dmu.c revision 263397
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/vm.h> 49#include <sys/zfs_znode.h> 50#endif 51 52/* 53 * Enable/disable nopwrite feature. 54 */ 55int zfs_nopwrite_enabled = 1; 56SYSCTL_DECL(_vfs_zfs); 57TUNABLE_INT("vfs.zfs.nopwrite_enabled", &zfs_nopwrite_enabled); 58SYSCTL_INT(_vfs_zfs, OID_AUTO, nopwrite_enabled, CTLFLAG_RDTUN, 59 &zfs_nopwrite_enabled, 0, "Enable nopwrite feature"); 60 61const dmu_object_type_info_t dmu_ot[DMU_OT_NUMTYPES] = { 62 { DMU_BSWAP_UINT8, TRUE, "unallocated" }, 63 { DMU_BSWAP_ZAP, TRUE, "object directory" }, 64 { DMU_BSWAP_UINT64, TRUE, "object array" }, 65 { DMU_BSWAP_UINT8, TRUE, "packed nvlist" }, 66 { DMU_BSWAP_UINT64, TRUE, "packed nvlist size" }, 67 { DMU_BSWAP_UINT64, TRUE, "bpobj" }, 68 { DMU_BSWAP_UINT64, TRUE, "bpobj header" }, 69 { DMU_BSWAP_UINT64, TRUE, "SPA space map header" }, 70 { DMU_BSWAP_UINT64, TRUE, "SPA space map" }, 71 { DMU_BSWAP_UINT64, TRUE, "ZIL intent log" }, 72 { DMU_BSWAP_DNODE, TRUE, "DMU dnode" }, 73 { DMU_BSWAP_OBJSET, TRUE, "DMU objset" }, 74 { DMU_BSWAP_UINT64, TRUE, "DSL directory" }, 75 { DMU_BSWAP_ZAP, TRUE, "DSL directory child map"}, 76 { DMU_BSWAP_ZAP, TRUE, "DSL dataset snap map" }, 77 { DMU_BSWAP_ZAP, TRUE, "DSL props" }, 78 { DMU_BSWAP_UINT64, TRUE, "DSL dataset" }, 79 { DMU_BSWAP_ZNODE, TRUE, "ZFS znode" }, 80 { DMU_BSWAP_OLDACL, TRUE, "ZFS V0 ACL" }, 81 { DMU_BSWAP_UINT8, FALSE, "ZFS plain file" }, 82 { DMU_BSWAP_ZAP, TRUE, "ZFS directory" }, 83 { DMU_BSWAP_ZAP, TRUE, "ZFS master node" }, 84 { DMU_BSWAP_ZAP, TRUE, "ZFS delete queue" }, 85 { DMU_BSWAP_UINT8, FALSE, "zvol object" }, 86 { DMU_BSWAP_ZAP, TRUE, "zvol prop" }, 87 { DMU_BSWAP_UINT8, FALSE, "other uint8[]" }, 88 { DMU_BSWAP_UINT64, FALSE, "other uint64[]" }, 89 { DMU_BSWAP_ZAP, TRUE, "other ZAP" }, 90 { DMU_BSWAP_ZAP, TRUE, "persistent error log" }, 91 { DMU_BSWAP_UINT8, TRUE, "SPA history" }, 92 { DMU_BSWAP_UINT64, TRUE, "SPA history offsets" }, 93 { DMU_BSWAP_ZAP, TRUE, "Pool properties" }, 94 { DMU_BSWAP_ZAP, TRUE, "DSL permissions" }, 95 { DMU_BSWAP_ACL, TRUE, "ZFS ACL" }, 96 { DMU_BSWAP_UINT8, TRUE, "ZFS SYSACL" }, 97 { DMU_BSWAP_UINT8, TRUE, "FUID table" }, 98 { DMU_BSWAP_UINT64, TRUE, "FUID table size" }, 99 { DMU_BSWAP_ZAP, TRUE, "DSL dataset next clones"}, 100 { DMU_BSWAP_ZAP, TRUE, "scan work queue" }, 101 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group used" }, 102 { DMU_BSWAP_ZAP, TRUE, "ZFS user/group quota" }, 103 { DMU_BSWAP_ZAP, TRUE, "snapshot refcount tags"}, 104 { DMU_BSWAP_ZAP, TRUE, "DDT ZAP algorithm" }, 105 { DMU_BSWAP_ZAP, TRUE, "DDT statistics" }, 106 { DMU_BSWAP_UINT8, TRUE, "System attributes" }, 107 { DMU_BSWAP_ZAP, TRUE, "SA master node" }, 108 { DMU_BSWAP_ZAP, TRUE, "SA attr registration" }, 109 { DMU_BSWAP_ZAP, TRUE, "SA attr layouts" }, 110 { DMU_BSWAP_ZAP, TRUE, "scan translations" }, 111 { DMU_BSWAP_UINT8, FALSE, "deduplicated block" }, 112 { DMU_BSWAP_ZAP, TRUE, "DSL deadlist map" }, 113 { DMU_BSWAP_UINT64, TRUE, "DSL deadlist map hdr" }, 114 { DMU_BSWAP_ZAP, TRUE, "DSL dir clones" }, 115 { DMU_BSWAP_UINT64, TRUE, "bpobj subobj" } 116}; 117 118const dmu_object_byteswap_info_t dmu_ot_byteswap[DMU_BSWAP_NUMFUNCS] = { 119 { byteswap_uint8_array, "uint8" }, 120 { byteswap_uint16_array, "uint16" }, 121 { byteswap_uint32_array, "uint32" }, 122 { byteswap_uint64_array, "uint64" }, 123 { zap_byteswap, "zap" }, 124 { dnode_buf_byteswap, "dnode" }, 125 { dmu_objset_byteswap, "objset" }, 126 { zfs_znode_byteswap, "znode" }, 127 { zfs_oldacl_byteswap, "oldacl" }, 128 { zfs_acl_byteswap, "acl" } 129}; 130 131int 132dmu_buf_hold(objset_t *os, uint64_t object, uint64_t offset, 133 void *tag, dmu_buf_t **dbp, int flags) 134{ 135 dnode_t *dn; 136 uint64_t blkid; 137 dmu_buf_impl_t *db; 138 int err; 139 int db_flags = DB_RF_CANFAIL; 140 141 if (flags & DMU_READ_NO_PREFETCH) 142 db_flags |= DB_RF_NOPREFETCH; 143 144 err = dnode_hold(os, object, FTAG, &dn); 145 if (err) 146 return (err); 147 blkid = dbuf_whichblock(dn, offset); 148 rw_enter(&dn->dn_struct_rwlock, RW_READER); 149 db = dbuf_hold(dn, blkid, tag); 150 rw_exit(&dn->dn_struct_rwlock); 151 if (db == NULL) { 152 err = SET_ERROR(EIO); 153 } else { 154 err = dbuf_read(db, NULL, db_flags); 155 if (err) { 156 dbuf_rele(db, tag); 157 db = NULL; 158 } 159 } 160 161 dnode_rele(dn, FTAG); 162 *dbp = &db->db; /* NULL db plus first field offset is NULL */ 163 return (err); 164} 165 166int 167dmu_bonus_max(void) 168{ 169 return (DN_MAX_BONUSLEN); 170} 171 172int 173dmu_set_bonus(dmu_buf_t *db_fake, int newsize, dmu_tx_t *tx) 174{ 175 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 176 dnode_t *dn; 177 int error; 178 179 DB_DNODE_ENTER(db); 180 dn = DB_DNODE(db); 181 182 if (dn->dn_bonus != db) { 183 error = SET_ERROR(EINVAL); 184 } else if (newsize < 0 || newsize > db_fake->db_size) { 185 error = SET_ERROR(EINVAL); 186 } else { 187 dnode_setbonuslen(dn, newsize, tx); 188 error = 0; 189 } 190 191 DB_DNODE_EXIT(db); 192 return (error); 193} 194 195int 196dmu_set_bonustype(dmu_buf_t *db_fake, dmu_object_type_t type, dmu_tx_t *tx) 197{ 198 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 199 dnode_t *dn; 200 int error; 201 202 DB_DNODE_ENTER(db); 203 dn = DB_DNODE(db); 204 205 if (!DMU_OT_IS_VALID(type)) { 206 error = SET_ERROR(EINVAL); 207 } else if (dn->dn_bonus != db) { 208 error = SET_ERROR(EINVAL); 209 } else { 210 dnode_setbonus_type(dn, type, tx); 211 error = 0; 212 } 213 214 DB_DNODE_EXIT(db); 215 return (error); 216} 217 218dmu_object_type_t 219dmu_get_bonustype(dmu_buf_t *db_fake) 220{ 221 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 222 dnode_t *dn; 223 dmu_object_type_t type; 224 225 DB_DNODE_ENTER(db); 226 dn = DB_DNODE(db); 227 type = dn->dn_bonustype; 228 DB_DNODE_EXIT(db); 229 230 return (type); 231} 232 233int 234dmu_rm_spill(objset_t *os, uint64_t object, dmu_tx_t *tx) 235{ 236 dnode_t *dn; 237 int error; 238 239 error = dnode_hold(os, object, FTAG, &dn); 240 dbuf_rm_spill(dn, tx); 241 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 242 dnode_rm_spill(dn, tx); 243 rw_exit(&dn->dn_struct_rwlock); 244 dnode_rele(dn, FTAG); 245 return (error); 246} 247 248/* 249 * returns ENOENT, EIO, or 0. 250 */ 251int 252dmu_bonus_hold(objset_t *os, uint64_t object, void *tag, dmu_buf_t **dbp) 253{ 254 dnode_t *dn; 255 dmu_buf_impl_t *db; 256 int error; 257 258 error = dnode_hold(os, object, FTAG, &dn); 259 if (error) 260 return (error); 261 262 rw_enter(&dn->dn_struct_rwlock, RW_READER); 263 if (dn->dn_bonus == NULL) { 264 rw_exit(&dn->dn_struct_rwlock); 265 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 266 if (dn->dn_bonus == NULL) 267 dbuf_create_bonus(dn); 268 } 269 db = dn->dn_bonus; 270 271 /* as long as the bonus buf is held, the dnode will be held */ 272 if (refcount_add(&db->db_holds, tag) == 1) { 273 VERIFY(dnode_add_ref(dn, db)); 274 (void) atomic_inc_32_nv(&dn->dn_dbufs_count); 275 } 276 277 /* 278 * Wait to drop dn_struct_rwlock until after adding the bonus dbuf's 279 * hold and incrementing the dbuf count to ensure that dnode_move() sees 280 * a dnode hold for every dbuf. 281 */ 282 rw_exit(&dn->dn_struct_rwlock); 283 284 dnode_rele(dn, FTAG); 285 286 VERIFY(0 == dbuf_read(db, NULL, DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH)); 287 288 *dbp = &db->db; 289 return (0); 290} 291 292/* 293 * returns ENOENT, EIO, or 0. 294 * 295 * This interface will allocate a blank spill dbuf when a spill blk 296 * doesn't already exist on the dnode. 297 * 298 * if you only want to find an already existing spill db, then 299 * dmu_spill_hold_existing() should be used. 300 */ 301int 302dmu_spill_hold_by_dnode(dnode_t *dn, uint32_t flags, void *tag, dmu_buf_t **dbp) 303{ 304 dmu_buf_impl_t *db = NULL; 305 int err; 306 307 if ((flags & DB_RF_HAVESTRUCT) == 0) 308 rw_enter(&dn->dn_struct_rwlock, RW_READER); 309 310 db = dbuf_hold(dn, DMU_SPILL_BLKID, tag); 311 312 if ((flags & DB_RF_HAVESTRUCT) == 0) 313 rw_exit(&dn->dn_struct_rwlock); 314 315 ASSERT(db != NULL); 316 err = dbuf_read(db, NULL, flags); 317 if (err == 0) 318 *dbp = &db->db; 319 else 320 dbuf_rele(db, tag); 321 return (err); 322} 323 324int 325dmu_spill_hold_existing(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp) 326{ 327 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus; 328 dnode_t *dn; 329 int err; 330 331 DB_DNODE_ENTER(db); 332 dn = DB_DNODE(db); 333 334 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_SA) { 335 err = SET_ERROR(EINVAL); 336 } else { 337 rw_enter(&dn->dn_struct_rwlock, RW_READER); 338 339 if (!dn->dn_have_spill) { 340 err = SET_ERROR(ENOENT); 341 } else { 342 err = dmu_spill_hold_by_dnode(dn, 343 DB_RF_HAVESTRUCT | DB_RF_CANFAIL, tag, dbp); 344 } 345 346 rw_exit(&dn->dn_struct_rwlock); 347 } 348 349 DB_DNODE_EXIT(db); 350 return (err); 351} 352 353int 354dmu_spill_hold_by_bonus(dmu_buf_t *bonus, void *tag, dmu_buf_t **dbp) 355{ 356 dmu_buf_impl_t *db = (dmu_buf_impl_t *)bonus; 357 dnode_t *dn; 358 int err; 359 360 DB_DNODE_ENTER(db); 361 dn = DB_DNODE(db); 362 err = dmu_spill_hold_by_dnode(dn, DB_RF_CANFAIL, tag, dbp); 363 DB_DNODE_EXIT(db); 364 365 return (err); 366} 367 368/* 369 * Note: longer-term, we should modify all of the dmu_buf_*() interfaces 370 * to take a held dnode rather than <os, object> -- the lookup is wasteful, 371 * and can induce severe lock contention when writing to several files 372 * whose dnodes are in the same block. 373 */ 374static int 375dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length, 376 int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp, uint32_t flags) 377{ 378 dmu_buf_t **dbp; 379 uint64_t blkid, nblks, i; 380 uint32_t dbuf_flags; 381 int err; 382 zio_t *zio; 383 384 ASSERT(length <= DMU_MAX_ACCESS); 385 386 dbuf_flags = DB_RF_CANFAIL | DB_RF_NEVERWAIT | DB_RF_HAVESTRUCT; 387 if (flags & DMU_READ_NO_PREFETCH || length > zfetch_array_rd_sz) 388 dbuf_flags |= DB_RF_NOPREFETCH; 389 390 rw_enter(&dn->dn_struct_rwlock, RW_READER); 391 if (dn->dn_datablkshift) { 392 int blkshift = dn->dn_datablkshift; 393 nblks = (P2ROUNDUP(offset+length, 1ULL<<blkshift) - 394 P2ALIGN(offset, 1ULL<<blkshift)) >> blkshift; 395 } else { 396 if (offset + length > dn->dn_datablksz) { 397 zfs_panic_recover("zfs: accessing past end of object " 398 "%llx/%llx (size=%u access=%llu+%llu)", 399 (longlong_t)dn->dn_objset-> 400 os_dsl_dataset->ds_object, 401 (longlong_t)dn->dn_object, dn->dn_datablksz, 402 (longlong_t)offset, (longlong_t)length); 403 rw_exit(&dn->dn_struct_rwlock); 404 return (SET_ERROR(EIO)); 405 } 406 nblks = 1; 407 } 408 dbp = kmem_zalloc(sizeof (dmu_buf_t *) * nblks, KM_SLEEP); 409 410 zio = zio_root(dn->dn_objset->os_spa, NULL, NULL, ZIO_FLAG_CANFAIL); 411 blkid = dbuf_whichblock(dn, offset); 412 for (i = 0; i < nblks; i++) { 413 dmu_buf_impl_t *db = dbuf_hold(dn, blkid+i, tag); 414 if (db == NULL) { 415 rw_exit(&dn->dn_struct_rwlock); 416 dmu_buf_rele_array(dbp, nblks, tag); 417 zio_nowait(zio); 418 return (SET_ERROR(EIO)); 419 } 420 /* initiate async i/o */ 421 if (read) 422 (void) dbuf_read(db, zio, dbuf_flags); 423#ifdef _KERNEL 424 else 425 curthread->td_ru.ru_oublock++; 426#endif 427 dbp[i] = &db->db; 428 } 429 rw_exit(&dn->dn_struct_rwlock); 430 431 /* wait for async i/o */ 432 err = zio_wait(zio); 433 if (err) { 434 dmu_buf_rele_array(dbp, nblks, tag); 435 return (err); 436 } 437 438 /* wait for other io to complete */ 439 if (read) { 440 for (i = 0; i < nblks; i++) { 441 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbp[i]; 442 mutex_enter(&db->db_mtx); 443 while (db->db_state == DB_READ || 444 db->db_state == DB_FILL) 445 cv_wait(&db->db_changed, &db->db_mtx); 446 if (db->db_state == DB_UNCACHED) 447 err = SET_ERROR(EIO); 448 mutex_exit(&db->db_mtx); 449 if (err) { 450 dmu_buf_rele_array(dbp, nblks, tag); 451 return (err); 452 } 453 } 454 } 455 456 *numbufsp = nblks; 457 *dbpp = dbp; 458 return (0); 459} 460 461static int 462dmu_buf_hold_array(objset_t *os, uint64_t object, uint64_t offset, 463 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp) 464{ 465 dnode_t *dn; 466 int err; 467 468 err = dnode_hold(os, object, FTAG, &dn); 469 if (err) 470 return (err); 471 472 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag, 473 numbufsp, dbpp, DMU_READ_PREFETCH); 474 475 dnode_rele(dn, FTAG); 476 477 return (err); 478} 479 480int 481dmu_buf_hold_array_by_bonus(dmu_buf_t *db_fake, uint64_t offset, 482 uint64_t length, int read, void *tag, int *numbufsp, dmu_buf_t ***dbpp) 483{ 484 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 485 dnode_t *dn; 486 int err; 487 488 DB_DNODE_ENTER(db); 489 dn = DB_DNODE(db); 490 err = dmu_buf_hold_array_by_dnode(dn, offset, length, read, tag, 491 numbufsp, dbpp, DMU_READ_PREFETCH); 492 DB_DNODE_EXIT(db); 493 494 return (err); 495} 496 497void 498dmu_buf_rele_array(dmu_buf_t **dbp_fake, int numbufs, void *tag) 499{ 500 int i; 501 dmu_buf_impl_t **dbp = (dmu_buf_impl_t **)dbp_fake; 502 503 if (numbufs == 0) 504 return; 505 506 for (i = 0; i < numbufs; i++) { 507 if (dbp[i]) 508 dbuf_rele(dbp[i], tag); 509 } 510 511 kmem_free(dbp, sizeof (dmu_buf_t *) * numbufs); 512} 513 514/* 515 * Issue prefetch i/os for the given blocks. 516 * 517 * Note: The assumption is that we *know* these blocks will be needed 518 * almost immediately. Therefore, the prefetch i/os will be issued at 519 * ZIO_PRIORITY_SYNC_READ 520 * 521 * Note: indirect blocks and other metadata will be read synchronously, 522 * causing this function to block if they are not already cached. 523 */ 524void 525dmu_prefetch(objset_t *os, uint64_t object, uint64_t offset, uint64_t len) 526{ 527 dnode_t *dn; 528 uint64_t blkid; 529 int nblks, err; 530 531 if (zfs_prefetch_disable) 532 return; 533 534 if (len == 0) { /* they're interested in the bonus buffer */ 535 dn = DMU_META_DNODE(os); 536 537 if (object == 0 || object >= DN_MAX_OBJECT) 538 return; 539 540 rw_enter(&dn->dn_struct_rwlock, RW_READER); 541 blkid = dbuf_whichblock(dn, object * sizeof (dnode_phys_t)); 542 dbuf_prefetch(dn, blkid, ZIO_PRIORITY_SYNC_READ); 543 rw_exit(&dn->dn_struct_rwlock); 544 return; 545 } 546 547 /* 548 * XXX - Note, if the dnode for the requested object is not 549 * already cached, we will do a *synchronous* read in the 550 * dnode_hold() call. The same is true for any indirects. 551 */ 552 err = dnode_hold(os, object, FTAG, &dn); 553 if (err != 0) 554 return; 555 556 rw_enter(&dn->dn_struct_rwlock, RW_READER); 557 if (dn->dn_datablkshift) { 558 int blkshift = dn->dn_datablkshift; 559 nblks = (P2ROUNDUP(offset + len, 1 << blkshift) - 560 P2ALIGN(offset, 1 << blkshift)) >> blkshift; 561 } else { 562 nblks = (offset < dn->dn_datablksz); 563 } 564 565 if (nblks != 0) { 566 blkid = dbuf_whichblock(dn, offset); 567 for (int i = 0; i < nblks; i++) 568 dbuf_prefetch(dn, blkid + i, ZIO_PRIORITY_SYNC_READ); 569 } 570 571 rw_exit(&dn->dn_struct_rwlock); 572 573 dnode_rele(dn, FTAG); 574} 575 576/* 577 * Get the next "chunk" of file data to free. We traverse the file from 578 * the end so that the file gets shorter over time (if we crashes in the 579 * middle, this will leave us in a better state). We find allocated file 580 * data by simply searching the allocated level 1 indirects. 581 * 582 * On input, *start should be the first offset that does not need to be 583 * freed (e.g. "offset + length"). On return, *start will be the first 584 * offset that should be freed. 585 */ 586static int 587get_next_chunk(dnode_t *dn, uint64_t *start, uint64_t minimum) 588{ 589 uint64_t maxblks = DMU_MAX_ACCESS >> (dn->dn_indblkshift + 1); 590 /* bytes of data covered by a level-1 indirect block */ 591 uint64_t iblkrange = 592 dn->dn_datablksz * EPB(dn->dn_indblkshift, SPA_BLKPTRSHIFT); 593 594 ASSERT3U(minimum, <=, *start); 595 596 if (*start - minimum <= iblkrange * maxblks) { 597 *start = minimum; 598 return (0); 599 } 600 ASSERT(ISP2(iblkrange)); 601 602 for (uint64_t blks = 0; *start > minimum && blks < maxblks; blks++) { 603 int err; 604 605 /* 606 * dnode_next_offset(BACKWARDS) will find an allocated L1 607 * indirect block at or before the input offset. We must 608 * decrement *start so that it is at the end of the region 609 * to search. 610 */ 611 (*start)--; 612 err = dnode_next_offset(dn, 613 DNODE_FIND_BACKWARDS, start, 2, 1, 0); 614 615 /* if there are no indirect blocks before start, we are done */ 616 if (err == ESRCH) { 617 *start = minimum; 618 break; 619 } else if (err != 0) { 620 return (err); 621 } 622 623 /* set start to the beginning of this L1 indirect */ 624 *start = P2ALIGN(*start, iblkrange); 625 } 626 if (*start < minimum) 627 *start = minimum; 628 return (0); 629} 630 631static int 632dmu_free_long_range_impl(objset_t *os, dnode_t *dn, uint64_t offset, 633 uint64_t length) 634{ 635 uint64_t object_size = (dn->dn_maxblkid + 1) * dn->dn_datablksz; 636 int err; 637 638 if (offset >= object_size) 639 return (0); 640 641 if (length == DMU_OBJECT_END || offset + length > object_size) 642 length = object_size - offset; 643 644 while (length != 0) { 645 uint64_t chunk_end, chunk_begin; 646 647 chunk_end = chunk_begin = offset + length; 648 649 /* move chunk_begin backwards to the beginning of this chunk */ 650 err = get_next_chunk(dn, &chunk_begin, offset); 651 if (err) 652 return (err); 653 ASSERT3U(chunk_begin, >=, offset); 654 ASSERT3U(chunk_begin, <=, chunk_end); 655 656 dmu_tx_t *tx = dmu_tx_create(os); 657 dmu_tx_hold_free(tx, dn->dn_object, 658 chunk_begin, chunk_end - chunk_begin); 659 err = dmu_tx_assign(tx, TXG_WAIT); 660 if (err) { 661 dmu_tx_abort(tx); 662 return (err); 663 } 664 dnode_free_range(dn, chunk_begin, chunk_end - chunk_begin, tx); 665 dmu_tx_commit(tx); 666 667 length -= chunk_end - chunk_begin; 668 } 669 return (0); 670} 671 672int 673dmu_free_long_range(objset_t *os, uint64_t object, 674 uint64_t offset, uint64_t length) 675{ 676 dnode_t *dn; 677 int err; 678 679 err = dnode_hold(os, object, FTAG, &dn); 680 if (err != 0) 681 return (err); 682 err = dmu_free_long_range_impl(os, dn, offset, length); 683 684 /* 685 * It is important to zero out the maxblkid when freeing the entire 686 * file, so that (a) subsequent calls to dmu_free_long_range_impl() 687 * will take the fast path, and (b) dnode_reallocate() can verify 688 * that the entire file has been freed. 689 */ 690 if (err == 0 && offset == 0 && length == DMU_OBJECT_END) 691 dn->dn_maxblkid = 0; 692 693 dnode_rele(dn, FTAG); 694 return (err); 695} 696 697int 698dmu_free_long_object(objset_t *os, uint64_t object) 699{ 700 dmu_tx_t *tx; 701 int err; 702 703 err = dmu_free_long_range(os, object, 0, DMU_OBJECT_END); 704 if (err != 0) 705 return (err); 706 707 tx = dmu_tx_create(os); 708 dmu_tx_hold_bonus(tx, object); 709 dmu_tx_hold_free(tx, object, 0, DMU_OBJECT_END); 710 err = dmu_tx_assign(tx, TXG_WAIT); 711 if (err == 0) { 712 err = dmu_object_free(os, object, tx); 713 dmu_tx_commit(tx); 714 } else { 715 dmu_tx_abort(tx); 716 } 717 718 return (err); 719} 720 721int 722dmu_free_range(objset_t *os, uint64_t object, uint64_t offset, 723 uint64_t size, dmu_tx_t *tx) 724{ 725 dnode_t *dn; 726 int err = dnode_hold(os, object, FTAG, &dn); 727 if (err) 728 return (err); 729 ASSERT(offset < UINT64_MAX); 730 ASSERT(size == -1ULL || size <= UINT64_MAX - offset); 731 dnode_free_range(dn, offset, size, tx); 732 dnode_rele(dn, FTAG); 733 return (0); 734} 735 736int 737dmu_read(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 738 void *buf, uint32_t flags) 739{ 740 dnode_t *dn; 741 dmu_buf_t **dbp; 742 int numbufs, err; 743 744 err = dnode_hold(os, object, FTAG, &dn); 745 if (err) 746 return (err); 747 748 /* 749 * Deal with odd block sizes, where there can't be data past the first 750 * block. If we ever do the tail block optimization, we will need to 751 * handle that here as well. 752 */ 753 if (dn->dn_maxblkid == 0) { 754 int newsz = offset > dn->dn_datablksz ? 0 : 755 MIN(size, dn->dn_datablksz - offset); 756 bzero((char *)buf + newsz, size - newsz); 757 size = newsz; 758 } 759 760 while (size > 0) { 761 uint64_t mylen = MIN(size, DMU_MAX_ACCESS / 2); 762 int i; 763 764 /* 765 * NB: we could do this block-at-a-time, but it's nice 766 * to be reading in parallel. 767 */ 768 err = dmu_buf_hold_array_by_dnode(dn, offset, mylen, 769 TRUE, FTAG, &numbufs, &dbp, flags); 770 if (err) 771 break; 772 773 for (i = 0; i < numbufs; i++) { 774 int tocpy; 775 int bufoff; 776 dmu_buf_t *db = dbp[i]; 777 778 ASSERT(size > 0); 779 780 bufoff = offset - db->db_offset; 781 tocpy = (int)MIN(db->db_size - bufoff, size); 782 783 bcopy((char *)db->db_data + bufoff, buf, tocpy); 784 785 offset += tocpy; 786 size -= tocpy; 787 buf = (char *)buf + tocpy; 788 } 789 dmu_buf_rele_array(dbp, numbufs, FTAG); 790 } 791 dnode_rele(dn, FTAG); 792 return (err); 793} 794 795void 796dmu_write(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 797 const void *buf, dmu_tx_t *tx) 798{ 799 dmu_buf_t **dbp; 800 int numbufs, i; 801 802 if (size == 0) 803 return; 804 805 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size, 806 FALSE, FTAG, &numbufs, &dbp)); 807 808 for (i = 0; i < numbufs; i++) { 809 int tocpy; 810 int bufoff; 811 dmu_buf_t *db = dbp[i]; 812 813 ASSERT(size > 0); 814 815 bufoff = offset - db->db_offset; 816 tocpy = (int)MIN(db->db_size - bufoff, size); 817 818 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 819 820 if (tocpy == db->db_size) 821 dmu_buf_will_fill(db, tx); 822 else 823 dmu_buf_will_dirty(db, tx); 824 825 bcopy(buf, (char *)db->db_data + bufoff, tocpy); 826 827 if (tocpy == db->db_size) 828 dmu_buf_fill_done(db, tx); 829 830 offset += tocpy; 831 size -= tocpy; 832 buf = (char *)buf + tocpy; 833 } 834 dmu_buf_rele_array(dbp, numbufs, FTAG); 835} 836 837void 838dmu_prealloc(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 839 dmu_tx_t *tx) 840{ 841 dmu_buf_t **dbp; 842 int numbufs, i; 843 844 if (size == 0) 845 return; 846 847 VERIFY(0 == dmu_buf_hold_array(os, object, offset, size, 848 FALSE, FTAG, &numbufs, &dbp)); 849 850 for (i = 0; i < numbufs; i++) { 851 dmu_buf_t *db = dbp[i]; 852 853 dmu_buf_will_not_fill(db, tx); 854 } 855 dmu_buf_rele_array(dbp, numbufs, FTAG); 856} 857 858/* 859 * DMU support for xuio 860 */ 861kstat_t *xuio_ksp = NULL; 862 863int 864dmu_xuio_init(xuio_t *xuio, int nblk) 865{ 866 dmu_xuio_t *priv; 867 uio_t *uio = &xuio->xu_uio; 868 869 uio->uio_iovcnt = nblk; 870 uio->uio_iov = kmem_zalloc(nblk * sizeof (iovec_t), KM_SLEEP); 871 872 priv = kmem_zalloc(sizeof (dmu_xuio_t), KM_SLEEP); 873 priv->cnt = nblk; 874 priv->bufs = kmem_zalloc(nblk * sizeof (arc_buf_t *), KM_SLEEP); 875 priv->iovp = uio->uio_iov; 876 XUIO_XUZC_PRIV(xuio) = priv; 877 878 if (XUIO_XUZC_RW(xuio) == UIO_READ) 879 XUIOSTAT_INCR(xuiostat_onloan_rbuf, nblk); 880 else 881 XUIOSTAT_INCR(xuiostat_onloan_wbuf, nblk); 882 883 return (0); 884} 885 886void 887dmu_xuio_fini(xuio_t *xuio) 888{ 889 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 890 int nblk = priv->cnt; 891 892 kmem_free(priv->iovp, nblk * sizeof (iovec_t)); 893 kmem_free(priv->bufs, nblk * sizeof (arc_buf_t *)); 894 kmem_free(priv, sizeof (dmu_xuio_t)); 895 896 if (XUIO_XUZC_RW(xuio) == UIO_READ) 897 XUIOSTAT_INCR(xuiostat_onloan_rbuf, -nblk); 898 else 899 XUIOSTAT_INCR(xuiostat_onloan_wbuf, -nblk); 900} 901 902/* 903 * Initialize iov[priv->next] and priv->bufs[priv->next] with { off, n, abuf } 904 * and increase priv->next by 1. 905 */ 906int 907dmu_xuio_add(xuio_t *xuio, arc_buf_t *abuf, offset_t off, size_t n) 908{ 909 struct iovec *iov; 910 uio_t *uio = &xuio->xu_uio; 911 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 912 int i = priv->next++; 913 914 ASSERT(i < priv->cnt); 915 ASSERT(off + n <= arc_buf_size(abuf)); 916 iov = uio->uio_iov + i; 917 iov->iov_base = (char *)abuf->b_data + off; 918 iov->iov_len = n; 919 priv->bufs[i] = abuf; 920 return (0); 921} 922 923int 924dmu_xuio_cnt(xuio_t *xuio) 925{ 926 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 927 return (priv->cnt); 928} 929 930arc_buf_t * 931dmu_xuio_arcbuf(xuio_t *xuio, int i) 932{ 933 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 934 935 ASSERT(i < priv->cnt); 936 return (priv->bufs[i]); 937} 938 939void 940dmu_xuio_clear(xuio_t *xuio, int i) 941{ 942 dmu_xuio_t *priv = XUIO_XUZC_PRIV(xuio); 943 944 ASSERT(i < priv->cnt); 945 priv->bufs[i] = NULL; 946} 947 948static void 949xuio_stat_init(void) 950{ 951 xuio_ksp = kstat_create("zfs", 0, "xuio_stats", "misc", 952 KSTAT_TYPE_NAMED, sizeof (xuio_stats) / sizeof (kstat_named_t), 953 KSTAT_FLAG_VIRTUAL); 954 if (xuio_ksp != NULL) { 955 xuio_ksp->ks_data = &xuio_stats; 956 kstat_install(xuio_ksp); 957 } 958} 959 960static void 961xuio_stat_fini(void) 962{ 963 if (xuio_ksp != NULL) { 964 kstat_delete(xuio_ksp); 965 xuio_ksp = NULL; 966 } 967} 968 969void 970xuio_stat_wbuf_copied() 971{ 972 XUIOSTAT_BUMP(xuiostat_wbuf_copied); 973} 974 975void 976xuio_stat_wbuf_nocopy() 977{ 978 XUIOSTAT_BUMP(xuiostat_wbuf_nocopy); 979} 980 981#ifdef _KERNEL 982int 983dmu_read_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size) 984{ 985 dmu_buf_t **dbp; 986 int numbufs, i, err; 987 xuio_t *xuio = NULL; 988 989 /* 990 * NB: we could do this block-at-a-time, but it's nice 991 * to be reading in parallel. 992 */ 993 err = dmu_buf_hold_array(os, object, uio->uio_loffset, size, TRUE, FTAG, 994 &numbufs, &dbp); 995 if (err) 996 return (err); 997 998#ifdef UIO_XUIO 999 if (uio->uio_extflg == UIO_XUIO) 1000 xuio = (xuio_t *)uio; 1001#endif 1002 1003 for (i = 0; i < numbufs; i++) { 1004 int tocpy; 1005 int bufoff; 1006 dmu_buf_t *db = dbp[i]; 1007 1008 ASSERT(size > 0); 1009 1010 bufoff = uio->uio_loffset - db->db_offset; 1011 tocpy = (int)MIN(db->db_size - bufoff, size); 1012 1013 if (xuio) { 1014 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 1015 arc_buf_t *dbuf_abuf = dbi->db_buf; 1016 arc_buf_t *abuf = dbuf_loan_arcbuf(dbi); 1017 err = dmu_xuio_add(xuio, abuf, bufoff, tocpy); 1018 if (!err) { 1019 uio->uio_resid -= tocpy; 1020 uio->uio_loffset += tocpy; 1021 } 1022 1023 if (abuf == dbuf_abuf) 1024 XUIOSTAT_BUMP(xuiostat_rbuf_nocopy); 1025 else 1026 XUIOSTAT_BUMP(xuiostat_rbuf_copied); 1027 } else { 1028 err = uiomove((char *)db->db_data + bufoff, tocpy, 1029 UIO_READ, uio); 1030 } 1031 if (err) 1032 break; 1033 1034 size -= tocpy; 1035 } 1036 dmu_buf_rele_array(dbp, numbufs, FTAG); 1037 1038 return (err); 1039} 1040 1041static int 1042dmu_write_uio_dnode(dnode_t *dn, uio_t *uio, uint64_t size, dmu_tx_t *tx) 1043{ 1044 dmu_buf_t **dbp; 1045 int numbufs; 1046 int err = 0; 1047 int i; 1048 1049 err = dmu_buf_hold_array_by_dnode(dn, uio->uio_loffset, size, 1050 FALSE, FTAG, &numbufs, &dbp, DMU_READ_PREFETCH); 1051 if (err) 1052 return (err); 1053 1054 for (i = 0; i < numbufs; i++) { 1055 int tocpy; 1056 int bufoff; 1057 dmu_buf_t *db = dbp[i]; 1058 1059 ASSERT(size > 0); 1060 1061 bufoff = uio->uio_loffset - db->db_offset; 1062 tocpy = (int)MIN(db->db_size - bufoff, size); 1063 1064 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 1065 1066 if (tocpy == db->db_size) 1067 dmu_buf_will_fill(db, tx); 1068 else 1069 dmu_buf_will_dirty(db, tx); 1070 1071 /* 1072 * XXX uiomove could block forever (eg. nfs-backed 1073 * pages). There needs to be a uiolockdown() function 1074 * to lock the pages in memory, so that uiomove won't 1075 * block. 1076 */ 1077 err = uiomove((char *)db->db_data + bufoff, tocpy, 1078 UIO_WRITE, uio); 1079 1080 if (tocpy == db->db_size) 1081 dmu_buf_fill_done(db, tx); 1082 1083 if (err) 1084 break; 1085 1086 size -= tocpy; 1087 } 1088 1089 dmu_buf_rele_array(dbp, numbufs, FTAG); 1090 return (err); 1091} 1092 1093int 1094dmu_write_uio_dbuf(dmu_buf_t *zdb, uio_t *uio, uint64_t size, 1095 dmu_tx_t *tx) 1096{ 1097 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zdb; 1098 dnode_t *dn; 1099 int err; 1100 1101 if (size == 0) 1102 return (0); 1103 1104 DB_DNODE_ENTER(db); 1105 dn = DB_DNODE(db); 1106 err = dmu_write_uio_dnode(dn, uio, size, tx); 1107 DB_DNODE_EXIT(db); 1108 1109 return (err); 1110} 1111 1112int 1113dmu_write_uio(objset_t *os, uint64_t object, uio_t *uio, uint64_t size, 1114 dmu_tx_t *tx) 1115{ 1116 dnode_t *dn; 1117 int err; 1118 1119 if (size == 0) 1120 return (0); 1121 1122 err = dnode_hold(os, object, FTAG, &dn); 1123 if (err) 1124 return (err); 1125 1126 err = dmu_write_uio_dnode(dn, uio, size, tx); 1127 1128 dnode_rele(dn, FTAG); 1129 1130 return (err); 1131} 1132 1133#ifdef sun 1134int 1135dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1136 page_t *pp, dmu_tx_t *tx) 1137{ 1138 dmu_buf_t **dbp; 1139 int numbufs, i; 1140 int err; 1141 1142 if (size == 0) 1143 return (0); 1144 1145 err = dmu_buf_hold_array(os, object, offset, size, 1146 FALSE, FTAG, &numbufs, &dbp); 1147 if (err) 1148 return (err); 1149 1150 for (i = 0; i < numbufs; i++) { 1151 int tocpy, copied, thiscpy; 1152 int bufoff; 1153 dmu_buf_t *db = dbp[i]; 1154 caddr_t va; 1155 1156 ASSERT(size > 0); 1157 ASSERT3U(db->db_size, >=, PAGESIZE); 1158 1159 bufoff = offset - db->db_offset; 1160 tocpy = (int)MIN(db->db_size - bufoff, size); 1161 1162 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 1163 1164 if (tocpy == db->db_size) 1165 dmu_buf_will_fill(db, tx); 1166 else 1167 dmu_buf_will_dirty(db, tx); 1168 1169 for (copied = 0; copied < tocpy; copied += PAGESIZE) { 1170 ASSERT3U(pp->p_offset, ==, db->db_offset + bufoff); 1171 thiscpy = MIN(PAGESIZE, tocpy - copied); 1172 va = zfs_map_page(pp, S_READ); 1173 bcopy(va, (char *)db->db_data + bufoff, thiscpy); 1174 zfs_unmap_page(pp, va); 1175 pp = pp->p_next; 1176 bufoff += PAGESIZE; 1177 } 1178 1179 if (tocpy == db->db_size) 1180 dmu_buf_fill_done(db, tx); 1181 1182 offset += tocpy; 1183 size -= tocpy; 1184 } 1185 dmu_buf_rele_array(dbp, numbufs, FTAG); 1186 return (err); 1187} 1188 1189#else 1190 1191int 1192dmu_write_pages(objset_t *os, uint64_t object, uint64_t offset, uint64_t size, 1193 vm_page_t *ma, dmu_tx_t *tx) 1194{ 1195 dmu_buf_t **dbp; 1196 struct sf_buf *sf; 1197 int numbufs, i; 1198 int err; 1199 1200 if (size == 0) 1201 return (0); 1202 1203 err = dmu_buf_hold_array(os, object, offset, size, 1204 FALSE, FTAG, &numbufs, &dbp); 1205 if (err) 1206 return (err); 1207 1208 for (i = 0; i < numbufs; i++) { 1209 int tocpy, copied, thiscpy; 1210 int bufoff; 1211 dmu_buf_t *db = dbp[i]; 1212 caddr_t va; 1213 1214 ASSERT(size > 0); 1215 ASSERT3U(db->db_size, >=, PAGESIZE); 1216 1217 bufoff = offset - db->db_offset; 1218 tocpy = (int)MIN(db->db_size - bufoff, size); 1219 1220 ASSERT(i == 0 || i == numbufs-1 || tocpy == db->db_size); 1221 1222 if (tocpy == db->db_size) 1223 dmu_buf_will_fill(db, tx); 1224 else 1225 dmu_buf_will_dirty(db, tx); 1226 1227 for (copied = 0; copied < tocpy; copied += PAGESIZE) { 1228 ASSERT3U(ptoa((*ma)->pindex), ==, db->db_offset + bufoff); 1229 thiscpy = MIN(PAGESIZE, tocpy - copied); 1230 va = zfs_map_page(*ma, &sf); 1231 bcopy(va, (char *)db->db_data + bufoff, thiscpy); 1232 zfs_unmap_page(sf); 1233 ma += 1; 1234 bufoff += PAGESIZE; 1235 } 1236 1237 if (tocpy == db->db_size) 1238 dmu_buf_fill_done(db, tx); 1239 1240 offset += tocpy; 1241 size -= tocpy; 1242 } 1243 dmu_buf_rele_array(dbp, numbufs, FTAG); 1244 return (err); 1245} 1246#endif /* sun */ 1247#endif 1248 1249/* 1250 * Allocate a loaned anonymous arc buffer. 1251 */ 1252arc_buf_t * 1253dmu_request_arcbuf(dmu_buf_t *handle, int size) 1254{ 1255 dmu_buf_impl_t *db = (dmu_buf_impl_t *)handle; 1256 1257 return (arc_loan_buf(db->db_objset->os_spa, size)); 1258} 1259 1260/* 1261 * Free a loaned arc buffer. 1262 */ 1263void 1264dmu_return_arcbuf(arc_buf_t *buf) 1265{ 1266 arc_return_buf(buf, FTAG); 1267 VERIFY(arc_buf_remove_ref(buf, FTAG)); 1268} 1269 1270/* 1271 * When possible directly assign passed loaned arc buffer to a dbuf. 1272 * If this is not possible copy the contents of passed arc buf via 1273 * dmu_write(). 1274 */ 1275void 1276dmu_assign_arcbuf(dmu_buf_t *handle, uint64_t offset, arc_buf_t *buf, 1277 dmu_tx_t *tx) 1278{ 1279 dmu_buf_impl_t *dbuf = (dmu_buf_impl_t *)handle; 1280 dnode_t *dn; 1281 dmu_buf_impl_t *db; 1282 uint32_t blksz = (uint32_t)arc_buf_size(buf); 1283 uint64_t blkid; 1284 1285 DB_DNODE_ENTER(dbuf); 1286 dn = DB_DNODE(dbuf); 1287 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1288 blkid = dbuf_whichblock(dn, offset); 1289 VERIFY((db = dbuf_hold(dn, blkid, FTAG)) != NULL); 1290 rw_exit(&dn->dn_struct_rwlock); 1291 DB_DNODE_EXIT(dbuf); 1292 1293 if (offset == db->db.db_offset && blksz == db->db.db_size) { 1294 dbuf_assign_arcbuf(db, buf, tx); 1295 dbuf_rele(db, FTAG); 1296 } else { 1297 objset_t *os; 1298 uint64_t object; 1299 1300 DB_DNODE_ENTER(dbuf); 1301 dn = DB_DNODE(dbuf); 1302 os = dn->dn_objset; 1303 object = dn->dn_object; 1304 DB_DNODE_EXIT(dbuf); 1305 1306 dbuf_rele(db, FTAG); 1307 dmu_write(os, object, offset, blksz, buf->b_data, tx); 1308 dmu_return_arcbuf(buf); 1309 XUIOSTAT_BUMP(xuiostat_wbuf_copied); 1310 } 1311} 1312 1313typedef struct { 1314 dbuf_dirty_record_t *dsa_dr; 1315 dmu_sync_cb_t *dsa_done; 1316 zgd_t *dsa_zgd; 1317 dmu_tx_t *dsa_tx; 1318} dmu_sync_arg_t; 1319 1320/* ARGSUSED */ 1321static void 1322dmu_sync_ready(zio_t *zio, arc_buf_t *buf, void *varg) 1323{ 1324 dmu_sync_arg_t *dsa = varg; 1325 dmu_buf_t *db = dsa->dsa_zgd->zgd_db; 1326 blkptr_t *bp = zio->io_bp; 1327 1328 if (zio->io_error == 0) { 1329 if (BP_IS_HOLE(bp)) { 1330 /* 1331 * A block of zeros may compress to a hole, but the 1332 * block size still needs to be known for replay. 1333 */ 1334 BP_SET_LSIZE(bp, db->db_size); 1335 } else { 1336 ASSERT(BP_GET_LEVEL(bp) == 0); 1337 bp->blk_fill = 1; 1338 } 1339 } 1340} 1341 1342static void 1343dmu_sync_late_arrival_ready(zio_t *zio) 1344{ 1345 dmu_sync_ready(zio, NULL, zio->io_private); 1346} 1347 1348/* ARGSUSED */ 1349static void 1350dmu_sync_done(zio_t *zio, arc_buf_t *buf, void *varg) 1351{ 1352 dmu_sync_arg_t *dsa = varg; 1353 dbuf_dirty_record_t *dr = dsa->dsa_dr; 1354 dmu_buf_impl_t *db = dr->dr_dbuf; 1355 1356 mutex_enter(&db->db_mtx); 1357 ASSERT(dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC); 1358 if (zio->io_error == 0) { 1359 dr->dt.dl.dr_nopwrite = !!(zio->io_flags & ZIO_FLAG_NOPWRITE); 1360 if (dr->dt.dl.dr_nopwrite) { 1361 blkptr_t *bp = zio->io_bp; 1362 blkptr_t *bp_orig = &zio->io_bp_orig; 1363 uint8_t chksum = BP_GET_CHECKSUM(bp_orig); 1364 1365 ASSERT(BP_EQUAL(bp, bp_orig)); 1366 ASSERT(zio->io_prop.zp_compress != ZIO_COMPRESS_OFF); 1367 ASSERT(zio_checksum_table[chksum].ci_dedup); 1368 } 1369 dr->dt.dl.dr_overridden_by = *zio->io_bp; 1370 dr->dt.dl.dr_override_state = DR_OVERRIDDEN; 1371 dr->dt.dl.dr_copies = zio->io_prop.zp_copies; 1372 if (BP_IS_HOLE(&dr->dt.dl.dr_overridden_by)) 1373 BP_ZERO(&dr->dt.dl.dr_overridden_by); 1374 } else { 1375 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 1376 } 1377 cv_broadcast(&db->db_changed); 1378 mutex_exit(&db->db_mtx); 1379 1380 dsa->dsa_done(dsa->dsa_zgd, zio->io_error); 1381 1382 kmem_free(dsa, sizeof (*dsa)); 1383} 1384 1385static void 1386dmu_sync_late_arrival_done(zio_t *zio) 1387{ 1388 blkptr_t *bp = zio->io_bp; 1389 dmu_sync_arg_t *dsa = zio->io_private; 1390 blkptr_t *bp_orig = &zio->io_bp_orig; 1391 1392 if (zio->io_error == 0 && !BP_IS_HOLE(bp)) { 1393 /* 1394 * If we didn't allocate a new block (i.e. ZIO_FLAG_NOPWRITE) 1395 * then there is nothing to do here. Otherwise, free the 1396 * newly allocated block in this txg. 1397 */ 1398 if (zio->io_flags & ZIO_FLAG_NOPWRITE) { 1399 ASSERT(BP_EQUAL(bp, bp_orig)); 1400 } else { 1401 ASSERT(BP_IS_HOLE(bp_orig) || !BP_EQUAL(bp, bp_orig)); 1402 ASSERT(zio->io_bp->blk_birth == zio->io_txg); 1403 ASSERT(zio->io_txg > spa_syncing_txg(zio->io_spa)); 1404 zio_free(zio->io_spa, zio->io_txg, zio->io_bp); 1405 } 1406 } 1407 1408 dmu_tx_commit(dsa->dsa_tx); 1409 1410 dsa->dsa_done(dsa->dsa_zgd, zio->io_error); 1411 1412 kmem_free(dsa, sizeof (*dsa)); 1413} 1414 1415static int 1416dmu_sync_late_arrival(zio_t *pio, objset_t *os, dmu_sync_cb_t *done, zgd_t *zgd, 1417 zio_prop_t *zp, zbookmark_t *zb) 1418{ 1419 dmu_sync_arg_t *dsa; 1420 dmu_tx_t *tx; 1421 1422 tx = dmu_tx_create(os); 1423 dmu_tx_hold_space(tx, zgd->zgd_db->db_size); 1424 if (dmu_tx_assign(tx, TXG_WAIT) != 0) { 1425 dmu_tx_abort(tx); 1426 /* Make zl_get_data do txg_waited_synced() */ 1427 return (SET_ERROR(EIO)); 1428 } 1429 1430 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP); 1431 dsa->dsa_dr = NULL; 1432 dsa->dsa_done = done; 1433 dsa->dsa_zgd = zgd; 1434 dsa->dsa_tx = tx; 1435 1436 zio_nowait(zio_write(pio, os->os_spa, dmu_tx_get_txg(tx), zgd->zgd_bp, 1437 zgd->zgd_db->db_data, zgd->zgd_db->db_size, zp, 1438 dmu_sync_late_arrival_ready, NULL, dmu_sync_late_arrival_done, dsa, 1439 ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, zb)); 1440 1441 return (0); 1442} 1443 1444/* 1445 * Intent log support: sync the block associated with db to disk. 1446 * N.B. and XXX: the caller is responsible for making sure that the 1447 * data isn't changing while dmu_sync() is writing it. 1448 * 1449 * Return values: 1450 * 1451 * EEXIST: this txg has already been synced, so there's nothing to do. 1452 * The caller should not log the write. 1453 * 1454 * ENOENT: the block was dbuf_free_range()'d, so there's nothing to do. 1455 * The caller should not log the write. 1456 * 1457 * EALREADY: this block is already in the process of being synced. 1458 * The caller should track its progress (somehow). 1459 * 1460 * EIO: could not do the I/O. 1461 * The caller should do a txg_wait_synced(). 1462 * 1463 * 0: the I/O has been initiated. 1464 * The caller should log this blkptr in the done callback. 1465 * It is possible that the I/O will fail, in which case 1466 * the error will be reported to the done callback and 1467 * propagated to pio from zio_done(). 1468 */ 1469int 1470dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd) 1471{ 1472 blkptr_t *bp = zgd->zgd_bp; 1473 dmu_buf_impl_t *db = (dmu_buf_impl_t *)zgd->zgd_db; 1474 objset_t *os = db->db_objset; 1475 dsl_dataset_t *ds = os->os_dsl_dataset; 1476 dbuf_dirty_record_t *dr; 1477 dmu_sync_arg_t *dsa; 1478 zbookmark_t zb; 1479 zio_prop_t zp; 1480 dnode_t *dn; 1481 1482 ASSERT(pio != NULL); 1483 ASSERT(txg != 0); 1484 1485 SET_BOOKMARK(&zb, ds->ds_object, 1486 db->db.db_object, db->db_level, db->db_blkid); 1487 1488 DB_DNODE_ENTER(db); 1489 dn = DB_DNODE(db); 1490 dmu_write_policy(os, dn, db->db_level, WP_DMU_SYNC, &zp); 1491 DB_DNODE_EXIT(db); 1492 1493 /* 1494 * If we're frozen (running ziltest), we always need to generate a bp. 1495 */ 1496 if (txg > spa_freeze_txg(os->os_spa)) 1497 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); 1498 1499 /* 1500 * Grabbing db_mtx now provides a barrier between dbuf_sync_leaf() 1501 * and us. If we determine that this txg is not yet syncing, 1502 * but it begins to sync a moment later, that's OK because the 1503 * sync thread will block in dbuf_sync_leaf() until we drop db_mtx. 1504 */ 1505 mutex_enter(&db->db_mtx); 1506 1507 if (txg <= spa_last_synced_txg(os->os_spa)) { 1508 /* 1509 * This txg has already synced. There's nothing to do. 1510 */ 1511 mutex_exit(&db->db_mtx); 1512 return (SET_ERROR(EEXIST)); 1513 } 1514 1515 if (txg <= spa_syncing_txg(os->os_spa)) { 1516 /* 1517 * This txg is currently syncing, so we can't mess with 1518 * the dirty record anymore; just write a new log block. 1519 */ 1520 mutex_exit(&db->db_mtx); 1521 return (dmu_sync_late_arrival(pio, os, done, zgd, &zp, &zb)); 1522 } 1523 1524 dr = db->db_last_dirty; 1525 while (dr && dr->dr_txg != txg) 1526 dr = dr->dr_next; 1527 1528 if (dr == NULL) { 1529 /* 1530 * There's no dr for this dbuf, so it must have been freed. 1531 * There's no need to log writes to freed blocks, so we're done. 1532 */ 1533 mutex_exit(&db->db_mtx); 1534 return (SET_ERROR(ENOENT)); 1535 } 1536 1537 ASSERT(dr->dr_next == NULL || dr->dr_next->dr_txg < txg); 1538 1539 /* 1540 * Assume the on-disk data is X, the current syncing data is Y, 1541 * and the current in-memory data is Z (currently in dmu_sync). 1542 * X and Z are identical but Y is has been modified. Normally, 1543 * when X and Z are the same we will perform a nopwrite but if Y 1544 * is different we must disable nopwrite since the resulting write 1545 * of Y to disk can free the block containing X. If we allowed a 1546 * nopwrite to occur the block pointing to Z would reference a freed 1547 * block. Since this is a rare case we simplify this by disabling 1548 * nopwrite if the current dmu_sync-ing dbuf has been modified in 1549 * a previous transaction. 1550 */ 1551 if (dr->dr_next) 1552 zp.zp_nopwrite = B_FALSE; 1553 1554 ASSERT(dr->dr_txg == txg); 1555 if (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC || 1556 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { 1557 /* 1558 * We have already issued a sync write for this buffer, 1559 * or this buffer has already been synced. It could not 1560 * have been dirtied since, or we would have cleared the state. 1561 */ 1562 mutex_exit(&db->db_mtx); 1563 return (SET_ERROR(EALREADY)); 1564 } 1565 1566 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); 1567 dr->dt.dl.dr_override_state = DR_IN_DMU_SYNC; 1568 mutex_exit(&db->db_mtx); 1569 1570 dsa = kmem_alloc(sizeof (dmu_sync_arg_t), KM_SLEEP); 1571 dsa->dsa_dr = dr; 1572 dsa->dsa_done = done; 1573 dsa->dsa_zgd = zgd; 1574 dsa->dsa_tx = NULL; 1575 1576 zio_nowait(arc_write(pio, os->os_spa, txg, 1577 bp, dr->dt.dl.dr_data, DBUF_IS_L2CACHEABLE(db), 1578 DBUF_IS_L2COMPRESSIBLE(db), &zp, dmu_sync_ready, 1579 NULL, dmu_sync_done, dsa, ZIO_PRIORITY_SYNC_WRITE, 1580 ZIO_FLAG_CANFAIL, &zb)); 1581 1582 return (0); 1583} 1584 1585int 1586dmu_object_set_blocksize(objset_t *os, uint64_t object, uint64_t size, int ibs, 1587 dmu_tx_t *tx) 1588{ 1589 dnode_t *dn; 1590 int err; 1591 1592 err = dnode_hold(os, object, FTAG, &dn); 1593 if (err) 1594 return (err); 1595 err = dnode_set_blksz(dn, size, ibs, tx); 1596 dnode_rele(dn, FTAG); 1597 return (err); 1598} 1599 1600void 1601dmu_object_set_checksum(objset_t *os, uint64_t object, uint8_t checksum, 1602 dmu_tx_t *tx) 1603{ 1604 dnode_t *dn; 1605 1606 /* XXX assumes dnode_hold will not get an i/o error */ 1607 (void) dnode_hold(os, object, FTAG, &dn); 1608 ASSERT(checksum < ZIO_CHECKSUM_FUNCTIONS); 1609 dn->dn_checksum = checksum; 1610 dnode_setdirty(dn, tx); 1611 dnode_rele(dn, FTAG); 1612} 1613 1614void 1615dmu_object_set_compress(objset_t *os, uint64_t object, uint8_t compress, 1616 dmu_tx_t *tx) 1617{ 1618 dnode_t *dn; 1619 1620 /* XXX assumes dnode_hold will not get an i/o error */ 1621 (void) dnode_hold(os, object, FTAG, &dn); 1622 ASSERT(compress < ZIO_COMPRESS_FUNCTIONS); 1623 dn->dn_compress = compress; 1624 dnode_setdirty(dn, tx); 1625 dnode_rele(dn, FTAG); 1626} 1627 1628int zfs_mdcomp_disable = 0; 1629TUNABLE_INT("vfs.zfs.mdcomp_disable", &zfs_mdcomp_disable); 1630SYSCTL_INT(_vfs_zfs, OID_AUTO, mdcomp_disable, CTLFLAG_RW, 1631 &zfs_mdcomp_disable, 0, "Disable metadata compression"); 1632 1633void 1634dmu_write_policy(objset_t *os, dnode_t *dn, int level, int wp, zio_prop_t *zp) 1635{ 1636 dmu_object_type_t type = dn ? dn->dn_type : DMU_OT_OBJSET; 1637 boolean_t ismd = (level > 0 || DMU_OT_IS_METADATA(type) || 1638 (wp & WP_SPILL)); 1639 enum zio_checksum checksum = os->os_checksum; 1640 enum zio_compress compress = os->os_compress; 1641 enum zio_checksum dedup_checksum = os->os_dedup_checksum; 1642 boolean_t dedup = B_FALSE; 1643 boolean_t nopwrite = B_FALSE; 1644 boolean_t dedup_verify = os->os_dedup_verify; 1645 int copies = os->os_copies; 1646 1647 /* 1648 * We maintain different write policies for each of the following 1649 * types of data: 1650 * 1. metadata 1651 * 2. preallocated blocks (i.e. level-0 blocks of a dump device) 1652 * 3. all other level 0 blocks 1653 */ 1654 if (ismd) { 1655 /* 1656 * XXX -- we should design a compression algorithm 1657 * that specializes in arrays of bps. 1658 */ 1659 compress = zfs_mdcomp_disable ? ZIO_COMPRESS_EMPTY : 1660 ZIO_COMPRESS_LZJB; 1661 1662 /* 1663 * Metadata always gets checksummed. If the data 1664 * checksum is multi-bit correctable, and it's not a 1665 * ZBT-style checksum, then it's suitable for metadata 1666 * as well. Otherwise, the metadata checksum defaults 1667 * to fletcher4. 1668 */ 1669 if (zio_checksum_table[checksum].ci_correctable < 1 || 1670 zio_checksum_table[checksum].ci_eck) 1671 checksum = ZIO_CHECKSUM_FLETCHER_4; 1672 } else if (wp & WP_NOFILL) { 1673 ASSERT(level == 0); 1674 1675 /* 1676 * If we're writing preallocated blocks, we aren't actually 1677 * writing them so don't set any policy properties. These 1678 * blocks are currently only used by an external subsystem 1679 * outside of zfs (i.e. dump) and not written by the zio 1680 * pipeline. 1681 */ 1682 compress = ZIO_COMPRESS_OFF; 1683 checksum = ZIO_CHECKSUM_NOPARITY; 1684 } else { 1685 compress = zio_compress_select(dn->dn_compress, compress); 1686 1687 checksum = (dedup_checksum == ZIO_CHECKSUM_OFF) ? 1688 zio_checksum_select(dn->dn_checksum, checksum) : 1689 dedup_checksum; 1690 1691 /* 1692 * Determine dedup setting. If we are in dmu_sync(), 1693 * we won't actually dedup now because that's all 1694 * done in syncing context; but we do want to use the 1695 * dedup checkum. If the checksum is not strong 1696 * enough to ensure unique signatures, force 1697 * dedup_verify. 1698 */ 1699 if (dedup_checksum != ZIO_CHECKSUM_OFF) { 1700 dedup = (wp & WP_DMU_SYNC) ? B_FALSE : B_TRUE; 1701 if (!zio_checksum_table[checksum].ci_dedup) 1702 dedup_verify = B_TRUE; 1703 } 1704 1705 /* 1706 * Enable nopwrite if we have a cryptographically secure 1707 * checksum that has no known collisions (i.e. SHA-256) 1708 * and compression is enabled. We don't enable nopwrite if 1709 * dedup is enabled as the two features are mutually exclusive. 1710 */ 1711 nopwrite = (!dedup && zio_checksum_table[checksum].ci_dedup && 1712 compress != ZIO_COMPRESS_OFF && zfs_nopwrite_enabled); 1713 } 1714 1715 zp->zp_checksum = checksum; 1716 zp->zp_compress = compress; 1717 zp->zp_type = (wp & WP_SPILL) ? dn->dn_bonustype : type; 1718 zp->zp_level = level; 1719 zp->zp_copies = MIN(copies + ismd, spa_max_replication(os->os_spa)); 1720 zp->zp_dedup = dedup; 1721 zp->zp_dedup_verify = dedup && dedup_verify; 1722 zp->zp_nopwrite = nopwrite; 1723} 1724 1725int 1726dmu_offset_next(objset_t *os, uint64_t object, boolean_t hole, uint64_t *off) 1727{ 1728 dnode_t *dn; 1729 int i, err; 1730 1731 err = dnode_hold(os, object, FTAG, &dn); 1732 if (err) 1733 return (err); 1734 /* 1735 * Sync any current changes before 1736 * we go trundling through the block pointers. 1737 */ 1738 for (i = 0; i < TXG_SIZE; i++) { 1739 if (list_link_active(&dn->dn_dirty_link[i])) 1740 break; 1741 } 1742 if (i != TXG_SIZE) { 1743 dnode_rele(dn, FTAG); 1744 txg_wait_synced(dmu_objset_pool(os), 0); 1745 err = dnode_hold(os, object, FTAG, &dn); 1746 if (err) 1747 return (err); 1748 } 1749 1750 err = dnode_next_offset(dn, (hole ? DNODE_FIND_HOLE : 0), off, 1, 1, 0); 1751 dnode_rele(dn, FTAG); 1752 1753 return (err); 1754} 1755 1756void 1757dmu_object_info_from_dnode(dnode_t *dn, dmu_object_info_t *doi) 1758{ 1759 dnode_phys_t *dnp; 1760 1761 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1762 mutex_enter(&dn->dn_mtx); 1763 1764 dnp = dn->dn_phys; 1765 1766 doi->doi_data_block_size = dn->dn_datablksz; 1767 doi->doi_metadata_block_size = dn->dn_indblkshift ? 1768 1ULL << dn->dn_indblkshift : 0; 1769 doi->doi_type = dn->dn_type; 1770 doi->doi_bonus_type = dn->dn_bonustype; 1771 doi->doi_bonus_size = dn->dn_bonuslen; 1772 doi->doi_indirection = dn->dn_nlevels; 1773 doi->doi_checksum = dn->dn_checksum; 1774 doi->doi_compress = dn->dn_compress; 1775 doi->doi_physical_blocks_512 = (DN_USED_BYTES(dnp) + 256) >> 9; 1776 doi->doi_max_offset = (dn->dn_maxblkid + 1) * dn->dn_datablksz; 1777 doi->doi_fill_count = 0; 1778 for (int i = 0; i < dnp->dn_nblkptr; i++) 1779 doi->doi_fill_count += dnp->dn_blkptr[i].blk_fill; 1780 1781 mutex_exit(&dn->dn_mtx); 1782 rw_exit(&dn->dn_struct_rwlock); 1783} 1784 1785/* 1786 * Get information on a DMU object. 1787 * If doi is NULL, just indicates whether the object exists. 1788 */ 1789int 1790dmu_object_info(objset_t *os, uint64_t object, dmu_object_info_t *doi) 1791{ 1792 dnode_t *dn; 1793 int err = dnode_hold(os, object, FTAG, &dn); 1794 1795 if (err) 1796 return (err); 1797 1798 if (doi != NULL) 1799 dmu_object_info_from_dnode(dn, doi); 1800 1801 dnode_rele(dn, FTAG); 1802 return (0); 1803} 1804 1805/* 1806 * As above, but faster; can be used when you have a held dbuf in hand. 1807 */ 1808void 1809dmu_object_info_from_db(dmu_buf_t *db_fake, dmu_object_info_t *doi) 1810{ 1811 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1812 1813 DB_DNODE_ENTER(db); 1814 dmu_object_info_from_dnode(DB_DNODE(db), doi); 1815 DB_DNODE_EXIT(db); 1816} 1817 1818/* 1819 * Faster still when you only care about the size. 1820 * This is specifically optimized for zfs_getattr(). 1821 */ 1822void 1823dmu_object_size_from_db(dmu_buf_t *db_fake, uint32_t *blksize, 1824 u_longlong_t *nblk512) 1825{ 1826 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1827 dnode_t *dn; 1828 1829 DB_DNODE_ENTER(db); 1830 dn = DB_DNODE(db); 1831 1832 *blksize = dn->dn_datablksz; 1833 /* add 1 for dnode space */ 1834 *nblk512 = ((DN_USED_BYTES(dn->dn_phys) + SPA_MINBLOCKSIZE/2) >> 1835 SPA_MINBLOCKSHIFT) + 1; 1836 DB_DNODE_EXIT(db); 1837} 1838 1839void 1840byteswap_uint64_array(void *vbuf, size_t size) 1841{ 1842 uint64_t *buf = vbuf; 1843 size_t count = size >> 3; 1844 int i; 1845 1846 ASSERT((size & 7) == 0); 1847 1848 for (i = 0; i < count; i++) 1849 buf[i] = BSWAP_64(buf[i]); 1850} 1851 1852void 1853byteswap_uint32_array(void *vbuf, size_t size) 1854{ 1855 uint32_t *buf = vbuf; 1856 size_t count = size >> 2; 1857 int i; 1858 1859 ASSERT((size & 3) == 0); 1860 1861 for (i = 0; i < count; i++) 1862 buf[i] = BSWAP_32(buf[i]); 1863} 1864 1865void 1866byteswap_uint16_array(void *vbuf, size_t size) 1867{ 1868 uint16_t *buf = vbuf; 1869 size_t count = size >> 1; 1870 int i; 1871 1872 ASSERT((size & 1) == 0); 1873 1874 for (i = 0; i < count; i++) 1875 buf[i] = BSWAP_16(buf[i]); 1876} 1877 1878/* ARGSUSED */ 1879void 1880byteswap_uint8_array(void *vbuf, size_t size) 1881{ 1882} 1883 1884void 1885dmu_init(void) 1886{ 1887 zfs_dbgmsg_init(); 1888 sa_cache_init(); 1889 xuio_stat_init(); 1890 dmu_objset_init(); 1891 dnode_init(); 1892 dbuf_init(); 1893 zfetch_init(); 1894 zio_compress_init(); 1895 l2arc_init(); 1896 arc_init(); 1897} 1898 1899void 1900dmu_fini(void) 1901{ 1902 arc_fini(); /* arc depends on l2arc, so arc must go first */ 1903 l2arc_fini(); 1904 zfetch_fini(); 1905 zio_compress_fini(); 1906 dbuf_fini(); 1907 dnode_fini(); 1908 dmu_objset_fini(); 1909 xuio_stat_fini(); 1910 sa_cache_fini(); 1911 zfs_dbgmsg_fini(); 1912} 1913