363 return (err); 1364} 1365 1366#else /* !illumos */ 1367 1368int 1369dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1370 vm_page_t *ma, dmu_tx_t *tx) 1371{ 1372 dmu_buf_t **dbp; 1373 struct sf_buf *sf; 1374 int numbufs, i; 1375 int err; 1376 1377 if (size == 0) 1378 return (0); 1379 1380 err = dmu_buf_hold_array(os, object, offset, size, 1381 FALSE, FTAG, &numbufs, &dbp); 1382 if (err) 1383 return (err); 1384 1385 for (i = 0; i < numbufs; i++) { 1386 int tocpy, copied, thiscpy; 1387 int bufoff; 1388 dmu_buf_t *db = dbp[i]; 1389 caddr_t va; 1390 1391 ASSERT(size > 0); 1392 ASSERT3U(db->db_size, >=, PAGESIZE); 1393 1394 bufoff = offset - db->db_offset; 1395 tocpy = (int)MIN(db->db_size - bufoff, size); 1396 1397 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 1398 1399 if (tocpy == db->db_size) 1400 dmu_buf_will_fill(db, tx); 1401 else 1402 dmu_buf_will_dirty(db, tx); 1403 1404 for (copied = 0; copied < tocpy; copied += PAGESIZE) { 1405 ASSERT3U(ptoa((*ma)->pindex), ==, db->db_offset + bufoff); 1406 thiscpy = MIN(PAGESIZE, tocpy - copied); 1407 va = zfs_map_page(*ma, &sf); 1408 bcopy(va, (char *)db->db_data + bufoff, thiscpy); 1409 zfs_unmap_page(sf); 1410 ma += 1; 1411 bufoff += PAGESIZE; 1412 } 1413 1414 if (tocpy == db->db_size) 1415 dmu_buf_fill_done(db, tx); 1416 1417 offset += tocpy; 1418 size -= tocpy; 1419 } 1420 dmu_buf_rele_array(dbp, numbufs, FTAG); 1421 return (err); 1422} 1423#endif /* illumos */ 1424#endif /* _KERNEL */ 1425 1426/* 1427 * Allocate a loaned anonymous arc buffer. 1428 */ 1429arc_buf_t * 1430dmu_request_arcbuf(dmu_buf_t *handle, int size) 1431{ 1432 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle; 1433 1434 return (arc_loan_buf(db->db_objset->os_spa, size)); 1435} 1436 1437/* 1438 * Free a loaned arc buffer. 1439 */ 1440void 1441dmu_return_arcbuf(arc_buf_t *buf) 1442{ 1443 arc_return_buf(buf, FTAG); 1444 arc_buf_destroy(buf, FTAG); 1445} 1446 1447/* 1448 * When possible directly assign passed loaned arc buffer to a dbuf. 1449 * If this is not possible copy the contents of passed arc buf via 1450 * dmu_write(). 1451 */ 1452void 1453dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf, 1454 dmu_tx_t *tx) 1455{ 1456 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle; 1457 dnode_t *dn; 1458 dmu_buf_impl_t *db; 1459 uint32_t blksz = (uint32_t)arc_buf_size(buf); 1460 uint64_t blkid; 1461 1462 DB_DNODE_ENTER(dbuf); 1463 dn = DB_DNODE(dbuf); 1464 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1465 blkid = dbuf_whichblock(dn, 0, offset); 1466 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL); 1467 rw_exit(&dn->dn_struct_rwlock); 1468 DB_DNODE_EXIT(dbuf); 1469 1470 /* 1471 * We can only assign if the offset is aligned, the arc buf is the 1472 * same size as the dbuf, and the dbuf is not metadata. It 1473 * can't be metadata because the loaned arc buf comes from the 1474 * user-data kmem arena. 1475 */ 1476 if (offset == db->db.db_offset && blksz == db->db.db_size && 1477 DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA) { 1478#ifdef _KERNEL 1479 curthread->td_ru.ru_oublock++; 1480#endif 1481 dbuf_assign_arcbuf(db, buf, tx); 1482 dbuf_rele(db, FTAG); 1483 } else { 1484 objset_t *os; 1485 uint64_t object; 1486 1487 DB_DNODE_ENTER(dbuf); 1488 dn = DB_DNODE(dbuf); 1489 os = dn->dn_objset; 1490 object = dn->dn_object; 1491 DB_DNODE_EXIT(dbuf); 1492 1493 dbuf_rele(db, FTAG); 1494 dmu_write(os, object, offset, blksz, buf->b_data, tx); 1495 dmu_return_arcbuf(buf); 1496 XUIOSTAT_BUMP(xuiostat_wbuf_copied); 1497 } 1498} 1499 1500typedef struct { 1501 dbuf_dirty_record_t *dsa_dr; 1502 dmu_sync_cb_t *dsa_done; 1503 zgd_t *dsa_zgd; 1504 dmu_tx_t *dsa_tx; 1505} dmu_sync_arg_t; 1506 1507/* ARGSUSED */ 1508static void 1509dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg) 1510{ 1511 dmu_sync_arg_t *dsa = varg; 1512 dmu_buf_t *db = dsa->dsa_zgd->zgd_db; 1513 blkptr_t *bp = zio->io_bp; 1514 1515 if (zio->io_error == 0) { 1516 if (BP_IS_HOLE(bp)) { 1517 /* 1518 * A block of zeros may compress to a hole, but the 1519 * block size still needs to be known for replay. 1520 */ 1521 BP_SET_LSIZE(bp, db->db_size); 1522 } else if (!BP_IS_EMBEDDED(bp)) { 1523 ASSERT(BP_GET_LEVEL(bp) == 0); 1524 bp->blk_fill = 1; 1525 } 1526 } 1527} 1528 1529static void 1530dmu_sync_late_arrival_ready(zio_t *zio) 1531{ 1532 dmu_sync_ready(zio, NULL, zio->io_private); 1533} 1534 1535/* ARGSUSED */ 1536static void 1537dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg) 1538{ 1539 dmu_sync_arg_t *dsa = varg; 1540 dbuf_dirty_record_t *dr = dsa->dsa_dr; 1541 dmu_buf_impl_t *db = dr->dr_dbuf; 1542 1543 mutex_enter(&db->db_mtx); 1544 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC); 1545 if (zio->io_error == 0) { 1546 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE); 1547 if (dr->dt.dl.dr_nopwrite) { 1548 blkptr_t *bp = zio->io_bp; 1549 blkptr_t *bp_orig = &zio->io_bp_orig; 1550 uint8_t chksum = BP_GET_CHECKSUM(bp_orig); 1551 1552 ASSERT(BP_EQUAL(bp, bp_orig)); 1553 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF); 1554 ASSERT(zio_checksum_table[chksum].ci_flags & 1555 ZCHECKSUM_FLAG_NOPWRITE); 1556 } 1557 dr->dt.dl.dr_overridden_by = *zio->io_bp; 1558 dr->dt.dl.dr_override_state = DR_OVERRIDDEN; 1559 dr->dt.dl.dr_copies = zio->io_prop.zp_copies; 1560 1561 /* 1562 * Old style holes are filled with all zeros, whereas 1563 * new-style holes maintain their lsize, type, level, 1564 * and birth time (see zio_write_compress). While we 1565 * need to reset the BP_SET_LSIZE() call that happened 1566 * in dmu_sync_ready for old style holes, we do *not* 1567 * want to wipe out the information contained in new 1568 * style holes. Thus, only zero out the block pointer if 1569 * it's an old style hole. 1570 */ 1571 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) && 1572 dr->dt.dl.dr_overridden_by.blk_birth == 0) 1573 BP_ZERO(&dr->dt.dl.dr_overridden_by); 1574 } else { 1575 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 1576 } 1577 cv_broadcast(&db->db_changed); 1578 mutex_exit(&db->db_mtx); 1579 1580 dsa->dsa_done(dsa->dsa_zgd, zio->io_error); 1581 1582 kmem_free(dsa, sizeof (*dsa)); 1583} 1584 1585static void 1586dmu_sync_late_arrival_done(zio_t *zio) 1587{ 1588 blkptr_t *bp = zio->io_bp; 1589 dmu_sync_arg_t *dsa = zio->io_private; 1590 blkptr_t *bp_orig = &zio->io_bp_orig; 1591 1592 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) { 1593 /* 1594 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE) 1595 * then there is nothing to do here. Otherwise, free the 1596 * newly allocated block in this txg. 1597 */ 1598 if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 1599 ASSERT(BP_EQUAL(bp, bp_orig)); 1600 } else { 1601 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig)); 1602 ASSERT(zio->io_bp->blk_birth == zio->io_txg); 1603 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa)); 1604 zio_free(zio->io_spa, zio->io_txg, zio->io_bp); 1605 } 1606 } 1607 1608 dmu_tx_commit(dsa->dsa_tx); 1609 1610 dsa->dsa_done(dsa->dsa_zgd, zio->io_error); 1611 1612 kmem_free(dsa, sizeof (*dsa)); 1613} 1614 1615static int 1616dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd, 1617 zio_prop_t *zp, zbookmark_phys_t *zb) 1618{ 1619 dmu_sync_arg_t *dsa; 1620 dmu_tx_t *tx; 1621 1622 tx = dmu_tx_create(os); 1623 dmu_tx_hold_space(tx, zgd->zgd_db->db_size); 1624 if (dmu_tx_assign(tx, TXG_WAIT) != 0) { 1625 dmu_tx_abort(tx); 1626 /* Make zl_get_data do txg_waited_synced() */ 1627 return (SET_ERROR(EIO)); 1628 } 1629 1630 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP); 1631 dsa->dsa_dr = NULL; 1632 dsa->dsa_done = done; 1633 dsa->dsa_zgd = zgd; 1634 dsa->dsa_tx = tx; 1635 1636 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), 1637 zgd->zgd_bp, zgd->zgd_db->db_data, zgd->zgd_db->db_size, 1638 zp, dmu_sync_late_arrival_ready, NULL, 1639 NULL, dmu_sync_late_arrival_done, dsa, ZIO_PRIORITY_SYNC_WRITE, 1640 ZIO_FLAG_CANFAIL, zb)); 1641 1642 return (0); 1643} 1644 1645/* 1646 * Intent log support: sync the block associated with db to disk. 1647 * N.B. and XXX: the caller is responsible for making sure that the 1648 * data isn't changing while dmu_sync() is writing it. 1649 * 1650 * Return values: 1651 * 1652 * EEXIST: this txg has already been synced, so there's nothing to do. 1653 * The caller should not log the write. 1654 * 1655 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do. 1656 * The caller should not log the write. 1657 * 1658 * EALREADY: this block is already in the process of being synced. 1659 * The caller should track its progress (somehow). 1660 * 1661 * EIO: could not do the I/O. 1662 * The caller should do a txg_wait_synced(). 1663 * 1664 * 0: the I/O has been initiated. 1665 * The caller should log this blkptr in the done callback. 1666 * It is possible that the I/O will fail, in which case 1667 * the error will be reported to the done callback and 1668 * propagated to pio from zio_done(). 1669 */ 1670int 1671dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd) 1672{ 1673 blkptr_t *bp = zgd->zgd_bp; 1674 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db; 1675 objset_t *os = db->db_objset; 1676 dsl_dataset_t *ds = os->os_dsl_dataset; 1677 dbuf_dirty_record_t *dr; 1678 dmu_sync_arg_t *dsa; 1679 zbookmark_phys_t zb; 1680 zio_prop_t zp; 1681 dnode_t *dn; 1682 1683 ASSERT(pio != NULL); 1684 ASSERT(txg != 0); 1685 1686 SET_BOOKMARK(&zb, ds->ds_object, 1687 db->db.db_object, db->db_level, db->db_blkid); 1688 1689 DB_DNODE_ENTER(db); 1690 dn = DB_DNODE(db); 1691 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp); 1692 DB_DNODE_EXIT(db); 1693 1694 /* 1695 * If we're frozen (running ziltest), we always need to generate a bp. 1696 */ 1697 if (txg > spa_freeze_txg(os->os_spa)) 1698 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); 1699 1700 /* 1701 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf() 1702 * and us. If we determine that this txg is not yet syncing, 1703 * but it begins to sync a moment later, that's OK because the 1704 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx. 1705 */ 1706 mutex_enter(&db->db_mtx); 1707 1708 if (txg <= spa_last_synced_txg(os->os_spa)) { 1709 /* 1710 * This txg has already synced. There's nothing to do. 1711 */ 1712 mutex_exit(&db->db_mtx); 1713 return (SET_ERROR(EEXIST)); 1714 } 1715 1716 if (txg <= spa_syncing_txg(os->os_spa)) { 1717 /* 1718 * This txg is currently syncing, so we can't mess with 1719 * the dirty record anymore; just write a new log block. 1720 */ 1721 mutex_exit(&db->db_mtx); 1722 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); 1723 } 1724 1725 dr = db->db_last_dirty; 1726 while (dr && dr->dr_txg != txg) 1727 dr = dr->dr_next; 1728 1729 if (dr == NULL) { 1730 /* 1731 * There's no dr for this dbuf, so it must have been freed. 1732 * There's no need to log writes to freed blocks, so we're done. 1733 */ 1734 mutex_exit(&db->db_mtx); 1735 return (SET_ERROR(ENOENT)); 1736 } 1737 1738 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg); 1739 1740 /* 1741 * Assume the on-disk data is X, the current syncing data (in 1742 * txg - 1) is Y, and the current in-memory data is Z (currently 1743 * in dmu_sync). 1744 * 1745 * We usually want to perform a nopwrite if X and Z are the 1746 * same. However, if Y is different (i.e. the BP is going to 1747 * change before this write takes effect), then a nopwrite will 1748 * be incorrect - we would override with X, which could have 1749 * been freed when Y was written. 1750 * 1751 * (Note that this is not a concern when we are nop-writing from 1752 * syncing context, because X and Y must be identical, because 1753 * all previous txgs have been synced.) 1754 * 1755 * Therefore, we disable nopwrite if the current BP could change 1756 * before this TXG. There are two ways it could change: by 1757 * being dirty (dr_next is non-NULL), or by being freed 1758 * (dnode_block_freed()). This behavior is verified by 1759 * zio_done(), which VERIFYs that the override BP is identical 1760 * to the on-disk BP. 1761 */ 1762 DB_DNODE_ENTER(db); 1763 dn = DB_DNODE(db); 1764 if (dr->dr_next != NULL || dnode_block_freed(dn, db->db_blkid)) 1765 zp.zp_nopwrite = B_FALSE; 1766 DB_DNODE_EXIT(db); 1767 1768 ASSERT(dr->dr_txg == txg); 1769 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC || 1770 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { 1771 /* 1772 * We have already issued a sync write for this buffer, 1773 * or this buffer has already been synced. It could not 1774 * have been dirtied since, or we would have cleared the state. 1775 */ 1776 mutex_exit(&db->db_mtx); 1777 return (SET_ERROR(EALREADY)); 1778 } 1779 1780 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); 1781 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC; 1782 mutex_exit(&db->db_mtx); 1783 1784 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP); 1785 dsa->dsa_dr = dr; 1786 dsa->dsa_done = done; 1787 dsa->dsa_zgd = zgd; 1788 dsa->dsa_tx = NULL; 1789 1790 zio_nowait(arc_write(pio, os->os_spa, txg, 1791 bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db), 1792 &zp, dmu_sync_ready, NULL, NULL, dmu_sync_done, dsa, 1793 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb)); 1794 1795 return (0); 1796} 1797 1798int 1799dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs, 1800 dmu_tx_t *tx) 1801{ 1802 dnode_t *dn; 1803 int err; 1804 1805 err = dnode_hold(os, object, FTAG, &dn); 1806 if (err) 1807 return (err); 1808 err = dnode_set_blksz(dn, size, ibs, tx); 1809 dnode_rele(dn, FTAG); 1810 return (err); 1811} 1812 1813void 1814dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum, 1815 dmu_tx_t *tx) 1816{ 1817 dnode_t *dn; 1818 1819 /* 1820 * Send streams include each object's checksum function. This 1821 * check ensures that the receiving system can understand the 1822 * checksum function transmitted. 1823 */ 1824 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS); 1825 1826 VERIFY0(dnode_hold(os, object, FTAG, &dn)); 1827 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS); 1828 dn->dn_checksum = checksum; 1829 dnode_setdirty(dn, tx); 1830 dnode_rele(dn, FTAG); 1831} 1832 1833void 1834dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress, 1835 dmu_tx_t *tx) 1836{ 1837 dnode_t *dn; 1838 1839 /* 1840 * Send streams include each object's compression function. This 1841 * check ensures that the receiving system can understand the 1842 * compression function transmitted. 1843 */ 1844 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS); 1845 1846 VERIFY0(dnode_hold(os, object, FTAG, &dn)); 1847 dn->dn_compress = compress; 1848 dnode_setdirty(dn, tx); 1849 dnode_rele(dn, FTAG); 1850} 1851 1852int zfs_mdcomp_disable = 0; 1853TUNABLE_INT("vfs.zfs.mdcomp_disable", &zfs_mdcomp_disable); 1854SYSCTL_INT(_vfs_zfs, OID_AUTO, mdcomp_disable, CTLFLAG_RW, 1855 &zfs_mdcomp_disable, 0, "Disable metadata compression"); 1856 1857/* 1858 * When the "redundant_metadata" property is set to "most", only indirect 1859 * blocks of this level and higher will have an additional ditto block. 1860 */ 1861int zfs_redundant_metadata_most_ditto_level = 2; 1862 1863void 1864dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp) 1865{ 1866 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET; 1867 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) || 1868 (wp & WP_SPILL)); 1869 enum zio_checksum checksum = os->os_checksum; 1870 enum zio_compress compress = os->os_compress; 1871 enum zio_checksum dedup_checksum = os->os_dedup_checksum; 1872 boolean_t dedup = B_FALSE; 1873 boolean_t nopwrite = B_FALSE; 1874 boolean_t dedup_verify = os->os_dedup_verify; 1875 int copies = os->os_copies; 1876 1877 /* 1878 * We maintain different write policies for each of the following 1879 * types of data: 1880 * 1. metadata 1881 * 2. preallocated blocks (i.e. level-0 blocks of a dump device) 1882 * 3. all other level 0 blocks 1883 */ 1884 if (ismd) { 1885 if (zfs_mdcomp_disable) { 1886 compress = ZIO_COMPRESS_EMPTY; 1887 } else { 1888 /* 1889 * XXX -- we should design a compression algorithm 1890 * that specializes in arrays of bps. 1891 */ 1892 compress = zio_compress_select(os->os_spa, 1893 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON); 1894 } 1895 1896 /* 1897 * Metadata always gets checksummed. If the data 1898 * checksum is multi-bit correctable, and it's not a 1899 * ZBT-style checksum, then it's suitable for metadata 1900 * as well. Otherwise, the metadata checksum defaults 1901 * to fletcher4. 1902 */ 1903 if (!(zio_checksum_table[checksum].ci_flags & 1904 ZCHECKSUM_FLAG_METADATA) || 1905 (zio_checksum_table[checksum].ci_flags & 1906 ZCHECKSUM_FLAG_EMBEDDED)) 1907 checksum = ZIO_CHECKSUM_FLETCHER_4; 1908 1909 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL || 1910 (os->os_redundant_metadata == 1911 ZFS_REDUNDANT_METADATA_MOST && 1912 (level >= zfs_redundant_metadata_most_ditto_level || 1913 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL)))) 1914 copies++; 1915 } else if (wp & WP_NOFILL) { 1916 ASSERT(level == 0); 1917 1918 /* 1919 * If we're writing preallocated blocks, we aren't actually 1920 * writing them so don't set any policy properties. These 1921 * blocks are currently only used by an external subsystem 1922 * outside of zfs (i.e. dump) and not written by the zio 1923 * pipeline. 1924 */ 1925 compress = ZIO_COMPRESS_OFF; 1926 checksum = ZIO_CHECKSUM_NOPARITY; 1927 } else { 1928 compress = zio_compress_select(os->os_spa, dn->dn_compress, 1929 compress); 1930 1931 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ? 1932 zio_checksum_select(dn->dn_checksum, checksum) : 1933 dedup_checksum; 1934 1935 /* 1936 * Determine dedup setting. If we are in dmu_sync(), 1937 * we won't actually dedup now because that's all 1938 * done in syncing context; but we do want to use the 1939 * dedup checkum. If the checksum is not strong 1940 * enough to ensure unique signatures, force 1941 * dedup_verify. 1942 */ 1943 if (dedup_checksum != ZIO_CHECKSUM_OFF) { 1944 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE; 1945 if (!(zio_checksum_table[checksum].ci_flags & 1946 ZCHECKSUM_FLAG_DEDUP)) 1947 dedup_verify = B_TRUE; 1948 } 1949 1950 /* 1951 * Enable nopwrite if we have secure enough checksum 1952 * algorithm (see comment in zio_nop_write) and 1953 * compression is enabled. We don't enable nopwrite if 1954 * dedup is enabled as the two features are mutually 1955 * exclusive. 1956 */ 1957 nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags & 1958 ZCHECKSUM_FLAG_NOPWRITE) && 1959 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled); 1960 } 1961 1962 zp->zp_checksum = checksum; 1963 zp->zp_compress = compress; 1964 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type; 1965 zp->zp_level = level; 1966 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa)); 1967 zp->zp_dedup = dedup; 1968 zp->zp_dedup_verify = dedup && dedup_verify; 1969 zp->zp_nopwrite = nopwrite; 1970} 1971 1972int 1973dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off) 1974{ 1975 dnode_t *dn; 1976 int err; 1977 1978 /* 1979 * Sync any current changes before 1980 * we go trundling through the block pointers. 1981 */ 1982 err = dmu_object_wait_synced(os, object); 1983 if (err) { 1984 return (err); 1985 } 1986 1987 err = dnode_hold(os, object, FTAG, &dn); 1988 if (err) { 1989 return (err); 1990 } 1991 1992 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0); 1993 dnode_rele(dn, FTAG); 1994 1995 return (err); 1996} 1997 1998/* 1999 * Given the ZFS object, if it contains any dirty nodes 2000 * this function flushes all dirty blocks to disk. This 2001 * ensures the DMU object info is updated. A more efficient 2002 * future version might just find the TXG with the maximum 2003 * ID and wait for that to be synced. 2004 */ 2005int 2006dmu_object_wait_synced(objset_t *os, uint64_t object) 2007{ 2008 dnode_t *dn; 2009 int error, i; 2010 2011 error = dnode_hold(os, object, FTAG, &dn); 2012 if (error) { 2013 return (error); 2014 } 2015 2016 for (i = 0; i < TXG_SIZE; i++) { 2017 if (list_link_active(&dn->dn_dirty_link[i])) { 2018 break; 2019 } 2020 } 2021 dnode_rele(dn, FTAG); 2022 if (i != TXG_SIZE) { 2023 txg_wait_synced(dmu_objset_pool(os), 0); 2024 } 2025 2026 return (0); 2027} 2028 2029void 2030dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi) 2031{ 2032 dnode_phys_t *dnp; 2033 2034 rw_enter(&dn->dn_struct_rwlock, RW_READER); 2035 mutex_enter(&dn->dn_mtx); 2036 2037 dnp = dn->dn_phys; 2038 2039 doi->doi_data_block_size = dn->dn_datablksz; 2040 doi->doi_metadata_block_size = dn->dn_indblkshift ? 2041 1ULL << dn->dn_indblkshift : 0; 2042 doi->doi_type = dn->dn_type; 2043 doi->doi_bonus_type = dn->dn_bonustype; 2044 doi->doi_bonus_size = dn->dn_bonuslen; 2045 doi->doi_indirection = dn->dn_nlevels; 2046 doi->doi_checksum = dn->dn_checksum; 2047 doi->doi_compress = dn->dn_compress; 2048 doi->doi_nblkptr = dn->dn_nblkptr; 2049 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9; 2050 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz; 2051 doi->doi_fill_count = 0; 2052 for (int i = 0; i < dnp->dn_nblkptr; i++) 2053 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]); 2054 2055 mutex_exit(&dn->dn_mtx); 2056 rw_exit(&dn->dn_struct_rwlock); 2057} 2058 2059/* 2060 * Get information on a DMU object. 2061 * If doi is NULL, just indicates whether the object exists. 2062 */ 2063int 2064dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi) 2065{ 2066 dnode_t *dn; 2067 int err = dnode_hold(os, object, FTAG, &dn); 2068 2069 if (err) 2070 return (err); 2071 2072 if (doi != NULL) 2073 dmu_object_info_from_dnode(dn, doi); 2074 2075 dnode_rele(dn, FTAG); 2076 return (0); 2077} 2078 2079/* 2080 * As above, but faster; can be used when you have a held dbuf in hand. 2081 */ 2082void 2083dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi) 2084{ 2085 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2086 2087 DB_DNODE_ENTER(db); 2088 dmu_object_info_from_dnode(DB_DNODE(db), doi); 2089 DB_DNODE_EXIT(db); 2090} 2091 2092/* 2093 * Faster still when you only care about the size. 2094 * This is specifically optimized for zfs_getattr(). 2095 */ 2096void 2097dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize, 2098 u_longlong_t *nblk512) 2099{ 2100 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2101 dnode_t *dn; 2102 2103 DB_DNODE_ENTER(db); 2104 dn = DB_DNODE(db); 2105 2106 *blksize = dn->dn_datablksz; 2107 /* add 1 for dnode space */ 2108 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >> 2109 SPA_MINBLOCKSHIFT) + 1; 2110 DB_DNODE_EXIT(db); 2111} 2112 2113void 2114byteswap_uint64_array(void *vbuf, size_t size) 2115{ 2116 uint64_t *buf = vbuf; 2117 size_t count = size >> 3; 2118 int i; 2119 2120 ASSERT((size & 7) == 0); 2121 2122 for (i = 0; i < count; i++) 2123 buf[i] = BSWAP_64(buf[i]); 2124} 2125 2126void 2127byteswap_uint32_array(void *vbuf, size_t size) 2128{ 2129 uint32_t *buf = vbuf; 2130 size_t count = size >> 2; 2131 int i; 2132 2133 ASSERT((size & 3) == 0); 2134 2135 for (i = 0; i < count; i++) 2136 buf[i] = BSWAP_32(buf[i]); 2137} 2138 2139void 2140byteswap_uint16_array(void *vbuf, size_t size) 2141{ 2142 uint16_t *buf = vbuf; 2143 size_t count = size >> 1; 2144 int i; 2145 2146 ASSERT((size & 1) == 0); 2147 2148 for (i = 0; i < count; i++) 2149 buf[i] = BSWAP_16(buf[i]); 2150} 2151 2152/* ARGSUSED */ 2153void 2154byteswap_uint8_array(void *vbuf, size_t size) 2155{ 2156} 2157 2158void 2159dmu_init(void) 2160{ 2161 zfs_dbgmsg_init(); 2162 sa_cache_init(); 2163 xuio_stat_init(); 2164 dmu_objset_init(); 2165 dnode_init(); 2166 zfetch_init(); 2167 zio_compress_init(); 2168 l2arc_init(); 2169 arc_init(); 2170 dbuf_init(); 2171} 2172 2173void 2174dmu_fini(void) 2175{ 2176 arc_fini(); /* arc depends on l2arc, so arc must go first */ 2177 l2arc_fini(); 2178 zfetch_fini(); 2179 zio_compress_fini(); 2180 dbuf_fini(); 2181 dnode_fini(); 2182 dmu_objset_fini(); 2183 xuio_stat_fini(); 2184 sa_cache_fini(); 2185 zfs_dbgmsg_fini(); 2186} 2187