xref: /openbmc/linux/kernel/bpf/hashtab.c (revision aac5987a)
1 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
2  * Copyright (c) 2016 Facebook
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of version 2 of the GNU General Public
6  * License as published by the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it will be useful, but
9  * WITHOUT ANY WARRANTY; without even the implied warranty of
10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11  * General Public License for more details.
12  */
13 #include <linux/bpf.h>
14 #include <linux/jhash.h>
15 #include <linux/filter.h>
16 #include "percpu_freelist.h"
17 #include "bpf_lru_list.h"
18 
19 struct bucket {
20 	struct hlist_head head;
21 	raw_spinlock_t lock;
22 };
23 
24 struct bpf_htab {
25 	struct bpf_map map;
26 	struct bucket *buckets;
27 	void *elems;
28 	union {
29 		struct pcpu_freelist freelist;
30 		struct bpf_lru lru;
31 	};
32 	void __percpu *extra_elems;
33 	atomic_t count;	/* number of elements in this hashtable */
34 	u32 n_buckets;	/* number of hash buckets */
35 	u32 elem_size;	/* size of each element in bytes */
36 };
37 
38 enum extra_elem_state {
39 	HTAB_NOT_AN_EXTRA_ELEM = 0,
40 	HTAB_EXTRA_ELEM_FREE,
41 	HTAB_EXTRA_ELEM_USED
42 };
43 
44 /* each htab element is struct htab_elem + key + value */
45 struct htab_elem {
46 	union {
47 		struct hlist_node hash_node;
48 		struct bpf_htab *htab;
49 		struct pcpu_freelist_node fnode;
50 	};
51 	union {
52 		struct rcu_head rcu;
53 		enum extra_elem_state state;
54 		struct bpf_lru_node lru_node;
55 	};
56 	u32 hash;
57 	char key[0] __aligned(8);
58 };
59 
60 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
61 
62 static bool htab_is_lru(const struct bpf_htab *htab)
63 {
64 	return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH ||
65 		htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
66 }
67 
68 static bool htab_is_percpu(const struct bpf_htab *htab)
69 {
70 	return htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH ||
71 		htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
72 }
73 
74 static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
75 				     void __percpu *pptr)
76 {
77 	*(void __percpu **)(l->key + key_size) = pptr;
78 }
79 
80 static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
81 {
82 	return *(void __percpu **)(l->key + key_size);
83 }
84 
85 static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
86 {
87 	return (struct htab_elem *) (htab->elems + i * htab->elem_size);
88 }
89 
90 static void htab_free_elems(struct bpf_htab *htab)
91 {
92 	int i;
93 
94 	if (!htab_is_percpu(htab))
95 		goto free_elems;
96 
97 	for (i = 0; i < htab->map.max_entries; i++) {
98 		void __percpu *pptr;
99 
100 		pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
101 					 htab->map.key_size);
102 		free_percpu(pptr);
103 	}
104 free_elems:
105 	bpf_map_area_free(htab->elems);
106 }
107 
108 static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
109 					  u32 hash)
110 {
111 	struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
112 	struct htab_elem *l;
113 
114 	if (node) {
115 		l = container_of(node, struct htab_elem, lru_node);
116 		memcpy(l->key, key, htab->map.key_size);
117 		return l;
118 	}
119 
120 	return NULL;
121 }
122 
123 static int prealloc_init(struct bpf_htab *htab)
124 {
125 	int err = -ENOMEM, i;
126 
127 	htab->elems = bpf_map_area_alloc(htab->elem_size *
128 					 htab->map.max_entries);
129 	if (!htab->elems)
130 		return -ENOMEM;
131 
132 	if (!htab_is_percpu(htab))
133 		goto skip_percpu_elems;
134 
135 	for (i = 0; i < htab->map.max_entries; i++) {
136 		u32 size = round_up(htab->map.value_size, 8);
137 		void __percpu *pptr;
138 
139 		pptr = __alloc_percpu_gfp(size, 8, GFP_USER | __GFP_NOWARN);
140 		if (!pptr)
141 			goto free_elems;
142 		htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
143 				  pptr);
144 	}
145 
146 skip_percpu_elems:
147 	if (htab_is_lru(htab))
148 		err = bpf_lru_init(&htab->lru,
149 				   htab->map.map_flags & BPF_F_NO_COMMON_LRU,
150 				   offsetof(struct htab_elem, hash) -
151 				   offsetof(struct htab_elem, lru_node),
152 				   htab_lru_map_delete_node,
153 				   htab);
154 	else
155 		err = pcpu_freelist_init(&htab->freelist);
156 
157 	if (err)
158 		goto free_elems;
159 
160 	if (htab_is_lru(htab))
161 		bpf_lru_populate(&htab->lru, htab->elems,
162 				 offsetof(struct htab_elem, lru_node),
163 				 htab->elem_size, htab->map.max_entries);
164 	else
165 		pcpu_freelist_populate(&htab->freelist, htab->elems,
166 				       htab->elem_size, htab->map.max_entries);
167 
168 	return 0;
169 
170 free_elems:
171 	htab_free_elems(htab);
172 	return err;
173 }
174 
175 static void prealloc_destroy(struct bpf_htab *htab)
176 {
177 	htab_free_elems(htab);
178 
179 	if (htab_is_lru(htab))
180 		bpf_lru_destroy(&htab->lru);
181 	else
182 		pcpu_freelist_destroy(&htab->freelist);
183 }
184 
185 static int alloc_extra_elems(struct bpf_htab *htab)
186 {
187 	void __percpu *pptr;
188 	int cpu;
189 
190 	pptr = __alloc_percpu_gfp(htab->elem_size, 8, GFP_USER | __GFP_NOWARN);
191 	if (!pptr)
192 		return -ENOMEM;
193 
194 	for_each_possible_cpu(cpu) {
195 		((struct htab_elem *)per_cpu_ptr(pptr, cpu))->state =
196 			HTAB_EXTRA_ELEM_FREE;
197 	}
198 	htab->extra_elems = pptr;
199 	return 0;
200 }
201 
202 /* Called from syscall */
203 static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
204 {
205 	bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
206 		       attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
207 	bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
208 		    attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
209 	/* percpu_lru means each cpu has its own LRU list.
210 	 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
211 	 * the map's value itself is percpu.  percpu_lru has
212 	 * nothing to do with the map's value.
213 	 */
214 	bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
215 	bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
216 	struct bpf_htab *htab;
217 	int err, i;
218 	u64 cost;
219 
220 	if (lru && !capable(CAP_SYS_ADMIN))
221 		/* LRU implementation is much complicated than other
222 		 * maps.  Hence, limit to CAP_SYS_ADMIN for now.
223 		 */
224 		return ERR_PTR(-EPERM);
225 
226 	if (attr->map_flags & ~(BPF_F_NO_PREALLOC | BPF_F_NO_COMMON_LRU))
227 		/* reserved bits should not be used */
228 		return ERR_PTR(-EINVAL);
229 
230 	if (!lru && percpu_lru)
231 		return ERR_PTR(-EINVAL);
232 
233 	if (lru && !prealloc)
234 		return ERR_PTR(-ENOTSUPP);
235 
236 	htab = kzalloc(sizeof(*htab), GFP_USER);
237 	if (!htab)
238 		return ERR_PTR(-ENOMEM);
239 
240 	/* mandatory map attributes */
241 	htab->map.map_type = attr->map_type;
242 	htab->map.key_size = attr->key_size;
243 	htab->map.value_size = attr->value_size;
244 	htab->map.max_entries = attr->max_entries;
245 	htab->map.map_flags = attr->map_flags;
246 
247 	/* check sanity of attributes.
248 	 * value_size == 0 may be allowed in the future to use map as a set
249 	 */
250 	err = -EINVAL;
251 	if (htab->map.max_entries == 0 || htab->map.key_size == 0 ||
252 	    htab->map.value_size == 0)
253 		goto free_htab;
254 
255 	if (percpu_lru) {
256 		/* ensure each CPU's lru list has >=1 elements.
257 		 * since we are at it, make each lru list has the same
258 		 * number of elements.
259 		 */
260 		htab->map.max_entries = roundup(attr->max_entries,
261 						num_possible_cpus());
262 		if (htab->map.max_entries < attr->max_entries)
263 			htab->map.max_entries = rounddown(attr->max_entries,
264 							  num_possible_cpus());
265 	}
266 
267 	/* hash table size must be power of 2 */
268 	htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
269 
270 	err = -E2BIG;
271 	if (htab->map.key_size > MAX_BPF_STACK)
272 		/* eBPF programs initialize keys on stack, so they cannot be
273 		 * larger than max stack size
274 		 */
275 		goto free_htab;
276 
277 	if (htab->map.value_size >= KMALLOC_MAX_SIZE -
278 	    MAX_BPF_STACK - sizeof(struct htab_elem))
279 		/* if value_size is bigger, the user space won't be able to
280 		 * access the elements via bpf syscall. This check also makes
281 		 * sure that the elem_size doesn't overflow and it's
282 		 * kmalloc-able later in htab_map_update_elem()
283 		 */
284 		goto free_htab;
285 
286 	if (percpu && round_up(htab->map.value_size, 8) > PCPU_MIN_UNIT_SIZE)
287 		/* make sure the size for pcpu_alloc() is reasonable */
288 		goto free_htab;
289 
290 	htab->elem_size = sizeof(struct htab_elem) +
291 			  round_up(htab->map.key_size, 8);
292 	if (percpu)
293 		htab->elem_size += sizeof(void *);
294 	else
295 		htab->elem_size += round_up(htab->map.value_size, 8);
296 
297 	/* prevent zero size kmalloc and check for u32 overflow */
298 	if (htab->n_buckets == 0 ||
299 	    htab->n_buckets > U32_MAX / sizeof(struct bucket))
300 		goto free_htab;
301 
302 	cost = (u64) htab->n_buckets * sizeof(struct bucket) +
303 	       (u64) htab->elem_size * htab->map.max_entries;
304 
305 	if (percpu)
306 		cost += (u64) round_up(htab->map.value_size, 8) *
307 			num_possible_cpus() * htab->map.max_entries;
308 	else
309 	       cost += (u64) htab->elem_size * num_possible_cpus();
310 
311 	if (cost >= U32_MAX - PAGE_SIZE)
312 		/* make sure page count doesn't overflow */
313 		goto free_htab;
314 
315 	htab->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
316 
317 	/* if map size is larger than memlock limit, reject it early */
318 	err = bpf_map_precharge_memlock(htab->map.pages);
319 	if (err)
320 		goto free_htab;
321 
322 	err = -ENOMEM;
323 	htab->buckets = bpf_map_area_alloc(htab->n_buckets *
324 					   sizeof(struct bucket));
325 	if (!htab->buckets)
326 		goto free_htab;
327 
328 	for (i = 0; i < htab->n_buckets; i++) {
329 		INIT_HLIST_HEAD(&htab->buckets[i].head);
330 		raw_spin_lock_init(&htab->buckets[i].lock);
331 	}
332 
333 	if (!percpu && !lru) {
334 		/* lru itself can remove the least used element, so
335 		 * there is no need for an extra elem during map_update.
336 		 */
337 		err = alloc_extra_elems(htab);
338 		if (err)
339 			goto free_buckets;
340 	}
341 
342 	if (prealloc) {
343 		err = prealloc_init(htab);
344 		if (err)
345 			goto free_extra_elems;
346 	}
347 
348 	return &htab->map;
349 
350 free_extra_elems:
351 	free_percpu(htab->extra_elems);
352 free_buckets:
353 	bpf_map_area_free(htab->buckets);
354 free_htab:
355 	kfree(htab);
356 	return ERR_PTR(err);
357 }
358 
359 static inline u32 htab_map_hash(const void *key, u32 key_len)
360 {
361 	return jhash(key, key_len, 0);
362 }
363 
364 static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
365 {
366 	return &htab->buckets[hash & (htab->n_buckets - 1)];
367 }
368 
369 static inline struct hlist_head *select_bucket(struct bpf_htab *htab, u32 hash)
370 {
371 	return &__select_bucket(htab, hash)->head;
372 }
373 
374 static struct htab_elem *lookup_elem_raw(struct hlist_head *head, u32 hash,
375 					 void *key, u32 key_size)
376 {
377 	struct htab_elem *l;
378 
379 	hlist_for_each_entry_rcu(l, head, hash_node)
380 		if (l->hash == hash && !memcmp(&l->key, key, key_size))
381 			return l;
382 
383 	return NULL;
384 }
385 
386 /* Called from syscall or from eBPF program */
387 static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
388 {
389 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
390 	struct hlist_head *head;
391 	struct htab_elem *l;
392 	u32 hash, key_size;
393 
394 	/* Must be called with rcu_read_lock. */
395 	WARN_ON_ONCE(!rcu_read_lock_held());
396 
397 	key_size = map->key_size;
398 
399 	hash = htab_map_hash(key, key_size);
400 
401 	head = select_bucket(htab, hash);
402 
403 	l = lookup_elem_raw(head, hash, key, key_size);
404 
405 	return l;
406 }
407 
408 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
409 {
410 	struct htab_elem *l = __htab_map_lookup_elem(map, key);
411 
412 	if (l)
413 		return l->key + round_up(map->key_size, 8);
414 
415 	return NULL;
416 }
417 
418 static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
419 {
420 	struct htab_elem *l = __htab_map_lookup_elem(map, key);
421 
422 	if (l) {
423 		bpf_lru_node_set_ref(&l->lru_node);
424 		return l->key + round_up(map->key_size, 8);
425 	}
426 
427 	return NULL;
428 }
429 
430 /* It is called from the bpf_lru_list when the LRU needs to delete
431  * older elements from the htab.
432  */
433 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
434 {
435 	struct bpf_htab *htab = (struct bpf_htab *)arg;
436 	struct htab_elem *l, *tgt_l;
437 	struct hlist_head *head;
438 	unsigned long flags;
439 	struct bucket *b;
440 
441 	tgt_l = container_of(node, struct htab_elem, lru_node);
442 	b = __select_bucket(htab, tgt_l->hash);
443 	head = &b->head;
444 
445 	raw_spin_lock_irqsave(&b->lock, flags);
446 
447 	hlist_for_each_entry_rcu(l, head, hash_node)
448 		if (l == tgt_l) {
449 			hlist_del_rcu(&l->hash_node);
450 			break;
451 		}
452 
453 	raw_spin_unlock_irqrestore(&b->lock, flags);
454 
455 	return l == tgt_l;
456 }
457 
458 /* Called from syscall */
459 static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
460 {
461 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
462 	struct hlist_head *head;
463 	struct htab_elem *l, *next_l;
464 	u32 hash, key_size;
465 	int i;
466 
467 	WARN_ON_ONCE(!rcu_read_lock_held());
468 
469 	key_size = map->key_size;
470 
471 	hash = htab_map_hash(key, key_size);
472 
473 	head = select_bucket(htab, hash);
474 
475 	/* lookup the key */
476 	l = lookup_elem_raw(head, hash, key, key_size);
477 
478 	if (!l) {
479 		i = 0;
480 		goto find_first_elem;
481 	}
482 
483 	/* key was found, get next key in the same bucket */
484 	next_l = hlist_entry_safe(rcu_dereference_raw(hlist_next_rcu(&l->hash_node)),
485 				  struct htab_elem, hash_node);
486 
487 	if (next_l) {
488 		/* if next elem in this hash list is non-zero, just return it */
489 		memcpy(next_key, next_l->key, key_size);
490 		return 0;
491 	}
492 
493 	/* no more elements in this hash list, go to the next bucket */
494 	i = hash & (htab->n_buckets - 1);
495 	i++;
496 
497 find_first_elem:
498 	/* iterate over buckets */
499 	for (; i < htab->n_buckets; i++) {
500 		head = select_bucket(htab, i);
501 
502 		/* pick first element in the bucket */
503 		next_l = hlist_entry_safe(rcu_dereference_raw(hlist_first_rcu(head)),
504 					  struct htab_elem, hash_node);
505 		if (next_l) {
506 			/* if it's not empty, just return it */
507 			memcpy(next_key, next_l->key, key_size);
508 			return 0;
509 		}
510 	}
511 
512 	/* iterated over all buckets and all elements */
513 	return -ENOENT;
514 }
515 
516 static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
517 {
518 	if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
519 		free_percpu(htab_elem_get_ptr(l, htab->map.key_size));
520 	kfree(l);
521 }
522 
523 static void htab_elem_free_rcu(struct rcu_head *head)
524 {
525 	struct htab_elem *l = container_of(head, struct htab_elem, rcu);
526 	struct bpf_htab *htab = l->htab;
527 
528 	/* must increment bpf_prog_active to avoid kprobe+bpf triggering while
529 	 * we're calling kfree, otherwise deadlock is possible if kprobes
530 	 * are placed somewhere inside of slub
531 	 */
532 	preempt_disable();
533 	__this_cpu_inc(bpf_prog_active);
534 	htab_elem_free(htab, l);
535 	__this_cpu_dec(bpf_prog_active);
536 	preempt_enable();
537 }
538 
539 static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
540 {
541 	if (l->state == HTAB_EXTRA_ELEM_USED) {
542 		l->state = HTAB_EXTRA_ELEM_FREE;
543 		return;
544 	}
545 
546 	if (!(htab->map.map_flags & BPF_F_NO_PREALLOC)) {
547 		pcpu_freelist_push(&htab->freelist, &l->fnode);
548 	} else {
549 		atomic_dec(&htab->count);
550 		l->htab = htab;
551 		call_rcu(&l->rcu, htab_elem_free_rcu);
552 	}
553 }
554 
555 static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
556 			    void *value, bool onallcpus)
557 {
558 	if (!onallcpus) {
559 		/* copy true value_size bytes */
560 		memcpy(this_cpu_ptr(pptr), value, htab->map.value_size);
561 	} else {
562 		u32 size = round_up(htab->map.value_size, 8);
563 		int off = 0, cpu;
564 
565 		for_each_possible_cpu(cpu) {
566 			bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
567 					value + off, size);
568 			off += size;
569 		}
570 	}
571 }
572 
573 static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
574 					 void *value, u32 key_size, u32 hash,
575 					 bool percpu, bool onallcpus,
576 					 bool old_elem_exists)
577 {
578 	u32 size = htab->map.value_size;
579 	bool prealloc = !(htab->map.map_flags & BPF_F_NO_PREALLOC);
580 	struct htab_elem *l_new;
581 	void __percpu *pptr;
582 	int err = 0;
583 
584 	if (prealloc) {
585 		l_new = (struct htab_elem *)pcpu_freelist_pop(&htab->freelist);
586 		if (!l_new)
587 			err = -E2BIG;
588 	} else {
589 		if (atomic_inc_return(&htab->count) > htab->map.max_entries) {
590 			atomic_dec(&htab->count);
591 			err = -E2BIG;
592 		} else {
593 			l_new = kmalloc(htab->elem_size,
594 					GFP_ATOMIC | __GFP_NOWARN);
595 			if (!l_new)
596 				return ERR_PTR(-ENOMEM);
597 		}
598 	}
599 
600 	if (err) {
601 		if (!old_elem_exists)
602 			return ERR_PTR(err);
603 
604 		/* if we're updating the existing element and the hash table
605 		 * is full, use per-cpu extra elems
606 		 */
607 		l_new = this_cpu_ptr(htab->extra_elems);
608 		if (l_new->state != HTAB_EXTRA_ELEM_FREE)
609 			return ERR_PTR(-E2BIG);
610 		l_new->state = HTAB_EXTRA_ELEM_USED;
611 	} else {
612 		l_new->state = HTAB_NOT_AN_EXTRA_ELEM;
613 	}
614 
615 	memcpy(l_new->key, key, key_size);
616 	if (percpu) {
617 		/* round up value_size to 8 bytes */
618 		size = round_up(size, 8);
619 
620 		if (prealloc) {
621 			pptr = htab_elem_get_ptr(l_new, key_size);
622 		} else {
623 			/* alloc_percpu zero-fills */
624 			pptr = __alloc_percpu_gfp(size, 8,
625 						  GFP_ATOMIC | __GFP_NOWARN);
626 			if (!pptr) {
627 				kfree(l_new);
628 				return ERR_PTR(-ENOMEM);
629 			}
630 		}
631 
632 		pcpu_copy_value(htab, pptr, value, onallcpus);
633 
634 		if (!prealloc)
635 			htab_elem_set_ptr(l_new, key_size, pptr);
636 	} else {
637 		memcpy(l_new->key + round_up(key_size, 8), value, size);
638 	}
639 
640 	l_new->hash = hash;
641 	return l_new;
642 }
643 
644 static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
645 		       u64 map_flags)
646 {
647 	if (l_old && map_flags == BPF_NOEXIST)
648 		/* elem already exists */
649 		return -EEXIST;
650 
651 	if (!l_old && map_flags == BPF_EXIST)
652 		/* elem doesn't exist, cannot update it */
653 		return -ENOENT;
654 
655 	return 0;
656 }
657 
658 /* Called from syscall or from eBPF program */
659 static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
660 				u64 map_flags)
661 {
662 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
663 	struct htab_elem *l_new = NULL, *l_old;
664 	struct hlist_head *head;
665 	unsigned long flags;
666 	struct bucket *b;
667 	u32 key_size, hash;
668 	int ret;
669 
670 	if (unlikely(map_flags > BPF_EXIST))
671 		/* unknown flags */
672 		return -EINVAL;
673 
674 	WARN_ON_ONCE(!rcu_read_lock_held());
675 
676 	key_size = map->key_size;
677 
678 	hash = htab_map_hash(key, key_size);
679 
680 	b = __select_bucket(htab, hash);
681 	head = &b->head;
682 
683 	/* bpf_map_update_elem() can be called in_irq() */
684 	raw_spin_lock_irqsave(&b->lock, flags);
685 
686 	l_old = lookup_elem_raw(head, hash, key, key_size);
687 
688 	ret = check_flags(htab, l_old, map_flags);
689 	if (ret)
690 		goto err;
691 
692 	l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
693 				!!l_old);
694 	if (IS_ERR(l_new)) {
695 		/* all pre-allocated elements are in use or memory exhausted */
696 		ret = PTR_ERR(l_new);
697 		goto err;
698 	}
699 
700 	/* add new element to the head of the list, so that
701 	 * concurrent search will find it before old elem
702 	 */
703 	hlist_add_head_rcu(&l_new->hash_node, head);
704 	if (l_old) {
705 		hlist_del_rcu(&l_old->hash_node);
706 		free_htab_elem(htab, l_old);
707 	}
708 	ret = 0;
709 err:
710 	raw_spin_unlock_irqrestore(&b->lock, flags);
711 	return ret;
712 }
713 
714 static int htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
715 				    u64 map_flags)
716 {
717 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
718 	struct htab_elem *l_new, *l_old = NULL;
719 	struct hlist_head *head;
720 	unsigned long flags;
721 	struct bucket *b;
722 	u32 key_size, hash;
723 	int ret;
724 
725 	if (unlikely(map_flags > BPF_EXIST))
726 		/* unknown flags */
727 		return -EINVAL;
728 
729 	WARN_ON_ONCE(!rcu_read_lock_held());
730 
731 	key_size = map->key_size;
732 
733 	hash = htab_map_hash(key, key_size);
734 
735 	b = __select_bucket(htab, hash);
736 	head = &b->head;
737 
738 	/* For LRU, we need to alloc before taking bucket's
739 	 * spinlock because getting free nodes from LRU may need
740 	 * to remove older elements from htab and this removal
741 	 * operation will need a bucket lock.
742 	 */
743 	l_new = prealloc_lru_pop(htab, key, hash);
744 	if (!l_new)
745 		return -ENOMEM;
746 	memcpy(l_new->key + round_up(map->key_size, 8), value, map->value_size);
747 
748 	/* bpf_map_update_elem() can be called in_irq() */
749 	raw_spin_lock_irqsave(&b->lock, flags);
750 
751 	l_old = lookup_elem_raw(head, hash, key, key_size);
752 
753 	ret = check_flags(htab, l_old, map_flags);
754 	if (ret)
755 		goto err;
756 
757 	/* add new element to the head of the list, so that
758 	 * concurrent search will find it before old elem
759 	 */
760 	hlist_add_head_rcu(&l_new->hash_node, head);
761 	if (l_old) {
762 		bpf_lru_node_set_ref(&l_new->lru_node);
763 		hlist_del_rcu(&l_old->hash_node);
764 	}
765 	ret = 0;
766 
767 err:
768 	raw_spin_unlock_irqrestore(&b->lock, flags);
769 
770 	if (ret)
771 		bpf_lru_push_free(&htab->lru, &l_new->lru_node);
772 	else if (l_old)
773 		bpf_lru_push_free(&htab->lru, &l_old->lru_node);
774 
775 	return ret;
776 }
777 
778 static int __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
779 					 void *value, u64 map_flags,
780 					 bool onallcpus)
781 {
782 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
783 	struct htab_elem *l_new = NULL, *l_old;
784 	struct hlist_head *head;
785 	unsigned long flags;
786 	struct bucket *b;
787 	u32 key_size, hash;
788 	int ret;
789 
790 	if (unlikely(map_flags > BPF_EXIST))
791 		/* unknown flags */
792 		return -EINVAL;
793 
794 	WARN_ON_ONCE(!rcu_read_lock_held());
795 
796 	key_size = map->key_size;
797 
798 	hash = htab_map_hash(key, key_size);
799 
800 	b = __select_bucket(htab, hash);
801 	head = &b->head;
802 
803 	/* bpf_map_update_elem() can be called in_irq() */
804 	raw_spin_lock_irqsave(&b->lock, flags);
805 
806 	l_old = lookup_elem_raw(head, hash, key, key_size);
807 
808 	ret = check_flags(htab, l_old, map_flags);
809 	if (ret)
810 		goto err;
811 
812 	if (l_old) {
813 		/* per-cpu hash map can update value in-place */
814 		pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
815 				value, onallcpus);
816 	} else {
817 		l_new = alloc_htab_elem(htab, key, value, key_size,
818 					hash, true, onallcpus, false);
819 		if (IS_ERR(l_new)) {
820 			ret = PTR_ERR(l_new);
821 			goto err;
822 		}
823 		hlist_add_head_rcu(&l_new->hash_node, head);
824 	}
825 	ret = 0;
826 err:
827 	raw_spin_unlock_irqrestore(&b->lock, flags);
828 	return ret;
829 }
830 
831 static int __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
832 					     void *value, u64 map_flags,
833 					     bool onallcpus)
834 {
835 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
836 	struct htab_elem *l_new = NULL, *l_old;
837 	struct hlist_head *head;
838 	unsigned long flags;
839 	struct bucket *b;
840 	u32 key_size, hash;
841 	int ret;
842 
843 	if (unlikely(map_flags > BPF_EXIST))
844 		/* unknown flags */
845 		return -EINVAL;
846 
847 	WARN_ON_ONCE(!rcu_read_lock_held());
848 
849 	key_size = map->key_size;
850 
851 	hash = htab_map_hash(key, key_size);
852 
853 	b = __select_bucket(htab, hash);
854 	head = &b->head;
855 
856 	/* For LRU, we need to alloc before taking bucket's
857 	 * spinlock because LRU's elem alloc may need
858 	 * to remove older elem from htab and this removal
859 	 * operation will need a bucket lock.
860 	 */
861 	if (map_flags != BPF_EXIST) {
862 		l_new = prealloc_lru_pop(htab, key, hash);
863 		if (!l_new)
864 			return -ENOMEM;
865 	}
866 
867 	/* bpf_map_update_elem() can be called in_irq() */
868 	raw_spin_lock_irqsave(&b->lock, flags);
869 
870 	l_old = lookup_elem_raw(head, hash, key, key_size);
871 
872 	ret = check_flags(htab, l_old, map_flags);
873 	if (ret)
874 		goto err;
875 
876 	if (l_old) {
877 		bpf_lru_node_set_ref(&l_old->lru_node);
878 
879 		/* per-cpu hash map can update value in-place */
880 		pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
881 				value, onallcpus);
882 	} else {
883 		pcpu_copy_value(htab, htab_elem_get_ptr(l_new, key_size),
884 				value, onallcpus);
885 		hlist_add_head_rcu(&l_new->hash_node, head);
886 		l_new = NULL;
887 	}
888 	ret = 0;
889 err:
890 	raw_spin_unlock_irqrestore(&b->lock, flags);
891 	if (l_new)
892 		bpf_lru_push_free(&htab->lru, &l_new->lru_node);
893 	return ret;
894 }
895 
896 static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
897 				       void *value, u64 map_flags)
898 {
899 	return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
900 }
901 
902 static int htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
903 					   void *value, u64 map_flags)
904 {
905 	return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
906 						 false);
907 }
908 
909 /* Called from syscall or from eBPF program */
910 static int htab_map_delete_elem(struct bpf_map *map, void *key)
911 {
912 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
913 	struct hlist_head *head;
914 	struct bucket *b;
915 	struct htab_elem *l;
916 	unsigned long flags;
917 	u32 hash, key_size;
918 	int ret = -ENOENT;
919 
920 	WARN_ON_ONCE(!rcu_read_lock_held());
921 
922 	key_size = map->key_size;
923 
924 	hash = htab_map_hash(key, key_size);
925 	b = __select_bucket(htab, hash);
926 	head = &b->head;
927 
928 	raw_spin_lock_irqsave(&b->lock, flags);
929 
930 	l = lookup_elem_raw(head, hash, key, key_size);
931 
932 	if (l) {
933 		hlist_del_rcu(&l->hash_node);
934 		free_htab_elem(htab, l);
935 		ret = 0;
936 	}
937 
938 	raw_spin_unlock_irqrestore(&b->lock, flags);
939 	return ret;
940 }
941 
942 static int htab_lru_map_delete_elem(struct bpf_map *map, void *key)
943 {
944 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
945 	struct hlist_head *head;
946 	struct bucket *b;
947 	struct htab_elem *l;
948 	unsigned long flags;
949 	u32 hash, key_size;
950 	int ret = -ENOENT;
951 
952 	WARN_ON_ONCE(!rcu_read_lock_held());
953 
954 	key_size = map->key_size;
955 
956 	hash = htab_map_hash(key, key_size);
957 	b = __select_bucket(htab, hash);
958 	head = &b->head;
959 
960 	raw_spin_lock_irqsave(&b->lock, flags);
961 
962 	l = lookup_elem_raw(head, hash, key, key_size);
963 
964 	if (l) {
965 		hlist_del_rcu(&l->hash_node);
966 		ret = 0;
967 	}
968 
969 	raw_spin_unlock_irqrestore(&b->lock, flags);
970 	if (l)
971 		bpf_lru_push_free(&htab->lru, &l->lru_node);
972 	return ret;
973 }
974 
975 static void delete_all_elements(struct bpf_htab *htab)
976 {
977 	int i;
978 
979 	for (i = 0; i < htab->n_buckets; i++) {
980 		struct hlist_head *head = select_bucket(htab, i);
981 		struct hlist_node *n;
982 		struct htab_elem *l;
983 
984 		hlist_for_each_entry_safe(l, n, head, hash_node) {
985 			hlist_del_rcu(&l->hash_node);
986 			if (l->state != HTAB_EXTRA_ELEM_USED)
987 				htab_elem_free(htab, l);
988 		}
989 	}
990 }
991 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
992 static void htab_map_free(struct bpf_map *map)
993 {
994 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
995 
996 	/* at this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
997 	 * so the programs (can be more than one that used this map) were
998 	 * disconnected from events. Wait for outstanding critical sections in
999 	 * these programs to complete
1000 	 */
1001 	synchronize_rcu();
1002 
1003 	/* some of free_htab_elem() callbacks for elements of this map may
1004 	 * not have executed. Wait for them.
1005 	 */
1006 	rcu_barrier();
1007 	if (htab->map.map_flags & BPF_F_NO_PREALLOC)
1008 		delete_all_elements(htab);
1009 	else
1010 		prealloc_destroy(htab);
1011 
1012 	free_percpu(htab->extra_elems);
1013 	bpf_map_area_free(htab->buckets);
1014 	kfree(htab);
1015 }
1016 
1017 static const struct bpf_map_ops htab_ops = {
1018 	.map_alloc = htab_map_alloc,
1019 	.map_free = htab_map_free,
1020 	.map_get_next_key = htab_map_get_next_key,
1021 	.map_lookup_elem = htab_map_lookup_elem,
1022 	.map_update_elem = htab_map_update_elem,
1023 	.map_delete_elem = htab_map_delete_elem,
1024 };
1025 
1026 static struct bpf_map_type_list htab_type __ro_after_init = {
1027 	.ops = &htab_ops,
1028 	.type = BPF_MAP_TYPE_HASH,
1029 };
1030 
1031 static const struct bpf_map_ops htab_lru_ops = {
1032 	.map_alloc = htab_map_alloc,
1033 	.map_free = htab_map_free,
1034 	.map_get_next_key = htab_map_get_next_key,
1035 	.map_lookup_elem = htab_lru_map_lookup_elem,
1036 	.map_update_elem = htab_lru_map_update_elem,
1037 	.map_delete_elem = htab_lru_map_delete_elem,
1038 };
1039 
1040 static struct bpf_map_type_list htab_lru_type __ro_after_init = {
1041 	.ops = &htab_lru_ops,
1042 	.type = BPF_MAP_TYPE_LRU_HASH,
1043 };
1044 
1045 /* Called from eBPF program */
1046 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1047 {
1048 	struct htab_elem *l = __htab_map_lookup_elem(map, key);
1049 
1050 	if (l)
1051 		return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1052 	else
1053 		return NULL;
1054 }
1055 
1056 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1057 {
1058 	struct htab_elem *l = __htab_map_lookup_elem(map, key);
1059 
1060 	if (l) {
1061 		bpf_lru_node_set_ref(&l->lru_node);
1062 		return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1063 	}
1064 
1065 	return NULL;
1066 }
1067 
1068 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
1069 {
1070 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1071 	struct htab_elem *l;
1072 	void __percpu *pptr;
1073 	int ret = -ENOENT;
1074 	int cpu, off = 0;
1075 	u32 size;
1076 
1077 	/* per_cpu areas are zero-filled and bpf programs can only
1078 	 * access 'value_size' of them, so copying rounded areas
1079 	 * will not leak any kernel data
1080 	 */
1081 	size = round_up(map->value_size, 8);
1082 	rcu_read_lock();
1083 	l = __htab_map_lookup_elem(map, key);
1084 	if (!l)
1085 		goto out;
1086 	if (htab_is_lru(htab))
1087 		bpf_lru_node_set_ref(&l->lru_node);
1088 	pptr = htab_elem_get_ptr(l, map->key_size);
1089 	for_each_possible_cpu(cpu) {
1090 		bpf_long_memcpy(value + off,
1091 				per_cpu_ptr(pptr, cpu), size);
1092 		off += size;
1093 	}
1094 	ret = 0;
1095 out:
1096 	rcu_read_unlock();
1097 	return ret;
1098 }
1099 
1100 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
1101 			   u64 map_flags)
1102 {
1103 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1104 	int ret;
1105 
1106 	rcu_read_lock();
1107 	if (htab_is_lru(htab))
1108 		ret = __htab_lru_percpu_map_update_elem(map, key, value,
1109 							map_flags, true);
1110 	else
1111 		ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
1112 						    true);
1113 	rcu_read_unlock();
1114 
1115 	return ret;
1116 }
1117 
1118 static const struct bpf_map_ops htab_percpu_ops = {
1119 	.map_alloc = htab_map_alloc,
1120 	.map_free = htab_map_free,
1121 	.map_get_next_key = htab_map_get_next_key,
1122 	.map_lookup_elem = htab_percpu_map_lookup_elem,
1123 	.map_update_elem = htab_percpu_map_update_elem,
1124 	.map_delete_elem = htab_map_delete_elem,
1125 };
1126 
1127 static struct bpf_map_type_list htab_percpu_type __ro_after_init = {
1128 	.ops = &htab_percpu_ops,
1129 	.type = BPF_MAP_TYPE_PERCPU_HASH,
1130 };
1131 
1132 static const struct bpf_map_ops htab_lru_percpu_ops = {
1133 	.map_alloc = htab_map_alloc,
1134 	.map_free = htab_map_free,
1135 	.map_get_next_key = htab_map_get_next_key,
1136 	.map_lookup_elem = htab_lru_percpu_map_lookup_elem,
1137 	.map_update_elem = htab_lru_percpu_map_update_elem,
1138 	.map_delete_elem = htab_lru_map_delete_elem,
1139 };
1140 
1141 static struct bpf_map_type_list htab_lru_percpu_type __ro_after_init = {
1142 	.ops = &htab_lru_percpu_ops,
1143 	.type = BPF_MAP_TYPE_LRU_PERCPU_HASH,
1144 };
1145 
1146 static int __init register_htab_map(void)
1147 {
1148 	bpf_register_map_type(&htab_type);
1149 	bpf_register_map_type(&htab_percpu_type);
1150 	bpf_register_map_type(&htab_lru_type);
1151 	bpf_register_map_type(&htab_lru_percpu_type);
1152 	return 0;
1153 }
1154 late_initcall(register_htab_map);
1155