dbuf.c revision 285202
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 2011 Nexenta Systems, Inc. All rights reserved. 24 * Copyright (c) 2012, 2015 by Delphix. All rights reserved. 25 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved. 26 * Copyright (c) 2013, Joyent, Inc. All rights reserved. 27 */ 28 29#include <sys/zfs_context.h> 30#include <sys/dmu.h> 31#include <sys/dmu_send.h> 32#include <sys/dmu_impl.h> 33#include <sys/dbuf.h> 34#include <sys/dmu_objset.h> 35#include <sys/dsl_dataset.h> 36#include <sys/dsl_dir.h> 37#include <sys/dmu_tx.h> 38#include <sys/spa.h> 39#include <sys/zio.h> 40#include <sys/dmu_zfetch.h> 41#include <sys/sa.h> 42#include <sys/sa_impl.h> 43#include <sys/zfeature.h> 44#include <sys/blkptr.h> 45#include <sys/range_tree.h> 46 47/* 48 * Number of times that zfs_free_range() took the slow path while doing 49 * a zfs receive. A nonzero value indicates a potential performance problem. 50 */ 51uint64_t zfs_free_range_recv_miss; 52 53static void dbuf_destroy(dmu_buf_impl_t *db); 54static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx); 55static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx); 56 57/* 58 * Global data structures and functions for the dbuf cache. 59 */ 60static kmem_cache_t *dbuf_cache; 61 62/* ARGSUSED */ 63static int 64dbuf_cons(void *vdb, void *unused, int kmflag) 65{ 66 dmu_buf_impl_t *db = vdb; 67 bzero(db, sizeof (dmu_buf_impl_t)); 68 69 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL); 70 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL); 71 refcount_create(&db->db_holds); 72 73 return (0); 74} 75 76/* ARGSUSED */ 77static void 78dbuf_dest(void *vdb, void *unused) 79{ 80 dmu_buf_impl_t *db = vdb; 81 mutex_destroy(&db->db_mtx); 82 cv_destroy(&db->db_changed); 83 refcount_destroy(&db->db_holds); 84} 85 86/* 87 * dbuf hash table routines 88 */ 89static dbuf_hash_table_t dbuf_hash_table; 90 91static uint64_t dbuf_hash_count; 92 93static uint64_t 94dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid) 95{ 96 uintptr_t osv = (uintptr_t)os; 97 uint64_t crc = -1ULL; 98 99 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 100 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF]; 101 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF]; 102 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF]; 103 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF]; 104 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF]; 105 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF]; 106 107 crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16); 108 109 return (crc); 110} 111 112#define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid); 113 114#define DBUF_EQUAL(dbuf, os, obj, level, blkid) \ 115 ((dbuf)->db.db_object == (obj) && \ 116 (dbuf)->db_objset == (os) && \ 117 (dbuf)->db_level == (level) && \ 118 (dbuf)->db_blkid == (blkid)) 119 120dmu_buf_impl_t * 121dbuf_find(dnode_t *dn, uint8_t level, uint64_t blkid) 122{ 123 dbuf_hash_table_t *h = &dbuf_hash_table; 124 objset_t *os = dn->dn_objset; 125 uint64_t obj = dn->dn_object; 126 uint64_t hv = DBUF_HASH(os, obj, level, blkid); 127 uint64_t idx = hv & h->hash_table_mask; 128 dmu_buf_impl_t *db; 129 130 mutex_enter(DBUF_HASH_MUTEX(h, idx)); 131 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) { 132 if (DBUF_EQUAL(db, os, obj, level, blkid)) { 133 mutex_enter(&db->db_mtx); 134 if (db->db_state != DB_EVICTING) { 135 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 136 return (db); 137 } 138 mutex_exit(&db->db_mtx); 139 } 140 } 141 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 142 return (NULL); 143} 144 145/* 146 * Insert an entry into the hash table. If there is already an element 147 * equal to elem in the hash table, then the already existing element 148 * will be returned and the new element will not be inserted. 149 * Otherwise returns NULL. 150 */ 151static dmu_buf_impl_t * 152dbuf_hash_insert(dmu_buf_impl_t *db) 153{ 154 dbuf_hash_table_t *h = &dbuf_hash_table; 155 objset_t *os = db->db_objset; 156 uint64_t obj = db->db.db_object; 157 int level = db->db_level; 158 uint64_t blkid = db->db_blkid; 159 uint64_t hv = DBUF_HASH(os, obj, level, blkid); 160 uint64_t idx = hv & h->hash_table_mask; 161 dmu_buf_impl_t *dbf; 162 163 mutex_enter(DBUF_HASH_MUTEX(h, idx)); 164 for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) { 165 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) { 166 mutex_enter(&dbf->db_mtx); 167 if (dbf->db_state != DB_EVICTING) { 168 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 169 return (dbf); 170 } 171 mutex_exit(&dbf->db_mtx); 172 } 173 } 174 175 mutex_enter(&db->db_mtx); 176 db->db_hash_next = h->hash_table[idx]; 177 h->hash_table[idx] = db; 178 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 179 atomic_inc_64(&dbuf_hash_count); 180 181 return (NULL); 182} 183 184/* 185 * Remove an entry from the hash table. It must be in the EVICTING state. 186 */ 187static void 188dbuf_hash_remove(dmu_buf_impl_t *db) 189{ 190 dbuf_hash_table_t *h = &dbuf_hash_table; 191 uint64_t hv = DBUF_HASH(db->db_objset, db->db.db_object, 192 db->db_level, db->db_blkid); 193 uint64_t idx = hv & h->hash_table_mask; 194 dmu_buf_impl_t *dbf, **dbp; 195 196 /* 197 * We musn't hold db_mtx to maintain lock ordering: 198 * DBUF_HASH_MUTEX > db_mtx. 199 */ 200 ASSERT(refcount_is_zero(&db->db_holds)); 201 ASSERT(db->db_state == DB_EVICTING); 202 ASSERT(!MUTEX_HELD(&db->db_mtx)); 203 204 mutex_enter(DBUF_HASH_MUTEX(h, idx)); 205 dbp = &h->hash_table[idx]; 206 while ((dbf = *dbp) != db) { 207 dbp = &dbf->db_hash_next; 208 ASSERT(dbf != NULL); 209 } 210 *dbp = db->db_hash_next; 211 db->db_hash_next = NULL; 212 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 213 atomic_dec_64(&dbuf_hash_count); 214} 215 216static arc_evict_func_t dbuf_do_evict; 217 218static void 219dbuf_evict_user(dmu_buf_impl_t *db) 220{ 221 ASSERT(MUTEX_HELD(&db->db_mtx)); 222 223 if (db->db_level != 0 || db->db_evict_func == NULL) 224 return; 225 226 db->db_evict_func(&db->db, db->db_user_ptr); 227 db->db_user_ptr = NULL; 228 db->db_evict_func = NULL; 229} 230 231boolean_t 232dbuf_is_metadata(dmu_buf_impl_t *db) 233{ 234 if (db->db_level > 0) { 235 return (B_TRUE); 236 } else { 237 boolean_t is_metadata; 238 239 DB_DNODE_ENTER(db); 240 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type); 241 DB_DNODE_EXIT(db); 242 243 return (is_metadata); 244 } 245} 246 247void 248dbuf_evict(dmu_buf_impl_t *db) 249{ 250 ASSERT(MUTEX_HELD(&db->db_mtx)); 251 ASSERT(db->db_buf == NULL); 252 ASSERT(db->db_data_pending == NULL); 253 254 dbuf_clear(db); 255 dbuf_destroy(db); 256} 257 258void 259dbuf_init(void) 260{ 261 uint64_t hsize = 1ULL << 16; 262 dbuf_hash_table_t *h = &dbuf_hash_table; 263 int i; 264 265 /* 266 * The hash table is big enough to fill all of physical memory 267 * with an average 4K block size. The table will take up 268 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers). 269 */ 270 while (hsize * 4096 < (uint64_t)physmem * PAGESIZE) 271 hsize <<= 1; 272 273retry: 274 h->hash_table_mask = hsize - 1; 275 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP); 276 if (h->hash_table == NULL) { 277 /* XXX - we should really return an error instead of assert */ 278 ASSERT(hsize > (1ULL << 10)); 279 hsize >>= 1; 280 goto retry; 281 } 282 283 dbuf_cache = kmem_cache_create("dmu_buf_impl_t", 284 sizeof (dmu_buf_impl_t), 285 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0); 286 287 for (i = 0; i < DBUF_MUTEXES; i++) 288 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL); 289} 290 291void 292dbuf_fini(void) 293{ 294 dbuf_hash_table_t *h = &dbuf_hash_table; 295 int i; 296 297 for (i = 0; i < DBUF_MUTEXES; i++) 298 mutex_destroy(&h->hash_mutexes[i]); 299 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *)); 300 kmem_cache_destroy(dbuf_cache); 301} 302 303/* 304 * Other stuff. 305 */ 306 307#ifdef ZFS_DEBUG 308static void 309dbuf_verify(dmu_buf_impl_t *db) 310{ 311 dnode_t *dn; 312 dbuf_dirty_record_t *dr; 313 314 ASSERT(MUTEX_HELD(&db->db_mtx)); 315 316 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY)) 317 return; 318 319 ASSERT(db->db_objset != NULL); 320 DB_DNODE_ENTER(db); 321 dn = DB_DNODE(db); 322 if (dn == NULL) { 323 ASSERT(db->db_parent == NULL); 324 ASSERT(db->db_blkptr == NULL); 325 } else { 326 ASSERT3U(db->db.db_object, ==, dn->dn_object); 327 ASSERT3P(db->db_objset, ==, dn->dn_objset); 328 ASSERT3U(db->db_level, <, dn->dn_nlevels); 329 ASSERT(db->db_blkid == DMU_BONUS_BLKID || 330 db->db_blkid == DMU_SPILL_BLKID || 331 !avl_is_empty(&dn->dn_dbufs)); 332 } 333 if (db->db_blkid == DMU_BONUS_BLKID) { 334 ASSERT(dn != NULL); 335 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 336 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID); 337 } else if (db->db_blkid == DMU_SPILL_BLKID) { 338 ASSERT(dn != NULL); 339 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 340 ASSERT0(db->db.db_offset); 341 } else { 342 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size); 343 } 344 345 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next) 346 ASSERT(dr->dr_dbuf == db); 347 348 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next) 349 ASSERT(dr->dr_dbuf == db); 350 351 /* 352 * We can't assert that db_size matches dn_datablksz because it 353 * can be momentarily different when another thread is doing 354 * dnode_set_blksz(). 355 */ 356 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) { 357 dr = db->db_data_pending; 358 /* 359 * It should only be modified in syncing context, so 360 * make sure we only have one copy of the data. 361 */ 362 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf); 363 } 364 365 /* verify db->db_blkptr */ 366 if (db->db_blkptr) { 367 if (db->db_parent == dn->dn_dbuf) { 368 /* db is pointed to by the dnode */ 369 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */ 370 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object)) 371 ASSERT(db->db_parent == NULL); 372 else 373 ASSERT(db->db_parent != NULL); 374 if (db->db_blkid != DMU_SPILL_BLKID) 375 ASSERT3P(db->db_blkptr, ==, 376 &dn->dn_phys->dn_blkptr[db->db_blkid]); 377 } else { 378 /* db is pointed to by an indirect block */ 379 int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT; 380 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1); 381 ASSERT3U(db->db_parent->db.db_object, ==, 382 db->db.db_object); 383 /* 384 * dnode_grow_indblksz() can make this fail if we don't 385 * have the struct_rwlock. XXX indblksz no longer 386 * grows. safe to do this now? 387 */ 388 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) { 389 ASSERT3P(db->db_blkptr, ==, 390 ((blkptr_t *)db->db_parent->db.db_data + 391 db->db_blkid % epb)); 392 } 393 } 394 } 395 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) && 396 (db->db_buf == NULL || db->db_buf->b_data) && 397 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID && 398 db->db_state != DB_FILL && !dn->dn_free_txg) { 399 /* 400 * If the blkptr isn't set but they have nonzero data, 401 * it had better be dirty, otherwise we'll lose that 402 * data when we evict this buffer. 403 */ 404 if (db->db_dirtycnt == 0) { 405 uint64_t *buf = db->db.db_data; 406 int i; 407 408 for (i = 0; i < db->db.db_size >> 3; i++) { 409 ASSERT(buf[i] == 0); 410 } 411 } 412 } 413 DB_DNODE_EXIT(db); 414} 415#endif 416 417static void 418dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf) 419{ 420 ASSERT(MUTEX_HELD(&db->db_mtx)); 421 db->db_buf = buf; 422 if (buf != NULL) { 423 ASSERT(buf->b_data != NULL); 424 db->db.db_data = buf->b_data; 425 if (!arc_released(buf)) 426 arc_set_callback(buf, dbuf_do_evict, db); 427 } else { 428 dbuf_evict_user(db); 429 db->db.db_data = NULL; 430 if (db->db_state != DB_NOFILL) 431 db->db_state = DB_UNCACHED; 432 } 433} 434 435/* 436 * Loan out an arc_buf for read. Return the loaned arc_buf. 437 */ 438arc_buf_t * 439dbuf_loan_arcbuf(dmu_buf_impl_t *db) 440{ 441 arc_buf_t *abuf; 442 443 mutex_enter(&db->db_mtx); 444 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) { 445 int blksz = db->db.db_size; 446 spa_t *spa = db->db_objset->os_spa; 447 448 mutex_exit(&db->db_mtx); 449 abuf = arc_loan_buf(spa, blksz); 450 bcopy(db->db.db_data, abuf->b_data, blksz); 451 } else { 452 abuf = db->db_buf; 453 arc_loan_inuse_buf(abuf, db); 454 dbuf_set_data(db, NULL); 455 mutex_exit(&db->db_mtx); 456 } 457 return (abuf); 458} 459 460uint64_t 461dbuf_whichblock(dnode_t *dn, uint64_t offset) 462{ 463 if (dn->dn_datablkshift) { 464 return (offset >> dn->dn_datablkshift); 465 } else { 466 ASSERT3U(offset, <, dn->dn_datablksz); 467 return (0); 468 } 469} 470 471static void 472dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb) 473{ 474 dmu_buf_impl_t *db = vdb; 475 476 mutex_enter(&db->db_mtx); 477 ASSERT3U(db->db_state, ==, DB_READ); 478 /* 479 * All reads are synchronous, so we must have a hold on the dbuf 480 */ 481 ASSERT(refcount_count(&db->db_holds) > 0); 482 ASSERT(db->db_buf == NULL); 483 ASSERT(db->db.db_data == NULL); 484 if (db->db_level == 0 && db->db_freed_in_flight) { 485 /* we were freed in flight; disregard any error */ 486 arc_release(buf, db); 487 bzero(buf->b_data, db->db.db_size); 488 arc_buf_freeze(buf); 489 db->db_freed_in_flight = FALSE; 490 dbuf_set_data(db, buf); 491 db->db_state = DB_CACHED; 492 } else if (zio == NULL || zio->io_error == 0) { 493 dbuf_set_data(db, buf); 494 db->db_state = DB_CACHED; 495 } else { 496 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 497 ASSERT3P(db->db_buf, ==, NULL); 498 VERIFY(arc_buf_remove_ref(buf, db)); 499 db->db_state = DB_UNCACHED; 500 } 501 cv_broadcast(&db->db_changed); 502 dbuf_rele_and_unlock(db, NULL); 503} 504 505static void 506dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t *flags) 507{ 508 dnode_t *dn; 509 zbookmark_phys_t zb; 510 arc_flags_t aflags = ARC_FLAG_NOWAIT; 511 512 DB_DNODE_ENTER(db); 513 dn = DB_DNODE(db); 514 ASSERT(!refcount_is_zero(&db->db_holds)); 515 /* We need the struct_rwlock to prevent db_blkptr from changing. */ 516 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 517 ASSERT(MUTEX_HELD(&db->db_mtx)); 518 ASSERT(db->db_state == DB_UNCACHED); 519 ASSERT(db->db_buf == NULL); 520 521 if (db->db_blkid == DMU_BONUS_BLKID) { 522 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen); 523 524 ASSERT3U(bonuslen, <=, db->db.db_size); 525 db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN); 526 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 527 if (bonuslen < DN_MAX_BONUSLEN) 528 bzero(db->db.db_data, DN_MAX_BONUSLEN); 529 if (bonuslen) 530 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen); 531 DB_DNODE_EXIT(db); 532 db->db_state = DB_CACHED; 533 mutex_exit(&db->db_mtx); 534 return; 535 } 536 537 /* 538 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync() 539 * processes the delete record and clears the bp while we are waiting 540 * for the dn_mtx (resulting in a "no" from block_freed). 541 */ 542 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) || 543 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) || 544 BP_IS_HOLE(db->db_blkptr)))) { 545 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 546 547 DB_DNODE_EXIT(db); 548 dbuf_set_data(db, arc_buf_alloc(db->db_objset->os_spa, 549 db->db.db_size, db, type)); 550 bzero(db->db.db_data, db->db.db_size); 551 db->db_state = DB_CACHED; 552 *flags |= DB_RF_CACHED; 553 mutex_exit(&db->db_mtx); 554 return; 555 } 556 557 DB_DNODE_EXIT(db); 558 559 db->db_state = DB_READ; 560 mutex_exit(&db->db_mtx); 561 562 if (DBUF_IS_L2CACHEABLE(db)) 563 aflags |= ARC_FLAG_L2CACHE; 564 if (DBUF_IS_L2COMPRESSIBLE(db)) 565 aflags |= ARC_FLAG_L2COMPRESS; 566 567 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ? 568 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET, 569 db->db.db_object, db->db_level, db->db_blkid); 570 571 dbuf_add_ref(db, NULL); 572 573 (void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr, 574 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ, 575 (*flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED, 576 &aflags, &zb); 577 if (aflags & ARC_FLAG_CACHED) 578 *flags |= DB_RF_CACHED; 579} 580 581int 582dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) 583{ 584 int err = 0; 585 boolean_t havepzio = (zio != NULL); 586 boolean_t prefetch; 587 dnode_t *dn; 588 589 /* 590 * We don't have to hold the mutex to check db_state because it 591 * can't be freed while we have a hold on the buffer. 592 */ 593 ASSERT(!refcount_is_zero(&db->db_holds)); 594 595 if (db->db_state == DB_NOFILL) 596 return (SET_ERROR(EIO)); 597 598 DB_DNODE_ENTER(db); 599 dn = DB_DNODE(db); 600 if ((flags & DB_RF_HAVESTRUCT) == 0) 601 rw_enter(&dn->dn_struct_rwlock, RW_READER); 602 603 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 604 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL && 605 DBUF_IS_CACHEABLE(db); 606 607 mutex_enter(&db->db_mtx); 608 if (db->db_state == DB_CACHED) { 609 mutex_exit(&db->db_mtx); 610 if (prefetch) 611 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset, 612 db->db.db_size, TRUE); 613 if ((flags & DB_RF_HAVESTRUCT) == 0) 614 rw_exit(&dn->dn_struct_rwlock); 615 DB_DNODE_EXIT(db); 616 } else if (db->db_state == DB_UNCACHED) { 617 spa_t *spa = dn->dn_objset->os_spa; 618 619 if (zio == NULL) 620 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); 621 dbuf_read_impl(db, zio, &flags); 622 623 /* dbuf_read_impl has dropped db_mtx for us */ 624 625 if (prefetch) 626 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset, 627 db->db.db_size, flags & DB_RF_CACHED); 628 629 if ((flags & DB_RF_HAVESTRUCT) == 0) 630 rw_exit(&dn->dn_struct_rwlock); 631 DB_DNODE_EXIT(db); 632 633 if (!havepzio) 634 err = zio_wait(zio); 635 } else { 636 /* 637 * Another reader came in while the dbuf was in flight 638 * between UNCACHED and CACHED. Either a writer will finish 639 * writing the buffer (sending the dbuf to CACHED) or the 640 * first reader's request will reach the read_done callback 641 * and send the dbuf to CACHED. Otherwise, a failure 642 * occurred and the dbuf went to UNCACHED. 643 */ 644 mutex_exit(&db->db_mtx); 645 if (prefetch) 646 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset, 647 db->db.db_size, TRUE); 648 if ((flags & DB_RF_HAVESTRUCT) == 0) 649 rw_exit(&dn->dn_struct_rwlock); 650 DB_DNODE_EXIT(db); 651 652 /* Skip the wait per the caller's request. */ 653 mutex_enter(&db->db_mtx); 654 if ((flags & DB_RF_NEVERWAIT) == 0) { 655 while (db->db_state == DB_READ || 656 db->db_state == DB_FILL) { 657 ASSERT(db->db_state == DB_READ || 658 (flags & DB_RF_HAVESTRUCT) == 0); 659 DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *, 660 db, zio_t *, zio); 661 cv_wait(&db->db_changed, &db->db_mtx); 662 } 663 if (db->db_state == DB_UNCACHED) 664 err = SET_ERROR(EIO); 665 } 666 mutex_exit(&db->db_mtx); 667 } 668 669 ASSERT(err || havepzio || db->db_state == DB_CACHED); 670 return (err); 671} 672 673static void 674dbuf_noread(dmu_buf_impl_t *db) 675{ 676 ASSERT(!refcount_is_zero(&db->db_holds)); 677 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 678 mutex_enter(&db->db_mtx); 679 while (db->db_state == DB_READ || db->db_state == DB_FILL) 680 cv_wait(&db->db_changed, &db->db_mtx); 681 if (db->db_state == DB_UNCACHED) { 682 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 683 spa_t *spa = db->db_objset->os_spa; 684 685 ASSERT(db->db_buf == NULL); 686 ASSERT(db->db.db_data == NULL); 687 dbuf_set_data(db, arc_buf_alloc(spa, db->db.db_size, db, type)); 688 db->db_state = DB_FILL; 689 } else if (db->db_state == DB_NOFILL) { 690 dbuf_set_data(db, NULL); 691 } else { 692 ASSERT3U(db->db_state, ==, DB_CACHED); 693 } 694 mutex_exit(&db->db_mtx); 695} 696 697/* 698 * This is our just-in-time copy function. It makes a copy of 699 * buffers, that have been modified in a previous transaction 700 * group, before we modify them in the current active group. 701 * 702 * This function is used in two places: when we are dirtying a 703 * buffer for the first time in a txg, and when we are freeing 704 * a range in a dnode that includes this buffer. 705 * 706 * Note that when we are called from dbuf_free_range() we do 707 * not put a hold on the buffer, we just traverse the active 708 * dbuf list for the dnode. 709 */ 710static void 711dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg) 712{ 713 dbuf_dirty_record_t *dr = db->db_last_dirty; 714 715 ASSERT(MUTEX_HELD(&db->db_mtx)); 716 ASSERT(db->db.db_data != NULL); 717 ASSERT(db->db_level == 0); 718 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT); 719 720 if (dr == NULL || 721 (dr->dt.dl.dr_data != 722 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf))) 723 return; 724 725 /* 726 * If the last dirty record for this dbuf has not yet synced 727 * and its referencing the dbuf data, either: 728 * reset the reference to point to a new copy, 729 * or (if there a no active holders) 730 * just null out the current db_data pointer. 731 */ 732 ASSERT(dr->dr_txg >= txg - 2); 733 if (db->db_blkid == DMU_BONUS_BLKID) { 734 /* Note that the data bufs here are zio_bufs */ 735 dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN); 736 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 737 bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN); 738 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) { 739 int size = db->db.db_size; 740 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 741 spa_t *spa = db->db_objset->os_spa; 742 743 dr->dt.dl.dr_data = arc_buf_alloc(spa, size, db, type); 744 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size); 745 } else { 746 dbuf_set_data(db, NULL); 747 } 748} 749 750void 751dbuf_unoverride(dbuf_dirty_record_t *dr) 752{ 753 dmu_buf_impl_t *db = dr->dr_dbuf; 754 blkptr_t *bp = &dr->dt.dl.dr_overridden_by; 755 uint64_t txg = dr->dr_txg; 756 757 ASSERT(MUTEX_HELD(&db->db_mtx)); 758 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC); 759 ASSERT(db->db_level == 0); 760 761 if (db->db_blkid == DMU_BONUS_BLKID || 762 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN) 763 return; 764 765 ASSERT(db->db_data_pending != dr); 766 767 /* free this block */ 768 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite) 769 zio_free(db->db_objset->os_spa, txg, bp); 770 771 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 772 dr->dt.dl.dr_nopwrite = B_FALSE; 773 774 /* 775 * Release the already-written buffer, so we leave it in 776 * a consistent dirty state. Note that all callers are 777 * modifying the buffer, so they will immediately do 778 * another (redundant) arc_release(). Therefore, leave 779 * the buf thawed to save the effort of freezing & 780 * immediately re-thawing it. 781 */ 782 arc_release(dr->dt.dl.dr_data, db); 783} 784 785/* 786 * Evict (if its unreferenced) or clear (if its referenced) any level-0 787 * data blocks in the free range, so that any future readers will find 788 * empty blocks. 789 * 790 * This is a no-op if the dataset is in the middle of an incremental 791 * receive; see comment below for details. 792 */ 793void 794dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid, 795 dmu_tx_t *tx) 796{ 797 dmu_buf_impl_t *db, *db_next, db_search; 798 uint64_t txg = tx->tx_txg; 799 avl_index_t where; 800 801 if (end_blkid > dn->dn_maxblkid && (end_blkid != DMU_SPILL_BLKID)) 802 end_blkid = dn->dn_maxblkid; 803 dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid); 804 805 db_search.db_level = 0; 806 db_search.db_blkid = start_blkid; 807 db_search.db_state = DB_SEARCH; 808 809 mutex_enter(&dn->dn_dbufs_mtx); 810 if (start_blkid >= dn->dn_unlisted_l0_blkid) { 811 /* There can't be any dbufs in this range; no need to search. */ 812#ifdef DEBUG 813 db = avl_find(&dn->dn_dbufs, &db_search, &where); 814 ASSERT3P(db, ==, NULL); 815 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER); 816 ASSERT(db == NULL || db->db_level > 0); 817#endif 818 mutex_exit(&dn->dn_dbufs_mtx); 819 return; 820 } else if (dmu_objset_is_receiving(dn->dn_objset)) { 821 /* 822 * If we are receiving, we expect there to be no dbufs in 823 * the range to be freed, because receive modifies each 824 * block at most once, and in offset order. If this is 825 * not the case, it can lead to performance problems, 826 * so note that we unexpectedly took the slow path. 827 */ 828 atomic_inc_64(&zfs_free_range_recv_miss); 829 } 830 831 db = avl_find(&dn->dn_dbufs, &db_search, &where); 832 ASSERT3P(db, ==, NULL); 833 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER); 834 835 for (; db != NULL; db = db_next) { 836 db_next = AVL_NEXT(&dn->dn_dbufs, db); 837 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 838 839 if (db->db_level != 0 || db->db_blkid > end_blkid) { 840 break; 841 } 842 ASSERT3U(db->db_blkid, >=, start_blkid); 843 844 /* found a level 0 buffer in the range */ 845 mutex_enter(&db->db_mtx); 846 if (dbuf_undirty(db, tx)) { 847 /* mutex has been dropped and dbuf destroyed */ 848 continue; 849 } 850 851 if (db->db_state == DB_UNCACHED || 852 db->db_state == DB_NOFILL || 853 db->db_state == DB_EVICTING) { 854 ASSERT(db->db.db_data == NULL); 855 mutex_exit(&db->db_mtx); 856 continue; 857 } 858 if (db->db_state == DB_READ || db->db_state == DB_FILL) { 859 /* will be handled in dbuf_read_done or dbuf_rele */ 860 db->db_freed_in_flight = TRUE; 861 mutex_exit(&db->db_mtx); 862 continue; 863 } 864 if (refcount_count(&db->db_holds) == 0) { 865 ASSERT(db->db_buf); 866 dbuf_clear(db); 867 continue; 868 } 869 /* The dbuf is referenced */ 870 871 if (db->db_last_dirty != NULL) { 872 dbuf_dirty_record_t *dr = db->db_last_dirty; 873 874 if (dr->dr_txg == txg) { 875 /* 876 * This buffer is "in-use", re-adjust the file 877 * size to reflect that this buffer may 878 * contain new data when we sync. 879 */ 880 if (db->db_blkid != DMU_SPILL_BLKID && 881 db->db_blkid > dn->dn_maxblkid) 882 dn->dn_maxblkid = db->db_blkid; 883 dbuf_unoverride(dr); 884 } else { 885 /* 886 * This dbuf is not dirty in the open context. 887 * Either uncache it (if its not referenced in 888 * the open context) or reset its contents to 889 * empty. 890 */ 891 dbuf_fix_old_data(db, txg); 892 } 893 } 894 /* clear the contents if its cached */ 895 if (db->db_state == DB_CACHED) { 896 ASSERT(db->db.db_data != NULL); 897 arc_release(db->db_buf, db); 898 bzero(db->db.db_data, db->db.db_size); 899 arc_buf_freeze(db->db_buf); 900 } 901 902 mutex_exit(&db->db_mtx); 903 } 904 mutex_exit(&dn->dn_dbufs_mtx); 905} 906 907static int 908dbuf_block_freeable(dmu_buf_impl_t *db) 909{ 910 dsl_dataset_t *ds = db->db_objset->os_dsl_dataset; 911 uint64_t birth_txg = 0; 912 913 /* 914 * We don't need any locking to protect db_blkptr: 915 * If it's syncing, then db_last_dirty will be set 916 * so we'll ignore db_blkptr. 917 * 918 * This logic ensures that only block births for 919 * filled blocks are considered. 920 */ 921 ASSERT(MUTEX_HELD(&db->db_mtx)); 922 if (db->db_last_dirty && (db->db_blkptr == NULL || 923 !BP_IS_HOLE(db->db_blkptr))) { 924 birth_txg = db->db_last_dirty->dr_txg; 925 } else if (db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) { 926 birth_txg = db->db_blkptr->blk_birth; 927 } 928 929 /* 930 * If this block don't exist or is in a snapshot, it can't be freed. 931 * Don't pass the bp to dsl_dataset_block_freeable() since we 932 * are holding the db_mtx lock and might deadlock if we are 933 * prefetching a dedup-ed block. 934 */ 935 if (birth_txg != 0) 936 return (ds == NULL || 937 dsl_dataset_block_freeable(ds, NULL, birth_txg)); 938 else 939 return (B_FALSE); 940} 941 942void 943dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx) 944{ 945 arc_buf_t *buf, *obuf; 946 int osize = db->db.db_size; 947 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 948 dnode_t *dn; 949 950 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 951 952 DB_DNODE_ENTER(db); 953 dn = DB_DNODE(db); 954 955 /* XXX does *this* func really need the lock? */ 956 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 957 958 /* 959 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held 960 * is OK, because there can be no other references to the db 961 * when we are changing its size, so no concurrent DB_FILL can 962 * be happening. 963 */ 964 /* 965 * XXX we should be doing a dbuf_read, checking the return 966 * value and returning that up to our callers 967 */ 968 dmu_buf_will_dirty(&db->db, tx); 969 970 /* create the data buffer for the new block */ 971 buf = arc_buf_alloc(dn->dn_objset->os_spa, size, db, type); 972 973 /* copy old block data to the new block */ 974 obuf = db->db_buf; 975 bcopy(obuf->b_data, buf->b_data, MIN(osize, size)); 976 /* zero the remainder */ 977 if (size > osize) 978 bzero((uint8_t *)buf->b_data + osize, size - osize); 979 980 mutex_enter(&db->db_mtx); 981 dbuf_set_data(db, buf); 982 VERIFY(arc_buf_remove_ref(obuf, db)); 983 db->db.db_size = size; 984 985 if (db->db_level == 0) { 986 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); 987 db->db_last_dirty->dt.dl.dr_data = buf; 988 } 989 mutex_exit(&db->db_mtx); 990 991 dnode_willuse_space(dn, size-osize, tx); 992 DB_DNODE_EXIT(db); 993} 994 995void 996dbuf_release_bp(dmu_buf_impl_t *db) 997{ 998 objset_t *os = db->db_objset; 999 1000 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os))); 1001 ASSERT(arc_released(os->os_phys_buf) || 1002 list_link_active(&os->os_dsl_dataset->ds_synced_link)); 1003 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf)); 1004 1005 (void) arc_release(db->db_buf, db); 1006} 1007 1008dbuf_dirty_record_t * 1009dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1010{ 1011 dnode_t *dn; 1012 objset_t *os; 1013 dbuf_dirty_record_t **drp, *dr; 1014 int drop_struct_lock = FALSE; 1015 boolean_t do_free_accounting = B_FALSE; 1016 int txgoff = tx->tx_txg & TXG_MASK; 1017 1018 ASSERT(tx->tx_txg != 0); 1019 ASSERT(!refcount_is_zero(&db->db_holds)); 1020 DMU_TX_DIRTY_BUF(tx, db); 1021 1022 DB_DNODE_ENTER(db); 1023 dn = DB_DNODE(db); 1024 /* 1025 * Shouldn't dirty a regular buffer in syncing context. Private 1026 * objects may be dirtied in syncing context, but only if they 1027 * were already pre-dirtied in open context. 1028 */ 1029 ASSERT(!dmu_tx_is_syncing(tx) || 1030 BP_IS_HOLE(dn->dn_objset->os_rootbp) || 1031 DMU_OBJECT_IS_SPECIAL(dn->dn_object) || 1032 dn->dn_objset->os_dsl_dataset == NULL); 1033 /* 1034 * We make this assert for private objects as well, but after we 1035 * check if we're already dirty. They are allowed to re-dirty 1036 * in syncing context. 1037 */ 1038 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 1039 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 1040 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 1041 1042 mutex_enter(&db->db_mtx); 1043 /* 1044 * XXX make this true for indirects too? The problem is that 1045 * transactions created with dmu_tx_create_assigned() from 1046 * syncing context don't bother holding ahead. 1047 */ 1048 ASSERT(db->db_level != 0 || 1049 db->db_state == DB_CACHED || db->db_state == DB_FILL || 1050 db->db_state == DB_NOFILL); 1051 1052 mutex_enter(&dn->dn_mtx); 1053 /* 1054 * Don't set dirtyctx to SYNC if we're just modifying this as we 1055 * initialize the objset. 1056 */ 1057 if (dn->dn_dirtyctx == DN_UNDIRTIED && 1058 !BP_IS_HOLE(dn->dn_objset->os_rootbp)) { 1059 dn->dn_dirtyctx = 1060 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN); 1061 ASSERT(dn->dn_dirtyctx_firstset == NULL); 1062 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP); 1063 } 1064 mutex_exit(&dn->dn_mtx); 1065 1066 if (db->db_blkid == DMU_SPILL_BLKID) 1067 dn->dn_have_spill = B_TRUE; 1068 1069 /* 1070 * If this buffer is already dirty, we're done. 1071 */ 1072 drp = &db->db_last_dirty; 1073 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg || 1074 db->db.db_object == DMU_META_DNODE_OBJECT); 1075 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg) 1076 drp = &dr->dr_next; 1077 if (dr && dr->dr_txg == tx->tx_txg) { 1078 DB_DNODE_EXIT(db); 1079 1080 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) { 1081 /* 1082 * If this buffer has already been written out, 1083 * we now need to reset its state. 1084 */ 1085 dbuf_unoverride(dr); 1086 if (db->db.db_object != DMU_META_DNODE_OBJECT && 1087 db->db_state != DB_NOFILL) 1088 arc_buf_thaw(db->db_buf); 1089 } 1090 mutex_exit(&db->db_mtx); 1091 return (dr); 1092 } 1093 1094 /* 1095 * Only valid if not already dirty. 1096 */ 1097 ASSERT(dn->dn_object == 0 || 1098 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 1099 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 1100 1101 ASSERT3U(dn->dn_nlevels, >, db->db_level); 1102 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) || 1103 dn->dn_phys->dn_nlevels > db->db_level || 1104 dn->dn_next_nlevels[txgoff] > db->db_level || 1105 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level || 1106 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level); 1107 1108 /* 1109 * We should only be dirtying in syncing context if it's the 1110 * mos or we're initializing the os or it's a special object. 1111 * However, we are allowed to dirty in syncing context provided 1112 * we already dirtied it in open context. Hence we must make 1113 * this assertion only if we're not already dirty. 1114 */ 1115 os = dn->dn_objset; 1116 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) || 1117 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp)); 1118 ASSERT(db->db.db_size != 0); 1119 1120 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 1121 1122 if (db->db_blkid != DMU_BONUS_BLKID) { 1123 /* 1124 * Update the accounting. 1125 * Note: we delay "free accounting" until after we drop 1126 * the db_mtx. This keeps us from grabbing other locks 1127 * (and possibly deadlocking) in bp_get_dsize() while 1128 * also holding the db_mtx. 1129 */ 1130 dnode_willuse_space(dn, db->db.db_size, tx); 1131 do_free_accounting = dbuf_block_freeable(db); 1132 } 1133 1134 /* 1135 * If this buffer is dirty in an old transaction group we need 1136 * to make a copy of it so that the changes we make in this 1137 * transaction group won't leak out when we sync the older txg. 1138 */ 1139 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP); 1140 if (db->db_level == 0) { 1141 void *data_old = db->db_buf; 1142 1143 if (db->db_state != DB_NOFILL) { 1144 if (db->db_blkid == DMU_BONUS_BLKID) { 1145 dbuf_fix_old_data(db, tx->tx_txg); 1146 data_old = db->db.db_data; 1147 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) { 1148 /* 1149 * Release the data buffer from the cache so 1150 * that we can modify it without impacting 1151 * possible other users of this cached data 1152 * block. Note that indirect blocks and 1153 * private objects are not released until the 1154 * syncing state (since they are only modified 1155 * then). 1156 */ 1157 arc_release(db->db_buf, db); 1158 dbuf_fix_old_data(db, tx->tx_txg); 1159 data_old = db->db_buf; 1160 } 1161 ASSERT(data_old != NULL); 1162 } 1163 dr->dt.dl.dr_data = data_old; 1164 } else { 1165 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL); 1166 list_create(&dr->dt.di.dr_children, 1167 sizeof (dbuf_dirty_record_t), 1168 offsetof(dbuf_dirty_record_t, dr_dirty_node)); 1169 } 1170 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL) 1171 dr->dr_accounted = db->db.db_size; 1172 dr->dr_dbuf = db; 1173 dr->dr_txg = tx->tx_txg; 1174 dr->dr_next = *drp; 1175 *drp = dr; 1176 1177 /* 1178 * We could have been freed_in_flight between the dbuf_noread 1179 * and dbuf_dirty. We win, as though the dbuf_noread() had 1180 * happened after the free. 1181 */ 1182 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 1183 db->db_blkid != DMU_SPILL_BLKID) { 1184 mutex_enter(&dn->dn_mtx); 1185 if (dn->dn_free_ranges[txgoff] != NULL) { 1186 range_tree_clear(dn->dn_free_ranges[txgoff], 1187 db->db_blkid, 1); 1188 } 1189 mutex_exit(&dn->dn_mtx); 1190 db->db_freed_in_flight = FALSE; 1191 } 1192 1193 /* 1194 * This buffer is now part of this txg 1195 */ 1196 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg); 1197 db->db_dirtycnt += 1; 1198 ASSERT3U(db->db_dirtycnt, <=, 3); 1199 1200 mutex_exit(&db->db_mtx); 1201 1202 if (db->db_blkid == DMU_BONUS_BLKID || 1203 db->db_blkid == DMU_SPILL_BLKID) { 1204 mutex_enter(&dn->dn_mtx); 1205 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1206 list_insert_tail(&dn->dn_dirty_records[txgoff], dr); 1207 mutex_exit(&dn->dn_mtx); 1208 dnode_setdirty(dn, tx); 1209 DB_DNODE_EXIT(db); 1210 return (dr); 1211 } else if (do_free_accounting) { 1212 blkptr_t *bp = db->db_blkptr; 1213 int64_t willfree = (bp && !BP_IS_HOLE(bp)) ? 1214 bp_get_dsize(os->os_spa, bp) : db->db.db_size; 1215 /* 1216 * This is only a guess -- if the dbuf is dirty 1217 * in a previous txg, we don't know how much 1218 * space it will use on disk yet. We should 1219 * really have the struct_rwlock to access 1220 * db_blkptr, but since this is just a guess, 1221 * it's OK if we get an odd answer. 1222 */ 1223 ddt_prefetch(os->os_spa, bp); 1224 dnode_willuse_space(dn, -willfree, tx); 1225 } 1226 1227 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) { 1228 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1229 drop_struct_lock = TRUE; 1230 } 1231 1232 if (db->db_level == 0) { 1233 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock); 1234 ASSERT(dn->dn_maxblkid >= db->db_blkid); 1235 } 1236 1237 if (db->db_level+1 < dn->dn_nlevels) { 1238 dmu_buf_impl_t *parent = db->db_parent; 1239 dbuf_dirty_record_t *di; 1240 int parent_held = FALSE; 1241 1242 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) { 1243 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1244 1245 parent = dbuf_hold_level(dn, db->db_level+1, 1246 db->db_blkid >> epbs, FTAG); 1247 ASSERT(parent != NULL); 1248 parent_held = TRUE; 1249 } 1250 if (drop_struct_lock) 1251 rw_exit(&dn->dn_struct_rwlock); 1252 ASSERT3U(db->db_level+1, ==, parent->db_level); 1253 di = dbuf_dirty(parent, tx); 1254 if (parent_held) 1255 dbuf_rele(parent, FTAG); 1256 1257 mutex_enter(&db->db_mtx); 1258 /* 1259 * Since we've dropped the mutex, it's possible that 1260 * dbuf_undirty() might have changed this out from under us. 1261 */ 1262 if (db->db_last_dirty == dr || 1263 dn->dn_object == DMU_META_DNODE_OBJECT) { 1264 mutex_enter(&di->dt.di.dr_mtx); 1265 ASSERT3U(di->dr_txg, ==, tx->tx_txg); 1266 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1267 list_insert_tail(&di->dt.di.dr_children, dr); 1268 mutex_exit(&di->dt.di.dr_mtx); 1269 dr->dr_parent = di; 1270 } 1271 mutex_exit(&db->db_mtx); 1272 } else { 1273 ASSERT(db->db_level+1 == dn->dn_nlevels); 1274 ASSERT(db->db_blkid < dn->dn_nblkptr); 1275 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf); 1276 mutex_enter(&dn->dn_mtx); 1277 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1278 list_insert_tail(&dn->dn_dirty_records[txgoff], dr); 1279 mutex_exit(&dn->dn_mtx); 1280 if (drop_struct_lock) 1281 rw_exit(&dn->dn_struct_rwlock); 1282 } 1283 1284 dnode_setdirty(dn, tx); 1285 DB_DNODE_EXIT(db); 1286 return (dr); 1287} 1288 1289/* 1290 * Undirty a buffer in the transaction group referenced by the given 1291 * transaction. Return whether this evicted the dbuf. 1292 */ 1293static boolean_t 1294dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1295{ 1296 dnode_t *dn; 1297 uint64_t txg = tx->tx_txg; 1298 dbuf_dirty_record_t *dr, **drp; 1299 1300 ASSERT(txg != 0); 1301 1302 /* 1303 * Due to our use of dn_nlevels below, this can only be called 1304 * in open context, unless we are operating on the MOS. 1305 * From syncing context, dn_nlevels may be different from the 1306 * dn_nlevels used when dbuf was dirtied. 1307 */ 1308 ASSERT(db->db_objset == 1309 dmu_objset_pool(db->db_objset)->dp_meta_objset || 1310 txg != spa_syncing_txg(dmu_objset_spa(db->db_objset))); 1311 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1312 ASSERT0(db->db_level); 1313 ASSERT(MUTEX_HELD(&db->db_mtx)); 1314 1315 /* 1316 * If this buffer is not dirty, we're done. 1317 */ 1318 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next) 1319 if (dr->dr_txg <= txg) 1320 break; 1321 if (dr == NULL || dr->dr_txg < txg) 1322 return (B_FALSE); 1323 ASSERT(dr->dr_txg == txg); 1324 ASSERT(dr->dr_dbuf == db); 1325 1326 DB_DNODE_ENTER(db); 1327 dn = DB_DNODE(db); 1328 1329 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 1330 1331 ASSERT(db->db.db_size != 0); 1332 1333 dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset), 1334 dr->dr_accounted, txg); 1335 1336 *drp = dr->dr_next; 1337 1338 /* 1339 * Note that there are three places in dbuf_dirty() 1340 * where this dirty record may be put on a list. 1341 * Make sure to do a list_remove corresponding to 1342 * every one of those list_insert calls. 1343 */ 1344 if (dr->dr_parent) { 1345 mutex_enter(&dr->dr_parent->dt.di.dr_mtx); 1346 list_remove(&dr->dr_parent->dt.di.dr_children, dr); 1347 mutex_exit(&dr->dr_parent->dt.di.dr_mtx); 1348 } else if (db->db_blkid == DMU_SPILL_BLKID || 1349 db->db_level + 1 == dn->dn_nlevels) { 1350 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf); 1351 mutex_enter(&dn->dn_mtx); 1352 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr); 1353 mutex_exit(&dn->dn_mtx); 1354 } 1355 DB_DNODE_EXIT(db); 1356 1357 if (db->db_state != DB_NOFILL) { 1358 dbuf_unoverride(dr); 1359 1360 ASSERT(db->db_buf != NULL); 1361 ASSERT(dr->dt.dl.dr_data != NULL); 1362 if (dr->dt.dl.dr_data != db->db_buf) 1363 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db)); 1364 } 1365 1366 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 1367 1368 ASSERT(db->db_dirtycnt > 0); 1369 db->db_dirtycnt -= 1; 1370 1371 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) { 1372 arc_buf_t *buf = db->db_buf; 1373 1374 ASSERT(db->db_state == DB_NOFILL || arc_released(buf)); 1375 dbuf_set_data(db, NULL); 1376 VERIFY(arc_buf_remove_ref(buf, db)); 1377 dbuf_evict(db); 1378 return (B_TRUE); 1379 } 1380 1381 return (B_FALSE); 1382} 1383 1384void 1385dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx) 1386{ 1387 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1388 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH; 1389 1390 ASSERT(tx->tx_txg != 0); 1391 ASSERT(!refcount_is_zero(&db->db_holds)); 1392 1393 DB_DNODE_ENTER(db); 1394 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock)) 1395 rf |= DB_RF_HAVESTRUCT; 1396 DB_DNODE_EXIT(db); 1397 (void) dbuf_read(db, NULL, rf); 1398 (void) dbuf_dirty(db, tx); 1399} 1400 1401void 1402dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) 1403{ 1404 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1405 1406 db->db_state = DB_NOFILL; 1407 1408 dmu_buf_will_fill(db_fake, tx); 1409} 1410 1411void 1412dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) 1413{ 1414 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1415 1416 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1417 ASSERT(tx->tx_txg != 0); 1418 ASSERT(db->db_level == 0); 1419 ASSERT(!refcount_is_zero(&db->db_holds)); 1420 1421 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT || 1422 dmu_tx_private_ok(tx)); 1423 1424 dbuf_noread(db); 1425 (void) dbuf_dirty(db, tx); 1426} 1427 1428#pragma weak dmu_buf_fill_done = dbuf_fill_done 1429/* ARGSUSED */ 1430void 1431dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx) 1432{ 1433 mutex_enter(&db->db_mtx); 1434 DBUF_VERIFY(db); 1435 1436 if (db->db_state == DB_FILL) { 1437 if (db->db_level == 0 && db->db_freed_in_flight) { 1438 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1439 /* we were freed while filling */ 1440 /* XXX dbuf_undirty? */ 1441 bzero(db->db.db_data, db->db.db_size); 1442 db->db_freed_in_flight = FALSE; 1443 } 1444 db->db_state = DB_CACHED; 1445 cv_broadcast(&db->db_changed); 1446 } 1447 mutex_exit(&db->db_mtx); 1448} 1449 1450void 1451dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data, 1452 bp_embedded_type_t etype, enum zio_compress comp, 1453 int uncompressed_size, int compressed_size, int byteorder, 1454 dmu_tx_t *tx) 1455{ 1456 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; 1457 struct dirty_leaf *dl; 1458 dmu_object_type_t type; 1459 1460 DB_DNODE_ENTER(db); 1461 type = DB_DNODE(db)->dn_type; 1462 DB_DNODE_EXIT(db); 1463 1464 ASSERT0(db->db_level); 1465 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1466 1467 dmu_buf_will_not_fill(dbuf, tx); 1468 1469 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); 1470 dl = &db->db_last_dirty->dt.dl; 1471 encode_embedded_bp_compressed(&dl->dr_overridden_by, 1472 data, comp, uncompressed_size, compressed_size); 1473 BPE_SET_ETYPE(&dl->dr_overridden_by, etype); 1474 BP_SET_TYPE(&dl->dr_overridden_by, type); 1475 BP_SET_LEVEL(&dl->dr_overridden_by, 0); 1476 BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder); 1477 1478 dl->dr_override_state = DR_OVERRIDDEN; 1479 dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg; 1480} 1481 1482/* 1483 * Directly assign a provided arc buf to a given dbuf if it's not referenced 1484 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf. 1485 */ 1486void 1487dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx) 1488{ 1489 ASSERT(!refcount_is_zero(&db->db_holds)); 1490 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1491 ASSERT(db->db_level == 0); 1492 ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA); 1493 ASSERT(buf != NULL); 1494 ASSERT(arc_buf_size(buf) == db->db.db_size); 1495 ASSERT(tx->tx_txg != 0); 1496 1497 arc_return_buf(buf, db); 1498 ASSERT(arc_released(buf)); 1499 1500 mutex_enter(&db->db_mtx); 1501 1502 while (db->db_state == DB_READ || db->db_state == DB_FILL) 1503 cv_wait(&db->db_changed, &db->db_mtx); 1504 1505 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED); 1506 1507 if (db->db_state == DB_CACHED && 1508 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) { 1509 mutex_exit(&db->db_mtx); 1510 (void) dbuf_dirty(db, tx); 1511 bcopy(buf->b_data, db->db.db_data, db->db.db_size); 1512 VERIFY(arc_buf_remove_ref(buf, db)); 1513 xuio_stat_wbuf_copied(); 1514 return; 1515 } 1516 1517 xuio_stat_wbuf_nocopy(); 1518 if (db->db_state == DB_CACHED) { 1519 dbuf_dirty_record_t *dr = db->db_last_dirty; 1520 1521 ASSERT(db->db_buf != NULL); 1522 if (dr != NULL && dr->dr_txg == tx->tx_txg) { 1523 ASSERT(dr->dt.dl.dr_data == db->db_buf); 1524 if (!arc_released(db->db_buf)) { 1525 ASSERT(dr->dt.dl.dr_override_state == 1526 DR_OVERRIDDEN); 1527 arc_release(db->db_buf, db); 1528 } 1529 dr->dt.dl.dr_data = buf; 1530 VERIFY(arc_buf_remove_ref(db->db_buf, db)); 1531 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) { 1532 arc_release(db->db_buf, db); 1533 VERIFY(arc_buf_remove_ref(db->db_buf, db)); 1534 } 1535 db->db_buf = NULL; 1536 } 1537 ASSERT(db->db_buf == NULL); 1538 dbuf_set_data(db, buf); 1539 db->db_state = DB_FILL; 1540 mutex_exit(&db->db_mtx); 1541 (void) dbuf_dirty(db, tx); 1542 dmu_buf_fill_done(&db->db, tx); 1543} 1544 1545/* 1546 * "Clear" the contents of this dbuf. This will mark the dbuf 1547 * EVICTING and clear *most* of its references. Unfortunately, 1548 * when we are not holding the dn_dbufs_mtx, we can't clear the 1549 * entry in the dn_dbufs list. We have to wait until dbuf_destroy() 1550 * in this case. For callers from the DMU we will usually see: 1551 * dbuf_clear()->arc_clear_callback()->dbuf_do_evict()->dbuf_destroy() 1552 * For the arc callback, we will usually see: 1553 * dbuf_do_evict()->dbuf_clear();dbuf_destroy() 1554 * Sometimes, though, we will get a mix of these two: 1555 * DMU: dbuf_clear()->arc_clear_callback() 1556 * ARC: dbuf_do_evict()->dbuf_destroy() 1557 * 1558 * This routine will dissociate the dbuf from the arc, by calling 1559 * arc_clear_callback(), but will not evict the data from the ARC. 1560 */ 1561void 1562dbuf_clear(dmu_buf_impl_t *db) 1563{ 1564 dnode_t *dn; 1565 dmu_buf_impl_t *parent = db->db_parent; 1566 dmu_buf_impl_t *dndb; 1567 boolean_t dbuf_gone = B_FALSE; 1568 1569 ASSERT(MUTEX_HELD(&db->db_mtx)); 1570 ASSERT(refcount_is_zero(&db->db_holds)); 1571 1572 dbuf_evict_user(db); 1573 1574 if (db->db_state == DB_CACHED) { 1575 ASSERT(db->db.db_data != NULL); 1576 if (db->db_blkid == DMU_BONUS_BLKID) { 1577 zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN); 1578 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 1579 } 1580 db->db.db_data = NULL; 1581 db->db_state = DB_UNCACHED; 1582 } 1583 1584 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL); 1585 ASSERT(db->db_data_pending == NULL); 1586 1587 db->db_state = DB_EVICTING; 1588 db->db_blkptr = NULL; 1589 1590 DB_DNODE_ENTER(db); 1591 dn = DB_DNODE(db); 1592 dndb = dn->dn_dbuf; 1593 if (db->db_blkid != DMU_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) { 1594 avl_remove(&dn->dn_dbufs, db); 1595 atomic_dec_32(&dn->dn_dbufs_count); 1596 membar_producer(); 1597 DB_DNODE_EXIT(db); 1598 /* 1599 * Decrementing the dbuf count means that the hold corresponding 1600 * to the removed dbuf is no longer discounted in dnode_move(), 1601 * so the dnode cannot be moved until after we release the hold. 1602 * The membar_producer() ensures visibility of the decremented 1603 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually 1604 * release any lock. 1605 */ 1606 dnode_rele(dn, db); 1607 db->db_dnode_handle = NULL; 1608 } else { 1609 DB_DNODE_EXIT(db); 1610 } 1611 1612 if (db->db_buf) 1613 dbuf_gone = arc_clear_callback(db->db_buf); 1614 1615 if (!dbuf_gone) 1616 mutex_exit(&db->db_mtx); 1617 1618 /* 1619 * If this dbuf is referenced from an indirect dbuf, 1620 * decrement the ref count on the indirect dbuf. 1621 */ 1622 if (parent && parent != dndb) 1623 dbuf_rele(parent, db); 1624} 1625 1626static int 1627dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse, 1628 dmu_buf_impl_t **parentp, blkptr_t **bpp) 1629{ 1630 int nlevels, epbs; 1631 1632 *parentp = NULL; 1633 *bpp = NULL; 1634 1635 ASSERT(blkid != DMU_BONUS_BLKID); 1636 1637 if (blkid == DMU_SPILL_BLKID) { 1638 mutex_enter(&dn->dn_mtx); 1639 if (dn->dn_have_spill && 1640 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) 1641 *bpp = &dn->dn_phys->dn_spill; 1642 else 1643 *bpp = NULL; 1644 dbuf_add_ref(dn->dn_dbuf, NULL); 1645 *parentp = dn->dn_dbuf; 1646 mutex_exit(&dn->dn_mtx); 1647 return (0); 1648 } 1649 1650 if (dn->dn_phys->dn_nlevels == 0) 1651 nlevels = 1; 1652 else 1653 nlevels = dn->dn_phys->dn_nlevels; 1654 1655 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1656 1657 ASSERT3U(level * epbs, <, 64); 1658 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1659 if (level >= nlevels || 1660 (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) { 1661 /* the buffer has no parent yet */ 1662 return (SET_ERROR(ENOENT)); 1663 } else if (level < nlevels-1) { 1664 /* this block is referenced from an indirect block */ 1665 int err = dbuf_hold_impl(dn, level+1, 1666 blkid >> epbs, fail_sparse, NULL, parentp); 1667 if (err) 1668 return (err); 1669 err = dbuf_read(*parentp, NULL, 1670 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL)); 1671 if (err) { 1672 dbuf_rele(*parentp, NULL); 1673 *parentp = NULL; 1674 return (err); 1675 } 1676 *bpp = ((blkptr_t *)(*parentp)->db.db_data) + 1677 (blkid & ((1ULL << epbs) - 1)); 1678 return (0); 1679 } else { 1680 /* the block is referenced from the dnode */ 1681 ASSERT3U(level, ==, nlevels-1); 1682 ASSERT(dn->dn_phys->dn_nblkptr == 0 || 1683 blkid < dn->dn_phys->dn_nblkptr); 1684 if (dn->dn_dbuf) { 1685 dbuf_add_ref(dn->dn_dbuf, NULL); 1686 *parentp = dn->dn_dbuf; 1687 } 1688 *bpp = &dn->dn_phys->dn_blkptr[blkid]; 1689 return (0); 1690 } 1691} 1692 1693static dmu_buf_impl_t * 1694dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid, 1695 dmu_buf_impl_t *parent, blkptr_t *blkptr) 1696{ 1697 objset_t *os = dn->dn_objset; 1698 dmu_buf_impl_t *db, *odb; 1699 1700 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1701 ASSERT(dn->dn_type != DMU_OT_NONE); 1702 1703 db = kmem_cache_alloc(dbuf_cache, KM_SLEEP); 1704 1705 db->db_objset = os; 1706 db->db.db_object = dn->dn_object; 1707 db->db_level = level; 1708 db->db_blkid = blkid; 1709 db->db_last_dirty = NULL; 1710 db->db_dirtycnt = 0; 1711 db->db_dnode_handle = dn->dn_handle; 1712 db->db_parent = parent; 1713 db->db_blkptr = blkptr; 1714 1715 db->db_user_ptr = NULL; 1716 db->db_evict_func = NULL; 1717 db->db_immediate_evict = 0; 1718 db->db_freed_in_flight = 0; 1719 1720 if (blkid == DMU_BONUS_BLKID) { 1721 ASSERT3P(parent, ==, dn->dn_dbuf); 1722 db->db.db_size = DN_MAX_BONUSLEN - 1723 (dn->dn_nblkptr-1) * sizeof (blkptr_t); 1724 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 1725 db->db.db_offset = DMU_BONUS_BLKID; 1726 db->db_state = DB_UNCACHED; 1727 /* the bonus dbuf is not placed in the hash table */ 1728 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 1729 return (db); 1730 } else if (blkid == DMU_SPILL_BLKID) { 1731 db->db.db_size = (blkptr != NULL) ? 1732 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE; 1733 db->db.db_offset = 0; 1734 } else { 1735 int blocksize = 1736 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz; 1737 db->db.db_size = blocksize; 1738 db->db.db_offset = db->db_blkid * blocksize; 1739 } 1740 1741 /* 1742 * Hold the dn_dbufs_mtx while we get the new dbuf 1743 * in the hash table *and* added to the dbufs list. 1744 * This prevents a possible deadlock with someone 1745 * trying to look up this dbuf before its added to the 1746 * dn_dbufs list. 1747 */ 1748 mutex_enter(&dn->dn_dbufs_mtx); 1749 db->db_state = DB_EVICTING; 1750 if ((odb = dbuf_hash_insert(db)) != NULL) { 1751 /* someone else inserted it first */ 1752 kmem_cache_free(dbuf_cache, db); 1753 mutex_exit(&dn->dn_dbufs_mtx); 1754 return (odb); 1755 } 1756 avl_add(&dn->dn_dbufs, db); 1757 if (db->db_level == 0 && db->db_blkid >= 1758 dn->dn_unlisted_l0_blkid) 1759 dn->dn_unlisted_l0_blkid = db->db_blkid + 1; 1760 db->db_state = DB_UNCACHED; 1761 mutex_exit(&dn->dn_dbufs_mtx); 1762 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 1763 1764 if (parent && parent != dn->dn_dbuf) 1765 dbuf_add_ref(parent, db); 1766 1767 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 1768 refcount_count(&dn->dn_holds) > 0); 1769 (void) refcount_add(&dn->dn_holds, db); 1770 atomic_inc_32(&dn->dn_dbufs_count); 1771 1772 dprintf_dbuf(db, "db=%p\n", db); 1773 1774 return (db); 1775} 1776 1777static int 1778dbuf_do_evict(void *private) 1779{ 1780 dmu_buf_impl_t *db = private; 1781 1782 if (!MUTEX_HELD(&db->db_mtx)) 1783 mutex_enter(&db->db_mtx); 1784 1785 ASSERT(refcount_is_zero(&db->db_holds)); 1786 1787 if (db->db_state != DB_EVICTING) { 1788 ASSERT(db->db_state == DB_CACHED); 1789 DBUF_VERIFY(db); 1790 db->db_buf = NULL; 1791 dbuf_evict(db); 1792 } else { 1793 mutex_exit(&db->db_mtx); 1794 dbuf_destroy(db); 1795 } 1796 return (0); 1797} 1798 1799static void 1800dbuf_destroy(dmu_buf_impl_t *db) 1801{ 1802 ASSERT(refcount_is_zero(&db->db_holds)); 1803 1804 if (db->db_blkid != DMU_BONUS_BLKID) { 1805 /* 1806 * If this dbuf is still on the dn_dbufs list, 1807 * remove it from that list. 1808 */ 1809 if (db->db_dnode_handle != NULL) { 1810 dnode_t *dn; 1811 1812 DB_DNODE_ENTER(db); 1813 dn = DB_DNODE(db); 1814 mutex_enter(&dn->dn_dbufs_mtx); 1815 avl_remove(&dn->dn_dbufs, db); 1816 atomic_dec_32(&dn->dn_dbufs_count); 1817 mutex_exit(&dn->dn_dbufs_mtx); 1818 DB_DNODE_EXIT(db); 1819 /* 1820 * Decrementing the dbuf count means that the hold 1821 * corresponding to the removed dbuf is no longer 1822 * discounted in dnode_move(), so the dnode cannot be 1823 * moved until after we release the hold. 1824 */ 1825 dnode_rele(dn, db); 1826 db->db_dnode_handle = NULL; 1827 } 1828 dbuf_hash_remove(db); 1829 } 1830 db->db_parent = NULL; 1831 db->db_buf = NULL; 1832 1833 ASSERT(db->db.db_data == NULL); 1834 ASSERT(db->db_hash_next == NULL); 1835 ASSERT(db->db_blkptr == NULL); 1836 ASSERT(db->db_data_pending == NULL); 1837 1838 kmem_cache_free(dbuf_cache, db); 1839 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 1840} 1841 1842void 1843dbuf_prefetch(dnode_t *dn, uint64_t blkid, zio_priority_t prio) 1844{ 1845 dmu_buf_impl_t *db = NULL; 1846 blkptr_t *bp = NULL; 1847 1848 ASSERT(blkid != DMU_BONUS_BLKID); 1849 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1850 1851 if (dnode_block_freed(dn, blkid)) 1852 return; 1853 1854 /* dbuf_find() returns with db_mtx held */ 1855 if (db = dbuf_find(dn, 0, blkid)) { 1856 /* 1857 * This dbuf is already in the cache. We assume that 1858 * it is already CACHED, or else about to be either 1859 * read or filled. 1860 */ 1861 mutex_exit(&db->db_mtx); 1862 return; 1863 } 1864 1865 if (dbuf_findbp(dn, 0, blkid, TRUE, &db, &bp) == 0) { 1866 if (bp && !BP_IS_HOLE(bp) && !BP_IS_EMBEDDED(bp)) { 1867 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset; 1868 arc_flags_t aflags = 1869 ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH; 1870 zbookmark_phys_t zb; 1871 1872 SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET, 1873 dn->dn_object, 0, blkid); 1874 1875 (void) arc_read(NULL, dn->dn_objset->os_spa, 1876 bp, NULL, NULL, prio, 1877 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 1878 &aflags, &zb); 1879 } 1880 if (db) 1881 dbuf_rele(db, NULL); 1882 } 1883} 1884 1885/* 1886 * Returns with db_holds incremented, and db_mtx not held. 1887 * Note: dn_struct_rwlock must be held. 1888 */ 1889int 1890dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse, 1891 void *tag, dmu_buf_impl_t **dbp) 1892{ 1893 dmu_buf_impl_t *db, *parent = NULL; 1894 1895 ASSERT(blkid != DMU_BONUS_BLKID); 1896 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1897 ASSERT3U(dn->dn_nlevels, >, level); 1898 1899 *dbp = NULL; 1900top: 1901 /* dbuf_find() returns with db_mtx held */ 1902 db = dbuf_find(dn, level, blkid); 1903 1904 if (db == NULL) { 1905 blkptr_t *bp = NULL; 1906 int err; 1907 1908 ASSERT3P(parent, ==, NULL); 1909 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp); 1910 if (fail_sparse) { 1911 if (err == 0 && bp && BP_IS_HOLE(bp)) 1912 err = SET_ERROR(ENOENT); 1913 if (err) { 1914 if (parent) 1915 dbuf_rele(parent, NULL); 1916 return (err); 1917 } 1918 } 1919 if (err && err != ENOENT) 1920 return (err); 1921 db = dbuf_create(dn, level, blkid, parent, bp); 1922 } 1923 1924 if (db->db_buf && refcount_is_zero(&db->db_holds)) { 1925 arc_buf_add_ref(db->db_buf, db); 1926 if (db->db_buf->b_data == NULL) { 1927 dbuf_clear(db); 1928 if (parent) { 1929 dbuf_rele(parent, NULL); 1930 parent = NULL; 1931 } 1932 goto top; 1933 } 1934 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data); 1935 } 1936 1937 ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf)); 1938 1939 /* 1940 * If this buffer is currently syncing out, and we are are 1941 * still referencing it from db_data, we need to make a copy 1942 * of it in case we decide we want to dirty it again in this txg. 1943 */ 1944 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 1945 dn->dn_object != DMU_META_DNODE_OBJECT && 1946 db->db_state == DB_CACHED && db->db_data_pending) { 1947 dbuf_dirty_record_t *dr = db->db_data_pending; 1948 1949 if (dr->dt.dl.dr_data == db->db_buf) { 1950 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 1951 1952 dbuf_set_data(db, 1953 arc_buf_alloc(dn->dn_objset->os_spa, 1954 db->db.db_size, db, type)); 1955 bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data, 1956 db->db.db_size); 1957 } 1958 } 1959 1960 (void) refcount_add(&db->db_holds, tag); 1961 DBUF_VERIFY(db); 1962 mutex_exit(&db->db_mtx); 1963 1964 /* NOTE: we can't rele the parent until after we drop the db_mtx */ 1965 if (parent) 1966 dbuf_rele(parent, NULL); 1967 1968 ASSERT3P(DB_DNODE(db), ==, dn); 1969 ASSERT3U(db->db_blkid, ==, blkid); 1970 ASSERT3U(db->db_level, ==, level); 1971 *dbp = db; 1972 1973 return (0); 1974} 1975 1976dmu_buf_impl_t * 1977dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag) 1978{ 1979 dmu_buf_impl_t *db; 1980 int err = dbuf_hold_impl(dn, 0, blkid, FALSE, tag, &db); 1981 return (err ? NULL : db); 1982} 1983 1984dmu_buf_impl_t * 1985dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag) 1986{ 1987 dmu_buf_impl_t *db; 1988 int err = dbuf_hold_impl(dn, level, blkid, FALSE, tag, &db); 1989 return (err ? NULL : db); 1990} 1991 1992void 1993dbuf_create_bonus(dnode_t *dn) 1994{ 1995 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 1996 1997 ASSERT(dn->dn_bonus == NULL); 1998 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL); 1999} 2000 2001int 2002dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx) 2003{ 2004 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2005 dnode_t *dn; 2006 2007 if (db->db_blkid != DMU_SPILL_BLKID) 2008 return (SET_ERROR(ENOTSUP)); 2009 if (blksz == 0) 2010 blksz = SPA_MINBLOCKSIZE; 2011 ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset))); 2012 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE); 2013 2014 DB_DNODE_ENTER(db); 2015 dn = DB_DNODE(db); 2016 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 2017 dbuf_new_size(db, blksz, tx); 2018 rw_exit(&dn->dn_struct_rwlock); 2019 DB_DNODE_EXIT(db); 2020 2021 return (0); 2022} 2023 2024void 2025dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx) 2026{ 2027 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx); 2028} 2029 2030#pragma weak dmu_buf_add_ref = dbuf_add_ref 2031void 2032dbuf_add_ref(dmu_buf_impl_t *db, void *tag) 2033{ 2034 int64_t holds = refcount_add(&db->db_holds, tag); 2035 ASSERT(holds > 1); 2036} 2037 2038/* 2039 * If you call dbuf_rele() you had better not be referencing the dnode handle 2040 * unless you have some other direct or indirect hold on the dnode. (An indirect 2041 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.) 2042 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the 2043 * dnode's parent dbuf evicting its dnode handles. 2044 */ 2045void 2046dbuf_rele(dmu_buf_impl_t *db, void *tag) 2047{ 2048 mutex_enter(&db->db_mtx); 2049 dbuf_rele_and_unlock(db, tag); 2050} 2051 2052void 2053dmu_buf_rele(dmu_buf_t *db, void *tag) 2054{ 2055 dbuf_rele((dmu_buf_impl_t *)db, tag); 2056} 2057 2058/* 2059 * dbuf_rele() for an already-locked dbuf. This is necessary to allow 2060 * db_dirtycnt and db_holds to be updated atomically. 2061 */ 2062void 2063dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag) 2064{ 2065 int64_t holds; 2066 2067 ASSERT(MUTEX_HELD(&db->db_mtx)); 2068 DBUF_VERIFY(db); 2069 2070 /* 2071 * Remove the reference to the dbuf before removing its hold on the 2072 * dnode so we can guarantee in dnode_move() that a referenced bonus 2073 * buffer has a corresponding dnode hold. 2074 */ 2075 holds = refcount_remove(&db->db_holds, tag); 2076 ASSERT(holds >= 0); 2077 2078 /* 2079 * We can't freeze indirects if there is a possibility that they 2080 * may be modified in the current syncing context. 2081 */ 2082 if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) 2083 arc_buf_freeze(db->db_buf); 2084 2085 if (holds == db->db_dirtycnt && 2086 db->db_level == 0 && db->db_immediate_evict) 2087 dbuf_evict_user(db); 2088 2089 if (holds == 0) { 2090 if (db->db_blkid == DMU_BONUS_BLKID) { 2091 mutex_exit(&db->db_mtx); 2092 2093 /* 2094 * If the dnode moves here, we cannot cross this barrier 2095 * until the move completes. 2096 */ 2097 DB_DNODE_ENTER(db); 2098 atomic_dec_32(&DB_DNODE(db)->dn_dbufs_count); 2099 DB_DNODE_EXIT(db); 2100 /* 2101 * The bonus buffer's dnode hold is no longer discounted 2102 * in dnode_move(). The dnode cannot move until after 2103 * the dnode_rele(). 2104 */ 2105 dnode_rele(DB_DNODE(db), db); 2106 } else if (db->db_buf == NULL) { 2107 /* 2108 * This is a special case: we never associated this 2109 * dbuf with any data allocated from the ARC. 2110 */ 2111 ASSERT(db->db_state == DB_UNCACHED || 2112 db->db_state == DB_NOFILL); 2113 dbuf_evict(db); 2114 } else if (arc_released(db->db_buf)) { 2115 arc_buf_t *buf = db->db_buf; 2116 /* 2117 * This dbuf has anonymous data associated with it. 2118 */ 2119 dbuf_set_data(db, NULL); 2120 VERIFY(arc_buf_remove_ref(buf, db)); 2121 dbuf_evict(db); 2122 } else { 2123 VERIFY(!arc_buf_remove_ref(db->db_buf, db)); 2124 2125 /* 2126 * A dbuf will be eligible for eviction if either the 2127 * 'primarycache' property is set or a duplicate 2128 * copy of this buffer is already cached in the arc. 2129 * 2130 * In the case of the 'primarycache' a buffer 2131 * is considered for eviction if it matches the 2132 * criteria set in the property. 2133 * 2134 * To decide if our buffer is considered a 2135 * duplicate, we must call into the arc to determine 2136 * if multiple buffers are referencing the same 2137 * block on-disk. If so, then we simply evict 2138 * ourselves. 2139 */ 2140 if (!DBUF_IS_CACHEABLE(db)) { 2141 if (db->db_blkptr != NULL && 2142 !BP_IS_HOLE(db->db_blkptr) && 2143 !BP_IS_EMBEDDED(db->db_blkptr)) { 2144 spa_t *spa = 2145 dmu_objset_spa(db->db_objset); 2146 blkptr_t bp = *db->db_blkptr; 2147 dbuf_clear(db); 2148 arc_freed(spa, &bp); 2149 } else { 2150 dbuf_clear(db); 2151 } 2152 } else if (arc_buf_eviction_needed(db->db_buf)) { 2153 dbuf_clear(db); 2154 } else { 2155 mutex_exit(&db->db_mtx); 2156 } 2157 } 2158 } else { 2159 mutex_exit(&db->db_mtx); 2160 } 2161} 2162 2163#pragma weak dmu_buf_refcount = dbuf_refcount 2164uint64_t 2165dbuf_refcount(dmu_buf_impl_t *db) 2166{ 2167 return (refcount_count(&db->db_holds)); 2168} 2169 2170void * 2171dmu_buf_set_user(dmu_buf_t *db_fake, void *user_ptr, 2172 dmu_buf_evict_func_t *evict_func) 2173{ 2174 return (dmu_buf_update_user(db_fake, NULL, user_ptr, evict_func)); 2175} 2176 2177void * 2178dmu_buf_set_user_ie(dmu_buf_t *db_fake, void *user_ptr, 2179 dmu_buf_evict_func_t *evict_func) 2180{ 2181 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2182 2183 db->db_immediate_evict = TRUE; 2184 return (dmu_buf_update_user(db_fake, NULL, user_ptr, evict_func)); 2185} 2186 2187void * 2188dmu_buf_update_user(dmu_buf_t *db_fake, void *old_user_ptr, void *user_ptr, 2189 dmu_buf_evict_func_t *evict_func) 2190{ 2191 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2192 ASSERT(db->db_level == 0); 2193 2194 ASSERT((user_ptr == NULL) == (evict_func == NULL)); 2195 2196 mutex_enter(&db->db_mtx); 2197 2198 if (db->db_user_ptr == old_user_ptr) { 2199 db->db_user_ptr = user_ptr; 2200 db->db_evict_func = evict_func; 2201 } else { 2202 old_user_ptr = db->db_user_ptr; 2203 } 2204 2205 mutex_exit(&db->db_mtx); 2206 return (old_user_ptr); 2207} 2208 2209void * 2210dmu_buf_get_user(dmu_buf_t *db_fake) 2211{ 2212 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2213 ASSERT(!refcount_is_zero(&db->db_holds)); 2214 2215 return (db->db_user_ptr); 2216} 2217 2218boolean_t 2219dmu_buf_freeable(dmu_buf_t *dbuf) 2220{ 2221 boolean_t res = B_FALSE; 2222 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; 2223 2224 if (db->db_blkptr) 2225 res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset, 2226 db->db_blkptr, db->db_blkptr->blk_birth); 2227 2228 return (res); 2229} 2230 2231blkptr_t * 2232dmu_buf_get_blkptr(dmu_buf_t *db) 2233{ 2234 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 2235 return (dbi->db_blkptr); 2236} 2237 2238static void 2239dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db) 2240{ 2241 /* ASSERT(dmu_tx_is_syncing(tx) */ 2242 ASSERT(MUTEX_HELD(&db->db_mtx)); 2243 2244 if (db->db_blkptr != NULL) 2245 return; 2246 2247 if (db->db_blkid == DMU_SPILL_BLKID) { 2248 db->db_blkptr = &dn->dn_phys->dn_spill; 2249 BP_ZERO(db->db_blkptr); 2250 return; 2251 } 2252 if (db->db_level == dn->dn_phys->dn_nlevels-1) { 2253 /* 2254 * This buffer was allocated at a time when there was 2255 * no available blkptrs from the dnode, or it was 2256 * inappropriate to hook it in (i.e., nlevels mis-match). 2257 */ 2258 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr); 2259 ASSERT(db->db_parent == NULL); 2260 db->db_parent = dn->dn_dbuf; 2261 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid]; 2262 DBUF_VERIFY(db); 2263 } else { 2264 dmu_buf_impl_t *parent = db->db_parent; 2265 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 2266 2267 ASSERT(dn->dn_phys->dn_nlevels > 1); 2268 if (parent == NULL) { 2269 mutex_exit(&db->db_mtx); 2270 rw_enter(&dn->dn_struct_rwlock, RW_READER); 2271 (void) dbuf_hold_impl(dn, db->db_level+1, 2272 db->db_blkid >> epbs, FALSE, db, &parent); 2273 rw_exit(&dn->dn_struct_rwlock); 2274 mutex_enter(&db->db_mtx); 2275 db->db_parent = parent; 2276 } 2277 db->db_blkptr = (blkptr_t *)parent->db.db_data + 2278 (db->db_blkid & ((1ULL << epbs) - 1)); 2279 DBUF_VERIFY(db); 2280 } 2281} 2282 2283static void 2284dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx) 2285{ 2286 dmu_buf_impl_t *db = dr->dr_dbuf; 2287 dnode_t *dn; 2288 zio_t *zio; 2289 2290 ASSERT(dmu_tx_is_syncing(tx)); 2291 2292 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 2293 2294 mutex_enter(&db->db_mtx); 2295 2296 ASSERT(db->db_level > 0); 2297 DBUF_VERIFY(db); 2298 2299 /* Read the block if it hasn't been read yet. */ 2300 if (db->db_buf == NULL) { 2301 mutex_exit(&db->db_mtx); 2302 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED); 2303 mutex_enter(&db->db_mtx); 2304 } 2305 ASSERT3U(db->db_state, ==, DB_CACHED); 2306 ASSERT(db->db_buf != NULL); 2307 2308 DB_DNODE_ENTER(db); 2309 dn = DB_DNODE(db); 2310 /* Indirect block size must match what the dnode thinks it is. */ 2311 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 2312 dbuf_check_blkptr(dn, db); 2313 DB_DNODE_EXIT(db); 2314 2315 /* Provide the pending dirty record to child dbufs */ 2316 db->db_data_pending = dr; 2317 2318 mutex_exit(&db->db_mtx); 2319 dbuf_write(dr, db->db_buf, tx); 2320 2321 zio = dr->dr_zio; 2322 mutex_enter(&dr->dt.di.dr_mtx); 2323 dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx); 2324 ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 2325 mutex_exit(&dr->dt.di.dr_mtx); 2326 zio_nowait(zio); 2327} 2328 2329static void 2330dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx) 2331{ 2332 arc_buf_t **datap = &dr->dt.dl.dr_data; 2333 dmu_buf_impl_t *db = dr->dr_dbuf; 2334 dnode_t *dn; 2335 objset_t *os; 2336 uint64_t txg = tx->tx_txg; 2337 2338 ASSERT(dmu_tx_is_syncing(tx)); 2339 2340 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 2341 2342 mutex_enter(&db->db_mtx); 2343 /* 2344 * To be synced, we must be dirtied. But we 2345 * might have been freed after the dirty. 2346 */ 2347 if (db->db_state == DB_UNCACHED) { 2348 /* This buffer has been freed since it was dirtied */ 2349 ASSERT(db->db.db_data == NULL); 2350 } else if (db->db_state == DB_FILL) { 2351 /* This buffer was freed and is now being re-filled */ 2352 ASSERT(db->db.db_data != dr->dt.dl.dr_data); 2353 } else { 2354 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL); 2355 } 2356 DBUF_VERIFY(db); 2357 2358 DB_DNODE_ENTER(db); 2359 dn = DB_DNODE(db); 2360 2361 if (db->db_blkid == DMU_SPILL_BLKID) { 2362 mutex_enter(&dn->dn_mtx); 2363 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR; 2364 mutex_exit(&dn->dn_mtx); 2365 } 2366 2367 /* 2368 * If this is a bonus buffer, simply copy the bonus data into the 2369 * dnode. It will be written out when the dnode is synced (and it 2370 * will be synced, since it must have been dirty for dbuf_sync to 2371 * be called). 2372 */ 2373 if (db->db_blkid == DMU_BONUS_BLKID) { 2374 dbuf_dirty_record_t **drp; 2375 2376 ASSERT(*datap != NULL); 2377 ASSERT0(db->db_level); 2378 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN); 2379 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen); 2380 DB_DNODE_EXIT(db); 2381 2382 if (*datap != db->db.db_data) { 2383 zio_buf_free(*datap, DN_MAX_BONUSLEN); 2384 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 2385 } 2386 db->db_data_pending = NULL; 2387 drp = &db->db_last_dirty; 2388 while (*drp != dr) 2389 drp = &(*drp)->dr_next; 2390 ASSERT(dr->dr_next == NULL); 2391 ASSERT(dr->dr_dbuf == db); 2392 *drp = dr->dr_next; 2393 if (dr->dr_dbuf->db_level != 0) { 2394 list_destroy(&dr->dt.di.dr_children); 2395 mutex_destroy(&dr->dt.di.dr_mtx); 2396 } 2397 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 2398 ASSERT(db->db_dirtycnt > 0); 2399 db->db_dirtycnt -= 1; 2400 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg); 2401 return; 2402 } 2403 2404 os = dn->dn_objset; 2405 2406 /* 2407 * This function may have dropped the db_mtx lock allowing a dmu_sync 2408 * operation to sneak in. As a result, we need to ensure that we 2409 * don't check the dr_override_state until we have returned from 2410 * dbuf_check_blkptr. 2411 */ 2412 dbuf_check_blkptr(dn, db); 2413 2414 /* 2415 * If this buffer is in the middle of an immediate write, 2416 * wait for the synchronous IO to complete. 2417 */ 2418 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) { 2419 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT); 2420 cv_wait(&db->db_changed, &db->db_mtx); 2421 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN); 2422 } 2423 2424 if (db->db_state != DB_NOFILL && 2425 dn->dn_object != DMU_META_DNODE_OBJECT && 2426 refcount_count(&db->db_holds) > 1 && 2427 dr->dt.dl.dr_override_state != DR_OVERRIDDEN && 2428 *datap == db->db_buf) { 2429 /* 2430 * If this buffer is currently "in use" (i.e., there 2431 * are active holds and db_data still references it), 2432 * then make a copy before we start the write so that 2433 * any modifications from the open txg will not leak 2434 * into this write. 2435 * 2436 * NOTE: this copy does not need to be made for 2437 * objects only modified in the syncing context (e.g. 2438 * DNONE_DNODE blocks). 2439 */ 2440 int blksz = arc_buf_size(*datap); 2441 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 2442 *datap = arc_buf_alloc(os->os_spa, blksz, db, type); 2443 bcopy(db->db.db_data, (*datap)->b_data, blksz); 2444 } 2445 db->db_data_pending = dr; 2446 2447 mutex_exit(&db->db_mtx); 2448 2449 dbuf_write(dr, *datap, tx); 2450 2451 ASSERT(!list_link_active(&dr->dr_dirty_node)); 2452 if (dn->dn_object == DMU_META_DNODE_OBJECT) { 2453 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr); 2454 DB_DNODE_EXIT(db); 2455 } else { 2456 /* 2457 * Although zio_nowait() does not "wait for an IO", it does 2458 * initiate the IO. If this is an empty write it seems plausible 2459 * that the IO could actually be completed before the nowait 2460 * returns. We need to DB_DNODE_EXIT() first in case 2461 * zio_nowait() invalidates the dbuf. 2462 */ 2463 DB_DNODE_EXIT(db); 2464 zio_nowait(dr->dr_zio); 2465 } 2466} 2467 2468void 2469dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx) 2470{ 2471 dbuf_dirty_record_t *dr; 2472 2473 while (dr = list_head(list)) { 2474 if (dr->dr_zio != NULL) { 2475 /* 2476 * If we find an already initialized zio then we 2477 * are processing the meta-dnode, and we have finished. 2478 * The dbufs for all dnodes are put back on the list 2479 * during processing, so that we can zio_wait() 2480 * these IOs after initiating all child IOs. 2481 */ 2482 ASSERT3U(dr->dr_dbuf->db.db_object, ==, 2483 DMU_META_DNODE_OBJECT); 2484 break; 2485 } 2486 if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID && 2487 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) { 2488 VERIFY3U(dr->dr_dbuf->db_level, ==, level); 2489 } 2490 list_remove(list, dr); 2491 if (dr->dr_dbuf->db_level > 0) 2492 dbuf_sync_indirect(dr, tx); 2493 else 2494 dbuf_sync_leaf(dr, tx); 2495 } 2496} 2497 2498/* ARGSUSED */ 2499static void 2500dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb) 2501{ 2502 dmu_buf_impl_t *db = vdb; 2503 dnode_t *dn; 2504 blkptr_t *bp = zio->io_bp; 2505 blkptr_t *bp_orig = &zio->io_bp_orig; 2506 spa_t *spa = zio->io_spa; 2507 int64_t delta; 2508 uint64_t fill = 0; 2509 int i; 2510 2511 ASSERT3P(db->db_blkptr, ==, bp); 2512 2513 DB_DNODE_ENTER(db); 2514 dn = DB_DNODE(db); 2515 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig); 2516 dnode_diduse_space(dn, delta - zio->io_prev_space_delta); 2517 zio->io_prev_space_delta = delta; 2518 2519 if (bp->blk_birth != 0) { 2520 ASSERT((db->db_blkid != DMU_SPILL_BLKID && 2521 BP_GET_TYPE(bp) == dn->dn_type) || 2522 (db->db_blkid == DMU_SPILL_BLKID && 2523 BP_GET_TYPE(bp) == dn->dn_bonustype) || 2524 BP_IS_EMBEDDED(bp)); 2525 ASSERT(BP_GET_LEVEL(bp) == db->db_level); 2526 } 2527 2528 mutex_enter(&db->db_mtx); 2529 2530#ifdef ZFS_DEBUG 2531 if (db->db_blkid == DMU_SPILL_BLKID) { 2532 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); 2533 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && 2534 db->db_blkptr == &dn->dn_phys->dn_spill); 2535 } 2536#endif 2537 2538 if (db->db_level == 0) { 2539 mutex_enter(&dn->dn_mtx); 2540 if (db->db_blkid > dn->dn_phys->dn_maxblkid && 2541 db->db_blkid != DMU_SPILL_BLKID) 2542 dn->dn_phys->dn_maxblkid = db->db_blkid; 2543 mutex_exit(&dn->dn_mtx); 2544 2545 if (dn->dn_type == DMU_OT_DNODE) { 2546 dnode_phys_t *dnp = db->db.db_data; 2547 for (i = db->db.db_size >> DNODE_SHIFT; i > 0; 2548 i--, dnp++) { 2549 if (dnp->dn_type != DMU_OT_NONE) 2550 fill++; 2551 } 2552 } else { 2553 if (BP_IS_HOLE(bp)) { 2554 fill = 0; 2555 } else { 2556 fill = 1; 2557 } 2558 } 2559 } else { 2560 blkptr_t *ibp = db->db.db_data; 2561 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 2562 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) { 2563 if (BP_IS_HOLE(ibp)) 2564 continue; 2565 fill += BP_GET_FILL(ibp); 2566 } 2567 } 2568 DB_DNODE_EXIT(db); 2569 2570 if (!BP_IS_EMBEDDED(bp)) 2571 bp->blk_fill = fill; 2572 2573 mutex_exit(&db->db_mtx); 2574} 2575 2576/* 2577 * The SPA will call this callback several times for each zio - once 2578 * for every physical child i/o (zio->io_phys_children times). This 2579 * allows the DMU to monitor the progress of each logical i/o. For example, 2580 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z 2581 * block. There may be a long delay before all copies/fragments are completed, 2582 * so this callback allows us to retire dirty space gradually, as the physical 2583 * i/os complete. 2584 */ 2585/* ARGSUSED */ 2586static void 2587dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg) 2588{ 2589 dmu_buf_impl_t *db = arg; 2590 objset_t *os = db->db_objset; 2591 dsl_pool_t *dp = dmu_objset_pool(os); 2592 dbuf_dirty_record_t *dr; 2593 int delta = 0; 2594 2595 dr = db->db_data_pending; 2596 ASSERT3U(dr->dr_txg, ==, zio->io_txg); 2597 2598 /* 2599 * The callback will be called io_phys_children times. Retire one 2600 * portion of our dirty space each time we are called. Any rounding 2601 * error will be cleaned up by dsl_pool_sync()'s call to 2602 * dsl_pool_undirty_space(). 2603 */ 2604 delta = dr->dr_accounted / zio->io_phys_children; 2605 dsl_pool_undirty_space(dp, delta, zio->io_txg); 2606} 2607 2608/* ARGSUSED */ 2609static void 2610dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb) 2611{ 2612 dmu_buf_impl_t *db = vdb; 2613 blkptr_t *bp_orig = &zio->io_bp_orig; 2614 blkptr_t *bp = db->db_blkptr; 2615 objset_t *os = db->db_objset; 2616 dmu_tx_t *tx = os->os_synctx; 2617 dbuf_dirty_record_t **drp, *dr; 2618 2619 ASSERT0(zio->io_error); 2620 ASSERT(db->db_blkptr == bp); 2621 2622 /* 2623 * For nopwrites and rewrites we ensure that the bp matches our 2624 * original and bypass all the accounting. 2625 */ 2626 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) { 2627 ASSERT(BP_EQUAL(bp, bp_orig)); 2628 } else { 2629 dsl_dataset_t *ds = os->os_dsl_dataset; 2630 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE); 2631 dsl_dataset_block_born(ds, bp, tx); 2632 } 2633 2634 mutex_enter(&db->db_mtx); 2635 2636 DBUF_VERIFY(db); 2637 2638 drp = &db->db_last_dirty; 2639 while ((dr = *drp) != db->db_data_pending) 2640 drp = &dr->dr_next; 2641 ASSERT(!list_link_active(&dr->dr_dirty_node)); 2642 ASSERT(dr->dr_dbuf == db); 2643 ASSERT(dr->dr_next == NULL); 2644 *drp = dr->dr_next; 2645 2646#ifdef ZFS_DEBUG 2647 if (db->db_blkid == DMU_SPILL_BLKID) { 2648 dnode_t *dn; 2649 2650 DB_DNODE_ENTER(db); 2651 dn = DB_DNODE(db); 2652 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); 2653 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && 2654 db->db_blkptr == &dn->dn_phys->dn_spill); 2655 DB_DNODE_EXIT(db); 2656 } 2657#endif 2658 2659 if (db->db_level == 0) { 2660 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 2661 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); 2662 if (db->db_state != DB_NOFILL) { 2663 if (dr->dt.dl.dr_data != db->db_buf) 2664 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, 2665 db)); 2666 else if (!arc_released(db->db_buf)) 2667 arc_set_callback(db->db_buf, dbuf_do_evict, db); 2668 } 2669 } else { 2670 dnode_t *dn; 2671 2672 DB_DNODE_ENTER(db); 2673 dn = DB_DNODE(db); 2674 ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 2675 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift); 2676 if (!BP_IS_HOLE(db->db_blkptr)) { 2677 int epbs = 2678 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 2679 ASSERT3U(db->db_blkid, <=, 2680 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs)); 2681 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, 2682 db->db.db_size); 2683 if (!arc_released(db->db_buf)) 2684 arc_set_callback(db->db_buf, dbuf_do_evict, db); 2685 } 2686 DB_DNODE_EXIT(db); 2687 mutex_destroy(&dr->dt.di.dr_mtx); 2688 list_destroy(&dr->dt.di.dr_children); 2689 } 2690 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 2691 2692 cv_broadcast(&db->db_changed); 2693 ASSERT(db->db_dirtycnt > 0); 2694 db->db_dirtycnt -= 1; 2695 db->db_data_pending = NULL; 2696 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg); 2697} 2698 2699static void 2700dbuf_write_nofill_ready(zio_t *zio) 2701{ 2702 dbuf_write_ready(zio, NULL, zio->io_private); 2703} 2704 2705static void 2706dbuf_write_nofill_done(zio_t *zio) 2707{ 2708 dbuf_write_done(zio, NULL, zio->io_private); 2709} 2710 2711static void 2712dbuf_write_override_ready(zio_t *zio) 2713{ 2714 dbuf_dirty_record_t *dr = zio->io_private; 2715 dmu_buf_impl_t *db = dr->dr_dbuf; 2716 2717 dbuf_write_ready(zio, NULL, db); 2718} 2719 2720static void 2721dbuf_write_override_done(zio_t *zio) 2722{ 2723 dbuf_dirty_record_t *dr = zio->io_private; 2724 dmu_buf_impl_t *db = dr->dr_dbuf; 2725 blkptr_t *obp = &dr->dt.dl.dr_overridden_by; 2726 2727 mutex_enter(&db->db_mtx); 2728 if (!BP_EQUAL(zio->io_bp, obp)) { 2729 if (!BP_IS_HOLE(obp)) 2730 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp); 2731 arc_release(dr->dt.dl.dr_data, db); 2732 } 2733 mutex_exit(&db->db_mtx); 2734 2735 dbuf_write_done(zio, NULL, db); 2736} 2737 2738/* Issue I/O to commit a dirty buffer to disk. */ 2739static void 2740dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx) 2741{ 2742 dmu_buf_impl_t *db = dr->dr_dbuf; 2743 dnode_t *dn; 2744 objset_t *os; 2745 dmu_buf_impl_t *parent = db->db_parent; 2746 uint64_t txg = tx->tx_txg; 2747 zbookmark_phys_t zb; 2748 zio_prop_t zp; 2749 zio_t *zio; 2750 int wp_flag = 0; 2751 2752 DB_DNODE_ENTER(db); 2753 dn = DB_DNODE(db); 2754 os = dn->dn_objset; 2755 2756 if (db->db_state != DB_NOFILL) { 2757 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) { 2758 /* 2759 * Private object buffers are released here rather 2760 * than in dbuf_dirty() since they are only modified 2761 * in the syncing context and we don't want the 2762 * overhead of making multiple copies of the data. 2763 */ 2764 if (BP_IS_HOLE(db->db_blkptr)) { 2765 arc_buf_thaw(data); 2766 } else { 2767 dbuf_release_bp(db); 2768 } 2769 } 2770 } 2771 2772 if (parent != dn->dn_dbuf) { 2773 /* Our parent is an indirect block. */ 2774 /* We have a dirty parent that has been scheduled for write. */ 2775 ASSERT(parent && parent->db_data_pending); 2776 /* Our parent's buffer is one level closer to the dnode. */ 2777 ASSERT(db->db_level == parent->db_level-1); 2778 /* 2779 * We're about to modify our parent's db_data by modifying 2780 * our block pointer, so the parent must be released. 2781 */ 2782 ASSERT(arc_released(parent->db_buf)); 2783 zio = parent->db_data_pending->dr_zio; 2784 } else { 2785 /* Our parent is the dnode itself. */ 2786 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 && 2787 db->db_blkid != DMU_SPILL_BLKID) || 2788 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0)); 2789 if (db->db_blkid != DMU_SPILL_BLKID) 2790 ASSERT3P(db->db_blkptr, ==, 2791 &dn->dn_phys->dn_blkptr[db->db_blkid]); 2792 zio = dn->dn_zio; 2793 } 2794 2795 ASSERT(db->db_level == 0 || data == db->db_buf); 2796 ASSERT3U(db->db_blkptr->blk_birth, <=, txg); 2797 ASSERT(zio); 2798 2799 SET_BOOKMARK(&zb, os->os_dsl_dataset ? 2800 os->os_dsl_dataset->ds_object : DMU_META_OBJSET, 2801 db->db.db_object, db->db_level, db->db_blkid); 2802 2803 if (db->db_blkid == DMU_SPILL_BLKID) 2804 wp_flag = WP_SPILL; 2805 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0; 2806 2807 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp); 2808 DB_DNODE_EXIT(db); 2809 2810 if (db->db_level == 0 && 2811 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { 2812 /* 2813 * The BP for this block has been provided by open context 2814 * (by dmu_sync() or dmu_buf_write_embedded()). 2815 */ 2816 void *contents = (data != NULL) ? data->b_data : NULL; 2817 2818 dr->dr_zio = zio_write(zio, os->os_spa, txg, 2819 db->db_blkptr, contents, db->db.db_size, &zp, 2820 dbuf_write_override_ready, NULL, dbuf_write_override_done, 2821 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); 2822 mutex_enter(&db->db_mtx); 2823 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 2824 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by, 2825 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite); 2826 mutex_exit(&db->db_mtx); 2827 } else if (db->db_state == DB_NOFILL) { 2828 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF || 2829 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY); 2830 dr->dr_zio = zio_write(zio, os->os_spa, txg, 2831 db->db_blkptr, NULL, db->db.db_size, &zp, 2832 dbuf_write_nofill_ready, NULL, dbuf_write_nofill_done, db, 2833 ZIO_PRIORITY_ASYNC_WRITE, 2834 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb); 2835 } else { 2836 ASSERT(arc_released(data)); 2837 dr->dr_zio = arc_write(zio, os->os_spa, txg, 2838 db->db_blkptr, data, DBUF_IS_L2CACHEABLE(db), 2839 DBUF_IS_L2COMPRESSIBLE(db), &zp, dbuf_write_ready, 2840 dbuf_write_physdone, dbuf_write_done, db, 2841 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); 2842 } 2843} 2844