xref: /linux/kernel/bpf/hashtab.c (revision 1f2367a39f17bd553a75e179a747f9b257bc9478)
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/btf.h>
15 #include <linux/jhash.h>
16 #include <linux/filter.h>
17 #include <linux/rculist_nulls.h>
18 #include <linux/random.h>
19 #include <uapi/linux/btf.h>
20 #include "percpu_freelist.h"
21 #include "bpf_lru_list.h"
22 #include "map_in_map.h"
23 
24 #define HTAB_CREATE_FLAG_MASK						\
25 	(BPF_F_NO_PREALLOC | BPF_F_NO_COMMON_LRU | BPF_F_NUMA_NODE |	\
26 	 BPF_F_RDONLY | BPF_F_WRONLY | BPF_F_ZERO_SEED)
27 
28 struct bucket {
29 	struct hlist_nulls_head head;
30 	raw_spinlock_t lock;
31 };
32 
33 struct bpf_htab {
34 	struct bpf_map map;
35 	struct bucket *buckets;
36 	void *elems;
37 	union {
38 		struct pcpu_freelist freelist;
39 		struct bpf_lru lru;
40 	};
41 	struct htab_elem *__percpu *extra_elems;
42 	atomic_t count;	/* number of elements in this hashtable */
43 	u32 n_buckets;	/* number of hash buckets */
44 	u32 elem_size;	/* size of each element in bytes */
45 	u32 hashrnd;
46 };
47 
48 /* each htab element is struct htab_elem + key + value */
49 struct htab_elem {
50 	union {
51 		struct hlist_nulls_node hash_node;
52 		struct {
53 			void *padding;
54 			union {
55 				struct bpf_htab *htab;
56 				struct pcpu_freelist_node fnode;
57 			};
58 		};
59 	};
60 	union {
61 		struct rcu_head rcu;
62 		struct bpf_lru_node lru_node;
63 	};
64 	u32 hash;
65 	char key[0] __aligned(8);
66 };
67 
68 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
69 
70 static bool htab_is_lru(const struct bpf_htab *htab)
71 {
72 	return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH ||
73 		htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
74 }
75 
76 static bool htab_is_percpu(const struct bpf_htab *htab)
77 {
78 	return htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH ||
79 		htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
80 }
81 
82 static bool htab_is_prealloc(const struct bpf_htab *htab)
83 {
84 	return !(htab->map.map_flags & BPF_F_NO_PREALLOC);
85 }
86 
87 static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
88 				     void __percpu *pptr)
89 {
90 	*(void __percpu **)(l->key + key_size) = pptr;
91 }
92 
93 static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
94 {
95 	return *(void __percpu **)(l->key + key_size);
96 }
97 
98 static void *fd_htab_map_get_ptr(const struct bpf_map *map, struct htab_elem *l)
99 {
100 	return *(void **)(l->key + roundup(map->key_size, 8));
101 }
102 
103 static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
104 {
105 	return (struct htab_elem *) (htab->elems + i * htab->elem_size);
106 }
107 
108 static void htab_free_elems(struct bpf_htab *htab)
109 {
110 	int i;
111 
112 	if (!htab_is_percpu(htab))
113 		goto free_elems;
114 
115 	for (i = 0; i < htab->map.max_entries; i++) {
116 		void __percpu *pptr;
117 
118 		pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
119 					 htab->map.key_size);
120 		free_percpu(pptr);
121 		cond_resched();
122 	}
123 free_elems:
124 	bpf_map_area_free(htab->elems);
125 }
126 
127 static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
128 					  u32 hash)
129 {
130 	struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
131 	struct htab_elem *l;
132 
133 	if (node) {
134 		l = container_of(node, struct htab_elem, lru_node);
135 		memcpy(l->key, key, htab->map.key_size);
136 		return l;
137 	}
138 
139 	return NULL;
140 }
141 
142 static int prealloc_init(struct bpf_htab *htab)
143 {
144 	u32 num_entries = htab->map.max_entries;
145 	int err = -ENOMEM, i;
146 
147 	if (!htab_is_percpu(htab) && !htab_is_lru(htab))
148 		num_entries += num_possible_cpus();
149 
150 	htab->elems = bpf_map_area_alloc(htab->elem_size * num_entries,
151 					 htab->map.numa_node);
152 	if (!htab->elems)
153 		return -ENOMEM;
154 
155 	if (!htab_is_percpu(htab))
156 		goto skip_percpu_elems;
157 
158 	for (i = 0; i < num_entries; i++) {
159 		u32 size = round_up(htab->map.value_size, 8);
160 		void __percpu *pptr;
161 
162 		pptr = __alloc_percpu_gfp(size, 8, GFP_USER | __GFP_NOWARN);
163 		if (!pptr)
164 			goto free_elems;
165 		htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
166 				  pptr);
167 		cond_resched();
168 	}
169 
170 skip_percpu_elems:
171 	if (htab_is_lru(htab))
172 		err = bpf_lru_init(&htab->lru,
173 				   htab->map.map_flags & BPF_F_NO_COMMON_LRU,
174 				   offsetof(struct htab_elem, hash) -
175 				   offsetof(struct htab_elem, lru_node),
176 				   htab_lru_map_delete_node,
177 				   htab);
178 	else
179 		err = pcpu_freelist_init(&htab->freelist);
180 
181 	if (err)
182 		goto free_elems;
183 
184 	if (htab_is_lru(htab))
185 		bpf_lru_populate(&htab->lru, htab->elems,
186 				 offsetof(struct htab_elem, lru_node),
187 				 htab->elem_size, num_entries);
188 	else
189 		pcpu_freelist_populate(&htab->freelist,
190 				       htab->elems + offsetof(struct htab_elem, fnode),
191 				       htab->elem_size, num_entries);
192 
193 	return 0;
194 
195 free_elems:
196 	htab_free_elems(htab);
197 	return err;
198 }
199 
200 static void prealloc_destroy(struct bpf_htab *htab)
201 {
202 	htab_free_elems(htab);
203 
204 	if (htab_is_lru(htab))
205 		bpf_lru_destroy(&htab->lru);
206 	else
207 		pcpu_freelist_destroy(&htab->freelist);
208 }
209 
210 static int alloc_extra_elems(struct bpf_htab *htab)
211 {
212 	struct htab_elem *__percpu *pptr, *l_new;
213 	struct pcpu_freelist_node *l;
214 	int cpu;
215 
216 	pptr = __alloc_percpu_gfp(sizeof(struct htab_elem *), 8,
217 				  GFP_USER | __GFP_NOWARN);
218 	if (!pptr)
219 		return -ENOMEM;
220 
221 	for_each_possible_cpu(cpu) {
222 		l = pcpu_freelist_pop(&htab->freelist);
223 		/* pop will succeed, since prealloc_init()
224 		 * preallocated extra num_possible_cpus elements
225 		 */
226 		l_new = container_of(l, struct htab_elem, fnode);
227 		*per_cpu_ptr(pptr, cpu) = l_new;
228 	}
229 	htab->extra_elems = pptr;
230 	return 0;
231 }
232 
233 /* Called from syscall */
234 static int htab_map_alloc_check(union bpf_attr *attr)
235 {
236 	bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
237 		       attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
238 	bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
239 		    attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
240 	/* percpu_lru means each cpu has its own LRU list.
241 	 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
242 	 * the map's value itself is percpu.  percpu_lru has
243 	 * nothing to do with the map's value.
244 	 */
245 	bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
246 	bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
247 	bool zero_seed = (attr->map_flags & BPF_F_ZERO_SEED);
248 	int numa_node = bpf_map_attr_numa_node(attr);
249 
250 	BUILD_BUG_ON(offsetof(struct htab_elem, htab) !=
251 		     offsetof(struct htab_elem, hash_node.pprev));
252 	BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
253 		     offsetof(struct htab_elem, hash_node.pprev));
254 
255 	if (lru && !capable(CAP_SYS_ADMIN))
256 		/* LRU implementation is much complicated than other
257 		 * maps.  Hence, limit to CAP_SYS_ADMIN for now.
258 		 */
259 		return -EPERM;
260 
261 	if (zero_seed && !capable(CAP_SYS_ADMIN))
262 		/* Guard against local DoS, and discourage production use. */
263 		return -EPERM;
264 
265 	if (attr->map_flags & ~HTAB_CREATE_FLAG_MASK)
266 		/* reserved bits should not be used */
267 		return -EINVAL;
268 
269 	if (!lru && percpu_lru)
270 		return -EINVAL;
271 
272 	if (lru && !prealloc)
273 		return -ENOTSUPP;
274 
275 	if (numa_node != NUMA_NO_NODE && (percpu || percpu_lru))
276 		return -EINVAL;
277 
278 	/* check sanity of attributes.
279 	 * value_size == 0 may be allowed in the future to use map as a set
280 	 */
281 	if (attr->max_entries == 0 || attr->key_size == 0 ||
282 	    attr->value_size == 0)
283 		return -EINVAL;
284 
285 	if (attr->key_size > MAX_BPF_STACK)
286 		/* eBPF programs initialize keys on stack, so they cannot be
287 		 * larger than max stack size
288 		 */
289 		return -E2BIG;
290 
291 	if (attr->value_size >= KMALLOC_MAX_SIZE -
292 	    MAX_BPF_STACK - sizeof(struct htab_elem))
293 		/* if value_size is bigger, the user space won't be able to
294 		 * access the elements via bpf syscall. This check also makes
295 		 * sure that the elem_size doesn't overflow and it's
296 		 * kmalloc-able later in htab_map_update_elem()
297 		 */
298 		return -E2BIG;
299 
300 	return 0;
301 }
302 
303 static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
304 {
305 	bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
306 		       attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
307 	bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
308 		    attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
309 	/* percpu_lru means each cpu has its own LRU list.
310 	 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
311 	 * the map's value itself is percpu.  percpu_lru has
312 	 * nothing to do with the map's value.
313 	 */
314 	bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
315 	bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
316 	struct bpf_htab *htab;
317 	int err, i;
318 	u64 cost;
319 
320 	htab = kzalloc(sizeof(*htab), GFP_USER);
321 	if (!htab)
322 		return ERR_PTR(-ENOMEM);
323 
324 	bpf_map_init_from_attr(&htab->map, attr);
325 
326 	if (percpu_lru) {
327 		/* ensure each CPU's lru list has >=1 elements.
328 		 * since we are at it, make each lru list has the same
329 		 * number of elements.
330 		 */
331 		htab->map.max_entries = roundup(attr->max_entries,
332 						num_possible_cpus());
333 		if (htab->map.max_entries < attr->max_entries)
334 			htab->map.max_entries = rounddown(attr->max_entries,
335 							  num_possible_cpus());
336 	}
337 
338 	/* hash table size must be power of 2 */
339 	htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
340 
341 	htab->elem_size = sizeof(struct htab_elem) +
342 			  round_up(htab->map.key_size, 8);
343 	if (percpu)
344 		htab->elem_size += sizeof(void *);
345 	else
346 		htab->elem_size += round_up(htab->map.value_size, 8);
347 
348 	err = -E2BIG;
349 	/* prevent zero size kmalloc and check for u32 overflow */
350 	if (htab->n_buckets == 0 ||
351 	    htab->n_buckets > U32_MAX / sizeof(struct bucket))
352 		goto free_htab;
353 
354 	cost = (u64) htab->n_buckets * sizeof(struct bucket) +
355 	       (u64) htab->elem_size * htab->map.max_entries;
356 
357 	if (percpu)
358 		cost += (u64) round_up(htab->map.value_size, 8) *
359 			num_possible_cpus() * htab->map.max_entries;
360 	else
361 	       cost += (u64) htab->elem_size * num_possible_cpus();
362 
363 	if (cost >= U32_MAX - PAGE_SIZE)
364 		/* make sure page count doesn't overflow */
365 		goto free_htab;
366 
367 	htab->map.pages = round_up(cost, PAGE_SIZE) >> PAGE_SHIFT;
368 
369 	/* if map size is larger than memlock limit, reject it early */
370 	err = bpf_map_precharge_memlock(htab->map.pages);
371 	if (err)
372 		goto free_htab;
373 
374 	err = -ENOMEM;
375 	htab->buckets = bpf_map_area_alloc(htab->n_buckets *
376 					   sizeof(struct bucket),
377 					   htab->map.numa_node);
378 	if (!htab->buckets)
379 		goto free_htab;
380 
381 	if (htab->map.map_flags & BPF_F_ZERO_SEED)
382 		htab->hashrnd = 0;
383 	else
384 		htab->hashrnd = get_random_int();
385 
386 	for (i = 0; i < htab->n_buckets; i++) {
387 		INIT_HLIST_NULLS_HEAD(&htab->buckets[i].head, i);
388 		raw_spin_lock_init(&htab->buckets[i].lock);
389 	}
390 
391 	if (prealloc) {
392 		err = prealloc_init(htab);
393 		if (err)
394 			goto free_buckets;
395 
396 		if (!percpu && !lru) {
397 			/* lru itself can remove the least used element, so
398 			 * there is no need for an extra elem during map_update.
399 			 */
400 			err = alloc_extra_elems(htab);
401 			if (err)
402 				goto free_prealloc;
403 		}
404 	}
405 
406 	return &htab->map;
407 
408 free_prealloc:
409 	prealloc_destroy(htab);
410 free_buckets:
411 	bpf_map_area_free(htab->buckets);
412 free_htab:
413 	kfree(htab);
414 	return ERR_PTR(err);
415 }
416 
417 static inline u32 htab_map_hash(const void *key, u32 key_len, u32 hashrnd)
418 {
419 	return jhash(key, key_len, hashrnd);
420 }
421 
422 static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
423 {
424 	return &htab->buckets[hash & (htab->n_buckets - 1)];
425 }
426 
427 static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
428 {
429 	return &__select_bucket(htab, hash)->head;
430 }
431 
432 /* this lookup function can only be called with bucket lock taken */
433 static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
434 					 void *key, u32 key_size)
435 {
436 	struct hlist_nulls_node *n;
437 	struct htab_elem *l;
438 
439 	hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
440 		if (l->hash == hash && !memcmp(&l->key, key, key_size))
441 			return l;
442 
443 	return NULL;
444 }
445 
446 /* can be called without bucket lock. it will repeat the loop in
447  * the unlikely event when elements moved from one bucket into another
448  * while link list is being walked
449  */
450 static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
451 					       u32 hash, void *key,
452 					       u32 key_size, u32 n_buckets)
453 {
454 	struct hlist_nulls_node *n;
455 	struct htab_elem *l;
456 
457 again:
458 	hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
459 		if (l->hash == hash && !memcmp(&l->key, key, key_size))
460 			return l;
461 
462 	if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
463 		goto again;
464 
465 	return NULL;
466 }
467 
468 /* Called from syscall or from eBPF program directly, so
469  * arguments have to match bpf_map_lookup_elem() exactly.
470  * The return value is adjusted by BPF instructions
471  * in htab_map_gen_lookup().
472  */
473 static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
474 {
475 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
476 	struct hlist_nulls_head *head;
477 	struct htab_elem *l;
478 	u32 hash, key_size;
479 
480 	/* Must be called with rcu_read_lock. */
481 	WARN_ON_ONCE(!rcu_read_lock_held());
482 
483 	key_size = map->key_size;
484 
485 	hash = htab_map_hash(key, key_size, htab->hashrnd);
486 
487 	head = select_bucket(htab, hash);
488 
489 	l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
490 
491 	return l;
492 }
493 
494 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
495 {
496 	struct htab_elem *l = __htab_map_lookup_elem(map, key);
497 
498 	if (l)
499 		return l->key + round_up(map->key_size, 8);
500 
501 	return NULL;
502 }
503 
504 /* inline bpf_map_lookup_elem() call.
505  * Instead of:
506  * bpf_prog
507  *   bpf_map_lookup_elem
508  *     map->ops->map_lookup_elem
509  *       htab_map_lookup_elem
510  *         __htab_map_lookup_elem
511  * do:
512  * bpf_prog
513  *   __htab_map_lookup_elem
514  */
515 static u32 htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
516 {
517 	struct bpf_insn *insn = insn_buf;
518 	const int ret = BPF_REG_0;
519 
520 	BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
521 		     (void *(*)(struct bpf_map *map, void *key))NULL));
522 	*insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
523 	*insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
524 	*insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
525 				offsetof(struct htab_elem, key) +
526 				round_up(map->key_size, 8));
527 	return insn - insn_buf;
528 }
529 
530 static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
531 {
532 	struct htab_elem *l = __htab_map_lookup_elem(map, key);
533 
534 	if (l) {
535 		bpf_lru_node_set_ref(&l->lru_node);
536 		return l->key + round_up(map->key_size, 8);
537 	}
538 
539 	return NULL;
540 }
541 
542 static u32 htab_lru_map_gen_lookup(struct bpf_map *map,
543 				   struct bpf_insn *insn_buf)
544 {
545 	struct bpf_insn *insn = insn_buf;
546 	const int ret = BPF_REG_0;
547 	const int ref_reg = BPF_REG_1;
548 
549 	BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
550 		     (void *(*)(struct bpf_map *map, void *key))NULL));
551 	*insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
552 	*insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
553 	*insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
554 			      offsetof(struct htab_elem, lru_node) +
555 			      offsetof(struct bpf_lru_node, ref));
556 	*insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
557 	*insn++ = BPF_ST_MEM(BPF_B, ret,
558 			     offsetof(struct htab_elem, lru_node) +
559 			     offsetof(struct bpf_lru_node, ref),
560 			     1);
561 	*insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
562 				offsetof(struct htab_elem, key) +
563 				round_up(map->key_size, 8));
564 	return insn - insn_buf;
565 }
566 
567 /* It is called from the bpf_lru_list when the LRU needs to delete
568  * older elements from the htab.
569  */
570 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
571 {
572 	struct bpf_htab *htab = (struct bpf_htab *)arg;
573 	struct htab_elem *l = NULL, *tgt_l;
574 	struct hlist_nulls_head *head;
575 	struct hlist_nulls_node *n;
576 	unsigned long flags;
577 	struct bucket *b;
578 
579 	tgt_l = container_of(node, struct htab_elem, lru_node);
580 	b = __select_bucket(htab, tgt_l->hash);
581 	head = &b->head;
582 
583 	raw_spin_lock_irqsave(&b->lock, flags);
584 
585 	hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
586 		if (l == tgt_l) {
587 			hlist_nulls_del_rcu(&l->hash_node);
588 			break;
589 		}
590 
591 	raw_spin_unlock_irqrestore(&b->lock, flags);
592 
593 	return l == tgt_l;
594 }
595 
596 /* Called from syscall */
597 static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
598 {
599 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
600 	struct hlist_nulls_head *head;
601 	struct htab_elem *l, *next_l;
602 	u32 hash, key_size;
603 	int i = 0;
604 
605 	WARN_ON_ONCE(!rcu_read_lock_held());
606 
607 	key_size = map->key_size;
608 
609 	if (!key)
610 		goto find_first_elem;
611 
612 	hash = htab_map_hash(key, key_size, htab->hashrnd);
613 
614 	head = select_bucket(htab, hash);
615 
616 	/* lookup the key */
617 	l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
618 
619 	if (!l)
620 		goto find_first_elem;
621 
622 	/* key was found, get next key in the same bucket */
623 	next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
624 				  struct htab_elem, hash_node);
625 
626 	if (next_l) {
627 		/* if next elem in this hash list is non-zero, just return it */
628 		memcpy(next_key, next_l->key, key_size);
629 		return 0;
630 	}
631 
632 	/* no more elements in this hash list, go to the next bucket */
633 	i = hash & (htab->n_buckets - 1);
634 	i++;
635 
636 find_first_elem:
637 	/* iterate over buckets */
638 	for (; i < htab->n_buckets; i++) {
639 		head = select_bucket(htab, i);
640 
641 		/* pick first element in the bucket */
642 		next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
643 					  struct htab_elem, hash_node);
644 		if (next_l) {
645 			/* if it's not empty, just return it */
646 			memcpy(next_key, next_l->key, key_size);
647 			return 0;
648 		}
649 	}
650 
651 	/* iterated over all buckets and all elements */
652 	return -ENOENT;
653 }
654 
655 static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
656 {
657 	if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
658 		free_percpu(htab_elem_get_ptr(l, htab->map.key_size));
659 	kfree(l);
660 }
661 
662 static void htab_elem_free_rcu(struct rcu_head *head)
663 {
664 	struct htab_elem *l = container_of(head, struct htab_elem, rcu);
665 	struct bpf_htab *htab = l->htab;
666 
667 	/* must increment bpf_prog_active to avoid kprobe+bpf triggering while
668 	 * we're calling kfree, otherwise deadlock is possible if kprobes
669 	 * are placed somewhere inside of slub
670 	 */
671 	preempt_disable();
672 	__this_cpu_inc(bpf_prog_active);
673 	htab_elem_free(htab, l);
674 	__this_cpu_dec(bpf_prog_active);
675 	preempt_enable();
676 }
677 
678 static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
679 {
680 	struct bpf_map *map = &htab->map;
681 
682 	if (map->ops->map_fd_put_ptr) {
683 		void *ptr = fd_htab_map_get_ptr(map, l);
684 
685 		map->ops->map_fd_put_ptr(ptr);
686 	}
687 
688 	if (htab_is_prealloc(htab)) {
689 		__pcpu_freelist_push(&htab->freelist, &l->fnode);
690 	} else {
691 		atomic_dec(&htab->count);
692 		l->htab = htab;
693 		call_rcu(&l->rcu, htab_elem_free_rcu);
694 	}
695 }
696 
697 static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
698 			    void *value, bool onallcpus)
699 {
700 	if (!onallcpus) {
701 		/* copy true value_size bytes */
702 		memcpy(this_cpu_ptr(pptr), value, htab->map.value_size);
703 	} else {
704 		u32 size = round_up(htab->map.value_size, 8);
705 		int off = 0, cpu;
706 
707 		for_each_possible_cpu(cpu) {
708 			bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
709 					value + off, size);
710 			off += size;
711 		}
712 	}
713 }
714 
715 static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
716 {
717 	return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
718 	       BITS_PER_LONG == 64;
719 }
720 
721 static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
722 					 void *value, u32 key_size, u32 hash,
723 					 bool percpu, bool onallcpus,
724 					 struct htab_elem *old_elem)
725 {
726 	u32 size = htab->map.value_size;
727 	bool prealloc = htab_is_prealloc(htab);
728 	struct htab_elem *l_new, **pl_new;
729 	void __percpu *pptr;
730 
731 	if (prealloc) {
732 		if (old_elem) {
733 			/* if we're updating the existing element,
734 			 * use per-cpu extra elems to avoid freelist_pop/push
735 			 */
736 			pl_new = this_cpu_ptr(htab->extra_elems);
737 			l_new = *pl_new;
738 			*pl_new = old_elem;
739 		} else {
740 			struct pcpu_freelist_node *l;
741 
742 			l = __pcpu_freelist_pop(&htab->freelist);
743 			if (!l)
744 				return ERR_PTR(-E2BIG);
745 			l_new = container_of(l, struct htab_elem, fnode);
746 		}
747 	} else {
748 		if (atomic_inc_return(&htab->count) > htab->map.max_entries)
749 			if (!old_elem) {
750 				/* when map is full and update() is replacing
751 				 * old element, it's ok to allocate, since
752 				 * old element will be freed immediately.
753 				 * Otherwise return an error
754 				 */
755 				l_new = ERR_PTR(-E2BIG);
756 				goto dec_count;
757 			}
758 		l_new = kmalloc_node(htab->elem_size, GFP_ATOMIC | __GFP_NOWARN,
759 				     htab->map.numa_node);
760 		if (!l_new) {
761 			l_new = ERR_PTR(-ENOMEM);
762 			goto dec_count;
763 		}
764 		check_and_init_map_lock(&htab->map,
765 					l_new->key + round_up(key_size, 8));
766 	}
767 
768 	memcpy(l_new->key, key, key_size);
769 	if (percpu) {
770 		size = round_up(size, 8);
771 		if (prealloc) {
772 			pptr = htab_elem_get_ptr(l_new, key_size);
773 		} else {
774 			/* alloc_percpu zero-fills */
775 			pptr = __alloc_percpu_gfp(size, 8,
776 						  GFP_ATOMIC | __GFP_NOWARN);
777 			if (!pptr) {
778 				kfree(l_new);
779 				l_new = ERR_PTR(-ENOMEM);
780 				goto dec_count;
781 			}
782 		}
783 
784 		pcpu_copy_value(htab, pptr, value, onallcpus);
785 
786 		if (!prealloc)
787 			htab_elem_set_ptr(l_new, key_size, pptr);
788 	} else if (fd_htab_map_needs_adjust(htab)) {
789 		size = round_up(size, 8);
790 		memcpy(l_new->key + round_up(key_size, 8), value, size);
791 	} else {
792 		copy_map_value(&htab->map,
793 			       l_new->key + round_up(key_size, 8),
794 			       value);
795 	}
796 
797 	l_new->hash = hash;
798 	return l_new;
799 dec_count:
800 	atomic_dec(&htab->count);
801 	return l_new;
802 }
803 
804 static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
805 		       u64 map_flags)
806 {
807 	if (l_old && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
808 		/* elem already exists */
809 		return -EEXIST;
810 
811 	if (!l_old && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
812 		/* elem doesn't exist, cannot update it */
813 		return -ENOENT;
814 
815 	return 0;
816 }
817 
818 /* Called from syscall or from eBPF program */
819 static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
820 				u64 map_flags)
821 {
822 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
823 	struct htab_elem *l_new = NULL, *l_old;
824 	struct hlist_nulls_head *head;
825 	unsigned long flags;
826 	struct bucket *b;
827 	u32 key_size, hash;
828 	int ret;
829 
830 	if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST))
831 		/* unknown flags */
832 		return -EINVAL;
833 
834 	WARN_ON_ONCE(!rcu_read_lock_held());
835 
836 	key_size = map->key_size;
837 
838 	hash = htab_map_hash(key, key_size, htab->hashrnd);
839 
840 	b = __select_bucket(htab, hash);
841 	head = &b->head;
842 
843 	if (unlikely(map_flags & BPF_F_LOCK)) {
844 		if (unlikely(!map_value_has_spin_lock(map)))
845 			return -EINVAL;
846 		/* find an element without taking the bucket lock */
847 		l_old = lookup_nulls_elem_raw(head, hash, key, key_size,
848 					      htab->n_buckets);
849 		ret = check_flags(htab, l_old, map_flags);
850 		if (ret)
851 			return ret;
852 		if (l_old) {
853 			/* grab the element lock and update value in place */
854 			copy_map_value_locked(map,
855 					      l_old->key + round_up(key_size, 8),
856 					      value, false);
857 			return 0;
858 		}
859 		/* fall through, grab the bucket lock and lookup again.
860 		 * 99.9% chance that the element won't be found,
861 		 * but second lookup under lock has to be done.
862 		 */
863 	}
864 
865 	/* bpf_map_update_elem() can be called in_irq() */
866 	raw_spin_lock_irqsave(&b->lock, flags);
867 
868 	l_old = lookup_elem_raw(head, hash, key, key_size);
869 
870 	ret = check_flags(htab, l_old, map_flags);
871 	if (ret)
872 		goto err;
873 
874 	if (unlikely(l_old && (map_flags & BPF_F_LOCK))) {
875 		/* first lookup without the bucket lock didn't find the element,
876 		 * but second lookup with the bucket lock found it.
877 		 * This case is highly unlikely, but has to be dealt with:
878 		 * grab the element lock in addition to the bucket lock
879 		 * and update element in place
880 		 */
881 		copy_map_value_locked(map,
882 				      l_old->key + round_up(key_size, 8),
883 				      value, false);
884 		ret = 0;
885 		goto err;
886 	}
887 
888 	l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
889 				l_old);
890 	if (IS_ERR(l_new)) {
891 		/* all pre-allocated elements are in use or memory exhausted */
892 		ret = PTR_ERR(l_new);
893 		goto err;
894 	}
895 
896 	/* add new element to the head of the list, so that
897 	 * concurrent search will find it before old elem
898 	 */
899 	hlist_nulls_add_head_rcu(&l_new->hash_node, head);
900 	if (l_old) {
901 		hlist_nulls_del_rcu(&l_old->hash_node);
902 		if (!htab_is_prealloc(htab))
903 			free_htab_elem(htab, l_old);
904 	}
905 	ret = 0;
906 err:
907 	raw_spin_unlock_irqrestore(&b->lock, flags);
908 	return ret;
909 }
910 
911 static int htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
912 				    u64 map_flags)
913 {
914 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
915 	struct htab_elem *l_new, *l_old = NULL;
916 	struct hlist_nulls_head *head;
917 	unsigned long flags;
918 	struct bucket *b;
919 	u32 key_size, hash;
920 	int ret;
921 
922 	if (unlikely(map_flags > BPF_EXIST))
923 		/* unknown flags */
924 		return -EINVAL;
925 
926 	WARN_ON_ONCE(!rcu_read_lock_held());
927 
928 	key_size = map->key_size;
929 
930 	hash = htab_map_hash(key, key_size, htab->hashrnd);
931 
932 	b = __select_bucket(htab, hash);
933 	head = &b->head;
934 
935 	/* For LRU, we need to alloc before taking bucket's
936 	 * spinlock because getting free nodes from LRU may need
937 	 * to remove older elements from htab and this removal
938 	 * operation will need a bucket lock.
939 	 */
940 	l_new = prealloc_lru_pop(htab, key, hash);
941 	if (!l_new)
942 		return -ENOMEM;
943 	memcpy(l_new->key + round_up(map->key_size, 8), value, map->value_size);
944 
945 	/* bpf_map_update_elem() can be called in_irq() */
946 	raw_spin_lock_irqsave(&b->lock, flags);
947 
948 	l_old = lookup_elem_raw(head, hash, key, key_size);
949 
950 	ret = check_flags(htab, l_old, map_flags);
951 	if (ret)
952 		goto err;
953 
954 	/* add new element to the head of the list, so that
955 	 * concurrent search will find it before old elem
956 	 */
957 	hlist_nulls_add_head_rcu(&l_new->hash_node, head);
958 	if (l_old) {
959 		bpf_lru_node_set_ref(&l_new->lru_node);
960 		hlist_nulls_del_rcu(&l_old->hash_node);
961 	}
962 	ret = 0;
963 
964 err:
965 	raw_spin_unlock_irqrestore(&b->lock, flags);
966 
967 	if (ret)
968 		bpf_lru_push_free(&htab->lru, &l_new->lru_node);
969 	else if (l_old)
970 		bpf_lru_push_free(&htab->lru, &l_old->lru_node);
971 
972 	return ret;
973 }
974 
975 static int __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
976 					 void *value, u64 map_flags,
977 					 bool onallcpus)
978 {
979 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
980 	struct htab_elem *l_new = NULL, *l_old;
981 	struct hlist_nulls_head *head;
982 	unsigned long flags;
983 	struct bucket *b;
984 	u32 key_size, hash;
985 	int ret;
986 
987 	if (unlikely(map_flags > BPF_EXIST))
988 		/* unknown flags */
989 		return -EINVAL;
990 
991 	WARN_ON_ONCE(!rcu_read_lock_held());
992 
993 	key_size = map->key_size;
994 
995 	hash = htab_map_hash(key, key_size, htab->hashrnd);
996 
997 	b = __select_bucket(htab, hash);
998 	head = &b->head;
999 
1000 	/* bpf_map_update_elem() can be called in_irq() */
1001 	raw_spin_lock_irqsave(&b->lock, flags);
1002 
1003 	l_old = lookup_elem_raw(head, hash, key, key_size);
1004 
1005 	ret = check_flags(htab, l_old, map_flags);
1006 	if (ret)
1007 		goto err;
1008 
1009 	if (l_old) {
1010 		/* per-cpu hash map can update value in-place */
1011 		pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1012 				value, onallcpus);
1013 	} else {
1014 		l_new = alloc_htab_elem(htab, key, value, key_size,
1015 					hash, true, onallcpus, NULL);
1016 		if (IS_ERR(l_new)) {
1017 			ret = PTR_ERR(l_new);
1018 			goto err;
1019 		}
1020 		hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1021 	}
1022 	ret = 0;
1023 err:
1024 	raw_spin_unlock_irqrestore(&b->lock, flags);
1025 	return ret;
1026 }
1027 
1028 static int __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1029 					     void *value, u64 map_flags,
1030 					     bool onallcpus)
1031 {
1032 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1033 	struct htab_elem *l_new = NULL, *l_old;
1034 	struct hlist_nulls_head *head;
1035 	unsigned long flags;
1036 	struct bucket *b;
1037 	u32 key_size, hash;
1038 	int ret;
1039 
1040 	if (unlikely(map_flags > BPF_EXIST))
1041 		/* unknown flags */
1042 		return -EINVAL;
1043 
1044 	WARN_ON_ONCE(!rcu_read_lock_held());
1045 
1046 	key_size = map->key_size;
1047 
1048 	hash = htab_map_hash(key, key_size, htab->hashrnd);
1049 
1050 	b = __select_bucket(htab, hash);
1051 	head = &b->head;
1052 
1053 	/* For LRU, we need to alloc before taking bucket's
1054 	 * spinlock because LRU's elem alloc may need
1055 	 * to remove older elem from htab and this removal
1056 	 * operation will need a bucket lock.
1057 	 */
1058 	if (map_flags != BPF_EXIST) {
1059 		l_new = prealloc_lru_pop(htab, key, hash);
1060 		if (!l_new)
1061 			return -ENOMEM;
1062 	}
1063 
1064 	/* bpf_map_update_elem() can be called in_irq() */
1065 	raw_spin_lock_irqsave(&b->lock, flags);
1066 
1067 	l_old = lookup_elem_raw(head, hash, key, key_size);
1068 
1069 	ret = check_flags(htab, l_old, map_flags);
1070 	if (ret)
1071 		goto err;
1072 
1073 	if (l_old) {
1074 		bpf_lru_node_set_ref(&l_old->lru_node);
1075 
1076 		/* per-cpu hash map can update value in-place */
1077 		pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1078 				value, onallcpus);
1079 	} else {
1080 		pcpu_copy_value(htab, htab_elem_get_ptr(l_new, key_size),
1081 				value, onallcpus);
1082 		hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1083 		l_new = NULL;
1084 	}
1085 	ret = 0;
1086 err:
1087 	raw_spin_unlock_irqrestore(&b->lock, flags);
1088 	if (l_new)
1089 		bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1090 	return ret;
1091 }
1092 
1093 static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1094 				       void *value, u64 map_flags)
1095 {
1096 	return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
1097 }
1098 
1099 static int htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1100 					   void *value, u64 map_flags)
1101 {
1102 	return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1103 						 false);
1104 }
1105 
1106 /* Called from syscall or from eBPF program */
1107 static int htab_map_delete_elem(struct bpf_map *map, void *key)
1108 {
1109 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1110 	struct hlist_nulls_head *head;
1111 	struct bucket *b;
1112 	struct htab_elem *l;
1113 	unsigned long flags;
1114 	u32 hash, key_size;
1115 	int ret = -ENOENT;
1116 
1117 	WARN_ON_ONCE(!rcu_read_lock_held());
1118 
1119 	key_size = map->key_size;
1120 
1121 	hash = htab_map_hash(key, key_size, htab->hashrnd);
1122 	b = __select_bucket(htab, hash);
1123 	head = &b->head;
1124 
1125 	raw_spin_lock_irqsave(&b->lock, flags);
1126 
1127 	l = lookup_elem_raw(head, hash, key, key_size);
1128 
1129 	if (l) {
1130 		hlist_nulls_del_rcu(&l->hash_node);
1131 		free_htab_elem(htab, l);
1132 		ret = 0;
1133 	}
1134 
1135 	raw_spin_unlock_irqrestore(&b->lock, flags);
1136 	return ret;
1137 }
1138 
1139 static int htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1140 {
1141 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1142 	struct hlist_nulls_head *head;
1143 	struct bucket *b;
1144 	struct htab_elem *l;
1145 	unsigned long flags;
1146 	u32 hash, key_size;
1147 	int ret = -ENOENT;
1148 
1149 	WARN_ON_ONCE(!rcu_read_lock_held());
1150 
1151 	key_size = map->key_size;
1152 
1153 	hash = htab_map_hash(key, key_size, htab->hashrnd);
1154 	b = __select_bucket(htab, hash);
1155 	head = &b->head;
1156 
1157 	raw_spin_lock_irqsave(&b->lock, flags);
1158 
1159 	l = lookup_elem_raw(head, hash, key, key_size);
1160 
1161 	if (l) {
1162 		hlist_nulls_del_rcu(&l->hash_node);
1163 		ret = 0;
1164 	}
1165 
1166 	raw_spin_unlock_irqrestore(&b->lock, flags);
1167 	if (l)
1168 		bpf_lru_push_free(&htab->lru, &l->lru_node);
1169 	return ret;
1170 }
1171 
1172 static void delete_all_elements(struct bpf_htab *htab)
1173 {
1174 	int i;
1175 
1176 	for (i = 0; i < htab->n_buckets; i++) {
1177 		struct hlist_nulls_head *head = select_bucket(htab, i);
1178 		struct hlist_nulls_node *n;
1179 		struct htab_elem *l;
1180 
1181 		hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1182 			hlist_nulls_del_rcu(&l->hash_node);
1183 			htab_elem_free(htab, l);
1184 		}
1185 	}
1186 }
1187 
1188 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1189 static void htab_map_free(struct bpf_map *map)
1190 {
1191 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1192 
1193 	/* at this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
1194 	 * so the programs (can be more than one that used this map) were
1195 	 * disconnected from events. Wait for outstanding critical sections in
1196 	 * these programs to complete
1197 	 */
1198 	synchronize_rcu();
1199 
1200 	/* some of free_htab_elem() callbacks for elements of this map may
1201 	 * not have executed. Wait for them.
1202 	 */
1203 	rcu_barrier();
1204 	if (!htab_is_prealloc(htab))
1205 		delete_all_elements(htab);
1206 	else
1207 		prealloc_destroy(htab);
1208 
1209 	free_percpu(htab->extra_elems);
1210 	bpf_map_area_free(htab->buckets);
1211 	kfree(htab);
1212 }
1213 
1214 static void htab_map_seq_show_elem(struct bpf_map *map, void *key,
1215 				   struct seq_file *m)
1216 {
1217 	void *value;
1218 
1219 	rcu_read_lock();
1220 
1221 	value = htab_map_lookup_elem(map, key);
1222 	if (!value) {
1223 		rcu_read_unlock();
1224 		return;
1225 	}
1226 
1227 	btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1228 	seq_puts(m, ": ");
1229 	btf_type_seq_show(map->btf, map->btf_value_type_id, value, m);
1230 	seq_puts(m, "\n");
1231 
1232 	rcu_read_unlock();
1233 }
1234 
1235 const struct bpf_map_ops htab_map_ops = {
1236 	.map_alloc_check = htab_map_alloc_check,
1237 	.map_alloc = htab_map_alloc,
1238 	.map_free = htab_map_free,
1239 	.map_get_next_key = htab_map_get_next_key,
1240 	.map_lookup_elem = htab_map_lookup_elem,
1241 	.map_update_elem = htab_map_update_elem,
1242 	.map_delete_elem = htab_map_delete_elem,
1243 	.map_gen_lookup = htab_map_gen_lookup,
1244 	.map_seq_show_elem = htab_map_seq_show_elem,
1245 };
1246 
1247 const struct bpf_map_ops htab_lru_map_ops = {
1248 	.map_alloc_check = htab_map_alloc_check,
1249 	.map_alloc = htab_map_alloc,
1250 	.map_free = htab_map_free,
1251 	.map_get_next_key = htab_map_get_next_key,
1252 	.map_lookup_elem = htab_lru_map_lookup_elem,
1253 	.map_update_elem = htab_lru_map_update_elem,
1254 	.map_delete_elem = htab_lru_map_delete_elem,
1255 	.map_gen_lookup = htab_lru_map_gen_lookup,
1256 	.map_seq_show_elem = htab_map_seq_show_elem,
1257 };
1258 
1259 /* Called from eBPF program */
1260 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1261 {
1262 	struct htab_elem *l = __htab_map_lookup_elem(map, key);
1263 
1264 	if (l)
1265 		return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1266 	else
1267 		return NULL;
1268 }
1269 
1270 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1271 {
1272 	struct htab_elem *l = __htab_map_lookup_elem(map, key);
1273 
1274 	if (l) {
1275 		bpf_lru_node_set_ref(&l->lru_node);
1276 		return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1277 	}
1278 
1279 	return NULL;
1280 }
1281 
1282 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
1283 {
1284 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1285 	struct htab_elem *l;
1286 	void __percpu *pptr;
1287 	int ret = -ENOENT;
1288 	int cpu, off = 0;
1289 	u32 size;
1290 
1291 	/* per_cpu areas are zero-filled and bpf programs can only
1292 	 * access 'value_size' of them, so copying rounded areas
1293 	 * will not leak any kernel data
1294 	 */
1295 	size = round_up(map->value_size, 8);
1296 	rcu_read_lock();
1297 	l = __htab_map_lookup_elem(map, key);
1298 	if (!l)
1299 		goto out;
1300 	if (htab_is_lru(htab))
1301 		bpf_lru_node_set_ref(&l->lru_node);
1302 	pptr = htab_elem_get_ptr(l, map->key_size);
1303 	for_each_possible_cpu(cpu) {
1304 		bpf_long_memcpy(value + off,
1305 				per_cpu_ptr(pptr, cpu), size);
1306 		off += size;
1307 	}
1308 	ret = 0;
1309 out:
1310 	rcu_read_unlock();
1311 	return ret;
1312 }
1313 
1314 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
1315 			   u64 map_flags)
1316 {
1317 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1318 	int ret;
1319 
1320 	rcu_read_lock();
1321 	if (htab_is_lru(htab))
1322 		ret = __htab_lru_percpu_map_update_elem(map, key, value,
1323 							map_flags, true);
1324 	else
1325 		ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
1326 						    true);
1327 	rcu_read_unlock();
1328 
1329 	return ret;
1330 }
1331 
1332 static void htab_percpu_map_seq_show_elem(struct bpf_map *map, void *key,
1333 					  struct seq_file *m)
1334 {
1335 	struct htab_elem *l;
1336 	void __percpu *pptr;
1337 	int cpu;
1338 
1339 	rcu_read_lock();
1340 
1341 	l = __htab_map_lookup_elem(map, key);
1342 	if (!l) {
1343 		rcu_read_unlock();
1344 		return;
1345 	}
1346 
1347 	btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1348 	seq_puts(m, ": {\n");
1349 	pptr = htab_elem_get_ptr(l, map->key_size);
1350 	for_each_possible_cpu(cpu) {
1351 		seq_printf(m, "\tcpu%d: ", cpu);
1352 		btf_type_seq_show(map->btf, map->btf_value_type_id,
1353 				  per_cpu_ptr(pptr, cpu), m);
1354 		seq_puts(m, "\n");
1355 	}
1356 	seq_puts(m, "}\n");
1357 
1358 	rcu_read_unlock();
1359 }
1360 
1361 const struct bpf_map_ops htab_percpu_map_ops = {
1362 	.map_alloc_check = htab_map_alloc_check,
1363 	.map_alloc = htab_map_alloc,
1364 	.map_free = htab_map_free,
1365 	.map_get_next_key = htab_map_get_next_key,
1366 	.map_lookup_elem = htab_percpu_map_lookup_elem,
1367 	.map_update_elem = htab_percpu_map_update_elem,
1368 	.map_delete_elem = htab_map_delete_elem,
1369 	.map_seq_show_elem = htab_percpu_map_seq_show_elem,
1370 };
1371 
1372 const struct bpf_map_ops htab_lru_percpu_map_ops = {
1373 	.map_alloc_check = htab_map_alloc_check,
1374 	.map_alloc = htab_map_alloc,
1375 	.map_free = htab_map_free,
1376 	.map_get_next_key = htab_map_get_next_key,
1377 	.map_lookup_elem = htab_lru_percpu_map_lookup_elem,
1378 	.map_update_elem = htab_lru_percpu_map_update_elem,
1379 	.map_delete_elem = htab_lru_map_delete_elem,
1380 	.map_seq_show_elem = htab_percpu_map_seq_show_elem,
1381 };
1382 
1383 static int fd_htab_map_alloc_check(union bpf_attr *attr)
1384 {
1385 	if (attr->value_size != sizeof(u32))
1386 		return -EINVAL;
1387 	return htab_map_alloc_check(attr);
1388 }
1389 
1390 static void fd_htab_map_free(struct bpf_map *map)
1391 {
1392 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1393 	struct hlist_nulls_node *n;
1394 	struct hlist_nulls_head *head;
1395 	struct htab_elem *l;
1396 	int i;
1397 
1398 	for (i = 0; i < htab->n_buckets; i++) {
1399 		head = select_bucket(htab, i);
1400 
1401 		hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1402 			void *ptr = fd_htab_map_get_ptr(map, l);
1403 
1404 			map->ops->map_fd_put_ptr(ptr);
1405 		}
1406 	}
1407 
1408 	htab_map_free(map);
1409 }
1410 
1411 /* only called from syscall */
1412 int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
1413 {
1414 	void **ptr;
1415 	int ret = 0;
1416 
1417 	if (!map->ops->map_fd_sys_lookup_elem)
1418 		return -ENOTSUPP;
1419 
1420 	rcu_read_lock();
1421 	ptr = htab_map_lookup_elem(map, key);
1422 	if (ptr)
1423 		*value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
1424 	else
1425 		ret = -ENOENT;
1426 	rcu_read_unlock();
1427 
1428 	return ret;
1429 }
1430 
1431 /* only called from syscall */
1432 int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
1433 				void *key, void *value, u64 map_flags)
1434 {
1435 	void *ptr;
1436 	int ret;
1437 	u32 ufd = *(u32 *)value;
1438 
1439 	ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
1440 	if (IS_ERR(ptr))
1441 		return PTR_ERR(ptr);
1442 
1443 	ret = htab_map_update_elem(map, key, &ptr, map_flags);
1444 	if (ret)
1445 		map->ops->map_fd_put_ptr(ptr);
1446 
1447 	return ret;
1448 }
1449 
1450 static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
1451 {
1452 	struct bpf_map *map, *inner_map_meta;
1453 
1454 	inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
1455 	if (IS_ERR(inner_map_meta))
1456 		return inner_map_meta;
1457 
1458 	map = htab_map_alloc(attr);
1459 	if (IS_ERR(map)) {
1460 		bpf_map_meta_free(inner_map_meta);
1461 		return map;
1462 	}
1463 
1464 	map->inner_map_meta = inner_map_meta;
1465 
1466 	return map;
1467 }
1468 
1469 static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
1470 {
1471 	struct bpf_map **inner_map  = htab_map_lookup_elem(map, key);
1472 
1473 	if (!inner_map)
1474 		return NULL;
1475 
1476 	return READ_ONCE(*inner_map);
1477 }
1478 
1479 static u32 htab_of_map_gen_lookup(struct bpf_map *map,
1480 				  struct bpf_insn *insn_buf)
1481 {
1482 	struct bpf_insn *insn = insn_buf;
1483 	const int ret = BPF_REG_0;
1484 
1485 	BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
1486 		     (void *(*)(struct bpf_map *map, void *key))NULL));
1487 	*insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
1488 	*insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
1489 	*insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
1490 				offsetof(struct htab_elem, key) +
1491 				round_up(map->key_size, 8));
1492 	*insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
1493 
1494 	return insn - insn_buf;
1495 }
1496 
1497 static void htab_of_map_free(struct bpf_map *map)
1498 {
1499 	bpf_map_meta_free(map->inner_map_meta);
1500 	fd_htab_map_free(map);
1501 }
1502 
1503 const struct bpf_map_ops htab_of_maps_map_ops = {
1504 	.map_alloc_check = fd_htab_map_alloc_check,
1505 	.map_alloc = htab_of_map_alloc,
1506 	.map_free = htab_of_map_free,
1507 	.map_get_next_key = htab_map_get_next_key,
1508 	.map_lookup_elem = htab_of_map_lookup_elem,
1509 	.map_delete_elem = htab_map_delete_elem,
1510 	.map_fd_get_ptr = bpf_map_fd_get_ptr,
1511 	.map_fd_put_ptr = bpf_map_fd_put_ptr,
1512 	.map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
1513 	.map_gen_lookup = htab_of_map_gen_lookup,
1514 	.map_check_btf = map_check_no_btf,
1515 };
1516