dnode.c revision 260763
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) 2013 by Delphix. All rights reserved. 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 39static int free_range_compar(const void *node1, const void *node2); 40 41static 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 56static dnode_phys_t dnode_phys_zero; 57 58int zfs_default_bs = SPA_MINBLOCKSHIFT; 59int zfs_default_ibs = DN_MAX_INDBLKSHIFT; 60 61#ifdef sun 62static kmem_cbrc_t dnode_move(void *, void *, size_t, void *); 63#endif 64 65/* ARGSUSED */ 66static int 67dnode_cons(void *arg, void *unused, int kmflag) 68{ 69 dnode_t *dn = arg; 70 int i; 71 72 rw_init(&dn->dn_struct_rwlock, NULL, RW_DEFAULT, NULL); 73 mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL); 74 mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL); 75 cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL); 76 77 /* 78 * Every dbuf has a reference, and dropping a tracked reference is 79 * O(number of references), so don't track dn_holds. 80 */ 81 refcount_create_untracked(&dn->dn_holds); 82 refcount_create(&dn->dn_tx_holds); 83 list_link_init(&dn->dn_link); 84 85 bzero(&dn->dn_next_nblkptr[0], sizeof (dn->dn_next_nblkptr)); 86 bzero(&dn->dn_next_nlevels[0], sizeof (dn->dn_next_nlevels)); 87 bzero(&dn->dn_next_indblkshift[0], sizeof (dn->dn_next_indblkshift)); 88 bzero(&dn->dn_next_bonustype[0], sizeof (dn->dn_next_bonustype)); 89 bzero(&dn->dn_rm_spillblk[0], sizeof (dn->dn_rm_spillblk)); 90 bzero(&dn->dn_next_bonuslen[0], sizeof (dn->dn_next_bonuslen)); 91 bzero(&dn->dn_next_blksz[0], sizeof (dn->dn_next_blksz)); 92 93 for (i = 0; i < TXG_SIZE; i++) { 94 list_link_init(&dn->dn_dirty_link[i]); 95 avl_create(&dn->dn_ranges[i], free_range_compar, 96 sizeof (free_range_t), 97 offsetof(struct free_range, fr_node)); 98 list_create(&dn->dn_dirty_records[i], 99 sizeof (dbuf_dirty_record_t), 100 offsetof(dbuf_dirty_record_t, dr_dirty_node)); 101 } 102 103 dn->dn_allocated_txg = 0; 104 dn->dn_free_txg = 0; 105 dn->dn_assigned_txg = 0; 106 dn->dn_dirtyctx = 0; 107 dn->dn_dirtyctx_firstset = NULL; 108 dn->dn_bonus = NULL; 109 dn->dn_have_spill = B_FALSE; 110 dn->dn_zio = NULL; 111 dn->dn_oldused = 0; 112 dn->dn_oldflags = 0; 113 dn->dn_olduid = 0; 114 dn->dn_oldgid = 0; 115 dn->dn_newuid = 0; 116 dn->dn_newgid = 0; 117 dn->dn_id_flags = 0; 118 119 dn->dn_dbufs_count = 0; 120 dn->dn_unlisted_l0_blkid = 0; 121 list_create(&dn->dn_dbufs, sizeof (dmu_buf_impl_t), 122 offsetof(dmu_buf_impl_t, db_link)); 123 124 dn->dn_moved = 0; 125 POINTER_INVALIDATE(&dn->dn_objset); 126 return (0); 127} 128 129/* ARGSUSED */ 130static void 131dnode_dest(void *arg, void *unused) 132{ 133 int i; 134 dnode_t *dn = arg; 135 136 rw_destroy(&dn->dn_struct_rwlock); 137 mutex_destroy(&dn->dn_mtx); 138 mutex_destroy(&dn->dn_dbufs_mtx); 139 cv_destroy(&dn->dn_notxholds); 140 refcount_destroy(&dn->dn_holds); 141 refcount_destroy(&dn->dn_tx_holds); 142 ASSERT(!list_link_active(&dn->dn_link)); 143 144 for (i = 0; i < TXG_SIZE; i++) { 145 ASSERT(!list_link_active(&dn->dn_dirty_link[i])); 146 avl_destroy(&dn->dn_ranges[i]); 147 list_destroy(&dn->dn_dirty_records[i]); 148 ASSERT0(dn->dn_next_nblkptr[i]); 149 ASSERT0(dn->dn_next_nlevels[i]); 150 ASSERT0(dn->dn_next_indblkshift[i]); 151 ASSERT0(dn->dn_next_bonustype[i]); 152 ASSERT0(dn->dn_rm_spillblk[i]); 153 ASSERT0(dn->dn_next_bonuslen[i]); 154 ASSERT0(dn->dn_next_blksz[i]); 155 } 156 157 ASSERT0(dn->dn_allocated_txg); 158 ASSERT0(dn->dn_free_txg); 159 ASSERT0(dn->dn_assigned_txg); 160 ASSERT0(dn->dn_dirtyctx); 161 ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL); 162 ASSERT3P(dn->dn_bonus, ==, NULL); 163 ASSERT(!dn->dn_have_spill); 164 ASSERT3P(dn->dn_zio, ==, NULL); 165 ASSERT0(dn->dn_oldused); 166 ASSERT0(dn->dn_oldflags); 167 ASSERT0(dn->dn_olduid); 168 ASSERT0(dn->dn_oldgid); 169 ASSERT0(dn->dn_newuid); 170 ASSERT0(dn->dn_newgid); 171 ASSERT0(dn->dn_id_flags); 172 173 ASSERT0(dn->dn_dbufs_count); 174 ASSERT0(dn->dn_unlisted_l0_blkid); 175 list_destroy(&dn->dn_dbufs); 176} 177 178void 179dnode_init(void) 180{ 181 ASSERT(dnode_cache == NULL); 182 dnode_cache = kmem_cache_create("dnode_t", 183 sizeof (dnode_t), 184 0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0); 185 kmem_cache_set_move(dnode_cache, dnode_move); 186} 187 188void 189dnode_fini(void) 190{ 191 kmem_cache_destroy(dnode_cache); 192 dnode_cache = NULL; 193} 194 195 196#ifdef ZFS_DEBUG 197void 198dnode_verify(dnode_t *dn) 199{ 200 int drop_struct_lock = FALSE; 201 202 ASSERT(dn->dn_phys); 203 ASSERT(dn->dn_objset); 204 ASSERT(dn->dn_handle->dnh_dnode == dn); 205 206 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type)); 207 208 if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY)) 209 return; 210 211 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) { 212 rw_enter(&dn->dn_struct_rwlock, RW_READER); 213 drop_struct_lock = TRUE; 214 } 215 if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) { 216 int i; 217 ASSERT3U(dn->dn_indblkshift, >=, 0); 218 ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT); 219 if (dn->dn_datablkshift) { 220 ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT); 221 ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT); 222 ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz); 223 } 224 ASSERT3U(dn->dn_nlevels, <=, 30); 225 ASSERT(DMU_OT_IS_VALID(dn->dn_type)); 226 ASSERT3U(dn->dn_nblkptr, >=, 1); 227 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR); 228 ASSERT3U(dn->dn_bonuslen, <=, DN_MAX_BONUSLEN); 229 ASSERT3U(dn->dn_datablksz, ==, 230 dn->dn_datablkszsec << SPA_MINBLOCKSHIFT); 231 ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0); 232 ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) + 233 dn->dn_bonuslen, <=, DN_MAX_BONUSLEN); 234 for (i = 0; i < TXG_SIZE; i++) { 235 ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels); 236 } 237 } 238 if (dn->dn_phys->dn_type != DMU_OT_NONE) 239 ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels); 240 ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL); 241 if (dn->dn_dbuf != NULL) { 242 ASSERT3P(dn->dn_phys, ==, 243 (dnode_phys_t *)dn->dn_dbuf->db.db_data + 244 (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT))); 245 } 246 if (drop_struct_lock) 247 rw_exit(&dn->dn_struct_rwlock); 248} 249#endif 250 251void 252dnode_byteswap(dnode_phys_t *dnp) 253{ 254 uint64_t *buf64 = (void*)&dnp->dn_blkptr; 255 int i; 256 257 if (dnp->dn_type == DMU_OT_NONE) { 258 bzero(dnp, sizeof (dnode_phys_t)); 259 return; 260 } 261 262 dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec); 263 dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen); 264 dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid); 265 dnp->dn_used = BSWAP_64(dnp->dn_used); 266 267 /* 268 * dn_nblkptr is only one byte, so it's OK to read it in either 269 * byte order. We can't read dn_bouslen. 270 */ 271 ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT); 272 ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR); 273 for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++) 274 buf64[i] = BSWAP_64(buf64[i]); 275 276 /* 277 * OK to check dn_bonuslen for zero, because it won't matter if 278 * we have the wrong byte order. This is necessary because the 279 * dnode dnode is smaller than a regular dnode. 280 */ 281 if (dnp->dn_bonuslen != 0) { 282 /* 283 * Note that the bonus length calculated here may be 284 * longer than the actual bonus buffer. This is because 285 * we always put the bonus buffer after the last block 286 * pointer (instead of packing it against the end of the 287 * dnode buffer). 288 */ 289 int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t); 290 size_t len = DN_MAX_BONUSLEN - off; 291 ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype)); 292 dmu_object_byteswap_t byteswap = 293 DMU_OT_BYTESWAP(dnp->dn_bonustype); 294 dmu_ot_byteswap[byteswap].ob_func(dnp->dn_bonus + off, len); 295 } 296 297 /* Swap SPILL block if we have one */ 298 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) 299 byteswap_uint64_array(&dnp->dn_spill, sizeof (blkptr_t)); 300 301} 302 303void 304dnode_buf_byteswap(void *vbuf, size_t size) 305{ 306 dnode_phys_t *buf = vbuf; 307 int i; 308 309 ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT)); 310 ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0); 311 312 size >>= DNODE_SHIFT; 313 for (i = 0; i < size; i++) { 314 dnode_byteswap(buf); 315 buf++; 316 } 317} 318 319static int 320free_range_compar(const void *node1, const void *node2) 321{ 322 const free_range_t *rp1 = node1; 323 const free_range_t *rp2 = node2; 324 325 if (rp1->fr_blkid < rp2->fr_blkid) 326 return (-1); 327 else if (rp1->fr_blkid > rp2->fr_blkid) 328 return (1); 329 else return (0); 330} 331 332void 333dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx) 334{ 335 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1); 336 337 dnode_setdirty(dn, tx); 338 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 339 ASSERT3U(newsize, <=, DN_MAX_BONUSLEN - 340 (dn->dn_nblkptr-1) * sizeof (blkptr_t)); 341 dn->dn_bonuslen = newsize; 342 if (newsize == 0) 343 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN; 344 else 345 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen; 346 rw_exit(&dn->dn_struct_rwlock); 347} 348 349void 350dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx) 351{ 352 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1); 353 dnode_setdirty(dn, tx); 354 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 355 dn->dn_bonustype = newtype; 356 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype; 357 rw_exit(&dn->dn_struct_rwlock); 358} 359 360void 361dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx) 362{ 363 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1); 364 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 365 dnode_setdirty(dn, tx); 366 dn->dn_rm_spillblk[tx->tx_txg&TXG_MASK] = DN_KILL_SPILLBLK; 367 dn->dn_have_spill = B_FALSE; 368} 369 370static void 371dnode_setdblksz(dnode_t *dn, int size) 372{ 373 ASSERT0(P2PHASE(size, SPA_MINBLOCKSIZE)); 374 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE); 375 ASSERT3U(size, >=, SPA_MINBLOCKSIZE); 376 ASSERT3U(size >> SPA_MINBLOCKSHIFT, <, 377 1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8)); 378 dn->dn_datablksz = size; 379 dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT; 380 dn->dn_datablkshift = ISP2(size) ? highbit(size - 1) : 0; 381} 382 383static dnode_t * 384dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db, 385 uint64_t object, dnode_handle_t *dnh) 386{ 387 dnode_t *dn = kmem_cache_alloc(dnode_cache, KM_SLEEP); 388 389 ASSERT(!POINTER_IS_VALID(dn->dn_objset)); 390 dn->dn_moved = 0; 391 392 /* 393 * Defer setting dn_objset until the dnode is ready to be a candidate 394 * for the dnode_move() callback. 395 */ 396 dn->dn_object = object; 397 dn->dn_dbuf = db; 398 dn->dn_handle = dnh; 399 dn->dn_phys = dnp; 400 401 if (dnp->dn_datablkszsec) { 402 dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT); 403 } else { 404 dn->dn_datablksz = 0; 405 dn->dn_datablkszsec = 0; 406 dn->dn_datablkshift = 0; 407 } 408 dn->dn_indblkshift = dnp->dn_indblkshift; 409 dn->dn_nlevels = dnp->dn_nlevels; 410 dn->dn_type = dnp->dn_type; 411 dn->dn_nblkptr = dnp->dn_nblkptr; 412 dn->dn_checksum = dnp->dn_checksum; 413 dn->dn_compress = dnp->dn_compress; 414 dn->dn_bonustype = dnp->dn_bonustype; 415 dn->dn_bonuslen = dnp->dn_bonuslen; 416 dn->dn_maxblkid = dnp->dn_maxblkid; 417 dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0); 418 dn->dn_id_flags = 0; 419 420 dmu_zfetch_init(&dn->dn_zfetch, dn); 421 422 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type)); 423 424 mutex_enter(&os->os_lock); 425 list_insert_head(&os->os_dnodes, dn); 426 membar_producer(); 427 /* 428 * Everything else must be valid before assigning dn_objset makes the 429 * dnode eligible for dnode_move(). 430 */ 431 dn->dn_objset = os; 432 mutex_exit(&os->os_lock); 433 434 arc_space_consume(sizeof (dnode_t), ARC_SPACE_OTHER); 435 return (dn); 436} 437 438/* 439 * Caller must be holding the dnode handle, which is released upon return. 440 */ 441static void 442dnode_destroy(dnode_t *dn) 443{ 444 objset_t *os = dn->dn_objset; 445 446 ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0); 447 448 mutex_enter(&os->os_lock); 449 POINTER_INVALIDATE(&dn->dn_objset); 450 list_remove(&os->os_dnodes, dn); 451 mutex_exit(&os->os_lock); 452 453 /* the dnode can no longer move, so we can release the handle */ 454 zrl_remove(&dn->dn_handle->dnh_zrlock); 455 456 dn->dn_allocated_txg = 0; 457 dn->dn_free_txg = 0; 458 dn->dn_assigned_txg = 0; 459 460 dn->dn_dirtyctx = 0; 461 if (dn->dn_dirtyctx_firstset != NULL) { 462 kmem_free(dn->dn_dirtyctx_firstset, 1); 463 dn->dn_dirtyctx_firstset = NULL; 464 } 465 if (dn->dn_bonus != NULL) { 466 mutex_enter(&dn->dn_bonus->db_mtx); 467 dbuf_evict(dn->dn_bonus); 468 dn->dn_bonus = NULL; 469 } 470 dn->dn_zio = NULL; 471 472 dn->dn_have_spill = B_FALSE; 473 dn->dn_oldused = 0; 474 dn->dn_oldflags = 0; 475 dn->dn_olduid = 0; 476 dn->dn_oldgid = 0; 477 dn->dn_newuid = 0; 478 dn->dn_newgid = 0; 479 dn->dn_id_flags = 0; 480 dn->dn_unlisted_l0_blkid = 0; 481 482 dmu_zfetch_rele(&dn->dn_zfetch); 483 kmem_cache_free(dnode_cache, dn); 484 arc_space_return(sizeof (dnode_t), ARC_SPACE_OTHER); 485} 486 487void 488dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs, 489 dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx) 490{ 491 int i; 492 493 if (blocksize == 0) 494 blocksize = 1 << zfs_default_bs; 495 else if (blocksize > SPA_MAXBLOCKSIZE) 496 blocksize = SPA_MAXBLOCKSIZE; 497 else 498 blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE); 499 500 if (ibs == 0) 501 ibs = zfs_default_ibs; 502 503 ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT); 504 505 dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d\n", dn->dn_objset, 506 dn->dn_object, tx->tx_txg, blocksize, ibs); 507 508 ASSERT(dn->dn_type == DMU_OT_NONE); 509 ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0); 510 ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE); 511 ASSERT(ot != DMU_OT_NONE); 512 ASSERT(DMU_OT_IS_VALID(ot)); 513 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) || 514 (bonustype == DMU_OT_SA && bonuslen == 0) || 515 (bonustype != DMU_OT_NONE && bonuslen != 0)); 516 ASSERT(DMU_OT_IS_VALID(bonustype)); 517 ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN); 518 ASSERT(dn->dn_type == DMU_OT_NONE); 519 ASSERT0(dn->dn_maxblkid); 520 ASSERT0(dn->dn_allocated_txg); 521 ASSERT0(dn->dn_assigned_txg); 522 ASSERT(refcount_is_zero(&dn->dn_tx_holds)); 523 ASSERT3U(refcount_count(&dn->dn_holds), <=, 1); 524 ASSERT3P(list_head(&dn->dn_dbufs), ==, NULL); 525 526 for (i = 0; i < TXG_SIZE; i++) { 527 ASSERT0(dn->dn_next_nblkptr[i]); 528 ASSERT0(dn->dn_next_nlevels[i]); 529 ASSERT0(dn->dn_next_indblkshift[i]); 530 ASSERT0(dn->dn_next_bonuslen[i]); 531 ASSERT0(dn->dn_next_bonustype[i]); 532 ASSERT0(dn->dn_rm_spillblk[i]); 533 ASSERT0(dn->dn_next_blksz[i]); 534 ASSERT(!list_link_active(&dn->dn_dirty_link[i])); 535 ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL); 536 ASSERT0(avl_numnodes(&dn->dn_ranges[i])); 537 } 538 539 dn->dn_type = ot; 540 dnode_setdblksz(dn, blocksize); 541 dn->dn_indblkshift = ibs; 542 dn->dn_nlevels = 1; 543 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */ 544 dn->dn_nblkptr = 1; 545 else 546 dn->dn_nblkptr = 1 + 547 ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT); 548 dn->dn_bonustype = bonustype; 549 dn->dn_bonuslen = bonuslen; 550 dn->dn_checksum = ZIO_CHECKSUM_INHERIT; 551 dn->dn_compress = ZIO_COMPRESS_INHERIT; 552 dn->dn_dirtyctx = 0; 553 554 dn->dn_free_txg = 0; 555 if (dn->dn_dirtyctx_firstset) { 556 kmem_free(dn->dn_dirtyctx_firstset, 1); 557 dn->dn_dirtyctx_firstset = NULL; 558 } 559 560 dn->dn_allocated_txg = tx->tx_txg; 561 dn->dn_id_flags = 0; 562 563 dnode_setdirty(dn, tx); 564 dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs; 565 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen; 566 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype; 567 dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz; 568} 569 570void 571dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, 572 dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx) 573{ 574 int nblkptr; 575 576 ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE); 577 ASSERT3U(blocksize, <=, SPA_MAXBLOCKSIZE); 578 ASSERT0(blocksize % SPA_MINBLOCKSIZE); 579 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx)); 580 ASSERT(tx->tx_txg != 0); 581 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) || 582 (bonustype != DMU_OT_NONE && bonuslen != 0) || 583 (bonustype == DMU_OT_SA && bonuslen == 0)); 584 ASSERT(DMU_OT_IS_VALID(bonustype)); 585 ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN); 586 587 /* clean up any unreferenced dbufs */ 588 dnode_evict_dbufs(dn); 589 590 dn->dn_id_flags = 0; 591 592 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 593 dnode_setdirty(dn, tx); 594 if (dn->dn_datablksz != blocksize) { 595 /* change blocksize */ 596 ASSERT(dn->dn_maxblkid == 0 && 597 (BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) || 598 dnode_block_freed(dn, 0))); 599 dnode_setdblksz(dn, blocksize); 600 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = blocksize; 601 } 602 if (dn->dn_bonuslen != bonuslen) 603 dn->dn_next_bonuslen[tx->tx_txg&TXG_MASK] = bonuslen; 604 605 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */ 606 nblkptr = 1; 607 else 608 nblkptr = 1 + ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT); 609 if (dn->dn_bonustype != bonustype) 610 dn->dn_next_bonustype[tx->tx_txg&TXG_MASK] = bonustype; 611 if (dn->dn_nblkptr != nblkptr) 612 dn->dn_next_nblkptr[tx->tx_txg&TXG_MASK] = nblkptr; 613 if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR) { 614 dbuf_rm_spill(dn, tx); 615 dnode_rm_spill(dn, tx); 616 } 617 rw_exit(&dn->dn_struct_rwlock); 618 619 /* change type */ 620 dn->dn_type = ot; 621 622 /* change bonus size and type */ 623 mutex_enter(&dn->dn_mtx); 624 dn->dn_bonustype = bonustype; 625 dn->dn_bonuslen = bonuslen; 626 dn->dn_nblkptr = nblkptr; 627 dn->dn_checksum = ZIO_CHECKSUM_INHERIT; 628 dn->dn_compress = ZIO_COMPRESS_INHERIT; 629 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR); 630 631 /* fix up the bonus db_size */ 632 if (dn->dn_bonus) { 633 dn->dn_bonus->db.db_size = 634 DN_MAX_BONUSLEN - (dn->dn_nblkptr-1) * sizeof (blkptr_t); 635 ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size); 636 } 637 638 dn->dn_allocated_txg = tx->tx_txg; 639 mutex_exit(&dn->dn_mtx); 640} 641 642#ifdef DNODE_STATS 643static struct { 644 uint64_t dms_dnode_invalid; 645 uint64_t dms_dnode_recheck1; 646 uint64_t dms_dnode_recheck2; 647 uint64_t dms_dnode_special; 648 uint64_t dms_dnode_handle; 649 uint64_t dms_dnode_rwlock; 650 uint64_t dms_dnode_active; 651} dnode_move_stats; 652#endif /* DNODE_STATS */ 653 654static void 655dnode_move_impl(dnode_t *odn, dnode_t *ndn) 656{ 657 int i; 658 659 ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock)); 660 ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx)); 661 ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx)); 662 ASSERT(!RW_LOCK_HELD(&odn->dn_zfetch.zf_rwlock)); 663 664 /* Copy fields. */ 665 ndn->dn_objset = odn->dn_objset; 666 ndn->dn_object = odn->dn_object; 667 ndn->dn_dbuf = odn->dn_dbuf; 668 ndn->dn_handle = odn->dn_handle; 669 ndn->dn_phys = odn->dn_phys; 670 ndn->dn_type = odn->dn_type; 671 ndn->dn_bonuslen = odn->dn_bonuslen; 672 ndn->dn_bonustype = odn->dn_bonustype; 673 ndn->dn_nblkptr = odn->dn_nblkptr; 674 ndn->dn_checksum = odn->dn_checksum; 675 ndn->dn_compress = odn->dn_compress; 676 ndn->dn_nlevels = odn->dn_nlevels; 677 ndn->dn_indblkshift = odn->dn_indblkshift; 678 ndn->dn_datablkshift = odn->dn_datablkshift; 679 ndn->dn_datablkszsec = odn->dn_datablkszsec; 680 ndn->dn_datablksz = odn->dn_datablksz; 681 ndn->dn_maxblkid = odn->dn_maxblkid; 682 bcopy(&odn->dn_next_nblkptr[0], &ndn->dn_next_nblkptr[0], 683 sizeof (odn->dn_next_nblkptr)); 684 bcopy(&odn->dn_next_nlevels[0], &ndn->dn_next_nlevels[0], 685 sizeof (odn->dn_next_nlevels)); 686 bcopy(&odn->dn_next_indblkshift[0], &ndn->dn_next_indblkshift[0], 687 sizeof (odn->dn_next_indblkshift)); 688 bcopy(&odn->dn_next_bonustype[0], &ndn->dn_next_bonustype[0], 689 sizeof (odn->dn_next_bonustype)); 690 bcopy(&odn->dn_rm_spillblk[0], &ndn->dn_rm_spillblk[0], 691 sizeof (odn->dn_rm_spillblk)); 692 bcopy(&odn->dn_next_bonuslen[0], &ndn->dn_next_bonuslen[0], 693 sizeof (odn->dn_next_bonuslen)); 694 bcopy(&odn->dn_next_blksz[0], &ndn->dn_next_blksz[0], 695 sizeof (odn->dn_next_blksz)); 696 for (i = 0; i < TXG_SIZE; i++) { 697 list_move_tail(&ndn->dn_dirty_records[i], 698 &odn->dn_dirty_records[i]); 699 } 700 bcopy(&odn->dn_ranges[0], &ndn->dn_ranges[0], sizeof (odn->dn_ranges)); 701 ndn->dn_allocated_txg = odn->dn_allocated_txg; 702 ndn->dn_free_txg = odn->dn_free_txg; 703 ndn->dn_assigned_txg = odn->dn_assigned_txg; 704 ndn->dn_dirtyctx = odn->dn_dirtyctx; 705 ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset; 706 ASSERT(refcount_count(&odn->dn_tx_holds) == 0); 707 refcount_transfer(&ndn->dn_holds, &odn->dn_holds); 708 ASSERT(list_is_empty(&ndn->dn_dbufs)); 709 list_move_tail(&ndn->dn_dbufs, &odn->dn_dbufs); 710 ndn->dn_dbufs_count = odn->dn_dbufs_count; 711 ndn->dn_unlisted_l0_blkid = odn->dn_unlisted_l0_blkid; 712 ndn->dn_bonus = odn->dn_bonus; 713 ndn->dn_have_spill = odn->dn_have_spill; 714 ndn->dn_zio = odn->dn_zio; 715 ndn->dn_oldused = odn->dn_oldused; 716 ndn->dn_oldflags = odn->dn_oldflags; 717 ndn->dn_olduid = odn->dn_olduid; 718 ndn->dn_oldgid = odn->dn_oldgid; 719 ndn->dn_newuid = odn->dn_newuid; 720 ndn->dn_newgid = odn->dn_newgid; 721 ndn->dn_id_flags = odn->dn_id_flags; 722 dmu_zfetch_init(&ndn->dn_zfetch, NULL); 723 list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream); 724 ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode; 725 ndn->dn_zfetch.zf_stream_cnt = odn->dn_zfetch.zf_stream_cnt; 726 ndn->dn_zfetch.zf_alloc_fail = odn->dn_zfetch.zf_alloc_fail; 727 728 /* 729 * Update back pointers. Updating the handle fixes the back pointer of 730 * every descendant dbuf as well as the bonus dbuf. 731 */ 732 ASSERT(ndn->dn_handle->dnh_dnode == odn); 733 ndn->dn_handle->dnh_dnode = ndn; 734 if (ndn->dn_zfetch.zf_dnode == odn) { 735 ndn->dn_zfetch.zf_dnode = ndn; 736 } 737 738 /* 739 * Invalidate the original dnode by clearing all of its back pointers. 740 */ 741 odn->dn_dbuf = NULL; 742 odn->dn_handle = NULL; 743 list_create(&odn->dn_dbufs, sizeof (dmu_buf_impl_t), 744 offsetof(dmu_buf_impl_t, db_link)); 745 odn->dn_dbufs_count = 0; 746 odn->dn_unlisted_l0_blkid = 0; 747 odn->dn_bonus = NULL; 748 odn->dn_zfetch.zf_dnode = NULL; 749 750 /* 751 * Set the low bit of the objset pointer to ensure that dnode_move() 752 * recognizes the dnode as invalid in any subsequent callback. 753 */ 754 POINTER_INVALIDATE(&odn->dn_objset); 755 756 /* 757 * Satisfy the destructor. 758 */ 759 for (i = 0; i < TXG_SIZE; i++) { 760 list_create(&odn->dn_dirty_records[i], 761 sizeof (dbuf_dirty_record_t), 762 offsetof(dbuf_dirty_record_t, dr_dirty_node)); 763 odn->dn_ranges[i].avl_root = NULL; 764 odn->dn_ranges[i].avl_numnodes = 0; 765 odn->dn_next_nlevels[i] = 0; 766 odn->dn_next_indblkshift[i] = 0; 767 odn->dn_next_bonustype[i] = 0; 768 odn->dn_rm_spillblk[i] = 0; 769 odn->dn_next_bonuslen[i] = 0; 770 odn->dn_next_blksz[i] = 0; 771 } 772 odn->dn_allocated_txg = 0; 773 odn->dn_free_txg = 0; 774 odn->dn_assigned_txg = 0; 775 odn->dn_dirtyctx = 0; 776 odn->dn_dirtyctx_firstset = NULL; 777 odn->dn_have_spill = B_FALSE; 778 odn->dn_zio = NULL; 779 odn->dn_oldused = 0; 780 odn->dn_oldflags = 0; 781 odn->dn_olduid = 0; 782 odn->dn_oldgid = 0; 783 odn->dn_newuid = 0; 784 odn->dn_newgid = 0; 785 odn->dn_id_flags = 0; 786 787 /* 788 * Mark the dnode. 789 */ 790 ndn->dn_moved = 1; 791 odn->dn_moved = (uint8_t)-1; 792} 793 794#ifdef sun 795#ifdef _KERNEL 796/*ARGSUSED*/ 797static kmem_cbrc_t 798dnode_move(void *buf, void *newbuf, size_t size, void *arg) 799{ 800 dnode_t *odn = buf, *ndn = newbuf; 801 objset_t *os; 802 int64_t refcount; 803 uint32_t dbufs; 804 805 /* 806 * The dnode is on the objset's list of known dnodes if the objset 807 * pointer is valid. We set the low bit of the objset pointer when 808 * freeing the dnode to invalidate it, and the memory patterns written 809 * by kmem (baddcafe and deadbeef) set at least one of the two low bits. 810 * A newly created dnode sets the objset pointer last of all to indicate 811 * that the dnode is known and in a valid state to be moved by this 812 * function. 813 */ 814 os = odn->dn_objset; 815 if (!POINTER_IS_VALID(os)) { 816 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_invalid); 817 return (KMEM_CBRC_DONT_KNOW); 818 } 819 820 /* 821 * Ensure that the objset does not go away during the move. 822 */ 823 rw_enter(&os_lock, RW_WRITER); 824 if (os != odn->dn_objset) { 825 rw_exit(&os_lock); 826 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck1); 827 return (KMEM_CBRC_DONT_KNOW); 828 } 829 830 /* 831 * If the dnode is still valid, then so is the objset. We know that no 832 * valid objset can be freed while we hold os_lock, so we can safely 833 * ensure that the objset remains in use. 834 */ 835 mutex_enter(&os->os_lock); 836 837 /* 838 * Recheck the objset pointer in case the dnode was removed just before 839 * acquiring the lock. 840 */ 841 if (os != odn->dn_objset) { 842 mutex_exit(&os->os_lock); 843 rw_exit(&os_lock); 844 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck2); 845 return (KMEM_CBRC_DONT_KNOW); 846 } 847 848 /* 849 * At this point we know that as long as we hold os->os_lock, the dnode 850 * cannot be freed and fields within the dnode can be safely accessed. 851 * The objset listing this dnode cannot go away as long as this dnode is 852 * on its list. 853 */ 854 rw_exit(&os_lock); 855 if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) { 856 mutex_exit(&os->os_lock); 857 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_special); 858 return (KMEM_CBRC_NO); 859 } 860 ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */ 861 862 /* 863 * Lock the dnode handle to prevent the dnode from obtaining any new 864 * holds. This also prevents the descendant dbufs and the bonus dbuf 865 * from accessing the dnode, so that we can discount their holds. The 866 * handle is safe to access because we know that while the dnode cannot 867 * go away, neither can its handle. Once we hold dnh_zrlock, we can 868 * safely move any dnode referenced only by dbufs. 869 */ 870 if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) { 871 mutex_exit(&os->os_lock); 872 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_handle); 873 return (KMEM_CBRC_LATER); 874 } 875 876 /* 877 * Ensure a consistent view of the dnode's holds and the dnode's dbufs. 878 * We need to guarantee that there is a hold for every dbuf in order to 879 * determine whether the dnode is actively referenced. Falsely matching 880 * a dbuf to an active hold would lead to an unsafe move. It's possible 881 * that a thread already having an active dnode hold is about to add a 882 * dbuf, and we can't compare hold and dbuf counts while the add is in 883 * progress. 884 */ 885 if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) { 886 zrl_exit(&odn->dn_handle->dnh_zrlock); 887 mutex_exit(&os->os_lock); 888 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_rwlock); 889 return (KMEM_CBRC_LATER); 890 } 891 892 /* 893 * A dbuf may be removed (evicted) without an active dnode hold. In that 894 * case, the dbuf count is decremented under the handle lock before the 895 * dbuf's hold is released. This order ensures that if we count the hold 896 * after the dbuf is removed but before its hold is released, we will 897 * treat the unmatched hold as active and exit safely. If we count the 898 * hold before the dbuf is removed, the hold is discounted, and the 899 * removal is blocked until the move completes. 900 */ 901 refcount = refcount_count(&odn->dn_holds); 902 ASSERT(refcount >= 0); 903 dbufs = odn->dn_dbufs_count; 904 905 /* We can't have more dbufs than dnode holds. */ 906 ASSERT3U(dbufs, <=, refcount); 907 DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount, 908 uint32_t, dbufs); 909 910 if (refcount > dbufs) { 911 rw_exit(&odn->dn_struct_rwlock); 912 zrl_exit(&odn->dn_handle->dnh_zrlock); 913 mutex_exit(&os->os_lock); 914 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_active); 915 return (KMEM_CBRC_LATER); 916 } 917 918 rw_exit(&odn->dn_struct_rwlock); 919 920 /* 921 * At this point we know that anyone with a hold on the dnode is not 922 * actively referencing it. The dnode is known and in a valid state to 923 * move. We're holding the locks needed to execute the critical section. 924 */ 925 dnode_move_impl(odn, ndn); 926 927 list_link_replace(&odn->dn_link, &ndn->dn_link); 928 /* If the dnode was safe to move, the refcount cannot have changed. */ 929 ASSERT(refcount == refcount_count(&ndn->dn_holds)); 930 ASSERT(dbufs == ndn->dn_dbufs_count); 931 zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */ 932 mutex_exit(&os->os_lock); 933 934 return (KMEM_CBRC_YES); 935} 936#endif /* _KERNEL */ 937#endif /* sun */ 938 939void 940dnode_special_close(dnode_handle_t *dnh) 941{ 942 dnode_t *dn = dnh->dnh_dnode; 943 944 /* 945 * Wait for final references to the dnode to clear. This can 946 * only happen if the arc is asyncronously evicting state that 947 * has a hold on this dnode while we are trying to evict this 948 * dnode. 949 */ 950 while (refcount_count(&dn->dn_holds) > 0) 951 delay(1); 952 zrl_add(&dnh->dnh_zrlock); 953 dnode_destroy(dn); /* implicit zrl_remove() */ 954 zrl_destroy(&dnh->dnh_zrlock); 955 dnh->dnh_dnode = NULL; 956} 957 958dnode_t * 959dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object, 960 dnode_handle_t *dnh) 961{ 962 dnode_t *dn = dnode_create(os, dnp, NULL, object, dnh); 963 dnh->dnh_dnode = dn; 964 zrl_init(&dnh->dnh_zrlock); 965 DNODE_VERIFY(dn); 966 return (dn); 967} 968 969static void 970dnode_buf_pageout(dmu_buf_t *db, void *arg) 971{ 972 dnode_children_t *children_dnodes = arg; 973 int i; 974 int epb = db->db_size >> DNODE_SHIFT; 975 976 ASSERT(epb == children_dnodes->dnc_count); 977 978 for (i = 0; i < epb; i++) { 979 dnode_handle_t *dnh = &children_dnodes->dnc_children[i]; 980 dnode_t *dn; 981 982 /* 983 * The dnode handle lock guards against the dnode moving to 984 * another valid address, so there is no need here to guard 985 * against changes to or from NULL. 986 */ 987 if (dnh->dnh_dnode == NULL) { 988 zrl_destroy(&dnh->dnh_zrlock); 989 continue; 990 } 991 992 zrl_add(&dnh->dnh_zrlock); 993 dn = dnh->dnh_dnode; 994 /* 995 * If there are holds on this dnode, then there should 996 * be holds on the dnode's containing dbuf as well; thus 997 * it wouldn't be eligible for eviction and this function 998 * would not have been called. 999 */ 1000 ASSERT(refcount_is_zero(&dn->dn_holds)); 1001 ASSERT(refcount_is_zero(&dn->dn_tx_holds)); 1002 1003 dnode_destroy(dn); /* implicit zrl_remove() */ 1004 zrl_destroy(&dnh->dnh_zrlock); 1005 dnh->dnh_dnode = NULL; 1006 } 1007 kmem_free(children_dnodes, sizeof (dnode_children_t) + 1008 (epb - 1) * sizeof (dnode_handle_t)); 1009} 1010 1011/* 1012 * errors: 1013 * EINVAL - invalid object number. 1014 * EIO - i/o error. 1015 * succeeds even for free dnodes. 1016 */ 1017int 1018dnode_hold_impl(objset_t *os, uint64_t object, int flag, 1019 void *tag, dnode_t **dnp) 1020{ 1021 int epb, idx, err; 1022 int drop_struct_lock = FALSE; 1023 int type; 1024 uint64_t blk; 1025 dnode_t *mdn, *dn; 1026 dmu_buf_impl_t *db; 1027 dnode_children_t *children_dnodes; 1028 dnode_handle_t *dnh; 1029 1030 /* 1031 * If you are holding the spa config lock as writer, you shouldn't 1032 * be asking the DMU to do *anything* unless it's the root pool 1033 * which may require us to read from the root filesystem while 1034 * holding some (not all) of the locks as writer. 1035 */ 1036 ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 || 1037 (spa_is_root(os->os_spa) && 1038 spa_config_held(os->os_spa, SCL_STATE, RW_WRITER))); 1039 1040 if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT) { 1041 dn = (object == DMU_USERUSED_OBJECT) ? 1042 DMU_USERUSED_DNODE(os) : DMU_GROUPUSED_DNODE(os); 1043 if (dn == NULL) 1044 return (SET_ERROR(ENOENT)); 1045 type = dn->dn_type; 1046 if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE) 1047 return (SET_ERROR(ENOENT)); 1048 if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE) 1049 return (SET_ERROR(EEXIST)); 1050 DNODE_VERIFY(dn); 1051 (void) refcount_add(&dn->dn_holds, tag); 1052 *dnp = dn; 1053 return (0); 1054 } 1055 1056 if (object == 0 || object >= DN_MAX_OBJECT) 1057 return (SET_ERROR(EINVAL)); 1058 1059 mdn = DMU_META_DNODE(os); 1060 ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT); 1061 1062 DNODE_VERIFY(mdn); 1063 1064 if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) { 1065 rw_enter(&mdn->dn_struct_rwlock, RW_READER); 1066 drop_struct_lock = TRUE; 1067 } 1068 1069 blk = dbuf_whichblock(mdn, object * sizeof (dnode_phys_t)); 1070 1071 db = dbuf_hold(mdn, blk, FTAG); 1072 if (drop_struct_lock) 1073 rw_exit(&mdn->dn_struct_rwlock); 1074 if (db == NULL) 1075 return (SET_ERROR(EIO)); 1076 err = dbuf_read(db, NULL, DB_RF_CANFAIL); 1077 if (err) { 1078 dbuf_rele(db, FTAG); 1079 return (err); 1080 } 1081 1082 ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT); 1083 epb = db->db.db_size >> DNODE_SHIFT; 1084 1085 idx = object & (epb-1); 1086 1087 ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE); 1088 children_dnodes = dmu_buf_get_user(&db->db); 1089 if (children_dnodes == NULL) { 1090 int i; 1091 dnode_children_t *winner; 1092 children_dnodes = kmem_zalloc(sizeof (dnode_children_t) + 1093 (epb - 1) * sizeof (dnode_handle_t), KM_SLEEP); 1094 children_dnodes->dnc_count = epb; 1095 dnh = &children_dnodes->dnc_children[0]; 1096 for (i = 0; i < epb; i++) { 1097 zrl_init(&dnh[i].dnh_zrlock); 1098 dnh[i].dnh_dnode = NULL; 1099 } 1100 if (winner = dmu_buf_set_user(&db->db, children_dnodes, NULL, 1101 dnode_buf_pageout)) { 1102 kmem_free(children_dnodes, sizeof (dnode_children_t) + 1103 (epb - 1) * sizeof (dnode_handle_t)); 1104 children_dnodes = winner; 1105 } 1106 } 1107 ASSERT(children_dnodes->dnc_count == epb); 1108 1109 dnh = &children_dnodes->dnc_children[idx]; 1110 zrl_add(&dnh->dnh_zrlock); 1111 if ((dn = dnh->dnh_dnode) == NULL) { 1112 dnode_phys_t *phys = (dnode_phys_t *)db->db.db_data+idx; 1113 dnode_t *winner; 1114 1115 dn = dnode_create(os, phys, db, object, dnh); 1116 winner = atomic_cas_ptr(&dnh->dnh_dnode, NULL, dn); 1117 if (winner != NULL) { 1118 zrl_add(&dnh->dnh_zrlock); 1119 dnode_destroy(dn); /* implicit zrl_remove() */ 1120 dn = winner; 1121 } 1122 } 1123 1124 mutex_enter(&dn->dn_mtx); 1125 type = dn->dn_type; 1126 if (dn->dn_free_txg || 1127 ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE) || 1128 ((flag & DNODE_MUST_BE_FREE) && 1129 (type != DMU_OT_NONE || !refcount_is_zero(&dn->dn_holds)))) { 1130 mutex_exit(&dn->dn_mtx); 1131 zrl_remove(&dnh->dnh_zrlock); 1132 dbuf_rele(db, FTAG); 1133 return (type == DMU_OT_NONE ? ENOENT : EEXIST); 1134 } 1135 mutex_exit(&dn->dn_mtx); 1136 1137 if (refcount_add(&dn->dn_holds, tag) == 1) 1138 dbuf_add_ref(db, dnh); 1139 /* Now we can rely on the hold to prevent the dnode from moving. */ 1140 zrl_remove(&dnh->dnh_zrlock); 1141 1142 DNODE_VERIFY(dn); 1143 ASSERT3P(dn->dn_dbuf, ==, db); 1144 ASSERT3U(dn->dn_object, ==, object); 1145 dbuf_rele(db, FTAG); 1146 1147 *dnp = dn; 1148 return (0); 1149} 1150 1151/* 1152 * Return held dnode if the object is allocated, NULL if not. 1153 */ 1154int 1155dnode_hold(objset_t *os, uint64_t object, void *tag, dnode_t **dnp) 1156{ 1157 return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, tag, dnp)); 1158} 1159 1160/* 1161 * Can only add a reference if there is already at least one 1162 * reference on the dnode. Returns FALSE if unable to add a 1163 * new reference. 1164 */ 1165boolean_t 1166dnode_add_ref(dnode_t *dn, void *tag) 1167{ 1168 mutex_enter(&dn->dn_mtx); 1169 if (refcount_is_zero(&dn->dn_holds)) { 1170 mutex_exit(&dn->dn_mtx); 1171 return (FALSE); 1172 } 1173 VERIFY(1 < refcount_add(&dn->dn_holds, tag)); 1174 mutex_exit(&dn->dn_mtx); 1175 return (TRUE); 1176} 1177 1178void 1179dnode_rele(dnode_t *dn, void *tag) 1180{ 1181 uint64_t refs; 1182 /* Get while the hold prevents the dnode from moving. */ 1183 dmu_buf_impl_t *db = dn->dn_dbuf; 1184 dnode_handle_t *dnh = dn->dn_handle; 1185 1186 mutex_enter(&dn->dn_mtx); 1187 refs = refcount_remove(&dn->dn_holds, tag); 1188 mutex_exit(&dn->dn_mtx); 1189 1190 /* 1191 * It's unsafe to release the last hold on a dnode by dnode_rele() or 1192 * indirectly by dbuf_rele() while relying on the dnode handle to 1193 * prevent the dnode from moving, since releasing the last hold could 1194 * result in the dnode's parent dbuf evicting its dnode handles. For 1195 * that reason anyone calling dnode_rele() or dbuf_rele() without some 1196 * other direct or indirect hold on the dnode must first drop the dnode 1197 * handle. 1198 */ 1199 ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread); 1200 1201 /* NOTE: the DNODE_DNODE does not have a dn_dbuf */ 1202 if (refs == 0 && db != NULL) { 1203 /* 1204 * Another thread could add a hold to the dnode handle in 1205 * dnode_hold_impl() while holding the parent dbuf. Since the 1206 * hold on the parent dbuf prevents the handle from being 1207 * destroyed, the hold on the handle is OK. We can't yet assert 1208 * that the handle has zero references, but that will be 1209 * asserted anyway when the handle gets destroyed. 1210 */ 1211 dbuf_rele(db, dnh); 1212 } 1213} 1214 1215void 1216dnode_setdirty(dnode_t *dn, dmu_tx_t *tx) 1217{ 1218 objset_t *os = dn->dn_objset; 1219 uint64_t txg = tx->tx_txg; 1220 1221 if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) { 1222 dsl_dataset_dirty(os->os_dsl_dataset, tx); 1223 return; 1224 } 1225 1226 DNODE_VERIFY(dn); 1227 1228#ifdef ZFS_DEBUG 1229 mutex_enter(&dn->dn_mtx); 1230 ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg); 1231 ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg); 1232 mutex_exit(&dn->dn_mtx); 1233#endif 1234 1235 /* 1236 * Determine old uid/gid when necessary 1237 */ 1238 dmu_objset_userquota_get_ids(dn, B_TRUE, tx); 1239 1240 mutex_enter(&os->os_lock); 1241 1242 /* 1243 * If we are already marked dirty, we're done. 1244 */ 1245 if (list_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) { 1246 mutex_exit(&os->os_lock); 1247 return; 1248 } 1249 1250 ASSERT(!refcount_is_zero(&dn->dn_holds) || list_head(&dn->dn_dbufs)); 1251 ASSERT(dn->dn_datablksz != 0); 1252 ASSERT0(dn->dn_next_bonuslen[txg&TXG_MASK]); 1253 ASSERT0(dn->dn_next_blksz[txg&TXG_MASK]); 1254 ASSERT0(dn->dn_next_bonustype[txg&TXG_MASK]); 1255 1256 dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n", 1257 dn->dn_object, txg); 1258 1259 if (dn->dn_free_txg > 0 && dn->dn_free_txg <= txg) { 1260 list_insert_tail(&os->os_free_dnodes[txg&TXG_MASK], dn); 1261 } else { 1262 list_insert_tail(&os->os_dirty_dnodes[txg&TXG_MASK], dn); 1263 } 1264 1265 mutex_exit(&os->os_lock); 1266 1267 /* 1268 * The dnode maintains a hold on its containing dbuf as 1269 * long as there are holds on it. Each instantiated child 1270 * dbuf maintains a hold on the dnode. When the last child 1271 * drops its hold, the dnode will drop its hold on the 1272 * containing dbuf. We add a "dirty hold" here so that the 1273 * dnode will hang around after we finish processing its 1274 * children. 1275 */ 1276 VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg)); 1277 1278 (void) dbuf_dirty(dn->dn_dbuf, tx); 1279 1280 dsl_dataset_dirty(os->os_dsl_dataset, tx); 1281} 1282 1283void 1284dnode_free(dnode_t *dn, dmu_tx_t *tx) 1285{ 1286 int txgoff = tx->tx_txg & TXG_MASK; 1287 1288 dprintf("dn=%p txg=%llu\n", dn, tx->tx_txg); 1289 1290 /* we should be the only holder... hopefully */ 1291 /* ASSERT3U(refcount_count(&dn->dn_holds), ==, 1); */ 1292 1293 mutex_enter(&dn->dn_mtx); 1294 if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) { 1295 mutex_exit(&dn->dn_mtx); 1296 return; 1297 } 1298 dn->dn_free_txg = tx->tx_txg; 1299 mutex_exit(&dn->dn_mtx); 1300 1301 /* 1302 * If the dnode is already dirty, it needs to be moved from 1303 * the dirty list to the free list. 1304 */ 1305 mutex_enter(&dn->dn_objset->os_lock); 1306 if (list_link_active(&dn->dn_dirty_link[txgoff])) { 1307 list_remove(&dn->dn_objset->os_dirty_dnodes[txgoff], dn); 1308 list_insert_tail(&dn->dn_objset->os_free_dnodes[txgoff], dn); 1309 mutex_exit(&dn->dn_objset->os_lock); 1310 } else { 1311 mutex_exit(&dn->dn_objset->os_lock); 1312 dnode_setdirty(dn, tx); 1313 } 1314} 1315 1316/* 1317 * Try to change the block size for the indicated dnode. This can only 1318 * succeed if there are no blocks allocated or dirty beyond first block 1319 */ 1320int 1321dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx) 1322{ 1323 dmu_buf_impl_t *db, *db_next; 1324 int err; 1325 1326 if (size == 0) 1327 size = SPA_MINBLOCKSIZE; 1328 if (size > SPA_MAXBLOCKSIZE) 1329 size = SPA_MAXBLOCKSIZE; 1330 else 1331 size = P2ROUNDUP(size, SPA_MINBLOCKSIZE); 1332 1333 if (ibs == dn->dn_indblkshift) 1334 ibs = 0; 1335 1336 if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0) 1337 return (0); 1338 1339 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 1340 1341 /* Check for any allocated blocks beyond the first */ 1342 if (dn->dn_phys->dn_maxblkid != 0) 1343 goto fail; 1344 1345 mutex_enter(&dn->dn_dbufs_mtx); 1346 for (db = list_head(&dn->dn_dbufs); db; db = db_next) { 1347 db_next = list_next(&dn->dn_dbufs, db); 1348 1349 if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID && 1350 db->db_blkid != DMU_SPILL_BLKID) { 1351 mutex_exit(&dn->dn_dbufs_mtx); 1352 goto fail; 1353 } 1354 } 1355 mutex_exit(&dn->dn_dbufs_mtx); 1356 1357 if (ibs && dn->dn_nlevels != 1) 1358 goto fail; 1359 1360 /* resize the old block */ 1361 err = dbuf_hold_impl(dn, 0, 0, TRUE, FTAG, &db); 1362 if (err == 0) 1363 dbuf_new_size(db, size, tx); 1364 else if (err != ENOENT) 1365 goto fail; 1366 1367 dnode_setdblksz(dn, size); 1368 dnode_setdirty(dn, tx); 1369 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size; 1370 if (ibs) { 1371 dn->dn_indblkshift = ibs; 1372 dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs; 1373 } 1374 /* rele after we have fixed the blocksize in the dnode */ 1375 if (db) 1376 dbuf_rele(db, FTAG); 1377 1378 rw_exit(&dn->dn_struct_rwlock); 1379 return (0); 1380 1381fail: 1382 rw_exit(&dn->dn_struct_rwlock); 1383 return (SET_ERROR(ENOTSUP)); 1384} 1385 1386/* read-holding callers must not rely on the lock being continuously held */ 1387void 1388dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read) 1389{ 1390 uint64_t txgoff = tx->tx_txg & TXG_MASK; 1391 int epbs, new_nlevels; 1392 uint64_t sz; 1393 1394 ASSERT(blkid != DMU_BONUS_BLKID); 1395 1396 ASSERT(have_read ? 1397 RW_READ_HELD(&dn->dn_struct_rwlock) : 1398 RW_WRITE_HELD(&dn->dn_struct_rwlock)); 1399 1400 /* 1401 * if we have a read-lock, check to see if we need to do any work 1402 * before upgrading to a write-lock. 1403 */ 1404 if (have_read) { 1405 if (blkid <= dn->dn_maxblkid) 1406 return; 1407 1408 if (!rw_tryupgrade(&dn->dn_struct_rwlock)) { 1409 rw_exit(&dn->dn_struct_rwlock); 1410 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 1411 } 1412 } 1413 1414 if (blkid <= dn->dn_maxblkid) 1415 goto out; 1416 1417 dn->dn_maxblkid = blkid; 1418 1419 /* 1420 * Compute the number of levels necessary to support the new maxblkid. 1421 */ 1422 new_nlevels = 1; 1423 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1424 for (sz = dn->dn_nblkptr; 1425 sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs) 1426 new_nlevels++; 1427 1428 if (new_nlevels > dn->dn_nlevels) { 1429 int old_nlevels = dn->dn_nlevels; 1430 dmu_buf_impl_t *db; 1431 list_t *list; 1432 dbuf_dirty_record_t *new, *dr, *dr_next; 1433 1434 dn->dn_nlevels = new_nlevels; 1435 1436 ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]); 1437 dn->dn_next_nlevels[txgoff] = new_nlevels; 1438 1439 /* dirty the left indirects */ 1440 db = dbuf_hold_level(dn, old_nlevels, 0, FTAG); 1441 ASSERT(db != NULL); 1442 new = dbuf_dirty(db, tx); 1443 dbuf_rele(db, FTAG); 1444 1445 /* transfer the dirty records to the new indirect */ 1446 mutex_enter(&dn->dn_mtx); 1447 mutex_enter(&new->dt.di.dr_mtx); 1448 list = &dn->dn_dirty_records[txgoff]; 1449 for (dr = list_head(list); dr; dr = dr_next) { 1450 dr_next = list_next(&dn->dn_dirty_records[txgoff], dr); 1451 if (dr->dr_dbuf->db_level != new_nlevels-1 && 1452 dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID && 1453 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) { 1454 ASSERT(dr->dr_dbuf->db_level == old_nlevels-1); 1455 list_remove(&dn->dn_dirty_records[txgoff], dr); 1456 list_insert_tail(&new->dt.di.dr_children, dr); 1457 dr->dr_parent = new; 1458 } 1459 } 1460 mutex_exit(&new->dt.di.dr_mtx); 1461 mutex_exit(&dn->dn_mtx); 1462 } 1463 1464out: 1465 if (have_read) 1466 rw_downgrade(&dn->dn_struct_rwlock); 1467} 1468 1469void 1470dnode_clear_range(dnode_t *dn, uint64_t blkid, uint64_t nblks, dmu_tx_t *tx) 1471{ 1472 avl_tree_t *tree = &dn->dn_ranges[tx->tx_txg&TXG_MASK]; 1473 avl_index_t where; 1474 free_range_t *rp; 1475 free_range_t rp_tofind; 1476 uint64_t endblk = blkid + nblks; 1477 1478 ASSERT(MUTEX_HELD(&dn->dn_mtx)); 1479 ASSERT(nblks <= UINT64_MAX - blkid); /* no overflow */ 1480 1481 dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n", 1482 blkid, nblks, tx->tx_txg); 1483 rp_tofind.fr_blkid = blkid; 1484 rp = avl_find(tree, &rp_tofind, &where); 1485 if (rp == NULL) 1486 rp = avl_nearest(tree, where, AVL_BEFORE); 1487 if (rp == NULL) 1488 rp = avl_nearest(tree, where, AVL_AFTER); 1489 1490 while (rp && (rp->fr_blkid <= blkid + nblks)) { 1491 uint64_t fr_endblk = rp->fr_blkid + rp->fr_nblks; 1492 free_range_t *nrp = AVL_NEXT(tree, rp); 1493 1494 if (blkid <= rp->fr_blkid && endblk >= fr_endblk) { 1495 /* clear this entire range */ 1496 avl_remove(tree, rp); 1497 kmem_free(rp, sizeof (free_range_t)); 1498 } else if (blkid <= rp->fr_blkid && 1499 endblk > rp->fr_blkid && endblk < fr_endblk) { 1500 /* clear the beginning of this range */ 1501 rp->fr_blkid = endblk; 1502 rp->fr_nblks = fr_endblk - endblk; 1503 } else if (blkid > rp->fr_blkid && blkid < fr_endblk && 1504 endblk >= fr_endblk) { 1505 /* clear the end of this range */ 1506 rp->fr_nblks = blkid - rp->fr_blkid; 1507 } else if (blkid > rp->fr_blkid && endblk < fr_endblk) { 1508 /* clear a chunk out of this range */ 1509 free_range_t *new_rp = 1510 kmem_alloc(sizeof (free_range_t), KM_SLEEP); 1511 1512 new_rp->fr_blkid = endblk; 1513 new_rp->fr_nblks = fr_endblk - endblk; 1514 avl_insert_here(tree, new_rp, rp, AVL_AFTER); 1515 rp->fr_nblks = blkid - rp->fr_blkid; 1516 } 1517 /* there may be no overlap */ 1518 rp = nrp; 1519 } 1520} 1521 1522void 1523dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx) 1524{ 1525 dmu_buf_impl_t *db; 1526 uint64_t blkoff, blkid, nblks; 1527 int blksz, blkshift, head, tail; 1528 int trunc = FALSE; 1529 int epbs; 1530 1531 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 1532 blksz = dn->dn_datablksz; 1533 blkshift = dn->dn_datablkshift; 1534 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 1535 1536 if (len == DMU_OBJECT_END) { 1537 len = UINT64_MAX - off; 1538 trunc = TRUE; 1539 } 1540 1541 /* 1542 * First, block align the region to free: 1543 */ 1544 if (ISP2(blksz)) { 1545 head = P2NPHASE(off, blksz); 1546 blkoff = P2PHASE(off, blksz); 1547 if ((off >> blkshift) > dn->dn_maxblkid) 1548 goto out; 1549 } else { 1550 ASSERT(dn->dn_maxblkid == 0); 1551 if (off == 0 && len >= blksz) { 1552 /* Freeing the whole block; fast-track this request */ 1553 blkid = 0; 1554 nblks = 1; 1555 goto done; 1556 } else if (off >= blksz) { 1557 /* Freeing past end-of-data */ 1558 goto out; 1559 } else { 1560 /* Freeing part of the block. */ 1561 head = blksz - off; 1562 ASSERT3U(head, >, 0); 1563 } 1564 blkoff = off; 1565 } 1566 /* zero out any partial block data at the start of the range */ 1567 if (head) { 1568 ASSERT3U(blkoff + head, ==, blksz); 1569 if (len < head) 1570 head = len; 1571 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, off), TRUE, 1572 FTAG, &db) == 0) { 1573 caddr_t data; 1574 1575 /* don't dirty if it isn't on disk and isn't dirty */ 1576 if (db->db_last_dirty || 1577 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) { 1578 rw_exit(&dn->dn_struct_rwlock); 1579 dbuf_will_dirty(db, tx); 1580 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 1581 data = db->db.db_data; 1582 bzero(data + blkoff, head); 1583 } 1584 dbuf_rele(db, FTAG); 1585 } 1586 off += head; 1587 len -= head; 1588 } 1589 1590 /* If the range was less than one block, we're done */ 1591 if (len == 0) 1592 goto out; 1593 1594 /* If the remaining range is past end of file, we're done */ 1595 if ((off >> blkshift) > dn->dn_maxblkid) 1596 goto out; 1597 1598 ASSERT(ISP2(blksz)); 1599 if (trunc) 1600 tail = 0; 1601 else 1602 tail = P2PHASE(len, blksz); 1603 1604 ASSERT0(P2PHASE(off, blksz)); 1605 /* zero out any partial block data at the end of the range */ 1606 if (tail) { 1607 if (len < tail) 1608 tail = len; 1609 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, off+len), 1610 TRUE, FTAG, &db) == 0) { 1611 /* don't dirty if not on disk and not dirty */ 1612 if (db->db_last_dirty || 1613 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) { 1614 rw_exit(&dn->dn_struct_rwlock); 1615 dbuf_will_dirty(db, tx); 1616 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 1617 bzero(db->db.db_data, tail); 1618 } 1619 dbuf_rele(db, FTAG); 1620 } 1621 len -= tail; 1622 } 1623 1624 /* If the range did not include a full block, we are done */ 1625 if (len == 0) 1626 goto out; 1627 1628 ASSERT(IS_P2ALIGNED(off, blksz)); 1629 ASSERT(trunc || IS_P2ALIGNED(len, blksz)); 1630 blkid = off >> blkshift; 1631 nblks = len >> blkshift; 1632 if (trunc) 1633 nblks += 1; 1634 1635 /* 1636 * Read in and mark all the level-1 indirects dirty, 1637 * so that they will stay in memory until syncing phase. 1638 * Always dirty the first and last indirect to make sure 1639 * we dirty all the partial indirects. 1640 */ 1641 if (dn->dn_nlevels > 1) { 1642 uint64_t i, first, last; 1643 int shift = epbs + dn->dn_datablkshift; 1644 1645 first = blkid >> epbs; 1646 if (db = dbuf_hold_level(dn, 1, first, FTAG)) { 1647 dbuf_will_dirty(db, tx); 1648 dbuf_rele(db, FTAG); 1649 } 1650 if (trunc) 1651 last = dn->dn_maxblkid >> epbs; 1652 else 1653 last = (blkid + nblks - 1) >> epbs; 1654 if (last > first && (db = dbuf_hold_level(dn, 1, last, FTAG))) { 1655 dbuf_will_dirty(db, tx); 1656 dbuf_rele(db, FTAG); 1657 } 1658 for (i = first + 1; i < last; i++) { 1659 uint64_t ibyte = i << shift; 1660 int err; 1661 1662 err = dnode_next_offset(dn, 1663 DNODE_FIND_HAVELOCK, &ibyte, 1, 1, 0); 1664 i = ibyte >> shift; 1665 if (err == ESRCH || i >= last) 1666 break; 1667 ASSERT(err == 0); 1668 db = dbuf_hold_level(dn, 1, i, FTAG); 1669 if (db) { 1670 dbuf_will_dirty(db, tx); 1671 dbuf_rele(db, FTAG); 1672 } 1673 } 1674 } 1675done: 1676 /* 1677 * Add this range to the dnode range list. 1678 * We will finish up this free operation in the syncing phase. 1679 */ 1680 mutex_enter(&dn->dn_mtx); 1681 dnode_clear_range(dn, blkid, nblks, tx); 1682 { 1683 free_range_t *rp, *found; 1684 avl_index_t where; 1685 avl_tree_t *tree = &dn->dn_ranges[tx->tx_txg&TXG_MASK]; 1686 1687 /* Add new range to dn_ranges */ 1688 rp = kmem_alloc(sizeof (free_range_t), KM_SLEEP); 1689 rp->fr_blkid = blkid; 1690 rp->fr_nblks = nblks; 1691 found = avl_find(tree, rp, &where); 1692 ASSERT(found == NULL); 1693 avl_insert(tree, rp, where); 1694 dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n", 1695 blkid, nblks, tx->tx_txg); 1696 } 1697 mutex_exit(&dn->dn_mtx); 1698 1699 dbuf_free_range(dn, blkid, blkid + nblks - 1, tx); 1700 dnode_setdirty(dn, tx); 1701out: 1702 if (trunc && dn->dn_maxblkid >= (off >> blkshift)) 1703 dn->dn_maxblkid = (off >> blkshift ? (off >> blkshift) - 1 : 0); 1704 1705 rw_exit(&dn->dn_struct_rwlock); 1706} 1707 1708static boolean_t 1709dnode_spill_freed(dnode_t *dn) 1710{ 1711 int i; 1712 1713 mutex_enter(&dn->dn_mtx); 1714 for (i = 0; i < TXG_SIZE; i++) { 1715 if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK) 1716 break; 1717 } 1718 mutex_exit(&dn->dn_mtx); 1719 return (i < TXG_SIZE); 1720} 1721 1722/* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */ 1723uint64_t 1724dnode_block_freed(dnode_t *dn, uint64_t blkid) 1725{ 1726 free_range_t range_tofind; 1727 void *dp = spa_get_dsl(dn->dn_objset->os_spa); 1728 int i; 1729 1730 if (blkid == DMU_BONUS_BLKID) 1731 return (FALSE); 1732 1733 /* 1734 * If we're in the process of opening the pool, dp will not be 1735 * set yet, but there shouldn't be anything dirty. 1736 */ 1737 if (dp == NULL) 1738 return (FALSE); 1739 1740 if (dn->dn_free_txg) 1741 return (TRUE); 1742 1743 if (blkid == DMU_SPILL_BLKID) 1744 return (dnode_spill_freed(dn)); 1745 1746 range_tofind.fr_blkid = blkid; 1747 mutex_enter(&dn->dn_mtx); 1748 for (i = 0; i < TXG_SIZE; i++) { 1749 free_range_t *range_found; 1750 avl_index_t idx; 1751 1752 range_found = avl_find(&dn->dn_ranges[i], &range_tofind, &idx); 1753 if (range_found) { 1754 ASSERT(range_found->fr_nblks > 0); 1755 break; 1756 } 1757 range_found = avl_nearest(&dn->dn_ranges[i], idx, AVL_BEFORE); 1758 if (range_found && 1759 range_found->fr_blkid + range_found->fr_nblks > blkid) 1760 break; 1761 } 1762 mutex_exit(&dn->dn_mtx); 1763 return (i < TXG_SIZE); 1764} 1765 1766/* call from syncing context when we actually write/free space for this dnode */ 1767void 1768dnode_diduse_space(dnode_t *dn, int64_t delta) 1769{ 1770 uint64_t space; 1771 dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n", 1772 dn, dn->dn_phys, 1773 (u_longlong_t)dn->dn_phys->dn_used, 1774 (longlong_t)delta); 1775 1776 mutex_enter(&dn->dn_mtx); 1777 space = DN_USED_BYTES(dn->dn_phys); 1778 if (delta > 0) { 1779 ASSERT3U(space + delta, >=, space); /* no overflow */ 1780 } else { 1781 ASSERT3U(space, >=, -delta); /* no underflow */ 1782 } 1783 space += delta; 1784 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) { 1785 ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0); 1786 ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT)); 1787 dn->dn_phys->dn_used = space >> DEV_BSHIFT; 1788 } else { 1789 dn->dn_phys->dn_used = space; 1790 dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES; 1791 } 1792 mutex_exit(&dn->dn_mtx); 1793} 1794 1795/* 1796 * Call when we think we're going to write/free space in open context to track 1797 * the amount of memory in use by the currently open txg. 1798 */ 1799void 1800dnode_willuse_space(dnode_t *dn, int64_t space, dmu_tx_t *tx) 1801{ 1802 objset_t *os = dn->dn_objset; 1803 dsl_dataset_t *ds = os->os_dsl_dataset; 1804 int64_t aspace = spa_get_asize(os->os_spa, space); 1805 1806 if (ds != NULL) { 1807 dsl_dir_willuse_space(ds->ds_dir, aspace, tx); 1808 dsl_pool_dirty_space(dmu_tx_pool(tx), space, tx); 1809 } 1810 1811 dmu_tx_willuse_space(tx, aspace); 1812} 1813 1814/* 1815 * Scans a block at the indicated "level" looking for a hole or data, 1816 * depending on 'flags'. 1817 * 1818 * If level > 0, then we are scanning an indirect block looking at its 1819 * pointers. If level == 0, then we are looking at a block of dnodes. 1820 * 1821 * If we don't find what we are looking for in the block, we return ESRCH. 1822 * Otherwise, return with *offset pointing to the beginning (if searching 1823 * forwards) or end (if searching backwards) of the range covered by the 1824 * block pointer we matched on (or dnode). 1825 * 1826 * The basic search algorithm used below by dnode_next_offset() is to 1827 * use this function to search up the block tree (widen the search) until 1828 * we find something (i.e., we don't return ESRCH) and then search back 1829 * down the tree (narrow the search) until we reach our original search 1830 * level. 1831 */ 1832static int 1833dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset, 1834 int lvl, uint64_t blkfill, uint64_t txg) 1835{ 1836 dmu_buf_impl_t *db = NULL; 1837 void *data = NULL; 1838 uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 1839 uint64_t epb = 1ULL << epbs; 1840 uint64_t minfill, maxfill; 1841 boolean_t hole; 1842 int i, inc, error, span; 1843 1844 dprintf("probing object %llu offset %llx level %d of %u\n", 1845 dn->dn_object, *offset, lvl, dn->dn_phys->dn_nlevels); 1846 1847 hole = ((flags & DNODE_FIND_HOLE) != 0); 1848 inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1; 1849 ASSERT(txg == 0 || !hole); 1850 1851 if (lvl == dn->dn_phys->dn_nlevels) { 1852 error = 0; 1853 epb = dn->dn_phys->dn_nblkptr; 1854 data = dn->dn_phys->dn_blkptr; 1855 } else { 1856 uint64_t blkid = dbuf_whichblock(dn, *offset) >> (epbs * lvl); 1857 error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FTAG, &db); 1858 if (error) { 1859 if (error != ENOENT) 1860 return (error); 1861 if (hole) 1862 return (0); 1863 /* 1864 * This can only happen when we are searching up 1865 * the block tree for data. We don't really need to 1866 * adjust the offset, as we will just end up looking 1867 * at the pointer to this block in its parent, and its 1868 * going to be unallocated, so we will skip over it. 1869 */ 1870 return (SET_ERROR(ESRCH)); 1871 } 1872 error = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_HAVESTRUCT); 1873 if (error) { 1874 dbuf_rele(db, FTAG); 1875 return (error); 1876 } 1877 data = db->db.db_data; 1878 } 1879 1880 if (db && txg && 1881 (db->db_blkptr == NULL || db->db_blkptr->blk_birth <= txg)) { 1882 /* 1883 * This can only happen when we are searching up the tree 1884 * and these conditions mean that we need to keep climbing. 1885 */ 1886 error = SET_ERROR(ESRCH); 1887 } else if (lvl == 0) { 1888 dnode_phys_t *dnp = data; 1889 span = DNODE_SHIFT; 1890 ASSERT(dn->dn_type == DMU_OT_DNODE); 1891 1892 for (i = (*offset >> span) & (blkfill - 1); 1893 i >= 0 && i < blkfill; i += inc) { 1894 if ((dnp[i].dn_type == DMU_OT_NONE) == hole) 1895 break; 1896 *offset += (1ULL << span) * inc; 1897 } 1898 if (i < 0 || i == blkfill) 1899 error = SET_ERROR(ESRCH); 1900 } else { 1901 blkptr_t *bp = data; 1902 uint64_t start = *offset; 1903 span = (lvl - 1) * epbs + dn->dn_datablkshift; 1904 minfill = 0; 1905 maxfill = blkfill << ((lvl - 1) * epbs); 1906 1907 if (hole) 1908 maxfill--; 1909 else 1910 minfill++; 1911 1912 *offset = *offset >> span; 1913 for (i = BF64_GET(*offset, 0, epbs); 1914 i >= 0 && i < epb; i += inc) { 1915 if (bp[i].blk_fill >= minfill && 1916 bp[i].blk_fill <= maxfill && 1917 (hole || bp[i].blk_birth > txg)) 1918 break; 1919 if (inc > 0 || *offset > 0) 1920 *offset += inc; 1921 } 1922 *offset = *offset << span; 1923 if (inc < 0) { 1924 /* traversing backwards; position offset at the end */ 1925 ASSERT3U(*offset, <=, start); 1926 *offset = MIN(*offset + (1ULL << span) - 1, start); 1927 } else if (*offset < start) { 1928 *offset = start; 1929 } 1930 if (i < 0 || i >= epb) 1931 error = SET_ERROR(ESRCH); 1932 } 1933 1934 if (db) 1935 dbuf_rele(db, FTAG); 1936 1937 return (error); 1938} 1939 1940/* 1941 * Find the next hole, data, or sparse region at or after *offset. 1942 * The value 'blkfill' tells us how many items we expect to find 1943 * in an L0 data block; this value is 1 for normal objects, 1944 * DNODES_PER_BLOCK for the meta dnode, and some fraction of 1945 * DNODES_PER_BLOCK when searching for sparse regions thereof. 1946 * 1947 * Examples: 1948 * 1949 * dnode_next_offset(dn, flags, offset, 1, 1, 0); 1950 * Finds the next/previous hole/data in a file. 1951 * Used in dmu_offset_next(). 1952 * 1953 * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg); 1954 * Finds the next free/allocated dnode an objset's meta-dnode. 1955 * Only finds objects that have new contents since txg (ie. 1956 * bonus buffer changes and content removal are ignored). 1957 * Used in dmu_object_next(). 1958 * 1959 * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0); 1960 * Finds the next L2 meta-dnode bp that's at most 1/4 full. 1961 * Used in dmu_object_alloc(). 1962 */ 1963int 1964dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset, 1965 int minlvl, uint64_t blkfill, uint64_t txg) 1966{ 1967 uint64_t initial_offset = *offset; 1968 int lvl, maxlvl; 1969 int error = 0; 1970 1971 if (!(flags & DNODE_FIND_HAVELOCK)) 1972 rw_enter(&dn->dn_struct_rwlock, RW_READER); 1973 1974 if (dn->dn_phys->dn_nlevels == 0) { 1975 error = SET_ERROR(ESRCH); 1976 goto out; 1977 } 1978 1979 if (dn->dn_datablkshift == 0) { 1980 if (*offset < dn->dn_datablksz) { 1981 if (flags & DNODE_FIND_HOLE) 1982 *offset = dn->dn_datablksz; 1983 } else { 1984 error = SET_ERROR(ESRCH); 1985 } 1986 goto out; 1987 } 1988 1989 maxlvl = dn->dn_phys->dn_nlevels; 1990 1991 for (lvl = minlvl; lvl <= maxlvl; lvl++) { 1992 error = dnode_next_offset_level(dn, 1993 flags, offset, lvl, blkfill, txg); 1994 if (error != ESRCH) 1995 break; 1996 } 1997 1998 while (error == 0 && --lvl >= minlvl) { 1999 error = dnode_next_offset_level(dn, 2000 flags, offset, lvl, blkfill, txg); 2001 } 2002 2003 if (error == 0 && (flags & DNODE_FIND_BACKWARDS ? 2004 initial_offset < *offset : initial_offset > *offset)) 2005 error = SET_ERROR(ESRCH); 2006out: 2007 if (!(flags & DNODE_FIND_HAVELOCK)) 2008 rw_exit(&dn->dn_struct_rwlock); 2009 2010 return (error); 2011} 2012