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>
9 #include <linux/btf_ids.h>
10 #include <linux/bpf_local_storage.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>
18 #define BPF_LOCAL_STORAGE_CREATE_FLAG_MASK (BPF_F_NO_PREALLOC | BPF_F_CLONE)
20 static struct bpf_local_storage_map_bucket *
21 select_bucket(struct bpf_local_storage_map *smap,
22 struct bpf_local_storage_elem *selem)
24 return &smap->buckets[hash_ptr(selem, smap->bucket_log)];
27 static int mem_charge(struct bpf_local_storage_map *smap, void *owner, u32 size)
29 struct bpf_map *map = &smap->map;
31 if (!map->ops->map_local_storage_charge)
34 return map->ops->map_local_storage_charge(smap, owner, size);
37 static void mem_uncharge(struct bpf_local_storage_map *smap, void *owner,
40 struct bpf_map *map = &smap->map;
42 if (map->ops->map_local_storage_uncharge)
43 map->ops->map_local_storage_uncharge(smap, owner, size);
46 static struct bpf_local_storage __rcu **
47 owner_storage(struct bpf_local_storage_map *smap, void *owner)
49 struct bpf_map *map = &smap->map;
51 return map->ops->map_owner_storage_ptr(owner);
54 static bool selem_linked_to_storage_lockless(const struct bpf_local_storage_elem *selem)
56 return !hlist_unhashed_lockless(&selem->snode);
59 static bool selem_linked_to_storage(const struct bpf_local_storage_elem *selem)
61 return !hlist_unhashed(&selem->snode);
64 static bool selem_linked_to_map_lockless(const struct bpf_local_storage_elem *selem)
66 return !hlist_unhashed_lockless(&selem->map_node);
69 static bool selem_linked_to_map(const struct bpf_local_storage_elem *selem)
71 return !hlist_unhashed(&selem->map_node);
74 struct bpf_local_storage_elem *
75 bpf_selem_alloc(struct bpf_local_storage_map *smap, void *owner,
76 void *value, bool charge_mem, gfp_t gfp_flags)
78 struct bpf_local_storage_elem *selem;
80 if (charge_mem && mem_charge(smap, owner, smap->elem_size))
83 selem = bpf_map_kzalloc(&smap->map, smap->elem_size,
84 gfp_flags | __GFP_NOWARN);
87 copy_map_value(&smap->map, SDATA(selem)->data, value);
88 /* No need to call check_and_init_map_value as memory is zero init */
93 mem_uncharge(smap, owner, smap->elem_size);
98 static void bpf_local_storage_free_rcu(struct rcu_head *rcu)
100 struct bpf_local_storage *local_storage;
102 /* If RCU Tasks Trace grace period implies RCU grace period, do
103 * kfree(), else do kfree_rcu().
105 local_storage = container_of(rcu, struct bpf_local_storage, rcu);
106 if (rcu_trace_implies_rcu_gp())
107 kfree(local_storage);
109 kfree_rcu(local_storage, rcu);
112 static void bpf_selem_free_trace_rcu(struct rcu_head *rcu)
114 struct bpf_local_storage_elem *selem;
116 selem = container_of(rcu, struct bpf_local_storage_elem, rcu);
117 if (rcu_trace_implies_rcu_gp())
120 kfree_rcu(selem, rcu);
123 /* local_storage->lock must be held and selem->local_storage == local_storage.
124 * The caller must ensure selem->smap is still valid to be
125 * dereferenced for its smap->elem_size and smap->cache_idx.
127 static bool bpf_selem_unlink_storage_nolock(struct bpf_local_storage *local_storage,
128 struct bpf_local_storage_elem *selem,
129 bool uncharge_mem, bool reuse_now)
131 struct bpf_local_storage_map *smap;
132 bool free_local_storage;
135 smap = rcu_dereference_check(SDATA(selem)->smap, bpf_rcu_lock_held());
136 owner = local_storage->owner;
138 /* All uncharging on the owner must be done first.
139 * The owner may be freed once the last selem is unlinked
140 * from local_storage.
143 mem_uncharge(smap, owner, smap->elem_size);
145 free_local_storage = hlist_is_singular_node(&selem->snode,
146 &local_storage->list);
147 if (free_local_storage) {
148 mem_uncharge(smap, owner, sizeof(struct bpf_local_storage));
149 local_storage->owner = NULL;
151 /* After this RCU_INIT, owner may be freed and cannot be used */
152 RCU_INIT_POINTER(*owner_storage(smap, owner), NULL);
154 /* local_storage is not freed now. local_storage->lock is
155 * still held and raw_spin_unlock_bh(&local_storage->lock)
156 * will be done by the caller.
158 * Although the unlock will be done under
159 * rcu_read_lock(), it is more intuitive to
160 * read if the freeing of the storage is done
161 * after the raw_spin_unlock_bh(&local_storage->lock).
163 * Hence, a "bool free_local_storage" is returned
164 * to the caller which then calls then frees the storage after
165 * all the RCU grace periods have expired.
168 hlist_del_init_rcu(&selem->snode);
169 if (rcu_access_pointer(local_storage->cache[smap->cache_idx]) ==
171 RCU_INIT_POINTER(local_storage->cache[smap->cache_idx], NULL);
173 bpf_obj_free_fields(smap->map.record, SDATA(selem)->data);
175 call_rcu_tasks_trace(&selem->rcu, bpf_selem_free_trace_rcu);
177 kfree_rcu(selem, rcu);
179 if (rcu_access_pointer(local_storage->smap) == smap)
180 RCU_INIT_POINTER(local_storage->smap, NULL);
182 return free_local_storage;
185 static void bpf_selem_unlink_storage(struct bpf_local_storage_elem *selem,
188 struct bpf_local_storage *local_storage;
189 bool free_local_storage = false;
192 if (unlikely(!selem_linked_to_storage_lockless(selem)))
193 /* selem has already been unlinked from sk */
196 local_storage = rcu_dereference_check(selem->local_storage,
197 bpf_rcu_lock_held());
198 raw_spin_lock_irqsave(&local_storage->lock, flags);
199 if (likely(selem_linked_to_storage(selem)))
200 free_local_storage = bpf_selem_unlink_storage_nolock(
201 local_storage, selem, true, reuse_now);
202 raw_spin_unlock_irqrestore(&local_storage->lock, flags);
204 if (free_local_storage) {
206 call_rcu_tasks_trace(&local_storage->rcu,
207 bpf_local_storage_free_rcu);
209 kfree_rcu(local_storage, rcu);
213 void bpf_selem_link_storage_nolock(struct bpf_local_storage *local_storage,
214 struct bpf_local_storage_elem *selem)
216 RCU_INIT_POINTER(selem->local_storage, local_storage);
217 hlist_add_head_rcu(&selem->snode, &local_storage->list);
220 static void bpf_selem_unlink_map(struct bpf_local_storage_elem *selem)
222 struct bpf_local_storage_map *smap;
223 struct bpf_local_storage_map_bucket *b;
226 if (unlikely(!selem_linked_to_map_lockless(selem)))
227 /* selem has already be unlinked from smap */
230 smap = rcu_dereference_check(SDATA(selem)->smap, bpf_rcu_lock_held());
231 b = select_bucket(smap, selem);
232 raw_spin_lock_irqsave(&b->lock, flags);
233 if (likely(selem_linked_to_map(selem)))
234 hlist_del_init_rcu(&selem->map_node);
235 raw_spin_unlock_irqrestore(&b->lock, flags);
238 void bpf_selem_link_map(struct bpf_local_storage_map *smap,
239 struct bpf_local_storage_elem *selem)
241 struct bpf_local_storage_map_bucket *b = select_bucket(smap, selem);
244 raw_spin_lock_irqsave(&b->lock, flags);
245 RCU_INIT_POINTER(SDATA(selem)->smap, smap);
246 hlist_add_head_rcu(&selem->map_node, &b->list);
247 raw_spin_unlock_irqrestore(&b->lock, flags);
250 void bpf_selem_unlink(struct bpf_local_storage_elem *selem, bool reuse_now)
252 /* Always unlink from map before unlinking from local_storage
253 * because selem will be freed after successfully unlinked from
256 bpf_selem_unlink_map(selem);
257 bpf_selem_unlink_storage(selem, reuse_now);
260 /* If cacheit_lockit is false, this lookup function is lockless */
261 struct bpf_local_storage_data *
262 bpf_local_storage_lookup(struct bpf_local_storage *local_storage,
263 struct bpf_local_storage_map *smap,
266 struct bpf_local_storage_data *sdata;
267 struct bpf_local_storage_elem *selem;
269 /* Fast path (cache hit) */
270 sdata = rcu_dereference_check(local_storage->cache[smap->cache_idx],
271 bpf_rcu_lock_held());
272 if (sdata && rcu_access_pointer(sdata->smap) == smap)
275 /* Slow path (cache miss) */
276 hlist_for_each_entry_rcu(selem, &local_storage->list, snode,
277 rcu_read_lock_trace_held())
278 if (rcu_access_pointer(SDATA(selem)->smap) == smap)
284 sdata = SDATA(selem);
285 if (cacheit_lockit) {
288 /* spinlock is needed to avoid racing with the
289 * parallel delete. Otherwise, publishing an already
290 * deleted sdata to the cache will become a use-after-free
291 * problem in the next bpf_local_storage_lookup().
293 raw_spin_lock_irqsave(&local_storage->lock, flags);
294 if (selem_linked_to_storage(selem))
295 rcu_assign_pointer(local_storage->cache[smap->cache_idx],
297 raw_spin_unlock_irqrestore(&local_storage->lock, flags);
303 static int check_flags(const struct bpf_local_storage_data *old_sdata,
306 if (old_sdata && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
307 /* elem already exists */
310 if (!old_sdata && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
311 /* elem doesn't exist, cannot update it */
317 int bpf_local_storage_alloc(void *owner,
318 struct bpf_local_storage_map *smap,
319 struct bpf_local_storage_elem *first_selem,
322 struct bpf_local_storage *prev_storage, *storage;
323 struct bpf_local_storage **owner_storage_ptr;
326 err = mem_charge(smap, owner, sizeof(*storage));
330 storage = bpf_map_kzalloc(&smap->map, sizeof(*storage),
331 gfp_flags | __GFP_NOWARN);
337 RCU_INIT_POINTER(storage->smap, smap);
338 INIT_HLIST_HEAD(&storage->list);
339 raw_spin_lock_init(&storage->lock);
340 storage->owner = owner;
342 bpf_selem_link_storage_nolock(storage, first_selem);
343 bpf_selem_link_map(smap, first_selem);
346 (struct bpf_local_storage **)owner_storage(smap, owner);
347 /* Publish storage to the owner.
348 * Instead of using any lock of the kernel object (i.e. owner),
349 * cmpxchg will work with any kernel object regardless what
350 * the running context is, bh, irq...etc.
352 * From now on, the owner->storage pointer (e.g. sk->sk_bpf_storage)
353 * is protected by the storage->lock. Hence, when freeing
354 * the owner->storage, the storage->lock must be held before
355 * setting owner->storage ptr to NULL.
357 prev_storage = cmpxchg(owner_storage_ptr, NULL, storage);
358 if (unlikely(prev_storage)) {
359 bpf_selem_unlink_map(first_selem);
363 /* Note that even first_selem was linked to smap's
364 * bucket->list, first_selem can be freed immediately
365 * (instead of kfree_rcu) because
366 * bpf_local_storage_map_free() does a
367 * synchronize_rcu_mult (waiting for both sleepable and
368 * normal programs) before walking the bucket->list.
369 * Hence, no one is accessing selem from the
370 * bucket->list under rcu_read_lock().
378 mem_uncharge(smap, owner, sizeof(*storage));
382 /* sk cannot be going away because it is linking new elem
383 * to sk->sk_bpf_storage. (i.e. sk->sk_refcnt cannot be 0).
384 * Otherwise, it will become a leak (and other memory issues
385 * during map destruction).
387 struct bpf_local_storage_data *
388 bpf_local_storage_update(void *owner, struct bpf_local_storage_map *smap,
389 void *value, u64 map_flags, gfp_t gfp_flags)
391 struct bpf_local_storage_data *old_sdata = NULL;
392 struct bpf_local_storage_elem *selem = NULL;
393 struct bpf_local_storage *local_storage;
397 /* BPF_EXIST and BPF_NOEXIST cannot be both set */
398 if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST) ||
399 /* BPF_F_LOCK can only be used in a value with spin_lock */
400 unlikely((map_flags & BPF_F_LOCK) &&
401 !btf_record_has_field(smap->map.record, BPF_SPIN_LOCK)))
402 return ERR_PTR(-EINVAL);
404 if (gfp_flags == GFP_KERNEL && (map_flags & ~BPF_F_LOCK) != BPF_NOEXIST)
405 return ERR_PTR(-EINVAL);
407 local_storage = rcu_dereference_check(*owner_storage(smap, owner),
408 bpf_rcu_lock_held());
409 if (!local_storage || hlist_empty(&local_storage->list)) {
410 /* Very first elem for the owner */
411 err = check_flags(NULL, map_flags);
415 selem = bpf_selem_alloc(smap, owner, value, true, gfp_flags);
417 return ERR_PTR(-ENOMEM);
419 err = bpf_local_storage_alloc(owner, smap, selem, gfp_flags);
422 mem_uncharge(smap, owner, smap->elem_size);
429 if ((map_flags & BPF_F_LOCK) && !(map_flags & BPF_NOEXIST)) {
430 /* Hoping to find an old_sdata to do inline update
431 * such that it can avoid taking the local_storage->lock
432 * and changing the lists.
435 bpf_local_storage_lookup(local_storage, smap, false);
436 err = check_flags(old_sdata, map_flags);
439 if (old_sdata && selem_linked_to_storage_lockless(SELEM(old_sdata))) {
440 copy_map_value_locked(&smap->map, old_sdata->data,
446 if (gfp_flags == GFP_KERNEL) {
447 selem = bpf_selem_alloc(smap, owner, value, true, gfp_flags);
449 return ERR_PTR(-ENOMEM);
452 raw_spin_lock_irqsave(&local_storage->lock, flags);
454 /* Recheck local_storage->list under local_storage->lock */
455 if (unlikely(hlist_empty(&local_storage->list))) {
456 /* A parallel del is happening and local_storage is going
457 * away. It has just been checked before, so very
458 * unlikely. Return instead of retry to keep things
465 old_sdata = bpf_local_storage_lookup(local_storage, smap, false);
466 err = check_flags(old_sdata, map_flags);
470 if (old_sdata && (map_flags & BPF_F_LOCK)) {
471 copy_map_value_locked(&smap->map, old_sdata->data, value,
473 selem = SELEM(old_sdata);
477 if (gfp_flags != GFP_KERNEL) {
478 /* local_storage->lock is held. Hence, we are sure
479 * we can unlink and uncharge the old_sdata successfully
480 * later. Hence, instead of charging the new selem now
481 * and then uncharge the old selem later (which may cause
482 * a potential but unnecessary charge failure), avoid taking
483 * a charge at all here (the "!old_sdata" check) and the
484 * old_sdata will not be uncharged later during
485 * bpf_selem_unlink_storage_nolock().
487 selem = bpf_selem_alloc(smap, owner, value, !old_sdata, gfp_flags);
494 /* First, link the new selem to the map */
495 bpf_selem_link_map(smap, selem);
497 /* Second, link (and publish) the new selem to local_storage */
498 bpf_selem_link_storage_nolock(local_storage, selem);
500 /* Third, remove old selem, SELEM(old_sdata) */
502 bpf_selem_unlink_map(SELEM(old_sdata));
503 bpf_selem_unlink_storage_nolock(local_storage, SELEM(old_sdata),
508 raw_spin_unlock_irqrestore(&local_storage->lock, flags);
512 raw_spin_unlock_irqrestore(&local_storage->lock, flags);
514 mem_uncharge(smap, owner, smap->elem_size);
520 static u16 bpf_local_storage_cache_idx_get(struct bpf_local_storage_cache *cache)
522 u64 min_usage = U64_MAX;
525 spin_lock(&cache->idx_lock);
527 for (i = 0; i < BPF_LOCAL_STORAGE_CACHE_SIZE; i++) {
528 if (cache->idx_usage_counts[i] < min_usage) {
529 min_usage = cache->idx_usage_counts[i];
532 /* Found a free cache_idx */
537 cache->idx_usage_counts[res]++;
539 spin_unlock(&cache->idx_lock);
544 static void bpf_local_storage_cache_idx_free(struct bpf_local_storage_cache *cache,
547 spin_lock(&cache->idx_lock);
548 cache->idx_usage_counts[idx]--;
549 spin_unlock(&cache->idx_lock);
552 int bpf_local_storage_map_alloc_check(union bpf_attr *attr)
554 if (attr->map_flags & ~BPF_LOCAL_STORAGE_CREATE_FLAG_MASK ||
555 !(attr->map_flags & BPF_F_NO_PREALLOC) ||
557 attr->key_size != sizeof(int) || !attr->value_size ||
558 /* Enforce BTF for userspace sk dumping */
559 !attr->btf_key_type_id || !attr->btf_value_type_id)
565 if (attr->value_size > BPF_LOCAL_STORAGE_MAX_VALUE_SIZE)
571 int bpf_local_storage_map_check_btf(const struct bpf_map *map,
572 const struct btf *btf,
573 const struct btf_type *key_type,
574 const struct btf_type *value_type)
578 if (BTF_INFO_KIND(key_type->info) != BTF_KIND_INT)
581 int_data = *(u32 *)(key_type + 1);
582 if (BTF_INT_BITS(int_data) != 32 || BTF_INT_OFFSET(int_data))
588 void bpf_local_storage_destroy(struct bpf_local_storage *local_storage)
590 struct bpf_local_storage_elem *selem;
591 bool free_storage = false;
592 struct hlist_node *n;
595 /* Neither the bpf_prog nor the bpf_map's syscall
596 * could be modifying the local_storage->list now.
597 * Thus, no elem can be added to or deleted from the
598 * local_storage->list by the bpf_prog or by the bpf_map's syscall.
600 * It is racing with bpf_local_storage_map_free() alone
601 * when unlinking elem from the local_storage->list and
602 * the map's bucket->list.
604 raw_spin_lock_irqsave(&local_storage->lock, flags);
605 hlist_for_each_entry_safe(selem, n, &local_storage->list, snode) {
606 /* Always unlink from map before unlinking from
609 bpf_selem_unlink_map(selem);
610 /* If local_storage list has only one element, the
611 * bpf_selem_unlink_storage_nolock() will return true.
612 * Otherwise, it will return false. The current loop iteration
613 * intends to remove all local storage. So the last iteration
614 * of the loop will set the free_cgroup_storage to true.
616 free_storage = bpf_selem_unlink_storage_nolock(
617 local_storage, selem, false, true);
619 raw_spin_unlock_irqrestore(&local_storage->lock, flags);
622 kfree_rcu(local_storage, rcu);
625 u64 bpf_local_storage_map_mem_usage(const struct bpf_map *map)
627 struct bpf_local_storage_map *smap = (struct bpf_local_storage_map *)map;
628 u64 usage = sizeof(*smap);
630 /* The dynamically callocated selems are not counted currently. */
631 usage += sizeof(*smap->buckets) * (1ULL << smap->bucket_log);
636 bpf_local_storage_map_alloc(union bpf_attr *attr,
637 struct bpf_local_storage_cache *cache)
639 struct bpf_local_storage_map *smap;
643 smap = bpf_map_area_alloc(sizeof(*smap), NUMA_NO_NODE);
645 return ERR_PTR(-ENOMEM);
646 bpf_map_init_from_attr(&smap->map, attr);
648 nbuckets = roundup_pow_of_two(num_possible_cpus());
649 /* Use at least 2 buckets, select_bucket() is undefined behavior with 1 bucket */
650 nbuckets = max_t(u32, 2, nbuckets);
651 smap->bucket_log = ilog2(nbuckets);
653 smap->buckets = bpf_map_kvcalloc(&smap->map, sizeof(*smap->buckets),
654 nbuckets, GFP_USER | __GFP_NOWARN);
655 if (!smap->buckets) {
656 bpf_map_area_free(smap);
657 return ERR_PTR(-ENOMEM);
660 for (i = 0; i < nbuckets; i++) {
661 INIT_HLIST_HEAD(&smap->buckets[i].list);
662 raw_spin_lock_init(&smap->buckets[i].lock);
665 smap->elem_size = offsetof(struct bpf_local_storage_elem,
666 sdata.data[attr->value_size]);
668 smap->cache_idx = bpf_local_storage_cache_idx_get(cache);
672 void bpf_local_storage_map_free(struct bpf_map *map,
673 struct bpf_local_storage_cache *cache,
674 int __percpu *busy_counter)
676 struct bpf_local_storage_map_bucket *b;
677 struct bpf_local_storage_elem *selem;
678 struct bpf_local_storage_map *smap;
681 smap = (struct bpf_local_storage_map *)map;
682 bpf_local_storage_cache_idx_free(cache, smap->cache_idx);
684 /* Note that this map might be concurrently cloned from
685 * bpf_sk_storage_clone. Wait for any existing bpf_sk_storage_clone
686 * RCU read section to finish before proceeding. New RCU
687 * read sections should be prevented via bpf_map_inc_not_zero.
691 /* bpf prog and the userspace can no longer access this map
692 * now. No new selem (of this map) can be added
693 * to the owner->storage or to the map bucket's list.
695 * The elem of this map can be cleaned up here
696 * or when the storage is freed e.g.
697 * by bpf_sk_storage_free() during __sk_destruct().
699 for (i = 0; i < (1U << smap->bucket_log); i++) {
700 b = &smap->buckets[i];
703 /* No one is adding to b->list now */
704 while ((selem = hlist_entry_safe(
705 rcu_dereference_raw(hlist_first_rcu(&b->list)),
706 struct bpf_local_storage_elem, map_node))) {
709 this_cpu_inc(*busy_counter);
711 bpf_selem_unlink(selem, true);
713 this_cpu_dec(*busy_counter);
721 /* While freeing the storage we may still need to access the map.
723 * e.g. when bpf_sk_storage_free() has unlinked selem from the map
724 * which then made the above while((selem = ...)) loop
727 * However, while freeing the storage one still needs to access the
728 * smap->elem_size to do the uncharging in
729 * bpf_selem_unlink_storage_nolock().
731 * Hence, wait another rcu grace period for the storage to be freed.
735 kvfree(smap->buckets);
736 bpf_map_area_free(smap);