xref: /openbmc/linux/kernel/bpf/hashtab.c (revision 8f67e87e)
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 <linux/btf_ids.h>
14 #include "percpu_freelist.h"
15 #include "bpf_lru_list.h"
16 #include "map_in_map.h"
17 #include <linux/bpf_mem_alloc.h>
18 
19 #define HTAB_CREATE_FLAG_MASK						\
20 	(BPF_F_NO_PREALLOC | BPF_F_NO_COMMON_LRU | BPF_F_NUMA_NODE |	\
21 	 BPF_F_ACCESS_MASK | BPF_F_ZERO_SEED)
22 
23 #define BATCH_OPS(_name)			\
24 	.map_lookup_batch =			\
25 	_name##_map_lookup_batch,		\
26 	.map_lookup_and_delete_batch =		\
27 	_name##_map_lookup_and_delete_batch,	\
28 	.map_update_batch =			\
29 	generic_map_update_batch,		\
30 	.map_delete_batch =			\
31 	generic_map_delete_batch
32 
33 /*
34  * The bucket lock has two protection scopes:
35  *
36  * 1) Serializing concurrent operations from BPF programs on different
37  *    CPUs
38  *
39  * 2) Serializing concurrent operations from BPF programs and sys_bpf()
40  *
41  * BPF programs can execute in any context including perf, kprobes and
42  * tracing. As there are almost no limits where perf, kprobes and tracing
43  * can be invoked from the lock operations need to be protected against
44  * deadlocks. Deadlocks can be caused by recursion and by an invocation in
45  * the lock held section when functions which acquire this lock are invoked
46  * from sys_bpf(). BPF recursion is prevented by incrementing the per CPU
47  * variable bpf_prog_active, which prevents BPF programs attached to perf
48  * events, kprobes and tracing to be invoked before the prior invocation
49  * from one of these contexts completed. sys_bpf() uses the same mechanism
50  * by pinning the task to the current CPU and incrementing the recursion
51  * protection across the map operation.
52  *
53  * This has subtle implications on PREEMPT_RT. PREEMPT_RT forbids certain
54  * operations like memory allocations (even with GFP_ATOMIC) from atomic
55  * contexts. This is required because even with GFP_ATOMIC the memory
56  * allocator calls into code paths which acquire locks with long held lock
57  * sections. To ensure the deterministic behaviour these locks are regular
58  * spinlocks, which are converted to 'sleepable' spinlocks on RT. The only
59  * true atomic contexts on an RT kernel are the low level hardware
60  * handling, scheduling, low level interrupt handling, NMIs etc. None of
61  * these contexts should ever do memory allocations.
62  *
63  * As regular device interrupt handlers and soft interrupts are forced into
64  * thread context, the existing code which does
65  *   spin_lock*(); alloc(GFP_ATOMIC); spin_unlock*();
66  * just works.
67  *
68  * In theory the BPF locks could be converted to regular spinlocks as well,
69  * but the bucket locks and percpu_freelist locks can be taken from
70  * arbitrary contexts (perf, kprobes, tracepoints) which are required to be
71  * atomic contexts even on RT. Before the introduction of bpf_mem_alloc,
72  * it is only safe to use raw spinlock for preallocated hash map on a RT kernel,
73  * because there is no memory allocation within the lock held sections. However
74  * after hash map was fully converted to use bpf_mem_alloc, there will be
75  * non-synchronous memory allocation for non-preallocated hash map, so it is
76  * safe to always use raw spinlock for bucket lock.
77  */
78 struct bucket {
79 	struct hlist_nulls_head head;
80 	raw_spinlock_t raw_lock;
81 };
82 
83 #define HASHTAB_MAP_LOCK_COUNT 8
84 #define HASHTAB_MAP_LOCK_MASK (HASHTAB_MAP_LOCK_COUNT - 1)
85 
86 struct bpf_htab {
87 	struct bpf_map map;
88 	struct bpf_mem_alloc ma;
89 	struct bpf_mem_alloc pcpu_ma;
90 	struct bucket *buckets;
91 	void *elems;
92 	union {
93 		struct pcpu_freelist freelist;
94 		struct bpf_lru lru;
95 	};
96 	struct htab_elem *__percpu *extra_elems;
97 	/* number of elements in non-preallocated hashtable are kept
98 	 * in either pcount or count
99 	 */
100 	struct percpu_counter pcount;
101 	atomic_t count;
102 	bool use_percpu_counter;
103 	u32 n_buckets;	/* number of hash buckets */
104 	u32 elem_size;	/* size of each element in bytes */
105 	u32 hashrnd;
106 	struct lock_class_key lockdep_key;
107 	int __percpu *map_locked[HASHTAB_MAP_LOCK_COUNT];
108 };
109 
110 /* each htab element is struct htab_elem + key + value */
111 struct htab_elem {
112 	union {
113 		struct hlist_nulls_node hash_node;
114 		struct {
115 			void *padding;
116 			union {
117 				struct pcpu_freelist_node fnode;
118 				struct htab_elem *batch_flink;
119 			};
120 		};
121 	};
122 	union {
123 		/* pointer to per-cpu pointer */
124 		void *ptr_to_pptr;
125 		struct bpf_lru_node lru_node;
126 	};
127 	u32 hash;
128 	char key[] __aligned(8);
129 };
130 
131 static inline bool htab_is_prealloc(const struct bpf_htab *htab)
132 {
133 	return !(htab->map.map_flags & BPF_F_NO_PREALLOC);
134 }
135 
136 static void htab_init_buckets(struct bpf_htab *htab)
137 {
138 	unsigned int i;
139 
140 	for (i = 0; i < htab->n_buckets; i++) {
141 		INIT_HLIST_NULLS_HEAD(&htab->buckets[i].head, i);
142 		raw_spin_lock_init(&htab->buckets[i].raw_lock);
143 		lockdep_set_class(&htab->buckets[i].raw_lock,
144 					  &htab->lockdep_key);
145 		cond_resched();
146 	}
147 }
148 
149 static inline int htab_lock_bucket(const struct bpf_htab *htab,
150 				   struct bucket *b, u32 hash,
151 				   unsigned long *pflags)
152 {
153 	unsigned long flags;
154 
155 	hash = hash & min_t(u32, HASHTAB_MAP_LOCK_MASK, htab->n_buckets - 1);
156 
157 	preempt_disable();
158 	local_irq_save(flags);
159 	if (unlikely(__this_cpu_inc_return(*(htab->map_locked[hash])) != 1)) {
160 		__this_cpu_dec(*(htab->map_locked[hash]));
161 		local_irq_restore(flags);
162 		preempt_enable();
163 		return -EBUSY;
164 	}
165 
166 	raw_spin_lock(&b->raw_lock);
167 	*pflags = flags;
168 
169 	return 0;
170 }
171 
172 static inline void htab_unlock_bucket(const struct bpf_htab *htab,
173 				      struct bucket *b, u32 hash,
174 				      unsigned long flags)
175 {
176 	hash = hash & min_t(u32, HASHTAB_MAP_LOCK_MASK, htab->n_buckets - 1);
177 	raw_spin_unlock(&b->raw_lock);
178 	__this_cpu_dec(*(htab->map_locked[hash]));
179 	local_irq_restore(flags);
180 	preempt_enable();
181 }
182 
183 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node);
184 
185 static bool htab_is_lru(const struct bpf_htab *htab)
186 {
187 	return htab->map.map_type == BPF_MAP_TYPE_LRU_HASH ||
188 		htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
189 }
190 
191 static bool htab_is_percpu(const struct bpf_htab *htab)
192 {
193 	return htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH ||
194 		htab->map.map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH;
195 }
196 
197 static inline void htab_elem_set_ptr(struct htab_elem *l, u32 key_size,
198 				     void __percpu *pptr)
199 {
200 	*(void __percpu **)(l->key + key_size) = pptr;
201 }
202 
203 static inline void __percpu *htab_elem_get_ptr(struct htab_elem *l, u32 key_size)
204 {
205 	return *(void __percpu **)(l->key + key_size);
206 }
207 
208 static void *fd_htab_map_get_ptr(const struct bpf_map *map, struct htab_elem *l)
209 {
210 	return *(void **)(l->key + roundup(map->key_size, 8));
211 }
212 
213 static struct htab_elem *get_htab_elem(struct bpf_htab *htab, int i)
214 {
215 	return (struct htab_elem *) (htab->elems + i * (u64)htab->elem_size);
216 }
217 
218 static bool htab_has_extra_elems(struct bpf_htab *htab)
219 {
220 	return !htab_is_percpu(htab) && !htab_is_lru(htab);
221 }
222 
223 static void htab_free_prealloced_timers(struct bpf_htab *htab)
224 {
225 	u32 num_entries = htab->map.max_entries;
226 	int i;
227 
228 	if (!btf_record_has_field(htab->map.record, BPF_TIMER))
229 		return;
230 	if (htab_has_extra_elems(htab))
231 		num_entries += num_possible_cpus();
232 
233 	for (i = 0; i < num_entries; i++) {
234 		struct htab_elem *elem;
235 
236 		elem = get_htab_elem(htab, i);
237 		bpf_obj_free_timer(htab->map.record, elem->key + round_up(htab->map.key_size, 8));
238 		cond_resched();
239 	}
240 }
241 
242 static void htab_free_prealloced_fields(struct bpf_htab *htab)
243 {
244 	u32 num_entries = htab->map.max_entries;
245 	int i;
246 
247 	if (IS_ERR_OR_NULL(htab->map.record))
248 		return;
249 	if (htab_has_extra_elems(htab))
250 		num_entries += num_possible_cpus();
251 	for (i = 0; i < num_entries; i++) {
252 		struct htab_elem *elem;
253 
254 		elem = get_htab_elem(htab, i);
255 		if (htab_is_percpu(htab)) {
256 			void __percpu *pptr = htab_elem_get_ptr(elem, htab->map.key_size);
257 			int cpu;
258 
259 			for_each_possible_cpu(cpu) {
260 				bpf_obj_free_fields(htab->map.record, per_cpu_ptr(pptr, cpu));
261 				cond_resched();
262 			}
263 		} else {
264 			bpf_obj_free_fields(htab->map.record, elem->key + round_up(htab->map.key_size, 8));
265 			cond_resched();
266 		}
267 		cond_resched();
268 	}
269 }
270 
271 static void htab_free_elems(struct bpf_htab *htab)
272 {
273 	int i;
274 
275 	if (!htab_is_percpu(htab))
276 		goto free_elems;
277 
278 	for (i = 0; i < htab->map.max_entries; i++) {
279 		void __percpu *pptr;
280 
281 		pptr = htab_elem_get_ptr(get_htab_elem(htab, i),
282 					 htab->map.key_size);
283 		free_percpu(pptr);
284 		cond_resched();
285 	}
286 free_elems:
287 	bpf_map_area_free(htab->elems);
288 }
289 
290 /* The LRU list has a lock (lru_lock). Each htab bucket has a lock
291  * (bucket_lock). If both locks need to be acquired together, the lock
292  * order is always lru_lock -> bucket_lock and this only happens in
293  * bpf_lru_list.c logic. For example, certain code path of
294  * bpf_lru_pop_free(), which is called by function prealloc_lru_pop(),
295  * will acquire lru_lock first followed by acquiring bucket_lock.
296  *
297  * In hashtab.c, to avoid deadlock, lock acquisition of
298  * bucket_lock followed by lru_lock is not allowed. In such cases,
299  * bucket_lock needs to be released first before acquiring lru_lock.
300  */
301 static struct htab_elem *prealloc_lru_pop(struct bpf_htab *htab, void *key,
302 					  u32 hash)
303 {
304 	struct bpf_lru_node *node = bpf_lru_pop_free(&htab->lru, hash);
305 	struct htab_elem *l;
306 
307 	if (node) {
308 		bpf_map_inc_elem_count(&htab->map);
309 		l = container_of(node, struct htab_elem, lru_node);
310 		memcpy(l->key, key, htab->map.key_size);
311 		return l;
312 	}
313 
314 	return NULL;
315 }
316 
317 static int prealloc_init(struct bpf_htab *htab)
318 {
319 	u32 num_entries = htab->map.max_entries;
320 	int err = -ENOMEM, i;
321 
322 	if (htab_has_extra_elems(htab))
323 		num_entries += num_possible_cpus();
324 
325 	htab->elems = bpf_map_area_alloc((u64)htab->elem_size * num_entries,
326 					 htab->map.numa_node);
327 	if (!htab->elems)
328 		return -ENOMEM;
329 
330 	if (!htab_is_percpu(htab))
331 		goto skip_percpu_elems;
332 
333 	for (i = 0; i < num_entries; i++) {
334 		u32 size = round_up(htab->map.value_size, 8);
335 		void __percpu *pptr;
336 
337 		pptr = bpf_map_alloc_percpu(&htab->map, size, 8,
338 					    GFP_USER | __GFP_NOWARN);
339 		if (!pptr)
340 			goto free_elems;
341 		htab_elem_set_ptr(get_htab_elem(htab, i), htab->map.key_size,
342 				  pptr);
343 		cond_resched();
344 	}
345 
346 skip_percpu_elems:
347 	if (htab_is_lru(htab))
348 		err = bpf_lru_init(&htab->lru,
349 				   htab->map.map_flags & BPF_F_NO_COMMON_LRU,
350 				   offsetof(struct htab_elem, hash) -
351 				   offsetof(struct htab_elem, lru_node),
352 				   htab_lru_map_delete_node,
353 				   htab);
354 	else
355 		err = pcpu_freelist_init(&htab->freelist);
356 
357 	if (err)
358 		goto free_elems;
359 
360 	if (htab_is_lru(htab))
361 		bpf_lru_populate(&htab->lru, htab->elems,
362 				 offsetof(struct htab_elem, lru_node),
363 				 htab->elem_size, num_entries);
364 	else
365 		pcpu_freelist_populate(&htab->freelist,
366 				       htab->elems + offsetof(struct htab_elem, fnode),
367 				       htab->elem_size, num_entries);
368 
369 	return 0;
370 
371 free_elems:
372 	htab_free_elems(htab);
373 	return err;
374 }
375 
376 static void prealloc_destroy(struct bpf_htab *htab)
377 {
378 	htab_free_elems(htab);
379 
380 	if (htab_is_lru(htab))
381 		bpf_lru_destroy(&htab->lru);
382 	else
383 		pcpu_freelist_destroy(&htab->freelist);
384 }
385 
386 static int alloc_extra_elems(struct bpf_htab *htab)
387 {
388 	struct htab_elem *__percpu *pptr, *l_new;
389 	struct pcpu_freelist_node *l;
390 	int cpu;
391 
392 	pptr = bpf_map_alloc_percpu(&htab->map, sizeof(struct htab_elem *), 8,
393 				    GFP_USER | __GFP_NOWARN);
394 	if (!pptr)
395 		return -ENOMEM;
396 
397 	for_each_possible_cpu(cpu) {
398 		l = pcpu_freelist_pop(&htab->freelist);
399 		/* pop will succeed, since prealloc_init()
400 		 * preallocated extra num_possible_cpus elements
401 		 */
402 		l_new = container_of(l, struct htab_elem, fnode);
403 		*per_cpu_ptr(pptr, cpu) = l_new;
404 	}
405 	htab->extra_elems = pptr;
406 	return 0;
407 }
408 
409 /* Called from syscall */
410 static int htab_map_alloc_check(union bpf_attr *attr)
411 {
412 	bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
413 		       attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
414 	bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
415 		    attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
416 	/* percpu_lru means each cpu has its own LRU list.
417 	 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
418 	 * the map's value itself is percpu.  percpu_lru has
419 	 * nothing to do with the map's value.
420 	 */
421 	bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
422 	bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
423 	bool zero_seed = (attr->map_flags & BPF_F_ZERO_SEED);
424 	int numa_node = bpf_map_attr_numa_node(attr);
425 
426 	BUILD_BUG_ON(offsetof(struct htab_elem, fnode.next) !=
427 		     offsetof(struct htab_elem, hash_node.pprev));
428 
429 	if (zero_seed && !capable(CAP_SYS_ADMIN))
430 		/* Guard against local DoS, and discourage production use. */
431 		return -EPERM;
432 
433 	if (attr->map_flags & ~HTAB_CREATE_FLAG_MASK ||
434 	    !bpf_map_flags_access_ok(attr->map_flags))
435 		return -EINVAL;
436 
437 	if (!lru && percpu_lru)
438 		return -EINVAL;
439 
440 	if (lru && !prealloc)
441 		return -ENOTSUPP;
442 
443 	if (numa_node != NUMA_NO_NODE && (percpu || percpu_lru))
444 		return -EINVAL;
445 
446 	/* check sanity of attributes.
447 	 * value_size == 0 may be allowed in the future to use map as a set
448 	 */
449 	if (attr->max_entries == 0 || attr->key_size == 0 ||
450 	    attr->value_size == 0)
451 		return -EINVAL;
452 
453 	if ((u64)attr->key_size + attr->value_size >= KMALLOC_MAX_SIZE -
454 	   sizeof(struct htab_elem))
455 		/* if key_size + value_size is bigger, the user space won't be
456 		 * able to access the elements via bpf syscall. This check
457 		 * also makes sure that the elem_size doesn't overflow and it's
458 		 * kmalloc-able later in htab_map_update_elem()
459 		 */
460 		return -E2BIG;
461 
462 	return 0;
463 }
464 
465 static struct bpf_map *htab_map_alloc(union bpf_attr *attr)
466 {
467 	bool percpu = (attr->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
468 		       attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
469 	bool lru = (attr->map_type == BPF_MAP_TYPE_LRU_HASH ||
470 		    attr->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH);
471 	/* percpu_lru means each cpu has its own LRU list.
472 	 * it is different from BPF_MAP_TYPE_PERCPU_HASH where
473 	 * the map's value itself is percpu.  percpu_lru has
474 	 * nothing to do with the map's value.
475 	 */
476 	bool percpu_lru = (attr->map_flags & BPF_F_NO_COMMON_LRU);
477 	bool prealloc = !(attr->map_flags & BPF_F_NO_PREALLOC);
478 	struct bpf_htab *htab;
479 	int err, i;
480 
481 	htab = bpf_map_area_alloc(sizeof(*htab), NUMA_NO_NODE);
482 	if (!htab)
483 		return ERR_PTR(-ENOMEM);
484 
485 	lockdep_register_key(&htab->lockdep_key);
486 
487 	bpf_map_init_from_attr(&htab->map, attr);
488 
489 	if (percpu_lru) {
490 		/* ensure each CPU's lru list has >=1 elements.
491 		 * since we are at it, make each lru list has the same
492 		 * number of elements.
493 		 */
494 		htab->map.max_entries = roundup(attr->max_entries,
495 						num_possible_cpus());
496 		if (htab->map.max_entries < attr->max_entries)
497 			htab->map.max_entries = rounddown(attr->max_entries,
498 							  num_possible_cpus());
499 	}
500 
501 	/* hash table size must be power of 2 */
502 	htab->n_buckets = roundup_pow_of_two(htab->map.max_entries);
503 
504 	htab->elem_size = sizeof(struct htab_elem) +
505 			  round_up(htab->map.key_size, 8);
506 	if (percpu)
507 		htab->elem_size += sizeof(void *);
508 	else
509 		htab->elem_size += round_up(htab->map.value_size, 8);
510 
511 	err = -E2BIG;
512 	/* prevent zero size kmalloc and check for u32 overflow */
513 	if (htab->n_buckets == 0 ||
514 	    htab->n_buckets > U32_MAX / sizeof(struct bucket))
515 		goto free_htab;
516 
517 	err = bpf_map_init_elem_count(&htab->map);
518 	if (err)
519 		goto free_htab;
520 
521 	err = -ENOMEM;
522 	htab->buckets = bpf_map_area_alloc(htab->n_buckets *
523 					   sizeof(struct bucket),
524 					   htab->map.numa_node);
525 	if (!htab->buckets)
526 		goto free_elem_count;
527 
528 	for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++) {
529 		htab->map_locked[i] = bpf_map_alloc_percpu(&htab->map,
530 							   sizeof(int),
531 							   sizeof(int),
532 							   GFP_USER);
533 		if (!htab->map_locked[i])
534 			goto free_map_locked;
535 	}
536 
537 	if (htab->map.map_flags & BPF_F_ZERO_SEED)
538 		htab->hashrnd = 0;
539 	else
540 		htab->hashrnd = get_random_u32();
541 
542 	htab_init_buckets(htab);
543 
544 /* compute_batch_value() computes batch value as num_online_cpus() * 2
545  * and __percpu_counter_compare() needs
546  * htab->max_entries - cur_number_of_elems to be more than batch * num_online_cpus()
547  * for percpu_counter to be faster than atomic_t. In practice the average bpf
548  * hash map size is 10k, which means that a system with 64 cpus will fill
549  * hashmap to 20% of 10k before percpu_counter becomes ineffective. Therefore
550  * define our own batch count as 32 then 10k hash map can be filled up to 80%:
551  * 10k - 8k > 32 _batch_ * 64 _cpus_
552  * and __percpu_counter_compare() will still be fast. At that point hash map
553  * collisions will dominate its performance anyway. Assume that hash map filled
554  * to 50+% isn't going to be O(1) and use the following formula to choose
555  * between percpu_counter and atomic_t.
556  */
557 #define PERCPU_COUNTER_BATCH 32
558 	if (attr->max_entries / 2 > num_online_cpus() * PERCPU_COUNTER_BATCH)
559 		htab->use_percpu_counter = true;
560 
561 	if (htab->use_percpu_counter) {
562 		err = percpu_counter_init(&htab->pcount, 0, GFP_KERNEL);
563 		if (err)
564 			goto free_map_locked;
565 	}
566 
567 	if (prealloc) {
568 		err = prealloc_init(htab);
569 		if (err)
570 			goto free_map_locked;
571 
572 		if (!percpu && !lru) {
573 			/* lru itself can remove the least used element, so
574 			 * there is no need for an extra elem during map_update.
575 			 */
576 			err = alloc_extra_elems(htab);
577 			if (err)
578 				goto free_prealloc;
579 		}
580 	} else {
581 		err = bpf_mem_alloc_init(&htab->ma, htab->elem_size, false);
582 		if (err)
583 			goto free_map_locked;
584 		if (percpu) {
585 			err = bpf_mem_alloc_init(&htab->pcpu_ma,
586 						 round_up(htab->map.value_size, 8), true);
587 			if (err)
588 				goto free_map_locked;
589 		}
590 	}
591 
592 	return &htab->map;
593 
594 free_prealloc:
595 	prealloc_destroy(htab);
596 free_map_locked:
597 	if (htab->use_percpu_counter)
598 		percpu_counter_destroy(&htab->pcount);
599 	for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++)
600 		free_percpu(htab->map_locked[i]);
601 	bpf_map_area_free(htab->buckets);
602 	bpf_mem_alloc_destroy(&htab->pcpu_ma);
603 	bpf_mem_alloc_destroy(&htab->ma);
604 free_elem_count:
605 	bpf_map_free_elem_count(&htab->map);
606 free_htab:
607 	lockdep_unregister_key(&htab->lockdep_key);
608 	bpf_map_area_free(htab);
609 	return ERR_PTR(err);
610 }
611 
612 static inline u32 htab_map_hash(const void *key, u32 key_len, u32 hashrnd)
613 {
614 	if (likely(key_len % 4 == 0))
615 		return jhash2(key, key_len / 4, hashrnd);
616 	return jhash(key, key_len, hashrnd);
617 }
618 
619 static inline struct bucket *__select_bucket(struct bpf_htab *htab, u32 hash)
620 {
621 	return &htab->buckets[hash & (htab->n_buckets - 1)];
622 }
623 
624 static inline struct hlist_nulls_head *select_bucket(struct bpf_htab *htab, u32 hash)
625 {
626 	return &__select_bucket(htab, hash)->head;
627 }
628 
629 /* this lookup function can only be called with bucket lock taken */
630 static struct htab_elem *lookup_elem_raw(struct hlist_nulls_head *head, u32 hash,
631 					 void *key, u32 key_size)
632 {
633 	struct hlist_nulls_node *n;
634 	struct htab_elem *l;
635 
636 	hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
637 		if (l->hash == hash && !memcmp(&l->key, key, key_size))
638 			return l;
639 
640 	return NULL;
641 }
642 
643 /* can be called without bucket lock. it will repeat the loop in
644  * the unlikely event when elements moved from one bucket into another
645  * while link list is being walked
646  */
647 static struct htab_elem *lookup_nulls_elem_raw(struct hlist_nulls_head *head,
648 					       u32 hash, void *key,
649 					       u32 key_size, u32 n_buckets)
650 {
651 	struct hlist_nulls_node *n;
652 	struct htab_elem *l;
653 
654 again:
655 	hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
656 		if (l->hash == hash && !memcmp(&l->key, key, key_size))
657 			return l;
658 
659 	if (unlikely(get_nulls_value(n) != (hash & (n_buckets - 1))))
660 		goto again;
661 
662 	return NULL;
663 }
664 
665 /* Called from syscall or from eBPF program directly, so
666  * arguments have to match bpf_map_lookup_elem() exactly.
667  * The return value is adjusted by BPF instructions
668  * in htab_map_gen_lookup().
669  */
670 static void *__htab_map_lookup_elem(struct bpf_map *map, void *key)
671 {
672 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
673 	struct hlist_nulls_head *head;
674 	struct htab_elem *l;
675 	u32 hash, key_size;
676 
677 	WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
678 		     !rcu_read_lock_bh_held());
679 
680 	key_size = map->key_size;
681 
682 	hash = htab_map_hash(key, key_size, htab->hashrnd);
683 
684 	head = select_bucket(htab, hash);
685 
686 	l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
687 
688 	return l;
689 }
690 
691 static void *htab_map_lookup_elem(struct bpf_map *map, void *key)
692 {
693 	struct htab_elem *l = __htab_map_lookup_elem(map, key);
694 
695 	if (l)
696 		return l->key + round_up(map->key_size, 8);
697 
698 	return NULL;
699 }
700 
701 /* inline bpf_map_lookup_elem() call.
702  * Instead of:
703  * bpf_prog
704  *   bpf_map_lookup_elem
705  *     map->ops->map_lookup_elem
706  *       htab_map_lookup_elem
707  *         __htab_map_lookup_elem
708  * do:
709  * bpf_prog
710  *   __htab_map_lookup_elem
711  */
712 static int htab_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
713 {
714 	struct bpf_insn *insn = insn_buf;
715 	const int ret = BPF_REG_0;
716 
717 	BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
718 		     (void *(*)(struct bpf_map *map, void *key))NULL));
719 	*insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
720 	*insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
721 	*insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
722 				offsetof(struct htab_elem, key) +
723 				round_up(map->key_size, 8));
724 	return insn - insn_buf;
725 }
726 
727 static __always_inline void *__htab_lru_map_lookup_elem(struct bpf_map *map,
728 							void *key, const bool mark)
729 {
730 	struct htab_elem *l = __htab_map_lookup_elem(map, key);
731 
732 	if (l) {
733 		if (mark)
734 			bpf_lru_node_set_ref(&l->lru_node);
735 		return l->key + round_up(map->key_size, 8);
736 	}
737 
738 	return NULL;
739 }
740 
741 static void *htab_lru_map_lookup_elem(struct bpf_map *map, void *key)
742 {
743 	return __htab_lru_map_lookup_elem(map, key, true);
744 }
745 
746 static void *htab_lru_map_lookup_elem_sys(struct bpf_map *map, void *key)
747 {
748 	return __htab_lru_map_lookup_elem(map, key, false);
749 }
750 
751 static int htab_lru_map_gen_lookup(struct bpf_map *map,
752 				   struct bpf_insn *insn_buf)
753 {
754 	struct bpf_insn *insn = insn_buf;
755 	const int ret = BPF_REG_0;
756 	const int ref_reg = BPF_REG_1;
757 
758 	BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
759 		     (void *(*)(struct bpf_map *map, void *key))NULL));
760 	*insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
761 	*insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 4);
762 	*insn++ = BPF_LDX_MEM(BPF_B, ref_reg, ret,
763 			      offsetof(struct htab_elem, lru_node) +
764 			      offsetof(struct bpf_lru_node, ref));
765 	*insn++ = BPF_JMP_IMM(BPF_JNE, ref_reg, 0, 1);
766 	*insn++ = BPF_ST_MEM(BPF_B, ret,
767 			     offsetof(struct htab_elem, lru_node) +
768 			     offsetof(struct bpf_lru_node, ref),
769 			     1);
770 	*insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
771 				offsetof(struct htab_elem, key) +
772 				round_up(map->key_size, 8));
773 	return insn - insn_buf;
774 }
775 
776 static void check_and_free_fields(struct bpf_htab *htab,
777 				  struct htab_elem *elem)
778 {
779 	if (htab_is_percpu(htab)) {
780 		void __percpu *pptr = htab_elem_get_ptr(elem, htab->map.key_size);
781 		int cpu;
782 
783 		for_each_possible_cpu(cpu)
784 			bpf_obj_free_fields(htab->map.record, per_cpu_ptr(pptr, cpu));
785 	} else {
786 		void *map_value = elem->key + round_up(htab->map.key_size, 8);
787 
788 		bpf_obj_free_fields(htab->map.record, map_value);
789 	}
790 }
791 
792 /* It is called from the bpf_lru_list when the LRU needs to delete
793  * older elements from the htab.
794  */
795 static bool htab_lru_map_delete_node(void *arg, struct bpf_lru_node *node)
796 {
797 	struct bpf_htab *htab = arg;
798 	struct htab_elem *l = NULL, *tgt_l;
799 	struct hlist_nulls_head *head;
800 	struct hlist_nulls_node *n;
801 	unsigned long flags;
802 	struct bucket *b;
803 	int ret;
804 
805 	tgt_l = container_of(node, struct htab_elem, lru_node);
806 	b = __select_bucket(htab, tgt_l->hash);
807 	head = &b->head;
808 
809 	ret = htab_lock_bucket(htab, b, tgt_l->hash, &flags);
810 	if (ret)
811 		return false;
812 
813 	hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
814 		if (l == tgt_l) {
815 			hlist_nulls_del_rcu(&l->hash_node);
816 			check_and_free_fields(htab, l);
817 			bpf_map_dec_elem_count(&htab->map);
818 			break;
819 		}
820 
821 	htab_unlock_bucket(htab, b, tgt_l->hash, flags);
822 
823 	return l == tgt_l;
824 }
825 
826 /* Called from syscall */
827 static int htab_map_get_next_key(struct bpf_map *map, void *key, void *next_key)
828 {
829 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
830 	struct hlist_nulls_head *head;
831 	struct htab_elem *l, *next_l;
832 	u32 hash, key_size;
833 	int i = 0;
834 
835 	WARN_ON_ONCE(!rcu_read_lock_held());
836 
837 	key_size = map->key_size;
838 
839 	if (!key)
840 		goto find_first_elem;
841 
842 	hash = htab_map_hash(key, key_size, htab->hashrnd);
843 
844 	head = select_bucket(htab, hash);
845 
846 	/* lookup the key */
847 	l = lookup_nulls_elem_raw(head, hash, key, key_size, htab->n_buckets);
848 
849 	if (!l)
850 		goto find_first_elem;
851 
852 	/* key was found, get next key in the same bucket */
853 	next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_next_rcu(&l->hash_node)),
854 				  struct htab_elem, hash_node);
855 
856 	if (next_l) {
857 		/* if next elem in this hash list is non-zero, just return it */
858 		memcpy(next_key, next_l->key, key_size);
859 		return 0;
860 	}
861 
862 	/* no more elements in this hash list, go to the next bucket */
863 	i = hash & (htab->n_buckets - 1);
864 	i++;
865 
866 find_first_elem:
867 	/* iterate over buckets */
868 	for (; i < htab->n_buckets; i++) {
869 		head = select_bucket(htab, i);
870 
871 		/* pick first element in the bucket */
872 		next_l = hlist_nulls_entry_safe(rcu_dereference_raw(hlist_nulls_first_rcu(head)),
873 					  struct htab_elem, hash_node);
874 		if (next_l) {
875 			/* if it's not empty, just return it */
876 			memcpy(next_key, next_l->key, key_size);
877 			return 0;
878 		}
879 	}
880 
881 	/* iterated over all buckets and all elements */
882 	return -ENOENT;
883 }
884 
885 static void htab_elem_free(struct bpf_htab *htab, struct htab_elem *l)
886 {
887 	check_and_free_fields(htab, l);
888 	if (htab->map.map_type == BPF_MAP_TYPE_PERCPU_HASH)
889 		bpf_mem_cache_free(&htab->pcpu_ma, l->ptr_to_pptr);
890 	bpf_mem_cache_free(&htab->ma, l);
891 }
892 
893 static void htab_put_fd_value(struct bpf_htab *htab, struct htab_elem *l)
894 {
895 	struct bpf_map *map = &htab->map;
896 	void *ptr;
897 
898 	if (map->ops->map_fd_put_ptr) {
899 		ptr = fd_htab_map_get_ptr(map, l);
900 		map->ops->map_fd_put_ptr(map, ptr, true);
901 	}
902 }
903 
904 static bool is_map_full(struct bpf_htab *htab)
905 {
906 	if (htab->use_percpu_counter)
907 		return __percpu_counter_compare(&htab->pcount, htab->map.max_entries,
908 						PERCPU_COUNTER_BATCH) >= 0;
909 	return atomic_read(&htab->count) >= htab->map.max_entries;
910 }
911 
912 static void inc_elem_count(struct bpf_htab *htab)
913 {
914 	bpf_map_inc_elem_count(&htab->map);
915 
916 	if (htab->use_percpu_counter)
917 		percpu_counter_add_batch(&htab->pcount, 1, PERCPU_COUNTER_BATCH);
918 	else
919 		atomic_inc(&htab->count);
920 }
921 
922 static void dec_elem_count(struct bpf_htab *htab)
923 {
924 	bpf_map_dec_elem_count(&htab->map);
925 
926 	if (htab->use_percpu_counter)
927 		percpu_counter_add_batch(&htab->pcount, -1, PERCPU_COUNTER_BATCH);
928 	else
929 		atomic_dec(&htab->count);
930 }
931 
932 
933 static void free_htab_elem(struct bpf_htab *htab, struct htab_elem *l)
934 {
935 	htab_put_fd_value(htab, l);
936 
937 	if (htab_is_prealloc(htab)) {
938 		bpf_map_dec_elem_count(&htab->map);
939 		check_and_free_fields(htab, l);
940 		__pcpu_freelist_push(&htab->freelist, &l->fnode);
941 	} else {
942 		dec_elem_count(htab);
943 		htab_elem_free(htab, l);
944 	}
945 }
946 
947 static void pcpu_copy_value(struct bpf_htab *htab, void __percpu *pptr,
948 			    void *value, bool onallcpus)
949 {
950 	if (!onallcpus) {
951 		/* copy true value_size bytes */
952 		copy_map_value(&htab->map, this_cpu_ptr(pptr), value);
953 	} else {
954 		u32 size = round_up(htab->map.value_size, 8);
955 		int off = 0, cpu;
956 
957 		for_each_possible_cpu(cpu) {
958 			copy_map_value_long(&htab->map, per_cpu_ptr(pptr, cpu), value + off);
959 			off += size;
960 		}
961 	}
962 }
963 
964 static void pcpu_init_value(struct bpf_htab *htab, void __percpu *pptr,
965 			    void *value, bool onallcpus)
966 {
967 	/* When not setting the initial value on all cpus, zero-fill element
968 	 * values for other cpus. Otherwise, bpf program has no way to ensure
969 	 * known initial values for cpus other than current one
970 	 * (onallcpus=false always when coming from bpf prog).
971 	 */
972 	if (!onallcpus) {
973 		int current_cpu = raw_smp_processor_id();
974 		int cpu;
975 
976 		for_each_possible_cpu(cpu) {
977 			if (cpu == current_cpu)
978 				copy_map_value_long(&htab->map, per_cpu_ptr(pptr, cpu), value);
979 			else /* Since elem is preallocated, we cannot touch special fields */
980 				zero_map_value(&htab->map, per_cpu_ptr(pptr, cpu));
981 		}
982 	} else {
983 		pcpu_copy_value(htab, pptr, value, onallcpus);
984 	}
985 }
986 
987 static bool fd_htab_map_needs_adjust(const struct bpf_htab *htab)
988 {
989 	return htab->map.map_type == BPF_MAP_TYPE_HASH_OF_MAPS &&
990 	       BITS_PER_LONG == 64;
991 }
992 
993 static struct htab_elem *alloc_htab_elem(struct bpf_htab *htab, void *key,
994 					 void *value, u32 key_size, u32 hash,
995 					 bool percpu, bool onallcpus,
996 					 struct htab_elem *old_elem)
997 {
998 	u32 size = htab->map.value_size;
999 	bool prealloc = htab_is_prealloc(htab);
1000 	struct htab_elem *l_new, **pl_new;
1001 	void __percpu *pptr;
1002 
1003 	if (prealloc) {
1004 		if (old_elem) {
1005 			/* if we're updating the existing element,
1006 			 * use per-cpu extra elems to avoid freelist_pop/push
1007 			 */
1008 			pl_new = this_cpu_ptr(htab->extra_elems);
1009 			l_new = *pl_new;
1010 			htab_put_fd_value(htab, old_elem);
1011 			*pl_new = old_elem;
1012 		} else {
1013 			struct pcpu_freelist_node *l;
1014 
1015 			l = __pcpu_freelist_pop(&htab->freelist);
1016 			if (!l)
1017 				return ERR_PTR(-E2BIG);
1018 			l_new = container_of(l, struct htab_elem, fnode);
1019 			bpf_map_inc_elem_count(&htab->map);
1020 		}
1021 	} else {
1022 		if (is_map_full(htab))
1023 			if (!old_elem)
1024 				/* when map is full and update() is replacing
1025 				 * old element, it's ok to allocate, since
1026 				 * old element will be freed immediately.
1027 				 * Otherwise return an error
1028 				 */
1029 				return ERR_PTR(-E2BIG);
1030 		inc_elem_count(htab);
1031 		l_new = bpf_mem_cache_alloc(&htab->ma);
1032 		if (!l_new) {
1033 			l_new = ERR_PTR(-ENOMEM);
1034 			goto dec_count;
1035 		}
1036 	}
1037 
1038 	memcpy(l_new->key, key, key_size);
1039 	if (percpu) {
1040 		if (prealloc) {
1041 			pptr = htab_elem_get_ptr(l_new, key_size);
1042 		} else {
1043 			/* alloc_percpu zero-fills */
1044 			pptr = bpf_mem_cache_alloc(&htab->pcpu_ma);
1045 			if (!pptr) {
1046 				bpf_mem_cache_free(&htab->ma, l_new);
1047 				l_new = ERR_PTR(-ENOMEM);
1048 				goto dec_count;
1049 			}
1050 			l_new->ptr_to_pptr = pptr;
1051 			pptr = *(void **)pptr;
1052 		}
1053 
1054 		pcpu_init_value(htab, pptr, value, onallcpus);
1055 
1056 		if (!prealloc)
1057 			htab_elem_set_ptr(l_new, key_size, pptr);
1058 	} else if (fd_htab_map_needs_adjust(htab)) {
1059 		size = round_up(size, 8);
1060 		memcpy(l_new->key + round_up(key_size, 8), value, size);
1061 	} else {
1062 		copy_map_value(&htab->map,
1063 			       l_new->key + round_up(key_size, 8),
1064 			       value);
1065 	}
1066 
1067 	l_new->hash = hash;
1068 	return l_new;
1069 dec_count:
1070 	dec_elem_count(htab);
1071 	return l_new;
1072 }
1073 
1074 static int check_flags(struct bpf_htab *htab, struct htab_elem *l_old,
1075 		       u64 map_flags)
1076 {
1077 	if (l_old && (map_flags & ~BPF_F_LOCK) == BPF_NOEXIST)
1078 		/* elem already exists */
1079 		return -EEXIST;
1080 
1081 	if (!l_old && (map_flags & ~BPF_F_LOCK) == BPF_EXIST)
1082 		/* elem doesn't exist, cannot update it */
1083 		return -ENOENT;
1084 
1085 	return 0;
1086 }
1087 
1088 /* Called from syscall or from eBPF program */
1089 static long htab_map_update_elem(struct bpf_map *map, void *key, void *value,
1090 				 u64 map_flags)
1091 {
1092 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1093 	struct htab_elem *l_new = NULL, *l_old;
1094 	struct hlist_nulls_head *head;
1095 	unsigned long flags;
1096 	struct bucket *b;
1097 	u32 key_size, hash;
1098 	int ret;
1099 
1100 	if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST))
1101 		/* unknown flags */
1102 		return -EINVAL;
1103 
1104 	WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1105 		     !rcu_read_lock_bh_held());
1106 
1107 	key_size = map->key_size;
1108 
1109 	hash = htab_map_hash(key, key_size, htab->hashrnd);
1110 
1111 	b = __select_bucket(htab, hash);
1112 	head = &b->head;
1113 
1114 	if (unlikely(map_flags & BPF_F_LOCK)) {
1115 		if (unlikely(!btf_record_has_field(map->record, BPF_SPIN_LOCK)))
1116 			return -EINVAL;
1117 		/* find an element without taking the bucket lock */
1118 		l_old = lookup_nulls_elem_raw(head, hash, key, key_size,
1119 					      htab->n_buckets);
1120 		ret = check_flags(htab, l_old, map_flags);
1121 		if (ret)
1122 			return ret;
1123 		if (l_old) {
1124 			/* grab the element lock and update value in place */
1125 			copy_map_value_locked(map,
1126 					      l_old->key + round_up(key_size, 8),
1127 					      value, false);
1128 			return 0;
1129 		}
1130 		/* fall through, grab the bucket lock and lookup again.
1131 		 * 99.9% chance that the element won't be found,
1132 		 * but second lookup under lock has to be done.
1133 		 */
1134 	}
1135 
1136 	ret = htab_lock_bucket(htab, b, hash, &flags);
1137 	if (ret)
1138 		return ret;
1139 
1140 	l_old = lookup_elem_raw(head, hash, key, key_size);
1141 
1142 	ret = check_flags(htab, l_old, map_flags);
1143 	if (ret)
1144 		goto err;
1145 
1146 	if (unlikely(l_old && (map_flags & BPF_F_LOCK))) {
1147 		/* first lookup without the bucket lock didn't find the element,
1148 		 * but second lookup with the bucket lock found it.
1149 		 * This case is highly unlikely, but has to be dealt with:
1150 		 * grab the element lock in addition to the bucket lock
1151 		 * and update element in place
1152 		 */
1153 		copy_map_value_locked(map,
1154 				      l_old->key + round_up(key_size, 8),
1155 				      value, false);
1156 		ret = 0;
1157 		goto err;
1158 	}
1159 
1160 	l_new = alloc_htab_elem(htab, key, value, key_size, hash, false, false,
1161 				l_old);
1162 	if (IS_ERR(l_new)) {
1163 		/* all pre-allocated elements are in use or memory exhausted */
1164 		ret = PTR_ERR(l_new);
1165 		goto err;
1166 	}
1167 
1168 	/* add new element to the head of the list, so that
1169 	 * concurrent search will find it before old elem
1170 	 */
1171 	hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1172 	if (l_old) {
1173 		hlist_nulls_del_rcu(&l_old->hash_node);
1174 		if (!htab_is_prealloc(htab))
1175 			free_htab_elem(htab, l_old);
1176 		else
1177 			check_and_free_fields(htab, l_old);
1178 	}
1179 	ret = 0;
1180 err:
1181 	htab_unlock_bucket(htab, b, hash, flags);
1182 	return ret;
1183 }
1184 
1185 static void htab_lru_push_free(struct bpf_htab *htab, struct htab_elem *elem)
1186 {
1187 	check_and_free_fields(htab, elem);
1188 	bpf_map_dec_elem_count(&htab->map);
1189 	bpf_lru_push_free(&htab->lru, &elem->lru_node);
1190 }
1191 
1192 static long htab_lru_map_update_elem(struct bpf_map *map, void *key, void *value,
1193 				     u64 map_flags)
1194 {
1195 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1196 	struct htab_elem *l_new, *l_old = NULL;
1197 	struct hlist_nulls_head *head;
1198 	unsigned long flags;
1199 	struct bucket *b;
1200 	u32 key_size, hash;
1201 	int ret;
1202 
1203 	if (unlikely(map_flags > BPF_EXIST))
1204 		/* unknown flags */
1205 		return -EINVAL;
1206 
1207 	WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1208 		     !rcu_read_lock_bh_held());
1209 
1210 	key_size = map->key_size;
1211 
1212 	hash = htab_map_hash(key, key_size, htab->hashrnd);
1213 
1214 	b = __select_bucket(htab, hash);
1215 	head = &b->head;
1216 
1217 	/* For LRU, we need to alloc before taking bucket's
1218 	 * spinlock because getting free nodes from LRU may need
1219 	 * to remove older elements from htab and this removal
1220 	 * operation will need a bucket lock.
1221 	 */
1222 	l_new = prealloc_lru_pop(htab, key, hash);
1223 	if (!l_new)
1224 		return -ENOMEM;
1225 	copy_map_value(&htab->map,
1226 		       l_new->key + round_up(map->key_size, 8), value);
1227 
1228 	ret = htab_lock_bucket(htab, b, hash, &flags);
1229 	if (ret)
1230 		goto err_lock_bucket;
1231 
1232 	l_old = lookup_elem_raw(head, hash, key, key_size);
1233 
1234 	ret = check_flags(htab, l_old, map_flags);
1235 	if (ret)
1236 		goto err;
1237 
1238 	/* add new element to the head of the list, so that
1239 	 * concurrent search will find it before old elem
1240 	 */
1241 	hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1242 	if (l_old) {
1243 		bpf_lru_node_set_ref(&l_new->lru_node);
1244 		hlist_nulls_del_rcu(&l_old->hash_node);
1245 	}
1246 	ret = 0;
1247 
1248 err:
1249 	htab_unlock_bucket(htab, b, hash, flags);
1250 
1251 err_lock_bucket:
1252 	if (ret)
1253 		htab_lru_push_free(htab, l_new);
1254 	else if (l_old)
1255 		htab_lru_push_free(htab, l_old);
1256 
1257 	return ret;
1258 }
1259 
1260 static long __htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1261 					  void *value, u64 map_flags,
1262 					  bool onallcpus)
1263 {
1264 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1265 	struct htab_elem *l_new = NULL, *l_old;
1266 	struct hlist_nulls_head *head;
1267 	unsigned long flags;
1268 	struct bucket *b;
1269 	u32 key_size, hash;
1270 	int ret;
1271 
1272 	if (unlikely(map_flags > BPF_EXIST))
1273 		/* unknown flags */
1274 		return -EINVAL;
1275 
1276 	WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1277 		     !rcu_read_lock_bh_held());
1278 
1279 	key_size = map->key_size;
1280 
1281 	hash = htab_map_hash(key, key_size, htab->hashrnd);
1282 
1283 	b = __select_bucket(htab, hash);
1284 	head = &b->head;
1285 
1286 	ret = htab_lock_bucket(htab, b, hash, &flags);
1287 	if (ret)
1288 		return ret;
1289 
1290 	l_old = lookup_elem_raw(head, hash, key, key_size);
1291 
1292 	ret = check_flags(htab, l_old, map_flags);
1293 	if (ret)
1294 		goto err;
1295 
1296 	if (l_old) {
1297 		/* per-cpu hash map can update value in-place */
1298 		pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1299 				value, onallcpus);
1300 	} else {
1301 		l_new = alloc_htab_elem(htab, key, value, key_size,
1302 					hash, true, onallcpus, NULL);
1303 		if (IS_ERR(l_new)) {
1304 			ret = PTR_ERR(l_new);
1305 			goto err;
1306 		}
1307 		hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1308 	}
1309 	ret = 0;
1310 err:
1311 	htab_unlock_bucket(htab, b, hash, flags);
1312 	return ret;
1313 }
1314 
1315 static long __htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1316 					      void *value, u64 map_flags,
1317 					      bool onallcpus)
1318 {
1319 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1320 	struct htab_elem *l_new = NULL, *l_old;
1321 	struct hlist_nulls_head *head;
1322 	unsigned long flags;
1323 	struct bucket *b;
1324 	u32 key_size, hash;
1325 	int ret;
1326 
1327 	if (unlikely(map_flags > BPF_EXIST))
1328 		/* unknown flags */
1329 		return -EINVAL;
1330 
1331 	WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1332 		     !rcu_read_lock_bh_held());
1333 
1334 	key_size = map->key_size;
1335 
1336 	hash = htab_map_hash(key, key_size, htab->hashrnd);
1337 
1338 	b = __select_bucket(htab, hash);
1339 	head = &b->head;
1340 
1341 	/* For LRU, we need to alloc before taking bucket's
1342 	 * spinlock because LRU's elem alloc may need
1343 	 * to remove older elem from htab and this removal
1344 	 * operation will need a bucket lock.
1345 	 */
1346 	if (map_flags != BPF_EXIST) {
1347 		l_new = prealloc_lru_pop(htab, key, hash);
1348 		if (!l_new)
1349 			return -ENOMEM;
1350 	}
1351 
1352 	ret = htab_lock_bucket(htab, b, hash, &flags);
1353 	if (ret)
1354 		goto err_lock_bucket;
1355 
1356 	l_old = lookup_elem_raw(head, hash, key, key_size);
1357 
1358 	ret = check_flags(htab, l_old, map_flags);
1359 	if (ret)
1360 		goto err;
1361 
1362 	if (l_old) {
1363 		bpf_lru_node_set_ref(&l_old->lru_node);
1364 
1365 		/* per-cpu hash map can update value in-place */
1366 		pcpu_copy_value(htab, htab_elem_get_ptr(l_old, key_size),
1367 				value, onallcpus);
1368 	} else {
1369 		pcpu_init_value(htab, htab_elem_get_ptr(l_new, key_size),
1370 				value, onallcpus);
1371 		hlist_nulls_add_head_rcu(&l_new->hash_node, head);
1372 		l_new = NULL;
1373 	}
1374 	ret = 0;
1375 err:
1376 	htab_unlock_bucket(htab, b, hash, flags);
1377 err_lock_bucket:
1378 	if (l_new) {
1379 		bpf_map_dec_elem_count(&htab->map);
1380 		bpf_lru_push_free(&htab->lru, &l_new->lru_node);
1381 	}
1382 	return ret;
1383 }
1384 
1385 static long htab_percpu_map_update_elem(struct bpf_map *map, void *key,
1386 					void *value, u64 map_flags)
1387 {
1388 	return __htab_percpu_map_update_elem(map, key, value, map_flags, false);
1389 }
1390 
1391 static long htab_lru_percpu_map_update_elem(struct bpf_map *map, void *key,
1392 					    void *value, u64 map_flags)
1393 {
1394 	return __htab_lru_percpu_map_update_elem(map, key, value, map_flags,
1395 						 false);
1396 }
1397 
1398 /* Called from syscall or from eBPF program */
1399 static long htab_map_delete_elem(struct bpf_map *map, void *key)
1400 {
1401 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1402 	struct hlist_nulls_head *head;
1403 	struct bucket *b;
1404 	struct htab_elem *l;
1405 	unsigned long flags;
1406 	u32 hash, key_size;
1407 	int ret;
1408 
1409 	WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1410 		     !rcu_read_lock_bh_held());
1411 
1412 	key_size = map->key_size;
1413 
1414 	hash = htab_map_hash(key, key_size, htab->hashrnd);
1415 	b = __select_bucket(htab, hash);
1416 	head = &b->head;
1417 
1418 	ret = htab_lock_bucket(htab, b, hash, &flags);
1419 	if (ret)
1420 		return ret;
1421 
1422 	l = lookup_elem_raw(head, hash, key, key_size);
1423 
1424 	if (l) {
1425 		hlist_nulls_del_rcu(&l->hash_node);
1426 		free_htab_elem(htab, l);
1427 	} else {
1428 		ret = -ENOENT;
1429 	}
1430 
1431 	htab_unlock_bucket(htab, b, hash, flags);
1432 	return ret;
1433 }
1434 
1435 static long htab_lru_map_delete_elem(struct bpf_map *map, void *key)
1436 {
1437 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1438 	struct hlist_nulls_head *head;
1439 	struct bucket *b;
1440 	struct htab_elem *l;
1441 	unsigned long flags;
1442 	u32 hash, key_size;
1443 	int ret;
1444 
1445 	WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_trace_held() &&
1446 		     !rcu_read_lock_bh_held());
1447 
1448 	key_size = map->key_size;
1449 
1450 	hash = htab_map_hash(key, key_size, htab->hashrnd);
1451 	b = __select_bucket(htab, hash);
1452 	head = &b->head;
1453 
1454 	ret = htab_lock_bucket(htab, b, hash, &flags);
1455 	if (ret)
1456 		return ret;
1457 
1458 	l = lookup_elem_raw(head, hash, key, key_size);
1459 
1460 	if (l)
1461 		hlist_nulls_del_rcu(&l->hash_node);
1462 	else
1463 		ret = -ENOENT;
1464 
1465 	htab_unlock_bucket(htab, b, hash, flags);
1466 	if (l)
1467 		htab_lru_push_free(htab, l);
1468 	return ret;
1469 }
1470 
1471 static void delete_all_elements(struct bpf_htab *htab)
1472 {
1473 	int i;
1474 
1475 	/* It's called from a worker thread, so disable migration here,
1476 	 * since bpf_mem_cache_free() relies on that.
1477 	 */
1478 	migrate_disable();
1479 	for (i = 0; i < htab->n_buckets; i++) {
1480 		struct hlist_nulls_head *head = select_bucket(htab, i);
1481 		struct hlist_nulls_node *n;
1482 		struct htab_elem *l;
1483 
1484 		hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1485 			hlist_nulls_del_rcu(&l->hash_node);
1486 			htab_elem_free(htab, l);
1487 		}
1488 	}
1489 	migrate_enable();
1490 }
1491 
1492 static void htab_free_malloced_timers(struct bpf_htab *htab)
1493 {
1494 	int i;
1495 
1496 	rcu_read_lock();
1497 	for (i = 0; i < htab->n_buckets; i++) {
1498 		struct hlist_nulls_head *head = select_bucket(htab, i);
1499 		struct hlist_nulls_node *n;
1500 		struct htab_elem *l;
1501 
1502 		hlist_nulls_for_each_entry(l, n, head, hash_node) {
1503 			/* We only free timer on uref dropping to zero */
1504 			bpf_obj_free_timer(htab->map.record, l->key + round_up(htab->map.key_size, 8));
1505 		}
1506 		cond_resched_rcu();
1507 	}
1508 	rcu_read_unlock();
1509 }
1510 
1511 static void htab_map_free_timers(struct bpf_map *map)
1512 {
1513 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1514 
1515 	/* We only free timer on uref dropping to zero */
1516 	if (!btf_record_has_field(htab->map.record, BPF_TIMER))
1517 		return;
1518 	if (!htab_is_prealloc(htab))
1519 		htab_free_malloced_timers(htab);
1520 	else
1521 		htab_free_prealloced_timers(htab);
1522 }
1523 
1524 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
1525 static void htab_map_free(struct bpf_map *map)
1526 {
1527 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1528 	int i;
1529 
1530 	/* bpf_free_used_maps() or close(map_fd) will trigger this map_free callback.
1531 	 * bpf_free_used_maps() is called after bpf prog is no longer executing.
1532 	 * There is no need to synchronize_rcu() here to protect map elements.
1533 	 */
1534 
1535 	/* htab no longer uses call_rcu() directly. bpf_mem_alloc does it
1536 	 * underneath and is reponsible for waiting for callbacks to finish
1537 	 * during bpf_mem_alloc_destroy().
1538 	 */
1539 	if (!htab_is_prealloc(htab)) {
1540 		delete_all_elements(htab);
1541 	} else {
1542 		htab_free_prealloced_fields(htab);
1543 		prealloc_destroy(htab);
1544 	}
1545 
1546 	bpf_map_free_elem_count(map);
1547 	free_percpu(htab->extra_elems);
1548 	bpf_map_area_free(htab->buckets);
1549 	bpf_mem_alloc_destroy(&htab->pcpu_ma);
1550 	bpf_mem_alloc_destroy(&htab->ma);
1551 	if (htab->use_percpu_counter)
1552 		percpu_counter_destroy(&htab->pcount);
1553 	for (i = 0; i < HASHTAB_MAP_LOCK_COUNT; i++)
1554 		free_percpu(htab->map_locked[i]);
1555 	lockdep_unregister_key(&htab->lockdep_key);
1556 	bpf_map_area_free(htab);
1557 }
1558 
1559 static void htab_map_seq_show_elem(struct bpf_map *map, void *key,
1560 				   struct seq_file *m)
1561 {
1562 	void *value;
1563 
1564 	rcu_read_lock();
1565 
1566 	value = htab_map_lookup_elem(map, key);
1567 	if (!value) {
1568 		rcu_read_unlock();
1569 		return;
1570 	}
1571 
1572 	btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
1573 	seq_puts(m, ": ");
1574 	btf_type_seq_show(map->btf, map->btf_value_type_id, value, m);
1575 	seq_puts(m, "\n");
1576 
1577 	rcu_read_unlock();
1578 }
1579 
1580 static int __htab_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
1581 					     void *value, bool is_lru_map,
1582 					     bool is_percpu, u64 flags)
1583 {
1584 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1585 	struct hlist_nulls_head *head;
1586 	unsigned long bflags;
1587 	struct htab_elem *l;
1588 	u32 hash, key_size;
1589 	struct bucket *b;
1590 	int ret;
1591 
1592 	key_size = map->key_size;
1593 
1594 	hash = htab_map_hash(key, key_size, htab->hashrnd);
1595 	b = __select_bucket(htab, hash);
1596 	head = &b->head;
1597 
1598 	ret = htab_lock_bucket(htab, b, hash, &bflags);
1599 	if (ret)
1600 		return ret;
1601 
1602 	l = lookup_elem_raw(head, hash, key, key_size);
1603 	if (!l) {
1604 		ret = -ENOENT;
1605 	} else {
1606 		if (is_percpu) {
1607 			u32 roundup_value_size = round_up(map->value_size, 8);
1608 			void __percpu *pptr;
1609 			int off = 0, cpu;
1610 
1611 			pptr = htab_elem_get_ptr(l, key_size);
1612 			for_each_possible_cpu(cpu) {
1613 				copy_map_value_long(&htab->map, value + off, per_cpu_ptr(pptr, cpu));
1614 				check_and_init_map_value(&htab->map, value + off);
1615 				off += roundup_value_size;
1616 			}
1617 		} else {
1618 			u32 roundup_key_size = round_up(map->key_size, 8);
1619 
1620 			if (flags & BPF_F_LOCK)
1621 				copy_map_value_locked(map, value, l->key +
1622 						      roundup_key_size,
1623 						      true);
1624 			else
1625 				copy_map_value(map, value, l->key +
1626 					       roundup_key_size);
1627 			/* Zeroing special fields in the temp buffer */
1628 			check_and_init_map_value(map, value);
1629 		}
1630 
1631 		hlist_nulls_del_rcu(&l->hash_node);
1632 		if (!is_lru_map)
1633 			free_htab_elem(htab, l);
1634 	}
1635 
1636 	htab_unlock_bucket(htab, b, hash, bflags);
1637 
1638 	if (is_lru_map && l)
1639 		htab_lru_push_free(htab, l);
1640 
1641 	return ret;
1642 }
1643 
1644 static int htab_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
1645 					   void *value, u64 flags)
1646 {
1647 	return __htab_map_lookup_and_delete_elem(map, key, value, false, false,
1648 						 flags);
1649 }
1650 
1651 static int htab_percpu_map_lookup_and_delete_elem(struct bpf_map *map,
1652 						  void *key, void *value,
1653 						  u64 flags)
1654 {
1655 	return __htab_map_lookup_and_delete_elem(map, key, value, false, true,
1656 						 flags);
1657 }
1658 
1659 static int htab_lru_map_lookup_and_delete_elem(struct bpf_map *map, void *key,
1660 					       void *value, u64 flags)
1661 {
1662 	return __htab_map_lookup_and_delete_elem(map, key, value, true, false,
1663 						 flags);
1664 }
1665 
1666 static int htab_lru_percpu_map_lookup_and_delete_elem(struct bpf_map *map,
1667 						      void *key, void *value,
1668 						      u64 flags)
1669 {
1670 	return __htab_map_lookup_and_delete_elem(map, key, value, true, true,
1671 						 flags);
1672 }
1673 
1674 static int
1675 __htab_map_lookup_and_delete_batch(struct bpf_map *map,
1676 				   const union bpf_attr *attr,
1677 				   union bpf_attr __user *uattr,
1678 				   bool do_delete, bool is_lru_map,
1679 				   bool is_percpu)
1680 {
1681 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
1682 	u32 bucket_cnt, total, key_size, value_size, roundup_key_size;
1683 	void *keys = NULL, *values = NULL, *value, *dst_key, *dst_val;
1684 	void __user *uvalues = u64_to_user_ptr(attr->batch.values);
1685 	void __user *ukeys = u64_to_user_ptr(attr->batch.keys);
1686 	void __user *ubatch = u64_to_user_ptr(attr->batch.in_batch);
1687 	u32 batch, max_count, size, bucket_size, map_id;
1688 	struct htab_elem *node_to_free = NULL;
1689 	u64 elem_map_flags, map_flags;
1690 	struct hlist_nulls_head *head;
1691 	struct hlist_nulls_node *n;
1692 	unsigned long flags = 0;
1693 	bool locked = false;
1694 	struct htab_elem *l;
1695 	struct bucket *b;
1696 	int ret = 0;
1697 
1698 	elem_map_flags = attr->batch.elem_flags;
1699 	if ((elem_map_flags & ~BPF_F_LOCK) ||
1700 	    ((elem_map_flags & BPF_F_LOCK) && !btf_record_has_field(map->record, BPF_SPIN_LOCK)))
1701 		return -EINVAL;
1702 
1703 	map_flags = attr->batch.flags;
1704 	if (map_flags)
1705 		return -EINVAL;
1706 
1707 	max_count = attr->batch.count;
1708 	if (!max_count)
1709 		return 0;
1710 
1711 	if (put_user(0, &uattr->batch.count))
1712 		return -EFAULT;
1713 
1714 	batch = 0;
1715 	if (ubatch && copy_from_user(&batch, ubatch, sizeof(batch)))
1716 		return -EFAULT;
1717 
1718 	if (batch >= htab->n_buckets)
1719 		return -ENOENT;
1720 
1721 	key_size = htab->map.key_size;
1722 	roundup_key_size = round_up(htab->map.key_size, 8);
1723 	value_size = htab->map.value_size;
1724 	size = round_up(value_size, 8);
1725 	if (is_percpu)
1726 		value_size = size * num_possible_cpus();
1727 	total = 0;
1728 	/* while experimenting with hash tables with sizes ranging from 10 to
1729 	 * 1000, it was observed that a bucket can have up to 5 entries.
1730 	 */
1731 	bucket_size = 5;
1732 
1733 alloc:
1734 	/* We cannot do copy_from_user or copy_to_user inside
1735 	 * the rcu_read_lock. Allocate enough space here.
1736 	 */
1737 	keys = kvmalloc_array(key_size, bucket_size, GFP_USER | __GFP_NOWARN);
1738 	values = kvmalloc_array(value_size, bucket_size, GFP_USER | __GFP_NOWARN);
1739 	if (!keys || !values) {
1740 		ret = -ENOMEM;
1741 		goto after_loop;
1742 	}
1743 
1744 again:
1745 	bpf_disable_instrumentation();
1746 	rcu_read_lock();
1747 again_nocopy:
1748 	dst_key = keys;
1749 	dst_val = values;
1750 	b = &htab->buckets[batch];
1751 	head = &b->head;
1752 	/* do not grab the lock unless need it (bucket_cnt > 0). */
1753 	if (locked) {
1754 		ret = htab_lock_bucket(htab, b, batch, &flags);
1755 		if (ret) {
1756 			rcu_read_unlock();
1757 			bpf_enable_instrumentation();
1758 			goto after_loop;
1759 		}
1760 	}
1761 
1762 	bucket_cnt = 0;
1763 	hlist_nulls_for_each_entry_rcu(l, n, head, hash_node)
1764 		bucket_cnt++;
1765 
1766 	if (bucket_cnt && !locked) {
1767 		locked = true;
1768 		goto again_nocopy;
1769 	}
1770 
1771 	if (bucket_cnt > (max_count - total)) {
1772 		if (total == 0)
1773 			ret = -ENOSPC;
1774 		/* Note that since bucket_cnt > 0 here, it is implicit
1775 		 * that the locked was grabbed, so release it.
1776 		 */
1777 		htab_unlock_bucket(htab, b, batch, flags);
1778 		rcu_read_unlock();
1779 		bpf_enable_instrumentation();
1780 		goto after_loop;
1781 	}
1782 
1783 	if (bucket_cnt > bucket_size) {
1784 		bucket_size = bucket_cnt;
1785 		/* Note that since bucket_cnt > 0 here, it is implicit
1786 		 * that the locked was grabbed, so release it.
1787 		 */
1788 		htab_unlock_bucket(htab, b, batch, flags);
1789 		rcu_read_unlock();
1790 		bpf_enable_instrumentation();
1791 		kvfree(keys);
1792 		kvfree(values);
1793 		goto alloc;
1794 	}
1795 
1796 	/* Next block is only safe to run if you have grabbed the lock */
1797 	if (!locked)
1798 		goto next_batch;
1799 
1800 	hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
1801 		memcpy(dst_key, l->key, key_size);
1802 
1803 		if (is_percpu) {
1804 			int off = 0, cpu;
1805 			void __percpu *pptr;
1806 
1807 			pptr = htab_elem_get_ptr(l, map->key_size);
1808 			for_each_possible_cpu(cpu) {
1809 				copy_map_value_long(&htab->map, dst_val + off, per_cpu_ptr(pptr, cpu));
1810 				check_and_init_map_value(&htab->map, dst_val + off);
1811 				off += size;
1812 			}
1813 		} else {
1814 			value = l->key + roundup_key_size;
1815 			if (map->map_type == BPF_MAP_TYPE_HASH_OF_MAPS) {
1816 				struct bpf_map **inner_map = value;
1817 
1818 				 /* Actual value is the id of the inner map */
1819 				map_id = map->ops->map_fd_sys_lookup_elem(*inner_map);
1820 				value = &map_id;
1821 			}
1822 
1823 			if (elem_map_flags & BPF_F_LOCK)
1824 				copy_map_value_locked(map, dst_val, value,
1825 						      true);
1826 			else
1827 				copy_map_value(map, dst_val, value);
1828 			/* Zeroing special fields in the temp buffer */
1829 			check_and_init_map_value(map, dst_val);
1830 		}
1831 		if (do_delete) {
1832 			hlist_nulls_del_rcu(&l->hash_node);
1833 
1834 			/* bpf_lru_push_free() will acquire lru_lock, which
1835 			 * may cause deadlock. See comments in function
1836 			 * prealloc_lru_pop(). Let us do bpf_lru_push_free()
1837 			 * after releasing the bucket lock.
1838 			 */
1839 			if (is_lru_map) {
1840 				l->batch_flink = node_to_free;
1841 				node_to_free = l;
1842 			} else {
1843 				free_htab_elem(htab, l);
1844 			}
1845 		}
1846 		dst_key += key_size;
1847 		dst_val += value_size;
1848 	}
1849 
1850 	htab_unlock_bucket(htab, b, batch, flags);
1851 	locked = false;
1852 
1853 	while (node_to_free) {
1854 		l = node_to_free;
1855 		node_to_free = node_to_free->batch_flink;
1856 		htab_lru_push_free(htab, l);
1857 	}
1858 
1859 next_batch:
1860 	/* If we are not copying data, we can go to next bucket and avoid
1861 	 * unlocking the rcu.
1862 	 */
1863 	if (!bucket_cnt && (batch + 1 < htab->n_buckets)) {
1864 		batch++;
1865 		goto again_nocopy;
1866 	}
1867 
1868 	rcu_read_unlock();
1869 	bpf_enable_instrumentation();
1870 	if (bucket_cnt && (copy_to_user(ukeys + total * key_size, keys,
1871 	    key_size * bucket_cnt) ||
1872 	    copy_to_user(uvalues + total * value_size, values,
1873 	    value_size * bucket_cnt))) {
1874 		ret = -EFAULT;
1875 		goto after_loop;
1876 	}
1877 
1878 	total += bucket_cnt;
1879 	batch++;
1880 	if (batch >= htab->n_buckets) {
1881 		ret = -ENOENT;
1882 		goto after_loop;
1883 	}
1884 	goto again;
1885 
1886 after_loop:
1887 	if (ret == -EFAULT)
1888 		goto out;
1889 
1890 	/* copy # of entries and next batch */
1891 	ubatch = u64_to_user_ptr(attr->batch.out_batch);
1892 	if (copy_to_user(ubatch, &batch, sizeof(batch)) ||
1893 	    put_user(total, &uattr->batch.count))
1894 		ret = -EFAULT;
1895 
1896 out:
1897 	kvfree(keys);
1898 	kvfree(values);
1899 	return ret;
1900 }
1901 
1902 static int
1903 htab_percpu_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1904 			     union bpf_attr __user *uattr)
1905 {
1906 	return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1907 						  false, true);
1908 }
1909 
1910 static int
1911 htab_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1912 					const union bpf_attr *attr,
1913 					union bpf_attr __user *uattr)
1914 {
1915 	return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1916 						  false, true);
1917 }
1918 
1919 static int
1920 htab_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1921 		      union bpf_attr __user *uattr)
1922 {
1923 	return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1924 						  false, false);
1925 }
1926 
1927 static int
1928 htab_map_lookup_and_delete_batch(struct bpf_map *map,
1929 				 const union bpf_attr *attr,
1930 				 union bpf_attr __user *uattr)
1931 {
1932 	return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1933 						  false, false);
1934 }
1935 
1936 static int
1937 htab_lru_percpu_map_lookup_batch(struct bpf_map *map,
1938 				 const union bpf_attr *attr,
1939 				 union bpf_attr __user *uattr)
1940 {
1941 	return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1942 						  true, true);
1943 }
1944 
1945 static int
1946 htab_lru_percpu_map_lookup_and_delete_batch(struct bpf_map *map,
1947 					    const union bpf_attr *attr,
1948 					    union bpf_attr __user *uattr)
1949 {
1950 	return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1951 						  true, true);
1952 }
1953 
1954 static int
1955 htab_lru_map_lookup_batch(struct bpf_map *map, const union bpf_attr *attr,
1956 			  union bpf_attr __user *uattr)
1957 {
1958 	return __htab_map_lookup_and_delete_batch(map, attr, uattr, false,
1959 						  true, false);
1960 }
1961 
1962 static int
1963 htab_lru_map_lookup_and_delete_batch(struct bpf_map *map,
1964 				     const union bpf_attr *attr,
1965 				     union bpf_attr __user *uattr)
1966 {
1967 	return __htab_map_lookup_and_delete_batch(map, attr, uattr, true,
1968 						  true, false);
1969 }
1970 
1971 struct bpf_iter_seq_hash_map_info {
1972 	struct bpf_map *map;
1973 	struct bpf_htab *htab;
1974 	void *percpu_value_buf; // non-zero means percpu hash
1975 	u32 bucket_id;
1976 	u32 skip_elems;
1977 };
1978 
1979 static struct htab_elem *
1980 bpf_hash_map_seq_find_next(struct bpf_iter_seq_hash_map_info *info,
1981 			   struct htab_elem *prev_elem)
1982 {
1983 	const struct bpf_htab *htab = info->htab;
1984 	u32 skip_elems = info->skip_elems;
1985 	u32 bucket_id = info->bucket_id;
1986 	struct hlist_nulls_head *head;
1987 	struct hlist_nulls_node *n;
1988 	struct htab_elem *elem;
1989 	struct bucket *b;
1990 	u32 i, count;
1991 
1992 	if (bucket_id >= htab->n_buckets)
1993 		return NULL;
1994 
1995 	/* try to find next elem in the same bucket */
1996 	if (prev_elem) {
1997 		/* no update/deletion on this bucket, prev_elem should be still valid
1998 		 * and we won't skip elements.
1999 		 */
2000 		n = rcu_dereference_raw(hlist_nulls_next_rcu(&prev_elem->hash_node));
2001 		elem = hlist_nulls_entry_safe(n, struct htab_elem, hash_node);
2002 		if (elem)
2003 			return elem;
2004 
2005 		/* not found, unlock and go to the next bucket */
2006 		b = &htab->buckets[bucket_id++];
2007 		rcu_read_unlock();
2008 		skip_elems = 0;
2009 	}
2010 
2011 	for (i = bucket_id; i < htab->n_buckets; i++) {
2012 		b = &htab->buckets[i];
2013 		rcu_read_lock();
2014 
2015 		count = 0;
2016 		head = &b->head;
2017 		hlist_nulls_for_each_entry_rcu(elem, n, head, hash_node) {
2018 			if (count >= skip_elems) {
2019 				info->bucket_id = i;
2020 				info->skip_elems = count;
2021 				return elem;
2022 			}
2023 			count++;
2024 		}
2025 
2026 		rcu_read_unlock();
2027 		skip_elems = 0;
2028 	}
2029 
2030 	info->bucket_id = i;
2031 	info->skip_elems = 0;
2032 	return NULL;
2033 }
2034 
2035 static void *bpf_hash_map_seq_start(struct seq_file *seq, loff_t *pos)
2036 {
2037 	struct bpf_iter_seq_hash_map_info *info = seq->private;
2038 	struct htab_elem *elem;
2039 
2040 	elem = bpf_hash_map_seq_find_next(info, NULL);
2041 	if (!elem)
2042 		return NULL;
2043 
2044 	if (*pos == 0)
2045 		++*pos;
2046 	return elem;
2047 }
2048 
2049 static void *bpf_hash_map_seq_next(struct seq_file *seq, void *v, loff_t *pos)
2050 {
2051 	struct bpf_iter_seq_hash_map_info *info = seq->private;
2052 
2053 	++*pos;
2054 	++info->skip_elems;
2055 	return bpf_hash_map_seq_find_next(info, v);
2056 }
2057 
2058 static int __bpf_hash_map_seq_show(struct seq_file *seq, struct htab_elem *elem)
2059 {
2060 	struct bpf_iter_seq_hash_map_info *info = seq->private;
2061 	u32 roundup_key_size, roundup_value_size;
2062 	struct bpf_iter__bpf_map_elem ctx = {};
2063 	struct bpf_map *map = info->map;
2064 	struct bpf_iter_meta meta;
2065 	int ret = 0, off = 0, cpu;
2066 	struct bpf_prog *prog;
2067 	void __percpu *pptr;
2068 
2069 	meta.seq = seq;
2070 	prog = bpf_iter_get_info(&meta, elem == NULL);
2071 	if (prog) {
2072 		ctx.meta = &meta;
2073 		ctx.map = info->map;
2074 		if (elem) {
2075 			roundup_key_size = round_up(map->key_size, 8);
2076 			ctx.key = elem->key;
2077 			if (!info->percpu_value_buf) {
2078 				ctx.value = elem->key + roundup_key_size;
2079 			} else {
2080 				roundup_value_size = round_up(map->value_size, 8);
2081 				pptr = htab_elem_get_ptr(elem, map->key_size);
2082 				for_each_possible_cpu(cpu) {
2083 					copy_map_value_long(map, info->percpu_value_buf + off,
2084 							    per_cpu_ptr(pptr, cpu));
2085 					check_and_init_map_value(map, info->percpu_value_buf + off);
2086 					off += roundup_value_size;
2087 				}
2088 				ctx.value = info->percpu_value_buf;
2089 			}
2090 		}
2091 		ret = bpf_iter_run_prog(prog, &ctx);
2092 	}
2093 
2094 	return ret;
2095 }
2096 
2097 static int bpf_hash_map_seq_show(struct seq_file *seq, void *v)
2098 {
2099 	return __bpf_hash_map_seq_show(seq, v);
2100 }
2101 
2102 static void bpf_hash_map_seq_stop(struct seq_file *seq, void *v)
2103 {
2104 	if (!v)
2105 		(void)__bpf_hash_map_seq_show(seq, NULL);
2106 	else
2107 		rcu_read_unlock();
2108 }
2109 
2110 static int bpf_iter_init_hash_map(void *priv_data,
2111 				  struct bpf_iter_aux_info *aux)
2112 {
2113 	struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
2114 	struct bpf_map *map = aux->map;
2115 	void *value_buf;
2116 	u32 buf_size;
2117 
2118 	if (map->map_type == BPF_MAP_TYPE_PERCPU_HASH ||
2119 	    map->map_type == BPF_MAP_TYPE_LRU_PERCPU_HASH) {
2120 		buf_size = round_up(map->value_size, 8) * num_possible_cpus();
2121 		value_buf = kmalloc(buf_size, GFP_USER | __GFP_NOWARN);
2122 		if (!value_buf)
2123 			return -ENOMEM;
2124 
2125 		seq_info->percpu_value_buf = value_buf;
2126 	}
2127 
2128 	bpf_map_inc_with_uref(map);
2129 	seq_info->map = map;
2130 	seq_info->htab = container_of(map, struct bpf_htab, map);
2131 	return 0;
2132 }
2133 
2134 static void bpf_iter_fini_hash_map(void *priv_data)
2135 {
2136 	struct bpf_iter_seq_hash_map_info *seq_info = priv_data;
2137 
2138 	bpf_map_put_with_uref(seq_info->map);
2139 	kfree(seq_info->percpu_value_buf);
2140 }
2141 
2142 static const struct seq_operations bpf_hash_map_seq_ops = {
2143 	.start	= bpf_hash_map_seq_start,
2144 	.next	= bpf_hash_map_seq_next,
2145 	.stop	= bpf_hash_map_seq_stop,
2146 	.show	= bpf_hash_map_seq_show,
2147 };
2148 
2149 static const struct bpf_iter_seq_info iter_seq_info = {
2150 	.seq_ops		= &bpf_hash_map_seq_ops,
2151 	.init_seq_private	= bpf_iter_init_hash_map,
2152 	.fini_seq_private	= bpf_iter_fini_hash_map,
2153 	.seq_priv_size		= sizeof(struct bpf_iter_seq_hash_map_info),
2154 };
2155 
2156 static long bpf_for_each_hash_elem(struct bpf_map *map, bpf_callback_t callback_fn,
2157 				   void *callback_ctx, u64 flags)
2158 {
2159 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2160 	struct hlist_nulls_head *head;
2161 	struct hlist_nulls_node *n;
2162 	struct htab_elem *elem;
2163 	u32 roundup_key_size;
2164 	int i, num_elems = 0;
2165 	void __percpu *pptr;
2166 	struct bucket *b;
2167 	void *key, *val;
2168 	bool is_percpu;
2169 	u64 ret = 0;
2170 
2171 	if (flags != 0)
2172 		return -EINVAL;
2173 
2174 	is_percpu = htab_is_percpu(htab);
2175 
2176 	roundup_key_size = round_up(map->key_size, 8);
2177 	/* disable migration so percpu value prepared here will be the
2178 	 * same as the one seen by the bpf program with bpf_map_lookup_elem().
2179 	 */
2180 	if (is_percpu)
2181 		migrate_disable();
2182 	for (i = 0; i < htab->n_buckets; i++) {
2183 		b = &htab->buckets[i];
2184 		rcu_read_lock();
2185 		head = &b->head;
2186 		hlist_nulls_for_each_entry_rcu(elem, n, head, hash_node) {
2187 			key = elem->key;
2188 			if (is_percpu) {
2189 				/* current cpu value for percpu map */
2190 				pptr = htab_elem_get_ptr(elem, map->key_size);
2191 				val = this_cpu_ptr(pptr);
2192 			} else {
2193 				val = elem->key + roundup_key_size;
2194 			}
2195 			num_elems++;
2196 			ret = callback_fn((u64)(long)map, (u64)(long)key,
2197 					  (u64)(long)val, (u64)(long)callback_ctx, 0);
2198 			/* return value: 0 - continue, 1 - stop and return */
2199 			if (ret) {
2200 				rcu_read_unlock();
2201 				goto out;
2202 			}
2203 		}
2204 		rcu_read_unlock();
2205 	}
2206 out:
2207 	if (is_percpu)
2208 		migrate_enable();
2209 	return num_elems;
2210 }
2211 
2212 static u64 htab_map_mem_usage(const struct bpf_map *map)
2213 {
2214 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2215 	u32 value_size = round_up(htab->map.value_size, 8);
2216 	bool prealloc = htab_is_prealloc(htab);
2217 	bool percpu = htab_is_percpu(htab);
2218 	bool lru = htab_is_lru(htab);
2219 	u64 num_entries;
2220 	u64 usage = sizeof(struct bpf_htab);
2221 
2222 	usage += sizeof(struct bucket) * htab->n_buckets;
2223 	usage += sizeof(int) * num_possible_cpus() * HASHTAB_MAP_LOCK_COUNT;
2224 	if (prealloc) {
2225 		num_entries = map->max_entries;
2226 		if (htab_has_extra_elems(htab))
2227 			num_entries += num_possible_cpus();
2228 
2229 		usage += htab->elem_size * num_entries;
2230 
2231 		if (percpu)
2232 			usage += value_size * num_possible_cpus() * num_entries;
2233 		else if (!lru)
2234 			usage += sizeof(struct htab_elem *) * num_possible_cpus();
2235 	} else {
2236 #define LLIST_NODE_SZ sizeof(struct llist_node)
2237 
2238 		num_entries = htab->use_percpu_counter ?
2239 					  percpu_counter_sum(&htab->pcount) :
2240 					  atomic_read(&htab->count);
2241 		usage += (htab->elem_size + LLIST_NODE_SZ) * num_entries;
2242 		if (percpu) {
2243 			usage += (LLIST_NODE_SZ + sizeof(void *)) * num_entries;
2244 			usage += value_size * num_possible_cpus() * num_entries;
2245 		}
2246 	}
2247 	return usage;
2248 }
2249 
2250 BTF_ID_LIST_SINGLE(htab_map_btf_ids, struct, bpf_htab)
2251 const struct bpf_map_ops htab_map_ops = {
2252 	.map_meta_equal = bpf_map_meta_equal,
2253 	.map_alloc_check = htab_map_alloc_check,
2254 	.map_alloc = htab_map_alloc,
2255 	.map_free = htab_map_free,
2256 	.map_get_next_key = htab_map_get_next_key,
2257 	.map_release_uref = htab_map_free_timers,
2258 	.map_lookup_elem = htab_map_lookup_elem,
2259 	.map_lookup_and_delete_elem = htab_map_lookup_and_delete_elem,
2260 	.map_update_elem = htab_map_update_elem,
2261 	.map_delete_elem = htab_map_delete_elem,
2262 	.map_gen_lookup = htab_map_gen_lookup,
2263 	.map_seq_show_elem = htab_map_seq_show_elem,
2264 	.map_set_for_each_callback_args = map_set_for_each_callback_args,
2265 	.map_for_each_callback = bpf_for_each_hash_elem,
2266 	.map_mem_usage = htab_map_mem_usage,
2267 	BATCH_OPS(htab),
2268 	.map_btf_id = &htab_map_btf_ids[0],
2269 	.iter_seq_info = &iter_seq_info,
2270 };
2271 
2272 const struct bpf_map_ops htab_lru_map_ops = {
2273 	.map_meta_equal = bpf_map_meta_equal,
2274 	.map_alloc_check = htab_map_alloc_check,
2275 	.map_alloc = htab_map_alloc,
2276 	.map_free = htab_map_free,
2277 	.map_get_next_key = htab_map_get_next_key,
2278 	.map_release_uref = htab_map_free_timers,
2279 	.map_lookup_elem = htab_lru_map_lookup_elem,
2280 	.map_lookup_and_delete_elem = htab_lru_map_lookup_and_delete_elem,
2281 	.map_lookup_elem_sys_only = htab_lru_map_lookup_elem_sys,
2282 	.map_update_elem = htab_lru_map_update_elem,
2283 	.map_delete_elem = htab_lru_map_delete_elem,
2284 	.map_gen_lookup = htab_lru_map_gen_lookup,
2285 	.map_seq_show_elem = htab_map_seq_show_elem,
2286 	.map_set_for_each_callback_args = map_set_for_each_callback_args,
2287 	.map_for_each_callback = bpf_for_each_hash_elem,
2288 	.map_mem_usage = htab_map_mem_usage,
2289 	BATCH_OPS(htab_lru),
2290 	.map_btf_id = &htab_map_btf_ids[0],
2291 	.iter_seq_info = &iter_seq_info,
2292 };
2293 
2294 /* Called from eBPF program */
2295 static void *htab_percpu_map_lookup_elem(struct bpf_map *map, void *key)
2296 {
2297 	struct htab_elem *l = __htab_map_lookup_elem(map, key);
2298 
2299 	if (l)
2300 		return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
2301 	else
2302 		return NULL;
2303 }
2304 
2305 static void *htab_percpu_map_lookup_percpu_elem(struct bpf_map *map, void *key, u32 cpu)
2306 {
2307 	struct htab_elem *l;
2308 
2309 	if (cpu >= nr_cpu_ids)
2310 		return NULL;
2311 
2312 	l = __htab_map_lookup_elem(map, key);
2313 	if (l)
2314 		return per_cpu_ptr(htab_elem_get_ptr(l, map->key_size), cpu);
2315 	else
2316 		return NULL;
2317 }
2318 
2319 static void *htab_lru_percpu_map_lookup_elem(struct bpf_map *map, void *key)
2320 {
2321 	struct htab_elem *l = __htab_map_lookup_elem(map, key);
2322 
2323 	if (l) {
2324 		bpf_lru_node_set_ref(&l->lru_node);
2325 		return this_cpu_ptr(htab_elem_get_ptr(l, map->key_size));
2326 	}
2327 
2328 	return NULL;
2329 }
2330 
2331 static void *htab_lru_percpu_map_lookup_percpu_elem(struct bpf_map *map, void *key, u32 cpu)
2332 {
2333 	struct htab_elem *l;
2334 
2335 	if (cpu >= nr_cpu_ids)
2336 		return NULL;
2337 
2338 	l = __htab_map_lookup_elem(map, key);
2339 	if (l) {
2340 		bpf_lru_node_set_ref(&l->lru_node);
2341 		return per_cpu_ptr(htab_elem_get_ptr(l, map->key_size), cpu);
2342 	}
2343 
2344 	return NULL;
2345 }
2346 
2347 int bpf_percpu_hash_copy(struct bpf_map *map, void *key, void *value)
2348 {
2349 	struct htab_elem *l;
2350 	void __percpu *pptr;
2351 	int ret = -ENOENT;
2352 	int cpu, off = 0;
2353 	u32 size;
2354 
2355 	/* per_cpu areas are zero-filled and bpf programs can only
2356 	 * access 'value_size' of them, so copying rounded areas
2357 	 * will not leak any kernel data
2358 	 */
2359 	size = round_up(map->value_size, 8);
2360 	rcu_read_lock();
2361 	l = __htab_map_lookup_elem(map, key);
2362 	if (!l)
2363 		goto out;
2364 	/* We do not mark LRU map element here in order to not mess up
2365 	 * eviction heuristics when user space does a map walk.
2366 	 */
2367 	pptr = htab_elem_get_ptr(l, map->key_size);
2368 	for_each_possible_cpu(cpu) {
2369 		copy_map_value_long(map, value + off, per_cpu_ptr(pptr, cpu));
2370 		check_and_init_map_value(map, value + off);
2371 		off += size;
2372 	}
2373 	ret = 0;
2374 out:
2375 	rcu_read_unlock();
2376 	return ret;
2377 }
2378 
2379 int bpf_percpu_hash_update(struct bpf_map *map, void *key, void *value,
2380 			   u64 map_flags)
2381 {
2382 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2383 	int ret;
2384 
2385 	rcu_read_lock();
2386 	if (htab_is_lru(htab))
2387 		ret = __htab_lru_percpu_map_update_elem(map, key, value,
2388 							map_flags, true);
2389 	else
2390 		ret = __htab_percpu_map_update_elem(map, key, value, map_flags,
2391 						    true);
2392 	rcu_read_unlock();
2393 
2394 	return ret;
2395 }
2396 
2397 static void htab_percpu_map_seq_show_elem(struct bpf_map *map, void *key,
2398 					  struct seq_file *m)
2399 {
2400 	struct htab_elem *l;
2401 	void __percpu *pptr;
2402 	int cpu;
2403 
2404 	rcu_read_lock();
2405 
2406 	l = __htab_map_lookup_elem(map, key);
2407 	if (!l) {
2408 		rcu_read_unlock();
2409 		return;
2410 	}
2411 
2412 	btf_type_seq_show(map->btf, map->btf_key_type_id, key, m);
2413 	seq_puts(m, ": {\n");
2414 	pptr = htab_elem_get_ptr(l, map->key_size);
2415 	for_each_possible_cpu(cpu) {
2416 		seq_printf(m, "\tcpu%d: ", cpu);
2417 		btf_type_seq_show(map->btf, map->btf_value_type_id,
2418 				  per_cpu_ptr(pptr, cpu), m);
2419 		seq_puts(m, "\n");
2420 	}
2421 	seq_puts(m, "}\n");
2422 
2423 	rcu_read_unlock();
2424 }
2425 
2426 const struct bpf_map_ops htab_percpu_map_ops = {
2427 	.map_meta_equal = bpf_map_meta_equal,
2428 	.map_alloc_check = htab_map_alloc_check,
2429 	.map_alloc = htab_map_alloc,
2430 	.map_free = htab_map_free,
2431 	.map_get_next_key = htab_map_get_next_key,
2432 	.map_lookup_elem = htab_percpu_map_lookup_elem,
2433 	.map_lookup_and_delete_elem = htab_percpu_map_lookup_and_delete_elem,
2434 	.map_update_elem = htab_percpu_map_update_elem,
2435 	.map_delete_elem = htab_map_delete_elem,
2436 	.map_lookup_percpu_elem = htab_percpu_map_lookup_percpu_elem,
2437 	.map_seq_show_elem = htab_percpu_map_seq_show_elem,
2438 	.map_set_for_each_callback_args = map_set_for_each_callback_args,
2439 	.map_for_each_callback = bpf_for_each_hash_elem,
2440 	.map_mem_usage = htab_map_mem_usage,
2441 	BATCH_OPS(htab_percpu),
2442 	.map_btf_id = &htab_map_btf_ids[0],
2443 	.iter_seq_info = &iter_seq_info,
2444 };
2445 
2446 const struct bpf_map_ops htab_lru_percpu_map_ops = {
2447 	.map_meta_equal = bpf_map_meta_equal,
2448 	.map_alloc_check = htab_map_alloc_check,
2449 	.map_alloc = htab_map_alloc,
2450 	.map_free = htab_map_free,
2451 	.map_get_next_key = htab_map_get_next_key,
2452 	.map_lookup_elem = htab_lru_percpu_map_lookup_elem,
2453 	.map_lookup_and_delete_elem = htab_lru_percpu_map_lookup_and_delete_elem,
2454 	.map_update_elem = htab_lru_percpu_map_update_elem,
2455 	.map_delete_elem = htab_lru_map_delete_elem,
2456 	.map_lookup_percpu_elem = htab_lru_percpu_map_lookup_percpu_elem,
2457 	.map_seq_show_elem = htab_percpu_map_seq_show_elem,
2458 	.map_set_for_each_callback_args = map_set_for_each_callback_args,
2459 	.map_for_each_callback = bpf_for_each_hash_elem,
2460 	.map_mem_usage = htab_map_mem_usage,
2461 	BATCH_OPS(htab_lru_percpu),
2462 	.map_btf_id = &htab_map_btf_ids[0],
2463 	.iter_seq_info = &iter_seq_info,
2464 };
2465 
2466 static int fd_htab_map_alloc_check(union bpf_attr *attr)
2467 {
2468 	if (attr->value_size != sizeof(u32))
2469 		return -EINVAL;
2470 	return htab_map_alloc_check(attr);
2471 }
2472 
2473 static void fd_htab_map_free(struct bpf_map *map)
2474 {
2475 	struct bpf_htab *htab = container_of(map, struct bpf_htab, map);
2476 	struct hlist_nulls_node *n;
2477 	struct hlist_nulls_head *head;
2478 	struct htab_elem *l;
2479 	int i;
2480 
2481 	for (i = 0; i < htab->n_buckets; i++) {
2482 		head = select_bucket(htab, i);
2483 
2484 		hlist_nulls_for_each_entry_safe(l, n, head, hash_node) {
2485 			void *ptr = fd_htab_map_get_ptr(map, l);
2486 
2487 			map->ops->map_fd_put_ptr(map, ptr, false);
2488 		}
2489 	}
2490 
2491 	htab_map_free(map);
2492 }
2493 
2494 /* only called from syscall */
2495 int bpf_fd_htab_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
2496 {
2497 	void **ptr;
2498 	int ret = 0;
2499 
2500 	if (!map->ops->map_fd_sys_lookup_elem)
2501 		return -ENOTSUPP;
2502 
2503 	rcu_read_lock();
2504 	ptr = htab_map_lookup_elem(map, key);
2505 	if (ptr)
2506 		*value = map->ops->map_fd_sys_lookup_elem(READ_ONCE(*ptr));
2507 	else
2508 		ret = -ENOENT;
2509 	rcu_read_unlock();
2510 
2511 	return ret;
2512 }
2513 
2514 /* only called from syscall */
2515 int bpf_fd_htab_map_update_elem(struct bpf_map *map, struct file *map_file,
2516 				void *key, void *value, u64 map_flags)
2517 {
2518 	void *ptr;
2519 	int ret;
2520 	u32 ufd = *(u32 *)value;
2521 
2522 	ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
2523 	if (IS_ERR(ptr))
2524 		return PTR_ERR(ptr);
2525 
2526 	ret = htab_map_update_elem(map, key, &ptr, map_flags);
2527 	if (ret)
2528 		map->ops->map_fd_put_ptr(map, ptr, false);
2529 
2530 	return ret;
2531 }
2532 
2533 static struct bpf_map *htab_of_map_alloc(union bpf_attr *attr)
2534 {
2535 	struct bpf_map *map, *inner_map_meta;
2536 
2537 	inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
2538 	if (IS_ERR(inner_map_meta))
2539 		return inner_map_meta;
2540 
2541 	map = htab_map_alloc(attr);
2542 	if (IS_ERR(map)) {
2543 		bpf_map_meta_free(inner_map_meta);
2544 		return map;
2545 	}
2546 
2547 	map->inner_map_meta = inner_map_meta;
2548 
2549 	return map;
2550 }
2551 
2552 static void *htab_of_map_lookup_elem(struct bpf_map *map, void *key)
2553 {
2554 	struct bpf_map **inner_map  = htab_map_lookup_elem(map, key);
2555 
2556 	if (!inner_map)
2557 		return NULL;
2558 
2559 	return READ_ONCE(*inner_map);
2560 }
2561 
2562 static int htab_of_map_gen_lookup(struct bpf_map *map,
2563 				  struct bpf_insn *insn_buf)
2564 {
2565 	struct bpf_insn *insn = insn_buf;
2566 	const int ret = BPF_REG_0;
2567 
2568 	BUILD_BUG_ON(!__same_type(&__htab_map_lookup_elem,
2569 		     (void *(*)(struct bpf_map *map, void *key))NULL));
2570 	*insn++ = BPF_EMIT_CALL(__htab_map_lookup_elem);
2571 	*insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 2);
2572 	*insn++ = BPF_ALU64_IMM(BPF_ADD, ret,
2573 				offsetof(struct htab_elem, key) +
2574 				round_up(map->key_size, 8));
2575 	*insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
2576 
2577 	return insn - insn_buf;
2578 }
2579 
2580 static void htab_of_map_free(struct bpf_map *map)
2581 {
2582 	bpf_map_meta_free(map->inner_map_meta);
2583 	fd_htab_map_free(map);
2584 }
2585 
2586 const struct bpf_map_ops htab_of_maps_map_ops = {
2587 	.map_alloc_check = fd_htab_map_alloc_check,
2588 	.map_alloc = htab_of_map_alloc,
2589 	.map_free = htab_of_map_free,
2590 	.map_get_next_key = htab_map_get_next_key,
2591 	.map_lookup_elem = htab_of_map_lookup_elem,
2592 	.map_delete_elem = htab_map_delete_elem,
2593 	.map_fd_get_ptr = bpf_map_fd_get_ptr,
2594 	.map_fd_put_ptr = bpf_map_fd_put_ptr,
2595 	.map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
2596 	.map_gen_lookup = htab_of_map_gen_lookup,
2597 	.map_check_btf = map_check_no_btf,
2598 	.map_mem_usage = htab_map_mem_usage,
2599 	BATCH_OPS(htab),
2600 	.map_btf_id = &htab_map_btf_ids[0],
2601 };
2602