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