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 static 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 static 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 static 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 static 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 static boolean_t 820 dmu_objset_zfs_unmounting(objset_t *os) 821 { 822 #ifdef _KERNEL 823 if (dmu_objset_type(os) == DMU_OST_ZFS) 824 return (zfs_get_vfs_flag_unmounted(os)); 825 #else 826 (void) os; 827 #endif 828 return (B_FALSE); 829 } 830 831 static int 832 dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset, 833 uint64_t length) 834 { 835 uint64_t object_size; 836 int err; 837 uint64_t dirty_frees_threshold; 838 dsl_pool_t *dp = dmu_objset_pool(os); 839 840 if (dn == NULL) 841 return (SET_ERROR(EINVAL)); 842 843 object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz; 844 if (offset >= object_size) 845 return (0); 846 847 if (zfs_per_txg_dirty_frees_percent <= 100) 848 dirty_frees_threshold = 849 zfs_per_txg_dirty_frees_percent * zfs_dirty_data_max / 100; 850 else 851 dirty_frees_threshold = zfs_dirty_data_max / 20; 852 853 if (length == DMU_OBJECT_END || offset + length > object_size) 854 length = object_size - offset; 855 856 while (length != 0) { 857 uint64_t chunk_end, chunk_begin, chunk_len; 858 uint64_t l1blks; 859 dmu_tx_t *tx; 860 861 if (dmu_objset_zfs_unmounting(dn->dn_objset)) 862 return (SET_ERROR(EINTR)); 863 864 chunk_end = chunk_begin = offset + length; 865 866 /* move chunk_begin backwards to the beginning of this chunk */ 867 err = get_next_chunk(dn, &chunk_begin, offset, &l1blks); 868 if (err) 869 return (err); 870 ASSERT3U(chunk_begin, >=, offset); 871 ASSERT3U(chunk_begin, <=, chunk_end); 872 873 chunk_len = chunk_end - chunk_begin; 874 875 tx = dmu_tx_create(os); 876 dmu_tx_hold_free(tx, dn->dn_object, chunk_begin, chunk_len); 877 878 /* 879 * Mark this transaction as typically resulting in a net 880 * reduction in space used. 881 */ 882 dmu_tx_mark_netfree(tx); 883 err = dmu_tx_assign(tx, TXG_WAIT); 884 if (err) { 885 dmu_tx_abort(tx); 886 return (err); 887 } 888 889 uint64_t txg = dmu_tx_get_txg(tx); 890 891 mutex_enter(&dp->dp_lock); 892 uint64_t long_free_dirty = 893 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK]; 894 mutex_exit(&dp->dp_lock); 895 896 /* 897 * To avoid filling up a TXG with just frees, wait for 898 * the next TXG to open before freeing more chunks if 899 * we have reached the threshold of frees. 900 */ 901 if (dirty_frees_threshold != 0 && 902 long_free_dirty >= dirty_frees_threshold) { 903 DMU_TX_STAT_BUMP(dmu_tx_dirty_frees_delay); 904 dmu_tx_commit(tx); 905 txg_wait_open(dp, 0, B_TRUE); 906 continue; 907 } 908 909 /* 910 * In order to prevent unnecessary write throttling, for each 911 * TXG, we track the cumulative size of L1 blocks being dirtied 912 * in dnode_free_range() below. We compare this number to a 913 * tunable threshold, past which we prevent new L1 dirty freeing 914 * blocks from being added into the open TXG. See 915 * dmu_free_long_range_impl() for details. The threshold 916 * prevents write throttle activation due to dirty freeing L1 917 * blocks taking up a large percentage of zfs_dirty_data_max. 918 */ 919 mutex_enter(&dp->dp_lock); 920 dp->dp_long_free_dirty_pertxg[txg & TXG_MASK] += 921 l1blks << dn->dn_indblkshift; 922 mutex_exit(&dp->dp_lock); 923 DTRACE_PROBE3(free__long__range, 924 uint64_t, long_free_dirty, uint64_t, chunk_len, 925 uint64_t, txg); 926 dnode_free_range(dn, chunk_begin, chunk_len, tx); 927 928 dmu_tx_commit(tx); 929 930 length -= chunk_len; 931 } 932 return (0); 933 } 934 935 int 936 dmu_free_long_range(objset_t *os, uint64_t object, 937 uint64_t offset, uint64_t length) 938 { 939 dnode_t *dn; 940 int err; 941 942 err = dnode_hold(os, object, FTAG, &dn); 943 if (err != 0) 944 return (err); 945 err = dmu_free_long_range_impl(os, dn, offset, length); 946 947 /* 948 * It is important to zero out the maxblkid when freeing the entire 949 * file, so that (a) subsequent calls to dmu_free_long_range_impl() 950 * will take the fast path, and (b) dnode_reallocate() can verify 951 * that the entire file has been freed. 952 */ 953 if (err == 0 && offset == 0 && length == DMU_OBJECT_END) 954 dn->dn_maxblkid = 0; 955 956 dnode_rele(dn, FTAG); 957 return (err); 958 } 959 960 int 961 dmu_free_long_object(objset_t *os, uint64_t object) 962 { 963 dmu_tx_t *tx; 964 int err; 965 966 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END); 967 if (err != 0) 968 return (err); 969 970 tx = dmu_tx_create(os); 971 dmu_tx_hold_bonus(tx, object); 972 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END); 973 dmu_tx_mark_netfree(tx); 974 err = dmu_tx_assign(tx, TXG_WAIT); 975 if (err == 0) { 976 err = dmu_object_free(os, object, tx); 977 dmu_tx_commit(tx); 978 } else { 979 dmu_tx_abort(tx); 980 } 981 982 return (err); 983 } 984 985 int 986 dmu_free_range(objset_t *os, uint64_t object, uint64_t offset, 987 uint64_t size, dmu_tx_t *tx) 988 { 989 dnode_t *dn; 990 int err = dnode_hold(os, object, FTAG, &dn); 991 if (err) 992 return (err); 993 ASSERT(offset < UINT64_MAX); 994 ASSERT(size == DMU_OBJECT_END || size <= UINT64_MAX - offset); 995 dnode_free_range(dn, offset, size, tx); 996 dnode_rele(dn, FTAG); 997 return (0); 998 } 999 1000 static int 1001 dmu_read_impl(dnode_t *dn, uint64_t offset, uint64_t size, 1002 void *buf, uint32_t flags) 1003 { 1004 dmu_buf_t **dbp; 1005 int numbufs, err = 0; 1006 1007 /* 1008 * Deal with odd block sizes, where there can't be data past the first 1009 * block. If we ever do the tail block optimization, we will need to 1010 * handle that here as well. 1011 */ 1012 if (dn->dn_maxblkid == 0) { 1013 uint64_t newsz = offset > dn->dn_datablksz ? 0 : 1014 MIN(size, dn->dn_datablksz - offset); 1015 memset((char *)buf + newsz, 0, size - newsz); 1016 size = newsz; 1017 } 1018 1019 while (size > 0) { 1020 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2); 1021 int i; 1022 1023 /* 1024 * NB: we could do this block-at-a-time, but it's nice 1025 * to be reading in parallel. 1026 */ 1027 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen, 1028 TRUE, FTAG, &numbufs, &dbp, flags); 1029 if (err) 1030 break; 1031 1032 for (i = 0; i < numbufs; i++) { 1033 uint64_t tocpy; 1034 int64_t bufoff; 1035 dmu_buf_t *db = dbp[i]; 1036 1037 ASSERT(size > 0); 1038 1039 bufoff = offset - db->db_offset; 1040 tocpy = MIN(db->db_size - bufoff, size); 1041 1042 (void) memcpy(buf, (char *)db->db_data + bufoff, tocpy); 1043 1044 offset += tocpy; 1045 size -= tocpy; 1046 buf = (char *)buf + tocpy; 1047 } 1048 dmu_buf_rele_array(dbp, numbufs, FTAG); 1049 } 1050 return (err); 1051 } 1052 1053 int 1054 dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1055 void *buf, uint32_t flags) 1056 { 1057 dnode_t *dn; 1058 int err; 1059 1060 err = dnode_hold(os, object, FTAG, &dn); 1061 if (err != 0) 1062 return (err); 1063 1064 err = dmu_read_impl(dn, offset, size, buf, flags); 1065 dnode_rele(dn, FTAG); 1066 return (err); 1067 } 1068 1069 int 1070 dmu_read_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, void *buf, 1071 uint32_t flags) 1072 { 1073 return (dmu_read_impl(dn, offset, size, buf, flags)); 1074 } 1075 1076 static void 1077 dmu_write_impl(dmu_buf_t **dbp, int numbufs, uint64_t offset, uint64_t size, 1078 const void *buf, dmu_tx_t *tx) 1079 { 1080 int i; 1081 1082 for (i = 0; i < numbufs; i++) { 1083 uint64_t tocpy; 1084 int64_t bufoff; 1085 dmu_buf_t *db = dbp[i]; 1086 1087 ASSERT(size > 0); 1088 1089 bufoff = offset - db->db_offset; 1090 tocpy = MIN(db->db_size - bufoff, size); 1091 1092 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 1093 1094 if (tocpy == db->db_size) 1095 dmu_buf_will_fill(db, tx); 1096 else 1097 dmu_buf_will_dirty(db, tx); 1098 1099 (void) memcpy((char *)db->db_data + bufoff, buf, tocpy); 1100 1101 if (tocpy == db->db_size) 1102 dmu_buf_fill_done(db, tx); 1103 1104 offset += tocpy; 1105 size -= tocpy; 1106 buf = (char *)buf + tocpy; 1107 } 1108 } 1109 1110 void 1111 dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1112 const void *buf, dmu_tx_t *tx) 1113 { 1114 dmu_buf_t **dbp; 1115 int numbufs; 1116 1117 if (size == 0) 1118 return; 1119 1120 VERIFY0(dmu_buf_hold_array(os, object, offset, size, 1121 FALSE, FTAG, &numbufs, &dbp)); 1122 dmu_write_impl(dbp, numbufs, offset, size, buf, tx); 1123 dmu_buf_rele_array(dbp, numbufs, FTAG); 1124 } 1125 1126 /* 1127 * Note: Lustre is an external consumer of this interface. 1128 */ 1129 void 1130 dmu_write_by_dnode(dnode_t *dn, uint64_t offset, uint64_t size, 1131 const void *buf, dmu_tx_t *tx) 1132 { 1133 dmu_buf_t **dbp; 1134 int numbufs; 1135 1136 if (size == 0) 1137 return; 1138 1139 VERIFY0(dmu_buf_hold_array_by_dnode(dn, offset, size, 1140 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH)); 1141 dmu_write_impl(dbp, numbufs, offset, size, buf, tx); 1142 dmu_buf_rele_array(dbp, numbufs, FTAG); 1143 } 1144 1145 void 1146 dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1147 dmu_tx_t *tx) 1148 { 1149 dmu_buf_t **dbp; 1150 int numbufs, i; 1151 1152 if (size == 0) 1153 return; 1154 1155 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size, 1156 FALSE, FTAG, &numbufs, &dbp)); 1157 1158 for (i = 0; i < numbufs; i++) { 1159 dmu_buf_t *db = dbp[i]; 1160 1161 dmu_buf_will_not_fill(db, tx); 1162 } 1163 dmu_buf_rele_array(dbp, numbufs, FTAG); 1164 } 1165 1166 void 1167 dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset, 1168 void *data, uint8_t etype, uint8_t comp, int uncompressed_size, 1169 int compressed_size, int byteorder, dmu_tx_t *tx) 1170 { 1171 dmu_buf_t *db; 1172 1173 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES); 1174 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS); 1175 VERIFY0(dmu_buf_hold_noread(os, object, offset, 1176 FTAG, &db)); 1177 1178 dmu_buf_write_embedded(db, 1179 data, (bp_embedded_type_t)etype, (enum zio_compress)comp, 1180 uncompressed_size, compressed_size, byteorder, tx); 1181 1182 dmu_buf_rele(db, FTAG); 1183 } 1184 1185 void 1186 dmu_redact(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1187 dmu_tx_t *tx) 1188 { 1189 int numbufs, i; 1190 dmu_buf_t **dbp; 1191 1192 VERIFY0(dmu_buf_hold_array(os, object, offset, size, FALSE, FTAG, 1193 &numbufs, &dbp)); 1194 for (i = 0; i < numbufs; i++) 1195 dmu_buf_redact(dbp[i], tx); 1196 dmu_buf_rele_array(dbp, numbufs, FTAG); 1197 } 1198 1199 #ifdef _KERNEL 1200 int 1201 dmu_read_uio_dnode(dnode_t *dn, zfs_uio_t *uio, uint64_t size) 1202 { 1203 dmu_buf_t **dbp; 1204 int numbufs, i, err; 1205 1206 /* 1207 * NB: we could do this block-at-a-time, but it's nice 1208 * to be reading in parallel. 1209 */ 1210 err = dmu_buf_hold_array_by_dnode(dn, zfs_uio_offset(uio), size, 1211 TRUE, FTAG, &numbufs, &dbp, 0); 1212 if (err) 1213 return (err); 1214 1215 for (i = 0; i < numbufs; i++) { 1216 uint64_t tocpy; 1217 int64_t bufoff; 1218 dmu_buf_t *db = dbp[i]; 1219 1220 ASSERT(size > 0); 1221 1222 bufoff = zfs_uio_offset(uio) - db->db_offset; 1223 tocpy = MIN(db->db_size - bufoff, size); 1224 1225 err = zfs_uio_fault_move((char *)db->db_data + bufoff, tocpy, 1226 UIO_READ, uio); 1227 1228 if (err) 1229 break; 1230 1231 size -= tocpy; 1232 } 1233 dmu_buf_rele_array(dbp, numbufs, FTAG); 1234 1235 return (err); 1236 } 1237 1238 /* 1239 * Read 'size' bytes into the uio buffer. 1240 * From object zdb->db_object. 1241 * Starting at zfs_uio_offset(uio). 1242 * 1243 * If the caller already has a dbuf in the target object 1244 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(), 1245 * because we don't have to find the dnode_t for the object. 1246 */ 1247 int 1248 dmu_read_uio_dbuf(dmu_buf_t *zdb, zfs_uio_t *uio, uint64_t size) 1249 { 1250 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb; 1251 dnode_t *dn; 1252 int err; 1253 1254 if (size == 0) 1255 return (0); 1256 1257 DB_DNODE_ENTER(db); 1258 dn = DB_DNODE(db); 1259 err = dmu_read_uio_dnode(dn, uio, size); 1260 DB_DNODE_EXIT(db); 1261 1262 return (err); 1263 } 1264 1265 /* 1266 * Read 'size' bytes into the uio buffer. 1267 * From the specified object 1268 * Starting at offset zfs_uio_offset(uio). 1269 */ 1270 int 1271 dmu_read_uio(objset_t *os, uint64_t object, zfs_uio_t *uio, uint64_t size) 1272 { 1273 dnode_t *dn; 1274 int err; 1275 1276 if (size == 0) 1277 return (0); 1278 1279 err = dnode_hold(os, object, FTAG, &dn); 1280 if (err) 1281 return (err); 1282 1283 err = dmu_read_uio_dnode(dn, uio, size); 1284 1285 dnode_rele(dn, FTAG); 1286 1287 return (err); 1288 } 1289 1290 int 1291 dmu_write_uio_dnode(dnode_t *dn, zfs_uio_t *uio, uint64_t size, dmu_tx_t *tx) 1292 { 1293 dmu_buf_t **dbp; 1294 int numbufs; 1295 int err = 0; 1296 int i; 1297 1298 err = dmu_buf_hold_array_by_dnode(dn, zfs_uio_offset(uio), size, 1299 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH); 1300 if (err) 1301 return (err); 1302 1303 for (i = 0; i < numbufs; i++) { 1304 uint64_t tocpy; 1305 int64_t bufoff; 1306 dmu_buf_t *db = dbp[i]; 1307 1308 ASSERT(size > 0); 1309 1310 bufoff = zfs_uio_offset(uio) - db->db_offset; 1311 tocpy = MIN(db->db_size - bufoff, size); 1312 1313 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 1314 1315 if (tocpy == db->db_size) 1316 dmu_buf_will_fill(db, tx); 1317 else 1318 dmu_buf_will_dirty(db, tx); 1319 1320 /* 1321 * XXX zfs_uiomove could block forever (eg.nfs-backed 1322 * pages). There needs to be a uiolockdown() function 1323 * to lock the pages in memory, so that zfs_uiomove won't 1324 * block. 1325 */ 1326 err = zfs_uio_fault_move((char *)db->db_data + bufoff, 1327 tocpy, UIO_WRITE, uio); 1328 1329 if (tocpy == db->db_size) 1330 dmu_buf_fill_done(db, tx); 1331 1332 if (err) 1333 break; 1334 1335 size -= tocpy; 1336 } 1337 1338 dmu_buf_rele_array(dbp, numbufs, FTAG); 1339 return (err); 1340 } 1341 1342 /* 1343 * Write 'size' bytes from the uio buffer. 1344 * To object zdb->db_object. 1345 * Starting at offset zfs_uio_offset(uio). 1346 * 1347 * If the caller already has a dbuf in the target object 1348 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(), 1349 * because we don't have to find the dnode_t for the object. 1350 */ 1351 int 1352 dmu_write_uio_dbuf(dmu_buf_t *zdb, zfs_uio_t *uio, uint64_t size, 1353 dmu_tx_t *tx) 1354 { 1355 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb; 1356 dnode_t *dn; 1357 int err; 1358 1359 if (size == 0) 1360 return (0); 1361 1362 DB_DNODE_ENTER(db); 1363 dn = DB_DNODE(db); 1364 err = dmu_write_uio_dnode(dn, uio, size, tx); 1365 DB_DNODE_EXIT(db); 1366 1367 return (err); 1368 } 1369 1370 /* 1371 * Write 'size' bytes from the uio buffer. 1372 * To the specified object. 1373 * Starting at offset zfs_uio_offset(uio). 1374 */ 1375 int 1376 dmu_write_uio(objset_t *os, uint64_t object, zfs_uio_t *uio, uint64_t size, 1377 dmu_tx_t *tx) 1378 { 1379 dnode_t *dn; 1380 int err; 1381 1382 if (size == 0) 1383 return (0); 1384 1385 err = dnode_hold(os, object, FTAG, &dn); 1386 if (err) 1387 return (err); 1388 1389 err = dmu_write_uio_dnode(dn, uio, size, tx); 1390 1391 dnode_rele(dn, FTAG); 1392 1393 return (err); 1394 } 1395 #endif /* _KERNEL */ 1396 1397 /* 1398 * Allocate a loaned anonymous arc buffer. 1399 */ 1400 arc_buf_t * 1401 dmu_request_arcbuf(dmu_buf_t *handle, int size) 1402 { 1403 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle; 1404 1405 return (arc_loan_buf(db->db_objset->os_spa, B_FALSE, size)); 1406 } 1407 1408 /* 1409 * Free a loaned arc buffer. 1410 */ 1411 void 1412 dmu_return_arcbuf(arc_buf_t *buf) 1413 { 1414 arc_return_buf(buf, FTAG); 1415 arc_buf_destroy(buf, FTAG); 1416 } 1417 1418 /* 1419 * A "lightweight" write is faster than a regular write (e.g. 1420 * dmu_write_by_dnode() or dmu_assign_arcbuf_by_dnode()), because it avoids the 1421 * CPU cost of creating a dmu_buf_impl_t and arc_buf_[hdr_]_t. However, the 1422 * data can not be read or overwritten until the transaction's txg has been 1423 * synced. This makes it appropriate for workloads that are known to be 1424 * (temporarily) write-only, like "zfs receive". 1425 * 1426 * A single block is written, starting at the specified offset in bytes. If 1427 * the call is successful, it returns 0 and the provided abd has been 1428 * consumed (the caller should not free it). 1429 */ 1430 int 1431 dmu_lightweight_write_by_dnode(dnode_t *dn, uint64_t offset, abd_t *abd, 1432 const zio_prop_t *zp, enum zio_flag flags, dmu_tx_t *tx) 1433 { 1434 dbuf_dirty_record_t *dr = 1435 dbuf_dirty_lightweight(dn, dbuf_whichblock(dn, 0, offset), tx); 1436 if (dr == NULL) 1437 return (SET_ERROR(EIO)); 1438 dr->dt.dll.dr_abd = abd; 1439 dr->dt.dll.dr_props = *zp; 1440 dr->dt.dll.dr_flags = flags; 1441 return (0); 1442 } 1443 1444 /* 1445 * When possible directly assign passed loaned arc buffer to a dbuf. 1446 * If this is not possible copy the contents of passed arc buf via 1447 * dmu_write(). 1448 */ 1449 int 1450 dmu_assign_arcbuf_by_dnode(dnode_t *dn, uint64_t offset, arc_buf_t *buf, 1451 dmu_tx_t *tx) 1452 { 1453 dmu_buf_impl_t *db; 1454 objset_t *os = dn->dn_objset; 1455 uint64_t object = dn->dn_object; 1456 uint32_t blksz = (uint32_t)arc_buf_lsize(buf); 1457 uint64_t blkid; 1458 1459 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1460 blkid = dbuf_whichblock(dn, 0, offset); 1461 db = dbuf_hold(dn, blkid, FTAG); 1462 if (db == NULL) 1463 return (SET_ERROR(EIO)); 1464 rw_exit(&dn->dn_struct_rwlock); 1465 1466 /* 1467 * We can only assign if the offset is aligned and the arc buf is the 1468 * same size as the dbuf. 1469 */ 1470 if (offset == db->db.db_offset && blksz == db->db.db_size) { 1471 zfs_racct_write(blksz, 1); 1472 dbuf_assign_arcbuf(db, buf, tx); 1473 dbuf_rele(db, FTAG); 1474 } else { 1475 /* compressed bufs must always be assignable to their dbuf */ 1476 ASSERT3U(arc_get_compression(buf), ==, ZIO_COMPRESS_OFF); 1477 ASSERT(!(buf->b_flags & ARC_BUF_FLAG_COMPRESSED)); 1478 1479 dbuf_rele(db, FTAG); 1480 dmu_write(os, object, offset, blksz, buf->b_data, tx); 1481 dmu_return_arcbuf(buf); 1482 } 1483 1484 return (0); 1485 } 1486 1487 int 1488 dmu_assign_arcbuf_by_dbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf, 1489 dmu_tx_t *tx) 1490 { 1491 int err; 1492 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle; 1493 1494 DB_DNODE_ENTER(dbuf); 1495 err = dmu_assign_arcbuf_by_dnode(DB_DNODE(dbuf), offset, buf, tx); 1496 DB_DNODE_EXIT(dbuf); 1497 1498 return (err); 1499 } 1500 1501 typedef struct { 1502 dbuf_dirty_record_t *dsa_dr; 1503 dmu_sync_cb_t *dsa_done; 1504 zgd_t *dsa_zgd; 1505 dmu_tx_t *dsa_tx; 1506 } dmu_sync_arg_t; 1507 1508 static void 1509 dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg) 1510 { 1511 (void) buf; 1512 dmu_sync_arg_t *dsa = varg; 1513 dmu_buf_t *db = dsa->dsa_zgd->zgd_db; 1514 blkptr_t *bp = zio->io_bp; 1515 1516 if (zio->io_error == 0) { 1517 if (BP_IS_HOLE(bp)) { 1518 /* 1519 * A block of zeros may compress to a hole, but the 1520 * block size still needs to be known for replay. 1521 */ 1522 BP_SET_LSIZE(bp, db->db_size); 1523 } else if (!BP_IS_EMBEDDED(bp)) { 1524 ASSERT(BP_GET_LEVEL(bp) == 0); 1525 BP_SET_FILL(bp, 1); 1526 } 1527 } 1528 } 1529 1530 static void 1531 dmu_sync_late_arrival_ready(zio_t *zio) 1532 { 1533 dmu_sync_ready(zio, NULL, zio->io_private); 1534 } 1535 1536 static void 1537 dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg) 1538 { 1539 (void) buf; 1540 dmu_sync_arg_t *dsa = varg; 1541 dbuf_dirty_record_t *dr = dsa->dsa_dr; 1542 dmu_buf_impl_t *db = dr->dr_dbuf; 1543 zgd_t *zgd = dsa->dsa_zgd; 1544 1545 /* 1546 * Record the vdev(s) backing this blkptr so they can be flushed after 1547 * the writes for the lwb have completed. 1548 */ 1549 if (zio->io_error == 0) { 1550 zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp); 1551 } 1552 1553 mutex_enter(&db->db_mtx); 1554 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC); 1555 if (zio->io_error == 0) { 1556 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE); 1557 if (dr->dt.dl.dr_nopwrite) { 1558 blkptr_t *bp = zio->io_bp; 1559 blkptr_t *bp_orig = &zio->io_bp_orig; 1560 uint8_t chksum = BP_GET_CHECKSUM(bp_orig); 1561 1562 ASSERT(BP_EQUAL(bp, bp_orig)); 1563 VERIFY(BP_EQUAL(bp, db->db_blkptr)); 1564 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF); 1565 VERIFY(zio_checksum_table[chksum].ci_flags & 1566 ZCHECKSUM_FLAG_NOPWRITE); 1567 } 1568 dr->dt.dl.dr_overridden_by = *zio->io_bp; 1569 dr->dt.dl.dr_override_state = DR_OVERRIDDEN; 1570 dr->dt.dl.dr_copies = zio->io_prop.zp_copies; 1571 1572 /* 1573 * Old style holes are filled with all zeros, whereas 1574 * new-style holes maintain their lsize, type, level, 1575 * and birth time (see zio_write_compress). While we 1576 * need to reset the BP_SET_LSIZE() call that happened 1577 * in dmu_sync_ready for old style holes, we do *not* 1578 * want to wipe out the information contained in new 1579 * style holes. Thus, only zero out the block pointer if 1580 * it's an old style hole. 1581 */ 1582 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) && 1583 dr->dt.dl.dr_overridden_by.blk_birth == 0) 1584 BP_ZERO(&dr->dt.dl.dr_overridden_by); 1585 } else { 1586 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 1587 } 1588 cv_broadcast(&db->db_changed); 1589 mutex_exit(&db->db_mtx); 1590 1591 dsa->dsa_done(dsa->dsa_zgd, zio->io_error); 1592 1593 kmem_free(dsa, sizeof (*dsa)); 1594 } 1595 1596 static void 1597 dmu_sync_late_arrival_done(zio_t *zio) 1598 { 1599 blkptr_t *bp = zio->io_bp; 1600 dmu_sync_arg_t *dsa = zio->io_private; 1601 zgd_t *zgd = dsa->dsa_zgd; 1602 1603 if (zio->io_error == 0) { 1604 /* 1605 * Record the vdev(s) backing this blkptr so they can be 1606 * flushed after the writes for the lwb have completed. 1607 */ 1608 zil_lwb_add_block(zgd->zgd_lwb, zgd->zgd_bp); 1609 1610 if (!BP_IS_HOLE(bp)) { 1611 blkptr_t *bp_orig __maybe_unused = &zio->io_bp_orig; 1612 ASSERT(!(zio->io_flags & ZIO_FLAG_NOPWRITE)); 1613 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig)); 1614 ASSERT(zio->io_bp->blk_birth == zio->io_txg); 1615 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa)); 1616 zio_free(zio->io_spa, zio->io_txg, zio->io_bp); 1617 } 1618 } 1619 1620 dmu_tx_commit(dsa->dsa_tx); 1621 1622 dsa->dsa_done(dsa->dsa_zgd, zio->io_error); 1623 1624 abd_free(zio->io_abd); 1625 kmem_free(dsa, sizeof (*dsa)); 1626 } 1627 1628 static int 1629 dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd, 1630 zio_prop_t *zp, zbookmark_phys_t *zb) 1631 { 1632 dmu_sync_arg_t *dsa; 1633 dmu_tx_t *tx; 1634 1635 tx = dmu_tx_create(os); 1636 dmu_tx_hold_space(tx, zgd->zgd_db->db_size); 1637 if (dmu_tx_assign(tx, TXG_WAIT) != 0) { 1638 dmu_tx_abort(tx); 1639 /* Make zl_get_data do txg_waited_synced() */ 1640 return (SET_ERROR(EIO)); 1641 } 1642 1643 /* 1644 * In order to prevent the zgd's lwb from being free'd prior to 1645 * dmu_sync_late_arrival_done() being called, we have to ensure 1646 * the lwb's "max txg" takes this tx's txg into account. 1647 */ 1648 zil_lwb_add_txg(zgd->zgd_lwb, dmu_tx_get_txg(tx)); 1649 1650 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP); 1651 dsa->dsa_dr = NULL; 1652 dsa->dsa_done = done; 1653 dsa->dsa_zgd = zgd; 1654 dsa->dsa_tx = tx; 1655 1656 /* 1657 * Since we are currently syncing this txg, it's nontrivial to 1658 * determine what BP to nopwrite against, so we disable nopwrite. 1659 * 1660 * When syncing, the db_blkptr is initially the BP of the previous 1661 * txg. We can not nopwrite against it because it will be changed 1662 * (this is similar to the non-late-arrival case where the dbuf is 1663 * dirty in a future txg). 1664 * 1665 * Then dbuf_write_ready() sets bp_blkptr to the location we will write. 1666 * We can not nopwrite against it because although the BP will not 1667 * (typically) be changed, the data has not yet been persisted to this 1668 * location. 1669 * 1670 * Finally, when dbuf_write_done() is called, it is theoretically 1671 * possible to always nopwrite, because the data that was written in 1672 * this txg is the same data that we are trying to write. However we 1673 * would need to check that this dbuf is not dirty in any future 1674 * txg's (as we do in the normal dmu_sync() path). For simplicity, we 1675 * don't nopwrite in this case. 1676 */ 1677 zp->zp_nopwrite = B_FALSE; 1678 1679 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp, 1680 abd_get_from_buf(zgd->zgd_db->db_data, zgd->zgd_db->db_size), 1681 zgd->zgd_db->db_size, zgd->zgd_db->db_size, zp, 1682 dmu_sync_late_arrival_ready, NULL, NULL, dmu_sync_late_arrival_done, 1683 dsa, ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb)); 1684 1685 return (0); 1686 } 1687 1688 /* 1689 * Intent log support: sync the block associated with db to disk. 1690 * N.B. and XXX: the caller is responsible for making sure that the 1691 * data isn't changing while dmu_sync() is writing it. 1692 * 1693 * Return values: 1694 * 1695 * EEXIST: this txg has already been synced, so there's nothing to do. 1696 * The caller should not log the write. 1697 * 1698 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do. 1699 * The caller should not log the write. 1700 * 1701 * EALREADY: this block is already in the process of being synced. 1702 * The caller should track its progress (somehow). 1703 * 1704 * EIO: could not do the I/O. 1705 * The caller should do a txg_wait_synced(). 1706 * 1707 * 0: the I/O has been initiated. 1708 * The caller should log this blkptr in the done callback. 1709 * It is possible that the I/O will fail, in which case 1710 * the error will be reported to the done callback and 1711 * propagated to pio from zio_done(). 1712 */ 1713 int 1714 dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd) 1715 { 1716 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db; 1717 objset_t *os = db->db_objset; 1718 dsl_dataset_t *ds = os->os_dsl_dataset; 1719 dbuf_dirty_record_t *dr, *dr_next; 1720 dmu_sync_arg_t *dsa; 1721 zbookmark_phys_t zb; 1722 zio_prop_t zp; 1723 dnode_t *dn; 1724 1725 ASSERT(pio != NULL); 1726 ASSERT(txg != 0); 1727 1728 SET_BOOKMARK(&zb, ds->ds_object, 1729 db->db.db_object, db->db_level, db->db_blkid); 1730 1731 DB_DNODE_ENTER(db); 1732 dn = DB_DNODE(db); 1733 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp); 1734 DB_DNODE_EXIT(db); 1735 1736 /* 1737 * If we're frozen (running ziltest), we always need to generate a bp. 1738 */ 1739 if (txg > spa_freeze_txg(os->os_spa)) 1740 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); 1741 1742 /* 1743 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf() 1744 * and us. If we determine that this txg is not yet syncing, 1745 * but it begins to sync a moment later, that's OK because the 1746 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx. 1747 */ 1748 mutex_enter(&db->db_mtx); 1749 1750 if (txg <= spa_last_synced_txg(os->os_spa)) { 1751 /* 1752 * This txg has already synced. There's nothing to do. 1753 */ 1754 mutex_exit(&db->db_mtx); 1755 return (SET_ERROR(EEXIST)); 1756 } 1757 1758 if (txg <= spa_syncing_txg(os->os_spa)) { 1759 /* 1760 * This txg is currently syncing, so we can't mess with 1761 * the dirty record anymore; just write a new log block. 1762 */ 1763 mutex_exit(&db->db_mtx); 1764 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); 1765 } 1766 1767 dr = dbuf_find_dirty_eq(db, txg); 1768 1769 if (dr == NULL) { 1770 /* 1771 * There's no dr for this dbuf, so it must have been freed. 1772 * There's no need to log writes to freed blocks, so we're done. 1773 */ 1774 mutex_exit(&db->db_mtx); 1775 return (SET_ERROR(ENOENT)); 1776 } 1777 1778 dr_next = list_next(&db->db_dirty_records, dr); 1779 ASSERT(dr_next == NULL || dr_next->dr_txg < txg); 1780 1781 if (db->db_blkptr != NULL) { 1782 /* 1783 * We need to fill in zgd_bp with the current blkptr so that 1784 * the nopwrite code can check if we're writing the same 1785 * data that's already on disk. We can only nopwrite if we 1786 * are sure that after making the copy, db_blkptr will not 1787 * change until our i/o completes. We ensure this by 1788 * holding the db_mtx, and only allowing nopwrite if the 1789 * block is not already dirty (see below). This is verified 1790 * by dmu_sync_done(), which VERIFYs that the db_blkptr has 1791 * not changed. 1792 */ 1793 *zgd->zgd_bp = *db->db_blkptr; 1794 } 1795 1796 /* 1797 * Assume the on-disk data is X, the current syncing data (in 1798 * txg - 1) is Y, and the current in-memory data is Z (currently 1799 * in dmu_sync). 1800 * 1801 * We usually want to perform a nopwrite if X and Z are the 1802 * same. However, if Y is different (i.e. the BP is going to 1803 * change before this write takes effect), then a nopwrite will 1804 * be incorrect - we would override with X, which could have 1805 * been freed when Y was written. 1806 * 1807 * (Note that this is not a concern when we are nop-writing from 1808 * syncing context, because X and Y must be identical, because 1809 * all previous txgs have been synced.) 1810 * 1811 * Therefore, we disable nopwrite if the current BP could change 1812 * before this TXG. There are two ways it could change: by 1813 * being dirty (dr_next is non-NULL), or by being freed 1814 * (dnode_block_freed()). This behavior is verified by 1815 * zio_done(), which VERIFYs that the override BP is identical 1816 * to the on-disk BP. 1817 */ 1818 DB_DNODE_ENTER(db); 1819 dn = DB_DNODE(db); 1820 if (dr_next != NULL || dnode_block_freed(dn, db->db_blkid)) 1821 zp.zp_nopwrite = B_FALSE; 1822 DB_DNODE_EXIT(db); 1823 1824 ASSERT(dr->dr_txg == txg); 1825 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC || 1826 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { 1827 /* 1828 * We have already issued a sync write for this buffer, 1829 * or this buffer has already been synced. It could not 1830 * have been dirtied since, or we would have cleared the state. 1831 */ 1832 mutex_exit(&db->db_mtx); 1833 return (SET_ERROR(EALREADY)); 1834 } 1835 1836 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); 1837 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC; 1838 mutex_exit(&db->db_mtx); 1839 1840 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP); 1841 dsa->dsa_dr = dr; 1842 dsa->dsa_done = done; 1843 dsa->dsa_zgd = zgd; 1844 dsa->dsa_tx = NULL; 1845 1846 zio_nowait(arc_write(pio, os->os_spa, txg, 1847 zgd->zgd_bp, dr->dt.dl.dr_data, dbuf_is_l2cacheable(db), 1848 &zp, dmu_sync_ready, NULL, NULL, dmu_sync_done, dsa, 1849 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb)); 1850 1851 return (0); 1852 } 1853 1854 int 1855 dmu_object_set_nlevels(objset_t *os, uint64_t object, int nlevels, dmu_tx_t *tx) 1856 { 1857 dnode_t *dn; 1858 int err; 1859 1860 err = dnode_hold(os, object, FTAG, &dn); 1861 if (err) 1862 return (err); 1863 err = dnode_set_nlevels(dn, nlevels, tx); 1864 dnode_rele(dn, FTAG); 1865 return (err); 1866 } 1867 1868 int 1869 dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs, 1870 dmu_tx_t *tx) 1871 { 1872 dnode_t *dn; 1873 int err; 1874 1875 err = dnode_hold(os, object, FTAG, &dn); 1876 if (err) 1877 return (err); 1878 err = dnode_set_blksz(dn, size, ibs, tx); 1879 dnode_rele(dn, FTAG); 1880 return (err); 1881 } 1882 1883 int 1884 dmu_object_set_maxblkid(objset_t *os, uint64_t object, uint64_t maxblkid, 1885 dmu_tx_t *tx) 1886 { 1887 dnode_t *dn; 1888 int err; 1889 1890 err = dnode_hold(os, object, FTAG, &dn); 1891 if (err) 1892 return (err); 1893 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 1894 dnode_new_blkid(dn, maxblkid, tx, B_FALSE, B_TRUE); 1895 rw_exit(&dn->dn_struct_rwlock); 1896 dnode_rele(dn, FTAG); 1897 return (0); 1898 } 1899 1900 void 1901 dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum, 1902 dmu_tx_t *tx) 1903 { 1904 dnode_t *dn; 1905 1906 /* 1907 * Send streams include each object's checksum function. This 1908 * check ensures that the receiving system can understand the 1909 * checksum function transmitted. 1910 */ 1911 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS); 1912 1913 VERIFY0(dnode_hold(os, object, FTAG, &dn)); 1914 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS); 1915 dn->dn_checksum = checksum; 1916 dnode_setdirty(dn, tx); 1917 dnode_rele(dn, FTAG); 1918 } 1919 1920 void 1921 dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress, 1922 dmu_tx_t *tx) 1923 { 1924 dnode_t *dn; 1925 1926 /* 1927 * Send streams include each object's compression function. This 1928 * check ensures that the receiving system can understand the 1929 * compression function transmitted. 1930 */ 1931 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS); 1932 1933 VERIFY0(dnode_hold(os, object, FTAG, &dn)); 1934 dn->dn_compress = compress; 1935 dnode_setdirty(dn, tx); 1936 dnode_rele(dn, FTAG); 1937 } 1938 1939 /* 1940 * When the "redundant_metadata" property is set to "most", only indirect 1941 * blocks of this level and higher will have an additional ditto block. 1942 */ 1943 static const int zfs_redundant_metadata_most_ditto_level = 2; 1944 1945 void 1946 dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp) 1947 { 1948 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET; 1949 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) || 1950 (wp & WP_SPILL)); 1951 enum zio_checksum checksum = os->os_checksum; 1952 enum zio_compress compress = os->os_compress; 1953 uint8_t complevel = os->os_complevel; 1954 enum zio_checksum dedup_checksum = os->os_dedup_checksum; 1955 boolean_t dedup = B_FALSE; 1956 boolean_t nopwrite = B_FALSE; 1957 boolean_t dedup_verify = os->os_dedup_verify; 1958 boolean_t encrypt = B_FALSE; 1959 int copies = os->os_copies; 1960 1961 /* 1962 * We maintain different write policies for each of the following 1963 * types of data: 1964 * 1. metadata 1965 * 2. preallocated blocks (i.e. level-0 blocks of a dump device) 1966 * 3. all other level 0 blocks 1967 */ 1968 if (ismd) { 1969 /* 1970 * XXX -- we should design a compression algorithm 1971 * that specializes in arrays of bps. 1972 */ 1973 compress = zio_compress_select(os->os_spa, 1974 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON); 1975 1976 /* 1977 * Metadata always gets checksummed. If the data 1978 * checksum is multi-bit correctable, and it's not a 1979 * ZBT-style checksum, then it's suitable for metadata 1980 * as well. Otherwise, the metadata checksum defaults 1981 * to fletcher4. 1982 */ 1983 if (!(zio_checksum_table[checksum].ci_flags & 1984 ZCHECKSUM_FLAG_METADATA) || 1985 (zio_checksum_table[checksum].ci_flags & 1986 ZCHECKSUM_FLAG_EMBEDDED)) 1987 checksum = ZIO_CHECKSUM_FLETCHER_4; 1988 1989 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL || 1990 (os->os_redundant_metadata == 1991 ZFS_REDUNDANT_METADATA_MOST && 1992 (level >= zfs_redundant_metadata_most_ditto_level || 1993 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL)))) 1994 copies++; 1995 } else if (wp & WP_NOFILL) { 1996 ASSERT(level == 0); 1997 1998 /* 1999 * If we're writing preallocated blocks, we aren't actually 2000 * writing them so don't set any policy properties. These 2001 * blocks are currently only used by an external subsystem 2002 * outside of zfs (i.e. dump) and not written by the zio 2003 * pipeline. 2004 */ 2005 compress = ZIO_COMPRESS_OFF; 2006 checksum = ZIO_CHECKSUM_OFF; 2007 } else { 2008 compress = zio_compress_select(os->os_spa, dn->dn_compress, 2009 compress); 2010 complevel = zio_complevel_select(os->os_spa, compress, 2011 complevel, complevel); 2012 2013 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ? 2014 zio_checksum_select(dn->dn_checksum, checksum) : 2015 dedup_checksum; 2016 2017 /* 2018 * Determine dedup setting. If we are in dmu_sync(), 2019 * we won't actually dedup now because that's all 2020 * done in syncing context; but we do want to use the 2021 * dedup checksum. If the checksum is not strong 2022 * enough to ensure unique signatures, force 2023 * dedup_verify. 2024 */ 2025 if (dedup_checksum != ZIO_CHECKSUM_OFF) { 2026 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE; 2027 if (!(zio_checksum_table[checksum].ci_flags & 2028 ZCHECKSUM_FLAG_DEDUP)) 2029 dedup_verify = B_TRUE; 2030 } 2031 2032 /* 2033 * Enable nopwrite if we have secure enough checksum 2034 * algorithm (see comment in zio_nop_write) and 2035 * compression is enabled. We don't enable nopwrite if 2036 * dedup is enabled as the two features are mutually 2037 * exclusive. 2038 */ 2039 nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags & 2040 ZCHECKSUM_FLAG_NOPWRITE) && 2041 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled); 2042 } 2043 2044 /* 2045 * All objects in an encrypted objset are protected from modification 2046 * via a MAC. Encrypted objects store their IV and salt in the last DVA 2047 * in the bp, so we cannot use all copies. Encrypted objects are also 2048 * not subject to nopwrite since writing the same data will still 2049 * result in a new ciphertext. Only encrypted blocks can be dedup'd 2050 * to avoid ambiguity in the dedup code since the DDT does not store 2051 * object types. 2052 */ 2053 if (os->os_encrypted && (wp & WP_NOFILL) == 0) { 2054 encrypt = B_TRUE; 2055 2056 if (DMU_OT_IS_ENCRYPTED(type)) { 2057 copies = MIN(copies, SPA_DVAS_PER_BP - 1); 2058 nopwrite = B_FALSE; 2059 } else { 2060 dedup = B_FALSE; 2061 } 2062 2063 if (level <= 0 && 2064 (type == DMU_OT_DNODE || type == DMU_OT_OBJSET)) { 2065 compress = ZIO_COMPRESS_EMPTY; 2066 } 2067 } 2068 2069 zp->zp_compress = compress; 2070 zp->zp_complevel = complevel; 2071 zp->zp_checksum = checksum; 2072 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type; 2073 zp->zp_level = level; 2074 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa)); 2075 zp->zp_dedup = dedup; 2076 zp->zp_dedup_verify = dedup && dedup_verify; 2077 zp->zp_nopwrite = nopwrite; 2078 zp->zp_encrypt = encrypt; 2079 zp->zp_byteorder = ZFS_HOST_BYTEORDER; 2080 memset(zp->zp_salt, 0, ZIO_DATA_SALT_LEN); 2081 memset(zp->zp_iv, 0, ZIO_DATA_IV_LEN); 2082 memset(zp->zp_mac, 0, ZIO_DATA_MAC_LEN); 2083 zp->zp_zpl_smallblk = DMU_OT_IS_FILE(zp->zp_type) ? 2084 os->os_zpl_special_smallblock : 0; 2085 2086 ASSERT3U(zp->zp_compress, !=, ZIO_COMPRESS_INHERIT); 2087 } 2088 2089 /* 2090 * This function is only called from zfs_holey_common() for zpl_llseek() 2091 * in order to determine the location of holes. In order to accurately 2092 * report holes all dirty data must be synced to disk. This causes extremely 2093 * poor performance when seeking for holes in a dirty file. As a compromise, 2094 * only provide hole data when the dnode is clean. When a dnode is dirty 2095 * report the dnode as having no holes which is always a safe thing to do. 2096 */ 2097 int 2098 dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off) 2099 { 2100 dnode_t *dn; 2101 int err; 2102 2103 restart: 2104 err = dnode_hold(os, object, FTAG, &dn); 2105 if (err) 2106 return (err); 2107 2108 rw_enter(&dn->dn_struct_rwlock, RW_READER); 2109 2110 if (dnode_is_dirty(dn)) { 2111 /* 2112 * If the zfs_dmu_offset_next_sync module option is enabled 2113 * then strict hole reporting has been requested. Dirty 2114 * dnodes must be synced to disk to accurately report all 2115 * holes. When disabled dirty dnodes are reported to not 2116 * have any holes which is always safe. 2117 * 2118 * When called by zfs_holey_common() the zp->z_rangelock 2119 * is held to prevent zfs_write() and mmap writeback from 2120 * re-dirtying the dnode after txg_wait_synced(). 2121 */ 2122 if (zfs_dmu_offset_next_sync) { 2123 rw_exit(&dn->dn_struct_rwlock); 2124 dnode_rele(dn, FTAG); 2125 txg_wait_synced(dmu_objset_pool(os), 0); 2126 goto restart; 2127 } 2128 2129 err = SET_ERROR(EBUSY); 2130 } else { 2131 err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK | 2132 (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0); 2133 } 2134 2135 rw_exit(&dn->dn_struct_rwlock); 2136 dnode_rele(dn, FTAG); 2137 2138 return (err); 2139 } 2140 2141 void 2142 __dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi) 2143 { 2144 dnode_phys_t *dnp = dn->dn_phys; 2145 2146 doi->doi_data_block_size = dn->dn_datablksz; 2147 doi->doi_metadata_block_size = dn->dn_indblkshift ? 2148 1ULL << dn->dn_indblkshift : 0; 2149 doi->doi_type = dn->dn_type; 2150 doi->doi_bonus_type = dn->dn_bonustype; 2151 doi->doi_bonus_size = dn->dn_bonuslen; 2152 doi->doi_dnodesize = dn->dn_num_slots << DNODE_SHIFT; 2153 doi->doi_indirection = dn->dn_nlevels; 2154 doi->doi_checksum = dn->dn_checksum; 2155 doi->doi_compress = dn->dn_compress; 2156 doi->doi_nblkptr = dn->dn_nblkptr; 2157 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9; 2158 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz; 2159 doi->doi_fill_count = 0; 2160 for (int i = 0; i < dnp->dn_nblkptr; i++) 2161 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]); 2162 } 2163 2164 void 2165 dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi) 2166 { 2167 rw_enter(&dn->dn_struct_rwlock, RW_READER); 2168 mutex_enter(&dn->dn_mtx); 2169 2170 __dmu_object_info_from_dnode(dn, doi); 2171 2172 mutex_exit(&dn->dn_mtx); 2173 rw_exit(&dn->dn_struct_rwlock); 2174 } 2175 2176 /* 2177 * Get information on a DMU object. 2178 * If doi is NULL, just indicates whether the object exists. 2179 */ 2180 int 2181 dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi) 2182 { 2183 dnode_t *dn; 2184 int err = dnode_hold(os, object, FTAG, &dn); 2185 2186 if (err) 2187 return (err); 2188 2189 if (doi != NULL) 2190 dmu_object_info_from_dnode(dn, doi); 2191 2192 dnode_rele(dn, FTAG); 2193 return (0); 2194 } 2195 2196 /* 2197 * As above, but faster; can be used when you have a held dbuf in hand. 2198 */ 2199 void 2200 dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi) 2201 { 2202 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2203 2204 DB_DNODE_ENTER(db); 2205 dmu_object_info_from_dnode(DB_DNODE(db), doi); 2206 DB_DNODE_EXIT(db); 2207 } 2208 2209 /* 2210 * Faster still when you only care about the size. 2211 */ 2212 void 2213 dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize, 2214 u_longlong_t *nblk512) 2215 { 2216 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2217 dnode_t *dn; 2218 2219 DB_DNODE_ENTER(db); 2220 dn = DB_DNODE(db); 2221 2222 *blksize = dn->dn_datablksz; 2223 /* add in number of slots used for the dnode itself */ 2224 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >> 2225 SPA_MINBLOCKSHIFT) + dn->dn_num_slots; 2226 DB_DNODE_EXIT(db); 2227 } 2228 2229 void 2230 dmu_object_dnsize_from_db(dmu_buf_t *db_fake, int *dnsize) 2231 { 2232 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2233 dnode_t *dn; 2234 2235 DB_DNODE_ENTER(db); 2236 dn = DB_DNODE(db); 2237 *dnsize = dn->dn_num_slots << DNODE_SHIFT; 2238 DB_DNODE_EXIT(db); 2239 } 2240 2241 void 2242 byteswap_uint64_array(void *vbuf, size_t size) 2243 { 2244 uint64_t *buf = vbuf; 2245 size_t count = size >> 3; 2246 int i; 2247 2248 ASSERT((size & 7) == 0); 2249 2250 for (i = 0; i < count; i++) 2251 buf[i] = BSWAP_64(buf[i]); 2252 } 2253 2254 void 2255 byteswap_uint32_array(void *vbuf, size_t size) 2256 { 2257 uint32_t *buf = vbuf; 2258 size_t count = size >> 2; 2259 int i; 2260 2261 ASSERT((size & 3) == 0); 2262 2263 for (i = 0; i < count; i++) 2264 buf[i] = BSWAP_32(buf[i]); 2265 } 2266 2267 void 2268 byteswap_uint16_array(void *vbuf, size_t size) 2269 { 2270 uint16_t *buf = vbuf; 2271 size_t count = size >> 1; 2272 int i; 2273 2274 ASSERT((size & 1) == 0); 2275 2276 for (i = 0; i < count; i++) 2277 buf[i] = BSWAP_16(buf[i]); 2278 } 2279 2280 void 2281 byteswap_uint8_array(void *vbuf, size_t size) 2282 { 2283 (void) vbuf, (void) size; 2284 } 2285 2286 void 2287 dmu_init(void) 2288 { 2289 abd_init(); 2290 zfs_dbgmsg_init(); 2291 sa_cache_init(); 2292 dmu_objset_init(); 2293 dnode_init(); 2294 zfetch_init(); 2295 dmu_tx_init(); 2296 l2arc_init(); 2297 arc_init(); 2298 dbuf_init(); 2299 } 2300 2301 void 2302 dmu_fini(void) 2303 { 2304 arc_fini(); /* arc depends on l2arc, so arc must go first */ 2305 l2arc_fini(); 2306 dmu_tx_fini(); 2307 zfetch_fini(); 2308 dbuf_fini(); 2309 dnode_fini(); 2310 dmu_objset_fini(); 2311 sa_cache_fini(); 2312 zfs_dbgmsg_fini(); 2313 abd_fini(); 2314 } 2315 2316 EXPORT_SYMBOL(dmu_bonus_hold); 2317 EXPORT_SYMBOL(dmu_bonus_hold_by_dnode); 2318 EXPORT_SYMBOL(dmu_buf_hold_array_by_bonus); 2319 EXPORT_SYMBOL(dmu_buf_rele_array); 2320 EXPORT_SYMBOL(dmu_prefetch); 2321 EXPORT_SYMBOL(dmu_free_range); 2322 EXPORT_SYMBOL(dmu_free_long_range); 2323 EXPORT_SYMBOL(dmu_free_long_object); 2324 EXPORT_SYMBOL(dmu_read); 2325 EXPORT_SYMBOL(dmu_read_by_dnode); 2326 EXPORT_SYMBOL(dmu_write); 2327 EXPORT_SYMBOL(dmu_write_by_dnode); 2328 EXPORT_SYMBOL(dmu_prealloc); 2329 EXPORT_SYMBOL(dmu_object_info); 2330 EXPORT_SYMBOL(dmu_object_info_from_dnode); 2331 EXPORT_SYMBOL(dmu_object_info_from_db); 2332 EXPORT_SYMBOL(dmu_object_size_from_db); 2333 EXPORT_SYMBOL(dmu_object_dnsize_from_db); 2334 EXPORT_SYMBOL(dmu_object_set_nlevels); 2335 EXPORT_SYMBOL(dmu_object_set_blocksize); 2336 EXPORT_SYMBOL(dmu_object_set_maxblkid); 2337 EXPORT_SYMBOL(dmu_object_set_checksum); 2338 EXPORT_SYMBOL(dmu_object_set_compress); 2339 EXPORT_SYMBOL(dmu_offset_next); 2340 EXPORT_SYMBOL(dmu_write_policy); 2341 EXPORT_SYMBOL(dmu_sync); 2342 EXPORT_SYMBOL(dmu_request_arcbuf); 2343 EXPORT_SYMBOL(dmu_return_arcbuf); 2344 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dnode); 2345 EXPORT_SYMBOL(dmu_assign_arcbuf_by_dbuf); 2346 EXPORT_SYMBOL(dmu_buf_hold); 2347 EXPORT_SYMBOL(dmu_ot); 2348 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 /* CSTYLED */ 2359 ZFS_MODULE_PARAM(zfs, , dmu_prefetch_max, INT, ZMOD_RW, 2360 "Limit one prefetch call to this size"); 2361