xref: /openbmc/linux/kernel/bpf/hashtab.c (revision f9834f18)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
3  * Copyright (c) 2016 Facebook
4  */
5 #include <linux/bpf.h>
6 #include <linux/btf.h>
7 #include <linux/jhash.h>
8 #include <linux/filter.h>
9 #include <linux/rculist_nulls.h>
10 #include <linux/random.h>
11 #include <uapi/linux/btf.h>
12 #include "percpu_freelist.h"
13 #include "bpf_lru_list.h"
14 #include "map_in_map.h"
15 
16 #define HTAB_CREATE_FLAG_MASK						\
17 	(BPF_F_NO_PREALLOC | BPF_F_NO_COMMON_LRU | BPF_F_NUMA_NODE |	\
18 	 BPF_F_ACCESS_MASK | BPF_F_ZERO_SEED)
19 
20 #define BATCH_OPS(_name)			\
21 	.map_lookup_batch =			\
22 	_name##_map_lookup_batch,		\
23 	.map_lookup_and_delete_batch =		\
24 	_name##_map_lookup_and_delete_batch,	\
25 	.map_update_batch =			\
26 	generic_map_update_batch,		\
27 	.map_delete_batch =			\
28 	generic_map_delete_batch
29 
30 struct bucket {
31 	struct hlist_nulls_head head;
32 	raw_spinlock_t lock;
33 };
34 
35 struct bpf_htab {
36 	struct bpf_map map;
37 	struct bucket *buckets;
38 	void *elems;
39 	union {
40 		struct pcpu_freelist freelist;
41 		struct bpf_lru lru;
42 	};
43 	struct htab_elem *__percpu *extra_elems;
44 	atomic_t count;	/* number of elements in this hashtable */
45 	u32 n_buckets;	/* number of hash buckets */
46 	u32 elem_size;	/* size of each element in bytes */
47 	u32 hashrnd;
48 };
49 
50 /* each htab element is struct htab_elem + key + value */
51 struct htab_elem {
52 	union {
53 		struct hlist_nulls_node hash_node;
54 		struct {
55 			void *padding;
56 			union {
57 				struct bpf_htab *htab;
58 				struct pcpu_freelist_node fnode;
59 				struct htab_elem *batch_flink;
60 			};
61 		};
62 	};
63 	union {
64 		struct rcu_head rcu;
65 		struct bpf_lru_node lru_node;
66 	};
67 	u32 hash;
68 	char key[0] __aligned(8);
69 };
70 
71 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
72 
73 static bool htab_is_lru(const struct bpf_htab *htab)
74 {
75 	return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH ||
76 		htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
77 }
78 
79 static bool htab_is_percpu(const struct bpf_htab *htab)
80 {
81 	return htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH ||
82 		htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
83 }
84 
85 static bool htab_is_prealloc(const struct bpf_htab *htab)
86 {
87 	return !(htab->map.map_flags & BPF_F_NO_PREALLOC);
88 }
89 
90 static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
91 				     void __percpu *pptr)
92 {
93 	*(void __percpu **)(l->key + key_size) = pptr;
94 }
95 
96 static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
97 {
98 	return *(void __percpu **)(l->key + key_size);
99 }
100 
101 static void *fd_htab_map_get_ptr(const struct bpf_map *map, struct htab_elem *l)
102 {
103 	return *(void **)(l->key + roundup(map->key_size, 8));
104 }
105 
106 static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
107 {
108 	return (struct htab_elem *) (htab->elems + i * htab->elem_size);
109 }
110 
111 static void htab_free_elems(struct bpf_htab *htab)
112 {
113 	int i;
114 
115 	if (!htab_is_percpu(htab))
116 		goto free_elems;
117 
118 	for (i = 0; i < htab->map.max_entries; i++) {
119 		void __percpu *pptr;
120 
121 		pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
122 					 htab->map.key_size);
123 		free_percpu(pptr);
124 		cond_resched();
125 	}
126 free_elems:
127 	bpf_map_area_free(htab->elems);
128 }
129 
130 /* The LRU list has a lock (lru_lock). Each htab bucket has a lock
131  * (bucket_lock). If both locks need to be acquired together, the lock
132  * order is always lru_lock -> bucket_lock and this only happens in
133  * bpf_lru_list.c logic. For example, certain code path of
134  * bpf_lru_pop_free(), which is called by function prealloc_lru_pop(),
135  * will acquire lru_lock first followed by acquiring bucket_lock.
136  *
137  * In hashtab.c, to avoid deadlock, lock acquisition of
138  * bucket_lock followed by lru_lock is not allowed. In such cases,
139  * bucket_lock needs to be released first before acquiring lru_lock.
140  */
141 static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
142 					  u32 hash)
143 {
144 	struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
145 	struct htab_elem *l;
146 
147 	if (node) {
148 		l = container_of(node, struct htab_elem, lru_node);
149 		memcpy(l->key, key, htab->map.key_size);
150 		return l;
151 	}
152 
153 	return NULL;
154 }
155 
156 static int prealloc_init(struct bpf_htab *htab)
157 {
158 	u32 num_entries = htab->map.max_entries;
159 	int err = -ENOMEM, i;
160 
161 	if (!htab_is_percpu(htab) && !htab_is_lru(htab))
162 		num_entries += num_possible_cpus();
163 
164 	htab->elems = bpf_map_area_alloc(htab->elem_size * num_entries,
165 					 htab->map.numa_node);
166 	if (!htab->elems)
167 		return -ENOMEM;
168 
169 	if (!htab_is_percpu(htab))
170 		goto skip_percpu_elems;
171 
172 	for (i = 0; i < num_entries; i++) {
173 		u32 size = round_up(htab->map.value_size, 8);
174 		void __percpu *pptr;
175 
176 		pptr = __alloc_percpu_gfp(size, 8, GFP_USER | __GFP_NOWARN);
177 		if (!pptr)
178 			goto free_elems;
179 		htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
180 				  pptr);
181 		cond_resched();
182 	}
183 
184 skip_percpu_elems:
185 	if (htab_is_lru(htab))
186 		err = bpf_lru_init(&htab->lru,
187 				   htab->map.map_flags & BPF_F_NO_COMMON_LRU,
188 				   offsetof(struct htab_elem, hash) -
189 				   offsetof(struct htab_elem, lru_node),
190 				   htab_lru_map_delete_node,
191 				   htab);
192 	else
193 		err = pcpu_freelist_init(&htab->freelist);
194 
195 	if (err)
196 		goto free_elems;
197 
198 	if (htab_is_lru(htab))
199 		bpf_lru_populate(&htab->lru, htab->elems,
200 				 offsetof(struct htab_elem, lru_node),
201 				 htab->elem_size, num_entries);
202 	else
203 		pcpu_freelist_populate(&htab->freelist,
204 				       htab->elems + offsetof(struct htab_elem, fnode),
205 				       htab->elem_size, num_entries);
206 
207 	return 0;
208 
209 free_elems:
210 	htab_free_elems(htab);
211 	return err;
212 }
213 
214 static void prealloc_destroy(struct bpf_htab *htab)
215 {
216 	htab_free_elems(htab);
217 
218 	if (htab_is_lru(htab))
219 		bpf_lru_destroy(&htab->lru);
220 	else
221 		pcpu_freelist_destroy(&htab->freelist);
222 }
223 
224 static int alloc_extra_elems(struct bpf_htab *htab)
225 {
226 	struct htab_elem *__percpu *pptr, *l_new;
227 	struct pcpu_freelist_node *l;
228 	int cpu;
229 
230 	pptr = __alloc_percpu_gfp(sizeof(struct htab_elem *), 8,
231 				  GFP_USER | __GFP_NOWARN);
232 	if (!pptr)
233 		return -ENOMEM;
234 
235 	for_each_possible_cpu(cpu) {
236 		l = pcpu_freelist_pop(&htab->freelist);
237 		/* pop will succeed, since prealloc_init()
238 		 * preallocated extra num_possible_cpus elements
239 		 */
240 		l_new = container_of(l, struct htab_elem, fnode);
241 		*per_cpu_ptr(pptr, cpu) = l_new;
242 	}
243 	htab->extra_elems = pptr;
244 	return 0;
245 }
246 
247 /* Called from syscall */
248 static int htab_map_alloc_check(union bpf_attr *attr)
249 {
250 	bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
251 		       attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
252 	bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
253 		    attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
254 	/* percpu_lru means each cpu has its own LRU list.
255 	 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
256 	 * the map's value itself is percpu.  percpu_lru has
257 	 * nothing to do with the map's value.
258 	 */
259 	bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
260 	bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
261 	bool zero_seed = (attr->map_flags & BPF_F_ZERO_SEED);
262 	int numa_node = bpf_map_attr_numa_node(attr);
263 
264 	BUILD_BUG_ON(offsetof(struct htab_elem, htab) !=
265 		     offsetof(struct htab_elem, hash_node.pprev));
266 	BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
267 		     offsetof(struct htab_elem, hash_node.pprev));
268 
269 	if (lru && !capable(CAP_SYS_ADMIN))
270 		/* LRU implementation is much complicated than other
271 		 * maps.  Hence, limit to CAP_SYS_ADMIN for now.
272 		 */
273 		return -EPERM;
274 
275 	if (zero_seed && !capable(CAP_SYS_ADMIN))
276 		/* Guard against local DoS, and discourage production use. */
277 		return -EPERM;
278 
279 	if (attr->map_flags & ~HTAB_CREATE_FLAG_MASK ||
280 	    !bpf_map_flags_access_ok(attr->map_flags))
281 		return -EINVAL;
282 
283 	if (!lru && percpu_lru)
284 		return -EINVAL;
285 
286 	if (lru && !prealloc)
287 		return -ENOTSUPP;
288 
289 	if (numa_node != NUMA_NO_NODE && (percpu || percpu_lru))
290 		return -EINVAL;
291 
292 	/* check sanity of attributes.
293 	 * value_size == 0 may be allowed in the future to use map as a set
294 	 */
295 	if (attr->max_entries == 0 || attr->key_size == 0 ||
296 	    attr->value_size == 0)
297 		return -EINVAL;
298 
299 	if (attr->key_size > MAX_BPF_STACK)
300 		/* eBPF programs initialize keys on stack, so they cannot be
301 		 * larger than max stack size
302 		 */
303 		return -E2BIG;
304 
305 	if (attr->value_size >= KMALLOC_MAX_SIZE -
306 	    MAX_BPF_STACK - sizeof(struct htab_elem))
307 		/* if value_size is bigger, the user space won't be able to
308 		 * access the elements via bpf syscall. This check also makes
309 		 * sure that the elem_size doesn't overflow and it's
310 		 * kmalloc-able later in htab_map_update_elem()
311 		 */
312 		return -E2BIG;
313 
314 	return 0;
315 }
316 
317 static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
318 {
319 	bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
320 		       attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
321 	bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
322 		    attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
323 	/* percpu_lru means each cpu has its own LRU list.
324 	 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
325 	 * the map's value itself is percpu.  percpu_lru has
326 	 * nothing to do with the map's value.
327 	 */
328 	bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
329 	bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
330 	struct bpf_htab *htab;
331 	int err, i;
332 	u64 cost;
333 
334 	htab = kzalloc(sizeof(*htab), GFP_USER);
335 	if (!htab)
336 		return ERR_PTR(-ENOMEM);
337 
338 	bpf_map_init_from_attr(&htab->map, attr);
339 
340 	if (percpu_lru) {
341 		/* ensure each CPU's lru list has >=1 elements.
342 		 * since we are at it, make each lru list has the same
343 		 * number of elements.
344 		 */
345 		htab->map.max_entries = roundup(attr->max_entries,
346 						num_possible_cpus());
347 		if (htab->map.max_entries < attr->max_entries)
348 			htab->map.max_entries = rounddown(attr->max_entries,
349 							  num_possible_cpus());
350 	}
351 
352 	/* hash table size must be power of 2 */
353 	htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
354 
355 	htab->elem_size = sizeof(struct htab_elem) +
356 			  round_up(htab->map.key_size, 8);
357 	if (percpu)
358 		htab->elem_size += sizeof(void *);
359 	else
360 		htab->elem_size += round_up(htab->map.value_size, 8);
361 
362 	err = -E2BIG;
363 	/* prevent zero size kmalloc and check for u32 overflow */
364 	if (htab->n_buckets == 0 ||
365 	    htab->n_buckets > U32_MAX / sizeof(struct bucket))
366 		goto free_htab;
367 
368 	cost = (u64) htab->n_buckets * sizeof(struct bucket) +
369 	       (u64) htab->elem_size * htab->map.max_entries;
370 
371 	if (percpu)
372 		cost += (u64) round_up(htab->map.value_size, 8) *
373 			num_possible_cpus() * htab->map.max_entries;
374 	else
375 	       cost += (u64) htab->elem_size * num_possible_cpus();
376 
377 	/* if map size is larger than memlock limit, reject it */
378 	err = bpf_map_charge_init(&htab->map.memory, cost);
379 	if (err)
380 		goto free_htab;
381 
382 	err = -ENOMEM;
383 	htab->buckets = bpf_map_area_alloc(htab->n_buckets *
384 					   sizeof(struct bucket),
385 					   htab->map.numa_node);
386 	if (!htab->buckets)
387 		goto free_charge;
388 
389 	if (htab->map.map_flags & BPF_F_ZERO_SEED)
390 		htab->hashrnd = 0;
391 	else
392 		htab->hashrnd = get_random_int();
393 
394 	for (i = 0; i < htab->n_buckets; i++) {
395 		INIT_HLIST_NULLS_HEAD(&htab->buckets[i].head, i);
396 		raw_spin_lock_init(&htab->buckets[i].lock);
397 	}
398 
399 	if (prealloc) {
400 		err = prealloc_init(htab);
401 		if (err)
402 			goto free_buckets;
403 
404 		if (!percpu && !lru) {
405 			/* lru itself can remove the least used element, so
406 			 * there is no need for an extra elem during map_update.
407 			 */
408 			err = alloc_extra_elems(htab);
409 			if (err)
410 				goto free_prealloc;
411 		}
412 	}
413 
414 	return &htab->map;
415 
416 free_prealloc:
417 	prealloc_destroy(htab);
418 free_buckets:
419 	bpf_map_area_free(htab->buckets);
420 free_charge:
421 	bpf_map_charge_finish(&htab->map.memory);
422 free_htab:
423 	kfree(htab);
424 	return ERR_PTR(err);
425 }
426 
427 static inline u32 htab_map_hash(const void *key, u32 key_len, u32 hashrnd)
428 {
429 	return jhash(key, key_len, hashrnd);
430 }
431 
432 static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
433 {
434 	return &htab->buckets[hash & (htab->n_buckets - 1)];
435 }
436 
437 static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
438 {
439 	return &__select_bucket(htab, hash)->head;
440 }
441 
442 /* this lookup function can only be called with bucket lock taken */
443 static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
444 					 void *key, u32 key_size)
445 {
446 	struct hlist_nulls_node *n;
447 	struct htab_elem *l;
448 
449 	hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
450 		if (l->hash == hash && !memcmp(&l->key, key, key_size))
451 			return l;
452 
453 	return NULL;
454 }
455 
456 /* can be called without bucket lock. it will repeat the loop in
457  * the unlikely event when elements moved from one bucket into another
458  * while link list is being walked
459  */
460 static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
461 					       u32 hash, void *key,
462 					       u32 key_size, u32 n_buckets)
463 {
464 	struct hlist_nulls_node *n;
465 	struct htab_elem *l;
466 
467 again:
468 	hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
469 		if (l->hash == hash && !memcmp(&l->key, key, key_size))
470 			return l;
471 
472 	if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
473 		goto again;
474 
475 	return NULL;
476 }
477 
478 /* Called from syscall or from eBPF program directly, so
479  * arguments have to match bpf_map_lookup_elem() exactly.
480  * The return value is adjusted by BPF instructions
481  * in htab_map_gen_lookup().
482  */
483 static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
484 {
485 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
486 	struct hlist_nulls_head *head;
487 	struct htab_elem *l;
488 	u32 hash, key_size;
489 
490 	/* Must be called with rcu_read_lock. */
491 	WARN_ON_ONCE(!rcu_read_lock_held());
492 
493 	key_size = map->key_size;
494 
495 	hash = htab_map_hash(key, key_size, htab->hashrnd);
496 
497 	head = select_bucket(htab, hash);
498 
499 	l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
500 
501 	return l;
502 }
503 
504 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
505 {
506 	struct htab_elem *l = __htab_map_lookup_elem(map, key);
507 
508 	if (l)
509 		return l->key + round_up(map->key_size, 8);
510 
511 	return NULL;
512 }
513 
514 /* inline bpf_map_lookup_elem() call.
515  * Instead of:
516  * bpf_prog
517  *   bpf_map_lookup_elem
518  *     map->ops->map_lookup_elem
519  *       htab_map_lookup_elem
520  *         __htab_map_lookup_elem
521  * do:
522  * bpf_prog
523  *   __htab_map_lookup_elem
524  */
525 static u32 htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
526 {
527 	struct bpf_insn *insn = insn_buf;
528 	const int ret = BPF_REG_0;
529 
530 	BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
531 		     (void *(*)(struct bpf_map *map, void *key))NULL));
532 	*insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
533 	*insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
534 	*insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
535 				offsetof(struct htab_elem, key) +
536 				round_up(map->key_size, 8));
537 	return insn - insn_buf;
538 }
539 
540 static __always_inline void *__htab_lru_map_lookup_elem(struct bpf_map *map,
541 							void *key, const bool mark)
542 {
543 	struct htab_elem *l = __htab_map_lookup_elem(map, key);
544 
545 	if (l) {
546 		if (mark)
547 			bpf_lru_node_set_ref(&l->lru_node);
548 		return l->key + round_up(map->key_size, 8);
549 	}
550 
551 	return NULL;
552 }
553 
554 static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
555 {
556 	return __htab_lru_map_lookup_elem(map, key, true);
557 }
558 
559 static void *htab_lru_map_lookup_elem_sys(struct bpf_map *map, void *key)
560 {
561 	return __htab_lru_map_lookup_elem(map, key, false);
562 }
563 
564 static u32 htab_lru_map_gen_lookup(struct bpf_map *map,
565 				   struct bpf_insn *insn_buf)
566 {
567 	struct bpf_insn *insn = insn_buf;
568 	const int ret = BPF_REG_0;
569 	const int ref_reg = BPF_REG_1;
570 
571 	BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
572 		     (void *(*)(struct bpf_map *map, void *key))NULL));
573 	*insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
574 	*insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
575 	*insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
576 			      offsetof(struct htab_elem, lru_node) +
577 			      offsetof(struct bpf_lru_node, ref));
578 	*insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
579 	*insn++ = BPF_ST_MEM(BPF_B, ret,
580 			     offsetof(struct htab_elem, lru_node) +
581 			     offsetof(struct bpf_lru_node, ref),
582 			     1);
583 	*insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
584 				offsetof(struct htab_elem, key) +
585 				round_up(map->key_size, 8));
586 	return insn - insn_buf;
587 }
588 
589 /* It is called from the bpf_lru_list when the LRU needs to delete
590  * older elements from the htab.
591  */
592 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
593 {
594 	struct bpf_htab *htab = (struct bpf_htab *)arg;
595 	struct htab_elem *l = NULL, *tgt_l;
596 	struct hlist_nulls_head *head;
597 	struct hlist_nulls_node *n;
598 	unsigned long flags;
599 	struct bucket *b;
600 
601 	tgt_l = container_of(node, struct htab_elem, lru_node);
602 	b = __select_bucket(htab, tgt_l->hash);
603 	head = &b->head;
604 
605 	raw_spin_lock_irqsave(&b->lock, flags);
606 
607 	hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
608 		if (l == tgt_l) {
609 			hlist_nulls_del_rcu(&l->hash_node);
610 			break;
611 		}
612 
613 	raw_spin_unlock_irqrestore(&b->lock, flags);
614 
615 	return l == tgt_l;
616 }
617 
618 /* Called from syscall */
619 static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
620 {
621 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
622 	struct hlist_nulls_head *head;
623 	struct htab_elem *l, *next_l;
624 	u32 hash, key_size;
625 	int i = 0;
626 
627 	WARN_ON_ONCE(!rcu_read_lock_held());
628 
629 	key_size = map->key_size;
630 
631 	if (!key)
632 		goto find_first_elem;
633 
634 	hash = htab_map_hash(key, key_size, htab->hashrnd);
635 
636 	head = select_bucket(htab, hash);
637 
638 	/* lookup the key */
639 	l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
640 
641 	if (!l)
642 		goto find_first_elem;
643 
644 	/* key was found, get next key in the same bucket */
645 	next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
646 				  struct htab_elem, hash_node);
647 
648 	if (next_l) {
649 		/* if next elem in this hash list is non-zero, just return it */
650 		memcpy(next_key, next_l->key, key_size);
651 		return 0;
652 	}
653 
654 	/* no more elements in this hash list, go to the next bucket */
655 	i = hash & (htab->n_buckets - 1);
656 	i++;
657 
658 find_first_elem:
659 	/* iterate over buckets */
660 	for (; i < htab->n_buckets; i++) {
661 		head = select_bucket(htab, i);
662 
663 		/* pick first element in the bucket */
664 		next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
665 					  struct htab_elem, hash_node);
666 		if (next_l) {
667 			/* if it's not empty, just return it */
668 			memcpy(next_key, next_l->key, key_size);
669 			return 0;
670 		}
671 	}
672 
673 	/* iterated over all buckets and all elements */
674 	return -ENOENT;
675 }
676 
677 static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
678 {
679 	if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
680 		free_percpu(htab_elem_get_ptr(l, htab->map.key_size));
681 	kfree(l);
682 }
683 
684 static void htab_elem_free_rcu(struct rcu_head *head)
685 {
686 	struct htab_elem *l = container_of(head, struct htab_elem, rcu);
687 	struct bpf_htab *htab = l->htab;
688 
689 	/* must increment bpf_prog_active to avoid kprobe+bpf triggering while
690 	 * we're calling kfree, otherwise deadlock is possible if kprobes
691 	 * are placed somewhere inside of slub
692 	 */
693 	preempt_disable();
694 	__this_cpu_inc(bpf_prog_active);
695 	htab_elem_free(htab, l);
696 	__this_cpu_dec(bpf_prog_active);
697 	preempt_enable();
698 }
699 
700 static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
701 {
702 	struct bpf_map *map = &htab->map;
703 
704 	if (map->ops->map_fd_put_ptr) {
705 		void *ptr = fd_htab_map_get_ptr(map, l);
706 
707 		map->ops->map_fd_put_ptr(ptr);
708 	}
709 
710 	if (htab_is_prealloc(htab)) {
711 		__pcpu_freelist_push(&htab->freelist, &l->fnode);
712 	} else {
713 		atomic_dec(&htab->count);
714 		l->htab = htab;
715 		call_rcu(&l->rcu, htab_elem_free_rcu);
716 	}
717 }
718 
719 static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
720 			    void *value, bool onallcpus)
721 {
722 	if (!onallcpus) {
723 		/* copy true value_size bytes */
724 		memcpy(this_cpu_ptr(pptr), value, htab->map.value_size);
725 	} else {
726 		u32 size = round_up(htab->map.value_size, 8);
727 		int off = 0, cpu;
728 
729 		for_each_possible_cpu(cpu) {
730 			bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
731 					value + off, size);
732 			off += size;
733 		}
734 	}
735 }
736 
737 static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
738 {
739 	return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
740 	       BITS_PER_LONG == 64;
741 }
742 
743 static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
744 					 void *value, u32 key_size, u32 hash,
745 					 bool percpu, bool onallcpus,
746 					 struct htab_elem *old_elem)
747 {
748 	u32 size = htab->map.value_size;
749 	bool prealloc = htab_is_prealloc(htab);
750 	struct htab_elem *l_new, **pl_new;
751 	void __percpu *pptr;
752 
753 	if (prealloc) {
754 		if (old_elem) {
755 			/* if we're updating the existing element,
756 			 * use per-cpu extra elems to avoid freelist_pop/push
757 			 */
758 			pl_new = this_cpu_ptr(htab->extra_elems);
759 			l_new = *pl_new;
760 			*pl_new = old_elem;
761 		} else {
762 			struct pcpu_freelist_node *l;
763 
764 			l = __pcpu_freelist_pop(&htab->freelist);
765 			if (!l)
766 				return ERR_PTR(-E2BIG);
767 			l_new = container_of(l, struct htab_elem, fnode);
768 		}
769 	} else {
770 		if (atomic_inc_return(&htab->count) > htab->map.max_entries)
771 			if (!old_elem) {
772 				/* when map is full and update() is replacing
773 				 * old element, it's ok to allocate, since
774 				 * old element will be freed immediately.
775 				 * Otherwise return an error
776 				 */
777 				l_new = ERR_PTR(-E2BIG);
778 				goto dec_count;
779 			}
780 		l_new = kmalloc_node(htab->elem_size, GFP_ATOMIC | __GFP_NOWARN,
781 				     htab->map.numa_node);
782 		if (!l_new) {
783 			l_new = ERR_PTR(-ENOMEM);
784 			goto dec_count;
785 		}
786 		check_and_init_map_lock(&htab->map,
787 					l_new->key + round_up(key_size, 8));
788 	}
789 
790 	memcpy(l_new->key, key, key_size);
791 	if (percpu) {
792 		size = round_up(size, 8);
793 		if (prealloc) {
794 			pptr = htab_elem_get_ptr(l_new, key_size);
795 		} else {
796 			/* alloc_percpu zero-fills */
797 			pptr = __alloc_percpu_gfp(size, 8,
798 						  GFP_ATOMIC | __GFP_NOWARN);
799 			if (!pptr) {
800 				kfree(l_new);
801 				l_new = ERR_PTR(-ENOMEM);
802 				goto dec_count;
803 			}
804 		}
805 
806 		pcpu_copy_value(htab, pptr, value, onallcpus);
807 
808 		if (!prealloc)
809 			htab_elem_set_ptr(l_new, key_size, pptr);
810 	} else if (fd_htab_map_needs_adjust(htab)) {
811 		size = round_up(size, 8);
812 		memcpy(l_new->key + round_up(key_size, 8), value, size);
813 	} else {
814 		copy_map_value(&htab->map,
815 			       l_new->key + round_up(key_size, 8),
816 			       value);
817 	}
818 
819 	l_new->hash = hash;
820 	return l_new;
821 dec_count:
822 	atomic_dec(&htab->count);
823 	return l_new;
824 }
825 
826 static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
827 		       u64 map_flags)
828 {
829 	if (l_old && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
830 		/* elem already exists */
831 		return -EEXIST;
832 
833 	if (!l_old && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
834 		/* elem doesn't exist, cannot update it */
835 		return -ENOENT;
836 
837 	return 0;
838 }
839 
840 /* Called from syscall or from eBPF program */
841 static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
842 				u64 map_flags)
843 {
844 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
845 	struct htab_elem *l_new = NULL, *l_old;
846 	struct hlist_nulls_head *head;
847 	unsigned long flags;
848 	struct bucket *b;
849 	u32 key_size, hash;
850 	int ret;
851 
852 	if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST))
853 		/* unknown flags */
854 		return -EINVAL;
855 
856 	WARN_ON_ONCE(!rcu_read_lock_held());
857 
858 	key_size = map->key_size;
859 
860 	hash = htab_map_hash(key, key_size, htab->hashrnd);
861 
862 	b = __select_bucket(htab, hash);
863 	head = &b->head;
864 
865 	if (unlikely(map_flags & BPF_F_LOCK)) {
866 		if (unlikely(!map_value_has_spin_lock(map)))
867 			return -EINVAL;
868 		/* find an element without taking the bucket lock */
869 		l_old = lookup_nulls_elem_raw(head, hash, key, key_size,
870 					      htab->n_buckets);
871 		ret = check_flags(htab, l_old, map_flags);
872 		if (ret)
873 			return ret;
874 		if (l_old) {
875 			/* grab the element lock and update value in place */
876 			copy_map_value_locked(map,
877 					      l_old->key + round_up(key_size, 8),
878 					      value, false);
879 			return 0;
880 		}
881 		/* fall through, grab the bucket lock and lookup again.
882 		 * 99.9% chance that the element won't be found,
883 		 * but second lookup under lock has to be done.
884 		 */
885 	}
886 
887 	/* bpf_map_update_elem() can be called in_irq() */
888 	raw_spin_lock_irqsave(&b->lock, flags);
889 
890 	l_old = lookup_elem_raw(head, hash, key, key_size);
891 
892 	ret = check_flags(htab, l_old, map_flags);
893 	if (ret)
894 		goto err;
895 
896 	if (unlikely(l_old && (map_flags & BPF_F_LOCK))) {
897 		/* first lookup without the bucket lock didn't find the element,
898 		 * but second lookup with the bucket lock found it.
899 		 * This case is highly unlikely, but has to be dealt with:
900 		 * grab the element lock in addition to the bucket lock
901 		 * and update element in place
902 		 */
903 		copy_map_value_locked(map,
904 				      l_old->key + round_up(key_size, 8),
905 				      value, false);
906 		ret = 0;
907 		goto err;
908 	}
909 
910 	l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
911 				l_old);
912 	if (IS_ERR(l_new)) {
913 		/* all pre-allocated elements are in use or memory exhausted */
914 		ret = PTR_ERR(l_new);
915 		goto err;
916 	}
917 
918 	/* add new element to the head of the list, so that
919 	 * concurrent search will find it before old elem
920 	 */
921 	hlist_nulls_add_head_rcu(&l_new->hash_node, head);
922 	if (l_old) {
923 		hlist_nulls_del_rcu(&l_old->hash_node);
924 		if (!htab_is_prealloc(htab))
925 			free_htab_elem(htab, l_old);
926 	}
927 	ret = 0;
928 err:
929 	raw_spin_unlock_irqrestore(&b->lock, flags);
930 	return ret;
931 }
932 
933 static int htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
934 				    u64 map_flags)
935 {
936 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
937 	struct htab_elem *l_new, *l_old = NULL;
938 	struct hlist_nulls_head *head;
939 	unsigned long flags;
940 	struct bucket *b;
941 	u32 key_size, hash;
942 	int ret;
943 
944 	if (unlikely(map_flags > BPF_EXIST))
945 		/* unknown flags */
946 		return -EINVAL;
947 
948 	WARN_ON_ONCE(!rcu_read_lock_held());
949 
950 	key_size = map->key_size;
951 
952 	hash = htab_map_hash(key, key_size, htab->hashrnd);
953 
954 	b = __select_bucket(htab, hash);
955 	head = &b->head;
956 
957 	/* For LRU, we need to alloc before taking bucket's
958 	 * spinlock because getting free nodes from LRU may need
959 	 * to remove older elements from htab and this removal
960 	 * operation will need a bucket lock.
961 	 */
962 	l_new = prealloc_lru_pop(htab, key, hash);
963 	if (!l_new)
964 		return -ENOMEM;
965 	memcpy(l_new->key + round_up(map->key_size, 8), value, map->value_size);
966 
967 	/* bpf_map_update_elem() can be called in_irq() */
968 	raw_spin_lock_irqsave(&b->lock, flags);
969 
970 	l_old = lookup_elem_raw(head, hash, key, key_size);
971 
972 	ret = check_flags(htab, l_old, map_flags);
973 	if (ret)
974 		goto err;
975 
976 	/* add new element to the head of the list, so that
977 	 * concurrent search will find it before old elem
978 	 */
979 	hlist_nulls_add_head_rcu(&l_new->hash_node, head);
980 	if (l_old) {
981 		bpf_lru_node_set_ref(&l_new->lru_node);
982 		hlist_nulls_del_rcu(&l_old->hash_node);
983 	}
984 	ret = 0;
985 
986 err:
987 	raw_spin_unlock_irqrestore(&b->lock, flags);
988 
989 	if (ret)
990 		bpf_lru_push_free(&htab->lru, &l_new->lru_node);
991 	else if (l_old)
992 		bpf_lru_push_free(&htab->lru, &l_old->lru_node);
993 
994 	return ret;
995 }
996 
997 static int __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
998 					 void *value, u64 map_flags,
999 					 bool onallcpus)
1000 {
1001 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1002 	struct htab_elem *l_new = NULL, *l_old;
1003 	struct hlist_nulls_head *head;
1004 	unsigned long flags;
1005 	struct bucket *b;
1006 	u32 key_size, hash;
1007 	int ret;
1008 
1009 	if (unlikely(map_flags > BPF_EXIST))
1010 		/* unknown flags */
1011 		return -EINVAL;
1012 
1013 	WARN_ON_ONCE(!rcu_read_lock_held());
1014 
1015 	key_size = map->key_size;
1016 
1017 	hash = htab_map_hash(key, key_size, htab->hashrnd);
1018 
1019 	b = __select_bucket(htab, hash);
1020 	head = &b->head;
1021 
1022 	/* bpf_map_update_elem() can be called in_irq() */
1023 	raw_spin_lock_irqsave(&b->lock, flags);
1024 
1025 	l_old = lookup_elem_raw(head, hash, key, key_size);
1026 
1027 	ret = check_flags(htab, l_old, map_flags);
1028 	if (ret)
1029 		goto err;
1030 
1031 	if (l_old) {
1032 		/* per-cpu hash map can update value in-place */
1033 		pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1034 				value, onallcpus);
1035 	} else {
1036 		l_new = alloc_htab_elem(htab, key, value, key_size,
1037 					hash, true, onallcpus, NULL);
1038 		if (IS_ERR(l_new)) {
1039 			ret = PTR_ERR(l_new);
1040 			goto err;
1041 		}
1042 		hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1043 	}
1044 	ret = 0;
1045 err:
1046 	raw_spin_unlock_irqrestore(&b->lock, flags);
1047 	return ret;
1048 }
1049 
1050 static int __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1051 					     void *value, u64 map_flags,
1052 					     bool onallcpus)
1053 {
1054 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1055 	struct htab_elem *l_new = NULL, *l_old;
1056 	struct hlist_nulls_head *head;
1057 	unsigned long flags;
1058 	struct bucket *b;
1059 	u32 key_size, hash;
1060 	int ret;
1061 
1062 	if (unlikely(map_flags > BPF_EXIST))
1063 		/* unknown flags */
1064 		return -EINVAL;
1065 
1066 	WARN_ON_ONCE(!rcu_read_lock_held());
1067 
1068 	key_size = map->key_size;
1069 
1070 	hash = htab_map_hash(key, key_size, htab->hashrnd);
1071 
1072 	b = __select_bucket(htab, hash);
1073 	head = &b->head;
1074 
1075 	/* For LRU, we need to alloc before taking bucket's
1076 	 * spinlock because LRU's elem alloc may need
1077 	 * to remove older elem from htab and this removal
1078 	 * operation will need a bucket lock.
1079 	 */
1080 	if (map_flags != BPF_EXIST) {
1081 		l_new = prealloc_lru_pop(htab, key, hash);
1082 		if (!l_new)
1083 			return -ENOMEM;
1084 	}
1085 
1086 	/* bpf_map_update_elem() can be called in_irq() */
1087 	raw_spin_lock_irqsave(&b->lock, flags);
1088 
1089 	l_old = lookup_elem_raw(head, hash, key, key_size);
1090 
1091 	ret = check_flags(htab, l_old, map_flags);
1092 	if (ret)
1093 		goto err;
1094 
1095 	if (l_old) {
1096 		bpf_lru_node_set_ref(&l_old->lru_node);
1097 
1098 		/* per-cpu hash map can update value in-place */
1099 		pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1100 				value, onallcpus);
1101 	} else {
1102 		pcpu_copy_value(htab, htab_elem_get_ptr(l_new, key_size),
1103 				value, onallcpus);
1104 		hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1105 		l_new = NULL;
1106 	}
1107 	ret = 0;
1108 err:
1109 	raw_spin_unlock_irqrestore(&b->lock, flags);
1110 	if (l_new)
1111 		bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1112 	return ret;
1113 }
1114 
1115 static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1116 				       void *value, u64 map_flags)
1117 {
1118 	return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
1119 }
1120 
1121 static int htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1122 					   void *value, u64 map_flags)
1123 {
1124 	return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1125 						 false);
1126 }
1127 
1128 /* Called from syscall or from eBPF program */
1129 static int htab_map_delete_elem(struct bpf_map *map, void *key)
1130 {
1131 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1132 	struct hlist_nulls_head *head;
1133 	struct bucket *b;
1134 	struct htab_elem *l;
1135 	unsigned long flags;
1136 	u32 hash, key_size;
1137 	int ret = -ENOENT;
1138 
1139 	WARN_ON_ONCE(!rcu_read_lock_held());
1140 
1141 	key_size = map->key_size;
1142 
1143 	hash = htab_map_hash(key, key_size, htab->hashrnd);
1144 	b = __select_bucket(htab, hash);
1145 	head = &b->head;
1146 
1147 	raw_spin_lock_irqsave(&b->lock, flags);
1148 
1149 	l = lookup_elem_raw(head, hash, key, key_size);
1150 
1151 	if (l) {
1152 		hlist_nulls_del_rcu(&l->hash_node);
1153 		free_htab_elem(htab, l);
1154 		ret = 0;
1155 	}
1156 
1157 	raw_spin_unlock_irqrestore(&b->lock, flags);
1158 	return ret;
1159 }
1160 
1161 static int htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1162 {
1163 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1164 	struct hlist_nulls_head *head;
1165 	struct bucket *b;
1166 	struct htab_elem *l;
1167 	unsigned long flags;
1168 	u32 hash, key_size;
1169 	int ret = -ENOENT;
1170 
1171 	WARN_ON_ONCE(!rcu_read_lock_held());
1172 
1173 	key_size = map->key_size;
1174 
1175 	hash = htab_map_hash(key, key_size, htab->hashrnd);
1176 	b = __select_bucket(htab, hash);
1177 	head = &b->head;
1178 
1179 	raw_spin_lock_irqsave(&b->lock, flags);
1180 
1181 	l = lookup_elem_raw(head, hash, key, key_size);
1182 
1183 	if (l) {
1184 		hlist_nulls_del_rcu(&l->hash_node);
1185 		ret = 0;
1186 	}
1187 
1188 	raw_spin_unlock_irqrestore(&b->lock, flags);
1189 	if (l)
1190 		bpf_lru_push_free(&htab->lru, &l->lru_node);
1191 	return ret;
1192 }
1193 
1194 static void delete_all_elements(struct bpf_htab *htab)
1195 {
1196 	int i;
1197 
1198 	for (i = 0; i < htab->n_buckets; i++) {
1199 		struct hlist_nulls_head *head = select_bucket(htab, i);
1200 		struct hlist_nulls_node *n;
1201 		struct htab_elem *l;
1202 
1203 		hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1204 			hlist_nulls_del_rcu(&l->hash_node);
1205 			htab_elem_free(htab, l);
1206 		}
1207 	}
1208 }
1209 
1210 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1211 static void htab_map_free(struct bpf_map *map)
1212 {
1213 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1214 
1215 	/* at this point bpf_prog->aux->refcnt == 0 and this map->refcnt == 0,
1216 	 * so the programs (can be more than one that used this map) were
1217 	 * disconnected from events. Wait for outstanding critical sections in
1218 	 * these programs to complete
1219 	 */
1220 	synchronize_rcu();
1221 
1222 	/* some of free_htab_elem() callbacks for elements of this map may
1223 	 * not have executed. Wait for them.
1224 	 */
1225 	rcu_barrier();
1226 	if (!htab_is_prealloc(htab))
1227 		delete_all_elements(htab);
1228 	else
1229 		prealloc_destroy(htab);
1230 
1231 	free_percpu(htab->extra_elems);
1232 	bpf_map_area_free(htab->buckets);
1233 	kfree(htab);
1234 }
1235 
1236 static void htab_map_seq_show_elem(struct bpf_map *map, void *key,
1237 				   struct seq_file *m)
1238 {
1239 	void *value;
1240 
1241 	rcu_read_lock();
1242 
1243 	value = htab_map_lookup_elem(map, key);
1244 	if (!value) {
1245 		rcu_read_unlock();
1246 		return;
1247 	}
1248 
1249 	btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1250 	seq_puts(m, ": ");
1251 	btf_type_seq_show(map->btf, map->btf_value_type_id, value, m);
1252 	seq_puts(m, "\n");
1253 
1254 	rcu_read_unlock();
1255 }
1256 
1257 static int
1258 __htab_map_lookup_and_delete_batch(struct bpf_map *map,
1259 				   const union bpf_attr *attr,
1260 				   union bpf_attr __user *uattr,
1261 				   bool do_delete, bool is_lru_map,
1262 				   bool is_percpu)
1263 {
1264 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1265 	u32 bucket_cnt, total, key_size, value_size, roundup_key_size;
1266 	void *keys = NULL, *values = NULL, *value, *dst_key, *dst_val;
1267 	void __user *uvalues = u64_to_user_ptr(attr->batch.values);
1268 	void __user *ukeys = u64_to_user_ptr(attr->batch.keys);
1269 	void *ubatch = u64_to_user_ptr(attr->batch.in_batch);
1270 	u32 batch, max_count, size, bucket_size;
1271 	struct htab_elem *node_to_free = NULL;
1272 	u64 elem_map_flags, map_flags;
1273 	struct hlist_nulls_head *head;
1274 	struct hlist_nulls_node *n;
1275 	unsigned long flags = 0;
1276 	bool locked = false;
1277 	struct htab_elem *l;
1278 	struct bucket *b;
1279 	int ret = 0;
1280 
1281 	elem_map_flags = attr->batch.elem_flags;
1282 	if ((elem_map_flags & ~BPF_F_LOCK) ||
1283 	    ((elem_map_flags & BPF_F_LOCK) && !map_value_has_spin_lock(map)))
1284 		return -EINVAL;
1285 
1286 	map_flags = attr->batch.flags;
1287 	if (map_flags)
1288 		return -EINVAL;
1289 
1290 	max_count = attr->batch.count;
1291 	if (!max_count)
1292 		return 0;
1293 
1294 	if (put_user(0, &uattr->batch.count))
1295 		return -EFAULT;
1296 
1297 	batch = 0;
1298 	if (ubatch && copy_from_user(&batch, ubatch, sizeof(batch)))
1299 		return -EFAULT;
1300 
1301 	if (batch >= htab->n_buckets)
1302 		return -ENOENT;
1303 
1304 	key_size = htab->map.key_size;
1305 	roundup_key_size = round_up(htab->map.key_size, 8);
1306 	value_size = htab->map.value_size;
1307 	size = round_up(value_size, 8);
1308 	if (is_percpu)
1309 		value_size = size * num_possible_cpus();
1310 	total = 0;
1311 	/* while experimenting with hash tables with sizes ranging from 10 to
1312 	 * 1000, it was observed that a bucket can have upto 5 entries.
1313 	 */
1314 	bucket_size = 5;
1315 
1316 alloc:
1317 	/* We cannot do copy_from_user or copy_to_user inside
1318 	 * the rcu_read_lock. Allocate enough space here.
1319 	 */
1320 	keys = kvmalloc(key_size * bucket_size, GFP_USER | __GFP_NOWARN);
1321 	values = kvmalloc(value_size * bucket_size, GFP_USER | __GFP_NOWARN);
1322 	if (!keys || !values) {
1323 		ret = -ENOMEM;
1324 		goto after_loop;
1325 	}
1326 
1327 again:
1328 	preempt_disable();
1329 	this_cpu_inc(bpf_prog_active);
1330 	rcu_read_lock();
1331 again_nocopy:
1332 	dst_key = keys;
1333 	dst_val = values;
1334 	b = &htab->buckets[batch];
1335 	head = &b->head;
1336 	/* do not grab the lock unless need it (bucket_cnt > 0). */
1337 	if (locked)
1338 		raw_spin_lock_irqsave(&b->lock, flags);
1339 
1340 	bucket_cnt = 0;
1341 	hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
1342 		bucket_cnt++;
1343 
1344 	if (bucket_cnt && !locked) {
1345 		locked = true;
1346 		goto again_nocopy;
1347 	}
1348 
1349 	if (bucket_cnt > (max_count - total)) {
1350 		if (total == 0)
1351 			ret = -ENOSPC;
1352 		/* Note that since bucket_cnt > 0 here, it is implicit
1353 		 * that the locked was grabbed, so release it.
1354 		 */
1355 		raw_spin_unlock_irqrestore(&b->lock, flags);
1356 		rcu_read_unlock();
1357 		this_cpu_dec(bpf_prog_active);
1358 		preempt_enable();
1359 		goto after_loop;
1360 	}
1361 
1362 	if (bucket_cnt > bucket_size) {
1363 		bucket_size = bucket_cnt;
1364 		/* Note that since bucket_cnt > 0 here, it is implicit
1365 		 * that the locked was grabbed, so release it.
1366 		 */
1367 		raw_spin_unlock_irqrestore(&b->lock, flags);
1368 		rcu_read_unlock();
1369 		this_cpu_dec(bpf_prog_active);
1370 		preempt_enable();
1371 		kvfree(keys);
1372 		kvfree(values);
1373 		goto alloc;
1374 	}
1375 
1376 	/* Next block is only safe to run if you have grabbed the lock */
1377 	if (!locked)
1378 		goto next_batch;
1379 
1380 	hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1381 		memcpy(dst_key, l->key, key_size);
1382 
1383 		if (is_percpu) {
1384 			int off = 0, cpu;
1385 			void __percpu *pptr;
1386 
1387 			pptr = htab_elem_get_ptr(l, map->key_size);
1388 			for_each_possible_cpu(cpu) {
1389 				bpf_long_memcpy(dst_val + off,
1390 						per_cpu_ptr(pptr, cpu), size);
1391 				off += size;
1392 			}
1393 		} else {
1394 			value = l->key + roundup_key_size;
1395 			if (elem_map_flags & BPF_F_LOCK)
1396 				copy_map_value_locked(map, dst_val, value,
1397 						      true);
1398 			else
1399 				copy_map_value(map, dst_val, value);
1400 			check_and_init_map_lock(map, dst_val);
1401 		}
1402 		if (do_delete) {
1403 			hlist_nulls_del_rcu(&l->hash_node);
1404 
1405 			/* bpf_lru_push_free() will acquire lru_lock, which
1406 			 * may cause deadlock. See comments in function
1407 			 * prealloc_lru_pop(). Let us do bpf_lru_push_free()
1408 			 * after releasing the bucket lock.
1409 			 */
1410 			if (is_lru_map) {
1411 				l->batch_flink = node_to_free;
1412 				node_to_free = l;
1413 			} else {
1414 				free_htab_elem(htab, l);
1415 			}
1416 		}
1417 		dst_key += key_size;
1418 		dst_val += value_size;
1419 	}
1420 
1421 	raw_spin_unlock_irqrestore(&b->lock, flags);
1422 	locked = false;
1423 
1424 	while (node_to_free) {
1425 		l = node_to_free;
1426 		node_to_free = node_to_free->batch_flink;
1427 		bpf_lru_push_free(&htab->lru, &l->lru_node);
1428 	}
1429 
1430 next_batch:
1431 	/* If we are not copying data, we can go to next bucket and avoid
1432 	 * unlocking the rcu.
1433 	 */
1434 	if (!bucket_cnt && (batch + 1 < htab->n_buckets)) {
1435 		batch++;
1436 		goto again_nocopy;
1437 	}
1438 
1439 	rcu_read_unlock();
1440 	this_cpu_dec(bpf_prog_active);
1441 	preempt_enable();
1442 	if (bucket_cnt && (copy_to_user(ukeys + total * key_size, keys,
1443 	    key_size * bucket_cnt) ||
1444 	    copy_to_user(uvalues + total * value_size, values,
1445 	    value_size * bucket_cnt))) {
1446 		ret = -EFAULT;
1447 		goto after_loop;
1448 	}
1449 
1450 	total += bucket_cnt;
1451 	batch++;
1452 	if (batch >= htab->n_buckets) {
1453 		ret = -ENOENT;
1454 		goto after_loop;
1455 	}
1456 	goto again;
1457 
1458 after_loop:
1459 	if (ret == -EFAULT)
1460 		goto out;
1461 
1462 	/* copy # of entries and next batch */
1463 	ubatch = u64_to_user_ptr(attr->batch.out_batch);
1464 	if (copy_to_user(ubatch, &batch, sizeof(batch)) ||
1465 	    put_user(total, &uattr->batch.count))
1466 		ret = -EFAULT;
1467 
1468 out:
1469 	kvfree(keys);
1470 	kvfree(values);
1471 	return ret;
1472 }
1473 
1474 static int
1475 htab_percpu_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1476 			     union bpf_attr __user *uattr)
1477 {
1478 	return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1479 						  false, true);
1480 }
1481 
1482 static int
1483 htab_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1484 					const union bpf_attr *attr,
1485 					union bpf_attr __user *uattr)
1486 {
1487 	return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1488 						  false, true);
1489 }
1490 
1491 static int
1492 htab_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1493 		      union bpf_attr __user *uattr)
1494 {
1495 	return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1496 						  false, false);
1497 }
1498 
1499 static int
1500 htab_map_lookup_and_delete_batch(struct bpf_map *map,
1501 				 const union bpf_attr *attr,
1502 				 union bpf_attr __user *uattr)
1503 {
1504 	return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1505 						  false, false);
1506 }
1507 
1508 static int
1509 htab_lru_percpu_map_lookup_batch(struct bpf_map *map,
1510 				 const union bpf_attr *attr,
1511 				 union bpf_attr __user *uattr)
1512 {
1513 	return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1514 						  true, true);
1515 }
1516 
1517 static int
1518 htab_lru_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1519 					    const union bpf_attr *attr,
1520 					    union bpf_attr __user *uattr)
1521 {
1522 	return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1523 						  true, true);
1524 }
1525 
1526 static int
1527 htab_lru_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1528 			  union bpf_attr __user *uattr)
1529 {
1530 	return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1531 						  true, false);
1532 }
1533 
1534 static int
1535 htab_lru_map_lookup_and_delete_batch(struct bpf_map *map,
1536 				     const union bpf_attr *attr,
1537 				     union bpf_attr __user *uattr)
1538 {
1539 	return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1540 						  true, false);
1541 }
1542 
1543 const struct bpf_map_ops htab_map_ops = {
1544 	.map_alloc_check = htab_map_alloc_check,
1545 	.map_alloc = htab_map_alloc,
1546 	.map_free = htab_map_free,
1547 	.map_get_next_key = htab_map_get_next_key,
1548 	.map_lookup_elem = htab_map_lookup_elem,
1549 	.map_update_elem = htab_map_update_elem,
1550 	.map_delete_elem = htab_map_delete_elem,
1551 	.map_gen_lookup = htab_map_gen_lookup,
1552 	.map_seq_show_elem = htab_map_seq_show_elem,
1553 	BATCH_OPS(htab),
1554 };
1555 
1556 const struct bpf_map_ops htab_lru_map_ops = {
1557 	.map_alloc_check = htab_map_alloc_check,
1558 	.map_alloc = htab_map_alloc,
1559 	.map_free = htab_map_free,
1560 	.map_get_next_key = htab_map_get_next_key,
1561 	.map_lookup_elem = htab_lru_map_lookup_elem,
1562 	.map_lookup_elem_sys_only = htab_lru_map_lookup_elem_sys,
1563 	.map_update_elem = htab_lru_map_update_elem,
1564 	.map_delete_elem = htab_lru_map_delete_elem,
1565 	.map_gen_lookup = htab_lru_map_gen_lookup,
1566 	.map_seq_show_elem = htab_map_seq_show_elem,
1567 	BATCH_OPS(htab_lru),
1568 };
1569 
1570 /* Called from eBPF program */
1571 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1572 {
1573 	struct htab_elem *l = __htab_map_lookup_elem(map, key);
1574 
1575 	if (l)
1576 		return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1577 	else
1578 		return NULL;
1579 }
1580 
1581 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1582 {
1583 	struct htab_elem *l = __htab_map_lookup_elem(map, key);
1584 
1585 	if (l) {
1586 		bpf_lru_node_set_ref(&l->lru_node);
1587 		return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1588 	}
1589 
1590 	return NULL;
1591 }
1592 
1593 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
1594 {
1595 	struct htab_elem *l;
1596 	void __percpu *pptr;
1597 	int ret = -ENOENT;
1598 	int cpu, off = 0;
1599 	u32 size;
1600 
1601 	/* per_cpu areas are zero-filled and bpf programs can only
1602 	 * access 'value_size' of them, so copying rounded areas
1603 	 * will not leak any kernel data
1604 	 */
1605 	size = round_up(map->value_size, 8);
1606 	rcu_read_lock();
1607 	l = __htab_map_lookup_elem(map, key);
1608 	if (!l)
1609 		goto out;
1610 	/* We do not mark LRU map element here in order to not mess up
1611 	 * eviction heuristics when user space does a map walk.
1612 	 */
1613 	pptr = htab_elem_get_ptr(l, map->key_size);
1614 	for_each_possible_cpu(cpu) {
1615 		bpf_long_memcpy(value + off,
1616 				per_cpu_ptr(pptr, cpu), size);
1617 		off += size;
1618 	}
1619 	ret = 0;
1620 out:
1621 	rcu_read_unlock();
1622 	return ret;
1623 }
1624 
1625 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
1626 			   u64 map_flags)
1627 {
1628 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1629 	int ret;
1630 
1631 	rcu_read_lock();
1632 	if (htab_is_lru(htab))
1633 		ret = __htab_lru_percpu_map_update_elem(map, key, value,
1634 							map_flags, true);
1635 	else
1636 		ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
1637 						    true);
1638 	rcu_read_unlock();
1639 
1640 	return ret;
1641 }
1642 
1643 static void htab_percpu_map_seq_show_elem(struct bpf_map *map, void *key,
1644 					  struct seq_file *m)
1645 {
1646 	struct htab_elem *l;
1647 	void __percpu *pptr;
1648 	int cpu;
1649 
1650 	rcu_read_lock();
1651 
1652 	l = __htab_map_lookup_elem(map, key);
1653 	if (!l) {
1654 		rcu_read_unlock();
1655 		return;
1656 	}
1657 
1658 	btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1659 	seq_puts(m, ": {\n");
1660 	pptr = htab_elem_get_ptr(l, map->key_size);
1661 	for_each_possible_cpu(cpu) {
1662 		seq_printf(m, "\tcpu%d: ", cpu);
1663 		btf_type_seq_show(map->btf, map->btf_value_type_id,
1664 				  per_cpu_ptr(pptr, cpu), m);
1665 		seq_puts(m, "\n");
1666 	}
1667 	seq_puts(m, "}\n");
1668 
1669 	rcu_read_unlock();
1670 }
1671 
1672 const struct bpf_map_ops htab_percpu_map_ops = {
1673 	.map_alloc_check = htab_map_alloc_check,
1674 	.map_alloc = htab_map_alloc,
1675 	.map_free = htab_map_free,
1676 	.map_get_next_key = htab_map_get_next_key,
1677 	.map_lookup_elem = htab_percpu_map_lookup_elem,
1678 	.map_update_elem = htab_percpu_map_update_elem,
1679 	.map_delete_elem = htab_map_delete_elem,
1680 	.map_seq_show_elem = htab_percpu_map_seq_show_elem,
1681 	BATCH_OPS(htab_percpu),
1682 };
1683 
1684 const struct bpf_map_ops htab_lru_percpu_map_ops = {
1685 	.map_alloc_check = htab_map_alloc_check,
1686 	.map_alloc = htab_map_alloc,
1687 	.map_free = htab_map_free,
1688 	.map_get_next_key = htab_map_get_next_key,
1689 	.map_lookup_elem = htab_lru_percpu_map_lookup_elem,
1690 	.map_update_elem = htab_lru_percpu_map_update_elem,
1691 	.map_delete_elem = htab_lru_map_delete_elem,
1692 	.map_seq_show_elem = htab_percpu_map_seq_show_elem,
1693 	BATCH_OPS(htab_lru_percpu),
1694 };
1695 
1696 static int fd_htab_map_alloc_check(union bpf_attr *attr)
1697 {
1698 	if (attr->value_size != sizeof(u32))
1699 		return -EINVAL;
1700 	return htab_map_alloc_check(attr);
1701 }
1702 
1703 static void fd_htab_map_free(struct bpf_map *map)
1704 {
1705 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1706 	struct hlist_nulls_node *n;
1707 	struct hlist_nulls_head *head;
1708 	struct htab_elem *l;
1709 	int i;
1710 
1711 	for (i = 0; i < htab->n_buckets; i++) {
1712 		head = select_bucket(htab, i);
1713 
1714 		hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1715 			void *ptr = fd_htab_map_get_ptr(map, l);
1716 
1717 			map->ops->map_fd_put_ptr(ptr);
1718 		}
1719 	}
1720 
1721 	htab_map_free(map);
1722 }
1723 
1724 /* only called from syscall */
1725 int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
1726 {
1727 	void **ptr;
1728 	int ret = 0;
1729 
1730 	if (!map->ops->map_fd_sys_lookup_elem)
1731 		return -ENOTSUPP;
1732 
1733 	rcu_read_lock();
1734 	ptr = htab_map_lookup_elem(map, key);
1735 	if (ptr)
1736 		*value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
1737 	else
1738 		ret = -ENOENT;
1739 	rcu_read_unlock();
1740 
1741 	return ret;
1742 }
1743 
1744 /* only called from syscall */
1745 int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
1746 				void *key, void *value, u64 map_flags)
1747 {
1748 	void *ptr;
1749 	int ret;
1750 	u32 ufd = *(u32 *)value;
1751 
1752 	ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
1753 	if (IS_ERR(ptr))
1754 		return PTR_ERR(ptr);
1755 
1756 	ret = htab_map_update_elem(map, key, &ptr, map_flags);
1757 	if (ret)
1758 		map->ops->map_fd_put_ptr(ptr);
1759 
1760 	return ret;
1761 }
1762 
1763 static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
1764 {
1765 	struct bpf_map *map, *inner_map_meta;
1766 
1767 	inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
1768 	if (IS_ERR(inner_map_meta))
1769 		return inner_map_meta;
1770 
1771 	map = htab_map_alloc(attr);
1772 	if (IS_ERR(map)) {
1773 		bpf_map_meta_free(inner_map_meta);
1774 		return map;
1775 	}
1776 
1777 	map->inner_map_meta = inner_map_meta;
1778 
1779 	return map;
1780 }
1781 
1782 static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
1783 {
1784 	struct bpf_map **inner_map  = htab_map_lookup_elem(map, key);
1785 
1786 	if (!inner_map)
1787 		return NULL;
1788 
1789 	return READ_ONCE(*inner_map);
1790 }
1791 
1792 static u32 htab_of_map_gen_lookup(struct bpf_map *map,
1793 				  struct bpf_insn *insn_buf)
1794 {
1795 	struct bpf_insn *insn = insn_buf;
1796 	const int ret = BPF_REG_0;
1797 
1798 	BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
1799 		     (void *(*)(struct bpf_map *map, void *key))NULL));
1800 	*insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
1801 	*insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
1802 	*insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
1803 				offsetof(struct htab_elem, key) +
1804 				round_up(map->key_size, 8));
1805 	*insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
1806 
1807 	return insn - insn_buf;
1808 }
1809 
1810 static void htab_of_map_free(struct bpf_map *map)
1811 {
1812 	bpf_map_meta_free(map->inner_map_meta);
1813 	fd_htab_map_free(map);
1814 }
1815 
1816 const struct bpf_map_ops htab_of_maps_map_ops = {
1817 	.map_alloc_check = fd_htab_map_alloc_check,
1818 	.map_alloc = htab_of_map_alloc,
1819 	.map_free = htab_of_map_free,
1820 	.map_get_next_key = htab_map_get_next_key,
1821 	.map_lookup_elem = htab_of_map_lookup_elem,
1822 	.map_delete_elem = htab_map_delete_elem,
1823 	.map_fd_get_ptr = bpf_map_fd_get_ptr,
1824 	.map_fd_put_ptr = bpf_map_fd_put_ptr,
1825 	.map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
1826 	.map_gen_lookup = htab_of_map_gen_lookup,
1827 	.map_check_btf = map_check_no_btf,
1828 };
1829