1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21 /* 22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. 23 * Copyright (c) 2012, 2015 by Delphix. All rights reserved. 24 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved. 25 */ 26 27 #include <sys/zfs_context.h> 28 #include <sys/dbuf.h> 29 #include <sys/dnode.h> 30 #include <sys/dmu.h> 31 #include <sys/dmu_impl.h> 32 #include <sys/dmu_tx.h> 33 #include <sys/dmu_objset.h> 34 #include <sys/dsl_dir.h> 35 #include <sys/dsl_dataset.h> 36 #include <sys/spa.h> 37 #include <sys/zio.h> 38 #include <sys/dmu_zfetch.h> 39 #include <sys/range_tree.h> 40 41 static kmem_cache_t *dnode_cache; 42 /* 43 * Define DNODE_STATS to turn on statistic gathering. By default, it is only 44 * turned on when DEBUG is also defined. 45 */ 46 #ifdef DEBUG 47 #define DNODE_STATS 48 #endif /* DEBUG */ 49 50 #ifdef DNODE_STATS 51 #define DNODE_STAT_ADD(stat) ((stat)++) 52 #else 53 #define DNODE_STAT_ADD(stat) /* nothing */ 54 #endif /* DNODE_STATS */ 55 56 static dnode_phys_t dnode_phys_zero; 57 58 int zfs_default_bs = SPA_MINBLOCKSHIFT; 59 int zfs_default_ibs = DN_MAX_INDBLKSHIFT; 60 61 static kmem_cbrc_t dnode_move(void *, void *, size_t, void *); 62 63 static int 64 dbuf_compare(const void *x1, const void *x2) 65 { 66 const dmu_buf_impl_t *d1 = x1; 67 const dmu_buf_impl_t *d2 = x2; 68 69 if (d1->db_level < d2->db_level) { 70 return (-1); 71 } 72 if (d1->db_level > d2->db_level) { 73 return (1); 74 } 75 76 if (d1->db_blkid < d2->db_blkid) { 77 return (-1); 78 } 79 if (d1->db_blkid > d2->db_blkid) { 80 return (1); 81 } 82 83 if (d1->db_state == DB_SEARCH) { 84 ASSERT3S(d2->db_state, !=, DB_SEARCH); 85 return (-1); 86 } else if (d2->db_state == DB_SEARCH) { 87 ASSERT3S(d1->db_state, !=, DB_SEARCH); 88 return (1); 89 } 90 91 if ((uintptr_t)d1 < (uintptr_t)d2) { 92 return (-1); 93 } 94 if ((uintptr_t)d1 > (uintptr_t)d2) { 95 return (1); 96 } 97 return (0); 98 } 99 100 /* ARGSUSED */ 101 static int 102 dnode_cons(void *arg, void *unused, int kmflag) 103 { 104 dnode_t *dn = arg; 105 int i; 106 107 rw_init(&dn->dn_struct_rwlock, NULL, RW_DEFAULT, NULL); 108 mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL); 109 mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL); 110 cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL); 111 112 /* 113 * Every dbuf has a reference, and dropping a tracked reference is 114 * O(number of references), so don't track dn_holds. 115 */ 116 refcount_create_untracked(&dn->dn_holds); 117 refcount_create(&dn->dn_tx_holds); 118 list_link_init(&dn->dn_link); 119 120 bzero(&dn->dn_next_nblkptr[0], sizeof (dn->dn_next_nblkptr)); 121 bzero(&dn->dn_next_nlevels[0], sizeof (dn->dn_next_nlevels)); 122 bzero(&dn->dn_next_indblkshift[0], sizeof (dn->dn_next_indblkshift)); 123 bzero(&dn->dn_next_bonustype[0], sizeof (dn->dn_next_bonustype)); 124 bzero(&dn->dn_rm_spillblk[0], sizeof (dn->dn_rm_spillblk)); 125 bzero(&dn->dn_next_bonuslen[0], sizeof (dn->dn_next_bonuslen)); 126 bzero(&dn->dn_next_blksz[0], sizeof (dn->dn_next_blksz)); 127 128 for (i = 0; i < TXG_SIZE; i++) { 129 list_link_init(&dn->dn_dirty_link[i]); 130 dn->dn_free_ranges[i] = NULL; 131 list_create(&dn->dn_dirty_records[i], 132 sizeof (dbuf_dirty_record_t), 133 offsetof(dbuf_dirty_record_t, dr_dirty_node)); 134 } 135 136 dn->dn_allocated_txg = 0; 137 dn->dn_free_txg = 0; 138 dn->dn_assigned_txg = 0; 139 dn->dn_dirtyctx = 0; 140 dn->dn_dirtyctx_firstset = NULL; 141 dn->dn_bonus = NULL; 142 dn->dn_have_spill = B_FALSE; 143 dn->dn_zio = NULL; 144 dn->dn_oldused = 0; 145 dn->dn_oldflags = 0; 146 dn->dn_olduid = 0; 147 dn->dn_oldgid = 0; 148 dn->dn_newuid = 0; 149 dn->dn_newgid = 0; 150 dn->dn_id_flags = 0; 151 152 dn->dn_dbufs_count = 0; 153 dn->dn_unlisted_l0_blkid = 0; 154 avl_create(&dn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t), 155 offsetof(dmu_buf_impl_t, db_link)); 156 157 dn->dn_moved = 0; 158 return (0); 159 } 160 161 /* ARGSUSED */ 162 static void 163 dnode_dest(void *arg, void *unused) 164 { 165 int i; 166 dnode_t *dn = arg; 167 168 rw_destroy(&dn->dn_struct_rwlock); 169 mutex_destroy(&dn->dn_mtx); 170 mutex_destroy(&dn->dn_dbufs_mtx); 171 cv_destroy(&dn->dn_notxholds); 172 refcount_destroy(&dn->dn_holds); 173 refcount_destroy(&dn->dn_tx_holds); 174 ASSERT(!list_link_active(&dn->dn_link)); 175 176 for (i = 0; i < TXG_SIZE; i++) { 177 ASSERT(!list_link_active(&dn->dn_dirty_link[i])); 178 ASSERT3P(dn->dn_free_ranges[i], ==, NULL); 179 list_destroy(&dn->dn_dirty_records[i]); 180 ASSERT0(dn->dn_next_nblkptr[i]); 181 ASSERT0(dn->dn_next_nlevels[i]); 182 ASSERT0(dn->dn_next_indblkshift[i]); 183 ASSERT0(dn->dn_next_bonustype[i]); 184 ASSERT0(dn->dn_rm_spillblk[i]); 185 ASSERT0(dn->dn_next_bonuslen[i]); 186 ASSERT0(dn->dn_next_blksz[i]); 187 } 188 189 ASSERT0(dn->dn_allocated_txg); 190 ASSERT0(dn->dn_free_txg); 191 ASSERT0(dn->dn_assigned_txg); 192 ASSERT0(dn->dn_dirtyctx); 193 ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL); 194 ASSERT3P(dn->dn_bonus, ==, NULL); 195 ASSERT(!dn->dn_have_spill); 196 ASSERT3P(dn->dn_zio, ==, NULL); 197 ASSERT0(dn->dn_oldused); 198 ASSERT0(dn->dn_oldflags); 199 ASSERT0(dn->dn_olduid); 200 ASSERT0(dn->dn_oldgid); 201 ASSERT0(dn->dn_newuid); 202 ASSERT0(dn->dn_newgid); 203 ASSERT0(dn->dn_id_flags); 204 205 ASSERT0(dn->dn_dbufs_count); 206 ASSERT0(dn->dn_unlisted_l0_blkid); 207 avl_destroy(&dn->dn_dbufs); 208 } 209 210 void 211 dnode_init(void) 212 { 213 ASSERT(dnode_cache == NULL); 214 dnode_cache = kmem_cache_create("dnode_t", 215 sizeof (dnode_t), 216 0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0); 217 kmem_cache_set_move(dnode_cache, dnode_move); 218 } 219 220 void 221 dnode_fini(void) 222 { 223 kmem_cache_destroy(dnode_cache); 224 dnode_cache = NULL; 225 } 226 227 228 #ifdef ZFS_DEBUG 229 void 230 dnode_verify(dnode_t *dn) 231 { 232 int drop_struct_lock = FALSE; 233 234 ASSERT(dn->dn_phys); 235 ASSERT(dn->dn_objset); 236 ASSERT(dn->dn_handle->dnh_dnode == dn); 237 238 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type)); 239 240 if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY)) 241 return; 242 243 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) { 244 rw_enter(&dn->dn_struct_rwlock, RW_READER); 245 drop_struct_lock = TRUE; 246 } 247 if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) { 248 int i; 249 ASSERT3U(dn->dn_indblkshift, >=, 0); 250 ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT); 251 if (dn->dn_datablkshift) { 252 ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT); 253 ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT); 254 ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz); 255 } 256 ASSERT3U(dn->dn_nlevels, <=, 30); 257 ASSERT(DMU_OT_IS_VALID(dn->dn_type)); 258 ASSERT3U(dn->dn_nblkptr, >=, 1); 259 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR); 260 ASSERT3U(dn->dn_bonuslen, <=, DN_MAX_BONUSLEN); 261 ASSERT3U(dn->dn_datablksz, ==, 262 dn->dn_datablkszsec << SPA_MINBLOCKSHIFT); 263 ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0); 264 ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) + 265 dn->dn_bonuslen, <=, DN_MAX_BONUSLEN); 266 for (i = 0; i < TXG_SIZE; i++) { 267 ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels); 268 } 269 } 270 if (dn->dn_phys->dn_type != DMU_OT_NONE) 271 ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels); 272 ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL); 273 if (dn->dn_dbuf != NULL) { 274 ASSERT3P(dn->dn_phys, ==, 275 (dnode_phys_t *)dn->dn_dbuf->db.db_data + 276 (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT))); 277 } 278 if (drop_struct_lock) 279 rw_exit(&dn->dn_struct_rwlock); 280 } 281 #endif 282 283 void 284 dnode_byteswap(dnode_phys_t *dnp) 285 { 286 uint64_t *buf64 = (void*)&dnp->dn_blkptr; 287 int i; 288 289 if (dnp->dn_type == DMU_OT_NONE) { 290 bzero(dnp, sizeof (dnode_phys_t)); 291 return; 292 } 293 294 dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec); 295 dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen); 296 dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid); 297 dnp->dn_used = BSWAP_64(dnp->dn_used); 298 299 /* 300 * dn_nblkptr is only one byte, so it's OK to read it in either 301 * byte order. We can't read dn_bouslen. 302 */ 303 ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT); 304 ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR); 305 for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++) 306 buf64[i] = BSWAP_64(buf64[i]); 307 308 /* 309 * OK to check dn_bonuslen for zero, because it won't matter if 310 * we have the wrong byte order. This is necessary because the 311 * dnode dnode is smaller than a regular dnode. 312 */ 313 if (dnp->dn_bonuslen != 0) { 314 /* 315 * Note that the bonus length calculated here may be 316 * longer than the actual bonus buffer. This is because 317 * we always put the bonus buffer after the last block 318 * pointer (instead of packing it against the end of the 319 * dnode buffer). 320 */ 321 int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t); 322 size_t len = DN_MAX_BONUSLEN - off; 323 ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype)); 324 dmu_object_byteswap_t byteswap = 325 DMU_OT_BYTESWAP(dnp->dn_bonustype); 326 dmu_ot_byteswap[byteswap].ob_func(dnp->dn_bonus + off, len); 327 } 328 329 /* Swap SPILL block if we have one */ 330 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) 331 byteswap_uint64_array(&dnp->dn_spill, sizeof (blkptr_t)); 332 333 } 334 335 void 336 dnode_buf_byteswap(void *vbuf, size_t size) 337 { 338 dnode_phys_t *buf = vbuf; 339 int i; 340 341 ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT)); 342 ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0); 343 344 size >>= DNODE_SHIFT; 345 for (i = 0; i < size; i++) { 346 dnode_byteswap(buf); 347 buf++; 348 } 349 } 350 351 void 352 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx) 353 { 354 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1); 355 356 dnode_setdirty(dn, tx); 357 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 358 ASSERT3U(newsize, <=, DN_MAX_BONUSLEN - 359 (dn->dn_nblkptr-1) * sizeof (blkptr_t)); 360 dn->dn_bonuslen = newsize; 361 if (newsize == 0) 362 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN; 363 else 364 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen; 365 rw_exit(&dn->dn_struct_rwlock); 366 } 367 368 void 369 dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx) 370 { 371 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1); 372 dnode_setdirty(dn, tx); 373 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 374 dn->dn_bonustype = newtype; 375 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype; 376 rw_exit(&dn->dn_struct_rwlock); 377 } 378 379 void 380 dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx) 381 { 382 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1); 383 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 384 dnode_setdirty(dn, tx); 385 dn->dn_rm_spillblk[tx->tx_txg&TXG_MASK] = DN_KILL_SPILLBLK; 386 dn->dn_have_spill = B_FALSE; 387 } 388 389 static void 390 dnode_setdblksz(dnode_t *dn, int size) 391 { 392 ASSERT0(P2PHASE(size, SPA_MINBLOCKSIZE)); 393 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE); 394 ASSERT3U(size, >=, SPA_MINBLOCKSIZE); 395 ASSERT3U(size >> SPA_MINBLOCKSHIFT, <, 396 1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8)); 397 dn->dn_datablksz = size; 398 dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT; 399 dn->dn_datablkshift = ISP2(size) ? highbit64(size - 1) : 0; 400 } 401 402 static dnode_t * 403 dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db, 404 uint64_t object, dnode_handle_t *dnh) 405 { 406 dnode_t *dn; 407 408 dn = kmem_cache_alloc(dnode_cache, KM_SLEEP); 409 ASSERT(!POINTER_IS_VALID(dn->dn_objset)); 410 dn->dn_moved = 0; 411 412 /* 413 * Defer setting dn_objset until the dnode is ready to be a candidate 414 * for the dnode_move() callback. 415 */ 416 dn->dn_object = object; 417 dn->dn_dbuf = db; 418 dn->dn_handle = dnh; 419 dn->dn_phys = dnp; 420 421 if (dnp->dn_datablkszsec) { 422 dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT); 423 } else { 424 dn->dn_datablksz = 0; 425 dn->dn_datablkszsec = 0; 426 dn->dn_datablkshift = 0; 427 } 428 dn->dn_indblkshift = dnp->dn_indblkshift; 429 dn->dn_nlevels = dnp->dn_nlevels; 430 dn->dn_type = dnp->dn_type; 431 dn->dn_nblkptr = dnp->dn_nblkptr; 432 dn->dn_checksum = dnp->dn_checksum; 433 dn->dn_compress = dnp->dn_compress; 434 dn->dn_bonustype = dnp->dn_bonustype; 435 dn->dn_bonuslen = dnp->dn_bonuslen; 436 dn->dn_maxblkid = dnp->dn_maxblkid; 437 dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0); 438 dn->dn_id_flags = 0; 439 440 dmu_zfetch_init(&dn->dn_zfetch, dn); 441 442 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type)); 443 444 mutex_enter(&os->os_lock); 445 if (dnh->dnh_dnode != NULL) { 446 /* Lost the allocation race. */ 447 mutex_exit(&os->os_lock); 448 kmem_cache_free(dnode_cache, dn); 449 return (dnh->dnh_dnode); 450 } 451 452 /* 453 * Exclude special dnodes from os_dnodes so an empty os_dnodes 454 * signifies that the special dnodes have no references from 455 * their children (the entries in os_dnodes). This allows 456 * dnode_destroy() to easily determine if the last child has 457 * been removed and then complete eviction of the objset. 458 */ 459 if (!DMU_OBJECT_IS_SPECIAL(object)) 460 list_insert_head(&os->os_dnodes, dn); 461 membar_producer(); 462 463 /* 464 * Everything else must be valid before assigning dn_objset 465 * makes the dnode eligible for dnode_move(). 466 */ 467 dn->dn_objset = os; 468 469 dnh->dnh_dnode = dn; 470 mutex_exit(&os->os_lock); 471 472 arc_space_consume(sizeof (dnode_t), ARC_SPACE_OTHER); 473 return (dn); 474 } 475 476 /* 477 * Caller must be holding the dnode handle, which is released upon return. 478 */ 479 static void 480 dnode_destroy(dnode_t *dn) 481 { 482 objset_t *os = dn->dn_objset; 483 boolean_t complete_os_eviction = B_FALSE; 484 485 ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0); 486 487 mutex_enter(&os->os_lock); 488 POINTER_INVALIDATE(&dn->dn_objset); 489 if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) { 490 list_remove(&os->os_dnodes, dn); 491 complete_os_eviction = 492 list_is_empty(&os->os_dnodes) && 493 list_link_active(&os->os_evicting_node); 494 } 495 mutex_exit(&os->os_lock); 496 497 /* the dnode can no longer move, so we can release the handle */ 498 zrl_remove(&dn->dn_handle->dnh_zrlock); 499 500 dn->dn_allocated_txg = 0; 501 dn->dn_free_txg = 0; 502 dn->dn_assigned_txg = 0; 503 504 dn->dn_dirtyctx = 0; 505 if (dn->dn_dirtyctx_firstset != NULL) { 506 kmem_free(dn->dn_dirtyctx_firstset, 1); 507 dn->dn_dirtyctx_firstset = NULL; 508 } 509 if (dn->dn_bonus != NULL) { 510 mutex_enter(&dn->dn_bonus->db_mtx); 511 dbuf_evict(dn->dn_bonus); 512 dn->dn_bonus = NULL; 513 } 514 dn->dn_zio = NULL; 515 516 dn->dn_have_spill = B_FALSE; 517 dn->dn_oldused = 0; 518 dn->dn_oldflags = 0; 519 dn->dn_olduid = 0; 520 dn->dn_oldgid = 0; 521 dn->dn_newuid = 0; 522 dn->dn_newgid = 0; 523 dn->dn_id_flags = 0; 524 dn->dn_unlisted_l0_blkid = 0; 525 526 dmu_zfetch_fini(&dn->dn_zfetch); 527 kmem_cache_free(dnode_cache, dn); 528 arc_space_return(sizeof (dnode_t), ARC_SPACE_OTHER); 529 530 if (complete_os_eviction) 531 dmu_objset_evict_done(os); 532 } 533 534 void 535 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs, 536 dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx) 537 { 538 int i; 539 540 ASSERT3U(blocksize, <=, 541 spa_maxblocksize(dmu_objset_spa(dn->dn_objset))); 542 if (blocksize == 0) 543 blocksize = 1 << zfs_default_bs; 544 else 545 blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE); 546 547 if (ibs == 0) 548 ibs = zfs_default_ibs; 549 550 ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT); 551 552 dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d\n", dn->dn_objset, 553 dn->dn_object, tx->tx_txg, blocksize, ibs); 554 555 ASSERT(dn->dn_type == DMU_OT_NONE); 556 ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0); 557 ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE); 558 ASSERT(ot != DMU_OT_NONE); 559 ASSERT(DMU_OT_IS_VALID(ot)); 560 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) || 561 (bonustype == DMU_OT_SA && bonuslen == 0) || 562 (bonustype != DMU_OT_NONE && bonuslen != 0)); 563 ASSERT(DMU_OT_IS_VALID(bonustype)); 564 ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN); 565 ASSERT(dn->dn_type == DMU_OT_NONE); 566 ASSERT0(dn->dn_maxblkid); 567 ASSERT0(dn->dn_allocated_txg); 568 ASSERT0(dn->dn_assigned_txg); 569 ASSERT(refcount_is_zero(&dn->dn_tx_holds)); 570 ASSERT3U(refcount_count(&dn->dn_holds), <=, 1); 571 ASSERT(avl_is_empty(&dn->dn_dbufs)); 572 573 for (i = 0; i < TXG_SIZE; i++) { 574 ASSERT0(dn->dn_next_nblkptr[i]); 575 ASSERT0(dn->dn_next_nlevels[i]); 576 ASSERT0(dn->dn_next_indblkshift[i]); 577 ASSERT0(dn->dn_next_bonuslen[i]); 578 ASSERT0(dn->dn_next_bonustype[i]); 579 ASSERT0(dn->dn_rm_spillblk[i]); 580 ASSERT0(dn->dn_next_blksz[i]); 581 ASSERT(!list_link_active(&dn->dn_dirty_link[i])); 582 ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL); 583 ASSERT3P(dn->dn_free_ranges[i], ==, NULL); 584 } 585 586 dn->dn_type = ot; 587 dnode_setdblksz(dn, blocksize); 588 dn->dn_indblkshift = ibs; 589 dn->dn_nlevels = 1; 590 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */ 591 dn->dn_nblkptr = 1; 592 else 593 dn->dn_nblkptr = 1 + 594 ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT); 595 dn->dn_bonustype = bonustype; 596 dn->dn_bonuslen = bonuslen; 597 dn->dn_checksum = ZIO_CHECKSUM_INHERIT; 598 dn->dn_compress = ZIO_COMPRESS_INHERIT; 599 dn->dn_dirtyctx = 0; 600 601 dn->dn_free_txg = 0; 602 if (dn->dn_dirtyctx_firstset) { 603 kmem_free(dn->dn_dirtyctx_firstset, 1); 604 dn->dn_dirtyctx_firstset = NULL; 605 } 606 607 dn->dn_allocated_txg = tx->tx_txg; 608 dn->dn_id_flags = 0; 609 610 dnode_setdirty(dn, tx); 611 dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs; 612 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen; 613 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype; 614 dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz; 615 } 616 617 void 618 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, 619 dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx) 620 { 621 int nblkptr; 622 623 ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE); 624 ASSERT3U(blocksize, <=, 625 spa_maxblocksize(dmu_objset_spa(dn->dn_objset))); 626 ASSERT0(blocksize % SPA_MINBLOCKSIZE); 627 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx)); 628 ASSERT(tx->tx_txg != 0); 629 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) || 630 (bonustype != DMU_OT_NONE && bonuslen != 0) || 631 (bonustype == DMU_OT_SA && bonuslen == 0)); 632 ASSERT(DMU_OT_IS_VALID(bonustype)); 633 ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN); 634 635 /* clean up any unreferenced dbufs */ 636 dnode_evict_dbufs(dn); 637 638 dn->dn_id_flags = 0; 639 640 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 641 dnode_setdirty(dn, tx); 642 if (dn->dn_datablksz != blocksize) { 643 /* change blocksize */ 644 ASSERT(dn->dn_maxblkid == 0 && 645 (BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) || 646 dnode_block_freed(dn, 0))); 647 dnode_setdblksz(dn, blocksize); 648 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = blocksize; 649 } 650 if (dn->dn_bonuslen != bonuslen) 651 dn->dn_next_bonuslen[tx->tx_txg&TXG_MASK] = bonuslen; 652 653 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */ 654 nblkptr = 1; 655 else 656 nblkptr = 1 + ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT); 657 if (dn->dn_bonustype != bonustype) 658 dn->dn_next_bonustype[tx->tx_txg&TXG_MASK] = bonustype; 659 if (dn->dn_nblkptr != nblkptr) 660 dn->dn_next_nblkptr[tx->tx_txg&TXG_MASK] = nblkptr; 661 if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR) { 662 dbuf_rm_spill(dn, tx); 663 dnode_rm_spill(dn, tx); 664 } 665 rw_exit(&dn->dn_struct_rwlock); 666 667 /* change type */ 668 dn->dn_type = ot; 669 670 /* change bonus size and type */ 671 mutex_enter(&dn->dn_mtx); 672 dn->dn_bonustype = bonustype; 673 dn->dn_bonuslen = bonuslen; 674 dn->dn_nblkptr = nblkptr; 675 dn->dn_checksum = ZIO_CHECKSUM_INHERIT; 676 dn->dn_compress = ZIO_COMPRESS_INHERIT; 677 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR); 678 679 /* fix up the bonus db_size */ 680 if (dn->dn_bonus) { 681 dn->dn_bonus->db.db_size = 682 DN_MAX_BONUSLEN - (dn->dn_nblkptr-1) * sizeof (blkptr_t); 683 ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size); 684 } 685 686 dn->dn_allocated_txg = tx->tx_txg; 687 mutex_exit(&dn->dn_mtx); 688 } 689 690 #ifdef DNODE_STATS 691 static struct { 692 uint64_t dms_dnode_invalid; 693 uint64_t dms_dnode_recheck1; 694 uint64_t dms_dnode_recheck2; 695 uint64_t dms_dnode_special; 696 uint64_t dms_dnode_handle; 697 uint64_t dms_dnode_rwlock; 698 uint64_t dms_dnode_active; 699 } dnode_move_stats; 700 #endif /* DNODE_STATS */ 701 702 static void 703 dnode_move_impl(dnode_t *odn, dnode_t *ndn) 704 { 705 int i; 706 707 ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock)); 708 ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx)); 709 ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx)); 710 ASSERT(!RW_LOCK_HELD(&odn->dn_zfetch.zf_rwlock)); 711 712 /* Copy fields. */ 713 ndn->dn_objset = odn->dn_objset; 714 ndn->dn_object = odn->dn_object; 715 ndn->dn_dbuf = odn->dn_dbuf; 716 ndn->dn_handle = odn->dn_handle; 717 ndn->dn_phys = odn->dn_phys; 718 ndn->dn_type = odn->dn_type; 719 ndn->dn_bonuslen = odn->dn_bonuslen; 720 ndn->dn_bonustype = odn->dn_bonustype; 721 ndn->dn_nblkptr = odn->dn_nblkptr; 722 ndn->dn_checksum = odn->dn_checksum; 723 ndn->dn_compress = odn->dn_compress; 724 ndn->dn_nlevels = odn->dn_nlevels; 725 ndn->dn_indblkshift = odn->dn_indblkshift; 726 ndn->dn_datablkshift = odn->dn_datablkshift; 727 ndn->dn_datablkszsec = odn->dn_datablkszsec; 728 ndn->dn_datablksz = odn->dn_datablksz; 729 ndn->dn_maxblkid = odn->dn_maxblkid; 730 bcopy(&odn->dn_next_nblkptr[0], &ndn->dn_next_nblkptr[0], 731 sizeof (odn->dn_next_nblkptr)); 732 bcopy(&odn->dn_next_nlevels[0], &ndn->dn_next_nlevels[0], 733 sizeof (odn->dn_next_nlevels)); 734 bcopy(&odn->dn_next_indblkshift[0], &ndn->dn_next_indblkshift[0], 735 sizeof (odn->dn_next_indblkshift)); 736 bcopy(&odn->dn_next_bonustype[0], &ndn->dn_next_bonustype[0], 737 sizeof (odn->dn_next_bonustype)); 738 bcopy(&odn->dn_rm_spillblk[0], &ndn->dn_rm_spillblk[0], 739 sizeof (odn->dn_rm_spillblk)); 740 bcopy(&odn->dn_next_bonuslen[0], &ndn->dn_next_bonuslen[0], 741 sizeof (odn->dn_next_bonuslen)); 742 bcopy(&odn->dn_next_blksz[0], &ndn->dn_next_blksz[0], 743 sizeof (odn->dn_next_blksz)); 744 for (i = 0; i < TXG_SIZE; i++) { 745 list_move_tail(&ndn->dn_dirty_records[i], 746 &odn->dn_dirty_records[i]); 747 } 748 bcopy(&odn->dn_free_ranges[0], &ndn->dn_free_ranges[0], 749 sizeof (odn->dn_free_ranges)); 750 ndn->dn_allocated_txg = odn->dn_allocated_txg; 751 ndn->dn_free_txg = odn->dn_free_txg; 752 ndn->dn_assigned_txg = odn->dn_assigned_txg; 753 ndn->dn_dirtyctx = odn->dn_dirtyctx; 754 ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset; 755 ASSERT(refcount_count(&odn->dn_tx_holds) == 0); 756 refcount_transfer(&ndn->dn_holds, &odn->dn_holds); 757 ASSERT(avl_is_empty(&ndn->dn_dbufs)); 758 avl_swap(&ndn->dn_dbufs, &odn->dn_dbufs); 759 ndn->dn_dbufs_count = odn->dn_dbufs_count; 760 ndn->dn_unlisted_l0_blkid = odn->dn_unlisted_l0_blkid; 761 ndn->dn_bonus = odn->dn_bonus; 762 ndn->dn_have_spill = odn->dn_have_spill; 763 ndn->dn_zio = odn->dn_zio; 764 ndn->dn_oldused = odn->dn_oldused; 765 ndn->dn_oldflags = odn->dn_oldflags; 766 ndn->dn_olduid = odn->dn_olduid; 767 ndn->dn_oldgid = odn->dn_oldgid; 768 ndn->dn_newuid = odn->dn_newuid; 769 ndn->dn_newgid = odn->dn_newgid; 770 ndn->dn_id_flags = odn->dn_id_flags; 771 dmu_zfetch_init(&ndn->dn_zfetch, NULL); 772 list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream); 773 ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode; 774 775 /* 776 * Update back pointers. Updating the handle fixes the back pointer of 777 * every descendant dbuf as well as the bonus dbuf. 778 */ 779 ASSERT(ndn->dn_handle->dnh_dnode == odn); 780 ndn->dn_handle->dnh_dnode = ndn; 781 if (ndn->dn_zfetch.zf_dnode == odn) { 782 ndn->dn_zfetch.zf_dnode = ndn; 783 } 784 785 /* 786 * Invalidate the original dnode by clearing all of its back pointers. 787 */ 788 odn->dn_dbuf = NULL; 789 odn->dn_handle = NULL; 790 avl_create(&odn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t), 791 offsetof(dmu_buf_impl_t, db_link)); 792 odn->dn_dbufs_count = 0; 793 odn->dn_unlisted_l0_blkid = 0; 794 odn->dn_bonus = NULL; 795 odn->dn_zfetch.zf_dnode = NULL; 796 797 /* 798 * Set the low bit of the objset pointer to ensure that dnode_move() 799 * recognizes the dnode as invalid in any subsequent callback. 800 */ 801 POINTER_INVALIDATE(&odn->dn_objset); 802 803 /* 804 * Satisfy the destructor. 805 */ 806 for (i = 0; i < TXG_SIZE; i++) { 807 list_create(&odn->dn_dirty_records[i], 808 sizeof (dbuf_dirty_record_t), 809 offsetof(dbuf_dirty_record_t, dr_dirty_node)); 810 odn->dn_free_ranges[i] = NULL; 811 odn->dn_next_nlevels[i] = 0; 812 odn->dn_next_indblkshift[i] = 0; 813 odn->dn_next_bonustype[i] = 0; 814 odn->dn_rm_spillblk[i] = 0; 815 odn->dn_next_bonuslen[i] = 0; 816 odn->dn_next_blksz[i] = 0; 817 } 818 odn->dn_allocated_txg = 0; 819 odn->dn_free_txg = 0; 820 odn->dn_assigned_txg = 0; 821 odn->dn_dirtyctx = 0; 822 odn->dn_dirtyctx_firstset = NULL; 823 odn->dn_have_spill = B_FALSE; 824 odn->dn_zio = NULL; 825 odn->dn_oldused = 0; 826 odn->dn_oldflags = 0; 827 odn->dn_olduid = 0; 828 odn->dn_oldgid = 0; 829 odn->dn_newuid = 0; 830 odn->dn_newgid = 0; 831 odn->dn_id_flags = 0; 832 833 /* 834 * Mark the dnode. 835 */ 836 ndn->dn_moved = 1; 837 odn->dn_moved = (uint8_t)-1; 838 } 839 840 #ifdef _KERNEL 841 /*ARGSUSED*/ 842 static kmem_cbrc_t 843 dnode_move(void *buf, void *newbuf, size_t size, void *arg) 844 { 845 dnode_t *odn = buf, *ndn = newbuf; 846 objset_t *os; 847 int64_t refcount; 848 uint32_t dbufs; 849 850 /* 851 * The dnode is on the objset's list of known dnodes if the objset 852 * pointer is valid. We set the low bit of the objset pointer when 853 * freeing the dnode to invalidate it, and the memory patterns written 854 * by kmem (baddcafe and deadbeef) set at least one of the two low bits. 855 * A newly created dnode sets the objset pointer last of all to indicate 856 * that the dnode is known and in a valid state to be moved by this 857 * function. 858 */ 859 os = odn->dn_objset; 860 if (!POINTER_IS_VALID(os)) { 861 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_invalid); 862 return (KMEM_CBRC_DONT_KNOW); 863 } 864 865 /* 866 * Ensure that the objset does not go away during the move. 867 */ 868 rw_enter(&os_lock, RW_WRITER); 869 if (os != odn->dn_objset) { 870 rw_exit(&os_lock); 871 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck1); 872 return (KMEM_CBRC_DONT_KNOW); 873 } 874 875 /* 876 * If the dnode is still valid, then so is the objset. We know that no 877 * valid objset can be freed while we hold os_lock, so we can safely 878 * ensure that the objset remains in use. 879 */ 880 mutex_enter(&os->os_lock); 881 882 /* 883 * Recheck the objset pointer in case the dnode was removed just before 884 * acquiring the lock. 885 */ 886 if (os != odn->dn_objset) { 887 mutex_exit(&os->os_lock); 888 rw_exit(&os_lock); 889 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck2); 890 return (KMEM_CBRC_DONT_KNOW); 891 } 892 893 /* 894 * At this point we know that as long as we hold os->os_lock, the dnode 895 * cannot be freed and fields within the dnode can be safely accessed. 896 * The objset listing this dnode cannot go away as long as this dnode is 897 * on its list. 898 */ 899 rw_exit(&os_lock); 900 if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) { 901 mutex_exit(&os->os_lock); 902 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_special); 903 return (KMEM_CBRC_NO); 904 } 905 ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */ 906 907 /* 908 * Lock the dnode handle to prevent the dnode from obtaining any new 909 * holds. This also prevents the descendant dbufs and the bonus dbuf 910 * from accessing the dnode, so that we can discount their holds. The 911 * handle is safe to access because we know that while the dnode cannot 912 * go away, neither can its handle. Once we hold dnh_zrlock, we can 913 * safely move any dnode referenced only by dbufs. 914 */ 915 if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) { 916 mutex_exit(&os->os_lock); 917 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_handle); 918 return (KMEM_CBRC_LATER); 919 } 920 921 /* 922 * Ensure a consistent view of the dnode's holds and the dnode's dbufs. 923 * We need to guarantee that there is a hold for every dbuf in order to 924 * determine whether the dnode is actively referenced. Falsely matching 925 * a dbuf to an active hold would lead to an unsafe move. It's possible 926 * that a thread already having an active dnode hold is about to add a 927 * dbuf, and we can't compare hold and dbuf counts while the add is in 928 * progress. 929 */ 930 if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) { 931 zrl_exit(&odn->dn_handle->dnh_zrlock); 932 mutex_exit(&os->os_lock); 933 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_rwlock); 934 return (KMEM_CBRC_LATER); 935 } 936 937 /* 938 * A dbuf may be removed (evicted) without an active dnode hold. In that 939 * case, the dbuf count is decremented under the handle lock before the 940 * dbuf's hold is released. This order ensures that if we count the hold 941 * after the dbuf is removed but before its hold is released, we will 942 * treat the unmatched hold as active and exit safely. If we count the 943 * hold before the dbuf is removed, the hold is discounted, and the 944 * removal is blocked until the move completes. 945 */ 946 refcount = refcount_count(&odn->dn_holds); 947 ASSERT(refcount >= 0); 948 dbufs = odn->dn_dbufs_count; 949 950 /* We can't have more dbufs than dnode holds. */ 951 ASSERT3U(dbufs, <=, refcount); 952 DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount, 953 uint32_t, dbufs); 954 955 if (refcount > dbufs) { 956 rw_exit(&odn->dn_struct_rwlock); 957 zrl_exit(&odn->dn_handle->dnh_zrlock); 958 mutex_exit(&os->os_lock); 959 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_active); 960 return (KMEM_CBRC_LATER); 961 } 962 963 rw_exit(&odn->dn_struct_rwlock); 964 965 /* 966 * At this point we know that anyone with a hold on the dnode is not 967 * actively referencing it. The dnode is known and in a valid state to 968 * move. We're holding the locks needed to execute the critical section. 969 */ 970 dnode_move_impl(odn, ndn); 971 972 list_link_replace(&odn->dn_link, &ndn->dn_link); 973 /* If the dnode was safe to move, the refcount cannot have changed. */ 974 ASSERT(refcount == refcount_count(&ndn->dn_holds)); 975 ASSERT(dbufs == ndn->dn_dbufs_count); 976 zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */ 977 mutex_exit(&os->os_lock); 978 979 return (KMEM_CBRC_YES); 980 } 981 #endif /* _KERNEL */ 982 983 void 984 dnode_special_close(dnode_handle_t *dnh) 985 { 986 dnode_t *dn = dnh->dnh_dnode; 987 988 /* 989 * Wait for final references to the dnode to clear. This can 990 * only happen if the arc is asyncronously evicting state that 991 * has a hold on this dnode while we are trying to evict this 992 * dnode. 993 */ 994 while (refcount_count(&dn->dn_holds) > 0) 995 delay(1); 996 ASSERT(dn->dn_dbuf == NULL || 997 dmu_buf_get_user(&dn->dn_dbuf->db) == NULL); 998 zrl_add(&dnh->dnh_zrlock); 999 dnode_destroy(dn); /* implicit zrl_remove() */ 1000 zrl_destroy(&dnh->dnh_zrlock); 1001 dnh->dnh_dnode = NULL; 1002 } 1003 1004 void 1005 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object, 1006 dnode_handle_t *dnh) 1007 { 1008 dnode_t *dn; 1009 1010 dn = dnode_create(os, dnp, NULL, object, dnh); 1011 zrl_init(&dnh->dnh_zrlock); 1012 DNODE_VERIFY(dn); 1013 } 1014 1015 static void 1016 dnode_buf_pageout(void *dbu) 1017 { 1018 dnode_children_t *children_dnodes = dbu; 1019 int i; 1020 1021 for (i = 0; i < children_dnodes->dnc_count; i++) { 1022 dnode_handle_t *dnh = &children_dnodes->dnc_children[i]; 1023 dnode_t *dn; 1024 1025 /* 1026 * The dnode handle lock guards against the dnode moving to 1027 * another valid address, so there is no need here to guard 1028 * against changes to or from NULL. 1029 */ 1030 if (dnh->dnh_dnode == NULL) { 1031 zrl_destroy(&dnh->dnh_zrlock); 1032 continue; 1033 } 1034 1035 zrl_add(&dnh->dnh_zrlock); 1036 dn = dnh->dnh_dnode; 1037 /* 1038 * If there are holds on this dnode, then there should 1039 * be holds on the dnode's containing dbuf as well; thus 1040 * it wouldn't be eligible for eviction and this function 1041 * would not have been called. 1042 */ 1043 ASSERT(refcount_is_zero(&dn->dn_holds)); 1044 ASSERT(refcount_is_zero(&dn->dn_tx_holds)); 1045 1046 dnode_destroy(dn); /* implicit zrl_remove() */ 1047 zrl_destroy(&dnh->dnh_zrlock); 1048 dnh->dnh_dnode = NULL; 1049 } 1050 kmem_free(children_dnodes, sizeof (dnode_children_t) + 1051 children_dnodes->dnc_count * sizeof (dnode_handle_t)); 1052 } 1053 1054 /* 1055 * errors: 1056 * EINVAL - invalid object number. 1057 * EIO - i/o error. 1058 * succeeds even for free dnodes. 1059 */ 1060 int 1061 dnode_hold_impl(objset_t *os, uint64_t object, int flag, 1062 void *tag, dnode_t **dnp) 1063 { 1064 int epb, idx, err; 1065 int drop_struct_lock = FALSE; 1066 int type; 1067 uint64_t blk; 1068 dnode_t *mdn, *dn; 1069 dmu_buf_impl_t *db; 1070 dnode_children_t *children_dnodes; 1071 dnode_handle_t *dnh; 1072 1073 /* 1074 * If you are holding the spa config lock as writer, you shouldn't 1075 * be asking the DMU to do *anything* unless it's the root pool 1076 * which may require us to read from the root filesystem while 1077 * holding some (not all) of the locks as writer. 1078 */ 1079 ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 || 1080 (spa_is_root(os->os_spa) && 1081 spa_config_held(os->os_spa, SCL_STATE, RW_WRITER))); 1082 1083 if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT) { 1084 dn = (object == DMU_USERUSED_OBJECT) ? 1085 DMU_USERUSED_DNODE(os) : DMU_GROUPUSED_DNODE(os); 1086 if (dn == NULL) 1087 return (SET_ERROR(ENOENT)); 1088 type = dn->dn_type; 1089 if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE) 1090 return (SET_ERROR(ENOENT)); 1091 if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE) 1092 return (SET_ERROR(EEXIST)); 1093 DNODE_VERIFY(dn); 1094 (void) refcount_add(&dn->dn_holds, tag); 1095 *dnp = dn; 1096 return (0); 1097 } 1098 1099 if (object == 0 || object >= DN_MAX_OBJECT) 1100 return (SET_ERROR(EINVAL)); 1101 1102 mdn = DMU_META_DNODE(os); 1103 ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT); 1104 1105 DNODE_VERIFY(mdn); 1106 1107 if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) { 1108 rw_enter(&mdn->dn_struct_rwlock, RW_READER); 1109 drop_struct_lock = TRUE; 1110 } 1111 1112 blk = dbuf_whichblock(mdn, 0, object * sizeof (dnode_phys_t)); 1113 1114 db = dbuf_hold(mdn, blk, FTAG); 1115 if (drop_struct_lock) 1116 rw_exit(&mdn->dn_struct_rwlock); 1117 if (db == NULL) 1118 return (SET_ERROR(EIO)); 1119 err = dbuf_read(db, NULL, DB_RF_CANFAIL); 1120 if (err) { 1121 dbuf_rele(db, FTAG); 1122 return (err); 1123 } 1124 1125 ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT); 1126 epb = db->db.db_size >> DNODE_SHIFT; 1127 1128 idx = object & (epb-1); 1129 1130 ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE); 1131 children_dnodes = dmu_buf_get_user(&db->db); 1132 if (children_dnodes == NULL) { 1133 int i; 1134 dnode_children_t *winner; 1135 children_dnodes = kmem_zalloc(sizeof (dnode_children_t) + 1136 epb * sizeof (dnode_handle_t), KM_SLEEP); 1137 children_dnodes->dnc_count = epb; 1138 dnh = &children_dnodes->dnc_children[0]; 1139 for (i = 0; i < epb; i++) { 1140 zrl_init(&dnh[i].dnh_zrlock); 1141 } 1142 dmu_buf_init_user(&children_dnodes->dnc_dbu, 1143 dnode_buf_pageout, NULL); 1144 winner = dmu_buf_set_user(&db->db, &children_dnodes->dnc_dbu); 1145 if (winner != NULL) { 1146 1147 for (i = 0; i < epb; i++) { 1148 zrl_destroy(&dnh[i].dnh_zrlock); 1149 } 1150 1151 kmem_free(children_dnodes, sizeof (dnode_children_t) + 1152 epb * sizeof (dnode_handle_t)); 1153 children_dnodes = winner; 1154 } 1155 } 1156 ASSERT(children_dnodes->dnc_count == epb); 1157 1158 dnh = &children_dnodes->dnc_children[idx]; 1159 zrl_add(&dnh->dnh_zrlock); 1160 dn = dnh->dnh_dnode; 1161 if (dn == NULL) { 1162 dnode_phys_t *phys = (dnode_phys_t *)db->db.db_data+idx; 1163 1164 dn = dnode_create(os, phys, db, object, dnh); 1165 } 1166 1167 mutex_enter(&dn->dn_mtx); 1168 type = dn->dn_type; 1169 if (dn->dn_free_txg || 1170 ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE) || 1171 ((flag & DNODE_MUST_BE_FREE) && 1172 (type != DMU_OT_NONE || !refcount_is_zero(&dn->dn_holds)))) { 1173 mutex_exit(&dn->dn_mtx); 1174 zrl_remove(&dnh->dnh_zrlock); 1175 dbuf_rele(db, FTAG); 1176 return (type == DMU_OT_NONE ? ENOENT : EEXIST); 1177 } 1178 if (refcount_add(&dn->dn_holds, tag) == 1) 1179 dbuf_add_ref(db, dnh); 1180 mutex_exit(&dn->dn_mtx); 1181 1182 /* Now we can rely on the hold to prevent the dnode from moving. */ 1183 zrl_remove(&dnh->dnh_zrlock); 1184 1185 DNODE_VERIFY(dn); 1186 ASSERT3P(dn->dn_dbuf, ==, db); 1187 ASSERT3U(dn->dn_object, ==, object); 1188 dbuf_rele(db, FTAG); 1189 1190 *dnp = dn; 1191 return (0); 1192 } 1193 1194 /* 1195 * Return held dnode if the object is allocated, NULL if not. 1196 */ 1197 int 1198 dnode_hold(objset_t *os, uint64_t object, void *tag, dnode_t **dnp) 1199 { 1200 return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, tag, dnp)); 1201 } 1202 1203 /* 1204 * Can only add a reference if there is already at least one 1205 * reference on the dnode. Returns FALSE if unable to add a 1206 * new reference. 1207 */ 1208 boolean_t 1209 dnode_add_ref(dnode_t *dn, void *tag) 1210 { 1211 mutex_enter(&dn->dn_mtx); 1212 if (refcount_is_zero(&dn->dn_holds)) { 1213 mutex_exit(&dn->dn_mtx); 1214 return (FALSE); 1215 } 1216 VERIFY(1 < refcount_add(&dn->dn_holds, tag)); 1217 mutex_exit(&dn->dn_mtx); 1218 return (TRUE); 1219 } 1220 1221 void 1222 dnode_rele(dnode_t *dn, void *tag) 1223 { 1224 mutex_enter(&dn->dn_mtx); 1225 dnode_rele_and_unlock(dn, tag); 1226 } 1227 1228 void 1229 dnode_rele_and_unlock(dnode_t *dn, void *tag) 1230 { 1231 uint64_t refs; 1232 /* Get while the hold prevents the dnode from moving. */ 1233 dmu_buf_impl_t *db = dn->dn_dbuf; 1234 dnode_handle_t *dnh = dn->dn_handle; 1235 1236 refs = refcount_remove(&dn->dn_holds, tag); 1237 mutex_exit(&dn->dn_mtx); 1238 1239 /* 1240 * It's unsafe to release the last hold on a dnode by dnode_rele() or 1241 * indirectly by dbuf_rele() while relying on the dnode handle to 1242 * prevent the dnode from moving, since releasing the last hold could 1243 * result in the dnode's parent dbuf evicting its dnode handles. For 1244 * that reason anyone calling dnode_rele() or dbuf_rele() without some 1245 * other direct or indirect hold on the dnode must first drop the dnode 1246 * handle. 1247 */ 1248 ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread); 1249 1250 /* NOTE: the DNODE_DNODE does not have a dn_dbuf */ 1251 if (refs == 0 && db != NULL) { 1252 /* 1253 * Another thread could add a hold to the dnode handle in 1254 * dnode_hold_impl() while holding the parent dbuf. Since the 1255 * hold on the parent dbuf prevents the handle from being 1256 * destroyed, the hold on the handle is OK. We can't yet assert 1257 * that the handle has zero references, but that will be 1258 * asserted anyway when the handle gets destroyed. 1259 */ 1260 dbuf_rele(db, dnh); 1261 } 1262 } 1263 1264 void 1265 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx) 1266 { 1267 objset_t *os = dn->dn_objset; 1268 uint64_t txg = tx->tx_txg; 1269 1270 if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) { 1271 dsl_dataset_dirty(os->os_dsl_dataset, tx); 1272 return; 1273 } 1274 1275 DNODE_VERIFY(dn); 1276 1277 #ifdef ZFS_DEBUG 1278 mutex_enter(&dn->dn_mtx); 1279 ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg); 1280 ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg); 1281 mutex_exit(&dn->dn_mtx); 1282 #endif 1283 1284 /* 1285 * Determine old uid/gid when necessary 1286 */ 1287 dmu_objset_userquota_get_ids(dn, B_TRUE, tx); 1288 1289 mutex_enter(&os->os_lock); 1290 1291 /* 1292 * If we are already marked dirty, we're done. 1293 */ 1294 if (list_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) { 1295 mutex_exit(&os->os_lock); 1296 return; 1297 } 1298 1299 ASSERT(!refcount_is_zero(&dn->dn_holds) || 1300 !avl_is_empty(&dn->dn_dbufs)); 1301 ASSERT(dn->dn_datablksz != 0); 1302 ASSERT0(dn->dn_next_bonuslen[txg&TXG_MASK]); 1303 ASSERT0(dn->dn_next_blksz[txg&TXG_MASK]); 1304 ASSERT0(dn->dn_next_bonustype[txg&TXG_MASK]); 1305 1306 dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n", 1307 dn->dn_object, txg); 1308 1309 if (dn->dn_free_txg > 0 && dn->dn_free_txg <= txg) { 1310 list_insert_tail(&os->os_free_dnodes[txg&TXG_MASK], dn); 1311 } else { 1312 list_insert_tail(&os->os_dirty_dnodes[txg&TXG_MASK], dn); 1313 } 1314 1315 mutex_exit(&os->os_lock); 1316 1317 /* 1318 * The dnode maintains a hold on its containing dbuf as 1319 * long as there are holds on it. Each instantiated child 1320 * dbuf maintains a hold on the dnode. When the last child 1321 * drops its hold, the dnode will drop its hold on the 1322 * containing dbuf. We add a "dirty hold" here so that the 1323 * dnode will hang around after we finish processing its 1324 * children. 1325 */ 1326 VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg)); 1327 1328 (void) dbuf_dirty(dn->dn_dbuf, tx); 1329 1330 dsl_dataset_dirty(os->os_dsl_dataset, tx); 1331 } 1332 1333 void 1334 dnode_free(dnode_t *dn, dmu_tx_t *tx) 1335 { 1336 int txgoff = tx->tx_txg & TXG_MASK; 1337 1338 dprintf("dn=%p txg=%llu\n", dn, tx->tx_txg); 1339 1340 /* we should be the only holder... hopefully */ 1341 /* ASSERT3U(refcount_count(&dn->dn_holds), ==, 1); */ 1342 1343 mutex_enter(&dn->dn_mtx); 1344 if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) { 1345 mutex_exit(&dn->dn_mtx); 1346 return; 1347 } 1348 dn->dn_free_txg = tx->tx_txg; 1349 mutex_exit(&dn->dn_mtx); 1350 1351 /* 1352 * If the dnode is already dirty, it needs to be moved from 1353 * the dirty list to the free list. 1354 */ 1355 mutex_enter(&dn->dn_objset->os_lock); 1356 if (list_link_active(&dn->dn_dirty_link[txgoff])) { 1357 list_remove(&dn->dn_objset->os_dirty_dnodes[txgoff], dn); 1358 list_insert_tail(&dn->dn_objset->os_free_dnodes[txgoff], dn); 1359 mutex_exit(&dn->dn_objset->os_lock); 1360 } else { 1361 mutex_exit(&dn->dn_objset->os_lock); 1362 dnode_setdirty(dn, tx); 1363 } 1364 } 1365 1366 /* 1367 * Try to change the block size for the indicated dnode. This can only 1368 * succeed if there are no blocks allocated or dirty beyond first block 1369 */ 1370 int 1371 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx) 1372 { 1373 dmu_buf_impl_t *db; 1374 int err; 1375 1376 ASSERT3U(size, <=, spa_maxblocksize(dmu_objset_spa(dn->dn_objset))); 1377 if (size == 0) 1378 size = SPA_MINBLOCKSIZE; 1379 else 1380 size = P2ROUNDUP(size, SPA_MINBLOCKSIZE); 1381 1382 if (ibs == dn->dn_indblkshift) 1383 ibs = 0; 1384 1385 if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0) 1386 return (0); 1387 1388 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 1389 1390 /* Check for any allocated blocks beyond the first */ 1391 if (dn->dn_maxblkid != 0) 1392 goto fail; 1393 1394 mutex_enter(&dn->dn_dbufs_mtx); 1395 for (db = avl_first(&dn->dn_dbufs); db != NULL; 1396 db = AVL_NEXT(&dn->dn_dbufs, db)) { 1397 if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID && 1398 db->db_blkid != DMU_SPILL_BLKID) { 1399 mutex_exit(&dn->dn_dbufs_mtx); 1400 goto fail; 1401 } 1402 } 1403 mutex_exit(&dn->dn_dbufs_mtx); 1404 1405 if (ibs && dn->dn_nlevels != 1) 1406 goto fail; 1407 1408 /* resize the old block */ 1409 err = dbuf_hold_impl(dn, 0, 0, TRUE, FALSE, FTAG, &db); 1410 if (err == 0) 1411 dbuf_new_size(db, size, tx); 1412 else if (err != ENOENT) 1413 goto fail; 1414 1415 dnode_setdblksz(dn, size); 1416 dnode_setdirty(dn, tx); 1417 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size; 1418 if (ibs) { 1419 dn->dn_indblkshift = ibs; 1420 dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs; 1421 } 1422 /* rele after we have fixed the blocksize in the dnode */ 1423 if (db) 1424 dbuf_rele(db, FTAG); 1425 1426 rw_exit(&dn->dn_struct_rwlock); 1427 return (0); 1428 1429 fail: 1430 rw_exit(&dn->dn_struct_rwlock); 1431 return (SET_ERROR(ENOTSUP)); 1432 } 1433 1434 /* read-holding callers must not rely on the lock being continuously held */ 1435 void 1436 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read) 1437 { 1438 uint64_t txgoff = tx->tx_txg & TXG_MASK; 1439 int epbs, new_nlevels; 1440 uint64_t sz; 1441 1442 ASSERT(blkid != DMU_BONUS_BLKID); 1443 1444 ASSERT(have_read ? 1445 RW_READ_HELD(&dn->dn_struct_rwlock) : 1446 RW_WRITE_HELD(&dn->dn_struct_rwlock)); 1447 1448 /* 1449 * if we have a read-lock, check to see if we need to do any work 1450 * before upgrading to a write-lock. 1451 */ 1452 if (have_read) { 1453 if (blkid <= dn->dn_maxblkid) 1454 return; 1455 1456 if (!rw_tryupgrade(&dn->dn_struct_rwlock)) { 1457 rw_exit(&dn->dn_struct_rwlock); 1458 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 1459 } 1460 } 1461 1462 if (blkid <= dn->dn_maxblkid) 1463 goto out; 1464 1465 dn->dn_maxblkid = blkid; 1466 1467 /* 1468 * Compute the number of levels necessary to support the new maxblkid. 1469 */ 1470 new_nlevels = 1; 1471 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1472 for (sz = dn->dn_nblkptr; 1473 sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs) 1474 new_nlevels++; 1475 1476 if (new_nlevels > dn->dn_nlevels) { 1477 int old_nlevels = dn->dn_nlevels; 1478 dmu_buf_impl_t *db; 1479 list_t *list; 1480 dbuf_dirty_record_t *new, *dr, *dr_next; 1481 1482 dn->dn_nlevels = new_nlevels; 1483 1484 ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]); 1485 dn->dn_next_nlevels[txgoff] = new_nlevels; 1486 1487 /* dirty the left indirects */ 1488 db = dbuf_hold_level(dn, old_nlevels, 0, FTAG); 1489 ASSERT(db != NULL); 1490 new = dbuf_dirty(db, tx); 1491 dbuf_rele(db, FTAG); 1492 1493 /* transfer the dirty records to the new indirect */ 1494 mutex_enter(&dn->dn_mtx); 1495 mutex_enter(&new->dt.di.dr_mtx); 1496 list = &dn->dn_dirty_records[txgoff]; 1497 for (dr = list_head(list); dr; dr = dr_next) { 1498 dr_next = list_next(&dn->dn_dirty_records[txgoff], dr); 1499 if (dr->dr_dbuf->db_level != new_nlevels-1 && 1500 dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID && 1501 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) { 1502 ASSERT(dr->dr_dbuf->db_level == old_nlevels-1); 1503 list_remove(&dn->dn_dirty_records[txgoff], dr); 1504 list_insert_tail(&new->dt.di.dr_children, dr); 1505 dr->dr_parent = new; 1506 } 1507 } 1508 mutex_exit(&new->dt.di.dr_mtx); 1509 mutex_exit(&dn->dn_mtx); 1510 } 1511 1512 out: 1513 if (have_read) 1514 rw_downgrade(&dn->dn_struct_rwlock); 1515 } 1516 1517 static void 1518 dnode_dirty_l1(dnode_t *dn, uint64_t l1blkid, dmu_tx_t *tx) 1519 { 1520 dmu_buf_impl_t *db = dbuf_hold_level(dn, 1, l1blkid, FTAG); 1521 if (db != NULL) { 1522 dmu_buf_will_dirty(&db->db, tx); 1523 dbuf_rele(db, FTAG); 1524 } 1525 } 1526 1527 void 1528 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx) 1529 { 1530 dmu_buf_impl_t *db; 1531 uint64_t blkoff, blkid, nblks; 1532 int blksz, blkshift, head, tail; 1533 int trunc = FALSE; 1534 int epbs; 1535 1536 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 1537 blksz = dn->dn_datablksz; 1538 blkshift = dn->dn_datablkshift; 1539 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1540 1541 if (len == DMU_OBJECT_END) { 1542 len = UINT64_MAX - off; 1543 trunc = TRUE; 1544 } 1545 1546 /* 1547 * First, block align the region to free: 1548 */ 1549 if (ISP2(blksz)) { 1550 head = P2NPHASE(off, blksz); 1551 blkoff = P2PHASE(off, blksz); 1552 if ((off >> blkshift) > dn->dn_maxblkid) 1553 goto out; 1554 } else { 1555 ASSERT(dn->dn_maxblkid == 0); 1556 if (off == 0 && len >= blksz) { 1557 /* 1558 * Freeing the whole block; fast-track this request. 1559 * Note that we won't dirty any indirect blocks, 1560 * which is fine because we will be freeing the entire 1561 * file and thus all indirect blocks will be freed 1562 * by free_children(). 1563 */ 1564 blkid = 0; 1565 nblks = 1; 1566 goto done; 1567 } else if (off >= blksz) { 1568 /* Freeing past end-of-data */ 1569 goto out; 1570 } else { 1571 /* Freeing part of the block. */ 1572 head = blksz - off; 1573 ASSERT3U(head, >, 0); 1574 } 1575 blkoff = off; 1576 } 1577 /* zero out any partial block data at the start of the range */ 1578 if (head) { 1579 ASSERT3U(blkoff + head, ==, blksz); 1580 if (len < head) 1581 head = len; 1582 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off), 1583 TRUE, FALSE, FTAG, &db) == 0) { 1584 caddr_t data; 1585 1586 /* don't dirty if it isn't on disk and isn't dirty */ 1587 if (db->db_last_dirty || 1588 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) { 1589 rw_exit(&dn->dn_struct_rwlock); 1590 dmu_buf_will_dirty(&db->db, tx); 1591 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 1592 data = db->db.db_data; 1593 bzero(data + blkoff, head); 1594 } 1595 dbuf_rele(db, FTAG); 1596 } 1597 off += head; 1598 len -= head; 1599 } 1600 1601 /* If the range was less than one block, we're done */ 1602 if (len == 0) 1603 goto out; 1604 1605 /* If the remaining range is past end of file, we're done */ 1606 if ((off >> blkshift) > dn->dn_maxblkid) 1607 goto out; 1608 1609 ASSERT(ISP2(blksz)); 1610 if (trunc) 1611 tail = 0; 1612 else 1613 tail = P2PHASE(len, blksz); 1614 1615 ASSERT0(P2PHASE(off, blksz)); 1616 /* zero out any partial block data at the end of the range */ 1617 if (tail) { 1618 if (len < tail) 1619 tail = len; 1620 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off+len), 1621 TRUE, FALSE, FTAG, &db) == 0) { 1622 /* don't dirty if not on disk and not dirty */ 1623 if (db->db_last_dirty || 1624 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) { 1625 rw_exit(&dn->dn_struct_rwlock); 1626 dmu_buf_will_dirty(&db->db, tx); 1627 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 1628 bzero(db->db.db_data, tail); 1629 } 1630 dbuf_rele(db, FTAG); 1631 } 1632 len -= tail; 1633 } 1634 1635 /* If the range did not include a full block, we are done */ 1636 if (len == 0) 1637 goto out; 1638 1639 ASSERT(IS_P2ALIGNED(off, blksz)); 1640 ASSERT(trunc || IS_P2ALIGNED(len, blksz)); 1641 blkid = off >> blkshift; 1642 nblks = len >> blkshift; 1643 if (trunc) 1644 nblks += 1; 1645 1646 /* 1647 * Dirty all the indirect blocks in this range. Note that only 1648 * the first and last indirect blocks can actually be written 1649 * (if they were partially freed) -- they must be dirtied, even if 1650 * they do not exist on disk yet. The interior blocks will 1651 * be freed by free_children(), so they will not actually be written. 1652 * Even though these interior blocks will not be written, we 1653 * dirty them for two reasons: 1654 * 1655 * - It ensures that the indirect blocks remain in memory until 1656 * syncing context. (They have already been prefetched by 1657 * dmu_tx_hold_free(), so we don't have to worry about reading 1658 * them serially here.) 1659 * 1660 * - The dirty space accounting will put pressure on the txg sync 1661 * mechanism to begin syncing, and to delay transactions if there 1662 * is a large amount of freeing. Even though these indirect 1663 * blocks will not be written, we could need to write the same 1664 * amount of space if we copy the freed BPs into deadlists. 1665 */ 1666 if (dn->dn_nlevels > 1) { 1667 uint64_t first, last; 1668 1669 first = blkid >> epbs; 1670 dnode_dirty_l1(dn, first, tx); 1671 if (trunc) 1672 last = dn->dn_maxblkid >> epbs; 1673 else 1674 last = (blkid + nblks - 1) >> epbs; 1675 if (last != first) 1676 dnode_dirty_l1(dn, last, tx); 1677 1678 int shift = dn->dn_datablkshift + dn->dn_indblkshift - 1679 SPA_BLKPTRSHIFT; 1680 for (uint64_t i = first + 1; i < last; i++) { 1681 /* 1682 * Set i to the blockid of the next non-hole 1683 * level-1 indirect block at or after i. Note 1684 * that dnode_next_offset() operates in terms of 1685 * level-0-equivalent bytes. 1686 */ 1687 uint64_t ibyte = i << shift; 1688 int err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK, 1689 &ibyte, 2, 1, 0); 1690 i = ibyte >> shift; 1691 if (i >= last) 1692 break; 1693 1694 /* 1695 * Normally we should not see an error, either 1696 * from dnode_next_offset() or dbuf_hold_level() 1697 * (except for ESRCH from dnode_next_offset). 1698 * If there is an i/o error, then when we read 1699 * this block in syncing context, it will use 1700 * ZIO_FLAG_MUSTSUCCEED, and thus hang/panic according 1701 * to the "failmode" property. dnode_next_offset() 1702 * doesn't have a flag to indicate MUSTSUCCEED. 1703 */ 1704 if (err != 0) 1705 break; 1706 1707 dnode_dirty_l1(dn, i, tx); 1708 } 1709 } 1710 1711 done: 1712 /* 1713 * Add this range to the dnode range list. 1714 * We will finish up this free operation in the syncing phase. 1715 */ 1716 mutex_enter(&dn->dn_mtx); 1717 int txgoff = tx->tx_txg & TXG_MASK; 1718 if (dn->dn_free_ranges[txgoff] == NULL) { 1719 dn->dn_free_ranges[txgoff] = 1720 range_tree_create(NULL, NULL, &dn->dn_mtx); 1721 } 1722 range_tree_clear(dn->dn_free_ranges[txgoff], blkid, nblks); 1723 range_tree_add(dn->dn_free_ranges[txgoff], blkid, nblks); 1724 dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n", 1725 blkid, nblks, tx->tx_txg); 1726 mutex_exit(&dn->dn_mtx); 1727 1728 dbuf_free_range(dn, blkid, blkid + nblks - 1, tx); 1729 dnode_setdirty(dn, tx); 1730 out: 1731 1732 rw_exit(&dn->dn_struct_rwlock); 1733 } 1734 1735 static boolean_t 1736 dnode_spill_freed(dnode_t *dn) 1737 { 1738 int i; 1739 1740 mutex_enter(&dn->dn_mtx); 1741 for (i = 0; i < TXG_SIZE; i++) { 1742 if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK) 1743 break; 1744 } 1745 mutex_exit(&dn->dn_mtx); 1746 return (i < TXG_SIZE); 1747 } 1748 1749 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */ 1750 uint64_t 1751 dnode_block_freed(dnode_t *dn, uint64_t blkid) 1752 { 1753 void *dp = spa_get_dsl(dn->dn_objset->os_spa); 1754 int i; 1755 1756 if (blkid == DMU_BONUS_BLKID) 1757 return (FALSE); 1758 1759 /* 1760 * If we're in the process of opening the pool, dp will not be 1761 * set yet, but there shouldn't be anything dirty. 1762 */ 1763 if (dp == NULL) 1764 return (FALSE); 1765 1766 if (dn->dn_free_txg) 1767 return (TRUE); 1768 1769 if (blkid == DMU_SPILL_BLKID) 1770 return (dnode_spill_freed(dn)); 1771 1772 mutex_enter(&dn->dn_mtx); 1773 for (i = 0; i < TXG_SIZE; i++) { 1774 if (dn->dn_free_ranges[i] != NULL && 1775 range_tree_contains(dn->dn_free_ranges[i], blkid, 1)) 1776 break; 1777 } 1778 mutex_exit(&dn->dn_mtx); 1779 return (i < TXG_SIZE); 1780 } 1781 1782 /* call from syncing context when we actually write/free space for this dnode */ 1783 void 1784 dnode_diduse_space(dnode_t *dn, int64_t delta) 1785 { 1786 uint64_t space; 1787 dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n", 1788 dn, dn->dn_phys, 1789 (u_longlong_t)dn->dn_phys->dn_used, 1790 (longlong_t)delta); 1791 1792 mutex_enter(&dn->dn_mtx); 1793 space = DN_USED_BYTES(dn->dn_phys); 1794 if (delta > 0) { 1795 ASSERT3U(space + delta, >=, space); /* no overflow */ 1796 } else { 1797 ASSERT3U(space, >=, -delta); /* no underflow */ 1798 } 1799 space += delta; 1800 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) { 1801 ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0); 1802 ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT)); 1803 dn->dn_phys->dn_used = space >> DEV_BSHIFT; 1804 } else { 1805 dn->dn_phys->dn_used = space; 1806 dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES; 1807 } 1808 mutex_exit(&dn->dn_mtx); 1809 } 1810 1811 /* 1812 * Call when we think we're going to write/free space in open context to track 1813 * the amount of memory in use by the currently open txg. 1814 */ 1815 void 1816 dnode_willuse_space(dnode_t *dn, int64_t space, dmu_tx_t *tx) 1817 { 1818 objset_t *os = dn->dn_objset; 1819 dsl_dataset_t *ds = os->os_dsl_dataset; 1820 int64_t aspace = spa_get_asize(os->os_spa, space); 1821 1822 if (ds != NULL) { 1823 dsl_dir_willuse_space(ds->ds_dir, aspace, tx); 1824 dsl_pool_dirty_space(dmu_tx_pool(tx), space, tx); 1825 } 1826 1827 dmu_tx_willuse_space(tx, aspace); 1828 } 1829 1830 /* 1831 * Scans a block at the indicated "level" looking for a hole or data, 1832 * depending on 'flags'. 1833 * 1834 * If level > 0, then we are scanning an indirect block looking at its 1835 * pointers. If level == 0, then we are looking at a block of dnodes. 1836 * 1837 * If we don't find what we are looking for in the block, we return ESRCH. 1838 * Otherwise, return with *offset pointing to the beginning (if searching 1839 * forwards) or end (if searching backwards) of the range covered by the 1840 * block pointer we matched on (or dnode). 1841 * 1842 * The basic search algorithm used below by dnode_next_offset() is to 1843 * use this function to search up the block tree (widen the search) until 1844 * we find something (i.e., we don't return ESRCH) and then search back 1845 * down the tree (narrow the search) until we reach our original search 1846 * level. 1847 */ 1848 static int 1849 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset, 1850 int lvl, uint64_t blkfill, uint64_t txg) 1851 { 1852 dmu_buf_impl_t *db = NULL; 1853 void *data = NULL; 1854 uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 1855 uint64_t epb = 1ULL << epbs; 1856 uint64_t minfill, maxfill; 1857 boolean_t hole; 1858 int i, inc, error, span; 1859 1860 dprintf("probing object %llu offset %llx level %d of %u\n", 1861 dn->dn_object, *offset, lvl, dn->dn_phys->dn_nlevels); 1862 1863 hole = ((flags & DNODE_FIND_HOLE) != 0); 1864 inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1; 1865 ASSERT(txg == 0 || !hole); 1866 1867 if (lvl == dn->dn_phys->dn_nlevels) { 1868 error = 0; 1869 epb = dn->dn_phys->dn_nblkptr; 1870 data = dn->dn_phys->dn_blkptr; 1871 } else { 1872 uint64_t blkid = dbuf_whichblock(dn, lvl, *offset); 1873 error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FALSE, FTAG, &db); 1874 if (error) { 1875 if (error != ENOENT) 1876 return (error); 1877 if (hole) 1878 return (0); 1879 /* 1880 * This can only happen when we are searching up 1881 * the block tree for data. We don't really need to 1882 * adjust the offset, as we will just end up looking 1883 * at the pointer to this block in its parent, and its 1884 * going to be unallocated, so we will skip over it. 1885 */ 1886 return (SET_ERROR(ESRCH)); 1887 } 1888 error = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_HAVESTRUCT); 1889 if (error) { 1890 dbuf_rele(db, FTAG); 1891 return (error); 1892 } 1893 data = db->db.db_data; 1894 } 1895 1896 1897 if (db != NULL && txg != 0 && (db->db_blkptr == NULL || 1898 db->db_blkptr->blk_birth <= txg || 1899 BP_IS_HOLE(db->db_blkptr))) { 1900 /* 1901 * This can only happen when we are searching up the tree 1902 * and these conditions mean that we need to keep climbing. 1903 */ 1904 error = SET_ERROR(ESRCH); 1905 } else if (lvl == 0) { 1906 dnode_phys_t *dnp = data; 1907 span = DNODE_SHIFT; 1908 ASSERT(dn->dn_type == DMU_OT_DNODE); 1909 1910 for (i = (*offset >> span) & (blkfill - 1); 1911 i >= 0 && i < blkfill; i += inc) { 1912 if ((dnp[i].dn_type == DMU_OT_NONE) == hole) 1913 break; 1914 *offset += (1ULL << span) * inc; 1915 } 1916 if (i < 0 || i == blkfill) 1917 error = SET_ERROR(ESRCH); 1918 } else { 1919 blkptr_t *bp = data; 1920 uint64_t start = *offset; 1921 span = (lvl - 1) * epbs + dn->dn_datablkshift; 1922 minfill = 0; 1923 maxfill = blkfill << ((lvl - 1) * epbs); 1924 1925 if (hole) 1926 maxfill--; 1927 else 1928 minfill++; 1929 1930 *offset = *offset >> span; 1931 for (i = BF64_GET(*offset, 0, epbs); 1932 i >= 0 && i < epb; i += inc) { 1933 if (BP_GET_FILL(&bp[i]) >= minfill && 1934 BP_GET_FILL(&bp[i]) <= maxfill && 1935 (hole || bp[i].blk_birth > txg)) 1936 break; 1937 if (inc > 0 || *offset > 0) 1938 *offset += inc; 1939 } 1940 *offset = *offset << span; 1941 if (inc < 0) { 1942 /* traversing backwards; position offset at the end */ 1943 ASSERT3U(*offset, <=, start); 1944 *offset = MIN(*offset + (1ULL << span) - 1, start); 1945 } else if (*offset < start) { 1946 *offset = start; 1947 } 1948 if (i < 0 || i >= epb) 1949 error = SET_ERROR(ESRCH); 1950 } 1951 1952 if (db) 1953 dbuf_rele(db, FTAG); 1954 1955 return (error); 1956 } 1957 1958 /* 1959 * Find the next hole, data, or sparse region at or after *offset. 1960 * The value 'blkfill' tells us how many items we expect to find 1961 * in an L0 data block; this value is 1 for normal objects, 1962 * DNODES_PER_BLOCK for the meta dnode, and some fraction of 1963 * DNODES_PER_BLOCK when searching for sparse regions thereof. 1964 * 1965 * Examples: 1966 * 1967 * dnode_next_offset(dn, flags, offset, 1, 1, 0); 1968 * Finds the next/previous hole/data in a file. 1969 * Used in dmu_offset_next(). 1970 * 1971 * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg); 1972 * Finds the next free/allocated dnode an objset's meta-dnode. 1973 * Only finds objects that have new contents since txg (ie. 1974 * bonus buffer changes and content removal are ignored). 1975 * Used in dmu_object_next(). 1976 * 1977 * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0); 1978 * Finds the next L2 meta-dnode bp that's at most 1/4 full. 1979 * Used in dmu_object_alloc(). 1980 */ 1981 int 1982 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset, 1983 int minlvl, uint64_t blkfill, uint64_t txg) 1984 { 1985 uint64_t initial_offset = *offset; 1986 int lvl, maxlvl; 1987 int error = 0; 1988 1989 if (!(flags & DNODE_FIND_HAVELOCK)) 1990 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1991 1992 if (dn->dn_phys->dn_nlevels == 0) { 1993 error = SET_ERROR(ESRCH); 1994 goto out; 1995 } 1996 1997 if (dn->dn_datablkshift == 0) { 1998 if (*offset < dn->dn_datablksz) { 1999 if (flags & DNODE_FIND_HOLE) 2000 *offset = dn->dn_datablksz; 2001 } else { 2002 error = SET_ERROR(ESRCH); 2003 } 2004 goto out; 2005 } 2006 2007 maxlvl = dn->dn_phys->dn_nlevels; 2008 2009 for (lvl = minlvl; lvl <= maxlvl; lvl++) { 2010 error = dnode_next_offset_level(dn, 2011 flags, offset, lvl, blkfill, txg); 2012 if (error != ESRCH) 2013 break; 2014 } 2015 2016 while (error == 0 && --lvl >= minlvl) { 2017 error = dnode_next_offset_level(dn, 2018 flags, offset, lvl, blkfill, txg); 2019 } 2020 2021 /* 2022 * There's always a "virtual hole" at the end of the object, even 2023 * if all BP's which physically exist are non-holes. 2024 */ 2025 if ((flags & DNODE_FIND_HOLE) && error == ESRCH && txg == 0 && 2026 minlvl == 1 && blkfill == 1 && !(flags & DNODE_FIND_BACKWARDS)) { 2027 error = 0; 2028 } 2029 2030 if (error == 0 && (flags & DNODE_FIND_BACKWARDS ? 2031 initial_offset < *offset : initial_offset > *offset)) 2032 error = SET_ERROR(ESRCH); 2033 out: 2034 if (!(flags & DNODE_FIND_HAVELOCK)) 2035 rw_exit(&dn->dn_struct_rwlock); 2036 2037 return (error); 2038 } 2039