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, 2023, Klara Inc. 30 * Copyright (c) 2019, Allan Jude 31 * Copyright (c) 2022 Hewlett Packard Enterprise Development LP. 32 * Copyright (c) 2021, 2022 by Pawel Jakub Dawidek 33 */ 34 35 #include <sys/dmu.h> 36 #include <sys/dmu_impl.h> 37 #include <sys/dmu_tx.h> 38 #include <sys/dbuf.h> 39 #include <sys/dnode.h> 40 #include <sys/zfs_context.h> 41 #include <sys/dmu_objset.h> 42 #include <sys/dmu_traverse.h> 43 #include <sys/dsl_dataset.h> 44 #include <sys/dsl_dir.h> 45 #include <sys/dsl_pool.h> 46 #include <sys/dsl_synctask.h> 47 #include <sys/dsl_prop.h> 48 #include <sys/dmu_zfetch.h> 49 #include <sys/zfs_ioctl.h> 50 #include <sys/zap.h> 51 #include <sys/zio_checksum.h> 52 #include <sys/zio_compress.h> 53 #include <sys/sa.h> 54 #include <sys/zfeature.h> 55 #include <sys/abd.h> 56 #include <sys/brt.h> 57 #include <sys/trace_zfs.h> 58 #include <sys/zfs_racct.h> 59 #include <sys/zfs_rlock.h> 60 #ifdef _KERNEL 61 #include <sys/vmsystm.h> 62 #include <sys/zfs_znode.h> 63 #endif 64 65 /* 66 * Enable/disable nopwrite feature. 67 */ 68 static int zfs_nopwrite_enabled = 1; 69 70 /* 71 * Tunable to control percentage of dirtied L1 blocks from frees allowed into 72 * one TXG. After this threshold is crossed, additional dirty blocks from frees 73 * will wait until the next TXG. 74 * A value of zero will disable this throttle. 75 */ 76 static uint_t zfs_per_txg_dirty_frees_percent = 30; 77 78 /* 79 * Enable/disable forcing txg sync when dirty checking for holes with lseek(). 80 * By default this is enabled to ensure accurate hole reporting, it can result 81 * in a significant performance penalty for lseek(SEEK_HOLE) heavy workloads. 82 * Disabling this option will result in holes never being reported in dirty 83 * files which is always safe. 84 */ 85 static int zfs_dmu_offset_next_sync = 1; 86 87 /* 88 * Limit the amount we can prefetch with one call to this amount. This 89 * helps to limit the amount of memory that can be used by prefetching. 90 * Larger objects should be prefetched a bit at a time. 91 */ 92 #ifdef _ILP32 93 uint_t dmu_prefetch_max = 8 * 1024 * 1024; 94 #else 95 uint_t dmu_prefetch_max = 8 * SPA_MAXBLOCKSIZE; 96 #endif 97 98 /* 99 * Override copies= for dedup state objects. 0 means the traditional behaviour 100 * (ie the default for the containing objset ie 3 for the MOS). 101 */ 102 uint_t dmu_ddt_copies = 0; 103 104 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = { 105 {DMU_BSWAP_UINT8, TRUE, FALSE, FALSE, "unallocated" }, 106 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "object directory" }, 107 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "object array" }, 108 {DMU_BSWAP_UINT8, TRUE, FALSE, FALSE, "packed nvlist" }, 109 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "packed nvlist size" }, 110 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "bpobj" }, 111 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "bpobj header" }, 112 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "SPA space map header" }, 113 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "SPA space map" }, 114 {DMU_BSWAP_UINT64, TRUE, FALSE, TRUE, "ZIL intent log" }, 115 {DMU_BSWAP_DNODE, TRUE, FALSE, TRUE, "DMU dnode" }, 116 {DMU_BSWAP_OBJSET, TRUE, TRUE, FALSE, "DMU objset" }, 117 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "DSL directory" }, 118 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL directory child map"}, 119 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL dataset snap map" }, 120 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL props" }, 121 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "DSL dataset" }, 122 {DMU_BSWAP_ZNODE, TRUE, FALSE, FALSE, "ZFS znode" }, 123 {DMU_BSWAP_OLDACL, TRUE, FALSE, TRUE, "ZFS V0 ACL" }, 124 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "ZFS plain file" }, 125 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS directory" }, 126 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "ZFS master node" }, 127 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS delete queue" }, 128 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "zvol object" }, 129 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "zvol prop" }, 130 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "other uint8[]" }, 131 {DMU_BSWAP_UINT64, FALSE, FALSE, TRUE, "other uint64[]" }, 132 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "other ZAP" }, 133 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "persistent error log" }, 134 {DMU_BSWAP_UINT8, TRUE, FALSE, FALSE, "SPA history" }, 135 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "SPA history offsets" }, 136 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "Pool properties" }, 137 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL permissions" }, 138 {DMU_BSWAP_ACL, TRUE, FALSE, TRUE, "ZFS ACL" }, 139 {DMU_BSWAP_UINT8, TRUE, FALSE, TRUE, "ZFS SYSACL" }, 140 {DMU_BSWAP_UINT8, TRUE, FALSE, TRUE, "FUID table" }, 141 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "FUID table size" }, 142 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL dataset next clones"}, 143 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "scan work queue" }, 144 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS user/group/project used" }, 145 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS user/group/project quota"}, 146 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "snapshot refcount tags"}, 147 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "DDT ZAP algorithm" }, 148 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "DDT statistics" }, 149 {DMU_BSWAP_UINT8, TRUE, FALSE, TRUE, "System attributes" }, 150 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "SA master node" }, 151 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "SA attr registration" }, 152 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "SA attr layouts" }, 153 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "scan translations" }, 154 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "deduplicated block" }, 155 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL deadlist map" }, 156 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "DSL deadlist map hdr" }, 157 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL dir clones" }, 158 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "bpobj subobj" } 159 }; 160 161 dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = { 162 { byteswap_uint8_array, "uint8" }, 163 { byteswap_uint16_array, "uint16" }, 164 { byteswap_uint32_array, "uint32" }, 165 { byteswap_uint64_array, "uint64" }, 166 { zap_byteswap, "zap" }, 167 { dnode_buf_byteswap, "dnode" }, 168 { dmu_objset_byteswap, "objset" }, 169 { zfs_znode_byteswap, "znode" }, 170 { zfs_oldacl_byteswap, "oldacl" }, 171 { zfs_acl_byteswap, "acl" } 172 }; 173 174 int 175 dmu_buf_hold_noread_by_dnode(dnode_t *dn, uint64_t offset, 176 const void *tag, dmu_buf_t **dbp) 177 { 178 uint64_t blkid; 179 dmu_buf_impl_t *db; 180 181 rw_enter(&dn->dn_struct_rwlock, RW_READER); 182 blkid = dbuf_whichblock(dn, 0, offset); 183 db = dbuf_hold(dn, blkid, tag); 184 rw_exit(&dn->dn_struct_rwlock); 185 186 if (db == NULL) { 187 *dbp = NULL; 188 return (SET_ERROR(EIO)); 189 } 190 191 *dbp = &db->db; 192 return (0); 193 } 194 195 int 196 dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset, 197 const void *tag, dmu_buf_t **dbp) 198 { 199 dnode_t *dn; 200 uint64_t blkid; 201 dmu_buf_impl_t *db; 202 int err; 203 204 err = dnode_hold(os, object, FTAG, &dn); 205 if (err) 206 return (err); 207 rw_enter(&dn->dn_struct_rwlock, RW_READER); 208 blkid = dbuf_whichblock(dn, 0, offset); 209 db = dbuf_hold(dn, blkid, tag); 210 rw_exit(&dn->dn_struct_rwlock); 211 dnode_rele(dn, FTAG); 212 213 if (db == NULL) { 214 *dbp = NULL; 215 return (SET_ERROR(EIO)); 216 } 217 218 *dbp = &db->db; 219 return (err); 220 } 221 222 int 223 dmu_buf_hold_by_dnode(dnode_t *dn, uint64_t offset, 224 const void *tag, dmu_buf_t **dbp, int flags) 225 { 226 int err; 227 int db_flags = DB_RF_CANFAIL; 228 229 if (flags & DMU_READ_NO_PREFETCH) 230 db_flags |= DB_RF_NOPREFETCH; 231 if (flags & DMU_READ_NO_DECRYPT) 232 db_flags |= DB_RF_NO_DECRYPT; 233 234 err = dmu_buf_hold_noread_by_dnode(dn, offset, tag, dbp); 235 if (err == 0) { 236 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp); 237 err = dbuf_read(db, NULL, db_flags); 238 if (err != 0) { 239 dbuf_rele(db, tag); 240 *dbp = NULL; 241 } 242 } 243 244 return (err); 245 } 246 247 int 248 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset, 249 const void *tag, dmu_buf_t **dbp, int flags) 250 { 251 int err; 252 int db_flags = DB_RF_CANFAIL; 253 254 if (flags & DMU_READ_NO_PREFETCH) 255 db_flags |= DB_RF_NOPREFETCH; 256 if (flags & DMU_READ_NO_DECRYPT) 257 db_flags |= DB_RF_NO_DECRYPT; 258 259 err = dmu_buf_hold_noread(os, object, offset, tag, dbp); 260 if (err == 0) { 261 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp); 262 err = dbuf_read(db, NULL, db_flags); 263 if (err != 0) { 264 dbuf_rele(db, tag); 265 *dbp = NULL; 266 } 267 } 268 269 return (err); 270 } 271 272 int 273 dmu_bonus_max(void) 274 { 275 return (DN_OLD_MAX_BONUSLEN); 276 } 277 278 int 279 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx) 280 { 281 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 282 dnode_t *dn; 283 int error; 284 285 if (newsize < 0 || newsize > db_fake->db_size) 286 return (SET_ERROR(EINVAL)); 287 288 DB_DNODE_ENTER(db); 289 dn = DB_DNODE(db); 290 291 if (dn->dn_bonus != db) { 292 error = SET_ERROR(EINVAL); 293 } else { 294 dnode_setbonuslen(dn, newsize, tx); 295 error = 0; 296 } 297 298 DB_DNODE_EXIT(db); 299 return (error); 300 } 301 302 int 303 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx) 304 { 305 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 306 dnode_t *dn; 307 int error; 308 309 if (!DMU_OT_IS_VALID(type)) 310 return (SET_ERROR(EINVAL)); 311 312 DB_DNODE_ENTER(db); 313 dn = DB_DNODE(db); 314 315 if (dn->dn_bonus != db) { 316 error = SET_ERROR(EINVAL); 317 } else { 318 dnode_setbonus_type(dn, type, tx); 319 error = 0; 320 } 321 322 DB_DNODE_EXIT(db); 323 return (error); 324 } 325 326 dmu_object_type_t 327 dmu_get_bonustype(dmu_buf_t *db_fake) 328 { 329 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 330 dmu_object_type_t type; 331 332 DB_DNODE_ENTER(db); 333 type = DB_DNODE(db)->dn_bonustype; 334 DB_DNODE_EXIT(db); 335 336 return (type); 337 } 338 339 int 340 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx) 341 { 342 dnode_t *dn; 343 int error; 344 345 error = dnode_hold(os, object, FTAG, &dn); 346 dbuf_rm_spill(dn, tx); 347 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 348 dnode_rm_spill(dn, tx); 349 rw_exit(&dn->dn_struct_rwlock); 350 dnode_rele(dn, FTAG); 351 return (error); 352 } 353 354 /* 355 * Lookup and hold the bonus buffer for the provided dnode. If the dnode 356 * has not yet been allocated a new bonus dbuf a will be allocated. 357 * Returns ENOENT, EIO, or 0. 358 */ 359 int dmu_bonus_hold_by_dnode(dnode_t *dn, const void *tag, dmu_buf_t **dbp, 360 uint32_t flags) 361 { 362 dmu_buf_impl_t *db; 363 int error; 364 uint32_t db_flags = DB_RF_MUST_SUCCEED; 365 366 if (flags & DMU_READ_NO_PREFETCH) 367 db_flags |= DB_RF_NOPREFETCH; 368 if (flags & DMU_READ_NO_DECRYPT) 369 db_flags |= DB_RF_NO_DECRYPT; 370 371 rw_enter(&dn->dn_struct_rwlock, RW_READER); 372 if (dn->dn_bonus == NULL) { 373 if (!rw_tryupgrade(&dn->dn_struct_rwlock)) { 374 rw_exit(&dn->dn_struct_rwlock); 375 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 376 } 377 if (dn->dn_bonus == NULL) 378 dbuf_create_bonus(dn); 379 } 380 db = dn->dn_bonus; 381 382 /* as long as the bonus buf is held, the dnode will be held */ 383 if (zfs_refcount_add(&db->db_holds, tag) == 1) { 384 VERIFY(dnode_add_ref(dn, db)); 385 atomic_inc_32(&dn->dn_dbufs_count); 386 } 387 388 /* 389 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's 390 * hold and incrementing the dbuf count to ensure that dnode_move() sees 391 * a dnode hold for every dbuf. 392 */ 393 rw_exit(&dn->dn_struct_rwlock); 394 395 error = dbuf_read(db, NULL, db_flags); 396 if (error) { 397 dnode_evict_bonus(dn); 398 dbuf_rele(db, tag); 399 *dbp = NULL; 400 return (error); 401 } 402 403 *dbp = &db->db; 404 return (0); 405 } 406 407 int 408 dmu_bonus_hold(objset_t *os, uint64_t object, const void *tag, dmu_buf_t **dbp) 409 { 410 dnode_t *dn; 411 int error; 412 413 error = dnode_hold(os, object, FTAG, &dn); 414 if (error) 415 return (error); 416 417 error = dmu_bonus_hold_by_dnode(dn, tag, dbp, DMU_READ_NO_PREFETCH); 418 dnode_rele(dn, FTAG); 419 420 return (error); 421 } 422 423 /* 424 * returns ENOENT, EIO, or 0. 425 * 426 * This interface will allocate a blank spill dbuf when a spill blk 427 * doesn't already exist on the dnode. 428 * 429 * if you only want to find an already existing spill db, then 430 * dmu_spill_hold_existing() should be used. 431 */ 432 int 433 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, const void *tag, 434 dmu_buf_t **dbp) 435 { 436 dmu_buf_impl_t *db = NULL; 437 int err; 438 439 if ((flags & DB_RF_HAVESTRUCT) == 0) 440 rw_enter(&dn->dn_struct_rwlock, RW_READER); 441 442 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag); 443 444 if ((flags & DB_RF_HAVESTRUCT) == 0) 445 rw_exit(&dn->dn_struct_rwlock); 446 447 if (db == NULL) { 448 *dbp = NULL; 449 return (SET_ERROR(EIO)); 450 } 451 err = dbuf_read(db, NULL, flags); 452 if (err == 0) 453 *dbp = &db->db; 454 else { 455 dbuf_rele(db, tag); 456 *dbp = NULL; 457 } 458 return (err); 459 } 460 461 int 462 dmu_spill_hold_existing(dmu_buf_t *bonus, const void *tag, dmu_buf_t **dbp) 463 { 464 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus; 465 dnode_t *dn; 466 int err; 467 468 DB_DNODE_ENTER(db); 469 dn = DB_DNODE(db); 470 471 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) { 472 err = SET_ERROR(EINVAL); 473 } else { 474 rw_enter(&dn->dn_struct_rwlock, RW_READER); 475 476 if (!dn->dn_have_spill) { 477 err = SET_ERROR(ENOENT); 478 } else { 479 err = dmu_spill_hold_by_dnode(dn, 480 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp); 481 } 482 483 rw_exit(&dn->dn_struct_rwlock); 484 } 485 486 DB_DNODE_EXIT(db); 487 return (err); 488 } 489 490 int 491 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, uint32_t flags, const void *tag, 492 dmu_buf_t **dbp) 493 { 494 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus; 495 int err; 496 uint32_t db_flags = DB_RF_CANFAIL; 497 498 if (flags & DMU_READ_NO_DECRYPT) 499 db_flags |= DB_RF_NO_DECRYPT; 500 501 DB_DNODE_ENTER(db); 502 err = dmu_spill_hold_by_dnode(DB_DNODE(db), db_flags, tag, dbp); 503 DB_DNODE_EXIT(db); 504 505 return (err); 506 } 507 508 /* 509 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces 510 * to take a held dnode rather than <os, object> -- the lookup is wasteful, 511 * and can induce severe lock contention when writing to several files 512 * whose dnodes are in the same block. 513 */ 514 int 515 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length, 516 boolean_t read, const void *tag, int *numbufsp, dmu_buf_t ***dbpp, 517 uint32_t flags) 518 { 519 dmu_buf_t **dbp; 520 zstream_t *zs = NULL; 521 uint64_t blkid, nblks, i; 522 uint32_t dbuf_flags; 523 int err; 524 zio_t *zio = NULL; 525 boolean_t missed = B_FALSE; 526 527 ASSERT(!read || length <= DMU_MAX_ACCESS); 528 529 /* 530 * Note: We directly notify the prefetch code of this read, so that 531 * we can tell it about the multi-block read. dbuf_read() only knows 532 * about the one block it is accessing. 533 */ 534 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT | 535 DB_RF_NOPREFETCH; 536 537 if ((flags & DMU_READ_NO_DECRYPT) != 0) 538 dbuf_flags |= DB_RF_NO_DECRYPT; 539 540 rw_enter(&dn->dn_struct_rwlock, RW_READER); 541 if (dn->dn_datablkshift) { 542 int blkshift = dn->dn_datablkshift; 543 nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) - 544 P2ALIGN_TYPED(offset, 1ULL << blkshift, uint64_t)) 545 >> blkshift; 546 } else { 547 if (offset + length > dn->dn_datablksz) { 548 zfs_panic_recover("zfs: accessing past end of object " 549 "%llx/%llx (size=%u access=%llu+%llu)", 550 (longlong_t)dn->dn_objset-> 551 os_dsl_dataset->ds_object, 552 (longlong_t)dn->dn_object, dn->dn_datablksz, 553 (longlong_t)offset, (longlong_t)length); 554 rw_exit(&dn->dn_struct_rwlock); 555 return (SET_ERROR(EIO)); 556 } 557 nblks = 1; 558 } 559 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP); 560 561 if (read) 562 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, 563 ZIO_FLAG_CANFAIL); 564 blkid = dbuf_whichblock(dn, 0, offset); 565 if ((flags & DMU_READ_NO_PREFETCH) == 0) { 566 /* 567 * Prepare the zfetch before initiating the demand reads, so 568 * that if multiple threads block on same indirect block, we 569 * base predictions on the original less racy request order. 570 */ 571 zs = dmu_zfetch_prepare(&dn->dn_zfetch, blkid, nblks, read, 572 B_TRUE); 573 } 574 for (i = 0; i < nblks; i++) { 575 dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag); 576 if (db == NULL) { 577 if (zs) { 578 dmu_zfetch_run(&dn->dn_zfetch, zs, missed, 579 B_TRUE); 580 } 581 rw_exit(&dn->dn_struct_rwlock); 582 dmu_buf_rele_array(dbp, nblks, tag); 583 if (read) 584 zio_nowait(zio); 585 return (SET_ERROR(EIO)); 586 } 587 588 /* 589 * Initiate async demand data read. 590 * We check the db_state after calling dbuf_read() because 591 * (1) dbuf_read() may change the state to CACHED due to a 592 * hit in the ARC, and (2) on a cache miss, a child will 593 * have been added to "zio" but not yet completed, so the 594 * state will not yet be CACHED. 595 */ 596 if (read) { 597 if (i == nblks - 1 && blkid + i < dn->dn_maxblkid && 598 offset + length < db->db.db_offset + 599 db->db.db_size) { 600 if (offset <= db->db.db_offset) 601 dbuf_flags |= DB_RF_PARTIAL_FIRST; 602 else 603 dbuf_flags |= DB_RF_PARTIAL_MORE; 604 } 605 (void) dbuf_read(db, zio, dbuf_flags); 606 if (db->db_state != DB_CACHED) 607 missed = B_TRUE; 608 } 609 dbp[i] = &db->db; 610 } 611 612 /* 613 * If we are doing O_DIRECT we still hold the dbufs, even for reads, 614 * but we do not issue any reads here. We do not want to account for 615 * writes in this case. 616 * 617 * O_DIRECT write/read accounting takes place in 618 * dmu_{write/read}_abd(). 619 */ 620 if (!read && ((flags & DMU_DIRECTIO) == 0)) 621 zfs_racct_write(dn->dn_objset->os_spa, length, nblks, flags); 622 623 if (zs) 624 dmu_zfetch_run(&dn->dn_zfetch, zs, missed, B_TRUE); 625 rw_exit(&dn->dn_struct_rwlock); 626 627 if (read) { 628 /* wait for async read i/o */ 629 err = zio_wait(zio); 630 if (err) { 631 dmu_buf_rele_array(dbp, nblks, tag); 632 return (err); 633 } 634 635 /* wait for other io to complete */ 636 for (i = 0; i < nblks; i++) { 637 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i]; 638 mutex_enter(&db->db_mtx); 639 while (db->db_state == DB_READ || 640 db->db_state == DB_FILL) 641 cv_wait(&db->db_changed, &db->db_mtx); 642 if (db->db_state == DB_UNCACHED) 643 err = SET_ERROR(EIO); 644 mutex_exit(&db->db_mtx); 645 if (err) { 646 dmu_buf_rele_array(dbp, nblks, tag); 647 return (err); 648 } 649 } 650 } 651 652 *numbufsp = nblks; 653 *dbpp = dbp; 654 return (0); 655 } 656 657 int 658 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset, 659 uint64_t length, int read, const void *tag, int *numbufsp, 660 dmu_buf_t ***dbpp) 661 { 662 dnode_t *dn; 663 int err; 664 665 err = dnode_hold(os, object, FTAG, &dn); 666 if (err) 667 return (err); 668 669 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag, 670 numbufsp, dbpp, DMU_READ_PREFETCH); 671 672 dnode_rele(dn, FTAG); 673 674 return (err); 675 } 676 677 int 678 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset, 679 uint64_t length, boolean_t read, const void *tag, int *numbufsp, 680 dmu_buf_t ***dbpp) 681 { 682 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 683 int err; 684 685 DB_DNODE_ENTER(db); 686 err = dmu_buf_hold_array_by_dnode(DB_DNODE(db), offset, length, read, 687 tag, numbufsp, dbpp, DMU_READ_PREFETCH); 688 DB_DNODE_EXIT(db); 689 690 return (err); 691 } 692 693 void 694 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, const void *tag) 695 { 696 int i; 697 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake; 698 699 if (numbufs == 0) 700 return; 701 702 for (i = 0; i < numbufs; i++) { 703 if (dbp[i]) 704 dbuf_rele(dbp[i], tag); 705 } 706 707 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs); 708 } 709 710 /* 711 * Issue prefetch I/Os for the given blocks. If level is greater than 0, the 712 * indirect blocks prefetched will be those that point to the blocks containing 713 * the data starting at offset, and continuing to offset + len. If the range 714 * is too long, prefetch the first dmu_prefetch_max bytes as requested, while 715 * for the rest only a higher level, also fitting within dmu_prefetch_max. It 716 * should primarily help random reads, since for long sequential reads there is 717 * a speculative prefetcher. 718 * 719 * Note that if the indirect blocks above the blocks being prefetched are not 720 * in cache, they will be asynchronously read in. Dnode read by dnode_hold() 721 * is currently synchronous. 722 */ 723 void 724 dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset, 725 uint64_t len, zio_priority_t pri) 726 { 727 dnode_t *dn; 728 729 if (dmu_prefetch_max == 0 || len == 0) { 730 dmu_prefetch_dnode(os, object, pri); 731 return; 732 } 733 734 if (dnode_hold(os, object, FTAG, &dn) != 0) 735 return; 736 737 dmu_prefetch_by_dnode(dn, level, offset, len, pri); 738 739 dnode_rele(dn, FTAG); 740 } 741 742 void 743 dmu_prefetch_by_dnode(dnode_t *dn, int64_t level, uint64_t offset, 744 uint64_t len, zio_priority_t pri) 745 { 746 int64_t level2 = level; 747 uint64_t start, end, start2, end2; 748 749 /* 750 * Depending on len we may do two prefetches: blocks [start, end) at 751 * level, and following blocks [start2, end2) at higher level2. 752 */ 753 rw_enter(&dn->dn_struct_rwlock, RW_READER); 754 if (dn->dn_datablkshift != 0) { 755 /* 756 * The object has multiple blocks. Calculate the full range 757 * of blocks [start, end2) and then split it into two parts, 758 * so that the first [start, end) fits into dmu_prefetch_max. 759 */ 760 start = dbuf_whichblock(dn, level, offset); 761 end2 = dbuf_whichblock(dn, level, offset + len - 1) + 1; 762 uint8_t ibs = dn->dn_indblkshift; 763 uint8_t bs = (level == 0) ? dn->dn_datablkshift : ibs; 764 uint_t limit = P2ROUNDUP(dmu_prefetch_max, 1 << bs) >> bs; 765 start2 = end = MIN(end2, start + limit); 766 767 /* 768 * Find level2 where [start2, end2) fits into dmu_prefetch_max. 769 */ 770 uint8_t ibps = ibs - SPA_BLKPTRSHIFT; 771 limit = P2ROUNDUP(dmu_prefetch_max, 1 << ibs) >> ibs; 772 do { 773 level2++; 774 start2 = P2ROUNDUP(start2, 1 << ibps) >> ibps; 775 end2 = P2ROUNDUP(end2, 1 << ibps) >> ibps; 776 } while (end2 - start2 > limit); 777 } else { 778 /* There is only one block. Prefetch it or nothing. */ 779 start = start2 = end2 = 0; 780 end = start + (level == 0 && offset < dn->dn_datablksz); 781 } 782 783 for (uint64_t i = start; i < end; i++) 784 dbuf_prefetch(dn, level, i, pri, 0); 785 for (uint64_t i = start2; i < end2; i++) 786 dbuf_prefetch(dn, level2, i, pri, 0); 787 rw_exit(&dn->dn_struct_rwlock); 788 } 789 790 typedef struct { 791 kmutex_t dpa_lock; 792 kcondvar_t dpa_cv; 793 uint64_t dpa_pending_io; 794 } dmu_prefetch_arg_t; 795 796 static void 797 dmu_prefetch_done(void *arg, uint64_t level, uint64_t blkid, boolean_t issued) 798 { 799 (void) level; (void) blkid; (void)issued; 800 dmu_prefetch_arg_t *dpa = arg; 801 802 ASSERT0(level); 803 804 mutex_enter(&dpa->dpa_lock); 805 ASSERT3U(dpa->dpa_pending_io, >, 0); 806 if (--dpa->dpa_pending_io == 0) 807 cv_broadcast(&dpa->dpa_cv); 808 mutex_exit(&dpa->dpa_lock); 809 } 810 811 static void 812 dmu_prefetch_wait_by_dnode(dnode_t *dn, uint64_t offset, uint64_t len) 813 { 814 dmu_prefetch_arg_t dpa; 815 816 mutex_init(&dpa.dpa_lock, NULL, MUTEX_DEFAULT, NULL); 817 cv_init(&dpa.dpa_cv, NULL, CV_DEFAULT, NULL); 818 819 rw_enter(&dn->dn_struct_rwlock, RW_READER); 820 821 uint64_t start = dbuf_whichblock(dn, 0, offset); 822 uint64_t end = dbuf_whichblock(dn, 0, offset + len - 1) + 1; 823 dpa.dpa_pending_io = end - start; 824 825 for (uint64_t blk = start; blk < end; blk++) { 826 (void) dbuf_prefetch_impl(dn, 0, blk, ZIO_PRIORITY_ASYNC_READ, 827 0, dmu_prefetch_done, &dpa); 828 } 829 830 rw_exit(&dn->dn_struct_rwlock); 831 832 /* wait for prefetch L0 reads to finish */ 833 mutex_enter(&dpa.dpa_lock); 834 while (dpa.dpa_pending_io > 0) { 835 cv_wait(&dpa.dpa_cv, &dpa.dpa_lock); 836 837 } 838 mutex_exit(&dpa.dpa_lock); 839 840 mutex_destroy(&dpa.dpa_lock); 841 cv_destroy(&dpa.dpa_cv); 842 } 843 844 /* 845 * Issue prefetch I/Os for the given L0 block range and wait for the I/O 846 * to complete. This does not enforce dmu_prefetch_max and will prefetch 847 * the entire range. The blocks are read from disk into the ARC but no 848 * decompression occurs (i.e., the dbuf cache is not required). 849 */ 850 int 851 dmu_prefetch_wait(objset_t *os, uint64_t object, uint64_t offset, uint64_t size) 852 { 853 dnode_t *dn; 854 int err = 0; 855 856 err = dnode_hold(os, object, FTAG, &dn); 857 if (err != 0) 858 return (err); 859 860 /* 861 * Chunk the requests (16 indirects worth) so that we can be interrupted 862 */ 863 uint64_t chunksize; 864 if (dn->dn_indblkshift) { 865 uint64_t nbps = bp_span_in_blocks(dn->dn_indblkshift, 1); 866 chunksize = (nbps * 16) << dn->dn_datablkshift; 867 } else { 868 chunksize = dn->dn_datablksz; 869 } 870 871 while (size > 0) { 872 uint64_t mylen = MIN(size, chunksize); 873 874 dmu_prefetch_wait_by_dnode(dn, offset, mylen); 875 876 offset += mylen; 877 size -= mylen; 878 879 if (issig()) { 880 err = SET_ERROR(EINTR); 881 break; 882 } 883 } 884 885 dnode_rele(dn, FTAG); 886 887 return (err); 888 } 889 890 /* 891 * Issue prefetch I/Os for the given object's dnode. 892 */ 893 void 894 dmu_prefetch_dnode(objset_t *os, uint64_t object, zio_priority_t pri) 895 { 896 if (object == 0 || object >= DN_MAX_OBJECT) 897 return; 898 899 dnode_t *dn = DMU_META_DNODE(os); 900 rw_enter(&dn->dn_struct_rwlock, RW_READER); 901 uint64_t blkid = dbuf_whichblock(dn, 0, object * sizeof (dnode_phys_t)); 902 dbuf_prefetch(dn, 0, blkid, pri, 0); 903 rw_exit(&dn->dn_struct_rwlock); 904 } 905 906 /* 907 * Get the next "chunk" of file data to free. We traverse the file from 908 * the end so that the file gets shorter over time (if we crash in the 909 * middle, this will leave us in a better state). We find allocated file 910 * data by simply searching the allocated level 1 indirects. 911 * 912 * On input, *start should be the first offset that does not need to be 913 * freed (e.g. "offset + length"). On return, *start will be the first 914 * offset that should be freed and l1blks is set to the number of level 1 915 * indirect blocks found within the chunk. 916 */ 917 static int 918 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum, uint64_t *l1blks) 919 { 920 uint64_t blks; 921 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1); 922 /* bytes of data covered by a level-1 indirect block */ 923 uint64_t iblkrange = (uint64_t)dn->dn_datablksz * 924 EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT); 925 926 ASSERT3U(minimum, <=, *start); 927 928 /* dn_nlevels == 1 means we don't have any L1 blocks */ 929 if (dn->dn_nlevels <= 1) { 930 *l1blks = 0; 931 *start = minimum; 932 return (0); 933 } 934 935 /* 936 * Check if we can free the entire range assuming that all of the 937 * L1 blocks in this range have data. If we can, we use this 938 * worst case value as an estimate so we can avoid having to look 939 * at the object's actual data. 940 */ 941 uint64_t total_l1blks = 942 (roundup(*start, iblkrange) - (minimum / iblkrange * iblkrange)) / 943 iblkrange; 944 if (total_l1blks <= maxblks) { 945 *l1blks = total_l1blks; 946 *start = minimum; 947 return (0); 948 } 949 ASSERT(ISP2(iblkrange)); 950 951 for (blks = 0; *start > minimum && blks < maxblks; blks++) { 952 int err; 953 954 /* 955 * dnode_next_offset(BACKWARDS) will find an allocated L1 956 * indirect block at or before the input offset. We must 957 * decrement *start so that it is at the end of the region 958 * to search. 959 */ 960 (*start)--; 961 962 err = dnode_next_offset(dn, 963 DNODE_FIND_BACKWARDS, start, 2, 1, 0); 964 965 /* if there are no indirect blocks before start, we are done */ 966 if (err == ESRCH) { 967 *start = minimum; 968 break; 969 } else if (err != 0) { 970 *l1blks = blks; 971 return (err); 972 } 973 974 /* set start to the beginning of this L1 indirect */ 975 *start = P2ALIGN_TYPED(*start, iblkrange, uint64_t); 976 } 977 if (*start < minimum) 978 *start = minimum; 979 *l1blks = blks; 980 981 return (0); 982 } 983 984 /* 985 * If this objset is of type OST_ZFS return true if vfs's unmounted flag is set, 986 * otherwise return false. 987 * Used below in dmu_free_long_range_impl() to enable abort when unmounting 988 */ 989 static boolean_t 990 dmu_objset_zfs_unmounting(objset_t *os) 991 { 992 #ifdef _KERNEL 993 if (dmu_objset_type(os) == DMU_OST_ZFS) 994 return (zfs_get_vfs_flag_unmounted(os)); 995 #else 996 (void) os; 997 #endif 998 return (B_FALSE); 999 } 1000 1001 static int 1002 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset, 1003 uint64_t length) 1004 { 1005 uint64_t object_size; 1006 int err; 1007 uint64_t dirty_frees_threshold; 1008 dsl_pool_t *dp = dmu_objset_pool(os); 1009 1010 if (dn == NULL) 1011 return (SET_ERROR(EINVAL)); 1012 1013 object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz; 1014 if (offset >= object_size) 1015 return (0); 1016 1017 if (zfs_per_txg_dirty_frees_percent <= 100) 1018 dirty_frees_threshold = 1019 zfs_per_txg_dirty_frees_percent * zfs_dirty_data_max / 100; 1020 else 1021 dirty_frees_threshold = zfs_dirty_data_max / 20; 1022 1023 if (length == DMU_OBJECT_END || offset + length > object_size) 1024 length = object_size - offset; 1025 1026 while (length != 0) { 1027 uint64_t chunk_end, chunk_begin, chunk_len; 1028 uint64_t l1blks; 1029 dmu_tx_t *tx; 1030 1031 if (dmu_objset_zfs_unmounting(dn->dn_objset)) 1032 return (SET_ERROR(EINTR)); 1033 1034 chunk_end = chunk_begin = offset + length; 1035 1036 /* move chunk_begin backwards to the beginning of this chunk */ 1037 err = get_next_chunk(dn, &chunk_begin, offset, &l1blks); 1038 if (err) 1039 return (err); 1040 ASSERT3U(chunk_begin, >=, offset); 1041 ASSERT3U(chunk_begin, <=, chunk_end); 1042 1043 chunk_len = chunk_end - chunk_begin; 1044 1045 tx = dmu_tx_create(os); 1046 dmu_tx_hold_free(tx, dn->dn_object, chunk_begin, chunk_len); 1047 1048 /* 1049 * Mark this transaction as typically resulting in a net 1050 * reduction in space used. 1051 */ 1052 dmu_tx_mark_netfree(tx); 1053 err = dmu_tx_assign(tx, TXG_WAIT); 1054 if (err) { 1055 dmu_tx_abort(tx); 1056 return (err); 1057 } 1058 1059 uint64_t txg = dmu_tx_get_txg(tx); 1060 1061 mutex_enter(&dp->dp_lock); 1062 uint64_t long_free_dirty = 1063 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK]; 1064 mutex_exit(&dp->dp_lock); 1065 1066 /* 1067 * To avoid filling up a TXG with just frees, wait for 1068 * the next TXG to open before freeing more chunks if 1069 * we have reached the threshold of frees. 1070 */ 1071 if (dirty_frees_threshold != 0 && 1072 long_free_dirty >= dirty_frees_threshold) { 1073 DMU_TX_STAT_BUMP(dmu_tx_dirty_frees_delay); 1074 dmu_tx_commit(tx); 1075 txg_wait_open(dp, 0, B_TRUE); 1076 continue; 1077 } 1078 1079 /* 1080 * In order to prevent unnecessary write throttling, for each 1081 * TXG, we track the cumulative size of L1 blocks being dirtied 1082 * in dnode_free_range() below. We compare this number to a 1083 * tunable threshold, past which we prevent new L1 dirty freeing 1084 * blocks from being added into the open TXG. See 1085 * dmu_free_long_range_impl() for details. The threshold 1086 * prevents write throttle activation due to dirty freeing L1 1087 * blocks taking up a large percentage of zfs_dirty_data_max. 1088 */ 1089 mutex_enter(&dp->dp_lock); 1090 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] += 1091 l1blks << dn->dn_indblkshift; 1092 mutex_exit(&dp->dp_lock); 1093 DTRACE_PROBE3(free__long__range, 1094 uint64_t, long_free_dirty, uint64_t, chunk_len, 1095 uint64_t, txg); 1096 dnode_free_range(dn, chunk_begin, chunk_len, tx); 1097 1098 dmu_tx_commit(tx); 1099 1100 length -= chunk_len; 1101 } 1102 return (0); 1103 } 1104 1105 int 1106 dmu_free_long_range(objset_t *os, uint64_t object, 1107 uint64_t offset, uint64_t length) 1108 { 1109 dnode_t *dn; 1110 int err; 1111 1112 err = dnode_hold(os, object, FTAG, &dn); 1113 if (err != 0) 1114 return (err); 1115 err = dmu_free_long_range_impl(os, dn, offset, length); 1116 1117 /* 1118 * It is important to zero out the maxblkid when freeing the entire 1119 * file, so that (a) subsequent calls to dmu_free_long_range_impl() 1120 * will take the fast path, and (b) dnode_reallocate() can verify 1121 * that the entire file has been freed. 1122 */ 1123 if (err == 0 && offset == 0 && length == DMU_OBJECT_END) 1124 dn->dn_maxblkid = 0; 1125 1126 dnode_rele(dn, FTAG); 1127 return (err); 1128 } 1129 1130 int 1131 dmu_free_long_object(objset_t *os, uint64_t object) 1132 { 1133 dmu_tx_t *tx; 1134 int err; 1135 1136 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END); 1137 if (err != 0) 1138 return (err); 1139 1140 tx = dmu_tx_create(os); 1141 dmu_tx_hold_bonus(tx, object); 1142 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END); 1143 dmu_tx_mark_netfree(tx); 1144 err = dmu_tx_assign(tx, TXG_WAIT); 1145 if (err == 0) { 1146 err = dmu_object_free(os, object, tx); 1147 dmu_tx_commit(tx); 1148 } else { 1149 dmu_tx_abort(tx); 1150 } 1151 1152 return (err); 1153 } 1154 1155 int 1156 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset, 1157 uint64_t size, dmu_tx_t *tx) 1158 { 1159 dnode_t *dn; 1160 int err = dnode_hold(os, object, FTAG, &dn); 1161 if (err) 1162 return (err); 1163 ASSERT(offset < UINT64_MAX); 1164 ASSERT(size == DMU_OBJECT_END || size <= UINT64_MAX - offset); 1165 dnode_free_range(dn, offset, size, tx); 1166 dnode_rele(dn, FTAG); 1167 return (0); 1168 } 1169 1170 static int 1171 dmu_read_impl(dnode_t *dn, uint64_t offset, uint64_t size, 1172 void *buf, uint32_t flags) 1173 { 1174 dmu_buf_t **dbp; 1175 int numbufs, err = 0; 1176 1177 /* 1178 * Deal with odd block sizes, where there can't be data past the first 1179 * block. If we ever do the tail block optimization, we will need to 1180 * handle that here as well. 1181 */ 1182 if (dn->dn_maxblkid == 0) { 1183 uint64_t newsz = offset > dn->dn_datablksz ? 0 : 1184 MIN(size, dn->dn_datablksz - offset); 1185 memset((char *)buf + newsz, 0, size - newsz); 1186 size = newsz; 1187 } 1188 1189 if (size == 0) 1190 return (0); 1191 1192 /* Allow Direct I/O when requested and properly aligned */ 1193 if ((flags & DMU_DIRECTIO) && zfs_dio_page_aligned(buf) && 1194 zfs_dio_aligned(offset, size, PAGESIZE)) { 1195 abd_t *data = abd_get_from_buf(buf, size); 1196 err = dmu_read_abd(dn, offset, size, data, flags); 1197 abd_free(data); 1198 return (err); 1199 } 1200 1201 while (size > 0) { 1202 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2); 1203 int i; 1204 1205 /* 1206 * NB: we could do this block-at-a-time, but it's nice 1207 * to be reading in parallel. 1208 */ 1209 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen, 1210 TRUE, FTAG, &numbufs, &dbp, flags); 1211 if (err) 1212 break; 1213 1214 for (i = 0; i < numbufs; i++) { 1215 uint64_t tocpy; 1216 int64_t bufoff; 1217 dmu_buf_t *db = dbp[i]; 1218 1219 ASSERT(size > 0); 1220 1221 bufoff = offset - db->db_offset; 1222 tocpy = MIN(db->db_size - bufoff, size); 1223 1224 (void) memcpy(buf, (char *)db->db_data + bufoff, tocpy); 1225 1226 offset += tocpy; 1227 size -= tocpy; 1228 buf = (char *)buf + tocpy; 1229 } 1230 dmu_buf_rele_array(dbp, numbufs, FTAG); 1231 } 1232 return (err); 1233 } 1234 1235 int 1236 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1237 void *buf, uint32_t flags) 1238 { 1239 dnode_t *dn; 1240 int err; 1241 1242 err = dnode_hold(os, object, FTAG, &dn); 1243 if (err != 0) 1244 return (err); 1245 1246 err = dmu_read_impl(dn, offset, size, buf, flags); 1247 dnode_rele(dn, FTAG); 1248 return (err); 1249 } 1250 1251 int 1252 dmu_read_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, void *buf, 1253 uint32_t flags) 1254 { 1255 return (dmu_read_impl(dn, offset, size, buf, flags)); 1256 } 1257 1258 static void 1259 dmu_write_impl(dmu_buf_t **dbp, int numbufs, uint64_t offset, uint64_t size, 1260 const void *buf, dmu_tx_t *tx) 1261 { 1262 int i; 1263 1264 for (i = 0; i < numbufs; i++) { 1265 uint64_t tocpy; 1266 int64_t bufoff; 1267 dmu_buf_t *db = dbp[i]; 1268 1269 ASSERT(size > 0); 1270 1271 bufoff = offset - db->db_offset; 1272 tocpy = MIN(db->db_size - bufoff, size); 1273 1274 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 1275 1276 if (tocpy == db->db_size) 1277 dmu_buf_will_fill(db, tx, B_FALSE); 1278 else 1279 dmu_buf_will_dirty(db, tx); 1280 1281 (void) memcpy((char *)db->db_data + bufoff, buf, tocpy); 1282 1283 if (tocpy == db->db_size) 1284 dmu_buf_fill_done(db, tx, B_FALSE); 1285 1286 offset += tocpy; 1287 size -= tocpy; 1288 buf = (char *)buf + tocpy; 1289 } 1290 } 1291 1292 void 1293 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1294 const void *buf, dmu_tx_t *tx) 1295 { 1296 dmu_buf_t **dbp; 1297 int numbufs; 1298 1299 if (size == 0) 1300 return; 1301 1302 VERIFY0(dmu_buf_hold_array(os, object, offset, size, 1303 FALSE, FTAG, &numbufs, &dbp)); 1304 dmu_write_impl(dbp, numbufs, offset, size, buf, tx); 1305 dmu_buf_rele_array(dbp, numbufs, FTAG); 1306 } 1307 1308 /* 1309 * This interface is not used internally by ZFS but is provided for 1310 * use by Lustre which is built on the DMU interfaces. 1311 */ 1312 int 1313 dmu_write_by_dnode_flags(dnode_t *dn, uint64_t offset, uint64_t size, 1314 const void *buf, dmu_tx_t *tx, uint32_t flags) 1315 { 1316 dmu_buf_t **dbp; 1317 int numbufs; 1318 int error; 1319 1320 if (size == 0) 1321 return (0); 1322 1323 /* Allow Direct I/O when requested and properly aligned */ 1324 if ((flags & DMU_DIRECTIO) && zfs_dio_page_aligned((void *)buf) && 1325 zfs_dio_aligned(offset, size, dn->dn_datablksz)) { 1326 abd_t *data = abd_get_from_buf((void *)buf, size); 1327 error = dmu_write_abd(dn, offset, size, data, DMU_DIRECTIO, tx); 1328 abd_free(data); 1329 return (error); 1330 } 1331 1332 VERIFY0(dmu_buf_hold_array_by_dnode(dn, offset, size, 1333 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH)); 1334 dmu_write_impl(dbp, numbufs, offset, size, buf, tx); 1335 dmu_buf_rele_array(dbp, numbufs, FTAG); 1336 return (0); 1337 } 1338 1339 int 1340 dmu_write_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, 1341 const void *buf, dmu_tx_t *tx) 1342 { 1343 return (dmu_write_by_dnode_flags(dn, offset, size, buf, tx, 0)); 1344 } 1345 1346 void 1347 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1348 dmu_tx_t *tx) 1349 { 1350 dmu_buf_t **dbp; 1351 int numbufs, i; 1352 1353 if (size == 0) 1354 return; 1355 1356 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size, 1357 FALSE, FTAG, &numbufs, &dbp)); 1358 1359 for (i = 0; i < numbufs; i++) { 1360 dmu_buf_t *db = dbp[i]; 1361 1362 dmu_buf_will_not_fill(db, tx); 1363 } 1364 dmu_buf_rele_array(dbp, numbufs, FTAG); 1365 } 1366 1367 void 1368 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset, 1369 void *data, uint8_t etype, uint8_t comp, int uncompressed_size, 1370 int compressed_size, int byteorder, dmu_tx_t *tx) 1371 { 1372 dmu_buf_t *db; 1373 1374 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES); 1375 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS); 1376 VERIFY0(dmu_buf_hold_noread(os, object, offset, 1377 FTAG, &db)); 1378 1379 dmu_buf_write_embedded(db, 1380 data, (bp_embedded_type_t)etype, (enum zio_compress)comp, 1381 uncompressed_size, compressed_size, byteorder, tx); 1382 1383 dmu_buf_rele(db, FTAG); 1384 } 1385 1386 void 1387 dmu_redact(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1388 dmu_tx_t *tx) 1389 { 1390 int numbufs, i; 1391 dmu_buf_t **dbp; 1392 1393 VERIFY0(dmu_buf_hold_array(os, object, offset, size, FALSE, FTAG, 1394 &numbufs, &dbp)); 1395 for (i = 0; i < numbufs; i++) 1396 dmu_buf_redact(dbp[i], tx); 1397 dmu_buf_rele_array(dbp, numbufs, FTAG); 1398 } 1399 1400 #ifdef _KERNEL 1401 int 1402 dmu_read_uio_dnode(dnode_t *dn, zfs_uio_t *uio, uint64_t size) 1403 { 1404 dmu_buf_t **dbp; 1405 int numbufs, i, err; 1406 1407 if (uio->uio_extflg & UIO_DIRECT) 1408 return (dmu_read_uio_direct(dn, uio, size)); 1409 1410 /* 1411 * NB: we could do this block-at-a-time, but it's nice 1412 * to be reading in parallel. 1413 */ 1414 err = dmu_buf_hold_array_by_dnode(dn, zfs_uio_offset(uio), size, 1415 TRUE, FTAG, &numbufs, &dbp, 0); 1416 if (err) 1417 return (err); 1418 1419 for (i = 0; i < numbufs; i++) { 1420 uint64_t tocpy; 1421 int64_t bufoff; 1422 dmu_buf_t *db = dbp[i]; 1423 1424 ASSERT(size > 0); 1425 1426 bufoff = zfs_uio_offset(uio) - db->db_offset; 1427 tocpy = MIN(db->db_size - bufoff, size); 1428 1429 err = zfs_uio_fault_move((char *)db->db_data + bufoff, tocpy, 1430 UIO_READ, uio); 1431 1432 if (err) 1433 break; 1434 1435 size -= tocpy; 1436 } 1437 dmu_buf_rele_array(dbp, numbufs, FTAG); 1438 1439 return (err); 1440 } 1441 1442 /* 1443 * Read 'size' bytes into the uio buffer. 1444 * From object zdb->db_object. 1445 * Starting at zfs_uio_offset(uio). 1446 * 1447 * If the caller already has a dbuf in the target object 1448 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(), 1449 * because we don't have to find the dnode_t for the object. 1450 */ 1451 int 1452 dmu_read_uio_dbuf(dmu_buf_t *zdb, zfs_uio_t *uio, uint64_t size) 1453 { 1454 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb; 1455 int err; 1456 1457 if (size == 0) 1458 return (0); 1459 1460 DB_DNODE_ENTER(db); 1461 err = dmu_read_uio_dnode(DB_DNODE(db), uio, size); 1462 DB_DNODE_EXIT(db); 1463 1464 return (err); 1465 } 1466 1467 /* 1468 * Read 'size' bytes into the uio buffer. 1469 * From the specified object 1470 * Starting at offset zfs_uio_offset(uio). 1471 */ 1472 int 1473 dmu_read_uio(objset_t *os, uint64_t object, zfs_uio_t *uio, uint64_t size) 1474 { 1475 dnode_t *dn; 1476 int err; 1477 1478 if (size == 0) 1479 return (0); 1480 1481 err = dnode_hold(os, object, FTAG, &dn); 1482 if (err) 1483 return (err); 1484 1485 err = dmu_read_uio_dnode(dn, uio, size); 1486 1487 dnode_rele(dn, FTAG); 1488 1489 return (err); 1490 } 1491 1492 int 1493 dmu_write_uio_dnode(dnode_t *dn, zfs_uio_t *uio, uint64_t size, dmu_tx_t *tx) 1494 { 1495 dmu_buf_t **dbp; 1496 int numbufs; 1497 int err = 0; 1498 uint64_t write_size; 1499 1500 top: 1501 write_size = size; 1502 1503 /* 1504 * We only allow Direct I/O writes to happen if we are block 1505 * sized aligned. Otherwise, we pass the write off to the ARC. 1506 */ 1507 if ((uio->uio_extflg & UIO_DIRECT) && 1508 (write_size >= dn->dn_datablksz)) { 1509 if (zfs_dio_aligned(zfs_uio_offset(uio), write_size, 1510 dn->dn_datablksz)) { 1511 return (dmu_write_uio_direct(dn, uio, size, tx)); 1512 } else if (write_size > dn->dn_datablksz && 1513 zfs_dio_offset_aligned(zfs_uio_offset(uio), 1514 dn->dn_datablksz)) { 1515 write_size = 1516 dn->dn_datablksz * (write_size / dn->dn_datablksz); 1517 err = dmu_write_uio_direct(dn, uio, write_size, tx); 1518 if (err == 0) { 1519 size -= write_size; 1520 goto top; 1521 } else { 1522 return (err); 1523 } 1524 } else { 1525 write_size = 1526 P2PHASE(zfs_uio_offset(uio), dn->dn_datablksz); 1527 } 1528 } 1529 1530 err = dmu_buf_hold_array_by_dnode(dn, zfs_uio_offset(uio), write_size, 1531 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH); 1532 if (err) 1533 return (err); 1534 1535 for (int i = 0; i < numbufs; i++) { 1536 uint64_t tocpy; 1537 int64_t bufoff; 1538 dmu_buf_t *db = dbp[i]; 1539 1540 ASSERT(write_size > 0); 1541 1542 offset_t off = zfs_uio_offset(uio); 1543 bufoff = off - db->db_offset; 1544 tocpy = MIN(db->db_size - bufoff, write_size); 1545 1546 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 1547 1548 if (tocpy == db->db_size) 1549 dmu_buf_will_fill(db, tx, B_TRUE); 1550 else 1551 dmu_buf_will_dirty(db, tx); 1552 1553 err = zfs_uio_fault_move((char *)db->db_data + bufoff, 1554 tocpy, UIO_WRITE, uio); 1555 1556 if (tocpy == db->db_size && dmu_buf_fill_done(db, tx, err)) { 1557 /* The fill was reverted. Undo any uio progress. */ 1558 zfs_uio_advance(uio, off - zfs_uio_offset(uio)); 1559 } 1560 1561 if (err) 1562 break; 1563 1564 write_size -= tocpy; 1565 size -= tocpy; 1566 } 1567 1568 IMPLY(err == 0, write_size == 0); 1569 1570 dmu_buf_rele_array(dbp, numbufs, FTAG); 1571 1572 if ((uio->uio_extflg & UIO_DIRECT) && size > 0) { 1573 goto top; 1574 } 1575 1576 return (err); 1577 } 1578 1579 /* 1580 * Write 'size' bytes from the uio buffer. 1581 * To object zdb->db_object. 1582 * Starting at offset zfs_uio_offset(uio). 1583 * 1584 * If the caller already has a dbuf in the target object 1585 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(), 1586 * because we don't have to find the dnode_t for the object. 1587 */ 1588 int 1589 dmu_write_uio_dbuf(dmu_buf_t *zdb, zfs_uio_t *uio, uint64_t size, 1590 dmu_tx_t *tx) 1591 { 1592 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb; 1593 int err; 1594 1595 if (size == 0) 1596 return (0); 1597 1598 DB_DNODE_ENTER(db); 1599 err = dmu_write_uio_dnode(DB_DNODE(db), uio, size, tx); 1600 DB_DNODE_EXIT(db); 1601 1602 return (err); 1603 } 1604 1605 /* 1606 * Write 'size' bytes from the uio buffer. 1607 * To the specified object. 1608 * Starting at offset zfs_uio_offset(uio). 1609 */ 1610 int 1611 dmu_write_uio(objset_t *os, uint64_t object, zfs_uio_t *uio, uint64_t size, 1612 dmu_tx_t *tx) 1613 { 1614 dnode_t *dn; 1615 int err; 1616 1617 if (size == 0) 1618 return (0); 1619 1620 err = dnode_hold(os, object, FTAG, &dn); 1621 if (err) 1622 return (err); 1623 1624 err = dmu_write_uio_dnode(dn, uio, size, tx); 1625 1626 dnode_rele(dn, FTAG); 1627 1628 return (err); 1629 } 1630 #endif /* _KERNEL */ 1631 1632 static void 1633 dmu_cached_bps(spa_t *spa, blkptr_t *bps, uint_t nbps, 1634 uint64_t *l1sz, uint64_t *l2sz) 1635 { 1636 int cached_flags; 1637 1638 if (bps == NULL) 1639 return; 1640 1641 for (size_t blk_off = 0; blk_off < nbps; blk_off++) { 1642 blkptr_t *bp = &bps[blk_off]; 1643 1644 if (BP_IS_HOLE(bp)) 1645 continue; 1646 1647 cached_flags = arc_cached(spa, bp); 1648 if (cached_flags == 0) 1649 continue; 1650 1651 if ((cached_flags & (ARC_CACHED_IN_L1 | ARC_CACHED_IN_L2)) == 1652 ARC_CACHED_IN_L2) 1653 *l2sz += BP_GET_LSIZE(bp); 1654 else 1655 *l1sz += BP_GET_LSIZE(bp); 1656 } 1657 } 1658 1659 /* 1660 * Estimate DMU object cached size. 1661 */ 1662 int 1663 dmu_object_cached_size(objset_t *os, uint64_t object, 1664 uint64_t *l1sz, uint64_t *l2sz) 1665 { 1666 dnode_t *dn; 1667 dmu_object_info_t doi; 1668 int err = 0; 1669 1670 *l1sz = *l2sz = 0; 1671 1672 if (dnode_hold(os, object, FTAG, &dn) != 0) 1673 return (0); 1674 1675 if (dn->dn_nlevels < 2) { 1676 dnode_rele(dn, FTAG); 1677 return (0); 1678 } 1679 1680 dmu_object_info_from_dnode(dn, &doi); 1681 1682 for (uint64_t off = 0; off < doi.doi_max_offset; 1683 off += dmu_prefetch_max) { 1684 /* dbuf_read doesn't prefetch L1 blocks. */ 1685 dmu_prefetch_by_dnode(dn, 1, off, 1686 dmu_prefetch_max, ZIO_PRIORITY_SYNC_READ); 1687 } 1688 1689 /* 1690 * Hold all valid L1 blocks, asking ARC the status of each BP 1691 * contained in each such L1 block. 1692 */ 1693 uint_t nbps = bp_span_in_blocks(dn->dn_indblkshift, 1); 1694 uint64_t l1blks = 1 + (dn->dn_maxblkid / nbps); 1695 1696 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1697 for (uint64_t blk = 0; blk < l1blks; blk++) { 1698 dmu_buf_impl_t *db = NULL; 1699 1700 if (issig()) { 1701 /* 1702 * On interrupt, get out, and bubble up EINTR 1703 */ 1704 err = EINTR; 1705 break; 1706 } 1707 1708 /* 1709 * If we get an i/o error here, the L1 can't be read, 1710 * and nothing under it could be cached, so we just 1711 * continue. Ignoring the error from dbuf_hold_impl 1712 * or from dbuf_read is then a reasonable choice. 1713 */ 1714 err = dbuf_hold_impl(dn, 1, blk, B_TRUE, B_FALSE, FTAG, &db); 1715 if (err != 0) { 1716 /* 1717 * ignore error and continue 1718 */ 1719 err = 0; 1720 continue; 1721 } 1722 1723 err = dbuf_read(db, NULL, DB_RF_CANFAIL); 1724 if (err == 0) { 1725 dmu_cached_bps(dmu_objset_spa(os), db->db.db_data, 1726 nbps, l1sz, l2sz); 1727 } 1728 /* 1729 * error may be ignored, and we continue 1730 */ 1731 err = 0; 1732 dbuf_rele(db, FTAG); 1733 } 1734 rw_exit(&dn->dn_struct_rwlock); 1735 1736 dnode_rele(dn, FTAG); 1737 return (err); 1738 } 1739 1740 /* 1741 * Allocate a loaned anonymous arc buffer. 1742 */ 1743 arc_buf_t * 1744 dmu_request_arcbuf(dmu_buf_t *handle, int size) 1745 { 1746 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle; 1747 1748 return (arc_loan_buf(db->db_objset->os_spa, B_FALSE, size)); 1749 } 1750 1751 /* 1752 * Free a loaned arc buffer. 1753 */ 1754 void 1755 dmu_return_arcbuf(arc_buf_t *buf) 1756 { 1757 arc_return_buf(buf, FTAG); 1758 arc_buf_destroy(buf, FTAG); 1759 } 1760 1761 /* 1762 * A "lightweight" write is faster than a regular write (e.g. 1763 * dmu_write_by_dnode() or dmu_assign_arcbuf_by_dnode()), because it avoids the 1764 * CPU cost of creating a dmu_buf_impl_t and arc_buf_[hdr_]_t. However, the 1765 * data can not be read or overwritten until the transaction's txg has been 1766 * synced. This makes it appropriate for workloads that are known to be 1767 * (temporarily) write-only, like "zfs receive". 1768 * 1769 * A single block is written, starting at the specified offset in bytes. If 1770 * the call is successful, it returns 0 and the provided abd has been 1771 * consumed (the caller should not free it). 1772 */ 1773 int 1774 dmu_lightweight_write_by_dnode(dnode_t *dn, uint64_t offset, abd_t *abd, 1775 const zio_prop_t *zp, zio_flag_t flags, dmu_tx_t *tx) 1776 { 1777 dbuf_dirty_record_t *dr = 1778 dbuf_dirty_lightweight(dn, dbuf_whichblock(dn, 0, offset), tx); 1779 if (dr == NULL) 1780 return (SET_ERROR(EIO)); 1781 dr->dt.dll.dr_abd = abd; 1782 dr->dt.dll.dr_props = *zp; 1783 dr->dt.dll.dr_flags = flags; 1784 return (0); 1785 } 1786 1787 /* 1788 * When possible directly assign passed loaned arc buffer to a dbuf. 1789 * If this is not possible copy the contents of passed arc buf via 1790 * dmu_write(). 1791 */ 1792 int 1793 dmu_assign_arcbuf_by_dnode(dnode_t *dn, uint64_t offset, arc_buf_t *buf, 1794 dmu_tx_t *tx) 1795 { 1796 dmu_buf_impl_t *db; 1797 objset_t *os = dn->dn_objset; 1798 uint64_t object = dn->dn_object; 1799 uint32_t blksz = (uint32_t)arc_buf_lsize(buf); 1800 uint64_t blkid; 1801 1802 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1803 blkid = dbuf_whichblock(dn, 0, offset); 1804 db = dbuf_hold(dn, blkid, FTAG); 1805 rw_exit(&dn->dn_struct_rwlock); 1806 if (db == NULL) 1807 return (SET_ERROR(EIO)); 1808 1809 /* 1810 * We can only assign if the offset is aligned and the arc buf is the 1811 * same size as the dbuf. 1812 */ 1813 if (offset == db->db.db_offset && blksz == db->db.db_size) { 1814 zfs_racct_write(os->os_spa, blksz, 1, 0); 1815 dbuf_assign_arcbuf(db, buf, tx); 1816 dbuf_rele(db, FTAG); 1817 } else { 1818 /* compressed bufs must always be assignable to their dbuf */ 1819 ASSERT3U(arc_get_compression(buf), ==, ZIO_COMPRESS_OFF); 1820 ASSERT(!(buf->b_flags & ARC_BUF_FLAG_COMPRESSED)); 1821 1822 dbuf_rele(db, FTAG); 1823 dmu_write(os, object, offset, blksz, buf->b_data, tx); 1824 dmu_return_arcbuf(buf); 1825 } 1826 1827 return (0); 1828 } 1829 1830 int 1831 dmu_assign_arcbuf_by_dbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf, 1832 dmu_tx_t *tx) 1833 { 1834 int err; 1835 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle; 1836 1837 DB_DNODE_ENTER(db); 1838 err = dmu_assign_arcbuf_by_dnode(DB_DNODE(db), offset, buf, tx); 1839 DB_DNODE_EXIT(db); 1840 1841 return (err); 1842 } 1843 1844 void 1845 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg) 1846 { 1847 (void) buf; 1848 dmu_sync_arg_t *dsa = varg; 1849 1850 if (zio->io_error == 0) { 1851 dbuf_dirty_record_t *dr = dsa->dsa_dr; 1852 blkptr_t *bp = zio->io_bp; 1853 1854 if (BP_IS_HOLE(bp)) { 1855 dmu_buf_t *db = NULL; 1856 if (dr) 1857 db = &(dr->dr_dbuf->db); 1858 else 1859 db = dsa->dsa_zgd->zgd_db; 1860 /* 1861 * A block of zeros may compress to a hole, but the 1862 * block size still needs to be known for replay. 1863 */ 1864 BP_SET_LSIZE(bp, db->db_size); 1865 } else if (!BP_IS_EMBEDDED(bp)) { 1866 ASSERT(BP_GET_LEVEL(bp) == 0); 1867 BP_SET_FILL(bp, 1); 1868 } 1869 } 1870 } 1871 1872 static void 1873 dmu_sync_late_arrival_ready(zio_t *zio) 1874 { 1875 dmu_sync_ready(zio, NULL, zio->io_private); 1876 } 1877 1878 void 1879 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg) 1880 { 1881 (void) buf; 1882 dmu_sync_arg_t *dsa = varg; 1883 dbuf_dirty_record_t *dr = dsa->dsa_dr; 1884 dmu_buf_impl_t *db = dr->dr_dbuf; 1885 zgd_t *zgd = dsa->dsa_zgd; 1886 1887 /* 1888 * Record the vdev(s) backing this blkptr so they can be flushed after 1889 * the writes for the lwb have completed. 1890 */ 1891 if (zgd && zio->io_error == 0) { 1892 zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp); 1893 } 1894 1895 mutex_enter(&db->db_mtx); 1896 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC); 1897 if (zio->io_error == 0) { 1898 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE); 1899 if (dr->dt.dl.dr_nopwrite) { 1900 blkptr_t *bp = zio->io_bp; 1901 blkptr_t *bp_orig = &zio->io_bp_orig; 1902 uint8_t chksum = BP_GET_CHECKSUM(bp_orig); 1903 1904 ASSERT(BP_EQUAL(bp, bp_orig)); 1905 VERIFY(BP_EQUAL(bp, db->db_blkptr)); 1906 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF); 1907 VERIFY(zio_checksum_table[chksum].ci_flags & 1908 ZCHECKSUM_FLAG_NOPWRITE); 1909 } 1910 dr->dt.dl.dr_overridden_by = *zio->io_bp; 1911 dr->dt.dl.dr_override_state = DR_OVERRIDDEN; 1912 dr->dt.dl.dr_copies = zio->io_prop.zp_copies; 1913 1914 /* 1915 * Old style holes are filled with all zeros, whereas 1916 * new-style holes maintain their lsize, type, level, 1917 * and birth time (see zio_write_compress). While we 1918 * need to reset the BP_SET_LSIZE() call that happened 1919 * in dmu_sync_ready for old style holes, we do *not* 1920 * want to wipe out the information contained in new 1921 * style holes. Thus, only zero out the block pointer if 1922 * it's an old style hole. 1923 */ 1924 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) && 1925 BP_GET_LOGICAL_BIRTH(&dr->dt.dl.dr_overridden_by) == 0) 1926 BP_ZERO(&dr->dt.dl.dr_overridden_by); 1927 } else { 1928 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 1929 } 1930 1931 cv_broadcast(&db->db_changed); 1932 mutex_exit(&db->db_mtx); 1933 1934 if (dsa->dsa_done) 1935 dsa->dsa_done(dsa->dsa_zgd, zio->io_error); 1936 1937 kmem_free(dsa, sizeof (*dsa)); 1938 } 1939 1940 static void 1941 dmu_sync_late_arrival_done(zio_t *zio) 1942 { 1943 blkptr_t *bp = zio->io_bp; 1944 dmu_sync_arg_t *dsa = zio->io_private; 1945 zgd_t *zgd = dsa->dsa_zgd; 1946 1947 if (zio->io_error == 0) { 1948 /* 1949 * Record the vdev(s) backing this blkptr so they can be 1950 * flushed after the writes for the lwb have completed. 1951 */ 1952 zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp); 1953 1954 if (!BP_IS_HOLE(bp)) { 1955 blkptr_t *bp_orig __maybe_unused = &zio->io_bp_orig; 1956 ASSERT(!(zio->io_flags & ZIO_FLAG_NOPWRITE)); 1957 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig)); 1958 ASSERT(BP_GET_LOGICAL_BIRTH(zio->io_bp) == zio->io_txg); 1959 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa)); 1960 zio_free(zio->io_spa, zio->io_txg, zio->io_bp); 1961 } 1962 } 1963 1964 dmu_tx_commit(dsa->dsa_tx); 1965 1966 dsa->dsa_done(dsa->dsa_zgd, zio->io_error); 1967 1968 abd_free(zio->io_abd); 1969 kmem_free(dsa, sizeof (*dsa)); 1970 } 1971 1972 static int 1973 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd, 1974 zio_prop_t *zp, zbookmark_phys_t *zb) 1975 { 1976 dmu_sync_arg_t *dsa; 1977 dmu_tx_t *tx; 1978 int error; 1979 1980 error = dbuf_read((dmu_buf_impl_t *)zgd->zgd_db, NULL, 1981 DB_RF_CANFAIL | DB_RF_NOPREFETCH); 1982 if (error != 0) 1983 return (error); 1984 1985 tx = dmu_tx_create(os); 1986 dmu_tx_hold_space(tx, zgd->zgd_db->db_size); 1987 /* 1988 * This transaction does not produce any dirty data or log blocks, so 1989 * it should not be throttled. All other cases wait for TXG sync, by 1990 * which time the log block we are writing will be obsolete, so we can 1991 * skip waiting and just return error here instead. 1992 */ 1993 if (dmu_tx_assign(tx, TXG_NOWAIT | TXG_NOTHROTTLE) != 0) { 1994 dmu_tx_abort(tx); 1995 /* Make zl_get_data do txg_waited_synced() */ 1996 return (SET_ERROR(EIO)); 1997 } 1998 1999 /* 2000 * In order to prevent the zgd's lwb from being free'd prior to 2001 * dmu_sync_late_arrival_done() being called, we have to ensure 2002 * the lwb's "max txg" takes this tx's txg into account. 2003 */ 2004 zil_lwb_add_txg(zgd->zgd_lwb, dmu_tx_get_txg(tx)); 2005 2006 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP); 2007 dsa->dsa_dr = NULL; 2008 dsa->dsa_done = done; 2009 dsa->dsa_zgd = zgd; 2010 dsa->dsa_tx = tx; 2011 2012 /* 2013 * Since we are currently syncing this txg, it's nontrivial to 2014 * determine what BP to nopwrite against, so we disable nopwrite. 2015 * 2016 * When syncing, the db_blkptr is initially the BP of the previous 2017 * txg. We can not nopwrite against it because it will be changed 2018 * (this is similar to the non-late-arrival case where the dbuf is 2019 * dirty in a future txg). 2020 * 2021 * Then dbuf_write_ready() sets bp_blkptr to the location we will write. 2022 * We can not nopwrite against it because although the BP will not 2023 * (typically) be changed, the data has not yet been persisted to this 2024 * location. 2025 * 2026 * Finally, when dbuf_write_done() is called, it is theoretically 2027 * possible to always nopwrite, because the data that was written in 2028 * this txg is the same data that we are trying to write. However we 2029 * would need to check that this dbuf is not dirty in any future 2030 * txg's (as we do in the normal dmu_sync() path). For simplicity, we 2031 * don't nopwrite in this case. 2032 */ 2033 zp->zp_nopwrite = B_FALSE; 2034 2035 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp, 2036 abd_get_from_buf(zgd->zgd_db->db_data, zgd->zgd_db->db_size), 2037 zgd->zgd_db->db_size, zgd->zgd_db->db_size, zp, 2038 dmu_sync_late_arrival_ready, NULL, dmu_sync_late_arrival_done, 2039 dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb)); 2040 2041 return (0); 2042 } 2043 2044 /* 2045 * Intent log support: sync the block associated with db to disk. 2046 * N.B. and XXX: the caller is responsible for making sure that the 2047 * data isn't changing while dmu_sync() is writing it. 2048 * 2049 * Return values: 2050 * 2051 * EEXIST: this txg has already been synced, so there's nothing to do. 2052 * The caller should not log the write. 2053 * 2054 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do. 2055 * The caller should not log the write. 2056 * 2057 * EALREADY: this block is already in the process of being synced. 2058 * The caller should track its progress (somehow). 2059 * 2060 * EIO: could not do the I/O. 2061 * The caller should do a txg_wait_synced(). 2062 * 2063 * 0: the I/O has been initiated. 2064 * The caller should log this blkptr in the done callback. 2065 * It is possible that the I/O will fail, in which case 2066 * the error will be reported to the done callback and 2067 * propagated to pio from zio_done(). 2068 */ 2069 int 2070 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd) 2071 { 2072 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db; 2073 objset_t *os = db->db_objset; 2074 dsl_dataset_t *ds = os->os_dsl_dataset; 2075 dbuf_dirty_record_t *dr, *dr_next; 2076 dmu_sync_arg_t *dsa; 2077 zbookmark_phys_t zb; 2078 zio_prop_t zp; 2079 2080 ASSERT(pio != NULL); 2081 ASSERT(txg != 0); 2082 2083 SET_BOOKMARK(&zb, ds->ds_object, 2084 db->db.db_object, db->db_level, db->db_blkid); 2085 2086 DB_DNODE_ENTER(db); 2087 dmu_write_policy(os, DB_DNODE(db), db->db_level, WP_DMU_SYNC, &zp); 2088 DB_DNODE_EXIT(db); 2089 2090 /* 2091 * If we're frozen (running ziltest), we always need to generate a bp. 2092 */ 2093 if (txg > spa_freeze_txg(os->os_spa)) 2094 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); 2095 2096 /* 2097 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf() 2098 * and us. If we determine that this txg is not yet syncing, 2099 * but it begins to sync a moment later, that's OK because the 2100 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx. 2101 */ 2102 mutex_enter(&db->db_mtx); 2103 2104 if (txg <= spa_last_synced_txg(os->os_spa)) { 2105 /* 2106 * This txg has already synced. There's nothing to do. 2107 */ 2108 mutex_exit(&db->db_mtx); 2109 return (SET_ERROR(EEXIST)); 2110 } 2111 2112 if (txg <= spa_syncing_txg(os->os_spa)) { 2113 /* 2114 * This txg is currently syncing, so we can't mess with 2115 * the dirty record anymore; just write a new log block. 2116 */ 2117 mutex_exit(&db->db_mtx); 2118 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); 2119 } 2120 2121 dr = dbuf_find_dirty_eq(db, txg); 2122 2123 if (dr == NULL) { 2124 /* 2125 * There's no dr for this dbuf, so it must have been freed. 2126 * There's no need to log writes to freed blocks, so we're done. 2127 */ 2128 mutex_exit(&db->db_mtx); 2129 return (SET_ERROR(ENOENT)); 2130 } 2131 2132 dr_next = list_next(&db->db_dirty_records, dr); 2133 ASSERT(dr_next == NULL || dr_next->dr_txg < txg); 2134 2135 if (db->db_blkptr != NULL) { 2136 /* 2137 * We need to fill in zgd_bp with the current blkptr so that 2138 * the nopwrite code can check if we're writing the same 2139 * data that's already on disk. We can only nopwrite if we 2140 * are sure that after making the copy, db_blkptr will not 2141 * change until our i/o completes. We ensure this by 2142 * holding the db_mtx, and only allowing nopwrite if the 2143 * block is not already dirty (see below). This is verified 2144 * by dmu_sync_done(), which VERIFYs that the db_blkptr has 2145 * not changed. 2146 */ 2147 *zgd->zgd_bp = *db->db_blkptr; 2148 } 2149 2150 /* 2151 * Assume the on-disk data is X, the current syncing data (in 2152 * txg - 1) is Y, and the current in-memory data is Z (currently 2153 * in dmu_sync). 2154 * 2155 * We usually want to perform a nopwrite if X and Z are the 2156 * same. However, if Y is different (i.e. the BP is going to 2157 * change before this write takes effect), then a nopwrite will 2158 * be incorrect - we would override with X, which could have 2159 * been freed when Y was written. 2160 * 2161 * (Note that this is not a concern when we are nop-writing from 2162 * syncing context, because X and Y must be identical, because 2163 * all previous txgs have been synced.) 2164 * 2165 * Therefore, we disable nopwrite if the current BP could change 2166 * before this TXG. There are two ways it could change: by 2167 * being dirty (dr_next is non-NULL), or by being freed 2168 * (dnode_block_freed()). This behavior is verified by 2169 * zio_done(), which VERIFYs that the override BP is identical 2170 * to the on-disk BP. 2171 */ 2172 if (dr_next != NULL) { 2173 zp.zp_nopwrite = B_FALSE; 2174 } else { 2175 DB_DNODE_ENTER(db); 2176 if (dnode_block_freed(DB_DNODE(db), db->db_blkid)) 2177 zp.zp_nopwrite = B_FALSE; 2178 DB_DNODE_EXIT(db); 2179 } 2180 2181 ASSERT(dr->dr_txg == txg); 2182 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC || 2183 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { 2184 /* 2185 * We have already issued a sync write for this buffer, 2186 * or this buffer has already been synced. It could not 2187 * have been dirtied since, or we would have cleared the state. 2188 */ 2189 mutex_exit(&db->db_mtx); 2190 return (SET_ERROR(EALREADY)); 2191 } 2192 2193 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); 2194 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC; 2195 mutex_exit(&db->db_mtx); 2196 2197 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP); 2198 dsa->dsa_dr = dr; 2199 dsa->dsa_done = done; 2200 dsa->dsa_zgd = zgd; 2201 dsa->dsa_tx = NULL; 2202 2203 zio_nowait(arc_write(pio, os->os_spa, txg, zgd->zgd_bp, 2204 dr->dt.dl.dr_data, !DBUF_IS_CACHEABLE(db), 2205 dbuf_is_l2cacheable(db, NULL), &zp, dmu_sync_ready, NULL, 2206 dmu_sync_done, dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, 2207 &zb)); 2208 2209 return (0); 2210 } 2211 2212 int 2213 dmu_object_set_nlevels(objset_t *os, uint64_t object, int nlevels, dmu_tx_t *tx) 2214 { 2215 dnode_t *dn; 2216 int err; 2217 2218 err = dnode_hold(os, object, FTAG, &dn); 2219 if (err) 2220 return (err); 2221 err = dnode_set_nlevels(dn, nlevels, tx); 2222 dnode_rele(dn, FTAG); 2223 return (err); 2224 } 2225 2226 int 2227 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs, 2228 dmu_tx_t *tx) 2229 { 2230 dnode_t *dn; 2231 int err; 2232 2233 err = dnode_hold(os, object, FTAG, &dn); 2234 if (err) 2235 return (err); 2236 err = dnode_set_blksz(dn, size, ibs, tx); 2237 dnode_rele(dn, FTAG); 2238 return (err); 2239 } 2240 2241 int 2242 dmu_object_set_maxblkid(objset_t *os, uint64_t object, uint64_t maxblkid, 2243 dmu_tx_t *tx) 2244 { 2245 dnode_t *dn; 2246 int err; 2247 2248 err = dnode_hold(os, object, FTAG, &dn); 2249 if (err) 2250 return (err); 2251 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 2252 dnode_new_blkid(dn, maxblkid, tx, B_FALSE, B_TRUE); 2253 rw_exit(&dn->dn_struct_rwlock); 2254 dnode_rele(dn, FTAG); 2255 return (0); 2256 } 2257 2258 void 2259 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum, 2260 dmu_tx_t *tx) 2261 { 2262 dnode_t *dn; 2263 2264 /* 2265 * Send streams include each object's checksum function. This 2266 * check ensures that the receiving system can understand the 2267 * checksum function transmitted. 2268 */ 2269 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS); 2270 2271 VERIFY0(dnode_hold(os, object, FTAG, &dn)); 2272 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS); 2273 dn->dn_checksum = checksum; 2274 dnode_setdirty(dn, tx); 2275 dnode_rele(dn, FTAG); 2276 } 2277 2278 void 2279 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress, 2280 dmu_tx_t *tx) 2281 { 2282 dnode_t *dn; 2283 2284 /* 2285 * Send streams include each object's compression function. This 2286 * check ensures that the receiving system can understand the 2287 * compression function transmitted. 2288 */ 2289 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS); 2290 2291 VERIFY0(dnode_hold(os, object, FTAG, &dn)); 2292 dn->dn_compress = compress; 2293 dnode_setdirty(dn, tx); 2294 dnode_rele(dn, FTAG); 2295 } 2296 2297 /* 2298 * When the "redundant_metadata" property is set to "most", only indirect 2299 * blocks of this level and higher will have an additional ditto block. 2300 */ 2301 static const int zfs_redundant_metadata_most_ditto_level = 2; 2302 2303 void 2304 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp) 2305 { 2306 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET; 2307 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) || 2308 (wp & WP_SPILL)); 2309 enum zio_checksum checksum = os->os_checksum; 2310 enum zio_compress compress = os->os_compress; 2311 uint8_t complevel = os->os_complevel; 2312 enum zio_checksum dedup_checksum = os->os_dedup_checksum; 2313 boolean_t dedup = B_FALSE; 2314 boolean_t nopwrite = B_FALSE; 2315 boolean_t dedup_verify = os->os_dedup_verify; 2316 boolean_t encrypt = B_FALSE; 2317 int copies = os->os_copies; 2318 2319 /* 2320 * We maintain different write policies for each of the following 2321 * types of data: 2322 * 1. metadata 2323 * 2. preallocated blocks (i.e. level-0 blocks of a dump device) 2324 * 3. all other level 0 blocks 2325 */ 2326 if (ismd) { 2327 /* 2328 * XXX -- we should design a compression algorithm 2329 * that specializes in arrays of bps. 2330 */ 2331 compress = zio_compress_select(os->os_spa, 2332 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON); 2333 2334 /* 2335 * Metadata always gets checksummed. If the data 2336 * checksum is multi-bit correctable, and it's not a 2337 * ZBT-style checksum, then it's suitable for metadata 2338 * as well. Otherwise, the metadata checksum defaults 2339 * to fletcher4. 2340 */ 2341 if (!(zio_checksum_table[checksum].ci_flags & 2342 ZCHECKSUM_FLAG_METADATA) || 2343 (zio_checksum_table[checksum].ci_flags & 2344 ZCHECKSUM_FLAG_EMBEDDED)) 2345 checksum = ZIO_CHECKSUM_FLETCHER_4; 2346 2347 switch (os->os_redundant_metadata) { 2348 case ZFS_REDUNDANT_METADATA_ALL: 2349 copies++; 2350 break; 2351 case ZFS_REDUNDANT_METADATA_MOST: 2352 if (level >= zfs_redundant_metadata_most_ditto_level || 2353 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL)) 2354 copies++; 2355 break; 2356 case ZFS_REDUNDANT_METADATA_SOME: 2357 if (DMU_OT_IS_CRITICAL(type)) 2358 copies++; 2359 break; 2360 case ZFS_REDUNDANT_METADATA_NONE: 2361 break; 2362 } 2363 2364 if (dmu_ddt_copies > 0) { 2365 /* 2366 * If this tuneable is set, and this is a write for a 2367 * dedup entry store (zap or log), then we treat it 2368 * something like ZFS_REDUNDANT_METADATA_MOST on a 2369 * regular dataset: this many copies, and one more for 2370 * "higher" indirect blocks. This specific exception is 2371 * necessary because dedup objects are stored in the 2372 * MOS, which always has the highest possible copies. 2373 */ 2374 dmu_object_type_t stype = 2375 dn ? dn->dn_storage_type : DMU_OT_NONE; 2376 if (stype == DMU_OT_NONE) 2377 stype = type; 2378 if (stype == DMU_OT_DDT_ZAP) { 2379 copies = dmu_ddt_copies; 2380 if (level >= 2381 zfs_redundant_metadata_most_ditto_level) 2382 copies++; 2383 } 2384 } 2385 } else if (wp & WP_NOFILL) { 2386 ASSERT(level == 0); 2387 2388 /* 2389 * If we're writing preallocated blocks, we aren't actually 2390 * writing them so don't set any policy properties. These 2391 * blocks are currently only used by an external subsystem 2392 * outside of zfs (i.e. dump) and not written by the zio 2393 * pipeline. 2394 */ 2395 compress = ZIO_COMPRESS_OFF; 2396 checksum = ZIO_CHECKSUM_OFF; 2397 } else { 2398 compress = zio_compress_select(os->os_spa, dn->dn_compress, 2399 compress); 2400 complevel = zio_complevel_select(os->os_spa, compress, 2401 complevel, complevel); 2402 2403 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ? 2404 zio_checksum_select(dn->dn_checksum, checksum) : 2405 dedup_checksum; 2406 2407 /* 2408 * Determine dedup setting. If we are in dmu_sync(), 2409 * we won't actually dedup now because that's all 2410 * done in syncing context; but we do want to use the 2411 * dedup checksum. If the checksum is not strong 2412 * enough to ensure unique signatures, force 2413 * dedup_verify. 2414 */ 2415 if (dedup_checksum != ZIO_CHECKSUM_OFF) { 2416 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE; 2417 if (!(zio_checksum_table[checksum].ci_flags & 2418 ZCHECKSUM_FLAG_DEDUP)) 2419 dedup_verify = B_TRUE; 2420 } 2421 2422 /* 2423 * Enable nopwrite if we have secure enough checksum 2424 * algorithm (see comment in zio_nop_write) and 2425 * compression is enabled. We don't enable nopwrite if 2426 * dedup is enabled as the two features are mutually 2427 * exclusive. 2428 */ 2429 nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags & 2430 ZCHECKSUM_FLAG_NOPWRITE) && 2431 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled); 2432 } 2433 2434 /* 2435 * All objects in an encrypted objset are protected from modification 2436 * via a MAC. Encrypted objects store their IV and salt in the last DVA 2437 * in the bp, so we cannot use all copies. Encrypted objects are also 2438 * not subject to nopwrite since writing the same data will still 2439 * result in a new ciphertext. Only encrypted blocks can be dedup'd 2440 * to avoid ambiguity in the dedup code since the DDT does not store 2441 * object types. 2442 */ 2443 if (os->os_encrypted && (wp & WP_NOFILL) == 0) { 2444 encrypt = B_TRUE; 2445 2446 if (DMU_OT_IS_ENCRYPTED(type)) { 2447 copies = MIN(copies, SPA_DVAS_PER_BP - 1); 2448 nopwrite = B_FALSE; 2449 } else { 2450 dedup = B_FALSE; 2451 } 2452 2453 if (level <= 0 && 2454 (type == DMU_OT_DNODE || type == DMU_OT_OBJSET)) { 2455 compress = ZIO_COMPRESS_EMPTY; 2456 } 2457 } 2458 2459 zp->zp_compress = compress; 2460 zp->zp_complevel = complevel; 2461 zp->zp_checksum = checksum; 2462 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type; 2463 zp->zp_level = level; 2464 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa)); 2465 zp->zp_dedup = dedup; 2466 zp->zp_dedup_verify = dedup && dedup_verify; 2467 zp->zp_nopwrite = nopwrite; 2468 zp->zp_encrypt = encrypt; 2469 zp->zp_byteorder = ZFS_HOST_BYTEORDER; 2470 zp->zp_direct_write = (wp & WP_DIRECT_WR) ? B_TRUE : B_FALSE; 2471 memset(zp->zp_salt, 0, ZIO_DATA_SALT_LEN); 2472 memset(zp->zp_iv, 0, ZIO_DATA_IV_LEN); 2473 memset(zp->zp_mac, 0, ZIO_DATA_MAC_LEN); 2474 zp->zp_zpl_smallblk = DMU_OT_IS_FILE(zp->zp_type) ? 2475 os->os_zpl_special_smallblock : 0; 2476 zp->zp_storage_type = dn ? dn->dn_storage_type : DMU_OT_NONE; 2477 2478 ASSERT3U(zp->zp_compress, !=, ZIO_COMPRESS_INHERIT); 2479 } 2480 2481 /* 2482 * Reports the location of data and holes in an object. In order to 2483 * accurately report holes all dirty data must be synced to disk. This 2484 * causes extremely poor performance when seeking for holes in a dirty file. 2485 * As a compromise, only provide hole data when the dnode is clean. When 2486 * a dnode is dirty report the dnode as having no holes by returning EBUSY 2487 * which is always safe to do. 2488 */ 2489 int 2490 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off) 2491 { 2492 dnode_t *dn; 2493 int restarted = 0, err; 2494 2495 restart: 2496 err = dnode_hold(os, object, FTAG, &dn); 2497 if (err) 2498 return (err); 2499 2500 rw_enter(&dn->dn_struct_rwlock, RW_READER); 2501 2502 if (dnode_is_dirty(dn)) { 2503 /* 2504 * If the zfs_dmu_offset_next_sync module option is enabled 2505 * then hole reporting has been requested. Dirty dnodes 2506 * must be synced to disk to accurately report holes. 2507 * 2508 * Provided a RL_READER rangelock spanning 0-UINT64_MAX is 2509 * held by the caller only a single restart will be required. 2510 * We tolerate callers which do not hold the rangelock by 2511 * returning EBUSY and not reporting holes after one restart. 2512 */ 2513 if (zfs_dmu_offset_next_sync) { 2514 rw_exit(&dn->dn_struct_rwlock); 2515 dnode_rele(dn, FTAG); 2516 2517 if (restarted) 2518 return (SET_ERROR(EBUSY)); 2519 2520 txg_wait_synced(dmu_objset_pool(os), 0); 2521 restarted = 1; 2522 goto restart; 2523 } 2524 2525 err = SET_ERROR(EBUSY); 2526 } else { 2527 err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK | 2528 (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0); 2529 } 2530 2531 rw_exit(&dn->dn_struct_rwlock); 2532 dnode_rele(dn, FTAG); 2533 2534 return (err); 2535 } 2536 2537 int 2538 dmu_read_l0_bps(objset_t *os, uint64_t object, uint64_t offset, uint64_t length, 2539 blkptr_t *bps, size_t *nbpsp) 2540 { 2541 dmu_buf_t **dbp, *dbuf; 2542 dmu_buf_impl_t *db; 2543 blkptr_t *bp; 2544 int error, numbufs; 2545 2546 error = dmu_buf_hold_array(os, object, offset, length, FALSE, FTAG, 2547 &numbufs, &dbp); 2548 if (error != 0) { 2549 if (error == ESRCH) { 2550 error = SET_ERROR(ENXIO); 2551 } 2552 return (error); 2553 } 2554 2555 ASSERT3U(numbufs, <=, *nbpsp); 2556 2557 for (int i = 0; i < numbufs; i++) { 2558 dbuf = dbp[i]; 2559 db = (dmu_buf_impl_t *)dbuf; 2560 2561 mutex_enter(&db->db_mtx); 2562 2563 if (!list_is_empty(&db->db_dirty_records)) { 2564 dbuf_dirty_record_t *dr; 2565 2566 dr = list_head(&db->db_dirty_records); 2567 if (dr->dt.dl.dr_brtwrite) { 2568 /* 2569 * This is very special case where we clone a 2570 * block and in the same transaction group we 2571 * read its BP (most likely to clone the clone). 2572 */ 2573 bp = &dr->dt.dl.dr_overridden_by; 2574 } else { 2575 /* 2576 * The block was modified in the same 2577 * transaction group. 2578 */ 2579 mutex_exit(&db->db_mtx); 2580 error = SET_ERROR(EAGAIN); 2581 goto out; 2582 } 2583 } else { 2584 bp = db->db_blkptr; 2585 } 2586 2587 mutex_exit(&db->db_mtx); 2588 2589 if (bp == NULL) { 2590 /* 2591 * The file size was increased, but the block was never 2592 * written, otherwise we would either have the block 2593 * pointer or the dirty record and would not get here. 2594 * It is effectively a hole, so report it as such. 2595 */ 2596 BP_ZERO(&bps[i]); 2597 continue; 2598 } 2599 /* 2600 * Make sure we clone only data blocks. 2601 */ 2602 if (BP_IS_METADATA(bp) && !BP_IS_HOLE(bp)) { 2603 error = SET_ERROR(EINVAL); 2604 goto out; 2605 } 2606 2607 /* 2608 * If the block was allocated in transaction group that is not 2609 * yet synced, we could clone it, but we couldn't write this 2610 * operation into ZIL, or it may be impossible to replay, since 2611 * the block may appear not yet allocated at that point. 2612 */ 2613 if (BP_GET_BIRTH(bp) > spa_freeze_txg(os->os_spa)) { 2614 error = SET_ERROR(EINVAL); 2615 goto out; 2616 } 2617 if (BP_GET_BIRTH(bp) > spa_last_synced_txg(os->os_spa)) { 2618 error = SET_ERROR(EAGAIN); 2619 goto out; 2620 } 2621 2622 bps[i] = *bp; 2623 } 2624 2625 *nbpsp = numbufs; 2626 out: 2627 dmu_buf_rele_array(dbp, numbufs, FTAG); 2628 2629 return (error); 2630 } 2631 2632 int 2633 dmu_brt_clone(objset_t *os, uint64_t object, uint64_t offset, uint64_t length, 2634 dmu_tx_t *tx, const blkptr_t *bps, size_t nbps) 2635 { 2636 spa_t *spa; 2637 dmu_buf_t **dbp, *dbuf; 2638 dmu_buf_impl_t *db; 2639 struct dirty_leaf *dl; 2640 dbuf_dirty_record_t *dr; 2641 const blkptr_t *bp; 2642 int error = 0, i, numbufs; 2643 2644 spa = os->os_spa; 2645 2646 VERIFY0(dmu_buf_hold_array(os, object, offset, length, FALSE, FTAG, 2647 &numbufs, &dbp)); 2648 ASSERT3U(nbps, ==, numbufs); 2649 2650 /* 2651 * Before we start cloning make sure that the dbufs sizes match new BPs 2652 * sizes. If they don't, that's a no-go, as we are not able to shrink 2653 * dbufs. 2654 */ 2655 for (i = 0; i < numbufs; i++) { 2656 dbuf = dbp[i]; 2657 db = (dmu_buf_impl_t *)dbuf; 2658 bp = &bps[i]; 2659 2660 ASSERT0(db->db_level); 2661 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 2662 ASSERT(db->db_blkid != DMU_SPILL_BLKID); 2663 2664 if (!BP_IS_HOLE(bp) && BP_GET_LSIZE(bp) != dbuf->db_size) { 2665 error = SET_ERROR(EXDEV); 2666 goto out; 2667 } 2668 } 2669 2670 for (i = 0; i < numbufs; i++) { 2671 dbuf = dbp[i]; 2672 db = (dmu_buf_impl_t *)dbuf; 2673 bp = &bps[i]; 2674 2675 ASSERT0(db->db_level); 2676 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 2677 ASSERT(db->db_blkid != DMU_SPILL_BLKID); 2678 ASSERT(BP_IS_HOLE(bp) || dbuf->db_size == BP_GET_LSIZE(bp)); 2679 2680 dmu_buf_will_clone_or_dio(dbuf, tx); 2681 2682 mutex_enter(&db->db_mtx); 2683 2684 dr = list_head(&db->db_dirty_records); 2685 VERIFY(dr != NULL); 2686 ASSERT3U(dr->dr_txg, ==, tx->tx_txg); 2687 dl = &dr->dt.dl; 2688 dl->dr_overridden_by = *bp; 2689 if (!BP_IS_HOLE(bp) || BP_GET_LOGICAL_BIRTH(bp) != 0) { 2690 if (!BP_IS_EMBEDDED(bp)) { 2691 BP_SET_BIRTH(&dl->dr_overridden_by, dr->dr_txg, 2692 BP_GET_BIRTH(bp)); 2693 } else { 2694 BP_SET_LOGICAL_BIRTH(&dl->dr_overridden_by, 2695 dr->dr_txg); 2696 } 2697 } 2698 dl->dr_brtwrite = B_TRUE; 2699 dl->dr_override_state = DR_OVERRIDDEN; 2700 2701 mutex_exit(&db->db_mtx); 2702 2703 /* 2704 * When data in embedded into BP there is no need to create 2705 * BRT entry as there is no data block. Just copy the BP as 2706 * it contains the data. 2707 */ 2708 if (!BP_IS_HOLE(bp) && !BP_IS_EMBEDDED(bp)) { 2709 brt_pending_add(spa, bp, tx); 2710 } 2711 } 2712 out: 2713 dmu_buf_rele_array(dbp, numbufs, FTAG); 2714 2715 return (error); 2716 } 2717 2718 void 2719 __dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi) 2720 { 2721 dnode_phys_t *dnp = dn->dn_phys; 2722 2723 doi->doi_data_block_size = dn->dn_datablksz; 2724 doi->doi_metadata_block_size = dn->dn_indblkshift ? 2725 1ULL << dn->dn_indblkshift : 0; 2726 doi->doi_type = dn->dn_type; 2727 doi->doi_bonus_type = dn->dn_bonustype; 2728 doi->doi_bonus_size = dn->dn_bonuslen; 2729 doi->doi_dnodesize = dn->dn_num_slots << DNODE_SHIFT; 2730 doi->doi_indirection = dn->dn_nlevels; 2731 doi->doi_checksum = dn->dn_checksum; 2732 doi->doi_compress = dn->dn_compress; 2733 doi->doi_nblkptr = dn->dn_nblkptr; 2734 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9; 2735 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz; 2736 doi->doi_fill_count = 0; 2737 for (int i = 0; i < dnp->dn_nblkptr; i++) 2738 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]); 2739 } 2740 2741 void 2742 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi) 2743 { 2744 rw_enter(&dn->dn_struct_rwlock, RW_READER); 2745 mutex_enter(&dn->dn_mtx); 2746 2747 __dmu_object_info_from_dnode(dn, doi); 2748 2749 mutex_exit(&dn->dn_mtx); 2750 rw_exit(&dn->dn_struct_rwlock); 2751 } 2752 2753 /* 2754 * Get information on a DMU object. 2755 * If doi is NULL, just indicates whether the object exists. 2756 */ 2757 int 2758 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi) 2759 { 2760 dnode_t *dn; 2761 int err = dnode_hold(os, object, FTAG, &dn); 2762 2763 if (err) 2764 return (err); 2765 2766 if (doi != NULL) 2767 dmu_object_info_from_dnode(dn, doi); 2768 2769 dnode_rele(dn, FTAG); 2770 return (0); 2771 } 2772 2773 /* 2774 * As above, but faster; can be used when you have a held dbuf in hand. 2775 */ 2776 void 2777 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi) 2778 { 2779 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2780 2781 DB_DNODE_ENTER(db); 2782 dmu_object_info_from_dnode(DB_DNODE(db), doi); 2783 DB_DNODE_EXIT(db); 2784 } 2785 2786 /* 2787 * Faster still when you only care about the size. 2788 */ 2789 void 2790 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize, 2791 u_longlong_t *nblk512) 2792 { 2793 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2794 dnode_t *dn; 2795 2796 DB_DNODE_ENTER(db); 2797 dn = DB_DNODE(db); 2798 2799 *blksize = dn->dn_datablksz; 2800 /* add in number of slots used for the dnode itself */ 2801 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >> 2802 SPA_MINBLOCKSHIFT) + dn->dn_num_slots; 2803 DB_DNODE_EXIT(db); 2804 } 2805 2806 void 2807 dmu_object_dnsize_from_db(dmu_buf_t *db_fake, int *dnsize) 2808 { 2809 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2810 2811 DB_DNODE_ENTER(db); 2812 *dnsize = DB_DNODE(db)->dn_num_slots << DNODE_SHIFT; 2813 DB_DNODE_EXIT(db); 2814 } 2815 2816 void 2817 byteswap_uint64_array(void *vbuf, size_t size) 2818 { 2819 uint64_t *buf = vbuf; 2820 size_t count = size >> 3; 2821 int i; 2822 2823 ASSERT((size & 7) == 0); 2824 2825 for (i = 0; i < count; i++) 2826 buf[i] = BSWAP_64(buf[i]); 2827 } 2828 2829 void 2830 byteswap_uint32_array(void *vbuf, size_t size) 2831 { 2832 uint32_t *buf = vbuf; 2833 size_t count = size >> 2; 2834 int i; 2835 2836 ASSERT((size & 3) == 0); 2837 2838 for (i = 0; i < count; i++) 2839 buf[i] = BSWAP_32(buf[i]); 2840 } 2841 2842 void 2843 byteswap_uint16_array(void *vbuf, size_t size) 2844 { 2845 uint16_t *buf = vbuf; 2846 size_t count = size >> 1; 2847 int i; 2848 2849 ASSERT((size & 1) == 0); 2850 2851 for (i = 0; i < count; i++) 2852 buf[i] = BSWAP_16(buf[i]); 2853 } 2854 2855 void 2856 byteswap_uint8_array(void *vbuf, size_t size) 2857 { 2858 (void) vbuf, (void) size; 2859 } 2860 2861 void 2862 dmu_init(void) 2863 { 2864 abd_init(); 2865 zfs_dbgmsg_init(); 2866 sa_cache_init(); 2867 dmu_objset_init(); 2868 dnode_init(); 2869 zfetch_init(); 2870 dmu_tx_init(); 2871 l2arc_init(); 2872 arc_init(); 2873 dbuf_init(); 2874 } 2875 2876 void 2877 dmu_fini(void) 2878 { 2879 arc_fini(); /* arc depends on l2arc, so arc must go first */ 2880 l2arc_fini(); 2881 dmu_tx_fini(); 2882 zfetch_fini(); 2883 dbuf_fini(); 2884 dnode_fini(); 2885 dmu_objset_fini(); 2886 sa_cache_fini(); 2887 zfs_dbgmsg_fini(); 2888 abd_fini(); 2889 } 2890 2891 EXPORT_SYMBOL(dmu_bonus_hold); 2892 EXPORT_SYMBOL(dmu_bonus_hold_by_dnode); 2893 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus); 2894 EXPORT_SYMBOL(dmu_buf_rele_array); 2895 EXPORT_SYMBOL(dmu_prefetch); 2896 EXPORT_SYMBOL(dmu_prefetch_by_dnode); 2897 EXPORT_SYMBOL(dmu_prefetch_dnode); 2898 EXPORT_SYMBOL(dmu_free_range); 2899 EXPORT_SYMBOL(dmu_free_long_range); 2900 EXPORT_SYMBOL(dmu_free_long_object); 2901 EXPORT_SYMBOL(dmu_read); 2902 EXPORT_SYMBOL(dmu_read_by_dnode); 2903 EXPORT_SYMBOL(dmu_read_uio); 2904 EXPORT_SYMBOL(dmu_read_uio_dbuf); 2905 EXPORT_SYMBOL(dmu_read_uio_dnode); 2906 EXPORT_SYMBOL(dmu_write); 2907 EXPORT_SYMBOL(dmu_write_by_dnode); 2908 EXPORT_SYMBOL(dmu_write_by_dnode_flags); 2909 EXPORT_SYMBOL(dmu_write_uio); 2910 EXPORT_SYMBOL(dmu_write_uio_dbuf); 2911 EXPORT_SYMBOL(dmu_write_uio_dnode); 2912 EXPORT_SYMBOL(dmu_prealloc); 2913 EXPORT_SYMBOL(dmu_object_info); 2914 EXPORT_SYMBOL(dmu_object_info_from_dnode); 2915 EXPORT_SYMBOL(dmu_object_info_from_db); 2916 EXPORT_SYMBOL(dmu_object_size_from_db); 2917 EXPORT_SYMBOL(dmu_object_dnsize_from_db); 2918 EXPORT_SYMBOL(dmu_object_set_nlevels); 2919 EXPORT_SYMBOL(dmu_object_set_blocksize); 2920 EXPORT_SYMBOL(dmu_object_set_maxblkid); 2921 EXPORT_SYMBOL(dmu_object_set_checksum); 2922 EXPORT_SYMBOL(dmu_object_set_compress); 2923 EXPORT_SYMBOL(dmu_offset_next); 2924 EXPORT_SYMBOL(dmu_write_policy); 2925 EXPORT_SYMBOL(dmu_sync); 2926 EXPORT_SYMBOL(dmu_request_arcbuf); 2927 EXPORT_SYMBOL(dmu_return_arcbuf); 2928 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dnode); 2929 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dbuf); 2930 EXPORT_SYMBOL(dmu_buf_hold); 2931 EXPORT_SYMBOL(dmu_ot); 2932 2933 ZFS_MODULE_PARAM(zfs, zfs_, nopwrite_enabled, INT, ZMOD_RW, 2934 "Enable NOP writes"); 2935 2936 ZFS_MODULE_PARAM(zfs, zfs_, per_txg_dirty_frees_percent, UINT, ZMOD_RW, 2937 "Percentage of dirtied blocks from frees in one TXG"); 2938 2939 ZFS_MODULE_PARAM(zfs, zfs_, dmu_offset_next_sync, INT, ZMOD_RW, 2940 "Enable forcing txg sync to find holes"); 2941 2942 /* CSTYLED */ 2943 ZFS_MODULE_PARAM(zfs, , dmu_prefetch_max, UINT, ZMOD_RW, 2944 "Limit one prefetch call to this size"); 2945 2946 /* CSTYLED */ 2947 ZFS_MODULE_PARAM(zfs, , dmu_ddt_copies, UINT, ZMOD_RW, 2948 "Override copies= for dedup objects"); 2949