1 // SPDX-License-Identifier: CDDL-1.0 2 /* 3 * CDDL HEADER START 4 * 5 * The contents of this file are subject to the terms of the 6 * Common Development and Distribution License (the "License"). 7 * You may not use this file except in compliance with the License. 8 * 9 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 10 * or https://opensource.org/licenses/CDDL-1.0. 11 * See the License for the specific language governing permissions 12 * and limitations under the License. 13 * 14 * When distributing Covered Code, include this CDDL HEADER in each 15 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 16 * If applicable, add the following below this CDDL HEADER, with the 17 * fields enclosed by brackets "[]" replaced with your own identifying 18 * information: Portions Copyright [yyyy] [name of copyright owner] 19 * 20 * CDDL HEADER END 21 */ 22 /* 23 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 24 * Copyright (c) 2011, 2020 by Delphix. All rights reserved. 25 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved. 26 * Copyright (c) 2013, Joyent, Inc. All rights reserved. 27 * Copyright (c) 2016, Nexenta Systems, Inc. All rights reserved. 28 * Copyright (c) 2015 by Chunwei Chen. All rights reserved. 29 * Copyright (c) 2019 Datto Inc. 30 * Copyright (c) 2019, 2023, Klara Inc. 31 * Copyright (c) 2019, Allan Jude 32 * Copyright (c) 2022 Hewlett Packard Enterprise Development LP. 33 * Copyright (c) 2021, 2022 by Pawel Jakub Dawidek 34 */ 35 36 #include <sys/dmu.h> 37 #include <sys/dmu_impl.h> 38 #include <sys/dmu_tx.h> 39 #include <sys/dbuf.h> 40 #include <sys/dnode.h> 41 #include <sys/zfs_context.h> 42 #include <sys/dmu_objset.h> 43 #include <sys/dmu_traverse.h> 44 #include <sys/dsl_dataset.h> 45 #include <sys/dsl_dir.h> 46 #include <sys/dsl_pool.h> 47 #include <sys/dsl_synctask.h> 48 #include <sys/dsl_prop.h> 49 #include <sys/dmu_zfetch.h> 50 #include <sys/zfs_ioctl.h> 51 #include <sys/zap.h> 52 #include <sys/zio_checksum.h> 53 #include <sys/zio_compress.h> 54 #include <sys/sa.h> 55 #include <sys/zfeature.h> 56 #include <sys/abd.h> 57 #include <sys/brt.h> 58 #include <sys/trace_zfs.h> 59 #include <sys/zfs_racct.h> 60 #include <sys/zfs_rlock.h> 61 #ifdef _KERNEL 62 #include <sys/vmsystm.h> 63 #include <sys/zfs_znode.h> 64 #endif 65 66 /* 67 * Enable/disable nopwrite feature. 68 */ 69 static int zfs_nopwrite_enabled = 1; 70 71 /* 72 * Tunable to control percentage of dirtied L1 blocks from frees allowed into 73 * one TXG. After this threshold is crossed, additional dirty blocks from frees 74 * will wait until the next TXG. 75 * A value of zero will disable this throttle. 76 */ 77 static uint_t zfs_per_txg_dirty_frees_percent = 30; 78 79 /* 80 * Enable/disable forcing txg sync when dirty checking for holes with lseek(). 81 * By default this is enabled to ensure accurate hole reporting, it can result 82 * in a significant performance penalty for lseek(SEEK_HOLE) heavy workloads. 83 * Disabling this option will result in holes never being reported in dirty 84 * files which is always safe. 85 */ 86 static int zfs_dmu_offset_next_sync = 1; 87 88 /* 89 * Limit the amount we can prefetch with one call to this amount. This 90 * helps to limit the amount of memory that can be used by prefetching. 91 * Larger objects should be prefetched a bit at a time. 92 */ 93 #ifdef _ILP32 94 uint_t dmu_prefetch_max = 8 * 1024 * 1024; 95 #else 96 uint_t dmu_prefetch_max = 8 * SPA_MAXBLOCKSIZE; 97 #endif 98 99 /* 100 * Override copies= for dedup state objects. 0 means the traditional behaviour 101 * (ie the default for the containing objset ie 3 for the MOS). 102 */ 103 uint_t dmu_ddt_copies = 0; 104 105 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = { 106 {DMU_BSWAP_UINT8, TRUE, FALSE, FALSE, "unallocated" }, 107 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "object directory" }, 108 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "object array" }, 109 {DMU_BSWAP_UINT8, TRUE, FALSE, FALSE, "packed nvlist" }, 110 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "packed nvlist size" }, 111 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "bpobj" }, 112 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "bpobj header" }, 113 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "SPA space map header" }, 114 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "SPA space map" }, 115 {DMU_BSWAP_UINT64, TRUE, FALSE, TRUE, "ZIL intent log" }, 116 {DMU_BSWAP_DNODE, TRUE, FALSE, TRUE, "DMU dnode" }, 117 {DMU_BSWAP_OBJSET, TRUE, TRUE, FALSE, "DMU objset" }, 118 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "DSL directory" }, 119 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL directory child map"}, 120 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL dataset snap map" }, 121 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL props" }, 122 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "DSL dataset" }, 123 {DMU_BSWAP_ZNODE, TRUE, FALSE, FALSE, "ZFS znode" }, 124 {DMU_BSWAP_OLDACL, TRUE, FALSE, TRUE, "ZFS V0 ACL" }, 125 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "ZFS plain file" }, 126 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS directory" }, 127 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "ZFS master node" }, 128 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS delete queue" }, 129 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "zvol object" }, 130 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "zvol prop" }, 131 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "other uint8[]" }, 132 {DMU_BSWAP_UINT64, FALSE, FALSE, TRUE, "other uint64[]" }, 133 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "other ZAP" }, 134 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "persistent error log" }, 135 {DMU_BSWAP_UINT8, TRUE, FALSE, FALSE, "SPA history" }, 136 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "SPA history offsets" }, 137 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "Pool properties" }, 138 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL permissions" }, 139 {DMU_BSWAP_ACL, TRUE, FALSE, TRUE, "ZFS ACL" }, 140 {DMU_BSWAP_UINT8, TRUE, FALSE, TRUE, "ZFS SYSACL" }, 141 {DMU_BSWAP_UINT8, TRUE, FALSE, TRUE, "FUID table" }, 142 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "FUID table size" }, 143 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL dataset next clones"}, 144 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "scan work queue" }, 145 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS user/group/project used" }, 146 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS user/group/project quota"}, 147 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "snapshot refcount tags"}, 148 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "DDT ZAP algorithm" }, 149 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "DDT statistics" }, 150 {DMU_BSWAP_UINT8, TRUE, FALSE, TRUE, "System attributes" }, 151 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "SA master node" }, 152 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "SA attr registration" }, 153 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "SA attr layouts" }, 154 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "scan translations" }, 155 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "deduplicated block" }, 156 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL deadlist map" }, 157 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "DSL deadlist map hdr" }, 158 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL dir clones" }, 159 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "bpobj subobj" } 160 }; 161 162 dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = { 163 { byteswap_uint8_array, "uint8" }, 164 { byteswap_uint16_array, "uint16" }, 165 { byteswap_uint32_array, "uint32" }, 166 { byteswap_uint64_array, "uint64" }, 167 { zap_byteswap, "zap" }, 168 { dnode_buf_byteswap, "dnode" }, 169 { dmu_objset_byteswap, "objset" }, 170 { zfs_znode_byteswap, "znode" }, 171 { zfs_oldacl_byteswap, "oldacl" }, 172 { zfs_acl_byteswap, "acl" } 173 }; 174 175 int 176 dmu_buf_hold_noread_by_dnode(dnode_t *dn, uint64_t offset, 177 const void *tag, dmu_buf_t **dbp) 178 { 179 uint64_t blkid; 180 dmu_buf_impl_t *db; 181 182 rw_enter(&dn->dn_struct_rwlock, RW_READER); 183 blkid = dbuf_whichblock(dn, 0, offset); 184 db = dbuf_hold(dn, blkid, tag); 185 rw_exit(&dn->dn_struct_rwlock); 186 187 if (db == NULL) { 188 *dbp = NULL; 189 return (SET_ERROR(EIO)); 190 } 191 192 *dbp = &db->db; 193 return (0); 194 } 195 196 int 197 dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset, 198 const void *tag, dmu_buf_t **dbp) 199 { 200 dnode_t *dn; 201 uint64_t blkid; 202 dmu_buf_impl_t *db; 203 int err; 204 205 err = dnode_hold(os, object, FTAG, &dn); 206 if (err) 207 return (err); 208 rw_enter(&dn->dn_struct_rwlock, RW_READER); 209 blkid = dbuf_whichblock(dn, 0, offset); 210 db = dbuf_hold(dn, blkid, tag); 211 rw_exit(&dn->dn_struct_rwlock); 212 dnode_rele(dn, FTAG); 213 214 if (db == NULL) { 215 *dbp = NULL; 216 return (SET_ERROR(EIO)); 217 } 218 219 *dbp = &db->db; 220 return (err); 221 } 222 223 int 224 dmu_buf_hold_by_dnode(dnode_t *dn, uint64_t offset, 225 const void *tag, dmu_buf_t **dbp, dmu_flags_t flags) 226 { 227 int err; 228 229 err = dmu_buf_hold_noread_by_dnode(dn, offset, tag, dbp); 230 if (err == 0) { 231 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp); 232 err = dbuf_read(db, NULL, flags | DB_RF_CANFAIL); 233 if (err != 0) { 234 dbuf_rele(db, tag); 235 *dbp = NULL; 236 } 237 } 238 239 return (err); 240 } 241 242 int 243 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset, 244 const void *tag, dmu_buf_t **dbp, dmu_flags_t flags) 245 { 246 int err; 247 248 err = dmu_buf_hold_noread(os, object, offset, tag, dbp); 249 if (err == 0) { 250 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp); 251 err = dbuf_read(db, NULL, flags | DB_RF_CANFAIL); 252 if (err != 0) { 253 dbuf_rele(db, tag); 254 *dbp = NULL; 255 } 256 } 257 258 return (err); 259 } 260 261 int 262 dmu_bonus_max(void) 263 { 264 return (DN_OLD_MAX_BONUSLEN); 265 } 266 267 int 268 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx) 269 { 270 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 271 dnode_t *dn; 272 int error; 273 274 if (newsize < 0 || newsize > db_fake->db_size) 275 return (SET_ERROR(EINVAL)); 276 277 DB_DNODE_ENTER(db); 278 dn = DB_DNODE(db); 279 280 if (dn->dn_bonus != db) { 281 error = SET_ERROR(EINVAL); 282 } else { 283 dnode_setbonuslen(dn, newsize, tx); 284 error = 0; 285 } 286 287 DB_DNODE_EXIT(db); 288 return (error); 289 } 290 291 int 292 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx) 293 { 294 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 295 dnode_t *dn; 296 int error; 297 298 if (!DMU_OT_IS_VALID(type)) 299 return (SET_ERROR(EINVAL)); 300 301 DB_DNODE_ENTER(db); 302 dn = DB_DNODE(db); 303 304 if (dn->dn_bonus != db) { 305 error = SET_ERROR(EINVAL); 306 } else { 307 dnode_setbonus_type(dn, type, tx); 308 error = 0; 309 } 310 311 DB_DNODE_EXIT(db); 312 return (error); 313 } 314 315 dmu_object_type_t 316 dmu_get_bonustype(dmu_buf_t *db_fake) 317 { 318 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 319 dmu_object_type_t type; 320 321 DB_DNODE_ENTER(db); 322 type = DB_DNODE(db)->dn_bonustype; 323 DB_DNODE_EXIT(db); 324 325 return (type); 326 } 327 328 int 329 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx) 330 { 331 dnode_t *dn; 332 int error; 333 334 error = dnode_hold(os, object, FTAG, &dn); 335 dbuf_rm_spill(dn, tx); 336 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 337 dnode_rm_spill(dn, tx); 338 rw_exit(&dn->dn_struct_rwlock); 339 dnode_rele(dn, FTAG); 340 return (error); 341 } 342 343 /* 344 * Lookup and hold the bonus buffer for the provided dnode. If the dnode 345 * has not yet been allocated a new bonus dbuf a will be allocated. 346 * Returns ENOENT, EIO, or 0. 347 */ 348 int dmu_bonus_hold_by_dnode(dnode_t *dn, const void *tag, dmu_buf_t **dbp, 349 dmu_flags_t flags) 350 { 351 dmu_buf_impl_t *db; 352 int error; 353 354 rw_enter(&dn->dn_struct_rwlock, RW_READER); 355 if (dn->dn_bonus == NULL) { 356 if (!rw_tryupgrade(&dn->dn_struct_rwlock)) { 357 rw_exit(&dn->dn_struct_rwlock); 358 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 359 } 360 if (dn->dn_bonus == NULL) 361 dbuf_create_bonus(dn); 362 } 363 db = dn->dn_bonus; 364 365 /* as long as the bonus buf is held, the dnode will be held */ 366 if (zfs_refcount_add(&db->db_holds, tag) == 1) { 367 VERIFY(dnode_add_ref(dn, db)); 368 atomic_inc_32(&dn->dn_dbufs_count); 369 } 370 371 /* 372 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's 373 * hold and incrementing the dbuf count to ensure that dnode_move() sees 374 * a dnode hold for every dbuf. 375 */ 376 rw_exit(&dn->dn_struct_rwlock); 377 378 error = dbuf_read(db, NULL, flags | DB_RF_CANFAIL); 379 if (error) { 380 dnode_evict_bonus(dn); 381 dbuf_rele(db, tag); 382 *dbp = NULL; 383 return (error); 384 } 385 386 *dbp = &db->db; 387 return (0); 388 } 389 390 int 391 dmu_bonus_hold(objset_t *os, uint64_t object, const void *tag, dmu_buf_t **dbp) 392 { 393 dnode_t *dn; 394 int error; 395 396 error = dnode_hold(os, object, FTAG, &dn); 397 if (error) 398 return (error); 399 400 error = dmu_bonus_hold_by_dnode(dn, tag, dbp, DMU_READ_NO_PREFETCH); 401 dnode_rele(dn, FTAG); 402 403 return (error); 404 } 405 406 /* 407 * returns ENOENT, EIO, or 0. 408 * 409 * This interface will allocate a blank spill dbuf when a spill blk 410 * doesn't already exist on the dnode. 411 * 412 * if you only want to find an already existing spill db, then 413 * dmu_spill_hold_existing() should be used. 414 */ 415 int 416 dmu_spill_hold_by_dnode(dnode_t *dn, dmu_flags_t flags, const void *tag, 417 dmu_buf_t **dbp) 418 { 419 dmu_buf_impl_t *db = NULL; 420 int err; 421 422 if ((flags & DB_RF_HAVESTRUCT) == 0) 423 rw_enter(&dn->dn_struct_rwlock, RW_READER); 424 425 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag); 426 427 if ((flags & DB_RF_HAVESTRUCT) == 0) 428 rw_exit(&dn->dn_struct_rwlock); 429 430 if (db == NULL) { 431 *dbp = NULL; 432 return (SET_ERROR(EIO)); 433 } 434 err = dbuf_read(db, NULL, flags); 435 if (err == 0) 436 *dbp = &db->db; 437 else { 438 dbuf_rele(db, tag); 439 *dbp = NULL; 440 } 441 return (err); 442 } 443 444 int 445 dmu_spill_hold_existing(dmu_buf_t *bonus, const void *tag, dmu_buf_t **dbp) 446 { 447 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus; 448 dnode_t *dn; 449 int err; 450 451 DB_DNODE_ENTER(db); 452 dn = DB_DNODE(db); 453 454 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) { 455 err = SET_ERROR(EINVAL); 456 } else { 457 rw_enter(&dn->dn_struct_rwlock, RW_READER); 458 459 if (!dn->dn_have_spill) { 460 err = SET_ERROR(ENOENT); 461 } else { 462 err = dmu_spill_hold_by_dnode(dn, 463 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp); 464 } 465 466 rw_exit(&dn->dn_struct_rwlock); 467 } 468 469 DB_DNODE_EXIT(db); 470 return (err); 471 } 472 473 int 474 dmu_spill_hold_by_bonus(dmu_buf_t *bonus, dmu_flags_t flags, const void *tag, 475 dmu_buf_t **dbp) 476 { 477 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus; 478 int err; 479 480 DB_DNODE_ENTER(db); 481 err = dmu_spill_hold_by_dnode(DB_DNODE(db), flags, tag, dbp); 482 DB_DNODE_EXIT(db); 483 484 return (err); 485 } 486 487 /* 488 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces 489 * to take a held dnode rather than <os, object> -- the lookup is wasteful, 490 * and can induce severe lock contention when writing to several files 491 * whose dnodes are in the same block. 492 */ 493 int 494 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length, 495 boolean_t read, const void *tag, int *numbufsp, dmu_buf_t ***dbpp, 496 dmu_flags_t flags) 497 { 498 dmu_buf_t **dbp; 499 zstream_t *zs = NULL; 500 uint64_t blkid, nblks, i; 501 dmu_flags_t dbuf_flags; 502 int err; 503 zio_t *zio = NULL; 504 boolean_t missed = B_FALSE; 505 506 ASSERT(!read || length <= DMU_MAX_ACCESS); 507 508 /* 509 * Note: We directly notify the prefetch code of this read, so that 510 * we can tell it about the multi-block read. dbuf_read() only knows 511 * about the one block it is accessing. 512 */ 513 dbuf_flags = (flags & ~DMU_READ_PREFETCH) | DMU_READ_NO_PREFETCH | 514 DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT; 515 516 rw_enter(&dn->dn_struct_rwlock, RW_READER); 517 if (dn->dn_datablkshift) { 518 int blkshift = dn->dn_datablkshift; 519 nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) - 520 P2ALIGN_TYPED(offset, 1ULL << blkshift, uint64_t)) 521 >> blkshift; 522 } else { 523 if (offset + length > dn->dn_datablksz) { 524 zfs_panic_recover("zfs: accessing past end of object " 525 "%llx/%llx (size=%u access=%llu+%llu)", 526 (longlong_t)dn->dn_objset-> 527 os_dsl_dataset->ds_object, 528 (longlong_t)dn->dn_object, dn->dn_datablksz, 529 (longlong_t)offset, (longlong_t)length); 530 rw_exit(&dn->dn_struct_rwlock); 531 return (SET_ERROR(EIO)); 532 } 533 nblks = 1; 534 } 535 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP); 536 537 if (read) 538 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, 539 ZIO_FLAG_CANFAIL); 540 blkid = dbuf_whichblock(dn, 0, offset); 541 if ((flags & DMU_READ_NO_PREFETCH) == 0) { 542 /* 543 * Prepare the zfetch before initiating the demand reads, so 544 * that if multiple threads block on same indirect block, we 545 * base predictions on the original less racy request order. 546 */ 547 zs = dmu_zfetch_prepare(&dn->dn_zfetch, blkid, nblks, 548 read && !(flags & DMU_DIRECTIO), B_TRUE); 549 } 550 for (i = 0; i < nblks; i++) { 551 dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag); 552 if (db == NULL) { 553 if (zs) { 554 dmu_zfetch_run(&dn->dn_zfetch, zs, missed, 555 B_TRUE, (flags & DMU_UNCACHEDIO)); 556 } 557 rw_exit(&dn->dn_struct_rwlock); 558 dmu_buf_rele_array(dbp, nblks, tag); 559 if (read) 560 zio_nowait(zio); 561 return (SET_ERROR(EIO)); 562 } 563 564 /* 565 * Initiate async demand data read. 566 * We check the db_state after calling dbuf_read() because 567 * (1) dbuf_read() may change the state to CACHED due to a 568 * hit in the ARC, and (2) on a cache miss, a child will 569 * have been added to "zio" but not yet completed, so the 570 * state will not yet be CACHED. 571 */ 572 if (read) { 573 if (i == nblks - 1 && blkid + i < dn->dn_maxblkid && 574 offset + length < db->db.db_offset + 575 db->db.db_size) { 576 if (offset <= db->db.db_offset) 577 dbuf_flags |= DMU_PARTIAL_FIRST; 578 else 579 dbuf_flags |= DMU_PARTIAL_MORE; 580 } 581 (void) dbuf_read(db, zio, dbuf_flags); 582 if (db->db_state != DB_CACHED) 583 missed = B_TRUE; 584 } 585 dbp[i] = &db->db; 586 } 587 588 /* 589 * If we are doing O_DIRECT we still hold the dbufs, even for reads, 590 * but we do not issue any reads here. We do not want to account for 591 * writes in this case. 592 * 593 * O_DIRECT write/read accounting takes place in 594 * dmu_{write/read}_abd(). 595 */ 596 if (!read && ((flags & DMU_DIRECTIO) == 0)) 597 zfs_racct_write(dn->dn_objset->os_spa, length, nblks, flags); 598 599 if (zs) { 600 dmu_zfetch_run(&dn->dn_zfetch, zs, missed, B_TRUE, 601 (flags & DMU_UNCACHEDIO)); 602 } 603 rw_exit(&dn->dn_struct_rwlock); 604 605 if (read) { 606 /* wait for async read i/o */ 607 err = zio_wait(zio); 608 if (err) { 609 dmu_buf_rele_array(dbp, nblks, tag); 610 return (err); 611 } 612 613 /* wait for other io to complete */ 614 for (i = 0; i < nblks; i++) { 615 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i]; 616 mutex_enter(&db->db_mtx); 617 while (db->db_state == DB_READ || 618 db->db_state == DB_FILL) 619 cv_wait(&db->db_changed, &db->db_mtx); 620 if (db->db_state == DB_UNCACHED) 621 err = SET_ERROR(EIO); 622 mutex_exit(&db->db_mtx); 623 if (err) { 624 dmu_buf_rele_array(dbp, nblks, tag); 625 return (err); 626 } 627 } 628 } 629 630 *numbufsp = nblks; 631 *dbpp = dbp; 632 return (0); 633 } 634 635 int 636 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset, 637 uint64_t length, int read, const void *tag, int *numbufsp, 638 dmu_buf_t ***dbpp, dmu_flags_t flags) 639 { 640 dnode_t *dn; 641 int err; 642 643 err = dnode_hold(os, object, FTAG, &dn); 644 if (err) 645 return (err); 646 647 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag, 648 numbufsp, dbpp, flags); 649 650 dnode_rele(dn, FTAG); 651 652 return (err); 653 } 654 655 int 656 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset, 657 uint64_t length, boolean_t read, const void *tag, int *numbufsp, 658 dmu_buf_t ***dbpp, dmu_flags_t flags) 659 { 660 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 661 int err; 662 663 DB_DNODE_ENTER(db); 664 err = dmu_buf_hold_array_by_dnode(DB_DNODE(db), offset, length, read, 665 tag, numbufsp, dbpp, flags); 666 DB_DNODE_EXIT(db); 667 668 return (err); 669 } 670 671 void 672 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, const void *tag) 673 { 674 int i; 675 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake; 676 677 if (numbufs == 0) 678 return; 679 680 for (i = 0; i < numbufs; i++) { 681 if (dbp[i]) 682 dbuf_rele(dbp[i], tag); 683 } 684 685 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs); 686 } 687 688 /* 689 * Issue prefetch I/Os for the given blocks. If level is greater than 0, the 690 * indirect blocks prefetched will be those that point to the blocks containing 691 * the data starting at offset, and continuing to offset + len. If the range 692 * is too long, prefetch the first dmu_prefetch_max bytes as requested, while 693 * for the rest only a higher level, also fitting within dmu_prefetch_max. It 694 * should primarily help random reads, since for long sequential reads there is 695 * a speculative prefetcher. 696 * 697 * Note that if the indirect blocks above the blocks being prefetched are not 698 * in cache, they will be asynchronously read in. Dnode read by dnode_hold() 699 * is currently synchronous. 700 */ 701 void 702 dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset, 703 uint64_t len, zio_priority_t pri) 704 { 705 dnode_t *dn; 706 707 if (dmu_prefetch_max == 0 || len == 0) { 708 dmu_prefetch_dnode(os, object, pri); 709 return; 710 } 711 712 if (dnode_hold(os, object, FTAG, &dn) != 0) 713 return; 714 715 dmu_prefetch_by_dnode(dn, level, offset, len, pri); 716 717 dnode_rele(dn, FTAG); 718 } 719 720 void 721 dmu_prefetch_by_dnode(dnode_t *dn, int64_t level, uint64_t offset, 722 uint64_t len, zio_priority_t pri) 723 { 724 int64_t level2 = level; 725 uint64_t start, end, start2, end2; 726 727 /* 728 * Depending on len we may do two prefetches: blocks [start, end) at 729 * level, and following blocks [start2, end2) at higher level2. 730 */ 731 rw_enter(&dn->dn_struct_rwlock, RW_READER); 732 if (dn->dn_datablkshift != 0) { 733 734 /* 735 * Limit prefetch to present blocks. 736 */ 737 uint64_t size = (dn->dn_maxblkid + 1) << dn->dn_datablkshift; 738 if (offset >= size) { 739 rw_exit(&dn->dn_struct_rwlock); 740 return; 741 } 742 if (offset + len < offset || offset + len > size) 743 len = size - offset; 744 745 /* 746 * The object has multiple blocks. Calculate the full range 747 * of blocks [start, end2) and then split it into two parts, 748 * so that the first [start, end) fits into dmu_prefetch_max. 749 */ 750 start = dbuf_whichblock(dn, level, offset); 751 end2 = dbuf_whichblock(dn, level, offset + len - 1) + 1; 752 uint8_t ibs = dn->dn_indblkshift; 753 uint8_t bs = (level == 0) ? dn->dn_datablkshift : ibs; 754 uint_t limit = P2ROUNDUP(dmu_prefetch_max, 1 << bs) >> bs; 755 start2 = end = MIN(end2, start + limit); 756 757 /* 758 * Find level2 where [start2, end2) fits into dmu_prefetch_max. 759 */ 760 uint8_t ibps = ibs - SPA_BLKPTRSHIFT; 761 limit = P2ROUNDUP(dmu_prefetch_max, 1 << ibs) >> ibs; 762 if (limit == 0) 763 end2 = start2; 764 do { 765 level2++; 766 start2 = P2ROUNDUP(start2, 1 << ibps) >> ibps; 767 end2 = P2ROUNDUP(end2, 1 << ibps) >> ibps; 768 } while (end2 - start2 > limit); 769 } else { 770 /* There is only one block. Prefetch it or nothing. */ 771 start = start2 = end2 = 0; 772 end = start + (level == 0 && offset < dn->dn_datablksz); 773 } 774 775 for (uint64_t i = start; i < end; i++) 776 dbuf_prefetch(dn, level, i, pri, 0); 777 for (uint64_t i = start2; i < end2; i++) 778 dbuf_prefetch(dn, level2, i, pri, 0); 779 rw_exit(&dn->dn_struct_rwlock); 780 } 781 782 typedef struct { 783 kmutex_t dpa_lock; 784 kcondvar_t dpa_cv; 785 uint64_t dpa_pending_io; 786 } dmu_prefetch_arg_t; 787 788 static void 789 dmu_prefetch_done(void *arg, uint64_t level, uint64_t blkid, boolean_t issued) 790 { 791 (void) level; (void) blkid; (void)issued; 792 dmu_prefetch_arg_t *dpa = arg; 793 794 ASSERT0(level); 795 796 mutex_enter(&dpa->dpa_lock); 797 ASSERT3U(dpa->dpa_pending_io, >, 0); 798 if (--dpa->dpa_pending_io == 0) 799 cv_broadcast(&dpa->dpa_cv); 800 mutex_exit(&dpa->dpa_lock); 801 } 802 803 static void 804 dmu_prefetch_wait_by_dnode(dnode_t *dn, uint64_t offset, uint64_t len) 805 { 806 dmu_prefetch_arg_t dpa; 807 808 mutex_init(&dpa.dpa_lock, NULL, MUTEX_DEFAULT, NULL); 809 cv_init(&dpa.dpa_cv, NULL, CV_DEFAULT, NULL); 810 811 rw_enter(&dn->dn_struct_rwlock, RW_READER); 812 813 uint64_t start = dbuf_whichblock(dn, 0, offset); 814 uint64_t end = dbuf_whichblock(dn, 0, offset + len - 1) + 1; 815 dpa.dpa_pending_io = end - start; 816 817 for (uint64_t blk = start; blk < end; blk++) { 818 (void) dbuf_prefetch_impl(dn, 0, blk, ZIO_PRIORITY_ASYNC_READ, 819 0, dmu_prefetch_done, &dpa); 820 } 821 822 rw_exit(&dn->dn_struct_rwlock); 823 824 /* wait for prefetch L0 reads to finish */ 825 mutex_enter(&dpa.dpa_lock); 826 while (dpa.dpa_pending_io > 0) { 827 cv_wait(&dpa.dpa_cv, &dpa.dpa_lock); 828 829 } 830 mutex_exit(&dpa.dpa_lock); 831 832 mutex_destroy(&dpa.dpa_lock); 833 cv_destroy(&dpa.dpa_cv); 834 } 835 836 /* 837 * Issue prefetch I/Os for the given L0 block range and wait for the I/O 838 * to complete. This does not enforce dmu_prefetch_max and will prefetch 839 * the entire range. The blocks are read from disk into the ARC but no 840 * decompression occurs (i.e., the dbuf cache is not required). 841 */ 842 int 843 dmu_prefetch_wait(objset_t *os, uint64_t object, uint64_t offset, uint64_t size) 844 { 845 dnode_t *dn; 846 int err = 0; 847 848 err = dnode_hold(os, object, FTAG, &dn); 849 if (err != 0) 850 return (err); 851 852 /* 853 * Chunk the requests (16 indirects worth) so that we can be 854 * interrupted. Prefetch at least SPA_MAXBLOCKSIZE at a time 855 * to better utilize pools with smaller block sizes. 856 */ 857 uint64_t chunksize; 858 if (dn->dn_indblkshift) { 859 uint64_t nbps = bp_span_in_blocks(dn->dn_indblkshift, 1); 860 chunksize = (nbps * 16) << dn->dn_datablkshift; 861 chunksize = MAX(chunksize, SPA_MAXBLOCKSIZE); 862 } else { 863 chunksize = dn->dn_datablksz; 864 } 865 866 while (size > 0) { 867 uint64_t mylen = MIN(size, chunksize); 868 869 dmu_prefetch_wait_by_dnode(dn, offset, mylen); 870 871 offset += mylen; 872 size -= mylen; 873 874 if (issig()) { 875 err = SET_ERROR(EINTR); 876 break; 877 } 878 } 879 880 dnode_rele(dn, FTAG); 881 882 return (err); 883 } 884 885 /* 886 * Issue prefetch I/Os for the given object's dnode. 887 */ 888 void 889 dmu_prefetch_dnode(objset_t *os, uint64_t object, zio_priority_t pri) 890 { 891 if (object == 0 || object >= DN_MAX_OBJECT) 892 return; 893 894 dnode_t *dn = DMU_META_DNODE(os); 895 rw_enter(&dn->dn_struct_rwlock, RW_READER); 896 uint64_t blkid = dbuf_whichblock(dn, 0, object * sizeof (dnode_phys_t)); 897 dbuf_prefetch(dn, 0, blkid, pri, 0); 898 rw_exit(&dn->dn_struct_rwlock); 899 } 900 901 /* 902 * Get the next "chunk" of file data to free. We traverse the file from 903 * the end so that the file gets shorter over time (if we crash in the 904 * middle, this will leave us in a better state). We find allocated file 905 * data by simply searching the allocated level 1 indirects. 906 * 907 * On input, *start should be the first offset that does not need to be 908 * freed (e.g. "offset + length"). On return, *start will be the first 909 * offset that should be freed and l1blks is set to the number of level 1 910 * indirect blocks found within the chunk. 911 */ 912 static int 913 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum, uint64_t *l1blks) 914 { 915 uint64_t blks; 916 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1); 917 /* bytes of data covered by a level-1 indirect block */ 918 uint64_t iblkrange = (uint64_t)dn->dn_datablksz * 919 EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT); 920 921 ASSERT3U(minimum, <=, *start); 922 923 /* dn_nlevels == 1 means we don't have any L1 blocks */ 924 if (dn->dn_nlevels <= 1) { 925 *l1blks = 0; 926 *start = minimum; 927 return (0); 928 } 929 930 /* 931 * Check if we can free the entire range assuming that all of the 932 * L1 blocks in this range have data. If we can, we use this 933 * worst case value as an estimate so we can avoid having to look 934 * at the object's actual data. 935 */ 936 uint64_t total_l1blks = 937 (roundup(*start, iblkrange) - (minimum / iblkrange * iblkrange)) / 938 iblkrange; 939 if (total_l1blks <= maxblks) { 940 *l1blks = total_l1blks; 941 *start = minimum; 942 return (0); 943 } 944 ASSERT(ISP2(iblkrange)); 945 946 for (blks = 0; *start > minimum && blks < maxblks; blks++) { 947 int err; 948 949 /* 950 * dnode_next_offset(BACKWARDS) will find an allocated L1 951 * indirect block at or before the input offset. We must 952 * decrement *start so that it is at the end of the region 953 * to search. 954 */ 955 (*start)--; 956 957 err = dnode_next_offset(dn, 958 DNODE_FIND_BACKWARDS, start, 2, 1, 0); 959 960 /* if there are no indirect blocks before start, we are done */ 961 if (err == ESRCH) { 962 *start = minimum; 963 break; 964 } else if (err != 0) { 965 *l1blks = blks; 966 return (err); 967 } 968 969 /* set start to the beginning of this L1 indirect */ 970 *start = P2ALIGN_TYPED(*start, iblkrange, uint64_t); 971 } 972 if (*start < minimum) 973 *start = minimum; 974 *l1blks = blks; 975 976 return (0); 977 } 978 979 /* 980 * If this objset is of type OST_ZFS return true if vfs's unmounted flag is set, 981 * otherwise return false. 982 * Used below in dmu_free_long_range_impl() to enable abort when unmounting 983 */ 984 static boolean_t 985 dmu_objset_zfs_unmounting(objset_t *os) 986 { 987 #ifdef _KERNEL 988 if (dmu_objset_type(os) == DMU_OST_ZFS) 989 return (zfs_get_vfs_flag_unmounted(os)); 990 #else 991 (void) os; 992 #endif 993 return (B_FALSE); 994 } 995 996 static int 997 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset, 998 uint64_t length) 999 { 1000 uint64_t object_size; 1001 int err; 1002 uint64_t dirty_frees_threshold; 1003 dsl_pool_t *dp = dmu_objset_pool(os); 1004 1005 if (dn == NULL) 1006 return (SET_ERROR(EINVAL)); 1007 1008 object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz; 1009 if (offset >= object_size) 1010 return (0); 1011 1012 if (zfs_per_txg_dirty_frees_percent <= 100) 1013 dirty_frees_threshold = 1014 zfs_per_txg_dirty_frees_percent * zfs_dirty_data_max / 100; 1015 else 1016 dirty_frees_threshold = zfs_dirty_data_max / 20; 1017 1018 if (length == DMU_OBJECT_END || offset + length > object_size) 1019 length = object_size - offset; 1020 1021 while (length != 0) { 1022 uint64_t chunk_end, chunk_begin, chunk_len; 1023 uint64_t l1blks; 1024 dmu_tx_t *tx; 1025 1026 if (dmu_objset_zfs_unmounting(dn->dn_objset)) 1027 return (SET_ERROR(EINTR)); 1028 1029 chunk_end = chunk_begin = offset + length; 1030 1031 /* move chunk_begin backwards to the beginning of this chunk */ 1032 err = get_next_chunk(dn, &chunk_begin, offset, &l1blks); 1033 if (err) 1034 return (err); 1035 ASSERT3U(chunk_begin, >=, offset); 1036 ASSERT3U(chunk_begin, <=, chunk_end); 1037 1038 chunk_len = chunk_end - chunk_begin; 1039 1040 tx = dmu_tx_create(os); 1041 dmu_tx_hold_free(tx, dn->dn_object, chunk_begin, chunk_len); 1042 1043 /* 1044 * Mark this transaction as typically resulting in a net 1045 * reduction in space used. 1046 */ 1047 dmu_tx_mark_netfree(tx); 1048 err = dmu_tx_assign(tx, DMU_TX_WAIT); 1049 if (err) { 1050 dmu_tx_abort(tx); 1051 return (err); 1052 } 1053 1054 uint64_t txg = dmu_tx_get_txg(tx); 1055 1056 mutex_enter(&dp->dp_lock); 1057 uint64_t long_free_dirty = 1058 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK]; 1059 mutex_exit(&dp->dp_lock); 1060 1061 /* 1062 * To avoid filling up a TXG with just frees, wait for 1063 * the next TXG to open before freeing more chunks if 1064 * we have reached the threshold of frees. 1065 */ 1066 if (dirty_frees_threshold != 0 && 1067 long_free_dirty >= dirty_frees_threshold) { 1068 DMU_TX_STAT_BUMP(dmu_tx_dirty_frees_delay); 1069 dmu_tx_commit(tx); 1070 txg_wait_open(dp, 0, B_TRUE); 1071 continue; 1072 } 1073 1074 /* 1075 * In order to prevent unnecessary write throttling, for each 1076 * TXG, we track the cumulative size of L1 blocks being dirtied 1077 * in dnode_free_range() below. We compare this number to a 1078 * tunable threshold, past which we prevent new L1 dirty freeing 1079 * blocks from being added into the open TXG. See 1080 * dmu_free_long_range_impl() for details. The threshold 1081 * prevents write throttle activation due to dirty freeing L1 1082 * blocks taking up a large percentage of zfs_dirty_data_max. 1083 */ 1084 mutex_enter(&dp->dp_lock); 1085 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] += 1086 l1blks << dn->dn_indblkshift; 1087 mutex_exit(&dp->dp_lock); 1088 DTRACE_PROBE3(free__long__range, 1089 uint64_t, long_free_dirty, uint64_t, chunk_len, 1090 uint64_t, txg); 1091 dnode_free_range(dn, chunk_begin, chunk_len, tx); 1092 1093 dmu_tx_commit(tx); 1094 1095 length -= chunk_len; 1096 } 1097 return (0); 1098 } 1099 1100 int 1101 dmu_free_long_range(objset_t *os, uint64_t object, 1102 uint64_t offset, uint64_t length) 1103 { 1104 dnode_t *dn; 1105 int err; 1106 1107 err = dnode_hold(os, object, FTAG, &dn); 1108 if (err != 0) 1109 return (err); 1110 err = dmu_free_long_range_impl(os, dn, offset, length); 1111 1112 /* 1113 * It is important to zero out the maxblkid when freeing the entire 1114 * file, so that (a) subsequent calls to dmu_free_long_range_impl() 1115 * will take the fast path, and (b) dnode_reallocate() can verify 1116 * that the entire file has been freed. 1117 */ 1118 if (err == 0 && offset == 0 && length == DMU_OBJECT_END) 1119 dn->dn_maxblkid = 0; 1120 1121 dnode_rele(dn, FTAG); 1122 return (err); 1123 } 1124 1125 int 1126 dmu_free_long_object(objset_t *os, uint64_t object) 1127 { 1128 dmu_tx_t *tx; 1129 int err; 1130 1131 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END); 1132 if (err != 0) 1133 return (err); 1134 1135 tx = dmu_tx_create(os); 1136 dmu_tx_hold_bonus(tx, object); 1137 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END); 1138 dmu_tx_mark_netfree(tx); 1139 err = dmu_tx_assign(tx, DMU_TX_WAIT); 1140 if (err == 0) { 1141 err = dmu_object_free(os, object, tx); 1142 dmu_tx_commit(tx); 1143 } else { 1144 dmu_tx_abort(tx); 1145 } 1146 1147 return (err); 1148 } 1149 1150 int 1151 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset, 1152 uint64_t size, dmu_tx_t *tx) 1153 { 1154 dnode_t *dn; 1155 int err = dnode_hold(os, object, FTAG, &dn); 1156 if (err) 1157 return (err); 1158 ASSERT(offset < UINT64_MAX); 1159 ASSERT(size == DMU_OBJECT_END || size <= UINT64_MAX - offset); 1160 dnode_free_range(dn, offset, size, tx); 1161 dnode_rele(dn, FTAG); 1162 return (0); 1163 } 1164 1165 static int 1166 dmu_read_impl(dnode_t *dn, uint64_t offset, uint64_t size, 1167 void *buf, dmu_flags_t flags) 1168 { 1169 dmu_buf_t **dbp; 1170 int numbufs, err = 0; 1171 1172 /* 1173 * Deal with odd block sizes, where there can't be data past the first 1174 * block. If we ever do the tail block optimization, we will need to 1175 * handle that here as well. 1176 */ 1177 if (dn->dn_maxblkid == 0) { 1178 uint64_t newsz = offset > dn->dn_datablksz ? 0 : 1179 MIN(size, dn->dn_datablksz - offset); 1180 memset((char *)buf + newsz, 0, size - newsz); 1181 size = newsz; 1182 } 1183 1184 if (size == 0) 1185 return (0); 1186 1187 /* Allow Direct I/O when requested and properly aligned */ 1188 if ((flags & DMU_DIRECTIO) && zfs_dio_page_aligned(buf) && 1189 zfs_dio_aligned(offset, size, PAGESIZE)) { 1190 abd_t *data = abd_get_from_buf(buf, size); 1191 err = dmu_read_abd(dn, offset, size, data, flags); 1192 abd_free(data); 1193 return (err); 1194 } 1195 flags &= ~DMU_DIRECTIO; 1196 1197 while (size > 0) { 1198 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2); 1199 int i; 1200 1201 /* 1202 * NB: we could do this block-at-a-time, but it's nice 1203 * to be reading in parallel. 1204 */ 1205 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen, 1206 TRUE, FTAG, &numbufs, &dbp, flags); 1207 if (err) 1208 break; 1209 1210 for (i = 0; i < numbufs; i++) { 1211 uint64_t tocpy; 1212 int64_t bufoff; 1213 dmu_buf_t *db = dbp[i]; 1214 1215 ASSERT(size > 0); 1216 1217 bufoff = offset - db->db_offset; 1218 tocpy = MIN(db->db_size - bufoff, size); 1219 1220 ASSERT(db->db_data != NULL); 1221 (void) memcpy(buf, (char *)db->db_data + bufoff, tocpy); 1222 1223 offset += tocpy; 1224 size -= tocpy; 1225 buf = (char *)buf + tocpy; 1226 } 1227 dmu_buf_rele_array(dbp, numbufs, FTAG); 1228 } 1229 return (err); 1230 } 1231 1232 int 1233 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1234 void *buf, dmu_flags_t flags) 1235 { 1236 dnode_t *dn; 1237 int err; 1238 1239 err = dnode_hold(os, object, FTAG, &dn); 1240 if (err != 0) 1241 return (err); 1242 1243 err = dmu_read_impl(dn, offset, size, buf, flags); 1244 dnode_rele(dn, FTAG); 1245 return (err); 1246 } 1247 1248 int 1249 dmu_read_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, void *buf, 1250 dmu_flags_t flags) 1251 { 1252 return (dmu_read_impl(dn, offset, size, buf, flags)); 1253 } 1254 1255 static void 1256 dmu_write_impl(dmu_buf_t **dbp, int numbufs, uint64_t offset, uint64_t size, 1257 const void *buf, dmu_tx_t *tx, dmu_flags_t flags) 1258 { 1259 int i; 1260 1261 for (i = 0; i < numbufs; i++) { 1262 uint64_t tocpy; 1263 int64_t bufoff; 1264 dmu_buf_t *db = dbp[i]; 1265 1266 ASSERT(size > 0); 1267 1268 bufoff = offset - db->db_offset; 1269 tocpy = MIN(db->db_size - bufoff, size); 1270 1271 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 1272 1273 if (tocpy == db->db_size) { 1274 dmu_buf_will_fill_flags(db, tx, B_FALSE, flags); 1275 } else { 1276 if (i == numbufs - 1 && bufoff + tocpy < db->db_size) { 1277 if (bufoff == 0) 1278 flags |= DMU_PARTIAL_FIRST; 1279 else 1280 flags |= DMU_PARTIAL_MORE; 1281 } 1282 dmu_buf_will_dirty_flags(db, tx, flags); 1283 } 1284 1285 ASSERT(db->db_data != NULL); 1286 (void) memcpy((char *)db->db_data + bufoff, buf, tocpy); 1287 1288 if (tocpy == db->db_size) 1289 dmu_buf_fill_done(db, tx, B_FALSE); 1290 1291 offset += tocpy; 1292 size -= tocpy; 1293 buf = (char *)buf + tocpy; 1294 } 1295 } 1296 1297 void 1298 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1299 const void *buf, dmu_tx_t *tx, dmu_flags_t flags) 1300 { 1301 dmu_buf_t **dbp; 1302 int numbufs; 1303 1304 if (size == 0) 1305 return; 1306 1307 VERIFY0(dmu_buf_hold_array(os, object, offset, size, 1308 FALSE, FTAG, &numbufs, &dbp, flags)); 1309 dmu_write_impl(dbp, numbufs, offset, size, buf, tx, flags); 1310 dmu_buf_rele_array(dbp, numbufs, FTAG); 1311 } 1312 1313 int 1314 dmu_write_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, 1315 const void *buf, dmu_tx_t *tx, dmu_flags_t flags) 1316 { 1317 dmu_buf_t **dbp; 1318 int numbufs; 1319 int error; 1320 1321 if (size == 0) 1322 return (0); 1323 1324 /* Allow Direct I/O when requested and properly aligned */ 1325 if ((flags & DMU_DIRECTIO) && zfs_dio_page_aligned((void *)buf) && 1326 zfs_dio_aligned(offset, size, dn->dn_datablksz)) { 1327 abd_t *data = abd_get_from_buf((void *)buf, size); 1328 error = dmu_write_abd(dn, offset, size, data, flags, tx); 1329 abd_free(data); 1330 return (error); 1331 } 1332 flags &= ~DMU_DIRECTIO; 1333 1334 VERIFY0(dmu_buf_hold_array_by_dnode(dn, offset, size, 1335 FALSE, FTAG, &numbufs, &dbp, flags)); 1336 dmu_write_impl(dbp, numbufs, offset, size, buf, tx, flags); 1337 dmu_buf_rele_array(dbp, numbufs, FTAG); 1338 return (0); 1339 } 1340 1341 void 1342 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1343 dmu_tx_t *tx) 1344 { 1345 dmu_buf_t **dbp; 1346 int numbufs, i; 1347 1348 if (size == 0) 1349 return; 1350 1351 VERIFY0(dmu_buf_hold_array(os, object, offset, size, 1352 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH)); 1353 1354 for (i = 0; i < numbufs; i++) { 1355 dmu_buf_t *db = dbp[i]; 1356 1357 dmu_buf_will_not_fill(db, tx); 1358 } 1359 dmu_buf_rele_array(dbp, numbufs, FTAG); 1360 } 1361 1362 void 1363 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset, 1364 void *data, uint8_t etype, uint8_t comp, int uncompressed_size, 1365 int compressed_size, int byteorder, dmu_tx_t *tx) 1366 { 1367 dmu_buf_t *db; 1368 1369 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES); 1370 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS); 1371 VERIFY0(dmu_buf_hold_noread(os, object, offset, 1372 FTAG, &db)); 1373 1374 dmu_buf_write_embedded(db, 1375 data, (bp_embedded_type_t)etype, (enum zio_compress)comp, 1376 uncompressed_size, compressed_size, byteorder, tx); 1377 1378 dmu_buf_rele(db, FTAG); 1379 } 1380 1381 void 1382 dmu_redact(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1383 dmu_tx_t *tx) 1384 { 1385 int numbufs, i; 1386 dmu_buf_t **dbp; 1387 1388 VERIFY0(dmu_buf_hold_array(os, object, offset, size, FALSE, FTAG, 1389 &numbufs, &dbp, DMU_READ_PREFETCH)); 1390 for (i = 0; i < numbufs; i++) 1391 dmu_buf_redact(dbp[i], tx); 1392 dmu_buf_rele_array(dbp, numbufs, FTAG); 1393 } 1394 1395 #ifdef _KERNEL 1396 int 1397 dmu_read_uio_dnode(dnode_t *dn, zfs_uio_t *uio, uint64_t size, 1398 dmu_flags_t flags) 1399 { 1400 dmu_buf_t **dbp; 1401 int numbufs, i, err; 1402 1403 if ((flags & DMU_DIRECTIO) && (uio->uio_extflg & UIO_DIRECT)) 1404 return (dmu_read_uio_direct(dn, uio, size, flags)); 1405 flags &= ~DMU_DIRECTIO; 1406 1407 /* 1408 * NB: we could do this block-at-a-time, but it's nice 1409 * to be reading in parallel. 1410 */ 1411 err = dmu_buf_hold_array_by_dnode(dn, zfs_uio_offset(uio), size, 1412 TRUE, FTAG, &numbufs, &dbp, flags); 1413 if (err) 1414 return (err); 1415 1416 for (i = 0; i < numbufs; i++) { 1417 uint64_t tocpy; 1418 int64_t bufoff; 1419 dmu_buf_t *db = dbp[i]; 1420 1421 ASSERT(size > 0); 1422 1423 bufoff = zfs_uio_offset(uio) - db->db_offset; 1424 tocpy = MIN(db->db_size - bufoff, size); 1425 1426 ASSERT(db->db_data != NULL); 1427 err = zfs_uio_fault_move((char *)db->db_data + bufoff, tocpy, 1428 UIO_READ, uio); 1429 1430 if (err) 1431 break; 1432 1433 size -= tocpy; 1434 } 1435 dmu_buf_rele_array(dbp, numbufs, FTAG); 1436 1437 return (err); 1438 } 1439 1440 /* 1441 * Read 'size' bytes into the uio buffer. 1442 * From object zdb->db_object. 1443 * Starting at zfs_uio_offset(uio). 1444 * 1445 * If the caller already has a dbuf in the target object 1446 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(), 1447 * because we don't have to find the dnode_t for the object. 1448 */ 1449 int 1450 dmu_read_uio_dbuf(dmu_buf_t *zdb, zfs_uio_t *uio, uint64_t size, 1451 dmu_flags_t flags) 1452 { 1453 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb; 1454 int err; 1455 1456 if (size == 0) 1457 return (0); 1458 1459 DB_DNODE_ENTER(db); 1460 err = dmu_read_uio_dnode(DB_DNODE(db), uio, size, flags); 1461 DB_DNODE_EXIT(db); 1462 1463 return (err); 1464 } 1465 1466 /* 1467 * Read 'size' bytes into the uio buffer. 1468 * From the specified object 1469 * Starting at offset zfs_uio_offset(uio). 1470 */ 1471 int 1472 dmu_read_uio(objset_t *os, uint64_t object, zfs_uio_t *uio, uint64_t size, 1473 dmu_flags_t flags) 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, flags); 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 dmu_flags_t flags) 1495 { 1496 dmu_buf_t **dbp; 1497 int numbufs; 1498 int err = 0; 1499 uint64_t write_size; 1500 dmu_flags_t oflags = flags; 1501 1502 top: 1503 write_size = size; 1504 1505 /* 1506 * We only allow Direct I/O writes to happen if we are block 1507 * sized aligned. Otherwise, we pass the write off to the ARC. 1508 */ 1509 if ((flags & DMU_DIRECTIO) && (uio->uio_extflg & UIO_DIRECT) && 1510 (write_size >= dn->dn_datablksz)) { 1511 if (zfs_dio_aligned(zfs_uio_offset(uio), write_size, 1512 dn->dn_datablksz)) { 1513 return (dmu_write_uio_direct(dn, uio, size, flags, tx)); 1514 } else if (write_size > dn->dn_datablksz && 1515 zfs_dio_offset_aligned(zfs_uio_offset(uio), 1516 dn->dn_datablksz)) { 1517 write_size = 1518 dn->dn_datablksz * (write_size / dn->dn_datablksz); 1519 err = dmu_write_uio_direct(dn, uio, write_size, flags, 1520 tx); 1521 if (err == 0) { 1522 size -= write_size; 1523 goto top; 1524 } else { 1525 return (err); 1526 } 1527 } else { 1528 write_size = 1529 P2PHASE(zfs_uio_offset(uio), dn->dn_datablksz); 1530 } 1531 } 1532 flags &= ~DMU_DIRECTIO; 1533 1534 err = dmu_buf_hold_array_by_dnode(dn, zfs_uio_offset(uio), write_size, 1535 FALSE, FTAG, &numbufs, &dbp, flags); 1536 if (err) 1537 return (err); 1538 1539 for (int i = 0; i < numbufs; i++) { 1540 uint64_t tocpy; 1541 int64_t bufoff; 1542 dmu_buf_t *db = dbp[i]; 1543 1544 ASSERT(write_size > 0); 1545 1546 offset_t off = zfs_uio_offset(uio); 1547 bufoff = off - db->db_offset; 1548 tocpy = MIN(db->db_size - bufoff, write_size); 1549 1550 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 1551 1552 if (tocpy == db->db_size) { 1553 dmu_buf_will_fill_flags(db, tx, B_TRUE, flags); 1554 } else { 1555 if (i == numbufs - 1 && bufoff + tocpy < db->db_size) { 1556 if (bufoff == 0) 1557 flags |= DMU_PARTIAL_FIRST; 1558 else 1559 flags |= DMU_PARTIAL_MORE; 1560 } 1561 dmu_buf_will_dirty_flags(db, tx, flags); 1562 } 1563 1564 ASSERT(db->db_data != NULL); 1565 err = zfs_uio_fault_move((char *)db->db_data + bufoff, 1566 tocpy, UIO_WRITE, uio); 1567 1568 if (tocpy == db->db_size && dmu_buf_fill_done(db, tx, err)) { 1569 /* The fill was reverted. Undo any uio progress. */ 1570 zfs_uio_advance(uio, off - zfs_uio_offset(uio)); 1571 } 1572 1573 if (err) 1574 break; 1575 1576 write_size -= tocpy; 1577 size -= tocpy; 1578 } 1579 1580 IMPLY(err == 0, write_size == 0); 1581 1582 dmu_buf_rele_array(dbp, numbufs, FTAG); 1583 1584 if ((oflags & DMU_DIRECTIO) && (uio->uio_extflg & UIO_DIRECT) && 1585 err == 0 && size > 0) { 1586 flags = oflags; 1587 goto top; 1588 } 1589 IMPLY(err == 0, size == 0); 1590 1591 return (err); 1592 } 1593 1594 /* 1595 * Write 'size' bytes from the uio buffer. 1596 * To object zdb->db_object. 1597 * Starting at offset zfs_uio_offset(uio). 1598 * 1599 * If the caller already has a dbuf in the target object 1600 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(), 1601 * because we don't have to find the dnode_t for the object. 1602 */ 1603 int 1604 dmu_write_uio_dbuf(dmu_buf_t *zdb, zfs_uio_t *uio, uint64_t size, 1605 dmu_tx_t *tx, dmu_flags_t flags) 1606 { 1607 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb; 1608 int err; 1609 1610 if (size == 0) 1611 return (0); 1612 1613 DB_DNODE_ENTER(db); 1614 err = dmu_write_uio_dnode(DB_DNODE(db), uio, size, tx, flags); 1615 DB_DNODE_EXIT(db); 1616 1617 return (err); 1618 } 1619 1620 /* 1621 * Write 'size' bytes from the uio buffer. 1622 * To the specified object. 1623 * Starting at offset zfs_uio_offset(uio). 1624 */ 1625 int 1626 dmu_write_uio(objset_t *os, uint64_t object, zfs_uio_t *uio, uint64_t size, 1627 dmu_tx_t *tx, dmu_flags_t flags) 1628 { 1629 dnode_t *dn; 1630 int err; 1631 1632 if (size == 0) 1633 return (0); 1634 1635 err = dnode_hold(os, object, FTAG, &dn); 1636 if (err) 1637 return (err); 1638 1639 err = dmu_write_uio_dnode(dn, uio, size, tx, flags); 1640 1641 dnode_rele(dn, FTAG); 1642 1643 return (err); 1644 } 1645 #endif /* _KERNEL */ 1646 1647 static void 1648 dmu_cached_bps(spa_t *spa, blkptr_t *bps, uint_t nbps, 1649 uint64_t *l1sz, uint64_t *l2sz) 1650 { 1651 int cached_flags; 1652 1653 if (bps == NULL) 1654 return; 1655 1656 for (size_t blk_off = 0; blk_off < nbps; blk_off++) { 1657 blkptr_t *bp = &bps[blk_off]; 1658 1659 if (BP_IS_HOLE(bp)) 1660 continue; 1661 1662 cached_flags = arc_cached(spa, bp); 1663 if (cached_flags == 0) 1664 continue; 1665 1666 if ((cached_flags & (ARC_CACHED_IN_L1 | ARC_CACHED_IN_L2)) == 1667 ARC_CACHED_IN_L2) 1668 *l2sz += BP_GET_LSIZE(bp); 1669 else 1670 *l1sz += BP_GET_LSIZE(bp); 1671 } 1672 } 1673 1674 /* 1675 * Estimate DMU object cached size. 1676 */ 1677 int 1678 dmu_object_cached_size(objset_t *os, uint64_t object, 1679 uint64_t *l1sz, uint64_t *l2sz) 1680 { 1681 dnode_t *dn; 1682 dmu_object_info_t doi; 1683 int err = 0; 1684 1685 *l1sz = *l2sz = 0; 1686 1687 if (dnode_hold(os, object, FTAG, &dn) != 0) 1688 return (0); 1689 1690 if (dn->dn_nlevels < 2) { 1691 dnode_rele(dn, FTAG); 1692 return (0); 1693 } 1694 1695 dmu_object_info_from_dnode(dn, &doi); 1696 1697 for (uint64_t off = 0; off < doi.doi_max_offset && 1698 dmu_prefetch_max > 0; off += dmu_prefetch_max) { 1699 /* dbuf_read doesn't prefetch L1 blocks. */ 1700 dmu_prefetch_by_dnode(dn, 1, off, 1701 dmu_prefetch_max, ZIO_PRIORITY_SYNC_READ); 1702 } 1703 1704 /* 1705 * Hold all valid L1 blocks, asking ARC the status of each BP 1706 * contained in each such L1 block. 1707 */ 1708 uint_t nbps = bp_span_in_blocks(dn->dn_indblkshift, 1); 1709 uint64_t l1blks = 1 + (dn->dn_maxblkid / nbps); 1710 1711 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1712 for (uint64_t blk = 0; blk < l1blks; blk++) { 1713 dmu_buf_impl_t *db = NULL; 1714 1715 if (issig()) { 1716 /* 1717 * On interrupt, get out, and bubble up EINTR 1718 */ 1719 err = EINTR; 1720 break; 1721 } 1722 1723 /* 1724 * If we get an i/o error here, the L1 can't be read, 1725 * and nothing under it could be cached, so we just 1726 * continue. Ignoring the error from dbuf_hold_impl 1727 * or from dbuf_read is then a reasonable choice. 1728 */ 1729 err = dbuf_hold_impl(dn, 1, blk, B_TRUE, B_FALSE, FTAG, &db); 1730 if (err != 0) { 1731 /* 1732 * ignore error and continue 1733 */ 1734 err = 0; 1735 continue; 1736 } 1737 1738 err = dbuf_read(db, NULL, DB_RF_CANFAIL); 1739 if (err == 0) { 1740 dmu_cached_bps(dmu_objset_spa(os), db->db.db_data, 1741 nbps, l1sz, l2sz); 1742 } 1743 /* 1744 * error may be ignored, and we continue 1745 */ 1746 err = 0; 1747 dbuf_rele(db, FTAG); 1748 } 1749 rw_exit(&dn->dn_struct_rwlock); 1750 1751 dnode_rele(dn, FTAG); 1752 return (err); 1753 } 1754 1755 /* 1756 * Allocate a loaned anonymous arc buffer. 1757 */ 1758 arc_buf_t * 1759 dmu_request_arcbuf(dmu_buf_t *handle, int size) 1760 { 1761 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle; 1762 1763 return (arc_loan_buf(db->db_objset->os_spa, B_FALSE, size)); 1764 } 1765 1766 /* 1767 * Free a loaned arc buffer. 1768 */ 1769 void 1770 dmu_return_arcbuf(arc_buf_t *buf) 1771 { 1772 arc_return_buf(buf, FTAG); 1773 arc_buf_destroy(buf, FTAG); 1774 } 1775 1776 /* 1777 * A "lightweight" write is faster than a regular write (e.g. 1778 * dmu_write_by_dnode() or dmu_assign_arcbuf_by_dnode()), because it avoids the 1779 * CPU cost of creating a dmu_buf_impl_t and arc_buf_[hdr_]_t. However, the 1780 * data can not be read or overwritten until the transaction's txg has been 1781 * synced. This makes it appropriate for workloads that are known to be 1782 * (temporarily) write-only, like "zfs receive". 1783 * 1784 * A single block is written, starting at the specified offset in bytes. If 1785 * the call is successful, it returns 0 and the provided abd has been 1786 * consumed (the caller should not free it). 1787 */ 1788 int 1789 dmu_lightweight_write_by_dnode(dnode_t *dn, uint64_t offset, abd_t *abd, 1790 const zio_prop_t *zp, zio_flag_t flags, dmu_tx_t *tx) 1791 { 1792 dbuf_dirty_record_t *dr = 1793 dbuf_dirty_lightweight(dn, dbuf_whichblock(dn, 0, offset), tx); 1794 if (dr == NULL) 1795 return (SET_ERROR(EIO)); 1796 dr->dt.dll.dr_abd = abd; 1797 dr->dt.dll.dr_props = *zp; 1798 dr->dt.dll.dr_flags = flags; 1799 return (0); 1800 } 1801 1802 /* 1803 * When possible directly assign passed loaned arc buffer to a dbuf. 1804 * If this is not possible copy the contents of passed arc buf via 1805 * dmu_write(). 1806 */ 1807 int 1808 dmu_assign_arcbuf_by_dnode(dnode_t *dn, uint64_t offset, arc_buf_t *buf, 1809 dmu_tx_t *tx, dmu_flags_t flags) 1810 { 1811 dmu_buf_impl_t *db; 1812 objset_t *os = dn->dn_objset; 1813 uint32_t blksz = (uint32_t)arc_buf_lsize(buf); 1814 uint64_t blkid; 1815 1816 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1817 blkid = dbuf_whichblock(dn, 0, offset); 1818 db = dbuf_hold(dn, blkid, FTAG); 1819 rw_exit(&dn->dn_struct_rwlock); 1820 if (db == NULL) 1821 return (SET_ERROR(EIO)); 1822 1823 /* 1824 * We can only assign if the offset is aligned and the arc buf is the 1825 * same size as the dbuf. 1826 */ 1827 if (offset == db->db.db_offset && blksz == db->db.db_size) { 1828 zfs_racct_write(os->os_spa, blksz, 1, flags); 1829 dbuf_assign_arcbuf(db, buf, tx, flags); 1830 dbuf_rele(db, FTAG); 1831 } else { 1832 /* compressed bufs must always be assignable to their dbuf */ 1833 ASSERT3U(arc_get_compression(buf), ==, ZIO_COMPRESS_OFF); 1834 ASSERT(!(buf->b_flags & ARC_BUF_FLAG_COMPRESSED)); 1835 1836 dbuf_rele(db, FTAG); 1837 dmu_write_by_dnode(dn, offset, blksz, buf->b_data, tx, flags); 1838 dmu_return_arcbuf(buf); 1839 } 1840 1841 return (0); 1842 } 1843 1844 int 1845 dmu_assign_arcbuf_by_dbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf, 1846 dmu_tx_t *tx, dmu_flags_t flags) 1847 { 1848 int err; 1849 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle; 1850 1851 DB_DNODE_ENTER(db); 1852 err = dmu_assign_arcbuf_by_dnode(DB_DNODE(db), offset, buf, tx, flags); 1853 DB_DNODE_EXIT(db); 1854 1855 return (err); 1856 } 1857 1858 void 1859 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg) 1860 { 1861 (void) buf; 1862 dmu_sync_arg_t *dsa = varg; 1863 1864 if (zio->io_error == 0) { 1865 dbuf_dirty_record_t *dr = dsa->dsa_dr; 1866 blkptr_t *bp = zio->io_bp; 1867 1868 if (BP_IS_HOLE(bp)) { 1869 dmu_buf_t *db = NULL; 1870 if (dr) 1871 db = &(dr->dr_dbuf->db); 1872 else 1873 db = dsa->dsa_zgd->zgd_db; 1874 /* 1875 * A block of zeros may compress to a hole, but the 1876 * block size still needs to be known for replay. 1877 */ 1878 BP_SET_LSIZE(bp, db->db_size); 1879 } else if (!BP_IS_EMBEDDED(bp)) { 1880 ASSERT0(BP_GET_LEVEL(bp)); 1881 BP_SET_FILL(bp, 1); 1882 } 1883 } 1884 } 1885 1886 static void 1887 dmu_sync_late_arrival_ready(zio_t *zio) 1888 { 1889 dmu_sync_ready(zio, NULL, zio->io_private); 1890 } 1891 1892 void 1893 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg) 1894 { 1895 (void) buf; 1896 dmu_sync_arg_t *dsa = varg; 1897 dbuf_dirty_record_t *dr = dsa->dsa_dr; 1898 dmu_buf_impl_t *db = dr->dr_dbuf; 1899 zgd_t *zgd = dsa->dsa_zgd; 1900 1901 /* 1902 * Record the vdev(s) backing this blkptr so they can be flushed after 1903 * the writes for the lwb have completed. 1904 */ 1905 if (zgd && zio->io_error == 0) { 1906 zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp); 1907 } 1908 1909 mutex_enter(&db->db_mtx); 1910 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC); 1911 if (zio->io_error == 0) { 1912 ASSERT0(dr->dt.dl.dr_has_raw_params); 1913 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE); 1914 if (dr->dt.dl.dr_nopwrite) { 1915 blkptr_t *bp = zio->io_bp; 1916 blkptr_t *bp_orig = &zio->io_bp_orig; 1917 uint8_t chksum = BP_GET_CHECKSUM(bp_orig); 1918 1919 ASSERT(BP_EQUAL(bp, bp_orig)); 1920 VERIFY(BP_EQUAL(bp, db->db_blkptr)); 1921 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF); 1922 VERIFY(zio_checksum_table[chksum].ci_flags & 1923 ZCHECKSUM_FLAG_NOPWRITE); 1924 } 1925 dr->dt.dl.dr_overridden_by = *zio->io_bp; 1926 dr->dt.dl.dr_override_state = DR_OVERRIDDEN; 1927 dr->dt.dl.dr_copies = zio->io_prop.zp_copies; 1928 dr->dt.dl.dr_gang_copies = zio->io_prop.zp_gang_copies; 1929 1930 /* 1931 * Old style holes are filled with all zeros, whereas 1932 * new-style holes maintain their lsize, type, level, 1933 * and birth time (see zio_write_compress). While we 1934 * need to reset the BP_SET_LSIZE() call that happened 1935 * in dmu_sync_ready for old style holes, we do *not* 1936 * want to wipe out the information contained in new 1937 * style holes. Thus, only zero out the block pointer if 1938 * it's an old style hole. 1939 */ 1940 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) && 1941 BP_GET_LOGICAL_BIRTH(&dr->dt.dl.dr_overridden_by) == 0) 1942 BP_ZERO(&dr->dt.dl.dr_overridden_by); 1943 } else { 1944 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 1945 } 1946 1947 cv_broadcast(&db->db_changed); 1948 mutex_exit(&db->db_mtx); 1949 1950 if (dsa->dsa_done) 1951 dsa->dsa_done(dsa->dsa_zgd, zio->io_error); 1952 1953 kmem_free(dsa, sizeof (*dsa)); 1954 } 1955 1956 static void 1957 dmu_sync_late_arrival_done(zio_t *zio) 1958 { 1959 blkptr_t *bp = zio->io_bp; 1960 dmu_sync_arg_t *dsa = zio->io_private; 1961 zgd_t *zgd = dsa->dsa_zgd; 1962 1963 if (zio->io_error == 0) { 1964 /* 1965 * Record the vdev(s) backing this blkptr so they can be 1966 * flushed after the writes for the lwb have completed. 1967 */ 1968 zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp); 1969 1970 if (!BP_IS_HOLE(bp)) { 1971 blkptr_t *bp_orig __maybe_unused = &zio->io_bp_orig; 1972 ASSERT(!(zio->io_flags & ZIO_FLAG_NOPWRITE)); 1973 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig)); 1974 ASSERT(BP_GET_BIRTH(zio->io_bp) == zio->io_txg); 1975 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa)); 1976 zio_free(zio->io_spa, zio->io_txg, zio->io_bp); 1977 } 1978 } 1979 1980 dmu_tx_commit(dsa->dsa_tx); 1981 1982 dsa->dsa_done(dsa->dsa_zgd, zio->io_error); 1983 1984 abd_free(zio->io_abd); 1985 kmem_free(dsa, sizeof (*dsa)); 1986 } 1987 1988 static int 1989 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd, 1990 zio_prop_t *zp, zbookmark_phys_t *zb) 1991 { 1992 dmu_sync_arg_t *dsa; 1993 dmu_tx_t *tx; 1994 int error; 1995 1996 error = dbuf_read((dmu_buf_impl_t *)zgd->zgd_db, NULL, 1997 DB_RF_CANFAIL | DMU_READ_NO_PREFETCH | DMU_KEEP_CACHING); 1998 if (error != 0) 1999 return (error); 2000 2001 tx = dmu_tx_create(os); 2002 dmu_tx_hold_space(tx, zgd->zgd_db->db_size); 2003 /* 2004 * This transaction does not produce any dirty data or log blocks, so 2005 * it should not be throttled. All other cases wait for TXG sync, by 2006 * which time the log block we are writing will be obsolete, so we can 2007 * skip waiting and just return error here instead. 2008 */ 2009 if (dmu_tx_assign(tx, DMU_TX_NOWAIT | DMU_TX_NOTHROTTLE) != 0) { 2010 dmu_tx_abort(tx); 2011 /* Make zl_get_data do txg_waited_synced() */ 2012 return (SET_ERROR(EIO)); 2013 } 2014 2015 /* 2016 * In order to prevent the zgd's lwb from being free'd prior to 2017 * dmu_sync_late_arrival_done() being called, we have to ensure 2018 * the lwb's "max txg" takes this tx's txg into account. 2019 */ 2020 zil_lwb_add_txg(zgd->zgd_lwb, dmu_tx_get_txg(tx)); 2021 2022 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP); 2023 dsa->dsa_dr = NULL; 2024 dsa->dsa_done = done; 2025 dsa->dsa_zgd = zgd; 2026 dsa->dsa_tx = tx; 2027 2028 /* 2029 * Since we are currently syncing this txg, it's nontrivial to 2030 * determine what BP to nopwrite against, so we disable nopwrite. 2031 * 2032 * When syncing, the db_blkptr is initially the BP of the previous 2033 * txg. We can not nopwrite against it because it will be changed 2034 * (this is similar to the non-late-arrival case where the dbuf is 2035 * dirty in a future txg). 2036 * 2037 * Then dbuf_write_ready() sets bp_blkptr to the location we will write. 2038 * We can not nopwrite against it because although the BP will not 2039 * (typically) be changed, the data has not yet been persisted to this 2040 * location. 2041 * 2042 * Finally, when dbuf_write_done() is called, it is theoretically 2043 * possible to always nopwrite, because the data that was written in 2044 * this txg is the same data that we are trying to write. However we 2045 * would need to check that this dbuf is not dirty in any future 2046 * txg's (as we do in the normal dmu_sync() path). For simplicity, we 2047 * don't nopwrite in this case. 2048 */ 2049 zp->zp_nopwrite = B_FALSE; 2050 2051 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp, 2052 abd_get_from_buf(zgd->zgd_db->db_data, zgd->zgd_db->db_size), 2053 zgd->zgd_db->db_size, zgd->zgd_db->db_size, zp, 2054 dmu_sync_late_arrival_ready, NULL, dmu_sync_late_arrival_done, 2055 dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb)); 2056 2057 return (0); 2058 } 2059 2060 /* 2061 * Intent log support: sync the block associated with db to disk. 2062 * N.B. and XXX: the caller is responsible for making sure that the 2063 * data isn't changing while dmu_sync() is writing it. 2064 * 2065 * Return values: 2066 * 2067 * EEXIST: this txg has already been synced, so there's nothing to do. 2068 * The caller should not log the write. 2069 * 2070 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do. 2071 * The caller should not log the write. 2072 * 2073 * EALREADY: this block is already in the process of being synced. 2074 * The caller should track its progress (somehow). 2075 * 2076 * EIO: could not do the I/O. 2077 * The caller should do a txg_wait_synced(). 2078 * 2079 * 0: the I/O has been initiated. 2080 * The caller should log this blkptr in the done callback. 2081 * It is possible that the I/O will fail, in which case 2082 * the error will be reported to the done callback and 2083 * propagated to pio from zio_done(). 2084 */ 2085 int 2086 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd) 2087 { 2088 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db; 2089 objset_t *os = db->db_objset; 2090 dsl_dataset_t *ds = os->os_dsl_dataset; 2091 dbuf_dirty_record_t *dr, *dr_next; 2092 dmu_sync_arg_t *dsa; 2093 zbookmark_phys_t zb; 2094 zio_prop_t zp; 2095 2096 ASSERT(pio != NULL); 2097 ASSERT(txg != 0); 2098 2099 SET_BOOKMARK(&zb, ds->ds_object, 2100 db->db.db_object, db->db_level, db->db_blkid); 2101 2102 DB_DNODE_ENTER(db); 2103 dmu_write_policy(os, DB_DNODE(db), db->db_level, WP_DMU_SYNC, &zp); 2104 DB_DNODE_EXIT(db); 2105 2106 /* 2107 * If we're frozen (running ziltest), we always need to generate a bp. 2108 */ 2109 if (txg > spa_freeze_txg(os->os_spa)) 2110 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); 2111 2112 /* 2113 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf() 2114 * and us. If we determine that this txg is not yet syncing, 2115 * but it begins to sync a moment later, that's OK because the 2116 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx. 2117 */ 2118 mutex_enter(&db->db_mtx); 2119 2120 if (txg <= spa_last_synced_txg(os->os_spa)) { 2121 /* 2122 * This txg has already synced. There's nothing to do. 2123 */ 2124 mutex_exit(&db->db_mtx); 2125 return (SET_ERROR(EEXIST)); 2126 } 2127 2128 if (txg <= spa_syncing_txg(os->os_spa)) { 2129 /* 2130 * This txg is currently syncing, so we can't mess with 2131 * the dirty record anymore; just write a new log block. 2132 */ 2133 mutex_exit(&db->db_mtx); 2134 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); 2135 } 2136 2137 dr = dbuf_find_dirty_eq(db, txg); 2138 2139 if (dr == NULL) { 2140 /* 2141 * There's no dr for this dbuf, so it must have been freed. 2142 * There's no need to log writes to freed blocks, so we're done. 2143 */ 2144 mutex_exit(&db->db_mtx); 2145 return (SET_ERROR(ENOENT)); 2146 } 2147 2148 dr_next = list_next(&db->db_dirty_records, dr); 2149 ASSERT(dr_next == NULL || dr_next->dr_txg < txg); 2150 2151 if (db->db_blkptr != NULL) { 2152 /* 2153 * We need to fill in zgd_bp with the current blkptr so that 2154 * the nopwrite code can check if we're writing the same 2155 * data that's already on disk. We can only nopwrite if we 2156 * are sure that after making the copy, db_blkptr will not 2157 * change until our i/o completes. We ensure this by 2158 * holding the db_mtx, and only allowing nopwrite if the 2159 * block is not already dirty (see below). This is verified 2160 * by dmu_sync_done(), which VERIFYs that the db_blkptr has 2161 * not changed. 2162 */ 2163 *zgd->zgd_bp = *db->db_blkptr; 2164 } 2165 2166 /* 2167 * Assume the on-disk data is X, the current syncing data (in 2168 * txg - 1) is Y, and the current in-memory data is Z (currently 2169 * in dmu_sync). 2170 * 2171 * We usually want to perform a nopwrite if X and Z are the 2172 * same. However, if Y is different (i.e. the BP is going to 2173 * change before this write takes effect), then a nopwrite will 2174 * be incorrect - we would override with X, which could have 2175 * been freed when Y was written. 2176 * 2177 * (Note that this is not a concern when we are nop-writing from 2178 * syncing context, because X and Y must be identical, because 2179 * all previous txgs have been synced.) 2180 * 2181 * Therefore, we disable nopwrite if the current BP could change 2182 * before this TXG. There are two ways it could change: by 2183 * being dirty (dr_next is non-NULL), or by being freed 2184 * (dnode_block_freed()). This behavior is verified by 2185 * zio_done(), which VERIFYs that the override BP is identical 2186 * to the on-disk BP. 2187 */ 2188 if (dr_next != NULL) { 2189 zp.zp_nopwrite = B_FALSE; 2190 } else { 2191 DB_DNODE_ENTER(db); 2192 if (dnode_block_freed(DB_DNODE(db), db->db_blkid)) 2193 zp.zp_nopwrite = B_FALSE; 2194 DB_DNODE_EXIT(db); 2195 } 2196 2197 ASSERT(dr->dr_txg == txg); 2198 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC || 2199 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { 2200 /* 2201 * We have already issued a sync write for this buffer, 2202 * or this buffer has already been synced. It could not 2203 * have been dirtied since, or we would have cleared the state. 2204 */ 2205 mutex_exit(&db->db_mtx); 2206 return (SET_ERROR(EALREADY)); 2207 } 2208 2209 ASSERT0(dr->dt.dl.dr_has_raw_params); 2210 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); 2211 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC; 2212 mutex_exit(&db->db_mtx); 2213 2214 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP); 2215 dsa->dsa_dr = dr; 2216 dsa->dsa_done = done; 2217 dsa->dsa_zgd = zgd; 2218 dsa->dsa_tx = NULL; 2219 2220 zio_nowait(arc_write(pio, os->os_spa, txg, zgd->zgd_bp, 2221 dr->dt.dl.dr_data, !DBUF_IS_CACHEABLE(db), 2222 dbuf_is_l2cacheable(db, NULL), &zp, dmu_sync_ready, NULL, 2223 dmu_sync_done, dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, 2224 &zb)); 2225 2226 return (0); 2227 } 2228 2229 int 2230 dmu_object_set_nlevels(objset_t *os, uint64_t object, int nlevels, dmu_tx_t *tx) 2231 { 2232 dnode_t *dn; 2233 int err; 2234 2235 err = dnode_hold(os, object, FTAG, &dn); 2236 if (err) 2237 return (err); 2238 err = dnode_set_nlevels(dn, nlevels, tx); 2239 dnode_rele(dn, FTAG); 2240 return (err); 2241 } 2242 2243 int 2244 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs, 2245 dmu_tx_t *tx) 2246 { 2247 dnode_t *dn; 2248 int err; 2249 2250 err = dnode_hold(os, object, FTAG, &dn); 2251 if (err) 2252 return (err); 2253 err = dnode_set_blksz(dn, size, ibs, tx); 2254 dnode_rele(dn, FTAG); 2255 return (err); 2256 } 2257 2258 int 2259 dmu_object_set_maxblkid(objset_t *os, uint64_t object, uint64_t maxblkid, 2260 dmu_tx_t *tx) 2261 { 2262 dnode_t *dn; 2263 int err; 2264 2265 err = dnode_hold(os, object, FTAG, &dn); 2266 if (err) 2267 return (err); 2268 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 2269 dnode_new_blkid(dn, maxblkid, tx, B_FALSE, B_TRUE); 2270 rw_exit(&dn->dn_struct_rwlock); 2271 dnode_rele(dn, FTAG); 2272 return (0); 2273 } 2274 2275 void 2276 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum, 2277 dmu_tx_t *tx) 2278 { 2279 dnode_t *dn; 2280 2281 /* 2282 * Send streams include each object's checksum function. This 2283 * check ensures that the receiving system can understand the 2284 * checksum function transmitted. 2285 */ 2286 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS); 2287 2288 VERIFY0(dnode_hold(os, object, FTAG, &dn)); 2289 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS); 2290 dn->dn_checksum = checksum; 2291 dnode_setdirty(dn, tx); 2292 dnode_rele(dn, FTAG); 2293 } 2294 2295 void 2296 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress, 2297 dmu_tx_t *tx) 2298 { 2299 dnode_t *dn; 2300 2301 /* 2302 * Send streams include each object's compression function. This 2303 * check ensures that the receiving system can understand the 2304 * compression function transmitted. 2305 */ 2306 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS); 2307 2308 VERIFY0(dnode_hold(os, object, FTAG, &dn)); 2309 dn->dn_compress = compress; 2310 dnode_setdirty(dn, tx); 2311 dnode_rele(dn, FTAG); 2312 } 2313 2314 /* 2315 * When the "redundant_metadata" property is set to "most", only indirect 2316 * blocks of this level and higher will have an additional ditto block. 2317 */ 2318 static const int zfs_redundant_metadata_most_ditto_level = 2; 2319 2320 void 2321 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp) 2322 { 2323 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET; 2324 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) || 2325 (wp & WP_SPILL)); 2326 enum zio_checksum checksum = os->os_checksum; 2327 enum zio_compress compress = os->os_compress; 2328 uint8_t complevel = os->os_complevel; 2329 enum zio_checksum dedup_checksum = os->os_dedup_checksum; 2330 boolean_t dedup = B_FALSE; 2331 boolean_t nopwrite = B_FALSE; 2332 boolean_t dedup_verify = os->os_dedup_verify; 2333 boolean_t encrypt = B_FALSE; 2334 int copies = os->os_copies; 2335 int gang_copies = os->os_copies; 2336 2337 /* 2338 * We maintain different write policies for each of the following 2339 * types of data: 2340 * 1. metadata 2341 * 2. preallocated blocks (i.e. level-0 blocks of a dump device) 2342 * 3. all other level 0 blocks 2343 */ 2344 if (ismd) { 2345 /* 2346 * XXX -- we should design a compression algorithm 2347 * that specializes in arrays of bps. 2348 */ 2349 compress = zio_compress_select(os->os_spa, 2350 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON); 2351 2352 /* 2353 * Metadata always gets checksummed. If the data 2354 * checksum is multi-bit correctable, and it's not a 2355 * ZBT-style checksum, then it's suitable for metadata 2356 * as well. Otherwise, the metadata checksum defaults 2357 * to fletcher4. 2358 */ 2359 if (!(zio_checksum_table[checksum].ci_flags & 2360 ZCHECKSUM_FLAG_METADATA) || 2361 (zio_checksum_table[checksum].ci_flags & 2362 ZCHECKSUM_FLAG_EMBEDDED)) 2363 checksum = ZIO_CHECKSUM_FLETCHER_4; 2364 2365 switch (os->os_redundant_metadata) { 2366 case ZFS_REDUNDANT_METADATA_ALL: 2367 copies++; 2368 gang_copies++; 2369 break; 2370 case ZFS_REDUNDANT_METADATA_MOST: 2371 if (level >= zfs_redundant_metadata_most_ditto_level || 2372 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL)) 2373 copies++; 2374 if (level + 1 >= 2375 zfs_redundant_metadata_most_ditto_level || 2376 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL)) 2377 gang_copies++; 2378 break; 2379 case ZFS_REDUNDANT_METADATA_SOME: 2380 if (DMU_OT_IS_CRITICAL(type, level)) { 2381 copies++; 2382 gang_copies++; 2383 } else if (DMU_OT_IS_METADATA(type)) { 2384 gang_copies++; 2385 } 2386 break; 2387 case ZFS_REDUNDANT_METADATA_NONE: 2388 break; 2389 } 2390 2391 if (dmu_ddt_copies > 0) { 2392 /* 2393 * If this tunable is set, and this is a write for a 2394 * dedup entry store (zap or log), then we treat it 2395 * something like ZFS_REDUNDANT_METADATA_MOST on a 2396 * regular dataset: this many copies, and one more for 2397 * "higher" indirect blocks. This specific exception is 2398 * necessary because dedup objects are stored in the 2399 * MOS, which always has the highest possible copies. 2400 */ 2401 dmu_object_type_t stype = 2402 dn ? dn->dn_storage_type : DMU_OT_NONE; 2403 if (stype == DMU_OT_NONE) 2404 stype = type; 2405 if (stype == DMU_OT_DDT_ZAP) { 2406 copies = dmu_ddt_copies; 2407 if (level >= 2408 zfs_redundant_metadata_most_ditto_level) 2409 copies++; 2410 } 2411 } 2412 } else if (wp & WP_NOFILL) { 2413 ASSERT0(level); 2414 2415 /* 2416 * If we're writing preallocated blocks, we aren't actually 2417 * writing them so don't set any policy properties. These 2418 * blocks are currently only used by an external subsystem 2419 * outside of zfs (i.e. dump) and not written by the zio 2420 * pipeline. 2421 */ 2422 compress = ZIO_COMPRESS_OFF; 2423 checksum = ZIO_CHECKSUM_OFF; 2424 } else { 2425 compress = zio_compress_select(os->os_spa, dn->dn_compress, 2426 compress); 2427 complevel = zio_complevel_select(os->os_spa, compress, 2428 complevel, complevel); 2429 2430 /* 2431 * Storing many references to an all zeros block in the dedup 2432 * table would be expensive. Instead, if dedup is enabled, 2433 * store them as holes even if compression is not enabled. 2434 */ 2435 if (compress == ZIO_COMPRESS_OFF && 2436 dedup_checksum != ZIO_CHECKSUM_OFF) 2437 compress = ZIO_COMPRESS_EMPTY; 2438 2439 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ? 2440 zio_checksum_select(dn->dn_checksum, checksum) : 2441 dedup_checksum; 2442 2443 /* 2444 * Determine dedup setting. If we are in dmu_sync(), 2445 * we won't actually dedup now because that's all 2446 * done in syncing context; but we do want to use the 2447 * dedup checksum. If the checksum is not strong 2448 * enough to ensure unique signatures, force 2449 * dedup_verify. 2450 */ 2451 if (dedup_checksum != ZIO_CHECKSUM_OFF) { 2452 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE; 2453 if (!(zio_checksum_table[checksum].ci_flags & 2454 ZCHECKSUM_FLAG_DEDUP)) 2455 dedup_verify = B_TRUE; 2456 } 2457 2458 /* 2459 * Enable nopwrite if we have secure enough checksum 2460 * algorithm (see comment in zio_nop_write) and 2461 * compression is enabled. We don't enable nopwrite if 2462 * dedup is enabled as the two features are mutually 2463 * exclusive. 2464 */ 2465 nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags & 2466 ZCHECKSUM_FLAG_NOPWRITE) && 2467 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled); 2468 2469 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL || 2470 (os->os_redundant_metadata == 2471 ZFS_REDUNDANT_METADATA_MOST && 2472 zfs_redundant_metadata_most_ditto_level <= 1)) 2473 gang_copies++; 2474 } 2475 2476 /* 2477 * All objects in an encrypted objset are protected from modification 2478 * via a MAC. Encrypted objects store their IV and salt in the last DVA 2479 * in the bp, so we cannot use all copies. Encrypted objects are also 2480 * not subject to nopwrite since writing the same data will still 2481 * result in a new ciphertext. Only encrypted blocks can be dedup'd 2482 * to avoid ambiguity in the dedup code since the DDT does not store 2483 * object types. 2484 */ 2485 if (os->os_encrypted && (wp & WP_NOFILL) == 0) { 2486 encrypt = B_TRUE; 2487 2488 if (DMU_OT_IS_ENCRYPTED(type)) { 2489 copies = MIN(copies, SPA_DVAS_PER_BP - 1); 2490 gang_copies = MIN(gang_copies, SPA_DVAS_PER_BP - 1); 2491 nopwrite = B_FALSE; 2492 } else { 2493 dedup = B_FALSE; 2494 } 2495 2496 if (level <= 0 && 2497 (type == DMU_OT_DNODE || type == DMU_OT_OBJSET)) { 2498 compress = ZIO_COMPRESS_EMPTY; 2499 } 2500 } 2501 2502 zp->zp_compress = compress; 2503 zp->zp_complevel = complevel; 2504 zp->zp_checksum = checksum; 2505 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type; 2506 zp->zp_level = level; 2507 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa)); 2508 zp->zp_gang_copies = MIN(MAX(gang_copies, copies), 2509 spa_max_replication(os->os_spa)); 2510 zp->zp_dedup = dedup; 2511 zp->zp_dedup_verify = dedup && dedup_verify; 2512 zp->zp_nopwrite = nopwrite; 2513 zp->zp_encrypt = encrypt; 2514 zp->zp_byteorder = ZFS_HOST_BYTEORDER; 2515 zp->zp_direct_write = (wp & WP_DIRECT_WR) ? B_TRUE : B_FALSE; 2516 zp->zp_rewrite = B_FALSE; 2517 memset(zp->zp_salt, 0, ZIO_DATA_SALT_LEN); 2518 memset(zp->zp_iv, 0, ZIO_DATA_IV_LEN); 2519 memset(zp->zp_mac, 0, ZIO_DATA_MAC_LEN); 2520 zp->zp_zpl_smallblk = (DMU_OT_IS_FILE(zp->zp_type) || 2521 zp->zp_type == DMU_OT_ZVOL) ? 2522 os->os_zpl_special_smallblock : 0; 2523 zp->zp_storage_type = dn ? dn->dn_storage_type : DMU_OT_NONE; 2524 2525 ASSERT3U(zp->zp_compress, !=, ZIO_COMPRESS_INHERIT); 2526 } 2527 2528 /* 2529 * Reports the location of data and holes in an object. In order to 2530 * accurately report holes all dirty data must be synced to disk. This 2531 * causes extremely poor performance when seeking for holes in a dirty file. 2532 * As a compromise, only provide hole data when the dnode is clean. When 2533 * a dnode is dirty report the dnode as having no holes by returning EBUSY 2534 * which is always safe to do. 2535 */ 2536 int 2537 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off) 2538 { 2539 dnode_t *dn; 2540 uint64_t txg, maxtxg = 0; 2541 int err; 2542 2543 restart: 2544 err = dnode_hold(os, object, FTAG, &dn); 2545 if (err) 2546 return (err); 2547 2548 rw_enter(&dn->dn_struct_rwlock, RW_READER); 2549 2550 if (dnode_is_dirty(dn)) { 2551 /* 2552 * If the zfs_dmu_offset_next_sync module option is enabled 2553 * then hole reporting has been requested. Dirty dnodes 2554 * must be synced to disk to accurately report holes. 2555 * 2556 * Provided a RL_READER rangelock spanning 0-UINT64_MAX is 2557 * held by the caller only limited restarts will be required. 2558 * We tolerate callers which do not hold the rangelock by 2559 * returning EBUSY and not reporting holes after at most 2560 * TXG_CONCURRENT_STATES (3) restarts. 2561 */ 2562 if (zfs_dmu_offset_next_sync) { 2563 rw_exit(&dn->dn_struct_rwlock); 2564 dnode_rele(dn, FTAG); 2565 2566 if (maxtxg == 0) { 2567 txg = spa_last_synced_txg(dmu_objset_spa(os)); 2568 maxtxg = txg + TXG_CONCURRENT_STATES; 2569 } else if (txg >= maxtxg) 2570 return (SET_ERROR(EBUSY)); 2571 2572 txg_wait_synced(dmu_objset_pool(os), ++txg); 2573 goto restart; 2574 } 2575 2576 err = SET_ERROR(EBUSY); 2577 } else { 2578 err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK | 2579 (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0); 2580 } 2581 2582 rw_exit(&dn->dn_struct_rwlock); 2583 dnode_rele(dn, FTAG); 2584 2585 return (err); 2586 } 2587 2588 int 2589 dmu_read_l0_bps(objset_t *os, uint64_t object, uint64_t offset, uint64_t length, 2590 blkptr_t *bps, size_t *nbpsp) 2591 { 2592 dmu_buf_t **dbp, *dbuf; 2593 dmu_buf_impl_t *db; 2594 blkptr_t *bp; 2595 int error, numbufs; 2596 2597 error = dmu_buf_hold_array(os, object, offset, length, FALSE, FTAG, 2598 &numbufs, &dbp, DMU_READ_PREFETCH); 2599 if (error != 0) { 2600 if (error == ESRCH) { 2601 error = SET_ERROR(ENXIO); 2602 } 2603 return (error); 2604 } 2605 2606 ASSERT3U(numbufs, <=, *nbpsp); 2607 2608 for (int i = 0; i < numbufs; i++) { 2609 dbuf = dbp[i]; 2610 db = (dmu_buf_impl_t *)dbuf; 2611 2612 mutex_enter(&db->db_mtx); 2613 2614 if (!list_is_empty(&db->db_dirty_records)) { 2615 dbuf_dirty_record_t *dr; 2616 2617 dr = list_head(&db->db_dirty_records); 2618 if (dr->dt.dl.dr_brtwrite) { 2619 /* 2620 * This is very special case where we clone a 2621 * block and in the same transaction group we 2622 * read its BP (most likely to clone the clone). 2623 */ 2624 bp = &dr->dt.dl.dr_overridden_by; 2625 } else { 2626 /* 2627 * The block was modified in the same 2628 * transaction group. 2629 */ 2630 mutex_exit(&db->db_mtx); 2631 error = SET_ERROR(EAGAIN); 2632 goto out; 2633 } 2634 } else { 2635 bp = db->db_blkptr; 2636 } 2637 2638 mutex_exit(&db->db_mtx); 2639 2640 if (bp == NULL) { 2641 /* 2642 * The file size was increased, but the block was never 2643 * written, otherwise we would either have the block 2644 * pointer or the dirty record and would not get here. 2645 * It is effectively a hole, so report it as such. 2646 */ 2647 BP_ZERO(&bps[i]); 2648 continue; 2649 } 2650 /* 2651 * Make sure we clone only data blocks. 2652 */ 2653 if (BP_IS_METADATA(bp) && !BP_IS_HOLE(bp)) { 2654 error = SET_ERROR(EINVAL); 2655 goto out; 2656 } 2657 2658 /* 2659 * If the block was allocated in transaction group that is not 2660 * yet synced, we could clone it, but we couldn't write this 2661 * operation into ZIL, or it may be impossible to replay, since 2662 * the block may appear not yet allocated at that point. 2663 */ 2664 if (BP_GET_PHYSICAL_BIRTH(bp) > spa_freeze_txg(os->os_spa)) { 2665 error = SET_ERROR(EINVAL); 2666 goto out; 2667 } 2668 if (BP_GET_PHYSICAL_BIRTH(bp) > 2669 spa_last_synced_txg(os->os_spa)) { 2670 error = SET_ERROR(EAGAIN); 2671 goto out; 2672 } 2673 2674 bps[i] = *bp; 2675 } 2676 2677 *nbpsp = numbufs; 2678 out: 2679 dmu_buf_rele_array(dbp, numbufs, FTAG); 2680 2681 return (error); 2682 } 2683 2684 int 2685 dmu_brt_clone(objset_t *os, uint64_t object, uint64_t offset, uint64_t length, 2686 dmu_tx_t *tx, const blkptr_t *bps, size_t nbps) 2687 { 2688 spa_t *spa; 2689 dmu_buf_t **dbp, *dbuf; 2690 dmu_buf_impl_t *db; 2691 struct dirty_leaf *dl; 2692 dbuf_dirty_record_t *dr; 2693 const blkptr_t *bp; 2694 int error = 0, i, numbufs; 2695 2696 spa = os->os_spa; 2697 2698 VERIFY0(dmu_buf_hold_array(os, object, offset, length, FALSE, FTAG, 2699 &numbufs, &dbp, DMU_READ_PREFETCH)); 2700 ASSERT3U(nbps, ==, numbufs); 2701 2702 /* 2703 * Before we start cloning make sure that the dbufs sizes match new BPs 2704 * sizes. If they don't, that's a no-go, as we are not able to shrink 2705 * dbufs. 2706 */ 2707 for (i = 0; i < numbufs; i++) { 2708 dbuf = dbp[i]; 2709 db = (dmu_buf_impl_t *)dbuf; 2710 bp = &bps[i]; 2711 2712 ASSERT3U(db->db.db_object, !=, DMU_META_DNODE_OBJECT); 2713 ASSERT0(db->db_level); 2714 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 2715 ASSERT(db->db_blkid != DMU_SPILL_BLKID); 2716 2717 if (!BP_IS_HOLE(bp) && BP_GET_LSIZE(bp) != dbuf->db_size) { 2718 error = SET_ERROR(EXDEV); 2719 goto out; 2720 } 2721 } 2722 2723 for (i = 0; i < numbufs; i++) { 2724 dbuf = dbp[i]; 2725 db = (dmu_buf_impl_t *)dbuf; 2726 bp = &bps[i]; 2727 2728 dmu_buf_will_clone_or_dio(dbuf, tx); 2729 2730 mutex_enter(&db->db_mtx); 2731 2732 dr = list_head(&db->db_dirty_records); 2733 VERIFY(dr != NULL); 2734 ASSERT3U(dr->dr_txg, ==, tx->tx_txg); 2735 dl = &dr->dt.dl; 2736 ASSERT0(dl->dr_has_raw_params); 2737 dl->dr_overridden_by = *bp; 2738 if (!BP_IS_HOLE(bp) || BP_GET_LOGICAL_BIRTH(bp) != 0) { 2739 if (!BP_IS_EMBEDDED(bp)) { 2740 BP_SET_BIRTH(&dl->dr_overridden_by, dr->dr_txg, 2741 BP_GET_PHYSICAL_BIRTH(bp)); 2742 BP_SET_REWRITE(&dl->dr_overridden_by, 0); 2743 } else { 2744 BP_SET_LOGICAL_BIRTH(&dl->dr_overridden_by, 2745 dr->dr_txg); 2746 } 2747 } 2748 dl->dr_brtwrite = B_TRUE; 2749 dl->dr_override_state = DR_OVERRIDDEN; 2750 2751 mutex_exit(&db->db_mtx); 2752 2753 /* 2754 * When data in embedded into BP there is no need to create 2755 * BRT entry as there is no data block. Just copy the BP as 2756 * it contains the data. 2757 */ 2758 if (!BP_IS_HOLE(bp) && !BP_IS_EMBEDDED(bp)) { 2759 brt_pending_add(spa, bp, tx); 2760 } 2761 } 2762 out: 2763 dmu_buf_rele_array(dbp, numbufs, FTAG); 2764 2765 return (error); 2766 } 2767 2768 void 2769 __dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi) 2770 { 2771 dnode_phys_t *dnp = dn->dn_phys; 2772 2773 doi->doi_data_block_size = dn->dn_datablksz; 2774 doi->doi_metadata_block_size = dn->dn_indblkshift ? 2775 1ULL << dn->dn_indblkshift : 0; 2776 doi->doi_type = dn->dn_type; 2777 doi->doi_bonus_type = dn->dn_bonustype; 2778 doi->doi_bonus_size = dn->dn_bonuslen; 2779 doi->doi_dnodesize = dn->dn_num_slots << DNODE_SHIFT; 2780 doi->doi_indirection = dn->dn_nlevels; 2781 doi->doi_checksum = dn->dn_checksum; 2782 doi->doi_compress = dn->dn_compress; 2783 doi->doi_nblkptr = dn->dn_nblkptr; 2784 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9; 2785 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz; 2786 doi->doi_fill_count = 0; 2787 for (int i = 0; i < dnp->dn_nblkptr; i++) 2788 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]); 2789 } 2790 2791 void 2792 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi) 2793 { 2794 rw_enter(&dn->dn_struct_rwlock, RW_READER); 2795 mutex_enter(&dn->dn_mtx); 2796 2797 __dmu_object_info_from_dnode(dn, doi); 2798 2799 mutex_exit(&dn->dn_mtx); 2800 rw_exit(&dn->dn_struct_rwlock); 2801 } 2802 2803 /* 2804 * Get information on a DMU object. 2805 * If doi is NULL, just indicates whether the object exists. 2806 */ 2807 int 2808 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi) 2809 { 2810 dnode_t *dn; 2811 int err = dnode_hold(os, object, FTAG, &dn); 2812 2813 if (err) 2814 return (err); 2815 2816 if (doi != NULL) 2817 dmu_object_info_from_dnode(dn, doi); 2818 2819 dnode_rele(dn, FTAG); 2820 return (0); 2821 } 2822 2823 /* 2824 * As above, but faster; can be used when you have a held dbuf in hand. 2825 */ 2826 void 2827 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi) 2828 { 2829 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2830 2831 DB_DNODE_ENTER(db); 2832 dmu_object_info_from_dnode(DB_DNODE(db), doi); 2833 DB_DNODE_EXIT(db); 2834 } 2835 2836 /* 2837 * Faster still when you only care about the size. 2838 */ 2839 void 2840 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize, 2841 u_longlong_t *nblk512) 2842 { 2843 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2844 dnode_t *dn; 2845 2846 DB_DNODE_ENTER(db); 2847 dn = DB_DNODE(db); 2848 2849 *blksize = dn->dn_datablksz; 2850 /* add in number of slots used for the dnode itself */ 2851 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >> 2852 SPA_MINBLOCKSHIFT) + dn->dn_num_slots; 2853 DB_DNODE_EXIT(db); 2854 } 2855 2856 void 2857 dmu_object_dnsize_from_db(dmu_buf_t *db_fake, int *dnsize) 2858 { 2859 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2860 2861 DB_DNODE_ENTER(db); 2862 *dnsize = DB_DNODE(db)->dn_num_slots << DNODE_SHIFT; 2863 DB_DNODE_EXIT(db); 2864 } 2865 2866 void 2867 byteswap_uint64_array(void *vbuf, size_t size) 2868 { 2869 uint64_t *buf = vbuf; 2870 size_t count = size >> 3; 2871 int i; 2872 2873 ASSERT0((size & 7)); 2874 2875 for (i = 0; i < count; i++) 2876 buf[i] = BSWAP_64(buf[i]); 2877 } 2878 2879 void 2880 byteswap_uint32_array(void *vbuf, size_t size) 2881 { 2882 uint32_t *buf = vbuf; 2883 size_t count = size >> 2; 2884 int i; 2885 2886 ASSERT0((size & 3)); 2887 2888 for (i = 0; i < count; i++) 2889 buf[i] = BSWAP_32(buf[i]); 2890 } 2891 2892 void 2893 byteswap_uint16_array(void *vbuf, size_t size) 2894 { 2895 uint16_t *buf = vbuf; 2896 size_t count = size >> 1; 2897 int i; 2898 2899 ASSERT0((size & 1)); 2900 2901 for (i = 0; i < count; i++) 2902 buf[i] = BSWAP_16(buf[i]); 2903 } 2904 2905 void 2906 byteswap_uint8_array(void *vbuf, size_t size) 2907 { 2908 (void) vbuf, (void) size; 2909 } 2910 2911 void 2912 dmu_init(void) 2913 { 2914 abd_init(); 2915 zfs_dbgmsg_init(); 2916 sa_cache_init(); 2917 dmu_objset_init(); 2918 dnode_init(); 2919 zfetch_init(); 2920 dmu_tx_init(); 2921 l2arc_init(); 2922 arc_init(); 2923 dbuf_init(); 2924 } 2925 2926 void 2927 dmu_fini(void) 2928 { 2929 arc_fini(); /* arc depends on l2arc, so arc must go first */ 2930 l2arc_fini(); 2931 dmu_tx_fini(); 2932 zfetch_fini(); 2933 dbuf_fini(); 2934 dnode_fini(); 2935 dmu_objset_fini(); 2936 sa_cache_fini(); 2937 zfs_dbgmsg_fini(); 2938 abd_fini(); 2939 } 2940 2941 EXPORT_SYMBOL(dmu_bonus_hold); 2942 EXPORT_SYMBOL(dmu_bonus_hold_by_dnode); 2943 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus); 2944 EXPORT_SYMBOL(dmu_buf_rele_array); 2945 EXPORT_SYMBOL(dmu_prefetch); 2946 EXPORT_SYMBOL(dmu_prefetch_by_dnode); 2947 EXPORT_SYMBOL(dmu_prefetch_dnode); 2948 EXPORT_SYMBOL(dmu_free_range); 2949 EXPORT_SYMBOL(dmu_free_long_range); 2950 EXPORT_SYMBOL(dmu_free_long_object); 2951 EXPORT_SYMBOL(dmu_read); 2952 EXPORT_SYMBOL(dmu_read_by_dnode); 2953 EXPORT_SYMBOL(dmu_read_uio); 2954 EXPORT_SYMBOL(dmu_read_uio_dbuf); 2955 EXPORT_SYMBOL(dmu_read_uio_dnode); 2956 EXPORT_SYMBOL(dmu_write); 2957 EXPORT_SYMBOL(dmu_write_by_dnode); 2958 EXPORT_SYMBOL(dmu_write_uio); 2959 EXPORT_SYMBOL(dmu_write_uio_dbuf); 2960 EXPORT_SYMBOL(dmu_write_uio_dnode); 2961 EXPORT_SYMBOL(dmu_prealloc); 2962 EXPORT_SYMBOL(dmu_object_info); 2963 EXPORT_SYMBOL(dmu_object_info_from_dnode); 2964 EXPORT_SYMBOL(dmu_object_info_from_db); 2965 EXPORT_SYMBOL(dmu_object_size_from_db); 2966 EXPORT_SYMBOL(dmu_object_dnsize_from_db); 2967 EXPORT_SYMBOL(dmu_object_set_nlevels); 2968 EXPORT_SYMBOL(dmu_object_set_blocksize); 2969 EXPORT_SYMBOL(dmu_object_set_maxblkid); 2970 EXPORT_SYMBOL(dmu_object_set_checksum); 2971 EXPORT_SYMBOL(dmu_object_set_compress); 2972 EXPORT_SYMBOL(dmu_offset_next); 2973 EXPORT_SYMBOL(dmu_write_policy); 2974 EXPORT_SYMBOL(dmu_sync); 2975 EXPORT_SYMBOL(dmu_request_arcbuf); 2976 EXPORT_SYMBOL(dmu_return_arcbuf); 2977 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dnode); 2978 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dbuf); 2979 EXPORT_SYMBOL(dmu_buf_hold); 2980 EXPORT_SYMBOL(dmu_ot); 2981 2982 ZFS_MODULE_PARAM(zfs, zfs_, nopwrite_enabled, INT, ZMOD_RW, 2983 "Enable NOP writes"); 2984 2985 ZFS_MODULE_PARAM(zfs, zfs_, per_txg_dirty_frees_percent, UINT, ZMOD_RW, 2986 "Percentage of dirtied blocks from frees in one TXG"); 2987 2988 ZFS_MODULE_PARAM(zfs, zfs_, dmu_offset_next_sync, INT, ZMOD_RW, 2989 "Enable forcing txg sync to find holes"); 2990 2991 ZFS_MODULE_PARAM(zfs, , dmu_prefetch_max, UINT, ZMOD_RW, 2992 "Limit one prefetch call to this size"); 2993 2994 ZFS_MODULE_PARAM(zfs, , dmu_ddt_copies, UINT, ZMOD_RW, 2995 "Override copies= for dedup objects"); 2996