dbuf.c revision 288594
1/* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21/* 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23 * Copyright 2011 Nexenta Systems, Inc. All rights reserved. 24 * Copyright (c) 2012, 2015 by Delphix. All rights reserved. 25 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved. 26 * Copyright (c) 2013, Joyent, Inc. All rights reserved. 27 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved. 28 */ 29 30#include <sys/zfs_context.h> 31#include <sys/dmu.h> 32#include <sys/dmu_send.h> 33#include <sys/dmu_impl.h> 34#include <sys/dbuf.h> 35#include <sys/dmu_objset.h> 36#include <sys/dsl_dataset.h> 37#include <sys/dsl_dir.h> 38#include <sys/dmu_tx.h> 39#include <sys/spa.h> 40#include <sys/zio.h> 41#include <sys/dmu_zfetch.h> 42#include <sys/sa.h> 43#include <sys/sa_impl.h> 44#include <sys/zfeature.h> 45#include <sys/blkptr.h> 46#include <sys/range_tree.h> 47 48/* 49 * Number of times that zfs_free_range() took the slow path while doing 50 * a zfs receive. A nonzero value indicates a potential performance problem. 51 */ 52uint64_t zfs_free_range_recv_miss; 53 54static void dbuf_destroy(dmu_buf_impl_t *db); 55static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx); 56static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx); 57 58/* 59 * Global data structures and functions for the dbuf cache. 60 */ 61static kmem_cache_t *dbuf_cache; 62static taskq_t *dbu_evict_taskq; 63 64/* ARGSUSED */ 65static int 66dbuf_cons(void *vdb, void *unused, int kmflag) 67{ 68 dmu_buf_impl_t *db = vdb; 69 bzero(db, sizeof (dmu_buf_impl_t)); 70 71 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL); 72 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL); 73 refcount_create(&db->db_holds); 74 75 return (0); 76} 77 78/* ARGSUSED */ 79static void 80dbuf_dest(void *vdb, void *unused) 81{ 82 dmu_buf_impl_t *db = vdb; 83 mutex_destroy(&db->db_mtx); 84 cv_destroy(&db->db_changed); 85 refcount_destroy(&db->db_holds); 86} 87 88/* 89 * dbuf hash table routines 90 */ 91static dbuf_hash_table_t dbuf_hash_table; 92 93static uint64_t dbuf_hash_count; 94 95static uint64_t 96dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid) 97{ 98 uintptr_t osv = (uintptr_t)os; 99 uint64_t crc = -1ULL; 100 101 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 102 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF]; 103 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF]; 104 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF]; 105 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF]; 106 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF]; 107 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF]; 108 109 crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16); 110 111 return (crc); 112} 113 114#define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid); 115 116#define DBUF_EQUAL(dbuf, os, obj, level, blkid) \ 117 ((dbuf)->db.db_object == (obj) && \ 118 (dbuf)->db_objset == (os) && \ 119 (dbuf)->db_level == (level) && \ 120 (dbuf)->db_blkid == (blkid)) 121 122dmu_buf_impl_t * 123dbuf_find(objset_t *os, uint64_t obj, uint8_t level, uint64_t blkid) 124{ 125 dbuf_hash_table_t *h = &dbuf_hash_table; 126 uint64_t hv = DBUF_HASH(os, obj, level, blkid); 127 uint64_t idx = hv & h->hash_table_mask; 128 dmu_buf_impl_t *db; 129 130 mutex_enter(DBUF_HASH_MUTEX(h, idx)); 131 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) { 132 if (DBUF_EQUAL(db, os, obj, level, blkid)) { 133 mutex_enter(&db->db_mtx); 134 if (db->db_state != DB_EVICTING) { 135 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 136 return (db); 137 } 138 mutex_exit(&db->db_mtx); 139 } 140 } 141 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 142 return (NULL); 143} 144 145static dmu_buf_impl_t * 146dbuf_find_bonus(objset_t *os, uint64_t object) 147{ 148 dnode_t *dn; 149 dmu_buf_impl_t *db = NULL; 150 151 if (dnode_hold(os, object, FTAG, &dn) == 0) { 152 rw_enter(&dn->dn_struct_rwlock, RW_READER); 153 if (dn->dn_bonus != NULL) { 154 db = dn->dn_bonus; 155 mutex_enter(&db->db_mtx); 156 } 157 rw_exit(&dn->dn_struct_rwlock); 158 dnode_rele(dn, FTAG); 159 } 160 return (db); 161} 162 163/* 164 * Insert an entry into the hash table. If there is already an element 165 * equal to elem in the hash table, then the already existing element 166 * will be returned and the new element will not be inserted. 167 * Otherwise returns NULL. 168 */ 169static dmu_buf_impl_t * 170dbuf_hash_insert(dmu_buf_impl_t *db) 171{ 172 dbuf_hash_table_t *h = &dbuf_hash_table; 173 objset_t *os = db->db_objset; 174 uint64_t obj = db->db.db_object; 175 int level = db->db_level; 176 uint64_t blkid = db->db_blkid; 177 uint64_t hv = DBUF_HASH(os, obj, level, blkid); 178 uint64_t idx = hv & h->hash_table_mask; 179 dmu_buf_impl_t *dbf; 180 181 mutex_enter(DBUF_HASH_MUTEX(h, idx)); 182 for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) { 183 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) { 184 mutex_enter(&dbf->db_mtx); 185 if (dbf->db_state != DB_EVICTING) { 186 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 187 return (dbf); 188 } 189 mutex_exit(&dbf->db_mtx); 190 } 191 } 192 193 mutex_enter(&db->db_mtx); 194 db->db_hash_next = h->hash_table[idx]; 195 h->hash_table[idx] = db; 196 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 197 atomic_inc_64(&dbuf_hash_count); 198 199 return (NULL); 200} 201 202/* 203 * Remove an entry from the hash table. It must be in the EVICTING state. 204 */ 205static void 206dbuf_hash_remove(dmu_buf_impl_t *db) 207{ 208 dbuf_hash_table_t *h = &dbuf_hash_table; 209 uint64_t hv = DBUF_HASH(db->db_objset, db->db.db_object, 210 db->db_level, db->db_blkid); 211 uint64_t idx = hv & h->hash_table_mask; 212 dmu_buf_impl_t *dbf, **dbp; 213 214 /* 215 * We musn't hold db_mtx to maintain lock ordering: 216 * DBUF_HASH_MUTEX > db_mtx. 217 */ 218 ASSERT(refcount_is_zero(&db->db_holds)); 219 ASSERT(db->db_state == DB_EVICTING); 220 ASSERT(!MUTEX_HELD(&db->db_mtx)); 221 222 mutex_enter(DBUF_HASH_MUTEX(h, idx)); 223 dbp = &h->hash_table[idx]; 224 while ((dbf = *dbp) != db) { 225 dbp = &dbf->db_hash_next; 226 ASSERT(dbf != NULL); 227 } 228 *dbp = db->db_hash_next; 229 db->db_hash_next = NULL; 230 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 231 atomic_dec_64(&dbuf_hash_count); 232} 233 234static arc_evict_func_t dbuf_do_evict; 235 236typedef enum { 237 DBVU_EVICTING, 238 DBVU_NOT_EVICTING 239} dbvu_verify_type_t; 240 241static void 242dbuf_verify_user(dmu_buf_impl_t *db, dbvu_verify_type_t verify_type) 243{ 244#ifdef ZFS_DEBUG 245 int64_t holds; 246 247 if (db->db_user == NULL) 248 return; 249 250 /* Only data blocks support the attachment of user data. */ 251 ASSERT(db->db_level == 0); 252 253 /* Clients must resolve a dbuf before attaching user data. */ 254 ASSERT(db->db.db_data != NULL); 255 ASSERT3U(db->db_state, ==, DB_CACHED); 256 257 holds = refcount_count(&db->db_holds); 258 if (verify_type == DBVU_EVICTING) { 259 /* 260 * Immediate eviction occurs when holds == dirtycnt. 261 * For normal eviction buffers, holds is zero on 262 * eviction, except when dbuf_fix_old_data() calls 263 * dbuf_clear_data(). However, the hold count can grow 264 * during eviction even though db_mtx is held (see 265 * dmu_bonus_hold() for an example), so we can only 266 * test the generic invariant that holds >= dirtycnt. 267 */ 268 ASSERT3U(holds, >=, db->db_dirtycnt); 269 } else { 270 if (db->db_immediate_evict == TRUE) 271 ASSERT3U(holds, >=, db->db_dirtycnt); 272 else 273 ASSERT3U(holds, >, 0); 274 } 275#endif 276} 277 278static void 279dbuf_evict_user(dmu_buf_impl_t *db) 280{ 281 dmu_buf_user_t *dbu = db->db_user; 282 283 ASSERT(MUTEX_HELD(&db->db_mtx)); 284 285 if (dbu == NULL) 286 return; 287 288 dbuf_verify_user(db, DBVU_EVICTING); 289 db->db_user = NULL; 290 291#ifdef ZFS_DEBUG 292 if (dbu->dbu_clear_on_evict_dbufp != NULL) 293 *dbu->dbu_clear_on_evict_dbufp = NULL; 294#endif 295 296 /* 297 * Invoke the callback from a taskq to avoid lock order reversals 298 * and limit stack depth. 299 */ 300 taskq_dispatch_ent(dbu_evict_taskq, dbu->dbu_evict_func, dbu, 0, 301 &dbu->dbu_tqent); 302} 303 304boolean_t 305dbuf_is_metadata(dmu_buf_impl_t *db) 306{ 307 if (db->db_level > 0) { 308 return (B_TRUE); 309 } else { 310 boolean_t is_metadata; 311 312 DB_DNODE_ENTER(db); 313 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type); 314 DB_DNODE_EXIT(db); 315 316 return (is_metadata); 317 } 318} 319 320void 321dbuf_evict(dmu_buf_impl_t *db) 322{ 323 ASSERT(MUTEX_HELD(&db->db_mtx)); 324 ASSERT(db->db_buf == NULL); 325 ASSERT(db->db_data_pending == NULL); 326 327 dbuf_clear(db); 328 dbuf_destroy(db); 329} 330 331void 332dbuf_init(void) 333{ 334 uint64_t hsize = 1ULL << 16; 335 dbuf_hash_table_t *h = &dbuf_hash_table; 336 int i; 337 338 /* 339 * The hash table is big enough to fill all of physical memory 340 * with an average 4K block size. The table will take up 341 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers). 342 */ 343 while (hsize * 4096 < (uint64_t)physmem * PAGESIZE) 344 hsize <<= 1; 345 346retry: 347 h->hash_table_mask = hsize - 1; 348 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP); 349 if (h->hash_table == NULL) { 350 /* XXX - we should really return an error instead of assert */ 351 ASSERT(hsize > (1ULL << 10)); 352 hsize >>= 1; 353 goto retry; 354 } 355 356 dbuf_cache = kmem_cache_create("dmu_buf_impl_t", 357 sizeof (dmu_buf_impl_t), 358 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0); 359 360 for (i = 0; i < DBUF_MUTEXES; i++) 361 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL); 362 363 /* 364 * All entries are queued via taskq_dispatch_ent(), so min/maxalloc 365 * configuration is not required. 366 */ 367 dbu_evict_taskq = taskq_create("dbu_evict", 1, minclsyspri, 0, 0, 0); 368} 369 370void 371dbuf_fini(void) 372{ 373 dbuf_hash_table_t *h = &dbuf_hash_table; 374 int i; 375 376 for (i = 0; i < DBUF_MUTEXES; i++) 377 mutex_destroy(&h->hash_mutexes[i]); 378 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *)); 379 kmem_cache_destroy(dbuf_cache); 380 taskq_destroy(dbu_evict_taskq); 381} 382 383/* 384 * Other stuff. 385 */ 386 387#ifdef ZFS_DEBUG 388static void 389dbuf_verify(dmu_buf_impl_t *db) 390{ 391 dnode_t *dn; 392 dbuf_dirty_record_t *dr; 393 394 ASSERT(MUTEX_HELD(&db->db_mtx)); 395 396 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY)) 397 return; 398 399 ASSERT(db->db_objset != NULL); 400 DB_DNODE_ENTER(db); 401 dn = DB_DNODE(db); 402 if (dn == NULL) { 403 ASSERT(db->db_parent == NULL); 404 ASSERT(db->db_blkptr == NULL); 405 } else { 406 ASSERT3U(db->db.db_object, ==, dn->dn_object); 407 ASSERT3P(db->db_objset, ==, dn->dn_objset); 408 ASSERT3U(db->db_level, <, dn->dn_nlevels); 409 ASSERT(db->db_blkid == DMU_BONUS_BLKID || 410 db->db_blkid == DMU_SPILL_BLKID || 411 !avl_is_empty(&dn->dn_dbufs)); 412 } 413 if (db->db_blkid == DMU_BONUS_BLKID) { 414 ASSERT(dn != NULL); 415 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 416 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID); 417 } else if (db->db_blkid == DMU_SPILL_BLKID) { 418 ASSERT(dn != NULL); 419 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 420 ASSERT0(db->db.db_offset); 421 } else { 422 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size); 423 } 424 425 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next) 426 ASSERT(dr->dr_dbuf == db); 427 428 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next) 429 ASSERT(dr->dr_dbuf == db); 430 431 /* 432 * We can't assert that db_size matches dn_datablksz because it 433 * can be momentarily different when another thread is doing 434 * dnode_set_blksz(). 435 */ 436 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) { 437 dr = db->db_data_pending; 438 /* 439 * It should only be modified in syncing context, so 440 * make sure we only have one copy of the data. 441 */ 442 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf); 443 } 444 445 /* verify db->db_blkptr */ 446 if (db->db_blkptr) { 447 if (db->db_parent == dn->dn_dbuf) { 448 /* db is pointed to by the dnode */ 449 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */ 450 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object)) 451 ASSERT(db->db_parent == NULL); 452 else 453 ASSERT(db->db_parent != NULL); 454 if (db->db_blkid != DMU_SPILL_BLKID) 455 ASSERT3P(db->db_blkptr, ==, 456 &dn->dn_phys->dn_blkptr[db->db_blkid]); 457 } else { 458 /* db is pointed to by an indirect block */ 459 int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT; 460 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1); 461 ASSERT3U(db->db_parent->db.db_object, ==, 462 db->db.db_object); 463 /* 464 * dnode_grow_indblksz() can make this fail if we don't 465 * have the struct_rwlock. XXX indblksz no longer 466 * grows. safe to do this now? 467 */ 468 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) { 469 ASSERT3P(db->db_blkptr, ==, 470 ((blkptr_t *)db->db_parent->db.db_data + 471 db->db_blkid % epb)); 472 } 473 } 474 } 475 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) && 476 (db->db_buf == NULL || db->db_buf->b_data) && 477 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID && 478 db->db_state != DB_FILL && !dn->dn_free_txg) { 479 /* 480 * If the blkptr isn't set but they have nonzero data, 481 * it had better be dirty, otherwise we'll lose that 482 * data when we evict this buffer. 483 */ 484 if (db->db_dirtycnt == 0) { 485 uint64_t *buf = db->db.db_data; 486 int i; 487 488 for (i = 0; i < db->db.db_size >> 3; i++) { 489 ASSERT(buf[i] == 0); 490 } 491 } 492 } 493 DB_DNODE_EXIT(db); 494} 495#endif 496 497static void 498dbuf_clear_data(dmu_buf_impl_t *db) 499{ 500 ASSERT(MUTEX_HELD(&db->db_mtx)); 501 dbuf_evict_user(db); 502 db->db_buf = NULL; 503 db->db.db_data = NULL; 504 if (db->db_state != DB_NOFILL) 505 db->db_state = DB_UNCACHED; 506} 507 508static void 509dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf) 510{ 511 ASSERT(MUTEX_HELD(&db->db_mtx)); 512 ASSERT(buf != NULL); 513 514 db->db_buf = buf; 515 ASSERT(buf->b_data != NULL); 516 db->db.db_data = buf->b_data; 517 if (!arc_released(buf)) 518 arc_set_callback(buf, dbuf_do_evict, db); 519} 520 521/* 522 * Loan out an arc_buf for read. Return the loaned arc_buf. 523 */ 524arc_buf_t * 525dbuf_loan_arcbuf(dmu_buf_impl_t *db) 526{ 527 arc_buf_t *abuf; 528 529 mutex_enter(&db->db_mtx); 530 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) { 531 int blksz = db->db.db_size; 532 spa_t *spa = db->db_objset->os_spa; 533 534 mutex_exit(&db->db_mtx); 535 abuf = arc_loan_buf(spa, blksz); 536 bcopy(db->db.db_data, abuf->b_data, blksz); 537 } else { 538 abuf = db->db_buf; 539 arc_loan_inuse_buf(abuf, db); 540 dbuf_clear_data(db); 541 mutex_exit(&db->db_mtx); 542 } 543 return (abuf); 544} 545 546/* 547 * Calculate which level n block references the data at the level 0 offset 548 * provided. 549 */ 550uint64_t 551dbuf_whichblock(dnode_t *dn, int64_t level, uint64_t offset) 552{ 553 if (dn->dn_datablkshift != 0 && dn->dn_indblkshift != 0) { 554 /* 555 * The level n blkid is equal to the level 0 blkid divided by 556 * the number of level 0s in a level n block. 557 * 558 * The level 0 blkid is offset >> datablkshift = 559 * offset / 2^datablkshift. 560 * 561 * The number of level 0s in a level n is the number of block 562 * pointers in an indirect block, raised to the power of level. 563 * This is 2^(indblkshift - SPA_BLKPTRSHIFT)^level = 564 * 2^(level*(indblkshift - SPA_BLKPTRSHIFT)). 565 * 566 * Thus, the level n blkid is: offset / 567 * ((2^datablkshift)*(2^(level*(indblkshift - SPA_BLKPTRSHIFT))) 568 * = offset / 2^(datablkshift + level * 569 * (indblkshift - SPA_BLKPTRSHIFT)) 570 * = offset >> (datablkshift + level * 571 * (indblkshift - SPA_BLKPTRSHIFT)) 572 */ 573 return (offset >> (dn->dn_datablkshift + level * 574 (dn->dn_indblkshift - SPA_BLKPTRSHIFT))); 575 } else { 576 ASSERT3U(offset, <, dn->dn_datablksz); 577 return (0); 578 } 579} 580 581static void 582dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb) 583{ 584 dmu_buf_impl_t *db = vdb; 585 586 mutex_enter(&db->db_mtx); 587 ASSERT3U(db->db_state, ==, DB_READ); 588 /* 589 * All reads are synchronous, so we must have a hold on the dbuf 590 */ 591 ASSERT(refcount_count(&db->db_holds) > 0); 592 ASSERT(db->db_buf == NULL); 593 ASSERT(db->db.db_data == NULL); 594 if (db->db_level == 0 && db->db_freed_in_flight) { 595 /* we were freed in flight; disregard any error */ 596 arc_release(buf, db); 597 bzero(buf->b_data, db->db.db_size); 598 arc_buf_freeze(buf); 599 db->db_freed_in_flight = FALSE; 600 dbuf_set_data(db, buf); 601 db->db_state = DB_CACHED; 602 } else if (zio == NULL || zio->io_error == 0) { 603 dbuf_set_data(db, buf); 604 db->db_state = DB_CACHED; 605 } else { 606 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 607 ASSERT3P(db->db_buf, ==, NULL); 608 VERIFY(arc_buf_remove_ref(buf, db)); 609 db->db_state = DB_UNCACHED; 610 } 611 cv_broadcast(&db->db_changed); 612 dbuf_rele_and_unlock(db, NULL); 613} 614 615static void 616dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) 617{ 618 dnode_t *dn; 619 zbookmark_phys_t zb; 620 arc_flags_t aflags = ARC_FLAG_NOWAIT; 621 622 DB_DNODE_ENTER(db); 623 dn = DB_DNODE(db); 624 ASSERT(!refcount_is_zero(&db->db_holds)); 625 /* We need the struct_rwlock to prevent db_blkptr from changing. */ 626 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 627 ASSERT(MUTEX_HELD(&db->db_mtx)); 628 ASSERT(db->db_state == DB_UNCACHED); 629 ASSERT(db->db_buf == NULL); 630 631 if (db->db_blkid == DMU_BONUS_BLKID) { 632 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen); 633 634 ASSERT3U(bonuslen, <=, db->db.db_size); 635 db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN); 636 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 637 if (bonuslen < DN_MAX_BONUSLEN) 638 bzero(db->db.db_data, DN_MAX_BONUSLEN); 639 if (bonuslen) 640 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen); 641 DB_DNODE_EXIT(db); 642 db->db_state = DB_CACHED; 643 mutex_exit(&db->db_mtx); 644 return; 645 } 646 647 /* 648 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync() 649 * processes the delete record and clears the bp while we are waiting 650 * for the dn_mtx (resulting in a "no" from block_freed). 651 */ 652 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) || 653 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) || 654 BP_IS_HOLE(db->db_blkptr)))) { 655 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 656 657 DB_DNODE_EXIT(db); 658 dbuf_set_data(db, arc_buf_alloc(db->db_objset->os_spa, 659 db->db.db_size, db, type)); 660 bzero(db->db.db_data, db->db.db_size); 661 db->db_state = DB_CACHED; 662 mutex_exit(&db->db_mtx); 663 return; 664 } 665 666 DB_DNODE_EXIT(db); 667 668 db->db_state = DB_READ; 669 mutex_exit(&db->db_mtx); 670 671 if (DBUF_IS_L2CACHEABLE(db)) 672 aflags |= ARC_FLAG_L2CACHE; 673 if (DBUF_IS_L2COMPRESSIBLE(db)) 674 aflags |= ARC_FLAG_L2COMPRESS; 675 676 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ? 677 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET, 678 db->db.db_object, db->db_level, db->db_blkid); 679 680 dbuf_add_ref(db, NULL); 681 682 (void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr, 683 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ, 684 (flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED, 685 &aflags, &zb); 686} 687 688int 689dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) 690{ 691 int err = 0; 692 boolean_t havepzio = (zio != NULL); 693 boolean_t prefetch; 694 dnode_t *dn; 695 696 /* 697 * We don't have to hold the mutex to check db_state because it 698 * can't be freed while we have a hold on the buffer. 699 */ 700 ASSERT(!refcount_is_zero(&db->db_holds)); 701 702 if (db->db_state == DB_NOFILL) 703 return (SET_ERROR(EIO)); 704 705 DB_DNODE_ENTER(db); 706 dn = DB_DNODE(db); 707 if ((flags & DB_RF_HAVESTRUCT) == 0) 708 rw_enter(&dn->dn_struct_rwlock, RW_READER); 709 710 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 711 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL && 712 DBUF_IS_CACHEABLE(db); 713 714 mutex_enter(&db->db_mtx); 715 if (db->db_state == DB_CACHED) { 716 mutex_exit(&db->db_mtx); 717 if (prefetch) 718 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1); 719 if ((flags & DB_RF_HAVESTRUCT) == 0) 720 rw_exit(&dn->dn_struct_rwlock); 721 DB_DNODE_EXIT(db); 722 } else if (db->db_state == DB_UNCACHED) { 723 spa_t *spa = dn->dn_objset->os_spa; 724 725 if (zio == NULL) 726 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); 727 dbuf_read_impl(db, zio, flags); 728 729 /* dbuf_read_impl has dropped db_mtx for us */ 730 731 if (prefetch) 732 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1); 733 734 if ((flags & DB_RF_HAVESTRUCT) == 0) 735 rw_exit(&dn->dn_struct_rwlock); 736 DB_DNODE_EXIT(db); 737 738 if (!havepzio) 739 err = zio_wait(zio); 740 } else { 741 /* 742 * Another reader came in while the dbuf was in flight 743 * between UNCACHED and CACHED. Either a writer will finish 744 * writing the buffer (sending the dbuf to CACHED) or the 745 * first reader's request will reach the read_done callback 746 * and send the dbuf to CACHED. Otherwise, a failure 747 * occurred and the dbuf went to UNCACHED. 748 */ 749 mutex_exit(&db->db_mtx); 750 if (prefetch) 751 dmu_zfetch(&dn->dn_zfetch, db->db_blkid, 1); 752 if ((flags & DB_RF_HAVESTRUCT) == 0) 753 rw_exit(&dn->dn_struct_rwlock); 754 DB_DNODE_EXIT(db); 755 756 /* Skip the wait per the caller's request. */ 757 mutex_enter(&db->db_mtx); 758 if ((flags & DB_RF_NEVERWAIT) == 0) { 759 while (db->db_state == DB_READ || 760 db->db_state == DB_FILL) { 761 ASSERT(db->db_state == DB_READ || 762 (flags & DB_RF_HAVESTRUCT) == 0); 763 DTRACE_PROBE2(blocked__read, dmu_buf_impl_t *, 764 db, zio_t *, zio); 765 cv_wait(&db->db_changed, &db->db_mtx); 766 } 767 if (db->db_state == DB_UNCACHED) 768 err = SET_ERROR(EIO); 769 } 770 mutex_exit(&db->db_mtx); 771 } 772 773 ASSERT(err || havepzio || db->db_state == DB_CACHED); 774 return (err); 775} 776 777static void 778dbuf_noread(dmu_buf_impl_t *db) 779{ 780 ASSERT(!refcount_is_zero(&db->db_holds)); 781 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 782 mutex_enter(&db->db_mtx); 783 while (db->db_state == DB_READ || db->db_state == DB_FILL) 784 cv_wait(&db->db_changed, &db->db_mtx); 785 if (db->db_state == DB_UNCACHED) { 786 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 787 spa_t *spa = db->db_objset->os_spa; 788 789 ASSERT(db->db_buf == NULL); 790 ASSERT(db->db.db_data == NULL); 791 dbuf_set_data(db, arc_buf_alloc(spa, db->db.db_size, db, type)); 792 db->db_state = DB_FILL; 793 } else if (db->db_state == DB_NOFILL) { 794 dbuf_clear_data(db); 795 } else { 796 ASSERT3U(db->db_state, ==, DB_CACHED); 797 } 798 mutex_exit(&db->db_mtx); 799} 800 801/* 802 * This is our just-in-time copy function. It makes a copy of 803 * buffers, that have been modified in a previous transaction 804 * group, before we modify them in the current active group. 805 * 806 * This function is used in two places: when we are dirtying a 807 * buffer for the first time in a txg, and when we are freeing 808 * a range in a dnode that includes this buffer. 809 * 810 * Note that when we are called from dbuf_free_range() we do 811 * not put a hold on the buffer, we just traverse the active 812 * dbuf list for the dnode. 813 */ 814static void 815dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg) 816{ 817 dbuf_dirty_record_t *dr = db->db_last_dirty; 818 819 ASSERT(MUTEX_HELD(&db->db_mtx)); 820 ASSERT(db->db.db_data != NULL); 821 ASSERT(db->db_level == 0); 822 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT); 823 824 if (dr == NULL || 825 (dr->dt.dl.dr_data != 826 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf))) 827 return; 828 829 /* 830 * If the last dirty record for this dbuf has not yet synced 831 * and its referencing the dbuf data, either: 832 * reset the reference to point to a new copy, 833 * or (if there a no active holders) 834 * just null out the current db_data pointer. 835 */ 836 ASSERT(dr->dr_txg >= txg - 2); 837 if (db->db_blkid == DMU_BONUS_BLKID) { 838 /* Note that the data bufs here are zio_bufs */ 839 dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN); 840 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 841 bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN); 842 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) { 843 int size = db->db.db_size; 844 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 845 spa_t *spa = db->db_objset->os_spa; 846 847 dr->dt.dl.dr_data = arc_buf_alloc(spa, size, db, type); 848 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size); 849 } else { 850 dbuf_clear_data(db); 851 } 852} 853 854void 855dbuf_unoverride(dbuf_dirty_record_t *dr) 856{ 857 dmu_buf_impl_t *db = dr->dr_dbuf; 858 blkptr_t *bp = &dr->dt.dl.dr_overridden_by; 859 uint64_t txg = dr->dr_txg; 860 861 ASSERT(MUTEX_HELD(&db->db_mtx)); 862 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC); 863 ASSERT(db->db_level == 0); 864 865 if (db->db_blkid == DMU_BONUS_BLKID || 866 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN) 867 return; 868 869 ASSERT(db->db_data_pending != dr); 870 871 /* free this block */ 872 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite) 873 zio_free(db->db_objset->os_spa, txg, bp); 874 875 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 876 dr->dt.dl.dr_nopwrite = B_FALSE; 877 878 /* 879 * Release the already-written buffer, so we leave it in 880 * a consistent dirty state. Note that all callers are 881 * modifying the buffer, so they will immediately do 882 * another (redundant) arc_release(). Therefore, leave 883 * the buf thawed to save the effort of freezing & 884 * immediately re-thawing it. 885 */ 886 arc_release(dr->dt.dl.dr_data, db); 887} 888 889/* 890 * Evict (if its unreferenced) or clear (if its referenced) any level-0 891 * data blocks in the free range, so that any future readers will find 892 * empty blocks. 893 * 894 * This is a no-op if the dataset is in the middle of an incremental 895 * receive; see comment below for details. 896 */ 897void 898dbuf_free_range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid, 899 dmu_tx_t *tx) 900{ 901 dmu_buf_impl_t db_search; 902 dmu_buf_impl_t *db, *db_next; 903 uint64_t txg = tx->tx_txg; 904 avl_index_t where; 905 906 if (end_blkid > dn->dn_maxblkid && (end_blkid != DMU_SPILL_BLKID)) 907 end_blkid = dn->dn_maxblkid; 908 dprintf_dnode(dn, "start=%llu end=%llu\n", start_blkid, end_blkid); 909 910 db_search.db_level = 0; 911 db_search.db_blkid = start_blkid; 912 db_search.db_state = DB_SEARCH; 913 914 mutex_enter(&dn->dn_dbufs_mtx); 915 if (start_blkid >= dn->dn_unlisted_l0_blkid) { 916 /* There can't be any dbufs in this range; no need to search. */ 917#ifdef DEBUG 918 db = avl_find(&dn->dn_dbufs, &db_search, &where); 919 ASSERT3P(db, ==, NULL); 920 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER); 921 ASSERT(db == NULL || db->db_level > 0); 922#endif 923 mutex_exit(&dn->dn_dbufs_mtx); 924 return; 925 } else if (dmu_objset_is_receiving(dn->dn_objset)) { 926 /* 927 * If we are receiving, we expect there to be no dbufs in 928 * the range to be freed, because receive modifies each 929 * block at most once, and in offset order. If this is 930 * not the case, it can lead to performance problems, 931 * so note that we unexpectedly took the slow path. 932 */ 933 atomic_inc_64(&zfs_free_range_recv_miss); 934 } 935 936 db = avl_find(&dn->dn_dbufs, &db_search, &where); 937 ASSERT3P(db, ==, NULL); 938 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER); 939 940 for (; db != NULL; db = db_next) { 941 db_next = AVL_NEXT(&dn->dn_dbufs, db); 942 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 943 944 if (db->db_level != 0 || db->db_blkid > end_blkid) { 945 break; 946 } 947 ASSERT3U(db->db_blkid, >=, start_blkid); 948 949 /* found a level 0 buffer in the range */ 950 mutex_enter(&db->db_mtx); 951 if (dbuf_undirty(db, tx)) { 952 /* mutex has been dropped and dbuf destroyed */ 953 continue; 954 } 955 956 if (db->db_state == DB_UNCACHED || 957 db->db_state == DB_NOFILL || 958 db->db_state == DB_EVICTING) { 959 ASSERT(db->db.db_data == NULL); 960 mutex_exit(&db->db_mtx); 961 continue; 962 } 963 if (db->db_state == DB_READ || db->db_state == DB_FILL) { 964 /* will be handled in dbuf_read_done or dbuf_rele */ 965 db->db_freed_in_flight = TRUE; 966 mutex_exit(&db->db_mtx); 967 continue; 968 } 969 if (refcount_count(&db->db_holds) == 0) { 970 ASSERT(db->db_buf); 971 dbuf_clear(db); 972 continue; 973 } 974 /* The dbuf is referenced */ 975 976 if (db->db_last_dirty != NULL) { 977 dbuf_dirty_record_t *dr = db->db_last_dirty; 978 979 if (dr->dr_txg == txg) { 980 /* 981 * This buffer is "in-use", re-adjust the file 982 * size to reflect that this buffer may 983 * contain new data when we sync. 984 */ 985 if (db->db_blkid != DMU_SPILL_BLKID && 986 db->db_blkid > dn->dn_maxblkid) 987 dn->dn_maxblkid = db->db_blkid; 988 dbuf_unoverride(dr); 989 } else { 990 /* 991 * This dbuf is not dirty in the open context. 992 * Either uncache it (if its not referenced in 993 * the open context) or reset its contents to 994 * empty. 995 */ 996 dbuf_fix_old_data(db, txg); 997 } 998 } 999 /* clear the contents if its cached */ 1000 if (db->db_state == DB_CACHED) { 1001 ASSERT(db->db.db_data != NULL); 1002 arc_release(db->db_buf, db); 1003 bzero(db->db.db_data, db->db.db_size); 1004 arc_buf_freeze(db->db_buf); 1005 } 1006 1007 mutex_exit(&db->db_mtx); 1008 } 1009 mutex_exit(&dn->dn_dbufs_mtx); 1010} 1011 1012static int 1013dbuf_block_freeable(dmu_buf_impl_t *db) 1014{ 1015 dsl_dataset_t *ds = db->db_objset->os_dsl_dataset; 1016 uint64_t birth_txg = 0; 1017 1018 /* 1019 * We don't need any locking to protect db_blkptr: 1020 * If it's syncing, then db_last_dirty will be set 1021 * so we'll ignore db_blkptr. 1022 * 1023 * This logic ensures that only block births for 1024 * filled blocks are considered. 1025 */ 1026 ASSERT(MUTEX_HELD(&db->db_mtx)); 1027 if (db->db_last_dirty && (db->db_blkptr == NULL || 1028 !BP_IS_HOLE(db->db_blkptr))) { 1029 birth_txg = db->db_last_dirty->dr_txg; 1030 } else if (db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) { 1031 birth_txg = db->db_blkptr->blk_birth; 1032 } 1033 1034 /* 1035 * If this block don't exist or is in a snapshot, it can't be freed. 1036 * Don't pass the bp to dsl_dataset_block_freeable() since we 1037 * are holding the db_mtx lock and might deadlock if we are 1038 * prefetching a dedup-ed block. 1039 */ 1040 if (birth_txg != 0) 1041 return (ds == NULL || 1042 dsl_dataset_block_freeable(ds, NULL, birth_txg)); 1043 else 1044 return (B_FALSE); 1045} 1046 1047void 1048dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx) 1049{ 1050 arc_buf_t *buf, *obuf; 1051 int osize = db->db.db_size; 1052 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 1053 dnode_t *dn; 1054 1055 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1056 1057 DB_DNODE_ENTER(db); 1058 dn = DB_DNODE(db); 1059 1060 /* XXX does *this* func really need the lock? */ 1061 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 1062 1063 /* 1064 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held 1065 * is OK, because there can be no other references to the db 1066 * when we are changing its size, so no concurrent DB_FILL can 1067 * be happening. 1068 */ 1069 /* 1070 * XXX we should be doing a dbuf_read, checking the return 1071 * value and returning that up to our callers 1072 */ 1073 dmu_buf_will_dirty(&db->db, tx); 1074 1075 /* create the data buffer for the new block */ 1076 buf = arc_buf_alloc(dn->dn_objset->os_spa, size, db, type); 1077 1078 /* copy old block data to the new block */ 1079 obuf = db->db_buf; 1080 bcopy(obuf->b_data, buf->b_data, MIN(osize, size)); 1081 /* zero the remainder */ 1082 if (size > osize) 1083 bzero((uint8_t *)buf->b_data + osize, size - osize); 1084 1085 mutex_enter(&db->db_mtx); 1086 dbuf_set_data(db, buf); 1087 VERIFY(arc_buf_remove_ref(obuf, db)); 1088 db->db.db_size = size; 1089 1090 if (db->db_level == 0) { 1091 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); 1092 db->db_last_dirty->dt.dl.dr_data = buf; 1093 } 1094 mutex_exit(&db->db_mtx); 1095 1096 dnode_willuse_space(dn, size-osize, tx); 1097 DB_DNODE_EXIT(db); 1098} 1099 1100void 1101dbuf_release_bp(dmu_buf_impl_t *db) 1102{ 1103 objset_t *os = db->db_objset; 1104 1105 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os))); 1106 ASSERT(arc_released(os->os_phys_buf) || 1107 list_link_active(&os->os_dsl_dataset->ds_synced_link)); 1108 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf)); 1109 1110 (void) arc_release(db->db_buf, db); 1111} 1112 1113dbuf_dirty_record_t * 1114dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1115{ 1116 dnode_t *dn; 1117 objset_t *os; 1118 dbuf_dirty_record_t **drp, *dr; 1119 int drop_struct_lock = FALSE; 1120 boolean_t do_free_accounting = B_FALSE; 1121 int txgoff = tx->tx_txg & TXG_MASK; 1122 1123 ASSERT(tx->tx_txg != 0); 1124 ASSERT(!refcount_is_zero(&db->db_holds)); 1125 DMU_TX_DIRTY_BUF(tx, db); 1126 1127 DB_DNODE_ENTER(db); 1128 dn = DB_DNODE(db); 1129 /* 1130 * Shouldn't dirty a regular buffer in syncing context. Private 1131 * objects may be dirtied in syncing context, but only if they 1132 * were already pre-dirtied in open context. 1133 */ 1134 ASSERT(!dmu_tx_is_syncing(tx) || 1135 BP_IS_HOLE(dn->dn_objset->os_rootbp) || 1136 DMU_OBJECT_IS_SPECIAL(dn->dn_object) || 1137 dn->dn_objset->os_dsl_dataset == NULL); 1138 /* 1139 * We make this assert for private objects as well, but after we 1140 * check if we're already dirty. They are allowed to re-dirty 1141 * in syncing context. 1142 */ 1143 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 1144 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 1145 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 1146 1147 mutex_enter(&db->db_mtx); 1148 /* 1149 * XXX make this true for indirects too? The problem is that 1150 * transactions created with dmu_tx_create_assigned() from 1151 * syncing context don't bother holding ahead. 1152 */ 1153 ASSERT(db->db_level != 0 || 1154 db->db_state == DB_CACHED || db->db_state == DB_FILL || 1155 db->db_state == DB_NOFILL); 1156 1157 mutex_enter(&dn->dn_mtx); 1158 /* 1159 * Don't set dirtyctx to SYNC if we're just modifying this as we 1160 * initialize the objset. 1161 */ 1162 if (dn->dn_dirtyctx == DN_UNDIRTIED && 1163 !BP_IS_HOLE(dn->dn_objset->os_rootbp)) { 1164 dn->dn_dirtyctx = 1165 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN); 1166 ASSERT(dn->dn_dirtyctx_firstset == NULL); 1167 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP); 1168 } 1169 mutex_exit(&dn->dn_mtx); 1170 1171 if (db->db_blkid == DMU_SPILL_BLKID) 1172 dn->dn_have_spill = B_TRUE; 1173 1174 /* 1175 * If this buffer is already dirty, we're done. 1176 */ 1177 drp = &db->db_last_dirty; 1178 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg || 1179 db->db.db_object == DMU_META_DNODE_OBJECT); 1180 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg) 1181 drp = &dr->dr_next; 1182 if (dr && dr->dr_txg == tx->tx_txg) { 1183 DB_DNODE_EXIT(db); 1184 1185 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) { 1186 /* 1187 * If this buffer has already been written out, 1188 * we now need to reset its state. 1189 */ 1190 dbuf_unoverride(dr); 1191 if (db->db.db_object != DMU_META_DNODE_OBJECT && 1192 db->db_state != DB_NOFILL) 1193 arc_buf_thaw(db->db_buf); 1194 } 1195 mutex_exit(&db->db_mtx); 1196 return (dr); 1197 } 1198 1199 /* 1200 * Only valid if not already dirty. 1201 */ 1202 ASSERT(dn->dn_object == 0 || 1203 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 1204 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 1205 1206 ASSERT3U(dn->dn_nlevels, >, db->db_level); 1207 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) || 1208 dn->dn_phys->dn_nlevels > db->db_level || 1209 dn->dn_next_nlevels[txgoff] > db->db_level || 1210 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level || 1211 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level); 1212 1213 /* 1214 * We should only be dirtying in syncing context if it's the 1215 * mos or we're initializing the os or it's a special object. 1216 * However, we are allowed to dirty in syncing context provided 1217 * we already dirtied it in open context. Hence we must make 1218 * this assertion only if we're not already dirty. 1219 */ 1220 os = dn->dn_objset; 1221 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) || 1222 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp)); 1223 ASSERT(db->db.db_size != 0); 1224 1225 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 1226 1227 if (db->db_blkid != DMU_BONUS_BLKID) { 1228 /* 1229 * Update the accounting. 1230 * Note: we delay "free accounting" until after we drop 1231 * the db_mtx. This keeps us from grabbing other locks 1232 * (and possibly deadlocking) in bp_get_dsize() while 1233 * also holding the db_mtx. 1234 */ 1235 dnode_willuse_space(dn, db->db.db_size, tx); 1236 do_free_accounting = dbuf_block_freeable(db); 1237 } 1238 1239 /* 1240 * If this buffer is dirty in an old transaction group we need 1241 * to make a copy of it so that the changes we make in this 1242 * transaction group won't leak out when we sync the older txg. 1243 */ 1244 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP); 1245 if (db->db_level == 0) { 1246 void *data_old = db->db_buf; 1247 1248 if (db->db_state != DB_NOFILL) { 1249 if (db->db_blkid == DMU_BONUS_BLKID) { 1250 dbuf_fix_old_data(db, tx->tx_txg); 1251 data_old = db->db.db_data; 1252 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) { 1253 /* 1254 * Release the data buffer from the cache so 1255 * that we can modify it without impacting 1256 * possible other users of this cached data 1257 * block. Note that indirect blocks and 1258 * private objects are not released until the 1259 * syncing state (since they are only modified 1260 * then). 1261 */ 1262 arc_release(db->db_buf, db); 1263 dbuf_fix_old_data(db, tx->tx_txg); 1264 data_old = db->db_buf; 1265 } 1266 ASSERT(data_old != NULL); 1267 } 1268 dr->dt.dl.dr_data = data_old; 1269 } else { 1270 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL); 1271 list_create(&dr->dt.di.dr_children, 1272 sizeof (dbuf_dirty_record_t), 1273 offsetof(dbuf_dirty_record_t, dr_dirty_node)); 1274 } 1275 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL) 1276 dr->dr_accounted = db->db.db_size; 1277 dr->dr_dbuf = db; 1278 dr->dr_txg = tx->tx_txg; 1279 dr->dr_next = *drp; 1280 *drp = dr; 1281 1282 /* 1283 * We could have been freed_in_flight between the dbuf_noread 1284 * and dbuf_dirty. We win, as though the dbuf_noread() had 1285 * happened after the free. 1286 */ 1287 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 1288 db->db_blkid != DMU_SPILL_BLKID) { 1289 mutex_enter(&dn->dn_mtx); 1290 if (dn->dn_free_ranges[txgoff] != NULL) { 1291 range_tree_clear(dn->dn_free_ranges[txgoff], 1292 db->db_blkid, 1); 1293 } 1294 mutex_exit(&dn->dn_mtx); 1295 db->db_freed_in_flight = FALSE; 1296 } 1297 1298 /* 1299 * This buffer is now part of this txg 1300 */ 1301 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg); 1302 db->db_dirtycnt += 1; 1303 ASSERT3U(db->db_dirtycnt, <=, 3); 1304 1305 mutex_exit(&db->db_mtx); 1306 1307 if (db->db_blkid == DMU_BONUS_BLKID || 1308 db->db_blkid == DMU_SPILL_BLKID) { 1309 mutex_enter(&dn->dn_mtx); 1310 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1311 list_insert_tail(&dn->dn_dirty_records[txgoff], dr); 1312 mutex_exit(&dn->dn_mtx); 1313 dnode_setdirty(dn, tx); 1314 DB_DNODE_EXIT(db); 1315 return (dr); 1316 } else if (do_free_accounting) { 1317 blkptr_t *bp = db->db_blkptr; 1318 int64_t willfree = (bp && !BP_IS_HOLE(bp)) ? 1319 bp_get_dsize(os->os_spa, bp) : db->db.db_size; 1320 /* 1321 * This is only a guess -- if the dbuf is dirty 1322 * in a previous txg, we don't know how much 1323 * space it will use on disk yet. We should 1324 * really have the struct_rwlock to access 1325 * db_blkptr, but since this is just a guess, 1326 * it's OK if we get an odd answer. 1327 */ 1328 ddt_prefetch(os->os_spa, bp); 1329 dnode_willuse_space(dn, -willfree, tx); 1330 } 1331 1332 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) { 1333 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1334 drop_struct_lock = TRUE; 1335 } 1336 1337 if (db->db_level == 0) { 1338 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock); 1339 ASSERT(dn->dn_maxblkid >= db->db_blkid); 1340 } 1341 1342 if (db->db_level+1 < dn->dn_nlevels) { 1343 dmu_buf_impl_t *parent = db->db_parent; 1344 dbuf_dirty_record_t *di; 1345 int parent_held = FALSE; 1346 1347 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) { 1348 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1349 1350 parent = dbuf_hold_level(dn, db->db_level+1, 1351 db->db_blkid >> epbs, FTAG); 1352 ASSERT(parent != NULL); 1353 parent_held = TRUE; 1354 } 1355 if (drop_struct_lock) 1356 rw_exit(&dn->dn_struct_rwlock); 1357 ASSERT3U(db->db_level+1, ==, parent->db_level); 1358 di = dbuf_dirty(parent, tx); 1359 if (parent_held) 1360 dbuf_rele(parent, FTAG); 1361 1362 mutex_enter(&db->db_mtx); 1363 /* 1364 * Since we've dropped the mutex, it's possible that 1365 * dbuf_undirty() might have changed this out from under us. 1366 */ 1367 if (db->db_last_dirty == dr || 1368 dn->dn_object == DMU_META_DNODE_OBJECT) { 1369 mutex_enter(&di->dt.di.dr_mtx); 1370 ASSERT3U(di->dr_txg, ==, tx->tx_txg); 1371 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1372 list_insert_tail(&di->dt.di.dr_children, dr); 1373 mutex_exit(&di->dt.di.dr_mtx); 1374 dr->dr_parent = di; 1375 } 1376 mutex_exit(&db->db_mtx); 1377 } else { 1378 ASSERT(db->db_level+1 == dn->dn_nlevels); 1379 ASSERT(db->db_blkid < dn->dn_nblkptr); 1380 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf); 1381 mutex_enter(&dn->dn_mtx); 1382 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1383 list_insert_tail(&dn->dn_dirty_records[txgoff], dr); 1384 mutex_exit(&dn->dn_mtx); 1385 if (drop_struct_lock) 1386 rw_exit(&dn->dn_struct_rwlock); 1387 } 1388 1389 dnode_setdirty(dn, tx); 1390 DB_DNODE_EXIT(db); 1391 return (dr); 1392} 1393 1394/* 1395 * Undirty a buffer in the transaction group referenced by the given 1396 * transaction. Return whether this evicted the dbuf. 1397 */ 1398static boolean_t 1399dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1400{ 1401 dnode_t *dn; 1402 uint64_t txg = tx->tx_txg; 1403 dbuf_dirty_record_t *dr, **drp; 1404 1405 ASSERT(txg != 0); 1406 1407 /* 1408 * Due to our use of dn_nlevels below, this can only be called 1409 * in open context, unless we are operating on the MOS. 1410 * From syncing context, dn_nlevels may be different from the 1411 * dn_nlevels used when dbuf was dirtied. 1412 */ 1413 ASSERT(db->db_objset == 1414 dmu_objset_pool(db->db_objset)->dp_meta_objset || 1415 txg != spa_syncing_txg(dmu_objset_spa(db->db_objset))); 1416 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1417 ASSERT0(db->db_level); 1418 ASSERT(MUTEX_HELD(&db->db_mtx)); 1419 1420 /* 1421 * If this buffer is not dirty, we're done. 1422 */ 1423 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next) 1424 if (dr->dr_txg <= txg) 1425 break; 1426 if (dr == NULL || dr->dr_txg < txg) 1427 return (B_FALSE); 1428 ASSERT(dr->dr_txg == txg); 1429 ASSERT(dr->dr_dbuf == db); 1430 1431 DB_DNODE_ENTER(db); 1432 dn = DB_DNODE(db); 1433 1434 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 1435 1436 ASSERT(db->db.db_size != 0); 1437 1438 dsl_pool_undirty_space(dmu_objset_pool(dn->dn_objset), 1439 dr->dr_accounted, txg); 1440 1441 *drp = dr->dr_next; 1442 1443 /* 1444 * Note that there are three places in dbuf_dirty() 1445 * where this dirty record may be put on a list. 1446 * Make sure to do a list_remove corresponding to 1447 * every one of those list_insert calls. 1448 */ 1449 if (dr->dr_parent) { 1450 mutex_enter(&dr->dr_parent->dt.di.dr_mtx); 1451 list_remove(&dr->dr_parent->dt.di.dr_children, dr); 1452 mutex_exit(&dr->dr_parent->dt.di.dr_mtx); 1453 } else if (db->db_blkid == DMU_SPILL_BLKID || 1454 db->db_level + 1 == dn->dn_nlevels) { 1455 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf); 1456 mutex_enter(&dn->dn_mtx); 1457 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr); 1458 mutex_exit(&dn->dn_mtx); 1459 } 1460 DB_DNODE_EXIT(db); 1461 1462 if (db->db_state != DB_NOFILL) { 1463 dbuf_unoverride(dr); 1464 1465 ASSERT(db->db_buf != NULL); 1466 ASSERT(dr->dt.dl.dr_data != NULL); 1467 if (dr->dt.dl.dr_data != db->db_buf) 1468 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db)); 1469 } 1470 1471 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 1472 1473 ASSERT(db->db_dirtycnt > 0); 1474 db->db_dirtycnt -= 1; 1475 1476 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) { 1477 arc_buf_t *buf = db->db_buf; 1478 1479 ASSERT(db->db_state == DB_NOFILL || arc_released(buf)); 1480 dbuf_clear_data(db); 1481 VERIFY(arc_buf_remove_ref(buf, db)); 1482 dbuf_evict(db); 1483 return (B_TRUE); 1484 } 1485 1486 return (B_FALSE); 1487} 1488 1489void 1490dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx) 1491{ 1492 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1493 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH; 1494 1495 ASSERT(tx->tx_txg != 0); 1496 ASSERT(!refcount_is_zero(&db->db_holds)); 1497 1498 DB_DNODE_ENTER(db); 1499 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock)) 1500 rf |= DB_RF_HAVESTRUCT; 1501 DB_DNODE_EXIT(db); 1502 (void) dbuf_read(db, NULL, rf); 1503 (void) dbuf_dirty(db, tx); 1504} 1505 1506void 1507dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) 1508{ 1509 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1510 1511 db->db_state = DB_NOFILL; 1512 1513 dmu_buf_will_fill(db_fake, tx); 1514} 1515 1516void 1517dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) 1518{ 1519 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1520 1521 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1522 ASSERT(tx->tx_txg != 0); 1523 ASSERT(db->db_level == 0); 1524 ASSERT(!refcount_is_zero(&db->db_holds)); 1525 1526 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT || 1527 dmu_tx_private_ok(tx)); 1528 1529 dbuf_noread(db); 1530 (void) dbuf_dirty(db, tx); 1531} 1532 1533#pragma weak dmu_buf_fill_done = dbuf_fill_done 1534/* ARGSUSED */ 1535void 1536dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx) 1537{ 1538 mutex_enter(&db->db_mtx); 1539 DBUF_VERIFY(db); 1540 1541 if (db->db_state == DB_FILL) { 1542 if (db->db_level == 0 && db->db_freed_in_flight) { 1543 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1544 /* we were freed while filling */ 1545 /* XXX dbuf_undirty? */ 1546 bzero(db->db.db_data, db->db.db_size); 1547 db->db_freed_in_flight = FALSE; 1548 } 1549 db->db_state = DB_CACHED; 1550 cv_broadcast(&db->db_changed); 1551 } 1552 mutex_exit(&db->db_mtx); 1553} 1554 1555void 1556dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data, 1557 bp_embedded_type_t etype, enum zio_compress comp, 1558 int uncompressed_size, int compressed_size, int byteorder, 1559 dmu_tx_t *tx) 1560{ 1561 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; 1562 struct dirty_leaf *dl; 1563 dmu_object_type_t type; 1564 1565 if (etype == BP_EMBEDDED_TYPE_DATA) { 1566 ASSERT(spa_feature_is_active(dmu_objset_spa(db->db_objset), 1567 SPA_FEATURE_EMBEDDED_DATA)); 1568 } 1569 1570 DB_DNODE_ENTER(db); 1571 type = DB_DNODE(db)->dn_type; 1572 DB_DNODE_EXIT(db); 1573 1574 ASSERT0(db->db_level); 1575 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1576 1577 dmu_buf_will_not_fill(dbuf, tx); 1578 1579 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); 1580 dl = &db->db_last_dirty->dt.dl; 1581 encode_embedded_bp_compressed(&dl->dr_overridden_by, 1582 data, comp, uncompressed_size, compressed_size); 1583 BPE_SET_ETYPE(&dl->dr_overridden_by, etype); 1584 BP_SET_TYPE(&dl->dr_overridden_by, type); 1585 BP_SET_LEVEL(&dl->dr_overridden_by, 0); 1586 BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder); 1587 1588 dl->dr_override_state = DR_OVERRIDDEN; 1589 dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg; 1590} 1591 1592/* 1593 * Directly assign a provided arc buf to a given dbuf if it's not referenced 1594 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf. 1595 */ 1596void 1597dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx) 1598{ 1599 ASSERT(!refcount_is_zero(&db->db_holds)); 1600 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1601 ASSERT(db->db_level == 0); 1602 ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA); 1603 ASSERT(buf != NULL); 1604 ASSERT(arc_buf_size(buf) == db->db.db_size); 1605 ASSERT(tx->tx_txg != 0); 1606 1607 arc_return_buf(buf, db); 1608 ASSERT(arc_released(buf)); 1609 1610 mutex_enter(&db->db_mtx); 1611 1612 while (db->db_state == DB_READ || db->db_state == DB_FILL) 1613 cv_wait(&db->db_changed, &db->db_mtx); 1614 1615 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED); 1616 1617 if (db->db_state == DB_CACHED && 1618 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) { 1619 mutex_exit(&db->db_mtx); 1620 (void) dbuf_dirty(db, tx); 1621 bcopy(buf->b_data, db->db.db_data, db->db.db_size); 1622 VERIFY(arc_buf_remove_ref(buf, db)); 1623 xuio_stat_wbuf_copied(); 1624 return; 1625 } 1626 1627 xuio_stat_wbuf_nocopy(); 1628 if (db->db_state == DB_CACHED) { 1629 dbuf_dirty_record_t *dr = db->db_last_dirty; 1630 1631 ASSERT(db->db_buf != NULL); 1632 if (dr != NULL && dr->dr_txg == tx->tx_txg) { 1633 ASSERT(dr->dt.dl.dr_data == db->db_buf); 1634 if (!arc_released(db->db_buf)) { 1635 ASSERT(dr->dt.dl.dr_override_state == 1636 DR_OVERRIDDEN); 1637 arc_release(db->db_buf, db); 1638 } 1639 dr->dt.dl.dr_data = buf; 1640 VERIFY(arc_buf_remove_ref(db->db_buf, db)); 1641 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) { 1642 arc_release(db->db_buf, db); 1643 VERIFY(arc_buf_remove_ref(db->db_buf, db)); 1644 } 1645 db->db_buf = NULL; 1646 } 1647 ASSERT(db->db_buf == NULL); 1648 dbuf_set_data(db, buf); 1649 db->db_state = DB_FILL; 1650 mutex_exit(&db->db_mtx); 1651 (void) dbuf_dirty(db, tx); 1652 dmu_buf_fill_done(&db->db, tx); 1653} 1654 1655/* 1656 * "Clear" the contents of this dbuf. This will mark the dbuf 1657 * EVICTING and clear *most* of its references. Unfortunately, 1658 * when we are not holding the dn_dbufs_mtx, we can't clear the 1659 * entry in the dn_dbufs list. We have to wait until dbuf_destroy() 1660 * in this case. For callers from the DMU we will usually see: 1661 * dbuf_clear()->arc_clear_callback()->dbuf_do_evict()->dbuf_destroy() 1662 * For the arc callback, we will usually see: 1663 * dbuf_do_evict()->dbuf_clear();dbuf_destroy() 1664 * Sometimes, though, we will get a mix of these two: 1665 * DMU: dbuf_clear()->arc_clear_callback() 1666 * ARC: dbuf_do_evict()->dbuf_destroy() 1667 * 1668 * This routine will dissociate the dbuf from the arc, by calling 1669 * arc_clear_callback(), but will not evict the data from the ARC. 1670 */ 1671void 1672dbuf_clear(dmu_buf_impl_t *db) 1673{ 1674 dnode_t *dn; 1675 dmu_buf_impl_t *parent = db->db_parent; 1676 dmu_buf_impl_t *dndb; 1677 boolean_t dbuf_gone = B_FALSE; 1678 1679 ASSERT(MUTEX_HELD(&db->db_mtx)); 1680 ASSERT(refcount_is_zero(&db->db_holds)); 1681 1682 dbuf_evict_user(db); 1683 1684 if (db->db_state == DB_CACHED) { 1685 ASSERT(db->db.db_data != NULL); 1686 if (db->db_blkid == DMU_BONUS_BLKID) { 1687 zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN); 1688 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 1689 } 1690 db->db.db_data = NULL; 1691 db->db_state = DB_UNCACHED; 1692 } 1693 1694 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL); 1695 ASSERT(db->db_data_pending == NULL); 1696 1697 db->db_state = DB_EVICTING; 1698 db->db_blkptr = NULL; 1699 1700 DB_DNODE_ENTER(db); 1701 dn = DB_DNODE(db); 1702 dndb = dn->dn_dbuf; 1703 if (db->db_blkid != DMU_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) { 1704 avl_remove(&dn->dn_dbufs, db); 1705 atomic_dec_32(&dn->dn_dbufs_count); 1706 membar_producer(); 1707 DB_DNODE_EXIT(db); 1708 /* 1709 * Decrementing the dbuf count means that the hold corresponding 1710 * to the removed dbuf is no longer discounted in dnode_move(), 1711 * so the dnode cannot be moved until after we release the hold. 1712 * The membar_producer() ensures visibility of the decremented 1713 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually 1714 * release any lock. 1715 */ 1716 dnode_rele(dn, db); 1717 db->db_dnode_handle = NULL; 1718 } else { 1719 DB_DNODE_EXIT(db); 1720 } 1721 1722 if (db->db_buf) 1723 dbuf_gone = arc_clear_callback(db->db_buf); 1724 1725 if (!dbuf_gone) 1726 mutex_exit(&db->db_mtx); 1727 1728 /* 1729 * If this dbuf is referenced from an indirect dbuf, 1730 * decrement the ref count on the indirect dbuf. 1731 */ 1732 if (parent && parent != dndb) 1733 dbuf_rele(parent, db); 1734} 1735 1736/* 1737 * Note: While bpp will always be updated if the function returns success, 1738 * parentp will not be updated if the dnode does not have dn_dbuf filled in; 1739 * this happens when the dnode is the meta-dnode, or a userused or groupused 1740 * object. 1741 */ 1742static int 1743dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse, 1744 dmu_buf_impl_t **parentp, blkptr_t **bpp) 1745{ 1746 int nlevels, epbs; 1747 1748 *parentp = NULL; 1749 *bpp = NULL; 1750 1751 ASSERT(blkid != DMU_BONUS_BLKID); 1752 1753 if (blkid == DMU_SPILL_BLKID) { 1754 mutex_enter(&dn->dn_mtx); 1755 if (dn->dn_have_spill && 1756 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) 1757 *bpp = &dn->dn_phys->dn_spill; 1758 else 1759 *bpp = NULL; 1760 dbuf_add_ref(dn->dn_dbuf, NULL); 1761 *parentp = dn->dn_dbuf; 1762 mutex_exit(&dn->dn_mtx); 1763 return (0); 1764 } 1765 1766 if (dn->dn_phys->dn_nlevels == 0) 1767 nlevels = 1; 1768 else 1769 nlevels = dn->dn_phys->dn_nlevels; 1770 1771 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1772 1773 ASSERT3U(level * epbs, <, 64); 1774 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1775 if (level >= nlevels || 1776 (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) { 1777 /* the buffer has no parent yet */ 1778 return (SET_ERROR(ENOENT)); 1779 } else if (level < nlevels-1) { 1780 /* this block is referenced from an indirect block */ 1781 int err = dbuf_hold_impl(dn, level+1, 1782 blkid >> epbs, fail_sparse, FALSE, NULL, parentp); 1783 if (err) 1784 return (err); 1785 err = dbuf_read(*parentp, NULL, 1786 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL)); 1787 if (err) { 1788 dbuf_rele(*parentp, NULL); 1789 *parentp = NULL; 1790 return (err); 1791 } 1792 *bpp = ((blkptr_t *)(*parentp)->db.db_data) + 1793 (blkid & ((1ULL << epbs) - 1)); 1794 return (0); 1795 } else { 1796 /* the block is referenced from the dnode */ 1797 ASSERT3U(level, ==, nlevels-1); 1798 ASSERT(dn->dn_phys->dn_nblkptr == 0 || 1799 blkid < dn->dn_phys->dn_nblkptr); 1800 if (dn->dn_dbuf) { 1801 dbuf_add_ref(dn->dn_dbuf, NULL); 1802 *parentp = dn->dn_dbuf; 1803 } 1804 *bpp = &dn->dn_phys->dn_blkptr[blkid]; 1805 return (0); 1806 } 1807} 1808 1809static dmu_buf_impl_t * 1810dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid, 1811 dmu_buf_impl_t *parent, blkptr_t *blkptr) 1812{ 1813 objset_t *os = dn->dn_objset; 1814 dmu_buf_impl_t *db, *odb; 1815 1816 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1817 ASSERT(dn->dn_type != DMU_OT_NONE); 1818 1819 db = kmem_cache_alloc(dbuf_cache, KM_SLEEP); 1820 1821 db->db_objset = os; 1822 db->db.db_object = dn->dn_object; 1823 db->db_level = level; 1824 db->db_blkid = blkid; 1825 db->db_last_dirty = NULL; 1826 db->db_dirtycnt = 0; 1827 db->db_dnode_handle = dn->dn_handle; 1828 db->db_parent = parent; 1829 db->db_blkptr = blkptr; 1830 1831 db->db_user = NULL; 1832 db->db_immediate_evict = 0; 1833 db->db_freed_in_flight = 0; 1834 1835 if (blkid == DMU_BONUS_BLKID) { 1836 ASSERT3P(parent, ==, dn->dn_dbuf); 1837 db->db.db_size = DN_MAX_BONUSLEN - 1838 (dn->dn_nblkptr-1) * sizeof (blkptr_t); 1839 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 1840 db->db.db_offset = DMU_BONUS_BLKID; 1841 db->db_state = DB_UNCACHED; 1842 /* the bonus dbuf is not placed in the hash table */ 1843 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 1844 return (db); 1845 } else if (blkid == DMU_SPILL_BLKID) { 1846 db->db.db_size = (blkptr != NULL) ? 1847 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE; 1848 db->db.db_offset = 0; 1849 } else { 1850 int blocksize = 1851 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz; 1852 db->db.db_size = blocksize; 1853 db->db.db_offset = db->db_blkid * blocksize; 1854 } 1855 1856 /* 1857 * Hold the dn_dbufs_mtx while we get the new dbuf 1858 * in the hash table *and* added to the dbufs list. 1859 * This prevents a possible deadlock with someone 1860 * trying to look up this dbuf before its added to the 1861 * dn_dbufs list. 1862 */ 1863 mutex_enter(&dn->dn_dbufs_mtx); 1864 db->db_state = DB_EVICTING; 1865 if ((odb = dbuf_hash_insert(db)) != NULL) { 1866 /* someone else inserted it first */ 1867 kmem_cache_free(dbuf_cache, db); 1868 mutex_exit(&dn->dn_dbufs_mtx); 1869 return (odb); 1870 } 1871 avl_add(&dn->dn_dbufs, db); 1872 if (db->db_level == 0 && db->db_blkid >= 1873 dn->dn_unlisted_l0_blkid) 1874 dn->dn_unlisted_l0_blkid = db->db_blkid + 1; 1875 db->db_state = DB_UNCACHED; 1876 mutex_exit(&dn->dn_dbufs_mtx); 1877 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 1878 1879 if (parent && parent != dn->dn_dbuf) 1880 dbuf_add_ref(parent, db); 1881 1882 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 1883 refcount_count(&dn->dn_holds) > 0); 1884 (void) refcount_add(&dn->dn_holds, db); 1885 atomic_inc_32(&dn->dn_dbufs_count); 1886 1887 dprintf_dbuf(db, "db=%p\n", db); 1888 1889 return (db); 1890} 1891 1892static int 1893dbuf_do_evict(void *private) 1894{ 1895 dmu_buf_impl_t *db = private; 1896 1897 if (!MUTEX_HELD(&db->db_mtx)) 1898 mutex_enter(&db->db_mtx); 1899 1900 ASSERT(refcount_is_zero(&db->db_holds)); 1901 1902 if (db->db_state != DB_EVICTING) { 1903 ASSERT(db->db_state == DB_CACHED); 1904 DBUF_VERIFY(db); 1905 db->db_buf = NULL; 1906 dbuf_evict(db); 1907 } else { 1908 mutex_exit(&db->db_mtx); 1909 dbuf_destroy(db); 1910 } 1911 return (0); 1912} 1913 1914static void 1915dbuf_destroy(dmu_buf_impl_t *db) 1916{ 1917 ASSERT(refcount_is_zero(&db->db_holds)); 1918 1919 if (db->db_blkid != DMU_BONUS_BLKID) { 1920 /* 1921 * If this dbuf is still on the dn_dbufs list, 1922 * remove it from that list. 1923 */ 1924 if (db->db_dnode_handle != NULL) { 1925 dnode_t *dn; 1926 1927 DB_DNODE_ENTER(db); 1928 dn = DB_DNODE(db); 1929 mutex_enter(&dn->dn_dbufs_mtx); 1930 avl_remove(&dn->dn_dbufs, db); 1931 atomic_dec_32(&dn->dn_dbufs_count); 1932 mutex_exit(&dn->dn_dbufs_mtx); 1933 DB_DNODE_EXIT(db); 1934 /* 1935 * Decrementing the dbuf count means that the hold 1936 * corresponding to the removed dbuf is no longer 1937 * discounted in dnode_move(), so the dnode cannot be 1938 * moved until after we release the hold. 1939 */ 1940 dnode_rele(dn, db); 1941 db->db_dnode_handle = NULL; 1942 } 1943 dbuf_hash_remove(db); 1944 } 1945 db->db_parent = NULL; 1946 db->db_buf = NULL; 1947 1948 ASSERT(db->db.db_data == NULL); 1949 ASSERT(db->db_hash_next == NULL); 1950 ASSERT(db->db_blkptr == NULL); 1951 ASSERT(db->db_data_pending == NULL); 1952 1953 kmem_cache_free(dbuf_cache, db); 1954 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 1955} 1956 1957typedef struct dbuf_prefetch_arg { 1958 spa_t *dpa_spa; /* The spa to issue the prefetch in. */ 1959 zbookmark_phys_t dpa_zb; /* The target block to prefetch. */ 1960 int dpa_epbs; /* Entries (blkptr_t's) Per Block Shift. */ 1961 int dpa_curlevel; /* The current level that we're reading */ 1962 zio_priority_t dpa_prio; /* The priority I/Os should be issued at. */ 1963 zio_t *dpa_zio; /* The parent zio_t for all prefetches. */ 1964 arc_flags_t dpa_aflags; /* Flags to pass to the final prefetch. */ 1965} dbuf_prefetch_arg_t; 1966 1967/* 1968 * Actually issue the prefetch read for the block given. 1969 */ 1970static void 1971dbuf_issue_final_prefetch(dbuf_prefetch_arg_t *dpa, blkptr_t *bp) 1972{ 1973 if (BP_IS_HOLE(bp) || BP_IS_EMBEDDED(bp)) 1974 return; 1975 1976 arc_flags_t aflags = 1977 dpa->dpa_aflags | ARC_FLAG_NOWAIT | ARC_FLAG_PREFETCH; 1978 1979 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp)); 1980 ASSERT3U(dpa->dpa_curlevel, ==, dpa->dpa_zb.zb_level); 1981 ASSERT(dpa->dpa_zio != NULL); 1982 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, bp, NULL, NULL, 1983 dpa->dpa_prio, ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 1984 &aflags, &dpa->dpa_zb); 1985} 1986 1987/* 1988 * Called when an indirect block above our prefetch target is read in. This 1989 * will either read in the next indirect block down the tree or issue the actual 1990 * prefetch if the next block down is our target. 1991 */ 1992static void 1993dbuf_prefetch_indirect_done(zio_t *zio, arc_buf_t *abuf, void *private) 1994{ 1995 dbuf_prefetch_arg_t *dpa = private; 1996 1997 ASSERT3S(dpa->dpa_zb.zb_level, <, dpa->dpa_curlevel); 1998 ASSERT3S(dpa->dpa_curlevel, >, 0); 1999 if (zio != NULL) { 2000 ASSERT3S(BP_GET_LEVEL(zio->io_bp), ==, dpa->dpa_curlevel); 2001 ASSERT3U(BP_GET_LSIZE(zio->io_bp), ==, zio->io_size); 2002 ASSERT3P(zio->io_spa, ==, dpa->dpa_spa); 2003 } 2004 2005 dpa->dpa_curlevel--; 2006 2007 uint64_t nextblkid = dpa->dpa_zb.zb_blkid >> 2008 (dpa->dpa_epbs * (dpa->dpa_curlevel - dpa->dpa_zb.zb_level)); 2009 blkptr_t *bp = ((blkptr_t *)abuf->b_data) + 2010 P2PHASE(nextblkid, 1ULL << dpa->dpa_epbs); 2011 if (BP_IS_HOLE(bp) || (zio != NULL && zio->io_error != 0)) { 2012 kmem_free(dpa, sizeof (*dpa)); 2013 } else if (dpa->dpa_curlevel == dpa->dpa_zb.zb_level) { 2014 ASSERT3U(nextblkid, ==, dpa->dpa_zb.zb_blkid); 2015 dbuf_issue_final_prefetch(dpa, bp); 2016 kmem_free(dpa, sizeof (*dpa)); 2017 } else { 2018 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT; 2019 zbookmark_phys_t zb; 2020 2021 ASSERT3U(dpa->dpa_curlevel, ==, BP_GET_LEVEL(bp)); 2022 2023 SET_BOOKMARK(&zb, dpa->dpa_zb.zb_objset, 2024 dpa->dpa_zb.zb_object, dpa->dpa_curlevel, nextblkid); 2025 2026 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, 2027 bp, dbuf_prefetch_indirect_done, dpa, dpa->dpa_prio, 2028 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2029 &iter_aflags, &zb); 2030 } 2031 (void) arc_buf_remove_ref(abuf, private); 2032} 2033 2034/* 2035 * Issue prefetch reads for the given block on the given level. If the indirect 2036 * blocks above that block are not in memory, we will read them in 2037 * asynchronously. As a result, this call never blocks waiting for a read to 2038 * complete. 2039 */ 2040void 2041dbuf_prefetch(dnode_t *dn, int64_t level, uint64_t blkid, zio_priority_t prio, 2042 arc_flags_t aflags) 2043{ 2044 blkptr_t bp; 2045 int epbs, nlevels, curlevel; 2046 uint64_t curblkid; 2047 2048 ASSERT(blkid != DMU_BONUS_BLKID); 2049 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 2050 2051 if (blkid > dn->dn_maxblkid) 2052 return; 2053 2054 if (dnode_block_freed(dn, blkid)) 2055 return; 2056 2057 /* 2058 * This dnode hasn't been written to disk yet, so there's nothing to 2059 * prefetch. 2060 */ 2061 nlevels = dn->dn_phys->dn_nlevels; 2062 if (level >= nlevels || dn->dn_phys->dn_nblkptr == 0) 2063 return; 2064 2065 epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 2066 if (dn->dn_phys->dn_maxblkid < blkid << (epbs * level)) 2067 return; 2068 2069 dmu_buf_impl_t *db = dbuf_find(dn->dn_objset, dn->dn_object, 2070 level, blkid); 2071 if (db != NULL) { 2072 mutex_exit(&db->db_mtx); 2073 /* 2074 * This dbuf already exists. It is either CACHED, or 2075 * (we assume) about to be read or filled. 2076 */ 2077 return; 2078 } 2079 2080 /* 2081 * Find the closest ancestor (indirect block) of the target block 2082 * that is present in the cache. In this indirect block, we will 2083 * find the bp that is at curlevel, curblkid. 2084 */ 2085 curlevel = level; 2086 curblkid = blkid; 2087 while (curlevel < nlevels - 1) { 2088 int parent_level = curlevel + 1; 2089 uint64_t parent_blkid = curblkid >> epbs; 2090 dmu_buf_impl_t *db; 2091 2092 if (dbuf_hold_impl(dn, parent_level, parent_blkid, 2093 FALSE, TRUE, FTAG, &db) == 0) { 2094 blkptr_t *bpp = db->db_buf->b_data; 2095 bp = bpp[P2PHASE(curblkid, 1 << epbs)]; 2096 dbuf_rele(db, FTAG); 2097 break; 2098 } 2099 2100 curlevel = parent_level; 2101 curblkid = parent_blkid; 2102 } 2103 2104 if (curlevel == nlevels - 1) { 2105 /* No cached indirect blocks found. */ 2106 ASSERT3U(curblkid, <, dn->dn_phys->dn_nblkptr); 2107 bp = dn->dn_phys->dn_blkptr[curblkid]; 2108 } 2109 if (BP_IS_HOLE(&bp)) 2110 return; 2111 2112 ASSERT3U(curlevel, ==, BP_GET_LEVEL(&bp)); 2113 2114 zio_t *pio = zio_root(dmu_objset_spa(dn->dn_objset), NULL, NULL, 2115 ZIO_FLAG_CANFAIL); 2116 2117 dbuf_prefetch_arg_t *dpa = kmem_zalloc(sizeof (*dpa), KM_SLEEP); 2118 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset; 2119 SET_BOOKMARK(&dpa->dpa_zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET, 2120 dn->dn_object, level, blkid); 2121 dpa->dpa_curlevel = curlevel; 2122 dpa->dpa_prio = prio; 2123 dpa->dpa_aflags = aflags; 2124 dpa->dpa_spa = dn->dn_objset->os_spa; 2125 dpa->dpa_epbs = epbs; 2126 dpa->dpa_zio = pio; 2127 2128 /* 2129 * If we have the indirect just above us, no need to do the asynchronous 2130 * prefetch chain; we'll just run the last step ourselves. If we're at 2131 * a higher level, though, we want to issue the prefetches for all the 2132 * indirect blocks asynchronously, so we can go on with whatever we were 2133 * doing. 2134 */ 2135 if (curlevel == level) { 2136 ASSERT3U(curblkid, ==, blkid); 2137 dbuf_issue_final_prefetch(dpa, &bp); 2138 kmem_free(dpa, sizeof (*dpa)); 2139 } else { 2140 arc_flags_t iter_aflags = ARC_FLAG_NOWAIT; 2141 zbookmark_phys_t zb; 2142 2143 SET_BOOKMARK(&zb, ds != NULL ? ds->ds_object : DMU_META_OBJSET, 2144 dn->dn_object, curlevel, curblkid); 2145 (void) arc_read(dpa->dpa_zio, dpa->dpa_spa, 2146 &bp, dbuf_prefetch_indirect_done, dpa, prio, 2147 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 2148 &iter_aflags, &zb); 2149 } 2150 /* 2151 * We use pio here instead of dpa_zio since it's possible that 2152 * dpa may have already been freed. 2153 */ 2154 zio_nowait(pio); 2155} 2156 2157/* 2158 * Returns with db_holds incremented, and db_mtx not held. 2159 * Note: dn_struct_rwlock must be held. 2160 */ 2161int 2162dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, 2163 boolean_t fail_sparse, boolean_t fail_uncached, 2164 void *tag, dmu_buf_impl_t **dbp) 2165{ 2166 dmu_buf_impl_t *db, *parent = NULL; 2167 2168 ASSERT(blkid != DMU_BONUS_BLKID); 2169 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 2170 ASSERT3U(dn->dn_nlevels, >, level); 2171 2172 *dbp = NULL; 2173top: 2174 /* dbuf_find() returns with db_mtx held */ 2175 db = dbuf_find(dn->dn_objset, dn->dn_object, level, blkid); 2176 2177 if (db == NULL) { 2178 blkptr_t *bp = NULL; 2179 int err; 2180 2181 if (fail_uncached) 2182 return (SET_ERROR(ENOENT)); 2183 2184 ASSERT3P(parent, ==, NULL); 2185 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp); 2186 if (fail_sparse) { 2187 if (err == 0 && bp && BP_IS_HOLE(bp)) 2188 err = SET_ERROR(ENOENT); 2189 if (err) { 2190 if (parent) 2191 dbuf_rele(parent, NULL); 2192 return (err); 2193 } 2194 } 2195 if (err && err != ENOENT) 2196 return (err); 2197 db = dbuf_create(dn, level, blkid, parent, bp); 2198 } 2199 2200 if (fail_uncached && db->db_state != DB_CACHED) { 2201 mutex_exit(&db->db_mtx); 2202 return (SET_ERROR(ENOENT)); 2203 } 2204 2205 if (db->db_buf && refcount_is_zero(&db->db_holds)) { 2206 arc_buf_add_ref(db->db_buf, db); 2207 if (db->db_buf->b_data == NULL) { 2208 dbuf_clear(db); 2209 if (parent) { 2210 dbuf_rele(parent, NULL); 2211 parent = NULL; 2212 } 2213 goto top; 2214 } 2215 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data); 2216 } 2217 2218 ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf)); 2219 2220 /* 2221 * If this buffer is currently syncing out, and we are are 2222 * still referencing it from db_data, we need to make a copy 2223 * of it in case we decide we want to dirty it again in this txg. 2224 */ 2225 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 2226 dn->dn_object != DMU_META_DNODE_OBJECT && 2227 db->db_state == DB_CACHED && db->db_data_pending) { 2228 dbuf_dirty_record_t *dr = db->db_data_pending; 2229 2230 if (dr->dt.dl.dr_data == db->db_buf) { 2231 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 2232 2233 dbuf_set_data(db, 2234 arc_buf_alloc(dn->dn_objset->os_spa, 2235 db->db.db_size, db, type)); 2236 bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data, 2237 db->db.db_size); 2238 } 2239 } 2240 2241 (void) refcount_add(&db->db_holds, tag); 2242 DBUF_VERIFY(db); 2243 mutex_exit(&db->db_mtx); 2244 2245 /* NOTE: we can't rele the parent until after we drop the db_mtx */ 2246 if (parent) 2247 dbuf_rele(parent, NULL); 2248 2249 ASSERT3P(DB_DNODE(db), ==, dn); 2250 ASSERT3U(db->db_blkid, ==, blkid); 2251 ASSERT3U(db->db_level, ==, level); 2252 *dbp = db; 2253 2254 return (0); 2255} 2256 2257dmu_buf_impl_t * 2258dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag) 2259{ 2260 return (dbuf_hold_level(dn, 0, blkid, tag)); 2261} 2262 2263dmu_buf_impl_t * 2264dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag) 2265{ 2266 dmu_buf_impl_t *db; 2267 int err = dbuf_hold_impl(dn, level, blkid, FALSE, FALSE, tag, &db); 2268 return (err ? NULL : db); 2269} 2270 2271void 2272dbuf_create_bonus(dnode_t *dn) 2273{ 2274 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 2275 2276 ASSERT(dn->dn_bonus == NULL); 2277 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL); 2278} 2279 2280int 2281dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx) 2282{ 2283 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2284 dnode_t *dn; 2285 2286 if (db->db_blkid != DMU_SPILL_BLKID) 2287 return (SET_ERROR(ENOTSUP)); 2288 if (blksz == 0) 2289 blksz = SPA_MINBLOCKSIZE; 2290 ASSERT3U(blksz, <=, spa_maxblocksize(dmu_objset_spa(db->db_objset))); 2291 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE); 2292 2293 DB_DNODE_ENTER(db); 2294 dn = DB_DNODE(db); 2295 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 2296 dbuf_new_size(db, blksz, tx); 2297 rw_exit(&dn->dn_struct_rwlock); 2298 DB_DNODE_EXIT(db); 2299 2300 return (0); 2301} 2302 2303void 2304dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx) 2305{ 2306 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx); 2307} 2308 2309#pragma weak dmu_buf_add_ref = dbuf_add_ref 2310void 2311dbuf_add_ref(dmu_buf_impl_t *db, void *tag) 2312{ 2313 int64_t holds = refcount_add(&db->db_holds, tag); 2314 ASSERT(holds > 1); 2315} 2316 2317#pragma weak dmu_buf_try_add_ref = dbuf_try_add_ref 2318boolean_t 2319dbuf_try_add_ref(dmu_buf_t *db_fake, objset_t *os, uint64_t obj, uint64_t blkid, 2320 void *tag) 2321{ 2322 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2323 dmu_buf_impl_t *found_db; 2324 boolean_t result = B_FALSE; 2325 2326 if (db->db_blkid == DMU_BONUS_BLKID) 2327 found_db = dbuf_find_bonus(os, obj); 2328 else 2329 found_db = dbuf_find(os, obj, 0, blkid); 2330 2331 if (found_db != NULL) { 2332 if (db == found_db && dbuf_refcount(db) > db->db_dirtycnt) { 2333 (void) refcount_add(&db->db_holds, tag); 2334 result = B_TRUE; 2335 } 2336 mutex_exit(&db->db_mtx); 2337 } 2338 return (result); 2339} 2340 2341/* 2342 * If you call dbuf_rele() you had better not be referencing the dnode handle 2343 * unless you have some other direct or indirect hold on the dnode. (An indirect 2344 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.) 2345 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the 2346 * dnode's parent dbuf evicting its dnode handles. 2347 */ 2348void 2349dbuf_rele(dmu_buf_impl_t *db, void *tag) 2350{ 2351 mutex_enter(&db->db_mtx); 2352 dbuf_rele_and_unlock(db, tag); 2353} 2354 2355void 2356dmu_buf_rele(dmu_buf_t *db, void *tag) 2357{ 2358 dbuf_rele((dmu_buf_impl_t *)db, tag); 2359} 2360 2361/* 2362 * dbuf_rele() for an already-locked dbuf. This is necessary to allow 2363 * db_dirtycnt and db_holds to be updated atomically. 2364 */ 2365void 2366dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag) 2367{ 2368 int64_t holds; 2369 2370 ASSERT(MUTEX_HELD(&db->db_mtx)); 2371 DBUF_VERIFY(db); 2372 2373 /* 2374 * Remove the reference to the dbuf before removing its hold on the 2375 * dnode so we can guarantee in dnode_move() that a referenced bonus 2376 * buffer has a corresponding dnode hold. 2377 */ 2378 holds = refcount_remove(&db->db_holds, tag); 2379 ASSERT(holds >= 0); 2380 2381 /* 2382 * We can't freeze indirects if there is a possibility that they 2383 * may be modified in the current syncing context. 2384 */ 2385 if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) 2386 arc_buf_freeze(db->db_buf); 2387 2388 if (holds == db->db_dirtycnt && 2389 db->db_level == 0 && db->db_immediate_evict) 2390 dbuf_evict_user(db); 2391 2392 if (holds == 0) { 2393 if (db->db_blkid == DMU_BONUS_BLKID) { 2394 dnode_t *dn; 2395 2396 /* 2397 * If the dnode moves here, we cannot cross this 2398 * barrier until the move completes. 2399 */ 2400 DB_DNODE_ENTER(db); 2401 2402 dn = DB_DNODE(db); 2403 atomic_dec_32(&dn->dn_dbufs_count); 2404 2405 /* 2406 * Decrementing the dbuf count means that the bonus 2407 * buffer's dnode hold is no longer discounted in 2408 * dnode_move(). The dnode cannot move until after 2409 * the dnode_rele_and_unlock() below. 2410 */ 2411 DB_DNODE_EXIT(db); 2412 2413 /* 2414 * Do not reference db after its lock is dropped. 2415 * Another thread may evict it. 2416 */ 2417 mutex_exit(&db->db_mtx); 2418 2419 /* 2420 * If the dnode has been freed, evict the bonus 2421 * buffer immediately. The data in the bonus 2422 * buffer is no longer relevant and this prevents 2423 * a stale bonus buffer from being associated 2424 * with this dnode_t should the dnode_t be reused 2425 * prior to being destroyed. 2426 */ 2427 mutex_enter(&dn->dn_mtx); 2428 if (dn->dn_type == DMU_OT_NONE || 2429 dn->dn_free_txg != 0) { 2430 /* 2431 * Drop dn_mtx. It is a leaf lock and 2432 * cannot be held when dnode_evict_bonus() 2433 * acquires other locks in order to 2434 * perform the eviction. 2435 * 2436 * Freed dnodes cannot be reused until the 2437 * last hold is released. Since this bonus 2438 * buffer has a hold, the dnode will remain 2439 * in the free state, even without dn_mtx 2440 * held, until the dnode_rele_and_unlock() 2441 * below. 2442 */ 2443 mutex_exit(&dn->dn_mtx); 2444 dnode_evict_bonus(dn); 2445 mutex_enter(&dn->dn_mtx); 2446 } 2447 dnode_rele_and_unlock(dn, db); 2448 } else if (db->db_buf == NULL) { 2449 /* 2450 * This is a special case: we never associated this 2451 * dbuf with any data allocated from the ARC. 2452 */ 2453 ASSERT(db->db_state == DB_UNCACHED || 2454 db->db_state == DB_NOFILL); 2455 dbuf_evict(db); 2456 } else if (arc_released(db->db_buf)) { 2457 arc_buf_t *buf = db->db_buf; 2458 /* 2459 * This dbuf has anonymous data associated with it. 2460 */ 2461 dbuf_clear_data(db); 2462 VERIFY(arc_buf_remove_ref(buf, db)); 2463 dbuf_evict(db); 2464 } else { 2465 VERIFY(!arc_buf_remove_ref(db->db_buf, db)); 2466 2467 /* 2468 * A dbuf will be eligible for eviction if either the 2469 * 'primarycache' property is set or a duplicate 2470 * copy of this buffer is already cached in the arc. 2471 * 2472 * In the case of the 'primarycache' a buffer 2473 * is considered for eviction if it matches the 2474 * criteria set in the property. 2475 * 2476 * To decide if our buffer is considered a 2477 * duplicate, we must call into the arc to determine 2478 * if multiple buffers are referencing the same 2479 * block on-disk. If so, then we simply evict 2480 * ourselves. 2481 */ 2482 if (!DBUF_IS_CACHEABLE(db)) { 2483 if (db->db_blkptr != NULL && 2484 !BP_IS_HOLE(db->db_blkptr) && 2485 !BP_IS_EMBEDDED(db->db_blkptr)) { 2486 spa_t *spa = 2487 dmu_objset_spa(db->db_objset); 2488 blkptr_t bp = *db->db_blkptr; 2489 dbuf_clear(db); 2490 arc_freed(spa, &bp); 2491 } else { 2492 dbuf_clear(db); 2493 } 2494 } else if (db->db_objset->os_evicting || 2495 arc_buf_eviction_needed(db->db_buf)) { 2496 dbuf_clear(db); 2497 } else { 2498 mutex_exit(&db->db_mtx); 2499 } 2500 } 2501 } else { 2502 mutex_exit(&db->db_mtx); 2503 } 2504} 2505 2506#pragma weak dmu_buf_refcount = dbuf_refcount 2507uint64_t 2508dbuf_refcount(dmu_buf_impl_t *db) 2509{ 2510 return (refcount_count(&db->db_holds)); 2511} 2512 2513void * 2514dmu_buf_replace_user(dmu_buf_t *db_fake, dmu_buf_user_t *old_user, 2515 dmu_buf_user_t *new_user) 2516{ 2517 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2518 2519 mutex_enter(&db->db_mtx); 2520 dbuf_verify_user(db, DBVU_NOT_EVICTING); 2521 if (db->db_user == old_user) 2522 db->db_user = new_user; 2523 else 2524 old_user = db->db_user; 2525 dbuf_verify_user(db, DBVU_NOT_EVICTING); 2526 mutex_exit(&db->db_mtx); 2527 2528 return (old_user); 2529} 2530 2531void * 2532dmu_buf_set_user(dmu_buf_t *db_fake, dmu_buf_user_t *user) 2533{ 2534 return (dmu_buf_replace_user(db_fake, NULL, user)); 2535} 2536 2537void * 2538dmu_buf_set_user_ie(dmu_buf_t *db_fake, dmu_buf_user_t *user) 2539{ 2540 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2541 2542 db->db_immediate_evict = TRUE; 2543 return (dmu_buf_set_user(db_fake, user)); 2544} 2545 2546void * 2547dmu_buf_remove_user(dmu_buf_t *db_fake, dmu_buf_user_t *user) 2548{ 2549 return (dmu_buf_replace_user(db_fake, user, NULL)); 2550} 2551 2552void * 2553dmu_buf_get_user(dmu_buf_t *db_fake) 2554{ 2555 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2556 2557 dbuf_verify_user(db, DBVU_NOT_EVICTING); 2558 return (db->db_user); 2559} 2560 2561void 2562dmu_buf_user_evict_wait() 2563{ 2564 taskq_wait(dbu_evict_taskq); 2565} 2566 2567boolean_t 2568dmu_buf_freeable(dmu_buf_t *dbuf) 2569{ 2570 boolean_t res = B_FALSE; 2571 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; 2572 2573 if (db->db_blkptr) 2574 res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset, 2575 db->db_blkptr, db->db_blkptr->blk_birth); 2576 2577 return (res); 2578} 2579 2580blkptr_t * 2581dmu_buf_get_blkptr(dmu_buf_t *db) 2582{ 2583 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 2584 return (dbi->db_blkptr); 2585} 2586 2587static void 2588dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db) 2589{ 2590 /* ASSERT(dmu_tx_is_syncing(tx) */ 2591 ASSERT(MUTEX_HELD(&db->db_mtx)); 2592 2593 if (db->db_blkptr != NULL) 2594 return; 2595 2596 if (db->db_blkid == DMU_SPILL_BLKID) { 2597 db->db_blkptr = &dn->dn_phys->dn_spill; 2598 BP_ZERO(db->db_blkptr); 2599 return; 2600 } 2601 if (db->db_level == dn->dn_phys->dn_nlevels-1) { 2602 /* 2603 * This buffer was allocated at a time when there was 2604 * no available blkptrs from the dnode, or it was 2605 * inappropriate to hook it in (i.e., nlevels mis-match). 2606 */ 2607 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr); 2608 ASSERT(db->db_parent == NULL); 2609 db->db_parent = dn->dn_dbuf; 2610 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid]; 2611 DBUF_VERIFY(db); 2612 } else { 2613 dmu_buf_impl_t *parent = db->db_parent; 2614 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 2615 2616 ASSERT(dn->dn_phys->dn_nlevels > 1); 2617 if (parent == NULL) { 2618 mutex_exit(&db->db_mtx); 2619 rw_enter(&dn->dn_struct_rwlock, RW_READER); 2620 parent = dbuf_hold_level(dn, db->db_level + 1, 2621 db->db_blkid >> epbs, db); 2622 rw_exit(&dn->dn_struct_rwlock); 2623 mutex_enter(&db->db_mtx); 2624 db->db_parent = parent; 2625 } 2626 db->db_blkptr = (blkptr_t *)parent->db.db_data + 2627 (db->db_blkid & ((1ULL << epbs) - 1)); 2628 DBUF_VERIFY(db); 2629 } 2630} 2631 2632static void 2633dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx) 2634{ 2635 dmu_buf_impl_t *db = dr->dr_dbuf; 2636 dnode_t *dn; 2637 zio_t *zio; 2638 2639 ASSERT(dmu_tx_is_syncing(tx)); 2640 2641 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 2642 2643 mutex_enter(&db->db_mtx); 2644 2645 ASSERT(db->db_level > 0); 2646 DBUF_VERIFY(db); 2647 2648 /* Read the block if it hasn't been read yet. */ 2649 if (db->db_buf == NULL) { 2650 mutex_exit(&db->db_mtx); 2651 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED); 2652 mutex_enter(&db->db_mtx); 2653 } 2654 ASSERT3U(db->db_state, ==, DB_CACHED); 2655 ASSERT(db->db_buf != NULL); 2656 2657 DB_DNODE_ENTER(db); 2658 dn = DB_DNODE(db); 2659 /* Indirect block size must match what the dnode thinks it is. */ 2660 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 2661 dbuf_check_blkptr(dn, db); 2662 DB_DNODE_EXIT(db); 2663 2664 /* Provide the pending dirty record to child dbufs */ 2665 db->db_data_pending = dr; 2666 2667 mutex_exit(&db->db_mtx); 2668 dbuf_write(dr, db->db_buf, tx); 2669 2670 zio = dr->dr_zio; 2671 mutex_enter(&dr->dt.di.dr_mtx); 2672 dbuf_sync_list(&dr->dt.di.dr_children, db->db_level - 1, tx); 2673 ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 2674 mutex_exit(&dr->dt.di.dr_mtx); 2675 zio_nowait(zio); 2676} 2677 2678static void 2679dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx) 2680{ 2681 arc_buf_t **datap = &dr->dt.dl.dr_data; 2682 dmu_buf_impl_t *db = dr->dr_dbuf; 2683 dnode_t *dn; 2684 objset_t *os; 2685 uint64_t txg = tx->tx_txg; 2686 2687 ASSERT(dmu_tx_is_syncing(tx)); 2688 2689 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 2690 2691 mutex_enter(&db->db_mtx); 2692 /* 2693 * To be synced, we must be dirtied. But we 2694 * might have been freed after the dirty. 2695 */ 2696 if (db->db_state == DB_UNCACHED) { 2697 /* This buffer has been freed since it was dirtied */ 2698 ASSERT(db->db.db_data == NULL); 2699 } else if (db->db_state == DB_FILL) { 2700 /* This buffer was freed and is now being re-filled */ 2701 ASSERT(db->db.db_data != dr->dt.dl.dr_data); 2702 } else { 2703 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL); 2704 } 2705 DBUF_VERIFY(db); 2706 2707 DB_DNODE_ENTER(db); 2708 dn = DB_DNODE(db); 2709 2710 if (db->db_blkid == DMU_SPILL_BLKID) { 2711 mutex_enter(&dn->dn_mtx); 2712 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR; 2713 mutex_exit(&dn->dn_mtx); 2714 } 2715 2716 /* 2717 * If this is a bonus buffer, simply copy the bonus data into the 2718 * dnode. It will be written out when the dnode is synced (and it 2719 * will be synced, since it must have been dirty for dbuf_sync to 2720 * be called). 2721 */ 2722 if (db->db_blkid == DMU_BONUS_BLKID) { 2723 dbuf_dirty_record_t **drp; 2724 2725 ASSERT(*datap != NULL); 2726 ASSERT0(db->db_level); 2727 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN); 2728 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen); 2729 DB_DNODE_EXIT(db); 2730 2731 if (*datap != db->db.db_data) { 2732 zio_buf_free(*datap, DN_MAX_BONUSLEN); 2733 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 2734 } 2735 db->db_data_pending = NULL; 2736 drp = &db->db_last_dirty; 2737 while (*drp != dr) 2738 drp = &(*drp)->dr_next; 2739 ASSERT(dr->dr_next == NULL); 2740 ASSERT(dr->dr_dbuf == db); 2741 *drp = dr->dr_next; 2742 if (dr->dr_dbuf->db_level != 0) { 2743 list_destroy(&dr->dt.di.dr_children); 2744 mutex_destroy(&dr->dt.di.dr_mtx); 2745 } 2746 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 2747 ASSERT(db->db_dirtycnt > 0); 2748 db->db_dirtycnt -= 1; 2749 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg); 2750 return; 2751 } 2752 2753 os = dn->dn_objset; 2754 2755 /* 2756 * This function may have dropped the db_mtx lock allowing a dmu_sync 2757 * operation to sneak in. As a result, we need to ensure that we 2758 * don't check the dr_override_state until we have returned from 2759 * dbuf_check_blkptr. 2760 */ 2761 dbuf_check_blkptr(dn, db); 2762 2763 /* 2764 * If this buffer is in the middle of an immediate write, 2765 * wait for the synchronous IO to complete. 2766 */ 2767 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) { 2768 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT); 2769 cv_wait(&db->db_changed, &db->db_mtx); 2770 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN); 2771 } 2772 2773 if (db->db_state != DB_NOFILL && 2774 dn->dn_object != DMU_META_DNODE_OBJECT && 2775 refcount_count(&db->db_holds) > 1 && 2776 dr->dt.dl.dr_override_state != DR_OVERRIDDEN && 2777 *datap == db->db_buf) { 2778 /* 2779 * If this buffer is currently "in use" (i.e., there 2780 * are active holds and db_data still references it), 2781 * then make a copy before we start the write so that 2782 * any modifications from the open txg will not leak 2783 * into this write. 2784 * 2785 * NOTE: this copy does not need to be made for 2786 * objects only modified in the syncing context (e.g. 2787 * DNONE_DNODE blocks). 2788 */ 2789 int blksz = arc_buf_size(*datap); 2790 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 2791 *datap = arc_buf_alloc(os->os_spa, blksz, db, type); 2792 bcopy(db->db.db_data, (*datap)->b_data, blksz); 2793 } 2794 db->db_data_pending = dr; 2795 2796 mutex_exit(&db->db_mtx); 2797 2798 dbuf_write(dr, *datap, tx); 2799 2800 ASSERT(!list_link_active(&dr->dr_dirty_node)); 2801 if (dn->dn_object == DMU_META_DNODE_OBJECT) { 2802 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr); 2803 DB_DNODE_EXIT(db); 2804 } else { 2805 /* 2806 * Although zio_nowait() does not "wait for an IO", it does 2807 * initiate the IO. If this is an empty write it seems plausible 2808 * that the IO could actually be completed before the nowait 2809 * returns. We need to DB_DNODE_EXIT() first in case 2810 * zio_nowait() invalidates the dbuf. 2811 */ 2812 DB_DNODE_EXIT(db); 2813 zio_nowait(dr->dr_zio); 2814 } 2815} 2816 2817void 2818dbuf_sync_list(list_t *list, int level, dmu_tx_t *tx) 2819{ 2820 dbuf_dirty_record_t *dr; 2821 2822 while (dr = list_head(list)) { 2823 if (dr->dr_zio != NULL) { 2824 /* 2825 * If we find an already initialized zio then we 2826 * are processing the meta-dnode, and we have finished. 2827 * The dbufs for all dnodes are put back on the list 2828 * during processing, so that we can zio_wait() 2829 * these IOs after initiating all child IOs. 2830 */ 2831 ASSERT3U(dr->dr_dbuf->db.db_object, ==, 2832 DMU_META_DNODE_OBJECT); 2833 break; 2834 } 2835 if (dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID && 2836 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) { 2837 VERIFY3U(dr->dr_dbuf->db_level, ==, level); 2838 } 2839 list_remove(list, dr); 2840 if (dr->dr_dbuf->db_level > 0) 2841 dbuf_sync_indirect(dr, tx); 2842 else 2843 dbuf_sync_leaf(dr, tx); 2844 } 2845} 2846 2847/* ARGSUSED */ 2848static void 2849dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb) 2850{ 2851 dmu_buf_impl_t *db = vdb; 2852 dnode_t *dn; 2853 blkptr_t *bp = zio->io_bp; 2854 blkptr_t *bp_orig = &zio->io_bp_orig; 2855 spa_t *spa = zio->io_spa; 2856 int64_t delta; 2857 uint64_t fill = 0; 2858 int i; 2859 2860 ASSERT3P(db->db_blkptr, ==, bp); 2861 2862 DB_DNODE_ENTER(db); 2863 dn = DB_DNODE(db); 2864 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig); 2865 dnode_diduse_space(dn, delta - zio->io_prev_space_delta); 2866 zio->io_prev_space_delta = delta; 2867 2868 if (bp->blk_birth != 0) { 2869 ASSERT((db->db_blkid != DMU_SPILL_BLKID && 2870 BP_GET_TYPE(bp) == dn->dn_type) || 2871 (db->db_blkid == DMU_SPILL_BLKID && 2872 BP_GET_TYPE(bp) == dn->dn_bonustype) || 2873 BP_IS_EMBEDDED(bp)); 2874 ASSERT(BP_GET_LEVEL(bp) == db->db_level); 2875 } 2876 2877 mutex_enter(&db->db_mtx); 2878 2879#ifdef ZFS_DEBUG 2880 if (db->db_blkid == DMU_SPILL_BLKID) { 2881 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); 2882 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && 2883 db->db_blkptr == &dn->dn_phys->dn_spill); 2884 } 2885#endif 2886 2887 if (db->db_level == 0) { 2888 mutex_enter(&dn->dn_mtx); 2889 if (db->db_blkid > dn->dn_phys->dn_maxblkid && 2890 db->db_blkid != DMU_SPILL_BLKID) 2891 dn->dn_phys->dn_maxblkid = db->db_blkid; 2892 mutex_exit(&dn->dn_mtx); 2893 2894 if (dn->dn_type == DMU_OT_DNODE) { 2895 dnode_phys_t *dnp = db->db.db_data; 2896 for (i = db->db.db_size >> DNODE_SHIFT; i > 0; 2897 i--, dnp++) { 2898 if (dnp->dn_type != DMU_OT_NONE) 2899 fill++; 2900 } 2901 } else { 2902 if (BP_IS_HOLE(bp)) { 2903 fill = 0; 2904 } else { 2905 fill = 1; 2906 } 2907 } 2908 } else { 2909 blkptr_t *ibp = db->db.db_data; 2910 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 2911 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) { 2912 if (BP_IS_HOLE(ibp)) 2913 continue; 2914 fill += BP_GET_FILL(ibp); 2915 } 2916 } 2917 DB_DNODE_EXIT(db); 2918 2919 if (!BP_IS_EMBEDDED(bp)) 2920 bp->blk_fill = fill; 2921 2922 mutex_exit(&db->db_mtx); 2923} 2924 2925/* 2926 * The SPA will call this callback several times for each zio - once 2927 * for every physical child i/o (zio->io_phys_children times). This 2928 * allows the DMU to monitor the progress of each logical i/o. For example, 2929 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z 2930 * block. There may be a long delay before all copies/fragments are completed, 2931 * so this callback allows us to retire dirty space gradually, as the physical 2932 * i/os complete. 2933 */ 2934/* ARGSUSED */ 2935static void 2936dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg) 2937{ 2938 dmu_buf_impl_t *db = arg; 2939 objset_t *os = db->db_objset; 2940 dsl_pool_t *dp = dmu_objset_pool(os); 2941 dbuf_dirty_record_t *dr; 2942 int delta = 0; 2943 2944 dr = db->db_data_pending; 2945 ASSERT3U(dr->dr_txg, ==, zio->io_txg); 2946 2947 /* 2948 * The callback will be called io_phys_children times. Retire one 2949 * portion of our dirty space each time we are called. Any rounding 2950 * error will be cleaned up by dsl_pool_sync()'s call to 2951 * dsl_pool_undirty_space(). 2952 */ 2953 delta = dr->dr_accounted / zio->io_phys_children; 2954 dsl_pool_undirty_space(dp, delta, zio->io_txg); 2955} 2956 2957/* ARGSUSED */ 2958static void 2959dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb) 2960{ 2961 dmu_buf_impl_t *db = vdb; 2962 blkptr_t *bp_orig = &zio->io_bp_orig; 2963 blkptr_t *bp = db->db_blkptr; 2964 objset_t *os = db->db_objset; 2965 dmu_tx_t *tx = os->os_synctx; 2966 dbuf_dirty_record_t **drp, *dr; 2967 2968 ASSERT0(zio->io_error); 2969 ASSERT(db->db_blkptr == bp); 2970 2971 /* 2972 * For nopwrites and rewrites we ensure that the bp matches our 2973 * original and bypass all the accounting. 2974 */ 2975 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) { 2976 ASSERT(BP_EQUAL(bp, bp_orig)); 2977 } else { 2978 dsl_dataset_t *ds = os->os_dsl_dataset; 2979 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE); 2980 dsl_dataset_block_born(ds, bp, tx); 2981 } 2982 2983 mutex_enter(&db->db_mtx); 2984 2985 DBUF_VERIFY(db); 2986 2987 drp = &db->db_last_dirty; 2988 while ((dr = *drp) != db->db_data_pending) 2989 drp = &dr->dr_next; 2990 ASSERT(!list_link_active(&dr->dr_dirty_node)); 2991 ASSERT(dr->dr_dbuf == db); 2992 ASSERT(dr->dr_next == NULL); 2993 *drp = dr->dr_next; 2994 2995#ifdef ZFS_DEBUG 2996 if (db->db_blkid == DMU_SPILL_BLKID) { 2997 dnode_t *dn; 2998 2999 DB_DNODE_ENTER(db); 3000 dn = DB_DNODE(db); 3001 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); 3002 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && 3003 db->db_blkptr == &dn->dn_phys->dn_spill); 3004 DB_DNODE_EXIT(db); 3005 } 3006#endif 3007 3008 if (db->db_level == 0) { 3009 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 3010 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); 3011 if (db->db_state != DB_NOFILL) { 3012 if (dr->dt.dl.dr_data != db->db_buf) 3013 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, 3014 db)); 3015 else if (!arc_released(db->db_buf)) 3016 arc_set_callback(db->db_buf, dbuf_do_evict, db); 3017 } 3018 } else { 3019 dnode_t *dn; 3020 3021 DB_DNODE_ENTER(db); 3022 dn = DB_DNODE(db); 3023 ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 3024 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift); 3025 if (!BP_IS_HOLE(db->db_blkptr)) { 3026 int epbs = 3027 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 3028 ASSERT3U(db->db_blkid, <=, 3029 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs)); 3030 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, 3031 db->db.db_size); 3032 if (!arc_released(db->db_buf)) 3033 arc_set_callback(db->db_buf, dbuf_do_evict, db); 3034 } 3035 DB_DNODE_EXIT(db); 3036 mutex_destroy(&dr->dt.di.dr_mtx); 3037 list_destroy(&dr->dt.di.dr_children); 3038 } 3039 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 3040 3041 cv_broadcast(&db->db_changed); 3042 ASSERT(db->db_dirtycnt > 0); 3043 db->db_dirtycnt -= 1; 3044 db->db_data_pending = NULL; 3045 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg); 3046} 3047 3048static void 3049dbuf_write_nofill_ready(zio_t *zio) 3050{ 3051 dbuf_write_ready(zio, NULL, zio->io_private); 3052} 3053 3054static void 3055dbuf_write_nofill_done(zio_t *zio) 3056{ 3057 dbuf_write_done(zio, NULL, zio->io_private); 3058} 3059 3060static void 3061dbuf_write_override_ready(zio_t *zio) 3062{ 3063 dbuf_dirty_record_t *dr = zio->io_private; 3064 dmu_buf_impl_t *db = dr->dr_dbuf; 3065 3066 dbuf_write_ready(zio, NULL, db); 3067} 3068 3069static void 3070dbuf_write_override_done(zio_t *zio) 3071{ 3072 dbuf_dirty_record_t *dr = zio->io_private; 3073 dmu_buf_impl_t *db = dr->dr_dbuf; 3074 blkptr_t *obp = &dr->dt.dl.dr_overridden_by; 3075 3076 mutex_enter(&db->db_mtx); 3077 if (!BP_EQUAL(zio->io_bp, obp)) { 3078 if (!BP_IS_HOLE(obp)) 3079 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp); 3080 arc_release(dr->dt.dl.dr_data, db); 3081 } 3082 mutex_exit(&db->db_mtx); 3083 3084 dbuf_write_done(zio, NULL, db); 3085} 3086 3087/* Issue I/O to commit a dirty buffer to disk. */ 3088static void 3089dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx) 3090{ 3091 dmu_buf_impl_t *db = dr->dr_dbuf; 3092 dnode_t *dn; 3093 objset_t *os; 3094 dmu_buf_impl_t *parent = db->db_parent; 3095 uint64_t txg = tx->tx_txg; 3096 zbookmark_phys_t zb; 3097 zio_prop_t zp; 3098 zio_t *zio; 3099 int wp_flag = 0; 3100 3101 DB_DNODE_ENTER(db); 3102 dn = DB_DNODE(db); 3103 os = dn->dn_objset; 3104 3105 if (db->db_state != DB_NOFILL) { 3106 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) { 3107 /* 3108 * Private object buffers are released here rather 3109 * than in dbuf_dirty() since they are only modified 3110 * in the syncing context and we don't want the 3111 * overhead of making multiple copies of the data. 3112 */ 3113 if (BP_IS_HOLE(db->db_blkptr)) { 3114 arc_buf_thaw(data); 3115 } else { 3116 dbuf_release_bp(db); 3117 } 3118 } 3119 } 3120 3121 if (parent != dn->dn_dbuf) { 3122 /* Our parent is an indirect block. */ 3123 /* We have a dirty parent that has been scheduled for write. */ 3124 ASSERT(parent && parent->db_data_pending); 3125 /* Our parent's buffer is one level closer to the dnode. */ 3126 ASSERT(db->db_level == parent->db_level-1); 3127 /* 3128 * We're about to modify our parent's db_data by modifying 3129 * our block pointer, so the parent must be released. 3130 */ 3131 ASSERT(arc_released(parent->db_buf)); 3132 zio = parent->db_data_pending->dr_zio; 3133 } else { 3134 /* Our parent is the dnode itself. */ 3135 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 && 3136 db->db_blkid != DMU_SPILL_BLKID) || 3137 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0)); 3138 if (db->db_blkid != DMU_SPILL_BLKID) 3139 ASSERT3P(db->db_blkptr, ==, 3140 &dn->dn_phys->dn_blkptr[db->db_blkid]); 3141 zio = dn->dn_zio; 3142 } 3143 3144 ASSERT(db->db_level == 0 || data == db->db_buf); 3145 ASSERT3U(db->db_blkptr->blk_birth, <=, txg); 3146 ASSERT(zio); 3147 3148 SET_BOOKMARK(&zb, os->os_dsl_dataset ? 3149 os->os_dsl_dataset->ds_object : DMU_META_OBJSET, 3150 db->db.db_object, db->db_level, db->db_blkid); 3151 3152 if (db->db_blkid == DMU_SPILL_BLKID) 3153 wp_flag = WP_SPILL; 3154 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0; 3155 3156 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp); 3157 DB_DNODE_EXIT(db); 3158 3159 if (db->db_level == 0 && 3160 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { 3161 /* 3162 * The BP for this block has been provided by open context 3163 * (by dmu_sync() or dmu_buf_write_embedded()). 3164 */ 3165 void *contents = (data != NULL) ? data->b_data : NULL; 3166 3167 dr->dr_zio = zio_write(zio, os->os_spa, txg, 3168 db->db_blkptr, contents, db->db.db_size, &zp, 3169 dbuf_write_override_ready, NULL, dbuf_write_override_done, 3170 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); 3171 mutex_enter(&db->db_mtx); 3172 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 3173 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by, 3174 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite); 3175 mutex_exit(&db->db_mtx); 3176 } else if (db->db_state == DB_NOFILL) { 3177 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF || 3178 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY); 3179 dr->dr_zio = zio_write(zio, os->os_spa, txg, 3180 db->db_blkptr, NULL, db->db.db_size, &zp, 3181 dbuf_write_nofill_ready, NULL, dbuf_write_nofill_done, db, 3182 ZIO_PRIORITY_ASYNC_WRITE, 3183 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb); 3184 } else { 3185 ASSERT(arc_released(data)); 3186 dr->dr_zio = arc_write(zio, os->os_spa, txg, 3187 db->db_blkptr, data, DBUF_IS_L2CACHEABLE(db), 3188 DBUF_IS_L2COMPRESSIBLE(db), &zp, dbuf_write_ready, 3189 dbuf_write_physdone, dbuf_write_done, db, 3190 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); 3191 } 3192} 3193