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 2007 Sun Microsystems, Inc. All rights reserved. 23 * Use is subject to license terms. 24 */ 25 26 #pragma ident "%Z%%M% %I% %E% SMI" 27 28 #include <sys/zfs_context.h> 29 #include <sys/dmu.h> 30 #include <sys/dmu_impl.h> 31 #include <sys/dbuf.h> 32 #include <sys/dmu_objset.h> 33 #include <sys/dsl_dataset.h> 34 #include <sys/dsl_dir.h> 35 #include <sys/dmu_tx.h> 36 #include <sys/spa.h> 37 #include <sys/zio.h> 38 #include <sys/dmu_zfetch.h> 39 40 static void dbuf_destroy(dmu_buf_impl_t *db); 41 static int dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx); 42 static void dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, int checksum, 43 int compress, dmu_tx_t *tx); 44 static arc_done_func_t dbuf_write_ready; 45 static arc_done_func_t dbuf_write_done; 46 47 int zfs_mdcomp_disable = 0; 48 49 /* 50 * Global data structures and functions for the dbuf cache. 51 */ 52 static kmem_cache_t *dbuf_cache; 53 54 /* ARGSUSED */ 55 static int 56 dbuf_cons(void *vdb, void *unused, int kmflag) 57 { 58 dmu_buf_impl_t *db = vdb; 59 bzero(db, sizeof (dmu_buf_impl_t)); 60 61 mutex_init(&db->db_mtx, NULL, MUTEX_DEFAULT, NULL); 62 cv_init(&db->db_changed, NULL, CV_DEFAULT, NULL); 63 refcount_create(&db->db_holds); 64 return (0); 65 } 66 67 /* ARGSUSED */ 68 static void 69 dbuf_dest(void *vdb, void *unused) 70 { 71 dmu_buf_impl_t *db = vdb; 72 mutex_destroy(&db->db_mtx); 73 cv_destroy(&db->db_changed); 74 refcount_destroy(&db->db_holds); 75 } 76 77 /* 78 * dbuf hash table routines 79 */ 80 static dbuf_hash_table_t dbuf_hash_table; 81 82 static uint64_t dbuf_hash_count; 83 84 static uint64_t 85 dbuf_hash(void *os, uint64_t obj, uint8_t lvl, uint64_t blkid) 86 { 87 uintptr_t osv = (uintptr_t)os; 88 uint64_t crc = -1ULL; 89 90 ASSERT(zfs_crc64_table[128] == ZFS_CRC64_POLY); 91 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (lvl)) & 0xFF]; 92 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (osv >> 6)) & 0xFF]; 93 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 0)) & 0xFF]; 94 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (obj >> 8)) & 0xFF]; 95 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 0)) & 0xFF]; 96 crc = (crc >> 8) ^ zfs_crc64_table[(crc ^ (blkid >> 8)) & 0xFF]; 97 98 crc ^= (osv>>14) ^ (obj>>16) ^ (blkid>>16); 99 100 return (crc); 101 } 102 103 #define DBUF_HASH(os, obj, level, blkid) dbuf_hash(os, obj, level, blkid); 104 105 #define DBUF_EQUAL(dbuf, os, obj, level, blkid) \ 106 ((dbuf)->db.db_object == (obj) && \ 107 (dbuf)->db_objset == (os) && \ 108 (dbuf)->db_level == (level) && \ 109 (dbuf)->db_blkid == (blkid)) 110 111 dmu_buf_impl_t * 112 dbuf_find(dnode_t *dn, uint8_t level, uint64_t blkid) 113 { 114 dbuf_hash_table_t *h = &dbuf_hash_table; 115 objset_impl_t *os = dn->dn_objset; 116 uint64_t obj = dn->dn_object; 117 uint64_t hv = DBUF_HASH(os, obj, level, blkid); 118 uint64_t idx = hv & h->hash_table_mask; 119 dmu_buf_impl_t *db; 120 121 mutex_enter(DBUF_HASH_MUTEX(h, idx)); 122 for (db = h->hash_table[idx]; db != NULL; db = db->db_hash_next) { 123 if (DBUF_EQUAL(db, os, obj, level, blkid)) { 124 mutex_enter(&db->db_mtx); 125 if (db->db_state != DB_EVICTING) { 126 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 127 return (db); 128 } 129 mutex_exit(&db->db_mtx); 130 } 131 } 132 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 133 return (NULL); 134 } 135 136 /* 137 * Insert an entry into the hash table. If there is already an element 138 * equal to elem in the hash table, then the already existing element 139 * will be returned and the new element will not be inserted. 140 * Otherwise returns NULL. 141 */ 142 static dmu_buf_impl_t * 143 dbuf_hash_insert(dmu_buf_impl_t *db) 144 { 145 dbuf_hash_table_t *h = &dbuf_hash_table; 146 objset_impl_t *os = db->db_objset; 147 uint64_t obj = db->db.db_object; 148 int level = db->db_level; 149 uint64_t blkid = db->db_blkid; 150 uint64_t hv = DBUF_HASH(os, obj, level, blkid); 151 uint64_t idx = hv & h->hash_table_mask; 152 dmu_buf_impl_t *dbf; 153 154 mutex_enter(DBUF_HASH_MUTEX(h, idx)); 155 for (dbf = h->hash_table[idx]; dbf != NULL; dbf = dbf->db_hash_next) { 156 if (DBUF_EQUAL(dbf, os, obj, level, blkid)) { 157 mutex_enter(&dbf->db_mtx); 158 if (dbf->db_state != DB_EVICTING) { 159 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 160 return (dbf); 161 } 162 mutex_exit(&dbf->db_mtx); 163 } 164 } 165 166 mutex_enter(&db->db_mtx); 167 db->db_hash_next = h->hash_table[idx]; 168 h->hash_table[idx] = db; 169 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 170 atomic_add_64(&dbuf_hash_count, 1); 171 172 return (NULL); 173 } 174 175 /* 176 * Remove an entry from the hash table. This operation will 177 * fail if there are any existing holds on the db. 178 */ 179 static void 180 dbuf_hash_remove(dmu_buf_impl_t *db) 181 { 182 dbuf_hash_table_t *h = &dbuf_hash_table; 183 uint64_t hv = DBUF_HASH(db->db_objset, db->db.db_object, 184 db->db_level, db->db_blkid); 185 uint64_t idx = hv & h->hash_table_mask; 186 dmu_buf_impl_t *dbf, **dbp; 187 188 /* 189 * We musn't hold db_mtx to maintin lock ordering: 190 * DBUF_HASH_MUTEX > db_mtx. 191 */ 192 ASSERT(refcount_is_zero(&db->db_holds)); 193 ASSERT(db->db_state == DB_EVICTING); 194 ASSERT(!MUTEX_HELD(&db->db_mtx)); 195 196 mutex_enter(DBUF_HASH_MUTEX(h, idx)); 197 dbp = &h->hash_table[idx]; 198 while ((dbf = *dbp) != db) { 199 dbp = &dbf->db_hash_next; 200 ASSERT(dbf != NULL); 201 } 202 *dbp = db->db_hash_next; 203 db->db_hash_next = NULL; 204 mutex_exit(DBUF_HASH_MUTEX(h, idx)); 205 atomic_add_64(&dbuf_hash_count, -1); 206 } 207 208 static arc_evict_func_t dbuf_do_evict; 209 210 static void 211 dbuf_evict_user(dmu_buf_impl_t *db) 212 { 213 ASSERT(MUTEX_HELD(&db->db_mtx)); 214 215 if (db->db_level != 0 || db->db_evict_func == NULL) 216 return; 217 218 if (db->db_user_data_ptr_ptr) 219 *db->db_user_data_ptr_ptr = db->db.db_data; 220 db->db_evict_func(&db->db, db->db_user_ptr); 221 db->db_user_ptr = NULL; 222 db->db_user_data_ptr_ptr = NULL; 223 db->db_evict_func = NULL; 224 } 225 226 void 227 dbuf_evict(dmu_buf_impl_t *db) 228 { 229 ASSERT(MUTEX_HELD(&db->db_mtx)); 230 ASSERT(db->db_buf == NULL); 231 ASSERT(db->db_data_pending == NULL); 232 233 dbuf_clear(db); 234 dbuf_destroy(db); 235 } 236 237 void 238 dbuf_init(void) 239 { 240 uint64_t hsize = 1ULL << 16; 241 dbuf_hash_table_t *h = &dbuf_hash_table; 242 int i; 243 244 /* 245 * The hash table is big enough to fill all of physical memory 246 * with an average 4K block size. The table will take up 247 * totalmem*sizeof(void*)/4K (i.e. 2MB/GB with 8-byte pointers). 248 */ 249 while (hsize * 4096 < physmem * PAGESIZE) 250 hsize <<= 1; 251 252 retry: 253 h->hash_table_mask = hsize - 1; 254 h->hash_table = kmem_zalloc(hsize * sizeof (void *), KM_NOSLEEP); 255 if (h->hash_table == NULL) { 256 /* XXX - we should really return an error instead of assert */ 257 ASSERT(hsize > (1ULL << 10)); 258 hsize >>= 1; 259 goto retry; 260 } 261 262 dbuf_cache = kmem_cache_create("dmu_buf_impl_t", 263 sizeof (dmu_buf_impl_t), 264 0, dbuf_cons, dbuf_dest, NULL, NULL, NULL, 0); 265 266 for (i = 0; i < DBUF_MUTEXES; i++) 267 mutex_init(&h->hash_mutexes[i], NULL, MUTEX_DEFAULT, NULL); 268 } 269 270 void 271 dbuf_fini(void) 272 { 273 dbuf_hash_table_t *h = &dbuf_hash_table; 274 int i; 275 276 for (i = 0; i < DBUF_MUTEXES; i++) 277 mutex_destroy(&h->hash_mutexes[i]); 278 kmem_free(h->hash_table, (h->hash_table_mask + 1) * sizeof (void *)); 279 kmem_cache_destroy(dbuf_cache); 280 } 281 282 /* 283 * Other stuff. 284 */ 285 286 #ifdef ZFS_DEBUG 287 static void 288 dbuf_verify(dmu_buf_impl_t *db) 289 { 290 dnode_t *dn = db->db_dnode; 291 292 ASSERT(MUTEX_HELD(&db->db_mtx)); 293 294 if (!(zfs_flags & ZFS_DEBUG_DBUF_VERIFY)) 295 return; 296 297 ASSERT(db->db_objset != NULL); 298 if (dn == NULL) { 299 ASSERT(db->db_parent == NULL); 300 ASSERT(db->db_blkptr == NULL); 301 } else { 302 ASSERT3U(db->db.db_object, ==, dn->dn_object); 303 ASSERT3P(db->db_objset, ==, dn->dn_objset); 304 ASSERT3U(db->db_level, <, dn->dn_nlevels); 305 ASSERT(db->db_blkid == DB_BONUS_BLKID || 306 list_head(&dn->dn_dbufs)); 307 } 308 if (db->db_blkid == DB_BONUS_BLKID) { 309 ASSERT(dn != NULL); 310 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 311 ASSERT3U(db->db.db_offset, ==, DB_BONUS_BLKID); 312 } else { 313 ASSERT3U(db->db.db_offset, ==, db->db_blkid * db->db.db_size); 314 } 315 316 if (db->db_level == 0) { 317 /* we can be momentarily larger in dnode_set_blksz() */ 318 if (db->db_blkid != DB_BONUS_BLKID && dn) { 319 ASSERT3U(db->db.db_size, >=, dn->dn_datablksz); 320 } 321 if (db->db.db_object == DMU_META_DNODE_OBJECT) { 322 dbuf_dirty_record_t *dr = db->db_data_pending; 323 /* 324 * it should only be modified in syncing 325 * context, so make sure we only have 326 * one copy of the data. 327 */ 328 ASSERT(dr == NULL || dr->dt.dl.dr_data == db->db_buf); 329 } 330 } 331 332 /* verify db->db_blkptr */ 333 if (db->db_blkptr) { 334 if (db->db_parent == dn->dn_dbuf) { 335 /* db is pointed to by the dnode */ 336 /* ASSERT3U(db->db_blkid, <, dn->dn_nblkptr); */ 337 if (db->db.db_object == DMU_META_DNODE_OBJECT) 338 ASSERT(db->db_parent == NULL); 339 else 340 ASSERT(db->db_parent != NULL); 341 ASSERT3P(db->db_blkptr, ==, 342 &dn->dn_phys->dn_blkptr[db->db_blkid]); 343 } else { 344 /* db is pointed to by an indirect block */ 345 int epb = db->db_parent->db.db_size >> SPA_BLKPTRSHIFT; 346 ASSERT3U(db->db_parent->db_level, ==, db->db_level+1); 347 ASSERT3U(db->db_parent->db.db_object, ==, 348 db->db.db_object); 349 /* 350 * dnode_grow_indblksz() can make this fail if we don't 351 * have the struct_rwlock. XXX indblksz no longer 352 * grows. safe to do this now? 353 */ 354 if (RW_WRITE_HELD(&db->db_dnode->dn_struct_rwlock)) { 355 ASSERT3P(db->db_blkptr, ==, 356 ((blkptr_t *)db->db_parent->db.db_data + 357 db->db_blkid % epb)); 358 } 359 } 360 } 361 if ((db->db_blkptr == NULL || BP_IS_HOLE(db->db_blkptr)) && 362 db->db.db_data && db->db_blkid != DB_BONUS_BLKID && 363 db->db_state != DB_FILL && !dn->dn_free_txg) { 364 /* 365 * If the blkptr isn't set but they have nonzero data, 366 * it had better be dirty, otherwise we'll lose that 367 * data when we evict this buffer. 368 */ 369 if (db->db_dirtycnt == 0) { 370 uint64_t *buf = db->db.db_data; 371 int i; 372 373 for (i = 0; i < db->db.db_size >> 3; i++) { 374 ASSERT(buf[i] == 0); 375 } 376 } 377 } 378 } 379 #endif 380 381 static void 382 dbuf_update_data(dmu_buf_impl_t *db) 383 { 384 ASSERT(MUTEX_HELD(&db->db_mtx)); 385 if (db->db_level == 0 && db->db_user_data_ptr_ptr) { 386 ASSERT(!refcount_is_zero(&db->db_holds)); 387 *db->db_user_data_ptr_ptr = db->db.db_data; 388 } 389 } 390 391 static void 392 dbuf_set_data(dmu_buf_impl_t *db, arc_buf_t *buf) 393 { 394 ASSERT(MUTEX_HELD(&db->db_mtx)); 395 ASSERT(db->db_buf == NULL || !arc_has_callback(db->db_buf)); 396 db->db_buf = buf; 397 if (buf != NULL) { 398 ASSERT(buf->b_data != NULL); 399 db->db.db_data = buf->b_data; 400 if (!arc_released(buf)) 401 arc_set_callback(buf, dbuf_do_evict, db); 402 dbuf_update_data(db); 403 } else { 404 dbuf_evict_user(db); 405 db->db.db_data = NULL; 406 db->db_state = DB_UNCACHED; 407 } 408 } 409 410 uint64_t 411 dbuf_whichblock(dnode_t *dn, uint64_t offset) 412 { 413 if (dn->dn_datablkshift) { 414 return (offset >> dn->dn_datablkshift); 415 } else { 416 ASSERT3U(offset, <, dn->dn_datablksz); 417 return (0); 418 } 419 } 420 421 static void 422 dbuf_read_done(zio_t *zio, arc_buf_t *buf, void *vdb) 423 { 424 dmu_buf_impl_t *db = vdb; 425 426 mutex_enter(&db->db_mtx); 427 ASSERT3U(db->db_state, ==, DB_READ); 428 /* 429 * All reads are synchronous, so we must have a hold on the dbuf 430 */ 431 ASSERT(refcount_count(&db->db_holds) > 0); 432 ASSERT(db->db_buf == NULL); 433 ASSERT(db->db.db_data == NULL); 434 if (db->db_level == 0 && db->db_freed_in_flight) { 435 /* we were freed in flight; disregard any error */ 436 arc_release(buf, db); 437 bzero(buf->b_data, db->db.db_size); 438 arc_buf_freeze(buf); 439 db->db_freed_in_flight = FALSE; 440 dbuf_set_data(db, buf); 441 db->db_state = DB_CACHED; 442 } else if (zio == NULL || zio->io_error == 0) { 443 dbuf_set_data(db, buf); 444 db->db_state = DB_CACHED; 445 } else { 446 ASSERT(db->db_blkid != DB_BONUS_BLKID); 447 ASSERT3P(db->db_buf, ==, NULL); 448 VERIFY(arc_buf_remove_ref(buf, db) == 1); 449 db->db_state = DB_UNCACHED; 450 } 451 cv_broadcast(&db->db_changed); 452 mutex_exit(&db->db_mtx); 453 dbuf_rele(db, NULL); 454 } 455 456 static void 457 dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t *flags) 458 { 459 blkptr_t *bp; 460 zbookmark_t zb; 461 uint32_t aflags = ARC_NOWAIT; 462 463 ASSERT(!refcount_is_zero(&db->db_holds)); 464 /* We need the struct_rwlock to prevent db_blkptr from changing. */ 465 ASSERT(RW_LOCK_HELD(&db->db_dnode->dn_struct_rwlock)); 466 ASSERT(MUTEX_HELD(&db->db_mtx)); 467 ASSERT(db->db_state == DB_UNCACHED); 468 ASSERT(db->db_buf == NULL); 469 470 if (db->db_blkid == DB_BONUS_BLKID) { 471 int bonuslen = db->db_dnode->dn_bonuslen; 472 473 ASSERT3U(bonuslen, <=, db->db.db_size); 474 db->db.db_data = zio_buf_alloc(DN_MAX_BONUSLEN); 475 arc_space_consume(DN_MAX_BONUSLEN); 476 if (bonuslen < DN_MAX_BONUSLEN) 477 bzero(db->db.db_data, DN_MAX_BONUSLEN); 478 bcopy(DN_BONUS(db->db_dnode->dn_phys), db->db.db_data, 479 bonuslen); 480 dbuf_update_data(db); 481 db->db_state = DB_CACHED; 482 mutex_exit(&db->db_mtx); 483 return; 484 } 485 486 if (db->db_level == 0 && dnode_block_freed(db->db_dnode, db->db_blkid)) 487 bp = NULL; 488 else 489 bp = db->db_blkptr; 490 491 if (bp == NULL) 492 dprintf_dbuf(db, "blkptr: %s\n", "NULL"); 493 else 494 dprintf_dbuf_bp(db, bp, "%s", "blkptr:"); 495 496 if (bp == NULL || BP_IS_HOLE(bp)) { 497 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 498 499 ASSERT(bp == NULL || BP_IS_HOLE(bp)); 500 dbuf_set_data(db, arc_buf_alloc(db->db_dnode->dn_objset->os_spa, 501 db->db.db_size, db, type)); 502 bzero(db->db.db_data, db->db.db_size); 503 db->db_state = DB_CACHED; 504 *flags |= DB_RF_CACHED; 505 mutex_exit(&db->db_mtx); 506 return; 507 } 508 509 db->db_state = DB_READ; 510 mutex_exit(&db->db_mtx); 511 512 zb.zb_objset = db->db_objset->os_dsl_dataset ? 513 db->db_objset->os_dsl_dataset->ds_object : 0; 514 zb.zb_object = db->db.db_object; 515 zb.zb_level = db->db_level; 516 zb.zb_blkid = db->db_blkid; 517 518 dbuf_add_ref(db, NULL); 519 /* ZIO_FLAG_CANFAIL callers have to check the parent zio's error */ 520 ASSERT3U(db->db_dnode->dn_type, <, DMU_OT_NUMTYPES); 521 (void) arc_read(zio, db->db_dnode->dn_objset->os_spa, bp, 522 db->db_level > 0 ? byteswap_uint64_array : 523 dmu_ot[db->db_dnode->dn_type].ot_byteswap, 524 dbuf_read_done, db, ZIO_PRIORITY_SYNC_READ, 525 (*flags & DB_RF_CANFAIL) ? ZIO_FLAG_CANFAIL : ZIO_FLAG_MUSTSUCCEED, 526 &aflags, &zb); 527 if (aflags & ARC_CACHED) 528 *flags |= DB_RF_CACHED; 529 } 530 531 int 532 dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags) 533 { 534 int err = 0; 535 int havepzio = (zio != NULL); 536 int prefetch; 537 538 /* 539 * We don't have to hold the mutex to check db_state because it 540 * can't be freed while we have a hold on the buffer. 541 */ 542 ASSERT(!refcount_is_zero(&db->db_holds)); 543 544 if ((flags & DB_RF_HAVESTRUCT) == 0) 545 rw_enter(&db->db_dnode->dn_struct_rwlock, RW_READER); 546 547 prefetch = db->db_level == 0 && db->db_blkid != DB_BONUS_BLKID && 548 (flags & DB_RF_NOPREFETCH) == 0 && db->db_dnode != NULL; 549 550 mutex_enter(&db->db_mtx); 551 if (db->db_state == DB_CACHED) { 552 mutex_exit(&db->db_mtx); 553 if (prefetch) 554 dmu_zfetch(&db->db_dnode->dn_zfetch, db->db.db_offset, 555 db->db.db_size, TRUE); 556 if ((flags & DB_RF_HAVESTRUCT) == 0) 557 rw_exit(&db->db_dnode->dn_struct_rwlock); 558 } else if (db->db_state == DB_UNCACHED) { 559 if (zio == NULL) { 560 zio = zio_root(db->db_dnode->dn_objset->os_spa, 561 NULL, NULL, ZIO_FLAG_CANFAIL); 562 } 563 dbuf_read_impl(db, zio, &flags); 564 565 /* dbuf_read_impl has dropped db_mtx for us */ 566 567 if (prefetch) 568 dmu_zfetch(&db->db_dnode->dn_zfetch, db->db.db_offset, 569 db->db.db_size, flags & DB_RF_CACHED); 570 571 if ((flags & DB_RF_HAVESTRUCT) == 0) 572 rw_exit(&db->db_dnode->dn_struct_rwlock); 573 574 if (!havepzio) 575 err = zio_wait(zio); 576 } else { 577 mutex_exit(&db->db_mtx); 578 if (prefetch) 579 dmu_zfetch(&db->db_dnode->dn_zfetch, db->db.db_offset, 580 db->db.db_size, TRUE); 581 if ((flags & DB_RF_HAVESTRUCT) == 0) 582 rw_exit(&db->db_dnode->dn_struct_rwlock); 583 584 mutex_enter(&db->db_mtx); 585 if ((flags & DB_RF_NEVERWAIT) == 0) { 586 while (db->db_state == DB_READ || 587 db->db_state == DB_FILL) { 588 ASSERT(db->db_state == DB_READ || 589 (flags & DB_RF_HAVESTRUCT) == 0); 590 cv_wait(&db->db_changed, &db->db_mtx); 591 } 592 if (db->db_state == DB_UNCACHED) 593 err = EIO; 594 } 595 mutex_exit(&db->db_mtx); 596 } 597 598 ASSERT(err || havepzio || db->db_state == DB_CACHED); 599 return (err); 600 } 601 602 static void 603 dbuf_noread(dmu_buf_impl_t *db) 604 { 605 ASSERT(!refcount_is_zero(&db->db_holds)); 606 ASSERT(db->db_blkid != DB_BONUS_BLKID); 607 mutex_enter(&db->db_mtx); 608 while (db->db_state == DB_READ || db->db_state == DB_FILL) 609 cv_wait(&db->db_changed, &db->db_mtx); 610 if (db->db_state == DB_UNCACHED) { 611 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 612 613 ASSERT(db->db_buf == NULL); 614 ASSERT(db->db.db_data == NULL); 615 dbuf_set_data(db, arc_buf_alloc(db->db_dnode->dn_objset->os_spa, 616 db->db.db_size, db, type)); 617 db->db_state = DB_FILL; 618 } else { 619 ASSERT3U(db->db_state, ==, DB_CACHED); 620 } 621 mutex_exit(&db->db_mtx); 622 } 623 624 /* 625 * This is our just-in-time copy function. It makes a copy of 626 * buffers, that have been modified in a previous transaction 627 * group, before we modify them in the current active group. 628 * 629 * This function is used in two places: when we are dirtying a 630 * buffer for the first time in a txg, and when we are freeing 631 * a range in a dnode that includes this buffer. 632 * 633 * Note that when we are called from dbuf_free_range() we do 634 * not put a hold on the buffer, we just traverse the active 635 * dbuf list for the dnode. 636 */ 637 static void 638 dbuf_fix_old_data(dmu_buf_impl_t *db, uint64_t txg) 639 { 640 dbuf_dirty_record_t *dr = db->db_last_dirty; 641 642 ASSERT(MUTEX_HELD(&db->db_mtx)); 643 ASSERT(db->db.db_data != NULL); 644 ASSERT(db->db_level == 0); 645 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT); 646 647 if (dr == NULL || 648 (dr->dt.dl.dr_data != 649 ((db->db_blkid == DB_BONUS_BLKID) ? db->db.db_data : db->db_buf))) 650 return; 651 652 /* 653 * If the last dirty record for this dbuf has not yet synced 654 * and its referencing the dbuf data, either: 655 * reset the reference to point to a new copy, 656 * or (if there a no active holders) 657 * just null out the current db_data pointer. 658 */ 659 ASSERT(dr->dr_txg >= txg - 2); 660 if (db->db_blkid == DB_BONUS_BLKID) { 661 /* Note that the data bufs here are zio_bufs */ 662 dr->dt.dl.dr_data = zio_buf_alloc(DN_MAX_BONUSLEN); 663 arc_space_consume(DN_MAX_BONUSLEN); 664 bcopy(db->db.db_data, dr->dt.dl.dr_data, DN_MAX_BONUSLEN); 665 } else if (refcount_count(&db->db_holds) > db->db_dirtycnt) { 666 int size = db->db.db_size; 667 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 668 dr->dt.dl.dr_data = arc_buf_alloc( 669 db->db_dnode->dn_objset->os_spa, size, db, type); 670 bcopy(db->db.db_data, dr->dt.dl.dr_data->b_data, size); 671 } else { 672 dbuf_set_data(db, NULL); 673 } 674 } 675 676 void 677 dbuf_unoverride(dbuf_dirty_record_t *dr) 678 { 679 dmu_buf_impl_t *db = dr->dr_dbuf; 680 uint64_t txg = dr->dr_txg; 681 682 ASSERT(MUTEX_HELD(&db->db_mtx)); 683 ASSERT(dr->dt.dl.dr_override_state != DR_IN_DMU_SYNC); 684 ASSERT(db->db_level == 0); 685 686 if (db->db_blkid == DB_BONUS_BLKID || 687 dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN) 688 return; 689 690 /* free this block */ 691 if (!BP_IS_HOLE(&dr->dt.dl.dr_overridden_by)) { 692 /* XXX can get silent EIO here */ 693 (void) arc_free(NULL, db->db_dnode->dn_objset->os_spa, 694 txg, &dr->dt.dl.dr_overridden_by, NULL, NULL, ARC_WAIT); 695 } 696 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 697 /* 698 * Release the already-written buffer, so we leave it in 699 * a consistent dirty state. Note that all callers are 700 * modifying the buffer, so they will immediately do 701 * another (redundant) arc_release(). Therefore, leave 702 * the buf thawed to save the effort of freezing & 703 * immediately re-thawing it. 704 */ 705 arc_release(dr->dt.dl.dr_data, db); 706 } 707 708 void 709 dbuf_free_range(dnode_t *dn, uint64_t blkid, uint64_t nblks, dmu_tx_t *tx) 710 { 711 dmu_buf_impl_t *db, *db_next; 712 uint64_t txg = tx->tx_txg; 713 714 dprintf_dnode(dn, "blkid=%llu nblks=%llu\n", blkid, nblks); 715 mutex_enter(&dn->dn_dbufs_mtx); 716 for (db = list_head(&dn->dn_dbufs); db; db = db_next) { 717 db_next = list_next(&dn->dn_dbufs, db); 718 ASSERT(db->db_blkid != DB_BONUS_BLKID); 719 if (db->db_level != 0) 720 continue; 721 dprintf_dbuf(db, "found buf %s\n", ""); 722 if (db->db_blkid < blkid || 723 db->db_blkid >= blkid+nblks) 724 continue; 725 726 /* found a level 0 buffer in the range */ 727 if (dbuf_undirty(db, tx)) 728 continue; 729 730 mutex_enter(&db->db_mtx); 731 if (db->db_state == DB_UNCACHED || 732 db->db_state == DB_EVICTING) { 733 ASSERT(db->db.db_data == NULL); 734 mutex_exit(&db->db_mtx); 735 continue; 736 } 737 if (db->db_state == DB_READ || db->db_state == DB_FILL) { 738 /* will be handled in dbuf_read_done or dbuf_rele */ 739 db->db_freed_in_flight = TRUE; 740 mutex_exit(&db->db_mtx); 741 continue; 742 } 743 if (refcount_count(&db->db_holds) == 0) { 744 ASSERT(db->db_buf); 745 dbuf_clear(db); 746 continue; 747 } 748 /* The dbuf is referenced */ 749 750 if (db->db_last_dirty != NULL) { 751 dbuf_dirty_record_t *dr = db->db_last_dirty; 752 753 if (dr->dr_txg == txg) { 754 /* 755 * This buffer is "in-use", re-adjust the file 756 * size to reflect that this buffer may 757 * contain new data when we sync. 758 */ 759 if (db->db_blkid > dn->dn_maxblkid) 760 dn->dn_maxblkid = db->db_blkid; 761 dbuf_unoverride(dr); 762 } else { 763 /* 764 * This dbuf is not dirty in the open context. 765 * Either uncache it (if its not referenced in 766 * the open context) or reset its contents to 767 * empty. 768 */ 769 dbuf_fix_old_data(db, txg); 770 } 771 } 772 /* clear the contents if its cached */ 773 if (db->db_state == DB_CACHED) { 774 ASSERT(db->db.db_data != NULL); 775 arc_release(db->db_buf, db); 776 bzero(db->db.db_data, db->db.db_size); 777 arc_buf_freeze(db->db_buf); 778 } 779 780 mutex_exit(&db->db_mtx); 781 } 782 mutex_exit(&dn->dn_dbufs_mtx); 783 } 784 785 static int 786 dbuf_block_freeable(dmu_buf_impl_t *db) 787 { 788 dsl_dataset_t *ds = db->db_objset->os_dsl_dataset; 789 uint64_t birth_txg = 0; 790 791 /* 792 * We don't need any locking to protect db_blkptr: 793 * If it's syncing, then db_last_dirty will be set 794 * so we'll ignore db_blkptr. 795 */ 796 ASSERT(MUTEX_HELD(&db->db_mtx)); 797 if (db->db_last_dirty) 798 birth_txg = db->db_last_dirty->dr_txg; 799 else if (db->db_blkptr) 800 birth_txg = db->db_blkptr->blk_birth; 801 802 /* If we don't exist or are in a snapshot, we can't be freed */ 803 if (birth_txg) 804 return (ds == NULL || 805 dsl_dataset_block_freeable(ds, birth_txg)); 806 else 807 return (FALSE); 808 } 809 810 void 811 dbuf_new_size(dmu_buf_impl_t *db, int size, dmu_tx_t *tx) 812 { 813 arc_buf_t *buf, *obuf; 814 int osize = db->db.db_size; 815 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 816 817 ASSERT(db->db_blkid != DB_BONUS_BLKID); 818 819 /* XXX does *this* func really need the lock? */ 820 ASSERT(RW_WRITE_HELD(&db->db_dnode->dn_struct_rwlock)); 821 822 /* 823 * This call to dbuf_will_dirty() with the dn_struct_rwlock held 824 * is OK, because there can be no other references to the db 825 * when we are changing its size, so no concurrent DB_FILL can 826 * be happening. 827 */ 828 /* 829 * XXX we should be doing a dbuf_read, checking the return 830 * value and returning that up to our callers 831 */ 832 dbuf_will_dirty(db, tx); 833 834 /* create the data buffer for the new block */ 835 buf = arc_buf_alloc(db->db_dnode->dn_objset->os_spa, size, db, type); 836 837 /* copy old block data to the new block */ 838 obuf = db->db_buf; 839 bcopy(obuf->b_data, buf->b_data, MIN(osize, size)); 840 /* zero the remainder */ 841 if (size > osize) 842 bzero((uint8_t *)buf->b_data + osize, size - osize); 843 844 mutex_enter(&db->db_mtx); 845 dbuf_set_data(db, buf); 846 VERIFY(arc_buf_remove_ref(obuf, db) == 1); 847 db->db.db_size = size; 848 849 if (db->db_level == 0) { 850 ASSERT3U(db->db_last_dirty->dr_txg, ==, tx->tx_txg); 851 db->db_last_dirty->dt.dl.dr_data = buf; 852 } 853 mutex_exit(&db->db_mtx); 854 855 dnode_willuse_space(db->db_dnode, size-osize, tx); 856 } 857 858 dbuf_dirty_record_t * 859 dbuf_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 860 { 861 dnode_t *dn = db->db_dnode; 862 objset_impl_t *os = dn->dn_objset; 863 dbuf_dirty_record_t **drp, *dr; 864 int drop_struct_lock = FALSE; 865 int txgoff = tx->tx_txg & TXG_MASK; 866 867 ASSERT(tx->tx_txg != 0); 868 ASSERT(!refcount_is_zero(&db->db_holds)); 869 DMU_TX_DIRTY_BUF(tx, db); 870 871 /* 872 * Shouldn't dirty a regular buffer in syncing context. Private 873 * objects may be dirtied in syncing context, but only if they 874 * were already pre-dirtied in open context. 875 * XXX We may want to prohibit dirtying in syncing context even 876 * if they did pre-dirty. 877 */ 878 ASSERT(!dmu_tx_is_syncing(tx) || 879 BP_IS_HOLE(dn->dn_objset->os_rootbp) || 880 dn->dn_object == DMU_META_DNODE_OBJECT || 881 dn->dn_objset->os_dsl_dataset == NULL || 882 dsl_dir_is_private(dn->dn_objset->os_dsl_dataset->ds_dir)); 883 884 /* 885 * We make this assert for private objects as well, but after we 886 * check if we're already dirty. They are allowed to re-dirty 887 * in syncing context. 888 */ 889 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 890 dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 891 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 892 893 mutex_enter(&db->db_mtx); 894 /* 895 * XXX make this true for indirects too? The problem is that 896 * transactions created with dmu_tx_create_assigned() from 897 * syncing context don't bother holding ahead. 898 */ 899 ASSERT(db->db_level != 0 || 900 db->db_state == DB_CACHED || db->db_state == DB_FILL); 901 902 mutex_enter(&dn->dn_mtx); 903 /* 904 * Don't set dirtyctx to SYNC if we're just modifying this as we 905 * initialize the objset. 906 */ 907 if (dn->dn_dirtyctx == DN_UNDIRTIED && 908 !BP_IS_HOLE(dn->dn_objset->os_rootbp)) { 909 dn->dn_dirtyctx = 910 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN); 911 ASSERT(dn->dn_dirtyctx_firstset == NULL); 912 dn->dn_dirtyctx_firstset = kmem_alloc(1, KM_SLEEP); 913 } 914 mutex_exit(&dn->dn_mtx); 915 916 /* 917 * If this buffer is already dirty, we're done. 918 */ 919 drp = &db->db_last_dirty; 920 ASSERT(*drp == NULL || (*drp)->dr_txg <= tx->tx_txg || 921 db->db.db_object == DMU_META_DNODE_OBJECT); 922 while ((dr = *drp) != NULL && dr->dr_txg > tx->tx_txg) 923 drp = &dr->dr_next; 924 if (dr && dr->dr_txg == tx->tx_txg) { 925 if (db->db_level == 0 && db->db_blkid != DB_BONUS_BLKID) { 926 /* 927 * If this buffer has already been written out, 928 * we now need to reset its state. 929 */ 930 dbuf_unoverride(dr); 931 if (db->db.db_object != DMU_META_DNODE_OBJECT) 932 arc_buf_thaw(db->db_buf); 933 } 934 mutex_exit(&db->db_mtx); 935 return (dr); 936 } 937 938 /* 939 * Only valid if not already dirty. 940 */ 941 ASSERT(dn->dn_dirtyctx == DN_UNDIRTIED || dn->dn_dirtyctx == 942 (dmu_tx_is_syncing(tx) ? DN_DIRTY_SYNC : DN_DIRTY_OPEN)); 943 944 ASSERT3U(dn->dn_nlevels, >, db->db_level); 945 ASSERT((dn->dn_phys->dn_nlevels == 0 && db->db_level == 0) || 946 dn->dn_phys->dn_nlevels > db->db_level || 947 dn->dn_next_nlevels[txgoff] > db->db_level || 948 dn->dn_next_nlevels[(tx->tx_txg-1) & TXG_MASK] > db->db_level || 949 dn->dn_next_nlevels[(tx->tx_txg-2) & TXG_MASK] > db->db_level); 950 951 /* 952 * We should only be dirtying in syncing context if it's the 953 * mos, a spa os, or we're initializing the os. However, we are 954 * allowed to dirty in syncing context provided we already 955 * dirtied it in open context. Hence we must make this 956 * assertion only if we're not already dirty. 957 */ 958 ASSERT(!dmu_tx_is_syncing(tx) || 959 os->os_dsl_dataset == NULL || 960 !dsl_dir_is_private(os->os_dsl_dataset->ds_dir) || 961 !BP_IS_HOLE(os->os_rootbp)); 962 ASSERT(db->db.db_size != 0); 963 964 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 965 966 if (db->db_blkid != DB_BONUS_BLKID) { 967 /* 968 * Update the accounting. 969 */ 970 if (dbuf_block_freeable(db)) { 971 blkptr_t *bp = db->db_blkptr; 972 int64_t willfree = (bp && !BP_IS_HOLE(bp)) ? 973 bp_get_dasize(os->os_spa, bp) : db->db.db_size; 974 /* 975 * This is only a guess -- if the dbuf is dirty 976 * in a previous txg, we don't know how much 977 * space it will use on disk yet. We should 978 * really have the struct_rwlock to access 979 * db_blkptr, but since this is just a guess, 980 * it's OK if we get an odd answer. 981 */ 982 dnode_willuse_space(dn, -willfree, tx); 983 } 984 dnode_willuse_space(dn, db->db.db_size, tx); 985 } 986 987 /* 988 * If this buffer is dirty in an old transaction group we need 989 * to make a copy of it so that the changes we make in this 990 * transaction group won't leak out when we sync the older txg. 991 */ 992 dr = kmem_zalloc(sizeof (dbuf_dirty_record_t), KM_SLEEP); 993 if (db->db_level == 0) { 994 void *data_old = db->db_buf; 995 996 if (db->db_blkid == DB_BONUS_BLKID) { 997 dbuf_fix_old_data(db, tx->tx_txg); 998 data_old = db->db.db_data; 999 } else if (db->db.db_object != DMU_META_DNODE_OBJECT) { 1000 /* 1001 * Release the data buffer from the cache so that we 1002 * can modify it without impacting possible other users 1003 * of this cached data block. Note that indirect 1004 * blocks and private objects are not released until the 1005 * syncing state (since they are only modified then). 1006 */ 1007 arc_release(db->db_buf, db); 1008 dbuf_fix_old_data(db, tx->tx_txg); 1009 data_old = db->db_buf; 1010 } 1011 ASSERT(data_old != NULL); 1012 dr->dt.dl.dr_data = data_old; 1013 } else { 1014 mutex_init(&dr->dt.di.dr_mtx, NULL, MUTEX_DEFAULT, NULL); 1015 list_create(&dr->dt.di.dr_children, 1016 sizeof (dbuf_dirty_record_t), 1017 offsetof(dbuf_dirty_record_t, dr_dirty_node)); 1018 } 1019 dr->dr_dbuf = db; 1020 dr->dr_txg = tx->tx_txg; 1021 dr->dr_next = *drp; 1022 *drp = dr; 1023 1024 /* 1025 * We could have been freed_in_flight between the dbuf_noread 1026 * and dbuf_dirty. We win, as though the dbuf_noread() had 1027 * happened after the free. 1028 */ 1029 if (db->db_level == 0 && db->db_blkid != DB_BONUS_BLKID) { 1030 mutex_enter(&dn->dn_mtx); 1031 dnode_clear_range(dn, db->db_blkid, 1, tx); 1032 mutex_exit(&dn->dn_mtx); 1033 db->db_freed_in_flight = FALSE; 1034 } 1035 1036 /* 1037 * This buffer is now part of this txg 1038 */ 1039 dbuf_add_ref(db, (void *)(uintptr_t)tx->tx_txg); 1040 db->db_dirtycnt += 1; 1041 ASSERT3U(db->db_dirtycnt, <=, 3); 1042 1043 mutex_exit(&db->db_mtx); 1044 1045 if (db->db_blkid == DB_BONUS_BLKID) { 1046 mutex_enter(&dn->dn_mtx); 1047 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1048 list_insert_tail(&dn->dn_dirty_records[txgoff], dr); 1049 mutex_exit(&dn->dn_mtx); 1050 dnode_setdirty(dn, tx); 1051 return (dr); 1052 } 1053 1054 if (db->db_level == 0) { 1055 dnode_new_blkid(dn, db->db_blkid, tx); 1056 ASSERT(dn->dn_maxblkid >= db->db_blkid); 1057 } 1058 1059 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) { 1060 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1061 drop_struct_lock = TRUE; 1062 } 1063 1064 if (db->db_level+1 < dn->dn_nlevels) { 1065 dmu_buf_impl_t *parent = db->db_parent; 1066 dbuf_dirty_record_t *di; 1067 int parent_held = FALSE; 1068 1069 if (db->db_parent == NULL || db->db_parent == dn->dn_dbuf) { 1070 int epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1071 1072 parent = dbuf_hold_level(dn, db->db_level+1, 1073 db->db_blkid >> epbs, FTAG); 1074 parent_held = TRUE; 1075 } 1076 if (drop_struct_lock) 1077 rw_exit(&dn->dn_struct_rwlock); 1078 ASSERT3U(db->db_level+1, ==, parent->db_level); 1079 di = dbuf_dirty(parent, tx); 1080 if (parent_held) 1081 dbuf_rele(parent, FTAG); 1082 1083 mutex_enter(&db->db_mtx); 1084 /* possible race with dbuf_undirty() */ 1085 if (db->db_last_dirty == dr || 1086 dn->dn_object == DMU_META_DNODE_OBJECT) { 1087 mutex_enter(&di->dt.di.dr_mtx); 1088 ASSERT3U(di->dr_txg, ==, tx->tx_txg); 1089 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1090 list_insert_tail(&di->dt.di.dr_children, dr); 1091 mutex_exit(&di->dt.di.dr_mtx); 1092 dr->dr_parent = di; 1093 } 1094 mutex_exit(&db->db_mtx); 1095 } else { 1096 ASSERT(db->db_level+1 == dn->dn_nlevels); 1097 ASSERT(db->db_blkid < dn->dn_nblkptr); 1098 ASSERT(db->db_parent == NULL || 1099 db->db_parent == db->db_dnode->dn_dbuf); 1100 mutex_enter(&dn->dn_mtx); 1101 ASSERT(!list_link_active(&dr->dr_dirty_node)); 1102 list_insert_tail(&dn->dn_dirty_records[txgoff], dr); 1103 mutex_exit(&dn->dn_mtx); 1104 if (drop_struct_lock) 1105 rw_exit(&dn->dn_struct_rwlock); 1106 } 1107 1108 dnode_setdirty(dn, tx); 1109 return (dr); 1110 } 1111 1112 static int 1113 dbuf_undirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1114 { 1115 dnode_t *dn = db->db_dnode; 1116 uint64_t txg = tx->tx_txg; 1117 dbuf_dirty_record_t *dr; 1118 1119 ASSERT(txg != 0); 1120 ASSERT(db->db_blkid != DB_BONUS_BLKID); 1121 1122 mutex_enter(&db->db_mtx); 1123 1124 /* 1125 * If this buffer is not dirty, we're done. 1126 */ 1127 for (dr = db->db_last_dirty; dr; dr = dr->dr_next) 1128 if (dr->dr_txg <= txg) 1129 break; 1130 if (dr == NULL || dr->dr_txg < txg) { 1131 mutex_exit(&db->db_mtx); 1132 return (0); 1133 } 1134 ASSERT(dr->dr_txg == txg); 1135 1136 /* 1137 * If this buffer is currently held, we cannot undirty 1138 * it, since one of the current holders may be in the 1139 * middle of an update. Note that users of dbuf_undirty() 1140 * should not place a hold on the dbuf before the call. 1141 */ 1142 if (refcount_count(&db->db_holds) > db->db_dirtycnt) { 1143 mutex_exit(&db->db_mtx); 1144 /* Make sure we don't toss this buffer at sync phase */ 1145 mutex_enter(&dn->dn_mtx); 1146 dnode_clear_range(dn, db->db_blkid, 1, tx); 1147 mutex_exit(&dn->dn_mtx); 1148 return (0); 1149 } 1150 1151 dprintf_dbuf(db, "size=%llx\n", (u_longlong_t)db->db.db_size); 1152 1153 ASSERT(db->db.db_size != 0); 1154 1155 /* XXX would be nice to fix up dn_towrite_space[] */ 1156 1157 db->db_last_dirty = dr->dr_next; 1158 1159 if (dr->dr_parent) { 1160 mutex_enter(&dr->dr_parent->dt.di.dr_mtx); 1161 list_remove(&dr->dr_parent->dt.di.dr_children, dr); 1162 mutex_exit(&dr->dr_parent->dt.di.dr_mtx); 1163 } else if (db->db_level+1 == dn->dn_nlevels) { 1164 ASSERT3P(db->db_parent, ==, dn->dn_dbuf); 1165 mutex_enter(&dn->dn_mtx); 1166 list_remove(&dn->dn_dirty_records[txg & TXG_MASK], dr); 1167 mutex_exit(&dn->dn_mtx); 1168 } 1169 1170 if (db->db_level == 0) { 1171 dbuf_unoverride(dr); 1172 1173 ASSERT(db->db_buf != NULL); 1174 ASSERT(dr->dt.dl.dr_data != NULL); 1175 if (dr->dt.dl.dr_data != db->db_buf) 1176 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db) == 1); 1177 } else { 1178 ASSERT(db->db_buf != NULL); 1179 ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 1180 mutex_destroy(&dr->dt.di.dr_mtx); 1181 list_destroy(&dr->dt.di.dr_children); 1182 } 1183 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 1184 1185 ASSERT(db->db_dirtycnt > 0); 1186 db->db_dirtycnt -= 1; 1187 1188 if (refcount_remove(&db->db_holds, (void *)(uintptr_t)txg) == 0) { 1189 arc_buf_t *buf = db->db_buf; 1190 1191 ASSERT(arc_released(buf)); 1192 dbuf_set_data(db, NULL); 1193 VERIFY(arc_buf_remove_ref(buf, db) == 1); 1194 dbuf_evict(db); 1195 return (1); 1196 } 1197 1198 mutex_exit(&db->db_mtx); 1199 return (0); 1200 } 1201 1202 #pragma weak dmu_buf_will_dirty = dbuf_will_dirty 1203 void 1204 dbuf_will_dirty(dmu_buf_impl_t *db, dmu_tx_t *tx) 1205 { 1206 int rf = DB_RF_MUST_SUCCEED; 1207 1208 ASSERT(tx->tx_txg != 0); 1209 ASSERT(!refcount_is_zero(&db->db_holds)); 1210 1211 if (RW_WRITE_HELD(&db->db_dnode->dn_struct_rwlock)) 1212 rf |= DB_RF_HAVESTRUCT; 1213 (void) dbuf_read(db, NULL, rf); 1214 (void) dbuf_dirty(db, tx); 1215 } 1216 1217 void 1218 dmu_buf_will_fill(dmu_buf_t *db_fake, dmu_tx_t *tx) 1219 { 1220 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1221 1222 ASSERT(db->db_blkid != DB_BONUS_BLKID); 1223 ASSERT(tx->tx_txg != 0); 1224 ASSERT(db->db_level == 0); 1225 ASSERT(!refcount_is_zero(&db->db_holds)); 1226 1227 ASSERT(db->db.db_object != DMU_META_DNODE_OBJECT || 1228 dmu_tx_private_ok(tx)); 1229 1230 dbuf_noread(db); 1231 (void) dbuf_dirty(db, tx); 1232 } 1233 1234 #pragma weak dmu_buf_fill_done = dbuf_fill_done 1235 /* ARGSUSED */ 1236 void 1237 dbuf_fill_done(dmu_buf_impl_t *db, dmu_tx_t *tx) 1238 { 1239 mutex_enter(&db->db_mtx); 1240 DBUF_VERIFY(db); 1241 1242 if (db->db_state == DB_FILL) { 1243 if (db->db_level == 0 && db->db_freed_in_flight) { 1244 ASSERT(db->db_blkid != DB_BONUS_BLKID); 1245 /* we were freed while filling */ 1246 /* XXX dbuf_undirty? */ 1247 bzero(db->db.db_data, db->db.db_size); 1248 db->db_freed_in_flight = FALSE; 1249 } 1250 db->db_state = DB_CACHED; 1251 cv_broadcast(&db->db_changed); 1252 } 1253 mutex_exit(&db->db_mtx); 1254 } 1255 1256 /* 1257 * "Clear" the contents of this dbuf. This will mark the dbuf 1258 * EVICTING and clear *most* of its references. Unfortunetely, 1259 * when we are not holding the dn_dbufs_mtx, we can't clear the 1260 * entry in the dn_dbufs list. We have to wait until dbuf_destroy() 1261 * in this case. For callers from the DMU we will usually see: 1262 * dbuf_clear()->arc_buf_evict()->dbuf_do_evict()->dbuf_destroy() 1263 * For the arc callback, we will usually see: 1264 * dbuf_do_evict()->dbuf_clear();dbuf_destroy() 1265 * Sometimes, though, we will get a mix of these two: 1266 * DMU: dbuf_clear()->arc_buf_evict() 1267 * ARC: dbuf_do_evict()->dbuf_destroy() 1268 */ 1269 void 1270 dbuf_clear(dmu_buf_impl_t *db) 1271 { 1272 dnode_t *dn = db->db_dnode; 1273 dmu_buf_impl_t *parent = db->db_parent; 1274 dmu_buf_impl_t *dndb = dn->dn_dbuf; 1275 int dbuf_gone = FALSE; 1276 1277 ASSERT(MUTEX_HELD(&db->db_mtx)); 1278 ASSERT(refcount_is_zero(&db->db_holds)); 1279 1280 dbuf_evict_user(db); 1281 1282 if (db->db_state == DB_CACHED) { 1283 ASSERT(db->db.db_data != NULL); 1284 if (db->db_blkid == DB_BONUS_BLKID) { 1285 zio_buf_free(db->db.db_data, DN_MAX_BONUSLEN); 1286 arc_space_return(DN_MAX_BONUSLEN); 1287 } 1288 db->db.db_data = NULL; 1289 db->db_state = DB_UNCACHED; 1290 } 1291 1292 ASSERT3U(db->db_state, ==, DB_UNCACHED); 1293 ASSERT(db->db_data_pending == NULL); 1294 1295 db->db_state = DB_EVICTING; 1296 db->db_blkptr = NULL; 1297 1298 if (db->db_blkid != DB_BONUS_BLKID && MUTEX_HELD(&dn->dn_dbufs_mtx)) { 1299 list_remove(&dn->dn_dbufs, db); 1300 dnode_rele(dn, db); 1301 db->db_dnode = NULL; 1302 } 1303 1304 if (db->db_buf) 1305 dbuf_gone = arc_buf_evict(db->db_buf); 1306 1307 if (!dbuf_gone) 1308 mutex_exit(&db->db_mtx); 1309 1310 /* 1311 * If this dbuf is referened from an indirect dbuf, 1312 * decrement the ref count on the indirect dbuf. 1313 */ 1314 if (parent && parent != dndb) 1315 dbuf_rele(parent, db); 1316 } 1317 1318 static int 1319 dbuf_findbp(dnode_t *dn, int level, uint64_t blkid, int fail_sparse, 1320 dmu_buf_impl_t **parentp, blkptr_t **bpp) 1321 { 1322 int nlevels, epbs; 1323 1324 *parentp = NULL; 1325 *bpp = NULL; 1326 1327 ASSERT(blkid != DB_BONUS_BLKID); 1328 1329 if (dn->dn_phys->dn_nlevels == 0) 1330 nlevels = 1; 1331 else 1332 nlevels = dn->dn_phys->dn_nlevels; 1333 1334 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1335 1336 ASSERT3U(level * epbs, <, 64); 1337 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1338 if (level >= nlevels || 1339 (blkid > (dn->dn_phys->dn_maxblkid >> (level * epbs)))) { 1340 /* the buffer has no parent yet */ 1341 return (ENOENT); 1342 } else if (level < nlevels-1) { 1343 /* this block is referenced from an indirect block */ 1344 int err = dbuf_hold_impl(dn, level+1, 1345 blkid >> epbs, fail_sparse, NULL, parentp); 1346 if (err) 1347 return (err); 1348 err = dbuf_read(*parentp, NULL, 1349 (DB_RF_HAVESTRUCT | DB_RF_NOPREFETCH | DB_RF_CANFAIL)); 1350 if (err) { 1351 dbuf_rele(*parentp, NULL); 1352 *parentp = NULL; 1353 return (err); 1354 } 1355 *bpp = ((blkptr_t *)(*parentp)->db.db_data) + 1356 (blkid & ((1ULL << epbs) - 1)); 1357 return (0); 1358 } else { 1359 /* the block is referenced from the dnode */ 1360 ASSERT3U(level, ==, nlevels-1); 1361 ASSERT(dn->dn_phys->dn_nblkptr == 0 || 1362 blkid < dn->dn_phys->dn_nblkptr); 1363 if (dn->dn_dbuf) { 1364 dbuf_add_ref(dn->dn_dbuf, NULL); 1365 *parentp = dn->dn_dbuf; 1366 } 1367 *bpp = &dn->dn_phys->dn_blkptr[blkid]; 1368 return (0); 1369 } 1370 } 1371 1372 static dmu_buf_impl_t * 1373 dbuf_create(dnode_t *dn, uint8_t level, uint64_t blkid, 1374 dmu_buf_impl_t *parent, blkptr_t *blkptr) 1375 { 1376 objset_impl_t *os = dn->dn_objset; 1377 dmu_buf_impl_t *db, *odb; 1378 1379 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1380 ASSERT(dn->dn_type != DMU_OT_NONE); 1381 1382 db = kmem_cache_alloc(dbuf_cache, KM_SLEEP); 1383 1384 db->db_objset = os; 1385 db->db.db_object = dn->dn_object; 1386 db->db_level = level; 1387 db->db_blkid = blkid; 1388 db->db_last_dirty = NULL; 1389 db->db_dirtycnt = 0; 1390 db->db_dnode = dn; 1391 db->db_parent = parent; 1392 db->db_blkptr = blkptr; 1393 1394 db->db_user_ptr = NULL; 1395 db->db_user_data_ptr_ptr = NULL; 1396 db->db_evict_func = NULL; 1397 db->db_immediate_evict = 0; 1398 db->db_freed_in_flight = 0; 1399 1400 if (blkid == DB_BONUS_BLKID) { 1401 ASSERT3P(parent, ==, dn->dn_dbuf); 1402 db->db.db_size = DN_MAX_BONUSLEN - 1403 (dn->dn_nblkptr-1) * sizeof (blkptr_t); 1404 ASSERT3U(db->db.db_size, >=, dn->dn_bonuslen); 1405 db->db.db_offset = DB_BONUS_BLKID; 1406 db->db_state = DB_UNCACHED; 1407 /* the bonus dbuf is not placed in the hash table */ 1408 arc_space_consume(sizeof (dmu_buf_impl_t)); 1409 return (db); 1410 } else { 1411 int blocksize = 1412 db->db_level ? 1<<dn->dn_indblkshift : dn->dn_datablksz; 1413 db->db.db_size = blocksize; 1414 db->db.db_offset = db->db_blkid * blocksize; 1415 } 1416 1417 /* 1418 * Hold the dn_dbufs_mtx while we get the new dbuf 1419 * in the hash table *and* added to the dbufs list. 1420 * This prevents a possible deadlock with someone 1421 * trying to look up this dbuf before its added to the 1422 * dn_dbufs list. 1423 */ 1424 mutex_enter(&dn->dn_dbufs_mtx); 1425 db->db_state = DB_EVICTING; 1426 if ((odb = dbuf_hash_insert(db)) != NULL) { 1427 /* someone else inserted it first */ 1428 kmem_cache_free(dbuf_cache, db); 1429 mutex_exit(&dn->dn_dbufs_mtx); 1430 return (odb); 1431 } 1432 list_insert_head(&dn->dn_dbufs, db); 1433 db->db_state = DB_UNCACHED; 1434 mutex_exit(&dn->dn_dbufs_mtx); 1435 arc_space_consume(sizeof (dmu_buf_impl_t)); 1436 1437 if (parent && parent != dn->dn_dbuf) 1438 dbuf_add_ref(parent, db); 1439 1440 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || 1441 refcount_count(&dn->dn_holds) > 0); 1442 (void) refcount_add(&dn->dn_holds, db); 1443 1444 dprintf_dbuf(db, "db=%p\n", db); 1445 1446 return (db); 1447 } 1448 1449 static int 1450 dbuf_do_evict(void *private) 1451 { 1452 arc_buf_t *buf = private; 1453 dmu_buf_impl_t *db = buf->b_private; 1454 1455 if (!MUTEX_HELD(&db->db_mtx)) 1456 mutex_enter(&db->db_mtx); 1457 1458 ASSERT(refcount_is_zero(&db->db_holds)); 1459 1460 if (db->db_state != DB_EVICTING) { 1461 ASSERT(db->db_state == DB_CACHED); 1462 DBUF_VERIFY(db); 1463 db->db_buf = NULL; 1464 dbuf_evict(db); 1465 } else { 1466 mutex_exit(&db->db_mtx); 1467 dbuf_destroy(db); 1468 } 1469 return (0); 1470 } 1471 1472 static void 1473 dbuf_destroy(dmu_buf_impl_t *db) 1474 { 1475 ASSERT(refcount_is_zero(&db->db_holds)); 1476 1477 if (db->db_blkid != DB_BONUS_BLKID) { 1478 /* 1479 * If this dbuf is still on the dn_dbufs list, 1480 * remove it from that list. 1481 */ 1482 if (db->db_dnode) { 1483 dnode_t *dn = db->db_dnode; 1484 1485 mutex_enter(&dn->dn_dbufs_mtx); 1486 list_remove(&dn->dn_dbufs, db); 1487 mutex_exit(&dn->dn_dbufs_mtx); 1488 1489 dnode_rele(dn, db); 1490 db->db_dnode = NULL; 1491 } 1492 dbuf_hash_remove(db); 1493 } 1494 db->db_parent = NULL; 1495 db->db_buf = NULL; 1496 1497 ASSERT(!list_link_active(&db->db_link)); 1498 ASSERT(db->db.db_data == NULL); 1499 ASSERT(db->db_hash_next == NULL); 1500 ASSERT(db->db_blkptr == NULL); 1501 ASSERT(db->db_data_pending == NULL); 1502 1503 kmem_cache_free(dbuf_cache, db); 1504 arc_space_return(sizeof (dmu_buf_impl_t)); 1505 } 1506 1507 void 1508 dbuf_prefetch(dnode_t *dn, uint64_t blkid) 1509 { 1510 dmu_buf_impl_t *db = NULL; 1511 blkptr_t *bp = NULL; 1512 1513 ASSERT(blkid != DB_BONUS_BLKID); 1514 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1515 1516 if (dnode_block_freed(dn, blkid)) 1517 return; 1518 1519 /* dbuf_find() returns with db_mtx held */ 1520 if (db = dbuf_find(dn, 0, blkid)) { 1521 if (refcount_count(&db->db_holds) > 0) { 1522 /* 1523 * This dbuf is active. We assume that it is 1524 * already CACHED, or else about to be either 1525 * read or filled. 1526 */ 1527 mutex_exit(&db->db_mtx); 1528 return; 1529 } 1530 mutex_exit(&db->db_mtx); 1531 db = NULL; 1532 } 1533 1534 if (dbuf_findbp(dn, 0, blkid, TRUE, &db, &bp) == 0) { 1535 if (bp && !BP_IS_HOLE(bp)) { 1536 uint32_t aflags = ARC_NOWAIT | ARC_PREFETCH; 1537 zbookmark_t zb; 1538 zb.zb_objset = dn->dn_objset->os_dsl_dataset ? 1539 dn->dn_objset->os_dsl_dataset->ds_object : 0; 1540 zb.zb_object = dn->dn_object; 1541 zb.zb_level = 0; 1542 zb.zb_blkid = blkid; 1543 1544 (void) arc_read(NULL, dn->dn_objset->os_spa, bp, 1545 dmu_ot[dn->dn_type].ot_byteswap, 1546 NULL, NULL, ZIO_PRIORITY_ASYNC_READ, 1547 ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, 1548 &aflags, &zb); 1549 } 1550 if (db) 1551 dbuf_rele(db, NULL); 1552 } 1553 } 1554 1555 /* 1556 * Returns with db_holds incremented, and db_mtx not held. 1557 * Note: dn_struct_rwlock must be held. 1558 */ 1559 int 1560 dbuf_hold_impl(dnode_t *dn, uint8_t level, uint64_t blkid, int fail_sparse, 1561 void *tag, dmu_buf_impl_t **dbp) 1562 { 1563 dmu_buf_impl_t *db, *parent = NULL; 1564 1565 ASSERT(blkid != DB_BONUS_BLKID); 1566 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 1567 ASSERT3U(dn->dn_nlevels, >, level); 1568 1569 *dbp = NULL; 1570 top: 1571 /* dbuf_find() returns with db_mtx held */ 1572 db = dbuf_find(dn, level, blkid); 1573 1574 if (db == NULL) { 1575 blkptr_t *bp = NULL; 1576 int err; 1577 1578 ASSERT3P(parent, ==, NULL); 1579 err = dbuf_findbp(dn, level, blkid, fail_sparse, &parent, &bp); 1580 if (fail_sparse) { 1581 if (err == 0 && bp && BP_IS_HOLE(bp)) 1582 err = ENOENT; 1583 if (err) { 1584 if (parent) 1585 dbuf_rele(parent, NULL); 1586 return (err); 1587 } 1588 } 1589 if (err && err != ENOENT) 1590 return (err); 1591 db = dbuf_create(dn, level, blkid, parent, bp); 1592 } 1593 1594 if (db->db_buf && refcount_is_zero(&db->db_holds)) { 1595 arc_buf_add_ref(db->db_buf, db); 1596 if (db->db_buf->b_data == NULL) { 1597 dbuf_clear(db); 1598 if (parent) { 1599 dbuf_rele(parent, NULL); 1600 parent = NULL; 1601 } 1602 goto top; 1603 } 1604 ASSERT3P(db->db.db_data, ==, db->db_buf->b_data); 1605 } 1606 1607 ASSERT(db->db_buf == NULL || arc_referenced(db->db_buf)); 1608 1609 /* 1610 * If this buffer is currently syncing out, and we are are 1611 * still referencing it from db_data, we need to make a copy 1612 * of it in case we decide we want to dirty it again in this txg. 1613 */ 1614 if (db->db_level == 0 && db->db_blkid != DB_BONUS_BLKID && 1615 dn->dn_object != DMU_META_DNODE_OBJECT && 1616 db->db_state == DB_CACHED && db->db_data_pending) { 1617 dbuf_dirty_record_t *dr = db->db_data_pending; 1618 1619 if (dr->dt.dl.dr_data == db->db_buf) { 1620 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 1621 1622 dbuf_set_data(db, 1623 arc_buf_alloc(db->db_dnode->dn_objset->os_spa, 1624 db->db.db_size, db, type)); 1625 bcopy(dr->dt.dl.dr_data->b_data, db->db.db_data, 1626 db->db.db_size); 1627 } 1628 } 1629 1630 (void) refcount_add(&db->db_holds, tag); 1631 dbuf_update_data(db); 1632 DBUF_VERIFY(db); 1633 mutex_exit(&db->db_mtx); 1634 1635 /* NOTE: we can't rele the parent until after we drop the db_mtx */ 1636 if (parent) 1637 dbuf_rele(parent, NULL); 1638 1639 ASSERT3P(db->db_dnode, ==, dn); 1640 ASSERT3U(db->db_blkid, ==, blkid); 1641 ASSERT3U(db->db_level, ==, level); 1642 *dbp = db; 1643 1644 return (0); 1645 } 1646 1647 dmu_buf_impl_t * 1648 dbuf_hold(dnode_t *dn, uint64_t blkid, void *tag) 1649 { 1650 dmu_buf_impl_t *db; 1651 int err = dbuf_hold_impl(dn, 0, blkid, FALSE, tag, &db); 1652 return (err ? NULL : db); 1653 } 1654 1655 dmu_buf_impl_t * 1656 dbuf_hold_level(dnode_t *dn, int level, uint64_t blkid, void *tag) 1657 { 1658 dmu_buf_impl_t *db; 1659 int err = dbuf_hold_impl(dn, level, blkid, FALSE, tag, &db); 1660 return (err ? NULL : db); 1661 } 1662 1663 void 1664 dbuf_create_bonus(dnode_t *dn) 1665 { 1666 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 1667 1668 ASSERT(dn->dn_bonus == NULL); 1669 dn->dn_bonus = dbuf_create(dn, 0, DB_BONUS_BLKID, dn->dn_dbuf, NULL); 1670 } 1671 1672 #pragma weak dmu_buf_add_ref = dbuf_add_ref 1673 void 1674 dbuf_add_ref(dmu_buf_impl_t *db, void *tag) 1675 { 1676 int64_t holds = refcount_add(&db->db_holds, tag); 1677 ASSERT(holds > 1); 1678 } 1679 1680 #pragma weak dmu_buf_rele = dbuf_rele 1681 void 1682 dbuf_rele(dmu_buf_impl_t *db, void *tag) 1683 { 1684 int64_t holds; 1685 1686 mutex_enter(&db->db_mtx); 1687 DBUF_VERIFY(db); 1688 1689 holds = refcount_remove(&db->db_holds, tag); 1690 ASSERT(holds >= 0); 1691 1692 /* 1693 * We can't freeze indirects if there is a possibility that they 1694 * may be modified in the current syncing context. 1695 */ 1696 if (db->db_buf && holds == (db->db_level == 0 ? db->db_dirtycnt : 0)) 1697 arc_buf_freeze(db->db_buf); 1698 1699 if (holds == db->db_dirtycnt && 1700 db->db_level == 0 && db->db_immediate_evict) 1701 dbuf_evict_user(db); 1702 1703 if (holds == 0) { 1704 if (db->db_blkid == DB_BONUS_BLKID) { 1705 mutex_exit(&db->db_mtx); 1706 dnode_rele(db->db_dnode, db); 1707 } else if (db->db_buf == NULL) { 1708 /* 1709 * This is a special case: we never associated this 1710 * dbuf with any data allocated from the ARC. 1711 */ 1712 ASSERT3U(db->db_state, ==, DB_UNCACHED); 1713 dbuf_evict(db); 1714 } else if (arc_released(db->db_buf)) { 1715 arc_buf_t *buf = db->db_buf; 1716 /* 1717 * This dbuf has anonymous data associated with it. 1718 */ 1719 dbuf_set_data(db, NULL); 1720 VERIFY(arc_buf_remove_ref(buf, db) == 1); 1721 dbuf_evict(db); 1722 } else { 1723 VERIFY(arc_buf_remove_ref(db->db_buf, db) == 0); 1724 mutex_exit(&db->db_mtx); 1725 } 1726 } else { 1727 mutex_exit(&db->db_mtx); 1728 } 1729 } 1730 1731 #pragma weak dmu_buf_refcount = dbuf_refcount 1732 uint64_t 1733 dbuf_refcount(dmu_buf_impl_t *db) 1734 { 1735 return (refcount_count(&db->db_holds)); 1736 } 1737 1738 void * 1739 dmu_buf_set_user(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr, 1740 dmu_buf_evict_func_t *evict_func) 1741 { 1742 return (dmu_buf_update_user(db_fake, NULL, user_ptr, 1743 user_data_ptr_ptr, evict_func)); 1744 } 1745 1746 void * 1747 dmu_buf_set_user_ie(dmu_buf_t *db_fake, void *user_ptr, void *user_data_ptr_ptr, 1748 dmu_buf_evict_func_t *evict_func) 1749 { 1750 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1751 1752 db->db_immediate_evict = TRUE; 1753 return (dmu_buf_update_user(db_fake, NULL, user_ptr, 1754 user_data_ptr_ptr, evict_func)); 1755 } 1756 1757 void * 1758 dmu_buf_update_user(dmu_buf_t *db_fake, void *old_user_ptr, void *user_ptr, 1759 void *user_data_ptr_ptr, dmu_buf_evict_func_t *evict_func) 1760 { 1761 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1762 ASSERT(db->db_level == 0); 1763 1764 ASSERT((user_ptr == NULL) == (evict_func == NULL)); 1765 1766 mutex_enter(&db->db_mtx); 1767 1768 if (db->db_user_ptr == old_user_ptr) { 1769 db->db_user_ptr = user_ptr; 1770 db->db_user_data_ptr_ptr = user_data_ptr_ptr; 1771 db->db_evict_func = evict_func; 1772 1773 dbuf_update_data(db); 1774 } else { 1775 old_user_ptr = db->db_user_ptr; 1776 } 1777 1778 mutex_exit(&db->db_mtx); 1779 return (old_user_ptr); 1780 } 1781 1782 void * 1783 dmu_buf_get_user(dmu_buf_t *db_fake) 1784 { 1785 dmu_buf_impl_t *db = (dmu_buf_impl_t *)db_fake; 1786 ASSERT(!refcount_is_zero(&db->db_holds)); 1787 1788 return (db->db_user_ptr); 1789 } 1790 1791 static void 1792 dbuf_check_blkptr(dnode_t *dn, dmu_buf_impl_t *db) 1793 { 1794 /* ASSERT(dmu_tx_is_syncing(tx) */ 1795 ASSERT(MUTEX_HELD(&db->db_mtx)); 1796 1797 if (db->db_blkptr != NULL) 1798 return; 1799 1800 if (db->db_level == dn->dn_phys->dn_nlevels-1) { 1801 /* 1802 * This buffer was allocated at a time when there was 1803 * no available blkptrs from the dnode, or it was 1804 * inappropriate to hook it in (i.e., nlevels mis-match). 1805 */ 1806 ASSERT(db->db_blkid < dn->dn_phys->dn_nblkptr); 1807 ASSERT(db->db_parent == NULL); 1808 db->db_parent = dn->dn_dbuf; 1809 db->db_blkptr = &dn->dn_phys->dn_blkptr[db->db_blkid]; 1810 DBUF_VERIFY(db); 1811 } else { 1812 dmu_buf_impl_t *parent = db->db_parent; 1813 int epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 1814 1815 ASSERT(dn->dn_phys->dn_nlevels > 1); 1816 if (parent == NULL) { 1817 mutex_exit(&db->db_mtx); 1818 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1819 (void) dbuf_hold_impl(dn, db->db_level+1, 1820 db->db_blkid >> epbs, FALSE, db, &parent); 1821 rw_exit(&dn->dn_struct_rwlock); 1822 mutex_enter(&db->db_mtx); 1823 db->db_parent = parent; 1824 } 1825 db->db_blkptr = (blkptr_t *)parent->db.db_data + 1826 (db->db_blkid & ((1ULL << epbs) - 1)); 1827 DBUF_VERIFY(db); 1828 } 1829 } 1830 1831 static void 1832 dbuf_sync_indirect(dbuf_dirty_record_t *dr, dmu_tx_t *tx) 1833 { 1834 dmu_buf_impl_t *db = dr->dr_dbuf; 1835 dnode_t *dn = db->db_dnode; 1836 zio_t *zio; 1837 1838 ASSERT(dmu_tx_is_syncing(tx)); 1839 1840 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 1841 1842 mutex_enter(&db->db_mtx); 1843 1844 ASSERT(db->db_level > 0); 1845 DBUF_VERIFY(db); 1846 1847 if (db->db_buf == NULL) { 1848 mutex_exit(&db->db_mtx); 1849 (void) dbuf_read(db, NULL, DB_RF_MUST_SUCCEED); 1850 mutex_enter(&db->db_mtx); 1851 } 1852 ASSERT3U(db->db_state, ==, DB_CACHED); 1853 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 1854 ASSERT(db->db_buf != NULL); 1855 1856 dbuf_check_blkptr(dn, db); 1857 1858 db->db_data_pending = dr; 1859 1860 arc_release(db->db_buf, db); 1861 mutex_exit(&db->db_mtx); 1862 1863 /* 1864 * XXX -- we should design a compression algorithm 1865 * that specializes in arrays of bps. 1866 */ 1867 dbuf_write(dr, db->db_buf, ZIO_CHECKSUM_FLETCHER_4, 1868 zfs_mdcomp_disable ? ZIO_COMPRESS_EMPTY : ZIO_COMPRESS_LZJB, tx); 1869 1870 zio = dr->dr_zio; 1871 mutex_enter(&dr->dt.di.dr_mtx); 1872 dbuf_sync_list(&dr->dt.di.dr_children, tx); 1873 ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 1874 mutex_exit(&dr->dt.di.dr_mtx); 1875 zio_nowait(zio); 1876 } 1877 1878 static void 1879 dbuf_sync_leaf(dbuf_dirty_record_t *dr, dmu_tx_t *tx) 1880 { 1881 arc_buf_t **datap = &dr->dt.dl.dr_data; 1882 dmu_buf_impl_t *db = dr->dr_dbuf; 1883 dnode_t *dn = db->db_dnode; 1884 objset_impl_t *os = dn->dn_objset; 1885 uint64_t txg = tx->tx_txg; 1886 int checksum, compress; 1887 int blksz; 1888 1889 ASSERT(dmu_tx_is_syncing(tx)); 1890 1891 dprintf_dbuf_bp(db, db->db_blkptr, "blkptr=%p", db->db_blkptr); 1892 1893 mutex_enter(&db->db_mtx); 1894 /* 1895 * To be synced, we must be dirtied. But we 1896 * might have been freed after the dirty. 1897 */ 1898 if (db->db_state == DB_UNCACHED) { 1899 /* This buffer has been freed since it was dirtied */ 1900 ASSERT(db->db.db_data == NULL); 1901 } else if (db->db_state == DB_FILL) { 1902 /* This buffer was freed and is now being re-filled */ 1903 ASSERT(db->db.db_data != dr->dt.dl.dr_data); 1904 } else { 1905 ASSERT3U(db->db_state, ==, DB_CACHED); 1906 } 1907 DBUF_VERIFY(db); 1908 1909 /* 1910 * If this is a bonus buffer, simply copy the bonus data into the 1911 * dnode. It will be written out when the dnode is synced (and it 1912 * will be synced, since it must have been dirty for dbuf_sync to 1913 * be called). 1914 */ 1915 if (db->db_blkid == DB_BONUS_BLKID) { 1916 dbuf_dirty_record_t **drp; 1917 1918 ASSERT(*datap != NULL); 1919 ASSERT3U(db->db_level, ==, 0); 1920 ASSERT3U(dn->dn_phys->dn_bonuslen, <=, DN_MAX_BONUSLEN); 1921 bcopy(*datap, DN_BONUS(dn->dn_phys), dn->dn_phys->dn_bonuslen); 1922 if (*datap != db->db.db_data) { 1923 zio_buf_free(*datap, DN_MAX_BONUSLEN); 1924 arc_space_return(DN_MAX_BONUSLEN); 1925 } 1926 db->db_data_pending = NULL; 1927 drp = &db->db_last_dirty; 1928 while (*drp != dr) 1929 drp = &(*drp)->dr_next; 1930 ASSERT((*drp)->dr_next == NULL); 1931 *drp = NULL; 1932 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 1933 ASSERT(db->db_dirtycnt > 0); 1934 db->db_dirtycnt -= 1; 1935 mutex_exit(&db->db_mtx); 1936 dbuf_rele(db, (void *)(uintptr_t)txg); 1937 return; 1938 } 1939 1940 /* 1941 * This function may have dropped the db_mtx lock allowing a dmu_sync 1942 * operation to sneak in. As a result, we need to ensure that we 1943 * don't check the dr_override_state until we have returned from 1944 * dbuf_check_blkptr. 1945 */ 1946 dbuf_check_blkptr(dn, db); 1947 1948 /* 1949 * If this buffer is in the middle of an immdiate write, 1950 * wait for the synchronous IO to complete. 1951 */ 1952 while (dr->dt.dl.dr_override_state == DR_IN_DMU_SYNC) { 1953 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT); 1954 cv_wait(&db->db_changed, &db->db_mtx); 1955 ASSERT(dr->dt.dl.dr_override_state != DR_NOT_OVERRIDDEN); 1956 } 1957 1958 /* 1959 * If this dbuf has already been written out via an immediate write, 1960 * just complete the write by copying over the new block pointer and 1961 * updating the accounting via the write-completion functions. 1962 */ 1963 if (dr->dt.dl.dr_override_state == DR_OVERRIDDEN) { 1964 zio_t zio_fake; 1965 1966 zio_fake.io_private = &db; 1967 zio_fake.io_error = 0; 1968 zio_fake.io_bp = db->db_blkptr; 1969 zio_fake.io_bp_orig = *db->db_blkptr; 1970 zio_fake.io_txg = txg; 1971 1972 *db->db_blkptr = dr->dt.dl.dr_overridden_by; 1973 dr->dt.dl.dr_override_state = DR_NOT_OVERRIDDEN; 1974 db->db_data_pending = dr; 1975 dr->dr_zio = &zio_fake; 1976 mutex_exit(&db->db_mtx); 1977 1978 if (BP_IS_OLDER(&zio_fake.io_bp_orig, txg)) 1979 dsl_dataset_block_kill(os->os_dsl_dataset, 1980 &zio_fake.io_bp_orig, dn->dn_zio, tx); 1981 1982 dbuf_write_ready(&zio_fake, db->db_buf, db); 1983 dbuf_write_done(&zio_fake, db->db_buf, db); 1984 1985 return; 1986 } 1987 1988 blksz = arc_buf_size(*datap); 1989 1990 if (dn->dn_object != DMU_META_DNODE_OBJECT) { 1991 /* 1992 * If this buffer is currently "in use" (i.e., there are 1993 * active holds and db_data still references it), then make 1994 * a copy before we start the write so that any modifications 1995 * from the open txg will not leak into this write. 1996 * 1997 * NOTE: this copy does not need to be made for objects only 1998 * modified in the syncing context (e.g. DNONE_DNODE blocks). 1999 */ 2000 if (refcount_count(&db->db_holds) > 1 && *datap == db->db_buf) { 2001 arc_buf_contents_t type = DBUF_GET_BUFC_TYPE(db); 2002 *datap = arc_buf_alloc(os->os_spa, blksz, db, type); 2003 bcopy(db->db.db_data, (*datap)->b_data, blksz); 2004 } 2005 } else { 2006 /* 2007 * Private object buffers are released here rather 2008 * than in dbuf_dirty() since they are only modified 2009 * in the syncing context and we don't want the 2010 * overhead of making multiple copies of the data. 2011 */ 2012 arc_release(db->db_buf, db); 2013 } 2014 2015 ASSERT(*datap != NULL); 2016 db->db_data_pending = dr; 2017 2018 mutex_exit(&db->db_mtx); 2019 2020 /* 2021 * Allow dnode settings to override objset settings, 2022 * except for metadata checksums. 2023 */ 2024 if (dmu_ot[dn->dn_type].ot_metadata) { 2025 checksum = os->os_md_checksum; 2026 compress = zio_compress_select(dn->dn_compress, 2027 os->os_md_compress); 2028 } else { 2029 checksum = zio_checksum_select(dn->dn_checksum, 2030 os->os_checksum); 2031 compress = zio_compress_select(dn->dn_compress, 2032 os->os_compress); 2033 } 2034 2035 dbuf_write(dr, *datap, checksum, compress, tx); 2036 2037 ASSERT(!list_link_active(&dr->dr_dirty_node)); 2038 if (dn->dn_object == DMU_META_DNODE_OBJECT) 2039 list_insert_tail(&dn->dn_dirty_records[txg&TXG_MASK], dr); 2040 else 2041 zio_nowait(dr->dr_zio); 2042 } 2043 2044 void 2045 dbuf_sync_list(list_t *list, dmu_tx_t *tx) 2046 { 2047 dbuf_dirty_record_t *dr; 2048 2049 while (dr = list_head(list)) { 2050 if (dr->dr_zio != NULL) { 2051 /* 2052 * If we find an already initialized zio then we 2053 * are processing the meta-dnode, and we have finished. 2054 * The dbufs for all dnodes are put back on the list 2055 * during processing, so that we can zio_wait() 2056 * these IOs after initiating all child IOs. 2057 */ 2058 ASSERT3U(dr->dr_dbuf->db.db_object, ==, 2059 DMU_META_DNODE_OBJECT); 2060 break; 2061 } 2062 list_remove(list, dr); 2063 if (dr->dr_dbuf->db_level > 0) 2064 dbuf_sync_indirect(dr, tx); 2065 else 2066 dbuf_sync_leaf(dr, tx); 2067 } 2068 } 2069 2070 static void 2071 dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, int checksum, 2072 int compress, dmu_tx_t *tx) 2073 { 2074 dmu_buf_impl_t *db = dr->dr_dbuf; 2075 dnode_t *dn = db->db_dnode; 2076 objset_impl_t *os = dn->dn_objset; 2077 dmu_buf_impl_t *parent = db->db_parent; 2078 uint64_t txg = tx->tx_txg; 2079 zbookmark_t zb; 2080 zio_t *zio; 2081 int zio_flags; 2082 2083 if (parent != dn->dn_dbuf) { 2084 ASSERT(parent && parent->db_data_pending); 2085 ASSERT(db->db_level == parent->db_level-1); 2086 ASSERT(arc_released(parent->db_buf)); 2087 zio = parent->db_data_pending->dr_zio; 2088 } else { 2089 ASSERT(db->db_level == dn->dn_phys->dn_nlevels-1); 2090 ASSERT3P(db->db_blkptr, ==, 2091 &dn->dn_phys->dn_blkptr[db->db_blkid]); 2092 zio = dn->dn_zio; 2093 } 2094 2095 ASSERT(db->db_level == 0 || data == db->db_buf); 2096 ASSERT3U(db->db_blkptr->blk_birth, <=, txg); 2097 ASSERT(zio); 2098 2099 zb.zb_objset = os->os_dsl_dataset ? os->os_dsl_dataset->ds_object : 0; 2100 zb.zb_object = db->db.db_object; 2101 zb.zb_level = db->db_level; 2102 zb.zb_blkid = db->db_blkid; 2103 2104 zio_flags = ZIO_FLAG_MUSTSUCCEED; 2105 if (dmu_ot[dn->dn_type].ot_metadata || zb.zb_level != 0) 2106 zio_flags |= ZIO_FLAG_METADATA; 2107 if (BP_IS_OLDER(db->db_blkptr, txg)) 2108 dsl_dataset_block_kill( 2109 os->os_dsl_dataset, db->db_blkptr, zio, tx); 2110 2111 dr->dr_zio = arc_write(zio, os->os_spa, checksum, compress, 2112 dmu_get_replication_level(os, &zb, dn->dn_type), txg, 2113 db->db_blkptr, data, dbuf_write_ready, dbuf_write_done, db, 2114 ZIO_PRIORITY_ASYNC_WRITE, zio_flags, &zb); 2115 } 2116 2117 /* ARGSUSED */ 2118 static void 2119 dbuf_write_ready(zio_t *zio, arc_buf_t *buf, void *vdb) 2120 { 2121 dmu_buf_impl_t *db = vdb; 2122 dnode_t *dn = db->db_dnode; 2123 objset_impl_t *os = dn->dn_objset; 2124 blkptr_t *bp_orig = &zio->io_bp_orig; 2125 uint64_t fill = 0; 2126 int old_size, new_size, i; 2127 2128 dprintf_dbuf_bp(db, bp_orig, "bp_orig: %s", ""); 2129 2130 old_size = bp_get_dasize(os->os_spa, bp_orig); 2131 new_size = bp_get_dasize(os->os_spa, zio->io_bp); 2132 2133 dnode_diduse_space(dn, new_size-old_size); 2134 2135 if (BP_IS_HOLE(zio->io_bp)) { 2136 dsl_dataset_t *ds = os->os_dsl_dataset; 2137 dmu_tx_t *tx = os->os_synctx; 2138 2139 if (bp_orig->blk_birth == tx->tx_txg) 2140 dsl_dataset_block_kill(ds, bp_orig, NULL, tx); 2141 ASSERT3U(db->db_blkptr->blk_fill, ==, 0); 2142 return; 2143 } 2144 2145 mutex_enter(&db->db_mtx); 2146 2147 if (db->db_level == 0) { 2148 mutex_enter(&dn->dn_mtx); 2149 if (db->db_blkid > dn->dn_phys->dn_maxblkid) 2150 dn->dn_phys->dn_maxblkid = db->db_blkid; 2151 mutex_exit(&dn->dn_mtx); 2152 2153 if (dn->dn_type == DMU_OT_DNODE) { 2154 dnode_phys_t *dnp = db->db.db_data; 2155 for (i = db->db.db_size >> DNODE_SHIFT; i > 0; 2156 i--, dnp++) { 2157 if (dnp->dn_type != DMU_OT_NONE) 2158 fill++; 2159 } 2160 } else { 2161 fill = 1; 2162 } 2163 } else { 2164 blkptr_t *bp = db->db.db_data; 2165 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 2166 for (i = db->db.db_size >> SPA_BLKPTRSHIFT; i > 0; i--, bp++) { 2167 if (BP_IS_HOLE(bp)) 2168 continue; 2169 ASSERT3U(BP_GET_LSIZE(bp), ==, 2170 db->db_level == 1 ? dn->dn_datablksz : 2171 (1<<dn->dn_phys->dn_indblkshift)); 2172 fill += bp->blk_fill; 2173 } 2174 } 2175 2176 db->db_blkptr->blk_fill = fill; 2177 BP_SET_TYPE(db->db_blkptr, dn->dn_type); 2178 BP_SET_LEVEL(db->db_blkptr, db->db_level); 2179 2180 mutex_exit(&db->db_mtx); 2181 2182 /* We must do this after we've set the bp's type and level */ 2183 if (!DVA_EQUAL(BP_IDENTITY(zio->io_bp), BP_IDENTITY(bp_orig))) { 2184 dsl_dataset_t *ds = os->os_dsl_dataset; 2185 dmu_tx_t *tx = os->os_synctx; 2186 2187 if (bp_orig->blk_birth == tx->tx_txg) 2188 dsl_dataset_block_kill(ds, bp_orig, NULL, tx); 2189 dsl_dataset_block_born(ds, zio->io_bp, tx); 2190 } 2191 } 2192 2193 /* ARGSUSED */ 2194 static void 2195 dbuf_write_done(zio_t *zio, arc_buf_t *buf, void *vdb) 2196 { 2197 dmu_buf_impl_t *db = vdb; 2198 uint64_t txg = zio->io_txg; 2199 dbuf_dirty_record_t **drp, *dr; 2200 2201 ASSERT3U(zio->io_error, ==, 0); 2202 2203 mutex_enter(&db->db_mtx); 2204 2205 drp = &db->db_last_dirty; 2206 while (*drp != db->db_data_pending) 2207 drp = &(*drp)->dr_next; 2208 ASSERT(!list_link_active(&(*drp)->dr_dirty_node)); 2209 ASSERT((*drp)->dr_txg == txg); 2210 ASSERT((*drp)->dr_next == NULL); 2211 dr = *drp; 2212 *drp = NULL; 2213 2214 if (db->db_level == 0) { 2215 ASSERT(db->db_blkid != DB_BONUS_BLKID); 2216 ASSERT(dr->dt.dl.dr_override_state == DR_NOT_OVERRIDDEN); 2217 2218 if (dr->dt.dl.dr_data != db->db_buf) 2219 VERIFY(arc_buf_remove_ref(dr->dt.dl.dr_data, db) == 1); 2220 else if (!BP_IS_HOLE(db->db_blkptr)) 2221 arc_set_callback(db->db_buf, dbuf_do_evict, db); 2222 else 2223 ASSERT(arc_released(db->db_buf)); 2224 } else { 2225 dnode_t *dn = db->db_dnode; 2226 2227 ASSERT(list_head(&dr->dt.di.dr_children) == NULL); 2228 ASSERT3U(db->db.db_size, ==, 1<<dn->dn_phys->dn_indblkshift); 2229 if (!BP_IS_HOLE(db->db_blkptr)) { 2230 int epbs = 2231 dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 2232 ASSERT3U(BP_GET_LSIZE(db->db_blkptr), ==, 2233 db->db.db_size); 2234 ASSERT3U(dn->dn_phys->dn_maxblkid 2235 >> (db->db_level * epbs), >=, db->db_blkid); 2236 arc_set_callback(db->db_buf, dbuf_do_evict, db); 2237 } 2238 mutex_destroy(&dr->dt.di.dr_mtx); 2239 list_destroy(&dr->dt.di.dr_children); 2240 } 2241 kmem_free(dr, sizeof (dbuf_dirty_record_t)); 2242 2243 cv_broadcast(&db->db_changed); 2244 ASSERT(db->db_dirtycnt > 0); 2245 db->db_dirtycnt -= 1; 2246 db->db_data_pending = NULL; 2247 mutex_exit(&db->db_mtx); 2248 2249 dprintf_dbuf_bp(db, zio->io_bp, "bp: %s", ""); 2250 2251 dbuf_rele(db, (void *)(uintptr_t)txg); 2252 } 2253