xref: /linux/kernel/bpf/bpf_local_storage.c (revision 151ebcf0797b1a3ba53c8843dc21748c80e098c7)
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_lockless(const struct bpf_local_storage_elem *selem)
55 {
56 	return !hlist_unhashed_lockless(&selem->snode);
57 }
58 
59 static bool selem_linked_to_storage(const struct bpf_local_storage_elem *selem)
60 {
61 	return !hlist_unhashed(&selem->snode);
62 }
63 
64 static bool selem_linked_to_map_lockless(const struct bpf_local_storage_elem *selem)
65 {
66 	return !hlist_unhashed_lockless(&selem->map_node);
67 }
68 
69 static bool selem_linked_to_map(const struct bpf_local_storage_elem *selem)
70 {
71 	return !hlist_unhashed(&selem->map_node);
72 }
73 
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)
77 {
78 	struct bpf_local_storage_elem *selem;
79 
80 	if (charge_mem && mem_charge(smap, owner, smap->elem_size))
81 		return NULL;
82 
83 	if (smap->bpf_ma) {
84 		migrate_disable();
85 		selem = bpf_mem_cache_alloc_flags(&smap->selem_ma, gfp_flags);
86 		migrate_enable();
87 		if (selem)
88 			/* Keep the original bpf_map_kzalloc behavior
89 			 * before started using the bpf_mem_cache_alloc.
90 			 *
91 			 * No need to use zero_map_value. The bpf_selem_free()
92 			 * only does bpf_mem_cache_free when there is
93 			 * no other bpf prog is using the selem.
94 			 */
95 			memset(SDATA(selem)->data, 0, smap->map.value_size);
96 	} else {
97 		selem = bpf_map_kzalloc(&smap->map, smap->elem_size,
98 					gfp_flags | __GFP_NOWARN);
99 	}
100 
101 	if (selem) {
102 		if (value)
103 			copy_map_value(&smap->map, SDATA(selem)->data, value);
104 		/* No need to call check_and_init_map_value as memory is zero init */
105 		return selem;
106 	}
107 
108 	if (charge_mem)
109 		mem_uncharge(smap, owner, smap->elem_size);
110 
111 	return NULL;
112 }
113 
114 /* rcu tasks trace callback for bpf_ma == false */
115 static void __bpf_local_storage_free_trace_rcu(struct rcu_head *rcu)
116 {
117 	struct bpf_local_storage *local_storage;
118 
119 	/* If RCU Tasks Trace grace period implies RCU grace period, do
120 	 * kfree(), else do kfree_rcu().
121 	 */
122 	local_storage = container_of(rcu, struct bpf_local_storage, rcu);
123 	if (rcu_trace_implies_rcu_gp())
124 		kfree(local_storage);
125 	else
126 		kfree_rcu(local_storage, rcu);
127 }
128 
129 static void bpf_local_storage_free_rcu(struct rcu_head *rcu)
130 {
131 	struct bpf_local_storage *local_storage;
132 
133 	local_storage = container_of(rcu, struct bpf_local_storage, rcu);
134 	bpf_mem_cache_raw_free(local_storage);
135 }
136 
137 static void bpf_local_storage_free_trace_rcu(struct rcu_head *rcu)
138 {
139 	if (rcu_trace_implies_rcu_gp())
140 		bpf_local_storage_free_rcu(rcu);
141 	else
142 		call_rcu(rcu, bpf_local_storage_free_rcu);
143 }
144 
145 /* Handle bpf_ma == false */
146 static void __bpf_local_storage_free(struct bpf_local_storage *local_storage,
147 				     bool vanilla_rcu)
148 {
149 	if (vanilla_rcu)
150 		kfree_rcu(local_storage, rcu);
151 	else
152 		call_rcu_tasks_trace(&local_storage->rcu,
153 				     __bpf_local_storage_free_trace_rcu);
154 }
155 
156 static void bpf_local_storage_free(struct bpf_local_storage *local_storage,
157 				   struct bpf_local_storage_map *smap,
158 				   bool bpf_ma, bool reuse_now)
159 {
160 	if (!local_storage)
161 		return;
162 
163 	if (!bpf_ma) {
164 		__bpf_local_storage_free(local_storage, reuse_now);
165 		return;
166 	}
167 
168 	if (!reuse_now) {
169 		call_rcu_tasks_trace(&local_storage->rcu,
170 				     bpf_local_storage_free_trace_rcu);
171 		return;
172 	}
173 
174 	if (smap) {
175 		migrate_disable();
176 		bpf_mem_cache_free(&smap->storage_ma, local_storage);
177 		migrate_enable();
178 	} else {
179 		/* smap could be NULL if the selem that triggered
180 		 * this 'local_storage' creation had been long gone.
181 		 * In this case, directly do call_rcu().
182 		 */
183 		call_rcu(&local_storage->rcu, bpf_local_storage_free_rcu);
184 	}
185 }
186 
187 /* rcu tasks trace callback for bpf_ma == false */
188 static void __bpf_selem_free_trace_rcu(struct rcu_head *rcu)
189 {
190 	struct bpf_local_storage_elem *selem;
191 
192 	selem = container_of(rcu, struct bpf_local_storage_elem, rcu);
193 	if (rcu_trace_implies_rcu_gp())
194 		kfree(selem);
195 	else
196 		kfree_rcu(selem, rcu);
197 }
198 
199 /* Handle bpf_ma == false */
200 static void __bpf_selem_free(struct bpf_local_storage_elem *selem,
201 			     bool vanilla_rcu)
202 {
203 	if (vanilla_rcu)
204 		kfree_rcu(selem, rcu);
205 	else
206 		call_rcu_tasks_trace(&selem->rcu, __bpf_selem_free_trace_rcu);
207 }
208 
209 static void bpf_selem_free_rcu(struct rcu_head *rcu)
210 {
211 	struct bpf_local_storage_elem *selem;
212 
213 	selem = container_of(rcu, struct bpf_local_storage_elem, rcu);
214 	bpf_mem_cache_raw_free(selem);
215 }
216 
217 static void bpf_selem_free_trace_rcu(struct rcu_head *rcu)
218 {
219 	if (rcu_trace_implies_rcu_gp())
220 		bpf_selem_free_rcu(rcu);
221 	else
222 		call_rcu(rcu, bpf_selem_free_rcu);
223 }
224 
225 void bpf_selem_free(struct bpf_local_storage_elem *selem,
226 		    struct bpf_local_storage_map *smap,
227 		    bool reuse_now)
228 {
229 	bpf_obj_free_fields(smap->map.record, SDATA(selem)->data);
230 
231 	if (!smap->bpf_ma) {
232 		__bpf_selem_free(selem, reuse_now);
233 		return;
234 	}
235 
236 	if (!reuse_now) {
237 		call_rcu_tasks_trace(&selem->rcu, bpf_selem_free_trace_rcu);
238 	} else {
239 		/* Instead of using the vanilla call_rcu(),
240 		 * bpf_mem_cache_free will be able to reuse selem
241 		 * immediately.
242 		 */
243 		migrate_disable();
244 		bpf_mem_cache_free(&smap->selem_ma, selem);
245 		migrate_enable();
246 	}
247 }
248 
249 /* local_storage->lock must be held and selem->local_storage == local_storage.
250  * The caller must ensure selem->smap is still valid to be
251  * dereferenced for its smap->elem_size and smap->cache_idx.
252  */
253 static bool bpf_selem_unlink_storage_nolock(struct bpf_local_storage *local_storage,
254 					    struct bpf_local_storage_elem *selem,
255 					    bool uncharge_mem, bool reuse_now)
256 {
257 	struct bpf_local_storage_map *smap;
258 	bool free_local_storage;
259 	void *owner;
260 
261 	smap = rcu_dereference_check(SDATA(selem)->smap, bpf_rcu_lock_held());
262 	owner = local_storage->owner;
263 
264 	/* All uncharging on the owner must be done first.
265 	 * The owner may be freed once the last selem is unlinked
266 	 * from local_storage.
267 	 */
268 	if (uncharge_mem)
269 		mem_uncharge(smap, owner, smap->elem_size);
270 
271 	free_local_storage = hlist_is_singular_node(&selem->snode,
272 						    &local_storage->list);
273 	if (free_local_storage) {
274 		mem_uncharge(smap, owner, sizeof(struct bpf_local_storage));
275 		local_storage->owner = NULL;
276 
277 		/* After this RCU_INIT, owner may be freed and cannot be used */
278 		RCU_INIT_POINTER(*owner_storage(smap, owner), NULL);
279 
280 		/* local_storage is not freed now.  local_storage->lock is
281 		 * still held and raw_spin_unlock_bh(&local_storage->lock)
282 		 * will be done by the caller.
283 		 *
284 		 * Although the unlock will be done under
285 		 * rcu_read_lock(),  it is more intuitive to
286 		 * read if the freeing of the storage is done
287 		 * after the raw_spin_unlock_bh(&local_storage->lock).
288 		 *
289 		 * Hence, a "bool free_local_storage" is returned
290 		 * to the caller which then calls then frees the storage after
291 		 * all the RCU grace periods have expired.
292 		 */
293 	}
294 	hlist_del_init_rcu(&selem->snode);
295 	if (rcu_access_pointer(local_storage->cache[smap->cache_idx]) ==
296 	    SDATA(selem))
297 		RCU_INIT_POINTER(local_storage->cache[smap->cache_idx], NULL);
298 
299 	bpf_selem_free(selem, smap, reuse_now);
300 
301 	if (rcu_access_pointer(local_storage->smap) == smap)
302 		RCU_INIT_POINTER(local_storage->smap, NULL);
303 
304 	return free_local_storage;
305 }
306 
307 static bool check_storage_bpf_ma(struct bpf_local_storage *local_storage,
308 				 struct bpf_local_storage_map *storage_smap,
309 				 struct bpf_local_storage_elem *selem)
310 {
311 
312 	struct bpf_local_storage_map *selem_smap;
313 
314 	/* local_storage->smap may be NULL. If it is, get the bpf_ma
315 	 * from any selem in the local_storage->list. The bpf_ma of all
316 	 * local_storage and selem should have the same value
317 	 * for the same map type.
318 	 *
319 	 * If the local_storage->list is already empty, the caller will not
320 	 * care about the bpf_ma value also because the caller is not
321 	 * responsible to free the local_storage.
322 	 */
323 
324 	if (storage_smap)
325 		return storage_smap->bpf_ma;
326 
327 	if (!selem) {
328 		struct hlist_node *n;
329 
330 		n = rcu_dereference_check(hlist_first_rcu(&local_storage->list),
331 					  bpf_rcu_lock_held());
332 		if (!n)
333 			return false;
334 
335 		selem = hlist_entry(n, struct bpf_local_storage_elem, snode);
336 	}
337 	selem_smap = rcu_dereference_check(SDATA(selem)->smap, bpf_rcu_lock_held());
338 
339 	return selem_smap->bpf_ma;
340 }
341 
342 static void bpf_selem_unlink_storage(struct bpf_local_storage_elem *selem,
343 				     bool reuse_now)
344 {
345 	struct bpf_local_storage_map *storage_smap;
346 	struct bpf_local_storage *local_storage;
347 	bool bpf_ma, free_local_storage = false;
348 	unsigned long flags;
349 
350 	if (unlikely(!selem_linked_to_storage_lockless(selem)))
351 		/* selem has already been unlinked from sk */
352 		return;
353 
354 	local_storage = rcu_dereference_check(selem->local_storage,
355 					      bpf_rcu_lock_held());
356 	storage_smap = rcu_dereference_check(local_storage->smap,
357 					     bpf_rcu_lock_held());
358 	bpf_ma = check_storage_bpf_ma(local_storage, storage_smap, selem);
359 
360 	raw_spin_lock_irqsave(&local_storage->lock, flags);
361 	if (likely(selem_linked_to_storage(selem)))
362 		free_local_storage = bpf_selem_unlink_storage_nolock(
363 			local_storage, selem, true, reuse_now);
364 	raw_spin_unlock_irqrestore(&local_storage->lock, flags);
365 
366 	if (free_local_storage)
367 		bpf_local_storage_free(local_storage, storage_smap, bpf_ma, reuse_now);
368 }
369 
370 void bpf_selem_link_storage_nolock(struct bpf_local_storage *local_storage,
371 				   struct bpf_local_storage_elem *selem)
372 {
373 	RCU_INIT_POINTER(selem->local_storage, local_storage);
374 	hlist_add_head_rcu(&selem->snode, &local_storage->list);
375 }
376 
377 static void bpf_selem_unlink_map(struct bpf_local_storage_elem *selem)
378 {
379 	struct bpf_local_storage_map *smap;
380 	struct bpf_local_storage_map_bucket *b;
381 	unsigned long flags;
382 
383 	if (unlikely(!selem_linked_to_map_lockless(selem)))
384 		/* selem has already be unlinked from smap */
385 		return;
386 
387 	smap = rcu_dereference_check(SDATA(selem)->smap, bpf_rcu_lock_held());
388 	b = select_bucket(smap, selem);
389 	raw_spin_lock_irqsave(&b->lock, flags);
390 	if (likely(selem_linked_to_map(selem)))
391 		hlist_del_init_rcu(&selem->map_node);
392 	raw_spin_unlock_irqrestore(&b->lock, flags);
393 }
394 
395 void bpf_selem_link_map(struct bpf_local_storage_map *smap,
396 			struct bpf_local_storage_elem *selem)
397 {
398 	struct bpf_local_storage_map_bucket *b = select_bucket(smap, selem);
399 	unsigned long flags;
400 
401 	raw_spin_lock_irqsave(&b->lock, flags);
402 	RCU_INIT_POINTER(SDATA(selem)->smap, smap);
403 	hlist_add_head_rcu(&selem->map_node, &b->list);
404 	raw_spin_unlock_irqrestore(&b->lock, flags);
405 }
406 
407 void bpf_selem_unlink(struct bpf_local_storage_elem *selem, bool reuse_now)
408 {
409 	/* Always unlink from map before unlinking from local_storage
410 	 * because selem will be freed after successfully unlinked from
411 	 * the local_storage.
412 	 */
413 	bpf_selem_unlink_map(selem);
414 	bpf_selem_unlink_storage(selem, reuse_now);
415 }
416 
417 void __bpf_local_storage_insert_cache(struct bpf_local_storage *local_storage,
418 				      struct bpf_local_storage_map *smap,
419 				      struct bpf_local_storage_elem *selem)
420 {
421 	unsigned long flags;
422 
423 	/* spinlock is needed to avoid racing with the
424 	 * parallel delete.  Otherwise, publishing an already
425 	 * deleted sdata to the cache will become a use-after-free
426 	 * problem in the next bpf_local_storage_lookup().
427 	 */
428 	raw_spin_lock_irqsave(&local_storage->lock, flags);
429 	if (selem_linked_to_storage(selem))
430 		rcu_assign_pointer(local_storage->cache[smap->cache_idx], SDATA(selem));
431 	raw_spin_unlock_irqrestore(&local_storage->lock, flags);
432 }
433 
434 static int check_flags(const struct bpf_local_storage_data *old_sdata,
435 		       u64 map_flags)
436 {
437 	if (old_sdata && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
438 		/* elem already exists */
439 		return -EEXIST;
440 
441 	if (!old_sdata && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
442 		/* elem doesn't exist, cannot update it */
443 		return -ENOENT;
444 
445 	return 0;
446 }
447 
448 int bpf_local_storage_alloc(void *owner,
449 			    struct bpf_local_storage_map *smap,
450 			    struct bpf_local_storage_elem *first_selem,
451 			    gfp_t gfp_flags)
452 {
453 	struct bpf_local_storage *prev_storage, *storage;
454 	struct bpf_local_storage **owner_storage_ptr;
455 	int err;
456 
457 	err = mem_charge(smap, owner, sizeof(*storage));
458 	if (err)
459 		return err;
460 
461 	if (smap->bpf_ma) {
462 		migrate_disable();
463 		storage = bpf_mem_cache_alloc_flags(&smap->storage_ma, gfp_flags);
464 		migrate_enable();
465 	} else {
466 		storage = bpf_map_kzalloc(&smap->map, sizeof(*storage),
467 					  gfp_flags | __GFP_NOWARN);
468 	}
469 
470 	if (!storage) {
471 		err = -ENOMEM;
472 		goto uncharge;
473 	}
474 
475 	RCU_INIT_POINTER(storage->smap, smap);
476 	INIT_HLIST_HEAD(&storage->list);
477 	raw_spin_lock_init(&storage->lock);
478 	storage->owner = owner;
479 
480 	bpf_selem_link_storage_nolock(storage, first_selem);
481 	bpf_selem_link_map(smap, first_selem);
482 
483 	owner_storage_ptr =
484 		(struct bpf_local_storage **)owner_storage(smap, owner);
485 	/* Publish storage to the owner.
486 	 * Instead of using any lock of the kernel object (i.e. owner),
487 	 * cmpxchg will work with any kernel object regardless what
488 	 * the running context is, bh, irq...etc.
489 	 *
490 	 * From now on, the owner->storage pointer (e.g. sk->sk_bpf_storage)
491 	 * is protected by the storage->lock.  Hence, when freeing
492 	 * the owner->storage, the storage->lock must be held before
493 	 * setting owner->storage ptr to NULL.
494 	 */
495 	prev_storage = cmpxchg(owner_storage_ptr, NULL, storage);
496 	if (unlikely(prev_storage)) {
497 		bpf_selem_unlink_map(first_selem);
498 		err = -EAGAIN;
499 		goto uncharge;
500 
501 		/* Note that even first_selem was linked to smap's
502 		 * bucket->list, first_selem can be freed immediately
503 		 * (instead of kfree_rcu) because
504 		 * bpf_local_storage_map_free() does a
505 		 * synchronize_rcu_mult (waiting for both sleepable and
506 		 * normal programs) before walking the bucket->list.
507 		 * Hence, no one is accessing selem from the
508 		 * bucket->list under rcu_read_lock().
509 		 */
510 	}
511 
512 	return 0;
513 
514 uncharge:
515 	bpf_local_storage_free(storage, smap, smap->bpf_ma, true);
516 	mem_uncharge(smap, owner, sizeof(*storage));
517 	return err;
518 }
519 
520 /* sk cannot be going away because it is linking new elem
521  * to sk->sk_bpf_storage. (i.e. sk->sk_refcnt cannot be 0).
522  * Otherwise, it will become a leak (and other memory issues
523  * during map destruction).
524  */
525 struct bpf_local_storage_data *
526 bpf_local_storage_update(void *owner, struct bpf_local_storage_map *smap,
527 			 void *value, u64 map_flags, gfp_t gfp_flags)
528 {
529 	struct bpf_local_storage_data *old_sdata = NULL;
530 	struct bpf_local_storage_elem *alloc_selem, *selem = NULL;
531 	struct bpf_local_storage *local_storage;
532 	unsigned long flags;
533 	int err;
534 
535 	/* BPF_EXIST and BPF_NOEXIST cannot be both set */
536 	if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST) ||
537 	    /* BPF_F_LOCK can only be used in a value with spin_lock */
538 	    unlikely((map_flags & BPF_F_LOCK) &&
539 		     !btf_record_has_field(smap->map.record, BPF_SPIN_LOCK)))
540 		return ERR_PTR(-EINVAL);
541 
542 	if (gfp_flags == GFP_KERNEL && (map_flags & ~BPF_F_LOCK) != BPF_NOEXIST)
543 		return ERR_PTR(-EINVAL);
544 
545 	local_storage = rcu_dereference_check(*owner_storage(smap, owner),
546 					      bpf_rcu_lock_held());
547 	if (!local_storage || hlist_empty(&local_storage->list)) {
548 		/* Very first elem for the owner */
549 		err = check_flags(NULL, map_flags);
550 		if (err)
551 			return ERR_PTR(err);
552 
553 		selem = bpf_selem_alloc(smap, owner, value, true, gfp_flags);
554 		if (!selem)
555 			return ERR_PTR(-ENOMEM);
556 
557 		err = bpf_local_storage_alloc(owner, smap, selem, gfp_flags);
558 		if (err) {
559 			bpf_selem_free(selem, smap, true);
560 			mem_uncharge(smap, owner, smap->elem_size);
561 			return ERR_PTR(err);
562 		}
563 
564 		return SDATA(selem);
565 	}
566 
567 	if ((map_flags & BPF_F_LOCK) && !(map_flags & BPF_NOEXIST)) {
568 		/* Hoping to find an old_sdata to do inline update
569 		 * such that it can avoid taking the local_storage->lock
570 		 * and changing the lists.
571 		 */
572 		old_sdata =
573 			bpf_local_storage_lookup(local_storage, smap, false);
574 		err = check_flags(old_sdata, map_flags);
575 		if (err)
576 			return ERR_PTR(err);
577 		if (old_sdata && selem_linked_to_storage_lockless(SELEM(old_sdata))) {
578 			copy_map_value_locked(&smap->map, old_sdata->data,
579 					      value, false);
580 			return old_sdata;
581 		}
582 	}
583 
584 	/* A lookup has just been done before and concluded a new selem is
585 	 * needed. The chance of an unnecessary alloc is unlikely.
586 	 */
587 	alloc_selem = selem = bpf_selem_alloc(smap, owner, value, true, gfp_flags);
588 	if (!alloc_selem)
589 		return ERR_PTR(-ENOMEM);
590 
591 	raw_spin_lock_irqsave(&local_storage->lock, flags);
592 
593 	/* Recheck local_storage->list under local_storage->lock */
594 	if (unlikely(hlist_empty(&local_storage->list))) {
595 		/* A parallel del is happening and local_storage is going
596 		 * away.  It has just been checked before, so very
597 		 * unlikely.  Return instead of retry to keep things
598 		 * simple.
599 		 */
600 		err = -EAGAIN;
601 		goto unlock;
602 	}
603 
604 	old_sdata = bpf_local_storage_lookup(local_storage, smap, false);
605 	err = check_flags(old_sdata, map_flags);
606 	if (err)
607 		goto unlock;
608 
609 	if (old_sdata && (map_flags & BPF_F_LOCK)) {
610 		copy_map_value_locked(&smap->map, old_sdata->data, value,
611 				      false);
612 		selem = SELEM(old_sdata);
613 		goto unlock;
614 	}
615 
616 	alloc_selem = NULL;
617 	/* First, link the new selem to the map */
618 	bpf_selem_link_map(smap, selem);
619 
620 	/* Second, link (and publish) the new selem to local_storage */
621 	bpf_selem_link_storage_nolock(local_storage, selem);
622 
623 	/* Third, remove old selem, SELEM(old_sdata) */
624 	if (old_sdata) {
625 		bpf_selem_unlink_map(SELEM(old_sdata));
626 		bpf_selem_unlink_storage_nolock(local_storage, SELEM(old_sdata),
627 						true, false);
628 	}
629 
630 unlock:
631 	raw_spin_unlock_irqrestore(&local_storage->lock, flags);
632 	if (alloc_selem) {
633 		mem_uncharge(smap, owner, smap->elem_size);
634 		bpf_selem_free(alloc_selem, smap, true);
635 	}
636 	return err ? ERR_PTR(err) : SDATA(selem);
637 }
638 
639 static u16 bpf_local_storage_cache_idx_get(struct bpf_local_storage_cache *cache)
640 {
641 	u64 min_usage = U64_MAX;
642 	u16 i, res = 0;
643 
644 	spin_lock(&cache->idx_lock);
645 
646 	for (i = 0; i < BPF_LOCAL_STORAGE_CACHE_SIZE; i++) {
647 		if (cache->idx_usage_counts[i] < min_usage) {
648 			min_usage = cache->idx_usage_counts[i];
649 			res = i;
650 
651 			/* Found a free cache_idx */
652 			if (!min_usage)
653 				break;
654 		}
655 	}
656 	cache->idx_usage_counts[res]++;
657 
658 	spin_unlock(&cache->idx_lock);
659 
660 	return res;
661 }
662 
663 static void bpf_local_storage_cache_idx_free(struct bpf_local_storage_cache *cache,
664 					     u16 idx)
665 {
666 	spin_lock(&cache->idx_lock);
667 	cache->idx_usage_counts[idx]--;
668 	spin_unlock(&cache->idx_lock);
669 }
670 
671 int bpf_local_storage_map_alloc_check(union bpf_attr *attr)
672 {
673 	if (attr->map_flags & ~BPF_LOCAL_STORAGE_CREATE_FLAG_MASK ||
674 	    !(attr->map_flags & BPF_F_NO_PREALLOC) ||
675 	    attr->max_entries ||
676 	    attr->key_size != sizeof(int) || !attr->value_size ||
677 	    /* Enforce BTF for userspace sk dumping */
678 	    !attr->btf_key_type_id || !attr->btf_value_type_id)
679 		return -EINVAL;
680 
681 	if (attr->value_size > BPF_LOCAL_STORAGE_MAX_VALUE_SIZE)
682 		return -E2BIG;
683 
684 	return 0;
685 }
686 
687 int bpf_local_storage_map_check_btf(const struct bpf_map *map,
688 				    const struct btf *btf,
689 				    const struct btf_type *key_type,
690 				    const struct btf_type *value_type)
691 {
692 	u32 int_data;
693 
694 	if (BTF_INFO_KIND(key_type->info) != BTF_KIND_INT)
695 		return -EINVAL;
696 
697 	int_data = *(u32 *)(key_type + 1);
698 	if (BTF_INT_BITS(int_data) != 32 || BTF_INT_OFFSET(int_data))
699 		return -EINVAL;
700 
701 	return 0;
702 }
703 
704 void bpf_local_storage_destroy(struct bpf_local_storage *local_storage)
705 {
706 	struct bpf_local_storage_map *storage_smap;
707 	struct bpf_local_storage_elem *selem;
708 	bool bpf_ma, free_storage = false;
709 	struct hlist_node *n;
710 	unsigned long flags;
711 
712 	storage_smap = rcu_dereference_check(local_storage->smap, bpf_rcu_lock_held());
713 	bpf_ma = check_storage_bpf_ma(local_storage, storage_smap, NULL);
714 
715 	/* Neither the bpf_prog nor the bpf_map's syscall
716 	 * could be modifying the local_storage->list now.
717 	 * Thus, no elem can be added to or deleted from the
718 	 * local_storage->list by the bpf_prog or by the bpf_map's syscall.
719 	 *
720 	 * It is racing with bpf_local_storage_map_free() alone
721 	 * when unlinking elem from the local_storage->list and
722 	 * the map's bucket->list.
723 	 */
724 	raw_spin_lock_irqsave(&local_storage->lock, flags);
725 	hlist_for_each_entry_safe(selem, n, &local_storage->list, snode) {
726 		/* Always unlink from map before unlinking from
727 		 * local_storage.
728 		 */
729 		bpf_selem_unlink_map(selem);
730 		/* If local_storage list has only one element, the
731 		 * bpf_selem_unlink_storage_nolock() will return true.
732 		 * Otherwise, it will return false. The current loop iteration
733 		 * intends to remove all local storage. So the last iteration
734 		 * of the loop will set the free_cgroup_storage to true.
735 		 */
736 		free_storage = bpf_selem_unlink_storage_nolock(
737 			local_storage, selem, true, true);
738 	}
739 	raw_spin_unlock_irqrestore(&local_storage->lock, flags);
740 
741 	if (free_storage)
742 		bpf_local_storage_free(local_storage, storage_smap, bpf_ma, true);
743 }
744 
745 u64 bpf_local_storage_map_mem_usage(const struct bpf_map *map)
746 {
747 	struct bpf_local_storage_map *smap = (struct bpf_local_storage_map *)map;
748 	u64 usage = sizeof(*smap);
749 
750 	/* The dynamically callocated selems are not counted currently. */
751 	usage += sizeof(*smap->buckets) * (1ULL << smap->bucket_log);
752 	return usage;
753 }
754 
755 /* When bpf_ma == true, the bpf_mem_alloc is used to allocate and free memory.
756  * A deadlock free allocator is useful for storage that the bpf prog can easily
757  * get a hold of the owner PTR_TO_BTF_ID in any context. eg. bpf_get_current_task_btf.
758  * The task and cgroup storage fall into this case. The bpf_mem_alloc reuses
759  * memory immediately. To be reuse-immediate safe, the owner destruction
760  * code path needs to go through a rcu grace period before calling
761  * bpf_local_storage_destroy().
762  *
763  * When bpf_ma == false, the kmalloc and kfree are used.
764  */
765 struct bpf_map *
766 bpf_local_storage_map_alloc(union bpf_attr *attr,
767 			    struct bpf_local_storage_cache *cache,
768 			    bool bpf_ma)
769 {
770 	struct bpf_local_storage_map *smap;
771 	unsigned int i;
772 	u32 nbuckets;
773 	int err;
774 
775 	smap = bpf_map_area_alloc(sizeof(*smap), NUMA_NO_NODE);
776 	if (!smap)
777 		return ERR_PTR(-ENOMEM);
778 	bpf_map_init_from_attr(&smap->map, attr);
779 
780 	nbuckets = roundup_pow_of_two(num_possible_cpus());
781 	/* Use at least 2 buckets, select_bucket() is undefined behavior with 1 bucket */
782 	nbuckets = max_t(u32, 2, nbuckets);
783 	smap->bucket_log = ilog2(nbuckets);
784 
785 	smap->buckets = bpf_map_kvcalloc(&smap->map, sizeof(*smap->buckets),
786 					 nbuckets, GFP_USER | __GFP_NOWARN);
787 	if (!smap->buckets) {
788 		err = -ENOMEM;
789 		goto free_smap;
790 	}
791 
792 	for (i = 0; i < nbuckets; i++) {
793 		INIT_HLIST_HEAD(&smap->buckets[i].list);
794 		raw_spin_lock_init(&smap->buckets[i].lock);
795 	}
796 
797 	smap->elem_size = offsetof(struct bpf_local_storage_elem,
798 				   sdata.data[attr->value_size]);
799 
800 	smap->bpf_ma = bpf_ma;
801 	if (bpf_ma) {
802 		err = bpf_mem_alloc_init(&smap->selem_ma, smap->elem_size, false);
803 		if (err)
804 			goto free_smap;
805 
806 		err = bpf_mem_alloc_init(&smap->storage_ma, sizeof(struct bpf_local_storage), false);
807 		if (err) {
808 			bpf_mem_alloc_destroy(&smap->selem_ma);
809 			goto free_smap;
810 		}
811 	}
812 
813 	smap->cache_idx = bpf_local_storage_cache_idx_get(cache);
814 	return &smap->map;
815 
816 free_smap:
817 	kvfree(smap->buckets);
818 	bpf_map_area_free(smap);
819 	return ERR_PTR(err);
820 }
821 
822 void bpf_local_storage_map_free(struct bpf_map *map,
823 				struct bpf_local_storage_cache *cache,
824 				int __percpu *busy_counter)
825 {
826 	struct bpf_local_storage_map_bucket *b;
827 	struct bpf_local_storage_elem *selem;
828 	struct bpf_local_storage_map *smap;
829 	unsigned int i;
830 
831 	smap = (struct bpf_local_storage_map *)map;
832 	bpf_local_storage_cache_idx_free(cache, smap->cache_idx);
833 
834 	/* Note that this map might be concurrently cloned from
835 	 * bpf_sk_storage_clone. Wait for any existing bpf_sk_storage_clone
836 	 * RCU read section to finish before proceeding. New RCU
837 	 * read sections should be prevented via bpf_map_inc_not_zero.
838 	 */
839 	synchronize_rcu();
840 
841 	/* bpf prog and the userspace can no longer access this map
842 	 * now.  No new selem (of this map) can be added
843 	 * to the owner->storage or to the map bucket's list.
844 	 *
845 	 * The elem of this map can be cleaned up here
846 	 * or when the storage is freed e.g.
847 	 * by bpf_sk_storage_free() during __sk_destruct().
848 	 */
849 	for (i = 0; i < (1U << smap->bucket_log); i++) {
850 		b = &smap->buckets[i];
851 
852 		rcu_read_lock();
853 		/* No one is adding to b->list now */
854 		while ((selem = hlist_entry_safe(
855 				rcu_dereference_raw(hlist_first_rcu(&b->list)),
856 				struct bpf_local_storage_elem, map_node))) {
857 			if (busy_counter) {
858 				migrate_disable();
859 				this_cpu_inc(*busy_counter);
860 			}
861 			bpf_selem_unlink(selem, true);
862 			if (busy_counter) {
863 				this_cpu_dec(*busy_counter);
864 				migrate_enable();
865 			}
866 			cond_resched_rcu();
867 		}
868 		rcu_read_unlock();
869 	}
870 
871 	/* While freeing the storage we may still need to access the map.
872 	 *
873 	 * e.g. when bpf_sk_storage_free() has unlinked selem from the map
874 	 * which then made the above while((selem = ...)) loop
875 	 * exit immediately.
876 	 *
877 	 * However, while freeing the storage one still needs to access the
878 	 * smap->elem_size to do the uncharging in
879 	 * bpf_selem_unlink_storage_nolock().
880 	 *
881 	 * Hence, wait another rcu grace period for the storage to be freed.
882 	 */
883 	synchronize_rcu();
884 
885 	if (smap->bpf_ma) {
886 		bpf_mem_alloc_destroy(&smap->selem_ma);
887 		bpf_mem_alloc_destroy(&smap->storage_ma);
888 	}
889 	kvfree(smap->buckets);
890 	bpf_map_area_free(smap);
891 }
892