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) 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, DMU_READ_PREFETCH); 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) 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, DMU_READ_PREFETCH); 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 interrupted 854 */ 855 uint64_t chunksize; 856 if (dn->dn_indblkshift) { 857 uint64_t nbps = bp_span_in_blocks(dn->dn_indblkshift, 1); 858 chunksize = (nbps * 16) << dn->dn_datablkshift; 859 } else { 860 chunksize = dn->dn_datablksz; 861 } 862 863 while (size > 0) { 864 uint64_t mylen = MIN(size, chunksize); 865 866 dmu_prefetch_wait_by_dnode(dn, offset, mylen); 867 868 offset += mylen; 869 size -= mylen; 870 871 if (issig()) { 872 err = SET_ERROR(EINTR); 873 break; 874 } 875 } 876 877 dnode_rele(dn, FTAG); 878 879 return (err); 880 } 881 882 /* 883 * Issue prefetch I/Os for the given object's dnode. 884 */ 885 void 886 dmu_prefetch_dnode(objset_t *os, uint64_t object, zio_priority_t pri) 887 { 888 if (object == 0 || object >= DN_MAX_OBJECT) 889 return; 890 891 dnode_t *dn = DMU_META_DNODE(os); 892 rw_enter(&dn->dn_struct_rwlock, RW_READER); 893 uint64_t blkid = dbuf_whichblock(dn, 0, object * sizeof (dnode_phys_t)); 894 dbuf_prefetch(dn, 0, blkid, pri, 0); 895 rw_exit(&dn->dn_struct_rwlock); 896 } 897 898 /* 899 * Get the next "chunk" of file data to free. We traverse the file from 900 * the end so that the file gets shorter over time (if we crash in the 901 * middle, this will leave us in a better state). We find allocated file 902 * data by simply searching the allocated level 1 indirects. 903 * 904 * On input, *start should be the first offset that does not need to be 905 * freed (e.g. "offset + length"). On return, *start will be the first 906 * offset that should be freed and l1blks is set to the number of level 1 907 * indirect blocks found within the chunk. 908 */ 909 static int 910 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum, uint64_t *l1blks) 911 { 912 uint64_t blks; 913 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1); 914 /* bytes of data covered by a level-1 indirect block */ 915 uint64_t iblkrange = (uint64_t)dn->dn_datablksz * 916 EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT); 917 918 ASSERT3U(minimum, <=, *start); 919 920 /* dn_nlevels == 1 means we don't have any L1 blocks */ 921 if (dn->dn_nlevels <= 1) { 922 *l1blks = 0; 923 *start = minimum; 924 return (0); 925 } 926 927 /* 928 * Check if we can free the entire range assuming that all of the 929 * L1 blocks in this range have data. If we can, we use this 930 * worst case value as an estimate so we can avoid having to look 931 * at the object's actual data. 932 */ 933 uint64_t total_l1blks = 934 (roundup(*start, iblkrange) - (minimum / iblkrange * iblkrange)) / 935 iblkrange; 936 if (total_l1blks <= maxblks) { 937 *l1blks = total_l1blks; 938 *start = minimum; 939 return (0); 940 } 941 ASSERT(ISP2(iblkrange)); 942 943 for (blks = 0; *start > minimum && blks < maxblks; blks++) { 944 int err; 945 946 /* 947 * dnode_next_offset(BACKWARDS) will find an allocated L1 948 * indirect block at or before the input offset. We must 949 * decrement *start so that it is at the end of the region 950 * to search. 951 */ 952 (*start)--; 953 954 err = dnode_next_offset(dn, 955 DNODE_FIND_BACKWARDS, start, 2, 1, 0); 956 957 /* if there are no indirect blocks before start, we are done */ 958 if (err == ESRCH) { 959 *start = minimum; 960 break; 961 } else if (err != 0) { 962 *l1blks = blks; 963 return (err); 964 } 965 966 /* set start to the beginning of this L1 indirect */ 967 *start = P2ALIGN_TYPED(*start, iblkrange, uint64_t); 968 } 969 if (*start < minimum) 970 *start = minimum; 971 *l1blks = blks; 972 973 return (0); 974 } 975 976 /* 977 * If this objset is of type OST_ZFS return true if vfs's unmounted flag is set, 978 * otherwise return false. 979 * Used below in dmu_free_long_range_impl() to enable abort when unmounting 980 */ 981 static boolean_t 982 dmu_objset_zfs_unmounting(objset_t *os) 983 { 984 #ifdef _KERNEL 985 if (dmu_objset_type(os) == DMU_OST_ZFS) 986 return (zfs_get_vfs_flag_unmounted(os)); 987 #else 988 (void) os; 989 #endif 990 return (B_FALSE); 991 } 992 993 static int 994 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset, 995 uint64_t length) 996 { 997 uint64_t object_size; 998 int err; 999 uint64_t dirty_frees_threshold; 1000 dsl_pool_t *dp = dmu_objset_pool(os); 1001 1002 if (dn == NULL) 1003 return (SET_ERROR(EINVAL)); 1004 1005 object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz; 1006 if (offset >= object_size) 1007 return (0); 1008 1009 if (zfs_per_txg_dirty_frees_percent <= 100) 1010 dirty_frees_threshold = 1011 zfs_per_txg_dirty_frees_percent * zfs_dirty_data_max / 100; 1012 else 1013 dirty_frees_threshold = zfs_dirty_data_max / 20; 1014 1015 if (length == DMU_OBJECT_END || offset + length > object_size) 1016 length = object_size - offset; 1017 1018 while (length != 0) { 1019 uint64_t chunk_end, chunk_begin, chunk_len; 1020 uint64_t l1blks; 1021 dmu_tx_t *tx; 1022 1023 if (dmu_objset_zfs_unmounting(dn->dn_objset)) 1024 return (SET_ERROR(EINTR)); 1025 1026 chunk_end = chunk_begin = offset + length; 1027 1028 /* move chunk_begin backwards to the beginning of this chunk */ 1029 err = get_next_chunk(dn, &chunk_begin, offset, &l1blks); 1030 if (err) 1031 return (err); 1032 ASSERT3U(chunk_begin, >=, offset); 1033 ASSERT3U(chunk_begin, <=, chunk_end); 1034 1035 chunk_len = chunk_end - chunk_begin; 1036 1037 tx = dmu_tx_create(os); 1038 dmu_tx_hold_free(tx, dn->dn_object, chunk_begin, chunk_len); 1039 1040 /* 1041 * Mark this transaction as typically resulting in a net 1042 * reduction in space used. 1043 */ 1044 dmu_tx_mark_netfree(tx); 1045 err = dmu_tx_assign(tx, DMU_TX_WAIT); 1046 if (err) { 1047 dmu_tx_abort(tx); 1048 return (err); 1049 } 1050 1051 uint64_t txg = dmu_tx_get_txg(tx); 1052 1053 mutex_enter(&dp->dp_lock); 1054 uint64_t long_free_dirty = 1055 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK]; 1056 mutex_exit(&dp->dp_lock); 1057 1058 /* 1059 * To avoid filling up a TXG with just frees, wait for 1060 * the next TXG to open before freeing more chunks if 1061 * we have reached the threshold of frees. 1062 */ 1063 if (dirty_frees_threshold != 0 && 1064 long_free_dirty >= dirty_frees_threshold) { 1065 DMU_TX_STAT_BUMP(dmu_tx_dirty_frees_delay); 1066 dmu_tx_commit(tx); 1067 txg_wait_open(dp, 0, B_TRUE); 1068 continue; 1069 } 1070 1071 /* 1072 * In order to prevent unnecessary write throttling, for each 1073 * TXG, we track the cumulative size of L1 blocks being dirtied 1074 * in dnode_free_range() below. We compare this number to a 1075 * tunable threshold, past which we prevent new L1 dirty freeing 1076 * blocks from being added into the open TXG. See 1077 * dmu_free_long_range_impl() for details. The threshold 1078 * prevents write throttle activation due to dirty freeing L1 1079 * blocks taking up a large percentage of zfs_dirty_data_max. 1080 */ 1081 mutex_enter(&dp->dp_lock); 1082 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] += 1083 l1blks << dn->dn_indblkshift; 1084 mutex_exit(&dp->dp_lock); 1085 DTRACE_PROBE3(free__long__range, 1086 uint64_t, long_free_dirty, uint64_t, chunk_len, 1087 uint64_t, txg); 1088 dnode_free_range(dn, chunk_begin, chunk_len, tx); 1089 1090 dmu_tx_commit(tx); 1091 1092 length -= chunk_len; 1093 } 1094 return (0); 1095 } 1096 1097 int 1098 dmu_free_long_range(objset_t *os, uint64_t object, 1099 uint64_t offset, uint64_t length) 1100 { 1101 dnode_t *dn; 1102 int err; 1103 1104 err = dnode_hold(os, object, FTAG, &dn); 1105 if (err != 0) 1106 return (err); 1107 err = dmu_free_long_range_impl(os, dn, offset, length); 1108 1109 /* 1110 * It is important to zero out the maxblkid when freeing the entire 1111 * file, so that (a) subsequent calls to dmu_free_long_range_impl() 1112 * will take the fast path, and (b) dnode_reallocate() can verify 1113 * that the entire file has been freed. 1114 */ 1115 if (err == 0 && offset == 0 && length == DMU_OBJECT_END) 1116 dn->dn_maxblkid = 0; 1117 1118 dnode_rele(dn, FTAG); 1119 return (err); 1120 } 1121 1122 int 1123 dmu_free_long_object(objset_t *os, uint64_t object) 1124 { 1125 dmu_tx_t *tx; 1126 int err; 1127 1128 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END); 1129 if (err != 0) 1130 return (err); 1131 1132 tx = dmu_tx_create(os); 1133 dmu_tx_hold_bonus(tx, object); 1134 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END); 1135 dmu_tx_mark_netfree(tx); 1136 err = dmu_tx_assign(tx, DMU_TX_WAIT); 1137 if (err == 0) { 1138 err = dmu_object_free(os, object, tx); 1139 dmu_tx_commit(tx); 1140 } else { 1141 dmu_tx_abort(tx); 1142 } 1143 1144 return (err); 1145 } 1146 1147 int 1148 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset, 1149 uint64_t size, dmu_tx_t *tx) 1150 { 1151 dnode_t *dn; 1152 int err = dnode_hold(os, object, FTAG, &dn); 1153 if (err) 1154 return (err); 1155 ASSERT(offset < UINT64_MAX); 1156 ASSERT(size == DMU_OBJECT_END || size <= UINT64_MAX - offset); 1157 dnode_free_range(dn, offset, size, tx); 1158 dnode_rele(dn, FTAG); 1159 return (0); 1160 } 1161 1162 static int 1163 dmu_read_impl(dnode_t *dn, uint64_t offset, uint64_t size, 1164 void *buf, dmu_flags_t flags) 1165 { 1166 dmu_buf_t **dbp; 1167 int numbufs, err = 0; 1168 1169 /* 1170 * Deal with odd block sizes, where there can't be data past the first 1171 * block. If we ever do the tail block optimization, we will need to 1172 * handle that here as well. 1173 */ 1174 if (dn->dn_maxblkid == 0) { 1175 uint64_t newsz = offset > dn->dn_datablksz ? 0 : 1176 MIN(size, dn->dn_datablksz - offset); 1177 memset((char *)buf + newsz, 0, size - newsz); 1178 size = newsz; 1179 } 1180 1181 if (size == 0) 1182 return (0); 1183 1184 /* Allow Direct I/O when requested and properly aligned */ 1185 if ((flags & DMU_DIRECTIO) && zfs_dio_page_aligned(buf) && 1186 zfs_dio_aligned(offset, size, PAGESIZE)) { 1187 abd_t *data = abd_get_from_buf(buf, size); 1188 err = dmu_read_abd(dn, offset, size, data, flags); 1189 abd_free(data); 1190 return (err); 1191 } 1192 flags &= ~DMU_DIRECTIO; 1193 1194 while (size > 0) { 1195 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2); 1196 int i; 1197 1198 /* 1199 * NB: we could do this block-at-a-time, but it's nice 1200 * to be reading in parallel. 1201 */ 1202 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen, 1203 TRUE, FTAG, &numbufs, &dbp, flags); 1204 if (err) 1205 break; 1206 1207 for (i = 0; i < numbufs; i++) { 1208 uint64_t tocpy; 1209 int64_t bufoff; 1210 dmu_buf_t *db = dbp[i]; 1211 1212 ASSERT(size > 0); 1213 1214 bufoff = offset - db->db_offset; 1215 tocpy = MIN(db->db_size - bufoff, size); 1216 1217 ASSERT(db->db_data != NULL); 1218 (void) memcpy(buf, (char *)db->db_data + bufoff, tocpy); 1219 1220 offset += tocpy; 1221 size -= tocpy; 1222 buf = (char *)buf + tocpy; 1223 } 1224 dmu_buf_rele_array(dbp, numbufs, FTAG); 1225 } 1226 return (err); 1227 } 1228 1229 int 1230 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1231 void *buf, dmu_flags_t flags) 1232 { 1233 dnode_t *dn; 1234 int err; 1235 1236 err = dnode_hold(os, object, FTAG, &dn); 1237 if (err != 0) 1238 return (err); 1239 1240 err = dmu_read_impl(dn, offset, size, buf, flags); 1241 dnode_rele(dn, FTAG); 1242 return (err); 1243 } 1244 1245 int 1246 dmu_read_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, void *buf, 1247 dmu_flags_t flags) 1248 { 1249 return (dmu_read_impl(dn, offset, size, buf, flags)); 1250 } 1251 1252 static void 1253 dmu_write_impl(dmu_buf_t **dbp, int numbufs, uint64_t offset, uint64_t size, 1254 const void *buf, dmu_tx_t *tx, dmu_flags_t flags) 1255 { 1256 int i; 1257 1258 for (i = 0; i < numbufs; i++) { 1259 uint64_t tocpy; 1260 int64_t bufoff; 1261 dmu_buf_t *db = dbp[i]; 1262 1263 ASSERT(size > 0); 1264 1265 bufoff = offset - db->db_offset; 1266 tocpy = MIN(db->db_size - bufoff, size); 1267 1268 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 1269 1270 if (tocpy == db->db_size) { 1271 dmu_buf_will_fill_flags(db, tx, B_FALSE, flags); 1272 } else { 1273 if (i == numbufs - 1 && bufoff + tocpy < db->db_size) { 1274 if (bufoff == 0) 1275 flags |= DMU_PARTIAL_FIRST; 1276 else 1277 flags |= DMU_PARTIAL_MORE; 1278 } 1279 dmu_buf_will_dirty_flags(db, tx, flags); 1280 } 1281 1282 ASSERT(db->db_data != NULL); 1283 (void) memcpy((char *)db->db_data + bufoff, buf, tocpy); 1284 1285 if (tocpy == db->db_size) 1286 dmu_buf_fill_done(db, tx, B_FALSE); 1287 1288 offset += tocpy; 1289 size -= tocpy; 1290 buf = (char *)buf + tocpy; 1291 } 1292 } 1293 1294 void 1295 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1296 const void *buf, dmu_tx_t *tx) 1297 { 1298 dmu_buf_t **dbp; 1299 int numbufs; 1300 1301 if (size == 0) 1302 return; 1303 1304 VERIFY0(dmu_buf_hold_array(os, object, offset, size, 1305 FALSE, FTAG, &numbufs, &dbp)); 1306 dmu_write_impl(dbp, numbufs, offset, size, buf, tx, DMU_READ_PREFETCH); 1307 dmu_buf_rele_array(dbp, numbufs, FTAG); 1308 } 1309 1310 int 1311 dmu_write_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, 1312 const void *buf, dmu_tx_t *tx, dmu_flags_t flags) 1313 { 1314 dmu_buf_t **dbp; 1315 int numbufs; 1316 int error; 1317 1318 if (size == 0) 1319 return (0); 1320 1321 /* Allow Direct I/O when requested and properly aligned */ 1322 if ((flags & DMU_DIRECTIO) && zfs_dio_page_aligned((void *)buf) && 1323 zfs_dio_aligned(offset, size, dn->dn_datablksz)) { 1324 abd_t *data = abd_get_from_buf((void *)buf, size); 1325 error = dmu_write_abd(dn, offset, size, data, flags, tx); 1326 abd_free(data); 1327 return (error); 1328 } 1329 flags &= ~DMU_DIRECTIO; 1330 1331 VERIFY0(dmu_buf_hold_array_by_dnode(dn, offset, size, 1332 FALSE, FTAG, &numbufs, &dbp, flags)); 1333 dmu_write_impl(dbp, numbufs, offset, size, buf, tx, flags); 1334 dmu_buf_rele_array(dbp, numbufs, FTAG); 1335 return (0); 1336 } 1337 1338 void 1339 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1340 dmu_tx_t *tx) 1341 { 1342 dmu_buf_t **dbp; 1343 int numbufs, i; 1344 1345 if (size == 0) 1346 return; 1347 1348 VERIFY0(dmu_buf_hold_array(os, object, offset, size, 1349 FALSE, FTAG, &numbufs, &dbp)); 1350 1351 for (i = 0; i < numbufs; i++) { 1352 dmu_buf_t *db = dbp[i]; 1353 1354 dmu_buf_will_not_fill(db, tx); 1355 } 1356 dmu_buf_rele_array(dbp, numbufs, FTAG); 1357 } 1358 1359 void 1360 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset, 1361 void *data, uint8_t etype, uint8_t comp, int uncompressed_size, 1362 int compressed_size, int byteorder, dmu_tx_t *tx) 1363 { 1364 dmu_buf_t *db; 1365 1366 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES); 1367 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS); 1368 VERIFY0(dmu_buf_hold_noread(os, object, offset, 1369 FTAG, &db)); 1370 1371 dmu_buf_write_embedded(db, 1372 data, (bp_embedded_type_t)etype, (enum zio_compress)comp, 1373 uncompressed_size, compressed_size, byteorder, tx); 1374 1375 dmu_buf_rele(db, FTAG); 1376 } 1377 1378 void 1379 dmu_redact(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1380 dmu_tx_t *tx) 1381 { 1382 int numbufs, i; 1383 dmu_buf_t **dbp; 1384 1385 VERIFY0(dmu_buf_hold_array(os, object, offset, size, FALSE, FTAG, 1386 &numbufs, &dbp)); 1387 for (i = 0; i < numbufs; i++) 1388 dmu_buf_redact(dbp[i], tx); 1389 dmu_buf_rele_array(dbp, numbufs, FTAG); 1390 } 1391 1392 #ifdef _KERNEL 1393 int 1394 dmu_read_uio_dnode(dnode_t *dn, zfs_uio_t *uio, uint64_t size, 1395 dmu_flags_t flags) 1396 { 1397 dmu_buf_t **dbp; 1398 int numbufs, i, err; 1399 1400 if ((flags & DMU_DIRECTIO) && (uio->uio_extflg & UIO_DIRECT)) 1401 return (dmu_read_uio_direct(dn, uio, size, flags)); 1402 flags &= ~DMU_DIRECTIO; 1403 1404 /* 1405 * NB: we could do this block-at-a-time, but it's nice 1406 * to be reading in parallel. 1407 */ 1408 err = dmu_buf_hold_array_by_dnode(dn, zfs_uio_offset(uio), size, 1409 TRUE, FTAG, &numbufs, &dbp, flags); 1410 if (err) 1411 return (err); 1412 1413 for (i = 0; i < numbufs; i++) { 1414 uint64_t tocpy; 1415 int64_t bufoff; 1416 dmu_buf_t *db = dbp[i]; 1417 1418 ASSERT(size > 0); 1419 1420 bufoff = zfs_uio_offset(uio) - db->db_offset; 1421 tocpy = MIN(db->db_size - bufoff, size); 1422 1423 ASSERT(db->db_data != NULL); 1424 err = zfs_uio_fault_move((char *)db->db_data + bufoff, tocpy, 1425 UIO_READ, uio); 1426 1427 if (err) 1428 break; 1429 1430 size -= tocpy; 1431 } 1432 dmu_buf_rele_array(dbp, numbufs, FTAG); 1433 1434 return (err); 1435 } 1436 1437 /* 1438 * Read 'size' bytes into the uio buffer. 1439 * From object zdb->db_object. 1440 * Starting at zfs_uio_offset(uio). 1441 * 1442 * If the caller already has a dbuf in the target object 1443 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(), 1444 * because we don't have to find the dnode_t for the object. 1445 */ 1446 int 1447 dmu_read_uio_dbuf(dmu_buf_t *zdb, zfs_uio_t *uio, uint64_t size, 1448 dmu_flags_t flags) 1449 { 1450 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb; 1451 int err; 1452 1453 if (size == 0) 1454 return (0); 1455 1456 DB_DNODE_ENTER(db); 1457 err = dmu_read_uio_dnode(DB_DNODE(db), uio, size, flags); 1458 DB_DNODE_EXIT(db); 1459 1460 return (err); 1461 } 1462 1463 /* 1464 * Read 'size' bytes into the uio buffer. 1465 * From the specified object 1466 * Starting at offset zfs_uio_offset(uio). 1467 */ 1468 int 1469 dmu_read_uio(objset_t *os, uint64_t object, zfs_uio_t *uio, uint64_t size, 1470 dmu_flags_t flags) 1471 { 1472 dnode_t *dn; 1473 int err; 1474 1475 if (size == 0) 1476 return (0); 1477 1478 err = dnode_hold(os, object, FTAG, &dn); 1479 if (err) 1480 return (err); 1481 1482 err = dmu_read_uio_dnode(dn, uio, size, flags); 1483 1484 dnode_rele(dn, FTAG); 1485 1486 return (err); 1487 } 1488 1489 int 1490 dmu_write_uio_dnode(dnode_t *dn, zfs_uio_t *uio, uint64_t size, dmu_tx_t *tx, 1491 dmu_flags_t flags) 1492 { 1493 dmu_buf_t **dbp; 1494 int numbufs; 1495 int err = 0; 1496 uint64_t write_size; 1497 dmu_flags_t oflags = flags; 1498 1499 top: 1500 write_size = size; 1501 1502 /* 1503 * We only allow Direct I/O writes to happen if we are block 1504 * sized aligned. Otherwise, we pass the write off to the ARC. 1505 */ 1506 if ((flags & DMU_DIRECTIO) && (uio->uio_extflg & UIO_DIRECT) && 1507 (write_size >= dn->dn_datablksz)) { 1508 if (zfs_dio_aligned(zfs_uio_offset(uio), write_size, 1509 dn->dn_datablksz)) { 1510 return (dmu_write_uio_direct(dn, uio, size, flags, tx)); 1511 } else if (write_size > dn->dn_datablksz && 1512 zfs_dio_offset_aligned(zfs_uio_offset(uio), 1513 dn->dn_datablksz)) { 1514 write_size = 1515 dn->dn_datablksz * (write_size / dn->dn_datablksz); 1516 err = dmu_write_uio_direct(dn, uio, write_size, flags, 1517 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 flags &= ~DMU_DIRECTIO; 1530 1531 err = dmu_buf_hold_array_by_dnode(dn, zfs_uio_offset(uio), write_size, 1532 FALSE, FTAG, &numbufs, &dbp, flags); 1533 if (err) 1534 return (err); 1535 1536 for (int i = 0; i < numbufs; i++) { 1537 uint64_t tocpy; 1538 int64_t bufoff; 1539 dmu_buf_t *db = dbp[i]; 1540 1541 ASSERT(write_size > 0); 1542 1543 offset_t off = zfs_uio_offset(uio); 1544 bufoff = off - db->db_offset; 1545 tocpy = MIN(db->db_size - bufoff, write_size); 1546 1547 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 1548 1549 if (tocpy == db->db_size) { 1550 dmu_buf_will_fill_flags(db, tx, B_TRUE, flags); 1551 } else { 1552 if (i == numbufs - 1 && bufoff + tocpy < db->db_size) { 1553 if (bufoff == 0) 1554 flags |= DMU_PARTIAL_FIRST; 1555 else 1556 flags |= DMU_PARTIAL_MORE; 1557 } 1558 dmu_buf_will_dirty_flags(db, tx, flags); 1559 } 1560 1561 ASSERT(db->db_data != NULL); 1562 err = zfs_uio_fault_move((char *)db->db_data + bufoff, 1563 tocpy, UIO_WRITE, uio); 1564 1565 if (tocpy == db->db_size && dmu_buf_fill_done(db, tx, err)) { 1566 /* The fill was reverted. Undo any uio progress. */ 1567 zfs_uio_advance(uio, off - zfs_uio_offset(uio)); 1568 } 1569 1570 if (err) 1571 break; 1572 1573 write_size -= tocpy; 1574 size -= tocpy; 1575 } 1576 1577 IMPLY(err == 0, write_size == 0); 1578 1579 dmu_buf_rele_array(dbp, numbufs, FTAG); 1580 1581 if ((oflags & DMU_DIRECTIO) && (uio->uio_extflg & UIO_DIRECT) && 1582 err == 0 && size > 0) { 1583 flags = oflags; 1584 goto top; 1585 } 1586 IMPLY(err == 0, size == 0); 1587 1588 return (err); 1589 } 1590 1591 /* 1592 * Write 'size' bytes from the uio buffer. 1593 * To object zdb->db_object. 1594 * Starting at offset zfs_uio_offset(uio). 1595 * 1596 * If the caller already has a dbuf in the target object 1597 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(), 1598 * because we don't have to find the dnode_t for the object. 1599 */ 1600 int 1601 dmu_write_uio_dbuf(dmu_buf_t *zdb, zfs_uio_t *uio, uint64_t size, 1602 dmu_tx_t *tx, dmu_flags_t flags) 1603 { 1604 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb; 1605 int err; 1606 1607 if (size == 0) 1608 return (0); 1609 1610 DB_DNODE_ENTER(db); 1611 err = dmu_write_uio_dnode(DB_DNODE(db), uio, size, tx, flags); 1612 DB_DNODE_EXIT(db); 1613 1614 return (err); 1615 } 1616 1617 /* 1618 * Write 'size' bytes from the uio buffer. 1619 * To the specified object. 1620 * Starting at offset zfs_uio_offset(uio). 1621 */ 1622 int 1623 dmu_write_uio(objset_t *os, uint64_t object, zfs_uio_t *uio, uint64_t size, 1624 dmu_tx_t *tx, dmu_flags_t flags) 1625 { 1626 dnode_t *dn; 1627 int err; 1628 1629 if (size == 0) 1630 return (0); 1631 1632 err = dnode_hold(os, object, FTAG, &dn); 1633 if (err) 1634 return (err); 1635 1636 err = dmu_write_uio_dnode(dn, uio, size, tx, flags); 1637 1638 dnode_rele(dn, FTAG); 1639 1640 return (err); 1641 } 1642 #endif /* _KERNEL */ 1643 1644 static void 1645 dmu_cached_bps(spa_t *spa, blkptr_t *bps, uint_t nbps, 1646 uint64_t *l1sz, uint64_t *l2sz) 1647 { 1648 int cached_flags; 1649 1650 if (bps == NULL) 1651 return; 1652 1653 for (size_t blk_off = 0; blk_off < nbps; blk_off++) { 1654 blkptr_t *bp = &bps[blk_off]; 1655 1656 if (BP_IS_HOLE(bp)) 1657 continue; 1658 1659 cached_flags = arc_cached(spa, bp); 1660 if (cached_flags == 0) 1661 continue; 1662 1663 if ((cached_flags & (ARC_CACHED_IN_L1 | ARC_CACHED_IN_L2)) == 1664 ARC_CACHED_IN_L2) 1665 *l2sz += BP_GET_LSIZE(bp); 1666 else 1667 *l1sz += BP_GET_LSIZE(bp); 1668 } 1669 } 1670 1671 /* 1672 * Estimate DMU object cached size. 1673 */ 1674 int 1675 dmu_object_cached_size(objset_t *os, uint64_t object, 1676 uint64_t *l1sz, uint64_t *l2sz) 1677 { 1678 dnode_t *dn; 1679 dmu_object_info_t doi; 1680 int err = 0; 1681 1682 *l1sz = *l2sz = 0; 1683 1684 if (dnode_hold(os, object, FTAG, &dn) != 0) 1685 return (0); 1686 1687 if (dn->dn_nlevels < 2) { 1688 dnode_rele(dn, FTAG); 1689 return (0); 1690 } 1691 1692 dmu_object_info_from_dnode(dn, &doi); 1693 1694 for (uint64_t off = 0; off < doi.doi_max_offset && 1695 dmu_prefetch_max > 0; off += dmu_prefetch_max) { 1696 /* dbuf_read doesn't prefetch L1 blocks. */ 1697 dmu_prefetch_by_dnode(dn, 1, off, 1698 dmu_prefetch_max, ZIO_PRIORITY_SYNC_READ); 1699 } 1700 1701 /* 1702 * Hold all valid L1 blocks, asking ARC the status of each BP 1703 * contained in each such L1 block. 1704 */ 1705 uint_t nbps = bp_span_in_blocks(dn->dn_indblkshift, 1); 1706 uint64_t l1blks = 1 + (dn->dn_maxblkid / nbps); 1707 1708 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1709 for (uint64_t blk = 0; blk < l1blks; blk++) { 1710 dmu_buf_impl_t *db = NULL; 1711 1712 if (issig()) { 1713 /* 1714 * On interrupt, get out, and bubble up EINTR 1715 */ 1716 err = EINTR; 1717 break; 1718 } 1719 1720 /* 1721 * If we get an i/o error here, the L1 can't be read, 1722 * and nothing under it could be cached, so we just 1723 * continue. Ignoring the error from dbuf_hold_impl 1724 * or from dbuf_read is then a reasonable choice. 1725 */ 1726 err = dbuf_hold_impl(dn, 1, blk, B_TRUE, B_FALSE, FTAG, &db); 1727 if (err != 0) { 1728 /* 1729 * ignore error and continue 1730 */ 1731 err = 0; 1732 continue; 1733 } 1734 1735 err = dbuf_read(db, NULL, DB_RF_CANFAIL); 1736 if (err == 0) { 1737 dmu_cached_bps(dmu_objset_spa(os), db->db.db_data, 1738 nbps, l1sz, l2sz); 1739 } 1740 /* 1741 * error may be ignored, and we continue 1742 */ 1743 err = 0; 1744 dbuf_rele(db, FTAG); 1745 } 1746 rw_exit(&dn->dn_struct_rwlock); 1747 1748 dnode_rele(dn, FTAG); 1749 return (err); 1750 } 1751 1752 /* 1753 * Allocate a loaned anonymous arc buffer. 1754 */ 1755 arc_buf_t * 1756 dmu_request_arcbuf(dmu_buf_t *handle, int size) 1757 { 1758 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle; 1759 1760 return (arc_loan_buf(db->db_objset->os_spa, B_FALSE, size)); 1761 } 1762 1763 /* 1764 * Free a loaned arc buffer. 1765 */ 1766 void 1767 dmu_return_arcbuf(arc_buf_t *buf) 1768 { 1769 arc_return_buf(buf, FTAG); 1770 arc_buf_destroy(buf, FTAG); 1771 } 1772 1773 /* 1774 * A "lightweight" write is faster than a regular write (e.g. 1775 * dmu_write_by_dnode() or dmu_assign_arcbuf_by_dnode()), because it avoids the 1776 * CPU cost of creating a dmu_buf_impl_t and arc_buf_[hdr_]_t. However, the 1777 * data can not be read or overwritten until the transaction's txg has been 1778 * synced. This makes it appropriate for workloads that are known to be 1779 * (temporarily) write-only, like "zfs receive". 1780 * 1781 * A single block is written, starting at the specified offset in bytes. If 1782 * the call is successful, it returns 0 and the provided abd has been 1783 * consumed (the caller should not free it). 1784 */ 1785 int 1786 dmu_lightweight_write_by_dnode(dnode_t *dn, uint64_t offset, abd_t *abd, 1787 const zio_prop_t *zp, zio_flag_t flags, dmu_tx_t *tx) 1788 { 1789 dbuf_dirty_record_t *dr = 1790 dbuf_dirty_lightweight(dn, dbuf_whichblock(dn, 0, offset), tx); 1791 if (dr == NULL) 1792 return (SET_ERROR(EIO)); 1793 dr->dt.dll.dr_abd = abd; 1794 dr->dt.dll.dr_props = *zp; 1795 dr->dt.dll.dr_flags = flags; 1796 return (0); 1797 } 1798 1799 /* 1800 * When possible directly assign passed loaned arc buffer to a dbuf. 1801 * If this is not possible copy the contents of passed arc buf via 1802 * dmu_write(). 1803 */ 1804 int 1805 dmu_assign_arcbuf_by_dnode(dnode_t *dn, uint64_t offset, arc_buf_t *buf, 1806 dmu_tx_t *tx, dmu_flags_t flags) 1807 { 1808 dmu_buf_impl_t *db; 1809 objset_t *os = dn->dn_objset; 1810 uint32_t blksz = (uint32_t)arc_buf_lsize(buf); 1811 uint64_t blkid; 1812 1813 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1814 blkid = dbuf_whichblock(dn, 0, offset); 1815 db = dbuf_hold(dn, blkid, FTAG); 1816 rw_exit(&dn->dn_struct_rwlock); 1817 if (db == NULL) 1818 return (SET_ERROR(EIO)); 1819 1820 /* 1821 * We can only assign if the offset is aligned and the arc buf is the 1822 * same size as the dbuf. 1823 */ 1824 if (offset == db->db.db_offset && blksz == db->db.db_size) { 1825 zfs_racct_write(os->os_spa, blksz, 1, flags); 1826 dbuf_assign_arcbuf(db, buf, tx, flags); 1827 dbuf_rele(db, FTAG); 1828 } else { 1829 /* compressed bufs must always be assignable to their dbuf */ 1830 ASSERT3U(arc_get_compression(buf), ==, ZIO_COMPRESS_OFF); 1831 ASSERT(!(buf->b_flags & ARC_BUF_FLAG_COMPRESSED)); 1832 1833 dbuf_rele(db, FTAG); 1834 dmu_write_by_dnode(dn, offset, blksz, buf->b_data, tx, flags); 1835 dmu_return_arcbuf(buf); 1836 } 1837 1838 return (0); 1839 } 1840 1841 int 1842 dmu_assign_arcbuf_by_dbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf, 1843 dmu_tx_t *tx, dmu_flags_t flags) 1844 { 1845 int err; 1846 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle; 1847 1848 DB_DNODE_ENTER(db); 1849 err = dmu_assign_arcbuf_by_dnode(DB_DNODE(db), offset, buf, tx, flags); 1850 DB_DNODE_EXIT(db); 1851 1852 return (err); 1853 } 1854 1855 void 1856 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg) 1857 { 1858 (void) buf; 1859 dmu_sync_arg_t *dsa = varg; 1860 1861 if (zio->io_error == 0) { 1862 dbuf_dirty_record_t *dr = dsa->dsa_dr; 1863 blkptr_t *bp = zio->io_bp; 1864 1865 if (BP_IS_HOLE(bp)) { 1866 dmu_buf_t *db = NULL; 1867 if (dr) 1868 db = &(dr->dr_dbuf->db); 1869 else 1870 db = dsa->dsa_zgd->zgd_db; 1871 /* 1872 * A block of zeros may compress to a hole, but the 1873 * block size still needs to be known for replay. 1874 */ 1875 BP_SET_LSIZE(bp, db->db_size); 1876 } else if (!BP_IS_EMBEDDED(bp)) { 1877 ASSERT0(BP_GET_LEVEL(bp)); 1878 BP_SET_FILL(bp, 1); 1879 } 1880 } 1881 } 1882 1883 static void 1884 dmu_sync_late_arrival_ready(zio_t *zio) 1885 { 1886 dmu_sync_ready(zio, NULL, zio->io_private); 1887 } 1888 1889 void 1890 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg) 1891 { 1892 (void) buf; 1893 dmu_sync_arg_t *dsa = varg; 1894 dbuf_dirty_record_t *dr = dsa->dsa_dr; 1895 dmu_buf_impl_t *db = dr->dr_dbuf; 1896 zgd_t *zgd = dsa->dsa_zgd; 1897 1898 /* 1899 * Record the vdev(s) backing this blkptr so they can be flushed after 1900 * the writes for the lwb have completed. 1901 */ 1902 if (zgd && zio->io_error == 0) { 1903 zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp); 1904 } 1905 1906 mutex_enter(&db->db_mtx); 1907 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC); 1908 if (zio->io_error == 0) { 1909 ASSERT0(dr->dt.dl.dr_has_raw_params); 1910 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE); 1911 if (dr->dt.dl.dr_nopwrite) { 1912 blkptr_t *bp = zio->io_bp; 1913 blkptr_t *bp_orig = &zio->io_bp_orig; 1914 uint8_t chksum = BP_GET_CHECKSUM(bp_orig); 1915 1916 ASSERT(BP_EQUAL(bp, bp_orig)); 1917 VERIFY(BP_EQUAL(bp, db->db_blkptr)); 1918 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF); 1919 VERIFY(zio_checksum_table[chksum].ci_flags & 1920 ZCHECKSUM_FLAG_NOPWRITE); 1921 } 1922 dr->dt.dl.dr_overridden_by = *zio->io_bp; 1923 dr->dt.dl.dr_override_state = DR_OVERRIDDEN; 1924 dr->dt.dl.dr_copies = zio->io_prop.zp_copies; 1925 dr->dt.dl.dr_gang_copies = zio->io_prop.zp_gang_copies; 1926 1927 /* 1928 * Old style holes are filled with all zeros, whereas 1929 * new-style holes maintain their lsize, type, level, 1930 * and birth time (see zio_write_compress). While we 1931 * need to reset the BP_SET_LSIZE() call that happened 1932 * in dmu_sync_ready for old style holes, we do *not* 1933 * want to wipe out the information contained in new 1934 * style holes. Thus, only zero out the block pointer if 1935 * it's an old style hole. 1936 */ 1937 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) && 1938 BP_GET_LOGICAL_BIRTH(&dr->dt.dl.dr_overridden_by) == 0) 1939 BP_ZERO(&dr->dt.dl.dr_overridden_by); 1940 } else { 1941 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 1942 } 1943 1944 cv_broadcast(&db->db_changed); 1945 mutex_exit(&db->db_mtx); 1946 1947 if (dsa->dsa_done) 1948 dsa->dsa_done(dsa->dsa_zgd, zio->io_error); 1949 1950 kmem_free(dsa, sizeof (*dsa)); 1951 } 1952 1953 static void 1954 dmu_sync_late_arrival_done(zio_t *zio) 1955 { 1956 blkptr_t *bp = zio->io_bp; 1957 dmu_sync_arg_t *dsa = zio->io_private; 1958 zgd_t *zgd = dsa->dsa_zgd; 1959 1960 if (zio->io_error == 0) { 1961 /* 1962 * Record the vdev(s) backing this blkptr so they can be 1963 * flushed after the writes for the lwb have completed. 1964 */ 1965 zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp); 1966 1967 if (!BP_IS_HOLE(bp)) { 1968 blkptr_t *bp_orig __maybe_unused = &zio->io_bp_orig; 1969 ASSERT(!(zio->io_flags & ZIO_FLAG_NOPWRITE)); 1970 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig)); 1971 ASSERT(BP_GET_BIRTH(zio->io_bp) == zio->io_txg); 1972 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa)); 1973 zio_free(zio->io_spa, zio->io_txg, zio->io_bp); 1974 } 1975 } 1976 1977 dmu_tx_commit(dsa->dsa_tx); 1978 1979 dsa->dsa_done(dsa->dsa_zgd, zio->io_error); 1980 1981 abd_free(zio->io_abd); 1982 kmem_free(dsa, sizeof (*dsa)); 1983 } 1984 1985 static int 1986 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd, 1987 zio_prop_t *zp, zbookmark_phys_t *zb) 1988 { 1989 dmu_sync_arg_t *dsa; 1990 dmu_tx_t *tx; 1991 int error; 1992 1993 error = dbuf_read((dmu_buf_impl_t *)zgd->zgd_db, NULL, 1994 DB_RF_CANFAIL | DMU_READ_NO_PREFETCH | DMU_KEEP_CACHING); 1995 if (error != 0) 1996 return (error); 1997 1998 tx = dmu_tx_create(os); 1999 dmu_tx_hold_space(tx, zgd->zgd_db->db_size); 2000 /* 2001 * This transaction does not produce any dirty data or log blocks, so 2002 * it should not be throttled. All other cases wait for TXG sync, by 2003 * which time the log block we are writing will be obsolete, so we can 2004 * skip waiting and just return error here instead. 2005 */ 2006 if (dmu_tx_assign(tx, DMU_TX_NOWAIT | DMU_TX_NOTHROTTLE) != 0) { 2007 dmu_tx_abort(tx); 2008 /* Make zl_get_data do txg_waited_synced() */ 2009 return (SET_ERROR(EIO)); 2010 } 2011 2012 /* 2013 * In order to prevent the zgd's lwb from being free'd prior to 2014 * dmu_sync_late_arrival_done() being called, we have to ensure 2015 * the lwb's "max txg" takes this tx's txg into account. 2016 */ 2017 zil_lwb_add_txg(zgd->zgd_lwb, dmu_tx_get_txg(tx)); 2018 2019 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP); 2020 dsa->dsa_dr = NULL; 2021 dsa->dsa_done = done; 2022 dsa->dsa_zgd = zgd; 2023 dsa->dsa_tx = tx; 2024 2025 /* 2026 * Since we are currently syncing this txg, it's nontrivial to 2027 * determine what BP to nopwrite against, so we disable nopwrite. 2028 * 2029 * When syncing, the db_blkptr is initially the BP of the previous 2030 * txg. We can not nopwrite against it because it will be changed 2031 * (this is similar to the non-late-arrival case where the dbuf is 2032 * dirty in a future txg). 2033 * 2034 * Then dbuf_write_ready() sets bp_blkptr to the location we will write. 2035 * We can not nopwrite against it because although the BP will not 2036 * (typically) be changed, the data has not yet been persisted to this 2037 * location. 2038 * 2039 * Finally, when dbuf_write_done() is called, it is theoretically 2040 * possible to always nopwrite, because the data that was written in 2041 * this txg is the same data that we are trying to write. However we 2042 * would need to check that this dbuf is not dirty in any future 2043 * txg's (as we do in the normal dmu_sync() path). For simplicity, we 2044 * don't nopwrite in this case. 2045 */ 2046 zp->zp_nopwrite = B_FALSE; 2047 2048 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp, 2049 abd_get_from_buf(zgd->zgd_db->db_data, zgd->zgd_db->db_size), 2050 zgd->zgd_db->db_size, zgd->zgd_db->db_size, zp, 2051 dmu_sync_late_arrival_ready, NULL, dmu_sync_late_arrival_done, 2052 dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb)); 2053 2054 return (0); 2055 } 2056 2057 /* 2058 * Intent log support: sync the block associated with db to disk. 2059 * N.B. and XXX: the caller is responsible for making sure that the 2060 * data isn't changing while dmu_sync() is writing it. 2061 * 2062 * Return values: 2063 * 2064 * EEXIST: this txg has already been synced, so there's nothing to do. 2065 * The caller should not log the write. 2066 * 2067 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do. 2068 * The caller should not log the write. 2069 * 2070 * EALREADY: this block is already in the process of being synced. 2071 * The caller should track its progress (somehow). 2072 * 2073 * EIO: could not do the I/O. 2074 * The caller should do a txg_wait_synced(). 2075 * 2076 * 0: the I/O has been initiated. 2077 * The caller should log this blkptr in the done callback. 2078 * It is possible that the I/O will fail, in which case 2079 * the error will be reported to the done callback and 2080 * propagated to pio from zio_done(). 2081 */ 2082 int 2083 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd) 2084 { 2085 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db; 2086 objset_t *os = db->db_objset; 2087 dsl_dataset_t *ds = os->os_dsl_dataset; 2088 dbuf_dirty_record_t *dr, *dr_next; 2089 dmu_sync_arg_t *dsa; 2090 zbookmark_phys_t zb; 2091 zio_prop_t zp; 2092 2093 ASSERT(pio != NULL); 2094 ASSERT(txg != 0); 2095 2096 SET_BOOKMARK(&zb, ds->ds_object, 2097 db->db.db_object, db->db_level, db->db_blkid); 2098 2099 DB_DNODE_ENTER(db); 2100 dmu_write_policy(os, DB_DNODE(db), db->db_level, WP_DMU_SYNC, &zp); 2101 DB_DNODE_EXIT(db); 2102 2103 /* 2104 * If we're frozen (running ziltest), we always need to generate a bp. 2105 */ 2106 if (txg > spa_freeze_txg(os->os_spa)) 2107 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); 2108 2109 /* 2110 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf() 2111 * and us. If we determine that this txg is not yet syncing, 2112 * but it begins to sync a moment later, that's OK because the 2113 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx. 2114 */ 2115 mutex_enter(&db->db_mtx); 2116 2117 if (txg <= spa_last_synced_txg(os->os_spa)) { 2118 /* 2119 * This txg has already synced. There's nothing to do. 2120 */ 2121 mutex_exit(&db->db_mtx); 2122 return (SET_ERROR(EEXIST)); 2123 } 2124 2125 if (txg <= spa_syncing_txg(os->os_spa)) { 2126 /* 2127 * This txg is currently syncing, so we can't mess with 2128 * the dirty record anymore; just write a new log block. 2129 */ 2130 mutex_exit(&db->db_mtx); 2131 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); 2132 } 2133 2134 dr = dbuf_find_dirty_eq(db, txg); 2135 2136 if (dr == NULL) { 2137 /* 2138 * There's no dr for this dbuf, so it must have been freed. 2139 * There's no need to log writes to freed blocks, so we're done. 2140 */ 2141 mutex_exit(&db->db_mtx); 2142 return (SET_ERROR(ENOENT)); 2143 } 2144 2145 dr_next = list_next(&db->db_dirty_records, dr); 2146 ASSERT(dr_next == NULL || dr_next->dr_txg < txg); 2147 2148 if (db->db_blkptr != NULL) { 2149 /* 2150 * We need to fill in zgd_bp with the current blkptr so that 2151 * the nopwrite code can check if we're writing the same 2152 * data that's already on disk. We can only nopwrite if we 2153 * are sure that after making the copy, db_blkptr will not 2154 * change until our i/o completes. We ensure this by 2155 * holding the db_mtx, and only allowing nopwrite if the 2156 * block is not already dirty (see below). This is verified 2157 * by dmu_sync_done(), which VERIFYs that the db_blkptr has 2158 * not changed. 2159 */ 2160 *zgd->zgd_bp = *db->db_blkptr; 2161 } 2162 2163 /* 2164 * Assume the on-disk data is X, the current syncing data (in 2165 * txg - 1) is Y, and the current in-memory data is Z (currently 2166 * in dmu_sync). 2167 * 2168 * We usually want to perform a nopwrite if X and Z are the 2169 * same. However, if Y is different (i.e. the BP is going to 2170 * change before this write takes effect), then a nopwrite will 2171 * be incorrect - we would override with X, which could have 2172 * been freed when Y was written. 2173 * 2174 * (Note that this is not a concern when we are nop-writing from 2175 * syncing context, because X and Y must be identical, because 2176 * all previous txgs have been synced.) 2177 * 2178 * Therefore, we disable nopwrite if the current BP could change 2179 * before this TXG. There are two ways it could change: by 2180 * being dirty (dr_next is non-NULL), or by being freed 2181 * (dnode_block_freed()). This behavior is verified by 2182 * zio_done(), which VERIFYs that the override BP is identical 2183 * to the on-disk BP. 2184 */ 2185 if (dr_next != NULL) { 2186 zp.zp_nopwrite = B_FALSE; 2187 } else { 2188 DB_DNODE_ENTER(db); 2189 if (dnode_block_freed(DB_DNODE(db), db->db_blkid)) 2190 zp.zp_nopwrite = B_FALSE; 2191 DB_DNODE_EXIT(db); 2192 } 2193 2194 ASSERT(dr->dr_txg == txg); 2195 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC || 2196 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { 2197 /* 2198 * We have already issued a sync write for this buffer, 2199 * or this buffer has already been synced. It could not 2200 * have been dirtied since, or we would have cleared the state. 2201 */ 2202 mutex_exit(&db->db_mtx); 2203 return (SET_ERROR(EALREADY)); 2204 } 2205 2206 ASSERT0(dr->dt.dl.dr_has_raw_params); 2207 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); 2208 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC; 2209 mutex_exit(&db->db_mtx); 2210 2211 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP); 2212 dsa->dsa_dr = dr; 2213 dsa->dsa_done = done; 2214 dsa->dsa_zgd = zgd; 2215 dsa->dsa_tx = NULL; 2216 2217 zio_nowait(arc_write(pio, os->os_spa, txg, zgd->zgd_bp, 2218 dr->dt.dl.dr_data, !DBUF_IS_CACHEABLE(db), 2219 dbuf_is_l2cacheable(db, NULL), &zp, dmu_sync_ready, NULL, 2220 dmu_sync_done, dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, 2221 &zb)); 2222 2223 return (0); 2224 } 2225 2226 int 2227 dmu_object_set_nlevels(objset_t *os, uint64_t object, int nlevels, dmu_tx_t *tx) 2228 { 2229 dnode_t *dn; 2230 int err; 2231 2232 err = dnode_hold(os, object, FTAG, &dn); 2233 if (err) 2234 return (err); 2235 err = dnode_set_nlevels(dn, nlevels, tx); 2236 dnode_rele(dn, FTAG); 2237 return (err); 2238 } 2239 2240 int 2241 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs, 2242 dmu_tx_t *tx) 2243 { 2244 dnode_t *dn; 2245 int err; 2246 2247 err = dnode_hold(os, object, FTAG, &dn); 2248 if (err) 2249 return (err); 2250 err = dnode_set_blksz(dn, size, ibs, tx); 2251 dnode_rele(dn, FTAG); 2252 return (err); 2253 } 2254 2255 int 2256 dmu_object_set_maxblkid(objset_t *os, uint64_t object, uint64_t maxblkid, 2257 dmu_tx_t *tx) 2258 { 2259 dnode_t *dn; 2260 int err; 2261 2262 err = dnode_hold(os, object, FTAG, &dn); 2263 if (err) 2264 return (err); 2265 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 2266 dnode_new_blkid(dn, maxblkid, tx, B_FALSE, B_TRUE); 2267 rw_exit(&dn->dn_struct_rwlock); 2268 dnode_rele(dn, FTAG); 2269 return (0); 2270 } 2271 2272 void 2273 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum, 2274 dmu_tx_t *tx) 2275 { 2276 dnode_t *dn; 2277 2278 /* 2279 * Send streams include each object's checksum function. This 2280 * check ensures that the receiving system can understand the 2281 * checksum function transmitted. 2282 */ 2283 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS); 2284 2285 VERIFY0(dnode_hold(os, object, FTAG, &dn)); 2286 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS); 2287 dn->dn_checksum = checksum; 2288 dnode_setdirty(dn, tx); 2289 dnode_rele(dn, FTAG); 2290 } 2291 2292 void 2293 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress, 2294 dmu_tx_t *tx) 2295 { 2296 dnode_t *dn; 2297 2298 /* 2299 * Send streams include each object's compression function. This 2300 * check ensures that the receiving system can understand the 2301 * compression function transmitted. 2302 */ 2303 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS); 2304 2305 VERIFY0(dnode_hold(os, object, FTAG, &dn)); 2306 dn->dn_compress = compress; 2307 dnode_setdirty(dn, tx); 2308 dnode_rele(dn, FTAG); 2309 } 2310 2311 /* 2312 * When the "redundant_metadata" property is set to "most", only indirect 2313 * blocks of this level and higher will have an additional ditto block. 2314 */ 2315 static const int zfs_redundant_metadata_most_ditto_level = 2; 2316 2317 void 2318 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp) 2319 { 2320 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET; 2321 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) || 2322 (wp & WP_SPILL)); 2323 enum zio_checksum checksum = os->os_checksum; 2324 enum zio_compress compress = os->os_compress; 2325 uint8_t complevel = os->os_complevel; 2326 enum zio_checksum dedup_checksum = os->os_dedup_checksum; 2327 boolean_t dedup = B_FALSE; 2328 boolean_t nopwrite = B_FALSE; 2329 boolean_t dedup_verify = os->os_dedup_verify; 2330 boolean_t encrypt = B_FALSE; 2331 int copies = os->os_copies; 2332 int gang_copies = os->os_copies; 2333 2334 /* 2335 * We maintain different write policies for each of the following 2336 * types of data: 2337 * 1. metadata 2338 * 2. preallocated blocks (i.e. level-0 blocks of a dump device) 2339 * 3. all other level 0 blocks 2340 */ 2341 if (ismd) { 2342 /* 2343 * XXX -- we should design a compression algorithm 2344 * that specializes in arrays of bps. 2345 */ 2346 compress = zio_compress_select(os->os_spa, 2347 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON); 2348 2349 /* 2350 * Metadata always gets checksummed. If the data 2351 * checksum is multi-bit correctable, and it's not a 2352 * ZBT-style checksum, then it's suitable for metadata 2353 * as well. Otherwise, the metadata checksum defaults 2354 * to fletcher4. 2355 */ 2356 if (!(zio_checksum_table[checksum].ci_flags & 2357 ZCHECKSUM_FLAG_METADATA) || 2358 (zio_checksum_table[checksum].ci_flags & 2359 ZCHECKSUM_FLAG_EMBEDDED)) 2360 checksum = ZIO_CHECKSUM_FLETCHER_4; 2361 2362 switch (os->os_redundant_metadata) { 2363 case ZFS_REDUNDANT_METADATA_ALL: 2364 copies++; 2365 gang_copies++; 2366 break; 2367 case ZFS_REDUNDANT_METADATA_MOST: 2368 if (level >= zfs_redundant_metadata_most_ditto_level || 2369 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL)) 2370 copies++; 2371 if (level + 1 >= 2372 zfs_redundant_metadata_most_ditto_level || 2373 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL)) 2374 gang_copies++; 2375 break; 2376 case ZFS_REDUNDANT_METADATA_SOME: 2377 if (DMU_OT_IS_CRITICAL(type, level)) { 2378 copies++; 2379 gang_copies++; 2380 } else if (DMU_OT_IS_METADATA(type)) { 2381 gang_copies++; 2382 } 2383 break; 2384 case ZFS_REDUNDANT_METADATA_NONE: 2385 break; 2386 } 2387 2388 if (dmu_ddt_copies > 0) { 2389 /* 2390 * If this tunable is set, and this is a write for a 2391 * dedup entry store (zap or log), then we treat it 2392 * something like ZFS_REDUNDANT_METADATA_MOST on a 2393 * regular dataset: this many copies, and one more for 2394 * "higher" indirect blocks. This specific exception is 2395 * necessary because dedup objects are stored in the 2396 * MOS, which always has the highest possible copies. 2397 */ 2398 dmu_object_type_t stype = 2399 dn ? dn->dn_storage_type : DMU_OT_NONE; 2400 if (stype == DMU_OT_NONE) 2401 stype = type; 2402 if (stype == DMU_OT_DDT_ZAP) { 2403 copies = dmu_ddt_copies; 2404 if (level >= 2405 zfs_redundant_metadata_most_ditto_level) 2406 copies++; 2407 } 2408 } 2409 } else if (wp & WP_NOFILL) { 2410 ASSERT0(level); 2411 2412 /* 2413 * If we're writing preallocated blocks, we aren't actually 2414 * writing them so don't set any policy properties. These 2415 * blocks are currently only used by an external subsystem 2416 * outside of zfs (i.e. dump) and not written by the zio 2417 * pipeline. 2418 */ 2419 compress = ZIO_COMPRESS_OFF; 2420 checksum = ZIO_CHECKSUM_OFF; 2421 } else { 2422 compress = zio_compress_select(os->os_spa, dn->dn_compress, 2423 compress); 2424 complevel = zio_complevel_select(os->os_spa, compress, 2425 complevel, complevel); 2426 2427 /* 2428 * Storing many references to an all zeros block in the dedup 2429 * table would be expensive. Instead, if dedup is enabled, 2430 * store them as holes even if compression is not enabled. 2431 */ 2432 if (compress == ZIO_COMPRESS_OFF && 2433 dedup_checksum != ZIO_CHECKSUM_OFF) 2434 compress = ZIO_COMPRESS_EMPTY; 2435 2436 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ? 2437 zio_checksum_select(dn->dn_checksum, checksum) : 2438 dedup_checksum; 2439 2440 /* 2441 * Determine dedup setting. If we are in dmu_sync(), 2442 * we won't actually dedup now because that's all 2443 * done in syncing context; but we do want to use the 2444 * dedup checksum. If the checksum is not strong 2445 * enough to ensure unique signatures, force 2446 * dedup_verify. 2447 */ 2448 if (dedup_checksum != ZIO_CHECKSUM_OFF) { 2449 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE; 2450 if (!(zio_checksum_table[checksum].ci_flags & 2451 ZCHECKSUM_FLAG_DEDUP)) 2452 dedup_verify = B_TRUE; 2453 } 2454 2455 /* 2456 * Enable nopwrite if we have secure enough checksum 2457 * algorithm (see comment in zio_nop_write) and 2458 * compression is enabled. We don't enable nopwrite if 2459 * dedup is enabled as the two features are mutually 2460 * exclusive. 2461 */ 2462 nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags & 2463 ZCHECKSUM_FLAG_NOPWRITE) && 2464 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled); 2465 2466 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL || 2467 (os->os_redundant_metadata == 2468 ZFS_REDUNDANT_METADATA_MOST && 2469 zfs_redundant_metadata_most_ditto_level <= 1)) 2470 gang_copies++; 2471 } 2472 2473 /* 2474 * All objects in an encrypted objset are protected from modification 2475 * via a MAC. Encrypted objects store their IV and salt in the last DVA 2476 * in the bp, so we cannot use all copies. Encrypted objects are also 2477 * not subject to nopwrite since writing the same data will still 2478 * result in a new ciphertext. Only encrypted blocks can be dedup'd 2479 * to avoid ambiguity in the dedup code since the DDT does not store 2480 * object types. 2481 */ 2482 if (os->os_encrypted && (wp & WP_NOFILL) == 0) { 2483 encrypt = B_TRUE; 2484 2485 if (DMU_OT_IS_ENCRYPTED(type)) { 2486 copies = MIN(copies, SPA_DVAS_PER_BP - 1); 2487 gang_copies = MIN(gang_copies, SPA_DVAS_PER_BP - 1); 2488 nopwrite = B_FALSE; 2489 } else { 2490 dedup = B_FALSE; 2491 } 2492 2493 if (level <= 0 && 2494 (type == DMU_OT_DNODE || type == DMU_OT_OBJSET)) { 2495 compress = ZIO_COMPRESS_EMPTY; 2496 } 2497 } 2498 2499 zp->zp_compress = compress; 2500 zp->zp_complevel = complevel; 2501 zp->zp_checksum = checksum; 2502 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type; 2503 zp->zp_level = level; 2504 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa)); 2505 zp->zp_gang_copies = MIN(MAX(gang_copies, copies), 2506 spa_max_replication(os->os_spa)); 2507 zp->zp_dedup = dedup; 2508 zp->zp_dedup_verify = dedup && dedup_verify; 2509 zp->zp_nopwrite = nopwrite; 2510 zp->zp_encrypt = encrypt; 2511 zp->zp_byteorder = ZFS_HOST_BYTEORDER; 2512 zp->zp_direct_write = (wp & WP_DIRECT_WR) ? B_TRUE : B_FALSE; 2513 zp->zp_rewrite = B_FALSE; 2514 memset(zp->zp_salt, 0, ZIO_DATA_SALT_LEN); 2515 memset(zp->zp_iv, 0, ZIO_DATA_IV_LEN); 2516 memset(zp->zp_mac, 0, ZIO_DATA_MAC_LEN); 2517 zp->zp_zpl_smallblk = (DMU_OT_IS_FILE(zp->zp_type) || 2518 zp->zp_type == DMU_OT_ZVOL) ? 2519 os->os_zpl_special_smallblock : 0; 2520 zp->zp_storage_type = dn ? dn->dn_storage_type : DMU_OT_NONE; 2521 2522 ASSERT3U(zp->zp_compress, !=, ZIO_COMPRESS_INHERIT); 2523 } 2524 2525 /* 2526 * Reports the location of data and holes in an object. In order to 2527 * accurately report holes all dirty data must be synced to disk. This 2528 * causes extremely poor performance when seeking for holes in a dirty file. 2529 * As a compromise, only provide hole data when the dnode is clean. When 2530 * a dnode is dirty report the dnode as having no holes by returning EBUSY 2531 * which is always safe to do. 2532 */ 2533 int 2534 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off) 2535 { 2536 dnode_t *dn; 2537 uint64_t txg, maxtxg = 0; 2538 int err; 2539 2540 restart: 2541 err = dnode_hold(os, object, FTAG, &dn); 2542 if (err) 2543 return (err); 2544 2545 rw_enter(&dn->dn_struct_rwlock, RW_READER); 2546 2547 if (dnode_is_dirty(dn)) { 2548 /* 2549 * If the zfs_dmu_offset_next_sync module option is enabled 2550 * then hole reporting has been requested. Dirty dnodes 2551 * must be synced to disk to accurately report holes. 2552 * 2553 * Provided a RL_READER rangelock spanning 0-UINT64_MAX is 2554 * held by the caller only limited restarts will be required. 2555 * We tolerate callers which do not hold the rangelock by 2556 * returning EBUSY and not reporting holes after at most 2557 * TXG_CONCURRENT_STATES (3) restarts. 2558 */ 2559 if (zfs_dmu_offset_next_sync) { 2560 rw_exit(&dn->dn_struct_rwlock); 2561 dnode_rele(dn, FTAG); 2562 2563 if (maxtxg == 0) { 2564 txg = spa_last_synced_txg(dmu_objset_spa(os)); 2565 maxtxg = txg + TXG_CONCURRENT_STATES; 2566 } else if (txg >= maxtxg) 2567 return (SET_ERROR(EBUSY)); 2568 2569 txg_wait_synced(dmu_objset_pool(os), ++txg); 2570 goto restart; 2571 } 2572 2573 err = SET_ERROR(EBUSY); 2574 } else { 2575 err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK | 2576 (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0); 2577 } 2578 2579 rw_exit(&dn->dn_struct_rwlock); 2580 dnode_rele(dn, FTAG); 2581 2582 return (err); 2583 } 2584 2585 int 2586 dmu_read_l0_bps(objset_t *os, uint64_t object, uint64_t offset, uint64_t length, 2587 blkptr_t *bps, size_t *nbpsp) 2588 { 2589 dmu_buf_t **dbp, *dbuf; 2590 dmu_buf_impl_t *db; 2591 blkptr_t *bp; 2592 int error, numbufs; 2593 2594 error = dmu_buf_hold_array(os, object, offset, length, FALSE, FTAG, 2595 &numbufs, &dbp); 2596 if (error != 0) { 2597 if (error == ESRCH) { 2598 error = SET_ERROR(ENXIO); 2599 } 2600 return (error); 2601 } 2602 2603 ASSERT3U(numbufs, <=, *nbpsp); 2604 2605 for (int i = 0; i < numbufs; i++) { 2606 dbuf = dbp[i]; 2607 db = (dmu_buf_impl_t *)dbuf; 2608 2609 mutex_enter(&db->db_mtx); 2610 2611 if (!list_is_empty(&db->db_dirty_records)) { 2612 dbuf_dirty_record_t *dr; 2613 2614 dr = list_head(&db->db_dirty_records); 2615 if (dr->dt.dl.dr_brtwrite) { 2616 /* 2617 * This is very special case where we clone a 2618 * block and in the same transaction group we 2619 * read its BP (most likely to clone the clone). 2620 */ 2621 bp = &dr->dt.dl.dr_overridden_by; 2622 } else { 2623 /* 2624 * The block was modified in the same 2625 * transaction group. 2626 */ 2627 mutex_exit(&db->db_mtx); 2628 error = SET_ERROR(EAGAIN); 2629 goto out; 2630 } 2631 } else { 2632 bp = db->db_blkptr; 2633 } 2634 2635 mutex_exit(&db->db_mtx); 2636 2637 if (bp == NULL) { 2638 /* 2639 * The file size was increased, but the block was never 2640 * written, otherwise we would either have the block 2641 * pointer or the dirty record and would not get here. 2642 * It is effectively a hole, so report it as such. 2643 */ 2644 BP_ZERO(&bps[i]); 2645 continue; 2646 } 2647 /* 2648 * Make sure we clone only data blocks. 2649 */ 2650 if (BP_IS_METADATA(bp) && !BP_IS_HOLE(bp)) { 2651 error = SET_ERROR(EINVAL); 2652 goto out; 2653 } 2654 2655 /* 2656 * If the block was allocated in transaction group that is not 2657 * yet synced, we could clone it, but we couldn't write this 2658 * operation into ZIL, or it may be impossible to replay, since 2659 * the block may appear not yet allocated at that point. 2660 */ 2661 if (BP_GET_PHYSICAL_BIRTH(bp) > spa_freeze_txg(os->os_spa)) { 2662 error = SET_ERROR(EINVAL); 2663 goto out; 2664 } 2665 if (BP_GET_PHYSICAL_BIRTH(bp) > 2666 spa_last_synced_txg(os->os_spa)) { 2667 error = SET_ERROR(EAGAIN); 2668 goto out; 2669 } 2670 2671 bps[i] = *bp; 2672 } 2673 2674 *nbpsp = numbufs; 2675 out: 2676 dmu_buf_rele_array(dbp, numbufs, FTAG); 2677 2678 return (error); 2679 } 2680 2681 int 2682 dmu_brt_clone(objset_t *os, uint64_t object, uint64_t offset, uint64_t length, 2683 dmu_tx_t *tx, const blkptr_t *bps, size_t nbps) 2684 { 2685 spa_t *spa; 2686 dmu_buf_t **dbp, *dbuf; 2687 dmu_buf_impl_t *db; 2688 struct dirty_leaf *dl; 2689 dbuf_dirty_record_t *dr; 2690 const blkptr_t *bp; 2691 int error = 0, i, numbufs; 2692 2693 spa = os->os_spa; 2694 2695 VERIFY0(dmu_buf_hold_array(os, object, offset, length, FALSE, FTAG, 2696 &numbufs, &dbp)); 2697 ASSERT3U(nbps, ==, numbufs); 2698 2699 /* 2700 * Before we start cloning make sure that the dbufs sizes match new BPs 2701 * sizes. If they don't, that's a no-go, as we are not able to shrink 2702 * dbufs. 2703 */ 2704 for (i = 0; i < numbufs; i++) { 2705 dbuf = dbp[i]; 2706 db = (dmu_buf_impl_t *)dbuf; 2707 bp = &bps[i]; 2708 2709 ASSERT3U(db->db.db_object, !=, DMU_META_DNODE_OBJECT); 2710 ASSERT0(db->db_level); 2711 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 2712 ASSERT(db->db_blkid != DMU_SPILL_BLKID); 2713 2714 if (!BP_IS_HOLE(bp) && BP_GET_LSIZE(bp) != dbuf->db_size) { 2715 error = SET_ERROR(EXDEV); 2716 goto out; 2717 } 2718 } 2719 2720 for (i = 0; i < numbufs; i++) { 2721 dbuf = dbp[i]; 2722 db = (dmu_buf_impl_t *)dbuf; 2723 bp = &bps[i]; 2724 2725 dmu_buf_will_clone_or_dio(dbuf, tx); 2726 2727 mutex_enter(&db->db_mtx); 2728 2729 dr = list_head(&db->db_dirty_records); 2730 VERIFY(dr != NULL); 2731 ASSERT3U(dr->dr_txg, ==, tx->tx_txg); 2732 dl = &dr->dt.dl; 2733 ASSERT0(dl->dr_has_raw_params); 2734 dl->dr_overridden_by = *bp; 2735 if (!BP_IS_HOLE(bp) || BP_GET_LOGICAL_BIRTH(bp) != 0) { 2736 if (!BP_IS_EMBEDDED(bp)) { 2737 BP_SET_BIRTH(&dl->dr_overridden_by, dr->dr_txg, 2738 BP_GET_PHYSICAL_BIRTH(bp)); 2739 BP_SET_REWRITE(&dl->dr_overridden_by, 0); 2740 } else { 2741 BP_SET_LOGICAL_BIRTH(&dl->dr_overridden_by, 2742 dr->dr_txg); 2743 } 2744 } 2745 dl->dr_brtwrite = B_TRUE; 2746 dl->dr_override_state = DR_OVERRIDDEN; 2747 2748 mutex_exit(&db->db_mtx); 2749 2750 /* 2751 * When data in embedded into BP there is no need to create 2752 * BRT entry as there is no data block. Just copy the BP as 2753 * it contains the data. 2754 */ 2755 if (!BP_IS_HOLE(bp) && !BP_IS_EMBEDDED(bp)) { 2756 brt_pending_add(spa, bp, tx); 2757 } 2758 } 2759 out: 2760 dmu_buf_rele_array(dbp, numbufs, FTAG); 2761 2762 return (error); 2763 } 2764 2765 void 2766 __dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi) 2767 { 2768 dnode_phys_t *dnp = dn->dn_phys; 2769 2770 doi->doi_data_block_size = dn->dn_datablksz; 2771 doi->doi_metadata_block_size = dn->dn_indblkshift ? 2772 1ULL << dn->dn_indblkshift : 0; 2773 doi->doi_type = dn->dn_type; 2774 doi->doi_bonus_type = dn->dn_bonustype; 2775 doi->doi_bonus_size = dn->dn_bonuslen; 2776 doi->doi_dnodesize = dn->dn_num_slots << DNODE_SHIFT; 2777 doi->doi_indirection = dn->dn_nlevels; 2778 doi->doi_checksum = dn->dn_checksum; 2779 doi->doi_compress = dn->dn_compress; 2780 doi->doi_nblkptr = dn->dn_nblkptr; 2781 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9; 2782 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz; 2783 doi->doi_fill_count = 0; 2784 for (int i = 0; i < dnp->dn_nblkptr; i++) 2785 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]); 2786 } 2787 2788 void 2789 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi) 2790 { 2791 rw_enter(&dn->dn_struct_rwlock, RW_READER); 2792 mutex_enter(&dn->dn_mtx); 2793 2794 __dmu_object_info_from_dnode(dn, doi); 2795 2796 mutex_exit(&dn->dn_mtx); 2797 rw_exit(&dn->dn_struct_rwlock); 2798 } 2799 2800 /* 2801 * Get information on a DMU object. 2802 * If doi is NULL, just indicates whether the object exists. 2803 */ 2804 int 2805 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi) 2806 { 2807 dnode_t *dn; 2808 int err = dnode_hold(os, object, FTAG, &dn); 2809 2810 if (err) 2811 return (err); 2812 2813 if (doi != NULL) 2814 dmu_object_info_from_dnode(dn, doi); 2815 2816 dnode_rele(dn, FTAG); 2817 return (0); 2818 } 2819 2820 /* 2821 * As above, but faster; can be used when you have a held dbuf in hand. 2822 */ 2823 void 2824 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi) 2825 { 2826 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2827 2828 DB_DNODE_ENTER(db); 2829 dmu_object_info_from_dnode(DB_DNODE(db), doi); 2830 DB_DNODE_EXIT(db); 2831 } 2832 2833 /* 2834 * Faster still when you only care about the size. 2835 */ 2836 void 2837 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize, 2838 u_longlong_t *nblk512) 2839 { 2840 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2841 dnode_t *dn; 2842 2843 DB_DNODE_ENTER(db); 2844 dn = DB_DNODE(db); 2845 2846 *blksize = dn->dn_datablksz; 2847 /* add in number of slots used for the dnode itself */ 2848 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >> 2849 SPA_MINBLOCKSHIFT) + dn->dn_num_slots; 2850 DB_DNODE_EXIT(db); 2851 } 2852 2853 void 2854 dmu_object_dnsize_from_db(dmu_buf_t *db_fake, int *dnsize) 2855 { 2856 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2857 2858 DB_DNODE_ENTER(db); 2859 *dnsize = DB_DNODE(db)->dn_num_slots << DNODE_SHIFT; 2860 DB_DNODE_EXIT(db); 2861 } 2862 2863 void 2864 byteswap_uint64_array(void *vbuf, size_t size) 2865 { 2866 uint64_t *buf = vbuf; 2867 size_t count = size >> 3; 2868 int i; 2869 2870 ASSERT0((size & 7)); 2871 2872 for (i = 0; i < count; i++) 2873 buf[i] = BSWAP_64(buf[i]); 2874 } 2875 2876 void 2877 byteswap_uint32_array(void *vbuf, size_t size) 2878 { 2879 uint32_t *buf = vbuf; 2880 size_t count = size >> 2; 2881 int i; 2882 2883 ASSERT0((size & 3)); 2884 2885 for (i = 0; i < count; i++) 2886 buf[i] = BSWAP_32(buf[i]); 2887 } 2888 2889 void 2890 byteswap_uint16_array(void *vbuf, size_t size) 2891 { 2892 uint16_t *buf = vbuf; 2893 size_t count = size >> 1; 2894 int i; 2895 2896 ASSERT0((size & 1)); 2897 2898 for (i = 0; i < count; i++) 2899 buf[i] = BSWAP_16(buf[i]); 2900 } 2901 2902 void 2903 byteswap_uint8_array(void *vbuf, size_t size) 2904 { 2905 (void) vbuf, (void) size; 2906 } 2907 2908 void 2909 dmu_init(void) 2910 { 2911 abd_init(); 2912 zfs_dbgmsg_init(); 2913 sa_cache_init(); 2914 dmu_objset_init(); 2915 dnode_init(); 2916 zfetch_init(); 2917 dmu_tx_init(); 2918 l2arc_init(); 2919 arc_init(); 2920 dbuf_init(); 2921 } 2922 2923 void 2924 dmu_fini(void) 2925 { 2926 arc_fini(); /* arc depends on l2arc, so arc must go first */ 2927 l2arc_fini(); 2928 dmu_tx_fini(); 2929 zfetch_fini(); 2930 dbuf_fini(); 2931 dnode_fini(); 2932 dmu_objset_fini(); 2933 sa_cache_fini(); 2934 zfs_dbgmsg_fini(); 2935 abd_fini(); 2936 } 2937 2938 EXPORT_SYMBOL(dmu_bonus_hold); 2939 EXPORT_SYMBOL(dmu_bonus_hold_by_dnode); 2940 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus); 2941 EXPORT_SYMBOL(dmu_buf_rele_array); 2942 EXPORT_SYMBOL(dmu_prefetch); 2943 EXPORT_SYMBOL(dmu_prefetch_by_dnode); 2944 EXPORT_SYMBOL(dmu_prefetch_dnode); 2945 EXPORT_SYMBOL(dmu_free_range); 2946 EXPORT_SYMBOL(dmu_free_long_range); 2947 EXPORT_SYMBOL(dmu_free_long_object); 2948 EXPORT_SYMBOL(dmu_read); 2949 EXPORT_SYMBOL(dmu_read_by_dnode); 2950 EXPORT_SYMBOL(dmu_read_uio); 2951 EXPORT_SYMBOL(dmu_read_uio_dbuf); 2952 EXPORT_SYMBOL(dmu_read_uio_dnode); 2953 EXPORT_SYMBOL(dmu_write); 2954 EXPORT_SYMBOL(dmu_write_by_dnode); 2955 EXPORT_SYMBOL(dmu_write_uio); 2956 EXPORT_SYMBOL(dmu_write_uio_dbuf); 2957 EXPORT_SYMBOL(dmu_write_uio_dnode); 2958 EXPORT_SYMBOL(dmu_prealloc); 2959 EXPORT_SYMBOL(dmu_object_info); 2960 EXPORT_SYMBOL(dmu_object_info_from_dnode); 2961 EXPORT_SYMBOL(dmu_object_info_from_db); 2962 EXPORT_SYMBOL(dmu_object_size_from_db); 2963 EXPORT_SYMBOL(dmu_object_dnsize_from_db); 2964 EXPORT_SYMBOL(dmu_object_set_nlevels); 2965 EXPORT_SYMBOL(dmu_object_set_blocksize); 2966 EXPORT_SYMBOL(dmu_object_set_maxblkid); 2967 EXPORT_SYMBOL(dmu_object_set_checksum); 2968 EXPORT_SYMBOL(dmu_object_set_compress); 2969 EXPORT_SYMBOL(dmu_offset_next); 2970 EXPORT_SYMBOL(dmu_write_policy); 2971 EXPORT_SYMBOL(dmu_sync); 2972 EXPORT_SYMBOL(dmu_request_arcbuf); 2973 EXPORT_SYMBOL(dmu_return_arcbuf); 2974 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dnode); 2975 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dbuf); 2976 EXPORT_SYMBOL(dmu_buf_hold); 2977 EXPORT_SYMBOL(dmu_ot); 2978 2979 ZFS_MODULE_PARAM(zfs, zfs_, nopwrite_enabled, INT, ZMOD_RW, 2980 "Enable NOP writes"); 2981 2982 ZFS_MODULE_PARAM(zfs, zfs_, per_txg_dirty_frees_percent, UINT, ZMOD_RW, 2983 "Percentage of dirtied blocks from frees in one TXG"); 2984 2985 ZFS_MODULE_PARAM(zfs, zfs_, dmu_offset_next_sync, INT, ZMOD_RW, 2986 "Enable forcing txg sync to find holes"); 2987 2988 ZFS_MODULE_PARAM(zfs, , dmu_prefetch_max, UINT, ZMOD_RW, 2989 "Limit one prefetch call to this size"); 2990 2991 ZFS_MODULE_PARAM(zfs, , dmu_ddt_copies, UINT, ZMOD_RW, 2992 "Override copies= for dedup objects"); 2993