dbuf.c revision 290750
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 */ 29 30#include <sys/zfs_context.h> 31#include <sys/dmu.h> 32#include <sys/dmu_send.h> 33#include <sys/dmu_impl.h> 34#include <sys/dbuf.h> 35#include <sys/dmu_objset.h> 36#include <sys/dsl_dataset.h> 37#include <sys/dsl_dir.h> 38#include <sys/dmu_tx.h> 39#include <sys/spa.h> 40#include <sys/zio.h> 41#include <sys/dmu_zfetch.h> 42#include <sys/sa.h> 43#include <sys/sa_impl.h> 44#include <sys/zfeature.h> 45#include <sys/blkptr.h> 46#include <sys/range_tree.h> 47 48/* 49 * Number of times that zfs_free_range() took the slow path while doing 50 * a zfs receive. A nonzero value indicates a potential performance problem. 51 */ 52uint64_t zfs_free_range_recv_miss; 53 54static void dbuf_destroy(dmu_buf_impl_t *db); 55static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx); 56static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx); 57 58/* 59 * Global data structures and functions for the dbuf cache. 60 */ 61static kmem_cache_t *dbuf_cache; 62static taskq_t *dbu_evict_taskq; 63 64/* ARGSUSED */ 65static int 66dbuf_cons(void *vdb, void *unused, int kmflag) 67{ 68 dmu_buf_impl_t *db = vdb; 69 bzero(db, sizeof (dmu_buf_impl_t)); 70 71 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL); 72 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL); 73 refcount_create(&db->db_holds); 74 75 return (0); 76} 77 78/* ARGSUSED */ 79static void 80dbuf_dest(void *vdb, void *unused) 81{ 82 dmu_buf_impl_t *db = vdb; 83 mutex_destroy(&db->db_mtx); 84 cv_destroy(&db->db_changed); 85 refcount_destroy(&db->db_holds); 86} 87 88/* 89 * dbuf hash table routines 90 */ 91static dbuf_hash_table_t dbuf_hash_table; 92 93static uint64_t dbuf_hash_count; 94 95static uint64_t 96dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid) 97{ 98 uintptr_t osv = (uintptr_t)os; 99 uint64_t crc = -1ULL; 100 101 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 102 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF]; 103 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF]; 104 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF]; 105 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF]; 106 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF]; 107 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF]; 108 109 crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16); 110 111 return (crc); 112} 113 114#define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid); 115 116#define DBUF_EQUAL(dbuf, os, obj, level, blkid) \ 117 ((dbuf)->db.db_object == (obj) && \ 118 (dbuf)->db_objset == (os) && \ 119 (dbuf)->db_level == (level) && \ 120 (dbuf)->db_blkid == (blkid)) 121 122dmu_buf_impl_t * 123dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid) 124{ 125 dbuf_hash_table_t *h = &dbuf_hash_table; 126 uint64_t hv = DBUF_HASH(os, obj, level, blkid); 127 uint64_t idx = hv & h->hash_table_mask; 128 dmu_buf_impl_t *db; 129 130 mutex_enter(DBUF_HASH_MUTEX(h, idx)); 131 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) { 132 if (DBUF_EQUAL(db, os, obj, level, blkid)) { 133 mutex_enter(&db->db_mtx); 134 if (db->db_state != DB_EVICTING) { 135 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 136 return (db); 137 } 138 mutex_exit(&db->db_mtx); 139 } 140 } 141 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 142 return (NULL); 143} 144 145static dmu_buf_impl_t * 146dbuf_find_bonus(objset_t *os, uint64_t object) 147{ 148 dnode_t *dn; 149 dmu_buf_impl_t *db = NULL; 150 151 if (dnode_hold(os, object, FTAG, &dn) == 0) { 152 rw_enter(&dn->dn_struct_rwlock, RW_READER); 153 if (dn->dn_bonus != NULL) { 154 db = dn->dn_bonus; 155 mutex_enter(&db->db_mtx); 156 } 157 rw_exit(&dn->dn_struct_rwlock); 158 dnode_rele(dn, FTAG); 159 } 160 return (db); 161} 162 163/* 164 * Insert an entry into the hash table. If there is already an element 165 * equal to elem in the hash table, then the already existing element 166 * will be returned and the new element will not be inserted. 167 * Otherwise returns NULL. 168 */ 169static dmu_buf_impl_t * 170dbuf_hash_insert(dmu_buf_impl_t *db) 171{ 172 dbuf_hash_table_t *h = &dbuf_hash_table; 173 objset_t *os = db->db_objset; 174 uint64_t obj = db->db.db_object; 175 int level = db->db_level; 176 uint64_t blkid = db->db_blkid; 177 uint64_t hv = DBUF_HASH(os, obj, level, blkid); 178 uint64_t idx = hv & h->hash_table_mask; 179 dmu_buf_impl_t *dbf; 180 181 mutex_enter(DBUF_HASH_MUTEX(h, idx)); 182 for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) { 183 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) { 184 mutex_enter(&dbf->db_mtx); 185 if (dbf->db_state != DB_EVICTING) { 186 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 187 return (dbf); 188 } 189 mutex_exit(&dbf->db_mtx); 190 } 191 } 192 193 mutex_enter(&db->db_mtx); 194 db->db_hash_next = h->hash_table[idx]; 195 h->hash_table[idx] = db; 196 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 197 atomic_inc_64(&dbuf_hash_count); 198 199 return (NULL); 200} 201 202/* 203 * Remove an entry from the hash table. It must be in the EVICTING state. 204 */ 205static void 206dbuf_hash_remove(dmu_buf_impl_t *db) 207{ 208 dbuf_hash_table_t *h = &dbuf_hash_table; 209 uint64_t hv = DBUF_HASH(db->db_objset, db->db.db_object, 210 db->db_level, db->db_blkid); 211 uint64_t idx = hv & h->hash_table_mask; 212 dmu_buf_impl_t *dbf, **dbp; 213 214 /* 215 * We musn't hold db_mtx to maintain lock ordering: 216 * DBUF_HASH_MUTEX > db_mtx. 217 */ 218 ASSERT(refcount_is_zero(&db->db_holds)); 219 ASSERT(db->db_state == DB_EVICTING); 220 ASSERT(!MUTEX_HELD(&db->db_mtx)); 221 222 mutex_enter(DBUF_HASH_MUTEX(h, idx)); 223 dbp = &h->hash_table[idx]; 224 while ((dbf = *dbp) != db) { 225 dbp = &dbf->db_hash_next; 226 ASSERT(dbf != NULL); 227 } 228 *dbp = db->db_hash_next; 229 db->db_hash_next = NULL; 230 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 231 atomic_dec_64(&dbuf_hash_count); 232} 233 234static arc_evict_func_t dbuf_do_evict; 235 236typedef enum { 237 DBVU_EVICTING, 238 DBVU_NOT_EVICTING 239} dbvu_verify_type_t; 240 241static void 242dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type) 243{ 244#ifdef ZFS_DEBUG 245 int64_t holds; 246 247 if (db->db_user == NULL) 248 return; 249 250 /* Only data blocks support the attachment of user data. */ 251 ASSERT(db->db_level == 0); 252 253 /* Clients must resolve a dbuf before attaching user data. */ 254 ASSERT(db->db.db_data != NULL); 255 ASSERT3U(db->db_state, ==, DB_CACHED); 256 257 holds = refcount_count(&db->db_holds); 258 if (verify_type == DBVU_EVICTING) { 259 /* 260 * Immediate eviction occurs when holds == dirtycnt. 261 * For normal eviction buffers, holds is zero on 262 * eviction, except when dbuf_fix_old_data() calls 263 * dbuf_clear_data(). However, the hold count can grow 264 * during eviction even though db_mtx is held (see 265 * dmu_bonus_hold() for an example), so we can only 266 * test the generic invariant that holds >= dirtycnt. 267 */ 268 ASSERT3U(holds, >=, db->db_dirtycnt); 269 } else { 270 if (db->db_immediate_evict == TRUE) 271 ASSERT3U(holds, >=, db->db_dirtycnt); 272 else 273 ASSERT3U(holds, >, 0); 274 } 275#endif 276} 277 278static void 279dbuf_evict_user(dmu_buf_impl_t *db) 280{ 281 dmu_buf_user_t *dbu = db->db_user; 282 283 ASSERT(MUTEX_HELD(&db->db_mtx)); 284 285 if (dbu == NULL) 286 return; 287 288 dbuf_verify_user(db, DBVU_EVICTING); 289 db->db_user = NULL; 290 291#ifdef ZFS_DEBUG 292 if (dbu->dbu_clear_on_evict_dbufp != NULL) 293 *dbu->dbu_clear_on_evict_dbufp = NULL; 294#endif 295 296 /* 297 * Invoke the callback from a taskq to avoid lock order reversals 298 * and limit stack depth. 299 */ 300 taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func, dbu, 0, 301 &dbu->dbu_tqent); 302} 303 304boolean_t 305dbuf_is_metadata(dmu_buf_impl_t *db) 306{ 307 if (db->db_level > 0) { 308 return (B_TRUE); 309 } else { 310 boolean_t is_metadata; 311 312 DB_DNODE_ENTER(db); 313 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type); 314 DB_DNODE_EXIT(db); 315 316 return (is_metadata); 317 } 318} 319 320void 321dbuf_evict(dmu_buf_impl_t *db) 322{ 323 ASSERT(MUTEX_HELD(&db->db_mtx)); 324 ASSERT(db->db_buf == NULL); 325 ASSERT(db->db_data_pending == NULL); 326 327 dbuf_clear(db); 328 dbuf_destroy(db); 329} 330 331void 332dbuf_init(void) 333{ 334 uint64_t hsize = 1ULL << 16; 335 dbuf_hash_table_t *h = &dbuf_hash_table; 336 int i; 337 338 /* 339 * The hash table is big enough to fill all of physical memory 340 * with an average 4K block size. The table will take up 341 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers). 342 */ 343 while (hsize * 4096 < (uint64_t)physmem * PAGESIZE) 344 hsize <<= 1; 345 346retry: 347 h->hash_table_mask = hsize - 1; 348 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP); 349 if (h->hash_table == NULL) { 350 /* XXX - we should really return an error instead of assert */ 351 ASSERT(hsize > (1ULL << 10)); 352 hsize >>= 1; 353 goto retry; 354 } 355 356 dbuf_cache = kmem_cache_create("dmu_buf_impl_t", 357 sizeof (dmu_buf_impl_t), 358 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0); 359 360 for (i = 0; i < DBUF_MUTEXES; i++) 361 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL); 362 363 /* 364 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc 365 * configuration is not required. 366 */ 367 dbu_evict_taskq = taskq_create("dbu_evict", 1, minclsyspri, 0, 0, 0); 368} 369 370void 371dbuf_fini(void) 372{ 373 dbuf_hash_table_t *h = &dbuf_hash_table; 374 int i; 375 376 for (i = 0; i < DBUF_MUTEXES; i++) 377 mutex_destroy(&h->hash_mutexes[i]); 378 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *)); 379 kmem_cache_destroy(dbuf_cache); 380 taskq_destroy(dbu_evict_taskq); 381} 382 383/* 384 * Other stuff. 385 */ 386 387#ifdef ZFS_DEBUG 388static void 389dbuf_verify(dmu_buf_impl_t *db) 390{ 391 dnode_t *dn; 392 dbuf_dirty_record_t *dr; 393 394 ASSERT(MUTEX_HELD(&db->db_mtx)); 395 396 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY)) 397 return; 398 399 ASSERT(db->db_objset != NULL); 400 DB_DNODE_ENTER(db); 401 dn = DB_DNODE(db); 402 if (dn == NULL) { 403 ASSERT(db->db_parent == NULL); 404 ASSERT(db->db_blkptr == NULL); 405 } else { 406 ASSERT3U(db->db.db_object, ==, dn->dn_object); 407 ASSERT3P(db->db_objset, ==, dn->dn_objset); 408 ASSERT3U(db->db_level, <, dn->dn_nlevels); 409 ASSERT(db->db_blkid == DMU_BONUS_BLKID || 410 db->db_blkid == DMU_SPILL_BLKID || 411 !avl_is_empty(&dn->dn_dbufs)); 412 } 413 if (db->db_blkid == DMU_BONUS_BLKID) { 414 ASSERT(dn != NULL); 415 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 416 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID); 417 } else if (db->db_blkid == DMU_SPILL_BLKID) { 418 ASSERT(dn != NULL); 419 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 420 ASSERT0(db->db.db_offset); 421 } else { 422 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size); 423 } 424 425 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next) 426 ASSERT(dr->dr_dbuf == db); 427 428 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next) 429 ASSERT(dr->dr_dbuf == db); 430 431 /* 432 * We can't assert that db_size matches dn_datablksz because it 433 * can be momentarily different when another thread is doing 434 * dnode_set_blksz(). 435 */ 436 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) { 437 dr = db->db_data_pending; 438 /* 439 * It should only be modified in syncing context, so 440 * make sure we only have one copy of the data. 441 */ 442 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf); 443 } 444 445 /* verify db->db_blkptr */ 446 if (db->db_blkptr) { 447 if (db->db_parent == dn->dn_dbuf) { 448 /* db is pointed to by the dnode */ 449 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */ 450 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object)) 451 ASSERT(db->db_parent == NULL); 452 else 453 ASSERT(db->db_parent != NULL); 454 if (db->db_blkid != DMU_SPILL_BLKID) 455 ASSERT3P(db->db_blkptr, ==, 456 &dn->dn_phys->dn_blkptr[db->db_blkid]); 457 } else { 458 /* db is pointed to by an indirect block */ 459 int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT; 460 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1); 461 ASSERT3U(db->db_parent->db.db_object, ==, 462 db->db.db_object); 463 /* 464 * dnode_grow_indblksz() can make this fail if we don't 465 * have the struct_rwlock. XXX indblksz no longer 466 * grows. safe to do this now? 467 */ 468 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) { 469 ASSERT3P(db->db_blkptr, ==, 470 ((blkptr_t *)db->db_parent->db.db_data + 471 db->db_blkid % epb)); 472 } 473 } 474 } 475 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) && 476 (db->db_buf == NULL || db->db_buf->b_data) && 477 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID && 478 db->db_state != DB_FILL && !dn->dn_free_txg) { 479 /* 480 * If the blkptr isn't set but they have nonzero data, 481 * it had better be dirty, otherwise we'll lose that 482 * data when we evict this buffer. 483 */ 484 if (db->db_dirtycnt == 0) { 485 uint64_t *buf = db->db.db_data; 486 int i; 487 488 for (i = 0; i < db->db.db_size >> 3; i++) { 489 ASSERT(buf[i] == 0); 490 } 491 } 492 } 493 DB_DNODE_EXIT(db); 494} 495#endif 496 497static void 498dbuf_clear_data(dmu_buf_impl_t *db) 499{ 500 ASSERT(MUTEX_HELD(&db->db_mtx)); 501 dbuf_evict_user(db); 502 db->db_buf = NULL; 503 db->db.db_data = NULL; 504 if (db->db_state != DB_NOFILL) 505 db->db_state = DB_UNCACHED; 506} 507 508static void 509dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf) 510{ 511 ASSERT(MUTEX_HELD(&db->db_mtx)); 512 ASSERT(buf != NULL); 513 514 db->db_buf = buf; 515 ASSERT(buf->b_data != NULL); 516 db->db.db_data = buf->b_data; 517 if (!arc_released(buf)) 518 arc_set_callback(buf, dbuf_do_evict, db); 519} 520 521/* 522 * Loan out an arc_buf for read. Return the loaned arc_buf. 523 */ 524arc_buf_t * 525dbuf_loan_arcbuf(dmu_buf_impl_t *db) 526{ 527 arc_buf_t *abuf; 528 529 mutex_enter(&db->db_mtx); 530 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) { 531 int blksz = db->db.db_size; 532 spa_t *spa = db->db_objset->os_spa; 533 534 mutex_exit(&db->db_mtx); 535 abuf = arc_loan_buf(spa, blksz); 536 bcopy(db->db.db_data, abuf->b_data, blksz); 537 } else { 538 abuf = db->db_buf; 539 arc_loan_inuse_buf(abuf, db); 540 dbuf_clear_data(db); 541 mutex_exit(&db->db_mtx); 542 } 543 return (abuf); 544} 545 546/* 547 * Calculate which level n block references the data at the level 0 offset 548 * provided. 549 */ 550uint64_t 551dbuf_whichblock(dnode_t *dn, int64_t level, uint64_t offset) 552{ 553 if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) { 554 /* 555 * The level n blkid is equal to the level 0 blkid divided by 556 * the number of level 0s in a level n block. 557 * 558 * The level 0 blkid is offset >> datablkshift = 559 * offset / 2^datablkshift. 560 * 561 * The number of level 0s in a level n is the number of block 562 * pointers in an indirect block, raised to the power of level. 563 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level = 564 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)). 565 * 566 * Thus, the level n blkid is: offset / 567 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT))) 568 * = offset / 2^(datablkshift + level * 569 * (indblkshift - SPA_BLKPTRSHIFT)) 570 * = offset >> (datablkshift + level * 571 * (indblkshift - SPA_BLKPTRSHIFT)) 572 */ 573 return (offset >> (dn->dn_datablkshift + level * 574 (dn->dn_indblkshift - SPA_BLKPTRSHIFT))); 575 } else { 576 ASSERT3U(offset, <, dn->dn_datablksz); 577 return (0); 578 } 579} 580 581static void 582dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb) 583{ 584 dmu_buf_impl_t *db = vdb; 585 586 mutex_enter(&db->db_mtx); 587 ASSERT3U(db->db_state, ==, DB_READ); 588 /* 589 * All reads are synchronous, so we must have a hold on the dbuf 590 */ 591 ASSERT(refcount_count(&db->db_holds) > 0); 592 ASSERT(db->db_buf == NULL); 593 ASSERT(db->db.db_data == NULL); 594 if (db->db_level == 0 && db->db_freed_in_flight) { 595 /* we were freed in flight; disregard any error */ 596 arc_release(buf, db); 597 bzero(buf->b_data, db->db.db_size); 598 arc_buf_freeze(buf); 599 db->db_freed_in_flight = FALSE; 600 dbuf_set_data(db, buf); 601 db->db_state = DB_CACHED; 602 } else if (zio == NULL || zio->io_error == 0) { 603 dbuf_set_data(db, buf); 604 db->db_state = DB_CACHED; 605 } else { 606 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 607 ASSERT3P(db->db_buf, ==, NULL); 608 VERIFY(arc_buf_remove_ref(buf, db)); 609 db->db_state = DB_UNCACHED; 610 } 611 cv_broadcast(&db->db_changed); 612 dbuf_rele_and_unlock(db, NULL); 613} 614 615static void 616dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) 617{ 618 dnode_t *dn; 619 zbookmark_phys_t zb; 620 arc_flags_t aflags = ARC_FLAG_NOWAIT; 621 622 DB_DNODE_ENTER(db); 623 dn = DB_DNODE(db); 624 ASSERT(!refcount_is_zero(&db->db_holds)); 625 /* We need the struct_rwlock to prevent db_blkptr from changing. */ 626 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 627 ASSERT(MUTEX_HELD(&db->db_mtx)); 628 ASSERT(db->db_state == DB_UNCACHED); 629 ASSERT(db->db_buf == NULL); 630 631 if (db->db_blkid == DMU_BONUS_BLKID) { 632 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen); 633 634 ASSERT3U(bonuslen, <=, db->db.db_size); 635 db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN); 636 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 637 if (bonuslen < DN_MAX_BONUSLEN) 638 bzero(db->db.db_data, DN_MAX_BONUSLEN); 639 if (bonuslen) 640 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen); 641 DB_DNODE_EXIT(db); 642 db->db_state = DB_CACHED; 643 mutex_exit(&db->db_mtx); 644 return; 645 } 646 647 /* 648 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync() 649 * processes the delete record and clears the bp while we are waiting 650 * for the dn_mtx (resulting in a "no" from block_freed). 651 */ 652 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) || 653 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) || 654 BP_IS_HOLE(db->db_blkptr)))) { 655 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 656 657 DB_DNODE_EXIT(db); 658 dbuf_set_data(db, arc_buf_alloc(db->db_objset->os_spa, 659 db->db.db_size, db, type)); 660 bzero(db->db.db_data, db->db.db_size); 661 db->db_state = DB_CACHED; 662 mutex_exit(&db->db_mtx); 663 return; 664 } 665 666 DB_DNODE_EXIT(db); 667 668 db->db_state = DB_READ; 669 mutex_exit(&db->db_mtx); 670 671 if (DBUF_IS_L2CACHEABLE(db)) 672 aflags |= ARC_FLAG_L2CACHE; 673 if (DBUF_IS_L2COMPRESSIBLE(db)) 674 aflags |= ARC_FLAG_L2COMPRESS; 675 676 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ? 677 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET, 678 db->db.db_object, db->db_level, db->db_blkid); 679 680 dbuf_add_ref(db, NULL); 681 682 (void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr, 683 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ, 684 (flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED, 685 &aflags, &zb); 686} 687 688int 689dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) 690{ 691 int err = 0; 692 boolean_t havepzio = (zio != NULL); 693 boolean_t prefetch; 694 dnode_t *dn; 695 696 /* 697 * We don't have to hold the mutex to check db_state because it 698 * can't be freed while we have a hold on the buffer. 699 */ 700 ASSERT(!refcount_is_zero(&db->db_holds)); 701 702 if (db->db_state == DB_NOFILL) 703 return (SET_ERROR(EIO)); 704 705 DB_DNODE_ENTER(db); 706 dn = DB_DNODE(db); 707 if ((flags & DB_RF_HAVESTRUCT) == 0) 708 rw_enter(&dn->dn_struct_rwlock, RW_READER); 709 710 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 711 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL && 712 DBUF_IS_CACHEABLE(db); 713 714 mutex_enter(&db->db_mtx); 715 if (db->db_state == DB_CACHED) { 716 mutex_exit(&db->db_mtx); 717 if (prefetch) 718 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1); 719 if ((flags & DB_RF_HAVESTRUCT) == 0) 720 rw_exit(&dn->dn_struct_rwlock); 721 DB_DNODE_EXIT(db); 722 } else if (db->db_state == DB_UNCACHED) { 723 spa_t *spa = dn->dn_objset->os_spa; 724 725 if (zio == NULL) 726 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); 727 dbuf_read_impl(db, zio, flags); 728 729 /* dbuf_read_impl has dropped db_mtx for us */ 730 731 if (prefetch) 732 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1); 733 734 if ((flags & DB_RF_HAVESTRUCT) == 0) 735 rw_exit(&dn->dn_struct_rwlock); 736 DB_DNODE_EXIT(db); 737 738 if (!havepzio) 739 err = zio_wait(zio); 740 } else { 741 /* 742 * Another reader came in while the dbuf was in flight 743 * between UNCACHED and CACHED. Either a writer will finish 744 * writing the buffer (sending the dbuf to CACHED) or the 745 * first reader's request will reach the read_done callback 746 * and send the dbuf to CACHED. Otherwise, a failure 747 * occurred and the dbuf went to UNCACHED. 748 */ 749 mutex_exit(&db->db_mtx); 750 if (prefetch) 751 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1); 752 if ((flags & DB_RF_HAVESTRUCT) == 0) 753 rw_exit(&dn->dn_struct_rwlock); 754 DB_DNODE_EXIT(db); 755 756 /* Skip the wait per the caller's request. */ 757 mutex_enter(&db->db_mtx); 758 if ((flags & DB_RF_NEVERWAIT) == 0) { 759 while (db->db_state == DB_READ || 760 db->db_state == DB_FILL) { 761 ASSERT(db->db_state == DB_READ || 762 (flags & DB_RF_HAVESTRUCT) == 0); 763 DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *, 764 db, zio_t *, zio); 765 cv_wait(&db->db_changed, &db->db_mtx); 766 } 767 if (db->db_state == DB_UNCACHED) 768 err = SET_ERROR(EIO); 769 } 770 mutex_exit(&db->db_mtx); 771 } 772 773 ASSERT(err || havepzio || db->db_state == DB_CACHED); 774 return (err); 775} 776 777static void 778dbuf_noread(dmu_buf_impl_t *db) 779{ 780 ASSERT(!refcount_is_zero(&db->db_holds)); 781 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 782 mutex_enter(&db->db_mtx); 783 while (db->db_state == DB_READ || db->db_state == DB_FILL) 784 cv_wait(&db->db_changed, &db->db_mtx); 785 if (db->db_state == DB_UNCACHED) { 786 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 787 spa_t *spa = db->db_objset->os_spa; 788 789 ASSERT(db->db_buf == NULL); 790 ASSERT(db->db.db_data == NULL); 791 dbuf_set_data(db, arc_buf_alloc(spa, db->db.db_size, db, type)); 792 db->db_state = DB_FILL; 793 } else if (db->db_state == DB_NOFILL) { 794 dbuf_clear_data(db); 795 } else { 796 ASSERT3U(db->db_state, ==, DB_CACHED); 797 } 798 mutex_exit(&db->db_mtx); 799} 800 801/* 802 * This is our just-in-time copy function. It makes a copy of 803 * buffers, that have been modified in a previous transaction 804 * group, before we modify them in the current active group. 805 * 806 * This function is used in two places: when we are dirtying a 807 * buffer for the first time in a txg, and when we are freeing 808 * a range in a dnode that includes this buffer. 809 * 810 * Note that when we are called from dbuf_free_range() we do 811 * not put a hold on the buffer, we just traverse the active 812 * dbuf list for the dnode. 813 */ 814static void 815dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg) 816{ 817 dbuf_dirty_record_t *dr = db->db_last_dirty; 818 819 ASSERT(MUTEX_HELD(&db->db_mtx)); 820 ASSERT(db->db.db_data != NULL); 821 ASSERT(db->db_level == 0); 822 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT); 823 824 if (dr == NULL || 825 (dr->dt.dl.dr_data != 826 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf))) 827 return; 828 829 /* 830 * If the last dirty record for this dbuf has not yet synced 831 * and its referencing the dbuf data, either: 832 * reset the reference to point to a new copy, 833 * or (if there a no active holders) 834 * just null out the current db_data pointer. 835 */ 836 ASSERT(dr->dr_txg >= txg - 2); 837 if (db->db_blkid == DMU_BONUS_BLKID) { 838 /* Note that the data bufs here are zio_bufs */ 839 dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN); 840 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 841 bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN); 842 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) { 843 int size = db->db.db_size; 844 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 845 spa_t *spa = db->db_objset->os_spa; 846 847 dr->dt.dl.dr_data = arc_buf_alloc(spa, size, db, type); 848 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size); 849 } else { 850 dbuf_clear_data(db); 851 } 852} 853 854void 855dbuf_unoverride(dbuf_dirty_record_t *dr) 856{ 857 dmu_buf_impl_t *db = dr->dr_dbuf; 858 blkptr_t *bp = &dr->dt.dl.dr_overridden_by; 859 uint64_t txg = dr->dr_txg; 860 861 ASSERT(MUTEX_HELD(&db->db_mtx)); 862 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC); 863 ASSERT(db->db_level == 0); 864 865 if (db->db_blkid == DMU_BONUS_BLKID || 866 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN) 867 return; 868 869 ASSERT(db->db_data_pending != dr); 870 871 /* free this block */ 872 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite) 873 zio_free(db->db_objset->os_spa, txg, bp); 874 875 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 876 dr->dt.dl.dr_nopwrite = B_FALSE; 877 878 /* 879 * Release the already-written buffer, so we leave it in 880 * a consistent dirty state. Note that all callers are 881 * modifying the buffer, so they will immediately do 882 * another (redundant) arc_release(). Therefore, leave 883 * the buf thawed to save the effort of freezing & 884 * immediately re-thawing it. 885 */ 886 arc_release(dr->dt.dl.dr_data, db); 887} 888 889/* 890 * Evict (if its unreferenced) or clear (if its referenced) any level-0 891 * data blocks in the free range, so that any future readers will find 892 * empty blocks. 893 * 894 * This is a no-op if the dataset is in the middle of an incremental 895 * receive; see comment below for details. 896 */ 897void 898dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid, 899 dmu_tx_t *tx) 900{ 901 dmu_buf_impl_t db_search; 902 dmu_buf_impl_t *db, *db_next; 903 uint64_t txg = tx->tx_txg; 904 avl_index_t where; 905 906 if (end_blkid > dn->dn_maxblkid && (end_blkid != DMU_SPILL_BLKID)) 907 end_blkid = dn->dn_maxblkid; 908 dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid); 909 910 db_search.db_level = 0; 911 db_search.db_blkid = start_blkid; 912 db_search.db_state = DB_SEARCH; 913 914 mutex_enter(&dn->dn_dbufs_mtx); 915 if (start_blkid >= dn->dn_unlisted_l0_blkid) { 916 /* There can't be any dbufs in this range; no need to search. */ 917#ifdef DEBUG 918 db = avl_find(&dn->dn_dbufs, &db_search, &where); 919 ASSERT3P(db, ==, NULL); 920 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER); 921 ASSERT(db == NULL || db->db_level > 0); 922#endif 923 mutex_exit(&dn->dn_dbufs_mtx); 924 return; 925 } else if (dmu_objset_is_receiving(dn->dn_objset)) { 926 /* 927 * If we are receiving, we expect there to be no dbufs in 928 * the range to be freed, because receive modifies each 929 * block at most once, and in offset order. If this is 930 * not the case, it can lead to performance problems, 931 * so note that we unexpectedly took the slow path. 932 */ 933 atomic_inc_64(&zfs_free_range_recv_miss); 934 } 935 936 db = avl_find(&dn->dn_dbufs, &db_search, &where); 937 ASSERT3P(db, ==, NULL); 938 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER); 939 940 for (; db != NULL; db = db_next) { 941 db_next = AVL_NEXT(&dn->dn_dbufs, db); 942 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 943 944 if (db->db_level != 0 || db->db_blkid > end_blkid) { 945 break; 946 } 947 ASSERT3U(db->db_blkid, >=, start_blkid); 948 949 /* found a level 0 buffer in the range */ 950 mutex_enter(&db->db_mtx); 951 if (dbuf_undirty(db, tx)) { 952 /* mutex has been dropped and dbuf destroyed */ 953 continue; 954 } 955 956 if (db->db_state == DB_UNCACHED || 957 db->db_state == DB_NOFILL || 958 db->db_state == DB_EVICTING) { 959 ASSERT(db->db.db_data == NULL); 960 mutex_exit(&db->db_mtx); 961 continue; 962 } 963 if (db->db_state == DB_READ || db->db_state == DB_FILL) { 964 /* will be handled in dbuf_read_done or dbuf_rele */ 965 db->db_freed_in_flight = TRUE; 966 mutex_exit(&db->db_mtx); 967 continue; 968 } 969 if (refcount_count(&db->db_holds) == 0) { 970 ASSERT(db->db_buf); 971 dbuf_clear(db); 972 continue; 973 } 974 /* The dbuf is referenced */ 975 976 if (db->db_last_dirty != NULL) { 977 dbuf_dirty_record_t *dr = db->db_last_dirty; 978 979 if (dr->dr_txg == txg) { 980 /* 981 * This buffer is "in-use", re-adjust the file 982 * size to reflect that this buffer may 983 * contain new data when we sync. 984 */ 985 if (db->db_blkid != DMU_SPILL_BLKID && 986 db->db_blkid > dn->dn_maxblkid) 987 dn->dn_maxblkid = db->db_blkid; 988 dbuf_unoverride(dr); 989 } else { 990 /* 991 * This dbuf is not dirty in the open context. 992 * Either uncache it (if its not referenced in 993 * the open context) or reset its contents to 994 * empty. 995 */ 996 dbuf_fix_old_data(db, txg); 997 } 998 } 999 /* clear the contents if its cached */ 1000 if (db->db_state == DB_CACHED) { 1001 ASSERT(db->db.db_data != NULL); 1002 arc_release(db->db_buf, db); 1003 bzero(db->db.db_data, db->db.db_size); 1004 arc_buf_freeze(db->db_buf); 1005 } 1006 1007 mutex_exit(&db->db_mtx); 1008 } 1009 mutex_exit(&dn->dn_dbufs_mtx); 1010} 1011 1012static int 1013dbuf_block_freeable(dmu_buf_impl_t *db) 1014{ 1015 dsl_dataset_t *ds = db->db_objset->os_dsl_dataset; 1016 uint64_t birth_txg = 0; 1017 1018 /* 1019 * We don't need any locking to protect db_blkptr: 1020 * If it's syncing, then db_last_dirty will be set 1021 * so we'll ignore db_blkptr. 1022 * 1023 * This logic ensures that only block births for 1024 * filled blocks are considered. 1025 */ 1026 ASSERT(MUTEX_HELD(&db->db_mtx)); 1027 if (db->db_last_dirty && (db->db_blkptr == NULL || 1028 !BP_IS_HOLE(db->db_blkptr))) { 1029 birth_txg = db->db_last_dirty->dr_txg; 1030 } else if (db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) { 1031 birth_txg = db->db_blkptr->blk_birth; 1032 } 1033 1034 /* 1035 * If this block don't exist or is in a snapshot, it can't be freed. 1036 * Don't pass the bp to dsl_dataset_block_freeable() since we 1037 * are holding the db_mtx lock and might deadlock if we are 1038 * prefetching a dedup-ed block. 1039 */ 1040 if (birth_txg != 0) 1041 return (ds == NULL || 1042 dsl_dataset_block_freeable(ds, NULL, birth_txg)); 1043 else 1044 return (B_FALSE); 1045} 1046 1047void 1048dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx) 1049{ 1050 arc_buf_t *buf, *obuf; 1051 int osize = db->db.db_size; 1052 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 1053 dnode_t *dn; 1054 1055 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1056 1057 DB_DNODE_ENTER(db); 1058 dn = DB_DNODE(db); 1059 1060 /* XXX does *this* func really need the lock? */ 1061 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 1062 1063 /* 1064 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held 1065 * is OK, because there can be no other references to the db 1066 * when we are changing its size, so no concurrent DB_FILL can 1067 * be happening. 1068 */ 1069 /* 1070 * XXX we should be doing a dbuf_read, checking the return 1071 * value and returning that up to our callers 1072 */ 1073 dmu_buf_will_dirty(&db->db, tx); 1074 1075 /* create the data buffer for the new block */ 1076 buf = arc_buf_alloc(dn->dn_objset->os_spa, size, db, type); 1077 1078 /* copy old block data to the new block */ 1079 obuf = db->db_buf; 1080 bcopy(obuf->b_data, buf->b_data, MIN(osize, size)); 1081 /* zero the remainder */ 1082 if (size > osize) 1083 bzero((uint8_t *)buf->b_data + osize, size - osize); 1084 1085 mutex_enter(&db->db_mtx); 1086 dbuf_set_data(db, buf); 1087 VERIFY(arc_buf_remove_ref(obuf, db)); 1088 db->db.db_size = size; 1089 1090 if (db->db_level == 0) { 1091 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); 1092 db->db_last_dirty->dt.dl.dr_data = buf; 1093 } 1094 mutex_exit(&db->db_mtx); 1095 1096 dnode_willuse_space(dn, size-osize, tx); 1097 DB_DNODE_EXIT(db); 1098} 1099 1100void 1101dbuf_release_bp(dmu_buf_impl_t *db) 1102{ 1103 objset_t *os = db->db_objset; 1104 1105 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os))); 1106 ASSERT(arc_released(os->os_phys_buf) || 1107 list_link_active(&os->os_dsl_dataset->ds_synced_link)); 1108 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf)); 1109 1110 (void) arc_release(db->db_buf, db); 1111} 1112 1113/* 1114 * We already have a dirty record for this TXG, and we are being 1115 * dirtied again. 1116 */ 1117static void 1118dbuf_redirty(dbuf_dirty_record_t *dr) 1119{ 1120 dmu_buf_impl_t *db = dr->dr_dbuf; 1121 1122 ASSERT(MUTEX_HELD(&db->db_mtx)); 1123 1124 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) { 1125 /* 1126 * If this buffer has already been written out, 1127 * we now need to reset its state. 1128 */ 1129 dbuf_unoverride(dr); 1130 if (db->db.db_object != DMU_META_DNODE_OBJECT && 1131 db->db_state != DB_NOFILL) { 1132 /* Already released on initial dirty, so just thaw. */ 1133 ASSERT(arc_released(db->db_buf)); 1134 arc_buf_thaw(db->db_buf); 1135 } 1136 } 1137} 1138 1139dbuf_dirty_record_t * 1140dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1141{ 1142 dnode_t *dn; 1143 objset_t *os; 1144 dbuf_dirty_record_t **drp, *dr; 1145 int drop_struct_lock = FALSE; 1146 boolean_t do_free_accounting = B_FALSE; 1147 int txgoff = tx->tx_txg & TXG_MASK; 1148 1149 ASSERT(tx->tx_txg != 0); 1150 ASSERT(!refcount_is_zero(&db->db_holds)); 1151 DMU_TX_DIRTY_BUF(tx, db); 1152 1153 DB_DNODE_ENTER(db); 1154 dn = DB_DNODE(db); 1155 /* 1156 * Shouldn't dirty a regular buffer in syncing context. Private 1157 * objects may be dirtied in syncing context, but only if they 1158 * were already pre-dirtied in open context. 1159 */ 1160 ASSERT(!dmu_tx_is_syncing(tx) || 1161 BP_IS_HOLE(dn->dn_objset->os_rootbp) || 1162 DMU_OBJECT_IS_SPECIAL(dn->dn_object) || 1163 dn->dn_objset->os_dsl_dataset == NULL); 1164 /* 1165 * We make this assert for private objects as well, but after we 1166 * check if we're already dirty. They are allowed to re-dirty 1167 * in syncing context. 1168 */ 1169 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 1170 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 1171 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 1172 1173 mutex_enter(&db->db_mtx); 1174 /* 1175 * XXX make this true for indirects too? The problem is that 1176 * transactions created with dmu_tx_create_assigned() from 1177 * syncing context don't bother holding ahead. 1178 */ 1179 ASSERT(db->db_level != 0 || 1180 db->db_state == DB_CACHED || db->db_state == DB_FILL || 1181 db->db_state == DB_NOFILL); 1182 1183 mutex_enter(&dn->dn_mtx); 1184 /* 1185 * Don't set dirtyctx to SYNC if we're just modifying this as we 1186 * initialize the objset. 1187 */ 1188 if (dn->dn_dirtyctx == DN_UNDIRTIED && 1189 !BP_IS_HOLE(dn->dn_objset->os_rootbp)) { 1190 dn->dn_dirtyctx = 1191 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN); 1192 ASSERT(dn->dn_dirtyctx_firstset == NULL); 1193 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP); 1194 } 1195 mutex_exit(&dn->dn_mtx); 1196 1197 if (db->db_blkid == DMU_SPILL_BLKID) 1198 dn->dn_have_spill = B_TRUE; 1199 1200 /* 1201 * If this buffer is already dirty, we're done. 1202 */ 1203 drp = &db->db_last_dirty; 1204 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg || 1205 db->db.db_object == DMU_META_DNODE_OBJECT); 1206 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg) 1207 drp = &dr->dr_next; 1208 if (dr && dr->dr_txg == tx->tx_txg) { 1209 DB_DNODE_EXIT(db); 1210 1211 dbuf_redirty(dr); 1212 mutex_exit(&db->db_mtx); 1213 return (dr); 1214 } 1215 1216 /* 1217 * Only valid if not already dirty. 1218 */ 1219 ASSERT(dn->dn_object == 0 || 1220 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 1221 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 1222 1223 ASSERT3U(dn->dn_nlevels, >, db->db_level); 1224 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) || 1225 dn->dn_phys->dn_nlevels > db->db_level || 1226 dn->dn_next_nlevels[txgoff] > db->db_level || 1227 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level || 1228 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level); 1229 1230 /* 1231 * We should only be dirtying in syncing context if it's the 1232 * mos or we're initializing the os or it's a special object. 1233 * However, we are allowed to dirty in syncing context provided 1234 * we already dirtied it in open context. Hence we must make 1235 * this assertion only if we're not already dirty. 1236 */ 1237 os = dn->dn_objset; 1238 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) || 1239 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp)); 1240 ASSERT(db->db.db_size != 0); 1241 1242 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 1243 1244 if (db->db_blkid != DMU_BONUS_BLKID) { 1245 /* 1246 * Update the accounting. 1247 * Note: we delay "free accounting" until after we drop 1248 * the db_mtx. This keeps us from grabbing other locks 1249 * (and possibly deadlocking) in bp_get_dsize() while 1250 * also holding the db_mtx. 1251 */ 1252 dnode_willuse_space(dn, db->db.db_size, tx); 1253 do_free_accounting = dbuf_block_freeable(db); 1254 } 1255 1256 /* 1257 * If this buffer is dirty in an old transaction group we need 1258 * to make a copy of it so that the changes we make in this 1259 * transaction group won't leak out when we sync the older txg. 1260 */ 1261 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP); 1262 if (db->db_level == 0) { 1263 void *data_old = db->db_buf; 1264 1265 if (db->db_state != DB_NOFILL) { 1266 if (db->db_blkid == DMU_BONUS_BLKID) { 1267 dbuf_fix_old_data(db, tx->tx_txg); 1268 data_old = db->db.db_data; 1269 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) { 1270 /* 1271 * Release the data buffer from the cache so 1272 * that we can modify it without impacting 1273 * possible other users of this cached data 1274 * block. Note that indirect blocks and 1275 * private objects are not released until the 1276 * syncing state (since they are only modified 1277 * then). 1278 */ 1279 arc_release(db->db_buf, db); 1280 dbuf_fix_old_data(db, tx->tx_txg); 1281 data_old = db->db_buf; 1282 } 1283 ASSERT(data_old != NULL); 1284 } 1285 dr->dt.dl.dr_data = data_old; 1286 } else { 1287 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL); 1288 list_create(&dr->dt.di.dr_children, 1289 sizeof (dbuf_dirty_record_t), 1290 offsetof(dbuf_dirty_record_t, dr_dirty_node)); 1291 } 1292 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL) 1293 dr->dr_accounted = db->db.db_size; 1294 dr->dr_dbuf = db; 1295 dr->dr_txg = tx->tx_txg; 1296 dr->dr_next = *drp; 1297 *drp = dr; 1298 1299 /* 1300 * We could have been freed_in_flight between the dbuf_noread 1301 * and dbuf_dirty. We win, as though the dbuf_noread() had 1302 * happened after the free. 1303 */ 1304 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 1305 db->db_blkid != DMU_SPILL_BLKID) { 1306 mutex_enter(&dn->dn_mtx); 1307 if (dn->dn_free_ranges[txgoff] != NULL) { 1308 range_tree_clear(dn->dn_free_ranges[txgoff], 1309 db->db_blkid, 1); 1310 } 1311 mutex_exit(&dn->dn_mtx); 1312 db->db_freed_in_flight = FALSE; 1313 } 1314 1315 /* 1316 * This buffer is now part of this txg 1317 */ 1318 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg); 1319 db->db_dirtycnt += 1; 1320 ASSERT3U(db->db_dirtycnt, <=, 3); 1321 1322 mutex_exit(&db->db_mtx); 1323 1324 if (db->db_blkid == DMU_BONUS_BLKID || 1325 db->db_blkid == DMU_SPILL_BLKID) { 1326 mutex_enter(&dn->dn_mtx); 1327 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1328 list_insert_tail(&dn->dn_dirty_records[txgoff], dr); 1329 mutex_exit(&dn->dn_mtx); 1330 dnode_setdirty(dn, tx); 1331 DB_DNODE_EXIT(db); 1332 return (dr); 1333 } else if (do_free_accounting) { 1334 blkptr_t *bp = db->db_blkptr; 1335 int64_t willfree = (bp && !BP_IS_HOLE(bp)) ? 1336 bp_get_dsize(os->os_spa, bp) : db->db.db_size; 1337 /* 1338 * This is only a guess -- if the dbuf is dirty 1339 * in a previous txg, we don't know how much 1340 * space it will use on disk yet. We should 1341 * really have the struct_rwlock to access 1342 * db_blkptr, but since this is just a guess, 1343 * it's OK if we get an odd answer. 1344 */ 1345 ddt_prefetch(os->os_spa, bp); 1346 dnode_willuse_space(dn, -willfree, tx); 1347 } 1348 1349 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) { 1350 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1351 drop_struct_lock = TRUE; 1352 } 1353 1354 if (db->db_level == 0) { 1355 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock); 1356 ASSERT(dn->dn_maxblkid >= db->db_blkid); 1357 } 1358 1359 if (db->db_level+1 < dn->dn_nlevels) { 1360 dmu_buf_impl_t *parent = db->db_parent; 1361 dbuf_dirty_record_t *di; 1362 int parent_held = FALSE; 1363 1364 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) { 1365 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1366 1367 parent = dbuf_hold_level(dn, db->db_level+1, 1368 db->db_blkid >> epbs, FTAG); 1369 ASSERT(parent != NULL); 1370 parent_held = TRUE; 1371 } 1372 if (drop_struct_lock) 1373 rw_exit(&dn->dn_struct_rwlock); 1374 ASSERT3U(db->db_level+1, ==, parent->db_level); 1375 di = dbuf_dirty(parent, tx); 1376 if (parent_held) 1377 dbuf_rele(parent, FTAG); 1378 1379 mutex_enter(&db->db_mtx); 1380 /* 1381 * Since we've dropped the mutex, it's possible that 1382 * dbuf_undirty() might have changed this out from under us. 1383 */ 1384 if (db->db_last_dirty == dr || 1385 dn->dn_object == DMU_META_DNODE_OBJECT) { 1386 mutex_enter(&di->dt.di.dr_mtx); 1387 ASSERT3U(di->dr_txg, ==, tx->tx_txg); 1388 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1389 list_insert_tail(&di->dt.di.dr_children, dr); 1390 mutex_exit(&di->dt.di.dr_mtx); 1391 dr->dr_parent = di; 1392 } 1393 mutex_exit(&db->db_mtx); 1394 } else { 1395 ASSERT(db->db_level+1 == dn->dn_nlevels); 1396 ASSERT(db->db_blkid < dn->dn_nblkptr); 1397 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf); 1398 mutex_enter(&dn->dn_mtx); 1399 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1400 list_insert_tail(&dn->dn_dirty_records[txgoff], dr); 1401 mutex_exit(&dn->dn_mtx); 1402 if (drop_struct_lock) 1403 rw_exit(&dn->dn_struct_rwlock); 1404 } 1405 1406 dnode_setdirty(dn, tx); 1407 DB_DNODE_EXIT(db); 1408 return (dr); 1409} 1410 1411/* 1412 * Undirty a buffer in the transaction group referenced by the given 1413 * transaction. Return whether this evicted the dbuf. 1414 */ 1415static boolean_t 1416dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1417{ 1418 dnode_t *dn; 1419 uint64_t txg = tx->tx_txg; 1420 dbuf_dirty_record_t *dr, **drp; 1421 1422 ASSERT(txg != 0); 1423 1424 /* 1425 * Due to our use of dn_nlevels below, this can only be called 1426 * in open context, unless we are operating on the MOS. 1427 * From syncing context, dn_nlevels may be different from the 1428 * dn_nlevels used when dbuf was dirtied. 1429 */ 1430 ASSERT(db->db_objset == 1431 dmu_objset_pool(db->db_objset)->dp_meta_objset || 1432 txg != spa_syncing_txg(dmu_objset_spa(db->db_objset))); 1433 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1434 ASSERT0(db->db_level); 1435 ASSERT(MUTEX_HELD(&db->db_mtx)); 1436 1437 /* 1438 * If this buffer is not dirty, we're done. 1439 */ 1440 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next) 1441 if (dr->dr_txg <= txg) 1442 break; 1443 if (dr == NULL || dr->dr_txg < txg) 1444 return (B_FALSE); 1445 ASSERT(dr->dr_txg == txg); 1446 ASSERT(dr->dr_dbuf == db); 1447 1448 DB_DNODE_ENTER(db); 1449 dn = DB_DNODE(db); 1450 1451 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 1452 1453 ASSERT(db->db.db_size != 0); 1454 1455 dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset), 1456 dr->dr_accounted, txg); 1457 1458 *drp = dr->dr_next; 1459 1460 /* 1461 * Note that there are three places in dbuf_dirty() 1462 * where this dirty record may be put on a list. 1463 * Make sure to do a list_remove corresponding to 1464 * every one of those list_insert calls. 1465 */ 1466 if (dr->dr_parent) { 1467 mutex_enter(&dr->dr_parent->dt.di.dr_mtx); 1468 list_remove(&dr->dr_parent->dt.di.dr_children, dr); 1469 mutex_exit(&dr->dr_parent->dt.di.dr_mtx); 1470 } else if (db->db_blkid == DMU_SPILL_BLKID || 1471 db->db_level + 1 == dn->dn_nlevels) { 1472 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf); 1473 mutex_enter(&dn->dn_mtx); 1474 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr); 1475 mutex_exit(&dn->dn_mtx); 1476 } 1477 DB_DNODE_EXIT(db); 1478 1479 if (db->db_state != DB_NOFILL) { 1480 dbuf_unoverride(dr); 1481 1482 ASSERT(db->db_buf != NULL); 1483 ASSERT(dr->dt.dl.dr_data != NULL); 1484 if (dr->dt.dl.dr_data != db->db_buf) 1485 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db)); 1486 } 1487 1488 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 1489 1490 ASSERT(db->db_dirtycnt > 0); 1491 db->db_dirtycnt -= 1; 1492 1493 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) { 1494 arc_buf_t *buf = db->db_buf; 1495 1496 ASSERT(db->db_state == DB_NOFILL || arc_released(buf)); 1497 dbuf_clear_data(db); 1498 VERIFY(arc_buf_remove_ref(buf, db)); 1499 dbuf_evict(db); 1500 return (B_TRUE); 1501 } 1502 1503 return (B_FALSE); 1504} 1505 1506void 1507dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx) 1508{ 1509 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1510 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH; 1511 1512 ASSERT(tx->tx_txg != 0); 1513 ASSERT(!refcount_is_zero(&db->db_holds)); 1514 1515 /* 1516 * Quick check for dirtyness. For already dirty blocks, this 1517 * reduces runtime of this function by >90%, and overall performance 1518 * by 50% for some workloads (e.g. file deletion with indirect blocks 1519 * cached). 1520 */ 1521 mutex_enter(&db->db_mtx); 1522 dbuf_dirty_record_t *dr; 1523 for (dr = db->db_last_dirty; 1524 dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) { 1525 /* 1526 * It's possible that it is already dirty but not cached, 1527 * because there are some calls to dbuf_dirty() that don't 1528 * go through dmu_buf_will_dirty(). 1529 */ 1530 if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) { 1531 /* This dbuf is already dirty and cached. */ 1532 dbuf_redirty(dr); 1533 mutex_exit(&db->db_mtx); 1534 return; 1535 } 1536 } 1537 mutex_exit(&db->db_mtx); 1538 1539 DB_DNODE_ENTER(db); 1540 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock)) 1541 rf |= DB_RF_HAVESTRUCT; 1542 DB_DNODE_EXIT(db); 1543 (void) dbuf_read(db, NULL, rf); 1544 (void) dbuf_dirty(db, tx); 1545} 1546 1547void 1548dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) 1549{ 1550 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1551 1552 db->db_state = DB_NOFILL; 1553 1554 dmu_buf_will_fill(db_fake, tx); 1555} 1556 1557void 1558dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) 1559{ 1560 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1561 1562 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1563 ASSERT(tx->tx_txg != 0); 1564 ASSERT(db->db_level == 0); 1565 ASSERT(!refcount_is_zero(&db->db_holds)); 1566 1567 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT || 1568 dmu_tx_private_ok(tx)); 1569 1570 dbuf_noread(db); 1571 (void) dbuf_dirty(db, tx); 1572} 1573 1574#pragma weak dmu_buf_fill_done = dbuf_fill_done 1575/* ARGSUSED */ 1576void 1577dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx) 1578{ 1579 mutex_enter(&db->db_mtx); 1580 DBUF_VERIFY(db); 1581 1582 if (db->db_state == DB_FILL) { 1583 if (db->db_level == 0 && db->db_freed_in_flight) { 1584 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1585 /* we were freed while filling */ 1586 /* XXX dbuf_undirty? */ 1587 bzero(db->db.db_data, db->db.db_size); 1588 db->db_freed_in_flight = FALSE; 1589 } 1590 db->db_state = DB_CACHED; 1591 cv_broadcast(&db->db_changed); 1592 } 1593 mutex_exit(&db->db_mtx); 1594} 1595 1596void 1597dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data, 1598 bp_embedded_type_t etype, enum zio_compress comp, 1599 int uncompressed_size, int compressed_size, int byteorder, 1600 dmu_tx_t *tx) 1601{ 1602 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; 1603 struct dirty_leaf *dl; 1604 dmu_object_type_t type; 1605 1606 if (etype == BP_EMBEDDED_TYPE_DATA) { 1607 ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset), 1608 SPA_FEATURE_EMBEDDED_DATA)); 1609 } 1610 1611 DB_DNODE_ENTER(db); 1612 type = DB_DNODE(db)->dn_type; 1613 DB_DNODE_EXIT(db); 1614 1615 ASSERT0(db->db_level); 1616 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1617 1618 dmu_buf_will_not_fill(dbuf, tx); 1619 1620 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); 1621 dl = &db->db_last_dirty->dt.dl; 1622 encode_embedded_bp_compressed(&dl->dr_overridden_by, 1623 data, comp, uncompressed_size, compressed_size); 1624 BPE_SET_ETYPE(&dl->dr_overridden_by, etype); 1625 BP_SET_TYPE(&dl->dr_overridden_by, type); 1626 BP_SET_LEVEL(&dl->dr_overridden_by, 0); 1627 BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder); 1628 1629 dl->dr_override_state = DR_OVERRIDDEN; 1630 dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg; 1631} 1632 1633/* 1634 * Directly assign a provided arc buf to a given dbuf if it's not referenced 1635 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf. 1636 */ 1637void 1638dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx) 1639{ 1640 ASSERT(!refcount_is_zero(&db->db_holds)); 1641 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1642 ASSERT(db->db_level == 0); 1643 ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA); 1644 ASSERT(buf != NULL); 1645 ASSERT(arc_buf_size(buf) == db->db.db_size); 1646 ASSERT(tx->tx_txg != 0); 1647 1648 arc_return_buf(buf, db); 1649 ASSERT(arc_released(buf)); 1650 1651 mutex_enter(&db->db_mtx); 1652 1653 while (db->db_state == DB_READ || db->db_state == DB_FILL) 1654 cv_wait(&db->db_changed, &db->db_mtx); 1655 1656 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED); 1657 1658 if (db->db_state == DB_CACHED && 1659 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) { 1660 mutex_exit(&db->db_mtx); 1661 (void) dbuf_dirty(db, tx); 1662 bcopy(buf->b_data, db->db.db_data, db->db.db_size); 1663 VERIFY(arc_buf_remove_ref(buf, db)); 1664 xuio_stat_wbuf_copied(); 1665 return; 1666 } 1667 1668 xuio_stat_wbuf_nocopy(); 1669 if (db->db_state == DB_CACHED) { 1670 dbuf_dirty_record_t *dr = db->db_last_dirty; 1671 1672 ASSERT(db->db_buf != NULL); 1673 if (dr != NULL && dr->dr_txg == tx->tx_txg) { 1674 ASSERT(dr->dt.dl.dr_data == db->db_buf); 1675 if (!arc_released(db->db_buf)) { 1676 ASSERT(dr->dt.dl.dr_override_state == 1677 DR_OVERRIDDEN); 1678 arc_release(db->db_buf, db); 1679 } 1680 dr->dt.dl.dr_data = buf; 1681 VERIFY(arc_buf_remove_ref(db->db_buf, db)); 1682 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) { 1683 arc_release(db->db_buf, db); 1684 VERIFY(arc_buf_remove_ref(db->db_buf, db)); 1685 } 1686 db->db_buf = NULL; 1687 } 1688 ASSERT(db->db_buf == NULL); 1689 dbuf_set_data(db, buf); 1690 db->db_state = DB_FILL; 1691 mutex_exit(&db->db_mtx); 1692 (void) dbuf_dirty(db, tx); 1693 dmu_buf_fill_done(&db->db, tx); 1694} 1695 1696/* 1697 * "Clear" the contents of this dbuf. This will mark the dbuf 1698 * EVICTING and clear *most* of its references. Unfortunately, 1699 * when we are not holding the dn_dbufs_mtx, we can't clear the 1700 * entry in the dn_dbufs list. We have to wait until dbuf_destroy() 1701 * in this case. For callers from the DMU we will usually see: 1702 * dbuf_clear()->arc_clear_callback()->dbuf_do_evict()->dbuf_destroy() 1703 * For the arc callback, we will usually see: 1704 * dbuf_do_evict()->dbuf_clear();dbuf_destroy() 1705 * Sometimes, though, we will get a mix of these two: 1706 * DMU: dbuf_clear()->arc_clear_callback() 1707 * ARC: dbuf_do_evict()->dbuf_destroy() 1708 * 1709 * This routine will dissociate the dbuf from the arc, by calling 1710 * arc_clear_callback(), but will not evict the data from the ARC. 1711 */ 1712void 1713dbuf_clear(dmu_buf_impl_t *db) 1714{ 1715 dnode_t *dn; 1716 dmu_buf_impl_t *parent = db->db_parent; 1717 dmu_buf_impl_t *dndb; 1718 boolean_t dbuf_gone = B_FALSE; 1719 1720 ASSERT(MUTEX_HELD(&db->db_mtx)); 1721 ASSERT(refcount_is_zero(&db->db_holds)); 1722 1723 dbuf_evict_user(db); 1724 1725 if (db->db_state == DB_CACHED) { 1726 ASSERT(db->db.db_data != NULL); 1727 if (db->db_blkid == DMU_BONUS_BLKID) { 1728 zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN); 1729 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 1730 } 1731 db->db.db_data = NULL; 1732 db->db_state = DB_UNCACHED; 1733 } 1734 1735 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL); 1736 ASSERT(db->db_data_pending == NULL); 1737 1738 db->db_state = DB_EVICTING; 1739 db->db_blkptr = NULL; 1740 1741 DB_DNODE_ENTER(db); 1742 dn = DB_DNODE(db); 1743 dndb = dn->dn_dbuf; 1744 if (db->db_blkid != DMU_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) { 1745 avl_remove(&dn->dn_dbufs, db); 1746 atomic_dec_32(&dn->dn_dbufs_count); 1747 membar_producer(); 1748 DB_DNODE_EXIT(db); 1749 /* 1750 * Decrementing the dbuf count means that the hold corresponding 1751 * to the removed dbuf is no longer discounted in dnode_move(), 1752 * so the dnode cannot be moved until after we release the hold. 1753 * The membar_producer() ensures visibility of the decremented 1754 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually 1755 * release any lock. 1756 */ 1757 dnode_rele(dn, db); 1758 db->db_dnode_handle = NULL; 1759 } else { 1760 DB_DNODE_EXIT(db); 1761 } 1762 1763 if (db->db_buf) 1764 dbuf_gone = arc_clear_callback(db->db_buf); 1765 1766 if (!dbuf_gone) 1767 mutex_exit(&db->db_mtx); 1768 1769 /* 1770 * If this dbuf is referenced from an indirect dbuf, 1771 * decrement the ref count on the indirect dbuf. 1772 */ 1773 if (parent && parent != dndb) 1774 dbuf_rele(parent, db); 1775} 1776 1777/* 1778 * Note: While bpp will always be updated if the function returns success, 1779 * parentp will not be updated if the dnode does not have dn_dbuf filled in; 1780 * this happens when the dnode is the meta-dnode, or a userused or groupused 1781 * object. 1782 */ 1783static int 1784dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse, 1785 dmu_buf_impl_t **parentp, blkptr_t **bpp) 1786{ 1787 int nlevels, epbs; 1788 1789 *parentp = NULL; 1790 *bpp = NULL; 1791 1792 ASSERT(blkid != DMU_BONUS_BLKID); 1793 1794 if (blkid == DMU_SPILL_BLKID) { 1795 mutex_enter(&dn->dn_mtx); 1796 if (dn->dn_have_spill && 1797 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) 1798 *bpp = &dn->dn_phys->dn_spill; 1799 else 1800 *bpp = NULL; 1801 dbuf_add_ref(dn->dn_dbuf, NULL); 1802 *parentp = dn->dn_dbuf; 1803 mutex_exit(&dn->dn_mtx); 1804 return (0); 1805 } 1806 1807 if (dn->dn_phys->dn_nlevels == 0) 1808 nlevels = 1; 1809 else 1810 nlevels = dn->dn_phys->dn_nlevels; 1811 1812 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1813 1814 ASSERT3U(level * epbs, <, 64); 1815 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1816 if (level >= nlevels || 1817 (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) { 1818 /* the buffer has no parent yet */ 1819 return (SET_ERROR(ENOENT)); 1820 } else if (level < nlevels-1) { 1821 /* this block is referenced from an indirect block */ 1822 int err = dbuf_hold_impl(dn, level+1, 1823 blkid >> epbs, fail_sparse, FALSE, NULL, parentp); 1824 if (err) 1825 return (err); 1826 err = dbuf_read(*parentp, NULL, 1827 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL)); 1828 if (err) { 1829 dbuf_rele(*parentp, NULL); 1830 *parentp = NULL; 1831 return (err); 1832 } 1833 *bpp = ((blkptr_t *)(*parentp)->db.db_data) + 1834 (blkid & ((1ULL << epbs) - 1)); 1835 return (0); 1836 } else { 1837 /* the block is referenced from the dnode */ 1838 ASSERT3U(level, ==, nlevels-1); 1839 ASSERT(dn->dn_phys->dn_nblkptr == 0 || 1840 blkid < dn->dn_phys->dn_nblkptr); 1841 if (dn->dn_dbuf) { 1842 dbuf_add_ref(dn->dn_dbuf, NULL); 1843 *parentp = dn->dn_dbuf; 1844 } 1845 *bpp = &dn->dn_phys->dn_blkptr[blkid]; 1846 return (0); 1847 } 1848} 1849 1850static dmu_buf_impl_t * 1851dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid, 1852 dmu_buf_impl_t *parent, blkptr_t *blkptr) 1853{ 1854 objset_t *os = dn->dn_objset; 1855 dmu_buf_impl_t *db, *odb; 1856 1857 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1858 ASSERT(dn->dn_type != DMU_OT_NONE); 1859 1860 db = kmem_cache_alloc(dbuf_cache, KM_SLEEP); 1861 1862 db->db_objset = os; 1863 db->db.db_object = dn->dn_object; 1864 db->db_level = level; 1865 db->db_blkid = blkid; 1866 db->db_last_dirty = NULL; 1867 db->db_dirtycnt = 0; 1868 db->db_dnode_handle = dn->dn_handle; 1869 db->db_parent = parent; 1870 db->db_blkptr = blkptr; 1871 1872 db->db_user = NULL; 1873 db->db_immediate_evict = 0; 1874 db->db_freed_in_flight = 0; 1875 1876 if (blkid == DMU_BONUS_BLKID) { 1877 ASSERT3P(parent, ==, dn->dn_dbuf); 1878 db->db.db_size = DN_MAX_BONUSLEN - 1879 (dn->dn_nblkptr-1) * sizeof (blkptr_t); 1880 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 1881 db->db.db_offset = DMU_BONUS_BLKID; 1882 db->db_state = DB_UNCACHED; 1883 /* the bonus dbuf is not placed in the hash table */ 1884 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 1885 return (db); 1886 } else if (blkid == DMU_SPILL_BLKID) { 1887 db->db.db_size = (blkptr != NULL) ? 1888 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE; 1889 db->db.db_offset = 0; 1890 } else { 1891 int blocksize = 1892 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz; 1893 db->db.db_size = blocksize; 1894 db->db.db_offset = db->db_blkid * blocksize; 1895 } 1896 1897 /* 1898 * Hold the dn_dbufs_mtx while we get the new dbuf 1899 * in the hash table *and* added to the dbufs list. 1900 * This prevents a possible deadlock with someone 1901 * trying to look up this dbuf before its added to the 1902 * dn_dbufs list. 1903 */ 1904 mutex_enter(&dn->dn_dbufs_mtx); 1905 db->db_state = DB_EVICTING; 1906 if ((odb = dbuf_hash_insert(db)) != NULL) { 1907 /* someone else inserted it first */ 1908 kmem_cache_free(dbuf_cache, db); 1909 mutex_exit(&dn->dn_dbufs_mtx); 1910 return (odb); 1911 } 1912 avl_add(&dn->dn_dbufs, db); 1913 if (db->db_level == 0 && db->db_blkid >= 1914 dn->dn_unlisted_l0_blkid) 1915 dn->dn_unlisted_l0_blkid = db->db_blkid + 1; 1916 db->db_state = DB_UNCACHED; 1917 mutex_exit(&dn->dn_dbufs_mtx); 1918 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 1919 1920 if (parent && parent != dn->dn_dbuf) 1921 dbuf_add_ref(parent, db); 1922 1923 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 1924 refcount_count(&dn->dn_holds) > 0); 1925 (void) refcount_add(&dn->dn_holds, db); 1926 atomic_inc_32(&dn->dn_dbufs_count); 1927 1928 dprintf_dbuf(db, "db=%p\n", db); 1929 1930 return (db); 1931} 1932 1933static int 1934dbuf_do_evict(void *private) 1935{ 1936 dmu_buf_impl_t *db = private; 1937 1938 if (!MUTEX_HELD(&db->db_mtx)) 1939 mutex_enter(&db->db_mtx); 1940 1941 ASSERT(refcount_is_zero(&db->db_holds)); 1942 1943 if (db->db_state != DB_EVICTING) { 1944 ASSERT(db->db_state == DB_CACHED); 1945 DBUF_VERIFY(db); 1946 db->db_buf = NULL; 1947 dbuf_evict(db); 1948 } else { 1949 mutex_exit(&db->db_mtx); 1950 dbuf_destroy(db); 1951 } 1952 return (0); 1953} 1954 1955static void 1956dbuf_destroy(dmu_buf_impl_t *db) 1957{ 1958 ASSERT(refcount_is_zero(&db->db_holds)); 1959 1960 if (db->db_blkid != DMU_BONUS_BLKID) { 1961 /* 1962 * If this dbuf is still on the dn_dbufs list, 1963 * remove it from that list. 1964 */ 1965 if (db->db_dnode_handle != NULL) { 1966 dnode_t *dn; 1967 1968 DB_DNODE_ENTER(db); 1969 dn = DB_DNODE(db); 1970 mutex_enter(&dn->dn_dbufs_mtx); 1971 avl_remove(&dn->dn_dbufs, db); 1972 atomic_dec_32(&dn->dn_dbufs_count); 1973 mutex_exit(&dn->dn_dbufs_mtx); 1974 DB_DNODE_EXIT(db); 1975 /* 1976 * Decrementing the dbuf count means that the hold 1977 * corresponding to the removed dbuf is no longer 1978 * discounted in dnode_move(), so the dnode cannot be 1979 * moved until after we release the hold. 1980 */ 1981 dnode_rele(dn, db); 1982 db->db_dnode_handle = NULL; 1983 } 1984 dbuf_hash_remove(db); 1985 } 1986 db->db_parent = NULL; 1987 db->db_buf = NULL; 1988 1989 ASSERT(db->db.db_data == NULL); 1990 ASSERT(db->db_hash_next == NULL); 1991 ASSERT(db->db_blkptr == NULL); 1992 ASSERT(db->db_data_pending == NULL); 1993 1994 kmem_cache_free(dbuf_cache, db); 1995 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 1996} 1997 1998typedef struct dbuf_prefetch_arg { 1999 spa_t *dpa_spa; /* The spa to issue the prefetch in. */ 2000 zbookmark_phys_t dpa_zb; /* The target block to prefetch. */ 2001 int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */ 2002 int dpa_curlevel; /* The current level that we're reading */ 2003 zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */ 2004 zio_t *dpa_zio; /* The parent zio_t for all prefetches. */ 2005 arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */ 2006} dbuf_prefetch_arg_t; 2007 2008/* 2009 * Actually issue the prefetch read for the block given. 2010 */ 2011static void 2012dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp) 2013{ 2014 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) 2015 return; 2016 2017 arc_flags_t aflags = 2018 dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH; 2019 2020 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp)); 2021 ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level); 2022 ASSERT(dpa->dpa_zio != NULL); 2023 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL, 2024 dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2025 &aflags, &dpa->dpa_zb); 2026} 2027 2028/* 2029 * Called when an indirect block above our prefetch target is read in. This 2030 * will either read in the next indirect block down the tree or issue the actual 2031 * prefetch if the next block down is our target. 2032 */ 2033static void 2034dbuf_prefetch_indirect_done(zio_t *zio, arc_buf_t *abuf, void *private) 2035{ 2036 dbuf_prefetch_arg_t *dpa = private; 2037 2038 ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel); 2039 ASSERT3S(dpa->dpa_curlevel, >, 0); 2040 if (zio != NULL) { 2041 ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel); 2042 ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size); 2043 ASSERT3P(zio->io_spa, ==, dpa->dpa_spa); 2044 } 2045 2046 dpa->dpa_curlevel--; 2047 2048 uint64_t nextblkid = dpa->dpa_zb.zb_blkid >> 2049 (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level)); 2050 blkptr_t *bp = ((blkptr_t *)abuf->b_data) + 2051 P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs); 2052 if (BP_IS_HOLE(bp) || (zio != NULL && zio->io_error != 0)) { 2053 kmem_free(dpa, sizeof (*dpa)); 2054 } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) { 2055 ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid); 2056 dbuf_issue_final_prefetch(dpa, bp); 2057 kmem_free(dpa, sizeof (*dpa)); 2058 } else { 2059 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT; 2060 zbookmark_phys_t zb; 2061 2062 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp)); 2063 2064 SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset, 2065 dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid); 2066 2067 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, 2068 bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio, 2069 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2070 &iter_aflags, &zb); 2071 } 2072 (void) arc_buf_remove_ref(abuf, private); 2073} 2074 2075/* 2076 * Issue prefetch reads for the given block on the given level. If the indirect 2077 * blocks above that block are not in memory, we will read them in 2078 * asynchronously. As a result, this call never blocks waiting for a read to 2079 * complete. 2080 */ 2081void 2082dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio, 2083 arc_flags_t aflags) 2084{ 2085 blkptr_t bp; 2086 int epbs, nlevels, curlevel; 2087 uint64_t curblkid; 2088 2089 ASSERT(blkid != DMU_BONUS_BLKID); 2090 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 2091 2092 if (blkid > dn->dn_maxblkid) 2093 return; 2094 2095 if (dnode_block_freed(dn, blkid)) 2096 return; 2097 2098 /* 2099 * This dnode hasn't been written to disk yet, so there's nothing to 2100 * prefetch. 2101 */ 2102 nlevels = dn->dn_phys->dn_nlevels; 2103 if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0) 2104 return; 2105 2106 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 2107 if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level)) 2108 return; 2109 2110 dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object, 2111 level, blkid); 2112 if (db != NULL) { 2113 mutex_exit(&db->db_mtx); 2114 /* 2115 * This dbuf already exists. It is either CACHED, or 2116 * (we assume) about to be read or filled. 2117 */ 2118 return; 2119 } 2120 2121 /* 2122 * Find the closest ancestor (indirect block) of the target block 2123 * that is present in the cache. In this indirect block, we will 2124 * find the bp that is at curlevel, curblkid. 2125 */ 2126 curlevel = level; 2127 curblkid = blkid; 2128 while (curlevel < nlevels - 1) { 2129 int parent_level = curlevel + 1; 2130 uint64_t parent_blkid = curblkid >> epbs; 2131 dmu_buf_impl_t *db; 2132 2133 if (dbuf_hold_impl(dn, parent_level, parent_blkid, 2134 FALSE, TRUE, FTAG, &db) == 0) { 2135 blkptr_t *bpp = db->db_buf->b_data; 2136 bp = bpp[P2PHASE(curblkid, 1 << epbs)]; 2137 dbuf_rele(db, FTAG); 2138 break; 2139 } 2140 2141 curlevel = parent_level; 2142 curblkid = parent_blkid; 2143 } 2144 2145 if (curlevel == nlevels - 1) { 2146 /* No cached indirect blocks found. */ 2147 ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr); 2148 bp = dn->dn_phys->dn_blkptr[curblkid]; 2149 } 2150 if (BP_IS_HOLE(&bp)) 2151 return; 2152 2153 ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp)); 2154 2155 zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL, 2156 ZIO_FLAG_CANFAIL); 2157 2158 dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP); 2159 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset; 2160 SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET, 2161 dn->dn_object, level, blkid); 2162 dpa->dpa_curlevel = curlevel; 2163 dpa->dpa_prio = prio; 2164 dpa->dpa_aflags = aflags; 2165 dpa->dpa_spa = dn->dn_objset->os_spa; 2166 dpa->dpa_epbs = epbs; 2167 dpa->dpa_zio = pio; 2168 2169 /* 2170 * If we have the indirect just above us, no need to do the asynchronous 2171 * prefetch chain; we'll just run the last step ourselves. If we're at 2172 * a higher level, though, we want to issue the prefetches for all the 2173 * indirect blocks asynchronously, so we can go on with whatever we were 2174 * doing. 2175 */ 2176 if (curlevel == level) { 2177 ASSERT3U(curblkid, ==, blkid); 2178 dbuf_issue_final_prefetch(dpa, &bp); 2179 kmem_free(dpa, sizeof (*dpa)); 2180 } else { 2181 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT; 2182 zbookmark_phys_t zb; 2183 2184 SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET, 2185 dn->dn_object, curlevel, curblkid); 2186 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, 2187 &bp, dbuf_prefetch_indirect_done, dpa, prio, 2188 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2189 &iter_aflags, &zb); 2190 } 2191 /* 2192 * We use pio here instead of dpa_zio since it's possible that 2193 * dpa may have already been freed. 2194 */ 2195 zio_nowait(pio); 2196} 2197 2198/* 2199 * Returns with db_holds incremented, and db_mtx not held. 2200 * Note: dn_struct_rwlock must be held. 2201 */ 2202int 2203dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, 2204 boolean_t fail_sparse, boolean_t fail_uncached, 2205 void *tag, dmu_buf_impl_t **dbp) 2206{ 2207 dmu_buf_impl_t *db, *parent = NULL; 2208 2209 ASSERT(blkid != DMU_BONUS_BLKID); 2210 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 2211 ASSERT3U(dn->dn_nlevels, >, level); 2212 2213 *dbp = NULL; 2214top: 2215 /* dbuf_find() returns with db_mtx held */ 2216 db = dbuf_find(dn->dn_objset, dn->dn_object, level, blkid); 2217 2218 if (db == NULL) { 2219 blkptr_t *bp = NULL; 2220 int err; 2221 2222 if (fail_uncached) 2223 return (SET_ERROR(ENOENT)); 2224 2225 ASSERT3P(parent, ==, NULL); 2226 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp); 2227 if (fail_sparse) { 2228 if (err == 0 && bp && BP_IS_HOLE(bp)) 2229 err = SET_ERROR(ENOENT); 2230 if (err) { 2231 if (parent) 2232 dbuf_rele(parent, NULL); 2233 return (err); 2234 } 2235 } 2236 if (err && err != ENOENT) 2237 return (err); 2238 db = dbuf_create(dn, level, blkid, parent, bp); 2239 } 2240 2241 if (fail_uncached && db->db_state != DB_CACHED) { 2242 mutex_exit(&db->db_mtx); 2243 return (SET_ERROR(ENOENT)); 2244 } 2245 2246 if (db->db_buf && refcount_is_zero(&db->db_holds)) { 2247 arc_buf_add_ref(db->db_buf, db); 2248 if (db->db_buf->b_data == NULL) { 2249 dbuf_clear(db); 2250 if (parent) { 2251 dbuf_rele(parent, NULL); 2252 parent = NULL; 2253 } 2254 goto top; 2255 } 2256 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data); 2257 } 2258 2259 ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf)); 2260 2261 /* 2262 * If this buffer is currently syncing out, and we are are 2263 * still referencing it from db_data, we need to make a copy 2264 * of it in case we decide we want to dirty it again in this txg. 2265 */ 2266 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 2267 dn->dn_object != DMU_META_DNODE_OBJECT && 2268 db->db_state == DB_CACHED && db->db_data_pending) { 2269 dbuf_dirty_record_t *dr = db->db_data_pending; 2270 2271 if (dr->dt.dl.dr_data == db->db_buf) { 2272 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 2273 2274 dbuf_set_data(db, 2275 arc_buf_alloc(dn->dn_objset->os_spa, 2276 db->db.db_size, db, type)); 2277 bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data, 2278 db->db.db_size); 2279 } 2280 } 2281 2282 (void) refcount_add(&db->db_holds, tag); 2283 DBUF_VERIFY(db); 2284 mutex_exit(&db->db_mtx); 2285 2286 /* NOTE: we can't rele the parent until after we drop the db_mtx */ 2287 if (parent) 2288 dbuf_rele(parent, NULL); 2289 2290 ASSERT3P(DB_DNODE(db), ==, dn); 2291 ASSERT3U(db->db_blkid, ==, blkid); 2292 ASSERT3U(db->db_level, ==, level); 2293 *dbp = db; 2294 2295 return (0); 2296} 2297 2298dmu_buf_impl_t * 2299dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag) 2300{ 2301 return (dbuf_hold_level(dn, 0, blkid, tag)); 2302} 2303 2304dmu_buf_impl_t * 2305dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag) 2306{ 2307 dmu_buf_impl_t *db; 2308 int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db); 2309 return (err ? NULL : db); 2310} 2311 2312void 2313dbuf_create_bonus(dnode_t *dn) 2314{ 2315 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 2316 2317 ASSERT(dn->dn_bonus == NULL); 2318 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL); 2319} 2320 2321int 2322dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx) 2323{ 2324 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2325 dnode_t *dn; 2326 2327 if (db->db_blkid != DMU_SPILL_BLKID) 2328 return (SET_ERROR(ENOTSUP)); 2329 if (blksz == 0) 2330 blksz = SPA_MINBLOCKSIZE; 2331 ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset))); 2332 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE); 2333 2334 DB_DNODE_ENTER(db); 2335 dn = DB_DNODE(db); 2336 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 2337 dbuf_new_size(db, blksz, tx); 2338 rw_exit(&dn->dn_struct_rwlock); 2339 DB_DNODE_EXIT(db); 2340 2341 return (0); 2342} 2343 2344void 2345dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx) 2346{ 2347 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx); 2348} 2349 2350#pragma weak dmu_buf_add_ref = dbuf_add_ref 2351void 2352dbuf_add_ref(dmu_buf_impl_t *db, void *tag) 2353{ 2354 int64_t holds = refcount_add(&db->db_holds, tag); 2355 ASSERT(holds > 1); 2356} 2357 2358#pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref 2359boolean_t 2360dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid, 2361 void *tag) 2362{ 2363 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2364 dmu_buf_impl_t *found_db; 2365 boolean_t result = B_FALSE; 2366 2367 if (db->db_blkid == DMU_BONUS_BLKID) 2368 found_db = dbuf_find_bonus(os, obj); 2369 else 2370 found_db = dbuf_find(os, obj, 0, blkid); 2371 2372 if (found_db != NULL) { 2373 if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) { 2374 (void) refcount_add(&db->db_holds, tag); 2375 result = B_TRUE; 2376 } 2377 mutex_exit(&db->db_mtx); 2378 } 2379 return (result); 2380} 2381 2382/* 2383 * If you call dbuf_rele() you had better not be referencing the dnode handle 2384 * unless you have some other direct or indirect hold on the dnode. (An indirect 2385 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.) 2386 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the 2387 * dnode's parent dbuf evicting its dnode handles. 2388 */ 2389void 2390dbuf_rele(dmu_buf_impl_t *db, void *tag) 2391{ 2392 mutex_enter(&db->db_mtx); 2393 dbuf_rele_and_unlock(db, tag); 2394} 2395 2396void 2397dmu_buf_rele(dmu_buf_t *db, void *tag) 2398{ 2399 dbuf_rele((dmu_buf_impl_t *)db, tag); 2400} 2401 2402/* 2403 * dbuf_rele() for an already-locked dbuf. This is necessary to allow 2404 * db_dirtycnt and db_holds to be updated atomically. 2405 */ 2406void 2407dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag) 2408{ 2409 int64_t holds; 2410 2411 ASSERT(MUTEX_HELD(&db->db_mtx)); 2412 DBUF_VERIFY(db); 2413 2414 /* 2415 * Remove the reference to the dbuf before removing its hold on the 2416 * dnode so we can guarantee in dnode_move() that a referenced bonus 2417 * buffer has a corresponding dnode hold. 2418 */ 2419 holds = refcount_remove(&db->db_holds, tag); 2420 ASSERT(holds >= 0); 2421 2422 /* 2423 * We can't freeze indirects if there is a possibility that they 2424 * may be modified in the current syncing context. 2425 */ 2426 if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) 2427 arc_buf_freeze(db->db_buf); 2428 2429 if (holds == db->db_dirtycnt && 2430 db->db_level == 0 && db->db_immediate_evict) 2431 dbuf_evict_user(db); 2432 2433 if (holds == 0) { 2434 if (db->db_blkid == DMU_BONUS_BLKID) { 2435 dnode_t *dn; 2436 2437 /* 2438 * If the dnode moves here, we cannot cross this 2439 * barrier until the move completes. 2440 */ 2441 DB_DNODE_ENTER(db); 2442 2443 dn = DB_DNODE(db); 2444 atomic_dec_32(&dn->dn_dbufs_count); 2445 2446 /* 2447 * Decrementing the dbuf count means that the bonus 2448 * buffer's dnode hold is no longer discounted in 2449 * dnode_move(). The dnode cannot move until after 2450 * the dnode_rele_and_unlock() below. 2451 */ 2452 DB_DNODE_EXIT(db); 2453 2454 /* 2455 * Do not reference db after its lock is dropped. 2456 * Another thread may evict it. 2457 */ 2458 mutex_exit(&db->db_mtx); 2459 2460 /* 2461 * If the dnode has been freed, evict the bonus 2462 * buffer immediately. The data in the bonus 2463 * buffer is no longer relevant and this prevents 2464 * a stale bonus buffer from being associated 2465 * with this dnode_t should the dnode_t be reused 2466 * prior to being destroyed. 2467 */ 2468 mutex_enter(&dn->dn_mtx); 2469 if (dn->dn_type == DMU_OT_NONE || 2470 dn->dn_free_txg != 0) { 2471 /* 2472 * Drop dn_mtx. It is a leaf lock and 2473 * cannot be held when dnode_evict_bonus() 2474 * acquires other locks in order to 2475 * perform the eviction. 2476 * 2477 * Freed dnodes cannot be reused until the 2478 * last hold is released. Since this bonus 2479 * buffer has a hold, the dnode will remain 2480 * in the free state, even without dn_mtx 2481 * held, until the dnode_rele_and_unlock() 2482 * below. 2483 */ 2484 mutex_exit(&dn->dn_mtx); 2485 dnode_evict_bonus(dn); 2486 mutex_enter(&dn->dn_mtx); 2487 } 2488 dnode_rele_and_unlock(dn, db); 2489 } else if (db->db_buf == NULL) { 2490 /* 2491 * This is a special case: we never associated this 2492 * dbuf with any data allocated from the ARC. 2493 */ 2494 ASSERT(db->db_state == DB_UNCACHED || 2495 db->db_state == DB_NOFILL); 2496 dbuf_evict(db); 2497 } else if (arc_released(db->db_buf)) { 2498 arc_buf_t *buf = db->db_buf; 2499 /* 2500 * This dbuf has anonymous data associated with it. 2501 */ 2502 dbuf_clear_data(db); 2503 VERIFY(arc_buf_remove_ref(buf, db)); 2504 dbuf_evict(db); 2505 } else { 2506 VERIFY(!arc_buf_remove_ref(db->db_buf, db)); 2507 2508 /* 2509 * A dbuf will be eligible for eviction if either the 2510 * 'primarycache' property is set or a duplicate 2511 * copy of this buffer is already cached in the arc. 2512 * 2513 * In the case of the 'primarycache' a buffer 2514 * is considered for eviction if it matches the 2515 * criteria set in the property. 2516 * 2517 * To decide if our buffer is considered a 2518 * duplicate, we must call into the arc to determine 2519 * if multiple buffers are referencing the same 2520 * block on-disk. If so, then we simply evict 2521 * ourselves. 2522 */ 2523 if (!DBUF_IS_CACHEABLE(db)) { 2524 if (db->db_blkptr != NULL && 2525 !BP_IS_HOLE(db->db_blkptr) && 2526 !BP_IS_EMBEDDED(db->db_blkptr)) { 2527 spa_t *spa = 2528 dmu_objset_spa(db->db_objset); 2529 blkptr_t bp = *db->db_blkptr; 2530 dbuf_clear(db); 2531 arc_freed(spa, &bp); 2532 } else { 2533 dbuf_clear(db); 2534 } 2535 } else if (db->db_objset->os_evicting || 2536 arc_buf_eviction_needed(db->db_buf)) { 2537 dbuf_clear(db); 2538 } else { 2539 mutex_exit(&db->db_mtx); 2540 } 2541 } 2542 } else { 2543 mutex_exit(&db->db_mtx); 2544 } 2545} 2546 2547#pragma weak dmu_buf_refcount = dbuf_refcount 2548uint64_t 2549dbuf_refcount(dmu_buf_impl_t *db) 2550{ 2551 return (refcount_count(&db->db_holds)); 2552} 2553 2554void * 2555dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user, 2556 dmu_buf_user_t *new_user) 2557{ 2558 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2559 2560 mutex_enter(&db->db_mtx); 2561 dbuf_verify_user(db, DBVU_NOT_EVICTING); 2562 if (db->db_user == old_user) 2563 db->db_user = new_user; 2564 else 2565 old_user = db->db_user; 2566 dbuf_verify_user(db, DBVU_NOT_EVICTING); 2567 mutex_exit(&db->db_mtx); 2568 2569 return (old_user); 2570} 2571 2572void * 2573dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user) 2574{ 2575 return (dmu_buf_replace_user(db_fake, NULL, user)); 2576} 2577 2578void * 2579dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user) 2580{ 2581 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2582 2583 db->db_immediate_evict = TRUE; 2584 return (dmu_buf_set_user(db_fake, user)); 2585} 2586 2587void * 2588dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user) 2589{ 2590 return (dmu_buf_replace_user(db_fake, user, NULL)); 2591} 2592 2593void * 2594dmu_buf_get_user(dmu_buf_t *db_fake) 2595{ 2596 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2597 2598 dbuf_verify_user(db, DBVU_NOT_EVICTING); 2599 return (db->db_user); 2600} 2601 2602void 2603dmu_buf_user_evict_wait() 2604{ 2605 taskq_wait(dbu_evict_taskq); 2606} 2607 2608boolean_t 2609dmu_buf_freeable(dmu_buf_t *dbuf) 2610{ 2611 boolean_t res = B_FALSE; 2612 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; 2613 2614 if (db->db_blkptr) 2615 res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset, 2616 db->db_blkptr, db->db_blkptr->blk_birth); 2617 2618 return (res); 2619} 2620 2621blkptr_t * 2622dmu_buf_get_blkptr(dmu_buf_t *db) 2623{ 2624 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 2625 return (dbi->db_blkptr); 2626} 2627 2628static void 2629dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db) 2630{ 2631 /* ASSERT(dmu_tx_is_syncing(tx) */ 2632 ASSERT(MUTEX_HELD(&db->db_mtx)); 2633 2634 if (db->db_blkptr != NULL) 2635 return; 2636 2637 if (db->db_blkid == DMU_SPILL_BLKID) { 2638 db->db_blkptr = &dn->dn_phys->dn_spill; 2639 BP_ZERO(db->db_blkptr); 2640 return; 2641 } 2642 if (db->db_level == dn->dn_phys->dn_nlevels-1) { 2643 /* 2644 * This buffer was allocated at a time when there was 2645 * no available blkptrs from the dnode, or it was 2646 * inappropriate to hook it in (i.e., nlevels mis-match). 2647 */ 2648 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr); 2649 ASSERT(db->db_parent == NULL); 2650 db->db_parent = dn->dn_dbuf; 2651 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid]; 2652 DBUF_VERIFY(db); 2653 } else { 2654 dmu_buf_impl_t *parent = db->db_parent; 2655 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 2656 2657 ASSERT(dn->dn_phys->dn_nlevels > 1); 2658 if (parent == NULL) { 2659 mutex_exit(&db->db_mtx); 2660 rw_enter(&dn->dn_struct_rwlock, RW_READER); 2661 parent = dbuf_hold_level(dn, db->db_level + 1, 2662 db->db_blkid >> epbs, db); 2663 rw_exit(&dn->dn_struct_rwlock); 2664 mutex_enter(&db->db_mtx); 2665 db->db_parent = parent; 2666 } 2667 db->db_blkptr = (blkptr_t *)parent->db.db_data + 2668 (db->db_blkid & ((1ULL << epbs) - 1)); 2669 DBUF_VERIFY(db); 2670 } 2671} 2672 2673static void 2674dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx) 2675{ 2676 dmu_buf_impl_t *db = dr->dr_dbuf; 2677 dnode_t *dn; 2678 zio_t *zio; 2679 2680 ASSERT(dmu_tx_is_syncing(tx)); 2681 2682 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 2683 2684 mutex_enter(&db->db_mtx); 2685 2686 ASSERT(db->db_level > 0); 2687 DBUF_VERIFY(db); 2688 2689 /* Read the block if it hasn't been read yet. */ 2690 if (db->db_buf == NULL) { 2691 mutex_exit(&db->db_mtx); 2692 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED); 2693 mutex_enter(&db->db_mtx); 2694 } 2695 ASSERT3U(db->db_state, ==, DB_CACHED); 2696 ASSERT(db->db_buf != NULL); 2697 2698 DB_DNODE_ENTER(db); 2699 dn = DB_DNODE(db); 2700 /* Indirect block size must match what the dnode thinks it is. */ 2701 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 2702 dbuf_check_blkptr(dn, db); 2703 DB_DNODE_EXIT(db); 2704 2705 /* Provide the pending dirty record to child dbufs */ 2706 db->db_data_pending = dr; 2707 2708 mutex_exit(&db->db_mtx); 2709 dbuf_write(dr, db->db_buf, tx); 2710 2711 zio = dr->dr_zio; 2712 mutex_enter(&dr->dt.di.dr_mtx); 2713 dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx); 2714 ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 2715 mutex_exit(&dr->dt.di.dr_mtx); 2716 zio_nowait(zio); 2717} 2718 2719static void 2720dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx) 2721{ 2722 arc_buf_t **datap = &dr->dt.dl.dr_data; 2723 dmu_buf_impl_t *db = dr->dr_dbuf; 2724 dnode_t *dn; 2725 objset_t *os; 2726 uint64_t txg = tx->tx_txg; 2727 2728 ASSERT(dmu_tx_is_syncing(tx)); 2729 2730 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 2731 2732 mutex_enter(&db->db_mtx); 2733 /* 2734 * To be synced, we must be dirtied. But we 2735 * might have been freed after the dirty. 2736 */ 2737 if (db->db_state == DB_UNCACHED) { 2738 /* This buffer has been freed since it was dirtied */ 2739 ASSERT(db->db.db_data == NULL); 2740 } else if (db->db_state == DB_FILL) { 2741 /* This buffer was freed and is now being re-filled */ 2742 ASSERT(db->db.db_data != dr->dt.dl.dr_data); 2743 } else { 2744 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL); 2745 } 2746 DBUF_VERIFY(db); 2747 2748 DB_DNODE_ENTER(db); 2749 dn = DB_DNODE(db); 2750 2751 if (db->db_blkid == DMU_SPILL_BLKID) { 2752 mutex_enter(&dn->dn_mtx); 2753 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR; 2754 mutex_exit(&dn->dn_mtx); 2755 } 2756 2757 /* 2758 * If this is a bonus buffer, simply copy the bonus data into the 2759 * dnode. It will be written out when the dnode is synced (and it 2760 * will be synced, since it must have been dirty for dbuf_sync to 2761 * be called). 2762 */ 2763 if (db->db_blkid == DMU_BONUS_BLKID) { 2764 dbuf_dirty_record_t **drp; 2765 2766 ASSERT(*datap != NULL); 2767 ASSERT0(db->db_level); 2768 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN); 2769 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen); 2770 DB_DNODE_EXIT(db); 2771 2772 if (*datap != db->db.db_data) { 2773 zio_buf_free(*datap, DN_MAX_BONUSLEN); 2774 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 2775 } 2776 db->db_data_pending = NULL; 2777 drp = &db->db_last_dirty; 2778 while (*drp != dr) 2779 drp = &(*drp)->dr_next; 2780 ASSERT(dr->dr_next == NULL); 2781 ASSERT(dr->dr_dbuf == db); 2782 *drp = dr->dr_next; 2783 if (dr->dr_dbuf->db_level != 0) { 2784 list_destroy(&dr->dt.di.dr_children); 2785 mutex_destroy(&dr->dt.di.dr_mtx); 2786 } 2787 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 2788 ASSERT(db->db_dirtycnt > 0); 2789 db->db_dirtycnt -= 1; 2790 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg); 2791 return; 2792 } 2793 2794 os = dn->dn_objset; 2795 2796 /* 2797 * This function may have dropped the db_mtx lock allowing a dmu_sync 2798 * operation to sneak in. As a result, we need to ensure that we 2799 * don't check the dr_override_state until we have returned from 2800 * dbuf_check_blkptr. 2801 */ 2802 dbuf_check_blkptr(dn, db); 2803 2804 /* 2805 * If this buffer is in the middle of an immediate write, 2806 * wait for the synchronous IO to complete. 2807 */ 2808 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) { 2809 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT); 2810 cv_wait(&db->db_changed, &db->db_mtx); 2811 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN); 2812 } 2813 2814 if (db->db_state != DB_NOFILL && 2815 dn->dn_object != DMU_META_DNODE_OBJECT && 2816 refcount_count(&db->db_holds) > 1 && 2817 dr->dt.dl.dr_override_state != DR_OVERRIDDEN && 2818 *datap == db->db_buf) { 2819 /* 2820 * If this buffer is currently "in use" (i.e., there 2821 * are active holds and db_data still references it), 2822 * then make a copy before we start the write so that 2823 * any modifications from the open txg will not leak 2824 * into this write. 2825 * 2826 * NOTE: this copy does not need to be made for 2827 * objects only modified in the syncing context (e.g. 2828 * DNONE_DNODE blocks). 2829 */ 2830 int blksz = arc_buf_size(*datap); 2831 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 2832 *datap = arc_buf_alloc(os->os_spa, blksz, db, type); 2833 bcopy(db->db.db_data, (*datap)->b_data, blksz); 2834 } 2835 db->db_data_pending = dr; 2836 2837 mutex_exit(&db->db_mtx); 2838 2839 dbuf_write(dr, *datap, tx); 2840 2841 ASSERT(!list_link_active(&dr->dr_dirty_node)); 2842 if (dn->dn_object == DMU_META_DNODE_OBJECT) { 2843 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr); 2844 DB_DNODE_EXIT(db); 2845 } else { 2846 /* 2847 * Although zio_nowait() does not "wait for an IO", it does 2848 * initiate the IO. If this is an empty write it seems plausible 2849 * that the IO could actually be completed before the nowait 2850 * returns. We need to DB_DNODE_EXIT() first in case 2851 * zio_nowait() invalidates the dbuf. 2852 */ 2853 DB_DNODE_EXIT(db); 2854 zio_nowait(dr->dr_zio); 2855 } 2856} 2857 2858void 2859dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx) 2860{ 2861 dbuf_dirty_record_t *dr; 2862 2863 while (dr = list_head(list)) { 2864 if (dr->dr_zio != NULL) { 2865 /* 2866 * If we find an already initialized zio then we 2867 * are processing the meta-dnode, and we have finished. 2868 * The dbufs for all dnodes are put back on the list 2869 * during processing, so that we can zio_wait() 2870 * these IOs after initiating all child IOs. 2871 */ 2872 ASSERT3U(dr->dr_dbuf->db.db_object, ==, 2873 DMU_META_DNODE_OBJECT); 2874 break; 2875 } 2876 if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID && 2877 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) { 2878 VERIFY3U(dr->dr_dbuf->db_level, ==, level); 2879 } 2880 list_remove(list, dr); 2881 if (dr->dr_dbuf->db_level > 0) 2882 dbuf_sync_indirect(dr, tx); 2883 else 2884 dbuf_sync_leaf(dr, tx); 2885 } 2886} 2887 2888/* ARGSUSED */ 2889static void 2890dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb) 2891{ 2892 dmu_buf_impl_t *db = vdb; 2893 dnode_t *dn; 2894 blkptr_t *bp = zio->io_bp; 2895 blkptr_t *bp_orig = &zio->io_bp_orig; 2896 spa_t *spa = zio->io_spa; 2897 int64_t delta; 2898 uint64_t fill = 0; 2899 int i; 2900 2901 ASSERT3P(db->db_blkptr, ==, bp); 2902 2903 DB_DNODE_ENTER(db); 2904 dn = DB_DNODE(db); 2905 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig); 2906 dnode_diduse_space(dn, delta - zio->io_prev_space_delta); 2907 zio->io_prev_space_delta = delta; 2908 2909 if (bp->blk_birth != 0) { 2910 ASSERT((db->db_blkid != DMU_SPILL_BLKID && 2911 BP_GET_TYPE(bp) == dn->dn_type) || 2912 (db->db_blkid == DMU_SPILL_BLKID && 2913 BP_GET_TYPE(bp) == dn->dn_bonustype) || 2914 BP_IS_EMBEDDED(bp)); 2915 ASSERT(BP_GET_LEVEL(bp) == db->db_level); 2916 } 2917 2918 mutex_enter(&db->db_mtx); 2919 2920#ifdef ZFS_DEBUG 2921 if (db->db_blkid == DMU_SPILL_BLKID) { 2922 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); 2923 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && 2924 db->db_blkptr == &dn->dn_phys->dn_spill); 2925 } 2926#endif 2927 2928 if (db->db_level == 0) { 2929 mutex_enter(&dn->dn_mtx); 2930 if (db->db_blkid > dn->dn_phys->dn_maxblkid && 2931 db->db_blkid != DMU_SPILL_BLKID) 2932 dn->dn_phys->dn_maxblkid = db->db_blkid; 2933 mutex_exit(&dn->dn_mtx); 2934 2935 if (dn->dn_type == DMU_OT_DNODE) { 2936 dnode_phys_t *dnp = db->db.db_data; 2937 for (i = db->db.db_size >> DNODE_SHIFT; i > 0; 2938 i--, dnp++) { 2939 if (dnp->dn_type != DMU_OT_NONE) 2940 fill++; 2941 } 2942 } else { 2943 if (BP_IS_HOLE(bp)) { 2944 fill = 0; 2945 } else { 2946 fill = 1; 2947 } 2948 } 2949 } else { 2950 blkptr_t *ibp = db->db.db_data; 2951 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 2952 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) { 2953 if (BP_IS_HOLE(ibp)) 2954 continue; 2955 fill += BP_GET_FILL(ibp); 2956 } 2957 } 2958 DB_DNODE_EXIT(db); 2959 2960 if (!BP_IS_EMBEDDED(bp)) 2961 bp->blk_fill = fill; 2962 2963 mutex_exit(&db->db_mtx); 2964} 2965 2966/* 2967 * The SPA will call this callback several times for each zio - once 2968 * for every physical child i/o (zio->io_phys_children times). This 2969 * allows the DMU to monitor the progress of each logical i/o. For example, 2970 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z 2971 * block. There may be a long delay before all copies/fragments are completed, 2972 * so this callback allows us to retire dirty space gradually, as the physical 2973 * i/os complete. 2974 */ 2975/* ARGSUSED */ 2976static void 2977dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg) 2978{ 2979 dmu_buf_impl_t *db = arg; 2980 objset_t *os = db->db_objset; 2981 dsl_pool_t *dp = dmu_objset_pool(os); 2982 dbuf_dirty_record_t *dr; 2983 int delta = 0; 2984 2985 dr = db->db_data_pending; 2986 ASSERT3U(dr->dr_txg, ==, zio->io_txg); 2987 2988 /* 2989 * The callback will be called io_phys_children times. Retire one 2990 * portion of our dirty space each time we are called. Any rounding 2991 * error will be cleaned up by dsl_pool_sync()'s call to 2992 * dsl_pool_undirty_space(). 2993 */ 2994 delta = dr->dr_accounted / zio->io_phys_children; 2995 dsl_pool_undirty_space(dp, delta, zio->io_txg); 2996} 2997 2998/* ARGSUSED */ 2999static void 3000dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb) 3001{ 3002 dmu_buf_impl_t *db = vdb; 3003 blkptr_t *bp_orig = &zio->io_bp_orig; 3004 blkptr_t *bp = db->db_blkptr; 3005 objset_t *os = db->db_objset; 3006 dmu_tx_t *tx = os->os_synctx; 3007 dbuf_dirty_record_t **drp, *dr; 3008 3009 ASSERT0(zio->io_error); 3010 ASSERT(db->db_blkptr == bp); 3011 3012 /* 3013 * For nopwrites and rewrites we ensure that the bp matches our 3014 * original and bypass all the accounting. 3015 */ 3016 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) { 3017 ASSERT(BP_EQUAL(bp, bp_orig)); 3018 } else { 3019 dsl_dataset_t *ds = os->os_dsl_dataset; 3020 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE); 3021 dsl_dataset_block_born(ds, bp, tx); 3022 } 3023 3024 mutex_enter(&db->db_mtx); 3025 3026 DBUF_VERIFY(db); 3027 3028 drp = &db->db_last_dirty; 3029 while ((dr = *drp) != db->db_data_pending) 3030 drp = &dr->dr_next; 3031 ASSERT(!list_link_active(&dr->dr_dirty_node)); 3032 ASSERT(dr->dr_dbuf == db); 3033 ASSERT(dr->dr_next == NULL); 3034 *drp = dr->dr_next; 3035 3036#ifdef ZFS_DEBUG 3037 if (db->db_blkid == DMU_SPILL_BLKID) { 3038 dnode_t *dn; 3039 3040 DB_DNODE_ENTER(db); 3041 dn = DB_DNODE(db); 3042 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); 3043 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && 3044 db->db_blkptr == &dn->dn_phys->dn_spill); 3045 DB_DNODE_EXIT(db); 3046 } 3047#endif 3048 3049 if (db->db_level == 0) { 3050 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 3051 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); 3052 if (db->db_state != DB_NOFILL) { 3053 if (dr->dt.dl.dr_data != db->db_buf) 3054 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, 3055 db)); 3056 else if (!arc_released(db->db_buf)) 3057 arc_set_callback(db->db_buf, dbuf_do_evict, db); 3058 } 3059 } else { 3060 dnode_t *dn; 3061 3062 DB_DNODE_ENTER(db); 3063 dn = DB_DNODE(db); 3064 ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 3065 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift); 3066 if (!BP_IS_HOLE(db->db_blkptr)) { 3067 int epbs = 3068 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 3069 ASSERT3U(db->db_blkid, <=, 3070 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs)); 3071 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, 3072 db->db.db_size); 3073 if (!arc_released(db->db_buf)) 3074 arc_set_callback(db->db_buf, dbuf_do_evict, db); 3075 } 3076 DB_DNODE_EXIT(db); 3077 mutex_destroy(&dr->dt.di.dr_mtx); 3078 list_destroy(&dr->dt.di.dr_children); 3079 } 3080 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 3081 3082 cv_broadcast(&db->db_changed); 3083 ASSERT(db->db_dirtycnt > 0); 3084 db->db_dirtycnt -= 1; 3085 db->db_data_pending = NULL; 3086 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg); 3087} 3088 3089static void 3090dbuf_write_nofill_ready(zio_t *zio) 3091{ 3092 dbuf_write_ready(zio, NULL, zio->io_private); 3093} 3094 3095static void 3096dbuf_write_nofill_done(zio_t *zio) 3097{ 3098 dbuf_write_done(zio, NULL, zio->io_private); 3099} 3100 3101static void 3102dbuf_write_override_ready(zio_t *zio) 3103{ 3104 dbuf_dirty_record_t *dr = zio->io_private; 3105 dmu_buf_impl_t *db = dr->dr_dbuf; 3106 3107 dbuf_write_ready(zio, NULL, db); 3108} 3109 3110static void 3111dbuf_write_override_done(zio_t *zio) 3112{ 3113 dbuf_dirty_record_t *dr = zio->io_private; 3114 dmu_buf_impl_t *db = dr->dr_dbuf; 3115 blkptr_t *obp = &dr->dt.dl.dr_overridden_by; 3116 3117 mutex_enter(&db->db_mtx); 3118 if (!BP_EQUAL(zio->io_bp, obp)) { 3119 if (!BP_IS_HOLE(obp)) 3120 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp); 3121 arc_release(dr->dt.dl.dr_data, db); 3122 } 3123 mutex_exit(&db->db_mtx); 3124 3125 dbuf_write_done(zio, NULL, db); 3126} 3127 3128/* Issue I/O to commit a dirty buffer to disk. */ 3129static void 3130dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx) 3131{ 3132 dmu_buf_impl_t *db = dr->dr_dbuf; 3133 dnode_t *dn; 3134 objset_t *os; 3135 dmu_buf_impl_t *parent = db->db_parent; 3136 uint64_t txg = tx->tx_txg; 3137 zbookmark_phys_t zb; 3138 zio_prop_t zp; 3139 zio_t *zio; 3140 int wp_flag = 0; 3141 3142 DB_DNODE_ENTER(db); 3143 dn = DB_DNODE(db); 3144 os = dn->dn_objset; 3145 3146 if (db->db_state != DB_NOFILL) { 3147 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) { 3148 /* 3149 * Private object buffers are released here rather 3150 * than in dbuf_dirty() since they are only modified 3151 * in the syncing context and we don't want the 3152 * overhead of making multiple copies of the data. 3153 */ 3154 if (BP_IS_HOLE(db->db_blkptr)) { 3155 arc_buf_thaw(data); 3156 } else { 3157 dbuf_release_bp(db); 3158 } 3159 } 3160 } 3161 3162 if (parent != dn->dn_dbuf) { 3163 /* Our parent is an indirect block. */ 3164 /* We have a dirty parent that has been scheduled for write. */ 3165 ASSERT(parent && parent->db_data_pending); 3166 /* Our parent's buffer is one level closer to the dnode. */ 3167 ASSERT(db->db_level == parent->db_level-1); 3168 /* 3169 * We're about to modify our parent's db_data by modifying 3170 * our block pointer, so the parent must be released. 3171 */ 3172 ASSERT(arc_released(parent->db_buf)); 3173 zio = parent->db_data_pending->dr_zio; 3174 } else { 3175 /* Our parent is the dnode itself. */ 3176 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 && 3177 db->db_blkid != DMU_SPILL_BLKID) || 3178 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0)); 3179 if (db->db_blkid != DMU_SPILL_BLKID) 3180 ASSERT3P(db->db_blkptr, ==, 3181 &dn->dn_phys->dn_blkptr[db->db_blkid]); 3182 zio = dn->dn_zio; 3183 } 3184 3185 ASSERT(db->db_level == 0 || data == db->db_buf); 3186 ASSERT3U(db->db_blkptr->blk_birth, <=, txg); 3187 ASSERT(zio); 3188 3189 SET_BOOKMARK(&zb, os->os_dsl_dataset ? 3190 os->os_dsl_dataset->ds_object : DMU_META_OBJSET, 3191 db->db.db_object, db->db_level, db->db_blkid); 3192 3193 if (db->db_blkid == DMU_SPILL_BLKID) 3194 wp_flag = WP_SPILL; 3195 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0; 3196 3197 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp); 3198 DB_DNODE_EXIT(db); 3199 3200 if (db->db_level == 0 && 3201 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { 3202 /* 3203 * The BP for this block has been provided by open context 3204 * (by dmu_sync() or dmu_buf_write_embedded()). 3205 */ 3206 void *contents = (data != NULL) ? data->b_data : NULL; 3207 3208 dr->dr_zio = zio_write(zio, os->os_spa, txg, 3209 db->db_blkptr, contents, db->db.db_size, &zp, 3210 dbuf_write_override_ready, NULL, dbuf_write_override_done, 3211 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); 3212 mutex_enter(&db->db_mtx); 3213 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 3214 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by, 3215 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite); 3216 mutex_exit(&db->db_mtx); 3217 } else if (db->db_state == DB_NOFILL) { 3218 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF || 3219 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY); 3220 dr->dr_zio = zio_write(zio, os->os_spa, txg, 3221 db->db_blkptr, NULL, db->db.db_size, &zp, 3222 dbuf_write_nofill_ready, NULL, dbuf_write_nofill_done, db, 3223 ZIO_PRIORITY_ASYNC_WRITE, 3224 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb); 3225 } else { 3226 ASSERT(arc_released(data)); 3227 dr->dr_zio = arc_write(zio, os->os_spa, txg, 3228 db->db_blkptr, data, DBUF_IS_L2CACHEABLE(db), 3229 DBUF_IS_L2COMPRESSIBLE(db), &zp, dbuf_write_ready, 3230 dbuf_write_physdone, dbuf_write_done, db, 3231 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); 3232 } 3233} 3234