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 https://opensource.org/licenses/CDDL-1.0. 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, 2020 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 #include <sys/trace_zfs.h> 41 #include <sys/zfs_project.h> 42 43 dnode_stats_t dnode_stats = { 44 { "dnode_hold_dbuf_hold", KSTAT_DATA_UINT64 }, 45 { "dnode_hold_dbuf_read", KSTAT_DATA_UINT64 }, 46 { "dnode_hold_alloc_hits", KSTAT_DATA_UINT64 }, 47 { "dnode_hold_alloc_misses", KSTAT_DATA_UINT64 }, 48 { "dnode_hold_alloc_interior", KSTAT_DATA_UINT64 }, 49 { "dnode_hold_alloc_lock_retry", KSTAT_DATA_UINT64 }, 50 { "dnode_hold_alloc_lock_misses", KSTAT_DATA_UINT64 }, 51 { "dnode_hold_alloc_type_none", KSTAT_DATA_UINT64 }, 52 { "dnode_hold_free_hits", KSTAT_DATA_UINT64 }, 53 { "dnode_hold_free_misses", KSTAT_DATA_UINT64 }, 54 { "dnode_hold_free_lock_misses", KSTAT_DATA_UINT64 }, 55 { "dnode_hold_free_lock_retry", KSTAT_DATA_UINT64 }, 56 { "dnode_hold_free_overflow", KSTAT_DATA_UINT64 }, 57 { "dnode_hold_free_refcount", KSTAT_DATA_UINT64 }, 58 { "dnode_free_interior_lock_retry", KSTAT_DATA_UINT64 }, 59 { "dnode_allocate", KSTAT_DATA_UINT64 }, 60 { "dnode_reallocate", KSTAT_DATA_UINT64 }, 61 { "dnode_buf_evict", KSTAT_DATA_UINT64 }, 62 { "dnode_alloc_next_chunk", KSTAT_DATA_UINT64 }, 63 { "dnode_alloc_race", KSTAT_DATA_UINT64 }, 64 { "dnode_alloc_next_block", KSTAT_DATA_UINT64 }, 65 { "dnode_move_invalid", KSTAT_DATA_UINT64 }, 66 { "dnode_move_recheck1", KSTAT_DATA_UINT64 }, 67 { "dnode_move_recheck2", KSTAT_DATA_UINT64 }, 68 { "dnode_move_special", KSTAT_DATA_UINT64 }, 69 { "dnode_move_handle", KSTAT_DATA_UINT64 }, 70 { "dnode_move_rwlock", KSTAT_DATA_UINT64 }, 71 { "dnode_move_active", KSTAT_DATA_UINT64 }, 72 }; 73 74 dnode_sums_t dnode_sums; 75 76 static kstat_t *dnode_ksp; 77 static kmem_cache_t *dnode_cache; 78 79 static dnode_phys_t dnode_phys_zero __maybe_unused; 80 81 int zfs_default_bs = SPA_MINBLOCKSHIFT; 82 int zfs_default_ibs = DN_MAX_INDBLKSHIFT; 83 84 #ifdef _KERNEL 85 static kmem_cbrc_t dnode_move(void *, void *, size_t, void *); 86 #endif /* _KERNEL */ 87 88 static int 89 dbuf_compare(const void *x1, const void *x2) 90 { 91 const dmu_buf_impl_t *d1 = x1; 92 const dmu_buf_impl_t *d2 = x2; 93 94 int cmp = TREE_CMP(d1->db_level, d2->db_level); 95 if (likely(cmp)) 96 return (cmp); 97 98 cmp = TREE_CMP(d1->db_blkid, d2->db_blkid); 99 if (likely(cmp)) 100 return (cmp); 101 102 if (d1->db_state == DB_SEARCH) { 103 ASSERT3S(d2->db_state, !=, DB_SEARCH); 104 return (-1); 105 } else if (d2->db_state == DB_SEARCH) { 106 ASSERT3S(d1->db_state, !=, DB_SEARCH); 107 return (1); 108 } 109 110 return (TREE_PCMP(d1, d2)); 111 } 112 113 static int 114 dnode_cons(void *arg, void *unused, int kmflag) 115 { 116 (void) unused, (void) kmflag; 117 dnode_t *dn = arg; 118 119 rw_init(&dn->dn_struct_rwlock, NULL, RW_NOLOCKDEP, NULL); 120 mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL); 121 mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL); 122 cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL); 123 cv_init(&dn->dn_nodnholds, NULL, CV_DEFAULT, NULL); 124 125 /* 126 * Every dbuf has a reference, and dropping a tracked reference is 127 * O(number of references), so don't track dn_holds. 128 */ 129 zfs_refcount_create_untracked(&dn->dn_holds); 130 zfs_refcount_create(&dn->dn_tx_holds); 131 list_link_init(&dn->dn_link); 132 133 memset(dn->dn_next_type, 0, sizeof (dn->dn_next_type)); 134 memset(dn->dn_next_nblkptr, 0, sizeof (dn->dn_next_nblkptr)); 135 memset(dn->dn_next_nlevels, 0, sizeof (dn->dn_next_nlevels)); 136 memset(dn->dn_next_indblkshift, 0, sizeof (dn->dn_next_indblkshift)); 137 memset(dn->dn_next_bonustype, 0, sizeof (dn->dn_next_bonustype)); 138 memset(dn->dn_rm_spillblk, 0, sizeof (dn->dn_rm_spillblk)); 139 memset(dn->dn_next_bonuslen, 0, sizeof (dn->dn_next_bonuslen)); 140 memset(dn->dn_next_blksz, 0, sizeof (dn->dn_next_blksz)); 141 memset(dn->dn_next_maxblkid, 0, sizeof (dn->dn_next_maxblkid)); 142 143 for (int i = 0; i < TXG_SIZE; i++) { 144 multilist_link_init(&dn->dn_dirty_link[i]); 145 dn->dn_free_ranges[i] = NULL; 146 list_create(&dn->dn_dirty_records[i], 147 sizeof (dbuf_dirty_record_t), 148 offsetof(dbuf_dirty_record_t, dr_dirty_node)); 149 } 150 151 dn->dn_allocated_txg = 0; 152 dn->dn_free_txg = 0; 153 dn->dn_assigned_txg = 0; 154 dn->dn_dirty_txg = 0; 155 dn->dn_dirtyctx = 0; 156 dn->dn_dirtyctx_firstset = NULL; 157 dn->dn_bonus = NULL; 158 dn->dn_have_spill = B_FALSE; 159 dn->dn_zio = NULL; 160 dn->dn_oldused = 0; 161 dn->dn_oldflags = 0; 162 dn->dn_olduid = 0; 163 dn->dn_oldgid = 0; 164 dn->dn_oldprojid = ZFS_DEFAULT_PROJID; 165 dn->dn_newuid = 0; 166 dn->dn_newgid = 0; 167 dn->dn_newprojid = ZFS_DEFAULT_PROJID; 168 dn->dn_id_flags = 0; 169 170 dn->dn_dbufs_count = 0; 171 avl_create(&dn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t), 172 offsetof(dmu_buf_impl_t, db_link)); 173 174 dn->dn_moved = 0; 175 return (0); 176 } 177 178 static void 179 dnode_dest(void *arg, void *unused) 180 { 181 (void) unused; 182 dnode_t *dn = arg; 183 184 rw_destroy(&dn->dn_struct_rwlock); 185 mutex_destroy(&dn->dn_mtx); 186 mutex_destroy(&dn->dn_dbufs_mtx); 187 cv_destroy(&dn->dn_notxholds); 188 cv_destroy(&dn->dn_nodnholds); 189 zfs_refcount_destroy(&dn->dn_holds); 190 zfs_refcount_destroy(&dn->dn_tx_holds); 191 ASSERT(!list_link_active(&dn->dn_link)); 192 193 for (int i = 0; i < TXG_SIZE; i++) { 194 ASSERT(!multilist_link_active(&dn->dn_dirty_link[i])); 195 ASSERT3P(dn->dn_free_ranges[i], ==, NULL); 196 list_destroy(&dn->dn_dirty_records[i]); 197 ASSERT0(dn->dn_next_nblkptr[i]); 198 ASSERT0(dn->dn_next_nlevels[i]); 199 ASSERT0(dn->dn_next_indblkshift[i]); 200 ASSERT0(dn->dn_next_bonustype[i]); 201 ASSERT0(dn->dn_rm_spillblk[i]); 202 ASSERT0(dn->dn_next_bonuslen[i]); 203 ASSERT0(dn->dn_next_blksz[i]); 204 ASSERT0(dn->dn_next_maxblkid[i]); 205 } 206 207 ASSERT0(dn->dn_allocated_txg); 208 ASSERT0(dn->dn_free_txg); 209 ASSERT0(dn->dn_assigned_txg); 210 ASSERT0(dn->dn_dirty_txg); 211 ASSERT0(dn->dn_dirtyctx); 212 ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL); 213 ASSERT3P(dn->dn_bonus, ==, NULL); 214 ASSERT(!dn->dn_have_spill); 215 ASSERT3P(dn->dn_zio, ==, NULL); 216 ASSERT0(dn->dn_oldused); 217 ASSERT0(dn->dn_oldflags); 218 ASSERT0(dn->dn_olduid); 219 ASSERT0(dn->dn_oldgid); 220 ASSERT0(dn->dn_oldprojid); 221 ASSERT0(dn->dn_newuid); 222 ASSERT0(dn->dn_newgid); 223 ASSERT0(dn->dn_newprojid); 224 ASSERT0(dn->dn_id_flags); 225 226 ASSERT0(dn->dn_dbufs_count); 227 avl_destroy(&dn->dn_dbufs); 228 } 229 230 static int 231 dnode_kstats_update(kstat_t *ksp, int rw) 232 { 233 dnode_stats_t *ds = ksp->ks_data; 234 235 if (rw == KSTAT_WRITE) 236 return (EACCES); 237 ds->dnode_hold_dbuf_hold.value.ui64 = 238 wmsum_value(&dnode_sums.dnode_hold_dbuf_hold); 239 ds->dnode_hold_dbuf_read.value.ui64 = 240 wmsum_value(&dnode_sums.dnode_hold_dbuf_read); 241 ds->dnode_hold_alloc_hits.value.ui64 = 242 wmsum_value(&dnode_sums.dnode_hold_alloc_hits); 243 ds->dnode_hold_alloc_misses.value.ui64 = 244 wmsum_value(&dnode_sums.dnode_hold_alloc_misses); 245 ds->dnode_hold_alloc_interior.value.ui64 = 246 wmsum_value(&dnode_sums.dnode_hold_alloc_interior); 247 ds->dnode_hold_alloc_lock_retry.value.ui64 = 248 wmsum_value(&dnode_sums.dnode_hold_alloc_lock_retry); 249 ds->dnode_hold_alloc_lock_misses.value.ui64 = 250 wmsum_value(&dnode_sums.dnode_hold_alloc_lock_misses); 251 ds->dnode_hold_alloc_type_none.value.ui64 = 252 wmsum_value(&dnode_sums.dnode_hold_alloc_type_none); 253 ds->dnode_hold_free_hits.value.ui64 = 254 wmsum_value(&dnode_sums.dnode_hold_free_hits); 255 ds->dnode_hold_free_misses.value.ui64 = 256 wmsum_value(&dnode_sums.dnode_hold_free_misses); 257 ds->dnode_hold_free_lock_misses.value.ui64 = 258 wmsum_value(&dnode_sums.dnode_hold_free_lock_misses); 259 ds->dnode_hold_free_lock_retry.value.ui64 = 260 wmsum_value(&dnode_sums.dnode_hold_free_lock_retry); 261 ds->dnode_hold_free_refcount.value.ui64 = 262 wmsum_value(&dnode_sums.dnode_hold_free_refcount); 263 ds->dnode_hold_free_overflow.value.ui64 = 264 wmsum_value(&dnode_sums.dnode_hold_free_overflow); 265 ds->dnode_free_interior_lock_retry.value.ui64 = 266 wmsum_value(&dnode_sums.dnode_free_interior_lock_retry); 267 ds->dnode_allocate.value.ui64 = 268 wmsum_value(&dnode_sums.dnode_allocate); 269 ds->dnode_reallocate.value.ui64 = 270 wmsum_value(&dnode_sums.dnode_reallocate); 271 ds->dnode_buf_evict.value.ui64 = 272 wmsum_value(&dnode_sums.dnode_buf_evict); 273 ds->dnode_alloc_next_chunk.value.ui64 = 274 wmsum_value(&dnode_sums.dnode_alloc_next_chunk); 275 ds->dnode_alloc_race.value.ui64 = 276 wmsum_value(&dnode_sums.dnode_alloc_race); 277 ds->dnode_alloc_next_block.value.ui64 = 278 wmsum_value(&dnode_sums.dnode_alloc_next_block); 279 ds->dnode_move_invalid.value.ui64 = 280 wmsum_value(&dnode_sums.dnode_move_invalid); 281 ds->dnode_move_recheck1.value.ui64 = 282 wmsum_value(&dnode_sums.dnode_move_recheck1); 283 ds->dnode_move_recheck2.value.ui64 = 284 wmsum_value(&dnode_sums.dnode_move_recheck2); 285 ds->dnode_move_special.value.ui64 = 286 wmsum_value(&dnode_sums.dnode_move_special); 287 ds->dnode_move_handle.value.ui64 = 288 wmsum_value(&dnode_sums.dnode_move_handle); 289 ds->dnode_move_rwlock.value.ui64 = 290 wmsum_value(&dnode_sums.dnode_move_rwlock); 291 ds->dnode_move_active.value.ui64 = 292 wmsum_value(&dnode_sums.dnode_move_active); 293 return (0); 294 } 295 296 void 297 dnode_init(void) 298 { 299 ASSERT(dnode_cache == NULL); 300 dnode_cache = kmem_cache_create("dnode_t", sizeof (dnode_t), 301 0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0); 302 kmem_cache_set_move(dnode_cache, dnode_move); 303 304 wmsum_init(&dnode_sums.dnode_hold_dbuf_hold, 0); 305 wmsum_init(&dnode_sums.dnode_hold_dbuf_read, 0); 306 wmsum_init(&dnode_sums.dnode_hold_alloc_hits, 0); 307 wmsum_init(&dnode_sums.dnode_hold_alloc_misses, 0); 308 wmsum_init(&dnode_sums.dnode_hold_alloc_interior, 0); 309 wmsum_init(&dnode_sums.dnode_hold_alloc_lock_retry, 0); 310 wmsum_init(&dnode_sums.dnode_hold_alloc_lock_misses, 0); 311 wmsum_init(&dnode_sums.dnode_hold_alloc_type_none, 0); 312 wmsum_init(&dnode_sums.dnode_hold_free_hits, 0); 313 wmsum_init(&dnode_sums.dnode_hold_free_misses, 0); 314 wmsum_init(&dnode_sums.dnode_hold_free_lock_misses, 0); 315 wmsum_init(&dnode_sums.dnode_hold_free_lock_retry, 0); 316 wmsum_init(&dnode_sums.dnode_hold_free_refcount, 0); 317 wmsum_init(&dnode_sums.dnode_hold_free_overflow, 0); 318 wmsum_init(&dnode_sums.dnode_free_interior_lock_retry, 0); 319 wmsum_init(&dnode_sums.dnode_allocate, 0); 320 wmsum_init(&dnode_sums.dnode_reallocate, 0); 321 wmsum_init(&dnode_sums.dnode_buf_evict, 0); 322 wmsum_init(&dnode_sums.dnode_alloc_next_chunk, 0); 323 wmsum_init(&dnode_sums.dnode_alloc_race, 0); 324 wmsum_init(&dnode_sums.dnode_alloc_next_block, 0); 325 wmsum_init(&dnode_sums.dnode_move_invalid, 0); 326 wmsum_init(&dnode_sums.dnode_move_recheck1, 0); 327 wmsum_init(&dnode_sums.dnode_move_recheck2, 0); 328 wmsum_init(&dnode_sums.dnode_move_special, 0); 329 wmsum_init(&dnode_sums.dnode_move_handle, 0); 330 wmsum_init(&dnode_sums.dnode_move_rwlock, 0); 331 wmsum_init(&dnode_sums.dnode_move_active, 0); 332 333 dnode_ksp = kstat_create("zfs", 0, "dnodestats", "misc", 334 KSTAT_TYPE_NAMED, sizeof (dnode_stats) / sizeof (kstat_named_t), 335 KSTAT_FLAG_VIRTUAL); 336 if (dnode_ksp != NULL) { 337 dnode_ksp->ks_data = &dnode_stats; 338 dnode_ksp->ks_update = dnode_kstats_update; 339 kstat_install(dnode_ksp); 340 } 341 } 342 343 void 344 dnode_fini(void) 345 { 346 if (dnode_ksp != NULL) { 347 kstat_delete(dnode_ksp); 348 dnode_ksp = NULL; 349 } 350 351 wmsum_fini(&dnode_sums.dnode_hold_dbuf_hold); 352 wmsum_fini(&dnode_sums.dnode_hold_dbuf_read); 353 wmsum_fini(&dnode_sums.dnode_hold_alloc_hits); 354 wmsum_fini(&dnode_sums.dnode_hold_alloc_misses); 355 wmsum_fini(&dnode_sums.dnode_hold_alloc_interior); 356 wmsum_fini(&dnode_sums.dnode_hold_alloc_lock_retry); 357 wmsum_fini(&dnode_sums.dnode_hold_alloc_lock_misses); 358 wmsum_fini(&dnode_sums.dnode_hold_alloc_type_none); 359 wmsum_fini(&dnode_sums.dnode_hold_free_hits); 360 wmsum_fini(&dnode_sums.dnode_hold_free_misses); 361 wmsum_fini(&dnode_sums.dnode_hold_free_lock_misses); 362 wmsum_fini(&dnode_sums.dnode_hold_free_lock_retry); 363 wmsum_fini(&dnode_sums.dnode_hold_free_refcount); 364 wmsum_fini(&dnode_sums.dnode_hold_free_overflow); 365 wmsum_fini(&dnode_sums.dnode_free_interior_lock_retry); 366 wmsum_fini(&dnode_sums.dnode_allocate); 367 wmsum_fini(&dnode_sums.dnode_reallocate); 368 wmsum_fini(&dnode_sums.dnode_buf_evict); 369 wmsum_fini(&dnode_sums.dnode_alloc_next_chunk); 370 wmsum_fini(&dnode_sums.dnode_alloc_race); 371 wmsum_fini(&dnode_sums.dnode_alloc_next_block); 372 wmsum_fini(&dnode_sums.dnode_move_invalid); 373 wmsum_fini(&dnode_sums.dnode_move_recheck1); 374 wmsum_fini(&dnode_sums.dnode_move_recheck2); 375 wmsum_fini(&dnode_sums.dnode_move_special); 376 wmsum_fini(&dnode_sums.dnode_move_handle); 377 wmsum_fini(&dnode_sums.dnode_move_rwlock); 378 wmsum_fini(&dnode_sums.dnode_move_active); 379 380 kmem_cache_destroy(dnode_cache); 381 dnode_cache = NULL; 382 } 383 384 385 #ifdef ZFS_DEBUG 386 void 387 dnode_verify(dnode_t *dn) 388 { 389 int drop_struct_lock = FALSE; 390 391 ASSERT(dn->dn_phys); 392 ASSERT(dn->dn_objset); 393 ASSERT(dn->dn_handle->dnh_dnode == dn); 394 395 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type)); 396 397 if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY)) 398 return; 399 400 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) { 401 rw_enter(&dn->dn_struct_rwlock, RW_READER); 402 drop_struct_lock = TRUE; 403 } 404 if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) { 405 int i; 406 int max_bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots); 407 ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT); 408 if (dn->dn_datablkshift) { 409 ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT); 410 ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT); 411 ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz); 412 } 413 ASSERT3U(dn->dn_nlevels, <=, 30); 414 ASSERT(DMU_OT_IS_VALID(dn->dn_type)); 415 ASSERT3U(dn->dn_nblkptr, >=, 1); 416 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR); 417 ASSERT3U(dn->dn_bonuslen, <=, max_bonuslen); 418 ASSERT3U(dn->dn_datablksz, ==, 419 dn->dn_datablkszsec << SPA_MINBLOCKSHIFT); 420 ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0); 421 ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) + 422 dn->dn_bonuslen, <=, max_bonuslen); 423 for (i = 0; i < TXG_SIZE; i++) { 424 ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels); 425 } 426 } 427 if (dn->dn_phys->dn_type != DMU_OT_NONE) 428 ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels); 429 ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL); 430 if (dn->dn_dbuf != NULL) { 431 ASSERT3P(dn->dn_phys, ==, 432 (dnode_phys_t *)dn->dn_dbuf->db.db_data + 433 (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT))); 434 } 435 if (drop_struct_lock) 436 rw_exit(&dn->dn_struct_rwlock); 437 } 438 #endif 439 440 void 441 dnode_byteswap(dnode_phys_t *dnp) 442 { 443 uint64_t *buf64 = (void*)&dnp->dn_blkptr; 444 int i; 445 446 if (dnp->dn_type == DMU_OT_NONE) { 447 memset(dnp, 0, sizeof (dnode_phys_t)); 448 return; 449 } 450 451 dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec); 452 dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen); 453 dnp->dn_extra_slots = BSWAP_8(dnp->dn_extra_slots); 454 dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid); 455 dnp->dn_used = BSWAP_64(dnp->dn_used); 456 457 /* 458 * dn_nblkptr is only one byte, so it's OK to read it in either 459 * byte order. We can't read dn_bouslen. 460 */ 461 ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT); 462 ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR); 463 for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++) 464 buf64[i] = BSWAP_64(buf64[i]); 465 466 /* 467 * OK to check dn_bonuslen for zero, because it won't matter if 468 * we have the wrong byte order. This is necessary because the 469 * dnode dnode is smaller than a regular dnode. 470 */ 471 if (dnp->dn_bonuslen != 0) { 472 dmu_object_byteswap_t byteswap; 473 ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype)); 474 byteswap = DMU_OT_BYTESWAP(dnp->dn_bonustype); 475 dmu_ot_byteswap[byteswap].ob_func(DN_BONUS(dnp), 476 DN_MAX_BONUS_LEN(dnp)); 477 } 478 479 /* Swap SPILL block if we have one */ 480 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) 481 byteswap_uint64_array(DN_SPILL_BLKPTR(dnp), sizeof (blkptr_t)); 482 } 483 484 void 485 dnode_buf_byteswap(void *vbuf, size_t size) 486 { 487 int i = 0; 488 489 ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT)); 490 ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0); 491 492 while (i < size) { 493 dnode_phys_t *dnp = (void *)(((char *)vbuf) + i); 494 dnode_byteswap(dnp); 495 496 i += DNODE_MIN_SIZE; 497 if (dnp->dn_type != DMU_OT_NONE) 498 i += dnp->dn_extra_slots * DNODE_MIN_SIZE; 499 } 500 } 501 502 void 503 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx) 504 { 505 ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1); 506 507 dnode_setdirty(dn, tx); 508 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 509 ASSERT3U(newsize, <=, DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) - 510 (dn->dn_nblkptr-1) * sizeof (blkptr_t)); 511 512 if (newsize < dn->dn_bonuslen) { 513 /* clear any data after the end of the new size */ 514 size_t diff = dn->dn_bonuslen - newsize; 515 char *data_end = ((char *)dn->dn_bonus->db.db_data) + newsize; 516 memset(data_end, 0, diff); 517 } 518 519 dn->dn_bonuslen = newsize; 520 if (newsize == 0) 521 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN; 522 else 523 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen; 524 rw_exit(&dn->dn_struct_rwlock); 525 } 526 527 void 528 dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx) 529 { 530 ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1); 531 dnode_setdirty(dn, tx); 532 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 533 dn->dn_bonustype = newtype; 534 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype; 535 rw_exit(&dn->dn_struct_rwlock); 536 } 537 538 void 539 dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx) 540 { 541 ASSERT3U(zfs_refcount_count(&dn->dn_holds), >=, 1); 542 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 543 dnode_setdirty(dn, tx); 544 dn->dn_rm_spillblk[tx->tx_txg & TXG_MASK] = DN_KILL_SPILLBLK; 545 dn->dn_have_spill = B_FALSE; 546 } 547 548 static void 549 dnode_setdblksz(dnode_t *dn, int size) 550 { 551 ASSERT0(P2PHASE(size, SPA_MINBLOCKSIZE)); 552 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE); 553 ASSERT3U(size, >=, SPA_MINBLOCKSIZE); 554 ASSERT3U(size >> SPA_MINBLOCKSHIFT, <, 555 1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8)); 556 dn->dn_datablksz = size; 557 dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT; 558 dn->dn_datablkshift = ISP2(size) ? highbit64(size - 1) : 0; 559 } 560 561 static dnode_t * 562 dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db, 563 uint64_t object, dnode_handle_t *dnh) 564 { 565 dnode_t *dn; 566 567 dn = kmem_cache_alloc(dnode_cache, KM_SLEEP); 568 dn->dn_moved = 0; 569 570 /* 571 * Defer setting dn_objset until the dnode is ready to be a candidate 572 * for the dnode_move() callback. 573 */ 574 dn->dn_object = object; 575 dn->dn_dbuf = db; 576 dn->dn_handle = dnh; 577 dn->dn_phys = dnp; 578 579 if (dnp->dn_datablkszsec) { 580 dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT); 581 } else { 582 dn->dn_datablksz = 0; 583 dn->dn_datablkszsec = 0; 584 dn->dn_datablkshift = 0; 585 } 586 dn->dn_indblkshift = dnp->dn_indblkshift; 587 dn->dn_nlevels = dnp->dn_nlevels; 588 dn->dn_type = dnp->dn_type; 589 dn->dn_nblkptr = dnp->dn_nblkptr; 590 dn->dn_checksum = dnp->dn_checksum; 591 dn->dn_compress = dnp->dn_compress; 592 dn->dn_bonustype = dnp->dn_bonustype; 593 dn->dn_bonuslen = dnp->dn_bonuslen; 594 dn->dn_num_slots = dnp->dn_extra_slots + 1; 595 dn->dn_maxblkid = dnp->dn_maxblkid; 596 dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0); 597 dn->dn_id_flags = 0; 598 599 dmu_zfetch_init(&dn->dn_zfetch, dn); 600 601 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type)); 602 ASSERT(zrl_is_locked(&dnh->dnh_zrlock)); 603 ASSERT(!DN_SLOT_IS_PTR(dnh->dnh_dnode)); 604 605 mutex_enter(&os->os_lock); 606 607 /* 608 * Exclude special dnodes from os_dnodes so an empty os_dnodes 609 * signifies that the special dnodes have no references from 610 * their children (the entries in os_dnodes). This allows 611 * dnode_destroy() to easily determine if the last child has 612 * been removed and then complete eviction of the objset. 613 */ 614 if (!DMU_OBJECT_IS_SPECIAL(object)) 615 list_insert_head(&os->os_dnodes, dn); 616 membar_producer(); 617 618 /* 619 * Everything else must be valid before assigning dn_objset 620 * makes the dnode eligible for dnode_move(). 621 */ 622 dn->dn_objset = os; 623 624 dnh->dnh_dnode = dn; 625 mutex_exit(&os->os_lock); 626 627 arc_space_consume(sizeof (dnode_t), ARC_SPACE_DNODE); 628 629 return (dn); 630 } 631 632 /* 633 * Caller must be holding the dnode handle, which is released upon return. 634 */ 635 static void 636 dnode_destroy(dnode_t *dn) 637 { 638 objset_t *os = dn->dn_objset; 639 boolean_t complete_os_eviction = B_FALSE; 640 641 ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0); 642 643 mutex_enter(&os->os_lock); 644 POINTER_INVALIDATE(&dn->dn_objset); 645 if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) { 646 list_remove(&os->os_dnodes, dn); 647 complete_os_eviction = 648 list_is_empty(&os->os_dnodes) && 649 list_link_active(&os->os_evicting_node); 650 } 651 mutex_exit(&os->os_lock); 652 653 /* the dnode can no longer move, so we can release the handle */ 654 if (!zrl_is_locked(&dn->dn_handle->dnh_zrlock)) 655 zrl_remove(&dn->dn_handle->dnh_zrlock); 656 657 dn->dn_allocated_txg = 0; 658 dn->dn_free_txg = 0; 659 dn->dn_assigned_txg = 0; 660 dn->dn_dirty_txg = 0; 661 662 dn->dn_dirtyctx = 0; 663 dn->dn_dirtyctx_firstset = NULL; 664 if (dn->dn_bonus != NULL) { 665 mutex_enter(&dn->dn_bonus->db_mtx); 666 dbuf_destroy(dn->dn_bonus); 667 dn->dn_bonus = NULL; 668 } 669 dn->dn_zio = NULL; 670 671 dn->dn_have_spill = B_FALSE; 672 dn->dn_oldused = 0; 673 dn->dn_oldflags = 0; 674 dn->dn_olduid = 0; 675 dn->dn_oldgid = 0; 676 dn->dn_oldprojid = ZFS_DEFAULT_PROJID; 677 dn->dn_newuid = 0; 678 dn->dn_newgid = 0; 679 dn->dn_newprojid = ZFS_DEFAULT_PROJID; 680 dn->dn_id_flags = 0; 681 682 dmu_zfetch_fini(&dn->dn_zfetch); 683 kmem_cache_free(dnode_cache, dn); 684 arc_space_return(sizeof (dnode_t), ARC_SPACE_DNODE); 685 686 if (complete_os_eviction) 687 dmu_objset_evict_done(os); 688 } 689 690 void 691 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs, 692 dmu_object_type_t bonustype, int bonuslen, int dn_slots, dmu_tx_t *tx) 693 { 694 int i; 695 696 ASSERT3U(dn_slots, >, 0); 697 ASSERT3U(dn_slots << DNODE_SHIFT, <=, 698 spa_maxdnodesize(dmu_objset_spa(dn->dn_objset))); 699 ASSERT3U(blocksize, <=, 700 spa_maxblocksize(dmu_objset_spa(dn->dn_objset))); 701 if (blocksize == 0) 702 blocksize = 1 << zfs_default_bs; 703 else 704 blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE); 705 706 if (ibs == 0) 707 ibs = zfs_default_ibs; 708 709 ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT); 710 711 dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d dn_slots=%d\n", 712 dn->dn_objset, (u_longlong_t)dn->dn_object, 713 (u_longlong_t)tx->tx_txg, blocksize, ibs, dn_slots); 714 DNODE_STAT_BUMP(dnode_allocate); 715 716 ASSERT(dn->dn_type == DMU_OT_NONE); 717 ASSERT0(memcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t))); 718 ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE); 719 ASSERT(ot != DMU_OT_NONE); 720 ASSERT(DMU_OT_IS_VALID(ot)); 721 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) || 722 (bonustype == DMU_OT_SA && bonuslen == 0) || 723 (bonustype == DMU_OTN_UINT64_METADATA && bonuslen == 0) || 724 (bonustype != DMU_OT_NONE && bonuslen != 0)); 725 ASSERT(DMU_OT_IS_VALID(bonustype)); 726 ASSERT3U(bonuslen, <=, DN_SLOTS_TO_BONUSLEN(dn_slots)); 727 ASSERT(dn->dn_type == DMU_OT_NONE); 728 ASSERT0(dn->dn_maxblkid); 729 ASSERT0(dn->dn_allocated_txg); 730 ASSERT0(dn->dn_assigned_txg); 731 ASSERT(zfs_refcount_is_zero(&dn->dn_tx_holds)); 732 ASSERT3U(zfs_refcount_count(&dn->dn_holds), <=, 1); 733 ASSERT(avl_is_empty(&dn->dn_dbufs)); 734 735 for (i = 0; i < TXG_SIZE; i++) { 736 ASSERT0(dn->dn_next_nblkptr[i]); 737 ASSERT0(dn->dn_next_nlevels[i]); 738 ASSERT0(dn->dn_next_indblkshift[i]); 739 ASSERT0(dn->dn_next_bonuslen[i]); 740 ASSERT0(dn->dn_next_bonustype[i]); 741 ASSERT0(dn->dn_rm_spillblk[i]); 742 ASSERT0(dn->dn_next_blksz[i]); 743 ASSERT0(dn->dn_next_maxblkid[i]); 744 ASSERT(!multilist_link_active(&dn->dn_dirty_link[i])); 745 ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL); 746 ASSERT3P(dn->dn_free_ranges[i], ==, NULL); 747 } 748 749 dn->dn_type = ot; 750 dnode_setdblksz(dn, blocksize); 751 dn->dn_indblkshift = ibs; 752 dn->dn_nlevels = 1; 753 dn->dn_num_slots = dn_slots; 754 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */ 755 dn->dn_nblkptr = 1; 756 else { 757 dn->dn_nblkptr = MIN(DN_MAX_NBLKPTR, 758 1 + ((DN_SLOTS_TO_BONUSLEN(dn_slots) - bonuslen) >> 759 SPA_BLKPTRSHIFT)); 760 } 761 762 dn->dn_bonustype = bonustype; 763 dn->dn_bonuslen = bonuslen; 764 dn->dn_checksum = ZIO_CHECKSUM_INHERIT; 765 dn->dn_compress = ZIO_COMPRESS_INHERIT; 766 dn->dn_dirtyctx = 0; 767 768 dn->dn_free_txg = 0; 769 dn->dn_dirtyctx_firstset = NULL; 770 dn->dn_dirty_txg = 0; 771 772 dn->dn_allocated_txg = tx->tx_txg; 773 dn->dn_id_flags = 0; 774 775 dnode_setdirty(dn, tx); 776 dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs; 777 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen; 778 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype; 779 dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz; 780 } 781 782 void 783 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, 784 dmu_object_type_t bonustype, int bonuslen, int dn_slots, 785 boolean_t keep_spill, dmu_tx_t *tx) 786 { 787 int nblkptr; 788 789 ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE); 790 ASSERT3U(blocksize, <=, 791 spa_maxblocksize(dmu_objset_spa(dn->dn_objset))); 792 ASSERT0(blocksize % SPA_MINBLOCKSIZE); 793 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx)); 794 ASSERT(tx->tx_txg != 0); 795 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) || 796 (bonustype != DMU_OT_NONE && bonuslen != 0) || 797 (bonustype == DMU_OT_SA && bonuslen == 0)); 798 ASSERT(DMU_OT_IS_VALID(bonustype)); 799 ASSERT3U(bonuslen, <=, 800 DN_BONUS_SIZE(spa_maxdnodesize(dmu_objset_spa(dn->dn_objset)))); 801 ASSERT3U(bonuslen, <=, DN_BONUS_SIZE(dn_slots << DNODE_SHIFT)); 802 803 dnode_free_interior_slots(dn); 804 DNODE_STAT_BUMP(dnode_reallocate); 805 806 /* clean up any unreferenced dbufs */ 807 dnode_evict_dbufs(dn); 808 809 dn->dn_id_flags = 0; 810 811 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 812 dnode_setdirty(dn, tx); 813 if (dn->dn_datablksz != blocksize) { 814 /* change blocksize */ 815 ASSERT0(dn->dn_maxblkid); 816 ASSERT(BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) || 817 dnode_block_freed(dn, 0)); 818 819 dnode_setdblksz(dn, blocksize); 820 dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = blocksize; 821 } 822 if (dn->dn_bonuslen != bonuslen) 823 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = bonuslen; 824 825 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */ 826 nblkptr = 1; 827 else 828 nblkptr = MIN(DN_MAX_NBLKPTR, 829 1 + ((DN_SLOTS_TO_BONUSLEN(dn_slots) - bonuslen) >> 830 SPA_BLKPTRSHIFT)); 831 if (dn->dn_bonustype != bonustype) 832 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = bonustype; 833 if (dn->dn_nblkptr != nblkptr) 834 dn->dn_next_nblkptr[tx->tx_txg & TXG_MASK] = nblkptr; 835 if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR && !keep_spill) { 836 dbuf_rm_spill(dn, tx); 837 dnode_rm_spill(dn, tx); 838 } 839 840 rw_exit(&dn->dn_struct_rwlock); 841 842 /* change type */ 843 dn->dn_type = ot; 844 845 /* change bonus size and type */ 846 mutex_enter(&dn->dn_mtx); 847 dn->dn_bonustype = bonustype; 848 dn->dn_bonuslen = bonuslen; 849 dn->dn_num_slots = dn_slots; 850 dn->dn_nblkptr = nblkptr; 851 dn->dn_checksum = ZIO_CHECKSUM_INHERIT; 852 dn->dn_compress = ZIO_COMPRESS_INHERIT; 853 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR); 854 855 /* fix up the bonus db_size */ 856 if (dn->dn_bonus) { 857 dn->dn_bonus->db.db_size = 858 DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) - 859 (dn->dn_nblkptr-1) * sizeof (blkptr_t); 860 ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size); 861 } 862 863 dn->dn_allocated_txg = tx->tx_txg; 864 mutex_exit(&dn->dn_mtx); 865 } 866 867 #ifdef _KERNEL 868 static void 869 dnode_move_impl(dnode_t *odn, dnode_t *ndn) 870 { 871 ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock)); 872 ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx)); 873 ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx)); 874 875 /* Copy fields. */ 876 ndn->dn_objset = odn->dn_objset; 877 ndn->dn_object = odn->dn_object; 878 ndn->dn_dbuf = odn->dn_dbuf; 879 ndn->dn_handle = odn->dn_handle; 880 ndn->dn_phys = odn->dn_phys; 881 ndn->dn_type = odn->dn_type; 882 ndn->dn_bonuslen = odn->dn_bonuslen; 883 ndn->dn_bonustype = odn->dn_bonustype; 884 ndn->dn_nblkptr = odn->dn_nblkptr; 885 ndn->dn_checksum = odn->dn_checksum; 886 ndn->dn_compress = odn->dn_compress; 887 ndn->dn_nlevels = odn->dn_nlevels; 888 ndn->dn_indblkshift = odn->dn_indblkshift; 889 ndn->dn_datablkshift = odn->dn_datablkshift; 890 ndn->dn_datablkszsec = odn->dn_datablkszsec; 891 ndn->dn_datablksz = odn->dn_datablksz; 892 ndn->dn_maxblkid = odn->dn_maxblkid; 893 ndn->dn_num_slots = odn->dn_num_slots; 894 memcpy(ndn->dn_next_type, odn->dn_next_type, 895 sizeof (odn->dn_next_type)); 896 memcpy(ndn->dn_next_nblkptr, odn->dn_next_nblkptr, 897 sizeof (odn->dn_next_nblkptr)); 898 memcpy(ndn->dn_next_nlevels, odn->dn_next_nlevels, 899 sizeof (odn->dn_next_nlevels)); 900 memcpy(ndn->dn_next_indblkshift, odn->dn_next_indblkshift, 901 sizeof (odn->dn_next_indblkshift)); 902 memcpy(ndn->dn_next_bonustype, odn->dn_next_bonustype, 903 sizeof (odn->dn_next_bonustype)); 904 memcpy(ndn->dn_rm_spillblk, odn->dn_rm_spillblk, 905 sizeof (odn->dn_rm_spillblk)); 906 memcpy(ndn->dn_next_bonuslen, odn->dn_next_bonuslen, 907 sizeof (odn->dn_next_bonuslen)); 908 memcpy(ndn->dn_next_blksz, odn->dn_next_blksz, 909 sizeof (odn->dn_next_blksz)); 910 memcpy(ndn->dn_next_maxblkid, odn->dn_next_maxblkid, 911 sizeof (odn->dn_next_maxblkid)); 912 for (int i = 0; i < TXG_SIZE; i++) { 913 list_move_tail(&ndn->dn_dirty_records[i], 914 &odn->dn_dirty_records[i]); 915 } 916 memcpy(ndn->dn_free_ranges, odn->dn_free_ranges, 917 sizeof (odn->dn_free_ranges)); 918 ndn->dn_allocated_txg = odn->dn_allocated_txg; 919 ndn->dn_free_txg = odn->dn_free_txg; 920 ndn->dn_assigned_txg = odn->dn_assigned_txg; 921 ndn->dn_dirty_txg = odn->dn_dirty_txg; 922 ndn->dn_dirtyctx = odn->dn_dirtyctx; 923 ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset; 924 ASSERT(zfs_refcount_count(&odn->dn_tx_holds) == 0); 925 zfs_refcount_transfer(&ndn->dn_holds, &odn->dn_holds); 926 ASSERT(avl_is_empty(&ndn->dn_dbufs)); 927 avl_swap(&ndn->dn_dbufs, &odn->dn_dbufs); 928 ndn->dn_dbufs_count = odn->dn_dbufs_count; 929 ndn->dn_bonus = odn->dn_bonus; 930 ndn->dn_have_spill = odn->dn_have_spill; 931 ndn->dn_zio = odn->dn_zio; 932 ndn->dn_oldused = odn->dn_oldused; 933 ndn->dn_oldflags = odn->dn_oldflags; 934 ndn->dn_olduid = odn->dn_olduid; 935 ndn->dn_oldgid = odn->dn_oldgid; 936 ndn->dn_oldprojid = odn->dn_oldprojid; 937 ndn->dn_newuid = odn->dn_newuid; 938 ndn->dn_newgid = odn->dn_newgid; 939 ndn->dn_newprojid = odn->dn_newprojid; 940 ndn->dn_id_flags = odn->dn_id_flags; 941 dmu_zfetch_init(&ndn->dn_zfetch, ndn); 942 943 /* 944 * Update back pointers. Updating the handle fixes the back pointer of 945 * every descendant dbuf as well as the bonus dbuf. 946 */ 947 ASSERT(ndn->dn_handle->dnh_dnode == odn); 948 ndn->dn_handle->dnh_dnode = ndn; 949 950 /* 951 * Invalidate the original dnode by clearing all of its back pointers. 952 */ 953 odn->dn_dbuf = NULL; 954 odn->dn_handle = NULL; 955 avl_create(&odn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t), 956 offsetof(dmu_buf_impl_t, db_link)); 957 odn->dn_dbufs_count = 0; 958 odn->dn_bonus = NULL; 959 dmu_zfetch_fini(&odn->dn_zfetch); 960 961 /* 962 * Set the low bit of the objset pointer to ensure that dnode_move() 963 * recognizes the dnode as invalid in any subsequent callback. 964 */ 965 POINTER_INVALIDATE(&odn->dn_objset); 966 967 /* 968 * Satisfy the destructor. 969 */ 970 for (int i = 0; i < TXG_SIZE; i++) { 971 list_create(&odn->dn_dirty_records[i], 972 sizeof (dbuf_dirty_record_t), 973 offsetof(dbuf_dirty_record_t, dr_dirty_node)); 974 odn->dn_free_ranges[i] = NULL; 975 odn->dn_next_nlevels[i] = 0; 976 odn->dn_next_indblkshift[i] = 0; 977 odn->dn_next_bonustype[i] = 0; 978 odn->dn_rm_spillblk[i] = 0; 979 odn->dn_next_bonuslen[i] = 0; 980 odn->dn_next_blksz[i] = 0; 981 } 982 odn->dn_allocated_txg = 0; 983 odn->dn_free_txg = 0; 984 odn->dn_assigned_txg = 0; 985 odn->dn_dirty_txg = 0; 986 odn->dn_dirtyctx = 0; 987 odn->dn_dirtyctx_firstset = NULL; 988 odn->dn_have_spill = B_FALSE; 989 odn->dn_zio = NULL; 990 odn->dn_oldused = 0; 991 odn->dn_oldflags = 0; 992 odn->dn_olduid = 0; 993 odn->dn_oldgid = 0; 994 odn->dn_oldprojid = ZFS_DEFAULT_PROJID; 995 odn->dn_newuid = 0; 996 odn->dn_newgid = 0; 997 odn->dn_newprojid = ZFS_DEFAULT_PROJID; 998 odn->dn_id_flags = 0; 999 1000 /* 1001 * Mark the dnode. 1002 */ 1003 ndn->dn_moved = 1; 1004 odn->dn_moved = (uint8_t)-1; 1005 } 1006 1007 static kmem_cbrc_t 1008 dnode_move(void *buf, void *newbuf, size_t size, void *arg) 1009 { 1010 dnode_t *odn = buf, *ndn = newbuf; 1011 objset_t *os; 1012 int64_t refcount; 1013 uint32_t dbufs; 1014 1015 /* 1016 * The dnode is on the objset's list of known dnodes if the objset 1017 * pointer is valid. We set the low bit of the objset pointer when 1018 * freeing the dnode to invalidate it, and the memory patterns written 1019 * by kmem (baddcafe and deadbeef) set at least one of the two low bits. 1020 * A newly created dnode sets the objset pointer last of all to indicate 1021 * that the dnode is known and in a valid state to be moved by this 1022 * function. 1023 */ 1024 os = odn->dn_objset; 1025 if (!POINTER_IS_VALID(os)) { 1026 DNODE_STAT_BUMP(dnode_move_invalid); 1027 return (KMEM_CBRC_DONT_KNOW); 1028 } 1029 1030 /* 1031 * Ensure that the objset does not go away during the move. 1032 */ 1033 rw_enter(&os_lock, RW_WRITER); 1034 if (os != odn->dn_objset) { 1035 rw_exit(&os_lock); 1036 DNODE_STAT_BUMP(dnode_move_recheck1); 1037 return (KMEM_CBRC_DONT_KNOW); 1038 } 1039 1040 /* 1041 * If the dnode is still valid, then so is the objset. We know that no 1042 * valid objset can be freed while we hold os_lock, so we can safely 1043 * ensure that the objset remains in use. 1044 */ 1045 mutex_enter(&os->os_lock); 1046 1047 /* 1048 * Recheck the objset pointer in case the dnode was removed just before 1049 * acquiring the lock. 1050 */ 1051 if (os != odn->dn_objset) { 1052 mutex_exit(&os->os_lock); 1053 rw_exit(&os_lock); 1054 DNODE_STAT_BUMP(dnode_move_recheck2); 1055 return (KMEM_CBRC_DONT_KNOW); 1056 } 1057 1058 /* 1059 * At this point we know that as long as we hold os->os_lock, the dnode 1060 * cannot be freed and fields within the dnode can be safely accessed. 1061 * The objset listing this dnode cannot go away as long as this dnode is 1062 * on its list. 1063 */ 1064 rw_exit(&os_lock); 1065 if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) { 1066 mutex_exit(&os->os_lock); 1067 DNODE_STAT_BUMP(dnode_move_special); 1068 return (KMEM_CBRC_NO); 1069 } 1070 ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */ 1071 1072 /* 1073 * Lock the dnode handle to prevent the dnode from obtaining any new 1074 * holds. This also prevents the descendant dbufs and the bonus dbuf 1075 * from accessing the dnode, so that we can discount their holds. The 1076 * handle is safe to access because we know that while the dnode cannot 1077 * go away, neither can its handle. Once we hold dnh_zrlock, we can 1078 * safely move any dnode referenced only by dbufs. 1079 */ 1080 if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) { 1081 mutex_exit(&os->os_lock); 1082 DNODE_STAT_BUMP(dnode_move_handle); 1083 return (KMEM_CBRC_LATER); 1084 } 1085 1086 /* 1087 * Ensure a consistent view of the dnode's holds and the dnode's dbufs. 1088 * We need to guarantee that there is a hold for every dbuf in order to 1089 * determine whether the dnode is actively referenced. Falsely matching 1090 * a dbuf to an active hold would lead to an unsafe move. It's possible 1091 * that a thread already having an active dnode hold is about to add a 1092 * dbuf, and we can't compare hold and dbuf counts while the add is in 1093 * progress. 1094 */ 1095 if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) { 1096 zrl_exit(&odn->dn_handle->dnh_zrlock); 1097 mutex_exit(&os->os_lock); 1098 DNODE_STAT_BUMP(dnode_move_rwlock); 1099 return (KMEM_CBRC_LATER); 1100 } 1101 1102 /* 1103 * A dbuf may be removed (evicted) without an active dnode hold. In that 1104 * case, the dbuf count is decremented under the handle lock before the 1105 * dbuf's hold is released. This order ensures that if we count the hold 1106 * after the dbuf is removed but before its hold is released, we will 1107 * treat the unmatched hold as active and exit safely. If we count the 1108 * hold before the dbuf is removed, the hold is discounted, and the 1109 * removal is blocked until the move completes. 1110 */ 1111 refcount = zfs_refcount_count(&odn->dn_holds); 1112 ASSERT(refcount >= 0); 1113 dbufs = DN_DBUFS_COUNT(odn); 1114 1115 /* We can't have more dbufs than dnode holds. */ 1116 ASSERT3U(dbufs, <=, refcount); 1117 DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount, 1118 uint32_t, dbufs); 1119 1120 if (refcount > dbufs) { 1121 rw_exit(&odn->dn_struct_rwlock); 1122 zrl_exit(&odn->dn_handle->dnh_zrlock); 1123 mutex_exit(&os->os_lock); 1124 DNODE_STAT_BUMP(dnode_move_active); 1125 return (KMEM_CBRC_LATER); 1126 } 1127 1128 rw_exit(&odn->dn_struct_rwlock); 1129 1130 /* 1131 * At this point we know that anyone with a hold on the dnode is not 1132 * actively referencing it. The dnode is known and in a valid state to 1133 * move. We're holding the locks needed to execute the critical section. 1134 */ 1135 dnode_move_impl(odn, ndn); 1136 1137 list_link_replace(&odn->dn_link, &ndn->dn_link); 1138 /* If the dnode was safe to move, the refcount cannot have changed. */ 1139 ASSERT(refcount == zfs_refcount_count(&ndn->dn_holds)); 1140 ASSERT(dbufs == DN_DBUFS_COUNT(ndn)); 1141 zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */ 1142 mutex_exit(&os->os_lock); 1143 1144 return (KMEM_CBRC_YES); 1145 } 1146 #endif /* _KERNEL */ 1147 1148 static void 1149 dnode_slots_hold(dnode_children_t *children, int idx, int slots) 1150 { 1151 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK); 1152 1153 for (int i = idx; i < idx + slots; i++) { 1154 dnode_handle_t *dnh = &children->dnc_children[i]; 1155 zrl_add(&dnh->dnh_zrlock); 1156 } 1157 } 1158 1159 static void 1160 dnode_slots_rele(dnode_children_t *children, int idx, int slots) 1161 { 1162 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK); 1163 1164 for (int i = idx; i < idx + slots; i++) { 1165 dnode_handle_t *dnh = &children->dnc_children[i]; 1166 1167 if (zrl_is_locked(&dnh->dnh_zrlock)) 1168 zrl_exit(&dnh->dnh_zrlock); 1169 else 1170 zrl_remove(&dnh->dnh_zrlock); 1171 } 1172 } 1173 1174 static int 1175 dnode_slots_tryenter(dnode_children_t *children, int idx, int slots) 1176 { 1177 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK); 1178 1179 for (int i = idx; i < idx + slots; i++) { 1180 dnode_handle_t *dnh = &children->dnc_children[i]; 1181 1182 if (!zrl_tryenter(&dnh->dnh_zrlock)) { 1183 for (int j = idx; j < i; j++) { 1184 dnh = &children->dnc_children[j]; 1185 zrl_exit(&dnh->dnh_zrlock); 1186 } 1187 1188 return (0); 1189 } 1190 } 1191 1192 return (1); 1193 } 1194 1195 static void 1196 dnode_set_slots(dnode_children_t *children, int idx, int slots, void *ptr) 1197 { 1198 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK); 1199 1200 for (int i = idx; i < idx + slots; i++) { 1201 dnode_handle_t *dnh = &children->dnc_children[i]; 1202 dnh->dnh_dnode = ptr; 1203 } 1204 } 1205 1206 static boolean_t 1207 dnode_check_slots_free(dnode_children_t *children, int idx, int slots) 1208 { 1209 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK); 1210 1211 /* 1212 * If all dnode slots are either already free or 1213 * evictable return B_TRUE. 1214 */ 1215 for (int i = idx; i < idx + slots; i++) { 1216 dnode_handle_t *dnh = &children->dnc_children[i]; 1217 dnode_t *dn = dnh->dnh_dnode; 1218 1219 if (dn == DN_SLOT_FREE) { 1220 continue; 1221 } else if (DN_SLOT_IS_PTR(dn)) { 1222 mutex_enter(&dn->dn_mtx); 1223 boolean_t can_free = (dn->dn_type == DMU_OT_NONE && 1224 zfs_refcount_is_zero(&dn->dn_holds) && 1225 !DNODE_IS_DIRTY(dn)); 1226 mutex_exit(&dn->dn_mtx); 1227 1228 if (!can_free) 1229 return (B_FALSE); 1230 else 1231 continue; 1232 } else { 1233 return (B_FALSE); 1234 } 1235 } 1236 1237 return (B_TRUE); 1238 } 1239 1240 static void 1241 dnode_reclaim_slots(dnode_children_t *children, int idx, int slots) 1242 { 1243 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK); 1244 1245 for (int i = idx; i < idx + slots; i++) { 1246 dnode_handle_t *dnh = &children->dnc_children[i]; 1247 1248 ASSERT(zrl_is_locked(&dnh->dnh_zrlock)); 1249 1250 if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) { 1251 ASSERT3S(dnh->dnh_dnode->dn_type, ==, DMU_OT_NONE); 1252 dnode_destroy(dnh->dnh_dnode); 1253 dnh->dnh_dnode = DN_SLOT_FREE; 1254 } 1255 } 1256 } 1257 1258 void 1259 dnode_free_interior_slots(dnode_t *dn) 1260 { 1261 dnode_children_t *children = dmu_buf_get_user(&dn->dn_dbuf->db); 1262 int epb = dn->dn_dbuf->db.db_size >> DNODE_SHIFT; 1263 int idx = (dn->dn_object & (epb - 1)) + 1; 1264 int slots = dn->dn_num_slots - 1; 1265 1266 if (slots == 0) 1267 return; 1268 1269 ASSERT3S(idx + slots, <=, DNODES_PER_BLOCK); 1270 1271 while (!dnode_slots_tryenter(children, idx, slots)) { 1272 DNODE_STAT_BUMP(dnode_free_interior_lock_retry); 1273 kpreempt(KPREEMPT_SYNC); 1274 } 1275 1276 dnode_set_slots(children, idx, slots, DN_SLOT_FREE); 1277 dnode_slots_rele(children, idx, slots); 1278 } 1279 1280 void 1281 dnode_special_close(dnode_handle_t *dnh) 1282 { 1283 dnode_t *dn = dnh->dnh_dnode; 1284 1285 /* 1286 * Ensure dnode_rele_and_unlock() has released dn_mtx, after final 1287 * zfs_refcount_remove() 1288 */ 1289 mutex_enter(&dn->dn_mtx); 1290 if (zfs_refcount_count(&dn->dn_holds) > 0) 1291 cv_wait(&dn->dn_nodnholds, &dn->dn_mtx); 1292 mutex_exit(&dn->dn_mtx); 1293 ASSERT3U(zfs_refcount_count(&dn->dn_holds), ==, 0); 1294 1295 ASSERT(dn->dn_dbuf == NULL || 1296 dmu_buf_get_user(&dn->dn_dbuf->db) == NULL); 1297 zrl_add(&dnh->dnh_zrlock); 1298 dnode_destroy(dn); /* implicit zrl_remove() */ 1299 zrl_destroy(&dnh->dnh_zrlock); 1300 dnh->dnh_dnode = NULL; 1301 } 1302 1303 void 1304 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object, 1305 dnode_handle_t *dnh) 1306 { 1307 dnode_t *dn; 1308 1309 zrl_init(&dnh->dnh_zrlock); 1310 VERIFY3U(1, ==, zrl_tryenter(&dnh->dnh_zrlock)); 1311 1312 dn = dnode_create(os, dnp, NULL, object, dnh); 1313 DNODE_VERIFY(dn); 1314 1315 zrl_exit(&dnh->dnh_zrlock); 1316 } 1317 1318 static void 1319 dnode_buf_evict_async(void *dbu) 1320 { 1321 dnode_children_t *dnc = dbu; 1322 1323 DNODE_STAT_BUMP(dnode_buf_evict); 1324 1325 for (int i = 0; i < dnc->dnc_count; i++) { 1326 dnode_handle_t *dnh = &dnc->dnc_children[i]; 1327 dnode_t *dn; 1328 1329 /* 1330 * The dnode handle lock guards against the dnode moving to 1331 * another valid address, so there is no need here to guard 1332 * against changes to or from NULL. 1333 */ 1334 if (!DN_SLOT_IS_PTR(dnh->dnh_dnode)) { 1335 zrl_destroy(&dnh->dnh_zrlock); 1336 dnh->dnh_dnode = DN_SLOT_UNINIT; 1337 continue; 1338 } 1339 1340 zrl_add(&dnh->dnh_zrlock); 1341 dn = dnh->dnh_dnode; 1342 /* 1343 * If there are holds on this dnode, then there should 1344 * be holds on the dnode's containing dbuf as well; thus 1345 * it wouldn't be eligible for eviction and this function 1346 * would not have been called. 1347 */ 1348 ASSERT(zfs_refcount_is_zero(&dn->dn_holds)); 1349 ASSERT(zfs_refcount_is_zero(&dn->dn_tx_holds)); 1350 1351 dnode_destroy(dn); /* implicit zrl_remove() for first slot */ 1352 zrl_destroy(&dnh->dnh_zrlock); 1353 dnh->dnh_dnode = DN_SLOT_UNINIT; 1354 } 1355 kmem_free(dnc, sizeof (dnode_children_t) + 1356 dnc->dnc_count * sizeof (dnode_handle_t)); 1357 } 1358 1359 /* 1360 * When the DNODE_MUST_BE_FREE flag is set, the "slots" parameter is used 1361 * to ensure the hole at the specified object offset is large enough to 1362 * hold the dnode being created. The slots parameter is also used to ensure 1363 * a dnode does not span multiple dnode blocks. In both of these cases, if 1364 * a failure occurs, ENOSPC is returned. Keep in mind, these failure cases 1365 * are only possible when using DNODE_MUST_BE_FREE. 1366 * 1367 * If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0. 1368 * dnode_hold_impl() will check if the requested dnode is already consumed 1369 * as an extra dnode slot by an large dnode, in which case it returns 1370 * ENOENT. 1371 * 1372 * If the DNODE_DRY_RUN flag is set, we don't actually hold the dnode, just 1373 * return whether the hold would succeed or not. tag and dnp should set to 1374 * NULL in this case. 1375 * 1376 * errors: 1377 * EINVAL - Invalid object number or flags. 1378 * ENOSPC - Hole too small to fulfill "slots" request (DNODE_MUST_BE_FREE) 1379 * EEXIST - Refers to an allocated dnode (DNODE_MUST_BE_FREE) 1380 * - Refers to a freeing dnode (DNODE_MUST_BE_FREE) 1381 * - Refers to an interior dnode slot (DNODE_MUST_BE_ALLOCATED) 1382 * ENOENT - The requested dnode is not allocated (DNODE_MUST_BE_ALLOCATED) 1383 * - The requested dnode is being freed (DNODE_MUST_BE_ALLOCATED) 1384 * EIO - I/O error when reading the meta dnode dbuf. 1385 * 1386 * succeeds even for free dnodes. 1387 */ 1388 int 1389 dnode_hold_impl(objset_t *os, uint64_t object, int flag, int slots, 1390 const void *tag, dnode_t **dnp) 1391 { 1392 int epb, idx, err; 1393 int drop_struct_lock = FALSE; 1394 int type; 1395 uint64_t blk; 1396 dnode_t *mdn, *dn; 1397 dmu_buf_impl_t *db; 1398 dnode_children_t *dnc; 1399 dnode_phys_t *dn_block; 1400 dnode_handle_t *dnh; 1401 1402 ASSERT(!(flag & DNODE_MUST_BE_ALLOCATED) || (slots == 0)); 1403 ASSERT(!(flag & DNODE_MUST_BE_FREE) || (slots > 0)); 1404 IMPLY(flag & DNODE_DRY_RUN, (tag == NULL) && (dnp == NULL)); 1405 1406 /* 1407 * If you are holding the spa config lock as writer, you shouldn't 1408 * be asking the DMU to do *anything* unless it's the root pool 1409 * which may require us to read from the root filesystem while 1410 * holding some (not all) of the locks as writer. 1411 */ 1412 ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 || 1413 (spa_is_root(os->os_spa) && 1414 spa_config_held(os->os_spa, SCL_STATE, RW_WRITER))); 1415 1416 ASSERT((flag & DNODE_MUST_BE_ALLOCATED) || (flag & DNODE_MUST_BE_FREE)); 1417 1418 if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT || 1419 object == DMU_PROJECTUSED_OBJECT) { 1420 if (object == DMU_USERUSED_OBJECT) 1421 dn = DMU_USERUSED_DNODE(os); 1422 else if (object == DMU_GROUPUSED_OBJECT) 1423 dn = DMU_GROUPUSED_DNODE(os); 1424 else 1425 dn = DMU_PROJECTUSED_DNODE(os); 1426 if (dn == NULL) 1427 return (SET_ERROR(ENOENT)); 1428 type = dn->dn_type; 1429 if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE) 1430 return (SET_ERROR(ENOENT)); 1431 if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE) 1432 return (SET_ERROR(EEXIST)); 1433 DNODE_VERIFY(dn); 1434 /* Don't actually hold if dry run, just return 0 */ 1435 if (!(flag & DNODE_DRY_RUN)) { 1436 (void) zfs_refcount_add(&dn->dn_holds, tag); 1437 *dnp = dn; 1438 } 1439 return (0); 1440 } 1441 1442 if (object == 0 || object >= DN_MAX_OBJECT) 1443 return (SET_ERROR(EINVAL)); 1444 1445 mdn = DMU_META_DNODE(os); 1446 ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT); 1447 1448 DNODE_VERIFY(mdn); 1449 1450 if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) { 1451 rw_enter(&mdn->dn_struct_rwlock, RW_READER); 1452 drop_struct_lock = TRUE; 1453 } 1454 1455 blk = dbuf_whichblock(mdn, 0, object * sizeof (dnode_phys_t)); 1456 db = dbuf_hold(mdn, blk, FTAG); 1457 if (drop_struct_lock) 1458 rw_exit(&mdn->dn_struct_rwlock); 1459 if (db == NULL) { 1460 DNODE_STAT_BUMP(dnode_hold_dbuf_hold); 1461 return (SET_ERROR(EIO)); 1462 } 1463 1464 /* 1465 * We do not need to decrypt to read the dnode so it doesn't matter 1466 * if we get the encrypted or decrypted version. 1467 */ 1468 err = dbuf_read(db, NULL, DB_RF_CANFAIL | 1469 DB_RF_NO_DECRYPT | DB_RF_NOPREFETCH); 1470 if (err) { 1471 DNODE_STAT_BUMP(dnode_hold_dbuf_read); 1472 dbuf_rele(db, FTAG); 1473 return (err); 1474 } 1475 1476 ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT); 1477 epb = db->db.db_size >> DNODE_SHIFT; 1478 1479 idx = object & (epb - 1); 1480 dn_block = (dnode_phys_t *)db->db.db_data; 1481 1482 ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE); 1483 dnc = dmu_buf_get_user(&db->db); 1484 dnh = NULL; 1485 if (dnc == NULL) { 1486 dnode_children_t *winner; 1487 int skip = 0; 1488 1489 dnc = kmem_zalloc(sizeof (dnode_children_t) + 1490 epb * sizeof (dnode_handle_t), KM_SLEEP); 1491 dnc->dnc_count = epb; 1492 dnh = &dnc->dnc_children[0]; 1493 1494 /* Initialize dnode slot status from dnode_phys_t */ 1495 for (int i = 0; i < epb; i++) { 1496 zrl_init(&dnh[i].dnh_zrlock); 1497 1498 if (skip) { 1499 skip--; 1500 continue; 1501 } 1502 1503 if (dn_block[i].dn_type != DMU_OT_NONE) { 1504 int interior = dn_block[i].dn_extra_slots; 1505 1506 dnode_set_slots(dnc, i, 1, DN_SLOT_ALLOCATED); 1507 dnode_set_slots(dnc, i + 1, interior, 1508 DN_SLOT_INTERIOR); 1509 skip = interior; 1510 } else { 1511 dnh[i].dnh_dnode = DN_SLOT_FREE; 1512 skip = 0; 1513 } 1514 } 1515 1516 dmu_buf_init_user(&dnc->dnc_dbu, NULL, 1517 dnode_buf_evict_async, NULL); 1518 winner = dmu_buf_set_user(&db->db, &dnc->dnc_dbu); 1519 if (winner != NULL) { 1520 1521 for (int i = 0; i < epb; i++) 1522 zrl_destroy(&dnh[i].dnh_zrlock); 1523 1524 kmem_free(dnc, sizeof (dnode_children_t) + 1525 epb * sizeof (dnode_handle_t)); 1526 dnc = winner; 1527 } 1528 } 1529 1530 ASSERT(dnc->dnc_count == epb); 1531 1532 if (flag & DNODE_MUST_BE_ALLOCATED) { 1533 slots = 1; 1534 1535 dnode_slots_hold(dnc, idx, slots); 1536 dnh = &dnc->dnc_children[idx]; 1537 1538 if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) { 1539 dn = dnh->dnh_dnode; 1540 } else if (dnh->dnh_dnode == DN_SLOT_INTERIOR) { 1541 DNODE_STAT_BUMP(dnode_hold_alloc_interior); 1542 dnode_slots_rele(dnc, idx, slots); 1543 dbuf_rele(db, FTAG); 1544 return (SET_ERROR(EEXIST)); 1545 } else if (dnh->dnh_dnode != DN_SLOT_ALLOCATED) { 1546 DNODE_STAT_BUMP(dnode_hold_alloc_misses); 1547 dnode_slots_rele(dnc, idx, slots); 1548 dbuf_rele(db, FTAG); 1549 return (SET_ERROR(ENOENT)); 1550 } else { 1551 dnode_slots_rele(dnc, idx, slots); 1552 while (!dnode_slots_tryenter(dnc, idx, slots)) { 1553 DNODE_STAT_BUMP(dnode_hold_alloc_lock_retry); 1554 kpreempt(KPREEMPT_SYNC); 1555 } 1556 1557 /* 1558 * Someone else won the race and called dnode_create() 1559 * after we checked DN_SLOT_IS_PTR() above but before 1560 * we acquired the lock. 1561 */ 1562 if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) { 1563 DNODE_STAT_BUMP(dnode_hold_alloc_lock_misses); 1564 dn = dnh->dnh_dnode; 1565 } else { 1566 dn = dnode_create(os, dn_block + idx, db, 1567 object, dnh); 1568 } 1569 } 1570 1571 mutex_enter(&dn->dn_mtx); 1572 if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg != 0) { 1573 DNODE_STAT_BUMP(dnode_hold_alloc_type_none); 1574 mutex_exit(&dn->dn_mtx); 1575 dnode_slots_rele(dnc, idx, slots); 1576 dbuf_rele(db, FTAG); 1577 return (SET_ERROR(ENOENT)); 1578 } 1579 1580 /* Don't actually hold if dry run, just return 0 */ 1581 if (flag & DNODE_DRY_RUN) { 1582 mutex_exit(&dn->dn_mtx); 1583 dnode_slots_rele(dnc, idx, slots); 1584 dbuf_rele(db, FTAG); 1585 return (0); 1586 } 1587 1588 DNODE_STAT_BUMP(dnode_hold_alloc_hits); 1589 } else if (flag & DNODE_MUST_BE_FREE) { 1590 1591 if (idx + slots - 1 >= DNODES_PER_BLOCK) { 1592 DNODE_STAT_BUMP(dnode_hold_free_overflow); 1593 dbuf_rele(db, FTAG); 1594 return (SET_ERROR(ENOSPC)); 1595 } 1596 1597 dnode_slots_hold(dnc, idx, slots); 1598 1599 if (!dnode_check_slots_free(dnc, idx, slots)) { 1600 DNODE_STAT_BUMP(dnode_hold_free_misses); 1601 dnode_slots_rele(dnc, idx, slots); 1602 dbuf_rele(db, FTAG); 1603 return (SET_ERROR(ENOSPC)); 1604 } 1605 1606 dnode_slots_rele(dnc, idx, slots); 1607 while (!dnode_slots_tryenter(dnc, idx, slots)) { 1608 DNODE_STAT_BUMP(dnode_hold_free_lock_retry); 1609 kpreempt(KPREEMPT_SYNC); 1610 } 1611 1612 if (!dnode_check_slots_free(dnc, idx, slots)) { 1613 DNODE_STAT_BUMP(dnode_hold_free_lock_misses); 1614 dnode_slots_rele(dnc, idx, slots); 1615 dbuf_rele(db, FTAG); 1616 return (SET_ERROR(ENOSPC)); 1617 } 1618 1619 /* 1620 * Allocated but otherwise free dnodes which would 1621 * be in the interior of a multi-slot dnodes need 1622 * to be freed. Single slot dnodes can be safely 1623 * re-purposed as a performance optimization. 1624 */ 1625 if (slots > 1) 1626 dnode_reclaim_slots(dnc, idx + 1, slots - 1); 1627 1628 dnh = &dnc->dnc_children[idx]; 1629 if (DN_SLOT_IS_PTR(dnh->dnh_dnode)) { 1630 dn = dnh->dnh_dnode; 1631 } else { 1632 dn = dnode_create(os, dn_block + idx, db, 1633 object, dnh); 1634 } 1635 1636 mutex_enter(&dn->dn_mtx); 1637 if (!zfs_refcount_is_zero(&dn->dn_holds) || dn->dn_free_txg) { 1638 DNODE_STAT_BUMP(dnode_hold_free_refcount); 1639 mutex_exit(&dn->dn_mtx); 1640 dnode_slots_rele(dnc, idx, slots); 1641 dbuf_rele(db, FTAG); 1642 return (SET_ERROR(EEXIST)); 1643 } 1644 1645 /* Don't actually hold if dry run, just return 0 */ 1646 if (flag & DNODE_DRY_RUN) { 1647 mutex_exit(&dn->dn_mtx); 1648 dnode_slots_rele(dnc, idx, slots); 1649 dbuf_rele(db, FTAG); 1650 return (0); 1651 } 1652 1653 dnode_set_slots(dnc, idx + 1, slots - 1, DN_SLOT_INTERIOR); 1654 DNODE_STAT_BUMP(dnode_hold_free_hits); 1655 } else { 1656 dbuf_rele(db, FTAG); 1657 return (SET_ERROR(EINVAL)); 1658 } 1659 1660 ASSERT0(dn->dn_free_txg); 1661 1662 if (zfs_refcount_add(&dn->dn_holds, tag) == 1) 1663 dbuf_add_ref(db, dnh); 1664 1665 mutex_exit(&dn->dn_mtx); 1666 1667 /* Now we can rely on the hold to prevent the dnode from moving. */ 1668 dnode_slots_rele(dnc, idx, slots); 1669 1670 DNODE_VERIFY(dn); 1671 ASSERT3P(dnp, !=, NULL); 1672 ASSERT3P(dn->dn_dbuf, ==, db); 1673 ASSERT3U(dn->dn_object, ==, object); 1674 dbuf_rele(db, FTAG); 1675 1676 *dnp = dn; 1677 return (0); 1678 } 1679 1680 /* 1681 * Return held dnode if the object is allocated, NULL if not. 1682 */ 1683 int 1684 dnode_hold(objset_t *os, uint64_t object, const void *tag, dnode_t **dnp) 1685 { 1686 return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, 0, tag, 1687 dnp)); 1688 } 1689 1690 /* 1691 * Can only add a reference if there is already at least one 1692 * reference on the dnode. Returns FALSE if unable to add a 1693 * new reference. 1694 */ 1695 boolean_t 1696 dnode_add_ref(dnode_t *dn, const void *tag) 1697 { 1698 mutex_enter(&dn->dn_mtx); 1699 if (zfs_refcount_is_zero(&dn->dn_holds)) { 1700 mutex_exit(&dn->dn_mtx); 1701 return (FALSE); 1702 } 1703 VERIFY(1 < zfs_refcount_add(&dn->dn_holds, tag)); 1704 mutex_exit(&dn->dn_mtx); 1705 return (TRUE); 1706 } 1707 1708 void 1709 dnode_rele(dnode_t *dn, const void *tag) 1710 { 1711 mutex_enter(&dn->dn_mtx); 1712 dnode_rele_and_unlock(dn, tag, B_FALSE); 1713 } 1714 1715 void 1716 dnode_rele_and_unlock(dnode_t *dn, const void *tag, boolean_t evicting) 1717 { 1718 uint64_t refs; 1719 /* Get while the hold prevents the dnode from moving. */ 1720 dmu_buf_impl_t *db = dn->dn_dbuf; 1721 dnode_handle_t *dnh = dn->dn_handle; 1722 1723 refs = zfs_refcount_remove(&dn->dn_holds, tag); 1724 if (refs == 0) 1725 cv_broadcast(&dn->dn_nodnholds); 1726 mutex_exit(&dn->dn_mtx); 1727 /* dnode could get destroyed at this point, so don't use it anymore */ 1728 1729 /* 1730 * It's unsafe to release the last hold on a dnode by dnode_rele() or 1731 * indirectly by dbuf_rele() while relying on the dnode handle to 1732 * prevent the dnode from moving, since releasing the last hold could 1733 * result in the dnode's parent dbuf evicting its dnode handles. For 1734 * that reason anyone calling dnode_rele() or dbuf_rele() without some 1735 * other direct or indirect hold on the dnode must first drop the dnode 1736 * handle. 1737 */ 1738 #ifdef ZFS_DEBUG 1739 ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread); 1740 #endif 1741 1742 /* NOTE: the DNODE_DNODE does not have a dn_dbuf */ 1743 if (refs == 0 && db != NULL) { 1744 /* 1745 * Another thread could add a hold to the dnode handle in 1746 * dnode_hold_impl() while holding the parent dbuf. Since the 1747 * hold on the parent dbuf prevents the handle from being 1748 * destroyed, the hold on the handle is OK. We can't yet assert 1749 * that the handle has zero references, but that will be 1750 * asserted anyway when the handle gets destroyed. 1751 */ 1752 mutex_enter(&db->db_mtx); 1753 dbuf_rele_and_unlock(db, dnh, evicting); 1754 } 1755 } 1756 1757 /* 1758 * Test whether we can create a dnode at the specified location. 1759 */ 1760 int 1761 dnode_try_claim(objset_t *os, uint64_t object, int slots) 1762 { 1763 return (dnode_hold_impl(os, object, DNODE_MUST_BE_FREE | DNODE_DRY_RUN, 1764 slots, NULL, NULL)); 1765 } 1766 1767 /* 1768 * Checks if the dnode contains any uncommitted dirty records. 1769 */ 1770 boolean_t 1771 dnode_is_dirty(dnode_t *dn) 1772 { 1773 mutex_enter(&dn->dn_mtx); 1774 1775 for (int i = 0; i < TXG_SIZE; i++) { 1776 if (multilist_link_active(&dn->dn_dirty_link[i])) { 1777 mutex_exit(&dn->dn_mtx); 1778 return (B_TRUE); 1779 } 1780 } 1781 1782 mutex_exit(&dn->dn_mtx); 1783 1784 return (B_FALSE); 1785 } 1786 1787 void 1788 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx) 1789 { 1790 objset_t *os = dn->dn_objset; 1791 uint64_t txg = tx->tx_txg; 1792 1793 if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) { 1794 dsl_dataset_dirty(os->os_dsl_dataset, tx); 1795 return; 1796 } 1797 1798 DNODE_VERIFY(dn); 1799 1800 #ifdef ZFS_DEBUG 1801 mutex_enter(&dn->dn_mtx); 1802 ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg); 1803 ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg); 1804 mutex_exit(&dn->dn_mtx); 1805 #endif 1806 1807 /* 1808 * Determine old uid/gid when necessary 1809 */ 1810 dmu_objset_userquota_get_ids(dn, B_TRUE, tx); 1811 1812 multilist_t *dirtylist = &os->os_dirty_dnodes[txg & TXG_MASK]; 1813 multilist_sublist_t *mls = multilist_sublist_lock_obj(dirtylist, dn); 1814 1815 /* 1816 * If we are already marked dirty, we're done. 1817 */ 1818 if (multilist_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) { 1819 multilist_sublist_unlock(mls); 1820 return; 1821 } 1822 1823 ASSERT(!zfs_refcount_is_zero(&dn->dn_holds) || 1824 !avl_is_empty(&dn->dn_dbufs)); 1825 ASSERT(dn->dn_datablksz != 0); 1826 ASSERT0(dn->dn_next_bonuslen[txg & TXG_MASK]); 1827 ASSERT0(dn->dn_next_blksz[txg & TXG_MASK]); 1828 ASSERT0(dn->dn_next_bonustype[txg & TXG_MASK]); 1829 1830 dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n", 1831 (u_longlong_t)dn->dn_object, (u_longlong_t)txg); 1832 1833 multilist_sublist_insert_head(mls, dn); 1834 1835 multilist_sublist_unlock(mls); 1836 1837 /* 1838 * The dnode maintains a hold on its containing dbuf as 1839 * long as there are holds on it. Each instantiated child 1840 * dbuf maintains a hold on the dnode. When the last child 1841 * drops its hold, the dnode will drop its hold on the 1842 * containing dbuf. We add a "dirty hold" here so that the 1843 * dnode will hang around after we finish processing its 1844 * children. 1845 */ 1846 VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg)); 1847 1848 (void) dbuf_dirty(dn->dn_dbuf, tx); 1849 1850 dsl_dataset_dirty(os->os_dsl_dataset, tx); 1851 } 1852 1853 void 1854 dnode_free(dnode_t *dn, dmu_tx_t *tx) 1855 { 1856 mutex_enter(&dn->dn_mtx); 1857 if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) { 1858 mutex_exit(&dn->dn_mtx); 1859 return; 1860 } 1861 dn->dn_free_txg = tx->tx_txg; 1862 mutex_exit(&dn->dn_mtx); 1863 1864 dnode_setdirty(dn, tx); 1865 } 1866 1867 /* 1868 * Try to change the block size for the indicated dnode. This can only 1869 * succeed if there are no blocks allocated or dirty beyond first block 1870 */ 1871 int 1872 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx) 1873 { 1874 dmu_buf_impl_t *db; 1875 int err; 1876 1877 ASSERT3U(size, <=, spa_maxblocksize(dmu_objset_spa(dn->dn_objset))); 1878 if (size == 0) 1879 size = SPA_MINBLOCKSIZE; 1880 else 1881 size = P2ROUNDUP(size, SPA_MINBLOCKSIZE); 1882 1883 if (ibs == dn->dn_indblkshift) 1884 ibs = 0; 1885 1886 if (size == dn->dn_datablksz && ibs == 0) 1887 return (0); 1888 1889 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 1890 1891 /* Check for any allocated blocks beyond the first */ 1892 if (dn->dn_maxblkid != 0) 1893 goto fail; 1894 1895 mutex_enter(&dn->dn_dbufs_mtx); 1896 for (db = avl_first(&dn->dn_dbufs); db != NULL; 1897 db = AVL_NEXT(&dn->dn_dbufs, db)) { 1898 if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID && 1899 db->db_blkid != DMU_SPILL_BLKID) { 1900 mutex_exit(&dn->dn_dbufs_mtx); 1901 goto fail; 1902 } 1903 } 1904 mutex_exit(&dn->dn_dbufs_mtx); 1905 1906 if (ibs && dn->dn_nlevels != 1) 1907 goto fail; 1908 1909 dnode_setdirty(dn, tx); 1910 if (size != dn->dn_datablksz) { 1911 /* resize the old block */ 1912 err = dbuf_hold_impl(dn, 0, 0, TRUE, FALSE, FTAG, &db); 1913 if (err == 0) { 1914 dbuf_new_size(db, size, tx); 1915 } else if (err != ENOENT) { 1916 goto fail; 1917 } 1918 1919 dnode_setdblksz(dn, size); 1920 dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = size; 1921 if (db) 1922 dbuf_rele(db, FTAG); 1923 } 1924 if (ibs) { 1925 dn->dn_indblkshift = ibs; 1926 dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs; 1927 } 1928 1929 rw_exit(&dn->dn_struct_rwlock); 1930 return (0); 1931 1932 fail: 1933 rw_exit(&dn->dn_struct_rwlock); 1934 return (SET_ERROR(ENOTSUP)); 1935 } 1936 1937 static void 1938 dnode_set_nlevels_impl(dnode_t *dn, int new_nlevels, dmu_tx_t *tx) 1939 { 1940 uint64_t txgoff = tx->tx_txg & TXG_MASK; 1941 int old_nlevels = dn->dn_nlevels; 1942 dmu_buf_impl_t *db; 1943 list_t *list; 1944 dbuf_dirty_record_t *new, *dr, *dr_next; 1945 1946 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock)); 1947 1948 ASSERT3U(new_nlevels, >, dn->dn_nlevels); 1949 dn->dn_nlevels = new_nlevels; 1950 1951 ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]); 1952 dn->dn_next_nlevels[txgoff] = new_nlevels; 1953 1954 /* dirty the left indirects */ 1955 db = dbuf_hold_level(dn, old_nlevels, 0, FTAG); 1956 ASSERT(db != NULL); 1957 new = dbuf_dirty(db, tx); 1958 dbuf_rele(db, FTAG); 1959 1960 /* transfer the dirty records to the new indirect */ 1961 mutex_enter(&dn->dn_mtx); 1962 mutex_enter(&new->dt.di.dr_mtx); 1963 list = &dn->dn_dirty_records[txgoff]; 1964 for (dr = list_head(list); dr; dr = dr_next) { 1965 dr_next = list_next(&dn->dn_dirty_records[txgoff], dr); 1966 1967 IMPLY(dr->dr_dbuf == NULL, old_nlevels == 1); 1968 if (dr->dr_dbuf == NULL || 1969 (dr->dr_dbuf->db_level == old_nlevels - 1 && 1970 dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID && 1971 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID)) { 1972 list_remove(&dn->dn_dirty_records[txgoff], dr); 1973 list_insert_tail(&new->dt.di.dr_children, dr); 1974 dr->dr_parent = new; 1975 } 1976 } 1977 mutex_exit(&new->dt.di.dr_mtx); 1978 mutex_exit(&dn->dn_mtx); 1979 } 1980 1981 int 1982 dnode_set_nlevels(dnode_t *dn, int nlevels, dmu_tx_t *tx) 1983 { 1984 int ret = 0; 1985 1986 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 1987 1988 if (dn->dn_nlevels == nlevels) { 1989 ret = 0; 1990 goto out; 1991 } else if (nlevels < dn->dn_nlevels) { 1992 ret = SET_ERROR(EINVAL); 1993 goto out; 1994 } 1995 1996 dnode_set_nlevels_impl(dn, nlevels, tx); 1997 1998 out: 1999 rw_exit(&dn->dn_struct_rwlock); 2000 return (ret); 2001 } 2002 2003 /* read-holding callers must not rely on the lock being continuously held */ 2004 void 2005 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read, 2006 boolean_t force) 2007 { 2008 int epbs, new_nlevels; 2009 uint64_t sz; 2010 2011 ASSERT(blkid != DMU_BONUS_BLKID); 2012 2013 ASSERT(have_read ? 2014 RW_READ_HELD(&dn->dn_struct_rwlock) : 2015 RW_WRITE_HELD(&dn->dn_struct_rwlock)); 2016 2017 /* 2018 * if we have a read-lock, check to see if we need to do any work 2019 * before upgrading to a write-lock. 2020 */ 2021 if (have_read) { 2022 if (blkid <= dn->dn_maxblkid) 2023 return; 2024 2025 if (!rw_tryupgrade(&dn->dn_struct_rwlock)) { 2026 rw_exit(&dn->dn_struct_rwlock); 2027 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 2028 } 2029 } 2030 2031 /* 2032 * Raw sends (indicated by the force flag) require that we take the 2033 * given blkid even if the value is lower than the current value. 2034 */ 2035 if (!force && blkid <= dn->dn_maxblkid) 2036 goto out; 2037 2038 /* 2039 * We use the (otherwise unused) top bit of dn_next_maxblkid[txgoff] 2040 * to indicate that this field is set. This allows us to set the 2041 * maxblkid to 0 on an existing object in dnode_sync(). 2042 */ 2043 dn->dn_maxblkid = blkid; 2044 dn->dn_next_maxblkid[tx->tx_txg & TXG_MASK] = 2045 blkid | DMU_NEXT_MAXBLKID_SET; 2046 2047 /* 2048 * Compute the number of levels necessary to support the new maxblkid. 2049 * Raw sends will ensure nlevels is set correctly for us. 2050 */ 2051 new_nlevels = 1; 2052 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 2053 for (sz = dn->dn_nblkptr; 2054 sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs) 2055 new_nlevels++; 2056 2057 ASSERT3U(new_nlevels, <=, DN_MAX_LEVELS); 2058 2059 if (!force) { 2060 if (new_nlevels > dn->dn_nlevels) 2061 dnode_set_nlevels_impl(dn, new_nlevels, tx); 2062 } else { 2063 ASSERT3U(dn->dn_nlevels, >=, new_nlevels); 2064 } 2065 2066 out: 2067 if (have_read) 2068 rw_downgrade(&dn->dn_struct_rwlock); 2069 } 2070 2071 static void 2072 dnode_dirty_l1(dnode_t *dn, uint64_t l1blkid, dmu_tx_t *tx) 2073 { 2074 dmu_buf_impl_t *db = dbuf_hold_level(dn, 1, l1blkid, FTAG); 2075 if (db != NULL) { 2076 dmu_buf_will_dirty(&db->db, tx); 2077 dbuf_rele(db, FTAG); 2078 } 2079 } 2080 2081 /* 2082 * Dirty all the in-core level-1 dbufs in the range specified by start_blkid 2083 * and end_blkid. 2084 */ 2085 static void 2086 dnode_dirty_l1range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid, 2087 dmu_tx_t *tx) 2088 { 2089 dmu_buf_impl_t *db_search; 2090 dmu_buf_impl_t *db; 2091 avl_index_t where; 2092 2093 db_search = kmem_zalloc(sizeof (dmu_buf_impl_t), KM_SLEEP); 2094 2095 mutex_enter(&dn->dn_dbufs_mtx); 2096 2097 db_search->db_level = 1; 2098 db_search->db_blkid = start_blkid + 1; 2099 db_search->db_state = DB_SEARCH; 2100 for (;;) { 2101 2102 db = avl_find(&dn->dn_dbufs, db_search, &where); 2103 if (db == NULL) 2104 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER); 2105 2106 if (db == NULL || db->db_level != 1 || 2107 db->db_blkid >= end_blkid) { 2108 break; 2109 } 2110 2111 /* 2112 * Setup the next blkid we want to search for. 2113 */ 2114 db_search->db_blkid = db->db_blkid + 1; 2115 ASSERT3U(db->db_blkid, >=, start_blkid); 2116 2117 /* 2118 * If the dbuf transitions to DB_EVICTING while we're trying 2119 * to dirty it, then we will be unable to discover it in 2120 * the dbuf hash table. This will result in a call to 2121 * dbuf_create() which needs to acquire the dn_dbufs_mtx 2122 * lock. To avoid a deadlock, we drop the lock before 2123 * dirtying the level-1 dbuf. 2124 */ 2125 mutex_exit(&dn->dn_dbufs_mtx); 2126 dnode_dirty_l1(dn, db->db_blkid, tx); 2127 mutex_enter(&dn->dn_dbufs_mtx); 2128 } 2129 2130 #ifdef ZFS_DEBUG 2131 /* 2132 * Walk all the in-core level-1 dbufs and verify they have been dirtied. 2133 */ 2134 db_search->db_level = 1; 2135 db_search->db_blkid = start_blkid + 1; 2136 db_search->db_state = DB_SEARCH; 2137 db = avl_find(&dn->dn_dbufs, db_search, &where); 2138 if (db == NULL) 2139 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER); 2140 for (; db != NULL; db = AVL_NEXT(&dn->dn_dbufs, db)) { 2141 if (db->db_level != 1 || db->db_blkid >= end_blkid) 2142 break; 2143 if (db->db_state != DB_EVICTING) 2144 ASSERT(db->db_dirtycnt > 0); 2145 } 2146 #endif 2147 kmem_free(db_search, sizeof (dmu_buf_impl_t)); 2148 mutex_exit(&dn->dn_dbufs_mtx); 2149 } 2150 2151 void 2152 dnode_set_dirtyctx(dnode_t *dn, dmu_tx_t *tx, const void *tag) 2153 { 2154 /* 2155 * Don't set dirtyctx to SYNC if we're just modifying this as we 2156 * initialize the objset. 2157 */ 2158 if (dn->dn_dirtyctx == DN_UNDIRTIED) { 2159 dsl_dataset_t *ds = dn->dn_objset->os_dsl_dataset; 2160 2161 if (ds != NULL) { 2162 rrw_enter(&ds->ds_bp_rwlock, RW_READER, tag); 2163 } 2164 if (!BP_IS_HOLE(dn->dn_objset->os_rootbp)) { 2165 if (dmu_tx_is_syncing(tx)) 2166 dn->dn_dirtyctx = DN_DIRTY_SYNC; 2167 else 2168 dn->dn_dirtyctx = DN_DIRTY_OPEN; 2169 dn->dn_dirtyctx_firstset = tag; 2170 } 2171 if (ds != NULL) { 2172 rrw_exit(&ds->ds_bp_rwlock, tag); 2173 } 2174 } 2175 } 2176 2177 static void 2178 dnode_partial_zero(dnode_t *dn, uint64_t off, uint64_t blkoff, uint64_t len, 2179 dmu_tx_t *tx) 2180 { 2181 dmu_buf_impl_t *db; 2182 int res; 2183 2184 rw_enter(&dn->dn_struct_rwlock, RW_READER); 2185 res = dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off), TRUE, FALSE, 2186 FTAG, &db); 2187 rw_exit(&dn->dn_struct_rwlock); 2188 if (res == 0) { 2189 db_lock_type_t dblt; 2190 boolean_t dirty; 2191 2192 dblt = dmu_buf_lock_parent(db, RW_READER, FTAG); 2193 /* don't dirty if not on disk and not dirty */ 2194 dirty = !list_is_empty(&db->db_dirty_records) || 2195 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr)); 2196 dmu_buf_unlock_parent(db, dblt, FTAG); 2197 if (dirty) { 2198 caddr_t data; 2199 2200 dmu_buf_will_dirty(&db->db, tx); 2201 data = db->db.db_data; 2202 memset(data + blkoff, 0, len); 2203 } 2204 dbuf_rele(db, FTAG); 2205 } 2206 } 2207 2208 void 2209 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx) 2210 { 2211 uint64_t blkoff, blkid, nblks; 2212 int blksz, blkshift, head, tail; 2213 int trunc = FALSE; 2214 int epbs; 2215 2216 blksz = dn->dn_datablksz; 2217 blkshift = dn->dn_datablkshift; 2218 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT; 2219 2220 if (len == DMU_OBJECT_END) { 2221 len = UINT64_MAX - off; 2222 trunc = TRUE; 2223 } 2224 2225 /* 2226 * First, block align the region to free: 2227 */ 2228 if (ISP2(blksz)) { 2229 head = P2NPHASE(off, blksz); 2230 blkoff = P2PHASE(off, blksz); 2231 if ((off >> blkshift) > dn->dn_maxblkid) 2232 return; 2233 } else { 2234 ASSERT(dn->dn_maxblkid == 0); 2235 if (off == 0 && len >= blksz) { 2236 /* 2237 * Freeing the whole block; fast-track this request. 2238 */ 2239 blkid = 0; 2240 nblks = 1; 2241 if (dn->dn_nlevels > 1) { 2242 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 2243 dnode_dirty_l1(dn, 0, tx); 2244 rw_exit(&dn->dn_struct_rwlock); 2245 } 2246 goto done; 2247 } else if (off >= blksz) { 2248 /* Freeing past end-of-data */ 2249 return; 2250 } else { 2251 /* Freeing part of the block. */ 2252 head = blksz - off; 2253 ASSERT3U(head, >, 0); 2254 } 2255 blkoff = off; 2256 } 2257 /* zero out any partial block data at the start of the range */ 2258 if (head) { 2259 ASSERT3U(blkoff + head, ==, blksz); 2260 if (len < head) 2261 head = len; 2262 dnode_partial_zero(dn, off, blkoff, head, tx); 2263 off += head; 2264 len -= head; 2265 } 2266 2267 /* If the range was less than one block, we're done */ 2268 if (len == 0) 2269 return; 2270 2271 /* If the remaining range is past end of file, we're done */ 2272 if ((off >> blkshift) > dn->dn_maxblkid) 2273 return; 2274 2275 ASSERT(ISP2(blksz)); 2276 if (trunc) 2277 tail = 0; 2278 else 2279 tail = P2PHASE(len, blksz); 2280 2281 ASSERT0(P2PHASE(off, blksz)); 2282 /* zero out any partial block data at the end of the range */ 2283 if (tail) { 2284 if (len < tail) 2285 tail = len; 2286 dnode_partial_zero(dn, off + len, 0, tail, tx); 2287 len -= tail; 2288 } 2289 2290 /* If the range did not include a full block, we are done */ 2291 if (len == 0) 2292 return; 2293 2294 ASSERT(IS_P2ALIGNED(off, blksz)); 2295 ASSERT(trunc || IS_P2ALIGNED(len, blksz)); 2296 blkid = off >> blkshift; 2297 nblks = len >> blkshift; 2298 if (trunc) 2299 nblks += 1; 2300 2301 /* 2302 * Dirty all the indirect blocks in this range. Note that only 2303 * the first and last indirect blocks can actually be written 2304 * (if they were partially freed) -- they must be dirtied, even if 2305 * they do not exist on disk yet. The interior blocks will 2306 * be freed by free_children(), so they will not actually be written. 2307 * Even though these interior blocks will not be written, we 2308 * dirty them for two reasons: 2309 * 2310 * - It ensures that the indirect blocks remain in memory until 2311 * syncing context. (They have already been prefetched by 2312 * dmu_tx_hold_free(), so we don't have to worry about reading 2313 * them serially here.) 2314 * 2315 * - The dirty space accounting will put pressure on the txg sync 2316 * mechanism to begin syncing, and to delay transactions if there 2317 * is a large amount of freeing. Even though these indirect 2318 * blocks will not be written, we could need to write the same 2319 * amount of space if we copy the freed BPs into deadlists. 2320 */ 2321 if (dn->dn_nlevels > 1) { 2322 rw_enter(&dn->dn_struct_rwlock, RW_WRITER); 2323 uint64_t first, last; 2324 2325 first = blkid >> epbs; 2326 dnode_dirty_l1(dn, first, tx); 2327 if (trunc) 2328 last = dn->dn_maxblkid >> epbs; 2329 else 2330 last = (blkid + nblks - 1) >> epbs; 2331 if (last != first) 2332 dnode_dirty_l1(dn, last, tx); 2333 2334 dnode_dirty_l1range(dn, first, last, tx); 2335 2336 int shift = dn->dn_datablkshift + dn->dn_indblkshift - 2337 SPA_BLKPTRSHIFT; 2338 for (uint64_t i = first + 1; i < last; i++) { 2339 /* 2340 * Set i to the blockid of the next non-hole 2341 * level-1 indirect block at or after i. Note 2342 * that dnode_next_offset() operates in terms of 2343 * level-0-equivalent bytes. 2344 */ 2345 uint64_t ibyte = i << shift; 2346 int err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK, 2347 &ibyte, 2, 1, 0); 2348 i = ibyte >> shift; 2349 if (i >= last) 2350 break; 2351 2352 /* 2353 * Normally we should not see an error, either 2354 * from dnode_next_offset() or dbuf_hold_level() 2355 * (except for ESRCH from dnode_next_offset). 2356 * If there is an i/o error, then when we read 2357 * this block in syncing context, it will use 2358 * ZIO_FLAG_MUSTSUCCEED, and thus hang/panic according 2359 * to the "failmode" property. dnode_next_offset() 2360 * doesn't have a flag to indicate MUSTSUCCEED. 2361 */ 2362 if (err != 0) 2363 break; 2364 2365 dnode_dirty_l1(dn, i, tx); 2366 } 2367 rw_exit(&dn->dn_struct_rwlock); 2368 } 2369 2370 done: 2371 /* 2372 * Add this range to the dnode range list. 2373 * We will finish up this free operation in the syncing phase. 2374 */ 2375 mutex_enter(&dn->dn_mtx); 2376 { 2377 int txgoff = tx->tx_txg & TXG_MASK; 2378 if (dn->dn_free_ranges[txgoff] == NULL) { 2379 dn->dn_free_ranges[txgoff] = range_tree_create(NULL, 2380 RANGE_SEG64, NULL, 0, 0); 2381 } 2382 range_tree_clear(dn->dn_free_ranges[txgoff], blkid, nblks); 2383 range_tree_add(dn->dn_free_ranges[txgoff], blkid, nblks); 2384 } 2385 dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n", 2386 (u_longlong_t)blkid, (u_longlong_t)nblks, 2387 (u_longlong_t)tx->tx_txg); 2388 mutex_exit(&dn->dn_mtx); 2389 2390 dbuf_free_range(dn, blkid, blkid + nblks - 1, tx); 2391 dnode_setdirty(dn, tx); 2392 } 2393 2394 static boolean_t 2395 dnode_spill_freed(dnode_t *dn) 2396 { 2397 int i; 2398 2399 mutex_enter(&dn->dn_mtx); 2400 for (i = 0; i < TXG_SIZE; i++) { 2401 if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK) 2402 break; 2403 } 2404 mutex_exit(&dn->dn_mtx); 2405 return (i < TXG_SIZE); 2406 } 2407 2408 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */ 2409 uint64_t 2410 dnode_block_freed(dnode_t *dn, uint64_t blkid) 2411 { 2412 int i; 2413 2414 if (blkid == DMU_BONUS_BLKID) 2415 return (FALSE); 2416 2417 if (dn->dn_free_txg) 2418 return (TRUE); 2419 2420 if (blkid == DMU_SPILL_BLKID) 2421 return (dnode_spill_freed(dn)); 2422 2423 mutex_enter(&dn->dn_mtx); 2424 for (i = 0; i < TXG_SIZE; i++) { 2425 if (dn->dn_free_ranges[i] != NULL && 2426 range_tree_contains(dn->dn_free_ranges[i], blkid, 1)) 2427 break; 2428 } 2429 mutex_exit(&dn->dn_mtx); 2430 return (i < TXG_SIZE); 2431 } 2432 2433 /* call from syncing context when we actually write/free space for this dnode */ 2434 void 2435 dnode_diduse_space(dnode_t *dn, int64_t delta) 2436 { 2437 uint64_t space; 2438 dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n", 2439 dn, dn->dn_phys, 2440 (u_longlong_t)dn->dn_phys->dn_used, 2441 (longlong_t)delta); 2442 2443 mutex_enter(&dn->dn_mtx); 2444 space = DN_USED_BYTES(dn->dn_phys); 2445 if (delta > 0) { 2446 ASSERT3U(space + delta, >=, space); /* no overflow */ 2447 } else { 2448 ASSERT3U(space, >=, -delta); /* no underflow */ 2449 } 2450 space += delta; 2451 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) { 2452 ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0); 2453 ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT)); 2454 dn->dn_phys->dn_used = space >> DEV_BSHIFT; 2455 } else { 2456 dn->dn_phys->dn_used = space; 2457 dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES; 2458 } 2459 mutex_exit(&dn->dn_mtx); 2460 } 2461 2462 /* 2463 * Scans a block at the indicated "level" looking for a hole or data, 2464 * depending on 'flags'. 2465 * 2466 * If level > 0, then we are scanning an indirect block looking at its 2467 * pointers. If level == 0, then we are looking at a block of dnodes. 2468 * 2469 * If we don't find what we are looking for in the block, we return ESRCH. 2470 * Otherwise, return with *offset pointing to the beginning (if searching 2471 * forwards) or end (if searching backwards) of the range covered by the 2472 * block pointer we matched on (or dnode). 2473 * 2474 * The basic search algorithm used below by dnode_next_offset() is to 2475 * use this function to search up the block tree (widen the search) until 2476 * we find something (i.e., we don't return ESRCH) and then search back 2477 * down the tree (narrow the search) until we reach our original search 2478 * level. 2479 */ 2480 static int 2481 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset, 2482 int lvl, uint64_t blkfill, uint64_t txg) 2483 { 2484 dmu_buf_impl_t *db = NULL; 2485 void *data = NULL; 2486 uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT; 2487 uint64_t epb = 1ULL << epbs; 2488 uint64_t minfill, maxfill; 2489 boolean_t hole; 2490 int i, inc, error, span; 2491 2492 ASSERT(RW_LOCK_HELD(&dn->dn_struct_rwlock)); 2493 2494 hole = ((flags & DNODE_FIND_HOLE) != 0); 2495 inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1; 2496 ASSERT(txg == 0 || !hole); 2497 2498 if (lvl == dn->dn_phys->dn_nlevels) { 2499 error = 0; 2500 epb = dn->dn_phys->dn_nblkptr; 2501 data = dn->dn_phys->dn_blkptr; 2502 } else { 2503 uint64_t blkid = dbuf_whichblock(dn, lvl, *offset); 2504 error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FALSE, FTAG, &db); 2505 if (error) { 2506 if (error != ENOENT) 2507 return (error); 2508 if (hole) 2509 return (0); 2510 /* 2511 * This can only happen when we are searching up 2512 * the block tree for data. We don't really need to 2513 * adjust the offset, as we will just end up looking 2514 * at the pointer to this block in its parent, and its 2515 * going to be unallocated, so we will skip over it. 2516 */ 2517 return (SET_ERROR(ESRCH)); 2518 } 2519 error = dbuf_read(db, NULL, 2520 DB_RF_CANFAIL | DB_RF_HAVESTRUCT | 2521 DB_RF_NO_DECRYPT | DB_RF_NOPREFETCH); 2522 if (error) { 2523 dbuf_rele(db, FTAG); 2524 return (error); 2525 } 2526 data = db->db.db_data; 2527 rw_enter(&db->db_rwlock, RW_READER); 2528 } 2529 2530 if (db != NULL && txg != 0 && (db->db_blkptr == NULL || 2531 db->db_blkptr->blk_birth <= txg || 2532 BP_IS_HOLE(db->db_blkptr))) { 2533 /* 2534 * This can only happen when we are searching up the tree 2535 * and these conditions mean that we need to keep climbing. 2536 */ 2537 error = SET_ERROR(ESRCH); 2538 } else if (lvl == 0) { 2539 dnode_phys_t *dnp = data; 2540 2541 ASSERT(dn->dn_type == DMU_OT_DNODE); 2542 ASSERT(!(flags & DNODE_FIND_BACKWARDS)); 2543 2544 for (i = (*offset >> DNODE_SHIFT) & (blkfill - 1); 2545 i < blkfill; i += dnp[i].dn_extra_slots + 1) { 2546 if ((dnp[i].dn_type == DMU_OT_NONE) == hole) 2547 break; 2548 } 2549 2550 if (i == blkfill) 2551 error = SET_ERROR(ESRCH); 2552 2553 *offset = (*offset & ~(DNODE_BLOCK_SIZE - 1)) + 2554 (i << DNODE_SHIFT); 2555 } else { 2556 blkptr_t *bp = data; 2557 uint64_t start = *offset; 2558 span = (lvl - 1) * epbs + dn->dn_datablkshift; 2559 minfill = 0; 2560 maxfill = blkfill << ((lvl - 1) * epbs); 2561 2562 if (hole) 2563 maxfill--; 2564 else 2565 minfill++; 2566 2567 if (span >= 8 * sizeof (*offset)) { 2568 /* This only happens on the highest indirection level */ 2569 ASSERT3U((lvl - 1), ==, dn->dn_phys->dn_nlevels - 1); 2570 *offset = 0; 2571 } else { 2572 *offset = *offset >> span; 2573 } 2574 2575 for (i = BF64_GET(*offset, 0, epbs); 2576 i >= 0 && i < epb; i += inc) { 2577 if (BP_GET_FILL(&bp[i]) >= minfill && 2578 BP_GET_FILL(&bp[i]) <= maxfill && 2579 (hole || bp[i].blk_birth > txg)) 2580 break; 2581 if (inc > 0 || *offset > 0) 2582 *offset += inc; 2583 } 2584 2585 if (span >= 8 * sizeof (*offset)) { 2586 *offset = start; 2587 } else { 2588 *offset = *offset << span; 2589 } 2590 2591 if (inc < 0) { 2592 /* traversing backwards; position offset at the end */ 2593 if (span < 8 * sizeof (*offset)) 2594 *offset = MIN(*offset + (1ULL << span) - 1, 2595 start); 2596 } else if (*offset < start) { 2597 *offset = start; 2598 } 2599 if (i < 0 || i >= epb) 2600 error = SET_ERROR(ESRCH); 2601 } 2602 2603 if (db != NULL) { 2604 rw_exit(&db->db_rwlock); 2605 dbuf_rele(db, FTAG); 2606 } 2607 2608 return (error); 2609 } 2610 2611 /* 2612 * Find the next hole, data, or sparse region at or after *offset. 2613 * The value 'blkfill' tells us how many items we expect to find 2614 * in an L0 data block; this value is 1 for normal objects, 2615 * DNODES_PER_BLOCK for the meta dnode, and some fraction of 2616 * DNODES_PER_BLOCK when searching for sparse regions thereof. 2617 * 2618 * Examples: 2619 * 2620 * dnode_next_offset(dn, flags, offset, 1, 1, 0); 2621 * Finds the next/previous hole/data in a file. 2622 * Used in dmu_offset_next(). 2623 * 2624 * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg); 2625 * Finds the next free/allocated dnode an objset's meta-dnode. 2626 * Only finds objects that have new contents since txg (ie. 2627 * bonus buffer changes and content removal are ignored). 2628 * Used in dmu_object_next(). 2629 * 2630 * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0); 2631 * Finds the next L2 meta-dnode bp that's at most 1/4 full. 2632 * Used in dmu_object_alloc(). 2633 */ 2634 int 2635 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset, 2636 int minlvl, uint64_t blkfill, uint64_t txg) 2637 { 2638 uint64_t initial_offset = *offset; 2639 int lvl, maxlvl; 2640 int error = 0; 2641 2642 if (!(flags & DNODE_FIND_HAVELOCK)) 2643 rw_enter(&dn->dn_struct_rwlock, RW_READER); 2644 2645 if (dn->dn_phys->dn_nlevels == 0) { 2646 error = SET_ERROR(ESRCH); 2647 goto out; 2648 } 2649 2650 if (dn->dn_datablkshift == 0) { 2651 if (*offset < dn->dn_datablksz) { 2652 if (flags & DNODE_FIND_HOLE) 2653 *offset = dn->dn_datablksz; 2654 } else { 2655 error = SET_ERROR(ESRCH); 2656 } 2657 goto out; 2658 } 2659 2660 maxlvl = dn->dn_phys->dn_nlevels; 2661 2662 for (lvl = minlvl; lvl <= maxlvl; lvl++) { 2663 error = dnode_next_offset_level(dn, 2664 flags, offset, lvl, blkfill, txg); 2665 if (error != ESRCH) 2666 break; 2667 } 2668 2669 while (error == 0 && --lvl >= minlvl) { 2670 error = dnode_next_offset_level(dn, 2671 flags, offset, lvl, blkfill, txg); 2672 } 2673 2674 /* 2675 * There's always a "virtual hole" at the end of the object, even 2676 * if all BP's which physically exist are non-holes. 2677 */ 2678 if ((flags & DNODE_FIND_HOLE) && error == ESRCH && txg == 0 && 2679 minlvl == 1 && blkfill == 1 && !(flags & DNODE_FIND_BACKWARDS)) { 2680 error = 0; 2681 } 2682 2683 if (error == 0 && (flags & DNODE_FIND_BACKWARDS ? 2684 initial_offset < *offset : initial_offset > *offset)) 2685 error = SET_ERROR(ESRCH); 2686 out: 2687 if (!(flags & DNODE_FIND_HAVELOCK)) 2688 rw_exit(&dn->dn_struct_rwlock); 2689 2690 return (error); 2691 } 2692 2693 #if defined(_KERNEL) 2694 EXPORT_SYMBOL(dnode_hold); 2695 EXPORT_SYMBOL(dnode_rele); 2696 EXPORT_SYMBOL(dnode_set_nlevels); 2697 EXPORT_SYMBOL(dnode_set_blksz); 2698 EXPORT_SYMBOL(dnode_free_range); 2699 EXPORT_SYMBOL(dnode_evict_dbufs); 2700 EXPORT_SYMBOL(dnode_evict_bonus); 2701 #endif 2702 2703 ZFS_MODULE_PARAM(zfs, zfs_, default_bs, INT, ZMOD_RW, 2704 "Default dnode block shift"); 2705 ZFS_MODULE_PARAM(zfs, zfs_, default_ibs, INT, ZMOD_RW, 2706 "Default dnode indirect block shift"); 2707