1 // SPDX-License-Identifier: GPL-2.0 2 /* Copyright (c) 2019 Facebook */ 3 #include <linux/rculist.h> 4 #include <linux/list.h> 5 #include <linux/hash.h> 6 #include <linux/types.h> 7 #include <linux/spinlock.h> 8 #include <linux/bpf.h> 9 #include <linux/btf_ids.h> 10 #include <linux/bpf_local_storage.h> 11 #include <net/sock.h> 12 #include <uapi/linux/sock_diag.h> 13 #include <uapi/linux/btf.h> 14 #include <linux/rcupdate.h> 15 #include <linux/rcupdate_trace.h> 16 #include <linux/rcupdate_wait.h> 17 18 #define BPF_LOCAL_STORAGE_CREATE_FLAG_MASK (BPF_F_NO_PREALLOC | BPF_F_CLONE) 19 20 static struct bpf_local_storage_map_bucket * 21 select_bucket(struct bpf_local_storage_map *smap, 22 struct bpf_local_storage_elem *selem) 23 { 24 return &smap->buckets[hash_ptr(selem, smap->bucket_log)]; 25 } 26 27 static int mem_charge(struct bpf_local_storage_map *smap, void *owner, u32 size) 28 { 29 struct bpf_map *map = &smap->map; 30 31 if (!map->ops->map_local_storage_charge) 32 return 0; 33 34 return map->ops->map_local_storage_charge(smap, owner, size); 35 } 36 37 static void mem_uncharge(struct bpf_local_storage_map *smap, void *owner, 38 u32 size) 39 { 40 struct bpf_map *map = &smap->map; 41 42 if (map->ops->map_local_storage_uncharge) 43 map->ops->map_local_storage_uncharge(smap, owner, size); 44 } 45 46 static struct bpf_local_storage __rcu ** 47 owner_storage(struct bpf_local_storage_map *smap, void *owner) 48 { 49 struct bpf_map *map = &smap->map; 50 51 return map->ops->map_owner_storage_ptr(owner); 52 } 53 54 static bool selem_linked_to_storage(const struct bpf_local_storage_elem *selem) 55 { 56 return !hlist_unhashed(&selem->snode); 57 } 58 59 static bool selem_linked_to_map(const struct bpf_local_storage_elem *selem) 60 { 61 return !hlist_unhashed(&selem->map_node); 62 } 63 64 struct bpf_local_storage_elem * 65 bpf_selem_alloc(struct bpf_local_storage_map *smap, void *owner, 66 void *value, bool charge_mem, gfp_t gfp_flags) 67 { 68 struct bpf_local_storage_elem *selem; 69 70 if (charge_mem && mem_charge(smap, owner, smap->elem_size)) 71 return NULL; 72 73 selem = bpf_map_kzalloc(&smap->map, smap->elem_size, 74 gfp_flags | __GFP_NOWARN); 75 if (selem) { 76 if (value) 77 memcpy(SDATA(selem)->data, value, smap->map.value_size); 78 return selem; 79 } 80 81 if (charge_mem) 82 mem_uncharge(smap, owner, smap->elem_size); 83 84 return NULL; 85 } 86 87 void bpf_local_storage_free_rcu(struct rcu_head *rcu) 88 { 89 struct bpf_local_storage *local_storage; 90 91 local_storage = container_of(rcu, struct bpf_local_storage, rcu); 92 kfree_rcu(local_storage, rcu); 93 } 94 95 static void bpf_selem_free_rcu(struct rcu_head *rcu) 96 { 97 struct bpf_local_storage_elem *selem; 98 99 selem = container_of(rcu, struct bpf_local_storage_elem, rcu); 100 kfree_rcu(selem, rcu); 101 } 102 103 /* local_storage->lock must be held and selem->local_storage == local_storage. 104 * The caller must ensure selem->smap is still valid to be 105 * dereferenced for its smap->elem_size and smap->cache_idx. 106 */ 107 bool bpf_selem_unlink_storage_nolock(struct bpf_local_storage *local_storage, 108 struct bpf_local_storage_elem *selem, 109 bool uncharge_mem, bool use_trace_rcu) 110 { 111 struct bpf_local_storage_map *smap; 112 bool free_local_storage; 113 void *owner; 114 115 smap = rcu_dereference_check(SDATA(selem)->smap, bpf_rcu_lock_held()); 116 owner = local_storage->owner; 117 118 /* All uncharging on the owner must be done first. 119 * The owner may be freed once the last selem is unlinked 120 * from local_storage. 121 */ 122 if (uncharge_mem) 123 mem_uncharge(smap, owner, smap->elem_size); 124 125 free_local_storage = hlist_is_singular_node(&selem->snode, 126 &local_storage->list); 127 if (free_local_storage) { 128 mem_uncharge(smap, owner, sizeof(struct bpf_local_storage)); 129 local_storage->owner = NULL; 130 131 /* After this RCU_INIT, owner may be freed and cannot be used */ 132 RCU_INIT_POINTER(*owner_storage(smap, owner), NULL); 133 134 /* local_storage is not freed now. local_storage->lock is 135 * still held and raw_spin_unlock_bh(&local_storage->lock) 136 * will be done by the caller. 137 * 138 * Although the unlock will be done under 139 * rcu_read_lock(), it is more intuitive to 140 * read if the freeing of the storage is done 141 * after the raw_spin_unlock_bh(&local_storage->lock). 142 * 143 * Hence, a "bool free_local_storage" is returned 144 * to the caller which then calls then frees the storage after 145 * all the RCU grace periods have expired. 146 */ 147 } 148 hlist_del_init_rcu(&selem->snode); 149 if (rcu_access_pointer(local_storage->cache[smap->cache_idx]) == 150 SDATA(selem)) 151 RCU_INIT_POINTER(local_storage->cache[smap->cache_idx], NULL); 152 153 if (use_trace_rcu) 154 call_rcu_tasks_trace(&selem->rcu, bpf_selem_free_rcu); 155 else 156 kfree_rcu(selem, rcu); 157 158 return free_local_storage; 159 } 160 161 static void __bpf_selem_unlink_storage(struct bpf_local_storage_elem *selem, 162 bool use_trace_rcu) 163 { 164 struct bpf_local_storage *local_storage; 165 bool free_local_storage = false; 166 unsigned long flags; 167 168 if (unlikely(!selem_linked_to_storage(selem))) 169 /* selem has already been unlinked from sk */ 170 return; 171 172 local_storage = rcu_dereference_check(selem->local_storage, 173 bpf_rcu_lock_held()); 174 raw_spin_lock_irqsave(&local_storage->lock, flags); 175 if (likely(selem_linked_to_storage(selem))) 176 free_local_storage = bpf_selem_unlink_storage_nolock( 177 local_storage, selem, true, use_trace_rcu); 178 raw_spin_unlock_irqrestore(&local_storage->lock, flags); 179 180 if (free_local_storage) { 181 if (use_trace_rcu) 182 call_rcu_tasks_trace(&local_storage->rcu, 183 bpf_local_storage_free_rcu); 184 else 185 kfree_rcu(local_storage, rcu); 186 } 187 } 188 189 void bpf_selem_link_storage_nolock(struct bpf_local_storage *local_storage, 190 struct bpf_local_storage_elem *selem) 191 { 192 RCU_INIT_POINTER(selem->local_storage, local_storage); 193 hlist_add_head_rcu(&selem->snode, &local_storage->list); 194 } 195 196 void bpf_selem_unlink_map(struct bpf_local_storage_elem *selem) 197 { 198 struct bpf_local_storage_map *smap; 199 struct bpf_local_storage_map_bucket *b; 200 unsigned long flags; 201 202 if (unlikely(!selem_linked_to_map(selem))) 203 /* selem has already be unlinked from smap */ 204 return; 205 206 smap = rcu_dereference_check(SDATA(selem)->smap, bpf_rcu_lock_held()); 207 b = select_bucket(smap, selem); 208 raw_spin_lock_irqsave(&b->lock, flags); 209 if (likely(selem_linked_to_map(selem))) 210 hlist_del_init_rcu(&selem->map_node); 211 raw_spin_unlock_irqrestore(&b->lock, flags); 212 } 213 214 void bpf_selem_link_map(struct bpf_local_storage_map *smap, 215 struct bpf_local_storage_elem *selem) 216 { 217 struct bpf_local_storage_map_bucket *b = select_bucket(smap, selem); 218 unsigned long flags; 219 220 raw_spin_lock_irqsave(&b->lock, flags); 221 RCU_INIT_POINTER(SDATA(selem)->smap, smap); 222 hlist_add_head_rcu(&selem->map_node, &b->list); 223 raw_spin_unlock_irqrestore(&b->lock, flags); 224 } 225 226 void bpf_selem_unlink(struct bpf_local_storage_elem *selem, bool use_trace_rcu) 227 { 228 /* Always unlink from map before unlinking from local_storage 229 * because selem will be freed after successfully unlinked from 230 * the local_storage. 231 */ 232 bpf_selem_unlink_map(selem); 233 __bpf_selem_unlink_storage(selem, use_trace_rcu); 234 } 235 236 struct bpf_local_storage_data * 237 bpf_local_storage_lookup(struct bpf_local_storage *local_storage, 238 struct bpf_local_storage_map *smap, 239 bool cacheit_lockit) 240 { 241 struct bpf_local_storage_data *sdata; 242 struct bpf_local_storage_elem *selem; 243 244 /* Fast path (cache hit) */ 245 sdata = rcu_dereference_check(local_storage->cache[smap->cache_idx], 246 bpf_rcu_lock_held()); 247 if (sdata && rcu_access_pointer(sdata->smap) == smap) 248 return sdata; 249 250 /* Slow path (cache miss) */ 251 hlist_for_each_entry_rcu(selem, &local_storage->list, snode, 252 rcu_read_lock_trace_held()) 253 if (rcu_access_pointer(SDATA(selem)->smap) == smap) 254 break; 255 256 if (!selem) 257 return NULL; 258 259 sdata = SDATA(selem); 260 if (cacheit_lockit) { 261 unsigned long flags; 262 263 /* spinlock is needed to avoid racing with the 264 * parallel delete. Otherwise, publishing an already 265 * deleted sdata to the cache will become a use-after-free 266 * problem in the next bpf_local_storage_lookup(). 267 */ 268 raw_spin_lock_irqsave(&local_storage->lock, flags); 269 if (selem_linked_to_storage(selem)) 270 rcu_assign_pointer(local_storage->cache[smap->cache_idx], 271 sdata); 272 raw_spin_unlock_irqrestore(&local_storage->lock, flags); 273 } 274 275 return sdata; 276 } 277 278 static int check_flags(const struct bpf_local_storage_data *old_sdata, 279 u64 map_flags) 280 { 281 if (old_sdata && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST) 282 /* elem already exists */ 283 return -EEXIST; 284 285 if (!old_sdata && (map_flags & ~BPF_F_LOCK) == BPF_EXIST) 286 /* elem doesn't exist, cannot update it */ 287 return -ENOENT; 288 289 return 0; 290 } 291 292 int bpf_local_storage_alloc(void *owner, 293 struct bpf_local_storage_map *smap, 294 struct bpf_local_storage_elem *first_selem, 295 gfp_t gfp_flags) 296 { 297 struct bpf_local_storage *prev_storage, *storage; 298 struct bpf_local_storage **owner_storage_ptr; 299 int err; 300 301 err = mem_charge(smap, owner, sizeof(*storage)); 302 if (err) 303 return err; 304 305 storage = bpf_map_kzalloc(&smap->map, sizeof(*storage), 306 gfp_flags | __GFP_NOWARN); 307 if (!storage) { 308 err = -ENOMEM; 309 goto uncharge; 310 } 311 312 INIT_HLIST_HEAD(&storage->list); 313 raw_spin_lock_init(&storage->lock); 314 storage->owner = owner; 315 316 bpf_selem_link_storage_nolock(storage, first_selem); 317 bpf_selem_link_map(smap, first_selem); 318 319 owner_storage_ptr = 320 (struct bpf_local_storage **)owner_storage(smap, owner); 321 /* Publish storage to the owner. 322 * Instead of using any lock of the kernel object (i.e. owner), 323 * cmpxchg will work with any kernel object regardless what 324 * the running context is, bh, irq...etc. 325 * 326 * From now on, the owner->storage pointer (e.g. sk->sk_bpf_storage) 327 * is protected by the storage->lock. Hence, when freeing 328 * the owner->storage, the storage->lock must be held before 329 * setting owner->storage ptr to NULL. 330 */ 331 prev_storage = cmpxchg(owner_storage_ptr, NULL, storage); 332 if (unlikely(prev_storage)) { 333 bpf_selem_unlink_map(first_selem); 334 err = -EAGAIN; 335 goto uncharge; 336 337 /* Note that even first_selem was linked to smap's 338 * bucket->list, first_selem can be freed immediately 339 * (instead of kfree_rcu) because 340 * bpf_local_storage_map_free() does a 341 * synchronize_rcu_mult (waiting for both sleepable and 342 * normal programs) before walking the bucket->list. 343 * Hence, no one is accessing selem from the 344 * bucket->list under rcu_read_lock(). 345 */ 346 } 347 348 return 0; 349 350 uncharge: 351 kfree(storage); 352 mem_uncharge(smap, owner, sizeof(*storage)); 353 return err; 354 } 355 356 /* sk cannot be going away because it is linking new elem 357 * to sk->sk_bpf_storage. (i.e. sk->sk_refcnt cannot be 0). 358 * Otherwise, it will become a leak (and other memory issues 359 * during map destruction). 360 */ 361 struct bpf_local_storage_data * 362 bpf_local_storage_update(void *owner, struct bpf_local_storage_map *smap, 363 void *value, u64 map_flags, gfp_t gfp_flags) 364 { 365 struct bpf_local_storage_data *old_sdata = NULL; 366 struct bpf_local_storage_elem *selem = NULL; 367 struct bpf_local_storage *local_storage; 368 unsigned long flags; 369 int err; 370 371 /* BPF_EXIST and BPF_NOEXIST cannot be both set */ 372 if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST) || 373 /* BPF_F_LOCK can only be used in a value with spin_lock */ 374 unlikely((map_flags & BPF_F_LOCK) && 375 !map_value_has_spin_lock(&smap->map))) 376 return ERR_PTR(-EINVAL); 377 378 if (gfp_flags == GFP_KERNEL && (map_flags & ~BPF_F_LOCK) != BPF_NOEXIST) 379 return ERR_PTR(-EINVAL); 380 381 local_storage = rcu_dereference_check(*owner_storage(smap, owner), 382 bpf_rcu_lock_held()); 383 if (!local_storage || hlist_empty(&local_storage->list)) { 384 /* Very first elem for the owner */ 385 err = check_flags(NULL, map_flags); 386 if (err) 387 return ERR_PTR(err); 388 389 selem = bpf_selem_alloc(smap, owner, value, true, gfp_flags); 390 if (!selem) 391 return ERR_PTR(-ENOMEM); 392 393 err = bpf_local_storage_alloc(owner, smap, selem, gfp_flags); 394 if (err) { 395 kfree(selem); 396 mem_uncharge(smap, owner, smap->elem_size); 397 return ERR_PTR(err); 398 } 399 400 return SDATA(selem); 401 } 402 403 if ((map_flags & BPF_F_LOCK) && !(map_flags & BPF_NOEXIST)) { 404 /* Hoping to find an old_sdata to do inline update 405 * such that it can avoid taking the local_storage->lock 406 * and changing the lists. 407 */ 408 old_sdata = 409 bpf_local_storage_lookup(local_storage, smap, false); 410 err = check_flags(old_sdata, map_flags); 411 if (err) 412 return ERR_PTR(err); 413 if (old_sdata && selem_linked_to_storage(SELEM(old_sdata))) { 414 copy_map_value_locked(&smap->map, old_sdata->data, 415 value, false); 416 return old_sdata; 417 } 418 } 419 420 if (gfp_flags == GFP_KERNEL) { 421 selem = bpf_selem_alloc(smap, owner, value, true, gfp_flags); 422 if (!selem) 423 return ERR_PTR(-ENOMEM); 424 } 425 426 raw_spin_lock_irqsave(&local_storage->lock, flags); 427 428 /* Recheck local_storage->list under local_storage->lock */ 429 if (unlikely(hlist_empty(&local_storage->list))) { 430 /* A parallel del is happening and local_storage is going 431 * away. It has just been checked before, so very 432 * unlikely. Return instead of retry to keep things 433 * simple. 434 */ 435 err = -EAGAIN; 436 goto unlock_err; 437 } 438 439 old_sdata = bpf_local_storage_lookup(local_storage, smap, false); 440 err = check_flags(old_sdata, map_flags); 441 if (err) 442 goto unlock_err; 443 444 if (old_sdata && (map_flags & BPF_F_LOCK)) { 445 copy_map_value_locked(&smap->map, old_sdata->data, value, 446 false); 447 selem = SELEM(old_sdata); 448 goto unlock; 449 } 450 451 if (gfp_flags != GFP_KERNEL) { 452 /* local_storage->lock is held. Hence, we are sure 453 * we can unlink and uncharge the old_sdata successfully 454 * later. Hence, instead of charging the new selem now 455 * and then uncharge the old selem later (which may cause 456 * a potential but unnecessary charge failure), avoid taking 457 * a charge at all here (the "!old_sdata" check) and the 458 * old_sdata will not be uncharged later during 459 * bpf_selem_unlink_storage_nolock(). 460 */ 461 selem = bpf_selem_alloc(smap, owner, value, !old_sdata, gfp_flags); 462 if (!selem) { 463 err = -ENOMEM; 464 goto unlock_err; 465 } 466 } 467 468 /* First, link the new selem to the map */ 469 bpf_selem_link_map(smap, selem); 470 471 /* Second, link (and publish) the new selem to local_storage */ 472 bpf_selem_link_storage_nolock(local_storage, selem); 473 474 /* Third, remove old selem, SELEM(old_sdata) */ 475 if (old_sdata) { 476 bpf_selem_unlink_map(SELEM(old_sdata)); 477 bpf_selem_unlink_storage_nolock(local_storage, SELEM(old_sdata), 478 false, true); 479 } 480 481 unlock: 482 raw_spin_unlock_irqrestore(&local_storage->lock, flags); 483 return SDATA(selem); 484 485 unlock_err: 486 raw_spin_unlock_irqrestore(&local_storage->lock, flags); 487 if (selem) { 488 mem_uncharge(smap, owner, smap->elem_size); 489 kfree(selem); 490 } 491 return ERR_PTR(err); 492 } 493 494 u16 bpf_local_storage_cache_idx_get(struct bpf_local_storage_cache *cache) 495 { 496 u64 min_usage = U64_MAX; 497 u16 i, res = 0; 498 499 spin_lock(&cache->idx_lock); 500 501 for (i = 0; i < BPF_LOCAL_STORAGE_CACHE_SIZE; i++) { 502 if (cache->idx_usage_counts[i] < min_usage) { 503 min_usage = cache->idx_usage_counts[i]; 504 res = i; 505 506 /* Found a free cache_idx */ 507 if (!min_usage) 508 break; 509 } 510 } 511 cache->idx_usage_counts[res]++; 512 513 spin_unlock(&cache->idx_lock); 514 515 return res; 516 } 517 518 void bpf_local_storage_cache_idx_free(struct bpf_local_storage_cache *cache, 519 u16 idx) 520 { 521 spin_lock(&cache->idx_lock); 522 cache->idx_usage_counts[idx]--; 523 spin_unlock(&cache->idx_lock); 524 } 525 526 void bpf_local_storage_map_free(struct bpf_local_storage_map *smap, 527 int __percpu *busy_counter) 528 { 529 struct bpf_local_storage_elem *selem; 530 struct bpf_local_storage_map_bucket *b; 531 unsigned int i; 532 533 /* Note that this map might be concurrently cloned from 534 * bpf_sk_storage_clone. Wait for any existing bpf_sk_storage_clone 535 * RCU read section to finish before proceeding. New RCU 536 * read sections should be prevented via bpf_map_inc_not_zero. 537 */ 538 synchronize_rcu(); 539 540 /* bpf prog and the userspace can no longer access this map 541 * now. No new selem (of this map) can be added 542 * to the owner->storage or to the map bucket's list. 543 * 544 * The elem of this map can be cleaned up here 545 * or when the storage is freed e.g. 546 * by bpf_sk_storage_free() during __sk_destruct(). 547 */ 548 for (i = 0; i < (1U << smap->bucket_log); i++) { 549 b = &smap->buckets[i]; 550 551 rcu_read_lock(); 552 /* No one is adding to b->list now */ 553 while ((selem = hlist_entry_safe( 554 rcu_dereference_raw(hlist_first_rcu(&b->list)), 555 struct bpf_local_storage_elem, map_node))) { 556 if (busy_counter) { 557 migrate_disable(); 558 __this_cpu_inc(*busy_counter); 559 } 560 bpf_selem_unlink(selem, false); 561 if (busy_counter) { 562 __this_cpu_dec(*busy_counter); 563 migrate_enable(); 564 } 565 cond_resched_rcu(); 566 } 567 rcu_read_unlock(); 568 } 569 570 /* While freeing the storage we may still need to access the map. 571 * 572 * e.g. when bpf_sk_storage_free() has unlinked selem from the map 573 * which then made the above while((selem = ...)) loop 574 * exit immediately. 575 * 576 * However, while freeing the storage one still needs to access the 577 * smap->elem_size to do the uncharging in 578 * bpf_selem_unlink_storage_nolock(). 579 * 580 * Hence, wait another rcu grace period for the storage to be freed. 581 */ 582 synchronize_rcu(); 583 584 kvfree(smap->buckets); 585 kfree(smap); 586 } 587 588 int bpf_local_storage_map_alloc_check(union bpf_attr *attr) 589 { 590 if (attr->map_flags & ~BPF_LOCAL_STORAGE_CREATE_FLAG_MASK || 591 !(attr->map_flags & BPF_F_NO_PREALLOC) || 592 attr->max_entries || 593 attr->key_size != sizeof(int) || !attr->value_size || 594 /* Enforce BTF for userspace sk dumping */ 595 !attr->btf_key_type_id || !attr->btf_value_type_id) 596 return -EINVAL; 597 598 if (!bpf_capable()) 599 return -EPERM; 600 601 if (attr->value_size > BPF_LOCAL_STORAGE_MAX_VALUE_SIZE) 602 return -E2BIG; 603 604 return 0; 605 } 606 607 struct bpf_local_storage_map *bpf_local_storage_map_alloc(union bpf_attr *attr) 608 { 609 struct bpf_local_storage_map *smap; 610 unsigned int i; 611 u32 nbuckets; 612 613 smap = kzalloc(sizeof(*smap), GFP_USER | __GFP_NOWARN | __GFP_ACCOUNT); 614 if (!smap) 615 return ERR_PTR(-ENOMEM); 616 bpf_map_init_from_attr(&smap->map, attr); 617 618 nbuckets = roundup_pow_of_two(num_possible_cpus()); 619 /* Use at least 2 buckets, select_bucket() is undefined behavior with 1 bucket */ 620 nbuckets = max_t(u32, 2, nbuckets); 621 smap->bucket_log = ilog2(nbuckets); 622 623 smap->buckets = kvcalloc(sizeof(*smap->buckets), nbuckets, 624 GFP_USER | __GFP_NOWARN | __GFP_ACCOUNT); 625 if (!smap->buckets) { 626 kfree(smap); 627 return ERR_PTR(-ENOMEM); 628 } 629 630 for (i = 0; i < nbuckets; i++) { 631 INIT_HLIST_HEAD(&smap->buckets[i].list); 632 raw_spin_lock_init(&smap->buckets[i].lock); 633 } 634 635 smap->elem_size = 636 sizeof(struct bpf_local_storage_elem) + attr->value_size; 637 638 return smap; 639 } 640 641 int bpf_local_storage_map_check_btf(const struct bpf_map *map, 642 const struct btf *btf, 643 const struct btf_type *key_type, 644 const struct btf_type *value_type) 645 { 646 u32 int_data; 647 648 if (BTF_INFO_KIND(key_type->info) != BTF_KIND_INT) 649 return -EINVAL; 650 651 int_data = *(u32 *)(key_type + 1); 652 if (BTF_INT_BITS(int_data) != 32 || BTF_INT_OFFSET(int_data)) 653 return -EINVAL; 654 655 return 0; 656 } 657