dmu.c revision 260763
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) 2013 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 28#include <sys/dmu.h> 29#include <sys/dmu_impl.h> 30#include <sys/dmu_tx.h> 31#include <sys/dbuf.h> 32#include <sys/dnode.h> 33#include <sys/zfs_context.h> 34#include <sys/dmu_objset.h> 35#include <sys/dmu_traverse.h> 36#include <sys/dsl_dataset.h> 37#include <sys/dsl_dir.h> 38#include <sys/dsl_pool.h> 39#include <sys/dsl_synctask.h> 40#include <sys/dsl_prop.h> 41#include <sys/dmu_zfetch.h> 42#include <sys/zfs_ioctl.h> 43#include <sys/zap.h> 44#include <sys/zio_checksum.h> 45#include <sys/zio_compress.h> 46#include <sys/sa.h> 47#ifdef _KERNEL 48#include <sys/zfs_znode.h> 49#endif 50 51/* 52 * Enable/disable nopwrite feature. 53 */ 54int zfs_nopwrite_enabled = 1; 55SYSCTL_DECL(_vfs_zfs); 56TUNABLE_INT("vfs.zfs.nopwrite_enabled", &zfs_nopwrite_enabled); 57SYSCTL_INT(_vfs_zfs, OID_AUTO, nopwrite_enabled, CTLFLAG_RDTUN, 58 &zfs_nopwrite_enabled, 0, "Enable nopwrite feature"); 59 60const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = { 61 { DMU_BSWAP_UINT8, TRUE, "unallocated" }, 62 { DMU_BSWAP_ZAP, TRUE, "object directory" }, 63 { DMU_BSWAP_UINT64, TRUE, "object array" }, 64 { DMU_BSWAP_UINT8, TRUE, "packed nvlist" }, 65 { DMU_BSWAP_UINT64, TRUE, "packed nvlist size" }, 66 { DMU_BSWAP_UINT64, TRUE, "bpobj" }, 67 { DMU_BSWAP_UINT64, TRUE, "bpobj header" }, 68 { DMU_BSWAP_UINT64, TRUE, "SPA space map header" }, 69 { DMU_BSWAP_UINT64, TRUE, "SPA space map" }, 70 { DMU_BSWAP_UINT64, TRUE, "ZIL intent log" }, 71 { DMU_BSWAP_DNODE, TRUE, "DMU dnode" }, 72 { DMU_BSWAP_OBJSET, TRUE, "DMU objset" }, 73 { DMU_BSWAP_UINT64, TRUE, "DSL directory" }, 74 { DMU_BSWAP_ZAP, TRUE, "DSL directory child map"}, 75 { DMU_BSWAP_ZAP, TRUE, "DSL dataset snap map" }, 76 { DMU_BSWAP_ZAP, TRUE, "DSL props" }, 77 { DMU_BSWAP_UINT64, TRUE, "DSL dataset" }, 78 { DMU_BSWAP_ZNODE, TRUE, "ZFS znode" }, 79 { DMU_BSWAP_OLDACL, TRUE, "ZFS V0 ACL" }, 80 { DMU_BSWAP_UINT8, FALSE, "ZFS plain file" }, 81 { DMU_BSWAP_ZAP, TRUE, "ZFS directory" }, 82 { DMU_BSWAP_ZAP, TRUE, "ZFS master node" }, 83 { DMU_BSWAP_ZAP, TRUE, "ZFS delete queue" }, 84 { DMU_BSWAP_UINT8, FALSE, "zvol object" }, 85 { DMU_BSWAP_ZAP, TRUE, "zvol prop" }, 86 { DMU_BSWAP_UINT8, FALSE, "other uint8[]" }, 87 { DMU_BSWAP_UINT64, FALSE, "other uint64[]" }, 88 { DMU_BSWAP_ZAP, TRUE, "other ZAP" }, 89 { DMU_BSWAP_ZAP, TRUE, "persistent error log" }, 90 { DMU_BSWAP_UINT8, TRUE, "SPA history" }, 91 { DMU_BSWAP_UINT64, TRUE, "SPA history offsets" }, 92 { DMU_BSWAP_ZAP, TRUE, "Pool properties" }, 93 { DMU_BSWAP_ZAP, TRUE, "DSL permissions" }, 94 { DMU_BSWAP_ACL, TRUE, "ZFS ACL" }, 95 { DMU_BSWAP_UINT8, TRUE, "ZFS SYSACL" }, 96 { DMU_BSWAP_UINT8, TRUE, "FUID table" }, 97 { DMU_BSWAP_UINT64, TRUE, "FUID table size" }, 98 { DMU_BSWAP_ZAP, TRUE, "DSL dataset next clones"}, 99 { DMU_BSWAP_ZAP, TRUE, "scan work queue" }, 100 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group used" }, 101 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group quota" }, 102 { DMU_BSWAP_ZAP, TRUE, "snapshot refcount tags"}, 103 { DMU_BSWAP_ZAP, TRUE, "DDT ZAP algorithm" }, 104 { DMU_BSWAP_ZAP, TRUE, "DDT statistics" }, 105 { DMU_BSWAP_UINT8, TRUE, "System attributes" }, 106 { DMU_BSWAP_ZAP, TRUE, "SA master node" }, 107 { DMU_BSWAP_ZAP, TRUE, "SA attr registration" }, 108 { DMU_BSWAP_ZAP, TRUE, "SA attr layouts" }, 109 { DMU_BSWAP_ZAP, TRUE, "scan translations" }, 110 { DMU_BSWAP_UINT8, FALSE, "deduplicated block" }, 111 { DMU_BSWAP_ZAP, TRUE, "DSL deadlist map" }, 112 { DMU_BSWAP_UINT64, TRUE, "DSL deadlist map hdr" }, 113 { DMU_BSWAP_ZAP, TRUE, "DSL dir clones" }, 114 { DMU_BSWAP_UINT64, TRUE, "bpobj subobj" } 115}; 116 117const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = { 118 { byteswap_uint8_array, "uint8" }, 119 { byteswap_uint16_array, "uint16" }, 120 { byteswap_uint32_array, "uint32" }, 121 { byteswap_uint64_array, "uint64" }, 122 { zap_byteswap, "zap" }, 123 { dnode_buf_byteswap, "dnode" }, 124 { dmu_objset_byteswap, "objset" }, 125 { zfs_znode_byteswap, "znode" }, 126 { zfs_oldacl_byteswap, "oldacl" }, 127 { zfs_acl_byteswap, "acl" } 128}; 129 130int 131dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset, 132 void *tag, dmu_buf_t **dbp, int flags) 133{ 134 dnode_t *dn; 135 uint64_t blkid; 136 dmu_buf_impl_t *db; 137 int err; 138 int db_flags = DB_RF_CANFAIL; 139 140 if (flags & DMU_READ_NO_PREFETCH) 141 db_flags |= DB_RF_NOPREFETCH; 142 143 err = dnode_hold(os, object, FTAG, &dn); 144 if (err) 145 return (err); 146 blkid = dbuf_whichblock(dn, offset); 147 rw_enter(&dn->dn_struct_rwlock, RW_READER); 148 db = dbuf_hold(dn, blkid, tag); 149 rw_exit(&dn->dn_struct_rwlock); 150 if (db == NULL) { 151 err = SET_ERROR(EIO); 152 } else { 153 err = dbuf_read(db, NULL, db_flags); 154 if (err) { 155 dbuf_rele(db, tag); 156 db = NULL; 157 } 158 } 159 160 dnode_rele(dn, FTAG); 161 *dbp = &db->db; /* NULL db plus first field offset is NULL */ 162 return (err); 163} 164 165int 166dmu_bonus_max(void) 167{ 168 return (DN_MAX_BONUSLEN); 169} 170 171int 172dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx) 173{ 174 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 175 dnode_t *dn; 176 int error; 177 178 DB_DNODE_ENTER(db); 179 dn = DB_DNODE(db); 180 181 if (dn->dn_bonus != db) { 182 error = SET_ERROR(EINVAL); 183 } else if (newsize < 0 || newsize > db_fake->db_size) { 184 error = SET_ERROR(EINVAL); 185 } else { 186 dnode_setbonuslen(dn, newsize, tx); 187 error = 0; 188 } 189 190 DB_DNODE_EXIT(db); 191 return (error); 192} 193 194int 195dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx) 196{ 197 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 198 dnode_t *dn; 199 int error; 200 201 DB_DNODE_ENTER(db); 202 dn = DB_DNODE(db); 203 204 if (!DMU_OT_IS_VALID(type)) { 205 error = SET_ERROR(EINVAL); 206 } else if (dn->dn_bonus != db) { 207 error = SET_ERROR(EINVAL); 208 } else { 209 dnode_setbonus_type(dn, type, tx); 210 error = 0; 211 } 212 213 DB_DNODE_EXIT(db); 214 return (error); 215} 216 217dmu_object_type_t 218dmu_get_bonustype(dmu_buf_t *db_fake) 219{ 220 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 221 dnode_t *dn; 222 dmu_object_type_t type; 223 224 DB_DNODE_ENTER(db); 225 dn = DB_DNODE(db); 226 type = dn->dn_bonustype; 227 DB_DNODE_EXIT(db); 228 229 return (type); 230} 231 232int 233dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx) 234{ 235 dnode_t *dn; 236 int error; 237 238 error = dnode_hold(os, object, FTAG, &dn); 239 dbuf_rm_spill(dn, tx); 240 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 241 dnode_rm_spill(dn, tx); 242 rw_exit(&dn->dn_struct_rwlock); 243 dnode_rele(dn, FTAG); 244 return (error); 245} 246 247/* 248 * returns ENOENT, EIO, or 0. 249 */ 250int 251dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp) 252{ 253 dnode_t *dn; 254 dmu_buf_impl_t *db; 255 int error; 256 257 error = dnode_hold(os, object, FTAG, &dn); 258 if (error) 259 return (error); 260 261 rw_enter(&dn->dn_struct_rwlock, RW_READER); 262 if (dn->dn_bonus == NULL) { 263 rw_exit(&dn->dn_struct_rwlock); 264 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 265 if (dn->dn_bonus == NULL) 266 dbuf_create_bonus(dn); 267 } 268 db = dn->dn_bonus; 269 270 /* as long as the bonus buf is held, the dnode will be held */ 271 if (refcount_add(&db->db_holds, tag) == 1) { 272 VERIFY(dnode_add_ref(dn, db)); 273 (void) atomic_inc_32_nv(&dn->dn_dbufs_count); 274 } 275 276 /* 277 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's 278 * hold and incrementing the dbuf count to ensure that dnode_move() sees 279 * a dnode hold for every dbuf. 280 */ 281 rw_exit(&dn->dn_struct_rwlock); 282 283 dnode_rele(dn, FTAG); 284 285 VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH)); 286 287 *dbp = &db->db; 288 return (0); 289} 290 291/* 292 * returns ENOENT, EIO, or 0. 293 * 294 * This interface will allocate a blank spill dbuf when a spill blk 295 * doesn't already exist on the dnode. 296 * 297 * if you only want to find an already existing spill db, then 298 * dmu_spill_hold_existing() should be used. 299 */ 300int 301dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp) 302{ 303 dmu_buf_impl_t *db = NULL; 304 int err; 305 306 if ((flags & DB_RF_HAVESTRUCT) == 0) 307 rw_enter(&dn->dn_struct_rwlock, RW_READER); 308 309 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag); 310 311 if ((flags & DB_RF_HAVESTRUCT) == 0) 312 rw_exit(&dn->dn_struct_rwlock); 313 314 ASSERT(db != NULL); 315 err = dbuf_read(db, NULL, flags); 316 if (err == 0) 317 *dbp = &db->db; 318 else 319 dbuf_rele(db, tag); 320 return (err); 321} 322 323int 324dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp) 325{ 326 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus; 327 dnode_t *dn; 328 int err; 329 330 DB_DNODE_ENTER(db); 331 dn = DB_DNODE(db); 332 333 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) { 334 err = SET_ERROR(EINVAL); 335 } else { 336 rw_enter(&dn->dn_struct_rwlock, RW_READER); 337 338 if (!dn->dn_have_spill) { 339 err = SET_ERROR(ENOENT); 340 } else { 341 err = dmu_spill_hold_by_dnode(dn, 342 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp); 343 } 344 345 rw_exit(&dn->dn_struct_rwlock); 346 } 347 348 DB_DNODE_EXIT(db); 349 return (err); 350} 351 352int 353dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp) 354{ 355 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus; 356 dnode_t *dn; 357 int err; 358 359 DB_DNODE_ENTER(db); 360 dn = DB_DNODE(db); 361 err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp); 362 DB_DNODE_EXIT(db); 363 364 return (err); 365} 366 367/* 368 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces 369 * to take a held dnode rather than <os, object> -- the lookup is wasteful, 370 * and can induce severe lock contention when writing to several files 371 * whose dnodes are in the same block. 372 */ 373static int 374dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length, 375 int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags) 376{ 377 dmu_buf_t **dbp; 378 uint64_t blkid, nblks, i; 379 uint32_t dbuf_flags; 380 int err; 381 zio_t *zio; 382 383 ASSERT(length <= DMU_MAX_ACCESS); 384 385 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT; 386 if (flags & DMU_READ_NO_PREFETCH || length > zfetch_array_rd_sz) 387 dbuf_flags |= DB_RF_NOPREFETCH; 388 389 rw_enter(&dn->dn_struct_rwlock, RW_READER); 390 if (dn->dn_datablkshift) { 391 int blkshift = dn->dn_datablkshift; 392 nblks = (P2ROUNDUP(offset+length, 1ULL<<blkshift) - 393 P2ALIGN(offset, 1ULL<<blkshift)) >> blkshift; 394 } else { 395 if (offset + length > dn->dn_datablksz) { 396 zfs_panic_recover("zfs: accessing past end of object " 397 "%llx/%llx (size=%u access=%llu+%llu)", 398 (longlong_t)dn->dn_objset-> 399 os_dsl_dataset->ds_object, 400 (longlong_t)dn->dn_object, dn->dn_datablksz, 401 (longlong_t)offset, (longlong_t)length); 402 rw_exit(&dn->dn_struct_rwlock); 403 return (SET_ERROR(EIO)); 404 } 405 nblks = 1; 406 } 407 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP); 408 409 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL); 410 blkid = dbuf_whichblock(dn, offset); 411 for (i = 0; i < nblks; i++) { 412 dmu_buf_impl_t *db = dbuf_hold(dn, blkid+i, tag); 413 if (db == NULL) { 414 rw_exit(&dn->dn_struct_rwlock); 415 dmu_buf_rele_array(dbp, nblks, tag); 416 zio_nowait(zio); 417 return (SET_ERROR(EIO)); 418 } 419 /* initiate async i/o */ 420 if (read) 421 (void) dbuf_read(db, zio, dbuf_flags); 422#ifdef _KERNEL 423 else 424 curthread->td_ru.ru_oublock++; 425#endif 426 dbp[i] = &db->db; 427 } 428 rw_exit(&dn->dn_struct_rwlock); 429 430 /* wait for async i/o */ 431 err = zio_wait(zio); 432 if (err) { 433 dmu_buf_rele_array(dbp, nblks, tag); 434 return (err); 435 } 436 437 /* wait for other io to complete */ 438 if (read) { 439 for (i = 0; i < nblks; i++) { 440 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i]; 441 mutex_enter(&db->db_mtx); 442 while (db->db_state == DB_READ || 443 db->db_state == DB_FILL) 444 cv_wait(&db->db_changed, &db->db_mtx); 445 if (db->db_state == DB_UNCACHED) 446 err = SET_ERROR(EIO); 447 mutex_exit(&db->db_mtx); 448 if (err) { 449 dmu_buf_rele_array(dbp, nblks, tag); 450 return (err); 451 } 452 } 453 } 454 455 *numbufsp = nblks; 456 *dbpp = dbp; 457 return (0); 458} 459 460static int 461dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset, 462 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp) 463{ 464 dnode_t *dn; 465 int err; 466 467 err = dnode_hold(os, object, FTAG, &dn); 468 if (err) 469 return (err); 470 471 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag, 472 numbufsp, dbpp, DMU_READ_PREFETCH); 473 474 dnode_rele(dn, FTAG); 475 476 return (err); 477} 478 479int 480dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset, 481 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp) 482{ 483 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 484 dnode_t *dn; 485 int err; 486 487 DB_DNODE_ENTER(db); 488 dn = DB_DNODE(db); 489 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag, 490 numbufsp, dbpp, DMU_READ_PREFETCH); 491 DB_DNODE_EXIT(db); 492 493 return (err); 494} 495 496void 497dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag) 498{ 499 int i; 500 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake; 501 502 if (numbufs == 0) 503 return; 504 505 for (i = 0; i < numbufs; i++) { 506 if (dbp[i]) 507 dbuf_rele(dbp[i], tag); 508 } 509 510 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs); 511} 512 513/* 514 * Issue prefetch i/os for the given blocks. 515 * 516 * Note: The assumption is that we *know* these blocks will be needed 517 * almost immediately. Therefore, the prefetch i/os will be issued at 518 * ZIO_PRIORITY_SYNC_READ 519 * 520 * Note: indirect blocks and other metadata will be read synchronously, 521 * causing this function to block if they are not already cached. 522 */ 523void 524dmu_prefetch(objset_t *os, uint64_t object, uint64_t offset, uint64_t len) 525{ 526 dnode_t *dn; 527 uint64_t blkid; 528 int nblks, err; 529 530 if (zfs_prefetch_disable) 531 return; 532 533 if (len == 0) { /* they're interested in the bonus buffer */ 534 dn = DMU_META_DNODE(os); 535 536 if (object == 0 || object >= DN_MAX_OBJECT) 537 return; 538 539 rw_enter(&dn->dn_struct_rwlock, RW_READER); 540 blkid = dbuf_whichblock(dn, object * sizeof (dnode_phys_t)); 541 dbuf_prefetch(dn, blkid, ZIO_PRIORITY_SYNC_READ); 542 rw_exit(&dn->dn_struct_rwlock); 543 return; 544 } 545 546 /* 547 * XXX - Note, if the dnode for the requested object is not 548 * already cached, we will do a *synchronous* read in the 549 * dnode_hold() call. The same is true for any indirects. 550 */ 551 err = dnode_hold(os, object, FTAG, &dn); 552 if (err != 0) 553 return; 554 555 rw_enter(&dn->dn_struct_rwlock, RW_READER); 556 if (dn->dn_datablkshift) { 557 int blkshift = dn->dn_datablkshift; 558 nblks = (P2ROUNDUP(offset + len, 1 << blkshift) - 559 P2ALIGN(offset, 1 << blkshift)) >> blkshift; 560 } else { 561 nblks = (offset < dn->dn_datablksz); 562 } 563 564 if (nblks != 0) { 565 blkid = dbuf_whichblock(dn, offset); 566 for (int i = 0; i < nblks; i++) 567 dbuf_prefetch(dn, blkid + i, ZIO_PRIORITY_SYNC_READ); 568 } 569 570 rw_exit(&dn->dn_struct_rwlock); 571 572 dnode_rele(dn, FTAG); 573} 574 575/* 576 * Get the next "chunk" of file data to free. We traverse the file from 577 * the end so that the file gets shorter over time (if we crashes in the 578 * middle, this will leave us in a better state). We find allocated file 579 * data by simply searching the allocated level 1 indirects. 580 * 581 * On input, *start should be the first offset that does not need to be 582 * freed (e.g. "offset + length"). On return, *start will be the first 583 * offset that should be freed. 584 */ 585static int 586get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum) 587{ 588 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1); 589 /* bytes of data covered by a level-1 indirect block */ 590 uint64_t iblkrange = 591 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT); 592 593 ASSERT3U(minimum, <=, *start); 594 595 if (*start - minimum <= iblkrange * maxblks) { 596 *start = minimum; 597 return (0); 598 } 599 ASSERT(ISP2(iblkrange)); 600 601 for (uint64_t blks = 0; *start > minimum && blks < maxblks; blks++) { 602 int err; 603 604 /* 605 * dnode_next_offset(BACKWARDS) will find an allocated L1 606 * indirect block at or before the input offset. We must 607 * decrement *start so that it is at the end of the region 608 * to search. 609 */ 610 (*start)--; 611 err = dnode_next_offset(dn, 612 DNODE_FIND_BACKWARDS, start, 2, 1, 0); 613 614 /* if there are no indirect blocks before start, we are done */ 615 if (err == ESRCH) { 616 *start = minimum; 617 break; 618 } else if (err != 0) { 619 return (err); 620 } 621 622 /* set start to the beginning of this L1 indirect */ 623 *start = P2ALIGN(*start, iblkrange); 624 } 625 if (*start < minimum) 626 *start = minimum; 627 return (0); 628} 629 630static int 631dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset, 632 uint64_t length) 633{ 634 uint64_t object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz; 635 int err; 636 637 if (offset >= object_size) 638 return (0); 639 640 if (length == DMU_OBJECT_END || offset + length > object_size) 641 length = object_size - offset; 642 643 while (length != 0) { 644 uint64_t chunk_end, chunk_begin; 645 646 chunk_end = chunk_begin = offset + length; 647 648 /* move chunk_begin backwards to the beginning of this chunk */ 649 err = get_next_chunk(dn, &chunk_begin, offset); 650 if (err) 651 return (err); 652 ASSERT3U(chunk_begin, >=, offset); 653 ASSERT3U(chunk_begin, <=, chunk_end); 654 655 dmu_tx_t *tx = dmu_tx_create(os); 656 dmu_tx_hold_free(tx, dn->dn_object, 657 chunk_begin, chunk_end - chunk_begin); 658 err = dmu_tx_assign(tx, TXG_WAIT); 659 if (err) { 660 dmu_tx_abort(tx); 661 return (err); 662 } 663 dnode_free_range(dn, chunk_begin, chunk_end - chunk_begin, tx); 664 dmu_tx_commit(tx); 665 666 length -= chunk_end - chunk_begin; 667 } 668 return (0); 669} 670 671int 672dmu_free_long_range(objset_t *os, uint64_t object, 673 uint64_t offset, uint64_t length) 674{ 675 dnode_t *dn; 676 int err; 677 678 err = dnode_hold(os, object, FTAG, &dn); 679 if (err != 0) 680 return (err); 681 err = dmu_free_long_range_impl(os, dn, offset, length); 682 683 /* 684 * It is important to zero out the maxblkid when freeing the entire 685 * file, so that (a) subsequent calls to dmu_free_long_range_impl() 686 * will take the fast path, and (b) dnode_reallocate() can verify 687 * that the entire file has been freed. 688 */ 689 if (offset == 0 && length == DMU_OBJECT_END) 690 dn->dn_maxblkid = 0; 691 692 dnode_rele(dn, FTAG); 693 return (err); 694} 695 696int 697dmu_free_long_object(objset_t *os, uint64_t object) 698{ 699 dmu_tx_t *tx; 700 int err; 701 702 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END); 703 if (err != 0) 704 return (err); 705 706 tx = dmu_tx_create(os); 707 dmu_tx_hold_bonus(tx, object); 708 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END); 709 err = dmu_tx_assign(tx, TXG_WAIT); 710 if (err == 0) { 711 err = dmu_object_free(os, object, tx); 712 dmu_tx_commit(tx); 713 } else { 714 dmu_tx_abort(tx); 715 } 716 717 return (err); 718} 719 720int 721dmu_free_range(objset_t *os, uint64_t object, uint64_t offset, 722 uint64_t size, dmu_tx_t *tx) 723{ 724 dnode_t *dn; 725 int err = dnode_hold(os, object, FTAG, &dn); 726 if (err) 727 return (err); 728 ASSERT(offset < UINT64_MAX); 729 ASSERT(size == -1ULL || size <= UINT64_MAX - offset); 730 dnode_free_range(dn, offset, size, tx); 731 dnode_rele(dn, FTAG); 732 return (0); 733} 734 735int 736dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 737 void *buf, uint32_t flags) 738{ 739 dnode_t *dn; 740 dmu_buf_t **dbp; 741 int numbufs, err; 742 743 err = dnode_hold(os, object, FTAG, &dn); 744 if (err) 745 return (err); 746 747 /* 748 * Deal with odd block sizes, where there can't be data past the first 749 * block. If we ever do the tail block optimization, we will need to 750 * handle that here as well. 751 */ 752 if (dn->dn_maxblkid == 0) { 753 int newsz = offset > dn->dn_datablksz ? 0 : 754 MIN(size, dn->dn_datablksz - offset); 755 bzero((char *)buf + newsz, size - newsz); 756 size = newsz; 757 } 758 759 while (size > 0) { 760 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2); 761 int i; 762 763 /* 764 * NB: we could do this block-at-a-time, but it's nice 765 * to be reading in parallel. 766 */ 767 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen, 768 TRUE, FTAG, &numbufs, &dbp, flags); 769 if (err) 770 break; 771 772 for (i = 0; i < numbufs; i++) { 773 int tocpy; 774 int bufoff; 775 dmu_buf_t *db = dbp[i]; 776 777 ASSERT(size > 0); 778 779 bufoff = offset - db->db_offset; 780 tocpy = (int)MIN(db->db_size - bufoff, size); 781 782 bcopy((char *)db->db_data + bufoff, buf, tocpy); 783 784 offset += tocpy; 785 size -= tocpy; 786 buf = (char *)buf + tocpy; 787 } 788 dmu_buf_rele_array(dbp, numbufs, FTAG); 789 } 790 dnode_rele(dn, FTAG); 791 return (err); 792} 793 794void 795dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 796 const void *buf, dmu_tx_t *tx) 797{ 798 dmu_buf_t **dbp; 799 int numbufs, i; 800 801 if (size == 0) 802 return; 803 804 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size, 805 FALSE, FTAG, &numbufs, &dbp)); 806 807 for (i = 0; i < numbufs; i++) { 808 int tocpy; 809 int bufoff; 810 dmu_buf_t *db = dbp[i]; 811 812 ASSERT(size > 0); 813 814 bufoff = offset - db->db_offset; 815 tocpy = (int)MIN(db->db_size - bufoff, size); 816 817 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 818 819 if (tocpy == db->db_size) 820 dmu_buf_will_fill(db, tx); 821 else 822 dmu_buf_will_dirty(db, tx); 823 824 bcopy(buf, (char *)db->db_data + bufoff, tocpy); 825 826 if (tocpy == db->db_size) 827 dmu_buf_fill_done(db, tx); 828 829 offset += tocpy; 830 size -= tocpy; 831 buf = (char *)buf + tocpy; 832 } 833 dmu_buf_rele_array(dbp, numbufs, FTAG); 834} 835 836void 837dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 838 dmu_tx_t *tx) 839{ 840 dmu_buf_t **dbp; 841 int numbufs, i; 842 843 if (size == 0) 844 return; 845 846 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size, 847 FALSE, FTAG, &numbufs, &dbp)); 848 849 for (i = 0; i < numbufs; i++) { 850 dmu_buf_t *db = dbp[i]; 851 852 dmu_buf_will_not_fill(db, tx); 853 } 854 dmu_buf_rele_array(dbp, numbufs, FTAG); 855} 856 857/* 858 * DMU support for xuio 859 */ 860kstat_t *xuio_ksp = NULL; 861 862int 863dmu_xuio_init(xuio_t *xuio, int nblk) 864{ 865 dmu_xuio_t *priv; 866 uio_t *uio = &xuio->xu_uio; 867 868 uio->uio_iovcnt = nblk; 869 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP); 870 871 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP); 872 priv->cnt = nblk; 873 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP); 874 priv->iovp = uio->uio_iov; 875 XUIO_XUZC_PRIV(xuio) = priv; 876 877 if (XUIO_XUZC_RW(xuio) == UIO_READ) 878 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk); 879 else 880 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk); 881 882 return (0); 883} 884 885void 886dmu_xuio_fini(xuio_t *xuio) 887{ 888 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 889 int nblk = priv->cnt; 890 891 kmem_free(priv->iovp, nblk * sizeof (iovec_t)); 892 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *)); 893 kmem_free(priv, sizeof (dmu_xuio_t)); 894 895 if (XUIO_XUZC_RW(xuio) == UIO_READ) 896 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk); 897 else 898 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk); 899} 900 901/* 902 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf } 903 * and increase priv->next by 1. 904 */ 905int 906dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n) 907{ 908 struct iovec *iov; 909 uio_t *uio = &xuio->xu_uio; 910 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 911 int i = priv->next++; 912 913 ASSERT(i < priv->cnt); 914 ASSERT(off + n <= arc_buf_size(abuf)); 915 iov = uio->uio_iov + i; 916 iov->iov_base = (char *)abuf->b_data + off; 917 iov->iov_len = n; 918 priv->bufs[i] = abuf; 919 return (0); 920} 921 922int 923dmu_xuio_cnt(xuio_t *xuio) 924{ 925 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 926 return (priv->cnt); 927} 928 929arc_buf_t * 930dmu_xuio_arcbuf(xuio_t *xuio, int i) 931{ 932 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 933 934 ASSERT(i < priv->cnt); 935 return (priv->bufs[i]); 936} 937 938void 939dmu_xuio_clear(xuio_t *xuio, int i) 940{ 941 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 942 943 ASSERT(i < priv->cnt); 944 priv->bufs[i] = NULL; 945} 946 947static void 948xuio_stat_init(void) 949{ 950 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc", 951 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t), 952 KSTAT_FLAG_VIRTUAL); 953 if (xuio_ksp != NULL) { 954 xuio_ksp->ks_data = &xuio_stats; 955 kstat_install(xuio_ksp); 956 } 957} 958 959static void 960xuio_stat_fini(void) 961{ 962 if (xuio_ksp != NULL) { 963 kstat_delete(xuio_ksp); 964 xuio_ksp = NULL; 965 } 966} 967 968void 969xuio_stat_wbuf_copied() 970{ 971 XUIOSTAT_BUMP(xuiostat_wbuf_copied); 972} 973 974void 975xuio_stat_wbuf_nocopy() 976{ 977 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy); 978} 979 980#ifdef _KERNEL 981int 982dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size) 983{ 984 dmu_buf_t **dbp; 985 int numbufs, i, err; 986 xuio_t *xuio = NULL; 987 988 /* 989 * NB: we could do this block-at-a-time, but it's nice 990 * to be reading in parallel. 991 */ 992 err = dmu_buf_hold_array(os, object, uio->uio_loffset, size, TRUE, FTAG, 993 &numbufs, &dbp); 994 if (err) 995 return (err); 996 997#ifdef UIO_XUIO 998 if (uio->uio_extflg == UIO_XUIO) 999 xuio = (xuio_t *)uio; 1000#endif 1001 1002 for (i = 0; i < numbufs; i++) { 1003 int tocpy; 1004 int bufoff; 1005 dmu_buf_t *db = dbp[i]; 1006 1007 ASSERT(size > 0); 1008 1009 bufoff = uio->uio_loffset - db->db_offset; 1010 tocpy = (int)MIN(db->db_size - bufoff, size); 1011 1012 if (xuio) { 1013 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 1014 arc_buf_t *dbuf_abuf = dbi->db_buf; 1015 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi); 1016 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy); 1017 if (!err) { 1018 uio->uio_resid -= tocpy; 1019 uio->uio_loffset += tocpy; 1020 } 1021 1022 if (abuf == dbuf_abuf) 1023 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy); 1024 else 1025 XUIOSTAT_BUMP(xuiostat_rbuf_copied); 1026 } else { 1027 err = uiomove((char *)db->db_data + bufoff, tocpy, 1028 UIO_READ, uio); 1029 } 1030 if (err) 1031 break; 1032 1033 size -= tocpy; 1034 } 1035 dmu_buf_rele_array(dbp, numbufs, FTAG); 1036 1037 return (err); 1038} 1039 1040static int 1041dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx) 1042{ 1043 dmu_buf_t **dbp; 1044 int numbufs; 1045 int err = 0; 1046 int i; 1047 1048 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size, 1049 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH); 1050 if (err) 1051 return (err); 1052 1053 for (i = 0; i < numbufs; i++) { 1054 int tocpy; 1055 int bufoff; 1056 dmu_buf_t *db = dbp[i]; 1057 1058 ASSERT(size > 0); 1059 1060 bufoff = uio->uio_loffset - db->db_offset; 1061 tocpy = (int)MIN(db->db_size - bufoff, size); 1062 1063 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 1064 1065 if (tocpy == db->db_size) 1066 dmu_buf_will_fill(db, tx); 1067 else 1068 dmu_buf_will_dirty(db, tx); 1069 1070 /* 1071 * XXX uiomove could block forever (eg. nfs-backed 1072 * pages). There needs to be a uiolockdown() function 1073 * to lock the pages in memory, so that uiomove won't 1074 * block. 1075 */ 1076 err = uiomove((char *)db->db_data + bufoff, tocpy, 1077 UIO_WRITE, uio); 1078 1079 if (tocpy == db->db_size) 1080 dmu_buf_fill_done(db, tx); 1081 1082 if (err) 1083 break; 1084 1085 size -= tocpy; 1086 } 1087 1088 dmu_buf_rele_array(dbp, numbufs, FTAG); 1089 return (err); 1090} 1091 1092int 1093dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size, 1094 dmu_tx_t *tx) 1095{ 1096 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb; 1097 dnode_t *dn; 1098 int err; 1099 1100 if (size == 0) 1101 return (0); 1102 1103 DB_DNODE_ENTER(db); 1104 dn = DB_DNODE(db); 1105 err = dmu_write_uio_dnode(dn, uio, size, tx); 1106 DB_DNODE_EXIT(db); 1107 1108 return (err); 1109} 1110 1111int 1112dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size, 1113 dmu_tx_t *tx) 1114{ 1115 dnode_t *dn; 1116 int err; 1117 1118 if (size == 0) 1119 return (0); 1120 1121 err = dnode_hold(os, object, FTAG, &dn); 1122 if (err) 1123 return (err); 1124 1125 err = dmu_write_uio_dnode(dn, uio, size, tx); 1126 1127 dnode_rele(dn, FTAG); 1128 1129 return (err); 1130} 1131 1132#ifdef sun 1133int 1134dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1135 page_t *pp, dmu_tx_t *tx) 1136{ 1137 dmu_buf_t **dbp; 1138 int numbufs, i; 1139 int err; 1140 1141 if (size == 0) 1142 return (0); 1143 1144 err = dmu_buf_hold_array(os, object, offset, size, 1145 FALSE, FTAG, &numbufs, &dbp); 1146 if (err) 1147 return (err); 1148 1149 for (i = 0; i < numbufs; i++) { 1150 int tocpy, copied, thiscpy; 1151 int bufoff; 1152 dmu_buf_t *db = dbp[i]; 1153 caddr_t va; 1154 1155 ASSERT(size > 0); 1156 ASSERT3U(db->db_size, >=, PAGESIZE); 1157 1158 bufoff = offset - db->db_offset; 1159 tocpy = (int)MIN(db->db_size - bufoff, size); 1160 1161 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 1162 1163 if (tocpy == db->db_size) 1164 dmu_buf_will_fill(db, tx); 1165 else 1166 dmu_buf_will_dirty(db, tx); 1167 1168 for (copied = 0; copied < tocpy; copied += PAGESIZE) { 1169 ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff); 1170 thiscpy = MIN(PAGESIZE, tocpy - copied); 1171 va = zfs_map_page(pp, S_READ); 1172 bcopy(va, (char *)db->db_data + bufoff, thiscpy); 1173 zfs_unmap_page(pp, va); 1174 pp = pp->p_next; 1175 bufoff += PAGESIZE; 1176 } 1177 1178 if (tocpy == db->db_size) 1179 dmu_buf_fill_done(db, tx); 1180 1181 offset += tocpy; 1182 size -= tocpy; 1183 } 1184 dmu_buf_rele_array(dbp, numbufs, FTAG); 1185 return (err); 1186} 1187#endif /* sun */ 1188#endif 1189 1190/* 1191 * Allocate a loaned anonymous arc buffer. 1192 */ 1193arc_buf_t * 1194dmu_request_arcbuf(dmu_buf_t *handle, int size) 1195{ 1196 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle; 1197 spa_t *spa; 1198 1199 DB_GET_SPA(&spa, db); 1200 return (arc_loan_buf(spa, size)); 1201} 1202 1203/* 1204 * Free a loaned arc buffer. 1205 */ 1206void 1207dmu_return_arcbuf(arc_buf_t *buf) 1208{ 1209 arc_return_buf(buf, FTAG); 1210 VERIFY(arc_buf_remove_ref(buf, FTAG)); 1211} 1212 1213/* 1214 * When possible directly assign passed loaned arc buffer to a dbuf. 1215 * If this is not possible copy the contents of passed arc buf via 1216 * dmu_write(). 1217 */ 1218void 1219dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf, 1220 dmu_tx_t *tx) 1221{ 1222 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle; 1223 dnode_t *dn; 1224 dmu_buf_impl_t *db; 1225 uint32_t blksz = (uint32_t)arc_buf_size(buf); 1226 uint64_t blkid; 1227 1228 DB_DNODE_ENTER(dbuf); 1229 dn = DB_DNODE(dbuf); 1230 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1231 blkid = dbuf_whichblock(dn, offset); 1232 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL); 1233 rw_exit(&dn->dn_struct_rwlock); 1234 DB_DNODE_EXIT(dbuf); 1235 1236 if (offset == db->db.db_offset && blksz == db->db.db_size) { 1237 dbuf_assign_arcbuf(db, buf, tx); 1238 dbuf_rele(db, FTAG); 1239 } else { 1240 objset_t *os; 1241 uint64_t object; 1242 1243 DB_DNODE_ENTER(dbuf); 1244 dn = DB_DNODE(dbuf); 1245 os = dn->dn_objset; 1246 object = dn->dn_object; 1247 DB_DNODE_EXIT(dbuf); 1248 1249 dbuf_rele(db, FTAG); 1250 dmu_write(os, object, offset, blksz, buf->b_data, tx); 1251 dmu_return_arcbuf(buf); 1252 XUIOSTAT_BUMP(xuiostat_wbuf_copied); 1253 } 1254} 1255 1256typedef struct { 1257 dbuf_dirty_record_t *dsa_dr; 1258 dmu_sync_cb_t *dsa_done; 1259 zgd_t *dsa_zgd; 1260 dmu_tx_t *dsa_tx; 1261} dmu_sync_arg_t; 1262 1263/* ARGSUSED */ 1264static void 1265dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg) 1266{ 1267 dmu_sync_arg_t *dsa = varg; 1268 dmu_buf_t *db = dsa->dsa_zgd->zgd_db; 1269 blkptr_t *bp = zio->io_bp; 1270 1271 if (zio->io_error == 0) { 1272 if (BP_IS_HOLE(bp)) { 1273 /* 1274 * A block of zeros may compress to a hole, but the 1275 * block size still needs to be known for replay. 1276 */ 1277 BP_SET_LSIZE(bp, db->db_size); 1278 } else { 1279 ASSERT(BP_GET_LEVEL(bp) == 0); 1280 bp->blk_fill = 1; 1281 } 1282 } 1283} 1284 1285static void 1286dmu_sync_late_arrival_ready(zio_t *zio) 1287{ 1288 dmu_sync_ready(zio, NULL, zio->io_private); 1289} 1290 1291/* ARGSUSED */ 1292static void 1293dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg) 1294{ 1295 dmu_sync_arg_t *dsa = varg; 1296 dbuf_dirty_record_t *dr = dsa->dsa_dr; 1297 dmu_buf_impl_t *db = dr->dr_dbuf; 1298 1299 mutex_enter(&db->db_mtx); 1300 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC); 1301 if (zio->io_error == 0) { 1302 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE); 1303 if (dr->dt.dl.dr_nopwrite) { 1304 blkptr_t *bp = zio->io_bp; 1305 blkptr_t *bp_orig = &zio->io_bp_orig; 1306 uint8_t chksum = BP_GET_CHECKSUM(bp_orig); 1307 1308 ASSERT(BP_EQUAL(bp, bp_orig)); 1309 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF); 1310 ASSERT(zio_checksum_table[chksum].ci_dedup); 1311 } 1312 dr->dt.dl.dr_overridden_by = *zio->io_bp; 1313 dr->dt.dl.dr_override_state = DR_OVERRIDDEN; 1314 dr->dt.dl.dr_copies = zio->io_prop.zp_copies; 1315 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by)) 1316 BP_ZERO(&dr->dt.dl.dr_overridden_by); 1317 } else { 1318 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 1319 } 1320 cv_broadcast(&db->db_changed); 1321 mutex_exit(&db->db_mtx); 1322 1323 dsa->dsa_done(dsa->dsa_zgd, zio->io_error); 1324 1325 kmem_free(dsa, sizeof (*dsa)); 1326} 1327 1328static void 1329dmu_sync_late_arrival_done(zio_t *zio) 1330{ 1331 blkptr_t *bp = zio->io_bp; 1332 dmu_sync_arg_t *dsa = zio->io_private; 1333 blkptr_t *bp_orig = &zio->io_bp_orig; 1334 1335 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) { 1336 /* 1337 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE) 1338 * then there is nothing to do here. Otherwise, free the 1339 * newly allocated block in this txg. 1340 */ 1341 if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 1342 ASSERT(BP_EQUAL(bp, bp_orig)); 1343 } else { 1344 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig)); 1345 ASSERT(zio->io_bp->blk_birth == zio->io_txg); 1346 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa)); 1347 zio_free(zio->io_spa, zio->io_txg, zio->io_bp); 1348 } 1349 } 1350 1351 dmu_tx_commit(dsa->dsa_tx); 1352 1353 dsa->dsa_done(dsa->dsa_zgd, zio->io_error); 1354 1355 kmem_free(dsa, sizeof (*dsa)); 1356} 1357 1358static int 1359dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd, 1360 zio_prop_t *zp, zbookmark_t *zb) 1361{ 1362 dmu_sync_arg_t *dsa; 1363 dmu_tx_t *tx; 1364 1365 tx = dmu_tx_create(os); 1366 dmu_tx_hold_space(tx, zgd->zgd_db->db_size); 1367 if (dmu_tx_assign(tx, TXG_WAIT) != 0) { 1368 dmu_tx_abort(tx); 1369 /* Make zl_get_data do txg_waited_synced() */ 1370 return (SET_ERROR(EIO)); 1371 } 1372 1373 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP); 1374 dsa->dsa_dr = NULL; 1375 dsa->dsa_done = done; 1376 dsa->dsa_zgd = zgd; 1377 dsa->dsa_tx = tx; 1378 1379 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp, 1380 zgd->zgd_db->db_data, zgd->zgd_db->db_size, zp, 1381 dmu_sync_late_arrival_ready, NULL, dmu_sync_late_arrival_done, dsa, 1382 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb)); 1383 1384 return (0); 1385} 1386 1387/* 1388 * Intent log support: sync the block associated with db to disk. 1389 * N.B. and XXX: the caller is responsible for making sure that the 1390 * data isn't changing while dmu_sync() is writing it. 1391 * 1392 * Return values: 1393 * 1394 * EEXIST: this txg has already been synced, so there's nothing to do. 1395 * The caller should not log the write. 1396 * 1397 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do. 1398 * The caller should not log the write. 1399 * 1400 * EALREADY: this block is already in the process of being synced. 1401 * The caller should track its progress (somehow). 1402 * 1403 * EIO: could not do the I/O. 1404 * The caller should do a txg_wait_synced(). 1405 * 1406 * 0: the I/O has been initiated. 1407 * The caller should log this blkptr in the done callback. 1408 * It is possible that the I/O will fail, in which case 1409 * the error will be reported to the done callback and 1410 * propagated to pio from zio_done(). 1411 */ 1412int 1413dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd) 1414{ 1415 blkptr_t *bp = zgd->zgd_bp; 1416 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db; 1417 objset_t *os = db->db_objset; 1418 dsl_dataset_t *ds = os->os_dsl_dataset; 1419 dbuf_dirty_record_t *dr; 1420 dmu_sync_arg_t *dsa; 1421 zbookmark_t zb; 1422 zio_prop_t zp; 1423 dnode_t *dn; 1424 1425 ASSERT(pio != NULL); 1426 ASSERT(txg != 0); 1427 1428 SET_BOOKMARK(&zb, ds->ds_object, 1429 db->db.db_object, db->db_level, db->db_blkid); 1430 1431 DB_DNODE_ENTER(db); 1432 dn = DB_DNODE(db); 1433 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp); 1434 DB_DNODE_EXIT(db); 1435 1436 /* 1437 * If we're frozen (running ziltest), we always need to generate a bp. 1438 */ 1439 if (txg > spa_freeze_txg(os->os_spa)) 1440 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); 1441 1442 /* 1443 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf() 1444 * and us. If we determine that this txg is not yet syncing, 1445 * but it begins to sync a moment later, that's OK because the 1446 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx. 1447 */ 1448 mutex_enter(&db->db_mtx); 1449 1450 if (txg <= spa_last_synced_txg(os->os_spa)) { 1451 /* 1452 * This txg has already synced. There's nothing to do. 1453 */ 1454 mutex_exit(&db->db_mtx); 1455 return (SET_ERROR(EEXIST)); 1456 } 1457 1458 if (txg <= spa_syncing_txg(os->os_spa)) { 1459 /* 1460 * This txg is currently syncing, so we can't mess with 1461 * the dirty record anymore; just write a new log block. 1462 */ 1463 mutex_exit(&db->db_mtx); 1464 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); 1465 } 1466 1467 dr = db->db_last_dirty; 1468 while (dr && dr->dr_txg != txg) 1469 dr = dr->dr_next; 1470 1471 if (dr == NULL) { 1472 /* 1473 * There's no dr for this dbuf, so it must have been freed. 1474 * There's no need to log writes to freed blocks, so we're done. 1475 */ 1476 mutex_exit(&db->db_mtx); 1477 return (SET_ERROR(ENOENT)); 1478 } 1479 1480 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg); 1481 1482 /* 1483 * Assume the on-disk data is X, the current syncing data is Y, 1484 * and the current in-memory data is Z (currently in dmu_sync). 1485 * X and Z are identical but Y is has been modified. Normally, 1486 * when X and Z are the same we will perform a nopwrite but if Y 1487 * is different we must disable nopwrite since the resulting write 1488 * of Y to disk can free the block containing X. If we allowed a 1489 * nopwrite to occur the block pointing to Z would reference a freed 1490 * block. Since this is a rare case we simplify this by disabling 1491 * nopwrite if the current dmu_sync-ing dbuf has been modified in 1492 * a previous transaction. 1493 */ 1494 if (dr->dr_next) 1495 zp.zp_nopwrite = B_FALSE; 1496 1497 ASSERT(dr->dr_txg == txg); 1498 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC || 1499 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { 1500 /* 1501 * We have already issued a sync write for this buffer, 1502 * or this buffer has already been synced. It could not 1503 * have been dirtied since, or we would have cleared the state. 1504 */ 1505 mutex_exit(&db->db_mtx); 1506 return (SET_ERROR(EALREADY)); 1507 } 1508 1509 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); 1510 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC; 1511 mutex_exit(&db->db_mtx); 1512 1513 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP); 1514 dsa->dsa_dr = dr; 1515 dsa->dsa_done = done; 1516 dsa->dsa_zgd = zgd; 1517 dsa->dsa_tx = NULL; 1518 1519 zio_nowait(arc_write(pio, os->os_spa, txg, 1520 bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db), 1521 DBUF_IS_L2COMPRESSIBLE(db), &zp, dmu_sync_ready, 1522 NULL, dmu_sync_done, dsa, ZIO_PRIORITY_SYNC_WRITE, 1523 ZIO_FLAG_CANFAIL, &zb)); 1524 1525 return (0); 1526} 1527 1528int 1529dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs, 1530 dmu_tx_t *tx) 1531{ 1532 dnode_t *dn; 1533 int err; 1534 1535 err = dnode_hold(os, object, FTAG, &dn); 1536 if (err) 1537 return (err); 1538 err = dnode_set_blksz(dn, size, ibs, tx); 1539 dnode_rele(dn, FTAG); 1540 return (err); 1541} 1542 1543void 1544dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum, 1545 dmu_tx_t *tx) 1546{ 1547 dnode_t *dn; 1548 1549 /* XXX assumes dnode_hold will not get an i/o error */ 1550 (void) dnode_hold(os, object, FTAG, &dn); 1551 ASSERT(checksum < ZIO_CHECKSUM_FUNCTIONS); 1552 dn->dn_checksum = checksum; 1553 dnode_setdirty(dn, tx); 1554 dnode_rele(dn, FTAG); 1555} 1556 1557void 1558dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress, 1559 dmu_tx_t *tx) 1560{ 1561 dnode_t *dn; 1562 1563 /* XXX assumes dnode_hold will not get an i/o error */ 1564 (void) dnode_hold(os, object, FTAG, &dn); 1565 ASSERT(compress < ZIO_COMPRESS_FUNCTIONS); 1566 dn->dn_compress = compress; 1567 dnode_setdirty(dn, tx); 1568 dnode_rele(dn, FTAG); 1569} 1570 1571int zfs_mdcomp_disable = 0; 1572TUNABLE_INT("vfs.zfs.mdcomp_disable", &zfs_mdcomp_disable); 1573SYSCTL_INT(_vfs_zfs, OID_AUTO, mdcomp_disable, CTLFLAG_RW, 1574 &zfs_mdcomp_disable, 0, "Disable metadata compression"); 1575 1576void 1577dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp) 1578{ 1579 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET; 1580 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) || 1581 (wp & WP_SPILL)); 1582 enum zio_checksum checksum = os->os_checksum; 1583 enum zio_compress compress = os->os_compress; 1584 enum zio_checksum dedup_checksum = os->os_dedup_checksum; 1585 boolean_t dedup = B_FALSE; 1586 boolean_t nopwrite = B_FALSE; 1587 boolean_t dedup_verify = os->os_dedup_verify; 1588 int copies = os->os_copies; 1589 1590 /* 1591 * We maintain different write policies for each of the following 1592 * types of data: 1593 * 1. metadata 1594 * 2. preallocated blocks (i.e. level-0 blocks of a dump device) 1595 * 3. all other level 0 blocks 1596 */ 1597 if (ismd) { 1598 /* 1599 * XXX -- we should design a compression algorithm 1600 * that specializes in arrays of bps. 1601 */ 1602 compress = zfs_mdcomp_disable ? ZIO_COMPRESS_EMPTY : 1603 ZIO_COMPRESS_LZJB; 1604 1605 /* 1606 * Metadata always gets checksummed. If the data 1607 * checksum is multi-bit correctable, and it's not a 1608 * ZBT-style checksum, then it's suitable for metadata 1609 * as well. Otherwise, the metadata checksum defaults 1610 * to fletcher4. 1611 */ 1612 if (zio_checksum_table[checksum].ci_correctable < 1 || 1613 zio_checksum_table[checksum].ci_eck) 1614 checksum = ZIO_CHECKSUM_FLETCHER_4; 1615 } else if (wp & WP_NOFILL) { 1616 ASSERT(level == 0); 1617 1618 /* 1619 * If we're writing preallocated blocks, we aren't actually 1620 * writing them so don't set any policy properties. These 1621 * blocks are currently only used by an external subsystem 1622 * outside of zfs (i.e. dump) and not written by the zio 1623 * pipeline. 1624 */ 1625 compress = ZIO_COMPRESS_OFF; 1626 checksum = ZIO_CHECKSUM_NOPARITY; 1627 } else { 1628 compress = zio_compress_select(dn->dn_compress, compress); 1629 1630 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ? 1631 zio_checksum_select(dn->dn_checksum, checksum) : 1632 dedup_checksum; 1633 1634 /* 1635 * Determine dedup setting. If we are in dmu_sync(), 1636 * we won't actually dedup now because that's all 1637 * done in syncing context; but we do want to use the 1638 * dedup checkum. If the checksum is not strong 1639 * enough to ensure unique signatures, force 1640 * dedup_verify. 1641 */ 1642 if (dedup_checksum != ZIO_CHECKSUM_OFF) { 1643 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE; 1644 if (!zio_checksum_table[checksum].ci_dedup) 1645 dedup_verify = B_TRUE; 1646 } 1647 1648 /* 1649 * Enable nopwrite if we have a cryptographically secure 1650 * checksum that has no known collisions (i.e. SHA-256) 1651 * and compression is enabled. We don't enable nopwrite if 1652 * dedup is enabled as the two features are mutually exclusive. 1653 */ 1654 nopwrite = (!dedup && zio_checksum_table[checksum].ci_dedup && 1655 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled); 1656 } 1657 1658 zp->zp_checksum = checksum; 1659 zp->zp_compress = compress; 1660 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type; 1661 zp->zp_level = level; 1662 zp->zp_copies = MIN(copies + ismd, spa_max_replication(os->os_spa)); 1663 zp->zp_dedup = dedup; 1664 zp->zp_dedup_verify = dedup && dedup_verify; 1665 zp->zp_nopwrite = nopwrite; 1666} 1667 1668int 1669dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off) 1670{ 1671 dnode_t *dn; 1672 int i, err; 1673 1674 err = dnode_hold(os, object, FTAG, &dn); 1675 if (err) 1676 return (err); 1677 /* 1678 * Sync any current changes before 1679 * we go trundling through the block pointers. 1680 */ 1681 for (i = 0; i < TXG_SIZE; i++) { 1682 if (list_link_active(&dn->dn_dirty_link[i])) 1683 break; 1684 } 1685 if (i != TXG_SIZE) { 1686 dnode_rele(dn, FTAG); 1687 txg_wait_synced(dmu_objset_pool(os), 0); 1688 err = dnode_hold(os, object, FTAG, &dn); 1689 if (err) 1690 return (err); 1691 } 1692 1693 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0); 1694 dnode_rele(dn, FTAG); 1695 1696 return (err); 1697} 1698 1699void 1700dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi) 1701{ 1702 dnode_phys_t *dnp; 1703 1704 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1705 mutex_enter(&dn->dn_mtx); 1706 1707 dnp = dn->dn_phys; 1708 1709 doi->doi_data_block_size = dn->dn_datablksz; 1710 doi->doi_metadata_block_size = dn->dn_indblkshift ? 1711 1ULL << dn->dn_indblkshift : 0; 1712 doi->doi_type = dn->dn_type; 1713 doi->doi_bonus_type = dn->dn_bonustype; 1714 doi->doi_bonus_size = dn->dn_bonuslen; 1715 doi->doi_indirection = dn->dn_nlevels; 1716 doi->doi_checksum = dn->dn_checksum; 1717 doi->doi_compress = dn->dn_compress; 1718 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9; 1719 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz; 1720 doi->doi_fill_count = 0; 1721 for (int i = 0; i < dnp->dn_nblkptr; i++) 1722 doi->doi_fill_count += dnp->dn_blkptr[i].blk_fill; 1723 1724 mutex_exit(&dn->dn_mtx); 1725 rw_exit(&dn->dn_struct_rwlock); 1726} 1727 1728/* 1729 * Get information on a DMU object. 1730 * If doi is NULL, just indicates whether the object exists. 1731 */ 1732int 1733dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi) 1734{ 1735 dnode_t *dn; 1736 int err = dnode_hold(os, object, FTAG, &dn); 1737 1738 if (err) 1739 return (err); 1740 1741 if (doi != NULL) 1742 dmu_object_info_from_dnode(dn, doi); 1743 1744 dnode_rele(dn, FTAG); 1745 return (0); 1746} 1747 1748/* 1749 * As above, but faster; can be used when you have a held dbuf in hand. 1750 */ 1751void 1752dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi) 1753{ 1754 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1755 1756 DB_DNODE_ENTER(db); 1757 dmu_object_info_from_dnode(DB_DNODE(db), doi); 1758 DB_DNODE_EXIT(db); 1759} 1760 1761/* 1762 * Faster still when you only care about the size. 1763 * This is specifically optimized for zfs_getattr(). 1764 */ 1765void 1766dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize, 1767 u_longlong_t *nblk512) 1768{ 1769 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1770 dnode_t *dn; 1771 1772 DB_DNODE_ENTER(db); 1773 dn = DB_DNODE(db); 1774 1775 *blksize = dn->dn_datablksz; 1776 /* add 1 for dnode space */ 1777 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >> 1778 SPA_MINBLOCKSHIFT) + 1; 1779 DB_DNODE_EXIT(db); 1780} 1781 1782void 1783byteswap_uint64_array(void *vbuf, size_t size) 1784{ 1785 uint64_t *buf = vbuf; 1786 size_t count = size >> 3; 1787 int i; 1788 1789 ASSERT((size & 7) == 0); 1790 1791 for (i = 0; i < count; i++) 1792 buf[i] = BSWAP_64(buf[i]); 1793} 1794 1795void 1796byteswap_uint32_array(void *vbuf, size_t size) 1797{ 1798 uint32_t *buf = vbuf; 1799 size_t count = size >> 2; 1800 int i; 1801 1802 ASSERT((size & 3) == 0); 1803 1804 for (i = 0; i < count; i++) 1805 buf[i] = BSWAP_32(buf[i]); 1806} 1807 1808void 1809byteswap_uint16_array(void *vbuf, size_t size) 1810{ 1811 uint16_t *buf = vbuf; 1812 size_t count = size >> 1; 1813 int i; 1814 1815 ASSERT((size & 1) == 0); 1816 1817 for (i = 0; i < count; i++) 1818 buf[i] = BSWAP_16(buf[i]); 1819} 1820 1821/* ARGSUSED */ 1822void 1823byteswap_uint8_array(void *vbuf, size_t size) 1824{ 1825} 1826 1827void 1828dmu_init(void) 1829{ 1830 zfs_dbgmsg_init(); 1831 sa_cache_init(); 1832 xuio_stat_init(); 1833 dmu_objset_init(); 1834 dnode_init(); 1835 dbuf_init(); 1836 zfetch_init(); 1837 zio_compress_init(); 1838 l2arc_init(); 1839 arc_init(); 1840} 1841 1842void 1843dmu_fini(void) 1844{ 1845 arc_fini(); /* arc depends on l2arc, so arc must go first */ 1846 l2arc_fini(); 1847 zfetch_fini(); 1848 zio_compress_fini(); 1849 dbuf_fini(); 1850 dnode_fini(); 1851 dmu_objset_fini(); 1852 xuio_stat_fini(); 1853 sa_cache_fini(); 1854 zfs_dbgmsg_fini(); 1855} 1856