Now available: The Design and Implementation of the FreeBSD Operating System (Second Edition)
[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]
FreeBSD/Linux Kernel Cross Reference
sys/contrib/openzfs/module/zfs/dmu.c
Version:
- FREEBSD - FREEBSD-13-STABLE - FREEBSD-13-0 - FREEBSD-12-STABLE - FREEBSD-12-0 - FREEBSD-11-STABLE - FREEBSD-11-0 - FREEBSD-10-STABLE - FREEBSD-10-0 - FREEBSD-9-STABLE - FREEBSD-9-0 - FREEBSD-8-STABLE - FREEBSD-8-0 - FREEBSD-7-STABLE - FREEBSD-7-0 - FREEBSD-6-STABLE - FREEBSD-6-0 - FREEBSD-5-STABLE - FREEBSD-5-0 - FREEBSD-4-STABLE - FREEBSD-3-STABLE - FREEBSD22 - l41 - OPENBSD - linux-2.6 - MK84 - PLAN9 - xnu-8792
SearchContext: - none - 3 - 10
SearchContext: - none - 3 - 10
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 https://opensource.org/licenses/CDDL-1.0. 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, 2020 by Delphix. All rights reserved. 24 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved. 25 * Copyright (c) 2013, Joyent, Inc. All rights reserved. 26 * Copyright (c) 2016, Nexenta Systems, Inc. All rights reserved. 27 * Copyright (c) 2015 by Chunwei Chen. All rights reserved. 28 * Copyright (c) 2019 Datto Inc. 29 * Copyright (c) 2019, Klara Inc. 30 * Copyright (c) 2019, Allan Jude 31 * Copyright (c) 2022 Hewlett Packard Enterprise Development LP. 32 */ 33 34 #include <sys/dmu.h> 35 #include <sys/dmu_impl.h> 36 #include <sys/dmu_tx.h> 37 #include <sys/dbuf.h> 38 #include <sys/dnode.h> 39 #include <sys/zfs_context.h> 40 #include <sys/dmu_objset.h> 41 #include <sys/dmu_traverse.h> 42 #include <sys/dsl_dataset.h> 43 #include <sys/dsl_dir.h> 44 #include <sys/dsl_pool.h> 45 #include <sys/dsl_synctask.h> 46 #include <sys/dsl_prop.h> 47 #include <sys/dmu_zfetch.h> 48 #include <sys/zfs_ioctl.h> 49 #include <sys/zap.h> 50 #include <sys/zio_checksum.h> 51 #include <sys/zio_compress.h> 52 #include <sys/sa.h> 53 #include <sys/zfeature.h> 54 #include <sys/abd.h> 55 #include <sys/trace_zfs.h> 56 #include <sys/zfs_racct.h> 57 #include <sys/zfs_rlock.h> 58 #ifdef _KERNEL 59 #include <sys/vmsystm.h> 60 #include <sys/zfs_znode.h> 61 #endif 62 63 /* 64 * Enable/disable nopwrite feature. 65 */ 66 static int zfs_nopwrite_enabled = 1; 67 68 /* 69 * Tunable to control percentage of dirtied L1 blocks from frees allowed into 70 * one TXG. After this threshold is crossed, additional dirty blocks from frees 71 * will wait until the next TXG. 72 * A value of zero will disable this throttle. 73 */ 74 static uint_t zfs_per_txg_dirty_frees_percent = 30; 75 76 /* 77 * Enable/disable forcing txg sync when dirty checking for holes with lseek(). 78 * By default this is enabled to ensure accurate hole reporting, it can result 79 * in a significant performance penalty for lseek(SEEK_HOLE) heavy workloads. 80 * Disabling this option will result in holes never being reported in dirty 81 * files which is always safe. 82 */ 83 static int zfs_dmu_offset_next_sync = 1; 84 85 /* 86 * Limit the amount we can prefetch with one call to this amount. This 87 * helps to limit the amount of memory that can be used by prefetching. 88 * Larger objects should be prefetched a bit at a time. 89 */ 90 uint_t dmu_prefetch_max = 8 * SPA_MAXBLOCKSIZE; 91 92 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = { 93 {DMU_BSWAP_UINT8, TRUE, FALSE, FALSE, "unallocated" }, 94 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "object directory" }, 95 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "object array" }, 96 {DMU_BSWAP_UINT8, TRUE, FALSE, FALSE, "packed nvlist" }, 97 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "packed nvlist size" }, 98 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "bpobj" }, 99 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "bpobj header" }, 100 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "SPA space map header" }, 101 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "SPA space map" }, 102 {DMU_BSWAP_UINT64, TRUE, FALSE, TRUE, "ZIL intent log" }, 103 {DMU_BSWAP_DNODE, TRUE, FALSE, TRUE, "DMU dnode" }, 104 {DMU_BSWAP_OBJSET, TRUE, TRUE, FALSE, "DMU objset" }, 105 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "DSL directory" }, 106 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL directory child map"}, 107 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL dataset snap map" }, 108 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL props" }, 109 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "DSL dataset" }, 110 {DMU_BSWAP_ZNODE, TRUE, FALSE, FALSE, "ZFS znode" }, 111 {DMU_BSWAP_OLDACL, TRUE, FALSE, TRUE, "ZFS V0 ACL" }, 112 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "ZFS plain file" }, 113 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS directory" }, 114 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "ZFS master node" }, 115 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS delete queue" }, 116 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "zvol object" }, 117 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "zvol prop" }, 118 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "other uint8[]" }, 119 {DMU_BSWAP_UINT64, FALSE, FALSE, TRUE, "other uint64[]" }, 120 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "other ZAP" }, 121 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "persistent error log" }, 122 {DMU_BSWAP_UINT8, TRUE, FALSE, FALSE, "SPA history" }, 123 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "SPA history offsets" }, 124 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "Pool properties" }, 125 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL permissions" }, 126 {DMU_BSWAP_ACL, TRUE, FALSE, TRUE, "ZFS ACL" }, 127 {DMU_BSWAP_UINT8, TRUE, FALSE, TRUE, "ZFS SYSACL" }, 128 {DMU_BSWAP_UINT8, TRUE, FALSE, TRUE, "FUID table" }, 129 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "FUID table size" }, 130 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL dataset next clones"}, 131 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "scan work queue" }, 132 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS user/group/project used" }, 133 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS user/group/project quota"}, 134 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "snapshot refcount tags"}, 135 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "DDT ZAP algorithm" }, 136 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "DDT statistics" }, 137 {DMU_BSWAP_UINT8, TRUE, FALSE, TRUE, "System attributes" }, 138 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "SA master node" }, 139 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "SA attr registration" }, 140 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "SA attr layouts" }, 141 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "scan translations" }, 142 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "deduplicated block" }, 143 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL deadlist map" }, 144 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "DSL deadlist map hdr" }, 145 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL dir clones" }, 146 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "bpobj subobj" } 147 }; 148 149 dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = { 150 { byteswap_uint8_array, "uint8" }, 151 { byteswap_uint16_array, "uint16" }, 152 { byteswap_uint32_array, "uint32" }, 153 { byteswap_uint64_array, "uint64" }, 154 { zap_byteswap, "zap" }, 155 { dnode_buf_byteswap, "dnode" }, 156 { dmu_objset_byteswap, "objset" }, 157 { zfs_znode_byteswap, "znode" }, 158 { zfs_oldacl_byteswap, "oldacl" }, 159 { zfs_acl_byteswap, "acl" } 160 }; 161 162 static int 163 dmu_buf_hold_noread_by_dnode(dnode_t *dn, uint64_t offset, 164 const void *tag, dmu_buf_t **dbp) 165 { 166 uint64_t blkid; 167 dmu_buf_impl_t *db; 168 169 rw_enter(&dn->dn_struct_rwlock, RW_READER); 170 blkid = dbuf_whichblock(dn, 0, offset); 171 db = dbuf_hold(dn, blkid, tag); 172 rw_exit(&dn->dn_struct_rwlock); 173 174 if (db == NULL) { 175 *dbp = NULL; 176 return (SET_ERROR(EIO)); 177 } 178 179 *dbp = &db->db; 180 return (0); 181 } 182 int 183 dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset, 184 const void *tag, dmu_buf_t **dbp) 185 { 186 dnode_t *dn; 187 uint64_t blkid; 188 dmu_buf_impl_t *db; 189 int err; 190 191 err = dnode_hold(os, object, FTAG, &dn); 192 if (err) 193 return (err); 194 rw_enter(&dn->dn_struct_rwlock, RW_READER); 195 blkid = dbuf_whichblock(dn, 0, offset); 196 db = dbuf_hold(dn, blkid, tag); 197 rw_exit(&dn->dn_struct_rwlock); 198 dnode_rele(dn, FTAG); 199 200 if (db == NULL) { 201 *dbp = NULL; 202 return (SET_ERROR(EIO)); 203 } 204 205 *dbp = &db->db; 206 return (err); 207 } 208 209 int 210 dmu_buf_hold_by_dnode(dnode_t *dn, uint64_t offset, 211 const void *tag, dmu_buf_t **dbp, int flags) 212 { 213 int err; 214 int db_flags = DB_RF_CANFAIL; 215 216 if (flags & DMU_READ_NO_PREFETCH) 217 db_flags |= DB_RF_NOPREFETCH; 218 if (flags & DMU_READ_NO_DECRYPT) 219 db_flags |= DB_RF_NO_DECRYPT; 220 221 err = dmu_buf_hold_noread_by_dnode(dn, offset, tag, dbp); 222 if (err == 0) { 223 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp); 224 err = dbuf_read(db, NULL, db_flags); 225 if (err != 0) { 226 dbuf_rele(db, tag); 227 *dbp = NULL; 228 } 229 } 230 231 return (err); 232 } 233 234 int 235 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset, 236 const void *tag, dmu_buf_t **dbp, int flags) 237 { 238 int err; 239 int db_flags = DB_RF_CANFAIL; 240 241 if (flags & DMU_READ_NO_PREFETCH) 242 db_flags |= DB_RF_NOPREFETCH; 243 if (flags & DMU_READ_NO_DECRYPT) 244 db_flags |= DB_RF_NO_DECRYPT; 245 246 err = dmu_buf_hold_noread(os, object, offset, tag, dbp); 247 if (err == 0) { 248 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp); 249 err = dbuf_read(db, NULL, db_flags); 250 if (err != 0) { 251 dbuf_rele(db, tag); 252 *dbp = NULL; 253 } 254 } 255 256 return (err); 257 } 258 259 int 260 dmu_bonus_max(void) 261 { 262 return (DN_OLD_MAX_BONUSLEN); 263 } 264 265 int 266 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx) 267 { 268 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 269 dnode_t *dn; 270 int error; 271 272 DB_DNODE_ENTER(db); 273 dn = DB_DNODE(db); 274 275 if (dn->dn_bonus != db) { 276 error = SET_ERROR(EINVAL); 277 } else if (newsize < 0 || newsize > db_fake->db_size) { 278 error = SET_ERROR(EINVAL); 279 } else { 280 dnode_setbonuslen(dn, newsize, tx); 281 error = 0; 282 } 283 284 DB_DNODE_EXIT(db); 285 return (error); 286 } 287 288 int 289 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx) 290 { 291 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 292 dnode_t *dn; 293 int error; 294 295 DB_DNODE_ENTER(db); 296 dn = DB_DNODE(db); 297 298 if (!DMU_OT_IS_VALID(type)) { 299 error = SET_ERROR(EINVAL); 300 } else if (dn->dn_bonus != db) { 301 error = SET_ERROR(EINVAL); 302 } else { 303 dnode_setbonus_type(dn, type, tx); 304 error = 0; 305 } 306 307 DB_DNODE_EXIT(db); 308 return (error); 309 } 310 311 dmu_object_type_t 312 dmu_get_bonustype(dmu_buf_t *db_fake) 313 { 314 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 315 dnode_t *dn; 316 dmu_object_type_t type; 317 318 DB_DNODE_ENTER(db); 319 dn = DB_DNODE(db); 320 type = dn->dn_bonustype; 321 DB_DNODE_EXIT(db); 322 323 return (type); 324 } 325 326 int 327 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx) 328 { 329 dnode_t *dn; 330 int error; 331 332 error = dnode_hold(os, object, FTAG, &dn); 333 dbuf_rm_spill(dn, tx); 334 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 335 dnode_rm_spill(dn, tx); 336 rw_exit(&dn->dn_struct_rwlock); 337 dnode_rele(dn, FTAG); 338 return (error); 339 } 340 341 /* 342 * Lookup and hold the bonus buffer for the provided dnode. If the dnode 343 * has not yet been allocated a new bonus dbuf a will be allocated. 344 * Returns ENOENT, EIO, or 0. 345 */ 346 int dmu_bonus_hold_by_dnode(dnode_t *dn, const void *tag, dmu_buf_t **dbp, 347 uint32_t flags) 348 { 349 dmu_buf_impl_t *db; 350 int error; 351 uint32_t db_flags = DB_RF_MUST_SUCCEED; 352 353 if (flags & DMU_READ_NO_PREFETCH) 354 db_flags |= DB_RF_NOPREFETCH; 355 if (flags & DMU_READ_NO_DECRYPT) 356 db_flags |= DB_RF_NO_DECRYPT; 357 358 rw_enter(&dn->dn_struct_rwlock, RW_READER); 359 if (dn->dn_bonus == NULL) { 360 rw_exit(&dn->dn_struct_rwlock); 361 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 362 if (dn->dn_bonus == NULL) 363 dbuf_create_bonus(dn); 364 } 365 db = dn->dn_bonus; 366 367 /* as long as the bonus buf is held, the dnode will be held */ 368 if (zfs_refcount_add(&db->db_holds, tag) == 1) { 369 VERIFY(dnode_add_ref(dn, db)); 370 atomic_inc_32(&dn->dn_dbufs_count); 371 } 372 373 /* 374 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's 375 * hold and incrementing the dbuf count to ensure that dnode_move() sees 376 * a dnode hold for every dbuf. 377 */ 378 rw_exit(&dn->dn_struct_rwlock); 379 380 error = dbuf_read(db, NULL, db_flags); 381 if (error) { 382 dnode_evict_bonus(dn); 383 dbuf_rele(db, tag); 384 *dbp = NULL; 385 return (error); 386 } 387 388 *dbp = &db->db; 389 return (0); 390 } 391 392 int 393 dmu_bonus_hold(objset_t *os, uint64_t object, const void *tag, dmu_buf_t **dbp) 394 { 395 dnode_t *dn; 396 int error; 397 398 error = dnode_hold(os, object, FTAG, &dn); 399 if (error) 400 return (error); 401 402 error = dmu_bonus_hold_by_dnode(dn, tag, dbp, DMU_READ_NO_PREFETCH); 403 dnode_rele(dn, FTAG); 404 405 return (error); 406 } 407 408 /* 409 * returns ENOENT, EIO, or 0. 410 * 411 * This interface will allocate a blank spill dbuf when a spill blk 412 * doesn't already exist on the dnode. 413 * 414 * if you only want to find an already existing spill db, then 415 * dmu_spill_hold_existing() should be used. 416 */ 417 int 418 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, const void *tag, 419 dmu_buf_t **dbp) 420 { 421 dmu_buf_impl_t *db = NULL; 422 int err; 423 424 if ((flags & DB_RF_HAVESTRUCT) == 0) 425 rw_enter(&dn->dn_struct_rwlock, RW_READER); 426 427 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag); 428 429 if ((flags & DB_RF_HAVESTRUCT) == 0) 430 rw_exit(&dn->dn_struct_rwlock); 431 432 if (db == NULL) { 433 *dbp = NULL; 434 return (SET_ERROR(EIO)); 435 } 436 err = dbuf_read(db, NULL, flags); 437 if (err == 0) 438 *dbp = &db->db; 439 else { 440 dbuf_rele(db, tag); 441 *dbp = NULL; 442 } 443 return (err); 444 } 445 446 int 447 dmu_spill_hold_existing(dmu_buf_t *bonus, const void *tag, dmu_buf_t **dbp) 448 { 449 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus; 450 dnode_t *dn; 451 int err; 452 453 DB_DNODE_ENTER(db); 454 dn = DB_DNODE(db); 455 456 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) { 457 err = SET_ERROR(EINVAL); 458 } else { 459 rw_enter(&dn->dn_struct_rwlock, RW_READER); 460 461 if (!dn->dn_have_spill) { 462 err = SET_ERROR(ENOENT); 463 } else { 464 err = dmu_spill_hold_by_dnode(dn, 465 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp); 466 } 467 468 rw_exit(&dn->dn_struct_rwlock); 469 } 470 471 DB_DNODE_EXIT(db); 472 return (err); 473 } 474 475 int 476 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, uint32_t flags, const void *tag, 477 dmu_buf_t **dbp) 478 { 479 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus; 480 dnode_t *dn; 481 int err; 482 uint32_t db_flags = DB_RF_CANFAIL; 483 484 if (flags & DMU_READ_NO_DECRYPT) 485 db_flags |= DB_RF_NO_DECRYPT; 486 487 DB_DNODE_ENTER(db); 488 dn = DB_DNODE(db); 489 err = dmu_spill_hold_by_dnode(dn, db_flags, tag, dbp); 490 DB_DNODE_EXIT(db); 491 492 return (err); 493 } 494 495 /* 496 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces 497 * to take a held dnode rather than <os, object> -- the lookup is wasteful, 498 * and can induce severe lock contention when writing to several files 499 * whose dnodes are in the same block. 500 */ 501 int 502 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length, 503 boolean_t read, const void *tag, int *numbufsp, dmu_buf_t ***dbpp, 504 uint32_t flags) 505 { 506 dmu_buf_t **dbp; 507 zstream_t *zs = NULL; 508 uint64_t blkid, nblks, i; 509 uint32_t dbuf_flags; 510 int err; 511 zio_t *zio = NULL; 512 boolean_t missed = B_FALSE; 513 514 ASSERT(length <= DMU_MAX_ACCESS); 515 516 /* 517 * Note: We directly notify the prefetch code of this read, so that 518 * we can tell it about the multi-block read. dbuf_read() only knows 519 * about the one block it is accessing. 520 */ 521 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT | 522 DB_RF_NOPREFETCH; 523 524 if ((flags & DMU_READ_NO_DECRYPT) != 0) 525 dbuf_flags |= DB_RF_NO_DECRYPT; 526 527 rw_enter(&dn->dn_struct_rwlock, RW_READER); 528 if (dn->dn_datablkshift) { 529 int blkshift = dn->dn_datablkshift; 530 nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) - 531 P2ALIGN(offset, 1ULL << blkshift)) >> blkshift; 532 } else { 533 if (offset + length > dn->dn_datablksz) { 534 zfs_panic_recover("zfs: accessing past end of object " 535 "%llx/%llx (size=%u access=%llu+%llu)", 536 (longlong_t)dn->dn_objset-> 537 os_dsl_dataset->ds_object, 538 (longlong_t)dn->dn_object, dn->dn_datablksz, 539 (longlong_t)offset, (longlong_t)length); 540 rw_exit(&dn->dn_struct_rwlock); 541 return (SET_ERROR(EIO)); 542 } 543 nblks = 1; 544 } 545 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP); 546 547 if (read) 548 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, 549 ZIO_FLAG_CANFAIL); 550 blkid = dbuf_whichblock(dn, 0, offset); 551 if ((flags & DMU_READ_NO_PREFETCH) == 0 && 552 length <= zfetch_array_rd_sz) { 553 /* 554 * Prepare the zfetch before initiating the demand reads, so 555 * that if multiple threads block on same indirect block, we 556 * base predictions on the original less racy request order. 557 */ 558 zs = dmu_zfetch_prepare(&dn->dn_zfetch, blkid, nblks, read, 559 B_TRUE); 560 } 561 for (i = 0; i < nblks; i++) { 562 dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag); 563 if (db == NULL) { 564 if (zs) 565 dmu_zfetch_run(zs, missed, B_TRUE); 566 rw_exit(&dn->dn_struct_rwlock); 567 dmu_buf_rele_array(dbp, nblks, tag); 568 if (read) 569 zio_nowait(zio); 570 return (SET_ERROR(EIO)); 571 } 572 573 /* 574 * Initiate async demand data read. 575 * We check the db_state after calling dbuf_read() because 576 * (1) dbuf_read() may change the state to CACHED due to a 577 * hit in the ARC, and (2) on a cache miss, a child will 578 * have been added to "zio" but not yet completed, so the 579 * state will not yet be CACHED. 580 */ 581 if (read) { 582 if (i == nblks - 1 && blkid + i < dn->dn_maxblkid && 583 offset + length < db->db.db_offset + 584 db->db.db_size) { 585 if (offset <= db->db.db_offset) 586 dbuf_flags |= DB_RF_PARTIAL_FIRST; 587 else 588 dbuf_flags |= DB_RF_PARTIAL_MORE; 589 } 590 (void) dbuf_read(db, zio, dbuf_flags); 591 if (db->db_state != DB_CACHED) 592 missed = B_TRUE; 593 } 594 dbp[i] = &db->db; 595 } 596 597 if (!read) 598 zfs_racct_write(length, nblks); 599 600 if (zs) 601 dmu_zfetch_run(zs, missed, B_TRUE); 602 rw_exit(&dn->dn_struct_rwlock); 603 604 if (read) { 605 /* wait for async read i/o */ 606 err = zio_wait(zio); 607 if (err) { 608 dmu_buf_rele_array(dbp, nblks, tag); 609 return (err); 610 } 611 612 /* wait for other io to complete */ 613 for (i = 0; i < nblks; i++) { 614 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i]; 615 mutex_enter(&db->db_mtx); 616 while (db->db_state == DB_READ || 617 db->db_state == DB_FILL) 618 cv_wait(&db->db_changed, &db->db_mtx); 619 if (db->db_state == DB_UNCACHED) 620 err = SET_ERROR(EIO); 621 mutex_exit(&db->db_mtx); 622 if (err) { 623 dmu_buf_rele_array(dbp, nblks, tag); 624 return (err); 625 } 626 } 627 } 628 629 *numbufsp = nblks; 630 *dbpp = dbp; 631 return (0); 632 } 633 634 int 635 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset, 636 uint64_t length, int read, const void *tag, int *numbufsp, 637 dmu_buf_t ***dbpp) 638 { 639 dnode_t *dn; 640 int err; 641 642 err = dnode_hold(os, object, FTAG, &dn); 643 if (err) 644 return (err); 645 646 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag, 647 numbufsp, dbpp, DMU_READ_PREFETCH); 648 649 dnode_rele(dn, FTAG); 650 651 return (err); 652 } 653 654 int 655 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset, 656 uint64_t length, boolean_t read, const void *tag, int *numbufsp, 657 dmu_buf_t ***dbpp) 658 { 659 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 660 dnode_t *dn; 661 int err; 662 663 DB_DNODE_ENTER(db); 664 dn = DB_DNODE(db); 665 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag, 666 numbufsp, dbpp, DMU_READ_PREFETCH); 667 DB_DNODE_EXIT(db); 668 669 return (err); 670 } 671 672 void 673 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, const void *tag) 674 { 675 int i; 676 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake; 677 678 if (numbufs == 0) 679 return; 680 681 for (i = 0; i < numbufs; i++) { 682 if (dbp[i]) 683 dbuf_rele(dbp[i], tag); 684 } 685 686 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs); 687 } 688 689 /* 690 * Issue prefetch i/os for the given blocks. If level is greater than 0, the 691 * indirect blocks prefetched will be those that point to the blocks containing 692 * the data starting at offset, and continuing to offset + len. 693 * 694 * Note that if the indirect blocks above the blocks being prefetched are not 695 * in cache, they will be asynchronously read in. 696 */ 697 void 698 dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset, 699 uint64_t len, zio_priority_t pri) 700 { 701 dnode_t *dn; 702 uint64_t blkid; 703 int nblks, err; 704 705 if (len == 0) { /* they're interested in the bonus buffer */ 706 dn = DMU_META_DNODE(os); 707 708 if (object == 0 || object >= DN_MAX_OBJECT) 709 return; 710 711 rw_enter(&dn->dn_struct_rwlock, RW_READER); 712 blkid = dbuf_whichblock(dn, level, 713 object * sizeof (dnode_phys_t)); 714 dbuf_prefetch(dn, level, blkid, pri, 0); 715 rw_exit(&dn->dn_struct_rwlock); 716 return; 717 } 718 719 /* 720 * See comment before the definition of dmu_prefetch_max. 721 */ 722 len = MIN(len, dmu_prefetch_max); 723 724 /* 725 * XXX - Note, if the dnode for the requested object is not 726 * already cached, we will do a *synchronous* read in the 727 * dnode_hold() call. The same is true for any indirects. 728 */ 729 err = dnode_hold(os, object, FTAG, &dn); 730 if (err != 0) 731 return; 732 733 /* 734 * offset + len - 1 is the last byte we want to prefetch for, and offset 735 * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the 736 * last block we want to prefetch, and dbuf_whichblock(dn, level, 737 * offset) is the first. Then the number we need to prefetch is the 738 * last - first + 1. 739 */ 740 rw_enter(&dn->dn_struct_rwlock, RW_READER); 741 if (level > 0 || dn->dn_datablkshift != 0) { 742 nblks = dbuf_whichblock(dn, level, offset + len - 1) - 743 dbuf_whichblock(dn, level, offset) + 1; 744 } else { 745 nblks = (offset < dn->dn_datablksz); 746 } 747 748 if (nblks != 0) { 749 blkid = dbuf_whichblock(dn, level, offset); 750 for (int i = 0; i < nblks; i++) 751 dbuf_prefetch(dn, level, blkid + i, pri, 0); 752 } 753 rw_exit(&dn->dn_struct_rwlock); 754 755 dnode_rele(dn, FTAG); 756 } 757 758 /* 759 * Get the next "chunk" of file data to free. We traverse the file from 760 * the end so that the file gets shorter over time (if we crashes in the 761 * middle, this will leave us in a better state). We find allocated file 762 * data by simply searching the allocated level 1 indirects. 763 * 764 * On input, *start should be the first offset that does not need to be 765 * freed (e.g. "offset + length"). On return, *start will be the first 766 * offset that should be freed and l1blks is set to the number of level 1 767 * indirect blocks found within the chunk. 768 */ 769 static int 770 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum, uint64_t *l1blks) 771 { 772 uint64_t blks; 773 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1); 774 /* bytes of data covered by a level-1 indirect block */ 775 uint64_t iblkrange = (uint64_t)dn->dn_datablksz * 776 EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT); 777 778 ASSERT3U(minimum, <=, *start); 779 780 /* 781 * Check if we can free the entire range assuming that all of the 782 * L1 blocks in this range have data. If we can, we use this 783 * worst case value as an estimate so we can avoid having to look 784 * at the object's actual data. 785 */ 786 uint64_t total_l1blks = 787 (roundup(*start, iblkrange) - (minimum / iblkrange * iblkrange)) / 788 iblkrange; 789 if (total_l1blks <= maxblks) { 790 *l1blks = total_l1blks; 791 *start = minimum; 792 return (0); 793 } 794 ASSERT(ISP2(iblkrange)); 795 796 for (blks = 0; *start > minimum && blks < maxblks; blks++) { 797 int err; 798 799 /* 800 * dnode_next_offset(BACKWARDS) will find an allocated L1 801 * indirect block at or before the input offset. We must 802 * decrement *start so that it is at the end of the region 803 * to search. 804 */ 805 (*start)--; 806 807 err = dnode_next_offset(dn, 808 DNODE_FIND_BACKWARDS, start, 2, 1, 0); 809 810 /* if there are no indirect blocks before start, we are done */ 811 if (err == ESRCH) { 812 *start = minimum; 813 break; 814 } else if (err != 0) { 815 *l1blks = blks; 816 return (err); 817 } 818 819 /* set start to the beginning of this L1 indirect */ 820 *start = P2ALIGN(*start, iblkrange); 821 } 822 if (*start < minimum) 823 *start = minimum; 824 *l1blks = blks; 825 826 return (0); 827 } 828 829 /* 830 * If this objset is of type OST_ZFS return true if vfs's unmounted flag is set, 831 * otherwise return false. 832 * Used below in dmu_free_long_range_impl() to enable abort when unmounting 833 */ 834 static boolean_t 835 dmu_objset_zfs_unmounting(objset_t *os) 836 { 837 #ifdef _KERNEL 838 if (dmu_objset_type(os) == DMU_OST_ZFS) 839 return (zfs_get_vfs_flag_unmounted(os)); 840 #else 841 (void) os; 842 #endif 843 return (B_FALSE); 844 } 845 846 static int 847 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset, 848 uint64_t length) 849 { 850 uint64_t object_size; 851 int err; 852 uint64_t dirty_frees_threshold; 853 dsl_pool_t *dp = dmu_objset_pool(os); 854 855 if (dn == NULL) 856 return (SET_ERROR(EINVAL)); 857 858 object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz; 859 if (offset >= object_size) 860 return (0); 861 862 if (zfs_per_txg_dirty_frees_percent <= 100) 863 dirty_frees_threshold = 864 zfs_per_txg_dirty_frees_percent * zfs_dirty_data_max / 100; 865 else 866 dirty_frees_threshold = zfs_dirty_data_max / 20; 867 868 if (length == DMU_OBJECT_END || offset + length > object_size) 869 length = object_size - offset; 870 871 while (length != 0) { 872 uint64_t chunk_end, chunk_begin, chunk_len; 873 uint64_t l1blks; 874 dmu_tx_t *tx; 875 876 if (dmu_objset_zfs_unmounting(dn->dn_objset)) 877 return (SET_ERROR(EINTR)); 878 879 chunk_end = chunk_begin = offset + length; 880 881 /* move chunk_begin backwards to the beginning of this chunk */ 882 err = get_next_chunk(dn, &chunk_begin, offset, &l1blks); 883 if (err) 884 return (err); 885 ASSERT3U(chunk_begin, >=, offset); 886 ASSERT3U(chunk_begin, <=, chunk_end); 887 888 chunk_len = chunk_end - chunk_begin; 889 890 tx = dmu_tx_create(os); 891 dmu_tx_hold_free(tx, dn->dn_object, chunk_begin, chunk_len); 892 893 /* 894 * Mark this transaction as typically resulting in a net 895 * reduction in space used. 896 */ 897 dmu_tx_mark_netfree(tx); 898 err = dmu_tx_assign(tx, TXG_WAIT); 899 if (err) { 900 dmu_tx_abort(tx); 901 return (err); 902 } 903 904 uint64_t txg = dmu_tx_get_txg(tx); 905 906 mutex_enter(&dp->dp_lock); 907 uint64_t long_free_dirty = 908 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK]; 909 mutex_exit(&dp->dp_lock); 910 911 /* 912 * To avoid filling up a TXG with just frees, wait for 913 * the next TXG to open before freeing more chunks if 914 * we have reached the threshold of frees. 915 */ 916 if (dirty_frees_threshold != 0 && 917 long_free_dirty >= dirty_frees_threshold) { 918 DMU_TX_STAT_BUMP(dmu_tx_dirty_frees_delay); 919 dmu_tx_commit(tx); 920 txg_wait_open(dp, 0, B_TRUE); 921 continue; 922 } 923 924 /* 925 * In order to prevent unnecessary write throttling, for each 926 * TXG, we track the cumulative size of L1 blocks being dirtied 927 * in dnode_free_range() below. We compare this number to a 928 * tunable threshold, past which we prevent new L1 dirty freeing 929 * blocks from being added into the open TXG. See 930 * dmu_free_long_range_impl() for details. The threshold 931 * prevents write throttle activation due to dirty freeing L1 932 * blocks taking up a large percentage of zfs_dirty_data_max. 933 */ 934 mutex_enter(&dp->dp_lock); 935 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] += 936 l1blks << dn->dn_indblkshift; 937 mutex_exit(&dp->dp_lock); 938 DTRACE_PROBE3(free__long__range, 939 uint64_t, long_free_dirty, uint64_t, chunk_len, 940 uint64_t, txg); 941 dnode_free_range(dn, chunk_begin, chunk_len, tx); 942 943 dmu_tx_commit(tx); 944 945 length -= chunk_len; 946 } 947 return (0); 948 } 949 950 int 951 dmu_free_long_range(objset_t *os, uint64_t object, 952 uint64_t offset, uint64_t length) 953 { 954 dnode_t *dn; 955 int err; 956 957 err = dnode_hold(os, object, FTAG, &dn); 958 if (err != 0) 959 return (err); 960 err = dmu_free_long_range_impl(os, dn, offset, length); 961 962 /* 963 * It is important to zero out the maxblkid when freeing the entire 964 * file, so that (a) subsequent calls to dmu_free_long_range_impl() 965 * will take the fast path, and (b) dnode_reallocate() can verify 966 * that the entire file has been freed. 967 */ 968 if (err == 0 && offset == 0 && length == DMU_OBJECT_END) 969 dn->dn_maxblkid = 0; 970 971 dnode_rele(dn, FTAG); 972 return (err); 973 } 974 975 int 976 dmu_free_long_object(objset_t *os, uint64_t object) 977 { 978 dmu_tx_t *tx; 979 int err; 980 981 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END); 982 if (err != 0) 983 return (err); 984 985 tx = dmu_tx_create(os); 986 dmu_tx_hold_bonus(tx, object); 987 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END); 988 dmu_tx_mark_netfree(tx); 989 err = dmu_tx_assign(tx, TXG_WAIT); 990 if (err == 0) { 991 err = dmu_object_free(os, object, tx); 992 dmu_tx_commit(tx); 993 } else { 994 dmu_tx_abort(tx); 995 } 996 997 return (err); 998 } 999 1000 int 1001 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset, 1002 uint64_t size, dmu_tx_t *tx) 1003 { 1004 dnode_t *dn; 1005 int err = dnode_hold(os, object, FTAG, &dn); 1006 if (err) 1007 return (err); 1008 ASSERT(offset < UINT64_MAX); 1009 ASSERT(size == DMU_OBJECT_END || size <= UINT64_MAX - offset); 1010 dnode_free_range(dn, offset, size, tx); 1011 dnode_rele(dn, FTAG); 1012 return (0); 1013 } 1014 1015 static int 1016 dmu_read_impl(dnode_t *dn, uint64_t offset, uint64_t size, 1017 void *buf, uint32_t flags) 1018 { 1019 dmu_buf_t **dbp; 1020 int numbufs, err = 0; 1021 1022 /* 1023 * Deal with odd block sizes, where there can't be data past the first 1024 * block. If we ever do the tail block optimization, we will need to 1025 * handle that here as well. 1026 */ 1027 if (dn->dn_maxblkid == 0) { 1028 uint64_t newsz = offset > dn->dn_datablksz ? 0 : 1029 MIN(size, dn->dn_datablksz - offset); 1030 memset((char *)buf + newsz, 0, size - newsz); 1031 size = newsz; 1032 } 1033 1034 while (size > 0) { 1035 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2); 1036 int i; 1037 1038 /* 1039 * NB: we could do this block-at-a-time, but it's nice 1040 * to be reading in parallel. 1041 */ 1042 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen, 1043 TRUE, FTAG, &numbufs, &dbp, flags); 1044 if (err) 1045 break; 1046 1047 for (i = 0; i < numbufs; i++) { 1048 uint64_t tocpy; 1049 int64_t bufoff; 1050 dmu_buf_t *db = dbp[i]; 1051 1052 ASSERT(size > 0); 1053 1054 bufoff = offset - db->db_offset; 1055 tocpy = MIN(db->db_size - bufoff, size); 1056 1057 (void) memcpy(buf, (char *)db->db_data + bufoff, tocpy); 1058 1059 offset += tocpy; 1060 size -= tocpy; 1061 buf = (char *)buf + tocpy; 1062 } 1063 dmu_buf_rele_array(dbp, numbufs, FTAG); 1064 } 1065 return (err); 1066 } 1067 1068 int 1069 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1070 void *buf, uint32_t flags) 1071 { 1072 dnode_t *dn; 1073 int err; 1074 1075 err = dnode_hold(os, object, FTAG, &dn); 1076 if (err != 0) 1077 return (err); 1078 1079 err = dmu_read_impl(dn, offset, size, buf, flags); 1080 dnode_rele(dn, FTAG); 1081 return (err); 1082 } 1083 1084 int 1085 dmu_read_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, void *buf, 1086 uint32_t flags) 1087 { 1088 return (dmu_read_impl(dn, offset, size, buf, flags)); 1089 } 1090 1091 static void 1092 dmu_write_impl(dmu_buf_t **dbp, int numbufs, uint64_t offset, uint64_t size, 1093 const void *buf, dmu_tx_t *tx) 1094 { 1095 int i; 1096 1097 for (i = 0; i < numbufs; i++) { 1098 uint64_t tocpy; 1099 int64_t bufoff; 1100 dmu_buf_t *db = dbp[i]; 1101 1102 ASSERT(size > 0); 1103 1104 bufoff = offset - db->db_offset; 1105 tocpy = MIN(db->db_size - bufoff, size); 1106 1107 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 1108 1109 if (tocpy == db->db_size) 1110 dmu_buf_will_fill(db, tx); 1111 else 1112 dmu_buf_will_dirty(db, tx); 1113 1114 (void) memcpy((char *)db->db_data + bufoff, buf, tocpy); 1115 1116 if (tocpy == db->db_size) 1117 dmu_buf_fill_done(db, tx); 1118 1119 offset += tocpy; 1120 size -= tocpy; 1121 buf = (char *)buf + tocpy; 1122 } 1123 } 1124 1125 void 1126 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1127 const void *buf, dmu_tx_t *tx) 1128 { 1129 dmu_buf_t **dbp; 1130 int numbufs; 1131 1132 if (size == 0) 1133 return; 1134 1135 VERIFY0(dmu_buf_hold_array(os, object, offset, size, 1136 FALSE, FTAG, &numbufs, &dbp)); 1137 dmu_write_impl(dbp, numbufs, offset, size, buf, tx); 1138 dmu_buf_rele_array(dbp, numbufs, FTAG); 1139 } 1140 1141 /* 1142 * Note: Lustre is an external consumer of this interface. 1143 */ 1144 void 1145 dmu_write_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, 1146 const void *buf, dmu_tx_t *tx) 1147 { 1148 dmu_buf_t **dbp; 1149 int numbufs; 1150 1151 if (size == 0) 1152 return; 1153 1154 VERIFY0(dmu_buf_hold_array_by_dnode(dn, offset, size, 1155 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH)); 1156 dmu_write_impl(dbp, numbufs, offset, size, buf, tx); 1157 dmu_buf_rele_array(dbp, numbufs, FTAG); 1158 } 1159 1160 void 1161 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1162 dmu_tx_t *tx) 1163 { 1164 dmu_buf_t **dbp; 1165 int numbufs, i; 1166 1167 if (size == 0) 1168 return; 1169 1170 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size, 1171 FALSE, FTAG, &numbufs, &dbp)); 1172 1173 for (i = 0; i < numbufs; i++) { 1174 dmu_buf_t *db = dbp[i]; 1175 1176 dmu_buf_will_not_fill(db, tx); 1177 } 1178 dmu_buf_rele_array(dbp, numbufs, FTAG); 1179 } 1180 1181 void 1182 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset, 1183 void *data, uint8_t etype, uint8_t comp, int uncompressed_size, 1184 int compressed_size, int byteorder, dmu_tx_t *tx) 1185 { 1186 dmu_buf_t *db; 1187 1188 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES); 1189 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS); 1190 VERIFY0(dmu_buf_hold_noread(os, object, offset, 1191 FTAG, &db)); 1192 1193 dmu_buf_write_embedded(db, 1194 data, (bp_embedded_type_t)etype, (enum zio_compress)comp, 1195 uncompressed_size, compressed_size, byteorder, tx); 1196 1197 dmu_buf_rele(db, FTAG); 1198 } 1199 1200 void 1201 dmu_redact(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1202 dmu_tx_t *tx) 1203 { 1204 int numbufs, i; 1205 dmu_buf_t **dbp; 1206 1207 VERIFY0(dmu_buf_hold_array(os, object, offset, size, FALSE, FTAG, 1208 &numbufs, &dbp)); 1209 for (i = 0; i < numbufs; i++) 1210 dmu_buf_redact(dbp[i], tx); 1211 dmu_buf_rele_array(dbp, numbufs, FTAG); 1212 } 1213 1214 #ifdef _KERNEL 1215 int 1216 dmu_read_uio_dnode(dnode_t *dn, zfs_uio_t *uio, uint64_t size) 1217 { 1218 dmu_buf_t **dbp; 1219 int numbufs, i, err; 1220 1221 /* 1222 * NB: we could do this block-at-a-time, but it's nice 1223 * to be reading in parallel. 1224 */ 1225 err = dmu_buf_hold_array_by_dnode(dn, zfs_uio_offset(uio), size, 1226 TRUE, FTAG, &numbufs, &dbp, 0); 1227 if (err) 1228 return (err); 1229 1230 for (i = 0; i < numbufs; i++) { 1231 uint64_t tocpy; 1232 int64_t bufoff; 1233 dmu_buf_t *db = dbp[i]; 1234 1235 ASSERT(size > 0); 1236 1237 bufoff = zfs_uio_offset(uio) - db->db_offset; 1238 tocpy = MIN(db->db_size - bufoff, size); 1239 1240 err = zfs_uio_fault_move((char *)db->db_data + bufoff, tocpy, 1241 UIO_READ, uio); 1242 1243 if (err) 1244 break; 1245 1246 size -= tocpy; 1247 } 1248 dmu_buf_rele_array(dbp, numbufs, FTAG); 1249 1250 return (err); 1251 } 1252 1253 /* 1254 * Read 'size' bytes into the uio buffer. 1255 * From object zdb->db_object. 1256 * Starting at zfs_uio_offset(uio). 1257 * 1258 * If the caller already has a dbuf in the target object 1259 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(), 1260 * because we don't have to find the dnode_t for the object. 1261 */ 1262 int 1263 dmu_read_uio_dbuf(dmu_buf_t *zdb, zfs_uio_t *uio, uint64_t size) 1264 { 1265 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb; 1266 dnode_t *dn; 1267 int err; 1268 1269 if (size == 0) 1270 return (0); 1271 1272 DB_DNODE_ENTER(db); 1273 dn = DB_DNODE(db); 1274 err = dmu_read_uio_dnode(dn, uio, size); 1275 DB_DNODE_EXIT(db); 1276 1277 return (err); 1278 } 1279 1280 /* 1281 * Read 'size' bytes into the uio buffer. 1282 * From the specified object 1283 * Starting at offset zfs_uio_offset(uio). 1284 */ 1285 int 1286 dmu_read_uio(objset_t *os, uint64_t object, zfs_uio_t *uio, uint64_t size) 1287 { 1288 dnode_t *dn; 1289 int err; 1290 1291 if (size == 0) 1292 return (0); 1293 1294 err = dnode_hold(os, object, FTAG, &dn); 1295 if (err) 1296 return (err); 1297 1298 err = dmu_read_uio_dnode(dn, uio, size); 1299 1300 dnode_rele(dn, FTAG); 1301 1302 return (err); 1303 } 1304 1305 int 1306 dmu_write_uio_dnode(dnode_t *dn, zfs_uio_t *uio, uint64_t size, dmu_tx_t *tx) 1307 { 1308 dmu_buf_t **dbp; 1309 int numbufs; 1310 int err = 0; 1311 int i; 1312 1313 err = dmu_buf_hold_array_by_dnode(dn, zfs_uio_offset(uio), size, 1314 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH); 1315 if (err) 1316 return (err); 1317 1318 for (i = 0; i < numbufs; i++) { 1319 uint64_t tocpy; 1320 int64_t bufoff; 1321 dmu_buf_t *db = dbp[i]; 1322 1323 ASSERT(size > 0); 1324 1325 bufoff = zfs_uio_offset(uio) - db->db_offset; 1326 tocpy = MIN(db->db_size - bufoff, size); 1327 1328 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 1329 1330 if (tocpy == db->db_size) 1331 dmu_buf_will_fill(db, tx); 1332 else 1333 dmu_buf_will_dirty(db, tx); 1334 1335 /* 1336 * XXX zfs_uiomove could block forever (eg.nfs-backed 1337 * pages). There needs to be a uiolockdown() function 1338 * to lock the pages in memory, so that zfs_uiomove won't 1339 * block. 1340 */ 1341 err = zfs_uio_fault_move((char *)db->db_data + bufoff, 1342 tocpy, UIO_WRITE, uio); 1343 1344 if (tocpy == db->db_size) 1345 dmu_buf_fill_done(db, tx); 1346 1347 if (err) 1348 break; 1349 1350 size -= tocpy; 1351 } 1352 1353 dmu_buf_rele_array(dbp, numbufs, FTAG); 1354 return (err); 1355 } 1356 1357 /* 1358 * Write 'size' bytes from the uio buffer. 1359 * To object zdb->db_object. 1360 * Starting at offset zfs_uio_offset(uio). 1361 * 1362 * If the caller already has a dbuf in the target object 1363 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(), 1364 * because we don't have to find the dnode_t for the object. 1365 */ 1366 int 1367 dmu_write_uio_dbuf(dmu_buf_t *zdb, zfs_uio_t *uio, uint64_t size, 1368 dmu_tx_t *tx) 1369 { 1370 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb; 1371 dnode_t *dn; 1372 int err; 1373 1374 if (size == 0) 1375 return (0); 1376 1377 DB_DNODE_ENTER(db); 1378 dn = DB_DNODE(db); 1379 err = dmu_write_uio_dnode(dn, uio, size, tx); 1380 DB_DNODE_EXIT(db); 1381 1382 return (err); 1383 } 1384 1385 /* 1386 * Write 'size' bytes from the uio buffer. 1387 * To the specified object. 1388 * Starting at offset zfs_uio_offset(uio). 1389 */ 1390 int 1391 dmu_write_uio(objset_t *os, uint64_t object, zfs_uio_t *uio, uint64_t size, 1392 dmu_tx_t *tx) 1393 { 1394 dnode_t *dn; 1395 int err; 1396 1397 if (size == 0) 1398 return (0); 1399 1400 err = dnode_hold(os, object, FTAG, &dn); 1401 if (err) 1402 return (err); 1403 1404 err = dmu_write_uio_dnode(dn, uio, size, tx); 1405 1406 dnode_rele(dn, FTAG); 1407 1408 return (err); 1409 } 1410 #endif /* _KERNEL */ 1411 1412 /* 1413 * Allocate a loaned anonymous arc buffer. 1414 */ 1415 arc_buf_t * 1416 dmu_request_arcbuf(dmu_buf_t *handle, int size) 1417 { 1418 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle; 1419 1420 return (arc_loan_buf(db->db_objset->os_spa, B_FALSE, size)); 1421 } 1422 1423 /* 1424 * Free a loaned arc buffer. 1425 */ 1426 void 1427 dmu_return_arcbuf(arc_buf_t *buf) 1428 { 1429 arc_return_buf(buf, FTAG); 1430 arc_buf_destroy(buf, FTAG); 1431 } 1432 1433 /* 1434 * A "lightweight" write is faster than a regular write (e.g. 1435 * dmu_write_by_dnode() or dmu_assign_arcbuf_by_dnode()), because it avoids the 1436 * CPU cost of creating a dmu_buf_impl_t and arc_buf_[hdr_]_t. However, the 1437 * data can not be read or overwritten until the transaction's txg has been 1438 * synced. This makes it appropriate for workloads that are known to be 1439 * (temporarily) write-only, like "zfs receive". 1440 * 1441 * A single block is written, starting at the specified offset in bytes. If 1442 * the call is successful, it returns 0 and the provided abd has been 1443 * consumed (the caller should not free it). 1444 */ 1445 int 1446 dmu_lightweight_write_by_dnode(dnode_t *dn, uint64_t offset, abd_t *abd, 1447 const zio_prop_t *zp, zio_flag_t flags, dmu_tx_t *tx) 1448 { 1449 dbuf_dirty_record_t *dr = 1450 dbuf_dirty_lightweight(dn, dbuf_whichblock(dn, 0, offset), tx); 1451 if (dr == NULL) 1452 return (SET_ERROR(EIO)); 1453 dr->dt.dll.dr_abd = abd; 1454 dr->dt.dll.dr_props = *zp; 1455 dr->dt.dll.dr_flags = flags; 1456 return (0); 1457 } 1458 1459 /* 1460 * When possible directly assign passed loaned arc buffer to a dbuf. 1461 * If this is not possible copy the contents of passed arc buf via 1462 * dmu_write(). 1463 */ 1464 int 1465 dmu_assign_arcbuf_by_dnode(dnode_t *dn, uint64_t offset, arc_buf_t *buf, 1466 dmu_tx_t *tx) 1467 { 1468 dmu_buf_impl_t *db; 1469 objset_t *os = dn->dn_objset; 1470 uint64_t object = dn->dn_object; 1471 uint32_t blksz = (uint32_t)arc_buf_lsize(buf); 1472 uint64_t blkid; 1473 1474 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1475 blkid = dbuf_whichblock(dn, 0, offset); 1476 db = dbuf_hold(dn, blkid, FTAG); 1477 if (db == NULL) 1478 return (SET_ERROR(EIO)); 1479 rw_exit(&dn->dn_struct_rwlock); 1480 1481 /* 1482 * We can only assign if the offset is aligned and the arc buf is the 1483 * same size as the dbuf. 1484 */ 1485 if (offset == db->db.db_offset && blksz == db->db.db_size) { 1486 zfs_racct_write(blksz, 1); 1487 dbuf_assign_arcbuf(db, buf, tx); 1488 dbuf_rele(db, FTAG); 1489 } else { 1490 /* compressed bufs must always be assignable to their dbuf */ 1491 ASSERT3U(arc_get_compression(buf), ==, ZIO_COMPRESS_OFF); 1492 ASSERT(!(buf->b_flags & ARC_BUF_FLAG_COMPRESSED)); 1493 1494 dbuf_rele(db, FTAG); 1495 dmu_write(os, object, offset, blksz, buf->b_data, tx); 1496 dmu_return_arcbuf(buf); 1497 } 1498 1499 return (0); 1500 } 1501 1502 int 1503 dmu_assign_arcbuf_by_dbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf, 1504 dmu_tx_t *tx) 1505 { 1506 int err; 1507 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle; 1508 1509 DB_DNODE_ENTER(dbuf); 1510 err = dmu_assign_arcbuf_by_dnode(DB_DNODE(dbuf), offset, buf, tx); 1511 DB_DNODE_EXIT(dbuf); 1512 1513 return (err); 1514 } 1515 1516 typedef struct { 1517 dbuf_dirty_record_t *dsa_dr; 1518 dmu_sync_cb_t *dsa_done; 1519 zgd_t *dsa_zgd; 1520 dmu_tx_t *dsa_tx; 1521 } dmu_sync_arg_t; 1522 1523 static void 1524 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg) 1525 { 1526 (void) buf; 1527 dmu_sync_arg_t *dsa = varg; 1528 dmu_buf_t *db = dsa->dsa_zgd->zgd_db; 1529 blkptr_t *bp = zio->io_bp; 1530 1531 if (zio->io_error == 0) { 1532 if (BP_IS_HOLE(bp)) { 1533 /* 1534 * A block of zeros may compress to a hole, but the 1535 * block size still needs to be known for replay. 1536 */ 1537 BP_SET_LSIZE(bp, db->db_size); 1538 } else if (!BP_IS_EMBEDDED(bp)) { 1539 ASSERT(BP_GET_LEVEL(bp) == 0); 1540 BP_SET_FILL(bp, 1); 1541 } 1542 } 1543 } 1544 1545 static void 1546 dmu_sync_late_arrival_ready(zio_t *zio) 1547 { 1548 dmu_sync_ready(zio, NULL, zio->io_private); 1549 } 1550 1551 static void 1552 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg) 1553 { 1554 (void) buf; 1555 dmu_sync_arg_t *dsa = varg; 1556 dbuf_dirty_record_t *dr = dsa->dsa_dr; 1557 dmu_buf_impl_t *db = dr->dr_dbuf; 1558 zgd_t *zgd = dsa->dsa_zgd; 1559 1560 /* 1561 * Record the vdev(s) backing this blkptr so they can be flushed after 1562 * the writes for the lwb have completed. 1563 */ 1564 if (zio->io_error == 0) { 1565 zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp); 1566 } 1567 1568 mutex_enter(&db->db_mtx); 1569 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC); 1570 if (zio->io_error == 0) { 1571 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE); 1572 if (dr->dt.dl.dr_nopwrite) { 1573 blkptr_t *bp = zio->io_bp; 1574 blkptr_t *bp_orig = &zio->io_bp_orig; 1575 uint8_t chksum = BP_GET_CHECKSUM(bp_orig); 1576 1577 ASSERT(BP_EQUAL(bp, bp_orig)); 1578 VERIFY(BP_EQUAL(bp, db->db_blkptr)); 1579 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF); 1580 VERIFY(zio_checksum_table[chksum].ci_flags & 1581 ZCHECKSUM_FLAG_NOPWRITE); 1582 } 1583 dr->dt.dl.dr_overridden_by = *zio->io_bp; 1584 dr->dt.dl.dr_override_state = DR_OVERRIDDEN; 1585 dr->dt.dl.dr_copies = zio->io_prop.zp_copies; 1586 1587 /* 1588 * Old style holes are filled with all zeros, whereas 1589 * new-style holes maintain their lsize, type, level, 1590 * and birth time (see zio_write_compress). While we 1591 * need to reset the BP_SET_LSIZE() call that happened 1592 * in dmu_sync_ready for old style holes, we do *not* 1593 * want to wipe out the information contained in new 1594 * style holes. Thus, only zero out the block pointer if 1595 * it's an old style hole. 1596 */ 1597 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) && 1598 dr->dt.dl.dr_overridden_by.blk_birth == 0) 1599 BP_ZERO(&dr->dt.dl.dr_overridden_by); 1600 } else { 1601 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 1602 } 1603 cv_broadcast(&db->db_changed); 1604 mutex_exit(&db->db_mtx); 1605 1606 dsa->dsa_done(dsa->dsa_zgd, zio->io_error); 1607 1608 kmem_free(dsa, sizeof (*dsa)); 1609 } 1610 1611 static void 1612 dmu_sync_late_arrival_done(zio_t *zio) 1613 { 1614 blkptr_t *bp = zio->io_bp; 1615 dmu_sync_arg_t *dsa = zio->io_private; 1616 zgd_t *zgd = dsa->dsa_zgd; 1617 1618 if (zio->io_error == 0) { 1619 /* 1620 * Record the vdev(s) backing this blkptr so they can be 1621 * flushed after the writes for the lwb have completed. 1622 */ 1623 zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp); 1624 1625 if (!BP_IS_HOLE(bp)) { 1626 blkptr_t *bp_orig __maybe_unused = &zio->io_bp_orig; 1627 ASSERT(!(zio->io_flags & ZIO_FLAG_NOPWRITE)); 1628 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig)); 1629 ASSERT(zio->io_bp->blk_birth == zio->io_txg); 1630 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa)); 1631 zio_free(zio->io_spa, zio->io_txg, zio->io_bp); 1632 } 1633 } 1634 1635 dmu_tx_commit(dsa->dsa_tx); 1636 1637 dsa->dsa_done(dsa->dsa_zgd, zio->io_error); 1638 1639 abd_free(zio->io_abd); 1640 kmem_free(dsa, sizeof (*dsa)); 1641 } 1642 1643 static int 1644 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd, 1645 zio_prop_t *zp, zbookmark_phys_t *zb) 1646 { 1647 dmu_sync_arg_t *dsa; 1648 dmu_tx_t *tx; 1649 1650 tx = dmu_tx_create(os); 1651 dmu_tx_hold_space(tx, zgd->zgd_db->db_size); 1652 if (dmu_tx_assign(tx, TXG_WAIT) != 0) { 1653 dmu_tx_abort(tx); 1654 /* Make zl_get_data do txg_waited_synced() */ 1655 return (SET_ERROR(EIO)); 1656 } 1657 1658 /* 1659 * In order to prevent the zgd's lwb from being free'd prior to 1660 * dmu_sync_late_arrival_done() being called, we have to ensure 1661 * the lwb's "max txg" takes this tx's txg into account. 1662 */ 1663 zil_lwb_add_txg(zgd->zgd_lwb, dmu_tx_get_txg(tx)); 1664 1665 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP); 1666 dsa->dsa_dr = NULL; 1667 dsa->dsa_done = done; 1668 dsa->dsa_zgd = zgd; 1669 dsa->dsa_tx = tx; 1670 1671 /* 1672 * Since we are currently syncing this txg, it's nontrivial to 1673 * determine what BP to nopwrite against, so we disable nopwrite. 1674 * 1675 * When syncing, the db_blkptr is initially the BP of the previous 1676 * txg. We can not nopwrite against it because it will be changed 1677 * (this is similar to the non-late-arrival case where the dbuf is 1678 * dirty in a future txg). 1679 * 1680 * Then dbuf_write_ready() sets bp_blkptr to the location we will write. 1681 * We can not nopwrite against it because although the BP will not 1682 * (typically) be changed, the data has not yet been persisted to this 1683 * location. 1684 * 1685 * Finally, when dbuf_write_done() is called, it is theoretically 1686 * possible to always nopwrite, because the data that was written in 1687 * this txg is the same data that we are trying to write. However we 1688 * would need to check that this dbuf is not dirty in any future 1689 * txg's (as we do in the normal dmu_sync() path). For simplicity, we 1690 * don't nopwrite in this case. 1691 */ 1692 zp->zp_nopwrite = B_FALSE; 1693 1694 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp, 1695 abd_get_from_buf(zgd->zgd_db->db_data, zgd->zgd_db->db_size), 1696 zgd->zgd_db->db_size, zgd->zgd_db->db_size, zp, 1697 dmu_sync_late_arrival_ready, NULL, NULL, dmu_sync_late_arrival_done, 1698 dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb)); 1699 1700 return (0); 1701 } 1702 1703 /* 1704 * Intent log support: sync the block associated with db to disk. 1705 * N.B. and XXX: the caller is responsible for making sure that the 1706 * data isn't changing while dmu_sync() is writing it. 1707 * 1708 * Return values: 1709 * 1710 * EEXIST: this txg has already been synced, so there's nothing to do. 1711 * The caller should not log the write. 1712 * 1713 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do. 1714 * The caller should not log the write. 1715 * 1716 * EALREADY: this block is already in the process of being synced. 1717 * The caller should track its progress (somehow). 1718 * 1719 * EIO: could not do the I/O. 1720 * The caller should do a txg_wait_synced(). 1721 * 1722 * 0: the I/O has been initiated. 1723 * The caller should log this blkptr in the done callback. 1724 * It is possible that the I/O will fail, in which case 1725 * the error will be reported to the done callback and 1726 * propagated to pio from zio_done(). 1727 */ 1728 int 1729 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd) 1730 { 1731 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db; 1732 objset_t *os = db->db_objset; 1733 dsl_dataset_t *ds = os->os_dsl_dataset; 1734 dbuf_dirty_record_t *dr, *dr_next; 1735 dmu_sync_arg_t *dsa; 1736 zbookmark_phys_t zb; 1737 zio_prop_t zp; 1738 dnode_t *dn; 1739 1740 ASSERT(pio != NULL); 1741 ASSERT(txg != 0); 1742 1743 SET_BOOKMARK(&zb, ds->ds_object, 1744 db->db.db_object, db->db_level, db->db_blkid); 1745 1746 DB_DNODE_ENTER(db); 1747 dn = DB_DNODE(db); 1748 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp); 1749 DB_DNODE_EXIT(db); 1750 1751 /* 1752 * If we're frozen (running ziltest), we always need to generate a bp. 1753 */ 1754 if (txg > spa_freeze_txg(os->os_spa)) 1755 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); 1756 1757 /* 1758 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf() 1759 * and us. If we determine that this txg is not yet syncing, 1760 * but it begins to sync a moment later, that's OK because the 1761 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx. 1762 */ 1763 mutex_enter(&db->db_mtx); 1764 1765 if (txg <= spa_last_synced_txg(os->os_spa)) { 1766 /* 1767 * This txg has already synced. There's nothing to do. 1768 */ 1769 mutex_exit(&db->db_mtx); 1770 return (SET_ERROR(EEXIST)); 1771 } 1772 1773 if (txg <= spa_syncing_txg(os->os_spa)) { 1774 /* 1775 * This txg is currently syncing, so we can't mess with 1776 * the dirty record anymore; just write a new log block. 1777 */ 1778 mutex_exit(&db->db_mtx); 1779 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); 1780 } 1781 1782 dr = dbuf_find_dirty_eq(db, txg); 1783 1784 if (dr == NULL) { 1785 /* 1786 * There's no dr for this dbuf, so it must have been freed. 1787 * There's no need to log writes to freed blocks, so we're done. 1788 */ 1789 mutex_exit(&db->db_mtx); 1790 return (SET_ERROR(ENOENT)); 1791 } 1792 1793 dr_next = list_next(&db->db_dirty_records, dr); 1794 ASSERT(dr_next == NULL || dr_next->dr_txg < txg); 1795 1796 if (db->db_blkptr != NULL) { 1797 /* 1798 * We need to fill in zgd_bp with the current blkptr so that 1799 * the nopwrite code can check if we're writing the same 1800 * data that's already on disk. We can only nopwrite if we 1801 * are sure that after making the copy, db_blkptr will not 1802 * change until our i/o completes. We ensure this by 1803 * holding the db_mtx, and only allowing nopwrite if the 1804 * block is not already dirty (see below). This is verified 1805 * by dmu_sync_done(), which VERIFYs that the db_blkptr has 1806 * not changed. 1807 */ 1808 *zgd->zgd_bp = *db->db_blkptr; 1809 } 1810 1811 /* 1812 * Assume the on-disk data is X, the current syncing data (in 1813 * txg - 1) is Y, and the current in-memory data is Z (currently 1814 * in dmu_sync). 1815 * 1816 * We usually want to perform a nopwrite if X and Z are the 1817 * same. However, if Y is different (i.e. the BP is going to 1818 * change before this write takes effect), then a nopwrite will 1819 * be incorrect - we would override with X, which could have 1820 * been freed when Y was written. 1821 * 1822 * (Note that this is not a concern when we are nop-writing from 1823 * syncing context, because X and Y must be identical, because 1824 * all previous txgs have been synced.) 1825 * 1826 * Therefore, we disable nopwrite if the current BP could change 1827 * before this TXG. There are two ways it could change: by 1828 * being dirty (dr_next is non-NULL), or by being freed 1829 * (dnode_block_freed()). This behavior is verified by 1830 * zio_done(), which VERIFYs that the override BP is identical 1831 * to the on-disk BP. 1832 */ 1833 DB_DNODE_ENTER(db); 1834 dn = DB_DNODE(db); 1835 if (dr_next != NULL || dnode_block_freed(dn, db->db_blkid)) 1836 zp.zp_nopwrite = B_FALSE; 1837 DB_DNODE_EXIT(db); 1838 1839 ASSERT(dr->dr_txg == txg); 1840 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC || 1841 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { 1842 /* 1843 * We have already issued a sync write for this buffer, 1844 * or this buffer has already been synced. It could not 1845 * have been dirtied since, or we would have cleared the state. 1846 */ 1847 mutex_exit(&db->db_mtx); 1848 return (SET_ERROR(EALREADY)); 1849 } 1850 1851 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); 1852 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC; 1853 mutex_exit(&db->db_mtx); 1854 1855 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP); 1856 dsa->dsa_dr = dr; 1857 dsa->dsa_done = done; 1858 dsa->dsa_zgd = zgd; 1859 dsa->dsa_tx = NULL; 1860 1861 zio_nowait(arc_write(pio, os->os_spa, txg, zgd->zgd_bp, 1862 dr->dt.dl.dr_data, !DBUF_IS_CACHEABLE(db), dbuf_is_l2cacheable(db), 1863 &zp, dmu_sync_ready, NULL, NULL, dmu_sync_done, dsa, 1864 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb)); 1865 1866 return (0); 1867 } 1868 1869 int 1870 dmu_object_set_nlevels(objset_t *os, uint64_t object, int nlevels, dmu_tx_t *tx) 1871 { 1872 dnode_t *dn; 1873 int err; 1874 1875 err = dnode_hold(os, object, FTAG, &dn); 1876 if (err) 1877 return (err); 1878 err = dnode_set_nlevels(dn, nlevels, tx); 1879 dnode_rele(dn, FTAG); 1880 return (err); 1881 } 1882 1883 int 1884 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs, 1885 dmu_tx_t *tx) 1886 { 1887 dnode_t *dn; 1888 int err; 1889 1890 err = dnode_hold(os, object, FTAG, &dn); 1891 if (err) 1892 return (err); 1893 err = dnode_set_blksz(dn, size, ibs, tx); 1894 dnode_rele(dn, FTAG); 1895 return (err); 1896 } 1897 1898 int 1899 dmu_object_set_maxblkid(objset_t *os, uint64_t object, uint64_t maxblkid, 1900 dmu_tx_t *tx) 1901 { 1902 dnode_t *dn; 1903 int err; 1904 1905 err = dnode_hold(os, object, FTAG, &dn); 1906 if (err) 1907 return (err); 1908 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 1909 dnode_new_blkid(dn, maxblkid, tx, B_FALSE, B_TRUE); 1910 rw_exit(&dn->dn_struct_rwlock); 1911 dnode_rele(dn, FTAG); 1912 return (0); 1913 } 1914 1915 void 1916 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum, 1917 dmu_tx_t *tx) 1918 { 1919 dnode_t *dn; 1920 1921 /* 1922 * Send streams include each object's checksum function. This 1923 * check ensures that the receiving system can understand the 1924 * checksum function transmitted. 1925 */ 1926 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS); 1927 1928 VERIFY0(dnode_hold(os, object, FTAG, &dn)); 1929 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS); 1930 dn->dn_checksum = checksum; 1931 dnode_setdirty(dn, tx); 1932 dnode_rele(dn, FTAG); 1933 } 1934 1935 void 1936 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress, 1937 dmu_tx_t *tx) 1938 { 1939 dnode_t *dn; 1940 1941 /* 1942 * Send streams include each object's compression function. This 1943 * check ensures that the receiving system can understand the 1944 * compression function transmitted. 1945 */ 1946 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS); 1947 1948 VERIFY0(dnode_hold(os, object, FTAG, &dn)); 1949 dn->dn_compress = compress; 1950 dnode_setdirty(dn, tx); 1951 dnode_rele(dn, FTAG); 1952 } 1953 1954 /* 1955 * When the "redundant_metadata" property is set to "most", only indirect 1956 * blocks of this level and higher will have an additional ditto block. 1957 */ 1958 static const int zfs_redundant_metadata_most_ditto_level = 2; 1959 1960 void 1961 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp) 1962 { 1963 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET; 1964 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) || 1965 (wp & WP_SPILL)); 1966 enum zio_checksum checksum = os->os_checksum; 1967 enum zio_compress compress = os->os_compress; 1968 uint8_t complevel = os->os_complevel; 1969 enum zio_checksum dedup_checksum = os->os_dedup_checksum; 1970 boolean_t dedup = B_FALSE; 1971 boolean_t nopwrite = B_FALSE; 1972 boolean_t dedup_verify = os->os_dedup_verify; 1973 boolean_t encrypt = B_FALSE; 1974 int copies = os->os_copies; 1975 1976 /* 1977 * We maintain different write policies for each of the following 1978 * types of data: 1979 * 1. metadata 1980 * 2. preallocated blocks (i.e. level-0 blocks of a dump device) 1981 * 3. all other level 0 blocks 1982 */ 1983 if (ismd) { 1984 /* 1985 * XXX -- we should design a compression algorithm 1986 * that specializes in arrays of bps. 1987 */ 1988 compress = zio_compress_select(os->os_spa, 1989 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON); 1990 1991 /* 1992 * Metadata always gets checksummed. If the data 1993 * checksum is multi-bit correctable, and it's not a 1994 * ZBT-style checksum, then it's suitable for metadata 1995 * as well. Otherwise, the metadata checksum defaults 1996 * to fletcher4. 1997 */ 1998 if (!(zio_checksum_table[checksum].ci_flags & 1999 ZCHECKSUM_FLAG_METADATA) || 2000 (zio_checksum_table[checksum].ci_flags & 2001 ZCHECKSUM_FLAG_EMBEDDED)) 2002 checksum = ZIO_CHECKSUM_FLETCHER_4; 2003 2004 switch (os->os_redundant_metadata) { 2005 case ZFS_REDUNDANT_METADATA_ALL: 2006 copies++; 2007 break; 2008 case ZFS_REDUNDANT_METADATA_MOST: 2009 if (level >= zfs_redundant_metadata_most_ditto_level || 2010 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL)) 2011 copies++; 2012 break; 2013 case ZFS_REDUNDANT_METADATA_SOME: 2014 if (DMU_OT_IS_CRITICAL(type)) 2015 copies++; 2016 break; 2017 case ZFS_REDUNDANT_METADATA_NONE: 2018 break; 2019 } 2020 } else if (wp & WP_NOFILL) { 2021 ASSERT(level == 0); 2022 2023 /* 2024 * If we're writing preallocated blocks, we aren't actually 2025 * writing them so don't set any policy properties. These 2026 * blocks are currently only used by an external subsystem 2027 * outside of zfs (i.e. dump) and not written by the zio 2028 * pipeline. 2029 */ 2030 compress = ZIO_COMPRESS_OFF; 2031 checksum = ZIO_CHECKSUM_OFF; 2032 } else { 2033 compress = zio_compress_select(os->os_spa, dn->dn_compress, 2034 compress); 2035 complevel = zio_complevel_select(os->os_spa, compress, 2036 complevel, complevel); 2037 2038 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ? 2039 zio_checksum_select(dn->dn_checksum, checksum) : 2040 dedup_checksum; 2041 2042 /* 2043 * Determine dedup setting. If we are in dmu_sync(), 2044 * we won't actually dedup now because that's all 2045 * done in syncing context; but we do want to use the 2046 * dedup checksum. If the checksum is not strong 2047 * enough to ensure unique signatures, force 2048 * dedup_verify. 2049 */ 2050 if (dedup_checksum != ZIO_CHECKSUM_OFF) { 2051 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE; 2052 if (!(zio_checksum_table[checksum].ci_flags & 2053 ZCHECKSUM_FLAG_DEDUP)) 2054 dedup_verify = B_TRUE; 2055 } 2056 2057 /* 2058 * Enable nopwrite if we have secure enough checksum 2059 * algorithm (see comment in zio_nop_write) and 2060 * compression is enabled. We don't enable nopwrite if 2061 * dedup is enabled as the two features are mutually 2062 * exclusive. 2063 */ 2064 nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags & 2065 ZCHECKSUM_FLAG_NOPWRITE) && 2066 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled); 2067 } 2068 2069 /* 2070 * All objects in an encrypted objset are protected from modification 2071 * via a MAC. Encrypted objects store their IV and salt in the last DVA 2072 * in the bp, so we cannot use all copies. Encrypted objects are also 2073 * not subject to nopwrite since writing the same data will still 2074 * result in a new ciphertext. Only encrypted blocks can be dedup'd 2075 * to avoid ambiguity in the dedup code since the DDT does not store 2076 * object types. 2077 */ 2078 if (os->os_encrypted && (wp & WP_NOFILL) == 0) { 2079 encrypt = B_TRUE; 2080 2081 if (DMU_OT_IS_ENCRYPTED(type)) { 2082 copies = MIN(copies, SPA_DVAS_PER_BP - 1); 2083 nopwrite = B_FALSE; 2084 } else { 2085 dedup = B_FALSE; 2086 } 2087 2088 if (level <= 0 && 2089 (type == DMU_OT_DNODE || type == DMU_OT_OBJSET)) { 2090 compress = ZIO_COMPRESS_EMPTY; 2091 } 2092 } 2093 2094 zp->zp_compress = compress; 2095 zp->zp_complevel = complevel; 2096 zp->zp_checksum = checksum; 2097 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type; 2098 zp->zp_level = level; 2099 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa)); 2100 zp->zp_dedup = dedup; 2101 zp->zp_dedup_verify = dedup && dedup_verify; 2102 zp->zp_nopwrite = nopwrite; 2103 zp->zp_encrypt = encrypt; 2104 zp->zp_byteorder = ZFS_HOST_BYTEORDER; 2105 memset(zp->zp_salt, 0, ZIO_DATA_SALT_LEN); 2106 memset(zp->zp_iv, 0, ZIO_DATA_IV_LEN); 2107 memset(zp->zp_mac, 0, ZIO_DATA_MAC_LEN); 2108 zp->zp_zpl_smallblk = DMU_OT_IS_FILE(zp->zp_type) ? 2109 os->os_zpl_special_smallblock : 0; 2110 2111 ASSERT3U(zp->zp_compress, !=, ZIO_COMPRESS_INHERIT); 2112 } 2113 2114 /* 2115 * This function is only called from zfs_holey_common() for zpl_llseek() 2116 * in order to determine the location of holes. In order to accurately 2117 * report holes all dirty data must be synced to disk. This causes extremely 2118 * poor performance when seeking for holes in a dirty file. As a compromise, 2119 * only provide hole data when the dnode is clean. When a dnode is dirty 2120 * report the dnode as having no holes which is always a safe thing to do. 2121 */ 2122 int 2123 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off) 2124 { 2125 dnode_t *dn; 2126 int err; 2127 2128 restart: 2129 err = dnode_hold(os, object, FTAG, &dn); 2130 if (err) 2131 return (err); 2132 2133 rw_enter(&dn->dn_struct_rwlock, RW_READER); 2134 2135 if (dnode_is_dirty(dn)) { 2136 /* 2137 * If the zfs_dmu_offset_next_sync module option is enabled 2138 * then strict hole reporting has been requested. Dirty 2139 * dnodes must be synced to disk to accurately report all 2140 * holes. When disabled dirty dnodes are reported to not 2141 * have any holes which is always safe. 2142 * 2143 * When called by zfs_holey_common() the zp->z_rangelock 2144 * is held to prevent zfs_write() and mmap writeback from 2145 * re-dirtying the dnode after txg_wait_synced(). 2146 */ 2147 if (zfs_dmu_offset_next_sync) { 2148 rw_exit(&dn->dn_struct_rwlock); 2149 dnode_rele(dn, FTAG); 2150 txg_wait_synced(dmu_objset_pool(os), 0); 2151 goto restart; 2152 } 2153 2154 err = SET_ERROR(EBUSY); 2155 } else { 2156 err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK | 2157 (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0); 2158 } 2159 2160 rw_exit(&dn->dn_struct_rwlock); 2161 dnode_rele(dn, FTAG); 2162 2163 return (err); 2164 } 2165 2166 void 2167 __dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi) 2168 { 2169 dnode_phys_t *dnp = dn->dn_phys; 2170 2171 doi->doi_data_block_size = dn->dn_datablksz; 2172 doi->doi_metadata_block_size = dn->dn_indblkshift ? 2173 1ULL << dn->dn_indblkshift : 0; 2174 doi->doi_type = dn->dn_type; 2175 doi->doi_bonus_type = dn->dn_bonustype; 2176 doi->doi_bonus_size = dn->dn_bonuslen; 2177 doi->doi_dnodesize = dn->dn_num_slots << DNODE_SHIFT; 2178 doi->doi_indirection = dn->dn_nlevels; 2179 doi->doi_checksum = dn->dn_checksum; 2180 doi->doi_compress = dn->dn_compress; 2181 doi->doi_nblkptr = dn->dn_nblkptr; 2182 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9; 2183 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz; 2184 doi->doi_fill_count = 0; 2185 for (int i = 0; i < dnp->dn_nblkptr; i++) 2186 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]); 2187 } 2188 2189 void 2190 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi) 2191 { 2192 rw_enter(&dn->dn_struct_rwlock, RW_READER); 2193 mutex_enter(&dn->dn_mtx); 2194 2195 __dmu_object_info_from_dnode(dn, doi); 2196 2197 mutex_exit(&dn->dn_mtx); 2198 rw_exit(&dn->dn_struct_rwlock); 2199 } 2200 2201 /* 2202 * Get information on a DMU object. 2203 * If doi is NULL, just indicates whether the object exists. 2204 */ 2205 int 2206 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi) 2207 { 2208 dnode_t *dn; 2209 int err = dnode_hold(os, object, FTAG, &dn); 2210 2211 if (err) 2212 return (err); 2213 2214 if (doi != NULL) 2215 dmu_object_info_from_dnode(dn, doi); 2216 2217 dnode_rele(dn, FTAG); 2218 return (0); 2219 } 2220 2221 /* 2222 * As above, but faster; can be used when you have a held dbuf in hand. 2223 */ 2224 void 2225 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi) 2226 { 2227 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2228 2229 DB_DNODE_ENTER(db); 2230 dmu_object_info_from_dnode(DB_DNODE(db), doi); 2231 DB_DNODE_EXIT(db); 2232 } 2233 2234 /* 2235 * Faster still when you only care about the size. 2236 */ 2237 void 2238 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize, 2239 u_longlong_t *nblk512) 2240 { 2241 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2242 dnode_t *dn; 2243 2244 DB_DNODE_ENTER(db); 2245 dn = DB_DNODE(db); 2246 2247 *blksize = dn->dn_datablksz; 2248 /* add in number of slots used for the dnode itself */ 2249 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >> 2250 SPA_MINBLOCKSHIFT) + dn->dn_num_slots; 2251 DB_DNODE_EXIT(db); 2252 } 2253 2254 void 2255 dmu_object_dnsize_from_db(dmu_buf_t *db_fake, int *dnsize) 2256 { 2257 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2258 dnode_t *dn; 2259 2260 DB_DNODE_ENTER(db); 2261 dn = DB_DNODE(db); 2262 *dnsize = dn->dn_num_slots << DNODE_SHIFT; 2263 DB_DNODE_EXIT(db); 2264 } 2265 2266 void 2267 byteswap_uint64_array(void *vbuf, size_t size) 2268 { 2269 uint64_t *buf = vbuf; 2270 size_t count = size >> 3; 2271 int i; 2272 2273 ASSERT((size & 7) == 0); 2274 2275 for (i = 0; i < count; i++) 2276 buf[i] = BSWAP_64(buf[i]); 2277 } 2278 2279 void 2280 byteswap_uint32_array(void *vbuf, size_t size) 2281 { 2282 uint32_t *buf = vbuf; 2283 size_t count = size >> 2; 2284 int i; 2285 2286 ASSERT((size & 3) == 0); 2287 2288 for (i = 0; i < count; i++) 2289 buf[i] = BSWAP_32(buf[i]); 2290 } 2291 2292 void 2293 byteswap_uint16_array(void *vbuf, size_t size) 2294 { 2295 uint16_t *buf = vbuf; 2296 size_t count = size >> 1; 2297 int i; 2298 2299 ASSERT((size & 1) == 0); 2300 2301 for (i = 0; i < count; i++) 2302 buf[i] = BSWAP_16(buf[i]); 2303 } 2304 2305 void 2306 byteswap_uint8_array(void *vbuf, size_t size) 2307 { 2308 (void) vbuf, (void) size; 2309 } 2310 2311 void 2312 dmu_init(void) 2313 { 2314 abd_init(); 2315 zfs_dbgmsg_init(); 2316 sa_cache_init(); 2317 dmu_objset_init(); 2318 dnode_init(); 2319 zfetch_init(); 2320 dmu_tx_init(); 2321 l2arc_init(); 2322 arc_init(); 2323 dbuf_init(); 2324 } 2325 2326 void 2327 dmu_fini(void) 2328 { 2329 arc_fini(); /* arc depends on l2arc, so arc must go first */ 2330 l2arc_fini(); 2331 dmu_tx_fini(); 2332 zfetch_fini(); 2333 dbuf_fini(); 2334 dnode_fini(); 2335 dmu_objset_fini(); 2336 sa_cache_fini(); 2337 zfs_dbgmsg_fini(); 2338 abd_fini(); 2339 } 2340 2341 EXPORT_SYMBOL(dmu_bonus_hold); 2342 EXPORT_SYMBOL(dmu_bonus_hold_by_dnode); 2343 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus); 2344 EXPORT_SYMBOL(dmu_buf_rele_array); 2345 EXPORT_SYMBOL(dmu_prefetch); 2346 EXPORT_SYMBOL(dmu_free_range); 2347 EXPORT_SYMBOL(dmu_free_long_range); 2348 EXPORT_SYMBOL(dmu_free_long_object); 2349 EXPORT_SYMBOL(dmu_read); 2350 EXPORT_SYMBOL(dmu_read_by_dnode); 2351 EXPORT_SYMBOL(dmu_write); 2352 EXPORT_SYMBOL(dmu_write_by_dnode); 2353 EXPORT_SYMBOL(dmu_prealloc); 2354 EXPORT_SYMBOL(dmu_object_info); 2355 EXPORT_SYMBOL(dmu_object_info_from_dnode); 2356 EXPORT_SYMBOL(dmu_object_info_from_db); 2357 EXPORT_SYMBOL(dmu_object_size_from_db); 2358 EXPORT_SYMBOL(dmu_object_dnsize_from_db); 2359 EXPORT_SYMBOL(dmu_object_set_nlevels); 2360 EXPORT_SYMBOL(dmu_object_set_blocksize); 2361 EXPORT_SYMBOL(dmu_object_set_maxblkid); 2362 EXPORT_SYMBOL(dmu_object_set_checksum); 2363 EXPORT_SYMBOL(dmu_object_set_compress); 2364 EXPORT_SYMBOL(dmu_offset_next); 2365 EXPORT_SYMBOL(dmu_write_policy); 2366 EXPORT_SYMBOL(dmu_sync); 2367 EXPORT_SYMBOL(dmu_request_arcbuf); 2368 EXPORT_SYMBOL(dmu_return_arcbuf); 2369 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dnode); 2370 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dbuf); 2371 EXPORT_SYMBOL(dmu_buf_hold); 2372 EXPORT_SYMBOL(dmu_ot); 2373 2374 ZFS_MODULE_PARAM(zfs, zfs_, nopwrite_enabled, INT, ZMOD_RW, 2375 "Enable NOP writes"); 2376 2377 ZFS_MODULE_PARAM(zfs, zfs_, per_txg_dirty_frees_percent, UINT, ZMOD_RW, 2378 "Percentage of dirtied blocks from frees in one TXG"); 2379 2380 ZFS_MODULE_PARAM(zfs, zfs_, dmu_offset_next_sync, INT, ZMOD_RW, 2381 "Enable forcing txg sync to find holes"); 2382 2383 /* CSTYLED */ 2384 ZFS_MODULE_PARAM(zfs, , dmu_prefetch_max, UINT, ZMOD_RW, 2385 "Limit one prefetch call to this size");
Cache object: a7b857aaf0ce90d90a9b9b5bb533f0d0
[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]