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, 2015 by Delphix. All rights reserved. 24 */ 25/* Copyright (c) 2013 by Saso Kiselkov. All rights reserved. */ 26/* Copyright (c) 2013, Joyent, Inc. All rights reserved. */ 27/* Copyright (c) 2014, Nexenta Systems, Inc. All rights reserved. */ 28 29#include <sys/dmu.h> 30#include <sys/dmu_impl.h> 31#include <sys/dmu_tx.h> 32#include <sys/dbuf.h> 33#include <sys/dnode.h> 34#include <sys/zfs_context.h> 35#include <sys/dmu_objset.h> 36#include <sys/dmu_traverse.h> 37#include <sys/dsl_dataset.h> 38#include <sys/dsl_dir.h> 39#include <sys/dsl_pool.h> 40#include <sys/dsl_synctask.h> 41#include <sys/dsl_prop.h> 42#include <sys/dmu_zfetch.h> 43#include <sys/zfs_ioctl.h> 44#include <sys/zap.h> 45#include <sys/zio_checksum.h> 46#include <sys/zio_compress.h> 47#include <sys/sa.h> 48#include <sys/zfeature.h> 49#ifdef _KERNEL 50#include <sys/racct.h> 51#include <sys/vm.h> 52#include <sys/zfs_znode.h> 53#endif 54 55/* 56 * Enable/disable nopwrite feature. 57 */ 58int zfs_nopwrite_enabled = 1; 59SYSCTL_DECL(_vfs_zfs); 60SYSCTL_INT(_vfs_zfs, OID_AUTO, nopwrite_enabled, CTLFLAG_RDTUN, 61 &zfs_nopwrite_enabled, 0, "Enable nopwrite feature"); 62 63const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = { 64 { DMU_BSWAP_UINT8, TRUE, "unallocated" }, 65 { DMU_BSWAP_ZAP, TRUE, "object directory" }, 66 { DMU_BSWAP_UINT64, TRUE, "object array" }, 67 { DMU_BSWAP_UINT8, TRUE, "packed nvlist" }, 68 { DMU_BSWAP_UINT64, TRUE, "packed nvlist size" }, 69 { DMU_BSWAP_UINT64, TRUE, "bpobj" }, 70 { DMU_BSWAP_UINT64, TRUE, "bpobj header" }, 71 { DMU_BSWAP_UINT64, TRUE, "SPA space map header" }, 72 { DMU_BSWAP_UINT64, TRUE, "SPA space map" }, 73 { DMU_BSWAP_UINT64, TRUE, "ZIL intent log" }, 74 { DMU_BSWAP_DNODE, TRUE, "DMU dnode" }, 75 { DMU_BSWAP_OBJSET, TRUE, "DMU objset" }, 76 { DMU_BSWAP_UINT64, TRUE, "DSL directory" }, 77 { DMU_BSWAP_ZAP, TRUE, "DSL directory child map"}, 78 { DMU_BSWAP_ZAP, TRUE, "DSL dataset snap map" }, 79 { DMU_BSWAP_ZAP, TRUE, "DSL props" }, 80 { DMU_BSWAP_UINT64, TRUE, "DSL dataset" }, 81 { DMU_BSWAP_ZNODE, TRUE, "ZFS znode" }, 82 { DMU_BSWAP_OLDACL, TRUE, "ZFS V0 ACL" }, 83 { DMU_BSWAP_UINT8, FALSE, "ZFS plain file" }, 84 { DMU_BSWAP_ZAP, TRUE, "ZFS directory" }, 85 { DMU_BSWAP_ZAP, TRUE, "ZFS master node" }, 86 { DMU_BSWAP_ZAP, TRUE, "ZFS delete queue" }, 87 { DMU_BSWAP_UINT8, FALSE, "zvol object" }, 88 { DMU_BSWAP_ZAP, TRUE, "zvol prop" }, 89 { DMU_BSWAP_UINT8, FALSE, "other uint8[]" }, 90 { DMU_BSWAP_UINT64, FALSE, "other uint64[]" }, 91 { DMU_BSWAP_ZAP, TRUE, "other ZAP" }, 92 { DMU_BSWAP_ZAP, TRUE, "persistent error log" }, 93 { DMU_BSWAP_UINT8, TRUE, "SPA history" }, 94 { DMU_BSWAP_UINT64, TRUE, "SPA history offsets" }, 95 { DMU_BSWAP_ZAP, TRUE, "Pool properties" }, 96 { DMU_BSWAP_ZAP, TRUE, "DSL permissions" }, 97 { DMU_BSWAP_ACL, TRUE, "ZFS ACL" }, 98 { DMU_BSWAP_UINT8, TRUE, "ZFS SYSACL" }, 99 { DMU_BSWAP_UINT8, TRUE, "FUID table" }, 100 { DMU_BSWAP_UINT64, TRUE, "FUID table size" }, 101 { DMU_BSWAP_ZAP, TRUE, "DSL dataset next clones"}, 102 { DMU_BSWAP_ZAP, TRUE, "scan work queue" }, 103 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group used" }, 104 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group quota" }, 105 { DMU_BSWAP_ZAP, TRUE, "snapshot refcount tags"}, 106 { DMU_BSWAP_ZAP, TRUE, "DDT ZAP algorithm" }, 107 { DMU_BSWAP_ZAP, TRUE, "DDT statistics" }, 108 { DMU_BSWAP_UINT8, TRUE, "System attributes" }, 109 { DMU_BSWAP_ZAP, TRUE, "SA master node" }, 110 { DMU_BSWAP_ZAP, TRUE, "SA attr registration" }, 111 { DMU_BSWAP_ZAP, TRUE, "SA attr layouts" }, 112 { DMU_BSWAP_ZAP, TRUE, "scan translations" }, 113 { DMU_BSWAP_UINT8, FALSE, "deduplicated block" }, 114 { DMU_BSWAP_ZAP, TRUE, "DSL deadlist map" }, 115 { DMU_BSWAP_UINT64, TRUE, "DSL deadlist map hdr" }, 116 { DMU_BSWAP_ZAP, TRUE, "DSL dir clones" }, 117 { DMU_BSWAP_UINT64, TRUE, "bpobj subobj" } 118}; 119 120const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = { 121 { byteswap_uint8_array, "uint8" }, 122 { byteswap_uint16_array, "uint16" }, 123 { byteswap_uint32_array, "uint32" }, 124 { byteswap_uint64_array, "uint64" }, 125 { zap_byteswap, "zap" }, 126 { dnode_buf_byteswap, "dnode" }, 127 { dmu_objset_byteswap, "objset" }, 128 { zfs_znode_byteswap, "znode" }, 129 { zfs_oldacl_byteswap, "oldacl" }, 130 { zfs_acl_byteswap, "acl" } 131}; 132 133int 134dmu_buf_hold_noread(objset_t *os, uint64_t object, uint64_t offset, 135 void *tag, dmu_buf_t **dbp) 136{ 137 dnode_t *dn; 138 uint64_t blkid; 139 dmu_buf_impl_t *db; 140 int err; 141 142 err = dnode_hold(os, object, FTAG, &dn); 143 if (err) 144 return (err); 145 blkid = dbuf_whichblock(dn, 0, offset); 146 rw_enter(&dn->dn_struct_rwlock, RW_READER); 147 db = dbuf_hold(dn, blkid, tag); 148 rw_exit(&dn->dn_struct_rwlock); 149 dnode_rele(dn, FTAG); 150 151 if (db == NULL) { 152 *dbp = NULL; 153 return (SET_ERROR(EIO)); 154 } 155 156 *dbp = &db->db; 157 return (err); 158} 159 160int 161dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset, 162 void *tag, dmu_buf_t **dbp, int flags) 163{ 164 int err; 165 int db_flags = DB_RF_CANFAIL; 166 167 if (flags & DMU_READ_NO_PREFETCH) 168 db_flags |= DB_RF_NOPREFETCH; 169 170 err = dmu_buf_hold_noread(os, object, offset, tag, dbp); 171 if (err == 0) { 172 dmu_buf_impl_t *db = (dmu_buf_impl_t *)(*dbp); 173 err = dbuf_read(db, NULL, db_flags); 174 if (err != 0) { 175 dbuf_rele(db, tag); 176 *dbp = NULL; 177 } 178 } 179 180 return (err); 181} 182 183int 184dmu_bonus_max(void) 185{ 186 return (DN_MAX_BONUSLEN); 187} 188 189int 190dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx) 191{ 192 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 193 dnode_t *dn; 194 int error; 195 196 DB_DNODE_ENTER(db); 197 dn = DB_DNODE(db); 198 199 if (dn->dn_bonus != db) { 200 error = SET_ERROR(EINVAL); 201 } else if (newsize < 0 || newsize > db_fake->db_size) { 202 error = SET_ERROR(EINVAL); 203 } else { 204 dnode_setbonuslen(dn, newsize, tx); 205 error = 0; 206 } 207 208 DB_DNODE_EXIT(db); 209 return (error); 210} 211 212int 213dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx) 214{ 215 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 216 dnode_t *dn; 217 int error; 218 219 DB_DNODE_ENTER(db); 220 dn = DB_DNODE(db); 221 222 if (!DMU_OT_IS_VALID(type)) { 223 error = SET_ERROR(EINVAL); 224 } else if (dn->dn_bonus != db) { 225 error = SET_ERROR(EINVAL); 226 } else { 227 dnode_setbonus_type(dn, type, tx); 228 error = 0; 229 } 230 231 DB_DNODE_EXIT(db); 232 return (error); 233} 234 235dmu_object_type_t 236dmu_get_bonustype(dmu_buf_t *db_fake) 237{ 238 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 239 dnode_t *dn; 240 dmu_object_type_t type; 241 242 DB_DNODE_ENTER(db); 243 dn = DB_DNODE(db); 244 type = dn->dn_bonustype; 245 DB_DNODE_EXIT(db); 246 247 return (type); 248} 249 250int 251dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx) 252{ 253 dnode_t *dn; 254 int error; 255 256 error = dnode_hold(os, object, FTAG, &dn); 257 dbuf_rm_spill(dn, tx); 258 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 259 dnode_rm_spill(dn, tx); 260 rw_exit(&dn->dn_struct_rwlock); 261 dnode_rele(dn, FTAG); 262 return (error); 263} 264 265/* 266 * returns ENOENT, EIO, or 0. 267 */ 268int 269dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp) 270{ 271 dnode_t *dn; 272 dmu_buf_impl_t *db; 273 int error; 274 275 error = dnode_hold(os, object, FTAG, &dn); 276 if (error) 277 return (error); 278 279 rw_enter(&dn->dn_struct_rwlock, RW_READER); 280 if (dn->dn_bonus == NULL) { 281 rw_exit(&dn->dn_struct_rwlock); 282 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 283 if (dn->dn_bonus == NULL) 284 dbuf_create_bonus(dn); 285 } 286 db = dn->dn_bonus; 287 288 /* as long as the bonus buf is held, the dnode will be held */ 289 if (refcount_add(&db->db_holds, tag) == 1) { 290 VERIFY(dnode_add_ref(dn, db)); 291 atomic_inc_32(&dn->dn_dbufs_count); 292 } 293 294 /* 295 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's 296 * hold and incrementing the dbuf count to ensure that dnode_move() sees 297 * a dnode hold for every dbuf. 298 */ 299 rw_exit(&dn->dn_struct_rwlock); 300 301 dnode_rele(dn, FTAG); 302 303 VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH)); 304 305 *dbp = &db->db; 306 return (0); 307} 308 309/* 310 * returns ENOENT, EIO, or 0. 311 * 312 * This interface will allocate a blank spill dbuf when a spill blk 313 * doesn't already exist on the dnode. 314 * 315 * if you only want to find an already existing spill db, then 316 * dmu_spill_hold_existing() should be used. 317 */ 318int 319dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp) 320{ 321 dmu_buf_impl_t *db = NULL; 322 int err; 323 324 if ((flags & DB_RF_HAVESTRUCT) == 0) 325 rw_enter(&dn->dn_struct_rwlock, RW_READER); 326 327 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag); 328 329 if ((flags & DB_RF_HAVESTRUCT) == 0) 330 rw_exit(&dn->dn_struct_rwlock); 331 332 ASSERT(db != NULL); 333 err = dbuf_read(db, NULL, flags); 334 if (err == 0) 335 *dbp = &db->db; 336 else 337 dbuf_rele(db, tag); 338 return (err); 339} 340 341int 342dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp) 343{ 344 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus; 345 dnode_t *dn; 346 int err; 347 348 DB_DNODE_ENTER(db); 349 dn = DB_DNODE(db); 350 351 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) { 352 err = SET_ERROR(EINVAL); 353 } else { 354 rw_enter(&dn->dn_struct_rwlock, RW_READER); 355 356 if (!dn->dn_have_spill) { 357 err = SET_ERROR(ENOENT); 358 } else { 359 err = dmu_spill_hold_by_dnode(dn, 360 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp); 361 } 362 363 rw_exit(&dn->dn_struct_rwlock); 364 } 365 366 DB_DNODE_EXIT(db); 367 return (err); 368} 369 370int 371dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp) 372{ 373 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus; 374 dnode_t *dn; 375 int err; 376 377 DB_DNODE_ENTER(db); 378 dn = DB_DNODE(db); 379 err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp); 380 DB_DNODE_EXIT(db); 381 382 return (err); 383} 384 385/* 386 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces 387 * to take a held dnode rather than <os, object> -- the lookup is wasteful, 388 * and can induce severe lock contention when writing to several files 389 * whose dnodes are in the same block. 390 */ 391static int 392dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length, 393 boolean_t read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags) 394{ 395 dmu_buf_t **dbp; 396 uint64_t blkid, nblks, i; 397 uint32_t dbuf_flags; 398 int err; 399 zio_t *zio; 400 401 ASSERT(length <= DMU_MAX_ACCESS); 402 403 /* 404 * Note: We directly notify the prefetch code of this read, so that 405 * we can tell it about the multi-block read. dbuf_read() only knows 406 * about the one block it is accessing. 407 */ 408 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT | 409 DB_RF_NOPREFETCH; 410 411 rw_enter(&dn->dn_struct_rwlock, RW_READER); 412 if (dn->dn_datablkshift) { 413 int blkshift = dn->dn_datablkshift; 414 nblks = (P2ROUNDUP(offset + length, 1ULL << blkshift) - 415 P2ALIGN(offset, 1ULL << blkshift)) >> blkshift; 416 } else { 417 if (offset + length > dn->dn_datablksz) { 418 zfs_panic_recover("zfs: accessing past end of object " 419 "%llx/%llx (size=%u access=%llu+%llu)", 420 (longlong_t)dn->dn_objset-> 421 os_dsl_dataset->ds_object, 422 (longlong_t)dn->dn_object, dn->dn_datablksz, 423 (longlong_t)offset, (longlong_t)length); 424 rw_exit(&dn->dn_struct_rwlock); 425 return (SET_ERROR(EIO)); 426 } 427 nblks = 1; 428 } 429 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP); 430 431#if defined(_KERNEL) && defined(RACCT) 432 if (racct_enable && !read) { 433 PROC_LOCK(curproc); 434 racct_add_force(curproc, RACCT_WRITEBPS, length); 435 racct_add_force(curproc, RACCT_WRITEIOPS, nblks); 436 PROC_UNLOCK(curproc); 437 } 438#endif 439 440 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL); 441 blkid = dbuf_whichblock(dn, 0, offset); 442 for (i = 0; i < nblks; i++) { 443 dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag); 444 if (db == NULL) { 445 rw_exit(&dn->dn_struct_rwlock); 446 dmu_buf_rele_array(dbp, nblks, tag); 447 zio_nowait(zio); 448 return (SET_ERROR(EIO)); 449 } 450 451 /* initiate async i/o */ 452 if (read) 453 (void) dbuf_read(db, zio, dbuf_flags); 454#ifdef _KERNEL 455 else 456 curthread->td_ru.ru_oublock++; 457#endif 458 dbp[i] = &db->db; 459 } 460 461 if ((flags & DMU_READ_NO_PREFETCH) == 0 && 462 DNODE_META_IS_CACHEABLE(dn) && length <= zfetch_array_rd_sz) { 463 dmu_zfetch(&dn->dn_zfetch, blkid, nblks, 464 read && DNODE_IS_CACHEABLE(dn)); 465 } 466 rw_exit(&dn->dn_struct_rwlock); 467 468 /* wait for async i/o */ 469 err = zio_wait(zio); 470 if (err) { 471 dmu_buf_rele_array(dbp, nblks, tag); 472 return (err); 473 } 474 475 /* wait for other io to complete */ 476 if (read) { 477 for (i = 0; i < nblks; i++) { 478 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i]; 479 mutex_enter(&db->db_mtx); 480 while (db->db_state == DB_READ || 481 db->db_state == DB_FILL) 482 cv_wait(&db->db_changed, &db->db_mtx); 483 if (db->db_state == DB_UNCACHED) 484 err = SET_ERROR(EIO); 485 mutex_exit(&db->db_mtx); 486 if (err) { 487 dmu_buf_rele_array(dbp, nblks, tag); 488 return (err); 489 } 490 } 491 } 492 493 *numbufsp = nblks; 494 *dbpp = dbp; 495 return (0); 496} 497 498static int 499dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset, 500 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp) 501{ 502 dnode_t *dn; 503 int err; 504 505 err = dnode_hold(os, object, FTAG, &dn); 506 if (err) 507 return (err); 508 509 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag, 510 numbufsp, dbpp, DMU_READ_PREFETCH); 511 512 dnode_rele(dn, FTAG); 513 514 return (err); 515} 516 517int 518dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset, 519 uint64_t length, boolean_t read, void *tag, int *numbufsp, 520 dmu_buf_t ***dbpp) 521{ 522 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 523 dnode_t *dn; 524 int err; 525 526 DB_DNODE_ENTER(db); 527 dn = DB_DNODE(db); 528 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag, 529 numbufsp, dbpp, DMU_READ_PREFETCH); 530 DB_DNODE_EXIT(db); 531 532 return (err); 533} 534 535void 536dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag) 537{ 538 int i; 539 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake; 540 541 if (numbufs == 0) 542 return; 543 544 for (i = 0; i < numbufs; i++) { 545 if (dbp[i]) 546 dbuf_rele(dbp[i], tag); 547 } 548 549 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs); 550} 551 552/* 553 * Issue prefetch i/os for the given blocks. If level is greater than 0, the 554 * indirect blocks prefeteched will be those that point to the blocks containing 555 * the data starting at offset, and continuing to offset + len. 556 * 557 * Note that if the indirect blocks above the blocks being prefetched are not in 558 * cache, they will be asychronously read in. 559 */ 560void 561dmu_prefetch(objset_t *os, uint64_t object, int64_t level, uint64_t offset, 562 uint64_t len, zio_priority_t pri) 563{ 564 dnode_t *dn; 565 uint64_t blkid; 566 int nblks, err; 567 568 if (len == 0) { /* they're interested in the bonus buffer */ 569 dn = DMU_META_DNODE(os); 570 571 if (object == 0 || object >= DN_MAX_OBJECT) 572 return; 573 574 rw_enter(&dn->dn_struct_rwlock, RW_READER); 575 blkid = dbuf_whichblock(dn, level, 576 object * sizeof (dnode_phys_t)); 577 dbuf_prefetch(dn, level, blkid, pri, 0); 578 rw_exit(&dn->dn_struct_rwlock); 579 return; 580 } 581 582 /* 583 * XXX - Note, if the dnode for the requested object is not 584 * already cached, we will do a *synchronous* read in the 585 * dnode_hold() call. The same is true for any indirects. 586 */ 587 err = dnode_hold(os, object, FTAG, &dn); 588 if (err != 0) 589 return; 590 591 rw_enter(&dn->dn_struct_rwlock, RW_READER); 592 /* 593 * offset + len - 1 is the last byte we want to prefetch for, and offset 594 * is the first. Then dbuf_whichblk(dn, level, off + len - 1) is the 595 * last block we want to prefetch, and dbuf_whichblock(dn, level, 596 * offset) is the first. Then the number we need to prefetch is the 597 * last - first + 1. 598 */ 599 if (level > 0 || dn->dn_datablkshift != 0) { 600 nblks = dbuf_whichblock(dn, level, offset + len - 1) - 601 dbuf_whichblock(dn, level, offset) + 1; 602 } else { 603 nblks = (offset < dn->dn_datablksz); 604 } 605 606 if (nblks != 0) { 607 blkid = dbuf_whichblock(dn, level, offset); 608 for (int i = 0; i < nblks; i++) 609 dbuf_prefetch(dn, level, blkid + i, pri, 0); 610 } 611 612 rw_exit(&dn->dn_struct_rwlock); 613 614 dnode_rele(dn, FTAG); 615} 616 617/* 618 * Get the next "chunk" of file data to free. We traverse the file from 619 * the end so that the file gets shorter over time (if we crashes in the 620 * middle, this will leave us in a better state). We find allocated file 621 * data by simply searching the allocated level 1 indirects. 622 * 623 * On input, *start should be the first offset that does not need to be 624 * freed (e.g. "offset + length"). On return, *start will be the first 625 * offset that should be freed. 626 */ 627static int 628get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum) 629{ 630 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1); 631 /* bytes of data covered by a level-1 indirect block */ 632 uint64_t iblkrange = 633 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT); 634 635 ASSERT3U(minimum, <=, *start); 636 637 if (*start - minimum <= iblkrange * maxblks) { 638 *start = minimum; 639 return (0); 640 } 641 ASSERT(ISP2(iblkrange)); 642 643 for (uint64_t blks = 0; *start > minimum && blks < maxblks; blks++) { 644 int err; 645 646 /* 647 * dnode_next_offset(BACKWARDS) will find an allocated L1 648 * indirect block at or before the input offset. We must 649 * decrement *start so that it is at the end of the region 650 * to search. 651 */ 652 (*start)--; 653 err = dnode_next_offset(dn, 654 DNODE_FIND_BACKWARDS, start, 2, 1, 0); 655 656 /* if there are no indirect blocks before start, we are done */ 657 if (err == ESRCH) { 658 *start = minimum; 659 break; 660 } else if (err != 0) { 661 return (err); 662 } 663 664 /* set start to the beginning of this L1 indirect */ 665 *start = P2ALIGN(*start, iblkrange); 666 } 667 if (*start < minimum) 668 *start = minimum; 669 return (0); 670} 671 672static int 673dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset, 674 uint64_t length) 675{ 676 uint64_t object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz; 677 int err; 678 679 if (offset >= object_size) 680 return (0); 681 682 if (length == DMU_OBJECT_END || offset + length > object_size) 683 length = object_size - offset; 684 685 while (length != 0) { 686 uint64_t chunk_end, chunk_begin; 687 688 chunk_end = chunk_begin = offset + length; 689 690 /* move chunk_begin backwards to the beginning of this chunk */ 691 err = get_next_chunk(dn, &chunk_begin, offset); 692 if (err) 693 return (err); 694 ASSERT3U(chunk_begin, >=, offset); 695 ASSERT3U(chunk_begin, <=, chunk_end); 696 697 dmu_tx_t *tx = dmu_tx_create(os); 698 dmu_tx_hold_free(tx, dn->dn_object, 699 chunk_begin, chunk_end - chunk_begin); 700 701 /* 702 * Mark this transaction as typically resulting in a net 703 * reduction in space used. 704 */ 705 dmu_tx_mark_netfree(tx); 706 err = dmu_tx_assign(tx, TXG_WAIT); 707 if (err) { 708 dmu_tx_abort(tx); 709 return (err); 710 } 711 dnode_free_range(dn, chunk_begin, chunk_end - chunk_begin, tx); 712 dmu_tx_commit(tx); 713 714 length -= chunk_end - chunk_begin; 715 } 716 return (0); 717} 718 719int 720dmu_free_long_range(objset_t *os, uint64_t object, 721 uint64_t offset, uint64_t length) 722{ 723 dnode_t *dn; 724 int err; 725 726 err = dnode_hold(os, object, FTAG, &dn); 727 if (err != 0) 728 return (err); 729 err = dmu_free_long_range_impl(os, dn, offset, length); 730 731 /* 732 * It is important to zero out the maxblkid when freeing the entire 733 * file, so that (a) subsequent calls to dmu_free_long_range_impl() 734 * will take the fast path, and (b) dnode_reallocate() can verify 735 * that the entire file has been freed. 736 */ 737 if (err == 0 && offset == 0 && length == DMU_OBJECT_END) 738 dn->dn_maxblkid = 0; 739 740 dnode_rele(dn, FTAG); 741 return (err); 742} 743 744int 745dmu_free_long_object(objset_t *os, uint64_t object) 746{ 747 dmu_tx_t *tx; 748 int err; 749 750 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END); 751 if (err != 0) 752 return (err); 753 754 tx = dmu_tx_create(os); 755 dmu_tx_hold_bonus(tx, object); 756 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END); 757 dmu_tx_mark_netfree(tx); 758 err = dmu_tx_assign(tx, TXG_WAIT); 759 if (err == 0) { 760 err = dmu_object_free(os, object, tx); 761 dmu_tx_commit(tx); 762 } else { 763 dmu_tx_abort(tx); 764 } 765 766 return (err); 767} 768 769int 770dmu_free_range(objset_t *os, uint64_t object, uint64_t offset, 771 uint64_t size, dmu_tx_t *tx) 772{ 773 dnode_t *dn; 774 int err = dnode_hold(os, object, FTAG, &dn); 775 if (err) 776 return (err); 777 ASSERT(offset < UINT64_MAX); 778 ASSERT(size == -1ULL || size <= UINT64_MAX - offset); 779 dnode_free_range(dn, offset, size, tx); 780 dnode_rele(dn, FTAG); 781 return (0); 782} 783 784int 785dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 786 void *buf, uint32_t flags) 787{ 788 dnode_t *dn; 789 dmu_buf_t **dbp; 790 int numbufs, err; 791 792 err = dnode_hold(os, object, FTAG, &dn); 793 if (err) 794 return (err); 795 796 /* 797 * Deal with odd block sizes, where there can't be data past the first 798 * block. If we ever do the tail block optimization, we will need to 799 * handle that here as well. 800 */ 801 if (dn->dn_maxblkid == 0) { 802 int newsz = offset > dn->dn_datablksz ? 0 : 803 MIN(size, dn->dn_datablksz - offset); 804 bzero((char *)buf + newsz, size - newsz); 805 size = newsz; 806 } 807 808 while (size > 0) { 809 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2); 810 int i; 811 812 /* 813 * NB: we could do this block-at-a-time, but it's nice 814 * to be reading in parallel. 815 */ 816 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen, 817 TRUE, FTAG, &numbufs, &dbp, flags); 818 if (err) 819 break; 820 821 for (i = 0; i < numbufs; i++) { 822 int tocpy; 823 int bufoff; 824 dmu_buf_t *db = dbp[i]; 825 826 ASSERT(size > 0); 827 828 bufoff = offset - db->db_offset; 829 tocpy = (int)MIN(db->db_size - bufoff, size); 830 831 bcopy((char *)db->db_data + bufoff, buf, tocpy); 832 833 offset += tocpy; 834 size -= tocpy; 835 buf = (char *)buf + tocpy; 836 } 837 dmu_buf_rele_array(dbp, numbufs, FTAG); 838 } 839 dnode_rele(dn, FTAG); 840 return (err); 841} 842 843void 844dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 845 const void *buf, dmu_tx_t *tx) 846{ 847 dmu_buf_t **dbp; 848 int numbufs, i; 849 850 if (size == 0) 851 return; 852 853 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size, 854 FALSE, FTAG, &numbufs, &dbp)); 855 856 for (i = 0; i < numbufs; i++) { 857 int tocpy; 858 int bufoff; 859 dmu_buf_t *db = dbp[i]; 860 861 ASSERT(size > 0); 862 863 bufoff = offset - db->db_offset; 864 tocpy = (int)MIN(db->db_size - bufoff, size); 865 866 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 867 868 if (tocpy == db->db_size) 869 dmu_buf_will_fill(db, tx); 870 else 871 dmu_buf_will_dirty(db, tx); 872 873 bcopy(buf, (char *)db->db_data + bufoff, tocpy); 874 875 if (tocpy == db->db_size) 876 dmu_buf_fill_done(db, tx); 877 878 offset += tocpy; 879 size -= tocpy; 880 buf = (char *)buf + tocpy; 881 } 882 dmu_buf_rele_array(dbp, numbufs, FTAG); 883} 884 885void 886dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 887 dmu_tx_t *tx) 888{ 889 dmu_buf_t **dbp; 890 int numbufs, i; 891 892 if (size == 0) 893 return; 894 895 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size, 896 FALSE, FTAG, &numbufs, &dbp)); 897 898 for (i = 0; i < numbufs; i++) { 899 dmu_buf_t *db = dbp[i]; 900 901 dmu_buf_will_not_fill(db, tx); 902 } 903 dmu_buf_rele_array(dbp, numbufs, FTAG); 904} 905 906void 907dmu_write_embedded(objset_t *os, uint64_t object, uint64_t offset, 908 void *data, uint8_t etype, uint8_t comp, int uncompressed_size, 909 int compressed_size, int byteorder, dmu_tx_t *tx) 910{ 911 dmu_buf_t *db; 912 913 ASSERT3U(etype, <, NUM_BP_EMBEDDED_TYPES); 914 ASSERT3U(comp, <, ZIO_COMPRESS_FUNCTIONS); 915 VERIFY0(dmu_buf_hold_noread(os, object, offset, 916 FTAG, &db)); 917 918 dmu_buf_write_embedded(db, 919 data, (bp_embedded_type_t)etype, (enum zio_compress)comp, 920 uncompressed_size, compressed_size, byteorder, tx); 921 922 dmu_buf_rele(db, FTAG); 923} 924 925/* 926 * DMU support for xuio 927 */ 928kstat_t *xuio_ksp = NULL; 929 930int 931dmu_xuio_init(xuio_t *xuio, int nblk) 932{ 933 dmu_xuio_t *priv; 934 uio_t *uio = &xuio->xu_uio; 935 936 uio->uio_iovcnt = nblk; 937 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP); 938 939 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP); 940 priv->cnt = nblk; 941 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP); 942 priv->iovp = uio->uio_iov; 943 XUIO_XUZC_PRIV(xuio) = priv; 944 945 if (XUIO_XUZC_RW(xuio) == UIO_READ) 946 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk); 947 else 948 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk); 949 950 return (0); 951} 952 953void 954dmu_xuio_fini(xuio_t *xuio) 955{ 956 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 957 int nblk = priv->cnt; 958 959 kmem_free(priv->iovp, nblk * sizeof (iovec_t)); 960 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *)); 961 kmem_free(priv, sizeof (dmu_xuio_t)); 962 963 if (XUIO_XUZC_RW(xuio) == UIO_READ) 964 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk); 965 else 966 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk); 967} 968 969/* 970 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf } 971 * and increase priv->next by 1. 972 */ 973int 974dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n) 975{ 976 struct iovec *iov; 977 uio_t *uio = &xuio->xu_uio; 978 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 979 int i = priv->next++; 980 981 ASSERT(i < priv->cnt); 982 ASSERT(off + n <= arc_buf_size(abuf)); 983 iov = uio->uio_iov + i; 984 iov->iov_base = (char *)abuf->b_data + off; 985 iov->iov_len = n; 986 priv->bufs[i] = abuf; 987 return (0); 988} 989 990int 991dmu_xuio_cnt(xuio_t *xuio) 992{ 993 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 994 return (priv->cnt); 995} 996 997arc_buf_t * 998dmu_xuio_arcbuf(xuio_t *xuio, int i) 999{ 1000 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 1001 1002 ASSERT(i < priv->cnt); 1003 return (priv->bufs[i]); 1004} 1005 1006void 1007dmu_xuio_clear(xuio_t *xuio, int i) 1008{ 1009 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 1010 1011 ASSERT(i < priv->cnt); 1012 priv->bufs[i] = NULL; 1013} 1014 1015static void 1016xuio_stat_init(void) 1017{ 1018 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc", 1019 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t), 1020 KSTAT_FLAG_VIRTUAL); 1021 if (xuio_ksp != NULL) { 1022 xuio_ksp->ks_data = &xuio_stats; 1023 kstat_install(xuio_ksp); 1024 } 1025} 1026 1027static void 1028xuio_stat_fini(void) 1029{ 1030 if (xuio_ksp != NULL) { 1031 kstat_delete(xuio_ksp); 1032 xuio_ksp = NULL; 1033 } 1034} 1035 1036void 1037xuio_stat_wbuf_copied() 1038{ 1039 XUIOSTAT_BUMP(xuiostat_wbuf_copied); 1040} 1041 1042void 1043xuio_stat_wbuf_nocopy() 1044{ 1045 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy); 1046} 1047 1048#ifdef _KERNEL 1049static int 1050dmu_read_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size) 1051{ 1052 dmu_buf_t **dbp; 1053 int numbufs, i, err; 1054 xuio_t *xuio = NULL; 1055 1056 /* 1057 * NB: we could do this block-at-a-time, but it's nice 1058 * to be reading in parallel. 1059 */ 1060 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size, 1061 TRUE, FTAG, &numbufs, &dbp, 0); 1062 if (err) 1063 return (err); 1064 1065#ifdef UIO_XUIO 1066 if (uio->uio_extflg == UIO_XUIO) 1067 xuio = (xuio_t *)uio; 1068#endif 1069 1070 for (i = 0; i < numbufs; i++) { 1071 int tocpy; 1072 int bufoff; 1073 dmu_buf_t *db = dbp[i]; 1074 1075 ASSERT(size > 0); 1076 1077 bufoff = uio->uio_loffset - db->db_offset; 1078 tocpy = (int)MIN(db->db_size - bufoff, size); 1079 1080 if (xuio) { 1081 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 1082 arc_buf_t *dbuf_abuf = dbi->db_buf; 1083 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi); 1084 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy); 1085 if (!err) { 1086 uio->uio_resid -= tocpy; 1087 uio->uio_loffset += tocpy; 1088 } 1089 1090 if (abuf == dbuf_abuf) 1091 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy); 1092 else 1093 XUIOSTAT_BUMP(xuiostat_rbuf_copied); 1094 } else { 1095#ifdef illumos 1096 err = uiomove((char *)db->db_data + bufoff, tocpy, 1097 UIO_READ, uio); 1098#else 1099 err = vn_io_fault_uiomove((char *)db->db_data + bufoff, 1100 tocpy, uio); 1101#endif 1102 } 1103 if (err) 1104 break; 1105 1106 size -= tocpy; 1107 } 1108 dmu_buf_rele_array(dbp, numbufs, FTAG); 1109 1110 return (err); 1111} 1112 1113/* 1114 * Read 'size' bytes into the uio buffer. 1115 * From object zdb->db_object. 1116 * Starting at offset uio->uio_loffset. 1117 * 1118 * If the caller already has a dbuf in the target object 1119 * (e.g. its bonus buffer), this routine is faster than dmu_read_uio(), 1120 * because we don't have to find the dnode_t for the object. 1121 */ 1122int 1123dmu_read_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size) 1124{ 1125 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb; 1126 dnode_t *dn; 1127 int err; 1128 1129 if (size == 0) 1130 return (0); 1131 1132 DB_DNODE_ENTER(db); 1133 dn = DB_DNODE(db); 1134 err = dmu_read_uio_dnode(dn, uio, size); 1135 DB_DNODE_EXIT(db); 1136 1137 return (err); 1138} 1139 1140/* 1141 * Read 'size' bytes into the uio buffer. 1142 * From the specified object 1143 * Starting at offset uio->uio_loffset. 1144 */ 1145int 1146dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size) 1147{ 1148 dnode_t *dn; 1149 int err; 1150 1151 if (size == 0) 1152 return (0); 1153 1154 err = dnode_hold(os, object, FTAG, &dn); 1155 if (err) 1156 return (err); 1157 1158 err = dmu_read_uio_dnode(dn, uio, size); 1159 1160 dnode_rele(dn, FTAG); 1161 1162 return (err); 1163} 1164 1165static int 1166dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx) 1167{ 1168 dmu_buf_t **dbp; 1169 int numbufs; 1170 int err = 0; 1171 int i; 1172 1173 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size, 1174 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH); 1175 if (err) 1176 return (err); 1177 1178 for (i = 0; i < numbufs; i++) { 1179 int tocpy; 1180 int bufoff; 1181 dmu_buf_t *db = dbp[i]; 1182 1183 ASSERT(size > 0); 1184 1185 bufoff = uio->uio_loffset - db->db_offset; 1186 tocpy = (int)MIN(db->db_size - bufoff, size); 1187 1188 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 1189 1190 if (tocpy == db->db_size) 1191 dmu_buf_will_fill(db, tx); 1192 else 1193 dmu_buf_will_dirty(db, tx); 1194 1195#ifdef illumos 1196 /* 1197 * XXX uiomove could block forever (eg. nfs-backed 1198 * pages). There needs to be a uiolockdown() function 1199 * to lock the pages in memory, so that uiomove won't 1200 * block. 1201 */ 1202 err = uiomove((char *)db->db_data + bufoff, tocpy, 1203 UIO_WRITE, uio); 1204#else 1205 err = vn_io_fault_uiomove((char *)db->db_data + bufoff, tocpy, 1206 uio); 1207#endif 1208 1209 if (tocpy == db->db_size) 1210 dmu_buf_fill_done(db, tx); 1211 1212 if (err) 1213 break; 1214 1215 size -= tocpy; 1216 } 1217 1218 dmu_buf_rele_array(dbp, numbufs, FTAG); 1219 return (err); 1220} 1221 1222/* 1223 * Write 'size' bytes from the uio buffer. 1224 * To object zdb->db_object. 1225 * Starting at offset uio->uio_loffset. 1226 * 1227 * If the caller already has a dbuf in the target object 1228 * (e.g. its bonus buffer), this routine is faster than dmu_write_uio(), 1229 * because we don't have to find the dnode_t for the object. 1230 */ 1231int 1232dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size, 1233 dmu_tx_t *tx) 1234{ 1235 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb; 1236 dnode_t *dn; 1237 int err; 1238 1239 if (size == 0) 1240 return (0); 1241 1242 DB_DNODE_ENTER(db); 1243 dn = DB_DNODE(db); 1244 err = dmu_write_uio_dnode(dn, uio, size, tx); 1245 DB_DNODE_EXIT(db); 1246 1247 return (err); 1248} 1249 1250/* 1251 * Write 'size' bytes from the uio buffer. 1252 * To the specified object. 1253 * Starting at offset uio->uio_loffset. 1254 */ 1255int 1256dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size, 1257 dmu_tx_t *tx) 1258{ 1259 dnode_t *dn; 1260 int err; 1261 1262 if (size == 0) 1263 return (0); 1264 1265 err = dnode_hold(os, object, FTAG, &dn); 1266 if (err) 1267 return (err); 1268 1269 err = dmu_write_uio_dnode(dn, uio, size, tx); 1270 1271 dnode_rele(dn, FTAG); 1272 1273 return (err); 1274} 1275 1276#ifdef illumos 1277int 1278dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1279 page_t *pp, dmu_tx_t *tx) 1280{ 1281 dmu_buf_t **dbp; 1282 int numbufs, i; 1283 int err; 1284 1285 if (size == 0) 1286 return (0); 1287 1288 err = dmu_buf_hold_array(os, object, offset, size, 1289 FALSE, FTAG, &numbufs, &dbp); 1290 if (err) 1291 return (err); 1292 1293 for (i = 0; i < numbufs; i++) { 1294 int tocpy, copied, thiscpy; 1295 int bufoff; 1296 dmu_buf_t *db = dbp[i]; 1297 caddr_t va; 1298 1299 ASSERT(size > 0); 1300 ASSERT3U(db->db_size, >=, PAGESIZE); 1301 1302 bufoff = offset - db->db_offset; 1303 tocpy = (int)MIN(db->db_size - bufoff, size); 1304 1305 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 1306 1307 if (tocpy == db->db_size) 1308 dmu_buf_will_fill(db, tx); 1309 else 1310 dmu_buf_will_dirty(db, tx); 1311 1312 for (copied = 0; copied < tocpy; copied += PAGESIZE) { 1313 ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff); 1314 thiscpy = MIN(PAGESIZE, tocpy - copied); 1315 va = zfs_map_page(pp, S_READ); 1316 bcopy(va, (char *)db->db_data + bufoff, thiscpy); 1317 zfs_unmap_page(pp, va); 1318 pp = pp->p_next; 1319 bufoff += PAGESIZE; 1320 } 1321 1322 if (tocpy == db->db_size) 1323 dmu_buf_fill_done(db, tx); 1324 1325 offset += tocpy; 1326 size -= tocpy; 1327 } 1328 dmu_buf_rele_array(dbp, numbufs, FTAG); 1329 return (err); 1330} 1331 1332#else /* !illumos */ 1333 1334int 1335dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1336 vm_page_t *ma, dmu_tx_t *tx) 1337{ 1338 dmu_buf_t **dbp; 1339 struct sf_buf *sf; 1340 int numbufs, i; 1341 int err; 1342 1343 if (size == 0) 1344 return (0); 1345 1346 err = dmu_buf_hold_array(os, object, offset, size, 1347 FALSE, FTAG, &numbufs, &dbp); 1348 if (err) 1349 return (err); 1350 1351 for (i = 0; i < numbufs; i++) { 1352 int tocpy, copied, thiscpy; 1353 int bufoff; 1354 dmu_buf_t *db = dbp[i]; 1355 caddr_t va; 1356 1357 ASSERT(size > 0); 1358 ASSERT3U(db->db_size, >=, PAGESIZE); 1359 1360 bufoff = offset - db->db_offset; 1361 tocpy = (int)MIN(db->db_size - bufoff, size); 1362 1363 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 1364 1365 if (tocpy == db->db_size) 1366 dmu_buf_will_fill(db, tx); 1367 else 1368 dmu_buf_will_dirty(db, tx); 1369 1370 for (copied = 0; copied < tocpy; copied += PAGESIZE) { 1371 ASSERT3U(ptoa((*ma)->pindex), ==, db->db_offset + bufoff); 1372 thiscpy = MIN(PAGESIZE, tocpy - copied); 1373 va = zfs_map_page(*ma, &sf); 1374 bcopy(va, (char *)db->db_data + bufoff, thiscpy); 1375 zfs_unmap_page(sf); 1376 ma += 1; 1377 bufoff += PAGESIZE; 1378 } 1379 1380 if (tocpy == db->db_size) 1381 dmu_buf_fill_done(db, tx); 1382 1383 offset += tocpy; 1384 size -= tocpy; 1385 } 1386 dmu_buf_rele_array(dbp, numbufs, FTAG); 1387 return (err); 1388} 1389#endif /* illumos */ 1390#endif /* _KERNEL */ 1391 1392/* 1393 * Allocate a loaned anonymous arc buffer. 1394 */ 1395arc_buf_t * 1396dmu_request_arcbuf(dmu_buf_t *handle, int size) 1397{ 1398 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle; 1399 1400 return (arc_loan_buf(db->db_objset->os_spa, size)); 1401} 1402 1403/* 1404 * Free a loaned arc buffer. 1405 */ 1406void 1407dmu_return_arcbuf(arc_buf_t *buf) 1408{ 1409 arc_return_buf(buf, FTAG); 1410 VERIFY(arc_buf_remove_ref(buf, FTAG)); 1411} 1412 1413/* 1414 * When possible directly assign passed loaned arc buffer to a dbuf. 1415 * If this is not possible copy the contents of passed arc buf via 1416 * dmu_write(). 1417 */ 1418void 1419dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf, 1420 dmu_tx_t *tx) 1421{ 1422 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle; 1423 dnode_t *dn; 1424 dmu_buf_impl_t *db; 1425 uint32_t blksz = (uint32_t)arc_buf_size(buf); 1426 uint64_t blkid; 1427 1428 DB_DNODE_ENTER(dbuf); 1429 dn = DB_DNODE(dbuf); 1430 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1431 blkid = dbuf_whichblock(dn, 0, offset); 1432 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL); 1433 rw_exit(&dn->dn_struct_rwlock); 1434 DB_DNODE_EXIT(dbuf); 1435 1436 /* 1437 * We can only assign if the offset is aligned, the arc buf is the 1438 * same size as the dbuf, and the dbuf is not metadata. It 1439 * can't be metadata because the loaned arc buf comes from the 1440 * user-data kmem arena. 1441 */ 1442 if (offset == db->db.db_offset && blksz == db->db.db_size && 1443 DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA) { 1444#ifdef _KERNEL 1445 curthread->td_ru.ru_oublock++; 1446#ifdef RACCT 1447 if (racct_enable) { 1448 PROC_LOCK(curproc); 1449 racct_add_force(curproc, RACCT_WRITEBPS, blksz); 1450 racct_add_force(curproc, RACCT_WRITEIOPS, 1); 1451 PROC_UNLOCK(curproc); 1452 } 1453#endif /* RACCT */ 1454#endif /* _KERNEL */ 1455 dbuf_assign_arcbuf(db, buf, tx); 1456 dbuf_rele(db, FTAG); 1457 } else { 1458 objset_t *os; 1459 uint64_t object; 1460 1461 DB_DNODE_ENTER(dbuf); 1462 dn = DB_DNODE(dbuf); 1463 os = dn->dn_objset; 1464 object = dn->dn_object; 1465 DB_DNODE_EXIT(dbuf); 1466 1467 dbuf_rele(db, FTAG); 1468 dmu_write(os, object, offset, blksz, buf->b_data, tx); 1469 dmu_return_arcbuf(buf); 1470 XUIOSTAT_BUMP(xuiostat_wbuf_copied); 1471 } 1472} 1473 1474typedef struct { 1475 dbuf_dirty_record_t *dsa_dr; 1476 dmu_sync_cb_t *dsa_done; 1477 zgd_t *dsa_zgd; 1478 dmu_tx_t *dsa_tx; 1479} dmu_sync_arg_t; 1480 1481/* ARGSUSED */ 1482static void 1483dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg) 1484{ 1485 dmu_sync_arg_t *dsa = varg; 1486 dmu_buf_t *db = dsa->dsa_zgd->zgd_db; 1487 blkptr_t *bp = zio->io_bp; 1488 1489 if (zio->io_error == 0) { 1490 if (BP_IS_HOLE(bp)) { 1491 /* 1492 * A block of zeros may compress to a hole, but the 1493 * block size still needs to be known for replay. 1494 */ 1495 BP_SET_LSIZE(bp, db->db_size); 1496 } else if (!BP_IS_EMBEDDED(bp)) { 1497 ASSERT(BP_GET_LEVEL(bp) == 0); 1498 bp->blk_fill = 1; 1499 } 1500 } 1501} 1502 1503static void 1504dmu_sync_late_arrival_ready(zio_t *zio) 1505{ 1506 dmu_sync_ready(zio, NULL, zio->io_private); 1507} 1508 1509/* ARGSUSED */ 1510static void 1511dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg) 1512{ 1513 dmu_sync_arg_t *dsa = varg; 1514 dbuf_dirty_record_t *dr = dsa->dsa_dr; 1515 dmu_buf_impl_t *db = dr->dr_dbuf; 1516 1517 mutex_enter(&db->db_mtx); 1518 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC); 1519 if (zio->io_error == 0) { 1520 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE); 1521 if (dr->dt.dl.dr_nopwrite) { 1522 blkptr_t *bp = zio->io_bp; 1523 blkptr_t *bp_orig = &zio->io_bp_orig; 1524 uint8_t chksum = BP_GET_CHECKSUM(bp_orig); 1525 1526 ASSERT(BP_EQUAL(bp, bp_orig)); 1527 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF); 1528 ASSERT(zio_checksum_table[chksum].ci_flags & 1529 ZCHECKSUM_FLAG_NOPWRITE); 1530 } 1531 dr->dt.dl.dr_overridden_by = *zio->io_bp; 1532 dr->dt.dl.dr_override_state = DR_OVERRIDDEN; 1533 dr->dt.dl.dr_copies = zio->io_prop.zp_copies; 1534 1535 /* 1536 * Old style holes are filled with all zeros, whereas 1537 * new-style holes maintain their lsize, type, level, 1538 * and birth time (see zio_write_compress). While we 1539 * need to reset the BP_SET_LSIZE() call that happened 1540 * in dmu_sync_ready for old style holes, we do *not* 1541 * want to wipe out the information contained in new 1542 * style holes. Thus, only zero out the block pointer if 1543 * it's an old style hole. 1544 */ 1545 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by) && 1546 dr->dt.dl.dr_overridden_by.blk_birth == 0) 1547 BP_ZERO(&dr->dt.dl.dr_overridden_by); 1548 } else { 1549 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 1550 } 1551 cv_broadcast(&db->db_changed); 1552 mutex_exit(&db->db_mtx); 1553 1554 dsa->dsa_done(dsa->dsa_zgd, zio->io_error); 1555 1556 kmem_free(dsa, sizeof (*dsa)); 1557} 1558 1559static void 1560dmu_sync_late_arrival_done(zio_t *zio) 1561{ 1562 blkptr_t *bp = zio->io_bp; 1563 dmu_sync_arg_t *dsa = zio->io_private; 1564 blkptr_t *bp_orig = &zio->io_bp_orig; 1565 1566 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) { 1567 /* 1568 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE) 1569 * then there is nothing to do here. Otherwise, free the 1570 * newly allocated block in this txg. 1571 */ 1572 if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 1573 ASSERT(BP_EQUAL(bp, bp_orig)); 1574 } else { 1575 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig)); 1576 ASSERT(zio->io_bp->blk_birth == zio->io_txg); 1577 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa)); 1578 zio_free(zio->io_spa, zio->io_txg, zio->io_bp); 1579 } 1580 } 1581 1582 dmu_tx_commit(dsa->dsa_tx); 1583 1584 dsa->dsa_done(dsa->dsa_zgd, zio->io_error); 1585 1586 kmem_free(dsa, sizeof (*dsa)); 1587} 1588 1589static int 1590dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd, 1591 zio_prop_t *zp, zbookmark_phys_t *zb) 1592{ 1593 dmu_sync_arg_t *dsa; 1594 dmu_tx_t *tx; 1595 1596 tx = dmu_tx_create(os); 1597 dmu_tx_hold_space(tx, zgd->zgd_db->db_size); 1598 if (dmu_tx_assign(tx, TXG_WAIT) != 0) { 1599 dmu_tx_abort(tx); 1600 /* Make zl_get_data do txg_waited_synced() */ 1601 return (SET_ERROR(EIO)); 1602 } 1603 1604 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP); 1605 dsa->dsa_dr = NULL; 1606 dsa->dsa_done = done; 1607 dsa->dsa_zgd = zgd; 1608 dsa->dsa_tx = tx; 1609 1610 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp, 1611 zgd->zgd_db->db_data, zgd->zgd_db->db_size, zp, 1612 dmu_sync_late_arrival_ready, NULL, dmu_sync_late_arrival_done, dsa, 1613 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb)); 1614 1615 return (0); 1616} 1617 1618/* 1619 * Intent log support: sync the block associated with db to disk. 1620 * N.B. and XXX: the caller is responsible for making sure that the 1621 * data isn't changing while dmu_sync() is writing it. 1622 * 1623 * Return values: 1624 * 1625 * EEXIST: this txg has already been synced, so there's nothing to do. 1626 * The caller should not log the write. 1627 * 1628 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do. 1629 * The caller should not log the write. 1630 * 1631 * EALREADY: this block is already in the process of being synced. 1632 * The caller should track its progress (somehow). 1633 * 1634 * EIO: could not do the I/O. 1635 * The caller should do a txg_wait_synced(). 1636 * 1637 * 0: the I/O has been initiated. 1638 * The caller should log this blkptr in the done callback. 1639 * It is possible that the I/O will fail, in which case 1640 * the error will be reported to the done callback and 1641 * propagated to pio from zio_done(). 1642 */ 1643int 1644dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd) 1645{ 1646 blkptr_t *bp = zgd->zgd_bp; 1647 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db; 1648 objset_t *os = db->db_objset; 1649 dsl_dataset_t *ds = os->os_dsl_dataset; 1650 dbuf_dirty_record_t *dr; 1651 dmu_sync_arg_t *dsa; 1652 zbookmark_phys_t zb; 1653 zio_prop_t zp; 1654 dnode_t *dn; 1655 1656 ASSERT(pio != NULL); 1657 ASSERT(txg != 0); 1658 1659 SET_BOOKMARK(&zb, ds->ds_object, 1660 db->db.db_object, db->db_level, db->db_blkid); 1661 1662 DB_DNODE_ENTER(db); 1663 dn = DB_DNODE(db); 1664 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp); 1665 DB_DNODE_EXIT(db); 1666 1667 /* 1668 * If we're frozen (running ziltest), we always need to generate a bp. 1669 */ 1670 if (txg > spa_freeze_txg(os->os_spa)) 1671 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); 1672 1673 /* 1674 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf() 1675 * and us. If we determine that this txg is not yet syncing, 1676 * but it begins to sync a moment later, that's OK because the 1677 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx. 1678 */ 1679 mutex_enter(&db->db_mtx); 1680 1681 if (txg <= spa_last_synced_txg(os->os_spa)) { 1682 /* 1683 * This txg has already synced. There's nothing to do. 1684 */ 1685 mutex_exit(&db->db_mtx); 1686 return (SET_ERROR(EEXIST)); 1687 } 1688 1689 if (txg <= spa_syncing_txg(os->os_spa)) { 1690 /* 1691 * This txg is currently syncing, so we can't mess with 1692 * the dirty record anymore; just write a new log block. 1693 */ 1694 mutex_exit(&db->db_mtx); 1695 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); 1696 } 1697 1698 dr = db->db_last_dirty; 1699 while (dr && dr->dr_txg != txg) 1700 dr = dr->dr_next; 1701 1702 if (dr == NULL) { 1703 /* 1704 * There's no dr for this dbuf, so it must have been freed. 1705 * There's no need to log writes to freed blocks, so we're done. 1706 */ 1707 mutex_exit(&db->db_mtx); 1708 return (SET_ERROR(ENOENT)); 1709 } 1710 1711 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg); 1712 1713 /* 1714 * Assume the on-disk data is X, the current syncing data (in 1715 * txg - 1) is Y, and the current in-memory data is Z (currently 1716 * in dmu_sync). 1717 * 1718 * We usually want to perform a nopwrite if X and Z are the 1719 * same. However, if Y is different (i.e. the BP is going to 1720 * change before this write takes effect), then a nopwrite will 1721 * be incorrect - we would override with X, which could have 1722 * been freed when Y was written. 1723 * 1724 * (Note that this is not a concern when we are nop-writing from 1725 * syncing context, because X and Y must be identical, because 1726 * all previous txgs have been synced.) 1727 * 1728 * Therefore, we disable nopwrite if the current BP could change 1729 * before this TXG. There are two ways it could change: by 1730 * being dirty (dr_next is non-NULL), or by being freed 1731 * (dnode_block_freed()). This behavior is verified by 1732 * zio_done(), which VERIFYs that the override BP is identical 1733 * to the on-disk BP. 1734 */ 1735 DB_DNODE_ENTER(db); 1736 dn = DB_DNODE(db); 1737 if (dr->dr_next != NULL || dnode_block_freed(dn, db->db_blkid)) 1738 zp.zp_nopwrite = B_FALSE; 1739 DB_DNODE_EXIT(db); 1740 1741 ASSERT(dr->dr_txg == txg); 1742 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC || 1743 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { 1744 /* 1745 * We have already issued a sync write for this buffer, 1746 * or this buffer has already been synced. It could not 1747 * have been dirtied since, or we would have cleared the state. 1748 */ 1749 mutex_exit(&db->db_mtx); 1750 return (SET_ERROR(EALREADY)); 1751 } 1752 1753 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); 1754 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC; 1755 mutex_exit(&db->db_mtx); 1756 1757 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP); 1758 dsa->dsa_dr = dr; 1759 dsa->dsa_done = done; 1760 dsa->dsa_zgd = zgd; 1761 dsa->dsa_tx = NULL; 1762 1763 zio_nowait(arc_write(pio, os->os_spa, txg, 1764 bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db), 1765 DBUF_IS_L2COMPRESSIBLE(db), &zp, dmu_sync_ready, 1766 NULL, dmu_sync_done, dsa, ZIO_PRIORITY_SYNC_WRITE, 1767 ZIO_FLAG_CANFAIL, &zb)); 1768 1769 return (0); 1770} 1771 1772int 1773dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs, 1774 dmu_tx_t *tx) 1775{ 1776 dnode_t *dn; 1777 int err; 1778 1779 err = dnode_hold(os, object, FTAG, &dn); 1780 if (err) 1781 return (err); 1782 err = dnode_set_blksz(dn, size, ibs, tx); 1783 dnode_rele(dn, FTAG); 1784 return (err); 1785} 1786 1787void 1788dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum, 1789 dmu_tx_t *tx) 1790{ 1791 dnode_t *dn; 1792 1793 /* 1794 * Send streams include each object's checksum function. This 1795 * check ensures that the receiving system can understand the 1796 * checksum function transmitted. 1797 */ 1798 ASSERT3U(checksum, <, ZIO_CHECKSUM_LEGACY_FUNCTIONS); 1799 1800 VERIFY0(dnode_hold(os, object, FTAG, &dn)); 1801 ASSERT3U(checksum, <, ZIO_CHECKSUM_FUNCTIONS); 1802 dn->dn_checksum = checksum; 1803 dnode_setdirty(dn, tx); 1804 dnode_rele(dn, FTAG); 1805} 1806 1807void 1808dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress, 1809 dmu_tx_t *tx) 1810{ 1811 dnode_t *dn; 1812 1813 /* 1814 * Send streams include each object's compression function. This 1815 * check ensures that the receiving system can understand the 1816 * compression function transmitted. 1817 */ 1818 ASSERT3U(compress, <, ZIO_COMPRESS_LEGACY_FUNCTIONS); 1819 1820 VERIFY0(dnode_hold(os, object, FTAG, &dn)); 1821 dn->dn_compress = compress; 1822 dnode_setdirty(dn, tx); 1823 dnode_rele(dn, FTAG); 1824} 1825 1826int zfs_mdcomp_disable = 0; 1827SYSCTL_INT(_vfs_zfs, OID_AUTO, mdcomp_disable, CTLFLAG_RWTUN, 1828 &zfs_mdcomp_disable, 0, "Disable metadata compression"); 1829 1830/* 1831 * When the "redundant_metadata" property is set to "most", only indirect 1832 * blocks of this level and higher will have an additional ditto block. 1833 */ 1834int zfs_redundant_metadata_most_ditto_level = 2; 1835 1836void 1837dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp) 1838{ 1839 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET; 1840 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) || 1841 (wp & WP_SPILL)); 1842 enum zio_checksum checksum = os->os_checksum; 1843 enum zio_compress compress = os->os_compress; 1844 enum zio_checksum dedup_checksum = os->os_dedup_checksum; 1845 boolean_t dedup = B_FALSE; 1846 boolean_t nopwrite = B_FALSE; 1847 boolean_t dedup_verify = os->os_dedup_verify; 1848 int copies = os->os_copies; 1849 1850 /* 1851 * We maintain different write policies for each of the following 1852 * types of data: 1853 * 1. metadata 1854 * 2. preallocated blocks (i.e. level-0 blocks of a dump device) 1855 * 3. all other level 0 blocks 1856 */ 1857 if (ismd) { 1858 if (zfs_mdcomp_disable) { 1859 compress = ZIO_COMPRESS_EMPTY; 1860 } else { 1861 /* 1862 * XXX -- we should design a compression algorithm 1863 * that specializes in arrays of bps. 1864 */ 1865 compress = zio_compress_select(os->os_spa, 1866 ZIO_COMPRESS_ON, ZIO_COMPRESS_ON); 1867 } 1868 1869 /* 1870 * Metadata always gets checksummed. If the data 1871 * checksum is multi-bit correctable, and it's not a 1872 * ZBT-style checksum, then it's suitable for metadata 1873 * as well. Otherwise, the metadata checksum defaults 1874 * to fletcher4. 1875 */ 1876 if (!(zio_checksum_table[checksum].ci_flags & 1877 ZCHECKSUM_FLAG_METADATA) || 1878 (zio_checksum_table[checksum].ci_flags & 1879 ZCHECKSUM_FLAG_EMBEDDED)) 1880 checksum = ZIO_CHECKSUM_FLETCHER_4; 1881 1882 if (os->os_redundant_metadata == ZFS_REDUNDANT_METADATA_ALL || 1883 (os->os_redundant_metadata == 1884 ZFS_REDUNDANT_METADATA_MOST && 1885 (level >= zfs_redundant_metadata_most_ditto_level || 1886 DMU_OT_IS_METADATA(type) || (wp & WP_SPILL)))) 1887 copies++; 1888 } else if (wp & WP_NOFILL) { 1889 ASSERT(level == 0); 1890 1891 /* 1892 * If we're writing preallocated blocks, we aren't actually 1893 * writing them so don't set any policy properties. These 1894 * blocks are currently only used by an external subsystem 1895 * outside of zfs (i.e. dump) and not written by the zio 1896 * pipeline. 1897 */ 1898 compress = ZIO_COMPRESS_OFF; 1899 checksum = ZIO_CHECKSUM_NOPARITY; 1900 } else { 1901 compress = zio_compress_select(os->os_spa, dn->dn_compress, 1902 compress); 1903 1904 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ? 1905 zio_checksum_select(dn->dn_checksum, checksum) : 1906 dedup_checksum; 1907 1908 /* 1909 * Determine dedup setting. If we are in dmu_sync(), 1910 * we won't actually dedup now because that's all 1911 * done in syncing context; but we do want to use the 1912 * dedup checkum. If the checksum is not strong 1913 * enough to ensure unique signatures, force 1914 * dedup_verify. 1915 */ 1916 if (dedup_checksum != ZIO_CHECKSUM_OFF) { 1917 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE; 1918 if (!(zio_checksum_table[checksum].ci_flags & 1919 ZCHECKSUM_FLAG_DEDUP)) 1920 dedup_verify = B_TRUE; 1921 } 1922 1923 /* 1924 * Enable nopwrite if we have secure enough checksum 1925 * algorithm (see comment in zio_nop_write) and 1926 * compression is enabled. We don't enable nopwrite if 1927 * dedup is enabled as the two features are mutually 1928 * exclusive. 1929 */ 1930 nopwrite = (!dedup && (zio_checksum_table[checksum].ci_flags & 1931 ZCHECKSUM_FLAG_NOPWRITE) && 1932 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled); 1933 } 1934 1935 zp->zp_checksum = checksum; 1936 zp->zp_compress = compress; 1937 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type; 1938 zp->zp_level = level; 1939 zp->zp_copies = MIN(copies, spa_max_replication(os->os_spa)); 1940 zp->zp_dedup = dedup; 1941 zp->zp_dedup_verify = dedup && dedup_verify; 1942 zp->zp_nopwrite = nopwrite; 1943} 1944 1945int 1946dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off) 1947{ 1948 dnode_t *dn; 1949 int err; 1950 1951 /* 1952 * Sync any current changes before 1953 * we go trundling through the block pointers. 1954 */ 1955 err = dmu_object_wait_synced(os, object); 1956 if (err) { 1957 return (err); 1958 } 1959 1960 err = dnode_hold(os, object, FTAG, &dn); 1961 if (err) { 1962 return (err); 1963 } 1964 1965 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0); 1966 dnode_rele(dn, FTAG); 1967 1968 return (err); 1969} 1970 1971/* 1972 * Given the ZFS object, if it contains any dirty nodes 1973 * this function flushes all dirty blocks to disk. This 1974 * ensures the DMU object info is updated. A more efficient 1975 * future version might just find the TXG with the maximum 1976 * ID and wait for that to be synced. 1977 */ 1978int 1979dmu_object_wait_synced(objset_t *os, uint64_t object) 1980{ 1981 dnode_t *dn; 1982 int error, i; 1983 1984 error = dnode_hold(os, object, FTAG, &dn); 1985 if (error) { 1986 return (error); 1987 } 1988 1989 for (i = 0; i < TXG_SIZE; i++) { 1990 if (list_link_active(&dn->dn_dirty_link[i])) { 1991 break; 1992 } 1993 } 1994 dnode_rele(dn, FTAG); 1995 if (i != TXG_SIZE) { 1996 txg_wait_synced(dmu_objset_pool(os), 0); 1997 } 1998 1999 return (0); 2000} 2001 2002void 2003dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi) 2004{ 2005 dnode_phys_t *dnp; 2006 2007 rw_enter(&dn->dn_struct_rwlock, RW_READER); 2008 mutex_enter(&dn->dn_mtx); 2009 2010 dnp = dn->dn_phys; 2011 2012 doi->doi_data_block_size = dn->dn_datablksz; 2013 doi->doi_metadata_block_size = dn->dn_indblkshift ? 2014 1ULL << dn->dn_indblkshift : 0; 2015 doi->doi_type = dn->dn_type; 2016 doi->doi_bonus_type = dn->dn_bonustype; 2017 doi->doi_bonus_size = dn->dn_bonuslen; 2018 doi->doi_indirection = dn->dn_nlevels; 2019 doi->doi_checksum = dn->dn_checksum; 2020 doi->doi_compress = dn->dn_compress; 2021 doi->doi_nblkptr = dn->dn_nblkptr; 2022 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9; 2023 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz; 2024 doi->doi_fill_count = 0; 2025 for (int i = 0; i < dnp->dn_nblkptr; i++) 2026 doi->doi_fill_count += BP_GET_FILL(&dnp->dn_blkptr[i]); 2027 2028 mutex_exit(&dn->dn_mtx); 2029 rw_exit(&dn->dn_struct_rwlock); 2030} 2031 2032/* 2033 * Get information on a DMU object. 2034 * If doi is NULL, just indicates whether the object exists. 2035 */ 2036int 2037dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi) 2038{ 2039 dnode_t *dn; 2040 int err = dnode_hold(os, object, FTAG, &dn); 2041 2042 if (err) 2043 return (err); 2044 2045 if (doi != NULL) 2046 dmu_object_info_from_dnode(dn, doi); 2047 2048 dnode_rele(dn, FTAG); 2049 return (0); 2050} 2051 2052/* 2053 * As above, but faster; can be used when you have a held dbuf in hand. 2054 */ 2055void 2056dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi) 2057{ 2058 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2059 2060 DB_DNODE_ENTER(db); 2061 dmu_object_info_from_dnode(DB_DNODE(db), doi); 2062 DB_DNODE_EXIT(db); 2063} 2064 2065/* 2066 * Faster still when you only care about the size. 2067 * This is specifically optimized for zfs_getattr(). 2068 */ 2069void 2070dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize, 2071 u_longlong_t *nblk512) 2072{ 2073 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2074 dnode_t *dn; 2075 2076 DB_DNODE_ENTER(db); 2077 dn = DB_DNODE(db); 2078 2079 *blksize = dn->dn_datablksz; 2080 /* add 1 for dnode space */ 2081 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >> 2082 SPA_MINBLOCKSHIFT) + 1; 2083 DB_DNODE_EXIT(db); 2084} 2085 2086void 2087byteswap_uint64_array(void *vbuf, size_t size) 2088{ 2089 uint64_t *buf = vbuf; 2090 size_t count = size >> 3; 2091 int i; 2092 2093 ASSERT((size & 7) == 0); 2094 2095 for (i = 0; i < count; i++) 2096 buf[i] = BSWAP_64(buf[i]); 2097} 2098 2099void 2100byteswap_uint32_array(void *vbuf, size_t size) 2101{ 2102 uint32_t *buf = vbuf; 2103 size_t count = size >> 2; 2104 int i; 2105 2106 ASSERT((size & 3) == 0); 2107 2108 for (i = 0; i < count; i++) 2109 buf[i] = BSWAP_32(buf[i]); 2110} 2111 2112void 2113byteswap_uint16_array(void *vbuf, size_t size) 2114{ 2115 uint16_t *buf = vbuf; 2116 size_t count = size >> 1; 2117 int i; 2118 2119 ASSERT((size & 1) == 0); 2120 2121 for (i = 0; i < count; i++) 2122 buf[i] = BSWAP_16(buf[i]); 2123} 2124 2125/* ARGSUSED */ 2126void 2127byteswap_uint8_array(void *vbuf, size_t size) 2128{ 2129} 2130 2131void 2132dmu_init(void) 2133{ 2134 zfs_dbgmsg_init(); 2135 sa_cache_init(); 2136 xuio_stat_init(); 2137 dmu_objset_init(); 2138 dnode_init(); 2139 dbuf_init(); 2140 zfetch_init(); 2141 zio_compress_init(); 2142 l2arc_init(); 2143 arc_init(); 2144} 2145 2146void 2147dmu_fini(void) 2148{ 2149 arc_fini(); /* arc depends on l2arc, so arc must go first */ 2150 l2arc_fini(); 2151 zfetch_fini(); 2152 zio_compress_fini(); 2153 dbuf_fini(); 2154 dnode_fini(); 2155 dmu_objset_fini(); 2156 xuio_stat_fini(); 2157 sa_cache_fini(); 2158 zfs_dbgmsg_fini(); 2159} 2160