dbuf.c revision 304139
1/* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21/* 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23 * Copyright 2011 Nexenta Systems, Inc. All rights reserved. 24 * Copyright (c) 2012, 2015 by Delphix. All rights reserved. 25 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved. 26 * Copyright (c) 2013, Joyent, Inc. All rights reserved. 27 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved. 28 * Copyright (c) 2014 Integros [integros.com] 29 */ 30 31#include <sys/zfs_context.h> 32#include <sys/dmu.h> 33#include <sys/dmu_send.h> 34#include <sys/dmu_impl.h> 35#include <sys/dbuf.h> 36#include <sys/dmu_objset.h> 37#include <sys/dsl_dataset.h> 38#include <sys/dsl_dir.h> 39#include <sys/dmu_tx.h> 40#include <sys/spa.h> 41#include <sys/zio.h> 42#include <sys/dmu_zfetch.h> 43#include <sys/sa.h> 44#include <sys/sa_impl.h> 45#include <sys/zfeature.h> 46#include <sys/blkptr.h> 47#include <sys/range_tree.h> 48 49/* 50 * Number of times that zfs_free_range() took the slow path while doing 51 * a zfs receive. A nonzero value indicates a potential performance problem. 52 */ 53uint64_t zfs_free_range_recv_miss; 54 55static void dbuf_destroy(dmu_buf_impl_t *db); 56static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx); 57static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx); 58 59/* 60 * Global data structures and functions for the dbuf cache. 61 */ 62static kmem_cache_t *dbuf_cache; 63static taskq_t *dbu_evict_taskq; 64 65/* ARGSUSED */ 66static int 67dbuf_cons(void *vdb, void *unused, int kmflag) 68{ 69 dmu_buf_impl_t *db = vdb; 70 bzero(db, sizeof (dmu_buf_impl_t)); 71 72 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL); 73 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL); 74 refcount_create(&db->db_holds); 75 76 return (0); 77} 78 79/* ARGSUSED */ 80static void 81dbuf_dest(void *vdb, void *unused) 82{ 83 dmu_buf_impl_t *db = vdb; 84 mutex_destroy(&db->db_mtx); 85 cv_destroy(&db->db_changed); 86 refcount_destroy(&db->db_holds); 87} 88 89/* 90 * dbuf hash table routines 91 */ 92static dbuf_hash_table_t dbuf_hash_table; 93 94static uint64_t dbuf_hash_count; 95 96static uint64_t 97dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid) 98{ 99 uintptr_t osv = (uintptr_t)os; 100 uint64_t crc = -1ULL; 101 102 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 103 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF]; 104 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF]; 105 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF]; 106 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF]; 107 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF]; 108 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF]; 109 110 crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16); 111 112 return (crc); 113} 114 115#define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid); 116 117#define DBUF_EQUAL(dbuf, os, obj, level, blkid) \ 118 ((dbuf)->db.db_object == (obj) && \ 119 (dbuf)->db_objset == (os) && \ 120 (dbuf)->db_level == (level) && \ 121 (dbuf)->db_blkid == (blkid)) 122 123dmu_buf_impl_t * 124dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid) 125{ 126 dbuf_hash_table_t *h = &dbuf_hash_table; 127 uint64_t hv = DBUF_HASH(os, obj, level, blkid); 128 uint64_t idx = hv & h->hash_table_mask; 129 dmu_buf_impl_t *db; 130 131 mutex_enter(DBUF_HASH_MUTEX(h, idx)); 132 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) { 133 if (DBUF_EQUAL(db, os, obj, level, blkid)) { 134 mutex_enter(&db->db_mtx); 135 if (db->db_state != DB_EVICTING) { 136 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 137 return (db); 138 } 139 mutex_exit(&db->db_mtx); 140 } 141 } 142 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 143 return (NULL); 144} 145 146static dmu_buf_impl_t * 147dbuf_find_bonus(objset_t *os, uint64_t object) 148{ 149 dnode_t *dn; 150 dmu_buf_impl_t *db = NULL; 151 152 if (dnode_hold(os, object, FTAG, &dn) == 0) { 153 rw_enter(&dn->dn_struct_rwlock, RW_READER); 154 if (dn->dn_bonus != NULL) { 155 db = dn->dn_bonus; 156 mutex_enter(&db->db_mtx); 157 } 158 rw_exit(&dn->dn_struct_rwlock); 159 dnode_rele(dn, FTAG); 160 } 161 return (db); 162} 163 164/* 165 * Insert an entry into the hash table. If there is already an element 166 * equal to elem in the hash table, then the already existing element 167 * will be returned and the new element will not be inserted. 168 * Otherwise returns NULL. 169 */ 170static dmu_buf_impl_t * 171dbuf_hash_insert(dmu_buf_impl_t *db) 172{ 173 dbuf_hash_table_t *h = &dbuf_hash_table; 174 objset_t *os = db->db_objset; 175 uint64_t obj = db->db.db_object; 176 int level = db->db_level; 177 uint64_t blkid = db->db_blkid; 178 uint64_t hv = DBUF_HASH(os, obj, level, blkid); 179 uint64_t idx = hv & h->hash_table_mask; 180 dmu_buf_impl_t *dbf; 181 182 mutex_enter(DBUF_HASH_MUTEX(h, idx)); 183 for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) { 184 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) { 185 mutex_enter(&dbf->db_mtx); 186 if (dbf->db_state != DB_EVICTING) { 187 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 188 return (dbf); 189 } 190 mutex_exit(&dbf->db_mtx); 191 } 192 } 193 194 mutex_enter(&db->db_mtx); 195 db->db_hash_next = h->hash_table[idx]; 196 h->hash_table[idx] = db; 197 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 198 atomic_inc_64(&dbuf_hash_count); 199 200 return (NULL); 201} 202 203/* 204 * Remove an entry from the hash table. It must be in the EVICTING state. 205 */ 206static void 207dbuf_hash_remove(dmu_buf_impl_t *db) 208{ 209 dbuf_hash_table_t *h = &dbuf_hash_table; 210 uint64_t hv = DBUF_HASH(db->db_objset, db->db.db_object, 211 db->db_level, db->db_blkid); 212 uint64_t idx = hv & h->hash_table_mask; 213 dmu_buf_impl_t *dbf, **dbp; 214 215 /* 216 * We musn't hold db_mtx to maintain lock ordering: 217 * DBUF_HASH_MUTEX > db_mtx. 218 */ 219 ASSERT(refcount_is_zero(&db->db_holds)); 220 ASSERT(db->db_state == DB_EVICTING); 221 ASSERT(!MUTEX_HELD(&db->db_mtx)); 222 223 mutex_enter(DBUF_HASH_MUTEX(h, idx)); 224 dbp = &h->hash_table[idx]; 225 while ((dbf = *dbp) != db) { 226 dbp = &dbf->db_hash_next; 227 ASSERT(dbf != NULL); 228 } 229 *dbp = db->db_hash_next; 230 db->db_hash_next = NULL; 231 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 232 atomic_dec_64(&dbuf_hash_count); 233} 234 235static arc_evict_func_t dbuf_do_evict; 236 237typedef enum { 238 DBVU_EVICTING, 239 DBVU_NOT_EVICTING 240} dbvu_verify_type_t; 241 242static void 243dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type) 244{ 245#ifdef ZFS_DEBUG 246 int64_t holds; 247 248 if (db->db_user == NULL) 249 return; 250 251 /* Only data blocks support the attachment of user data. */ 252 ASSERT(db->db_level == 0); 253 254 /* Clients must resolve a dbuf before attaching user data. */ 255 ASSERT(db->db.db_data != NULL); 256 ASSERT3U(db->db_state, ==, DB_CACHED); 257 258 holds = refcount_count(&db->db_holds); 259 if (verify_type == DBVU_EVICTING) { 260 /* 261 * Immediate eviction occurs when holds == dirtycnt. 262 * For normal eviction buffers, holds is zero on 263 * eviction, except when dbuf_fix_old_data() calls 264 * dbuf_clear_data(). However, the hold count can grow 265 * during eviction even though db_mtx is held (see 266 * dmu_bonus_hold() for an example), so we can only 267 * test the generic invariant that holds >= dirtycnt. 268 */ 269 ASSERT3U(holds, >=, db->db_dirtycnt); 270 } else { 271 if (db->db_user_immediate_evict == TRUE) 272 ASSERT3U(holds, >=, db->db_dirtycnt); 273 else 274 ASSERT3U(holds, >, 0); 275 } 276#endif 277} 278 279static void 280dbuf_evict_user(dmu_buf_impl_t *db) 281{ 282 dmu_buf_user_t *dbu = db->db_user; 283 284 ASSERT(MUTEX_HELD(&db->db_mtx)); 285 286 if (dbu == NULL) 287 return; 288 289 dbuf_verify_user(db, DBVU_EVICTING); 290 db->db_user = NULL; 291 292#ifdef ZFS_DEBUG 293 if (dbu->dbu_clear_on_evict_dbufp != NULL) 294 *dbu->dbu_clear_on_evict_dbufp = NULL; 295#endif 296 297 /* 298 * Invoke the callback from a taskq to avoid lock order reversals 299 * and limit stack depth. 300 */ 301 taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func, dbu, 0, 302 &dbu->dbu_tqent); 303} 304 305boolean_t 306dbuf_is_metadata(dmu_buf_impl_t *db) 307{ 308 if (db->db_level > 0) { 309 return (B_TRUE); 310 } else { 311 boolean_t is_metadata; 312 313 DB_DNODE_ENTER(db); 314 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type); 315 DB_DNODE_EXIT(db); 316 317 return (is_metadata); 318 } 319} 320 321void 322dbuf_evict(dmu_buf_impl_t *db) 323{ 324 ASSERT(MUTEX_HELD(&db->db_mtx)); 325 ASSERT(db->db_buf == NULL); 326 ASSERT(db->db_data_pending == NULL); 327 328 dbuf_clear(db); 329 dbuf_destroy(db); 330} 331 332void 333dbuf_init(void) 334{ 335 uint64_t hsize = 1ULL << 16; 336 dbuf_hash_table_t *h = &dbuf_hash_table; 337 int i; 338 339 /* 340 * The hash table is big enough to fill all of physical memory 341 * with an average 4K block size. The table will take up 342 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers). 343 */ 344 while (hsize * 4096 < (uint64_t)physmem * PAGESIZE) 345 hsize <<= 1; 346 347retry: 348 h->hash_table_mask = hsize - 1; 349 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP); 350 if (h->hash_table == NULL) { 351 /* XXX - we should really return an error instead of assert */ 352 ASSERT(hsize > (1ULL << 10)); 353 hsize >>= 1; 354 goto retry; 355 } 356 357 dbuf_cache = kmem_cache_create("dmu_buf_impl_t", 358 sizeof (dmu_buf_impl_t), 359 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0); 360 361 for (i = 0; i < DBUF_MUTEXES; i++) 362 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL); 363 364 /* 365 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc 366 * configuration is not required. 367 */ 368 dbu_evict_taskq = taskq_create("dbu_evict", 1, minclsyspri, 0, 0, 0); 369} 370 371void 372dbuf_fini(void) 373{ 374 dbuf_hash_table_t *h = &dbuf_hash_table; 375 int i; 376 377 for (i = 0; i < DBUF_MUTEXES; i++) 378 mutex_destroy(&h->hash_mutexes[i]); 379 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *)); 380 kmem_cache_destroy(dbuf_cache); 381 taskq_destroy(dbu_evict_taskq); 382} 383 384/* 385 * Other stuff. 386 */ 387 388#ifdef ZFS_DEBUG 389static void 390dbuf_verify(dmu_buf_impl_t *db) 391{ 392 dnode_t *dn; 393 dbuf_dirty_record_t *dr; 394 395 ASSERT(MUTEX_HELD(&db->db_mtx)); 396 397 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY)) 398 return; 399 400 ASSERT(db->db_objset != NULL); 401 DB_DNODE_ENTER(db); 402 dn = DB_DNODE(db); 403 if (dn == NULL) { 404 ASSERT(db->db_parent == NULL); 405 ASSERT(db->db_blkptr == NULL); 406 } else { 407 ASSERT3U(db->db.db_object, ==, dn->dn_object); 408 ASSERT3P(db->db_objset, ==, dn->dn_objset); 409 ASSERT3U(db->db_level, <, dn->dn_nlevels); 410 ASSERT(db->db_blkid == DMU_BONUS_BLKID || 411 db->db_blkid == DMU_SPILL_BLKID || 412 !avl_is_empty(&dn->dn_dbufs)); 413 } 414 if (db->db_blkid == DMU_BONUS_BLKID) { 415 ASSERT(dn != NULL); 416 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 417 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID); 418 } else if (db->db_blkid == DMU_SPILL_BLKID) { 419 ASSERT(dn != NULL); 420 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 421 ASSERT0(db->db.db_offset); 422 } else { 423 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size); 424 } 425 426 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next) 427 ASSERT(dr->dr_dbuf == db); 428 429 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next) 430 ASSERT(dr->dr_dbuf == db); 431 432 /* 433 * We can't assert that db_size matches dn_datablksz because it 434 * can be momentarily different when another thread is doing 435 * dnode_set_blksz(). 436 */ 437 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) { 438 dr = db->db_data_pending; 439 /* 440 * It should only be modified in syncing context, so 441 * make sure we only have one copy of the data. 442 */ 443 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf); 444 } 445 446 /* verify db->db_blkptr */ 447 if (db->db_blkptr) { 448 if (db->db_parent == dn->dn_dbuf) { 449 /* db is pointed to by the dnode */ 450 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */ 451 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object)) 452 ASSERT(db->db_parent == NULL); 453 else 454 ASSERT(db->db_parent != NULL); 455 if (db->db_blkid != DMU_SPILL_BLKID) 456 ASSERT3P(db->db_blkptr, ==, 457 &dn->dn_phys->dn_blkptr[db->db_blkid]); 458 } else { 459 /* db is pointed to by an indirect block */ 460 int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT; 461 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1); 462 ASSERT3U(db->db_parent->db.db_object, ==, 463 db->db.db_object); 464 /* 465 * dnode_grow_indblksz() can make this fail if we don't 466 * have the struct_rwlock. XXX indblksz no longer 467 * grows. safe to do this now? 468 */ 469 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) { 470 ASSERT3P(db->db_blkptr, ==, 471 ((blkptr_t *)db->db_parent->db.db_data + 472 db->db_blkid % epb)); 473 } 474 } 475 } 476 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) && 477 (db->db_buf == NULL || db->db_buf->b_data) && 478 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID && 479 db->db_state != DB_FILL && !dn->dn_free_txg) { 480 /* 481 * If the blkptr isn't set but they have nonzero data, 482 * it had better be dirty, otherwise we'll lose that 483 * data when we evict this buffer. 484 * 485 * There is an exception to this rule for indirect blocks; in 486 * this case, if the indirect block is a hole, we fill in a few 487 * fields on each of the child blocks (importantly, birth time) 488 * to prevent hole birth times from being lost when you 489 * partially fill in a hole. 490 */ 491 if (db->db_dirtycnt == 0) { 492 if (db->db_level == 0) { 493 uint64_t *buf = db->db.db_data; 494 int i; 495 496 for (i = 0; i < db->db.db_size >> 3; i++) { 497 ASSERT(buf[i] == 0); 498 } 499 } else { 500 blkptr_t *bps = db->db.db_data; 501 ASSERT3U(1 << DB_DNODE(db)->dn_indblkshift, ==, 502 db->db.db_size); 503 /* 504 * We want to verify that all the blkptrs in the 505 * indirect block are holes, but we may have 506 * automatically set up a few fields for them. 507 * We iterate through each blkptr and verify 508 * they only have those fields set. 509 */ 510 for (int i = 0; 511 i < db->db.db_size / sizeof (blkptr_t); 512 i++) { 513 blkptr_t *bp = &bps[i]; 514 ASSERT(ZIO_CHECKSUM_IS_ZERO( 515 &bp->blk_cksum)); 516 ASSERT( 517 DVA_IS_EMPTY(&bp->blk_dva[0]) && 518 DVA_IS_EMPTY(&bp->blk_dva[1]) && 519 DVA_IS_EMPTY(&bp->blk_dva[2])); 520 ASSERT0(bp->blk_fill); 521 ASSERT0(bp->blk_pad[0]); 522 ASSERT0(bp->blk_pad[1]); 523 ASSERT(!BP_IS_EMBEDDED(bp)); 524 ASSERT(BP_IS_HOLE(bp)); 525 ASSERT0(bp->blk_phys_birth); 526 } 527 } 528 } 529 } 530 DB_DNODE_EXIT(db); 531} 532#endif 533 534static void 535dbuf_clear_data(dmu_buf_impl_t *db) 536{ 537 ASSERT(MUTEX_HELD(&db->db_mtx)); 538 dbuf_evict_user(db); 539 db->db_buf = NULL; 540 db->db.db_data = NULL; 541 if (db->db_state != DB_NOFILL) 542 db->db_state = DB_UNCACHED; 543} 544 545static void 546dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf) 547{ 548 ASSERT(MUTEX_HELD(&db->db_mtx)); 549 ASSERT(buf != NULL); 550 551 db->db_buf = buf; 552 ASSERT(buf->b_data != NULL); 553 db->db.db_data = buf->b_data; 554 if (!arc_released(buf)) 555 arc_set_callback(buf, dbuf_do_evict, db); 556} 557 558/* 559 * Loan out an arc_buf for read. Return the loaned arc_buf. 560 */ 561arc_buf_t * 562dbuf_loan_arcbuf(dmu_buf_impl_t *db) 563{ 564 arc_buf_t *abuf; 565 566 mutex_enter(&db->db_mtx); 567 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) { 568 int blksz = db->db.db_size; 569 spa_t *spa = db->db_objset->os_spa; 570 571 mutex_exit(&db->db_mtx); 572 abuf = arc_loan_buf(spa, blksz); 573 bcopy(db->db.db_data, abuf->b_data, blksz); 574 } else { 575 abuf = db->db_buf; 576 arc_loan_inuse_buf(abuf, db); 577 dbuf_clear_data(db); 578 mutex_exit(&db->db_mtx); 579 } 580 return (abuf); 581} 582 583/* 584 * Calculate which level n block references the data at the level 0 offset 585 * provided. 586 */ 587uint64_t 588dbuf_whichblock(dnode_t *dn, int64_t level, uint64_t offset) 589{ 590 if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) { 591 /* 592 * The level n blkid is equal to the level 0 blkid divided by 593 * the number of level 0s in a level n block. 594 * 595 * The level 0 blkid is offset >> datablkshift = 596 * offset / 2^datablkshift. 597 * 598 * The number of level 0s in a level n is the number of block 599 * pointers in an indirect block, raised to the power of level. 600 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level = 601 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)). 602 * 603 * Thus, the level n blkid is: offset / 604 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT))) 605 * = offset / 2^(datablkshift + level * 606 * (indblkshift - SPA_BLKPTRSHIFT)) 607 * = offset >> (datablkshift + level * 608 * (indblkshift - SPA_BLKPTRSHIFT)) 609 */ 610 return (offset >> (dn->dn_datablkshift + level * 611 (dn->dn_indblkshift - SPA_BLKPTRSHIFT))); 612 } else { 613 ASSERT3U(offset, <, dn->dn_datablksz); 614 return (0); 615 } 616} 617 618static void 619dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb) 620{ 621 dmu_buf_impl_t *db = vdb; 622 623 mutex_enter(&db->db_mtx); 624 ASSERT3U(db->db_state, ==, DB_READ); 625 /* 626 * All reads are synchronous, so we must have a hold on the dbuf 627 */ 628 ASSERT(refcount_count(&db->db_holds) > 0); 629 ASSERT(db->db_buf == NULL); 630 ASSERT(db->db.db_data == NULL); 631 if (db->db_level == 0 && db->db_freed_in_flight) { 632 /* we were freed in flight; disregard any error */ 633 arc_release(buf, db); 634 bzero(buf->b_data, db->db.db_size); 635 arc_buf_freeze(buf); 636 db->db_freed_in_flight = FALSE; 637 dbuf_set_data(db, buf); 638 db->db_state = DB_CACHED; 639 } else if (zio == NULL || zio->io_error == 0) { 640 dbuf_set_data(db, buf); 641 db->db_state = DB_CACHED; 642 } else { 643 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 644 ASSERT3P(db->db_buf, ==, NULL); 645 VERIFY(arc_buf_remove_ref(buf, db)); 646 db->db_state = DB_UNCACHED; 647 } 648 cv_broadcast(&db->db_changed); 649 dbuf_rele_and_unlock(db, NULL); 650} 651 652static void 653dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) 654{ 655 dnode_t *dn; 656 zbookmark_phys_t zb; 657 arc_flags_t aflags = ARC_FLAG_NOWAIT; 658 659 DB_DNODE_ENTER(db); 660 dn = DB_DNODE(db); 661 ASSERT(!refcount_is_zero(&db->db_holds)); 662 /* We need the struct_rwlock to prevent db_blkptr from changing. */ 663 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 664 ASSERT(MUTEX_HELD(&db->db_mtx)); 665 ASSERT(db->db_state == DB_UNCACHED); 666 ASSERT(db->db_buf == NULL); 667 668 if (db->db_blkid == DMU_BONUS_BLKID) { 669 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen); 670 671 ASSERT3U(bonuslen, <=, db->db.db_size); 672 db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN); 673 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 674 if (bonuslen < DN_MAX_BONUSLEN) 675 bzero(db->db.db_data, DN_MAX_BONUSLEN); 676 if (bonuslen) 677 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen); 678 DB_DNODE_EXIT(db); 679 db->db_state = DB_CACHED; 680 mutex_exit(&db->db_mtx); 681 return; 682 } 683 684 /* 685 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync() 686 * processes the delete record and clears the bp while we are waiting 687 * for the dn_mtx (resulting in a "no" from block_freed). 688 */ 689 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) || 690 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) || 691 BP_IS_HOLE(db->db_blkptr)))) { 692 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 693 694 dbuf_set_data(db, arc_buf_alloc(db->db_objset->os_spa, 695 db->db.db_size, db, type)); 696 bzero(db->db.db_data, db->db.db_size); 697 698 if (db->db_blkptr != NULL && db->db_level > 0 && 699 BP_IS_HOLE(db->db_blkptr) && 700 db->db_blkptr->blk_birth != 0) { 701 blkptr_t *bps = db->db.db_data; 702 for (int i = 0; i < ((1 << 703 DB_DNODE(db)->dn_indblkshift) / sizeof (blkptr_t)); 704 i++) { 705 blkptr_t *bp = &bps[i]; 706 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, 707 1 << dn->dn_indblkshift); 708 BP_SET_LSIZE(bp, 709 BP_GET_LEVEL(db->db_blkptr) == 1 ? 710 dn->dn_datablksz : 711 BP_GET_LSIZE(db->db_blkptr)); 712 BP_SET_TYPE(bp, BP_GET_TYPE(db->db_blkptr)); 713 BP_SET_LEVEL(bp, 714 BP_GET_LEVEL(db->db_blkptr) - 1); 715 BP_SET_BIRTH(bp, db->db_blkptr->blk_birth, 0); 716 } 717 } 718 DB_DNODE_EXIT(db); 719 db->db_state = DB_CACHED; 720 mutex_exit(&db->db_mtx); 721 return; 722 } 723 724 DB_DNODE_EXIT(db); 725 726 db->db_state = DB_READ; 727 mutex_exit(&db->db_mtx); 728 729 if (DBUF_IS_L2CACHEABLE(db)) 730 aflags |= ARC_FLAG_L2CACHE; 731 if (DBUF_IS_L2COMPRESSIBLE(db)) 732 aflags |= ARC_FLAG_L2COMPRESS; 733 734 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ? 735 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET, 736 db->db.db_object, db->db_level, db->db_blkid); 737 738 dbuf_add_ref(db, NULL); 739 740 (void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr, 741 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ, 742 (flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED, 743 &aflags, &zb); 744} 745 746int 747dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) 748{ 749 int err = 0; 750 boolean_t havepzio = (zio != NULL); 751 boolean_t prefetch; 752 dnode_t *dn; 753 754 /* 755 * We don't have to hold the mutex to check db_state because it 756 * can't be freed while we have a hold on the buffer. 757 */ 758 ASSERT(!refcount_is_zero(&db->db_holds)); 759 760 if (db->db_state == DB_NOFILL) 761 return (SET_ERROR(EIO)); 762 763 DB_DNODE_ENTER(db); 764 dn = DB_DNODE(db); 765 if ((flags & DB_RF_HAVESTRUCT) == 0) 766 rw_enter(&dn->dn_struct_rwlock, RW_READER); 767 768 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 769 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL && 770 DBUF_IS_CACHEABLE(db); 771 772 mutex_enter(&db->db_mtx); 773 if (db->db_state == DB_CACHED) { 774 mutex_exit(&db->db_mtx); 775 if (prefetch) 776 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE); 777 if ((flags & DB_RF_HAVESTRUCT) == 0) 778 rw_exit(&dn->dn_struct_rwlock); 779 DB_DNODE_EXIT(db); 780 } else if (db->db_state == DB_UNCACHED) { 781 spa_t *spa = dn->dn_objset->os_spa; 782 783 if (zio == NULL) 784 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); 785 dbuf_read_impl(db, zio, flags); 786 787 /* dbuf_read_impl has dropped db_mtx for us */ 788 789 if (prefetch) 790 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE); 791 792 if ((flags & DB_RF_HAVESTRUCT) == 0) 793 rw_exit(&dn->dn_struct_rwlock); 794 DB_DNODE_EXIT(db); 795 796 if (!havepzio) 797 err = zio_wait(zio); 798 } else { 799 /* 800 * Another reader came in while the dbuf was in flight 801 * between UNCACHED and CACHED. Either a writer will finish 802 * writing the buffer (sending the dbuf to CACHED) or the 803 * first reader's request will reach the read_done callback 804 * and send the dbuf to CACHED. Otherwise, a failure 805 * occurred and the dbuf went to UNCACHED. 806 */ 807 mutex_exit(&db->db_mtx); 808 if (prefetch) 809 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1, B_TRUE); 810 if ((flags & DB_RF_HAVESTRUCT) == 0) 811 rw_exit(&dn->dn_struct_rwlock); 812 DB_DNODE_EXIT(db); 813 814 /* Skip the wait per the caller's request. */ 815 mutex_enter(&db->db_mtx); 816 if ((flags & DB_RF_NEVERWAIT) == 0) { 817 while (db->db_state == DB_READ || 818 db->db_state == DB_FILL) { 819 ASSERT(db->db_state == DB_READ || 820 (flags & DB_RF_HAVESTRUCT) == 0); 821 DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *, 822 db, zio_t *, zio); 823 cv_wait(&db->db_changed, &db->db_mtx); 824 } 825 if (db->db_state == DB_UNCACHED) 826 err = SET_ERROR(EIO); 827 } 828 mutex_exit(&db->db_mtx); 829 } 830 831 ASSERT(err || havepzio || db->db_state == DB_CACHED); 832 return (err); 833} 834 835static void 836dbuf_noread(dmu_buf_impl_t *db) 837{ 838 ASSERT(!refcount_is_zero(&db->db_holds)); 839 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 840 mutex_enter(&db->db_mtx); 841 while (db->db_state == DB_READ || db->db_state == DB_FILL) 842 cv_wait(&db->db_changed, &db->db_mtx); 843 if (db->db_state == DB_UNCACHED) { 844 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 845 spa_t *spa = db->db_objset->os_spa; 846 847 ASSERT(db->db_buf == NULL); 848 ASSERT(db->db.db_data == NULL); 849 dbuf_set_data(db, arc_buf_alloc(spa, db->db.db_size, db, type)); 850 db->db_state = DB_FILL; 851 } else if (db->db_state == DB_NOFILL) { 852 dbuf_clear_data(db); 853 } else { 854 ASSERT3U(db->db_state, ==, DB_CACHED); 855 } 856 mutex_exit(&db->db_mtx); 857} 858 859/* 860 * This is our just-in-time copy function. It makes a copy of 861 * buffers, that have been modified in a previous transaction 862 * group, before we modify them in the current active group. 863 * 864 * This function is used in two places: when we are dirtying a 865 * buffer for the first time in a txg, and when we are freeing 866 * a range in a dnode that includes this buffer. 867 * 868 * Note that when we are called from dbuf_free_range() we do 869 * not put a hold on the buffer, we just traverse the active 870 * dbuf list for the dnode. 871 */ 872static void 873dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg) 874{ 875 dbuf_dirty_record_t *dr = db->db_last_dirty; 876 877 ASSERT(MUTEX_HELD(&db->db_mtx)); 878 ASSERT(db->db.db_data != NULL); 879 ASSERT(db->db_level == 0); 880 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT); 881 882 if (dr == NULL || 883 (dr->dt.dl.dr_data != 884 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf))) 885 return; 886 887 /* 888 * If the last dirty record for this dbuf has not yet synced 889 * and its referencing the dbuf data, either: 890 * reset the reference to point to a new copy, 891 * or (if there a no active holders) 892 * just null out the current db_data pointer. 893 */ 894 ASSERT(dr->dr_txg >= txg - 2); 895 if (db->db_blkid == DMU_BONUS_BLKID) { 896 /* Note that the data bufs here are zio_bufs */ 897 dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN); 898 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 899 bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN); 900 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) { 901 int size = db->db.db_size; 902 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 903 spa_t *spa = db->db_objset->os_spa; 904 905 dr->dt.dl.dr_data = arc_buf_alloc(spa, size, db, type); 906 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size); 907 } else { 908 dbuf_clear_data(db); 909 } 910} 911 912void 913dbuf_unoverride(dbuf_dirty_record_t *dr) 914{ 915 dmu_buf_impl_t *db = dr->dr_dbuf; 916 blkptr_t *bp = &dr->dt.dl.dr_overridden_by; 917 uint64_t txg = dr->dr_txg; 918 919 ASSERT(MUTEX_HELD(&db->db_mtx)); 920 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC); 921 ASSERT(db->db_level == 0); 922 923 if (db->db_blkid == DMU_BONUS_BLKID || 924 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN) 925 return; 926 927 ASSERT(db->db_data_pending != dr); 928 929 /* free this block */ 930 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite) 931 zio_free(db->db_objset->os_spa, txg, bp); 932 933 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 934 dr->dt.dl.dr_nopwrite = B_FALSE; 935 936 /* 937 * Release the already-written buffer, so we leave it in 938 * a consistent dirty state. Note that all callers are 939 * modifying the buffer, so they will immediately do 940 * another (redundant) arc_release(). Therefore, leave 941 * the buf thawed to save the effort of freezing & 942 * immediately re-thawing it. 943 */ 944 arc_release(dr->dt.dl.dr_data, db); 945} 946 947/* 948 * Evict (if its unreferenced) or clear (if its referenced) any level-0 949 * data blocks in the free range, so that any future readers will find 950 * empty blocks. 951 * 952 * This is a no-op if the dataset is in the middle of an incremental 953 * receive; see comment below for details. 954 */ 955void 956dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid, 957 dmu_tx_t *tx) 958{ 959 dmu_buf_impl_t db_search; 960 dmu_buf_impl_t *db, *db_next; 961 uint64_t txg = tx->tx_txg; 962 avl_index_t where; 963 964 if (end_blkid > dn->dn_maxblkid && (end_blkid != DMU_SPILL_BLKID)) 965 end_blkid = dn->dn_maxblkid; 966 dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid); 967 968 db_search.db_level = 0; 969 db_search.db_blkid = start_blkid; 970 db_search.db_state = DB_SEARCH; 971 972 mutex_enter(&dn->dn_dbufs_mtx); 973 if (start_blkid >= dn->dn_unlisted_l0_blkid) { 974 /* There can't be any dbufs in this range; no need to search. */ 975#ifdef DEBUG 976 db = avl_find(&dn->dn_dbufs, &db_search, &where); 977 ASSERT3P(db, ==, NULL); 978 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER); 979 ASSERT(db == NULL || db->db_level > 0); 980#endif 981 mutex_exit(&dn->dn_dbufs_mtx); 982 return; 983 } else if (dmu_objset_is_receiving(dn->dn_objset)) { 984 /* 985 * If we are receiving, we expect there to be no dbufs in 986 * the range to be freed, because receive modifies each 987 * block at most once, and in offset order. If this is 988 * not the case, it can lead to performance problems, 989 * so note that we unexpectedly took the slow path. 990 */ 991 atomic_inc_64(&zfs_free_range_recv_miss); 992 } 993 994 db = avl_find(&dn->dn_dbufs, &db_search, &where); 995 ASSERT3P(db, ==, NULL); 996 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER); 997 998 for (; db != NULL; db = db_next) { 999 db_next = AVL_NEXT(&dn->dn_dbufs, db); 1000 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1001 1002 if (db->db_level != 0 || db->db_blkid > end_blkid) { 1003 break; 1004 } 1005 ASSERT3U(db->db_blkid, >=, start_blkid); 1006 1007 /* found a level 0 buffer in the range */ 1008 mutex_enter(&db->db_mtx); 1009 if (dbuf_undirty(db, tx)) { 1010 /* mutex has been dropped and dbuf destroyed */ 1011 continue; 1012 } 1013 1014 if (db->db_state == DB_UNCACHED || 1015 db->db_state == DB_NOFILL || 1016 db->db_state == DB_EVICTING) { 1017 ASSERT(db->db.db_data == NULL); 1018 mutex_exit(&db->db_mtx); 1019 continue; 1020 } 1021 if (db->db_state == DB_READ || db->db_state == DB_FILL) { 1022 /* will be handled in dbuf_read_done or dbuf_rele */ 1023 db->db_freed_in_flight = TRUE; 1024 mutex_exit(&db->db_mtx); 1025 continue; 1026 } 1027 if (refcount_count(&db->db_holds) == 0) { 1028 ASSERT(db->db_buf); 1029 dbuf_clear(db); 1030 continue; 1031 } 1032 /* The dbuf is referenced */ 1033 1034 if (db->db_last_dirty != NULL) { 1035 dbuf_dirty_record_t *dr = db->db_last_dirty; 1036 1037 if (dr->dr_txg == txg) { 1038 /* 1039 * This buffer is "in-use", re-adjust the file 1040 * size to reflect that this buffer may 1041 * contain new data when we sync. 1042 */ 1043 if (db->db_blkid != DMU_SPILL_BLKID && 1044 db->db_blkid > dn->dn_maxblkid) 1045 dn->dn_maxblkid = db->db_blkid; 1046 dbuf_unoverride(dr); 1047 } else { 1048 /* 1049 * This dbuf is not dirty in the open context. 1050 * Either uncache it (if its not referenced in 1051 * the open context) or reset its contents to 1052 * empty. 1053 */ 1054 dbuf_fix_old_data(db, txg); 1055 } 1056 } 1057 /* clear the contents if its cached */ 1058 if (db->db_state == DB_CACHED) { 1059 ASSERT(db->db.db_data != NULL); 1060 arc_release(db->db_buf, db); 1061 bzero(db->db.db_data, db->db.db_size); 1062 arc_buf_freeze(db->db_buf); 1063 } 1064 1065 mutex_exit(&db->db_mtx); 1066 } 1067 mutex_exit(&dn->dn_dbufs_mtx); 1068} 1069 1070static int 1071dbuf_block_freeable(dmu_buf_impl_t *db) 1072{ 1073 dsl_dataset_t *ds = db->db_objset->os_dsl_dataset; 1074 uint64_t birth_txg = 0; 1075 1076 /* 1077 * We don't need any locking to protect db_blkptr: 1078 * If it's syncing, then db_last_dirty will be set 1079 * so we'll ignore db_blkptr. 1080 * 1081 * This logic ensures that only block births for 1082 * filled blocks are considered. 1083 */ 1084 ASSERT(MUTEX_HELD(&db->db_mtx)); 1085 if (db->db_last_dirty && (db->db_blkptr == NULL || 1086 !BP_IS_HOLE(db->db_blkptr))) { 1087 birth_txg = db->db_last_dirty->dr_txg; 1088 } else if (db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) { 1089 birth_txg = db->db_blkptr->blk_birth; 1090 } 1091 1092 /* 1093 * If this block don't exist or is in a snapshot, it can't be freed. 1094 * Don't pass the bp to dsl_dataset_block_freeable() since we 1095 * are holding the db_mtx lock and might deadlock if we are 1096 * prefetching a dedup-ed block. 1097 */ 1098 if (birth_txg != 0) 1099 return (ds == NULL || 1100 dsl_dataset_block_freeable(ds, NULL, birth_txg)); 1101 else 1102 return (B_FALSE); 1103} 1104 1105void 1106dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx) 1107{ 1108 arc_buf_t *buf, *obuf; 1109 int osize = db->db.db_size; 1110 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 1111 dnode_t *dn; 1112 1113 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1114 1115 DB_DNODE_ENTER(db); 1116 dn = DB_DNODE(db); 1117 1118 /* XXX does *this* func really need the lock? */ 1119 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 1120 1121 /* 1122 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held 1123 * is OK, because there can be no other references to the db 1124 * when we are changing its size, so no concurrent DB_FILL can 1125 * be happening. 1126 */ 1127 /* 1128 * XXX we should be doing a dbuf_read, checking the return 1129 * value and returning that up to our callers 1130 */ 1131 dmu_buf_will_dirty(&db->db, tx); 1132 1133 /* create the data buffer for the new block */ 1134 buf = arc_buf_alloc(dn->dn_objset->os_spa, size, db, type); 1135 1136 /* copy old block data to the new block */ 1137 obuf = db->db_buf; 1138 bcopy(obuf->b_data, buf->b_data, MIN(osize, size)); 1139 /* zero the remainder */ 1140 if (size > osize) 1141 bzero((uint8_t *)buf->b_data + osize, size - osize); 1142 1143 mutex_enter(&db->db_mtx); 1144 dbuf_set_data(db, buf); 1145 VERIFY(arc_buf_remove_ref(obuf, db)); 1146 db->db.db_size = size; 1147 1148 if (db->db_level == 0) { 1149 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); 1150 db->db_last_dirty->dt.dl.dr_data = buf; 1151 } 1152 mutex_exit(&db->db_mtx); 1153 1154 dnode_willuse_space(dn, size-osize, tx); 1155 DB_DNODE_EXIT(db); 1156} 1157 1158void 1159dbuf_release_bp(dmu_buf_impl_t *db) 1160{ 1161 objset_t *os = db->db_objset; 1162 1163 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os))); 1164 ASSERT(arc_released(os->os_phys_buf) || 1165 list_link_active(&os->os_dsl_dataset->ds_synced_link)); 1166 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf)); 1167 1168 (void) arc_release(db->db_buf, db); 1169} 1170 1171/* 1172 * We already have a dirty record for this TXG, and we are being 1173 * dirtied again. 1174 */ 1175static void 1176dbuf_redirty(dbuf_dirty_record_t *dr) 1177{ 1178 dmu_buf_impl_t *db = dr->dr_dbuf; 1179 1180 ASSERT(MUTEX_HELD(&db->db_mtx)); 1181 1182 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) { 1183 /* 1184 * If this buffer has already been written out, 1185 * we now need to reset its state. 1186 */ 1187 dbuf_unoverride(dr); 1188 if (db->db.db_object != DMU_META_DNODE_OBJECT && 1189 db->db_state != DB_NOFILL) { 1190 /* Already released on initial dirty, so just thaw. */ 1191 ASSERT(arc_released(db->db_buf)); 1192 arc_buf_thaw(db->db_buf); 1193 } 1194 } 1195} 1196 1197dbuf_dirty_record_t * 1198dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1199{ 1200 dnode_t *dn; 1201 objset_t *os; 1202 dbuf_dirty_record_t **drp, *dr; 1203 int drop_struct_lock = FALSE; 1204 boolean_t do_free_accounting = B_FALSE; 1205 int txgoff = tx->tx_txg & TXG_MASK; 1206 1207 ASSERT(tx->tx_txg != 0); 1208 ASSERT(!refcount_is_zero(&db->db_holds)); 1209 DMU_TX_DIRTY_BUF(tx, db); 1210 1211 DB_DNODE_ENTER(db); 1212 dn = DB_DNODE(db); 1213 /* 1214 * Shouldn't dirty a regular buffer in syncing context. Private 1215 * objects may be dirtied in syncing context, but only if they 1216 * were already pre-dirtied in open context. 1217 */ 1218 ASSERT(!dmu_tx_is_syncing(tx) || 1219 BP_IS_HOLE(dn->dn_objset->os_rootbp) || 1220 DMU_OBJECT_IS_SPECIAL(dn->dn_object) || 1221 dn->dn_objset->os_dsl_dataset == NULL); 1222 /* 1223 * We make this assert for private objects as well, but after we 1224 * check if we're already dirty. They are allowed to re-dirty 1225 * in syncing context. 1226 */ 1227 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 1228 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 1229 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 1230 1231 mutex_enter(&db->db_mtx); 1232 /* 1233 * XXX make this true for indirects too? The problem is that 1234 * transactions created with dmu_tx_create_assigned() from 1235 * syncing context don't bother holding ahead. 1236 */ 1237 ASSERT(db->db_level != 0 || 1238 db->db_state == DB_CACHED || db->db_state == DB_FILL || 1239 db->db_state == DB_NOFILL); 1240 1241 mutex_enter(&dn->dn_mtx); 1242 /* 1243 * Don't set dirtyctx to SYNC if we're just modifying this as we 1244 * initialize the objset. 1245 */ 1246 if (dn->dn_dirtyctx == DN_UNDIRTIED && 1247 !BP_IS_HOLE(dn->dn_objset->os_rootbp)) { 1248 dn->dn_dirtyctx = 1249 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN); 1250 ASSERT(dn->dn_dirtyctx_firstset == NULL); 1251 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP); 1252 } 1253 mutex_exit(&dn->dn_mtx); 1254 1255 if (db->db_blkid == DMU_SPILL_BLKID) 1256 dn->dn_have_spill = B_TRUE; 1257 1258 /* 1259 * If this buffer is already dirty, we're done. 1260 */ 1261 drp = &db->db_last_dirty; 1262 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg || 1263 db->db.db_object == DMU_META_DNODE_OBJECT); 1264 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg) 1265 drp = &dr->dr_next; 1266 if (dr && dr->dr_txg == tx->tx_txg) { 1267 DB_DNODE_EXIT(db); 1268 1269 dbuf_redirty(dr); 1270 mutex_exit(&db->db_mtx); 1271 return (dr); 1272 } 1273 1274 /* 1275 * Only valid if not already dirty. 1276 */ 1277 ASSERT(dn->dn_object == 0 || 1278 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 1279 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 1280 1281 ASSERT3U(dn->dn_nlevels, >, db->db_level); 1282 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) || 1283 dn->dn_phys->dn_nlevels > db->db_level || 1284 dn->dn_next_nlevels[txgoff] > db->db_level || 1285 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level || 1286 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level); 1287 1288 /* 1289 * We should only be dirtying in syncing context if it's the 1290 * mos or we're initializing the os or it's a special object. 1291 * However, we are allowed to dirty in syncing context provided 1292 * we already dirtied it in open context. Hence we must make 1293 * this assertion only if we're not already dirty. 1294 */ 1295 os = dn->dn_objset; 1296 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) || 1297 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp)); 1298 ASSERT(db->db.db_size != 0); 1299 1300 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 1301 1302 if (db->db_blkid != DMU_BONUS_BLKID) { 1303 /* 1304 * Update the accounting. 1305 * Note: we delay "free accounting" until after we drop 1306 * the db_mtx. This keeps us from grabbing other locks 1307 * (and possibly deadlocking) in bp_get_dsize() while 1308 * also holding the db_mtx. 1309 */ 1310 dnode_willuse_space(dn, db->db.db_size, tx); 1311 do_free_accounting = dbuf_block_freeable(db); 1312 } 1313 1314 /* 1315 * If this buffer is dirty in an old transaction group we need 1316 * to make a copy of it so that the changes we make in this 1317 * transaction group won't leak out when we sync the older txg. 1318 */ 1319 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP); 1320 if (db->db_level == 0) { 1321 void *data_old = db->db_buf; 1322 1323 if (db->db_state != DB_NOFILL) { 1324 if (db->db_blkid == DMU_BONUS_BLKID) { 1325 dbuf_fix_old_data(db, tx->tx_txg); 1326 data_old = db->db.db_data; 1327 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) { 1328 /* 1329 * Release the data buffer from the cache so 1330 * that we can modify it without impacting 1331 * possible other users of this cached data 1332 * block. Note that indirect blocks and 1333 * private objects are not released until the 1334 * syncing state (since they are only modified 1335 * then). 1336 */ 1337 arc_release(db->db_buf, db); 1338 dbuf_fix_old_data(db, tx->tx_txg); 1339 data_old = db->db_buf; 1340 } 1341 ASSERT(data_old != NULL); 1342 } 1343 dr->dt.dl.dr_data = data_old; 1344 } else { 1345 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL); 1346 list_create(&dr->dt.di.dr_children, 1347 sizeof (dbuf_dirty_record_t), 1348 offsetof(dbuf_dirty_record_t, dr_dirty_node)); 1349 } 1350 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL) 1351 dr->dr_accounted = db->db.db_size; 1352 dr->dr_dbuf = db; 1353 dr->dr_txg = tx->tx_txg; 1354 dr->dr_next = *drp; 1355 *drp = dr; 1356 1357 /* 1358 * We could have been freed_in_flight between the dbuf_noread 1359 * and dbuf_dirty. We win, as though the dbuf_noread() had 1360 * happened after the free. 1361 */ 1362 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 1363 db->db_blkid != DMU_SPILL_BLKID) { 1364 mutex_enter(&dn->dn_mtx); 1365 if (dn->dn_free_ranges[txgoff] != NULL) { 1366 range_tree_clear(dn->dn_free_ranges[txgoff], 1367 db->db_blkid, 1); 1368 } 1369 mutex_exit(&dn->dn_mtx); 1370 db->db_freed_in_flight = FALSE; 1371 } 1372 1373 /* 1374 * This buffer is now part of this txg 1375 */ 1376 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg); 1377 db->db_dirtycnt += 1; 1378 ASSERT3U(db->db_dirtycnt, <=, 3); 1379 1380 mutex_exit(&db->db_mtx); 1381 1382 if (db->db_blkid == DMU_BONUS_BLKID || 1383 db->db_blkid == DMU_SPILL_BLKID) { 1384 mutex_enter(&dn->dn_mtx); 1385 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1386 list_insert_tail(&dn->dn_dirty_records[txgoff], dr); 1387 mutex_exit(&dn->dn_mtx); 1388 dnode_setdirty(dn, tx); 1389 DB_DNODE_EXIT(db); 1390 return (dr); 1391 } else if (do_free_accounting) { 1392 blkptr_t *bp = db->db_blkptr; 1393 int64_t willfree = (bp && !BP_IS_HOLE(bp)) ? 1394 bp_get_dsize(os->os_spa, bp) : db->db.db_size; 1395 /* 1396 * This is only a guess -- if the dbuf is dirty 1397 * in a previous txg, we don't know how much 1398 * space it will use on disk yet. We should 1399 * really have the struct_rwlock to access 1400 * db_blkptr, but since this is just a guess, 1401 * it's OK if we get an odd answer. 1402 */ 1403 ddt_prefetch(os->os_spa, bp); 1404 dnode_willuse_space(dn, -willfree, tx); 1405 } 1406 1407 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) { 1408 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1409 drop_struct_lock = TRUE; 1410 } 1411 1412 if (db->db_level == 0) { 1413 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock); 1414 ASSERT(dn->dn_maxblkid >= db->db_blkid); 1415 } 1416 1417 if (db->db_level+1 < dn->dn_nlevels) { 1418 dmu_buf_impl_t *parent = db->db_parent; 1419 dbuf_dirty_record_t *di; 1420 int parent_held = FALSE; 1421 1422 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) { 1423 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1424 1425 parent = dbuf_hold_level(dn, db->db_level+1, 1426 db->db_blkid >> epbs, FTAG); 1427 ASSERT(parent != NULL); 1428 parent_held = TRUE; 1429 } 1430 if (drop_struct_lock) 1431 rw_exit(&dn->dn_struct_rwlock); 1432 ASSERT3U(db->db_level+1, ==, parent->db_level); 1433 di = dbuf_dirty(parent, tx); 1434 if (parent_held) 1435 dbuf_rele(parent, FTAG); 1436 1437 mutex_enter(&db->db_mtx); 1438 /* 1439 * Since we've dropped the mutex, it's possible that 1440 * dbuf_undirty() might have changed this out from under us. 1441 */ 1442 if (db->db_last_dirty == dr || 1443 dn->dn_object == DMU_META_DNODE_OBJECT) { 1444 mutex_enter(&di->dt.di.dr_mtx); 1445 ASSERT3U(di->dr_txg, ==, tx->tx_txg); 1446 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1447 list_insert_tail(&di->dt.di.dr_children, dr); 1448 mutex_exit(&di->dt.di.dr_mtx); 1449 dr->dr_parent = di; 1450 } 1451 mutex_exit(&db->db_mtx); 1452 } else { 1453 ASSERT(db->db_level+1 == dn->dn_nlevels); 1454 ASSERT(db->db_blkid < dn->dn_nblkptr); 1455 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf); 1456 mutex_enter(&dn->dn_mtx); 1457 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1458 list_insert_tail(&dn->dn_dirty_records[txgoff], dr); 1459 mutex_exit(&dn->dn_mtx); 1460 if (drop_struct_lock) 1461 rw_exit(&dn->dn_struct_rwlock); 1462 } 1463 1464 dnode_setdirty(dn, tx); 1465 DB_DNODE_EXIT(db); 1466 return (dr); 1467} 1468 1469/* 1470 * Undirty a buffer in the transaction group referenced by the given 1471 * transaction. Return whether this evicted the dbuf. 1472 */ 1473static boolean_t 1474dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1475{ 1476 dnode_t *dn; 1477 uint64_t txg = tx->tx_txg; 1478 dbuf_dirty_record_t *dr, **drp; 1479 1480 ASSERT(txg != 0); 1481 1482 /* 1483 * Due to our use of dn_nlevels below, this can only be called 1484 * in open context, unless we are operating on the MOS. 1485 * From syncing context, dn_nlevels may be different from the 1486 * dn_nlevels used when dbuf was dirtied. 1487 */ 1488 ASSERT(db->db_objset == 1489 dmu_objset_pool(db->db_objset)->dp_meta_objset || 1490 txg != spa_syncing_txg(dmu_objset_spa(db->db_objset))); 1491 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1492 ASSERT0(db->db_level); 1493 ASSERT(MUTEX_HELD(&db->db_mtx)); 1494 1495 /* 1496 * If this buffer is not dirty, we're done. 1497 */ 1498 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next) 1499 if (dr->dr_txg <= txg) 1500 break; 1501 if (dr == NULL || dr->dr_txg < txg) 1502 return (B_FALSE); 1503 ASSERT(dr->dr_txg == txg); 1504 ASSERT(dr->dr_dbuf == db); 1505 1506 DB_DNODE_ENTER(db); 1507 dn = DB_DNODE(db); 1508 1509 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 1510 1511 ASSERT(db->db.db_size != 0); 1512 1513 dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset), 1514 dr->dr_accounted, txg); 1515 1516 *drp = dr->dr_next; 1517 1518 /* 1519 * Note that there are three places in dbuf_dirty() 1520 * where this dirty record may be put on a list. 1521 * Make sure to do a list_remove corresponding to 1522 * every one of those list_insert calls. 1523 */ 1524 if (dr->dr_parent) { 1525 mutex_enter(&dr->dr_parent->dt.di.dr_mtx); 1526 list_remove(&dr->dr_parent->dt.di.dr_children, dr); 1527 mutex_exit(&dr->dr_parent->dt.di.dr_mtx); 1528 } else if (db->db_blkid == DMU_SPILL_BLKID || 1529 db->db_level + 1 == dn->dn_nlevels) { 1530 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf); 1531 mutex_enter(&dn->dn_mtx); 1532 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr); 1533 mutex_exit(&dn->dn_mtx); 1534 } 1535 DB_DNODE_EXIT(db); 1536 1537 if (db->db_state != DB_NOFILL) { 1538 dbuf_unoverride(dr); 1539 1540 ASSERT(db->db_buf != NULL); 1541 ASSERT(dr->dt.dl.dr_data != NULL); 1542 if (dr->dt.dl.dr_data != db->db_buf) 1543 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db)); 1544 } 1545 1546 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 1547 1548 ASSERT(db->db_dirtycnt > 0); 1549 db->db_dirtycnt -= 1; 1550 1551 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) { 1552 arc_buf_t *buf = db->db_buf; 1553 1554 ASSERT(db->db_state == DB_NOFILL || arc_released(buf)); 1555 dbuf_clear_data(db); 1556 VERIFY(arc_buf_remove_ref(buf, db)); 1557 dbuf_evict(db); 1558 return (B_TRUE); 1559 } 1560 1561 return (B_FALSE); 1562} 1563 1564void 1565dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx) 1566{ 1567 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1568 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH; 1569 1570 ASSERT(tx->tx_txg != 0); 1571 ASSERT(!refcount_is_zero(&db->db_holds)); 1572 1573 /* 1574 * Quick check for dirtyness. For already dirty blocks, this 1575 * reduces runtime of this function by >90%, and overall performance 1576 * by 50% for some workloads (e.g. file deletion with indirect blocks 1577 * cached). 1578 */ 1579 mutex_enter(&db->db_mtx); 1580 dbuf_dirty_record_t *dr; 1581 for (dr = db->db_last_dirty; 1582 dr != NULL && dr->dr_txg >= tx->tx_txg; dr = dr->dr_next) { 1583 /* 1584 * It's possible that it is already dirty but not cached, 1585 * because there are some calls to dbuf_dirty() that don't 1586 * go through dmu_buf_will_dirty(). 1587 */ 1588 if (dr->dr_txg == tx->tx_txg && db->db_state == DB_CACHED) { 1589 /* This dbuf is already dirty and cached. */ 1590 dbuf_redirty(dr); 1591 mutex_exit(&db->db_mtx); 1592 return; 1593 } 1594 } 1595 mutex_exit(&db->db_mtx); 1596 1597 DB_DNODE_ENTER(db); 1598 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock)) 1599 rf |= DB_RF_HAVESTRUCT; 1600 DB_DNODE_EXIT(db); 1601 (void) dbuf_read(db, NULL, rf); 1602 (void) dbuf_dirty(db, tx); 1603} 1604 1605void 1606dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) 1607{ 1608 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1609 1610 db->db_state = DB_NOFILL; 1611 1612 dmu_buf_will_fill(db_fake, tx); 1613} 1614 1615void 1616dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) 1617{ 1618 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1619 1620 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1621 ASSERT(tx->tx_txg != 0); 1622 ASSERT(db->db_level == 0); 1623 ASSERT(!refcount_is_zero(&db->db_holds)); 1624 1625 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT || 1626 dmu_tx_private_ok(tx)); 1627 1628 dbuf_noread(db); 1629 (void) dbuf_dirty(db, tx); 1630} 1631 1632#pragma weak dmu_buf_fill_done = dbuf_fill_done 1633/* ARGSUSED */ 1634void 1635dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx) 1636{ 1637 mutex_enter(&db->db_mtx); 1638 DBUF_VERIFY(db); 1639 1640 if (db->db_state == DB_FILL) { 1641 if (db->db_level == 0 && db->db_freed_in_flight) { 1642 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1643 /* we were freed while filling */ 1644 /* XXX dbuf_undirty? */ 1645 bzero(db->db.db_data, db->db.db_size); 1646 db->db_freed_in_flight = FALSE; 1647 } 1648 db->db_state = DB_CACHED; 1649 cv_broadcast(&db->db_changed); 1650 } 1651 mutex_exit(&db->db_mtx); 1652} 1653 1654void 1655dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data, 1656 bp_embedded_type_t etype, enum zio_compress comp, 1657 int uncompressed_size, int compressed_size, int byteorder, 1658 dmu_tx_t *tx) 1659{ 1660 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; 1661 struct dirty_leaf *dl; 1662 dmu_object_type_t type; 1663 1664 if (etype == BP_EMBEDDED_TYPE_DATA) { 1665 ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset), 1666 SPA_FEATURE_EMBEDDED_DATA)); 1667 } 1668 1669 DB_DNODE_ENTER(db); 1670 type = DB_DNODE(db)->dn_type; 1671 DB_DNODE_EXIT(db); 1672 1673 ASSERT0(db->db_level); 1674 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1675 1676 dmu_buf_will_not_fill(dbuf, tx); 1677 1678 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); 1679 dl = &db->db_last_dirty->dt.dl; 1680 encode_embedded_bp_compressed(&dl->dr_overridden_by, 1681 data, comp, uncompressed_size, compressed_size); 1682 BPE_SET_ETYPE(&dl->dr_overridden_by, etype); 1683 BP_SET_TYPE(&dl->dr_overridden_by, type); 1684 BP_SET_LEVEL(&dl->dr_overridden_by, 0); 1685 BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder); 1686 1687 dl->dr_override_state = DR_OVERRIDDEN; 1688 dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg; 1689} 1690 1691/* 1692 * Directly assign a provided arc buf to a given dbuf if it's not referenced 1693 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf. 1694 */ 1695void 1696dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx) 1697{ 1698 ASSERT(!refcount_is_zero(&db->db_holds)); 1699 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1700 ASSERT(db->db_level == 0); 1701 ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA); 1702 ASSERT(buf != NULL); 1703 ASSERT(arc_buf_size(buf) == db->db.db_size); 1704 ASSERT(tx->tx_txg != 0); 1705 1706 arc_return_buf(buf, db); 1707 ASSERT(arc_released(buf)); 1708 1709 mutex_enter(&db->db_mtx); 1710 1711 while (db->db_state == DB_READ || db->db_state == DB_FILL) 1712 cv_wait(&db->db_changed, &db->db_mtx); 1713 1714 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED); 1715 1716 if (db->db_state == DB_CACHED && 1717 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) { 1718 mutex_exit(&db->db_mtx); 1719 (void) dbuf_dirty(db, tx); 1720 bcopy(buf->b_data, db->db.db_data, db->db.db_size); 1721 VERIFY(arc_buf_remove_ref(buf, db)); 1722 xuio_stat_wbuf_copied(); 1723 return; 1724 } 1725 1726 xuio_stat_wbuf_nocopy(); 1727 if (db->db_state == DB_CACHED) { 1728 dbuf_dirty_record_t *dr = db->db_last_dirty; 1729 1730 ASSERT(db->db_buf != NULL); 1731 if (dr != NULL && dr->dr_txg == tx->tx_txg) { 1732 ASSERT(dr->dt.dl.dr_data == db->db_buf); 1733 if (!arc_released(db->db_buf)) { 1734 ASSERT(dr->dt.dl.dr_override_state == 1735 DR_OVERRIDDEN); 1736 arc_release(db->db_buf, db); 1737 } 1738 dr->dt.dl.dr_data = buf; 1739 VERIFY(arc_buf_remove_ref(db->db_buf, db)); 1740 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) { 1741 arc_release(db->db_buf, db); 1742 VERIFY(arc_buf_remove_ref(db->db_buf, db)); 1743 } 1744 db->db_buf = NULL; 1745 } 1746 ASSERT(db->db_buf == NULL); 1747 dbuf_set_data(db, buf); 1748 db->db_state = DB_FILL; 1749 mutex_exit(&db->db_mtx); 1750 (void) dbuf_dirty(db, tx); 1751 dmu_buf_fill_done(&db->db, tx); 1752} 1753 1754/* 1755 * "Clear" the contents of this dbuf. This will mark the dbuf 1756 * EVICTING and clear *most* of its references. Unfortunately, 1757 * when we are not holding the dn_dbufs_mtx, we can't clear the 1758 * entry in the dn_dbufs list. We have to wait until dbuf_destroy() 1759 * in this case. For callers from the DMU we will usually see: 1760 * dbuf_clear()->arc_clear_callback()->dbuf_do_evict()->dbuf_destroy() 1761 * For the arc callback, we will usually see: 1762 * dbuf_do_evict()->dbuf_clear();dbuf_destroy() 1763 * Sometimes, though, we will get a mix of these two: 1764 * DMU: dbuf_clear()->arc_clear_callback() 1765 * ARC: dbuf_do_evict()->dbuf_destroy() 1766 * 1767 * This routine will dissociate the dbuf from the arc, by calling 1768 * arc_clear_callback(), but will not evict the data from the ARC. 1769 */ 1770void 1771dbuf_clear(dmu_buf_impl_t *db) 1772{ 1773 dnode_t *dn; 1774 dmu_buf_impl_t *parent = db->db_parent; 1775 dmu_buf_impl_t *dndb; 1776 boolean_t dbuf_gone = B_FALSE; 1777 1778 ASSERT(MUTEX_HELD(&db->db_mtx)); 1779 ASSERT(refcount_is_zero(&db->db_holds)); 1780 1781 dbuf_evict_user(db); 1782 1783 if (db->db_state == DB_CACHED) { 1784 ASSERT(db->db.db_data != NULL); 1785 if (db->db_blkid == DMU_BONUS_BLKID) { 1786 zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN); 1787 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 1788 } 1789 db->db.db_data = NULL; 1790 db->db_state = DB_UNCACHED; 1791 } 1792 1793 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL); 1794 ASSERT(db->db_data_pending == NULL); 1795 1796 db->db_state = DB_EVICTING; 1797 db->db_blkptr = NULL; 1798 1799 DB_DNODE_ENTER(db); 1800 dn = DB_DNODE(db); 1801 dndb = dn->dn_dbuf; 1802 if (db->db_blkid != DMU_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) { 1803 avl_remove(&dn->dn_dbufs, db); 1804 atomic_dec_32(&dn->dn_dbufs_count); 1805 membar_producer(); 1806 DB_DNODE_EXIT(db); 1807 /* 1808 * Decrementing the dbuf count means that the hold corresponding 1809 * to the removed dbuf is no longer discounted in dnode_move(), 1810 * so the dnode cannot be moved until after we release the hold. 1811 * The membar_producer() ensures visibility of the decremented 1812 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually 1813 * release any lock. 1814 */ 1815 dnode_rele(dn, db); 1816 db->db_dnode_handle = NULL; 1817 } else { 1818 DB_DNODE_EXIT(db); 1819 } 1820 1821 if (db->db_buf) 1822 dbuf_gone = arc_clear_callback(db->db_buf); 1823 1824 if (!dbuf_gone) 1825 mutex_exit(&db->db_mtx); 1826 1827 /* 1828 * If this dbuf is referenced from an indirect dbuf, 1829 * decrement the ref count on the indirect dbuf. 1830 */ 1831 if (parent && parent != dndb) 1832 dbuf_rele(parent, db); 1833} 1834 1835/* 1836 * Note: While bpp will always be updated if the function returns success, 1837 * parentp will not be updated if the dnode does not have dn_dbuf filled in; 1838 * this happens when the dnode is the meta-dnode, or a userused or groupused 1839 * object. 1840 */ 1841static int 1842dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse, 1843 dmu_buf_impl_t **parentp, blkptr_t **bpp) 1844{ 1845 int nlevels, epbs; 1846 1847 *parentp = NULL; 1848 *bpp = NULL; 1849 1850 ASSERT(blkid != DMU_BONUS_BLKID); 1851 1852 if (blkid == DMU_SPILL_BLKID) { 1853 mutex_enter(&dn->dn_mtx); 1854 if (dn->dn_have_spill && 1855 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) 1856 *bpp = &dn->dn_phys->dn_spill; 1857 else 1858 *bpp = NULL; 1859 dbuf_add_ref(dn->dn_dbuf, NULL); 1860 *parentp = dn->dn_dbuf; 1861 mutex_exit(&dn->dn_mtx); 1862 return (0); 1863 } 1864 1865 if (dn->dn_phys->dn_nlevels == 0) 1866 nlevels = 1; 1867 else 1868 nlevels = dn->dn_phys->dn_nlevels; 1869 1870 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1871 1872 ASSERT3U(level * epbs, <, 64); 1873 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1874 if (level >= nlevels || 1875 (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) { 1876 /* the buffer has no parent yet */ 1877 return (SET_ERROR(ENOENT)); 1878 } else if (level < nlevels-1) { 1879 /* this block is referenced from an indirect block */ 1880 int err = dbuf_hold_impl(dn, level+1, 1881 blkid >> epbs, fail_sparse, FALSE, NULL, parentp); 1882 if (err) 1883 return (err); 1884 err = dbuf_read(*parentp, NULL, 1885 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL)); 1886 if (err) { 1887 dbuf_rele(*parentp, NULL); 1888 *parentp = NULL; 1889 return (err); 1890 } 1891 *bpp = ((blkptr_t *)(*parentp)->db.db_data) + 1892 (blkid & ((1ULL << epbs) - 1)); 1893 return (0); 1894 } else { 1895 /* the block is referenced from the dnode */ 1896 ASSERT3U(level, ==, nlevels-1); 1897 ASSERT(dn->dn_phys->dn_nblkptr == 0 || 1898 blkid < dn->dn_phys->dn_nblkptr); 1899 if (dn->dn_dbuf) { 1900 dbuf_add_ref(dn->dn_dbuf, NULL); 1901 *parentp = dn->dn_dbuf; 1902 } 1903 *bpp = &dn->dn_phys->dn_blkptr[blkid]; 1904 return (0); 1905 } 1906} 1907 1908static dmu_buf_impl_t * 1909dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid, 1910 dmu_buf_impl_t *parent, blkptr_t *blkptr) 1911{ 1912 objset_t *os = dn->dn_objset; 1913 dmu_buf_impl_t *db, *odb; 1914 1915 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1916 ASSERT(dn->dn_type != DMU_OT_NONE); 1917 1918 db = kmem_cache_alloc(dbuf_cache, KM_SLEEP); 1919 1920 db->db_objset = os; 1921 db->db.db_object = dn->dn_object; 1922 db->db_level = level; 1923 db->db_blkid = blkid; 1924 db->db_last_dirty = NULL; 1925 db->db_dirtycnt = 0; 1926 db->db_dnode_handle = dn->dn_handle; 1927 db->db_parent = parent; 1928 db->db_blkptr = blkptr; 1929 1930 db->db_user = NULL; 1931 db->db_user_immediate_evict = FALSE; 1932 db->db_freed_in_flight = FALSE; 1933 db->db_pending_evict = FALSE; 1934 1935 if (blkid == DMU_BONUS_BLKID) { 1936 ASSERT3P(parent, ==, dn->dn_dbuf); 1937 db->db.db_size = DN_MAX_BONUSLEN - 1938 (dn->dn_nblkptr-1) * sizeof (blkptr_t); 1939 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 1940 db->db.db_offset = DMU_BONUS_BLKID; 1941 db->db_state = DB_UNCACHED; 1942 /* the bonus dbuf is not placed in the hash table */ 1943 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 1944 return (db); 1945 } else if (blkid == DMU_SPILL_BLKID) { 1946 db->db.db_size = (blkptr != NULL) ? 1947 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE; 1948 db->db.db_offset = 0; 1949 } else { 1950 int blocksize = 1951 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz; 1952 db->db.db_size = blocksize; 1953 db->db.db_offset = db->db_blkid * blocksize; 1954 } 1955 1956 /* 1957 * Hold the dn_dbufs_mtx while we get the new dbuf 1958 * in the hash table *and* added to the dbufs list. 1959 * This prevents a possible deadlock with someone 1960 * trying to look up this dbuf before its added to the 1961 * dn_dbufs list. 1962 */ 1963 mutex_enter(&dn->dn_dbufs_mtx); 1964 db->db_state = DB_EVICTING; 1965 if ((odb = dbuf_hash_insert(db)) != NULL) { 1966 /* someone else inserted it first */ 1967 kmem_cache_free(dbuf_cache, db); 1968 mutex_exit(&dn->dn_dbufs_mtx); 1969 return (odb); 1970 } 1971 avl_add(&dn->dn_dbufs, db); 1972 if (db->db_level == 0 && db->db_blkid >= 1973 dn->dn_unlisted_l0_blkid) 1974 dn->dn_unlisted_l0_blkid = db->db_blkid + 1; 1975 db->db_state = DB_UNCACHED; 1976 mutex_exit(&dn->dn_dbufs_mtx); 1977 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 1978 1979 if (parent && parent != dn->dn_dbuf) 1980 dbuf_add_ref(parent, db); 1981 1982 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 1983 refcount_count(&dn->dn_holds) > 0); 1984 (void) refcount_add(&dn->dn_holds, db); 1985 atomic_inc_32(&dn->dn_dbufs_count); 1986 1987 dprintf_dbuf(db, "db=%p\n", db); 1988 1989 return (db); 1990} 1991 1992static int 1993dbuf_do_evict(void *private) 1994{ 1995 dmu_buf_impl_t *db = private; 1996 1997 if (!MUTEX_HELD(&db->db_mtx)) 1998 mutex_enter(&db->db_mtx); 1999 2000 ASSERT(refcount_is_zero(&db->db_holds)); 2001 2002 if (db->db_state != DB_EVICTING) { 2003 ASSERT(db->db_state == DB_CACHED); 2004 DBUF_VERIFY(db); 2005 db->db_buf = NULL; 2006 dbuf_evict(db); 2007 } else { 2008 mutex_exit(&db->db_mtx); 2009 dbuf_destroy(db); 2010 } 2011 return (0); 2012} 2013 2014static void 2015dbuf_destroy(dmu_buf_impl_t *db) 2016{ 2017 ASSERT(refcount_is_zero(&db->db_holds)); 2018 2019 if (db->db_blkid != DMU_BONUS_BLKID) { 2020 /* 2021 * If this dbuf is still on the dn_dbufs list, 2022 * remove it from that list. 2023 */ 2024 if (db->db_dnode_handle != NULL) { 2025 dnode_t *dn; 2026 2027 DB_DNODE_ENTER(db); 2028 dn = DB_DNODE(db); 2029 mutex_enter(&dn->dn_dbufs_mtx); 2030 avl_remove(&dn->dn_dbufs, db); 2031 atomic_dec_32(&dn->dn_dbufs_count); 2032 mutex_exit(&dn->dn_dbufs_mtx); 2033 DB_DNODE_EXIT(db); 2034 /* 2035 * Decrementing the dbuf count means that the hold 2036 * corresponding to the removed dbuf is no longer 2037 * discounted in dnode_move(), so the dnode cannot be 2038 * moved until after we release the hold. 2039 */ 2040 dnode_rele(dn, db); 2041 db->db_dnode_handle = NULL; 2042 } 2043 dbuf_hash_remove(db); 2044 } 2045 db->db_parent = NULL; 2046 db->db_buf = NULL; 2047 2048 ASSERT(db->db.db_data == NULL); 2049 ASSERT(db->db_hash_next == NULL); 2050 ASSERT(db->db_blkptr == NULL); 2051 ASSERT(db->db_data_pending == NULL); 2052 2053 kmem_cache_free(dbuf_cache, db); 2054 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 2055} 2056 2057typedef struct dbuf_prefetch_arg { 2058 spa_t *dpa_spa; /* The spa to issue the prefetch in. */ 2059 zbookmark_phys_t dpa_zb; /* The target block to prefetch. */ 2060 int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */ 2061 int dpa_curlevel; /* The current level that we're reading */ 2062 zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */ 2063 zio_t *dpa_zio; /* The parent zio_t for all prefetches. */ 2064 arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */ 2065} dbuf_prefetch_arg_t; 2066 2067/* 2068 * Actually issue the prefetch read for the block given. 2069 */ 2070static void 2071dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp) 2072{ 2073 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) 2074 return; 2075 2076 arc_flags_t aflags = 2077 dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH; 2078 2079 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp)); 2080 ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level); 2081 ASSERT(dpa->dpa_zio != NULL); 2082 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL, 2083 dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2084 &aflags, &dpa->dpa_zb); 2085} 2086 2087/* 2088 * Called when an indirect block above our prefetch target is read in. This 2089 * will either read in the next indirect block down the tree or issue the actual 2090 * prefetch if the next block down is our target. 2091 */ 2092static void 2093dbuf_prefetch_indirect_done(zio_t *zio, arc_buf_t *abuf, void *private) 2094{ 2095 dbuf_prefetch_arg_t *dpa = private; 2096 2097 ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel); 2098 ASSERT3S(dpa->dpa_curlevel, >, 0); 2099 if (zio != NULL) { 2100 ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel); 2101 ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size); 2102 ASSERT3P(zio->io_spa, ==, dpa->dpa_spa); 2103 } 2104 2105 dpa->dpa_curlevel--; 2106 2107 uint64_t nextblkid = dpa->dpa_zb.zb_blkid >> 2108 (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level)); 2109 blkptr_t *bp = ((blkptr_t *)abuf->b_data) + 2110 P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs); 2111 if (BP_IS_HOLE(bp) || (zio != NULL && zio->io_error != 0)) { 2112 kmem_free(dpa, sizeof (*dpa)); 2113 } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) { 2114 ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid); 2115 dbuf_issue_final_prefetch(dpa, bp); 2116 kmem_free(dpa, sizeof (*dpa)); 2117 } else { 2118 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT; 2119 zbookmark_phys_t zb; 2120 2121 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp)); 2122 2123 SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset, 2124 dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid); 2125 2126 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, 2127 bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio, 2128 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2129 &iter_aflags, &zb); 2130 } 2131 (void) arc_buf_remove_ref(abuf, private); 2132} 2133 2134/* 2135 * Issue prefetch reads for the given block on the given level. If the indirect 2136 * blocks above that block are not in memory, we will read them in 2137 * asynchronously. As a result, this call never blocks waiting for a read to 2138 * complete. 2139 */ 2140void 2141dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio, 2142 arc_flags_t aflags) 2143{ 2144 blkptr_t bp; 2145 int epbs, nlevels, curlevel; 2146 uint64_t curblkid; 2147 2148 ASSERT(blkid != DMU_BONUS_BLKID); 2149 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 2150 2151 if (blkid > dn->dn_maxblkid) 2152 return; 2153 2154 if (dnode_block_freed(dn, blkid)) 2155 return; 2156 2157 /* 2158 * This dnode hasn't been written to disk yet, so there's nothing to 2159 * prefetch. 2160 */ 2161 nlevels = dn->dn_phys->dn_nlevels; 2162 if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0) 2163 return; 2164 2165 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 2166 if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level)) 2167 return; 2168 2169 dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object, 2170 level, blkid); 2171 if (db != NULL) { 2172 mutex_exit(&db->db_mtx); 2173 /* 2174 * This dbuf already exists. It is either CACHED, or 2175 * (we assume) about to be read or filled. 2176 */ 2177 return; 2178 } 2179 2180 /* 2181 * Find the closest ancestor (indirect block) of the target block 2182 * that is present in the cache. In this indirect block, we will 2183 * find the bp that is at curlevel, curblkid. 2184 */ 2185 curlevel = level; 2186 curblkid = blkid; 2187 while (curlevel < nlevels - 1) { 2188 int parent_level = curlevel + 1; 2189 uint64_t parent_blkid = curblkid >> epbs; 2190 dmu_buf_impl_t *db; 2191 2192 if (dbuf_hold_impl(dn, parent_level, parent_blkid, 2193 FALSE, TRUE, FTAG, &db) == 0) { 2194 blkptr_t *bpp = db->db_buf->b_data; 2195 bp = bpp[P2PHASE(curblkid, 1 << epbs)]; 2196 dbuf_rele(db, FTAG); 2197 break; 2198 } 2199 2200 curlevel = parent_level; 2201 curblkid = parent_blkid; 2202 } 2203 2204 if (curlevel == nlevels - 1) { 2205 /* No cached indirect blocks found. */ 2206 ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr); 2207 bp = dn->dn_phys->dn_blkptr[curblkid]; 2208 } 2209 if (BP_IS_HOLE(&bp)) 2210 return; 2211 2212 ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp)); 2213 2214 zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL, 2215 ZIO_FLAG_CANFAIL); 2216 2217 dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP); 2218 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset; 2219 SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET, 2220 dn->dn_object, level, blkid); 2221 dpa->dpa_curlevel = curlevel; 2222 dpa->dpa_prio = prio; 2223 dpa->dpa_aflags = aflags; 2224 dpa->dpa_spa = dn->dn_objset->os_spa; 2225 dpa->dpa_epbs = epbs; 2226 dpa->dpa_zio = pio; 2227 2228 /* 2229 * If we have the indirect just above us, no need to do the asynchronous 2230 * prefetch chain; we'll just run the last step ourselves. If we're at 2231 * a higher level, though, we want to issue the prefetches for all the 2232 * indirect blocks asynchronously, so we can go on with whatever we were 2233 * doing. 2234 */ 2235 if (curlevel == level) { 2236 ASSERT3U(curblkid, ==, blkid); 2237 dbuf_issue_final_prefetch(dpa, &bp); 2238 kmem_free(dpa, sizeof (*dpa)); 2239 } else { 2240 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT; 2241 zbookmark_phys_t zb; 2242 2243 SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET, 2244 dn->dn_object, curlevel, curblkid); 2245 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, 2246 &bp, dbuf_prefetch_indirect_done, dpa, prio, 2247 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2248 &iter_aflags, &zb); 2249 } 2250 /* 2251 * We use pio here instead of dpa_zio since it's possible that 2252 * dpa may have already been freed. 2253 */ 2254 zio_nowait(pio); 2255} 2256 2257/* 2258 * Returns with db_holds incremented, and db_mtx not held. 2259 * Note: dn_struct_rwlock must be held. 2260 */ 2261int 2262dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, 2263 boolean_t fail_sparse, boolean_t fail_uncached, 2264 void *tag, dmu_buf_impl_t **dbp) 2265{ 2266 dmu_buf_impl_t *db, *parent = NULL; 2267 2268 ASSERT(blkid != DMU_BONUS_BLKID); 2269 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 2270 ASSERT3U(dn->dn_nlevels, >, level); 2271 2272 *dbp = NULL; 2273top: 2274 /* dbuf_find() returns with db_mtx held */ 2275 db = dbuf_find(dn->dn_objset, dn->dn_object, level, blkid); 2276 2277 if (db == NULL) { 2278 blkptr_t *bp = NULL; 2279 int err; 2280 2281 if (fail_uncached) 2282 return (SET_ERROR(ENOENT)); 2283 2284 ASSERT3P(parent, ==, NULL); 2285 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp); 2286 if (fail_sparse) { 2287 if (err == 0 && bp && BP_IS_HOLE(bp)) 2288 err = SET_ERROR(ENOENT); 2289 if (err) { 2290 if (parent) 2291 dbuf_rele(parent, NULL); 2292 return (err); 2293 } 2294 } 2295 if (err && err != ENOENT) 2296 return (err); 2297 db = dbuf_create(dn, level, blkid, parent, bp); 2298 } 2299 2300 if (fail_uncached && db->db_state != DB_CACHED) { 2301 mutex_exit(&db->db_mtx); 2302 return (SET_ERROR(ENOENT)); 2303 } 2304 2305 if (db->db_buf && refcount_is_zero(&db->db_holds)) { 2306 arc_buf_add_ref(db->db_buf, db); 2307 if (db->db_buf->b_data == NULL) { 2308 dbuf_clear(db); 2309 if (parent) { 2310 dbuf_rele(parent, NULL); 2311 parent = NULL; 2312 } 2313 goto top; 2314 } 2315 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data); 2316 } 2317 2318 ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf)); 2319 2320 /* 2321 * If this buffer is currently syncing out, and we are are 2322 * still referencing it from db_data, we need to make a copy 2323 * of it in case we decide we want to dirty it again in this txg. 2324 */ 2325 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 2326 dn->dn_object != DMU_META_DNODE_OBJECT && 2327 db->db_state == DB_CACHED && db->db_data_pending) { 2328 dbuf_dirty_record_t *dr = db->db_data_pending; 2329 2330 if (dr->dt.dl.dr_data == db->db_buf) { 2331 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 2332 2333 dbuf_set_data(db, 2334 arc_buf_alloc(dn->dn_objset->os_spa, 2335 db->db.db_size, db, type)); 2336 bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data, 2337 db->db.db_size); 2338 } 2339 } 2340 2341 (void) refcount_add(&db->db_holds, tag); 2342 DBUF_VERIFY(db); 2343 mutex_exit(&db->db_mtx); 2344 2345 /* NOTE: we can't rele the parent until after we drop the db_mtx */ 2346 if (parent) 2347 dbuf_rele(parent, NULL); 2348 2349 ASSERT3P(DB_DNODE(db), ==, dn); 2350 ASSERT3U(db->db_blkid, ==, blkid); 2351 ASSERT3U(db->db_level, ==, level); 2352 *dbp = db; 2353 2354 return (0); 2355} 2356 2357dmu_buf_impl_t * 2358dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag) 2359{ 2360 return (dbuf_hold_level(dn, 0, blkid, tag)); 2361} 2362 2363dmu_buf_impl_t * 2364dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag) 2365{ 2366 dmu_buf_impl_t *db; 2367 int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db); 2368 return (err ? NULL : db); 2369} 2370 2371void 2372dbuf_create_bonus(dnode_t *dn) 2373{ 2374 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 2375 2376 ASSERT(dn->dn_bonus == NULL); 2377 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL); 2378} 2379 2380int 2381dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx) 2382{ 2383 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2384 dnode_t *dn; 2385 2386 if (db->db_blkid != DMU_SPILL_BLKID) 2387 return (SET_ERROR(ENOTSUP)); 2388 if (blksz == 0) 2389 blksz = SPA_MINBLOCKSIZE; 2390 ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset))); 2391 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE); 2392 2393 DB_DNODE_ENTER(db); 2394 dn = DB_DNODE(db); 2395 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 2396 dbuf_new_size(db, blksz, tx); 2397 rw_exit(&dn->dn_struct_rwlock); 2398 DB_DNODE_EXIT(db); 2399 2400 return (0); 2401} 2402 2403void 2404dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx) 2405{ 2406 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx); 2407} 2408 2409#pragma weak dmu_buf_add_ref = dbuf_add_ref 2410void 2411dbuf_add_ref(dmu_buf_impl_t *db, void *tag) 2412{ 2413 int64_t holds = refcount_add(&db->db_holds, tag); 2414 ASSERT(holds > 1); 2415} 2416 2417#pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref 2418boolean_t 2419dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid, 2420 void *tag) 2421{ 2422 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2423 dmu_buf_impl_t *found_db; 2424 boolean_t result = B_FALSE; 2425 2426 if (db->db_blkid == DMU_BONUS_BLKID) 2427 found_db = dbuf_find_bonus(os, obj); 2428 else 2429 found_db = dbuf_find(os, obj, 0, blkid); 2430 2431 if (found_db != NULL) { 2432 if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) { 2433 (void) refcount_add(&db->db_holds, tag); 2434 result = B_TRUE; 2435 } 2436 mutex_exit(&db->db_mtx); 2437 } 2438 return (result); 2439} 2440 2441/* 2442 * If you call dbuf_rele() you had better not be referencing the dnode handle 2443 * unless you have some other direct or indirect hold on the dnode. (An indirect 2444 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.) 2445 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the 2446 * dnode's parent dbuf evicting its dnode handles. 2447 */ 2448void 2449dbuf_rele(dmu_buf_impl_t *db, void *tag) 2450{ 2451 mutex_enter(&db->db_mtx); 2452 dbuf_rele_and_unlock(db, tag); 2453} 2454 2455void 2456dmu_buf_rele(dmu_buf_t *db, void *tag) 2457{ 2458 dbuf_rele((dmu_buf_impl_t *)db, tag); 2459} 2460 2461/* 2462 * dbuf_rele() for an already-locked dbuf. This is necessary to allow 2463 * db_dirtycnt and db_holds to be updated atomically. 2464 */ 2465void 2466dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag) 2467{ 2468 int64_t holds; 2469 2470 ASSERT(MUTEX_HELD(&db->db_mtx)); 2471 DBUF_VERIFY(db); 2472 2473 /* 2474 * Remove the reference to the dbuf before removing its hold on the 2475 * dnode so we can guarantee in dnode_move() that a referenced bonus 2476 * buffer has a corresponding dnode hold. 2477 */ 2478 holds = refcount_remove(&db->db_holds, tag); 2479 ASSERT(holds >= 0); 2480 2481 /* 2482 * We can't freeze indirects if there is a possibility that they 2483 * may be modified in the current syncing context. 2484 */ 2485 if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) 2486 arc_buf_freeze(db->db_buf); 2487 2488 if (holds == db->db_dirtycnt && 2489 db->db_level == 0 && db->db_user_immediate_evict) 2490 dbuf_evict_user(db); 2491 2492 if (holds == 0) { 2493 if (db->db_blkid == DMU_BONUS_BLKID) { 2494 dnode_t *dn; 2495 boolean_t evict_dbuf = db->db_pending_evict; 2496 2497 /* 2498 * If the dnode moves here, we cannot cross this 2499 * barrier until the move completes. 2500 */ 2501 DB_DNODE_ENTER(db); 2502 2503 dn = DB_DNODE(db); 2504 atomic_dec_32(&dn->dn_dbufs_count); 2505 2506 /* 2507 * Decrementing the dbuf count means that the bonus 2508 * buffer's dnode hold is no longer discounted in 2509 * dnode_move(). The dnode cannot move until after 2510 * the dnode_rele() below. 2511 */ 2512 DB_DNODE_EXIT(db); 2513 2514 /* 2515 * Do not reference db after its lock is dropped. 2516 * Another thread may evict it. 2517 */ 2518 mutex_exit(&db->db_mtx); 2519 2520 if (evict_dbuf) 2521 dnode_evict_bonus(dn); 2522 2523 dnode_rele(dn, db); 2524 } else if (db->db_buf == NULL) { 2525 /* 2526 * This is a special case: we never associated this 2527 * dbuf with any data allocated from the ARC. 2528 */ 2529 ASSERT(db->db_state == DB_UNCACHED || 2530 db->db_state == DB_NOFILL); 2531 dbuf_evict(db); 2532 } else if (arc_released(db->db_buf)) { 2533 arc_buf_t *buf = db->db_buf; 2534 /* 2535 * This dbuf has anonymous data associated with it. 2536 */ 2537 dbuf_clear_data(db); 2538 VERIFY(arc_buf_remove_ref(buf, db)); 2539 dbuf_evict(db); 2540 } else { 2541 VERIFY(!arc_buf_remove_ref(db->db_buf, db)); 2542 2543 /* 2544 * A dbuf will be eligible for eviction if either the 2545 * 'primarycache' property is set or a duplicate 2546 * copy of this buffer is already cached in the arc. 2547 * 2548 * In the case of the 'primarycache' a buffer 2549 * is considered for eviction if it matches the 2550 * criteria set in the property. 2551 * 2552 * To decide if our buffer is considered a 2553 * duplicate, we must call into the arc to determine 2554 * if multiple buffers are referencing the same 2555 * block on-disk. If so, then we simply evict 2556 * ourselves. 2557 */ 2558 if (!DBUF_IS_CACHEABLE(db)) { 2559 if (db->db_blkptr != NULL && 2560 !BP_IS_HOLE(db->db_blkptr) && 2561 !BP_IS_EMBEDDED(db->db_blkptr)) { 2562 spa_t *spa = 2563 dmu_objset_spa(db->db_objset); 2564 blkptr_t bp = *db->db_blkptr; 2565 dbuf_clear(db); 2566 arc_freed(spa, &bp); 2567 } else { 2568 dbuf_clear(db); 2569 } 2570 } else if (db->db_pending_evict || 2571 arc_buf_eviction_needed(db->db_buf)) { 2572 dbuf_clear(db); 2573 } else { 2574 mutex_exit(&db->db_mtx); 2575 } 2576 } 2577 } else { 2578 mutex_exit(&db->db_mtx); 2579 } 2580} 2581 2582#pragma weak dmu_buf_refcount = dbuf_refcount 2583uint64_t 2584dbuf_refcount(dmu_buf_impl_t *db) 2585{ 2586 return (refcount_count(&db->db_holds)); 2587} 2588 2589void * 2590dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user, 2591 dmu_buf_user_t *new_user) 2592{ 2593 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2594 2595 mutex_enter(&db->db_mtx); 2596 dbuf_verify_user(db, DBVU_NOT_EVICTING); 2597 if (db->db_user == old_user) 2598 db->db_user = new_user; 2599 else 2600 old_user = db->db_user; 2601 dbuf_verify_user(db, DBVU_NOT_EVICTING); 2602 mutex_exit(&db->db_mtx); 2603 2604 return (old_user); 2605} 2606 2607void * 2608dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user) 2609{ 2610 return (dmu_buf_replace_user(db_fake, NULL, user)); 2611} 2612 2613void * 2614dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user) 2615{ 2616 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2617 2618 db->db_user_immediate_evict = TRUE; 2619 return (dmu_buf_set_user(db_fake, user)); 2620} 2621 2622void * 2623dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user) 2624{ 2625 return (dmu_buf_replace_user(db_fake, user, NULL)); 2626} 2627 2628void * 2629dmu_buf_get_user(dmu_buf_t *db_fake) 2630{ 2631 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2632 2633 dbuf_verify_user(db, DBVU_NOT_EVICTING); 2634 return (db->db_user); 2635} 2636 2637void 2638dmu_buf_user_evict_wait() 2639{ 2640 taskq_wait(dbu_evict_taskq); 2641} 2642 2643boolean_t 2644dmu_buf_freeable(dmu_buf_t *dbuf) 2645{ 2646 boolean_t res = B_FALSE; 2647 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; 2648 2649 if (db->db_blkptr) 2650 res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset, 2651 db->db_blkptr, db->db_blkptr->blk_birth); 2652 2653 return (res); 2654} 2655 2656blkptr_t * 2657dmu_buf_get_blkptr(dmu_buf_t *db) 2658{ 2659 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 2660 return (dbi->db_blkptr); 2661} 2662 2663static void 2664dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db) 2665{ 2666 /* ASSERT(dmu_tx_is_syncing(tx) */ 2667 ASSERT(MUTEX_HELD(&db->db_mtx)); 2668 2669 if (db->db_blkptr != NULL) 2670 return; 2671 2672 if (db->db_blkid == DMU_SPILL_BLKID) { 2673 db->db_blkptr = &dn->dn_phys->dn_spill; 2674 BP_ZERO(db->db_blkptr); 2675 return; 2676 } 2677 if (db->db_level == dn->dn_phys->dn_nlevels-1) { 2678 /* 2679 * This buffer was allocated at a time when there was 2680 * no available blkptrs from the dnode, or it was 2681 * inappropriate to hook it in (i.e., nlevels mis-match). 2682 */ 2683 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr); 2684 ASSERT(db->db_parent == NULL); 2685 db->db_parent = dn->dn_dbuf; 2686 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid]; 2687 DBUF_VERIFY(db); 2688 } else { 2689 dmu_buf_impl_t *parent = db->db_parent; 2690 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 2691 2692 ASSERT(dn->dn_phys->dn_nlevels > 1); 2693 if (parent == NULL) { 2694 mutex_exit(&db->db_mtx); 2695 rw_enter(&dn->dn_struct_rwlock, RW_READER); 2696 parent = dbuf_hold_level(dn, db->db_level + 1, 2697 db->db_blkid >> epbs, db); 2698 rw_exit(&dn->dn_struct_rwlock); 2699 mutex_enter(&db->db_mtx); 2700 db->db_parent = parent; 2701 } 2702 db->db_blkptr = (blkptr_t *)parent->db.db_data + 2703 (db->db_blkid & ((1ULL << epbs) - 1)); 2704 DBUF_VERIFY(db); 2705 } 2706} 2707 2708static void 2709dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx) 2710{ 2711 dmu_buf_impl_t *db = dr->dr_dbuf; 2712 dnode_t *dn; 2713 zio_t *zio; 2714 2715 ASSERT(dmu_tx_is_syncing(tx)); 2716 2717 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 2718 2719 mutex_enter(&db->db_mtx); 2720 2721 ASSERT(db->db_level > 0); 2722 DBUF_VERIFY(db); 2723 2724 /* Read the block if it hasn't been read yet. */ 2725 if (db->db_buf == NULL) { 2726 mutex_exit(&db->db_mtx); 2727 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED); 2728 mutex_enter(&db->db_mtx); 2729 } 2730 ASSERT3U(db->db_state, ==, DB_CACHED); 2731 ASSERT(db->db_buf != NULL); 2732 2733 DB_DNODE_ENTER(db); 2734 dn = DB_DNODE(db); 2735 /* Indirect block size must match what the dnode thinks it is. */ 2736 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 2737 dbuf_check_blkptr(dn, db); 2738 DB_DNODE_EXIT(db); 2739 2740 /* Provide the pending dirty record to child dbufs */ 2741 db->db_data_pending = dr; 2742 2743 mutex_exit(&db->db_mtx); 2744 dbuf_write(dr, db->db_buf, tx); 2745 2746 zio = dr->dr_zio; 2747 mutex_enter(&dr->dt.di.dr_mtx); 2748 dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx); 2749 ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 2750 mutex_exit(&dr->dt.di.dr_mtx); 2751 zio_nowait(zio); 2752} 2753 2754static void 2755dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx) 2756{ 2757 arc_buf_t **datap = &dr->dt.dl.dr_data; 2758 dmu_buf_impl_t *db = dr->dr_dbuf; 2759 dnode_t *dn; 2760 objset_t *os; 2761 uint64_t txg = tx->tx_txg; 2762 2763 ASSERT(dmu_tx_is_syncing(tx)); 2764 2765 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 2766 2767 mutex_enter(&db->db_mtx); 2768 /* 2769 * To be synced, we must be dirtied. But we 2770 * might have been freed after the dirty. 2771 */ 2772 if (db->db_state == DB_UNCACHED) { 2773 /* This buffer has been freed since it was dirtied */ 2774 ASSERT(db->db.db_data == NULL); 2775 } else if (db->db_state == DB_FILL) { 2776 /* This buffer was freed and is now being re-filled */ 2777 ASSERT(db->db.db_data != dr->dt.dl.dr_data); 2778 } else { 2779 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL); 2780 } 2781 DBUF_VERIFY(db); 2782 2783 DB_DNODE_ENTER(db); 2784 dn = DB_DNODE(db); 2785 2786 if (db->db_blkid == DMU_SPILL_BLKID) { 2787 mutex_enter(&dn->dn_mtx); 2788 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR; 2789 mutex_exit(&dn->dn_mtx); 2790 } 2791 2792 /* 2793 * If this is a bonus buffer, simply copy the bonus data into the 2794 * dnode. It will be written out when the dnode is synced (and it 2795 * will be synced, since it must have been dirty for dbuf_sync to 2796 * be called). 2797 */ 2798 if (db->db_blkid == DMU_BONUS_BLKID) { 2799 dbuf_dirty_record_t **drp; 2800 2801 ASSERT(*datap != NULL); 2802 ASSERT0(db->db_level); 2803 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN); 2804 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen); 2805 DB_DNODE_EXIT(db); 2806 2807 if (*datap != db->db.db_data) { 2808 zio_buf_free(*datap, DN_MAX_BONUSLEN); 2809 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 2810 } 2811 db->db_data_pending = NULL; 2812 drp = &db->db_last_dirty; 2813 while (*drp != dr) 2814 drp = &(*drp)->dr_next; 2815 ASSERT(dr->dr_next == NULL); 2816 ASSERT(dr->dr_dbuf == db); 2817 *drp = dr->dr_next; 2818 if (dr->dr_dbuf->db_level != 0) { 2819 list_destroy(&dr->dt.di.dr_children); 2820 mutex_destroy(&dr->dt.di.dr_mtx); 2821 } 2822 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 2823 ASSERT(db->db_dirtycnt > 0); 2824 db->db_dirtycnt -= 1; 2825 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg); 2826 return; 2827 } 2828 2829 os = dn->dn_objset; 2830 2831 /* 2832 * This function may have dropped the db_mtx lock allowing a dmu_sync 2833 * operation to sneak in. As a result, we need to ensure that we 2834 * don't check the dr_override_state until we have returned from 2835 * dbuf_check_blkptr. 2836 */ 2837 dbuf_check_blkptr(dn, db); 2838 2839 /* 2840 * If this buffer is in the middle of an immediate write, 2841 * wait for the synchronous IO to complete. 2842 */ 2843 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) { 2844 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT); 2845 cv_wait(&db->db_changed, &db->db_mtx); 2846 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN); 2847 } 2848 2849 if (db->db_state != DB_NOFILL && 2850 dn->dn_object != DMU_META_DNODE_OBJECT && 2851 refcount_count(&db->db_holds) > 1 && 2852 dr->dt.dl.dr_override_state != DR_OVERRIDDEN && 2853 *datap == db->db_buf) { 2854 /* 2855 * If this buffer is currently "in use" (i.e., there 2856 * are active holds and db_data still references it), 2857 * then make a copy before we start the write so that 2858 * any modifications from the open txg will not leak 2859 * into this write. 2860 * 2861 * NOTE: this copy does not need to be made for 2862 * objects only modified in the syncing context (e.g. 2863 * DNONE_DNODE blocks). 2864 */ 2865 int blksz = arc_buf_size(*datap); 2866 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 2867 *datap = arc_buf_alloc(os->os_spa, blksz, db, type); 2868 bcopy(db->db.db_data, (*datap)->b_data, blksz); 2869 } 2870 db->db_data_pending = dr; 2871 2872 mutex_exit(&db->db_mtx); 2873 2874 dbuf_write(dr, *datap, tx); 2875 2876 ASSERT(!list_link_active(&dr->dr_dirty_node)); 2877 if (dn->dn_object == DMU_META_DNODE_OBJECT) { 2878 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr); 2879 DB_DNODE_EXIT(db); 2880 } else { 2881 /* 2882 * Although zio_nowait() does not "wait for an IO", it does 2883 * initiate the IO. If this is an empty write it seems plausible 2884 * that the IO could actually be completed before the nowait 2885 * returns. We need to DB_DNODE_EXIT() first in case 2886 * zio_nowait() invalidates the dbuf. 2887 */ 2888 DB_DNODE_EXIT(db); 2889 zio_nowait(dr->dr_zio); 2890 } 2891} 2892 2893void 2894dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx) 2895{ 2896 dbuf_dirty_record_t *dr; 2897 2898 while (dr = list_head(list)) { 2899 if (dr->dr_zio != NULL) { 2900 /* 2901 * If we find an already initialized zio then we 2902 * are processing the meta-dnode, and we have finished. 2903 * The dbufs for all dnodes are put back on the list 2904 * during processing, so that we can zio_wait() 2905 * these IOs after initiating all child IOs. 2906 */ 2907 ASSERT3U(dr->dr_dbuf->db.db_object, ==, 2908 DMU_META_DNODE_OBJECT); 2909 break; 2910 } 2911 if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID && 2912 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) { 2913 VERIFY3U(dr->dr_dbuf->db_level, ==, level); 2914 } 2915 list_remove(list, dr); 2916 if (dr->dr_dbuf->db_level > 0) 2917 dbuf_sync_indirect(dr, tx); 2918 else 2919 dbuf_sync_leaf(dr, tx); 2920 } 2921} 2922 2923/* ARGSUSED */ 2924static void 2925dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb) 2926{ 2927 dmu_buf_impl_t *db = vdb; 2928 dnode_t *dn; 2929 blkptr_t *bp = zio->io_bp; 2930 blkptr_t *bp_orig = &zio->io_bp_orig; 2931 spa_t *spa = zio->io_spa; 2932 int64_t delta; 2933 uint64_t fill = 0; 2934 int i; 2935 2936 ASSERT3P(db->db_blkptr, !=, NULL); 2937 ASSERT3P(&db->db_data_pending->dr_bp_copy, ==, bp); 2938 2939 DB_DNODE_ENTER(db); 2940 dn = DB_DNODE(db); 2941 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig); 2942 dnode_diduse_space(dn, delta - zio->io_prev_space_delta); 2943 zio->io_prev_space_delta = delta; 2944 2945 if (bp->blk_birth != 0) { 2946 ASSERT((db->db_blkid != DMU_SPILL_BLKID && 2947 BP_GET_TYPE(bp) == dn->dn_type) || 2948 (db->db_blkid == DMU_SPILL_BLKID && 2949 BP_GET_TYPE(bp) == dn->dn_bonustype) || 2950 BP_IS_EMBEDDED(bp)); 2951 ASSERT(BP_GET_LEVEL(bp) == db->db_level); 2952 } 2953 2954 mutex_enter(&db->db_mtx); 2955 2956#ifdef ZFS_DEBUG 2957 if (db->db_blkid == DMU_SPILL_BLKID) { 2958 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); 2959 ASSERT(!(BP_IS_HOLE(bp)) && 2960 db->db_blkptr == &dn->dn_phys->dn_spill); 2961 } 2962#endif 2963 2964 if (db->db_level == 0) { 2965 mutex_enter(&dn->dn_mtx); 2966 if (db->db_blkid > dn->dn_phys->dn_maxblkid && 2967 db->db_blkid != DMU_SPILL_BLKID) 2968 dn->dn_phys->dn_maxblkid = db->db_blkid; 2969 mutex_exit(&dn->dn_mtx); 2970 2971 if (dn->dn_type == DMU_OT_DNODE) { 2972 dnode_phys_t *dnp = db->db.db_data; 2973 for (i = db->db.db_size >> DNODE_SHIFT; i > 0; 2974 i--, dnp++) { 2975 if (dnp->dn_type != DMU_OT_NONE) 2976 fill++; 2977 } 2978 } else { 2979 if (BP_IS_HOLE(bp)) { 2980 fill = 0; 2981 } else { 2982 fill = 1; 2983 } 2984 } 2985 } else { 2986 blkptr_t *ibp = db->db.db_data; 2987 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 2988 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) { 2989 if (BP_IS_HOLE(ibp)) 2990 continue; 2991 fill += BP_GET_FILL(ibp); 2992 } 2993 } 2994 DB_DNODE_EXIT(db); 2995 2996 if (!BP_IS_EMBEDDED(bp)) 2997 bp->blk_fill = fill; 2998 2999 mutex_exit(&db->db_mtx); 3000 3001 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 3002 *db->db_blkptr = *bp; 3003 rw_exit(&dn->dn_struct_rwlock); 3004} 3005 3006/* ARGSUSED */ 3007/* 3008 * This function gets called just prior to running through the compression 3009 * stage of the zio pipeline. If we're an indirect block comprised of only 3010 * holes, then we want this indirect to be compressed away to a hole. In 3011 * order to do that we must zero out any information about the holes that 3012 * this indirect points to prior to before we try to compress it. 3013 */ 3014static void 3015dbuf_write_children_ready(zio_t *zio, arc_buf_t *buf, void *vdb) 3016{ 3017 dmu_buf_impl_t *db = vdb; 3018 dnode_t *dn; 3019 blkptr_t *bp; 3020 uint64_t i; 3021 int epbs; 3022 3023 ASSERT3U(db->db_level, >, 0); 3024 DB_DNODE_ENTER(db); 3025 dn = DB_DNODE(db); 3026 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 3027 3028 /* Determine if all our children are holes */ 3029 for (i = 0, bp = db->db.db_data; i < 1 << epbs; i++, bp++) { 3030 if (!BP_IS_HOLE(bp)) 3031 break; 3032 } 3033 3034 /* 3035 * If all the children are holes, then zero them all out so that 3036 * we may get compressed away. 3037 */ 3038 if (i == 1 << epbs) { 3039 /* didn't find any non-holes */ 3040 bzero(db->db.db_data, db->db.db_size); 3041 } 3042 DB_DNODE_EXIT(db); 3043} 3044 3045/* 3046 * The SPA will call this callback several times for each zio - once 3047 * for every physical child i/o (zio->io_phys_children times). This 3048 * allows the DMU to monitor the progress of each logical i/o. For example, 3049 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z 3050 * block. There may be a long delay before all copies/fragments are completed, 3051 * so this callback allows us to retire dirty space gradually, as the physical 3052 * i/os complete. 3053 */ 3054/* ARGSUSED */ 3055static void 3056dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg) 3057{ 3058 dmu_buf_impl_t *db = arg; 3059 objset_t *os = db->db_objset; 3060 dsl_pool_t *dp = dmu_objset_pool(os); 3061 dbuf_dirty_record_t *dr; 3062 int delta = 0; 3063 3064 dr = db->db_data_pending; 3065 ASSERT3U(dr->dr_txg, ==, zio->io_txg); 3066 3067 /* 3068 * The callback will be called io_phys_children times. Retire one 3069 * portion of our dirty space each time we are called. Any rounding 3070 * error will be cleaned up by dsl_pool_sync()'s call to 3071 * dsl_pool_undirty_space(). 3072 */ 3073 delta = dr->dr_accounted / zio->io_phys_children; 3074 dsl_pool_undirty_space(dp, delta, zio->io_txg); 3075} 3076 3077/* ARGSUSED */ 3078static void 3079dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb) 3080{ 3081 dmu_buf_impl_t *db = vdb; 3082 blkptr_t *bp_orig = &zio->io_bp_orig; 3083 blkptr_t *bp = db->db_blkptr; 3084 objset_t *os = db->db_objset; 3085 dmu_tx_t *tx = os->os_synctx; 3086 dbuf_dirty_record_t **drp, *dr; 3087 3088 ASSERT0(zio->io_error); 3089 ASSERT(db->db_blkptr == bp); 3090 3091 /* 3092 * For nopwrites and rewrites we ensure that the bp matches our 3093 * original and bypass all the accounting. 3094 */ 3095 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) { 3096 ASSERT(BP_EQUAL(bp, bp_orig)); 3097 } else { 3098 dsl_dataset_t *ds = os->os_dsl_dataset; 3099 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE); 3100 dsl_dataset_block_born(ds, bp, tx); 3101 } 3102 3103 mutex_enter(&db->db_mtx); 3104 3105 DBUF_VERIFY(db); 3106 3107 drp = &db->db_last_dirty; 3108 while ((dr = *drp) != db->db_data_pending) 3109 drp = &dr->dr_next; 3110 ASSERT(!list_link_active(&dr->dr_dirty_node)); 3111 ASSERT(dr->dr_dbuf == db); 3112 ASSERT(dr->dr_next == NULL); 3113 *drp = dr->dr_next; 3114 3115#ifdef ZFS_DEBUG 3116 if (db->db_blkid == DMU_SPILL_BLKID) { 3117 dnode_t *dn; 3118 3119 DB_DNODE_ENTER(db); 3120 dn = DB_DNODE(db); 3121 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); 3122 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && 3123 db->db_blkptr == &dn->dn_phys->dn_spill); 3124 DB_DNODE_EXIT(db); 3125 } 3126#endif 3127 3128 if (db->db_level == 0) { 3129 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 3130 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); 3131 if (db->db_state != DB_NOFILL) { 3132 if (dr->dt.dl.dr_data != db->db_buf) 3133 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, 3134 db)); 3135 else if (!arc_released(db->db_buf)) 3136 arc_set_callback(db->db_buf, dbuf_do_evict, db); 3137 } 3138 } else { 3139 dnode_t *dn; 3140 3141 DB_DNODE_ENTER(db); 3142 dn = DB_DNODE(db); 3143 ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 3144 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift); 3145 if (!BP_IS_HOLE(db->db_blkptr)) { 3146 int epbs = 3147 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 3148 ASSERT3U(db->db_blkid, <=, 3149 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs)); 3150 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, 3151 db->db.db_size); 3152 if (!arc_released(db->db_buf)) 3153 arc_set_callback(db->db_buf, dbuf_do_evict, db); 3154 } 3155 DB_DNODE_EXIT(db); 3156 mutex_destroy(&dr->dt.di.dr_mtx); 3157 list_destroy(&dr->dt.di.dr_children); 3158 } 3159 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 3160 3161 cv_broadcast(&db->db_changed); 3162 ASSERT(db->db_dirtycnt > 0); 3163 db->db_dirtycnt -= 1; 3164 db->db_data_pending = NULL; 3165 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg); 3166} 3167 3168static void 3169dbuf_write_nofill_ready(zio_t *zio) 3170{ 3171 dbuf_write_ready(zio, NULL, zio->io_private); 3172} 3173 3174static void 3175dbuf_write_nofill_done(zio_t *zio) 3176{ 3177 dbuf_write_done(zio, NULL, zio->io_private); 3178} 3179 3180static void 3181dbuf_write_override_ready(zio_t *zio) 3182{ 3183 dbuf_dirty_record_t *dr = zio->io_private; 3184 dmu_buf_impl_t *db = dr->dr_dbuf; 3185 3186 dbuf_write_ready(zio, NULL, db); 3187} 3188 3189static void 3190dbuf_write_override_done(zio_t *zio) 3191{ 3192 dbuf_dirty_record_t *dr = zio->io_private; 3193 dmu_buf_impl_t *db = dr->dr_dbuf; 3194 blkptr_t *obp = &dr->dt.dl.dr_overridden_by; 3195 3196 mutex_enter(&db->db_mtx); 3197 if (!BP_EQUAL(zio->io_bp, obp)) { 3198 if (!BP_IS_HOLE(obp)) 3199 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp); 3200 arc_release(dr->dt.dl.dr_data, db); 3201 } 3202 mutex_exit(&db->db_mtx); 3203 3204 dbuf_write_done(zio, NULL, db); 3205} 3206 3207/* Issue I/O to commit a dirty buffer to disk. */ 3208static void 3209dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx) 3210{ 3211 dmu_buf_impl_t *db = dr->dr_dbuf; 3212 dnode_t *dn; 3213 objset_t *os; 3214 dmu_buf_impl_t *parent = db->db_parent; 3215 uint64_t txg = tx->tx_txg; 3216 zbookmark_phys_t zb; 3217 zio_prop_t zp; 3218 zio_t *zio; 3219 int wp_flag = 0; 3220 3221 ASSERT(dmu_tx_is_syncing(tx)); 3222 3223 DB_DNODE_ENTER(db); 3224 dn = DB_DNODE(db); 3225 os = dn->dn_objset; 3226 3227 if (db->db_state != DB_NOFILL) { 3228 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) { 3229 /* 3230 * Private object buffers are released here rather 3231 * than in dbuf_dirty() since they are only modified 3232 * in the syncing context and we don't want the 3233 * overhead of making multiple copies of the data. 3234 */ 3235 if (BP_IS_HOLE(db->db_blkptr)) { 3236 arc_buf_thaw(data); 3237 } else { 3238 dbuf_release_bp(db); 3239 } 3240 } 3241 } 3242 3243 if (parent != dn->dn_dbuf) { 3244 /* Our parent is an indirect block. */ 3245 /* We have a dirty parent that has been scheduled for write. */ 3246 ASSERT(parent && parent->db_data_pending); 3247 /* Our parent's buffer is one level closer to the dnode. */ 3248 ASSERT(db->db_level == parent->db_level-1); 3249 /* 3250 * We're about to modify our parent's db_data by modifying 3251 * our block pointer, so the parent must be released. 3252 */ 3253 ASSERT(arc_released(parent->db_buf)); 3254 zio = parent->db_data_pending->dr_zio; 3255 } else { 3256 /* Our parent is the dnode itself. */ 3257 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 && 3258 db->db_blkid != DMU_SPILL_BLKID) || 3259 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0)); 3260 if (db->db_blkid != DMU_SPILL_BLKID) 3261 ASSERT3P(db->db_blkptr, ==, 3262 &dn->dn_phys->dn_blkptr[db->db_blkid]); 3263 zio = dn->dn_zio; 3264 } 3265 3266 ASSERT(db->db_level == 0 || data == db->db_buf); 3267 ASSERT3U(db->db_blkptr->blk_birth, <=, txg); 3268 ASSERT(zio); 3269 3270 SET_BOOKMARK(&zb, os->os_dsl_dataset ? 3271 os->os_dsl_dataset->ds_object : DMU_META_OBJSET, 3272 db->db.db_object, db->db_level, db->db_blkid); 3273 3274 if (db->db_blkid == DMU_SPILL_BLKID) 3275 wp_flag = WP_SPILL; 3276 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0; 3277 3278 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp); 3279 DB_DNODE_EXIT(db); 3280 3281 /* 3282 * We copy the blkptr now (rather than when we instantiate the dirty 3283 * record), because its value can change between open context and 3284 * syncing context. We do not need to hold dn_struct_rwlock to read 3285 * db_blkptr because we are in syncing context. 3286 */ 3287 dr->dr_bp_copy = *db->db_blkptr; 3288 3289 if (db->db_level == 0 && 3290 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { 3291 /* 3292 * The BP for this block has been provided by open context 3293 * (by dmu_sync() or dmu_buf_write_embedded()). 3294 */ 3295 void *contents = (data != NULL) ? data->b_data : NULL; 3296 3297 dr->dr_zio = zio_write(zio, os->os_spa, txg, 3298 &dr->dr_bp_copy, contents, db->db.db_size, &zp, 3299 dbuf_write_override_ready, NULL, NULL, 3300 dbuf_write_override_done, 3301 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); 3302 mutex_enter(&db->db_mtx); 3303 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 3304 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by, 3305 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite); 3306 mutex_exit(&db->db_mtx); 3307 } else if (db->db_state == DB_NOFILL) { 3308 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF || 3309 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY); 3310 dr->dr_zio = zio_write(zio, os->os_spa, txg, 3311 &dr->dr_bp_copy, NULL, db->db.db_size, &zp, 3312 dbuf_write_nofill_ready, NULL, NULL, 3313 dbuf_write_nofill_done, db, 3314 ZIO_PRIORITY_ASYNC_WRITE, 3315 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb); 3316 } else { 3317 ASSERT(arc_released(data)); 3318 3319 /* 3320 * For indirect blocks, we want to setup the children 3321 * ready callback so that we can properly handle an indirect 3322 * block that only contains holes. 3323 */ 3324 arc_done_func_t *children_ready_cb = NULL; 3325 if (db->db_level != 0) 3326 children_ready_cb = dbuf_write_children_ready; 3327 3328 dr->dr_zio = arc_write(zio, os->os_spa, txg, 3329 &dr->dr_bp_copy, data, DBUF_IS_L2CACHEABLE(db), 3330 DBUF_IS_L2COMPRESSIBLE(db), &zp, dbuf_write_ready, 3331 children_ready_cb, 3332 dbuf_write_physdone, dbuf_write_done, db, 3333 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); 3334 } 3335} 3336