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, 2014 by Delphix. All rights reserved. 25 * Copyright (c) 2013 by Saso Kiselkov. All rights reserved. 26 * Copyright (c) 2013, Joyent, Inc. All rights reserved. 27 */ 28 29 #include <sys/zfs_context.h> 30 #include <sys/dmu.h> 31 #include <sys/dmu_send.h> 32 #include <sys/dmu_impl.h> 33 #include <sys/dbuf.h> 34 #include <sys/dmu_objset.h> 35 #include <sys/dsl_dataset.h> 36 #include <sys/dsl_dir.h> 37 #include <sys/dmu_tx.h> 38 #include <sys/spa.h> 39 #include <sys/zio.h> 40 #include <sys/dmu_zfetch.h> 41 #include <sys/sa.h> 42 #include <sys/sa_impl.h> 43 #include <sys/zfeature.h> 44 #include <sys/blkptr.h> 45 #include <sys/range_tree.h> 46 47 /* 48 * Number of times that zfs_free_range() took the slow path while doing 49 * a zfs receive. A nonzero value indicates a potential performance problem. 50 */ 51 uint64_t zfs_free_range_recv_miss; 52 53 static void dbuf_destroy(dmu_buf_impl_t *db); 54 static boolean_t dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx); 55 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx); 56 57 /* 58 * Global data structures and functions for the dbuf cache. 59 */ 60 static kmem_cache_t *dbuf_cache; 61 62 /* ARGSUSED */ 63 static int 64 dbuf_cons(void *vdb, void *unused, int kmflag) 65 { 66 dmu_buf_impl_t *db = vdb; 67 bzero(db, sizeof (dmu_buf_impl_t)); 68 69 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL); 70 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL); 71 refcount_create(&db->db_holds); 72 return (0); 73 } 74 75 /* ARGSUSED */ 76 static void 77 dbuf_dest(void *vdb, void *unused) 78 { 79 dmu_buf_impl_t *db = vdb; 80 mutex_destroy(&db->db_mtx); 81 cv_destroy(&db->db_changed); 82 refcount_destroy(&db->db_holds); 83 } 84 85 /* 86 * dbuf hash table routines 87 */ 88 static dbuf_hash_table_t dbuf_hash_table; 89 90 static uint64_t dbuf_hash_count; 91 92 static uint64_t 93 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid) 94 { 95 uintptr_t osv = (uintptr_t)os; 96 uint64_t crc = -1ULL; 97 98 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 99 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF]; 100 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF]; 101 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF]; 102 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF]; 103 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF]; 104 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF]; 105 106 crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16); 107 108 return (crc); 109 } 110 111 #define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid); 112 113 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \ 114 ((dbuf)->db.db_object == (obj) && \ 115 (dbuf)->db_objset == (os) && \ 116 (dbuf)->db_level == (level) && \ 117 (dbuf)->db_blkid == (blkid)) 118 119 dmu_buf_impl_t * 120 dbuf_find(dnode_t *dn, uint8_t level, uint64_t blkid) 121 { 122 dbuf_hash_table_t *h = &dbuf_hash_table; 123 objset_t *os = dn->dn_objset; 124 uint64_t obj = dn->dn_object; 125 uint64_t hv = DBUF_HASH(os, obj, level, blkid); 126 uint64_t idx = hv & h->hash_table_mask; 127 dmu_buf_impl_t *db; 128 129 mutex_enter(DBUF_HASH_MUTEX(h, idx)); 130 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) { 131 if (DBUF_EQUAL(db, os, obj, level, blkid)) { 132 mutex_enter(&db->db_mtx); 133 if (db->db_state != DB_EVICTING) { 134 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 135 return (db); 136 } 137 mutex_exit(&db->db_mtx); 138 } 139 } 140 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 141 return (NULL); 142 } 143 144 /* 145 * Insert an entry into the hash table. If there is already an element 146 * equal to elem in the hash table, then the already existing element 147 * will be returned and the new element will not be inserted. 148 * Otherwise returns NULL. 149 */ 150 static dmu_buf_impl_t * 151 dbuf_hash_insert(dmu_buf_impl_t *db) 152 { 153 dbuf_hash_table_t *h = &dbuf_hash_table; 154 objset_t *os = db->db_objset; 155 uint64_t obj = db->db.db_object; 156 int level = db->db_level; 157 uint64_t blkid = db->db_blkid; 158 uint64_t hv = DBUF_HASH(os, obj, level, blkid); 159 uint64_t idx = hv & h->hash_table_mask; 160 dmu_buf_impl_t *dbf; 161 162 mutex_enter(DBUF_HASH_MUTEX(h, idx)); 163 for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) { 164 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) { 165 mutex_enter(&dbf->db_mtx); 166 if (dbf->db_state != DB_EVICTING) { 167 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 168 return (dbf); 169 } 170 mutex_exit(&dbf->db_mtx); 171 } 172 } 173 174 mutex_enter(&db->db_mtx); 175 db->db_hash_next = h->hash_table[idx]; 176 h->hash_table[idx] = db; 177 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 178 atomic_add_64(&dbuf_hash_count, 1); 179 180 return (NULL); 181 } 182 183 /* 184 * Remove an entry from the hash table. This operation will 185 * fail if there are any existing holds on the db. 186 */ 187 static void 188 dbuf_hash_remove(dmu_buf_impl_t *db) 189 { 190 dbuf_hash_table_t *h = &dbuf_hash_table; 191 uint64_t hv = DBUF_HASH(db->db_objset, db->db.db_object, 192 db->db_level, db->db_blkid); 193 uint64_t idx = hv & h->hash_table_mask; 194 dmu_buf_impl_t *dbf, **dbp; 195 196 /* 197 * We musn't hold db_mtx to maintin lock ordering: 198 * DBUF_HASH_MUTEX > db_mtx. 199 */ 200 ASSERT(refcount_is_zero(&db->db_holds)); 201 ASSERT(db->db_state == DB_EVICTING); 202 ASSERT(!MUTEX_HELD(&db->db_mtx)); 203 204 mutex_enter(DBUF_HASH_MUTEX(h, idx)); 205 dbp = &h->hash_table[idx]; 206 while ((dbf = *dbp) != db) { 207 dbp = &dbf->db_hash_next; 208 ASSERT(dbf != NULL); 209 } 210 *dbp = db->db_hash_next; 211 db->db_hash_next = NULL; 212 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 213 atomic_add_64(&dbuf_hash_count, -1); 214 } 215 216 static arc_evict_func_t dbuf_do_evict; 217 218 static void 219 dbuf_evict_user(dmu_buf_impl_t *db) 220 { 221 ASSERT(MUTEX_HELD(&db->db_mtx)); 222 223 if (db->db_level != 0 || db->db_evict_func == NULL) 224 return; 225 226 if (db->db_user_data_ptr_ptr) 227 *db->db_user_data_ptr_ptr = db->db.db_data; 228 db->db_evict_func(&db->db, db->db_user_ptr); 229 db->db_user_ptr = NULL; 230 db->db_user_data_ptr_ptr = NULL; 231 db->db_evict_func = NULL; 232 } 233 234 boolean_t 235 dbuf_is_metadata(dmu_buf_impl_t *db) 236 { 237 if (db->db_level > 0) { 238 return (B_TRUE); 239 } else { 240 boolean_t is_metadata; 241 242 DB_DNODE_ENTER(db); 243 is_metadata = DMU_OT_IS_METADATA(DB_DNODE(db)->dn_type); 244 DB_DNODE_EXIT(db); 245 246 return (is_metadata); 247 } 248 } 249 250 void 251 dbuf_evict(dmu_buf_impl_t *db) 252 { 253 ASSERT(MUTEX_HELD(&db->db_mtx)); 254 ASSERT(db->db_buf == NULL); 255 ASSERT(db->db_data_pending == NULL); 256 257 dbuf_clear(db); 258 dbuf_destroy(db); 259 } 260 261 void 262 dbuf_init(void) 263 { 264 uint64_t hsize = 1ULL << 16; 265 dbuf_hash_table_t *h = &dbuf_hash_table; 266 int i; 267 268 /* 269 * The hash table is big enough to fill all of physical memory 270 * with an average 4K block size. The table will take up 271 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers). 272 */ 273 while (hsize * 4096 < physmem * PAGESIZE) 274 hsize <<= 1; 275 276 retry: 277 h->hash_table_mask = hsize - 1; 278 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP); 279 if (h->hash_table == NULL) { 280 /* XXX - we should really return an error instead of assert */ 281 ASSERT(hsize > (1ULL << 10)); 282 hsize >>= 1; 283 goto retry; 284 } 285 286 dbuf_cache = kmem_cache_create("dmu_buf_impl_t", 287 sizeof (dmu_buf_impl_t), 288 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0); 289 290 for (i = 0; i < DBUF_MUTEXES; i++) 291 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL); 292 } 293 294 void 295 dbuf_fini(void) 296 { 297 dbuf_hash_table_t *h = &dbuf_hash_table; 298 int i; 299 300 for (i = 0; i < DBUF_MUTEXES; i++) 301 mutex_destroy(&h->hash_mutexes[i]); 302 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *)); 303 kmem_cache_destroy(dbuf_cache); 304 } 305 306 /* 307 * Other stuff. 308 */ 309 310 #ifdef ZFS_DEBUG 311 static void 312 dbuf_verify(dmu_buf_impl_t *db) 313 { 314 dnode_t *dn; 315 dbuf_dirty_record_t *dr; 316 317 ASSERT(MUTEX_HELD(&db->db_mtx)); 318 319 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY)) 320 return; 321 322 ASSERT(db->db_objset != NULL); 323 DB_DNODE_ENTER(db); 324 dn = DB_DNODE(db); 325 if (dn == NULL) { 326 ASSERT(db->db_parent == NULL); 327 ASSERT(db->db_blkptr == NULL); 328 } else { 329 ASSERT3U(db->db.db_object, ==, dn->dn_object); 330 ASSERT3P(db->db_objset, ==, dn->dn_objset); 331 ASSERT3U(db->db_level, <, dn->dn_nlevels); 332 ASSERT(db->db_blkid == DMU_BONUS_BLKID || 333 db->db_blkid == DMU_SPILL_BLKID || 334 !list_is_empty(&dn->dn_dbufs)); 335 } 336 if (db->db_blkid == DMU_BONUS_BLKID) { 337 ASSERT(dn != NULL); 338 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 339 ASSERT3U(db->db.db_offset, ==, DMU_BONUS_BLKID); 340 } else if (db->db_blkid == DMU_SPILL_BLKID) { 341 ASSERT(dn != NULL); 342 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 343 ASSERT0(db->db.db_offset); 344 } else { 345 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size); 346 } 347 348 for (dr = db->db_data_pending; dr != NULL; dr = dr->dr_next) 349 ASSERT(dr->dr_dbuf == db); 350 351 for (dr = db->db_last_dirty; dr != NULL; dr = dr->dr_next) 352 ASSERT(dr->dr_dbuf == db); 353 354 /* 355 * We can't assert that db_size matches dn_datablksz because it 356 * can be momentarily different when another thread is doing 357 * dnode_set_blksz(). 358 */ 359 if (db->db_level == 0 && db->db.db_object == DMU_META_DNODE_OBJECT) { 360 dr = db->db_data_pending; 361 /* 362 * It should only be modified in syncing context, so 363 * make sure we only have one copy of the data. 364 */ 365 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf); 366 } 367 368 /* verify db->db_blkptr */ 369 if (db->db_blkptr) { 370 if (db->db_parent == dn->dn_dbuf) { 371 /* db is pointed to by the dnode */ 372 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */ 373 if (DMU_OBJECT_IS_SPECIAL(db->db.db_object)) 374 ASSERT(db->db_parent == NULL); 375 else 376 ASSERT(db->db_parent != NULL); 377 if (db->db_blkid != DMU_SPILL_BLKID) 378 ASSERT3P(db->db_blkptr, ==, 379 &dn->dn_phys->dn_blkptr[db->db_blkid]); 380 } else { 381 /* db is pointed to by an indirect block */ 382 int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT; 383 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1); 384 ASSERT3U(db->db_parent->db.db_object, ==, 385 db->db.db_object); 386 /* 387 * dnode_grow_indblksz() can make this fail if we don't 388 * have the struct_rwlock. XXX indblksz no longer 389 * grows. safe to do this now? 390 */ 391 if (RW_WRITE_HELD(&dn->dn_struct_rwlock)) { 392 ASSERT3P(db->db_blkptr, ==, 393 ((blkptr_t *)db->db_parent->db.db_data + 394 db->db_blkid % epb)); 395 } 396 } 397 } 398 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) && 399 (db->db_buf == NULL || db->db_buf->b_data) && 400 db->db.db_data && db->db_blkid != DMU_BONUS_BLKID && 401 db->db_state != DB_FILL && !dn->dn_free_txg) { 402 /* 403 * If the blkptr isn't set but they have nonzero data, 404 * it had better be dirty, otherwise we'll lose that 405 * data when we evict this buffer. 406 */ 407 if (db->db_dirtycnt == 0) { 408 uint64_t *buf = db->db.db_data; 409 int i; 410 411 for (i = 0; i < db->db.db_size >> 3; i++) { 412 ASSERT(buf[i] == 0); 413 } 414 } 415 } 416 DB_DNODE_EXIT(db); 417 } 418 #endif 419 420 static void 421 dbuf_update_data(dmu_buf_impl_t *db) 422 { 423 ASSERT(MUTEX_HELD(&db->db_mtx)); 424 if (db->db_level == 0 && db->db_user_data_ptr_ptr) { 425 ASSERT(!refcount_is_zero(&db->db_holds)); 426 *db->db_user_data_ptr_ptr = db->db.db_data; 427 } 428 } 429 430 static void 431 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf) 432 { 433 ASSERT(MUTEX_HELD(&db->db_mtx)); 434 ASSERT(db->db_buf == NULL || !arc_has_callback(db->db_buf)); 435 db->db_buf = buf; 436 if (buf != NULL) { 437 ASSERT(buf->b_data != NULL); 438 db->db.db_data = buf->b_data; 439 if (!arc_released(buf)) 440 arc_set_callback(buf, dbuf_do_evict, db); 441 dbuf_update_data(db); 442 } else { 443 dbuf_evict_user(db); 444 db->db.db_data = NULL; 445 if (db->db_state != DB_NOFILL) 446 db->db_state = DB_UNCACHED; 447 } 448 } 449 450 /* 451 * Loan out an arc_buf for read. Return the loaned arc_buf. 452 */ 453 arc_buf_t * 454 dbuf_loan_arcbuf(dmu_buf_impl_t *db) 455 { 456 arc_buf_t *abuf; 457 458 mutex_enter(&db->db_mtx); 459 if (arc_released(db->db_buf) || refcount_count(&db->db_holds) > 1) { 460 int blksz = db->db.db_size; 461 spa_t *spa = db->db_objset->os_spa; 462 463 mutex_exit(&db->db_mtx); 464 abuf = arc_loan_buf(spa, blksz); 465 bcopy(db->db.db_data, abuf->b_data, blksz); 466 } else { 467 abuf = db->db_buf; 468 arc_loan_inuse_buf(abuf, db); 469 dbuf_set_data(db, NULL); 470 mutex_exit(&db->db_mtx); 471 } 472 return (abuf); 473 } 474 475 uint64_t 476 dbuf_whichblock(dnode_t *dn, uint64_t offset) 477 { 478 if (dn->dn_datablkshift) { 479 return (offset >> dn->dn_datablkshift); 480 } else { 481 ASSERT3U(offset, <, dn->dn_datablksz); 482 return (0); 483 } 484 } 485 486 static void 487 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb) 488 { 489 dmu_buf_impl_t *db = vdb; 490 491 mutex_enter(&db->db_mtx); 492 ASSERT3U(db->db_state, ==, DB_READ); 493 /* 494 * All reads are synchronous, so we must have a hold on the dbuf 495 */ 496 ASSERT(refcount_count(&db->db_holds) > 0); 497 ASSERT(db->db_buf == NULL); 498 ASSERT(db->db.db_data == NULL); 499 if (db->db_level == 0 && db->db_freed_in_flight) { 500 /* we were freed in flight; disregard any error */ 501 arc_release(buf, db); 502 bzero(buf->b_data, db->db.db_size); 503 arc_buf_freeze(buf); 504 db->db_freed_in_flight = FALSE; 505 dbuf_set_data(db, buf); 506 db->db_state = DB_CACHED; 507 } else if (zio == NULL || zio->io_error == 0) { 508 dbuf_set_data(db, buf); 509 db->db_state = DB_CACHED; 510 } else { 511 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 512 ASSERT3P(db->db_buf, ==, NULL); 513 VERIFY(arc_buf_remove_ref(buf, db)); 514 db->db_state = DB_UNCACHED; 515 } 516 cv_broadcast(&db->db_changed); 517 dbuf_rele_and_unlock(db, NULL); 518 } 519 520 static void 521 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t *flags) 522 { 523 dnode_t *dn; 524 zbookmark_phys_t zb; 525 uint32_t aflags = ARC_NOWAIT; 526 527 DB_DNODE_ENTER(db); 528 dn = DB_DNODE(db); 529 ASSERT(!refcount_is_zero(&db->db_holds)); 530 /* We need the struct_rwlock to prevent db_blkptr from changing. */ 531 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 532 ASSERT(MUTEX_HELD(&db->db_mtx)); 533 ASSERT(db->db_state == DB_UNCACHED); 534 ASSERT(db->db_buf == NULL); 535 536 if (db->db_blkid == DMU_BONUS_BLKID) { 537 int bonuslen = MIN(dn->dn_bonuslen, dn->dn_phys->dn_bonuslen); 538 539 ASSERT3U(bonuslen, <=, db->db.db_size); 540 db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN); 541 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 542 if (bonuslen < DN_MAX_BONUSLEN) 543 bzero(db->db.db_data, DN_MAX_BONUSLEN); 544 if (bonuslen) 545 bcopy(DN_BONUS(dn->dn_phys), db->db.db_data, bonuslen); 546 DB_DNODE_EXIT(db); 547 dbuf_update_data(db); 548 db->db_state = DB_CACHED; 549 mutex_exit(&db->db_mtx); 550 return; 551 } 552 553 /* 554 * Recheck BP_IS_HOLE() after dnode_block_freed() in case dnode_sync() 555 * processes the delete record and clears the bp while we are waiting 556 * for the dn_mtx (resulting in a "no" from block_freed). 557 */ 558 if (db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr) || 559 (db->db_level == 0 && (dnode_block_freed(dn, db->db_blkid) || 560 BP_IS_HOLE(db->db_blkptr)))) { 561 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 562 563 DB_DNODE_EXIT(db); 564 dbuf_set_data(db, arc_buf_alloc(db->db_objset->os_spa, 565 db->db.db_size, db, type)); 566 bzero(db->db.db_data, db->db.db_size); 567 db->db_state = DB_CACHED; 568 *flags |= DB_RF_CACHED; 569 mutex_exit(&db->db_mtx); 570 return; 571 } 572 573 DB_DNODE_EXIT(db); 574 575 db->db_state = DB_READ; 576 mutex_exit(&db->db_mtx); 577 578 if (DBUF_IS_L2CACHEABLE(db)) 579 aflags |= ARC_L2CACHE; 580 if (DBUF_IS_L2COMPRESSIBLE(db)) 581 aflags |= ARC_L2COMPRESS; 582 583 SET_BOOKMARK(&zb, db->db_objset->os_dsl_dataset ? 584 db->db_objset->os_dsl_dataset->ds_object : DMU_META_OBJSET, 585 db->db.db_object, db->db_level, db->db_blkid); 586 587 dbuf_add_ref(db, NULL); 588 589 (void) arc_read(zio, db->db_objset->os_spa, db->db_blkptr, 590 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ, 591 (*flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED, 592 &aflags, &zb); 593 if (aflags & ARC_CACHED) 594 *flags |= DB_RF_CACHED; 595 } 596 597 int 598 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) 599 { 600 int err = 0; 601 boolean_t havepzio = (zio != NULL); 602 boolean_t prefetch; 603 dnode_t *dn; 604 605 /* 606 * We don't have to hold the mutex to check db_state because it 607 * can't be freed while we have a hold on the buffer. 608 */ 609 ASSERT(!refcount_is_zero(&db->db_holds)); 610 611 if (db->db_state == DB_NOFILL) 612 return (SET_ERROR(EIO)); 613 614 DB_DNODE_ENTER(db); 615 dn = DB_DNODE(db); 616 if ((flags & DB_RF_HAVESTRUCT) == 0) 617 rw_enter(&dn->dn_struct_rwlock, RW_READER); 618 619 prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 620 (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL && 621 DBUF_IS_CACHEABLE(db); 622 623 mutex_enter(&db->db_mtx); 624 if (db->db_state == DB_CACHED) { 625 mutex_exit(&db->db_mtx); 626 if (prefetch) 627 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset, 628 db->db.db_size, TRUE); 629 if ((flags & DB_RF_HAVESTRUCT) == 0) 630 rw_exit(&dn->dn_struct_rwlock); 631 DB_DNODE_EXIT(db); 632 } else if (db->db_state == DB_UNCACHED) { 633 spa_t *spa = dn->dn_objset->os_spa; 634 635 if (zio == NULL) 636 zio = zio_root(spa, NULL, NULL, ZIO_FLAG_CANFAIL); 637 dbuf_read_impl(db, zio, &flags); 638 639 /* dbuf_read_impl has dropped db_mtx for us */ 640 641 if (prefetch) 642 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset, 643 db->db.db_size, flags & DB_RF_CACHED); 644 645 if ((flags & DB_RF_HAVESTRUCT) == 0) 646 rw_exit(&dn->dn_struct_rwlock); 647 DB_DNODE_EXIT(db); 648 649 if (!havepzio) 650 err = zio_wait(zio); 651 } else { 652 /* 653 * Another reader came in while the dbuf was in flight 654 * between UNCACHED and CACHED. Either a writer will finish 655 * writing the buffer (sending the dbuf to CACHED) or the 656 * first reader's request will reach the read_done callback 657 * and send the dbuf to CACHED. Otherwise, a failure 658 * occurred and the dbuf went to UNCACHED. 659 */ 660 mutex_exit(&db->db_mtx); 661 if (prefetch) 662 dmu_zfetch(&dn->dn_zfetch, db->db.db_offset, 663 db->db.db_size, TRUE); 664 if ((flags & DB_RF_HAVESTRUCT) == 0) 665 rw_exit(&dn->dn_struct_rwlock); 666 DB_DNODE_EXIT(db); 667 668 /* Skip the wait per the caller's request. */ 669 mutex_enter(&db->db_mtx); 670 if ((flags & DB_RF_NEVERWAIT) == 0) { 671 while (db->db_state == DB_READ || 672 db->db_state == DB_FILL) { 673 ASSERT(db->db_state == DB_READ || 674 (flags & DB_RF_HAVESTRUCT) == 0); 675 cv_wait(&db->db_changed, &db->db_mtx); 676 } 677 if (db->db_state == DB_UNCACHED) 678 err = SET_ERROR(EIO); 679 } 680 mutex_exit(&db->db_mtx); 681 } 682 683 ASSERT(err || havepzio || db->db_state == DB_CACHED); 684 return (err); 685 } 686 687 static void 688 dbuf_noread(dmu_buf_impl_t *db) 689 { 690 ASSERT(!refcount_is_zero(&db->db_holds)); 691 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 692 mutex_enter(&db->db_mtx); 693 while (db->db_state == DB_READ || db->db_state == DB_FILL) 694 cv_wait(&db->db_changed, &db->db_mtx); 695 if (db->db_state == DB_UNCACHED) { 696 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 697 spa_t *spa = db->db_objset->os_spa; 698 699 ASSERT(db->db_buf == NULL); 700 ASSERT(db->db.db_data == NULL); 701 dbuf_set_data(db, arc_buf_alloc(spa, db->db.db_size, db, type)); 702 db->db_state = DB_FILL; 703 } else if (db->db_state == DB_NOFILL) { 704 dbuf_set_data(db, NULL); 705 } else { 706 ASSERT3U(db->db_state, ==, DB_CACHED); 707 } 708 mutex_exit(&db->db_mtx); 709 } 710 711 /* 712 * This is our just-in-time copy function. It makes a copy of 713 * buffers, that have been modified in a previous transaction 714 * group, before we modify them in the current active group. 715 * 716 * This function is used in two places: when we are dirtying a 717 * buffer for the first time in a txg, and when we are freeing 718 * a range in a dnode that includes this buffer. 719 * 720 * Note that when we are called from dbuf_free_range() we do 721 * not put a hold on the buffer, we just traverse the active 722 * dbuf list for the dnode. 723 */ 724 static void 725 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg) 726 { 727 dbuf_dirty_record_t *dr = db->db_last_dirty; 728 729 ASSERT(MUTEX_HELD(&db->db_mtx)); 730 ASSERT(db->db.db_data != NULL); 731 ASSERT(db->db_level == 0); 732 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT); 733 734 if (dr == NULL || 735 (dr->dt.dl.dr_data != 736 ((db->db_blkid == DMU_BONUS_BLKID) ? db->db.db_data : db->db_buf))) 737 return; 738 739 /* 740 * If the last dirty record for this dbuf has not yet synced 741 * and its referencing the dbuf data, either: 742 * reset the reference to point to a new copy, 743 * or (if there a no active holders) 744 * just null out the current db_data pointer. 745 */ 746 ASSERT(dr->dr_txg >= txg - 2); 747 if (db->db_blkid == DMU_BONUS_BLKID) { 748 /* Note that the data bufs here are zio_bufs */ 749 dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN); 750 arc_space_consume(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 751 bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN); 752 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) { 753 int size = db->db.db_size; 754 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 755 spa_t *spa = db->db_objset->os_spa; 756 757 dr->dt.dl.dr_data = arc_buf_alloc(spa, size, db, type); 758 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size); 759 } else { 760 dbuf_set_data(db, NULL); 761 } 762 } 763 764 void 765 dbuf_unoverride(dbuf_dirty_record_t *dr) 766 { 767 dmu_buf_impl_t *db = dr->dr_dbuf; 768 blkptr_t *bp = &dr->dt.dl.dr_overridden_by; 769 uint64_t txg = dr->dr_txg; 770 771 ASSERT(MUTEX_HELD(&db->db_mtx)); 772 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC); 773 ASSERT(db->db_level == 0); 774 775 if (db->db_blkid == DMU_BONUS_BLKID || 776 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN) 777 return; 778 779 ASSERT(db->db_data_pending != dr); 780 781 /* free this block */ 782 if (!BP_IS_HOLE(bp) && !dr->dt.dl.dr_nopwrite) 783 zio_free(db->db_objset->os_spa, txg, bp); 784 785 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 786 dr->dt.dl.dr_nopwrite = B_FALSE; 787 788 /* 789 * Release the already-written buffer, so we leave it in 790 * a consistent dirty state. Note that all callers are 791 * modifying the buffer, so they will immediately do 792 * another (redundant) arc_release(). Therefore, leave 793 * the buf thawed to save the effort of freezing & 794 * immediately re-thawing it. 795 */ 796 arc_release(dr->dt.dl.dr_data, db); 797 } 798 799 /* 800 * Evict (if its unreferenced) or clear (if its referenced) any level-0 801 * data blocks in the free range, so that any future readers will find 802 * empty blocks. 803 * 804 * This is a no-op if the dataset is in the middle of an incremental 805 * receive; see comment below for details. 806 */ 807 void 808 dbuf_free_range(dnode_t *dn, uint64_t start, uint64_t end, dmu_tx_t *tx) 809 { 810 dmu_buf_impl_t *db, *db_next; 811 uint64_t txg = tx->tx_txg; 812 813 if (end > dn->dn_maxblkid && (end != DMU_SPILL_BLKID)) 814 end = dn->dn_maxblkid; 815 dprintf_dnode(dn, "start=%llu end=%llu\n", start, end); 816 817 mutex_enter(&dn->dn_dbufs_mtx); 818 if (start >= dn->dn_unlisted_l0_blkid * dn->dn_datablksz) { 819 /* There can't be any dbufs in this range; no need to search. */ 820 mutex_exit(&dn->dn_dbufs_mtx); 821 return; 822 } else if (dmu_objset_is_receiving(dn->dn_objset)) { 823 /* 824 * If we are receiving, we expect there to be no dbufs in 825 * the range to be freed, because receive modifies each 826 * block at most once, and in offset order. If this is 827 * not the case, it can lead to performance problems, 828 * so note that we unexpectedly took the slow path. 829 */ 830 atomic_inc_64(&zfs_free_range_recv_miss); 831 } 832 833 for (db = list_head(&dn->dn_dbufs); db != NULL; db = db_next) { 834 db_next = list_next(&dn->dn_dbufs, db); 835 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 836 837 if (db->db_level != 0) 838 continue; 839 if (db->db_blkid < start || db->db_blkid > end) 840 continue; 841 842 /* found a level 0 buffer in the range */ 843 mutex_enter(&db->db_mtx); 844 if (dbuf_undirty(db, tx)) { 845 /* mutex has been dropped and dbuf destroyed */ 846 continue; 847 } 848 849 if (db->db_state == DB_UNCACHED || 850 db->db_state == DB_NOFILL || 851 db->db_state == DB_EVICTING) { 852 ASSERT(db->db.db_data == NULL); 853 mutex_exit(&db->db_mtx); 854 continue; 855 } 856 if (db->db_state == DB_READ || db->db_state == DB_FILL) { 857 /* will be handled in dbuf_read_done or dbuf_rele */ 858 db->db_freed_in_flight = TRUE; 859 mutex_exit(&db->db_mtx); 860 continue; 861 } 862 if (refcount_count(&db->db_holds) == 0) { 863 ASSERT(db->db_buf); 864 dbuf_clear(db); 865 continue; 866 } 867 /* The dbuf is referenced */ 868 869 if (db->db_last_dirty != NULL) { 870 dbuf_dirty_record_t *dr = db->db_last_dirty; 871 872 if (dr->dr_txg == txg) { 873 /* 874 * This buffer is "in-use", re-adjust the file 875 * size to reflect that this buffer may 876 * contain new data when we sync. 877 */ 878 if (db->db_blkid != DMU_SPILL_BLKID && 879 db->db_blkid > dn->dn_maxblkid) 880 dn->dn_maxblkid = db->db_blkid; 881 dbuf_unoverride(dr); 882 } else { 883 /* 884 * This dbuf is not dirty in the open context. 885 * Either uncache it (if its not referenced in 886 * the open context) or reset its contents to 887 * empty. 888 */ 889 dbuf_fix_old_data(db, txg); 890 } 891 } 892 /* clear the contents if its cached */ 893 if (db->db_state == DB_CACHED) { 894 ASSERT(db->db.db_data != NULL); 895 arc_release(db->db_buf, db); 896 bzero(db->db.db_data, db->db.db_size); 897 arc_buf_freeze(db->db_buf); 898 } 899 900 mutex_exit(&db->db_mtx); 901 } 902 mutex_exit(&dn->dn_dbufs_mtx); 903 } 904 905 static int 906 dbuf_block_freeable(dmu_buf_impl_t *db) 907 { 908 dsl_dataset_t *ds = db->db_objset->os_dsl_dataset; 909 uint64_t birth_txg = 0; 910 911 /* 912 * We don't need any locking to protect db_blkptr: 913 * If it's syncing, then db_last_dirty will be set 914 * so we'll ignore db_blkptr. 915 * 916 * This logic ensures that only block births for 917 * filled blocks are considered. 918 */ 919 ASSERT(MUTEX_HELD(&db->db_mtx)); 920 if (db->db_last_dirty && (db->db_blkptr == NULL || 921 !BP_IS_HOLE(db->db_blkptr))) { 922 birth_txg = db->db_last_dirty->dr_txg; 923 } else if (db->db_blkptr != NULL && !BP_IS_HOLE(db->db_blkptr)) { 924 birth_txg = db->db_blkptr->blk_birth; 925 } 926 927 /* 928 * If this block don't exist or is in a snapshot, it can't be freed. 929 * Don't pass the bp to dsl_dataset_block_freeable() since we 930 * are holding the db_mtx lock and might deadlock if we are 931 * prefetching a dedup-ed block. 932 */ 933 if (birth_txg != 0) 934 return (ds == NULL || 935 dsl_dataset_block_freeable(ds, NULL, birth_txg)); 936 else 937 return (B_FALSE); 938 } 939 940 void 941 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx) 942 { 943 arc_buf_t *buf, *obuf; 944 int osize = db->db.db_size; 945 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 946 dnode_t *dn; 947 948 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 949 950 DB_DNODE_ENTER(db); 951 dn = DB_DNODE(db); 952 953 /* XXX does *this* func really need the lock? */ 954 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 955 956 /* 957 * This call to dmu_buf_will_dirty() with the dn_struct_rwlock held 958 * is OK, because there can be no other references to the db 959 * when we are changing its size, so no concurrent DB_FILL can 960 * be happening. 961 */ 962 /* 963 * XXX we should be doing a dbuf_read, checking the return 964 * value and returning that up to our callers 965 */ 966 dmu_buf_will_dirty(&db->db, tx); 967 968 /* create the data buffer for the new block */ 969 buf = arc_buf_alloc(dn->dn_objset->os_spa, size, db, type); 970 971 /* copy old block data to the new block */ 972 obuf = db->db_buf; 973 bcopy(obuf->b_data, buf->b_data, MIN(osize, size)); 974 /* zero the remainder */ 975 if (size > osize) 976 bzero((uint8_t *)buf->b_data + osize, size - osize); 977 978 mutex_enter(&db->db_mtx); 979 dbuf_set_data(db, buf); 980 VERIFY(arc_buf_remove_ref(obuf, db)); 981 db->db.db_size = size; 982 983 if (db->db_level == 0) { 984 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); 985 db->db_last_dirty->dt.dl.dr_data = buf; 986 } 987 mutex_exit(&db->db_mtx); 988 989 dnode_willuse_space(dn, size-osize, tx); 990 DB_DNODE_EXIT(db); 991 } 992 993 void 994 dbuf_release_bp(dmu_buf_impl_t *db) 995 { 996 objset_t *os = db->db_objset; 997 998 ASSERT(dsl_pool_sync_context(dmu_objset_pool(os))); 999 ASSERT(arc_released(os->os_phys_buf) || 1000 list_link_active(&os->os_dsl_dataset->ds_synced_link)); 1001 ASSERT(db->db_parent == NULL || arc_released(db->db_parent->db_buf)); 1002 1003 (void) arc_release(db->db_buf, db); 1004 } 1005 1006 dbuf_dirty_record_t * 1007 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1008 { 1009 dnode_t *dn; 1010 objset_t *os; 1011 dbuf_dirty_record_t **drp, *dr; 1012 int drop_struct_lock = FALSE; 1013 boolean_t do_free_accounting = B_FALSE; 1014 int txgoff = tx->tx_txg & TXG_MASK; 1015 1016 ASSERT(tx->tx_txg != 0); 1017 ASSERT(!refcount_is_zero(&db->db_holds)); 1018 DMU_TX_DIRTY_BUF(tx, db); 1019 1020 DB_DNODE_ENTER(db); 1021 dn = DB_DNODE(db); 1022 /* 1023 * Shouldn't dirty a regular buffer in syncing context. Private 1024 * objects may be dirtied in syncing context, but only if they 1025 * were already pre-dirtied in open context. 1026 */ 1027 ASSERT(!dmu_tx_is_syncing(tx) || 1028 BP_IS_HOLE(dn->dn_objset->os_rootbp) || 1029 DMU_OBJECT_IS_SPECIAL(dn->dn_object) || 1030 dn->dn_objset->os_dsl_dataset == NULL); 1031 /* 1032 * We make this assert for private objects as well, but after we 1033 * check if we're already dirty. They are allowed to re-dirty 1034 * in syncing context. 1035 */ 1036 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 1037 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 1038 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 1039 1040 mutex_enter(&db->db_mtx); 1041 /* 1042 * XXX make this true for indirects too? The problem is that 1043 * transactions created with dmu_tx_create_assigned() from 1044 * syncing context don't bother holding ahead. 1045 */ 1046 ASSERT(db->db_level != 0 || 1047 db->db_state == DB_CACHED || db->db_state == DB_FILL || 1048 db->db_state == DB_NOFILL); 1049 1050 mutex_enter(&dn->dn_mtx); 1051 /* 1052 * Don't set dirtyctx to SYNC if we're just modifying this as we 1053 * initialize the objset. 1054 */ 1055 if (dn->dn_dirtyctx == DN_UNDIRTIED && 1056 !BP_IS_HOLE(dn->dn_objset->os_rootbp)) { 1057 dn->dn_dirtyctx = 1058 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN); 1059 ASSERT(dn->dn_dirtyctx_firstset == NULL); 1060 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP); 1061 } 1062 mutex_exit(&dn->dn_mtx); 1063 1064 if (db->db_blkid == DMU_SPILL_BLKID) 1065 dn->dn_have_spill = B_TRUE; 1066 1067 /* 1068 * If this buffer is already dirty, we're done. 1069 */ 1070 drp = &db->db_last_dirty; 1071 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg || 1072 db->db.db_object == DMU_META_DNODE_OBJECT); 1073 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg) 1074 drp = &dr->dr_next; 1075 if (dr && dr->dr_txg == tx->tx_txg) { 1076 DB_DNODE_EXIT(db); 1077 1078 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID) { 1079 /* 1080 * If this buffer has already been written out, 1081 * we now need to reset its state. 1082 */ 1083 dbuf_unoverride(dr); 1084 if (db->db.db_object != DMU_META_DNODE_OBJECT && 1085 db->db_state != DB_NOFILL) 1086 arc_buf_thaw(db->db_buf); 1087 } 1088 mutex_exit(&db->db_mtx); 1089 return (dr); 1090 } 1091 1092 /* 1093 * Only valid if not already dirty. 1094 */ 1095 ASSERT(dn->dn_object == 0 || 1096 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 1097 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 1098 1099 ASSERT3U(dn->dn_nlevels, >, db->db_level); 1100 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) || 1101 dn->dn_phys->dn_nlevels > db->db_level || 1102 dn->dn_next_nlevels[txgoff] > db->db_level || 1103 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level || 1104 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level); 1105 1106 /* 1107 * We should only be dirtying in syncing context if it's the 1108 * mos or we're initializing the os or it's a special object. 1109 * However, we are allowed to dirty in syncing context provided 1110 * we already dirtied it in open context. Hence we must make 1111 * this assertion only if we're not already dirty. 1112 */ 1113 os = dn->dn_objset; 1114 ASSERT(!dmu_tx_is_syncing(tx) || DMU_OBJECT_IS_SPECIAL(dn->dn_object) || 1115 os->os_dsl_dataset == NULL || BP_IS_HOLE(os->os_rootbp)); 1116 ASSERT(db->db.db_size != 0); 1117 1118 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 1119 1120 if (db->db_blkid != DMU_BONUS_BLKID) { 1121 /* 1122 * Update the accounting. 1123 * Note: we delay "free accounting" until after we drop 1124 * the db_mtx. This keeps us from grabbing other locks 1125 * (and possibly deadlocking) in bp_get_dsize() while 1126 * also holding the db_mtx. 1127 */ 1128 dnode_willuse_space(dn, db->db.db_size, tx); 1129 do_free_accounting = dbuf_block_freeable(db); 1130 } 1131 1132 /* 1133 * If this buffer is dirty in an old transaction group we need 1134 * to make a copy of it so that the changes we make in this 1135 * transaction group won't leak out when we sync the older txg. 1136 */ 1137 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP); 1138 if (db->db_level == 0) { 1139 void *data_old = db->db_buf; 1140 1141 if (db->db_state != DB_NOFILL) { 1142 if (db->db_blkid == DMU_BONUS_BLKID) { 1143 dbuf_fix_old_data(db, tx->tx_txg); 1144 data_old = db->db.db_data; 1145 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) { 1146 /* 1147 * Release the data buffer from the cache so 1148 * that we can modify it without impacting 1149 * possible other users of this cached data 1150 * block. Note that indirect blocks and 1151 * private objects are not released until the 1152 * syncing state (since they are only modified 1153 * then). 1154 */ 1155 arc_release(db->db_buf, db); 1156 dbuf_fix_old_data(db, tx->tx_txg); 1157 data_old = db->db_buf; 1158 } 1159 ASSERT(data_old != NULL); 1160 } 1161 dr->dt.dl.dr_data = data_old; 1162 } else { 1163 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL); 1164 list_create(&dr->dt.di.dr_children, 1165 sizeof (dbuf_dirty_record_t), 1166 offsetof(dbuf_dirty_record_t, dr_dirty_node)); 1167 } 1168 if (db->db_blkid != DMU_BONUS_BLKID && os->os_dsl_dataset != NULL) 1169 dr->dr_accounted = db->db.db_size; 1170 dr->dr_dbuf = db; 1171 dr->dr_txg = tx->tx_txg; 1172 dr->dr_next = *drp; 1173 *drp = dr; 1174 1175 /* 1176 * We could have been freed_in_flight between the dbuf_noread 1177 * and dbuf_dirty. We win, as though the dbuf_noread() had 1178 * happened after the free. 1179 */ 1180 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 1181 db->db_blkid != DMU_SPILL_BLKID) { 1182 mutex_enter(&dn->dn_mtx); 1183 if (dn->dn_free_ranges[txgoff] != NULL) { 1184 range_tree_clear(dn->dn_free_ranges[txgoff], 1185 db->db_blkid, 1); 1186 } 1187 mutex_exit(&dn->dn_mtx); 1188 db->db_freed_in_flight = FALSE; 1189 } 1190 1191 /* 1192 * This buffer is now part of this txg 1193 */ 1194 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg); 1195 db->db_dirtycnt += 1; 1196 ASSERT3U(db->db_dirtycnt, <=, 3); 1197 1198 mutex_exit(&db->db_mtx); 1199 1200 if (db->db_blkid == DMU_BONUS_BLKID || 1201 db->db_blkid == DMU_SPILL_BLKID) { 1202 mutex_enter(&dn->dn_mtx); 1203 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1204 list_insert_tail(&dn->dn_dirty_records[txgoff], dr); 1205 mutex_exit(&dn->dn_mtx); 1206 dnode_setdirty(dn, tx); 1207 DB_DNODE_EXIT(db); 1208 return (dr); 1209 } else if (do_free_accounting) { 1210 blkptr_t *bp = db->db_blkptr; 1211 int64_t willfree = (bp && !BP_IS_HOLE(bp)) ? 1212 bp_get_dsize(os->os_spa, bp) : db->db.db_size; 1213 /* 1214 * This is only a guess -- if the dbuf is dirty 1215 * in a previous txg, we don't know how much 1216 * space it will use on disk yet. We should 1217 * really have the struct_rwlock to access 1218 * db_blkptr, but since this is just a guess, 1219 * it's OK if we get an odd answer. 1220 */ 1221 ddt_prefetch(os->os_spa, bp); 1222 dnode_willuse_space(dn, -willfree, tx); 1223 } 1224 1225 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) { 1226 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1227 drop_struct_lock = TRUE; 1228 } 1229 1230 if (db->db_level == 0) { 1231 dnode_new_blkid(dn, db->db_blkid, tx, drop_struct_lock); 1232 ASSERT(dn->dn_maxblkid >= db->db_blkid); 1233 } 1234 1235 if (db->db_level+1 < dn->dn_nlevels) { 1236 dmu_buf_impl_t *parent = db->db_parent; 1237 dbuf_dirty_record_t *di; 1238 int parent_held = FALSE; 1239 1240 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) { 1241 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1242 1243 parent = dbuf_hold_level(dn, db->db_level+1, 1244 db->db_blkid >> epbs, FTAG); 1245 ASSERT(parent != NULL); 1246 parent_held = TRUE; 1247 } 1248 if (drop_struct_lock) 1249 rw_exit(&dn->dn_struct_rwlock); 1250 ASSERT3U(db->db_level+1, ==, parent->db_level); 1251 di = dbuf_dirty(parent, tx); 1252 if (parent_held) 1253 dbuf_rele(parent, FTAG); 1254 1255 mutex_enter(&db->db_mtx); 1256 /* 1257 * Since we've dropped the mutex, it's possible that 1258 * dbuf_undirty() might have changed this out from under us. 1259 */ 1260 if (db->db_last_dirty == dr || 1261 dn->dn_object == DMU_META_DNODE_OBJECT) { 1262 mutex_enter(&di->dt.di.dr_mtx); 1263 ASSERT3U(di->dr_txg, ==, tx->tx_txg); 1264 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1265 list_insert_tail(&di->dt.di.dr_children, dr); 1266 mutex_exit(&di->dt.di.dr_mtx); 1267 dr->dr_parent = di; 1268 } 1269 mutex_exit(&db->db_mtx); 1270 } else { 1271 ASSERT(db->db_level+1 == dn->dn_nlevels); 1272 ASSERT(db->db_blkid < dn->dn_nblkptr); 1273 ASSERT(db->db_parent == NULL || db->db_parent == dn->dn_dbuf); 1274 mutex_enter(&dn->dn_mtx); 1275 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1276 list_insert_tail(&dn->dn_dirty_records[txgoff], dr); 1277 mutex_exit(&dn->dn_mtx); 1278 if (drop_struct_lock) 1279 rw_exit(&dn->dn_struct_rwlock); 1280 } 1281 1282 dnode_setdirty(dn, tx); 1283 DB_DNODE_EXIT(db); 1284 return (dr); 1285 } 1286 1287 /* 1288 * Undirty a buffer in the transaction group referenced by the given 1289 * transaction. Return whether this evicted the dbuf. 1290 */ 1291 static boolean_t 1292 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1293 { 1294 dnode_t *dn; 1295 uint64_t txg = tx->tx_txg; 1296 dbuf_dirty_record_t *dr, **drp; 1297 1298 ASSERT(txg != 0); 1299 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1300 ASSERT0(db->db_level); 1301 ASSERT(MUTEX_HELD(&db->db_mtx)); 1302 1303 /* 1304 * If this buffer is not dirty, we're done. 1305 */ 1306 for (drp = &db->db_last_dirty; (dr = *drp) != NULL; drp = &dr->dr_next) 1307 if (dr->dr_txg <= txg) 1308 break; 1309 if (dr == NULL || dr->dr_txg < txg) 1310 return (B_FALSE); 1311 ASSERT(dr->dr_txg == txg); 1312 ASSERT(dr->dr_dbuf == db); 1313 1314 DB_DNODE_ENTER(db); 1315 dn = DB_DNODE(db); 1316 1317 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 1318 1319 ASSERT(db->db.db_size != 0); 1320 1321 /* 1322 * Any space we accounted for in dp_dirty_* will be cleaned up by 1323 * dsl_pool_sync(). This is relatively rare so the discrepancy 1324 * is not a big deal. 1325 */ 1326 1327 *drp = dr->dr_next; 1328 1329 /* 1330 * Note that there are three places in dbuf_dirty() 1331 * where this dirty record may be put on a list. 1332 * Make sure to do a list_remove corresponding to 1333 * every one of those list_insert calls. 1334 */ 1335 if (dr->dr_parent) { 1336 mutex_enter(&dr->dr_parent->dt.di.dr_mtx); 1337 list_remove(&dr->dr_parent->dt.di.dr_children, dr); 1338 mutex_exit(&dr->dr_parent->dt.di.dr_mtx); 1339 } else if (db->db_blkid == DMU_SPILL_BLKID || 1340 db->db_level+1 == dn->dn_nlevels) { 1341 ASSERT(db->db_blkptr == NULL || db->db_parent == dn->dn_dbuf); 1342 mutex_enter(&dn->dn_mtx); 1343 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr); 1344 mutex_exit(&dn->dn_mtx); 1345 } 1346 DB_DNODE_EXIT(db); 1347 1348 if (db->db_state != DB_NOFILL) { 1349 dbuf_unoverride(dr); 1350 1351 ASSERT(db->db_buf != NULL); 1352 ASSERT(dr->dt.dl.dr_data != NULL); 1353 if (dr->dt.dl.dr_data != db->db_buf) 1354 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db)); 1355 } 1356 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 1357 1358 ASSERT(db->db_dirtycnt > 0); 1359 db->db_dirtycnt -= 1; 1360 1361 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) { 1362 arc_buf_t *buf = db->db_buf; 1363 1364 ASSERT(db->db_state == DB_NOFILL || arc_released(buf)); 1365 dbuf_set_data(db, NULL); 1366 VERIFY(arc_buf_remove_ref(buf, db)); 1367 dbuf_evict(db); 1368 return (B_TRUE); 1369 } 1370 1371 return (B_FALSE); 1372 } 1373 1374 void 1375 dmu_buf_will_dirty(dmu_buf_t *db_fake, dmu_tx_t *tx) 1376 { 1377 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1378 int rf = DB_RF_MUST_SUCCEED | DB_RF_NOPREFETCH; 1379 1380 ASSERT(tx->tx_txg != 0); 1381 ASSERT(!refcount_is_zero(&db->db_holds)); 1382 1383 DB_DNODE_ENTER(db); 1384 if (RW_WRITE_HELD(&DB_DNODE(db)->dn_struct_rwlock)) 1385 rf |= DB_RF_HAVESTRUCT; 1386 DB_DNODE_EXIT(db); 1387 (void) dbuf_read(db, NULL, rf); 1388 (void) dbuf_dirty(db, tx); 1389 } 1390 1391 void 1392 dmu_buf_will_not_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) 1393 { 1394 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1395 1396 db->db_state = DB_NOFILL; 1397 1398 dmu_buf_will_fill(db_fake, tx); 1399 } 1400 1401 void 1402 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) 1403 { 1404 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1405 1406 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1407 ASSERT(tx->tx_txg != 0); 1408 ASSERT(db->db_level == 0); 1409 ASSERT(!refcount_is_zero(&db->db_holds)); 1410 1411 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT || 1412 dmu_tx_private_ok(tx)); 1413 1414 dbuf_noread(db); 1415 (void) dbuf_dirty(db, tx); 1416 } 1417 1418 #pragma weak dmu_buf_fill_done = dbuf_fill_done 1419 /* ARGSUSED */ 1420 void 1421 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx) 1422 { 1423 mutex_enter(&db->db_mtx); 1424 DBUF_VERIFY(db); 1425 1426 if (db->db_state == DB_FILL) { 1427 if (db->db_level == 0 && db->db_freed_in_flight) { 1428 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1429 /* we were freed while filling */ 1430 /* XXX dbuf_undirty? */ 1431 bzero(db->db.db_data, db->db.db_size); 1432 db->db_freed_in_flight = FALSE; 1433 } 1434 db->db_state = DB_CACHED; 1435 cv_broadcast(&db->db_changed); 1436 } 1437 mutex_exit(&db->db_mtx); 1438 } 1439 1440 void 1441 dmu_buf_write_embedded(dmu_buf_t *dbuf, void *data, 1442 bp_embedded_type_t etype, enum zio_compress comp, 1443 int uncompressed_size, int compressed_size, int byteorder, 1444 dmu_tx_t *tx) 1445 { 1446 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; 1447 struct dirty_leaf *dl; 1448 dmu_object_type_t type; 1449 1450 DB_DNODE_ENTER(db); 1451 type = DB_DNODE(db)->dn_type; 1452 DB_DNODE_EXIT(db); 1453 1454 ASSERT0(db->db_level); 1455 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1456 1457 dmu_buf_will_not_fill(dbuf, tx); 1458 1459 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); 1460 dl = &db->db_last_dirty->dt.dl; 1461 encode_embedded_bp_compressed(&dl->dr_overridden_by, 1462 data, comp, uncompressed_size, compressed_size); 1463 BPE_SET_ETYPE(&dl->dr_overridden_by, etype); 1464 BP_SET_TYPE(&dl->dr_overridden_by, type); 1465 BP_SET_LEVEL(&dl->dr_overridden_by, 0); 1466 BP_SET_BYTEORDER(&dl->dr_overridden_by, byteorder); 1467 1468 dl->dr_override_state = DR_OVERRIDDEN; 1469 dl->dr_overridden_by.blk_birth = db->db_last_dirty->dr_txg; 1470 } 1471 1472 /* 1473 * Directly assign a provided arc buf to a given dbuf if it's not referenced 1474 * by anybody except our caller. Otherwise copy arcbuf's contents to dbuf. 1475 */ 1476 void 1477 dbuf_assign_arcbuf(dmu_buf_impl_t *db, arc_buf_t *buf, dmu_tx_t *tx) 1478 { 1479 ASSERT(!refcount_is_zero(&db->db_holds)); 1480 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 1481 ASSERT(db->db_level == 0); 1482 ASSERT(DBUF_GET_BUFC_TYPE(db) == ARC_BUFC_DATA); 1483 ASSERT(buf != NULL); 1484 ASSERT(arc_buf_size(buf) == db->db.db_size); 1485 ASSERT(tx->tx_txg != 0); 1486 1487 arc_return_buf(buf, db); 1488 ASSERT(arc_released(buf)); 1489 1490 mutex_enter(&db->db_mtx); 1491 1492 while (db->db_state == DB_READ || db->db_state == DB_FILL) 1493 cv_wait(&db->db_changed, &db->db_mtx); 1494 1495 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_UNCACHED); 1496 1497 if (db->db_state == DB_CACHED && 1498 refcount_count(&db->db_holds) - 1 > db->db_dirtycnt) { 1499 mutex_exit(&db->db_mtx); 1500 (void) dbuf_dirty(db, tx); 1501 bcopy(buf->b_data, db->db.db_data, db->db.db_size); 1502 VERIFY(arc_buf_remove_ref(buf, db)); 1503 xuio_stat_wbuf_copied(); 1504 return; 1505 } 1506 1507 xuio_stat_wbuf_nocopy(); 1508 if (db->db_state == DB_CACHED) { 1509 dbuf_dirty_record_t *dr = db->db_last_dirty; 1510 1511 ASSERT(db->db_buf != NULL); 1512 if (dr != NULL && dr->dr_txg == tx->tx_txg) { 1513 ASSERT(dr->dt.dl.dr_data == db->db_buf); 1514 if (!arc_released(db->db_buf)) { 1515 ASSERT(dr->dt.dl.dr_override_state == 1516 DR_OVERRIDDEN); 1517 arc_release(db->db_buf, db); 1518 } 1519 dr->dt.dl.dr_data = buf; 1520 VERIFY(arc_buf_remove_ref(db->db_buf, db)); 1521 } else if (dr == NULL || dr->dt.dl.dr_data != db->db_buf) { 1522 arc_release(db->db_buf, db); 1523 VERIFY(arc_buf_remove_ref(db->db_buf, db)); 1524 } 1525 db->db_buf = NULL; 1526 } 1527 ASSERT(db->db_buf == NULL); 1528 dbuf_set_data(db, buf); 1529 db->db_state = DB_FILL; 1530 mutex_exit(&db->db_mtx); 1531 (void) dbuf_dirty(db, tx); 1532 dmu_buf_fill_done(&db->db, tx); 1533 } 1534 1535 /* 1536 * "Clear" the contents of this dbuf. This will mark the dbuf 1537 * EVICTING and clear *most* of its references. Unfortunately, 1538 * when we are not holding the dn_dbufs_mtx, we can't clear the 1539 * entry in the dn_dbufs list. We have to wait until dbuf_destroy() 1540 * in this case. For callers from the DMU we will usually see: 1541 * dbuf_clear()->arc_buf_evict()->dbuf_do_evict()->dbuf_destroy() 1542 * For the arc callback, we will usually see: 1543 * dbuf_do_evict()->dbuf_clear();dbuf_destroy() 1544 * Sometimes, though, we will get a mix of these two: 1545 * DMU: dbuf_clear()->arc_buf_evict() 1546 * ARC: dbuf_do_evict()->dbuf_destroy() 1547 */ 1548 void 1549 dbuf_clear(dmu_buf_impl_t *db) 1550 { 1551 dnode_t *dn; 1552 dmu_buf_impl_t *parent = db->db_parent; 1553 dmu_buf_impl_t *dndb; 1554 int dbuf_gone = FALSE; 1555 1556 ASSERT(MUTEX_HELD(&db->db_mtx)); 1557 ASSERT(refcount_is_zero(&db->db_holds)); 1558 1559 dbuf_evict_user(db); 1560 1561 if (db->db_state == DB_CACHED) { 1562 ASSERT(db->db.db_data != NULL); 1563 if (db->db_blkid == DMU_BONUS_BLKID) { 1564 zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN); 1565 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 1566 } 1567 db->db.db_data = NULL; 1568 db->db_state = DB_UNCACHED; 1569 } 1570 1571 ASSERT(db->db_state == DB_UNCACHED || db->db_state == DB_NOFILL); 1572 ASSERT(db->db_data_pending == NULL); 1573 1574 db->db_state = DB_EVICTING; 1575 db->db_blkptr = NULL; 1576 1577 DB_DNODE_ENTER(db); 1578 dn = DB_DNODE(db); 1579 dndb = dn->dn_dbuf; 1580 if (db->db_blkid != DMU_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) { 1581 list_remove(&dn->dn_dbufs, db); 1582 (void) atomic_dec_32_nv(&dn->dn_dbufs_count); 1583 membar_producer(); 1584 DB_DNODE_EXIT(db); 1585 /* 1586 * Decrementing the dbuf count means that the hold corresponding 1587 * to the removed dbuf is no longer discounted in dnode_move(), 1588 * so the dnode cannot be moved until after we release the hold. 1589 * The membar_producer() ensures visibility of the decremented 1590 * value in dnode_move(), since DB_DNODE_EXIT doesn't actually 1591 * release any lock. 1592 */ 1593 dnode_rele(dn, db); 1594 db->db_dnode_handle = NULL; 1595 } else { 1596 DB_DNODE_EXIT(db); 1597 } 1598 1599 if (db->db_buf) 1600 dbuf_gone = arc_buf_evict(db->db_buf); 1601 1602 if (!dbuf_gone) 1603 mutex_exit(&db->db_mtx); 1604 1605 /* 1606 * If this dbuf is referenced from an indirect dbuf, 1607 * decrement the ref count on the indirect dbuf. 1608 */ 1609 if (parent && parent != dndb) 1610 dbuf_rele(parent, db); 1611 } 1612 1613 static int 1614 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse, 1615 dmu_buf_impl_t **parentp, blkptr_t **bpp) 1616 { 1617 int nlevels, epbs; 1618 1619 *parentp = NULL; 1620 *bpp = NULL; 1621 1622 ASSERT(blkid != DMU_BONUS_BLKID); 1623 1624 if (blkid == DMU_SPILL_BLKID) { 1625 mutex_enter(&dn->dn_mtx); 1626 if (dn->dn_have_spill && 1627 (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR)) 1628 *bpp = &dn->dn_phys->dn_spill; 1629 else 1630 *bpp = NULL; 1631 dbuf_add_ref(dn->dn_dbuf, NULL); 1632 *parentp = dn->dn_dbuf; 1633 mutex_exit(&dn->dn_mtx); 1634 return (0); 1635 } 1636 1637 if (dn->dn_phys->dn_nlevels == 0) 1638 nlevels = 1; 1639 else 1640 nlevels = dn->dn_phys->dn_nlevels; 1641 1642 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1643 1644 ASSERT3U(level * epbs, <, 64); 1645 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1646 if (level >= nlevels || 1647 (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) { 1648 /* the buffer has no parent yet */ 1649 return (SET_ERROR(ENOENT)); 1650 } else if (level < nlevels-1) { 1651 /* this block is referenced from an indirect block */ 1652 int err = dbuf_hold_impl(dn, level+1, 1653 blkid >> epbs, fail_sparse, NULL, parentp); 1654 if (err) 1655 return (err); 1656 err = dbuf_read(*parentp, NULL, 1657 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL)); 1658 if (err) { 1659 dbuf_rele(*parentp, NULL); 1660 *parentp = NULL; 1661 return (err); 1662 } 1663 *bpp = ((blkptr_t *)(*parentp)->db.db_data) + 1664 (blkid & ((1ULL << epbs) - 1)); 1665 return (0); 1666 } else { 1667 /* the block is referenced from the dnode */ 1668 ASSERT3U(level, ==, nlevels-1); 1669 ASSERT(dn->dn_phys->dn_nblkptr == 0 || 1670 blkid < dn->dn_phys->dn_nblkptr); 1671 if (dn->dn_dbuf) { 1672 dbuf_add_ref(dn->dn_dbuf, NULL); 1673 *parentp = dn->dn_dbuf; 1674 } 1675 *bpp = &dn->dn_phys->dn_blkptr[blkid]; 1676 return (0); 1677 } 1678 } 1679 1680 static dmu_buf_impl_t * 1681 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid, 1682 dmu_buf_impl_t *parent, blkptr_t *blkptr) 1683 { 1684 objset_t *os = dn->dn_objset; 1685 dmu_buf_impl_t *db, *odb; 1686 1687 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1688 ASSERT(dn->dn_type != DMU_OT_NONE); 1689 1690 db = kmem_cache_alloc(dbuf_cache, KM_SLEEP); 1691 1692 db->db_objset = os; 1693 db->db.db_object = dn->dn_object; 1694 db->db_level = level; 1695 db->db_blkid = blkid; 1696 db->db_last_dirty = NULL; 1697 db->db_dirtycnt = 0; 1698 db->db_dnode_handle = dn->dn_handle; 1699 db->db_parent = parent; 1700 db->db_blkptr = blkptr; 1701 1702 db->db_user_ptr = NULL; 1703 db->db_user_data_ptr_ptr = NULL; 1704 db->db_evict_func = NULL; 1705 db->db_immediate_evict = 0; 1706 db->db_freed_in_flight = 0; 1707 1708 if (blkid == DMU_BONUS_BLKID) { 1709 ASSERT3P(parent, ==, dn->dn_dbuf); 1710 db->db.db_size = DN_MAX_BONUSLEN - 1711 (dn->dn_nblkptr-1) * sizeof (blkptr_t); 1712 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 1713 db->db.db_offset = DMU_BONUS_BLKID; 1714 db->db_state = DB_UNCACHED; 1715 /* the bonus dbuf is not placed in the hash table */ 1716 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 1717 return (db); 1718 } else if (blkid == DMU_SPILL_BLKID) { 1719 db->db.db_size = (blkptr != NULL) ? 1720 BP_GET_LSIZE(blkptr) : SPA_MINBLOCKSIZE; 1721 db->db.db_offset = 0; 1722 } else { 1723 int blocksize = 1724 db->db_level ? 1 << dn->dn_indblkshift : dn->dn_datablksz; 1725 db->db.db_size = blocksize; 1726 db->db.db_offset = db->db_blkid * blocksize; 1727 } 1728 1729 /* 1730 * Hold the dn_dbufs_mtx while we get the new dbuf 1731 * in the hash table *and* added to the dbufs list. 1732 * This prevents a possible deadlock with someone 1733 * trying to look up this dbuf before its added to the 1734 * dn_dbufs list. 1735 */ 1736 mutex_enter(&dn->dn_dbufs_mtx); 1737 db->db_state = DB_EVICTING; 1738 if ((odb = dbuf_hash_insert(db)) != NULL) { 1739 /* someone else inserted it first */ 1740 kmem_cache_free(dbuf_cache, db); 1741 mutex_exit(&dn->dn_dbufs_mtx); 1742 return (odb); 1743 } 1744 list_insert_head(&dn->dn_dbufs, db); 1745 if (db->db_level == 0 && db->db_blkid >= 1746 dn->dn_unlisted_l0_blkid) 1747 dn->dn_unlisted_l0_blkid = db->db_blkid + 1; 1748 db->db_state = DB_UNCACHED; 1749 mutex_exit(&dn->dn_dbufs_mtx); 1750 arc_space_consume(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 1751 1752 if (parent && parent != dn->dn_dbuf) 1753 dbuf_add_ref(parent, db); 1754 1755 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 1756 refcount_count(&dn->dn_holds) > 0); 1757 (void) refcount_add(&dn->dn_holds, db); 1758 (void) atomic_inc_32_nv(&dn->dn_dbufs_count); 1759 1760 dprintf_dbuf(db, "db=%p\n", db); 1761 1762 return (db); 1763 } 1764 1765 static int 1766 dbuf_do_evict(void *private) 1767 { 1768 arc_buf_t *buf = private; 1769 dmu_buf_impl_t *db = buf->b_private; 1770 1771 if (!MUTEX_HELD(&db->db_mtx)) 1772 mutex_enter(&db->db_mtx); 1773 1774 ASSERT(refcount_is_zero(&db->db_holds)); 1775 1776 if (db->db_state != DB_EVICTING) { 1777 ASSERT(db->db_state == DB_CACHED); 1778 DBUF_VERIFY(db); 1779 db->db_buf = NULL; 1780 dbuf_evict(db); 1781 } else { 1782 mutex_exit(&db->db_mtx); 1783 dbuf_destroy(db); 1784 } 1785 return (0); 1786 } 1787 1788 static void 1789 dbuf_destroy(dmu_buf_impl_t *db) 1790 { 1791 ASSERT(refcount_is_zero(&db->db_holds)); 1792 1793 if (db->db_blkid != DMU_BONUS_BLKID) { 1794 /* 1795 * If this dbuf is still on the dn_dbufs list, 1796 * remove it from that list. 1797 */ 1798 if (db->db_dnode_handle != NULL) { 1799 dnode_t *dn; 1800 1801 DB_DNODE_ENTER(db); 1802 dn = DB_DNODE(db); 1803 mutex_enter(&dn->dn_dbufs_mtx); 1804 list_remove(&dn->dn_dbufs, db); 1805 (void) atomic_dec_32_nv(&dn->dn_dbufs_count); 1806 mutex_exit(&dn->dn_dbufs_mtx); 1807 DB_DNODE_EXIT(db); 1808 /* 1809 * Decrementing the dbuf count means that the hold 1810 * corresponding to the removed dbuf is no longer 1811 * discounted in dnode_move(), so the dnode cannot be 1812 * moved until after we release the hold. 1813 */ 1814 dnode_rele(dn, db); 1815 db->db_dnode_handle = NULL; 1816 } 1817 dbuf_hash_remove(db); 1818 } 1819 db->db_parent = NULL; 1820 db->db_buf = NULL; 1821 1822 ASSERT(!list_link_active(&db->db_link)); 1823 ASSERT(db->db.db_data == NULL); 1824 ASSERT(db->db_hash_next == NULL); 1825 ASSERT(db->db_blkptr == NULL); 1826 ASSERT(db->db_data_pending == NULL); 1827 1828 kmem_cache_free(dbuf_cache, db); 1829 arc_space_return(sizeof (dmu_buf_impl_t), ARC_SPACE_OTHER); 1830 } 1831 1832 void 1833 dbuf_prefetch(dnode_t *dn, uint64_t blkid, zio_priority_t prio) 1834 { 1835 dmu_buf_impl_t *db = NULL; 1836 blkptr_t *bp = NULL; 1837 1838 ASSERT(blkid != DMU_BONUS_BLKID); 1839 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1840 1841 if (dnode_block_freed(dn, blkid)) 1842 return; 1843 1844 /* dbuf_find() returns with db_mtx held */ 1845 if (db = dbuf_find(dn, 0, blkid)) { 1846 /* 1847 * This dbuf is already in the cache. We assume that 1848 * it is already CACHED, or else about to be either 1849 * read or filled. 1850 */ 1851 mutex_exit(&db->db_mtx); 1852 return; 1853 } 1854 1855 if (dbuf_findbp(dn, 0, blkid, TRUE, &db, &bp) == 0) { 1856 if (bp && !BP_IS_HOLE(bp) && !BP_IS_EMBEDDED(bp)) { 1857 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset; 1858 uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH; 1859 zbookmark_phys_t zb; 1860 1861 SET_BOOKMARK(&zb, ds ? ds->ds_object : DMU_META_OBJSET, 1862 dn->dn_object, 0, blkid); 1863 1864 (void) arc_read(NULL, dn->dn_objset->os_spa, 1865 bp, NULL, NULL, prio, 1866 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 1867 &aflags, &zb); 1868 } 1869 if (db) 1870 dbuf_rele(db, NULL); 1871 } 1872 } 1873 1874 /* 1875 * Returns with db_holds incremented, and db_mtx not held. 1876 * Note: dn_struct_rwlock must be held. 1877 */ 1878 int 1879 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse, 1880 void *tag, dmu_buf_impl_t **dbp) 1881 { 1882 dmu_buf_impl_t *db, *parent = NULL; 1883 1884 ASSERT(blkid != DMU_BONUS_BLKID); 1885 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1886 ASSERT3U(dn->dn_nlevels, >, level); 1887 1888 *dbp = NULL; 1889 top: 1890 /* dbuf_find() returns with db_mtx held */ 1891 db = dbuf_find(dn, level, blkid); 1892 1893 if (db == NULL) { 1894 blkptr_t *bp = NULL; 1895 int err; 1896 1897 ASSERT3P(parent, ==, NULL); 1898 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp); 1899 if (fail_sparse) { 1900 if (err == 0 && bp && BP_IS_HOLE(bp)) 1901 err = SET_ERROR(ENOENT); 1902 if (err) { 1903 if (parent) 1904 dbuf_rele(parent, NULL); 1905 return (err); 1906 } 1907 } 1908 if (err && err != ENOENT) 1909 return (err); 1910 db = dbuf_create(dn, level, blkid, parent, bp); 1911 } 1912 1913 if (db->db_buf && refcount_is_zero(&db->db_holds)) { 1914 arc_buf_add_ref(db->db_buf, db); 1915 if (db->db_buf->b_data == NULL) { 1916 dbuf_clear(db); 1917 if (parent) { 1918 dbuf_rele(parent, NULL); 1919 parent = NULL; 1920 } 1921 goto top; 1922 } 1923 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data); 1924 } 1925 1926 ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf)); 1927 1928 /* 1929 * If this buffer is currently syncing out, and we are are 1930 * still referencing it from db_data, we need to make a copy 1931 * of it in case we decide we want to dirty it again in this txg. 1932 */ 1933 if (db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID && 1934 dn->dn_object != DMU_META_DNODE_OBJECT && 1935 db->db_state == DB_CACHED && db->db_data_pending) { 1936 dbuf_dirty_record_t *dr = db->db_data_pending; 1937 1938 if (dr->dt.dl.dr_data == db->db_buf) { 1939 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 1940 1941 dbuf_set_data(db, 1942 arc_buf_alloc(dn->dn_objset->os_spa, 1943 db->db.db_size, db, type)); 1944 bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data, 1945 db->db.db_size); 1946 } 1947 } 1948 1949 (void) refcount_add(&db->db_holds, tag); 1950 dbuf_update_data(db); 1951 DBUF_VERIFY(db); 1952 mutex_exit(&db->db_mtx); 1953 1954 /* NOTE: we can't rele the parent until after we drop the db_mtx */ 1955 if (parent) 1956 dbuf_rele(parent, NULL); 1957 1958 ASSERT3P(DB_DNODE(db), ==, dn); 1959 ASSERT3U(db->db_blkid, ==, blkid); 1960 ASSERT3U(db->db_level, ==, level); 1961 *dbp = db; 1962 1963 return (0); 1964 } 1965 1966 dmu_buf_impl_t * 1967 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag) 1968 { 1969 dmu_buf_impl_t *db; 1970 int err = dbuf_hold_impl(dn, 0, blkid, FALSE, tag, &db); 1971 return (err ? NULL : db); 1972 } 1973 1974 dmu_buf_impl_t * 1975 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag) 1976 { 1977 dmu_buf_impl_t *db; 1978 int err = dbuf_hold_impl(dn, level, blkid, FALSE, tag, &db); 1979 return (err ? NULL : db); 1980 } 1981 1982 void 1983 dbuf_create_bonus(dnode_t *dn) 1984 { 1985 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 1986 1987 ASSERT(dn->dn_bonus == NULL); 1988 dn->dn_bonus = dbuf_create(dn, 0, DMU_BONUS_BLKID, dn->dn_dbuf, NULL); 1989 } 1990 1991 int 1992 dbuf_spill_set_blksz(dmu_buf_t *db_fake, uint64_t blksz, dmu_tx_t *tx) 1993 { 1994 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1995 dnode_t *dn; 1996 1997 if (db->db_blkid != DMU_SPILL_BLKID) 1998 return (SET_ERROR(ENOTSUP)); 1999 if (blksz == 0) 2000 blksz = SPA_MINBLOCKSIZE; 2001 if (blksz > SPA_MAXBLOCKSIZE) 2002 blksz = SPA_MAXBLOCKSIZE; 2003 else 2004 blksz = P2ROUNDUP(blksz, SPA_MINBLOCKSIZE); 2005 2006 DB_DNODE_ENTER(db); 2007 dn = DB_DNODE(db); 2008 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 2009 dbuf_new_size(db, blksz, tx); 2010 rw_exit(&dn->dn_struct_rwlock); 2011 DB_DNODE_EXIT(db); 2012 2013 return (0); 2014 } 2015 2016 void 2017 dbuf_rm_spill(dnode_t *dn, dmu_tx_t *tx) 2018 { 2019 dbuf_free_range(dn, DMU_SPILL_BLKID, DMU_SPILL_BLKID, tx); 2020 } 2021 2022 #pragma weak dmu_buf_add_ref = dbuf_add_ref 2023 void 2024 dbuf_add_ref(dmu_buf_impl_t *db, void *tag) 2025 { 2026 int64_t holds = refcount_add(&db->db_holds, tag); 2027 ASSERT(holds > 1); 2028 } 2029 2030 /* 2031 * If you call dbuf_rele() you had better not be referencing the dnode handle 2032 * unless you have some other direct or indirect hold on the dnode. (An indirect 2033 * hold is a hold on one of the dnode's dbufs, including the bonus buffer.) 2034 * Without that, the dbuf_rele() could lead to a dnode_rele() followed by the 2035 * dnode's parent dbuf evicting its dnode handles. 2036 */ 2037 void 2038 dbuf_rele(dmu_buf_impl_t *db, void *tag) 2039 { 2040 mutex_enter(&db->db_mtx); 2041 dbuf_rele_and_unlock(db, tag); 2042 } 2043 2044 void 2045 dmu_buf_rele(dmu_buf_t *db, void *tag) 2046 { 2047 dbuf_rele((dmu_buf_impl_t *)db, tag); 2048 } 2049 2050 /* 2051 * dbuf_rele() for an already-locked dbuf. This is necessary to allow 2052 * db_dirtycnt and db_holds to be updated atomically. 2053 */ 2054 void 2055 dbuf_rele_and_unlock(dmu_buf_impl_t *db, void *tag) 2056 { 2057 int64_t holds; 2058 2059 ASSERT(MUTEX_HELD(&db->db_mtx)); 2060 DBUF_VERIFY(db); 2061 2062 /* 2063 * Remove the reference to the dbuf before removing its hold on the 2064 * dnode so we can guarantee in dnode_move() that a referenced bonus 2065 * buffer has a corresponding dnode hold. 2066 */ 2067 holds = refcount_remove(&db->db_holds, tag); 2068 ASSERT(holds >= 0); 2069 2070 /* 2071 * We can't freeze indirects if there is a possibility that they 2072 * may be modified in the current syncing context. 2073 */ 2074 if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) 2075 arc_buf_freeze(db->db_buf); 2076 2077 if (holds == db->db_dirtycnt && 2078 db->db_level == 0 && db->db_immediate_evict) 2079 dbuf_evict_user(db); 2080 2081 if (holds == 0) { 2082 if (db->db_blkid == DMU_BONUS_BLKID) { 2083 mutex_exit(&db->db_mtx); 2084 2085 /* 2086 * If the dnode moves here, we cannot cross this barrier 2087 * until the move completes. 2088 */ 2089 DB_DNODE_ENTER(db); 2090 (void) atomic_dec_32_nv(&DB_DNODE(db)->dn_dbufs_count); 2091 DB_DNODE_EXIT(db); 2092 /* 2093 * The bonus buffer's dnode hold is no longer discounted 2094 * in dnode_move(). The dnode cannot move until after 2095 * the dnode_rele(). 2096 */ 2097 dnode_rele(DB_DNODE(db), db); 2098 } else if (db->db_buf == NULL) { 2099 /* 2100 * This is a special case: we never associated this 2101 * dbuf with any data allocated from the ARC. 2102 */ 2103 ASSERT(db->db_state == DB_UNCACHED || 2104 db->db_state == DB_NOFILL); 2105 dbuf_evict(db); 2106 } else if (arc_released(db->db_buf)) { 2107 arc_buf_t *buf = db->db_buf; 2108 /* 2109 * This dbuf has anonymous data associated with it. 2110 */ 2111 dbuf_set_data(db, NULL); 2112 VERIFY(arc_buf_remove_ref(buf, db)); 2113 dbuf_evict(db); 2114 } else { 2115 VERIFY(!arc_buf_remove_ref(db->db_buf, db)); 2116 2117 /* 2118 * A dbuf will be eligible for eviction if either the 2119 * 'primarycache' property is set or a duplicate 2120 * copy of this buffer is already cached in the arc. 2121 * 2122 * In the case of the 'primarycache' a buffer 2123 * is considered for eviction if it matches the 2124 * criteria set in the property. 2125 * 2126 * To decide if our buffer is considered a 2127 * duplicate, we must call into the arc to determine 2128 * if multiple buffers are referencing the same 2129 * block on-disk. If so, then we simply evict 2130 * ourselves. 2131 */ 2132 if (!DBUF_IS_CACHEABLE(db) || 2133 arc_buf_eviction_needed(db->db_buf)) 2134 dbuf_clear(db); 2135 else 2136 mutex_exit(&db->db_mtx); 2137 } 2138 } else { 2139 mutex_exit(&db->db_mtx); 2140 } 2141 } 2142 2143 #pragma weak dmu_buf_refcount = dbuf_refcount 2144 uint64_t 2145 dbuf_refcount(dmu_buf_impl_t *db) 2146 { 2147 return (refcount_count(&db->db_holds)); 2148 } 2149 2150 void * 2151 dmu_buf_set_user(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr, 2152 dmu_buf_evict_func_t *evict_func) 2153 { 2154 return (dmu_buf_update_user(db_fake, NULL, user_ptr, 2155 user_data_ptr_ptr, evict_func)); 2156 } 2157 2158 void * 2159 dmu_buf_set_user_ie(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr, 2160 dmu_buf_evict_func_t *evict_func) 2161 { 2162 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2163 2164 db->db_immediate_evict = TRUE; 2165 return (dmu_buf_update_user(db_fake, NULL, user_ptr, 2166 user_data_ptr_ptr, evict_func)); 2167 } 2168 2169 void * 2170 dmu_buf_update_user(dmu_buf_t *db_fake, void *old_user_ptr, void *user_ptr, 2171 void *user_data_ptr_ptr, dmu_buf_evict_func_t *evict_func) 2172 { 2173 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2174 ASSERT(db->db_level == 0); 2175 2176 ASSERT((user_ptr == NULL) == (evict_func == NULL)); 2177 2178 mutex_enter(&db->db_mtx); 2179 2180 if (db->db_user_ptr == old_user_ptr) { 2181 db->db_user_ptr = user_ptr; 2182 db->db_user_data_ptr_ptr = user_data_ptr_ptr; 2183 db->db_evict_func = evict_func; 2184 2185 dbuf_update_data(db); 2186 } else { 2187 old_user_ptr = db->db_user_ptr; 2188 } 2189 2190 mutex_exit(&db->db_mtx); 2191 return (old_user_ptr); 2192 } 2193 2194 void * 2195 dmu_buf_get_user(dmu_buf_t *db_fake) 2196 { 2197 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 2198 ASSERT(!refcount_is_zero(&db->db_holds)); 2199 2200 return (db->db_user_ptr); 2201 } 2202 2203 boolean_t 2204 dmu_buf_freeable(dmu_buf_t *dbuf) 2205 { 2206 boolean_t res = B_FALSE; 2207 dmu_buf_impl_t *db = (dmu_buf_impl_t *)dbuf; 2208 2209 if (db->db_blkptr) 2210 res = dsl_dataset_block_freeable(db->db_objset->os_dsl_dataset, 2211 db->db_blkptr, db->db_blkptr->blk_birth); 2212 2213 return (res); 2214 } 2215 2216 blkptr_t * 2217 dmu_buf_get_blkptr(dmu_buf_t *db) 2218 { 2219 dmu_buf_impl_t *dbi = (dmu_buf_impl_t *)db; 2220 return (dbi->db_blkptr); 2221 } 2222 2223 static void 2224 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db) 2225 { 2226 /* ASSERT(dmu_tx_is_syncing(tx) */ 2227 ASSERT(MUTEX_HELD(&db->db_mtx)); 2228 2229 if (db->db_blkptr != NULL) 2230 return; 2231 2232 if (db->db_blkid == DMU_SPILL_BLKID) { 2233 db->db_blkptr = &dn->dn_phys->dn_spill; 2234 BP_ZERO(db->db_blkptr); 2235 return; 2236 } 2237 if (db->db_level == dn->dn_phys->dn_nlevels-1) { 2238 /* 2239 * This buffer was allocated at a time when there was 2240 * no available blkptrs from the dnode, or it was 2241 * inappropriate to hook it in (i.e., nlevels mis-match). 2242 */ 2243 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr); 2244 ASSERT(db->db_parent == NULL); 2245 db->db_parent = dn->dn_dbuf; 2246 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid]; 2247 DBUF_VERIFY(db); 2248 } else { 2249 dmu_buf_impl_t *parent = db->db_parent; 2250 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 2251 2252 ASSERT(dn->dn_phys->dn_nlevels > 1); 2253 if (parent == NULL) { 2254 mutex_exit(&db->db_mtx); 2255 rw_enter(&dn->dn_struct_rwlock, RW_READER); 2256 (void) dbuf_hold_impl(dn, db->db_level+1, 2257 db->db_blkid >> epbs, FALSE, db, &parent); 2258 rw_exit(&dn->dn_struct_rwlock); 2259 mutex_enter(&db->db_mtx); 2260 db->db_parent = parent; 2261 } 2262 db->db_blkptr = (blkptr_t *)parent->db.db_data + 2263 (db->db_blkid & ((1ULL << epbs) - 1)); 2264 DBUF_VERIFY(db); 2265 } 2266 } 2267 2268 static void 2269 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx) 2270 { 2271 dmu_buf_impl_t *db = dr->dr_dbuf; 2272 dnode_t *dn; 2273 zio_t *zio; 2274 2275 ASSERT(dmu_tx_is_syncing(tx)); 2276 2277 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 2278 2279 mutex_enter(&db->db_mtx); 2280 2281 ASSERT(db->db_level > 0); 2282 DBUF_VERIFY(db); 2283 2284 /* Read the block if it hasn't been read yet. */ 2285 if (db->db_buf == NULL) { 2286 mutex_exit(&db->db_mtx); 2287 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED); 2288 mutex_enter(&db->db_mtx); 2289 } 2290 ASSERT3U(db->db_state, ==, DB_CACHED); 2291 ASSERT(db->db_buf != NULL); 2292 2293 DB_DNODE_ENTER(db); 2294 dn = DB_DNODE(db); 2295 /* Indirect block size must match what the dnode thinks it is. */ 2296 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 2297 dbuf_check_blkptr(dn, db); 2298 DB_DNODE_EXIT(db); 2299 2300 /* Provide the pending dirty record to child dbufs */ 2301 db->db_data_pending = dr; 2302 2303 mutex_exit(&db->db_mtx); 2304 dbuf_write(dr, db->db_buf, tx); 2305 2306 zio = dr->dr_zio; 2307 mutex_enter(&dr->dt.di.dr_mtx); 2308 dbuf_sync_list(&dr->dt.di.dr_children, tx); 2309 ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 2310 mutex_exit(&dr->dt.di.dr_mtx); 2311 zio_nowait(zio); 2312 } 2313 2314 static void 2315 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx) 2316 { 2317 arc_buf_t **datap = &dr->dt.dl.dr_data; 2318 dmu_buf_impl_t *db = dr->dr_dbuf; 2319 dnode_t *dn; 2320 objset_t *os; 2321 uint64_t txg = tx->tx_txg; 2322 2323 ASSERT(dmu_tx_is_syncing(tx)); 2324 2325 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 2326 2327 mutex_enter(&db->db_mtx); 2328 /* 2329 * To be synced, we must be dirtied. But we 2330 * might have been freed after the dirty. 2331 */ 2332 if (db->db_state == DB_UNCACHED) { 2333 /* This buffer has been freed since it was dirtied */ 2334 ASSERT(db->db.db_data == NULL); 2335 } else if (db->db_state == DB_FILL) { 2336 /* This buffer was freed and is now being re-filled */ 2337 ASSERT(db->db.db_data != dr->dt.dl.dr_data); 2338 } else { 2339 ASSERT(db->db_state == DB_CACHED || db->db_state == DB_NOFILL); 2340 } 2341 DBUF_VERIFY(db); 2342 2343 DB_DNODE_ENTER(db); 2344 dn = DB_DNODE(db); 2345 2346 if (db->db_blkid == DMU_SPILL_BLKID) { 2347 mutex_enter(&dn->dn_mtx); 2348 dn->dn_phys->dn_flags |= DNODE_FLAG_SPILL_BLKPTR; 2349 mutex_exit(&dn->dn_mtx); 2350 } 2351 2352 /* 2353 * If this is a bonus buffer, simply copy the bonus data into the 2354 * dnode. It will be written out when the dnode is synced (and it 2355 * will be synced, since it must have been dirty for dbuf_sync to 2356 * be called). 2357 */ 2358 if (db->db_blkid == DMU_BONUS_BLKID) { 2359 dbuf_dirty_record_t **drp; 2360 2361 ASSERT(*datap != NULL); 2362 ASSERT0(db->db_level); 2363 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN); 2364 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen); 2365 DB_DNODE_EXIT(db); 2366 2367 if (*datap != db->db.db_data) { 2368 zio_buf_free(*datap, DN_MAX_BONUSLEN); 2369 arc_space_return(DN_MAX_BONUSLEN, ARC_SPACE_OTHER); 2370 } 2371 db->db_data_pending = NULL; 2372 drp = &db->db_last_dirty; 2373 while (*drp != dr) 2374 drp = &(*drp)->dr_next; 2375 ASSERT(dr->dr_next == NULL); 2376 ASSERT(dr->dr_dbuf == db); 2377 *drp = dr->dr_next; 2378 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 2379 ASSERT(db->db_dirtycnt > 0); 2380 db->db_dirtycnt -= 1; 2381 dbuf_rele_and_unlock(db, (void *)(uintptr_t)txg); 2382 return; 2383 } 2384 2385 os = dn->dn_objset; 2386 2387 /* 2388 * This function may have dropped the db_mtx lock allowing a dmu_sync 2389 * operation to sneak in. As a result, we need to ensure that we 2390 * don't check the dr_override_state until we have returned from 2391 * dbuf_check_blkptr. 2392 */ 2393 dbuf_check_blkptr(dn, db); 2394 2395 /* 2396 * If this buffer is in the middle of an immediate write, 2397 * wait for the synchronous IO to complete. 2398 */ 2399 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) { 2400 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT); 2401 cv_wait(&db->db_changed, &db->db_mtx); 2402 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN); 2403 } 2404 2405 if (db->db_state != DB_NOFILL && 2406 dn->dn_object != DMU_META_DNODE_OBJECT && 2407 refcount_count(&db->db_holds) > 1 && 2408 dr->dt.dl.dr_override_state != DR_OVERRIDDEN && 2409 *datap == db->db_buf) { 2410 /* 2411 * If this buffer is currently "in use" (i.e., there 2412 * are active holds and db_data still references it), 2413 * then make a copy before we start the write so that 2414 * any modifications from the open txg will not leak 2415 * into this write. 2416 * 2417 * NOTE: this copy does not need to be made for 2418 * objects only modified in the syncing context (e.g. 2419 * DNONE_DNODE blocks). 2420 */ 2421 int blksz = arc_buf_size(*datap); 2422 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 2423 *datap = arc_buf_alloc(os->os_spa, blksz, db, type); 2424 bcopy(db->db.db_data, (*datap)->b_data, blksz); 2425 } 2426 db->db_data_pending = dr; 2427 2428 mutex_exit(&db->db_mtx); 2429 2430 dbuf_write(dr, *datap, tx); 2431 2432 ASSERT(!list_link_active(&dr->dr_dirty_node)); 2433 if (dn->dn_object == DMU_META_DNODE_OBJECT) { 2434 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr); 2435 DB_DNODE_EXIT(db); 2436 } else { 2437 /* 2438 * Although zio_nowait() does not "wait for an IO", it does 2439 * initiate the IO. If this is an empty write it seems plausible 2440 * that the IO could actually be completed before the nowait 2441 * returns. We need to DB_DNODE_EXIT() first in case 2442 * zio_nowait() invalidates the dbuf. 2443 */ 2444 DB_DNODE_EXIT(db); 2445 zio_nowait(dr->dr_zio); 2446 } 2447 } 2448 2449 void 2450 dbuf_sync_list(list_t *list, dmu_tx_t *tx) 2451 { 2452 dbuf_dirty_record_t *dr; 2453 2454 while (dr = list_head(list)) { 2455 if (dr->dr_zio != NULL) { 2456 /* 2457 * If we find an already initialized zio then we 2458 * are processing the meta-dnode, and we have finished. 2459 * The dbufs for all dnodes are put back on the list 2460 * during processing, so that we can zio_wait() 2461 * these IOs after initiating all child IOs. 2462 */ 2463 ASSERT3U(dr->dr_dbuf->db.db_object, ==, 2464 DMU_META_DNODE_OBJECT); 2465 break; 2466 } 2467 list_remove(list, dr); 2468 if (dr->dr_dbuf->db_level > 0) 2469 dbuf_sync_indirect(dr, tx); 2470 else 2471 dbuf_sync_leaf(dr, tx); 2472 } 2473 } 2474 2475 /* ARGSUSED */ 2476 static void 2477 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb) 2478 { 2479 dmu_buf_impl_t *db = vdb; 2480 dnode_t *dn; 2481 blkptr_t *bp = zio->io_bp; 2482 blkptr_t *bp_orig = &zio->io_bp_orig; 2483 spa_t *spa = zio->io_spa; 2484 int64_t delta; 2485 uint64_t fill = 0; 2486 int i; 2487 2488 ASSERT3P(db->db_blkptr, ==, bp); 2489 2490 DB_DNODE_ENTER(db); 2491 dn = DB_DNODE(db); 2492 delta = bp_get_dsize_sync(spa, bp) - bp_get_dsize_sync(spa, bp_orig); 2493 dnode_diduse_space(dn, delta - zio->io_prev_space_delta); 2494 zio->io_prev_space_delta = delta; 2495 2496 if (bp->blk_birth != 0) { 2497 ASSERT((db->db_blkid != DMU_SPILL_BLKID && 2498 BP_GET_TYPE(bp) == dn->dn_type) || 2499 (db->db_blkid == DMU_SPILL_BLKID && 2500 BP_GET_TYPE(bp) == dn->dn_bonustype) || 2501 BP_IS_EMBEDDED(bp)); 2502 ASSERT(BP_GET_LEVEL(bp) == db->db_level); 2503 } 2504 2505 mutex_enter(&db->db_mtx); 2506 2507 #ifdef ZFS_DEBUG 2508 if (db->db_blkid == DMU_SPILL_BLKID) { 2509 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); 2510 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && 2511 db->db_blkptr == &dn->dn_phys->dn_spill); 2512 } 2513 #endif 2514 2515 if (db->db_level == 0) { 2516 mutex_enter(&dn->dn_mtx); 2517 if (db->db_blkid > dn->dn_phys->dn_maxblkid && 2518 db->db_blkid != DMU_SPILL_BLKID) 2519 dn->dn_phys->dn_maxblkid = db->db_blkid; 2520 mutex_exit(&dn->dn_mtx); 2521 2522 if (dn->dn_type == DMU_OT_DNODE) { 2523 dnode_phys_t *dnp = db->db.db_data; 2524 for (i = db->db.db_size >> DNODE_SHIFT; i > 0; 2525 i--, dnp++) { 2526 if (dnp->dn_type != DMU_OT_NONE) 2527 fill++; 2528 } 2529 } else { 2530 if (BP_IS_HOLE(bp)) { 2531 fill = 0; 2532 } else { 2533 fill = 1; 2534 } 2535 } 2536 } else { 2537 blkptr_t *ibp = db->db.db_data; 2538 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 2539 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, ibp++) { 2540 if (BP_IS_HOLE(ibp)) 2541 continue; 2542 fill += BP_GET_FILL(ibp); 2543 } 2544 } 2545 DB_DNODE_EXIT(db); 2546 2547 if (!BP_IS_EMBEDDED(bp)) 2548 bp->blk_fill = fill; 2549 2550 mutex_exit(&db->db_mtx); 2551 } 2552 2553 /* 2554 * The SPA will call this callback several times for each zio - once 2555 * for every physical child i/o (zio->io_phys_children times). This 2556 * allows the DMU to monitor the progress of each logical i/o. For example, 2557 * there may be 2 copies of an indirect block, or many fragments of a RAID-Z 2558 * block. There may be a long delay before all copies/fragments are completed, 2559 * so this callback allows us to retire dirty space gradually, as the physical 2560 * i/os complete. 2561 */ 2562 /* ARGSUSED */ 2563 static void 2564 dbuf_write_physdone(zio_t *zio, arc_buf_t *buf, void *arg) 2565 { 2566 dmu_buf_impl_t *db = arg; 2567 objset_t *os = db->db_objset; 2568 dsl_pool_t *dp = dmu_objset_pool(os); 2569 dbuf_dirty_record_t *dr; 2570 int delta = 0; 2571 2572 dr = db->db_data_pending; 2573 ASSERT3U(dr->dr_txg, ==, zio->io_txg); 2574 2575 /* 2576 * The callback will be called io_phys_children times. Retire one 2577 * portion of our dirty space each time we are called. Any rounding 2578 * error will be cleaned up by dsl_pool_sync()'s call to 2579 * dsl_pool_undirty_space(). 2580 */ 2581 delta = dr->dr_accounted / zio->io_phys_children; 2582 dsl_pool_undirty_space(dp, delta, zio->io_txg); 2583 } 2584 2585 /* ARGSUSED */ 2586 static void 2587 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb) 2588 { 2589 dmu_buf_impl_t *db = vdb; 2590 blkptr_t *bp_orig = &zio->io_bp_orig; 2591 blkptr_t *bp = db->db_blkptr; 2592 objset_t *os = db->db_objset; 2593 dmu_tx_t *tx = os->os_synctx; 2594 dbuf_dirty_record_t **drp, *dr; 2595 2596 ASSERT0(zio->io_error); 2597 ASSERT(db->db_blkptr == bp); 2598 2599 /* 2600 * For nopwrites and rewrites we ensure that the bp matches our 2601 * original and bypass all the accounting. 2602 */ 2603 if (zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE)) { 2604 ASSERT(BP_EQUAL(bp, bp_orig)); 2605 } else { 2606 dsl_dataset_t *ds = os->os_dsl_dataset; 2607 (void) dsl_dataset_block_kill(ds, bp_orig, tx, B_TRUE); 2608 dsl_dataset_block_born(ds, bp, tx); 2609 } 2610 2611 mutex_enter(&db->db_mtx); 2612 2613 DBUF_VERIFY(db); 2614 2615 drp = &db->db_last_dirty; 2616 while ((dr = *drp) != db->db_data_pending) 2617 drp = &dr->dr_next; 2618 ASSERT(!list_link_active(&dr->dr_dirty_node)); 2619 ASSERT(dr->dr_dbuf == db); 2620 ASSERT(dr->dr_next == NULL); 2621 *drp = dr->dr_next; 2622 2623 #ifdef ZFS_DEBUG 2624 if (db->db_blkid == DMU_SPILL_BLKID) { 2625 dnode_t *dn; 2626 2627 DB_DNODE_ENTER(db); 2628 dn = DB_DNODE(db); 2629 ASSERT(dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR); 2630 ASSERT(!(BP_IS_HOLE(db->db_blkptr)) && 2631 db->db_blkptr == &dn->dn_phys->dn_spill); 2632 DB_DNODE_EXIT(db); 2633 } 2634 #endif 2635 2636 if (db->db_level == 0) { 2637 ASSERT(db->db_blkid != DMU_BONUS_BLKID); 2638 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); 2639 if (db->db_state != DB_NOFILL) { 2640 if (dr->dt.dl.dr_data != db->db_buf) 2641 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, 2642 db)); 2643 else if (!arc_released(db->db_buf)) 2644 arc_set_callback(db->db_buf, dbuf_do_evict, db); 2645 } 2646 } else { 2647 dnode_t *dn; 2648 2649 DB_DNODE_ENTER(db); 2650 dn = DB_DNODE(db); 2651 ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 2652 ASSERT3U(db->db.db_size, ==, 1 << dn->dn_phys->dn_indblkshift); 2653 if (!BP_IS_HOLE(db->db_blkptr)) { 2654 int epbs = 2655 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 2656 ASSERT3U(db->db_blkid, <=, 2657 dn->dn_phys->dn_maxblkid >> (db->db_level * epbs)); 2658 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, 2659 db->db.db_size); 2660 if (!arc_released(db->db_buf)) 2661 arc_set_callback(db->db_buf, dbuf_do_evict, db); 2662 } 2663 DB_DNODE_EXIT(db); 2664 mutex_destroy(&dr->dt.di.dr_mtx); 2665 list_destroy(&dr->dt.di.dr_children); 2666 } 2667 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 2668 2669 cv_broadcast(&db->db_changed); 2670 ASSERT(db->db_dirtycnt > 0); 2671 db->db_dirtycnt -= 1; 2672 db->db_data_pending = NULL; 2673 dbuf_rele_and_unlock(db, (void *)(uintptr_t)tx->tx_txg); 2674 } 2675 2676 static void 2677 dbuf_write_nofill_ready(zio_t *zio) 2678 { 2679 dbuf_write_ready(zio, NULL, zio->io_private); 2680 } 2681 2682 static void 2683 dbuf_write_nofill_done(zio_t *zio) 2684 { 2685 dbuf_write_done(zio, NULL, zio->io_private); 2686 } 2687 2688 static void 2689 dbuf_write_override_ready(zio_t *zio) 2690 { 2691 dbuf_dirty_record_t *dr = zio->io_private; 2692 dmu_buf_impl_t *db = dr->dr_dbuf; 2693 2694 dbuf_write_ready(zio, NULL, db); 2695 } 2696 2697 static void 2698 dbuf_write_override_done(zio_t *zio) 2699 { 2700 dbuf_dirty_record_t *dr = zio->io_private; 2701 dmu_buf_impl_t *db = dr->dr_dbuf; 2702 blkptr_t *obp = &dr->dt.dl.dr_overridden_by; 2703 2704 mutex_enter(&db->db_mtx); 2705 if (!BP_EQUAL(zio->io_bp, obp)) { 2706 if (!BP_IS_HOLE(obp)) 2707 dsl_free(spa_get_dsl(zio->io_spa), zio->io_txg, obp); 2708 arc_release(dr->dt.dl.dr_data, db); 2709 } 2710 mutex_exit(&db->db_mtx); 2711 2712 dbuf_write_done(zio, NULL, db); 2713 } 2714 2715 /* Issue I/O to commit a dirty buffer to disk. */ 2716 static void 2717 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx) 2718 { 2719 dmu_buf_impl_t *db = dr->dr_dbuf; 2720 dnode_t *dn; 2721 objset_t *os; 2722 dmu_buf_impl_t *parent = db->db_parent; 2723 uint64_t txg = tx->tx_txg; 2724 zbookmark_phys_t zb; 2725 zio_prop_t zp; 2726 zio_t *zio; 2727 int wp_flag = 0; 2728 2729 DB_DNODE_ENTER(db); 2730 dn = DB_DNODE(db); 2731 os = dn->dn_objset; 2732 2733 if (db->db_state != DB_NOFILL) { 2734 if (db->db_level > 0 || dn->dn_type == DMU_OT_DNODE) { 2735 /* 2736 * Private object buffers are released here rather 2737 * than in dbuf_dirty() since they are only modified 2738 * in the syncing context and we don't want the 2739 * overhead of making multiple copies of the data. 2740 */ 2741 if (BP_IS_HOLE(db->db_blkptr)) { 2742 arc_buf_thaw(data); 2743 } else { 2744 dbuf_release_bp(db); 2745 } 2746 } 2747 } 2748 2749 if (parent != dn->dn_dbuf) { 2750 /* Our parent is an indirect block. */ 2751 /* We have a dirty parent that has been scheduled for write. */ 2752 ASSERT(parent && parent->db_data_pending); 2753 /* Our parent's buffer is one level closer to the dnode. */ 2754 ASSERT(db->db_level == parent->db_level-1); 2755 /* 2756 * We're about to modify our parent's db_data by modifying 2757 * our block pointer, so the parent must be released. 2758 */ 2759 ASSERT(arc_released(parent->db_buf)); 2760 zio = parent->db_data_pending->dr_zio; 2761 } else { 2762 /* Our parent is the dnode itself. */ 2763 ASSERT((db->db_level == dn->dn_phys->dn_nlevels-1 && 2764 db->db_blkid != DMU_SPILL_BLKID) || 2765 (db->db_blkid == DMU_SPILL_BLKID && db->db_level == 0)); 2766 if (db->db_blkid != DMU_SPILL_BLKID) 2767 ASSERT3P(db->db_blkptr, ==, 2768 &dn->dn_phys->dn_blkptr[db->db_blkid]); 2769 zio = dn->dn_zio; 2770 } 2771 2772 ASSERT(db->db_level == 0 || data == db->db_buf); 2773 ASSERT3U(db->db_blkptr->blk_birth, <=, txg); 2774 ASSERT(zio); 2775 2776 SET_BOOKMARK(&zb, os->os_dsl_dataset ? 2777 os->os_dsl_dataset->ds_object : DMU_META_OBJSET, 2778 db->db.db_object, db->db_level, db->db_blkid); 2779 2780 if (db->db_blkid == DMU_SPILL_BLKID) 2781 wp_flag = WP_SPILL; 2782 wp_flag |= (db->db_state == DB_NOFILL) ? WP_NOFILL : 0; 2783 2784 dmu_write_policy(os, dn, db->db_level, wp_flag, &zp); 2785 DB_DNODE_EXIT(db); 2786 2787 if (db->db_level == 0 && 2788 dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { 2789 /* 2790 * The BP for this block has been provided by open context 2791 * (by dmu_sync() or dmu_buf_write_embedded()). 2792 */ 2793 void *contents = (data != NULL) ? data->b_data : NULL; 2794 2795 dr->dr_zio = zio_write(zio, os->os_spa, txg, 2796 db->db_blkptr, contents, db->db.db_size, &zp, 2797 dbuf_write_override_ready, NULL, dbuf_write_override_done, 2798 dr, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); 2799 mutex_enter(&db->db_mtx); 2800 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 2801 zio_write_override(dr->dr_zio, &dr->dt.dl.dr_overridden_by, 2802 dr->dt.dl.dr_copies, dr->dt.dl.dr_nopwrite); 2803 mutex_exit(&db->db_mtx); 2804 } else if (db->db_state == DB_NOFILL) { 2805 ASSERT(zp.zp_checksum == ZIO_CHECKSUM_OFF || 2806 zp.zp_checksum == ZIO_CHECKSUM_NOPARITY); 2807 dr->dr_zio = zio_write(zio, os->os_spa, txg, 2808 db->db_blkptr, NULL, db->db.db_size, &zp, 2809 dbuf_write_nofill_ready, NULL, dbuf_write_nofill_done, db, 2810 ZIO_PRIORITY_ASYNC_WRITE, 2811 ZIO_FLAG_MUSTSUCCEED | ZIO_FLAG_NODATA, &zb); 2812 } else { 2813 ASSERT(arc_released(data)); 2814 dr->dr_zio = arc_write(zio, os->os_spa, txg, 2815 db->db_blkptr, data, DBUF_IS_L2CACHEABLE(db), 2816 DBUF_IS_L2COMPRESSIBLE(db), &zp, dbuf_write_ready, 2817 dbuf_write_physdone, dbuf_write_done, db, 2818 ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb); 2819 } 2820 } 2821