xref: /linux/kernel/bpf/arraymap.c (revision 23b0f90ba871f096474e1c27c3d14f455189d2d9)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2011-2014 PLUMgrid, http://plumgrid.com
3  * Copyright (c) 2016,2017 Facebook
4  */
5 #include <linux/bpf.h>
6 #include <linux/btf.h>
7 #include <linux/err.h>
8 #include <linux/slab.h>
9 #include <linux/mm.h>
10 #include <linux/filter.h>
11 #include <linux/perf_event.h>
12 #include <uapi/linux/btf.h>
13 #include <linux/rcupdate_trace.h>
14 #include <linux/btf_ids.h>
15 #include <crypto/sha2.h>
16 
17 #include "map_in_map.h"
18 
19 #define ARRAY_CREATE_FLAG_MASK \
20 	(BPF_F_NUMA_NODE | BPF_F_MMAPABLE | BPF_F_ACCESS_MASK | \
21 	 BPF_F_PRESERVE_ELEMS | BPF_F_INNER_MAP)
22 
23 static void bpf_array_free_percpu(struct bpf_array *array)
24 {
25 	int i;
26 
27 	for (i = 0; i < array->map.max_entries; i++) {
28 		free_percpu(array->pptrs[i]);
29 		cond_resched();
30 	}
31 }
32 
33 static int bpf_array_alloc_percpu(struct bpf_array *array)
34 {
35 	void __percpu *ptr;
36 	int i;
37 
38 	for (i = 0; i < array->map.max_entries; i++) {
39 		ptr = bpf_map_alloc_percpu(&array->map, array->elem_size, 8,
40 					   GFP_USER | __GFP_NOWARN);
41 		if (!ptr) {
42 			bpf_array_free_percpu(array);
43 			return -ENOMEM;
44 		}
45 		array->pptrs[i] = ptr;
46 		cond_resched();
47 	}
48 
49 	return 0;
50 }
51 
52 /* Called from syscall */
53 int array_map_alloc_check(union bpf_attr *attr)
54 {
55 	bool percpu = attr->map_type == BPF_MAP_TYPE_PERCPU_ARRAY;
56 	int numa_node = bpf_map_attr_numa_node(attr);
57 
58 	/* check sanity of attributes */
59 	if (attr->max_entries == 0 || attr->key_size != 4 ||
60 	    attr->value_size == 0 ||
61 	    attr->map_flags & ~ARRAY_CREATE_FLAG_MASK ||
62 	    !bpf_map_flags_access_ok(attr->map_flags) ||
63 	    (percpu && numa_node != NUMA_NO_NODE))
64 		return -EINVAL;
65 
66 	if (attr->map_type != BPF_MAP_TYPE_ARRAY &&
67 	    attr->map_flags & (BPF_F_MMAPABLE | BPF_F_INNER_MAP))
68 		return -EINVAL;
69 
70 	if (attr->map_type != BPF_MAP_TYPE_PERF_EVENT_ARRAY &&
71 	    attr->map_flags & BPF_F_PRESERVE_ELEMS)
72 		return -EINVAL;
73 
74 	/* avoid overflow on round_up(map->value_size) */
75 	if (attr->value_size > INT_MAX)
76 		return -E2BIG;
77 	/* percpu map value size is bound by PCPU_MIN_UNIT_SIZE */
78 	if (percpu && round_up(attr->value_size, 8) > PCPU_MIN_UNIT_SIZE)
79 		return -E2BIG;
80 
81 	return 0;
82 }
83 
84 static struct bpf_map *array_map_alloc(union bpf_attr *attr)
85 {
86 	bool percpu = attr->map_type == BPF_MAP_TYPE_PERCPU_ARRAY;
87 	int numa_node = bpf_map_attr_numa_node(attr);
88 	u32 elem_size, index_mask, max_entries;
89 	bool bypass_spec_v1 = bpf_bypass_spec_v1(NULL);
90 	u64 array_size, mask64;
91 	struct bpf_array *array;
92 
93 	elem_size = round_up(attr->value_size, 8);
94 
95 	max_entries = attr->max_entries;
96 
97 	/* On 32 bit archs roundup_pow_of_two() with max_entries that has
98 	 * upper most bit set in u32 space is undefined behavior due to
99 	 * resulting 1U << 32, so do it manually here in u64 space.
100 	 */
101 	mask64 = fls_long(max_entries - 1);
102 	mask64 = 1ULL << mask64;
103 	mask64 -= 1;
104 
105 	index_mask = mask64;
106 	if (!bypass_spec_v1) {
107 		/* round up array size to nearest power of 2,
108 		 * since cpu will speculate within index_mask limits
109 		 */
110 		max_entries = index_mask + 1;
111 		/* Check for overflows. */
112 		if (max_entries < attr->max_entries)
113 			return ERR_PTR(-E2BIG);
114 	}
115 
116 	array_size = sizeof(*array);
117 	if (percpu) {
118 		array_size += (u64) max_entries * sizeof(void *);
119 	} else {
120 		/* rely on vmalloc() to return page-aligned memory and
121 		 * ensure array->value is exactly page-aligned
122 		 */
123 		if (attr->map_flags & BPF_F_MMAPABLE) {
124 			array_size = PAGE_ALIGN(array_size);
125 			array_size += PAGE_ALIGN((u64) max_entries * elem_size);
126 		} else {
127 			array_size += (u64) max_entries * elem_size;
128 		}
129 	}
130 
131 	/* allocate all map elements and zero-initialize them */
132 	if (attr->map_flags & BPF_F_MMAPABLE) {
133 		void *data;
134 
135 		/* kmalloc'ed memory can't be mmap'ed, use explicit vmalloc */
136 		data = bpf_map_area_mmapable_alloc(array_size, numa_node);
137 		if (!data)
138 			return ERR_PTR(-ENOMEM);
139 		array = data + PAGE_ALIGN(sizeof(struct bpf_array))
140 			- offsetof(struct bpf_array, value);
141 	} else {
142 		array = bpf_map_area_alloc(array_size, numa_node);
143 	}
144 	if (!array)
145 		return ERR_PTR(-ENOMEM);
146 	array->index_mask = index_mask;
147 	array->map.bypass_spec_v1 = bypass_spec_v1;
148 
149 	/* copy mandatory map attributes */
150 	bpf_map_init_from_attr(&array->map, attr);
151 	array->elem_size = elem_size;
152 
153 	if (percpu && bpf_array_alloc_percpu(array)) {
154 		bpf_map_area_free(array);
155 		return ERR_PTR(-ENOMEM);
156 	}
157 
158 	return &array->map;
159 }
160 
161 static void *array_map_elem_ptr(struct bpf_array* array, u32 index)
162 {
163 	return array->value + (u64)array->elem_size * index;
164 }
165 
166 /* Called from syscall or from eBPF program */
167 static void *array_map_lookup_elem(struct bpf_map *map, void *key)
168 {
169 	struct bpf_array *array = container_of(map, struct bpf_array, map);
170 	u32 index = *(u32 *)key;
171 
172 	if (unlikely(index >= array->map.max_entries))
173 		return NULL;
174 
175 	return array->value + (u64)array->elem_size * (index & array->index_mask);
176 }
177 
178 static int array_map_get_hash(struct bpf_map *map, u32 hash_buf_size,
179 			       void *hash_buf)
180 {
181 	struct bpf_array *array = container_of(map, struct bpf_array, map);
182 
183 	sha256(array->value, (u64)array->elem_size * array->map.max_entries,
184 	       hash_buf);
185 	memcpy(array->map.sha, hash_buf, sizeof(array->map.sha));
186 	return 0;
187 }
188 
189 static int array_map_direct_value_addr(const struct bpf_map *map, u64 *imm,
190 				       u32 off)
191 {
192 	struct bpf_array *array = container_of(map, struct bpf_array, map);
193 
194 	if (map->max_entries != 1)
195 		return -ENOTSUPP;
196 	if (off >= map->value_size)
197 		return -EINVAL;
198 
199 	*imm = (unsigned long)array->value;
200 	return 0;
201 }
202 
203 static int array_map_direct_value_meta(const struct bpf_map *map, u64 imm,
204 				       u32 *off)
205 {
206 	struct bpf_array *array = container_of(map, struct bpf_array, map);
207 	u64 base = (unsigned long)array->value;
208 	u64 range = array->elem_size;
209 
210 	if (map->max_entries != 1)
211 		return -ENOTSUPP;
212 	if (imm < base || imm >= base + range)
213 		return -ENOENT;
214 
215 	*off = imm - base;
216 	return 0;
217 }
218 
219 /* emit BPF instructions equivalent to C code of array_map_lookup_elem() */
220 static int array_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
221 {
222 	struct bpf_array *array = container_of(map, struct bpf_array, map);
223 	struct bpf_insn *insn = insn_buf;
224 	u32 elem_size = array->elem_size;
225 	const int ret = BPF_REG_0;
226 	const int map_ptr = BPF_REG_1;
227 	const int index = BPF_REG_2;
228 
229 	if (map->map_flags & BPF_F_INNER_MAP)
230 		return -EOPNOTSUPP;
231 
232 	*insn++ = BPF_ALU64_IMM(BPF_ADD, map_ptr, offsetof(struct bpf_array, value));
233 	*insn++ = BPF_LDX_MEM(BPF_W, ret, index, 0);
234 	if (!map->bypass_spec_v1) {
235 		*insn++ = BPF_JMP_IMM(BPF_JGE, ret, map->max_entries, 4);
236 		*insn++ = BPF_ALU32_IMM(BPF_AND, ret, array->index_mask);
237 	} else {
238 		*insn++ = BPF_JMP_IMM(BPF_JGE, ret, map->max_entries, 3);
239 	}
240 
241 	if (is_power_of_2(elem_size)) {
242 		*insn++ = BPF_ALU64_IMM(BPF_LSH, ret, ilog2(elem_size));
243 	} else {
244 		*insn++ = BPF_ALU64_IMM(BPF_MUL, ret, elem_size);
245 	}
246 	*insn++ = BPF_ALU64_REG(BPF_ADD, ret, map_ptr);
247 	*insn++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1);
248 	*insn++ = BPF_MOV64_IMM(ret, 0);
249 	return insn - insn_buf;
250 }
251 
252 /* Called from eBPF program */
253 static void *percpu_array_map_lookup_elem(struct bpf_map *map, void *key)
254 {
255 	struct bpf_array *array = container_of(map, struct bpf_array, map);
256 	u32 index = *(u32 *)key;
257 
258 	if (unlikely(index >= array->map.max_entries))
259 		return NULL;
260 
261 	return this_cpu_ptr(array->pptrs[index & array->index_mask]);
262 }
263 
264 /* emit BPF instructions equivalent to C code of percpu_array_map_lookup_elem() */
265 static int percpu_array_map_gen_lookup(struct bpf_map *map, struct bpf_insn *insn_buf)
266 {
267 	struct bpf_array *array = container_of(map, struct bpf_array, map);
268 	struct bpf_insn *insn = insn_buf;
269 
270 	if (!bpf_jit_supports_percpu_insn())
271 		return -EOPNOTSUPP;
272 
273 	if (map->map_flags & BPF_F_INNER_MAP)
274 		return -EOPNOTSUPP;
275 
276 	BUILD_BUG_ON(offsetof(struct bpf_array, map) != 0);
277 	*insn++ = BPF_ALU64_IMM(BPF_ADD, BPF_REG_1, offsetof(struct bpf_array, pptrs));
278 
279 	*insn++ = BPF_LDX_MEM(BPF_W, BPF_REG_0, BPF_REG_2, 0);
280 	if (!map->bypass_spec_v1) {
281 		*insn++ = BPF_JMP_IMM(BPF_JGE, BPF_REG_0, map->max_entries, 6);
282 		*insn++ = BPF_ALU32_IMM(BPF_AND, BPF_REG_0, array->index_mask);
283 	} else {
284 		*insn++ = BPF_JMP_IMM(BPF_JGE, BPF_REG_0, map->max_entries, 5);
285 	}
286 
287 	*insn++ = BPF_ALU64_IMM(BPF_LSH, BPF_REG_0, 3);
288 	*insn++ = BPF_ALU64_REG(BPF_ADD, BPF_REG_0, BPF_REG_1);
289 	*insn++ = BPF_LDX_MEM(BPF_DW, BPF_REG_0, BPF_REG_0, 0);
290 	*insn++ = BPF_MOV64_PERCPU_REG(BPF_REG_0, BPF_REG_0);
291 	*insn++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1);
292 	*insn++ = BPF_MOV64_IMM(BPF_REG_0, 0);
293 	return insn - insn_buf;
294 }
295 
296 static void *percpu_array_map_lookup_percpu_elem(struct bpf_map *map, void *key, u32 cpu)
297 {
298 	struct bpf_array *array = container_of(map, struct bpf_array, map);
299 	u32 index = *(u32 *)key;
300 
301 	if (cpu >= nr_cpu_ids)
302 		return NULL;
303 
304 	if (unlikely(index >= array->map.max_entries))
305 		return NULL;
306 
307 	return per_cpu_ptr(array->pptrs[index & array->index_mask], cpu);
308 }
309 
310 int bpf_percpu_array_copy(struct bpf_map *map, void *key, void *value, u64 map_flags)
311 {
312 	struct bpf_array *array = container_of(map, struct bpf_array, map);
313 	u32 index = *(u32 *)key;
314 	void __percpu *pptr;
315 	int cpu, off = 0;
316 	u32 size;
317 
318 	if (unlikely(index >= array->map.max_entries))
319 		return -ENOENT;
320 
321 	/* per_cpu areas are zero-filled and bpf programs can only
322 	 * access 'value_size' of them, so copying rounded areas
323 	 * will not leak any kernel data
324 	 */
325 	size = array->elem_size;
326 	rcu_read_lock();
327 	pptr = array->pptrs[index & array->index_mask];
328 	if (map_flags & BPF_F_CPU) {
329 		cpu = map_flags >> 32;
330 		copy_map_value(map, value, per_cpu_ptr(pptr, cpu));
331 		check_and_init_map_value(map, value);
332 		goto unlock;
333 	}
334 	for_each_possible_cpu(cpu) {
335 		copy_map_value_long(map, value + off, per_cpu_ptr(pptr, cpu));
336 		check_and_init_map_value(map, value + off);
337 		off += size;
338 	}
339 unlock:
340 	rcu_read_unlock();
341 	return 0;
342 }
343 
344 /* Called from syscall */
345 int bpf_array_get_next_key(struct bpf_map *map, void *key, void *next_key)
346 {
347 	u32 index = key ? *(u32 *)key : U32_MAX;
348 	u32 *next = (u32 *)next_key;
349 
350 	if (index >= map->max_entries) {
351 		*next = 0;
352 		return 0;
353 	}
354 
355 	if (index == map->max_entries - 1)
356 		return -ENOENT;
357 
358 	*next = index + 1;
359 	return 0;
360 }
361 
362 /* Called from syscall or from eBPF program */
363 static long array_map_update_elem(struct bpf_map *map, void *key, void *value,
364 				  u64 map_flags)
365 {
366 	struct bpf_array *array = container_of(map, struct bpf_array, map);
367 	u32 index = *(u32 *)key;
368 	char *val;
369 
370 	if (unlikely((map_flags & ~BPF_F_LOCK) > BPF_EXIST))
371 		/* unknown flags */
372 		return -EINVAL;
373 
374 	if (unlikely(index >= array->map.max_entries))
375 		/* all elements were pre-allocated, cannot insert a new one */
376 		return -E2BIG;
377 
378 	if (unlikely(map_flags & BPF_NOEXIST))
379 		/* all elements already exist */
380 		return -EEXIST;
381 
382 	if (unlikely((map_flags & BPF_F_LOCK) &&
383 		     !btf_record_has_field(map->record, BPF_SPIN_LOCK)))
384 		return -EINVAL;
385 
386 	if (array->map.map_type == BPF_MAP_TYPE_PERCPU_ARRAY) {
387 		val = this_cpu_ptr(array->pptrs[index & array->index_mask]);
388 		copy_map_value(map, val, value);
389 		bpf_obj_free_fields(array->map.record, val);
390 	} else {
391 		val = array->value +
392 			(u64)array->elem_size * (index & array->index_mask);
393 		if (map_flags & BPF_F_LOCK)
394 			copy_map_value_locked(map, val, value, false);
395 		else
396 			copy_map_value(map, val, value);
397 		bpf_obj_free_fields(array->map.record, val);
398 	}
399 	return 0;
400 }
401 
402 int bpf_percpu_array_update(struct bpf_map *map, void *key, void *value,
403 			    u64 map_flags)
404 {
405 	struct bpf_array *array = container_of(map, struct bpf_array, map);
406 	u32 index = *(u32 *)key;
407 	void __percpu *pptr;
408 	void *ptr, *val;
409 	u32 size;
410 	int cpu;
411 
412 	if (unlikely((map_flags & BPF_F_LOCK) || (u32)map_flags > BPF_F_ALL_CPUS))
413 		/* unknown flags */
414 		return -EINVAL;
415 
416 	if (unlikely(index >= array->map.max_entries))
417 		/* all elements were pre-allocated, cannot insert a new one */
418 		return -E2BIG;
419 
420 	if (unlikely(map_flags == BPF_NOEXIST))
421 		/* all elements already exist */
422 		return -EEXIST;
423 
424 	/* the user space will provide round_up(value_size, 8) bytes that
425 	 * will be copied into per-cpu area. bpf programs can only access
426 	 * value_size of it. During lookup the same extra bytes will be
427 	 * returned or zeros which were zero-filled by percpu_alloc,
428 	 * so no kernel data leaks possible
429 	 */
430 	size = array->elem_size;
431 	rcu_read_lock();
432 	pptr = array->pptrs[index & array->index_mask];
433 	if (map_flags & BPF_F_CPU) {
434 		cpu = map_flags >> 32;
435 		ptr = per_cpu_ptr(pptr, cpu);
436 		copy_map_value(map, ptr, value);
437 		bpf_obj_free_fields(array->map.record, ptr);
438 		goto unlock;
439 	}
440 	for_each_possible_cpu(cpu) {
441 		ptr = per_cpu_ptr(pptr, cpu);
442 		val = (map_flags & BPF_F_ALL_CPUS) ? value : value + size * cpu;
443 		copy_map_value(map, ptr, val);
444 		bpf_obj_free_fields(array->map.record, ptr);
445 	}
446 unlock:
447 	rcu_read_unlock();
448 	return 0;
449 }
450 
451 /* Called from syscall or from eBPF program */
452 static long array_map_delete_elem(struct bpf_map *map, void *key)
453 {
454 	return -EINVAL;
455 }
456 
457 static void *array_map_vmalloc_addr(struct bpf_array *array)
458 {
459 	return (void *)round_down((unsigned long)array, PAGE_SIZE);
460 }
461 
462 static void array_map_free_internal_structs(struct bpf_map *map)
463 {
464 	struct bpf_array *array = container_of(map, struct bpf_array, map);
465 	int i;
466 
467 	/* We only free internal structs on uref dropping to zero */
468 	if (!bpf_map_has_internal_structs(map))
469 		return;
470 
471 	for (i = 0; i < array->map.max_entries; i++)
472 		bpf_map_free_internal_structs(map, array_map_elem_ptr(array, i));
473 }
474 
475 /* Called when map->refcnt goes to zero, either from workqueue or from syscall */
476 static void array_map_free(struct bpf_map *map)
477 {
478 	struct bpf_array *array = container_of(map, struct bpf_array, map);
479 	int i;
480 
481 	if (!IS_ERR_OR_NULL(map->record)) {
482 		if (array->map.map_type == BPF_MAP_TYPE_PERCPU_ARRAY) {
483 			for (i = 0; i < array->map.max_entries; i++) {
484 				void __percpu *pptr = array->pptrs[i & array->index_mask];
485 				int cpu;
486 
487 				for_each_possible_cpu(cpu) {
488 					bpf_obj_free_fields(map->record, per_cpu_ptr(pptr, cpu));
489 					cond_resched();
490 				}
491 			}
492 		} else {
493 			for (i = 0; i < array->map.max_entries; i++)
494 				bpf_obj_free_fields(map->record, array_map_elem_ptr(array, i));
495 		}
496 	}
497 
498 	if (array->map.map_type == BPF_MAP_TYPE_PERCPU_ARRAY)
499 		bpf_array_free_percpu(array);
500 
501 	if (array->map.map_flags & BPF_F_MMAPABLE)
502 		bpf_map_area_free(array_map_vmalloc_addr(array));
503 	else
504 		bpf_map_area_free(array);
505 }
506 
507 static void array_map_seq_show_elem(struct bpf_map *map, void *key,
508 				    struct seq_file *m)
509 {
510 	void *value;
511 
512 	rcu_read_lock();
513 
514 	value = array_map_lookup_elem(map, key);
515 	if (!value) {
516 		rcu_read_unlock();
517 		return;
518 	}
519 
520 	if (map->btf_key_type_id)
521 		seq_printf(m, "%u: ", *(u32 *)key);
522 	btf_type_seq_show(map->btf, map->btf_value_type_id, value, m);
523 	seq_putc(m, '\n');
524 
525 	rcu_read_unlock();
526 }
527 
528 static void percpu_array_map_seq_show_elem(struct bpf_map *map, void *key,
529 					   struct seq_file *m)
530 {
531 	struct bpf_array *array = container_of(map, struct bpf_array, map);
532 	u32 index = *(u32 *)key;
533 	void __percpu *pptr;
534 	int cpu;
535 
536 	rcu_read_lock();
537 
538 	seq_printf(m, "%u: {\n", *(u32 *)key);
539 	pptr = array->pptrs[index & array->index_mask];
540 	for_each_possible_cpu(cpu) {
541 		seq_printf(m, "\tcpu%d: ", cpu);
542 		btf_type_seq_show(map->btf, map->btf_value_type_id,
543 				  per_cpu_ptr(pptr, cpu), m);
544 		seq_putc(m, '\n');
545 	}
546 	seq_puts(m, "}\n");
547 
548 	rcu_read_unlock();
549 }
550 
551 static int array_map_check_btf(const struct bpf_map *map,
552 			       const struct btf *btf,
553 			       const struct btf_type *key_type,
554 			       const struct btf_type *value_type)
555 {
556 	/* One exception for keyless BTF: .bss/.data/.rodata map */
557 	if (btf_type_is_void(key_type)) {
558 		if (map->map_type != BPF_MAP_TYPE_ARRAY ||
559 		    map->max_entries != 1)
560 			return -EINVAL;
561 
562 		if (BTF_INFO_KIND(value_type->info) != BTF_KIND_DATASEC)
563 			return -EINVAL;
564 
565 		return 0;
566 	}
567 
568 	/*
569 	 * Bpf array can only take a u32 key. This check makes sure
570 	 * that the btf matches the attr used during map_create.
571 	 */
572 	if (!btf_type_is_i32(key_type))
573 		return -EINVAL;
574 
575 	return 0;
576 }
577 
578 static int array_map_mmap(struct bpf_map *map, struct vm_area_struct *vma)
579 {
580 	struct bpf_array *array = container_of(map, struct bpf_array, map);
581 	pgoff_t pgoff = PAGE_ALIGN(sizeof(*array)) >> PAGE_SHIFT;
582 
583 	if (!(map->map_flags & BPF_F_MMAPABLE))
584 		return -EINVAL;
585 
586 	if (vma->vm_pgoff * PAGE_SIZE + (vma->vm_end - vma->vm_start) >
587 	    PAGE_ALIGN((u64)array->map.max_entries * array->elem_size))
588 		return -EINVAL;
589 
590 	return remap_vmalloc_range(vma, array_map_vmalloc_addr(array),
591 				   vma->vm_pgoff + pgoff);
592 }
593 
594 static bool array_map_meta_equal(const struct bpf_map *meta0,
595 				 const struct bpf_map *meta1)
596 {
597 	if (!bpf_map_meta_equal(meta0, meta1))
598 		return false;
599 	return meta0->map_flags & BPF_F_INNER_MAP ? true :
600 	       meta0->max_entries == meta1->max_entries;
601 }
602 
603 struct bpf_iter_seq_array_map_info {
604 	struct bpf_map *map;
605 	void *percpu_value_buf;
606 	u32 index;
607 };
608 
609 static void *bpf_array_map_seq_start(struct seq_file *seq, loff_t *pos)
610 {
611 	struct bpf_iter_seq_array_map_info *info = seq->private;
612 	struct bpf_map *map = info->map;
613 	struct bpf_array *array;
614 	u32 index;
615 
616 	if (info->index >= map->max_entries)
617 		return NULL;
618 
619 	if (*pos == 0)
620 		++*pos;
621 	array = container_of(map, struct bpf_array, map);
622 	index = info->index & array->index_mask;
623 	if (info->percpu_value_buf)
624 		return (void *)(uintptr_t)array->pptrs[index];
625 	return array_map_elem_ptr(array, index);
626 }
627 
628 static void *bpf_array_map_seq_next(struct seq_file *seq, void *v, loff_t *pos)
629 {
630 	struct bpf_iter_seq_array_map_info *info = seq->private;
631 	struct bpf_map *map = info->map;
632 	struct bpf_array *array;
633 	u32 index;
634 
635 	++*pos;
636 	++info->index;
637 	if (info->index >= map->max_entries)
638 		return NULL;
639 
640 	array = container_of(map, struct bpf_array, map);
641 	index = info->index & array->index_mask;
642 	if (info->percpu_value_buf)
643 		return (void *)(uintptr_t)array->pptrs[index];
644 	return array_map_elem_ptr(array, index);
645 }
646 
647 static int __bpf_array_map_seq_show(struct seq_file *seq, void *v)
648 {
649 	struct bpf_iter_seq_array_map_info *info = seq->private;
650 	struct bpf_iter__bpf_map_elem ctx = {};
651 	struct bpf_map *map = info->map;
652 	struct bpf_array *array = container_of(map, struct bpf_array, map);
653 	struct bpf_iter_meta meta;
654 	struct bpf_prog *prog;
655 	int off = 0, cpu = 0;
656 	void __percpu *pptr;
657 	u32 size;
658 
659 	meta.seq = seq;
660 	prog = bpf_iter_get_info(&meta, v == NULL);
661 	if (!prog)
662 		return 0;
663 
664 	ctx.meta = &meta;
665 	ctx.map = info->map;
666 	if (v) {
667 		ctx.key = &info->index;
668 
669 		if (!info->percpu_value_buf) {
670 			ctx.value = v;
671 		} else {
672 			pptr = (void __percpu *)(uintptr_t)v;
673 			size = array->elem_size;
674 			for_each_possible_cpu(cpu) {
675 				copy_map_value_long(map, info->percpu_value_buf + off,
676 						    per_cpu_ptr(pptr, cpu));
677 				check_and_init_map_value(map, info->percpu_value_buf + off);
678 				off += size;
679 			}
680 			ctx.value = info->percpu_value_buf;
681 		}
682 	}
683 
684 	return bpf_iter_run_prog(prog, &ctx);
685 }
686 
687 static int bpf_array_map_seq_show(struct seq_file *seq, void *v)
688 {
689 	return __bpf_array_map_seq_show(seq, v);
690 }
691 
692 static void bpf_array_map_seq_stop(struct seq_file *seq, void *v)
693 {
694 	if (!v)
695 		(void)__bpf_array_map_seq_show(seq, NULL);
696 }
697 
698 static int bpf_iter_init_array_map(void *priv_data,
699 				   struct bpf_iter_aux_info *aux)
700 {
701 	struct bpf_iter_seq_array_map_info *seq_info = priv_data;
702 	struct bpf_map *map = aux->map;
703 	struct bpf_array *array = container_of(map, struct bpf_array, map);
704 	void *value_buf;
705 	u32 buf_size;
706 
707 	if (map->map_type == BPF_MAP_TYPE_PERCPU_ARRAY) {
708 		buf_size = array->elem_size * num_possible_cpus();
709 		value_buf = kmalloc(buf_size, GFP_USER | __GFP_NOWARN);
710 		if (!value_buf)
711 			return -ENOMEM;
712 
713 		seq_info->percpu_value_buf = value_buf;
714 	}
715 
716 	/* bpf_iter_attach_map() acquires a map uref, and the uref may be
717 	 * released before or in the middle of iterating map elements, so
718 	 * acquire an extra map uref for iterator.
719 	 */
720 	bpf_map_inc_with_uref(map);
721 	seq_info->map = map;
722 	return 0;
723 }
724 
725 static void bpf_iter_fini_array_map(void *priv_data)
726 {
727 	struct bpf_iter_seq_array_map_info *seq_info = priv_data;
728 
729 	bpf_map_put_with_uref(seq_info->map);
730 	kfree(seq_info->percpu_value_buf);
731 }
732 
733 static const struct seq_operations bpf_array_map_seq_ops = {
734 	.start	= bpf_array_map_seq_start,
735 	.next	= bpf_array_map_seq_next,
736 	.stop	= bpf_array_map_seq_stop,
737 	.show	= bpf_array_map_seq_show,
738 };
739 
740 static const struct bpf_iter_seq_info iter_seq_info = {
741 	.seq_ops		= &bpf_array_map_seq_ops,
742 	.init_seq_private	= bpf_iter_init_array_map,
743 	.fini_seq_private	= bpf_iter_fini_array_map,
744 	.seq_priv_size		= sizeof(struct bpf_iter_seq_array_map_info),
745 };
746 
747 static long bpf_for_each_array_elem(struct bpf_map *map, bpf_callback_t callback_fn,
748 				    void *callback_ctx, u64 flags)
749 {
750 	u32 i, key, num_elems = 0;
751 	struct bpf_array *array;
752 	bool is_percpu;
753 	u64 ret = 0;
754 	void *val;
755 
756 	cant_migrate();
757 
758 	if (flags != 0)
759 		return -EINVAL;
760 
761 	is_percpu = map->map_type == BPF_MAP_TYPE_PERCPU_ARRAY;
762 	array = container_of(map, struct bpf_array, map);
763 	for (i = 0; i < map->max_entries; i++) {
764 		if (is_percpu)
765 			val = this_cpu_ptr(array->pptrs[i]);
766 		else
767 			val = array_map_elem_ptr(array, i);
768 		num_elems++;
769 		key = i;
770 		ret = callback_fn((u64)(long)map, (u64)(long)&key,
771 				  (u64)(long)val, (u64)(long)callback_ctx, 0);
772 		/* return value: 0 - continue, 1 - stop and return */
773 		if (ret)
774 			break;
775 	}
776 
777 	return num_elems;
778 }
779 
780 static u64 array_map_mem_usage(const struct bpf_map *map)
781 {
782 	struct bpf_array *array = container_of(map, struct bpf_array, map);
783 	bool percpu = map->map_type == BPF_MAP_TYPE_PERCPU_ARRAY;
784 	u32 elem_size = array->elem_size;
785 	u64 entries = map->max_entries;
786 	u64 usage = sizeof(*array);
787 
788 	if (percpu) {
789 		usage += entries * sizeof(void *);
790 		usage += entries * elem_size * num_possible_cpus();
791 	} else {
792 		if (map->map_flags & BPF_F_MMAPABLE) {
793 			usage = PAGE_ALIGN(usage);
794 			usage += PAGE_ALIGN(entries * elem_size);
795 		} else {
796 			usage += entries * elem_size;
797 		}
798 	}
799 	return usage;
800 }
801 
802 BTF_ID_LIST_SINGLE(array_map_btf_ids, struct, bpf_array)
803 const struct bpf_map_ops array_map_ops = {
804 	.map_meta_equal = array_map_meta_equal,
805 	.map_alloc_check = array_map_alloc_check,
806 	.map_alloc = array_map_alloc,
807 	.map_free = array_map_free,
808 	.map_get_next_key = bpf_array_get_next_key,
809 	.map_release_uref = array_map_free_internal_structs,
810 	.map_lookup_elem = array_map_lookup_elem,
811 	.map_update_elem = array_map_update_elem,
812 	.map_delete_elem = array_map_delete_elem,
813 	.map_gen_lookup = array_map_gen_lookup,
814 	.map_direct_value_addr = array_map_direct_value_addr,
815 	.map_direct_value_meta = array_map_direct_value_meta,
816 	.map_mmap = array_map_mmap,
817 	.map_seq_show_elem = array_map_seq_show_elem,
818 	.map_check_btf = array_map_check_btf,
819 	.map_lookup_batch = generic_map_lookup_batch,
820 	.map_update_batch = generic_map_update_batch,
821 	.map_set_for_each_callback_args = map_set_for_each_callback_args,
822 	.map_for_each_callback = bpf_for_each_array_elem,
823 	.map_mem_usage = array_map_mem_usage,
824 	.map_btf_id = &array_map_btf_ids[0],
825 	.iter_seq_info = &iter_seq_info,
826 	.map_get_hash = &array_map_get_hash,
827 };
828 
829 const struct bpf_map_ops percpu_array_map_ops = {
830 	.map_meta_equal = bpf_map_meta_equal,
831 	.map_alloc_check = array_map_alloc_check,
832 	.map_alloc = array_map_alloc,
833 	.map_free = array_map_free,
834 	.map_get_next_key = bpf_array_get_next_key,
835 	.map_lookup_elem = percpu_array_map_lookup_elem,
836 	.map_gen_lookup = percpu_array_map_gen_lookup,
837 	.map_update_elem = array_map_update_elem,
838 	.map_delete_elem = array_map_delete_elem,
839 	.map_lookup_percpu_elem = percpu_array_map_lookup_percpu_elem,
840 	.map_seq_show_elem = percpu_array_map_seq_show_elem,
841 	.map_check_btf = array_map_check_btf,
842 	.map_lookup_batch = generic_map_lookup_batch,
843 	.map_update_batch = generic_map_update_batch,
844 	.map_set_for_each_callback_args = map_set_for_each_callback_args,
845 	.map_for_each_callback = bpf_for_each_array_elem,
846 	.map_mem_usage = array_map_mem_usage,
847 	.map_btf_id = &array_map_btf_ids[0],
848 	.iter_seq_info = &iter_seq_info,
849 };
850 
851 static int fd_array_map_alloc_check(union bpf_attr *attr)
852 {
853 	/* only file descriptors can be stored in this type of map */
854 	if (attr->value_size != sizeof(u32))
855 		return -EINVAL;
856 	/* Program read-only/write-only not supported for special maps yet. */
857 	if (attr->map_flags & (BPF_F_RDONLY_PROG | BPF_F_WRONLY_PROG))
858 		return -EINVAL;
859 	return array_map_alloc_check(attr);
860 }
861 
862 static void fd_array_map_free(struct bpf_map *map)
863 {
864 	struct bpf_array *array = container_of(map, struct bpf_array, map);
865 	int i;
866 
867 	/* make sure it's empty */
868 	for (i = 0; i < array->map.max_entries; i++)
869 		BUG_ON(array->ptrs[i] != NULL);
870 
871 	bpf_map_area_free(array);
872 }
873 
874 static void *fd_array_map_lookup_elem(struct bpf_map *map, void *key)
875 {
876 	return ERR_PTR(-EOPNOTSUPP);
877 }
878 
879 /* only called from syscall */
880 int bpf_fd_array_map_lookup_elem(struct bpf_map *map, void *key, u32 *value)
881 {
882 	void **elem, *ptr;
883 	int ret =  0;
884 
885 	if (!map->ops->map_fd_sys_lookup_elem)
886 		return -ENOTSUPP;
887 
888 	rcu_read_lock();
889 	elem = array_map_lookup_elem(map, key);
890 	if (elem && (ptr = READ_ONCE(*elem)))
891 		*value = map->ops->map_fd_sys_lookup_elem(ptr);
892 	else
893 		ret = -ENOENT;
894 	rcu_read_unlock();
895 
896 	return ret;
897 }
898 
899 /* only called from syscall */
900 int bpf_fd_array_map_update_elem(struct bpf_map *map, struct file *map_file,
901 				 void *key, void *value, u64 map_flags)
902 {
903 	struct bpf_array *array = container_of(map, struct bpf_array, map);
904 	void *new_ptr, *old_ptr;
905 	u32 index = *(u32 *)key, ufd;
906 
907 	if (map_flags != BPF_ANY)
908 		return -EINVAL;
909 
910 	if (index >= array->map.max_entries)
911 		return -E2BIG;
912 
913 	ufd = *(u32 *)value;
914 	new_ptr = map->ops->map_fd_get_ptr(map, map_file, ufd);
915 	if (IS_ERR(new_ptr))
916 		return PTR_ERR(new_ptr);
917 
918 	if (map->ops->map_poke_run) {
919 		mutex_lock(&array->aux->poke_mutex);
920 		old_ptr = xchg(array->ptrs + index, new_ptr);
921 		map->ops->map_poke_run(map, index, old_ptr, new_ptr);
922 		mutex_unlock(&array->aux->poke_mutex);
923 	} else {
924 		old_ptr = xchg(array->ptrs + index, new_ptr);
925 	}
926 
927 	if (old_ptr)
928 		map->ops->map_fd_put_ptr(map, old_ptr, true);
929 	return 0;
930 }
931 
932 static long __fd_array_map_delete_elem(struct bpf_map *map, void *key, bool need_defer)
933 {
934 	struct bpf_array *array = container_of(map, struct bpf_array, map);
935 	void *old_ptr;
936 	u32 index = *(u32 *)key;
937 
938 	if (index >= array->map.max_entries)
939 		return -E2BIG;
940 
941 	if (map->ops->map_poke_run) {
942 		mutex_lock(&array->aux->poke_mutex);
943 		old_ptr = xchg(array->ptrs + index, NULL);
944 		map->ops->map_poke_run(map, index, old_ptr, NULL);
945 		mutex_unlock(&array->aux->poke_mutex);
946 	} else {
947 		old_ptr = xchg(array->ptrs + index, NULL);
948 	}
949 
950 	if (old_ptr) {
951 		map->ops->map_fd_put_ptr(map, old_ptr, need_defer);
952 		return 0;
953 	} else {
954 		return -ENOENT;
955 	}
956 }
957 
958 static long fd_array_map_delete_elem(struct bpf_map *map, void *key)
959 {
960 	return __fd_array_map_delete_elem(map, key, true);
961 }
962 
963 static void *prog_fd_array_get_ptr(struct bpf_map *map,
964 				   struct file *map_file, int fd)
965 {
966 	struct bpf_prog *prog = bpf_prog_get(fd);
967 	bool is_extended;
968 
969 	if (IS_ERR(prog))
970 		return prog;
971 
972 	if (prog->type == BPF_PROG_TYPE_EXT ||
973 	    !bpf_prog_map_compatible(map, prog)) {
974 		bpf_prog_put(prog);
975 		return ERR_PTR(-EINVAL);
976 	}
977 
978 	mutex_lock(&prog->aux->ext_mutex);
979 	is_extended = prog->aux->is_extended;
980 	if (!is_extended)
981 		prog->aux->prog_array_member_cnt++;
982 	mutex_unlock(&prog->aux->ext_mutex);
983 	if (is_extended) {
984 		/* Extended prog can not be tail callee. It's to prevent a
985 		 * potential infinite loop like:
986 		 * tail callee prog entry -> tail callee prog subprog ->
987 		 * freplace prog entry --tailcall-> tail callee prog entry.
988 		 */
989 		bpf_prog_put(prog);
990 		return ERR_PTR(-EBUSY);
991 	}
992 
993 	return prog;
994 }
995 
996 static void prog_fd_array_put_ptr(struct bpf_map *map, void *ptr, bool need_defer)
997 {
998 	struct bpf_prog *prog = ptr;
999 
1000 	mutex_lock(&prog->aux->ext_mutex);
1001 	prog->aux->prog_array_member_cnt--;
1002 	mutex_unlock(&prog->aux->ext_mutex);
1003 	/* bpf_prog is freed after one RCU or tasks trace grace period */
1004 	bpf_prog_put(prog);
1005 }
1006 
1007 static u32 prog_fd_array_sys_lookup_elem(void *ptr)
1008 {
1009 	return ((struct bpf_prog *)ptr)->aux->id;
1010 }
1011 
1012 /* decrement refcnt of all bpf_progs that are stored in this map */
1013 static void bpf_fd_array_map_clear(struct bpf_map *map, bool need_defer)
1014 {
1015 	struct bpf_array *array = container_of(map, struct bpf_array, map);
1016 	int i;
1017 
1018 	for (i = 0; i < array->map.max_entries; i++)
1019 		__fd_array_map_delete_elem(map, &i, need_defer);
1020 }
1021 
1022 static void prog_array_map_seq_show_elem(struct bpf_map *map, void *key,
1023 					 struct seq_file *m)
1024 {
1025 	void **elem, *ptr;
1026 	u32 prog_id;
1027 
1028 	rcu_read_lock();
1029 
1030 	elem = array_map_lookup_elem(map, key);
1031 	if (elem) {
1032 		ptr = READ_ONCE(*elem);
1033 		if (ptr) {
1034 			seq_printf(m, "%u: ", *(u32 *)key);
1035 			prog_id = prog_fd_array_sys_lookup_elem(ptr);
1036 			btf_type_seq_show(map->btf, map->btf_value_type_id,
1037 					  &prog_id, m);
1038 			seq_putc(m, '\n');
1039 		}
1040 	}
1041 
1042 	rcu_read_unlock();
1043 }
1044 
1045 struct prog_poke_elem {
1046 	struct list_head list;
1047 	struct bpf_prog_aux *aux;
1048 };
1049 
1050 static int prog_array_map_poke_track(struct bpf_map *map,
1051 				     struct bpf_prog_aux *prog_aux)
1052 {
1053 	struct prog_poke_elem *elem;
1054 	struct bpf_array_aux *aux;
1055 	int ret = 0;
1056 
1057 	aux = container_of(map, struct bpf_array, map)->aux;
1058 	mutex_lock(&aux->poke_mutex);
1059 	list_for_each_entry(elem, &aux->poke_progs, list) {
1060 		if (elem->aux == prog_aux)
1061 			goto out;
1062 	}
1063 
1064 	elem = kmalloc(sizeof(*elem), GFP_KERNEL);
1065 	if (!elem) {
1066 		ret = -ENOMEM;
1067 		goto out;
1068 	}
1069 
1070 	INIT_LIST_HEAD(&elem->list);
1071 	/* We must track the program's aux info at this point in time
1072 	 * since the program pointer itself may not be stable yet, see
1073 	 * also comment in prog_array_map_poke_run().
1074 	 */
1075 	elem->aux = prog_aux;
1076 
1077 	list_add_tail(&elem->list, &aux->poke_progs);
1078 out:
1079 	mutex_unlock(&aux->poke_mutex);
1080 	return ret;
1081 }
1082 
1083 static void prog_array_map_poke_untrack(struct bpf_map *map,
1084 					struct bpf_prog_aux *prog_aux)
1085 {
1086 	struct prog_poke_elem *elem, *tmp;
1087 	struct bpf_array_aux *aux;
1088 
1089 	aux = container_of(map, struct bpf_array, map)->aux;
1090 	mutex_lock(&aux->poke_mutex);
1091 	list_for_each_entry_safe(elem, tmp, &aux->poke_progs, list) {
1092 		if (elem->aux == prog_aux) {
1093 			list_del_init(&elem->list);
1094 			kfree(elem);
1095 			break;
1096 		}
1097 	}
1098 	mutex_unlock(&aux->poke_mutex);
1099 }
1100 
1101 void __weak bpf_arch_poke_desc_update(struct bpf_jit_poke_descriptor *poke,
1102 				      struct bpf_prog *new, struct bpf_prog *old)
1103 {
1104 	WARN_ON_ONCE(1);
1105 }
1106 
1107 static void prog_array_map_poke_run(struct bpf_map *map, u32 key,
1108 				    struct bpf_prog *old,
1109 				    struct bpf_prog *new)
1110 {
1111 	struct prog_poke_elem *elem;
1112 	struct bpf_array_aux *aux;
1113 
1114 	aux = container_of(map, struct bpf_array, map)->aux;
1115 	WARN_ON_ONCE(!mutex_is_locked(&aux->poke_mutex));
1116 
1117 	list_for_each_entry(elem, &aux->poke_progs, list) {
1118 		struct bpf_jit_poke_descriptor *poke;
1119 		int i;
1120 
1121 		for (i = 0; i < elem->aux->size_poke_tab; i++) {
1122 			poke = &elem->aux->poke_tab[i];
1123 
1124 			/* Few things to be aware of:
1125 			 *
1126 			 * 1) We can only ever access aux in this context, but
1127 			 *    not aux->prog since it might not be stable yet and
1128 			 *    there could be danger of use after free otherwise.
1129 			 * 2) Initially when we start tracking aux, the program
1130 			 *    is not JITed yet and also does not have a kallsyms
1131 			 *    entry. We skip these as poke->tailcall_target_stable
1132 			 *    is not active yet. The JIT will do the final fixup
1133 			 *    before setting it stable. The various
1134 			 *    poke->tailcall_target_stable are successively
1135 			 *    activated, so tail call updates can arrive from here
1136 			 *    while JIT is still finishing its final fixup for
1137 			 *    non-activated poke entries.
1138 			 * 3) Also programs reaching refcount of zero while patching
1139 			 *    is in progress is okay since we're protected under
1140 			 *    poke_mutex and untrack the programs before the JIT
1141 			 *    buffer is freed.
1142 			 */
1143 			if (!READ_ONCE(poke->tailcall_target_stable))
1144 				continue;
1145 			if (poke->reason != BPF_POKE_REASON_TAIL_CALL)
1146 				continue;
1147 			if (poke->tail_call.map != map ||
1148 			    poke->tail_call.key != key)
1149 				continue;
1150 
1151 			bpf_arch_poke_desc_update(poke, new, old);
1152 		}
1153 	}
1154 }
1155 
1156 static void prog_array_map_clear_deferred(struct work_struct *work)
1157 {
1158 	struct bpf_map *map = container_of(work, struct bpf_array_aux,
1159 					   work)->map;
1160 	bpf_fd_array_map_clear(map, true);
1161 	bpf_map_put(map);
1162 }
1163 
1164 static void prog_array_map_clear(struct bpf_map *map)
1165 {
1166 	struct bpf_array_aux *aux = container_of(map, struct bpf_array,
1167 						 map)->aux;
1168 	bpf_map_inc(map);
1169 	schedule_work(&aux->work);
1170 }
1171 
1172 static struct bpf_map *prog_array_map_alloc(union bpf_attr *attr)
1173 {
1174 	struct bpf_array_aux *aux;
1175 	struct bpf_map *map;
1176 
1177 	aux = kzalloc(sizeof(*aux), GFP_KERNEL_ACCOUNT);
1178 	if (!aux)
1179 		return ERR_PTR(-ENOMEM);
1180 
1181 	INIT_WORK(&aux->work, prog_array_map_clear_deferred);
1182 	INIT_LIST_HEAD(&aux->poke_progs);
1183 	mutex_init(&aux->poke_mutex);
1184 
1185 	map = array_map_alloc(attr);
1186 	if (IS_ERR(map)) {
1187 		kfree(aux);
1188 		return map;
1189 	}
1190 
1191 	container_of(map, struct bpf_array, map)->aux = aux;
1192 	aux->map = map;
1193 
1194 	return map;
1195 }
1196 
1197 static void prog_array_map_free(struct bpf_map *map)
1198 {
1199 	struct prog_poke_elem *elem, *tmp;
1200 	struct bpf_array_aux *aux;
1201 
1202 	aux = container_of(map, struct bpf_array, map)->aux;
1203 	list_for_each_entry_safe(elem, tmp, &aux->poke_progs, list) {
1204 		list_del_init(&elem->list);
1205 		kfree(elem);
1206 	}
1207 	kfree(aux);
1208 	fd_array_map_free(map);
1209 }
1210 
1211 /* prog_array->aux->{type,jited} is a runtime binding.
1212  * Doing static check alone in the verifier is not enough.
1213  * Thus, prog_array_map cannot be used as an inner_map
1214  * and map_meta_equal is not implemented.
1215  */
1216 const struct bpf_map_ops prog_array_map_ops = {
1217 	.map_alloc_check = fd_array_map_alloc_check,
1218 	.map_alloc = prog_array_map_alloc,
1219 	.map_free = prog_array_map_free,
1220 	.map_poke_track = prog_array_map_poke_track,
1221 	.map_poke_untrack = prog_array_map_poke_untrack,
1222 	.map_poke_run = prog_array_map_poke_run,
1223 	.map_get_next_key = bpf_array_get_next_key,
1224 	.map_lookup_elem = fd_array_map_lookup_elem,
1225 	.map_delete_elem = fd_array_map_delete_elem,
1226 	.map_fd_get_ptr = prog_fd_array_get_ptr,
1227 	.map_fd_put_ptr = prog_fd_array_put_ptr,
1228 	.map_fd_sys_lookup_elem = prog_fd_array_sys_lookup_elem,
1229 	.map_release_uref = prog_array_map_clear,
1230 	.map_seq_show_elem = prog_array_map_seq_show_elem,
1231 	.map_mem_usage = array_map_mem_usage,
1232 	.map_btf_id = &array_map_btf_ids[0],
1233 };
1234 
1235 static struct bpf_event_entry *bpf_event_entry_gen(struct file *perf_file,
1236 						   struct file *map_file)
1237 {
1238 	struct bpf_event_entry *ee;
1239 
1240 	ee = kzalloc(sizeof(*ee), GFP_KERNEL);
1241 	if (ee) {
1242 		ee->event = perf_file->private_data;
1243 		ee->perf_file = perf_file;
1244 		ee->map_file = map_file;
1245 	}
1246 
1247 	return ee;
1248 }
1249 
1250 static void __bpf_event_entry_free(struct rcu_head *rcu)
1251 {
1252 	struct bpf_event_entry *ee;
1253 
1254 	ee = container_of(rcu, struct bpf_event_entry, rcu);
1255 	fput(ee->perf_file);
1256 	kfree(ee);
1257 }
1258 
1259 static void bpf_event_entry_free_rcu(struct bpf_event_entry *ee)
1260 {
1261 	call_rcu(&ee->rcu, __bpf_event_entry_free);
1262 }
1263 
1264 static void *perf_event_fd_array_get_ptr(struct bpf_map *map,
1265 					 struct file *map_file, int fd)
1266 {
1267 	struct bpf_event_entry *ee;
1268 	struct perf_event *event;
1269 	struct file *perf_file;
1270 	u64 value;
1271 
1272 	perf_file = perf_event_get(fd);
1273 	if (IS_ERR(perf_file))
1274 		return perf_file;
1275 
1276 	ee = ERR_PTR(-EOPNOTSUPP);
1277 	event = perf_file->private_data;
1278 	if (perf_event_read_local(event, &value, NULL, NULL) == -EOPNOTSUPP)
1279 		goto err_out;
1280 
1281 	ee = bpf_event_entry_gen(perf_file, map_file);
1282 	if (ee)
1283 		return ee;
1284 	ee = ERR_PTR(-ENOMEM);
1285 err_out:
1286 	fput(perf_file);
1287 	return ee;
1288 }
1289 
1290 static void perf_event_fd_array_put_ptr(struct bpf_map *map, void *ptr, bool need_defer)
1291 {
1292 	/* bpf_perf_event is freed after one RCU grace period */
1293 	bpf_event_entry_free_rcu(ptr);
1294 }
1295 
1296 static void perf_event_fd_array_release(struct bpf_map *map,
1297 					struct file *map_file)
1298 {
1299 	struct bpf_array *array = container_of(map, struct bpf_array, map);
1300 	struct bpf_event_entry *ee;
1301 	int i;
1302 
1303 	if (map->map_flags & BPF_F_PRESERVE_ELEMS)
1304 		return;
1305 
1306 	rcu_read_lock();
1307 	for (i = 0; i < array->map.max_entries; i++) {
1308 		ee = READ_ONCE(array->ptrs[i]);
1309 		if (ee && ee->map_file == map_file)
1310 			__fd_array_map_delete_elem(map, &i, true);
1311 	}
1312 	rcu_read_unlock();
1313 }
1314 
1315 static void perf_event_fd_array_map_free(struct bpf_map *map)
1316 {
1317 	if (map->map_flags & BPF_F_PRESERVE_ELEMS)
1318 		bpf_fd_array_map_clear(map, false);
1319 	fd_array_map_free(map);
1320 }
1321 
1322 const struct bpf_map_ops perf_event_array_map_ops = {
1323 	.map_meta_equal = bpf_map_meta_equal,
1324 	.map_alloc_check = fd_array_map_alloc_check,
1325 	.map_alloc = array_map_alloc,
1326 	.map_free = perf_event_fd_array_map_free,
1327 	.map_get_next_key = bpf_array_get_next_key,
1328 	.map_lookup_elem = fd_array_map_lookup_elem,
1329 	.map_delete_elem = fd_array_map_delete_elem,
1330 	.map_fd_get_ptr = perf_event_fd_array_get_ptr,
1331 	.map_fd_put_ptr = perf_event_fd_array_put_ptr,
1332 	.map_release = perf_event_fd_array_release,
1333 	.map_check_btf = map_check_no_btf,
1334 	.map_mem_usage = array_map_mem_usage,
1335 	.map_btf_id = &array_map_btf_ids[0],
1336 };
1337 
1338 #ifdef CONFIG_CGROUPS
1339 static void *cgroup_fd_array_get_ptr(struct bpf_map *map,
1340 				     struct file *map_file /* not used */,
1341 				     int fd)
1342 {
1343 	return cgroup_get_from_fd(fd);
1344 }
1345 
1346 static void cgroup_fd_array_put_ptr(struct bpf_map *map, void *ptr, bool need_defer)
1347 {
1348 	/* cgroup_put free cgrp after a rcu grace period */
1349 	cgroup_put(ptr);
1350 }
1351 
1352 static void cgroup_fd_array_free(struct bpf_map *map)
1353 {
1354 	bpf_fd_array_map_clear(map, false);
1355 	fd_array_map_free(map);
1356 }
1357 
1358 const struct bpf_map_ops cgroup_array_map_ops = {
1359 	.map_meta_equal = bpf_map_meta_equal,
1360 	.map_alloc_check = fd_array_map_alloc_check,
1361 	.map_alloc = array_map_alloc,
1362 	.map_free = cgroup_fd_array_free,
1363 	.map_get_next_key = bpf_array_get_next_key,
1364 	.map_lookup_elem = fd_array_map_lookup_elem,
1365 	.map_delete_elem = fd_array_map_delete_elem,
1366 	.map_fd_get_ptr = cgroup_fd_array_get_ptr,
1367 	.map_fd_put_ptr = cgroup_fd_array_put_ptr,
1368 	.map_check_btf = map_check_no_btf,
1369 	.map_mem_usage = array_map_mem_usage,
1370 	.map_btf_id = &array_map_btf_ids[0],
1371 };
1372 #endif
1373 
1374 static struct bpf_map *array_of_map_alloc(union bpf_attr *attr)
1375 {
1376 	struct bpf_map *map, *inner_map_meta;
1377 
1378 	inner_map_meta = bpf_map_meta_alloc(attr->inner_map_fd);
1379 	if (IS_ERR(inner_map_meta))
1380 		return inner_map_meta;
1381 
1382 	map = array_map_alloc(attr);
1383 	if (IS_ERR(map)) {
1384 		bpf_map_meta_free(inner_map_meta);
1385 		return map;
1386 	}
1387 
1388 	map->inner_map_meta = inner_map_meta;
1389 
1390 	return map;
1391 }
1392 
1393 static void array_of_map_free(struct bpf_map *map)
1394 {
1395 	/* map->inner_map_meta is only accessed by syscall which
1396 	 * is protected by fdget/fdput.
1397 	 */
1398 	bpf_map_meta_free(map->inner_map_meta);
1399 	bpf_fd_array_map_clear(map, false);
1400 	fd_array_map_free(map);
1401 }
1402 
1403 static void *array_of_map_lookup_elem(struct bpf_map *map, void *key)
1404 {
1405 	struct bpf_map **inner_map = array_map_lookup_elem(map, key);
1406 
1407 	if (!inner_map)
1408 		return NULL;
1409 
1410 	return READ_ONCE(*inner_map);
1411 }
1412 
1413 static int array_of_map_gen_lookup(struct bpf_map *map,
1414 				   struct bpf_insn *insn_buf)
1415 {
1416 	struct bpf_array *array = container_of(map, struct bpf_array, map);
1417 	u32 elem_size = array->elem_size;
1418 	struct bpf_insn *insn = insn_buf;
1419 	const int ret = BPF_REG_0;
1420 	const int map_ptr = BPF_REG_1;
1421 	const int index = BPF_REG_2;
1422 
1423 	*insn++ = BPF_ALU64_IMM(BPF_ADD, map_ptr, offsetof(struct bpf_array, value));
1424 	*insn++ = BPF_LDX_MEM(BPF_W, ret, index, 0);
1425 	if (!map->bypass_spec_v1) {
1426 		*insn++ = BPF_JMP_IMM(BPF_JGE, ret, map->max_entries, 6);
1427 		*insn++ = BPF_ALU32_IMM(BPF_AND, ret, array->index_mask);
1428 	} else {
1429 		*insn++ = BPF_JMP_IMM(BPF_JGE, ret, map->max_entries, 5);
1430 	}
1431 	if (is_power_of_2(elem_size))
1432 		*insn++ = BPF_ALU64_IMM(BPF_LSH, ret, ilog2(elem_size));
1433 	else
1434 		*insn++ = BPF_ALU64_IMM(BPF_MUL, ret, elem_size);
1435 	*insn++ = BPF_ALU64_REG(BPF_ADD, ret, map_ptr);
1436 	*insn++ = BPF_LDX_MEM(BPF_DW, ret, ret, 0);
1437 	*insn++ = BPF_JMP_IMM(BPF_JEQ, ret, 0, 1);
1438 	*insn++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1);
1439 	*insn++ = BPF_MOV64_IMM(ret, 0);
1440 
1441 	return insn - insn_buf;
1442 }
1443 
1444 const struct bpf_map_ops array_of_maps_map_ops = {
1445 	.map_alloc_check = fd_array_map_alloc_check,
1446 	.map_alloc = array_of_map_alloc,
1447 	.map_free = array_of_map_free,
1448 	.map_get_next_key = bpf_array_get_next_key,
1449 	.map_lookup_elem = array_of_map_lookup_elem,
1450 	.map_delete_elem = fd_array_map_delete_elem,
1451 	.map_fd_get_ptr = bpf_map_fd_get_ptr,
1452 	.map_fd_put_ptr = bpf_map_fd_put_ptr,
1453 	.map_fd_sys_lookup_elem = bpf_map_fd_sys_lookup_elem,
1454 	.map_gen_lookup = array_of_map_gen_lookup,
1455 	.map_lookup_batch = generic_map_lookup_batch,
1456 	.map_update_batch = generic_map_update_batch,
1457 	.map_check_btf = map_check_no_btf,
1458 	.map_mem_usage = array_map_mem_usage,
1459 	.map_btf_id = &array_map_btf_ids[0],
1460 };
1461