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 2008 Sun Microsystems, Inc. All rights reserved. 23 * Use is subject to license terms. 24 */ 25 26 #include <sys/zfs_context.h> 27 #include <sys/dbuf.h> 28 #include <sys/dnode.h> 29 #include <sys/dmu.h> 30 #include <sys/dmu_impl.h> 31 #include <sys/dmu_tx.h> 32 #include <sys/dmu_objset.h> 33 #include <sys/dsl_dir.h> 34 #include <sys/dsl_dataset.h> 35 #include <sys/spa.h> 36 #include <sys/zio.h> 37 #include <sys/dmu_zfetch.h> 38 39 static int free_range_compar(const void *node1, const void *node2); 40 41 static kmem_cache_t *dnode_cache; 42 43 static dnode_phys_t dnode_phys_zero; 44 45 int zfs_default_bs = SPA_MINBLOCKSHIFT; 46 int zfs_default_ibs = DN_MAX_INDBLKSHIFT; 47 48 /* ARGSUSED */ 49 static int 50 dnode_cons(void *arg, void *unused, int kmflag) 51 { 52 int i; 53 dnode_t *dn = arg; 54 bzero(dn, sizeof (dnode_t)); 55 56 rw_init(&dn->dn_struct_rwlock, NULL, RW_DEFAULT, NULL); 57 mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL); 58 mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL); 59 refcount_create(&dn->dn_holds); 60 refcount_create(&dn->dn_tx_holds); 61 62 for (i = 0; i < TXG_SIZE; i++) { 63 avl_create(&dn->dn_ranges[i], free_range_compar, 64 sizeof (free_range_t), 65 offsetof(struct free_range, fr_node)); 66 list_create(&dn->dn_dirty_records[i], 67 sizeof (dbuf_dirty_record_t), 68 offsetof(dbuf_dirty_record_t, dr_dirty_node)); 69 } 70 71 list_create(&dn->dn_dbufs, sizeof (dmu_buf_impl_t), 72 offsetof(dmu_buf_impl_t, db_link)); 73 74 return (0); 75 } 76 77 /* ARGSUSED */ 78 static void 79 dnode_dest(void *arg, void *unused) 80 { 81 int i; 82 dnode_t *dn = arg; 83 84 rw_destroy(&dn->dn_struct_rwlock); 85 mutex_destroy(&dn->dn_mtx); 86 mutex_destroy(&dn->dn_dbufs_mtx); 87 refcount_destroy(&dn->dn_holds); 88 refcount_destroy(&dn->dn_tx_holds); 89 90 for (i = 0; i < TXG_SIZE; i++) { 91 avl_destroy(&dn->dn_ranges[i]); 92 list_destroy(&dn->dn_dirty_records[i]); 93 } 94 95 list_destroy(&dn->dn_dbufs); 96 } 97 98 void 99 dnode_init(void) 100 { 101 dnode_cache = kmem_cache_create("dnode_t", 102 sizeof (dnode_t), 103 0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0); 104 } 105 106 void 107 dnode_fini(void) 108 { 109 kmem_cache_destroy(dnode_cache); 110 } 111 112 113 #ifdef ZFS_DEBUG 114 void 115 dnode_verify(dnode_t *dn) 116 { 117 int drop_struct_lock = FALSE; 118 119 ASSERT(dn->dn_phys); 120 ASSERT(dn->dn_objset); 121 122 ASSERT(dn->dn_phys->dn_type < DMU_OT_NUMTYPES); 123 124 if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY)) 125 return; 126 127 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) { 128 rw_enter(&dn->dn_struct_rwlock, RW_READER); 129 drop_struct_lock = TRUE; 130 } 131 if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) { 132 int i; 133 ASSERT3U(dn->dn_indblkshift, >=, 0); 134 ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT); 135 if (dn->dn_datablkshift) { 136 ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT); 137 ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT); 138 ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz); 139 } 140 ASSERT3U(dn->dn_nlevels, <=, 30); 141 ASSERT3U(dn->dn_type, <=, DMU_OT_NUMTYPES); 142 ASSERT3U(dn->dn_nblkptr, >=, 1); 143 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR); 144 ASSERT3U(dn->dn_bonuslen, <=, DN_MAX_BONUSLEN); 145 ASSERT3U(dn->dn_datablksz, ==, 146 dn->dn_datablkszsec << SPA_MINBLOCKSHIFT); 147 ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0); 148 ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) + 149 dn->dn_bonuslen, <=, DN_MAX_BONUSLEN); 150 for (i = 0; i < TXG_SIZE; i++) { 151 ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels); 152 } 153 } 154 if (dn->dn_phys->dn_type != DMU_OT_NONE) 155 ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels); 156 ASSERT(dn->dn_object == DMU_META_DNODE_OBJECT || dn->dn_dbuf != NULL); 157 if (dn->dn_dbuf != NULL) { 158 ASSERT3P(dn->dn_phys, ==, 159 (dnode_phys_t *)dn->dn_dbuf->db.db_data + 160 (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT))); 161 } 162 if (drop_struct_lock) 163 rw_exit(&dn->dn_struct_rwlock); 164 } 165 #endif 166 167 void 168 dnode_byteswap(dnode_phys_t *dnp) 169 { 170 uint64_t *buf64 = (void*)&dnp->dn_blkptr; 171 int i; 172 173 if (dnp->dn_type == DMU_OT_NONE) { 174 bzero(dnp, sizeof (dnode_phys_t)); 175 return; 176 } 177 178 dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec); 179 dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen); 180 dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid); 181 dnp->dn_used = BSWAP_64(dnp->dn_used); 182 183 /* 184 * dn_nblkptr is only one byte, so it's OK to read it in either 185 * byte order. We can't read dn_bouslen. 186 */ 187 ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT); 188 ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR); 189 for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++) 190 buf64[i] = BSWAP_64(buf64[i]); 191 192 /* 193 * OK to check dn_bonuslen for zero, because it won't matter if 194 * we have the wrong byte order. This is necessary because the 195 * dnode dnode is smaller than a regular dnode. 196 */ 197 if (dnp->dn_bonuslen != 0) { 198 /* 199 * Note that the bonus length calculated here may be 200 * longer than the actual bonus buffer. This is because 201 * we always put the bonus buffer after the last block 202 * pointer (instead of packing it against the end of the 203 * dnode buffer). 204 */ 205 int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t); 206 size_t len = DN_MAX_BONUSLEN - off; 207 ASSERT3U(dnp->dn_bonustype, <, DMU_OT_NUMTYPES); 208 dmu_ot[dnp->dn_bonustype].ot_byteswap(dnp->dn_bonus + off, len); 209 } 210 } 211 212 void 213 dnode_buf_byteswap(void *vbuf, size_t size) 214 { 215 dnode_phys_t *buf = vbuf; 216 int i; 217 218 ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT)); 219 ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0); 220 221 size >>= DNODE_SHIFT; 222 for (i = 0; i < size; i++) { 223 dnode_byteswap(buf); 224 buf++; 225 } 226 } 227 228 static int 229 free_range_compar(const void *node1, const void *node2) 230 { 231 const free_range_t *rp1 = node1; 232 const free_range_t *rp2 = node2; 233 234 if (rp1->fr_blkid < rp2->fr_blkid) 235 return (-1); 236 else if (rp1->fr_blkid > rp2->fr_blkid) 237 return (1); 238 else return (0); 239 } 240 241 void 242 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx) 243 { 244 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1); 245 246 dnode_setdirty(dn, tx); 247 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 248 ASSERT3U(newsize, <=, DN_MAX_BONUSLEN - 249 (dn->dn_nblkptr-1) * sizeof (blkptr_t)); 250 dn->dn_bonuslen = newsize; 251 if (newsize == 0) 252 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN; 253 else 254 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen; 255 rw_exit(&dn->dn_struct_rwlock); 256 } 257 258 static void 259 dnode_setdblksz(dnode_t *dn, int size) 260 { 261 ASSERT3U(P2PHASE(size, SPA_MINBLOCKSIZE), ==, 0); 262 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE); 263 ASSERT3U(size, >=, SPA_MINBLOCKSIZE); 264 ASSERT3U(size >> SPA_MINBLOCKSHIFT, <, 265 1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8)); 266 dn->dn_datablksz = size; 267 dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT; 268 dn->dn_datablkshift = ISP2(size) ? highbit(size - 1) : 0; 269 } 270 271 static dnode_t * 272 dnode_create(objset_impl_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db, 273 uint64_t object) 274 { 275 dnode_t *dn = kmem_cache_alloc(dnode_cache, KM_SLEEP); 276 (void) dnode_cons(dn, NULL, 0); /* XXX */ 277 278 dn->dn_objset = os; 279 dn->dn_object = object; 280 dn->dn_dbuf = db; 281 dn->dn_phys = dnp; 282 283 if (dnp->dn_datablkszsec) 284 dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT); 285 dn->dn_indblkshift = dnp->dn_indblkshift; 286 dn->dn_nlevels = dnp->dn_nlevels; 287 dn->dn_type = dnp->dn_type; 288 dn->dn_nblkptr = dnp->dn_nblkptr; 289 dn->dn_checksum = dnp->dn_checksum; 290 dn->dn_compress = dnp->dn_compress; 291 dn->dn_bonustype = dnp->dn_bonustype; 292 dn->dn_bonuslen = dnp->dn_bonuslen; 293 dn->dn_maxblkid = dnp->dn_maxblkid; 294 295 dmu_zfetch_init(&dn->dn_zfetch, dn); 296 297 ASSERT(dn->dn_phys->dn_type < DMU_OT_NUMTYPES); 298 mutex_enter(&os->os_lock); 299 list_insert_head(&os->os_dnodes, dn); 300 mutex_exit(&os->os_lock); 301 302 arc_space_consume(sizeof (dnode_t)); 303 return (dn); 304 } 305 306 static void 307 dnode_destroy(dnode_t *dn) 308 { 309 objset_impl_t *os = dn->dn_objset; 310 311 #ifdef ZFS_DEBUG 312 int i; 313 314 for (i = 0; i < TXG_SIZE; i++) { 315 ASSERT(!list_link_active(&dn->dn_dirty_link[i])); 316 ASSERT(NULL == list_head(&dn->dn_dirty_records[i])); 317 ASSERT(0 == avl_numnodes(&dn->dn_ranges[i])); 318 } 319 ASSERT(NULL == list_head(&dn->dn_dbufs)); 320 #endif 321 322 mutex_enter(&os->os_lock); 323 list_remove(&os->os_dnodes, dn); 324 mutex_exit(&os->os_lock); 325 326 if (dn->dn_dirtyctx_firstset) { 327 kmem_free(dn->dn_dirtyctx_firstset, 1); 328 dn->dn_dirtyctx_firstset = NULL; 329 } 330 dmu_zfetch_rele(&dn->dn_zfetch); 331 if (dn->dn_bonus) { 332 mutex_enter(&dn->dn_bonus->db_mtx); 333 dbuf_evict(dn->dn_bonus); 334 dn->dn_bonus = NULL; 335 } 336 kmem_cache_free(dnode_cache, dn); 337 arc_space_return(sizeof (dnode_t)); 338 } 339 340 void 341 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs, 342 dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx) 343 { 344 int i; 345 346 if (blocksize == 0) 347 blocksize = 1 << zfs_default_bs; 348 else if (blocksize > SPA_MAXBLOCKSIZE) 349 blocksize = SPA_MAXBLOCKSIZE; 350 else 351 blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE); 352 353 if (ibs == 0) 354 ibs = zfs_default_ibs; 355 356 ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT); 357 358 dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d\n", dn->dn_objset, 359 dn->dn_object, tx->tx_txg, blocksize, ibs); 360 361 ASSERT(dn->dn_type == DMU_OT_NONE); 362 ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0); 363 ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE); 364 ASSERT(ot != DMU_OT_NONE); 365 ASSERT3U(ot, <, DMU_OT_NUMTYPES); 366 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) || 367 (bonustype != DMU_OT_NONE && bonuslen != 0)); 368 ASSERT3U(bonustype, <, DMU_OT_NUMTYPES); 369 ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN); 370 ASSERT(dn->dn_type == DMU_OT_NONE); 371 ASSERT3U(dn->dn_maxblkid, ==, 0); 372 ASSERT3U(dn->dn_allocated_txg, ==, 0); 373 ASSERT3U(dn->dn_assigned_txg, ==, 0); 374 ASSERT(refcount_is_zero(&dn->dn_tx_holds)); 375 ASSERT3U(refcount_count(&dn->dn_holds), <=, 1); 376 ASSERT3P(list_head(&dn->dn_dbufs), ==, NULL); 377 378 for (i = 0; i < TXG_SIZE; i++) { 379 ASSERT3U(dn->dn_next_nlevels[i], ==, 0); 380 ASSERT3U(dn->dn_next_indblkshift[i], ==, 0); 381 ASSERT3U(dn->dn_next_bonuslen[i], ==, 0); 382 ASSERT3U(dn->dn_next_blksz[i], ==, 0); 383 ASSERT(!list_link_active(&dn->dn_dirty_link[i])); 384 ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL); 385 ASSERT3U(avl_numnodes(&dn->dn_ranges[i]), ==, 0); 386 } 387 388 dn->dn_type = ot; 389 dnode_setdblksz(dn, blocksize); 390 dn->dn_indblkshift = ibs; 391 dn->dn_nlevels = 1; 392 dn->dn_nblkptr = 1 + ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT); 393 dn->dn_bonustype = bonustype; 394 dn->dn_bonuslen = bonuslen; 395 dn->dn_checksum = ZIO_CHECKSUM_INHERIT; 396 dn->dn_compress = ZIO_COMPRESS_INHERIT; 397 dn->dn_dirtyctx = 0; 398 399 dn->dn_free_txg = 0; 400 if (dn->dn_dirtyctx_firstset) { 401 kmem_free(dn->dn_dirtyctx_firstset, 1); 402 dn->dn_dirtyctx_firstset = NULL; 403 } 404 405 dn->dn_allocated_txg = tx->tx_txg; 406 407 dnode_setdirty(dn, tx); 408 dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs; 409 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen; 410 dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz; 411 } 412 413 void 414 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, 415 dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx) 416 { 417 int i, old_nblkptr; 418 dmu_buf_impl_t *db = NULL; 419 420 ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE); 421 ASSERT3U(blocksize, <=, SPA_MAXBLOCKSIZE); 422 ASSERT3U(blocksize % SPA_MINBLOCKSIZE, ==, 0); 423 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx)); 424 ASSERT(tx->tx_txg != 0); 425 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) || 426 (bonustype != DMU_OT_NONE && bonuslen != 0)); 427 ASSERT3U(bonustype, <, DMU_OT_NUMTYPES); 428 ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN); 429 430 for (i = 0; i < TXG_SIZE; i++) 431 ASSERT(!list_link_active(&dn->dn_dirty_link[i])); 432 433 /* clean up any unreferenced dbufs */ 434 dnode_evict_dbufs(dn); 435 ASSERT3P(list_head(&dn->dn_dbufs), ==, NULL); 436 437 /* 438 * XXX I should really have a generation number to tell if we 439 * need to do this... 440 */ 441 if (blocksize != dn->dn_datablksz || 442 dn->dn_bonustype != bonustype || dn->dn_bonuslen != bonuslen) { 443 /* free all old data */ 444 dnode_free_range(dn, 0, -1ULL, tx); 445 } 446 447 /* change blocksize */ 448 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 449 if (blocksize != dn->dn_datablksz && 450 (!BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) || 451 list_head(&dn->dn_dbufs) != NULL)) { 452 db = dbuf_hold(dn, 0, FTAG); 453 dbuf_new_size(db, blocksize, tx); 454 } 455 dnode_setdblksz(dn, blocksize); 456 dnode_setdirty(dn, tx); 457 dn->dn_next_bonuslen[tx->tx_txg&TXG_MASK] = bonuslen; 458 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = blocksize; 459 rw_exit(&dn->dn_struct_rwlock); 460 if (db) 461 dbuf_rele(db, FTAG); 462 463 /* change type */ 464 dn->dn_type = ot; 465 466 /* change bonus size and type */ 467 mutex_enter(&dn->dn_mtx); 468 old_nblkptr = dn->dn_nblkptr; 469 dn->dn_bonustype = bonustype; 470 dn->dn_bonuslen = bonuslen; 471 dn->dn_nblkptr = 1 + ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT); 472 dn->dn_checksum = ZIO_CHECKSUM_INHERIT; 473 dn->dn_compress = ZIO_COMPRESS_INHERIT; 474 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR); 475 476 /* XXX - for now, we can't make nblkptr smaller */ 477 ASSERT3U(dn->dn_nblkptr, >=, old_nblkptr); 478 479 /* fix up the bonus db_size if dn_nblkptr has changed */ 480 if (dn->dn_bonus && dn->dn_bonuslen != old_nblkptr) { 481 dn->dn_bonus->db.db_size = 482 DN_MAX_BONUSLEN - (dn->dn_nblkptr-1) * sizeof (blkptr_t); 483 ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size); 484 } 485 486 dn->dn_allocated_txg = tx->tx_txg; 487 mutex_exit(&dn->dn_mtx); 488 } 489 490 void 491 dnode_special_close(dnode_t *dn) 492 { 493 /* 494 * Wait for final references to the dnode to clear. This can 495 * only happen if the arc is asyncronously evicting state that 496 * has a hold on this dnode while we are trying to evict this 497 * dnode. 498 */ 499 while (refcount_count(&dn->dn_holds) > 0) 500 delay(1); 501 dnode_destroy(dn); 502 } 503 504 dnode_t * 505 dnode_special_open(objset_impl_t *os, dnode_phys_t *dnp, uint64_t object) 506 { 507 dnode_t *dn = dnode_create(os, dnp, NULL, object); 508 DNODE_VERIFY(dn); 509 return (dn); 510 } 511 512 static void 513 dnode_buf_pageout(dmu_buf_t *db, void *arg) 514 { 515 dnode_t **children_dnodes = arg; 516 int i; 517 int epb = db->db_size >> DNODE_SHIFT; 518 519 for (i = 0; i < epb; i++) { 520 dnode_t *dn = children_dnodes[i]; 521 int n; 522 523 if (dn == NULL) 524 continue; 525 #ifdef ZFS_DEBUG 526 /* 527 * If there are holds on this dnode, then there should 528 * be holds on the dnode's containing dbuf as well; thus 529 * it wouldn't be eligable for eviction and this function 530 * would not have been called. 531 */ 532 ASSERT(refcount_is_zero(&dn->dn_holds)); 533 ASSERT(list_head(&dn->dn_dbufs) == NULL); 534 ASSERT(refcount_is_zero(&dn->dn_tx_holds)); 535 536 for (n = 0; n < TXG_SIZE; n++) 537 ASSERT(!list_link_active(&dn->dn_dirty_link[n])); 538 #endif 539 children_dnodes[i] = NULL; 540 dnode_destroy(dn); 541 } 542 kmem_free(children_dnodes, epb * sizeof (dnode_t *)); 543 } 544 545 /* 546 * errors: 547 * EINVAL - invalid object number. 548 * EIO - i/o error. 549 * succeeds even for free dnodes. 550 */ 551 int 552 dnode_hold_impl(objset_impl_t *os, uint64_t object, int flag, 553 void *tag, dnode_t **dnp) 554 { 555 int epb, idx, err; 556 int drop_struct_lock = FALSE; 557 int type; 558 uint64_t blk; 559 dnode_t *mdn, *dn; 560 dmu_buf_impl_t *db; 561 dnode_t **children_dnodes; 562 563 /* 564 * If you are holding the spa config lock as writer, you shouldn't 565 * be asking the DMU to do *anything*. 566 */ 567 ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0); 568 569 if (object == 0 || object >= DN_MAX_OBJECT) 570 return (EINVAL); 571 572 mdn = os->os_meta_dnode; 573 574 DNODE_VERIFY(mdn); 575 576 if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) { 577 rw_enter(&mdn->dn_struct_rwlock, RW_READER); 578 drop_struct_lock = TRUE; 579 } 580 581 blk = dbuf_whichblock(mdn, object * sizeof (dnode_phys_t)); 582 583 db = dbuf_hold(mdn, blk, FTAG); 584 if (drop_struct_lock) 585 rw_exit(&mdn->dn_struct_rwlock); 586 if (db == NULL) 587 return (EIO); 588 err = dbuf_read(db, NULL, DB_RF_CANFAIL); 589 if (err) { 590 dbuf_rele(db, FTAG); 591 return (err); 592 } 593 594 ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT); 595 epb = db->db.db_size >> DNODE_SHIFT; 596 597 idx = object & (epb-1); 598 599 children_dnodes = dmu_buf_get_user(&db->db); 600 if (children_dnodes == NULL) { 601 dnode_t **winner; 602 children_dnodes = kmem_zalloc(epb * sizeof (dnode_t *), 603 KM_SLEEP); 604 if (winner = dmu_buf_set_user(&db->db, children_dnodes, NULL, 605 dnode_buf_pageout)) { 606 kmem_free(children_dnodes, epb * sizeof (dnode_t *)); 607 children_dnodes = winner; 608 } 609 } 610 611 if ((dn = children_dnodes[idx]) == NULL) { 612 dnode_phys_t *dnp = (dnode_phys_t *)db->db.db_data+idx; 613 dnode_t *winner; 614 615 dn = dnode_create(os, dnp, db, object); 616 winner = atomic_cas_ptr(&children_dnodes[idx], NULL, dn); 617 if (winner != NULL) { 618 dnode_destroy(dn); 619 dn = winner; 620 } 621 } 622 623 mutex_enter(&dn->dn_mtx); 624 type = dn->dn_type; 625 if (dn->dn_free_txg || 626 ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE) || 627 ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)) { 628 mutex_exit(&dn->dn_mtx); 629 dbuf_rele(db, FTAG); 630 return (type == DMU_OT_NONE ? ENOENT : EEXIST); 631 } 632 mutex_exit(&dn->dn_mtx); 633 634 if (refcount_add(&dn->dn_holds, tag) == 1) 635 dbuf_add_ref(db, dn); 636 637 DNODE_VERIFY(dn); 638 ASSERT3P(dn->dn_dbuf, ==, db); 639 ASSERT3U(dn->dn_object, ==, object); 640 dbuf_rele(db, FTAG); 641 642 *dnp = dn; 643 return (0); 644 } 645 646 /* 647 * Return held dnode if the object is allocated, NULL if not. 648 */ 649 int 650 dnode_hold(objset_impl_t *os, uint64_t object, void *tag, dnode_t **dnp) 651 { 652 return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, tag, dnp)); 653 } 654 655 /* 656 * Can only add a reference if there is already at least one 657 * reference on the dnode. Returns FALSE if unable to add a 658 * new reference. 659 */ 660 boolean_t 661 dnode_add_ref(dnode_t *dn, void *tag) 662 { 663 mutex_enter(&dn->dn_mtx); 664 if (refcount_is_zero(&dn->dn_holds)) { 665 mutex_exit(&dn->dn_mtx); 666 return (FALSE); 667 } 668 VERIFY(1 < refcount_add(&dn->dn_holds, tag)); 669 mutex_exit(&dn->dn_mtx); 670 return (TRUE); 671 } 672 673 void 674 dnode_rele(dnode_t *dn, void *tag) 675 { 676 uint64_t refs; 677 678 mutex_enter(&dn->dn_mtx); 679 refs = refcount_remove(&dn->dn_holds, tag); 680 mutex_exit(&dn->dn_mtx); 681 /* NOTE: the DNODE_DNODE does not have a dn_dbuf */ 682 if (refs == 0 && dn->dn_dbuf) 683 dbuf_rele(dn->dn_dbuf, dn); 684 } 685 686 void 687 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx) 688 { 689 objset_impl_t *os = dn->dn_objset; 690 uint64_t txg = tx->tx_txg; 691 692 if (dn->dn_object == DMU_META_DNODE_OBJECT) 693 return; 694 695 DNODE_VERIFY(dn); 696 697 #ifdef ZFS_DEBUG 698 mutex_enter(&dn->dn_mtx); 699 ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg); 700 /* ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg); */ 701 mutex_exit(&dn->dn_mtx); 702 #endif 703 704 mutex_enter(&os->os_lock); 705 706 /* 707 * If we are already marked dirty, we're done. 708 */ 709 if (list_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) { 710 mutex_exit(&os->os_lock); 711 return; 712 } 713 714 ASSERT(!refcount_is_zero(&dn->dn_holds) || list_head(&dn->dn_dbufs)); 715 ASSERT(dn->dn_datablksz != 0); 716 ASSERT3U(dn->dn_next_bonuslen[txg&TXG_MASK], ==, 0); 717 ASSERT3U(dn->dn_next_blksz[txg&TXG_MASK], ==, 0); 718 719 dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n", 720 dn->dn_object, txg); 721 722 if (dn->dn_free_txg > 0 && dn->dn_free_txg <= txg) { 723 list_insert_tail(&os->os_free_dnodes[txg&TXG_MASK], dn); 724 } else { 725 list_insert_tail(&os->os_dirty_dnodes[txg&TXG_MASK], dn); 726 } 727 728 mutex_exit(&os->os_lock); 729 730 /* 731 * The dnode maintains a hold on its containing dbuf as 732 * long as there are holds on it. Each instantiated child 733 * dbuf maintaines a hold on the dnode. When the last child 734 * drops its hold, the dnode will drop its hold on the 735 * containing dbuf. We add a "dirty hold" here so that the 736 * dnode will hang around after we finish processing its 737 * children. 738 */ 739 VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg)); 740 741 (void) dbuf_dirty(dn->dn_dbuf, tx); 742 743 dsl_dataset_dirty(os->os_dsl_dataset, tx); 744 } 745 746 void 747 dnode_free(dnode_t *dn, dmu_tx_t *tx) 748 { 749 int txgoff = tx->tx_txg & TXG_MASK; 750 751 dprintf("dn=%p txg=%llu\n", dn, tx->tx_txg); 752 753 /* we should be the only holder... hopefully */ 754 /* ASSERT3U(refcount_count(&dn->dn_holds), ==, 1); */ 755 756 mutex_enter(&dn->dn_mtx); 757 if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) { 758 mutex_exit(&dn->dn_mtx); 759 return; 760 } 761 dn->dn_free_txg = tx->tx_txg; 762 mutex_exit(&dn->dn_mtx); 763 764 /* 765 * If the dnode is already dirty, it needs to be moved from 766 * the dirty list to the free list. 767 */ 768 mutex_enter(&dn->dn_objset->os_lock); 769 if (list_link_active(&dn->dn_dirty_link[txgoff])) { 770 list_remove(&dn->dn_objset->os_dirty_dnodes[txgoff], dn); 771 list_insert_tail(&dn->dn_objset->os_free_dnodes[txgoff], dn); 772 mutex_exit(&dn->dn_objset->os_lock); 773 } else { 774 mutex_exit(&dn->dn_objset->os_lock); 775 dnode_setdirty(dn, tx); 776 } 777 } 778 779 /* 780 * Try to change the block size for the indicated dnode. This can only 781 * succeed if there are no blocks allocated or dirty beyond first block 782 */ 783 int 784 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx) 785 { 786 dmu_buf_impl_t *db, *db_next; 787 int err; 788 789 if (size == 0) 790 size = SPA_MINBLOCKSIZE; 791 if (size > SPA_MAXBLOCKSIZE) 792 size = SPA_MAXBLOCKSIZE; 793 else 794 size = P2ROUNDUP(size, SPA_MINBLOCKSIZE); 795 796 if (ibs == dn->dn_indblkshift) 797 ibs = 0; 798 799 if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0) 800 return (0); 801 802 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 803 804 /* Check for any allocated blocks beyond the first */ 805 if (dn->dn_phys->dn_maxblkid != 0) 806 goto fail; 807 808 mutex_enter(&dn->dn_dbufs_mtx); 809 for (db = list_head(&dn->dn_dbufs); db; db = db_next) { 810 db_next = list_next(&dn->dn_dbufs, db); 811 812 if (db->db_blkid != 0 && db->db_blkid != DB_BONUS_BLKID) { 813 mutex_exit(&dn->dn_dbufs_mtx); 814 goto fail; 815 } 816 } 817 mutex_exit(&dn->dn_dbufs_mtx); 818 819 if (ibs && dn->dn_nlevels != 1) 820 goto fail; 821 822 /* resize the old block */ 823 err = dbuf_hold_impl(dn, 0, 0, TRUE, FTAG, &db); 824 if (err == 0) 825 dbuf_new_size(db, size, tx); 826 else if (err != ENOENT) 827 goto fail; 828 829 dnode_setdblksz(dn, size); 830 dnode_setdirty(dn, tx); 831 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size; 832 if (ibs) { 833 dn->dn_indblkshift = ibs; 834 dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs; 835 } 836 /* rele after we have fixed the blocksize in the dnode */ 837 if (db) 838 dbuf_rele(db, FTAG); 839 840 rw_exit(&dn->dn_struct_rwlock); 841 return (0); 842 843 fail: 844 rw_exit(&dn->dn_struct_rwlock); 845 return (ENOTSUP); 846 } 847 848 /* read-holding callers must not rely on the lock being continuously held */ 849 void 850 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read) 851 { 852 uint64_t txgoff = tx->tx_txg & TXG_MASK; 853 int epbs, new_nlevels; 854 uint64_t sz; 855 856 ASSERT(blkid != DB_BONUS_BLKID); 857 858 ASSERT(have_read ? 859 RW_READ_HELD(&dn->dn_struct_rwlock) : 860 RW_WRITE_HELD(&dn->dn_struct_rwlock)); 861 862 /* 863 * if we have a read-lock, check to see if we need to do any work 864 * before upgrading to a write-lock. 865 */ 866 if (have_read) { 867 if (blkid <= dn->dn_maxblkid) 868 return; 869 870 if (!rw_tryupgrade(&dn->dn_struct_rwlock)) { 871 rw_exit(&dn->dn_struct_rwlock); 872 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 873 } 874 } 875 876 if (blkid <= dn->dn_maxblkid) 877 goto out; 878 879 dn->dn_maxblkid = blkid; 880 881 /* 882 * Compute the number of levels necessary to support the new maxblkid. 883 */ 884 new_nlevels = 1; 885 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 886 for (sz = dn->dn_nblkptr; 887 sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs) 888 new_nlevels++; 889 890 if (new_nlevels > dn->dn_nlevels) { 891 int old_nlevels = dn->dn_nlevels; 892 dmu_buf_impl_t *db; 893 list_t *list; 894 dbuf_dirty_record_t *new, *dr, *dr_next; 895 896 dn->dn_nlevels = new_nlevels; 897 898 ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]); 899 dn->dn_next_nlevels[txgoff] = new_nlevels; 900 901 /* dirty the left indirects */ 902 db = dbuf_hold_level(dn, old_nlevels, 0, FTAG); 903 new = dbuf_dirty(db, tx); 904 dbuf_rele(db, FTAG); 905 906 /* transfer the dirty records to the new indirect */ 907 mutex_enter(&dn->dn_mtx); 908 mutex_enter(&new->dt.di.dr_mtx); 909 list = &dn->dn_dirty_records[txgoff]; 910 for (dr = list_head(list); dr; dr = dr_next) { 911 dr_next = list_next(&dn->dn_dirty_records[txgoff], dr); 912 if (dr->dr_dbuf->db_level != new_nlevels-1 && 913 dr->dr_dbuf->