dbuf.c revision 297112
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 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved. 28 * Copyright (c) 2014 Integros [integros.com] 29 */ 30 31#include <sys/zfs_context.h> 32#include <sys/dmu.h> 33#include <sys/dmu_send.h> 34#include <sys/dmu_impl.h> 35#include <sys/dbuf.h> 36#include <sys/dmu_objset.h> 37#include <sys/dsl_dataset.h> 38#include <sys/dsl_dir.h> 39#include <sys/dmu_tx.h> 40#include <sys/spa.h> 41#include <sys/zio.h> 42#include <sys/dmu_zfetch.h> 43#include <sys/sa.h> 44#include <sys/sa_impl.h> 45#include <sys/zfeature.h> 46#include <sys/blkptr.h> 47#include <sys/range_tree.h> 48 49/* 50 * Number of times that zfs_free_range() took the slow path while doing 51 * a zfs receive. A nonzero value indicates a potential performance problem. 52 */ 53uint64_t zfs_free_range_recv_miss; 54 55static void dbuf_destroy(dmu_buf_impl_t *db); 56static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx); 57static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx); 58 59/* 60 * Global data structures and functions for the dbuf cache. 61 */ 62static kmem_cache_t *dbuf_cache; 63static taskq_t *dbu_evict_taskq; 64 65/* ARGSUSED */ 66static int 67dbuf_cons(void *vdb, void *unused, int kmflag) 68{ 69 dmu_buf_impl_t *db = vdb; 70 bzero(db, sizeof (dmu_buf_impl_t)); 71 72 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL); 73 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL); 74 refcount_create(&db->db_holds); 75 76 return (0); 77} 78 79/* ARGSUSED */ 80static void 81dbuf_dest(void *vdb, void *unused) 82{ 83 dmu_buf_impl_t *db = vdb; 84 mutex_destroy(&db->db_mtx); 85 cv_destroy(&db->db_changed); 86 refcount_destroy(&db->db_holds); 87} 88 89/* 90 * dbuf hash table routines 91 */ 92static dbuf_hash_table_t dbuf_hash_table; 93 94static uint64_t dbuf_hash_count; 95 96static uint64_t 97dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid) 98{ 99 uintptr_t osv = (uintptr_t)os; 100 uint64_t crc = -1ULL; 101 102 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 103 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF]; 104 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF]; 105 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF]; 106 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF]; 107 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF]; 108 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF]; 109 110 crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16); 111 112 return (crc); 113} 114 115#define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid); 116 117#define DBUF_EQUAL(dbuf, os, obj, level, blkid) \ 118 ((dbuf)->db.db_object == (obj) && \ 119 (dbuf)->db_objset == (os) && \ 120 (dbuf)->db_level == (level) && \ 121 (dbuf)->db_blkid == (blkid)) 122 123dmu_buf_impl_t * 124dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid) 125{ 126 dbuf_hash_table_t *h = &dbuf_hash_table; 127 uint64_t hv = DBUF_HASH(os, obj, level, blkid); 128 uint64_t idx = hv & h->hash_table_mask; 129 dmu_buf_impl_t *db; 130 131 mutex_enter(DBUF_HASH_MUTEX(h, idx)); 132 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) { 133 if (DBUF_EQUAL(db, os, obj, level, blkid)) { 134 mutex_enter(&db->db_mtx); 135 if (db->db_state != DB_EVICTING) { 136 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 137 return (db); 138 } 139 mutex_exit(&db->db_mtx); 140 } 141 } 142 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 143 return (NULL); 144} 145 146static dmu_buf_impl_t * 147dbuf_find_bonus(objset_t *os, uint64_t object) 148{ 149 dnode_t *dn; 150 dmu_buf_impl_t *db = NULL; 151 152 if (dnode_hold(os, object, FTAG, &dn) == 0) { 153 rw_enter(&dn->dn_struct_rwlock, RW_READER); 154 if (dn->dn_bonus != NULL) { 155 db = dn->dn_bonus; 156 mutex_enter(&db->db_mtx); 157 } 158 rw_exit(&dn->dn_struct_rwlock); 159 dnode_rele(dn, FTAG); 160 } 161 return (db); 162} 163 164/* 165 * Insert an entry into the hash table. If there is already an element 166 * equal to elem in the hash table, then the already existing element 167 * will be returned and the new element will not be inserted. 168 * Otherwise returns NULL. 169 */ 170static dmu_buf_impl_t * 171dbuf_hash_insert(dmu_buf_impl_t *db) 172{ 173 dbuf_hash_table_t *h = &dbuf_hash_table; 174 objset_t *os = db->db_objset; 175 uint64_t obj = db->db.db_object; 176 int level = db->db_level; 177 uint64_t blkid = db->db_blkid; 178 uint64_t hv = DBUF_HASH(os, obj, level, blkid); 179 uint64_t idx = hv & h->hash_table_mask; 180 dmu_buf_impl_t *dbf; 181 182 mutex_enter(DBUF_HASH_MUTEX(h, idx)); 183 for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) { 184 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) { 185 mutex_enter(&dbf->db_mtx); 186 if (dbf->db_state != DB_EVICTING) { 187 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 188 return (dbf); 189 } 190 mutex_exit(&dbf->db_mtx); 191 } 192 } 193 194 mutex_enter(&db->db_mtx); 195 db->db_hash_next = h->hash_table[idx]; 196 h->hash_table[idx] = db; 197 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 198 atomic_inc_64(&dbuf_hash_count); 199 200 return (NULL); 201} 202 203/* 204 * Remove an entry from the hash table. It must be in the EVICTING state. 205 */ 206static void 207dbuf_hash_remove(dmu_buf_impl_t *db) 208{ 209 dbuf_hash_table_t *h = &dbuf_hash_table; 210 uint64_t hv = DBUF_HASH(db->db_objset, db->db.db_object, 211 db->db_level, db->db_blkid); 212 uint64_t idx = hv & h->hash_table_mask; 213 dmu_buf_impl_t *dbf, **dbp; 214 215 /* 216 * We musn't hold db_mtx to maintain lock ordering: 217 * DBUF_HASH_MUTEX > db_mtx. 218 */ 219 ASSERT(refcount_is_zero(&db->db_holds)); 220 ASSERT(db->db_state == DB_EVICTING); 221 ASSERT(!MUTEX_HELD(&db->db_mtx)); 222 223 mutex_enter(DBUF_HASH_MUTEX(h, idx)); 224 dbp = &h->hash_table[idx]; 225 while ((dbf = *dbp) != db) { 226 dbp = &dbf->db_hash_next; 227 ASSERT(dbf != NULL); 228 } 229 *dbp = db->db_hash_next; 230 db->db_hash_next = NULL; 231 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 232 atomic_dec_64(&dbuf_hash_count); 233} 234 235static arc_evict_func_t dbuf_do_evict; 236 237typedef enum { 238 DBVU_EVICTING, 239 DBVU_NOT_EVICTING 240} dbvu_verify_type_t; 241 242static void 243dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type) 244{ 245#ifdef ZFS_DEBUG 246 int64_t holds; 247 248 if (db->db_user == NULL) 249 return; 250 251 /* Only data blocks support the attachment of user data. */ 252 ASSERT(db->db_level == 0); 253 254 /* Clients must resolve a dbuf before attaching user data. */ 255 ASSERT(db->db.db_data != NULL); 256 ASSERT3U(db->db_state, ==, DB_CACHED); 257 258 holds = refcount_count(&db->db_holds); 259 if (verify_type == DBVU_EVICTING) { 260 /* 261 * Immediate eviction occurs when holds == dirtycnt. 262 * For normal eviction buffers, holds is zero on 263 * eviction, except when dbuf_fix_old_data() calls 264 * dbuf_clear_data(). However, the hold count can grow 265 * during eviction even though db_mtx is held (see 266 * dmu_bonus_hold() for an example), so we can only 267 * test the generic invariant that holds >= dirtycnt. 268 */ 269 ASSERT3U(holds, >=, db->db_dirtycnt); 270 } else { 271 if (db->db_user_immediate_evict == TRUE) 272 ASSERT3U(holds, >=, db->db_dirtycnt); 273 else 274 ASSERT3U(holds, >, 0); 275 } 276#endif 277} 278 279static void 280dbuf_evict_user(dmu_buf_impl_t *db) 281{ 282 dmu_buf_user_t *dbu = db->db_user; 283 284 ASSERT(MUTEX_HELD(&db->db_mtx)); 285 286 if (dbu == NULL) 287 return; 288 289 dbuf_verify_user(db, DBVU_EVICTING); 290 db->db_user = NULL; 291 292#ifdef ZFS_DEBUG 293 if (dbu->dbu_clear_on_evict_dbufp != NULL) 294 *dbu->dbu_clear_on_evict_dbufp = NULL; 295#endif 296 297 /* 298 * Invoke the callback from a taskq to avoid lock order reversals 299 * and limit stack depth. 300 */ 301 taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func, dbu, 0, 302 &dbu->dbu_tqent); 303} 304 305boolean_t 306dbuf_is_metadata(dmu_buf_impl_t *db) 307{ 308 if (db->db_level > 0) { 309 return (B_TRUE); 310 } else { 311 boolean_t is_metadata; 312 313 DB_DNODE_ENTER(db); 314 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type); 315 DB_DNODE_EXIT(db); 316 317 return (is_metadata); 318 } 319} 320 321void 322dbuf_evict(dmu_buf_impl_t *db) 323{ 324 ASSERT(MUTEX_HELD(&db->db_mtx)); 325 ASSERT(db->db_buf == NULL); 326 ASSERT(db->db_data_pending == NULL); 327 328 dbuf_clear(db); 329 dbuf_destroy(db); 330} 331 332void 333dbuf_init(void) 334{ 335 uint64_t hsize = 1ULL << 16; 336 dbuf_hash_table_t *h = &dbuf_hash_table; 337 int i; 338 339 /* 340 * The hash table is big enough to fill all of physical memory 341 * with an average 4K block size. The table will take up 342 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers). 343 */ 344 while (hsize * 4096 < (uint64_t)physmem * PAGESIZE) 345 hsize <<= 1; 346 347retry: 348 h->hash_table_mask = hsize - 1; 349 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP); 350 if (h->hash_table == NULL) { 351 /* XXX - we should really return an error instead of assert */ 352 ASSERT(hsize > (1ULL << 10)); 353 hsize >>= 1; 354 goto retry; 355 } 356 357 dbuf_cache = kmem_cache_create("dmu_buf_impl_t", 358 sizeof (dmu_buf_impl_t), 359 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0); 360 361 for (i = 0; i < DBUF_MUTEXES; i++) 362 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL); 363 364 /* 365 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc 366 * configuration is not required. 367 */ 368 dbu_evict_taskq = taskq_create("dbu_evict", 1, minclsyspri, 0, 0, 0); 369} 370 371void 372dbuf_fini(void) 373{ 374 dbuf_hash_table_t *h = &dbuf_hash_table; 375 int i; 376 377 for (i = 0; i < DBUF_MUTEXES; i++) 378 mutex_destroy(&h->hash_mutexes[i]); 379 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *)); 380 kmem_cache_destroy(dbuf_cache); 381 taskq_destroy(dbu_evict_taskq); 382} 383 384/* 385 * Other stuff. 386 */ 387 388#ifdef ZFS_DEBUG 389static void 390dbuf_verify(dmu_buf_impl_t *db) 391{ 392 dnode_t *dn; 393 dbuf_dirty_record_t *dr; 394 395 ASSERT(MUTEX_HELD(&db->db_mtx)); 396 397 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY)) 398 return; 399 400 ASSERT(db->db_objset != NULL); 401 DB_DNODE_ENTER(db); 402 dn = DB_DNODE(db); 403 if (dn == NULL) { 404 ASSERT(db->db_parent == NULL); 405 ASSERT(db->db_blkptr == NULL); 406 } else { 407 ASSERT3U(db->db.db_object, ==, dn->dn_object); 408 ASSERT3P(db->db_objset, ==, dn->dn_objset); 409 ASSERT3U(db->db_level, <, dn->dn_nlevels); 410 ASSERT(db->db_blkid == DMU_BONUS_BLKID || 411 db->db_blkid == DMU_SPILL_BLKID || 412 !avl_is_empty(&dn->dn_dbufs)); 413 } 414 if (db->db_blkid == DMU_BONUS_BLKID) { 415 ASSERT(dn != NULL); 416 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 417 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID); 418 } else if (db->db_blkid == DMU_SPILL_BLKID) { 419 ASSERT(dn != NULL); 420 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 421 ASSERT0(db->db.db_offset); 422 } else { 423 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size); 424 } 425 426 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next) 427 ASSERT(dr->dr_dbuf == db); 428 429 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next) 430 ASSERT(dr->dr_dbuf == db); 431 432 /* 433 * We can't assert that db_size matches dn_datablksz because it 434 * can be momentarily different when another thread is doing 435 * dnode_set_blksz(). 436 */ 437 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) { 438 dr = db->db_data_pending; 439 /* 440 * It should only be modified in syncing context, so 441 * make sure we only have one copy of the data. 442 */ 443 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf); 444 } 445 446 /* verify db->db_blkptr */ 447 if (db->db_blkptr) { 448 if (db->db_parent == dn->dn_dbuf) { 449 /* db is pointed to by the dnode */ 450 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */ 451 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object)) 452 ASSERT(db->db_parent == NULL); 453 else 454 ASSERT(db->db_parent != NULL); 455 if (db->db_blkid != DMU_SPILL_BLKID) 456 ASSERT3P(db->db_blkptr, ==, 457 &dn->dn_phys->dn_blkptr[db->db_blkid]); 458 } else { 459 /* db is pointed to by an indirect block */ 460 int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT; 461 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1); 462 ASSERT3U(db->db_parent->db.db_object, ==, 463 db->db.db_object); 464 /* 465 * dnode_grow_indblksz() can make this fail if we don't 466 * have the struct_rwlock. XXX indblksz no longer 467 * grows. safe to do this now? 468 */ 469 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) { 470 ASSERT3P(db->db_blkptr, ==, 471 ((blkptr_t *)db->db_parent->db.db_data + 472 db->db_blkid % epb)); 473 } 474 } 475 } 476 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) && 477 (db->db_buf == NULL || db->db_buf->b_data) && 478 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID && 479 db->db_state != DB_FILL && !dn->dn_free_txg) { 480 /* 481 * If the blkptr isn't set but they have nonzero data, 482 * it had better be dirty, otherwise we'll lose that 483 * data when we evict this buffer. 484 */ 485 if (db->db_dirtycnt == 0) { 486 uint64_t *buf = db->db.db_data; 487 int i; 488 489 for (i = 0; i < db->db.db_size >> 3; i++) { 490 ASSERT(buf[i] == 0); 491 } 492 } 493 } 494 DB_DNODE_EXIT(db); 495} 496#endif 497 498static void 499dbuf_clear_data(dmu_buf_impl_t *db) 500{ 501 ASSERT(MUTEX_HELD(&db->db_mtx)); 502 dbuf_evict_user(db); 503 db->db_buf = NULL; 504 db->db.db_data = NULL; 505 if (db->db_state != DB_NOFILL) 506 db->db_state = DB_UNCACHED; 507} 508 509static void 510dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf) 511{ 512 ASSERT(MUTEX_HELD(&db->db_mtx)); 513 ASSERT(buf != NULL); 514 515 db->db_buf = buf; 516 ASSERT(buf->b_data != NULL); 517 db->db.db_data = buf->b_data; 518 if (!arc_released(buf)) 519 arc_set_callback(buf, dbuf_do_evict, db); 520} 521 522/* 523 * Loan out an arc_buf for read. Return the loaned arc_buf. 524 */ 525arc_buf_t * 526dbuf_loan_arcbuf(dmu_buf_impl_t *db) 527{ 528 arc_buf_t *abuf; 529 530 mutex_enter(&db->db_mtx); 531 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) { 532 int blksz = db->db.db_size; 533 spa_t *spa = db->db_objset->os_spa; 534 535 mutex_exit(&db->db_mtx); 536 abuf = arc_loan_buf(spa, blksz); 537 bcopy(db->db.db_data, abuf->b_data, blksz); 538 } else { 539 abuf = db->db_buf; 540 arc_loan_inuse_buf(abuf, db); 541 dbuf_clear_data(db); 542 mutex_exit(&db->db_mtx); 543 } 544 return (abuf); 545} 546 547/* 548 * Calculate which level n block references the data at the level 0 offset 549 * provided. 550 */ 551uint64_t 552dbuf_whichblock(dnode_t *dn, int64_t level, uint64_t offset) 553{ 554 if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) { 555 /* 556 * The level n blkid is equal to the level 0 blkid divided by 557 * the number of level 0s in a level n block. 558 * 559 * The level 0 blkid is offset >> datablkshift = 560 * offset / 2^datablkshift. 561 * 562 * The number of level 0s in a level n is the number of block 563 * pointers in an indirect block, raised to the power of level. 564 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level = 565 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)). 566 * 567 * Thus, the level n blkid is: offset / 568 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT))) 569 * = offset / 2^(datablkshift + level * 570 * (indblkshift - SPA_BLKPTRSHIFT)) 571 * = offset >> (datablkshift + level * 572 * (indblkshift - SPA_BLKPTRSHIFT)) 573 */ 574 return (offset >> (dn->dn_datablkshift + level * 575 (dn->dn_indblkshift - SPA_BLKPTRSHIFT))); 576 } else { 577 ASSERT3U(offset, <, dn->dn_datablksz); 578 return (0); 579 } 580} 581 582static void 583dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb) 584{ 585 dmu_buf_impl_t *db = vdb; 586 587 mutex_enter(&db->db_mtx); 588 ASSERT3U(db->db_state, ==, DB_READ); 589 /* 590 * All reads are synchronous, so we must have a hold on the dbuf 591 */ 592 ASSERT(refcount_count(&db->db_holds) > 0); 593 ASSERT(db->db_buf == NULL); 594 ASSERT(db->db.db_data == NULL); 595 if (db->db_level == 0 && db->db_freed_in_flight) { 596 /* we were freed in flight; disregard any error */ 597 arc_release(buf, db); 598 bzero(buf->b_data, db->db.db_size); 599 arc_buf_freeze(buf); 600 db->db_freed_in_flight = FALSE; 601 dbuf_set_data(db, buf); 602 db->db_state = DB_CACHED; 603 } else if (zio == NULL || zio->io_error == 0) { 604 dbuf_set_data(db, buf); 605 db->db_state = DB_CACHED; 606 } else { 607 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 608 ASSERT3P(db->db_buf, ==, NULL); 609 VERIFY(arc_buf_remove_ref(buf, db)); 610 db->db_state = DB_UNCACHED; 611 } 612 cv_broadcast(&db->db_changed); 613 dbuf_rele_and_unlock(db, NULL); 614} 615 616static void 617dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) 618{ 619 dnode_t *dn; 620 zbookmark_phys_t zb; 621 arc_flags_t aflags = ARC_FLAG_NOWAIT; 622 623 DB_DNODE_ENTER(db); 624 dn = DB_DNODE(db); 625 ASSERT(!refcount_is_zero(&db->db_holds)); 626 /* We need the struct_rwlock to prevent db_blkptr from changing. */ 627 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 628 ASSERT(MUTEX_HELD(&db->db_mtx)); 629 ASSERT(db->db_state == DB_UNCACHED); 630 ASSERT(db->db_buf == NULL); 631 632 if (db->db_blkid == DMU_BONUS_BLKID) { 633 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen); 634 635 ASSERT3U(bonuslen, <=, db->db.db_size); 636 db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN); 637 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 638 if (bonuslen < DN_MAX_BONUSLEN) 639 bzero(db->db.db_data, DN_MAX_BONUSLEN); 640 if (bonuslen) 641 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen); 642 DB_DNODE_EXIT(db); 643 db->db_state = DB_CACHED; 644 mutex_exit(&db->db_mtx); 645 return; 646 } 647 648 /* 649 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync() 650 * processes the delete record and clears the bp while we are waiting 651 * for the dn_mtx (resulting in a "no" from block_freed). 652 */ 653 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) || 654 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) || 655 BP_IS_HOLE(db->db_blkptr)))) { 656 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 657 658 DB_DNODE_EXIT(db); 659 dbuf_set_data(db, arc_buf_alloc(db->db_objset->os_spa, 660 db->db.db_size, db, type)); 661 bzero(db->db.db_data, db->db.db_size); 662 db->db_state = DB_CACHED; 663 mutex_exit(&db->db_mtx); 664 return; 665 } 666 667 DB_DNODE_EXIT(db); 668 669 db->db_state = DB_READ; 670 mutex_exit(&db->db_mtx); 671 672 if (DBUF_IS_L2CACHEABLE(db)) 673 aflags |= ARC_FLAG_L2CACHE; 674 if (DBUF_IS_L2COMPRESSIBLE(db)) 675 aflags |= ARC_FLAG_L2COMPRESS; 676 677 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ? 678 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET, 679 db->db.db_object, db->db_level, db->db_blkid); 680 681 dbuf_add_ref(db, NULL); 682 683 (void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr, 684 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ, 685 (flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED, 686 &aflags, &zb); 687} 688 689int 690dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) 691{ 692 int err = 0; 693 boolean_t havepzio = (zio != NULL); 694 boolean_t prefetch; 695 dnode_t *dn; 696 697 /* 698 * We don't have to hold the mutex to check db_state because it 699 * can't be freed while we have a hold on the buffer. 700 */ 701 ASSERT(!refcount_is_zero(&db->db_holds)); 702 703 if (db->db_state == DB_NOFILL) 704 return (SET_ERROR(EIO)); 705 706 DB_DNODE_ENTER(db); 707 dn = DB_DNODE(db); 708 if ((flags & DB_RF_HAVESTRUCT) == 0) 709 rw_enter(&dn->dn_struct_rwlock, RW_READER); 710 711 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 712 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL && 713 DBUF_IS_CACHEABLE(db); 714 715 mutex_enter(&db->db_mtx); 716 if (db->db_state == DB_CACHED) { 717 mutex_exit(&db->db_mtx); 718 if (prefetch) 719 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1); 720 if ((flags & DB_RF_HAVESTRUCT) == 0) 721 rw_exit(&dn->dn_struct_rwlock); 722 DB_DNODE_EXIT(db); 723 } else if (db->db_state == DB_UNCACHED) { 724 spa_t *spa = dn->dn_objset->os_spa; 725 726 if (zio == NULL) 727 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); 728 dbuf_read_impl(db, zio, flags); 729 730 /* dbuf_read_impl has dropped db_mtx for us */ 731 732 if (prefetch) 733 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1); 734 735 if ((flags & DB_RF_HAVESTRUCT) == 0) 736 rw_exit(&dn->dn_struct_rwlock); 737 DB_DNODE_EXIT(db); 738 739 if (!havepzio) 740 err = zio_wait(zio); 741 } else { 742 /* 743 * Another reader came in while the dbuf was in flight 744 * between UNCACHED and CACHED. Either a writer will finish 745 * writing the buffer (sending the dbuf to CACHED) or the 746 * first reader's request will reach the read_done callback 747 * and send the dbuf to CACHED. Otherwise, a failure 748 * occurred and the dbuf went to UNCACHED. 749 */ 750 mutex_exit(&db->db_mtx); 751 if (prefetch) 752 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1); 753 if ((flags & DB_RF_HAVESTRUCT) == 0) 754 rw_exit(&dn->dn_struct_rwlock); 755 DB_DNODE_EXIT(db); 756 757 /* Skip the wait per the caller's request. */ 758 mutex_enter(&db->db_mtx); 759 if ((flags & DB_RF_NEVERWAIT) == 0) { 760 while (db->db_state == DB_READ || 761 db->db_state == DB_FILL) { 762 ASSERT(db->db_state == DB_READ || 763 (flags & DB_RF_HAVESTRUCT) == 0); 764 DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *, 765 db, zio_t *, zio); 766 cv_wait(&db->db_changed, &db->db_mtx); 767 } 768 if (db->db_state == DB_UNCACHED) 769 err = SET_ERROR(EIO); 770 } 771 mutex_exit(&db->db_mtx); 772 } 773 774 ASSERT(err || havepzio || db->db_state == DB_CACHED); 775 return (err); 776} 777 778static void 779dbuf_noread(dmu_buf_impl_t *db) 780{ 781 ASSERT(!refcount_is_zero(&db->db_holds)); 782 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 783 mutex_enter(&db->db_mtx); 784 while (db->db_state == DB_READ || db->db_state == DB_FILL) 785 cv_wait(&db->db_changed, &db->db_mtx); 786 if (db->db_state == DB_UNCACHED) { 787 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 788 spa_t *spa = db->db_objset->os_spa; 789 790 ASSERT(db->db_buf == NULL); 791 ASSERT(db->db.db_data == NULL); 792 dbuf_set_data(db, arc_buf_alloc(spa, db->db.db_size, db, type)); 793 db->db_state = DB_FILL; 794 } else if (db->db_state == DB_NOFILL) { 795 dbuf_clear_data(db); 796 } else { 797 ASSERT3U(db->db_state, ==, DB_CACHED); 798 } 799 mutex_exit(&db->db_mtx); 800} 801 802/* 803 * This is our just-in-time copy function. It makes a copy of 804 * buffers, that have been modified in a previous transaction 805 * group, before we modify them in the current active group. 806 * 807 * This function is used in two places: when we are dirtying a 808 * buffer for the first time in a txg, and when we are freeing 809 * a range in a dnode that includes this buffer. 810 * 811 * Note that when we are called from dbuf_free_range() we do 812 * not put a hold on the buffer, we just traverse the active 813 * dbuf list for the dnode. 814 */ 815static void 816dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg) 817{ 818 dbuf_dirty_record_t *dr = db->db_last_dirty; 819 820 ASSERT(MUTEX_HELD(&db->db_mtx)); 821 ASSERT(db->db.db_data != NULL); 822 ASSERT(db->db_level == 0); 823 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT); 824 825 if (dr == NULL || 826 (dr->dt.dl.dr_data != 827 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf))) 828 return; 829 830 /* 831 * If the last dirty record for this dbuf has not yet synced 832 * and its referencing the dbuf data, either: 833 * reset the reference to point to a new copy, 834 * or (if there a no active holders) 835 * just null out the current db_data pointer. 836 */ 837 ASSERT(dr->dr_txg >= txg - 2); 838 if (db->db_blkid == DMU_BONUS_BLKID) { 839 /* Note that the data bufs here are zio_bufs */ 840 dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN); 841 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 842 bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN); 843 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) { 844 int size = db->db.db_size; 845 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 846 spa_t *spa = db->db_objset->os_spa; 847 848 dr->dt.dl.dr_data = arc_buf_alloc(spa, size, db, type); 849 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size); 850 } else { 851 dbuf_clear_data(db); 852 } 853} 854 855void 856dbuf_unoverride(dbuf_dirty_record_t *dr) 857{ 858 dmu_buf_impl_t *db = dr->dr_dbuf; 859 blkptr_t *bp = &dr->dt.dl.dr_overridden_by; 860 uint64_t txg = dr->dr_txg; 861 862 ASSERT(MUTEX_HELD(&db->db_mtx)); 863 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC); 864 ASSERT(db->db_level == 0); 865 866 if (db->db_blkid == DMU_BONUS_BLKID || 867 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN) 868 return; 869 870 ASSERT(db->db_data_pending != dr); 871 872 /* free this block */ 873 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite) 874 zio_free(db->db_objset->os_spa, txg, bp); 875 876 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 877 dr->dt.dl.dr_nopwrite = B_FALSE; 878 879 /* 880 * Release the already-written buffer, so we leave it in 881 * a consistent dirty state. Note that all callers are 882 * modifying the buffer, so they will immediately do 883 * another (redundant) arc_release(). Therefore, leave 884 * the buf thawed to save the effort of freezing & 885 * immediately re-thawing it. 886 */ 887 arc_release(dr->dt.dl.dr_data, db); 888} 889 890/* 891 * Evict (if its unreferenced) or clear (if its referenced) any level-0 892 * data blocks in the free range, so that any future readers will find 893 * empty blocks. 894 * 895 * This is a no-op if the dataset is in the middle of an incremental 896 * receive; see comment below for details. 897 */ 898void 899dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid, 900 dmu_tx_t *tx) 901{ 902 dmu_buf_impl_t db_search; 903 dmu_buf_impl_t *db, *db_next; 904 uint64_t txg = tx->tx_txg; 905 avl_index_t where; 906 907 if (end_blkid > dn->dn_maxblkid && (end_blkid != DMU_SPILL_BLKID)) 908 end_blkid = dn->dn_maxblkid; 909 dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid); 910 911 db_search.db_level = 0; 912 db_search.db_blkid = start_blkid; 913 db_search.db_state = DB_SEARCH; 914 915 mutex_enter(&dn->dn_dbufs_mtx); 916 if (start_blkid >= dn->dn_unlisted_l0_blkid) { 917 /* There can't be any dbufs in this range; no need to search. */ 918#ifdef DEBUG 919 db = avl_find(&dn->dn_dbufs, &db_search, &where); 920 ASSERT3P(db, ==, NULL); 921 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER); 922 ASSERT(db == NULL || db->db_level > 0); 923#endif 924 mutex_exit(&dn->dn_dbufs_mtx); 925 return; 926 } else if (dmu_objset_is_receiving(dn->dn_objset)) { 927 /* 928 * If we are receiving, we expect there to be no dbufs in 929 * the range to be freed, because receive modifies each 930 * block at most once, and in offset order. If this is 931 * not the case, it can lead to performance problems, 932 * so note that we unexpectedly took the slow path. 933 */ 934 atomic_inc_64(&zfs_free_range_recv_miss); 935 } 936 937 db = avl_find(&dn->dn_dbufs, &db_search, &where); 938 ASSERT3P(db, ==, NULL); 939 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER); 940 941 for (; db != NULL; db = db_next) { 942 db_next = AVL_NEXT(&dn->dn_dbufs, db); 943 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 944 945 if (db->db_level != 0 || db->db_blkid > end_blkid) { 946 break; 947 } 948 ASSERT3U(db->db_blkid, >=, start_blkid); 949 950 /* found a level 0 buffer in the range */ 951 mutex_enter(&db->db_mtx); 952 if (dbuf_undirty(db, tx)) { 953 /* mutex has been dropped and dbuf destroyed */ 954 continue; 955 } 956 957 if (db->db_state == DB_UNCACHED || 958 db->db_state == DB_NOFILL || 959 db->db_state == DB_EVICTING) { 960 ASSERT(db->db.db_data == NULL); 961 mutex_exit(&db->db_mtx); 962 continue; 963 } 964 if (db->db_state == DB_READ || db->db_state == DB_FILL) { 965 /* will be handled in dbuf_read_done or dbuf_rele */ 966 db->db_freed_in_flight = TRUE; 967 mutex_exit(&db->db_mtx); 968 continue; 969 } 970 if (refcount_count(&db->db_holds) == 0) { 971 ASSERT(db->db_buf); 972 dbuf_clear(db); 973 continue; 974 } 975 /* The dbuf is referenced */ 976 977 if (db->db_last_dirty != NULL) { 978 dbuf_dirty_record_t *dr = db->db_last_dirty; 979 980 if (dr->dr_txg == txg) { 981 /* 982 * This buffer is "in-use", re-adjust the file 983 * size to reflect that this buffer may 984 * contain new data when we sync. 985 */ 986 if (db->db_blkid != DMU_SPILL_BLKID && 987 db->db_blkid > dn->dn_maxblkid) 988 dn->dn_maxblkid = db->db_blkid; 989 dbuf_unoverride(dr); 990 } else { 991 /* 992 * This dbuf is not dirty in the open context. 993 * Either uncache it (if its not referenced in 994 * the open context) or reset its contents to 995 * empty. 996 */ 997 dbuf_fix_old_data(db, txg); 998 } 999 } 1000 /* clear the contents if its cached */ 1001 if (db->db_state == DB_CACHED) { 1002 ASSERT(db->db.db_data != NULL); 1003 arc_release(db->db_buf, db); 1004 bzero(db->db.db_data, db->db.db_size); 1005 arc_buf_freeze(db->db_buf); 1006 } 1007 1008 mutex_exit(&db->db_mtx); 1009 } 1010 mutex_exit(&dn->dn_dbufs_mtx); 1011} 1012 1013static int 1014dbuf_block_freeable(dmu_buf_impl_t *db) 1015{ 1016 dsl_dataset_t *ds = db->db_objset->os_dsl_dataset; 1017 uint64_t birth_txg = 0; 1018 1019 /* 1020 * We don't need any locking to protect db_blkptr: 1021 * If it's syncing, then db_last_dirty will be set 1022 * so we'll ignore db_blkptr. 1023 * 1024 * This logic ensures that only block births for 1025 * filled blocks are considered. 1026 */ 1027 ASSERT(MUTEX_HELD(&db->db_mtx)); 1028 if (db->db_last_dirty && (db->db_blkptr == NULL || 1029 !BP_IS_HOLE(db->db_blkptr))) { 1030 birth_txg = db->db_last_dirty->dr_txg; 1031 } else if (db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) { 1032 birth_txg = db->db_blkptr->blk_birth; 1033 } 1034 1035 /* 1036 * If this block don't exist or is in a snapshot, it can't be freed. 1037 * Don't pass the bp to dsl_dataset_block_freeable() since we 1038 * are holding the db_mtx lock and might deadlock if we are 1039 * prefetching a dedup-ed block. 1040 */ 1041 if (birth_txg != 0) 1042 return (ds == NULL || 1043 dsl_dataset_block_freeable(ds, NULL, birth_txg)); 1044 else 1045 return (B_FALSE); 1046} 1047 1048void 1049dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx) 1050{ 1051 arc_buf_t *buf, *obuf; 1052 int osize = db->db.db_size; 1053 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 1054 dnode_t *dn; 1055 1056 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1057 1058 DB_DNODE_ENTER(db); 1059 dn = DB_DNODE(db); 1060 1061 /* XXX does *this* func really need the lock? */ 1062 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 1063 1064 /* 1065 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held 1066 * is OK, because there can be no other references to the db 1067 * when we are changing its size, so no concurrent DB_FILL can 1068 * be happening. 1069 */ 1070 /* 1071 * XXX we should be doing a dbuf_read, checking the return 1072 * value and returning that up to our callers 1073 */ 1074 dmu_buf_will_dirty(&db->db, tx); 1075 1076 /* create the data buffer for the new block */ 1077 buf = arc_buf_alloc(dn->dn_objset->os_spa, size, db, type); 1078 1079 /* copy old block data to the new block */ 1080 obuf = db->db_buf; 1081 bcopy(obuf->b_data, buf->b_data, MIN(osize, size)); 1082 /* zero the remainder */ 1083 if (size > osize) 1084 bzero((uint8_t *)buf->b_data + osize, size - osize); 1085 1086 mutex_enter(&db->db_mtx); 1087 dbuf_set_data(db, buf); 1088 VERIFY(arc_buf_remove_ref(obuf, db)); 1089 db->db.db_size = size; 1090 1091 if (db->db_level == 0) { 1092 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); 1093 db->db_last_dirty->dt.dl.dr_data = buf; 1094 } 1095 mutex_exit(&db->db_mtx); 1096 1097 dnode_willuse_space(dn, size-osize, tx); 1098 DB_DNODE_EXIT(db); 1099} 1100 1101void 1102dbuf_release_bp(dmu_buf_impl_t *db) 1103{ 1104 objset_t *os = db->db_objset; 1105 1106 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os))); 1107 ASSERT(arc_released(os->os_phys_buf) || 1108 list_link_active(&os->os_dsl_dataset->ds_synced_link)); 1109 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf)); 1110 1111 (void) arc_release(db->db_buf, db); 1112} 1113 1114/* 1115 * We already have a dirty record for this TXG, and we are being 1116 * dirtied again. 1117 */ 1118static void 1119dbuf_redirty(dbuf_dirty_record_t *dr) 1120{ 1121 dmu_buf_impl_t *db = dr->dr_dbuf; 1122 1123 ASSERT(MUTEX_HELD(&db->db_mtx)); 1124 1125 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) { 1126 /* 1127 * If this buffer has already been written out, 1128 * we now need to reset its state. 1129 */ 1130 dbuf_unoverride(dr); 1131 if (db->db.db_object != DMU_META_DNODE_OBJECT && 1132 db->db_state != DB_NOFILL) { 1133 /* Already released on initial dirty, so just thaw. */ 1134 ASSERT(arc_released(db->db_buf)); 1135 arc_buf_thaw(db->db_buf); 1136 } 1137 } 1138} 1139 1140dbuf_dirty_record_t * 1141dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1142{ 1143 dnode_t *dn; 1144 objset_t *os; 1145 dbuf_dirty_record_t **drp, *dr; 1146 int drop_struct_lock = FALSE; 1147 boolean_t do_free_accounting = B_FALSE; 1148 int txgoff = tx->tx_txg & TXG_MASK; 1149 1150 ASSERT(tx->tx_txg != 0); 1151 ASSERT(!refcount_is_zero(&db->db_holds)); 1152 DMU_TX_DIRTY_BUF(tx, db); 1153 1154 DB_DNODE_ENTER(db); 1155 dn = DB_DNODE(db); 1156 /* 1157 * Shouldn't dirty a regular buffer in syncing context. Private 1158 * objects may be dirtied in syncing context, but only if they 1159 * were already pre-dirtied in open context. 1160 */ 1161 ASSERT(!dmu_tx_is_syncing(tx) || 1162 BP_IS_HOLE(dn->dn_objset->os_rootbp) || 1163 DMU_OBJECT_IS_SPECIAL(dn->dn_object) || 1164 dn->dn_objset->os_dsl_dataset == NULL); 1165 /* 1166 * We make this assert for private objects as well, but after we 1167 * check if we're already dirty. They are allowed to re-dirty 1168 * in syncing context. 1169 */ 1170 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 1171 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 1172 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 1173 1174 mutex_enter(&db->db_mtx); 1175 /* 1176 * XXX make this true for indirects too? The problem is that 1177 * transactions created with dmu_tx_create_assigned() from 1178 * syncing context don't bother holding ahead. 1179 */ 1180 ASSERT(db->db_level != 0 || 1181 db->db_state == DB_CACHED || db->db_state == DB_FILL || 1182 db->db_state == DB_NOFILL); 1183 1184 mutex_enter(&dn->dn_mtx); 1185 /* 1186 * Don't set dirtyctx to SYNC if we're just modifying this as we 1187 * initialize the objset. 1188 */ 1189 if (dn->dn_dirtyctx == DN_UNDIRTIED && 1190 !BP_IS_HOLE(dn->dn_objset->os_rootbp)) { 1191 dn->dn_dirtyctx = 1192 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN); 1193 ASSERT(dn->dn_dirtyctx_firstset == NULL); 1194 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP); 1195 } 1196 mutex_exit(&dn->dn_mtx); 1197 1198 if (db->db_blkid == DMU_SPILL_BLKID) 1199 dn->dn_have_spill = B_TRUE; 1200 1201 /* 1202 * If this buffer is already dirty, we're done. 1203 */ 1204 drp = &db->db_last_dirty; 1205 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg || 1206 db->db.db_object == DMU_META_DNODE_OBJECT); 1207 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg) 1208 drp = &dr->dr_next; 1209 if (dr && dr->dr_txg == tx->tx_txg) { 1210 DB_DNODE_EXIT(db); 1211 1212 dbuf_redirty(dr); 1213 mutex_exit(&db->db_mtx); 1214 return (dr); 1215 } 1216 1217 /* 1218 * Only valid if not already dirty. 1219 */ 1220 ASSERT(dn->dn_object == 0 || 1221 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 1222 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 1223 1224 ASSERT3U(dn->dn_nlevels, >, db->db_level); 1225 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) || 1226 dn->dn_phys->dn_nlevels > db->db_level || 1227 dn->dn_next_nlevels[txgoff] > db->db_level || 1228 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level || 1229 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level); 1230 1231 /* 1232 * We should only be dirtying in syncing context if it's the 1233 * mos or we're initializing the os or it's a special object. 1234 * However, we are allowed to dirty in syncing context provided 1235 * we already dirtied it in open context. Hence we must make 1236 * this assertion only if we're not already dirty. 1237 */ 1238 os = dn->dn_objset; 1239 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) || 1240 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp)); 1241 ASSERT(db->db.db_size != 0); 1242 1243 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 1244 1245 if (db->db_blkid != DMU_BONUS_BLKID) { 1246 /* 1247 * Update the accounting. 1248 * Note: we delay "free accounting" until after we drop 1249 * the db_mtx. This keeps us from grabbing other locks 1250 * (and possibly deadlocking) in bp_get_dsize() while 1251 * also holding the db_mtx. 1252 */ 1253 dnode_willuse_space(dn, db->db.db_size, tx); 1254 do_free_accounting = dbuf_block_freeable(db); 1255 } 1256 1257 /* 1258 * If this buffer is dirty in an old transaction group we need 1259 * to make a copy of it so that the changes we make in this 1260 * transaction group won't leak out when we sync the older txg. 1261 */ 1262 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP); 1263 if (db->db_level == 0) { 1264 void *data_old = db->db_buf; 1265 1266 if (db->db_state != DB_NOFILL) { 1267 if (db->db_blkid == DMU_BONUS_BLKID) { 1268 dbuf_fix_old_data(db, tx->tx_txg); 1269 data_old = db->db.db_data; 1270 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) { 1271 /* 1272 * Release the data buffer from the cache so 1273 * that we can modify it without impacting 1274 * possible other users of this cached data 1275 * block. Note that indirect blocks and 1276 * private objects are not released until the 1277 * syncing state (since they are only modified 1278 * then). 1279 */ 1280 arc_release(db->db_buf, db); 1281 dbuf_fix_old_data(db, tx->tx_txg); 1282 data_old = db->db_buf; 1283 } 1284 ASSERT(data_old != NULL); 1285 } 1286 dr->dt.dl.dr_data = data_old; 1287 } else { 1288 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL); 1289 list_create(&dr->dt.di.dr_children, 1290 sizeof (dbuf_dirty_record_t), 1291 offsetof(dbuf_dirty_record_t, dr_dirty_node)); 1292 } 1293 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL) 1294 dr->dr_accounted = db->db.db_size; 1295 dr->dr_dbuf = db; 1296 dr->dr_txg = tx->tx_txg; 1297 dr->dr_next = *drp; 1298 *drp = dr; 1299 1300 /* 1301 * We could have been freed_in_flight between the dbuf_noread 1302 * and dbuf_dirty. We win, as though the dbuf_noread() had 1303 * happened after the free. 1304 */ 1305 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 1306 db->db_blkid != DMU_SPILL_BLKID) { 1307 mutex_enter(&dn->dn_mtx); 1308 if (dn->dn_free_ranges[txgoff] != NULL) { 1309 range_tree_clear(dn->dn_free_ranges[txgoff], 1310 db->db_blkid, 1); 1311 } 1312 mutex_exit(&dn->dn_mtx); 1313 db->db_freed_in_flight = FALSE; 1314 } 1315 1316 /* 1317 * This buffer is now part of this txg 1318 */ 1319 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg); 1320 db->db_dirtycnt += 1; 1321 ASSERT3U(db->db_dirtycnt, <=, 3); 1322 1323 mutex_exit(&db->db_mtx); 1324 1325 if (db->db_blkid == DMU_BONUS_BLKID || 1326 db->db_blkid == DMU_SPILL_BLKID) { 1327 mutex_enter(&dn->dn_mtx); 1328 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1329 list_insert_tail(&dn->dn_dirty_records[txgoff], dr); 1330 mutex_exit(&dn->dn_mtx); 1331 dnode_setdirty(dn, tx); 1332 DB_DNODE_EXIT(db); 1333 return (dr); 1334 } else if (do_free_accounting) { 1335 blkptr_t *bp = db->db_blkptr; 1336 int64_t willfree = (bp && !BP_IS_HOLE(bp)) ? 1337 bp_get_dsize(os->os_spa, bp) : db->db.db_size; 1338 /* 1339 * This is only a guess -- if the dbuf is dirty 1340 * in a previous txg, we don't know how much 1341 * space it will use on disk yet. We should 1342 * really have the struct_rwlock to access 1343 * db_blkptr, but since this is just a guess, 1344 * it's OK if we get an odd answer. 1345 */ 1346 ddt_prefetch(os->os_spa, bp); 1347 dnode_willuse_space(dn, -willfree, tx); 1348 } 1349 1350 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) { 1351 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1352 drop_struct_lock = TRUE; 1353 } 1354 1355 if (db->db_level == 0) { 1356 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock); 1357 ASSERT(dn->dn_maxblkid >= db->db_blkid); 1358 } 1359 1360 if (db->db_level+1 < dn->dn_nlevels) { 1361 dmu_buf_impl_t *parent = db->db_parent; 1362 dbuf_dirty_record_t *di; 1363 int parent_held = FALSE; 1364 1365 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) { 1366 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1367 1368 parent = dbuf_hold_level(dn, db->db_level+1, 1369 db->db_blkid >> epbs, FTAG); 1370 ASSERT(parent != NULL); 1371 parent_held = TRUE; 1372 } 1373 if (drop_struct_lock) 1374 rw_exit(&dn->dn_struct_rwlock); 1375 ASSERT3U(db->db_level+1, ==, parent->db_level); 1376 di = dbuf_dirty(parent, tx); 1377 if (parent_held) 1378 dbuf_rele(parent, FTAG); 1379 1380 mutex_enter(&db->db_mtx); 1381 /* 1382 * Since we've dropped the mutex, it's possible that 1383 * dbuf_undirty() might have changed this out from under us. 1384 */ 1385 if (db->db_last_dirty == dr || 1386 dn->dn_object == DMU_META_DNODE_OBJECT) { 1387 mutex_enter(&di->dt.di.dr_mtx); 1388 ASSERT3U(di->dr_txg, ==, tx->tx_txg); 1389 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1390 list_insert_tail(&di->dt.di.dr_children, dr); 1391 mutex_exit(&di->dt.di.dr_mtx); 1392 dr->dr_parent = di; 1393 } 1394 mutex_exit(&db->db_mtx); 1395 } else { 1396 ASSERT(db->db_level+1 == dn->dn_nlevels); 1397 ASSERT(db->db_blkid < dn->dn_nblkptr); 1398 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf); 1399 mutex_enter(&dn->dn_mtx); 1400 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1401 list_insert_tail(&dn->dn_dirty_records[txgoff], dr); 1402 mutex_exit(&dn->dn_mtx); 1403 if (drop_struct_lock) 1404 rw_exit(&dn->dn_struct_rwlock); 1405 } 1406 1407 dnode_setdirty(dn, tx); 1408 DB_DNODE_EXIT(db); 1409 return (dr); 1410} 1411 1412/* 1413 * Undirty a buffer in the transaction group referenced by the given 1414 * transaction. Return whether this evicted the dbuf. 1415 */ 1416static boolean_t 1417dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1418{ 1419 dnode_t *dn; 1420 uint64_t txg = tx->tx_txg; 1421 dbuf_dirty_record_t *dr, **drp; 1422 1423 ASSERT(txg != 0); 1424 1425 /* 1426 * Due to our use of dn_nlevels below, this can only be called 1427 * in open context, unless we are operating on the MOS. 1428 * From syncing context, dn_nlevels may be different from the 1429 * dn_nlevels used when dbuf was dirtied. 1430 */ 1431 ASSERT(db->db_objset == 1432 dmu_objset_pool(db->db_objset)->dp_meta_objset || 1433 txg != spa_syncing_txg(dmu_objset_spa(db->db_objset))); 1434 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1435 ASSERT0(db->db_level); 1436 ASSERT(MUTEX_HELD(&db->db_mtx)); 1437 1438 /* 1439 * If this buffer is not dirty, we're done. 1440 */ 1441 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next) 1442 if (dr->dr_txg <= txg) 1443 break; 1444 if (dr == NULL || dr->dr_txg < txg) 1445 return (B_FALSE); 1446 ASSERT(dr->dr_txg == txg); 1447 ASSERT(dr->dr_dbuf == db); 1448 1449 DB_DNODE_ENTER(db); 1450 dn = DB_DNODE(db); 1451 1452 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 1453 1454 ASSERT(db->db.db_size != 0); 1455 1456 dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset), 1457 dr->dr_accounted, txg); 1458 1459 *drp = dr->dr_next; 1460 1461 /* 1462 * Note that there are three places in dbuf_dirty() 1463 * where this dirty record may be put on a list. 1464 * Make sure to do a list_remove corresponding to 1465 * every one of those list_insert calls. 1466 */ 1467 if (dr->dr_parent) { 1468 mutex_enter(&dr->dr_parent->dt.di.dr_mtx); 1469 list_remove(&dr->dr_parent->dt.di.dr_children, dr); 1470 mutex_exit(&dr->dr_parent->dt.di.dr_mtx); 1471 } else if (db->db_blkid == DMU_SPILL_BLKID || 1472 db->db_level + 1 == dn->dn_nlevels) { 1473 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf); 1474 mutex_enter(&dn->dn_mtx); 1475 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr); 1476 mutex_exit(&dn->dn_mtx); 1477 } 1478 DB_DNODE_EXIT(db); 1479 1480 if (db->db_state != DB_NOFILL) { 1481 dbuf_unoverride(dr); 1482 1483 ASSERT(db->db_buf != NULL); 1484 ASSERT(dr->dt.dl.dr_data != NULL); 1485 if (dr->dt.dl.dr_data != db->db_buf) 1486 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db)); 1487 } 1488 1489 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 1490 1491 ASSERT(db->db_dirtycnt > 0); 1492 db->db_dirtycnt -= 1; 1493 1494 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) { 1495 arc_buf_t *buf = db->db_buf; 1496 1497 ASSERT(db->db_state == DB_NOFILL || arc_released(buf)); 1498 dbuf_clear_data(db); 1499 VERIFY(arc_buf_remove_ref(buf, db)); 1500 dbuf_evict(db); 1501 return (B_TRUE); 1502 } 1503 1504 return (B_FALSE); 1505} 1506 1507void 1508dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx) 1509{ 1510 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1511 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH; 1512 1513 ASSERT(tx->tx_txg != 0); 1514 ASSERT(!refcount_is_zero(&db->db_holds)); 1515 1516 /* 1517 * Quick check for dirtyness. For already dirty blocks, this 1518 * reduces runtime of this function by >90%, and overall performance 1519 * by 50% for some workloads (e.g. file deletion with indirect blocks 1520 * cached). 1521 */ 1522 mutex_enter(&db->db_mtx); 1523 dbuf_dirty_record_t *dr; 1524 for (dr = db->db_last_dirty; 1525 dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) { 1526 /* 1527 * It's possible that it is already dirty but not cached, 1528 * because there are some calls to dbuf_dirty() that don't 1529 * go through dmu_buf_will_dirty(). 1530 */ 1531 if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) { 1532 /* This dbuf is already dirty and cached. */ 1533 dbuf_redirty(dr); 1534 mutex_exit(&db->db_mtx); 1535 return; 1536 } 1537 } 1538 mutex_exit(&db->db_mtx); 1539 1540 DB_DNODE_ENTER(db); 1541 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock)) 1542 rf |= DB_RF_HAVESTRUCT; 1543 DB_DNODE_EXIT(db); 1544 (void) dbuf_read(db, NULL, rf); 1545 (void) dbuf_dirty(db, tx); 1546} 1547 1548void 1549dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) 1550{ 1551 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1552 1553 db->db_state = DB_NOFILL; 1554 1555 dmu_buf_will_fill(db_fake, tx); 1556} 1557 1558void 1559dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) 1560{ 1561 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1562 1563 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1564 ASSERT(tx->tx_txg != 0); 1565 ASSERT(db->db_level == 0); 1566 ASSERT(!refcount_is_zero(&db->db_holds)); 1567 1568 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT || 1569 dmu_tx_private_ok(tx)); 1570 1571 dbuf_noread(db); 1572 (void) dbuf_dirty(db, tx); 1573} 1574 1575#pragma weak dmu_buf_fill_done = dbuf_fill_done 1576/* ARGSUSED */ 1577void 1578dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx) 1579{ 1580 mutex_enter(&db->db_mtx); 1581 DBUF_VERIFY(db); 1582 1583 if (db->db_state == DB_FILL) { 1584 if (db->db_level == 0 && db->db_freed_in_flight) { 1585 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1586 /* we were freed while filling */ 1587 /* XXX dbuf_undirty? */ 1588 bzero(db->db.db_data, db->db.db_size); 1589 db->db_freed_in_flight = FALSE; 1590 } 1591 db->db_state = DB_CACHED; 1592 cv_broadcast(&db->db_changed); 1593 } 1594 mutex_exit(&db->db_mtx); 1595} 1596 1597void 1598dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data, 1599 bp_embedded_type_t etype, enum zio_compress comp, 1600 int uncompressed_size, int compressed_size, int byteorder, 1601 dmu_tx_t *tx) 1602{ 1603 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; 1604 struct dirty_leaf *dl; 1605 dmu_object_type_t type; 1606 1607 if (etype == BP_EMBEDDED_TYPE_DATA) { 1608 ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset), 1609 SPA_FEATURE_EMBEDDED_DATA)); 1610 } 1611 1612 DB_DNODE_ENTER(db); 1613 type = DB_DNODE(db)->dn_type; 1614 DB_DNODE_EXIT(db); 1615 1616 ASSERT0(db->db_level); 1617 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1618 1619 dmu_buf_will_not_fill(dbuf, tx); 1620 1621 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); 1622 dl = &db->db_last_dirty->dt.dl; 1623 encode_embedded_bp_compressed(&dl->dr_overridden_by, 1624 data, comp, uncompressed_size, compressed_size); 1625 BPE_SET_ETYPE(&dl->dr_overridden_by, etype); 1626 BP_SET_TYPE(&dl->dr_overridden_by, type); 1627 BP_SET_LEVEL(&dl->dr_overridden_by, 0); 1628 BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder); 1629 1630 dl->dr_override_state = DR_OVERRIDDEN; 1631 dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg; 1632} 1633 1634/* 1635 * Directly assign a provided arc buf to a given dbuf if it's not referenced 1636 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf. 1637 */ 1638void 1639dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx) 1640{ 1641 ASSERT(!refcount_is_zero(&db->db_holds)); 1642 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1643 ASSERT(db->db_level == 0); 1644 ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA); 1645 ASSERT(buf != NULL); 1646 ASSERT(arc_buf_size(buf) == db->db.db_size); 1647 ASSERT(tx->tx_txg != 0); 1648 1649 arc_return_buf(buf, db); 1650 ASSERT(arc_released(buf)); 1651 1652 mutex_enter(&db->db_mtx); 1653 1654 while (db->db_state == DB_READ || db->db_state == DB_FILL) 1655 cv_wait(&db->db_changed, &db->db_mtx); 1656 1657 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED); 1658 1659 if (db->db_state == DB_CACHED && 1660 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) { 1661 mutex_exit(&db->db_mtx); 1662 (void) dbuf_dirty(db, tx); 1663 bcopy(buf->b_data, db->db.db_data, db->db.db_size); 1664 VERIFY(arc_buf_remove_ref(buf, db)); 1665 xuio_stat_wbuf_copied(); 1666 return; 1667 } 1668 1669 xuio_stat_wbuf_nocopy(); 1670 if (db->db_state == DB_CACHED) { 1671 dbuf_dirty_record_t *dr = db->db_last_dirty; 1672 1673 ASSERT(db->db_buf != NULL); 1674 if (dr != NULL && dr->dr_txg == tx->tx_txg) { 1675 ASSERT(dr->dt.dl.dr_data == db->db_buf); 1676 if (!arc_released(db->db_buf)) { 1677 ASSERT(dr->dt.dl.dr_override_state == 1678 DR_OVERRIDDEN); 1679 arc_release(db->db_buf, db); 1680 } 1681 dr->dt.dl.dr_data = buf; 1682 VERIFY(arc_buf_remove_ref(db->db_buf, db)); 1683 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) { 1684 arc_release(db->db_buf, db); 1685 VERIFY(arc_buf_remove_ref(db->db_buf, db)); 1686 } 1687 db->db_buf = NULL; 1688 } 1689 ASSERT(db->db_buf == NULL); 1690 dbuf_set_data(db, buf); 1691 db->db_state = DB_FILL; 1692 mutex_exit(&db->db_mtx); 1693 (void) dbuf_dirty(db, tx); 1694 dmu_buf_fill_done(&db->db, tx); 1695} 1696 1697/* 1698 * "Clear" the contents of this dbuf. This will mark the dbuf 1699 * EVICTING and clear *most* of its references. Unfortunately, 1700 * when we are not holding the dn_dbufs_mtx, we can't clear the 1701 * entry in the dn_dbufs list. We have to wait until dbuf_destroy() 1702 * in this case. For callers from the DMU we will usually see: 1703 * dbuf_clear()->arc_clear_callback()->dbuf_do_evict()->dbuf_destroy() 1704 * For the arc callback, we will usually see: 1705 * dbuf_do_evict()->dbuf_clear();dbuf_destroy() 1706 * Sometimes, though, we will get a mix of these two: 1707 * DMU: dbuf_clear()->arc_clear_callback() 1708 * ARC: dbuf_do_evict()->dbuf_destroy() 1709 * 1710 * This routine will dissociate the dbuf from the arc, by calling 1711 * arc_clear_callback(), but will not evict the data from the ARC. 1712 */ 1713void 1714dbuf_clear(dmu_buf_impl_t *db) 1715{ 1716 dnode_t *dn; 1717 dmu_buf_impl_t *parent = db->db_parent; 1718 dmu_buf_impl_t *dndb; 1719 boolean_t dbuf_gone = B_FALSE; 1720 1721 ASSERT(MUTEX_HELD(&db->db_mtx)); 1722 ASSERT(refcount_is_zero(&db->db_holds)); 1723 1724 dbuf_evict_user(db); 1725 1726 if (db->db_state == DB_CACHED) { 1727 ASSERT(db->db.db_data != NULL); 1728 if (db->db_blkid == DMU_BONUS_BLKID) { 1729 zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN); 1730 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 1731 } 1732 db->db.db_data = NULL; 1733 db->db_state = DB_UNCACHED; 1734 } 1735 1736 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL); 1737 ASSERT(db->db_data_pending == NULL); 1738 1739 db->db_state = DB_EVICTING; 1740 db->db_blkptr = NULL; 1741 1742 DB_DNODE_ENTER(db); 1743 dn = DB_DNODE(db); 1744 dndb = dn->dn_dbuf; 1745 if (db->db_blkid != DMU_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) { 1746 avl_remove(&dn->dn_dbufs, db); 1747 atomic_dec_32(&dn->dn_dbufs_count); 1748 membar_producer(); 1749 DB_DNODE_EXIT(db); 1750 /* 1751 * Decrementing the dbuf count means that the hold corresponding 1752 * to the removed dbuf is no longer discounted in dnode_move(), 1753 * so the dnode cannot be moved until after we release the hold. 1754 * The membar_producer() ensures visibility of the decremented 1755 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually 1756 * release any lock. 1757 */ 1758 dnode_rele(dn, db); 1759 db->db_dnode_handle = NULL; 1760 } else { 1761 DB_DNODE_EXIT(db); 1762 } 1763 1764 if (db->db_buf) 1765 dbuf_gone = arc_clear_callback(db->db_buf); 1766 1767 if (!dbuf_gone) 1768 mutex_exit(&db->db_mtx); 1769 1770 /* 1771 * If this dbuf is referenced from an indirect dbuf, 1772 * decrement the ref count on the indirect dbuf. 1773 */ 1774 if (parent && parent != dndb) 1775 dbuf_rele(parent, db); 1776} 1777 1778/* 1779 * Note: While bpp will always be updated if the function returns success, 1780 * parentp will not be updated if the dnode does not have dn_dbuf filled in; 1781 * this happens when the dnode is the meta-dnode, or a userused or groupused 1782 * object. 1783 */ 1784static int 1785dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse, 1786 dmu_buf_impl_t **parentp, blkptr_t **bpp) 1787{ 1788 int nlevels, epbs; 1789 1790 *parentp = NULL; 1791 *bpp = NULL; 1792 1793 ASSERT(blkid != DMU_BONUS_BLKID); 1794 1795 if (blkid == DMU_SPILL_BLKID) { 1796 mutex_enter(&dn->dn_mtx); 1797 if (dn->dn_have_spill && 1798 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) 1799 *bpp = &dn->dn_phys->dn_spill; 1800 else 1801 *bpp = NULL; 1802 dbuf_add_ref(dn->dn_dbuf, NULL); 1803 *parentp = dn->dn_dbuf; 1804 mutex_exit(&dn->dn_mtx); 1805 return (0); 1806 } 1807 1808 if (dn->dn_phys->dn_nlevels == 0) 1809 nlevels = 1; 1810 else 1811 nlevels = dn->dn_phys->dn_nlevels; 1812 1813 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1814 1815 ASSERT3U(level * epbs, <, 64); 1816 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1817 if (level >= nlevels || 1818 (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) { 1819 /* the buffer has no parent yet */ 1820 return (SET_ERROR(ENOENT)); 1821 } else if (level < nlevels-1) { 1822 /* this block is referenced from an indirect block */ 1823 int err = dbuf_hold_impl(dn, level+1, 1824 blkid >> epbs, fail_sparse, FALSE, NULL, parentp); 1825 if (err) 1826 return (err); 1827 err = dbuf_read(*parentp, NULL, 1828 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL)); 1829 if (err) { 1830 dbuf_rele(*parentp, NULL); 1831 *parentp = NULL; 1832 return (err); 1833 } 1834 *bpp = ((blkptr_t *)(*parentp)->db.db_data) + 1835 (blkid & ((1ULL << epbs) - 1)); 1836 return (0); 1837 } else { 1838 /* the block is referenced from the dnode */ 1839 ASSERT3U(level, ==, nlevels-1); 1840 ASSERT(dn->dn_phys->dn_nblkptr == 0 || 1841 blkid < dn->dn_phys->dn_nblkptr); 1842 if (dn->dn_dbuf) { 1843 dbuf_add_ref(dn->dn_dbuf, NULL); 1844 *parentp = dn->dn_dbuf; 1845 } 1846 *bpp = &dn->dn_phys->dn_blkptr[blkid]; 1847 return (0); 1848 } 1849} 1850 1851static dmu_buf_impl_t * 1852dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid, 1853 dmu_buf_impl_t *parent, blkptr_t *blkptr) 1854{ 1855 objset_t *os = dn->dn_objset; 1856 dmu_buf_impl_t *db, *odb; 1857 1858 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1859 ASSERT(dn->dn_type != DMU_OT_NONE); 1860 1861 db = kmem_cache_alloc(dbuf_cache, KM_SLEEP); 1862 1863 db->db_objset = os; 1864 db->db.db_object = dn->dn_object; 1865 db->db_level = level; 1866 db->db_blkid = blkid; 1867 db->db_last_dirty = NULL; 1868 db->db_dirtycnt = 0; 1869 db->db_dnode_handle = dn->dn_handle; 1870 db->db_parent = parent; 1871 db->db_blkptr = blkptr; 1872 1873 db->db_user = NULL; 1874 db->db_user_immediate_evict = FALSE; 1875 db->db_freed_in_flight = FALSE; 1876 db->db_pending_evict = FALSE; 1877 1878 if (blkid == DMU_BONUS_BLKID) { 1879 ASSERT3P(parent, ==, dn->dn_dbuf); 1880 db->db.db_size = DN_MAX_BONUSLEN - 1881 (dn->dn_nblkptr-1) * sizeof (blkptr_t); 1882 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 1883 db->db.db_offset = DMU_BONUS_BLKID; 1884 db->db_state = DB_UNCACHED; 1885 /* the bonus dbuf is not placed in the hash table */ 1886 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 1887 return (db); 1888 } else if (blkid == DMU_SPILL_BLKID) { 1889 db->db.db_size = (blkptr != NULL) ? 1890 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE; 1891 db->db.db_offset = 0; 1892 } else { 1893 int blocksize = 1894 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz; 1895 db->db.db_size = blocksize; 1896 db->db.db_offset = db->db_blkid * blocksize; 1897 } 1898 1899 /* 1900 * Hold the dn_dbufs_mtx while we get the new dbuf 1901 * in the hash table *and* added to the dbufs list. 1902 * This prevents a possible deadlock with someone 1903 * trying to look up this dbuf before its added to the 1904 * dn_dbufs list. 1905 */ 1906 mutex_enter(&dn->dn_dbufs_mtx); 1907 db->db_state = DB_EVICTING; 1908 if ((odb = dbuf_hash_insert(db)) != NULL) { 1909 /* someone else inserted it first */ 1910 kmem_cache_free(dbuf_cache, db); 1911 mutex_exit(&dn->dn_dbufs_mtx); 1912 return (odb); 1913 } 1914 avl_add(&dn->dn_dbufs, db); 1915 if (db->db_level == 0 && db->db_blkid >= 1916 dn->dn_unlisted_l0_blkid) 1917 dn->dn_unlisted_l0_blkid = db->db_blkid + 1; 1918 db->db_state = DB_UNCACHED; 1919 mutex_exit(&dn->dn_dbufs_mtx); 1920 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 1921 1922 if (parent && parent != dn->dn_dbuf) 1923 dbuf_add_ref(parent, db); 1924 1925 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 1926 refcount_count(&dn->dn_holds) > 0); 1927 (void) refcount_add(&dn->dn_holds, db); 1928 atomic_inc_32(&dn->dn_dbufs_count); 1929 1930 dprintf_dbuf(db, "db=%p\n", db); 1931 1932 return (db); 1933} 1934 1935static int 1936dbuf_do_evict(void *private) 1937{ 1938 dmu_buf_impl_t *db = private; 1939 1940 if (!MUTEX_HELD(&db->db_mtx)) 1941 mutex_enter(&db->db_mtx); 1942 1943 ASSERT(refcount_is_zero(&db->db_holds)); 1944 1945 if (db->db_state != DB_EVICTING) { 1946 ASSERT(db->db_state == DB_CACHED); 1947 DBUF_VERIFY(db); 1948 db->db_buf = NULL; 1949 dbuf_evict(db); 1950 } else { 1951 mutex_exit(&db->db_mtx); 1952 dbuf_destroy(db); 1953 } 1954 return (0); 1955} 1956 1957static void 1958dbuf_destroy(dmu_buf_impl_t *db) 1959{ 1960 ASSERT(refcount_is_zero(&db->db_holds)); 1961 1962 if (db->db_blkid != DMU_BONUS_BLKID) { 1963 /* 1964 * If this dbuf is still on the dn_dbufs list, 1965 * remove it from that list. 1966 */ 1967 if (db->db_dnode_handle != NULL) { 1968 dnode_t *dn; 1969 1970 DB_DNODE_ENTER(db); 1971 dn = DB_DNODE(db); 1972 mutex_enter(&dn->dn_dbufs_mtx); 1973 avl_remove(&dn->dn_dbufs, db); 1974 atomic_dec_32(&dn->dn_dbufs_count); 1975 mutex_exit(&dn->dn_dbufs_mtx); 1976 DB_DNODE_EXIT(db); 1977 /* 1978 * Decrementing the dbuf count means that the hold 1979 * corresponding to the removed dbuf is no longer 1980 * discounted in dnode_move(), so the dnode cannot be 1981 * moved until after we release the hold. 1982 */ 1983 dnode_rele(dn, db); 1984 db->db_dnode_handle = NULL; 1985 } 1986 dbuf_hash_remove(db); 1987 } 1988 db->db_parent = NULL; 1989 db->db_buf = NULL; 1990 1991 ASSERT(db->db.db_data == NULL); 1992 ASSERT(db->db_hash_next == NULL); 1993 ASSERT(db->db_blkptr == NULL); 1994 ASSERT(db->db_data_pending == NULL); 1995 1996 kmem_cache_free(dbuf_cache, db); 1997 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 1998} 1999 2000typedef struct dbuf_prefetch_arg { 2001 spa_t *dpa_spa; /* The spa to issue the prefetch in. */ 2002 zbookmark_phys_t dpa_zb; /* The target block to prefetch. */ 2003 int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */ 2004 int dpa_curlevel; /* The current level that we're reading */ 2005 zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */ 2006 zio_t *dpa_zio; /* The parent zio_t for all prefetches. */ 2007 arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */ 2008} dbuf_prefetch_arg_t; 2009 2010/* 2011 * Actually issue the prefetch read for the block given. 2012 */ 2013static void 2014dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp) 2015{ 2016 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) 2017 return; 2018 2019 arc_flags_t aflags = 2020 dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH; 2021 2022 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp)); 2023 ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level); 2024 ASSERT(dpa->dpa_zio != NULL); 2025 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL, 2026 dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2027 &aflags, &dpa->dpa_zb); 2028} 2029 2030/* 2031 * Called when an indirect block above our prefetch target is read in. This 2032 * will either read in the next indirect block down the tree or issue the actual 2033 * prefetch if the next block down is our target. 2034 */ 2035static void 2036dbuf_prefetch_indirect_done(zio_t *zio, arc_buf_t *abuf, void *private) 2037{ 2038 dbuf_prefetch_arg_t *dpa = private; 2039 2040 ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel); 2041 ASSERT3S(dpa->dpa_curlevel, >, 0); 2042 if (zio != NULL) { 2043 ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel); 2044 ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size); 2045 ASSERT3P(zio->io_spa, ==, dpa->dpa_spa); 2046 } 2047 2048 dpa->dpa_curlevel--; 2049 2050 uint64_t nextblkid = dpa->dpa_zb.zb_blkid >> 2051 (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level)); 2052 blkptr_t *bp = ((blkptr_t *)abuf->b_data) + 2053 P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs); 2054 if (BP_IS_HOLE(bp) || (zio != NULL && zio->io_error != 0)) { 2055 kmem_free(dpa, sizeof (*dpa)); 2056 } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) { 2057 ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid); 2058 dbuf_issue_final_prefetch(dpa, bp); 2059 kmem_free(dpa, sizeof (*dpa)); 2060 } else { 2061 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT; 2062 zbookmark_phys_t zb; 2063 2064 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp)); 2065 2066 SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset, 2067 dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid); 2068 2069 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, 2070 bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio, 2071 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2072 &iter_aflags, &zb); 2073 } 2074 (void) arc_buf_remove_ref(abuf, private); 2075} 2076 2077/* 2078 * Issue prefetch reads for the given block on the given level. If the indirect 2079 * blocks above that block are not in memory, we will read them in 2080 * asynchronously. As a result, this call never blocks waiting for a read to 2081 * complete. 2082 */ 2083void 2084dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio, 2085 arc_flags_t aflags) 2086{ 2087 blkptr_t bp; 2088 int epbs, nlevels, curlevel; 2089 uint64_t curblkid; 2090 2091 ASSERT(blkid != DMU_BONUS_BLKID); 2092 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 2093 2094 if (blkid > dn->dn_maxblkid) 2095 return; 2096 2097 if (dnode_block_freed(dn, blkid)) 2098 return; 2099 2100 /* 2101 * This dnode hasn't been written to disk yet, so there's nothing to 2102 * prefetch. 2103 */ 2104 nlevels = dn->dn_phys->dn_nlevels; 2105 if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0) 2106 return; 2107 2108 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 2109 if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level)) 2110 return; 2111 2112 dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object, 2113 level, blkid); 2114 if (db != NULL) { 2115 mutex_exit(&db->db_mtx); 2116 /* 2117 * This dbuf already exists. It is either CACHED, or 2118 * (we assume) about to be read or filled. 2119 */ 2120 return; 2121 } 2122 2123 /* 2124 * Find the closest ancestor (indirect block) of the target block 2125 * that is present in the cache. In this indirect block, we will 2126 * find the bp that is at curlevel, curblkid. 2127 */ 2128 curlevel = level; 2129 curblkid = blkid; 2130 while (curlevel < nlevels - 1) { 2131 int parent_level = curlevel + 1; 2132 uint64_t parent_blkid = curblkid >> epbs; 2133 dmu_buf_impl_t *db; 2134 2135 if (dbuf_hold_impl(dn, parent_level, parent_blkid, 2136 FALSE, TRUE, FTAG, &db) == 0) { 2137 blkptr_t *bpp = db->db_buf->b_data; 2138 bp = bpp[P2PHASE(curblkid, 1 << epbs)]; 2139 dbuf_rele(db, FTAG); 2140 break; 2141 } 2142 2143 curlevel = parent_level; 2144 curblkid = parent_blkid; 2145 } 2146 2147 if (curlevel == nlevels - 1) { 2148 /* No cached indirect blocks found. */ 2149 ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr); 2150 bp = dn->dn_phys->dn_blkptr[curblkid]; 2151 } 2152 if (BP_IS_HOLE(&bp)) 2153 return; 2154 2155 ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp)); 2156 2157 zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL, 2158 ZIO_FLAG_CANFAIL); 2159 2160 dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP); 2161 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset; 2162 SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET, 2163 dn->dn_object, level, blkid); 2164 dpa->dpa_curlevel = curlevel; 2165 dpa->dpa_prio = prio; 2166 dpa->dpa_aflags = aflags; 2167 dpa->dpa_spa = dn->dn_objset->os_spa; 2168 dpa->dpa_epbs = epbs; 2169 dpa->dpa_zio = pio; 2170 2171 /* 2172 * If we have the indirect just above us, no need to do the asynchronous 2173 * prefetch chain; we'll just run the last step ourselves. If we're at 2174 * a higher level, though, we want to issue the prefetches for all the 2175 * indirect blocks asynchronously, so we can go on with whatever we were 2176 * doing. 2177 */ 2178 if (curlevel == level) { 2179 ASSERT3U(curblkid, ==, blkid); 2180 dbuf_issue_final_prefetch(dpa, &bp); 2181 kmem_free(dpa, sizeof (*dpa)); 2182 } else { 2183 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT; 2184 zbookmark_phys_t zb; 2185 2186 SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET, 2187 dn->dn_object, curlevel, curblkid); 2188 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, 2189 &bp, dbuf_prefetch_indirect_done, dpa, prio, 2190 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2191 &iter_aflags, &zb); 2192 } 2193 /* 2194 * We use pio here instead of dpa_zio since it's possible that 2195 * dpa may have already been freed. 2196 */ 2197 zio_nowait(pio); 2198} 2199 2200/* 2201 * Returns with db_holds incremented, and db_mtx not held. 2202 * Note: dn_struct_rwlock must be held. 2203 */ 2204int 2205dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, 2206 boolean_t fail_sparse, boolean_t fail_uncached, 2207 void *tag, dmu_buf_impl_t **dbp) 2208{ 2209 dmu_buf_impl_t *db, *parent = NULL; 2210 2211 ASSERT(blkid != DMU_BONUS_BLKID); 2212 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 2213 ASSERT3U(dn->dn_nlevels, >, level); 2214 2215 *dbp = NULL; 2216top: 2217 /* dbuf_find() returns with db_mtx held */ 2218 db = dbuf_find(dn->dn_objset, dn->dn_object, level, blkid); 2219 2220 if (db == NULL) { 2221 blkptr_t *bp = NULL; 2222 int err; 2223 2224 if (fail_uncached) 2225 return (SET_ERROR(ENOENT)); 2226 2227 ASSERT3P(parent, ==, NULL); 2228 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp); 2229 if (fail_sparse) { 2230 if (err == 0 && bp && BP_IS_HOLE(bp)) 2231 err = SET_ERROR(ENOENT); 2232 if (err) { 2233 if (parent) 2234 dbuf_rele(parent, NULL); 2235 return (err); 2236 } 2237 } 2238 if (err && err != ENOENT) 2239 return (err); 2240 db = dbuf_create(dn, level, blkid, parent, bp); 2241 } 2242 2243 if (fail_uncached && db->db_state != DB_CACHED) { 2244 mutex_exit(&db->db_mtx); 2245 return (SET_ERROR(ENOENT)); 2246 } 2247 2248 if (db->db_buf && refcount_is_zero(&db->db_holds)) { 2249 arc_buf_add_ref(db->db_buf, db); 2250 if (db->db_buf->b_data == NULL) { 2251 dbuf_clear(db); 2252 if (parent) { 2253 dbuf_rele(parent, NULL); 2254 parent = NULL; 2255 } 2256 goto top; 2257 } 2258 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data); 2259 } 2260 2261 ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf)); 2262 2263 /* 2264 * If this buffer is currently syncing out, and we are are 2265 * still referencing it from db_data, we need to make a copy 2266 * of it in case we decide we want to dirty it again in this txg. 2267 */ 2268 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 2269 dn->dn_object != DMU_META_DNODE_OBJECT && 2270 db->db_state == DB_CACHED && db->db_data_pending) { 2271 dbuf_dirty_record_t *dr = db->db_data_pending; 2272 2273 if (dr->dt.dl.dr_data == db->db_buf) { 2274 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 2275 2276 dbuf_set_data(db, 2277 arc_buf_alloc(dn->dn_objset->os_spa, 2278 db->db.db_size, db, type)); 2279 bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data, 2280 db->db.db_size); 2281 } 2282 } 2283 2284 (void) refcount_add(&db->db_holds, tag); 2285 DBUF_VERIFY(db); 2286 mutex_exit(&db->db_mtx); 2287 2288 /* NOTE: we can't rele the parent until after we drop the db_mtx */ 2289 if (parent) 2290 dbuf_rele(parent, NULL); 2291 2292 ASSERT3P(DB_DNODE(db), ==, dn); 2293 ASSERT3U(db->db_blkid, ==, blkid); 2294 ASSERT3U(db->db_level, ==, level); 2295 *dbp = db; 2296 2297 return (0); 2298} 2299 2300dmu_buf_impl_t * 2301dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag) 2302{ 2303 return (dbuf_hold_level(dn, 0, blkid, tag)); 2304} 2305 2306dmu_buf_impl_t * 2307dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag) 2308{ 2309 dmu_buf_impl_t *db; 2310 int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db); 2311 return (err ? NULL : db); 2312} 2313 2314void 2315dbuf_create_bonus(dnode_t *dn) 2316{ 2317 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 2318 2319 ASSERT(dn->dn_bonus == NULL); 2320 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL); 2321} 2322 2323int 2324dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx) 2325{ 2326 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2327 dnode_t *dn; 2328 2329 if (db->db_blkid != DMU_SPILL_BLKID) 2330 return (SET_ERROR(ENOTSUP)); 2331 if (blksz == 0) 2332 blksz = SPA_MINBLOCKSIZE; 2333 ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset))); 2334 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE); 2335 2336 DB_DNODE_ENTER(db); 2337 dn = DB_DNODE(db); 2338 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 2339 dbuf_new_size(db, blksz, tx); 2340 rw_exit(&dn->dn_struct_rwlock); 2341 DB_DNODE_EXIT(db); 2342 2343 return (0); 2344} 2345 2346void 2347dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx) 2348{ 2349 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx); 2350} 2351 2352#pragma weak dmu_buf_add_ref = dbuf_add_ref 2353void 2354dbuf_add_ref(dmu_buf_impl_t *db, void *tag) 2355{ 2356 int64_t holds = refcount_add(&db->db_holds, tag); 2357 ASSERT(holds > 1); 2358} 2359 2360#pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref 2361boolean_t 2362dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid, 2363 void *tag) 2364{ 2365 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2366 dmu_buf_impl_t *found_db; 2367 boolean_t result = B_FALSE; 2368 2369 if (db->db_blkid == DMU_BONUS_BLKID) 2370 found_db = dbuf_find_bonus(os, obj); 2371 else 2372 found_db = dbuf_find(os, obj, 0, blkid); 2373 2374 if (found_db != NULL) { 2375 if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) { 2376 (void) refcount_add(&db->db_holds, tag); 2377 result = B_TRUE; 2378 } 2379 mutex_exit(&db->db_mtx); 2380 } 2381 return (result); 2382} 2383 2384/* 2385 * If you call dbuf_rele() you had better not be referencing the dnode handle 2386 * unless you have some other direct or indirect hold on the dnode. (An indirect 2387 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.) 2388 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the 2389 * dnode's parent dbuf evicting its dnode handles. 2390 */ 2391void 2392dbuf_rele(dmu_buf_impl_t *db, void *tag) 2393{ 2394 mutex_enter(&db->db_mtx); 2395 dbuf_rele_and_unlock(db, tag); 2396} 2397 2398void 2399dmu_buf_rele(dmu_buf_t *db, void *tag) 2400{ 2401 dbuf_rele((dmu_buf_impl_t *)db, tag); 2402} 2403 2404/* 2405 * dbuf_rele() for an already-locked dbuf. This is necessary to allow 2406 * db_dirtycnt and db_holds to be updated atomically. 2407 */ 2408void 2409dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag) 2410{ 2411 int64_t holds; 2412 2413 ASSERT(MUTEX_HELD(&db->db_mtx)); 2414 DBUF_VERIFY(db); 2415 2416 /* 2417 * Remove the reference to the dbuf before removing its hold on the 2418 * dnode so we can guarantee in dnode_move() that a referenced bonus 2419 * buffer has a corresponding dnode hold. 2420 */ 2421 holds = refcount_remove(&db->db_holds, tag); 2422 ASSERT(holds >= 0); 2423 2424 /* 2425 * We can't freeze indirects if there is a possibility that they 2426 * may be modified in the current syncing context. 2427 */ 2428 if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) 2429 arc_buf_freeze(db->db_buf); 2430 2431 if (holds == db->db_dirtycnt && 2432 db->db_level == 0 && db->db_user_immediate_evict) 2433 dbuf_evict_user(db); 2434 2435 if (holds == 0) { 2436 if (db->db_blkid == DMU_BONUS_BLKID) { 2437 dnode_t *dn; 2438 boolean_t evict_dbuf = db->db_pending_evict; 2439 2440 /* 2441 * If the dnode moves here, we cannot cross this 2442 * barrier until the move completes. 2443 */ 2444 DB_DNODE_ENTER(db); 2445 2446 dn = DB_DNODE(db); 2447 atomic_dec_32(&dn->dn_dbufs_count); 2448 2449 /* 2450 * Decrementing the dbuf count means that the bonus 2451 * buffer's dnode hold is no longer discounted in 2452 * dnode_move(). The dnode cannot move until after 2453 * the dnode_rele() below. 2454 */ 2455 DB_DNODE_EXIT(db); 2456 2457 /* 2458 * Do not reference db after its lock is dropped. 2459 * Another thread may evict it. 2460 */ 2461 mutex_exit(&db->db_mtx); 2462 2463 if (evict_dbuf) 2464 dnode_evict_bonus(dn); 2465 2466 dnode_rele(dn, db); 2467 } else if (db->db_buf == NULL) { 2468 /* 2469 * This is a special case: we never associated this 2470 * dbuf with any data allocated from the ARC. 2471 */ 2472 ASSERT(db->db_state == DB_UNCACHED || 2473 db->db_state == DB_NOFILL); 2474 dbuf_evict(db); 2475 } else if (arc_released(db->db_buf)) { 2476 arc_buf_t *buf = db->db_buf; 2477 /* 2478 * This dbuf has anonymous data associated with it. 2479 */ 2480 dbuf_clear_data(db); 2481 VERIFY(arc_buf_remove_ref(buf, db)); 2482 dbuf_evict(db); 2483 } else { 2484 VERIFY(!arc_buf_remove_ref(db->db_buf, db)); 2485 2486 /* 2487 * A dbuf will be eligible for eviction if either the 2488 * 'primarycache' property is set or a duplicate 2489 * copy of this buffer is already cached in the arc. 2490 * 2491 * In the case of the 'primarycache' a buffer 2492 * is considered for eviction if it matches the 2493 * criteria set in the property. 2494 * 2495 * To decide if our buffer is considered a 2496 * duplicate, we must call into the arc to determine 2497 * if multiple buffers are referencing the same 2498 * block on-disk. If so, then we simply evict 2499 * ourselves. 2500 */ 2501 if (!DBUF_IS_CACHEABLE(db)) { 2502 if (db->db_blkptr != NULL && 2503 !BP_IS_HOLE(db->db_blkptr) && 2504 !BP_IS_EMBEDDED(db->db_blkptr)) { 2505 spa_t *spa = 2506 dmu_objset_spa(db->db_objset); 2507 blkptr_t bp = *db->db_blkptr; 2508 dbuf_clear(db); 2509 arc_freed(spa, &bp); 2510 } else { 2511 dbuf_clear(db); 2512 } 2513 } else if (db->db_pending_evict || 2514 arc_buf_eviction_needed(db->db_buf)) { 2515 dbuf_clear(db); 2516 } else { 2517 mutex_exit(&db->db_mtx); 2518 } 2519 } 2520 } else { 2521 mutex_exit(&db->db_mtx); 2522 } 2523} 2524 2525#pragma weak dmu_buf_refcount = dbuf_refcount 2526uint64_t 2527dbuf_refcount(dmu_buf_impl_t *db) 2528{ 2529 return (refcount_count(&db->db_holds)); 2530} 2531 2532void * 2533dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user, 2534 dmu_buf_user_t *new_user) 2535{ 2536 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2537 2538 mutex_enter(&db->db_mtx); 2539 dbuf_verify_user(db, DBVU_NOT_EVICTING); 2540 if (db->db_user == old_user) 2541 db->db_user = new_user; 2542 else 2543 old_user = db->db_user; 2544 dbuf_verify_user(db, DBVU_NOT_EVICTING); 2545 mutex_exit(&db->db_mtx); 2546 2547 return (old_user); 2548} 2549 2550void * 2551dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user) 2552{ 2553 return (dmu_buf_replace_user(db_fake, NULL, user)); 2554} 2555 2556void * 2557dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user) 2558{ 2559 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2560 2561 db->db_user_immediate_evict = TRUE; 2562 return (dmu_buf_set_user(db_fake, user)); 2563} 2564 2565void * 2566dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user) 2567{ 2568 return (dmu_buf_replace_user(db_fake, user, NULL)); 2569} 2570 2571void * 2572dmu_buf_get_user(dmu_buf_t *db_fake) 2573{ 2574 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2575 2576 dbuf_verify_user(db, DBVU_NOT_EVICTING); 2577 return (db->db_user); 2578} 2579 2580void 2581dmu_buf_user_evict_wait() 2582{ 2583 taskq_wait(dbu_evict_taskq); 2584} 2585 2586boolean_t 2587dmu_buf_freeable(dmu_buf_t *dbuf) 2588{ 2589 boolean_t res = B_FALSE; 2590 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; 2591 2592 if (db->db_blkptr) 2593 res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset, 2594 db->db_blkptr, db->db_blkptr->blk_birth); 2595 2596 return (res); 2597} 2598 2599blkptr_t * 2600dmu_buf_get_blkptr(dmu_buf_t *db) 2601{ 2602 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 2603 return (dbi->db_blkptr); 2604} 2605 2606static void 2607dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db) 2608{ 2609 /* ASSERT(dmu_tx_is_syncing(tx) */ 2610 ASSERT(MUTEX_HELD(&db->db_mtx)); 2611 2612 if (db->db_blkptr != NULL) 2613 return; 2614 2615 if (db->db_blkid == DMU_SPILL_BLKID) { 2616 db->db_blkptr = &dn->dn_phys->dn_spill; 2617 BP_ZERO(db->db_blkptr); 2618 return; 2619 } 2620 if (db->db_level == dn->dn_phys->dn_nlevels-1) { 2621 /* 2622 * This buffer was allocated at a time when there was 2623 * no available blkptrs from the dnode, or it was 2624 * inappropriate to hook it in (i.e., nlevels mis-match). 2625 */ 2626 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr); 2627 ASSERT(db->db_parent == NULL); 2628 db->db_parent = dn->dn_dbuf; 2629 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid]; 2630 DBUF_VERIFY(db); 2631 } else { 2632 dmu_buf_impl_t *parent = db->db_parent; 2633 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 2634 2635 ASSERT(dn->dn_phys->dn_nlevels > 1); 2636 if (parent == NULL) { 2637 mutex_exit(&db->db_mtx); 2638 rw_enter(&dn->dn_struct_rwlock, RW_READER); 2639 parent = dbuf_hold_level(dn, db->db_level + 1, 2640 db->db_blkid >> epbs, db); 2641 rw_exit(&dn->dn_struct_rwlock); 2642 mutex_enter(&db->db_mtx); 2643 db->db_parent = parent; 2644 } 2645 db->db_blkptr = (blkptr_t *)parent->db.db_data + 2646 (db->db_blkid & ((1ULL << epbs) - 1)); 2647 DBUF_VERIFY(db); 2648 } 2649} 2650 2651static void 2652dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx) 2653{ 2654 dmu_buf_impl_t *db = dr->dr_dbuf; 2655 dnode_t *dn; 2656 zio_t *zio; 2657 2658 ASSERT(dmu_tx_is_syncing(tx)); 2659 2660 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 2661 2662 mutex_enter(&db->db_mtx); 2663 2664 ASSERT(db->db_level > 0); 2665 DBUF_VERIFY(db); 2666 2667 /* Read the block if it hasn't been read yet. */ 2668 if (db->db_buf == NULL) { 2669 mutex_exit(&db->db_mtx); 2670 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED); 2671 mutex_enter(&db->db_mtx); 2672 } 2673 ASSERT3U(db->db_state, ==, DB_CACHED); 2674 ASSERT(db->db_buf != NULL); 2675 2676 DB_DNODE_ENTER(db); 2677 dn = DB_DNODE(db); 2678 /* Indirect block size must match what the dnode thinks it is. */ 2679 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 2680 dbuf_check_blkptr(dn, db); 2681 DB_DNODE_EXIT(db); 2682 2683 /* Provide the pending dirty record to child dbufs */ 2684 db->db_data_pending = dr; 2685 2686 mutex_exit(&db->db_mtx); 2687 dbuf_write(dr, db->db_buf, tx); 2688 2689 zio = dr->dr_zio; 2690 mutex_enter(&dr->dt.di.dr_mtx); 2691 dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx); 2692 ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 2693 mutex_exit(&dr->dt.di.dr_mtx); 2694 zio_nowait(zio); 2695} 2696 2697static void 2698dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx) 2699{ 2700 arc_buf_t **datap = &dr->dt.dl.dr_data; 2701 dmu_buf_impl_t *db = dr->dr_dbuf; 2702 dnode_t *dn; 2703 objset_t *os; 2704 uint64_t txg = tx->tx_txg; 2705 2706 ASSERT(dmu_tx_is_syncing(tx)); 2707 2708 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 2709 2710 mutex_enter(&db->db_mtx); 2711 /* 2712 * To be synced, we must be dirtied. But we 2713 * might have been freed after the dirty. 2714 */ 2715 if (db->db_state == DB_UNCACHED) { 2716 /* This buffer has been freed since it was dirtied */ 2717 ASSERT(db->db.db_data == NULL); 2718 } else if (db->db_state == DB_FILL) { 2719 /* This buffer was freed and is now being re-filled */ 2720 ASSERT(db->db.db_data != dr->dt.dl.dr_data); 2721 } else { 2722 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL); 2723 } 2724 DBUF_VERIFY(db); 2725 2726 DB_DNODE_ENTER(db); 2727 dn = DB_DNODE(db); 2728 2729 if (db->db_blkid == DMU_SPILL_BLKID) { 2730 mutex_enter(&dn->dn_mtx); 2731 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR; 2732 mutex_exit(&dn->dn_mtx); 2733 } 2734 2735 /* 2736 * If this is a bonus buffer, simply copy the bonus data into the 2737 * dnode. It will be written out when the dnode is synced (and it 2738 * will be synced, since it must have been dirty for dbuf_sync to 2739 * be called). 2740 */ 2741 if (db->db_blkid == DMU_BONUS_BLKID) { 2742 dbuf_dirty_record_t **drp; 2743 2744 ASSERT(*datap != NULL); 2745 ASSERT0(db->db_level); 2746 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN); 2747 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen); 2748 DB_DNODE_EXIT(db); 2749 2750 if (*datap != db->db.db_data) { 2751 zio_buf_free(*datap, DN_MAX_BONUSLEN); 2752 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 2753 } 2754 db->db_data_pending = NULL; 2755 drp = &db->db_last_dirty; 2756 while (*drp != dr) 2757 drp = &(*drp)->dr_next; 2758 ASSERT(dr->dr_next == NULL); 2759 ASSERT(dr->dr_dbuf == db); 2760 *drp = dr->dr_next; 2761 if (dr->dr_dbuf->db_level != 0) { 2762 list_destroy(&dr->dt.di.dr_children); 2763 mutex_destroy(&dr->dt.di.dr_mtx); 2764 } 2765 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 2766 ASSERT(db->db_dirtycnt > 0); 2767 db->db_dirtycnt -= 1; 2768 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg); 2769 return; 2770 } 2771 2772 os = dn->dn_objset; 2773 2774 /* 2775 * This function may have dropped the db_mtx lock allowing a dmu_sync 2776 * operation to sneak in. As a result, we need to ensure that we 2777 * don't check the dr_override_state until we have returned from 2778 * dbuf_check_blkptr. 2779 */ 2780 dbuf_check_blkptr(dn, db); 2781 2782 /* 2783 * If this buffer is in the middle of an immediate write, 2784 * wait for the synchronous IO to complete. 2785 */ 2786 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) { 2787 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT); 2788 cv_wait(&db->db_changed, &db->db_mtx); 2789 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN); 2790 } 2791 2792 if (db->db_state != DB_NOFILL && 2793 dn->dn_object != DMU_META_DNODE_OBJECT && 2794 refcount_count(&db->db_holds) > 1 && 2795 dr->dt.dl.dr_override_state != DR_OVERRIDDEN && 2796 *datap == db->db_buf) { 2797 /* 2798 * If this buffer is currently "in use" (i.e., there 2799 * are active holds and db_data still references it), 2800 * then make a copy before we start the write so that 2801 * any modifications from the open txg will not leak 2802 * into this write. 2803 * 2804 * NOTE: this copy does not need to be made for 2805 * objects only modified in the syncing context (e.g. 2806 * DNONE_DNODE blocks). 2807 */ 2808 int blksz = arc_buf_size(*datap); 2809 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 2810 *datap = arc_buf_alloc(os->os_spa, blksz, db, type); 2811 bcopy(db->db.db_data, (*datap)->b_data, blksz); 2812 } 2813 db->db_data_pending = dr; 2814 2815 mutex_exit(&db->db_mtx); 2816 2817 dbuf_write(dr, *datap, tx); 2818 2819 ASSERT(!list_link_active(&dr->dr_dirty_node)); 2820 if (dn->dn_object == DMU_META_DNODE_OBJECT) { 2821 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr); 2822 DB_DNODE_EXIT(db); 2823 } else { 2824 /* 2825 * Although zio_nowait() does not "wait for an IO", it does 2826 * initiate the IO. If this is an empty write it seems plausible 2827 * that the IO could actually be completed before the nowait 2828 * returns. We need to DB_DNODE_EXIT() first in case 2829 * zio_nowait() invalidates the dbuf. 2830 */ 2831 DB_DNODE_EXIT(db); 2832 zio_nowait(dr->dr_zio); 2833 } 2834} 2835 2836void 2837dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx) 2838{ 2839 dbuf_dirty_record_t *dr; 2840 2841 while (dr = list_head(list)) { 2842 if (dr->dr_zio != NULL) { 2843 /* 2844 * If we find an already initialized zio then we 2845 * are processing the meta-dnode, and we have finished. 2846 * The dbufs for all dnodes are put back on the list 2847 * during processing, so that we can zio_wait() 2848 * these IOs after initiating all child IOs. 2849 */ 2850 ASSERT3U(dr->dr_dbuf->db.db_object, ==, 2851 DMU_META_DNODE_OBJECT); 2852 break; 2853 } 2854 if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID && 2855 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) { 2856 VERIFY3U(dr->dr_dbuf->db_level, ==, level); 2857 } 2858 list_remove(list, dr); 2859 if (dr->dr_dbuf->db_level > 0) 2860 dbuf_sync_indirect(dr, tx); 2861 else 2862 dbuf_sync_leaf(dr, tx); 2863 } 2864} 2865 2866/* ARGSUSED */ 2867static void 2868dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb) 2869{ 2870 dmu_buf_impl_t *db = vdb; 2871 dnode_t *dn; 2872 blkptr_t *bp = zio->io_bp; 2873 blkptr_t *bp_orig = &zio->io_bp_orig; 2874 spa_t *spa = zio->io_spa; 2875 int64_t delta; 2876 uint64_t fill = 0; 2877 int i; 2878 2879 ASSERT3P(db->db_blkptr, ==, bp); 2880 2881 DB_DNODE_ENTER(db); 2882 dn = DB_DNODE(db); 2883 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig); 2884 dnode_diduse_space(dn, delta - zio->io_prev_space_delta); 2885 zio->io_prev_space_delta = delta; 2886 2887 if (bp->blk_birth != 0) { 2888 ASSERT((db->db_blkid != DMU_SPILL_BLKID && 2889 BP_GET_TYPE(bp) == dn->dn_type) || 2890 (db->db_blkid == DMU_SPILL_BLKID && 2891 BP_GET_TYPE(bp) == dn->dn_bonustype) || 2892 BP_IS_EMBEDDED(bp)); 2893 ASSERT(BP_GET_LEVEL(bp) == db->db_level); 2894 } 2895 2896 mutex_enter(&db->db_mtx); 2897 2898#ifdef ZFS_DEBUG 2899 if (db->db_blkid == DMU_SPILL_BLKID) { 2900 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); 2901 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && 2902 db->db_blkptr == &dn->dn_phys->dn_spill); 2903 } 2904#endif 2905 2906 if (db->db_level == 0) { 2907 mutex_enter(&dn->dn_mtx); 2908 if (db->db_blkid > dn->dn_phys->dn_maxblkid && 2909 db->db_blkid != DMU_SPILL_BLKID) 2910 dn->dn_phys->dn_maxblkid = db->db_blkid; 2911 mutex_exit(&dn->dn_mtx); 2912 2913 if (dn->dn_type == DMU_OT_DNODE) { 2914 dnode_phys_t *dnp = db->db.db_data; 2915 for (i = db->db.db_size >> DNODE_SHIFT; i > 0; 2916 i--, dnp++) { 2917 if (dnp->dn_type != DMU_OT_NONE) 2918 fill++; 2919 } 2920 } else { 2921 if (BP_IS_HOLE(bp)) { 2922 fill = 0; 2923 } else { 2924 fill = 1; 2925 } 2926 } 2927 } else { 2928 blkptr_t *ibp = db->db.db_data; 2929 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 2930 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) { 2931 if (BP_IS_HOLE(ibp)) 2932 continue; 2933 fill += BP_GET_FILL(ibp); 2934 } 2935 } 2936 DB_DNODE_EXIT(db); 2937 2938 if (!BP_IS_EMBEDDED(bp)) 2939 bp->blk_fill = fill; 2940 2941 mutex_exit(&db->db_mtx); 2942} 2943 2944/* 2945 * The SPA will call this callback several times for each zio - once 2946 * for every physical child i/o (zio->io_phys_children times). This 2947 * allows the DMU to monitor the progress of each logical i/o. For example, 2948 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z 2949 * block. There may be a long delay before all copies/fragments are completed, 2950 * so this callback allows us to retire dirty space gradually, as the physical 2951 * i/os complete. 2952 */ 2953/* ARGSUSED */ 2954static void 2955dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg) 2956{ 2957 dmu_buf_impl_t *db = arg; 2958 objset_t *os = db->db_objset; 2959 dsl_pool_t *dp = dmu_objset_pool(os); 2960 dbuf_dirty_record_t *dr; 2961 int delta = 0; 2962 2963 dr = db->db_data_pending; 2964 ASSERT3U(dr->dr_txg, ==, zio->io_txg); 2965 2966 /* 2967 * The callback will be called io_phys_children times. Retire one 2968 * portion of our dirty space each time we are called. Any rounding 2969 * error will be cleaned up by dsl_pool_sync()'s call to 2970 * dsl_pool_undirty_space(). 2971 */ 2972 delta = dr->dr_accounted / zio->io_phys_children; 2973 dsl_pool_undirty_space(dp, delta, zio->io_txg); 2974} 2975 2976/* ARGSUSED */ 2977static void 2978dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb) 2979{ 2980 dmu_buf_impl_t *db = vdb; 2981 blkptr_t *bp_orig = &zio->io_bp_orig; 2982 blkptr_t *bp = db->db_blkptr; 2983 objset_t *os = db->db_objset; 2984 dmu_tx_t *tx = os->os_synctx; 2985 dbuf_dirty_record_t **drp, *dr; 2986 2987 ASSERT0(zio->io_error); 2988 ASSERT(db->db_blkptr == bp); 2989 2990 /* 2991 * For nopwrites and rewrites we ensure that the bp matches our 2992 * original and bypass all the accounting. 2993 */ 2994 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) { 2995 ASSERT(BP_EQUAL(bp, bp_orig)); 2996 } else { 2997 dsl_dataset_t *ds = os->os_dsl_dataset; 2998 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE); 2999 dsl_dataset_block_born(ds, bp, tx); 3000 } 3001 3002 mutex_enter(&db->db_mtx); 3003 3004 DBUF_VERIFY(db); 3005 3006 drp = &db->db_last_dirty; 3007 while ((dr = *drp) != db->db_data_pending) 3008 drp = &dr->dr_next; 3009 ASSERT(!list_link_active(&dr->dr_dirty_node)); 3010 ASSERT(dr->dr_dbuf == db); 3011 ASSERT(dr->dr_next == NULL); 3012 *drp = dr->dr_next; 3013 3014#ifdef ZFS_DEBUG 3015 if (db->db_blkid == DMU_SPILL_BLKID) { 3016 dnode_t *dn; 3017 3018 DB_DNODE_ENTER(db); 3019 dn = DB_DNODE(db); 3020 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); 3021 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && 3022 db->db_blkptr == &dn->dn_phys->dn_spill); 3023 DB_DNODE_EXIT(db); 3024 } 3025#endif 3026 3027 if (db->db_level == 0) { 3028 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 3029 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); 3030 if (db->db_state != DB_NOFILL) { 3031 if (dr->dt.dl.dr_data != db->db_buf) 3032 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, 3033 db)); 3034 else if (!arc_released(db->db_buf)) 3035 arc_set_callback(db->db_buf, dbuf_do_evict, db); 3036 } 3037 } else { 3038 dnode_t *dn; 3039 3040 DB_DNODE_ENTER(db); 3041 dn = DB_DNODE(db); 3042 ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 3043 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift); 3044 if (!BP_IS_HOLE(db->db_blkptr)) { 3045 int epbs = 3046 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 3047 ASSERT3U(db->db_blkid, <=, 3048 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs)); 3049 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, 3050 db->db.db_size); 3051 if (!arc_released(db->db_buf)) 3052 arc_set_callback(db->db_buf, dbuf_do_evict, db); 3053 } 3054 DB_DNODE_EXIT(db); 3055 mutex_destroy(&dr->dt.di.dr_mtx); 3056 list_destroy(&dr->dt.di.dr_children); 3057 } 3058 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 3059 3060 cv_broadcast(&db->db_changed); 3061 ASSERT(db->db_dirtycnt > 0); 3062 db->db_dirtycnt -= 1; 3063 db->db_data_pending = NULL; 3064 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg); 3065} 3066 3067static void 3068dbuf_write_nofill_ready(zio_t *zio) 3069{ 3070 dbuf_write_ready(zio, NULL, zio->io_private); 3071} 3072 3073static void 3074dbuf_write_nofill_done(zio_t *zio) 3075{ 3076 dbuf_write_done(zio, NULL, zio->io_private); 3077} 3078 3079static void 3080dbuf_write_override_ready(zio_t *zio) 3081{ 3082 dbuf_dirty_record_t *dr = zio->io_private; 3083 dmu_buf_impl_t *db = dr->dr_dbuf; 3084 3085 dbuf_write_ready(zio, NULL, db); 3086} 3087 3088static void 3089dbuf_write_override_done(zio_t *zio) 3090{ 3091 dbuf_dirty_record_t *dr = zio->io_private; 3092 dmu_buf_impl_t *db = dr->dr_dbuf; 3093 blkptr_t *obp = &dr->dt.dl.dr_overridden_by; 3094 3095 mutex_enter(&db->db_mtx); 3096 if (!BP_EQUAL(zio->io_bp, obp)) { 3097 if (!BP_IS_HOLE(obp)) 3098 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp); 3099 arc_release(dr->dt.dl.dr_data, db); 3100 } 3101 mutex_exit(&db->db_mtx); 3102 3103 dbuf_write_done(zio, NULL, db); 3104} 3105 3106/* Issue I/O to commit a dirty buffer to disk. */ 3107static void 3108dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx) 3109{ 3110 dmu_buf_impl_t *db = dr->dr_dbuf; 3111 dnode_t *dn; 3112 objset_t *os; 3113 dmu_buf_impl_t *parent = db->db_parent; 3114 uint64_t txg = tx->tx_txg; 3115 zbookmark_phys_t zb; 3116 zio_prop_t zp; 3117 zio_t *zio; 3118 int wp_flag = 0; 3119 3120 DB_DNODE_ENTER(db); 3121 dn = DB_DNODE(db); 3122 os = dn->dn_objset; 3123 3124 if (db->db_state != DB_NOFILL) { 3125 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) { 3126 /* 3127 * Private object buffers are released here rather 3128 * than in dbuf_dirty() since they are only modified 3129 * in the syncing context and we don't want the 3130 * overhead of making multiple copies of the data. 3131 */ 3132 if (BP_IS_HOLE(db->db_blkptr)) { 3133 arc_buf_thaw(data); 3134 } else { 3135 dbuf_release_bp(db); 3136 } 3137 } 3138 } 3139 3140 if (parent != dn->dn_dbuf) { 3141 /* Our parent is an indirect block. */ 3142 /* We have a dirty parent that has been scheduled for write. */ 3143 ASSERT(parent && parent->db_data_pending); 3144 /* Our parent's buffer is one level closer to the dnode. */ 3145 ASSERT(db->db_level == parent->db_level-1); 3146 /* 3147 * We're about to modify our parent's db_data by modifying 3148 * our block pointer, so the parent must be released. 3149 */ 3150 ASSERT(arc_released(parent->db_buf)); 3151 zio = parent->db_data_pending->dr_zio; 3152 } else { 3153 /* Our parent is the dnode itself. */ 3154 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 && 3155 db->db_blkid != DMU_SPILL_BLKID) || 3156 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0)); 3157 if (db->db_blkid != DMU_SPILL_BLKID) 3158 ASSERT3P(db->db_blkptr, ==, 3159 &dn->dn_phys->dn_blkptr[db->db_blkid]); 3160 zio = dn->dn_zio; 3161 } 3162 3163 ASSERT(db->db_level == 0 || data == db->db_buf); 3164 ASSERT3U(db->db_blkptr->blk_birth, <=, txg); 3165 ASSERT(zio); 3166 3167 SET_BOOKMARK(&zb, os->os_dsl_dataset ? 3168 os->os_dsl_dataset->ds_object : DMU_META_OBJSET, 3169 db->db.db_object, db->db_level, db->db_blkid); 3170 3171 if (db->db_blkid == DMU_SPILL_BLKID) 3172 wp_flag = WP_SPILL; 3173 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0; 3174 3175 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp); 3176 DB_DNODE_EXIT(db); 3177 3178 if (db->db_level == 0 && 3179 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { 3180 /* 3181 * The BP for this block has been provided by open context 3182 * (by dmu_sync() or dmu_buf_write_embedded()). 3183 */ 3184 void *contents = (data != NULL) ? data->b_data : NULL; 3185 3186 dr->dr_zio = zio_write(zio, os->os_spa, txg, 3187 db->db_blkptr, contents, db->db.db_size, &zp, 3188 dbuf_write_override_ready, NULL, dbuf_write_override_done, 3189 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); 3190 mutex_enter(&db->db_mtx); 3191 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 3192 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by, 3193 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite); 3194 mutex_exit(&db->db_mtx); 3195 } else if (db->db_state == DB_NOFILL) { 3196 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF || 3197 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY); 3198 dr->dr_zio = zio_write(zio, os->os_spa, txg, 3199 db->db_blkptr, NULL, db->db.db_size, &zp, 3200 dbuf_write_nofill_ready, NULL, dbuf_write_nofill_done, db, 3201 ZIO_PRIORITY_ASYNC_WRITE, 3202 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb); 3203 } else { 3204 ASSERT(arc_released(data)); 3205 dr->dr_zio = arc_write(zio, os->os_spa, txg, 3206 db->db_blkptr, data, DBUF_IS_L2CACHEABLE(db), 3207 DBUF_IS_L2COMPRESSIBLE(db), &zp, dbuf_write_ready, 3208 dbuf_write_physdone, dbuf_write_done, db, 3209 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); 3210 } 3211} 3212