1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21 /* 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23 * Copyright (c) 2011, 2020 by Delphix. All rights reserved. 24 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved. 25 * Copyright (c) 2013, Joyent, Inc. All rights reserved. 26 * Copyright (c) 2016, Nexenta Systems, Inc. All rights reserved. 27 * Copyright (c) 2015 by Chunwei Chen. All rights reserved. 28 * Copyright (c) 2019 Datto Inc. 29 * Copyright (c) 2019, Klara Inc. 30 * Copyright (c) 2019, Allan Jude 31 */ 32 33 #include <sys/dmu.h> 34 #include <sys/dmu_impl.h> 35 #include <sys/dmu_tx.h> 36 #include <sys/dbuf.h> 37 #include <sys/dnode.h> 38 #include <sys/zfs_context.h> 39 #include <sys/dmu_objset.h> 40 #include <sys/dmu_traverse.h> 41 #include <sys/dsl_dataset.h> 42 #include <sys/dsl_dir.h> 43 #include <sys/dsl_pool.h> 44 #include <sys/dsl_synctask.h> 45 #include <sys/dsl_prop.h> 46 #include <sys/dmu_zfetch.h> 47 #include <sys/zfs_ioctl.h> 48 #include <sys/zap.h> 49 #include <sys/zio_checksum.h> 50 #include <sys/zio_compress.h> 51 #include <sys/sa.h> 52 #include <sys/zfeature.h> 53 #include <sys/abd.h> 54 #include <sys/trace_zfs.h> 55 #include <sys/zfs_racct.h> 56 #include <sys/zfs_rlock.h> 57 #ifdef _KERNEL 58 #include <sys/vmsystm.h> 59 #include <sys/zfs_znode.h> 60 #endif 61 62 /* 63 * Enable/disable nopwrite feature. 64 */ 65 int zfs_nopwrite_enabled = 1; 66 67 /* 68 * Tunable to control percentage of dirtied L1 blocks from frees allowed into 69 * one TXG. After this threshold is crossed, additional dirty blocks from frees 70 * will wait until the next TXG. 71 * A value of zero will disable this throttle. 72 */ 73 unsigned long zfs_per_txg_dirty_frees_percent = 5; 74 75 /* 76 * Enable/disable forcing txg sync when dirty checking for holes with lseek(). 77 * By default this is enabled to ensure accurate hole reporting, it can result 78 * in a significant performance penalty for lseek(SEEK_HOLE) heavy workloads. 79 * Disabling this option will result in holes never being reported in dirty 80 * files which is always safe. 81 */ 82 int zfs_dmu_offset_next_sync = 1; 83 84 /* 85 * Limit the amount we can prefetch with one call to this amount. This 86 * helps to limit the amount of memory that can be used by prefetching. 87 * Larger objects should be prefetched a bit at a time. 88 */ 89 int dmu_prefetch_max = 8 * SPA_MAXBLOCKSIZE; 90 91 const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = { 92 {DMU_BSWAP_UINT8, TRUE, FALSE, FALSE, "unallocated" }, 93 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "object directory" }, 94 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "object array" }, 95 {DMU_BSWAP_UINT8, TRUE, FALSE, FALSE, "packed nvlist" }, 96 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "packed nvlist size" }, 97 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "bpobj" }, 98 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "bpobj header" }, 99 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "SPA space map header" }, 100 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "SPA space map" }, 101 {DMU_BSWAP_UINT64, TRUE, FALSE, TRUE, "ZIL intent log" }, 102 {DMU_BSWAP_DNODE, TRUE, FALSE, TRUE, "DMU dnode" }, 103 {DMU_BSWAP_OBJSET, TRUE, TRUE, FALSE, "DMU objset" }, 104 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "DSL directory" }, 105 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL directory child map"}, 106 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL dataset snap map" }, 107 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL props" }, 108 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "DSL dataset" }, 109 {DMU_BSWAP_ZNODE, TRUE, FALSE, FALSE, "ZFS znode" }, 110 {DMU_BSWAP_OLDACL, TRUE, FALSE, TRUE, "ZFS V0 ACL" }, 111 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "ZFS plain file" }, 112 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS directory" }, 113 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "ZFS master node" }, 114 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS delete queue" }, 115 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "zvol object" }, 116 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "zvol prop" }, 117 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "other uint8[]" }, 118 {DMU_BSWAP_UINT64, FALSE, FALSE, TRUE, "other uint64[]" }, 119 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "other ZAP" }, 120 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "persistent error log" }, 121 {DMU_BSWAP_UINT8, TRUE, FALSE, FALSE, "SPA history" }, 122 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "SPA history offsets" }, 123 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "Pool properties" }, 124 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL permissions" }, 125 {DMU_BSWAP_ACL, TRUE, FALSE, TRUE, "ZFS ACL" }, 126 {DMU_BSWAP_UINT8, TRUE, FALSE, TRUE, "ZFS SYSACL" }, 127 {DMU_BSWAP_UINT8, TRUE, FALSE, TRUE, "FUID table" }, 128 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "FUID table size" }, 129 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL dataset next clones"}, 130 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "scan work queue" }, 131 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS user/group/project used" }, 132 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "ZFS user/group/project quota"}, 133 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "snapshot refcount tags"}, 134 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "DDT ZAP algorithm" }, 135 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "DDT statistics" }, 136 {DMU_BSWAP_UINT8, TRUE, FALSE, TRUE, "System attributes" }, 137 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "SA master node" }, 138 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "SA attr registration" }, 139 {DMU_BSWAP_ZAP, TRUE, FALSE, TRUE, "SA attr layouts" }, 140 {DMU_BSWAP_ZAP, TRUE, FALSE, FALSE, "scan translations" }, 141 {DMU_BSWAP_UINT8, FALSE, FALSE, TRUE, "deduplicated block" }, 142 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL deadlist map" }, 143 {DMU_BSWAP_UINT64, TRUE, TRUE, FALSE, "DSL deadlist map hdr" }, 144 {DMU_BSWAP_ZAP, TRUE, TRUE, FALSE, "DSL dir clones" }, 145 {DMU_BSWAP_UINT64, TRUE, FALSE, FALSE, "bpobj subobj" } 146 }; 147 148 const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = { 149 { byteswap_uint8_array, "uint8" }, 150 { byteswap_uint16_array, "uint16" }, 151 { byteswap_uint32_array, "uint32" }, 152 { byteswap_uint64_array, "uint64" }, 153 { zap_byteswap, "zap" }, 154 { dnode_buf_byteswap, "dnode" }, 155 { dmu_objset_byteswap, "objset" }, 156 { zfs_znode_byteswap, "znode" }, 157 { zfs_oldacl_byteswap, "oldacl" }, 158 { zfs_acl_byteswap, "acl" } 159 }; 160 161 static int 162 dmu_buf_hold_noread_by_dnode(dnode_t *dn, uint64_t offset, 163 void *tag, dmu_buf_t **dbp) 164 { 165 uint64_t blkid; 166 dmu_buf_impl_t *db; 167 168 rw_enter(&dn->dn_struct_rwlock, RW_READER); 169 blkid = dbuf_whichblock(dn, 0, offset); 170 db = dbuf_hold(dn, blkid, tag); 171 rw_exit(&dn->dn_struct_rwlock); 172 173 if (db == NULL) { 174 *dbp = NULL; 175 return (SET_ERROR(EIO)); 176 } 177 178 *dbp = &db->db; 179 return (0); 180 } 181 int 182 dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset, 183 void *tag, dmu_buf_t **dbp) 184 { 185 dnode_t *dn; 186 uint64_t blkid; 187 dmu_buf_impl_t *db; 188 int err; 189 190 err = dnode_hold(os, object, FTAG, &dn); 191 if (err) 192 return (err); 193 rw_enter(&dn->dn_struct_rwlock, RW_READER); 194 blkid = dbuf_whichblock(dn, 0, offset); 195 db = dbuf_hold(dn, blkid, tag); 196 rw_exit(&dn->dn_struct_rwlock); 197 dnode_rele(dn, FTAG); 198 199 if (db == NULL) { 200 *dbp = NULL; 201 return (SET_ERROR(EIO)); 202 } 203 204 *dbp = &db->db; 205 return (err); 206 } 207 208 int 209 dmu_buf_hold_by_dnode(dnode_t *dn, uint64_t offset, 210 void *tag, dmu_buf_t **dbp, int flags) 211 { 212 int err; 213 int db_flags = DB_RF_CANFAIL; 214 215 if (flags & DMU_READ_NO_PREFETCH) 216 db_flags |= DB_RF_NOPREFETCH; 217 if (flags & DMU_READ_NO_DECRYPT) 218 db_flags |= DB_RF_NO_DECRYPT; 219 220 err = dmu_buf_hold_noread_by_dnode(dn, offset, tag, dbp); 221 if (err == 0) { 222 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp); 223 err = dbuf_read(db, NULL, db_flags); 224 if (err != 0) { 225 dbuf_rele(db, tag); 226 *dbp = NULL; 227 } 228 } 229 230 return (err); 231 } 232 233 int 234 dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset, 235 void *tag, dmu_buf_t **dbp, int flags) 236 { 237 int err; 238 int db_flags = DB_RF_CANFAIL; 239 240 if (flags & DMU_READ_NO_PREFETCH) 241 db_flags |= DB_RF_NOPREFETCH; 242 if (flags & DMU_READ_NO_DECRYPT) 243 db_flags |= DB_RF_NO_DECRYPT; 244 245 err = dmu_buf_hold_noread(os, object, offset, tag, dbp); 246 if (err == 0) { 247 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp); 248 err = dbuf_read(db, NULL, db_flags); 249 if (err != 0) { 250 dbuf_rele(db, tag); 251 *dbp = NULL; 252 } 253 } 254 255 return (err); 256 } 257 258 int 259 dmu_bonus_max(void) 260 { 261 return (DN_OLD_MAX_BONUSLEN); 262 } 263 264 int 265 dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx) 266 { 267 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 268 dnode_t *dn; 269 int error; 270 271 DB_DNODE_ENTER(db); 272 dn = DB_DNODE(db); 273 274 if (dn->dn_bonus != db) { 275 error = SET_ERROR(EINVAL); 276 } else if (newsize < 0 || newsize > db_fake->db_size) { 277 error = SET_ERROR(EINVAL); 278 } else { 279 dnode_setbonuslen(dn, newsize, tx); 280 error = 0; 281 } 282 283 DB_DNODE_EXIT(db); 284 return (error); 285 } 286 287 int 288 dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx) 289 { 290 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 291 dnode_t *dn; 292 int error; 293 294 DB_DNODE_ENTER(db); 295 dn = DB_DNODE(db); 296 297 if (!DMU_OT_IS_VALID(type)) { 298 error = SET_ERROR(EINVAL); 299 } else if (dn->dn_bonus != db) { 300 error = SET_ERROR(EINVAL); 301 } else { 302 dnode_setbonus_type(dn, type, tx); 303 error = 0; 304 } 305 306 DB_DNODE_EXIT(db); 307 return (error); 308 } 309 310 dmu_object_type_t 311 dmu_get_bonustype(dmu_buf_t *db_fake) 312 { 313 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 314 dnode_t *dn; 315 dmu_object_type_t type; 316 317 DB_DNODE_ENTER(db); 318 dn = DB_DNODE(db); 319 type = dn->dn_bonustype; 320 DB_DNODE_EXIT(db); 321 322 return (type); 323 } 324 325 int 326 dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx) 327 { 328 dnode_t *dn; 329 int error; 330 331 error = dnode_hold(os, object, FTAG, &dn); 332 dbuf_rm_spill(dn, tx); 333 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 334 dnode_rm_spill(dn, tx); 335 rw_exit(&dn->dn_struct_rwlock); 336 dnode_rele(dn, FTAG); 337 return (error); 338 } 339 340 /* 341 * Lookup and hold the bonus buffer for the provided dnode. If the dnode 342 * has not yet been allocated a new bonus dbuf a will be allocated. 343 * Returns ENOENT, EIO, or 0. 344 */ 345 int dmu_bonus_hold_by_dnode(dnode_t *dn, void *tag, dmu_buf_t **dbp, 346 uint32_t flags) 347 { 348 dmu_buf_impl_t *db; 349 int error; 350 uint32_t db_flags = DB_RF_MUST_SUCCEED; 351 352 if (flags & DMU_READ_NO_PREFETCH) 353 db_flags |= DB_RF_NOPREFETCH; 354 if (flags & DMU_READ_NO_DECRYPT) 355 db_flags |= DB_RF_NO_DECRYPT; 356 357 rw_enter(&dn->dn_struct_rwlock, RW_READER); 358 if (dn->dn_bonus == NULL) { 359 rw_exit(&dn->dn_struct_rwlock); 360 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 361 if (dn->dn_bonus == NULL) 362 dbuf_create_bonus(dn); 363 } 364 db = dn->dn_bonus; 365 366 /* as long as the bonus buf is held, the dnode will be held */ 367 if (zfs_refcount_add(&db->db_holds, tag) == 1) { 368 VERIFY(dnode_add_ref(dn, db)); 369 atomic_inc_32(&dn->dn_dbufs_count); 370 } 371 372 /* 373 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's 374 * hold and incrementing the dbuf count to ensure that dnode_move() sees 375 * a dnode hold for every dbuf. 376 */ 377 rw_exit(&dn->dn_struct_rwlock); 378 379 error = dbuf_read(db, NULL, db_flags); 380 if (error) { 381 dnode_evict_bonus(dn); 382 dbuf_rele(db, tag); 383 *dbp = NULL; 384 return (error); 385 } 386 387 *dbp = &db->db; 388 return (0); 389 } 390 391 int 392 dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp) 393 { 394 dnode_t *dn; 395 int error; 396 397 error = dnode_hold(os, object, FTAG, &dn); 398 if (error) 399 return (error); 400 401 error = dmu_bonus_hold_by_dnode(dn, tag, dbp, DMU_READ_NO_PREFETCH); 402 dnode_rele(dn, FTAG); 403 404 return (error); 405 } 406 407 /* 408 * returns ENOENT, EIO, or 0. 409 * 410 * This interface will allocate a blank spill dbuf when a spill blk 411 * doesn't already exist on the dnode. 412 * 413 * if you only want to find an already existing spill db, then 414 * dmu_spill_hold_existing() should be used. 415 */ 416 int 417 dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, 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, 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, uint32_t flags, void *tag, 475 dmu_buf_t **dbp) 476 { 477 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus; 478 dnode_t *dn; 479 int err; 480 uint32_t db_flags = DB_RF_CANFAIL; 481 482 if (flags & DMU_READ_NO_DECRYPT) 483 db_flags |= DB_RF_NO_DECRYPT; 484 485 DB_DNODE_ENTER(db); 486 dn = DB_DNODE(db); 487 err = dmu_spill_hold_by_dnode(dn, db_flags, tag, dbp); 488 DB_DNODE_EXIT(db); 489 490 return (err); 491 } 492 493 /* 494 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces 495 * to take a held dnode rather than <os, object> -- the lookup is wasteful, 496 * and can induce severe lock contention when writing to several files 497 * whose dnodes are in the same block. 498 */ 499 int 500 dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length, 501 boolean_t read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags) 502 { 503 dmu_buf_t **dbp; 504 zstream_t *zs = NULL; 505 uint64_t blkid, nblks, i; 506 uint32_t dbuf_flags; 507 int err; 508 zio_t *zio = NULL; 509 boolean_t missed = B_FALSE; 510 511 ASSERT(length <= DMU_MAX_ACCESS); 512 513 /* 514 * Note: We directly notify the prefetch code of this read, so that 515 * we can tell it about the multi-block read. dbuf_read() only knows 516 * about the one block it is accessing. 517 */ 518 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT | 519 DB_RF_NOPREFETCH; 520 521 rw_enter(&dn->dn_struct_rwlock, RW_READER); 522 if (dn->dn_datablkshift) { 523 int blkshift = dn->dn_datablkshift; 524 nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) - 525 P2ALIGN(offset, 1ULL << blkshift)) >> blkshift; 526 } else { 527 if (offset + length > dn->dn_datablksz) { 528 zfs_panic_recover("zfs: accessing past end of object " 529 "%llx/%llx (size=%u access=%llu+%llu)", 530 (longlong_t)dn->dn_objset-> 531 os_dsl_dataset->ds_object, 532 (longlong_t)dn->dn_object, dn->dn_datablksz, 533 (longlong_t)offset, (longlong_t)length); 534 rw_exit(&dn->dn_struct_rwlock); 535 return (SET_ERROR(EIO)); 536 } 537 nblks = 1; 538 } 539 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP); 540 541 if (read) 542 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, 543 ZIO_FLAG_CANFAIL); 544 blkid = dbuf_whichblock(dn, 0, offset); 545 if ((flags & DMU_READ_NO_PREFETCH) == 0 && 546 DNODE_META_IS_CACHEABLE(dn) && length <= zfetch_array_rd_sz) { 547 /* 548 * Prepare the zfetch before initiating the demand reads, so 549 * that if multiple threads block on same indirect block, we 550 * base predictions on the original less racy request order. 551 */ 552 zs = dmu_zfetch_prepare(&dn->dn_zfetch, blkid, nblks, 553 read && DNODE_IS_CACHEABLE(dn), B_TRUE); 554 } 555 for (i = 0; i < nblks; i++) { 556 dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag); 557 if (db == NULL) { 558 if (zs) 559 dmu_zfetch_run(zs, missed, B_TRUE); 560 rw_exit(&dn->dn_struct_rwlock); 561 dmu_buf_rele_array(dbp, nblks, tag); 562 if (read) 563 zio_nowait(zio); 564 return (SET_ERROR(EIO)); 565 } 566 567 /* 568 * Initiate async demand data read. 569 * We check the db_state after calling dbuf_read() because 570 * (1) dbuf_read() may change the state to CACHED due to a 571 * hit in the ARC, and (2) on a cache miss, a child will 572 * have been added to "zio" but not yet completed, so the 573 * state will not yet be CACHED. 574 */ 575 if (read) { 576 (void) dbuf_read(db, zio, dbuf_flags); 577 if (db->db_state != DB_CACHED) 578 missed = B_TRUE; 579 } 580 dbp[i] = &db->db; 581 } 582 583 if (!read) 584 zfs_racct_write(length, nblks); 585 586 if (zs) 587 dmu_zfetch_run(zs, missed, B_TRUE); 588 rw_exit(&dn->dn_struct_rwlock); 589 590 if (read) { 591 /* wait for async read i/o */ 592 err = zio_wait(zio); 593 if (err) { 594 dmu_buf_rele_array(dbp, nblks, tag); 595 return (err); 596 } 597 598 /* wait for other io to complete */ 599 for (i = 0; i < nblks; i++) { 600 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i]; 601 mutex_enter(&db->db_mtx); 602 while (db->db_state == DB_READ || 603 db->db_state == DB_FILL) 604 cv_wait(&db->db_changed, &db->db_mtx); 605 if (db->db_state == DB_UNCACHED) 606 err = SET_ERROR(EIO); 607 mutex_exit(&db->db_mtx); 608 if (err) { 609 dmu_buf_rele_array(dbp, nblks, tag); 610 return (err); 611 } 612 } 613 } 614 615 *numbufsp = nblks; 616 *dbpp = dbp; 617 return (0); 618 } 619 620 int 621 dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset, 622 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp) 623 { 624 dnode_t *dn; 625 int err; 626 627 err = dnode_hold(os, object, FTAG, &dn); 628 if (err) 629 return (err); 630 631 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag, 632 numbufsp, dbpp, DMU_READ_PREFETCH); 633 634 dnode_rele(dn, FTAG); 635 636 return (err); 637 } 638 639 int 640 dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset, 641 uint64_t length, boolean_t read, void *tag, int *numbufsp, 642 dmu_buf_t ***dbpp) 643 { 644 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 645 dnode_t *dn; 646 int err; 647 648 DB_DNODE_ENTER(db); 649 dn = DB_DNODE(db); 650 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag, 651 numbufsp, dbpp, DMU_READ_PREFETCH); 652 DB_DNODE_EXIT(db); 653 654 return (err); 655 } 656 657 void 658 dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag) 659 { 660 int i; 661 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake; 662 663 if (numbufs == 0) 664 return; 665 666 for (i = 0; i < numbufs; i++) { 667 if (dbp[i]) 668 dbuf_rele(dbp[i], tag); 669 } 670 671 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs); 672 } 673 674 /* 675 * Issue prefetch i/os for the given blocks. If level is greater than 0, the 676 * indirect blocks prefetched will be those that point to the blocks containing 677 * the data starting at offset, and continuing to offset + len. 678 * 679 * Note that if the indirect blocks above the blocks being prefetched are not 680 * in cache, they will be asynchronously read in. 681 */ 682 void 683 dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset, 684 uint64_t len, zio_priority_t pri) 685 { 686 dnode_t *dn; 687 uint64_t blkid; 688 int nblks, err; 689 690 if (len == 0) { /* they're interested in the bonus buffer */ 691 dn = DMU_META_DNODE(os); 692 693 if (object == 0 || object >= DN_MAX_OBJECT) 694 return; 695 696 rw_enter(&dn->dn_struct_rwlock, RW_READER); 697 blkid = dbuf_whichblock(dn, level, 698 object * sizeof (dnode_phys_t)); 699 dbuf_prefetch(dn, level, blkid, pri, 0); 700 rw_exit(&dn->dn_struct_rwlock); 701 return; 702 } 703 704 /* 705 * See comment before the definition of dmu_prefetch_max. 706 */ 707 len = MIN(len, dmu_prefetch_max); 708 709 /* 710 * XXX - Note, if the dnode for the requested object is not 711 * already cached, we will do a *synchronous* read in the 712 * dnode_hold() call. The same is true for any indirects. 713 */ 714 err = dnode_hold(os, object, FTAG, &dn); 715 if (err != 0) 716 return; 717 718 /* 719 * offset + len - 1 is the last byte we want to prefetch for, and offset 720 * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the 721 * last block we want to prefetch, and dbuf_whichblock(dn, level, 722 * offset) is the first. Then the number we need to prefetch is the 723 * last - first + 1. 724 */ 725 rw_enter(&dn->dn_struct_rwlock, RW_READER); 726 if (level > 0 || dn->dn_datablkshift != 0) { 727 nblks = dbuf_whichblock(dn, level, offset + len - 1) - 728 dbuf_whichblock(dn, level, offset) + 1; 729 } else { 730 nblks = (offset < dn->dn_datablksz); 731 } 732 733 if (nblks != 0) { 734 blkid = dbuf_whichblock(dn, level, offset); 735 for (int i = 0; i < nblks; i++) 736 dbuf_prefetch(dn, level, blkid + i, pri, 0); 737 } 738 rw_exit(&dn->dn_struct_rwlock); 739 740 dnode_rele(dn, FTAG); 741 } 742 743 /* 744 * Get the next "chunk" of file data to free. We traverse the file from 745 * the end so that the file gets shorter over time (if we crashes in the 746 * middle, this will leave us in a better state). We find allocated file 747 * data by simply searching the allocated level 1 indirects. 748 * 749 * On input, *start should be the first offset that does not need to be 750 * freed (e.g. "offset + length"). On return, *start will be the first 751 * offset that should be freed and l1blks is set to the number of level 1 752 * indirect blocks found within the chunk. 753 */ 754 static int 755 get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum, uint64_t *l1blks) 756 { 757 uint64_t blks; 758 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1); 759 /* bytes of data covered by a level-1 indirect block */ 760 uint64_t iblkrange = (uint64_t)dn->dn_datablksz * 761 EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT); 762 763 ASSERT3U(minimum, <=, *start); 764 765 /* 766 * Check if we can free the entire range assuming that all of the 767 * L1 blocks in this range have data. If we can, we use this 768 * worst case value as an estimate so we can avoid having to look 769 * at the object's actual data. 770 */ 771 uint64_t total_l1blks = 772 (roundup(*start, iblkrange) - (minimum / iblkrange * iblkrange)) / 773 iblkrange; 774 if (total_l1blks <= maxblks) { 775 *l1blks = total_l1blks; 776 *start = minimum; 777 return (0); 778 } 779 ASSERT(ISP2(iblkrange)); 780 781 for (blks = 0; *start > minimum && blks < maxblks; blks++) { 782 int err; 783 784 /* 785 * dnode_next_offset(BACKWARDS) will find an allocated L1 786 * indirect block at or before the input offset. We must 787 * decrement *start so that it is at the end of the region 788 * to search. 789 */ 790 (*start)--; 791 792 err = dnode_next_offset(dn, 793 DNODE_FIND_BACKWARDS, start, 2, 1, 0); 794 795 /* if there are no indirect blocks before start, we are done */ 796 if (err == ESRCH) { 797 *start = minimum; 798 break; 799 } else if (err != 0) { 800 *l1blks = blks; 801 return (err); 802 } 803 804 /* set start to the beginning of this L1 indirect */ 805 *start = P2ALIGN(*start, iblkrange); 806 } 807 if (*start < minimum) 808 *start = minimum; 809 *l1blks = blks; 810 811 return (0); 812 } 813 814 /* 815 * If this objset is of type OST_ZFS return true if vfs's unmounted flag is set, 816 * otherwise return false. 817 * Used below in dmu_free_long_range_impl() to enable abort when unmounting 818 */ 819 /*ARGSUSED*/ 820 static boolean_t 821 dmu_objset_zfs_unmounting(objset_t *os) 822 { 823 #ifdef _KERNEL 824 if (dmu_objset_type(os) == DMU_OST_ZFS) 825 return (zfs_get_vfs_flag_unmounted(os)); 826 #endif 827 return (B_FALSE); 828 } 829 830 static int 831 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset, 832 uint64_t length) 833 { 834 uint64_t object_size; 835 int err; 836 uint64_t dirty_frees_threshold; 837 dsl_pool_t *dp = dmu_objset_pool(os); 838 839 if (dn == NULL) 840 return (SET_ERROR(EINVAL)); 841 842 object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz; 843 if (offset >= object_size) 844 return (0); 845 846 if (zfs_per_txg_dirty_frees_percent <= 100) 847 dirty_frees_threshold = 848 zfs_per_txg_dirty_frees_percent * zfs_dirty_data_max / 100; 849 else 850 dirty_frees_threshold = zfs_dirty_data_max / 20; 851 852 if (length == DMU_OBJECT_END || offset + length > object_size) 853 length = object_size - offset; 854 855 while (length != 0) { 856 uint64_t chunk_end, chunk_begin, chunk_len; 857 uint64_t l1blks; 858 dmu_tx_t *tx; 859 860 if (dmu_objset_zfs_unmounting(dn->dn_objset)) 861 return (SET_ERROR(EINTR)); 862 863 chunk_end = chunk_begin = offset + length; 864 865 /* move chunk_begin backwards to the beginning of this chunk */ 866 err = get_next_chunk(dn, &chunk_begin, offset, &l1blks); 867 if (err) 868 return (err); 869 ASSERT3U(chunk_begin, >=, offset); 870 ASSERT3U(chunk_begin, <=, chunk_end); 871 872 chunk_len = chunk_end - chunk_begin; 873 874 tx = dmu_tx_create(os); 875 dmu_tx_hold_free(tx, dn->dn_object, chunk_begin, chunk_len); 876 877 /* 878 * Mark this transaction as typically resulting in a net 879 * reduction in space used. 880 */ 881 dmu_tx_mark_netfree(tx); 882 err = dmu_tx_assign(tx, TXG_WAIT); 883 if (err) { 884 dmu_tx_abort(tx); 885 return (err); 886 } 887 888 uint64_t txg = dmu_tx_get_txg(tx); 889 890 mutex_enter(&dp->dp_lock); 891 uint64_t long_free_dirty = 892 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK]; 893 mutex_exit(&dp->dp_lock); 894 895 /* 896 * To avoid filling up a TXG with just frees, wait for 897 * the next TXG to open before freeing more chunks if 898 * we have reached the threshold of frees. 899 */ 900 if (dirty_frees_threshold != 0 && 901 long_free_dirty >= dirty_frees_threshold) { 902 DMU_TX_STAT_BUMP(dmu_tx_dirty_frees_delay); 903 dmu_tx_commit(tx); 904 txg_wait_open(dp, 0, B_TRUE); 905 continue; 906 } 907 908 /* 909 * In order to prevent unnecessary write throttling, for each 910 * TXG, we track the cumulative size of L1 blocks being dirtied 911 * in dnode_free_range() below. We compare this number to a 912 * tunable threshold, past which we prevent new L1 dirty freeing 913 * blocks from being added into the open TXG. See 914 * dmu_free_long_range_impl() for details. The threshold 915 * prevents write throttle activation due to dirty freeing L1 916 * blocks taking up a large percentage of zfs_dirty_data_max. 917 */ 918 mutex_enter(&dp->dp_lock); 919 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] += 920 l1blks << dn->dn_indblkshift; 921 mutex_exit(&dp->dp_lock); 922 DTRACE_PROBE3(free__long__range, 923 uint64_t, long_free_dirty, uint64_t, chunk_len, 924 uint64_t, txg); 925 dnode_free_range(dn, chunk_begin, chunk_len, tx); 926 927 dmu_tx_commit(tx); 928 929 length -= chunk_len; 930 } 931 return (0); 932 } 933 934 int 935 dmu_free_long_range(objset_t *os, uint64_t object, 936 uint64_t offset, uint64_t length) 937 { 938 dnode_t *dn; 939 int err; 940 941 err = dnode_hold(os, object, FTAG, &dn); 942 if (err != 0) 943 return (err); 944 err = dmu_free_long_range_impl(os, dn, offset, length); 945 946 /* 947 * It is important to zero out the maxblkid when freeing the entire 948 * file, so that (a) subsequent calls to dmu_free_long_range_impl() 949 * will take the fast path, and (b) dnode_reallocate() can verify 950 * that the entire file has been freed. 951 */ 952 if (err == 0 && offset == 0 && length == DMU_OBJECT_END) 953 dn->dn_maxblkid = 0; 954 955 dnode_rele(dn, FTAG); 956 return (err); 957 } 958 959 int 960 dmu_free_long_object(objset_t *os, uint64_t object) 961 { 962 dmu_tx_t *tx; 963 int err; 964 965 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END); 966 if (err != 0) 967 return (err); 968 969 tx = dmu_tx_create(os); 970 dmu_tx_hold_bonus(tx, object); 971 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END); 972 dmu_tx_mark_netfree(tx); 973 err = dmu_tx_assign(tx, TXG_WAIT); 974 if (err == 0) { 975 err = dmu_object_free(os, object, tx); 976 dmu_tx_commit(tx); 977 } else { 978 dmu_tx_abort(tx); 979 } 980 981 return (err); 982 } 983 984 int 985 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset, 986 uint64_t size, dmu_tx_t *tx) 987 { 988 dnode_t *dn; 989 int err = dnode_hold(os, object, FTAG, &dn); 990 if (err) 991 return (err); 992 ASSERT(offset < UINT64_MAX); 993 ASSERT(size == DMU_OBJECT_END || size <= UINT64_MAX - offset); 994 dnode_free_range(dn, offset, size, tx); 995 dnode_rele(dn, FTAG); 996 return (0); 997 } 998 999 static int 1000 dmu_read_impl(dnode_t *dn, uint64_t offset, uint64_t size, 1001 void *buf, uint32_t flags) 1002 { 1003 dmu_buf_t **dbp; 1004 int numbufs, err = 0; 1005 1006 /* 1007 * Deal with odd block sizes, where there can't be data past the first 1008 * block. If we ever do the tail block optimization, we will need to 1009 * handle that here as well. 1010 */ 1011 if (dn->dn_maxblkid == 0) { 1012 uint64_t newsz = offset > dn->dn_datablksz ? 0 : 1013 MIN(size, dn->dn_datablksz - offset); 1014 bzero((char *)buf + newsz, size - newsz); 1015 size = newsz; 1016 } 1017 1018 while (size > 0) { 1019 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2); 1020 int i; 1021 1022 /* 1023 * NB: we could do this block-at-a-time, but it's nice 1024 * to be reading in parallel. 1025 */ 1026 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen, 1027 TRUE, FTAG, &numbufs, &dbp, flags); 1028 if (err) 1029 break; 1030 1031 for (i = 0; i < numbufs; i++) { 1032 uint64_t tocpy; 1033 int64_t bufoff; 1034 dmu_buf_t *db = dbp[i]; 1035 1036 ASSERT(size > 0); 1037 1038 bufoff = offset - db->db_offset; 1039 tocpy = MIN(db->db_size - bufoff, size); 1040 1041 (void) memcpy(buf, (char *)db->db_data + bufoff, tocpy); 1042 1043 offset += tocpy; 1044 size -= tocpy; 1045 buf = (char *)buf + tocpy; 1046 } 1047 dmu_buf_rele_array(dbp, numbufs, FTAG); 1048 } 1049 return (err); 1050 } 1051 1052 int 1053 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1054 void *buf, uint32_t flags) 1055 { 1056 dnode_t *dn; 1057 int err; 1058 1059 err = dnode_hold(os, object, FTAG, &dn); 1060 if (err != 0) 1061 return (err); 1062 1063 err = dmu_read_impl(dn, offset, size, buf, flags); 1064 dnode_rele(dn, FTAG); 1065 return (err); 1066 } 1067 1068 int 1069 dmu_read_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, void *buf, 1070 uint32_t flags) 1071 { 1072 return (dmu_read_impl(dn, offset, size, buf, flags)); 1073 } 1074 1075 static void 1076 dmu_write_impl(dmu_buf_t **dbp, int numbufs, uint64_t offset, uint64_t size, 1077 const void *buf, dmu_tx_t *tx) 1078 { 1079 int i; 1080 1081 for (i = 0; i < numbufs; i++) { 1082 uint64_t tocpy; 1083 int64_t bufoff; 1084 dmu_buf_t *db = dbp[i]; 1085 1086 ASSERT(size > 0); 1087 1088 bufoff = offset - db->db_offset; 1089 tocpy = MIN(db->db_size - bufoff, size); 1090 1091 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 1092 1093 if (tocpy == db->db_size) 1094 dmu_buf_will_fill(db, tx); 1095 else 1096 dmu_buf_will_dirty(db, tx); 1097 1098 (void) memcpy((char *)db->db_data + bufoff, buf, tocpy); 1099 1100 if (tocpy == db->db_size) 1101 dmu_buf_fill_done(db, tx); 1102 1103 offset += tocpy; 1104 size -= tocpy; 1105 buf = (char *)buf + tocpy; 1106 } 1107 } 1108 1109 void 1110 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1111 const void *buf, dmu_tx_t *tx) 1112 { 1113 dmu_buf_t **dbp; 1114 int numbufs; 1115 1116 if (size == 0) 1117 return; 1118 1119 VERIFY0(dmu_buf_hold_array(os, object, offset, size, 1120 FALSE, FTAG, &numbufs, &dbp)); 1121 dmu_write_impl(dbp, numbufs, offset, size, buf, tx); 1122 dmu_buf_rele_array(dbp, numbufs, FTAG); 1123 } 1124 1125 /* 1126 * Note: Lustre is an external consumer of this interface. 1127 */ 1128 void 1129 dmu_write_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, 1130 const void *buf, dmu_tx_t *tx) 1131 { 1132 dmu_buf_t **dbp; 1133 int numbufs; 1134 1135 if (size == 0) 1136 return; 1137 1138 VERIFY0(dmu_buf_hold_array_by_dnode(dn, offset, size, 1139 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH)); 1140 dmu_write_impl(dbp, numbufs, offset, size, buf, tx); 1141 dmu_buf_rele_array(dbp, numbufs, FTAG); 1142 } 1143 1144 void 1145 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1146 dmu_tx_t *tx) 1147 { 1148 dmu_buf_t **dbp; 1149 int numbufs, i; 1150 1151 if (size == 0) 1152 return; 1153 1154 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size, 1155 FALSE, FTAG, &numbufs, &dbp)); 1156 1157 for (i = 0; i < numbufs; i++) { 1158 dmu_buf_t *db = dbp[i]; 1159 1160 dmu_buf_will_not_fill(db, tx); 1161 } 1162 dmu_buf_rele_array(dbp, numbufs, FTAG); 1163 } 1164 1165 void 1166 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset, 1167 void *data, uint8_t etype, uint8_t comp, int uncompressed_size, 1168 int compressed_size, int byteorder, dmu_tx_t *tx) 1169 { 1170 dmu_buf_t *db; 1171 1172 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES); 1173 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS); 1174 VERIFY0(dmu_buf_hold_noread(os, object, offset, 1175 FTAG, &db)); 1176 1177 dmu_buf_write_embedded(db, 1178 data, (bp_embedded_type_t)etype, (enum zio_compress)comp, 1179 uncompressed_size, compressed_size, byteorder, tx); 1180 1181 dmu_buf_rele(db, FTAG); 1182 } 1183 1184 void 1185 dmu_redact(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1186 dmu_tx_t *tx) 1187 { 1188 int numbufs, i; 1189 dmu_buf_t **dbp; 1190 1191 VERIFY0(dmu_buf_hold_array(os, object, offset, size, FALSE, FTAG, 1192 &numbufs, &dbp)); 1193 for (i = 0; i < numbufs; i++) 1194 dmu_buf_redact(dbp[i], tx); 1195 dmu_buf_rele_array(dbp, numbufs, FTAG); 1196 } 1197 1198 #ifdef _KERNEL 1199 int 1200 dmu_read_uio_dnode(dnode_t *dn, zfs_uio_t *uio, uint64_t size) 1201 { 1202 dmu_buf_t **dbp; 1203 int numbufs, i, err; 1204 1205 /* 1206 * NB: we could do this block-at-a-time, but it's nice 1207 * to be reading in parallel. 1208 */ 1209 err = dmu_buf_hold_array_by_dnode(dn, zfs_uio_offset(uio), size, 1210 TRUE, FTAG, &numbufs, &dbp, 0); 1211 if (err) 1212 return (err); 1213 1214 for (i = 0; i < numbufs; i++) { 1215 uint64_t tocpy; 1216 int64_t bufoff; 1217 dmu_buf_t *db = dbp[i]; 1218 1219 ASSERT(size > 0); 1220 1221 bufoff = zfs_uio_offset(uio) - db->db_offset; 1222 tocpy = MIN(db->db_size - bufoff, size); 1223 1224 err = zfs_uio_fault_move((char *)db->db_data + bufoff, tocpy, 1225 UIO_READ, uio); 1226 1227 if (err) 1228 break; 1229 1230 size -= tocpy; 1231 } 1232 dmu_buf_rele_array(dbp, numbufs, FTAG); 1233 1234 return (err); 1235 } 1236 1237 /* 1238 * Read 'size' bytes into the uio buffer. 1239 * From object zdb->db_object. 1240 * Starting at zfs_uio_offset(uio). 1241 * 1242 * If the caller already has a dbuf in the target object 1243 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(), 1244 * because we don't have to find the dnode_t for the object. 1245 */ 1246 int 1247 dmu_read_uio_dbuf(dmu_buf_t *zdb, zfs_uio_t *uio, uint64_t size) 1248 { 1249 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb; 1250 dnode_t *dn; 1251 int err; 1252 1253 if (size == 0) 1254 return (0); 1255 1256 DB_DNODE_ENTER(db); 1257 dn = DB_DNODE(db); 1258 err = dmu_read_uio_dnode(dn, uio, size); 1259 DB_DNODE_EXIT(db); 1260 1261 return (err); 1262 } 1263 1264 /* 1265 * Read 'size' bytes into the uio buffer. 1266 * From the specified object 1267 * Starting at offset zfs_uio_offset(uio). 1268 */ 1269 int 1270 dmu_read_uio(objset_t *os, uint64_t object, zfs_uio_t *uio, uint64_t size) 1271 { 1272 dnode_t *dn; 1273 int err; 1274 1275 if (size == 0) 1276 return (0); 1277 1278 err = dnode_hold(os, object, FTAG, &dn); 1279 if (err) 1280 return (err); 1281 1282 err = dmu_read_uio_dnode(dn, uio, size); 1283 1284 dnode_rele(dn, FTAG); 1285 1286 return (err); 1287 } 1288 1289 int 1290 dmu_write_uio_dnode(dnode_t *dn, zfs_uio_t *uio, uint64_t size, dmu_tx_t *tx) 1291 { 1292 dmu_buf_t **dbp; 1293 int numbufs; 1294 int err = 0; 1295 int i; 1296 1297 err = dmu_buf_hold_array_by_dnode(dn, zfs_uio_offset(uio), size, 1298 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH); 1299 if (err) 1300 return (err); 1301 1302 for (i = 0; i < numbufs; i++) { 1303 uint64_t tocpy; 1304 int64_t bufoff; 1305 dmu_buf_t *db = dbp[i]; 1306 1307 ASSERT(size > 0); 1308 1309 bufoff = zfs_uio_offset(uio) - db->db_offset; 1310 tocpy = MIN(db->db_size - bufoff, size); 1311 1312 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 1313 1314 if (tocpy == db->db_size) 1315 dmu_buf_will_fill(db, tx); 1316 else 1317 dmu_buf_will_dirty(db, tx); 1318 1319 /* 1320 * XXX zfs_uiomove could block forever (eg.nfs-backed 1321 * pages). There needs to be a uiolockdown() function 1322 * to lock the pages in memory, so that zfs_uiomove won't 1323 * block. 1324 */ 1325 err = zfs_uio_fault_move((char *)db->db_data + bufoff, 1326 tocpy, UIO_WRITE, uio); 1327 1328 if (tocpy == db->db_size) 1329 dmu_buf_fill_done(db, tx); 1330 1331 if (err) 1332 break; 1333 1334 size -= tocpy; 1335 } 1336 1337 dmu_buf_rele_array(dbp, numbufs, FTAG); 1338 return (err); 1339 } 1340 1341 /* 1342 * Write 'size' bytes from the uio buffer. 1343 * To object zdb->db_object. 1344 * Starting at offset zfs_uio_offset(uio). 1345 * 1346 * If the caller already has a dbuf in the target object 1347 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(), 1348 * because we don't have to find the dnode_t for the object. 1349 */ 1350 int 1351 dmu_write_uio_dbuf(dmu_buf_t *zdb, zfs_uio_t *uio, uint64_t size, 1352 dmu_tx_t *tx) 1353 { 1354 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb; 1355 dnode_t *dn; 1356 int err; 1357 1358 if (size == 0) 1359 return (0); 1360 1361 DB_DNODE_ENTER(db); 1362 dn = DB_DNODE(db); 1363 err = dmu_write_uio_dnode(dn, uio, size, tx); 1364 DB_DNODE_EXIT(db); 1365 1366 return (err); 1367 } 1368 1369 /* 1370 * Write 'size' bytes from the uio buffer. 1371 * To the specified object. 1372 * Starting at offset zfs_uio_offset(uio). 1373 */ 1374 int 1375 dmu_write_uio(objset_t *os, uint64_t object, zfs_uio_t *uio, uint64_t size, 1376 dmu_tx_t *tx) 1377 { 1378 dnode_t *dn; 1379 int err; 1380 1381 if (size == 0) 1382 return (0); 1383 1384 err = dnode_hold(os, object, FTAG, &dn); 1385 if (err) 1386 return (err); 1387 1388 err = dmu_write_uio_dnode(dn, uio, size, tx); 1389 1390 dnode_rele(dn, FTAG); 1391 1392 return (err); 1393 } 1394 #endif /* _KERNEL */ 1395 1396 /* 1397 * Allocate a loaned anonymous arc buffer. 1398 */ 1399 arc_buf_t * 1400 dmu_request_arcbuf(dmu_buf_t *handle, int size) 1401 { 1402 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle; 1403 1404 return (arc_loan_buf(db->db_objset->os_spa, B_FALSE, size)); 1405 } 1406 1407 /* 1408 * Free a loaned arc buffer. 1409 */ 1410 void 1411 dmu_return_arcbuf(arc_buf_t *buf) 1412 { 1413 arc_return_buf(buf, FTAG); 1414 arc_buf_destroy(buf, FTAG); 1415 } 1416 1417 /* 1418 * A "lightweight" write is faster than a regular write (e.g. 1419 * dmu_write_by_dnode() or dmu_assign_arcbuf_by_dnode()), because it avoids the 1420 * CPU cost of creating a dmu_buf_impl_t and arc_buf_[hdr_]_t. However, the 1421 * data can not be read or overwritten until the transaction's txg has been 1422 * synced. This makes it appropriate for workloads that are known to be 1423 * (temporarily) write-only, like "zfs receive". 1424 * 1425 * A single block is written, starting at the specified offset in bytes. If 1426 * the call is successful, it returns 0 and the provided abd has been 1427 * consumed (the caller should not free it). 1428 */ 1429 int 1430 dmu_lightweight_write_by_dnode(dnode_t *dn, uint64_t offset, abd_t *abd, 1431 const zio_prop_t *zp, enum zio_flag flags, dmu_tx_t *tx) 1432 { 1433 dbuf_dirty_record_t *dr = 1434 dbuf_dirty_lightweight(dn, dbuf_whichblock(dn, 0, offset), tx); 1435 if (dr == NULL) 1436 return (SET_ERROR(EIO)); 1437 dr->dt.dll.dr_abd = abd; 1438 dr->dt.dll.dr_props = *zp; 1439 dr->dt.dll.dr_flags = flags; 1440 return (0); 1441 } 1442 1443 /* 1444 * When possible directly assign passed loaned arc buffer to a dbuf. 1445 * If this is not possible copy the contents of passed arc buf via 1446 * dmu_write(). 1447 */ 1448 int 1449 dmu_assign_arcbuf_by_dnode(dnode_t *dn, uint64_t offset, arc_buf_t *buf, 1450 dmu_tx_t *tx) 1451 { 1452 dmu_buf_impl_t *db; 1453 objset_t *os = dn->dn_objset; 1454 uint64_t object = dn->dn_object; 1455 uint32_t blksz = (uint32_t)arc_buf_lsize(buf); 1456 uint64_t blkid; 1457 1458 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1459 blkid = dbuf_whichblock(dn, 0, offset); 1460 db = dbuf_hold(dn, blkid, FTAG); 1461 if (db == NULL) 1462 return (SET_ERROR(EIO)); 1463 rw_exit(&dn->dn_struct_rwlock); 1464 1465 /* 1466 * We can only assign if the offset is aligned and the arc buf is the 1467 * same size as the dbuf. 1468 */ 1469 if (offset == db->db.db_offset && blksz == db->db.db_size) { 1470 zfs_racct_write(blksz, 1); 1471 dbuf_assign_arcbuf(db, buf, tx); 1472 dbuf_rele(db, FTAG); 1473 } else { 1474 /* compressed bufs must always be assignable to their dbuf */ 1475 ASSERT3U(arc_get_compression(buf), ==, ZIO_COMPRESS_OFF); 1476 ASSERT(!(buf->b_flags & ARC_BUF_FLAG_COMPRESSED)); 1477 1478 dbuf_rele(db, FTAG); 1479 dmu_write(os, object, offset, blksz, buf->b_data, tx); 1480 dmu_return_arcbuf(buf); 1481 } 1482 1483 return (0); 1484 } 1485 1486 int 1487 dmu_assign_arcbuf_by_dbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf, 1488 dmu_tx_t *tx) 1489 { 1490 int err; 1491 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle; 1492 1493 DB_DNODE_ENTER(dbuf); 1494 err = dmu_assign_arcbuf_by_dnode(DB_DNODE(dbuf), offset, buf, tx); 1495 DB_DNODE_EXIT(dbuf); 1496 1497 return (err); 1498 } 1499 1500 typedef struct { 1501 dbuf_dirty_record_t *dsa_dr; 1502 dmu_sync_cb_t *dsa_done; 1503 zgd_t *dsa_zgd; 1504 dmu_tx_t *dsa_tx; 1505 } dmu_sync_arg_t; 1506 1507 /* ARGSUSED */ 1508 static void 1509 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg) 1510 { 1511 dmu_sync_arg_t *dsa = varg; 1512 dmu_buf_t *db = dsa->dsa_zgd->zgd_db; 1513 blkptr_t *bp = zio->io_bp; 1514 1515 if (zio->io_error == 0) { 1516 if (BP_IS_HOLE(bp)) { 1517 /* 1518 * A block of zeros may compress to a hole, but the 1519 * block size still needs to be known for replay. 1520 */ 1521 BP_SET_LSIZE(bp, db->db_size); 1522 } else if (!BP_IS_EMBEDDED(bp)) { 1523 ASSERT(BP_GET_LEVEL(bp) == 0); 1524 BP_SET_FILL(bp, 1); 1525 } 1526 } 1527 } 1528 1529 static void 1530 dmu_sync_late_arrival_ready(zio_t *zio) 1531 { 1532 dmu_sync_ready(zio, NULL, zio->io_private); 1533 } 1534 1535 /* ARGSUSED */ 1536 static void 1537 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg) 1538 { 1539 dmu_sync_arg_t *dsa = varg; 1540 dbuf_dirty_record_t *dr = dsa->dsa_dr; 1541 dmu_buf_impl_t *db = dr->dr_dbuf; 1542 zgd_t *zgd = dsa->dsa_zgd; 1543 1544 /* 1545 * Record the vdev(s) backing this blkptr so they can be flushed after 1546 * the writes for the lwb have completed. 1547 */ 1548 if (zio->io_error == 0) { 1549 zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp); 1550 } 1551 1552 mutex_enter(&db->db_mtx); 1553 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC); 1554 if (zio->io_error == 0) { 1555 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE); 1556 if (dr->dt.dl.dr_nopwrite) { 1557 blkptr_t *bp = zio->io_bp; 1558 blkptr_t *bp_orig = &zio->io_bp_orig; 1559 uint8_t chksum = BP_GET_CHECKSUM(bp_orig); 1560 1561 ASSERT(BP_EQUAL(bp, bp_orig)); 1562 VERIFY(BP_EQUAL(bp, db->db_blkptr)); 1563 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF); 1564 VERIFY(zio_checksum_table[chksum].ci_flags & 1565 ZCHECKSUM_FLAG_NOPWRITE); 1566 } 1567 dr->dt.dl.dr_overridden_by = *zio->io_bp; 1568 dr->dt.dl.dr_override_state = DR_OVERRIDDEN; 1569 dr->dt.dl.dr_copies = zio->io_prop.zp_copies; 1570 1571 /* 1572 * Old style holes are filled with all zeros, whereas 1573 * new-style holes maintain their lsize, type, level, 1574 * and birth time (see zio_write_compress). While we 1575 * need to reset the BP_SET_LSIZE() call that happened 1576 * in dmu_sync_ready for old style holes, we do *not* 1577 * want to wipe out the information contained in new 1578 * style holes. Thus, only zero out the block pointer if 1579 * it's an old style hole. 1580 */ 1581 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) && 1582 dr->dt.dl.dr_overridden_by.blk_birth == 0) 1583 BP_ZERO(&dr->dt.dl.dr_overridden_by); 1584 } else { 1585 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 1586 } 1587 cv_broadcast(&db->db_changed); 1588 mutex_exit(&db->db_mtx); 1589 1590 dsa->dsa_done(dsa->dsa_zgd, zio->io_error); 1591 1592 kmem_free(dsa, sizeof (*dsa)); 1593 } 1594 1595 static void 1596 dmu_sync_late_arrival_done(zio_t *zio) 1597 { 1598 blkptr_t *bp = zio->io_bp; 1599 dmu_sync_arg_t *dsa = zio->io_private; 1600 zgd_t *zgd = dsa->dsa_zgd; 1601 1602 if (zio->io_error == 0) { 1603 /* 1604 * Record the vdev(s) backing this blkptr so they can be 1605 * flushed after the writes for the lwb have completed. 1606 */ 1607 zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp); 1608 1609 if (!BP_IS_HOLE(bp)) { 1610 blkptr_t *bp_orig __maybe_unused = &zio->io_bp_orig; 1611 ASSERT(!(zio->io_flags & ZIO_FLAG_NOPWRITE)); 1612 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig)); 1613 ASSERT(zio->io_bp->blk_birth == zio->io_txg); 1614 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa)); 1615 zio_free(zio->io_spa, zio->io_txg, zio->io_bp); 1616 } 1617 } 1618 1619 dmu_tx_commit(dsa->dsa_tx); 1620 1621 dsa->dsa_done(dsa->dsa_zgd, zio->io_error); 1622 1623 abd_free(zio->io_abd); 1624 kmem_free(dsa, sizeof (*dsa)); 1625 } 1626 1627 static int 1628 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd, 1629 zio_prop_t *zp, zbookmark_phys_t *zb) 1630 { 1631 dmu_sync_arg_t *dsa; 1632 dmu_tx_t *tx; 1633 1634 tx = dmu_tx_create(os); 1635 dmu_tx_hold_space(tx, zgd->zgd_db->db_size); 1636 if (dmu_tx_assign(tx, TXG_WAIT) != 0) { 1637 dmu_tx_abort(tx); 1638 /* Make zl_get_data do txg_waited_synced() */ 1639 return (SET_ERROR(EIO)); 1640 } 1641 1642 /* 1643 * In order to prevent the zgd's lwb from being free'd prior to 1644 * dmu_sync_late_arrival_done() being called, we have to ensure 1645 * the lwb's "max txg" takes this tx's txg into account. 1646 */ 1647 zil_lwb_add_txg(zgd->zgd_lwb, dmu_tx_get_txg(tx)); 1648 1649 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP); 1650 dsa->dsa_dr = NULL; 1651 dsa->dsa_done = done; 1652 dsa->dsa_zgd = zgd; 1653 dsa->dsa_tx = tx; 1654 1655 /* 1656 * Since we are currently syncing this txg, it's nontrivial to 1657 * determine what BP to nopwrite against, so we disable nopwrite. 1658 * 1659 * When syncing, the db_blkptr is initially the BP of the previous 1660 * txg. We can not nopwrite against it because it will be changed 1661 * (this is similar to the non-late-arrival case where the dbuf is 1662 * dirty in a future txg). 1663 * 1664 * Then dbuf_write_ready() sets bp_blkptr to the location we will write. 1665 * We can not nopwrite against it because although the BP will not 1666 * (typically) be changed, the data has not yet been persisted to this 1667 * location. 1668 * 1669 * Finally, when dbuf_write_done() is called, it is theoretically 1670 * possible to always nopwrite, because the data that was written in 1671 * this txg is the same data that we are trying to write. However we 1672 * would need to check that this dbuf is not dirty in any future 1673 * txg's (as we do in the normal dmu_sync() path). For simplicity, we 1674 * don't nopwrite in this case. 1675 */ 1676 zp->zp_nopwrite = B_FALSE; 1677 1678 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp, 1679 abd_get_from_buf(zgd->zgd_db->db_data, zgd->zgd_db->db_size), 1680 zgd->zgd_db->db_size, zgd->zgd_db->db_size, zp, 1681 dmu_sync_late_arrival_ready, NULL, NULL, dmu_sync_late_arrival_done, 1682 dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb)); 1683 1684 return (0); 1685 } 1686 1687 /* 1688 * Intent log support: sync the block associated with db to disk. 1689 * N.B. and XXX: the caller is responsible for making sure that the 1690 * data isn't changing while dmu_sync() is writing it. 1691 * 1692 * Return values: 1693 * 1694 * EEXIST: this txg has already been synced, so there's nothing to do. 1695 * The caller should not log the write. 1696 * 1697 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do. 1698 * The caller should not log the write. 1699 * 1700 * EALREADY: this block is already in the process of being synced. 1701 * The caller should track its progress (somehow). 1702 * 1703 * EIO: could not do the I/O. 1704 * The caller should do a txg_wait_synced(). 1705 * 1706 * 0: the I/O has been initiated. 1707 * The caller should log this blkptr in the done callback. 1708 * It is possible that the I/O will fail, in which case 1709 * the error will be reported to the done callback and 1710 * propagated to pio from zio_done(). 1711 */ 1712 int 1713 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd) 1714 { 1715 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db; 1716 objset_t *os = db->db_objset; 1717 dsl_dataset_t *ds = os->os_dsl_dataset; 1718 dbuf_dirty_record_t *dr, *dr_next; 1719 dmu_sync_arg_t *dsa; 1720 zbookmark_phys_t zb; 1721 zio_prop_t zp; 1722 dnode_t *dn; 1723 1724 ASSERT(pio != NULL); 1725 ASSERT(txg != 0); 1726 1727 SET_BOOKMARK(&zb, ds->ds_object, 1728 db->db.db_object, db->db_level, db->db_blkid); 1729 1730 DB_DNODE_ENTER(db); 1731 dn = DB_DNODE(db); 1732 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp); 1733 DB_DNODE_EXIT(db); 1734 1735 /* 1736 * If we're frozen (running ziltest), we always need to generate a bp. 1737 */ 1738 if (txg > spa_freeze_txg(os->os_spa)) 1739 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); 1740 1741 /* 1742 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf() 1743 * and us. If we determine that this txg is not yet syncing, 1744 * but it begins to sync a moment later, that's OK because the 1745 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx. 1746 */ 1747 mutex_enter(&db->db_mtx); 1748 1749 if (txg <= spa_last_synced_txg(os->os_spa)) { 1750 /* 1751 * This txg has already synced. There's nothing to do. 1752 */ 1753 mutex_exit(&db->db_mtx); 1754 return (SET_ERROR(EEXIST)); 1755 } 1756 1757 if (txg <= spa_syncing_txg(os->os_spa)) { 1758 /* 1759 * This txg is currently syncing, so we can't mess with 1760 * the dirty record anymore; just write a new log block. 1761 */ 1762 mutex_exit(&db->db_mtx); 1763 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); 1764 } 1765 1766 dr = dbuf_find_dirty_eq(db, txg); 1767 1768 if (dr == NULL) { 1769 /* 1770 * There's no dr for this dbuf, so it must have been freed. 1771 * There's no need to log writes to freed blocks, so we're done. 1772 */ 1773 mutex_exit(&db->db_mtx); 1774 return (SET_ERROR(ENOENT)); 1775 } 1776 1777 dr_next = list_next(&db->db_dirty_records, dr); 1778 ASSERT(dr_next == NULL || dr_next->dr_txg < txg); 1779 1780 if (db->db_blkptr != NULL) { 1781 /* 1782 * We need to fill in zgd_bp with the current blkptr so that 1783 * the nopwrite code can check if we're writing the same 1784 * data that's already on disk. We can only nopwrite if we 1785 * are sure that after making the copy, db_blkptr will not 1786 * change until our i/o completes. We ensure this by 1787 * holding the db_mtx, and only allowing nopwrite if the 1788 * block is not already dirty (see below). This is verified 1789 * by dmu_sync_done(), which VERIFYs that the db_blkptr has 1790 * not changed. 1791 */ 1792 *zgd->zgd_bp = *db->db_blkptr; 1793 } 1794 1795 /* 1796 * Assume the on-disk data is X, the current syncing data (in 1797 * txg - 1) is Y, and the current in-memory data is Z (currently 1798 * in dmu_sync). 1799 * 1800 * We usually want to perform a nopwrite if X and Z are the 1801 * same. However, if Y is different (i.e. the BP is going to 1802 * change before this write takes effect), then a nopwrite will 1803 * be incorrect - we would override with X, which could have 1804 * been freed when Y was written. 1805 * 1806 * (Note that this is not a concern when we are nop-writing from 1807 * syncing context, because X and Y must be identical, because 1808 * all previous txgs have been synced.) 1809 * 1810 * Therefore, we disable nopwrite if the current BP could change 1811 * before this TXG. There are two ways it could change: by 1812 * being dirty (dr_next is non-NULL), or by being freed 1813 * (dnode_block_freed()). This behavior is verified by 1814 * zio_done(), which VERIFYs that the override BP is identical 1815 * to the on-disk BP. 1816 */ 1817 DB_DNODE_ENTER(db); 1818 dn = DB_DNODE(db); 1819 if (dr_next != NULL || dnode_block_freed(dn, db->db_blkid)) 1820 zp.zp_nopwrite = B_FALSE; 1821 DB_DNODE_EXIT(db); 1822 1823 ASSERT(dr->dr_txg == txg); 1824 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC || 1825 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { 1826 /* 1827 * We have already issued a sync write for this buffer, 1828 * or this buffer has already been synced. It could not 1829 * have been dirtied since, or we would have cleared the state. 1830 */ 1831 mutex_exit(&db->db_mtx); 1832 return (SET_ERROR(EALREADY)); 1833 } 1834 1835 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); 1836 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC; 1837 mutex_exit(&db->db_mtx); 1838 1839 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP); 1840 dsa->dsa_dr = dr; 1841 dsa->dsa_done = done; 1842 dsa->dsa_zgd = zgd; 1843 dsa->dsa_tx = NULL; 1844 1845 zio_nowait(arc_write(pio, os->os_spa, txg, 1846 zgd->zgd_bp, dr->dt.dl.dr_data, dbuf_is_l2cacheable(db), 1847 &zp, dmu_sync_ready, NULL, NULL, dmu_sync_done, dsa, 1848 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb)); 1849 1850 return (0); 1851 } 1852 1853 int 1854 dmu_object_set_nlevels(objset_t *os, uint64_t object, int nlevels, dmu_tx_t *tx) 1855 { 1856 dnode_t *dn; 1857 int err; 1858 1859 err = dnode_hold(os, object, FTAG, &dn); 1860 if (err) 1861 return (err); 1862 err = dnode_set_nlevels(dn, nlevels, tx); 1863 dnode_rele(dn, FTAG); 1864 return (err); 1865 } 1866 1867 int 1868 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs, 1869 dmu_tx_t *tx) 1870 { 1871 dnode_t *dn; 1872 int err; 1873 1874 err = dnode_hold(os, object, FTAG, &dn); 1875 if (err) 1876 return (err); 1877 err = dnode_set_blksz(dn, size, ibs, tx); 1878 dnode_rele(dn, FTAG); 1879 return (err); 1880 } 1881 1882 int 1883 dmu_object_set_maxblkid(objset_t *os, uint64_t object, uint64_t maxblkid, 1884 dmu_tx_t *tx) 1885 { 1886 dnode_t *dn; 1887 int err; 1888 1889 err = dnode_hold(os, object, FTAG, &dn); 1890 if (err) 1891 return (err); 1892 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 1893 dnode_new_blkid(dn, maxblkid, tx, B_FALSE, B_TRUE); 1894 rw_exit(&dn->dn_struct_rwlock); 1895 dnode_rele(dn, FTAG); 1896 return (0); 1897 } 1898 1899 void 1900 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum, 1901 dmu_tx_t *tx) 1902 { 1903 dnode_t *dn; 1904 1905 /* 1906 * Send streams include each object's checksum function. This 1907 * check ensures that the receiving system can understand the 1908 * checksum function transmitted. 1909 */ 1910 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS); 1911 1912 VERIFY0(dnode_hold(os, object, FTAG, &dn)); 1913 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS); 1914 dn->dn_checksum = checksum; 1915 dnode_setdirty(dn, tx); 1916 dnode_rele(dn, FTAG); 1917 } 1918 1919 void 1920 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress, 1921 dmu_tx_t *tx) 1922 { 1923 dnode_t *dn; 1924 1925 /* 1926 * Send streams include each object's compression function. This 1927 * check ensures that the receiving system can understand the 1928 * compression function transmitted. 1929 */ 1930 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS); 1931 1932 VERIFY0(dnode_hold(os, object, FTAG, &dn)); 1933 dn->dn_compress = compress; 1934 dnode_setdirty(dn, tx); 1935 dnode_rele(dn, FTAG); 1936 } 1937 1938 /* 1939 * When the "redundant_metadata" property is set to "most", only indirect 1940 * blocks of this level and higher will have an additional ditto block. 1941 */ 1942 int zfs_redundant_metadata_most_ditto_level = 2; 1943 1944 void 1945 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp) 1946 { 1947 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET; 1948 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) || 1949 (wp & WP_SPILL)); 1950 enum zio_checksum checksum = os->os_checksum; 1951 enum zio_compress compress = os->os_compress; 1952 uint8_t complevel = os->os_complevel; 1953 enum zio_checksum dedup_checksum = os->os_dedup_checksum; 1954 boolean_t dedup = B_FALSE; 1955 boolean_t nopwrite = B_FALSE; 1956 boolean_t dedup_verify = os->os_dedup_verify; 1957 boolean_t encrypt = B_FALSE; 1958 int copies = os->os_copies; 1959 1960 /* 1961 * We maintain different write policies for each of the following 1962 * types of data: 1963 * 1. metadata 1964 * 2. preallocated blocks (i.e. level-0 blocks of a dump device) 1965 * 3. all other level 0 blocks 1966 */ 1967 if (ismd) { 1968 /* 1969 * XXX -- we should design a compression algorithm 1970 * that specializes in arrays of bps. 1971 */ 1972 compress = zio_compress_select(os->os_spa, 1973 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON); 1974 1975 /* 1976 * Metadata always gets checksummed. If the data 1977 * checksum is multi-bit correctable, and it's not a 1978 * ZBT-style checksum, then it's suitable for metadata 1979 * as well. Otherwise, the metadata checksum defaults 1980 * to fletcher4. 1981 */ 1982 if (!(zio_checksum_table[checksum].ci_flags & 1983 ZCHECKSUM_FLAG_METADATA) || 1984 (zio_checksum_table[checksum].ci_flags & 1985 ZCHECKSUM_FLAG_EMBEDDED)) 1986 checksum = ZIO_CHECKSUM_FLETCHER_4; 1987 1988 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL || 1989 (os->os_redundant_metadata == 1990 ZFS_REDUNDANT_METADATA_MOST && 1991 (level >= zfs_redundant_metadata_most_ditto_level || 1992 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL)))) 1993 copies++; 1994 } else if (wp & WP_NOFILL) { 1995 ASSERT(level == 0); 1996 1997 /* 1998 * If we're writing preallocated blocks, we aren't actually 1999 * writing them so don't set any policy properties. These 2000 * blocks are currently only used by an external subsystem 2001 * outside of zfs (i.e. dump) and not written by the zio 2002 * pipeline. 2003 */ 2004 compress = ZIO_COMPRESS_OFF; 2005 checksum = ZIO_CHECKSUM_OFF; 2006 } else { 2007 compress = zio_compress_select(os->os_spa, dn->dn_compress, 2008 compress); 2009 complevel = zio_complevel_select(os->os_spa, compress, 2010 complevel, complevel); 2011 2012 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ? 2013 zio_checksum_select(dn->dn_checksum, checksum) : 2014 dedup_checksum; 2015 2016 /* 2017 * Determine dedup setting. If we are in dmu_sync(), 2018 * we won't actually dedup now because that's all 2019 * done in syncing context; but we do want to use the 2020 * dedup checksum. If the checksum is not strong 2021 * enough to ensure unique signatures, force 2022 * dedup_verify. 2023 */ 2024 if (dedup_checksum != ZIO_CHECKSUM_OFF) { 2025 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE; 2026 if (!(zio_checksum_table[checksum].ci_flags & 2027 ZCHECKSUM_FLAG_DEDUP)) 2028 dedup_verify = B_TRUE; 2029 } 2030 2031 /* 2032 * Enable nopwrite if we have secure enough checksum 2033 * algorithm (see comment in zio_nop_write) and 2034 * compression is enabled. We don't enable nopwrite if 2035 * dedup is enabled as the two features are mutually 2036 * exclusive. 2037 */ 2038 nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags & 2039 ZCHECKSUM_FLAG_NOPWRITE) && 2040 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled); 2041 } 2042 2043 /* 2044 * All objects in an encrypted objset are protected from modification 2045 * via a MAC. Encrypted objects store their IV and salt in the last DVA 2046 * in the bp, so we cannot use all copies. Encrypted objects are also 2047 * not subject to nopwrite since writing the same data will still 2048 * result in a new ciphertext. Only encrypted blocks can be dedup'd 2049 * to avoid ambiguity in the dedup code since the DDT does not store 2050 * object types. 2051 */ 2052 if (os->os_encrypted && (wp & WP_NOFILL) == 0) { 2053 encrypt = B_TRUE; 2054 2055 if (DMU_OT_IS_ENCRYPTED(type)) { 2056 copies = MIN(copies, SPA_DVAS_PER_BP - 1); 2057 nopwrite = B_FALSE; 2058 } else { 2059 dedup = B_FALSE; 2060 } 2061 2062 if (level <= 0 && 2063 (type == DMU_OT_DNODE || type == DMU_OT_OBJSET)) { 2064 compress = ZIO_COMPRESS_EMPTY; 2065 } 2066 } 2067 2068 zp->zp_compress = compress; 2069 zp->zp_complevel = complevel; 2070 zp->zp_checksum = checksum; 2071 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type; 2072 zp->zp_level = level; 2073 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa)); 2074 zp->zp_dedup = dedup; 2075 zp->zp_dedup_verify = dedup && dedup_verify; 2076 zp->zp_nopwrite = nopwrite; 2077 zp->zp_encrypt = encrypt; 2078 zp->zp_byteorder = ZFS_HOST_BYTEORDER; 2079 bzero(zp->zp_salt, ZIO_DATA_SALT_LEN); 2080 bzero(zp->zp_iv, ZIO_DATA_IV_LEN); 2081 bzero(zp->zp_mac, ZIO_DATA_MAC_LEN); 2082 zp->zp_zpl_smallblk = DMU_OT_IS_FILE(zp->zp_type) ? 2083 os->os_zpl_special_smallblock : 0; 2084 2085 ASSERT3U(zp->zp_compress, !=, ZIO_COMPRESS_INHERIT); 2086 } 2087 2088 /* 2089 * This function is only called from zfs_holey_common() for zpl_llseek() 2090 * in order to determine the location of holes. In order to accurately 2091 * report holes all dirty data must be synced to disk. This causes extremely 2092 * poor performance when seeking for holes in a dirty file. As a compromise, 2093 * only provide hole data when the dnode is clean. When a dnode is dirty 2094 * report the dnode as having no holes which is always a safe thing to do. 2095 */ 2096 int 2097 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off) 2098 { 2099 dnode_t *dn; 2100 int err; 2101 2102 restart: 2103 err = dnode_hold(os, object, FTAG, &dn); 2104 if (err) 2105 return (err); 2106 2107 rw_enter(&dn->dn_struct_rwlock, RW_READER); 2108 2109 if (dnode_is_dirty(dn)) { 2110 /* 2111 * If the zfs_dmu_offset_next_sync module option is enabled 2112 * then strict hole reporting has been requested. Dirty 2113 * dnodes must be synced to disk to accurately report all 2114 * holes. When disabled dirty dnodes are reported to not 2115 * have any holes which is always safe. 2116 * 2117 * When called by zfs_holey_common() the zp->z_rangelock 2118 * is held to prevent zfs_write() and mmap writeback from 2119 * re-dirtying the dnode after txg_wait_synced(). 2120 */ 2121 if (zfs_dmu_offset_next_sync) { 2122 rw_exit(&dn->dn_struct_rwlock); 2123 dnode_rele(dn, FTAG); 2124 txg_wait_synced(dmu_objset_pool(os), 0); 2125 goto restart; 2126 } 2127 2128 err = SET_ERROR(EBUSY); 2129 } else { 2130 err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK | 2131 (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0); 2132 } 2133 2134 rw_exit(&dn->dn_struct_rwlock); 2135 dnode_rele(dn, FTAG); 2136 2137 return (err); 2138 } 2139 2140 void 2141 __dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi) 2142 { 2143 dnode_phys_t *dnp = dn->dn_phys; 2144 2145 doi->doi_data_block_size = dn->dn_datablksz; 2146 doi->doi_metadata_block_size = dn->dn_indblkshift ? 2147 1ULL << dn->dn_indblkshift : 0; 2148 doi->doi_type = dn->dn_type; 2149 doi->doi_bonus_type = dn->dn_bonustype; 2150 doi->doi_bonus_size = dn->dn_bonuslen; 2151 doi->doi_dnodesize = dn->dn_num_slots << DNODE_SHIFT; 2152 doi->doi_indirection = dn->dn_nlevels; 2153 doi->doi_checksum = dn->dn_checksum; 2154 doi->doi_compress = dn->dn_compress; 2155 doi->doi_nblkptr = dn->dn_nblkptr; 2156 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9; 2157 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz; 2158 doi->doi_fill_count = 0; 2159 for (int i = 0; i < dnp->dn_nblkptr; i++) 2160 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]); 2161 } 2162 2163 void 2164 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi) 2165 { 2166 rw_enter(&dn->dn_struct_rwlock, RW_READER); 2167 mutex_enter(&dn->dn_mtx); 2168 2169 __dmu_object_info_from_dnode(dn, doi); 2170 2171 mutex_exit(&dn->dn_mtx); 2172 rw_exit(&dn->dn_struct_rwlock); 2173 } 2174 2175 /* 2176 * Get information on a DMU object. 2177 * If doi is NULL, just indicates whether the object exists. 2178 */ 2179 int 2180 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi) 2181 { 2182 dnode_t *dn; 2183 int err = dnode_hold(os, object, FTAG, &dn); 2184 2185 if (err) 2186 return (err); 2187 2188 if (doi != NULL) 2189 dmu_object_info_from_dnode(dn, doi); 2190 2191 dnode_rele(dn, FTAG); 2192 return (0); 2193 } 2194 2195 /* 2196 * As above, but faster; can be used when you have a held dbuf in hand. 2197 */ 2198 void 2199 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi) 2200 { 2201 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2202 2203 DB_DNODE_ENTER(db); 2204 dmu_object_info_from_dnode(DB_DNODE(db), doi); 2205 DB_DNODE_EXIT(db); 2206 } 2207 2208 /* 2209 * Faster still when you only care about the size. 2210 */ 2211 void 2212 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize, 2213 u_longlong_t *nblk512) 2214 { 2215 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2216 dnode_t *dn; 2217 2218 DB_DNODE_ENTER(db); 2219 dn = DB_DNODE(db); 2220 2221 *blksize = dn->dn_datablksz; 2222 /* add in number of slots used for the dnode itself */ 2223 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >> 2224 SPA_MINBLOCKSHIFT) + dn->dn_num_slots; 2225 DB_DNODE_EXIT(db); 2226 } 2227 2228 void 2229 dmu_object_dnsize_from_db(dmu_buf_t *db_fake, int *dnsize) 2230 { 2231 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2232 dnode_t *dn; 2233 2234 DB_DNODE_ENTER(db); 2235 dn = DB_DNODE(db); 2236 *dnsize = dn->dn_num_slots << DNODE_SHIFT; 2237 DB_DNODE_EXIT(db); 2238 } 2239 2240 void 2241 byteswap_uint64_array(void *vbuf, size_t size) 2242 { 2243 uint64_t *buf = vbuf; 2244 size_t count = size >> 3; 2245 int i; 2246 2247 ASSERT((size & 7) == 0); 2248 2249 for (i = 0; i < count; i++) 2250 buf[i] = BSWAP_64(buf[i]); 2251 } 2252 2253 void 2254 byteswap_uint32_array(void *vbuf, size_t size) 2255 { 2256 uint32_t *buf = vbuf; 2257 size_t count = size >> 2; 2258 int i; 2259 2260 ASSERT((size & 3) == 0); 2261 2262 for (i = 0; i < count; i++) 2263 buf[i] = BSWAP_32(buf[i]); 2264 } 2265 2266 void 2267 byteswap_uint16_array(void *vbuf, size_t size) 2268 { 2269 uint16_t *buf = vbuf; 2270 size_t count = size >> 1; 2271 int i; 2272 2273 ASSERT((size & 1) == 0); 2274 2275 for (i = 0; i < count; i++) 2276 buf[i] = BSWAP_16(buf[i]); 2277 } 2278 2279 /* ARGSUSED */ 2280 void 2281 byteswap_uint8_array(void *vbuf, size_t size) 2282 { 2283 } 2284 2285 void 2286 dmu_init(void) 2287 { 2288 abd_init(); 2289 zfs_dbgmsg_init(); 2290 sa_cache_init(); 2291 dmu_objset_init(); 2292 dnode_init(); 2293 zfetch_init(); 2294 dmu_tx_init(); 2295 l2arc_init(); 2296 arc_init(); 2297 dbuf_init(); 2298 } 2299 2300 void 2301 dmu_fini(void) 2302 { 2303 arc_fini(); /* arc depends on l2arc, so arc must go first */ 2304 l2arc_fini(); 2305 dmu_tx_fini(); 2306 zfetch_fini(); 2307 dbuf_fini(); 2308 dnode_fini(); 2309 dmu_objset_fini(); 2310 sa_cache_fini(); 2311 zfs_dbgmsg_fini(); 2312 abd_fini(); 2313 } 2314 2315 EXPORT_SYMBOL(dmu_bonus_hold); 2316 EXPORT_SYMBOL(dmu_bonus_hold_by_dnode); 2317 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus); 2318 EXPORT_SYMBOL(dmu_buf_rele_array); 2319 EXPORT_SYMBOL(dmu_prefetch); 2320 EXPORT_SYMBOL(dmu_free_range); 2321 EXPORT_SYMBOL(dmu_free_long_range); 2322 EXPORT_SYMBOL(dmu_free_long_object); 2323 EXPORT_SYMBOL(dmu_read); 2324 EXPORT_SYMBOL(dmu_read_by_dnode); 2325 EXPORT_SYMBOL(dmu_write); 2326 EXPORT_SYMBOL(dmu_write_by_dnode); 2327 EXPORT_SYMBOL(dmu_prealloc); 2328 EXPORT_SYMBOL(dmu_object_info); 2329 EXPORT_SYMBOL(dmu_object_info_from_dnode); 2330 EXPORT_SYMBOL(dmu_object_info_from_db); 2331 EXPORT_SYMBOL(dmu_object_size_from_db); 2332 EXPORT_SYMBOL(dmu_object_dnsize_from_db); 2333 EXPORT_SYMBOL(dmu_object_set_nlevels); 2334 EXPORT_SYMBOL(dmu_object_set_blocksize); 2335 EXPORT_SYMBOL(dmu_object_set_maxblkid); 2336 EXPORT_SYMBOL(dmu_object_set_checksum); 2337 EXPORT_SYMBOL(dmu_object_set_compress); 2338 EXPORT_SYMBOL(dmu_offset_next); 2339 EXPORT_SYMBOL(dmu_write_policy); 2340 EXPORT_SYMBOL(dmu_sync); 2341 EXPORT_SYMBOL(dmu_request_arcbuf); 2342 EXPORT_SYMBOL(dmu_return_arcbuf); 2343 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dnode); 2344 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dbuf); 2345 EXPORT_SYMBOL(dmu_buf_hold); 2346 EXPORT_SYMBOL(dmu_ot); 2347 2348 /* BEGIN CSTYLED */ 2349 ZFS_MODULE_PARAM(zfs, zfs_, nopwrite_enabled, INT, ZMOD_RW, 2350 "Enable NOP writes"); 2351 2352 ZFS_MODULE_PARAM(zfs, zfs_, per_txg_dirty_frees_percent, ULONG, ZMOD_RW, 2353 "Percentage of dirtied blocks from frees in one TXG"); 2354 2355 ZFS_MODULE_PARAM(zfs, zfs_, dmu_offset_next_sync, INT, ZMOD_RW, 2356 "Enable forcing txg sync to find holes"); 2357 2358 ZFS_MODULE_PARAM(zfs, , dmu_prefetch_max, INT, ZMOD_RW, 2359 "Limit one prefetch call to this size"); 2360 /* END CSTYLED */ 2361