xref: /openbmc/linux/kernel/bpf/hashtab.c (revision 505b0877)
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 <linux/rcupdate_trace.h>
13 #include "percpu_freelist.h"
14 #include "bpf_lru_list.h"
15 #include "map_in_map.h"
16 
17 #define HTAB_CREATE_FLAG_MASK						\
18 	(BPF_F_NO_PREALLOC | BPF_F_NO_COMMON_LRU | BPF_F_NUMA_NODE |	\
19 	 BPF_F_ACCESS_MASK | BPF_F_ZERO_SEED)
20 
21 #define BATCH_OPS(_name)			\
22 	.map_lookup_batch =			\
23 	_name##_map_lookup_batch,		\
24 	.map_lookup_and_delete_batch =		\
25 	_name##_map_lookup_and_delete_batch,	\
26 	.map_update_batch =			\
27 	generic_map_update_batch,		\
28 	.map_delete_batch =			\
29 	generic_map_delete_batch
30 
31 /*
32  * The bucket lock has two protection scopes:
33  *
34  * 1) Serializing concurrent operations from BPF programs on differrent
35  *    CPUs
36  *
37  * 2) Serializing concurrent operations from BPF programs and sys_bpf()
38  *
39  * BPF programs can execute in any context including perf, kprobes and
40  * tracing. As there are almost no limits where perf, kprobes and tracing
41  * can be invoked from the lock operations need to be protected against
42  * deadlocks. Deadlocks can be caused by recursion and by an invocation in
43  * the lock held section when functions which acquire this lock are invoked
44  * from sys_bpf(). BPF recursion is prevented by incrementing the per CPU
45  * variable bpf_prog_active, which prevents BPF programs attached to perf
46  * events, kprobes and tracing to be invoked before the prior invocation
47  * from one of these contexts completed. sys_bpf() uses the same mechanism
48  * by pinning the task to the current CPU and incrementing the recursion
49  * protection accross the map operation.
50  *
51  * This has subtle implications on PREEMPT_RT. PREEMPT_RT forbids certain
52  * operations like memory allocations (even with GFP_ATOMIC) from atomic
53  * contexts. This is required because even with GFP_ATOMIC the memory
54  * allocator calls into code pathes which acquire locks with long held lock
55  * sections. To ensure the deterministic behaviour these locks are regular
56  * spinlocks, which are converted to 'sleepable' spinlocks on RT. The only
57  * true atomic contexts on an RT kernel are the low level hardware
58  * handling, scheduling, low level interrupt handling, NMIs etc. None of
59  * these contexts should ever do memory allocations.
60  *
61  * As regular device interrupt handlers and soft interrupts are forced into
62  * thread context, the existing code which does
63  *   spin_lock*(); alloc(GPF_ATOMIC); spin_unlock*();
64  * just works.
65  *
66  * In theory the BPF locks could be converted to regular spinlocks as well,
67  * but the bucket locks and percpu_freelist locks can be taken from
68  * arbitrary contexts (perf, kprobes, tracepoints) which are required to be
69  * atomic contexts even on RT. These mechanisms require preallocated maps,
70  * so there is no need to invoke memory allocations within the lock held
71  * sections.
72  *
73  * BPF maps which need dynamic allocation are only used from (forced)
74  * thread context on RT and can therefore use regular spinlocks which in
75  * turn allows to invoke memory allocations from the lock held section.
76  *
77  * On a non RT kernel this distinction is neither possible nor required.
78  * spinlock maps to raw_spinlock and the extra code is optimized out by the
79  * compiler.
80  */
81 struct bucket {
82 	struct hlist_nulls_head head;
83 	union {
84 		raw_spinlock_t raw_lock;
85 		spinlock_t     lock;
86 	};
87 };
88 
89 struct bpf_htab {
90 	struct bpf_map map;
91 	struct bucket *buckets;
92 	void *elems;
93 	union {
94 		struct pcpu_freelist freelist;
95 		struct bpf_lru lru;
96 	};
97 	struct htab_elem *__percpu *extra_elems;
98 	atomic_t count;	/* number of elements in this hashtable */
99 	u32 n_buckets;	/* number of hash buckets */
100 	u32 elem_size;	/* size of each element in bytes */
101 	u32 hashrnd;
102 };
103 
104 /* each htab element is struct htab_elem + key + value */
105 struct htab_elem {
106 	union {
107 		struct hlist_nulls_node hash_node;
108 		struct {
109 			void *padding;
110 			union {
111 				struct bpf_htab *htab;
112 				struct pcpu_freelist_node fnode;
113 				struct htab_elem *batch_flink;
114 			};
115 		};
116 	};
117 	union {
118 		struct rcu_head rcu;
119 		struct bpf_lru_node lru_node;
120 	};
121 	u32 hash;
122 	char key[] __aligned(8);
123 };
124 
125 static inline bool htab_is_prealloc(const struct bpf_htab *htab)
126 {
127 	return !(htab->map.map_flags & BPF_F_NO_PREALLOC);
128 }
129 
130 static inline bool htab_use_raw_lock(const struct bpf_htab *htab)
131 {
132 	return (!IS_ENABLED(CONFIG_PREEMPT_RT) || htab_is_prealloc(htab));
133 }
134 
135 static void htab_init_buckets(struct bpf_htab *htab)
136 {
137 	unsigned i;
138 
139 	for (i = 0; i < htab->n_buckets; i++) {
140 		INIT_HLIST_NULLS_HEAD(&htab->buckets[i].head, i);
141 		if (htab_use_raw_lock(htab))
142 			raw_spin_lock_init(&htab->buckets[i].raw_lock);
143 		else
144 			spin_lock_init(&htab->buckets[i].lock);
145 	}
146 }
147 
148 static inline unsigned long htab_lock_bucket(const struct bpf_htab *htab,
149 					     struct bucket *b)
150 {
151 	unsigned long flags;
152 
153 	if (htab_use_raw_lock(htab))
154 		raw_spin_lock_irqsave(&b->raw_lock, flags);
155 	else
156 		spin_lock_irqsave(&b->lock, flags);
157 	return flags;
158 }
159 
160 static inline void htab_unlock_bucket(const struct bpf_htab *htab,
161 				      struct bucket *b,
162 				      unsigned long flags)
163 {
164 	if (htab_use_raw_lock(htab))
165 		raw_spin_unlock_irqrestore(&b->raw_lock, flags);
166 	else
167 		spin_unlock_irqrestore(&b->lock, flags);
168 }
169 
170 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
171 
172 static bool htab_is_lru(const struct bpf_htab *htab)
173 {
174 	return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH ||
175 		htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
176 }
177 
178 static bool htab_is_percpu(const struct bpf_htab *htab)
179 {
180 	return htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH ||
181 		htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
182 }
183 
184 static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
185 				     void __percpu *pptr)
186 {
187 	*(void __percpu **)(l->key + key_size) = pptr;
188 }
189 
190 static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
191 {
192 	return *(void __percpu **)(l->key + key_size);
193 }
194 
195 static void *fd_htab_map_get_ptr(const struct bpf_map *map, struct htab_elem *l)
196 {
197 	return *(void **)(l->key + roundup(map->key_size, 8));
198 }
199 
200 static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
201 {
202 	return (struct htab_elem *) (htab->elems + i * htab->elem_size);
203 }
204 
205 static void htab_free_elems(struct bpf_htab *htab)
206 {
207 	int i;
208 
209 	if (!htab_is_percpu(htab))
210 		goto free_elems;
211 
212 	for (i = 0; i < htab->map.max_entries; i++) {
213 		void __percpu *pptr;
214 
215 		pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
216 					 htab->map.key_size);
217 		free_percpu(pptr);
218 		cond_resched();
219 	}
220 free_elems:
221 	bpf_map_area_free(htab->elems);
222 }
223 
224 /* The LRU list has a lock (lru_lock). Each htab bucket has a lock
225  * (bucket_lock). If both locks need to be acquired together, the lock
226  * order is always lru_lock -> bucket_lock and this only happens in
227  * bpf_lru_list.c logic. For example, certain code path of
228  * bpf_lru_pop_free(), which is called by function prealloc_lru_pop(),
229  * will acquire lru_lock first followed by acquiring bucket_lock.
230  *
231  * In hashtab.c, to avoid deadlock, lock acquisition of
232  * bucket_lock followed by lru_lock is not allowed. In such cases,
233  * bucket_lock needs to be released first before acquiring lru_lock.
234  */
235 static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
236 					  u32 hash)
237 {
238 	struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
239 	struct htab_elem *l;
240 
241 	if (node) {
242 		l = container_of(node, struct htab_elem, lru_node);
243 		memcpy(l->key, key, htab->map.key_size);
244 		return l;
245 	}
246 
247 	return NULL;
248 }
249 
250 static int prealloc_init(struct bpf_htab *htab)
251 {
252 	u32 num_entries = htab->map.max_entries;
253 	int err = -ENOMEM, i;
254 
255 	if (!htab_is_percpu(htab) && !htab_is_lru(htab))
256 		num_entries += num_possible_cpus();
257 
258 	htab->elems = bpf_map_area_alloc(htab->elem_size * num_entries,
259 					 htab->map.numa_node);
260 	if (!htab->elems)
261 		return -ENOMEM;
262 
263 	if (!htab_is_percpu(htab))
264 		goto skip_percpu_elems;
265 
266 	for (i = 0; i < num_entries; i++) {
267 		u32 size = round_up(htab->map.value_size, 8);
268 		void __percpu *pptr;
269 
270 		pptr = __alloc_percpu_gfp(size, 8, GFP_USER | __GFP_NOWARN);
271 		if (!pptr)
272 			goto free_elems;
273 		htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
274 				  pptr);
275 		cond_resched();
276 	}
277 
278 skip_percpu_elems:
279 	if (htab_is_lru(htab))
280 		err = bpf_lru_init(&htab->lru,
281 				   htab->map.map_flags & BPF_F_NO_COMMON_LRU,
282 				   offsetof(struct htab_elem, hash) -
283 				   offsetof(struct htab_elem, lru_node),
284 				   htab_lru_map_delete_node,
285 				   htab);
286 	else
287 		err = pcpu_freelist_init(&htab->freelist);
288 
289 	if (err)
290 		goto free_elems;
291 
292 	if (htab_is_lru(htab))
293 		bpf_lru_populate(&htab->lru, htab->elems,
294 				 offsetof(struct htab_elem, lru_node),
295 				 htab->elem_size, num_entries);
296 	else
297 		pcpu_freelist_populate(&htab->freelist,
298 				       htab->elems + offsetof(struct htab_elem, fnode),
299 				       htab->elem_size, num_entries);
300 
301 	return 0;
302 
303 free_elems:
304 	htab_free_elems(htab);
305 	return err;
306 }
307 
308 static void prealloc_destroy(struct bpf_htab *htab)
309 {
310 	htab_free_elems(htab);
311 
312 	if (htab_is_lru(htab))
313 		bpf_lru_destroy(&htab->lru);
314 	else
315 		pcpu_freelist_destroy(&htab->freelist);
316 }
317 
318 static int alloc_extra_elems(struct bpf_htab *htab)
319 {
320 	struct htab_elem *__percpu *pptr, *l_new;
321 	struct pcpu_freelist_node *l;
322 	int cpu;
323 
324 	pptr = __alloc_percpu_gfp(sizeof(struct htab_elem *), 8,
325 				  GFP_USER | __GFP_NOWARN);
326 	if (!pptr)
327 		return -ENOMEM;
328 
329 	for_each_possible_cpu(cpu) {
330 		l = pcpu_freelist_pop(&htab->freelist);
331 		/* pop will succeed, since prealloc_init()
332 		 * preallocated extra num_possible_cpus elements
333 		 */
334 		l_new = container_of(l, struct htab_elem, fnode);
335 		*per_cpu_ptr(pptr, cpu) = l_new;
336 	}
337 	htab->extra_elems = pptr;
338 	return 0;
339 }
340 
341 /* Called from syscall */
342 static int htab_map_alloc_check(union bpf_attr *attr)
343 {
344 	bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
345 		       attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
346 	bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
347 		    attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
348 	/* percpu_lru means each cpu has its own LRU list.
349 	 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
350 	 * the map's value itself is percpu.  percpu_lru has
351 	 * nothing to do with the map's value.
352 	 */
353 	bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
354 	bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
355 	bool zero_seed = (attr->map_flags & BPF_F_ZERO_SEED);
356 	int numa_node = bpf_map_attr_numa_node(attr);
357 
358 	BUILD_BUG_ON(offsetof(struct htab_elem, htab) !=
359 		     offsetof(struct htab_elem, hash_node.pprev));
360 	BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
361 		     offsetof(struct htab_elem, hash_node.pprev));
362 
363 	if (lru && !bpf_capable())
364 		/* LRU implementation is much complicated than other
365 		 * maps.  Hence, limit to CAP_BPF.
366 		 */
367 		return -EPERM;
368 
369 	if (zero_seed && !capable(CAP_SYS_ADMIN))
370 		/* Guard against local DoS, and discourage production use. */
371 		return -EPERM;
372 
373 	if (attr->map_flags & ~HTAB_CREATE_FLAG_MASK ||
374 	    !bpf_map_flags_access_ok(attr->map_flags))
375 		return -EINVAL;
376 
377 	if (!lru && percpu_lru)
378 		return -EINVAL;
379 
380 	if (lru && !prealloc)
381 		return -ENOTSUPP;
382 
383 	if (numa_node != NUMA_NO_NODE && (percpu || percpu_lru))
384 		return -EINVAL;
385 
386 	/* check sanity of attributes.
387 	 * value_size == 0 may be allowed in the future to use map as a set
388 	 */
389 	if (attr->max_entries == 0 || attr->key_size == 0 ||
390 	    attr->value_size == 0)
391 		return -EINVAL;
392 
393 	if (attr->key_size > MAX_BPF_STACK)
394 		/* eBPF programs initialize keys on stack, so they cannot be
395 		 * larger than max stack size
396 		 */
397 		return -E2BIG;
398 
399 	if (attr->value_size >= KMALLOC_MAX_SIZE -
400 	    MAX_BPF_STACK - sizeof(struct htab_elem))
401 		/* if value_size is bigger, the user space won't be able to
402 		 * access the elements via bpf syscall. This check also makes
403 		 * sure that the elem_size doesn't overflow and it's
404 		 * kmalloc-able later in htab_map_update_elem()
405 		 */
406 		return -E2BIG;
407 
408 	return 0;
409 }
410 
411 static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
412 {
413 	bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
414 		       attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
415 	bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
416 		    attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
417 	/* percpu_lru means each cpu has its own LRU list.
418 	 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
419 	 * the map's value itself is percpu.  percpu_lru has
420 	 * nothing to do with the map's value.
421 	 */
422 	bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
423 	bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
424 	struct bpf_htab *htab;
425 	u64 cost;
426 	int err;
427 
428 	htab = kzalloc(sizeof(*htab), GFP_USER);
429 	if (!htab)
430 		return ERR_PTR(-ENOMEM);
431 
432 	bpf_map_init_from_attr(&htab->map, attr);
433 
434 	if (percpu_lru) {
435 		/* ensure each CPU's lru list has >=1 elements.
436 		 * since we are at it, make each lru list has the same
437 		 * number of elements.
438 		 */
439 		htab->map.max_entries = roundup(attr->max_entries,
440 						num_possible_cpus());
441 		if (htab->map.max_entries < attr->max_entries)
442 			htab->map.max_entries = rounddown(attr->max_entries,
443 							  num_possible_cpus());
444 	}
445 
446 	/* hash table size must be power of 2 */
447 	htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
448 
449 	htab->elem_size = sizeof(struct htab_elem) +
450 			  round_up(htab->map.key_size, 8);
451 	if (percpu)
452 		htab->elem_size += sizeof(void *);
453 	else
454 		htab->elem_size += round_up(htab->map.value_size, 8);
455 
456 	err = -E2BIG;
457 	/* prevent zero size kmalloc and check for u32 overflow */
458 	if (htab->n_buckets == 0 ||
459 	    htab->n_buckets > U32_MAX / sizeof(struct bucket))
460 		goto free_htab;
461 
462 	cost = (u64) htab->n_buckets * sizeof(struct bucket) +
463 	       (u64) htab->elem_size * htab->map.max_entries;
464 
465 	if (percpu)
466 		cost += (u64) round_up(htab->map.value_size, 8) *
467 			num_possible_cpus() * htab->map.max_entries;
468 	else
469 	       cost += (u64) htab->elem_size * num_possible_cpus();
470 
471 	/* if map size is larger than memlock limit, reject it */
472 	err = bpf_map_charge_init(&htab->map.memory, cost);
473 	if (err)
474 		goto free_htab;
475 
476 	err = -ENOMEM;
477 	htab->buckets = bpf_map_area_alloc(htab->n_buckets *
478 					   sizeof(struct bucket),
479 					   htab->map.numa_node);
480 	if (!htab->buckets)
481 		goto free_charge;
482 
483 	if (htab->map.map_flags & BPF_F_ZERO_SEED)
484 		htab->hashrnd = 0;
485 	else
486 		htab->hashrnd = get_random_int();
487 
488 	htab_init_buckets(htab);
489 
490 	if (prealloc) {
491 		err = prealloc_init(htab);
492 		if (err)
493 			goto free_buckets;
494 
495 		if (!percpu && !lru) {
496 			/* lru itself can remove the least used element, so
497 			 * there is no need for an extra elem during map_update.
498 			 */
499 			err = alloc_extra_elems(htab);
500 			if (err)
501 				goto free_prealloc;
502 		}
503 	}
504 
505 	return &htab->map;
506 
507 free_prealloc:
508 	prealloc_destroy(htab);
509 free_buckets:
510 	bpf_map_area_free(htab->buckets);
511 free_charge:
512 	bpf_map_charge_finish(&htab->map.memory);
513 free_htab:
514 	kfree(htab);
515 	return ERR_PTR(err);
516 }
517 
518 static inline u32 htab_map_hash(const void *key, u32 key_len, u32 hashrnd)
519 {
520 	return jhash(key, key_len, hashrnd);
521 }
522 
523 static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
524 {
525 	return &htab->buckets[hash & (htab->n_buckets - 1)];
526 }
527 
528 static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
529 {
530 	return &__select_bucket(htab, hash)->head;
531 }
532 
533 /* this lookup function can only be called with bucket lock taken */
534 static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
535 					 void *key, u32 key_size)
536 {
537 	struct hlist_nulls_node *n;
538 	struct htab_elem *l;
539 
540 	hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
541 		if (l->hash == hash && !memcmp(&l->key, key, key_size))
542 			return l;
543 
544 	return NULL;
545 }
546 
547 /* can be called without bucket lock. it will repeat the loop in
548  * the unlikely event when elements moved from one bucket into another
549  * while link list is being walked
550  */
551 static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
552 					       u32 hash, void *key,
553 					       u32 key_size, u32 n_buckets)
554 {
555 	struct hlist_nulls_node *n;
556 	struct htab_elem *l;
557 
558 again:
559 	hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
560 		if (l->hash == hash && !memcmp(&l->key, key, key_size))
561 			return l;
562 
563 	if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
564 		goto again;
565 
566 	return NULL;
567 }
568 
569 /* Called from syscall or from eBPF program directly, so
570  * arguments have to match bpf_map_lookup_elem() exactly.
571  * The return value is adjusted by BPF instructions
572  * in htab_map_gen_lookup().
573  */
574 static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
575 {
576 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
577 	struct hlist_nulls_head *head;
578 	struct htab_elem *l;
579 	u32 hash, key_size;
580 
581 	WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held());
582 
583 	key_size = map->key_size;
584 
585 	hash = htab_map_hash(key, key_size, htab->hashrnd);
586 
587 	head = select_bucket(htab, hash);
588 
589 	l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
590 
591 	return l;
592 }
593 
594 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
595 {
596 	struct htab_elem *l = __htab_map_lookup_elem(map, key);
597 
598 	if (l)
599 		return l->key + round_up(map->key_size, 8);
600 
601 	return NULL;
602 }
603 
604 /* inline bpf_map_lookup_elem() call.
605  * Instead of:
606  * bpf_prog
607  *   bpf_map_lookup_elem
608  *     map->ops->map_lookup_elem
609  *       htab_map_lookup_elem
610  *         __htab_map_lookup_elem
611  * do:
612  * bpf_prog
613  *   __htab_map_lookup_elem
614  */
615 static int htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
616 {
617 	struct bpf_insn *insn = insn_buf;
618 	const int ret = BPF_REG_0;
619 
620 	BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
621 		     (void *(*)(struct bpf_map *map, void *key))NULL));
622 	*insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
623 	*insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
624 	*insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
625 				offsetof(struct htab_elem, key) +
626 				round_up(map->key_size, 8));
627 	return insn - insn_buf;
628 }
629 
630 static __always_inline void *__htab_lru_map_lookup_elem(struct bpf_map *map,
631 							void *key, const bool mark)
632 {
633 	struct htab_elem *l = __htab_map_lookup_elem(map, key);
634 
635 	if (l) {
636 		if (mark)
637 			bpf_lru_node_set_ref(&l->lru_node);
638 		return l->key + round_up(map->key_size, 8);
639 	}
640 
641 	return NULL;
642 }
643 
644 static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
645 {
646 	return __htab_lru_map_lookup_elem(map, key, true);
647 }
648 
649 static void *htab_lru_map_lookup_elem_sys(struct bpf_map *map, void *key)
650 {
651 	return __htab_lru_map_lookup_elem(map, key, false);
652 }
653 
654 static int htab_lru_map_gen_lookup(struct bpf_map *map,
655 				   struct bpf_insn *insn_buf)
656 {
657 	struct bpf_insn *insn = insn_buf;
658 	const int ret = BPF_REG_0;
659 	const int ref_reg = BPF_REG_1;
660 
661 	BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
662 		     (void *(*)(struct bpf_map *map, void *key))NULL));
663 	*insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
664 	*insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
665 	*insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
666 			      offsetof(struct htab_elem, lru_node) +
667 			      offsetof(struct bpf_lru_node, ref));
668 	*insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
669 	*insn++ = BPF_ST_MEM(BPF_B, ret,
670 			     offsetof(struct htab_elem, lru_node) +
671 			     offsetof(struct bpf_lru_node, ref),
672 			     1);
673 	*insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
674 				offsetof(struct htab_elem, key) +
675 				round_up(map->key_size, 8));
676 	return insn - insn_buf;
677 }
678 
679 /* It is called from the bpf_lru_list when the LRU needs to delete
680  * older elements from the htab.
681  */
682 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
683 {
684 	struct bpf_htab *htab = (struct bpf_htab *)arg;
685 	struct htab_elem *l = NULL, *tgt_l;
686 	struct hlist_nulls_head *head;
687 	struct hlist_nulls_node *n;
688 	unsigned long flags;
689 	struct bucket *b;
690 
691 	tgt_l = container_of(node, struct htab_elem, lru_node);
692 	b = __select_bucket(htab, tgt_l->hash);
693 	head = &b->head;
694 
695 	flags = htab_lock_bucket(htab, b);
696 
697 	hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
698 		if (l == tgt_l) {
699 			hlist_nulls_del_rcu(&l->hash_node);
700 			break;
701 		}
702 
703 	htab_unlock_bucket(htab, b, flags);
704 
705 	return l == tgt_l;
706 }
707 
708 /* Called from syscall */
709 static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
710 {
711 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
712 	struct hlist_nulls_head *head;
713 	struct htab_elem *l, *next_l;
714 	u32 hash, key_size;
715 	int i = 0;
716 
717 	WARN_ON_ONCE(!rcu_read_lock_held());
718 
719 	key_size = map->key_size;
720 
721 	if (!key)
722 		goto find_first_elem;
723 
724 	hash = htab_map_hash(key, key_size, htab->hashrnd);
725 
726 	head = select_bucket(htab, hash);
727 
728 	/* lookup the key */
729 	l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
730 
731 	if (!l)
732 		goto find_first_elem;
733 
734 	/* key was found, get next key in the same bucket */
735 	next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
736 				  struct htab_elem, hash_node);
737 
738 	if (next_l) {
739 		/* if next elem in this hash list is non-zero, just return it */
740 		memcpy(next_key, next_l->key, key_size);
741 		return 0;
742 	}
743 
744 	/* no more elements in this hash list, go to the next bucket */
745 	i = hash & (htab->n_buckets - 1);
746 	i++;
747 
748 find_first_elem:
749 	/* iterate over buckets */
750 	for (; i < htab->n_buckets; i++) {
751 		head = select_bucket(htab, i);
752 
753 		/* pick first element in the bucket */
754 		next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
755 					  struct htab_elem, hash_node);
756 		if (next_l) {
757 			/* if it's not empty, just return it */
758 			memcpy(next_key, next_l->key, key_size);
759 			return 0;
760 		}
761 	}
762 
763 	/* iterated over all buckets and all elements */
764 	return -ENOENT;
765 }
766 
767 static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
768 {
769 	if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
770 		free_percpu(htab_elem_get_ptr(l, htab->map.key_size));
771 	kfree(l);
772 }
773 
774 static void htab_elem_free_rcu(struct rcu_head *head)
775 {
776 	struct htab_elem *l = container_of(head, struct htab_elem, rcu);
777 	struct bpf_htab *htab = l->htab;
778 
779 	htab_elem_free(htab, l);
780 }
781 
782 static void htab_put_fd_value(struct bpf_htab *htab, struct htab_elem *l)
783 {
784 	struct bpf_map *map = &htab->map;
785 	void *ptr;
786 
787 	if (map->ops->map_fd_put_ptr) {
788 		ptr = fd_htab_map_get_ptr(map, l);
789 		map->ops->map_fd_put_ptr(ptr);
790 	}
791 }
792 
793 static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
794 {
795 	htab_put_fd_value(htab, l);
796 
797 	if (htab_is_prealloc(htab)) {
798 		__pcpu_freelist_push(&htab->freelist, &l->fnode);
799 	} else {
800 		atomic_dec(&htab->count);
801 		l->htab = htab;
802 		call_rcu(&l->rcu, htab_elem_free_rcu);
803 	}
804 }
805 
806 static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
807 			    void *value, bool onallcpus)
808 {
809 	if (!onallcpus) {
810 		/* copy true value_size bytes */
811 		memcpy(this_cpu_ptr(pptr), value, htab->map.value_size);
812 	} else {
813 		u32 size = round_up(htab->map.value_size, 8);
814 		int off = 0, cpu;
815 
816 		for_each_possible_cpu(cpu) {
817 			bpf_long_memcpy(per_cpu_ptr(pptr, cpu),
818 					value + off, size);
819 			off += size;
820 		}
821 	}
822 }
823 
824 static void pcpu_init_value(struct bpf_htab *htab, void __percpu *pptr,
825 			    void *value, bool onallcpus)
826 {
827 	/* When using prealloc and not setting the initial value on all cpus,
828 	 * zero-fill element values for other cpus (just as what happens when
829 	 * not using prealloc). Otherwise, bpf program has no way to ensure
830 	 * known initial values for cpus other than current one
831 	 * (onallcpus=false always when coming from bpf prog).
832 	 */
833 	if (htab_is_prealloc(htab) && !onallcpus) {
834 		u32 size = round_up(htab->map.value_size, 8);
835 		int current_cpu = raw_smp_processor_id();
836 		int cpu;
837 
838 		for_each_possible_cpu(cpu) {
839 			if (cpu == current_cpu)
840 				bpf_long_memcpy(per_cpu_ptr(pptr, cpu), value,
841 						size);
842 			else
843 				memset(per_cpu_ptr(pptr, cpu), 0, size);
844 		}
845 	} else {
846 		pcpu_copy_value(htab, pptr, value, onallcpus);
847 	}
848 }
849 
850 static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
851 {
852 	return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
853 	       BITS_PER_LONG == 64;
854 }
855 
856 static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
857 					 void *value, u32 key_size, u32 hash,
858 					 bool percpu, bool onallcpus,
859 					 struct htab_elem *old_elem)
860 {
861 	u32 size = htab->map.value_size;
862 	bool prealloc = htab_is_prealloc(htab);
863 	struct htab_elem *l_new, **pl_new;
864 	void __percpu *pptr;
865 
866 	if (prealloc) {
867 		if (old_elem) {
868 			/* if we're updating the existing element,
869 			 * use per-cpu extra elems to avoid freelist_pop/push
870 			 */
871 			pl_new = this_cpu_ptr(htab->extra_elems);
872 			l_new = *pl_new;
873 			htab_put_fd_value(htab, old_elem);
874 			*pl_new = old_elem;
875 		} else {
876 			struct pcpu_freelist_node *l;
877 
878 			l = __pcpu_freelist_pop(&htab->freelist);
879 			if (!l)
880 				return ERR_PTR(-E2BIG);
881 			l_new = container_of(l, struct htab_elem, fnode);
882 		}
883 	} else {
884 		if (atomic_inc_return(&htab->count) > htab->map.max_entries)
885 			if (!old_elem) {
886 				/* when map is full and update() is replacing
887 				 * old element, it's ok to allocate, since
888 				 * old element will be freed immediately.
889 				 * Otherwise return an error
890 				 */
891 				l_new = ERR_PTR(-E2BIG);
892 				goto dec_count;
893 			}
894 		l_new = kmalloc_node(htab->elem_size, GFP_ATOMIC | __GFP_NOWARN,
895 				     htab->map.numa_node);
896 		if (!l_new) {
897 			l_new = ERR_PTR(-ENOMEM);
898 			goto dec_count;
899 		}
900 		check_and_init_map_lock(&htab->map,
901 					l_new->key + round_up(key_size, 8));
902 	}
903 
904 	memcpy(l_new->key, key, key_size);
905 	if (percpu) {
906 		size = round_up(size, 8);
907 		if (prealloc) {
908 			pptr = htab_elem_get_ptr(l_new, key_size);
909 		} else {
910 			/* alloc_percpu zero-fills */
911 			pptr = __alloc_percpu_gfp(size, 8,
912 						  GFP_ATOMIC | __GFP_NOWARN);
913 			if (!pptr) {
914 				kfree(l_new);
915 				l_new = ERR_PTR(-ENOMEM);
916 				goto dec_count;
917 			}
918 		}
919 
920 		pcpu_init_value(htab, pptr, value, onallcpus);
921 
922 		if (!prealloc)
923 			htab_elem_set_ptr(l_new, key_size, pptr);
924 	} else if (fd_htab_map_needs_adjust(htab)) {
925 		size = round_up(size, 8);
926 		memcpy(l_new->key + round_up(key_size, 8), value, size);
927 	} else {
928 		copy_map_value(&htab->map,
929 			       l_new->key + round_up(key_size, 8),
930 			       value);
931 	}
932 
933 	l_new->hash = hash;
934 	return l_new;
935 dec_count:
936 	atomic_dec(&htab->count);
937 	return l_new;
938 }
939 
940 static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
941 		       u64 map_flags)
942 {
943 	if (l_old && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
944 		/* elem already exists */
945 		return -EEXIST;
946 
947 	if (!l_old && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
948 		/* elem doesn't exist, cannot update it */
949 		return -ENOENT;
950 
951 	return 0;
952 }
953 
954 /* Called from syscall or from eBPF program */
955 static int htab_map_update_elem(struct bpf_map *map, void *key, void *value,
956 				u64 map_flags)
957 {
958 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
959 	struct htab_elem *l_new = NULL, *l_old;
960 	struct hlist_nulls_head *head;
961 	unsigned long flags;
962 	struct bucket *b;
963 	u32 key_size, hash;
964 	int ret;
965 
966 	if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST))
967 		/* unknown flags */
968 		return -EINVAL;
969 
970 	WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held());
971 
972 	key_size = map->key_size;
973 
974 	hash = htab_map_hash(key, key_size, htab->hashrnd);
975 
976 	b = __select_bucket(htab, hash);
977 	head = &b->head;
978 
979 	if (unlikely(map_flags & BPF_F_LOCK)) {
980 		if (unlikely(!map_value_has_spin_lock(map)))
981 			return -EINVAL;
982 		/* find an element without taking the bucket lock */
983 		l_old = lookup_nulls_elem_raw(head, hash, key, key_size,
984 					      htab->n_buckets);
985 		ret = check_flags(htab, l_old, map_flags);
986 		if (ret)
987 			return ret;
988 		if (l_old) {
989 			/* grab the element lock and update value in place */
990 			copy_map_value_locked(map,
991 					      l_old->key + round_up(key_size, 8),
992 					      value, false);
993 			return 0;
994 		}
995 		/* fall through, grab the bucket lock and lookup again.
996 		 * 99.9% chance that the element won't be found,
997 		 * but second lookup under lock has to be done.
998 		 */
999 	}
1000 
1001 	flags = htab_lock_bucket(htab, b);
1002 
1003 	l_old = lookup_elem_raw(head, hash, key, key_size);
1004 
1005 	ret = check_flags(htab, l_old, map_flags);
1006 	if (ret)
1007 		goto err;
1008 
1009 	if (unlikely(l_old && (map_flags & BPF_F_LOCK))) {
1010 		/* first lookup without the bucket lock didn't find the element,
1011 		 * but second lookup with the bucket lock found it.
1012 		 * This case is highly unlikely, but has to be dealt with:
1013 		 * grab the element lock in addition to the bucket lock
1014 		 * and update element in place
1015 		 */
1016 		copy_map_value_locked(map,
1017 				      l_old->key + round_up(key_size, 8),
1018 				      value, false);
1019 		ret = 0;
1020 		goto err;
1021 	}
1022 
1023 	l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
1024 				l_old);
1025 	if (IS_ERR(l_new)) {
1026 		/* all pre-allocated elements are in use or memory exhausted */
1027 		ret = PTR_ERR(l_new);
1028 		goto err;
1029 	}
1030 
1031 	/* add new element to the head of the list, so that
1032 	 * concurrent search will find it before old elem
1033 	 */
1034 	hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1035 	if (l_old) {
1036 		hlist_nulls_del_rcu(&l_old->hash_node);
1037 		if (!htab_is_prealloc(htab))
1038 			free_htab_elem(htab, l_old);
1039 	}
1040 	ret = 0;
1041 err:
1042 	htab_unlock_bucket(htab, b, flags);
1043 	return ret;
1044 }
1045 
1046 static int htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
1047 				    u64 map_flags)
1048 {
1049 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1050 	struct htab_elem *l_new, *l_old = NULL;
1051 	struct hlist_nulls_head *head;
1052 	unsigned long flags;
1053 	struct bucket *b;
1054 	u32 key_size, hash;
1055 	int ret;
1056 
1057 	if (unlikely(map_flags > BPF_EXIST))
1058 		/* unknown flags */
1059 		return -EINVAL;
1060 
1061 	WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held());
1062 
1063 	key_size = map->key_size;
1064 
1065 	hash = htab_map_hash(key, key_size, htab->hashrnd);
1066 
1067 	b = __select_bucket(htab, hash);
1068 	head = &b->head;
1069 
1070 	/* For LRU, we need to alloc before taking bucket's
1071 	 * spinlock because getting free nodes from LRU may need
1072 	 * to remove older elements from htab and this removal
1073 	 * operation will need a bucket lock.
1074 	 */
1075 	l_new = prealloc_lru_pop(htab, key, hash);
1076 	if (!l_new)
1077 		return -ENOMEM;
1078 	memcpy(l_new->key + round_up(map->key_size, 8), value, map->value_size);
1079 
1080 	flags = htab_lock_bucket(htab, b);
1081 
1082 	l_old = lookup_elem_raw(head, hash, key, key_size);
1083 
1084 	ret = check_flags(htab, l_old, map_flags);
1085 	if (ret)
1086 		goto err;
1087 
1088 	/* add new element to the head of the list, so that
1089 	 * concurrent search will find it before old elem
1090 	 */
1091 	hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1092 	if (l_old) {
1093 		bpf_lru_node_set_ref(&l_new->lru_node);
1094 		hlist_nulls_del_rcu(&l_old->hash_node);
1095 	}
1096 	ret = 0;
1097 
1098 err:
1099 	htab_unlock_bucket(htab, b, flags);
1100 
1101 	if (ret)
1102 		bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1103 	else if (l_old)
1104 		bpf_lru_push_free(&htab->lru, &l_old->lru_node);
1105 
1106 	return ret;
1107 }
1108 
1109 static int __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1110 					 void *value, u64 map_flags,
1111 					 bool onallcpus)
1112 {
1113 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1114 	struct htab_elem *l_new = NULL, *l_old;
1115 	struct hlist_nulls_head *head;
1116 	unsigned long flags;
1117 	struct bucket *b;
1118 	u32 key_size, hash;
1119 	int ret;
1120 
1121 	if (unlikely(map_flags > BPF_EXIST))
1122 		/* unknown flags */
1123 		return -EINVAL;
1124 
1125 	WARN_ON_ONCE(!rcu_read_lock_held());
1126 
1127 	key_size = map->key_size;
1128 
1129 	hash = htab_map_hash(key, key_size, htab->hashrnd);
1130 
1131 	b = __select_bucket(htab, hash);
1132 	head = &b->head;
1133 
1134 	flags = htab_lock_bucket(htab, b);
1135 
1136 	l_old = lookup_elem_raw(head, hash, key, key_size);
1137 
1138 	ret = check_flags(htab, l_old, map_flags);
1139 	if (ret)
1140 		goto err;
1141 
1142 	if (l_old) {
1143 		/* per-cpu hash map can update value in-place */
1144 		pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1145 				value, onallcpus);
1146 	} else {
1147 		l_new = alloc_htab_elem(htab, key, value, key_size,
1148 					hash, true, onallcpus, NULL);
1149 		if (IS_ERR(l_new)) {
1150 			ret = PTR_ERR(l_new);
1151 			goto err;
1152 		}
1153 		hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1154 	}
1155 	ret = 0;
1156 err:
1157 	htab_unlock_bucket(htab, b, flags);
1158 	return ret;
1159 }
1160 
1161 static int __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1162 					     void *value, u64 map_flags,
1163 					     bool onallcpus)
1164 {
1165 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1166 	struct htab_elem *l_new = NULL, *l_old;
1167 	struct hlist_nulls_head *head;
1168 	unsigned long flags;
1169 	struct bucket *b;
1170 	u32 key_size, hash;
1171 	int ret;
1172 
1173 	if (unlikely(map_flags > BPF_EXIST))
1174 		/* unknown flags */
1175 		return -EINVAL;
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 
1183 	b = __select_bucket(htab, hash);
1184 	head = &b->head;
1185 
1186 	/* For LRU, we need to alloc before taking bucket's
1187 	 * spinlock because LRU's elem alloc may need
1188 	 * to remove older elem from htab and this removal
1189 	 * operation will need a bucket lock.
1190 	 */
1191 	if (map_flags != BPF_EXIST) {
1192 		l_new = prealloc_lru_pop(htab, key, hash);
1193 		if (!l_new)
1194 			return -ENOMEM;
1195 	}
1196 
1197 	flags = htab_lock_bucket(htab, b);
1198 
1199 	l_old = lookup_elem_raw(head, hash, key, key_size);
1200 
1201 	ret = check_flags(htab, l_old, map_flags);
1202 	if (ret)
1203 		goto err;
1204 
1205 	if (l_old) {
1206 		bpf_lru_node_set_ref(&l_old->lru_node);
1207 
1208 		/* per-cpu hash map can update value in-place */
1209 		pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1210 				value, onallcpus);
1211 	} else {
1212 		pcpu_init_value(htab, htab_elem_get_ptr(l_new, key_size),
1213 				value, onallcpus);
1214 		hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1215 		l_new = NULL;
1216 	}
1217 	ret = 0;
1218 err:
1219 	htab_unlock_bucket(htab, b, flags);
1220 	if (l_new)
1221 		bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1222 	return ret;
1223 }
1224 
1225 static int htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1226 				       void *value, u64 map_flags)
1227 {
1228 	return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
1229 }
1230 
1231 static int htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1232 					   void *value, u64 map_flags)
1233 {
1234 	return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1235 						 false);
1236 }
1237 
1238 /* Called from syscall or from eBPF program */
1239 static int htab_map_delete_elem(struct bpf_map *map, void *key)
1240 {
1241 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1242 	struct hlist_nulls_head *head;
1243 	struct bucket *b;
1244 	struct htab_elem *l;
1245 	unsigned long flags;
1246 	u32 hash, key_size;
1247 	int ret = -ENOENT;
1248 
1249 	WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held());
1250 
1251 	key_size = map->key_size;
1252 
1253 	hash = htab_map_hash(key, key_size, htab->hashrnd);
1254 	b = __select_bucket(htab, hash);
1255 	head = &b->head;
1256 
1257 	flags = htab_lock_bucket(htab, b);
1258 
1259 	l = lookup_elem_raw(head, hash, key, key_size);
1260 
1261 	if (l) {
1262 		hlist_nulls_del_rcu(&l->hash_node);
1263 		free_htab_elem(htab, l);
1264 		ret = 0;
1265 	}
1266 
1267 	htab_unlock_bucket(htab, b, flags);
1268 	return ret;
1269 }
1270 
1271 static int htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1272 {
1273 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1274 	struct hlist_nulls_head *head;
1275 	struct bucket *b;
1276 	struct htab_elem *l;
1277 	unsigned long flags;
1278 	u32 hash, key_size;
1279 	int ret = -ENOENT;
1280 
1281 	WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held());
1282 
1283 	key_size = map->key_size;
1284 
1285 	hash = htab_map_hash(key, key_size, htab->hashrnd);
1286 	b = __select_bucket(htab, hash);
1287 	head = &b->head;
1288 
1289 	flags = htab_lock_bucket(htab, b);
1290 
1291 	l = lookup_elem_raw(head, hash, key, key_size);
1292 
1293 	if (l) {
1294 		hlist_nulls_del_rcu(&l->hash_node);
1295 		ret = 0;
1296 	}
1297 
1298 	htab_unlock_bucket(htab, b, flags);
1299 	if (l)
1300 		bpf_lru_push_free(&htab->lru, &l->lru_node);
1301 	return ret;
1302 }
1303 
1304 static void delete_all_elements(struct bpf_htab *htab)
1305 {
1306 	int i;
1307 
1308 	for (i = 0; i < htab->n_buckets; i++) {
1309 		struct hlist_nulls_head *head = select_bucket(htab, i);
1310 		struct hlist_nulls_node *n;
1311 		struct htab_elem *l;
1312 
1313 		hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1314 			hlist_nulls_del_rcu(&l->hash_node);
1315 			htab_elem_free(htab, l);
1316 		}
1317 	}
1318 }
1319 
1320 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1321 static void htab_map_free(struct bpf_map *map)
1322 {
1323 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1324 
1325 	/* bpf_free_used_maps() or close(map_fd) will trigger this map_free callback.
1326 	 * bpf_free_used_maps() is called after bpf prog is no longer executing.
1327 	 * There is no need to synchronize_rcu() here to protect map elements.
1328 	 */
1329 
1330 	/* some of free_htab_elem() callbacks for elements of this map may
1331 	 * not have executed. Wait for them.
1332 	 */
1333 	rcu_barrier();
1334 	if (!htab_is_prealloc(htab))
1335 		delete_all_elements(htab);
1336 	else
1337 		prealloc_destroy(htab);
1338 
1339 	free_percpu(htab->extra_elems);
1340 	bpf_map_area_free(htab->buckets);
1341 	kfree(htab);
1342 }
1343 
1344 static void htab_map_seq_show_elem(struct bpf_map *map, void *key,
1345 				   struct seq_file *m)
1346 {
1347 	void *value;
1348 
1349 	rcu_read_lock();
1350 
1351 	value = htab_map_lookup_elem(map, key);
1352 	if (!value) {
1353 		rcu_read_unlock();
1354 		return;
1355 	}
1356 
1357 	btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1358 	seq_puts(m, ": ");
1359 	btf_type_seq_show(map->btf, map->btf_value_type_id, value, m);
1360 	seq_puts(m, "\n");
1361 
1362 	rcu_read_unlock();
1363 }
1364 
1365 static int
1366 __htab_map_lookup_and_delete_batch(struct bpf_map *map,
1367 				   const union bpf_attr *attr,
1368 				   union bpf_attr __user *uattr,
1369 				   bool do_delete, bool is_lru_map,
1370 				   bool is_percpu)
1371 {
1372 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1373 	u32 bucket_cnt, total, key_size, value_size, roundup_key_size;
1374 	void *keys = NULL, *values = NULL, *value, *dst_key, *dst_val;
1375 	void __user *uvalues = u64_to_user_ptr(attr->batch.values);
1376 	void __user *ukeys = u64_to_user_ptr(attr->batch.keys);
1377 	void *ubatch = u64_to_user_ptr(attr->batch.in_batch);
1378 	u32 batch, max_count, size, bucket_size;
1379 	struct htab_elem *node_to_free = NULL;
1380 	u64 elem_map_flags, map_flags;
1381 	struct hlist_nulls_head *head;
1382 	struct hlist_nulls_node *n;
1383 	unsigned long flags = 0;
1384 	bool locked = false;
1385 	struct htab_elem *l;
1386 	struct bucket *b;
1387 	int ret = 0;
1388 
1389 	elem_map_flags = attr->batch.elem_flags;
1390 	if ((elem_map_flags & ~BPF_F_LOCK) ||
1391 	    ((elem_map_flags & BPF_F_LOCK) && !map_value_has_spin_lock(map)))
1392 		return -EINVAL;
1393 
1394 	map_flags = attr->batch.flags;
1395 	if (map_flags)
1396 		return -EINVAL;
1397 
1398 	max_count = attr->batch.count;
1399 	if (!max_count)
1400 		return 0;
1401 
1402 	if (put_user(0, &uattr->batch.count))
1403 		return -EFAULT;
1404 
1405 	batch = 0;
1406 	if (ubatch && copy_from_user(&batch, ubatch, sizeof(batch)))
1407 		return -EFAULT;
1408 
1409 	if (batch >= htab->n_buckets)
1410 		return -ENOENT;
1411 
1412 	key_size = htab->map.key_size;
1413 	roundup_key_size = round_up(htab->map.key_size, 8);
1414 	value_size = htab->map.value_size;
1415 	size = round_up(value_size, 8);
1416 	if (is_percpu)
1417 		value_size = size * num_possible_cpus();
1418 	total = 0;
1419 	/* while experimenting with hash tables with sizes ranging from 10 to
1420 	 * 1000, it was observed that a bucket can have upto 5 entries.
1421 	 */
1422 	bucket_size = 5;
1423 
1424 alloc:
1425 	/* We cannot do copy_from_user or copy_to_user inside
1426 	 * the rcu_read_lock. Allocate enough space here.
1427 	 */
1428 	keys = kvmalloc(key_size * bucket_size, GFP_USER | __GFP_NOWARN);
1429 	values = kvmalloc(value_size * bucket_size, GFP_USER | __GFP_NOWARN);
1430 	if (!keys || !values) {
1431 		ret = -ENOMEM;
1432 		goto after_loop;
1433 	}
1434 
1435 again:
1436 	bpf_disable_instrumentation();
1437 	rcu_read_lock();
1438 again_nocopy:
1439 	dst_key = keys;
1440 	dst_val = values;
1441 	b = &htab->buckets[batch];
1442 	head = &b->head;
1443 	/* do not grab the lock unless need it (bucket_cnt > 0). */
1444 	if (locked)
1445 		flags = htab_lock_bucket(htab, b);
1446 
1447 	bucket_cnt = 0;
1448 	hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
1449 		bucket_cnt++;
1450 
1451 	if (bucket_cnt && !locked) {
1452 		locked = true;
1453 		goto again_nocopy;
1454 	}
1455 
1456 	if (bucket_cnt > (max_count - total)) {
1457 		if (total == 0)
1458 			ret = -ENOSPC;
1459 		/* Note that since bucket_cnt > 0 here, it is implicit
1460 		 * that the locked was grabbed, so release it.
1461 		 */
1462 		htab_unlock_bucket(htab, b, flags);
1463 		rcu_read_unlock();
1464 		bpf_enable_instrumentation();
1465 		goto after_loop;
1466 	}
1467 
1468 	if (bucket_cnt > bucket_size) {
1469 		bucket_size = bucket_cnt;
1470 		/* Note that since bucket_cnt > 0 here, it is implicit
1471 		 * that the locked was grabbed, so release it.
1472 		 */
1473 		htab_unlock_bucket(htab, b, flags);
1474 		rcu_read_unlock();
1475 		bpf_enable_instrumentation();
1476 		kvfree(keys);
1477 		kvfree(values);
1478 		goto alloc;
1479 	}
1480 
1481 	/* Next block is only safe to run if you have grabbed the lock */
1482 	if (!locked)
1483 		goto next_batch;
1484 
1485 	hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1486 		memcpy(dst_key, l->key, key_size);
1487 
1488 		if (is_percpu) {
1489 			int off = 0, cpu;
1490 			void __percpu *pptr;
1491 
1492 			pptr = htab_elem_get_ptr(l, map->key_size);
1493 			for_each_possible_cpu(cpu) {
1494 				bpf_long_memcpy(dst_val + off,
1495 						per_cpu_ptr(pptr, cpu), size);
1496 				off += size;
1497 			}
1498 		} else {
1499 			value = l->key + roundup_key_size;
1500 			if (elem_map_flags & BPF_F_LOCK)
1501 				copy_map_value_locked(map, dst_val, value,
1502 						      true);
1503 			else
1504 				copy_map_value(map, dst_val, value);
1505 			check_and_init_map_lock(map, dst_val);
1506 		}
1507 		if (do_delete) {
1508 			hlist_nulls_del_rcu(&l->hash_node);
1509 
1510 			/* bpf_lru_push_free() will acquire lru_lock, which
1511 			 * may cause deadlock. See comments in function
1512 			 * prealloc_lru_pop(). Let us do bpf_lru_push_free()
1513 			 * after releasing the bucket lock.
1514 			 */
1515 			if (is_lru_map) {
1516 				l->batch_flink = node_to_free;
1517 				node_to_free = l;
1518 			} else {
1519 				free_htab_elem(htab, l);
1520 			}
1521 		}
1522 		dst_key += key_size;
1523 		dst_val += value_size;
1524 	}
1525 
1526 	htab_unlock_bucket(htab, b, flags);
1527 	locked = false;
1528 
1529 	while (node_to_free) {
1530 		l = node_to_free;
1531 		node_to_free = node_to_free->batch_flink;
1532 		bpf_lru_push_free(&htab->lru, &l->lru_node);
1533 	}
1534 
1535 next_batch:
1536 	/* If we are not copying data, we can go to next bucket and avoid
1537 	 * unlocking the rcu.
1538 	 */
1539 	if (!bucket_cnt && (batch + 1 < htab->n_buckets)) {
1540 		batch++;
1541 		goto again_nocopy;
1542 	}
1543 
1544 	rcu_read_unlock();
1545 	bpf_enable_instrumentation();
1546 	if (bucket_cnt && (copy_to_user(ukeys + total * key_size, keys,
1547 	    key_size * bucket_cnt) ||
1548 	    copy_to_user(uvalues + total * value_size, values,
1549 	    value_size * bucket_cnt))) {
1550 		ret = -EFAULT;
1551 		goto after_loop;
1552 	}
1553 
1554 	total += bucket_cnt;
1555 	batch++;
1556 	if (batch >= htab->n_buckets) {
1557 		ret = -ENOENT;
1558 		goto after_loop;
1559 	}
1560 	goto again;
1561 
1562 after_loop:
1563 	if (ret == -EFAULT)
1564 		goto out;
1565 
1566 	/* copy # of entries and next batch */
1567 	ubatch = u64_to_user_ptr(attr->batch.out_batch);
1568 	if (copy_to_user(ubatch, &batch, sizeof(batch)) ||
1569 	    put_user(total, &uattr->batch.count))
1570 		ret = -EFAULT;
1571 
1572 out:
1573 	kvfree(keys);
1574 	kvfree(values);
1575 	return ret;
1576 }
1577 
1578 static int
1579 htab_percpu_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1580 			     union bpf_attr __user *uattr)
1581 {
1582 	return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1583 						  false, true);
1584 }
1585 
1586 static int
1587 htab_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1588 					const union bpf_attr *attr,
1589 					union bpf_attr __user *uattr)
1590 {
1591 	return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1592 						  false, true);
1593 }
1594 
1595 static int
1596 htab_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1597 		      union bpf_attr __user *uattr)
1598 {
1599 	return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1600 						  false, false);
1601 }
1602 
1603 static int
1604 htab_map_lookup_and_delete_batch(struct bpf_map *map,
1605 				 const union bpf_attr *attr,
1606 				 union bpf_attr __user *uattr)
1607 {
1608 	return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1609 						  false, false);
1610 }
1611 
1612 static int
1613 htab_lru_percpu_map_lookup_batch(struct bpf_map *map,
1614 				 const union bpf_attr *attr,
1615 				 union bpf_attr __user *uattr)
1616 {
1617 	return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1618 						  true, true);
1619 }
1620 
1621 static int
1622 htab_lru_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1623 					    const union bpf_attr *attr,
1624 					    union bpf_attr __user *uattr)
1625 {
1626 	return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1627 						  true, true);
1628 }
1629 
1630 static int
1631 htab_lru_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1632 			  union bpf_attr __user *uattr)
1633 {
1634 	return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1635 						  true, false);
1636 }
1637 
1638 static int
1639 htab_lru_map_lookup_and_delete_batch(struct bpf_map *map,
1640 				     const union bpf_attr *attr,
1641 				     union bpf_attr __user *uattr)
1642 {
1643 	return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1644 						  true, false);
1645 }
1646 
1647 struct bpf_iter_seq_hash_map_info {
1648 	struct bpf_map *map;
1649 	struct bpf_htab *htab;
1650 	void *percpu_value_buf; // non-zero means percpu hash
1651 	u32 bucket_id;
1652 	u32 skip_elems;
1653 };
1654 
1655 static struct htab_elem *
1656 bpf_hash_map_seq_find_next(struct bpf_iter_seq_hash_map_info *info,
1657 			   struct htab_elem *prev_elem)
1658 {
1659 	const struct bpf_htab *htab = info->htab;
1660 	u32 skip_elems = info->skip_elems;
1661 	u32 bucket_id = info->bucket_id;
1662 	struct hlist_nulls_head *head;
1663 	struct hlist_nulls_node *n;
1664 	struct htab_elem *elem;
1665 	struct bucket *b;
1666 	u32 i, count;
1667 
1668 	if (bucket_id >= htab->n_buckets)
1669 		return NULL;
1670 
1671 	/* try to find next elem in the same bucket */
1672 	if (prev_elem) {
1673 		/* no update/deletion on this bucket, prev_elem should be still valid
1674 		 * and we won't skip elements.
1675 		 */
1676 		n = rcu_dereference_raw(hlist_nulls_next_rcu(&prev_elem->hash_node));
1677 		elem = hlist_nulls_entry_safe(n, struct htab_elem, hash_node);
1678 		if (elem)
1679 			return elem;
1680 
1681 		/* not found, unlock and go to the next bucket */
1682 		b = &htab->buckets[bucket_id++];
1683 		rcu_read_unlock();
1684 		skip_elems = 0;
1685 	}
1686 
1687 	for (i = bucket_id; i < htab->n_buckets; i++) {
1688 		b = &htab->buckets[i];
1689 		rcu_read_lock();
1690 
1691 		count = 0;
1692 		head = &b->head;
1693 		hlist_nulls_for_each_entry_rcu(elem, n, head, hash_node) {
1694 			if (count >= skip_elems) {
1695 				info->bucket_id = i;
1696 				info->skip_elems = count;
1697 				return elem;
1698 			}
1699 			count++;
1700 		}
1701 
1702 		rcu_read_unlock();
1703 		skip_elems = 0;
1704 	}
1705 
1706 	info->bucket_id = i;
1707 	info->skip_elems = 0;
1708 	return NULL;
1709 }
1710 
1711 static void *bpf_hash_map_seq_start(struct seq_file *seq, loff_t *pos)
1712 {
1713 	struct bpf_iter_seq_hash_map_info *info = seq->private;
1714 	struct htab_elem *elem;
1715 
1716 	elem = bpf_hash_map_seq_find_next(info, NULL);
1717 	if (!elem)
1718 		return NULL;
1719 
1720 	if (*pos == 0)
1721 		++*pos;
1722 	return elem;
1723 }
1724 
1725 static void *bpf_hash_map_seq_next(struct seq_file *seq, void *v, loff_t *pos)
1726 {
1727 	struct bpf_iter_seq_hash_map_info *info = seq->private;
1728 
1729 	++*pos;
1730 	++info->skip_elems;
1731 	return bpf_hash_map_seq_find_next(info, v);
1732 }
1733 
1734 static int __bpf_hash_map_seq_show(struct seq_file *seq, struct htab_elem *elem)
1735 {
1736 	struct bpf_iter_seq_hash_map_info *info = seq->private;
1737 	u32 roundup_key_size, roundup_value_size;
1738 	struct bpf_iter__bpf_map_elem ctx = {};
1739 	struct bpf_map *map = info->map;
1740 	struct bpf_iter_meta meta;
1741 	int ret = 0, off = 0, cpu;
1742 	struct bpf_prog *prog;
1743 	void __percpu *pptr;
1744 
1745 	meta.seq = seq;
1746 	prog = bpf_iter_get_info(&meta, elem == NULL);
1747 	if (prog) {
1748 		ctx.meta = &meta;
1749 		ctx.map = info->map;
1750 		if (elem) {
1751 			roundup_key_size = round_up(map->key_size, 8);
1752 			ctx.key = elem->key;
1753 			if (!info->percpu_value_buf) {
1754 				ctx.value = elem->key + roundup_key_size;
1755 			} else {
1756 				roundup_value_size = round_up(map->value_size, 8);
1757 				pptr = htab_elem_get_ptr(elem, map->key_size);
1758 				for_each_possible_cpu(cpu) {
1759 					bpf_long_memcpy(info->percpu_value_buf + off,
1760 							per_cpu_ptr(pptr, cpu),
1761 							roundup_value_size);
1762 					off += roundup_value_size;
1763 				}
1764 				ctx.value = info->percpu_value_buf;
1765 			}
1766 		}
1767 		ret = bpf_iter_run_prog(prog, &ctx);
1768 	}
1769 
1770 	return ret;
1771 }
1772 
1773 static int bpf_hash_map_seq_show(struct seq_file *seq, void *v)
1774 {
1775 	return __bpf_hash_map_seq_show(seq, v);
1776 }
1777 
1778 static void bpf_hash_map_seq_stop(struct seq_file *seq, void *v)
1779 {
1780 	if (!v)
1781 		(void)__bpf_hash_map_seq_show(seq, NULL);
1782 	else
1783 		rcu_read_unlock();
1784 }
1785 
1786 static int bpf_iter_init_hash_map(void *priv_data,
1787 				  struct bpf_iter_aux_info *aux)
1788 {
1789 	struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
1790 	struct bpf_map *map = aux->map;
1791 	void *value_buf;
1792 	u32 buf_size;
1793 
1794 	if (map->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
1795 	    map->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH) {
1796 		buf_size = round_up(map->value_size, 8) * num_possible_cpus();
1797 		value_buf = kmalloc(buf_size, GFP_USER | __GFP_NOWARN);
1798 		if (!value_buf)
1799 			return -ENOMEM;
1800 
1801 		seq_info->percpu_value_buf = value_buf;
1802 	}
1803 
1804 	seq_info->map = map;
1805 	seq_info->htab = container_of(map, struct bpf_htab, map);
1806 	return 0;
1807 }
1808 
1809 static void bpf_iter_fini_hash_map(void *priv_data)
1810 {
1811 	struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
1812 
1813 	kfree(seq_info->percpu_value_buf);
1814 }
1815 
1816 static const struct seq_operations bpf_hash_map_seq_ops = {
1817 	.start	= bpf_hash_map_seq_start,
1818 	.next	= bpf_hash_map_seq_next,
1819 	.stop	= bpf_hash_map_seq_stop,
1820 	.show	= bpf_hash_map_seq_show,
1821 };
1822 
1823 static const struct bpf_iter_seq_info iter_seq_info = {
1824 	.seq_ops		= &bpf_hash_map_seq_ops,
1825 	.init_seq_private	= bpf_iter_init_hash_map,
1826 	.fini_seq_private	= bpf_iter_fini_hash_map,
1827 	.seq_priv_size		= sizeof(struct bpf_iter_seq_hash_map_info),
1828 };
1829 
1830 static int htab_map_btf_id;
1831 const struct bpf_map_ops htab_map_ops = {
1832 	.map_meta_equal = bpf_map_meta_equal,
1833 	.map_alloc_check = htab_map_alloc_check,
1834 	.map_alloc = htab_map_alloc,
1835 	.map_free = htab_map_free,
1836 	.map_get_next_key = htab_map_get_next_key,
1837 	.map_lookup_elem = htab_map_lookup_elem,
1838 	.map_update_elem = htab_map_update_elem,
1839 	.map_delete_elem = htab_map_delete_elem,
1840 	.map_gen_lookup = htab_map_gen_lookup,
1841 	.map_seq_show_elem = htab_map_seq_show_elem,
1842 	BATCH_OPS(htab),
1843 	.map_btf_name = "bpf_htab",
1844 	.map_btf_id = &htab_map_btf_id,
1845 	.iter_seq_info = &iter_seq_info,
1846 };
1847 
1848 static int htab_lru_map_btf_id;
1849 const struct bpf_map_ops htab_lru_map_ops = {
1850 	.map_meta_equal = bpf_map_meta_equal,
1851 	.map_alloc_check = htab_map_alloc_check,
1852 	.map_alloc = htab_map_alloc,
1853 	.map_free = htab_map_free,
1854 	.map_get_next_key = htab_map_get_next_key,
1855 	.map_lookup_elem = htab_lru_map_lookup_elem,
1856 	.map_lookup_elem_sys_only = htab_lru_map_lookup_elem_sys,
1857 	.map_update_elem = htab_lru_map_update_elem,
1858 	.map_delete_elem = htab_lru_map_delete_elem,
1859 	.map_gen_lookup = htab_lru_map_gen_lookup,
1860 	.map_seq_show_elem = htab_map_seq_show_elem,
1861 	BATCH_OPS(htab_lru),
1862 	.map_btf_name = "bpf_htab",
1863 	.map_btf_id = &htab_lru_map_btf_id,
1864 	.iter_seq_info = &iter_seq_info,
1865 };
1866 
1867 /* Called from eBPF program */
1868 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1869 {
1870 	struct htab_elem *l = __htab_map_lookup_elem(map, key);
1871 
1872 	if (l)
1873 		return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1874 	else
1875 		return NULL;
1876 }
1877 
1878 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
1879 {
1880 	struct htab_elem *l = __htab_map_lookup_elem(map, key);
1881 
1882 	if (l) {
1883 		bpf_lru_node_set_ref(&l->lru_node);
1884 		return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
1885 	}
1886 
1887 	return NULL;
1888 }
1889 
1890 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
1891 {
1892 	struct htab_elem *l;
1893 	void __percpu *pptr;
1894 	int ret = -ENOENT;
1895 	int cpu, off = 0;
1896 	u32 size;
1897 
1898 	/* per_cpu areas are zero-filled and bpf programs can only
1899 	 * access 'value_size' of them, so copying rounded areas
1900 	 * will not leak any kernel data
1901 	 */
1902 	size = round_up(map->value_size, 8);
1903 	rcu_read_lock();
1904 	l = __htab_map_lookup_elem(map, key);
1905 	if (!l)
1906 		goto out;
1907 	/* We do not mark LRU map element here in order to not mess up
1908 	 * eviction heuristics when user space does a map walk.
1909 	 */
1910 	pptr = htab_elem_get_ptr(l, map->key_size);
1911 	for_each_possible_cpu(cpu) {
1912 		bpf_long_memcpy(value + off,
1913 				per_cpu_ptr(pptr, cpu), size);
1914 		off += size;
1915 	}
1916 	ret = 0;
1917 out:
1918 	rcu_read_unlock();
1919 	return ret;
1920 }
1921 
1922 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
1923 			   u64 map_flags)
1924 {
1925 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1926 	int ret;
1927 
1928 	rcu_read_lock();
1929 	if (htab_is_lru(htab))
1930 		ret = __htab_lru_percpu_map_update_elem(map, key, value,
1931 							map_flags, true);
1932 	else
1933 		ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
1934 						    true);
1935 	rcu_read_unlock();
1936 
1937 	return ret;
1938 }
1939 
1940 static void htab_percpu_map_seq_show_elem(struct bpf_map *map, void *key,
1941 					  struct seq_file *m)
1942 {
1943 	struct htab_elem *l;
1944 	void __percpu *pptr;
1945 	int cpu;
1946 
1947 	rcu_read_lock();
1948 
1949 	l = __htab_map_lookup_elem(map, key);
1950 	if (!l) {
1951 		rcu_read_unlock();
1952 		return;
1953 	}
1954 
1955 	btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1956 	seq_puts(m, ": {\n");
1957 	pptr = htab_elem_get_ptr(l, map->key_size);
1958 	for_each_possible_cpu(cpu) {
1959 		seq_printf(m, "\tcpu%d: ", cpu);
1960 		btf_type_seq_show(map->btf, map->btf_value_type_id,
1961 				  per_cpu_ptr(pptr, cpu), m);
1962 		seq_puts(m, "\n");
1963 	}
1964 	seq_puts(m, "}\n");
1965 
1966 	rcu_read_unlock();
1967 }
1968 
1969 static int htab_percpu_map_btf_id;
1970 const struct bpf_map_ops htab_percpu_map_ops = {
1971 	.map_meta_equal = bpf_map_meta_equal,
1972 	.map_alloc_check = htab_map_alloc_check,
1973 	.map_alloc = htab_map_alloc,
1974 	.map_free = htab_map_free,
1975 	.map_get_next_key = htab_map_get_next_key,
1976 	.map_lookup_elem = htab_percpu_map_lookup_elem,
1977 	.map_update_elem = htab_percpu_map_update_elem,
1978 	.map_delete_elem = htab_map_delete_elem,
1979 	.map_seq_show_elem = htab_percpu_map_seq_show_elem,
1980 	BATCH_OPS(htab_percpu),
1981 	.map_btf_name = "bpf_htab",
1982 	.map_btf_id = &htab_percpu_map_btf_id,
1983 	.iter_seq_info = &iter_seq_info,
1984 };
1985 
1986 static int htab_lru_percpu_map_btf_id;
1987 const struct bpf_map_ops htab_lru_percpu_map_ops = {
1988 	.map_meta_equal = bpf_map_meta_equal,
1989 	.map_alloc_check = htab_map_alloc_check,
1990 	.map_alloc = htab_map_alloc,
1991 	.map_free = htab_map_free,
1992 	.map_get_next_key = htab_map_get_next_key,
1993 	.map_lookup_elem = htab_lru_percpu_map_lookup_elem,
1994 	.map_update_elem = htab_lru_percpu_map_update_elem,
1995 	.map_delete_elem = htab_lru_map_delete_elem,
1996 	.map_seq_show_elem = htab_percpu_map_seq_show_elem,
1997 	BATCH_OPS(htab_lru_percpu),
1998 	.map_btf_name = "bpf_htab",
1999 	.map_btf_id = &htab_lru_percpu_map_btf_id,
2000 	.iter_seq_info = &iter_seq_info,
2001 };
2002 
2003 static int fd_htab_map_alloc_check(union bpf_attr *attr)
2004 {
2005 	if (attr->value_size != sizeof(u32))
2006 		return -EINVAL;
2007 	return htab_map_alloc_check(attr);
2008 }
2009 
2010 static void fd_htab_map_free(struct bpf_map *map)
2011 {
2012 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2013 	struct hlist_nulls_node *n;
2014 	struct hlist_nulls_head *head;
2015 	struct htab_elem *l;
2016 	int i;
2017 
2018 	for (i = 0; i < htab->n_buckets; i++) {
2019 		head = select_bucket(htab, i);
2020 
2021 		hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
2022 			void *ptr = fd_htab_map_get_ptr(map, l);
2023 
2024 			map->ops->map_fd_put_ptr(ptr);
2025 		}
2026 	}
2027 
2028 	htab_map_free(map);
2029 }
2030 
2031 /* only called from syscall */
2032 int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
2033 {
2034 	void **ptr;
2035 	int ret = 0;
2036 
2037 	if (!map->ops->map_fd_sys_lookup_elem)
2038 		return -ENOTSUPP;
2039 
2040 	rcu_read_lock();
2041 	ptr = htab_map_lookup_elem(map, key);
2042 	if (ptr)
2043 		*value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
2044 	else
2045 		ret = -ENOENT;
2046 	rcu_read_unlock();
2047 
2048 	return ret;
2049 }
2050 
2051 /* only called from syscall */
2052 int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
2053 				void *key, void *value, u64 map_flags)
2054 {
2055 	void *ptr;
2056 	int ret;
2057 	u32 ufd = *(u32 *)value;
2058 
2059 	ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
2060 	if (IS_ERR(ptr))
2061 		return PTR_ERR(ptr);
2062 
2063 	ret = htab_map_update_elem(map, key, &ptr, map_flags);
2064 	if (ret)
2065 		map->ops->map_fd_put_ptr(ptr);
2066 
2067 	return ret;
2068 }
2069 
2070 static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
2071 {
2072 	struct bpf_map *map, *inner_map_meta;
2073 
2074 	inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
2075 	if (IS_ERR(inner_map_meta))
2076 		return inner_map_meta;
2077 
2078 	map = htab_map_alloc(attr);
2079 	if (IS_ERR(map)) {
2080 		bpf_map_meta_free(inner_map_meta);
2081 		return map;
2082 	}
2083 
2084 	map->inner_map_meta = inner_map_meta;
2085 
2086 	return map;
2087 }
2088 
2089 static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
2090 {
2091 	struct bpf_map **inner_map  = htab_map_lookup_elem(map, key);
2092 
2093 	if (!inner_map)
2094 		return NULL;
2095 
2096 	return READ_ONCE(*inner_map);
2097 }
2098 
2099 static int htab_of_map_gen_lookup(struct bpf_map *map,
2100 				  struct bpf_insn *insn_buf)
2101 {
2102 	struct bpf_insn *insn = insn_buf;
2103 	const int ret = BPF_REG_0;
2104 
2105 	BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
2106 		     (void *(*)(struct bpf_map *map, void *key))NULL));
2107 	*insn++ = BPF_EMIT_CALL(BPF_CAST_CALL(__htab_map_lookup_elem));
2108 	*insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
2109 	*insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
2110 				offsetof(struct htab_elem, key) +
2111 				round_up(map->key_size, 8));
2112 	*insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
2113 
2114 	return insn - insn_buf;
2115 }
2116 
2117 static void htab_of_map_free(struct bpf_map *map)
2118 {
2119 	bpf_map_meta_free(map->inner_map_meta);
2120 	fd_htab_map_free(map);
2121 }
2122 
2123 static int htab_of_maps_map_btf_id;
2124 const struct bpf_map_ops htab_of_maps_map_ops = {
2125 	.map_alloc_check = fd_htab_map_alloc_check,
2126 	.map_alloc = htab_of_map_alloc,
2127 	.map_free = htab_of_map_free,
2128 	.map_get_next_key = htab_map_get_next_key,
2129 	.map_lookup_elem = htab_of_map_lookup_elem,
2130 	.map_delete_elem = htab_map_delete_elem,
2131 	.map_fd_get_ptr = bpf_map_fd_get_ptr,
2132 	.map_fd_put_ptr = bpf_map_fd_put_ptr,
2133 	.map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
2134 	.map_gen_lookup = htab_of_map_gen_lookup,
2135 	.map_check_btf = map_check_no_btf,
2136 	.map_btf_name = "bpf_htab",
2137 	.map_btf_id = &htab_of_maps_map_btf_id,
2138 };
2139